Applied Statistics and Probability for Engineers [5th Ed][Montgomery & Runger][2011]

SAVE OFFLINE

JWCL232_fm_i-xvi.qxd

1/23/10

1:18 PM

Page xvi

JWCL232_fm_i-xvi.qxd

1/21/10

7:40 PM

Page i

This online teaching and learning environment integrates the entire digital textbook with the most effective instructor and student resources WR ÀW HYHU\ OHDUQLQJ VW\OH

:LWK WileyPLUS ‡ 6WXGHQWV DFKLHYH FRQFHSW PDVWHU\ LQ D ULFK VWUXFWXUHG HQYLURQPHQW WKDW·V DYDLODEOH 

‡ ,QVWUXFWRUV SHUVRQDOL]H DQG PDQDJH WKHLU FRXUVH PRUH HIIHFWLYHO\ ZLWK DVVHVVPHQW DVVLJQPHQWV JUDGH WUDFNLQJ DQG PRUH

‡ PDQDJH WLPH EHWWHU ‡ VWXG\ VPDUWHU ‡ VDYH PRQH\

From multiple study paths, to self-assessment, to a wealth of interactive visual and audio resources, WileyPLUS gives you everything you need to personalize the teaching and learning experience.

» ) L Q G R X W K RZ W R 0 $ . ( , 7 30 m. The risk of oxygen toxicity is always considered when deep diving is planned. The data shown below demonstrate shortened latencies in a dry atmosphere ( 35? 9-37. Output from a software package is given below:

One-Sample Z:

One-Sample Z:

Test of mu ⫽ 35 ⫽ vs not ⫽ 35 The assumed standard deviation ⫽ 1.8

Test of mu ⫽ 99 vs > 99 The assumed standard deviation ⫽ 2.5 Variable N Mean StDev SE Mean x 12 100.039 2.365 ?

Variable x

N 25

Mean 35.710

StDev 1.475

SE Mean ?

Z ?

P ?

(a) Fill in the missing items. What conclusions would you draw? (b) Is this a one-sided or a two-sided test?

One-Sample Z: Test of mu ⫽ 20 vs > 20 The assumed standard deviation ⫽ 0.75 Variable N Mean StDev SE Mean x 10 19.889 ? 0.237

Z ?

P ?

(a) Fill in the missing items. What conclusions would you draw? (b) Is this a one-sided or a two-sided test? (c) Use the normal table and the above data to construct a 95% two-sided CI on the mean. (d) What would the P-value be if the alternative hypothesis is H1: ␮ ⫽ 20? 9-38. Output from a software package is given below: One-Sample Z: Test of mu ⫽ 14.5 vs > 14.5 The assumed standard deviation ⫽ 1.1 Variable N Mean StDev SE Mean x 16 15.016 1.015 ?

Z ?

P ?

(a) Fill in the missing items. What conclusions would you draw? (b) Is this a one-sided or a two-sided test? (c) Use the normal table and the above data to construct a 95% lower bound on the mean. (d) What would the P-value be if the alternative hypothesis is H1: ␮ ⫽ 14.5? 9-39. Output from a software package is given below:

Z 1.44

P 0.075

(a) Fill in the missing items. What conclusions would you draw? (b) Is this a one-sided or a two-sided test?

JWCL232_c09_283-350.qxd

1/21/10

8:30 PM

Page 309

9-2 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE KNOWN

(c) If the hypothesis had been H0:   98 versus H1:  > 98, would you reject the null hypothesis at the 0.05 level of significance? Can you answer this without referring to the normal table? (d) Use the normal table and the above data to construct a 95% lower bound on the mean. (e) What would the P-value be if the alternative hypothesis is H1:   99? 9-40. The mean water temperature downstream from a power plant cooling tower discharge pipe should be no more than 100°F. Past experience has indicated that the standard deviation of temperature is 2°F. The water temperature is measured on nine randomly chosen days, and the average temperature is found to be 98°F. (a) Is there evidence that the water temperature is acceptable at   0.05? (b) What is the P-value for this test? (c) What is the probability of accepting the null hypothesis at   0.05 if the water has a true mean temperature of 104°F? 9-41. A manufacturer produces crankshafts for an automobile engine. The wear of the crankshaft after 100,000 miles (0.0001 inch) is of interest because it is likely to have an impact on warranty claims. A random sample of n  15 shafts is tested and x  2.78. It is known that   0.9 and that wear is normally distributed. (a) Test H0:   3 versus H1:  Z 3 using   0.05. (b) What is the power of this test if   3.25? (c) What sample size would be required to detect a true mean of 3.75 if we wanted the power to be at least 0.9? 9-42. A melting point test of n  10 samples of a binder used in manufacturing a rocket propellant resulted in x  154.2 F. Assume that the melting point is normally distributed with   1.5 F. (a) Test H0:   155 versus H1:   155 using   0.01. (b) What is the P-value for this test? (c) What is the -error if the true mean is   150? (d) What value of n would be required if we want   0.1 when   150? Assume that   0.01. 9-43. The life in hours of a battery is known to be approximately normally distributed, with standard deviation   1.25 hours. A random sample of 10 batteries has a mean life of x  40.5 hours. (a) Is there evidence to support the claim that battery life exceeds 40 hours? Use   0.05. (b) What is the P-value for the test in part (a)? (c) What is the -error for the test in part (a) if the true mean life is 42 hours? (d) What sample size would be required to ensure that  does not exceed 0.10 if the true mean life is 44 hours? (e) Explain how you could answer the question in part (a) by calculating an appropriate confidence bound on life. 9-44. An engineer who is studying the tensile strength of a steel alloy intended for use in golf club shafts knows that

309

tensile strength is approximately normally distributed with   60 psi. A random sample of 12 specimens has a mean tensile strength of x  3450 psi. (a) Test the hypothesis that mean strength is 3500 psi. Use   0.01. (b) What is the smallest level of significance at which you would be willing to reject the null hypothesis? (c) What is the -error for the test in part (a) if the true mean is 3470? (d) Suppose that we wanted to reject the null hypothesis with probability at least 0.8 if mean strength   3500. What sample size should be used? (e) Explain how you could answer the question in part (a) with a two-sided confidence interval on mean tensile strength. 9-45. Supercavitation is a propulsion technology for undersea vehicles that can greatly increase their speed. It occurs above approximately 50 meters per second, when pressure drops sufficiently to allow the water to dissociate into water vapor, forming a gas bubble behind the vehicle. When the gas bubble completely encloses the vehicle, supercavitation is said to occur. Eight tests were conducted on a scale model of an undersea vehicle in a towing basin with the average observed speed x  102.2 meters per second. Assume that speed is normally distributed with known standard deviation   4 meters per second. (a) Test the hypothesis H0:   100 versus H1:   100 using   0.05. (b) What is the P-value for the test in part (a)? (c) Compute the power of the test if the true mean speed is as low as 95 meters per second. (d) What sample size would be required to detect a true mean speed as low as 95 meters per second if we wanted the power of the test to be at least 0.85? (e) Explain how the question in part (a) could be answered by constructing a one-sided confidence bound on the mean speed. 9-46. A bearing used in an automotive application is supposed to have a nominal inside diameter of 1.5 inches. A random sample of 25 bearings is selected and the average inside diameter of these bearings is 1.4975 inches. Bearing diameter is known to be normally distributed with standard deviation   0.01 inch. (a) Test the hypothesis H0:   1.5 versus H1:   1.5 using   0.01. (b) What is the P-value for the test in part (a)? (c) Compute the power of the test if the true mean diameter is 1.495 inches. (d) What sample size would be required to detect a true mean diameter as low as 1.495 inches if we wanted the power of the test to be at least 0.9? (e) Explain how the question in part (a) could be answered by constructing a two-sided confidence interval on the mean diameter. 9-47. Medical researchers have developed a new artificial heart constructed primarily of titanium and plastic. The heart

JWCL232_c09_283-350.qxd

310

1/14/10

3:07 PM

Page 310

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

will last and operate almost indefinitely once it is implanted in the patient’s body, but the battery pack needs to be recharged about every four hours. A random sample of 50 battery packs is selected and subjected to a life test. The average life of these batteries is 4.05 hours. Assume that battery life is normally distributed with standard deviation  0.2 hour. (a) Is there evidence to support the claim that mean battery life exceeds 4 hours? Use   0.05. (b) What is the P-value for the test in part (a)?

(c) Compute the power of the test if the true mean battery life is 4.5 hours. (d) What sample size would be required to detect a true mean battery life of 4.5 hours if we wanted the power of the test to be at least 0.9? (e) Explain how the question in part (a) could be answered by constructing a one-sided confidence bound on the mean life.

9-3 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE UNKNOWN 9-3.1 Hypothesis Tests on the Mean We now consider the case of hypothesis testing on the mean of a population with unknown variance 2. The situation is analogous to Section 8-2, where we considered a confidence interval on the mean for the same situation. As in that section, the validity of the test procedure we will describe rests on the assumption that the population distribution is at least approximately normal. The important result upon which the test procedure relies is that if X1, X2, p , Xn is a random sample from a normal distribution with mean  and variance 2, the random variable T

X  S 1n

has a t distribution with n 1 degrees of freedom. Recall that we used this result in Section 8-2 to devise the t-confidence interval for . Now consider testing the hypotheses H0:   0 H1:   0 We will use the test statistic:

Test Statistic

T0 

X 0 S  1n

(9-26)

If the null hypothesis is true, T0 has a t distribution with n 1 degrees of freedom. When we know the distribution of the test statistic when H0 is true (this is often called the reference distribution or the null distribution), we can calculate the P-value from this distribution, or, if we use a fixed significance level approach, we can locate the critical region to control the type I error probability at the desired level. To test H0:   0 against the two-sided alternative H1 :  0, the value of the test statistic t0 in Equation 9-26 is calculated, and the P-value is found from the t distribution with n 1 degrees of freedom. Because the test is two-tailed, the P-value is the sum of the probabilities in

JWCL232_c09_283-350.qxd

1/21/10

8:30 PM

Page 311

9-3 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE UNKNOWN Two-tailed test

One-tailed test

tn – 1

–t0

0

t0

(a)

Figure 9-10

One-tailed test

tn – 1

P-value = probability in both tails

311

tn – 1 P-value

0

t0

(b)

t0

0 (c)

Calculating the P-value for a t-test: (a) H1: ␮ Z ␮0: (b) H1: ␮ ⬎ ␮0; (c) H1: ␮ ⬍ ␮0.

the two tails of the t distribution. Refer to Fig. 9-10(a). The P-value is the probability above |t0| plus the probability below ⫺|t0|. Because the t distribution is symmetric around zero, a simple way to write this is P ⫽ 2P1Tn⫺1 ⬎ 冟t0冟2 (9-27) A small P-value is evidence against H0, so if P is of sufficiently small value (typically ⬍ 0.05), reject the null hypothesis. For the one-sided alternative hypotheses H0: ␮ ⫽ ␮0 H1: ␮ ⬎ ␮0

(9-28)

we calculate the test statistic t0 from Equation 9-26 and calculate the P-value as P ⫽ P1Tn⫺1 ⬎ t0 2

(9-29)

For the other one-sided alternative H0: ␮ ⫽ ␮0 H1: ␮ ⬍ ␮0

(9-30)

we calculate the P-value as P ⫽ P1Tn⫺1 ⬍ t0 2

(9-31)

Figure 9-10(b) and (c) show how these P-values are calculated. Statistics software packages calculate and display P-values. However, in working problems by hand, it is useful to be able to find the P-value for a t-test. Because the t-table in Appendix A Table II contains only 10 critical values for each t distribution, determining the exact P-value from this table is usually impossible. Fortunately, it’s easy to find lower and upper bounds on the P-value by using this table. To illustrate, suppose that we are conducting an upper-tailed t-test (so H1: ␮ > ␮0) with 14 degrees of freedom. The relevant critical values from Appendix A Table II are as follows: Critical Value: 0.258 0.692 1.345 1.761 2.145 2.624 2.977 3.326 3.787 4.140 Tail Area: 0.40 0.25 0.10 0.05 0.025 0.01 0.005 0.0025 0.001 0.0005 After calculating the test statistic, we find that t0 ⫽ 2.8. Now, t0 ⫽ 2.8 is between two tabulated values, 2.624 and 2.977. Therefore, the P-value must be between 0.01 and 0.005. Refer to Fig. 9-11. These are effectively the upper and lower bounds on the P-value. This illustrates the procedure for an upper-tailed test. If the test is lower-tailed, just change the sign on the lower and upper bounds for t0 and proceed as above. Remember that for a

JWCL232_c09_283-350.qxd

312

1/14/10

8:45 PM

Page 312

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE t distribution with 14 degrees of freedom P(T14 > 2.624) = 0.01 P(T14 > 2.977) = 0.005

Figure 9-11 P-value for t0  2.8; an upper-tailed test is shown to be between 0.005 and 0.01.

0 2.624 t0 = 2.8 2.977

two-tailed test, the level of significance associated with a particular critical value is twice the corresponding tail area in the column heading. This consideration must be taken into account when we compute the bound on the P-value. For example, suppose that t0  2.8 for a two-tailed alternative based on 14 degrees of freedom. The value of the test statistic t0  2.624 (corresponding to   2  0.01  0.02) and t0  2.977 (corresponding to   2  0.005  0.01), so the lower and upper bounds on the P-value would be 0.01  P  0.02 for this case. Some statistics software packages can help you calculate P-values. For example, Minitab has the capability to find cumulative probabilities from many standard probability distributions, including the t distribution. Simply enter the value of the test statistic t0 along with the appropriate number of degrees of freedom. Minitab will display the probability P(T  t0) where  is the degrees of freedom for the test statistic t0. From the cumulative probability, the P-value can be determined. The single-sample t-test we have just described can also be conducted using the fixed significance level approach. Consider the two-sided alternative hypothesis. The null hypothesis would be rejected if the value of the test statistic t0 falls in the critical region defined by the lower and upper /2 percentage points of the t distribution with n 1 degrees of freedom. That is, reject H0 if t0  t/2,n 1

or t0  t/2,n 1

For the one-tailed tests, the location of the critical region is determined by the direction that the inequality in the alternative hypothesis “points.” So if the alternative is H1:  0, reject H0 if t0  t,n 1 and if the alternative is H1:  0, reject H0 if t0  t,n 1 Figure 9-12 shows the locations of these critical regions.

tn – 1 α /2

α /2

– tα /2, n – 1

tn – 1

0 (a)

tα /2, n – 1

tn – 1 α

0 (b)

tα , n – 1

α

–tα , n – 1

0 (c)

Figure 9-12 The distribution of T0 when H0:  0 is true, with critical region for (a) H1: Z 0, (b) H1:  0, and (c) H1:  0.

T0

JWCL232_c09_283-350.qxd

1/18/10

12:20 PM

Page 313

9-3 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE UNKNOWN

Summary for the OneSample t-Test

313

Testing Hypotheses on the Mean of a Normal Distribution, Variance Unknown Null hypothesis: Test statistic:

H0:   0 T0 

X  0 S 1n Rejection Criterion for Fixed-Level Tests

Alternative Hypotheses

P-Value

H1:  Z 0

Probability above |t0| and probability below |t0| Probability above t0 Probability below t0

H1:   0 H1:   0

t0  t/2,n1 or t0  t/2,n1 t0  t,n1 t0  t,n1

The calculations of the P-values and the locations of the critical regions for these situations are shown in Figs. 9-10 and 9-12, respectively.

EXAMPLE 9-6 Golf Club Design The increased availability of light materials with high strength has revolutionized the design and manufacture of golf clubs, particularly drivers. Clubs with hollow heads and very thin faces can result in much longer tee shots, especially for players of modest skills. This is due partly to the “springlike effect” that the thin face imparts to the ball. Firing a golf ball at the head of the club and measuring the ratio of the outgoing velocity of the ball to the incoming velocity can quantify this spring-like effect. The ratio of velocities is called the coefficient of restitution of the club. An experiment was performed in which 15 drivers produced by a particular club maker were selected at random and their coefficients of restitution measured. In the experiment the golf balls were fired from an air cannon so that the incoming velocity and spin rate of the ball could be precisely controlled. It is of interest to determine if there is evidence (with   0.05) to support a claim that the mean coefficient of restitution exceeds 0.82. The observations follow: 0.8411 0.8580 0.8042

0.8191 0.8532 0.8730

0.8182 0.8483 0.8282

0.8125 0.8276 0.8359

0.8750 0.7983 0.8660

The sample mean and sample standard deviation are x  0.83725 and s  0.02456. The normal probability plot of the data in Fig. 9-13 supports the assumption that the coefficient of restitution is normally distributed. Since the objective of the experimenter is to demonstrate that the mean coefficient of restitution exceeds 0.82, a one-sided alternative hypothesis is appropriate.

The solution using the seven-step procedure for hypothesis testing is as follows: 1. Parameter of interest: The parameter of interest is the mean coefficient of restitution, . 2. Null hypothesis: H0:   0.82 3. Alternative hypothesis: .H1:   0.82 We want to reject H0 if the mean coefficient of restitution exceeds 0.82. 4. Test Statistic: The test statistic is t0 

x  0 s 1n

5. Reject H0 if : Reject H0 if the P-value is less than 0.05. 6. Computations: Since x  0.83725, s  0.02456, 0  0.82, and n  15, we have t0 

0.83725  0.82  2.72 0.02456 115

7. Conclusions: From Appendix A Table II we find, for a t distribution with 14 degrees of freedom, that t0  2.72 falls between two values: 2.624, for which   0.01, and 2.977, for which   0.005. Because this is a one-tailed test, we know that the P-value is between those two values, that is, 0.005  P  0.01. Therefore, since P  0.05, we reject H0 and conclude that the mean coefficient of restitution exceeds 0.82. To use Minitab to compute the P-value, use the Calc

JWCL232_c09_283-350.qxd

314

1/18/10

11:14 AM

Page 314

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

99 95

Percentage

90 80 70 60 50 40 30 20 10 5 1 0.78

0.83 Coefficient of restitution

0.88

Figure 9-13. Normal probability plot of the coefficient of restitution data from Example 9-6. menu and select the probability distribution option. Then, for the t distribution, enter 14 degrees of freedom and the value of the test statistic t0  2.72 as the input constant. Minitab returns the probability P (T14

 2.72)  0.991703. The P-value is P(T14 2.72) or P  1  P(T14  2.72)  1  0.991703  0.008297. Practical Interpretation: There is strong evidence to conclude that the mean coefficient of restitution exceeds 0.82.

Minitab will conduct the one-sample t-test. The output from this software package is in the following display: Minitab Computations One-Sample T: COR Test of mu  0.82 vs mu  0.82 Variable COR

N 15

Mean 0.83725

Variable COR

95.0% Lower Bound 0.82608

StDev 0.02456

SE Mean 0.00634

T 2.72

P 0.008

Notice that Minitab computes both the test statistic T0 and a 95% lower confidence bound for the coefficient of restitution. The reported P-value is 0.008. Because the 95% lower confidence bound exceeds 0.82, we would reject the hypothesis that H0:   0.82 and conclude that the alternative hypothesis H1:   0.82 is true.

9-3.2 Type II Error and Choice of Sample Size The type II error probability for the t-test depends on the distribution of the test statistic in Equation 9-26 when the null hypothesis H0:   0 is false. When the true value of the mean is   0  , the distribution for T0 is called the noncentral t distribution with n  1

JWCL232_c09_283-350.qxd

1/14/10

3:07 PM

Page 315

9-3 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE UNKNOWN

315

degrees of freedom and noncentrality parameter 1n . Note that if   0, the noncentral t distribution reduces to the usual central t distribution. Therefore, the type II error of the twosided alternative (for example) would be  P5 t2,n 1  T0  t2,n 1 0   06  P5 t2,n 1  T 0¿  t2,n 1 6 where T¿0 denotes the noncentral t random variable. Finding the type II error probability for the t-test involves finding the probability contained between two points of the noncentral t distribution. Because the noncentral t-random variable has a messy density function, this integration must be done numerically. Fortunately, this ugly task has already been done, and the results are summarized in a series of O.C. curves in Appendix Charts VIIe, VIIf, VIIg, and VIIh that plot for the t-test against a parameter d for various sample sizes n. Curves are provided for two-sided alternatives on Charts VIIe and VIIf. The abscissa scale factor d on these charts is defined as d

00 0  0 0 



(9-32)

For the one-sided alternative   0 or   0 , we use charts VIG and VIH with d

00 0  0 0 



(9-33)

We note that d depends on the unknown parameter 2. We can avoid this difficulty in several ways. In some cases, we may use the results of a previous experiment or prior information to make a rough initial estimate of 2. If we are interested in evaluating test performance after the data have been collected, we could use the sample variance s2 to estimate

2. If there is no previous experience on which to draw in estimating 2, we then define the difference in the mean d that we wish to detect relative to . For example, if we wish to detect a small difference in the mean, we might use a value of d  0  0   1 (for example), whereas if we are interested in detecting only moderately large differences in the mean, we might select d  0  0   2 (for example). That is, it is the value of the ratio 0  0  that is important in determining sample size, and if it is possible to specify the relative size of the difference in means that we are interested in detecting, then a proper value of d can usually be selected. EXAMPLE 9-7 Golf Club Design Sample Size Consider the golf club testing problem from Example 9-6. If the mean coefficient of restitution exceeds 0.82 by as much as 0.02, is the sample size n  15 adequate to ensure that H0:   0.82 will be rejected with probability at least 0.8? To solve this problem, we will use the sample standard deviation s  0.02456 to estimate . Then d  0  0   0.02 0.02456  0.81. By referring to the operating charac-

teristic curves in Appendix Chart VIIg (for   0.05) with d  0.81 and n  15, we find that  0.10, approximately. Thus, the probability of rejecting H0:   0.82 if the true mean exceeds this by 0.02 is approximately 1  1 0.10  0.90, and we conclude that a sample size of n  15 is adequate to provide the desired sensitivity.

Minitab will also perform power and sample size computations for the one-sample t-test. Below are several calculations based on the golf club testing problem:

JWCL232_c09_283-350.qxd

316

1/14/10

3:07 PM

Page 316

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

Minitab Computations Power and Sample Size 1-Sample t Test Testing mean  null (versus  null) Calculating power for mean  null difference Alpha  0.05 Sigma  0.02456 Difference 0.02

Sample Size 15

Power 0.9117

Power and Sample Size 1-Sample t Test Testing mean  null (versus  null) Calculating power for mean  null difference Alpha  0.05 Sigma  0.02456 Difference 0.01

Sample Size 15

Power 0.4425

Power and Sample Size 1-Sample t Test Testing mean  null (versus  null) Calculating power for mean  null difference Alpha  0.05 Sigma  0.02456 Difference 0.01

Sample Size 39

Target Power 0.8000

Actual Power 0.8029

In the first portion of the computer output, Minitab reproduces the solution to Example 9-7, verifying that a sample size of n  15 is adequate to give power of at least 0.8 if the mean coefficient of restitution exceeds 0.82 by at least 0.02. In the middle section of the output, we used Minitab to compute the power to detect a difference between  and 0  0.82 of 0.01. Notice that with n  15, the power drops considerably to 0.4425. The final portion of the output is the sample size required for a power of at least 0.8 if the difference between  and 0 of interest is actually 0.01. A much larger n is required to detect this smaller difference. EXERCISES FOR SECTION 9-3 9-48. A hypothesis will be used to test that a population mean equals 7 against the alternative that the population mean does not equal 7 with unknown variance . What are the critical values for the test statistic T0 for the following significance levels and sample sizes? (a)   0.01 and n  20 (b)   0.05 and n  12 (c)   0.10 and n  15

9-49. A hypothesis will be used to test that a population mean equals 10 against the alternative that the population mean is greater than 10 with known variance . What is the critical value for the test statistic Z0 for the following significance levels? (a)   0.01 and n  20 (b)   0.05 and n  12 (c)   0.10 and n  15

JWCL232_c09_283-350.qxd

1/14/10

3:07 PM

Page 317

317

9-3 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE UNKNOWN

9-50. A hypothesis will be used to test that a population mean equals 5 against the alternative that the population mean is less than 5 with known variance . What is the critical value for the test statistic Z0 for the following significance levels? (a)   0.01 and n  20 (b)   0.05 and n  12 (c)   0.10 and n  15

9-56. Consider the computer output below.

9-51. For the hypothesis test H0:   7 against H1:   7 with variance unknown and n  20, approximate the P-value for each of the following test statistics. (a) t0  2.05 (b) t0  1.84 (c) t0  0.4

(a) How many degrees of freedom are there on the t-test statistic? (b) Fill in the missing quantities. (c) At what level of significance can the null hypothesis be rejected? (d) If the hypothesis had been H0:   34 versus H1:   34, would the P-value have been larger or smaller? (e) If the hypothesis had been H0:   34.5 versus H1:   34.5, would you have rejected the null hypothesis at the 0.05 level? 9-57. An article in Growth: A Journal Devoted to Problems of Normal and Abnormal Growth [“Comparison of Measured and Estimated Fat-Free Weight, Fat, Potassium and Nitrogen of Growing Guinea Pigs” (Vol. 46, No. 4, 1982, pp. 306–321)] reported the results of a study that measured the body weight (in grams) for guinea pigs at birth.

9-52. For the hypothesis test H0:   10 against H1:   10 with variance unknown and n  15, approximate the P-value for each of the following test statistics. (a) t0  2.05 (b) t0  1.84 (c) t0  0.4 9-53. For the hypothesis test H0:   5 against H1:   5 with variance unknown and n  12, approximate the P-value for each of the following test statistics. (a) t0  2.05 (b) t0  1.84 (c) t0  0.4 9-54. Consider the computer output below. One-Sample T: Test of mu  91 vs  91 95% Lower Variable N Mean StDev SE Mean Bound T P x 20 92.379 0.717 ? ? ? ? (a) Fill in the missing values. You may calculate bounds on the P-value. What conclusions would you draw? (b) Is this a one-sided or a two-sided test? (c) If the hypothesis had been H0:   90 versus H1:  > 90, would your conclusions change? 9-55. Consider the computer output below. One-Sample T: Test of mu  12 vs not  12 Variable x

N 10

Mean 12.564

StDev ?

SE Mean 0.296

T ?

P ?

(a) How many degrees of freedom are there on the t-test statistic? (b) Fill in the missing values. You may calculate bounds on the P-value. What conclusions would you draw? (c) Is this a one-sided or a two-sided test? (d) Construct a 95% two-sided CI on the mean. (e) If the hypothesis had been H0:   12 versus H1:   12, would your conclusions change? (f) If the hypothesis had been H0:   11.5, versus H1:   11.5, would your conclusions change? Answer this question by using the CI computed in part (d).

One-Sample T: Test of mu  34 vs not  34 Variable N x

Mean

StDev SE Mean

16 35.274 1.783

?

95% CI

T

P

(34.324, 36.224) ? 0.012

421.0

452.6

456.1

494.6

373.8

90.5

110.7

96.4

81.7

102.4

241.0

296.0

317.0

290.9

256.5

447.8

687.6

705.7

879.0

88.8

296.0

273.0

268.0

227.5

279.3

258.5

296.0

(a) Test the hypothesis that mean body weight is 300 grams. Use   0.05. (b) What is the smallest level of significance at which you would be willing to reject the null hypothesis? (c) Explain how you could answer the question in part (a) with a two-sided confidence interval on mean body weight. 9-58. An article in the ASCE Journal of Energy Engineering (1999, Vol. 125, pp. 59–75) describes a study of the thermal inertia properties of autoclaved aerated concrete used as a building material. Five samples of the material were tested in a structure, and the average interior temperatures (°C) reported were as follows: 23.01, 22.22, 22.04, 22.62, and 22.59. (a) Test the hypotheses H0:   22.5 versus H1:   22.5, using   0.05. Find the P-value. (b) Check the assumption that interior temperature is normally distributed. (c) Compute the power of the test if the true mean interior temperature is as high as 22.75. (d) What sample size would be required to detect a true mean interior temperature as high as 22.75 if we wanted the power of the test to be at least 0.9?

JWCL232_c09_283-350.qxd

318

1/14/10

3:07 PM

Page 318

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

(e) Explain how the question in part (a) could be answered by constructing a two-sided confidence interval on the mean interior temperature. 9-59. A 1992 article in the Journal of the American Medical Association (“A Critical Appraisal of 98.6 Degrees F, the Upper Limit of the Normal Body Temperature, and Other Legacies of Carl Reinhold August Wunderlich”) reported body temperature, gender, and heart rate for a number of subjects. The body temperatures for 25 female subjects follow: 97.8, 97.2, 97.4, 97.6, 97.8, 97.9, 98.0, 98.0, 98.0, 98.1, 98.2, 98.3, 98.3, 98.4, 98.4, 98.4, 98.5, 98.6, 98.6, 98.7, 98.8, 98.8, 98.9, 98.9, and 99.0. (a) Test the hypothesis H0:   98.6 versus H1:   98.6, using   0.05. Find the P-value. (b) Check the assumption that female body temperature is normally distributed. (c) Compute the power of the test if the true mean female body temperature is as low as 98.0. (d) What sample size would be required to detect a true mean female body temperature as low as 98.2 if we wanted the power of the test to be at least 0.9? (e) Explain how the question in part (a) could be answered by constructing a two-sided confidence interval on the mean female body temperature. 9-60. Cloud seeding has been studied for many decades as a weather modification procedure (for an interesting study of this subject, see the article in Technometrics, “A Bayesian Analysis of a Multiplicative Treatment Effect in Weather Modification,” Vol. 17, pp. 161–166). The rainfall in acre-feet from 20 clouds that were selected at random and seeded with silver nitrate follows: 18.0, 30.7, 19.8, 27.1, 22.3, 18.8, 31.8, 23.4, 21.2, 27.9, 31.9, 27.1, 25.0, 24.7, 26.9, 21.8, 29.2, 34.8, 26.7, and 31.6. (a) Can you support a claim that mean rainfall from seeded clouds exceeds 25 acre-feet? Use   0.01. Find the P-value. (b) Check that rainfall is normally distributed. (c) Compute the power of the test if the true mean rainfall is 27 acre-feet. (d) What sample size would be required to detect a true mean rainfall of 27.5 acre-feet if we wanted the power of the test to be at least 0.9? (e) Explain how the question in part (a) could be answered by constructing a one-sided confidence bound on the mean diameter. 9-61. The sodium content of twenty 300-gram boxes of organic cornflakes was determined. The data (in milligrams) are as follows: 131.15, 130.69, 130.91, 129.54, 129.64, 128.77, 130.72, 128.33, 128.24, 129.65, 130.14, 129.29, 128.71, 129.00, 129.39, 130.42, 129.53, 130.12, 129.78, 130.92. (a) Can you support a claim that mean sodium content of this brand of cornflakes differs from 130 milligrams? Use   0.05. Find the P-value. (b) Check that sodium content is normally distributed. (c) Compute the power of the test if the true mean sodium content is 130.5 milligrams.

(d) What sample size would be required to detect a true mean sodium content of 130.1 milligrams if we wanted the power of the test to be at least 0.75? (e) Explain how the question in part (a) could be answered by constructing a two-sided confidence interval on the mean sodium content. 9-62. Consider the baseball coefficient of restitution data first presented in Exercise 8-92. (a) Do the data support the claim that the mean coefficient of restitution of baseballs exceeds 0.635? Use   0.05. Find the P-value. (b) Check the normality assumption. (c) Compute the power of the test if the true mean coefficient of restitution is as high as 0.64. (d) What sample size would be required to detect a true mean coefficient of restitution as high as 0.64 if we wanted the power of the test to be at least 0.75? (e) Explain how the question in part (a) could be answered with a confidence interval. 9-63. Consider the dissolved oxygen concentration at TVA dams first presented in Exercise 8-94. (a) Test the hypothesis H0:   4 versus H1:   4 . Use   0.01. Find the P-value. (b) Check the normality assumption. (c) Compute the power of the test if the true mean dissolved oxygen concentration is as low as 3. (d) What sample size would be required to detect a true mean dissolved oxygen concentration as low as 2.5 if we wanted the power of the test to be at least 0.9? (e) Explain how the question in part (a) could be answered with a confidence interval. 9-64. Reconsider the data from Medicine and Science in Sports and Exercise described in Exercise 8-30. The sample size was seven and the sample mean and sample standard deviation were 315 watts and 16 watts, respectively. (a) Is there evidence that leg strength exceeds 300 watts at significance level 0.05? Find the P-value. (b) Compute the power of the test if the true strength is 305 watts. (c) What sample size would be required to detect a true mean of 305 watts if the power of the test should be at least 0.90? (d) Explain how the question in part (a) could be answered with a confidence interval. 9-65. Reconsider the tire testing experiment described in Exercise 8-27. (a) The engineer would like to demonstrate that the mean life of this new tire is in excess of 60,000 kilometers. Formulate and test appropriate hypotheses, and draw conclusions using   0.05. (b) Suppose that if the mean life is as long as 61,000 kilometers, the engineer would like to detect this difference with probability at least 0.90. Was the sample size n  16 used in part (a) adequate?

JWCL232_c09_283-350.qxd

1/14/10

3:07 PM

Page 319

319

9-4 TESTS ON THE VARIANCE AND STANDARD DEVIATION OF A NORMAL DISTRIBUTION

9-66. Reconsider the Izod impact test on PVC pipe described in Exercise 8-28. Suppose that you want to use the data from this experiment to support a claim that the mean impact strength exceeds the ASTM standard (one foot-pound per inch). Formulate and test the appropriate hypotheses using   0.05. 9-67. Reconsider the television tube brightness experiment in Exercise 8-35. Suppose that the design engineer claims that this tube will require at least 300 microamps of current to produce the desired brightness level. Formulate and test an appropriate hypothesis to confirm this claim using   0.05. Find the P-value for this test. State any necessary assumptions about the underlying distribution of the data. 9-68. Exercise 6-30 gave data on the heights of female engineering students at ASU. (a) Can you support a claim that the mean height of female engineering students at ASU is at least 65 inches? Use   0.05. Find the P-value. (b) Check the normality assumption. (c) Compute the power of the test if the true mean height is 62 inches. (d) What sample size would be required to detect a true mean height of 64 inches if we wanted the power of the test to be at least 0.8?

9-69. Exercise 6-33 describes testing golf balls for an overall distance standard. (a) Can you support a claim that mean distance achieved by this particular golf ball exceeds 280 yards? Use   0.05. Find the P-value. (b) Check the normality assumption. (c) Compute the power of the test if the true mean distance is 290 yards. (d) What sample size would be required to detect a true mean distance of 290 yards if we wanted the power of the test to be at least 0.8? 9-70. Exercise 6-32 presented data on the concentration of suspended solids in lake water. (a) Test the hypothesis H0:   55 versus H1:   55; use   0.05. Find the P-value. (b) Check the normality assumption. (c) Compute the power of the test if the true mean concentration is as low as 50. (d) What sample size would be required to detect a true mean concentration as low as 50 if we wanted the power of the test to be at least 0.9?

9-4 TESTS ON THE VARIANCE AND STANDARD DEVIATION OF A NORMAL DISTRIBUTION Sometimes hypothesis tests on the population variance or standard deviation are needed. When the population is modeled by a normal distribution, the tests and intervals described in this section are applicable.

9-4.1 Hypothesis Tests on the Variance Suppose that we wish to test the hypothesis that the variance of a normal population 2 equals a specified value, say 20, or equivalently, that the standard deviation is equal to 0. Let X1, X2, p , Xn be a random sample of n observations from this population. To test H0: 2  20 H1: 2  20

(9-34)

we will use the test statistic:

Test Statistic

X 20 

1n 12S 2

02

(9-35)

JWCL232_c09_283-350.qxd

320

1/14/10

3:07 PM

Page 320

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE f (x)

f (x)

f (x)

 2n – 1

 2n – 1

α /2

0

 2n – 1

α /2

 21 – α /2, n – 1

 2α /2, n – 1

α

α

x

 2α , n – 1

0

(a)

x

0

x

 21 – α , n – 1

(b)

(c)

Figure 9-14 Reference distribution for the test of H0: 2  20 with critical region values for (a) H1: 2  20 , (b) H1: 2  20 , and (c) H1: 2  20 .

If the null hypothesis H0: 2  20 is true, the test statistic 20 defined in Equation 9-35 follows the chi-square distribution with n 1 degrees of freedom. This is the reference distribution for this test procedure. To perform a fixed significance level test, we would take a random sample from the population of interest, calculate 20 , the value of the test statistic X 02, and the null hypothesis H0: 2  20 would be rejected if 20  22, n 1 or if 20  21 2,n 1 where 22,n 1 and 21 2,n 1 are the upper and lower 100兾2 percentage points of the chisquare distribution with n 1 degrees of freedom, respectively. Figure 9-14(a) shows the critical region. The same test statistic is used for one-sided alternative hypotheses. For the one-sided hypotheses H0: 2  20 H1: 2  20

(9-33)

we would reject H0 if 20  2,n 1, whereas for the other one-sided hypotheses H0: 2  20 H1: 2  20

(9-34)

we would reject H0 if 20  21 ,n 1. The one-sided critical regions are shown in Fig. 9-14(b) and (c).

Tests on the Variance of a Normal Distribution

Null hypothesis:

H0: 2  20 20 

Test statistic:

1n 12S2

20

Alternative hypothesis H1: Z H1: 2  H1: 2  2

20

20

20

Rejection criteria 20 20 20

  

2/2,n 1 2,n 1

2,n 1

or 20  2/2,n 1

JWCL232_c09_283-350.qxd

1/14/10

3:07 PM

Page 321

9-4 TESTS ON THE VARIANCE AND STANDARD DEVIATION OF A NORMAL DISTRIBUTION

321

EXAMPLE 9-8 Automated Filling 4. Test statistic: The test statistic is

An automated filling machine is used to fill bottles with liquid detergent. A random sample of 20 bottles results in a sample variance of fill volume of s2  0.0153 (fluid ounces)2. If the variance of fill volume exceeds 0.01 (fluid ounces)2, an unacceptable proportion of bottles will be underfilled or overfilled. Is there evidence in the sample data to suggest that the manufacturer has a problem with underfilled or overfilled bottles? Use   0.05, and assume that fill volume has a normal distribution. Using the seven-step procedure results in the following:

20 

1n 12s2

20

5. Reject H0: Use   0.05, and reject 20  20.05,19  30.14.

H0 if

6. Computations: 20 

1. Parameter of Interest: The parameter of interest is the population variance 2.

1910.01532 0.01

 29.07

7. Conclusions: Since 20  29.07  20.05,19  30.14, we conclude that there is no strong evidence that the variance of fill volume exceeds 0.01 (fluid ounces)2. So there is no strong evidence of a problem with incorrectly filled bottles.

2. Null hypothesis: H0: 2  0.01 3. Alternative hypothesis: H1: 2  0.01

We can also use the P-value approach. Using Appendix Table III, it is easy to place bounds on the P-value of a chi-square test. From inspection of the table, we find that 20.10,19  27.20 and 20.05,19  30.14. Since 27.20  29.07  30.14, we conclude that the P-value for the test in Example 9-8 is in the interval 0.05  P-value  0.10. The actual P-value can be computed from Minitab. For 19 degrees of freedom, Minitab calculates the cumulative chi-square probability that is less than or equal to the value of the test statistic as 0.935108 (use the cumulative distribution function in the Calc menu). This is the probability to the left of (or below) 29.07, and the P-value is the probability above or beyond 29.07, or P  1 0.935108  0.064892. The P-value for a lower-tail test would be found as the area (probability) in the lower tail of the chi-square distribution to the left of (or below) the computed value of the test statistic 20. For the two-sided alternative, find the tail area associated with the computed value of the test statistic and double it to obtain the P-value. Minitab will perform the test on a variance of a normal distribution described in this section. The output for Example 9-8 is as follows: Test and CI for One Variance Method Null hypothesis Alternative hypothesis

Sigma-squared  0.01 Sigma-squared > 0.01

Statistics N 20

Variance 0.0153

StDev 0.124

95% One-Sided Confidence Intervals Lower Bound Method for StDev Standard 0.098

Lower Bound for Variance 0.0096

Tests Method Standard

DF 19

Chi-Square 29.07

P-Value 0.065

The standard method that is referred to is the method described in this section. Minitab also has an adjusted method that can be employed with continuous nonnormal distributions.

JWCL232_c09_283-350.qxd

322

1/14/10

3:07 PM

Page 322

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

9-4.2 Type II Error and Choice of Sample Size Operating characteristic curves for the chi-square tests in Section 9-4.1 are provided in Appendix Charts VIi through VIn for   0.05 and   0.01. For the two-sided alternative hypothesis of Equation 9-34, Charts VIIi and VIIj plot against an abscissa parameter

␭

0

(9-38)

for various sample sizes n, where denotes the true value of the standard deviation. Charts VIk and VIl are for the one-sided alternative H1: 2  20, while Charts VIIm and VIIn are for the other one-sided alternative H1: 2  20. In using these charts, we think of as the value of the standard deviation that we want to detect. These curves can be used to evaluate the -error (or power) associated with a particular test. Alternatively, they can be used to design a test—that is, to determine what sample size is necessary to detect a particular value of that differs from the hypothesized value 0.

EXAMPLE 9-9

Automated Filling Sample Size

Consider the bottle-filling problem from Example 9-8. If the variance of the filling process exceeds 0.01 (fluid ounces)2, too many bottles will be underfilled. Thus, the hypothesized value of the standard deviation is 0  0.10. Suppose that if the true standard deviation of the filling process exceeds this value by 25%, we would like to detect this with probability at least 0.8. Is the sample size of n  20 adequate? To solve this problem, note that we require

0.125 ␭   1.25 0 0.10

This is the abscissa parameter for Chart VIIk. From this chart, with n  20 and   1.25, we find that ⯝ 0.6. Therefore, there is only about a 40% chance that the null hypothesis will be rejected if the true standard deviation is really as large as

 0.125 fluid ounce. To reduce the -error, a larger sample size must be used. From the operating characteristic curve with  0.20 and   1.25, we find that n  75, approximately. Thus, if we want the test to perform as required above, the sample size must be at least 75 bottles.

EXERCISES FOR SECTION 9-4 9-71. Consider the test of H0: 2  7 against H1: 2  7. What are the critical values for the test statistic X02 for the following significance levels and sample sizes? (a)   0.01 and n  20 (b)   0.05 and n  12 (c)   0.10 and n  15 9-72. Consider the test of H0: 2  10 against H1: 2  10. What are the critical values for the test statistic X20 for the following significance levels and sample sizes? (a)   0.01 and n  20 (b)   0.05 and n  12 (c)   0.10 and n  15

9-73. Consider the test of H0: 2  5 against H1: 2  5. What are the critical values for the test statistic X20 for the following significance levels and sample sizes? (a)   0.01 and n  20 (b)   0.05 and n  12 (c)   0.10 and n  15 9-74. Consider the hypothesis test of H0: 2  7 against H1: 2  7. Approximate the P-value for each of the following test statistics. (a) x20  25.2 and n  20 (b) x20  15.2 and n  12 2 (c) x0  23.0 and n  15

JWCL232_c09_283-350.qxd

1/14/10

3:07 PM

Page 323

9-5 TESTS ON A POPULATION PROPORTION

9-75. Consider the test of H0: 2  5 against H1: 2  5. Approximate the P-value for each of the following test statistics. (a) x20  25.2 and n  20 (b) x20  15.2 and n  12 (c) x20  4.2 and n  15 9-76. Consider the hypothesis test of H0: 2  10 against H1: 2  10. Approximate the P-value for each of the following test statistics. (a) x20  25.2 and n  20 (b) x20  15.2 and n  12 (c) x20  4.2 and n  15 9-77. The data from Medicine and Science in Sports and Exercise described in Exercise 8-48 considered ice hockey player performance after electrostimulation training. In summary, there were 17 players and the sample standard deviation of performance was 0.09 seconds. (a) Is there strong evidence to conclude that the standard deviation of performance time exceeds the historical value of 0.75 seconds? Use   0.05. Find the P-value for this test. (b) Discuss how part (a) could be answered by constructing a 95% one-sided confidence interval for . 9-78. The data from Technometrics described in Exercise 8-51 considered the variability in repeated measurements of the weight of a sheet of paper. In summary, the sample standard deviation from 15 measurements was 0.0083 grams. (a) Does the measurement standard deviation differ from 0.01 grams at   0.05? Find the P-value for this test. (b) Discuss how part (a) could be answered by constructing a confidence interval for . 9-79. Reconsider the percentage of titanium in an alloy used in aerospace castings from Exercise 8-47. Recall that s  0.37 and n  51. (a) Test the hypothesis H0:  0.25 versus H1:  0.25 using   0.05. State any necessary assumptions about the underlying distribution of the data. Find the P-value. (b) Explain how you could answer the question in part (a) by constructing a 95% two-sided confidence interval for . 9-80. Data from an Izod impact test was described in Exercise 8-28. The sample standard deviation was 0.25 and n  20 specimens were tested. (a) Test the hypothesis that  0.10 against an alternative specifying that  0.10, using   0.01, and draw a

323

conclusion. State any necessary assumptions about the underlying distribution of the data. (b) What is the P-value for this test? (c) Could the question in part (a) have been answered by constructing a 99% two-sided confidence interval for 2? 9-81. Data for tire life was described in Exercise 8-27. The sample standard deviation was 3645.94 kilometers and n  16. (a) Can you conclude, using   0.05, that the standard deviation of tire life is less than 4000 kilometers? State any necessary assumptions about the underlying distribution of the data. Find the P-value for this test. (b) Explain how you could answer the question in part (a) by constructing a 95% one-sided confidence interval for . 9-82. If the standard deviation of hole diameter exceeds 0.01 millimeters, there is an unacceptably high probability that the rivet will not fit. Suppose that n  15 and s  0.008 millimeter. (a) Is there strong evidence to indicate that the standard deviation of hole diameter exceeds 0.01 millimeter? Use   0.01. State any necessary assumptions about the underlying distribution of the data. Find the P-value for this test. (b) Suppose that the actual standard deviation of hole diameter exceeds the hypothesized value by 50%. What is the probability that this difference will be detected by the test described in part (a)? (c) If is really as large as 0.0125 millimeters, what sample size will be required to detect this with power of at least 0.8? 9-83. Recall the sugar content of the syrup in canned peaches from Exercise 8-46. Suppose that the variance is thought to be 2  18 (milligrams)2. Recall that a random sample of n  10 cans yields a sample standard deviation of s  4.8 milligrams. (a) Test the hypothesis H0: 2  18 versus H1: 2  18 using   0.05. Find the P-value for this test. (b) Suppose that the actual standard deviation is twice as large as the hypothesized value. What is the probability that this difference will be detected by the test described in part (a)? (c) Suppose that the true variance is 2  40. How large a sample would be required to detect this difference with probability at least 0.90?

9-5 TESTS ON A POPULATION PROPORTION It is often necessary to test hypotheses on a population proportion. For example, suppose that a random sample of size n has been taken from a large (possibly infinite) population and that X( n) observations in this sample belong to a class of interest. Then Pˆ  Xn is a point estimator of the proportion of the population p that belongs to this class. Note that n and p are the parameters of a binomial distribution. Furthermore, from Chapter 7 we know that the sampling distribution of Pˆ is approximately normal with mean p and variance p(1 p)兾n, if

JWCL232_c09_283-350.qxd

324

1/14/10

8:56 PM

Page 324

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

p is not too close to either 0 or 1 and if n is relatively large. Typically, to apply this approximation we require that np and n(1 ⫺ p) be greater than or equal to 5. We will give a largesample test that makes use of the normal approximation to the binomial distribution.

9-5.1 Large-Sample Tests on a Proportion In many engineering problems, we are concerned with a random variable that follows the binomial distribution. For example, consider a production process that manufactures items that are classified as either acceptable or defective. It is usually reasonable to model the occurrence of defectives with the binomial distribution, where the binomial parameter p represents the proportion of defective items produced. Consequently, many engineering decision problems involve hypothesis testing about p. We will consider testing H0: p ⫽ p0 H1: p ⫽ p0

(9-39)

An approximate test based on the normal approximation to the binomial will be given. As noted above, this approximate procedure will be valid as long as p is not extremely close to zero or one, and if the sample size is relatively large. Let X be the number of observations in a random sample of size n that belongs to the class associated with p. Then, if the null hypothesis H0: p ⫽ p0 is true, we have X ⬃ N[np0, np0(1 ⫺ p0)], approximately. To test H0: p ⫽ p0, calculate the test statistic Test Statistic

Z0 ⫽

X ⫺ np0 1np0 11 ⫺ p0 2

(9-40)

and determine the P-value. Because the test statistic follows a standard normal distribution if H0 is true, the P-value is calculated exactly like the P-value for the z-tests in Section 9-2. So for the two-sided alternative hypothesis, the P-value is the sum of the probability in the standard normal distribution above |z0| and the probability below the negative value ⫺|z0|, or P ⫽ 231 ⫺ ⌽1|z0|2 4 For the one-sided alternative hypothesis H0: p > p0, the P-value is the probability above z0, or P ⫽ 1 ⫺ ⌽1z0 2

and for the one-sided alternative hypothesis H0: p < p0, the P-value is the probability below z0, or P ⫽ ⌽1z0 2 We can also perform a fixed-significance-level test. For the two-sided alternative hypothesis, we would reject H0: p ⫽ p0 if z0 ⬎ z␣/2 or z0 ⬍ ⫺z␣/2 Critical regions for the one-sided alternative hypotheses would be constructed in the usual manner.

JWCL232_c09_283-350.qxd

1/18/10

12:29 PM

Page 325

325

9-5 TESTS ON A POPULATION PROPORTION

Summary of Approximate Tests on a Binomial Proportion

EXAMPLE 9-10

Testing Hypotheses on a Binomial Proportion Null hypotheses:

H0: p ⫽ p0

Test statistic:

Z0 ⫽

X ⫺ np0

2np0 11 ⫺ p0 2

Rejection Criterion for Fixed-Level Tests

Alternative Hypotheses

P-Value

H1: p ⫽ p0

Probability above |z0| and probability below ⫺|z0| P ⫽ 231 ⫺ ⌽1z0 2

z0 > z␣/2 or z0 < ⫺z␣/2

H1: p > p0

Probability above z0,

z0 > z␣

H1: p < p0

Probability below z0,

z0 < ⫺ z␣

P ⫽ 1 ⫺ ⌽1z0 2 P ⫽ ⌽1z0 2

Automobile Engine Controller

A semiconductor manufacturer produces controllers used in automobile engine applications. The customer requires that the process fallout or fraction defective at a critical manufacturing step not exceed 0.05 and that the manufacturer demonstrate process capability at this level of quality using ␣ ⫽ 0.05. The semiconductor manufacturer takes a random sample of 200 devices and finds that four of them are defective. Can the manufacturer demonstrate process capability for the customer? We may solve this problem using the seven-step hypothesis-testing procedure as follows:

4.

z0 ⫽

x ⫺ np0 1np0 11 ⫺ p0 2

where x ⫽ 4, n ⫽ 200, and p0 ⫽ 0.05. 5. Reject H0 if: Reject H0: p ⫽ 0.05 if the p-value is less than 0.05. 6. Computations: The test statistic is z0 ⫽

1. Parameter of Interest: The parameter of interest is the process fraction defective p.

4 ⫺ 20010.052 120010.05210.952

⫽ ⫺1.95

7. Conclusions: Since z0 ⫽ ⫺1.95, the P-value is ⌽(⫺1.95) ⫽ 0.0256, so we reject H0 and conclude that the process fraction defective p is less than 0.05.

2. Null hypothesis: H0: p ⫽ 0.05 3. Alternative hypothesis: H1: p ⬍ 0.05 This formulation of the problem will allow the manufacturer to make a strong claim about process capa bility if the null hypothesis H0: p ⫽ 0.05 is rejected.

The test statistic is (from Equation 9-40)

Practical Interpretation: We conclude that the process is capable.

Another form of the test statistic Z0 in Equation 9-40 is occasionally encountered. Note that if X is the number of observations in a random sample of size n that belongs to a class of interest, then Pˆ ⫽ XⲐn is the sample proportion that belongs to that class. Now divide both numerator and denominator of Z0 in Equation 9-40 by n, giving Z0 ⫽

XⲐn ⫺ p0 1p0 11 ⫺ p0 2 Ⲑn

Z0 ⫽

Pˆ ⫺ p0 1p0 11 ⫺ p0 2 Ⲑn

or (9-41)

JWCL232_c09_283-350.qxd

326

1/14/10

3:07 PM

Page 326

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

This presents the test statistic in terms of the sample proportion instead of the number of items X in the sample that belongs to the class of interest. Minitab can be used to perform the test on a binomial proportion. The following Minitab output shows the results for Example 9-10. Test and CI for One Proportion Test of p  0.05 vs p  0.05 Sample

X

N

Sample p

95% Upper Bound

Z-Value

P-Value

1

4

200

0.020000

0.036283

1.95

0.026

* Note * The normal approximation may be inaccurate for small samples.

This output also shows a 95% one-sided upper-confidence bound on P. In Section 8-4 we showed how CIs on a binomial proportion are computed. This Minitab display shows the result of using the normal approximation for tests and CIs. When the sample size is small, this may be inappropriate. Small Sample Tests on a Binomial Proportion Tests on a proportion when the sample size n is small are based on the binomial distribution, not the normal approximation to the binomial. To illustrate, suppose we wish to test H0: p < p0. Let X be the number of successes in the sample. The P-value for this test would be found from the lower tail of a binomial distribution with parameters n and p0. Specifically, the P-value would be the probability that a binomial random variable with parameters n and p0 is less than or equal to X. P-values for the upper-tail one-sided test and the two-sided alternative are computed similarly. Minitab will calculate the exact P-value for a binomial test. The output below contains the exact P-value results for Example 9-10. Test of p  0.05 vs p  0.05 Sample

X

N

Sample p

95% Upper Bound

Exact P-Value

1

4

200

0.020000

0.045180

0.026

The P-value is the same as that reported for the normal approximation, because the sample size is fairly large. Notice that the CI is different from the one found using the normal approximation.

9-5.2 Type II Error and Choice of Sample Size It is possible to obtain closed-form equations for the approximate -error for the tests in Section 9-5.1. Suppose that p is the true value of the population proportion. The approximate -error for the two-sided alternative H1 : p  p0 is  a

p0 p z 2 1p0 11 p0 2 n p0 p z 2 1p0 11 p0 2 n b a b 1p11 p2 n 1p11 p2 n

(9-42)

If the alternative is H1: p  p0,  1 a

p0 p z 1p0 11 p0 2 n b 1p11 p2 n

(9-43)

JWCL232_c09_283-350.qxd

1/14/10

3:07 PM

Page 327

327

9-5 TESTS ON A POPULATION PROPORTION

whereas if the alternative is H1: p  p0,  a

p0 p z 1p0 11 p0 2 n b 1p11 p2 n

(9-44)

These equations can be solved to find the approximate sample size n that gives a test of level  that has a specified risk. The sample size equations are Approximate Sample Size for a Two-Sided Test on a Binomial Proportion

n c

z2 1p0 11 p0 2 z 1p11 p2 2 d p p0

(9-45)

for a two-sided alternative and

Approximate Sample Size for a One-Sided Test on a Binomial Proportion

n c

z 1p0 11 p0 2 z 1p11 p2 2 d p p0

(9-46)

for a one-sided alternative. EXAMPLE 9-11

Automobile Engine Controller Type II Error

Consider the semiconductor manufacturer from Example 9-10. Suppose that its process fallout is really p  0.03. What is the -error for a test of process capability that uses n  200 and   0.05? The -error can be computed using Equation 9-43 as follows: 1 c

0.05 0.03 11.6452 10.0510.952 200 10.0311 0.032 200

d

 1 1 0.442  0.67

Suppose that the semiconductor manufacturer was willing to accept a -error as large as 0.10 if the true value of the process fraction defective was p  0.03. If the manufacturer continues to use   0.05, what sample size would be required? The required sample size can be computed from Equation 9-46 as follows: n c

1.64510.0510.952 1.2810.0310.972 0.03 0.05

d

2

⯝ 832 Thus, the probability is about 0.7 that the semiconductor manufacturer will fail to conclude that the process is capable if the true process fraction defective is p  0.03 (3%). That is, the power of the test against this particular alternative is only about 0.3. This appears to be a large -error (or small power), but the difference between p  0.05 and p  0.03 is fairly small, and the sample size n  200 is not particularly large.

where we have used p  0.03 in Equation 9-46. Conclusion: Note that n  832 is a very large sample size. However, we are trying to detect a fairly small deviation from the null value p0  0.05.

Minitab will also perform power and sample size calculations for the one-sample Z-test on a proportion. Output from Minitab for the engine controllers tested in Example 9-10 follows.

JWCL232_c09_283-350.qxd

328

1/14/10

3:07 PM

Page 328

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

Power and Sample Size Test for One Proportion Testing proportion  0.05 (versus  0.05) Alpha  0.05 Alternative Proportion 3.00E-02

Sample Size 200

Power 0.3287

Power and Sample Size Test for One Proportion Testing proportion  0.05 (versus  0.05) Alpha  0.05 Alternative Proportion 3.00E-02

Sample Size 833

Target Power 0.9000

Actual Power 0.9001

Power and Sample Size Test for One Proportion Testing proportion  0.05 (versus  0.05) Alpha  0.05 Alternative Proportion 3.00E-02

Sample Size 561

Target Power 0.7500

Actual Power 0.7503

The first part of the output shows the power calculation based on the situation described in Example 9-11, where the true proportion is really 0.03. The power calculation from Minitab agrees with the results from Equation 9-43 in Example 9-11. The second part of the output computes the sample size necessary for a power of 0.9 (  0.1) if p  0.03. Again, the results agree closely with those obtained from Equation 9-46. The final portion of the display shows the sample size that would be required if p  0.03 and the power requirement is relaxed to 0.75. Notice that the sample size of n  561 is still quite large because the difference between p  0.05 and p  0.03 is fairly small. EXERCISES FOR SECTION 9-5 9-84. Consider the computer output below.

9-85. Consider the computer output below.

Test and Cl for One Proportion Test of p  0.4 vs p not  0.4

Test and Cl for One Proportion Test of p  0.6 vs p < 0.6

X 98

N 275

Sample p ?

95% CI (0.299759, 0.412968)

Z-Value ?

P-Value ?

Using the normal approximation. (a) Is this a one-sided or a two-sided test? (b) Complete the missing items. (c) The normal approximation was used in the problem. Was that appropriate?

X 287

N 500

Sample p ?

95% Upper Bound ?

Z-Value ?

P-Value ?

(a) Is this a one-sided or a two-sided test? (b) Is this a test based on the normal approximation? Is that appropriate? (c) Complete the missing items. (d) Suppose that the alternative hypothesis was two-sided. What is the P-value for this situation?

JWCL232_c09_283-350.qxd

1/14/10

3:07 PM

Page 329

9-6 SUMMARY TABLE OF INFERENCE PROCEDURES FOR A SINGLE SAMPLE

9-86. Suppose that 1000 customers are surveyed and 850 are satisfied or very satisfied with a corporation’s products and services. (a) Test the hypothesis H0: p  0.9 against H1: p  0.9 at   0.05. Find the P-value. (b) Explain how the question in part (a) could be answered by constructing a 95% two-sided confidence interval for p. 9-87. Suppose that 500 parts are tested in manufacturing and 10 are rejected. (a) Test the hypothesis H0: p  0.03 against H1: p  0.03 at   0.05. Find the P-value. (b) Explain how the question in part (a) could be answered by constructing a 95% one-sided confidence interval for p. 9-88. A random sample of 300 circuits generated 13 defectives. (a) Use the data to test H0: p  0.05 versus H1: p  0.05. Use   0.05. Find the P-value for the test. (b) Explain how the question in part (a) could be answered with a confidence interval. 9-89. An article in the British Medical Journal [“Comparison of Treatment of Renal Calculi by Operative Surgery, Percutaneous Nephrolithotomy, and Extra-Corporeal Shock Wave Lithotrips,” (1986, Vol. 292, pp. 879–882)] found that percutaneous nephrolithotomy (PN) had a success rate in removing kidney stones of 289 out of 350 patients. The traditional method was 78% effective. (a) Is there evidence that the success rate for PN is greater than the historical success rate? Find the P-value. (b) Explain how the question in part (a) could be answered with a confidence interval. 9-90. A manufacturer of interocular lenses is qualifying a new grinding machine and will qualify the machine if there is evidence that the percentage of polished lenses that contain surface defects does not exceed 2%. A random sample of 250 lenses contains six defective lenses. (a) Formulate and test an appropriate set of hypotheses to determine if the machine can be qualified. Use   0.05. Find the P-value. (b) Explain how the question in part (a) could be answered with a confidence interval. 9-91. A researcher claims that at least 10% of all football helmets have manufacturing flaws that could potentially cause injury to the wearer. A sample of 200 helmets revealed that 16 helmets contained such defects.

329

(a) Does this finding support the researcher’s claim? Use   0.01. Find the P-value. (b) Explain how the question in part (a) could be answered with a confidence interval. 9-92. An article in Fortune (September 21, 1992) claimed that nearly one-half of all engineers continue academic studies beyond the B.S. degree, ultimately receiving either an M.S. or a Ph.D. degree. Data from an article in Engineering Horizons (Spring 1990) indicated that 117 of 484 new engineering graduates were planning graduate study. (a) Are the data from Engineering Horizons consistent with the claim reported by Fortune? Use   0.05 in reaching your conclusions. Find the P-value for this test. (b) Discuss how you could have answered the question in part (a) by constructing a two-sided confidence interval on p. 9-93. The advertised claim for batteries for cell phones is set at 48 operating hours, with proper charging procedures. A study of 5000 batteries is carried out and 15 stop operating prior to 48 hours. Do these experimental results support the claim that less than 0.2 percent of the company’s batteries will fail during the advertised time period, with proper charging procedures? Use a hypothesis-testing procedure with   0.01. 9-94. A random sample of 500 registered voters in Phoenix is asked if they favor the use of oxygenated fuels year-round to reduce air pollution. If more than 315 voters respond positively, we will conclude that at least 60% of the voters favor the use of these fuels. (a) Find the probability of type I error if exactly 60% of the voters favor the use of these fuels. (b) What is the type II error probability if 75% of the voters favor this action? 9-95. In a random sample of 85 automobile engine crankshaft bearings, 10 have a surface finish roughness that exceeds the specifications. Does this data present strong evidence that the proportion of crankshaft bearings exhibiting excess surface roughness exceeds 0.10? (a) State and test the appropriate hypotheses using   0.05. (b) If it is really the situation that p  0.15, how likely is it that the test procedure in part (a) will not reject the null hypothesis? (c) If p  0.15, how large would the sample size have to be for us to have a probability of correctly rejecting the null hypothesis of 0.9?

9-6 SUMMARY TABLE OF INFERENCE PROCEDURES FOR A SINGLE SAMPLE The table in the end papers of this book (inside back cover) presents a summary of all the single-sample inference procedures from Chapters 8 and 9. The table contains the null hypothesis statement, the test statistic, the various alternative hypotheses and the criteria

JWCL232_c09_283-350.qxd

330

1/14/10

9:12 PM

Page 330

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

for rejecting H0, and the formulas for constructing the 100(1 ⫺ ␣)% two-sided confidence interval. It would also be helpful to refer to the roadmap table in Chapter 8 that provides guidance to match the problem type to the information inside the back cover.

9-7 TESTING FOR GOODNESS OF FIT The hypothesis-testing procedures that we have discussed in previous sections are designed for problems in which the population or probability distribution is known and the hypotheses involve the parameters of the distribution. Another kind of hypothesis is often encountered: We do not know the underlying distribution of the population, and we wish to test the hypothesis that a particular distribution will be satisfactory as a population model. For example, we might wish to test the hypothesis that the population is normal. We have previously discussed a very useful graphical technique for this problem called probability plotting and illustrated how it was applied in the case of a normal distribution. In this section, we describe a formal goodness-of-fit test procedure based on the chi-square distribution. The test procedure requires a random sample of size n from the population whose probability distribution is unknown. These n observations are arranged in a frequency histogram, having k bins or class intervals. Let Oi be the observed frequency in the ith class interval. From the hypothesized probability distribution, we compute the expected frequency in the ith class interval, denoted Ei. The test statistic is Goodness of Fit Test Statistic

k

X 02 ⫽ a i⫽1

1Oi ⫺ Ei 2 2 Ei

(9-47)

It can be shown that, if the population follows the hypothesized distribution, X 20 has, approximately, a chi-square distribution with k ⫺ p ⫺ 1 degrees of freedom, where p represents the number of parameters of the hypothesized distribution estimated by sample statistics. This approximation improves as n increases. We should reject the null hypothesis that the population is the hypothesized distribution if the test statistic is too large. Therefore, the P-value would be the probability under the chi-square distribution with k ⫺ p ⫺ 1 degrees of freedom above the computed value of the test statistic ␹20 or P ⫽ P1␹2k⫺p⫺1 ⬎ ␹20 2 . For a fixed-level test, we would reject the hypothesis that the distribution of the population is the hypothesized distribution if the calculated value of the test statistic ␹20 ⬎ ␹2␣,k⫺p⫺1. One point to be noted in the application of this test procedure concerns the magnitude of the expected frequencies. If these expected frequencies are too small, the test statistic X02 will not reflect the departure of observed from expected, but only the small magnitude of the expected frequencies. There is no general agreement regarding the minimum value of expected frequencies, but values of 3, 4, and 5 are widely used as minimal. Some writers suggest that an expected frequency could be as small as 1 or 2, so long as most of them exceed 5. Should an expected frequency be too small, it can be combined with the expected frequency in an adjacent class interval. The corresponding observed frequencies would then also be combined, and k would be reduced by 1. Class intervals are not required to be of equal width. We now give two examples of the test procedure.

JWCL232_c09_283-350.qxd

1/14/10

3:07 PM

Page 331

9-7 TESTING FOR GOODNESS OF FIT

EXAMPLE 9-12 Printed Circuit Board Defects Poisson Distribution The number of defects in printed circuit boards is hypothesized to follow a Poisson distribution. A random sample of n  60 printed boards has been collected, and the following number of defects observed. Number of Defects

Observed Frequency

0 1 2 3

32 15 9 4

The mean of the assumed Poisson distribution in this example is unknown and must be estimated from the sample data. The estimate of the mean number of defects per board is the sample average, that is, (320 151 92 43)兾60  0.75. From the Poisson distribution with parameter 0.75, we may compute pi, the theoretical, hypothesized probability associated with the ith class interval. Since each class interval corresponds to a particular number of defects, we may find the pi as follows:

331

Since the expected frequency in the last cell is less than 3, we combine the last two cells: Number of Defects

Observed Frequency

Expected Frequency

0 1 2 (or more)

32 15 13

28.32 21.24 10.44

The chi-square test statistic in Equation 9-47 will have k p 1  3 1 1  1 degree of freedom, because the mean of the Poisson distribution was estimated from the data. The seven-step hypothesis-testing procedure may now be applied, using   0.05, as follows: 1. Parameter of interest: The variable of interest is the form of the distribution of defects in printed circuit boards. 2. Null hypothesis: H0: The form of the distribution of defects is Poisson. 3. Alternative hypothesis: H1: The form of the distribution of defects is not Poisson. 4. Test statistic: The test statistic is

p1  P1X  02 

e 0.75 10.752 0

p2  P1X  12 

e 0.75 10.752 1

 0.354

p3  P1X  22 

e 0.75 10.752 2

5. Reject H0 if: Reject H0 if the P-value is less than 0.05.

 0.133

6. Computations:

0! 1! 2!

k

 0.472

20  a

1oi Ei 2 2

i1

p4  P1X  32  1 1 p1 p2 p3 2  0.041 The expected frequencies are computed by multiplying the sample size n  60 times the probabilities pi. That is, Ei  npi. The expected frequencies follow: Number of Defects

Probability

Expected Frequency

0 1 2 3 (or more)

0.472 0.354 0.133 0.041

28.32 21.24 7.98 2.46

EXAMPLE 9-13 Power Supply Distribution Continuous Distribution A manufacturing engineer is testing a power supply used in a notebook computer and, using   0.05, wishes to determine whether output voltage is adequately described by a normal

20 

132 28.322 2

Ei



28.32 113 10.442 2 10.44

115 21.242 2 21.24  2.94

7. Conclusions: We find from Appendix Table III that 20.10,1  2.71 and 20.05,1  3.84. Because 20  2.94 lies between these values, we conclude that the P-value is between 0.05 and 0.10. Therefore, since the P-value exceeds 0.05 we are unable to reject the null hypothesis that the distribution of defects in printed circuit boards is Poisson. The exact P-value computed from Minitab is 0.0864.

distribution. Sample estimates of the mean and standard deviation of x  5.04 V and s  0.08 V are obtained from a random sample of n  100 units.

JWCL232_c09_283-350.qxd

332

1/14/10

9:15 PM

Page 332

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

A common practice in constructing the class intervals for the frequency distribution used in the chi-square goodness-offit test is to choose the cell boundaries so that the expected frequencies Ei  npi are equal for all cells. To use this method, we want to choose the cell boundaries a0, a1, p , ak for the k cells so that all the probabilities ai

are equal. Suppose we decide to use k  8 cells. For the standard normal distribution, the intervals that divide the scale into eight equally likely segments are [0, 0.32), [0.32, 0.675), [0.675, 1.15), [1.15, ), and their four “mirror image” intervals on the other side of zero. For each interval pi  1兾8  0.125, so the expected cell frequencies are Ei  npi 100(0.125)  12.5. The complete table of observed and expected frequencies is as follows: Class Interval

Observed Frequency oi

Expected Frequency Ei

x  4.948 4.948 x  4.986 4.986 x  5.014 5.014 x  5.040 5.040 x  5.066 5.066 x  5.094 5.094 x  5.132 5.132 x

12 14 12 13 12 11 12 14

12.5 12.5 12.5 12.5 12.5 12.5 12.5 12.5

100

2. Null hypothesis: H0: The form of the distribution is normal.

4. Test statistic: The test statistic is

ai1

Totals

1. Parameter of Interest: The variable of interest is the form of the distribution of power supply voltage.

3. Alternative hypothesis: H1: The form of the distribution is nonnormal.

冮 f 1x2 dx

pi  P1ai1 X ai 2 

forth. We may apply the seven-step hypothesis-testing procedure to this problem.

k

20  a

1oi  Ei 2 2 Ei

i1

5. Reject H0 if: Since two parameters in the normal distribution have been estimated, the chi-square statistic above will have k  p  1  8  2  1  5 degrees of freedom. We will use a fixed significance level test with   0.05. Therefore, we will reject H0 if  20   20.05,5  11.07. 6. Computations: 8

20  a

1oi  Ei 2 2

Ei 112  12.52 2 i1



 0.64

100

The boundary of the first class interval is x  1.15s  4.948. The second class interval is 3x  1.15s, x  0.675s2 and so

12.5



114  12.52 2 12.5

p

114  12.52 2 12.5

7. Conclusions: Since  20  0.64   20.05,5  11.07, we are unable to reject H0, and there is no strong evidence to indicate that output voltage is not normally distributed. The P-value for the chi-square statistic  20  0.64 is P  0.9861.

EXERCISES FOR SECTION 9-7 9-96. Consider the following frequency table of observations on the random variable X.

spected and the following data were observed for the values of X:

Values Observed Frequency

Values Observed Frequency

0 24

1 30

2 31

3 11

4 4

(a) Based on these 100 observations, is a Poisson distribution with a mean of 1.2 an appropriate model? Perform a goodness-of-fit procedure with   0.05. (b) Calculate the P-value for this test. 9-97. Let X denote the number of flaws observed on a large coil of galvanized steel. Seventy-five coils are in-

1

2

3

4

5

6

7

8

1

11

8

13

11

12

10

9

(a) Does the assumption of the Poisson distribution seem appropriate as a probability model for these data? Use   0.01. (b) Calculate the P-value for this test.

JWCL232_c09_283-350.qxd

1/14/10

3:07 PM

Page 333

9-8 CONTINGENCY TABLE TESTS

9-98. The number of calls arriving at a switchboard from noon to 1:00 P.M. during the business days Monday through Friday is monitored for six weeks (i.e., 30 days). Let X be defined as the number of calls during that one-hour period. The relative frequency of calls was recorded and reported as Value Relative Frequency Value Relative Frequency

5

6

0.067 11 0.133

8

0.067 12 0.133

0.100 13 0.067

9

10

0.133 14

0.200 15

0.033

0.067

(a) Does the assumption of a Poisson distribution seem appropriate as a probability model for this data? Use   0.05. (b) Calculate the P-value for this test. 9-99. Consider the following frequency table of observations on the random variable X: Values Frequency

0 4

1 21

2 10

3 13

4 2

(a) Based on these 50 observations, is a binomial distribution with n  6 and p  0.25 an appropriate model? Perform a goodness-of-fit procedure with   0.05. (b) Calculate the P-value for this test. 9-100. Define X as the number of underfilled bottles from a filling operation in a carton of 24 bottles. Seventy-five cartons are inspected and the following observations on X are recorded: Values Frequency

0 39

1 23

2 12

3 1

(a) Based on these 75 observations, is a binomial distribution an appropriate model? Perform a goodness-of-fit procedure with   0.05. (b) Calculate the P-value for this test.

333

9-101. The number of cars passing eastbound through the intersection of Mill and University Avenues has been tabulated by a group of civil engineering students. They have obtained the data in the adjacent table: (a) Does the assumption of a Poisson distribution seem appropriate as a probability model for this process? Use   0.05. (b) Calculate the P-value for this test. Vehicles per Minute

Observed Frequency

Vehicles per Minute

Observed Frequency

40 41 42 43 44 45 46 47 48 49 50 51 52

14 24 57 111 194 256 296 378 250 185 171 150 110

53 54 55 56 57 58 59 60 61 62 63 64 65

102 96 90 81 73 64 61 59 50 42 29 18 15

9-102. Reconsider Exercise 6-71. The data were the number of earthquakes per year of magnitude 7.0 and greater since 1900. (a) Use computer software to summarize these data into a frequency distribution. Test the hypothesis that the number of earthquakes of magnitude 7.0 or greater each year follows a Poisson distribution at   0.05. (b) Calculate the P-value for the test.

9-8 CONTINGENCY TABLE TESTS Many times, the n elements of a sample from a population may be classified according to two different criteria. It is then of interest to know whether the two methods of classification are statistically independent; for example, we may consider the population of graduating engineers, and we may wish to determine whether starting salary is independent of academic disciplines. Assume that the first method of classification has r levels and that the second method has c levels. We will let Oij be the observed frequency for level i of the first classification method and level j on the second classification method. The data would, in general, appear as shown in Table 9-2. Such a table is usually called an r  c contingency table.

JWCL232_c09_283-350.qxd

334

1/14/10

9:17 PM

Page 334

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

Table 9-2 An r  c Contingency Table Columns 1

2

p

c

1

O11

O12

p

O1c

2

O21

O22

p

O2c

o

o

o

o

o

r

Or1

Or2

p

Orc

Rows

We are interested in testing the hypothesis that the row-and-column methods of classification are independent. If we reject this hypothesis, we conclude there is some interaction between the two criteria of classification. The exact test procedures are difficult to obtain, but an approximate test statistic is valid for large n. Let pij be the probability that a randomly selected element falls in the ijth cell, given that the two classifications are independent. Then pij  uivj, where ui is the probability that a randomly selected element falls in row class i and vj is the probability that a randomly selected element falls in column class j. Now, assuming independence, the estimators of ui and vj are c uˆ i  1 a Oij n j1

1 r vˆ j  n a Oij i1

(9-48)

Therefore, the expected frequency of each cell is r 1 c Eij  nuˆ ivˆj  n a Oij a Oij j1 i1

(9-49)

Then, for large n, the statistic r

c

20  a a

i1 j1

1Oij  Eij 2 2 Eij

(9-50)

has an approximate chi-square distribution with (r  1)(c  1) degrees of freedom if the null hypothesis is true. We should reject the null hypothesis if the value of the test statistic 20 is too large. The P-value would be calculated as the probability beyond 20 on the 21r121c12 distribution, or P  p121r121c12  20 2. For a fixed-level test, we would reject the hypothesis of independence if the observed value of the test statistic 20 exceeded 2,(r1)(c1).

EXAMPLE 9-14

Health Insurance Plan Preference

A company has to choose among three health insurance plans. Management wishes to know whether the preference for plans is independent of job classification and wants to use   0.05.

The opinions of a random sample of 500 employees are shown in Table 9-3.

JWCL232_c09_283-350.qxd

1/14/10

9:22 PM

Page 335

335

9-8 CONTINGENCY TABLE TESTS

Table 9-4 Expected Frequencies for Example 9-14

Table 9-3 Observed Data for Example 9-14

Health Insurance Plan

Health Insurance Plan Job Classification

1

2

3

Totals

Salaried workers Hourly workers

160 40

140 60

40 60

340 160

Totals

200

200

100

500

To find the expected frequencies, we must first compute uˆ1 (340兾500)  0.68, uˆ2  (160兾500)  0.32, vˆ1  (200兾500)  0.40, vˆ2  (200兾500)  0.40, and vˆ3  (100兾500)  0.20. The expected frequencies may now be computed from Equation 9-49. For example, the expected number of salaried workers favoring health insurance plan 1 is E11  nuˆ 1vˆ 1  50010.68210.402  136 The expected frequencies are shown in Table 9-4. The seven-step hypothesis-testing procedure may now be applied to this problem. 1. Parameter of Interest: The variable of interest is employee preference among health insurance plans. 2. Null hypothesis: H0: Preference is independent of salaried versus hourly job classification. 3. Alternative hypothesis: H1: Preference is not independent of salaried versus hourly job classification. 4. Test statistic: The test statistic is r

c

20  a a

i1 j1

1oij  Eij 2 2 Eij

5. Reject H0 if: We will use a fixed-significance level test with   0.05. Therefore, since r  2 and c  3, the

Job Classification

1

2

3

Totals

Salaried workers Hourly workers

136 64

136 64

68 32

340 160

Totals

200

200

100

500

degrees of freedom for chi-square are (r  1)(c  1) 2  (1)(2)  2, and we would reject H0 if  02   0.05,2  5.99. 6. Computations: 2

3

20  a a

1oij  Eij 2 2 Eij

i1 j1



1160  1362 2 



136 140  642 2 64

1140  1362 2



136 160  642 2 64





140  682 2

68 160  322 2 32

 49.63 7. Conclusions: Since 20  49.63  20.05,2  5.99, we reject the hypothesis of independence and conclude that the preference for health insurance plans is not independent of job classification. The P-value for 20  49.63 is P  1.671  1011. (This value was computed from computer software.) Further analysis would be necessary to explore the nature of the association between these factors. It might be helpful to examine the table of observed minus expected frequencies.

Using the two-way contingency table to test independence between two variables of classification in a sample from a single population of interest is only one application of contingency table methods. Another common situation occurs when there are r populations of interest and each population is divided into the same c categories. A sample is then taken from the ith population, and the counts are entered in the appropriate columns of the ith row. In this situation we want to investigate whether or not the proportions in the c categories are the same for all populations. The null hypothesis in this problem states that the populations are homogeneous with respect to the categories. For example, when there are only two categories, such as success and failure, defective and nondefective, and so on, the test for homogeneity is really a test of the equality of r binomial parameters. Calculation of expected frequencies, determination of degrees of freedom, and computation of the chi-square statistic for the test for homogeneity are identical to the test for independence.

JWCL232_c09_283-350.qxd

336

1/14/10

3:08 PM

Page 336

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

EXERCISES FOR SECTION 9-8 9-103. A company operates four machines in three shifts each day. From production records, the following data on the number of breakdowns are collected:

Machines Shift

A

B

C

D

1 2 3

41 31 15

20 11 17

12 9 16

16 14 10

Patient Category Surgical

Medical

Yes No

46 36

52 43

Range (yards)

Left

Normal

Right

0–1,999 2,000–5,999 6,000–11,999

6 9 8

14 11 17

8 4 6

9-107. A study is being made of the failures of an electronic component. There are four types of failures possible and two mounting positions for the device. The following data have been taken:

Test the hypothesis (using   0.05) that breakdowns are independent of the shift. Find the P-value for this test. 9-104. Patients in a hospital are classified as surgical or medical. A record is kept of the number of times patients require nursing service during the night and whether or not these patients are on Medicare. The data are presented here:

Medicare

Lateral Deflection

Failure Type Mounting Position

A

B

C

D

1 2

22 4

46 17

18 6

9 12

Would you conclude that the type of failure is independent of the mounting position? Use   0.01. Find the P-value for this test. 9-108. A random sample of students is asked their opinions on a proposed core curriculum change. The results are as follows.

Test the hypothesis (using   0.01) that calls by surgicalmedical patients are independent of whether the patients are receiving Medicare. Find the P-value for this test. 9-105. Grades in a statistics course and an operations research course taken simultaneously were as follows for a group of students.

Opinion Class Freshman Sophomore Junior Senior

Favoring

Opposing

120 70 60 40

80 130 70 60

Operation Research Grade Statistics Grade

A

B

C

Other

A B C Other

25 17 18 10

6 16 4 8

17 15 18 11

13 6 10 20

Are the grades in statistics and operations research related? Use   0.01 in reaching your conclusion. What is the P-value for this test? 9-106. An experiment with artillery shells yields the following data on the characteristics of lateral deflections and ranges. Would you conclude that deflection and range are independent? Use   0.05. What is the P-value for this test?

Test the hypothesis that opinion on the change is independent of class standing. Use   0.05. What is the P-value for this test? 9-109. An article in the British Medical Journal [“Comparison of Treatment of Renal Calculi by Operative Surgery, Percutaneous Nephrolithotomy, and Extracorporeal Shock Wave Lithotripsy” (1986, Vol. 292, pp. 879–882)] found that percutaneous nephrolithotomy (PN) had a success rate in removing kidney stones of 289 out of 350 (83%) patients. However, when the stone diameter was considered, the results looked different. For stones of 2 cm, 87% (234兾270) of cases were successful. For stones of 2 cm, a success rate of 69% (55兾80) was observed for PN. (a) Are the successes and size of stones independent? Use   0.05. (b) Find the P-value for this test.

JWCL232_c09_283-350.qxd

1/14/10

9:34 PM

Page 337

9-9

NONPARAMETRIC PROCEDURES

337

9-9 NONPARAMETRIC PROCEDURES Most of the hypothesis-testing and confidence interval procedures discussed previously are based on the assumption that we are working with random samples from normal populations. Traditionally, we have called these procedures parametric methods because they are based on a particular parametric family of distributions—in this case, the normal. Alternately, sometimes we say that these procedures are not distribution-free because they depend on the assumption of normality. Fortunately, most of these procedures are relatively insensitive to moderate departures from normality. In general, the t- and F-tests and the t-confidence intervals will have actual levels of significance or confidence levels that differ from the nominal or advertised levels chosen by the experimenter, although the difference between the actual and advertised levels is usually fairly small when the underlying population is not too different from the normal. In this section we describe procedures called nonparametric and distribution-free methods, and we usually make no assumptions about the distribution of the underlying population other than that it is continuous. These procedures have actual level of significance ␣ or confidence level 100(1 ⫺ ␣) % for many different types of distributions. These procedures have some appeal. One of their advantages is that the data need not be quantitative but can be categorical (such as yes or no, defective or nondefective) or rank data. Another advantage is that nonparametric procedures are usually very quick and easy to perform. The procedures described in this chapter are alternatives to the parametric t- and F-procedures described earlier. Consequently, it is important to compare the performance of both parametric and nonparametric methods under the assumptions of both normal and nonnormal populations. In general, nonparametric procedures do not utilize all the information provided by the sample. As a result, a nonparametric procedure will be less efficient than the corresponding parametric procedure when the underlying population is normal. This loss of efficiency is reflected by a requirement of a larger sample size for the nonparametric procedure than would be required by the parametric procedure in order to achieve the same power. On the other hand, this loss of efficiency is usually not large, and often the difference in sample size is very small. When the underlying distributions are not close to normal, nonparametric methods may have much to offer. They often provide improvement over the normal-theory parametric methods. Generally, if both parametric and nonparametric methods are applicable to a particular problem, we should use the more efficient parametric procedure. Another approach that can be used is to transform the original data, say, by taking logarithms, square roots, or a reciprocal, and then analyze the transformed data using a parametric technique. A normal probability plot often works well to see if the transformation has been successful. When this approach is successful, it is usually preferable to using a nonparametric technique. However, sometimes transformations are not satisfactory. That is, no transformation makes the sample observations look very close to a sample from a normal distribution. One situation where this happens is when the data are in the form of ranks. These situations frequently occur in practice. For instance, a panel of judges may be used to evaluate 10 different formulations of a soft-drink beverage for overall quality, with the “best” formulation assigned rank 1, the “next-best” formulation assigned rank 2, and so forth. It is unlikely that rank data satisfy the normality assumption. Transformations may not prove satisfactory either. Many nonparametric methods involve the analysis of ranks and consequently are directly suited to this type of problem.

9-9.1 The Sign Test ~ of ␣ continuous distribution. The The sign test is used to test hypotheses about the median ␮ median of a distribution is a value of the random variable X such that the probability is 0.5 that

JWCL232_c09_283-350.qxd

338

1/14/10

3:08 PM

Page 338

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

an observed value of X is less than or equal to the median, and the probability is 0.5 that an ob~ 2  P1X   ~ 2  0.5 served value of X is greater than or equal to the median. That is, P1X   Since the normal distribution is symmetric, the mean of a normal distribution equals the median. Therefore, the sign test can be used to test hypotheses about the mean of a normal distribution. This is the same problem for which we previously used the t-test. We will briefly discuss the relative merits of the two procedures in Section 9-9.3 Note that, although the t-test was designed for samples from a normal distribution, the sign test is appropriate for samples from any continuous distribution. Thus, the sign test is a nonparametric procedure. Suppose that the hypotheses are ~  ~ H0:  0 ~ ~ H1:   0

(9-51)

The test procedure is easy to describe. Suppose that X1, X2, . . . , Xn is a random sample from the population of interest. Form the differences ~ , Xi  0

i  1, 2, . . . , n

(9-52)

~ is true, any difference X  ~ is equally likely to Now if the null hypothesis H0: ~   0 i 0 be positive or negative. An appropriate test statistic is the number of these differences that are positive, say, R+. Therefore, to test the null hypothesis we are really testing that the number of plus signs is a value of a binomial random variable that has the parameter p = 1/2. A P-value for the observed number of plus signs r+ can be calculated directly from the binomial distribution. For instance, in testing the hypotheses in Equation 9-51, we will reject H0 in favor of H1 only if the proportion of plus signs is sufficiently less than 1/2 (or equivalently, whenever the observed number of plus signs r+ is too small). Thus, if the computed P-value 1 P  P aR  r when p  b 2 is less than or equal to some preselected significance level , we will reject H0 and conclude H1 is true. To test the other one-sided hypotheses ~ H : ~   0

0

~  ~ H1:  0

(9-53) +

we will reject H0 in favor of H1 only if the observed number of plus signs, say, r , is large or, equivalently, whenever the observed fraction of plus signs is significantly greater than 1/2. Thus, if the computed P-value 1 P  P aR  r when p  b 2 is less than , we will reject H0 and conclude that H1 is true. The two-sided alternative may also be tested. If the hypotheses are ~ H0: ~   0 ~ ~ H : 1

0

(9-54)

~ if the proportion of plus signs is significantly different from we should reject H0: ~   0 (either less than or greater than) 1/2. This is equivalent to the observed number of plus signs r+

JWCL232_c09_283-350.qxd

1/14/10

9:36 PM

Page 339

9-9

NONPARAMETRIC PROCEDURES

339

being either sufficiently large or sufficiently small. Thus, if r ⫹ ⬍ n/2, the P-value is 1 P ⫽ 2PaR ⫹ ⱕ r ⫹ when p ⫽ b 2 and if r+ ⬎ n/2, the P-value is 1 P ⫽ 2PaR ⫹ ⱖ r ⫹ when p ⫽ b 2 If the P-value is less than some preselected level ␣, we will reject H0 and conclude that H1 is true.

EXAMPLE 9-15

Propellant Shear Strength Sign Test

Montgomery, Peck, and Vining (2006) report on a study in which a rocket motor is formed by binding an igniter propellant and a sustainer propellant together inside a metal housing. The shear strength of the bond between the two propellant types is an important characteristic. The results of testing 20 randomly selected motors are shown in Table 9-5. We would like to test the hypothesis that the median shear strength is 2000 psi, using ␣ ⫽ 0.05. This problem can be solved using the eight-step hypothesistesting procedure:

1. Parameter of Interest: The parameter of interest is the median of the distribution of propellant shear strength. 苲 ⫽ 2000 psi 2. Null hypothesis: H : ␮ 0

苲 ⫽ 2000 psi 3. Alternative hypothesis: H1: ␮

4. Test statistic: The test statistic is the observed number of plus differences in Table 9-5, or r⫹ ⫽ 14. 5. Reject H0 if: We will reject H0 if the P-value corresponding to r⫹ ⫽ 14 is less than or equal to ␣ ⫽ 0.05.

Table 9-5 Propellant Shear Strength Data Observation i

Shear Strength xi

Differences xi ⫺ 2000

Sign

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

2158.70 1678.15 2316.00 2061.30 2207.50 1708.30 1784.70 2575.10 2357.90 2256.70 2165.20 2399.55 1779.80 2336.75 1765.30 2053.50 2414.40 2200.50 2654.20 1753.70

⫹158.70 ⫺321.85 ⫹316.00 ⫹61.30 ⫹207.50 ⫺291.70 ⫺215.30 ⫹575.10 ⫹357.90 ⫹256.70 ⫹165.20 ⫹399.55 ⫺220.20 ⫹336.75 ⫺234.70 ⫹53.50 ⫹414.40 ⫹200.50 ⫹654.20 ⫺246.30

⫹ ⫺ ⫹ ⫹ ⫹ ⫺ ⫺ ⫹ ⫹ ⫹ ⫹ ⫹ ⫺ ⫹ ⫺ ⫹ ⫹ ⫹ ⫹ ⫺

JWCL232_c09_283-350.qxd

340

1/14/10

3:08 PM

Page 340

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

6. Computations: Since r  14 is greater than n兾2  20兾2  10, we calculate the P-value from 1 P  2P aR  14 when p  b 2 20 20  2 a a b 10.52 r 10.52 20 r r14 r

7. Conclusions: Since P  0.1153 is not less than   0.05, we cannot reject the null hypothesis that the median shear strength is 2000 psi. Another way to say this is that the observed number of plus signs r  14 was not large or small enough to indicate that median shear strength is different from 2000 psi at the   0.05 level of significance.

 0.1153

It is also possible to construct a table of critical values for the sign test. This table is shown as Appendix Table VIII. The use of this table for the two-sided alternative hypothesis in Equation  0) that are positive 9-54 is simple. As before, let R denote the number of the differences (Xi 

and let R denote the number of these differences that are negative. Let R  min (R , R ). Appendix Table VIII presents critical values r* for the sign test that ensure that P (type I error)  P (reject H0 when H0 is true)   for   0.01,   0.05 and   0.10. If the   0 should be rejected. observed value of the test statistic r  r*, the null hypothesis H0:  To illustrate how this table is used, refer to the data in Table 9-5 that were used in Example 9-15. Now r  14 and r  6; therefore, r  min (14, 6)  6. From Appendix Table VIII with n  20 and   0.05, we find that r*0.05  5. Since r  6 is not less than or equal to the critical value r*0.05  5, we cannot reject the null hypothesis that the median shear strength is 2000 psi. We can also use Appendix Table VIII for the sign test when a one-sided alternative   0, reject H0:    0 if r  r*; hypothesis is appropriate. If the alternative is H1:      if the alternative is H1:   0, reject H0:   0 if r  r*. The level of significance of a one-sided test is one-half the value for a two-sided test. Appendix Table VIII shows the onesided significance levels in the column headings immediately below the two-sided levels. Finally, note that when a test statistic has a discrete distribution such as R does in the sign test, it may be impossible to choose a critical value r* that has a level of significance exactly equal to . The approach used in Appendix Table VIII is to choose r* to yield an  that is as close to the advertised significance level  as possible. Ties in the Sign Test Since the underlying population is assumed to be continuous, there is a zero probability that  . However, this may sometimes we will find a “tie”—that is, a value of Xi exactly equal to  0 happen in practice because of the way the data are collected. When ties occur, they should be set aside and the sign test applied to the remaining data. The Normal Approximation When p  0.5, the binomial distribution is well approximated by a normal distribution when n is at least 10. Thus, since the mean of the binomial is np and the variance is np(1 p), the distribution of R is approximately normal with mean 0.5n and variance 0.25n whenever n is   can be tested using moderately large. Therefore, in these cases the null hypothesis H0:  0 the statistic Normal Approximation for Sign Test Statistic

Z0 

R 0.5n 0.51n

(9-55)

JWCL232_c09_283-350.qxd

1/14/10

3:08 PM

Page 341

9-9

NONPARAMETRIC PROCEDURES

341

A P-value approach could be used for decision making. The fixed significance level approach could also be used. The two-sided alternative would be rejected if the observed value of the test statistic 0 z0 0  z 2, and the critical regions of the one-sided alternative would be chosen to reflect the   , reject H if z  z , for example.) sense of the alternative. (If the alternative is H1:  0 0 0  Type II Error for the Sign Test The sign test will control the probability of type I error at an advertised level  for testing the   for any continuous distribution. As with any hypothesis-testing null hypothesis H0:  procedure, it is important to investigate the probability of a type II error, . The test should be able to effectively detect departures from the null hypothesis, and a good measure of this effectiveness is the value of for departures that are important. A small value of implies an effective test procedure.  , say,   , In determining , it is important to realize not only that a particular value of  0 must be used but also that the form of the underlying distribution will affect the calculations. To illustrate, suppose that the underlying distribution is normal with  1 and we are testing the   2 versus H :    2. (Since     in the normal distribution, this is equivhypothesis H0:  1 alent to testing that the mean equals 2.) Suppose that it is important to detect a departure from   2 to    3. The situation is illustrated graphically in Fig. 9-15(a). When the alternative    3), the probability that the random variable X is less than or equal to the hypothesis is true (H1:  value 2 is P1X  22  P1Z  12  1 12  0.1587 Suppose we have taken a random sample of size 12. At the   0.05 level, Appendix Table VIII   2 if r  r*  2. Therefore, is the probability that indicates that we would reject H0:  0.05   2 when in fact    3, or we do not reject H0:  2 12  1 a a b 10.15872 x 10.84132 12 x  0.2944 x0 x

If the distribution of X had been exponential rather than normal, the situation would be as shown in Fig. 9-15(b), and the probability that the random variable X is less than or equal   3 (note that when the median of an exponential distribution to the value x  2 when  is 3, the mean is 4.33) is 2

P1X  22 

冮 4.33 e 1

1

4.33 x

dx  0.3699

0

In this case, 2 12  1 a a b 10.36992 x 10.63012 12 x  0.8794 x x0

 but also on the area Thus, for the sign test depends not only on the alternative value of  to the right of the value specified in the null hypothesis under the population probability

JWCL232_c09_283-350.qxd

342

1/14/10

3:08 PM

Page 342

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

σ =1

σ =1 0.1587

–1

0

1

2

3

4

5

x

–1

∼ Under H0 : μ = 2

0

1

2

3

4

5

6

x

∼ Under H1 : μ = 3 (a)

0.3699

∼ μ=2

μ = 2.89

x

2

∼ Under H0 : μ = 2

μ = 4.33

x

∼ Under H1 : μ = 3 (b)

Figure 9-15 Calculation of for the sign test. (a) Normal distributions. (b) Exponential distributions.

distribution. This area is highly dependent on the shape of that particular probability distribution. In this example, is large so the ability of the test to detect this departure from the null hypothesis with the current sample size is poor.

9-9.2 The Wilcoxon Signed-Rank Test The sign makes use only of the plus and minus signs of the differences between the observa (or the plus and minus signs of the differences between the observations and the median  0 tions in the paired case). It does not take into account the size or magnitude of these differences. Frank Wilcoxon devised a test procedure that uses both direction (sign) and magnitude. This procedure, now called the Wilcoxon signed-rank test, is discussed and illustrated in this section. The Wilcoxon signed-rank test applies to the case of symmetric continuous distributions. Under these assumptions, the mean equals the median, and we can use this procedure to test the null hypothesis   0. The Test Procedure We are interested in testing H0:   0 against the usual alternatives. Assume that X1, X2, . . . , Xn is a random sample from a continuous and symmetric distribution with mean (and median) . Compute the differences Xi 0, i  1, 2, . . . , n. Rank the absolute differences 0 Xi 0 0 , i  1, 2, . . . , n in ascending order, and then give the ranks the signs of their corresponding differences. Let W be the sum of the positive ranks and W be the absolute

JWCL232_c09_283-350.qxd

1/19/10

7:42 PM

Page 343

9-9

343

NONPARAMETRIC PROCEDURES

value of the sum of the negative ranks, and let W ⫽ min(W ⫹, W). Appendix Table IX contains critical values of W, say, w*␣. If the alternative hypothesis is H1: ␮ ⫽ ␮0 , then if the observed value of the statistic w ⱕ w*␣, the null hypothesis H0: ␮ ⫽ ␮0 is rejected. Appendix Table IX provides significance levels of ␣ ⫽ 0.10, ␣ ⫽ 0.05, ␣ ⫽ 0.02, ␣ ⫽ 0.01 for the two-sided test. For one-sided tests, if the alternative is H1: ␮ ⬎ ␮0, reject H0: ␮ ⫽ ␮0 if w⫺ ⱕ w*␣; and if the alternative is H1: ␮ ⬍ ␮0, reject H0: ␮ ⫽ ␮0 if w⫹ ⱕ w*␣. The significance levels for onesided tests provided in Appendix Table IX are ␣ ⫽ 0.05, 0.025, 0.01, and 0.005.

EXAMPLE 9-16

Propellant Shear Strength Wilcoxon Signed-Rank Test

We will illustrate the Wilcoxon signed-rank test by applying it to the propellant shear strength data from Table 9-5. Assume that the underlying distribution is a continuous symmetric distribution. The seven-step procedure is applied as follows: 1. Parameter of Interest: The parameter of interest is the mean (or median) of the distribution of propellant shear strength. 2. Null hypothesis: H0: ␮ ⫽ 2000 psi 3. Alternative hypothesis: H1: ␮ ⬆ 2000 psi 4. Test statistic: The test statistic is w ⫽ min1w ⫹ , w⫺ 2

5. Reject H0 if: We will reject H0 if w ⱕ w*0.05 ⫽ 52 from Appendix Table IX.

15 20 10 6 3 2 14 9 12 17 8 19

⫺234.70 ⫺246.30 ⫹256.70 ⫺291.70 ⫹316.00 ⫺321.85 ⫹336.75 ⫹357.90 ⫹399.55 ⫹414.40 ⫹575.10 ⫹654.20

⫺9 ⫺10 ⫹11 ⫺12 ⫹13 ⫺14 ⫹15 ⫹16 ⫹17 ⫹18 ⫹19 ⫹20

6. Computations: The signed ranks from Table 9-5 are shown in the following display:

Observation

Difference xi ⫺ 2000

16 4 1 11 18 5 7 13

⫹53.50 ⫹61.30 ⫹158.70 ⫹165.20 ⫹200.50 ⫹207.50 ⫺215.30 ⫺220.20

Signed Rank ⫹1 ⫹2 ⫹3 ⫹4 ⫹5 ⫹6 ⫺7 ⫺8 continued

The sum of the positive ranks is w⫹ ⫽ (1 ⫹ 2 ⫹ 3 ⫹ 4 ⫹ 5 ⫹ 6 ⫹ 11 ⫹ 13 ⫹ 15 ⫹ 16 ⫹ 17 ⫹ 18 ⫹ 19 ⫹ 20) ⫽ 150, and the sum of the absolute values of the negative ranks is w⫺ ⫽ (7 ⫹ 8 ⫹ 9 ⫹ 10 ⫹ 12 ⫹ 14) ⫽ 60. Therefore, w ⫽ min1150, 602 ⫽ 60 7. Conclusions: Since w ⫽ 60 is not less than or equal to the critical value w0.05 ⫽ 52, we cannot reject the null hypothesis that the mean (or median, since the population is assumed to be symmetric) shear strength is 2000 psi.

Ties in the Wilcoxon Signed-Rank Test Because the underlying population is continuous, ties are theoretically impossible, although they will sometimes occur in practice. If several observations have the same absolute magnitude, they are assigned the average of the ranks that they would receive if they differed slightly from one another.

JWCL232_c09_283-350.qxd

344

1/14/10

3:08 PM

Page 344

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

Large Sample Approximation If the sample size is moderately large, say, n  20, it can be shown that W (or W ) has approximately a normal distribution with mean w 

n1n 12 4

and variance

2w 

n1n 1212n 12 24

Therefore, a test of H0:   0 can be based on the statistic:

Normal Approximation for Wilcoxon Signed-Rank Statistic

Z0 

W n1n 12/4 1n1n 1212n 12/24

(9-56)

An appropriate critical region for either the two-sided or one-sided alternative hypotheses can be chosen from a table of the standard normal distribution.

9-9.3 Comparison to the t-Test If the underlying population is normal, either the sign test or the t-test could be used to test a hypothesis about the population median. The t-test is known to have the smallest value of possible among all tests that have significance level  for the one-sided alternative and for tests with symmetric critical regions for the two-sided alternative, so it is superior to the sign test in the normal distribution case. When the population distribution is symmetric and nonnormal (but with finite mean), the t-test will have a smaller (or a higher power) than the sign test, unless the distribution has very heavy tails compared with the normal. Thus, the sign test is usually considered a test procedure for the median rather than as a serious competitor for the t-test. The Wilcoxon signed-rank test is preferable to the sign test and compares well with the t-test for symmetric distributions. It can be useful in situations where a transformation on the observations does not produce a distribution that is reasonably close to the normal.

EXERCISES FOR SECTION 9-9 9-110. Ten samples were taken from a plating bath used in an electronics manufacturing process, and the bath pH was determined. The sample pH values are 7.91, 7.85, 6.82, 8.01, 7.46, 6.95, 7.05, 7.35, 7.25, and 7.42. Manufacturing engineering believes that pH has a median value of 7.0.

(a) Do the sample data indicate that this statement is correct? Use the sign test with   0.05 to investigate this hypothesis. Find the P-value for this test. (b) Use the normal approximation for the sign test to test H0: ~  7.0 versus H1: ~  7.0. What is the P-value for this test?

JWCL232_c09_283-350.qxd

1/20/10

10:28 PM

Page 345

9-9

9-111. The titanium content in an aircraft-grade alloy is an important determinant of strength. A sample of 20 test coupons reveals the following titanium content (in percent): 8.32, 8.05, 8.93, 8.65, 8.25, 8.46, 8.52, 8.35, 8.36, 8.41, 8.42, 8.30, 8.71, 8.75, 8.60, 8.83, 8.50, 8.38, 8.29, 8.46 The median titanium content should be 8.5%. (a) Use the sign test with   0.05 to investigate this hypothesis. Find the P-value for this test. (b) Use the normal approximation for the sign test to test ~  8.5, with   0.05. What is H0: ~  8.5 versus H1:  the P-value for this test? 9-112. The impurity level (in ppm) is routinely measured in an intermediate chemical product. The following data were observed in a recent test:

345

NONPARAMETRIC PROCEDURES

The Rockwell C-scale hardness readings are 63, 65, 58, 60, 55, 57, 53, and 59. Do the results support the claim that the mean hardness exceeds 60 at a 0.05 level? 9-117. A primer paint can be used on aluminum panels. The drying time of the primer is an important consideration in the manufacturing process. Twenty panels are selected and the drying times are as follows: 1.6, 1.3, 1.5, 1.6, 1.7, 1.9, 1.8, 1.6, 1.4, 1.8, 1.9, 1.8, 1.7, 1.5, 1.6, 1.4, 1.3, 1.6, 1.5, and 1.8. Is there evidence that the mean drying time of the primer exceeds 1.5 hr?

Supplemental Exercises 9-118.

Consider the computer output below.

One-Sample Z: Test of mu  26 vs  26 The assumed standard deviation  1.5

2.4, 2.5, 1.7, 1.6, 1.9, 2.6, 1.3, 1.9, 2.0, 2.5, 2.6, 2.3, 2.0, 1.8, 1.3, 1.7, 2.0, 1.9, 2.3, 1.9, 2.4, 1.6

Variable X

N ?

Mean 26.541

StDev 2.032

SE Mean 0.401

Can you claim that the median impurity level is less than 2.5 ppm? (a) State and test the appropriate hypothesis using the sign test with   0.05. What is the P-value for this test? (b) Use the normal approximation for the sign test to test ~  2.5 versus H : ~  2.5. What is the P-value for H0:  1 this test? 9-113. Consider the margarine fat content data in Exercise 8-36. Use the sign test to test H0:~  17.0 versus H1: ~  17.0 with   0.05. (a) Find the P-value for the test statistic and use this quantity to make your decision. (b) Use the normal approximation to test the same hypothesis that you formulated in part (a). What is the P-value for this test? 9-114. Consider the compressive strength data in Exercise 8-37. (a) Use the sign test to investigate the claim that the median strength is at least 2250 psi. Use   0.05. (b) Use the normal approximation to test the same hypothesis that you formulated in part (a). What is the P-value for this test? 9-115. The diameter of a ball bearing was measured by an inspector using a new type of caliper. The results were as follows (in mm): 0.265, 0.263, 0.266, 0.267, 0.267, 0.265, 0.267, 0.267, 0.265, 0.268, 0.268, and 0.263. (a) Use the Wilcoxon signed-rank test to evaluate the claim that the mean ball diameter is 0.265 mm. Use   0.05. (b) Use the normal approximation for the test. With   0.05, what conclusions can you draw? 9-116. A new type of tip can be used in a Rockwell hardness tester. Eight coupons from test ingots of a nickel-based alloy are selected, and each coupon is tested using the new tip.

(a) Fill in the missing information. (b) Is this a one-sided or a two-sided test? (c) What are your conclusions if   0.05? (d) Find a 95% two-sided CI on the mean. 9-119. Consider the computer output below.

Z ?

P ?

One-Sample T: Test of mu  100 vs not  100 Variable X

N 16

Mean 98.33

StDev 4.61

SE Mean ?

95% CI (?, ?)

T P ? ?

(a) How many degrees of freedom are there on the t-statistic? (b) Fill in the missing information. You may use bounds on the P-value. (c) What are your conclusions if   0.05? (d) What are your conclusions if the hypothesis is H0:   100 versus H0:  > 100? 9-120. Consider the computer output below. One-Sample T: Test of mu  85 vs  85 Variable X

N 25

Mean 84.331

StDev ?

SE Mean 0.631

T ?

P ?

(a) How many degrees of freedom are there on the t-statistic? (b) Fill in the missing information. You may use bounds on the P-value. (c) What are your conclusions if   0.05? (d) Find a 95% upper-confidence bound on the mean. (e) What are your conclusions if the hypothesis is H0:   100 versus H0:   100? 9-121. An article in Transfusion Science [“Early Total White Blood Cell Recovery Is a Predictor of Low Number of

JWCL232_c09_283-350.qxd

346

1/14/10

9:46 PM

Page 346

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

Apheresis and Good CD34 Cell Yield” (Vol. 23, 2000, pp. 91–100)] studied the white blood cell recovery of patients with haematological malignancies after a new chemotherapy treatment. Data (in days) on white blood cell recovery (WBC) for 19 patients consistent with summary data reported in the paper follow: 18, 16, 13, 16, 15, 12, 9, 14, 12, 8, 16, 12, 10, 8, 14, 9, 5, 18, and 12. (a) Is there sufficient evidence to support a claim that the mean WBC recovery exceeds 12 days? (b) Find a 95% two-sided CI on the mean WBC recovery. 9-122. An article in Fire Technology [“An Experimental Examination of Dead Air Space for Smoke Alarms” (Vol. 45, 2009, pp. 97–115)] studied the performance of smoke detectors installed not less than 100 mm from any adjoining wall if mounted on a flat ceiling, and not closer than 100 mm and not farther than 300 mm from the adjoining ceiling surface if mounted on walls. The purpose of this rule is to avoid installation of smoke alarms in the “dead air space,” where it is assumed to be difficult for smoke to reach. A number of interesting experiments were described in the paper. Results on the time to signal (in seconds) for one such experiment with pine stick fuel in an open bedroom using photoelectric smoke alarms are as follows: 220, 225, 297, 315, 282, and 313. (a) Is there sufficient evidence to support a claim that the mean time to signal is less than 300 seconds? (b) Is there practical concern about the assumption of a normal distribution as a model for the time-to-signal data? (c) Find a 95% two-sided CI on the mean time to signal. 9-123. Suppose we wish to test the hypothesis H0:   85 versus the alternative H1:   85 where   16. Suppose that the true mean is   86 and that in the practical context of the problem this is not a departure from 0  85 that has practical significance. (a) For a test with   0.01, compute  for the sample sizes n  25, 100, 400, and 2500 assuming that   86. (b) Suppose the sample average is x  86 . Find the P-value for the test statistic for the different sample sizes specified in part (a). Would the data be statistically significant at   0.01? (c) Comment on the use of a large sample size in this problem. 9-124. A manufacturer of semiconductor devices takes a random sample of size n of chips and tests them, classifying each chip as defective or nondefective. Let Xi  0 if the chip is nondefective and Xi  1 if the chip is defective. The sample fraction defective is pˆ 

X1  X2  p  Xn n

What are the sampling distribution, the sample mean, and sample variance estimates of pˆ when (a) The sample size is n  50? (b) The sample size is n  80? (c) The sample size is n  100?

(d) Compare your answers to parts (a)–(c) and comment on the effect of sample size on the variance of the sampling distribution. 9-125. Consider the situation of Exercise 9-124. After collecting a sample, we are interested in testing H0: p  0.10 versus H1: p 0.10 with   0.05. For each of the following situations, compute the p-value for this test: (a) n  50, pˆ  0.095 (b) n  100, pˆ  0.095 (c) n  500, pˆ  0.095 (d) n  1000, pˆ  0.095 (e) Comment on the effect of sample size on the observed P-value of the test. 9-126. An inspector of flow metering devices used to administer fluid intravenously will perform a hypothesis test to determine whether the mean flow rate is different from the flow rate setting of 200 milliliters per hour. Based on prior information, the standard deviation of the flow rate is assumed to be known and equal to 12 milliliters per hour. For each of the following sample sizes, and a fixed   0.05, find the probability of a type II error if the true mean is 205 milliliters per hour. (a) n  20 (b) n  50 (c) n  100 (d) Does the probability of a type II error increase or decrease as the sample size increases? Explain your answer. 9-127. Suppose that in Exercise 9-126, the experimenter had believed that   14. For each of the following sample sizes, and a fixed   0.05, find the probability of a type II error if the true mean is 205 milliliters per hour. (a) n  20 (b) n  50 (c) n  100 (d) Comparing your answers to those in Exercise 9-126, does the probability of a type II error increase or decrease with the increase in standard deviation? Explain your answer. 9-128. The marketers of shampoo products know that customers like their product to have a lot of foam. A manufacturer of shampoo claims that the foam height of his product exceeds 200 millimeters. It is known from prior experience that the standard deviation of foam height is 8 millimeters. For each of the following sample sizes, and a fixed   0.05, find the power of the test if the true mean is 204 millimeters. (a) n  20 (b) n  50 (c) n  100 (d) Does the power of the test increase or decrease as the sample size increases? Explain your answer. 9-129. Suppose we are testing H0: p  0.5 versus H0: p 0.5. Suppose that p is the true value of the population proportion. (a) Using   0.05, find the power of the test for n  100, 150, and 300 assuming that p  0.6. Comment on the effect of sample size on the power of the test.

JWCL232_c09_283-350.qxd

1/14/10

3:08 PM

Page 347

9-9

(b) Using   0.01, find the power of the test for n  100, 150, and 300 assuming that p  0.6. Compare your answers to those from part (a) and comment on the effect of  on the power of the test for different sample sizes. (c) Using   0.05, find the power of the test for n  100, assuming p  0.08. Compare your answer to part (a) and comment on the effect of the true value of p on the power of the test for the same sample size and  level. (d) Using   0.01, what sample size is required if p  0.6 and we want  0.05? What sample is required if p  0.8 and we want  0.05? Compare the two sample sizes and comment on the effect of the true value of p on sample size required when is held approximately constant. 9-130. The cooling system in a nuclear submarine consists of an assembly of welded pipes through which a coolant is circulated. Specifications require that weld strength must meet or exceed 150 psi. (a) Suppose that the design engineers decide to test the hypothesis H0:   150 versus H1:   150. Explain why this choice of alternative hypothesis is better than H1:   150. (b) A random sample of 20 welds results in x  153.7 psi and s  11.3 psi. What conclusions can you draw about the hypothesis in part (a)? State any necessary assumptions about the underlying distribution of the data. 9-131. The mean pull-off force of an adhesive used in manufacturing a connector for an automotive engine application should be at least 75 pounds. This adhesive will be used unless there is strong evidence that the pull-off force does not meet this requirement. A test of an appropriate hypothesis is to be conducted with sample size n  10 and   0.05. Assume that the pull-off force is normally distributed, and is not known. (a) If the true standard deviation is  1, what is the risk that the adhesive will be judged acceptable when the true mean pull-off force is only 73 pounds? Only 72 pounds? (b) What sample size is required to give a 90% chance of detecting that the true mean is only 72 pounds when  1? (c) Rework parts (a) and (b) assuming that  2. How much impact does increasing the value of have on the answers you obtain? 9-132. A manufacturer of precision measuring instruments claims that the standard deviation in the use of the instruments is at most 0.00002 millimeter. An analyst, who is unaware of the claim, uses the instrument eight times and obtains a sample standard deviation of 0.00001 millimeter. (a) Confirm using a test procedure and an  level of 0.01 that there is insufficient evidence to support the claim that the standard deviation of the instruments is at most 0.00002. State any necessary assumptions about the underlying distribution of the data. (b) Explain why the sample standard deviation, s  0.00001, is less than 0.00002, yet the statistical test procedure results do not support the claim.

NONPARAMETRIC PROCEDURES

347

9-133. A biotechnology company produces a therapeutic drug whose concentration has a standard deviation of 4 grams per liter. A new method of producing this drug has been proposed, although some additional cost is involved. Management will authorize a change in production technique only if the standard deviation of the concentration in the new process is less than 4 grams per liter. The researchers chose n  10 and obtained the following data in grams per liter. Perform the necessary analysis to determine whether a change in production technique should be implemented. 16.628 16.622 16.627 16.623 16.618

16.630 16.631 16.624 16.622 16.626

9-134. Consider the 40 observations collected on the number of nonconforming coil springs in production batches of size 50 given in Exercise 6-93. (a) Based on the description of the random variable and these 40 observations, is a binomial distribution an appropriate model? Perform a goodness of fit procedure with   0.05. (b) Calculate the P-value for this test. 9-135. Consider the 20 observations collected on the number of errors in a string of 1000 bits of a communication channel given in Exercise 6-94. (a) Based on the description of the random variable and these 20 observations, is a binomial distribution an appropriate model? Perform a goodness of fit procedure with   0.05. (b) Calculate the P-value for this test. 9-136. Consider the spot weld shear strength data in Exercise 6-31. Does the normal distribution seem to be a reasonable model for these data? Perform an appropriate goodness of fit test to answer this question. 9-137. Consider the water quality data in Exercise 6-32. (a) Do these data support the claim that mean concentration of suspended solids does not exceed 50 parts per million? Use   0.05. (b) What is the P-value for the test in part (a)? (c) Does the normal distribution seem to be a reasonable model for these data? Perform an appropriate goodness of fit test to answer this question. 9-138. Consider the golf ball overall distance data in Exercise 6-33. (a) Do these data support the claim that the mean overall distance for this brand of ball does not exceed 270 yards? Use   0.05. (b) What is the P-value for the test in part (a)? (c) Do these data appear to be well modeled by a normal distribution? Use a formal goodness of fit test in answering this question.

JWCL232_c09_283-350.qxd

348

1/14/10

9:50 PM

Page 348

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

9-139. Consider the baseball coefficient of restitution data in Exercise 8-92. If the mean coefficient of restitution exceeds 0.635, the population of balls from which the sample has been taken will be too “lively” and considered unacceptable for play. (a) Formulate an appropriate hypothesis testing procedure to answer this question. (b) Test these hypotheses and draw conclusions, using   0.01. (c) Find the P-value for this test. (d) In Exercise 8-92(b), you found a 99% confidence interval on the mean coefficient of restitution. Does this interval, or a one-sided CI, provide additional useful information to the decision maker? Explain why or why not. 9-140. Consider the dissolved oxygen data in Exercise 8-94. Water quality engineers are interested in knowing whether these data support a claim that mean dissolved oxygen concentration is 2.5 milligrams per liter. (a) Formulate an appropriate hypothesis testing procedure to investigate this claim. (b) Test these hypotheses and draw conclusions, using   0.05. (c) Find the P-value for this test. (d) In Exercise 8-94(b) you found a 95% CI on the mean dissolved oxygen concentration. Does this interval provide useful additional information beyond that of the hypothesis testing results? Explain your answer. 9-141. An article in Food Testing and Analysis [“Improving Reproducibility of Refractometry Measurements of Fruit Juices” (1999, Vol. 4, No. 4, pp. 13–17)] measured the sugar concentration (Brix) in clear apple juice. All readings were taken at 20C: 11.48 11.50

11.45 11.42

11.48 11.49

11.47 11.45

11.48 11.44

11.45 11.50

11.47 11.49

11.46 11.45

11.47 11.46

11.43 11.47

(a) Test the hypothesis H0:   11.5 versus H1:   11.5 using   0.05. Find the P-value. (b) Compute the power of the test if the true mean is 11.4. (c) What sample size would be required to detect a true mean sugar concentration of 11.45 if we wanted the power of the test to be at least 0.9? (d) Explain how the question in part (a) could be answered by constructing a two-sided confidence interval on the mean sugar concentration. (e) Is there evidence to support the assumption that the sugar concentration is normally distributed? 9-142. Consider the computer output below Test and Cl for One Proportion Test of p  0.25 vs p  0.25 X 53

N 225

Sample p 0.235556

Bound 0.282088

Z-Value ?

P-Value ?

Using the normal approximation. (a) Fill in the missing information. (b) What are your conclusions if   0.05? (c) The normal approximation to the binomial was used here. Was that appropriate? (d) Find a 95% upper-confidence bound on the true proportion. (e) What are the P-value and your conclusions if the alternative hypothesis is H1: p  0.25? 9-143. An article in Food Chemistry [“A Study of Factors Affecting Extraction of Peanut (Arachis Hypgaea L.) Solids with Water” (1991, Vol. 42, No. 2, pp. 153–165)] found the percent protein extracted from peanut milk as follows: 78.3 78.2

77.1 91.2

71.3 86.2

84.5 80.9

87.8 82.1

75.7 89.3

64.8 89.4

72.5 81.6

(a) Can you support a claim that mean percent protein extracted exceeds 80 percent? Use   0.05. (b) Is there evidence that percent protein extracted is normally distributed? (c) What is the P-value of the test statistic computed in part (a)? 9-144. An article in Biological Trace Element Research [“Interaction of Dietary Calcium, Manganese, and Manganese Source (Mn Oxide or Mn Methionine Complex) or Chick Performance and Manganese Utilization” (1991, Vol. 29, No. 3, pp. 217–228)] showed the following results of tissue assay for liver manganese (ppm) in chicks fed high-Ca diets. 6.02 5.29

6.08 5.84

7.11 6.03

5.73 5.99

5.32 4.53

7.10 6.81

(a) Test the hypothesis H0:  2  0.6 versus H1: 2  0.6 using   0.01. (b) What is the P-value for this test? (c) Discuss how part (a) could be answered by constructing a 99% two-sided confidence interval for . 9-145. An article in Experimental Brain Research [“Synapses in the Granule Cell Layer of the Rat Dentate Gyrus: Serial-Sectionin Study” (1996, Vol. 112, No. 2, pp. 237–243)] showed the ratio between the numbers of symmetrical and total synapses on somata and azon initial segments of reconstructed granule cells in the dentate gyrus of a 12-week-old rat: 0.65 0.91 0.50

0.90 0.86 0.68

0.78 0.53 1.00

0.94 0.84 0.57

0.40 0.42 1.00

0.84 0.96 0.89

0.9 0.56 0.60

0.91 0.67 0.54

0.92 0.96

0.96 0.52

0.94 0.50 1.00

(a) Use the data to test H0:  2  0.02 versus H1:  2  0.02 using   0.05. (b) Find the P-value for the test.

JWCL232_c09_283-350.qxd

1/14/10

3:08 PM

Page 349

9-9

9-146. An article in the Journal of Electronic Material [“Progress in CdZnTe Substrate Producibility and Critical Drive of IRFPA Yield Originating with CdZnTe Substrates” (1998, Vol. 27, No. 6, pp. 564–572)] improved the quality of CdZnTe substrates used to produce the HgCdTe infrared focal plane arrays (IRFPAs), also defined as sensor chip assemblies (SCAs). The cut-on wavelength 1m2 on 11 wafers was measured and is shown below: 6.06 6.16 6.57 6.67 6.98 6.17 6.17 6.93 6.73 6.87 6.76 (a) Is there evidence that the mean of cut-on wave length is not 6.50 m? (b) What is the P-value for this test? (c) What sample size would be required to detect a true mean cut-on wavelength of 6.25 m with probability 95%? (d) What is the type II error probability if the true mean cut-on wavelength is 6.95 m?

NONPARAMETRIC PROCEDURES

9-147. Consider the fatty acid measurements for the diet margarine described in Exercise 8-36. (a) For the sample size n  6, using a two-sided alternative hypothesis and   0.01, test H0: 2  1.0. (b) Suppose that instead of n  6, the sample size was n  51. Repeat the analysis performed in part (a) using n  51. (c) Compare your answers and comment on how sample size affects your conclusions drawn in parts (a) and (b). 9-148. Consider the television picture tube brightness experiment described in Exercise 8-35. (a) For the sample size n  10, do the data support the claim that the standard deviation of current is less than 20 microamps? (b) Suppose that instead of n  10, the sample size was 51. Repeat the analysis performed in part (a) using n  51. (c) Compare your answers and comment on how sample size affects your conclusions drawn in parts (a) and (b).

MIND-EXPANDING EXERCISES 9-149. Suppose that we wish to test H0:   0 versus H1:   0, where the population is normal with known

. Let 0    , and define the critical region so that we will reject H0 if z0  z or if z0  z , where z0 is the value of the usual test statistic for these hypotheses. (a) Show that the probability of type I error for this test is . (b) Suppose that the true mean is 1  0 . Derive an expression for for the above test. 9-150. Derive an expression for for the test on the variance of a normal distribution. Assume that the twosided alternative is specified. 9-151. When X1, X2, p , Xn are independent Poisson random variables, each with parameter , and n is large, the sample mean X has an approximate normal distribution with mean  and variance ␭n. Therefore, Z

X ␭ 1␭n

has approximately a standard normal distribution. Thus we can test H0:   0 by replacing  in Z by 0. When Xi are Poisson variables, this test is preferable to the largesample test of Section 9-2.3, which would use S  1n in the denominator, because it is designed just for the

349

Poisson distribution. Suppose that the number of open circuits on a semiconductor wafer has a Poisson distribution. Test data for 500 wafers indicate a total of 1038 opens. Using   0.05, does this suggest that the mean number of open circuits per wafer exceeds 2.0? 9-152. When X1, X2, p , Xn is a random sample from a normal distribution and n is large, the sample standard deviation has approximately a normal distribution with mean and variance 2 12n2 . Therefore, a large-sample test for H0:  0 can be based on the statistic Z

S 0

2 20 12n2

(a) Use this result to test H0:  10 versus H1:  10 for the golf ball overall distance data in Exercise 6-33. (b) Find an approximately unbiased estimator of the 95 percentile    1.645 . From the fact that X and S are independent random variables, find the standard error of the estimator of . How would you estimate the standard error? (c) Consider the golf ball overall distance data in Exercise 6-33. We wish to investigate a claim that

JWCL232_c09_283-350.qxd

350

1/14/10

9:52 PM

Page 350

CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE

MIND-EXPANDING EXERCISES the 95 percentile of overall distance does not exceed 285 yards. Construct a test statistic that can be used for testing the appropriate hypotheses. Apply this procedure to the data from Exercise 6-33. What are your conclusions? 9-153. Let X1, X2, p , Xn be a sample from an exponential distribution with parameter . It can be shown that

2␭ ni1 Xi has a chi-square distribution with 2n degrees of freedom. Use this fact to devise a test statistic and critical region for H0:   0 versus the three usual alternatives.

IMPORTANT TERMS AND CONCEPTS ␣ and ␤ Connection between hypothesis tests and confidence intervals Critical region for a test statistic Goodness of fit test Homogeneity test Hypothesis test Inference Independence test

Nonparametric or distribution free methods Normal approximation to nonparametric tests Null distribution Null hypothesis One- and two-sided alternative hypotheses

Operating characteristic (OC) curves Power of a test P-value Ranks Reference distribution for a test statistic Sample size determination for hypothesis tests

Significance level of a test Sign test Statistical hypotheses Statistical versus practical significance Test statistic Type I and type II errors Wilcoxon signed-rank test

JWCL232_c10_351-400.qxd

1/15/10

2:13 PM

Page 351

10

© Robert Dant/iStockphoto

Statistical Inference for Two Samples

The safety of drinking water is a serious public health issue. An article appeared in the Arizona Republic on May 27, 2001, that reported on arsenic contamination in the water sampled from 10 communities in the metropolitan Phoenix area and 10 communities from rural Arizona. The data showed dramatic differences in the arsenic concentration, ranging from 3 parts per billion (ppb) to 48 ppb. There are some important questions suggested by this article. Is there a real difference between the arsenic concentrations in the Phoenix area and in the rural communities in Arizona? How large is this difference? Is it large enough to require action on the part of the public health service and other state agencies to correct the problem? Are the levels of reported arsenic concentration large enough to constitute a public health risk? Some of these questions can be answered by statistical methods. If we think of the metropolitan Phoenix communities as one population and the rural Arizona communities as a second population, we could determine whether there is a statistically significant difference in the mean arsenic concentration between the two populations by testing the hypothesis that the two means, say, 1 and 2, are different. This is a relatively simple extension to two samples of the one-sample hypothesis testing procedures of Chapter 9. We could also use a confidence interval to estimate the difference in the two means, say, 1  2. The arsenic concentration problem is very typical of many problems in engineering and science that involve statistics. Some of the questions can be answered by the application of appropriate statistical tools, while other questions require using engineering or scientific knowledge and expertise to answer satisfactorily.

CHAPTER OUTLINE 10-1 INFERENCE ON THE DIFFERENCE IN MEANS OF TWO NORMAL DISTRIBUTIONS, VARIANCES KNOWN 10-1.1 Hypothesis Tests on the Difference in Means, Variances Known

10-1.2 Type II Error and Choice of Sample Size 10-1.3 Confidence Interval on the Difference in Means, Variances Known

351

JWCL232_c10_351-400.qxd

352

1/15/10

2:13 PM

Page 352

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

10-2 INFERENCE ON THE DIFFERENCE IN MEANS OF TWO NORMAL DISTRIBUTIONS, VARIANCES UNKNOWN 10-2.1 Hypothesis Tests on the Difference in Means, Variances Unknown 10-2.2 Type II Error and Choice of Sample Size 10-2.3 Confidence Interval on the Difference in Means, Variances Unknown 10-3 A NONPARAMETRIC TEST ON THE DIFFERENCE IN TWO MEANS 10-3.1 Description of the Wilcoxon Rank-Sum Test 10-3.2 Large-Sample Approximation 10-3.3 Comparison to the t-Test 10-4 PAIRED t-TEST 10-5 INFERENCE ON THE VARIANCES OF TWO NORMAL DISTRIBUTIONS

10-5.1 F Distribution 10-5.2 Hypothesis Tests on the Ratio of Two Variances 10-5.3 Type II Error and Choice of Sample Size 10-5.4 Confidence Interval on the Ratio of Two Variances 10-6 INFERENCE ON TWO POPULATION PROPORTIONS 10-6.1 Large-Sample Tests on the Difference in Population Proportions 10-6.2 Type II Error and Choice of Sample Size 10-6.3 Confidence Interval on the Difference in Population Proportions 10-7 SUMMARY TABLE AND ROADMAP FOR INFERENCE PROCEDURES FOR TWO SAMPLES

LEARNING OBJECTIVES After careful study of this chapter you should be able to do the following: 1. Structure comparative experiments involving two samples as hypothesis tests 2. Test hypotheses and construct confidence intervals on the difference in means of two normal distributions 3. Test hypotheses and construct confidence intervals on the ratio of the variances or standard deviations of two normal distributions 4. Test hypotheses and construct confidence intervals on the difference in two population proportions 5. Use the P-value approach for making decisions in hypotheses tests 6. Compute power, type II error probability, and make sample size decisions for two-sample tests on means, variances, and proportions 7. Explain and use the relationship between confidence intervals and hypothesis tests

10-1 INFERENCE ON THE DIFFERENCE IN MEANS OF TWO NORMAL DISTRIBUTIONS, VARIANCES KNOWN The previous two chapters presented hypothesis tests and confidence intervals for a single population parameter (the mean , the variance 2, or a proportion p). This chapter extends those results to the case of two independent populations. The general situation is shown in Fig. 10-1. Population 1 has mean 1 and variance  21, while population 2 has mean 2 and variance  22. Inferences will be based on two random samples of sizes n1 and n2, respectively. That is, X11, X12, p, X1n1 is a random sample of n1 observations from population 1, and X21, X22, p, X2n2 is a random sample of n2 observations from population 2. Most of the practical applications of the procedures in this chapter

JWCL232_c10_351-400.qxd

1/15/10

2:13 PM

Page 353

10-1 INFERENCE ON THE DIFFERENCE IN MEANS OF TWO NORMAL DISTRIBUTIONS, VARIANCES KNOWN Population 1

Population 2

2 ␴1

Figure 10-1 Two independent populations.

353

␮1

2

␴2

Sample 1: x11, x12,…, x1n

␮2

1

Sample 2: x21, x22,…, x2n

2

arise in the context of simple comparative experiments in which the objective is to study the difference in the parameters of the two populations. Engineers and scientists are often interested in comparing two different conditions to determine whether either condition produces a significant effect on the response that is observed. These conditions are sometimes called treatments. Example 10-1 illustrates such an experiment; the two different treatments are two paint formulations, and the response is the drying time. The purpose of the study is to determine whether the new formulation results in a significant effect—reducing drying time. In this situation, the product developer (the experimenter) randomly assigned 10 test specimens to one formulation and 10 test specimens to the other formulation. Then the paints were applied to the test specimens in random order until all 20 specimens were painted. This is an example of a completely randomized experiment. When statistical significance is observed in a randomized experiment, the experimenter can be confident in the conclusion that it was the difference in treatments that resulted in the difference in response. That is, we can be confident that a cause-and-effect relationship has been found. Sometimes the objects to be used in the comparison are not assigned at random to the treatments. For example, the September 1992 issue of Circulation (a medical journal published by the American Heart Association) reports a study linking high iron levels in the body with increased risk of heart attack. The study, done in Finland, tracked 1931 men for five years and showed a statistically significant effect of increasing iron levels on the incidence of heart attacks. In this study, the comparison was not performed by randomly selecting a sample of men and then assigning some to a “low iron level” treatment and the others to a “high iron level” treatment. The researchers just tracked the subjects over time. Recall from Chapter 1 that this type of study is called an observational study. It is difficult to identify causality in observational studies, because the observed statistically significant difference in response between the two groups may be due to some other underlying factor (or group of factors) that was not equalized by randomization and not due to the treatments. For example, the difference in heart attack risk could be attributable to the difference in iron levels, or to other underlying factors that form a reasonable explanation for the observed results—such as cholesterol levels or hypertension. In this section we consider statistical inferences on the difference in means 1  2 of two normal distributions, where the variances  21 and  22 are known. The assumptions for this section are summarized as follows. Assumptions for Two-Sample Inference

1. X11, X12, p , X1n1 is a random sample from population 1. 2. X21, X22, p , X2n2 is a random sample from population 2. 3. The two populations represented by X1 and X2 are independent. 4. Both populations are normal.

JWCL232_c10_351-400.qxd

354

1/15/10

2:13 PM

Page 354

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

A logical point estimator of 1  2 is the difference in sample means X1  X2. Based on the properties of expected values, E1X1  X2 2  E1X1 2  E1X 2 2  1  2 and the variance of X1  X2 is  21  22 V1X1  X2 2  V1X1 2 V1X2 2  n n 1 2 Based on the assumptions and the preceding results, we may state the following.

The quantity Z

X1  X2  11  2 2  21  22 n2 B n1

(10-1)

has a N(0, 1) distribution.

This result will be used to form tests of hypotheses and confidence intervals on 1  2. ˆ X X Essentially, we may think of 1  2 as a parameter , and its estimator is 1 2 2 with variance  ˆ  12n1  22 n2. If 0 is the null hypothesis value specified for , the test ˆ  2   ˆ . Notice how similar this is to the test statistic for a single mean statistic will be 1 0 used in Equation 9-8 of Chapter 9.

10-1.1 Hypothesis Tests on the Difference in Means, Variances Known We now consider hypothesis testing on the difference in the means 1  2 of two normal populations. Suppose that we are interested in testing that the difference in means 1  2 is equal to a specified value 0. Thus, the null hypothesis will be stated as H0: 1  2  0. Obviously, in many cases, we will specify 0  0 so that we are testing the equality of two means (i.e., H0: 1  2). The appropriate test statistic would be found by replacing 1  2 in Equation 10-1 by 0, and this test statistic would have a standard normal distribution under H0. That is, the standard normal distribution is the reference distribution for the test statistic. Suppose that the alternative hypothesis is H1: 1  2  0. Now, a sample value of x1  x2 that is considerably different from 0 is evidence that H1 is true. Because Z0 has the N(0, 1) distribution when H0 is true, we would calculate the P-value as the sum of the probabilities beyond the test statistic value z0 and z0 in the standard normal distribution. That is, P  231  1 0 z0 0 24 . This is exactly what we did in the one-sample z-test of Section 4-4.1. If we wanted to perform a fixed-significance-level test, we would take z/2 and z/2 as the boundaries of the critical region just as we did in the single-sample z-test. This would give a test with level of significance . P-values or critical regions for the one-sided

JWCL232_c10_351-400.qxd

1/18/10

11:44 AM

Page 355

10-1 INFERENCE ON THE DIFFERENCE IN MEANS OF TWO NORMAL DISTRIBUTIONS, VARIANCES KNOWN

355

alternatives would be determined similarly. Formally, we summarize these results in the following display.

Tests on the Difference in Means, Variances Known

Null hypothesis:

H0: ␮1 ⫺ ␮2 ⫽ ⌬0

Test statistic:

Z0 ⫽

X1 ⫺ X2 ⫺ ⌬0

Rejection Criterion For for Fixed-Level Tests

Alternative Hypotheses

P-Value

H 1: ␮1 ⫺ ␮2 ⫽ ⌬0

Probability above |z0| and probability below ⫺|z0|,

z0 ⬎ z␣Ⲑ 2 or z0 ⬍ ⫺z␣Ⲑ2

H 1: ␮1 ⫺ ␮2 ⬎ ⌬0

Probability above z0,

z0 ⬎ z␣

H 1: ␮1 ⫺ ␮2 ⬍ ⌬0

Probability below z0,

z0 ⬍ ⫺z␣

˛

˛

˛

EXAMPLE 10-1

(10-2)

␴21 ␴22 ⫹ n2 B n1

P ⫽ 231 ⫺ ⌽1|z0|2 4 P ⫽ 1 ⫺ ⌽1z0 2 P ⫽ ⌽1z0 2

Paint Drying Time

A product developer is interested in reducing the drying time of a primer paint. Two formulations of the paint are tested; formulation 1 is the standard chemistry, and formulation 2 has a new drying ingredient that should reduce the drying time. From experience, it is known that the standard deviation of drying time is 8 minutes, and this inherent variability should be unaffected by the addition of the new ingredient. Ten specimens are painted with formulation 1, and another 10 specimens are painted with formulation 2; the 20 specimens are painted in random order. The two sample average drying times are x 1 ⫽ 121 minutes and x 2 ⫽ 112 minutes, respectively. What conclusions can the product developer draw about the effectiveness of the new ingredient, using ␣ ⫽ 0.05? We apply the seven-step procedure to this problem as follows: ˛

˛

1. Parameter of interest: The quantity of interest is the difference in mean drying times, ␮1 ⫺ ␮2, and ⌬0 ⫽ 0. 2. Non hypothesis: H 0: ␮1 ⫺ ␮2 ⫽ 0, or H 0: ␮1 ⫽ ␮2. ˛

˛

˛

3. Alternative hypothesis: H 1: ␮1 ⬎ ␮2. We want to reject H0 if the new ingredient reduces mean drying time. ˛

4. Test statistic: The test statistic is z0 ⫽ ˛

x1 ⫺ x2 ⫺ 0 ␴21 ␴2 ⫹ 2 n2 B n1

where ␴21 ⫽ ␴22 ⫽ 182 2 ⫽ 64 and n1 ⫽ n2 ⫽ 10.

5. Reject H0 if: Reject H0: ␮1 ⫽ ␮2 if the P-value is less than 0.05. 6. Computations: Since x1 ⫽ 121 minutes and x2 ⫽ 112 minutes, the test statistic is z0 ⫽

121 ⫺ 112 182 2

B 10

⫹

182 2

⫽ 2.52

10

7. Conclusion: Since z0 ⫽ 2.52, the P-value is P ⫽ 1 ⫺ ⌽12.522 ⫽ 0.0059, so we reject H0 at the ␣ ⫽ 0.05 level Practical Interpretation: We conclude that adding the new ingredient to the paint significantly reduces the drying time. This is a strong conclusion.

When the population variances are unknown, the sample variances s12 and s22 can be substituted into the test statistic Equation 10-2 to produce a large-sample test for the difference in means. This procedure will also work well when the populations are not necessarily normally distributed. However, both n1 and n2 should exceed 40 for this large-sample test to be valid.

JWCL232_c10_351-400.qxd

356

1/15/10

2:13 PM

Page 356

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

10-1.2 Type II Error and Choice of Sample Size Use of Operating Characteristic Curves The operating characteristic curves (OC) in Appendix Charts VIIa, VIIb, VIIc, and VIId may be used to evaluate the type II error probability for the hypotheses in the display (10-2). These curves are also useful in determining sample size. Curves are provided for   0.05 and   0.01. For the two-sided alternative hypothesis, the abscissa scale of the operating characteristic curve in charts VIIa and VIIb is d, where d

ƒ 1  2  0 ƒ 221



22



ƒ   0 ƒ 221 22

(10-3)

and one must choose equal sample sizes, say, n  n1  n2. The one-sided alternative hypotheses require the use of Charts VIIc and VIId. For the one-sided alternatives H1: 1  2 0 or H1: 1  2  0, the abscissa scale is also given by d

ƒ 1  2  0 ƒ 221



22



ƒ   0 ƒ 221 22

It is not unusual to encounter problems where the costs of collecting data differ substantially between the two populations, or where one population variance is much greater than the other. In those cases, we often use unequal sample sizes. If n1  n2, the operating characteristic curves may be entered with an equivalent value of n computed from n

21

21 22 n1 22n2

(10-4)

If n1  n2, and their values are fixed in advance, Equation 10-4 is used directly to calculate n, and the operating characteristic curves are entered with a specified d to obtain . If we are given d and it is necessary to determine n1 and n2 to obtain a specified , say, *, we guess at trial values of n1 and n2, calculate n in Equation 10-4, and enter the curves with the specified value of d to find . If   *, the trial values of n1 and n2 are satisfactory. If   *, adjustments to n1 and n2 are made and the process is repeated. EXAMPLE 10-2

Paint Drying Time, Sample Size from OC Curves

Consider the paint drying time experiment from Example 10-1. If the true difference in mean drying times is as much as 10 minutes, find the sample sizes required to detect this difference with probability at least 0.90. The appropriate value of the abscissa parameter is (since 0  0, and   10)

d

10 ƒ 1  2 ƒ   0.88 221 22 282 82

and since the detection probability or power of the test must be at least 0.9, with   0.05, we find from Appendix Chart VIIc that n  n1  n2 ⯝ 11.

Sample Size Formulas It is also possible to obtain formulas for calculating the sample sizes directly. Suppose that the null hypothesis H0: 1  2  0 is false and that the true difference in means is 1  2  , where  0. One may find formulas for the sample size required to obtain a specific value of the type II error probability  for a given difference in means  and level of significance .

JWCL232_c10_351-400.qxd

1/15/10

2:13 PM

Page 357

10-1 INFERENCE ON THE DIFFERENCE IN MEANS OF TWO NORMAL DISTRIBUTIONS, VARIANCES KNOWN

357

For example, we first write the expression for the -error for the two-sided alternative, which is    ± z 2 

  0 21

22

n 2 B n1

≤   ± z 2 

  0 21 22 n n 2 B 1

≤

The derivation for sample size closely follows the single-sample case in Section 9-2.2. Sample Size for a Two-Sided Test on the Difference in Means with n1 ⴝ n2, Variances Known

For the two-sided alternative hypothesis with significance level , the sample size n1  n2  n required to detect a true difference in means of  with power at least 1   is 1z2 z 2 2 121 22 2

n⯝

1  0 2 2

(10-5)

This approximation is valid when 1z2  1  0 2 1n 121 22 2 is small compared to . Sample Size for a One-Sided Test on the Difference in Means with n1 ⴝ n2, Variances Known

For a one-sided alternative hypothesis with significance level , the sample size n1  n2  n required to detect a true difference in means of (0) with power at least 1   is n

1z z 2 2 121 22 2 1  0 2 2

(10-6)

where  is the true difference in means of interest. Then by following a procedure similar to that used to obtain Equation 9-17, the expression for  can be obtained for the case where n  n1  n2. EXAMPLE 10-3

Paint Drying Time Sample Size

To illustrate the use of these sample size equations, consider the situation described in Example 10-1, and suppose that if the true difference in drying times is as much as 10 minutes, we want to detect this with probability at least 0.90. Under the null hypothesis, 0  0. We have a one-sided alternative hypothesis with   10,   0.05 (so z  z0.05  1.645), and since the power is 0.9,   0.10 (so z  z0.10  1.28). Therefore, we may find the required sample size from Equation 10-6 as follows:

n 

1z z 2 2 121 22 2 1  0 2 2

11.645 1.282 2 3 182 2 182 2 4 110  02 2

 11

This is exactly the same as the result obtained from using the OC curves.

10-1.3 Confidence Interval on the Difference in Means, Variances Known The 100(1  )% confidence interval on the difference in two means 1  2 when the variances are known can be found directly from results given previously in this section. Recall that X11, X12, p , X1n1 is a random sample of n1 observations from the first population and X21,

JWCL232_c10_351-400.qxd

358

1/15/10

2:13 PM

Page 358

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

X22, p , X2n2 is a random sample of n2 observations from the second population. The difference in sample means X1  X2 is a point estimator of 1  2, and Z

X1  X2  11  2 2 21 22 n2 B n1

has a standard normal distribution if the two populations are normal or is approximately standard normal if the conditions of the central limit theorem apply, respectively. This implies that P1z 2  Z  z 2 2  1  , or P ≥ z 2 

X1  X2  11  2 2 21 22 n2 B n1

 z 2 ¥  1  

This can be rearranged as 21 22 21 22      X  X z 1 2 1 2  2  n2 n2 b  1   B n1 B n1

P aX1  X2  z 2

Therefore, the 100(1  )% confidence interval for 1  2 is defined as follows. Confidence Interval on the Difference in Means, Variances Known

If x1 and x2 are the means of independent random samples of sizes n1 and n2 from two independent normal populations with known variances 21 and 22, respectively, a 100(1 ⴚ ␣)% confidence interval for ␮1 ⴚ ␮2 is 21 22 21 22 x1  x2  z2 n n  1  2  x1  x2 z2 n n 2 2 B 1 B 1

(10-7)

where z兾2 is the upper 兾2 percentage point of the standard normal distribution.

The confidence level 1   is exact when the populations are normal. For nonnormal populations, the confidence level is approximately valid for large sample sizes. EXAMPLE 10-4

Aluminum Tensile Strength

Tensile strength tests were performed on two different grades of aluminum spars used in manufacturing the wing of a commercial transport aircraft. From past experience with the spar manufacturing process and the testing procedure, the standard deviations of tensile strengths are assumed to be known. The data obtained are as follows: n1  10, x1  87.6, 1  1, n2  12, x2  74.5, and 2  1.5. If 1 and 2 denote the true mean tensile strengths for the two grades of spars, we may find a 90% confidence interval on the difference in mean strength 1  2 as follows:

21 22 x1  x2  z2 n n  1  2 2 B 1 ˛

21 22  x1  x2 z2 n n 2 B 1 ˛

87.6  74.5  1.645

˛

112 2

B 10



11.52 2 12

 1  2

 87.6  74.5 1.645

112 2

B 10



11.52 2 12

JWCL232_c10_351-400.qxd

1/15/10

2:13 PM

Page 359

10-1 INFERENCE ON THE DIFFERENCE IN MEANS OF TWO NORMAL DISTRIBUTIONS, VARIANCES KNOWN

Therefore, the 90% confidence interval on the difference in mean tensile strength (in kilograms per square millimeter) is 12.22  1  2  13.98 (in kilograms per square millimeter)

359

Practical Interpretation: Notice that the confidence interval does not include zero, implying that the mean strength of aluminum grade 1 (1) exceeds the mean strength of aluminum grade 2 (2). In fact, we can state that we are 90% confident that the mean tensile strength of aluminum grade 1 exceeds that of aluminum grade 2 by between 12.22 and 13.98 kilograms per square millimeter.

Choice of Sample Size If the standard deviations 1 and 2 are known (at least approximately) and the two sample sizes n1 and n2 are equal (n1  n2  n, say), we can determine the sample size required so that the error in estimating 1  2 by x1  x2 will be less than E at 100(1  )% confidence. The required sample size from each population is Sample Size for a Confidence Interval on the Difference in Means, Variances Known

na

z  2 E

2

b 121 22 2

(10-8)

Remember to round up if n is not an integer. This will ensure that the level of confidence does not drop below 100(1  )%. One-Sided Confidence Bounds One-sided confidence bounds on 1  2 may also be obtained. A 100(1  )% upperconfidence bound on 1  2 is One-Sided Upper Confidence Bound

21 22 n2 B n1

(10-9)

21 22 n2  1  2 B n1

(10-10)

1  2  x1  x2 z

and a 100(1  )% lower-confidence bound is One-Sided Lower Confidence Bound

x1  x2  z

EXERCISES FOR SECTION 10-1 10-1. Consider the hypothesis test H0 : 1  2 against H1 : 1  2 with known variances 1  10 and 2  5. Suppose that sample sizes n1  10 and n2  15 and that x1  4.7 and x2  7.8. Use   0.05. (a) Test the hypothesis and find the P-value. (b) Explain how the test could be conducted with a confidence interval. (c) What is the power of the test in part (a) for a true difference in means of 3? (d) Assuming equal sample sizes, what sample size should be used to obtain   0.05 if the true difference in means is 3? Assume that   0.05.

10-2. Consider the hypothesis test H0 : 1  2 against H1 : 1  2 with known variances 1  10 and 2  5. Suppose that sample sizes n1  10 and n2  15 and that x1  14.2 and x2  19.7. Use   0.05. (a) Test the hypothesis and find the P-value. (b) Explain how the test could be conducted with a confidence interval. (c) What is the power of the test in part (a) if 1 is 4 units less than 2? (d) Assuming equal sample sizes, what sample size should be used to obtain   0.05 if 1 is 4 units less than 2? Assume that   0.05.

JWCL232_c10_351-400.qxd

360

1/15/10

2:13 PM

Page 360

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

10-3. Consider the hypothesis test H0 : 1  2 against H1 : 1 2 with known variances 1  10 and 2  5. Suppose that sample sizes n1  10 and n2  15 and that x1  24.5 and x2  21.3. Use   0.01. (a) Test the hypothesis and find the P-value. (b) Explain how the test could be conducted with a confidence interval. (c) What is the power of the test in part (a) if 1 is 2 units greater than 2? (d) Assuming equal sample sizes, what sample size should be used to obtain   0.05 if 1 is 2 units greater than 2? Assume that   0.05. 10-4. Two machines are used for filling plastic bottles with a net volume of 16.0 ounces. The fill volume can be assumed normal, with standard deviation 1  0.020 and 2  0.025 ounces. A member of the quality engineering staff suspects that both machines fill to the same mean net volume, whether or not this volume is 16.0 ounces. A random sample of 10 bottles is taken from the output of each machine. Machine 1 16.03 16.04 16.05 16.05 16.02

16.01 15.96 15.98 16.02 15.99

Machine 2 16.02 15.97 15.96 16.01 15.99

16.03 16.04 16.02 16.01 16.00

(a) Do you think the engineer is correct? Use   0.05. What is the P-value for this test? (b) Calculate a 95% confidence interval on the difference in means. Provide a practical interpretation of this interval. (c) What is the power of the test in part (a) for a true difference in means of 0.04? (d) Assuming equal sample sizes, what sample size should be used to assure that   0.05 if the true difference in means is 0.04? Assume that   0.05. 10-5. Two types of plastic are suitable for use by an electronics component manufacturer. The breaking strength of this plastic is important. It is known that 1  2  1.0 psi. From a random sample of size n1  10 and n2  12, we obtain x1  162.5 and x2  155.0 . The company will not adopt plastic 1 unless its mean breaking strength exceeds that of plastic 2 by at least 10 psi. (a) Based on the sample information, should it use plastic 1? Use   0.05 in reaching a decision. Find the P-value. (b) Calculate a 95% confidence interval on the difference in means. Suppose that the true difference in means is really 12 psi. (c) Find the power of the test assuming that   0.05. (d) If it is really important to detect a difference of 12 psi, are the sample sizes employed in part (a) adequate, in your opinion?

10-6. The burning rates of two different solid-fuel propellants used in aircrew escape systems are being studied. It is known that both propellants have approximately the same standard deviation of burning rate; that is 1  2  3 centimeters per second. Two random samples of n1  20 and n2  20 specimens are tested; the sample mean burning rates are x1  18 centimeters per second and x2  24 centimeters per second. (a) Test the hypothesis that both propellants have the same mean burning rate. Use   0.05. What is the P-value? (b) Construct a 95% confidence interval on the difference in means 1  2. What is the practical meaning of this interval? (c) What is the -error of the test in part (a) if the true difference in mean burning rate is 2.5 centimeters per second? (d) Assuming equal sample sizes, what sample size is needed to obtain power of 0.9 at a true difference in means of 14 cm/s? 10-7. Two different formulations of an oxygenated motor fuel are being tested to study their road octane numbers. The variance of road octane number for formulation 1 is  21  1.5, and for formulation 2 it is 22  1.2. Two random samples of size n1  15 and n2  20 are tested, and the mean road octane numbers observed are x1  89.6 and x2  92.5. Assume normality. (a) If formulation 2 produces a higher road octane number than formulation 1, the manufacturer would like to detect it. Formulate and test an appropriate hypothesis, using   0.05. What is the P-value? (b) Explain how the question in part (a) could be answered with a 95% confidence interval on the difference in mean road octane number. (c) What sample size would be required in each population if we wanted to be 95% confident that the error in estimating the difference in mean road octane number is less than 1? 10-8. A polymer is manufactured in a batch chemical process. Viscosity measurements are normally made on each batch, and long experience with the process has indicated that the variability in the process is fairly stable with   20. Fifteen batch viscosity measurements are given as follows: 724, 718, 776, 760, 745, 759, 795, 756, 742, 740, 761, 749, 739, 747, 742 A process change is made which involves switching the type of catalyst used in the process. Following the process change, eight batch viscosity measurements are taken: 735, 775, 729, 755, 783, 760, 738, 780 Assume that process variability is unaffected by the catalyst change. If the difference in mean batch viscosity is 10 or less, the manufacturer would like to detect it with a high probability. (a) Formulate and test an appropriate hypothesis using   0.10. What are your conclusions? Find the P-value.

JWCL232_c10_351-400.qxd

1/15/10

2:14 PM

Page 361

10-2 INFERENCE ON THE DIFFERENCE IN MEANS OF TWO NORMAL DISTRIBUTIONS, VARIANCES UNKNOWN

(b) Find a 90% confidence interval on the difference in mean batch viscosity resulting from the process change. (c) Compare the results of parts (a) and (b) and discuss your findings. 10-9. The concentration of active ingredient in a liquid laundry detergent is thought to be affected by the type of catalyst used in the process. The standard deviation of active concentration is known to be 3 grams per liter, regardless of the catalyst type. Ten observations on concentration are taken with each catalyst, and the data follow: Catalyst 1: 57.9, 66.2, 65.4, 65.4, 65.2, 62.6, 67.6, 63.7, 67.2, 71.0 Catalyst 2: 66.4, 71.7, 70.3, 69.3, 64.8, 69.6, 68.6, 69.4, 65.3, 68.8

361

(a) Find a 95% confidence interval on the difference in mean active concentrations for the two catalysts. Find the P-value. (b) Is there any evidence to indicate that the mean active concentrations depend on the choice of catalyst? Base your answer on the results of part (a). (c) Suppose that the true mean difference in active concentration is 5 grams per liter. What is the power of the test to detect this difference if   0.05? (d) If this difference of 5 grams per liter is really important, do you consider the sample sizes used by the experimenter to be adequate? Does the assumption of normality seem reasonable for both samples?

10-2 INFERENCE ON THE DIFFERENCE IN MEANS OF TWO NORMAL DISTRIBUTIONS, VARIANCES UNKNOWN We now extend the results of the previous section to the difference in means of the two distributions in Fig. 10-1 when the variances of both distributions  21 and  22 are unknown. If the sample sizes n1 and n2 exceed 40, the normal distribution procedures in Section 10-1 could be used. However, when small samples are taken, we will assume that the populations are normally distributed and base our hypotheses tests and confidence intervals on the t distribution. This nicely parallels the case of inference on the mean of a single sample with unknown variance.

10-2.1 Hypotheses Tests on the Difference in Means, Variances Unknown We now consider tests of hypotheses on the difference in means 1  2 of two normal distributions where the variances  21 and  22 are unknown. A t-statistic will be used to test these hypotheses. As noted above and in Section 9-3, the normality assumption is required to develop the test procedure, but moderate departures from normality do not adversely affect the procedure. Two different situations must be treated. In the first case, we assume that the variances of the two normal distributions are unknown but equal; that is,  21  22  2. In the second, we assume that  21 and 22 are unknown and not necessarily equal. Case 1: 12  22  2 Suppose we have two independent normal populations with unknown means 1 and 2, and unknown but equal variances,  21  22  2. We wish to test H0 : 1  2  0 H1 : 1  2  0

(10-11)

Let X11, X12, p , X1n1 be a random sample of n1 observations from the first population and X21, X22, p , X2n2 be a random sample of n2 observations from the second population. Let X1, X2, S 21, and S 22 be the sample means and sample variances, respectively. Now the expected value of the difference in sample means X1  X2 is E1X1  X2 2  1  2, so X1  X2 is an unbiased estimator of the difference in means. The variance of X1  X2 is 2 2 1 1 V1X1  X2 2  n n  2 a n n b 1 2 1 2

JWCL232_c10_351-400.qxd

362

1/15/10

2:14 PM

Page 362

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

It seems reasonable to combine the two sample variances S 21 and S 22 to form an estimator of  . The pooled estimator of 2 is defined as follows. 2

Pooled Estimator of Variance

The pooled estimator of 2, denoted by S 2p, is defined by S p2 

1n1  12S21 1n2  12S22 n1 n2  2

(10-12)

It is easy to see that the pooled estimator S 2p can be written as S 2p 

n1  1 n2  1 S 21 S 2  wS 21 11  w2S 22 n1 n2  2 n1 n2  2 2

where 0  w  1. Thus S p2 is a weighted average of the two sample variances S12 and S22, where the weights w and 1  w depend on the two sample sizes n1 and n2. Obviously, if n1  n2  n, w  0.5 and Sp2 is just the arithmetic average of S12 and S22. If n1  10 and n2  20 (say), w  0.32 and 1  w  0.68. The first sample contributes n1  1 degrees of freedom to Sp2 and the second sample contributes n2  1 degrees of freedom. Therefore, Sp2 has n1 n2  2 degrees of freedom. Now we know that Z

X1  X2  11  2 2 1 1 n 2 B n1



has a N(0, 1) distribution. Replacing  by Sp gives the following.

Given the assumptions of this section, the quantity T

X1  X2  11  2 2 Sp

1 1 n 2 B n1

(10-13)

has a t distribution with n1 n2  2 degrees of freedom.

The use of this information to test the hypotheses in Equation 10-11 is now straightforward: Simply replace 1  2 by 0, and the resulting test statistic has a t distribution with n1 n2  2 degrees of freedom under H0: 1  2  0. Therefore, the reference distribution for the test statistic is the t distribution with n1 n2  2 degrees of freedom. The calculation of P-values and the location of the critical region for fixed-significance-level testing for both two- and one-sided alternatives parallels those in the one-sample case. Because a pooled estimate of variance is used, the procedure is often called the pooled t-test.

JWCL232_c10_351-400.qxd

1/15/10

2:14 PM

Page 363

363

10-2 INFERENCE ON THE DIFFERENCE IN MEANS OF TWO NORMAL DISTRIBUTIONS, VARIANCES UNKNOWN

Tests on the Difference in Means of Two Normal Distributions, Variances Unknown and Equal*

Null hypothesis:

H0: 1  2  0

Test statistic:

T0 

Sp

Alternative Hypothesis H1: 1  2  0 H1: 1  2 0 H1: 1  2  0 EXAMPLE 10-5

(10-14)

1 1 n 2 B n1

Rejection Criterion for Fixed-Level Tests t0 t2,n1 n22 or t0  t2,n1 n22 t0 t,n1 n22 t0  t,n1 n22

P-Value Probability above |t0| and probability below|t0| Probability above t0 Probability below t0

Yield from a Catalyst

Two catalysts are being analyzed to determine how they affect the mean yield of a chemical process. Specifically, catalyst 1 is currently in use, but catalyst 2 is acceptable. Since catalyst 2 is cheaper, it should be adopted, providing it does not change the process yield. A test is run in the pilot plant and results in the data shown in Table 10-1. Is there any difference between the mean yields? Use   0.05, and assume equal variances. The solution using the seven-step hypothesis-testing procedure is as follows: 1.

X1  X2  0

Parameter of interest: The parameters of interest are 1 and 2, the mean process yield using catalysts 1 and 2, respectively, and we want to know if 1  2  0.

2.

Null hypothesis: H0: 1  2  0, or H0: 1  2

3.

Alternative hypothesis: H1: 1  2

4.

Test statistic: The test statistic is t0 

x1  x2  0 sp

1 1 n n 1 2 B

5.

Reject H0 if: Reject H0 if the P-value is less than 0.05.

6.

Computations: From Table 10-1 we have x1  92.255, s1  2.39, n1  8, x2  92.733, s2  2.98, and n2  8. Therefore

s2p 

1n1  12s12 1n2  12s22 n1 n2  2



17212.392 2 712.982 2 8 82

 7.30

sp  27.30  2.70

Table 10-1 Catalyst Yield Data, Example 10-5 Observation Number

Catalyst 1

Catalyst 2

1 2 3 4 5 6 7 8

91.50 94.18 92.18 95.39 91.79 89.07 94.72 89.21

89.19 90.95 90.46 93.21 97.19 97.04 91.07 92.75

x1  92.255 s1  2.39

x2  92.733 s2  2.98

*While we have given the development of this procedure for the case where the sample sizes could be different, there is an advantage to using equal sample sizes n1  n2  n. When the sample sizes are the same from both populations, the t-test is more robust to the assumption of equal variances.

JWCL232_c10_351-400.qxd

364

1/15/10

2:14 PM

Page 364

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

since 0.258  0.35  0.692, we conclude that lower and upper bounds on the P-value are 0.50  P  0.80. Therefore, since the P-value exceeds   0.05, the null hypothesis cannot be rejected.

and t0 

x1  x2 2.70

˛

92.255  92.733 1 1 8 B8

 0.35

2.70

Conclusions: Since 2.145, from Appendix Table V we find that t0.40,14  0.258 and t0.25,14  0.692. Therefore,

Practical Interpretation: At the 0.05 level of significance, we do not have strong evidence to conclude that catalyst 2 results in a mean yield that differs from the mean yield when catalyst 1 is used.

The Minitab two-sample t-test and confidence interval procedure for Example 10-5 follows: Minitab Computations Two-Sample T-Test and CI: Cat 1, Cat 2 Two-sample T for Cat 1 vs Cat 2 N 8 8

Cat 1 Cat 2

Mean 92.26 92.73

StDev 2.39 2.99

SE Mean 0.84 1.1

Difference  mu Cat 1  mu Cat 2 Estimate for difference: 0.48 95% CI for difference: (3.37, 2.42) T-Test of difference  0 (vs not  ): T-Value  0.35 P-Value  0.730 DF  14 Both use Pooled StDev  2.70

Notice that the numerical results are essentially the same as the manual computations in Example 10-5. The P-value is reported as P  0.73. The two-sided CI on 1  2 is also reported. We will give the computing formula for the CI in Section 10-2.3. Figure 10-2 shows the normal probability plot of the two samples of yield data and comparative box plots. The normal probability plots indicate that there is no problem with the normality assumption. 99 98 95 90 96

80 70 60 50 40 30 20

94 Yield

Percentage

7.

1 1 n 2 B n1



92

10 5

Cat 1 Cat 2

1 88

93 Yield data (a)

98

90

88

2

1 Catalyst type (b)

Figure 10-2 Normal probability plot and comparative box plot for the catalyst yield data in Example 10-5. (a) Normal probability plot, (b) Box plots.

JWCL232_c10_351-400.qxd

1/15/10

2:14 PM

Page 365

10-2 INFERENCE ON THE DIFFERENCE IN MEANS OF TWO NORMAL DISTRIBUTIONS, VARIANCES UNKNOWN

365

Furthermore, both straight lines have similar slopes, providing some verification of the assumption of equal variances. The comparative box plots indicate that there is no obvious difference in the two catalysts, although catalyst 2 has slightly greater sample variability. Case 2: ␴ 21 ⴝ ␴22 In some situations, we cannot reasonably assume that the unknown variances  21 and  22 are equal. There is not an exact t-statistic available for testing H0: 1  2   0 in this case. However, an approximate result can be applied. Case 2: Test Statistic for the Difference in Means, Variances Unknown and Not Assumed Equal

If H0: 1  2  0 is true, the statistic T0* 

X1  X2  0

(10-15)

S21 S22 n2 B n1

is distributed approximately as t with degrees of freedom given by

v

s21 s22 2 an n b 1 2

1s22 n2 2 2 1s21 n1 2 2 n1  1 n2  1

(10-16)

If v is not an integer, round down to the nearest integer.

Therefore, if 12  22, the hypotheses on differences in the means of two normal distributions are tested as in the equal variances case, except that T 0* is used as the test statistic and n1 n2  2 is replaced by v in determining the degrees of freedom for the test. EXAMPLE 10-6

Arsenic in Drinking Water

Arsenic concentration in public drinking water supplies is a potential health risk. An article in the Arizona Republic (May 27, 2001) reported drinking water arsenic concentrations in parts per billion (ppb) for 10 metropolitan Phoenix communities and 10 communities in rural Arizona. The data follow: Metro Phoenix 1x1 ⴝ 12.5, s1 ⴝ 7.632

Rural Arizona 1x2 ⴝ 27.5, s2 ⴝ 15.32

Phoenix, 3 Chandler, 7 Gilbert, 25 Glendale, 10 Mesa, 15 Paradise Valley, 6 Peoria, 12 Scottsdale, 25 Tempe, 15 Sun City, 7

Rimrock, 48 Goodyear, 44 New River, 40 Apache Junction, 38 Buckeye, 33 Nogales, 21 Black Canyon City, 20 Sedona, 12 Payson, 1 Casa Grande, 18

We wish to determine if there is any difference in mean arsenic concentrations between metropolitan Phoenix communities and communities in rural Arizona. Figure 10-3 shows a normal probability plot for the two samples of arsenic concentration. The assumption of normality appears quite reasonable, but since the slopes of the two straight lines are very different, it is unlikely that the population variances are the same. Applying the seven-step procedure gives the following: 1.

Parameter of interest: The parameters of interest are the mean arsenic concentrations for the two geographic regions, say, 1 and 2, and we are interested in determining whether 1  2  0. 2. Non hypothesis: H0: 1  2  0, or H0: 1  2 3. Alternative hypothesis: H1: 1  2 4. Test statistic: The test statistic is t*0 

x1  x2  0 s12 s22 n2 B n1

JWCL232_c10_351-400.qxd

366

1/15/10

2:14 PM

Page 366

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

99 95

Percentage

90 80 70 60 50 40 30 20 10 5

Figure 10-3 Normal probability plot of the arsenic concentration data from Example 10-6.

5.

PHX RuralAZ

1 0

10 20 30 40 50 Arsenic concentration in parts per billion

The degrees of freedom on t*0 are found from Equation 10-16 as

v

s22 2 s21 an n b 1 2

1s21 n1 2 2 n1  1



6.

c



1s22 n2 2 2

17.632 2

10 3 17.632 210 4 2 9

Computations: Using the sample data we find t*0 

n2  1

115.32 2

7. 2

d 10  13.2 ⯝ 13 3 115.32 2 10 4 2 9

Therefore, using   0.05 and a fixed-significance-level test, we would reject H0: 1  2 if t*0 t0.025,13  2.160 or if t*0  t0.025,13  2.160.

60

x1  x2 s21 s22 n2 B n1



12.5  27.5  2.77 115.32 2 17.632 2 10 B 10

Conclusions: Because t*0  2.77  t0.025,13  2.160, we reject the null hypothesis.

Practical Interpretation: There is strong evidence to conclude that mean arsenic concentration in the drinking water in rural Arizona is different from the mean arsenic concentration in metropolitan Phoenix drinking water. Furthermore, the mean arsenic concentration is higher in rural Arizona communities. The P-value for this test is approximately P  0.016.

The Minitab output for this example follows: Minitab Computations Two-Sample T-Test and CI: PHX, RuralAZ Two-sample T for PHX vs RuralAZ PHX RuralAZ

N 10 10

Mean 12.50 27.5

StDev 7.63 15.3

Difference  mu PHX  mu RuralAZ Estimate for difference: 15.00 95% CI for difference: (26.71, 3.29) T-Test of difference  0 (vs not  ): T-Value  2.77

SE Mean 2.4 4.9

P-Value  0.016 DF  13

JWCL232_c10_351-400.qxd

1/15/10

2:14 PM

Page 367

10-2 INFERENCE ON THE DIFFERENCE IN MEANS OF TWO NORMAL DISTRIBUTIONS, VARIANCES UNKNOWN

367

The numerical results from Minitab exactly match the calculations from Example 10-6. Note that a two-sided 95% CI on 1  2 is also reported. We will discuss its computation in Section 10-2.3; however, note that the interval does not include zero. Indeed, the upper 95% of confidence limit is 3.29 ppb, well below zero, and the mean observed difference is x1  x2  12.5  27.5  15 ppb.

10-2.2 Type II Error and Choice of Sample Size The operating characteristic curves in Appendix Charts VIIe, VIIf, VIIg, and VIIh are used to evaluate the type II error for the case where 21  22  2. Unfortunately, when  21   22, the distribution of T *0 is unknown if the null hypothesis is false, and no operating characteristic curves are available for this case. For the two-sided alternative H1: 1  2    0, when 21  22  2 and n1  n2  n, Charts VIIe and VIIf are used with ƒ   0 ƒ d (10-17) 2 where  is the true difference in means that is of interest. To use these curves, they must be entered with the sample size n*  2n  1. For the one-sided alternative hypothesis, we use Charts VIIg and VIIh and define d and  as in Equation 10-17. It is noted that the parameter d is a function of , which is unknown. As in the single-sample t-test, we may have to rely on a prior estimate of  or use a subjective estimate. Alternatively, we could define the differences in the mean that we wish to detect relative to . EXAMPLE 10-7

Yield from Catalyst Sample Size

Consider the catalyst experiment in Example 10-5. Suppose that, if catalyst 2 produces a mean yield that differs from the mean yield of catalyst 1 by 4.0%, we would like to reject the null hypothesis with probability at least 0.85. What sample size is required? Using sp  2.70 as a rough estimate of the common standard deviation , we have d  ƒ  ƒ 2  ƒ 4.0 ƒ  3 12212.702 4 

0.74. From Appendix Chart VIIe with d  0.74 and   0.15, we find n*  20, approximately. Therefore, since n*  2n  1, n

20 1 n* 1   10.5 ⯝ 111say2 2 2

and we would use sample sizes of n1  n2  n  11.

Minitab will also perform power and sample size calculations for the two-sample t-test (equal variances). The output from Example 10-7 is as follows: Minitab Computations Power and Sample Size 2-Sample t Test Testing mean 1  mean 2 (versus not ) Calculating power for mean 1  mean 2 difference Alpha  0.05 Sigma  2.7 Difference 4

Sample Size 10

Target Power 0.8500

Actual Power 0.8793

The results agree fairly closely with the results obtained from the O.C. curve.

JWCL232_c10_351-400.qxd

368

1/15/10

2:14 PM

Page 368

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

10-2.3 Confidence Interval on the Difference in Means, Variances Unknown Case 1: ␴21 ⴝ ␴22 ⴝ ␴2 To develop the confidence interval for the difference in means 1  2 when both variances are equal, note that the distribution of the statistic T

X1  X2  11  2 2

(10-18)

1 1 Sp n n 1 2 B

is the t distribution with n1 n2  2 degrees of freedom. Therefore P1t 2,n1 n22  T  t 2,n1 n22 2  1  . Now substituting Equation 10-18 for T and manipulating the quantities inside the probability statement will lead to the 10011  2 % confidence interval on 1  2.

Case 1: Confidence Interval on the Difference in Means, Variances Unknowns and Equal

If x1, x2 , s12, and s22 are the sample means and variances of two random samples of sizes n1 and n2, respectively, from two independent normal populations with unknown but equal variances, then a 100(1 ⴚ ␣)% confidence interval on the difference in means ␮1 ⴚ ␮2 is 1 1 x1  x2  t 2, n1 n22 sp n n 2 B 1 ˛

1 1  1  2  x1  x2 t 2, n1 n22 sp n n 2 B 1 ˛

(10-19)

where sp  23 1n1  12 s21 1n2  12 s22 4  1n1 n2  22 is the pooled estimate of the common population standard deviation, and t 2, n1 n2 2 is the upper 2 percentage point of the t distribution with n1 n2  2 degrees of freedom.

EXAMPLE 10-8

Cement Hydration

An article in the journal Hazardous Waste and Hazardous Materials (Vol. 6, 1989) reported the results of an analysis of the weight of calcium in standard cement and cement doped with lead. Reduced levels of calcium would indicate that the hydration mechanism in the cement is blocked and would allow water to attack various locations in the cement structure. Ten samples of standard cement had an average weight percent calcium of x1  90.0, with a sample standard deviation of s1  5.0, while 15 samples of the lead-doped cement had an average weight percent calcium of x2  87.0, with a sample standard deviation of s2  4.0. We will assume that weight percent calcium is normally distributed and find a 95% confidence interval on the difference

in means, 1  2, for the two types of cement. Furthermore, we will assume that both normal populations have the same standard deviation. The pooled estimate of the common standard deviation is found using Equation 10-12 as follows:

s2p 

1n1  12 s21 1n2  12 s22 n1 n2  2

915.02 1414.02 2 2



10 15  2

 19.52

JWCL232_c10_351-400.qxd

1/15/10

2:14 PM

Page 369

10-2 INFERENCE ON THE DIFFERENCE IN MEANS OF TWO NORMAL DISTRIBUTIONS, VARIANCES UNKNOWN

Therefore, the pooled standard deviation estimate is sp  119.52  4.4. The 95% confidence interval is found using Equation 10-19: x1  x2  t0.025,23 sp

which reduces to 0.72  1  2  6.72

1 1 n  1  2 2 B n1  x1  x2 t0.025,23 sp

Practical Interpretation: Notice that the 95% confidence interval includes zero; therefore, at this level of confidence we cannot conclude that there is a difference in the means. Put another way, there is no evidence that doping the cement with lead affected the mean weight percent of calcium; therefore, we cannot claim that the presence of lead affects this aspect of the hydration mechanism at the 95% level of confidence.

1 1 n n 1 2 B

or upon substituting the sample values and using t0.025,23  2.069, 90.0  87.0  2.06914.42

1 1  1  2 15 B 10

 90.0  87.0 2.06914.42

369

1 1 15 B 10

Case 2: ␴21 ⴝ ␴22 In many situations it is not reasonable to assume that 21  22. When this assumption is unwarranted, we may still find a 100(1  )% confidence interval on 1  2 using the fact that T *  3X1  X2  11  2 2 4  2S21n1 S22n2 is distributed approximately as t with degrees of freedom v given by Equation 10-16. The CI expression follows. ˛

Case 2: Approximate Confidence Interval on the Difference in Means, Variances Unknown Are Not Assumed Equal

If x1, x2, s21, and s 22 are the means and variances of two random samples of sizes n1 and n2, respectively, from two independent normal populations with unknown and unequal variances, an approximate 100(1  )% confidence interval on the difference in means 1  2 is ˛

s21 s22 s21 s22 x1  x2  t 2,  n n  1  2  x1  x2 t 2,  n n 2 2 B 1 B 1 ˛

˛

(10-20)

where v is given by Equation 10-16 and t 2, is the upper  2 percentage point of the t distribution with v degrees of freedom. ˛

EXERCISES FOR SECTION 10-2 10-10.

Consider the computer output below.

Two-Sample T-Test and CI Sample 1 2

N 12 16

Mean 10.94 12.15

StDev 1.26 1.99

SE Mean 0.36 0.50

Difference  mu (1)  mu (2) Estimate for difference: 1.210 95% CI for difference: (2.560, 0.140) T-Test of difference  0 (vs not ) : T-Value  ? P-Value  ? DF  ? Both use Pooled StDev  ?

(a) Fill in the missing values. Is this a one-sided or a two-sided test? Use lower and upper bounds for the P-value. (b) What are your conclusions if   0.05? What if   0.01?

(c) This test was done assuming that the two population variances were equal. Does this seem reasonable? (d) Suppose that the hypothesis had been H0: 1  2 versus H0: 1  2. What would your conclusions be if   0.05? 10-11. Consider the computer output below. Two-Sample T-Test and Cl Sample 1 2

N 15 20

Mean 54.73 58.64

StDev 2.13 5.28

SE Mean 0.55 1.2

Difference  mu (1)  mu (2) Estimate for difference: 3.91 95% upper bound for difference: ? T-Test of difference  0(vs 2. Use   0.025. (b) Use the normal approximation for the Wilcoxon rank-sum test. Assume that   0.05. Find the approximate P-value for this test statistic. 10-32. One of the authors travels regularly to Seattle, Washington. He uses either Delta or Alaska. Flight delays are sometimes unavoidable, but he would be willing to give most of his business to the airline with the best on-time arrival record. The number of minutes that his flight arrived late for the last six trips on each airline follows. Is there evidence that either airline has superior on-time arrival performance? Use   0.01 and the Wilcoxon rank-sum test. Delta: 13, 10, 1, 4, 0, 9 (minutes late) Alaska: 15, 8, 3, 1, 2, 4 (minutes late) 10-33. The manufacturer of a hot tub is interested in testing two different heating elements for his product. The element that produces the maximum heat gain after 15 minutes would be preferable. He obtains 10 samples of each heating unit and tests each one. The heat gain after 15 minutes (in F) follows.

Unit 1: 25, 27, 29, 31, 30, 26, 24, 32, 33, 38 Unit 2: 31, 33, 32, 35, 34, 29, 38, 35, 37, 30 (a) Is there any reason to suspect that one unit is superior to the other? Use   0.05 and the Wilcoxon rank-sum test. (b) Use the normal approximation for the Wilcoxon rank-sum test. Assume that   0.05. What is the approximate P-value for this test statistic? 10-34. Consider the chemical etch rate data in Exercise 10-19. (a) Use the Wilcoxon rank-sum test to investigate the claim that the mean etch rate is the same for both solutions. If   0.05, what are your conclusions? (b) Use the normal approximation for the Wilcoxon rank-sum test. Assume that   0.05. Find the approximate P-value for this test. 10-35. Consider the pipe deflection data in Exercise 10-18. (a) Use the Wilcoxon rank-sum test for the pipe deflection temperature experiment. If   0.05, what are your conclusions? (b) Use the normal approximation for the Wilcoxon rank-sum test. Assume that   0.05. Find the approximate P-value for this test. 10-36. Consider the distance traveled by a golf ball in Exercise 10-29. (a) Use the Wilcoxon rank-sum test to investigate if the means differ. Use   0.05. (b) Use the normal approximation for the Wilcoxon rank-sum test with   0.05. Find the approximate P-value for this test.

JWCL232_c10_351-400.qxd

376

1/21/10

9:33 PM

Page 376

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

10-4 PAIRED t-TEST A special case of the two-sample t-tests of Section 10-2 occurs when the observations on the two populations of interest are collected in pairs. Each pair of observations, say (X1j, X2j), is taken under homogeneous conditions, but these conditions may change from one pair to another. For example, suppose that we are interested in comparing two different types of tips for a hardness-testing machine. This machine presses the tip into a metal specimen with a known force. By measuring the depth of the depression caused by the tip, the hardness of the specimen can be determined. If several specimens were selected at random, half tested with tip 1, half tested with tip 2, and the pooled or independent t-test in Section 10-2 was applied, the results of the test could be erroneous. The metal specimens could have been cut from bar stock that was produced in different heats, or they might not be homogeneous in some other way that might affect hardness. Then the observed difference between mean hardness readings for the two tip types also includes hardness differences between specimens. A more powerful experimental procedure is to collect the data in pairs—that is, to make two hardness readings on each specimen, one with each tip. The test procedure would then consist of analyzing the differences between hardness readings on each specimen. If there is no difference between tips, the mean of the differences should be zero. This test procedure is called the paired t-test. Let (X11, X21), (X12, X22), p , (X1n, X2n) be a set of n paired observations where we assume that the mean and variance of the population represented by X1 are 1 and 21, and the mean and variance of the population represented by X2 are 2 and 22. Define the differences between each pair of observations as Dj  X1j  X2j, j  1, 2, p , n. The Dj’s are assumed to be normally distributed with mean D  E1X1  X2 2  E1X1 2  E1X2 2  1  2 and variance 2D, so testing hypotheses about the difference between 1 and 2 can be accomplished by performing a one-sample t-test on D. Specifically, testing H0: 1  2  0 against H1: 1  2  0 is equivalent to testing H0: D  0 H1: D  0

(10-23)

The test statistic and decision procedure are given below. Paired t-Test

Null hypothesis: H0: D  0 Test statistic:

T0 

D  0 SD 1n

(10-24)

Alternative Hypothesis

P-Value

H1: D  0

Probability above |t0| and probability below |t0| Probability above t0 Probability below t0

H1: D 0 H1: D  0

Rejection Region for Fixed-Level Tests

t0 t 2, n1 or t0  t 2, n1 t0 t , n1 t0  t , n1

JWCL232_c10_351-400.qxd

1/15/10

2:14 PM

Page 377

10-4

PAIRED t-TEST

377

In Equation 10-24, D is the sample average of the n differences D1, D2, p , Dn, and SD is the sample standard deviation of these differences. EXAMPLE 10-10 Shear Strength of Steel Girders An article in the Journal of Strain Analysis (Vol. 18, No. 2 1983) reports a comparison of several methods for predicting the shear strength for steel plate girders. Data for two of these methods, the Karlsruhe and Lehigh procedures, when applied to nine specific girders, are shown in Table 10-3. We wish to determine whether there is any difference (on the average) between the two methods. The seven-step procedure is applied as follows: 1. Parameter of Interest: The parameter of interest is the difference in mean shear strength between the two methods—say, D  1  2  0. 2. Null hypothesis: H0: D  0 3. Alternative hypothesis: H1: D  0

6. Computations: The sample average and standard deviation of the differences dj are d  0.2769 and sd  0.1350, and so the test statistic is t0 

d 0.2769   6.15 sd  1n 0.1350  19

7. Conclusions: Because t0.0005.8  5.041 and the value of the test statistic t0  6.15 exceeds this value, the P-value is less than 2(0.0005)  0.001. Therefore, we conclude that the strength prediction methods yield different results. Practical Interpretation: Specifically, the data indicate that the Karlsruhe method produces, on the average, higher strength predictions than does the Lehigh method. This is a strong conclusion.

4. Test statistic: The test statistic is t0 

5. Reject H0 if: Reject H0 if the P-value is 0.05.

d sd 1n

Minitab can perform the paired t-test. The Minitab output for Example 10-10 is shown below:

Paired T for Karlsruhe–Lehigh Karlsruhe Lehigh Difference

N 9 9 9

Mean 1.34011 1.06322 0.276889

StDev 0.14603 0.05041 0.135027

SE Mean 0.04868 0.01680 0.045009

95% CI for mean difference: (0.173098, 0.380680) T-Test of mean difference  0 (vs not  0): T-Value  6.15, P-Value  0.000

Table 10-3 Strength Predictions for Nine Steel Plate Girders (Predicted Load/Observed Load) Girder

Karlsruhe Method

Lehigh Method

Difference dj

S1兾1 S2兾1 S3兾1 S4兾1 S5兾1 S2兾1 S2兾2 S2兾3 S2兾4

1.186 1.151 1.322 1.339 1.200 1.402 1.365 1.537 1.559

1.061 0.992 1.063 1.062 1.065 1.178 1.037 1.086 1.052

0.125 0.159 0.259 0.277 0.135 0.224 0.328 0.451 0.507

JWCL232_c10_351-400.qxd

378

1/15/10

2:14 PM

Page 378

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

The results essentially agree with the manual calculations. In addition to the hypothesis test results. Minitab reports a two-sided CI on the difference in means. This Cl was found by constructing a single-sample CI on D. We will give the details later. Paired Versus Unpaired Comparisons In performing a comparative experiment, the investigator can sometimes choose between the paired experiment and the two-sample (or unpaired) experiment. If n measurements are to be made on each population, the two-sample t-statistic is T0 

X1  X2  0 Sp

1 1 n Bn

which would be compared to t2n2, and of course, the paired t-statistic is T0 

D  0 SD 1n

which is compared to tn1. Notice that since n D n 1X  X 2 n X n X j 1j 2j 1j 2j   D a n  a a n a n  X1  X2 n j1 j1 j1 j1

the numerators of both statistics are identical. However, the denominator of the two-sample t-test is based on the assumption that X1 and X2 are independent. In many paired experiments, a strong positive correlation  exists between X1 and X2. Then it can be shown that V 1D2  V1X1  X2  0 2  V1X1 2 V1X2 2  2 cov 1X1, X2 2 2 2 11  2  n assuming that both populations X1 and X2 have identical variances  2. Furthermore, SD2 兾n estimates the variance of D. Whenever there is positive correlation within the pairs, the denominator for the paired t-test will be smaller than the denominator of the two-sample t-test. This can cause the two-sample t-test to considerably understate the significance of the data if it is incorrectly applied to paired samples. Although pairing will often lead to a smaller value of the variance of X1  X2 , it does have a disadvantage—namely, the paired t-test leads to a loss of n  1 degrees of freedom in comparison to the two-sample t-test. Generally, we know that increasing the degrees of freedom of a test increases the power against any fixed alternative values of the parameter. So how do we decide to conduct the experiment? Should we pair the observations or not? Although there is no general answer to this question, we can give some guidelines based on the above discussion. 1.

If the experimental units are relatively homogeneous (small ) and the correlation within pairs is small, the gain in precision attributable to pairing will be offset by the loss of degrees of freedom, so an independent-sample experiment should be used.

JWCL232_c10_351-400.qxd

1/15/10

2:14 PM

Page 379

10-4

PAIRED t-TEST

379

If the experimental units are relatively heterogeneous (large ) and there is large positive correlation within pairs, the paired experiment should be used. Typically, this case occurs when the experimental units are the same for both treatments; as in Example 10-10, the same girders were used to test the two methods.

2.

Implementing the rules still requires judgment, because  and  are never known precisely. Furthermore, if the number of degrees of freedom is large (say, 40 or 50), the loss of n  1 of them for pairing may not be serious. However, if the number of degrees of freedom is small (say, 10 or 20), losing half of them is potentially serious if not compensated for by increased precision from pairing. Confidence Interval for ␮D To construct the confidence interval for D  1  2, note that T

D  D SD 1n

follows a t distribution with n  1 degrees of freedom. Then, since P(t兾2,n1  T  t兾2,n1)  1  , we can substitute for T in the above expression and perform the necessary steps to isolate D  1  2 between the inequalities. This leads to the following 100(1  )% confidence interval on 1  2.

Confidence Interval for ␮D from Paired Samples



If d and sD are the sample mean and standard deviation of the difference of n random pairs of normally distributed measurements, a 100(1 ⴚ ␣)% confidence interval on the difference in means ␮D ⴝ ␮1 ⴚ ␮2 is d  t2, n1 sD 1n  D  d t2, n1 sD 1n

(10-25)

where t/2,n1 is the upper 兾2% point of the t distribution with n  1 degrees of freedom.

This confidence interval is also valid for the case where 12  22, because sD2 estimates D2  V(X1  X2). Also, for large samples (say, n  30 pairs), the explicit assumption of normality is unnecessary because of the central limit theorem.

EXAMPLE 10-11 Parallel Park Cars The journal Human Factors (1962, pp. 375–380) reported a study in which n  14 subjects were asked to parallel park two cars having very different wheel bases and turning radii. The time in seconds for each subject was recorded and is given in Table 10-4. From the column of observed differences we calculate d  1.21 and sD  12.68. The 90% confidence interval for D  1  2 is found from Equation 10-25 as follows:

d  t0.05,13 sD 1n  D  d t0.05,13 sD 1n 1.21  1.771112.682  114  D  1.21 1.771112.682  114 4.79  D  7.21 Notice that the confidence interval on D includes zero. This implies that, at the 90% level of confidence, the data do not support the claim that the two cars have different mean parking times 1 and 2. That is, the value D  1  2  0 is not inconsistent with the observed data.

JWCL232_c10_351-400.qxd

380

1/15/10

2:14 PM

Page 380

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

Table 10-4 Time in Seconds to Parallel Park Two Automobiles Automobile

Subject

11x1j2

21x2j2

1 2 3 4 5 6 7 8 9 10 11 12 13 14

37.0 25.8 16.2 24.2 22.0 33.4 23.8 58.2 33.6 24.4 23.4 21.2 36.2 29.8

17.8 20.2 16.8 41.4 21.4 38.4 16.8 32.2 27.8 23.2 29.6 20.6 32.2 53.8

Difference 1dj2

19.2 5.6 0.6 17.2 0.6 5.0 7.0 26.0 5.8 1.2 6.2 0.6 4.0 24.0

Nonparametric Approach to Paired Comparisons Both the sign test and the Wilcoxon signed-rank test discussed in Section 9-9 can be applied to paired observations. In the case of the sign test, the null hypothesis is that the median of the ~  0). The Wilcoxon signed-rank test is for the null differences is equal to zero (that is, H0:  D hypothesis that the mean of the differences is equal to zero. The procedures are applied to the observed differences as described in Sections 9-9.1 and 9-9.2. EXERCISES FOR SECTION 10-4 10-37. Consider the shear strength experiment described in Example 10-10. (a) Construct a 95% confidence interval on the difference in mean shear strength for the two methods. Is the result you obtained consistent with the findings in Example 10-10? Explain why. (b) Do each of the individual shear strengths have to be normally distributed for the paired t-test to be appropriate, or is it only the difference in shear strengths that must be normal? Use a normal probability plot to investigate the normality assumption. 10-38. Consider the parking data in Example 10-11. (a) Use the paired t-test to investigate the claim that the two types of cars have different levels of difficulty to parallel park. Use   0.10. (b) Compare your results with the confidence interval constructed in Example 10-11 and comment on why they are the same or different. (c) Investigate the assumption that the differences in parking times are normally distributed. 10-39. The manager of a fleet of automobiles is testing two brands of radial tires. He assigns one tire of each brand

at random to the two rear wheels of eight cars and runs the cars until the tires wear out. The data (in kilometers) follow. Find a 99% confidence interval on the difference in mean life. Which brand would you prefer, based on this calculation? Car

Brand 1

Brand 2

1 2 3 4 5 6 7 8

36,925 45,300 36,240 32,100 37,210 48,360 38,200 33,500

34,318 42,280 35,500 31,950 38,015 47,800 37,810 33,215

10-40. A computer scientist is investigating the usefulness of two different design languages in improving programming tasks. Twelve expert programmers, familiar with both languages, are asked to code a standard function in both languages, and the time (in minutes) is recorded. The data follow:

JWCL232_c10_351-400.qxd

1/15/10

2:14 PM

Page 381

10-4

Time

Programmer

Design Language 1

Design Language 2

1 2 3 4 5 6 7 8 9 10 11 12

17 16 21 14 18 24 16 14 21 23 13 18

18 14 19 11 23 21 10 13 19 24 15 20

(a) Is the assumption that the difference in coding time is normally distributed reasonable? (b) Find a 95% confidence interval on the difference in mean coding times. Is there any indication that one design language is preferable? 10-41. Fifteen adult males between the ages of 35 and 50 participated in a study to evaluate the effect of diet and exercise on blood cholesterol levels. The total cholesterol was measured in each subject initially and then three months after participating in an aerobic exercise program and switching to a low-fat diet. The data are shown in the accompanying table Blood Cholesterol Level Subject 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Before

After

265 240 258 295 251 245 287 314 260 279 283 240 238 225 247

229 231 227 240 238 241 234 256 247 239 246 218 219 226 233

PAIRED t-TEST

381

(a) Do the data support the claim that low-fat diet and aerobic exercise are of value in producing a mean reduction in blood cholesterol levels? Use   0.05. Find the P-value. (b) Calculate a one-sided confidence limit that can be used to answer the question in part (a). 10-42. An article in the Journal of Aircraft (Vol. 23, 1986, pp. 859–864) described a new equivalent plate analysis method formulation that is capable of modeling aircraft structures such as cranked wing boxes, and that produces results similar to the more computationally intensive finite element analysis method. Natural vibration frequencies for the cranked wing box structure are calculated using both methods, and results for the first seven natural frequencies follow:

Freq. 1 2 3 4 5 6 7

Finite Element Cycle/s

Equivalent Plate, Cycle/s

14.58 48.52 97.22 113.99 174.73 212.72 277.38

14.76 49.10 99.99 117.53 181.22 220.14 294.80

(a) Do the data suggest that the two methods provide the same mean value for natural vibration frequency? Use   0.05. Find the P-value. (b) Find a 95% confidence interval on the mean difference between the two methods. 10-43. Ten individuals have participated in a diet-modification program to stimulate weight loss. Their weight both before and after participation in the program is shown in the following list. Subject 1 2 3 4 5 6 7 8 9 10

Before

After

195 213 247 201 187 210 215 246 294 310

187 195 221 190 175 197 199 221 278 285

(a) Is there evidence to support the claim that this particular diet-modification program is effective in producing a mean weight reduction? Use   0.05. (b) Is there evidence to support the claim that this particular diet-modification program will result in a mean weight loss of at least 10 pounds? Use   0.05.

JWCL232_c10_351-400.qxd

382

1/15/10

2:14 PM

Page 382

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

(c) Suppose that, if the diet-modification program results in mean weight loss of at least 10 pounds, it is important to detect this with probability of at least 0.90. Was the use of 10 subjects an adequate sample size? If not, how many subjects should have been used? 10-44. Two different analytical tests can be used to determine the impurity level in steel alloys. Eight specimens are tested using both procedures, and the results are shown in the following tabulation. Specimen

Test 1

Test 2

1 2 3 4 5 6 7 8

1.2 1.3 1.5 1.4 1.7 1.8 1.4 1.3

1.4 1.7 1.5 1.3 2.0 2.1 1.7 1.6

(a) Is there sufficient evidence to conclude that tests differ in the mean impurity level, using   0.01? (b) Is there evidence to support the claim that Test 1 generates a mean difference 0.1 units lower than Test 2? Use   0.05. (c) If the mean from Test 1 is 0.1 less than the mean from Test 2, it is important to detect this with probability at least 0.90. Was the use of eight alloys an adequate sample size? If not, how many alloys should have been used? 10-45. An article in Neurology (1998, Vol. 50, pp. 1246–1252) discussed that monozygotic twins share numerous physical, psychological, and pathological traits. The investigators measured an intelligence score of 10 pairs of twins, and the data are as follows: Pair

Birth Order: 1

Birth Order: 2

1 2 3 4 5 6 7 8 9 10

6.08 6.22 7.99 7.44 6.48 7.99 6.32 7.60 6.03 7.52

5.73 5.80 8.42 6.84 6.43 8.76 6.32 7.62 6.59 7.67

(a) Is the assumption that the difference in score is normally distributed reasonable? Show results to support your answer. (b) Find a 95% confidence interval on the difference in mean score. Is there any evidence that mean score depends on birth order? (c) It is important to detect a mean difference in score of one point, with a probability of at least 0.90. Was the use of 10 pairs an adequate sample size? If not, how many pairs should have been used? 10-46. In Biometrics (1990, Vol. 46, pp. 673–87), the authors analyzed the circumference of five orange trees (labeled as A–E) measured on seven occasions (xi). Tree

x1

x2

x3

x4

x5

x6

x7

A B C D E

30 33 30 32 30

58 69 51 62 49

87 111 75 112 81

115 156 108 167 125

120 172 115 179 142

142 203 139 209 174

145 203 140 214 177

(a) Compare the mean increase in circumference in periods 1 to 2 to the mean increase in periods 2 to 3. The increase is the difference in circumference in the two periods. Are these means significantly different at   0.10? (b) Is there evidence that the mean increase in period 1 to period 2 is greater than the mean increase in period 6 to period 7 at   0.05? (c) Are the assumptions of the test in part (a) violated because the same data (period 2 circumference) is used to calculate both mean increases? 10-47. Use the sign test on the blood cholesterol data in Exercise 10-41. Is there evidence that diet and exercise reduce the median cholesterol level? 10-48. Repeat Exercise 10-47 using the Wilcoxon signedrank test. State carefully what hypothesis is being tested and how it differs from the one tested in Exercise 10-47.

10-5 INFERENCE ON THE VARIANCES OF TWO NORMAL DISTRIBUTIONS We now introduce tests and confidence intervals for the two population variances shown in Fig. 10-1. We will assume that both populations are normal. Both the hypothesis-testing and confidence interval procedures are relatively sensitive to the normality assumption.

JWCL232_c10_351-400.qxd

1/15/10

2:14 PM

Page 383

383

10-5 INFERENCE ON THE VARIANCES OF TWO NORMAL DISTRIBUTIONS

10-5.1 F Distribution Suppose that two independent normal populations are of interest, where the population means and variances, say, 1, 21, 2, and 22, are unknown. We wish to test hypotheses about the equality of the two variances, say, H0: 21  22. Assume that two random samples of size n1 from population 1 and of size n2 from population 2 are available, and let S 21 and S 22 be the sample variances. We wish to test the hypotheses H0: 21  22 H1: 21  22

(10-26)

The development of a test procedure for these hypotheses requires a new probability distribution, the F distribution. The random variable F is defined to be the ratio of two independent chi-square random variables, each divided by its number of degrees of freedom. That is, F

Wu Yv

(10-27)

where W and Y are independent chi-square random variables with u and v degrees of freedom, respectively. We now formally state the sampling distribution of F. F Distribution

Let W and Y be independent chi-square random variables with u and v degrees of freedom, respectively. Then the ratio F

Wu Yv

(10-28)

has the probability density function u v u u2 1u221 bavb x 2 f 1x2  1u v22 , u u v  a b  a b c avb x 1d 2 2 a

0x

(10-29)

and is said to follow the F distribution with u degrees of freedom in the numerator and v degrees of freedom in the denominator. It is usually abbreviated as Fu,v.

The mean and variance of the F distribution are   v兾(v  2) for v 2, and 2 

2v 2 1u v  22 , u1v  22 2 1v  42

v 4

Two F distributions are shown in Fig. 10-4. The F random variable is nonnegative, and the distribution is skewed to the right. The F distribution looks very similar to the chi-square distribution; however, the two parameters u and v provide extra flexibility regarding shape.

JWCL232_c10_351-400.qxd

384

1/15/10

2:14 PM

Page 384

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES f(x) u = 5, v = 5

f (x)

u = 5, v = 15

α

α

0

2

4

6

8

10

f1 – α , u, v

x

Figure 10-4 Probability density functions of two F distributions.

f α , u, v

x

Figure 10-5 Upper and lower percentage points of the F distribution.

The percentage points of the F distribution are given in Table VI of the Appendix. Let f,u,v be the percentage point of the F distribution, with numerator degrees of freedom u and denominator degrees of freedom v such that the probability that the random variable F exceeds this value is P1F f, u, v 2 



冮 f 1x2 dx   f,u,v

This is illustrated in Fig. 10-5. For example, if u  5 and v  10, we find from Table V of the Appendix that P1F f0.05,5,10 2  P1F5,10 3.332  0.05 That is, the upper 5 percentage point of F5,10 is f0.05,5,10  3.33. Table V contains only upper-tail percentage points (for selected values of f,u,v for   0.25) of the F distribution. The lower-tail percentage points f1,u,v can be found as follows.

f1,u,v 

1 f,v,u

(10-30)

For example, to find the lower-tail percentage point f0.95,5,10, note that f0.95, 5,10 

1 1  0.211  4.74 f0.05,10, 5

10-5.2 Hypothesis Tests on the Ratio of Two Variances A hypothesis-testing procedure for the equality of two variances is based on the following result.

JWCL232_c10_351-400.qxd

1/15/10

2:14 PM

Page 385

10-5 INFERENCE ON THE VARIANCES OF TWO NORMAL DISTRIBUTIONS

Distribution of the Ratio of Sample Variances from Two Normal Distributions

385

Let X11, X12, p , X1n1 be a random sample from a normal population with mean 1 and variance 21, and let X21, X22, p , X2n2 be a random sample from a second normal population with mean 2 and variance 22. Assume that both normal populations are independent. Let S 12 and S 22 be the sample variances. Then the ratio F

S 12 21 S 22 22

has an F distribution with n1  1 numerator degrees of freedom and n2  1 denominator degrees of freedom.

This result is based on the fact that (n1  1)S 21/21 is a chi-square random variable with n1  1 degrees of freedom, that (n2  1)S 22 兾22 is a chi-square random variable with n2  1 degrees of freedom, and that the two normal populations are independent. Clearly under the null hypothesis H0: 21  22 the ratio F0  S 21 S 22 has an Fn11,n21 distribution. This is the basis of the following test procedure. Tests on the Ratio of Variances from Two Normal Distributions

Null hypothesis:

H0: 21  22

Test statistic:

F0 

S12 S 22

(10-31)

Alternative Hypotheses

Rejection Criterion

H1: 21  22 H1: 21 22 H1: 21  22

f0 f2,n11,n21 or f0  f12,n11,n21 f0 f,n11,n21 f0  f1, n11,n21

The critical regions for these fixed-significance-level tests are shown in Figure 10-6. f (x)

f (x)

f (x)

 2n – 1

 2n – 1

α /2

0

 2n – 1

α /2

 21 – α /2, n – 1

 2α /2, n – 1

x

 2α , n – 1

0

(a)

Figure 10-6 The F distribution for the test of and (c) H1: 12  22.

α

α x

0

(b)

H0: 12



22

with critical region values for (a)

x

 21 – α , n – 1 (c)

H1: 12



22,

(b) H1: 12 22,

EXAMPLE 10-12 Semiconductor Etch Variability Oxide layers on semiconductor wafers are etched in a mixture of gases to achieve the proper thickness. The variability in the thickness of these oxide layers is a critical characteristic of the wafer, and low variability is desirable for subsequent processing steps. Two different mixtures of gases are being studied to determine whether one is superior in reducing the variability

of the oxide thickness. Sixteen wafers are etched in each gas. The sample standard deviations of oxide thickness are s1  1.96 angstroms and s2  2.13 angstroms, respectively. Is there any evidence to indicate that either gas is preferable? Use a fixed-level test with   0.05.

JWCL232_c10_351-400.qxd

386

1/15/10

2:14 PM

Page 386

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

The seven-step hypothesis-testing procedure may be applied to this problem as follows: 1. Parameter of interest: The parameter of interest are the variances of oxide thickness 21 and 22. We will assume that oxide thickness is a normal random variable for both gas mixtures.

f0  f0.975,15,15  1f0.025,15,15  12.86  0.35. Refer to Figure 10-6(a). 7. Computations: Because s21  (1.96)2  3.84 and s22  (2.13)2  4.54, the test statistic is f0 

2. Null hypothesis: H0: 21  22 3. Alternative hypothesis: H1: 21  22 4. Test statistic: The test statistic is given by equation 10-31: f0 

s21 s22

6. Reject H0 if : Because n1  n2  16 and   0.05, we will reject H0: 21  22 if f0 f0.025,15,15  2.86 or if

s21 3.84   0.85 4.54 s22

8. Conclusions: Because f0.975,15,15  0.35  0.85  f0.025,15,15  2.86, we cannot reject the null hypothesis H0: 21  22 at the 0.05 level of significance. Practical Interpretation: There is no strong evidence to indicate that either gas results in a smaller variance of oxide thickness.

P-Values for the F-Test The P-value approach can also be used with F-tests. To show how to do this, consider the upper-tailed one-tailed test. The P-value is the area (probability) under the F distribution with n1  1 and n2  1 degrees of freedom that lies beyond the computed value of the test statistic f0. Appendix A Table IV can be used to obtain upper and lower bounds on the P-value. For example, consider an F-test with 9 numerator and 14 denominator degrees of freedom for which f0  3.05. From Appendix A Table IV we find that f0.05,9,14  2.65 and f0.025,9,14  3.21, so because f0 = 3.05 lies between these two values, the P-value is between 0.05 and 0.025; that is, 0.025  P  0.05. The P-value for a lower-tailed test would be found similarly, although since Appendix A Table IV contains only upper-tail points of the F distribution, equation 10-30 would have to be used to find the necessary lower-tail points. For a two-tailed test, the bounds obtained from a one-tail test would be doubled to obtain the P-value. To illustrate calculating bounds on the P-value for a two-tailed F-test, reconsider Finding the P-Value for Example 10-12. The computed value of the test statistic in this example is f0  0.85. This Example 10-12 value falls in the lower tail of the F15,15 distribution. The lower-tail point that has 0.25 probability to the left of it is f0.75,15,15  1/ f0.25,15,15  1/1.43  0.70 and since 0.70  0.85, the probability that lies to the left of 0.85 exceeds 0.25. Therefore, we would conclude that the P-value for f0  0.85 is greater than 2(0.25)  0.5, so there is insufficient evidence to reject the null hypothesis. This is consistent with the original conclusions from Example 10-12. The actual P-value is 0.7570. This value was obtained from a calculator from which we found that P(F15,15  0.85)  0.3785 and 2(0.3785)  0.7570. Minitab can also be used to calculate the required probabilities. Minitab will perform the F-test on the equality of two variances of independent normal distributions. The Minitab output is shown below. Test for Equal Variances 95% Bonferroni confidence intervals for standard deviations Sample 1 2

N 16 16

Lower 1.38928 1.51061

StDev 1.95959 2.13073

F-Test (Normal Distribution) Test statistic  0.85, P-value  0.750

Upper 3.24891 3.53265

JWCL232_c10_351-400.qxd

1/15/10

2:14 PM

Page 387

10-5 INFERENCE ON THE VARIANCES OF TWO NORMAL DISTRIBUTIONS

387

Minitab also gives confidence intervals on the individual variances. These are the confidence intervals originally given in Equation 8-19, except that a Bonferroni “adjustment” has been applied to make the confidence level for both intervals simultaneously at least equal to 95%. This consists of using /2  0.05/2  0.025 to construct the individual intervals. That is, each individual confidence interval is a 97.5% CI. In Section 10-5.4, we will show how to construct a CI on the ratio of the two variances.

10-5.3 Type II Error and Choice of Sample Size Appendix Charts VIIo, VIIp, VIIq, and VIIr provide operating characteristic curves for the F-test given in Section 10-5.1 for   0.05 and   0.01, assuming that n1  n2  n. Charts VIIo and VIIp are used with the two-sided alternate hypothesis. They plot  against the abscissa parameter 1 ␭ (10-32) 2 for various n1  n2  n. Charts VIIq and VIIr are used for the one-sided alternative hypotheses. EXAMPLE 10-13 Semiconductor Etch Variability Sample Size For the semiconductor wafer oxide etching problem in Example 10-12, suppose that one gas resulted in a standard deviation of oxide thickness that is half the standard deviation of oxide thickness of the other gas. If we wish to detect such a situation with probability at least 0.80, is the sample size n1  n2  20 adequate? Note that if one standard deviation is half the other,

1 ␭ 2 2

By referring to Appendix Chart VIIo with n1  n2  n  20 and   2, we find that  ⯝ 0.20. Therefore, if   0.20, the power of the test (which is the probability that the difference in standard deviations will be detected by the test) is 0.80, and we conclude that the sample sizes n1  n2  20 are adequate.

10-5.4 Confidence Interval on the Ratio of Two Variances To find the confidence interval on 21 22, recall that the sampling distribution of F

S 22 22 S 21 21

is an F with n2  1 and n1  1 degrees of freedom. Therefore, P1 f12, n21, n11  F  f2,n21, n11 2  1  . Substitution for F and manipulation of the inequalities will lead to the 10011  2 % confidence interval for 21 22. Confidence Interval on the Ratio of Variances from Two Normal Distributions

If s21 and s22 are the sample variances of random samples of sizes n1 and n2, respectively, from two independent normal populations with unknown variances  21 and 22, then a 100(1 ⴚ ␣)% confidence interval on the ratio ␴21 ␴22 is s21 21 s 21 f   f s22 12,n21,n11 22 s 22 2,n21,n11

(10-33)

where f2,n21,n11 and f12,n21,n11 are the upper and lower 兾2 percentage points of the F distribution with n2  1 numerator and n1  1 denominator degrees of freedom, respectively. A confidence interval on the ratio of the standard deviations can be obtained by taking square roots in Equation 10-33.

JWCL232_c10_351-400.qxd

388

1/15/10

2:14 PM

Page 388

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

EXAMPLE 10-14 Surface Finish for Titanium Alloy A company manufactures impellers for use in jet-turbine engines. One of the operations involves grinding a particular surface finish on a titanium alloy component. Two different grinding processes can be used, and both processes can produce parts at identical mean surface roughness. The manufacturing engineer would like to select the process having the least variability in surface roughness. A random sample of n1  11 parts from the first process results in a sample standard deviation s1  5.1 microinches, and a random sample of n2  16 parts from the second process results in a sample standard deviation of s2  4.7 microinches. We will find a 90% confidence interval on the ratio of the two standard deviations, 1  2. Assuming that the two processes are independent and that surface roughness is normally distributed, we can use Equation 10-33 as follows: s21 21 s 21 2 f0.95,15,10  2  2 f0.05,15,10 s2 2 s2

15.12 2

14.72

2

0.39 

15.12 2 21 2.85 2  2 14.72 2

or upon completing the implied calculations and taking square roots, 1 0.678    1.832 2 Notice that we have used Equation 10-30 to find f0.95,15,10  1兾f0.05,10,15  1兾2.54  0.39. Practical Interpretation: Since this confidence interval includes unity, we cannot claim that the standard deviations of surface roughness for the two processes are different at the 90% level of confidence.

EXERCISES FOR SECTION 10-5 10-49. For an F distribution, find the following: (a) f0.25,5,10 (b) f0.10,24,9 (c) f0.05,8,15 (d) f0.75,5,10 (e) f0.90,24,9 (f ) f0.95,8,15 10-50. For an F distribution, find the following: (a) f0.25,7,15 (b) f0.10,10,12 (c) f0.01,20,10 (d) f0.75,7,15 (e) f0.90,10,12 (f ) f0.99,20,10 10-51. Consider the hypothesis test H0 : 21  22 against H1 : 21  22. Suppose that the sample sizes are n1  5 and n2  10, and that s21  23.2 and s22  28.8. Use   0.05. Test the hypothesis and explain how the test could be conducted with a confidence interval on 1兾2. 10-52. Consider the hypothesis test H0 : 21  22 against H0 : 21 22. Suppose that the sample sizes are n1  20 and n2  8, and that s21  4.5 and s22  2.3. Use   0.01. Test the hypothesis and explain how the test could be conducted with a confidence interval on 1兾2. 10-53. Consider the hypothesis test H0 : 21  22 against H1 : 21  22. Suppose that the sample sizes are n1  15 and n2  15, and the sample variances are s21  2.3 and s22  1.9. Use   0.05. (a) Test the hypothesis and explain how the test could be conducted with a confidence interval on 1兾2. (b) What is the power of the test in part (a) if 1 is twice as large as 2? (c) Assuming equal sample sizes, what sample size should be used to obtain   0.05 if the 2 is half of 1? 10-54. Two chemical companies can supply a raw material. The concentration of a particular element in this material is

important. The mean concentration for both suppliers is the same, but we suspect that the variability in concentration may differ between the two companies. The standard deviation of concentration in a random sample of n1  10 batches produced by company 1 is s1  4.7 grams per liter, while for company 2, a random sample of n2  16 batches yields s2  5.8 grams per liter. Is there sufficient evidence to conclude that the two population variances differ? Use   0.05. 10-55. A study was performed to determine whether men and women differ in their repeatability in assembling components on printed circuit boards. Random samples of 25 men and 21 women were selected, and each subject assembled the units. The two sample standard deviations of assembly time were smen  0.98 minutes and swomen  1.02 minutes. (a) Is there evidence to support the claim that men and women differ in repeatability for this assembly task? Use   0.02 and state any necessary assumptions about the underlying distribution of the data. (b) Find a 98% confidence interval on the ratio of the two variances. Provide an interpretation of the interval. 10-56. Consider the foam data in Exercise 10-16. Construct the following: (a) A 90% two-sided confidence interval on 21兾22. (b) A 95% two-sided confidence interval on 21兾22. Comment on the comparison of the width of this interval with the width of the interval in part (a). (c) A 90% lower-confidence bound on 1兾2. 10-57. Consider the diameter data in Exercise 10-15. Construct the following: (a) A 90% two-sided confidence interval on 1兾2.

JWCL232_c10_351-400.qxd

1/15/10

2:14 PM

Page 389

10-6 INFERENCE ON TWO POPULATION PROPORTIONS

(b) A 95% two-sided confidence interval on 12. Comment on the comparison of the width of this interval with the width of the interval in part (a). (c) A 90% lower-confidence bound on 1兾2. 10-58. Consider the gear impact strength data in Exercise 10-20. Is there sufficient evidence to conclude that the variance of impact strength is different for the two suppliers? Use   0.05. 10-59. Consider the melting-point data in Exercise 10-21. Do the sample data support a claim that both alloys have the same variance of melting point? Use   0.05 in reaching your conclusion. 10-60. Exercise 10-24 presented measurements of plastic coating thickness at two different application temperatures. Test H0: 21  22 against H1: 21  22 using   0.01. 10-61. Reconsider the overall distance data for golf balls in Exercise 10-29. Is there evidence to support the claim that the standard deviation of overall distance is the same for both brands of balls (use   0.05)? Explain how this question can be answered with a 95% confidence interval on 12 .

389

10-62. Reconsider the coefficient of restitution data in Exercise 10-30. Do the data suggest that the standard deviation is the same for both brands of drivers (use   0.05)? Explain how to answer this question with a confidence interval on 12 . 10-63. Consider the weight of paper data from Technometrics in Exercise 10-28. Is there evidence that the variance of the weight measurement differs between the sheets of paper? Use   0.05. Explain how this test can be conducted with a confidence interval. 10-64. Consider the film speed data in Exercise 10-22. (a) Test H0: 21  22 versus H1: 21  22 using   0.02. (b) Suppose that one population standard deviation is 50% larger than the other. Is the sample size n1  n2  8 adequate to detect this difference with high probability? Use   0.01 in answering this question. 10-65. Consider the etch rate data in Exercise 10-19. (a) Test the hypothesis H0: 21  22 against H1: 21  22 using   0.05, and draw conclusions. (b) Suppose that if one population variance is twice as large as the other, we want to detect this with probability at least 0.90 (using   0.05). Are the sample sizes n1  n2  10 adequate?

10-6 INFERENCE ON TWO POPULATION PROPORTIONS We now consider the case where there are two binomial parameters of interest, say, p1 and p2, and we wish to draw inferences about these proportions. We will present large-sample hypothesis testing and confidence interval procedures based on the normal approximation to the binomial.

10-6.1 Large-Sample Tests on the Difference in Population Proportions Suppose that two independent random samples of sizes n1 and n2 are taken from two populations, and let X1 and X2 represent the number of observations that belong to the class of interest in samples 1 and 2, respectively. Furthermore, suppose that the normal approximation to the binomial is applied to each population, so the estimators of the population proportions P1  X1 n1 and P2  X2 n2 have approximate normal distributions. We are interested in testing the hypotheses H0: p1  p2 H1: p1  p2 The statistic Test Statistic for the Difference of Two Population Proportions

Z

Pˆ 1  Pˆ 2  1 p1  p2 2

p1 11  p1 2 p2 11  p2 2 n1 n2 B

(10-34)

JWCL232_c10_351-400.qxd

390

1/21/10

8:32 PM

Page 390

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

is distributed approximately as standard normal and is the basis of a test for H0: p1 ⫽ p2. Specifically, if the null hypothesis H0: p1 ⫽ p2 is true, using the fact that p1 ⫽ p2 ⫽ p, the random variable Pˆ 1 ⫺ Pˆ 2

Z⫽

1 1 p11 ⫺ p2 a n ⫹ n b 1 2 B

is distributed approximately N(0, 1). A pooled estimator of the common parameter p is Pˆ ⫽

X1 ⫹ X2 n1 ⫹ n2

The test statistic for H0: p1 ⫽ p2 is then Pˆ 1 ⫺ Pˆ 2

Z0 ⫽ B

1 1 Pˆ 11 ⫺ Pˆ 2 a n ⫹ n b 1 2

This leads to the test procedures described below. Approximate Tests on the Difference of Two Population Proportions

Null hypothesis:

H0: p1 ⫽ p2

Test statistic:

Z0 ⫽

Pˆ 1 ⫺ Pˆ 2

(10-35)

1 1 Pˆ 11 ⫺ Pˆ 2 a n ⫹ n b 1 2 B

Alternative Hypothesis H1: p1 ⫽ p2

P-Value Probability above | z0| and probability below ⫺| z0 |. P ⫽ 2 31 ⫺ ⌽1 0 z0 0 24

Rejection Criterion for Fixed-Level Tests z0 ⬎ z␣Ⲑ2 or z0 ⬍ ⫺z␣Ⲑ2

H1: p1 ⬎ p2

Probability above z0.

z0 ⬎ z␣

H1: p1 ⬍ p2

Probability below z0. P ⫽ ⌽1z0 2

z0 ⬍ ⫺z␣

P ⫽ 1 ⫺ ⌽1z0 2

EXAMPLE 10-15 St. John’s Wort Extracts of St. John’s Wort are widely used to treat depression. An article in the April 18, 2001, issue of the Journal of the American Medical Association (“Effectiveness of St. John’s Wort on Major Depression: A Randomized Controlled Trial”) compared the efficacy of a standard extract of St. John’s Wort with a placebo in 200 outpatients diagnosed with major depression. Patients were randomly assigned to two groups; one group received the St. John’s Wort, and the other received the placebo. After eight weeks, 19 of the placebo-treated patients showed improvement, whereas 27 of those treated with St. John’s Wort

improved. Is there any reason to believe that St. John’s Wort is effective in treating major depression? Use ␣ ⫽ 0.05. The seven-step hypothesis testing procedure leads to the following results:

1.

Parameter of interest: The parameters of interest are p1 and p2, the proportion of patients who improve following treatment with St. John’s Wort ( p1) or the placebo ( p2).

2. 3.

Null hypothesis: H0: p1 ⫽ p2 Alternative hypothesis: H1: p1 ⫽ p2

JWCL232_c10_351-400.qxd

1/15/10

2:14 PM

Page 391

10-6 INFERENCE ON TWO POPULATION PROPORTIONS

4.

6.

Test statistic: The test statistic is z0 

pˆ 1  pˆ 2

where pˆ 1  27100  0.27, pˆ 2  19100  0.19, n1  n2  100, and pˆ 

5.

Computations: The value of the test statistic is z0 

1 1 pˆ 11  pˆ 2 a n n b 1 2 B

7.

x1 x2 19 27   0.23 n1 n2 100 100

391

0.27  0.19 1 1 0.2310.772 a b 100 100 B

 1.34

Conclusions: Since z0  1.34, the P-value is P  23 1  11.342 4  0.18, so, we cannot reject the null hypothesis.

Practical Interpretation: There is insufficient evidence to support the claim that St. John’s Wort is effective in treating major depression.

Reject H0 if: Reject H0: p1  p2 if the P-value is less than 0.05.

The following box shows the Minitab two-sample hypothesis test and CI procedure for proportions. Notice that the 95% CI on p1  p2 includes zero. The equation for constructing the CI will be given in Section 10-6.3. Minitab Computations Test and CI for Two Proportions Sample X N 1 27 100 2 19 100 Estimate for p(1)  p(2): 0.08 95% CI for p(1)  p(2): (0.0361186, 0.196119) Test for p(1)  p(2)  0 (vs not  0): Z  1.35 P-Value

Sample p 0.270000 0.190000

 0.177

10-6.2 Type II Error and Choice of Sample Size The computation of the -error for the large-sample test of H0: p1  p2 is somewhat more involved than in the single-sample case. The problem is that the denominator of the test statistic Z0 is an estimate of the standard deviation of Pˆ 1  Pˆ 2 under the assumption that p1  p2  p. When H0: p1  p2 is false, the standard deviation of Pˆ 1  Pˆ 2 is Pˆ1Pˆ 2  Approximate Type II Error for a Two-Sided Test on the Difference of Two Population Proportions

B

p1 11  p1 2 p2 11  p2 2 n1 n2

(10-36)

If the alternative hypothesis is two sided, the -error is c

z  2 2pq11 n1 1n2 2  1 p1  p2 2 d Pˆ1Pˆ2

c

z  2 2pq 11 n1 1n2 2  1 p1  p2 2 d Pˆ1Pˆ2

(10-37)

JWCL232_c10_351-400.qxd

392

1/15/10

2:14 PM

Page 392

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

where p

n1 p1 n2 p2 n1 n2

and

q

n1 11  p1 2 n2 11  p2 2 n1 n2

and Pˆ1 Pˆ 2 is given by Equation 10-36. Approximate Type II Error for a One-Sided Test on the Difference of Two Population Proportions

If the alternative hypothesis is H1: p1 p2, c

z 2pq 11n1 1n2 2  1 p1  p2 2 d Pˆ1Pˆ2

(10-38)

and if the alternative hypothesis is H1: p1  p2, 1c

z 2pq 11n1 1n2 2  1 p1  p2 2 d Pˆ1Pˆ2

(10-39)

For a specified pair of values p1 and p2, we can find the sample sizes n1  n2  n required to give the test of size  that has specified type II error . Approximate Sample Size for a Two-Sided Test on the Difference in Population Proportions

For the two-sided alternative, the common sample size is n

3z  2 11 p1 p2 21q1 q2 2 2 z 1p1q1 p 2q2 4 2 1 p1  p2 2 2

(10-40)

where q1  1  p1 and q2  1  p2.

For a one-sided alternative, replace z  2 in Equation 10-40 by z.

10-6.3 Confidence Interval on the Difference in Population Proportions The confidence interval for p1  p2 can be found directly, since we know that Z

Pˆ 1  Pˆ 2  1 p1  p2 2

p1 11  p1 2 p2 11  p2 2 n n2 1 B

is a standard normal random variable. Thus P(z兾2  Z  z兾2) ⯝ 1  , so we can substitute for Z in this last expression and use an approach similar to the one employed previously to find an approximate 100(1  )% two-sided confidence interval for p1  p2.

JWCL232_c10_351-400.qxd

1/15/10

2:15 PM

Page 393

393

10-6 INFERENCE ON TWO POPULATION PROPORTIONS

Approximate Confidence Interval on the Difference in Population Proportions

If pˆ 1 and pˆ 2 are the sample proportions of observations in two independent random samples of sizes n1 and n2 that belong to a class of interest, an approximate twosided 100(1 ⴚ ␣)% confidence interval on the difference in the true proportions p1 ⴚ p2 is pˆ1 11  pˆ1 2 pˆ2 11  pˆ 2 2 n n2 1 B

pˆ 1  pˆ2  z  2

 p1  p2  pˆ 1  pˆ 2 z  2 B

pˆ 1 11  pˆ 1 2 pˆ 2 11  pˆ 2 2 n1 n2

(10-41)

where z兾2 is the upper 兾2 percentage point of the standard normal distribution.

EXAMPLE 10-16 Defective Bearings Consider the process of manufacturing crankshaft bearings described in Example 8-7. Suppose that a modification is made in the surface finishing process and that, subsequently, a second random sample of 85 bearings is obtained. The number of defective bearings in this second sample is 8. Therefore, since n1  85, pˆ 1  0.12, n2  85, and pˆ 2  885  0.09, we can obtain an approximate 95% confidence interval on the difference in the proportion of defective bearings produced under the two processes from Equation 10-41 as follows: pˆ 1  pˆ 2  z0.025

pˆ 1 11  pˆ 1 2 pˆ 2 11  pˆ 2 2 n n2 1 B

 p1  p2  pˆ 1  pˆ 2 z0.025

B

pˆ 1 11  pˆ 1 2 pˆ 2 11  pˆ 2 2 n1 n2

or 0.12  0.09  1.96

0.1210.882 B

85

 p1  p2  0.12  0.09 1.96

B

0.0910.912 85 0.1210.882 85



0.0910.912 85

This simplifies to 0.06  p1  p2  0.12 Practical Interpretation: This confidence interval includes zero, so, based on the sample data, it seems unlikely that the changes made in the surface finish process have reduced the proportion of defective crankshaft bearings being produced.

EXERCISES FOR SECTION 10-6 10-66.

Consider the computer output below.

Sample 1 2

X 54 60

N 250 290

Sample p 0.216000 0.206897

Difference  p (1)  p (2) Estimate for difference: 0.00910345 95% CI for difference: (0.0600031, 0.0782100) Test for difference  0 (vs not  0): Z  ? P-Value  ?

(a) (b) (c) (d)

10-67. Consider the computer output below. Test and CI for Two Proportions

Test and Cl for Two Proportions

Is this a one-sided or a two-sided test? Fill in the missing values. Can the null hypothesis be rejected? Construct an approximate 90% CI for the difference in the two proportions.

Sample 1 2

X 188 245

N 250 350

Sample p 0.752000 0.700000

Difference  p (1)  p (2) Estimate for difference: 0.052 95% lower bound for difference: ? Test for difference  0 (vs 0) : Z  ? P-Value  ?

(a) Is this a one-sided or a two-sided test? (b) Fill in the missing values (c) Can the null hypothesis be rejected if   0.10? What if   0.05?

JWCL232_c10_351-400.qxd

394

1/15/10

2:15 PM

Page 394

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

10-68. An article in Knee Surgery, Sports Traumatology, Arthroscopy (2005, Vol. 13, pp. 273–279), considered arthroscopic meniscal repair with an absorbable screw. Results showed that for tears greater than 25 millimeters, 14 of 18 (78%) repairs were successful while for shorter tears, 22 of 30 (73%) repairs were successful. (a) Is there evidence that the success rate is greater for longer tears? Use   0.05. What is the P-value? (b) Calculate a one-sided 95% confidence bound on the difference in proportions that can be used to answer the question in part (a). 10-69. In the 2004 presidential election, exit polls from the critical state of Ohio provided the following results: For respondents with college degrees, 53% voted for Bush and 46% voted for Kerry. There were 2020 respondents. (a) Is there a significant difference in these proportions? Use   0.05. What is the P-value? (b) Calculate a 95% confidence interval for the difference in the two proportions and comment on the use of this interval to answer the question in part (a). 10-70. Two different types of injection-molding machines are used to form plastic parts. A part is considered defective if it has excessive shrinkage or is discolored. Two random samples, each of size 300, are selected, and 15 defective parts are found in the sample from machine 1 while 8 defective parts are found in the sample from machine 2. (a) Is it reasonable to conclude that both machines produce the same fraction of defective parts, using   0.05? Find the P-value for this test. (b) Construct a 95% confidence interval on the difference in the two fractions defective. (c) Suppose that p1  0.05 and p2  0.01. With the sample sizes given here, what is the power of the test for this twosided alternate?

(d) Suppose that p1  0.05 and p2  0.01. Determine the sample size needed to detect this difference with a probability of at least 0.9. (e) Suppose that p1  0.05 and p2  0.02. With the sample sizes given here, what is the power of the test for this twosided alternate? (f ) Suppose that p1  0.05 and p2  0.02. Determine the sample size needed to detect this difference with a probability of at least 0.9. 10-71. Two different types of polishing solutions are being evaluated for possible use in a tumble-polish operation for manufacturing interocular lenses used in the human eye following cataract surgery. Three hundred lenses were tumble polished using the first polishing solution, and of this number 253 had no polishing-induced defects. Another 300 lenses were tumble-polished using the second polishing solution, and 196 lenses were satisfactory upon completion. (a) Is there any reason to believe that the two polishing solutions differ? Use   0.01. What is the P-value for this test? (b) Discuss how this question could be answered with a confidence interval on p1  p2. 10-72. A random sample of 500 adult residents of Maricopa County found that 385 were in favor of increasing the highway speed limit to 75 mph, while another sample of 400 adult residents of Pima County found that 267 were in favor of the increased speed limit. (a) Do these data indicate that there is a difference in the support for increasing the speed limit between the residents of the two counties? Use   0.05. What is the P-value for this test? (b) Construct a 95% confidence interval on the difference in the two proportions. Provide a practical interpretation of this interval.

10-7 SUMMARY TABLE AND ROADMAP FOR INFERENCE PROCEDURES FOR TWO SAMPLES The table in the end papers of the book summarizes all of the two-sample parametric inference procedures given in this chapter. The table contains the null hypothesis statements, the test statistics, the criteria for rejection of the various alternative hypotheses, and the formulas for constructing the 100(1  )% confidence intervals. The roadmap to select the appropriate parametric confidence interval formula or hypothesis test method for one-sample problems was presented in Table 8-1. In Table 10-5, we extend the road map to two-sample problems. The primary comments stated previously also apply here (except we usually apply conclusions to a function of the parameters from each sample, such as the difference in means): 1. 2.

Determine the function of the parameters (and the distribution of the data) that is to be bounded by the confidence interval or tested by the hypothesis. Check if other parameters are known or need to be estimated (and if any assumptions are made).

JWCL232_c10_351-400.qxd

1/15/10

2:15 PM

Page 395

10-7 SUMMARY TABLE AND ROADMAPS FOR INFERENCE PROCEDURES FOR TWO SAMPLES

395

Table 10-5 Roadmap to Construct Confidence Intervals and Hypothesis Tests, Two-Sample Case Function of the Parameters to be Bounded by the Confidence Interval or Tested with a Hypothesis

Confidence Interval Section

Hypothesis Test Section

Symbol

Other Parameters?

Difference in means from two normal distributions Difference in means from two arbitrary distributions with large sample sizes Difference in means from two normal distributions Difference in means from two symmetric distributions

12

Standard deviations 1 and 2 known

10-1.3

10-1.1

12

Sample sizes large enough that 1 and 2 are essentially known

10-1.3

10-1.1

Large sample size is often taken to be n1 and n2  40

12

Standard deviations 1 and 2 are unknown, and assumed equal

10-2.3

10-2.1

Case 1: 1  2

10-3

Difference in means from two normal distributions

12

The Wilcoxon rank-sum test is a nonparametric procedure Case 2: 1  2

Difference in means from two normal distributions in a paired analysis

D 12

Ratio of variances of two normal distributions

21/22

Difference in two population proportions

12

Standard deviations 1 and 2 are unknown, and NOT assumed equal Standard deviation of differences are unknown

Means 1 and 2 unknown and estimated None

p1p2

Supplemental Exercises

N 20 20

Mean 11.87 12.73

StDev 2.23 3.19

10-4

10-2.1

10-4

10-5.4

10-5.2

10-6.3

10-6.1

Paired analysis calculates differences and uses a one-sample method for inference on the mean difference

Normal approximation to the binomial distribution used for the tests and confidence intervals

(a) Fill in the missing values. You may use bounds for the P-value. (b) Is this a two-sided test or a one-sided test? (c) What are your conclusions if   0.05? What if   0.10?

10-73. Consider the computer output below. Two-Sample T-Test and Cl Sample 1 2

10-2.3

Comments

SE Mean ? 0.71

Difference  mu (1)  mu (2) Estimate for difference: 0.860 95% CI for difference: (?, ?) T-Test of difference  0(vs not ) : T-Value  ? P-Value  ? DF  ? Both use Pooled StDev  ?

10-74. Consider the computer output below. Two-Sample T-Test CI Sample N Mean StDev SE Mean 1 16 22.45 2.98 0.75 2 25 24.61 5.36 1.1 Difference  mu (1)  mu (2) Estimate for difference: 2.16 T-Test of difference  0 (vs ): T-Value  1.65 P-Value  ? DF  ?

JWCL232_c10_351-400.qxd

396

1/15/10

2:15 PM

Page 396

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

(a) Is this a one-sided or a two-sided test? (b) Fill in the missing values. You may use bounds for the P-value. (c) What are your conclusions if   0.05? What if   0.10? (d) Find a 95% upper-confidence bound on the difference in the two means. 10-75. An article in the Journal of Materials Engineering (1989, Vol. 11, No. 4, pp. 275–282) reported the results of an experiment to determine failure mechanisms for plasmasprayed thermal barrier coatings. The failure stress for one particular coating (NiCrAlZr) under two different test conditions is as follows: Failure stress 1 106 Pa2 after nine 1-hour cycles: 19.8, 18.5, 17.6, 16.7, 16.7, 14.8, 15.4, 14.1, 13.6 Failure stress 1 106 Pa2 after six 1-hour cycles: 14.9, 12.7, 11.9, 11.4, 10.1, 7.9

(a) What assumptions are needed to construct confidence intervals for the difference in mean failure stress under the two different test conditions? Use normal probability plots of the data to check these assumptions. (b) Find a 99% confidence interval on the difference in mean failure stress under the two different test conditions. (c) Using the confidence interval constructed in part (b), does the evidence support the claim that the first test conditions yield higher results, on the average, than the second? Explain your answer. (d) Construct a 95% confidence interval on the ratio of the variances, 21/22, of failure stress under the two different test conditions. (e) Use your answer in part (b) to determine whether there is a significant difference in variances of the two different test conditions. Explain your answer. 10-76. A procurement specialist has purchased 25 resistors from vendor 1 and 35 resistors from vendor 2. Each resistor’s resistance is measured with the following results: Vendor 1 96.8 99.6 99.7 99.4 98.6

100.0 99.4 101.1 99.8

100.3 99.9 97.7 99.1

98.5 101.1 98.6 99.6

98.3 103.7 101.9 101.2

98.2 97.7 101.0 98.2

108.0 104.0 104.3 104.2 107.2 107.7

102.2 106.2 105.8 103.4 105.4

Vendor 2 106.8 103.2 102.6 104.0 104.6 106.4

106.8 103.7 100.3 106.3 103.5 106.8

104.7 106.8 104.0 102.2 106.3 104.1

104.7 105.1 107.0 102.8 109.2 107.1

(a) What distributional assumption is needed to test the claim that the variance of resistance of product from vendor 1 is not significantly different from the variance of resistance of product from vendor 2? Perform a graphical procedure to check this assumption. (b) Perform an appropriate statistical hypothesis-testing procedure to determine whether the procurement specialist can claim that the variance of resistance of product from vendor 1 is significantly different from the variance of resistance of product from vendor 2. 10-77. A liquid dietary product implies in its advertising that use of the product for one month results in an average weight loss of at least 3 pounds. Eight subjects use the product for one month, and the resulting weight loss data are reported below. Use hypothesis-testing procedures to answer the following questions.

Subject

Initial Weight (lb)

Final Weight (lb)

1 2 3 4 5 6 7 8

165 201 195 198 155 143 150 187

161 195 192 193 150 141 146 183

(a) Do the data support the claim of the producer of the dietary product with the probability of a type I error set to 0.05? (b) Do the data support the claim of the producer of the dietary product with the probability of a type I error set to 0.01? (c) In an effort to improve sales, the producer is considering changing its claim from “at least 3 pounds” to “at least 5 pounds.” Repeat parts (a) and (b) to test this new claim. 10-78. The breaking strength of yarn supplied by two manufacturers is being investigated. We know from experience with the manufacturers’ processes that 1  5 psi and 2  4 psi. A random sample of 20 test specimens from each manufacturer results in x1  88 psi and x2  91 psi, respectively. (a) Using a 90% confidence interval on the difference in mean breaking strength, comment on whether or not there is evidence to support the claim that manufacturer 2 produces yarn with higher mean breaking strength. (b) Using a 98% confidence interval on the difference in mean breaking strength, comment on whether or not there is evidence to support the claim that manufacturer 2 produces yarn with higher mean breaking strength. (c) Comment on why the results from parts (a) and (b) are different or the same. Which would you choose to make your decision and why?

JWCL232_c10_351-400.qxd

1/15/10

2:15 PM

Page 397

10-7 SUMMARY TABLE AND ROADMAPS FOR INFERENCE PROCEDURES FOR TWO SAMPLES

10-79. The Salk polio vaccine experiment in 1954 focused on the effectiveness of the vaccine in combating paralytic polio. Because it was felt that without a control group of children there would be no sound basis for evaluating the efficacy of the Salk vaccine, the vaccine was administered to one group, and a placebo (visually identical to the vaccine but known to have no effect) was administered to a second group. For ethical reasons, and because it was suspected that knowledge of vaccine administration would affect subsequent diagnoses, the experiment was conducted in a doubleblind fashion. That is, neither the subjects nor the administrators knew who received the vaccine and who received the placebo. The actual data for this experiment are as follows: Placebo group: n  201,299: 110 cases of polio observed Vaccine group: n  200,745: 33 cases of polio observed (a) Use a hypothesis-testing procedure to determine if the proportion of children in the two groups who contracted paralytic polio is statistically different. Use a probability of a type I error equal to 0.05. (b) Repeat part (a) using a probability of a type I error equal to 0.01. (c) Compare your conclusions from parts (a) and (b) and explain why they are the same or different. 10-80. Consider Supplemental Exercise 10-78. Suppose that prior to collecting the data, you decide that you want the error in estimating 1  2 by x1  x2 to be less than 1.5 psi. Specify the sample size for the following percentage confidence: (a) 90% (b) 98% (c) Comment on the effect of increasing the percentage confidence on the sample size needed. (d) Repeat parts (a)–(c) with an error of less than 0.75 psi instead of 1.5 psi. (e) Comment on the effect of decreasing the error on the sample size needed. 10-81. A random sample of 1500 residential telephones in Phoenix in 1990 found that 387 of the numbers were unlisted. A random sample in the same year of 1200 telephones in Scottsdale found that 310 were unlisted. (a) Find a 95% confidence interval on the difference in the two proportions and use this confidence interval to determine if there is a statistically significant difference in proportions of unlisted numbers between the two cities. (b) Find a 90% confidence interval on the difference in the two proportions and use this confidence interval to determine if there is a statistically significant difference in proportions of unlisted numbers between the two cities. (c) Suppose that all the numbers in the problem description were doubled. That is, 774 residents out of 3000 sampled in Phoenix and 620 residents out of 2400 in Scottsdale had unlisted phone numbers. Repeat parts (a) and (b) and comment on the effect of increasing the sample size without changing the proportions on your results.

397

10-82. In a random sample of 200 Phoenix residents who drive a domestic car, 165 reported wearing their seat belt regularly, while another sample of 250 Phoenix residents who drive a foreign car revealed 198 who regularly wore their seat belt. (a) Perform a hypothesis-testing procedure to determine if there is a statistically significant difference in seat belt usage between domestic and foreign car drivers. Set your probability of a type I error to 0.05. (b) Perform a hypothesis-testing procedure to determine if there is a statistically significant difference in seat belt usage between domestic and foreign car drivers. Set your probability of a type I error to 0.1. (c) Compare your answers for parts (a) and (b) and explain why they are the same or different. (d) Suppose that all the numbers in the problem description were doubled. That is, in a random sample of 400 Phoenix residents who drive a domestic car, 330 reported wearing their seat belt regularly, while another sample of 500 Phoenix residents who drive a foreign car revealed 396 who regularly wore their seat belt. Repeat parts (a) and (b) and comment on the effect of increasing the sample size without changing the proportions on your results. 10-83. Consider the previous exercise, which summarized data collected from drivers about their seat belt usage. (a) Do you think there is a reason not to believe these data? Explain your answer. (b) Is it reasonable to use the hypothesis-testing results from the previous problem to draw an inference about the difference in proportion of seat belt usage (i) of the spouses of these drivers of domestic and foreign cars? Explain your answer. (ii) of the children of these drivers of domestic and foreign cars? Explain your answer. (iii) of all drivers of domestic and foreign cars? Explain your answer. (iv) of all drivers of domestic and foreign trucks? Explain your answer. 10-84. A manufacturer of a new pain relief tablet would like to demonstrate that its product works twice as fast as the competitor’s product. Specifically, the manufacturer would like to test H0: 1  22 H1: 1 22 where 1 is the mean absorption time of the competitive product and 2 is the mean absorption time of the new product. Assuming that the variances  21 and  22 are known, develop a procedure for testing this hypothesis. 10-85. Two machines are used to fill plastic bottles with dishwashing detergent. The standard deviations of fill volume are known to be 1  0.10 fluid ounces and 2  0.15 fluid ounces for the two machines, respectively. Two random samples of n1  12 bottles from machine 1 and n2  10 bottles from machine 2 are selected, and the sample mean fill volumes are

JWCL232_c10_351-400.qxd

398

1/15/10

2:15 PM

Page 398

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

x1  30.87 fluid ounces and x2  30.68 fluid ounces. Assume normality. (a) Construct a 90% two-sided confidence interval on the mean difference in fill volume. Interpret this interval. (b) Construct a 95% two-sided confidence interval on the mean difference in fill volume. Compare and comment on the width of this interval to the width of the interval in part (a). (c) Construct a 95% upper-confidence interval on the mean difference in fill volume. Interpret this interval. (d) Test the hypothesis that both machines fill to the same mean volume. Use   0.05. What is the P-value? (e) If the -error of the test when the true difference in fill volume is 0.2 fluid ounces should not exceed 0.1, what sample sizes must be used? Use   0.05. 10-86. Suppose that we are testing H0: 1  2 versus H1 : 1  2, and we plan to use equal sample sizes from the two populations. Both populations are assumed to be normal with unknown but equal variances. If we use   0.05 and if the true mean 1  2 , what sample size must be used for the power of this test to be at least 0.90? 10-87. Consider the situation described in Exercise 10-71. (a) Redefine the parameters of interest to be the proportion of lenses that are unsatisfactory following tumble polishing with polishing fluids 1 or 2. Test the hypothesis that the two polishing solutions give different results using   0.01. (b) Compare your answer in part (a) with that for Exercise 10-71. Explain why they are the same or different. (c) We wish to use   0.01. Suppose that if p1  0.9 and p2  0.6, we wish to detect this with a high probability, say, at least 0.9. What sample sizes are required to meet this objective? 10-88. Consider the fire-fighting foam expanding agents investigated in Exercise 10-16, in which five observations of each agent were recorded. Suppose that, if agent 1 produces a mean expansion that differs from the mean expansion of agent 1 by 1.5, we would like to reject the null hypothesis with probability at least 0.95. (a) What sample size is required? (b) Do you think that the original sample size in Exercise 10-16 was appropriate to detect this difference? Explain your answer. 10-89. A fuel-economy study was conducted for two German automobiles, Mercedes and Volkswagen. One vehicle of each brand was selected, and the mileage performance was observed for 10 tanks of fuel in each car. The data are as follows (in miles per gallon): Mercedes 24.7 24.8 24.9 24.7 24.5

24.9 24.6 23.9 24.9 24.8

Volkswagen 41.7 42.3 41.6 39.5 41.9

42.8 42.4 39.9 40.8 29.6

(a) Construct a normal probability plot of each of the data sets. Based on these plots, is it reasonable to assume that they are each drawn from a normal population? (b) Suppose that it was determined that the lowest observation of the Mercedes data was erroneously recorded and should be 24.6. Furthermore, the lowest observation of the Volkswagen data was also mistaken and should be 39.6. Again construct normal probability plots of each of the data sets with the corrected values. Based on these new plots, is it reasonable to assume that they are each drawn from a normal population? (c) Compare your answers from parts (a) and (b) and comment on the effect of these mistaken observations on the normality assumption. (d) Using the corrected data from part (b) and a 95% confidence interval, is there evidence to support the claim that the variability in mileage performance is greater for a Volkswagen than for a Mercedes? (e) Rework part (d) of this problem using an appropriate hypothesis-testing procedure. Did you get the same answer as you did originally? Why? 10-90. An experiment was conducted to compare the filling capability of packaging equipment at two different wineries. Ten bottles of pinot noir from Ridgecrest Vineyards were randomly selected and measured, along with 10 bottles of pinot noir from Valley View Vineyards. The data are as follows (fill volume is in milliliters): Ridgecrest 755 753 752

751 753 751

752 753

Valley View 753 754

756 755 755

754 756 756

757 756

756 755

(a) What assumptions are necessary to perform a hypothesistesting procedure for equality of means of these data? Check these assumptions. (b) Perform the appropriate hypothesis-testing procedure to determine whether the data support the claim that both wineries will fill bottles to the same mean volume. (c) Suppose that the true difference in mean fill volume is as much as 2 fluid ounces; did the sample sizes of 10 from each vineyard provide good detection capability when   0.05? Explain your answer. 10-91. A Rockwell hardness-testing machine presses a tip into a test coupon and uses the depth of the resulting depression to indicate hardness. Two different tips are being compared to determine whether they provide the same Rockwell C-scale hardness readings. Nine coupons are tested, with both tips being tested on each coupon. The data are shown in the accompanying table. (a) State any assumptions necessary to test the claim that both tips produce the same Rockwell C-scale hardness readings. Check those assumptions for which you have the information.

JWCL232_c10_351-400.qxd

1/15/10

2:15 PM

Page 399

10-7 SUMMARY TABLE AND ROADMAPS FOR INFERENCE PROCEDURES FOR TWO SAMPLES

Coupon

Tip 1

Tip 2

Coupon

Tip 1

Tip 2

1 2 3 4 5

47 42 43 40 42

46 40 45 41 43

6 7 8 9

41 45 45 49

41 46 46 48

(b) Apply an appropriate statistical method to determine if the data support the claim that the difference in Rockwell C-scale hardness readings of the two tips is significantly different from zero. (c) Suppose that if the two tips differ in mean hardness readings by as much as 1.0, we want the power of the test to be at least 0.9. For an   0.01, how many coupons should have been used in the test? 10-92. Two different gauges can be used to measure the depth of bath material in a Hall cell used in smelting aluminum. Each gauge is used once in 15 cells by the same operator. Cell

Gauge 1

Gauge 2

Cell

Gauge 1

Gauge 2

1 2 3 4 5 6 7 8

46 in. 50 47 53 49 48 53 56

47 in. 53 45 50 51 48 54 53

9 10 11 12 13 14 15

52 47 49 45 47 46 50

51 45 51 45 49 43 51

(a) State any assumptions necessary to test the claim that both gauges produce the same mean bath depth readings. Check those assumptions for which you have the information. (b) Apply an appropriate statistical procedure to determine if the data support the claim that the two gauges produce different mean bath depth readings. (c) Suppose that if the two gauges differ in mean bath depth readings by as much as 1.65 inch, we want the power of the test to be at least 0.8. For   0.01, how many cells should have been used? 10-93. An article in the Journal of the Environmental Engineering Division [“Distribution of Toxic Substances in Rivers” (1982, Vol. 108, pp. 639–649)] investigated the concentration of several hydrophobic organic substances in the Wolf River in Tennessee. Measurements on hexachlorobenzene (HCB) in nanograms per liter were taken at different depths downstream of an abandoned dump site. Data for two depths follow: Surface: 3.74, 4.61, 4.00, 4.67, 4.87, 5.12, 4.52, 5.29, 5.74, 5.48 Bottom: 5.44, 6.88, 5.37, 5.44, 5.03, 6.48, 3.89, 5.85, 6.85, 7.16 (a) What assumptions are required to test the claim that mean HCB concentration is the same at both depths? Check those assumptions for which you have the information. (b) Apply an appropriate procedure to determine if the data support the claim in part a. (c) Suppose that the true difference in mean concentrations is 2.0 nanograms per liter. For   0.05, what is the power of a statistical test for H0: 1  2 versus H1: 1  2? (d) What sample size would be required to detect a difference of 1.0 nanograms per liter at   0.05 if the power must be at least 0.9?

MIND-EXPANDING EXERCISES 10-94. Three different pesticides can be used to control infestation of grapes. It is suspected that pesticide 3 is more effective than the other two. In a particular vineyard, three different plantings of pinot noir grapes are selected for study. The following results on yield are obtained:

Pesticide

xi (Bushels/ Plant)

si

ni (Number of Plants)

1 2 3

4.6 5.2 6.1

0.7 0.6 0.8

100 120 130

399

If i is the true mean yield after treatment with the i th pesticide, we are interested in the quantity 

1 1 2 2  3 2 1

which measures the difference in mean yields between pesticides 1 and 2 and pesticide 3. If the samˆ ) obtained ple sizes ni are large, the estimator (say,  by replacing each individual i by Xi is approximately normal. (a) Find an approximate 100(1  )% large-sample confidence interval for .

JWCL232_c10_351-400.qxd

400

1/15/10

2:15 PM

Page 400

CHAPTER 10 STATISTICAL INFERENCE FOR TWO SAMPLES

MIND-EXPANDING EXERCISES (b) Do these data support the claim that pesticide 3 is more effective than the other two? Use   0.05 in determining your answer. 10-95. Suppose that we wish to test H0: 1  2 versus H1: 1  2, where 12 and 22 are known. The total sample size N is to be determined, and the allocation of observations to the two populations such that n1 n2  N is to be made on the basis of cost. If the cost of sampling for populations 1 and 2 are C1 and C2, respectively, find the minimum cost sample sizes that provide a specified variance for the difference in sample means. 10-96. Suppose that we wish to test the hypothesis H0: 1  2 versus H1: 1  2, where both variances 12 and  22 are known. A total of n1 n2  N observations can be taken. How should these observations be allocated to the two populations to maximize the probability that H0 will be rejected if H1 is true and 1  2    0? 10-97. Suppose that we wish to test H0:   0 versus H1:   0, where the population is normal with known . Let 0    , and define the critical region so that we will reject H0 if z0 z or if z0  z, where z0 is the value of the usual test statistic for these hypotheses. (a) Show that the probability of type I error for this test is .

(b) Suppose that the true mean is 1  0 . Derive an expression for  for the above test. 10-98. Construct a data set for which the paired t-test statistic is very large, indicating that when this analysis is used the two population means are different, but t0 for the two-sample t-test is very small so that the incorrect analysis would indicate that there is no significant difference between the means. 10-99. In some situations involving proportions, we are interested in the ratio  p1兾p2 rather than the difference p1  p2. Let ˆ  pˆ 1兾 pˆ 2 . We can show that ln( ˆ ) has an approximate normal distribution with the mean (n  ) and variance 3 1n1  x1 2  1n1x1 2 1n2  x2 2  1n2x2 2 4 1 2. (a) Use the information above to derive a large-sample confidence interval for ln . (b) Show how to find a large-sample CI for . (c) Use the data from the St. John’s Wort study in Example 10-15, and find a 95% CI on  p1兾p2. Provide a practical interpretation for this CI. 10-100. Derive an expression for  for the test of the equality of the variances of two normal distributions. Assume that the two-sided alternative is specified.

IMPORTANT TERMS AND CONCEPTS Comparative experiments Confidence intervals on differences and ratios Critical region for a test statistic

Identifying cause and effect Null and alternative hypotheses One-sided and twosided alternative hypotheses

Operating characteristic curves Paired t-test Pooled t-test P-value Reference distribution for a test statistic

Sample size determination for hypothesis tests and confidence intervals Statistical hypotheses Test statistic Wilcoxon rank-sum test

JWCL232_c11_401-448.qxd

1/15/10

5:04 PM

Page 401

11

(15 July 2009)—Space Shuttle Endeavour and its seven-member STS-127 crew head toward Earth orbit and rendezvous with the International Space Station Courtesy NASA

Simple Linear Regression and Correlation The space shuttle Challenger accident in January 1986 was the result of the failure of O-rings used to seal field joints in the solid rocket motor due to the extremely low ambient temperatures at the time of launch. Prior to the launch there were data on the occurrence of O-ring failure and the corresponding temperature on 24 prior launches or static firings of the motor. In this chapter we will see how to build a statistical model relating the probability of O-ring failure to temperature. This model provides a measure of the risk associated with launching the shuttle at the low temperature occurring when Challenger was launched.

CHAPTER OUTLINE 11-1

EMPIRICAL MODELS

11-2

SIMPLE LINEAR REGRESSION

11-3

PROPERTIES OF THE LEAST SQUARES ESTIMATORS

11-6

PREDICTION OF NEW OBSERVATIONS

11-4

HYPOTHESIS TESTS IN SIMPLE LINEAR REGRESSION

11-7

ADEQUACY OF THE REGRESSION MODEL

11-5.2 Confidence Interval on the Mean Response

11-7.1 Residual Analysis

11-4.1 Use of t-Tests

11-5

11-7.2 Coefficient of Determination (R2)

11-4.2 Analysis of Variance Approach to Test Significance of Regression

11-8

CORRELATION

CONFIDENCE INTERVALS

11-9

11-5.1 Confidence Intervals on the Slope and Intercept

REGRESSION ON TRANSFORMED VARIABLES

11-10 LOGISTIC REGRESSION

401

JWCL232_c11_401-448.qxd

402

1/15/10

5:36 PM

Page 402

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

LEARNING OBJECTIVES After careful study of this chapter, you should be able to do the following: 1. Use simple linear regression for building empirical models to engineering and scientific data 2. Understand how the method of least squares is used to estimate the parameters in a linear regression model 3. Analyze residuals to determine if the regression model is an adequate fit to the data or to see if any underlying assumptions are violated 4. Test statistical hypotheses and construct confidence intervals on regression model parameters 5. Use the regression model to make a prediction of a future observation and construct an appropriate prediction interval on the future observation 6. Apply the correlation model 7. Use simple transformations to achieve a linear regression model

11-1 EMPIRICAL MODELS Many problems in engineering and the sciences involve a study or analysis of the relationship between two or more variables. For example, the pressure of a gas in a container is related to the temperature, the velocity of water in an open channel is related to the width of the channel, and the displacement of a particle at a certain time is related to its velocity. In this last example, if we let d0 be the displacement of the particle from the origin at time t = 0 and v be the velocity, then the displacement at time t is dt = d0 + vt. This is an example of a deterministic linear relationship, because (apart from measurement errors) the model predicts displacement perfectly. However, there are many situations where the relationship between variables is not deterministic. For example, the electrical energy consumption of a house ( y) is related to the size of the house (x, in square feet), but it is unlikely to be a deterministic relationship. Similarly, the fuel usage of an automobile ( y) is related to the vehicle weight x, but the relationship is not a deterministic one. In both of these examples the value of the response of interest y (energy consumption, fuel usage) cannot be predicted perfectly from knowledge of the corresponding x. It is possible for different automobiles to have different fuel usage even if they weigh the same, and it is possible for different houses to use different amounts of electricity even if they are the same size. The collection of statistical tools that are used to model and explore relationships between variables that are related in a nondeterministic manner is called regression analysis. Because problems of this type occur so frequently in many branches of engineering and science, regression analysis is one of the most widely used statistical tools. In this chapter we present the situation where there is only one independent or predictor variable x and the relationship with the response y is assumed to be linear. While this seems to be a simple scenario, there are many practical problems that fall into this framework. For example, in a chemical process, suppose that the yield of the product is related to the process-operating temperature. Regression analysis can be used to build a model to predict yield at a given temperature level. This model can also be used for process optimization, such as finding the level of temperature that maximizes yield, or for process control purposes.

JWCL232_c11_401-448.qxd

1/14/10

8:01 PM

Page 403

11-1 EMPIRICAL MODELS

403

As an illustration, consider the data in Table 11-1. In this table y is the purity of oxygen produced in a chemical distillation process, and x is the percentage of hydrocarbons that are present in the main condenser of the distillation unit. Figure 11-1 presents a scatter diagram of the data in Table 11-1. This is just a graph on which each (xi, yi) pair is represented as a point plotted in a two-dimensional coordinate system. This scatter diagram was produced by Minitab, and we selected an option that shows dot diagrams of the x and y variables along the top and right margins of the graph, respectively, making it easy to see the distributions of the individual variables (box plots or histograms could also be selected). Inspection of this scatter diagram indicates that, although no simple curve will pass exactly through all the points, there is a strong indication that the points lie scattered randomly around a straight line. Therefore, it is probably reasonable to assume that the mean of the random variable Y is related to x by the following straight-line relationship: E1Y 0 x2  Y 0 x  0  1x where the slope and intercept of the line are called regression coefficients. While the mean of Y is a linear function of x, the actual observed value y does not fall exactly on a straight line. The appropriate way to generalize this to a probabilistic linear model is to assume that the expected value of Y is a linear function of x, but that for a fixed value of x the actual value of Y is determined by the mean value function (the linear model) plus a random error term, say, Y  0  1x  

(11-1)

Table 11-1 Oxygen and Hydrocarbon Levels Hydrocarbon Level x (%)

Purity y (%)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0.99 1.02 1.15 1.29 1.46 1.36 0.87 1.23 1.55 1.40 1.19 1.15 0.98 1.01 1.11 1.20 1.26 1.32 1.43 0.95

90.01 89.05 91.43 93.74 96.73 94.45 87.59 91.77 99.42 93.65 93.54 92.52 90.56 89.54 89.85 90.39 93.25 93.41 94.98 87.33

100 98 96 Purity ( y)

Observation Number

94 92 90 88 86 0.85

0.95

1.05

1.15 1.25 1.35 Hydrocarbon level (x)

1.45

1.55

Figure 11-1 Scatter diagram of oxygen purity versus hydrocarbon level from Table 11-1.

JWCL232_c11_401-448.qxd

404

1/14/10

8:01 PM

Page 404

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

where  is the random error term. We will call this model the simple linear regression model, because it has only one independent variable or regressor. Sometimes a model like this will arise from a theoretical relationship. At other times, we will have no theoretical knowledge of the relationship between x and y, and the choice of the model is based on inspection of a scatter diagram, such as we did with the oxygen purity data. We then think of the regression model as an empirical model. To gain more insight into this model, suppose that we can fix the value of x and observe the value of the random variable Y. Now if x is fixed, the random component  on the righthand side of the model in Equation 11-1 determines the properties of Y. Suppose that the mean and variance of  are 0 and 2, respectively. Then, E1Y 0 x2  E10  1x  2  0  1x  E12  0  1x Notice that this is the same relationship that we initially wrote down empirically from inspection of the scatter diagram in Fig. 11-1. The variance of Y given x is V 1Y 0 x2  V 10  1x  2  V 10  1x2  V 12  0  2  2 Thus, the true regression model Y 0 x  0  1x is a line of mean values; that is, the height of the regression line at any value of x is just the expected value of Y for that x. The slope, 1, can be interpreted as the change in the mean of Y for a unit change in x. Furthermore, the variability of Y at a particular value of x is determined by the error variance 2. This implies that there is a distribution of Y-values at each x and that the variance of this distribution is the same at each x. For example, suppose that the true regression model relating oxygen purity to hydrocarbon level is Y 0 x  75  15x, and suppose that the variance is 2  2. Figure 11-2 illustrates this situation. Notice that we have used a normal distribution to describe the random variation in . Since Y is the sum of a constant 0  1x (the mean) and a normally distributed random variable, Y is a normally distributed random variable. The variance 2 determines the variability in the observations Y on oxygen purity. Thus, when 2 is small, the observed values of Y will fall close to the line, and when 2 is large, the observed values of Y may deviate considerably from the line. Because 2 is constant, the variability in Y at any value of x is the same. The regression model describes the relationship between oxygen purity Y and hydrocarbon level x. Thus, for any value of hydrocarbon level, oxygen purity has a normal distribution y (Oxygen purity) β 0 + β 1 (1.25)

True regression line μ Y ⎜x = β 0 + β1x = 75 + 15x

β 0 + β 1 (1.00)

x = 1.00

x = 1.25

x (Hydrocarbon level)

Figure 11-2 The distribution of Y for a given value of x for the oxygen purity–hydrocarbon data.

JWCL232_c11_401-448.qxd

1/14/10

8:01 PM

Page 405

11-2 SIMPLE LINEAR REGRESSION

405

with mean 75  15x and variance 2. For example, if x  1.25, Y has mean value Y  x  75  15(1.25)  93.75 and variance 2. In most real-world problems, the values of the intercept and slope (0, 1) and the error variance 2 will not be known, and they must be estimated from sample data. Then this fitted regression equation or model is typically used in prediction of future observations of Y, or for estimating the mean response at a particular level of x. To illustrate, a chemical engineer might be interested in estimating the mean purity of oxygen produced when the hydrocarbon level is x  1.25%. This chapter discusses such procedures and applications for the simple linear regression model. Chapter 12 will discuss multiple linear regression models that involve more than one regressor. Historical Note Sir Francis Galton first used the term regression analysis in a study of the heights of fathers (x) and sons ( y). Galton fit a least squares line and used it to predict the son’s height from the father’s height. He found that if a father’s height was above average, the son’s height would also be above average, but not by as much as the father’s height was. A similar effect was observed for below average heights. That is, the son’s height “regressed” toward the average. Consequently, Galton referred to the least squares line as a regression line. Abuses of Regression Regression is widely used and frequently misused; several common abuses of regression are briefly mentioned here. Care should be taken in selecting variables with which to construct regression equations and in determining the form of the model. It is possible to develop statistically significant relationships among variables that are completely unrelated in a causal sense. For example, we might attempt to relate the shear strength of spot welds with the number of empty parking spaces in the visitor parking lot. A straight line may even appear to provide a good fit to the data, but the relationship is an unreasonable one on which to rely. You can’t increase the weld strength by blocking off parking spaces. A strong observed association between variables does not necessarily imply that a causal relationship exists between those variables. This type of effect is encountered fairly often in retrospective data analysis, and even in observational studies. Designed experiments are the only way to determine causeand-effect relationships. Regression relationships are valid only for values of the regressor variable within the range of the original data. The linear relationship that we have tentatively assumed may be valid over the original range of x, but it may be unlikely to remain so as we extrapolate—that is, if we use values of x beyond that range. In other words, as we move beyond the range of values of x for which data were collected, we become less certain about the validity of the assumed model. Regression models are not necessarily valid for extrapolation purposes. Now this does not mean don’t ever extrapolate. There are many problem situations in science and engineering where extrapolation of a regression model is the only way to even approach the problem. However, there is a strong warning to be careful. A modest extrapolation may be perfectly all right in many cases, but a large extrapolation will almost never produce acceptable results.

11-2 SIMPLE LINEAR REGRESSION The case of simple linear regression considers a single regressor variable or predictor variable x and a dependent or response variable Y. Suppose that the true relationship between Y and x is a straight line and that the observation Y at each level of x is a random variable. As noted

JWCL232_c11_401-448.qxd

406

1/14/10

8:01 PM

Page 406

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

previously, the expected value of Y for each value of x is E1Y 0 x2  0  1 x where the intercept 0 and the slope 1 are unknown regression coefficients. We assume that each observation, Y, can be described by the model Y  0  1 x  

(11-2)

where  is a random error with mean zero and (unknown) variance 2. The random errors corresponding to different observations are also assumed to be uncorrelated random variables. Suppose that we have n pairs of observations (x1, y1), (x2, y2), p , (xn, yn). Figure 11-3 shows a typical scatter plot of observed data and a candidate for the estimated regression line. The estimates of 0 and 1 should result in a line that is (in some sense) a “best fit” to the data. The German scientist Karl Gauss (1777–1855) proposed estimating the parameters 0 and 1 in Equation 11-2 to minimize the sum of the squares of the vertical deviations in Fig. 11-3. We call this criterion for estimating the regression coefficients the method of least squares. Using Equation 11-2, we may express the n observations in the sample as yi  0  1 xi  i,

i  1, 2, p , n

(11-3)

and the sum of the squares of the deviations of the observations from the true regression line is n

n

i1

i1

L  a 2i  a 1 yi 0 1xi 2 2

(11-4)

ˆ and  ˆ , must satisfy The least squares estimators of 0 and 1, say,  0 1 n L ˆ  ˆ x2 0 `  2 a 1 yi  0 1 i 0 ˆ 0,ˆ 1 i1 n L ˆ  ˆ x 2x  0 `  2 a 1 yi  0 1 i i 1 ˆ 0,ˆ 1 i1 y Observed value Data (y)

Estimated regression line

x

Figure 11-3 Deviations of the data from the estimated regression model.

(11-5)

JWCL232_c11_401-448.qxd

1/15/10

4:53 PM

Page 407

407

11-2 SIMPLE LINEAR REGRESSION

Simplifying these two equations yields n

n

i1 n

i1 n

ˆ  ˆ n 0 1 a xi  a yi n

2 ˆ ˆ  0 a xi  1 a x i  a yi xi i1

i1

(11-6)

i1

Equations 11-6 are called the least squares normal equations. The solution to the normal ˆ and  ˆ . equations results in the least squares estimators  0 1 Least Squares Estimates

The least squares estimates of the intercept and slope in the simple linear regression model are ˆ y  ˆ x  0 1

(11-7)

n

n

a a yi b a a xi b

i1

n

a yi xi

ˆ   1

n

i1

i1

n

n

i1

n

2 a xi

(11-8)

2

a a xi b i1

where y  11 n2 g i1 yi and x  11 n2 g i1 xi. n

n

The fitted or estimated regression line is therefore ˆ  ˆ x yˆ   0 1

(11-9)

Note that each pair of observations satisfies the relationship ˆ  ˆ x  e, yi   0 1 i i

i  1, 2, p , n

where ei  yi yˆi is called the residual. The residual describes the error in the fit of the model to the ith observation yi. Later in this chapter we will use the residuals to provide information about the adequacy of the fitted model. Notationally, it is occasionally convenient to give special symbols to the numerator and denominator of Equation 11-8. Given data (x1, y1), (x2, y2), p , (xn, yn), let n

2

n

n

a a xi b

i1

i1

n

Sx x  a 1xi x2 2  a x 2i

i1

(11-10)

and n

n

n

i1

n

Sx y  a 1yi y21xi x2  a xi yi i1

n

a a xi b a a yi b i1

i1

(11-11)

JWCL232_c11_401-448.qxd

408

1/14/10

8:02 PM

Page 408

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

EXAMPLE 11-1

Oxygen Purity

We will fit a simple linear regression model to the oxygen purity data in Table 11-1. The following quantities may be computed: n  20

20

Therefore, the least squares estimates of the slope and intercept are ˆ   1

20

a xi  23.92

a yi  1,843.21

i1

i1

2 a x i  29.2892

i1

10.17744  14.94748 0.68088

ˆ y  ˆ x  92.1605 114.9474821.196  74.28331  0 1

20

2 a y i  170,044.5321



and

x  1.1960 y  92.1605 20

Sx y Sx x

i1

The fitted simple linear regression model (with the coefficients reported to three decimal places) is

20

a xi yi  2,214.6566 i1

20

2

20

a a xi b

i1

20

Sx x  a x 2i

i1

 29.2892

123.922

yˆ  74.283  14.947 x 2

20

 0.68088 and 20

20

20

a a xi b a a yi b

i1

20

Sx y  a xi yi

i1

 2,214.6566

i1

123.922 11,843.212 20

 10.17744

This model is plotted in Fig. 11-4, along with the sample data. Practical Interpretation: Using the regression model, we would predict oxygen purity of yˆ  89.23% when the hydrocarbon level is x  1.00%. The purity 89.23% may be interpreted as an estimate of the true population mean purity when x  1.00%, or as an estimate of a new observation when x = 1.00%. These estimates are, of course, subject to error; that is, it is unlikely that a future observation on purity would be exactly 89.23% when the hydrocarbon level is 1.00%. In subsequent sections we will see how to use confidence intervals and prediction intervals to describe the error in estimation from a regression model.

Computer software programs are widely used in regression modeling. These programs typically carry more decimal places in the calculations. Table 11-2 shows a portion of the ˆ and  ˆ are highlighted. In subseoutput from Minitab for this problem. The estimates  0 1 quent sections we will provide explanations for the information provided in this computer output. 102

Oxygen purity y (%)

99

Figure 11-4 Scatter plot of oxygen purity y versus hydrocarbon level x and regression model yˆ  74.283  14.947x.

96

93

90

87

0.87

1.07

1.27 Hydrocarbon level (%) x

1.47

1.67

JWCL232_c11_401-448.qxd

1/14/10

8:02 PM

Page 409

409

11-2 SIMPLE LINEAR REGRESSION

Table 11-2 Minitab Output for the Oxygen Purity Data in Example 11-1 Regression Analysis The regression equation is Purity  74.3  14.9 HC Level Predictor Constant HC Level

Coef 74.283 14.947

S  1.087

SE Coef 1.593 1.317

ˆ 0 ˆ 1

T 46.62 11.35

R-Sq  87.7%

P 0.000 0.000

R-Sq (adj)  87.1%

Analysis of Variance Source Regression Residual Error Total

DF 1 18 19

SS 152.13 21.25 173.38

MS 152.13 1.18

SSE

F 128.86

P 0.000

ˆ 2

Predicted Values for New Observations New Obs 1

Fit 89.231

SE Fit 0.354

95.0% CI (88.486, 89.975)

95.0% PI (86.830, 91.632)

Values of Predictors for New Observations New Obs 1

HC Level 1.00

Estimating 2 There is actually another unknown parameter in our regression model, 2 (the variance of the error term ). The residuals ei  yi yˆi are used to obtain an estimate of 2. The sum of squares of the residuals, often called the error sum of squares, is n

n

i1

i1

SSE  a e2i  a 1 yi yˆi 2 2

(11-12)

We can show that the expected value of the error sum of squares is E(SSE)  (n 2)2. Therefore an unbiased estimator of 2 is Estimator of Variance

ˆ2 

SSE n 2

(11-13)

Computing SSE using Equation 11-12 would be fairly tedious. A more convenient computing ˆ  ˆ x into Equation 11-12 and simplifying. formula can be obtained by substituting yˆi   0 1 i The resulting computing formula is

ˆ S SSE  SST  1 xy

(11-14)

JWCL232_c11_401-448.qxd

410

1/15/10

4:53 PM

Page 410

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

where SST  g i1 1 yi y 2 2  g i1 y 2i ny 2 is the total sum of squares of the response variable y. Formulas such as this are presented in Section 11-4. The error sum of squares and the estimate of 2 for the oxygen purity data, ˆ 2  1.18, are highlighted in the Minitab output in Table 11-2. n

n

EXERCISES FOR SECTION 11-2 11-1. An article in Concrete Research [“Near Surface Characteristics of Concrete: Intrinsic Permeability” (Vol. 41, 1989)] presented data on compressive strength x and intrinsic permeability y of various concrete mixes and cures. Summary quantities are n  14, gyi  572, g y2i  23,530, g xi  43, g x 2i  157.42, and g xi yi  1697.80. Assume that the two variables are related according to the simple linear regression model. (a) Calculate the least squares estimates of the slope and intercept. Estimate 2. Graph the regression line. (b) Use the equation of the fitted line to predict what permeability would be observed when the compressive strength is x  4.3. (c) Give a point estimate of the mean permeability when compressive strength is x  3.7. (d) Suppose that the observed value of permeability at x  3.7 is y  46.1. Calculate the value of the corresponding residual. 11-2. Regression methods were used to analyze the data from a study investigating the relationship between roadway surface temperature (x) and pavement deflection ( y). Summary quantities were n  20, g yi  12.75, g y 2i  8.86, g xi  1478, g x 2i  143,215.8, and g xi yi  1083.67. (a) Calculate the least squares estimates of the slope and intercept. Graph the regression line. Estimate 2. (b) Use the equation of the fitted line to predict what pavement deflection would be observed when the surface temperature is 85F. (c) What is the mean pavement deflection when the surface temperature is 90F? (d) What change in mean pavement deflection would be expected for a 1F change in surface temperature? 11-3. The following table presents data on the ratings of quarterbacks for the 2008 National Football League season (source: The Sports Network). It is suspected that the rating (y) is related to the average number of yards gained per pass attempt (x). (a) Calculate the least squares estimates of the slope and intercept. What is the estimate of 2? Graph the regression model. (b) Find an estimate of the mean rating if a quarterback averages 7.5 yards per attempt. (c) What change in the mean rating is associated with a decrease of one yard per attempt? (d) To increase the mean rating by 10 points, how much increase in the average yards per attempt must be generated?

(e) Given that x  7.21 yards, find the fitted value of y and the corresponding residual.

Player Philip Chad Kurt Drew Peyton Aaron Matt Tony Jeff Matt Matt Shaun Seneca Eli Donovan Jay Trent Jake Jason David Brett Joe Kerry Ben Kyle JaMarcus Tyler Gus Dan Marc Ryan Derek

Rivers Pennington Warner Brees Manning Rodgers Schaub Romo Garcia Cassel Ryan Hill Wallace Manning McNabb Cutler Edwards Delhomme Campbell Garrard Favre Flacco Collins Roethlisberger Orton Russell Thigpen Freotte Orlovsky Bulger Fitzpatrick Anderson

Team

Yards per Attempt

Rating Points

SD MIA ARI NO IND GB HOU DAL TB NE ATL SF SEA NYG PHI DEN BUF CAR WAS JAC NYJ BAL TEN PIT CHI OAK KC MIN DET STL CIN CLE

8.39 7.67 7.66 7.98 7.21 7.53 8.01 7.66 7.21 7.16 7.93 7.10 6.33 6.76 6.86 7.35 7.22 7.94 6.41 6.77 6.65 6.94 6.45 7.04 6.39 6.58 6.21 7.17 6.34 6.18 5.12 5.71

105.5 97.4 96.9 96.2 95 93.8 92.7 91.4 90.2 89.4 87.7 87.5 87 86.4 86.4 86 85.4 84.7 84.3 81.7 81 80.3 80.2 80.1 79.6 77.1 76 73.7 72.6 71.4 70 66.5

JWCL232_c11_401-448.qxd

1/14/10

8:02 PM

Page 411

411

11-2 SIMPLE LINEAR REGRESSION

11-4. An article in Technometrics by S. C. Narula and J. F. Wellington [“Prediction, Linear Regression, and a Minimum Sum of Relative Errors” (Vol. 19, 1977)] presents data on the selling price and annual taxes for 24 houses. The data are shown in the following table.

Taxes Sale (Local, School), Price/1000 County)/1000 25.9 29.5 27.9 25.9 29.9 29.9 30.9 28.9 35.9 31.5 31.0 30.9

4.9176 5.0208 4.5429 4.5573 5.0597 3.8910 5.8980 5.6039 5.8282 5.3003 6.2712 5.9592

Taxes Sale (Local, School), Price/1000 County)/1000 30.0 36.9 41.9 40.5 43.9 37.5 37.9 44.5 37.9 38.9 36.9 45.8

5.0500 8.2464 6.6969 7.7841 9.0384 5.9894 7.5422 8.7951 6.0831 8.3607 8.1400 9.1416

(a) Assuming that a simple linear regression model is appropriate, obtain the least squares fit relating selling price to taxes paid. What is the estimate of 2? (b) Find the mean selling price given that the taxes paid are x  7.50. (c) Calculate the fitted value of y corresponding to x  5.8980. Find the corresponding residual. (d) Calculate the fitted yˆi for each value of xi used to fit the model. Then construct a graph of yˆi versus the corresponding observed value yi and comment on what this plot would look like if the relationship between y and x was a deterministic (no random error) straight line. Does the plot actually obtained indicate that taxes paid is an effective regressor variable in predicting selling price? 11-5. The number of pounds of steam used per month by a chemical plant is thought to be related to the average ambient temperature (in F) for that month. The past year’s usage and temperature are shown in the following table: Month Temp. Usage/1000

Month Temp. Usage/1000

Jan. Feb. Mar. Apr. May June

July Aug. Sept. Oct. Nov. Dec.

21 24 32 47 50 59

185.79 214.47 288.03 424.84 454.58 539.03

68 74 62 50 41 30

621.55 675.06 562.03 452.93 369.95 273.98

(a) Assuming that a simple linear regression model is appropriate, fit the regression model relating steam usage ( y) to the average temperature (x). What is the estimate of  2? Graph the regression line. (b) What is the estimate of expected steam usage when the average temperature is 55F? (c) What change in mean steam usage is expected when the monthly average temperature changes by 1F? (d) Suppose the monthly average temperature is 47F. Calculate the fitted value of y and the corresponding residual. 11-6. The following table presents the highway gasoline mileage performance and engine displacement for DaimlerChrysler vehicles for model year 2005 (source: U.S. Environmental Protection Agency). (a) Fit a simple linear model relating highway miles per gallon ( y) to engine displacement (x) in cubic inches using least squares. (b) Find an estimate of the mean highway gasoline mileage performance for a car with 150 cubic inches engine displacement. (c) Obtain the fitted value of y and the corresponding residual for a car, the Neon, with an engine displacement of 122 cubic inches.

Carline 300C/SRT-8 CARAVAN 2WD CROSSFIRE ROADSTER DAKOTA PICKUP 2WD DAKOTA PICKUP 4WD DURANGO 2WD GRAND CHEROKEE 2WD GRAND CHEROKEE 4WD LIBERTY/CHEROKEE 2WD LIBERTY/CHEROKEE 4WD NEON/SRT-4/SX 2.0 PACIFICA 2WD PACIFICA AWD PT CRUISER RAM 1500 PICKUP 2WD RAM 1500 PICKUP 4WD SEBRING 4-DR STRATUS 4-DR TOWN & COUNTRY 2WD VIPER CONVERTIBLE WRANGLER/TJ 4WD

Engine Displacement (in3)

MPG (highway)

215 201 196 226 226 348 226 348 148 226 122 215 215 148 500 348 165 148 148 500 148

30.8 32.5 35.4 28.1 24.4 24.1 28.5 24.2 32.8 28 41.3 30.0 28.2 34.1 18.7 20.3 35.1 37.9 33.8 25.9 26.4

JWCL232_c11_401-448.qxd

412

1/14/10

8:02 PM

Page 412

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

11-7. An article in the Tappi Journal (March, 1986) presented data on green liquor Na2S concentration (in grams per liter) and paper machine production (in tons per day). The data (read from a graph) are shown as follows: y

40

42

49

46

44

48

x

825

830

890

895

890

910

y

46

43

53

52

54

57

58

x

915

960

990

1010

1012

1030

1050

(a) Fit a simple linear regression model with y  green liquor Na2S concentration and x  production. Find an estimate of 2. Draw a scatter diagram of the data and the resulting least squares fitted model. (b) Find the fitted value of y corresponding to x  910 and the associated residual. (c) Find the mean green liquor Na2S concentration when the production rate is 950 tons per day. 11-8. An article in the Journal of Sound and Vibration (Vol. 151, 1991, pp. 383–394) described a study investigating the relationship between noise exposure and hypertension. The following data are representative of those reported in the article. y

1

0

1

2

5

1

4

6

2

3

x

60

63

65

70

70

70

80

90

80

80

y

5

4

6

8

4

5

7

9

7

6

x

85

89

90

90

90

90

94 100 100 100

(a) Draw a scatter diagram of y (blood pressure rise in millimeters of mercury) versus x (sound pressure level in decibels). Does a simple linear regression model seem reasonable in this situation? (b) Fit the simple linear regression model using least squares. Find an estimate of 2. (c) Find the predicted mean rise in blood pressure level associated with a sound pressure level of 85 decibels. 11-9. An article in Wear (Vol. 152, 1992, pp. 171–181) presents data on the fretting wear of mild steel and oil viscosity. Representative data follow, with x  oil viscosity and y  wear volume (10 4 cubic millimeters). y

240

181

193

155

172

x

1.6

9.4

15.5

20.0

22.0

y

110

113

75

94

x

35.5

43.0

40.5

33.0

(a) Construct a scatter plot of the data. Does a simple linear regression model appear to be plausible? (b) Fit the simple linear regression model using least squares. Find an estimate of 2. (c) Predict fretting wear when viscosity x  30. (d) Obtain the fitted value of y when x  22.0 and calculate the corresponding residual. 11-10. An article in the Journal of Environmental Engineering (Vol. 115, No. 3, 1989, pp. 608–619) reported the results of a study on the occurrence of sodium and chloride in surface streams in central Rhode Island. The following data are chloride concentration y (in milligrams per liter) and roadway area in the watershed x (in percentage). y

4.4

6.6

9.7

10.6

10.8

10.9

x

0.19

0.15

0.57

0.70

0.67

0.63

y

11.8

12.1

14.3

14.7

15.0

17.3

x

0.47

0.70

0.60

0.78

0.81

0.78

y

19.2

23.1

27.4

27.7

31.8

39.5

x

0.69

1.30

1.05

1.06

1.74

1.62

(a) Draw a scatter diagram of the data. Does a simple linear regression model seem appropriate here? (b) Fit the simple linear regression model using the method of least squares. Find an estimate of 2. (c) Estimate the mean chloride concentration for a watershed that has 1% roadway area. (d) Find the fitted value corresponding to x  0.47 and the associated residual. 11-11. A rocket motor is manufactured by bonding together two types of propellants, an igniter and a sustainer. The shear strength of the bond y is thought to be a linear function of the age of the propellant x when the motor is cast. Twenty observations are shown in the following table. (a) Draw a scatter diagram of the data. Does the straight-line regression model seem to be plausible? (b) Find the least squares estimates of the slope and intercept in the simple linear regression model. Find an estimate of 2. (c) Estimate the mean shear strength of a motor made from propellant that is 20 weeks old. (d) Obtain the fitted values yˆ i that correspond to each observed value yi. Plot yˆ i versus yi and comment on what this plot would look like if the linear relationship between shear strength and age were perfectly deterministic (no error). Does this plot indicate that age is a reasonable choice of regressor variable in this model?

JWCL232_c11_401-448.qxd

1/14/10

8:02 PM

Page 413

413

11-2 SIMPLE LINEAR REGRESSION

Observation Number

Strength y (psi)

Age x (weeks)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

2158.70 1678.15 2316.00 2061.30 2207.50 1708.30 1784.70 2575.00 2357.90 2277.70 2165.20 2399.55 1779.80 2336.75 1765.30 2053.50 2414.40 2200.50 2654.20 1753.70

15.50 23.75 8.00 17.00 5.00 19.00 24.00 2.50 7.50 11.00 13.00 3.75 25.00 9.75 22.00 18.00 6.00 12.50 2.00 21.50

11-12. An article in the Journal of the American Ceramic Society [“Rapid Hot-Pressing of Ultrafine PSZ Powders” (1991, Vol. 74, pp. 1547–1553)] considered the microstructure of the ultrafine powder of partially stabilized zirconia as a function of temperature. The data are shown below: x  Temperature (C): 1100 1200 y  Porosity (%): 30.8 7.7

1200 1300 1100 1500 1300 19.2 6.0 13.5 11.4 3.6

(a) Fit the simple linear regression model using the method of least squares. Find an estimate of 2. (b) Estimate the mean porosity for a temperature of 1400C. (c) Find the fitted value corresponding to y  11.4 and the associated residual. (d) Draw a scatter diagram of the data. Does a simple linear regression model seem appropriate here? Explain. 11-13. An article in the Journal of the Environmental Engineering Division [“Least Squares Estimates of BOD Parameters” (1980, Vol. 106, pp. 1197–1202)] took a sample from the Holston River below Kingport, Tennessee, during August 1977. The biochemical oxygen demand (BOD) test is

conducted over a period of time in days. The resulting data are shown below: Time (days):

1 2 4 18 20

6

8

10

12

14

16

BOD (mg/liter): 0.6 0.7 1.5 1.9 2.1 2.6 2.9 3.7 3.5 3.7 3.8 (a) Assuming that a simple linear regression model is appropriate, fit the regression model relating BOD (y) to the time (x). What is the estimate of 2? (b) What is the estimate of expected BOD level when the time is 15 days? (c) What change in mean BOD is expected when the time changes by three days? (d) Suppose the time used is six days. Calculate the fitted value of y and the corresponding residual. (e) Calculate the fitted yˆ i for each value of xi used to fit the model. Then construct a graph of yˆ i versus the corresponding observed values yi and comment on what this plot would look like if the relationship between y and x was a deterministic (no random error) straight line. Does the plot actually obtained indicate that time is an effective regressor variable in predicting BOD? 11-14. An article in Wood Science and Technology [“Creep in Chipboard, Part 3: Initial Assessment of the Influence of Moisture Content and Level of Stressing on Rate of Creep and Time to Failure” (1981, Vol. 15, pp. 125–144)] studied the deflection (mm) of particleboard from stress levels of relative humidity. Assume that the two variables are related according to the simple linear regression model. The data are shown below: x  Stress level (%): 54 54 61 61 68 y  Deflection (mm): 16.473 18.693 14.305 15.121 13.505 x  Stress level (%): 68 75 75 75 y  Deflection (mm): 11.640 11.168 12.534 11.224 (a) Calculate the least square estimates of the slope and intercept. What is the estimate of 2? Graph the regression model and the data. (b) Find the estimate of the mean deflection if the stress level can be limited to 65%. (c) Estimate the change in the mean deflection associated with a 5% increment in stress level. (d) To decrease the mean deflection by one millimeter, how much increase in stress level must be generated? (e) Given that the stress level is 68%, find the fitted value of deflection and the corresponding residual. 11-15. In an article in Statistics and Computing [“An Iterative Monte Carlo Method for Nonconjugate Bayesian Analysis” (1991, pp. 119–128)] Carlin and Gelfand investigated the age (x) and length (y) of 27 captured dugongs (sea cows).

JWCL232_c11_401-448.qxd

414

1/14/10

8:02 PM

Page 414

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

x  1.0, 1.5, 1.5, 1.5, 2.5, 4.0, 5.0, 5.0, 7.0, 8.0, 8.5, 9.0, 9.5, 9.5, 10.0, 12.0, 12.0, 13.0, 13.0, 14.5, 15.5, 15.5, 16.5, 17.0, 22.5, 29.0, 31.5

(a) Write the new regression model. (b) What change in gasoline mileage is associated with a 1 cm3 change is engine displacement?

y  1.80, 1.85, 1.87, 1.77, 2.02, 2.27, 2.15, 2.26, 2.47, 2.19, 2.26, 2.40, 2.39, 2.41, 2.50, 2.32, 2.32, 2.43, 2.47, 2.56, 2.65, 2.47, 2.64, 2.56, 2.70, 2.72, 2.57

11-18. Show that in a simple linear regression model the point ( x, y ) lies exactly on the least squares regression line. 11-19. Consider the simple linear regression model Y  0  1x  . Suppose that the analyst wants to use z  x x as the regressor variable. (a) Using the data in Exercise 11-11, construct one scatter plot of the (xi, yi) points and then another of the (zi  xi x, yi) points. Use the two plots to intuitively explain how the two models, Y  0  1x   and Y  *0  *1 z  , are related. (b) Find the least squares estimates of *0 and *1 in the model Y  *0  *1 z  . How do they relate to the least squares estimates ˆ 0 and ˆ 1?

(a) Find the least squares estimates of the slope and the intercept in the simple linear regression model. Find an estimate of 2. (b) Estimate the mean length of dugongs at age 11. (c) Obtain the fitted values yˆ i that correspond to each observed value yi. Plot yˆ i versus yi, and comment on what this plot would look like if the linear relationship between length and age were perfectly deterministic (no error). Does this plot indicate that age is a reasonable choice of regressor variable in this model? 11-16. Consider the regression model developed in Exercise 11-2. (a) Suppose that temperature is measured in C rather than F. Write the new regression model. (b) What change in expected pavement deflection is associated with a 1C change in surface temperature? 11-17. Consider the regression model developed in Exercise 11-6. Suppose that engine displacement is measured in cubic centimeters instead of cubic inches.

11-20. Suppose we wish to fit a regression model for which the true regression line passes through the point (0, 0). The appropriate model is Y  x  . Assume that we have n pairs of data (x1, y1), (x2, y2), p , (xn, yn). (a) Find the least squares estimate of . (b) Fit the model Y  x   to the chloride concentrationroadway area data in Exercise 11-10. Plot the fitted model on a scatter diagram of the data and comment on the appropriateness of the model.

11-3 PROPERTIES OF THE LEAST SQUARES ESTIMATORS ˆ and  ˆ may be easily described. The statistical properties of the least squares estimators  0 1 Recall that we have assumed that the error term  in the model Y  0  1x   is a random variable with mean zero and variance 2. Since the values of x are fixed, Y is a random variˆ and  ˆ depend able with mean Y 0 x  0  1x and variance 2. Therefore, the values of  0 1 on the observed y’s; thus, the least squares estimators of the regression coefficients may be viewed as random variables. We will investigate the bias and variance properties of the least ˆ and  ˆ . squares estimators  0 1 ˆ ˆ is a linear combination of the observations Y , we can use Consider first 1. Because  i 1 ˆ is properties of expectation to show that the expected value of  1 ˆ 2 E1 1 1

(11-15)

ˆ is an unbiased estimator of the true slope  . Thus,  1 1 ˆ . Since we have assumed that V( )  2, it follows that Now consider the variance of  i 1 ˆ is a linear combination of the observations Y , the results in V(Yi)  2. Because  i 1 Section 5-5 can be applied to show that ˆ 2 V1 1

2 Sxx

(11-16)

JWCL232_c11_401-448.qxd

1/14/10

8:02 PM

Page 415

415

11-4 HYPOTHESIS TESTS IN SIMPLE LINEAR REGRESSION

For the intercept, we can show in a similar manner that ˆ 2 E1 0 0

2 ˆ 2  2 c 1  x d and V1 0 n Sxx

(11-17)

ˆ is an unbiased estimator of the intercept  . The covariance of the random variThus,  0 0 ˆ ˆ is not zero. It can be shown (see Exercise 11-98) that cov( ˆ , ˆ ) ables 0 and  1 0 1 2  x Sxx. The estimate of 2 could be used in Equations 11-16 and 11-17 to provide estimates of the variance of the slope and the intercept. We call the square roots of the resulting variance estimators the estimated standard errors of the slope and intercept, respectively.

Estimated Standard Errors

In simple linear regression the estimated standard error of the slope and the estimated standard error of the intercept are ˆ 2 se1 1

ˆ 2 B Sxx

and

ˆ 2 se1 0

x2 1 ˆ 2 c n  d Sxx B

ˆ 2 is computed from Equation 11-13. respectively, where 

The Minitab computer output in Table 11-2 reports the estimated standard errors of the slope and intercept under the column heading “SE coeff.”

11-4 HYPOTHESIS TESTS IN SIMPLE LINEAR REGRESSION An important part of assessing the adequacy of a linear regression model is testing statistical hypotheses about the model parameters and constructing certain confidence intervals. Hypothesis testing in simple linear regression is discussed in this section, and Section 11-5 presents methods for constructing confidence intervals. To test hypotheses about the slope and intercept of the regression model, we must make the additional assumption that the error component in the model, , is normally distributed. Thus, the complete assumptions are that the errors are normally and independently distributed with mean zero and variance 2, abbreviated NID(0, 2).

11-4.1 Use of t-Tests Suppose we wish to test the hypothesis that the slope equals a constant, say, 1,0. The appropriate hypotheses are H0: 1  1,0 H1: 1 1,0

(11-18)

where we have assumed a two-sided alternative. Since the errors i are NID(0, 2), it follows directly that the observations Yi are NID(0  1xi, 2). Now ˆ 1 is a linear combination of

JWCL232_c11_401-448.qxd

416

1/14/10

8:02 PM

Page 416

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

independent normal random variables, and consequently, ˆ 1 is N(1, 2Sxx), using the bias ˆ 2 2 has and variance properties of the slope discussed in Section 11-3. In addition, 1n 22 ˆ ˆ 2 . As a a chi-square distribution with n 2 degrees of freedom, and 1 is independent of  result of those properties, the statistic T0 

Test Statistic

ˆ   1 1,0

(11-19)

2ˆ 2 Sxx

follows the t distribution with n 2 degrees of freedom under H0: 1  1,0. We would reject H0: 1  1,0 if 0 t0 0 t 2,n 2

(11-20)

where t0 is computed from Equation 11-19. The denominator of Equation 11-19 is the standard error of the slope, so we could write the test statistic as T0 

ˆ   1 1,0 ˆ 2 se1 1

A similar procedure can be used to test hypotheses about the intercept. To test H0: 0  0,0 H1: 0 0,0

(11-21)

we would use the statistic

Test Statistic

T0 

ˆ   0 0,0

x2 1  cn  d Sxx B



ˆ2

ˆ   0 0,0 ˆ 2 se1 0

(11-22)

and reject the null hypothesis if the computed value of this test statistic, t0, is such that 0 t0 0 t 2,n 2. Note that the denominator of the test statistic in Equation 11-22 is just the standard error of the intercept. A very important special case of the hypotheses of Equation 11-18 is H0: 1  0 H1: 1 0

(11-23)

These hypotheses relate to the significance of regression. Failure to reject H0: 1  0 is equivalent to concluding that there is no linear relationship between x and Y. This situation is illustrated in Fig. 11-5. Note that this may imply either that x is of little value in explaining the variation in Y and that the best estimator of Y for any x is yˆ  Y [Fig. 11-5(a)] or that the true relationship between x and Y is not linear [Fig. 11-5(b)]. Alternatively, if H0: 1  0 is rejected, this implies that x is of value in explaining the variability in Y (see Fig. 11-6). Rejecting H0: 1  0 could mean either that the straight-line model is adequate [Fig. 11-6(a)] or that, although there is a linear effect of x, better results could be obtained with the addition of higher order polynomial terms in x [Fig. 11-6(b)].

JWCL232_c11_401-448.qxd

1/15/10

5:37 PM

Page 417

11-4 HYPOTHESIS TESTS IN SIMPLE LINEAR REGRESSION y

y

Figure 11-5 The hypothesis H0: 1  0 is not rejected.

x

(a)

EXAMPLE 11-2

H0: 1  0 H1: 1  0 and we will use   0.01. From Example 11-1 and Table 11-2 we have ˆ  14.947 n  20, S  0.68088,  ˆ 2  1.18  1 xx

so the t-statistic in Equation 10-20 becomes ˆ  1

2 Sxx ˆ2



x

(b)

Oxygen Purity Tests of Coefficients

We will test for significance of regression using the model for the oxygen purity data from Example 11-1. The hypotheses are

t0 

417

ˆ  1

ˆ 2 se1 1



14.947 21.18 0.68088

 11.35

Practical Interpretation: Since the reference value of t is t0.005,18  2.88, the value of the test statistic is very far into the critical region, implying that H0: 1  0 should be rejected. There is strong evidence to support this claim. The P-value for this test is P ⯝ 1.23  10 9. This was obtained manually with a calculator. Table 11-2 presents the Minitab output for this problem. Notice that the t-statistic value for the slope is computed as 11.35 and that the reported P-value is P  0.000. Minitab also reports the t-statistic for testing the hypothesis H0: 0  0. This statistic is computed from Equation 11-22, with 0,0  0, as t0  46.62. Clearly, then, the hypothesis that the intercept is zero is rejected.

11-4.2 Analysis of Variance Approach to Test Significance of Regression A method called the analysis of variance can be used to test for significance of regression. The procedure partitions the total variability in the response variable into meaningful components as the basis for the test. The analysis of variance identity is as follows: Analysis of Variance Identity

n

n

n

i1

i1

i1

2 2 2 a 1 yi y 2  a 1 yˆ i y 2 a 1 yi yˆi 2

y

Figure 11-6 The hypothesis H0: 1  0 is rejected.

y

x (a)

x (b)

(11-24)

JWCL232_c11_401-448.qxd

418

1/14/10

8:02 PM

Page 418

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

The two components on the right-hand-side of Equation 11-24 measure, respectively, the amount of variability in yi accounted for by the regression line and the residual variation left n unexplained by the regression line. We usually call SSE  g i1 1yi yˆ i 2 2 the error sum of n squares and SSR  g i1 1 yˆi y 2 2 the regression sum of squares. Symbolically, Equation 11-24 may be written as

SST  SSR  SSE

(11-25)

where SST  g i1 1 yi y2 2 is the total corrected sum of squares of y. In Section 11-2 we ˆ S (see Equation 11-14), so since SS   ˆ S  SS , we note that the noted that SSE  SST  1 xy T 1 xy E ˆ S . The total sum of squares SS has regression sum of squares in Equation 11-25 is SSR   1 xy T n 1 degrees of freedom, and SSR and SSE have 1 and n 2 degrees of freedom, respectively. We may show that E3SSE 1n 22 4  2, E1SSR 2  2  21Sx x and that SSE 2 and SSR 2 are independent chi-square random variables with n 2 and 1 degrees of freedom, respectively. Thus, if the null hypothesis H0: 1  0 is true, the statistic n

Test for Significance of Regression

F0 

SSR 1 MSR  MSE SSE 1n 22

(11-26)

follows the F1,n 2 distribution, and we would reject H0 if f0 f,1,n 2. The quantities MSR  SSR1 and MSE  SSE(n 2) are called mean squares. In general, a mean square is always computed by dividing a sum of squares by its number of degrees of freedom. The test procedure is usually arranged in an analysis of variance table, such as Table 11-3. EXAMPLE 11-3

Oxygen Purity ANOVA

We will use the analysis of variance approach to test for significance of regression using the oxygen purity data model from ˆ  14.947, S  Example 11-1. Recall that SST  173.38,  xy 1 10.17744, and n  20. The regression sum of squares is ˆ S  114.947210.17744  152.13 SSR   1 xy

and the error sum of squares is SSE  SST SSR  173.38 152.13  21.25

The analysis of variance for testing H0: 1  0 is summarized in the Minitab output in Table 11-2. The test statistic is f0  MSRMSE  152.131.18  128.86, for which we find that the P-value is P  1.23  10 9, so we conclude that 1 is not zero. There are frequently minor differences in terminology among computer packages. For example, sometimes the regression sum of squares is called the “model” sum of squares, and the error sum of squares is called the “residual” sum of squares.

Table 11-3 Analysis of Variance for Testing Significance of Regression Source of Variation Regression Error Total ˆ 2. Note that MSE  

Sum of Squares SSR  ˆ 1Sx y SSE  SST ˆ 1Sxy SST

Degrees of Freedom

Mean Square

1 n 2 n 1

MSR MSE

F0 MSRMSE

JWCL232_c11_401-448.qxd

1/14/10

8:02 PM

Page 419

11-4 HYPOTHESIS TESTS IN SIMPLE LINEAR REGRESSION

419

Note that the analysis of variance procedure for testing for significance of regression is equivalent to the t-test in Section 11-4.1. That is, either procedure will lead to the same conclusions. This is easy to demonstrate by starting with the t-test statistic in Equation 11-19 with 1,0  0, say T0 

ˆ  1

2ˆ 2 Sx x Squaring both sides of Equation 11-27 and using the fact that ˆ 2  MSE results in T 20 

ˆ 2S ˆ S   MSR 1 xx 1 xy   MSE MSE MSE

(11-27)

(11-28)

Note that T 02 in Equation 11-28 is identical to F0 in Equation 11-26. It is true, in general, that the square of a t random variable with v degrees of freedom is an F random variable, with one and v degrees of freedom in the numerator and denominator, respectively. Thus, the test using T0 is equivalent to the test based on F0. Note, however, that the t-test is somewhat more flexible in that it would allow testing against a one-sided alternative hypothesis, while the F-test is restricted to a two-sided alternative. EXERCISES FOR SECTION 11-4 11-21.

Consider the computer output below.

The regression equation is Y  12.9  2.34 x Predictor Constant X S  1.48111

Coef 12.857 2.3445

SE Coef 1.032 0.1150

R Sq  98.1%

T ? ?

P ? ?

R Sq(adj)  97.9%

Analysis of Variance Source Regression Residual Error Total

DF 1 8 9

SS 912.43 17.55 929.98

MS 912.43 ?

F ?

P ?

(a) Fill in the missing information. You may use bounds for the P-values. (b) Can you conclude that the model defines a useful linear relationship? (c) What is your estimate of 2? 11-22. Consider the computer output below. The regression equation is Y = 26.8  1.48 x Predictor Constant X

Coef 26.753 1.4756

SE Coef 2.373 0.1063

S  2.70040

R Sq  93.7%

T ? ?

P ? ?

R-Sq (adj)  93.2%

Analysis of Variance Source Regression Residual Error Total

DF 1 ? 15

SS ? 94.8 1500.0

MS ? 7.3

F ?

P ?

(a) Fill in the missing information. You may use bounds for the P-values. (b) Can you conclude that the model defines a useful linear relationship? (c) What is your estimate of 2? 11-23. Consider the data from Exercise 11-1 on x  compressive strength and y  intrinsic permeability of concrete. (a) Test for significance of regression using   0.05. Find the P-value for this test. Can you conclude that the model specifies a useful linear relationship between these two variables? (b) Estimate 2 and the standard deviation of ˆ 1. (c) What is the standard error of the intercept in this model? 11-24. Consider the data from Exercise 11-2 on x  roadway surface temperature and y  pavement deflection. (a) Test for significance of regression using   0.05. Find the P-value for this test. What conclusions can you draw? (b) Estimate the standard errors of the slope and intercept. 11-25. Consider the National Football League data in Exercise 11-3. (a) Test for significance of regression using   0.01. Find the P-value for this test. What conclusions can you draw? (b) Estimate the standard errors of the slope and intercept. (c) Test H0: 1  10 versus H1: 1 10 with   0.01. Would you agree with the statement that this is a test of the hypothesis that a one-yard increase in the average yards per attempt results in a mean increase of 10 rating points? 11-26. Consider the data from Exercise 11-4 on y  sales price and x  taxes paid. (a) Test H0: 1  0 using the t-test; use   0.05. (b) Test H0: 1  0 using the analysis of variance with   0.05. Discuss the relationship of this test to the test from part (a).

JWCL232_c11_401-448.qxd

420

1/14/10

8:02 PM

Page 420

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

(c) Estimate the standard errors of the slope and intercept. (d) Test the hypothesis that 0  0. 11-27. Consider the data from Exercise 11-5 on y  steam usage and x  average temperature. (a) Test for significance of regression using   0.01. What is the P-value for this test? State the conclusions that result from this test. (b) Estimate the standard errors of the slope and intercept. (c) Test the hypothesis H0: 1  10 versus H1: 1  10 using   0.01. Find the P-value for this test. (d) Test H0: 0  0 versus H1: 0  0 using   0.01. Find the P-value for this test and draw conclusions. 11-28. Consider the data from Exercise 11-6 on y  highway gasoline mileage and x  engine displacement. (a) Test for significance of regression using   0.01. Find the P-value for this test. What conclusions can you reach? (b) Estimate the standard errors of the slope and intercept. (c) Test H0: 1  0.05 versus H1: 1  0.05 using   0.01 and draw conclusions. What is the P-value for this test? (d) Test the hypothesis H0: 0  0 versus H1: 0  0 using   0.01. What is the P-value for this test? 11-29. Consider the data from Exercise 11-7 on y  green liquor Na2S concentration and x  production in a paper mill. (a) Test for significance of regression using   0.05. Find the P-value for this test. (b) Estimate the standard errors of the slope and intercept. (c) Test H0: 0  0 versus H1: 0  0 using   0.05. What is the P-value for this test? 11-30. Consider the data from Exercise 11-8 on y  blood pressure rise and x  sound pressure level. (a) Test for significance of regression using   0.05. What is the P-value for this test? (b) Estimate the standard errors of the slope and intercept. (c) Test H0: 0  0 versus H1: 0  0 using   0.05. Find the P-value for this test. 11-31. Consider the data from Exercise 11-11, on y  shear strength of a propellant and x  propellant age. (a) Test for significance of regression with   0.01. Find the P-value for this test. (b) Estimate the standard errors of ˆ 0 and ˆ 1. (c) Test H0: 1  30 versus H1: 1  30 using   0.01. What is the P-value for this test? (d) Test H0: 0  0 versus H1: 0  0 using   0.01. What is the P-value for this test? (e) Test H0: 0  2500 versus H1: 0 2500 using   0.01. What is the P-value for this test? 11-32. Consider the data from Exercise 11-10 on y  chloride concentration in surface streams and x  roadway area. (a) Test the hypothesis H0: 1  0 versus H1: 1  0 using the analysis of variance procedure with   0.01. (b) Find the P-value for the test in part (a). (c) Estimate the standard errors of ˆ 1 and ˆ 0.

(d) Test H0: 0  0 versus H1: 0  0 using   0.01. What conclusions can you draw? Does it seem that the model might be a better fit to the data if the intercept were removed? 11-33. Consider the data in Exercise 11-13 on y  oxygen demand and x  time. (a) Test for significance of regression using   0.01. Find the P-value for this test. What conclusions can you draw? (b) Estimate the standard errors of the slope and intercept. (c) Test the hypothesis that 0  0. 11-34. Consider the data in Exercise 11-14 on y  deflection and x  stress level. (a) Test for significance of regression using   0.01. What is the P-value for this test? State the conclusions that result from this test. (b) Does this model appear to be adequate? (c) Estimate the standard errors of the slope and intercept. 11-35. An article in The Journal of Clinical Endocrinology and Metabolism [“Simultaneous and Continuous 24-Hour Plasma and Cerebrospinal Fluid Leptin Measurements: Dissociation of Concentrations in Central and Peripheral Compartments” (2004, Vol. 89, pp. 258–265)] studied the demographics of simultaneous and continuous 24-hour plasma and cerebrospinal fluid leptin measurements. The data follow: y  BMI (kg/m2): 19.92 20.39 x  Age (yr):

20.59 23.29

29.02 17.27

45.5 34.6 40.6 52.1 33.3 47.0

32.9

20.78 35.24 28.2

25.97 30.1

(a) Test for significance of regression using   0.05. Find the P-value for this test. Can you conclude that the model specifies a useful linear relationship between these two variables? (b) Estimate 2 and the standard deviation of ˆ 1. (c) What is the standard error of the intercept in this model? 11-36. Suppose that each value of xi is multiplied by a positive constant a, and each value of yi is multiplied by another positive constant b. Show that the t-statistic for testing H0: 1  0 versus H1: 1  0 is unchanged in value. 11-37. The type II error probability for the t-test for H0: 1  1,0 can be computed in a similar manner to the t-tests of Chapter 9. If the true value of 1 is 1œ , the value d  01,0 ¿1 0 111n 12 Sxx is calculated and used as the horizontal scale factor on the operating characteristic curves for the t-test (Appendix Charts VIIe through VIIh) and the type II error probability is read from the vertical scale using the curve for n 2 degrees of freedom. Apply this procedure to the football data of Exercise 11-3, using   5.5 and 1œ  12.5, where the hypotheses are H0: 1  10 versus H1: 1 10. 11-38. Consider the no-intercept model Y  x   with the ’s NID(0, 2). The estimate of 2 is s2  n n g i1 1 yi ˆ xi 2 2 1n 12 and V 1ˆ 2  2g i1x 2i . (a) Devise a test statistic for H0:   0 versus H1:   0. (b) Apply the test in (a) to the model from Exercise 11-20.

JWCL232_c11_401-448.qxd

1/15/10

4:53 PM

Page 421

11-5 CONFIDENCE INTERVALS

421

11-5 CONFIDENCE INTERVALS 11-5.1 Confidence Intervals on the Slope and Intercept In addition to point estimates of the slope and intercept, it is possible to obtain confidence interval estimates of these parameters. The width of these confidence intervals is a measure of the overall quality of the regression line. If the error terms, i, in the regression model are normally and independently distributed, ˆ  2 2 ˆ 2 Sx x and 1 1 1

ˆ  2

1 0 0

1 x2 ˆ 2 c n  d Sx x B

are both distributed as t random variables with n 2 degrees of freedom. This leads to the following definition of 100(1 )% confidence intervals on the slope and intercept. Confidence Intervals on Parameters

Under the assumption that the observations are normally and independently distributed, a 100(1 )% confidence interval on the slope 1 in simple linear regression is ˆ 2 ˆ 2 ˆ t ˆ t   1   1  2, n 2 1  2, n 2 BS BS xx

(11-29)

xx

Similarly, a 100(1 )% confidence interval on the intercept 0 is x2 1 ˆ t  ˆ 2 c n  d 0  2, n 2 Sx x B 1 x2 ˆ t  0   ˆ 2 c n  d 0  2, n 2 Sx x B EXAMPLE 11-4

Oxygen Purity Confidence Interval on the Slope

We will find a 95% confidence interval on the slope of the regression line using the data in Example 11-1. Recall that ˆ  14.947, S  0.68088, and  ˆ 2  1.18 (see Table 11-2).  xx 1 Then, from Equation 11-29 we find ˆ t  1 0.025,18

(11-30)

ˆ 2 ˆ 2 ˆ t  1   1 0.025,18 B Sxx B Sxx

or 14.947 2.101  2.101

This simplifies to 12.181  1  17.713 Practical Interpretation: This CI does not include zero, so there is strong evidence (at   0.05) that the slope is not zero. The CI is reasonably narrow (2.766) because the error variance is fairly small.

1.18  1  14.947 A 0.68088 1.18 A 0.68088

11-5.2 Confidence Interval on the Mean Response A confidence interval may be constructed on the mean response at a specified value of x, say, x0. This is a confidence interval about E(Y x0)  Y  x0 and is often called a confidence interval

JWCL232_c11_401-448.qxd

422

1/14/10

8:03 PM

Page 422

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

about the regression line. Since E(Y x0)  Y  x0  0  1x0, we may obtain a point estimate of the mean of Y at x  x0 (Y  x0) from the fitted model as ˆ  ˆ x ˆ Y 0 x0    0 1 0 ˆ and  ˆ are unbiased estimators of ˆ Y 0 x0 is an unbiased point estimator of Y  x0, since  Now  0 1 0 and 1. The variance of ˆ Y 0 x0 is

1x0 x2 2 1 ˆ Y 0 x0 2  2 c  V 1 d n Sx x

ˆ 1x x2 and cov 1Y,  ˆ 2  0. The zero ˆ Y | x0  y   This last result follows from the fact that  1 0 1 ˆ covariance result is left as a mind-expanding exercise. Also, Y 0 x0 is normally distributed, because ˆ and  ˆ are normally distributed, and if we ˆ 2 use as an estimate of 2, it is easy to show that  1 0 ˆ Y 0 x0 Y 0 x0 

1x0 x 2 2 1 ˆ 2 c n  d Sx x B has a t distribution with n 2 degrees of freedom. This leads to the following confidence interval definition. Confidence Interval on the Mean Response

A 100(1 )% confidence interval about the mean response at the value of x  x0, say Y 0 x0, is given by 1x0 x 2 2 1 ˆ Y 0 x0 t 2, n 2  ˆ2 c   d n Sx x B

1x0 x 2 2 1 ˆ Y 0 x0  t 2, n 2  ˆ2 c   Y 0 x0   d n Sx x B

(11-31)

ˆ  ˆ x is computed from the fitted regression model. ˆ Y 0 x0   where  0 1 0

Note that the width of the CI for Y 0 x0 is a function of the value specified for x0. The interval width is a minimum for x0  x and widens as 0 x0 x 0 increases. EXAMPLE 11-5

Oxygen Purity Confidence Interval on the Mean Response

We will construct a 95% confidence interval about the mean response for the data in Example 11-1. The fitted model is ˆ Y 0 x0  74.283  14.947x0, and the 95% confidence interval  on Y 0 x0 is found from Equation 11-31 as ˆ Y 0 x0  2.101 

B

1.18 c

1x0 1.19602 2 1  d 20 0.68088

Suppose that we are interested in predicting mean oxygen purity when x0  1.00%. Then ˆ Y 0 x1.00  74.283  14.94711.002  89.23 

and the 95% confidence interval is 89.23  2.101

B

1.18 c

11.00 1.19602 2 1  d 20 0.68088

or 89.23  0.75 Therefore, the 95% CI on Y 0 1.00 is 88.48  Y 0 1.00  89.98

JWCL232_c11_401-448.qxd

1/14/10

8:03 PM

Page 423

11-6 PREDICTION OF NEW OBSERVATIONS

423

102

Figure 11-7 Scatter diagram of oxygen purity data from Example 11-1 with fitted regression line and 95 percent confidence limits on Y 0 x0 .

Oxygen purity y (%)

99

96

93

90

87

0.87

1.07

1.27

1.47

1.67

Hydrocarbon level (%) x

This is a reasonable narrow CI. Minitab will also perform these calculations. Refer to Table 11-2. The predicted value of y at x  1.00 is shown along with the 95% CI on the mean of y at this level of x. By repeating these calculations for several different values for x0, we can obtain confidence limits for each corresponding value of Y 0 x0. Figure 11-7 displays the scatter diagram

with the fitted model and the corresponding 95% confidence limits plotted as the upper and lower lines. The 95% confidence level applies only to the interval obtained at one value of x and not to the entire set of x-levels. Notice that the width of the confidence interval on Y 0 x0 increases as 0 x0 x 0 increases.

11-6 PREDICTION OF NEW OBSERVATIONS An important application of a regression model is predicting new or future observations Y corresponding to a specified level of the regressor variable x. If x0 is the value of the regressor variable of interest, Yˆ0  ˆ 0  ˆ 1x0

(11-32)

is the point estimator of the new or future value of the response Y0. Now consider obtaining an interval estimate for this future observation Y0. This new observation is independent of the observations used to develop the regression model. Therefore, the confidence interval for Y 0 x0 in Equation 11-31 is inappropriate, since it is based only on the data used to fit the regression model. The confidence interval about Y 0 x0 refers to the true mean response at x  x0 (that is, a population parameter), not to future observations. Let Y0 be the future observation at x  x0, and let Yˆ0 given by Equation 11-32 be the estimator of Y0. Note that the error in prediction epˆ  Y0 Yˆ 0 is a normally distributed random variable with mean zero and variance 1x0 x 2 2 1 d V 1epˆ 2  V1Y0 Yˆ 0 2  2 c 1  n  Sx x

JWCL232_c11_401-448.qxd

424

1/14/10

8:03 PM

Page 424

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

because Y0 is independent of Yˆ0. If we use ˆ 2 to estimate 2, we can show that Y0 Yˆ 0

1x0 x 2 2 1 ˆ 2 c 1  n  d Sx x B has a t distribution with n 2 degrees of freedom. From this we can develop the following prediction interval definition.

Prediction Interval

A 100(1 ) % prediction interval on a future observation Y0 at the value x0 is given by 1x0 x 2 2 1 yˆ 0 t 2, n 2 ˆ 2 c 1  n  d Sx x B

1x0 x 2 2 1  Y0  yˆ0  t 2, n 2 ˆ 2 c 1  n  d Sx x B

(11-33)

ˆ  ˆ x. The value yˆ 0 is computed from the regression model yˆ0   0 1 0

Notice that the prediction interval is of minimum width at x0  x and widens as 0 x0 x 0 increases. By comparing Equation 11-33 with Equation 11-31, we observe that the prediction interval at the point x0 is always wider than the confidence interval at x0. This results because the prediction interval depends on both the error from the fitted model and the error associated with future observations.

EXAMPLE 11-6

Oxygen Purity Prediction Interval

To illustrate the construction of a prediction interval, suppose we use the data in Example 11-1 and find a 95% prediction interval on the next observation of oxygen purity at x0  1.00%. Using Equation 11-33 and recalling from Example 11-5 that yˆ 0  89.23, we find that the prediction interval is 89.23 2.101

B

1.18 c 1 

 Y0  89.23  2.101

11.00 1.19602 2 1  d 20 0.68088

11.00 1.19602 2 1 d  1.18 c 1  B 20 0.68088

which simplifies to 86.83  y0  91.63

This is a reasonably narrow prediction interval. Minitab will also calculate prediction intervals. Refer to the output in Table 11-2. The 95% PI on the future observation at x0  1.00 is shown in the display. By repeating the foregoing calculations at different levels of x0, we may obtain the 95% prediction intervals shown graphically as the lower and upper lines about the fitted regression model in Fig. 11-8. Notice that this graph also shows the 95% confidence limits on Y 0 x0 calculated in Example 11-5. It illustrates that the prediction limits are always wider than the confidence limits.

JWCL232_c11_401-448.qxd

1/18/10

2:12 PM

Page 425

11-6 PREDICTION OF NEW OBSERVATIONS

425

102

Figure 11-8 Scatter diagram of oxygen purity data from Example 11-1 with fitted regression line, 95% prediction limits (outer lines) and 95% confidence limits on ␮Y 0 x0.

Oxygen purity y (%)

99

96

93

90

87

0.87

1.07

1.27 Hydrocarbon level (%) x

1.47

1.67

EXERCISES FOR SECTIONS 11-5 AND 11-6 11-39. Refer to the data in Exercise 11-1 on y ⫽ intrinsic permeability of concrete and x ⫽ compressive strength. Find a 95% confidence interval on each of the following: (a) Slope (b) Intercept (c) Mean permeability when x ⫽ 2.5 (d) Find a 95% prediction interval on permeability when x ⫽ 2.5. Explain why this interval is wider than the interval in part (c). 11-40. Exercise 11-2 presented data on roadway surface temperature x and pavement deflection y. Find a 99% confidence interval on each of the following: (a) Slope (b) Intercept (c) Mean deflection when temperature x ⫽ 85⬚F (d) Find a 99% prediction interval on pavement deflection when the temperature is 90⬚F. 11-41. Refer to the NFL quarterback ratings data in Exercise 11-3. Find a 95% confidence interval on each of the following: (a) Slope (b) Intercept (c) Mean rating when the average yards per attempt is 8.0 (d) Find a 95% prediction interval on the rating when the average yards per attempt is 8.0. 11-42. Refer to the data on y ⫽ house selling price and x ⫽ taxes paid in Exercise 11-4. Find a 95% confidence interval on each of the following: (a) ␤1 (b) ␤0 (c) Mean selling price when the taxes paid are x ⫽ 7.50 (d) Compute the 95% prediction interval for selling price when the taxes paid are x ⫽ 7.50. 11-43. Exercise 11-5 presented data on y ⫽ steam usage and x ⫽ monthly average temperature.

(a) Find a 99% confidence interval for ␤1. (b) Find a 99% confidence interval for ␤0. (c) Find a 95% confidence interval on mean steam usage when the average temperature is 55⬚F. (d) Find a 95% prediction interval on steam usage when temperature is 55⬚F. Explain why this interval is wider than the interval in part (c). 11-44. Exercise 11-6 presented gasoline mileage performance for 21 cars, along with information about the engine displacement. Find a 95% confidence interval on each of the following: (a) Slope (b) Intercept (c) Mean highway gasoline mileage when the engine displacement is x ⫽ 150 in3 (d) Construct a 95% prediction interval on highway gasoline mileage when the engine displacement is x ⫽ 150 in3. 11-45. Consider the data in Exercise 11-7 on y ⫽ green liquor Na2S concentration and x ⫽ production in a paper mill. Find a 99% confidence interval on each of the following: (a) ␤1 (b) ␤0 (c) Mean Na2S concentration when production x ⫽ 910 tons Ⲑday (d) Find a 99% prediction interval on Na2S concentration when x ⫽ 910 tonsⲐday. 11-46. Exercise 11-8 presented data on y ⫽ blood pressure rise and x ⫽ sound pressure level. Find a 95% confidence interval on each of the following: (a) ␤1 (b) ␤0 (c) Mean blood pressure rise when the sound pressure level is 85 decibels (d) Find a 95% prediction interval on blood pressure rise when the sound pressure level is 85 decibels.

JWCL232_c11_401-448.qxd

426

1/14/10

8:03 PM

Page 426

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

11-47. Refer to the data in Exercise 11-9 on y  wear volume of mild steel and x  oil viscosity. Find a 95% confidence interval on each of the following: (a) Intercept (b) Slope (c) Mean wear when oil viscosity x  30 11-48. Exercise 11-10 presented data on chloride concentration y and roadway area x on watersheds in central Rhode Island. Find a 99% confidence interval on each of the following: (a) 1 (b) 0 (c) Mean chloride concentration when roadway area x  1.0% (d) Find a 99% prediction interval on chloride concentration when roadway area x  1.0%. 11-49. Refer to the data in Exercise 11-11 on rocket motor shear strength y and propellant age x. Find a 95% confidence interval on each of the following: (a) Slope 1 (b) Intercept 0 (c) Mean shear strength when age x  20 weeks

(d) Find a 95% prediction interval on shear strength when age x  20 weeks. 11-50. Refer to the data in Exercise 11-12 on the microstructure of zirconia. Find a 95% confidence interval on each of the following: (a) Slope (b) Intercept (c) Mean length when x  1500 (d) Find a 95% prediction interval on length when x  1500. Explain why this interval is wider than the interval in part (c). 11-51. Refer to the data in Exercise 11-13 on oxygen demand. Find a 99% confidence interval on each of the following: (a) 1 (b) 0 (c) Find a 95% confidence interval on mean BOD when the time is 8 days.

11-7 ADEQUACY OF THE REGRESSION MODEL Fitting a regression model requires several assumptions. Estimation of the model parameters requires the assumption that the errors are uncorrelated random variables with mean zero and constant variance. Tests of hypotheses and interval estimation require that the errors be normally distributed. In addition, we assume that the order of the model is correct; that is, if we fit a simple linear regression model, we are assuming that the phenomenon actually behaves in a linear or first-order manner. The analyst should always consider the validity of these assumptions to be doubtful and conduct analyses to examine the adequacy of the model that has been tentatively entertained. In this section we discuss methods useful in this respect.

11-7.1 Residual Analysis The residuals from a regression model are ei  yi yˆ i, i  1, 2, p , n, where yi is an actual observation and yˆi is the corresponding fitted value from the regression model. Analysis of the residuals is frequently helpful in checking the assumption that the errors are approximately normally distributed with constant variance, and in determining whether additional terms in the model would be useful. As an approximate check of normality, the experimenter can construct a frequency histogram of the residuals or a normal probability plot of residuals. Many computer programs will produce a normal probability plot of residuals, and since the sample sizes in regression are often too small for a histogram to be meaningful, the normal probability plotting method is preferred. It requires judgment to assess the abnormality of such plots. (Refer to the discussion of the “fat pencil” method in Section 6-6). We may also standardize the residuals by computing di  ei 2ˆ 2, i  1, 2, p , n. If the errors are normally distributed, approximately 95% of the standardized residuals should fall in the interval ( 2, 2). Residuals that are far outside this interval may indicate the presence of an outlier, that is, an observation that is not typical of the rest of the data. Various rules have been proposed for discarding outliers. However, outliers sometimes provide

JWCL232_c11_401-448.qxd

1/14/10

8:03 PM

Page 427

11-7 ADEQUACY OF THE REGRESSION MODEL ei

427

ei

0

0

(a) ei

(b) ei

0

0

(c)

(d)

Figure 11-9 Patterns for residual plots. (a) Satisfactory, (b) Funnel, (c) Double bow, (d) Nonlinear. [Adapted from Montgomery, Peck, and Vining (2006).]

important information about unusual circumstances of interest to experimenters and should not be automatically discarded. For further discussion of outliers, see Montgomery, Peck, and Vining (2006). It is frequently helpful to plot the residuals (1) in time sequence (if known), (2), against the yˆ i, and (3) against the independent variable x. These graphs will usually look like one of the four general patterns shown in Fig. 11-9. Pattern (a) in Fig. 11-9 represents the ideal situation, while patterns (b), (c), and (d ) represent anomalies. If the residuals appear as in (b), the variance of the observations may be increasing with time or with the magnitude of yi or xi. Data transformation on the response y is often used to eliminate this problem. Widely used variance-stabilizing transformations include the use of 1y, ln y, or 1y as the response. See Montgomery, Peck, and Vining (2006) for more details regarding methods for selecting an appropriate transformation. Plots of residuals against yˆ i and xi that look like (c) also indicate inequality of variance. Residual plots that look like (d) indicate model inadequacy; that is, higher order terms should be added to the model, a transformation on the x-variable or the y-variable (or both) should be considered, or other regressors should be considered.

EXAMPLE 11-7

Oxygen Purity Residuals

The regression model for the oxygen purity data in Example 11-1 is yˆ  74.283  14.947x. Table 11-4 presents the observed and predicted values of y at each value of x from this data set, along with the corresponding residual. These values were computed using Minitab and show the number of decimal places typical of computer output. A normal probability

plot of the residuals is shown in Fig. 11-10. Since the residuals fall approximately along a straight line in the figure, we conclude that there is no severe departure from normality. The residuals are also plotted against the predicted value yˆi in Fig. 11-11 and against the hydrocarbon levels xi in Fig. 11-12. These plots do not indicate any serious model inadequacies.

JWCL232_c11_401-448.qxd

428

1/14/10

8:03 PM

Page 428

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

Table 11-4 Oxygen Purity Data from Example 11-1, Predicted Values, and Residuals Hydrocarbon Level, x

Oxygen Purity, y

Predicted Value, yˆ

0.99 1.02 1.15 1.29 1.46 1.36 0.87 1.23 1.55 1.40

90.01 89.05 91.43 93.74 96.73 94.45 87.59 91.77 99.42 93.65

89.081 89.530 91.473 93.566 96.107 94.612 87.288 92.669 97.452 95.210

1 2 3 4 5 6 7 8 9 10

Residual e  y yˆ 0.929 0.480 0.043 0.174 0.623 0.162 0.302 0.899 1.968 1.560

Hydrocarbon Level, x

Oxygen Purity, y

Predicted Value, yˆ

1.19 1.15 0.98 1.01 1.11 1.20 1.26 1.32 1.43 0.95

93.54 92.52 90.56 89.54 89.85 90.39 93.25 93.41 94.98 87.33

92.071 91.473 88.932 89.380 90.875 92.220 93.117 94.014 95.658 88.483

11 12 13 14 15 16 17 18 19 20

Residual e  y yˆ 1.469 1.047 1.628 0.160 1.025 1.830 0.133 0.604 0.678 1.153

11-7.2 Coefficient of Determination (R2) A widely used measure for a regression model is the following ratio of sum of squares. R2

The coefficient of determination is R2 

SSR SSE 1 SST SST

(11-34)

The coefficient is often used to judge the adequacy of a regression model. Subsequently, we will see that in the case where X and Y are jointly distributed random variables, R2 is the square of the correlation coefficient between X and Y. From the analysis of variance identity in Equations 11-24 and 11-25, 0  R2  1. We often refer loosely to R2 as the amount of variability in the data explained or accounted for by the regression model. For the oxygen purity regression model, we have R2  SSR SST  152.13 173.38  0.877; that is, the model accounts for 87.7% of the variability in the data. 99.9 2.5

95

2 1.5

80

1

50

Residuals

Cumulative normal probability

99

20

0.5 0 – 0.5

5

–1 – 1.5

1

–2 – 2.5 87

0.1 –1.9

–0.9

0.1

1.1

Residuals

Figure 11-10 Normal probability plot of residuals, Example 11-7.

2.1

89

91

93

95

97

99

Predicted values, ^ y

Figure 11-11 Plot of residuals versus predicted oxygen purity yˆ , Example 11-7.

JWCL232_c11_401-448.qxd

1/14/10

8:03 PM

Page 429

11-7 ADEQUACY OF THE REGRESSION MODEL

429

2.1

Residuals

1.1

0.1

–0.9

Figure 11-12 Plot of residuals versus hydrocarbon level x, Example 11-8.

–1,9

0.87

1.07

1.27

1.47

1.67

Hydrocarbon level (%) x

The statistic R2 should be used with caution, because it is always possible to make R2 unity by simply adding enough terms to the model. For example, we can obtain a “perfect” fit to n data points with a polynomial of degree n 1. In addition, R2 will always increase if we add a variable to the model, but this does not necessarily imply that the new model is superior to the old one. Unless the error sum of squares in the new model is reduced by an amount equal to the original error mean square, the new model will have a larger error mean square than the old one, because of the loss of one error degree of freedom. Thus, the new model will actually be worse than the old one. There are several misconceptions about R2. In general, R2 does not measure the magnitude of the slope of the regression line. A large value of R2 does not imply a steep slope. Furthermore, R2 does not measure the appropriateness of the model, since it can be artificially inflated by adding higher order polynomial terms in x to the model. Even if y and x are related in a nonlinear fashion, R2 will often be large. For example, R2 for the regression equation in Fig. 11-6(b) will be relatively large, even though the linear approximation is poor. Finally, even though R2 is large, this does not necessarily imply that the regression model will provide accurate predictions of future observations.

EXERCISES FOR SECTION 11-7 11-52. Refer to the compressive strength data in Exercise 11-1. Use the summary statistics provided to calculate R2 and provide a practical interpretation of this quantity. 11-53. Refer to the NFL quarterback ratings data in Exercise 11-3. (a) Calculate R2 for this model and provide a practical interpretation of this quantity. (b) Prepare a normal probability plot of the residuals from the least squares model. Does the normality assumption seem to be satisfied? (c) Plot the residuals versus the fitted values and against x. Interpret these graphs. 11-54. Refer to the data in Exercise 11-4 on house selling price y and taxes paid x. (a) Find the residuals for the least squares model. (b) Prepare a normal probability plot of the residuals and interpret this display.

(c) Plot the residuals versus yˆ and versus x. Does the assumption of constant variance seem to be satisfied? (d) What proportion of total variability is explained by the regression model? 11-55. Refer to the data in Exercise 11-5 on y  steam usage and x  average monthly temperature. (a) What proportion of total variability is accounted for by the simple linear regression model? (b) Prepare a normal probability plot of the residuals and interpret this graph. (c) Plot residuals versus yˆ and x. Do the regression assumptions appear to be satisfied? 11-56. Refer to the gasoline mileage data in Exercise 11-6. (a) What proportion of total variability in highway gasoline mileage performance is accounted for by engine displacement? (b) Plot the residuals versus yˆ and x, and comment on the graphs.

JWCL232_c11_401-448.qxd

430

1/14/10

8:03 PM

Page 430

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

(c) Prepare a normal probability plot of the residuals. Does the normality assumption appear to be satisfied? 11-57. Exercise 11-9 presents data on wear volume y and oil viscosity x. (a) Calculate R2 for this model. Provide an interpretation of this quantity. (b) Plot the residuals from this model versus yˆ and versus x. Interpret these plots. (c) Prepare a normal probability plot of the residuals. Does the normality assumption appear to be satisfied? 11-58. Refer to Exercise 11-8, which presented data on blood pressure rise y and sound pressure level x. (a) What proportion of total variability in blood pressure rise is accounted for by sound pressure level? (b) Prepare a normal probability plot of the residuals from this least squares model. Interpret this plot. (c) Plot residuals versus yˆ and versus x. Comment on these plots. 11-59. Refer to Exercise 11-10, which presented data on chloride concentration y and roadway area x. (a) What proportion of the total variability in chloride concentration is accounted for by the regression model? (b) Plot the residuals versus yˆ and versus x. Interpret these plots. (c) Prepare a normal probability plot of the residuals. Does the normality assumption appear to be satisfied? 11-60. An article in the Journal of the American Statistical Association [“Markov Chain Monte Carlo Methods for

Compressive Strength

Density

Compressive Strength

Density

3040 2470 3610 3480 3810 2330 1800 3110 3160 2310 4360 1880 3670 1740 2250 2650 4970 2620 2900 1670 2540

29.2 24.7 32.3 31.3 31.5 24.5 19.9 27.3 27.1 24.0 33.8 21.5 32.2 22.5 27.5 25.6 34.5 26.2 26.7 21.1 24.1

3840 3800 4600 1900 2530 2920 4990 1670 3310 3450 3600 2850 1590 3770 3850 2480 3570 2620 1890 3030 3030

30.7 32.7 32.6 22.1 25.3 30.8 38.9 22.1 29.2 30.1 31.4 26.7 22.1 30.3 32.0 23.2 30.3 29.9 20.8 33.2 28.2

Computing Bayes Factors: A Comparative Review” (2001, Vol. 96, pp. 1122–1132)] analyzed the tabulated data on compressive strength parallel to the grain versus resin-adjusted density for specimens of radiata pine. (a) Fit a regression model relating compressive strength to density. (b) Test for significance of regression with   0.05. (c) Estimate 2 for this model. (d) Calculate R2 for this model. Provide an interpretation of this quantity. (e) Prepare a normal probability plot of the residuals and interpret this display. (f ) Plot the residuals versus yˆ and versus x. Does the assumption of constant variance seem to be satisfied? 11-61. Consider the rocket propellant data in Exercise 11-11. (a) Calculate R2 for this model. Provide an interpretation of this quantity. (b) Plot the residuals on a normal probability scale. Do any points seem unusual on this plot? (c) Delete the two points identified in part (b) from the sample and fit the simple linear regression model to the remaining 18 points. Calculate the value of R2 for the new model. Is it larger or smaller than the value of R2 computed in part (a)? Why? (d) Did the value of ˆ 2 change dramatically when the two points identified above were deleted and the model fit to the remaining points? Why? 11-62. Consider the data in Exercise 11-7 on y  green liquor Na2S concentration and x  paper machine production. Suppose that a 14th sample point is added to the original data, where y14  59 and x14  855. (a) Prepare a scatter diagram of y versus x. Fit the simple linear regression model to all 14 observations. (b) Test for significance of regression with   0.05. (c) Estimate 2 for this model. (d) Compare the estimate of 2 obtained in part (c) above with the estimate of 2 obtained from the original 13 points. Which estimate is larger and why? (e) Compute the residuals for this model. Does the value of e14 appear unusual? (f ) Prepare and interpret a normal probability plot of the residuals. (g) Plot the residuals versus yˆ and versus x. Comment on these graphs. 11-63. Consider the rocket propellant data in Exercise 11-11. Calculate the standardized residuals for these data. Does this provide any helpful information about the magnitude of the residuals? 11-64. Studentized Residuals. Show that the variance of the ith residual is 1xi x2 2 1 V1ei 2  2 c 1 a n  bd Sxx

JWCL232_c11_401-448.qxd

1/14/10

8:03 PM

Page 431

11-8 CORRELATION

Hint:

(d) Discuss the behavior of the studentized residual when the sample value xi is very near one end of the range of x. 11-65. Show that an equivalent way to define the test for significance of regression in simple linear regression is to base the test on R2 as follows: to test H0: 1  0 versus H1: 1  0, calculate

1xi x 2 2 1 cov1Yi, Yˆ i 2  2 c n  d. Sxx The ith studentized residual is defined as ri 

431

F0 

ei

1xi x 2 2 1 bd  c 1 an  Sxx B ˆ2

(a) Explain why ri has unit standard deviation. (b) Do the standardized residuals have unit standard deviation? (c) Discuss the behavior of the studentized residual when the sample value xi is very close to the middle of the range of x.

R2 1n 22 1 R2

and to reject H0: 1  0 if the computed value f0 f,1,n 2. Suppose that a simple linear regression model has been fit to n  25 observations and R2  0.90. (a) Test for significance of regression at   0.05. (b) What is the smallest value of R2 that would lead to the conclusion of a significant regression if   0.05?

11-8 CORRELATION Our development of regression analysis has assumed that x is a mathematical variable, measured with negligible error, and that Y is a random variable. Many applications of regression analysis involve situations in which both X and Y are random variables. In these situations, it is usually assumed that the observations (Xi, Yi), i  1, 2, p , n are jointly distributed random variables obtained from the distribution f (x, y). For example, suppose we wish to develop a regression model relating the shear strength of spot welds to the weld diameter. In this example, weld diameter cannot be controlled. We would randomly select n spot welds and observe a diameter (Xi) and a shear strength (Yi) for each. Therefore (Xi, Yi) are jointly distributed random variables. We assume that the joint distribution of Xi and Yi is the bivariate normal distribution pre2 sented in Chapter 5, and Y and Y2 are the mean and variance of Y, X and X are the mean and variance of X, and  is the correlation coefficient between Y and X. Recall that the correlation coefficient is defined as XY   X Y

(11-35)

where XY is the covariance between Y and X. The conditional distribution of Y for a given value of X  x is fY 0 x 1 y2 

1 1 y 0 1x 2 exp c a bd Y 0 x 2 12Y 0 x

(11-36)

where Y 0  Y X  

(11-37)

Y 1    X

(11-38)

X

and the variance of the conditional distribution of Y given X  x is 2Y 0 x   2Y 11 2 2

(11-39)

JWCL232_c11_401-448.qxd

432

1/14/10

8:04 PM

Page 432

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

That is, the conditional distribution of Y given X  x is normal with mean E1Y 0 x2  0  1x

(11-40)

and variance 2Y 0 x . Thus, the mean of the conditional distribution of Y given X  x is a simple linear regression model. Furthermore, there is a relationship between the correlation coefficient  and the slope 1. From Equation 11-38 we see that if   0, then 1  0, which implies that there is no regression of Y on X. That is, knowledge of X does not assist us in predicting Y. The method of maximum likelihood may be used to estimate the parameters 0 and 1. It can be shown that the maximum likelihood estimators of those parameters are ˆ Y  ˆ X  0 1

(11-41)

and n

ˆ   1

a Yi 1Xi X 2

i1 n

2 a 1Xi X 2



SXY SX X

(11-42)

i1

We note that the estimators of the intercept and slope in Equations 11-41 and 11-42 are identical to those given by the method of least squares in the case where X was assumed to be a mathematical variable. That is, the regression model with Y and X jointly normally distributed is equivalent to the model with X considered as a mathematical variable. This follows because the random variables Y given X  x are independently and normally distributed with mean 0  1x and constant variance 2Y 0 x . These results will also hold for any joint distribution of Y and X such that the conditional distribution of Y given X is normal. It is possible to draw inferences about the correlation coefficient  in this model. The estimator of  is the sample correlation coefficient n

R

a Yi 1Xi X 2

i1 n

n

c a 1Xi X 2 a 1Yi Y 2 d 2

i1

2

1 2



SX Y 1SX X SST 21 2

(11-43)

i1

Note that ˆ a  1

SST 1 2 b R SX X

(11-44)

ˆ is just the sample correlation coefficient R multiplied by a scale factor that is so the slope  1 the square root of the “spread” of the Y values divided by the “spread” of the X values.Thus, ˆ and R are closely related, although they provide somewhat different information. The  1 ˆ sample correlation coefficient R measures the linear association between Y and X, while  1 measures the predicted change in the mean of Y for a unit change in X. In the case of a mathematical variable x, R has no meaning because the magnitude of R depends on the choice of spacing of x. We may also write, from Equation 11-44, ˆ2 R2   1

ˆ S  SX X SSR 1 XY   SST SST SST

JWCL232_c11_401-448.qxd

1/14/10

8:04 PM

Page 433

433

11-8 CORRELATION

which is just the coefficient of determination. That is, the coefficient of determination R2 is just the square of the correlation coefficient between Y and X. It is often useful to test the hypotheses H0 :   0 H1:  0

(11-45)

The appropriate test statistic for these hypotheses is Test Statistic for Zero Correlation

T0 

R1n 2

(11-46)

21 R2

which has the t distribution with n 2 degrees of freedom if H0:   0 is true. Therefore, we would reject the null hypothesis if t0 t2,n 2. This test is equivalent to the test of the hypothesis H0: 1  0 given in Section 11-5.1. This equivalence follows directly from Equation 11-46. The test procedure for the hypotheses H0:   0 H1:  0

(11-47)

where 0  0 is somewhat more complicated. For moderately large samples (say, n  25), the statistic Z  arctanh R 

1 1R ln 2 1 R

(11-48)

is approximately normally distributed with mean and variance Z  arctanh  

1 1 ln 2 1 

and

Z2 

1 n 3

respectively. Therefore, to test the hypothesis H0:   0, we may use the test statistic Z0  1arctanh R arctanh 0 2 1n 321 2

(11-49)

and reject H0:   0 if the value of the test statistic in Equation 11-49 is such that z0 z2. It is also possible to construct an approximate 100(1 )% confidence interval for , using the transformation in Equation 11-48. The approximate 100(1 )% confidence interval is Confidence Interval for a Correlation Coefficient

tanh aarctanh r

z 2

1n 3

b    tanh aarctanh r 

where tanh u  (eu e u )(eu  e u ).

z 2

1n 3

b

(11-50)

JWCL232_c11_401-448.qxd

434

1/14/10

8:04 PM

Page 434

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

70 60

Strength

50 40 30 20

Figure 11-13 Scatter plot of wire bond strength versus wire length, Example 11-8.

EXAMPLE 11-8

10 0

0

5

10 Wire length

Wire Bond Pull Strength

In Chapter 1 (Section 1-3) an application of regression analysis is described in which an engineer at a semiconductor assembly plant is investigating the relationship between pull strength of a wire bond and two factors: wire length and die height. In this example, we will consider only one of the factors, the wire length. A random sample of 25 units is selected and tested, and the wire bond pull strength and wire length are observed for each unit. The data are shown in Table 1-2. We assume that pull strength and wire length are jointly normally distributed. Figure 11-13 shows a scatter diagram of wire bond strength versus wire length. We have used the Minitab option of displaying box plots of each individual variable on the scatter diagram. There is evidence of a linear relationship between the two variables. The Minitab output for fitting a simple linear regression model to the data is shown below.

15

20

Now Sxx  698.56 and Sxy  2027.7132, and the sample correlation coefficient is r

Sxy

3Sx x SST 4 1 2



2027.7132  0.9818 3 1698.560216105.92 4 1 2

Note that r 2  (0.9818)2  0.9640 (which is reported in the Minitab output), or that approximately 96.40% of the variability in pull strength is explained by the linear relationship to wire length. Now suppose that we wish to test the hypotheses H0:   0 H1:  0

Regression Analysis: Strength versus Length The regression equation is Strength  5.11  2.90 Length Predictor Constant Length

Coef 5.115 2.9027

S  3.093 PRESS  272.144

SE Coef 1.146 0.1170

T 4.46 24.80

R-Sq  96.4% R-Sq(pred)  95.54%

P 0.000 0.000 R-Sq(adj)  96.2%

Analysis of Variance Source Regression Residual Error Total

DF 1 23 24

SS 5885.9 220.1 6105.9

MS 5885.9 9.6

F 615.08

P 0.000

JWCL232_c11_401-448.qxd

1/14/10

8:04 PM

Page 435

11-8 CORRELATION

with   0.05. We can compute the t-statistic of Equation 11-46 as t0 

r 1n 2 21 r 2



0.9818123  24.8 11 0.9640

This statistic is also reported in the Minitab output as a test of H0: 1  0. Because t0.025,23  2.069, we reject H0 and conclude that the correlation coefficient  0.

435

Finally, we may construct an approximate 95% confidence interval on  from Equation 11-50. Since arctanh r  arctanh 0.9818  2.3452, Equation 11-50 becomes tanh a2.3452

1.96 1.96 b    tanh a2.3452  b 122 122

which reduces to 0.9585    0.9921

EXERCISES FOR SECTION 11–8 11-66. Suppose data is obtained from 20 pairs of (x, y) and the sample correlation coefficient is 0.8. (a) Test the hypothesis that H0 :   0 against H1:  0 with   0.05. Calculate the P-value. (b) Test the hypothesis that H1:   0.5 against H1:  0.5 with   0.05. Calculate the P-value. (c) Construct a 95% two-sided confidence interval for the correlation coefficient. Explain how the questions in parts (a) and (b) could be answered with a confidence interval. 11-67. Suppose data are obtained from 20 pairs of (x, y) and the sample correlation coefficient is 0.75. (a) Test the hypothesis that H0 :   0 against H1:  0 with   0.05. Calculate the P-value. (b) Test the hypothesis that H1:   0.5 against H1:  0.5 with   0.05. Calculate the P-value. (c) Construct a 95% one-sided confidence interval for the correlation coefficient. Explain how the questions in parts (a) and (b) could be answered with a confidence interval. 11-68. A random sample of n  25 observations was made on the time to failure of an electronic component and the temperature in the application environment in which the component was used. (a) Given that r  0.83, test the hypothesis that   0, using   0.05. What is the P-value for this test? (b) Find a 95% confidence interval on . (c) Test the hypothesis H0:   0.8 versus H1:   0.8, using   0.05. Find the P-value for this test. 11-69. A random sample of 50 observations was made on the diameter of spot welds and the corresponding weld shear strength. (a) Given that r  0.62, test the hypothesis that   0, using   0.01. What is the P-value for this test? (b) Find a 99% confidence interval for . (c) Based on the confidence interval in part (b), can you conclude that   0.5 at the 0.01 level of significance? 11-70. The following data gave X  the water content of snow on April 1 and Y  the yield from April to July (in inches) on the Snake River watershed in Wyoming for 1919 to 1935. (The data were taken from an article in Research Notes, Vol. 61, 1950, Pacific Northwest Forest Range Experiment Station, Oregon.)

x

y

x

y

23.1 32.8 31.8 32.0 30.4 24.0 39.5 24.2 52.5

10.5 16.7 18.2 17.0 16.3 10.5 23.1 12.4 24.9

37.9 30.5 25.1 12.4 35.1 31.5 21.1 27.6

22.8 14.1 12.9 8.8 17.4 14.9 10.5 16.1

(a) Estimate the correlation between Y and X. (b) Test the hypothesis that   0, using   0.05. (c) Fit a simple linear regression model and test for significance of regression using   0.05. What conclusions can you draw? How is the test for significance of regression related to the test on  in part (b)? (d) Analyze the residuals and comment on model adequacy. 11-71. The final test and exam averages for 20 randomly selected students taking a course in engineering statistics and a course in operations research follow. Assume that the final averages are jointly normally distributed. (a) Find the regression line relating the statistics final average to the OR final average. Graph the data. (b) Test for significance of regression using   0.05. (c) Estimate the correlation coefficient. (d) Test the hypothesis that   0, using   0.05. (e) Test the hypothesis that   0.5, using   0.05. (f) Construct a 95% confidence interval for the correlation coefficient. Statistics

OR

Statistics

OR

Statistics

OR

86 75 69 75 90 94 83

80 81 75 81 92 95 80

86 71 65 84 71 62 90

81 76 72 85 72 65 93

83 75 71 76 84 97

81 70 73 72 80 98

JWCL232_c11_401-448.qxd

436

1/14/10

8:04 PM

Page 436

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

11-72. The weight and systolic blood pressure of 26 randomly selected males in the age group 25 to 30 are shown in the following table. Assume that weight and blood pressure are jointly normally distributed.

Subject Weight 1 2 3 4 5 6 7 8 9 10 11 12 13

Systolic BP

Subject

Weight

Systolic BP

130 133 150 128 151 146 150 140 148 125 133 135 150

14 15 16 17 18 19 20 21 22 23 24 25 26

172 159 168 174 183 215 195 180 143 240 235 192 187

153 128 132 149 158 150 163 156 124 170 165 160 159

165 167 180 155 212 175 190 210 200 149 158 169 170

(a) Find a regression line relating systolic blood pressure to weight. (b) Test for significance of regression using   0.05. (c) Estimate the correlation coefficient. (d) Test the hypothesis that   0, using   0.05. (e) Test the hypothesis that   0.6, using   0.05. (f) Construct a 95% confidence interval for the correlation coefficient. 11-73. In an article in IEEE Transactions on Instrumentation and Measurement (2001, Vol. 50, pp. 986–990), researchers studied the effects of reducing current draw in a magnetic core by electronic means. They measured the current in a magnetic winding with and without the electronics in a paired experiment. Data for the case without electronics are provided in the following table.

Supply Voltage

Current Without Electronics (mA)

0.66 1.32 1.98 2.64 3.3 3.96 4.62 3.28 5.94 6.6

7.32 12.22 16.34 23.66 28.06 33.39 34.12 39.21 44.21 47.48

(a) Graph the data and fit a regression line to predict current without electronics to supply voltage. Is there a significant regression at   0.05? What is the P-value? (b) Estimate the correlation coefficient. (c) Test the hypothesis that   0 against the alternative  0 with   0.05. What is the P-value? (d) Compute a 95% confidence interval for the correlation coefficient. 11-74. The monthly absolute estimate of global (land and ocean combined) temperature indexes (degrees C) in 2000 and 2001 are (source: http://www.ncdc.noaa.gov/ oa/climate/): 2000: 12.28, 12.63, 13.22, 14.21, 15.13, 15.82, 16.05, 16.02, 15.29, 14.29, 13.16, 12.47 2001: 12.44, 12.55, 13.35, 14.22, 15.28, 15.99, 16.23, 16.17, 15.44, 14.52, 13.52, 12.61 (a) Graph the data and fit a regression line to predict 2001 temperatures from those in 2000. Is there a significant regression at   0.05? What is the P-value? (b) Estimate the correlation coefficient. (c) Test the hypothesis that   0.9 against the alternative  0.9 with   0.05. What is the P-value? (d) Compute a 95% confidence interval for the correlation coefficient. 11-75 Refer to the NFL quarterback ratings data in Exercise 11-3. (a) Estimate the correlation coefficient between the ratings and the average yards per attempt. (b) Test the hypothesis H0 :   0 versus H1:  0 using   0.05. What is the P-value for this test? (c) Construct a 95% confidence interval for . (d) Test the hypothesis H0:   0.7 versus H1:  0.7 using   0.05. Find the P-value for this test. 11-76. Consider the following (x, y) data. Calculate the correlation coefficient. Graph the data and comment on the relationship between x and y. Explain why the correlation coefficient does not detect the relationship between x and y.

x

y

x

y

4 3 3 2 2 1 1 0

0 2.65 2.65 3.46 3.46 3.87 3.87 4

0 1 1 2 2 3 3 4

4 3.87 3.87 3.46 3.46 2.65 2.65 0

JWCL232_c11_401-448.qxd

1/1/70

6:05 AM

Page 437

11-9 REGRESSION ON TRANSFORMD VARIABLE

437

11-9 REGRESSION ON TRANSFORMED VARIABLES We occasionally find that the straight-line regression model Y ⫽ ␤0 ⫹ ␤1x ⫹ ⑀ is inappropriate because the true regression function is nonlinear. Sometimes nonlinearity is visually determined from the scatter diagram, and sometimes, because of prior experience or underlying theory, we know in advance that the model is nonlinear. Occasionally, a scatter diagram will exhibit an apparent nonlinear relationship between Y and x. In some of these situations, a nonlinear function can be expressed as a straight line by using a suitable transformation. Such nonlinear models are called intrinsically linear. As an example of a nonlinear model that is intrinsically linear, consider the exponential function Y ⫽ ␤0 e ␤1 x⑀ This function is intrinsically linear, since it can be transformed to a straight line by a logarithmic transformation ln Y ⫽ ln ␤0 ⫹ ␤1 x ⫹ ln ⑀ This transformation requires that the transformed error terms ln ⑀ are normally and independently distributed with mean 0 and variance ␴2. Another intrinsically linear function is 1 Y ⫽ ␤0 ⫹ ␤1 a x b ⫹ ⑀ By using the reciprocal transformation z ⫽ 1兾x, the model is linearized to Y ⫽ ␤0 ⫹ ␤1z ⫹ ⑀ Sometimes several transformations can be employed jointly to linearize a function. For example, consider the function Y⫽

1 exp 1␤0 ⫹ ␤1 x ⫹ ⑀2

Letting Y* ⫽ 1ⲐY, we have the linearized form ln Y * ⫽ ␤0 ⫹ ␤1x ⫹ ⑀ For examples of fitting these models, refer to Montgomery, Peck, and Vining (2006) or Myers (1990). Transformations can be very useful in many situations where the true relationship between the response Y and the regressor x is not well approximated by a straight line. The utility of a transformation is illustrated in the following example.

EXAMPLE 11-9

Windmill Power

A research engineer is investigating the use of a windmill to generate electricity and has collected data on the DC output from this windmill and the corresponding wind velocity. The data are plotted in Figure 11-14 and listed in Table 11-5 (p.439). Inspection of the scatter diagram indicates that the relationship between DC output Y and wind velocity (x) may be nonlinear. However, we initially fit a straight-line model to the

data. The regression model is yˆ ⫽ 0.1309 ⫹ 0.2411 x The summary statistics for this model are R 2 ⫽ 0.8745, MSE ⫽ ␴ˆ 2 ⫽ 0.0557, and F0 ⫽ 160.26 (the P-value is ⬍0.0001).

JWCL232_c11_401-448.qxd

438

1/18/10

2:17 PM

Page 438

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

DC output, y

3.0

0.4

0.2

2.0

0.0 ei

1.0

–0.2 0.0

0

2

4

6 8 Wind velocity, x

–0.4

10

Figure 11-14 Plot of DC output y versus wind velocity x for the windmill data.

–0.6 0.8

0.4

1.2

1.6

2.0

2.4

y

Figure 11-15 Plot of residuals ei versus fitted values yˆ i for the windmill data. A plot of the residuals versus yˆ i is shown in Figure 11-15. This residual plot indicates model inadequacy and implies that the linear relationship has not captured all of the information in the wind speed variable. Note that the curvature that was apparent in the scatter diagram of Figure 11-14 is greatly amplified in the residual plots. Clearly some other model form must be considered. We might initially consider using a quadratic model such as y ⫽ ␤0 ⫹ ␤1 x ⫹ ␤2 x 2 ⫹ ⑀ to account for the apparent curvature. However, the scatter diagram of Figure 11-14 suggests that as wind speed increases, DC output approaches an upper limit of approximately 2.5. This is also consistent with the theory of windmill operation. Since the quadratic model will eventually bend downward as wind speed increases, it would not be appropriate for these data. A more reasonable model for the windmill data that incorporates an upper asymptote would be

Figure 11-16 is a scatter diagram with the transformed variable x¿ ⫽ 1Ⲑx. This plot appears linear, indicating that the reciprocal transformation is appropriate. The fitted regression model is yˆ ⫽ 2.9789 ⫺ 6.9345 x¿ The summary statistics for this model are R2 ⫽ 0.9800, MSE ⫽ ␴ˆ 2 ⫽ 0.0089, and F0 ⫽ 1128.43 (the P value is ⬍0.0001). A plot of the residuals from the transformed model versus yˆ is shown in Figure 11-17. This plot does not reveal any serious problem with inequality of variance. The normal probability plot, shown in Figure 11-18, gives a mild indication that the errors come from a distribution with heavier tails than the normal (notice the slight upward and downward curve at the extremes). This normal probability plot has the z-score value plotted on the horizontal axis. Since there is no strong signal of model inadequacy, we conclude that the transformed model is satisfactory.

1 y ⫽ ␤0 ⫹ ␤1 a x b ⫹ ⑀ 0.4 3.0

DC output, y

0.2 2.0

0 ei –0.2

1.0 –0.4

0.0

–0.6 0.10

0.20

0.30

0.40

0.50

1 x' = x

Figure 11-16 Plot of DC output versus x¿ ⫽ 1Ⲑx for the windmill data.

0

2

1 yi

Figure 11-17 Plot of residuals versus fitted values yˆ i for the transformed model for the windmill data.

3

JWCL232_c11_401-448.qxd

1/1/70

6:04 AM

Page 439

11-9 REGRESSION ON TRANSFORMED VARIABLES 0.4 0.2

0 ei – 0.2

– 0.4

– 0.6 –2

–1

0 zi

1

2

Figure 11-18 Normal probability plot of the residuals for the transformed model for the windmill data. Table 11-5 Observed Values yi and Regressor Variable xi for Example 11-9 Observation Number, i

Wind Velocity (mph), xi

DC Output, yi

1 2 3

5.00 6.00 3.40

1.582 1.822 1.057 continued

439

Observation Number, i

Wind Velocity (mph), xi

DC Output, yi

4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

2.70 10.00 9.70 9.55 3.05 8.15 6.20 2.90 6.35 4.60 5.80 7.40 3.60 7.85 8.80 7.00 5.45 9.10 10.20 4.10 3.95 2.45

0.500 2.236 2.386 2.294 0.558 2.166 1.866 0.653 1.930 1.562 1.737 2.088 1.137 2.179 2.112 1.800 1.501 2.303 2.310 1.194 1.144 0.123

EXERCISES FOR SECTION 11–9 11-77. Determine if the following models are intrinsically linear. If yes, determine the appropriate transformation to generate the linear model. 3 ⫹ 5x (a) Y ⫽ ␤0x␤1⑀ (b) Y ⫽ ⫹⑀ x x (c) Y ⫽ ␤0␤x1⑀ (d) Y ⫽ ␤0x ⫹ ␤1 ⫹ x⑀ 11-78. The vapor pressure of water at various temperatures follows: Observation Number, i

Temperature (K)

Vapor pressure (mm Hg)

1 2 3 4 5 6 7 8 9 10 11

273 283 293 303 313 323 333 343 353 363 373

4.6 9.2 17.5 31.8 55.3 92.5 149.4 233.7 355.1 525.8 760.0

(a) Draw a scatter diagram of these data. What type of relationship seems appropriate in relating y to x?

(b) Fit a simple linear regression model to these data. (c) Test for significance of regression using ␣ ⫽ 0.05. What conclusions can you draw? (d) Plot the residuals from the simple linear regression model versus yˆi. What do you conclude about model adequacy? (e) The Clausis–Clapeyron relationship states that ln 1Pv 2⬀⫺ T1 , where Pv is the vapor pressure of water. Repeat parts (a)–(d). using an appropriate transformation. 11-79. An electric utility is interested in developing a model relating peak hour demand ( y in kilowatts) to total monthly energy usage during the month (x, in kilowatt hours). Data for 50 residential customers are shown in the following table. Customer

x

y

Customer

x

y

1 2 3 4 5 6 7 8 9 10

679 292 1012 493 582 1156 997 2189 1097 2078

0.79 0.44 0.56 0.79 2.70 3.64 4.73 9.50 5.34 6.85

26 27 28 29 30 31 32 33 34 35

1434 837 1748 1381 1428 1255 1777 370 2316 1130

0.31 4.20 4.88 3.48 7.58 2.63 4.99 0.59 8.19 4.79 continued

JWCL232_c11_401-448.qxd

440

1/14/10

8:04 PM

Page 440

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

Customer

x

y

Customer

x

y

11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

1818 1700 747 2030 1643 414 354 1276 745 795 540 874 1543 1029 710

5.84 5.21 3.25 4.43 3.16 0.50 0.17 1.88 0.77 3.70 0.56 1.56 5.28 0.64 4.00

36 37 38 39 40 41 42 43 44 45 46 47 48 49 50

463 770 724 808 790 783 406 1242 658 1746 895 1114 413 1787 3560

0.51 1.74 4.10 3.94 0.96 3.29 0.44 3.24 2.14 5.71 4.12 1.90 0.51 8.33 14.94

(a) (b) (c) (d)

Draw a scatter diagram of y versus x. Fit the simple linear regression model. Test for significance of regression using   0.05. Plot the residuals versus yˆi and comment on the underlying regression assumptions. Specifically, does it seem that the equality of variance assumption is satisfied? (e) Find a simple linear regression model using 1y as the response. Does this transformation on y stabilize the inequality of variance problem noted in part (d) above?

11-10 LOGISTIC REGRESSION Linear regression often works very well when the response variable is quantitative. We now consider the situation where the response variable takes on only two possible values, 0 and 1. These could be arbitrary assignments resulting from observing a qualitative response. For example, the response could be the outcome of a functional electrical test on a semiconductor device for which the results are either a “success,” which means the device works properly, or a “failure,” which could be due to a short, an open, or some other functional problem. Suppose that the model has the form Yi  0  1 xi  i

(11-51)

and the response variable Yi takes on the values either 0 or 1. We will assume that the response variable Yi is a Bernoulli random variable with probability distribution as follows:

Yi

Probability

1 0

P1Yi  12  i P1Yi  02  1 i

Now since E 1i 2  0, the expected value of the response variable is E 1Yi 2  1 1i 2  0 11 i 2  i This implies that E 1Yi 2  0  1xi  i

JWCL232_c11_401-448.qxd

1/14/10

8:05 PM

Page 441

441

11-10 LOGISTIC REGRESSION

This means that the expected response given by the response function E(Yi)  0  1xi is just the probability that the response variable takes on the value 1. There are some substantive problems with the regression model in Equation 11-51. First, note that if the response is binary, the error terms i can only take on two values, namely, i  1 10  1 xi 2

when Yi  1

i  10  1 xi 2

when Yi  0

Consequently, the errors in this model cannot possibly be normal. Second, the error variance is not constant, since 2Yi  E5Yi E1Yi 26 2

 11 i 2 2i  10 i 2 2 11 i 2

 i 11 i 2 Notice that this last expression is just

2yi  E1Yi 2 3 1 E1Yi 2 4

since E1Yi 2  0  1xi  i. This indicates that the variance of the observations (which is the same as the variance of the errors because i  Yi i, and i is a constant) is a function of the mean. Finally, there is a constraint on the response function, because 0  E 1Yi 2  i  1 This restriction can cause serious problems with the choice of a linear response function, as we have initially assumed in Equation 11-51. It would be possible to fit a model to the data for which the predicted values of the response lie outside the 0, 1 interval. Generally, when the response variable is binary, there is considerable empirical evidence indicating that the shape of the response function should be nonlinear. A monotonically increasing (or decreasing) S-shaped (or reverse S-shaped) function, such as shown in Figure 11-19, is usually employed. This function is called the logit response function, and has the form E1Y 2 

exp 10  1 x2 1  exp 10  1 x2

1.2

1.2

1

1

0.8

0.8

E( Y) 0.6

E( Y) 0.6

0.4

0.4

0.2

0.2

0

0

2

4

6

8

10

12

14

0

0

2

4

x

(11-52)

6

8

10

12

14

x

(a)

Figure 11-19 Examples of the logistic response function. (a) E1Y 2  1 11  e

6.0 1.0x

(b)

2, (b) E1Y 2  1 11  e 6.01.0x 2 .

JWCL232_c11_401-448.qxd

442

1/15/10

4:54 PM

Page 442

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

or equivalently, E1Y 2 

1 1  exp 3 10  1 x2 4

(11-53)

In logistic regression we assume that E(Y) is related to x by the logit function. It is easy to show that E1Y 2

1 E1Y 2

 exp10  1x2

(11-54)

The quantity exp( 0  1x) on the right-hand side of Equation 11-54 is called the odds ratio. It has a straightforward interpretation: If the odds ratio is 2 for a particular value of x, it means that a success is twice as likely as a failure at that value of the regressor x. Notice that the natural logarithm of the odds ratio is a linear function of the regressor variable. Therefore the slope 1 is the change in the log odds that results from a one-unit increase in x. This means that the odds ratio changes by e1 when x increases by one unit. The parameters in this logistic regression model are usually estimated by the method of maximum likelihood. For details of the procedure, see Montgomery, Peck, and Vining (2006). Minitab will fit logistic regression models and provide useful information on the quality of the fit. We will illustrate logistic regression using the data on launch temperature and O-ring failure for the 24 space shuttle launches prior to the Challenger disaster of January 1986. There are six O-rings used to seal field joints on the rocket motor assembly. The table below presents the launch temperatures. A 1 in the “O-Ring Failure” column indicates that at least one O-ring failure had occurred on that launch.

Temperature

O-Ring Failure

Temperature

O-Ring Failure

Temperature

O-Ring Failure

53 56 57 63 66 67 67 67

1 1 1 0 0 0 0 0

68 69 70 70 70 70 72 73

0 0 0 1 1 1 0 0

75 75 76 76 78 79 80 81

0 1 0 0 0 0 0 0

Figure 11-20 is a scatter plot of the data. Note that failures tend to occur at lower temperatures. The logistic regression model fit to this data from Minitab is shown in the following boxed display. The fitted logistic regression model is yˆ 

1 1  exp 3 110.875 0.17132x2 4

JWCL232_c11_401-448.qxd

1/14/10

8:05 PM

Page 443

11-10 LOGISTIC REGRESSION

443

Binary Logistic Regression: O-Ring Failure versus Temperature Link Function: Logit Response Information Variable O-Ring F

Value 1 0 Total

Count 7 17 24

(Event)

Logistic Regression Table Predictor Constant Temperat

Coef 10.875 0.17132

SE Coef 5.703 0.08344

Z 1.91 2.05

P 0.057 0.040

Odds Ratio

95% Lower

CI Upper

0.84

0.72

0.99

Log-Likelihood  11.515 Test that all slopes are zero: G  5.944, DF  1, P-Value  0.015

The standard error of the slope ˆ 1 is se(ˆ 1)  0.08344. For large samples, ˆ 1 has an approximate normal distribution, and so ˆ 1se(ˆ 1) can be compared to the standard normal distribution to test H0: 1  0. Minitab performs this test. The P-value is 0.04, indicating that temperature has a significant effect on the probability of O-ring failure. The odds ratio is 0.84, so every one degree increase in temperature reduces the odds of failure by 0.84. Figure 11-21 shows the fitted logistic regression model. The sharp increase in the probability of O-ring failure is very evident in this graph. The actual temperature at the Challenger launch was 31F . This is well outside the range of other launch temperatures, so our logistic regression model is not likely to provide highly accurate predictions at that temperature, but it is clear that a launch at 31F is almost certainly going to result in O-ring failure. It is interesting to note that all of these data were available prior to launch. However, engineers were unable to effectively analyze the data and use them to provide a convincing argument against launching Challenger to NASA managers. Yet a simple regression analysis

1.0

P(O-ring failure)

O-ring failure

1.0

0.5

0.0

0.5

0.0 50

60

70 Temperature

Figure 11-20 Scatter plot of O-ring failures versus launch temperature for 24 space shuttle flights.

80

50

60

70

80

Temperature

Figure 11-21 Probability of O-ring failure versus launch temperature (based on a logistic regression model).

JWCL232_c11_401-448.qxd

444

1/19/10

8:59 PM

Page 444

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

of the data would have provided a strong quantitative basis for this argument. This is one of the more dramatic instances that points out why engineers and scientists need a strong background in basic statistical techniques.

EXERCISES FOR SECTION 11–10 11-80 A study was conducted attempting to relate home ownership to family income. Twenty households were selected and family income was estimated, along with information concerning home ownership (y ⫽ 1 indicates yes and y ⫽ 0 indicates no). The data are shown below.

Household

Income

Home Ownership Status

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

38,000 51,200 39,600 43,400 47,700 53,000 41,500 40,800 45,400 52,400 38,700 40,100 49,500 38,000 42,000 54,000 51,700 39,400 40,900 52,800

0 1 0 1 0 0 1 0 1 1 1 0 1 0 1 1 1 0 0 1

(a) Fit a logistic regression model to the response variable y. Use a simple linear regression model as the structure for the linear predictor. (b) Is the logistic regression model in part (a) adequate? (c) Provide an interpretation of the parameter ␤1 in this model. 11-81 The compressive strength of an alloy fastener used in aircraft construction is being studied. Ten loads were selected over the range 2500– 4300 psi and a number of fasteners were tested at those loads. The numbers of fasteners failing at each load were recorded. The complete test data follow.

Load, x (psi)

Sample Size, n

Number Failing, r

2500 2700 2900 3100 3300 3500 3700 3900 4100 4300

50 70 100 60 40 85 90 50 80 65

10 17 30 21 18 43 54 33 60 51

(a) Fit a logistic regression model to the data. Use a simple linear regression model as the structure for the linear predictor. (b) Is the logistic regression model in part (a) adequate? 11-82 The market research department of a soft drink manufacturer is investigating the effectiveness of a price discount coupon on the purchase of a two-liter beverage product. A sample of 5500 customers was given coupons for varying price discounts between 5 and 25 cents. The response variable was the number of coupons in each price discount category redeemed after one month. The data are shown below.

Discount, x

Sample Size, n

Number Redeemed, r

5 7 9 11 13 15 17 19 21 23 25

500 500 500 500 500 500 500 500 500 500 500

100 122 147 176 211 244 277 310 343 372 391

(a) Fit a logistic regression model to the data. Use a simple linear regression model as the structure for the linear predictor. (b) Is the logistic regression model in part (a) adequate? (c) Draw a graph of the data and the fitted logistic regression model.

JWCL232_c11_401-448.qxd

1/14/10

8:05 PM

Page 445

445

11-10 LOGISTIC REGRESSION

(d) Expand the linear predictor to include a quadratic term. Is there any evidence that this quadratic term is required in the model? (e) Draw a graph of this new model on the same plot that you prepared in part (c). Does the expanded model visually provide a better fit to the data than the original model from part (a)? 11-83 A study was performed to investigate new automobile purchases. A sample of 20 families was selected. Each family was surveyed to determine the age of their oldest vehicle and their total family income. A follow-up survey was conducted six months later to determine if they had actually purchased a new vehicle during that time period ( y = 1 indicates yes and y = 0 indicates no). The data from this study are shown in the following table. Income, x1 Age, x2 45,000 40,000 60,000 50,000 55,000 50,000 35,000 65,000 53,000 48,000

2 4 3 2 2 5 7 2 2 1

y

Income, x1 Age, x2

0 0 1 1 0 1 1 1 0 0

37,000 31,000 40,000 75,000 43,000 49,000 37,500 71,000 34,000 27,000

5 7 4 2 9 2 4 1 5 6

y 1 1 1 0 1 0 1 0 0 0

y

x

y

x

0.734 0.886 1.04 1.19 1.35

1.1 1.2 1.3 1.4 1.5

1.50 1.66 1.81 1.97 2.12

1.6 1.7 1.8 1.9 2.0

(a) Draw a scatter diagram of these data. Does a straight-line relationship seem plausible? (b) Fit a simple linear regression model to these data. (c) Test for significance of regression using   0.05. What is the P-value for this test? (d) Find a 95% confidence interval estimate on the slope. (e) Test the hypothesis H0: 0  0 versus H1: 0  0 using   0.05. What conclusions can you draw? 11-86. The strength of paper used in the manufacture of cardboard boxes ( y) is related to the percentage of hardwood concentration in the original pulp (x). Under controlled conditions, a pilot plant manufactures 16 samples, each from a different batch of pulp, and measures the tensile strength. The data are shown in the table that follows:

(a) (b) (c) (d)

Fit a logistic regression model to the data. Is the logistic regression model in part (a) adequate? Interpret the model coefficients 1 and 2. What is the estimated probability that a family with an income of $45,000 and a car that is five years old will purchase a new vehicle in the next six months? (e) Expand the linear predictor to include an interaction term. Is there any evidence that this term is required in the model?

y

101.4

117.4

117.1

106.2

x

1.0

1.5

1.5

1.5

y

131.9

146.9

146.8

133.9

x

2.0

2.0

2.2

2.4

y

111.0

123.0

125.1

145.2

x

2.5

2.5

2.8

2.8

y

134.3

144.5

143.7

146.9

x

3.0

3.0

3.2

3.3

Supplemental Exercises 11-84. Show that, for the simple linear regression model, the following statements are true: n

(a) a 1 yi yˆi 2  0 i1

n

(b) a 1 yi yˆi 2 xi  0 i1

n

1 (c) n a yˆi  y i1 11-85. An article in the IEEE Transactions on Instrumentation and Measurement [“Direct, Fast, and Accurate Measurement of VT and K of MOS Transistor Using VT-Sift Circuit” (1991, Vol. 40, pp. 951–955)] described the use of a simple linear regression model to express drain current y (in milliamperes) as a function of ground-to-source voltage x (in volts). The data are as follows:

(a) (b) (c) (d) (e)

Fit a simple linear regression model to the data. Test for significance of regression using   0.05. Construct a 90% confidence interval on the slope 1. Construct a 90% confidence interval on the intercept 0. Construct a 95% confidence interval on the mean strength at x  2.5. (f) Analyze the residuals and comment on model adequacy. 11-87. Consider the following data. Suppose that the relationship between Y and x is hypothesized to be Y  (0  1x  ) 1. Fit an appropriate model to the data. Does the assumed model form seem reasonable? x

10

15

18

12

9

8

11

6

y

0.1

0.13

0.09

0.15

0.20

0.21

0.18

0.24

JWCL232_c11_401-448.qxd

446

1/14/10

8:05 PM

Page 446

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

11-88. The following data, adapted from Montgomery, Peck, and Vining (2006), present the number of certified mental defectives per 10,000 of estimated population in the United Kingdom ( y) and the number of radio receiver licenses issued (x) by the BBC (in millions) for the years 1924 through 1937. Fit a regression model relating y and x. Comment on the model. Specifically, does the existence of a strong correlation imply a cause-and-effect relationship?

Year

y

1924 1925 1926 1927 1928 1929 1930

8 8 9 10 11 11 12

x

Year

y

x

1.350 1.960 2.270 2.483 2.730 3.091 3.674

1931 1932 1933 1934 1935 1936 1937

16 18 19 20 21 22 23

4.620 5.497 6.260 7.012 7.618 8.131 8.593

11-89. Consider the weight and blood pressure data in Exercise 11-72. Fit a no-intercept model to the data, and compare it to the model obtained in Exercise 11-70. Which model is superior? 11-90. An article in Air and Waste [“Update on Ozone Trends in California’s South Coast Air Basin” (Vol. 43, 1993)] studied the ozone levels on the South Coast air basin of California for the years 1976–1991. The author believes that the number of days that the ozone level exceeds 0.20 parts per million depends on the seasonal meteorological index (the seasonal average 850 millibar temperature). The data follow:

Year

Days

Index

Year

Days

Index

1976 1977 1978 1979 1980 1981 1982 1983

91 105 106 108 88 91 58 82

16.7 17.1 18.2 18.1 17.2 18.2 16.0 17.2

1984 1985 1986 1987 1988 1989 1990 1991

81 65 61 48 61 43 33 36

18.0 17.2 16.9 17.1 18.2 17.3 17.5 16.6

(a) Construct a scatter diagram of the data. (b) Fit a simple linear regression model to the data. Test for significance of regression. (c) Find a 95% CI on the slope 1. (d) Analyze the residuals and comment on model adequacy. 11-91. An article in the Journal of Applied Polymer Science (Vol. 56, pp. 471–476, 1995) studied the effect of the mole

ratio of sebacic acid on the intrinsic viscosity of copolyesters. The data follow: Mole ratio x 1.0

0.9

0.8

0.7

0.6

0.5

0.4

0.3

Viscosity y 0.45 0.20 0.34 0.58 0.70 0.57 0.55 0.44 (a) Construct a scatter diagram of the data. (b) Fit a simple linear repression model. (c) Test for significance of regression. Calculate R2 for the model. (d) Analyze the residuals and comment on model adequacy. 11-92. Two different methods can be used for measuring the temperature of the solution in a Hall cell used in aluminum smelting, a thermocouple implanted in the cell and an indirect measurement produced from an IR device. The indirect method is preferable because the thermocouples are eventually destroyed by the solution. Consider the following 10 measurements: Thermocouple

921

935

916

920

940

IR

918

934

924

921

945

Thermocouple

936

925

940

933

927

IR

930

919

943

932

935

(a) Construct a scatter diagram for these data, letting x  thermocouple measurement and y  IR measurement. (b) Fit a simple linear regression model. (c) Test for significance a regression and calculate R2. What conclusions can you draw? (d) Is there evidence to support a claim that both devices produce equivalent temperature measurements? Formulate and test an appropriate hypothesis to support this claim. (e) Analyze the residuals and comment on model adequacy. 11-93. The grams of solids removed from a material ( y) is thought to be related to the drying time. Ten observations obtained from an experimental study follow: y 4.3

1.5

1.8

4.9

4.2

4.8

5.8

6.2

7.0

7.9

x 2.5

3.0

3.5

4.0

4.5

5.0

5.5

6.0

6.5

7.0

(a) (b) (c) (d)

Construct a scatter diagram for these data. Fit a simple linear regression model. Test for significance of regression. Based on these data, what is your estimate of the mean grams of solids removed at 4.25 hours? Find a 95% confidence interval on the mean. (e) Analyze the residuals and comment on model adequacy.

JWCL232_c11_401-448.qxd

1/14/10

8:05 PM

Page 447

11-10 LOGISTIC REGRESSION

11-94. Cesium atoms cooled by laser light could be used to build inexpensive atomic clocks. In a study in IEEE Transactions on Instrumentation and Measurement (2001, Vol. 50, pp. 1224–1228), the number of atoms cooled by lasers of various powers were counted.

Power (mW)

Number of Atoms (10E9)

11 12 18 21 22 24 28 32 37 39 41 46 48 50 51

0 0.02 0.08 0.13 0.15 0.18 0.31 0.4 0.49 0.57 0.64 0.71 0.79 0.82 0.83

(a) Graph the data and fit a regression line to predict the number of atoms from laser power. Comment on the adequacy of a linear model. (b) Is there a significant regression at   0.05? What is the P-value? (c) Estimate the correlation coefficient. (d) Test the hypothesis that   0 against the alternative  0 with   0.05. What is the P-value? (e) Compute a 95% confidence interval for the slope coefficient. 11-95. The following data related diamond carats to purchase prices. It appeared in Singapore’s Business Times, February 18, 2000.

Carat

Price

Carat

Price

0.3 0.3 0.3 0.3 0.31 0.31 0.31 0.31 0.31

1302 1510 1510 1260 1641 1555 1427 1427 1126

0.33 0.33 0.34 0.34 0.34 0.34 0.34 0.34 0.35

1327 1098 1693 1551 1410 1269 1316 1222 1738

Carat

Price

Carat

Price

0.31 0.32 0.32 0.36 0.36 0.37 0.37 0.4 0.4 0.41 0.43

1126 1468 1202 1635 1485 1420 1420 1911 1525 1956 1747

0.35 0.35 0.35 0.45 0.46 0.48 0.5 0.5 0.5 0.5 0.5

1593 1447 1255 1572 2942 2532 3501 3501 3501 3293 3016

447

(a) Graph the data. What is the relation between carat and price? Is there an outlier? (b) What would you say to the person who purchased the diamond that was an outlier? (c) Fit two regression models, one with all the data and the other with unusual data omitted. Estimate the slope coefficient with a 95% confidence interval in both cases. Comment on any difference. 11-96. The following table shows the population and the average count of wood storks sighted per sample period for South Carolina from 1991 to 2004. Fit a regression line with population as the response and the count of wood storks as the predictor. Such an analysis might be used to evaluate the relationship between storks and babies. Is regression significant at   0.05? What do you conclude about the role of regression analysis to establish a cause-and-effect relationship?

Year

Population

Stork Count

1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004

3,559,470 3,600,576 3,634,507 3,666,456 3,699,943 3,738,974 3,790,066 3,839,578 3,885,736 4,012,012 4,061,209 4,105,848 4,148,744 4,198,068

0.342 0.291 0.291 0.291 0.291 0.509 0.294 0.799 0.542 0.495 0.859 0.364 0.501 0.656

JWCL232_c11_401-448.qxd

448

1/14/10

8:05 PM

Page 448

CHAPTER 11 SIMPLE LINEAR REGRESSION AND CORRELATION

MIND-EXPANDING EXERCISES 11-97. Suppose that we have n pairs of observations (xi, yi) such that the sample correlation coefficient r is unity (approximately). Now let zi  y 2i and consider the sample correlation coefficient for the n-pairs of data (xi, zi). Will this sample correlation coefficient be approximately unity? Explain why or why not. 11-98. Consider the simple linear regression model Y  0  1x  , with E()  0, V()  2, and the errors  uncorrelated. (a) Show that cov 1ˆ 0, ˆ 1 2  x2 Sx x. (b) Show that cov 1Y, ˆ 1 2  0 . 11-99. Consider the simple linear regression model Y  0  1x  , with E()  0, V()  2, and the errors  uncorrelated. ˆ 2)  E(MSE)  2. (a) Show that E( (b) Show that E(MSR)  2  12Sx x. 11-100. Suppose that we have assumed the straightline regression model Y  0  1x1   but the response is affected by a second variable x2 such that the true regression function is

of Y depends on the level of x; that is, 2 V1Yi 0 xi 2   2i  w i

i  1, 2, p , n

where the wi are constants, often called weights. Show that for an objective function in which each squared residual is multiplied by the reciprocal of the variance of the corresponding observation, the resulting weighted least squares normal equations are n

n

n

i1

n

i1 n

i1 n

i1

i1

i1

ˆ 0 a wi  ˆ 1 a wi xi  a wi yi ˆ 0 a wi xi  ˆ 1 a wi x2i  a wi xi yi Find the solution to these normal equations. The solutions are weighted least squares estimators of 0 and 1. 11-103. Consider a situation where both Y and X are random variables. Let sx and sy be the sample standard deviations of the observed x’s and y’s, respectively. Show that an alternative expression for the fitted simple linear regression model yˆ  ˆ 0  ˆ 1x is

E1Y 2  0  1x1  2x2

sy yˆ  y  r s 1x x2 x

Is the estimator of the slope in the simple linear regression model unbiased? 11-101. Suppose that we are fitting a line and we wish to make the variance of the regression coefficient ˆ 1 as small as possible. Where should the observations xi, i  1, 2, p , n, be taken so as to minimize V(ˆ 1)? Discuss the practical implications of this allocation of the xi. 11-102. Weighted Least Squares. Suppose that we are fitting the line Y  0  1x  , but the variance

11-104. Suppose that we are interested in fitting a simple linear regression model Y  0  1x  , where the intercept, 0, is known. (a) Find the least squares estimator of 1. (b) What is the variance of the estimator of the slope in part (a)? (c) Find an expression for a 100(1 )% confidence interval for the slope 1. Is this interval longer than the corresponding interval for the case where both the intercept and slope are unknown? Justify your answer.

IMPORTANT TERMS AND CONCEPTS Analysis of variance test in regression Confidence interval on mean response Correlation coefficient Empirical model

Confidence intervals on model parameters Intrinsically linear model Least squares estimation of regression model parameters Logistic regression

Model adequacy checking Odds ratio Prediction interval on a future observation Regression analysis Residual plots Residuals

Scatter diagram Significance of regression Simple linear regression model standard errors Statistical tests on model parameters Transformations

JWCL232_c12_449-512.qxd

1/16/10

10:43 AM

Page 449

12

© David Lewis/ iStockphoto

Multiple Linear Regression This chapter generalizes the simple linear regression to a situation where there is more than one predictor or regressor variable. This situation occurs frequently in science and engineering; for example, in Chapter 1 we provided data on the pull strength of a wire bond on a semiconductor package and illustrated its relationship to the wire length and the die height. Understanding the relationship between strength and the other two variables may provide important insight to the engineer when the package is designed, or to the manufacturing personnel who assemble the die into the package. We used a multiple linear regression model to relate strength to wire length and die height. There are many examples of such relationships: The life of a cutting tool is related to the cutting speed and the tool angle; patient satisfaction in a hospital is related to patient age, type of procedure performed, and length of stay; and the fuel economy of a vehicle is related to the type of vehicle (car versus truck), engine displacement, horsepower, type of transmission, and vehicle weight. Multiple regression models give insight into the relationships between these variables that can have important practical implications. This chapter shows how to fit multiple linear regression models, perform the statistical tests and confidence procedures that are analogous to those for simple linear regression, and check for model adequacy. We also show how models that have polynomial terms in the regressor variables are just multiple linear regression models. We also discuss some aspects of building a good regression model from a collection of candidate regressors.

CHAPTER OUTLINE 12-1 MULTIPLE LINEAR REGRESSION MODEL 12-1.1 Introduction 12-1.2 Least Squares Estimation of the Parameters

12-1.3 Matrix Approach to Multiple Linear Regression 12-1.4 Properties of the Least Squares Estimators

449

JWCL232_c12_449-512.qxd

450

1/15/10

10:06 PM

Page 450

CHAPTER 12 MULTIPLE LINEAR REGRESSION

12-2 HYPOTHESIS TESTS IN MULTIPLE LINEAR REGRESSION 12-2.1 Test for Significance of Regression 12-2.2 Tests on Individual Regression Coefficients and Subsets of Coefficients 12-3 CONFIDENCE INTERVALS IN MULTIPLE LINEAR REGRESSION 12-3.1 Confidence Intervals on Individual Regression Coefficients 12-3.2 Confidence Interval on the Mean Response

12-4 PREDICTION OF NEW OBSERVATIONS 12-5 MODEL ADEQUACY CHECKING 12-5.1 Residual Analysis 12-5.2 Influential Observations 12-6 ASPECTS OF MULTIPLE REGRESSION MODELING 12-6.1 Polynomial Regression Models 12-6.2 Categorical Regressors and Indicator Variables 12-6.3 Selection of Variables and Model Building 12-6.4 Multicollinearity

LEARNING OBJECTIVES After careful study of this chapter you should be able to do the following: 1. Use multiple regression techniques to build empirical models to engineering and scientific data 2. Understand how the method of least squares extends to fitting multiple regression models 3. Assess regression model adequacy 4. Test hypotheses and construct confidence intervals on the regression coefficients 5. Use the regression model to estimate the mean response and to make predictions and to construct confidence intervals and prediction intervals 6. Build regression models with polynomial terms 7. Use indicator variables to model categorical regressors 8. Use stepwise regression and other model building techniques to select the appropriate set of variables for a regression model

12-1 MULTIPLE LINEAR REGRESSION MODEL 12-1.1 Introduction Many applications of regression analysis involve situations in which there are more than one regressor or predictor variable. A regression model that contains more than one regressor variable is called a multiple regression model. As an example, suppose that the effective life of a cutting tool depends on the cutting speed and the tool angle. A multiple regression model that might describe this relationship is Y  0  1x1  2x2  

(12-1)

where Y represents the tool life, x1 represents the cutting speed, x2 represents the tool angle, and  is a random error term. This is a multiple linear regression model with two regressors. The term linear is used because Equation 12-1 is a linear function of the unknown parameters 0, 1, and 2.

JWCL232_c12_449-512.qxd

1/15/10

10:07 PM

Page 451

12-1 MULTIPLE LINEAR REGRESSION MODEL

451

x2 10 220 240

8

203

200 160 E(Y) 120

6

80

4

186

40 0

0

Figure 12-1

2

4 x1

6

8 (a)

10 0

2

4

6 x2

8

10

169

2 0

67 0

84

2

101 4

118 6

135 8

152 10 x1

(b)

(a) The regression plane for the model E(Y )  50  10x1  7x2. (b) The contour plot.

The regression model in Equation 12-1 describes a plane in the three-dimensional space of Y, x1, and x2. Figure 12-1(a) shows this plane for the regression model E1Y 2  50  10x1  7x2 where we have assumed that the expected value of the error term is zero; that is E()  0. The parameter 0 is the intercept of the plane. We sometimes call 1 and 2 partial regression coefficients, because 1 measures the expected change in Y per unit change in x1 when x2 is held constant, and 2 measures the expected change in Y per unit change in x2 when x1 is held constant. Figure 12-1(b) shows a contour plot of the regression model—that is, lines of constant E(Y ) as a function of x1 and x2. Notice that the contour lines in this plot are straight lines. In general, the dependent variable or response Y may be related to k independent or regressor variables. The model Y  0  1x1  2x2  p   k x k   ˛

˛

(12-2)

is called a multiple linear regression model with k regressor variables. The parameters j, j  0, 1, p , k, are called the regression coefficients. This model describes a hyperplane in the k-dimensional space of the regressor variables {xj}. The parameter j represents the expected change in response Y per unit change in xj when all the remaining regressors xi (i  j) are held constant. Multiple linear regression models are often used as approximating functions. That is, the true functional relationship between Y and x1, x2, p , xk is unknown, but over certain ranges of the independent variables the linear regression model is an adequate approximation. Models that are more complex in structure than Equation 12-2 may often still be analyzed by multiple linear regression techniques. For example, consider the cubic polynomial model in one regressor variable. Y  0  1x  2x 2  3 x 3  

(12-3)

If we let x1  x, x2  x 2, x3  x 3, Equation 12-3 can be written as Y  0  1x1  2x2  3 x3   which is a multiple linear regression model with three regressor variables.

(12-4)

JWCL232_c12_449-512.qxd

452

1/15/10

10:07 PM

Page 452

CHAPTER 12 MULTIPLE LINEAR REGRESSION

Models that include interaction effects may also be analyzed by multiple linear regression methods. An interaction between two variables can be represented by a cross-product term in the model, such as Y  0  1x1  2x2  12 x1x2  

(12-5)

If we let x3  x1x2 and 3  12, Equation 12-5 can be written as Y  0  1x1  2x2  3 x3   which is a linear regression model. Figure 12-2(a) and (b) shows the three-dimensional plot of the regression model Y  50  10x1  7x2  5x1x2 and the corresponding two-dimensional contour plot. Notice that, although this model is a linear regression model, the shape of the surface that is generated by the model is not linear. In general, any regression model that is linear in parameters (the ’s) is a linear regression model, regardless of the shape of the surface that it generates. Figure 12-2 provides a nice graphical interpretation of an interaction. Generally, interaction implies that the effect produced by changing one variable (x1, say) depends on the level of the other variable (x2). For example, Fig. 12-2 shows that changing x1 from 2 to 8 produces a much smaller change in E(Y ) when x2  2 than when x2  10. Interaction effects occur frequently in the study and analysis of real-world systems, and regression methods are one of the techniques that we can use to describe them. As a final example, consider the second-order model with interaction Y  0  1 x1  2 x2  11x 21  22 x 22  12 x1x2  

(12-6)

If we let x3  x 21, x4  x 22, x5  x1x2, 3  11, 4  22, and 5  12, Equation 12-6 can be written as a multiple linear regression model as follows: Y  0  1x1  2 x2  3 x3  4 x4  5 x5   Figure 12-3(a) and (b) show the three-dimensional plot and the corresponding contour plot for E1Y 2  800  10x1  7x2  8.5x 21  5x 22  4x 1x2 ˛

These plots indicate that the expected change in Y when x1 is changed by one unit (say) is a function of both x1 and x2. The quadratic and interaction terms in this model produce a moundshaped function. Depending on the values of the regression coefficients, the second-order model with interaction is capable of assuming a wide variety of shapes; thus, it is a very flexible regression model.

12-1.2 Least Squares Estimation of the Parameters The method of least squares may be used to estimate the regression coefficients in the multiple regression model, Equation 12-2. Suppose that n  k observations are available, and let

JWCL232_c12_449-512.qxd

1/15/10

10:07 PM

Page 453

453

12-1 MULTIPLE LINEAR REGRESSION MODEL

800 600 1000

E(Y) 400

800

200 0

0

2

4 x1

6

8

10

0

4

2

6

8

10

E(Y)

600 400

x2

200 0

(a)

0

2

4 x1

x2 10

8

10

2

4

586

25

8

100

519

6

6

452 385

4

4

318 2

251 117 0

2

4

6

10

x2

10

653

0

0

8

(a)

x2

720

8

6

6

550 700 625 800 750

2

325250 400 475

175

184 0

10 x1

8

0

2

4

6

10 x1

8

(b)

(b)

Figure 12-3 (a) Three-dimensional plot of the regression model E(Y )  800  10x1  7x2  8.5x21  5x22  4x1x2. (b) The contour plot.

Figure 12-2 (a) Three-dimensional plot of the regression model E(Y )  50  10x1  7x2  5x1x2. (b) The contour plot.

xij denote the ith observation or level of variable xj. The observations are 1xi 1, xi 2, p , xik, yi 2, ˛

˛

i  1, 2, p , n

nk

and

˛

It is customary to present the data for multiple regression in a table such as Table 12-1. Each observation (xi1, xi2, p , xik , yi ), satisfies the model in Equation 12-2, or y i  0  1xi1  2 xi 2  p  k xik  i ˛

k

 0  a j xij  i

i  1, 2, p , n ˛

˛

j1

Table 12-1 Data for Multiple Linear Regression y y1 y2

x1 x11 x21

x2 x12 x22

... ... ...

xk x1k x2k

o yn

o xn1

o xn2

...

o xnk

(12-7)

JWCL232_c12_449-512.qxd

454

1/15/10

10:07 PM

Page 454

CHAPTER 12 MULTIPLE LINEAR REGRESSION

The least squares function is n

n

k

2

L  a 2i  a ayi  0  a j xij b ˛

i1

˛

(12-8)

˛

i1

j1

We want to minimize L with respect to 0, 1, p , k. The least squares estimates of 0, 1, p , k must satisfy n k L `  2 a ayi  ˆ 0  a ˆ j xij b  0 0 ˆ 0,ˆ 1, p , ˆ k i1 j1

(12-9a)

and n k L `  2 a ayi  ˆ 0  a ˆ j xij b xij  0 j ˆ 0,ˆ 1, p, ˆ k i1 j1

j  1, 2, p , k

(12-9b)

Simplifying Equation 12-9, we obtain the least squares normal equations n

nˆ 0  ˆ 1 a xi1 ˛

i1

n

n

ˆ 2 ˆ  0 a xi1  1 a xi1 ˛

˛

n

n

 ˆ 2 a xi 2 ˛

 p  ˆ k a xik

˛

˛

i1

i1

n

˛

i1

n

n

i1

i1

o

o

 ˆ 2 a xi1 xi2  p  ˆ k a xi1xik  a xi1 yi

i1

i1

o

o

o

n

n

n

i1

o

n

2 ˆ 0 a xik  ˆ 1 a xik xi1  ˆ 2 a xik xi2  p  ˆ k a xik i1 i1 ˛

n

 a yi

˛

i1

i1

n

 a xik yi

(12-10)

i1

Note that there are p  k  1 normal equations, one for each of the unknown regression coefficients. The solution to the normal equations will be the least squares estimators of the ˆ , ˆ ,p, ˆ . The normal equations can be solved by any method regression coefficients,  0 1 k appropriate for solving a system of linear equations.

EXAMPLE 12-1

Wire Bond Strength

In Chapter 1, we used data on pull strength of a wire bond in a semiconductor manufacturing process, wire length, and die height to illustrate building an empirical model. We will use the same data, repeated for convenience in Table 12-2, and show the details of estimating the model parameters. A threedimensional scatter plot of the data is presented in Fig. 1-15. Figure 12-4 shows a matrix of two-dimensional scatter plots of the data. These displays can be helpful in visualizing the relationships among variables in a multivariable data set. For example, the plot indicates that there is a strong linear relationship between strength and wire length. Specifically, we will fit the multiple linear regression model Y  0  1x1  2 x2   where Y  pull strength, x1  wire length, and x2  die height. From the data in Table 12-2 we calculate

25

n  25, a yi  725.82 ˛

i1

25

25

i1

i1

a xi1  206, a xi 2  8,294 ˛

25

25

i1

i1

2 2 a x i1  2,396, a x i2  3,531,848 ˛

25

25

a xi1xi2  77,177, a xi1 yi  8,008.47, ˛

i1

25

˛

i1

a xi2 yi  274,816.71

i1

JWCL232_c12_449-512.qxd

1/15/10

10:07 PM

Page 455

455

12-1 MULTIPLE LINEAR REGRESSION MODEL

Table 12-2 Wire Bond Data for Example 12-1 Observation Number

Pull Strength y

Wire Length x1

Die Height x2

Observation Number

Pull Strength y

Wire Length Die Height x1 x2

1

9.95

2

50

14

11.66

2

360

2

24.45

8

110

15

21.65

4

205

3

31.75

11

120

16

17.89

4

400

4

35.00

10

550

17

69.00

20

600

5

25.02

8

295

18

10.30

1

585

6

16.86

4

200

19

34.93

10

540

7

14.38

2

375

20

46.59

15

250

8

9.60

2

52

21

44.88

15

290

9

24.35

9

100

22

54.12

16

510

10

27.50

8

300

23

56.63

17

590

11

17.08

4

412

24

22.13

6

100

12 13

37.00 41.95

11 12

400 500

25

21.15

5

400

For the model Y  0  1 x1   2 x 2  , the normal equations 12-10 are n

n

nˆ 0  ˆ 1 a xi1  ˆ 2 a xi 2 ˛

i1

n

n

ˆ 0 a xi1  ˆ 1 a xi12 ˛

˛

˛

i1

i1

n

n

˛

i1

n

˛

i1

i1

8294ˆ 2  725.82

206ˆ 0 

2396ˆ 1 

77,177ˆ 2  8,008.47

8294ˆ 0  77,177ˆ 1  3,531,848ˆ 2  274,816.71

i1

˛

206ˆ 1 

˛

i1 n

25ˆ 0 

i1

n

˛

i1

n

 a yi

 ˆ 2 a xi1xi2  a xi1 yi

ˆ 0 a xi 2  ˆ 1 a xi1xi 2  ˆ 2 a x2i 2 ˛

Inserting the computed summations into the normal equations, we obtain

n

 a xi 2 yi ˛

i1

54.15 Strength 24.45

15.25 Length 5.75

462.5 Height 187.5

24.45

54.15

5.75

15.25

187.5

462.5

Figure 12-4 Matrix of scatter plots (from Minitab) for the wire bond pull strength data in Table 12-2.

JWCL232_c12_449-512.qxd

456

1/15/10

10:07 PM

Page 456

CHAPTER 12 MULTIPLE LINEAR REGRESSION

The solution to this set of equations is

Practical Interpretation: This equation can be used to predict pull strength for pairs of values of the regressor variables wire length (x1) and die height (x2). This is essentially the same regression model given in Section 1-3. Figure 1-16 shows a three-dimensional plot of the plane of predicted values yˆ generated from this equation.

ˆ 0  2.26379, ˆ 1  2.74427, ˆ 2  0.01253 Therefore, the fitted regression equation is yˆ  2.26379  2.74427x1  0.01253x2

12-1.3 Matrix Approach to Multiple Linear Regression In fitting a multiple regression model, it is much more convenient to express the mathematical operations using matrix notation. Suppose that there are k regressor variables and n observations, (xi1, xi2, p , xik , yi), i  1, 2, p , n and that the model relating the regressors to the response is yi  0  1xi1  2 xi 2  p  k xik  i

i  1, 2, p , n

This model is a system of n equations that can be expressed in matrix notation as y ⴝ X␤ ⴙ ⑀

(12-11)

where

yⴝ ≥

y1 y2 o yn

¥

Xⴝ ≥

1 1

x11 x21

x12 x22

o 1

o

o

xn1

xn2

p p

x1k x2k o

p

¥

xnk

␤ⴝ ≥

0 1 o k

¥

and

⑀ⴝ ≥

1 2 o n

¥

In general, y is an (n 1) vector of the observations, X is an (n p) matrix of the levels of the independent variables (assuming that the intercept is always multiplied by a constant value—unity), ␤ is a ( p 1) vector of the regression coefficients, and ⑀ is a (n 1) vector of random errors. The X matrix is often called the model matrix. ˆ , that minimizes We wish to find the vector of least squares estimators, ␤ n

L  a 2i  ¿  1y  X␤2¿1y  X␤2 ˛

i1

ˆ is the solution for ␤ in the equations The least squares estimator ␤ L 0 ␤ We will not give the details of taking the derivatives above; however, the resulting equations that must be solved are Normal Equations

ˆ ⴝ Xⴕy XⴕX␤

(12-12)

JWCL232_c12_449-512.qxd

1/15/10

10:07 PM

Page 457

457

12-1 MULTIPLE LINEAR REGRESSION MODEL

Equations 12-12 are the least squares normal equations in matrix form. They are identical to the scalar form of the normal equations given earlier in Equations 12-10. To solve the normal equations, multiply both sides of Equations 12-12 by the inverse of X¿X. Therefore, the least squares estimate of ␤ is Least Square Estimate of ␤

ˆ ⴝ (XⴕX)⫺1 Xⴕy ␤

(12-13)

Note that there are p  k  1 normal equations in p  k  1 unknowns (the values of ˆ 0, ˆ 1, p , ˆ k 2. Furthermore, the matrix XⴕX is always nonsingular, as was assumed above, so the methods described in textbooks on determinants and matrices for inverting these matrices can be used to find 1X¿X2 1. In practice, multiple regression calculations are almost always performed using a computer. It is easy to see that the matrix form of the normal equations is identical to the scalar form. Writing out Equation 12-12 in detail, we obtain n

n

i1 n

n

a xi1

H

n

a xi1

i1

o n

a xik

i1

2

a xi1

i1

a xik

p

i1 n

i1 n

a xi1xi2

a yi

i1 n

i1

o n

a xik xi1

n

ˆ 0

a xi1xik ˆ 1 a xi1 yi X X H X  H i1 o o o

p

i1

o n i1

n

a xi2

n

a xik xi2

2

a xik

p

i1

i1

ˆ k

n

a xik yi ˛

i1

If the indicated matrix multiplication is performed, the scalar form of the normal equations (that is, Equation 12-10) will result. In this form it is easy to see that XⴕX is a ( p p) symmetric matrix and Xⴕy is a ( p 1) column vector. Note the special structure of the XⴕX matrix. The diagonal elements of XⴕX are the sums of squares of the elements in the columns of X, and the off-diagonal elements are the sums of cross-products of the elements in the columns of X. Furthermore, note that the elements of Xⴕy are the sums of cross-products of the columns of X and the observations 5 yi 6. The fitted regression model is k

yˆ i  ˆ 0  a ˆ j xi j ˛

˛

i  1, 2, p , n ˛

˛

(12-14)

j1

In matrix notation, the fitted model is ˆ yˆ  X␤

The difference between the observation yi and the fitted value yˆ i is a residual, say, ei  yi  yˆ i. The (n 1) vector of residuals is denoted by e  y  yˆ

(12-15)

JWCL232_c12_449-512.qxd

458

1/15/10

10:07 PM

Page 458

CHAPTER 12 MULTIPLE LINEAR REGRESSION

EXAMPLE 12-2

Wire Bond Strength with Matrix Notation

In Example 12-1, we illustrated fitting the multiple regression model y  0  1 x1  2 x2   ˛

where y is the observed pull strength for a wire bond, x1 is the wire length, and x2 is the die height. The 25 observations are in Table 12-2. We will now use the matrix approach to fit the regression model above to these data. The model matrix X and y vector for this model are

X

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

2 8 11 10 8 4 2 2 9 8 4 11 12 2 4 4 20 1 10 15 15 16 17 6 5

50 110 120 550 295 200 375 52 100 300 412 400 500 360 205 400 600 585 540 250 290 510 590 100 400

y

9.95 24.45 31.75 35.00 25.02 16.86 14.38 9.60 24.35 27.50 17.08 37.00 41.95 11.66 21.65 17.89 69.00 10.30 34.93 46.59 44.88 54.12 56.63 22.13 21.15

The X¿X matrix is

1 X¿X  £ 2 50

1 8 110

25  £ 206 8,294

p p p

206 2,396 77,177

1 2 1 1 8 5 §≥ o o 400 1 5 ˛

8,294 77,177 § 3,531,848

50 110 ¥ o 400

and the X¿y vector is

1 X¿y  £ 2 50

1 8 110

9.95 1 725.82 24.45 5 §≥ ¥  £ 8,008.47 § o 400 274,816.71 21.15

p p p

The least squares estimates are found from Equation 12-13 as ˆ ⴝ (XⴕX)⫺1Xⴕy ␤ or 25 ˆ 0 £ ˆ 1 §  £ 206 ˆ 2 8,294 0.214653  £ 0.007491 0.000340

206 2,396 77,177

1 8,294 725.82 77,177 § £ 8,008.37 § 3,531,848 274,811.31

0.007491 0.001671 0.000019

0.000340 725.82 0.000019 § £ 8,008.47 § 0.0000015 274,811.31

2.26379143  £ 2.74426964 § 0.01252781

Therefore, the fitted regression model with the regression coefficients rounded to five decimal places is yˆ  2.26379  2.74427x1  0.01253x2 This is identical to the results obtained in Example 12-1. This regression model can be used to predict values of pull strength for various values of wire length (x1) and die height (x2). We can also obtain the fitted values yˆ i by substituting each observation (xi1, xi2), i  1, 2, . . . , n, into the equation. For example, the first observation has x11  2 and x12  50, and the fitted value is yˆ1  2.26379  2.74427x11  0.01253x12  2.26379  2.74427122  0.012531502  8.38 The corresponding observed value is y1  9.95. The residual corresponding to the first observation is e1  y1  yˆ1  9.95  8.38  1.57 Table 12-3 displays all 25 fitted values yˆi and the corresponding residuals. The fitted values and residuals are calculated to the same accuracy as the original data.

JWCL232_c12_449-512.qxd

1/15/10

10:07 PM

Page 459

459

12-1 MULTIPLE LINEAR REGRESSION MODEL

Table 12-3 Observations, Fitted Values, and Residuals for Example 12-2 Observation Number

yi

yˆi

ei  yi  yˆi

1 2 3 4 5 6 7 8 9 10 11 12 13

9.95 24.45 31.75 35.00 25.02 16.86 14.38 9.60 24.35 27.50 17.08 37.00 41.95

8.38 25.60 33.95 36.60 27.91 15.75 12.45 8.40 28.21 27.98 18.40 37.46 41.46

1.57 1.15 2.20 1.60 2.89 1.11 1.93 1.20 3.86 0.48 1.32 0.46 0.49

Observation Number

yi

yˆi

ei  yi  yˆi

14 15 16 17 18 19 20 21 22 23 24 25

11.66 21.65 17.89 69.00 10.30 34.93 46.59 44.88 54.12 56.63 22.13 21.15

12.26 15.81 18.25 64.67 12.34 36.47 46.56 47.06 52.56 56.31 19.98 21.00

0.60 5.84 0.36 4.33 2.04 1.54 0.03 2.18 1.56 0.32 2.15 0.15

Computers are almost always used in fitting multiple regression models. Table 12-4 presents some annotated output from Minitab for the least squares regression model for wire bond pull strength data. The upper part of the table contains the numerical estimates of the regression coefficients. The computer also calculates several other quantities that reflect important information about the regression model. In subsequent sections, we will define and explain the quantities in this output. Estimating ␴2 Just as in simple linear regression, it is important to estimate 2, the variance of the error term , in a multiple regression model. Recall that in simple linear regression the estimate of 2 was obtained by dividing the sum of the squared residuals by n  2. Now there are two parameters in the simple linear regression model, so in multiple linear regression with p parameters a logical estimator for 2 is Estimator of Variance

n 2

a ei ˛

SSE ˆ 2  n  p  n  p i1

(12-16)

This is an unbiased estimator of  2. Just as in simple linear regression, the estimate of 2 is usually obtained from the analysis of variance for the regression model. The numerator of Equation 12-16 is called the error or residual sum of squares, and the denominator n  p is called the error or residual degrees of freedom. We can find a computing formula for SSE as follows: n

n

i1

i1

SSE  a 1 yi  yˆ i 2 2  a e2i  e¿e ˆ into the above, we obtain Substituting e  y  yˆ  y  X␤ ˆ ¿X¿y SSE  y¿y  ␤  27,178.5316  27,063.3581  115.174

(12-17)

JWCL232_c12_449-512.qxd

460

1/15/10

10:07 PM

Page 460

CHAPTER 12 MULTIPLE LINEAR REGRESSION

Table 12-4 Minitab Multiple Regression Output for the Wire Bond Pull Strength Data Regression Analysis: Strength versus Length, Height The regression equation is Strength  2.26  2.74 Length  0.0125 Height Predictor Constant Length Height

ˆ 0 ˆ 1 ˆ 2

Coef 2.264 2.74427 0.012528

S  2.288 PRESS  156.163

SE Coef 1.060 0.09352 0.002798

T 2.14 29.34 4.48

R-Sq  98.1% R-Sq (pred)  97.44%

P 0.044 0.000 0.000

VIF 1.2 1.2

R-Sq (adj)  97.9%

Analysis of Variance Source Regression Residual Error Total

DF 2 22 24

Source Length Height

Seq SS 5885.9 104.9

DF 1 1

SS 5990.8 115.2 6105.9

MS 2995.4 5.2

F 572.17

P 0.000

ˆ 2

Predicted Values for New Observations New Obs 1

Fit 27.663

SE Fit 0.482

95.0% CI (26.663, 28.663)

95.0% PI (22.814, 32.512)

Values of Predictors for New Observations New Obs 1

Length 8.00

Height 275

Table 12-4 shows that the estimate of 2 for the wire bond pull strength regression model is ˆ  115.2兾22  5.2364. The Minitab output rounds the estimate to ˆ 2  5.2. 2

12-1.4 Properties of the Least Squares Estimators The statistical properties of the least squares estimators ˆ 0, ˆ 1, p , ˆ k may be easily found, under certain assumptions on the error terms 1, 2, p , n, in the regression model. Paralleling the assumptions made in Chapter 11, we assume that the errors i are statistically independent with mean zero and variance 2. Under these assumptions, the least squares estimators ˆ 0, ˆ 1, p , ˆ k are unbiased estimators of the regression coefficients 0, 1, p , k. This property may be shown as follows: ˛

ˆ 2  E3 1X¿X2 1X¿Y4 E1␤  E3 1X¿X2 1X¿1X␤  ⑀2 4  E3 1X¿X2 1X¿X␤  1X¿X2 1X¿⑀4 ␤ ˆ is an unbiased estimator of ␤. since E(⑀) ⴝ 0 and (X X)1X X ⴝ I, the identity matrix. Thus, ␤

JWCL232_c12_449-512.qxd

1/16/10

10:47 AM

Page 461

12-1 MULTIPLE LINEAR REGRESSION MODEL

461

ˆ ’s are expressed in terms of the elements of the inverse of the X¿X The variances of the ␤ matrix. The inverse of X¿X times the constant ␴2 represents the covariance matrix of the regression coefficients ␤ˆ . The diagonal elements of ␴2 1X¿X2 ⫺1 are the variances of ␤ˆ 0, ␤ˆ 1, p , ␤ˆ k, and the off-diagonal elements of this matrix are the covariances. For example, if we have k ⫽ 2 regressors, such as in the pull-strength problem,

C00 C ⫽ 1X¿X2 ⫺1 ⫽ £ C10 C20

C01 C11 C21

C02 C12 § C22

which is symmetric (C10 ⫽ C01, C20 ⫽ C02, and C21 ⫽ C12) because (X⬘X)⫺1 is symmetric, and we have V 1␤ˆ j 2 ⫽ ␴2C jj, ˛

˛

cov1␤ˆ i, ␤ˆ j 2 ⫽ ␴ C ij, 2

˛

j ⫽ 0, 1, 2 i⫽j

ˆ is a ( p ⫻ p) symmetric matrix whose jjth element is the In general, the covariance matrix of ␤ ˆ variance of ␤j and whose i, jth element is the covariance between ␤ˆ i and ␤ˆ j, that is, ˆ 2 ⫽ ␴2 1X¿X2 ⫺1 ⫽ ␴2 C cov1␤

The estimates of the variances of these regression coefficients are obtained by replacing ␴2 with an estimate. When ␴2 is replaced by its estimate ␴ˆ 2, the square root of the estimated variance of the jth regression coefficient is called the estimated standard error of ␤ˆ j or se 1␤ˆ j 2 ⫽ 2␴ˆ 2Cjj . These standard errors are a useful measure of the precision of estimation for the regression coefficients; small standard errors imply good precision. Multiple regression computer programs usually display these standard errors. For example, the Minitab output in Table 12-4 reports se 1␤ˆ 0 2 ⫽ 1.060, se 1␤ˆ 1 2 ⫽ 0.09352, and se1␤ˆ 2 2 ⫽ 0.002798. The intercept estimate is about twice the magnitude of its standard error, and ␤ˆ 1 and ␤ˆ 2 are considerably larger than se 1␤ˆ 1 2 and se 1␤ˆ 2 2. This implies reasonable precision of estimation, although the parameters ␤1 and ␤2 are much more precisely estimated than the intercept (this is not unusual in multiple regression).

EXERCISES FOR SECTION 12-1 12-1. A study was performed to investigate the shear strength of soil ( y) as it related to depth in feet (x1) and % moisture content (x2). Ten observations were collected, and the following summary quantities obtained: n ⫽ 10, gxi1 ⫽ 223, gxi2 ⫽ 553, g yi ⫽ 1,916, g x2i1 ⫽ 5,200.9, gx2i2 ⫽ 31,729, g xi1xi2 ⫽ 12,352, g xi1 yi ⫽ 43,550.8, gxi2 yi ⫽ 104,736.8, and g y2i ⫽ 371,595.6. (a) Set up the least squares normal equations for the model Y ⫽ ␤0 ⫹ ␤1x1 ⫹ ␤2x2 ⫹ ⑀. (b) Estimate the parameters in the model in part (a). (c) What is the predicted strength when x1 ⫽ 18 feet and x2 ⫽ 43%? 12-2. A regression model is to be developed for predicting the ability of soil to absorb chemical contaminants. Ten observations have been taken on a soil absorption index ( y) and two regressors: x1 ⫽ amount of extractable iron ore and x2 ⫽

amount of bauxite. We wish to fit the model Y ⫽ ␤0 ⫹ ␤1x1 ⫹ ␤2x2 ⫹ ⑀. Some necessary quantities are: 1.17991 ⫺7.30982 E-3 7.3006 E-4 1XⴕX2 ⫺1 ⫽ £⫺7.30982 E-3 7.9799 E-5 ⫺1.23713 E-4§, 7.3006 E-4 ⫺1.23713 E-4 4.6576 E-4 220 Xⴕy ⫽ £ 36,768 § 9,965 (a) Estimate the regression coefficients in the model specified above. (b) What is the predicted value of the absorption index y when x1 ⫽ 200 and x2 ⫽ 50? 12-3. A chemical engineer is investigating how the amount of conversion of a product from a raw material (y) depends on

JWCL232_c12_449-512.qxd

462

1/15/10

10:07 PM

Page 462

CHAPTER 12 MULTIPLE LINEAR REGRESSION

reaction temperature (x1) and the reaction time (x2). He has developed the following regression models: 1. yˆ  100  2 x1  4 x2 2. yˆ  95  1.5x1  3x2  2 x1x2 Both models have been built over the range 0.5 x2 10. (a) What is the predicted value of conversion when x2  2? Repeat this calculation for x2  8. Draw a graph of the predicted values for both conversion models. Comment on the effect of the interaction term in model 2. (b) Find the expected change in the mean conversion for a unit change in temperature x1 for model 1 when x2  5. Does this quantity depend on the specific value of reaction time selected? Why? (c) Find the expected change in the mean conversion for a unit change in temperature x1 for model 2 when x2  5. Repeat this calculation for x2  2 and x2  8. Does the result depend on the value selected for x2? Why? 12-4. You have fit a multiple linear regression model and the (X⬘X)1 matrix is: 0.893758 1XⴕX2 1  £0.028245 0.017564

0.0282448 0.0013329 0.0001547

0.0175641 0.0001547 § 0.0009108

(a) How many regressor variables are in this model? (b) If the error sum of squares is 307 and there are 15 observations, what is the estimate of 2? (c) What is the standard error of the regression coefficient ˆ 1? 12-5. Data from a patient satisfaction survey in a hospital are shown in the following table:

Observation

Age

Severity

Surg-Med

Anxiety

Satisfaction

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

55 46 30 35 59 61 74 38 27 51 53 41 37 24 42 50

50 24 46 48 58 60 65 42 42 50 38 30 31 34 30 48

0 1 1 1 0 0 1 1 0 1 1 0 0 0 0 1

2.1 2.8 3.3 4.5 2.0 5.1 5.5 3.2 3.1 2.4 2.2 2.1 1.9 3.1 3.0 4.2

68 77 96 80 43 44 26 88 75 57 56 88 88 102 88 70

17 18 19 20 21 22 23 24 25

58 60 62 68 70 79 63 39 49

61 71 62 38 41 66 31 42 40

1 1 0 0 1 1 1 0 1

4.6 5.3 7.2 7.8 7.0 6.2 4.1 3.5 2.1

52 43 46 56 59 26 52 83 75

The regressor variables are the patient’s age, an illness severity index (larger values indicate greater severity), an indicator variable denoting whether the patient is a medical patient (0) or a surgical patient (1), and an anxiety index (larger values indicate greater anxiety). (a) Fit a multiple linear regression model to the satisfaction response using age, illness severity, and the anxiety index as the regressors. (b) Estimate 2. (c) Find the standard errors of the regression coefficients. (d) Are all of the model parameters estimated with nearly the same precision? Why or why not? 12-6. The electric power consumed each month by a chemical plant is thought to be related to the average ambient temperature (x1), the number of days in the month (x2), the average product purity (x3), and the tons of product produced (x4). The past year’s historical data are available and are presented in the following table:

y

x1

x2

x3

x4

240 236 270 274 301 316 300 296 267 276 288 261

25 31 45 60 65 72 80 84 75 60 50 38

24 21 24 25 25 26 25 25 24 25 25 23

91 90 88 87 91 94 87 86 88 91 90 89

100 95 110 88 94 99 97 96 110 105 100 98

(a) Fit a multiple linear regression model to these data. (b) Estimate 2.

JWCL232_c12_449-512.qxd

1/16/10

10:48 AM

Page 463

463

12-1 MULTIPLE LINEAR REGRESSION MODEL

Table 12-5 DaimlerChrysler Fuel Economy and Emissions mfr carline

car/truck

cid

rhp

trns

drv

etw

cmp

axle

n/v

a/c

hc

co

co2 mpg

20

300C/SRT-8

C

215

253

L5

2

4500

9.9

3.07

30.9

Y

0.011

0.09

288

20

CARAVAN 2WD

T

201

180

L4

F

2

4500

9.3

2.49

32.3

Y

0.014

0.11

274

32.5

20

CROSSFIRE ROADSTER

C

196

168

L5

R

2

3375

10

3.27

37.1

Y

0.001

0.02

250

35.4

20

DAKOTA PICKUP 2WD

T

226

210

L4

R

2

4500

9.2

3.55

29.6

Y

0.012

0.04

316

28.1

20

DAKOTA PICKUP 4WD

T

226

210

L4

2

5000

9.2

3.55

29.6

Y

0.011

0.05

365

24.4

20

DURANGO 2WD

T

348

345

L5

R

2

5250

8.6

3.55

27.2

Y

0.023

0.15

367

24.1

20

GRAND CHEROKEE 2WD

T

226

210

L4

R

2

4500

9.2

3.07

30.4

Y

0.006

0.09

312

28.5

20

GRAND CHEROKEE 4WD

T

348

230

L5

4

2

5000

9

3.07

24.7

Y

0.008

0.11

369

24.2

20

LIBERTY/CHEROKEE 2WD

T

148

150

M6 R

2

4000

9.5

4.1

41

Y

0.004

0.41

270

32.8

20

LIBERTY/CHEROKEE 4WD

T

226

210

L4

4

2

4250

9.2

3.73

31.2

Y

0.003

0.04

317

28

20

NEON/SRT-4/SX 2.0

C

122

132

L4

F

2

3000

9.8

2.69

39.2

Y

0.003

0.16

214

41.3

20

PACIFICA 2WD

T

215

249

L4

F

2

4750

9.9

2.95

35.3

Y

0.022

0.01

295

30

20

PACIFICA AWD

T

215

249

L4

2

5000

9.9

2.95

35.3

Y

0.024

0.05

314

28.2

20

PT CRUISER

T

148

220

L4

F

2

3625

9.5

2.69

37.3

Y

0.002

0.03

260

34.1

20

RAM 1500 PICKUP 2WD

T

500

500

M6 R

2

5250

9.6

4.1

22.3

Y

0.01

0.1

474

18.7

20

RAM 1500 PICKUP 4WD

T

348

345

L5

2

6000

8.6

3.92

29

Y

0

0

0

20.3

20

SEBRING 4-DR

C

165

200

L4

F

2

3625

9.7

2.69

36.8

Y

0.011

0.12

252

35.1

20

STRATUS 4-DR

C

148

167

L4

F

2

3500

9.5

2.69

36.8

Y

0.002

0.06

233

37.9

20

TOWN & COUNTRY 2WD

T

148

150

L4

F

2

4250

9.4

2.69

34.9

Y

0

0.09

262

33.8

20

VIPER CONVERTIBLE

C

500

501

M6 R

2

3750

9.6

3.07

19.4

Y

0.007

0.05

342

25.9

20

WRANGLER/TJ 4WD

T

148

150

M6

2

3625

9.5

3.73

40.1

Y

0.004

0.43

337

26.4

mfr-mfr code carline-car line name (test vehicle model name) car/truck-‘C’ for passenger vehicle and ‘T’ for truck cid-cubic inch displacement of test vehicle rhp-rated horsepower trns-transmission code drv-drive system code od-overdrive code etw-equivalent test weight

(c) Compute the standard errors of the regression coefficients. Are all of the model parameters estimated with the same precision? Why or why not? (d) Predict power consumption for a month in which x1 ⫽ 75⬚F, x2 ⫽ 24 days, x3 ⫽ 90%, and x4 ⫽ 98 tons. 12-7. Table 12-5 provides the highway gasoline mileage test results for 2005 model year vehicles from DaimlerChrysler. The full table of data (available on the book’s Web site) contains the same data for 2005 models from over 250 vehicles from many manufacturers (source: Environmental Protection Agency Web site www.epa.gov/ otaq/cert/mpg/testcars/database). (a) Fit a multiple linear regression model to these data to estimate gasoline mileage that uses the following regressors: cid, rhp, etw, cmp, axle, n/v.

4

od

4

4

4

4

30.8

cmp-compression ratio axle-axle ratio n/v-n/v ratio (engine speed versus vehicle speed at 50 mph) a/c-indicates air conditioning simulation hc-HC(hydrocarbon emissions) Test level composite results co-CO(carbon monoxide emissions) Test level composite results co2-CO2(carbon dioxide emissions) Test level composite results mpg-mpg(fuel economy, miles per gallon)

(b) Estimate ␴2 and the standard errors of the regression coefficients. (c) Predict the gasoline mileage for the first vehicle in the table. 12-8. The pull strength of a wire bond is an important characteristic. The following table gives information on pull strength ( y), die height (x1), post height (x2), loop height (x3), wire length (x4), bond width on the die (x5), and bond width on the post (x6). (a) Fit a multiple linear regression model using x2, x3, x4, and x5 as the regressors. (b) Estimate ␴2. (c) Find the se(␤ˆ j). How precisely are the regression coefficients estimated, in your opinion?

JWCL232_c12_449-512.qxd

464

1/15/10

10:07 PM

Page 464

CHAPTER 12 MULTIPLE LINEAR REGRESSION

y

x1

x2

x3

x4

x5

x6

8.0 8.3 8.5 8.8 9.0 9.3 9.3 9.5 9.8 10.0 10.3 10.5 10.8 11.0 11.3 11.5 11.8 12.3 12.5

5.2 5.2 5.8 6.4 5.8 5.2 5.6 6.0 5.2 5.8 6.4 6.0 6.2 6.2 6.2 5.6 6.0 5.8 5.6

19.6 19.8 19.6 19.4 18.6 18.8 20.4 19.0 20.8 19.9 18.0 20.6 20.2 20.2 19.2 17.0 19.8 18.8 18.6

29.6 32.4 31.0 32.4 28.6 30.6 32.4 32.6 32.2 31.8 32.6 33.4 31.8 32.4 31.4 33.2 35.4 34.0 34.2

94.9 89.7 96.2 95.6 86.5 84.5 88.8 85.7 93.6 86.0 87.1 93.1 83.4 94.5 83.4 85.2 84.1 86.9 83.0

2.1 2.1 2.0 2.2 2.0 2.1 2.2 2.1 2.3 2.1 2.0 2.1 2.2 2.1 1.9 2.1 2.0 2.1 1.9

2.3 1.8 2.0 2.1 1.8 2.1 1.9 1.9 2.1 1.8 1.6 2.1 2.1 1.9 1.8 2.1 1.8 1.8 2.0

(d) Use the model from part (a) to predict pull strength when x2 = 20, x3 = 30, x4 = 90, and x5 = 2.0. 12-9. An engineer at a semiconductor company wants to model the relationship between the device HFE ( y) and three parameters: Emitter-RS (x1), Base-RS (x2), and Emitter-to-Base RS (x3). The data are shown in the following table.

x1 Emitter-RS

x2 Base-RS

x3 E-B-RS

y HFE-1M-5V

14.620 15.630 14.620 15.000 14.500 15.250 16.120 15.130 15.500 15.130 15.500 16.120 15.130 15.630 15.380 14.380

226.00 220.00 217.40 220.00 226.50 224.10 220.50 223.50 217.60 228.50 230.20 226.50 226.60 225.60 229.70 234.00

7.000 3.375 6.375 6.000 7.625 6.000 3.375 6.125 5.000 6.625 5.750 3.750 6.125 5.375 5.875 8.875

128.40 52.62 113.90 98.01 139.90 102.60 48.14 109.60 82.68 112.60 97.52 59.06 111.80 89.09 101.00 171.90

15.500 14.250 14.500 14.620

230.00 224.30 240.50 223.70

4.000 8.000 10.870 7.375

66.80 157.10 208.40 133.40

(a) Fit a multiple linear regression model to the data. (b) Estimate 2. (c) Find the standard errors se 1ˆ j 2. Are all of the model parameters estimated with the same precision? Justify your answer. (d) Predict HFE when x1  14.5, x2  220, and x3  5.0. 12-10. Heat treating is often used to carburize metal parts, such as gears. The thickness of the carburized layer is considered a crucial feature of the gear and contributes to the overall reliability of the part. Because of the critical nature of this feature, two different lab tests are performed on each furnace load. One test is run on a sample pin that accompanies each load. The other test is a destructive test, where an actual part is cross-sectioned. This test involves running a carbon analysis on the surface of both the gear pitch (top of the gear tooth) and the gear root (between the gear teeth). Table 12-6 shows the results of the pitch carbon analysis test for 32 parts. The regressors are furnace temperature (TEMP), carbon concentration and duration of the carburizing cycle (SOAKPCT, SOAKTIME), and carbon concentration and duration of the diffuse cycle (DIFFPCT, DIFFTIME). (a) Fit a linear regression model relating the results of the pitch carbon analysis test (PITCH) to the five regressor variables. (b) Estimate 2. (c) Find the standard errors se 1ˆ j 2. (d) Use the model in part (a) to predict PITCH when TEMP  1650, SOAKTIME  1.00, SOAKPCT  1.10, DIFFTIME  1.00, and DIFFPCT  0.80. 12-11. An article in Electronic Packaging and Production (2002, Vol. 42) considered the effect of X-ray inspection of integrated circuits. The rads (radiation dose) were studied as a function of current (in milliamps) and exposure time (in minutes). Rads

mAmps

Exposure Time

7.4 14.8 29.6 59.2 88.8 296 444 592 11.1 22.2 44.4

10 10 10 10 10 10 10 10 15 15 15

0.25 0.5 1 2 3 10 15 20 0.25 0.5 1 continued

JWCL232_c12_449-512.qxd

1/15/10

10:07 PM

Page 465

12-1 MULTIPLE LINEAR REGRESSION MODEL

Rads

mAmps

88.8 133.2 444 666 888 14.8 29.6 59.2 118.4 177.6 592 888 1184 22.2 44.4 88.8 177.6 266.4 888 1332 1776 29.6 59.2 118.4 236.8 355.2 1184 1776 2368

15 15 15 15 15 20 20 20 20 20 20 20 20 30 30 30 30 30 30 30 30 40 40 40 40 40 40 40 40

Exposure Time 2 3 10 15 20 0.25 0.5 1 2 3 10 15 20 0.25 0.5 1 2 3 10 15 20 0.25 0.5 1 2 3 10 15 20

(a) Fit a multiple linear regression model to these data with rads as the response. (b) Estimate 2 and the standard errors of the regression coefficients. (c) Use the model to predict rads when the current is 15 milliamps and the exposure time is 5 seconds. 12-12. An article in Cancer Epidemiology, Biomarkers and Prevention (1996, Vol. 5, pp. 849–852) conducted a pilot study to assess the use of toenail arsenic concentrations as an indicator of ingestion of arsenic-containing water. Twenty-one participants were interviewed regarding use of their private (unregulated) wells for drinking and cooking, and each provided a sample of water and toenail clippings. The table below showed the data of age (years), sex of person (1  male, 2  female), proportion of times household well used for drinking (1 1/4, 2  1/4, 3  1/2, 4  3/4, 5  3/4), proportion of times household well used for cooking (1 1/4, 2  1/4, 3  1/2, 4  3/4, 5  3/4), arsenic in water (ppm), and arsenic in toenails (ppm) respectively.

465

Age

Sex

Drink Use

Cook Use

Arsenic Water

Arsenic Nails

44 45 44 66 37 45 47 38 41 49 72 45 53 86 8 32 44 63 42 62 36

2 2 1 2 1 2 1 2 2 2 2 2 1 2 2 2 1 2 1 1 1

5 4 5 3 2 5 5 4 3 4 5 1 5 5 5 5 5 5 5 5 5

5 5 5 5 5 5 5 5 2 5 5 5 5 5 5 5 5 5 5 5 5

0.00087 0.00021 0 0.00115 0 0 0.00013 0.00069 0.00039 0 0 0.046 0.0194 0.137 0.0214 0.0175 0.0764 0 0.0165 0.00012 0.0041

0.119 0.118 0.099 0.118 0.277 0.358 0.08 0.158 0.31 0.105 0.073 0.832 0.517 2.252 0.851 0.269 0.433 0.141 0.275 0.135 0.175

(a) Fit a multiple linear regression model using arsenic concentration in nails as the response and age, drink use, cook use, and arsenic in the water as the regressors. (b) Estimate 2 and the standard errors of the regression coefficients. (c) Use the model to predict the arsenic in nails when the age is 30, the drink use is category 5, the cook use is category 5, and arsenic in the water is 0.135 ppm. 12-13. In an article in IEEE Transactions on Instrumentation and Measurement (2001, Vol. 50, pp. 2033–2040) powdered mixtures of coal and limestone were analyzed for permittivity. The errors in the density measurement was the response. Density

Dielectric Constant

Loss Factor

0.749 0.798 0.849 0.877 0.929 0.963 0.997 1.046 1.133 1.17 1.215

2.05 2.15 2.25 2.3 2.4 2.47 2.54 2.64 2.85 2.94 3.05

0.016 0.02 0.022 0.023 0.026 0.028 0.031 0.034 0.039 0.042 0.045

JWCL232_c12_449-512.qxd

466

1/15/10

10:07 PM

Page 466

CHAPTER 12 MULTIPLE LINEAR REGRESSION

Table 12-6 TEMP

SOAKTIME

SOAKPCT

DIFFTIME

DIFFPCT

PITCH

1650 1650 1650 1650 1600 1600 1650 1650 1650 1650 1650 1650 1650 1650 1650 1650 1650 1650 1650 1650 1650 1650 1650 1650 1650 1650 1650 1700 1650 1650 1700 1700

0.58 0.66 0.66 0.66 0.66 0.66 1.00 1.17 1.17 1.17 1.17 1.17 1.17 1.20 2.00 2.00 2.20 2.20 2.20 2.20 2.20 2.20 3.00 3.00 3.00 3.00 3.33 4.00 4.00 4.00 12.50 18.50

1.10 1.10 1.10 1.10 1.15 1.15 1.10 1.10 1.10 1.10 1.10 1.10 1.15 1.15 1.15 1.10 1.10 1.10 1.15 1.10 1.10 1.10 1.15 1.10 1.10 1.15 1.10 1.10 1.10 1.15 1.00 1.00

0.25 0.33 0.33 0.33 0.33 0.33 0.50 0.58 0.58 0.58 0.58 0.58 0.58 1.10 1.00 1.10 1.10 1.10 1.10 1.10 1.10 1.50 1.50 1.50 1.50 1.66 1.50 1.50 1.50 1.50 1.50 1.50

0.90 0.90 0.90 0.95 1.00 1.00 0.80 0.80 0.80 0.80 0.90 0.90 0.90 0.80 0.80 0.80 0.80 0.80 0.80 0.90 0.90 0.90 0.80 0.70 0.75 0.85 0.80 0.70 0.70 0.85 0.70 0.70

0.013 0.016 0.015 0.016 0.015 0.016 0.014 0.021 0.018 0.019 0.021 0.019 0.021 0.025 0.025 0.026 0.024 0.025 0.024 0.025 0.027 0.026 0.029 0.030 0.028 0.032 0.033 0.039 0.040 0.035 0.056 0.068

(a) Fit a multiple linear regression model to these data with the density as the response. (b) Estimate 2 and the standard errors of the regression coefficients. (c) Use the model to predict the density when the dielectric constant is 2.5 and the loss factor is 0.03. 12-14. An article in Biotechnology Progress (2001, Vol. 17, pp. 366–368) reported on an experiment to investigate

and optimize nisin extraction in aqueous two-phase systems (ATPS). The nisin recovery was the dependent variable ( y). The two regressor variables were concentration (%) of PEG 4000 (denoted as x1 2 and concentration (%) of Na2SO4 (denoted as x2 2. (a) Fit a multiple linear regression model to these data. (b) Estimate 2 and the standard errors of the regression coefficients.

JWCL232_c12_449-512.qxd

1/15/10

10:07 PM

Page 467

12-1 MULTIPLE LINEAR REGRESSION MODEL

x1

x2

y

13 15 13 15 14 14 14 14 14

11 11 13 13 12 12 12 12 12

62.8739 76.1328 87.4667 102.3236 76.1872 77.5287 76.7824 77.4381 78.7417

(c) Use the model to predict the nisin recovery when x1  14.5 and x2  12.5. 12-15. An article in Optical Engineering [“Operating Curve Extraction of a Correlator’s Filter” (2004, Vol. 43, pp. 2775–2779)] reported on use of an optical correlator to perform an experiment by varying brightness and contrast. The resulting modulation is characterized by the useful range of gray levels. The data are shown below: Brightness (%): 54 61 65 100 100 100 50 57 54 Contrast (%): 56 80 70 50 65 80 25 35 26 Useful range (ng): 96 50 50 112 96 80 155 144 255 (a) (b) (c) (d)

Fit a multiple linear regression model to these data. Estimate 2. Compute the standard errors of the regression coefficients. Predict the useful range when brightness  80 and contrast  75. 12-16. An article in Technometrics (1974, Vol. 16, pp. 523–531) considered the following stack-loss data from a plant oxidizing ammonia to nitric acid. Twenty-one daily responses of stack loss y (the amount of ammonia escaping) were measured with air flow x1, temperature x2, and acid concentration x3. Stack loss y  42, 37, 37, 28, 18, 18, 19, 20, 15, 14, 14, 13, 11, 12, 8, 7, 8, 8, 9, 15, 15 x1  80, 80, 75, 62, 62, 62, 62, 62, 58, 58, 58, 58, 58, 58, 50, 50, 50, 50, 50, 56, 70 x2  27, 27, 25, 24, 22, 23, 24, 24, 23, 18, 18, 17, 18, 19, 18, 18, 19, 19, 20, 20, 20 x3  89, 88, 90, 87, 87, 87, 93, 93, 87, 80, 89, 88, 82, 93, 89, 86, 72, 79, 80, 82, 91 (a) Fit a linear regression model relating the results of the stack loss to the three regressor varilables. (b) Estimate 2. (c) Find the standard error se1ˆ j 2. (d) Use the model in part (a) to predict stack loss when x1  60, x2  26, and x3  85.

467

12-17. Table 12-7 presents quarterback ratings for the 2008 National Football League season (source: The Sports Network). (a) Fit a multiple regression model to relate the quarterback rating to the percentage of completions, the percentage of TDs, and the percentage of interceptions. (b) Estimate 2. (c) What are the standard errors of the regression coefficients? (d) Use the model to predict the rating when the percentage of completions is 60%, the percentage of TDs is 4%, and the percentage of interceptions is 3%. 12-18. Table 12-8 presents statistics for the National Hockey League teams from the 2008–2009 season (source: The Sports Network). Fit a multiple linear regression model that relates Wins to the variables GF through FG. Because teams play 82 games W  82  L  T  OTL, but such a model does not help build a better team. Estimate 2 and find the standard errors of the regression coefficients for your model. 12-19. A study was performed on wear of a bearing y and its relationship to x1  oil viscosity and x2  load. The following data were obtained. y

x1

x2

293 230 172 91 113 125

1.6 15.5 22.0 43.0 33.0 40.0

851 816 1058 1201 1357 1115

(a) Fit a multiple linear regression model to these data. (b) Estimate 2 and the standard errors of the regression coefficients. (c) Use the model to predict wear when x1  25 and x2  1000. (d) Fit a multiple linear regression model with an interaction term to these data. (e) Estimate 2 and se(ˆ j) for this new model. How did these quantities change? Does this tell you anything about the value of adding the interaction term to the model? (f) Use the model in (d) to predict when x1  25 and x2  1000. Compare this prediction with the predicted value from part (c) above. 12-20. Consider the linear regression model Yi  ¿0  1 1xi1  x1 2  2 1xi2  x2 2  i

where x1  g xi1n and x2  g xi2n. (a) Write out the least squares normal equations for this model. (b) Verify that the least squares estimate of the intercept in this model is ˆ ¿0  g yin  y. (c) Suppose that we use yi  y as the response variable in the model above. What effect will this have on the least squares estimate of the intercept?

JWCL232_c12_449-512.qxd

468

1/21/10

9:46 PM

Page 468

CHAPTER 12 MULTIPLE LINEAR REGRESSION

Table 12-7 Quarterback Ratings for the 2008 National Football League Season Player Philip Chad Kurt Drew Peyton Aaron Matt Tony Jeff Matt Matt Shaun Seneca Eli Donovan Jay Trent Jake Jason David Brett Joe Kerry Ben Kyle JaMarcus Tyler Gus Dan Marc Ryan Derek Att Comp Pct Comp Yds Yds per Att TD Pct TD Long Int Pct Int Rating Pts

Rivers Pennington Warner Brees Manning Rodgers Schaub Romo Garcia Cassel Ryan Hill Wallace Manning McNabb Cutler Edwards Delhomme Campbell Garrard Favre Flacco Collins Roethlisberger Orton Russell Thigpen Frerotte Orlovsky Bulger Fitzpatrick Anderson

Team

Att

Comp

Pct Comp

Yds

Yds per Att

TD

Pct TD

Lng

Int

Pct Int

Rating Pts

SD MIA ARI NO IND GB HOU DAL TB NE ATL SF SEA NYG PHI DEN BUF CAR WAS JAC NYJ BAL TEN PIT CHI OAK KC MIN DET STL CIN CLE

478 476 598 635 555 536 380 450 376 516 434 288 242 479 571 616 374 414 506 535 522 428 415 469 465 368 420 301 255 440 372 283

312 321 401 413 371 341 251 276 244 327 265 181 141 289 345 384 245 246 315 335 343 257 242 281 272 198 230 178 143 251 221 142

65.3 67.4 67.1 65 66.8 63.6 66.1 61.3 64.9 63.4 61.1 62.8 58.3 60.3 60.4 62.3 65.5 59.4 62.3 62.6 65.7 60 58.3 59.9 58.5 53.8 54.8 59.1 56.1 57 59.4 50.2

4,009 3,653 4,583 5,069 4,002 4,038 3,043 3,448 2,712 3,693 3,440 2,046 1,532 3,238 3,916 4,526 2,699 3,288 3,245 3,620 3,472 2,971 2,676 3,301 2,972 2,423 2,608 2,157 1,616 2,720 1,905 1,615

8.39 7.67 7.66 7.98 7.21 7.53 8.01 7.66 7.21 7.16 7.93 7.10 6.33 6.76 6.86 7.35 7.22 7.94 6.41 6.77 6.65 6.94 6.45 7.04 6.39 6.58 6.21 7.17 6.34 6.18 5.12 5.71

34 19 30 34 27 28 15 26 12 21 16 13 11 21 23 25 11 15 13 15 22 14 12 17 18 13 18 12 8 11 8 9

7.1 4.0 5.0 5.4 4.9 5.2 3.9 5.8 3.2 4.1 3.7 4.5 4.5 4.4 4.0 4.1 2.9 3.6 2.6 2.8 4.2 3.3 2.9 3.6 3.9 3.5 4.3 4.0 3.1 2.5 2.2 3.2

67 80 79 84 75 71 65 75 71 76 70 48 90 48 90 93 65 65 67 41 56 70 56 65 65 84 75 99 96 80 79 70

11 7 14 17 12 13 10 14 6 11 11 8 3 10 11 18 10 12 6 13 22 12 7 15 12 8 12 15 8 13 9 8

2.3 1.5 2.3 2.7 2.2 2.4 2.6 3.1 1.6 2.1 2.5 2.8 1.2 2.1 1.9 2.9 2.7 2.9 1.2 2.4 4.2 2.8 1.7 3.2 2.6 2.2 2.9 5.0 3.1 3.0 2.4 2.8

105.5 97.4 96.9 96.2 95 93.8 92.7 91.4 90.2 89.4 87.7 87.5 87 86.4 86.4 86 85.4 84.7 84.3 81.7 81 80.3 80.2 80.1 79.6 77.1 76 73.7 72.6 71.4 70 66.5

Attempts (number of pass attempts) Completed passes Percentage of completed passes Yards gained passing Yards gained per pass attempt Number of touchdown passes Percentage of attempts that are touchdowns Longest pass completion Number of interceptions Percentage of attempts that are interceptions Rating points

JWCL232_c12_449-512.qxd

1/15/10

10:07 PM

Page 469

469

12-1 MULTIPLE LINEAR REGRESSION MODEL

Table 12-8 Team Statistics for the 2008–2009 National Hockey League Season Team

W

L

OTL PTS

GF

GA ADV PPGF PCTG PEN BMI AVG SHT PPGA PKPCT SHGF SHGA FG

Anaheim Atlanta Boston Buffalo Carolina Columbus Calgary Chicago Colorado Dallas Detroit Edmonton Florida Los Angeles Minnesota Montreal New Jersey Nashville NI Islanders NY Rangers Ottawa Philadelphia Phoenix Pittsburgh San Jose St. Louis Tampa Bay Toronto Vancouver Washington

42 35 53 41 45 41 46 46 32 36 51 38 41 34 40 41 51 40 26 43 36 44 36 45 53 41 24 34 45 50

33 41 19 32 30 31 30 24 45 35 21 35 30 37 33 30 27 34 47 30 35 27 39 28 18 31 40 35 27 24

7 6 10 9 7 10 6 12 5 11 10 9 11 11 9 11 4 8 9 9 11 11 7 9 11 10 18 13 10 8

238 250 270 242 236 220 251 260 190 224 289 228 231 202 214 242 238 207 198 200 213 260 205 258 251 227 207 244 243 268

235 279 190 229 221 223 246 209 253 251 240 244 223 226 197 240 207 228 274 212 231 232 249 233 199 227 269 286 213 240

W L OTL PTS

GF GA ADV PPGF PCTG

91 76 116 91 97 92 98 104 69 83 112 85 93 79 89 93 106 88 61 95 83 99 79 99 117 92 66 81 100 108

309 357 313 358 374 322 358 363 318 351 353 354 308 360 328 374 307 318 320 346 339 316 344 360 360 351 343 330 357 337

73 69 74 75 70 41 61 70 50 54 90 60 51 69 66 72 58 50 54 48 66 71 50 62 87 72 61 62 67 85

23.6 19.3 23.6 21 18.7 12.7 17 19.3 15.7 15.4 25.5 17 16.6 19.2 20.1 19.2 18.9 15.7 16.9 13.9 19.5 22.5 14.5 17.2 24.2 20.5 17.8 18.8 18.8 25.2

Wins Losses during regular time Overtime losses Points. Two points for winning a game, one point for a tie or losing in overtime, zero points for losing in regular time. Goals for Goals against Total advantages. Power play opportunities. Power-play goals for. Goals scored while on power play. Power play percentage. Power-play goals divided by total advantages.

1418 1244 1016 1105 786 1207 1281 1129 1044 1134 810 1227 884 1191 869 1223 1038 982 1198 1175 1084 1408 1074 1106 1037 1226 1280 1113 1323 1021 PEN BMI AVG SHT PPGA PKPCT

SHGF SHGA FG

8 12 12 16 16 20 18 28 18 10 14 20 16 16 20 6 20 12 18 24 14 26 18 8 16 22 26 12 28 20

17.4 15.3 12.5 13.7 9.8 15 15.8 14.1 13 14 10 15.2 11 14.7 10.8 15 12.9 12.1 14.8 14.6 13.4 17.5 13.3 13.6 12.8 15.2 15.9 13.7 16.5 12.7

385 366 306 336 301 346 349 330 318 327 327 338 311 362 291 370 324 338 361 329 346 393 293 347 306 357 405 308 371 387

78 88 54 61 59 62 58 64 64 70 71 76 54 62 36 65 65 59 73 40 64 67 68 60 51 58 89 78 69 75

79.7 76 82.4 81.8 80.4 82.1 83.4 80.6 79.9 78.6 78.3 77.5 82.6 82.9 87.6 82.4 79.9 82.5 79.8 87.8 81.5 83 76.8 82.7 83.3 83.8 78 74.7 81.4 80.6

6 13 8 7 8 8 6 10 4 2 6 3 7 4 9 10 12 9 12 9 8 16 5 7 12 10 4 6 7 7

6 9 7 4 7 9 13 5 5 2 4 8 6 7 6 10 3 8 5 13 5 1 4 11 10 8 8 7 5 9

43 39 47 44 39 41 37 43 31 38 46 39 39 39 39 38 44 41 37 42 46 43 36 46 46 35 34 40 47 45

Total penalty minutes including bench minutes Total bench minor minutes Average penalty minutes per game Total times short-handed. Measures opponent opportunities. Power-play goals against Penalty killing percentage. Measures a team’s ability to prevent goals while its opponent is on a power play. Opponent opportunities minus power play goals divided by opponent’s opportunities. Short-handed goals for Short-handed goals against Games scored first

JWCL232_c12_449-512.qxd

470

1/15/10

10:07 PM

Page 470

CHAPTER 12 MULTIPLE LINEAR REGRESSION

12-2 HYPOTHESIS TESTS IN MULTIPLE LINEAR REGRESSION In multiple linear regression problems, certain tests of hypotheses about the model parameters are useful in measuring model adequacy. In this section, we describe several important hypothesis-testing procedures. As in the simple linear regression case, hypothesis testing requires that the error terms i in the regression model are normally and independently distributed with mean zero and variance 2.

12-2.1 Test for Significance of Regression The test for significance of regression is a test to determine whether a linear relationship exists between the response variable y and a subset of the regressor variables x1, x2, p , xk. The appropriate hypotheses are Hypotheses for ANOVA Test

H0: 1  2  # # #  k  0 H1: j Z 0 for at least one j

(12-18)

Rejection of H0: 1  2  p  k  0 implies that at least one of the regressor variables x1, x2, p , xk contributes significantly to the model. The test for significance of regression is a generalization of the procedure used in simple linear regression. The total sum of squares SST is partitioned into a sum of squares due to the model or to regression and a sum of squares due to error, say, SST  SSR  SSE Now if H0: 1  2  p  k  0 is true, SSR2 is a chi-square random variable with k degrees of freedom. Note that the number of degrees of freedom for this chi-square random variable is equal to the number of regressor variables in the model. We can also show that the SSE兾2 is a chi-square random variable with n  p degrees of freedom, and that SSE and SSR are independent. The test statistic for H0: 1  2  p  k  0 is Test Statistic for ANOVA

F0 

SSRk MSR  MS SSE 1n  p2 E

(12-19)

We should reject H0 if the computed value of the test statistic in Equation 12-19, f0, is greater than f,k,np. The procedure is usually summarized in an analysis of variance table such as Table 12-9. A computational formula for SSR may be found easily. Now since SST  g ni1 y2i  n n 1g i1 yi 2 2n  y¿y  1g i1 yi 2 2 n, we may rewrite Equation 12-19 as n

SSE  y¿y 

2

a a yi b i1

n

n

 ≥ ␤ˆ ¿X¿y 

a a yi b i1

n

2

¥

JWCL232_c12_449-512.qxd

1/15/10

10:07 PM

Page 471

471

12-2 HYPOTHESIS TESTS IN MULTIPLE LINEAR REGRESSION

Table 12-9 Analysis of Variance for Testing Significance of Regression in Multiple Regression Source of Variation

Sum of Squares

Degrees of Freedom

Regression Error or residual Total

SSR SSE SST

k np n1

Mean Square

F0

MSR MSE

MSR兾MSE

or

SSE  SST  SSR Therefore, the regression sum of squares is n

ˆ ¿X¿y  SSR  ␤

EXAMPLE 12-3

n

2

a a yi b i1

 27,178.5316 

n

1725.822 2

(12-21)

n

The regression or model sum of squares is computed from Equation 12-20 as follows: n

2

a a yi b i1

n

 27,063.3581 

and by subtraction ˆ ¿X¿y  115.1716 SSE  SST  SSR  y¿y  ␤

The analysis of variance is shown in Table 12-10. To test H0: 1  2  0, we calculate the statistic f0 

25

 6105.9447

ˆ ¿X¿y  SSR  ␤

i1

Wire Bond Strength ANOVA

We will test for significance of regression (with   0.05) using the wire bond pull strength data from Example 12-1. The total sum of squares is

SST  y¿y 

2

a a yi b

1725.822 2 25

 5990.7712

MSR 2995.3856   572.17 5.2352 MSE

Since f0  f0.05,2,22  3.44 (or since the P-value is considerably smaller than  = 0.05), we reject the null hypothesis and conclude that pull strength is linearly related to either wire length or die height, or both. Practical Interpretation: Rejection of H0 does not necessarily imply that the relationship found is an appropriate model for predicting pull strength as a function of wire length and die height. Further tests of model adequacy are required before we can be comfortable using this model in practice.

Most multiple regression computer programs provide the test for significance of regression in their output display. The middle portion of Table 12-4 is the Minitab output for this example. Compare Tables 12-4 and 12-10 and note their equivalence apart from rounding. The P-value is rounded to zero in the computer output. Table 12-10

Test for Significance of Regression for Example 12-3

Source of Variation

Sum of Squares

Degrees of Freedom

Regression Error or residual Total

5990.7712 115.1735 6105.9447

2 22 24

Mean Square

f0

P-value

2995.3856 5.2352

572.17

1.08E-19

JWCL232_c12_449-512.qxd

472

1/15/10

10:07 PM

Page 472

CHAPTER 12 MULTIPLE LINEAR REGRESSION

R2 and Adjusted R2 We may also use the coefficient of multiple determination R2 as a global statistic to assess the fit of the model. Computationally, R2 

SSR SSE 1 SST SST

(12-22)

For the wire bond pull strength data, we find that R2  SSR兾SST  5990.7712兾6105.9447  0.9811. Thus the model accounts for about 98% of the variability in the pull strength response (refer to the Minitab output in Table 12-4). The R2 statistic is somewhat problematic as a measure of the quality of the fit for a multiple regression model because it never decreases when a variable is added to a model. To illustrate, consider the model fit to the wire bond pull strength data in Example 11-8. This was a simple linear regression model with x1  wire length as the regressor. The value of R2 for this model is R2  0.9640. Therefore, adding x2  die height to the model increases R2 by 0.9811  0.9640  0.0171, a very small amount. Since R2 can never decrease when a regressor is added, it can be difficult to judge whether the increase is telling us anything useful about the new regressor. It is particularly hard to interpret a small increase, such as observed in the pull strength data. Many regression users prefer to use an adjusted R2 statistic: Adjusted R2

R2adj  1 

SSE 1n  p2 SST  1n  12

(12-23)

Because SSE 1n  p2 is the error or residual mean square and SST  1n  12 is a constant, R2adj will only increase when a variable is added to the model if the new variable reduces the error mean square. Note that for the multiple regression model for the pull strength data R2adj  0.979 (see the Minitab output in Table 12-4), whereas in Example 11-8 the adjusted R2 for the one-variable model is R2adj  0.962. Therefore, we would conclude that adding x2  die height to the model does result in a meaningful reduction in unexplained variability in the response. The adjusted R2 statistic essentially penalizes the analyst for adding terms to the model. It is an easy way to guard against overfitting, that is, including regressors that are not really useful. Consequently, it is very useful in comparing and evaluating competing regression models. We will use R2adj for this when we discuss variable selection in regression in Section 12-6.3.

12-2.2 Tests on Individual Regression Coefficients and Subsets of Coefficients We are frequently interested in testing hypotheses on the individual regression coefficients. Such tests would be useful in determining the potential value of each of the regressor variables in the regression model. For example, the model might be more effective with the inclusion of additional variables or perhaps with the deletion of one or more of the regressors presently in the model.

JWCL232_c12_449-512.qxd

1/15/10

10:07 PM

Page 473

12-2 HYPOTHESIS TESTS IN MULTIPLE LINEAR REGRESSION

473

The hypothesis to test if an individual regression coefficient, say j equals a value j0 is

H0: j  j0 H1: j  j0

(12-24)

The test statistic for this hypothesis is

T0 

ˆ j  j0 22Cjj



ˆ j  j0 se1 ˆ j 2

(12-25)

where Cjj is the diagonal element of 1X¿X2 1 corresponding to ˆ j. Notice that the denominator of Equation 12-24 is the standard error of the regression coefficient ˆ j. The null hypothesis H0: j  j 0 is rejected if 0 t0 0  t2,np. This is called a partial or marginal test because the regression coefficient ˆ j depends on all the other regressor variables xi (i  j) that are in the model. More will be said about this in the following example. An important special case of the previous hypothesis occurs for j0  0. If H0: j  0 is not rejected, this indicates that the regressor xj can be deleted from the model. Adding a variable to a regression model always causes the sum of squares for regression to increase and the error sum of squares to decrease (this is why R2 always increases when a variable is added). We must decide whether the increase in the regression sum of squares is large enough to justify using the additional variable in the model. Furthermore, adding an unimportant variable to the model can actually increase the error mean square, indicating that adding such a variable has actually made the model a poorer fit to the data (this is why R2adj is a better measure of global model fit then the ordinary R2). EXAMPLE 12-4

Wire Bond Strength Coefficient Test

Consider the wire bond pull strength data, and suppose that we want to test the hypothesis that the regression coefficient for x2 (die height) is zero. The hypotheses are H0: 2  0 H1: 2  0 The main diagonal element of the 1X¿X2 1 matrix corresponding to ˆ 2 is C22  0.0000015, so the t-statistic in Equation 12-25 is

t0 

ˆ 2 2ˆ 2C22



0.01253

215.23522 10.00000152

 4.477

Note that we have used the estimate of 2 reported to four decimal places in Table 12-10. Since t0.025,22  2.074, we reject H0: 2  0 and conclude that the variable x2 (die height) contributes significantly to the model. We could also have used a P-value to draw conclusions. The P-value for t0  4.477 is P  0.0002, so with  = 0.05 we would reject the null hypothesis. Practical Interpretation: Note that this test measures the marginal or partial contribution of x2 given that x1 is in the model. That is, the t-test measures the contribution of adding the variable x2  die height to a model that already contains x1  wire length. Table 12-4 shows the value of the t-test computed by Minitab. The Minitab t-test statistic is reported to two decimal places. Note that the computer produces a t-test for each regression coefficient in the model. These t-tests indicate that both regressors contribute to the model.

JWCL232_c12_449-512.qxd

474

1/15/10

10:07 PM

Page 474

CHAPTER 12 MULTIPLE LINEAR REGRESSION

EXAMPLE 12-5

Wire Bond Strength One-Sided Coefficient Test

There is an interest in the effect of die height on strength. This can be evaluated by the magnitude of the coefficient for die height. To conclude that the coefficient for die height exceeds 0.01 the hypotheses become H0: 2  0.01

H1: 2  0.01

For such a test, computer software can complete much of the hard work. We only need to assemble the pieces. From the Minitab output in Table 12-4, ˆ 2  0.012528 and the standard

error of ˆ 2  0.002798. Therefore the t-statistic is t0 

0.012528  0.01  0.9035 0.002798

with 22 degrees of freedom (error degrees of freedom). From Table IV in Appendix A, t0.25, 22  0.686 and t0.1, 22  1.321. Therefore, the P-value can be bounded as 0.1  P-value  0.25. One cannot conclude that the coefficient exceeds 0.01 at common levels of significance.

There is another way to test the contribution of an individual regressor variable to the model. This approach determines the increase in the regression sum of squares obtained by adding a variable xj (say) to the model, given that other variables xi (i ⬆ j) are already included in the regression equation. The procedure used to do this is called the general regression significance test, or the extra sum of squares method. This procedure can also be used to investigate the contribution of a subset of the regressor variables to the model. Consider the regression model with k regressor variables y  X␤  

(12-26)

where y is (n 1), X is (n p), ␤ is (p 1), ⑀ is (n 1), and p  k  1. We would like to determine if the subset of regressor variables x1, x2, . . . , xr (r  k) as a whole contributes significantly to the regression model. Let the vector of regression coefficients be partitioned as follows: ␤ c

␤1 d ␤2

(12-27)

where ␤1 is (r 1) and ␤2 is [(p  r) 1]. We wish to test the hypotheses Hypotheses for General Regression Test

H0: ␤1  0 H1: ␤1  0

(12-28)

where 0 denotes a vector of zeroes. The model may be written as y  X␤    X1␤1  X2␤2  

(12-29)

where X1 represents the columns of X associated with ␤1 and X2 represents the columns of X associated with ␤2. ˆ  1X¿X2 1 X¿y. In For the full model (including both ␤1 and ␤2), we know that ␤ addition, the regression sum of squares for all variables including the intercept is ˆ ¿X¿y SSR 1␤2  ␤

1 p  k  1 degrees of freedom2

and MSE 

ˆ X¿y y¿y  ␤ np

JWCL232_c12_449-512.qxd

1/15/10

10:07 PM

Page 475

475

12-2 HYPOTHESIS TESTS IN MULTIPLE LINEAR REGRESSION

SSR(␤) is called the regression sum of squares due to ␤. To find the contribution of the terms in ␤1 to the regression, fit the model assuming the null hypothesis H0: ␤1  0 to be true. The reduced model is found from Equation 12-29 as y  X2␤2  

(12-30)

ˆ  1X¿ X 2 1X¿ y, and The least squares estimate of ␤2 is ␤ 2 2 2 2 ˆ ¿ X¿ y SSR 1␤2 2  ␤ 2 2

1p  r degrees of freedom2

(12-31)

The regression sum of squares due to ␤1 given that ␤2 is already in the model is SSR 1␤1 0 ␤2 2  SSR 1␤2  SSR 1␤2 2

(12-32)

This sum of squares has r degrees of freedom. It is sometimes called the extra sum of squares due to ␤1. Note that SSR 1␤1 0 ␤2 2 is the increase in the regression sum of squares due to including the variables x1, x2, p , xr in the model. Now SSR 1␤1 0 ␤2 2 is independent of MSE, and the null hypothesis ␤1  0 may be tested by the statistic. F Statistic for General Regression Test

F0 

SSR 1␤1 | ␤2 2 r MSE

(12-33)

If the computed value of the test statistic f0  f,r,np, we reject H0, concluding that at least one of the parameters in ␤1 is not zero and, consequently, at least one of the variables x1, x2, p , xr in X1 contributes significantly to the regression model. Some authors call the test in Equation 12-33 a partial F-test. The partial F-test is very useful. We can use it to measure the contribution of each individual regressor xj as if it were the last variable added to the model by computing SSR 1j 0 0, 1, p , j1, j1, p , k 2,

j  1, 2, p , k

This is the increase in the regression sum of squares due to adding xj to a model that already includes x1, . . . , xj1, xj1, . . . , xk. The partial F-test is a more general procedure in that we can measure the effect of sets of variables. In Section 12-6.3 we show how the partial F-test plays a major role in model building—that is, in searching for the best set of regressor variables to use in the model.

EXAMPLE 12-6

Wire Bond Strength General Regression Test

Consider the wire bond pull-strength data in Example 12-1. We will investigate the contribution of two new variables, x3 and x4, to the model using the partial F-test approach. The new variables are explained at the end of this example. That is, we wish to test H0 : 3  4  0

H1 : 3  0 or 4  0

To test this hypothesis, we need the extra sum of squares due to 3 and 4 or SSR 14, 3 0 2, 1, 0 2  SSR 14, 3, 2, 1, 0 2  SSR 12, 1, 0 2  SSR 14, 3, 2, 1 0 0 2  SSR 12, 1 0 0 2

JWCL232_c12_449-512.qxd

476

1/21/10

9:23 PM

Page 476

CHAPTER 12 MULTIPLE LINEAR REGRESSION

In Example 12-3 we calculated n

SSR 12, 1 0 0 2  ␤¿Xⴕy 

2

a a yib i1

n

 5990.7712 (two degrees of freedom)

Also, Table 12-4 shows the Minitab output for the model with only x1 and x2 as predictors. In the analysis of variance table, we can see that SSR  5990.8 and this agrees with our calculation. In practice, the computer output would be used to obtain this sum of squares. If we fit the model Y  0  1x1  2x2  3x3  4x4, we can use the same matrix formula. Alternatively, we can look at SSR from computer output for this model. The analysis of variance table for this model is shown in Table 12-11 and we see that SSR 14, 3, 2, 1 0 0 2  6024.0 (four degrees of freedom) Therefore, SSR 14, 3 0 2, 1, 0 2  6024.0  5990.8  33.2 (two degrees of freedom)

This is the increase in the regression sum of squares due to adding x3 and x4 to a model already containing x1 and x2. To test H0, calculate the test statistic f0 

SSR 14, 3 0 2, 1, 0 2 2 MSE



33.22  4.05 4.1

Note that MSE from the full model using x1, x2, x3 and x4 is used in the denominator of the test statistic. Because f0.05, 2, 20  3.49, we reject H0 and conclude that at least one of the new variables contributes significantly to the model. Further analysis and tests will be needed to refine the model and determine if one or both of x3 and x4 are important. The mystery of the new variables can now be explained. These are quadratic powers of the original predictors wire length and wire height. That is, x3  x21 and x4  x22. A test for quadratic terms is a common use of partial F-tests. With this information and the original data for x1 and x2, you can use computer software to reproduce these calculations. Multiple regression allows models to be extended in such a simple manner that the real meaning of x3 and x4 did not even enter into the test procedure. Polynomial models such as this are discussed further in Section 12-6.

If a partial F-test is applied to a single variable, it is equivalent to a t-test. To see this, consider the Minitab regression output for the wire bond pull strength in Table 12-4. Just below the analysis of variance summary in this table, the quantity labeled ” ‘SeqSS”’ shows the sum Table 12-11

Regression Analysis: y versus x1, x2, x3, x4

The regression equation is y  5.00  1.90 x1 + 0.0151 x2 + 0.0460 x3  0.000008 x4 Predictor Constant x1 x2 x3 x4

Coef 4.996 1.9049 0.01513 0.04595 0.00000766

S  2.02474

SE Coef 1.655 0.3126 0.01051 0.01666 0.00001641

RSq  98.7%

T 3.02 6.09 1.44 2.76 0.47

P 0.007 0.000 0.165 0.012 0.646

RSq (adj)  98.4%

Analysis of Variance Source Regression Residual Error Total Source x1 x2 x3 x4

DF 4 20 24 DF 1 1 1 1

Seq SS 5885.9 104.9 32.3 0.9

SS 6024.0 82.0 6105.9

MS 1506.0 4.1

F 367.35

P 0.000

JWCL232_c12_449-512.qxd

1/16/10

10:52 AM

Page 477

12-2 HYPOTHESIS TESTS IN MULTIPLE LINEAR REGRESSION

477

of squares obtained by fitting x1 alone (5885.9) and the sum of squares obtained by fitting x2 after x1 (104.9). In out notation, these are referred to as SSR 11 0 0 2 and SSR 12, 1 0 0 2, respectively. Therefore, to test H0 : 2  0, H1 : 2  0 the partial F-test is f0 

SSR 12 0 1, 0 2 1 104.92  20.2  5.24 MSE

where MSE is the mean square for residual in the computer output in Table 12.4. This statistic should be compared to an F-distribution with 1 and 22 degrees of freedom in the numerator and denominator, respectively. From Table 12-4, the t-test for the same hypothesis is t0  4.48. Note that t02  4.482  20.07  f0, except for round-off error. Furthermore, the square of a t-random variable with  degrees of freedom is an F-random variable with one and  degrees of freedom. Consequently, the t-test provides an equivalent method to test a single variable for contribution to a model. Because the t-test is typically provided by computer output, it is the preferred method to test a single variable.

EXERCISES FOR SECTION 12-2 12-21.

(a) Test for significance of regression using  0.05. What is the P-value for this test? (b) Construct the t-test on each regression coefficient. What are your conclusions, using  0.05? Calculate P-values.

Consider the computer output below.

The regression equation is Y  254  2.77 x1  3.58 x2 Predictor Constant x1 x2 S  5.05756

Coef 253.810 2.7738 3.5753 R-Sq  ?

SE Coef 4.781 0.1846 0.1526

T ? 15.02 ?

P ? ? ?

R-Sq (adj)  98.4%

Analysis of Variance Source Regression Residual Error Total

DF 2 ? 14

SS 22784 ? 23091

MS 11392 ?

F ?

P ?

(a) Fill in the missing quantities. You may use bounds for the P-values. (b) What conclusions can you draw about the significance of regression? (c) What conclusions can you draw about the contributions of the individual regressors to the model? 12-22. You have fit a regression model with two regressors to a data set that has 20 observations. The total sum of squares is 1000 and the model sum of squares is 750. (a) What is the value of R2 for this model? (b) What is the adjusted R2 for this model? (c) What is the value of the F-statistic for testing the significance of regression? What conclusions would you draw about this model if  0.05? What if  0.01? (d) Suppose that you add a third regressor to the model and as a result the model sum of squares is now 785. Does it seem to you that adding this factor has improved the model? 12-23. Consider the regression model fit to the soil shear strength data in Exercise 12-1.

12-24. Consider the absorption index data in Exercise 12-2. The total sum of squares for y is SST  742.00. (a) Test for significance of regression using  0.01. What is the P-value for this test? (b) Test the hypothesis H0: 1  0 versus H1: 1  0 using  0.01. What is the P-value for this test? (c) What conclusion can you draw about the usefulness of x1 as a regressor in this model? 12-25. A regression model Y  0  1x1  2 x2  3 x3 

has been fit to a sample of n  25 observations. The calculated t-ratios ˆ j se 1ˆ j 2, j  1, 2, 3 are as follows: for 1, t0  4.82, for 2, t0  8.21 and for 3, t0  0.98. (a) Find P-values for each of the t-statistics. (b) Using  0.05, what conclusions can you draw about the regressor x3? Does it seem likely that this regressor contributes significantly to the model? 12-26. Consider the electric power consumption data in Exercise 12-6. (a) Test for significance of regression using  0.05. What is the P-value for this test? (b) Use the t-test to assess the contribution of each regressor to the model. Using  0.05, what conclusions can you draw? 12-27. Consider the gasoline mileage data in Exercise 12-7. (a) Test for significance of regression using  0.05. What conclusions can you draw? (b) Find the t-test statistic for each regressor. Using  0.05, what conclusions can you draw? Does each regressor contribute to the model?

JWCL232_c12_449-512.qxd

478

1/15/10

10:07 PM

Page 478

CHAPTER 12 MULTIPLE LINEAR REGRESSION

12-28. Consider the wire bond pull strength data in Exercise 12-8. (a) Test for significance of regression using   0.05. Find the P-value for this test. What conclusions can you draw? (b) Calculate the t-test statistic for each regression coefficient. Using   0.05, what conclusions can you draw? Do all variables contribute to the model? 12-29. Reconsider the semiconductor data in Exercise 12-9. (a) Test for significance of regression using   0.05. What conclusions can you draw? (b) Calculate the t-test statistic and P-value for each regression coefficient. Using   0.05, what conclusions can you draw? 12-30. Consider the regression model fit to the arsenic data in Exercise 12-12. Use arsenic in nails as the response and age, drink use, and cook use as the regressors. (a) Test for significance of regression using   0.05. What is the P-value for this test? (b) Construct a t-test on each regression coefficient. What conclusions can you draw about the variables in this model? Use   0.05. 12-31. Consider the regression model fit to the X-ray inspection data in Exercise 12-11. Use rads as the response. (a) Test for significance of regression using   0.05. What is the P-value for this test? (b) Construct a t-test on each regression coefficient. What conclusions can you draw about the variables in this model? Use   0.05. 12-32. Consider the regression model fit to the nisin extraction data in Exercise 12-14. Use nisin extraction as the response. (a) Test for significance of regression using   0.05. What is the P-value for this test? (b) Construct a t-test on each regression coefficient. What conclusions can you draw about the variables in this model? Use   0.05. (c) Comment on the effect of a small sample size to the tests in the previous parts. 12-33. Consider the regression model fit to the grey range modulation data in Exercise 12-15. Use the useful range as the response. (a) Test for significance of regression using   0.05. What is the P-value for this test? (b) Construct a t-test on each regression coefficient. What conclusions can you draw about the variables in this model? Use   0.05. 12-34. Consider the regression model fit to the stack loss data in Exercise 12-16. Use stack loss as the response. (a) Test for significance of regression using   0.05. What is the P-value for this test? (b) Construct a t-test on each regression coefficient. What conclusions can you draw about the variables in this model? Use   0.05.

12-35. Consider the NFL data in Exercise 12-17. (a) Test for significance of regression using   0.05. What is the P-value for this test? (b) Conduct the t-test for each regression coefficient. Using   0.05, what conclusions can you draw about the variables in this model? (c) Find the amount by which the regressor x2 (TD percentage) increases the regression sum of squares, and conduct an F-test for H0: 2  0 versus H1: 2 ⬆ 0 using   0.05. What is the P-value for this test? What conclusions can you draw? 12-36. Exercise 12-10 presents data on heat treating gears. (a) Test the regression model for significance of regression. Using   0.05, find the P-value for the test and draw conclusions. (b) Evaluate the contribution of each regressor to the model using the t-test with   0.05. (c) Fit a new model to the response PITCH using new regressors x1  SOAKTIME SOAKPCT and x2  DIFFTIME DIFFPCT. (d) Test the model in part (c) for significance of regression using   0.05. Also calculate the t-test for each regressor and draw conclusions. (e) Estimate 2 for the model from part (c) and compare this to the estimate of 2 for the model in part (a). Which estimate is smaller? Does this offer any insight regarding which model might be preferable? 12-37. Consider the bearing wear data in Exercise 12-19. (a) For the model with no interaction, test for significance of regression using   0.05. What is the P-value for this test? What are your conclusions? (b) For the model with no interaction, compute the t-statistics for each regression coefficient. Using   0.05, what conclusions can you draw? (c) For the model with no interaction, use the extra sum of squares method to investigate the usefulness of adding x2  load to a model that already contains x1  oil viscosity. Use   0.05. (d) Refit the model with an interaction term. Test for significance of regression using   0.05. (e) Use the extra sum of squares method to determine whether the interaction term contributes significantly to the model. Use   0.05. (f) Estimate 2 for the interaction model. Compare this to the estimate of 2 from the model in part (a). 12-38. Data on National Hockey League team performance was presented in Exercise 12-18. (a) Test the model from this exercise for significance of regression using   0.05. What conclusions can you draw? (b) Use the t-test to evaluate the contribution of each regressor to the model. Does it seem that all regressors are necessary? Use   0.05. (c) Fit a regression model relating the number of games won to the number of goals for and the number of power play goals

JWCL232_c12_449-512.qxd

1/15/10

10:07 PM

Page 479

12-3 CONFIDENCE INTERVALS IN MULTIPLE LINEAR REGRESSION

for. Does this seem to be a logical choice of regressors, considering your answer to part (b)? Test this new model for significance of regression and evaluate the contribution of each regressor to the model using the t-test. Use   0.05. 12-39. Data from a hospital patient satisfaction survey were presented in Exercise 12-5. (a) Test the model from this exercise for significance of regression. What conclusions can you draw if   0.05? What if   0.01? (b) Test the contribution of the individual regressors using the t-test. Does it seem that all regressors used in the model are really necessary? 12-40. Data from a hospital patient satisfaction survey were presented in Exercise 12-5.

479

(a) Fit a regression model using only the patient age and severity regressors. Test the model from this exercise for significance of regression. What conclusions can you draw if   0.05? What if   0.01? (b) Test the contribution of the individual regressors using the t-test. Does it seem that all regressors used in the model are really necessary? (c) Find an estimate of the error variance 2. Compare this estimate of 2 with the estimate obtained from the model containing the third regressor, anxiety. Which estimate is smaller? Does this tell you anything about which model might be preferred?

12-3 CONFIDENCE INTERVALS IN MULTIPLE LINEAR REGRESSION 12-3.1 Confidence Intervals on Individual Regression Coefficients In multiple regression models, it is often useful to construct confidence interval estimates for the regression coefficients 5 j 6. The development of a procedure for obtaining these confidence intervals requires that the errors 5i 6 are normally and independently distributed with mean zero and variance 2. This is the same assumption required in hypothesis testing. Therefore, the observations {Yi} are normally and independently distributed with mean 0  k g j1 j xij and variance 2. Since the least squares estimator ␤ˆ is a linear combination of the observations, it follows that ␤ˆ is normally distributed with mean vector ␤ and covariance matrix 2 1X¿X2 1. Then each of the statistics

T

ˆ j  j 2ˆ 2Cjj

j  0, 1, p , k

(12-34)

has a t distribution with n  p degrees of freedom, where Cjj is the jjth element of the 1X¿X2 1 matrix, and ˆ 2 is the estimate of the error variance, obtained from Equation 12-16. This leads to the following 100(1  )% confidence interval for the regression coefficient j, j  0, 1, p , k. Confidence Interval on a Regression Coefficient

A 100(1  ) % confidence interval on the regression coefficient j, j  0, 1, p , k in the multiple linear regression model is given by ˆ j  t2,np 2ˆ 2Cjj j ˆ j  t2,np 2ˆ 2Cjj

(12-35)

Because 2ˆ 2Cjj is the standard error of the regression coefficient ˆ j, we would also write the CI formula as ˆ j  t2,np se1ˆ j 2 j ˆ j  t2,np se1ˆ j 2.

JWCL232_c12_449-512.qxd

480

1/15/10

10:08 PM

Page 480

CHAPTER 12 MULTIPLE LINEAR REGRESSION

EXAMPLE 12-7

Wire Bond Strength Confidence Interval

We will construct a 95% confidence interval on the parameter 1 in the wire bond pull strength problem. The point estimate of 1 is ˆ 1  2.74427 and the diagonal element of 1X¿X2 1 corresponding to 1 is C11  0.001671. The estimate of 2 is ˆ 2  5.2352, and t0.025,22  2.074. Therefore, the 95% CI on 1 is computed from Equation 12-35 as 2.74427  12.0742 215.23522 1.0016712 1 2.74427  12.0742 215.23522 1.0016712

Also, computer software such as Minitab can be used to help calculate this confidence interval. From the regression output in Table 10-4, ˆ 1  2.74427 and the standard error of ˆ 1  0.0935. This standard error is the multiplier of the t-table constant in the confidence interval. That is, 0.0935  115.2352210.0016712. Consequently, all the numbers are available from the computer output to construct the interval and this is the typical method used in practice.

which reduces to 2.55029 1 2.93825

12-3.2 Confidence Interval on the Mean Response We may also obtain a confidence interval on the mean response at a particular point, say, x01, x02, p , x0k. To estimate the mean response at this point, define the vector 1 x01 x0  M x02 m o x0k The mean response at this point is E1Y 0 x0 2  Y 0 x0  x¿0␤, which is estimated by ˆ ˆ Y 0 x0  x¿0␤

(12-36)

ˆ Y 0 x0 2  2x¿0 1X¿X2 1x0 V1

(12-37)

This estimator is unbiased, since E1x¿0␤ˆ 2  x¿0␤  E1Y 0 x0 2  Y 0 x0 and the variance of ˆ Y 0 x0 is A 100(1  ) % CI on Y 0 x0 can be constructed from the statistic ˆ Y 0 x0  Y 0 x0

2ˆ 2 x¿0 1X¿X21 x0

(12-38)

˛

Confidence Interval on the Mean Response

For the multiple linear regression model, a 100(1  )% confidence interval on the mean response at the point x01, x02, . . . , x0k is ˆ Y 0 x0  t2,np 2ˆ 2 x¿0 1X¿X2 1 x0 ˛

Y 0 x0 ˆ Y 0 x0  t2,np 2ˆ 2 x¿0 1⌾¿⌾21 x0 ˛

(12-39)

JWCL232_c12_449-512.qxd

1/15/10

10:08 PM

Page 481

481

12-4 PREDICTION OF NEW OBSERVATIONS

Equation 12-39 is a CI about the regression plane (or hyperplane). It is the multiple regression generalization of Equation 11-32.

EXAMPLE 12-8

Wire Bond Strength Confidence Interval on the Mean Response

The engineer in Example 12-1 would like to construct a 95% CI on the mean pull strength for a wire bond with wire length x1  8 and die height x2  275. Therefore,

Therefore, a 95% CI on the mean pull strength at this point is found from Equation 12-39 as 27.66  2.074 10.23244 Y 0 x0 27.66

1 x0  £ 8 § 275

 2.074 10.23244

which reduces to

The estimated mean response at this point is found from Equation 12-36 as ˆ Y |x0  x0¿ ˆ  31 8

2.26379 275 4 £ 2.74427 §  27.66 0.01253

ˆ Y 0 x0 is estimated by The variance of 

26.66 Y |x0 28.66 Some computer software packages will provide estimates of the mean for a point of interest x0 and the associated CI. Table 12-4 shows the Minitab output for Example 12-8. Both the estimate of the mean and the 95% CI are provided.

ˆ 2x0¿ 1⌾¿⌾2 1x0  5.2352 31 8 275 4 .214653 .007491 .000340 1

£ .007491 .001671 .000019 § £ 8 § .000340 .000019 .0000015 275

 5.2352 10.04442  0.23244

12-4 PREDICTION OF NEW OBSERVATIONS A regression model can be used to predict new or future observations on the response variable Y corresponding to particular values of the independent variables, say, x01, x02, p , x0k. If x¿0  31, x01, x02, p , x0k 4 , a point estimate of the future observation Y0 at the point x01, x02, p , x0k is yˆ 0  x¿0 ␤ˆ ˛

Prediction Interval

(12-40)

A 100(1  )% prediction interval for this future observation is yˆ 0  t2,np 2ˆ 2 11  x¿0 1⌾¿⌾2 1 x0 2 ˛

Y0 yˆ 0  t2,np 2ˆ 2 11  x¿0 1⌾¿⌾2 1 x0 2 ˛

(12-41)

This prediction interval is a generalization of the prediction interval given in Equation 11-33 for a future observation in simple linear regression. If you compare the prediction interval Equation 12-41 with the expression for the confidence interval on the mean, Equation 12-39,

JWCL232_c12_449-512.qxd

482

1/15/10

10:08 PM

Page 482

CHAPTER 12 MULTIPLE LINEAR REGRESSION

Original range for x2

x2

Joint region of original area x01

x01 Original range for x1

x1

Figure 12-5 An example of extrapolation in multiple regression.

you will observe that the prediction interval is always wider than the confidence interval. The confidence interval expresses the error in estimating the mean of a distribution, while the prediction interval expresses the error in predicting a future observation from the distribution at the point x0. This must include the error in estimating the mean at that point, as well as the inherent variability in the random variable Y at the same value x  x0. Also, one might want to predict the mean of several values of Y, say m, all at the same value x = x0. Because the variance of a sample mean is 2/m, Equation 12-41 is modified as follows. Replace the constant 1 under the square root with 1/m to reflect the lower variability in the mean of m observations. This results in a narrower interval. In predicting new observations and in estimating the mean response at a given point x01, x02, . . . , x0k, we must be careful about extrapolating beyond the region containing the original observations. It is very possible that a model that fits well in the region of the original data will no longer fit well outside of that region. In multiple regression it is often easy to inadvertently extrapolate, since the levels of the variables (xi1, xi2, . . . , xik), i  1, 2, . . . , n, jointly define the region containing the data. As an example, consider Fig. 12-5, which illustrates the region containing the observations for a two-variable regression model. Note that the point (x01, x02) lies within the ranges of both regressor variables x1 and x2, but it is outside the region that is actually spanned by the original observations. This is sometimes called a hidden extrapolation. Either predicting the value of a new observation or estimating the mean response at this point is an extrapolation of the original regression model. EXAMPLE 12-9

Wire Bond Strength Confidence Interval

Suppose that the engineer in Example 12-1 wishes to construct a 95% prediction interval on the wire bond pull strength when the wire length is x1  8 and the die height is x2  275. Note that x¿0  [1 8 275], and the point estimate of the ˆ  27.66. Also, in Example 12-8 pull strength is yˆ 0  x¿0  we calculated x¿0 1⌾¿⌾2 1x0  0.04444. Therefore, from Equation 12-41 we have 27.66  2.074 25.235211  0.04442 Y0 27.66 ˛

 2.074 25.235211  0.04442

and the 95% prediction interval is 22.81 Y0 32.51 Notice that the prediction interval is wider than the confidence interval on the mean response at the same point, calculated in Example 12-8. The Minitab output in Table 12-4 also displays this prediction interval.

JWCL232_c12_449-512.qxd

1/15/10

10:08 PM

Page 483

12-4 PREDICTION OF NEW OBSERVATIONS

483

EXERCISES FOR SECTIONS 12-3 AND 12-4 12-41. Consider the regression model fit to the shear strength of soil in Exercise 12-1. (a) Calculate 95% confidence intervals on each regression coefficient. (b) Calculate a 95% confidence interval on mean strength when x1  18 feet and x2  43%. (c) Calculate 95% prediction interval on strength for the same values of the regressors used in the previous part. 12-42. Consider the soil absorption data in Exercise 12-2. (a) Find 95% confidence intervals on the regression coefficients. (b) Find a 95% confidence interval on mean soil absorption index when x1  200 and x2  50. (c) Find a 95% prediction interval on the soil absorption index when x1  200 and x2  50. 12-43. Consider the semiconductor data in Exercise 12-9. (a) Find 99% confidence intervals on the regression coefficients. (b) Find a 99% prediction interval on HFE when x1  14.5, x2  220, and x3  5.0. (c) Find a 99% confidence interval on mean HFE when x1  14.5, x2  220, and x3  5.0. 12-44. Consider the electric power consumption data in Exercise 12-6. (a) Find 95% confidence intervals on 1, 2, 3, and 4. (b) Find a 95% confidence interval on the mean of Y when x1  75, x2  24, x3  90, and x4  98. (c) Find a 95% prediction interval on the power consumption when x1  75, x2  24, x3  90, and x4  98. 12-45. Consider the bearing wear data in Exercise 12-19. (a) Find 99% confidence intervals on 1 and 2. (b) Recompute the confidence intervals in part (a) after the interaction term x1x2 is added to the model. Compare the lengths of these confidence intervals with those computed in part (a). Do the lengths of these intervals provide any information about the contribution of the interaction term in the model? 12-46. Consider the wire bond pull strength data in Exercise 12-8. (a) Find 95% confidence interval on the regression coefficients. (b) Find a 95% confidence interval on mean pull strength when x2  20, x3  30, x4  90, and x5  2.0. (c) Find a 95% prediction interval on pull strength when x2  20, x3  30, x4  90, and x5  2.0. 12-47. Consider the regression model fit to the X-ray inspection data in Exercise 12-11. Use rads as the response. (a) Calculate 95% confidence intervals on each regression coefficient. (b) Calculate a 99% confidence interval on mean rads at 15 milliamps and 1 second on exposure time. (c) Calculate a 99% prediction interval on rads for the same values of the regressors used in the previous part.

12-48. Consider the regression model fit to the arsenic data in Exercise 12-12. Use arsenic in nails as the response and age, drink use, and cook use as the regressors. (a) Calculate 99% confidence intervals on each regression coefficient. (b) Calculate a 99% confidence interval on mean arsenic concentration in nails when age  30, drink use  4, and cook use  4. (c) Calculate a prediction interval on arsenic concentration in nails for the same values of the regressors used in the previous part. 12-49. Consider the regression model fit to the coal and limestone mixture data in Exercise 12-13. Use density as the response. (a) Calculate 90% confidence intervals on each regression coefficient. (b) Calculate a 90% confidence interval on mean density when the dielectric constant  2.3 and the loss factor  0.025. (c) Calculate a prediction interval on density for the same values of the regressors used in the previous part. 12-50. Consider the regression model fit to the nisin extraction data in Exercise 12-14. (a) Calculate 95% confidence intervals on each regression coefficient. (b) Calculate a 95% confidence interval on mean nisin extraction when x1  15.5 and x2  16. (c) Calculate a prediction interval on nisin extraction for the same values of the regressors used in the previous part. (d) Comment on the effect of a small sample size to the widths of these intervals. 12-51. Consider the regression model fit to the grey range modulation data in Exercise 12-15. Use the useful range as the response. (a) Calculate 99% confidence intervals on each regression coefficient. (b) Calculate a 99% confidence interval on mean useful range when brightness  70 and contrast  80. (c) Calculate a prediction interval on useful range for the same values of the regressors used in the previous part. (d) Calculate a 99% confidence interval and a 99% a prediction interval on useful range when brightness  50 and contrast  25. Compare the widths of these intervals to those calculated in parts (b) and (c). Explain any differences in widths. 12-52. Consider the stack loss data in Exercise 12-16. (a) Calculate 95% confidence intervals on each regression coefficient. (b) Calculate a 95% confidence interval on mean stack loss when x1  80, x2  25 and x3  90. (c) Calculate a prediction interval on stack loss for the same values of the regressors used in the previous part.

JWCL232_c12_449-512.qxd

484

1/15/10

10:08 PM

Page 484

CHAPTER 12 MULTIPLE LINEAR REGRESSION

(d) Calculate a 95% confidence interval and a 95% prediction interval on stack loss when x1  80, x2  19, and x3  93. Compare the widths of these intervals to those calculated in parts (b) and (c). Explain any differences in widths. 12-53. Consider the NFL data in Exercise 12-17. (a) Find 95% confidence intervals on the regression coefficients. (b) What is the estimated standard error of ˆ Y |x0 when the percentage of completions is 60%, the percentage of TDs is 4%, and the percentage of interceptions is 3%. (c) Find a 95% confidence interval on the mean rating when the percentage of completions is 60%, the percentage of TDs is 4%, and the percentage of interceptions is 3%. 12-54. Consider the heat treating data from Exercise 12-10. (a) Find 95% confidence intervals on the regression coefficients. (b) Find a 95% confidence interval on mean PITCH when TEMP  1650, SOAKTIME  1.00, SOAKPCT  1.10, DIFFTIME  1.00, and DIFFPCT  0.80. (c) Fit a model to PITCH using regressors x1  SOAKTIME SOAKPCT and x2  DIFFTIME DIFFPCT. Using the model with regressors x1 and x2, find a 95% confidence interval on mean PITCH when SOAKTIME  1.00, SOAKPCT  1.10, DIFFTIME  1.00, and DIFFPCT  0.80. (d) Compare the length of this confidence interval with the length of the confidence interval on mean PITCH at

the same point from part (b), where an additive model in SOAKTIME, SOAKPCT, DIFFTIME, and DIFFPCT was used. Which confidence interval is shorter? Does this tell you anything about which model is preferable? 12-55. Consider the gasoline mileage data in Exercise 12-7. (a) Find 99% confidence intervals on the regression coefficients. (b) Find a 99% confidence interval on the mean of Y for the regressor values in the first row of data. (c) Fit a new regression model to these data using cid, etw, and axle as the regressors. Find 99% confidence intervals on the regression coefficients in this new model. (d) Compare the lengths of the confidence intervals from part (c) with those found in part (a). Which intervals are longer? Does this offer any insight about which model is preferable? 12-56. Consider the NHL data in Exercise 12-18. (a) Find a 95% confidence interval on the regression coefficient for the variable GF. (b) Fit a simple linear regression model relating the response variable W to the regressor GF. (c) Find a 95% confidence interval on the slope for the simple linear regression model from part (b). (d) Compare the lengths of the two confidence intervals computed in parts (a) and (c). Which interval is shorter? Does this tell you anything about which model is preferable?

12-5 MODEL ADEQUACY CHECKING 12-5.1 Residual Analysis The residuals from the multiple regression model, defined by ei  yi  yˆ i, play an important role in judging model adequacy just as they do in simple linear regression. As noted in Section 11-7.1, several residual plots are often useful; these are illustrated in Example 12-10. It is also helpful to plot the residuals against variables not presently in the model that are possible candidates for inclusion. Patterns in these plots may indicate that the model may be improved by adding the candidate variable. EXAMPLE 12-10 Wire Bond Strength Residuals The residuals for the model from Example 12-1 are shown in Table 12-3. A normal probability plot of these residuals is shown in Fig. 12-6. No severe deviations from normality are

obviously apparent, although the two largest residuals (e15  5.84 and e17  4.33) do not fall extremely close to a straight line drawn through the remaining residuals.

The standardized residuals Standardized Residual

di 

ei 2MSE



ei 2ˆ 2

(12-42)

JWCL232_c12_449-512.qxd

1/15/10

10:08 PM

Page 485

485

12-5 MODEL ADEQUACY CHECKING

2 5

6

10

5 4 3

Probability

20 30 40 50 60 70

ei

2 1 0 –1 –2

80

–3

90

–4

98

–5

99

10 –6

Figure 12-6

–4

–2

0

2

4

6

20

30

7

Normal probability plot of residuals.

Figure 12-7

40

50

60

70

^ yi

Plot of residuals against yˆ.

are often more useful than the ordinary residuals when assessing residual magnitude. For the wire bond strength example, the standardized residuals corresponding to e15 and e17 are d15  5.84 15.2352  2.55 and d17  4.33 15.2352  1.89, and they do not seem unusually large. Inspection of the data does not reveal any error in collecting observations 15 and 17, nor does it produce any other reason to discard or modify these two points. The residuals are plotted against yˆ in Fig. 12-7, and against x1 and x2 in Figs. 12-8 and 12-9, respectively.* The two largest residuals, e15 and e17, are apparent. Figure 12-8 gives some indication that the model underpredicts the pull strength for assemblies with short wire length 1x1 62 and long wire length 1x1  152 and overpredicts the strength for assemblies with intermediate wire length 17 x1 142 . The same impression is obtained from Fig. 12-7. Either the relationship between strength and wire length is not linear (requiring that a term involving x12, say, be added to the model), or other regressor variables not presently in the model affected the response. In the wire bond strength example we used the standardized residuals di  ei 2ˆ 2 as a measure of residual magnitude. Some analysts prefer to plot standardized residuals instead of ordinary residuals, because the standardized residuals are scaled so that their standard

ei

6

6

5

5

4

4

3

3

2

2

1 0

ei 1 0

–1

–1

–2

–2

–3

–3

–4

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 x1

Figure 12-8

Plot of residuals against x1.

–4

100

200

300

400

500

600

700

x2

Figure 12-9

Plot of residuals against x2.

*There are other methods, described in Montgomery, Peck, and Vining (2006) and Myers (1990), that plot a modified version of the residual, called a partial residual, against each regressor. These partial residual plots are useful in displaying the relationship between the response y and each individual regressor.

JWCL232_c12_449-512.qxd

486

1/15/10

10:08 PM

Page 486

CHAPTER 12 MULTIPLE LINEAR REGRESSION

deviation is approximately unity. Consequently, large residuals (that may indicate possible outliers or unusual observations) will be more obvious from inspection of the residual plots. Many regression computer programs compute other types of scaled residuals. One of the most popular are the studentized residuals

Studentized Residual

ei

ri 

i  1, 2, p , n

2 11  hii 2 ˆ2

(12-43)

where hii is the ith diagonal element of the matrix

H  ⌾ 1⌾¿⌾2 1 ⌾¿

The H matrix is sometimes called the “hat” matrix, since ˆ  X 1X¿X2 1 X¿y  Hy yˆ  X␤ ˛

˛

Thus H transforms the observed values of y into a vector of fitted values yˆ . Since each row of the matrix X corresponds to a vector, say x¿i  31, xi1, xi2, p , xik 4 , another way to write the diagonal elements of the hat matrix is

Diagonal Elements of Hat Matrix

hii  x¿i 1X¿X2 1xi

(12-44)

Note that apart from 2, hii is the variance of the fitted value yˆ i. The quantities hii were used in the computation of the confidence interval on the mean response in Section 12-3.2. Under the usual assumptions that the model errors are independently distributed with mean zero and variance 2, we can show that the variance of the ith residual ei is V 1ei 2  2 11  hii 2, ˛

i  1, 2, p , n

Furthermore, the hii elements must fall in the interval 0  hii 1. This implies that the standardized residuals understate the true residual magnitude; thus, the studentized residuals would be a better statistic to examine in evaluating potential outliers. To illustrate, consider the two observations identified in the wire bond strength data (Example 12-10) as having residuals that might be unusually large, observations 15 and 17. The standardized residuals are d15 

e15 2

ˆ2



5.84 25.2352

 2.55

and

d17 

e17 2MSE



4.33 25.2352

Now h15,15  0.0737 and h17,17  0.2593, so the studentized residuals are r15 

e15

2 11  h15,15 2 ˆ2



5.84 25.235211  0.07372

 2.65

 1.89

JWCL232_c12_449-512.qxd

1/15/10

10:08 PM

Page 487

12-5 MODEL ADEQUACY CHECKING

487

and r17 

e17

2 11  h17,17 2 ˆ2



4.33 25.235211  0.25932

 2.20

Notice that the studentized residuals are larger than the corresponding standardized residuals. However, the studentized residuals are still not so large as to cause us serious concern about possible outliers.

12-5.2 Influential Observations When using multiple regression, we occasionally find that some subset of the observations is unusually influential. Sometimes these influential observations are relatively far away from the vicinity where the rest of the data were collected. A hypothetical situation for two variables is depicted in Fig. 12-10, where one observation in x-space is remote from the rest of the data. The disposition of points in the x-space is important in determining the properties of the model. For example, point (xi1, xi2) in Fig. 12-10 may be very influential in determining R2, the estimates of the regression coefficients, and the magnitude of the error mean square. We would like to examine the influential points to determine whether they control many model properties. If these influential points are “bad” points, or erroneous in any way, they should be eliminated. On the other hand, there may be nothing wrong with these points, but at least we would like to determine whether or not they produce results consistent with the rest of the data. In any event, even if an influential point is a valid one, if it controls important model properties, we would like to know this, since it could have an impact on the use of the model. Montgomery, Peck, and Vining (2006) and Myers (1990) describe several methods for detecting influential observations. An excellent diagnostic is the distance measure developed by Dennis R. Cook. This is a measure of the squared distance between the usual least squares estimate of ␤ based on all n observations and the estimate obtained when the ith point is removed, say, ␤ˆ 1i2. The Cook’s distance measure is Cook’s Distance

Di 

1␤ˆ 1i2  ␤ˆ 2 ¿X¿X 1␤ˆ 1i2  ␤ˆ 2 pˆ 2

x2 x i2 Region containing all observations except the ith

Figure 12-10 A point that is remote in x-space.

x i1

x1

i  1, 2, p , n

JWCL232_c12_449-512.qxd

488

1/15/10

10:08 PM

Page 488

CHAPTER 12 MULTIPLE LINEAR REGRESSION

Clearly, if the ith point is influential, its removal will result in ␤ˆ 1i2 changing considerably from ˆ . Thus, a large value of D implies that the ith point is influential. The statistic D is the value ␤ i i actually computed using

Cook’s Distance Formula

hii r i2 Di  p 11  hii 2

i  1, 2, p , n

(12-45)

From Equation 12-44 we see that Di consists of the squared studentized residual, which reflects how well the model fits the ith observation yi [recall that ri  ei 2ˆ 2 11  hii 2 4 and a component that measures how far that point is from the rest of the data 3hii 11  hii 2 is a measure of the distance of the ith point from the centroid of the remaining n  1 points]. A value of Di  1 would indicate that the point is influential. Either component of Di (or both) may contribute to a large value.

EXAMPLE 12-11 Wire Bond Strength Cook’s Distances Table 12-12 lists the values of the hat matrix diagonals hii and Cook’s distance measure Di for the wire bond pull strength data in Example 12-1. To illustrate the calculations, consider the first observation:



31.57  25.235211  0.15732 4 2 3

 0.035

˛

ⴢ

˛

0.1573 11  0.15732

The Cook distance measure Di does not identify any potentially influential observations in the data, for no value of Di exceeds unity.

h11 r12 D1  p ⴢ 11  h11 2 ˛



˛

3e1  2MSE 11  h11 2 4 2 h11 ⴢ p 11  h11 2 ˛

Table 12-12 Influence Diagnostics for the Wire Bond Pull Strength Data 2 Observations i

hii

Cook’s Distance Measure Di

Observations i

hii

Cook’s Distance Measure Di

1 2 3 4 5 6 7 8 9 10 11 12 13

0.1573 0.1116 0.1419 0.1019 0.0418 0.0749 0.1181 0.1561 0.1280 0.0413 0.0925 0.0526 0.0820

0.035 0.012 0.060 0.021 0.024 0.007 0.036 0.020 0.160 0.001 0.013 0.001 0.001

14 15 16 17 18 19 20 21 22 23 24 25

0.1129 0.0737 0.0879 0.2593 0.2929 0.0962 0.1473 0.1296 0.1358 0.1824 0.1091 0.0729

0.003 0.187 0.001 0.565 0.155 0.018 0.000 0.052 0.028 0.002 0.040 0.000

JWCL232_c12_449-512.qxd

1/15/10

10:08 PM

Page 489

12-5 MODEL ADEQUACY CHECKING

489

EXERCISES FOR SECTION 12-5 12-57. Consider the gasoline mileage data in Exercise 12-7. (a) What proportion of total variability is explained by this model? (b) Construct a normal probability plot of the residuals and comment on the normality assumption. (c) Plot residuals versus yˆ and versus each regressor. Discuss these residual plots. (d) Calculate Cook’s distance for the observations in this data set. Are any observations influential? 12-58. Consider the electric power consumption data in Exercise 12-6. (a) Calculate R2 for this model. Interpret this quantity. (b) Plot the residuals versus yˆ and versus each regressor. Interpret this plot. (c) Construct a normal probability plot of the residuals and comment on the normality assumption. 12-59. Consider the regression model for the NFL data in Exercise 12-17. (a) What proportion of total variability is explained by this model? (b) Construct a normal probability plot of the residuals. What conclusion can you draw from this plot? (c) Plot the residuals versus yˆ and versus each regressor, and comment on model adequacy. (d) Are there any influential points in these data? 12-60. Consider the regression model for the heat treating data in Exercise 12-10. (a) Calculate the percent of variability explained by this model. (b) Construct a normal probability plot for the residuals. Comment on the normality assumption. (c) Plot the residuals versus yˆ and interpret the display. (d) Calculate Cook’s distance for each observation and provide an interpretation of this statistic. 12-61. Consider the regression model fit to the X-ray inspection data in Exercise 12-11. Use rads as the response. (a) What proportion of total variability is explained by this model? (b) Construct a normal probability plot of the residuals. What conclusion can you draw from this plot? (c) Plot the residuals versus yˆ and versus each regressor, and comment on model adequacy. (d) Calculate Cook’s distance for the observations in this data set. Are there any influential points in these data? 12-62. Consider the regression model fit to the arsenic data in Exercise 12-12. Use arsenic in nails as the response and age, drink use, and cook use as the regressors. (a) What proportion of total variability is explained by this model? (b) Construct a normal probability plot of the residuals. What conclusion can you draw from this plot? (c) Plot the residuals versus yˆ and versus each regressor, and comment on model adequacy.

(d) Calculate Cook’s distance for the observations in this data set. Are there any influential points in these data? 12-63. Consider the regression model fit to the coal and limestone mixture data in Exercise 12-13. Use density as the response. (a) What proportion of total variability is explained by this model? (b) Construct a normal probability plot of the residuals. What conclusion can you draw from this plot? (c) Plot the residuals versus yˆ and versus each regressor, and comment on model adequacy. (d) Calculate Cook’s distance for the observations in this data set. Are there any influential points in these data? 12-64. Consider the regression model fit to the nisin extraction data in Exercise 12-14. (a) What proportion of total variability is explained by this model? (b) Construct a normal probability plot of the residuals. What conclusion can you draw from this plot? (c) Plot the residuals versus yˆ and versus each regressor, and comment on model adequacy. (d) Calculate Cook’s distance for the observations in this data set. Are there any influential points in these data? 12-65. Consider the regression model fit to the grey range modulation data in Exercise 12-15. Use the useful range as the response. (a) What proportion of total variability is explained by this model? (b) Construct a normal probability plot of the residuals. What conclusion can you draw from this plot? (c) Plot the residuals versus yˆ and versus each regressor, and comment on model adequacy. (d) Calculate Cook’s distance for the observations in this data set. Are there any influential points in these data? 12-66. Consider the stack loss data in Exercise 12-16. (a) What proportion of total variability is explained by this model? (b) Construct a normal probability plot of the residuals. What conclusion can you draw from this plot? (c) Plot the residuals versus yˆ and versus each regressor, and comment on model adequacy. (d) Calculate Cook’s distance for the observations in this data set. Are there any influential points in these data? 12-67. Consider the bearing wear data in Exercise 12-19. (a) Find the value of R2 when the model uses the regressors x1 and x2. (b) What happens to the value of R2 when an interaction term x1 x2 is added to the model? Does this necessarily imply that adding the interaction term is a good idea? 12-68. Fit a model to the response PITCH in the heat treating data of Exercise 12-10 using new regressors x1  SOAKTIME

SOAKPCT and x2  DIFFTIME DIFFPCT.

JWCL232_c12_449-512.qxd

490

1/16/10

11:16 AM

Page 490

CHAPTER 12 MULTIPLE LINEAR REGRESSION

(a) Calculate the R2 for this model and compare it to the value of R2 from the original model in Exercise 12-10. Does this provide some information about which model is preferable? (b) Plot the residuals from this model versus yˆ and on a normal probability scale. Comment on model adequacy. (c) Find the values of Cook’s distance measure. Are any observations unusually influential? 12-69. Consider the semiconductor HFE data in Exercise 12-9. (a) Plot the residuals from this model versus yˆ . Comment on the information in this plot. (b) What is the value of R2 for this model? (c) Refit the model using log HFE as the response variable. (d) Plot the residuals versus predicted log HFE for the model in part (c). Does this give any information about which model is preferable? (e) Plot the residuals from the model in part (d) versus the regressor x3. Comment on this plot. (f) Refit the model to log HFE using x1, x2, and 1兾x3, as the regressors. Comment on the effect of this change in the model.

12-70. Consider the regression model for the NHL data from Exercise 12-18. (a) Fit a model using GF as the only regressor. (b) How much variability is explained by this model? (c) Plot the residuals versus yˆ and comment on model adequacy. (d) Plot the residuals from part (a) versus PPGF, the points scored while in power play. Does this indicate that the model would be better if this variable were included? 12-71. The diagonal elements of the hat matrix are often used to denote leverage —that is, a point that is unusual in its location in the x-space and that may be influential. Generally, the ith point is called a leverage point if its hat diagonal hii exceeds 2p/n, which is twice the average size of all the hat diagonals. Recall that p ⫽ k ⫹ 1. (a) Table 12-12 contains the hat diagonal for the wire bond pull strength data used in Example 12-1. Find the average size of these elements. (b) Based on the criterion above, are there any observations that are leverage points in the data set?

12-6 ASPECTS OF MULTIPLE REGRESSION MODELING In this section we briefly discuss several other aspects of building multiple regression models. For more extensive presentations of these topics and additional examples refer to Montgomery, Peck, and Vining (2006) and Myers (1990).

12-6.1 Polynomial Regression Models The linear model Y ⫽ X␤ ⫹ ⑀ is a general model that can be used to fit any relationship that is linear in the unknown parameters ␤. This includes the important class of polynomial regression models. For example, the second-degree polynomial in one variable Y ⫽ ␤0 ⫹ ␤1x ⫹ ␤11x2 ⫹ ⑀

(12-46)

and the second-degree polynomial in two variables Y ⫽ ␤0 ⫹ ␤1x1 ⫹ ␤2 x2 ⫹ ␤11 x21 ⫹ ␤22 x22 ⫹ ␤12 x1x2 ⫹ ⑀

(12-47)

are linear regression models. Polynomial regression models are widely used when the response is curvilinear, because the general principles of multiple regression can be applied. The following example illustrates some of the types of analyses that can be performed.

EXAMPLE 12-12 Airplane Sidewall Panels Sidewall panels for the interior of an airplane are formed in a 1500-ton press. The unit manufacturing cost varies with the production lot size. The data shown below give the average cost per unit (in hundreds of dollars) for this product ( y) and the production lot size (x). The scatter diagram, shown in Fig. 12-11, indicates that a second-order polynomial may be appropriate.

y

1.81

1.70

1.65

1.55

1.48

1.40

x

20

25

30

35

40

50

y

1.30

1.26

1.24

1.21

1.20

1.18

x

60

65

70

75

80

90

JWCL232_c12_449-512.qxd

1/15/10

10:08 PM

Page 491

491

12-6 ASPECTS OF MULTIPLE REGRESSION MODELING 1.90 1.80

Average cost per unit, y

1.70 1.60 1.50 1.40 1.30 1.20 1.10 1.00

Figure 12-11 Data for Example 12-11.

20

30

40

50

60

70

80

ˆ  X¿y gives the fitted Solving the normal equations X¿X␤ model

We will fit the model Y  0  1x  11x2  

yˆ  2.19826629  0.02252236x  0.00012507 x 2

The y vector, the model matrix X and the ␤ vector are as follows:

y

1.81 1.70 1.65 1.55 1.48 1.40 1.30 1.26 1.24 1.21 1.20 1.18

X

90

Lot size, x

1 1 1 1 1 1 1 1 1 1 1 1

20 25 30 35 40 50 60 65 70 75 80 90

400 625 900 1225 1600 2500 3600 4225 4900 5625 6400 8100

0 ␤  £ 1 § 11

˛

Conclusions: The test for significance of regression is shown in Table 12-13. Since f0  1762.3 is significant at 1%, we conclude that at least one of the parameters 1 and 11 is not zero. Furthermore, the standard tests for model adequacy do not reveal any unusual behavior, and we would conclude that this is a reasonable model for the sidewall panel cost data.

In fitting polynomials, we generally like to use the lowest-degree model consistent with the data. In this example, it would seem logical to investigate the possibility of dropping the quadratic term from the model. That is, we would like to test H0: 11  0 H1: 11  0 Table 12-13 Test for Significance of Regression for the Second-Order Model in Example 12-12 Source of Variation

Sum of Squares

Degrees of Freedom

Mean Square

Regression Error Total

0.52516 0.00134 0.5265

2 9 11

0.26258 0.00015

f0

P-value

1762.28

2.12E-12

JWCL232_c12_449-512.qxd

492

1/15/10

10:08 PM

Page 492

CHAPTER 12 MULTIPLE LINEAR REGRESSION

Table 12-14 Analysis of Variance for Example 12-12, Showing the Test for H0: 11  0 Source of Variation Regression Linear Quadratic Error Total

Sum of Squares

SSR 11,11 0 0 2  0.52516 SSR 11 0 0 2  0.49416 SSR 111 0 0,1 2  0.03100 0.00133 0.5265

Degrees of Freedom

Mean Square

2 1 1 9 11

0.26258 0.49416 0.03100 0.00015

f0

P-value

1767.40 2236.12 208.67

2.09E-12 7.13E-13 1.56E-7

The general regression significance test can be used to test this hypothesis. We need to determine the “extra sum of squares” due to 11, or SSR 111 0 1,0 2  SSR 11,11 0 0 2  SSR 11 0 0 2

The sum of squares SSR 11,11 0 0 2  0.52516 from Table 12-13. To find SSR 11 0 0 2 , we fit a simple linear regression model to the original data, yielding yˆ  1.90036313  0.00910056x It can be easily verified that the regression sum of squares for this model is SSR 11 0 0 2  0.4942 Therefore, the extra sum of the squares due to 11, given that 1 and 0 are in the model, is SSR 111 0 1,0 2  SSR 11,11 0 0 2  SSR 11 0 0 2  0.5252  0.4942  0.031 The analysis of variance, with the test of H0: 11  0 incorporated into the procedure, is displayed in Table 12-14. Note that the quadratic term contributes significantly to the model.

12-6.2 Categorical Regressors and Indicator Variables The regression models presented in previous sections have been based on quantitative variables, that is, variables that are measured on a numerical scale. For example, variables such as temperature, pressure, distance, and voltage are quantitative variables. Occasionally, we need to incorporate categorical, or qualitative, variables in a regression model. For example, suppose that one of the variables in a regression model is the operator who is associated with each observation yi. Assume that only two operators are involved. We may wish to assign different levels to the two operators to account for the possibility that each operator may have a different effect on the response. The usual method of accounting for the different levels of a qualitative variable is to use indicator variables. For example, to introduce the effect of two different operators into a regression model, we could define an indicator variable as follows: x e

0 if the observation is from operator 1 1 if the observation is from operator 2

JWCL232_c12_449-512.qxd

1/16/10

8:29 AM

Page 493

12-6 ASPECTS OF MULTIPLE REGRESSION MODELING

493

In general, a qualitative variable with r-levels can be modeled by r  1 indicator variables, which are assigned the value of either zero or one. Thus, if there are three operators, the different levels will be accounted for by the two indicator variables defined as follows: x1

x2

0 1 0

0 0 1

if the observation is from operator 1 if the observation is from operator 2 if the observation is from operator 3

Indicator variables are also referred to as dummy variables. The following example [from Montgomery, Peck, and Vining (2006)] illustrates some of the uses of indicator variables; for other applications, see Montgomery, Peck, and Vining (2006). EXAMPLE 12-13 Surface Finish A mechanical engineer is investigating the surface finish of metal parts produced on a lathe and its relationship to the speed (in revolutions per minute) of the lathe. The data are shown in Table 12-15. Note that the data have been collected using two different types of cutting tools. Since the type of cutting tool likely affects the surface finish, we will fit the model Y  0  1x1  2x2   where Y is the surface finish, x1 is the lathe speed in revolutions per minute, and x2 is an indicator variable denoting the type of cutting tool used; that is, x2  e

0, for tool type 302 1, for tool type 416

The parameters in this model may be easily interpreted. If x2  0, the model becomes Y  0  1x1   which is a straight-line model with slope 1 and intercept 0. However, if x2  1, the model becomes Y  0  1x1  2 112    10  2 2  1x1   which is a straight-line model with slope 1 and intercept 0  2. Thus, the model Y  0  1 x  2 x2   implies that surface finish is linearly related to lathe speed and that the slope 1 does not depend on the type of cutting tool used. However, the type of cutting tool does affect the intercept, and 2 indicates the change in the intercept associated with a change in tool type from 302 to 416. The model matrix X and y vector for this problem are as follows:

X

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

225 200 250 245 235 237 265 259 221 218 224 212 248 260 243 238 224 251 232 216

0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1

y

45.44 42.03 50.10 48.75 47.92 47.79 52.26 50.52 45.58 44.78 33.50 31.23 37.52 37.13 34.70 33.92 32.13 35.47 33.49 32.29

The fitted model is yˆ  14.27620  0.14115x1  13.28020x2 Conclusions: The analysis of variance for this model is shown in Table 12-16. Note that the hypothesis H0: 1 2  0 (significance of regression) would be rejected at any reasonable level of significance because the P-value is very small. This table also contains the sums of squares SSR  SSR 11,2 0 0 2  SSR 11 0 0 2  SSR 12 0 1,0 2

so a test of the hypothesis H0: 2  0 can be made. Since this hypothesis is also rejected, we conclude that tool type has an effect on surface finish.

JWCL232_c12_449-512.qxd

494

1/16/10

8:29 AM

Page 494

CHAPTER 12 MULTIPLE LINEAR REGRESSION

Table 12-15 Surface Finish Data for Example 12-13 Observation Number, i

Surface Finish yi

RPM

Type of Cutting Tool

Observation Number, i

Surface Finish yi

RPM

Type of Cutting Tool

1 2 3 4 5 6 7 8 9 10

45.44 42.03 50.10 48.75 47.92 47.79 52.26 50.52 45.58 44.78

225 200 250 245 235 237 265 259 221 218

302 302 302 302 302 302 302 302 302 302

11 12 13 14 15 16 17 18 19 20

33.50 31.23 37.52 37.13 34.70 33.92 32.13 35.47 33.49 32.29

224 212 248 260 243 238 224 251 232 216

416 416 416 416 416 416 416 416 416 416

It is also possible to use indicator variables to investigate whether tool type affects both the slope and intercept. Let the model be Y  0  1 x1  2 x2  3 x1 x2   where x2 is the indicator variable. Now if tool type 302 is used, x2  0, and the model is Y  0  1x1   If tool type 416 is used, x2  1, and the model becomes Y  0  1 x1  2  3 x1    10  2 2  11  3 2 x1   Note that 2 is the change in the intercept and that 3 is the change in slope produced by a change in tool type. Another method of analyzing these data is to fit separate regression models to the data for each tool type. However, the indicator variable approach has several advantages. First, only one regression model must be fit. Second, by pooling the data on both tool types, more degrees of freedom for error are obtained. Third, tests of both hypotheses on the parameters 2 and 3 are just special cases of the extra sum of squares method.

12-6.3 Selection of Variables and Model Building An important problem in many applications of regression analysis involves selecting the set of regressor variables to be used in the model. Sometimes previous experience or underlying theoretical considerations can help the analyst specify the set of regressor variables to use in a particular situation. Usually, however, the problem consists of selecting an appropriate set of Table 12-16 Analysis of Variance for Example 12-13 Source of Variation

Sum of Squares

Degrees of Freedom

Mean Square

Regression SSR 11 0 0 2 SSR 12 0 1,0 2 Error Total

1012.0595 130.6091 881.4504 7.7943 1019.8538

2 1 1 17 19

506.0297 130.6091 881.4504 0.4585

f0

P-value

1103.69 284.87 1922.52

1.02E-18 4.70E-12 6.24E-19

JWCL232_c12_449-512.qxd

1/16/10

8:29 AM

Page 495

12-6 ASPECTS OF MULTIPLE REGRESSION MODELING

495

regressors from a set that quite likely includes all the important variables, but we are sure that not all these candidate regressors are necessary to adequately model the response Y. In such a situation, we are interested in variable selection; that is, screening the candidate variables to obtain a regression model that contains the “best” subset of regressor variables. We would like the final model to contain enough regressor variables so that in the intended use of the model (prediction, for example) it will perform satisfactorily. On the other hand, to keep model maintenance costs to a minimum and to make the model easy to use, we would like the model to use as few regressor variables as possible. The compromise between these conflicting objectives is often called finding the “best” regression equation. However, in most problems, no single regression model is “best” in terms of the various evaluation criteria that have been proposed. A great deal of judgment and experience with the system being modeled is usually necessary to select an appropriate set of regressor variables for a regression equation. No single algorithm will always produce a good solution to the variable selection problem. Most of the currently available procedures are search techniques, and to perform satisfactorily, they require interaction with judgment by the analyst. We now briefly discuss some of the more popular variable selection techniques. We assume that there are K candidate regressors, x1, x2, p , xK, and a single response variable y. All models will include an intercept term 0, so the model with all variables included would have K  1 terms. Furthermore, the functional form of each candidate variable (for example, x1  1兾x, x2  ln x, etc.) is assumed to be correct. All Possible Regressions This approach requires that the analyst fit all the regression equations involving one candidate variable, all regression equations involving two candidate variables, and so on. Then these equations are evaluated according to some suitable criteria to select the “best” regression model. If there are K candidate regressors, there are 2K total equations to be examined. For example, if K  4, there are 24  16 possible regression equations; while if K  10, there are 210  1024 possible regression equations. Hence, the number of equations to be examined increases rapidly as the number of candidate variables increases. However, there are some very efficient computing algorithms for all possible regressions available and they are widely implemented in statistical software, so it is a very practical procedure unless the number of candidate regressors is fairly large. Look for a menu choice such as “Best Subsets” regression. Several criteria may be used for evaluating and comparing the different regression models obtained. A commonly used criterion is based on the value of R2 or the value of the 2 adjusted R2, Radj . Basically, the analyst continues to increase the number of variables in the 2 2 model until the increase in R2 or the adjusted Radj is small. Often, we will find that the Radj will stabilize and actually begin to decrease as the number of variables in the model increases. 2 Usually, the model that maximizes Radj is considered to be a good candidate for the best re2 gression equation. Because we can write Radj  1  {MSE 兾[SST兾(n  1)]} and SST兾(n  1) 2 is a constant, the model that maximizes the Radj value also minimizes the mean square error, so this is a very attractive criterion. Another criterion used to evaluate regression models is the Cp statistic, which is a measure of the total mean square error for the regression model. We define the total standardized mean square error for the regression model as 1 n 2 ˆ a E3Yi  E1Yi 2 4  2 i1 n n 1  2 e a 3E1Yi 2  E1Yˆi 2 4 2  a V 1Yˆi 2 f  i1 i1 1  2 3 1bias2 2  variance4 

p 

JWCL232_c12_449-512.qxd

496

1/16/10

8:29 AM

Page 496

CHAPTER 12 MULTIPLE LINEAR REGRESSION

We use the mean square error from the full K  1 term model as an estimate of  2; that is, ˆ 2  MSE 1K  12. Then an estimator of p is [see Montgomery, Peck, and Vining (2006) or Myers (1990) for the details]: Cp Statistic

Cp 

SSE 1p2  n  2p ˆ 2

(12-48)

If the p-term model has negligible bias, it can be shown that E 1Cp 0 zero bias2  p Therefore, the values of Cp for each regression model under consideration should be evaluated relative to p. The regression equations that have negligible bias will have values of Cp that are close to p, while those with significant bias will have values of Cp that are significantly greater than p. We then choose as the “best” regression equation either a model with minimum Cp or a model with a slightly larger Cp, that does not contain as much bias (i.e., Cp ⬵ p). The PRESS statistic can also be used to evaluate competing regression models. PRESS is an acronym for Prediction Error Sum of Squares, and it is defined as the sum of the squares of the differences between each observation yi and the corresponding predicted value based on a model fit to the remaining n  1 points, say yˆ 1i2. So PRESS provides a measure of how well the model is likely to perform when predicting new data, or data that was not used to fit the regression model. The computing formula for PRESS is Prediction Error Sum of Squares (PRESS)

n n 2 ei PRESS  a 1 yi  yˆ 1i2 2 2  a a b i1 i1 1  hii

where ei  yi  yˆ i is the usual residual. Thus PRESS is easy to calculate from the standard least squares regression results. Models that have small values of PRESS are preferred.

EXAMPLE 12-14 Wine Quality Table 12-17 presents data on taste-testing 38 brands of pinot noir wine (the data were first reported in an article by Kwan, Kowalski, and Skogenboe in an article in the Journal of Agricultural and Food Chemistry, Vol. 27, 1979, and it also appears as one of the default data sets in Minitab). The response variable is y  quality, and we wish to find the “best” regression equation that relates quality to the other five parameters. Figure 12-12 is the matrix of scatter plots for the wine quality data, as constructed by Minitab. We notice that there are some indications of possible linear relationships between quality and the regressors, but there is no obvious visual impression of which regressors would be appropriate. Table 12-18 lists the all possible regressions output from Minitab. In this analysis,

we asked Minitab to present the best three equations for each 2 subset size. Note that Minitab reports the values of R2, Radj , Cp, and S  1MSE for each model. From Table 12-18 we see that the three-variable equation with x2  aroma, x4  flavor, and x5  oakiness produces the minimum Cp equation, whereas the four-variable model, which adds x1  clarity to the previous 2 three regressors, results in maximum Radj (or minimum MSE). The three-variable model is yˆ  6.47  0.580 x2  1.20 x4  0.602 x5 and the four-variable model is yˆ  4.99  1.79 x1  0.530 x2  1.26 x4  0.659 x5

JWCL232_c12_449-512.qxd

1/16/10

8:29 AM

Page 497

12-6 ASPECTS OF MULTIPLE REGRESSION MODELING

497

Table 12-17 Wine Quality Data

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38

x1 Clarity

x2 Aroma

x3 Body

x4 Flavor

x5 Oakiness

y Quality

1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 0.5 0.8 0.7 1.0 0.9 1.0 1.0 1.0 0.9 0.9 1.0 0.7 0.7 1.0 1.0 1.0 1.0 1.0 1.0 1.0 0.8 1.0 1.0 0.8 0.8 0.8 0.8

3.3 4.4 3.9 3.9 5.6 4.6 4.8 5.3 4.3 4.3 5.1 3.3 5.9 7.7 7.1 5.5 6.3 5.0 4.6 3.4 6.4 5.5 4.7 4.1 6.0 4.3 3.9 5.1 3.9 4.5 5.2 4.2 3.3 6.8 5.0 3.5 4.3 5.2

2.8 4.9 5.3 2.6 5.1 4.7 4.8 4.5 4.3 3.9 4.3 5.4 5.7 6.6 4.4 5.6 5.4 5.5 4.1 5.0 5.4 5.3 4.1 4.0 5.4 4.6 4.0 4.9 4.4 3.7 4.3 3.8 3.5 5.0 5.7 4.7 5.5 4.8

3.1 3.5 4.8 3.1 5.5 5.0 4.8 4.3 3.9 4.7 4.5 4.3 7.0 6.7 5.8 5.6 4.8 5.5 4.3 3.4 6.6 5.3 5.0 4.1 5.7 4.7 5.1 5.0 5.0 2.9 5.0 3.0 4.3 6.0 5.5 4.2 3.5 5.7

4.1 3.9 4.7 3.6 5.1 4.1 3.3 5.2 2.9 3.9 3.6 3.6 4.1 3.7 4.1 4.4 4.6 4.1 3.1 3.4 4.8 3.8 3.7 4.0 4.7 4.9 5.1 5.1 4.4 3.9 6.0 4.7 4.5 5.2 4.8 3.3 5.8 3.5

9.8 12.6 11.9 11.1 13.3 12.8 12.8 12.0 13.6 13.9 14.4 12.3 16.1 16.1 15.5 15.5 13.8 13.8 11.3 7.9 15.1 13.5 10.8 9.5 12.7 11.6 11.7 11.9 10.8 8.5 10.7 9.1 12.1 14.9 13.5 12.2 10.3 13.2

These models should now be evaluated further using residuals plots and the other techniques discussed earlier in the chapter, to see if either model is satisfactory with respect to the underlying assumptions and to determine if one of them is preferable. It turns out that the residual plots do not reveal any major problems with either model. The value of PRESS for the three-variable model is 56.0524 and for the four-variable model it is 60.3327. Since PRESS is smaller in the model with three regressors, and since it is the model with the smallest number of predictors, it would likely be the preferred choice.

JWCL232_c12_449-512.qxd

498

1/16/10

8:29 AM

Page 498

CHAPTER 12 MULTIPLE LINEAR REGRESSION

14.05

Quality

9.95 0.875

Clarity

0.625 6.6

Aroma

4.4 5.6

Body

3.6 5.975

Flavor

3.925 5.225

Oakiness

3.675 5

9.9

.05

14

25

0.6

75

0.8

4.4

6.6

3.6

5.6

25

3.9

75

5.9

75

3.6

Figure 12-12 A matrix of scatter plots from Minitab for the wine quality data.

Table 12-18 Minitab All Possible Regressions Output for the Wine Quality Data Best Subsets Regression: Quality versus Clarity, Aroma, . . . Response is Quality

Vars 1 1 1 2 2 2 3 3 3 4 4 4 5

R-Sq 62.4 50.0 30.1 66.1 65.9 63.3 70.4 68.0 66.5 71.5 70.5 69.3 72.1

R-Sq (adj) 61.4 48.6 28.2 64.2 63.9 61.2 67.8 65.2 63.5 68.0 66.9 65.6 67.7

C–p 9.0 23.2 46.0 6.8 7.1 10.0 3.9 6.6 8.4 4.7 5.8 7.1 6.0

S 1.2712 1.4658 1.7335 1.2242 1.2288 1.2733 1.1613 1.2068 1.2357 1.1568 1.1769 1.1996 1.1625

C l a r i t y X X X X X X X X X X X X X

A r B o o md a y X X X X X X X X X X X X X X X X X X X X X X X X X X

F l a v o r X X X X X X X X X X X X X

O a k i n e s s X X X X X X X X X X X X X

25

5.2

JWCL232_c12_449-512.qxd

1/16/10

8:29 AM

Page 499

12-6 ASPECTS OF MULTIPLE REGRESSION MODELING

499

Stepwise Regression Stepwise regression is probably the most widely used variable selection technique. The procedure iteratively constructs a sequence of regression models by adding or removing variables at each step. The criterion for adding or removing a variable at any step is usually expressed in terms of a partial F-test. Let fin be the value of the F-random variable for adding a variable to the model, and let fout be the value of the F-random variable for removing a variable from the model. We must have fin fout, and usually fin  fout. Stepwise regression begins by forming a one-variable model using the regressor variable that has the highest correlation with the response variable Y. This will also be the regressor producing the largest F-statistic. For example, suppose that at this step, x1 is selected. At the second step, the remaining K  1 candidate variables are examined, and the variable for which the partial F-statistic Fj 

SSR 1j 0 1,0 2 MSE 1xj, x1 2

(12-49)

is a maximum is added to the equation, provided that fj fin. In equation 12-49, MSE (xj, x1) denotes the mean square for error for the model containing both x1 and xj. Suppose that this procedure indicates that x2 should be added to the model. Now the stepwise regression algorithm determines whether the variable x1 added at the first step should be removed. This is done by calculating the F-statistic F1 

SSR 11 0 2,0 2 MSE 1x1, x2 2

(12-50)

If the calculated value f1  fout, the variable x1 is removed; otherwise it is retained, and we would attempt to add a regressor to the model containing both x1 and x2. In general, at each step the set of remaining candidate regressors is examined, and the regressor with the largest partial F-statistic is entered, provided that the observed value of f exceeds fin. Then the partial F-statistic for each regressor in the model is calculated, and the regressor with the smallest observed value of F is deleted if the observed f  fout. The procedure continues until no other regressors can be added to or removed from the model. Stepwise regression is almost always performed using a computer program. The analyst exercises control over the procedure by the choice of fin and fout. Some stepwise regression computer programs require that numerical values be specified for fin and fout. Since the number of degrees of freedom on MSE depends on the number of variables in the model, which changes from step to step, a fixed value of fin and fout causes the type I and type II error rates to vary. Some computer programs allow the analyst to specify the type I error levels for fin and fout. However, the “advertised” significance level is not the true level, because the variable selected is the one that maximizes (or minimizes) the partial F-statistic at that stage. Sometimes it is useful to experiment with different values of fin and fout (or different advertised type I error rates) in several different runs to see if this substantially affects the choice of the final model. EXAMPLE 12-15 Wine Quality Stepwise Regression Table 12-19 gives the Minitab stepwise regression output for the wine quality data. Minitab uses fixed values of for entering and removing variables. The default level is  0.15 for both decisions. The output in Table 12-19 uses the default value. Notice that the variables were entered in the order Flavor (step 1),

Oakiness (step 2), and Aroma (step 3) and that no variables were removed. No other variable could be entered, so the algorithm terminated. This is the three-variable model found by all possible regressions that results in a minimum value of Cp.

JWCL232_c12_449-512.qxd

500

1/16/10

8:29 AM

Page 500

CHAPTER 12 MULTIPLE LINEAR REGRESSION

Table 12-19 Minitab Stepwise Regression Output for the Wine Quality Data Stepwise Regression: Quality versus Clarity, Aroma, . . . Alpha-to-Enter: 0.15 Alpha-to-Remove: 0.15 Response is Quality on 5 predictors, with N  38 Step Constant

1 4.941

2 6.912

3 6.467

Flavor T-Value P-Value

1.57 7.73 0.000

1.64 8.25 0.000

1.20 4.36 0.000

0.54 1.95 0.059

0.60 2.28 0.029

Oakiness T-Value P-Value Aroma T-Value P-Value S R-Sq R-Sq(adj) C–p

0.58 2.21 0.034 1.27 62.42 61.37 9.0

1.22 66.11 64.17 6.8

1.16 70.38 67.76 3.9

Forward Selection The forward selection procedure is a variation of stepwise regression and is based on the principle that regressors should be added to the model one at a time until there are no remaining candidate regressors that produce a significant increase in the regression sum of squares. That is, variables are added one at a time as long as their partial F-value exceeds fin. Forward selection is a simplification of stepwise regression that omits the partial F-test for deleting variables from the model that have been added at previous steps. This is a potential weakness of forward selection; that is, the procedure does not explore the effect that adding a regressor at the current step has on regressor variables added at earlier steps. Notice that if we were to apply forward selection to the wine quality data, we would obtain exactly the same results as we did with stepwise regression in Example 12-15, since stepwise regression terminated without deleting a variable. Backward Elimination The backward elimination algorithm begins with all K candidate regressors in the model. Then the regressor with the smallest partial F-statistic is deleted if this F-statistic is insignificant, that is, if f  fout. Next, the model with K  1 regressors is fit, and the next regressor for potential elimination is found. The algorithm terminates when no further regressor can be deleted. Table 12-20 shows the Minitab output for backward elimination applied to the wine quality data. The value for removing a variable is  0.10. Notice that this procedure removes Body at step 1 and then Clarity at step 2, terminating with the three-variable model found previously. Some Comments on Final Model Selection We have illustrated several different approaches to the selection of variables in multiple linear regression. The final model obtained from any model-building procedure should be subjected

JWCL232_c12_449-512.qxd

1/16/10

8:29 AM

Page 501

12-6 ASPECTS OF MULTIPLE REGRESSION MODELING

501

Table 12-20 Minitab Backward Elimination Output for the Wine Quality Data Stepwise Regression: Quality versus Clarity, Aroma, . . . Backward elimination. Alpha-to-Remove: 0.1 Response is Quality on 5 predictors, with N = 38 Step Constant

1 3.997

2 4.986

3 6.467

Clarity T-Value P-Value

2.3 1.35 0.187

1.8 1.12 0.269

Aroma T-Value P-Value

0.48 1.77 0.086

0.53 2.00 0.054

0.58 2.21 0.034

Body T-Value P-Value

0.27 0.82 0.418

Flavor T-Value P-Value

1.17 3.84 0.001

1.26 4.52 0.000

1.20 4.36 0.000

Oakiness T-Value P-Value

0.68 2.52 0.017

0.66 2.46 0.019

0.60 2.28 0.029

S R-Sq R-Sq(adj) C–p

1.16 72.06 67.69 6.0

1.16 71.47 68.01 4.7

1.16 70.38 67.76 3.9

to the usual adequacy checks, such as residual analysis, lack-of-fit testing, and examination of the effects of influential points. The analyst may also consider augmenting the original set of candidate variables with cross-products, polynomial terms, or other transformations of the original variables that might improve the model. A major criticism of variable selection methods such as stepwise regression is that the analyst may conclude there is one “best” regression equation. Generally, this is not the case, because several equally good regression models can often be used. One way to avoid this problem is to use several different model-building techniques and see if different models result. For example, we have found the same model for the wine quality data using stepwise regression, forward selection, and backward elimination. The same model was also one of the two best found from all possible regressions. The results from variable selection methods frequently do not agree, so this is a good indication that the threevariable model is the best regression equation. If the number of candidate regressors is not too large, the all-possible regressions method is recommended. We usually recommend using the minimum MSE and Cp evaluation criteria in conjunction with this procedure. The all-possible regressions approach can find the “best” regression equation with respect to these criteria, while stepwise-type methods offer no such assurance. Furthermore, the all-possible regressions procedure is not distorted by dependencies among the regressors, as stepwise-type methods are.

JWCL232_c12_449-512.qxd

502

1/16/10

8:29 AM

Page 502

CHAPTER 12 MULTIPLE LINEAR REGRESSION

12-6.4 Multicollinearity In multiple regression problems, we expect to find dependencies between the response variable Y and the regressors xj. In most regression problems, however, we find that there are also dependencies among the regressor variables xj. In situations where these dependencies are strong, we say that multicollinearity exists. Multicollinearity can have serious effects on the estimates of the regression coefficients and on the general applicability of the estimated model. The effects of multicollinearity may be easily demonstrated. The diagonal elements of the matrix C  (X X)1 can be written as Cjj 

1 11  R2j 2

j  1, 2, p , k

where R 2j is the coefficient of multiple determination resulting from regressing xj on the other k  1 regressor variables. We can think of Rj2 as a measure of the correlation between xj and the other regressors. Clearly, the stronger the linear dependency of xj on the remaining regressor variables, and hence the stronger the multicollinearity, the larger the value of R2j will ˆ is “inflated’’ ˆ 2  2 C . Therefore, we say that the variance of  be. Recall that V 1 j jj j 2 1 by the quantity 11  R j 2 . Consequently, we define the variance inflation factor for j as Variance Inflation Factor (VIF)

VIF 1j 2 

1 11  R 2j 2

j  1, 2, . . . , k

(12-51)

These factors are an important measure of the extent to which multicollinearity is present. If the columns of the model matrix X are orthogonal, then the regressors are completely uncorrelated, and the variance inflation factors will all be unity. So any VIF that exceeds one indicates some level of multicollinearity in the data. Although the estimates of the regression coefficients are very imprecise when multicollinearity is present, the fitted model equation may still be useful. For example, suppose we wish to predict new observations on the response. If these predictions are interpolations in the original region of the x-space where the multicollinearity is in effect, satisfactory predictions will often be obtained, because while individual j may be poorly estimated, the function k g j1  j x ij may be estimated quite well. On the other hand, if the prediction of new observations requires extrapolation beyond the original region of the x-space where the data were collected, generally we would expect to obtain poor results. Extrapolation usually requires good estimates of the individual model parameters. Multicollinearity arises for several reasons. It will occur when the analyst collects data such that a linear constraint holds approximately among the columns of the X matrix. For example, if four regressor variables are the components of a mixture, such a constraint will always exist because the sum of the components is always constant. Usually, these constraints do not hold exactly, and the analyst might not know that they exist. The presence of multicollinearity can be detected in several ways. Two of the more easily understood of these will be discussed briefly. 1.

The variance inflation factors, defined in Equation 12-51, are very useful measures of multicollinearity. The larger the variance inflation factor, the more severe the multicollinearity. Some authors have suggested that if any variance inflation factor exceeds 10, multicollinearity is a problem. Other authors consider this value too liberal and suggest that the variance inflation factors should not exceed 4 or 5. Minitab will calculate the variance inflation factors. Table 12-4 presents the Minitab

JWCL232_c12_449-512.qxd

1/16/10

8:29 AM

Page 503

12-6 ASPECTS OF MULTIPLE REGRESSION MODELING

503

multiple regression output for the wire bond pull strength data. Since both VIF1 and VIF2 are small, there is no problem with multicollinearity. 2. If the F-test for significance of regression is significant, but tests on the individual regression coefficients are not significant, multicollinearity may be present. Several remedial measures have been proposed for solving the problem of multicollinearity. Augmenting the data with new observations specifically designed to break up the approximate linear dependencies that currently exist is often suggested. However, this is sometimes impossible because of economic reasons or because of the physical constraints that relate the xj. Another possibility is to delete certain variables from the model, but this approach has the disadvantage of discarding the information contained in the deleted variables. Since multicollinearity primarily affects the stability of the regression coefficients, it would seem that estimating these parameters by some method that is less sensitive to multicollinearity than ordinary least squares would be helpful. Several methods have been suggested. One alternative to ordinary least squares, ridge regression, can be useful in combating multicollinearity. For more details on ridge regression, there are more extensive presentations in Montgomery, Peck, and Vining (2006) and Myers (1990). EXERCISES FOR SECTION 12-6 12-72. An article entitled “A Method for Improving the Accuracy of Polynomial Regression Analysis’’ in the Journal of Quality Technology (1971, pp. 149–155) reported the following data on y  ultimate shear strength of a rubber compound (psi) and x  cure temperature (°F). y

770

800

840

810

x

280

284

292

295

y

735

640

590

560

x

298

305

308

315

(a) (b) (c) (d)

Fit a second-order polynomial to these data. Test for significance of regression using  0.05. Test the hypothesis that 11  0 using  0.05. Compute the residuals from part (a) and use them to evaluate model adequacy. 12-73. Consider the following data, which result from an experiment to determine the effect of x  test time in hours at a particular temperature on y  change in oil viscosity: (a) Fit a second-order polynomial to the data. y

1.42

1.39

1.55

1.89

2.43

x

.25

.50

.75

1.00

1.25

y

3.15

4.05

5.15

6.43

7.89

x

1.50

1.75

2.00

2.25

2.50

(b) Test for significance of regression using  0.05. (c) Test the hypothesis that 11  0 using  0.05. (d) Compute the residuals from part (a) and use them to evaluate model adequacy.

12-74. The following data were collected during an experiment to determine the change in thrust efficiency ( y, in percent) as the divergence angle of a rocket nozzle (x) changes: y

24.60

24.71

23.90

39.50

39.60

57.12

x

4.0

4.0

4.0

5.0

5.0

6.0

y

67.11

67.24

67.15

77.87

80.11

84.67

x

6.5

6.5

6.75

7.0

7.1

7.3

(a) Fit a second-order model to the data. (b) Test for significance of regression and lack of fit using  0.05. (c) Test the hypothesis that 11  0, using  0.05. (d) Plot the residuals and comment on model adequacy. (e) Fit a cubic model, and test for the significance of the cubic term using  0.05. 12-75. An article in the Journal of Pharmaceuticals Sciences (Vol. 80, 1991, pp. 971–977) presents data on the observed mole fraction solubility of a solute at a constant temperature and the dispersion, dipolar, and hydrogen bonding Hansen partial solubility parameters. The data are as shown in the following table, where y is the negative logarithm of the mole fraction solubility, x1 is the dispersion partial solubility, x2 is the dipolar partial solubility, and x3 is the hydrogen bonding partial solubility. (a) Fit the model Y  0  1 x 1   2 x 2  3 x3  12 x1 x2 13 x1 x 3  23 x 2 x 3  11 x 21   22 x 22  33x23  . (b) Test for significance of regression using  0.05. (c) Plot the residuals and comment on model adequacy. (d) Use the extra sum of squares method to test the contribution of the second-order terms using  0.05.

JWCL232_c12_449-512.qxd

504

1/16/10

8:29 AM

Page 504

CHAPTER 12 MULTIPLE LINEAR REGRESSION

Observation Number

y

x1

x2

x3

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26

0.22200 0.39500 0.42200 0.43700 0.42800 0.46700 0.44400 0.37800 0.49400 0.45600 0.45200 0.11200 0.43200 0.10100 0.23200 0.30600 0.09230 0.11600 0.07640 0.43900 0.09440 0.11700 0.07260 0.04120 0.25100 0.00002

7.3 8.7 8.8 8.1 9.0 8.7 9.3 7.6 10.0 8.4 9.3 7.7 9.8 7.3 8.5 9.5 7.4 7.8 7.7 10.3 7.8 7.1 7.7 7.4 7.3 7.6

0.0 0.0 0.7 4.0 0.5 1.5 2.1 5.1 0.0 3.7 3.6 2.8 4.2 2.5 2.0 2.5 2.8 2.8 3.0 1.7 3.3 3.9 4.3 6.0 2.0 7.8

0.0 0.3 1.0 0.2 1.0 2.8 1.0 3.4 0.3 4.1 2.0 7.1 2.0 6.8 6.6 5.0 7.8 7.7 8.0 4.2 8.5 6.6 9.5 10.9 5.2 20.7

12-76. Consider the arsenic concentration data in Exercise 12-10. (a) Discuss how you would model the information about the person’s sex. (b) Fit a regression model to the arsenic in nails using age, drink use, cook use, and the person’s sex as the regressors. (c) Is there evidence that the person’s sex affects arsenic in the nails? Why? 12-77. Consider the gasoline mileage data in Exercise 12-7. (a) Discuss how you would model the information about the type of transmission in the car. (b) Fit a regression model to the gasoline mileage using cid, etw, and the type of transmission in the car as the regressors. (c) Is there evidence that the type of transmission (L4, L5, or M6) affects gasoline mileage performance? 12-78. Consider the surface finish data in Example 12-15. Test the hypothesis that two different regression models (with different slopes and intercepts) are required to adequately model the data. Use indicator variables in answering this question.

12-79. Consider the X-ray inspection data in Exercise 12-11. Use rads as the response. Build regression models for the data using the following techniques: (a) All possible regressions. (b) Stepwise regression. (c) Forward selection. (d) Backward elimination. (e) Comment on the models obtained. Which model would you prefer? Why? 12-80. Consider the electric power data in Exercise 12-6. Build regression models for the data using the following techniques: (a) All possible regressions. Find the minimum Cp and minimum MSE equations. (b) Stepwise regression. (c) Forward selection. (d) Backward elimination. (e) Comment on the models obtained. Which model would you prefer? 12-81. Consider the regression model fit to the coal and limestone mixture data in Exercise 12-13. Use density as the response. Build regression models for the data using the following techniques: (a) All possible regressions. (b) Stepwise regression. (c) Forward selection. (d) Backward elimination. (e) Comment on the models obtained. Which model would you prefer? Why? 12-82. Consider the wire bond pull strength data in Exercise 12-8. Build regression models for the data using the following methods: (a) All possible regressions. Find the minimum Cp and minimum MSE equations. (b) Stepwise regression. (c) Forward selection. (d) Backward elimination. (e) Comment on the models obtained. Which model would you prefer? 12-83. Consider the grey range modulation data in Exercise 12-15. Use the useful range as the response. Build regression models for the data using the following techniques: (a) All possible regressions. (b) Stepwise regression. (c) Forward selection. (d) Backward elimination. (e) Comment on the models obtained. Which model would you prefer? Why? 12-84. Consider the nisin extraction data in Exercise 12-14. Build regression models for the data using the following techniques: (a) All possible regressions. (b) Stepwise regression. (c) Forward selection.

JWCL232_c12_449-512.qxd

1/16/10

8:29 AM

Page 505

12-6 ASPECTS OF MULTIPLE REGRESSION MODELING

(d) Backward elimination. (e) Comment on the models obtained. Which model would you prefer? Why? 12.85. Consider the stack loss data in Exercise 12-16. Build regression models for the data using the following techniques: (a) All possible regressions. (b) Stepwise regression. (c) Forward selection. (d) Backward elimination. (e) Comment on the models obtained. Which model would you prefer? Why? (f) Remove any influential data points and repeat the model building in the previous parts? Does your conclusion in part (e) change? 12-86. Consider the NHL data in Exercise 12-18. Build regression models for these data with regressors GF through FG using the following methods: (a) All possible regressions. Find the minimum Cp and minimum MSE equations. (b) Stepwise regression. (c) Forward selection. (d) Backward elimination. (e) Which model would you prefer? 12-87. Use the football data in Exercise 12-17 to build regression models using the following techniques: (a) All possible regressions. Find the equations that minimize MSE and that minimize Cp. (b) Stepwise regression. (c) Forward selection. (d) Backward elimination. (e) Comment on the various models obtained. Which model seems “best,’’ and why? 12-88. Consider the arsenic data in Exercise 12-12. Use arsenic in nails as the response and age, drink use, and cook use as the regressors. Build regression models for the data using the following techniques: (a) All possible regressions. (b) Stepwise regression. (c) Forward selection. (d) Backward elimination. (e) Comment on the models obtained. Which model would you prefer? Why? (f) Now construct an indicator variable and add the person’s sex to the list of regressors. Repeat the model building in the previous parts. Does your conclusion in part (e) change? 12-89. Consider the gas mileage data in Exercise 12-7. Build regression models for the data from the numerical regressors using the following techniques: (a) All possible regressions. (b) Stepwise regression. (c) Forward selection. (d) Backward elimination.

505

(e) Comment on the models obtained. Which model would you prefer? Why? (f) Now construct indicator variable for trns and drv and add these to the list of regressors. Repeat the model building in the previous parts. Does your conclusion in part (e) change? 12-90. When fitting polynomial regression models, we often subtract x from each x value to produce a “centered’’ regressor x¿  x  x. This reduces the effects of dependencies among the model terms and often leads to more accurate estimates of the regression coefficients. Using the data from Exercise 12-72, fit the model Y   *0   *1 x¿   *11 1x¿ 2 2  . (a) Use the results to estimate the coefficients in the uncentered model Y   0   1 x   11 x 2  . Predict y when x  285F. Suppose that we use a standardized variable x¿  1x  x 2 sx , where sx is the standard deviation of x, in constructing a polynomial regression model. Fit the model Y  *0  *1 x¿   *11 1x¿2 2  . (b) What value of y do you predict when x  285F ? (c) Estimate the regression coefficients in the unstandardized model Y  0  1x   11x 2  . (d) What can you say about the relationship between SSE and R2 for the standardized and unstandardized models? (e) Suppose that y ¿  1 y  y 2 sy is used in the model along with x¿ . Fit the model and comment on the relationship between SSE and R2 in the standardized model and the unstandardized model. 12-91. Consider the data in Exercise 12-75. Use all the terms in the full quadratic model as the candidate regressors. (a) Use forward selection to identify a model. (b) Use backward elimination to identify a model. (c) Compare the two models obtained in parts (a) and (b). Which model would you prefer and why? 12-92. We have used a sample of 30 observations to fit a regression model. The full model has nine regressors, the variance estimate is ˆ 2  MSE  100, and R 2  0.92. (a) Calculate the F-statistic for testing significance of regression. Using = 0.05, what would you conclude? (b) Suppose that we fit another model using only four of the original regressors and that the error sum of squares for this new model is 2200. Find the estimate of 2 for this new reduced model. Would you conclude that the reduced model is superior to the old one? Why? (c) Find the value of Cp for the reduced model in part (b). Would you conclude that the reduced model is better than the old model? 12-93. A sample of 25 observations is used to fit a regression model in seven variables. The estimate of 2 for this full model is MSE  10. (a) A forward selection algorithm has put three of the original seven regressors in the model. The error sum of squares for the three-variable model is SSE  300. Based on Cp, would you conclude that the three-variable model has any remaining bias?

JWCL232_c12_449-512.qxd

506

1/16/10

8:29 AM

Page 506

CHAPTER 12 MULTIPLE LINEAR REGRESSION

(b) After looking at the forward selection model in part (a), suppose you could add one more regressor to the model. This regressor will reduce the error sum of squares to 275. Will the addition of this variable improve the model? Why?

Supplemental Exercises 12-94.

Consider the computer output below.

The regression equation is Y  517  11.5 x1  8.14 x2  10.9 x3 Predictor Constant x1 x2 x3 S  10.2560

Coef 517.46 11.4720 8.1378 10.8565

SE Coef 11.76 ? 0.1969 0.6652

RSq  ?

T ? 36.50 ? ?

P ? ? ? ?

RSq (adj)  ?

Analysis of Variance Source Regression Residual Error Total

DF ? 16 19

SS 347300 ? 348983

MS 115767 105

F ?

P ?

0.028245 0.0013329 0.0001547

12-97. Consider the engine thrust data in Exercise 12-96. Refit the model using y*  ln y as the response variable and x*3  ln x3 as the regressor (along with x4 and x5). (a) Test for significance of regression using  0.01. Find the P-value for this test and state your conclusions. (b) Use the t-statistic to test H0 : j  0 versus H1: j  0 for each variable in the model. If  0.01, what conclusions can you draw? (c) Plot the residuals versus yˆ * and versus x*3 . Comment on these plots. How do they compare with their counterparts obtained in Exercise 12-96 parts (f) and (g)? 12-98. Transient points of an electronic inverter are influenced by many factors. Table 12-21 gives data on the transient point (y, in volts) of PMOS-NMOS inverters and five candidate regressors: x1  width of the NMOS device, x2  length Table 12-21

(a) Fill in the missing values. Use bounds for the P-values. (b) Is the overall model significant at  0.05? Is it significant at  0.01? (c) Discuss the contribution of the individual regressors to the model. 12-95. Consider the following inverse of the model matrix: 0.893758 1X¿X2 1  £ 0.028245 0.017564

(f ) Plot the residuals versus yˆ. Are there any indications of inequality of variance or nonlinearity? (g) Plot the residuals versus x3. Is there any indication of nonlinearity? (h) Predict the thrust for an engine for which x3  28900, x4  170, and x5  1589.

0.0175641 0.0001547 § 0.0009108

(a) How many variables are in the regression model? (b) If the estimate of 2 is 50, what is the estimate of the variance of each regression coefficient? (c) What is the standard error of the intercept? 12-96. The data shown in Table 12-22 represent the thrust of a jet-turbine engine (y) and six candidate regressors: x1 = primary speed of rotation, x2  secondary speed of rotation, x3  fuel flow rate, x4  pressure, x5  exhaust temperature, and x6  ambient temperature at time of test. (a) Fit a multiple linear regression model using x3  fuel flow rate, x4  pressure, and x5  exhaust temperature as the regressors. (b) Test for significance of regression using  0.01. Find the P-value for this test. What are your conclusions? (c) Find the t-test statistic for each regressor. Using  0.01, explain carefully the conclusion you can draw from these statistics. (d) Find R2 and the adjusted statistic for this model. (e) Construct a normal probability plot of the residuals and interpret this graph.

Transient Point of an Electronic Inverter

Observation Number

x1

x2

x3

x4

x5

y

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

3 8 3 4 8 10 8 6 4 16 3 8 3 3 4 5 2 10 15 15 10 3 6 2 3

3 30 6 4 7 20 6 24 10 12 10 3 6 8 8 2 2 15 6 6 4 8 6 3 3

3 8 6 4 6 5 3 4 12 8 8 3 3 8 4 2 2 3 2 2 3 2 6 8 8

3 8 6 12 5 5 3 4 4 4 8 3 3 3 8 2 3 3 3 3 3 2 4 6 8

0 0 0 0 0 0 25 25 25 25 25 25 50 50 50 50 50 50 50 75 75 75 75 75 75

0.787 0.293 1.710 0.203 0.806 4.713 0.607 9.107 9.210 1.365 4.554 0.293 2.252 9.167 0.694 0.379 0.485 3.345 0.208 0.201 0.329 4.966 1.362 1.515 0.751

JWCL232_c12_449-512.qxd

1/16/10

8:29 AM

Page 507

12-6 ASPECTS OF MULTIPLE REGRESSION MODELING

Table 12-22

507

Thrust of a Jet-Turbine Engine

Observation Number

y

x1

x2

x3

x4

x5

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40

4540 4315 4095 3650 3200 4833 4617 4340 3820 3368 4445 4188 3981 3622 3125 4560 4340 4115 3630 3210 4330 4119 3891 3467 3045 4411 4203 3968 3531 3074 4350 4128 3940 3480 3064 4402 4180 3973 3530 3080

2140 2016 1905 1675 1474 2239 2120 1990 1702 1487 2107 1973 1864 1674 1440 2165 2048 1916 1658 1489 2062 1929 1815 1595 1400 2047 1935 1807 1591 1388 2071 1944 1831 1612 1410 2066 1954 1835 1616 1407

20640 20280 19860 18980 18100 20740 20305 19961 18916 18012 20520 20130 19780 19020 18030 20680 20340 19860 18950 18700 20500 20050 19680 18890 17870 20540 20160 19750 18890 17870 20460 20010 19640 18710 17780 20520 20150 19750 18850 17910

30250 30010 29780 29330 28960 30083 29831 29604 29088 28675 30120 29920 29720 29370 28940 30160 29960 29710 29250 28890 30190 29960 29770 29360 28960 30160 29940 29760 29350 28910 30180 29940 29750 29360 28900 30170 29950 29740 29320 28910

205 195 184 164 144 216 206 196 171 149 195 190 180 161 139 208 199 187 164 145 193 183 173 153 134 193 184 173 153 133 198 186 178 156 136 197 188 178 156 137

1732 1697 1662 1598 1541 1709 1669 1640 1572 1522 1740 1711 1682 1630 1572 1704 1679 1642 1576 1528 1748 1713 1684 1624 1569 1746 1714 1679 1621 1561 1729 1692 1667 1609 1552 1758 1729 1690 1616 1569

x6 99 100 97 97 97 87 87 87 85 85 101 100 100 100 101 98 96 94 94 94 101 100 100 99 100 99 99 99 99 99 102 101 101 101 101 100 99 99 99 100

JWCL232_c12_449-512.qxd

508

1/16/10

8:29 AM

Page 508

CHAPTER 12 MULTIPLE LINEAR REGRESSION

of the NMOS device, x3  width of the PMOS device, x4  length of the PMOS device, and x5  temperature (°C). (a) Fit a multiple linear regression model that uses all regressors to these data. Test for significance of regression using  0.01. Find the P-value for this test and use it to draw your conclusions. (b) Test the contribution of each variable to the model using the t-test with  0.05. What are your conclusions? (c) Delete x5 from the model. Test the new model for significance of regression. Also test the relative contribution of each regressor to the new model with the t-test. Using  0.05, what are your conclusions? (d) Notice that the MSE for the model in part (c) is smaller than the MSE for the full model in part (a). Explain why this has occurred. (e) Calculate the studentized residuals. Do any of these seem unusually large? (f ) Suppose that you learn that the second observation was recorded incorrectly. Delete this observation and refit the model using x1, x2, x3, and x4 as the regressors. Notice that the R2 for this model is considerably higher than the R2 for either of the models fitted previously. Explain why the R2 for this model has increased. (g) Test the model from part (f ) for significance of regression using  0.05. Also investigate the contribution of each regressor to the model using the t-test with  0.05. What conclusions can you draw? (h) Plot the residuals from the model in part (f ) versus yˆ and versus each of the regressors x1, x2, x3, and x4. Comment on the plots. 12-99. Consider the inverter data in Exercise 12-98. Delete observation 2 from the original data. Define new variables as follows: y*  ln y, x*1  1 1x1, x*2  1x2, x *3  1 1x3, and x*4  1x4. (a) Fit a regression model using these transformed regressors (do not use x5). (b) Test the model for significance of regression using  0.05. Use the t-test to investigate the contribution of each variable to the model (  0.05). What are your conclusions? (c) Plot the residuals versus yˆ * and versus each of the transformed regressors. Comment on the plots. 12-100. Following are data on y  green liquor (g/l) and x  paper machine speed (feet per minute) from a Kraft paper machine. (The data were read from a graph in an article in the Tappi Journal, March 1986.) y

16.0

15.8

15.6

15.5

14.8

x

1700

1720

1730

1740

1750

y

14.0

13.5

13.0

12.0

11.0

x

1760

1770

1780

1790

1795

(a) Fit the model Y   0  1 x   2 x 2   using least squares.

(b) Test for significance of regression using  0.05. What are your conclusions? (c) Test the contribution of the quadratic term to the model, over the contribution of the linear term, using an F-statistic. If  0.05, what conclusion can you draw? (d) Plot the residuals from the model in part (a) versus yˆ . Does the plot reveal any inadequacies? (e) Construct a normal probability plot of the residuals. Comment on the normality assumption. 12-101. Consider the jet engine thrust data in Exercise 12-96 and 12-97. Define the response and regressors as in Exercise 12-97. (a) Use all possible regressions to select the best regression equation, where the model with the minimum value of MSE is to be selected as “best.’’ (b) Repeat part (a) using the CP criterion to identify the best equation. (c) Use stepwise regression to select a subset regression model. (d) Compare the models obtained in parts (a), (b), and (c) above. (e) Consider the three-variable regression model. Calculate the variance inflation factors for this model. Would you conclude that multicollinearity is a problem in this model? 12-102. Consider the electronic inverter data in Exercise 12-98 and 12-99. Define the response and regressors variables as in Exercise 12-99, and delete the second observation in the sample. (a) Use all possible regressions to find the equation that minimizes Cp. (b) Use all possible regressions to find the equation that minimizes MSE. (c) Use stepwise regression to select a subset regression model. (d) Compare the models you have obtained. 12-103. A multiple regression model was used to relate y  viscosity of a chemical product to x1  temperature and x2  reaction time. The data set consisted of n  15 observations. (a) The estimated regression coefficients were ˆ 0  300.00, ˆ 1  0.85, and ˆ 2  10.40. Calculate an estimate of mean viscosity when x1  100°F and x2  2 hours. (b) The sums of squares were SST  1230.50 and SSE  120.30. Test for significance of regression using  0.05. What conclusion can you draw? (c) What proportion of total variability in viscosity is accounted for by the variables in this model? (d) Suppose that another regressor, x3  stirring rate, is added to the model. The new value of the error sum of squares is SSE  117.20. Has adding the new variable resulted in a smaller value of MSE? Discuss the significance of this result. (e) Calculate an F-statistic to assess the contribution of x3 to the model. Using  0.05, what conclusions do you reach? 12-104. Tables 12-23 and 12-24 present statistics for the Major League Baseball 2005 season (source: The Sports Network). (a) Consider the batting data. Use model-building methods to predict Wins from the other variables. Check that the assumptions for your model are valid.

JWCL232_c12_449-512.qxd

1/16/10

8:29 AM

Page 509

12-6 ASPECTS OF MULTIPLE REGRESSION MODELING

Table 12-23

509

Major League Baseball 2005 Season American League Batting

Team

W

AVG

R

H

2B

3B

HR

RBI

BB

SO

SB

GIDP

LOB

OBP

Chicago Boston LA Angels New York Cleveland Oakland Minnesota Toronto Texas Baltimore Detroit Seattle Tampa Bay Kansas City

99 95 95 95 93 88 83 80 79 74 71 69 67 56

0.262 0.281 0.27 0.276 0.271 0.262 0.259 0.265 0.267 0.269 0.272 0.256 0.274 0.263

741 910 761 886 790 772 688 775 865 729 723 699 750 701

1450 1579 1520 1552 1522 1476 1441 1480 1528 1492 1521 1408 1519 1445

253 339 278 259 337 310 269 307 311 296 283 289 289 289

23 21 30 16 30 20 32 39 29 27 45 34 40 34

200 199 147 229 207 155 134 136 260 189 168 130 157 126

713 863 726 847 760 739 644 735 834 700 678 657 717 653

435 653 447 637 503 537 485 486 495 447 384 466 412 424

1002 1044 848 989 1093 819 978 955 1112 902 1038 986 990 1008

137 45 161 84 62 31 102 72 67 83 66 102 151 53

122 135 125 125 128 148 155 126 123 145 137 115 133 139

1032 1249 1086 1264 1148 1170 1109 1118 1104 1103 1077 1076 1065 1062

0.322 0.357 0.325 0.355 0.334 0.33 0.323 0.331 0.329 0.327 0.321 0.317 0.329 0.32

National League Batting Team

W

AVG

R

H

2B

3B

HR

RBI

BB

SO

SB

GIDP

LOB

OBP

St. Louis Atlanta Houston Philadelphia Florida New York San Diego Milwaukee Washington Chicago Arizona San Francisco Cincinnati Los Angeles Colorado Pittsburgh

100 90 89 88 83 83 82 81 81 79 77 75 73 71 67 67

0.27 0.265 0.256 0.27 0.272 0.258 0.257 0.259 0.252 0.27 0.256 0.261 0.261 0.253 0.267 0.259

805 769 693 807 717 722 684 726 639 703 696 649 820 685 740 680

1494 1453 1400 1494 1499 1421 1416 1413 1367 1506 1419 1427 1453 1374 1477 1445

287 308 281 282 306 279 269 327 311 323 291 299 335 284 280 292

26 37 32 35 32 32 39 19 32 23 27 26 15 21 34 38

170 184 161 167 128 175 130 175 117 194 191 128 222 149 150 139

757 733 654 760 678 683 655 689 615 674 670 617 784 653 704 656

534 534 481 639 512 486 600 531 491 419 606 431 611 541 509 471

947 1084 1037 1083 918 1075 977 1162 1090 920 1094 901 1303 1094 1103 1092

83 92 115 116 96 153 99 79 45 65 67 71 72 58 65 73

127 146 116 107 144 103 122 137 130 131 132 147 116 139 125 130

1152 1114 1136 1251 1181 1122 1220 1120 1137 1133 1247 1093 1176 1135 1197 1193

0.339 0.333 0.322 0.348 0.339 0.322 0.333 0.331 0.322 0.324 0.332 0.319 0.339 0.326 0.333 0.322

Batting W AVG R H 2B 3B HR RBI BB SO SB GIDP

Wins Batting average Runs Hits Doubles Triples Home runs Runs batted in Walks Strikeouts Stolen bases Grounded into double play

LOB OBP

Left on base On-base percentage

Pitching ERA SV H R ER HR BB SO AVG

Earned run average Saves Hits Runs Earned runs Home runs Walks Strikeouts Opponent batting average

JWCL232_c12_449-512.qxd

510

1/16/10

8:29 AM

Page 510

CHAPTER 12 MULTIPLE LINEAR REGRESSION

Table 12-24

Major League Baseball 2005 Season American League Pitching

Team

W

ERA

SV

H

R

ER

HR

BB

SO

AVG

Chicago Boston LA Angels New York Cleveland Oakland Minnesota Toronto Texas Baltimore Detroit Seattle Tampa Bay Kansas City

99 95 95 95 93 88 83 80 79 74 71 69 67 56

3.61 4.74 3.68 4.52 3.61 3.69 3.71 4.06 4.96 4.56 4.51 4.49 5.39 5.49

54 38 54 46 51 38 44 35 46 38 37 39 43 25

1392 1550 1419 1495 1363 1315 1458 1475 1589 1458 1504 1483 1570 1640

645 805 643 789 642 658 662 705 858 800 787 751 936 935

592 752 598 718 582 594 604 653 794 724 719 712 851 862

167 164 158 164 157 154 169 185 159 180 193 179 194 178

459 440 443 463 413 504 348 444 522 580 461 496 615 580

1040 959 1126 985 1050 1075 965 958 932 1052 907 892 949 924

0.249 0.276 0.254 0.269 0.247 0.241 0.261 0.264 0.279 0263 0.272 0.268 0.28 0.291

National League Pitching Team

W

ERA

SV

H

R

ER

HR

BB

SO

AVG

St. Louis Atlanta Houston Philadelphia Florida New York San Diego Milwaukee Washington Chicago Arizona San Francisco Cincinnati Los Angeles Colorado Pittsburgh

100 90 89 88 83 83 82 81 81 79 77 75 73 71 67 67

3.49 3.98 3.51 4.21 4.16 3.76 4.13 3.97 3.87 4.19 4.84 4.33 5.15 4.38 5.13 4.42

48 38 45 40 42 38 45 46 51 39 45 46 31 40 37 35

1399 1487 1336 1379 1459 1390 1452 1382 1456 1357 1580 1456 1657 1434 1600 1456

634 674 609 726 732 648 726 697 673 714 856 745 889 755 862 769

560 639 563 672 666 599 668 635 627 671 783 695 820 695 808 706

153 145 155 189 116 135 146 169 140 186 193 151 219 182 175 162

443 520 440 487 563 491 503 569 539 576 537 592 492 471 604 612

974 929 1164 1159 1125 1012 1133 1173 997 1256 1038 972 955 1004 981 958

0.257 0.268 0.246 0.253 0.266 0.255 0.259 0.251 0.262 0.25 0.278 0.263 0.29 0.263 0.287 0.267

Batting W AVG R H 2B 3B HR RBI BB SO SB GID

Wins Batting average Runs Hits Doubles Triples Home runs Runs batted in Walks Strikeouts Stolen bases Grounded into double play

LOB OBP

Left on base On-base percentage

Pitching ERA SV H R ER HR BB SO AVG

Earned run average Saves Hits Runs Earned runs Home runs Walks Strikeouts Opponent batting average

JWCL232_c12_449-512.qxd

1/16/10

8:29 AM

Page 511

12-6 ASPECTS OF MULTIPLE REGRESSION MODELING

(b) Repeat part (a) for the pitching data. (c) Use both the batting and pitching data to build a model to predict Wins. What variables are most important? Check that the assumptions for your model are valid. 12-105. An article in the Journal of the American Ceramics Society (1992, Vol. 75, pp. 112–116) describes a process for immobilizing chemical or nuclear wastes in soil by dissolving the contaminated soil into a glass block. The authors mix CaO and Na2O with soil and model viscosity and electrical conductivity. The electrical conductivity model involves six regressors, and the sample consists of n  14 observations. (a) For the six-regressor model, suppose that SST  0.50 and R2  0.94. Find SSE and SSR , and use this information to test for significance of regression with  0.05. What are your conclusions? (b) Suppose that one of the original regressors is deleted from the model, resulting in R2  0.92. What can you conclude about the contribution of the variable that was removed? Answer this question by calculating an F-statistic. (c) Does deletion of the regressor variable in part (b) result in a smaller value of MSE for the five-variable model, in

comparison to the original six-variable model? Comment on the significance of your answer. 12-106. Exercise 12-5 introduced the hospital patient satisfaction survey data. One of the variables in that data set is a categorical variable indicating whether the patient is a medical patient or a surgical patient. Fit a model including this indicator variable to the data, using all three of the other regressors. Is there any evidence that the service the patient is on (medical versus surgical) has an impact on the reported satisfaction? 12-107. Consider the inverse model matrix shown below.

1X¿X2 1

0.125 0  ≥ 0 0

0 0.125 0 0

0 0 0.125 0

0 0 ¥ 0 0.125

(a) How many regressors are in this model? (b) What was the sample size? (c) Notice the special diagonal structure of the matrix. What does that tell you about the columns in the original X matrix?

MIND-EXPANDING EXERCISES 12-108. Consider a multiple regression model with k regressors. Show that the test statistic for significance of regression can be written as R 2k F0  11  R 2 2  1n  k  12

Suppose that n  20, k  4, and R  0.90. If  0.05, what conclusion would you draw about the relationship between y and the four regressors? 12-109. A regression model is used to relate a response y to k  4 regressors with n  20. What is the smallest value of R2 that will result in a significant regression if  0.05? Use the results of the previous exercise. Are you surprised by how small the value of R2 is? 12-110. Show that we can express the residuals from a multiple regression model as e  (I  H)y, where H  X(X ¿ X)1X ¿ . 12-111. Show that the variance of the ith residual ei in a multiple regression model is 2 11  hii 2 and that the covariance between ei and ej is 2hij, where the h’s are the elements of H  X(X X)1X ¿ . 12-112. Consider the multiple linear regression model ˆ denotes the least squares estimator of y  X  . If  ˆ , show that     R, where R  1X¿X2 1X¿ . 12-113. Constrained Least Squares. Suppose we wish to find the least squares estimator of  in the model 2

511

y  X   subject to a set of equality constraints, say, T  c. (a) Show that the estimator is ˆ  1X¿X2 1 ˆ c    T[T(XX)–1T]–1(c  Tˆ ) ˆ  (XX)–1Xy. where  (b) Discuss situations where this model might be appropriate. 12-114. Piecewise Linear Regression. Suppose that y is piecewise linearly related to x. That is, different linear relationships are appropriate over the intervals   x  x* and x*  x  . (a) Show how indicator variables can be used to fit such a piecewise linear regression model, assuming that the point x* is known. (b) Suppose that at the point x* a discontinuity occurs in the regression function. Show how indicator variables can be used to incorporate the discontinuity into the model. (c) Suppose that the point x* is not known with certainty and must be estimated. Suggest an approach that could be used to fit the piecewise linear regression model.

JWCL232_c12_449-512.qxd

512

1/16/10

8:29 AM

Page 512

CHAPTER 12 MULTIPLE LINEAR REGRESSION

IMPORTANT TERMS AND CONCEPTS All possible regressions Analysis of variance test in multiple regression Categorical variables Confidence interval on the mean response Cp statistic Extra sum of squares method Hidden extrapolation

Indicator variables Inference (test and intervals) on individual model parameters Influential observations Model parameters and their interpretation in multiple regression Multicollinearity

Multiple Regression Outliers Polynomial regression model Prediction interval on a future observation PRESS statistic Residual analysis and model adequacy checking

Significance of regression Stepwise regression and related methods Variance Inflation Factor (VIF)

JWCL232_c13_513-550.qxd

1/18/10

1:59 PM

Page 513

13

© Vasko Miokovic/iStockphoto

Design and Analysis of Single-Factor Experiments: The Analysis of Variance

Experiments are a natural part of the engineering and scientific decision-making process. Suppose, for example, that a civil engineer is investigating the effects of different curing methods on the mean compressive strength of concrete. The experiment would consist of making up several test specimens of concrete using each of the proposed curing methods and then testing the compressive strength of each specimen. The data from this experiment could be used to determine which curing method should be used to provide maximum mean compressive strength. If there are only two curing methods of interest, this experiment could be designed and analyzed using the statistical hypothesis methods for two samples introduced in Chapter 10. That is, the experimenter has a single factor of interest—curing methods— and there are only two levels of the factor. If the experimenter is interested in determining which curing method produces the maximum compressive strength, the number of specimens to test can be determined from the operating characteristic curves in Appendix Chart VII, and the t-test can be used to decide if the two means differ. Many single-factor experiments require that more than two levels of the factor be considered. For example, the civil engineer may want to investigate five different curing methods. In this chapter we show how the analysis of variance (frequently abbreviated ANOVA) can be used for comparing means when there are more than two levels of a single factor. We will also discuss randomization of the experimental runs and the important role this concept plays in the overall experimentation strategy. In the next chapter, we will show how to design and analyze experiments with several factors.

513

JWCL232_c13_513-550.qxd

514

1/18/10

10:40 AM

Page 514

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

CHAPTER OUTLINE 13-1 DESIGNING ENGINEERING EXPERIMENTS 13-2 COMPLETELY RANDOMIZED SINGLE-FACTOR EXPERIMENT 13-2.1 Example: Tensile Strength 13-2.2 Analysis of Variance

13-3 THE RANDOM-EFFECTS MODEL 13-3.1 Fixed Versus Random Factors 13-3.2 ANOVA and Variance Components 13-4 RANDOMIZED COMPLETE BLOCK DESIGN

13-2.3 Multiple Comparisons Following the ANOVA

13-4.1 Design and Statistical Analysis

13-2.4 Residual Analysis and Model Checking

13-4.3 Residual Analysis and Model Checking

13-4.2 Multiple Comparisons

13-2.5 Determining Sample Size

LEARNING OBJECTIVES After careful study of this chapter you should be able to do the following: 1. Design and conduct engineering experiments involving a single factor with an arbitrary number of levels 2. Understand how the analysis of variance is used to analyze the data from these experiments 3. Assess model adequacy with residual plots 4. Use multiple comparison procedures to identify specific differences between means 5. Make decisions about sample size in single-factor experiments 6. Understand the difference between fixed and random factors 7. Estimate variance components in an experiment involving random factors 8. Understand the blocking principle and how it is used to isolate the effect of nuisance factors 9. Design and conduct experiments involving the randomized complete block design

13-1 DESIGNING ENGINEERING EXPERIMENTS Statistically based experimental design techniques are particularly useful in the engineering world for solving many important problems: discovery of new basic phenomena that can lead to new products, and commercialization of new technology including new product development, new process development, and improvement of existing products and processes. For example, consider the development of a new process. Most processes can be described in terms of several controllable variables, such as temperature, pressure, and feed rate. By using designed experiments, engineers can determine which subset of the process variables has the greatest influence on process performance. The results of such an experiment can lead to Improved process yield Reduced variability in the process and closer conformance to nominal or target requirements Reduced design and development time Reduced cost of operation

JWCL232_c13_513-550.qxd

1/18/10

10:40 AM

Page 515

13-2 COMPLETELY RANDOMIZED SINGLE-FACTOR EXPERIMENT

515

Experimental design methods are also useful in engineering design activities, where new products are developed and existing ones are improved. Some typical applications of statistically designed experiments in engineering design include Evaluation and comparison of basic design configurations Evaluation of different materials Selection of design parameters so that the product will work well under a wide variety of field conditions (or so that the design will be robust) Determination of key product design parameters that affect product performance The use of experimental design in the engineering design process can result in products that are easier to manufacture, products that have better field performance and reliability than their competitors, and products that can be designed, developed, and produced in less time. Designed experiments are usually employed sequentially. That is, the first experiment with a complex system (perhaps a manufacturing process) that has many controllable variables is often a screening experiment designed to determine which variables are most important. Subsequent experiments are used to refine this information and determine which adjustments to these critical variables are required to improve the process. Finally, the objective of the experimenter is optimization, that is, to determine which levels of the critical variables result in the best process performance. Every experiment involves a sequence of activities: 1. 2. 3. 4.

Conjecture—the original hypothesis that motivates the experiment. Experiment—the test performed to investigate the conjecture. Analysis—the statistical analysis of the data from the experiment. Conclusion—what has been learned about the original conjecture from the experiment. Often the experiment will lead to a revised conjecture, and a new experiment, and so forth.

The statistical methods introduced in this chapter and Chapter 14 are essential to good experimentation. All experiments are designed experiments; unfortunately, some of them are poorly designed, and as a result, valuable resources are used ineffectively. Statistically designed experiments permit efficiency and economy in the experimental process, and the use of statistical methods in examining the data results in scientific objectivity when drawing conclusions.

13-2 COMPLETELY RANDOMIZED SINGLE-FACTOR EXPERIMENT 13-2.1 Example: Tensile Strength A manufacturer of paper used for making grocery bags is interested in improving the tensile strength of the product. Product engineering thinks that tensile strength is a function of the hardwood concentration in the pulp and that the range of hardwood concentrations of practical interest is between 5 and 20%. A team of engineers responsible for the study decides to investigate four levels of hardwood concentration: 5%, 10%, 15%, and 20%. They decide to make up six test specimens at each concentration level, using a pilot plant. All 24 specimens are tested on a laboratory tensile tester, in random order. The data from this experiment are shown in Table 13-1. This is an example of a completely randomized single-factor experiment with four levels of the factor. The levels of the factor are sometimes called treatments, and each treatment has six observations or replicates. The role of randomization in this experiment is extremely

JWCL232_c13_513-550.qxd

516

1/18/10

1:59 PM

Page 516

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

Table 13-1 Tensile Strength of Paper (psi) Hardwood Concentration (%) 5 10 15 20

1 7 12 14 19

Observations 3 4 15 11 13 18 19 17 22 23

2 8 17 18 25

5 9 19 16 18

6 10 15 18 20

Totals 60 94 102 127 383

Averages 10.00 15.67 17.00 21.17 15.96

important. By randomizing the order of the 24 runs, the effect of any nuisance variable that may influence the observed tensile strength is approximately balanced out. For example, suppose that there is a warm-up effect on the tensile testing machine; that is, the longer the machine is on, the greater the observed tensile strength. If all 24 runs are made in order of increasing hardwood concentration (that is, all six 5% concentration specimens are tested first, followed by all six 10% concentration specimens, etc.), any observed differences in tensile strength could also be due to the warm-up effect. The role of randomization to identify causality was discussed in Section 10-1. It is important to graphically analyze the data from a designed experiment. Figure 13-1(a) presents box plots of tensile strength at the four hardwood concentration levels. This figure indicates that changing the hardwood concentration has an effect on tensile strength; specifically, higher hardwood concentrations produce higher observed tensile strength. Furthermore, the distribution of tensile strength at a particular hardwood level is reasonably symmetric, and the variability in tensile strength does not change dramatically as the hardwood concentration changes. Graphical interpretation of the data is always useful. Box plots show the variability of the observations within a treatment (factor level) and the variability between treatments. We now discuss how the data from a single-factor randomized experiment can be analyzed statistically. 30

Tensile strength (psi)

25

20

15

10

σ2

5

0

5

10 15 20 Hardwood concentration (%) (a)

μ

+ τ1 μ1

σ2

μ

+ τ2 μ2

σ2

σ2

μ

μ

+ τ3 μ3

μ

+ τ4 μ4

(b)

Figure 13-1 (a) Box plots of hardwood concentration data. (b) Display of the model in Equation 13-1 for the completely randomized single-factor experiment.

JWCL232_c13_513-550.qxd

1/18/10

10:40 AM

Page 517

13-2 COMPLETELY RANDOMIZED SINGLE-FACTOR EXPERIMENT

517

13-2.2 Analysis of Variance Suppose we have a different levels of a single factor that we wish to compare. Sometimes, each factor level is called a treatment, a very general term that can be traced to the early applications of experimental design methodology in the agricultural sciences. The response for each of the a treatments is a random variable. The observed data would appear as shown in Table 13-2. An entry in Table 13-2, say yij, represents the jth observation taken under treatment i. We initially consider the case in which there are an equal number of observations, n, on each treatment. We may describe the observations in Table 13-2 by the linear statistical model Yij    i  ij e

i  1, 2, p , a j  1, 2, p , n

(13-1)

where Yij is a random variable denoting the (ij)th observation,  is a parameter common to all treatments called the overall mean, i is a parameter associated with the ith treatment called the ith treatment effect, and ij is a random error component. Notice that the model could have been written as Yij  i  ij e

i  1, 2, p , a j  1, 2, p , n

where i    i is the mean of the ith treatment. In this form of the model, we see that each treatment defines a population that has mean i, consisting of the overall mean  plus an effect i that is due to that particular treatment. We will assume that the errors ij are normally and independently distributed with mean zero and variance 2. Therefore, each treatment can be thought of as a normal population with mean i and variance 2. See Fig. 13-1(b). Equation 13-1 is the underlying model for a single-factor experiment. Furthermore, since we require that the observations are taken in random order and that the environment (often called the experimental units) in which the treatments are used is as uniform as possible, this experimental design is called a completely randomized design (CRD). The a factor levels in the experiment could have been chosen in two different ways. First, the experimenter could have specifically chosen the a treatments. In this situation, we wish to test hypotheses about the treatment means, and conclusions cannot be extended to similar treatments that were not considered. In addition, we may wish to estimate the treatment effects. This is called the fixed-effects model. Alternatively, the a treatments could be a random sample from a larger population of treatments. In this situation, we would like to be able to extend the conclusions (which are based on the sample of treatments) to all treatments in the population, whether or not they were explicitly considered in the experiment. Here the treatment effects i are random variables, and knowledge about the particular ones investigated is relatively Table 13-2 Typical Data for a Single-Factor Experiment Treatment

Totals

Averages

1 2

y11 y21

Observations y12 y22

p p

y1n y2n

y1. y2.

y1. y2.

o a

o ya1

o ya2

ooo p

 yan

o ya.

o ya.

y..

y..

JWCL232_c13_513-550.qxd

518

1/18/10

10:40 AM

Page 518

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

unimportant. Instead, we test hypotheses about the variability of the i and try to estimate this variability. This is called the random effects, or components of variance, model. In this section we develop the analysis of variance for the fixed-effects model. The analysis of variance is not new to us; it was used previously in the presentation of regression analysis. However, in this section we show how it can be used to test for equality of treatment effects. In the fixed-effects model, the treatment effects i are usually defined as deviations from the overall mean , so that a

a i  0

(13-2)

i1

Let yi. represent the total of the observations under the ith treatment and yi. represent the average of the observations under the ith treatment. Similarly, let y.. represent the grand total of all observations and y.. represent the grand mean of all observations. Expressed mathematically, n

yi.  yi. n

yi.  a yij j1 a

n

i1

j1

i  1, 2, . . . , a

y..  y.. N

y..  a a yij

(13-3)

where N  an is the total number of observations. Thus, the “dot” subscript notation implies summation over the subscript that it replaces. We are interested in testing the equality of the a treatment means 1, 2, . . . , a. Using Equation 13-2, we find that this is equivalent to testing the hypotheses H0: 1  2  p  a  0 H1: i 0 for at least one i

(13-4)

Thus, if the null hypothesis is true, each observation consists of the overall mean  plus a realization of the random error component ij. This is equivalent to saying that all N observations are taken from a normal distribution with mean  and variance 2. Therefore, if the null hypothesis is true, changing the levels of the factor has no effect on the mean response. The ANOVA partitions the total variability in the sample data into two component parts. Then, the test of the hypothesis in Equation 13-4 is based on a comparison of two independent estimates of the population variance. The total variability in the data is described by the total sum of squares a

n

SST  a a 1 yij  y..2 2 i1 j1

The partition of the total sum of squares is given in the following definition. ANOVA Sum of Squares Identity: Single Factor Experiment

The sum of squares identity is a

n

a

a

n

2 2 2 a a 1 yij  y..2  n a 1 yi.  y..2  a a 1 yij  yi.2

i1 j1

i1

(13-5)

i1 j1

or symbolically SST  SS Treatments  SSE

(13-6)

JWCL232_c13_513-550.qxd

1/18/10

10:40 AM

Page 519

13-2 COMPLETELY RANDOMIZED SINGLE-FACTOR EXPERIMENT

519

The identity in Equation 13-5 shows that the total variability in the data, measured by the total corrected sum of squares SST, can be partitioned into a sum of squares of differences between treatment means and the grand mean denoted SSTreatments and a sum of squares of differences of observations within a treatment from the treatment mean denoted SSE. Differences between observed treatment means and the grand mean measure the differences between treatments, while differences of observations within a treatment from the treatment mean can be due only to random error. We can gain considerable insight into how the analysis of variance works by examining the expected values of SSTreatments and SSE. This will lead us to an appropriate statistic for testing the hypothesis of no differences among treatment means (or all i  0). Expected Values of Sums of Squares: Single Factor Experiment

The expected value of the treatment sum of squares is a

E1SS Treatments 2  1a  122  n a 2i i1

and the expected value of the error sum of squares is E1SSE 2  a1n  122 There is also a partition of the number of degrees of freedom that corresponds to the sum of squares identity in Equation 13-5. That is, there are an  N observations; thus, SST has an  1 degrees of freedom. There are a levels of the factor, so SSTreatments has a  1 degrees of freedom. Finally, within any treatment there are n replicates providing n  1 degrees of freedom with which to estimate the experimental error. Since there are a treatments, we have a(n  1) degrees of freedom for error. Therefore, the degrees of freedom partition is an  1  a  1  a1n  12 The ratio MS Treatments  SS Treatments 1a  12 is called the mean square for treatments. Now if the null hypothesis H0: 1  2  p  a  0 is true, MSTreatments is an unbiased estimator of 2 because g ai1 i  0. However, if H1 is true, MSTreatments estimates 2 plus a positive term that incorporates variation due to the systematic difference in treatment means. Note that the error mean square MSE  SSE 3a1n  12 4 is an unbiased estimator of 2 regardless of whether or not H0 is true. We can also show that MSTreatments and MSE are independent. Consequently, we can show that if the null hypothesis H0 is true, the ratio ANOVA F-Test

F0 

SS Treatments 1a  12 MS Treatments  MSE SSE 3a1n  12 4

(13-7)

JWCL232_c13_513-550.qxd

520

1/18/10

10:40 AM

Page 520

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

has an F-distribution with a  1 and a (n  1) degrees of freedom. Furthermore, from the expected mean squares, we know that MSE is an unbiased estimator of 2. Also, under the null hypothesis, MSTreatments is an unbiased estimator of 2. However, if the null hypothesis is false, the expected value of MSTreatments is greater than 2. Therefore, under the alternative hypothesis, the expected value of the numerator of the test statistic (Equation 13-7) is greater than the expected value of the denominator. Consequently, we should reject H0 if the statistic is large. This implies an upper-tail, one-tail critical region. Therefore, we would reject H 0 if f0  f , a1, a1n12 where f0 is the computed value of F 0 from Equation 13-7. Efficient computational formulas for the sums of squares may be obtained by expanding and simplifying the definitions of SSTreatments and SST. This yields the following results. Computing Formulas for ANOVA: Single Factor with Equal Sample Sizes

The sums of squares computing formulas for the ANOVA with equal sample sizes in each treatment are a n y..2 SS T  a a yij2  N i1 j1

(13-8)

a y 2. y..2 i SS Treatments  a n  N i1

(13-9)

and

The error sum of squares is obtained by subtraction as SSE  SST  SS Treatments

(13-10)

The computations for this test procedure are usually summarized in tabular form as shown in Table 13-3. This is called an analysis of variance (or ANOVA) table. EXAMPLE 13-1

Tensile Strength ANOVA

Consider the paper tensile strength experiment described in Section 13-2.1. This experiment is a CRD. We can use the analysis of variance to test the hypothesis that different hardwood concentrations do not affect the mean tensile strength of the paper. The hypotheses are H0: 1  2  3  4  0

We will use  0.01. The sums of squares for the analysis of variance are computed from Equations 13-8, 13-9, and 13-10 as follows: 4 6 y..2 SST  a a y2ij  N i1 j1

 172 2  182 2  p  1202 2 

H1: i 0 for at least one i

13832 2 24

 512.96

Table 13-3 The Analysis of Variance for a Single-Factor Experiment, Fixed-Effects Model Source of Variation

Sum of Squares

Degrees of Freedom

Mean Square

F0

Treatments

SSTreatments

a1

MSTreatments

MS Treatments MSE

Error Total

SSE SST

a(n  1) an  1

MSE

JWCL232_c13_513-550.qxd

1/18/10

10:40 AM

Page 521

521

13-2 COMPLETELY RANDOMIZED SINGLE-FACTOR EXPERIMENT

Table 13-4 ANOVA for the Tensile Strength Data Source of Variation

Sum of Squares

Degrees of Freedom

382.79 130.17 512.96

3 20 23

Hardwood concentration Error Total 4 y2. y2.. i SS Treatments  a n  N i1



Mean Square

f0

P-value

127.60 6.51

19.60

3.59 E-6

the paper. We can also find a P-value for this test statistic as follows:

1602 2  1942 2  11022 2  11272 2 6



13832 2 24

 382.79 SSE  SST  SS Treatments  512.96  382.79  130.17 The ANOVA is summarized in Table 13-4. Since f0.01,3,20  4.94, we reject H0 and conclude that hardwood concentration in the pulp significantly affects the mean strength of

P  P1F3,20  19.602  3.59  106 Since P  3.59  106 is considerably smaller than  0.01, we have strong evidence to conclude that H0 is not true. Practical Interpretation: There is strong evidence to conclude that hardwood concentration has an effect on tensile strength. However, the ANOVA does not tell as which levels of hardwood concentration result in different tensile strength means. We will see how to answer this question below.

Minitab Output Many software packages have the capability to analyze data from designed experiments using the analysis of variance. Table 13-5 presents the output from the Minitab one-way analysis of variance routine for the paper tensile strength experiment in Example 13-1. The results agree closely with the manual calculations reported previously in Table 13-4. The Minitab output also presents 95% confidence intervals on each individual treatment mean. The mean of the ith treatment is defined as i    i

i  1, 2, p , a

ˆ i  Yi. . Now, if we assume that the errors are normally distributed, A point estimator of i is  each treatment average is normally distributed with mean i and variance 2 n. Thus, if 2 were known, we could use the normal distribution to construct a CI. Using MSE as an estimator of 2 (the square root of MSE is the “Pooled StDev” referred to in the Minitab output), we would base the CI on the t distribution, since

T

Yi.  i 1MSE n

has a t distribution with a(n  1) degrees of freedom. This leads to the following definition of the confidence interval. Confidence Interval on a Treatment Mean

A 100(1  ) percent confidence interval on the mean of the ith treatment i is MSE MSE

i yi.  t 2,a1n12 yi.  t 2,a1n12 n B B n

(13-11)

JWCL232_c13_513-550.qxd

522

1/18/10

10:40 AM

Page 522

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

Table 13-5 Minitab Analysis of Variance Output for Example 13-1 One-Way ANOVA: Strength versus CONC Analysis of Variance for Strength Source Conc Error Total Level 5 10 15 20

DF 3 20 23 N 6 6 6 6

Pooled StDev 

SS 382.79 130.17 512.96 Mean 10.000 15.667 17.000 21.167

MS 127.60 6.51

F 19.61

P 0.000

Individual 95% CIs For Mean Based on Pooled StDev —-  ——-  ——-  ——-  (—*—) (—*—) (—*—) (—*—) —-  ———-  ———-  ———-  10.0 15.0 20.0 25.0

StDev 2.828 2.805 1.789 2.639

2.551

Fisher’s pairwise comparisons Family error rate  0.192 Individual error rate  0.0500 Critical value  2.086 Intervals for (column level mean)  (row level mean) 5 10 15 10

8.739 2.594

15

10.072 3.928

4.406 1.739

20

14.239 8.094

8.572 2.428

7.239 1.094

Equation 13-11 is used to calculate the 95% CIs shown graphically in the Minitab output of Table 13-5. For example, at 20% hardwood the point estimate of the mean is y4.  21.167, MSE  6.51, and t0.025,20  2.086, so the 95% CI is 3y4.  t0.025,20 1MSE n4 ˛

321.167  12.0862 16.51 64 ˛

or 19.00 psi 4 23.34 psi It can also be interesting to find confidence intervals on the difference in two treatment means, say, i  j. The point estimator of i  j is Yi.  Yj., and the variance of this estimator is 2 2 2 2 V1Yi.  Yj.2  n  n  n

JWCL232_c13_513-550.qxd

1/18/10

10:40 AM

Page 523

13-2 COMPLETELY RANDOMIZED SINGLE-FACTOR EXPERIMENT

523

Now if we use MSE to estimate 2, T

Yi.  Yj.  1i  j 2 12MSE n

has a t distribution with a(n  1) degrees of freedom. Therefore, a CI on i  j may be based on the t distribution. Confidence Interval on a Difference in Treatment Means

A 100(1  ) percent confidence interval on the difference in two treatment means i  j is yi.  yj.  t 2,a1n12 B

2MSE 2MSE n i  j yi.  yj.  t 2,a1n12 B n (13-12)

A 95% CI on the difference in means 3  2 is computed from Equation 13-12 as follows: 3y3.  y2.  t0.025,20 12MSE n 4

317.00  15.67  12.0862 1216.512 64 ˛

or 1.74 3  2 4.40 Since the CI includes zero, we would conclude that there is no difference in mean tensile strength at these two particular hardwood levels. The bottom portion of the computer output in Table 13-5 provides additional information concerning which specific means are different. We will discuss this in more detail in Section 13-2.3. An Unbalanced Experiment In some single-factor experiments, the number of observations taken under each treatment may be different. We then say that the design is unbalanced. In this situation, slight modifications must be made in the sums of squares formulas. Let ni observations be taken a under treatment i (i  1, 2, . . . , a), and let the total number of observations N  g i1 ni. The computational formulas for SST and SSTreatments are as shown in the following definition. Computing Formulas for ANOVA: Single Factor with Unequal Sample Sizes

The sums of squares computing formulas for the ANOVA with unequal sample sizes ni in each treatment are a ni y2.. SST  a a yij2  N i1 j1 a y 2. y2.. i SS Treatments  a n  i N i1

(13-13) (13-14)

and SSE  SST  SSTreatments

(13-15)

JWCL232_c13_513-550.qxd

524

1/18/10

10:40 AM

Page 524

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

Choosing a balanced design has two important advantages. First, the ANOVA is relatively insensitive to small departures from the assumption of equality of variances if the sample sizes are equal. This is not the case for unequal sample sizes. Second, the power of the test is maximized if the samples are of equal size.

13-2.3 Multiple Comparisons Following the ANOVA When the null hypothesis H0: 1  2  p  a  0 is rejected in the ANOVA, we know that some of the treatment or factor level means are different. However, the ANOVA doesn’t identify which means are different. Methods for investigating this issue are called multiple comparisons methods. Many of these procedures are available. Here we describe a very simple one, Fisher’s least significant difference (LSD) method and a graphical method. Montgomery (2009) presents these and other methods and provides a comparative discussion. The Fisher LSD method compares all pairs of means with the null hypotheses H0: i  j (for all i j) using the t-statistic t0 

yi.  yj. B

2MSE n

Assuming a two-sided alternative hypothesis, the pair of means i and j would be declared significantly different if 0 yi.  yj. 0  LSD where LSD, the least significant difference, is Least Significant Difference for Multiple Comparisons

LSD  t 2,a 1n12 B

2MSE n

(13-16)

If the sample sizes are different in each treatment, the LSD is defined as 1 1 LSD  t 2,Na MS E a n  n b i j B EXAMPLE 13-2 We will apply the Fisher LSD method to the hardwood concentration experiment. There are a  4 means, n  6, MSE  6.51, and t0.025,20  2.086. The treatment means are y1. y2. y3. y4.

   

10.00 psi 15.67 psi 17.00 psi 21.17 psi

The value of LSD is LSD  t 0.025,20 12MSE n  2.086 1216.512 6  3.07. Therefore, any pair of treatment aver-

ages that differs by more than 3.07 implies that the corresponding pair of treatment means are different. The comparisons among the observed treatment averages are as follows: 4 vs. 1  21.17  10.00  11.17  3.07 4 vs. 2  21.17  15.67  5.50  3.07 4 vs. 3  21.17  17.00  4.17  3.07 3 vs. 1  17.00  10.00  7.00  3.07 3 vs. 2  17.00  15.67  1.33  3.07 2 vs. 1  15.67  10.00  5.67  3.07

JWCL232_c13_513-550.qxd

1/18/10

10:40 AM

Page 525

13-2 COMPLETELY RANDOMIZED SINGLE-FACTOR EXPERIMENT 5%

0

5

10

10% 15%

15

525

20%

20

25 psi

Figure 13-2 Results of Fisher’s LSD method in Example 13-2. Conclusions: From this analysis, we see that there are significant differences between all pairs of means except 2 and 3. This implies that 10% and 15% hardwood concentration produce approximately the same tensile strength and that all other concentration levels tested produce different tensile strengths.

It is often helpful to draw a graph of the treatment means, such as in Fig. 13-2, with the means that are not different underlined. This graph clearly reveals the results of the experiment and shows that 20% hardwood produces the maximum tensile strength.

The Minitab output in Table 13-5 shows the Fisher LSD method under the heading “Fisher’s pairwise comparisons.” The critical value reported is actually the value of t0.025,20  2.086. Minitab implements Fisher’s LSD method by computing confidence intervals on all pairs of treatment means using Equation 13-12. The lower and upper 95% confidence limits are shown at the bottom of the table. Notice that the only pair of means for which the confidence interval includes zero is for 10 and 15. This implies that 10 and 15 are not significantly different, the same result found in Example 13-2. Table 13-5 also provides a “family error rate,” equal to 0.192 in this example. When all possible pairs of means are tested, the probability of at least one type I error can be much greater than for a single test. We can interpret the family error rate as follows. The probability is 1  0.192  0.808 that there are no type I errors in the six comparisons. The family error rate in Table 13-5 is based on the distribution of the range of the sample means. See Montgomery (2009) for details. Alternatively, Minitab permits you to specify a family error rate and will then calculate an individual error rate for each comparison. Graphical Comparison of Means It is easy to compare treatment means graphically, following the analysis of variance. Suppose that the factor has a levels and that y1., y2., …, ya. are the observed averages for these factor levels. Each treatment average has standard deviation / 2n, where  is the standard deviation of an individual observation. If all treatment means are equal, the observed means yi. would behave as if they were a set of observations drawn at random from a normal distribution with mean  and standard deviation / 2n. Visualize this normal distribution capable of being slid along an axis below which the treatment means y1., y2., …, ya. are plotted. If all treatment means are equal, there should be some position for this distribution that makes it obvious that the yi. values were drawn from the same distribution. If this is not the case, the yi. values that do not appear to have been drawn from this distribution are associated with treatments that produce different mean responses. The only flaw in this logic is that  is unknown. However, we can use 2MSE from the analysis of variance to estimate . This implies that a t distribution should be used instead of the normal in making the plot, but since the t looks so much like the normal, sketching a normal curve that is approximately 62MSE/n units wide will usually work very well. Figure 13-3 shows this arrangement for the hardwood concentration experiment in Example 13-1. The standard deviation of this normal distribution is 2MSE/n  26.51/6  1.04

If we visualize sliding this distribution along the horizontal axis, we note that there is no location for the distribution that would suggest that all four observations (the plotted means) are typical, randomly selected values from that distribution. This, of course, should be expected, because the

JWCL232_c13_513-550.qxd

526

1/18/10

10:40 AM

Page 526

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE ∧ σ / √n = 1.04

1 0

5

2

10

3

15

4 20

25

30

Figure 13-3 Tensile strength averages from the hardwood concentration experiment in relation to a normal distribution with standard deviation 1MSE n  16.51 6  1.04.

analysis of variance has indicated that the means differ, and the display in Fig. 13-3 is just a graphical representation of the analysis of variance results. The figure does indicate that treatment 4 (20% hardwood) produces paper with higher mean tensile strength than do the other treatments, and treatment 1 (5% hardwood) results in lower mean tensile strength than do the other treatments. The means of treatments 2 and 3 (10% and 15% hardwood, respectively) do not differ. This simple procedure is a rough but very effective multiple comparison technique. It works well in many situations.

13-2.4 Residual Analysis and Model Checking The analysis of variance assumes that the observations are normally and independently distributed with the same variance for each treatment or factor level. These assumptions should be checked by examining the residuals. A residual is the difference between an observation yij and its estimated (or fitted) value from the statistical model being studied, denoted as yˆ ij. For the completely randomized design yˆij  yi. and each residual is eij  yij  yi., that is, the difference between an observation and the corresponding observed treatment mean. The residuals for the paper tensile strength experiment are shown in Table 13-6. Using yi. to calculate each residual essentially removes the effect of hardwood concentration from the data; consequently, the residuals contain information about unexplained variability. The normality assumption can be checked by constructing a normal probability plot of the residuals. To check the assumption of equal variances at each factor level, plot the residuals against the factor levels and compare the spread in the residuals. It is also useful to plot the residuals against yi. (sometimes called the fitted value); the variability in the residuals should not depend in any way on the value of yi. Most statistical software packages will construct these plots on request. When a pattern appears in these plots, it usually suggests the need for a transformation, that is, analyzing the data in a different metric. For example, if the variability in the residuals increases with yi., a transformation such as log y or 1y should be considered. In some problems, the dependency of residual scatter on the observed mean yi. is very important information. It may be desirable to select the factor level that results in maximum response; however, this level may also cause more variation in response from run to run. Table 13-6 Residuals for the Tensile Strength Experiment Hardwood Concentration (%) 5 10 15 20

Residuals 3.00 3.67 3.00 2.17

2.00 1.33 1.00 3.83

5.00 2.67 2.00 0.83

1.00 2.33 0.00 1.83

1.00 3.33 1.00 3.17

0.00 0.67 1.00 1.17

JWCL232_c13_513-550.qxd

1/18/10

10:40 AM

Page 527

527

13-2 COMPLETELY RANDOMIZED SINGLE-FACTOR EXPERIMENT

The independence assumption can be checked by plotting the residuals against the time or run order in which the experiment was performed. A pattern in this plot, such as sequences of positive and negative residuals, may indicate that the observations are not independent. This suggests that time or run order is important or that variables that change over time are important and have not been included in the experimental design. A normal probability plot of the residuals from the paper tensile strength experiment is shown in Fig. 13-4. Figures 13-5 and 13-6 present the residuals plotted against the factor levels and the fitted value yi. respectively. These plots do not reveal any model inadequacy or unusual problem with the assumptions.

13-2.5 Determining Sample Size In any experimental design problem, the choice of the sample size or number of replicates to use is important. Operating characteristic curves can be used to provide guidance in making this selection. Recall that an operating characteristic curve is a plot of the probability of a type II

4

Residual value

2

2

10%

5%

15%

20%

–2

1 Normal score zj

0

0

–4

Figure 13-5 Plot of residuals versus factor levels (hardwood concentration). –1 4 –2

–2

0

2

4

Residual value

Figure 13-4 Normal probability plot of residuals from the hardwood concentration experiment.

2

6 Residual value

–4

0 10.0

15.0

20.0

–2

–4

Figure 13-6

Plot of residuals versus yi.

25.0

y–i •

JWCL232_c13_513-550.qxd

528

1/18/10

10:40 AM

Page 528

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

error () for various sample sizes against values of the parameters under test. The operating characteristic curves can be used to determine how many replicates are required to achieve adequate sensitivity. The power of the ANOVA test is 1    P5Reject H0 0 H0 is false6  P5F0  f ,a1, a 1n12 0 H0 is false6

(13-17)

To evaluate this probability statement, we need to know the distribution of the test statistic F0 if the null hypothesis is false. Because ANOVA compares several means, the null hypothesis can be false in different ways. For example, possibly 1  0, 2  0, 3  0, and so forth. It can be shown that the power for ANOVA in Equation 13-17 depends on the i’s only through the function a

2 

n a i12i a2

Therefore, alternative hypotheses for the i’s can be used to calculate 2 and this in turn can be used to calculate the power. Specifically, it can be shown that if H0 is false, the statistic F0  MSTreatments/MSE has a noncentral F distribution, with a  1 and n1a  12 degrees of freedom and a noncentrality parameter that depends on 2. Instead of tables for the noncentral F distribution, operating characteristic curves are used to evaluate  defined in Equation 13-17. These curves plot  against . Curves are available for  0.05 and  0.01 and for several values of the number of degrees of freedom for numerator (denoted v1) and denominator (denoted v2). Figure 13-7 gives representative O.C. curves, one for a  4 (v1  3) and one for a  5 (v1  4) treatments. Notice that for each value of a there are curves for  0.05 and  0.01. In using the operating curves, we must define the difference in means that we wish to detect a in terms of g i1 2i . Also, the error variance 2 is usually unknown. In such cases, we a must choose ratios of g i1 2i 2 that we wish to detect. Alternatively, if an estimate of 2 is available, one may replace 2 with this estimate. For example, if we were interested in the sensitivity of an experiment that has already been performed, we might use MSE as the estimate of 2. EXAMPLE 13-3 Suppose that five means are being compared in a completely randomized experiment with  0.01. The experimenter would like to know how many replicates to run if it is important to reject H0 with probability at least 0.90 if 5 g i1 2i 2  5.0 . The parameter  2 is, in this case,

Figure 13-7, we find that   0.38. Therefore, the power of the test is approximately 1    1  0.38  0.62, which is less than the required 0.90, and so we conclude that n  4 replicates is not sufficient. Proceeding in a similar manner, we can construct the following table:

a

n a 2i   2

i1 2

a

n  152  n 5

and for the operating characteristic curve with v1  a  1  5  1  4, and v2  a (n  1)  5(n  1) error degrees of freedom refer to the lower curve in Figure 13-7. As a first guess, try n  4 replicates. This yields 2  4,   2, and v2  5(3)  15 error degrees of freedom. Consequently, from

n

2



a(n  1)



Power  (1  )

4 5 6

4 5 6

2.00 2.24 2.45

15 20 25

0.38 0.18 0.06

0.62 0.82 0.94

Conclusions: At least n  6 replicates must be run in order to obtain a test with the required power.

1/18/10

10:40 AM

Page 529

13-2 COMPLETELY RANDOMIZED SINGLE-FACTOR EXPERIMENT 1.00

ν1 = 3

α

=0

.01

α = 05 0.

0.80 0.70 0.60 0.50

Probability of accepting the hypothesis

0.40 ∞

0.30

60 30

0.20

20 15 12 0.10

10 9

0.08 0.07 0.06

8 7

0.05

ν2 = 6

0.04 0.03

0.02

0.01

1.00

1 φ (for α = 0.01)

2 1

3 2

φ (for α = 0.05) 3

4

5

φ (for α = 0.05) 3

4

5

ν1 = 4

α

=0

.01

α = 05

0.

0.80 0.70 0.60 0.50 0.40 Probability of accepting the hypothesis

JWCL232_c13_513-550.qxd

∞

0.30

60 30

0.20

20 15 12 0.10

10 9

0.08 0.07 0.06

8 7

0.05

ν2 = 6

0.04 0.03

0.02

0.01

1 φ (for α = 0.01)

2 1

3 2

Figure 13-7 Two Operating Characteristic curves for the fixed-effects model analysis of variance. Top curves for four treatments and bottom curves for five treatments.

529

JWCL232_c13_513-550.qxd

530

1/18/10

2:00 PM

Page 530

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

EXERCISES FOR SECTION 13-2 13-1. Consider the computer output below. One-way ANOVA: y versus Factor Source DF SS MS Factor ? 117.4 39.1 Error 16 396.8 ? Total 19 514.2

F ?

13-2. An article in Nature describes an experiment to investigate the effect on consuming chocolate on cardiovascular health (“Plasma Antioxidants from Chocolate,” Vol. 424, 2003, pp. 1013). The experiment consisted of using three different types of chocolates: 100 g of dark chocolate, 100 g of dark chocolate with 200 ml of full-fat milk, and 200 g of milk chocolate. Twelve subjects were used, seven women and five men, with an average age range of 32.2 ⫾ 1 years, an average weight of 65.8 ⫾ 3.1 kg, and body-mass index of 21.9 ⫾ 0.4 kg m⫺2. On different days, a subject consumed one of the chocolate-factor levels, and one hour later the total antioxidant capacity of their blood plasma was measured in an assay. Data similar to those summarized in the article are shown below. Factor DC DC⫹MK MC

1

2

3

4

5

15 20 25 30 35

7 12 14 19 7

7 17 18 25 10

15 12 18 22 11

11 18 19 19 15

9 18 19 23 11

P ?

(a) How many levels of the factor were used in this experiment? (b) How many replicates did the experimenter use? (c) Fill in the missing information in the ANOVA table. Use bounds for the P-value. (d) What conclusions can you draw about differences in the factor level means?

Observations

Cotton Percentage

(a) Does cotton percentage affect breaking strength? Draw comparative box plots and perform an analysis of variance. Use ␣ ⫽ 0.05. (b) Plot average tensile strength against cotton percentage and interpret the results. (c) Analyze the residuals and comment on model adequacy. 13-4. In “Orthogonal Design for Process Optimization and Its Application to Plasma Etching” (Solid State Technology, May 1987), G. Z. Yin and D. W. Jillie describe an experiment to determine the effect of C2F6 flow rate on the uniformity of the etch on a silicon wafer used in integrated circuit manufacturing. Three flow rates are used in the experiment, and the resulting uniformity (in percent) for six replicates is shown below. Observations

C2F6 Flow (SCCM)

1

2

3

4

5

6

125 160 200

2.7 4.9 4.6

4.6 4.6 3.4

2.6 5.0 2.9

3.0 4.2 3.5

3.2 3.6 4.1

3.8 4.2 5.1

Subjects (Observations) 1 118.8 105.4 102.1

2 122.6 101.1 105.8

3 115.6 102.7 99.6

4 113.6 97.1 102.7

5 119.5 101.9 98.8

(a) Construct comparative box plots and study the data. What visual impression do you have from examining these plots? (b) Analyze the experimental data using an ANOVA. If ␣ ⫽ 0.05, what conclusions would you draw? What would you conclude if ␣ ⫽ 0.01? (c) Is there evidence that the dark chocolate increases the mean antioxidant capacity of the subjects’ blood plasma? (d) Analyze the residuals from this experiment. 13-3. In Design and Analysis of Experiments, 7th edition (John Wiley & Sons, 2009) D. C. Montgomery described an experiment in which the tensile strength of a synthetic fiber was of interest to the manufacturer. It is suspected that strength is related to the percentage of cotton in the fiber. Five levels of cotton percentage were used, and five replicates were run in random order, resulting in the data below.

6 115.9 98.9 100.9

7 115.8 100.0 102.8

8 115.1 99.8 98.7

9 116.9 102.6 94.7

10 115.4 100.9 97.8

11 115.6 104.5 99.7

12 107.9 93.5 98.6

(a) Does C2F6 flow rate affect etch uniformity? Construct box plots to compare the factor levels and perform the analysis of variance. Use ␣ ⫽ 0.05. (b) Do the residuals indicate any problems with the underlying assumptions? 13-5. The compressive strength of concrete is being studied, and four different mixing techniques are being investigated. The following data have been collected. Mixing Technique 1 2 3 4

Compressive Strength (psi) 3129 3200 2800 2600

3000 3300 2900 2700

2865 2975 2985 2600

2890 3150 3050 2765

JWCL232_c13_513-550.qxd

1/18/10

10:40 AM

Page 531

13-2 COMPLETELY RANDOMIZED SINGLE-FACTOR EXPERIMENT

(a) Test the hypothesis that mixing techniques affect the strength of the concrete. Use  0.05. (b) Find the P-value for the F-statistic computed in part (a). (c) Analyze the residuals from this experiment. 13-6. The response time in milliseconds was determined for three different types of circuits in an electronic calculator. The results are recorded here. Circuit Type 1 2 3

Response 19 20 16

22 21 15

20 33 18

18 27 26

25 40 17

(a) Using  0.01, test the hypothesis that the three circuit types have the same response time. (b) Analyze the residuals from this experiment. (c) Find a 95% confidence interval on the response time for circuit three. 13-7. An electronics engineer is interested in the effect on tube conductivity of five different types of coating for cathode ray tubes in a telecommunications system display device. The following conductivity data are obtained. Coating Type 1 2 3 4 5

Conductivity 143 152 134 129 147

141 149 133 127 148

150 137 132 132 144

146 143 127 129 142

(a) Is there any difference in conductivity due to coating type? Use  0.01. (b) Analyze the residuals from this experiment. (c) Construct a 95% interval estimate of the coating type 1 mean. Construct a 99% interval estimate of the mean difference between coating types 1 and 4. 13-8. An article in Environment International (Vol. 18, No. 4, 1992) described an experiment in which the amount of radon released in showers was investigated. Radon-enriched water was used in the experiment, and six different orifice diameters were tested in shower heads. The data from the experiment are shown in the following table. Orifice Diameter 0.37 0.51 0.71 1.02 1.40 1.99

531

(a) Does the size of the orifice affect the mean percentage of radon released? Use  0.05. (b) Find the P-value for the F-statistic in part (a). (c) Analyze the residuals from this experiment. (d) Find a 95% confidence interval on the mean percent of radon released when the orifice diameter is 1.40. 13-9. An article in the ACI Materials Journal (Vol. 84, 1987, pp. 213–216) described several experiments investigating the rodding of concrete to remove entrapped air. A 3-inch  6-inch cylinder was used, and the number of times this rod was used is the design variable. The resulting compressive strength of the concrete specimen is the response. The data are shown in the following table. Rodding Level

Compressive Strength

10 15 20 25

1530 1610 1560 1500

1530 1650 1730 1490

1440 1500 1530 1510

(a) Is there any difference in compressive strength due to the rodding level? (b) Find the P-value for the F-statistic in part (a). (c) Analyze the residuals from this experiment. What conclusions can you draw about the underlying model assumptions? 13-10. An article in the Materials Research Bulletin (Vol. 26, No. 11, 1991) investigated four different methods of preparing the superconducting compound PbMo6S8. The authors contend that the presence of oxygen during the preparation process affects the material’s superconducting transition temperature Tc. Preparation methods 1 and 2 use techniques that are designed to eliminate the presence of oxygen, while methods 3 and 4 allow oxygen to be present. Five observations on Tc (in °K) were made for each method, and the results are as follows: Preparation Method 1 2 3 4

Transition Temperature Tc(K) 14.8 14.6 12.7 14.2

14.8 15.0 11.6 14.4

14.7 14.9 12.4 14.4

14.8 14.8 12.7 12.2

14.9 14.7 12.1 11.7

Radon Released (%) 80 75 74 67 62 60

83 75 73 72 62 61

83 79 76 74 67 64

85 79 77 74 69 66

(a) Is there evidence to support the claim that the presence of oxygen during preparation affects the mean transition temperature? Use  0.05. (b) What is the P-value for the F-test in part (a)? (c) Analyze the residuals from this experiment. (d) Find a 95% confidence interval on mean Tc when method 1 is used to prepare the material.

JWCL232_c13_513-550.qxd

532

1/18/10

10:40 AM

Page 532

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

13-11. A paper in the Journal of the Association of Asphalt Paving Technologists (Vol. 59, 1990) describes an experiment to determine the effect of air voids on percentage retained strength of asphalt. For purposes of the experiment, air voids are controlled at three levels; low (2–4%), medium (4–6%), and high (6–8%). The data are shown in the following table. Air Voids Low Medium High

Retained Strength (%) 106 80 78

90 69 80

103 94 62

90 91 69

79 70 76

88 83 85

92 87 69

95 83 85

(a) Do the different levels of air voids significantly affect mean retained strength? Use  0.01.

(a) Is there a difference in the cross-linker level? Draw comparative boxplots and perform an analysis of variance. Use  0.05. (b) Find the P-value of the test. Estimate the variability due to random error. (c) Plot average domain spacing against cross-linker level and interpret the results. (d) Analyze the residuals from this experiment and comment on model adequacy. 13-13. In the book Analysis of Longitudinal Data, 2nd ed., (2002, Oxford University Press), by Diggle, Heagerty, Liang, and Zeger, the authors analyzed the effects of three diets on the protein content of cow’s milk. The data shown here were collected after one week and include 25 cows on the barley diet, and 27 cows each on the other two diets:

Diet

Protein Content of Cow’s Milk

Barley Barleylupins Lupins Diet (continued) Barley Barleylupins Lupins

3.63 3.38 3.69

3.24 3.8 4.2

3.98 4.17 3.31

3.66 4.59 3.13

4.34 4.07 3.73

4.36 4.32 4.32

4.17 3.56 3.04

4.4 3.67 3.84

3.4 4.15 3.98

3.75 3.51 4.18

4.2 4.2 4.2

4.02 4.12 4.1

4.02 3.52 3.25

3.81 4.02 3.5

3.62 3.18 4.13

3.66 4.11 3.21

4.44 3.27 3.9

4.23 3.27 3.5

3.82 3.97 4.1

3.53 3.31 2.69

4.47 4.12 4.3

3.93 3.92 4.06

3.27 3.78 3.88

3.3 4 4

4.37 3.67

3.79 4.27

(b) Find the P-value for the F-statistic in part (a). (c) Analyze the residuals from this experiment. (d) Find a 95% confidence interval on mean retained strength where there is a high level of air voids. (e) Find a 95% confidence interval on the difference in mean retained strength at the low and high levels of air voids. 13-12. An article in Quality Engineering [“Estimating Sources of Variation: A Case Study from Polyurethane Product Research” (1999–2000, Vol. 12, pp. 89–96)] studied the effects of additives on final polymer properties. In this case, polyurethane additives were referred to as cross-linkers. The average domain spacing was the measurement of the polymer property. The data are as follows:

Cross-Linker Level 1 0.75 0.5 0 0.5 1

Domain Spacing (nm) 8.2 8.3 8.9 8.5 8.8 8.6

8 8.4 8.7 8.7 9.1 8.5

8.2 8.3 8.9 8.7 9.0 8.6

7.9 8.2 8.4 8.7 8.7 8.7

8.1 8.3 8.3 8.8 8.9 8.8

8 8.1 8.5 8.8 8.5 8.8

3.9 4.08 3.34

(a) Does diet affect the protein content of cow’s milk? Draw comparative boxplots and perform an analysis of variance. Use  0.05. (b) Find the P-value of the test. Estimate the variability due to random error. (c) Plot average protein content against diets and interpret the results. (d) Analyze the residuals and comment on model adequacy. 13-14. An article in Journal of Food Science (2001, Vol. 66, No. 3, pp. 472–477) studied potato spoilage based on different conditions of acidified oxine (AO), which is a mixture of chlorite and chlorine dioxide. The data are shown below: AO Solution (ppm) 50 100 200 400

% Spoilage 100 60 60 25

50 30 50 30

60 30 29 15

(a) Do the AO solutions differ in the spoilage percentage? Use  0.05. (b) Find the P-value of the test. Estimate the variability due to random error.

JWCL232_c13_513-550.qxd

1/18/10

10:40 AM

Page 533

13-3 THE RANDOM-EFFECTS MODEL

(c) Plot average spoilage against AO solution and interpret the results. Which AO solution would you recommend for use in practice? (d) Analyze the residuals from this experiment. 13-15. An experiment was run to determine whether four specific firing temperatures affect the density of a certain type of brick. The experiment led to the following data. Temperature (°F) 100 125 150 175

21.8 21.7 21.9 21.9

Density 21.9 21.4 21.8 21.7

21.7 21.5 21.8 21.8

21.6 21.5 21.6 21.7

21.7 — 21.5 21.6

21.5 — — 21.8

21.8 — — —

(a) Does the firing temperature affect the density of the bricks? Use  0.05. (b) Find the P-value for the F-statistic computed in part (a). (c) Analyze the residuals from the experiment. 13-16. (a) Use Fisher’s LSD method with  0.05 to analyze the means of the three types of chocolate in Exercise 13-4. (b) Use the graphical method to compare means described in this section and compare your conclusions to those from Fisher’s LSD method. 13-17. (a) Use Fisher’s LSD method with  0.05 to analyze the means of the five different levels of cotton content in Exercise 13-3. (b) Use the graphical method to compare means described in this section and compare your conclusions to those from Fisher’s LSD method. 13-18. (a) Use Fisher’s LSD method with  0.01 to analyze the mean response times for the three circuits described in Exercise 13-6. (b) Use the graphical method to compare means described in this section and compare your conclusions to those from Fisher’s LSD method. 13-19. (a) Use Fisher’s LSD method with  0.05 to analyze the mean compressive strength of the four mixing techniques in Exercise 13-5. (b) Use the graphical method to compare means described in this section and compare your conclusions to those from Fisher’s LSD method.

533

13-20. (a) Use Fisher’s LSD method with  0.05 to analyze the mean amounts of radon released in the experiment described in Exercise 13-8. (b) Use the graphical method to compare means described in this section and compare your conclusions to those from Fisher’s LSD method. 13-21. (a) Use Fisher’s LSD method with  0.01 to analyze the five means for the coating types described in Exercise 13-7. (b) Use the graphical method to compare means described in this section and compare your conclusions to those from Fisher’s LSD method. 13-22. (a) Apply Fisher’s LSD method with  0.05 to the superconducting material experiment described in Exercise 13-10. Which preparation methods differ? (b) Use the graphical method to compare means described in this section and compare your conclusions to those from Fisher’s LSD method. 13-23. (a) Apply Fisher’s LSD method to the air void experiment described in Exercise 13-11. Using  0.05, which treatment means are different? (b) Use the graphical method to compare means described in this section and compare your conclusions to those from Fisher’s LSD method. 13-24. (a) Apply Fisher’s LSD method to the domain spacing data in Exercise 13-12. Which cross-linker levels differ? Use  0.05. (b) Use the graphical method to compare means described in this section and compare your conclusions to those from Fisher’s LSD method. 13-25. (a) Apply Fisher’s LSD method to the data on protein content of milk in Exercise 13-13. Which diets differ? Use  0.01. (b) Use the graphical method to compare means described in this section and compare your conclusions to those from Fisher’s LSD method. 13-26. Suppose that four normal populations have common variance 2  25 and means 1  50, 2  60, 3  50, and 4  60. How many observations should be taken on each population so that the probability of rejecting the hypothesis of equality of means is at least 0.90? Use  0.05. 13-27. Suppose that five normal populations have common variance 2  100 and means 1  175, 2  190, 3  160, 4  200, and 5  215. How many observations per population must be taken so that the probability of rejecting the hypothesis of equality of means is at least 0.95? Use  0.01.

13-3 THE RANDOM-EFFECTS MODEL 13-3.1 Fixed Versus Random Factors In many situations, the factor of interest has a large number of possible levels. The analyst is interested in drawing conclusions about the entire population of factor levels. If the experimenter randomly selects a of these levels from the population of factor levels, we say that the factor is a

JWCL232_c13_513-550.qxd

534

1/18/10

10:40 AM

Page 534

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

random factor. Because the levels of the factor actually used in the experiment were chosen randomly, the conclusions reached will be valid for the entire population of factor levels. We will assume that the population of factor levels is either of infinite size or is large enough to be considered infinite. Notice that this is a very different situation than we encountered in the fixedeffects case, where the conclusions apply only for the factor levels used in the experiment.

13-3.2 ANOVA and Variance Components The linear statistical model is Yij    i  ij e

i  1, 2, p , a j  1, 2, p , n

(13-18)

where the treatment effects i and the errors ij are independent random variables. Note that the model is identical in structure to the fixed-effects case, but the parameters have a different interpretation. If the variance of the treatment effects i is 2 , by independence the variance of the response is V1Yij 2  2  2 (13-19) The variances 2 and 2 are called variance components, and the model, Equation 13-19, is called the components of variance model or the random-effects model. To test hypotheses in this model, we assume that the errors ij are normally and independently distributed with mean 0 and variance 2 and that the treatment effects i are normally and independently distributed with mean zero and variance 2 .* For the random-effects model, testing the hypothesis that the individual treatment effects are zero is meaningless. It is more appropriate to test hypotheses about 2 . Specifically, H0: 2  0 H1: 2  0 If 2  0, all treatments are identical; but if 2  0, there is variability between treatments. The ANOVA decomposition of total variability is still valid; that is, SST  SSTreatments  SSE

(13-20)

However, the expected values of the mean squares for treatments and error are somewhat different than in the fixed-effects case. Expected Values of Mean Squares: Random Effects

In the random-effects model for a single-factor, completely randomized experiment, the expected mean square for treatments is E1MS Treatments 2  E a

SS Treatments b a1

 2  n2

(13-21)

and the expected mean square for error is E1MSE 2  E c  2

SSE d a1n  12 (13-22)

*The assumption that the {i} are independent random variables implies that the usual assumption of g i1 i  0 from the fixed-effects model does not apply to the random-effects model. a

JWCL232_c13_513-550.qxd

1/18/10

10:40 AM

Page 535

535

13-3 THE RANDOM-EFFECTS MODEL

From examining the expected mean squares, it is clear that both MSE and MSTreatments estimate 2 when H0: 2  0 is true. Furthermore, MSE and MSTreatments are independent. Consequently, the ratio F0 

MS Treatments MSE

(13-23)

is an F random variable with a  1 and a(n  1) degrees of freedom when H0 is true. The null hypothesis would be rejected at the -level of significance if the computed value of the test statistic f0  f ,a1, a(n1). The computational procedure and construction of the ANOVA table for the randomeffects model are identical to the fixed-effects case. The conclusions, however, are quite different because they apply to the entire population of treatments. Usually, we also want to estimate the variance components (2 and 2 ) in the model. The procedure that we will use to estimate 2 and 2 is called the analysis of variance method because it uses the information in the analysis of variance table. It does not require the normality assumption on the observations. The procedure consists of equating the expected mean squares to their observed values in the ANOVA table and solving for the variance components. When equating observed and expected mean squares in the one-way classification random-effects model, we obtain MS Treatments  2  n2

and

MSE  2

Therefore, the estimators of the variance components are Variance Components Estimates

ˆ 2  MSE

(13-24)

and ˆ 2 

MS Treatments  MSE n

(13-25)

Sometimes the analysis of variance method produces a negative estimate of a variance component. Since variance components are by definition nonnegative, a negative estimate of a variance component is disturbing. One course of action is to accept the estimate and use it as evidence that the true value of the variance component is zero, assuming that sampling variation led to the negative estimate. While this approach has intuitive appeal, it will disturb the statistical properties of other estimates. Another alternative is to reestimate the negative variance component with a method that always yields nonnegative estimates. Still another possibility is to consider the negative estimate as evidence that the assumed linear model is incorrect, requiring that a study of the model and its assumptions be made to find a more appropriate model. EXAMPLE 13-4

Textile Manufacturing

In Design and Analysis of Experiments, 7th edition (John Wiley, 2009), D. C. Montgomery describes a single-factor experiment involving the random-effects model in which a textile manufacturing company weaves a fabric on a large

number of looms. The company is interested in loom-to-loom variability in tensile strength. To investigate this variability, a manufacturing engineer selects four looms at random and makes four strength determinations on fabric samples chosen

JWCL232_c13_513-550.qxd

536

1/18/10

10:40 AM

Page 536

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

Table 13-7 Strength Data for Example 13-4 Observations 1

2

3

4

Total

Average

1 2 3 4

98 91 96 95

97 90 95 96

99 93 97 99

96 92 95 98

390 366 383 388 1527

97.5 91.5 95.8 97.0 95.45

Table 13-8 Analysis of Variance for the Strength Data Source of Variation

Sum of Squares

Degrees of Freedom

Mean Square

Looms Error Total

89.19 22.75 111.94

3 12 15

29.73 1.90

at random from each loom. The data are shown in Table 13-7 and the ANOVA is summarized in Table 13-8. From the analysis of variance, we conclude that the looms in the plant differ significantly in their ability to produce fabric of uniform strength. The variance components are estimated by ˆ 2  1.90 and ˆ 2 

29.73  1.90  6.96 4

f0

P-value

15.68

1.88 E-4

Therefore, the variance of strength in the manufacturing process is estimated by V1Yij 2  ˆ 2  ˆ 2  6.96  1.90  8.86 i

Loom

Conclusions: Most of the variability in strength in the output product is attributable to differences between looms.

This example illustrates an important application of the analysis of variance—the isolation of different sources of variability in a manufacturing process. Problems of excessive variability in critical functional parameters or properties frequently arise in qualityimprovement programs. For example, in the previous fabric strength example, the process mean is estimated by y  95.45 psi, and the process standard deviation is estimated by ˆ y  2Vˆ 1Yij 2  18.86  2.98 psi. If strength is approximately normally distributed, the distribution of strength in the outgoing product would look like the normal distribution shown in Fig. 13-8(a). If the lower specification limit (LSL) on strength is at 90 psi, a substantial proportion of the process output is fallout—that is, scrap or defective material that must be sold as second quality, and so on. This fallout is directly related to the excess variability resulting from differences between looms. Variability in loom performance could be caused by faulty setup, poor maintenance, inadequate supervision, poorly trained operators, and so forth. The engineer or manager responsible for quality improvement must identify and remove these sources of variability from the process. If this can be done, strength variability will be greatly reduced, perhaps as low as ˆ Y  2ˆ 2  21.90  1.38 psi, as shown in Fig. 13-8(b). In this improved process, reducing the variability in strength has greatly reduced the fallout, resulting in lower cost, higher quality, a more satisfied customer, and enhanced competitive position for the company.

Figure 13-8 The distribution of fabric strength. (a) Current process, (b) improved process.

Process fallout 80

85

90 LSL

95 (a)

100

105

110 psi

80

85

90 LSL

95 (b)

100

105

110 psi

JWCL232_c13_513-550.qxd

1/18/10

10:40 AM

Page 537

537

13-3 THE RANDOM-EFFECTS MODEL

EXERCISES FOR SECTION 13-3 13-28. An article in the Journal of the Electrochemical Society (Vol. 139, No. 2, 1992, pp. 524–532) describes an experiment to investigate the low-pressure vapor deposition of polysilicon. The experiment was carried out in a large-capacity reactor at Sematech in Austin, Texas. The reactor has several wafer positions, and four of these positions are selected at random. The response variable is film thickness uniformity. Three replicates of the experiment were run, and the data are as follows: Wafer Position 1 2 3 4

(a) (b) (c) (d)

5.67 1.70 1.97 1.36

4.49 2.19 1.47 1.65

Is there a difference in the wafer positions? Use  0.05. Estimate the variability due to wafer positions. Estimate the random error component. Analyze the residuals from this experiment and comment on model adequacy.

13-29. A textile mill has a large number of looms. Each loom is supposed to provide the same output of cloth per minute. To investigate this assumption, five looms are chosen at random, and their output is measured at different times. The following data are obtained:

Loom 1 2 3 4 5

(a) (b) (c) (d)

4.1 3.8 4.2 3.8 3.6

4.2 3.9 4.1 4.0 3.9

4.0 4.0 4.0 3.9 3.8

Yield (in grams) 1545 1540 1595 1445 1595 1520

1440 1555 1550 1440 1630 1455

1440 1490 1605 1595 1515 1450

1520 1560 1510 1465 1635 1480

1580 1495 1560 1545 1625 1445

13-31. An article in the Journal of Quality Technology (Vol. 13, No. 2, 1981, pp. 111–114) described an experiment that investigated the effects of four bleaching chemicals on pulp brightness. These four chemicals were selected at random from a large population of potential bleaching agents. The data are as follows: Chemical 1 2 3 4

4.1 4.0 3.9 3.7 4.0

Are the looms similar in output? Use  0.05. Estimate the variability between looms. Estimate the experimental error variance. Analyze the residuals from this experiment and check for model adequacy.

13-30. In the book Bayesian Inference in Statistical Analysis (1973, John Wiley and Sons) by Box and Tiao, the total product yield was determined for five samples randomly selected from each of six randomly chosen batches of raw material.

Pulp Brightness 77.199 80.522 79.417 78.001

74.466 79.306 78.017 78.358

92.746 81.914 91.596 77.544

76.208 80.346 80.802 77.364

82.876 73.385 80.626 77.386

Is there a difference in the chemical types? Use  0.05. Estimate the variability due to chemical types. Estimate the variability due to random error. Analyze the residuals from this experiment and comment on model adequacy. 13-32. Consider the vapor-deposition experiment described in Exercise 13-28. (a) Estimate the total variability in the uniformity response. (b) How much of the total variability in the uniformity response is due to the difference between positions in the reactor? (c) To what level could the variability in the uniformity response be reduced if the position-to-position variability in the reactor could be eliminated? Do you believe this is a significant reduction? 13-33. Reconsider Exercise 13-13 in which the effect of different diets on the protein content of cow’s milk was investigated. Suppose that the three diets reported were selected at random from a large number of diets. To simplify, delete the last two observations in the diets with n  27 (to make equal sample sizes). (a) How does this change the interpretation of the experiment? (b) What is an appropriate statistical model for this experiment? (c) Estimate the parameters of this model. (a) (b) (c) (d)

Output (lb/min) 4.0 3.9 4.1 3.6 3.8

1 2 3 4 5 6

(a) Do the different batches of raw material significantly affect mean yield? Use  0.01. (b) Estimate the variability between batches. (c) Estimate the variability between samples within batches. (d) Analyze the residuals from this experiment and check for model adequacy.

Uniformity 2.76 1.43 2.34 0.94

Batch

JWCL232_c13_513-550.qxd

538

1/18/10

10:40 AM

Page 538

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

13-4 RANDOMIZED COMPLETE BLOCK DESIGN 13-4.1 Design and Statistical Analysis In many experimental design problems, it is necessary to design the experiment so that the variability arising from a nuisance factor can be controlled. For example, consider the situation of Example 10-9, where two different methods were used to predict the shear strength of steel plate girders. Because each girder has different strength (potentially), and this variability in strength was not of direct interest, we designed the experiment by using the two test methods on each girder and then comparing the average difference in strength readings on each girder to zero using the paired t-test. The paired t-test is a procedure for comparing two treatment means when all experimental runs cannot be made under homogeneous conditions. Alternatively, we can view the paired t-test as a method for reducing the background noise in the experiment by blocking out a nuisance factor effect. The block is the nuisance factor, and in this case, the nuisance factor is the actual experimental unit—the steel girder specimens used in the experiment. The randomized block design is an extension of the paired t-test to situations where the factor of interest has more than two levels; that is, more than two treatments must be compared. For example, suppose that three methods could be used to evaluate the strength readings on steel plate girders. We may think of these as three treatments, say t1, t2, and t3. If we use four girders as the experimental units, a randomized complete block design (RCBD) would appear as shown in Fig. 13-9. The design is called a randomized complete block design because each block is large enough to hold all the treatments and because the actual assignment of each of the three treatments within each block is done randomly. Once the experiment has been conducted, the data are recorded in a table, such as is shown in Table 13-9. The observations in this table, say, yij, represent the response obtained when method i is used on girder j. The general procedure for a randomized complete block design consists of selecting b blocks and running a complete replicate of the experiment in each block. The data that result from running a RCBD for investigating a single factor with a levels and b blocks are shown in Table 13-10. There will be a observations (one per factor level) in each block, and the order in which these observations are run is randomly assigned within the block. We will now describe the statistical analysis for the RCBD. Suppose that a single factor with a levels is of interest and that the experiment is run in b blocks. The observations may be represented by the linear statistical model Yij    i  j  ij e

i  1, 2, p , a j  1, 2, p , b

(13-26)

where  is an overall mean, i is the effect of the ith treatment, j is the effect of the jth block, and ij is the random error term, which is assumed to be normally and independently distributed Block 1

Block 2

Block 3

Block 4

t1

t1

t1

t1

t2

t2

t2

t2

t3

t3

t3

t3

Figure 13-9 A randomized complete block design.

Table 13-9

A Randomized Complete Block Design Block (Girder)

Treatments (Method)

1

2

3

4

1 2 3

y11 y21 y31

y12 y22 y32

y13 y23 y33

y14 y24 y34

JWCL232_c13_513-550.qxd

1/18/10

10:40 AM

Page 539

13-4 RANDOMIZED COMPLETE BLOCK DESIGN

539

Table 13-10 A Randomized Complete Block Design with a Treatments and b Blocks Blocks Treatments

1

2

p

b

Totals

Averages

1 2

y11 y21

y12 y22

p p

y1b y2b

y1. y2.

y1. y2.

o a Totals Averages

o ya1 y.1 y.1

o ya2 y.2 y.2

p p p

o yab y.b y.b

o ya. y..

o ya. y..

with mean zero and variance 2. Furthermore, the treatment and block effects are defined as dea b viations from the overall mean, so g i1 i  0 and g j1 j  0 . This was the same type of definition used for completely randomized experiments in Section 13-2. We also assume that treatments and blocks do not interact. That is, the effect of treatment i is the same regardless of which block (or blocks) it is tested in. We are interested in testing the equality of the treatment effects. That is, H0: 1  2  p  a  0 H1: i 0 at least one i The analysis of variance can be extended to the RCBD. The procedure uses a sum of squares identity that partitions the total sum of squares into three components. ANOVA Sums of Squares Identity: Randomized Block Experiment

The sum of squares identity for the randomized complete block design is a

b

a

b

2 2 2 a a 1 yij  y..2  b a 1 yi.  y..2  a a 1 y.j  y..2

i1 j1

i1 a

j1

b

 a a 1 yij  y.j  yi.  y..2 2 i1 j1

or symbolically SST  SSTreatments  SSBlocks  SSE Furthermore, the degrees of freedom corresponding to these sums of squares are ab  1  1a  12  1b  12  1a  121b  12 For the randomized block design, the relevant mean squares are SSTreatments a1 SSBlocks MSBlocks  b1 SSE MSE  1a  121b  12

MSTreatments 

(13-27)

JWCL232_c13_513-550.qxd

540

1/18/10

10:41 AM

Page 540

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

The expected values of these mean squares can be shown to be as follows: Expected Mean Squares: Randomized Block Experiment

a

E1MSTreatments 2  2 

b a 2i i1

a1 b

E1MSBlocks 2  2 

a a 2j

E1MSE 2  

j1

b1

2

Therefore, if the null hypothesis H0 is true so that all treatment effects i  0, MSTreatments is an unbiased estimator of 2, while if H0 is false, MSTreatments overestimates 2. The mean square for error is always an unbiased estimate of 2. To test the null hypothesis that the treatment effects are all zero, we use the ratio F0 

MSTreatments MSE

(13-28)

which has an F-distribution with a  1 and (a  1)(b  1) degrees of freedom if the null hypothesis is true. We would reject the null hypothesis at the -level of significance if the computed value of the test statistic in Equation 13-28 is f0  f ,a1,(a1)(b1). In practice, we compute SST, SSTreatments and SSBlocks and then obtain the error sum of squares SSE by subtraction. The appropriate computing formulas are as follows. Computing Formulas for ANOVA: Randomized Block Experiment

The computing formulas for the sums of squares in the analysis of variance for a randomized complete block design are a b y..2 SST  a a y2ij  ab i1 j1

(13-29)

y..2 1 a 2 y .  a i b i1 ab

(13-30)

y..2 1 b SSBlocks  a a y2. j  ab j1

(13-31)

SSE  SST  SS Treatments  SSBlocks

(13-32)

SSTreatments 

and

The computations are usually arranged in an ANOVA table, such as is shown in Table 13-11. Generally, a computer software package will be used to perform the analysis of variance for the randomized complete block design.

JWCL232_c13_513-550.qxd

1/18/10

10:41 AM

Page 541

541

13-4 RANDOMIZED COMPLETE BLOCK DESIGN

Table 13-11

ANOVA for a Randomized Complete Block Design

Source of Variation

EXAMPLE 13-5

Degrees of Freedom

Sum of Squares

Treatments

SSTreatments

a1

Blocks

SSBlocks

b1

Error

SSE (by subtraction)

Total

SST

SSTreatments a1 SSBlocks b1

F0 MSTreatments MSE

SSE 1a  121b  12

(a  1)(b  1) ab  1

Fabric Strength 5 y.2j y..2 SSBlocks  a a  ab j1

An experiment was performed to determine the effect of four different chemicals on the strength of a fabric. These chemicals are used as part of the permanent press finishing process. Five fabric samples were selected, and a RCBD was run by testing each chemical type once in random order on each fabric sample. The data are shown in Table 13-12. We will test for differences in means using an ANOVA with  0.01. The sums of squares for the analysis of variance are computed as follows: 4 5 y..2 SST  a a y 2ij  ab i1 j1

 11.32  11.62  p  13.42  2

2

2

1 39.22 2 20



19.22 2  110.12 2  13.52 2  18.82 2  17.62 2



139.22

4 2

 6.69 20 SSE  SST  SSBlocks  SSTreatments  25.69  6.69  18.04  0.96 The ANOVA is summarized in Table 13-13. Since f0  75.13  f0.01,3,12  5.95 (the P-value is 4.79  108), we conclude that there is a significant difference in the chemical types so far as their effect on strength is concerned.

 25.69

4 y 2. y..2 i  SSTreatments  a ab i1 b



Mean Square

15.72 2  18.82 2  16.92 2  117.82 2



139.22 2 20

5  18.04

Table 13-12

Fabric Strength Data—Randomized Complete Block Design Treatment Totals

Fabric Sample Chemical Type

1

2

3

4

5

1 2 3 4

1.3 2.2 1.8 3.9

1.6 2.4 1.7 4.4

0.5 0.4 0.6 2.0

1.2 2.0 1.5 4.1

1.1 1.8 1.3 3.4

5.7 8.8 6.9 17.8

9.2 2.30

10.1 2.53

3.5 0.88

8.8 2.20

7.6 1.90

39.2(y..)

Block totals y.j Block averages y.j

yi.

Treatment Averages yi. 1.14 1.76 1.38 3.56 1.96( y..)

JWCL232_c13_513-550.qxd

542

1/18/10

10:41 AM

Page 542

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

Table 13-13

Analysis of Variance for the Randomized Complete Block Experiment

Source of Variation Chemical types (treatments) Fabric samples (blocks) Error Total

Sum of Squares

Degrees of Freedom

Mean Square

f0

P-value

18.04

3

6.01

75.13

4.79 E-8

6.69 0.96 25.69

4 12 19

1.67 0.08

When Is Blocking Necessary? Suppose an experiment is conducted as a randomized block design, and blocking was not really necessary. There are ab observations and (a  1)(b  1) degrees of freedom for error. If the experiment had been run as a completely randomized single-factor design with b replicates, we would have had a(b  1) degrees of freedom for error. Therefore, blocking has cost a(b  1)  (a  1)(b  1)  b  1 degrees of freedom for error. Thus, since the loss in error degrees of freedom is usually small, if there is a reasonable chance that block effects may be important, the experimenter should use the randomized block design. For example, consider the experiment described in Example 13-5 as a single-factor experiment with no blocking. We would then have 16 degrees of freedom for error. In the randomized block design, there are 12 degrees of freedom for error. Therefore, blocking has cost only 4 degrees of freedom, which is a very small loss considering the possible gain in information that would be achieved if block effects are really important. The block effect in Example 13-5 is large, and if we had not blocked, SSBlocks would have been included in the error sum of squares for the completely randomized analysis. This would have resulted in a much larger MSE, making it more difficult to detect treatment differences. As a general rule, when in doubt as to the importance of block effects, the experimenter should block and gamble that the block effect does exist. If the experimenter is wrong, the slight loss in the degrees of freedom for error will have a negligible effect, unless the number of degrees of freedom is very small. Computer Solution Table 13-14 presents the computer output from Minitab for the randomized complete block design in Example 13-5. We used the analysis of variance menu for balanced designs to solve this problem. The results agree closely with the hand calculations from Table 13-13. Notice that Minitab computes an F-statistic for the blocks (the fabric samples). The validity of this ratio as a test statistic for the null hypothesis of no block effects is doubtful because the blocks represent a restriction on randomization; that is, we have only randomized within the blocks. If the blocks are not chosen at random, or if they are not run in random order, the F-ratio for blocks may not provide reliable information about block effects. For more discussion see Montgomery (2009, Chapter 4).

13-4.2 Multiple Comparisons When the ANOVA indicates that a difference exists between the treatment means, we may need to perform some follow-up tests to isolate the specific differences. Any multiple comparison method, such as Fisher’s LSD method, could be used for this purpose.

JWCL232_c13_513-550.qxd

1/18/10

10:41 AM

Page 543

13-4 RANDOMIZED COMPLETE BLOCK DESIGN

543

Table 13-14 Minitab Analysis of Variance for the Randomized Complete Block Design in Example 13-5 Analysis of Variance (Balanced Designs) Factor Chemical Fabric S

Type fixed fixed

Levels 4 5

Values 1 1

2 2

3 3

4 4

5

Analysis of Variance for strength Source Chemical Fabric S Error Total

DF 3 4 12 19

SS 18.0440 6.6930 0.9510 25.6880

MS 6.0147 1.6733 0.0792

F 75.89 21.11

P 0.000 0.000

F-test with denominator: Error Denominator MS  0.079250 with 12 degrees of freedom Numerator Chemical Fabric S

DF 3 4

MS 6.015 1.673

F 75.89 21.11

P 0.000 0.000

We will illustrate Fisher’s LSD method. The four chemical type averages from Example 13-5 are: y1.  1.14

y2.  1.76

y3.  1.38

y4.  3.56

Each treatment average uses b  5 observations (one from each block). We will use  0.05, so t0.025,12  2.179. Therefore the value of the LSD is LSD  t0.025,12

210.082 2MSE  2.179  0.39 B b B 5

Any pair of treatment averages that differ by 0.39 or more indicates that this pair of treatment means is significantly different. The comparisons are shown below: 4 vs. 1  y4. 4 vs. 3  y4. 4 vs. 2  y4. 2 vs. 1  y2. 2 vs. 3  y2. 3 vs. 1  y3.

     

y1. y3. y2. y1. y3. y1.

     

3.56 3.56 3.56 1.76 1.76 1.38

     

1.14 1.38 1.76 1.14 1.38 1.14

     

2.42 2.18 1.80 0.62 0.38 0.24

     

0.39 0.39 0.39 0.39 0.39 0.39

Figure 13-10 presents the results graphically. The underlined pairs of means are not different. The LSD procedure indicates that chemical type 4 results in significantly different strengths than the other three types do. Chemical types 2 and 3 do not differ, and types 1 and 3 do not differ. There may be a small difference in strength between types 1 and 2.

JWCL232_c13_513-550.qxd

544

1/18/10

10:41 AM

Page 544

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE Chemical type 1 3

Figure 13-10 Results of Fisher’s LSD method.

0

2

1

4

2

3

4

5

6

13-4.3 Residual Analysis and Model Checking In any designed experiment, it is always important to examine the residuals and to check for violation of basic assumptions that could invalidate the results. As usual, the residuals for the RCBD are just the difference between the observed and estimated (or fitted) values from the statistical model, say, eij  yij  yˆ ij

(13-33)

and the fitted values are yˆij  yi.  y.j  y.. The fitted value represents the estimate of the mean response when the ith treatment is run in the jth block. The residuals from the chemical type experiment are shown in Table 13-15. Figures 13-11, 13-12, 13-13, and 13-14 present the important residual plots for the experiment. These residual plots are usually constructed by computer software packages. Table 13-15 Residuals from the Randomized Complete Block Design Fabric Sample

Chemical Type

1

2

3

4

5

1 2 3 4

0.18 0.10 0.08 0.00

0.10 0.08 0.24 0.28

0.44 0.28 0.30 0.48

0.18 0.00 0.12 0.30

0.02 0.10 0.02 0.10

Normal score zj

2

1 0.5 0

ei j

–1

0

1

2

3

4

–2 –0.50

–0.25

0

0.25

0.50

Residual value

Figure 13-11 Normal probability plot of residuals from the randomized complete block design.

–0.5

Figure 13-12 Residuals by treatment.

JWCL232_c13_513-550.qxd

1/18/10

10:41 AM

Page 545

13-4 RANDOMIZED COMPLETE BLOCK DESIGN 0.5

ei j

545

0.5

ei j 0

1

2

3

5

4

–0.5

0

2

4

6

yi j

–0.5

Figure 13-13 Residuals by block.

Figure 13-14 Residuals versus yˆ ij.

There is some indication that fabric sample (block) 3 has greater variability in strength when treated with the four chemicals than the other samples. Chemical type 4, which provides the greatest strength, also has somewhat more variability in strength. Followup experiments may be necessary to confirm these findings, if they are potentially important.

EXERCISES FOR SECTION 13-4 13-34.

Consider the computer output below.

Randomized Block ANOVA: y versus Factor, Block Source DF SS MS F Factor ? 193.800 64.600 ? Block 3 464.218 154.739 Error ? ? 4.464 Total 15 698.190

P ?

(a) How many levels of the factor were used in this experiment? (b) How many blocks were used in this experiment? (c) Fill in the missing information. Use bounds for the P-value. (d) What conclusions would you draw if  0.05? What would you conclude if  0.01? 13-35. Exercise 13-2 introduced you to an experiment to investigate the potential effect of consuming chocolate on cardiovascular health. The experiment was conducted as a completely randomized design, and the exercise asked you to use the ANOVA to analyze the data and draw conclusions. Now assume that the experiment had been conducted as a RCBD, with the subjects considered as blocks. Analyze the data using this assumption. What conclusions would you draw (using  0.05) about the effect of the different types of chocolate on cardiovascular health? Would your conclusions change if  0.01? 13-36. An article in Quality Engineering [“Designed Experiment to Stabilize Blood Glucose Levels” (1999–2000, Vol. 12, pp. 83–87)] described an experiment to minimize

variations in blood glucose levels. The treatment was the exercise time on a Nordic Track cross-country skier (10 or 20 min). The experiment was blocked for time of day. The data are as follows: Exercise (min)

Time of Day

Average Blood Glucose

10 10 20 20 10 10 20 20

pm am am pm am pm pm am

71.5 103 83.5 126 125.5 129.5 95 93

(a) Is there an effect of exercise time on the average blood glucose? Use  0.05. (b) Find the P-value for the test in part (a). (c) Analyze the residuals from this experiment. 13-37. In “The Effect of Nozzle Design on the Stability and Performance of Turbulent Water Jets” (Fire Safety Journal, Vol. 4, August 1981), C. Theobald described an experiment in which a shape measurement was determined for several different nozzle types at different levels of jet efflux velocity. Interest in this experiment focuses primarily on nozzle type, and velocity is a nuisance factor. The data are as follows:

JWCL232_c13_513-550.qxd

546

1/18/10

10:41 AM

Page 546

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

a field test for detecting the presence of arsenic in urine samples. The test has been proposed for use among forestry workers because of the increasing use of organic arsenics in that industry. The experiment compared the test as performed by both a trainee and an experienced trainer to an analysis at a remote laboratory. Four subjects were selected for testing and are considered as blocks. The response variable is arsenic content (in ppm) in the subject’s urine. The data are as follows:

Jet Efflux Velocity (m/s)

Nozzle Type

11.73

14.37

16.59

20.43

23.46

28.74

1 2 3 4 5

0.78 0.85 0.93 1.14 0.97

0.80 0.85 0.92 0.97 0.86

0.81 0.92 0.95 0.98 0.78

0.75 0.86 0.89 0.88 0.76

0.77 0.81 0.89 0.86 0.76

0.78 0.83 0.83 0.83 0.75

(a) Does nozzle type affect shape measurement? Compare the nozzles with box plots and the analysis of variance. (b) Use Fisher’s LSD method to determine specific differences between the nozzles. Does a graph of the average (or standard deviation) of the shape measurements versus nozzle type assist with the conclusions? (c) Analyze the residuals from this experiment.

Subject Test Trainee Trainer Lab

13-38. In Design and Analysis of Experiments, 7th edition (John Wiley & Sons, 2009), D. C. Montgomery described an experiment that determined the effect of four different types of tips in a hardness tester on the observed hardness of a metal alloy. Four specimens of the alloy were obtained, and each tip was tested once on each specimen, producing the following data:

1

2

3

4

1 2 3 4

9.3 9.4 9.2 9.7

9.4 9.3 9.4 9.6

9.6 9.8 9.5 10.0

10.0 9.9 9.7 10.2

3

4

0.05 0.05 0.04

0.05 0.05 0.04

0.04 0.04 0.03

0.15 0.17 0.10

13-40. An article in the Food Technology Journal (Vol. 10, 1956, pp. 39–42) described a study on the protopectin content of tomatoes during storage. Four storage times were selected, and samples from nine lots of tomatoes were analyzed. The protopectin content (expressed as hydrochloric acid soluble fraction mg/kg) is in Table 13-16. (a) The researchers in this study hypothesized that mean protopectin content would be different at different storage times. Can you confirm this hypothesis with a statistical test using  0.05? (b) Find the P-value for the test in part (a). (c) Which specific storage times are different? Would you agree with the statement that protopectin content decreases as storage time increases? (d) Analyze the residuals from this experiment.

(a) Is there any difference in hardness measurements between the tips? (b) Use Fisher’s LSD method to investigate specific differences between the tips. (c) Analyze the residuals from this experiment.

13-41. An experiment was conducted to investigate leaking current in a SOS MOSFETS device. The purpose of the experiment was to investigate how leakage current varies as the channel length changes. Four channel lengths were selected. For each channel length, five different widths were

13-39. An article in the American Industrial Hygiene Association Journal (Vol. 37, 1976, pp. 418–422) described

Table 13-16

2

(a) Is there any difference in the arsenic test procedure? (b) Analyze the residuals from this experiment.

Specimen

Type of Tip

1

Protopectin Content of Tomatoes in Storage Lot

Storage Time

1

2

3

4

5

6

7

8

9

0 days 7 days 14 days 21 days

1694.0 1802.0 1568.0 415.5

989.0 1074.0 646.2 845.4

917.3 278.8 1820.0 377.6

346.1 1375.0 1150.0 279.4

1260.0 544.0 983.7 447.8

965.6 672.2 395.3 272.1

1123.0 818.0 422.3 394.1

1106.0 406.8 420.0 356.4

1116.0 461.6 409.5 351.2

JWCL232_c13_513-550.qxd

1/18/10

10:41 AM

Page 547

547

13-4 RANDOMIZED COMPLETE BLOCK DESIGN

also used, and width is to be considered a nuisance factor. The data are as follows: Width

Channel Length

1

2

3

4

5

1 2 3 4

0.7 0.8 0.9 1.0

0.8 0.8 1.0 1.5

0.8 0.9 1.7 2.0

0.9 0.9 2.0 3.0

1.0 1.0 4.0 20.0

(a) Test the hypothesis that mean leakage voltage does not depend on the channel length, using  0.05. (b) Analyze the residuals from this experiment. Comment on the residual plots. (c) The observed leakage voltage for channel length 4 and width 5 was erroneously recorded. The correct observation is 4.0. Analyze the corrected data from this experiment. Is there evidence to conclude that mean leakage voltage increases with channel length?

Supplemental Exercises 13-42.

Consider the computer output below.

One-way ANOVA: y versus Factor Source DF SS Factor ? ? Error 15 167.5 Total 19 326.2 S  3.342 (a) (b) (c) (d)

are used extensively in equipment such as air turbine starters. Five different carbon materials were tested, and the surface roughness was measured. The data are as follows:

RSq  ?

MS ? ?

F ?

P ?

RSq(adj)  34.96%

How many levels of the factor were used in this experiment? How many replicates were used? Fill in the missing information. Use bounds for the P-value. What conclusions would you draw if  0.05? What if  0.01?

Carbon Material Type EC10 EC10A EC4 EC1

(a) (b) (c) (d)

P ?

How many levels of the factor were used in this experiment? How many blocks were used? Fill in the missing information. Use bounds for the P-value. What conclusions would you draw if  0.05? What if  0.01? 13-44. An article in Lubrication Engineering (December 1990) described the results of an experiment designed to investigate the effects of carbon material properties on the progression of blisters on carbon face seals. The carbon face seals

0.55 0.07 0.28 0.16

0.55 0.25 0.12

0.36 0.18

0.56

0.20

(a) Does carbon material type have an effect on mean surface roughness? Use  0.05. (b) Find the residuals for this experiment. Does a normal probability plot of the residuals indicate any problem with the normality assumption? (c) Plot the residuals versus yˆ ij. Comment on the plot. (d) Find a 95% confidence interval on the difference between the mean surface roughness between the EC10 and the EC1 carbon grades. (e) Apply the Fisher LSD method to this experiment. Summarize your conclusions regarding the effect of material type on surface roughness. 13-45. An article in the IEEE Transactions on Components, Hybrids, and Manufacturing Technology (Vol. 15, No. 2, 1992, pp. 146–153) described an experiment in which the contact resistance of a brake-only relay was studied for three different materials (all were silver-based alloys). The data are as follows.

Alloy

Contact Resistance

1

95 99

97 99

99 94

98 95

99 98

2

104 102

102 111

102 103

105 100

99 103

3

119 172

130 145

132 150

136 144

141 135

13-43. Consider the computer output below. Randomized Block ANOVA: y versus Factor, Block Source DF SS MS F Factor ? 126.880 63.4401 ? Block ? 54.825 18.2751 Error 6 ? 2.7403 Total 11 198.147

Surface Roughness 0.50 0.31 0.20 0.10

(a) Does the type of alloy affect mean contact resistance? Use  0.01. (b) Use Fisher’s LSD method to determine which means differ. (c) Find a 99% confidence interval on the mean contact resistance for alloy 3. (d) Analyze the residuals for this experiment. 13-46. An article in the Journal of Quality Technology (Vol. 14, No. 2, 1982, pp. 80–89) described an experiment in which three different methods of preparing fish are evaluated on the basis of sensory criteria and a quality score is

JWCL232_c13_513-550.qxd

548

1/18/10

10:41 AM

Page 548

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

(d) Analyze the residuals from this experiment and comment on model adequacy.

assigned. Assume that these methods have been randomly selected from a large population of preparation methods. The data are in the following table:

Method

13-48. An article in Agricultural Engineering (December 1964, pp. 672–673) described an experiment in which the daily weight gain of swine is evaluated at different levels of housing temperature. The mean weight of each group of swine at the start of the experiment is considered to be a nuisance factor. The data from this experiment are as follows:

Score

1

24.4 22.2

23.2 24.4

25.0 23.8

19.7 18.0

2

22.1 22.3

19.5 23.2

17.3 21.4

19.7 22.6

3

23.3 20.4

22.8 23.5

22.4 20.8

23.7 24.1

Mean Weight (lbs) 100 150 200

(a) Is there any difference in preparation methods? Use  0.05. (b) Calculate the P-value for the F-statistic in part (a). (c) Analyze the residuals from this experiment and comment on model adequacy. (d) Estimate the components of variance.

70.0 75.0 80.0

1235 1240 1200

1285 1200 1170

1245 1220 1155

1235 1210 1095

(a) Does drying temperature affect mean bread volume? Use  0.01. (b) Find the P-value for this test. (c) Use the Fisher LSD method to determine which means are different. Table 13-17

60

70

80

90

100

1.37 1.47 1.19

1.58 1.75 1.91

2.00 2.16 2.22

1.97 1.82 1.67

1.40 1.14 0.88

0.39 0.19 0.77

13-49. An article in Communications of the ACM (Vol. 30, No. 5, 1987) studied different algorithms for estimating software development costs. Six algorithms were applied to eight software development projects and the percent error in estimating the development cost was observed. The data are in Table 13-17. (a) Do the algorithms differ in their mean cost estimation accuracy? Use  0.05. (b) Analyze the residuals from this experiment. (c) Which algorithm would you recommend for use in practice?

Volume (CC) 1245 1235 1225

50

(a) Does housing air temperature affect mean weight gain? Use  0.05. (b) Use Fisher’s LSD method to determine which temperature levels are different. (c) Analyze the residuals from this experiment and comment on model adequacy.

13-47. An article in the Journal of Agricultural Engineering Research (Vol. 52, 1992, pp. 53–76) described an experiment to investigate the effect of drying temperature of wheat grain on the baking quality of bread. Three temperature levels were used, and the response variable measured was the volume of the loaf of bread produced. The data are as follows: Temperature (°C)

Housing Air Temperatures (degrees F)

13-50. An article in Nature Genetics (2003, Vol. 34(1), pp. 85–90) “Treatment-specific changes in gene expression discriminate in vivo drug response in human leukemia cells” studied gene expression as a function of different treatments

Software Development Costs Project

Algorithm 1(SLIM) 2(COCOMO-A) 3(COCOMO-R) 4(COCOMO-C) 5(FUNCTION POINTS) 6(ESTIMALS)

1

2

3

4

5

6

7

8

1244 281 220 225 19 20

21 129 84 83 11 35

82 396 458 425 34 53

2221 1306 543 552 121 170

905 336 300 291 15 104

839 910 794 826 103 199

527 473 488 509 87 142

122 199 142 153 17 41

JWCL232_c13_513-550.qxd

1/18/10

10:41 AM

Page 549

549

13-4 RANDOMIZED COMPLETE BLOCK DESIGN

Table 13-18

Treatment-Specific Changes in Gene Expression

Treatments MP ONLY MPHDMTX MPLDMTX

Observations 334.5 919.4 108.4

31.6 404.2 26.1

701 1024.8 240.8

41.2 54.1 191.1

for leukemia. Three treatment groups are: mercaptopurine (MP) only; low-dose methotrexate (LDMTX) and MP; and high-dose methotrexate (HDMTX) and MP. Each group contained ten subjects. The responses from a specific gene are shown in Table 13-18. (a) Check the normality of the data. Can we assume these samples are from normal populations? (b) Take the logarithm of the raw data and check the normality of the transformed data. Is there evidence to support the claim that the treatment means differ for the transformed data? Use  0.1. (c) Analyze the residuals from the transformed data and comment on model adequacy.

61.2 62.8 69.7

69.6 671.6 242.8

67.5 882.1 62.7

66.6 354.2 396.9

120.7 321.9 23.6

881.9 91.1 290.4

13-51. Consider an ANOVA situation with a  5 treatments. Let 2  9 and  0.05, and suppose that n  4. (a) Find the power of the ANOVA F-test when 1  2  3  1, 4  3, and 5  2. (b) What sample size is required if we want the power of the F-test in this situation to be at least 0.90? 13-52. Consider an ANOVA situation with a  4 means 1  1, 2  5, 3  8, and 4  4. Suppose that 2  4, n  4, and  0.05. (a) Find the power of the ANOVA F-test. (b) How large would the sample size have to be if we want the power of the F-test for detecting this difference in means to be at least 0.90?

MIND-EXPANDING EXERCISES 13-53. Show that in the fixed-effects model analysis of variance E(MSE)  2. How would your development change if the random-effects model had been specified? 13-54. Consider testing the equality of the means of two normal populations where the variances are unknown but are assumed equal. The appropriate test procedure is the two-sample t-test. Show that the two-sample t-test is equivalent to the single-factor analysis of variance F-test. 13-55. Consider the ANOVA with a  2 treatments. Show that the MSE in this analysis is equal to the pooled variance estimate used in the two-sample t-test. 13-56. Show that the variance of the linear combination a

a

a ciYi. is  a 2

i1

nic2i .

means is as large as D, the minimum value that the OC curve parameter 2 can take is 2 

nD2 2a2

13-59. Consider the single-factor completely randomized design. Show that a 100(1  ) percent confidence interval for 2 is 1N  a2MSE  2 2, Na

2

1N  a2MSE  21 2, Na

i1

13-57. In a fixed-effects model, suppose that there are n observations for each of four treatments. Let Q 21, Q22, and Q23 be single-degree-of-freedom sums of squares for orthogonal contrasts. A contrast is a linear combination of the treatment means with coefficients that sum to zero. The coefficient vectors of orthogonal contrasts are orthogonal vectors. Prove that SSTreatments  Q21  Q22  Q23. 13-58. Consider the single-factor completely randomized design with a treatments and n replicates. Show that if the difference between any two treatment

where N is the total number of observations in the experimental design. 13-60. Consider the random-effect model for the single-factor completely randomized design. Show that a 100(1  )% confidence interval on the ratio of variance components 22 is given by L

 2

U 2

JWCL232_c13_513-550.qxd

550

1/18/10

10:41 AM

Page 550

CHAPTER 13 DESIGN AND ANALYSIS OF SINGLE-FACTOR EXPERIMENTS: THE ANALYSIS OF VARIANCE

MIND-EXPANDING EXERCISES where 1 1 MSTreatments a b  1d Ln c MSE f 2,a1,Na

the model parameters.) For the case of unequal sample a size n1, n2, p , na, the restriction is g i1 ni i  0 . Use this to show that a

E1MSTreatments 2  2 

and 1 1 MSTreatments a b  1d Un c MSE f1 2,a1,Na 13-61. Consider a random-effects model for the single-factor completely randomized design. (a) Show that a 100(1  )% confidence interval on the ratio 2 (2  2) is 2 L U

2 2 1L 1  U    where L and U are as defined in Exercise 13-60. (b) Use the results of part (a) to find a 100(1  )% confidence interval for 2(2  2). 13-62. Consider the fixed-effect model of the completely randomized single-factor design. The model a parameters are restricted by the constraint g i1 i  0 . (Actually, other restrictions could be used, but this one is simple and results in intuitively pleasing estimates for

2 a nii

i1

a1

Does this suggest that the null hypothesis in this model is H0: n11  n22  p  naa  0? 13-63. Sample Size Determination. In the singlefactor completely randomized design, the accuracy of a 100(1  )% confidence interval on the difference in any two treatment means is t 2,a1n12 12MSE n. (a) Show that if A is the desired accuracy of the interval, the sample size required is

n

2F 2,1,a1n12 MSE A2

(b) Suppose that in comparing a  5 means we have a preliminary estimate of 2 of 4. If we want the 95% confidence interval on the difference in means to have an accuracy of 2, how many replicates should we use?

IMPORTANT TERMS AND CONCEPTS Analysis of variance (ANOVA) Blocking Completely randomized experiment Expected mean squares Fisher’s least significant difference (LSD) method

Fixed factor Graphical comparison of means Levels of a factor Mean square Multiple comparisons Nuisance factors

Random factor Randomization Randomized complete block design Residual analysis and model adequacy checking

Sample size and replication in an experiment Treatment effect Variance component

JWCL232_c14_551-636.qxd

1/1/70

5:38 AM

Page 551

14

©Oleksiy Kondratyuk/iStockphoto

Design of Experiments with Several Factors

Carotenoids are fat-soluble pigments that occur naturally in fruits in vegetables that are recommended for healthy diets. A well-known carotenoid is beta-carotene. Astaxanthin is another carotenoid that is a strong antioxidant and commercially produced. An exercise later in this chapter describes an experiment in Biotechnology Progress to promote astaxanthin production. Seven variables were considered important to production: photon flux density, and concentrations of nitrogen, phosphorous, magnesium, acetate, ferrous, NaCl. It was important to study the effects of these factors, but also the effects of combinations on the production. Even with only a high and low setting for each variable, an experiment that uses all possible combinations requires 27 ⫽ 128 tests. There are a number of disadvantages of such a large experiment and a question is whether a fraction of the full set of tests can be selected to provide the most important information about the effects of these variables in many fewer runs. The example used a surprisingly small set of 16 runs (16/128 ⫽ 1/8 fraction). The design and analysis of experiments of this type is the focus of this chapter. Such experiments are widely-used throughout modern engineering development and scientific studies.

CHAPTER OUTLINE 14-1 INTRODUCTION 14-2 FACTORIAL EXPERIMENTS 14-3 TWO-FACTOR FACTORIAL EXPERIMENTS 14-3.1 Statistical Analysis of the Fixed-Effects Model 14-3.2 Model Adequacy Checking 14-3.3 One Observation per Cell

14-4 GENERAL FACTORIAL EXPERIMENTS 14-5 2k FACTORIAL DESIGNS 14-5.1 22 Design 14-5.2 2k Design for k ⱖ 3 Factors 14-5.3 Single Replicate of the 2k Design 14-5.4 Addition of Center Points to a 2k Design

551

JWCL232_c14_551-636.qxd

552

1/16/10

9:55 AM

Page 552

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

14-6 BLOCKING AND CONFOUNDING IN THE 2k DESIGN 14-7 FRACTIONAL REPLICATION OF THE 2k DESIGN

14-7.2 Smaller Fractions: The 2kⴚp Fractional Factorial 14-8

RESPONSE SURFACE METHODS AND DESIGNS

14-7.1 One-Half Fraction of the 2k Design

LEARNING OBJECTIVES After careful study of this chapter you should be able to do the following: 1. Design and conduct engineering experiments involving several factors using the factorial design approach 2. Know how to analyze and interpret main effects and interactions 3. Understand how the ANOVA is used to analyze the data from these experiments 4. Assess model adequacy with residual plots 5. Know how to use the two-level series of factorial designs 6. Understand how two-level factorial designs can be run in blocks 7. Design and conduct two-level fractional factorial designs 8. Test for curvature in two-level factorial designs by using center points 9. Use response surface methodology for process optimization experiments

14-1 INTRODUCTION An experiment is just a test or series of tests. Experiments are performed in all engineering and scientific disciplines and are an important part of the way we learn about how systems and processes work. The validity of the conclusions that are drawn from an experiment depends to a large extent on how the experiment was conducted. Therefore, the design of the experiment plays a major role in the eventual solution of the problem that initially motivated the experiment. In this chapter we focus on experiments that include two or more factors that the experimenter thinks may be important. A factorial experiment is a powerful technique for this type of problem. Generally, in a factorial experimental design, experimental trials (or runs) are performed at all combinations of factor levels. For example, if a chemical engineer is interested in investigating the effects of reaction time and reaction temperature on the yield of a process, and if two levels of time (1 and 1.5 hours) and two levels of temperature (125 and 150F) are considered important, a factorial experiment would consist of making experimental runs at each of the four possible combinations of these levels of reaction time and reaction temperature. Experimental design is an extremely important tool for engineers and scientists who are interested in improving the performance of a manufacturing process. It also has extensive application in the development of new processes and in new product design. We now give some examples. Process Characterization Experiment In an article in IEEE Transactions on “Electronics Packaging Manufacturing” (2001, Vol. 24(4), pp. 249–254), the authors discussed the change to lead-free solder in surface mount technology (SMT). SMT is a process to assemble electronic components to a printed circuit board. Solder paste is printed through a stencil onto the printed circuit board. The stencil-printing machine has squeegees; the paste rolls in front of the squeegee and fills the apertures in the stencil. The squeegee shears off the paste in the apertures as it moves over the stencil. Once the print stroke

JWCL232_c14_551-636.qxd

1/16/10

9:55 AM

Page 553

14-1 INTRODUCTION

553

Controllable factors x1 x2 xp

...

Input (printed circuit boards)

Figure 14-1 flow solder experiment.

Output SMT Process

(defects, y)

...

The z1 z2

zq

Uncontrollable (noise) factors

is completed, the board is separated mechanically from the stencil. Electronic components are placed on the deposits and the board is heated so that the paste reflows to form the solder joints. The current SMT soldering process is based on tin–lead solders, and it has been well developed and refined over the years to operate at a competitive cost. The process will have several (perhaps many) variables, and all of them are not equally important. The initial list of candidate variables to be included in an experiment is constructed by combining the knowledge and information about the process from all team members. For example, engineers would conduct a brainstorming session and invite manufacturing personnel with SMT experience to participate. SMT has several variables that can be controlled. These include (1) squeegee speed, (2) squeegee pressure, (3) squeegee angle, (4) metal or polyurethane squeegee, (5) squeegee vibration, (6) delay time before the squeegee lifts from the stencil, (7) stencil separation speed, (8) print gap, (9) solder paste alloy, (10) paste pretreatment (11) paste particle size, (12) flux type, (13) reflow temperature, (14) reflow time, and so forth.

In addition to these controllable factors, there are several other factors that cannot be easily controlled during routine manufacturing, including (1) thickness of the printed circuit board, (2) types of components used on the board and aperture width and length, (3) layout of the components on the board, (4) paste density variation, (5) environmental factors, (6) squeegee wear, (7) cleanliness, and so forth.

Sometimes we call the uncontrollable factors noise factors. A schematic representation of the process is shown in Fig. 14-1. In this situation, the engineer wants to characterize the SMT process; that is, to determine the factors (both controllable and uncontrollable) that affect the occurrence of defects on the printed circuit boards. To determine these factors, an experiment can be designed to estimate the magnitude and direction of the factor effects. Sometimes we call such an experiment a screening experiment. The information from this characterization study, or screening experiment, can help determine the critical process variables as well as the direction of adjustment for these factors in order to reduce the number of defects, and assist in determining which process variables should be carefully controlled during manufacturing to prevent high defect levels and erratic process performance. Optimization Experiment In a characterization experiment, we are interested in determining which factors affect the response. A logical next step is to determine the region in the important factors that leads to an optimum response. For example, if the response is cost, we will look for a region of minimum cost. As an illustration, suppose that the yield of a chemical process is influenced by the operating temperature and the reaction time. We are currently operating the process at 155F and 1.7 hours of reaction time, and the current process yield is around 75%. Figure 14-2 shows a view of the time–temperature space from above. In this graph we have connected points of constant yield with lines. These lines are yield contours, and we have shown the

JWCL232_c14_551-636.qxd

554

1/16/10

9:55 AM

Page 554

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

200 Path leading to region of higher yield

190

95%

Temperature (°F)

180

90%

60% 70%

170 82%

58% Current operating conditions

80% 160 75% 150

Figure 14-2 Contour plot of yield as a function of reaction time and reaction temperature, illustrating an optimization experiment.

69%

56% 140

0.5

1.0

1.5

2.0

2.5

Time (hr)

contours at 60, 70, 80, 90, and 95% yield. To locate the optimum, we might begin with a factorial experiment such as we describe below, with the two factors, time and temperature, run at two levels each at 10F and 0.5 hours above and below the current operating conditions. This two-factor factorial design is shown in Fig. 14-2. The average responses observed at the four points in the experiment (145F, 1.2 hours; 145F, 2.2 hours; 165F, 1.2 hours; and 165F, 2.2 hours) indicate that we should move in the general direction of increased temperature and lower reaction time to increase yield. A few additional runs could be performed in this direction to locate the region of maximum yield. A Product Design Example We can also use experimental design in the development of new products. For example, suppose that a group of engineers are designing a door hinge for an automobile. The product characteristic is the check effort, or the holding ability, of the latch that prevents the door from swinging closed when the vehicle is parked on a hill. The check mechanism consists of a leaf spring and a roller. When the door is opened, the roller travels through an arc causing the leaf spring to be compressed. To close the door, the spring must be forced aside, and this creates the check effort. The engineering team thinks that check effort is a function of the following factors: (1) roller travel distance, (2) spring height from pivot to base, (3) horizontal distance from pivot to spring, (4) free height of the reinforcement spring, (5) free height of the main spring.

The engineers can build a prototype hinge mechanism in which all these factors can be varied over certain ranges. Once appropriate levels for these five factors have been identified, an experiment can be designed consisting of various combinations of the factor levels, and the prototype can be tested at these combinations. This will produce information concerning which factors are most influential on the latch check effort, and through analysis of this information, the latch design can be improved.

JWCL232_c14_551-636.qxd

1/16/10

9:55 AM

Page 555

14-2 FACTORIAL EXPERIMENTS

555

Most of the statistical concepts introduced in Chapter 13 for single-factor experiments can be extended to the factorial experiments of this chapter. The analysis of variance (ANOVA), in particular, will continue to be used as a tool for statistical data analysis. We will also introduce several graphical methods that are useful in analyzing the data from designed experiments.

14-2 FACTORIAL EXPERIMENTS When several factors are of interest in an experiment, a factorial experiment should be used. As noted previously, in these experiments factors are varied together. Factorial Experiment

By a factorial experiment we mean that in each complete trial or replicate of the experiment all possible combinations of the levels of the factors are investigated.

Thus, if there are two factors A and B with a levels of factor A and b levels of factor B, each replicate contains all ab treatment combinations. The effect of a factor is defined as the change in response produced by a change in the level of the factor. It is called a main effect because it refers to the primary factors in the study. For example, consider the data in Table 14-1. This is a factorial experiment with two factors, A and B, each at two levels (Alow, Ahigh, and Blow, Bhigh). The main effect of factor A is the difference between the average response at the high level of A and the average response at the low level of A, or A

30  40 10  20   20 2 2

That is, changing factor A from the low level to the high level causes an average response increase of 20 units. Similarly, the main effect of B is B

10  30 20  40   10 2 2

In some experiments, the difference in response between the levels of one factor is not the same at all levels of the other factors. When this occurs, there is an interaction between the factors. For example, consider the data in Table 14-2. At the low level of factor B, the A effect is A  30  10  20 and at the high level of factor B, the A effect is A  0  20  20 Since the effect of A depends on the level chosen for factor B, there is interaction between A and B. Table 14-1 A Factorial Experiment with Two Factors

Table 14-2 A Factorial Experiment with Interaction

Factor B

Factor B

Factor A

B low

B high

Factor A

B low

B high

Alow Ahigh

10 30

20 40

Alow Ahigh

10 30

20 40

JWCL232_c14_551-636.qxd

9:55 AM

Page 556

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

When an interaction is large, the corresponding main effects have very little practical meaning. For example, by using the data in Table 14-2, we find the main effect of A as A

30  0 10  20  0 2 2

and we would be tempted to conclude that there is no factor A effect. However, when we examined the effects of A at different levels of factor B, we saw that this was not the case. The effect of factor A depends on the levels of factor B. Thus, knowledge of the AB interaction is more useful than knowledge of the main effect. A significant interaction can mask the significance of main effects. Consequently, when interaction is present, the main effects of the factors involved in the interaction may not have much meaning. It is easy to estimate the interaction effect in factorial experiments such as those illustrated in Tables 14-1 and 14-2. In this type of experiment, when both factors have two levels, the AB interaction effect is the difference in the diagonal averages. This represents one-half the difference between the A effects at the two levels of B. For example, in Table 14-1, we find the AB interaction effect to be AB 

20  30 10  40  0 2 2

Thus, there is no interaction between A and B. In Table 14-2, the AB interaction effect is AB 

10  0 20  30   20 2 2

As we noted before, the interaction effect in these data is very large. The concept of interaction can be illustrated graphically in several ways. Figure 14-3 plots the data in Table 14-1 against the levels of A for both levels of B. Note that the Blow and Bhigh lines are approximately parallel, indicating that factors A and B do not interact significantly. Figure 14-4 presents a similar plot for the data in Table 14-2. In this graph, the Blow and Bhigh lines are not parallel, indicating the interaction between factors A and B. Such graphical displays are called two-factor interaction plots. They are often useful in presenting the results of experiments, and many computer software programs used for analyzing data from designed experiments will construct these graphs automatically.

50

50

40

Bhigh

30

Blow

20 10 0

40 Observation

Observation

556

1/16/10

20 10 Bhigh

0 Alow

Ahigh

Alow

Factor A

Figure 14-3 interaction.

Blow

30

Factorial experiment, no

Ahigh Factor A

Figure 14-4 interaction.

Factorial experiment, with

JWCL232_c14_551-636.qxd

1/16/10

9:55 AM

Page 557

557

14-2 FACTORIAL EXPERIMENTS

45 40

35

35

25

30 y

y 15

25

5

20 1 0.6 0.2

15 10

–1

–0.6

–0.2 A

0.2

0.6

1

0 –0.2 B –0.6 –1

1 0.6 0.2 – 0.2 – 0.6 B –1

–5 –15

–1

–0.6

–0.2 A

0.2

0.6

1

Figure 14-5 Three-dimensional surface plot of the data from Figure 14-6 Three-dimensional surface plot of the data from Table 14-1, showing the main effects of the two factors A and B. Table 14-2 showing the effect of the A and B interaction.

80

80

70

70 Yield (%)

Yield (%)

Figures 14-5 and 14-6 present another graphical illustration of the data from Tables 14-1 and 14-2. In Fig. 14-3 we have shown a three-dimensional surface plot of the data from Table 14-1. These data contain no interaction, and the surface plot is a plane lying above the A-B space. The slope of the plane in the A and B directions is proportional to the main effects of factors A and B, respectively. Figure 14-6 is a surface plot of the data from Table 14-2. Notice that the effect of the interaction in these data is to “twist” the plane, so that there is curvature in the response function. Factorial experiments are the only way to discover interactions between variables. An alternative to the factorial design that is (unfortunately) used in practice is to change the factors one at a time rather than to vary them simultaneously. To illustrate this one-factor-ata-time procedure, suppose that an engineer is interested in finding the values of temperature and pressure that maximize yield in a chemical process. Suppose that we fix temperature at 155F (the current operating level) and perform five runs at different levels of time, say, 0.5, 1.0, 1.5, 2.0, and 2.5 hours. The results of this series of runs are shown in Fig. 14-7. This figure indicates that maximum yield is achieved at about 1.7 hours of reaction time. To optimize temperature, the engineer then fixes time at 1.7 hours (the apparent optimum) and performs five runs at different temperatures, say, 140, 150, 160, 170, and 180F. The results of this set of runs are plotted in Fig. 14-8. Maximum yield occurs at about 155F. Therefore, we would conclude that running the process at 155F and 1.7 hours is the best set of operating conditions, resulting in yields of around 75%.

60

50

60

50

0.5

1.0

1.5

2.0

2.5

Time (hr)

Figure 14-7 Yield versus reaction time with temperature constant at 155F.

140

150

160

170

180

Temperature (°F)

Figure 14-8 Yield versus temperature with reaction time constant at 1.7 hours.

JWCL232_c14_551-636.qxd

558

1/16/10

9:55 AM

Page 558

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

200

190

95%

Temperature (°F)

180

90% 170 80% 160

150 70%

Figure 14-9 Optimization experiment using the one-factor-at-atime method.

140 60% 0.5

1.0

1.5 Time (hr)

2.0

2.5

Figure 14-9 displays the contour plot of actual process yield as a function of temperature and time with the one-factor-at-a-time experiments superimposed on the contours. Clearly, this one-factor-at-a-time approach has failed dramatically here, as the true optimum is at least 20 yield points higher and occurs at much lower reaction times and higher temperatures. The failure to discover the importance of the shorter reaction times is particularly important because this could have significant impact on production volume or capacity, production planning, manufacturing cost, and total productivity. The one-factor-at-a-time approach has failed here because it cannot detect the interaction between temperature and time. Factorial experiments are the only way to detect interactions. Furthermore, the one-factor-at-a-time method is inefficient. It will require more experimentation than a factorial, and as we have just seen, there is no assurance that it will produce the correct results.

14-3 TWO-FACTOR FACTORIAL EXPERIMENTS The simplest type of factorial experiment involves only two factors, say A, and B. There are a levels of factor A and b levels of factor B. This two-factor factorial is shown in Table 14-3. The experiment has n replicates, and each replicate contains all ab treatment combinations. The observation in the ij th cell for the kth replicate is denoted by yijk. In performing the experiment, the abn observations would be run in random order. Thus, like the single-factor experiment studied in Chapter 13, the two-factor factorial is a completely randomized design. The observations may be described by the linear statistical model i  1, 2, p , a Yijk    i  j  12 ij  ijk • j  1, 2, p , b k  1, 2, p , n

(14-1)

JWCL232_c14_551-636.qxd

1/16/10

9:55 AM

Page 559

14-3 TWO-FACTOR FACTORIAL EXPERIMENTS

559

Table 14-3 Data Arrangement for a Two-Factor Factorial Design Factor B

1

Factor A

2

p

1 y111, y112, p , y11n

2 y121, y122, p , y12n

b y1b1, y1b2, p , y1bn

y211, y212, p , y21n

y221, y222, p , y22n

y2b1, y2b2, p , y2bn

ya11, ya12, p , ya1n

ya21, ya22, p , ya2n

yab1, yab2, p , yabn

y.1. y.1.

y.2. y.2.

y.b . y.b.

Totals

Averages

y1..

y1..

y2..

y2..

ya..

ya..

o a Totals Averages

yp y...

where  is the overall mean effect, i is the effect of the i th level of factor A, j is the effect of the jth level of factor B, ()ij is the effect of the interaction between A and B, and ijk is a random error component having a normal distribution with mean zero and variance 2. We are interested in testing the hypotheses of no main effect for factor A, no main effect for B, and no AB interaction effect. As with the single-factor experiments of Chapter 13, the analysis of variance (ANOVA) will be used to test these hypotheses. Since there are two factors in the experiment, the test procedure is sometimes called the two-way analysis of variance.

14-3.1 Statistical Analysis of the Fixed-Effects Model Suppose that A and B are fixed factors. That is, the a levels of factor A and the b levels of factor B are specifically chosen by the experimenter, and inferences are confined to these levels only. In this model, it is customary to define the effects i, j, and ()ij as deviations from the a b a b mean, so that g i1 i  0, g j1 j  0, g i1 12 ij  0, and g j1 12 ij  0. The analysis of variance can be used to test hypotheses about the main factor effects of A and B and the AB interaction. To present the ANOVA, we will need some symbols, some of which are illustrated in Table 14-3. Let yi.. denote the total of the observations taken at the ith level of factor A; y.j . denote the total of the observations taken at the jth level of factor B; yij. denote the total of the observations in the ij th cell of Table 14-3; and y... denote the grand total of all the observations. Define yi.., y.j., yij., and y... as the corresponding row, column, cell, and grand averages. That is,

Notation for Totals and Means

b

n

y .j.  a a yijk

yi.. bn y.j. y. j .  an

yij ..  a yijk

y ij. 

yij. n

i  1, 2, p , a

y... 

y... abn

j  1, 2, p , b

yi ..  a a yijk j1 k1 a n i1 k1 n

k1 a

b

n

y ...  a a a yijk i1 j1 k1

yi .. 

i  1, 2, p , a j  1, 2, p , b

JWCL232_c14_551-636.qxd

560

1/16/10

9:55 AM

Page 560

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

The hypotheses that we will test are as follows: H0: 1  2  p  a  0 (no main effect of factor A) H1: at least one i  0 2. H0: 1  2  p  b  0 (no main effect of factor B) H1: at least one j  0 3. H0: ()11  ()12  p  ()ab  0 (no interaction) H1: at least one ()ij  0 1.

(14-2)

As before, the ANOVA tests these hypotheses by decomposing the total variability in the data into component parts and then comparing the various elements in this decomposition. Total variability is measured by the total sum of squares of the observations a

b

n

SST  a a a 1 yijk  y...2 2 i1 j1 k1

and the sum of squares decomposition is defined below. ANOVA Sum of Squares Identity: Two Factors

The sum of squares identity for a two-factor ANOVA is a

b

n

a

b

2 2 2 a a a 1 yijk  y...2  bn a 1 yi..  y...2  an a 1 y.j.  y...2

i1 j1 k1

i1

a

b

i1

j1

j1

a

b

n

 n a a 1 yij.  yi..  y.j.  y...2 2  a a a 1 yijk  yij.2 2

(14-3)

i1 j1 k1

or symbolically, SST  SSA  SSB  SSAB  SSE

(14-4)

Equations 14-3 and 14-4 state that the total sum of squares SST is partitioned into a sum of squares for the row factor A (SSA), a sum of squares for the column factor B (SSB), a sum of squares for the interaction between A and B (SSAB), and an error sum of squares (SSE). There are abn  1 total degrees of freedom. The main effects A and B have a  1 and b  1 degrees of freedom, while the interaction effect AB has (a  1)(b  1) degrees of freedom. Within each of the ab cells in Table 14-3, there are n  1 degrees of freedom between the n replicates, and observations in the same cell can differ only because of random error. Therefore, there are ab(n  1) degrees of freedom for error. Therefore, the degrees of freedom are partitioned according to abn  1  1a  12  1b  12  1a  121b  12  ab1n  12 If we divide each of the sum of squares on the right-hand side of Equation 14-4 by the corresponding number of degrees of freedom, we obtain the mean squares for A, B, the interaction, and error: MSA 

SSA a1

MSB 

SSB b1

MSAB 

SSAB 1a  121b  12

MSE 

SSE ab1n  12

JWCL232_c14_551-636.qxd

1/16/10

9:55 AM

Page 561

561

14-3 TWO-FACTOR FACTORIAL EXPERIMENTS

Assuming that factors A and B are fixed factors, it is not difficult to show that the expected values of these mean squares are Expected Values of Mean Squares: Two Factors

b

a

E1MSA 2  E a

bn g 2i

SSA i1 b  2  a1 a1

E1MSB 2  E a

E1MSE 2  E a

j1 SSB b  2  b1 b1

n g g 12 2ij a

E1MSAB 2  E a

an g 2j

b

i1 j1 SSAB b  2  1a  121b  12 1a  121b  12

SSE b  2 ab1n  12

From examining these expected mean squares, it is clear that if the null hypotheses about main effects H0: i  0, H0: j  0, and the interaction hypothesis H0: ()ij  0 are all true, all four mean squares are unbiased estimates of 2. To test that the row factor effects are all equal to zero (H0: i  0), we would use the ratio F Test for Factor A

F0 

MSA MSE

which has an F distribution with a  1 and ab(n  1) degrees of freedom if H0: i  0 is true. This null hypothesis is rejected at the level of significance if f0 f ,a1,ab(n1). Similarly, to test the hypothesis that all the column factor effects are equal to zero (H0: j  0), we would use the ratio F Test for Factor B

F0 

MSB MSE

which has an F distribution with b  1 and ab(n  1) degrees of freedom if H0: j  0 is true. This null hypothesis is rejected at the level of significance if f0 f ,b1,ab(n1). Finally, to test the hypothesis H0: ()ij  0, which is the hypothesis that all interaction effects are zero, we use the ratio F Test for AB Interaction

F0 

MSAB MSE

which has an F distribution with (a  1)(b  1) and ab(n  1) degrees of freedom if the null hypothesis H0: ()ij  0. This hypothesis is rejected at the level of significance if f0 f ,(a1)(b1),ab(n1).

JWCL232_c14_551-636.qxd

562

1/16/10

9:55 AM

Page 562

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

It is usually best to conduct the test for interaction first and then to evaluate the main effects. If interaction is not significant, interpretation of the tests on the main effects is straightforward. However, as noted in Section 14-3, when interaction is significant, the main effects of the factors involved in the interaction may not have much practical interpretative value. Knowledge of the interaction is usually more important than knowledge about the main effects. Computational formulas for the sums of squares are easily obtained. Computing Formulas for ANOVA: Two Factors

Computing formulas for the sums of squares in a two-factor analysis of variance. a b n y2... 2 SST  a a a yijk  abn i1 j1 k1

(14-5)

2 a y 2# .. y... i SSA  a  abn i1 bn

(14-6)

b y.2j. y 2... SSB  a an  abn j1

(14-7)

2 a b y2ij. y...  SSA  SSB SSAB  a a n  abn i1 j1

SSE  SST  SSAB  SSA  SSB

(14-8) (14-9)

The computations are usually displayed in an ANOVA table, such as Table 14-4. EXAMPLE 14-1

Aircraft Primer Paint

Aircraft primer paints are applied to aluminum surfaces by two methods: dipping and spraying. The purpose of the primer is to improve paint adhesion, and some parts can be primed using either application method. The process engineering group responsible for this operation is interested in learning whether three different primers differ in their adhesion proper-

ties. A factorial experiment was performed to investigate the effect of paint primer type and application method on paint adhesion. For each combination of primer type and application method, three specimens were painted, then a finish paint was applied, and the adhesion force was measured. The data from the experiment are shown in Table 14-5. The circled numbers

Table 14-4 ANOVA Table for a Two-Factor Factorial, Fixed-Effects Model Source of Variation

Sum of Squares

Degrees of Freedom

Mean Square SSA MSA  a1

A treatments

SSA

a1

B treatments

SSB

b1

MSB 

Interaction

SSAB

(a  1)(b  1)

MSAB 

Error

SSE

ab(n  1)

Total

SST

abn  1

MSE 

SSB b1 SSAB 1a  121b  12

SSE ab1n  12

F0 MSA MSE MSB MSE MSAB MSE

JWCL232_c14_551-636.qxd

1/16/10

9:55 AM

Page 563

14-3 TWO-FACTOR FACTORIAL EXPERIMENTS

563

Table 14-5 Adhesion Force Data for Example 14-1 Primer Type

Dipping

1

4.0, 4.5, 4.3 12.8

5.4, 4.9, 5.6

15.9

28.7

2

5.6, 4.9, 5.4 15.9

5.8, 6.1, 6.3

18.2

34.1

3

3.8, 3.7, 4.0 11.5

5.5, 5.0, 5.0

15.5

27.0

y.j.

40.2

in the cells are the cell totals yij .. The sums of squares required to perform the ANOVA are computed as follows: 2 a b n y... 2 SST  a a a yijk  abn i1 j1 k1

 14.02 2  14.52 2  p  15.02  2

189.82 2 18

 10.72

2 a y 2.. y... i SStypes  a  abn i1 bn



128.72 2  134.12 2  127.02 2 

189.82 18

6 2

 4.58

2 b y .2j . y... SSmethods  a an  abn j1



140.22 2  149.62 2 9



189.82 2 18

 4.91

112.82 2  115.92 2  111.52 2  115.92 2  118.22 2  115.52 2 

189.82 2 18

49.6

yi..

89.8  y...

The ANOVA is summarized in Table 14-6. The experimenter has decided to use  0.05. Since f0.05,2,12  3.89 and f0.05,1,12  4.75, we conclude that the main effects of primer type and application method affect adhesion force. Furthermore, since 1.5  f0.05,2,12, there is no indication of interaction between these factors. The last column of Table 14-6 shows the P-value for each F-ratio. Notice that the P-values for the two test statistics for the main effects are considerably less than 0.05, while the P-value for the test statistic for the interaction is greater than 0.05. Practical Interpretation: A graph of the cell adhesion force averages 5 yij.6 versus levels of primer type for each application method is shown in Fig. 14-10. The no-interaction conclusion is obvious in this graph, because the two lines are nearly parallel. Furthermore, since a large response indicates greater adhesion force, we conclude that spraying is the best application method and that primer type 2 is most effective.

7.0

a b y2 ij. y 2... SSinteraction  a a n   SS types  SS methods abn i1 j1



Spraying

3  4.58  4.91  0.24

and SSE  SST  SStypes  SSmethods  SSinteraction

6.0

Spraying

yi j • 5.0

Dipping

4.0 3.0 1

2

3

Primer type

Figure 14-10 Graph of average adhesion force versus primer types for both application methods.

 10.72  4.58  4.91  0.24  0.99

Tests on Individual Means When both factors are fixed, comparisons between the individual means of either factor may be made using any multiple comparison technique such as Fisher’s LSD method (described in

JWCL232_c14_551-636.qxd

564

1/16/10

9:55 AM

Page 564

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

Table 14-6 ANOVA for Example 14-1 Source of Variation

Sum of Squares

Degrees of Freedom

Mean Square

Primer types Application methods Interaction Error Total

4.58 4.91 0.24 0.99 10.72

2 1 2 12 17

2.29 4.91 0.12 0.08

f0

P-Value

28.63 61.38 1.50

2.7  E-5 4.7  E-6 0.2621

Chapter 13). When there is no interaction, these comparisons may be made using either the row averages yi.. or the column averages y.j.. However, when interaction is significant, comparisons between the means of one factor (say, A) may be obscured by the AB interaction. In this case, we could apply a procedure such as Fisher’s LSD method to the means of factor A, with factor B set at a particular level. Minitab Output Table 14-7 shows some of the output from the Minitab analysis of variance procedure for the aircraft primer paint experiment in Example 14-1. The upper portion of the table gives factor name and level information, and the lower portion of the table presents the analysis of variance for the adhesion force response. The results are identical to the manual calculations displayed in Table 14-6 apart from rounding.

14-3.2 Model Adequacy Checking Just as in the single-factor experiments discussed in Chapter 13, the residuals from a factorial experiment play an important role in assessing model adequacy. The residuals from a two-factor factorial are eijk  yijk  yij. That is, the residuals are just the difference between the observations and the corresponding cell averages. Table 14-7 Analysis of Variance from Minitab for Example 14-1 ANOVA (Balanced Designs) Factor Primer Method

Type fixed fixed

Levels 3 2

Values 1 Dip

2 Spray

3

MS 2.2906 4.9089 0.1206 0.0822

F 27.86 59.70 1.47

Analysis of Variance for Adhesion Source Primer Method Primer *Method Error Total

DF 2 1 2 12 17

SS 4.5811 4.9089 0.2411 0.9867 10.7178

P 0.000 0.000 0.269

JWCL232_c14_551-636.qxd

1/16/10

9:55 AM

Page 565

14-3 TWO-FACTOR FACTORIAL EXPERIMENTS

565

Table 14-8 Residuals for the Aircraft Primer Experiment in Example 14-1 Application Method Primer Type

Dipping

Spraying

1 2 3

0.27, 0.23, 0.03 0.30, 0.40, 0.10 0.03, 0.13, 0.17

0.10, 0.40, 0.30 0.27, 0.03, 0.23 0.33, 0.17, 0.17

Table 14-8 presents the residuals for the aircraft primer paint data in Example 14-1. The normal probability plot of these residuals is shown in Fig. 14-11. This plot has tails that do not fall exactly along a straight line passing through the center of the plot, indicating some potential problems with the normality assumption, but the deviation from normality does not appear severe. Figures 14-12 and 14-13 plot the residuals versus the levels of primer types and application methods, respectively. There is some indication that primer type 3 results in slightly lower variability in adhesion force than the other two primers. The graph of residuals versus fitted values in Fig. 14-14 does not reveal any unusual or diagnostic pattern. 2.0

1.0 +0.5 zj

0.0 eij k

–1.0

–2.0 –0.5

–0.3

–0.1 +0.1 ei j k, residual

0

1

2

3

Primer type

+0.3

Figure 14-11 Normal probability plot of the residuals from Example 14-1.

–0.5

Figure 14-12 Plot of residuals versus primer type.

+0.5 +0.5

eijk

0

D

S

Application method

– 0.5

Figure 14-13 Plot of residuals versus application method.

ei j k

0

4

5

–0.5

Figure 14-14 Plot of residuals versus predicted values yˆijk.

6

^ yij k

JWCL232_c14_551-636.qxd

566

1/16/10

9:56 AM

Page 566

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

14-3.3 One Observation per Cell In some cases involving a two-factor factorial experiment, we may have only one replicate—that is, only one observation per cell. In this situation, there are exactly as many parameters in the analysis of variance model as observations, and the error degrees of freedom are zero. Thus, we cannot test hypotheses about the main effects and interactions unless some additional assumptions are made. One possible assumption is to assume the interaction effect is negligible and use the interaction mean square as an error mean square. Thus, the analysis is equivalent to the analysis used in the randomized block design. This no-interaction assumption can be dangerous, and the experimenter should carefully examine the data and the residuals for indications as to whether or not interaction is present. For more details, see Montgomery (2009). EXERCISES FOR SECTION 14-3 14-1. An article in Industrial Quality Control (1956, pp. 5–8) describes an experiment to investigate the effect of two factors (glass type and phosphor type) on the brightness of a television tube. The response variable measured is the current (in microamps) necessary to obtain a specified brightness level. The data are shown in the following table: (a) State the hypotheses of interest in this experiment. (b) Test the above hypotheses and draw conclusions using the analysis of variance with  0.05. (c) Analyze the residuals from this experiment.

Glass Type 1

2

Phosphor Type 1

2

3

280 290 285

300 310 295

290 285 290

230 235 240

260 240 235

220 225 230

14-2. An engineer suspects that the surface finish of metal parts is influenced by the type of paint used and the drying time. He selected three drying times—20, 25, and 30 minutes—and used two types of paint. Three parts are tested with each combination of paint type and drying time. The data are as follows: Drying Time (min) Paint

20

25

30

1

74 64 50

73 61 44

78 85 92

2

92 86 68

98 73 88

66 45 85

(a) State the hypotheses of interest in this experiment. (b) Test the above hypotheses and draw conclusions using the analysis of variance with = 0.05. (c) Analyze the residuals from this experiment. 14-3. In the book Design and Analysis of Experiments, 7th edition (2009, John Wiley & Sons), the results of an experiment involving a storage battery used in the launching mechanism of a shoulder-fired ground-to-air missile were presented. Three material types can be used to make the battery plates. The objective is to design a battery that is relatively unaffected by the ambient temperature. The output response from the battery is effective life in hours. Three temperature levels are selected, and a factorial experiment with four replicates is run. The data are as follows: Temperature (ⴗF) Material

Low

Medium

High

1

130 74

155 180

34 80

40 75

20 82

70 58

2

150 159

188 126

136 106

122 115

25 58

70 45

3

138 168

110 160

174 150

120 139

96 82

104 60

(a) Test the appropriate hypotheses and draw conclusions using the analysis of variance with = 0.05. (b) Graphically analyze the interaction. (c) Analyze the residuals from this experiment. 14-4. An experiment was conducted to determine whether either firing temperature or furnace position affects the baked density of a carbon anode. The data are as follows: Temperature (ⴗC) Position 1

800

825

850

570 565

1063 1080

565 510

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 567

567

14-3 TWO-FACTOR FACTORIAL EXPERIMENTS

2

583

1043

590

528 547 521

988 1026 1004

526 538 532

(a) State the hypotheses of interest. (b) Test the above hypotheses using the analysis of variance with = 0.05. What are your conclusions? (c) Analyze the residuals from this experiment. (d) Using Fisher’s LSD method, investigate the differences between the mean baked anode density at the three different levels of temperature. Use = 0.05. 14-5. An article in Technometrics [“Exact Analysis of Means with Unequal Variances” (2002, Vol. 44, pp. 152–160)] described the technique of the analysis of means (ANOM) and presented the results of an experiment on insulation. Four insulation types were tested at three different temperatures. The data are as follows:

Temperature (F) Insulation

1

2

3

1

6.6 2.7 6

4 6.2 5

4.5 5.5 4.8

2.2 2.7 5.8

2.3 5.6 2.2

0.9 4.9 3.4

2

3 2.1 5.9

3.2 4.1 2.5

3 2.5 0.4

1.5 2.6 3.5

1.3 0.5 1.7

3.3 1.1 0.1

3

5.7 3.2 5.3

4.4 3.2 9.7

8.9 7 8

7.7 7.3 2.2

2.6 11.5 3.4

9.9 10.5 6.7

4

7 7.3 8.6

8.9 9 11.3

12 8.5 7.9

9.7 10.8 7.3

8.3 10.4 10.6

8 9.7 7.4

(a) Write down a model for this experiment. (b) Test the appropriate hypotheses and draw conclusions using the analysis of variance with  0.05. (c) Graphically analyze the interaction. (d) Analyze the residuals from the experiment. (e) Use Fisher’s LSD method to investigate the differences between mean effects of insulation type. Use  0.05. 14-6. Johnson and Leone (Statistics and Experimental Design in Engineering and the Physical Sciences, John Wiley, 1977) described an experiment conducted to investigate warping of copper plates. The two factors studied were temperature and the copper content of the plates. The response variable is the amount of warping. The data are as follows:

Copper Content (%)

Temperature (ⴗC)

40

60

80

100

50 75 100 125

17, 20 12, 9 16, 12 21, 17

16, 21 18, 13 18, 21 23, 21

24, 22 17, 12 25, 23 23, 22

28, 27 27, 31 30, 23 29, 31

(a) Is there any indication that either factor affects the amount of warping? Is there any interaction between the factors? Use = 0.05. (b) Analyze the residuals from this experiment. (c) Plot the average warping at each level of copper content and compare the levels using Fisher’s LSD method. Describe the differences in the effects of the different levels of copper content on warping. If low warping is desirable, what level of copper content would you specify? (d) Suppose that temperature cannot be easily controlled in the environment in which the copper plates are to be used. Does this change your answer for part (c)? 14-7. An article in the IEEE Transactions on Electron Devices (November 1986, p. 1754) describes a study on the effects of two variables—polysilicon doping and anneal conditions (time and temperature)—on the base current of a bipolar transistor. The data from this experiment follows below. (a) Is there any evidence to support the claim that either polysilicon doping level or anneal conditions affect base current? Do these variables interact? Use  0.05. (b) Graphically analyze the interaction. (c) Analyze the residuals from this experiment. (d) Use Fisher’s LSD method to isolate the effects of anneal conditions on base current, with  0.05. 14-8. An article in the Journal of Testing and Evaluation (1988, Vol. 16, pp. 508–515) investigated the effects of cyclic loading frequency and environment conditions on fatigue crack growth at a constant 22 MPa stress for a particular material.

Air

Environment Salt H2O H2O

10

2.29 2.47 2.48 2.12

2.06 2.05 2.23 2.03

1.90 1.93 1.75 2.06

Frequency 1

2.65 2.68 2.06 2.38

3.20 3.18 3.96 3.64

3.10 3.24 3.98 3.24

0.1

2.24 2.71 2.81 2.08

11.00 11.00 9.06 11.30

9.96 10.01 9.36 10.40

JWCL232_c14_551-636.qxd

568

1/16/10

9:56 AM

Page 568

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

The data from the experiment follow. The response variable is fatigue crack growth rate. (a) Is there indication that either factor affects crack growth rate? Is there any indication of interaction? Use  0.05. (b) Analyze the residuals from this experiment. (c) Repeat the analysis in part (a) using ln(y) as the response. Analyze the residuals from this new response variable and comment on the results.

14-9. Consider a two-factor factorial experiment. Develop a formula for finding a 100(1  )% confidence interval on the difference between any two means for either a row or column factor. Apply this formula to find a 95% CI on the difference in mean warping at the levels of copper content 60 and 80% in Exercise 14-6.

Anneal (temperature/time) 900 950 1000 180 60 15

900 60 Polysilicon doping

1000 30

1  1020

4.40 4.60

8.30 8.90

10.15 10.20

10.29 10.30

11.01 10.58

2  1020

3.20 3.50

7.81 7.75

9.38 10.02

10.19 10.10

10.81 10.60

14-4 GENERAL FACTORIAL EXPERIMENTS Many experiments involve more than two factors. In this section we introduce the case where there are a levels of factor A, b levels of factor B, c levels of factor C, and so on, arranged in a factorial experiment. In general, there will be abc p n total observations, if there are n replicates of the complete experiment. For example, consider the three-factor-factorial experiment, with underlying model Yijkl    i  j  k  12 ij  12 ik  12 jk  12 ijk  ijkl

i  1, 2, p , a j  1, 2, p , b μ k  1, 2, p , c l  1, 2, p , n

(14-10)

Notice that the model contains three main effects, three two-factor interactions, a three-factor interaction, and an error term. Assuming that A, B, and C are fixed factors, the analysis of variance is shown in Table 14-9. Note that there must be at least two replicates (n  2) to compute an error sum of squares. The F-test on main effects and interactions follows directly from the expected mean squares. These ratios follow F distributions under the respective null hypotheses.

EXAMPLE 14-2

Surface Roughness

A mechanical engineer is studying the surface roughness of a part produced in a metal-cutting operation. Three factors, feed rate (A), depth of cut (B), and tool angle (C ), are of interest. All three factors have been assigned two levels, and two replicates of a factorial design are run. The coded data are shown in Table 14-10. The ANOVA is summarized in Table 14-11. Since manual ANOVA computations are tedious for three-factor ex-

periments, we have used Minitab for the solution of this problem. The F-ratios for all three main effects and the interactions are formed by dividing the mean square for the effect of interest by the error mean square. Since the experimenter has selected  0.05, the critical value for each of these F-ratios is f0.05,1,8  5.32. Alternately, we could use the Pvalue approach. The P-values for all the test statistics are shown in the last column of Table 14-11. Inspection of these

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 569

14-4 GENERAL FACTORIAL EXPERIMENTS

569

Table 14-9 Analysis of Variance Table for the Three-Factor Fixed Effects Model Source of Variation

Sum of Squares

Degrees of Freedom

Expected Mean Squares

Mean Square

A

SSA

a1

MSA

2 

B

SSB

b1

MSB

2 

C

SSC

c1

MSC

2 

AB

SSAB

1a  121b  12

MSAB

2 

AC

SSAC

1a  121c  12

MSAC

2 

BC

SSBC

1b  121c  12

MSBC

2 

ABC

SSABC

1a  121b  121c  12

MSABC

2 

MSE

2

Error Total

abc1n  12 abcn  1

SSE SST

P-values is revealing. There is a strong main effect of feed rate, since the F-ratio is well into the critical region. However, there is some indication of an effect due to the depth of cut, since P = 0.0710 is not much greater than  0.05. The next largest effect is the AB or feed rate  depth of cut interaction.

F0

bcn 兺2i

MSA MSE

a1 acn 兺2j b1

MSB MSE

abn 兺2k c1

MSC MSE

cn 兺 兺12 2ij

MSAB MSE

bn 兺 兺12 2ik

MSAC MSE

an 兺 兺12 jk2

MSBC MSE

1a  121b  12 1a  121c  12 1b  121c  12 n 兺兺兺12 2ijk

MSABC MSE

1a  121b  121c  12

Most likely, both feed rate and depth of cut are important process variables. Practical Interpretation: Further experiments might study the important factors in more detail to improve the surface roughness.

Obviously, factorial experiments with three or more factors can require many runs, particularly if some of the factors have several (more than two) levels. This point of view leads us to the class of factorial designs considered in Section 14-5 with all factors at two levels. These designs are easy to set up and analyze, and they may be used as the basis of many other useful experimental designs. Table 14-10

Coded Surface Roughness Data for Example 14-2 Depth of Cut (B) 0.025 inch 0.040 inch Tool Angle (C)

Tool Angle (C )

15

25

15

25

yi p

20 inches per minute

9 7

11 10

9 11

10 8

75

30 inches per minute

10 12

10 13

12 15

16 14

102

Feed Rate (A)

JWCL232_c14_551-636.qxd

570

1/16/10

9:56 AM

Page 570

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

Table 14-11

Minitab ANOVA for Example 14-2

ANOVA (Balanced Designs) Factor Feed Depth Angle

Type fixed fixed fixed

Levels 2 2 2

Values 20 0.025 15

30 0.040 25

Analysis of Variance for Roughness Source Feed Depth Angle Feed*Depth Feed*Angle Depth*Angle Feed*Depth*Angle Error Total

DF 1 1 1 1 1 1 1 8 15

SS 45.563 10.563 3.063 7.563 0.062 1.563 5.062 19.500 92.938

MS 45.563 10.563 3.063 7.563 0.062 1.563 5.062 2.437

F 18.69 4.33 1.26 3.10 0.03 0.64 2.08

P 0.003 0.071 0.295 0.116 0.877 0.446 0.188

EXERCISES FOR SECTION 14-4 14-10. The quality control department of a fabric finishing plant is studying the effects of several factors on dyeing for a blended cotton/synthetic cloth used to manufacture shirts. Three operators, three cycle times, and two temperatures were selected, and three small specimens of cloth were dyed under each set of conditions. The finished cloth was compared to a standard, and a numerical score was assigned. The results are shown in the following table.

(a) State and test the appropriate hypotheses using the analysis of variance with  0.05. (b) The residuals may be obtained from eijkl  yijkl  yijk .. Graphically analyze the residuals from this experiment. 14-11. The percentage of hardwood concentration in raw pulp, the freeness, and the cooking time of the pulp are being investigated for their effects on the strength of paper. The data

Temperature

Cycle Time

1

300ⴗ Operator 2

3

1

350ⴗ Operator 2

40

23 24 25

27 28 26

31 32 28

24 23 28

38 36 35

34 36 39

50

36 35 36

34 38 39

33 34 35

37 39 35

34 38 36

34 36 31

60

28 24 27

35 35 34

26 27 25

26 29 25

36 37 34

28 26 34

3

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 571

14-5 2k FACTORIAL DESIGNS

from a three-factor factorial experiment are shown in the following table. (a) Analyze the data using the analysis of variance assuming that all factors are fixed. Use  0.05.

Percentage of Hardwood Concentration

571

(b) Find P-values for the F-ratios in part (a). (c) The residuals are found by eijkl  yijkl  yijk.. Graphically analyze the residuals from this experiment.

Cooking Time 1.5 hours

Cooking Time 2.0 hours

Freeness

Freeness

350

500

650

350

500

650

10

96.6 96.0

97.7 96.0

99.4 99.8

98.4 98.6

99.6 100.4

100.6 100.9

15

98.5 97.2

96.0 96.9

98.4 97.6

97.5 98.1

98.7 96.0

99.6 99.0

20

97.5 96.6

95.6 96.2

97.4 98.1

97.6 98.4

97.0 97.8

98.5 99.8

14-5 2k FACTORIAL DESIGNS Factorial designs are frequently used in experiments involving several factors where it is necessary to study the joint effect of the factors on a response. However, several special cases of the general factorial design are important because they are widely employed in research work and because they form the basis of other designs of considerable practical value. The most important of these special cases is that of k factors, each at only two levels. These levels may be quantitative, such as two values of temperature, pressure, or time; or they may be qualitative, such as two machines, two operators, the “high’’ and “low’’ levels of a factor, or perhaps the presence and absence of a factor. A complete replicate of such a design requires 2  2      2  2k observations and is called a 2k factorial design. The 2k design is particularly useful in the early stages of experimental work, when many factors are likely to be investigated. It provides the smallest number of runs for which k factors can be studied in a complete factorial design. Because there are only two levels for each factor, we must assume that the response is approximately linear over the range of the factor levels chosen.

14-5.1 22 Design The simplest type of 2k design is the 22—that is, two factors A and B, each at two levels. We usually think of these levels as the low and high levels of the factor. The 22 design is shown in Fig. 14-15. Note that the design can be represented geometrically as a square with the 22  4 runs, or treatment combinations, forming the corners of the square. In the 22 design it is customary to denote the low and high levels of the factors A and B by the signs  and , respectively. This is sometimes called the geometric notation for the design. A special notation is used to label the treatment combinations. In general, a treatment combination is represented by a series of lowercase letters. If a letter is present, the corresponding factor is run at the high level in that treatment combination; if it is absent, the factor is run at its low level. For example, treatment combination a indicates that factor A is at the

JWCL232_c14_551-636.qxd

572

1/16/10

9:56 AM

Page 572

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS High (+)

b

ab

Treatment A – (1) + a – b + ab

B

Low (1) (–) Low (–)

B – – + +

a A

High (+)

Figure 14-15 The 22 factorial design.

high level and factor B is at the low level. The treatment combination with both factors at the low level is represented by (1). This notation is used throughout the 2k design series. For example, the treatment combination in a 24 with A and C at the high level and B and D at the low level is denoted by ac. The effects of interest in the 22 design are the main effects A and B and the two-factor interaction AB. Let the letters (1), a, b, and ab also represent the totals of all n observations taken at these design points. It is easy to estimate the effects of these factors. To estimate the main effect of A, we would average the observations on the right side of the square in Fig. 14-15 where A is at the high level, and subtract from this the average of the observations on the left side of the square, where A is at the low level, or Main Effect of Factor A: 22 Design

A  yA  yA 

b  112 a  ab 1   3a  ab  b  112 4 2n 2n 2n

(14-11)

Similarly, the main effect of B is found by averaging the observations on the top of the square, where B is at the high level, and subtracting the average of the observations on the bottom of the square, where B is at the low level: Main Effect of Factor B: 22 Design

B  yB  yB 

a  112 b  ab 1   3b  ab  a  112 4 2n 2n 2n

(14-12)

Finally, the AB interaction is estimated by taking the difference in the diagonal averages in Fig. 14-15, or Interaction Effect AB: 22 Design

AB 

ab  112 1 ab  3ab  112  a  b 4  2n 2n 2n

(14-13)

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 573

14-5 2k FACTORIAL DESIGNS

573

Table 14-12 Signs for Effects in the 22 Design Treatment Combination 112 a b ab

Factorial Effect I

A

B

AB

   

   

   

   

The quantities in brackets in Equations 14-11, 14-12, and 14-13 are called contrasts. For example, the A contrast is ContrastA  a  ab  b  112 In these equations, the contrast coefficients are always either 1 or 1. A table of plus and minus signs, such as Table 14-12, can be used to determine the sign on each treatment combination for a particular contrast. The column headings for Table 14-12 are the main effects A and B, the AB interaction, and I, which represents the total. The row headings are the treatment combinations. Note that the signs in the AB column are the product of signs from columns A and B. To generate a contrast from this table, multiply the signs in the appropriate column of Table 14-12 by the treatment combinations listed in the rows and add. For example, contrastAB  31124  3a4  3b4  3ab4  ab  112  a  b. Contrasts are used in calculating both the effect estimates and the sums of squares for A, B, and the AB interaction. For any 2k design with n replicates, the effect estimates are computed from Relationship Between a Contrast and an Effect

Contrast n2k1

(14-14)

1Contrast2 2 n2k

(14-15)

Effect 

and the sum of squares for any effect is

Sum of Squares for an Effect

SS 

There is one degree of freedom associated with each effect (two levels minus one) so that the mean square error of each effect equals the sum of squares. The analysis of variance is completed by computing the total sum of squares SST (with 4n  1 degrees of freedom) as usual, and obtaining the error sum of squares SSE (with 4(n  1) degrees of freedom) by subtraction.

JWCL232_c14_551-636.qxd

574

1/16/10

9:56 AM

Page 574

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

EXAMPLE 14-3

Wafer Layer

An article in the AT&T Technical Journal (Vol. 65, March /April 1986, pp. 39–50) describes the application of two-level factorial designs to integrated circuit manufacturing. A basic processing step in this industry is to grow an epitaxial layer on polished silicon wafers. The wafers are mounted on a susceptor and positioned inside a bell jar. Chemical vapors are introduced through nozzles near the top of the jar. The susceptor is rotated, and heat is applied. These conditions are maintained until the epitaxial layer is thick enough. Table 14-13 presents the results of a 22 factorial design with n  4 replicates using the factors A  deposition time and B  arsenic flow rate. The two levels of deposition time are  short and  long, and the two levels of arsenic flow rate are  55% and  59%. The response variable is epitaxial layer thickness ( m). We may find the estimates of the effects using Equations 14-11, 14-12, and 14-13 as follows: A

The numerical estimates of the effects indicate that the effect of deposition time is large and has a positive direction (increasing deposition time increases thickness), since changing deposition time from low to high changes the mean epitaxial layer thickness by 0.836 m. The effects of arsenic flow rate (B) and the AB interaction appear small. The importance of these effects may be confirmed with the analysis of variance. The sums of squares for A, B, and AB are computed as follows:

1 3a  ab  b  112 4 2n

3a  ab  b  112 4 2

SSB 

3b  ab  a  112 4 2

SSAB 

3ab  112  a  b4 2

16

16

16



36.688 4 2



30.538 4 2



30.252 4 2

16

16

16

 2.7956  0.0181

 0.0040

SST  14.0372  p  14.9322

1  359.299  59.156  55.686  56.0814  0.836 2142 B

SSA 



1 3b  ab  a  112 4 2n

156.081  p  59.1562 2 16

 3.0672

1  355.686  59.156  59.299  56.0814 2142

Practical Interpretation: The analysis of variance is summarized in Table 14-14 and confirms our conclusions obtained by examining the magnitude and direction of the effects. Deposition time is the only factor that significantly affects epitaxial layer thickness, and from the direction of the effect estimates we know that longer deposition times lead to thicker epitaxial layers.

 0.067 AB 

1 3ab  112  a  b 4 2n

AB 

1 359.156  56.081  59.299  55.6864 2142

 0.032

Models and Residual Analysis It is easy to obtain a model for the response and residuals from a 2k design by fitting a regression model to the data. For the epitaxial process experiment, the regression model is Y  0  1 x1  Table 14-13 Treatment Combination 112 a b ab

The 22 Design for the Epitaxial Process Experiment Design Factors A B AB            

14.037 14.821 13.880 14.888

Thickness (m) 14.165 13.972 14.757 14.843 13.860 14.032 14.921 14.415

13.907 14.878 13.914 14.932

Thickness (␮m) Total Average 56.081 14.020 59.299 14.825 55.686 13.922 59.156 14.789

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 575

575

14-5 2k FACTORIAL DESIGNS

Table 14-14

Analysis of Variance for the Epitaxial Process Experiment

Source of Variation

Sum of Squares

Degrees of Freedom

A (deposition time) B (arsenic flow) AB Error Total

2.7956 0.0181 0.0040 0.2495 3.0672

1 1 1 12 15

Mean Square

f0

P-Value

2.7956 0.0181 0.0040 0.0208

134.40 0.87 0.19

7.07 E-8 0.38 0.67

since the only active variable is deposition time, which is represented by a coded variable x1. The low and high levels of deposition time are assigned values x1  1 and x1  1, respectively. The least squares fitted model is yˆ  14.389  a

0.836 b x1 2

where the intercept ˆ 0 is the grand average of all 16 observations ( y ) and the slope ˆ 1 is onehalf the effect estimate for deposition time. The regression coefficient is one-half the effect estimate because regression coefficients measure the effect of a unit change in x1 on the mean of Y, and the effect estimate is based on a two-unit change from 1 to 1. A coefficient relates a factor to the response and, similar to regression analysis, interest centers on whether or not a coefficient estimate is significantly different from zero. A t-test for a coefficient can also be used to test the significance of an effect. Each effect estimate in Equations 14-11 through 14-13 is the difference between two averages (that we denote in general as y   y ). In a 2k experiment with n replicates, half the observations appear in each average so that there are n2k1 observations in each average. The associated coefficient estimate, say ˆ , equals half the associated effect estimate so that Coefficient and Effect

y   y effect  2 2

ˆ 

(14-16)

The standard error of ˆ equals half the standard error of the effect and an effect is simply the difference between two averages. Therefore, Standard Error of a Coefficient

standard error ˆ 

1

ˆ 1 1  k1  ˆ 2 B n2k1 n2 B n2k

(14-17)

where ˆ is estimated from the square root of mean square error. The t-statistic to test H0:   0 in a 2k experiment is t-statistic for a Coefficient

t

ˆ

 standard error ˆ

1 y   y  2 2

ˆ

1 B n2k

(14-18)

JWCL232_c14_551-636.qxd

9:56 AM

Page 576

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

with degrees of freedom equal to those associated with mean square error. This statistic is similar to a two-sample t-test, but is estimated from root mean square error. The estimate ˆ accounts for the multiple treatments in an experiment and generally differs from the estimate used in a two-sample t-test. Some algebra can be used to show that for a 2k experiment the square of the t-statistic for the coefficient test equals the F-statistic used for the effect test in the analysis of variance. Also, the square of a t random variable with d degrees of freedom is an F random variable with 1 numerator and d denominator degrees of freedom. Thus, the test that compares the absolute value of the t-statistic to the t distribution is equivalent to the F-test, and either method may be used to test an effect. This model can also be used to obtain the predicted values at the four points that form the corners of the square in the design. For example, consider the point with low deposition time (x1  1) and low arsenic flow rate. The predicted value is yˆ  14.389  a

0.836 b 112  13.971 m 2

and the residuals for the four runs at that design point are e1  14.037  13.971  0.066 e2  14.165  13.971  0.194 e3  13.972  13.971  0.001 e4  13.907  13.971  0.064 The remaining predicted values and residuals at the other three design points are calculated in a similar manner. A normal probability plot of these residuals is shown in Fig. 14-16. This plot indicates that one residual e15  0.392 is an outlier. Examining the four runs with high deposition time and high arsenic flow rate reveals that observation y15  14.415 is considerably smaller 99 95 90 80 70 Normal probability

576

1/16/10

60 30 20 10 8 1

–0.392 –0.294 –0.197 –0.099

–0.001

0.096

0.194

Residual

Figure 14-16 Normal probability plot of residuals for the epitaxial process experiment.

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 577

14-5 2k FACTORIAL DESIGNS 0.5

577

0.5

e

e 0

Low

High

Deposition time, A

–0.5

0

Low

High

Arsenic flow rate, B

–0.5

Figure 14-17 Plot of residuals versus deposition time.

Figure 14-18 Plot of residuals versus arsenic flow rate.

than the other three observations at that treatment combination. This adds some additional evidence to the tentative conclusion that observation 15 is an outlier. Another possibility is that some process variables affect the variability in epitaxial layer thickness. If we could discover which variables produce this effect, we could perhaps adjust these variables to levels that would minimize the variability in epitaxial layer thickness. This could have important implications in subsequent manufacturing stages. Figures 14-17 and 14-18 are plots of residuals versus deposition time and arsenic flow rate, respectively. Apart from that unusually large residual associated with y15, there is no strong evidence that either deposition time or arsenic flow rate influences the variability in epitaxial layer thickness. Figure 14-19 shows the standard deviation of epitaxial layer thickness at all four runs in the 22 design. These standard deviations were calculated using the data in Table 14-13. Notice that the standard deviation of the four observations with A and B at the high level is considerably larger than the standard deviations at any of the other three design points. Most of this difference is attributable to the unusually low thickness measurement associated with y15. The standard deviation of the four observations with A and B at the low level is also somewhat larger than the standard deviations at the remaining two runs. This could indicate that other process variables not included in this experiment may affect the variability in epitaxial layer thickness. Another experiment to study this possibility, involving other process variables, could be designed and conducted. (The original paper in the AT&T Technical Journal shows that two additional factors, not considered in this example, affect process variability.)

14-5.2 2k Design for k ⱖ 3 Factors The methods presented in the previous section for factorial designs with k  2 factors each at two levels can be easily extended to more than two factors. For example, consider k  3 factors, each at two levels. This design is a 23 factorial design, and it has eight runs or treatment combinations. Geometrically, the design is a cube as shown in Fig. 14-20(a), with the eight runs forming the corners of the cube. Figure 14-20(b) lists the eight runs in a table, with each row representing one of the runs and the  and  settings indicating the low and high levels for each of the three factors. This table is sometimes called the design matrix. This design allows three main effects to be estimated (A, B, and C) along with three two-factor interactions (AB, AC, and BC ) and a three-factor interaction (ABC). The main effects can easily be estimated. Remember that the lowercase letters (1), a, b, ab, c, ac, bc, and abc represent the total of all n replicates at each of the eight runs in the design. As seen in Fig. 14-21(a), the main effect of A can be estimated by averaging the four

JWCL232_c14_551-636.qxd

578

1/16/10

9:56 AM

Page 578

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

+

0.077

bc

0.250

b

ab +

abc

c

C

B

ab

b –

–

a

(1) 0.110 –

A

0.051 +

Figure 14-19 The standard deviation of epitaxial layer thickness at the four runs in the 22 design.

(1)

+

a – –

A

B

+

+

Figure 14-20 The 23 design.

+ – –

A

B Main effects

C

(a)

+

–

+

+

+

–

– –

– +

AB

AC

BC

Two-factor interactions (b)

C

= + runs = – runs

B

B

C

1

–

–

–

2

+

–

–

3

–

+

–

4

+

+

–

5

–

–

+

6

+

–

+

7

–

+

+

8

+

+

+

(b) The 23 design matrix

(a) Geometric view

+ –

A

Run

ac

A ABC Three-factor interaction (c)

Figure 14-21 Geometric presentation of contrasts corresponding to the main effects and interaction in the 23 design. (a) Main effects. (b) Two-factor interactions. (c) Three-factor interaction.

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 579

14-5 2k FACTORIAL DESIGNS

579

treatment combinations on the right-hand side of the cube, where A is at the high level, and by subtracting from this quantity the average of the four treatment combinations on the left-hand side of the cube where A is at the low level. This gives A  yA  yA 112  b  c  bc a  ab  ac  abc   4n 4n This equation can be rearranged as Main Effect of Factor A: 23 Design

A  yA  yA 1  3a  ab  ac  abc  112  b  c  bc4 4n

In a similar manner, the effect of B is the difference in averages between the four treatment combinations in the back face of the cube [Fig. 14-19(a)], and the four in the front. This yields Main Effect of Factor B: 23 Design

B  yB  yB 1  3b  ab  bc  abc  112  a  c  ac4 4n

The effect of C is the difference in average response between the four treatment combinations in the top face of the cube in Figure 14-19(a) and the four in the bottom, that is, Main Effect of Factor C: 23 Design

C  yC  yC 1  3c  ac  bc  abc  112  a  b  ab4 4n

The two-factor interaction effects may be computed easily. A measure of the AB interaction is the difference between the average A effects at the two levels of B. By convention, one-half of this difference is called the AB interaction. Symbolically, B High () Low () Difference

Average A Effect 3 1abc  bc2  1ab  b2 4 2n 51ac  c2  3a  112 4 6 2n 3abc  bc  ab  b  ac  c  a  112 4 2n

JWCL232_c14_551-636.qxd

580

1/16/10

9:56 AM

Page 580

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

Because the AB interaction is one-half of this difference, Two-Factor Interaction Effect: 23 Design

AB 

1 3abc  bc  ab  b  ac  c  a  112 4 4n

We could write the AB effect as follows: AB 

abc  ab  c  112 bc  b  ac  a  4n 4n

In this form, the AB interaction is easily seen to be the difference in averages between runs on two diagonal planes in the cube in Fig. 14-19(b). Using similar logic and referring to Fig. 14-19(b), we find that the AC and BC interactions are

Two-Factor Interaction Effect: 23 Design

1 3 112  a  b  ab  c  ac  bc  abc4 4n 1 BC  3 112  a  b  ab  c  ac  bc  abc4 4n AC 

The ABC interaction is defined as the average difference between the AB interaction for the two different levels of C. Thus, ABC 

1 5 3abc  bc4  3ac  c 4  3ab  b 4  3a  112 4 6 4n

or Three-Factor Interaction Effect: 23 Design

ABC 

1 3abc  bc  ac  c  ab  b  a  112 4 4n

As before, we can think of the ABC interaction as the difference in two averages. If the runs in the two averages are isolated, they define the vertices of the two tetrahedra that comprise the cube in Fig. 14-21(c). In the equations for the effects, the quantities in brackets are contrasts in the treatment combinations. A table of plus and minus signs can be developed from the contrasts and is shown in Table 14-15. Signs for the main effects are determined directly from the test matrix in Figure 14-20(b). Once the signs for the main effect columns have been established, the signs for the remaining columns can be obtained by multiplying the appropriate main effect

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 581

14-5 2k FACTORIAL DESIGNS

Table 14-15

581

Algebraic Signs for Calculating Effects in the 23 Design Factorial Effect

Treatment Combination

I

A

B

AB

C

AC

BC

ABC

a b ab c ac bc abc

       

       

       

       

       

       

       

       

112

row by row. For example, the signs in the AB column are the products of the A and B column signs in each row. The contrast for any effect can easily be obtained from this table. Table 14-15 has several interesting properties: 1.

Except for the identity column I, each column has an equal number of plus and minus signs. 2. The sum of products of signs in any two columns is zero; that is, the columns in the table are orthogonal. 3. Multiplying any column by column I leaves the column unchanged; that is, I is an identity element. 4. The product of any two columns yields a column in the table, for example A  B  AB, and AB  ABC  A2B2C  C, since any column multiplied by itself is the identity column. The estimate of any main effect or interaction in a 2k design is determined by multiplying the treatment combinations in the first column of the table by the signs in the corresponding main effect or interaction column, by adding the result to produce a contrast, and then by dividing the contrast by one-half the total number of runs in the experiment.

EXAMPLE 14-4

Surface Roughness

Consider the surface roughness experiment originally described in Example 14-2. This is a 23 factorial design in the factors feed rate (A), depth of cut (B), and tool angle (C ), with n  2 replicates. Table 14-16 presents the observed surface roughness data. The effect of A, for example, is

A

1 3a  ab  ac  abc  112  b  c  bc4 4n



1 322  27  23  30  16  20  21  184 4122



1 327 4  3.375 8

and the sum of squares for A is found using Equation 14-15:

SSA 

1ContrastA 2 2 n2k



1272 2 2182

 45.5625

It is easy to verify that the other effects are B C AB AC BC ABC

 1.625  0.875  1.375  0.125  0.625  1.125

Examining the magnitude of the effects clearly shows that feed rate (factor A) is dominant, followed by depth of cut (B)

JWCL232_c14_551-636.qxd

582

1/16/10

9:56 AM

Page 582

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

Table 14-16

Surface Roughness Data for Example 14-4

Treatment Combinations 112 a b ab c ac bc abc

Design Factors A

B

C

Surface Roughness

Totals

1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1

9, 7 10, 12 9, 11 12, 15 11, 10 10, 13 10, 8 16, 14

16 22 20 27 21 23 18 30

and the AB interaction, although the interaction effect is relatively small. The analysis of variance, summarized in Table 14-17, confirms our interpretation of the effect estimates. Minitab will analyze 2k factorial designs. The output from the Minitab DOE (Design of Experiments) module for this experiment is shown in Table 14-18. The upper portion of the table displays the effect estimates and regression coefficients for each factorial effect. However, the t-statistic computed from Equation 14-18 is reported for each effect instead of the F-statistic used in Table 14-17. To illustrate, for the main effect of feed Minitab reports t  4.32 (with eight degrees of freedom), and t 2  (4.32)2  18.66, which is approximately equal to the F-ratio for feed reported in Table 14-17 (F  18.69). This F-ratio has one numerator and eight denominator degrees of freedom. The lower panel of the Minitab output in Table 14-18 is an analysis of variance summary focusing on the types of terms in the model. A regression model approach is used in the presentation. You might find it helpful to review Section 12-2.2, particularly the material on the partial F-test. The row entitled “main effects’’ under source refers to the three main effects feed, depth, and angle, each having a single degree of

Table 14-17 Source of Variation A B C AB AC BC ABC Error Total

freedom, giving the total 3 in the column headed “DF.’’ The column headed “Seq SS’’ (an abbreviation for sequential sum of squares) reports how much the model sum of squares increases when each group of terms is added to a model that contains the terms listed above the groups. The first number in the “Seq SS’’ column presents the model sum of squares for fitting a model having only the three main effects. The row labeled “2-Way Interactions’’ refers to AB, AC, and BC, and the sequential sum of squares reported here is the increase in the model sum of squares if the interaction terms are added to a model containing only the main effects. Similarly, the sequential sum of squares for the three-way interaction is the increase in the model sum of squares that results from adding the term ABC to a model containing all other effects. The column headed “Adj SS’’ (an abbreviation for adjusted sum of squares) reports how much the model sum of squares increases when each group of terms is added to a model that contains all the other terms. Now since any 2k design with an equal number of replicates in each cell is an orthogonal design, the adjusted sum of squares will equal the sequential sum of squares. Therefore, the F-tests for each row in the Minitab analysis of variance table are testing the significance

Analysis of Variance for the Surface Finish Experiment Sum of Squares

Degrees of Freedom

45.5625 10.5625 3.0625 7.5625 0.0625 1.5625 5.0625 19.5000 92.9375

1 1 1 1 1 1 1 8 15

Mean Square

f0

P-Value

45.5625 10.5625 3.0625 7.5625 0.0625 1.5625 5.0625 2.4375

18.69 4.33 1.26 3.10 0.03 0.64 2.08

0.0025 0.0709 0.2948 0.1162 0.8784 0.4548 0.1875

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Table 14-18

Page 583

14-5 2k FACTORIAL DESIGNS

583

StDev Coef 0.3903 0.3903 0.3903 0.3903 0.3903 0.3903 0.3903 0.3903

T 28.34 4.32 2.08 1.12 1.76 0.16 0.80 1.44

P 0.000 0.003 0.071 0.295 0.116 0.877 0.446 0.188

F 8.09 1.26 2.08

P 0.008 0.352 0.188

Minitab Analysis for Example 14-4

Estimated Effects and Coefficients for Roughness Term Constant Feed Depth Angle Feed*Depth Feed*Angle Depth*Angle Feed*Depth*Angle

Effect 3.3750 1.6250 0.8750 1.3750 0.1250 0.6250 1.1250

Coef 11.0625 1.6875 0.8125 0.4375 0.6875 0.0625 0.3125 0.5625

Analysis of Variance for Roughness Source Main Effects 2-Way Interactions 3-Way Interactions Residual Error Pure Error Total

DF 3 3 1 8 8 15

of each group of terms (main effects, two-factor interactions, and three-factor interactions) as if they were the last terms to be included in the model. Clearly, only the main effect terms are significant. The t-tests on the individual factor effects indicate

Seq SS 59.188 9.187 5.062 19.500 19.500 92.938

Adj SS 59.188 9.187 5.062 19.500 19.500

Adj MS 19.729 3.062 5.062 2.437 2.437

that feed rate and depth of cut have large main effects, and there may be some mild interaction between these two factors. Therefore, the Minitab output is in agreement with the results given previously.

Models and Residual Analysis We may obtain the residuals from a 2k design by using the method demonstrated earlier for the 22 design. As an example, consider the surface roughness experiment. The three largest effects are A, B, and the AB interaction. The regression model used to obtain the predicted values is Y  0  1x1  2 x2  12 x1x2  where x1 represents factor A, x2 represents factor B, and x1 x2 represents the AB interaction. The regression coefficients 1, 2, and 12 are estimated by one-half the corresponding effect estimates, and 0 is the grand average. Thus, 3.375 1.625 1.375 b x1  a b x2  a b x1x2 2 2 2  11.0625  1.6875x1  0.8125x2  0.6875x1x2

yˆ  11.0625  a

Note that the regression coefficients are presented by Minitab in the upper panel of Table 14-18. The predicted values would be obtained by substituting the low and high levels of A and B into

JWCL232_c14_551-636.qxd

584

1/16/10

9:56 AM

Page 584

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS 99 95 90

Normal probability

80 70 60 30 20 10 8

Figure 14-22 Normal probability plot of residuals from the surface roughness experiment.

1

–2.250 –1.653 –0.917 –0.250

0.417

1.083

1.750

Residual

this equation. To illustrate this, at the treatment combination where A, B, and C are all at the low level, the predicted value is yˆ  11.0625  1.6875112  0.8125112  0.6875112112  9.25 Since the observed values at this run are 9 and 7, the residuals are 9  9.25  0.25 and 7  9.25  2.25. Residuals for the other 14 runs are obtained similarly. A normal probability plot of the residuals is shown in Fig. 14-22. Since the residuals lie approximately along a straight line, we do not suspect any problem with normality in the data. There are no indications of severe outliers. It would also be helpful to plot the residuals versus the predicted values and against each of the factors A, B, and C. Projection of 2k Designs Any 2k design will collapse or project into another 2k design in fewer variables if one or more of the original factors are dropped. Sometimes this can provide additional insight into the remaining factors. For example, consider the surface roughness experiment. Since factor C and all its interactions are negligible, we could eliminate factor C from the design. The result is to collapse the cube in Fig. 14-20 into a square in the A  B plane; therefore, each of the four runs in the new design has four replicates. In general, if we delete h factors so that r  k  h factors remain, the original 2k design with n replicates will project into a 2r design with n2h replicates.

14-5.3 Single Replicate of the 2k Design As the number of factors in a factorial experiment grows, the number of effects that can be estimated also grows. For example, a 24 experiment has 4 main effects, 6 two-factor interactions, 4 three-factor interactions, and 1 four-factor interaction, while a 26 experiment has 6 main effects, 15 two-factor interactions, 20 three-factor interactions, 15 four-factor interactions, 6 five-factor interactions, and 1 six-factor interaction. In most situations the

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 585

14-5 2k FACTORIAL DESIGNS

585

sparsity of effects principle applies; that is, the system is usually dominated by the main effects and low-order interactions. The three-factor and higher order interactions are usually negligible. Therefore, when the number of factors is moderately large, say, k  4 or 5, a common practice is to run only a single replicate of the 2k design and then pool or combine the higher order interactions as an estimate of error. Sometimes a single replicate of a 2k design is called an unreplicated 2k factorial design. When analyzing data from unreplicated factorial designs, occasionally real high-order interactions occur. The use of an error mean square obtained by pooling high-order interactions is inappropriate in these cases. A simple method of analysis can be used to overcome this problem. Construct a plot of the estimates of the effects on a normal probability scale. The effects that are negligible are normally distributed, with mean zero and variance 2 and will tend to fall along a straight line on this plot, whereas significant effects will have nonzero means and will not lie along the straight line. We will illustrate this method in the next example.

EXAMPLE 14-5

Plasma Etch

An article in Solid State Technology [“Orthogonal Design for Process Optimization and Its Application in Plasma Etching” (May 1987, pp. 127–132)] describes the application of factorial designs in developing a nitride etch process on a singlewafer plasma etcher. The process uses C2F6 as the reactant gas. It is possible to vary the gas flow, the power applied to the cathode, the pressure in the reactor chamber, and the spacing between the anode and the cathode (gap). Several response

variables would usually be of interest in this process, but in this example we will concentrate on etch rate for silicon nitride. We will use a single replicate of a 24 design to investigate this process. Since it is unlikely that the three- and four-factor interactions are significant, we will tentatively plan to combine them as an estimate of error. The factor levels used in the design are shown below:

Design Factor Level

Gap (cm)

Pressure (mTorr)

C2F6 Flow (SCCM)

Power (w)

Low () High ()

0.80 1.20

450 550

125 200

275 325

Table 14-19 presents the data from the 16 runs of the 24 design. Table 14-20 is the table of plus and minus signs for the 24 design. The signs in the columns of this table can be used to estimate the factor effects. For example, the estimate of factor A is 1 A  3a  ab  ac  abc  ad  abd  acd 8  abcd  112  b  c  bc  d  bd  cd  bcd 4 1  3669  650  642  635  749  868  860 8  729  550  604  633  601  1037  1052  1075  10634  101.625 Thus, the effect of increasing the gap between the anode and the cathode from 0.80 to 1.20 centimeters is to decrease the etch rate by 101.625 angstroms per minute.

It is easy to verify (using Minitab, for example) that the complete set of effect estimates is A B AB C AC BC ABC D

 101.625  1.625  7.875  7.375  24.875  43.875  15.625  306.125

AD BD ABD CD ACD BCD ABCD

 153.625  0.625  4.125  2.125  5.625  25.375  40.125

The normal probability plot of these effects from the plasma etch experiment is shown in Fig. 14-23. Clearly, the main effects of A and D and the AD interaction are significant, because they fall far from the line passing through the other

JWCL232_c14_551-636.qxd

586

1/16/10

9:56 AM

Page 586

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

Table 14-19

The 24 Design for the Plasma Etch Experiment

A (Gap)

B (Pressure)

C (C2F6 Flow)

D (Power)

Etch Rate (Å/min)

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

550 669 604 650 633 642 601 635 1037 749 1052 868 1075 860 1063 729

points. The analysis of variance summarized in Table 14-21 confirms these findings. Notice that in the analysis of variance we have pooled the three- and four-factor interactions to form the error mean square. If the normal probability plot had indicated that any of these interactions were important, they would not have been included in the error term. Practical Interpretation: Since A  101.625, the effect of increasing the gap between the cathode and anode

Table 14-20 112 a b ab c ac bc abc d ad bd abd cd acd bcd abcd

is to decrease the etch rate. However, D  306.125; thus, applying higher power levels will increase the etch rate. Figure 14-24 is a plot of the AD interaction. This plot indicates that the effect of changing the gap width at low power settings is small, but that increasing the gap at high power settings dramatically reduces the etch rate. High etch rates are obtained at high power settings and narrow gap widths.

Contrast Constants for the 24 Design

A

B

AB

C

AC

BC

ABC

D

AD

BD

ABD

CD

ACD

BCD

ABCD

               

               

               

               

               

               

               

               

               

               

               

               

               

               

               

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 587

587

14-5 2k FACTORIAL DESIGNS

99

D

96 80 70

1400 1200

60 Etch rate Å/min

Normal probability

90

30 20 10

A

6 1

AD

1000

D(Power)high = 325 w

800 600

D(Power)low = 275 w

400 200

–141.1 –64.5 12.1

88.8 165.4 242.0 318.6 Effect

Figure 14-23 Normal probability plot of effects from the plasma etch experiment.

0

High (1.20 cm)

Low (0.80 cm) A (Gap)

Figure 14-24 AD (Gap-Power) interaction from the plasma etch experiment.

The residuals from the experiment in Example 14-5 can be obtained from the regression model yˆ  776.0625  a

101.625 306.125 153.625 b x1  a b x4  a b x1x4 2 2 2

For example, when both A and D are at the low level, the predicted value is yˆ  776.0625  a

101.625 306.125 153.625 b 112  a b 112  a b 112112 2 2 2

 597

Table 14-21 Source of Variation A B C D AB AC AD BC BD CD Error Total

Analysis of Variance for the Plasma Etch Experiment Sum of Squares

Degrees of Freedom

41,310.563 10.563 217.563 374,850.063 248.063 2,475.063 94,402.563 7,700.063 1.563 18.063 10,186.813 531,420.938

1 1 1 1 1 1 1 1 1 1 5 15

Mean Square

f0

P-Value

41,310.563 10.563 217.563 374,850.063 248.063 2,475.063 94,402.563 7,700.063 1.563 18.063 2,037.363

20.28 1 1 183.99 1 1.21 46.34 3.78 1 1

0.0064 — — 0.0000 — 0.3206 0.0010 0.1095 — —

JWCL232_c14_551-636.qxd

588

1/16/10

9:56 AM

Page 588

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS 99 96 90

Normal probability

80 70 60 30 20 10 8 1

–72.50

–49.33 –26.17

–3.00 Residual

20.17

43.33

66.50

Figure 14-25 Normal probability plot of residuals from the plasma etch experiment.

and the four residuals at this treatment combination are e1  550  597  47

e2  604  597  7

e3  633  597 

e4  601  597  4

36

The residuals at the other three treatment combinations (A high, D low), (A low, D high), and (A high, D high) are obtained similarly. A normal probability plot of the residuals is shown in Fig. 14-25. The plot is satisfactory.

14-5.4 Addition of Center Points to a 2k Design A potential concern in the use of two-level factorial designs is the assumption of linearity in the factor effects. Of course, perfect linearity is unnecessary, and the 2k system will work quite well even when the linearity assumption holds only approximately. However, there is a method of replicating certain points in the 2k factorial that will provide protection against curvature as well as allow an independent estimate of error to be obtained. The method consists of adding center points to the 2k design. These consist of nC replicates run at the point xi  0 (i  1, 2, . . . , k). One important reason for adding the replicate runs at the design center is that center points do not affect the usual effects estimates in a 2k design. We assume that the k factors are quantitative. To illustrate the approach, consider a 22 design with one observation at each of the factorial points (, ), (, ), (, ), and (, ) and nC observations at the center points (0, 0). Figure 14-26 illustrates the situation. Let yF be the average of the four runs at the four factorial points, and let yC be the average of the nC run at the center point. If the difference yF  yC is small, the center points lie on or near the plane passing through the factorial points,

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 589

14-5 2k FACTORIAL DESIGNS

589

y

+1

–1 0

0

+1 A

–1 B

Figure 14-26 A 22 design with center points.

and there is no curvature. On the other hand, if yF  yC is large, curvature is present. A singledegree-of-freedom sum of squares for curvature is given by Curvature Sum of Squares

SSCurvature  

nFnC 1 yF  yC 2 2 nF  nC °

yF  yC

2

1 1 ¢  nC B nF

(14-19)

where, in general, nF is the number of factorial design points. This quantity may be compared to the error mean square to test for curvature. Notice that when Equation 14-19 is divided by

ˆ 2  MSE, the result is similar to the square of the t statistic used to compare two means. More specifically, when points are added to the center of the 2k design, the model we may entertain is k

k

Y  0  a j x j  b ij x i xj  a jj x2j  j1

i j

j1

where the jj are pure quadratic effects. The test for curvature actually tests the hypotheses k

H0: a jj  0 j1 k

H1: a jj  0 j1

Furthermore, if the factorial points in the design are unreplicated, we may use the nC center points to construct an estimate of error with nC  1 degrees of freedom.

JWCL232_c14_551-636.qxd

590

1/16/10

9:56 AM

Page 590

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS 40.0

B = Temperature (°C)

160

41.5

1

155

0

150

–1

40.3 40.5 40.7 40.2 40.6

39.3

40.9

–1 30

0

+1

35

40

A = Reaction time (min)

Figure 14-27 The 22 design with five center points for Example 14-6.

EXAMPLE 14-6

Process Yield

A chemical engineer is studying the percentage of conversion or yield of a process. There are two variables of interest, reaction time and reaction temperature. Because she is uncertain about the assumption of linearity over the region of exploration, the engineer decides to conduct a 22 design (with a single replicate of each factorial run) augmented with five center points. The design and the yield data are shown in Fig. 14-27. Table 14-22 summarizes the analysis of variance for this experiment. The mean square error is calculated from the center points as follows:

SSE MSE   nC  1

a

Center points

1 yi  yC 2 2

5



i1

4



SSCurvature  

nC  1

2 a 1 yi  40.462

The average of the points in the factorial portion of the design is yF  40.425, and the average of the points at the center is yC  40.46. The difference yF  yC  40.425  40.46  0.035 appears to be small. The curvature sum of squares in the analysis of variance table is computed from Equation 14-19 as follows:

0.1720  0.0430 4

Table 14-22

nF n C 1 yF  yC 2 2 nF  nC

14215210.0352 2  0.0027 45

Practical Interpretation: The analysis of variance indicates that both factors exhibit significant main effects, that there is no interaction, and that there is no evidence of curvature in the response over the region of exploration. That is, the null k hypothesis H0: g j1 jj  0 cannot be rejected.

Analysis of Variance for Example 14-6

Source of Variation

Sum of Squares

Degrees of Freedom

Mean Square

A (Time) B (Temperature) AB Curvature Error Total

2.4025 0.4225 0.0025 0.0027 0.1720 3.0022

1 1 1 1 4 8

2.4025 0.4225 0.0025 0.0027 0.0430

f0

P-Value

55.87 9.83 0.06 0.06

0.0017 0.0350 0.8237 0.8163

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 591

14-5 2k FACTORIAL DESIGNS

591

EXERCISES FOR SECTION 14-5 14-12. An engineer is interested in the effect of cutting speed (A), metal hardness (B), and cutting angle (C) on the life of a cutting tool. Two levels of each factor are chosen, and two replicates of a 23 factorial design are run. The tool life data (in hours) are shown in the table at right.

Treatment Combination 112 a b ab c ac bc abc

Replicate I

II

221 325 354 552 440 406 605 392

311 435 348 472 453 377 500 419

(a) Analyze the data from this experiment. (b) Find an appropriate regression model that explains tool life in terms of the variables used in the experiment. (c) Analyze the residuals from this experiment. 14-13. Four factors are thought to influence the taste of a soft-drink beverage: type of sweetener (A), ratio of syrup to water (B), carbonation level (C ), and temperature (D). Each factor can be run at two levels, producing a 24 design. At each run in the design, samples of the beverage are given to a test

Treatment Combination 112 a b ab c ac bc abc d ad bd abd cd acd bcd abcd

Replicate I

II

159 168 158 166 175 179 173 179 164 187 163 185 168 197 170 194

163 175 163 168 178 183 168 182 159 189 159 191 174 199 174 198

panel consisting of 20 people. Each tester assigns the beverage a point score from 1 to 10. Total score is the response variable, and the objective is to find a formulation that maximizes total score. Two replicates of this design are run, and the results are shown in the table. Analyze the data and draw conclusions. Use  0.05 in the statistical tests. 14-14. The data shown here represent a single replicate of a 25 design that is used in an experiment to study the compressive strength of concrete. The factors are mix (A), time (B), laboratory (C), temperature (D), and drying time (E). 112 a b ab c ac bc abc d ad bd abd cd acd bcd abcd

               

700 900 3400 5500 600 1000 3000 5300 1000 1100 3000 6100 800 1100 3300 6000

e  ae  be  abe  ce  ace  bce  abce  de  ade  bde  abde  cde  acde  bcde  abcde 

800 1200 3500 6200 600 1200 3006 5500 1900 1500 4000 6500 1500 2000 3400 6800

(a) Estimate the factor effects. (b) Which effects appear important? Use a normal probability plot. (c) If it is desirable to maximize the strength, in which direction would you adjust the process variables? (d) Analyze the residuals from this experiment. 14-15. An article in IEEE Transactions on Semiconduc-tor Manufacturing (Vol. 5, 1992, pp. 214–222) describes an experiment to investigate the surface charge on a silicon wafer. The factors thought to influence induced surface charge are cleaning method (spin rinse dry or SRD and spin dry or SD) and the position on the wafer where the charge was measured. The surface charge (1011 q/cm3) response data are as shown. Test Position L R SD

1.66 1.90 1.92

1.84 1.84 1.62

SRD

4.21 1.35 2.08

7.58 2.20 5.36

Cleaning Method

JWCL232_c14_551-636.qxd

592

1/16/10

9:56 AM

Page 592

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

(a) Estimate the factor effects. (b) Which factors appear important? Use  0.05. (c) Analyze the residuals from this experiment. 14-16. An article in Oikos: A Journal of Ecology [“Regulation of Root Vole Population Dynamics by Food Supply and Predation: A Two-Factor Experiment” (2005, Vol. 109, pp. 387–395)] investigated how food supply interacts with predation in the regulation of root vole (Microtus oeconomus Pallas) population dynamics. A replicated two-factor field experiment manipulating both food supply and predation condition for root voles was conducted. Four treatments were applied: P, F (no-predator, food-supplemented); P, F (predator-access, food-supplemented); P, F (no-predator, nonsupplemented); P, F (predator-access, food-supplemented). The population density of root voles (voles ha1 2 for each treatment combination in each is shown below. Food Supply (F) 1 1 1 1

Predation (P) 1 1 1 1

Replicates 88.589 56.949 65.439 40.799

114.059 97.079 89.089 47.959

200.979 78.759 172.339 74.439

(a) What is an appropriate statistical model for this experiment? (b) Analyze the data and draw conclusions. (c) Analyze the residuals from this experiment. Are there any problems with model adequacy? 14-17. An experiment was run in a semiconductor fabrication plant in an effort to increase yield. Five factors, each at two levels, were studied. The factors (and levels) were A  aperture setting (small, large), B  exposure time (20% below nominal, 20% above nominal), C  development time (30 and 45 seconds), D  mask dimension (small, large), and E  etch time (14.5 and 15.5 minutes). The following unreplicated 25 design was run: 112  a  b  ab  c  ac  bc  abc  d  ad  bd  abd  cd  acd  bcd  abcd 

7 9 34 55 16 20 40 60 8 10 32 50 18 21 44 61

e  8 ae  12 be  35 abe  52 ce  15 ace  22 bce  45 abce  65 de  6 ade  10 bde  30 abde  53 cde  15 acde  20 bcde  41 abcde  63

(a) Construct a normal probability plot of the effect estimates. Which effects appear to be large? (b) Conduct an analysis of variance to confirm your findings for part (a). (c) Construct a normal probability plot of the residuals. Is the plot satisfactory? (d) Plot the residuals versus the predicted yields and versus each of the five factors. Comment on the plots. (e) Interpret any significant interactions. (f ) What are your recommendations regarding process operating conditions? (g) Project the 25 design in this problem into a 2r for r  5 design in the important factors. Sketch the design and show the average and range of yields at each run. Does this sketch aid in data interpretation? 14-18. An experiment described by M. G. Natrella in the National Bureau of Standards’ Handbook of Experimental Statistics (No. 91, 1963) involves flame-testing fabrics after applying fire-retardant treatments. The four factors considered are type of fabric (A), type of fire-retardant treatment (B), laundering condition (C—the low level is no laundering, the high level is after one laundering), and method of conducting the flame test (D). All factors are run at two levels, and the response variable is the inches of fabric burned on a standard size test sample. The data are:

112 a b ab c ac bc abc

       

42 31 45 29 39 28 46 32

d ad bd abd cd acd bcd abcd

       

40 30 50 25 40 25 50 23

(a) Estimate the effects and prepare a normal plot of the effects. (b) Construct an analysis of variance table based on the model tentatively identified in part (a). (c) Construct a normal probability plot of the residuals and comment on the results. 14-19. Consider the data from Exercise 14-12. Suppose that the data from the second replicate was not available. Analyze the data from replicate I only and comment on your findings. 14-20. A 24 factorial design was run in a chemical process. The design factors are A  time, B  concentration, C  pressure, and D  temperature. The response variable is yield. The data follow:

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 593

14-5 2k FACTORIAL DESIGNS

Run

A

B

C

D

Yield (pounds)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

               

               

               

               

12 18 13 16 17 15 20 15 10 25 13 24 19 21 17 23

Factor Levels  A (hours) 2 B (%) 14 C (psi) 60 D (C) 200

 3 18 80 250

 80.25  65.50  9.25  20.50

AB AC AD BC BD CD

 53.25  11.00  9.75  18.36  15.10  1.25

American Supplier Institute, Dearborn, MI, 1986, pp. 13–21). The purpose was to reduce the number of defects in the finish of sheet-molded grill opening panels. A portion of the experimental design, and the resulting number of defects, yi observed on each run is shown in the table following. This is a single replicate of the 24 design. (a) Estimate the factor effects and use a normal probability plot to tentatively identify the important factors. (b) Fit an appropriate model using the factors identified in part (a) above. (c) Plot the residuals from this model versus the predicted number of defects. Also, prepare a normal probability plot of the residuals. Comment on the adequacy of these plots. (d) The table also shows the square root of the number of defects. Repeat parts (a) and (c) of the analysis using the square root of the number of defects as the response. Does this change the conclusions? The Gril Defects Experiment

(a) Estimate the factor effects. Based on a normal probability plot of the effect estimates, identify a model for the data from this experiment. (b) Conduct an ANOVA based on the model identified in part (a). What are your conclusions? (c) Analyze the residuals and comment on model adequacy. (d) Find a regression model to predict yield in terms of the actual factor levels. (e) Can this design be projected into a 23 design with two replicates? If so, sketch the design and show the average and range of the two yield values at each cube corner. Discuss the practical value of this plot. 14-21. An experiment has run a single replicate of a 24 design and calculated the following factor effects: A B C D

593

ABC  2.95 ABD  8.00 ACD  10.25 BCD  7.95 ABCD  6.25

(a) Construct a normal probability plot of the effects. (b) Identify a tentative model, based on the plot of effects in part (a). (c) Estimate the regression coefficients in this model, assuming that y  400. 14-22. A two-level factorial experiment in four factors was conducted by Chrysler and described in the article “Sheet Molded Compound Process Improvement” by P. I. Hsieh and D. E. Goodwin (Fourth Symposium on Taguchi Methods,

Run

A

B

C

D

y

1y

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

               

               

               

               

56 17 2 4 3 4 50 2 1 0 3 12 3 4 0 0

7.48 4.12 1.41 2.00 1.73 2.00 7.07 1.41 1.00 0.00 1.73 3.46 1.73 2.00 0.00 0.00

14-23. Consider a 22 factorial experiment with four center points. The data are 112  21, a  125, b  154, ab  352, and the responses at the center point are 92, 130, 98, 152. Compute an ANOVA with the sum of squares for curvature and conduct an F-test for curvature. Use  0.05. 14-24. Consider the experiment in Exercise 14-14. Suppose that a center point with five replicates is added to the factorial runs and the responses are 2800, 5600, 4500, 5400, 3600. Compute an ANOVA with the sum of squares for curvature and conduct an F-test for curvature. Use  0.05. 14-25. Consider the experiment in Exercise 14-17. Suppose that a center point with five replicates is added to the factorial runs and the responses are 45, 40, 41, 47, and 43.

JWCL232_c14_551-636.qxd

594

1/19/10

9:02 PM

Page 594

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

(a) Estimate the experimental error using the center points. Compare this to the estimate obtained originally in Exercise 14-17 by pooling apparently nonsignificant effects. (b) Test for curvature with ␣ ⫽ 0.05. 14-26. An article in Talanta (2005, Vol. 65, pp. 895–899) presented a 23 factorial design to find lead level by using flame atomic absorption spectrometry (FAAS). The data are shown in the following table.

Factors Run 1 2 3 4 5 6 7 8

ST ⫺ ⫹ ⫺ ⫹ ⫺ ⫹ ⫺ ⫹

pH ⫺ ⫺ ⫹ ⫹ ⫺ ⫺ ⫹ ⫹

Lead Recovery (%) RC ⫺ ⫺ ⫺ ⫺ ⫹ ⫹ ⫹ ⫹

R1 39.8 51.3 57.9 78.9 78.9 84.2 94.4 94.7

R2 42.1 48 58.1 85.9 84.2 84.2 90.9 105.3

The factors and levels are shown in the following table. Factor

Low (⫺)

High (⫹)

Reagent concentration (RC) (mol 1⫺1) pH Shaking time (ST) (min)

5 ⫻ 10⫺6

5 ⫻ 10⫺5

6.0 10

8.0 30

(a) Construct a normal probability plot of the effect estimates. Which effects appear to be large? (b) Conduct an analysis of variance to confirm your findings for part (a). (c) Analyze the residuals from this experiment. Are there any problems with model adequacy? 14-27. An experiment to study the effect of machining factors on ceramic strength was described at http://www.itl. nist.gov/div898/handbook/. Five factors were considered at two levels each: A ⫽ Table Speed, B ⫽ Down Feed Rate, C ⫽ Wheel Grit, D ⫽ Direction, E ⫽ Batch. The response is the average of the ceramic strength over 15 repetitions. The following data are from a single replicate of a 25 factorial design.

A

B

C

D

E

Strength

⫺1 1 ⫺1 1

⫺1 ⫺1 1 1

⫺1 ⫺1 ⫺1 ⫺1

⫺1 ⫺1 ⫺1 ⫺1

⫺1 ⫺1 ⫺1 ⫺1

680.45 722.48 702.14 666.93

⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1

⫺1 ⫺1 1 1 ⫺1 ⫺1 1 1 ⫺1 ⫺1 1 1 ⫺1 ⫺1 1 1 ⫺1 ⫺1 1 1 ⫺1 ⫺1 1 1 ⫺1 ⫺1 1 1

1 1 1 1 ⫺1 ⫺1 ⫺1 ⫺1 1 1 1 1 ⫺1 ⫺1 ⫺1 ⫺1 1 1 1 1 ⫺1 ⫺1 ⫺1 ⫺1 1 1 1 1

⫺1 ⫺1 ⫺1 ⫺1 1 1 1 1 1 1 1 1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 1 1 1 1 1 1 1 1

⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

703.67 642.14 692.98 669.26 491.58 475.52 478.76 568.23 444.72 410.37 428.51 491.47 607.34 620.8 610.55 638.04 585.19 586.17 601.67 608.31 442.9 434.41 417.66 510.84 392.11 343.22 385.52 446.73

(a) Estimate the factor effects and use a normal probability plot of the effects. Identify which effects appear to be large. (b) Fit an appropriate model using the factors identified in part (a) above. (c) Prepare a normal probability plot of the residuals. Also, plot the residuals versus the predicted ceramic strength. Comment on the adequacy of these plots. (d) Identify and interpret any significant interactions. (e) What are your recommendations regarding process operating conditions? 14-28. Consider the following Minitab output for a 23 factorial experiment. (a) How many replicates were used in the experiment? (b) Use Equation 14-17 to calculate the standard error of a coefficient. (c) Calculate the entries marked with “?” in the output.

JWCL232_c14_551-636.qxd

1/19/10

9:03 PM

Page 595

595

14-6 BLOCKING AND CONFOUNDING IN THE 2k DESIGN

Configuration

Factorial Fit: y versus A, B, C Estimated Effects and Coefficients for y (coded units) Term

Effect

Coef

Constant A 2.95 B 15.92 C ⫺37.87 A*B 20.43 A*C ⫺17.11 B*C 4.41 A*B*C 13.35

579.33 1.47 ? ⫺18.94 10.21 ⫺8.55 2.21 6.68

S ⫽ 153.832

SE Coef

T

P

38.46 15.06 38.46 0.04 38.46 0.21 38.46 ⫺0.49 38.46 ? 38.46 ⫺0.22 38.46 0.06 ? 0.17

0.000 0.970 0.841 0.636 0.797 0.830 0.956 0.866

R⫺Sq ⫽ 5.22% R⫺Sq (adj) ⫽ 0.00%

Analysis of Variance for y (coded units) Source

DF Seq SS Adj SS Adj MS

F

P

Main Effects 3 6785 6785 2261.8 ? 0.960 2-Way 3 ? 2918 972.5 0.04 0.988 Interactions 3-Way 1 ? 713 713.3 0.03 0.866 Interactions Residual 8 189314 189314 23664.2 Error Pure Error 8 189314 189314 23664.2 Total 15 199730

14-29. An article in Analytica Chimica Acta [“Designof-Experiment Optimization of Exhaled Breath Condensate Analysis Using a Miniature Differential Mobility Spectrometer (DMS)” (2008, Vol. 628, No. 2, pp. 155–161)] examined four parameters that affect the sensitivity and detection of the analytical instruments used to measure clinical samples. They optimized the sensor function using EBC samples spiked with acetone, a known clinical biomarker in breath. The following table shows the results for a single replicate of a 24 factorial experiment for one of the outputs, the average amplitude of acetone peak over three repetitions.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

A

B

C

D

Y

⫹ ⫹ ⫹ ⫹ ⫹ ⫹ ⫹ ⫹ ⫺ ⫺ ⫺ ⫺ ⫺ ⫺ ⫺ ⫺

⫹ ⫹ ⫹ ⫹ ⫺ ⫺ ⫺ ⫺ ⫹ ⫹ ⫹ ⫹ ⫺ ⫺ ⫺ ⫺

⫹ ⫹ ⫺ ⫺ ⫹ ⫹ ⫺ ⫺ ⫹ ⫹ ⫺ ⫺ ⫹ ⫹ ⫺ ⫺

⫹ ⫺ ⫹ ⫺ ⫹ ⫺ ⫹ ⫺ ⫹ ⫺ ⫹ ⫺ ⫹ ⫺ ⫹ ⫺

0.12 0.1193 0.1196 0.1192 0.1186 0.1188 0.1191 0.1186 0.121 0.1195 0.1196 0.1191 0.1192 0.1194 0.1188 0.1188

The factors and levels are shown in the following table. A RF voltage of the DMS sensor (1200 or 1400 V) B Nitrogen carrier gas flow rate (250 or 500mLmin⫺1) C Solid phase microextraction (SPME) filter type (polyacrylate or PDMS–DVB) D GC cooling profile (cryogenic and noncryogenic) (a) Estimate the factor effects and use a normal probability plot of the effects. Identify which effects appear to be large, and identify a model for the data from this experiment. (b) Conduct an ANOVA based on the model identified in part (a). What are your conclusions? (c) Analyze the residuals from this experiment. Are there any problems with model adequacy? (d) Project the design in this problem into a 2r design for r ⬍ 4 in the important factors. Sketch the design and show the average and range of yields at each run. Does this sketch aid in data representation?

14-6 BLOCKING AND CONFOUNDING IN THE 2k DESIGN It is often impossible to run all the observations in a 2k factorial design under homogeneous conditions. Blocking is the design technique that is appropriate for this general situation. However, in many situations the block size is smaller than the number of runs in the complete replicate. In these cases, confounding is a useful procedure for running the 2k design in 2p blocks where the number of runs in a block is less than the number of treatment combinations in one complete replicate. The technique causes certain interaction effects to be indistinguishable from

JWCL232_c14_551-636.qxd

596

1/16/10

9:56 AM

Page 596

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

+

b

ab

= Run in block 1 = Run in block 2

Figure 14-28 A 22 design in two blocks. (a) Geometric view. (b) Assignment of the four runs to two blocks.

(1) – +

–

Block 1

Block 2

(1)

a

ab

b

a

A Geometric view

Assignment of the four runs to two blocks

(a)

(b)

blocks or confounded with blocks. We will illustrate confounding in the 2k factorial design in 2p blocks, where p  k. Consider a 22 design. Suppose that each of the 22  4 treatment combinations requires four hours of laboratory analysis. Thus, two days are required to perform the experiment. If days are considered as blocks, we must assign two of the four treatment combinations to each day. This design is shown in Fig. 14-28. Notice that block 1 contains the treatment combinations (1) and ab and that block 2 contains a and b. The contrasts for estimating the main effects of factors A and B are ContrastA  ab  a  b  112

ContrastB  ab  b  a  112

Note that these contrasts are unaffected by blocking since in each contrast there is one plus and one minus treatment combination from each block. That is, any difference between block 1 and block 2 that increases the readings in one block by an additive constant cancels out. The contrast for the AB interaction is ContrastAB  ab  112  a  b Since the two treatment combinations with the plus signs, ab and (1), are in block 1 and the two with the minus signs, a and b, are in block 2, the block effect and the AB interaction are identical. That is, the AB interaction is confounded with blocks. The reason for this is apparent from the table of plus and minus signs for the 22 design shown in Table 14-12. From the table we see that all treatment combinations that have a plus on AB are assigned to block 1, whereas all treatment combinations that have a minus sign on AB are assigned to block 2. This scheme can be used to confound any 2k design in two blocks. As a second example, consider a 23 design, run in two blocks. From the table of plus and minus signs, shown in Table 14-15, we assign the treatment combinations that are minus in the ABC column to block 1 and those that are plus in the ABC column to block 2. The resulting design is shown in Fig. 14-29. There is a more general method of constructing the blocks. The method employs a defining contrast, say, L  1x1  2 x 2      k x k

(14-20)

where xi is the level of the ith factor appearing in a treatment combination and i is the exponent appearing on the ith factor in the effect that is to be confounded with blocks. For the 2k system, we have either i  0 or 1, and either xi  0 (low level) or xi  1 (high level).

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 597

597

14-6 BLOCKING AND CONFOUNDING IN THE 2k DESIGN abc

bc

= Run in block 1 c

3

Figure 14-29 The 2 design in two blocks with ABC confounded. (a) Geometric view. (b) Assignment of the eight runs to two blocks.

Block 1

Block 2

(1)

a

ab

b

= Run in block 2

ac

C

b

ab

ac

c

bc

abc

A

a

(1)

B

Geometric view

Assignment of the eight runs to two blocks

(a)

(b)

Treatment combinations that produce the same value of L (modulus 2) will be placed in the same block. Since the only possible values of L (mod 2) are 0 and 1, this will assign the 2k treatment combinations to exactly two blocks. As an example, consider the 23 design with ABC confounded with blocks. Here x1 corresponds to A, x2 to B, x3 to C, and 1  2  3  1. Thus, the defining contrast that would be used to confound ABC with blocks is L  x1  x2  x3 To assign the treatment combinations to the two blocks, we substitute the treatment combinations into the defining contrast as follows: 112: L  1102  1102  1102  0  0 1mod 22 a: b:

L  1112  1102  1102  1  1 1mod 22

L  1102  1112  1102  1  1 1mod 22

ab: L  1112  1112  1102  2  0 1mod 22 c:

L  1102  1102  1112  1  1 1mod 22

ac: L  1112  1102  1112  2  0 1mod 22

bc: L  1102  1112  1112  2  0 1mod 22

abc: L  1112  1112  1112  3  1 1mod 22 Thus (1), ab, ac, and bc are run in block 1, and a, b, c, and abc are run in block 2. This same design is shown in Fig. 14-29. A shortcut method is useful in constructing these designs. The block containing the treatment combination (1) is called the principal block. Any element [except (1)] in the principal block may be generated by multiplying two other elements in the principal block modulus 2 on the exponents. For example, consider the principal block of the 23 design with ABC confounded, shown in Fig. 14-29. Note that ab ⴢ ac  a2bc  bc ab ⴢ bc  ab2c  ac ac ⴢ bc  abc2  ab

JWCL232_c14_551-636.qxd

598

1/16/10

9:56 AM

Page 598

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

Treatment combinations in the other block (or blocks) may be generated by multiplying one element in the new block by each element in the principal block modulus 2 on the exponents. For the 23 with ABC confounded, since the principal block is (1), ab, ac, and bc, we know that the treatment combination b is in the other block. Thus, elements of this second block are b ⴢ 112

b

b ⴢ ab  ab2  a b ⴢ ac

 abc

b ⴢ bc  b2c  c

EXAMPLE 14-7

Missile Miss Distance

An experiment is performed to investigate the effect of four factors on the terminal miss distance of a shoulder-fired ground-to-air missile. The four factors are target type (A), seeker type (B), target altitude (C), and target range (D). Each factor may be conveniently run at two levels, and the optical tracking system will allow terminal miss distance to be measured to the nearest foot. Two different operators or gunners are used in the flight test and, since there may be differences between operators, the test engineers decided to conduct the 24 design in two blocks with ABCD confounded. Thus, the defining contrast is L  x1  x2  x3  x4

D

–

c

abcd

bcd cd

ac

b

(1)

+

abc

bc

The experimental design and the resulting data are shown in Fig. 14-30. The effect estimates obtained from Minitab are shown in Table 14-23. A normal probability plot of the effects in Fig. 14-31 reveals that A (target type), D (target range), AD, and AC have large effects. A confirming analysis of variance, pooling the three-factor interactions as error, is shown in Table 14-24. Practical Interpretation: Since the AC and AD interactions are significant, it is logical to conclude that A (target type), C (target altitude), and D (target range) all have important effects on the miss distance and that there are interactions between target type and altitude and target type and range. Notice that the ABCD effect is treated as blocks in this analysis.

acd

bd

ab d

a C = Run in block 1

abd ad

B

Block 2

Block 1 (1) ab ac bc ad bd cd abcd

= = = = = = = =

a b c d abc bcd acd abd

3 7 6 8 10 4 8 9

(b)

Geometric view (a) 4

7 5 6 4 6 7 9 12

Assignment of the sixteen runs to two blocks

= Run in block 2 A

= = = = = = = =

Figure 14-30 The 2 design in two blocks for Example 14-7. (a) Geometric view. (b) Assignment of the 16 runs to two blocks.

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 599

14-6 BLOCKING AND CONFOUNDING IN THE 2k DESIGN

Table 14-23

599

Minitab Effect Estimates for Example 14-7

Estimated Effects and Coefficients for Distance Effect

Coef 6.938 0.063 1.312 0.313 0.438 0.938 0.063 1.187 0.813 0.188 0.187 0.062 0.063 0.438 0.187 0.187

2.625 0.625 0.875 1.875 0.125 2.375 1.625 0.375 0.375 0.125 0.125 0.875 0.375 0.375

A D AD

1 Normal score

Term Constant Block A B C D AB AC AD BC BD CD ABC ABD ACD BCD

0

_1 AC _2

0 Effect

2

Figure 14-31 Normal probability plot of the effects from Minitab, Example 14-6.

It is possible to confound the 2k design in four blocks of 2k2 observations each. To construct the design, two effects are chosen to confound with blocks, and their defining contrasts are obtained. A third effect, the generalized interaction of the two effects initially chosen, is also confounded with blocks. The generalized interaction of two effects is found by multiplying their respective letters and reducing the exponents modulus 2. For example, consider the 24 design in four blocks. If AC and BD are confounded with blocks, their generalized interaction is (AC)(BD) = ABCD. The design is constructed by using Table 14-24

Analysis of Variance for Example 14-7 Source of Variation

Blocks (ABCD) A B C D AB AC AD BC BD CD Error (ABC  ABD  ACD  BCD) Total

Sum of Squares

Degrees of Freedom

Mean Square

0.0625 27.5625 1.5625 3.0625 14.0625 0.0625 22.5625 10.5625 0.5625 0.5625 0.0625 4.2500 84.9375

1 1 1 1 1 1 1 1 1 1 1 4 15

0.0625 27.5625 1.5625 3.0625 14.0625 0.0625 22.5625 10.5625 0.5625 0.5625 0.0625 1.0625

f0

P-Value

0.06 25.94 1.47 2.88 13.24 0.06 21.24 9.94 0.53 0.53 0.06

— 0.0070 0.2920 0.1648 0.0220 — 0.0100 0.0344 — — —

JWCL232_c14_551-636.qxd

600

1/16/10

9:56 AM

Page 600

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

the defining contrasts for AC and BD: L1  x1  x3 L2  x2  x4 It is easy to verify that the four blocks are Block 1 L1  0, L2  0 112 ac bd abcd

Block 2 L1  1, L2  0

Block 3 L1  0, L2  1

Block 4 L1  1, L2  1

a c abd bcd

b abc d acd

ab bc ad cd

This general procedure can be extended to confounding the 2k design in 2p blocks, where p  k. Start by selecting p effects to be confounded, such that no effect chosen is a generalized interaction of the others. Then the blocks can be constructed from the p defining contrasts L1, L2, . . . , Lp that are associated with these effects. In addition to the p effects chosen to be confounded, exactly 2p  p  1 additional effects are confounded with blocks; these are the generalized interactions of the original p effects chosen. Care should be taken so as not to confound effects of potential interest. For more information on confounding in the 2k factorial design, refer to Montgomery (2009). This book contains guidelines for selecting factors to confound with blocks so that main effects and low-order interactions are not confounded. In particular, the book contains a table of suggested confounding schemes for designs with up to seven factors and a range of block sizes, some of which are as small as two runs. EXERCISES FOR SECTION 14-6 14-30. Consider the data from the first replicate of Exercise 14-12. (a) Suppose that these observations could not all be run under the same conditions. Set up a design to run these observations in two blocks of four observations each, with ABC confounded. (b) Analyze the data. 14-31. Consider the data from the first replicate of Exercise 14-13. (a) Construct a design with two blocks of eight observations each, with ABCD confounded. (b) Analyze the data. 14-32. Consider the data from Exercise 14-18. (a) Construct the design that would have been used to run this experiment in two blocks of eight runs each. (b) Analyze the data and draw conclusions. 14-33. Construct a 25 design in two blocks. Select the ABCDE interaction to be confounded with blocks. 14-34. Consider the data from the first replicate of Exercise 14-13, assuming that four blocks are required. Confound ABD and ABC (and consequently CD) with blocks.

(a) Construct a design with four blocks of four observations each. (b) Analyze the data. 14-35. Construct a 25 design in four blocks. Select the appropriate effects to confound so that the highest possible interactions are confounded with blocks. 14-36. Consider the 26 factorial design. Set up a design to be run in four blocks of 16 runs each. Show that a design that confounds three of the four-factor interactions with blocks is the best possible blocking arrangement. 14-37. An article in Quality Engineering [“Designed Experiment to Stabilize Blood Glucose Levels” (1999–2000, Vol. 12, pp. 83–87)] reported on an experiment to minimize variations in blood glucose levels. The factors were: volume of juice intake before exercise (4 or 8 oz), amount of exercise on a Nordic Track cross-country skier (10 or 20 min), and delay between the time of juice intake (0 or 20 min) and the beginning of the exercise period. The experiment was blocked for time of day. The data follow. (a) What effects are confounded with blocks? Comment on any concerns with the confounding in this design. (b) Analyze the data and draw conclusions.

JWCL232_c14_551-636.qxd

1/16/10

9:56 AM

Page 601

601

14-6 BLOCKING AND CONFOUNDING IN THE 2k DESIGN

Run

Juice (oz)

Exercise (min)

Delay (min)

Time of Day

Average Blood Glucose

1 2 3 4 5 6 7 8

4 8 4 8 4 8 4 8

10 10 20 20 10 10 20 20

0 0 0 0 20 20 20 20

pm am am pm am pm pm am

71.5 103 83.5 126 125.5 129.5 95 93

14-38. An article in Industrial and Engineering Chemistry [“Factorial Experiments in Pilot Plant Studies” (1951, pp. 1300–1306)] reports on an experiment to investigate the effect of temperature (A), gas throughput (B), and concentration (C) on the strength of product solution in a recirculation unit. Two blocks were used with ABC confounded, and the experiment was replicated twice. The data follow. (a) Analyze the data from this experiment.

Replicate 1 Block 1

Block 2

(1)  99 ab  52 ac  42 bc  95

a  18 b  51 c 108 abc  35 Replicate 2

Block 3 (1)  46 ab  47 ac  22 bc  67

Block 4 a  18 b  62 c 104 abc  36

(b) Analyze the residuals and comment on model adequacy. (c) Comment on the efficiency of this design. Note that we have replicated the experiment twice, yet we have no information on the ABC interaction. (d) Suggest a better design, specifically, one that would provide some information on all interactions. 14-39. Consider the following Minitab output from a single replicate of a 24 experiment in two blocks with ABCD confounded. (a) Comment on the value of blocking in this experiment. (b) What effects were used to generate the residual error in the ANOVA? (c) Calculate the entries marked with “?” in the output.

Factorial Fit: y versus Block, A, B, C, D Estimated Effects and Coefficients for y (coded units) Term

Effect

Coef

Se Coef

T

P

Constant 579.33 Block 105.68 A 15.41 7.70 B 2.95 1.47 C 15.92 7.96 D 37.87 18.94 A*B 8.16 4.08 A*C 5.91 2.95 A*D 30.28 ? B*C 20.43 10.21 B*D 17.11 8.55 C*D 4.41 2.21

9.928 9.928 9.928 9.928 9.928 9.928 9.928 9.928 9.928 9.928 9.928 9.928

58.35 10.64 0.78 0.15 0.80 1.91 0.41 0.30 ? 1.03 0.86 0.22

0.000 0.000 0.481 0.889 0.468 0.129 0.702 0.781 0.202 0.362 0.437 0.835

S  39.7131

R-Sq  96.84% R-Sq (adj)  88.16%

Analysis of Variance for y (coded units) Source

DF Seq SS

Blocks Main Effects 2-Way Interactions Residual Error Total

Adj SS Adj MS

? 4 6

178694 7735 6992

178694 7735 6992

4

6309

6309

F

P

178694 113.30 1934 1.23 ? 0.74

0.000 0.424 0.648

1577

15 199730

14-40. An article in Advanced Semiconductor Manufacturing Conference (ASMC) (May 2004, pp. 325–29) stated that dispatching rules and rework strategies are two major operational elements that impact productivity in a semiconductor fabrication plant (fab). A four-factor experiment was conducted to determine the effect of dispatching rule time (5 or 10 min), rework delay (0 or 15 min), fab temperature (60 or 80°F), and rework levels (level 0 or level 1) on key fab performance measures. The performance measure that was analyzed was the average cycle time. The experiment was blocked for the fab temperature. Data modified from the original study are shown in the following table. Dispatching Rework Fab Average Rule Time Delay Rework Temperature Cycle Time Run (min) (min) Level (°F) (min) 1 2 3 4 5 6 7 8

5 10 5 10 5 10 5 10

0 0 0 0 15 15 15 15

0 0 1 1 0 0 1 1

60 80 80 60 80 60 60 80

218 256.5 231 302.5 298.5 314 249 241

(a) What effects are confounded with blocks? Do you find any concerns with confounding in this design? If so, comment on it. (b) Analyze the data and draw conclusions.

JWCL232_c14_551-636.qxd

602

1/16/10

9:57 AM

Page 602

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

14-7 FRACTIONAL REPLICATION OF THE 2k DESIGN As the number of factors in a 2k factorial design increases, the number of runs required increases rapidly. For example, a 25 requires 32 runs. In this design, only 5 degrees of freedom correspond to main effects, and 10 degrees of freedom correspond to two-factor interactions. Sixteen of the 31 degrees of freedom are used to estimate high-order interactions— that is, three-factor and higher order interactions. Often there is little interest in these highorder interactions, particularly when we first begin to study a process or system. If we can assume that certain high-order interactions are negligible, a fractional factorial design involving fewer than the complete set of 2k runs can be used to obtain information on the main effects and low-order interactions. In this section, we will introduce fractional replications of the 2k design. A major use of fractional factorials is in screening experiments. These are experiments in which many factors are considered with the purpose of identifying those factors (if any) that have large effects. Screening experiments are usually performed in the early stages of a project when it is likely that many of the factors initially considered have little or no effect on the response. The factors that are identified as important are then investigated more thoroughly in subsequent experiments.

14-7.1 One-Half Fraction of the 2k Design A one-half fraction of the 2k design contains 2k1 runs and is often called a 2k1 fractional factorial design. As an example, consider the 231 design—that is, a one-half fraction of the 23. This design has only four runs, in contrast to the full factorial that would require eight runs. The table of plus and minus signs for the 23 design is shown in Table 14-25. Suppose we select the four treatment combinations a, b, c, and abc, as our one-half fraction. These treatment combinations are shown in the top half of Table 14-25 and in Fig. 14-32(a). Notice that the 231 design is formed by selecting only those treatment combinations that yield a plus on the ABC effect. Thus, ABC is called the generator of this particular fraction. Furthermore, the identity element I is also plus for the four runs, so we call I  ABC the defining relation for the design. The treatment combinations in the 231 design yields three degrees of freedom associated with the main effects. From the upper half of Table 14-25, we obtain the estimates of the main Table 14-25

Plus and Minus Signs for the 23 Factorial Design Factorial Effect

Treatment Combination

I

A

B

C

AB

AC

BC

ABC

a b c abc

   

   

   

   

   

   

   

   

ab ac bc 112

   

   

   

   

   

   

   

   

JWCL232_c14_551-636.qxd

1/16/10

9:57 AM

Page 603

603

14-7 FRACTIONAL REPLICATION OF THE 2k DESIGN abc c

bc

C

ac B

Figure 14-32 The one-half fractions of the 23 design. (a) The principal fraction, I  ABC. (b) The alternate fraction, I  ABC.

b

A

a

ab (1)

The principal fraction, I = +ABC

The alternate fraction, I = –ABC

(a)

(b)

effects as linear combinations of the observations, say, A B C

2 3a  b  c  abc4 1 3a  b  c  abc4 2 1 3a  b  c  abc4 2 1

It is also easy to verify that the estimates of the two-factor interactions should be the following linear combinations of the observations: BC  AC  AB 

2 3a  b  c  abc4 1 3a  b  c  abc4 2 1 3a  b  c  abc4 2 1

Thus, the linear combination of observations in column A, /A, estimates both the main effect of A and the BC interaction. That is, the linear combination /A estimates the sum of these two effects A  BC. Similarly, /B estimates B  AC, and /C estimates C  AB. Two or more effects that have this property are called aliases. In our 231 design, A and BC are aliases, B and AC are aliases, and C and AB are aliases. Aliasing is the direct result of fractional replication. In many practical situations, it will be possible to select the fraction so that the main effects and low-order interactions that are of interest will be aliased only with high-order interactions (which are probably negligible). The alias structure for this design is found by using the defining relation I  ABC. Multiplying any effect by the defining relation yields the aliases for that effect. In our example, the alias of A is A  A ⴢ ABC  A2BC  BC since A ⴢ I  A and A2  I . The aliases of B and C are B  B ⴢ ABC  AB2C  AC and C  C ⴢ ABC  ABC2  AB Now suppose that we had chosen the other one-half fraction, that is, the treatment combinations in Table 14-25 associated with minus on ABC. These four runs are shown in the

JWCL232_c14_551-636.qxd

604

1/16/10

9:57 AM

Page 604

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

lower half of Table 14-25 and in Fig. 14-32(b). The defining relation for this design is I  ABC. The aliases are A  BC, B  AC, and C  AB. Thus, estimates of A, B, and C that result from this fraction really estimate A  BC, B  AC, and C  AB. In practice, it usually does not matter which one-half fraction we select. The fraction with the plus sign in the defining relation is usually called the principal fraction, and the other fraction is usually called the alternate fraction. Note that if we had chosen AB as the generator for the fractional factorial, A  A ⴢ AB  B and the two main effects of A and B would be aliased. This typically loses important information. Sometimes we use sequences of fractional factorial designs to estimate effects. For example, suppose we had run the principal fraction of the 231 design with generator ABC. From this design we have the following effect estimates: /A  A  BC /B  B  AC /C  C  AB Suppose that we are willing to assume at this point that the two-factor interactions are negligible. If they are, the 231 design has produced estimates of the three main effects A, B, and C. However, if after running the principal fraction we are uncertain about the interactions, it is possible to estimate them by running the alternate fraction. The alternate fraction produces the following effect estimates: /A¿  A  BC /B¿  B  AC /C¿  C  AB We may now obtain de-aliased estimates of the main effects and two-factor interactions by adding and subtracting the linear combinations of effects estimated in the two individual fractions. For example, suppose we want to de-alias A from the two-factor interaction BC. Since /A  A  BC and /A¿  A  BC , we can combine these effect estimates as follows: 1 1 1/A  /A¿ 2  1A  BC  A  BC2  A 2 2 and 1 1 1/A  /A¿ 2  1A  BC  A  BC 2  BC 2 2 For all three pairs of effect estimates, we would obtain the following results: Effect, i iA iB iC

from 12 (li ⴙ liⴕ)

2 1A  BC  A  BC2  A 1 2 1B  AC  B  AC2  B 1 2 1C  AB  C  AB2  C 1

from

2 (li ⴚ liⴕ)

1

2 3A  BC  2 3B  AC  1 2 3C  AB 

1

1

1A  BC2 4  BC

1B  AC2 4  AC 1C  AB2 4  AB

JWCL232_c14_551-636.qxd

1/16/10

9:57 AM

Page 605

14-7 FRACTIONAL REPLICATION OF THE 2k DESIGN

605

Thus, by combining a sequence of two fractional factorial designs, we can isolate both the main effects and the two-factor interactions. This property makes the fractional factorial design highly useful in experimental problems since we can run sequences of small, efficient experiments, combine information across several experiments, and take advantage of learning about the process we are experimenting with as we go along. This is an illustration of the concept of sequential experimentation. A 2k1 design may be constructed by writing down the treatment combinations for a full factorial with k  1 factors, called the basic design, and then adding the kth factor by identifying its plus and minus levels with the plus and minus signs of the highest order interaction. Therefore, a 231 fractional factorial is constructed by writing down the basic design as a full 22 factorial and then equating factor C with the AB interaction. Thus, to construct the principal fraction, we would use C  AB as follows:

Basic Design

Fractional Design 23ⴚ1, I ⴝ ⴙABC

2

Full 2 A

B

A

B

C  AB

   

   

   

   

   

To obtain the alternate fraction we would equate the last column to C  AB. EXAMPLE 14-8

Plasma Etch

To illustrate the use of a one-half fraction, consider the plasma etch experiment described in Example 14-5. Suppose that we decide to use a 241 design with I  ABCD to investigate the four factors gap (A), pressure (B), C2F6 flow rate (C ), and power setting (D). This design would be constructed by writing down as the basic design a 23 in the factors A, B, and C and then setting the levels of the fourth factor D  ABC. The design and the resulting etch rates are shown in Table 14-26. The design is shown graphically in Fig. 14-33.

In this design, the main effects are aliased with the threefactor interactions; note that the alias of A is A # I  A # ABCD

or

A  A2BCD  BCD

and similarly B  ACD, C  ABD, and D  ABC. The two-factor interactions are aliased with each other. For example, the alias of AB is CD: AB # I  AB # ABCD

or

AB  A 2B 2CD  CD

Table 14-26 The 241 Design with Defining Relation I  ABCD A

B

C

D  ABC

       

       

       

       

Treatment Combination 112 ad bd ab cd ac bc abcd

Etch Rate 550 749 1052 650 1075 642 601 729

JWCL232_c14_551-636.qxd

606

1/16/10

9:57 AM

Page 606

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS D

–

+ abcd = 729

bc = 601 cd = 1075

ac = 642

C B A

bd = 1052 ab = 650 ad = 749

(1) = 550

Figure 14-33 The 2

41

design for the experiment of Example 14-8.

The other aliases are AC  BD and AD  BC. The estimates of the main effects and their aliases are found using the four columns of signs in Table 14-26. For example, from column A we obtain the estimated effect /A  A  BCD  14 1550  749  1052  650  1075  642  601  7292

the signs in the AB column are , , , , , , , , and this column produces the estimate /AB  AB  CD  14 1550  749  1052  650  1075  642  601  7292  10 From the AC and AD columns we find

 127.00

/AC  AC  BD  25.50

The other columns produce /B  B  ACD  4.00 /C  C  ABD  11.50 and /D  D  ABC  290.50 Clearly, /A and /D are large, and if we believe that the threefactor interactions are negligible, the main effects A (gap) and D (power setting) significantly affect etch rate. The interactions are estimated by forming the AB, AC, and AD columns and adding them to the table. For example,

and /AD  AD  BC  197.50 The /AD estimate is large; the most straightforward interpretation of the results is that since A and D are large, this is the AD interaction. Thus, the results obtained from the 241 design agree with the full factorial results in Example 14-5. Practical Interpretation: Often a fraction of a 2k design is satisfactory when an experiment uses four or more factors.

Computer Solution Fractional factorial designs are usually analyzed with a software package. Table 14-26 shows the effect estimates obtained from Minitab for Example 14-8. They are in agreement with the hand calculation reported earlier. Normal Probability Plot of Effects The normal probability plot is very useful in assessing the significance of effects from a fractional factorial design, particularly when many effects are to be estimated. We strongly recommend examining this plot. Figure 14-34 presents the normal probability plot of the effects from Example 14-8. This plot was obtained from Minitab. Notice that the A, D, and AD interaction effects stand out clearly in this graph.

JWCL232_c14_551-636.qxd

1/16/10

9:57 AM

Page 607

607

14-7 FRACTIONAL REPLICATION OF THE 2k DESIGN

Table 14-27 Effect Estimates from Minitab, Example 14-8 1.5

D

Normal score

1.0 0.5 0.0 _ 0.5

A

_ 1.0 _ 1.5

AD _ 200 _ 100

0

100

200

300

Effect

Figure 14-34 Normal probability plot of the effects from Minitab, Example 14-8.

Fractional Factorial Fit Estimated Effects and Coefficients for Etch Rt Term Constant Gap Pressure F Power Gap*Pressure Gap*F Gap*Power

Effect 127.00 4.00 11.50 290.50 10.00 25.50 197.50

Coef 756.00 63.50 2.00 5.75 145.25 5.00 12.75 98.75

Residual Analysis The residuals can be obtained from a fractional factorial by the regression model method shown previously. Note that the Minitab output for Example 14-8 in Table 14-27 shows the regression coefficients. The residuals should be graphically analyzed as we have discussed before, both to assess the validity of the underlying model assumptions and to gain additional insight into the experimental situation. Projection of the 2kⴚ1 Design If one or more factors from a one-half fraction of a 2k can be dropped, the design will project into a full factorial design. For example, Fig. 14-35 presents a 231 design. Notice that this design will project into a full factorial in any two of the three original factors. Thus, if we think that at most two of the three factors are important, the 231 design is an excellent design for identifying the significant factors. This projection property is highly useful in factor screening,

B

b

abc a c

Figure 14-35 Projection of a 231 design into three 22 designs.

C

A

JWCL232_c14_551-636.qxd

608

1/16/10

9:57 AM

Page 608

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS (749, 729)

(1052, 1075) +1

Figure 14-36 The 22 design obtained by dropping factors B and C from the plasma etch experiment in Example 14-7.

D (Power)

(550, 601) –1 –1

(650, 642) +1 A (Gap)

because it allows negligible factors to be eliminated, resulting in a stronger experiment in the active factors that remain. In the 241 design used in the plasma etch experiment in Example 14-8, we found that two of the four factors (B and C) could be dropped. If we eliminate these two factors, the remaining columns in Table 14-24 form a 22 design in the factors A and D, with two replicates. This design is shown in Fig. 14-36. The main effects of A and D and the strong two-factor AD interaction are clearly evident from this graph. Design Resolution The concept of design resolution is a useful way to catalog fractional factorial designs according to the alias patterns they produce. Designs of resolution III, IV, and V are particularly important. The definitions of these terms and an example of each follow. 1.

Resolution III Designs. These are designs in which no main effects are aliased with any other main effect, but main effects are aliased with two-factor interactions and some two-factor interactions may be aliased with each other. The 231 design with I  ABC is a resolution III design. We usually employ a Roman numeral subscript to indicate design resolution; thus, this one-half fraction is a 2 31 III design. 2. Resolution IV Designs. These are designs in which no main effect is aliased with any other main effect or two-factor interactions, but two-factor interactions are aliased with each other. The 241 design with I  ABCD used in Example 14-8 is a resolution IV design (2 41 IV ). 3. Resolution V Designs. These are designs in which no main effect or two-factor interaction is aliased with any other main effect or two-factor interaction, but twofactor interactions are aliased with three-factor interactions. The 251 design with I  ABCDE is a resolution V design (251 V ).

Resolution III and IV designs are particularly useful in factor screening experiments. A resolution IV design provides good information about main effects and will provide some information about all two-factor interactions.

14-7.2 Smaller Fractions: The 2kⴚp Fractional Factorial Although the 2k1 design is valuable in reducing the number of runs required for an experiment, we frequently find that smaller fractions will provide almost as much useful information at even

JWCL232_c14_551-636.qxd

1/16/10

9:57 AM

Page 609

14-7 FRACTIONAL REPLICATION OF THE 2k DESIGN

609

greater economy. In general, a 2k design may be run in a 12p fraction called a 2kp fractional factorial design. Thus, a 1兾4 fraction is called a 2k2 design, a 1兾8 fraction is called a 2k3 design, a 1兾16 fraction a 2k4 design, and so on. To illustrate the 1兾4 fraction, consider an experiment with six factors and suppose that the engineer is primarily interested in main effects but would also like to get some information about the two-factor interactions. A 261 design would require 32 runs and would have 31 degrees of freedom for estimating effects. Since there are only six main effects and 15 twofactor interactions, the one-half fraction is inefficient—it requires too many runs. Suppose we consider a 1兾4 fraction, or a 262 design. This design contains 16 runs and, with 15 degrees of freedom, will allow all six main effects to be estimated, with some capability for examining the two-factor interactions. To generate this design, we would write down a 24 design in the factors A, B, C, and D as the basic design and then add two columns for E and F. To find the new columns we could select the two design generators I  ABCE and I  BCDF. Thus, column E would be found from E  ABC, and column F would be F  BCD. That is, columns ABCE and BCDF are equal to the identity column. However, we know that the product of any two columns in the table of plus and minus signs for a 2k design is just another column in the table; therefore, the product of ABCE and BCDF or ABCE(BCDF)  AB2C2DEF  ADEF is also an identity column. Consequently, the complete defining relation for the 262 design is I  ABCE  BCDF  ADEF We refer to each term in a defining relation (such as ABCE above) as a word. To find the alias of any effect, simply multiply the effect by each word in the foregoing defining relation. For example, the alias of A is A  BCE  ABCDF  DEF The complete alias relationships for this design are shown in Table 14-28. In general, the resolution of a 2kp design is equal to the number of letters in the shortest word in the complete defining relation. Therefore, this is a resolution IV design; main effects are aliased with three-factor and higher interactions, and two-factor interactions are aliased with each other. This design would provide good information on the main effects and would give some idea about the strength of the two-factor interactions. The construction and analysis of the design are illustrated in Example 14-9.

Table 14-28 Alias Structure for the 262 IV Design with I  ABCE  BCDF  ADEF A  BCE  DEF  ABCDF B  ACE  CDF  ABDEF C  ABE  BDF  ACDEF D  BCF  AEF  ABCDE E  ABC  ADF  BCDEF F  BCD  ADE  ABCEF ABD  CDE  ACF  BEF ACD  BDE  ABF  CEF

AB  CE  ACDF  BDEF AC  BE  ABDF  CDEF AD  EF  BCDE  ABCF AE  BC  DF  ABCDEF AF  DE  BCEF  ABCD BD  CF  ACDE  ABEF BF  CD  ACEF  ABDE

JWCL232_c14_551-636.qxd

610

1/16/10

2:03 PM

Page 610

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

Table 14-29

Run 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

EXAMPLE 14-9

A 262 IV Design for the Injection-Molding Experiment

A                

B                

C                

D                

E  ABC                

F  BCD                

Observed Shrinkage (10) 6 10 32 60 4 15 26 60 8 12 34 60 16 5 37 52

Injection Molding

Parts manufactured in an injection-molding process are showing excessive shrinkage, which is causing problems in assembly operations upstream from the injection-molding area. In an effort to reduce the shrinkage, a quality-improvement team has decided to use a designed experiment to study the injectionmolding process. The team investigates six factors—mold temperature (A), screw speed (B), holding time (C), cycle time (D), gate size (E), and holding pressure (F )—each at two levels, with the objective of learning how each factor affects shrinkage and obtaining preliminary information about how the factors interact. The team decides to use a 16-run two-level fractional factorial design for these six factors. The design is constructed by writing down a 24 as the basic design in the factors A, B, C, and D and then setting E  ABC and F  BCD as discussed above. Table 14-29 shows the design, along with the observed shrinkage (10) for the test part produced at each of the 16 runs in the design. A normal probability plot of the effect estimates from this experiment is shown in Fig. 14-37. The only large effects are A (mold temperature), B (screw speed), and the AB interaction. In light of the alias relationship in Table 14-28, it seems reasonable to tentatively adopt these conclusions. The plot of the AB interaction in Fig. 14-38 shows that the process is insensitive to temperature if the screw speed is at the low level but sensitive to temperature if the screw speed is at the high level. With the screw speed at a low level, the

process should produce an average shrinkage of around 10% regardless of the temperature level chosen. Based on this initial analysis, the team decides to set both the mold temperature and the screw speed at the low level. This set of conditions should reduce the mean shrinkage of parts to around 10%. However, the variability in shrinkage from part to part is still a potential problem. In effect, the mean shrinkage can be adequately reduced by the above modifications; however, the part-to-part variability in shrinkage over a production run could still cause problems in assembly. One way to address this issue is to see if any of the process factors affect the variability in parts shrinkage. Figure 14-39 presents the normal probability plot of the residuals. This plot appears satisfactory. The plots of residuals versus each factor were then constructed. One of these plots, that for residuals versus factor C (holding time), is shown in Fig. 14-40. The plot reveals much less scatter in the residuals at the low holding time than at the high holding time. These residuals were obtained in the usual way from a model for predicted shrinkage yˆ  ˆ 0  ˆ 1x1  ˆ 2x2  ˆ 12x1x2  27.3125  6.9375x1  17.8125x2  5.9375x1x2 where x1, x2, and x1x2 are coded variables that correspond to the factors A and B and the AB interaction. The regression

JWCL232_c14_551-636.qxd

1/16/10

9:57 AM

Page 611

611

14-7 FRACTIONAL REPLICATION OF THE 2k DESIGN

99

B

95 AB

60

A

80 60

Shrinkage (× 10)

Normal probability

90

50 30 20 10 5

B+

B+

B–

B–

1

4 –5

0

5

10 15

20 25

30 35

Low

40

High Mold temperature, A

Effect

Figure 14-37 Normal probability plot of effects for Example 14-9.

model used to produce the residuals essentially removes the location effects of A, B, and AB from the data; the residuals therefore contain information about unexplained variability. Figure 14-40 indicates that there is a pattern in the variability and that the variability in the shrinkage of parts may be smaller when the holding time is at the low level. Practical Interpretation: Figure 14-41 shows the data from this experiment projected onto a cube in the factors A, B, and C. The average observed shrinkage and the range of observed shrinkage are shown at each corner of the cube.

Figure 14-38 Plot of AB (mold temperature– screw speed) interaction for Example 14-9.

From inspection of this figure, we see that running the process with the screw speed (B) at the low level is the key to reducing average parts shrinkage. If B is low, virtually any combination of temperature (A) and holding time (C ) will result in low values of average parts shrinkage. However, from examining the ranges of the shrinkage values at each corner of the cube, it is immediately clear that setting the holding time (C ) at the low level is the most appropriate choice if we wish to keep the part-to-part variability in shrinkage low during a production run.

6 99

4 2

80 70

Residuals

Normal probability

95 90

50

0

30 20

–2

10 5

–4

1

–6 –6

–3

0

3

6

Residual

Figure 14-39 Normal probability plot of residuals for Example 14-9.

Low

High Holding time (C)

Figure 14-40 Residuals versus holding time (C ) for Example 14-9.

JWCL232_c14_551-636.qxd

612

1/16/10

9:57 AM

Page 612

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS y = 56.0 R=8

y = 31.5 R = 11 +

y = 33.0 R=2

y = 60.0 R=0

B, screw speed

Figure 14-41 Average shrinkage and range of shrinkage in factors A, B, and C for Example 14-9.

–

y = 10.0 R = 12

y = 10.0 R = 10

y = 7.0 R=2 –

+

y = 11.0 R=2

–

+

C, holding time

A, mold temperature

The concepts used in constructing the 262 fractional factorial design in Example 14-9 can be extended to the construction of any 2kp fractional factorial design. In general, a 2k fractional factorial design containing 2kp runs is called a 1兾2p fraction of the 2k design or, more simply, a 2kp fractional factorial design. These designs require the selection of p independent generators. The defining relation for the design consists of the p generators initially chosen and their 2p  p  1 generalized interactions. The alias structure may be found by multiplying each effect column by the defining relation. Care should be exercised in choosing the generators so that effects of potential interest are not aliased with each other. Each effect has 2p  1 aliases. For moderately large values of k, we usually assume higher order interactions (say, third- or fourth-order or higher) to be negligible, and this greatly simplifies the alias structure. It is important to select the p generators for the 2kp fractional factorial design in such a way that we obtain the best possible alias relationships. A reasonable criterion is to select the generators so that the resulting 2kp design has the highest possible design resolution. Montgomery (2009) presented a table of recommended generators for 2kp fractional factorial designs for k 15 factors and up to as many as n 128 runs. A portion of his table is reproduced here as Table 14-30. In this table, the generators are shown with either  or  choices; selection of all generators as  will give a principal fraction, while if any generators are  choices, the design will be one of the alternate fractions for the same family. The suggested generators in this table will result in a design of the highest possible resolution. Montgomery (2009) also provided a table of alias relationships for these designs.

EXAMPLE 14-10 Aliases with Seven Factors To illustrate the use of Table 14-30, suppose that we have seven factors and that we are interested in estimating the seven main effects and obtaining some insight regarding the twofactor interactions. We are willing to assume that three-factor and higher interactions are negligible. This information suggests that a resolution IV design would be appropriate. Table 14-30 shows that two resolution IV fractions are 73 available: the 272 IV with 32 runs and the 2IV with 16 runs. The

aliases involving main effects and two- and three-factor interactions for the 16-run design are presented in Table 14-31. Notice that all seven main effects are aliased with three-factor interactions. All the two-factor interactions are aliased in groups of three. Therefore, this design will satisfy our objectives; that is, it will allow the estimation of the main effects, and it will give some insight regarding two-factor interactions. It is not necessary to run the 272 IV design, which would require

JWCL232_c14_551-636.qxd

1/16/10

9:57 AM

Page 613

14-7 FRACTIONAL REPLICATION OF THE 2k DESIGN

Table 14-30 Number of Factors k

Selected 2kp Fractional Factorial Designs

Fraction

Number of Runs

Design Generators

3 4 5

231 III 241 IV 251 V 252 III

4 8 16 8

6

261 VI 262 IV

32 16

263 III

8

271 VII 272 IV

64 32

273 IV

16

274 III

8

C  AB D  ABC E  ABCD D  AB E  AC F  ABCDE E  ABC F  BCD D  AB E  AC F  BC G  ABCDEF F  ABCD G  ABDE E  ABC F  BCD G  ACD D  AB E  AC F  BC G  ABC G  ABCD H  ABEF F  ABC G  ABD H  BCDE E  BCD F  ACD G  ABC H  ABD H  ACDFG J  BCEFG G  ABCD H  ACEF J  CDEF F  BCDE G  ACDE H  ABDE J  ABCE E  ABC F  BCD G  ACD H  ABD J  ABCD

7

8

9

613

282 V

64

283 IV

32

284 IV

16

292 VI

128

293 IV

64

294 IV

32

295 III

16

Source: Montgomery (2009)

Number of Factors k

Fraction

Number of Runs

10 2103 V

128

2104 IV

64

2105 IV

32

2106 III

16

2115 IV

64

2116 IV

32

2117 III

16

11

Design Generators H  ABCG J  ACDE K  ACDF G  BCDF H  ACDF J  ABDE K  ABCE F  ABCD G  ABCE H  ABDE J  ACDE K  BCDE E  ABC F  BCD G  ACD H  ABD J  ABCD K  AB G  CDE H  ABCD J  ABF K  BDEF L  ADEF F  ABC G  BCD H  CDE J  ACD K  ADE L  BDE E  ABC F  BCD G  ACD H  ABD J  ABCD K  AB L  AC

JWCL232_c14_551-636.qxd

614

1/16/10

9:57 AM

Page 614

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

Table 14-31 Generators, Defining Relation, and Aliases for the 273 IV Fractional Factorial Design Generators and Defining Relation E  ABC, F  BCD, G  ACD I  ABCE  BCDF  ADEF  ACDG  BDEG  ABFG  CEFG Aliases A  BCE  DEF  CDG  BFG B  ACE  CDF  DEG  AFG C  ABE  BDF  ADG  EFG D  BCF  AEF  ACG  BEG E  ABC  ADF  BDG  CFG F  BCD  ADE  ABG  CEG G  ACD  BDE  ABF  CEF

AB  CE  FG AC  BE  DG AD  EF  CG AE  BC  DF AF  DE  BG AG  CD  BF BD  CF  EG

ABD  CDE  ACF  BEF  BCG  AEG  DFG

32 runs. The construction of the 273 IV design is shown in Table 14-32. Notice that it was constructed by starting with the 16run 24 design in A, B, C, and D as the basic design and then adding the three columns E  ABC, F  BCD, and G  ACD as suggested in Table 14-30. Thus, the generators for this design are I  ABCE, I  BCDF, and I  ACDG. The complete defining relation is I  ABCE  BCDF  ADEF  ACDG  BDEG  CEFG  ABFG. This defining relation was used to

produce the aliases in Table 14-31. For example, the alias relationship of A is A  BCE  ABCDF  DEF  CDG  ABDEG  ACEFG  BFG which, if we ignore interactions higher than three factors, agrees with Table 14-31.

For seven factors, we can reduce the number of runs even further. The 274 design is an eight-run experiment accommodating seven variables. This is a 1兾16th fraction and is obtained by first writing down a 23 design as the basic design in the factors A, B, and C, and then Table 14-32

A 273 IV Fractional Factorial Design Basic Design

Run

A

B

C

D

E  ABC

F  BCD

G  ACD

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

               

               

               

               

               

               

               

JWCL232_c14_551-636.qxd

1/16/10

9:57 AM

Page 615

14-7 FRACTIONAL REPLICATION OF THE 2k DESIGN

Table 14-33 A        

615

A 274 III Fractional Factorial Design B        

C        

D  AB        

E  AC        

F  BC        

G  ABC        

forming the four new columns from I  ABD, I  ACE, I  BCF, and I  ABCG, as suggested in Table 14-30. The design is shown in Table 14-33. The complete defining relation is found by multiplying the generators together two, three, and finally four at a time, producing I  ABD  ACE  BCF  ABCG  BCDE  ACDF  CDG  ABEF  BEG  AFG  DEF  ADEG  CEFG  BDFG  ABCDEFG The alias of any main effect is found by multiplying that effect through each term in the defining relation. For example, the alias of A is A  BD  CE  ABCF  BCG  ABCDE  CDF  ACDG  BEF  ABEG  FG  ADEF  DEG  ACEFG  ABDFG  BCDEFG This design is of resolution III, since the main effect is aliased with two-factor interactions. If we assume that all three-factor and higher interactions are negligible, the aliases of the seven main effects are /A  A  BD  CE  FG /B  B  AD  CF  EG /C  C  AE  BF  DG /D  D  AB  CG  EF /E  E  AC  BG  DF /F  F  BC  AG  DE /G  G  CD  BE  AF 74 This 2III design is called a saturated fractional factorial, because all the available degrees of freedom are used to estimate main effects. It is possible to combine sequences of these resolution III fractional factorials to separate the main effects from the two-factor interactions. The procedure is illustrated in Montgomery (2009) and in Box, Hunter, and Hunter (2005).

JWCL232_c14_551-636.qxd

616

1/16/10

1:27 PM

Page 616

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

EXERCISES FOR SECTION 14-7 14-41. Consider the problem in Exercise 14-17. Suppose that only half of the 32 runs could be made. (a) Choose the half that you think should be run. (b) Write out the alias relationships for your design. (c) Estimate the factor effects. (d) Plot the effect estimates on normal probability paper and interpret the results. (e) Set up an analysis of variance for the factors identified as potentially interesting from the normal probability plot in part (d). (f ) Analyze the residuals from the model. (g) Provide a practical interpretation of the results. 14-42. Suppose that in Exercise 14-20 it was possible to run only a 12 fraction of the 24 design. Construct the design and use only the data from the eight runs you have generated to perform the analysis. 14-43. An article by L. B. Hare [“In the Soup: A Case Study to Identify Contributors to Filling Variability,” Journal of Quality Technology (Vol. 20, pp. 36–43)] describes a factorial experiment used to study filling variability of dry soup mix packages. The factors are A  number of mixing ports through which the vegetable oil was added (1, 2), B  temperature surrounding the mixer (cooled, ambient), C  mixing time (60, 80 sec), D  batch weight (1500, 2000 lb), and E  number of days of delay between mixing and packaging (1, 7). Between 125 and 150 packages of soup were sampled over an eight-hour period for each run in the design, and the standard deviation of package weight was used as the response variable. The design and resulting data follow.

A Mixer Ports

B

C

Std Order

Time

D Batch Weight

Temp

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

               

               

E Delay

v Std Dev

               

               

               

1.13 1.25 0.97 1.7 1.47 1.28 1.18 0.98 0.78 1.36 1.85 0.62 1.09 1.1 0.76 2.1

(a) (b) (c) (d)

What is the generator for this design? What is the resolution of this design? Estimate the factor effects. Which effects are large? Does a residual analysis indicate any problems with the underlying assumptions? (e) Draw conclusions about this filling process. 14-44. Montgomery (2009) described a 241 fractional factorial design used to study four factors in a chemical process. The factors are A  temperature, B  pressure, C  concentration, and D  stirring rate, and the response is filtration rate. The design and the data are as follows:

Run

A

B

C

D  ABC

1 2 3 4 5 6 7 8

       

       

       

       

Treatment Combination 112 ad bd ab cd ac bc abcd

Filtration Rate 45 100 45 65 75 60 80 96

(a) Write down the alias relationships. (b) Estimate the factor effects. Which factor effects appear large? (c) Project this design into a full factorial in the three apparently important factors and provide a practical interpretation of the results. 14-45. R. D. Snee (“Experimenting with a Large Number of Variables,” in Experiments in Industry: Design, Analysis and Interpretation of Results, Snee, Hare, and Trout, eds., ASQC, 1985) described an experiment in which a 251 design with I  ABCDE was used to investigate the effects of five factors on the color of a chemical product. The factors are A  solvent/reactant, B  catalyst/reactant, C  temperature, D  reactant purity, and E  reactant pH. The results obtained are as follows: e   0.63 a  2.51 b   2.68 abe  1.66 c  2.06 ace  1.22 bce   2.09 abc  1.93

d ade bde abd cde acd bcd abcde

 6.79  6.47  3.45  5.68  5.22  4.38  4.30  4.05

(a) Prepare a normal probability plot of the effects. Which factors are active?

JWCL232_c14_551-636.qxd

1/16/10

9:58 AM

Page 617

14-7 FRACTIONAL REPLICATION OF THE 2k DESIGN

(b) Calculate the residuals. Construct a normal probability plot of the residuals and plot the residuals versus the fitted values. Comment on the plots. (c) If any factors are negligible, collapse the 251 design into a full factorial in the active factors. Comment on the resulting design, and interpret the results. 14-46. An article in Quality Engineering [“A Comparison of Multi-response Optimization: Sensitivity to Parameter Selection” (1999, Vol. 11, pp. 405–415)] conducted a half replicate of a 25 factorial design to optimize the retort process of beef stew MREs, a military ration. The design factors are x1  Sauce Viscosity, x2  Residual Gas, x 3  Solid / Liquid Ratio, x 4  Net Weight, x 5  Rotation Speed. The response variable is the heating rate index, a measure of heat penetration, and there are two replicates.

Run

x1

x2

x3

x4

x5

Heating Rate Index I II

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

8.46 15.68 14.94 12.52 17 11.44 10.45 19.73 17.37 14.98 8.4 19.08 13.07 18.57 20.59 14.03

9.61 14.68 13.09 12.71 16.36 11.83 9.22 16.94 16.36 11.93 8.16 15.40 10.55 20.53 21.19 11.31

(a) Estimate the factor effects. Based on a normal probability plot of the effect estimates, identify a model for the data from this experiment. (b) Conduct an ANOVA based on the model identified in part (a). What are your conclusions? (c) Analyze the residuals and comment on model adequacy. (d) Find a regression model to predict yield in terms of the coded factor levels. (e) This experiment was replicated, so an ANOVA could have been conducted without using a normal plot of the effects to tentatively identify a model. What model would be

617

appropriate? Use the ANOVA to analyze this model and compare the results with those obtained from the normal probability plot approach. 14-47. An article in Industrial and Engineering Chemistry [“More on Planning Experiments to Increase Research Efficiency” (1970, pp. 60–65)] uses a 252 design to investigate the effect on process yield of A  condensation temperature, B  amount of material 1, C  solvent volume, D  condensation time, and E  amount of material 2. The results obtained are as follows:

ae ab ad bc

   

23.2 15.5 16.9 16.2

cd ace bde abcde

   

23.8 23.4 16.8 18.1

(a) Verify that the design generators used were I  ACE and I  BDE. (b) Write down the complete defining relation and the aliases from the design. (c) Estimate the main effects. (d) Prepare an analysis of variance table. Verify that the AB and AD interactions are available to use as error. (e) Plot the residuals versus the fitted values. Also construct a normal probability plot of the residuals. Comment on the results. 14-48. Suppose that in Exercise 14-14 only a 1⁄4 fraction of the 25 design could be run. Construct the design and analyze the data that are obtained by selecting only the response for the eight runs in your design. 14-49. For each of the following designs write down the aliases, assuming that only main effects and two factor interactions are of interest. (a) 263 (b) 284 IV III 14-50. Consider the 262 design in Table 14-29. (a) Suppose that after analyzing the original data, we find that factors C and E can be dropped. What type of 2k design is left in the remaining variables? (b) Suppose that after the original data analysis, we find that factors D and F can be dropped. What type of 2k design is left in the remaining variables? Compare the results with part (a). Can you explain why the answers are different? 14-51. An article in the Journal of Radioanalytical and Nuclear Chemistry (2008, Vol. 276, No. 2, pp. 323–328) presented a 284 fractional factorial design to identify sources of Pu contamination in the radioactivity material analysis of dried shellfish at the National Institute of Standards and Technology (NIST). The data are shown in the following table. No contamination occurred at runs 1, 4, and 9.

JWCL232_c14_551-636.qxd

618

1/20/10

10:14 PM

Page 618

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

28⫺4

Glassware

Reagent

Sample Prep

Tracer

Dissolution

Hood

Chemistry

Ashing

Response, mBq

Run

x1

x2

x3

x4

x5

x6

x7

x8

y

1

⫺1

⫺1

⫺1

⫺1

⫺1

⫺1

⫺1

⫺1

0

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1

⫺1 ⫹1 ⫹1 ⫺1 ⫺1 ⫹1 ⫹1 ⫺1 ⫺1 ⫹1 ⫹1 ⫺1 ⫺1 ⫹1 ⫹1

⫺1 ⫺1 ⫺1 ⫹1 ⫹1 ⫹1 ⫹1 ⫺1 ⫺1 ⫺1 ⫺1 ⫹1 ⫹1 ⫹1 ⫹1

⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫹1 ⫹1 ⫹1 ⫹1 ⫹1 ⫹1 ⫹1 ⫹1

⫺1 ⫹1 ⫹1 ⫹1 ⫹1 ⫺1 ⫺1 ⫹1 ⫹1 ⫺1 ⫺1 ⫺1 ⫺1 ⫹1 ⫹1

⫹1 ⫺1 ⫹1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫺1 ⫹1 ⫺1 ⫹1

⫹1 ⫹1 ⫺1 ⫹1 ⫺1 ⫺1 ⫹1 ⫺1 ⫹1 ⫹1 ⫺1 ⫹1 ⫺1 ⫺1 ⫹1

⫹1 ⫹1 ⫺1 ⫺1 ⫹1 ⫹1 ⫺1 ⫹1 ⫺1 ⫺1 ⫹1 ⫹1 ⫺1 ⫺1 ⫹1

3.31 0.0373 0 0.0649 0.133 0.0461 0.0297 0 0.287 0.133 0.0476 0.133 5.75 0.0153 2.47

The factors and levels are shown in the following table. ⫺1

⫹1

Glassware

Distilled water

Soap, acid, stored

Reagent

New

Old

Sample prep

Co-precipitation

Electrodeposition

Tracer

Stock

Fresh

Dissolution

Without

With

Hood

B

A

Factor

Chemistry

Without

With

Ashing

Without

With

1 2

A

B

C

D

E

⫺1 ⫺1 ⫺1

1

1

1 ⫺1 ⫺1 ⫺1 ⫺1

3

⫺1

4

1

G

1 ⫺1

1 ⫺1

1

1 ⫺1 ⫺1

1 ⫺1

1

14.7

1 ⫺1 ⫺1

22.3

7

⫺1

1

1 ⫺1 ⫺1

1 ⫺1

16.1

8

1

1

1

1

22.1

1

1

1

⫺1

⫹1

A

Television advertising

No advertising

Advertising

B

Billboard advertising

No advertising

Advertising

C

Newspaper advertising

No advertising

Advertising

D

Candy wrapper design

Conservative design

Flashy design

E

Display design

Normal shelf display

Special aisle display

F

Free sample program

No free samples

Free samples

G

Size of candy bar

Factor

Sales for a 6-week period (in $1000)

1 oz. bar

2 1⁄2 oz. bar

8.7

1

1

1 ⫺1

1

9.7

1 ⫺1 ⫺1 ⫺1

11.3

1 ⫺1 ⫺1 1 ⫺1

F

6

⫺1 ⫺1

The factors and levels are shown in the following table.

(a) Write down the alias relationships. (b) Estimate the main effects. (c) Prepare a normal probability plot for the effects and interpret the results. 14-52. An article in the Journal of Marketing Research (1973, Vol. 10, No. 3, pp. 270–276) presented a 27⫺4 fractional factorial design to conduct marketing research:

Runs

5

15.1

(a) Write down the alias relationships. (b) Estimate the main effects. (c) Prepare a normal probability plot for the effects and interpret the results.

JWCL232_c14_551-636.qxd

1/16/10

9:58 AM

Page 619

619

14-8 RESPONSE SURFACE METHODS AND DESIGNS

14-8 RESPONSE SURFACE METHODS AND DESIGNS Response surface methodology, or RSM, is a collection of mathematical and statistical techniques that are useful for modeling and analysis in applications where a response of interest is influenced by several variables and the objective is to optimize this response. For example, suppose that a chemical engineer wishes to find the levels of temperature (x1) and feed concentration (x2) that maximize the yield (y) of a process. The process yield is a function of the levels of temperature and feed concentration, say, Y  f 1x1, x2 2   where  represents the noise or error observed in the response Y. If we denote the expected response by E(Y )  f (x1, x2)  , then the surface represented by  f 1x1, x2 2 is called a response surface. We may represent the response surface graphically as shown in Fig. 14-42, where is plotted versus the levels of x1 and x2. Notice that the response is represented as a surface plot in a three-dimensional space. To help visualize the shape of a response surface, we often plot the contours of the response surface as shown in Fig. 14-43. In the contour plot, lines of constant response are drawn in the x1, x2 plane. Each contour corresponds to a particular height of the response surface. The contour plot is helpful in studying the levels of x1 and x2 that result in changes in the shape or height of the response surface. In most RSM problems, the form of the relationship between the response and the independent variables is unknown. Thus, the first step in RSM is to find a suitable approximation for the true relationship between Y and the independent variables. Usually, a low-order polynomial in some region of the independent variables is employed. If the response is well modeled by a linear function of the independent variables, the approximating function is the first-order model Y  0  1x1  2 x2  p  k x k  

(14-21)

If there is curvature in the system, then a polynomial of higher degree must be used, such as the second-order model k

k

i1

i1

Y  0  a i xi  a ii x 2i  b ij x i xj  

Yield

74 64 54 44 34 100

120

140

160 Temperature, °C

3.0 2.6 2.2 Feed concentration, % 1.8 1.4 1.0 180

Figure 14-42 A three-dimensional response surface showing the expected yield as a function of temperature and feed concentration.

Feed concentration, %

3.0 84

(14-22)

i j

55

60

65 70

2.6 2.2

75

80

Current operating conditions

1.8

Region of the optimum

1.4 1.0 100

85

120

140 160 Temperature, °C

180

Figure 14-43 A contour plot of the yield response surface in Figure 14-42.

JWCL232_c14_551-636.qxd

620

1/16/10

9:58 AM

Page 620

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

Many RSM problems use one or both of these approximating polynomials. Of course, it is unlikely that a polynomial model will be a reasonable approximation of the true functional relationship over the entire space of the independent variables, but for a relatively small region they usually work quite well. The method of least squares, discussed in Chapters 11 and 12, is used to estimate the parameters in the approximating polynomials. The response surface analysis is then done in terms of the fitted surface. If the fitted surface is an adequate approximation of the true response function, analysis of the fitted surface will be approximately equivalent to analysis of the actual system. RSM is a sequential procedure. Often, when we are at a point on the response surface that is remote from the optimum, such as the current operating conditions in Fig. 14-43, there is little curvature in the system and the first-order model will be appropriate. Our objective here is to lead the experimenter rapidly and efficiently to the general vicinity of the optimum. Once the region of the optimum has been found, a more elaborate model such as the second-order model may be employed, and an analysis may be performed to locate the optimum. From Fig. 14-43, we see that the analysis of a response surface can be thought of as “climbing a hill,” where the top of the hill represents the point of maximum response. If the true optimum is a point of minimum response, we may think of “descending into a valley.” The eventual objective of RSM is to determine the optimum operating conditions for the system or to determine a region of the factor space in which operating specifications are satisfied. Also, note that the word “optimum” in RSM is used in a special sense. The “hill climbing” procedures of RSM guarantee convergence to a local optimum only. Method of Steepest Ascent Frequently, the initial estimate of the optimum operating conditions for the system will be far from the actual optimum. In such circumstances, the objective of the experimenter is to move rapidly to the general vicinity of the optimum. We wish to use a simple and economically efficient experimental procedure. When we are remote from the optimum, we usually assume that a first-order model is an adequate approximation to the true surface in a small region of the x’s. The method of steepest ascent is a procedure for moving sequentially along the path of steepest ascent, that is, in the direction of the maximum increase in the response. Of course, if minimization is desired, we are talking about the method of steepest descent. The fitted first-order model is k

yˆ  ˆ 0  a ˆ i xi

(14-23)

i1

and the first-order response surface, that is, the contours of yˆ , is a series of parallel lines such as that shown in Fig. 14-44. The direction of steepest ascent is the direction in which yˆ increases most rapidly. This direction is normal to the fitted response surface contours. We usually take as the path of steepest ascent the line through the center of the region of interest and normal to the fitted surface contours. Thus, the steps along the path are proportional to the regression coefficients 5ˆ i 6 . The experimenter determines the actual step size based on process knowledge or other practical considerations. Experiments are conducted along the path of steepest ascent until no further increase in response is observed. Then a new first-order model may be fit, a new direction of steepest ascent determined, and further experiments conducted in that direction until the experimenter feels that the process is near the optimum.

JWCL232_c14_551-636.qxd

1/16/10

1:29 PM

Page 621

14-8 RESPONSE SURFACE METHODS AND DESIGNS

621

x2

Path of steepest ascent Region of fitted first-order response surface ∧

y = 50

∧

Figure 14-44 Firstorder response surface and path of steepest ascent.

y = 40

∧

y = 20

∧

y = 10

∧

y = 30 x1

EXAMPLE 14-11 Process Yield Steepest Ascent In Example 14-6 we described an experiment on a chemical process in which two factors, reaction time (x1) and reaction temperature (x2), affect the percent conversion or yield (Y ). Figure 14-27 shows the 22 design plus five center points used in this study. The engineer found that both factors were important, there was no interaction, and there was no curvature in the response surface. Therefore, the first-order model

yˆ ⫽ 40.44 ⫹ 0.775x1 ⫹ 0.325x2 Figure 14-45(a) and (b) show the contour plot and threedimensional surface plot of this model. Figure 14-45 also

Y ⫽ ␤0 ⫹ ␤1 x1 ⫹ ␤2 x2 ⫹ ⑀

1

should be appropriate. Now the effect estimate of time is 1.55 hours and the effect estimate of temperature is 0.65⬚F, and since the regression coefficients ␤ˆ 1 and ␤ˆ 2 are one-half of the corresponding effect estimates, the fitted first-order model is

160.0 41.50

x2 (temperature)

158.3

45.00

41.00

43.00

156.7 40.50 0

155.0

150.0 30.00 –1

41.00

39.00

40.00

153.3

151.7 –1

Conversion

160.0 +1

39.50

31.67

33.33

35.00 0 x1 (time)

36.67

38.33

40.00 +1

40.00 158.0 38.00 +1 156.0 36.00 0 154.0 34.00 0 152.0 32.00 x2 (temperature) x1 (time) 150.0 30.00 –1 –1

Contour plot

Three-dimensional surface plot

(a)

(b)

Figure 14-45 Response surface plots for the first-order model in Example 14-11.

JWCL232_c14_551-636.qxd

622

1/16/10

9:58 AM

Page 622

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS 170 Original region of experimentation

Temperature

160 A

B

F E C

D

Point Point Point Point Point Point

Path of steepest ascent

150

A: 40 minutes, 157°F, y = 40.5 B: 45 minutes, 159°F, y = 51.3 C: 50 minutes, 161°F, y = 59.6 D: 55 minutes, 163°F, y = 67.1 E: 60 minutes, 165°F, y = 63.6 F: 65 minutes, 167°F, y = 60.7

Original fitted contours

30

40

50

60

70

Time

Figure 14-46 Steepest ascent experiment for Example 14-11. (0.325兾0.775) x1  0.42. A change of  x2  0.42 in the coded variable x2 is equivalent to about 2F in the original variable temperature. Therefore, the engineer will move along the path of steepest ascent by increasing reaction time by 5 minutes and temperature by 2F. An actual observation on yield will be determined at each point. Next Steps: Figure 14-46 shows several points along this path of steepest ascent and the yields actually observed from the process at those points. At points A–D the observed yield increases steadily, but beyond point D, the yield decreases. Therefore, steepest ascent would terminate in the vicinity of 55 minutes of reaction time and 163F with an observed percent conversion of 67%.

shows the relationship between the coded variables x1 and x2 (that defined the high and low levels of the factors) and the original variables, time (in minutes) and temperature (in F). From examining these plots (or the fitted model), we see that to move away from the design center—the point (x1  0, x2  0)—along the path of steepest ascent, we would move 0.775 unit in the x1 direction for every 0.325 unit in the x2 direction. Thus, the path of steepest ascent passes through the point (x1  0, x2  0) and has a slope 0.325兾0.775. The engineer decides to use 5 minutes of reaction time as the basic step size. Now, 5 minutes of reaction time is equivalent to a step in the coded variable x1 of x1  1. Therefore, the steps along the path of steepest ascent are x1  1.0000 and  x2 

Analysis of a Second-Order Response Surface When the experimenter is relatively close to the optimum, a second-order model is usually required to approximate the response because of curvature in the true response surface. The fitted second-order model is k

k

i1

i1

ˆ xx yˆ  ˆ 0  a ˆ i xi  a ˆ ii x2i  b  ij i j i j

where ˆ denotes the least squares estimate of . In this section we show how to use this fitted model to find the optimum set of operating conditions for the x’s and to characterize the nature of the response surface. EXAMPLE 14-12 Process Yield Central Composite Design Continuation of Example 14-11 Consider the chemical process from Example 14-11, where the method of steepest ascent terminated at a reaction time of 55 minutes and a temperature of 163F. The experimenter decides to fit a second-order model in this region. Table 14-34 and Fig. 14-47 show the experimental design, which consists of

a 22 design centered at 55 minutes and 165F, five center points, and four runs along the coordinate axes called axial runs. This type of design is called a central composite design, and it is a very popular design for fitting second-order response surfaces. Two response variables were measured during this phase of the experiment: percentage conversion (yield) and

JWCL232_c14_551-636.qxd

1/16/10

9:58 AM

Page 623

623

14-8 RESPONSE SURFACE METHODS AND DESIGNS

Table 14-34

Central Composite Design for Example 14-12

Observation Number

Time (minutes)

Temperature (F)

1 2 3 4 5 6 7 8 9 10 11 12 13

50 60 50 60 48 62 55 55 55 55 55 55 55

160 160 170 170 165 165 158 172 165 165 165 165 165

Coded Variables x1 x2 1 1 1 1 1.414 1.414 0 0 0 0 0 0 0

1 1 1 1 0 0 1.414 1.414 0 0 0 0 0

Conversion (percent) Response 1

Viscosity (mPa-sec) Response 2

65.3 68.2 66 69.8 64.5 69 64 68.5 68.9 69.7 68.5 69.4 69

35 39 36 43 30 44 31 45 37 34 35 36 37

viscosity. The least-squares quadratic model for the yield response is

The viscosity response is adequately described by the first-order model

yˆ  69.1  1.633x1 1.083x2  0.969x21  1.219x22  0.225x1x2

yˆ 2  37.08  3.85 x1  3.10x2

The analysis of variance for this model is shown in Table 14-35. Figure 14-48 shows the response surface contour plot and the three-dimensional surface plot for this model. From examination of these plots, the maximum yield is about 70%, obtained at approximately 60 minutes of reaction time and 167F.

Table 14-36 summarizes the analysis of variance for this model. The response surface is shown graphically in Fig. 14-49. Notice that viscosity increases as both time and temperature increase. Practical Interpretation: As in most response surface problems, the experimenter in this example had conflicting objectives regarding the two responses. The objective was to maximize yield, but the acceptable range for viscosity was

x2 +2 (0, 1.414) (1, 1)

(–1, 1)

(–1.414, 0)

(1.414, 0)

–2

(0, 0)

(1, –1)

(–1, –1) (0, –1.414)

Figure 14-47 Central composite design for Example 14-12.

+2

–2

x1

JWCL232_c14_551-636.qxd

624

1/16/10

9:58 AM

Page 624

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

Table 14-35

Analysis of Variance for the Quadratic Model, Yield Response

Source of Variation

Sum of Squares

Degrees of Freedom

Mean Square

Model Residual Total

45.89 4.30 50.19

5 7 12

9.178 0.615

Independent Variable Intercept x1 x2 x21 x22 x1x2

1 172.0

f0

P-Value

14.93

0.0013

Coefficient Estimate

Standard Error

t for H0 Coefficient  0

P-Value

69.100 1.633 1.083 0.969 1.219 0.225

0.351 0.277 0.277 0.297 0.297 0.392

197.1 5.891 3.907 3.259 4.100 0.5740

0.0006 0.0058 0.0139 0.0046 0.5839

64.00

x2 (temperature)

65.00 169.7

70.12

167.3

67.23 70.00

0 165.0

69.00

Conversion

68.00 162.7

160.3

61.45

67.00 66.00

172.0

65.00 64.00 63.00

–1 158.0 48.00 –1

50.33

64.34

62.00 169.2 59.20 +1 166.4 56.40 0 163.6 53.60 0 160.8 50.80 x2 (temperature) x1 (time) 158.0 48.00 –1 –1 +1

52.67

55.00 0 x1 (time)

57.33

59.67

62.00 +1

Contour plot

Surface plot

(a)

(b)

Figure 14-48 Response surface plots for the yield response, Example 14-12.

Table 14-36

Analysis of Variance for the First-Order Model, Viscosity Response

Source of Variation

Sum of Squares

Degrees of Freedom

Mean Square

Model Residual Total

195.4 61.5 256.9

2 10 12

97.72 6.15

Independent Variable Intercept x1 x2

f0

P-Value

15.89

0.0008

Coefficient Estimate

Degrees of Freedom

Standard Error

t for H0 Coefficient  0

P-Value

37.08 3.85 3.10

1 1 1

0.69 0.88 0.88

53.91 4.391 3.536

0.0014 0.0054

JWCL232_c14_551-636.qxd

1/16/10

9:58 AM

Page 625

14-8 RESPONSE SURFACE METHODS AND DESIGNS

1

625

172.0 46.00

x2 (temperature)

169.7

46.81

44.00 42.00

167.3

40.32

40.00 0

38.00

165.0

Viscosity

36.00 27.35

162.7

34.00 32.00

160.3

172.0 +1

x1 (time)

62.00 169.2 59.20 +1 166.4 56.40 0 163.6 53.60 0 160.8 x2 (temperature) 50.80 x1 (time) 158.0 48.00 –1 –1

Contour plot

Surface plot

(a)

(b)

30.00 –1

33.83

158.0 48.00 –1

50.33

52.67

55.00 0

57.33

59.67

62.00 +1

Figure 14-49 Response surface plots for the viscosity response, Example 14-12. conversion, y2  38, and y2  42 highlighted. The shaded areas on this plot identify unfeasible combinations of time and temperature. This graph shows that several combinations of time and temperature will be satisfactory.

38 y2 42. When there are only a few independent variables, an easy way to solve this problem is to overlay the response surfaces to find the optimum. Figure 14-50 shows the overlay plot of both responses, with the contours y1  69%

Example 14-12 illustrates the use of a central composite design (CCD) for fitting a second-order response surface model. These designs are widely used in practice because they are relatively efficient with respect to the number of runs required. In general, a CCD in k factors requires 2k factorial runs, 2k axial runs, and at least one center point (three to five center points are typically used). Designs for k  2 and k  3 factors are shown in Fig. 14-51. The central composite design may be made rotatable by proper choice of the axial spacing  in Fig. 14-51. If the design is rotatable, the standard deviation of predicted response 1

172.0

169.7

Viscosity 42.00

x2 (temperature)

Viscosity 38.00 167.3

0

165.0 Conversion 69.00 162.7

160.3

Figure 14-50 Overlay of yield and viscosity response surfaces, Example 14-12.

–1

158.0 48.00 –1

50.33

52.67

55.00 0 x1 (time)

57.33

59.67

62.00 +1

JWCL232_c14_551-636.qxd

626

1/16/10

9:58 AM

Page 626

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS x2

x2 (0, α )

(–1, +1)

(+1, +1)

(0, 0)

(– α , 0)

(α , 0)

(–1, –1)

x2

x1

x1

(+1, –1)

Figure 14-51 Central composite designs for k  2 and k  3.

(0, – α )

yˆ is constant at all points that are the same distance from the center of the design. For rotatability, choose   (F)1兾4, where F is the number of points in the factorial part of the design (usually F  2k ). For the case of k  2 factors,   (22)1兾4  1.414, as was used in the design in Example 14-12. Figure 14-52 presents a contour plot and a surface plot of the standard deviation of prediction for the quadratic model used for the yield response. Notice that the contours are concentric circles, implying that yield is predicted with equal precision for all points that are the same distance from the center of the design. Also, as one would expect, the precision decreases with increasing distance from the design center.

172.0

1.000 0.8500 169.7 0.7000

1.000 0.8500 0.7000

1.140

0.8741

Temperature

167.3

∧ std(y)

165.0 0.3500

160.3

0.3419

0.3500

162.7

0.4000 0.5000 0.8500 0.6000 0.7000

158.0 48.00

Figure 14-52

50.33

52.67

0.6080

172.0 0.7000 0.8500 1.000 55.00 Time

57.33

59.67

62.00

62.00 169.2 59.20 166.4 56.40 163.6 53.60 160.8 Time 50.80 Temperature 158.0 48.00

Contour plot

Surface plot

(a)

(b)

Plots of constant 2V1yˆ 2 for a rotatable central composite design.

EXERCISES FOR SECTION 14-8 14-53. An article in Rubber Age (Vol. 89, 1961, pp. 453– 458) describes an experiment on the manufacture of a product in which two factors were varied. The factors are reaction time (hr)

and temperature (C). These factors are coded as x1  (time  12)兾8 and x2  (temperature  250)兾30. The following data were observed where y is the yield (in percent):

JWCL232_c14_551-636.qxd

1/16/10

9:58 AM

Page 627

627

14-8 RESPONSE SURFACE METHODS AND DESIGNS

14-55.

Consider the first-order model

Run Number

x1

x2

y

yˆ  50  1.5x1  0.8x2

1 2 3 4 5 6 7 8 9 10 11 12

1 1 0 0 0 0 0 1.414 1.414 1.414 1.414 0

0 0 0 0 1 1 0 1.414 1.414 1.414 1.414 0

83.8 81.7 82.4 82.9 84.7 75.9 81.2 81.3 83.1 85.3 72.7 82.0

where 1 xi 1. Find the direction of steepest ascent. 14-56. A manufacturer of cutting tools has developed two empirical equations for tool life ( y1) and tool cost ( y2). Both models are functions of tool hardness (x1) and manufacturing time (x2). The equations are

(a) Plot the points at which the experimental runs were made. (b) Fit a second-order model to the data. Is the second-order model adequate? (c) Plot the yield response surface. What recommendations would you make about the operating conditions for this process? 14-54. An article in Quality Engineering [“Mean and Variance Modeling with Qualitative Responses: A Case Study” (1998– 1999, Vol. 11, pp. 141–148)] studied how three active ingredients of a particular food affect the overall taste of the product. The measure of the overall taste is the overall mean liking score (MLS). The three ingredients are identified by the variables x1, x2, and x3. The data are shown in the following table. Run

x1

x2

x3

MLS

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

1 1 0 0 1 1 0 0 1 0 1 1 1 1 1

1 1 0 1 1 1 0 1 1 0 0 1 1 1 0

1 1 0 0 1 1 1 0 1 1 0 1 1 1 0

6.3261 6.2444 6.5909 6.3409 5.907 6.488 5.9773 6.8605 6.0455 6.3478 6.7609 5.7727 6.1805 6.4894 6.8182

(a) Fit a second-order response surface model to the data. (b) Construct contour plots and response surface plots for MLS. What are your conclusions? (c) Analyze the residuals from this experiment. Does your analysis indicate any potential problems? (d) This design has only a single center point. Is this a good design in your opinion?

yˆ 1  10  5x1  2x2 yˆ 2  23  3x1  4x2 and both equations are valid over the range 1.5 xi 1.5. Suppose that tool life must exceed 12 hours and cost must be below $27.50. (a) Is there a feasible set of operating conditions? (b) Where would you run this process? 14-57. An article in Tappi (1960, Vol. 43, pp. 38–44) describes an experiment that investigated the ash value of paper pulp (a measure of inorganic impurities). Two variables, temperature T in degrees Celsius and time t in hours, were studied, and some of the results are shown in the following table. The coded predictor variables shown are x1 

1T  7752 115

,

x2 

1t  32 1.5

and the response y is (dry ash value in %)  103. x1

x2

y

x1

x2

y

1 1 1 1 1.5 1.5

1 1 1 1 0 0

211 92 216 99 222 48

0 0 0 0 0 0

1.5 1.5 0 0 0 0

168 179 122 175 157 146

(a) What type of design has been used in this study? Is the design rotatable? (b) Fit a quadratic model to the data. Is this model satisfactory? (c) If it is important to minimize the ash value, where would you run the process? 14-58. In their book Empirical Model Building and Response Surfaces (John Wiley, 1987), Box and Draper described an experiment with three factors. The data shown in the following table are a variation of the original experiment on page 247 of their book. Suppose that these data were collected in a semiconductor manufacturing process. (a) The response y1 is the average of three readings on resistivity for a single wafer. Fit a quadratic model to this response. (b) The response y2 is the standard deviation of the three resistivity measurements. Fit a linear model to this response. (c) Where would you recommend that we set x1, x2, and x3 if the objective is to hold mean resistivity at 500 and minimize the standard deviation?

JWCL232_c14_551-636.qxd

628

1/19/10

9:16 PM

Page 628

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

x1

x2

x3

y1

y2

x1

x2

x3

y1

y2

⫺1 0 1 ⫺1 0 1 ⫺1 0 1 ⫺1 0 1 ⫺1 0

⫺1 ⫺1 ⫺1 0 0 0 1 1 1 ⫺1 ⫺1 ⫺1 0 0

⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 0 0 0 0 0

24.00 120.33 213.67 86.00 136.63 340.67 112.33 256.33 271.67 81.00 101.67 357.00 171.33 372.00

12.49 8.39 42.83 3.46 80.41 16.17 27.57 4.62 23.63 0.00 17.67 32.91 15.01 0.00

1 ⫺1 0 1 ⫺1 0 1 ⫺1 0 1 ⫺1 0 1

0 1 1 1 ⫺1 ⫺1 ⫺1 0 0 0 1 1 1

0 0 0 0 1 1 1 1 1 1 1 1 1

501.67 264.00 427.00 730.67 220.67 239.67 422.00 199.00 485.33 673.67 176.67 501.00 1010.00

92.50 63.50 88.61 21.08 133.82 23.46 18.52 29.44 44.67 158.21 55.51 138.94 142.45

14-59.

Consider the first-order model y ⫽ 12 ⫹ 1.2x1 ⫺ 2.1x2 ⫹ 1.6x3 ⫺ 0.6x4

where ⫺1 ⱕ xi ⱕ 1. (a) Find the direction of steepest ascent. (b) Assume the current design is centered at the point (0, 0, 0, 0). Determine the point that is three units from the current center point in the direction of steepest ascent.

Trial 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

14-60. An article in the Journal of Materials Processing Technology (1997, Vol. 67, pp. 55–61) used response surface methodology to generate surface roughness prediction models for turning EN 24T steel (290 BHN). The data are shown in the following table.

Speed, V (m min⫺1)

Feed, f (mm rev⫺1)

Depth of cut, d (mm)

x1

Coding x2

x3

Surface roughness, Ra (␮m)

36 117 36 117 36 117 36 117 65 65 65 65 28 150 65 65 65 65 28 150 65 65 65 65

0.15 0.15 0.40 0.40 0.15 0.15 0.40 0.40 0.25 0.25 0.25 0.25 0.25 0.25 0.12 0.50 0.25 0.25 0.25 0.25 0.12 0.50 0.25 0.25

0.50 0.50 0.50 0.50 1.125 1.125 1.125 1.125 0.75 0.75 0.75 0.75 0.75 0.75 0.75 0.75 0.42 1.33 0.75 0.75 0.75 0.75 0.42 1.33

⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1 0 0 0 0 ⫺12 12 0 0 0 0 ⫺12 12 0 0 0 0

⫺1 ⫺1 1 1 ⫺1 ⫺1 1 1 0 0 0 0 0 0 ⫺12 12 0 0 0 0 ⫺12 12 0 0

⫺1 ⫺1 ⫺1 ⫺1 1 1 1 1 0 0 0 0 0 0 0 0 ⫺12 12 0 0 0 0 ⫺12 12

1.8 1.233 5.3 5.067 2.133 1.45 6.233 5.167 2.433 2.3 2.367 2.467 3.633 2.767 1.153 6.333 2.533 3.20 3.233 2.967 1.21 6.733 2.833 3.267

JWCL232_c14_551-636.qxd

1/16/10

1:35 PM

Page 629

629

14-8 RESPONSE SURFACE METHODS AND DESIGNS

The factors and levels for the experiment are shown in the following table. (a) Plot the points at which the experimental runs were made.

(b) Fit both first-and second-order models to the data. Comment on the adequacies of these models. (c) Plot the roughness response surface for the second-order model and comment.

Levels

Lowest

Low

Center

High

Highest

Coding

12

1

0

1

12

Speed, V (m min1)

28

36

65

117

150

Feed, f (mm rev 1)

0.12

0.15

0.25

0.40

0.50

Depth of cut, d (mm)

0.42

0.50

0.75

1.125

1.33

Supplemental Exercises

Time (minutes) Gear Type

14-61. An article in Process Engineering (1992, No. 71, pp. 46–47) presents a two-factor factorial experiment used to investigate the effect of pH and catalyst concentration on product viscosity (cSt). The data are as follows:

90

120

20-tooth

0.0265 0.0340

0.0560 0.0650

24-tooth

0.0430 0.0510

0.0720 0.0880

28-tooth

0.0405 0.0575

0.0620 0.0825

Catalyst Concentration pH

2.5

2.7

5.6 5.9

192, 199, 189, 198 185, 193, 185, 192

178, 186, 179, 188 197, 196, 204, 204

(a) Test for main effects and interactions using   0.05. What are your conclusions? (b) Graph the interaction and discuss the information provided by this plot. (c) Analyze the residuals from this experiment. 14-62. Heat treating of metal parts is a widely used manufacturing process. An article in the Journal of Metals (Vol. 41, 1989) describes an experiment to investigate flatness distortion from heat treating for three types of gears and two heat-treating times. The data are as follows: (a) Is there any evidence that flatness distortion is different for the different gear types? Is there any indication that

heat treating time affects the flatness distortion? Do these factors interact? Use   0.05. (b) Construct graphs of the factor effects that aid in drawing conclusions from this experiment. (c) Analyze the residuals from this experiment. Comment on the validity of the underlying assumptions. 14-63. An article in the Textile Research Institute Journal (1984, Vol. 54, pp. 171–179) reported the results of an experiment that studied the effects of treating fabric with selected inorganic salts on the flammability of the material. Two application levels of each salt were used, and a vertical burn test was used on each sample. (This finds the temperature at which each sample ignites.) The burn test data follow. Salt

Level

Untreated

MgCl2

NaCl

CaCO3

CaCl2

Na2CO3

1

812 827 876

752 728 764

739 731 726

733 728 720

725 727 719

751 761 755

2

945 881 919

794 760 757

741 744 727

786 771 779

756 781 814

910 854 848

(a) Test for differences between salts, application levels, and interactions. Use   0.01. (b) Draw a graph of the interaction between salt and application level. What conclusions can you draw from this graph?

(c) Analyze the residuals from this experiment. 14-64. An article in the IEEE Transactions on Components, Hybrids, and Manufacturing Technology (1992, Vol. 15) describes an experiment for investigating a method for aligning

JWCL232_c14_551-636.qxd

630

1/16/10

9:58 AM

Page 630

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

optical chips onto circuit boards. The method involves placing solder bumps onto the bottom of the chip. The experiment used three solder bump sizes and three alignment methods. The response variable is alignment accuracy (in micrometers). The data are as follows:

Alignment Method

Solder Bump Size (diameter in m)

1

2

3

75

4.60 4.53

1.55 1.45

1.05 1.00

130

2.33 2.44

1.72 1.76

0.82 0.95

260

4.95 4.55

2.73 2.60

2.36 2.46

(a) Is there any indication that either solder bump size or alignment method affects the alignment accuracy? Is there any evidence of interaction between these factors? Use   0.05. (b) What recommendations would you make about this process? (c) Analyze the residuals from this experiment. Comment on model adequacy. 14-65. An article in Solid State Technology (Vol. 29, 1984, pp. 281–284) describes the use of factorial experiments in photolithography, an important step in the process of manufacturing integrated circuits. The variables in this experiment (all at two levels) are prebake temperature (A), prebake time (B), and exposure energy (C), and the response variable is delta line width, the difference between the line on the mask and the printed line on the device. The data are as follows: (1)  2.30, a  9.87, b  18.20, ab  30.20, c  23.80, ac  4.30, bc  3.80, and abc  14.70. (a) Estimate the factor effects. (b) Use a normal probability plot of the effect estimates to identity factors that may be important. (c) What model would you recommend for predicting the delta line width response, based on the results of this experiment? (d) Analyze the residuals from this experiment, and comment on model adequacy. 14-66. An article in the Journal of Coatings Technology (Vol. 60, 1988, pp. 27–32) describes a 24 factorial design used for studying a silver automobile basecoat. The response variable is distinctness of image (DOI). The variables used in the experiment are A  Percentage of polyester by weight of polyester/melamine (low value  50%, high value  70%) B  Percentage of cellulose acetate butyrate carboxylate (low value  15%, high value  30%)

C  Percentage of aluminum stearate (low value  1%, high value  3%) D  Percentage of acid catalyst (low value  0.25%, high value  0.50%) The responses are (1)  63.8, a  77.6, b  68.8, ab  76.5, c  72.5, ac  77.2, bc  77.7, abc  84.5, d  60.6, ad  64.9, bd  72.7, abd  73.3, cd  68.0, acd  76.3, bcd  76.0, and abcd  75.9. (a) Estimate the factor effects. (b) From a normal probability plot of the effects, identify a tentative model for the data from this experiment. (c) Using the apparently negligible factors as an estimate of error, test for significance of the factors identified in part (b). Use   0.05. (d) What model would you use to describe the process, based on this experiment? Interpret the model. (e) Analyze the residuals from the model in part (d) and comment on your findings. 14-67. An article in the Journal of Manufacturing Systems (Vol. 10, 1991, pp. 32– 40) describes an experiment to investigate the effect of four factors, P  waterjet pressure, F  abrasive flow rate, G  abrasive grain size, and V  jet traverse speed, on the surface roughness of a waterjet cutter. A 24 design follows. (a) Estimate the factor effects. (b) Form a tentative model by examining a normal probability plot of the effects. (c) Is the model in part (b) a reasonable description of the process? Is lack of fit significant? Use   0.05. (d) Interpret the results of this experiment. (e) Analyze the residuals from this experiment.

Factors Surface V F P G Roughness Run (in/min) (lb/min) (kpsi) (Mesh No.) (m ) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

6 2 6 2 6 2 6 2 6 2 6 2 6 2 6 2

2.0 2.0 2.0 2.0 1.0 1.0 1.0 1.0 2.0 2.0 2.0 2.0 1.0 1.0 1.0 1.0

38 38 30 30 38 38 30 30 38 38 30 30 38 38 30 30

80 80 80 80 80 80 80 80 170 170 170 170 170 170 170 170

104 98 103 96 137 112 143 129 88 70 110 110 102 76 98 68

JWCL232_c14_551-636.qxd

1/16/10

9:58 AM

Page 631

14-8 RESPONSE SURFACE METHODS AND DESIGNS

14-68. Construct a 241 IV design for the problem in Exercise 14-66. Select the data for the eight runs that would have been required for this design. Analyze these runs and compare your conclusions to those obtained in Exercise 14-66 for the full factorial. 14-69. Construct a 241 IV design for the problem in Exercise 14-67. Select the data for the eight runs that would have been required for this design. Analyze these data and compare your conclusions to those obtained in Exercise 14-67 for the full factorial. 14-70. Construct a 284 IV design in 16 runs. What are the alias relationships in this design? 14-71. Construct a 252 III design in eight runs. What are the alias relationships in this design? 14-72. An article in the Journal of Quality Technology (Vol. 17, 1985, pp. 198–206) describes the use of a replicated fractional

factorial to investigate the effect of five factors on the free height of leaf springs used in an automotive application. The factors are A  furnace temperature, B  heating time, C  transfer time, D  hold down time, and E  quench oil temperature. The data are shown in the following table. (a) What is the generator for this fraction? Write out the alias structure. (b) Analyze the data. What factors influence mean free height? (c) Calculate the range of free height for each run. Is there any indication that any of these factors affect variability in free height? (d) Analyze the residuals from this experiment and comment on your findings.

A

B

C

D

E

               

               

               

               

               

14-73. An article in Rubber Chemistry and Technology (Vol. 47, 1974, pp. 825–836) describes an experiment that studies the Mooney viscosity of rubber to several variables, including silica filler (parts per hundred) and oil filler (parts per hundred). Data typical of that reported in this experiment follow, where x1 

silica  60 , 15

x2 

oil  21 15

(a) What type of experimental design has been used? (b) Analyze the data and draw appropriate conclusions.

631

Free Height 7.78 8.15 7.50 7.59 7.54 7.69 7.56 7.56 7.50 7.88 7.50 7.63 7.32 7.56 7.18 7.81

7.78 8.18 7.56 7.56 8.00 8.09 7.52 7.81 7.56 7.88 7.56 7.75 7.44 7.69 7.18 7.50

7.81 7.88 7.50 7.75 7.88 8.06 7.44 7.69 7.50 7.44 7.50 7.56 7.44 7.62 7.25 7.59

Coded levels x1

x2

y

1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1

13.71 14.15 12.87 13.53 13.90 14.88 12.25 13.35

JWCL232_c14_551-636.qxd

632

1/16/10

9:58 AM

Page 632

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

14-74. An article in Tropical Science [“Proximate Composition of the Seeds of Acacia Nilotica var Adansonti (Bagaruwa) and Extraction of Its Protein” (1992, Vol. 32, No. 3, pp. 263–268)] reported on research extracting and concentrating the proteins of the bagaruwa seeds in livestock feeding in Nigeria to eliminate the toxic substances from the seeds. The following are the effects of extraction time and flour to solvent ratio on protein extractability of the bagaruwa seeds in distilled water:

with center points was used to verify the most significant factors affecting the nisin recovery. The factor x1 was the concentration (% w/w) of PEG 4000 and x2 was the concentration (% w/w) of Na2SO4. The range and levels of the variables investigated in this study are presented below. Nisin extraction is a ratio representing the concentration of nisin and this was the response y. Trial

x1

x2

y

1 2 3 4 5 6 7 8 9

13 15 13 15 14 14 14 14 14

11 11 13 13 12 12 12 12 12

62.874 76.133 87.467 102.324 76.187 77.523 76.782 77.438 78.742

Percentage of Protein Extracted at Time (min) 30 60 90 120

Flour: Solvent Ratio (w/v) (%) 3

30.5 36.9

45.7 44.3

30.5 29.5

31.0 22.1

7

32.9 37.5

42.4 40.9

28.2 27.3

23.5 34.1

11

29.0 32.7

39.5 43.6

29.0 30.5

29.0 28.4

All values are means of three determinations. (a) Test the appropriate hypotheses and draw conclusions using the analysis of variance with   0.5. (b) Graphically analyze the interaction. (c) Analyze the residuals from this experiment. 14-75. An article in Plant Disease [“Effect of Nitrogen and Potassium Fertilizer Rates on Severity of Xanthomonas Blight of Syngonium Podophyllum” (1989, Vol. 73, No. 12, pp. 972–975)] showed the effect of the variable nitrogen and potassium rates on the growth of “White Butterfly” and the mean percentage of disease. Data representative of that collected in this experiment is provided in the following table.

Nitrogen (mg/pot/wk)

30

Potassium (mg/pot/wk) 90 120

50

61.0

61.3

45.5

42.5

59.5

58.2

150

54.5

55.9

53.5

51.9

34.0

35.9

250

42.7

40.4

36.5

37.4

32.5

33.8

(a) State the appropriate hypotheses. (b) Use the analysis of variance to test these hypotheses with   0.05. (c) Graphically analyze the residuals from this experiment. (d) Estimate the appropriate variance components. 14-76. An article in Biotechnology Progress (2001, Vol. 17, pp. 366–368) reported on an experiment to investigate and optimize the operating conditions of the nisin extraction in aqueous two-phase systems (ATPS). A 22 full factorial design

(a) Compute an ANOVA table for the effects and test for curvature with   0.05. Is curvature important in this region of the factors? (b) Calculate residuals from the linear model and test for adequacy of your model. (c) In a new region of factor space a central composite design (CCD) was used to perform second order optimization. The results are shown in the following table. Fit a second order model to this data and make conclusions. Coded Trial 1 2 3 4 5 6 7 8 9 10 11 12 13

Uncoded

x1

x2

x1

x2

y

1 1 1 1 1.414 1.414 0 0 0 0 0 0 0

1 1 1 1 0 0 1.414 1.414 0 0 0 0 0

15 16 15 16 14.793 16.207 15.5 15.5 15.5 15.5 15.5 15.5 15.5

14 14 16 16 15 15 13.586 16.414 15 15 15 15 15

102.015 106.868 108.13 110.176 105.236 110.289 103.999 110.171 108.044 109.098 107.824 108.978 109.169

14-77. An article in the Journal of Applied Electrochemistry (May 2008, Vol. 38, No. 5, pp. 583–590) presented a 273 fractional factorial design to perform optimization of polybenzimidazole-based membrane electrode assemblies for H2 / O2 fuel cells. The design and data are shown in the following table.

JWCL232_c14_551-636.qxd

1/19/10

9:08 PM

Page 633

14-8 RESPONSE SURFACE METHODS AND DESIGNS

Runs

A

B

C

D

E

F

G

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1

⫺1 ⫺1 ⫹1 ⫹1 ⫺1 ⫺1 ⫹1 ⫹1 ⫺1 ⫺1 ⫹1 ⫹1 ⫺1 ⫺1 ⫹1 ⫹1

⫺1 ⫺1 ⫺1 ⫺1 ⫹1 ⫹1 ⫹1 ⫹1 ⫺1 ⫺1 ⫺1 ⫺1 ⫹1 ⫹1 ⫹1 ⫹1

⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫺1 ⫹1 ⫹1 ⫹1 ⫹1 ⫹1 ⫹1 ⫹1 ⫹1

⫺1 ⫹1 ⫹1 ⫺1 ⫹1 ⫺1 ⫺1 ⫹1 ⫺1 ⫹1 ⫹1 ⫺1 ⫹1 ⫺1 ⫺1 ⫹1

⫺1 ⫹1 ⫹1 ⫺1 ⫺1 ⫹1 ⫹1 ⫺1 ⫹1 ⫺1 ⫺1 ⫹1 ⫹1 ⫺1 ⫺1 ⫹1

⫺1 ⫹1 ⫺1 ⫹1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫹1 ⫺1 ⫺1 ⫹1 ⫺1 ⫹1

The factors and levels are shown in the following table. ⫺1

Factor A B C D E F G

Amount of binder in the catalyst layer Electrocatalyst loading Amount of carbon in the gas diffusion layer Hot compaction time Compaction temperature Hot compaction load Amount of PTFE in the gas diffusion layer

0.2 mg cm

⫹1 2

1 mg cm2

0.1 mg cm2 2 mg cm2

1 mg cm2 4.5 mg cm2

1 min 100°C 0.04 ton cm2 0.1 mg cm2

10 min 150°C 0.2 ton cm2 1 mg cm2

633

Current Density (CD mA cm2) 160 20 80 317 19 4 20 87.7 1100 12 552 880 16 20 8 15

(a) Write down the alias relationships. (b) Estimate the main effects. (c) Prepare a normal probability plot for the effects and interpret the results. (d) Calculate the sum of squares for the alias set that contains the ABG interaction from the corresponding effect estimate. 14-78. An article in Biotechnology Progress (December 2002, Vol. 18, No. 6, pp. 1170–1175) presented a 27⫺3 fractional factorial to evaluate factors promoting astaxanthin production. The data are shown in the following table.

Runs

A

B

C

D

E

F

G

Weight Content (mg/g)

Cellular Content (pg/cell)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1 1 ⫺1 1 ⫺1 1 ⫺1 1 ⫺1

⫺1 ⫺1 1 1 ⫺1 ⫺1 1 1 1 1 ⫺1 ⫺1 1 1 ⫺1 ⫺1

⫺1 ⫺1 ⫺1 ⫺1 1 1 1 1 1 1 1 1 ⫺1 ⫺1 ⫺1 ⫺1

1 ⫺1 ⫺1 1 1 ⫺1 ⫺1 1 ⫺1 1 1 ⫺1 ⫺1 1 1 ⫺1

1 ⫺1 1 ⫺1 ⫺1 1 ⫺1 1 ⫺1 1 ⫺1 1 1 ⫺1 1 ⫺1

1 1 ⫺1 ⫺1 ⫺1 ⫺1 1 1 ⫺1 ⫺1 1 1 1 1 ⫺1 ⫺1

⫺1 1 1 ⫺1 1 ⫺1 ⫺1 1 1 ⫺1 ⫺1 1 ⫺1 1 1 ⫺1

4.2 4.4 7.8 14.9 25.3 26.7 23.9 21.9 24.3 20.5 10.8 20.8 13.5 10.3 23 12.1

10.8 24.9 27.3 36.3 112.6 159.3 145.2 243.2 72.1 112.2 22.5 149.7 140.1 47.3 153.2 35.2

JWCL232_c14_551-636.qxd

634

1/16/10

9:58 AM

Page 634

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

The factors and levels are shown in the following table.

Factor

1

1

A

Nitrogen concentration (mM)

4.06

0

B

Phosphorus concentration (mM)

0.21

0

2 2

C

Photon flux density (u E m

100

500

D

Magnesium concentration (mM)

s )

1

0

E

Acetate concentration (mM)

0

15

F

Ferrous concentration (mM)

0

0.45

G

NaCl concentration (mM)

Optimal haematococcus medium

25

(a) Write down the complete defining relation and the aliases from the design. (b) Estimate the main effects. (c) Plot the effect estimates on normal probability paper and interpret the results. 14-79 The rework time required for a machine was found to depend on the speed at which the machine was run (A), the lubricant used while working (B), and the hardness of the metal used in the machine (C ). Two levels of each factor are chosen and a single replicate of a 23 experiment is run. The rework time data (in hours) are shown in the following table.

Treatment Combination

Time (in hours)

(1)

27

a

34

b

38

ab

59

c

44

ac

40

bc

63

abc

37

(a) These treatments cannot all be run under the same conditions. Set up a design to run these observations in two blocks of four observations each, with ABC confounded with blocks. (b) Analyze the data. 14-80 Consider the following results from a two-factor experiment with two levels for factor A and three levels for factor B. Each treatment has three replicates.

A

B

Mean

StDev

1

1

21.33333

6.027714

1

2

20

7.549834

1

3

32.66667

3.511885

2

1

31

6.244998

2

2

33

6.557439

2

3

23

10

(a) Calculate the sum of squares for each factor and the interaction. (b) Calculate the sum of squares total and error. (c) Complete an ANOVA table with F-statistics. 14-81. Consider the following ANOVA table from a twofactor factorial experiment. Two-way ANOVA: y versus A, B Source DF SS MS A 3 1213770 404590 B 2 ? 17335441 Error ? 1784195 ? Total 11 37668847

F ? 58.30

P 0.341 0.000

(a) How many levels of each factor were used in the experiment? (b) How many replicates were used? (c) What assumption is made in order to obtain an estimate of error? (d) Calculate the missing entries (denoted with “?”) in the ANOVA table. 14-82. An article in Process Biochemistry (Dec. 1996, Vol. 31, No. 8, pp. 773–785) presented a 273 fractional factorial to perform optimization of manganese dioxide bioleaching media. The data are shown in the following table.

JWCL232_c14_551-636.qxd

1/16/10

9:58 AM

Page 635

635

14-8 RESPONSE SURFACE METHODS AND DESIGNS

Runs

A

B

C

D

E

F

G

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

The factors and levels are shown in the following table.

A B C D E F G

Manganese Extraction Yield (%)

Factor

1

1

Mineral concentration (%) Molasses (g/liter) NH4NO3 (g/liter) KH2PO4 (g/liter) MgSO4 (g/liter) Yeast extract (g/liter) NaHCO3 (g/liter)

10 100 1.25 0.75 0.5 0.20 2.00

20 200 2.50 1.50 1.00 0.50 4.00

(a) Write down the complete defining relation and the aliases from the design. (b) Estimate the main effects.

99.0 97.4 97.7 90.0 100.0 98.0 90.0 93.5 100.0 98.6 97.1 92.4 93.0 95.0 97.0 98.0

(c) Plot the effect estimates on normal probability paper and interpret the results (d) Calculate the missing entries (denoted with “?”) in the following ANOVA table from Minitab with two-way interactions. Source Main Effects 2-Way Interactions Residual Error Total

DF

SS

MS

F

P

7 7

95.934 67.884

13.7049 ?

? 53.69

? 0.105

1 15

? 163.999

?

MIND-EXPANDING EXERCISES 14-83. Consider an unreplicated 2k factorial, and suppose that one of the treatment combinations is missing. One logical approach to this problem is to estimate the missing value with a number that makes the highest order interaction estimate zero. Apply this technique to the data in Example 14-5, assuming that ab is missing. Compare the results of the analysis of these data with the results in Example 14-5.

14-84. What blocking scheme would you recommend if it were necessary to run a 24 design in four blocks of four runs each? 14-85. Consider a 22 design in two blocks with AB confounded with blocks. Prove algebraically that SSAB  SSBlocks.

JWCL232_c14_551-636.qxd

636

1/16/10

9:58 AM

Page 636

CHAPTER 14 DESIGN OF EXPERIMENTS WITH SEVERAL FACTORS

MIND-EXPANDING EXERCISES 14-86. Consider a 23 design. Suppose that the largest number of runs that can be made in one block is four, but we can afford to perform a total of 32 observations. (a) Suggest a blocking scheme that will provide some information on all interactions. (b) Show an outline (source of variability, degrees of freedom only) for the analysis of variance for this design. 14-87. Construct a 251 design. Suppose that it is necessary to run this design in two blocks of eight runs each. Show how this can be done by confounding a twofactor interaction (and its aliased three-factor interaction) with blocks. 14-88. Construct a 272 IV design. Show how this design may be confounded in four blocks of eight runs each. Are any two-factor interactions confounded with blocks? 14-89. Construct a 273 design. Show how this deIV sign can be confounded in two blocks of eight runs each

without losing information on any of the two-factor interactions. 14-90. Set up a 274 III design using D  AB, E  AC, F  BC, and G  ABC as the design generators. Ignore all interaction above the two factors. (a) Verify that each main effect is aliased with three two-factor interactions. (b) Suppose that a second 274 III design with generators D  AB, E  AC, F  BC, and G  ABC is run. What are the aliases of the main effects in this design? (c) What factors may be estimated if the two sets of factor effect estimates above are combined? 14-91. Consider the square root of the sum of squares for curvature and divide by the square root of mean squared error. Explain why the statistic that results has a t distribution and why it can be used to conduct a t test for curvature that is equivalent to the F test in the ANOVA.

IMPORTANT TERMS AND CONCEPTS Analysis of variance (ANOVA) Blocking and nuisance factors Center points Central composite design Confounding

Contrast Defining relation Design matrix Factorial experiment Fractional factorial design Generator Interaction Main effect

Normal probability plot of factor effects Optimization experiment Orthogonal design Regression model Residual analysis

Resolution Response surface Screening experiment Steepest ascent (or descent) 2k factorial design Two-level factorial design

JWCL232_c15_637-701.qxd

1/20/10

8:08 AM

Page 637

15

©doc.stock/©Corbis Courtesy of Mike Johnson, www.redbead.com

Statistical Quality Control BOWL OF BEADS The quality guru Edward Deming conducted a simple experiment in his seminars with a bowl of beads. Many were colored white but a percentage of red beads were randomly mixed in the bowl. A participant from the seminar was provided with a paddle made with indentations so that 50 beads at a time could be scooped from the bowl. The participant was only allowed to use the paddle and instructed to only scoop white beads (repeated multiple times with beads replaced). The red beads were considered to be defectives. Of course, this was difficult to do, and each scoop resulted in a count of red beads. Deming plotted the fraction of red beads from each scoop and used the results to make several points. As was clear from the scenario, this process is beyond the participant’s ability to make simple improvements. It is the process that needs to be changed (reduce the number of red beads) and that is the responsibility of management. Furthermore, many business processes have this type of characteristic and it is important to learn from the data whether the variability is common, intrinsic to the process or whether some special cause has occurred. This distinction is important for the type of process control or improvements to be applied. Refer to the example of control adjustments in Chapter 1. Control charts are primary tools to understand process variability and that is main topic of this chapter.

CHAPTER OUTLINE 15-1

QUALITY IMPROVEMENT AND STATISTICS

15-2.3 Rational Subgroups 15-2.4 Analysis of Patterns on Control Charts

15-1.1 Statistical Quality Control 15-2

15-1.2 Statistical Process Control

15-3

INTRODUCTION TO CONTROL CHARTS

X AND R OR S CONTROL CHARTS

15-4

CONTROL CHARTS FOR INDIVIDUAL MEASUREMENTS

15-5

PROCESS CAPABILITY

15-2.1 Basic Principles 15-2.2 Design of a Control Chart

637

JWCL232_c15_637-701.qxd

638

1/11/10

4:57 PM

Page 638

CHAPTER 15 STATISTICAL QUALITY CONTROL

15-6

ATTRIBUTE CONTROL CHARTS

15-8

15-8.1 Cumulative Sum Control Chart

15-6.1 P Chart (Control Chart for Proportions)

15-7

TIME-WEIGHTED CHARTS

15-8.2 Exponentially Weighted Moving Average Control Chart

15-6.2 U Chart (Control Chart for Defects per Unit)

15-9

CONTROL CHART PERFORMANCE

15-10 IMPLEMENTING SPC

OTHER SPC PROBLEM-SOLVING TOOLS

LEARNING OBJECTIVES After careful study of this chapter you should be able to do the following: 1. Understand the role of statistical tools in quality improvement 2. Understand the different types of variability, rational subgroups, and how a control chart is used to detect assignable causes 3. Understand the general form of a Shewhart control chart and how to apply zone rules (such as the Western Electric rules) and pattern analysis to detect assignable causes 4. Construct and interpret control charts for variables such as X , R, S, and individuals charts 5. 6. 7. 8. 9.

Construct and interpret control charts for attributes such as P and U charts Calculate and interpret process capability ratios Calculate the ARL performance for a Shewhart control chart Construct and interpret a cumulative sum and exponentially weighted moving average control chart Use other statistical process control problem-solving tools

15-1 QUALITY IMPROVEMENT AND STATISTICS The quality of products and services has become a major decision factor in most businesses today. Regardless of whether the consumer is an individual, a corporation, a military defense program, or a retail store, when the consumer is making purchase decisions, he or she is likely to consider quality of equal importance to cost and schedule. Consequently, quality improvement has become a major concern to many U.S. corporations. This chapter is about statistical quality control, a collection of tools that are essential in quality-improvement activities. Quality means fitness for use. For example, you or I may purchase automobiles that we expect to be free of manufacturing defects and that should provide reliable and economical transportation, a retailer buys finished goods with the expectation that they are properly packaged and arranged for easy storage and display, or a manufacturer buys raw material and expects to process it with no rework or scrap. In other words, all consumers expect that the products and services they buy will meet their requirements. Those requirements define fitness for use. Quality or fitness for use is determined through the interaction of quality of design and quality of conformance. By quality of design we mean the different grades or levels of performance, reliability, serviceability, and function that are the result of deliberate engineering and management decisions. By quality of conformance, we mean the systematic reduction of variability and elimination of defects until every unit produced is identical and defect-free. Some confusion exists in our society about quality improvement; some people still think that it means gold-plating a product or spending more money to develop a product or process.

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 639

15-1 QUALITY IMPROVEMENT AND STATISTICS

639

This thinking is wrong. Quality improvement means the systematic elimination of waste. Examples of waste include scrap and rework in manufacturing, inspection and testing, errors on documents (such as engineering drawings, checks, purchase orders, and plans), customer complaint hotlines, warranty costs, and the time required to do things over again that could have been done right the first time. A successful quality-improvement effort can eliminate much of this waste and lead to lower costs, higher productivity, increased customer satisfaction, increased business reputation, higher market share, and ultimately higher profits for the company. Statistical methods play a vital role in quality improvement. Some applications are outlined below: 1.

2.

3. 4.

In product design and development, statistical methods, including designed experiments, can be used to compare different materials, components, or ingredients, and to help determine both system and component tolerances. This application can significantly lower development costs and reduce development time. Statistical methods can be used to determine the capability of a manufacturing process. Statistical process control can be used to systematically improve a process by reducing variability. Experimental design methods can be used to investigate improvements in the process. These improvements can lead to higher yields and lower manufacturing costs. Life testing provides reliability and other performance data about the product. This can lead to new and improved designs and products that have longer useful lives and lower operating and maintenance costs.

Some of these applications have been illustrated in earlier chapters of this book. It is essential that engineers, scientists, and managers have an in-depth understanding of these statistical tools in any industry or business that wants to be a high-quality, low-cost producer. In this chapter we provide an introduction to the basic methods of statistical quality control that, along with experimental design, form the basis of a successful quality-improvement effort.

15-1.1 Statistical Quality Control The field of statistical quality control can be broadly defined as those statistical and engineering methods that are used in measuring, monitoring, controlling, and improving quality. Statistical quality control is a field that dates back to the 1920s. Dr. Walter A. Shewhart of the Bell Telephone Laboratories was one of the early pioneers of the field. In 1924 he wrote a memorandum showing a modern control chart, one of the basic tools of statistical process control. Harold F. Dodge and Harry G. Romig, two other Bell System employees, provided much of the leadership in the development of statistically based sampling and inspection methods. The work of these three men forms much of the basis of the modern field of statistical quality control. World War II saw the widespread introduction of these methods to U.S. industry. Dr. W. Edwards Deming and Dr. Joseph M. Juran have been instrumental in spreading statistical quality-control methods since World War II. The Japanese have been particularly successful in deploying statistical quality-control methods and have used statistical methods to gain significant advantage over their competitors. In the 1970s American industry suffered extensively from Japanese (and other foreign) competition; that has led, in turn, to renewed interest in statistical quality-control methods in the United States. Much of this interest focuses on statistical process control and experimental design. Many U.S. companies have implemented these methods in their manufacturing, engineering, and other business organizations.

JWCL232_c15_637-701.qxd

640

1/11/10

4:57 PM

Page 640

CHAPTER 15 STATISTICAL QUALITY CONTROL

15-1.2 Statistical Process Control It is impractical to inspect quality into a product; the product must be built right the first time. The manufacturing process must therefore be stable or repeatable and capable of operating with little variability around the target or nominal dimension. Online statistical process control is a powerful tool for achieving process stability and improving capability through the reduction of variability. It is customary to think of statistical process control (SPC) as a set of problem-solving tools that may be applied to any process. The major tools of SPC* are 1. 2. 3. 4. 5. 6. 7.

Histogram Pareto chart Cause-and-effect diagram Defect-concentration diagram Control chart Scatter diagram Check sheet

Although these tools are an important part of SPC, they comprise only the technical aspect of the subject. An equally important element of SPC is attitude—a desire of all individuals in the organization for continuous improvement in quality and productivity through the systematic reduction of variability. The control chart is the most powerful of the SPC tools.

15-2 INTRODUCTION TO CONTROL CHARTS 15-2.1 Basic Principles In any production process, regardless of how well-designed or carefully maintained it is, a certain amount of inherent or natural variability will always exist. This natural variability or “background noise” is the cumulative effect of many small, essentially unavoidable causes. When the background noise in a process is relatively small, we usually consider it an acceptable level of process performance. In the framework of statistical quality control, this natural variability is often called a “stable system of chance causes.” A process that is operating with only chance causes of variation present is said to be in statistical control. In other words, the chance causes are an inherent part of the process. Other kinds of variability may occasionally be present in the output of a process. This variability in key quality characteristics usually arises from three sources: improperly adjusted machines, operator errors, or defective raw materials. Such variability is generally large when compared to the background noise, and it usually represents an unacceptable level of process performance. We refer to these sources of variability that are not part of the chance cause pattern as assignable causes. A process that is operating in the presence of assignable causes is said to be out of control.† * Some prefer to include the experimental design methods discussed previously as part of the SPC toolkit. We did not do so, because we think of SPC as an online approach to quality improvement using techniques founded on passive observation of the process, while design of experiments is an active approach in which deliberate changes are made to the process variables. As such, designed experiments are often referred to as offline quality control. † The terminology chance and assignable causes was developed by Dr. Walter A. Shewhart. Today, some writers use common cause instead of chance cause and special cause instead of assignable cause.

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 641

Sample quality characteristic

15-2 INTRODUCTION TO CONTROL CHARTS

Figure 15-1 A typical control chart.

641

Upper control limit

Center line

Lower control limit

Sample number or time

Production processes will often operate in the in-control state, producing acceptable product for relatively long periods of time. Occasionally, however, assignable causes will occur, seemingly at random, resulting in a “shift” to an out-of-control state where a large proportion of the process output does not conform to requirements. A major objective of statistical process control is to quickly detect the occurrence of assignable causes or process shifts so that investigation of the process and corrective action may be undertaken before many nonconforming units are manufactured. The control chart is an online process-monitoring technique widely used for this purpose. Recall the following from Chapter 1. Figure 1-11 illustrates that adjustments to common causes of variation increase the variation of a process whereas Fig. 1-12 illustrates that actions should be taken in response to assignable causes of variation. Control charts may also be used to estimate the parameters of a production process and, through this information, to determine the capability of a process to meet specifications. The control chart can also provide information that is useful in improving the process. Finally, remember that the eventual goal of statistical process control is the elimination of variability in the process. Although it may not be possible to eliminate variability completely, the control chart helps reduce it as much as possible. A typical control chart is shown in Fig. 15-1, which is a graphical display of a quality characteristic that has been measured or computed from a sample versus the sample number or time. Often, the samples are selected at periodic intervals such as every hour. The chart contains a center line (CL) that represents the average value of the quality characteristic corresponding to the in-control state. (That is, only chance causes are present.) Two other horizontal lines, called the upper control limit (UCL) and the lower control limit (LCL), are also shown on the chart. These control limits are chosen so that if the process is in control, nearly all of the sample points will fall between them. In general, as long as the points plot within the control limits, the process is assumed to be in control, and no action is necessary. However, a point that plots outside of the control limits is interpreted as evidence that the process is out of control, and investigation and corrective action are required to find and eliminate the assignable cause or causes responsible for this behavior. The sample points on the control chart are usually connected with straight-line segments so that it is easier to visualize how the sequence of points has evolved over time. Even if all the points plot inside the control limits, if they behave in a systematic or nonrandom manner, this is an indication that the process is out of control. For example, if 18 of the last 20 points plotted above the center line but below the upper control limit, and only two of these points plotted below the center line but above the lower control limit, we would be very suspicious that something was wrong. If the process is in control, all the plotted points should have an essentially random pattern. Methods designed to find

JWCL232_c15_637-701.qxd

642

1/11/10

4:57 PM

Page 642

CHAPTER 15 STATISTICAL QUALITY CONTROL

sequences or nonrandom patterns can be applied to control charts as an aid in detecting out-of-control conditions. A particular nonrandom pattern usually appears on a control chart for a reason, and if that reason can be found and eliminated, process performance can be improved. There is a close connection between control charts and hypothesis testing. Essentially, the control chart is a test of the hypothesis that the process is in a state of statistical control. A point plotting within the control limits is equivalent to failing to reject the hypothesis of statistical control, and a point plotting outside the control limits is equivalent to rejecting the hypothesis of statistical control. We give a general model for a control chart. Let W be a sample statistic that measures some quality characteristic of interest, and suppose that the mean of W is W and the standard deviation of W is W.* Then the center line, the upper control limit, and the lower control limit become Control Chart Model

UCL  W  kW CL  W LCL  W  kW

(15-1)

where k is the “distance” of the control limits from the center line, expressed in standard deviation units. A common choice is k  3. This general theory of control charts was first proposed by Dr. Walter A. Shewhart, and control charts developed according to these principles are often called Shewhart control charts. The control chart is a device for describing exactly what is meant by statistical control; as such, it may be used in a variety of ways. In many applications, it is used for online process monitoring. That is, sample data are collected and used to construct the control chart, and if the sample values of x (say) fall within the control limits and do not exhibit any systematic pattern, we say the process is in control at the level indicated by the chart. Note that we may be interested here in determining both whether the past data came from a process that was in control and whether future samples from this process indicate statistical control. The most important use of a control chart is to improve the process. We have found that, generally 1. 2.

Most processes do not operate in a state of statistical control. Consequently, the routine and attentive use of control charts will identify assignable causes. If these causes can be eliminated from the process, variability will be reduced and the process will be improved.

This process-improvement activity using the control chart is illustrated in Fig. 15-2. Notice that: 3.

The control chart will only detect assignable causes. Management, operator, and engineering action will usually be necessary to eliminate the assignable cause. An action plan for responding to control chart signals is vital.

* Note that “sigma” refers to the standard deviation of the statistic plotted on the chart (i.e., W), not the standard deviation of the quality characteristic.

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 643

15-2 INTRODUCTION TO CONTROL CHARTS Input

643

Output Process

Measurement system

Verify and follow up

Figure 15-2 Process improvement using the control chart.

Implement corrective action

Detect assignable cause

Identify root cause of problem

In identifying and eliminating assignable causes, it is important to find the underlying root cause of the problem and to attack it. A cosmetic solution will not result in any real, long-term process improvement. Developing an effective system for corrective action is an essential component of an effective SPC implementation. We may also use the control chart as an estimating device. That is, from a control chart that exhibits statistical control, we may estimate certain process parameters, such as the mean, standard deviation, and fraction nonconforming or fallout. These estimates may then be used to determine the capability of the process to produce acceptable products. Such process capability studies have considerable impact on many management decision problems that occur over the product cycle, including make-or-buy decisions, plant and process improvements that reduce process variability, and contractual agreements with customers or suppliers regarding product quality. Control charts may be classified into two general types. Many quality characteristics can be measured and expressed as numbers on some continuous scale of measurement. In such cases, it is convenient to describe the quality characteristic with a measure of central tendency and a measure of variability. Control charts for central tendency and variability are collectively called variables control charts. The X chart is the most widely used chart for monitoring central tendency, whereas charts based on either the sample range or the sample standard deviation are used to control process variability. Many quality characteristics are not measured on a continuous scale or even a quantitative scale. In these cases, we may judge each unit of product as either conforming or nonconforming on the basis of whether or not it possesses certain attributes, or we may count the number of nonconformities (defects) appearing on a unit of product. Control charts for such quality characteristics are called attributes control charts. Control charts have had a long history of use in industry. There are at least five reasons for their popularity: 1. Control charts are a proven technique for improving productivity. A successful control chart program will reduce scrap and rework, which are the primary productivity killers in any operation. If you reduce scrap and rework, productivity increases, cost decreases, and production capacity (measured in the number of good parts per hour) increases. 2. Control charts are effective in defect prevention. The control chart helps keep the process in control, which is consistent with the “do it right the first time” philosophy.

JWCL232_c15_637-701.qxd

644

1/11/10

4:57 PM

Page 644

CHAPTER 15 STATISTICAL QUALITY CONTROL

It is never cheaper to sort out the “good” units from the “bad” later on than it is to build them correctly initially. If you do not have effective process control, you are paying someone to make a nonconforming product. 3. Control charts prevent unnecessary process adjustments. A control chart can distinguish between background noise and abnormal variation; no other device, including a human operator, is as effective in making this distinction. If process operators adjust the process based on periodic tests unrelated to a control chart program, they will often overreact to the background noise and make unneeded adjustments. These unnecessary adjustments can result in a deterioration of process performance. In other words, the control chart is consistent with the “if it isn’t broken, don’t fix it” philosophy. 4. Control charts provide diagnostic information. Frequently, the pattern of points on the control chart will contain information that is of diagnostic value to an experienced operator or engineer. This information allows the operator to implement a change in the process that will improve its performance. 5. Control charts provide information about process capability. The control chart provides information about the value of important process parameters and their stability over time. This allows an estimate of process capability to be made. This information is of tremendous use to product and process designers. Control charts are among the most effective management control tools, and they are as important as cost controls and material controls. Modern computer technology has made it easy to implement control charts in any type of process, because data collection and analysis can be performed on a microcomputer or a local area network terminal in real time, online at the work center.

15-2.2 Design of a Control Chart To illustrate these ideas, we give a simplified example of a control chart. In manufacturing automobile engine piston rings, the inside diameter of the rings is a critical quality characteristic. The process mean inside ring diameter is 74 millimeters, and it is known that the standard deviation of ring diameter is 0.01 millimeters. A control chart for average ring diameter is shown in Fig. 15-3. Every hour a random sample of five rings is taken, the average ring diameter of the sample (say x) is computed, and x is plotted on the chart. Because this control chart utilizes the sample mean X to monitor the process mean, it is usually called an X control chart. Note that all the points fall within the control limits, so the chart indicates that the process is in statistical control.

Average ring diameter x

74.0180

Figure 15-3 X control chart for piston ring diameter.

74.0135

UCL = 74.0135

74.0090 74.0045 74.0000 73.9955 73.9910 73.9865

LCL = 73.9865

73.9820 1 2 3 4 5 6 7

8 9 10 11 12 13 14 15 16

Sample number

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 645

15-2 INTRODUCTION TO CONTROL CHARTS

645

Consider how the control limits were determined. The process average is 74 millimeters, and the process standard deviation is   0.01 millimeters. Now if samples of size n  5 are taken, the standard deviation of the sample average X is  X 

 0.01   0.0045 1n 15

Therefore, if the process is in control with a mean diameter of 74 millimeters, by using the central limit theorem to assume that X is approximately normally distributed, we would expect approximately 100(1  )% of the sample mean diameters X to fall between 74  z兾2(0.0045) and 74  z兾2(0.0045). As discussed above, we customarily choose the constant z兾2 to be 3, so the upper and lower control limits become UCL  74  310.00452  74.0135 and LCL  74  310.00452  73.9865 as shown on the control chart. These are the 3-sigma control limits referred to above. Note that the use of 3-sigma limits implies that   0.0027; that is, the probability that the point plots outside the control limits when the process is in control is 0.0027. The width of the control limits is inversely related to the sample size n for a given multiple of sigma. Choosing the control limits is equivalent to setting up the critical region for testing the hypotheses H0:   74 H1:   74 where   0.01 is known. Essentially, the control chart tests this hypothesis repeatedly at different points in time. In designing a control chart, we must specify both the sample size to use and the frequency of sampling. In general, larger samples will make it easier to detect small shifts in the process. When choosing the sample size, we must keep in mind the size of the shift that we are trying to detect. If we are interested in detecting a relatively large process shift, we use smaller sample sizes than those that would be employed if the shift of interest were relatively small. We must also determine the frequency of sampling. The most desirable situation from the point of view of detecting shifts would be to take large samples very frequently; however, this is usually not economically feasible. The general problem is one of allocating sampling effort. That is, either we take small samples at short intervals or larger samples at longer intervals. Current industry practice tends to favor smaller, more frequent samples, particularly in highvolume manufacturing processes or where a great many types of assignable causes can occur. Furthermore, as automatic sensing and measurement technology develops, it is becoming possible to greatly increase frequencies. Ultimately, every unit can be tested as it is manufactured. This capability will not eliminate the need for control charts because the test system will not prevent defects. The increased data will increase the effectiveness of process control and improve quality. When preliminary samples are used to construct limits for control charts, these limits are customarily treated as trial values. Therefore, the sample statistics should be plotted on the appropriate charts, and any points that exceed the control limits should be investigated. If

JWCL232_c15_637-701.qxd

646

1/11/10

4:57 PM

Page 646

CHAPTER 15 STATISTICAL QUALITY CONTROL

assignable causes for these points are discovered, they should be eliminated and new limits for the control charts determined. In this way, the process may be eventually brought into statistical control and its inherent capabilities assessed. Other changes in process centering and dispersion may then be contemplated.

15-2.3 Rational Subgroups A fundamental idea in the use of control charts is to collect sample data according to what Shewhart called the rational subgroup concept. Generally, this means that subgroups or samples should be selected so that to the extent possible, the variability of the observations within a subgroup should include all the chance or natural variability and exclude the assignable variability. Then, the control limits will represent bounds for all the chance variability and not the assignable variability. Consequently, assignable causes will tend to generate points that are outside of the control limits, while chance variability will tend to generate points that are within the control limits. When control charts are applied to production processes, the time order of production is a logical basis for rational subgrouping. Even though time order is preserved, it is still possible to form subgroups erroneously. If some of the observations in the subgroup are taken at the end of one eight-hour shift and the remaining observations are taken at the start of the next eighthour shift, any differences between shifts might not be detected. Time order is frequently a good basis for forming subgroups because it allows us to detect assignable causes that occur over time. Two general approaches to constructing rational subgroups are used. In the first approach, each subgroup consists of units that were produced at the same time (or as closely together as possible). This approach is used when the primary purpose of the control chart is to detect process shifts. It minimizes variability due to assignable causes within a sample, and it maximizes variability between samples if assignable causes are present. It also provides better estimates of the standard deviation of the process in the case of variables control charts. This approach to rational subgrouping essentially gives a “snapshot” of the process at each point in time where a sample is collected. In the second approach, each sample consists of units of product that are representative of all units that have been produced since the last sample was taken. Essentially, each subgroup is a random sample of all process output over the sampling interval. This method of rational subgrouping is often used when the control chart is employed to make decisions about the acceptance of all units of product that have been produced since the last sample. In fact, if the process shifts to an out-of-control state and then back in control again between samples, it is sometimes argued that the first method of rational subgrouping defined above will be ineffective against these types of shifts, and so the second method must be used. When the rational subgroup is a random sample of all units produced over the sampling interval, considerable care must be taken in interpreting the control charts. If the process mean drifts between several levels during the interval between samples, the range of observations within the sample may consequently be relatively large. It is the within-sample variability that determines the width of the control limits on an X chart, so this practice will result in wider limits on the X chart. This makes it harder to detect shifts in the mean. In fact, we can often make any process appear to be in statistical control just by stretching out the interval between observations in the sample. It is also possible for shifts in the process average to cause points on a control chart for the range or standard deviation to plot out of control, even though no shift in process variability has taken place. There are other bases for forming rational subgroups. For example, suppose a process consists of several machines that pool their output into a common stream. If we sample from this

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 647

15-2 INTRODUCTION TO CONTROL CHARTS

647

UCL

x Center line

LCL

Figure 15-4 An X control chart.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 Sample number

common stream of output, it will be very difficult to detect whether or not some of the machines are out of control. A logical approach to rational subgrouping here is to apply control chart techniques to the output for each individual machine. Sometimes this concept needs to be applied to different heads on the same machine, different workstations, different operators, and so forth. The rational subgroup concept is very important. The proper selection of samples requires careful consideration of the process, with the objective of obtaining as much useful information as possible from the control chart analysis.

15-2.4 Analysis of Patterns on Control Charts A control chart may indicate an out-of-control condition either when one or more points fall beyond the control limits, or when the plotted points exhibit some nonrandom pattern of behavior. For example, consider the X chart shown in Fig. 15-4. Although all 25 points fall within the control limits, the points do not indicate statistical control because their pattern is very nonrandom in appearance. Specifically, we note that 19 of the 25 points plot below the center line, while only 6 of them plot above. If the points are truly random, we should expect a more even distribution of them above and below the center line. We also observe that following the fourth point, five points in a row increase in magnitude. This arrangement of points is called a run. Since the observations are increasing, we could call it a run up; similarly, a sequence of decreasing points is called a run down. This control chart has an unusually long run up (beginning with the fourth point) and an unusually long run down (beginning with the eighteenth point). In general, we define a run as a sequence of observations of the same type. In addition to runs up and runs down, we could define the types of observations as those above and below the center line, respectively, so two points in a row above the center line would be a run of length 2. A run of length 8 or more points has a very low probability of occurrence in a random sample of points. Consequently, any type of run of length 8 or more is often taken as a signal of an out-of-control condition. For example, eight consecutive points on one side of the center line will indicate that the process is out of control. Although runs are an important measure of nonrandom behavior on a control chart, other types of patterns may also indicate an out-of-control condition. For example, consider the X

JWCL232_c15_637-701.qxd

648

1/11/10

4:57 PM

Page 648

CHAPTER 15 STATISTICAL QUALITY CONTROL

UCL x Center line LCL

Figure 15-5 An X chart with a cyclic pattern.

1

2

3

4

5

6

7 8 9 10 11 12 13 14 15 Sample number

chart in Fig. 15-5. Note that the plotted sample averages exhibit a cyclic behavior, yet they all fall within the control limits. Such a pattern may indicate a problem with the process, such as operator fatigue, raw material deliveries, and heat or stress buildup. The yield may be improved by eliminating or reducing the sources of variability causing this cyclic behavior (see Fig. 15-6). In Fig. 15-6, LSL and USL denote the lower and upper specification limits of the process. These limits represent bounds within which acceptable product must fall and they are often based on customer requirements. The problem is one of pattern recognition, that is, recognizing systematic or nonrandom patterns on the control chart and identifying the reason for this behavior. The ability to interpret a particular pattern in terms of assignable causes requires experience and knowledge of the process. That is, we must not only know the statistical principles of control charts, but we must also have a good understanding of the process. The Western Electric Handbook (1956) suggests a set of decision rules for detecting nonrandom patterns on control charts. Specifically, the Western Electric rules would conclude that the process is out of control if either 1. One point plots outside 3-sigma control limits. 2. Two out of three consecutive points plot beyond a 2-sigma limit. 3. Four out of five consecutive points plot at a distance of 1-sigma or beyond from the center line. 4. Eight consecutive points plot on one side of the center line.

LSL

μ

USL

(a)

Figure 15-6 (a) Variability with the cyclic pattern. (b) Variability with the cyclic pattern eliminated.

LSL

μ (b)

USL

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 649

15-3 X AND R OR S CONTROL CHARTS

649

74.0180 74.0135

UCL = 74.0135 Zone A

74.0090

Zone B

74.0045

Zone C

3σ X 2σ X 1σ X

74.0000 Zone C

73.9955

Zone B 73.9910 Zone A 73.9865

Figure 15-7 The Western Electric zone rules.

LCL = 73.9865

1σ X 2σ X 3σ X

73.9820 1

2

3

4

5

6

7

8

9

10 11 12

Sample number

We have found these rules very effective in practice for enhancing the sensitivity of control charts. Rules 2 and 3 apply to one side of the center line at a time. That is, a point above the upper 2-sigma limit followed immediately by a point below the lower 2-sigma limit would not signal an out-of-control alarm. Figure 15-7 shows an X control chart for the piston ring process with the 1-sigma, 2-sigma, and 3-sigma limits used in the Western Electric procedure. Notice that these inner limits (sometimes called warning limits) partition the control chart into three zones A, B, and C on each side of the center line. Consequently, the Western Electric rules are sometimes called the run rules for control charts. Notice that the last four points fall in zone B or beyond. Thus, since four of five consecutive points exceed the 1-sigma limit, the Western Electric procedure will conclude that the pattern is nonrandom and the process is out of control.

15-3

X AND R OR S CONTROL CHARTS When dealing with a quality characteristic that can be expressed as a measurement, it is customary to monitor both the mean value of the quality characteristic and its variability. Control over the average quality is exercised by the control chart for averages, usually called the X chart. Process variability can be controlled by either a range chart (R chart) or a standard deviation chart (S chart), depending on how the population standard deviation is estimated. Suppose that the process mean and standard deviation  and  are known and that we can assume that the quality characteristic has a normal distribution. Consider the X chart. As discussed previously, we can use  as the center line for the control chart, and we can place the upper and lower 3-sigma limits at

UCL    3 1n LCL    3 1n CL  

(15-2)

When the parameters  and  are unknown, we usually estimate them on the basis of preliminary samples, taken when the process is thought to be in control. We recommend the

JWCL232_c15_637-701.qxd

650

1/11/10

4:57 PM

Page 650

CHAPTER 15 STATISTICAL QUALITY CONTROL

use of at least 20 to 25 preliminary samples. Suppose m preliminary samples are available, each of size n. Typically, n will be 4, 5, or 6; these relatively small sample sizes are widely used and often arise from the construction of rational subgroups. Let the sample mean for the ith sample be Xi. Then we estimate the mean of the population, , by the grand mean 1 m ˆ X  a Xi m i1

(15-3)

Thus, we may take X as the center line on the X control chart. We may estimate  from either the standard deviation or the range of the observations within each sample. The sample size is relatively small, so there is little loss in efficiency in estimating  from the sample ranges. The relationship between the range R of a sample from a normal population with known parameters and the standard deviation of that population is needed. Since R is a random variable, the quantity W  R兾, called the relative range, is also a random variable. The parameters of the distribution of W have been determined for any sample size n. The mean and standard deviation of the distribution of W are called d2 and d3 respectively. Because R  W, R  d2

R  d3

(15-4)

Let Ri be the range of the ith sample, and let 1 m R  m a Ri i1

(15-5)

be the average range. Then R is an estimator of R and from Equation 15-4 an unbiased estimator of  is

Estimator of  from R Chart

ˆ 

R d2

(15-6)

where the constant d2 is tabulated for various sample sizes in Appendix Table XI.

Therefore, we may use as our upper and lower control limits for the X chart UCL  X 

3 R d2 1n

LCL  X 

3 R d2 1n

(15-7)

Define the constant A2 

3 d2 1n

(15-8)

Now, once we have computed the sample values x and r, the X control chart may be defined as follows.

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 651

651

15-3 X AND R OR S CONTROL CHARTS

X Control Chart (from R)

The center line and upper and lower control limits for an X control chart are UCL  x  A2 r

CL  x

LCL  x  A2 r

(15-9)

where the constant A2 is tabulated for various sample sizes in Appendix Table XI.

The parameters of the R chart may also be easily determined. The center line is R. To determine the control limits, we need an estimate of R, the standard deviation of R. Once again, assuming the process is in control, the distribution of the relative range, W, is useful. We may estimate R from Equation 15-4 as ˆ R  d3ˆ  d3

R d2

(15-10)

and the upper and lower control limits on the R chart are UCL  R 

3d3 3d3 R  a1  bR d2 d2

LCL  R 

3d3 3d3 R  a1  bR d2 d2

(15-11)

Setting D3  1  3d3兾d2 and D4  1  3d3兾d2 leads to the following definition. R Chart

The center line and upper and lower control limits for an R chart are UCL  D4 r

CL  r

LCL  D3 r

(15-12)

where r is the sample average range, and the constants D3 and D4 are tabulated for various sample sizes in Appendix Table XI.

The LCL for an R chart can be a negative number. In that case, it is customary to set LCL to zero. Because the points plotted on an R chart are nonnegative, no points can fall below an LCL of zero. Also, we often study the R chart first because if the process variability is not constant over time the control limits calculated for the X chart can be misleading. Rather than base control charts on ranges, a more modern approach is to calculate the standard deviation of each subgroup and plot these standard deviations to monitor the process standard deviation . This is called an S chart. When an S chart is used, it is common to use these standard deviations to develop control limits for the X chart. Typically, the sample size used for subgroups is small (fewer than 10) and in that case there is usually little difference in the X chart generated from ranges or standard deviations. However, because computer software is often used to implement control charts, S charts are quite common. Details to construct these charts follow. In Section 7-3, it was stated that S is a biased estimator of . That is, E(S)  c4 where c4 is a constant that is near, but not equal to, 1. Furthermore, a calculation similar to the one used for E(S) can derive the standard deviation of the statistic S with the result 21  c24. Therefore, the center line and three-sigma control limits for S are LCL  c4  3 21  c24

CL  c4

UCL  c4  3 21  c24

(15-13)

JWCL232_c15_637-701.qxd

652

1/11/10

4:57 PM

Page 652

CHAPTER 15 STATISTICAL QUALITY CONTROL

Assume that there are m preliminary samples available, each of size n, and let Si denote the standard deviation of the ith sample. Define 1 m S  m a Si i1

(15-14)

Because E1S2  c4, an unbiased estimator of  is S c4 That is, Estimator of  from S Chart

ˆ  S c4

(15-15)

where the constant c4 is tabulated for various sample sizes in Appendix Table XI.

A control chart for standard deviations follows. S Chart

s UCL  s  3 c 21  c24 4

s LCL  s  3 c 21  c24 4

CL  s

(15-16)

The LCL for an S chart can be a negative number, in that case, it is customary to set LCL to zero. When an S chart is used, the estimate for  in Equation 15-15 is commonly used to calculate the control limits for an X chart. This produces the following control limits for an X chart. X Control Chart (from S)

EXAMPLE 15-1

UCL  x  3

s c4 1n

Vane Opening

A component part for a jet aircraft engine is manufactured by an investment casting process. The vane opening on this casting is an important functional parameter of the part. We will illustrate the use of X and R control charts to assess the statistical stability of this process. Table 15-1 presents 20 samples of five parts each. The values given in the table have been coded by using the last three digits of the dimension; that is, 31.6 should be 0.50316 inch. The quantities x  33.3 and r  5.8 are shown at the foot of Table 15-1. The value of A2 for samples of size 5 is A2  0.577. Then the trial control limits for the X chart are x A2 r  33.32 10.577215.82  33.32 3.35

CL  x

LCL  s  3

s c4 1n

UCL  D4 r  12.115215.82  12.27 LCL  D3 r  10215.82  0

The X and R control charts with these trial control limits are shown in Fig. 15-8. Notice that samples 6, 8, 11, and 19 are out of control on the X chart and that sample 9 is out of control on the R chart. (These points are labeled with a “1” because they violate the first Western Electric rule.) For the S chart, the value of c4  0.94. Therefore, 312.3452 3s 2 2 c4 21  c4  0.94 21  0.94  2.553

or and the trial control limits are UCL  36.67

LCL  29.97

For the R chart, the trial control limits are

(15-17)

UCL  2.345  2.553  4.898 LCL  2.345  2.553  0.208

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 653

16-5 X AND R OR S CONTROL CHARTS

Table 15-1 Vane-Opening Measurements Sample Number

x1

x2

x3

x4

x5

x

r

s

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

33 33 35 30 33 38 30 29 28 38 28 31 27 33 35 33 35 32 25 35

29 31 37 31 34 37 31 39 33 33 30 35 32 33 37 33 34 33 27 35

31 35 33 33 35 39 32 38 35 32 28 35 34 35 32 27 34 30 34 36

32 37 34 34 33 40 34 39 36 35 32 35 35 37 35 31 30 30 27 33

33 31 36 33 34 38 31 39 43 32 31 34 37 36 39 30 32 33 28 30

31.6 33.4 35.0 32.2 33.8 38.4 31.6 36.8 35.0 34.0 29.8 34.0 33.0 34.8 35.6 30.8 33.0 31.6 28.2 33.8 x  33.32

4 6 4 4 2 3 4 10 15 6 4 4 10 4 7 6 5 3 9 6 r  5.8

1.67332 2.60768 1.58114 1.64317 0.83666 1.14018 1.51658 4.38178 5.43139 2.54951 1.78885 1.73205 3.80789 1.78885 2.60768 2.48998 2.00000 1.51658 3.42053 2.38747 s  2.345

42 40

Means

38 36 34

1 1

UCL = 36.67 x= 33.32

32 30

LCL = 29.97 1

28 26

1 1

2

3

4

5

6

7

8 9 10 11 12 13 14 15 16 17 18 19 20 Subgroup

16 14

Ranges

12

1

UCL = 12.27

10 8 6

a = 5.80

4 2

Figure 15-8 The X and R control charts for vane opening.

0

LCL = 0.00 1

2

3

4

5

6

7

8 9 10 11 12 13 14 15 16 17 18 19 20 Subgroup

653

JWCL232_c15_637-701.qxd

654

1/11/10

4:57 PM

Page 654

CHAPTER 15 STATISTICAL QUALITY CONTROL

Figure 15-9 The S control chart for vane opening.

Sample standard deviation

6

1

UCL = 4.899

5 4

S = 2.345

3 2 1 0

LCL = 0.00 0

5

10

The LCL is set to zero. If s is used to determine the control limits for the X chart, x

15

20

Subgroup

and for the R chart, UCL  D4 r  12.115215.02  10.57

312.3452 3s  33.32

 33.32 3.35 c4 1n 0.94 15

LCL  D3 r  10215.02  0

and this result is nearly the same as from r . The S chart is shown in Fig. 15-9. Because the control limits for the X chart calculated from s are nearly the same as from r, the chart is not shown. Suppose that all of these assignable causes can be traced to a defective tool in the wax-molding area. We should discard these five samples and recompute the limits for the X and R charts. These new revised limits are, for the X chart, UCL  x  A2 r  33.21  10.577215.02  36.10 LCL  x  A2 r  33.21  10.577215.02  30.33

The revised control charts are shown in Fig. 15-10. Practical Interpretation: Notice that we have treated the first 20 preliminary samples as estimation data with which to establish control limits. These limits can now be used to judge the statistical control of future production. As each new sample becomes available, the values of x and r should be computed and plotted on the control charts. It may be desirable to revise the limits periodically, even if the process remains stable. The limits should always be revised when process improvements are made.

42 = not used in computing control limits

40 38 x

Revised UCL = 36.10

36

CL = 33.21

34 32

Revised LCL = 30.33

30 28

1

2

3

4

5

6

7

8 9 10 11 12 13 14 15 16 17 18 19 20

Sample number

Estimation of limits X Chart 16 = not used in computing control limits

14 12 r

Revised UCL = 10.57

10 8 6

CL = 5.0

4

Figure 15-10 The X and R control charts for vane opening, revised limits.

2

1

2

3

4

5

6

7

8 9 10 11 12 13 14 15 16 17 18 19 20 Estimation of limits R Chart

Sample number

JWCL232_c15_637-701.qxd

1/19/10

9:10 PM

Page 655

16-5 X AND R OR S CONTROL CHARTS

Table 15-2

655

Summary Report from Minitab for the Vane-Opening Data

Test Results for Xbar Chart TEST 1. One point more than 3.00 sigmas from center line. Test Failed at points: 6 8 11 19 Test Results for R Chart TEST 1. One point more than 3.00 sigmas from center line. Test Failed at points: 9

Computer Construction of X and R Control Charts Many computer programs construct X and R control charts. Figures 15-8 and 15-10 show charts similar to those produced by Minitab for the vane-opening data. This program will allow the user to select any multiple of sigma as the width of the control limits and use the Western Electric rules to detect out-of-control points. The program will also prepare a summary report as in Table 15-2 and exclude subgroups from the calculation of the control limits.

EXERCISES FOR SECTION 15-3 15-1. Control charts for X and R are to be set up for an important quality characteristic. The sample size is n ⫽ 5, and x and r are computed for each of 35 preliminary samples. The summary data are 35

a xi ⫽ 7805

i⫽1

35

a ri ⫽ 1200

i⫽1

(a) Calculate trial control limits for X and R charts. (b) Assuming that the process is in control, estimate the process mean and standard deviation. 15-2. Twenty-five samples of size 5 are drawn from a process at one-hour intervals, and the following data are obtained: 25

a xi ⫽ 362.75

i⫽1

25

a ri ⫽ 8.60

i⫽1

25

a si ⫽ 3.64

i⫽1

(a) Calculate trial control limits for X and R charts. (b) Repeat part (a) for X and S charts. 15-3. Control charts are to be constructed for samples of size n ⫽ 4, and x and s are computed for each of 20 preliminary samples as follows: 20

a xi ⫽ 4460

i⫽1

20

a si ⫽ 271.6

i⫽1

(a) Calculate trial control limits for X and S charts. (b) Assuming the process is in control, estimate the process mean and standard deviation. 15-4. Samples of size n ⫽ 6 are collected from a process every hour. After 20 samples have been collected, we calculate x ⫽ 20.0 and rⲐd2 ⫽ 1.4. (a) Calculate trial control limits for X and R charts.

(b) If s兾c4 ⫽ 1.5, determine trial control limits for X and S charts. 15-5. The level of cholesterol (in mg/dL) is an important index for human health. The sample size is n ⫽ 5. The following summary statistics are obtained from cholesterol measurements: 30

a xi ⫽ 140.03,

i⫽1

30

a ri ⫽ 13.63,

i⫽1

30

a si ⫽ 5.10

i⫽1

(a) Find trial control limits for X and R charts. (b) Repeat part (a) for X and S charts. 15-6. An X control chart with three-sigma control limits has UCL ⫽ 48.75 and LCL ⫽ 42.71. Suppose the process standard deviation is ␴ ⫽ 2.25. What subgroup size was used for the chart? 15-7. An extrusion die is used to produce aluminum rods. The diameter of the rods is a critical quality characteristic. The following table shows x and r values for 20 samples of five rods each. Specifications on the rods are 0.5035 ⫾ 0.0010 inch. The values given are the last three digits of the measurement; that is, 34.2 is read as 0.50342. Sample

x

r

1 2 3 4 5 6 7 8 9

34.2 31.6 31.8 33.4 35.0 32.1 32.6 33.8 34.8

3 4 4 5 4 2 7 9 10 continued

JWCL232_c15_637-701.qxd

656

1/11/10

4:57 PM

Page 656

CHAPTER 15 STATISTICAL QUALITY CONTROL

Sample

x

r

Sample Number

x1

x2

x3

10 11 12 13 14 15 16 17 18 19 20

38.6 35.4 34.0 36.0 37.2 35.2 33.4 35.0 34.4 33.9 34.0

4 8 6 4 7 3 10 4 7 8 4

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

15.4 15.4 16.1 13.5 18.3 19.2 14.1 15.6 13.9 18.7 15.3 16.6 17.0 16.3 8.4 11.1 16.5 18.0 17.8 11.5

15.6 17.1 16.1 12.5 16.1 17.2 12.4 13.3 14.9 21.2 13.1 18.0 15.2 16.5 7.7 13.8 17.1 14.1 17.3 10.8

15.3 15.2 13.5 10.2 17.0 19.4 11.7 13.6 15.5 20.1 13.7 18.0 18.1 17.7 8.4 11.9 18.5 15.9 12.0 11.2

(a) Using all the data, find trial control limits for X and R charts, construct the chart, and plot the data. (b) Use the trial control limits from part (a) to identify out-of-control points. If necessary, revise your control limits, assuming that any samples that plot outside the control limits can be eliminated. Estimate . 15-8. The copper content of a plating bath is measured three times per day, and the results are reported in ppm. The x and r values for 25 days are shown in the following table:

Day

x

r

Day

x

r

1 2 3 4 5 6 7 8 9 10 11 12 13

5.45 5.39 6.85 6.74 5.83 7.22 6.39 6.50 7.15 5.92 6.45 5.38 6.03

1.21 0.95 1.43 1.29 1.35 0.88 0.92 1.13 1.25 1.05 0.98 1.36 0.83

14 15 16 17 18 19 20 21 22 23 24 25

7.01 5.83 6.35 6.05 7.11 7.32 5.90 5.50 6.32 6.55 5.90 5.95

1.45 1.37 1.04 0.83 1.35 1.09 1.22 0.98 1.21 0.76 1.20 1.19

(a) Using all the data, find trial control limits for X and R charts, construct the chart, and plot the data. Is the process in statistical control? (b) If necessary, revise the control limits computed in part (a), assuming that any samples that plot outside the control limits can be eliminated. 15-9. The pull strength of a wire-bonded lead for an integrated circuit is monitored. The following table provides data for 20 samples each of size three.

(a) Use all the data to determine trial control limits for X and R charts, construct the control limits, and plot the data. (b) Use the control limits from part (a) to identify out-of-control points. If necessary, revise your control limits assuming that any samples that plot outside of the control limits can be eliminated. (c) Repeat parts (a) and (b) for X and S charts. 15-10. The following data were considered in Quality Engineering [“An SPC Case Study on Stabilizing Syringe Lengths” (1999–2000, Vol. 12(1))]. The syringe length is measured during a pharmaceutical manufacturing process. The following table provides data (in inches) for 20 samples each of size five. Sample

x1

x2

x3

x4

x5

1 2 3 4 5 6 7 8 9 10 11 12

4.960 4.958 4.971 4.940 4.964 4.969 4.960 4.969 4.984 4.970 4.975 4.945

4.946 4.927 4.929 4.982 4.950 4.951 4.944 4.949 4.928 4.934 4.959 4.977

4.950 4.935 4.965 4.970 4.953 4.955 4.957 4.963 4.960 4.961 4.962 4.950

4.956 4.940 4.952 4.953 4.962 4.966 4.948 4.952 4.943 4.940 4.971 4.969

4.958 4.950 4.938 4.960 4.956 4.954 4.951 4.962 4.955 4.965 4.968 4.954

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 657

657

16-5 X AND R OR S CONTROL CHARTS

13 14 15 16 17 18 19 20

4.976 4.970 4.982 4.961 4.980 4.975 4.977 4.975

4.964 4.954 4.962 4.943 4.970 4.968 4.966 4.967

4.970 4.964 4.968 4.950 4.975 4.971 4.969 4.969

4.968 4.959 4.975 4.949 4.978 4.969 4.973 4.972

4.972 4.968 4.963 4.957 4.977 4.972 4.970 4.972

(a) Using all the data, find trial control limits for X and R charts, construct the chart, and plot the data. Is this process in statistical control? (b) Use the trial control limits from part (a) to identify outof-control points. If necessary, revise your control limits assuming that any samples that plot outside of the control limits can be eliminated. (c) Repeat parts (a) and (b) for X and S charts. 15-11. The thickness of a metal part is an important quality parameter. Data on thickness (in inches) are given in the following table, for 25 samples of five parts each.

(a) Using all the data, find trial control limits for X and R charts, construct the chart, and plot the data. Is the process in statistical control? (b) Use the trial control limits from part (a) to identify out-ofcontrol points. If necessary, revise your control limits assuming that any samples that plot outside of the control limits can be eliminated. (c) Repeat parts (a) and (b) for X and S charts. 15-12. Apply the Western Electric Rules to the following x control chart. The warning limits are shown as dotted lines. Describe any rule violations.

UCL=16 14 12 _ X=10 8 6 LCL=4 2

Sample Number

x1

x2

x3

x4

x5

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

0.0629 0.0630 0.0628 0.0634 0.0619 0.0613 0.0630 0.0628 0.0623 0.0631 0.0635 0.0623 0.0635 0.0645 0.0619 0.0631 0.0616 0.0630 0.0636 0.0640 0.0628 0.0615 0.0630 0.0635 0.0623

0.0636 0.0631 0.0631 0.0630 0.0628 0.0629 0.0639 0.0627 0.0626 0.0631 0.0630 0.0630 0.0631 0.0640 0.0644 0.0627 0.0623 0.0630 0.0631 0.0635 0.0625 0.0625 0.0632 0.0629 0.0629

0.0640 0.0622 0.0633 0.0631 0.0630 0.0634 0.0625 0.0622 0.0633 0.0633 0.0638 0.0630 0.0630 0.0631 0.0632 0.0630 0.0631 0.0626 0.0629 0.0629 0.0616 0.0619 0.0630 0.0635 0.0630

0.0635 0.0625 0.0633 0.0632 0.0619 0.0625 0.0629 0.0625 0.0630 0.0631 0.0635 0.0627 0.0630 0.0640 0.0622 0.0628 0.0620 0.0629 0.0635 0.0635 0.0620 0.0619 0.0631 0.0631 0.0626

0.0640 0.0627 0.0630 0.0633 0.0625 0.0628 0.0627 0.0627 0.0624 0.0630 0.0633 0.0629 0.0630 0.0642 0.0635 0.0629 0.0625 0.0628 0.0634 0.0634 0.0623 0.0622 0.0630 0.0633 0.0628

4

6

8

10

12

14

16

18

20

Observation

15-13. An X control chart with three-sigma control limits and subgroup size n  4 has control limits UCL  48.75 and LCL  40.55. (a) Estimate the process standard deviation. (b) Does the answer to part (a) depend on whether r or s was used to construct the X control chart? 15-14. Web traffic can be measured to help highlight security problems or indicate a potential lack of bandwidth. Data on Web traffic (in thousand hits) from http://en.wikipedia. org/wiki/Web_traffic are given in the following table for 25 samples each of size four. Sample

x1

x2

x3

x4

1 2 3 4 5 6 7 8 9 10 11

163.95 163.30 163.13 164.08 165.44 163.83 162.94 164.97 165.04 164.74 164.72

164.54 162.85 165.14 163.43 163.63 164.14 163.64 163.68 164.06 163.74 165.75

163.87 163.18 162.80 164.03 163.95 165.22 162.30 164.73 164.40 165.10 163.07

165.10 165.10 163.81 163.77 164.78 164.91 163.78 162.32 163.69 164.32 163.84 continued

JWCL232_c15_637-701.qxd

658

1/11/10

4:57 PM

Page 658

CHAPTER 15 STATISTICAL QUALITY CONTROL

Sample

x1

x2

x3

x4

12 13 14 15 16 17 18 19 20 21 22 23 24 25

164.25 164.71 166.61 165.23 164.27 163.59 164.90 163.98 164.08 165.71 164.03 160.52 164.22 163.93

162.72 162.63 167.07 163.40 163.42 164.84 164.20 163.53 164.33 162.63 163.36 161.68 164.27 163.96

163.25 165.07 167.41 164.94 164.73 164.45 164.32 163.34 162.38 164.42 164.55 161.18 164.35 165.05

164.14 162.59 166.10 163.74 164.88 164.12 163.98 163.82 164.08 165.27 165.77 161.33 165.12 164.52

(a) Use all the data to determine trial control limits for X and R charts, construct the chart, and plot the data. (b) Use the trial control limits from part (a) to identify outof-control points. If necessary, revise your control limits, assuming that any samples that plot outside the control limits can be eliminated. 15-15. Consider the data in Exercise 15-9. Calculate the sample standard deviation of all 60 measurements and compare this result to the estimate of  obtained from your revised X and R charts. Explain any differences. 15-16. Consider the data in Exercise 15-10. Calculate the sample standard deviation of all 100 measurements and compare this result to the estimate of  obtained from your revised X and R charts. Explain any differences.

15-4 CONTROL CHARTS FOR INDIVIDUAL MEASUREMENTS In many situations, the sample size used for process control is n  1; that is, the sample consists of an individual unit. Some examples of these situations are as follows: 1. 2. 3. 4.

Automated inspection and measurement technology is used, and every unit manufactured is analyzed. The production rate is very slow, and it is inconvenient to allow sample sizes of n  1 to accumulate before being analyzed. Repeat measurements on the process differ only because of laboratory or analysis error, as in many chemical processes. In process plants, such as papermaking, measurements on some parameters such as coating thickness across the roll will differ very little and produce a standard deviation that is much too small if the objective is to control coating thickness along the roll.

In such situations, the individuals control chart (also called an X chart) is useful. The control chart for individuals uses the moving range of two successive observations to estimate the process variability. The moving range is defined as MRi  0Xi  Xi10 and for m observations the average moving range is m MR 

m 1 0 X  Xi10 a m  1 i2 i

An estimate of  is ˆ 

MR MR  1.128 d2

(15-18)

because each moving range is the range between two consecutive observations. Note that there are only m  1 moving ranges. It is also possible to establish a control chart on the moving range using D3 and D4 for n  2. The parameters for these charts are defined as follows.

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 659

15-4 CONTROL CHARTS FOR INDIVIDUAL MEASUREMENTS

Individuals Control Chart

659

The center line and upper and lower control limits for a control chart for individuals are UCL  x  3

mr mr x3 1.128 d2

CL  x

(15-19)

mr mr LCL  x  3 x3 1.128 d2 and for a control chart for moving ranges UCL  D4mr  3.267mr CL  mr LCL  D3mr  0

Note that LCL for this moving range chart is always zero because D3  0 for n = 2. The procedure is illustrated in the following example. EXAMPLE 15-2

Chemical Process Concentration

Table 15-3 shows 20 observations on concentration for the output of a chemical process. The observations are taken at one-hour intervals. If several observations are taken at the

same time, the observed concentration reading will differ only because of measurement error. Since the measurement error is small, only one observation is taken each hour.

Table 15-3 Chemical Process Concentration Measurements Observation 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Concentration x 102.0 94.8 98.3 98.4 102.0 98.5 99.0 97.7 100.0 98.1 101.3 98.7 101.1 98.4 97.0 96.7 100.3 101.4 97.2 101.0 x  99.1

Moving Range mr 7.2 3.5 0.1 3.6 3.5 0.5 1.3 2.3 1.9 3.2 2.6 2.4 2.7 1.4 0.3 3.6 1.1 4.2 3.8 mr  2.59

JWCL232_c15_637-701.qxd

660

1/11/10

4:57 PM

Page 660

CHAPTER 15 STATISTICAL QUALITY CONTROL

107

10 UCL = 105.99

UCL = 8.46 8

101

Moving range (2)

Individuals

104 x= 99.10

98 95

6 4

r= 2.59

2 LCL = 92.21

92 0

4

LCL = 0.000

0 8

12 Subgroup

16

20

0

4

8

12

16

20

Subgroup

Figure 15-11 Control charts for individuals and the moving range (from Minitab) for the chemical process concentration data.

To set up the control chart for individuals, note that the sample average of the 20 concentration readings is x  99.1 and that the average of the moving ranges of two observations shown in the last column of Table 16-3 is mr  2.59. To set up the moving-range chart, we note that D3  0 and D4  3.267 for n  2. Therefore, the moving-range chart has center line mr  2.59, LCL  0, and UCL  D4mr  13.267212.592  8.46. The control chart is shown as the lower control chart in Fig. 15-11. This control chart was constructed by Minitab. Because no points exceed the upper control limit, we may now set up the control chart for individual concentration measurements. If a moving range of n  2 observations is used, d2  1.128. For the data in Table 15-3 we have

UCL  x  3

mr 2.59  105.99  99.1  3 1.128 d2

CL  x  99.1 LCL  x  3

mr 2.59  92.21  99.1  3 1.128 d2

The control chart for individual concentration measurements is shown as the upper control chart in Fig. 15-11. There is no indication of an out-of-control condition. Practical Interpretation: These calculated control limits are used to monitor future production.

The chart for individuals can be interpreted much like an ordinary X control chart. A shift in the process average will result in either a point (or points) outside the control limits, or a pattern consisting of a run on one side of the center line. Some care should be exercised in interpreting patterns on the moving-range chart. The moving ranges are correlated, and this correlation may often induce a pattern of runs or cycles on the chart. The individual measurements are assumed to be uncorrelated, however, and any apparent pattern on the individuals’ control chart should be carefully investigated. The control chart for individuals is not very sensitive to small shifts in the process mean. For example, if the size of the shift in the mean is one standard deviation, the average number of points to detect this shift is 43.9. This result is shown later in the chapter. While the performance of the control chart for individuals is much better for large shifts, in many situations the shift of interest is not large and more rapid shift detection is desirable. In these cases, we recommend time-weighted charts such as the cumulative sum control chart or an exponentially weighted moving-average chart (discussed in Section 15-8). Some individuals have suggested that limits narrower than 3-sigma be used on the chart for individuals to enhance its ability to detect small process shifts. This is a dangerous suggestion, for narrower limits dramatically increase false alarms and the charts may be ignored and become useless. If you are interested in detecting small shifts, consider the time-weighted charts.

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 661

661

15-4 CONTROL CHARTS FOR INDIVIDUAL MEASUREMENTS

EXERCISES FOR SECTION 15-4 15-17. Twenty successive hardness measurements are made on a metal alloy, and the data are shown in the following table. Observation

Hardness

Observation

Hardness

1 2 3 4 5 6 7 8 9 10

51 52 54 55 55 51 52 50 51 56

11 12 13 14 15 16 17 18 19 20

51 57 58 50 53 52 54 50 56 53

(a) Using all the data, compute trial control limits for individual observations and moving-range charts. Construct the chart and plot the data. Determine whether the process is in statistical control. If not, assume assignable causes can be found to eliminate these samples and revise the control limits. (b) Estimate the process mean and standard deviation for the in-control process. 15-18. In a semiconductor manufacturing process CVD metal thickness was measured on 30 wafers obtained over approximately two weeks. Data are shown in the following table. (a) Using all the data, compute trial control limits for individual observations and moving-range charts. Construct the chart and plot the data. Determine whether the process is in statistical control. If not, assume assignable causes can be found to eliminate these samples and revise the control limits. (b) Estimate the process mean and standard deviation for the in-control process. Wafer 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

x 16.8 14.9 18.3 16.5 17.1 17.4 15.9 14.4 15.0 15.7 17.1 15.9 16.4 15.8 15.4

Wafer 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

15-19. The diameter of holes is measured in consecutive order by an automatic sensor. The results of measuring 25 holes are in the following table. Sample

Diameter

Sample

Diameter

1 2 3 4 5 6 7 8 9 10 11 12 13

9.94 9.93 10.09 9.98 10.11 9.99 10.11 9.84 9.82 10.38 9.99 10.41 10.36

14 15 16 17 18 19 20 21 22 23 24 25

9.99 10.12 9.81 9.73 10.14 9.96 10.06 10.11 9.95 9.92 10.09 9.85

(a) Using all the data, compute trial control limits for individual observations and moving-range charts. Construct the control chart and plot the data. Determine whether the process is in statistical control. If not, assume assignable causes can be found to eliminate these samples and revise the control limits. (b) Estimate the process mean and standard deviation for the in-control process. 15-20. The viscosity of a chemical intermediate is measured every hour. Twenty samples each of size n  1, are in the following table.

x

Sample

Viscosity

Sample

Viscosity

15.4 14.3 16.1 15.8 15.9 15.2 16.7 15.2 14.7 17.9 14.8 17.0 16.2 15.6 16.3

1 2 3 4 5 6 7 8 9 10

495 491 501 501 512 540 492 504 542 508

11 12 13 14 15 16 17 18 19 20

493 507 503 475 497 499 468 486 511 487

(a) Using all the data, compute trial control limits for individual observations and moving-range charts. Determine whether the process is in statistical control. If not, assume assignable causes can be found to eliminate these samples and revise the control limits.

JWCL232_c15_637-701.qxd

662

1/11/10

4:57 PM

Page 662

CHAPTER 15 STATISTICAL QUALITY CONTROL

(b) Estimate the process mean and standard deviation for the in-control process. 15-21. The following table of data was analyzed in Quality Engineering (1991–1992, Vol. 4(1)). The average particle size of raw material was obtained from 25 successive samples. Observation

Size

Observation

Size

1 2 3 4 5 6 7 8 9 10 11 12 13

96.1 94.4 116.2 98.8 95.0 120.3 104.8 88.4 106.8 96.8 100.9 117.7 115.6

14 15 16 17 18 19 20 21 22 23 24 25

100.5 103.1 93.1 93.7 72.4 87.4 96.1 97.1 95.7 94.2 102.4 131.9

(a) Using all the data, compute trial control limits for individual observations and moving-range charts. Construct the chart and plot the data. Determine whether the process is in statistical control. If not, assume assignable causes can be found to eliminate these samples and revise the control limits. (b) Estimate the process mean and standard deviation for the in-control process. 15-22. Pulsed laser deposition technique is a thin film deposition technique with a high-powered laser beam. Twenty-five films were deposited through this technique. The thicknesses of the films obtained are shown in the following table. Film

Thickness (in nm)

Film

Thickness (in nm)

1 2 3 4 5 6 7

28 45 34 29 37 52 29

8 9 10 11 12 13 14

51 23 35 47 50 32 40

15 16 17 18 19 20

46 59 20 33 56 49

21 22 23 24 25

21 62 34 31 98

(a) Using all the data, compute trial control limits for individual observations and moving-range charts. Determine whether the process is in statistical control. If not, assume assignable causes can be found to eliminate these samples, and revise the control limits. (b) Estimate the process mean and standard deviation for the in-control process. 15-23. The production manager of a soap manufacturing company wants to monitor the weights of the bars produced on the line. Twenty bars are taken during a stable period of the process. The weights of the bars are shown in the following table.

Bar

Weight (in g)

Bar

Weight (in g)

1 2 3 4 5 6 7 8 9 10

74 82 97 86 71 68 83 90 88 64

11 12 13 14 15 16 17 18 19 20

99 75 77 82 93 70 87 76 84 94

(a) Using all the data, compute trial control limits for individual observations and moving-range charts. Determine whether the process is in statistical control. If not, assume assignable causes can be found to eliminate these samples, and revise the control limits. (b) Estimate the process mean and standard deviation for the in-control process.

15-5 PROCESS CAPABILITY It is usually necessary to obtain some information about the process capability, that is, the performance of the process when it is operating in control. Two graphical tools, the tolerance chart (or tier chart) and the histogram, are helpful in assessing process capability. The tolerance chart for all 20 samples from the vane-manufacturing process is shown in Fig. 15-12. The specifications on vane opening are 0.5030 0.0010 in. In terms of the coded data, the upper specification limit is USL  40 and the lower specification limit is LSL  20, and these limits

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 663

15-5 PROCESS CAPABILITY

663

45

40

USL = 40

Vane opening

35

30 Nominal dimension = 30 25

20

Figure 15-12 Tolerance diagram of vane openings.

15

LSL = 20

5

10

15

20

Sample number

are shown on the chart in Fig. 15-12. Each measurement is plotted on the tolerance chart. Measurements from the same subgroup are connected with lines. The tolerance chart is useful in revealing patterns over time in the individual measurements, or it may show that a particular value of x or r was produced by one or two unusual observations in the sample. For example, note the two unusual observations in sample 9 and the single unusual observation in sample 8. Note also that it is appropriate to plot the specification limits on the tolerance chart, since it is a chart of individual measurements. It is never appropriate to plot specification limits on a control chart or to use the specifications in determining the control limits. Specification limits and control limits are unrelated. Finally, note from Fig. 15-12 that the process is running off-center from the nominal dimension of 30 (or 0.5030 inch). The histogram for the vane-opening measurements is shown in Fig. 15-13. The observations from samples 6, 8, 9, 11, and 19 (corresponding to out of-control points on either the X or R chart) have been deleted from this histogram. The general impression from examining this histogram is that the process is capable of meeting the specification but that it is running off-center. Another way to express process capability is in terms of an index that is defined as follows. Process Capability Ratio

The process capability ratio (PCR) is PCR 

USL  LSL 6

(15-20)

JWCL232_c15_637-701.qxd

664

1/11/10

4:57 PM

Page 664

CHAPTER 15 STATISTICAL QUALITY CONTROL 20

Frequency

15

10

5

Figure 15-13 Histogram for vane opening.

18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 LSL

Nominal dimension Vane opening

USL

The numerator of PCR is the width of the specifications. The limits 3 on either side of the process mean are sometimes called natural tolerance limits, for these represent limits that an in-control process should meet with most of the units produced. Consequently, 6 is often referred to as the width of the process. For the vane opening, where our sample size is 5, we could estimate  as ˆ 

r 5.0  2.15  2.326 d2

Therefore, the PCR is estimated to be PCR 

USL  LSL 40  20  1.55  6ˆ 612.152

The PCR has a natural interpretation: (1兾PCR)100% is just the percentage of the specifications’ width used by the process. Thus, the vane-opening process uses approximately (1兾1.55)100%  64.5% of the specifications’ width. Figure 15-14(a) shows a process for which the PCR exceeds unity. Since the process natural tolerance limits lie inside the specifications, very few defective or nonconforming units will be produced. If PCR  1, as shown in Fig.15-14(b), more nonconforming units result. In fact, for a normally distributed process, if PCR  1, the fraction nonconforming is 0.27%, or 2700 parts per million. Finally, when the PCR is less than unity, as in Fig. 15-14(c), the process is very yield-sensitive and a large number of nonconforming units will be produced. The definition of the PCR given in Equation 15-20 implicitly assumes that the process is centered at the nominal dimension. If the process is running off-center, its actual capability will be less than indicated by the PCR. It is convenient to think of PCR as a measure of potential capability, that is, capability with a centered process. If the

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 665

15-5 PROCESS CAPABILITY

665

PCR > 1

μ

LSL 3σ

USL 3σ

(a)

PCR = 1 Nonconforming units

Nonconforming units

μ 3σ LSL

3σ USL

(b)

PCR < 1 Nonconforming units

Figure 15-14 Process fallout and the process capability ratio (PCR).

Nonconforming units

LSL

μ

3σ

USL 3σ

(c)

process is not centered, a measure of actual capability is often used. This ratio, called PCRk , is defined below. PCRk

PCRk  min c

USL     LSL , d 3 3

(15-21)

In effect, PCRk is a one-sided process capability ratio that is calculated relative to the specification limit nearest to the process mean. For the vane-opening process, we find that the estimate of the process capability ratio PCRk (after deleting the samples corresponding to outof-control points) is USL  x x  LSL ^ PCR k  min c , d 3ˆ 3ˆ  min c

40  33.21 33.21  20  1.06,  2.04 d  1.05 312.152 312.152

Note that if PCR  PCRk, the process is centered at the nominal dimension. Since ^ PCR k  1.05 for the vane-opening process and ^ PCR  1.55, the process is obviously running off-center, as was first noted in Figs. 15-10 and 15-13. This off-center operation was

JWCL232_c15_637-701.qxd

666

1/11/10

4:57 PM

Page 666

CHAPTER 15 STATISTICAL QUALITY CONTROL

ultimately traced to an oversized wax tool. Changing the tooling resulted in a substantial improvement in the process. The fractions of nonconforming output (or fallout) below the lower specification limit and above the upper specification limit are often of interest. Suppose that the output from a normally distributed process in statistical control is denoted as X. The fractions are determined from P1X LSL2  P1Z 1LSL  2 2 EXAMPLE 15-3

P1X  USL2  P1Z  1USL  2 2

Electrical Current

For an electronic manufacturing process a current has specifications of 100 10 milliamperes. The process mean  and standard deviation  are 107.0 and 1.5, respectively. The process mean is nearer to the USL. Consequently,

distribution in Appendix Table II, P1X LSL2  P1Z 190  1072 1.52  P1Z 11.332 ⬵ 0

P1X  USL2  P1Z  1110  1072 1.52

110  90 PCR   2.22 6ⴢ5

 P1Z  22  0.023

and Practical Interpretation: The probability a current is less than the LSL is nearly zero. Consequently, the nonconforming output exceeds the USL. The PCRk would improve if the process mean were centered in the specifications at 100 milliamperes.

110  107  0.67 PCRk  3 ⴢ 15 The small PCRk indicates that the process is likely to produce currents outside of the specification limits. From the normal

For this example, the relatively large probability of exceeding the USL is a warning of potential problems with this criterion even if none of the measured observations in a preliminary sample exceed this limit. We emphasize that the fraction-nonconforming calculation assumes that the observations are normally distributed and the process is in control. Departures from normality can seriously affect the results. The calculation should be interpreted as an approximate guideline for process performance. To make matters worse,  and  need to be estimated from the data available and a small sample size can result in poor estimates that further degrade the calculation. Montgomery (2009) provides guidelines on appropriate values of the PCR and a table relating fallout for a normally distributed process in statistical control to the value of PCR. Many U.S. companies use PCR  1.33 as a minimum acceptable target and PCR  1.66 as a minimum target for strength, safety, or critical characteristics. Some companies require that internal processes and those at suppliers achieve a PCRk  2.0. Figure 15-15 illustrates a process with PCR  PCRk  2.0. Assuming a normal distribution, the calculated fallout for this process is 0.0018 parts per million. A process with PCRk  2.0 is referred to as a six-sigma process because the distance from the process mean to the nearest specification is six standard deviations. The reason that such a large process capability is often required is that it is difficult to maintain a process mean at the center of the specifications for long periods of time. A common model that 1.5 σ

PCRk = 2

Figure 15-15 Mean of a six-sigma process shifts by 1.5 standard deviations.

PCRk = 1.5

μ

LSL 3σ

USL 3σ

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 667

15-5 PROCESS CAPABILITY

667

is used to justify the importance of a six-sigma process is illustrated by referring to Fig. 15-15. If the process mean shifts off-center by 1.5 standard deviations, the PCRk decreases to PCRk 

USL   6  1.5 4.5    1.5 3 3 3

Assuming a normally distributed process, the fallout of the shifted process is 3.4 parts per million. Consequently, the mean of a 6-sigma process can shift 1.5 standard deviations from the center of the specifications and still maintain a fallout of 3.4 parts per million. In addition, some U.S. companies, particularly the automobile industry, have adopted the terminology Cp  PCR and Cpk  PCRk. Because Cp has another meaning in statistics (in multiple regression) we prefer the traditional notation PCR and PCRk. We repeat that process capability calculations are meaningful only for stable processes; that is, processes that are in control. A process capability ratio indicates whether or not the natural or chance variability in a process is acceptable relative to the specifications. EXERCISES FOR SECTION 15-5 15-24. Suppose that a quality characteristic is normally distributed with specifications at 100 20. The process standard deviation is 6. (a) Suppose that the process mean is 100. What are the natural tolerance limits? What is the fraction defective? Calculate PCR and PCRk and interpret these ratios. (b) Suppose that the process mean is 106. What are the natural tolerance limits? What is the fraction defective? Calculate PCR and PCRk and interpret these ratios. 15-25. Suppose that a quality characteristic is normally distributed with specifications from 20 to 32 units. (a) What value is needed for  to achieve a PCR of 1.5? (b) What value for the process mean minimizes the fraction defective? Does this choice for the mean depend on the value of ? 15-26. Suppose that a quality characteristic is normally distributed with specifications from 10 to 30 units. The process standard deviation is 2 units. (a) Calculate the natural tolerance limits, fraction defective, PCR, and PCRk when the process mean is 20. (b) Suppose the process mean shifts higher by 1.5 standard deviations. Recalculate the quantities in part (b). (c) Compare the results in parts (a) and (b) and comment on any differences. 15-27. A normally distributed process uses 66.7% of the specification band. It is centered at the nominal dimension, located halfway between the upper and lower specification limits. (a) Estimate PCR and PCRk. Interpret these ratios. (b) What fallout level (fraction defective) is produced? 15-28. A normally distributed process uses 85% of the specification band. It is centered at the nominal dimension, located halfway between the upper and lower specification limits. (a) Estimate PCR and PCRk. Interpret these ratios. (b) What fallout level (fraction defective) is produced?

15-29. Reconsider Exercise 15-1. Suppose that the quality characteristic is normally distributed with specification at 220 40. What is the fallout level? Estimate PCR and PCRk and interpret these ratios. 15-30. Reconsider Exercise 15-2, where the specification limits are 14.50 0.50. (a) What conclusions can you draw about the ability of the process to operate within these limits? Estimate the percentage of defective items that will be produced. (b) Estimate PCR and PCRk. Interpret these ratios. 15-31. Reconsider Exercise 15-3. Suppose that the variable is normally distributed with specifications at 220 50. What is the proportion out of specifications? Estimate and interpret PCR and PCRk. 15-32. Reconsider Exercise 15-4(a). Assuming that both charts exhibit statistical control and that the process specifications are at 20 5, estimate PCR and PCRk and interpret these ratios. 15-33. Reconsider Exercise 15-7. Use the revised control limits and process estimates. (a) Estimate PCR and PCRk. Interpret these ratios. (b) What percentage of defectives is being produced by this process? 15-34. Reconsider Exercise 15-8. Given that the specifications are at 6.0 1.0, estimate PCR and PCRk and interpret these ratios. 15-35. Reconsider Exercise 15-9. Using the process estimates, what is the fallout level if the specifications are 16 5? Estimate PCR and interpret this ratio. 15-36. Reconsider 15-20. The viscosity specifications are at 500 25. Calculate estimates of the process capability ratios PCR and PCRk for this process and provide an interpretation. 15-37. Suppose that a quality characteristic is normally distributed with specifications at 120 20. The process standard deviation is 6.5.

JWCL232_c15_637-701.qxd

668

1/11/10

4:57 PM

Page 668

CHAPTER 15 STATISTICAL QUALITY CONTROL

(c) What fallout level (fraction defective) is produced? 15-39. An X control chart with three-sigma control limits and subgroup size n = 4 has control limits UCL = 28.8 and LCL = 24.6. The process specification limits are (24, 32). (a) Estimate the process standard deviation. (b) Calculate PCR and PCRk for the process. 15-40. A control chart for individual observations has three-sigma control limits UCL = 1.80 and LCL = 1.62. The process specification limits are (1.64, 1.84). (a) Estimate the process standard deviation. (b) Calculate PCR and PCRk for the process.

(a) Suppose that the process mean is 120. What are the natural tolerance limits? What is the fraction defective? Calculate PCR and PCRk and interpret these ratios. (b) Suppose the process mean shifts off-center by 1.5 standard deviations toward the upper specification limit. Recalculate the quantities in part (a). (c) Compare the results in parts (a) and (b) and comment on any differences. 15-38. Suppose that a quality characteristic is normally distributed with specifications at 150 20. Natural tolerance limits for the process are 150 18. (a) Calculate the process standard deviation. (b) Calculate PCR and PCRk of the process. Calculate the percentage of the specification width used by the process.

15-6 ATTRIBUTE CONTROL CHARTS 15-6.1

P Chart (Control Chart for Proportions) Often it is desirable to classify a product as either defective or nondefective on the basis of comparison with a standard. This classification is usually done to achieve economy and simplicity in the inspection operation. For example, the diameter of a ball bearing may be checked by determining whether it will pass through a gauge consisting of circular holes cut in a template. This kind of measurement would be much simpler than directly measuring the diameter with a device such as a micrometer. Control charts for attributes are used in these situations. Attribute control charts often require a considerably larger sample size than do their variable measurements counterparts. In this section, we discuss the fraction-defective control chart, or P chart. Sometimes the P chart is called the control chart for fraction nonconforming. Suppose D is the number of defective units in a random sample of size n. We assume that D is a binomial random variable with unknown parameter p. The fraction defective D P n of each sample is plotted on the chart. Furthermore, the variance of the statistic P is 2P 

p11  p2 n

Therefore, a P chart for fraction defective could be constructed using p as the center line and control limits at UCL  p  3

B

p11  p2 n

LCL  p  3

B

p11  p2 n

(15-22)

However, the true process fraction defective is almost always unknown and must be estimated using the data from preliminary samples. Suppose that m preliminary samples each of size n are available, and let Di be the number of defectives in the ith sample. Then Pi  Di n is the sample fraction defective in the ith sample. The average fraction defective is

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 669

669

15-6 ATTRIBUTE CONTROL CHARTS

1 m 1 m P  m a Pi  mn a Di i1 i1

(15-23)

Now P may be used as an estimator of p in the center line and control limit formulas. P Chart

The center line and upper and lower control limits for the P chart are UCL  p  3

p11  p2 n B

CL  p LCL  p  3

p11  p2 n B

(15-24)

where p is the observed value of the average fraction defective.

These control limits are based on the normal approximation to the binomial distribution. When p is small, the normal approximation may not always be adequate. In such cases, we may use control limits obtained directly from a table of binomial probabilities. If p is small, the lower control limit obtained from the normal approximation may be a negative number. If this should occur, it is customary to consider zero as the lower control limit.

EXAMPLE 15-4 Ceramic Substrate Suppose we wish to construct a fraction-defective control chart for a ceramic substrate production line. We have 20 preliminary samples, each of size 100; the number of defectives in each sample is shown in Table 15-4. Assume that the samples are numbered in the sequence of production. Note that p  (800兾2000)  0.40; therefore, the trial parameters for the control chart are UCL  0.40  3 LCL  0.40  3

B

B

10.402 10.602

 0.55

10.402 10.602

 0.25

100

100

CL  0.40

The control chart is shown in Fig. 15-16. All samples are in control. If they were not, we would search for assignable causes of variation and revise the limits accordingly. This chart can be used for controlling future production. Practical Interpretation: Although this process exhibits statistical control, its defective rate ( p  0.40) is very poor. We should take appropriate steps to investigate the process to determine why such a large number of defective units is being produced. Defective units should be analyzed to determine the specific types of defects present. Once the defect types are known, process changes should be investigated to determine their impact on defect levels. Designed experiments may be useful in this regard.

Table 15-4 Number of Defectives in Samples of 100 Ceramic Substrates Sample

No. of Defectives

Sample

No. of Defectives

1 2 3 4 5 6 7 8 9 10

44 48 32 50 29 31 46 52 44 48

11 12 13 14 15 16 17 18 19 20

36 52 35 41 42 30 46 38 26 30

JWCL232_c15_637-701.qxd

5:19 PM

Page 670

CHAPTER 15 STATISTICAL QUALITY CONTROL

Sample fraction defective, p

670

1/20/10

Figure 15-16 P chart for a ceramic substrate.

0.6

UCL

0.5 0.4 0.3

p LCL

0.2 0.1 1 2

3

4

5

6

7

8 9 10 11 12 13 14 15 16 17 18 19 20 Sample number

Computer software also produces an NP chart. This is just a control chart of nP ⫽ D, the number of defectives in a sample. The points, center line, and control limits for this chart are simply multiples (times n) of the corresponding elements of a P chart. The use of an NP chart avoids the fractions in a P chart but it is otherwise equivalent.

15-6.2

U Chart (Control Chart for Defects per Unit) It is sometimes necessary to monitor the number of defects in a unit of product rather than the fraction defective. Suppose that in the production of cloth it is necessary to control the number of defects per yard or that in assembling an aircraft wing the number of missing rivets must be controlled. In these situations we may use the control chart for defects per unit, or the U chart. Many defects-per-unit situations can be modeled by the Poisson distribution. If each sample consists of n units and there are C total defects in the sample, C U⫽ n is the average number of defects per unit. A U chart may be constructed for such data. If the number of defects in a unit is a Poisson random variable with parameter ␭, the mean and variance of this distribution are both ␭. Each point on the chart is an observed value of U, the average number of defects per unit from a sample of n units. The mean of U is ␭ and the variance of U is ␭兾n. Therefore, the control limits for the U chart with known ␭ are: UCL ⫽ ␭ ⫹ 3

␭ Bn

LCL ⫽ ␭ ⫺ 3

␭ Bn

(15-25)

If there are m preliminary samples, and the number of defects per unit in these samples are U1, U2, . . . , Um, the estimator of the average number of defects per unit is 1 m U ⫽ m a Ui i⫽1 Now U is used as an estimator of ␭ in the centerline and control limit formulas.

(15-26)

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 671

671

15-6 ATTRIBUTE CONTROL CHARTS

U Chart

The center line and upper and lower control limits on the U chart are UCL  u  3

u Bn

CL  u

LCL  u  3

u Bn

(15-27)

where u is the average number of defects per unit. These control limits are based on the normal approximation to the Poisson distribution. When is small, the normal approximation may not always be adequate. In such cases, we may use control limits obtained directly from a table of Poisson probabilities. If u is small, the lower control limit obtained from the normal approximation may be a negative number. If this should occur, it is customary to use zero as the lower control limit. EXAMPLE 15-5 Printed Circuit Boards Printed circuit boards are assembled by a combination of manual assembly and automation. Surface Mount Technology (SMT) is used to make the mechanical and electrical connections of the components to the board. Every hour, five boards are selected and inspected for process-control purposes. The number of defects in each sample of five boards is noted. Results for 20 samples are shown in Table 15-5. The center line for the U chart is u

1 20 32.0 a ui  20  1.6 20 i1

and the upper and lower control limits are

UCL  u  3

u 1.6  1.6  3  3.3 n B B 5

LCL  u  3

u 1.6  1.6  3 0 Bn B 5

The control chart is plotted in Fig. 15-17. Because LCL is negative, it is set to 0. From the control chart in Fig. 15-17, we see that the process is in control. Practical Interpretation: Eight defects per group of five circuit boards are too many (about 8兾5  1.6 defects/board), and the process needs improvement. An investigation needs to be made of the specific types of defects found on the printed circuit boards. This will usually suggest potential avenues for process improvement.

Computer software also produces a C chart. This is just a control chart of C, the total of defects in a sample. The points, center line, and control limits for this chart are simply multiples (times n) of the corresponding elements of a U chart. The use of a C chart avoids the fractions that can occur in a U chart but it is otherwise equivalent. Table 15-5 Number of Defects in Samples of Five Printed Circuit Boards Sample

Number of Defects

Defects per Unit ui

Sample

Number of Defects

Defects per Unit ui

1 2 3 4 5 6 7 8 9 10

6 4 8 10 9 12 16 2 3 10

1.2 0.8 1.6 2.0 1.8 2.4 3.2 0.4 0.6 2.0

11 12 13 14 15 16 17 18 19 20

9 15 8 10 8 2 7 1 7 13

1.8 3.0 1.6 2.0 1.6 0.4 1.4 0.2 1.4 2.6

JWCL232_c15_637-701.qxd

672

1/11/10

4:57 PM

Page 672

CHAPTER 15 STATISTICAL QUALITY CONTROL 4

Defects per unit, u

UCL = 3.3

Figure 15-17 U chart of defects per unit on printed circuit boards.

2

0

u = 1.6

1 2

3

4

5

6

7

8 9 10 11 12 13 14 15 16 17 18 19 20 Sample number

EXERCISES FOR SECTION 15-6 15-41. An early example of SPC was described in Industrial Quality Control [“The Introduction of Quality Control at Colonial Radio Corporation” (1944, Vol. 1(1), pp. 4– 9)]. The following are the fractions defective of shaft and washer assemblies during the month of April in samples of n  1500 each:

Sample

Fraction Defective

Sample

Fraction Defective

1 2 3 4 5 6 7 8 9 10

0.11 0.06 0.1 0.11 0.14 0.11 0.14 0.03 0.02 0.03

11 12 13 14 15 16 17 18 19 20

0.03 0.03 0.04 0.07 0.04 0.04 0.04 0.03 0.06 0.06

(a) Set up a P chart for this process. Is this process in statistical control? (b) Suppose that instead of n  1500, n  100. Use the data given to set up a P chart for this process. Revise the control limits if necessary. (c) Compare your control limits for the P charts in parts (a) and (b). Explain why they differ. Also, explain why your assessment about statistical control differs for the two sizes of n. 15-42. Suppose the following fraction defective has been found in successive samples of size 100 (read down): (a) Using all the data, compute trial control limits for a fractiondefective control chart, construct the chart, and plot the data. 0.09 0.10 0.13

0.03 0.05 0.13

0.12 0.14 0.06

0.08 0.14 0.09 0.10 0.15 0.13 0.06

0.10 0.14 0.07 0.06 0.09 0.08 0.11

0.05 0.14 0.11 0.09 0.13 0.12 0.09

(b) Determine whether the process is in statistical control. If not, assume assignable causes can be found and out-ofcontrol points eliminated. Revise the control limits. 15-43. The following are the numbers of defective solder joints found during successive samples of 500 solder joints: Day

No. of Defectives

Day

1 2 3 4 5 6 7 8 9 10 11

106 116 164 89 99 40 112 36 69 74 42

12 13 14 15 16 17 18 19 20 21

No. of Defectives 37 25 88 101 64 51 74 71 43 80

(a) Using all the data, compute trial control limits for a fraction-defective control chart, construct the chart, and plot the data. (b) Determine whether the process is in statistical control. If not, assume assignable causes can be found and out-ofcontrol points eliminated. Revise the control limits. 15-44. The following represent the number of defects per 1000 feet in rubber-covered wire: 1, 1, 3, 7, 8, 10, 5, 13, 0,

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 673

673

15-7 CONTROL CHART PERFORMANCE

19, 24, 6, 9, 11, 15, 8, 3, 6, 7, 4, 9, 20, 11, 7, 18, 10, 6, 4, 0, 9, 7, 3, 1, 8, 12. Do the data come from a controlled process? 15-45. The following represent the number of solder defects observed on 24 samples of five printed circuit boards: 7, 6, 8, 10, 24, 6, 5, 4, 8, 11, 15, 8, 4, 16, 11, 12, 8, 6, 5, 9, 7, 14, 8, 21. (a) Using all the data, compute trial control limits for a U control chart, construct the chart, and plot the data. (b) Can we conclude that the process is in control using a U chart? If not, assume assignable causes can be found, list points and revise the control limits. 15-46. Consider the data on the number of earthquakes of magnitude 7.0 or greater by year in Exercise 6-71. (a) Construct a U chart for this data with a sample size of n  1. (b) Do the data appear to be generated by an in-control process? Explain. 15-47. In a semiconductor manufacturing company, samples of 200 wafers are tested for defectives in the lot. The number of defectives in 20 such samples is shown in the following table.

Sample

No. of Defectives

Sample

No. of Defectives

1 2 3 4 5 6 7 8 9 10

44 63 40 35 29 56 40 38 74 66

11 12 13 14 15 16 17 18 19 20

52 74 43 50 60 38 36 65 41 95

(a) Set up a P chart for this process. Is the process in statistical control? (b) Suppose that instead of samples of size 200, we have samples of size 100. Use the data to set up a P chart for this process. Revise the control limits if necessary. (c) Compare the control limits in parts (a) and (b). Explain why they differ. 15-48. The following data are the number of spelling errors detected for every 1000 words on a news Web site over 20 weeks.

Week

No. of Spelling Errors

Week

1 2 3 4 5 6 7 8 9 10

3 6 0 5 9 5 2 2 3 2

11 12 13 14 15 16 17 18 19 20

No. of Spelling Errors 1 6 9 8 6 4 13 3 0 7

(a) What control chart is most appropriate for these data? (b) Using all the data, compute trial control limits for the chart in part (a), construct the chart, and plot the data. (c) Determine whether the process is in statistical control. If not, assume assignable causes can be found and out-ofcontrol points eliminated. Revise the control limits.

15-7 CONTROL CHART PERFORMANCE Specifying the control limits is one of the critical decisions that must be made in designing a control chart. By moving the control limits further from the center line, we decrease the risk of a type I error—that is, the risk of a point falling beyond the control limits, indicating an out-of-control condition when no assignable cause is present. However, widening the control limits will also increase the risk of a type II error—that is, the risk of a point falling between the control limits when the process is really out of control. If we move the control limits closer to the center line, the opposite effect is obtained: The risk of type I error is increased, while the risk of type II error is decreased. The control limits on a Shewhart control chart are customarily located a distance of plus or minus three standard deviations of the variable plotted on the chart from the center line. That is, the constant k in Equation 15-1 should be set equal to 3. These limits are called 3-sigma control limits.

JWCL232_c15_637-701.qxd

674

1/11/10

4:57 PM

Page 674

CHAPTER 15 STATISTICAL QUALITY CONTROL

A way to evaluate decisions regarding sample size and sampling frequency is through the average run length (ARL) of the control chart. Essentially, the ARL is the average number of points that must be plotted before a point indicates an out-of-control condition. For any Shewhart control chart, the ARL can be calculated from the mean of a geometric random variable. Suppose that p is the probability that any point exceeds the control limits. Then Average Run Length

1 ARL  p

(15-28)

Thus, for an X chart with 3-sigma limits, p  0.0027 is the probability that a normally distributed point falls outside the limits when the process is in control, so 1 1 ⬵ 370 ARL  p  0.0027 is the average run length of the X chart when the process is in control. That is, even if the process remains in control, an out-of-control signal will be generated every 370 points, on the average. Consider the piston ring process discussed in Section 15-2.2, and suppose we are sampling every hour. Thus, we will have a false alarm about every 370 hours on the average. Suppose we are using a sample size of n  5 and that when the process goes out of control the mean shifts to 74.0135 millimeters. Then, the probability that X falls between the control limits of Fig. 15-3 is equal to P 373.9865  X  74.0135 when   74.01354 74.0135  74.0135 73.9865  74.0135 Z d Pc 0.0045 0.0045  P 36  Z  04  0.5 Therefore, p in Equation 15-28 is 0.50, and the out-of-control ARL is 1 1 2 ARL  p  0.5 That is, the control chart will require two samples to detect the process shift, on the average, so two hours will elapse between the shift and its detection (again, on the average). Suppose this approach is unacceptable, because production of piston rings with a mean diameter of 74.0135 millimeters results in excessive scrap costs and delays final engine assembly. How can we reduce the time needed to detect the out-of-control condition? One method is to sample more frequently. For example, if we sample every half hour, only one hour will elapse (on the average) between the shift and its detection. The second possibility is to increase the sample size. For example, if we use n  10, the control limits in Fig. 15-3 narrow to 73.9905 and 74.0095. The probability of X falling between the control limits when the process mean is 74.0135 millimeters is approximately 0.1, so p  0.9, and the out-ofcontrol ARL is 1 1  1.11 ARL  p  0.9

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 675

15-7 CONTROL CHART PERFORMANCE

675

Table 15-6 Average Run Length (ARL) for an X Chart with 3-Sigma Control Limits Magnitude of Process Shift

ARL n1

ARL n4

0 0.5 1.0 1.5 2.0 3.0

370.4 155.2 43.9 15.0 6.3 2.0

370.4 43.9 6.3 2.0 1.2 1.0

Thus, the larger sample size would allow the shift to be detected about twice as quickly as the old one. If it became important to detect the shift in approximately the first hour after it occurred, two control chart designs would work: Design 1

Design 2

Sample size: n  5 Sampling frequency: every half hour

Sample size: n  10 Sampling frequency: every hour

Table 15-6 provides average run lengths for an X chart with 3-sigma control limits. The average run lengths are calculated for shifts in the process mean from 0 to 3.0 and for sample sizes of n  1 and n  4 by using 1兾p, where p is the probability that a point plots outside of the control limits. Figure 15-18 illustrates a shift in the process mean of 2. EXERCISES FOR SECTION 15-7 15-49. An X chart uses samples of size 1. The center line is at 100 and the upper and lower 3-sigma limits are at 112 and 88, respectively. (a) What is the process ? (b) Suppose the process mean shifts to 96. Find the probability that this shift will be detected on the next sample. (c) Find the ARL to detect the shift in part (b). 15-50. An X chart uses samples of size 4. The center line is at 100, and the upper and lower 3-sigma control limits are at 106 and 94, respectively. (a) What is the process ? (b) Suppose the process mean shifts to 96. Find the probability that this shift will be detected on the next sample. (c) Find the ARL to detect the shift in part (b). 15-51. Consider the X control chart in Fig. 15-3. Suppose that the mean shifts to 74.010 millimeters.

Figure 15-18 Process mean shift of 2.

μ

μ + 2σ

(a) What is the probability that this shift will be detected on the next sample? (b) What is the ARL after the shift? 15-52. Consider an X control chart with r  0.344 , UCL  14.708, LCL  14.312, and n  5. Suppose that the mean shifts to 14.6. (a) What is the probability that this shift will be detected on the next sample? (b) What is the ARL after the shift? 15-53. Consider an X control chart with r  34.286 , UCL  242.780, LCL  203.220, and n  5. Suppose that the mean shifts to 210. (a) What is the probability that this shift will be detected on the next sample? (b) What is the ARL after the shift?

JWCL232_c15_637-701.qxd

676

1/21/10

12:13 AM

Page 676

CHAPTER 15 STATISTICAL QUALITY CONTROL

15-54. Consider an X control chart with ␴ˆ ⫽ 1.40 , UCL ⫽ 21.71, LCL ⫽ 18.29, and n ⫽ 6. Suppose that the mean shifts to 17. (a) What is the probability that this shift will be detected on the next sample? (b) What is the ARL after the shift? 15-55. Consider an X control chart with ␴ˆ ⫽ 2.466, UCL ⫽ 37.404, LCL ⫽ 30.780, and n ⫽ 5. Suppose that the mean shifts to 36. (a) What is the probability that this shift will be detected on the next sample? (b) What is the ARL after the shift? 15-56. Consider an X control chart with r ⫽ 2.25, UCL ⫽ 17.40, LCL ⫽ 12.79, and n ⫽ 3. Suppose that the mean shifts to 13. (a) What is the probability that this shift will be detected on the next sample? (b) What is the ARL after the shift? 15-57. Consider an X control chart with r ⫽ 0.000924, UCL ⫽ 0.0635, LCL ⫽ 0.0624, and n ⫽ 5. Suppose that the mean shifts to 0.0625.

(a) What is the probability that this shift will be detected on the next sample? (b) What is the ARL after the shift? 15-58. Consider the revised X control chart in Exercise 15-8 with ␴ˆ ⫽ 0.669 , UCL ⫽ 7.443, LCL ⫽ 5.125, and n ⫽ 3. Suppose that the mean shifts to 5.5. (a) What is the probability that this shift will be detected on the next sample? (b) What is the ARL after the shift? 15-59. An X chart uses a subgroup of size three. The center line is at 200, and the upper and lower three-sigma control limits are at 212 and 188, respectively. (a) Estimate the process ␴. (b) Suppose the process mean shifts to 195. Determine the probability that this shift will be detected on the next sample. (c) Find the ARL to detect the shift in part (b). 15-60. Consider an X control chart with UCL = 24.802, LCL = 23.792, and n = 3. Suppose the mean shifts to 24.2. (a) What is the probability that this shift will be detected on the next sample? (b) What is the ARL to detect the shift?

15-8 TIME-WEIGHTED CHARTS 15-8.1 Cumulative Sum Control Chart In Sections 15-3 and 15-4 we have presented basic types of Shewhart control charts. A major disadvantage of any Shewhart control chart is that the chart is relatively insensitive to small shifts in the process, say, on the order of about 1.5␴ or less. One reason for this relatively poor performance in detecting small process shifts is that the Shewhart chart makes use of only the information in the last plotted point, and it ignores the information in the sequence of points. This problem can be addressed, to some extent by adding criteria such as the Western Electric rules to a Shewhart chart, but the use of these rules reduces the simplicity and ease of interpretation of the chart. These rules would also cause the in-control average run length of a Shewhart chart to drop below 370. This increase in the false alarm rate can have serious practical consequences. A very effective alternative to the Shewhart control chart is the cumulative sum control chart (or CUSUM). This chart has much better performance (in terms of ARL) for detecting small shifts than the Shewhart chart, but it does not cause the in-control ARL to drop significantly. This section will illustrate the use of the CUSUM for sample averages and individual measurements. The CUSUM chart plots the cumulative sums of the deviations of the sample values from a target value. For example, suppose that samples of size n ⱖ 1 are collected, and Xj is the average of the jth sample. Then if ␮0 is the target for the process mean, the cumulative sum control chart is formed by plotting the quantity i

Si ⫽ a 1Xj ⫺ ␮0 2

(15-29)

j⫽1

against the sample number i. Now, Si is called the cumulative sum up to and including the ith sample. Because they combine information from several samples, cumulative sum charts are

JWCL232_c15_637-701.qxd

1/11/10

4:57 PM

Page 677

15-8 TIME-WEIGHTED CHARTS

677

more effective than Shewhart charts for detecting small process shifts. Furthermore, they are particularly effective with samples of n  1. This makes the cumulative sum control chart a good candidate for use in the chemical and process industries where rational subgroups are frequently of size 1, as well as in discrete parts manufacturing with automatic measurement of each part and online control using a computer directly at the work center. If the process remains in control at the target value 0, the cumulative sum defined in Equation 15-29 should fluctuate around zero. However, if the mean shifts upward to some value 1  0, say, an upward or positive drift develops in the cumulative sum Si. Conversely, if the mean shifts downward to some 1 0, a downward or negative drift in Si develops. Therefore, if a trend develops in the plotted points either upward or downward, we should consider this as evidence that the process mean has shifted, and a search for the assignable cause should be performed. This theory can easily be demonstrated by applying the CUSUM to the chemical process concentration data in Table 15-3. Since the concentration readings are individual measurements, we would take Xj  Xj in computing the CUSUM. Suppose that the target value for the concentration is 0  99. Then the CUSUM is i

Si  a 1Xj  992 j1

i1

 1Xi  992  a 1Xj  992 j1

 1Xi  992  Si1

Table 15-7 shows the computation of this CUSUM, where the starting value of the CUSUM, S0, is taken to be zero. Figure 15-19 plots the CUSUM from the last column of Table 15-7. Notice that the CUSUM fluctuates) around the value of 0. Table 15-7 CUSUM Computations for the Chemical Process Concentration Data in Table 15-3 Observation, i 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

xi

xi  99

s i  (x i  99)  s i1

102.0 94.8 98.3 98.4 102.0 98.5 99.0 97.7 100.0 98.1 101.3 98.7 101.1 98.4 97.0 96.7 100.3 101.4 97.2 101.0

3.0 4.2 0.7 0.6 3.0 0.5 0.0 1.3 1.0 0.9 2.3 0.3 2.1 0.6 2.0 2.3 1.3 2.4 1.8 2.0

3.0 1.2 1.9 2.5 0.5 0.0 0.0 1.3 0.3 1.2 1.1 0.8 2.9 2.3 0.3 2.0 0.7 1.7 0.1 1.9

JWCL232_c15_637-701.qxd

678

1/11/10

4:57 PM

Page 678

CHAPTER 15 STATISTICAL QUALITY CONTROL +4

+2

si 0

–2

Figure 15-19 Plot of the cumulative sum for the concentration data, Table 15-7.

–4

1 2

3

4

5

6

7

8 9 10 11 12 13 14 15 16 17 18 19 20 Observation, i

The graph in Fig. 15-19 is not a control chart because it lacks control limits. There are two general approaches to devising control limits for CUSUMs. The older of these two methods is the V-mask procedure. A typical V mask is shown in Fig. 15-20(a). It is a V-shaped notch in a plane that can be placed at different locations on the CUSUM chart. The decision procedure consists of placing the V mask on the cumulative sum control chart with the point O on the last value of si and the line OP parallel to the horizontal axis. If all the previous cumulative sums, s1, s2, . . . , si1, lie within the two arms of the V mask, the process is in control. The arms are the lines that make angles  with segment OP in Figure 15-20(a) and they are assumed to extend infinitely in length. However, if any si lies outside the arms of the mask, the process is considered to be out of control. In actual use, the V mask would be

+6

+4 K

U

θ

O

si

d L

3A

si

2A

+2

P 0

–2

A 1

2

3

4

...

i

–4

1

5

10

15

20

25

30

Observation, i (a)

(b)

Figure 15-20 The cumulative sum control chart. (a) The V-mask and scaling. (b) The cumulative sum control chart in operation.

JWCL232_c15_637-701.qxd

1/11/10

9:21 PM

Page 679

15-8 TIME-WEIGHTED CHARTS

679

applied to each new point on the CUSUM chart as soon as it was plotted. In the example shown in Fig. 15-20(b), an upward shift in the mean is indicated, since at least one of the points that have occurred earlier than sample 22 now lies below the lower arm of the mask, when the V mask is centered on sample 30. If the point lies above the upper arm, a downward shift in the mean is indicated. Thus, the V mask forms a visual frame of reference similar to the control limits on an ordinary Shewhart control chart. For the technical details of designing the V mask, see Montgomery (2009). While some computer programs plot CUSUMs with the V-mask control scheme, we feel that the other approach to CUSUM control, the tabular CUSUM, is superior. The tabular procedure is particularly attractive when the CUSUM is implemented on a computer. Let SH (i) be an upper one-sided CUSUM for period i and SL(i) be a lower one-sided CUSUM for period i. These quantities are calculated from CUSUM Control Chart

sH 1i2 ⫽ max 30, xi ⫺ 1␮0 ⫹ K2 ⫹ sH 1i ⫺ 12 4

(15-30)

sL 1i2 ⫽ max 30, 1␮0 ⫺ K2 ⫺ xi ⫹ sL 1i ⫺ 12 4

(15-31)

and

where the starting values sH 102 ⫽ sL 102 ⫽ 0.

In Equations 15-30 and 15-31 K is called the reference value, which is usually chosen about halfway between the target ␮0 and the value of the mean corresponding to the out-of-control state, ␮1 ⫽ ␮0 ⫹ ⌬. That is, K is about one-half the magnitude of the shift we are interested in, or K⫽

⌬ 2

Notice that SH (i) and SL (i) accumulate deviations from the target value that are greater than K, with both quantities reset to zero upon becoming negative. If either SH (i) or SL(i) exceeds a constant H, the process is out of control. This constant H is usually called the decision interval. EXAMPLE 15-6

Chemical Process Concentration CUSUM

A Tabular CUSUM We will illustrate the tabular CUSUM by applying it to the chemical process concentration data in Table 15-7. The process target is ␮0 ⫽ 99, and we will use K ⫽ 1 as the reference value and H ⫽ 10 as the decision interval. The reasons for these choices will be explained later. Table 15-8 shows the tabular CUSUM scheme for the chemical process concentration data. To illustrate the calculations, note that sH 1i2 ⫽ max 30, xi ⫺ 1␮0 ⫹ K2 ⫹ sH 1i ⫺ 12 4 ⫽ max 30, xi ⫺ 199 ⫹ 12 ⫹ sH 1i ⫺ 12 4

⫽ max 30, xi ⫺ 100 ⫹ sH 1i ⫺ 12 4

sL 1i2 ⫽ max 30, 1␮0 ⫺ K2 ⫺ xi ⫹ sL 1i ⫺ 12 4 ⫽ max 30, 199 ⫺ 12 ⫺ xi ⫹ sL 1i ⫺ 12 4

⫽ max 30, 98 ⫺ xi ⫹ sL 1i ⫺ 12 4

Therefore, for observation 1 the CUSUMs are sH 112 ⫽ max 3 0, x1 ⫺ 100 ⫹ sH 102 4

⫽ max 30, 102.0 ⫺ 100 ⫹ 04 ⫽ 2.0

and

sL 11 2 ⫽ max 30, 98 ⫺ x1 ⫹ sL 102 4

⫽ max 30, 98 ⫺ 102.0 ⫹ 0 4 ⫽ 0

JWCL232_c15_637-701.qxd

680

1/11/10

4:57 PM

Page 680

CHAPTER 15 STATISTICAL QUALITY CONTROL

Table 15-8 The Tabular CUSUM for the Chemical Process Concentration Data Observation i 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Upper CUSUM

Lower CUSUM

xi

xi  100

sH (i)

nH

98  xi

sL(i)

nL

102.0 94.8 98.3 98.4 102.0 98.5 99.0 97.7 100.0 98.1 101.3 98.7 101.1 98.4 97.0 96.7 100.3 101.4 97.2 101.0

2.0 5.2 1.7 1.6 2.0 1.5 1.0 2.3 0.0 1.9 1.3 1.3 1.1 1.6 3.0 3.3 0.3 1.4 2.8 1.0

2.0 0.0 0.0 0.0 2.0 0.5 0.0 0.0 0.0 0.0 1.3 0.0 1.1 0.0 0.0 0.0 0.3 1.7 0.0 1.0

1 0 0 0 1 2 0 0 0 0 1 0 1 0 0 0 1 2 0 0

4.0 3.2 0.3 0.4 4.0 0.5 1.0 0.3 2.0 0.1 3.3 0.7 3.1 0.4 1.0 1.3 2.3 3.4 0.8 3.0

0.0 3.2 2.9 2.5 0.0 0.0 0.0 0.3 0.0 0.0 0.0 0.0 0.0 0.0 1.0 2.3 0.0 0.0 0.8 0.0

0 1 2 3 0 0 0 1 0 0 0 0 0 0 1 2 0 0 1 0

as shown in Table 15-8. The quantities nH and nL in Table 15-8 indicate the number of periods that the CUSUM sH (i) or sL(i) have been nonzero. Notice that the CUSUMs in this example never exceed the decision interval H  10. We would therefore conclude that the process is in control.

Next Steps: The limits for the CUSUM charts may be used to continue to operate the chart in order to monitor future productions.

When the tabular CUSUM indicates that the process is out of control, we should search for the assignable cause, take any corrective actions indicated, and restart the CUSUMs at zero. It may be helpful to have an estimate of the new process mean following the shift. This can be computed from

ˆ  μ 

sH 1i2 0  K  n , H sL 1i2 0  K  n , L

if sH 1i2  H if sL 1i2  H

(15-32)

It is also useful to present a graphical display of the tabular CUSUMs, which are sometimes called CUSUM status charts. They are constructed by plotting sH (i) and sL(i) versus the sample number. Figure 15-21 shows the CUSUM status chart for the data in Example 15-6. Each vertical bar represents the value of sH (i) and sL(i) in period i. With the decision interval plotted on the chart, the CUSUM status chart resembles a Shewhart control chart. We have also plotted the sample statistics xi for each period on the CUSUM status chart as the

JWCL232_c15_637-701.qxd

1/11/10

4:58 PM

Page 681

15-8 TIME-WEIGHTED CHARTS 106

6

105

5

104

4

103

3

102 101 x

1 0

99

Figure 15-21 The CUSUM status chart for Example 15-6.

H=5

2 sH (i)

100

98

681

1 sL(i)

2

97

3

96

4

95

5

94

6

H=5

1

2

3

4

5

6

7

8

9 10 11 12 13 14 15 16 17 18 19 20 Sample number

solid dots. This frequently helps the user of the control chart to visualize the actual process performance that has led to a particular value of the CUSUM. The tabular CUSUM is designed by choosing values for the reference value K and the decision interval H. We recommend that these parameters be selected to provide good average run-length values. There have been many analytical studies of CUSUM ARL performance. Based on these studies, we may give some general recommendations for selecting H and K. Define H  hX and K  kX , where X is the standard deviation of the sample variable used in forming the CUSUM (if n  1, X  X ). Using h  4 or h  5 and k  1兾2 will generally provide a CUSUM that has good ARL properties against a shift of about 1X (or 1X) in the process mean. If much larger or smaller shifts are of interest, set k  兾2, where  is the size of the shift in standard deviation units. To illustrate how well the recommendations of h  4 or h  5 with k  1兾2 work, consider these average run lengths in Table 15-9. Notice that a shift of 1X would be detected in Table 15-9 Average Run Lengths for a CUSUM Control Chart with k = 1兾2 Shift in Mean (multiple of  X )

h4

h5

0 0.25 0.50 0.75 1.00 1.50 2.00 2.50 3.00 4.00

168 74.2 26.6 13.3 8.38 4.75 3.34 2.62 2.19 1.71

465 139 38.0 17.0 10.4 5.75 4.01 3.11 2.57 2.01

JWCL232_c15_637-701.qxd

682

1/11/10

9:26 PM

Page 682

CHAPTER 15 STATISTICAL QUALITY CONTROL

either 8.38 samples (with k ⫽ 1兾2 and h ⫽ 4) or 10.4 samples (with k ⫽ 1兾2 and h ⫽ 5). By comparison, Table 15-6 shows that an X chart would require approximately 43.9 samples, on the average, to detect this shift. These design rules were used for the CUSUM in Example 15-6. We assumed that the process standard deviation ␴ ⫽ 2. (This is a reasonable value; see Example 15-2.) Then with k ⫽ 1兾2 and h ⫽ 5, we would use K ⫽ k␴ ⫽ 1⁄2 122 ⫽ 1

and

H ⫽ h␴ ⫽ 5122 ⫽ 10

in the tabular CUSUM procedure. Finally, we should note that supplemental procedures such as the Western Electric rules cannot be safely applied to the CUSUM, because successive values of SH (i) and SL(i) are not independent. In fact, the CUSUM can be thought of as a weighted average, where the weights are stochastic or random. In effect, all the CUSUM values are highly correlated, thereby causing the Western Electric rules to give too many false alarms.

15-8.2 Exponentially Weighted Moving Average Control Chart Data collected in time order is often averaged over several time periods. For example, economic data is often presented as an average over the last four quarters. That is, at time t the average of the last four measurements can be written as xt 142 ⫽

1 1 1 1 x ⫹ x ⫹ x ⫹ x 4 t 4 t⫺1 4 t⫺2 4 t⫺3

This average places weight of 1兾4 on each of the most recent observations, and zero weight on older observations. It is called a moving average and in this case a window of size 4 is used. An average of the recent data is used to smooth the noise in the data to generate a better estimate of the process mean than only the most recent observation. However, in a dynamic environment where the process mean may change, the number of observations used to construct the average is kept to a modest size so that the estimate can adjust to any change in the process mean. Therefore, the window size is a compromise between a better statistical estimate from an average and a response to a mean change. If a window of size 10 were used in a moving average, the statistic xt 1102 would have lower variability, but it would not adjust as well to a change in mean. For statistical process control, rather than use a fixed window size it is useful to place the greatest weight on the most recent observation or subgroup average, and then gradually decrease the weight on older observations. One average of this type can be constructed by a multiplicative decrease in the weights. Let ␭ ⱕ 1 denote a constant and ␮0 denote the process target or historical mean. Suppose that samples of size n ⱖ 1 are collected and xt is the average of the sample at time t. The exponentially weighted moving average (EWMA) is zt ⫽ ␭xt ⫹ ␭11 ⫺ ␭2xt⫺1 ⫹ ␭11 ⫺ ␭2 2 xt⫺2 ⫹ p ⫹ ␭11 ⫺ ␭2 t⫺1 x1 ⫹ 11 ⫺ ␭2 t␮0 t

⫽ a ␭11 ⫺ ␭2 k xt⫺k ⫹ 11 ⫺ ␭2 t␮0 k⫽0

Each older observation has its weight decreased by the factor 11 ⫺ ␭2. The weight on the starting value ␮0 is selected so that the weights sum to one. Here zt is also sometimes called a geometric average.

JWCL232_c15_637-701.qxd

1/11/10

4:58 PM

Page 683

15-8 TIME-WEIGHTED CHARTS

683

16 14 12 10

Data

8

Figure 15-22 EWMAs with  0.8 and  0.2 show a compromise between a smooth curve and a response to a shift.

0.8 0.2

6 4 2 0

1

4

7

10 13 16 19 22 25 28 31 34 37 40 43 46 49

The value of determines the compromise between noise reduction and response to a mean change. For example, the series of weights when  0.8 are 0.8, 0.16, 0.032, 0.0064, 0.00128, . . . and when ␭  0.2 the weights are 0.2, 0.16, 0.128, 0.1024, 0.0819, . . . When  0.8 the weights decrease rapidly. Most of the weight is placed on the most recent observation, with modest contributions to the EWMA from older measurements. In this case, the EWMA does not average noise much, but it responds quickly to a mean change. However, when  0.2 the weights decrease much more slowly and the EWMA has substantial contributions from the more recent observations. In this case, the EWMA averages noise more, but it responds more slowly to a change in the mean. Fig. 15-22 displays a series of observations with a mean shift in the middle on the series. Notice that the EWMA with  0.2 smooths the data more, but that the EWMA with  0.8 adjusts the estimate to the mean shift more quickly. It appears that it is difficult to calculate an EWMA because at every time t a new weighted average of all previous data is required. However, there is an easy method to calculate an EWMA zt based on a simple recursive equation. Let z0  0. Then it can be shown that EWMA Update Equation

zt  xt  11  2zt1

(15-33)

Consequently, only a brief computation is needed at each time t. To develop a control chart from an EWMA, control limits are needed for Zt. The control limits are defined in a straightforward manner. They are placed at three standard deviations around the mean of the plotted statistic Zt. This follows the general approach for a control chart in Equation 15-1. An EWMA control chart may be applied to individual measurements as an extension to an X chart or to subgroup averages. Formulas here are developed for the more general case with an average from a subgroup of size n. For individual measurements n  1.

JWCL232_c15_637-701.qxd

684

1/11/10

4:58 PM

Page 684

CHAPTER 15 STATISTICAL QUALITY CONTROL

Because Zt is a linear function of the independent observations X1, X2, . . . , Xt (and 0), the results from Chapter 5 can be used to show that E1Zt 2  0

2 V1Zt 2  n 冤1  11  2 2t冥 2

and

where n is the subgroup size. Therefore an EWMA control chart uses estimates of 0 and  in the following formulas: EWMA Control Chart

LCL  0  3



31  11  2 2t 4 2 

1n A

CL  0 UCL  0  3

(15-34)



31  11  2 2t 4 1n A 2 

Note that the control limits are not of equal width about the centerline. The control limits are calculated from the variance of Zt and that changes with time. However, for large t the variance of Zt converges to

2 b lim V1Zt 2  n a 2 

tS so that the control limits tend to be parallel lines about the centerline as t increases. The parameters 0 and  are estimated by the same statistics used in X or X charts. That is, for subgroups ˆ0  X 

ˆ  R d2

and

ˆ  S c4 

or

and for n  1 ˆ0  X 

EXAMPLE 15-7

and

ˆ  MR 1.128

Chemical Process Concentration EWMA

Consider the concentration data shown in Table 15-3. Construct an EWMA control chart with  0.2 with n  1. It was deterˆ 0  99.1 and mined that x  99.1 and mr  2.59. Therefore,  ˆ  2.59 1.128  2.30. The control limits for z1 are LCL  99.1  312.302

0.2 冤1  11  0.22 2冥  98.19 B 2  0.2

LCL  99.1  312.302

0.2 冤1  11  0.22 2冥  100.01 B 2  0.2

The first few values of zt along with the corresponding control limits are

t

1

2

3

4

5

xt zt LCL UCL

102.0 99.68 97.72 100.48

94.8 98.70 97.33 100.87

98.3 98.62 97.12 101.08

98.4 98.58 97.00 101.20

102.0 99.26 96.93 101.27

The chart generated by Minitab is shown in Figure 15-23. Notice that the control limits widen as time increases but quickly stabilize. Each point is within its set of corresponding control limits so there are no signals from the chart.

JWCL232_c15_637-701.qxd

1/11/10

4:58 PM

Page 685

15-8 TIME-WEIGHTED CHARTS

685

EWMA Chart of X 102 UCL = 101.390

101

EWMA

100

Figure 15-23 EWMA control chart for the chemical process concentration data from Minitab.

= X = 99.095

99 98 97

LCL = 96.800 2

4

6

8

10 12 Sample

14

16

18

20

Similar to a CUSUM chart, the points plotted on an EWMA control chart are not independent. Therefore, run rules should not be applied to an EWMA control chart. Information in the history of the data that is considered by run rules is to a large extent incorporated into the EWMA that is calculated at each time t. The value of is usually chosen from the range 0.1 0.5. A common choice is

 0.2. Smaller values for provide more sensitivity for small shifts and larger values better tune the chart for larger shifts. This performance can be seen in the average run lengths in Table 15-10. These calculations are more difficult than those used for Shewhart charts, and details are omitted. Here,  0.1 and 0.5 are compared. The multiplier of the standard deviation, denoted L in the table, is adjusted so that the average run length (ARL) equals 500 for both choices for . That is, the control limits are placed at E1Zt 2 L2V1Zt 2 and L is chosen so the ARL without a mean shift is 500 in both cases. The EWMA ARLs in the table indicate that the smaller value for is preferred when the magnitude of the shift is small. Also, the EWMA performance is in general much better than results for a Shewhart control chart (in Table 15-6) and the results are comparable to a CUSUM control chart (in Table 15-9). However, these are average results. At the time of an increase in the process mean, zt might be negative and there would be some performance penalty to first increase zt to near zero and then further increase it to a signal above the UCL. Such a penalty provides an advantage to CUSUM control charts that is not accounted for in these ARL tables. A more refined analysis can be used to quantify this penalty, but the conclusion is that the EWMA penalty is moderate to small in most applications.

Table 15-10 Average Run Lengths for an EWMA Control Chart Shift in Mean (multiple of  X 2

 0.5 L  3.07

 0.1 L  2.81

0 0.25 0.5 0.75 1 1.5 2 3

500 255 88.8 35.9 17.5 6.53 3.63 1.93

500 106 31.3 15.9 10.3 6.09 4.36 2.87

JWCL232_c15_637-701.qxd

686

1/11/10

4:58 PM

Page 686

CHAPTER 15 STATISTICAL QUALITY CONTROL

EXERCISES FOR SECTION 15-8 15-61. The following data were considered in Quality Engineering [“Parabolic Control Limits for The Exponentially Weighted Moving Average Control Charts in Quality Engineering” (1992, Vol. 4(4))]. In a chemical plant, the data for one of the quality characteristics (viscosity) were obtained for each 12-hour batch at the completion of the batch. The results of 15 consecutive measurements are shown in the table below. Batch

Viscosity

Batch

Viscosity

1 2 3 4 5 6 7 8 9 10

13.3 14.5 15.3 15.3 14.3 14.8 15.2 14.9 14.6 14.1

11 12 13 14 15

14.3 16.1 13.1 15.5 12.6

(a) Set up a CUSUM control chart for this process. Assume the desired process target 14.1. Does the process appear to be in control? (b) Suppose that the next five observations are 14.6, 15.3, 15.7, 16.1, and 16.8. Apply the CUSUM in part (a) to these new observations. Is there any evidence that the process has shifted out of control? 15-62. The purity of a chemical product is measured every two hours. The results of 20 consecutive measurements are as follows: Sample 1 2 3 4 5 6 7 8 9 10

Purity

Sample

Purity

89.11 90.59 91.03 89.46 89.78 90.05 90.63 90.75 89.65 90.15

11 12 13 14 15 16 17 18 19 20

88.55 90.43 91.04 88.17 91.23 90.92 88.86 90.87 90.73 89.78

(a) Set up a CUSUM control chart for this process. Use   0.8 in setting up the procedure, and assume that the desired process target is 90. Does the process appear to be in control? (b) Suppose that the next five observations are 90.75, 90.00, 91.15, 90.95, and 90.86. Apply the CUSUM in part (a) to these new observations. Is there any evidence that the process has shifted out of control? 15-63. The diameter of holes is measured in consecutive order by an automatic sensor. The results of measuring 25 holes follow.

Sample

Diameter

Sample

Diameter

1 2 3 4 5 6 7 8 9 10 11 12 13

9.94 9.93 10.09 9.98 10.11 9.99 10.11 9.84 9.82 10.38 9.99 10.41 10.36

14 15 16 17 18 19 20 21 22 23 24 25

9.99 10.12 9.81 9.73 10.14 9.96 10.06 10.11 9.95 9.92 10.09 9.85

(a) Estimate the process standard deviation. (b) Set up a CUSUM control procedure, assuming that the target diameter is 10.0 millimeters. Does the process appear to be operating in a state of statistical control at the desired target level? 15-64. The concentration of a chemical product is measured by taking four samples from each batch of material. The average concentration of these measurements is shown for the last 20 batches in the following table: Batch

Concentration

Batch

Concentration

1 2 3 4 5 6 7 8 9 10

104.5 99.9 106.7 105.2 94.8 94.6 104.4 99.4 100.3 100.3

11 12 13 14 15 16 17 18 19 20

95.4 94.5 104.5 99.7 97.7 97.0 95.8 97.4 99.0 102.6

(a) Suppose that the process standard deviation is   8 and that the target value of concentration for this process is 100. Design a CUSUM scheme for the process. Does the process appear to be in control at the target? (b) How many batches would you expect to be produced with off-target concentration before it would be detected by the CUSUM control chart if the concentration shifted to 104? Use Table 15-9. 15-65. Consider a CUSUM with h  5 and k  1兾2. Samples are taken every two hours from the process. The target value for the process is 0  50 and   2. Use Table 15-9.

JWCL232_c15_637-701.qxd

1/11/10

4:58 PM

Page 687

687

15-8 TIME-WEIGHTED CHARTS

(a) If the sample size is n  1, how many samples would be required to detect a shift in the process mean to   51 on average? (b) If the sample size is increased to n  4, how does this affect the average run length to detect the shift to   51 that you determined in part (a)? 15-66. Consider the purity data in Exercise 15-62. Use   0.8 and assume that the desired process target is 90. (a) Construct an EWMA control chart with ␭  0.2. Does the process appear to be in control? (b) Construct an EWMA control chart with ␭  0.5. Compare your results to part (a). (c) Suppose that the next five observations are 90.75, 90.00, 91.15, 90.95, and 90.86. Apply the EWMAs in part (a) and (b) to these new observations. Is there any evidence that the process has shifted out of control? 15-67. Consider the diameter data in Exercise 15-63. Assume that the desired process target is 10.0 millimeters. (a) Estimate the process standard deviation. (b) Construct an EWMA control chart with  0.2. Does the process appear to be in control? (c) Construct an EWMA control chart with  0.5. Compare your results to part (a). 15-68. Consider the concentration data in Exercise 15-64. Use   8 and assume that the desired process target is 100. (a) Construct an EWMA control chart with ␭  0.2. Does the process appear to be in control? (b) Construct an EWMA control chart with ␭  0.5. Compare your results to part (a). (c) If the concentration shifted to 104, would you prefer the chart in part (a) or (b)? Explain. 15-69. Consider an EMWA control chart. The target value for the process is 0  50 and   2. Use Table 15-10. (a) If the sample size is n  1, would you prefer an EWMA chart with  0.1 and L  2.81 or  0.5 and L  3.07 to detect a shift in the process mean to   52 on average? Why? (b) If the sample size is increased to n  4, which chart in part (a) do you prefer? Why? (c) If an EWMA chart with ␭  0.1 and L  2.81 is used, what sample size is needed to detect a shift to   52 in approximately 3 samples on average? 15-70. A process has a target of 0  100 and a standard deviation of   4. Samples of size n  1 are taken every two hours. Use Table 15-9. (a) Suppose the process mean shifts to   102. How many hours of production will occur before the process shift is detected by a CUSUM with h  5 and k  1兾2? (b) It is important to detect the shift defined in part (a) more quickly. A proposal is made to reduce the sampling frequency to 0.5 hour. How will this affect the CUSUM control procedure? How much more quickly will the shift be detected? (c) Suppose that the 0.5 hour sampling interval in part (b) is adopted. How often will false alarms occur with this new

sampling interval? How often did they occur with the old interval of two hours? (d) A proposal is made to increase the sample size to n  4 and retain the two-hour sampling interval. How does this suggestion compare in terms of average detection time to the suggestion of decreasing the sampling interval to 0.5 hour? 15-71. Heart rate (in counts/minute) is measured every 30 minutes. The results of 20 consecutive measurements are as follows: Sample

Heart Rate

Sample

Heart Rate

1 2 3 4 5 6 7 8 9 10

68 71 67 69 71 70 69 67 70 70

11 12 13 14 15 16 17 18 19 20

79 79 78 78 78 79 79 82 82 81

Suppose that the standard deviation of the heart rate is  = 3 and the target value is 70. (a) Design a CUSUM scheme for the heart rate process. Does the process appear to be in control at the target? (b) How many samples on average would be required to detect a shift of the mean heart rate to 80? 15-72. The number of influenza patients (in thousands) visiting hospitals weekly are shown in the following table. Suppose that the standard deviation is  = 2 and the target value is 160.

Sample

Number of Patients

Sample

Number of Patients

1 2 3 4 5 6 7 8 9 10 11 12

162.27 157.47 157.065 160.45 157.993 162.27 160.652 159.09 157.442 160.78 159.138 161.08

13 14 15 16 17 18 19 20 21 22 23 24

159.989 159.09 162.699 163.89 164.247 162.70 164.859 163.65 165.99 163.22 164.338 164.83

(a) Design a CUSUM scheme for the process. Does the process appear to be in control at the target? (b) How many samples on average would be required to detect a shift of the mean to 165?

JWCL232_c15_637-701.qxd

688 15-9

1/11/10

4:58 PM

Page 688

CHAPTER 15 STATISTICAL QUALITY CONTROL

OTHER SPC PROBLEM-SOLVING TOOLS While the control chart is a very powerful tool for investigating the causes of variation in a process, it is most effective when used with other SPC problem-solving tools. In this section we illustrate some of these tools, using the printed circuit board defect data in Example 15-5. Figure 15-17 shows a U chart for the number of defects in samples of five printed circuit boards. The chart exhibits statistical control, but the number of defects must be reduced. The average number of defects per board is 8兾5  1.6, and this level of defects would require extensive rework. The first step in solving this problem is to construct a Pareto diagram of the individual defect types. The Pareto diagram, shown in Fig. 15-24, indicates that insufficient solder and solder balls are the most frequently occurring defects, accounting for (109兾160) 100  68% of the observed defects. Furthermore, the first five defect categories on the Pareto chart are all solder-related defects. This points to the flow solder process as a potential opportunity for improvement. To improve the surface mount process, a team consisting of the operator, the shop supervisor, the manufacturing engineer responsible for the process, and a quality engineer meets to study potential causes of solder defects. They conduct a brainstorming session and produce the cause-and-effect diagram shown in Fig. 15-25. The cause-and-effect diagram is widely used to display the various potential causes of defects in products and their interrelationships. They are useful in summarizing knowledge about the process. As a result of the brainstorming session, the team tentatively identifies the following variables as potentially influential in creating solder defects: 1. 2. 3.

Flux specific gravity Reflow temperature Squeegee speed

75 64

Number of defects

50

45

25

18 8

6

5

4

4

3

2

ne

lur

nt

e

red

po om

sol Un

gc on

de

on

fai

ne

nt Wr

nts

Co mp o

ed

co

mp o

ign sal

ing Mi ss

Mi

s s

ole wh

ort Sh

Blo

g les

ttin

ho Pin

we

co

s all

De

rb lde

So

mp

ne

er old ts en ici uff Ins

Figure 15-24 Pareto diagram for printed circuit board defects.

en

ts

0

1

JWCL232_c15_637-701.qxd

1/11/10

4:58 PM

Page 689

15-9 OTHER SPC PROBLEM-SOLVING TOOLS

Machine

Solder

Flux Temperature

Squeegee pressure

689

Amount

Stencil removal Specific gravity

Paste height

Squeegee speed Maintenance Squeegee angle

Density

Type Solder defects

Orientation

Alignment of pallet Solderability Pallet loading

Operator

Temperature

Contaminated solder

Components

Reflow

Figure 15-25 Cause-and-effect diagram for the printed circuit board flow solder process.

4. 5. 6. 7.

Squeegee angle Paste height Reflow temperature Board loading method

A statistically designed experiment could be used to investigate the effect of these seven variables on solder defects. In addition, the team constructed a defect concentration diagram for the product. A defect concentration diagram is just a sketch or drawing of the product, with the most frequently occurring defects shown on the part. This diagram is used to determine whether defects occur in the same location on the part. The defect concentration diagram for the printed circuit board is shown in Fig. 15-26. This diagram indicates that most of the insufficient solder defects are near the front edge of the board. Further investigation showed that one of the pallets used to carry the boards was bent, causing the front edge of the board to make poor contact with the squeegee. When the defective pallet was replaced, a designed experiment was used to investigate the seven variables discussed earlier. The results of this experiment indicated that several of these factors were influential and could be adjusted to reduce solder defects. After the results of the experiment were implemented, the percentage of solder joints requiring rework was reduced from 1% to under 100 parts per million (0.01%).

Front

Region of insufficient solder

Figure 15-26 Defect concentration diagram for a printed circuit board.

Back

JWCL232_c15_637-701.qxd

690

1/11/10

4:58 PM

Page 690

CHAPTER 15 STATISTICAL QUALITY CONTROL

15-10 IMPLEMENTING SPC The methods of statistical process control can provide significant payback to those companies that can successfully implement them. While SPC seems to be a collection of statistically based problem-solving tools, there is more to the successful use of SPC than simply learning and using these tools. Management involvement and commitment to the quality-improvement process is the most vital component of SPC’s potential success. Management is a role model, and others in the organization will look to management for guidance and as an example. A team approach is also important, for it is usually difficult for one person alone to introduce process improvements. Many of the “magnificent seven’’ problem-solving tools are helpful in building an improvement team, including cause-and-effect diagrams, Pareto charts, and defect concentration diagrams. The basic SPC problem-solving tools must become widely known and widely used throughout the organization. Continuous training in SPC and quality improvement is necessary to achieve this widespread knowledge of the tools. The objective of an SPC-based quality-improvement program is continuous improvement on a weekly, quarterly, and annual basis. SPC is not a one-time program to be applied when the business is in trouble and later abandoned. Quality improvement must become part of the culture of the organization. The control chart is an important tool for process improvement. Processes do not naturally operate in an in-control state, and the use of control charts is an important step that must be taken early in an SPC program to eliminate assignable causes, reduce process variability, and stabilize process performance. To improve quality and productivity, we must begin to manage with facts and data, and not just rely on judgment. Control charts are an important part of this change in management approach. In implementing a company-wide SPC program, we have found that the following elements are usually present in all successful efforts: 1. 2. 3. 4. 5.

Management leadership A team approach Education of employees at all levels Emphasis on continuous improvement A mechanism for recognizing success

We cannot overemphasize the importance of management leadership and the team approach. Successful quality improvement is a “top-down” management-driven activity. It is also important to measure progress and success and to spread knowledge of this success throughout the organization. When successful improvements are communicated throughout the company, this can provide motivation and incentive to improve other processes and to make continuous improvement a normal part of the way of doing business. The philosophy of W. Edwards Deming provides an important framework for implementing quality and productivity improvement. Deming’s philosophy is summarized in his 14 points for management. The adherence to these management principles has been an important factor in Japan’s industrial success and continues to be the catalyst in that nation’s quality- and productivity-improvement efforts. This philosophy has also now spread rapidly in the West. Deming’s 14 points are as follows. 1. Create a constancy of purpose focused on the improvement of products and services. Constantly try to improve product design and performance. Investment in research, development, and innovation will have a long-term payback to the organization.

JWCL232_c15_637-701.qxd

1/11/10

4:58 PM

Page 691

15-10 IMPLEMENTING SPC

691

2. Adopt a new philosophy of rejecting poor workmanship, defective products, or bad service. It costs as much to produce a defective unit as it does to produce a good one (and sometimes more). The cost of dealing with scrap, rework, and other losses created by defectives is an enormous drain on company resources. 3. Do not rely on mass inspection to “control” quality. All inspection can do is sort out defectives, and at this point it is too late because we have already paid to produce these defectives. Inspection occurs too late in the process, it is expensive, and it is often ineffective. Quality results from the prevention of defectives through process improvement, not inspection. 4. Do not award business to suppliers on the basis of price alone, but also consider quality. Price is a meaningful measure of a supplier’s product only if it is considered in relation to a measure of quality. In other words, the total cost of the item must be considered, not just the purchase price. When quality is considered, the lowest bidder is frequently not the low-cost supplier. Preference should be given to suppliers who use modern methods of quality improvement in their business and who can demonstrate process control and capability. 5. Focus on continuous improvement. Constantly try to improve the production and service system. Involve the workforce in these activities and make use of statistical methods, particularly the SPC problem-solving tools discussed in the previous section. 6. Practice modern training methods and invest in training for all employees. Everyone should be trained in the technical aspects of their job, as well as in modern quality- and productivity-improvement methods. The training should encourage all employees to practice these methods every day. 7. Practice modern supervision methods. Supervision should not consist merely of passive surveillance of workers, but should be focused on helping the employees improve the system in which they work. The first goal of supervision should be to improve the work system and the product. 8. Drive out fear. Many workers are afraid to ask questions, report problems, or point out conditions that are barriers to quality and effective production. In many organizations the economic loss associated with fear is large; only management can eliminate fear. 9. Break down the barriers between functional areas of the business. Teamwork among different organizational units is essential for effective quality and productivity improvement to take place. 10. Eliminate targets, slogans, and numerical goals for the workforce. A target such as “zero defects” is useless without a plan as to how to achieve this objective. In fact, these slogans and “programs” are usually counterproductive. Work to improve the system and provide information on that. 11. Eliminate numerical quotas and work standards. These standards have historically been set without regard to quality. Work standards are often symptoms of management’s inability to understand the work process and to provide an effective management system focused on improving this process. 12. Remove the barriers that discourage employees from doing their jobs. Management must listen to employee suggestions, comments, and complaints. The

JWCL232_c15_637-701.qxd

692

1/12/10

4:33 PM

Page 692

CHAPTER 15 STATISTICAL QUALITY CONTROL

person who is doing the job is the one who knows the most about it, and usually has valuable ideas about how to make the process work more effectively. The workforce is an important participant in the business, and not just an opponent in collective bargaining. 13.

Institute an ongoing program of training and education for all employees. Education in simple, powerful statistical techniques should be mandatory for all employees. Use of the basic SPC problem-solving tools, particularly the control chart, should become widespread in the business. As these charts become widespread, and as employees understand their uses, they will be more likely to look for the causes of poor quality and to identify process improvements. Education is a way of making everyone partners in the quality-improvement process.

14. Create a structure in top management that will vigorously advocate the first 13 points. As we read Deming’s 14 points, we notice two things. First, there is a strong emphasis on change. Second, the role of management in guiding this change process is of dominating importance. But what should be changed, and how should this change process be started? For example, if we want to improve the yield of a semiconductor manufacturing process, what should we do? It is in this area that statistical methods most frequently come into play. To improve the semiconductor process, we must determine which controllable factors in the process influence the number of defective units produced. To answer this question, we must collect data on the process and see how the system reacts to changes in the process variables. Statistical methods, including the SPC and experimental design techniques in this book, can contribute to this knowledge. SUPPLEMENTAL EXERCISES 15-73. The diameter of fuse pins used in an aircraft engine application is an important quality characteristic. Twenty-five samples of three pins each are shown as follows: Sample Number 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

Diameter 64.030 63.995 63.988 64.002 63.992 64.009 63.995 63.985 64.008 63.998 63.994 64.004 63.983 64.006 64.012 64.000 63.994

64.002 63.992 64.024 63.996 64.007 63.994 64.006 64.003 63.995 74.000 63.998 64.000 64.002 63.967 64.014 63.984 64.012

64.019 64.001 64.021 63.993 64.015 63.997 63.994 63.993 64.009 63.990 63.994 64.007 63.998 63.994 63.998 64.005 63.986

18 19 20 21 22 23 24 25

64.006 63.984 64.000 63.988 64.004 64.010 64.015 63.982

64.010 64.002 64.010 64.001 63.999 63.989 64.008 63.984

64.018 64.003 64.013 64.009 63.990 63.990 63.993 63.995

(a) Set up X and R charts for this process. If necessary, revise limits so that no observations are out-of-control. (b) Estimate the process mean and standard deviation. (c) Suppose the process specifications are at 64 ⫾ 0.02. Calculate an estimate of PCR. Does the process meet a minimum capability level of PCR ⱖ 1.33? (d) Calculate an estimate of PCRk. Use this ratio to draw conclusions about process capability. (e) To make this process a six-sigma process, the variance ␴2 would have to be decreased such that PCRk ⫽ 2.0. What should this new variance value be? (f ) Suppose the mean shifts to 64.01. What is the probability that this shift will be detected on the next sample? What is the ARL after the shift?

JWCL232_c15_637-701.qxd

1/11/10

4:58 PM

Page 693

693

15-10 IMPLEMENTING SPC

15-74. Rework Exercise 15-73 with X and S charts. 15-75. Plastic bottles for liquid laundry detergent are formed by blow molding. Twenty samples of n  100 bottles are inspected in time order of production, and the fraction defective in each sample is reported. The data are as follows:

Sample

Fraction Defective

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0.12 0.15 0.18 0.10 0.12 0.11 0.05 0.09 0.13 0.13 0.10 0.07 0.12 0.08 0.09 0.15 0.10 0.06 0.12 0.13

13 14 15 16 17 18 19

2 8 0 2 4 3 5

20 21 22 23 24 25

0 2 1 9 3 2

(a) Using all the data, find trial control limits for a U chart for the process. (b) Use the trial control limits from part (a) to identify out-ofcontrol points. If necessary, revise your control limits. (c) Suppose that instead of samples of five cases, the sample size was 10. Repeat parts (a) and (b). Explain how this change alters your answers to parts (a) and (b). 15-77. An article in Quality Engineering [“Is the Process Capable? Tables and Graphs in Assessing Cpm” (1992, Vol. 4(4))]. Considered manufacturing data. Specifications for the outer diameter of the hubs were 60.3265 0.001 mm. A random sample with size n  20 was taken and the data are shown in the following table.

(a) Set up a P chart for this process. Is the process in statistical control? (b) Suppose that instead of n  100, n  200. Use the data given to set up a P chart for this process. Revise the control limits if necessary. (c) Compare your control limits for the P charts in parts (a) and (b). Explain why they differ. Also, explain why your assessment about statistical control differs for the two sizes of n. 15-76. Cover cases for a personal computer are manufactured by injection molding. Samples of five cases are taken from the process periodically, and the number of defects is noted. Twenty-five samples follow:

Sample

No. of Defects

Sample

No. of Defects

1 2 3 4 5 6

3 2 0 1 4 3

7 8 9 10 11 12

2 4 1 0 2 3

Sample

x

Sample

x

1 2 3 4 5 6 7 8 9 10

60.3262 60.3262 60.3262 60.3266 60.3263 60.3260 60.3262 60.3267 60.3263 60.3269

11 12 13 14 15 16 17 18 19 20

60.3262 60.3262 60.3269 60.3261 60.3265 60.3266 60.3265 60.3268 60.3262 60.3266

(a) Construct a control chart for individual measurements. Revise the control limits if necessary. (b) Compare your chart in part (a) to one that uses only the last (least significant) digit of each diameter as the measurement. Explain your conclusion. (c) Estimate  and  from the moving range of the revised chart and use this value to estimate PCR and PCRk and interpret these ratios. 15-78. The following data from the U.S. Department of Energy Web site (http://www.eia.doe.gov) reported the total U.S. renewable energy consumption by year (trillion BTU) from 1973 to 2004.

JWCL232_c15_637-701.qxd

694

1/11/10

4:58 PM

Page 694

CHAPTER 15 STATISTICAL QUALITY CONTROL

Year

Total Renewable Energy Consumption (Trillion BTU)

Year

Total Renewable Energy Consumption (Trillion Btu)

1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988

4433.121 4769.395 4723.494 4767.792 4249.002 5038.938 5166.379 5494.42 5470.574 5985.352 6487.898 6430.646 6032.728 6131.542 5686.932 5488.649

1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004

6294.209 6132.572 6158.087 5907.147 6155.959 6064.779 6669.261 7136.799 7075.152 6560.632 6598.63 6158.232 5328.335 5835.339 6081.722 6116.287

(a) Using all the data, find trial control limits for a control chart for individual measurements, construct the chart, and plot the data. (b) Do the data appear to be generated from an in-control process? Comment on any patterns on the chart. 15-79. The following dataset was considered in Quality Engineering [“Analytic Examination of Variance Components” (1994–1995, Vol. 7(2))]. A quality characteristic for cement mortar briquettes was monitored. Samples of size n  6 were taken from the process, and 25 samples from the process are shown in the following table. (a) Using all the data, find trial control limits for X and S charts. Is the process in control? Batch

x

s

1 2 3 4 5 6 7 8 9 10 11 12

572.00 583.83 720.50 368.67 374.00 580.33 388.33 559.33 562.00 729.00 469.00 566.67

73.25 79.30 86.44 98.62 92.36 93.50 110.23 74.79 76.53 49.80 40.52 113.82

13 14 15 16 17 18 19 20 21 22 23 24 25

578.33 485.67 746.33 436.33 556.83 390.33 562.33 675.00 416.50 568.33 762.67 786.17 530.67

58.03 103.33 107.88 98.69 99.25 117.35 75.69 90.10 89.27 61.36 105.94 65.05 99.42

(b) Suppose that the specifications are at 580 250. What statements can you make about process capability? Compute estimates of the appropriate process capability ratios. (c) To make this process a “6-sigma process,” the variance 2 would have to be decreased such that PCRk  2.0. What should this new variance value be? (d) Suppose the mean shifts to 600. What is the probability that this shift will be detected on the next sample? What is the ARL after the shift? 15-80. Suppose that an X control chart with 2-sigma limits is used to control a process. Find the probability that a false out-of-control signal will be produced on the next sample. Compare this with the corresponding probability for the chart with 3-sigma limits and discuss. Comment on when you would prefer to use 2-sigma limits instead of 3-sigma limits. 15-81. Consider the diameter data in Exercise 15-73. (a) Construct an EWMA control chart with  0.2 and L  3. Comment on process control. (b) Construct an EWMA control chart with  0.5 and L  3 and compare your conclusion to part (a). 15-82. Consider the renewable energy data in Exercise 15-78. (a) Construct an EWMA control chart with ␭  0.2 and L  3. Do the data appear to be generated from an incontrol process? (b) Construct an EWMA control chart with  0.5 and L  3 and compare your conclusion to part (a). 15-83. Consider the hub data in Exercise 15-77. (a) Construct an EWMA control chart with  0.2 and L  3. Comment on process control. (b) Construct an EWMA control chart with  0.5 and L  3 and compare your conclusion to part (a). 15-84. Consider the data in Exercise 15-18. Set up a CUSUM scheme for this process assuming that   16 is the process target. Explain how you determined your estimate of  and the CUSUM parameters K and H.

JWCL232_c15_637-701.qxd

1/11/10

4:58 PM

Page 695

15-10 IMPLEMENTING SPC

15-85. Consider the hardness measurement data in Exercise 15-17. Set up a CUSUM scheme for this process using   50 and   2, so that K  1 and H  10. Is the process in control? 15-86. Reconsider the viscosity data in Exercise 15-20. Construct a CUSUM control chart for this process using 0  500 as the process target. Explain how you determined your estimate of  and the CUSUM parameters H and K. 15-87. The following data were considered in Quality Progress [“Digidot Plots for Process Surveillance” (1990, May, pp. 66–68)]. Measurements of center thickness (in mils) from 25 contact lenses sampled from the production process at regular intervals are shown in the following table. Sample

x

Sample

x

1 2 3 4 5 6 7 8 9 10 11 12 13

0.3978 0.4019 0.4031 0.4044 0.3984 0.3972 0.3981 0.3947 0.4012 0.4043 0.4051 0.4016 0.3994

14 15 16 17 18 19 20 21 22 23 24 25

0.3999 0.4062 0.4048 0.4071 0.4015 0.3991 0.4021 0.4009 0.3988 0.3994 0.4016 0.4010

(a) Construct a CUSUM scheme for this process with the target 0  0.4. Explain how you determined your estimate of  and the CUSUM parameters H and K. Is the process in control? (b) Construct an EWMA control chart with  0.5 and L  3 and compare your conclusions to part (a). 15-88. Suppose that a process is in control and an X chart is used with a sample size of 4 to monitor the process. Suddenly there is a mean shift of 1.5. (a) If 3-sigma control limits are in use on the X chart, what is the probability that this shift will remain undetected for three consecutive samples? (b) If 2-sigma control limits are in use on the X chart, what is the probability that this shift will remain undetected for three consecutive samples? (c) Compare your answers to parts (a) and (b) and explain why they differ. Also, which limits you would recommend using and why? 15-89. Consider the control chart for individuals with 3-sigma limits. (a) Suppose that a shift in the process mean of magnitude  occurs. Verify that the ARL for detecting the shift is ARL  43.9.

695

(b) Find the ARL for detecting a shift of magnitude 2 in the process mean. (c) Find the ARL for detecting a shift of magnitude 3 in the process mean. (d) Compare your answers to parts (a), (b), and (c) and explain why the ARL for detection is decreasing as the magnitude of the shift increases. 15-90. Consider a control chart for individuals, applied to a continuous 24-hour chemical process with observations taken every hour. (a) If the chart has 3-sigma limits, verify that the in-control ARL is ARL  370. How many false alarms would occur each 30-day month, on the average, with this chart? (b) Suppose that the chart has 2-sigma limits. Does this reduce the ARL for detecting a shift in the mean of magnitude ? (Recall that the ARL for detecting this shift with 3-sigma limits is 43.9.) (c) Find the in-control ARL if 2-sigma limits are used on the chart. How many false alarms would occur each month with this chart? Is this in-control ARL performance satisfactory? Explain your answer. 15-91. The depth of a keyway is an important part quality characteristic. Samples of size n  5 are taken every four hours from the process and 20 samples are summarized in the following table. (a) Using all the data, find trial control limits for X and R charts. Is the process in control? (b) Use the trial control limits from part (a) to identify out-ofcontrol points. If necessary, revise your control limits. Then, estimate the process standard deviation. (c) Suppose that the specifications are at 140 2. Using the results from part (b), what statements can you make about process capability? Compute estimates of the appropriate process capability ratios. (d) To make this process a “6-sigma process,” the variance 2 would have to be decreased such that PCRk  2.0. What should this new variance value be? Sample

X

r

1 2 3 4 5 6 7 8 9 10 11 12

139.7 139.8 140.0 140.1 139.8 139.9 139.7 140.2 139.3 140.7 138.4 138.5

1.1 1.4 1.3 1.6 0.9 1.0 1.4 1.2 1.1 1.0 0.8 0.9 continued

JWCL232_c15_637-701.qxd

696

1/11/10

4:58 PM

Page 696

CHAPTER 15 STATISTICAL QUALITY CONTROL

Sample

X

r

13 14 15 16 17 18 19 20

137.9 138.5 140.8 140.5 139.4 139.9 137.5 139.2

1.2 1.1 1.0 1.3 1.4 1.0 1.5 1.3

(e) Suppose the mean shifts to 139.7. What is the probability that this shift will be detected on the next sample? What is the ARL after the shift? 15-92. Consider a control chart for individuals with 3-sigma limits. What is the probability that there will not be a signal in three samples? In six samples? In 10 samples? 15-93. Suppose a process has a PCR  2, but the mean is exactly three standard deviations above the upper specification limit. What is the probability of making a product outside the specification limits? 15-94. A process is controlled by a P chart using samples of size 100. The center line on the chart is 0.05. (a) What is the probability that the control chart detects a shift to 0.08 on the first sample following the shift? (b) What is the probability that the control chart does not detect a shift to 0.08 on the first sample following the shift, but does detect it on the second sample? (c) Suppose that instead of a shift in the mean to 0.08, the mean shifts to 0.10. Repeat parts (a) and (b). (d) Compare your answers for a shift to 0.08 and for a shift to 0.10. Explain why they differ. Also, explain why a shift to 0.10 is easier to detect. 15-95. Suppose the average number of defects in a unit is known to be 8. If the mean number of defects in a unit shifts to 16, what is the probability that it will be detected by a U chart on the first sample following the shift (a) if the sample size is n  4? (b) if the sample size is n  10? Use a normal approximation for U.

15-96. Suppose the average number of defects in a unit is known to be 10. If the mean number of defects in a unit shifts to 14, what is the probability that it will be detected by a U chart on the first sample following the shift (a) if the sample size is n  1? (b) if the sample size is n  4? Use a normal approximation for U. 15-97. An EWMA chart with  0.5 and L  3.07 is to be used to monitor a process. Suppose that the process mean is 0  10 and   2. (a) Assume that n  1. What is the ARL without any shift in the process mean? What is the ARL to detect a shift to   12. (b) Assume that n  4. Repeat part (a) and comment on your conclusions. 15-98. The following table provides the costs for gasoline by month in the U.S. over recent years and the percentage of the cost due to refining, distribution and marketing, taxes, and crude oil. The table is from the U.S. Department of Energy Web site (http://tonto.eia.doe.gov/oog/info/gdu/ gaspump.html). There is some concern that the refining or distribution and marketing percentages of the retail price have shown patterns over time. (a) Construct separate control charts for the refining percentage of the retail price and the distribution and marketing percentage of the retail price. Use control charts for individual measurements. Comment on any signs of assignable causes on these charts. (b) Construct a control chart for the crude oil percentage of the retail price. Use a control chart for individual measurements. Comment on any signs of assignable causes on this chart. (c) Another way to study the data is to calculate refining, distribution and marketing, and tax as costs directly. The costs of these categories might not depend strongly on the crude oil cost. Use the percentages provided in the table to calculate the cost each month associated with refining and distribution and marketing. Construct separate control charts for the refining and the distribution and marketing costs each month. Use control charts for individual measurements. Comment on any signs of assignable causes on these charts and comment on any differences between these charts and the ones constructed in part (a).

What We Pay For in a Gallon of Regular Gasoline Mo/Year

Retail Price (Dollars per gallon)

Refining (percentage)

Distribution & Marketing (percentage)

Taxes (percentage)

Crude Oil (percentage)

Jan-00 Feb-00 Mar-00 Apr-00 May-00

1.289 1.377 1.517 1.465 1.485

7.8 17.9 15.4 10.1 20.2

13.0 7.5 12.8 20.2 9.2

32.1 30.1 27.3 28.3 27.9

47.1 44.6 44.6 41.4 42.7

JWCL232_c15_637-701.qxd

1/11/10

4:58 PM

Page 697

697

15-10 IMPLEMENTING SPC

What We Pay For in a Gallon of Regular Gasoline Mo/Year

Retail Price (Dollars per gallon)

Refining (percentage)

Distribution & Marketing (percentage)

Taxes (percentage)

Crude Oil (percentage)

Jun-00 Jul-00 Aug-00 Sep-00 Oct-00 Nov-00 Dec-00 Jan-01 Feb-01 Mar-01 Apr-01 May-01 Jun-01 Jul-01 Aug-01 Sep-01 Oct-01 Nov-01 Dec-01 Jan-02 Feb-02 Mar-02 Apr-02 May-02 Jun-02 Jul-02 Aug-02 Sep-02 Oct-02 Nov-02 Dec-02 Jan-03 Feb-03 Mar-03 Apr-03 May-03 Jun-03 Jul-03 Aug-03 Sept-03 Oct-03 Nov-03 Dec-03 Jan-04 Feb-04 Mar-04

1.633 1.551 1.465 1.550 1.532 1.517 1.443 1.447 1.450 1.409 1.552 1.702 1.616 1.421 1.421 1.522 1.315 1.171 1.086 1.107 1.114 1.249 1.397 1.392 1.382 1.397 1.396 1.400 1.445 1.419 1.386 1.458 1.613 1.693 1.589 1.497 1.493 1.513 1.620 1.679 1.564 1.512 1.479 1.572 1.648 1.736

22.2 13.2 15.8 15.4 13.7 10.4 8.0 17.8 17.3 18.8 31.6 26.4 13.2 10.0 20.0 18.0 10.0 10.0 11.7 13.0 12.1 19.4 15.5 11.9 15.0 15.0 11.4 10.8 13.9 11.1 11.7 11.5 15.0 14.8 13.2 15.3 15.1 15.3 22.5 13.9 14.9 11.7 11.5 15.9 19.1 19.0

8.8 15.8 7.5 9.0 10.1 11.8 17.9 10.4 11.0 9.7 4.6 14.0 24.1 20.0 9.0 17.0 20.8 18.0 12.7 11.8 11.2 6.1 13.0 14.2 13.0 12.6 13.4 12.6 11.7 18.0 12.3 10.3 9.5 14.8 19.8 16.3 12.3 11.9 8.2 22.7 16.1 15.3 12.6 9.9 9.2 11.3

25.8 27.2 28.8 27.2 27.5 27.8 29.2 29.2 29.1 30.0 27.1 24.7 26.0 30.0 30.0 28.0 31.9 36.0 38.7 37.9 37.7 33.6 30.1 30.2 30.4 30.1 30.0 30.0 29.1 29.6 30.3 28.8 26.0 24.8 26.4 28.1 28.1 27.8 25.9 25.0 26.9 27.8 28.4 26.7 25.5 24.2

43.1 43.8 47.8 48.3 48.6 50.0 44.8 42.7 42.6 41.5 36.7 35.0 36.7 40.0 41.0 37.0 37.2 36.0 36.9 37.2 39.1 40.9 41.4 43.7 41.6 42.3 45.0 46.7 45.3 41.3 45.7 49.4 49.5 45.5 40.5 40.4 44.5 44.9 43.3 38.3 42.2 45.2 47.5 47.5 46.2 45.5 continued

JWCL232_c15_637-701.qxd

698

1/11/10

4:58 PM

Page 698

CHAPTER 15 STATISTICAL QUALITY CONTROL

What We Pay For in a Gallon of Regular Gasoline Mo/Year Apr-04 May-04 Jun-04 Jul-04 Aug-04 Sep-04 Oct-04 Nov-04 Dec-04 Jan-05 Feb-05 Mar-05 Apr-05 May-05 Jun-05 Jul-05 Aug-05 Sep-05 Oct-05 Nov-05 Dec-05 Jan-06 Feb-06 Mar-06 Apr-06 May-06 Jun-06 Jul-06 Aug-06 Sep-06 Oct-06 Nov-06 Dec-06 Jan-07 Feb-07 Mar-07 Apr-07 May-07 Jun-07 Jul-07 Aug-07 Sep-07 Oct-07 Nov-07 Dec-07 Jan-08

Retail Price (Dollars per gallon) 1.798 1.983 1.969 1.911 1.878 1.870 2.000 1.979 1.841 1.831 1.910 2.079 2.243 2.161 2.156 2.290 2.486 2.903 2.717 2.257 2.185 2.316 2.280 2.425 2.742 2.907 2.885 2.981 2.952 2.555 2.245 2.229 2.313 2.240 2.278 2.563 2.845 3.146 3.056 2.965 2.786 2.803 2.803 3.080 3.018 3.043

Refining Distribution & Marketing MIND-EXPANDING EXERCISES (percentage) 22.0 30.6 21.3 20.9 13.9 14.8 13.0 10.7 8.9 17.7 16.1 19.3 20.9 17.9 18.5 17.9 24.3 27.3 15.1 8.3 13.5 13.4 9.8 21.7 25.8 21.9 22.0 26.3 15.2 6.3 10.9 14.6 12.9 10.6 18.0 23.6 28.1 27.9 22.7 18.4 13.5 12.8 10.1 10.0 8.1 7.8

(percentage) 9.9 7.8 16.7 11.3 12.2 9.1 9.3 14.6 18.1 7.3 9.3 6.2 9.6 12.8 6.9 8.0 2.1 7.5 17.8 13.1 7.9 6.6 11.4 4.5 3.1 8.8 7.9 6.3 13.5 18.8 10.6 7.5 9.4 15.2 5.8 8.5 7.6 13.3 13.7 11.4 11.8 8.6 8.1 8.7 10.5 11.1

Taxes (percentage) 23.4 21.2 21.3 21.9 22.4 22.5 21.0 21.2 23.9 24.0 23.0 21.2 19.6 20.4 20.4 19.2 17.7 15.2 16.2 19.5 20.1 19.8 20.1 18.9 16.7 15.8 15.9 15.4 15.9 18.3 20.8 20.4 19.7 20.3 20.0 15.5 14.0 12.7 13.0 13.4 14.3 14.2 14.2 13.0 13.2 13.1

Crude Oil (percentage) 44.6 40.4 40.7 45.8 51.5 53.6 56.7 53.6 49.1 50.9 51.6 53.4 49.8 49.0 54.2 54.9 55.9 50.0 50.9 57.1 58.4 60.1 58.6 54.8 54.2 53.4 54.1 52.0 55.4 56.7 57.7 57.5 58.0 53.9 56.3 52.3 50.3 46.1 50.5 56.8 60.4 64.3 67.6 68.3 68.1 67.9

JWCL232_c15_637-701.qxd

1/11/10

4:58 PM

Page 699

699

15-10 IMPLEMENTING SPC

What We Pay For in a Gallon of Regular Gasoline Mo/Year

Retail Price (Dollars per gallon)

Refining (percentage)

Distribution & Marketing (percentage)

Taxes (percentage)

Crude Oil (percentage)

Feb-08 Mar-08 Apr-08 May-08 Jun-08 Jul-08 Aug-08 Sep-08 Oct-08 Nov-08 Dec-08 Jan-09 Feb-09 Mar-09 Apr-09 May-09 Jun-09 Jul-09 Aug-09 Sep-09

3.028 3.244 3.458 3.766 4.054 4.062 3.779 3.703 3.051 2.147 1.687 1.788 1.923 1.959 2.049 2.266 2.631 2.527 2.616 2.554

9.9 8.0 10.0 10.0 8.5 3.2 6.1 14.2 3.3 3.7 0.7 13.4 16.8 12.0 12.0 17.6 13.7 10.1 11.0 6.7

7.2 7.9 5.8 4.7 6.8 11.2 10.2 8.2 25.0 24.7 19.5 10.7 14.9 12.3 12.1 3.9 10.3 14.0 9.7 13.5

13.2 12.3 11.5 10.6 9.8 9.8 10.6 10.8 13.1 18.6 23.6 22.3 20.7 20.4 19.5 19.5 15.1 15.8 15.4 15.7

69.7 71.8 72.7 74.7 74.8 75.8 73.1 66.8 58.6 60.4 56.2 53.6 47.6 55.3 56.4 66.8 60.9 60.1 63.9 64.0

15-99. The following table shows the number of e-mails a student received at each hour from 8:00 A.M. to 6:00 P.M. The samples are collected for five days from Monday to Friday. Hour

M

T

W

Th

F

1 2 3 4 5 6 7 8 9 10

2 2 2 4 1 1 3 2 1 2

2 4 2 4 1 3 2 3 3 3

2 0 2 3 2 2 1 2 3 2

3 1 1 3 2 2 1 3 2 3

1 2 2 2 1 1 0 1 0 0

(a) Use the rational subgrouping principle to comment on why an X chart that plots one point each hour with a subgroup of size 5 is not appropriate. (b) Construct an appropriate attribute control chart. Use all the data to find trial control limits, construct the chart, and plot the data.

(c) Use the trial control limits from part (b) to identify out-ofcontrol points. If necessary, revise your control limits, assuming that any samples that plot outside the control limits can be eliminated. 15-100. The following are the number of defects observed on 15 samples of transmission units in an automotive manufacturing company. Each lot contains five transmission units.

Sample

No. of Defects

1 2 3 4 5 6 7 8 9 10

8 10 24 6 5 21 10 7 9 15

Sample

No. of Defects

11 12 13 14 15

6 10 11 17 9

(a) Using all the data, compute trial control limits for a U control chart, construct the chart, and plot the data.

JWCL232_c15_637-701.qxd

700

1/11/10

4:58 PM

Page 700

CHAPTER 15 STATISTICAL QUALITY CONTROL

(b) Determine whether the process is in statistical control. If not, assume assignable causes can be found and out-ofcontrol points eliminated. Revise the control limits. 15-101. Consider an X control chart with UCL = 32.802, UCL = 24.642, and n = 5. Suppose the mean shifts to 30. (a) What is the probability that this shift will be detected on the next sample? (b) What is the ARL to detect the shift? 15-102. The number of visits (in millions) on a Web site is recorded every day. The following table shows the samples for 25 consecutive days.

Sample

Number of Visits

Sample

Number of Visits

1 2 3

10.12 9.92 9.76

4 5 6

9.35 9.60 8.60

7 8 9 10 11 12 13 14 15 16 17 18

10.46 10.58 9.95 9.50 11.26 10.02 10.95 8.99 9.50 9.66 10.42 11.30

19 20 21 22 23 24 25 26 27 28 29 30

12.53 10.76 11.92 13.24 10.64 11.31 11.26 11.79 10.53 11.82 11.47 11.76

(a) Estimate the process standard estimation. (b) Set up a CUSUM control chart for this process, assuming the target is 10. Does the process appear to be in control?

MIND-EXPANDING EXERCISES 15-103. Suppose a process is in control, and 3-sigma control limits are in use on the X chart. Let the mean shift by 1.5. What is the probability that this shift will remain undetected for three consecutive samples? What would its probability be if 2-sigma control limits were used? The sample size is 4. 15-104. Consider an X control chart with k-sigma control limits. Develop a general expression for the probability that a point will plot outside the control limits when the process mean has shifted by  units from the center line. 15-105. Suppose that an X chart is used to control a normally distributed process and that samples of size n are taken every n hours and plotted on the chart, which has k-sigma limits. (a) Find a general expression for the expected number of samples and time that will be taken until a false action signal is generated. (b) Suppose that the process mean shifts to an out-ofcontrol state, say 1  0  . Find an expression for the expected number of samples that will be taken until a false action is generated. (c) Evaluate the in-control ARL for k  3. How, does this change if k  2? What do you think about the use of 2-sigma limits in practice? (d) Evaluate the out-of-control ARL for a shift of 1 sigma, given that n  5.

15-106. Suppose a P chart with center line at p with k-sigma control limits is used to control a process. There is a critical fraction defective pc that must be detected with probability 0.50 on the first sample following the shift to this state. Derive a general formula for the sample size that should be used on this chart. 15-107. Suppose that a P chart with center line at p and k-sigma control limits is used to control a process. What is the smallest sample size that can be used on this control chart to ensure that the lower control limit is positive? 15-108. A process is controlled by a P chart using samples of size 100. The center line on the chart is 0.05. What is the probability that the control chart detects a shift to 0.08 on the first sample following the shift? What is the probability that the shift is detected by at least the third sample following the shift? 15-109. Consider a process where specifications on a quality characteristic are 100 15. We know that the standard deviation of this normally distributed quality characteristic is 5. Where should we center the process to minimize the fraction defective produced? Now suppose the mean shifts to 105 and we are using a sample size of 4 on an X chart. What is the probability that such a shift will be detected on the first sample following the shift? What is the average number of samples until an out-of-control point occurs? Compare this result to the

JWCL232_c15_637-701.qxd

1/11/10

4:58 PM

Page 701

15-10 IMPLEMENTING SPC

701

MIND-EXPANDING EXERCISES average number of observations until a defective occurs (assuming normality). 15-110. NP Control Chart. An alternative to the control chart for fraction defective is a control chart based on the number of defectives, or the NP control chart. The chart has centerline at n p, and the control limits are

(a) Apply this chart to the data in Example 15-5. (b) Will this chart always provide results equivalent to the U chart? 15-112. Standardized Control Chart. Consider the P chart with the usual 3-sigma control limits. Suppose that we define a new variable:

UCL  np  3 2np11  p2

Zi 

LCL  np  3 2np11  p2 and the number of defectives for each sample is plotted on the chart. (a) Verify that the control limits given above are correct. (b) Apply this control chart to the data in Example 15-4. (c) Will this chart always provide results that are equivalent to the usual P chart? 15-111. C Control Chart. An alternative to the U chart is a chart based on the number of defects. The chart has center line at nu, and the control limits are UCL  nu  3 2nu LCL  nu  3 2nu

Pˆi  P

P 11  P2 n C

as the quantity to plot on a control chart. It is proposed that this new chart will have a center line at 0 with the upper and lower control limits at 3. Verify that this standardized control chart will be equivalent to the original P chart. 15-113. Unequal Sample Sizes. One application of the standardized control chart introduced in Exercise 15-112 is to allow unequal sample sizes on the control chart. Provide details concerning how this procedure would be implemented and illustrate using the following data:

Sample, i

1

2

3

4

5

6

7

8

9

10

ni pi

20 0.2

25 0.16

20 0.25

25 0.08

50 0.3

30 0.1

25 0.12

25 0.16

25 0.12

20 0.15

IMPORTANT TERMS AND CONCEPTS ARL Assignable causes Attributes control charts Average run length C chart Cause-and-effect diagram Center line Chance causes Control chart Control limits Cumulative sum control chart

Defect concentration diagram Defects-per-unit chart Deming’s 14 points Exponentially weighted moving average control chart (EWMA) False alarm Fraction-defective control chart Implementing SPC Individuals control chart ( ⌾ chart)

Moving range NP chart P chart Pareto diagram PCR PCRk Problem-solving tools Process capability Process capability ratio Quality control R chart Rational subgroup Run rule S chart

Shewhart control chart Six-sigma process Specification limits Statistical process control (SPC) Statistical quality control U chart V mask Variables control charts Warning limits Western Electric rules ⌾ chart

JWCL232_AppA_702-730.qxd

1/18/10

1:21 PM

Page 702

APPENDICES APPENDIX A. STATISTICAL TABLES AND CHARTS 703 Table I Table II Table III Table IV Table V Table VI Chart VII Table VIII Table IX Table X Table XI Table XII

Summary of Common Probability Distributions 704 Cumulative Binomial Probabilities P (X  x) 705 Cumulative Standard Normal Distribution 708 2 Percentage Points , of the Chi-Squared Distribution 710 Percentage Points t, of the t Distribution 711 Percentage Points f,1,2 of the F Distribution 712 Operating Characteristic Curves 717 Critical Values for the Sign Test 726 Critical Values for the Wilcoxon Signed-Rank Test 726 Critical Values for the Wilcoxon Rank-Sum Test 727 Factors for Constructing Variables Control Charts 728 Factors for Tolerance Intervals 729

APPENDIX B. ANSWERS TO SELECTED EXERCISES 731 APPENDIX C. BIBLIOGRAPHY 747

702

JWCL232_AppA_702-730.qxd

1/18/10

1:21 PM

Page 703

Appendix A Statistical Tables and Charts

703

JWCL232_AppA_702-730.qxd

704

1/18/10

1:21 PM

Page 704

APPENDIX A STATISTICAL TABLES AND CHARTS

Table I Summary of Common Probability Distributions Probability Distribution

Name

Mean

Variance

Section in Book

1b  a2

1b a  12 2 1

3-5

np

np11 p2

3-6

1p

11 p2 p2

3-7

rp

r11 p2 p2

3-7

Discrete Uniform

1 n, a  b

Binomial

n a b px 11 p2 n x x

2

12

x  0, 1, . . . , n, 0  p  1 11 p2 x 1p x  1, 2, . . . , 0  p  1

Geometric

Negative binomial

x 1 b 11 p2 x rpr r 1 x  r, r  1, r  2, . . . , 0  p  1 a

K N K b a ba n x x N a b n

Hypergeometric

np11 p2 a

np where p 

x  max10, n N  K2, 1, . . .

N n b N 1

3-8

K N

min1K, n2, K  N, n  N Poisson

e

␭ x

␭ , x  0, 1, 2, . . . , 0 ␭ x!

Continuous 1 ,axb b a

Uniform

1

Normal

 12

e 2 1 1





3-9

1b  a2

1b a2 2

4-5



2

4-6

1␭

1␭2

4-8

r␭

r␭2

4-9.1

r␭

r␭2

4-9.2

2

x  2  2

12

 x ,   , 0  Exponential

␭e ␭x, 0  x, 0 ␭ ␭x e , 0 x, r  1, 2, . . . 1r 12! r r 1 ␭x

Erlang Gamma

␭xr 1e ␭x , 0 x, 0 r, 0 ␭ 1r2

Weibull

x

1 1x 2

a b e   

 a1 

1 b



2 c a1 

0 x, 0 , 0  Lognormal

Beta

1 x 22

exp a

3ln1x2  4 2 2

2

1  2  1 x 11 x2

1 12 1 2 0  x  1, 0 , 0

2 a1 

b

2 b

4-10

1 2 bd

e 2

e2 1e 12

4-11

  



1  2 2 1  12

4-12

2

2

2

JWCL232_AppA_702-730.qxd

1/18/10

1:21 PM

Page 705

APPENDIX A

705

Table II Cumulative Binomial Probabilities P(X  x) P n

x

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0.95

0.99

1 2

0 0 1 0 1 2 0 1 2 3 0 1 2 3 4 0 1 2 3 4 5 0 1 2 3 4 5 6 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 8

0.9000 0.8100 0.9900 0.7290 0.9720 0.9990 0.6561 0.9477 0.9963 0.9999 0.5905 0.9185 0.9914 0.9995 1.0000 0.5314 0.8857 0.9842 0.9987 0.9999 1.0000 0.4783 0.8503 0.9743 0.9973 0.9998 1.0000 1.0000 0.4305 0.8131 0.9619 0.9950 0.9996 1.0000 1.0000 1.0000 0.3874 0.7748 0.9470 0.9917 0.9991 0.9999 1.0000 1.0000 1.0000

0.8000 0.6400 0.9600 0.5120 0.8960 0.9920 0.4096 0.8192 0.9728 0.9984 0.3277 0.7373 0.9421 0.9933 0.9997 0.2621 0.6554 0.9011 0.9830 0.9984 0.9999 0.2097 0.5767 0.8520 0.9667 0.9953 0.9996 1.0000 0.1678 0.5033 0.7969 0.9437 0.9896 0.9988 0.9999 1.0000 0.1342 0.4362 0.7382 0.9144 0.9804 0.9969 0.9997 1.0000 1.0000

0.7000 0.4900 0.9100 0.3430 0.7840 0.9730 0.2401 0.6517 0.9163 0.9919 0.1681 0.5282 0.8369 0.9692 0.9976 0.1176 0.4202 0.7443 0.9295 0.9891 0.9993 0.0824 0.3294 0.6471 0.8740 0.9712 0.9962 0.9998 0.0576 0.2553 0.5518 0.8059 0.9420 0.9887 0.9987 0.9999 0.0404 0.1960 0.4628 0.7297 0.9012 0.9747 0.9957 0.9996 1.0000

0.6000 0.3600 0.8400 0.2160 0.6480 0.9360 0.1296 0.4752 0.8208 0.9744 0.0778 0.3370 0.6826 0.9130 0.9898 0.0467 0.2333 0.5443 0.8208 0.9590 0.9959 0.0280 0.1586 0.4199 0.7102 0.9037 0.9812 0.9984 0.0168 0.1064 0.3154 0.5941 0.8263 0.9502 0.9915 0.9993 0.0101 0.0705 0.2318 0.4826 0.7334 0.9006 0.9750 0.9962 0.9997

0.5000 0.2500 0.7500 0.1250 0.5000 0.8750 0.0625 0.3125 0.6875 0.9375 0.0313 0.1875 0.5000 0.8125 0.6988 0.0156 0.1094 0.3438 0.6563 0.9806 0.9844 0.0078 0.0625 0.2266 0.5000 0.7734 0.9375 0.9922 0.0039 0.0352 0.1445 0.3633 0.6367 0.8555 0.9648 0.9961 0.0020 0.0195 0.0889 0.2539 0.5000 0.7461 0.9102 0.9805 0.9980

0.4000 0.1600 0.6400 0.0640 0.3520 0.7840 0.0256 0.1792 0.5248 0.8704 0.0102 0.0870 0.3174 0.6630 0.9222 0.0041 0.0410 0.1792 0.4557 0.7667 0.9533 0.0016 0.0188 0.0963 0.2898 0.5801 0.8414 0.9720 0.0007 0.0085 0.0498 0.1737 0.4059 0.6846 0.8936 0.9832 0.0003 0.0038 0.0250 0.0994 0.2666 0.5174 0.7682 0.9295 0.9899

0.3000 0.0900 0.5100 0.0270 0.2160 0.6570 0.0081 0.0837 0.3483 0.7599 0.0024 0.0308 0.1631 0.4718 0.8319 0.0007 0.0109 0.0705 0.2557 0.5798 0.8824 0.0002 0.0038 0.0288 0.1260 0.3529 0.6706 0.9176 0.0001 0.0013 0.0113 0.0580 0.1941 0.4482 0.7447 0.9424 0.0000 0.0004 0.0043 0.0253 0.0988 0.2703 0.5372 0.8040 0.9596

0.2000 0.0400 0.3600 0.0080 0.1040 0.4880 0.0016 0.0272 0.1808 0.5904 0.0003 0.0067 0.0579 0.2627 0.6723 0.0001 0.0016 0.0170 0.0989 0.3446 0.7379 0.0000 0.0004 0.0047 0.0333 0.1480 0.4233 0.7903 0.0000 0.0001 0.0012 0.0104 0.0563 0.2031 0.4967 0.8322 0.0000 0.0000 0.0003 0.0031 0.0196 0.0856 0.2618 0.5638 0.8658

0.1000 0.0100 0.1900 0.0010 0.0280 0.2710 0.0001 0.0037 0.0523 0.3439 0.0000 0.0005 0.0086 0.0815 0.4095 0.0000 0.0001 0.0013 0.0159 0.1143 0.4686 0.0000 0.0000 0.0002 0.0027 0.0257 0.1497 0.5217 0.0000 0.0000 0.0000 0.0004 0.0050 0.0381 0.1869 0.5695 0.0000 0.0000 0.0000 0.0001 0.0009 0.0083 0.0530 0.2252 0.6126

0.0500 0.0025 0.0975 0.0001 0.0073 0.1426 0.0000 0.0005 0.0140 0.1855 0.0000 0.0000 0.0012 0.0226 0.2262 0.0000 0.0000 0.0001 0.0022 0.0328 0.2649 0.0000 0.0000 0.0000 0.0002 0.0038 0.0444 0.3017 0.0000 0.0000 0.0000 0.0000 0.0004 0.0058 0.0572 0.3366 0.0000 0.0000 0.0000 0.0000 0.0000 0.0006 0.0084 0.0712 0.3698

0.0100 0.0001 0.0199 0.0000 0.0003 0.0297 0.0000 0.0000 0.0006 0.0394 0.0000 0.0000 0.0000 0.0010 0.0490 0.0000 0.0000 0.0000 0.0000 0.0015 0.0585 0.0000 0.0000 0.0000 0.0000 0.0000 0.0020 0.0679 0.0000 0.0000 0.0000 0.0000 0.0000 0.0001 0.0027 0.0773 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0001 0.0034 0.0865

3

4

5

6

7

8

9

JWCL232_AppA_702-730.qxd

706

1/18/10

1:21 PM

Page 706

APPENDIX A STATISTICAL TABLES AND CHARTS

Table II Cumulative Binomial Probabilities P(X  x) (continued) P n

x

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0.95

0.99

10

0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 11 0 1 2 3 4 5 6 7 8 9 10 11

0.3487 0.7361 0.9298 0.9872 0.9984 0.9999 1.0000 1.0000 1.0000 1.0000 0.3138 0.6974 0.9104 0.9815 0.9972 0.9997 1.0000 1.0000 1.0000 1.0000 1.0000 0.2824 0.6590 0.8891 0.9744 0.9957 0.9995 0.9999 1.0000 1.0000 1.0000 1.0000 1.0000 0.2542 0.6213 0.8661 0.9658 0.9935 0.9991 0.9999 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

0.1074 0.3758 0.6778 0.8791 0.9672 0.9936 0.9991 0.9999 1.0000 1.0000 0.0859 0.3221 0.6174 0.8389 0.9496 0.9883 0.9980 0.9998 1.0000 1.0000 1.0000 0.0687 0.2749 0.5583 0.7946 0.9274 0.9806 0.9961 0.9994 0.9999 1.0000 1.0000 1.0000 0.0550 0.2336 0.5017 0.7473 0.9009 0.9700 0.9930 0.9988 0.9988 1.0000 1.0000 1.0000 1.0000

0.0282 0.1493 0.3828 0.6496 0.8497 0.9527 0.9894 0.9984 0.9999 1.0000 0.0198 0.1130 0.3127 0.5696 0.7897 0.9218 0.9784 0.9957 0.9994 1.0000 1.0000 0.0138 0.0850 0.2528 0.4925 0.7237 0.8822 0.9614 0.9905 0.9983 0.9998 1.0000 1.0000 0.0097 0.0637 0.2025 0.4206 0.6543 0.8346 0.9376 0.9818 0.9960 0.9993 0.9999 1.0000 1.0000

0.0060 0.0464 0.1673 0.3823 0.6331 0.8338 0.9452 0.9877 0.9983 0.9999 0.0036 0.0302 0.1189 0.2963 0.5328 0.7535 0.9006 0.9707 0.9941 0.9993 1.0000 0.0022 0.0196 0.0834 0.2253 0.4382 0.6652 0.8418 0.9427 0.9847 0.9972 0.9997 1.0000 0.0013 0.0126 0.0579 0.1686 0.3530 0.5744 0.7712 0.9023 0.9679 0.9922 0.9987 0.9999 1.0000

0.0010 0.0107 0.0547 0.1719 0.3770 0.6230 0.8281 0.9453 0.9893 0.9990 0.0005 0.0059 0.0327 0.1133 0.2744 0.5000 0.7256 0.8867 0.9673 0.9941 0.9995 0.0002 0.0032 0.0193 0.0730 0.1938 0.3872 0.6128 0.8062 0.9270 0.9807 0.9968 0.9998 0.0001 0.0017 0.0112 0.0461 0.1334 0.2905 0.5000 0.7095 0.8666 0.9539 0.9888 0.9983 0.9999

0.0001 0.0017 0.0123 0.0548 0.1662 0.3669 0.6177 0.8327 0.9536 0.9940 0.0000 0.0007 0.0059 0.0293 0.0994 0.2465 0.4672 0.7037 0.8811 0.9698 0.9964 0.0000 0.0003 0.0028 0.0153 0.0573 0.1582 0.3348 0.5618 0.7747 0.9166 0.9804 0.9978 0.0000 0.0001 0.0013 0.0078 0.0321 0.0977 0.2288 0.4256 0.6470 0.8314 0.9421 0.9874 0.9987

0.0000 0.0001 0.0016 0.0106 0.0473 0.1503 0.3504 0.6172 0.8507 0.9718 0.0000 0.0000 0.0006 0.0043 0.0216 0.0782 0.2103 0.4304 0.6873 0.8870 0.9802 0.0000 0.0000 0.0002 0.0017 0.0095 0.0386 0.1178 0.2763 0.5075 0.7472 0.9150 0.9862 0.0000 0.0000 0.0001 0.0007 0.0040 0.0182 0.0624 0.1654 0.3457 0.5794 0.7975 0.9363 0.9903

0.0000 0.0000 0.0001 0.0009 0.0064 0.0328 0.1209 0.3222 0.6242 0.8926 0.0000 0.0000 0.0000 0.0002 0.0020 0.0117 0.0504 0.1611 0.3826 0.6779 0.9141 0.0000 0.0000 0.0000 0.0001 0.0006 0.0039 0.0194 0.0726 0.2054 0.4417 0.7251 0.9313 0.0000 0.0000 0.0000 0.0000 0.0002 0.0012 0.0070 0.0300 0.0991 0.2527 0.4983 0.7664 0.9450

0.0000 0.0000 0.0000 0.0000 0.0001 0.0016 0.0128 0.0702 0.2639 0.6513 0.0000 0.0000 0.0000 0.0000 0.0000 0.0003 0.0028 0.0185 0.0896 0.3026 0.6862 0.0000 0.0000 0.0000 0.0000 0.0000 0.0001 0.0005 0.0043 0.0256 0.1109 0.3410 0.7176 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0001 0.0009 0.0065 0.0342 0.1339 0.3787 0.7458

0.0000 0.0000 0.0000 0.0000 0.0000 0.0001 0.0010 0.0115 0.0861 0.4013 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0001 0.0016 0.0152 0.1019 0.4312 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0002 0.0022 0.0196 0.1184 0.4596 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0003 0.0031 0.0245 0.1354 0.4867

0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0001 0.0043 0.0956 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0002 0.0052 0.1047 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0002 0.0062 0.1136 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0003 0.0072 0.1225

11

12

13

12

JWCL232_AppA_702-730.qxd

1/18/10

1:21 PM

Page 707

Table II Cumulative Binomial Probabilities P(X  x) (continued) P n

x

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0.95

0.99

14

0 1 2 3 4 5 6 7 8

0.2288 0.5846 0.8416 0.9559 0.9908 0.9985 0.9998 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.2059 0.5490 0.8159 0.9444 0.9873 0.9978 0.9997 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.1216 0.3917 0.6769 0.8670 0.9568 0.9887 0.9976 0.9996 0.9999 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

0.0440 0.1979 0.4481 0.6982 0.8702 0.9561 0.9884 0.9976 0.9996 1.0000 1.0000 1.0000 1.0000 1.0000 0.0352 0.1671 0.3980 0.6482 0.8358 0.9389 0.9819 0.9958 0.9992 0.9999 1.0000 1.0000 1.0000 1.0000 1.0000 0.0115 0.0692 0.2061 0.4114 0.6296 0.8042 0.9133 0.9679 0.9900 0.9974 0.9994 0.9999 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

0.0068 0.0475 0.1608 0.3552 0.5842 0.7805 0.9067 0.9685 0.9917 0.9983 0.9998 1.0000 1.0000 1.0000 0.0047 0.0353 0.1268 0.2969 0.5155 0.7216 0.8689 0.9500 0.9848 0.9963 0.9993 0.9999 1.0000 1.0000 1.0000 0.0008 0.0076 0.0355 0.1071 0.2375 0.4164 0.6080 0.7723 0.8867 0.9520 0.9829 0.9949 0.9987 0.9997 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

0.0008 0.0081 0.0398 0.1243 0.2793 0.4859 0.6925 0.8499 0.9417 0.9825 0.9961 0.9994 0.9999 1.0000 0.0005 0.0052 0.0271 0.0905 0.2173 0.4032 0.6098 0.7869 0.9050 0.9662 0.9907 0.9981 0.9997 1.0000 1.0000 0.0000 0.0005 0.0036 0.0160 0.0510 0.1256 0.2500 0.4159 0.5956 0.7553 0.8725 0.9435 0.9790 0.9935 0.9984 0.9997 1.0000 1.0000 1.0000 1.0000

0.0001 0.0009 0.0065 0.0287 0.0898 0.2120 0.3953 0.6047 0.7880 0.9102 0.9713 0.9935 0.9991 0.9999 0.0000 0.0005 0.0037 0.0176 0.0592 0.1509 0.3036 0.5000 0.6964 0.8491 0.9408 0.9824 0.9963 0.9995 1.0000 0.0000 0.0000 0.0002 0.0013 0.0059 0.0207 0.0577 0.1316 0.2517 0.4119 0.5881 0.7483 0.8684 0.9423 0.9793 0.9941 0.9987 0.9998 1.0000 1.0000

0.0000 0.0001 0.0006 0.0039 0.0175 0.0583 0.1501 0.3075 0.5141 0.7207 0.8757 0.9602 0.9919 0.9992 0.0000 0.0000 0.0003 0.0019 0.0093 0.0338 0.0950 0.2131 0.3902 0.5968 0.7827 0.9095 0.9729 0.9948 0.9995 0.0000 0.0000 0.0000 0.0000 0.0003 0.0016 0.0065 0.0210 0.0565 0.1275 0.2447 0.4044 0.5841 0.7500 0.8744 0.9490 0.9840 0.9964 0.9995 1.0000

0.0000 0.0000 0.0000 0.0002 0.0017 0.0083 0.0315 0.0933 0.2195 0.4158 0.6448 0.8392 0.9525 0.9932 0.0000 0.0000 0.0000 0.0001 0.0007 0.0037 0.0152 0.0500 0.1311 0.2784 0.4845 0.7031 0.8732 0.9647 0.9953 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0003 0.0013 0.0051 0.0171 0.0480 0.1133 0.2277 0.3920 0.5836 0.7625 0.8929 0.9645 0.9924 0.9992

0.0000 0.0000 0.0000 0.0000 0.0000 0.0004 0.0024 0.0116 0.0439 0.1298 0.3018 0.5519 0.8021 0.9560 0.0000 0.0000 0.0000 0.0000 0.0000 0.0001 0.0008 0.0042 0.0181 0.0611 0.1642 0.3518 0.6020 0.8329 0.9648 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0001 0.0006 0.0026 0.0100 0.0321 0.0867 0.1958 0.3704 0.5886 0.7939 0.9308 0.9885

0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0002 0.0015 0.0092 0.0441 0.1584 0.4154 0.7712 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0003 0.0022 0.0127 0.0556 01841 0.4510 0.7941 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0001 0.0004 0.0024 0.0113 0.0432 0.1330 0.3231 0.6083 0.8784

0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0004 0.0042 0.0301 0.1530 0.5123 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0001 0.0006 0.0055 0.0362 0.1710 0.5367 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0003 0.0026 0.0159 0.0755 0.2642 0.6415

0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0003 0.0084 0.1313 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0004 0.0096 0.1399 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0010 0.0169 0.1821

15

20

9 10 11 12 13 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

JWCL232_AppA_702-730.qxd

708

1/18/10

1:21 PM

Page 708

APPENDIX A STATISTICAL TABLES AND CHARTS

1z2  P1Z  z2 

冮

z

1

 22

e

12 u

2

du

Φ (z)

z

0

Table III Cumulative Standard Normal Distribution z

0.09

0.08

0.07

0.06

0.05

0.04

0.03

0.02

0.01

0.00

3.9

3.8

3.7

3.6

3.5

3.4

3.3

3.2

3.1

3.0

2.9

2.8

2.7

2.6

2.5

2.4

2.3

2.2

2.1

2.0

1.9

1.8

1.7

1.6

1.5

1.4

1.3

1.2

1.1

1.0

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1 0.0

0.000033 0.000050 0.000075 0.000112 0.000165 0.000242 0.000350 0.000501 0.000711 0.001001 0.001395 0.001926 0.002635 0.003573 0.004799 0.006387 0.008424 0.011011 0.014262 0.018309 0.023295 0.029379 0.036727 0.045514 0.055917 0.068112 0.082264 0.098525 0.117023 0.137857 0.161087 0.186733 0.214764 0.245097 0.277595 0.312067 0.348268 0.385908 0.424655 0.464144

0.000034 0.000052 0.000078 0.000117 0.000172 0.000251 0.000362 0.000519 0.000736 0.001035 0.001441 0.001988 0.002718 0.003681 0.004940 0.006569 0.008656 0.011304 0.014629 0.018763 0.023852 0.030054 0.037538 0.046479 0.057053 0.069437 0.083793 0.100273 0.119000 0.140071 0.163543 0.189430 0.217695 0.248252 0.280957 0.315614 0.351973 0.389739 0.428576 0.468119

0.000036 0.000054 0.000082 0.000121 0.000179 0.000260 0.000376 0.000538 0.000762 0.001070 0.001489 0.002052 0.002803 0.003793 0.005085 0.006756 0.008894 0.011604 0.015003 0.019226 0.024419 0.030742 0.038364 0.047460 0.058208 0.070781 0.085343 0.102042 0.121001 0.142310 0.166023 0.192150 0.220650 0.251429 0.284339 0.319178 0.355691 0.393580 0.432505 0.472097

0.000037 0.000057 0.000085 0.000126 0.000185 0.000270 0.000390 0.000557 0.000789 0.001107 0.001538 0.002118 0.002890 0.003907 0.005234 0.006947 0.009137 0.011911 0.015386 0.019699 0.024998 0.031443 0.039204 0.048457 0.059380 0.072145 0.086915 0.103835 0.123024 0.144572 0.168528 0.194894 0.223627 0.254627 0.287740 0.322758 0.359424 0.397432 0.436441 0.476078

0.000039 0.000059 0.000088 0.000131 0.000193 0.000280 0.000404 0.000577 0.000816 0.001144 0.001589 0.002186 0.002980 0.004025 0.005386 0.007143 0.009387 0.012224 0.015778 0.020182 0.025588 0.032157 0.040059 0.049471 0.060571 0.073529 0.088508 0.105650 0.125072 0.146859 0.171056 0.197662 0.226627 0.257846 0.291160 0.326355 0.363169 0.401294 0.440382 0.480061

0.000041 0.000062 0.000092 0.000136 0.000200 0.000291 0.000419 0.000598 0.000845 0.001183 0.001641 0.002256 0.003072 0.004145 0.005543 0.007344 0.009642 0.012545 0.016177 0.020675 0.026190 0.032884 0.040929 0.050503 0.061780 0.074934 0.090123 0.107488 0.127143 0.149170 0.173609 0.200454 0.229650 0.261086 0.294599 0.329969 0.366928 0.405165 0.444330 0.484047

0.000042 0.000064 0.000096 0.000142 0.000208 0.000302 0.000434 0.000619 0.000874 0.001223 0.001695 0.002327 0.003167 0.004269 0.005703 0.007549 0.009903 0.012874 0.016586 0.021178 0.026803 0.033625 0.041815 0.051551 0.063008 0.076359 0.091759 0.109349 0.129238 0.151505 0.176185 0.203269 0.232695 0.264347 0.298056 0.333598 0.370700 0.409046 0.448283 0.488033

0.000044 0.000067 0.000100 0.000147 0.000216 0.000313 0.000450 0.000641 0.000904 0.001264 0.001750 0.002401 0.003264 0.004396 0.005868 0.007760 0.010170 0.013209 0.017003 0.021692 0.027429 0.034379 0.042716 0.052616 0.064256 0.077804 0.093418 0.111233 0.131357 0.153864 0.178786 0.206108 0.235762 0.267629 0.301532 0.337243 0.374484 0.412936 0.452242 0.492022

0.000046 0.000069 0.000104 0.000153 0.000224 0.000325 0.000467 0.000664 0.000935 0.001306 0.001807 0.002477 0.003364 0.004527 0.006037 0.007976 0.010444 0.013553 0.017429 0.022216 0.028067 0.035148 0.043633 0.053699 0.065522 0.079270 0.095098 0.113140 0.133500 0.156248 0.181411 0.208970 0.238852 0.270931 0.305026 0.340903 0.378281 0.416834 0.456205 0.496011

0.000048 0.000072 0.000108 0.000159 0.000233 0.000337 0.000483 0.000687 0.000968 0.001350 0.001866 0.002555 0.003467 0.004661 0.006210 0.008198 0.010724 0.013903 0.017864 0.022750 0.028717 0.035930 0.044565 0.054799 0.066807 0.080757 0.096801 0.115070 0.135666 0.158655 0.184060 0.211855 0.241964 0.274253 0.308538 0.344578 0.382089 0.420740 0.460172 0.500000

JWCL232_AppA_702-730.qxd

1/18/10

1:21 PM

Page 709

APPENDIX A

1z2  P1Z  z2 

冮

z

1

 22

e

12 u

709

2

du

Φ (z)

z

0

Table III Cumulative Standard Normal Distribution (continued) z

0.00

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3.0 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9

0.500000 0.539828 0.579260 0.617911 0.655422 0.691462 0.725747 0.758036 0.788145 0.815940 0.841345 0.864334 0.884930 0.903199 0.919243 0.933193 0.945201 0.955435 0.964070 0.971283 0.977250 0.982136 0.986097 0.989276 0.991802 0.993790 0.995339 0.996533 0.997445 0.998134 0.998650 0.999032 0.999313 0.999517 0.999663 0.999767 0.999841 0.999892 0.999928 0.999952

0.503989 0.543795 0.583166 0.621719 0.659097 0.694974 0.729069 0.761148 0.791030 0.818589 0.843752 0.866500 0.886860 0.904902 0.920730 0.934478 0.946301 0.956367 0.964852 0.971933 0.977784 0.982571 0.986447 0.989556 0.992024 0.993963 0.995473 0.996636 0.997523 0.998193 0.998694 0.999065 0.999336 0.999533 0.999675 0.999776 0.999847 0.999896 0.999931 0.999954

0.507978 0.547758 0.587064 0.625516 0.662757 0.698468 0.732371 0.764238 0.793892 0.821214 0.846136 0.868643 0.888767 0.906582 0.922196 0.935744 0.947384 0.957284 0.965621 0.972571 0.978308 0.982997 0.986791 0.989830 0.992240 0.994132 0.995604 0.996736 0.997599 0.998250 0.998736 0.999096 0.999359 0.999550 0.999687 0.999784 0.999853 0.999900 0.999933 0.999956

0.511967 0.551717 0.590954 0.629300 0.666402 0.701944 0.735653 0.767305 0.796731 0.823815 0.848495 0.870762 0.890651 0.908241 0.923641 0.936992 0.948449 0.958185 0.966375 0.973197 0.978822 0.983414 0.987126 0.990097 0.992451 0.994297 0.995731 0.996833 0.997673 0.998305 0.998777 0.999126 0.999381 0.999566 0.999698 0.999792 0.999858 0.999904 0.999936 0.999958

0.515953 0.555760 0.594835 0.633072 0.670031 0.705401 0.738914 0.770350 0.799546 0.826391 0.850830 0.872857 0.892512 0.909877 0.925066 0.938220 0.949497 0.959071 0.967116 0.973810 0.979325 0.983823 0.987455 0.990358 0.992656 0.994457 0.995855 0.996928 0.997744 0.998359 0.998817 0.999155 0.999402 0.999581 0.999709 0.999800 0.999864 0.999908 0.999938 0.999959

0.519939 0.559618 0.598706 0.636831 0.673645 0.708840 0.742154 0.773373 0.802338 0.828944 0.853141 0.874928 0.894350 0.911492 0.926471 0.939429 0.950529 0.959941 0.967843 0.974412 0.979818 0.984222 0.987776 0.990613 0.992857 0.994614 0.995975 0.997020 0.997814 0.998411 0.998856 0.999184 0.999423 0.999596 0.999720 0.999807 0.999869 0.999912 0.999941 0.999961

0.532922 0.563559 0.602568 0.640576 0.677242 0.712260 0.745373 0.776373 0.805106 0.831472 0.855428 0.876976 0.896165 0.913085 0.927855 0.940620 0.951543 0.960796 0.968557 0.975002 0.980301 0.984614 0.988089 0.990863 0.993053 0.994766 0.996093 0.997110 0.997882 0.998462 0.998893 0.999211 0.999443 0.999610 0.999730 0.999815 0.999874 0.999915 0.999943 0.999963

0.527903 0.567495 0.606420 0.644309 0.680822 0.715661 0.748571 0.779350 0.807850 0.833977 0.857690 0.878999 0.897958 0.914657 0.929219 0.941792 0.952540 0.961636 0.969258 0.975581 0.980774 0.984997 0.988396 0.991106 0.993244 0.994915 0.996207 0.997197 0.997948 0.998511 0.998930 0.999238 0.999462 0.999624 0.999740 0.999821 0.999879 0.999918 0.999946 0.999964

0.531881 0.571424 0.610261 0.648027 0.684386 0.719043 0.751748 0.782305 0.810570 0.836457 0.859929 0.881000 0.899727 0.916207 0.930563 0.942947 0.953521 0.962462 0.969946 0.976148 0.981237 0.985371 0.988696 0.991344 0.993431 0.995060 0.996319 0.997282 0.998012 0.998559 0.998965 0.999264 0.999481 0.999638 0.999749 0.999828 0.999883 0.999922 0.999948 0.999966

0.535856 0.575345 0.614092 0.651732 0.687933 0.722405 0.754903 0.785236 0.813267 0.838913 0.862143 0.882977 0.901475 0.917736 0.931888 0.944083 0.954486 0.963273 0.970621 0.976705 0.981691 0.985738 0.988989 0.991576 0.993613 0.995201 0.996427 0.997365 0.998074 0.998605 0.998999 0.999289 0.999499 0.999650 0.999758 0.999835 0.999888 0.999925 0.999950 0.999967

JWCL232_AppA_702-730.qxd

710

1/18/10

1:21 PM

Page 710

APPENDIX A STATISTICAL TABLES AND CHARTS

α

2 χα, ν

Table IV Percentage Points 2, of the Chi-Squared Distribution  1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 40 50 60 70 80 90 100

 .995

.990

.975

.950

.900

.500

.100

.050

.025

.010

.005

.00 .01 .07 .21 .41 .68 .99 1.34 1.73 2.16 2.60 3.07 3.57 4.07 4.60 5.14 5.70 6.26 6.84 7.43 8.03 8.64 9.26 9.89 10.52 11.16 11.81 12.46 13.12 13.79 20.71 27.99 35.53 43.28 51.17 59.20 67.33

.00 .02 .11 .30 .55 .87 1.24 1.65 2.09 2.56 3.05 3.57 4.11 4.66 5.23 5.81 6.41 7.01 7.63 8.26 8.90 9.54 10.20 10.86 11.52 12.20 12.88 13.57 14.26 14.95 22.16 29.71 37.48 45.44 53.54 61.75 70.06

.00 .05 .22 .48 .83 1.24 1.69 2.18 2.70 3.25 3.82 4.40 5.01 5.63 6.27 6.91 7.56 8.23 8.91 9.59 10.28 10.98 11.69 12.40 13.12 13.84 14.57 15.31 16.05 16.79 24.43 32.36 40.48 48.76 57.15 65.65 74.22

.00 .10 .35 .71 1.15 1.64 2.17 2.73 3.33 3.94 4.57 5.23 5.89 6.57 7.26 7.96 8.67 9.39 10.12 10.85 11.59 12.34 13.09 13.85 14.61 15.38 16.15 16.93 17.71 18.49 26.51 34.76 43.19 51.74 60.39 69.13 77.93

.02 .21 .58 1.06 1.61 2.20 2.83 3.49 4.17 4.87 5.58 6.30 7.04 7.79 8.55 9.31 10.09 10.87 11.65 12.44 13.24 14.04 14.85 15.66 16.47 17.29 18.11 18.94 19.77 20.60 29.05 37.69 46.46 55.33 64.28 73.29 82.36

.45 1.39 2.37 3.36 4.35 5.35 6.35 7.34 8.34 9.34 10.34 11.34 12.34 13.34 14.34 15.34 16.34 17.34 18.34 19.34 20.34 21.34 22.34 23.34 24.34 25.34 26.34 27.34 28.34 29.34 39.34 49.33 59.33 69.33 79.33 89.33 99.33

2.71 4.61 6.25 7.78 9.24 10.65 12.02 13.36 14.68 15.99 17.28 18.55 19.81 21.06 22.31 23.54 24.77 25.99 27.20 28.41 29.62 30.81 32.01 33.20 34.28 35.56 36.74 37.92 39.09 40.26 51.81 63.17 74.40 85.53 96.58 107.57 118.50

3.84 5.99 7.81 9.49 11.07 12.59 14.07 15.51 16.92 18.31 19.68 21.03 22.36 23.68 25.00 26.30 27.59 28.87 30.14 31.41 32.67 33.92 35.17 36.42 37.65 38.89 40.11 41.34 42.56 43.77 55.76 67.50 79.08 90.53 101.88 113.14 124.34

5.02 7.38 9.35 11.14 12.83 14.45 16.01 17.53 19.02 20.48 21.92 23.34 24.74 26.12 27.49 28.85 30.19 31.53 32.85 34.17 35.48 36.78 38.08 39.36 40.65 41.92 43.19 44.46 45.72 46.98 59.34 71.42 83.30 95.02 106.63 118.14 129.56

6.63 9.21 11.34 13.28 15.09 16.81 18.48 20.09 21.67 23.21 24.72 26.22 27.69 29.14 30.58 32.00 33.41 34.81 36.19 37.57 38.93 40.29 41.64 42.98 44.31 45.64 46.96 48.28 49.59 50.89 63.69 76.15 88.38 100.42 112.33 124.12 135.81

7.88 10.60 12.84 14.86 16.75 18.55 20.28 21.96 23.59 25.19 26.76 28.30 29.82 31.32 32.80 34.27 35.72 37.16 38.58 40.00 41.40 42.80 44.18 45.56 46.93 48.29 49.65 50.99 52.34 53.67 66.77 79.49 91.95 104.22 116.32 128.30 140.17

  degrees of freedom.

JWCL232_AppA_702-730.qxd

1/18/10

1:21 PM

Page 711

APPENDIX A

711

α t α, ν

0

Table V  1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 40 60 120 

Percentage Points t, of the t Distribution

 .40

.25

.10

.05

.025

.01

.005

.0025

.001

.0005

.325 .289 .277 .271 .267 .265 .263 .262 .261 .260 .260 .259 .259 .258 .258 .258 .257 .257 .257 .257 .257 .256 .256 .256 .256 .256 .256 .256 .256 .256 .255 .254 .254 .253

1.000 .816 .765 .741 .727 .718 .711 .706 .703 .700 .697 .695 .694 .692 .691 .690 .689 .688 .688 .687 .686 .686 .685 .685 .684 .684 .684 .683 .683 .683 .681 .679 .677 .674

3.078 1.886 1.638 1.533 1.476 1.440 1.415 1.397 1.383 1.372 1.363 1.356 1.350 1.345 1.341 1.337 1.333 1.330 1.328 1.325 1.323 1.321 1.319 1.318 1.316 1.315 1.314 1.313 1.311 1.310 1.303 1.296 1.289 1.282

6.314 2.920 2.353 2.132 2.015 1.943 1.895 1.860 1.833 1.812 1.796 1.782 1.771 1.761 1.753 1.746 1.740 1.734 1.729 1.725 1.721 1.717 1.714 1.711 1.708 1.706 1.703 1.701 1.699 1.697 1.684 1.671 1.658 1.645

12.706 4.303 3.182 2.776 2.571 2.447 2.365 2.306 2.262 2.228 2.201 2.179 2.160 2.145 2.131 2.120 2.110 2.101 2.093 2.086 2.080 2.074 2.069 2.064 2.060 2.056 2.052 2.048 2.045 2.042 2.021 2.000 1.980 1.960

31.821 6.965 4.541 3.747 3.365 3.143 2.998 2.896 2.821 2.764 2.718 2.681 2.650 2.624 2.602 2.583 2.567 2.552 2.539 2.528 2.518 2.508 2.500 2.492 2.485 2.479 2.473 2.467 2.462 2.457 2.423 2.390 2.358 2.326

63.657 9.925 5.841 4.604 4.032 3.707 3.499 3.355 3.250 3.169 3.106 3.055 3.012 2.977 2.947 2.921 2.898 2.878 2.861 2.845 2.831 2.819 2.807 2.797 2.787 2.779 2.771 2.763 2.756 2.750 2.704 2.660 2.617 2.576

127.32 14.089 7.453 5.598 4.773 4.317 4.029 3.833 3.690 3.581 3.497 3.428 3.372 3.326 3.286 3.252 3.222 3.197 3.174 3.153 3.135 3.119 3.104 3.091 3.078 3.067 3.057 3.047 3.038 3.030 2.971 2.915 2.860 2.807

318.31 23.326 10.213 7.173 5.893 5.208 4.785 4.501 4.297 4.144 4.025 3.930 3.852 3.787 3.733 3.686 3.646 3.610 3.579 3.552 3.527 3.505 3.485 3.467 3.450 3.435 3.421 3.408 3.396 3.385 3.307 3.232 3.160 3.090

636.62 31.598 12.924 8.610 6.869 5.959 5.408 5.041 4.781 4.587 4.437 4.318 4.221 4.140 4.073 4.015 3.965 3.922 3.883 3.850 3.819 3.792 3.767 3.745 3.725 3.707 3.690 3.674 3.659 3.646 3.551 3.460 3.373 3.291

  degrees of freedom.

v2

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 40 60 120 

v1

5.83 2.57 2.02 1.81 1.69 1.62 1.57 1.54 1.51 1.49 1.47 1.46 1.45 1.44 1.43 1.42 1.42 1.41 1.41 1.40 1.40 1.40 1.39 1.39 1.39 1.38 1.38 1.38 1.38 1.38 1.36 1.35 1.34 1.32

1 7.50 3.00 2.28 2.00 1.85 1.76 1.70 1.66 1.62 1.60 1.58 1.56 1.55 1.53 1.52 1.51 1.51 1.50 1.49 1.49 1.48 1.48 1.47 1.47 1.47 1.46 1.46 1.46 1.45 1.45 1.44 1.42 1.40 1.39

2 8.20 3.15 2.36 2.05 1.88 1.78 1.72 1.67 1.63 1.60 1.58 1.56 1.55 1.53 1.52 1.51 1.50 1.49 1.49 1.48 1.48 1.47 1.47 1.46 1.46 1.45 1.45 1.45 1.45 1.44 1.42 1.41 1.39 1.37

3

1

2

8.58 3.23 2.39 2.06 1.89 1.79 1.72 1.66 1.63 1.59 1.57 1.55 1.53 1.52 1.51 1.50 1.49 1.48 1.47 1.47 1.46 1.45 1.45 1.44 1.44 1.44 1.43 1.43 1.43 1.42 1.40 1.38 1.37 1.35

4 8.82 3.28 2.41 2.07 1.89 1.79 1.71 1.66 1.62 1.59 1.56 1.54 1.52 1.51 1.49 1.48 1.47 1.46 1.46 1.45 1.44 1.44 1.43 1.43 1.42 1.42 1.42 1.41 1.41 1.41 1.39 1.37 1.35 1.33

5 8.98 3.31 2.42 2.08 1.89 1.78 1.71 1.65 1.61 1.58 1.55 1.53 1.51 1.50 1.48 1.47 1.46 1.45 1.44 1.44 1.43 1.42 1.42 1.41 1.41 1.41 1.40 1.40 1.40 1.39 1.37 1.35 1.33 1.31

6

Table VI Percentage Points f,v ,v of the F Distribution

9.10 3.34 2.43 2.08 1.89 1.78 1.70 1.64 1.60 1.57 1.54 1.52 1.50 1.49 1.47 1.46 1.45 1.44 1.43 1.43 1.42 1.41 1.41 1.40 1.40 1.39 1.39 1.39 1.38 1.38 1.36 1.33 1.31 1.29

7 9.19 3.35 2.44 2.08 1.89 1.78 1.70 1.64 1.60 1.56 1.53 1.51 1.49 1.48 1.46 1.45 1.44 1.43 1.42 1.42 1.41 1.40 1.40 1.39 1.39 1.38 1.38 1.38 1.37 1.37 1.35 1.32 1.30 1.28

8 9.26 3.37 2.44 2.08 1.89 1.77 1.70 1.63 1.59 1.56 1.53 1.51 1.49 1.47 1.46 1.44 1.43 1.42 1.41 1.41 1.40 1.39 1.39 1.38 1.38 1.37 1.37 1.37 1.36 1.36 1.34 1.31 1.29 1.27

9 9.32 3.38 2.44 2.08 1.89 1.77 1.69 1.63 1.59 1.55 1.52 1.50 1.48 1.46 1.45 1.44 1.43 1.42 1.41 1.40 1.39 1.39 1.38 1.38 1.37 1.37 1.36 1.36 1.35 1.35 1.33 1.30 1.28 1.25

10 9.41 3.39 2.45 2.08 1.89 1.77 1.68 1.62 1.58 1.54 1.51 1.49 1.47 1.45 1.44 1.43 1.41 1.40 1.40 1.39 1.38 1.37 1.37 1.36 1.36 1.35 1.35 1.34 1.34 1.34 1.31 1.29 1.26 1.24

12 9.49 3.41 2.46 2.08 1.89 1.76 1.68 1.62 1.57 1.53 1.50 1.48 1.46 1.44 1.43 1.41 1.40 1.39 1.38 1.37 1.37 1.36 1.35 1.35 1.34 1.34 1.33 1.33 1.32 1.32 1.30 1.27 1.24 1.22

15

Degrees of freedom for the numerator (v1)

f0.25,v1,v2

f0.25, ␯ , ␯ 1 2

α = 0.25

9.58 3.43 2.46 2.08 1.88 1.76 1.67 1.61 1.56 1.52 1.49 1.47 1.45 1.43 1.41 1.40 1.39 1.38 1.37 1.36 1.35 1.34 1.34 1.33 1.33 1.32 1.32 1.31 1.31 1.30 1.28 1.25 1.22 1.19

20 9.63 3.43 2.46 2.08 1.88 1.75 1.67 1.60 1.56 1.52 1.49 1.46 1.44 1.42 1.41 1.39 1.38 1.37 1.36 1.35 1.34 1.33 1.33 1.32 1.32 1.31 1.31 1.30 1.30 1.29 1.26 1.24 1.21 1.18

24 9.67 3.44 2.47 2.08 1.88 1.75 1.66 1.60 1.55 1.51 1.48 1.45 1.43 1.41 1.40 1.38 1.37 1.36 1.35 1.34 1.33 1.32 1.32 1.31 1.31 1.30 1.30 1.29 1.29 1.28 1.25 1.22 1.19 1.16

30

9.71 3.45 2.47 2.08 1.88 1.75 1.66 1.59 1.54 1.51 1.47 1.45 1.42 1.41 1.39 1.37 1.36 1.35 1.34 1.33 1.32 1.31 1.31 1.30 1.29 1.29 1.28 1.28 1.27 1.27 1.24 1.21 1.18 1.14

40

9.76 3.46 2.47 2.08 1.87 1.74 1.65 1.59 1.54 1.50 1.47 1.44 1.42 1.40 1.38 1.36 1.35 1.34 1.33 1.32 1.31 1.30 1.30 1.29 1.28 1.28 1.27 1.27 1.26 1.26 1.22 1.19 1.16 1.12

60

9.80 3.47 2.47 2.08 1.87 1.74 1.65 1.58 1.53 1.49 1.46 1.43 1.41 1.39 1.37 1.35 1.34 1.33 1.32 1.31 1.30 1.29 1.28 1.28 1.27 1.26 1.26 1.25 1.25 1.24 1.21 1.17 1.13 1.08

120

9.85 3.48 2.47 2.08 1.87 1.74 1.65 1.58 1.53 1.48 1.45 1.42 1.40 1.38 1.36 1.34 1.33 1.32 1.30 1.29 1.28 1.28 1.27 1.26 1.25 1.25 1.24 1.24 1.23 1.23 1.19 1.15 1.10 1.00



1/18/10 1:21 PM

Degrees of freedom for the denominator (v2)

JWCL232_AppA_702-730.qxd Page 712

v2

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 40 60 120 

v1 1

39.86 8.53 5.54 4.54 4.06 3.78 3.59 3.46 3.36 3.29 3.23 3.18 3.14 3.10 3.07 3.05 3.03 3.01 2.99 2.97 2.96 2.95 2.94 2.93 2.92 2.91 2.90 2.89 2.89 2.88 2.84 2.79 2.75 2.71

2 49.50 9.00 5.46 4.32 3.78 3.46 3.26 3.11 3.01 2.92 2.86 2.81 2.76 2.73 2.70 2.67 2.64 2.62 2.61 2.59 2.57 2.56 2.55 2.54 2.53 2.52 2.51 2.50 2.50 2.49 2.44 2.39 2.35 2.30

3 53.59 9.16 5.39 4.19 3.62 3.29 3.07 2.92 2.81 2.73 2.66 2.61 2.56 2.52 2.49 2.46 2.44 2.42 2.40 2.38 2.36 2.35 2.34 2.33 2.32 2.31 2.30 2.29 2.28 2.28 2.23 2.18 2.13 2.08

1

2

4 55.83 9.24 5.34 4.11 3.52 3.18 2.96 2.81 2.69 2.61 2.54 2.48 2.43 2.39 2.36 2.33 2.31 2.29 2.27 2.25 2.23 2.22 2.21 2.19 2.18 2.17 2.17 2.16 2.15 2.14 2.09 2.04 1.99 1.94

5 57.24 9.29 5.31 4.05 3.45 3.11 2.88 2.73 2.61 2.52 2.45 2.39 2.35 2.31 2.27 2.24 2.22 2.20 2.18 2.16 2.14 2.13 2.11 2.10 2.09 2.08 2.07 2.06 2.06 2.03 2.00 1.95 1.90 1.85

6 58.20 9.33 5.28 4.01 3.40 3.05 2.83 2.67 2.55 2.46 2.39 2.33 2.28 2.24 2.21 2.18 2.15 2.13 2.11 2.09 2.08 2.06 2.05 2.04 2.02 2.01 2.00 2.00 1.99 1.98 1.93 1.87 1.82 1.77

7 58.91 9.35 5.27 3.98 3.37 3.01 2.78 2.62 2.51 2.41 2.34 2.28 2.23 2.19 2.16 2.13 2.10 2.08 2.06 2.04 2.02 2.01 1.99 1.98 1.97 1.96 1.95 1.94 1.93 1.93 1.87 1.82 1.77 1.72

f0.10,v1,v2

59.44 9.37 5.25 3.95 3.34 2.98 2.75 2.59 2.47 2.38 2.30 2.24 2.20 2.15 2.12 2.09 2.06 2.04 2.02 2.00 1.98 1.97 1.95 1.94 1.93 1.92 1.91 1.90 1.89 1.88 1.83 1.77 1.72 1.67

8 59.86 9.38 5.24 3.94 3.32 2.96 2.72 2.56 2.44 2.35 2.27 2.21 2.16 2.12 2.09 2.06 2.03 2.00 1.98 1.96 1.95 1.93 1.92 1.91 1.89 1.88 1.87 1.87 1.86 1.85 1.79 1.74 1.68 1.63

9 60.19 9.39 5.23 3.92 3.30 2.94 2.70 2.54 2.42 2.32 2.25 2.19 2.14 2.10 2.06 2.03 2.00 1.98 1.96 1.94 1.92 1.90 1.89 1.88 1.87 1.86 1.85 1.84 1.83 1.82 1.76 1.71 1.65 1.60

10 60.71 9.41 5.22 3.90 3.27 2.90 2.67 2.50 2.38 2.28 2.21 2.15 2.10 2.05 2.02 1.99 1.96 1.93 1.91 1.89 1.87 1.86 1.84 1.83 1.82 1.81 1.80 1.79 1.78 1.77 1.71 1.66 1.60 1.55

12 61.22 9.42 5.20 3.87 3.24 2.87 2.63 2.46 2.34 2.24 2.17 2.10 2.05 2.01 1.97 1.94 1.91 1.89 1.86 1.84 1.83 1.81 1.80 1.78 1.77 1.76 1.75 1.74 1.73 1.72 1.66 1.60 1.55 1.49

15

Degrees of freedom for the numerator (v1)

Table VI Percentage Points f,v ,v of the F Distribution (continued)

f0.10, ␯ , ␯ 1 2

20 61.74 9.44 5.18 3.84 3.21 2.84 2.59 2.42 2.30 2.20 2.12 2.06 2.01 1.96 1.92 1.89 1.86 1.84 1.81 1.79 1.78 1.76 1.74 1.73 1.72 1.71 1.70 1.69 1.68 1.67 1.61 1.54 1.48 1.42

α = 0.10

24 62.00 9.45 5.18 3.83 3.19 2.82 2.58 2.40 2.28 2.18 2.10 2.04 1.98 1.94 1.90 1.87 1.84 1.81 1.79 1.77 1.75 1.73 1.72 1.70 1.69 1.68 1.67 1.66 1.65 1.64 1.57 1.51 1.45 1.38

30 62.26 9.46 5.17 3.82 3.17 2.80 2.56 2.38 2.25 2.16 2.08 2.01 1.96 1.91 1.87 1.84 1.81 1.78 1.76 1.74 1.72 1.70 1.69 1.67 1.66 1.65 1.64 1.63 1.62 1.61 1.54 1.48 1.41 1.34

40 62.53 9.47 5.16 3.80 3.16 2.78 2.54 2.36 2.23 2.13 2.05 1.99 1.93 1.89 1.85 1.81 1.78 1.75 1.73 1.71 1.69 1.67 1.66 1.64 1.63 1.61 1.60 1.59 1.58 1.57 1.51 1.44 1.37 1.30

60 62.79 9.47 5.15 3.79 3.14 2.76 2.51 2.34 2.21 2.11 2.03 1.96 1.90 1.86 1.82 1.78 1.75 1.72 1.70 1.68 1.66 1.64 1.62 1.61 1.59 1.58 1.57 1.56 1.55 1.54 1.47 1.40 1.32 1.24

120 63.06 9.48 5.14 3.78 3.12 2.74 2.49 2.32 2.18 2.08 2.00 1.93 1.88 1.83 1.79 1.75 1.72 1.69 1.67 1.64 1.62 1.60 1.59 1.57 1.56 1.54 1.53 1.52 1.51 1.50 1.42 1.35 1.26 1.17

 63.33 9.49 5.13 3.76 3.10 2.72 2.47 2.29 2.16 2.06 1.97 1.90 1.85 1.80 1.76 1.72 1.69 1.66 1.63 1.61 1.59 1.57 1.55 1.53 1.52 1.50 1.49 1.48 1.47 1.46 1.38 1.29 1.19 1.00

1/18/10 1:21 PM

Degrees of freedom for the denominator (v2)

JWCL232_AppA_702-730.qxd Page 713

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 40 60 120 

v2

v1 2

3

2

4

5

6

7

8

9

10

12

15

Degrees of freedom for the numerator (v1)

f0.05,v1,v2

20

24

30

40

60

120



161.4 199.5 215.7 224.6 230.2 234.0 236.8 238.9 240.5 241.9 243.9 245.9 248.0 249.1 250.1 251.1 252.2 253.3 254.3 18.51 19.00 19.16 19.25 19.30 19.33 19.35 19.37 19.38 19.40 19.41 19.43 19.45 19.45 19.46 19.47 19.48 19.49 19.50 10.13 9.55 9.28 9.12 9.01 8.94 8.89 8.85 8.81 8.79 8.74 8.70 8.66 8.64 8.62 8.59 8.57 8.55 8.53 7.71 6.94 6.59 6.39 6.26 6.16 6.09 6.04 6.00 5.96 5.91 5.86 5.80 5.77 5.75 5.72 5.69 5.66 5.63 6.61 5.79 5.41 5.19 5.05 4.95 4.88 4.82 4.77 4.74 4.68 4.62 4.56 4.53 4.50 4.46 4.43 4.40 4.36 5.99 5.14 4.76 4.53 4.39 4.28 4.21 4.15 4.10 4.06 4.00 3.94 3.87 3.84 3.81 3.77 3.74 3.70 3.67 5.59 4.74 4.35 4.12 3.97 3.87 3.79 3.73 3.68 3.64 3.57 3.51 3.44 3.41 3.38 3.34 3.30 3.27 3.23 5.32 4.46 4.07 3.84 3.69 3.58 3.50 3.44 3.39 3.35 3.28 3.22 3.15 3.12 3.08 3.04 3.01 2.97 2.93 5.12 4.26 3.86 3.63 3.48 3.37 3.29 3.23 3.18 3.14 3.07 3.01 2.94 2.90 2.86 2.83 2.79 2.75 2.71 4.96 4.10 3.71 3.48 3.33 3.22 3.14 3.07 3.02 2.98 2.91 2.85 2.77 2.74 2.70 2.66 2.62 2.58 2.54 4.84 3.98 3.59 3.36 3.20 3.09 3.01 2.95 2.90 2.85 2.79 2.72 2.65 2.61 2.57 2.53 2.49 2.45 2.40 4.75 3.89 3.49 3.26 3.11 3.00 2.91 2.85 2.80 2.75 2.69 2.62 2.54 2.51 2.47 2.43 2.38 2.34 2.30 4.67 3.81 3.41 3.18 3.03 2.92 2.83 2.77 2.71 2.67 2.60 2.53 2.46 2.42 2.38 2.34 2.30 2.25 2.21 4.60 3.74 3.34 3.11 2.96 2.85 2.76 2.70 2.65 2.60 2.53 2.46 2.39 2.35 2.31 2.27 2.22 2.18 2.13 4.54 3.68 3.29 3.06 2.90 2.79 2.71 2.64 2.59 2.54 2.48 2.40 2.33 2.29 2.25 2.20 2.16 2.11 2.07 4.49 3.63 3.24 3.01 2.85 2.74 2.66 2.59 2.54 2.49 2.42 2.35 2.28 2.24 2.19 2.15 2.11 2.06 2.01 4.45 3.59 3.20 2.96 2.81 2.70 2.61 2.55 2.49 2.45 2.38 2.31 2.23 2.19 2.15 2.10 2.06 2.01 1.96 4.41 3.55 3.16 2.93 2.77 2.66 2.58 2.51 2.46 2.41 2.34 2.27 2.19 2.15 2.11 2.06 2.02 1.97 1.92 4.38 3.52 3.13 2.90 2.74 2.63 2.54 2.48 2.42 2.38 2.31 2.23 2.16 2.11 2.07 2.03 1.98 1.93 1.88 4.35 3.49 3.10 2.87 2.71 2.60 2.51 2.45 2.39 2.35 2.28 2.20 2.12 2.08 2.04 1.99 1.95 1.90 1.84 4.32 3.47 3.07 2.84 2.68 2.57 2.49 2.42 2.37 2.32 2.25 2.18 2.10 2.05 2.01 1.96 1.92 1.87 1.81 4.30 3.44 3.05 2.82 2.66 2.55 2.46 2.40 2.34 2.30 2.23 2.15 2.07 2.03 1.98 1.94 1.89 1.84 1.78 4.28 3.42 3.03 2.80 2.64 2.53 2.44 2.37 2.32 2.27 2.20 2.13 2.05 2.01 1.96 1.91 1.86 1.81 1.76 4.26 3.40 3.01 2.78 2.62 2.51 2.42 2.36 2.30 2.25 2.18 2.11 2.03 1.98 1.94 1.89 1.84 1.79 1.73 4.24 3.39 2.99 2.76 2.60 2.49 2.40 2.34 2.28 2.24 2.16 2.09 2.01 1.96 1.92 1.87 1.82 1.77 1.71 4.23 3.37 2.98 2.74 2.59 2.47 2.39 2.32 2.27 2.22 2.15 2.07 1.99 1.95 1.90 1.85 1.80 1.75 1.69 4.21 3.35 2.96 2.73 2.57 2.46 2.37 2.31 2.25 2.20 2.13 2.06 1.97 1.93 1.88 1.84 1.79 1.73 1.67 4.20 3.34 2.95 2.71 2.56 2.45 2.36 2.29 2.24 2.19 2.12 2.04 1.96 1.91 1.87 1.82 1.77 1.71 1.65 4.18 3.33 2.93 2.70 2.55 2.43 2.35 2.28 2.22 2.18 2.10 2.03 1.94 1.90 1.85 1.81 1.75 1.70 1.64 4.17 3.32 2.92 2.69 2.53 2.42 2.33 2.27 2.21 2.16 2.09 2.01 1.93 1.89 1.84 1.79 1.74 1.68 1.62 4.08 3.23 2.84 2.61 2.45 2.34 2.25 2.18 2.12 2.08 2.00 1.92 1.84 1.79 1.74 1.69 1.64 1.58 1.51 4.00 3.15 2.76 2.53 2.37 2.25 2.17 2.10 2.04 1.99 1.92 1.84 1.75 1.70 1.65 1.59 1.53 1.47 1.39 3.92 3.07 2.68 2.45 2.29 2.17 2.09 2.02 1.96 1.91 1.83 1.75 1.66 1.61 1.55 1.55 1.43 1.35 1.25 3.84 3.00 2.60 2.37 2.21 2.10 2.01 1.94 1.88 1.83 1.75 1.67 1.57 1.52 1.46 1.39 1.32 1.22 1.00

1

1

Table VI Percentage Points f,v ,v of the F Distribution (continued)

f0.05, ␯ , ␯ 1 2

α = 0.05

1/18/10 1:21 PM

Degrees of freedom for the denominator (v2)

JWCL232_AppA_702-730.qxd Page 714

v2

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 40 60 120 

v1 2

3

2

4

5

6

7

8

9

10

12

15

Degrees of freedom for the numerator (v1)

f0.025,v1,v2

20

24

30

40

60

120



647.8 799.5 864.2 899.6 921.8 937.1 948.2 956.7 963.3 968.6 976.7 984.9 993.1 997.2 1001 1006 1010 1014 1018 38.51 39.00 39.17 39.25 39.30 39.33 39.36 39.37 39.39 39.40 39.41 39.43 39.45 39.46 39.46 39.47 39.48 39.49 39.50 17.44 16.04 15.44 15.10 14.88 14.73 14.62 14.54 14.47 14.42 14.34 14.25 14.17 14.12 14.08 14.04 13.99 13.95 13.90 12.22 10.65 9.98 9.60 9.36 9.20 9.07 8.98 8.90 8.84 8.75 8.66 8.56 8.51 8.46 8.41 8.36 8.31 8.26 10.01 8.43 7.76 7.39 7.15 6.98 6.85 6.76 6.68 6.62 6.52 6.43 6.33 6.28 6.23 6.18 6.12 6.07 6.02 8.81 7.26 6.60 6.23 5.99 5.82 5.70 5.60 5.52 5.46 5.37 5.27 5.17 5.12 5.07 5.01 4.96 4.90 4.85 8.07 6.54 5.89 5.52 5.29 5.12 4.99 4.90 4.82 4.76 4.67 4.57 4.47 4.42 4.36 4.31 4.25 4.20 4.14 7.57 6.06 5.42 5.05 4.82 4.65 4.53 4.43 4.36 4.30 4.20 4.10 4.00 3.95 3.89 3.84 3.78 3.73 3.67 7.21 5.71 5.08 4.72 4.48 4.32 4.20 4.10 4.03 3.96 3.87 3.77 3.67 3.61 3.56 3.51 3.45 3.39 3.33 6.94 5.46 4.83 4.47 4.24 4.07 3.95 3.85 3.78 3.72 3.62 3.52 3.42 3.37 3.31 3.26 3.20 3.14 3.08 6.72 5.26 4.63 4.28 4.04 3.88 3.76 3.66 3.59 3.53 3.43 3.33 3.23 3.17 3.12 3.06 3.00 2.94 2.88 6.55 5.10 4.47 4.12 3.89 3.73 3.61 3.51 3.44 3.37 3.28 3.18 3.07 3.02 2.96 2.91 2.85 2.79 2.72 6.41 4.97 4.35 4.00 3.77 3.60 3.48 3.39 3.31 3.25 3.15 3.05 2.95 2.89 2.84 2.78 2.72 2.66 2.60 6.30 4.86 4.24 3.89 3.66 3.50 3.38 3.29 3.21 3.15 3.05 2.95 2.84 2.79 2.73 2.67 2.61 2.55 2.49 6.20 4.77 4.15 3.80 3.58 3.41 3.29 3.20 3.12 3.06 2.96 2.86 2.76 2.70 2.64 2.59 2.52 2.46 2.40 6.12 4.69 4.08 3.73 3.50 3.34 3.22 3.12 3.05 2.99 2.89 2.79 2.68 2.63 2.57 2.51 2.45 2.38 2.32 6.04 4.62 4.01 3.66 3.44 3.28 3.16 3.06 2.98 2.92 2.82 2.72 2.62 2.56 2.50 2.44 2.38 2.32 2.25 5.98 4.56 3.95 3.61 3.38 3.22 3.10 3.01 2.93 2.87 2.77 2.67 2.56 2.50 2.44 2.38 2.32 2.26 2.19 5.92 4.51 3.90 3.56 3.33 3.17 3.05 2.96 2.88 2.82 2.72 2.62 2.51 2.45 2.39 2.33 2.27 2.20 2.13 5.87 4.46 3.86 3.51 3.29 3.13 3.01 2.91 2.84 2.77 2.68 2.57 2.46 2.41 2.35 2.29 2.22 2.16 2.09 5.83 4.42 3.82 3.48 3.25 3.09 2.97 2.87 2.80 2.73 2.64 2.53 2.42 2.37 2.31 2.25 2.18 2.11 2.04 5.79 4.38 3.78 3.44 3.22 3.05 2.93 2.84 2.76 2.70 2.60 2.50 2.39 2.33 2.27 2.21 2.14 2.08 2.00 5.75 4.35 3.75 3.41 3.18 3.02 2.90 2.81 2.73 2.67 2.57 2.47 2.36 2.30 2.24 2.18 2.11 2.04 1.97 5.72 4.32 3.72 3.38 3.15 2.99 2.87 2.78 2.70 2.64 2.54 2.44 2.33 2.27 2.21 2.15 2.08 2.01 1.94 5.69 4.29 3.69 3.35 3.13 2.97 2.85 2.75 2.68 2.61 2.51 2.41 2.30 2.24 2.18 2.12 2.05 1.98 1.91 5.66 4.27 3.67 3.33 3.10 2.94 2.82 2.73 2.65 2.59 2.49 2.39 2.28 2.22 2.16 2.09 2.03 1.95 1.88 5.63 4.24 3.65 3.31 3.08 2.92 2.80 2.71 2.63 2.57 2.47 2.36 2.25 2.19 2.13 2.07 2.00 1.93 1.85 5.61 4.22 3.63 3.29 3.06 2.90 2.78 2.69 2.61 2.55 2.45 2.34 2.23 2.17 2.11 2.05 1.98 1.91 1.83 5.59 4.20 3.61 3.27 3.04 2.88 2.76 2.67 2.59 2.53 2.43 2.32 2.21 2.15 2.09 2.03 1.96 1.89 1.81 5.57 4.18 3.59 3.25 3.03 2.87 2.75 2.65 2.57 2.51 2.41 2.31 2.20 2.14 2.07 2.01 1.94 1.87 1.79 5.42 4.05 3.46 3.13 2.90 2.74 2.62 2.53 2.45 2.39 2.29 2.18 2.07 2.01 1.94 1.88 1.80 1.72 1.64 5.29 3.93 3.34 3.01 2.79 2.63 2.51 2.41 2.33 2.27 2.17 2.06 1.94 1.88 1.82 1.74 1.67 1.58 1.48 5.15 3.80 3.23 2.89 2.67 2.52 2.39 2.30 2.22 2.16 2.05 1.94 1.82 1.76 1.69 1.61 1.53 1.43 1.31 5.02 3.69 3.12 2.79 2.57 2.41 2.29 2.19 2.11 2.05 1.94 1.83 1.71 1.64 1.57 1.48 1.39 1.27 1.00

1

1

Table VI Percentage Points f,v ,v of the F Distribution (continued)

f0.025, ␯ , ␯ 1 2

α = 0.025

1/18/10 1:21 PM

Degrees of freedom for the denominator (v2)

JWCL232_AppA_702-730.qxd Page 715

v2

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 40 60 120 

v1 2

3

2

4

5

6

7

8

9

10

12

15

Degrees of freedom for the numerator (v1)

f0.01,v1,v2

20

24

30

40

60

120



4052 4999.5 5403 5625 5764 5859 5928 5982 6022 6056 6106 6157 6209 6235 6261 6287 6313 6339 6366 98.50 99.00 99.17 99.25 99.30 99.33 99.36 99.37 99.39 99.40 99.42 99.43 99.45 99.46 99.47 99.47 99.48 99.49 99.50 34.12 30.82 29.46 28.71 28.24 27.91 27.67 27.49 27.35 27.23 27.05 26.87 26.69 26.00 26.50 26.41 26.32 26.22 26.13 21.20 18.00 16.69 15.98 15.52 15.21 14.98 14.80 14.66 14.55 14.37 14.20 14.02 13.93 13.84 13.75 13.65 13.56 13.46 16.26 13.27 12.06 11.39 10.97 10.67 10.46 10.29 10.16 10.05 9.89 9.72 9.55 9.47 9.38 9.29 9.20 9.11 9.02 13.75 10.92 9.78 9.15 8.75 8.47 8.26 8.10 7.98 7.87 7.72 7.56 7.40 7.31 7.23 7.14 7.06 6.97 6.88 12.25 9.55 8.45 7.85 7.46 7.19 6.99 6.84 6.72 6.62 6.47 6.31 6.16 6.07 5.99 5.91 5.82 5.74 5.65 11.26 8.65 7.59 7.01 6.63 6.37 6.18 6.03 5.91 5.81 5.67 5.52 5.36 5.28 5.20 5.12 5.03 4.95 4.46 10.56 8.02 6.99 6.42 6.06 5.80 5.61 5.47 5.35 5.26 5.11 4.96 4.81 4.73 4.65 4.57 4.48 4.40 4.31 10.04 7.56 6.55 5.99 5.64 5.39 5.20 5.06 4.94 4.85 4.71 4.56 4.41 4.33 4.25 4.17 4.08 4.00 3.91 9.65 7.21 6.22 5.67 5.32 5.07 4.89 4.74 4.63 4.54 4.40 4.25 4.10 4.02 3.94 3.86 3.78 3.69 3.60 9.33 6.93 5.95 5.41 5.06 4.82 4.64 4.50 4.39 4.30 4.16 4.01 3.86 3.78 3.70 3.62 3.54 3.45 3.36 9.07 6.70 5.74 5.21 4.86 4.62 4.44 4.30 4.19 4.10 3.96 3.82 3.66 3.59 3.51 3.43 3.34 3.25 3.17 8.86 6.51 5.56 5.04 4.69 4.46 4.28 4.14 4.03 3.94 3.80 3.66 3.51 3.43 3.35 3.27 3.18 3.09 3.00 8.68 6.36 5.42 4.89 4.36 4.32 4.14 4.00 3.89 3.80 3.67 3.52 3.37 3.29 3.21 3.13 3.05 2.96 2.87 8.53 6.23 5.29 4.77 4.44 4.20 4.03 3.89 3.78 3.69 3.55 3.41 3.26 3.18 3.10 3.02 2.93 2.84 2.75 8.40 6.11 5.18 4.67 4.34 4.10 3.93 3.79 3.68 3.59 3.46 3.31 3.16 3.08 3.00 2.92 2.83 2.75 2.65 8.29 6.01 5.09 4.58 4.25 4.01 3.84 3.71 3.60 3.51 3.37 3.23 3.08 3.00 2.92 2.84 2.75 2.66 2.57 8.18 5.93 5.01 4.50 4.17 3.94 3.77 3.63 3.52 3.43 3.30 3.15 3.00 2.92 2.84 2.76 2.67 2.58 2.59 8.10 5.85 4.94 4.43 4.10 3.87 3.70 3.56 3.46 3.37 3.23 3.09 2.94 2.86 2.78 2.69 2.61 2.52 2.42 8.02 5.78 4.87 4.37 4.04 3.81 3.64 3.51 3.40 3.31 3.17 3.03 2.88 2.80 2.72 2.64 2.55 2.46 2.36 7.95 5.72 4.82 4.31 3.99 3.76 3.59 3.45 3.35 3.26 3.12 2.98 2.83 2.75 2.67 2.58 2.50 2.40 2.31 7.88 5.66 4.76 4.26 3.94 3.71 3.54 3.41 3.30 3.21 3.07 2.93 2.78 2.70 2.62 2.54 2.45 2.35 2.26 7.82 5.61 4.72 4.22 3.90 3.67 3.50 3.36 3.26 3.17 3.03 2.89 2.74 2.66 2.58 2.49 2.40 2.31 2.21 7.77 5.57 4.68 4.18 3.85 3.63 3.46 3.32 3.22 3.13 2.99 2.85 2.70 2.62 2.54 2.45 2.36 2.27 2.17 7.72 5.53 4.64 4.14 3.82 3.59 3.42 3.29 3.18 3.09 2.96 2.81 2.66 2.58 2.50 2.42 2.33 2.23 2.13 7.68 5.49 4.60 4.11 3.78 3.56 3.39 3.26 3.15 3.06 2.93 2.78 2.63 2.55 2.47 2.38 2.29 2.20 2.10 7.64 5.45 4.57 4.07 3.75 3.53 3.36 3.23 3.12 3.03 2.90 2.75 2.60 2.52 2.44 2.35 2.26 2.17 2.06 7.60 5.42 4.54 4.04 3.73 3.50 3.33 3.20 3.09 3.00 2.87 2.73 2.57 2.49 2.41 2.33 2.23 2.14 2.03 7.56 5.39 4.51 4.02 3.70 3.47 3.30 3.17 3.07 2.98 2.84 2.70 2.55 2.47 2.39 2.30 2.21 2.11 2.01 7.31 5.18 4.31 3.83 3.51 3.29 3.12 2.99 2.89 2.80 2.66 2.52 2.37 2.29 2.20 2.11 2.02 1.92 1.80 7.08 4.98 4.13 3.65 3.34 3.12 2.95 2.82 2.72 2.63 2.50 2.35 2.20 2.12 2.03 1.94 1.84 1.73 1.60 6.85 4.79 3.95 3.48 3.17 2.96 2.79 2.66 2.56 2.47 2.34 2.19 2.03 1.95 1.86 1.76 1.66 1.53 1.38 6.63 4.61 3.78 3.32 3.02 2.80 2.64 2.51 2.41 2.32 2.18 2.04 1.88 1.79 1.70 1.59 1.47 1.32 1.00

1

1

Table VI Percentage Points f,v ,v of the F Distribution (continued)

f0.01, ␯ , ␯ 1 2

α = 0.01

1/18/10 1:21 PM

Degrees of freedom for the denominator (v2)

JWCL232_AppA_702-730.qxd Page 716

1/18/10

1:21 PM

Page 717

717

APPENDIX A

Chart VII Operating Characteristic Curves 1.0

0.6 n = 1

0.4 2

3 4

0.2

15

20 30 40 50 75 100

0

5 6 7 8 10

Probability of accepting H0

0.8

1

0

2

3

4

5

d

(a) O.C. curves for different values of n for the two-sided normal test for a level of significance   0.05.

1.00

0.80 Probability of accepting H0

JWCL232_AppA_702-730.qxd

0.60

0.40

n

1

2

0.20

=

3 4 5

20

1

6 7 8 9 10

15

0

30 40 50 75

100

0

2

3

4

d

(b) O.C. curves for different values of n for the two-sided normal test for a level of significance   0.01. Source: Charts VIa, e, f, k, m, and q are reproduced with permission from “Operating Characteristics for the Common Statistical Tests of Significance,” by C. L. Ferris, F. E. Grubbs, and C. L. Weaver, Annals of Mathematical Statistics, June 1946. Charts VIb, c, d, g, h, i, j, l, n, o, p, and r are reproduced with permission from Engineering Statistics, 2nd Edition, by A. H. Bowker and G. J. Lieberman, Prentice-Hall, 1972.

JWCL232_AppA_702-730.qxd

1:21 PM

Page 718

APPENDIX A STATISTICAL TABLES AND CHARTS

Chart VII Operating Characteristic Curves (continued) 1.00

Probability of accepting H0

0.80

0.60

0.40 n

=

1

2

0.20 3 4

10

5 6 7

0.50

9

0.0

15 20

–0.50

30 40 50 75

100

0 –1.00

8

1.00 d

1.50

2.00

2.50

3.00

(c) O.C. curves for different values of n for the one-sided normal test for a level of significance   0.05.

1.00

0.80

0.60

n =

0.40

1

Probability of accepting H0

718

1/18/10

2

3

0.20 4 5

6

7 8 9 10

20

0.50

15

0.0

50 75

–0.50

30 40

100

0 –1.00

1.00

1.50

2.00

2.50

3.00

d

(d) O.C. curves for different values of n for the one-sided normal test for a level of significance   0.01.

1/18/10

1:21 PM

Page 719

719

APPENDIX A

Chart VII Operating Characteristic Curves (continued) 1.0

n=2

3

0.6 4

5

0.4 7

Probability of accepting H0

0.8

10 15

20

100

0

30 40 50 75

0.2

0

1

2

3

d

(e) O.C. curves for different values of n for the two-sided t-test for a level of significance   0.05.

1.00 0.90 n=3

0.80 Probability of accepting H0

JWCL232_AppA_702-730.qxd

0.70 0.60 n=

0.50 0.40

n

=

4

5

0.30 n = 7

n

0.20

=

15

20

1.0

=

=

0.8

10

n

30

0.6

n

0.4

n=

0.2

40 50

0

n=

100

0

n= 75

n=

n=

0.10

1.2

1.4

1.6 d

1.8

2.0

2.2

2.4

2.6

2.8

3.0

3.2

( f ) O.C. curves for different values of n for the two-sided t-test for a level of significance   0.01.

JWCL232_AppA_702-730.qxd

1:21 PM

Page 720

APPENDIX A STATISTICAL TABLES AND CHARTS

Chart VII Operating Characteristic Curves (continued) 1.00 0.90

0.70 0.60 0.50 0.40

n

0.30 0.20 0.10 0 –0.8 –0.6 –0.4 –0.2 0

n

7 = n 10 = n 15 n= 20 n= 0 3 n= 40 50 n= n= 5 n=7 0 n = 10

Probability of accepting H0

0.80

n =

=

=

3

4

5

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 d

(g) O.C. curves for different values of n for the one-sided t-test for a level of significance   0.05.

1.00 0.90 0.80 n=

0.70

3

0.60 n

0.50 n

0.40

4

5

= 7

= 10

0 –0.8 –0.6 –0.4 –0.2 0

n

0.10

15 n= 20 n= 30 n= 40 n = n = 50

0.20

=

=

n

0.30

n = 75 0 n = 10

Probability of accepting H0

720

1/18/10

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 d

(h) O.C. curves for different values of n for the one-sided t-test for a level of significance   0.01.

JWCL232_AppA_702-730.qxd

1/18/10

1:21 PM

Page 721

APPENDIX A

721

Chart VII Operating Characteristic Curves (continued)

2

1.00

3

n

=

0.90

4

n=

2

5

0.70

6

0.60 0.50 0.40 0.30 0.20 0.10 0

0

0.40

3

7 8 9 10 15 20 30 40 50 75 100

0.80 1.00 1.20

4 5

6

Probability of accepting H0

0.80

1.60

2.00 λ

2.40

2.80

3.20

3.60

4.00

(i) O.C. curves for different values of n for the two-sided chi-square test for a level of significance   0.05.

3

1.00

4

n=

0.90

n=

2

0.70 6

3 4

0.60

7 8 9 10 15 20 30 40 50 75 100

0.40 0.30 0.20 0.10 0

0

0.40

0.80 1.00 1.20

6

0.50

5

Probability of accepting H0

5

0.80

1.60

2.00 λ

2.40

2.80

3.20

3.60

( j) O.C. curves for different values of n for the two-sided chi-square test for a level of significance   0.01.

4.00

JWCL232_AppA_702-730.qxd

1:21 PM

Page 722

APPENDIX A STATISTICAL TABLES AND CHARTS

Chart VII Operating Characteristic Curves (continued) 1.00

Probability of accepting H0

0.80

0.60 n=

0.40

2

3

4 5

0

6

10

15 20 30 40 50 75 100

0

7

8

0.20

1.0

2.0

3.0

4.0

λ

(k) O.C. curves for different values of n for the one-sided (upper-tail) chi-square test for a level of significance   0.05.

1.00

0.80 Probability of accepting H0

722

1/18/10

0.60

0.40

100 75 50

n=

40

0.20

30

9

20

10

15

0

0

3

4

1.0

2.0

8

7

3.0

6

5

4.0

5.0

6.0

7.0

8.0

9.0

λ

(l) O.C. curves for different values of n for the one-sided (upper-tail) chi-square test for a level of significance   0.01.

1/18/10

1:21 PM

Page 723

723

APPENDIX A

Chart VII Operating Characteristic Curves (continued) 1.0

n=

0.6

5

4

3

7

6

0.4

0.2

0

0

100

40 30 50

75

15 20

8 10

Probability of accepting H0

2

0.8

0.5

01.

1.5

2.0

λ

(m) O.C. curves for different values of n for the one-sided (lower-tail) chi-square test for a level of significance   0.05.

n

=

2

1.00

3

0.80

0.60

5

4

0.40

0

0

0.20

0.40

0.60

75

100

30 40 50

15

0.20

20

9 10

8

7

6

Probability of accepting H0

JWCL232_AppA_702-730.qxd

0.80

1.00

1.20

1.40

λ

(n) O.C. curves for different values of n for the one-sided (lower-tail) chi-square test for a level of significance   0.01.

JWCL232_AppA_702-730.qxd

1:21 PM

Page 724

APPENDIX A STATISTICAL TABLES AND CHARTS

Chart VII Operating Characteristic Curves (continued) 1.00

n1 = n2 = 3

4

4

0.60 5

5

0.40

0.80 1.00

16

0.40

21

0

101

0

9

31

0.20

7 8

51

6 7 8 9 10 16 21 31 51 101

6

10

Probability of accepting H0

3

0.80

1.40

1.80

2.20

2.60

3.00

3.40

3.80 4.00

λ

(o) O.C. curves for different values of n for the two-sided F-test for a level of significance   0.05.

1.00 n1 = n = 2 3 4

0.80

5

6

0.40

0.80 1.00 1.20

21

0

51

0

31

0.20

9

10

0.40

7 8

101

3 4 5 6 7 8 9 10 16 21 31 51 101

0.60

16

Probability of accepting H0

724

1/18/10

1.60

2.00 λ

2.40

2.80

3.20

3.60

( p) O.C. curves for different values of n for the two-sided F-test for a level of significance   0.01.

4.00

1/18/10

1:21 PM

Page 725

APPENDIX A

725

Chart VII Operating Characteristic Curves (continued) 1.00 n1 = n 2 =2

3

0.60 4

5 6

0.40 7 8

10

Probability of accepting H0

0.80

20

1

15

100

0

30 40 50 75

0.20

2

3

4

λ

(q) O.C. curves for different values of n for the one-sided F-test for a level of significance   0.05.

1.00

n1 = n

2 =2

0.80 Probability of accepting H0

JWCL232_AppA_702-730.qxd

0.60

3

0.40 4 5 6

7

0.20

1.00 2.00

10

0

8 9

16 21 31 41 51 121

0

4.00

5.00

8.00 λ

10.00

12.00

14.00

16.00

(r) O.C. curves for different values of n for the one-sided F-test for a level of significance   0.01.

JWCL232_AppA_702-730.qxd

726

1/18/10

1:21 PM

Page 726

APPENDIX A STATISTICAL TABLES AND CHARTS

Table VIII Critical Values for the Sign Test r*  n 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

0.10 0.05

0.05 0.025

0 0 0 1 1 1 2 2 3 3 3 4 4 5 5 5 6 6

0 0 0 1 1 1 2 2 2 3 3 4 4 4 5 5 5

0.01 0.005

Two-sided tests One-sided tests

 n 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40

0 0 0 0 1 1 1 2 2 2 3 3 3 4 4

Table IX Critical Values for the Wilcoxon Signed-Rank Test w*  n* 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

0.10 0.05

0.05 0.025

0.02 0.01

0.01 0.005

0 2 3 5 8 10 13 17 21 25 30 35 41 47 53 60 67 75 83 91 100

0 2 3 5 8 10 13 17 21 25 29 34 40 46 52 58 65 73 81 89

0 1 3 5 7 9 12 15 19 23 27 32 37 43 49 55 62 69 76

0 1 3 5 7 9 12 15 19 23 27 32 37 42 48 54 61 68

Two-sided tests One-sided tests

* If n  25, W (or W ) is approximately normally distributed with mean n(n  1)兾4 and variance n(n  1)(2n  1)兾24.

0.10 0.05

0.05 0.025

0.01 0.005

7 7 7 8 8 9 9 10 10 10 11 11 12 12 13 13 13 14

6 6 7 7 7 8 8 9 9 9 10 10 11 11 12 12 12 13

4 5 5 6 6 6 7 7 7 8 8 9 9 9 10 10 11 11

Two-sided tests One-sided tests

JWCL232_AppA_702-730.qxd

1/18/10

1:21 PM

Page 727

APPENDIX A

727

Table X Critical Values for the Wilcoxon Rank-Sum Test w0.05 n1* n2 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28

4

5

6

7

8

9

10

11

12

13

14

15

10 11 12 13 14 15 15 16 17 18 19 20 21 21 22 23 24 25 26 27 28 28 29

17 18 20 21 22 23 24 26 27 28 29 31 32 33 34 35 37 38 39 40 42

26 27 29 31 32 34 35 37 38 40 42 43 45 46 48 50 51 53 55

36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 66 68

49 51 53 55 58 60 63 65 67 70 72 74 77 79 82

63 65 68 71 73 76 79 82 84 87 90 93 95

78 81 85 88 91 94 97 100 103 107 110

96 99 103 106 110 114 117 121 124

115 119 123 127 131 135 139

137 141 145 150 154

160 164 169

185

*For n1 and n2  8, W1 is approximately normally distributed with mean 12 n1 1 n1  n2  12 and variance n1n2(n1  n2  1)兾12.

Table X Critical Values for the Wilcoxon Rank-Sum Test (continued) w0.01 n1 n2 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28

4 10 10 11 11 12 12 13 14 14 15 15 16 16 17 18 18 19 19 20 20 21

5

6

7

8

9

10

11

12

13

14

15

15 16 17 17 18 19 20 21 22 22 23 24 25 26 27 28 29 29 30 31 32

23 24 25 26 27 28 30 31 32 33 34 36 37 38 39 40 42 43 44

32 34 35 37 38 40 41 43 44 46 47 49 50 52 53 55 57

43 45 47 49 51 53 54 56 58 60 62 64 66 68 70

56 58 61 63 65 67 70 72 74 76 78 81 83

71 74 76 79 81 84 86 89 92 94 97

87 90 93 96 99 102 105 108 111

106 109 112 115 119 122 125

125 129 133 137 140

147 151 155

171

JWCL232_AppA_702-730.qxd

728

1/18/10

1:21 PM

Page 728

APPENDIX A STATISTICAL TABLES AND CHARTS

Table XI Factors for Constructing Variables Control Charts Factor for Control Limits X Chart

R Chart

S Chart

n*

A1

A2

d2

D3

D4

c4

n

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

3.760 2.394 1.880 1.596 1.410 1.277 1.175 1.094 1.028 .973 .925 .884 .848 .816 .788 .762 .738 .717 .697 .679 .662 .647 .632 .619

1.880 1.023 .729 .577 .483 .419 .373 .337 .308 .285 .266 .249 .235 .223 .212 .203 .194 .187 .180 .173 .167 .162 .157 .153

1.128 1.693 2.059 2.326 2.534 2.704 2.847 2.970 3.078 3.173 3.258 3.336 3.407 3.472 3.532 3.588 3.640 3.689 3.735 3.778 3.819 3.858 3.895 3.931

0 0 0 0 0 .076 .136 .184 .223 .256 .284 .308 .329 .348 .364 .379 .392 .404 .414 .425 .434 .443 .452 .459

3.267 2.575 2.282 2.115 2.004 1.924 1.864 1.816 1.777 1.744 1.716 1.692 1.671 1.652 1.636 1.621 1.608 1.596 1.586 1.575 1.566 1.557 1.548 1.541

0.7979 0.8862 0.9213 0.9400 0.9515 0.9594 0.9650 0.9693 0.9727 0.9754 0.9776 0.9794 0.9810 0.9823 0.9835 0.9845 0.9854 0.9862 0.9869 0.9876 0.9882 0.9887 0.9892 0.9896

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

*n  25: A1  3 1n where n  number of observations in sample.

JWCL232_AppA_702-730.qxd

1/18/10

1:21 PM

Page 729

APPENDIX A

729

Table XII Factors for Tolerance Intervals Values of k for Two-Sided Intervals Confidence Level 0.95

0.90 Sample Size

0.90

0.95

0.99

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 30 40 50 60 70 80 90 100

15.978 5.847 4.166 3.949 3.131 2.902 2.743 2.626 2.535 2.463 2.404 2.355 2.314 2.278 2.246 2.219 2.194 2.172 2.152 2.135 2.118 2.103 2.089 2.077 2.025 1.959 1.916 1.887 1.865 1.848 1.834 1.822

18.800 6.919 4.943 4.152 3.723 3.452 3.264 3.125 3.018 2.933 2.863 2.805 2.756 2.713 2.676 2.643 2.614 2.588 2.564 2.543 2.524 2.506 2.489 2.474 2.413 2.334 2.284 2.248 2.222 2.202 2.185 2.172

24.167 8.974 6.440 5.423 4.870 4.521 4.278 4.098 3.959 3.849 3.758 3.682 3.618 3.562 3.514 3.471 3.433 3.399 3.368 3.340 3.315 3.292 3.270 3.251 3.170 3.066 3.001 2.955 2.920 2.894 2.872 2.854

Probability of Coverage 0.90 0.95 0.99 32.019 8.380 5.369 4.275 3.712 3.369 3.136 2.967 2.839 2.737 2.655 2.587 2.529 2.480 2.437 2.400 2.366 2.337 2.310 2.286 2.264 2.244 2.225 2.208 2.140 2.052 1.996 1.958 1.929 1.907 1.889 1.874

37.674 9.916 6.370 5.079 4.414 4.007 3.732 3.532 3.379 3.259 3.162 3.081 3.012 2.954 2.903 2.858 2.819 2.784 2.752 2.723 2.697 2.673 2.651 2.631 2.529 2.445 2.379 2.333 2.299 2.272 2.251 2.233

48.430 12.861 8.299 6.634 5.775 5.248 4.891 4.631 4.433 4.277 4.150 4.044 3.955 3.878 3.812 3.754 3.702 3.656 3.615 3.577 3.543 3.512 3.483 3.457 3.350 3.213 3.126 3.066 3.021 2.986 2.958 2.934

0.99 0.90

0.95

160.193 18.930 9.398 6.612 5.337 4.613 4.147 3.822 3.582 3.397 3.250 3.130 3.029 2.945 2.872 2.808 2.753 2.703 2.659 2.620 2.584 2.551 2.522 2.494 2.385 2.247 2.162 2.103 2.060 2.026 1.999 1.977

188.491 22.401 11.150 7.855 6.345 5.488 4.936 4.550 4.265 4.045 3.870 3.727 3.608 3.507 3.421 3.345 3.279 3.221 3.168 3.121 3.078 3.040 3.004 2.972 2.841 2.677 2.576 2.506 2.454 2.414 2.382 2.355

0.99 242.300 29.055 14.527 10.260 8.301 7.187 6.468 5.966 5.594 5.308 5.079 4.893 4.737 4.605 4.492 4.393 4.307 4.230 4.161 4.100 4.044 3.993 3.947 3.904 3.733 3.518 3.385 3.293 3.225 3.173 3.130 3.096

JWCL232_AppA_702-730.qxd

730

1/18/10

1:21 PM

Page 730

APPENDIX A STATISTICAL TABLES AND CHARTS

Table XII Factors for Tolerance Intervals (continued) Values of k for One-Sided Intervals Confidence Level 0.90

0.95

Sample Size

0.90

0.95

0.99

2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 30 40 50 60 70 80 90 100

10.253 4.258 3.188 2.742 2.494 2.333 2.219 2.133 2.066 2.011 1.966 1.928 1.895 1.867 1.842 1.819 1.800 1.782 1.765 1.750 1.737 1.724 1.712 1.702 1.657 1.598 1.559 1.532 1.511 1.495 1.481 1.470

13.090 5.311 3.957 3.400 3.092 2.894 2.754 2.650 2.568 2.503 2.448 2.402 2.363 2.329 2.299 2.272 2.249 2.227 2.028 2.190 2.174 2.159 2.145 2.132 2.080 2.010 1.965 1.933 1.909 1.890 1.874 1.861

18.500 7.340 5.438 4.666 4.243 3.972 3.783 3.641 3.532 3.443 3.371 3.309 3.257 3.212 3.172 3.137 3.105 3.077 3.052 3.028 3.007 2.987 2.969 2.952 2.884 2.793 2.735 2.694 2.662 2.638 2.618 2.601

0.99

Probability of Coverage 0.90 0.95 0.99 20.581 6.155 4.162 3.407 3.006 2.755 2.582 2.454 2.355 2.275 2.210 2.155 2.109 2.068 2.033 2.002 1.974 1.949 1.926 1.905 1.886 1.869 1.853 1.838 1.777 1.697 1.646 1.609 1.581 1.559 1.542 1.527

26.260 7.656 5.144 4.203 3.708 3.399 3.187 3.031 2.911 2.815 2.736 2.671 2.614 2.566 2.524 2.486 2.453 2.423 2.396 2.371 2.349 2.328 2.309 2.292 2.220 2.125 2.065 2.022 1.990 1.964 1.944 1.927

37.094 10.553 7.042 5.741 5.062 4.642 4.354 4.143 3.981 3.852 3.747 3.659 3.585 3.520 3.464 3.414 3.370 3.331 3.295 3.263 3.233 3.206 3.181 3.158 3.064 2.941 2.862 2.807 2.765 2.733 2.706 2.684

0.90

0.95

0.99

103.029 13.995 7.380 5.362 4.411 3.859 3.497 3.240 3.048 2.898 2.777 2.677 2.593 2.521 2.459 2.405 2.357 2.314 2.276 2.241 2.209 2.180 2.154 2.129 2.030 1.902 1.821 1.764 1.722 1.688 1.661 1.639

131.426 17.370 9.083 6.578 5.406 4.728 4.285 3.972 3.738 3.556 3.410 3.290 3.189 3.102 3.028 2.963 2.905 2.854 2.808 2.766 2.729 2.694 2.662 2.633 2.515 2.364 2.269 2.202 2.153 2.114 2.082 2.056

185.617 23.896 12.387 8.939 7.335 6.412 5.812 5.389 5.074 4.829 4.633 4.472 4.337 4.222 4.123 4.037 3.960 3.892 3.832 3.777 3.727 3.681 3.640 3.601 3.447 3.249 3.125 3.038 2.974 2.924 2.883 2.850

JWCL232_AppB_731-748.qxd

1/23/10

10:04 AM

Page 731

Appendix B Answers to Selected Exercises

CHAPTER 2 Section 2-1 2-1. Let a, b denote a part above, below the specification, respectively S  {aaa, aab, aba, abb, baa, bab, bba, bbb} 2-3. Let a denote an acceptable power supply Let f, m, c denote a supply with a functional, minor, or cosmetic error, respectively. S  {a, f, m, c} 2-5. Sequences of y or n of length 24 with 224 outcomes 2-7. S is the sample space of 100 possible two digit integers. 2-9. S  {0, 1, 2, . . . , 1E09} in ppb 2-11. S  {1.0, 1.1, 1.2, . . . , 14.0} 2-13. S  {0, 1, 2, . . . ,} in milliseconds 2-17. c  connect, b  busy, S  {c, bc, bbc, bbbc, bbbbc, . . .} 2-21. (a) S  nonnegative integers from 0 to the largest integer that can be displayed by the scale S  {0, 1, 2, 3, . . .}

(b) S (c) {12, 13, 14, 15} (d) {0, 1, 2, . . . , 11} (e) S (f) {0, 1, 2, . . . , 7} (g)  (h)  (i) {8, 9, 10, . . .} 2-23. Let d denoted a distorted bit and let o denote a bit that is not distorted. dddd, dodd, oddd, oodd, dddo, dodo, oddo, oodo, (a) S  μ ∂ ddod, dood, odod, oood, ddoo, dooo, odoo, oooo (b) No, for example A1 ¨ A2  {dddd, dddo, ddod, ddoo} dddd, dodd, dddo, dodo, (c) A1  μ ∂ ddod, dood, ddoo, dooo oddd, oodd, oddo, oodo, (d) A¿1  μ ∂ odod, oood, odoo, oooo (e) A1 ¨ A2 ¨ A3 ¨ A4  {dddd}

2-25.

2-27. 2-29.

2-31.

(f) 1A1 ¨ A2 2 ´ 1A3 ¨ A4 2  {dddd, dodd, dddo, oddd, ddod, oodd, ddoo} Let P denote being positive and let N denote being negative. The sample space is {PPP, PPN, PNP, NPP, PNN, NPN, NNP, NNN}. (a) A  {PPP} (b) B  {NNN} (c) A ¨ B   (d) A ´ B  {PPP, NNN} (a) A¿ ¨ B  10, B¿  10, A ´ B  92 (a) A¿  5x 0 x  72.56 (b) B¿  5x 0 x  52.56 (c) A ¨ B  5x 0 52.5  x  72.5} (d) A ´ B  5x 0 x  06 Let g denote a good board, m a board with minor defects, and j a board with major defects. (a) S  {gg, gm, gj, mg, mm, mj, jg, jm, jj} (b) S  {gg, gm, gj, mg, mm, mj, jg, jm}

731

JWCL232_AppB_731-748.qxd

732 2-35. 2-37. 2-39. 2-41.

2-43. 2-45. 2-47. 2-49.

1/23/10

10:04 AM

Page 732

APPENDIX B ANSWERS TO SELECTED EXERCISES

120 144 14,400 (a) 416,965,528 (b) 113,588,800 (c) 130,721,752 (a) 21 (b) 2520 (c) 720 (a) 1000 (b) 160 (c) 720 (a) 0.416 (b) 0.712 (c) 0.206 0.0082

Section 2-2 2-51. 900 2-53. (a) 673 (b) 1672 (c) 6915 (d) 8399 (b) 1578 2-55. (a) 0.4 (b) 0.8 (c) 0.6 (d) 1 (e) 0.2 2-57. (a) 110 (b) 510 2-59. (a) S  {1, 2, 3, 4, 5, 6, 7, 8} (b) 28 (c) 68 2-61. (a) 0.83 (b) 0.85 2-63. (1103)*(1263)  5.7  10 8 2-65. (a) 4 4  3 4  3  3  52 (b) 3652 (c) No 2-67. (a) 0.30 (b) 0.77 (c) 0.70 (d) 0.22 (e) 0.85 (f ) 0.92 2-71. 0.9889 2-73. (a) 0.0792 (b) 0.1969 (c) 0.8142 (d) 0.9889 (e) 0.1858 Section 2-3 2-75. (a) 0.9 (b) 0 (c) 0 (d) 0 (e) 0.1 2-77. (a) 0.70 (b) 0.95 (c) No 2-79. (a) 350370 (b) 362370 (c) 358370 (d) 345370 2-81. (a) 13130 (b) 0.90, No 2-83. (a) 0.7255 (b) 0.8235 (c) 0.7255 2-85. (a) 0.2264 (b) 0.9680 (c) 0.9891 Section 2-4 2-87. (a) 86100 (b) 79100 (c) 7079 (d) 7086 2-89. (a) 0.903 (b) 0.591 2-91. (a) 12100 (b) 1228 (c) 34122 2-93. (a) 0.5625 (b) 0.1918 (c) 0.3333 2-95. (a) 20100 (b) 1999 (c) 0.038 (d) 0.2

2-97. (a) (c) 2-99. No 2-101. (a) (c) 2-103. (a)

0.02 (b) 0.000458 0.9547 0.6087 (b) 0.3913 0.5 (d) 0.5 0.0987 (b) 0.0650

Section 2-5 2-105. (a) 0.2 (b) 2-107. 0.014 2-109. 0.028 2-111. (a) 0.2376 2-113. (a) 0.2 (b) 2-117. (a) 0.0109 (c) 0.9891 2-119. (a) 0.0792 (c) 0.9208 2-121. 0.2

0.3

(b) 0.0078 0.2 (b) 0.2264 (d) 0.1945 (b) 0.8142 (d) 0.8031

Section 2-6 2-123. independent 2-125. (a) not independent. (b) yes 2-127. (a) not independent. (b) 0.733 2-129. (a) 0.59 (b) 0.328 (c) 0.41 2-131. (a) 0.00307 (b) 0.04096 2-133. (a) 0.01 (b) 0.49 (c) 0.09 2-135. (a) 0.00003 (b) 0.00024 (c) 0.00107 2-137. 0.9702 2-139. not independent. 2-141. independent. Section 2-7 2-143. 0.89 2-145. (a) 0.97638 (b) 0.20755 2-147. (a) 0.615 (b) 0.618 (c) 0.052 2-149. (a) 0.9847 (b) 0.1184 2-151. 0.2540 2-153. 0.5 Supplemental Exercises 2-155. 0.014 2-157. (a) 0.82 (b) 0.90 (c) 0.18 (d) 0.80 (e) 0.92 (f ) 0.98 2-161. (a) 0.2 (b) 0.202 (c) 0.638 (d) 0.2 2-163. (a) 0.03 (b) 0.97 (c) 0.40 (d) 0.05 (e) 0.012 (f ) 0.018 (g) 0.0605 2-165. (a) 0.18143 (b) 0.005976 (c) 0.86494 2-167. 0.000008 2-169. (a) 50 (b) 37 (c) 93

2-171. S  5A, A¿D1, A¿D2, A¿D3, A¿D4, A¿D5 6 2-173. (a) 0.19 (b) 0.15 (c) 0.99 (d) 0.80 (e) 0.158 2-175. (a) No (b) No (c) 40240 (d) 200240 (e) 234240 (f ) 1 2-177. (a) 0.282 (b) 0.718 2-179. 0.996 2-181. (a) 0.0037 (b) 0.8108 2-183. (a) 0.0778 (b) 0.00108 (c) 0.947 2-185. (a) 0.9764 (b) 0.3159 2-187. (a) 0.207 (b) 0.625 2-189. (a) 0.453 (b) 0.262 (c) 0.881 (d) 0.547 (e) 0.783 (f ) 0.687 2-191. 1.58  10 7 2-193. (a) 0.67336 (b) 2.646  10 8 (c) 0.99973 2-195. (a) 367 (b) 70(266) (c) 100(265) 2-197. (a) 0.994, 0.995 (b) 0.99, 0.985 (c) 0.998, 0.9975 Mind-Expanding Exercises 2-199. (a) n  3 (b) n  3 2-201. 0.306, 0.694

CHAPTER 3 Section 3-1 3-1. {0, 1, 2, . . . , 1000} 3-3. {0, 1, 2, . . . , 99999} 3-5. {1, 2, . . . , 491} 3-7. {0, 1, 2, . . .} 3-9. {0, 1, 2, . . . , 15} 3-11. {0, 1, 2, . . . , 10000} 3-13. {0, 1, 2, . . . , 40000} Section 3-2 3-15. (a) 1 (b) 78 (c) 34 (d) 12 3-17. (a) 925 (b) 425 (c) 1225 (d) 1 3-19. f(0)  0.033, f(1)  0.364, f(2)  0.603 3-21. P(X  0)  0.008, P(X  1)  0.096, P(X  2)  0.384, P(X  3)  0.512 3-23. P(X  50)  0.5, P(X  25)  0.3, P(X  10)  0.2

JWCL232_AppB_731-748.qxd

1/23/10

10:04 AM

Page 733

APPENDIX B

3-25. P(X  15)  0.6, P(X  5)  0.3, P(X  0.5) = 0.1 3-27. P(X  0)  0.00001, P(X  1)  0.00167, P(X  2)  0.07663, P(X  3)  0.92169

0, x  266 0.24, 266  x  271 3-43. F1x2  μ ∂ 0.54, 271  x  274 1, 274  x

Section 3-3 3-29. X  waiting time (hours) 0.038, x  1 0.102, x  2 0.172, x  3 0.204, x  4 0.174, x  5 f (x)  0.124, x  6 0.080, x  7 0.036, x  8 0.028, x  9 0.022, x  10 0.020, x  15 3-31. X  Non-failed well depth P(X  255)  (1515 1343)/ 7726  0.370 P(X  218)  26/ 7726  0.003 P(X  317)  3290/ 7726  0.426 P(X  231)  349/ 7726  0.045 P(X  267)  (280 887)/ 7726  0.151 P(X  217)  36/ 7726  0.005 3-33. (a) 78 (b) 1 (c) 34 (d) 38 0, x0 0.008, 0  x  1 3-35. F1x2  0.104, 1  x  2 0.488, 2  x  3 1, 3x where fx (0)  0.008, fx (1)  0.096, fx (2)  0.384, fx (3)  0.512 0, x  10 0.2, 10  x  25 ∂ 3-37. F1x2  μ 0.5, 25  x  50 1, 50  x where P(X = 50 million) = 0.5, P(X = 25 million) = 0.3, P(X = 10 million) = 0.2 3-39. (a) 1 (b) 0.5 (c) 0.5 (d) 0.5 3-41. (a) 1 (b) 0.75 (c) 0.25 (d) 0.25 (e) 0 (f) 0

3-45. F1x2 

u

u

v

u

0, 0.05, 0.30, 0.65, 0.85, 1

x  1.5 1.5  x  3 3  x  4.5 4.5  x  5 5x7 7x

v

Section 3-4 3-47.   2, 2  2 3-49.   0, 2  1.5 3-51.   2.8, 2  1.36 3-53.   1.57, 2  0.311 3-55. 24 3-57.   0.0004, 2  0.00039996 3-59. (a)   18.694, 2  735.9644,  27.1287 (b)   37.172, 2  2947.996,  54.2955 3-61. E(X )  4.808, V(X )  6.147 3-63. E(X )  281.83, V(X )  976.24 Section 3-5 3-65.   2, 2  0.667 3-67.   3.5, 2  1.25 3-69. (a)   687.5, 2  56.25 (b)   87.5, 2  56.25 3-71. E(X )  4.5, E(Y )  22.5, Y  14.36 3-73.   7,  1.414 Section 3-6 3-77. (a) 0.9298 (b) 0 (c) 0.0112 (d) 0.0016 3-79. (a) 2.40  10 8 (b) 0.9999 (c) 9.91  10 18 (d) 1.138  10 4 3-81. (a) 0 (b) 10 3-85. (a) 0.215 (b) 0.9999 (c) 4 3-87. (a) 0.410 (b) 0.218 (c) 0.37 3-89. (a) 1 (b) 0.999997 (c) E(X )  12.244, V(X )  2.179 3-91. (a) Binomial, p  104369, n  1E09 (b) 0 (c) E(X )  4593.9, V(X )  4593.9

3-93. (a) 0.9961 3-95. (a) 0.142 (c) 0.963 3-97. (a) 0.009 (c) 0.972

733

(b) 0.9886 (b) 0.322 (b) 0.382 (d) 3

Section 3-7 3-99. (a) 0.5 (b) 0.0625 (c) 0.0039 (d) 0.75 (e) 0.25 3-101. (a) 5 (b) 5 3-103. (a) 0.0064 (b) 0.9984 (c) 0.008 3-105. (a) 0.0167 (b) 0.9039 (c) 50 3-107. (a) 0.13 (b) 0.098 (c) 7.69  8 3-109. (a) 3.91  10 19 (b) 200 (c) 2.56  1018 3-111. (a) 3  108 (b) 3  1016 3-115. (a) 10 (b) 0.039 (c) 0.039 (d) 0.271 Section 3-8 3-117. (a) 0.4191 (b) 0 (c) 0.001236 (d) E(X )  0.8, V(X )  0.6206 3-121. (a) 0.1201 (b) 0.8523 3-123. (a) 0.087 (b) 0.9934 (c) 0.297 (d) 0.9998 3-125. (a) 0.7069 (b) 0.0607 (c) 0.2811. 3-127. (a) 0.0041 (b) 0.3091 (c) 0.0165 Section 3-9 3-129. (a) 0.0183 (b) 0.2381 (c) 0.1954 (d) 0.0298 3-131. E(X )  V(X )  2.996. 3-133. (a) 0.264 (b) 48 3-135. (a) 0.4566 (b) 0.047 3-137. (a) 0.2 (b) 99.89% 3-139. (a) 0.6065 (b) 0.0067 (c) P(W  0)  0.0067, P(W  1)  0.0437, P(W  1)  0.0504 3-141. (a) 0.026 (b) 0.287 (c) 0.868 Supplemental Exercises 3-143. E(X )  14, V(X )  0.0104 3-145. (a) n  50, p  0.1 (b) 0.112 (c) P(X  49)  4.51  10 48 3-147. (a) 0.000224 (b) 0.2256 (c) 0.4189

JWCL232_AppB_731-748.qxd

734

1/23/10

10:04 AM

Page 734

APPENDIX B ANSWERS TO SELECTED EXERCISES

3-149. (a) 0.1024 (b) 0.1074 3-151. (a) 3000 (b) 1731.18 3-153. (a) P(X  0)  0.0498 (b) 0.5768 (c) P(X  x)  0.9, x  5 (d) 2   6. Not appropriate. 3-155. (a) 0.1877 (b) 0.4148 (c) 15 3-157. (a) 0.0110 (b) 83 3-159. 40000 3-161. 0.1330 3-163. (a) 500 (b) 222.49 3-165. 0.1 3-167. fX (0)  0.16, fX (1)  0.19, fX (2)  0.20, fX (3)  0.31, fX (4)  0.14 3-169. fx (2)  0.2, fx(5.7)  0.3, fx (6.5)  0.3, fx(8.5)  0.2 3-171. (a) 0.0433 (b) 3.58 3-173. (a) fX (0)  0.2357, fX (1)  0.3971, fX (2)  0.2647, fX (3)  0.0873, fX (4)  0.01424, fX (5)  0.00092 (b) fX (0)  0.0546, fX (1)  0.1866, fX (2)  0.2837, fX (3)  0.2528, fX (4)  0.1463, fX (5)  0.0574, fX (6)  0.0155, fX (7)  0.0028, fX (8)  0.0003, fX (9)  0.0000, fX (10)  0.0000 3-175. 37.8 seconds Mind-Expanding Exercises 3-181. (a) 131 (b) 123

CHAPTER 4 Section 4-2 4-1. (a) 0.3679 (b) 0.2858 (c) 0 (d) 0.9817 (e) 0.0498 4-3. (a) 0.5 (b) 0.1464 (c) 0.7072 (d) 0.8536 (e) 1.12 radians 4-5. (a) 0.4375 (b) 0.7969 (c) 0.5625 (d) 0.7031 (e) 0.5 4-7. (a) 0.5 (b) 0.4375 (c) 0.125 (d) 0 (e) 1 (f ) 0.9655 4-9. (a) 0.5 (b) 49.8 4-11. (a) 0.1

Section 4-3 4-13. (a) 0.56 (b) 0.7 (c) 0 (d) 0 4-15. F(x)  0 for x  0; 1 e x for x  0  4-17. F 1x2  0 for x  , 2 0.5 sin x  

0.5 for  x  , 2 2  1 for x  2 4-19. (a) 0.56 (b) 0.7 (c) 0 (d) 0 4-21. 0.2 4-23. F(x)  0 for x  0; 0.25x2 for 0  x  2; 1 for 2  x 4-25. f (x)  0.2 for 0  x  4; f(x)  0.04 for 4  x  9 Section 4-4 4-27. E(X )  2, V(X )  43 4-29. E(X )  0, V(X )  0.6 4-31. E(X )  4, V(X )  3.2 4-33. E(X )  2 4-35. (a) E(X )  109.39, V(X ) 33.19 (b) 54.70 4-37. (a) E(X )  5.1, V(X )  0.01 (b) 0.3679 Section 4-5 4-39. (a) E(X)  0, V(X)  0.577 (b) 0.90 (c) F(x)  0 for x  1; 0.5x 0.5 for 1  x  1; 1 for 1  x 4-41. (a) F(x)  0 for x  0.95; 10x 9.5 for 0.95  x  1.05; 1 for 1.05  x (b) 0.3 (c) 0.96 (d) E(X)  1.00, V(X)  0.00083 4-43. (a) F(x)  0 for x  0.2050; 100x 20.50 for 0.2050  x  0.2150; 1 for 0.21.50  x (b) 0.25 (c) 0.2140 (d) E(X)  0.2100, V(X)  8.33  10 6 4-45. (a) F(X )  x90 for 0  x  90 (b) E(X)  45, V(X )  675 (c) 1/ 3 (d) 0.333 4-47. (a) X  34.64 (b) 1/ 3 (c) 1/ 2 Section 4-6 4-49. (a) 0.90658 (b) 0.99865 (c) 0.07353 (d) 0.98422 (e) 0.95116

4-51. (a) 0.90 (b) 0.5 (c) 1.28 (d) 1.28 (e) 1.33 4-53. (a) 0.93319 (b) 0.69146 (c) 0.9545 (d) 0.00132 (e) 0.15866 4-55. (a) 0.93319 (b) 0.89435 (c) 0.38292 (d) 0.80128 (e) 0.54674 4-57. (a) 0.99379 (b) 0.13591 (c) 5835 4-59. (a) 0.0228 (b) 0.019 (c) 152.028 (d) small (less than 5%) 4-61. (a) 0.0082 (b) 0.72109 (c) 0.564 4-63. (a) 12.309 (b) 12.1545 4-65. (a) 0.00621 (b) 0.308538 (c) 133.33 4-67. (a) 0.1587 (b) 1.3936 (c) 0.9545 4-69. (a) 0.00043 (b) 6016 (c) 18 4-71. (a) 0.02275 (b) 0.324 (c) 11.455 4-73. [23.5, 24.5], no effect from stdev 4-75. (a) 0.0248 (b) 0.1501 (c) 92.0213 Section 4-7 4-77. (a) 0.0853 (b) 0.8293 (c) 0.0575 4-79. (a) 0.1446 (b) 0.4761 (c) 0.3823 4-81. (a) 0.2743 (b) 0.8413 4-83. 0.022 4-85. 0.5 4-87. (a) 0 (b) 0.156 (c) 13,300 (d) 8.3 days/year (e) 0.0052 4-89. (a) 0.012 (b) 0.9732 (c) 536.78 Section 4-8 4-91. (a) 0.3679 (b) 0.1353 (c) 0.9502 (d) 0.95, x  29.96 4-93. (a) 0.3679 (b) 0.2835 (c) 0.1170 4-95. (a) 0.1353 (b) 0.4866 (c) 0.2031 (d) 34.54 4-97. (a) 0.0498 (b) 0.8775 4-99. (a) 0.0025 (b) 0.6321 (c) 23.03 (d) same as part (c) (e) 6.93

JWCL232_AppB_731-748.qxd

1/23/10

10:04 AM

Page 735

APPENDIX B

4-101. (a) (c) 4-103. (a) (c) (e) 4-105. (a) (c) (e) 4-107. (a) (c) (d) 4-109. (a) (c)

15.625 (b) 0.1629 3  10 6 0.2212 (b) 0.2865 0.2212 (d) 0.9179 0.2337 0.3528 (b) 0.04979 46.05 (d) 6.14  10 6 e 12 (f ) same 0.3679 (b) 0.1353 0.0498 does not depend on  0.435 (b) 0.135 0.369 (d) 0.865

Section 4-9 4-111. (a) 120 (b) 1.32934 (c) 11.6317 4-113. (a) Erlang  5 calls/min, r  10 (b) E(X)  2, V(X)  0.4 (c) 0.2 minute (d) 0.1755 (e) 0.2643 4-115. (a) 50000 (b) 0.6767 4-117. (a) 5  105 (b) V(X)  5  1010,  223607 (c) 0.0803 4-119. (a) 0.1429 (b) 0.1847 4-123. (a) 1.54 (b) 0.632 Section 4-10 4-125. E(X)  12000, V(X)  3.61  1010 4-127. 1000 4-129. (a) 803.68 hours (b) 85319.64 (c) 0.1576 4-131. (a) 443.11 (b) 53650.5 (c) 0.2212 4-135. (a) 0.5698 (b) 0.1850 (c) 0.4724 4-137. (a) 0.0468 (b) 0.1388 Section 4-11 4-139. (a) 0.0016 (b) 0.0029 (c) E(X)  12.1825, V(X)  1202455.87 4-141. (a) 0.03593 (b) 1.65 (c) 2.7183 12.6965 4-143. (a)   8.4056, 2  1.6094 (b) 0.2643 (c) 881.65 4-147. (a) E(X)  4.855, V(X)  4.090 (b) 0.9263 (c) 0.008 Section 4-12 4-149. (a) 0.0313 (b) 0.4559 (c) E(X)  0.7143, V(X)  0.0454

4-151. (a) Mode  0.8333, E(X)  0.6818, V(X)  0.0402 (b) Mode  0.6316, E(X)  0.6154, V(X)  0.0137 4-153. 0.0272 Supplemental Exercises 4-155. (a) 0.99379 (b) 0.621% 4-157. (a) 0.15866 (b) 90.0 (c) 0.9973 (d) (0.9973)10 (e) 9.973 4-159. (a) 0.0217 (b) 0.9566 (c) 229.5 4-161. 0.8488 4-163. (a) 620.4 (b) 105154.9 (c) 0.4559 4-165. (a) 0.0625 (b) 0.75 (c) 0.5 (d) F(x)  0 for x  2; x4 x 1 for 2  x  4; 1 for 4  x (e) E(X)  103, V(X)  0.2222 4-167. (a) 0.3935 (b) 0.9933 4-169. (a)   3.43, 2  0.96 (b) 0.946301 4-171. (a) 0.6915 (b) 0.683 (c) 1.86 4-173. (a) 0.0062 (b) 0.012 (c) 5.33 4-175. 0.0008 to 0.0032 4-177. (a) 0.5633 (b) 737.5 4-179. (a) 0.9906 (b) 0.8790 Mind-Expanding Exercises 4-183. (a) k 2 (b) k 2 k( m)2

5-5.

5-7.

5-9.

5-11. 5-13.

5-15.

CHAPTER 5 Section 5-1 5-1. (a) 38 (b) 58 (c) 38 (d) 18 (e) V(X)  0.4961 V(Y)  1.8594 (f ) f (1)  14, f (1.5)  38, f (2.5)  14, f(3)  18 (g) f(2)  13, f(3)  23 (h) 1 (i) 2 13 ( j) Not independent 5-3. (a) 38 (b) 38 (c) 78 (d) 58 (e) V(X)  0.4219 V(Y)  1.6875 (f ) f( 1)  18, f( 0.5)  14, f(0.5)  12, f(1)  18 (g) 1 (h) 1 (i) 0.5 (j) Not independent

5-17.

5-19. 5-21.

735

(b) f X (0)  0.970299, f X (1)  0.029403, f X (2)  0.000297, f X (3)  0.000001 (c) 0.03 (d) f(0)  0.920824, f (1)  0.077543, f (2)  0.001632 (e) 0.080807 (g) Not independent (b) f(0)  2.40  10 9, f(1)  1.36  10 6, f(2)  2.899  10 4, f(3)  0.0274, f (4)  0.972 (c) 3.972 (d) equals f( y) (e) 3.988 (f ) 0.0120 (g) Independent (b) f X (0)  0.2511, f X (1)  0.0405, f X (2)  0.0063, f X (3)  0.0009, f X (4)  0.0001 (c) 0.0562 (d) f Y|3(0)  23, f Y|3(1)  13, f Y|3(2)  f Y|3(3)  f Y|3(4)  0 (e) 0.0003 (f ) 0.0741 (g) Not independent (c) 0.308 (d) 5.7 (a) 0.4444 (b) 0.6944 (c) 0.5833 (d) 0.3733 (e) 2 (f ) 0 (g) fX(x)  2x9; 0  x  3 (h) f Y|1.5(y)  2y9; 0  y  3 (i) 2 (j) 49 (k) f X|2(x)  2x9; 0  x  3 (a) 181 (b) 527 (c) 0.790 (d) 1681 (e) 125 (f ) 85 (g) f (x)  4x381; 0  x  3 (h) f Y|X1(y)  2y; 0  y  1 ( i) 1 ( j) 0 (k) f X |Y2(x)  2x9; 0  x  3 (a) 0.9879 (b) 0.0067 (c) 0.000308 (d) 0.9939 (e) 0.04 (f) 815 (g) f (x)  5e 5x; x  0 (h) f Y|X1(y)  3e3 3y; 1  y (i) 43 (j) 0.9502, f Y (2)  15e 62; 0y (k) fX|Y2(y)  2e 2x; 0  x  2 (a) 130 (b) 112 (c) 199 (d) 9745 (g) 1 (h) 0.25 (a) P(X  5, Y  5)  0.0439, P(X  10, Y  10)  0.0019 (b) 0.0655

JWCL232_AppB_731-748.qxd

736

1/23/10

10:04 AM

Page 736

APPENDIX B ANSWERS TO SELECTED EXERCISES

5-23. (a) 0.25 (b) 0.0625 (c) 1 (d) 1 (e) 23 (f ) 0.25 (g) 0.0625 (h) fX|YZ(y, z)  2x; 0  x  1 (i) 0.25 5-25. (a) 0.75 (b) 34 (c) 0.875 (d) 0.25 (g) 1 for x  0 5-27. (a) 0.032 (b) 0.0267 Section 5-2 5-29. 0.8851 5-31. c  136,   0.0435 5-33.   0.5, negative 5-35. c  881,   0.4924 5-37. XY  XY  0 5-39. XY  XY  0 Section 5-3 5-43. (a) p1  0.05, p2  0.85, p3  0.10 (d) E(X)  1, V(X)  0.95 (f ) 0.07195 (g) 0.7358 (h) E(Y)  17 (i) 0 ( j) P(X  2, Y  17)  0.0540, P(X  2  Y  17)  0.2224 (k) E(X  Y  17)  1 5-45. (b) 0.1944 (c) 0.0001 (e) E(X )  2.4 (f ) E(Y )  1.2 (g) 0.7347 (h) 0 (i) P(X  0  Y  2)  0.0204, P(X  1  Y  2)  0.2449, P(X  2  Y  2)  0.7347 ( j) 1.7143 5-47. (a) 0.7887 (b) 0.7887 (c) 0.6220 5-49. 0.8270 Section 5-4 5-55. (a) 18 (b) 77 (c) 0.5 (d) 0.873 5-57. (a) E(T)  4, T  0.1414 5-59. (a) 0 (b) 1 5-61. E(X)  1290, V(X)  19600 5-63. (a) 0.002 (b) n  6 (c) 0.9612 5-65. (a) 0.0027 (b) No (c) 0

Section 5-5 5-67. fY (y)  1⁄4; y  3, 5, 7, 9 5-69. (b) 18 Supplemental Exercises 5-75. (a) 38 (b) 34 (c) 34 (d) 38 (e) E(X )  78, V(X )  3964, E(Y )  78, V(Y )  3964 (h) 23 ( i) not independent ( j) 0.7949 5-77. (a) 0.0560 ( b) Z~Bin(20, 0.1) (c) 2 (d) 0.863 (e) Z|X~Bin(4, 0.25) (f ) 1 ( g) not independent 5-79. (a) 1108 (b) 0.5787 (c) 34 (d) 0.2199 (e) 94 (f ) 43 5-81. 34 5-83. (a) 0.085 ( b) Z~Bin(10, 0.3) (c) 3 5-85. (a) 0.499 (b) 0.5 5-87. (a) 0.057 (b) 0.057 5-91. (a) E(T)  1.5, V(T)  0.078 (b) 0.0367 (c) E(P)  4, V(P)  0.568 5-95. (a) p1  0.13, p2  0.72, p3  0.15, x y z  12 (b) not possible (c) 0.736 (d) 0 (e) 0.970 (f) 0.285 (g) 0.345

CHAPTER 6 Section 6-1 6-1. No, usually not, Ex: {2, 3} 6-3. No, usually not, Ex: {1, 2, 3, 1000}. 6-5. Yes, Ex: {5, 5, 5, 5, 5, 5, 5}. 6-7. x  74.0044, s  0.00473 6-9. x  7068.1, s  226.5 6-11. x  43.975, s  12.294 6-13.   5.44 6-15. x1  287.89, s1  160.154 x2  325.01, s2  121.20 6-17. x  7.184, s  0.02066 6-19. (a) x  65.86, s  12.16 (b) x  66.86, s  10.74 Section 6-2 6-21. Symmetric with a single mode 6-23. ~x  1436.5, lower quartile: Q1  1097.8, and upper quartile: Q3  1735.0

6-25. ~x  89.250, lower quartile: Q1 86.100, and upper quartile: Q3  93.125 6-27. median: ~x  1436.5, mode: 1102, 1315, and 1750, mean: x  1403.7 6-29. x  366.57, s  940.02, and ~x  41.455 6-31. 95th percentile  5479 6-33. x  260.3, s  13.41, ~x  260.85, and 90th percentile  277.2 6-35. x  89.45, s  2.8, ~x  90, and proportion  2240  55% Section 6-4 6-53. (a) x  2.415, s2  0.285, and s  0.543 6-55. (a) x  952.44, s2  9.55, and s  3.09 (b) ~x  953 6-57. (a) ~x  67.50, lower quartile: Q1  58.50, and upper quartile: Q3  75.00 (b) ~x  68.00, lower quartile: Q1  60.00, and upper quartile: Q3  75.00 Supplemental Exercises 6-87. (a) Sample 1 Range  4 Sample 2 Range  4 (b) Sample 1: s  1.604 Sample 2: s  1.852 6-93. (b) x  9.325, s  4.486 Mind-Expanding Exercises 6-105. s2 (old)  50.61, s2 (new)  5061.1 6-109. y  431.89, s2y  34.028 6-111. ~x  69 6-113. (a) x  89.29 (b) x  89.19 (c) No

CHAPTER 7 Section 7-2 7-1. 8/ 103 7-3. 0.8186 7-5. 0.4306 7-7. 0.191 7-9. n  12 7-11. 0.2312 7-13. (a) 0.5885 (b) 0.1759 7-15. 0.983

JWCL232_AppB_731-748.qxd

1/23/10

10:05 AM

Page 737

APPENDIX B

Section 7-3 7-17. (a) N  25, Mean  150.47, S2  105.06, SS  2521.5 (b) 150.468 7-19. Bias  2 1 n c 1 12 7-21. V1 ˆ 1 2  2  7 is smallest 7-27. Bias  2n 7-29. (a) 423.33 (b) 9.08 (c) 1.85 (d) 424 (e) 0.2917 7-33. (d) 0.01 (e) 0.0413 Section 7-4 7-35. x 7-39. unbiased 7-41. (a) ˆ  xi2 12n2 (b) same as part (a) 7-47. (b) 0(m x 1) (m 0 1) 7-49. (a) 5.046 (b) 5.05

8-25. (a) Mean  25.1848, Variance  2.5760 (b) 24.037    26.333 8-27. [58197.33, 62082.07] 8-29. [1.094, 1.106] 8-31. ( , 125.312] 8-33. [443.520, 528.080] 8-35. [301.06, 333.34] 8-37. (b) [2237.3, 2282.5] (c) [2241.4, ) 8-39. (b) [4.051, 4.575] (c) [4.099, ) 8-41. (b) [2.813, 2.991] Section 8-3 8-43. (a) 18.31 (b) 13.85 (c) 10.12, 30.14 8-45. [0.0055, ) 8-47. [0.31, 0.46] 8-49. [0.626, 1.926] 8-51.  0.0122

Supplemental Exercises 7-55. 0.8664 7-57. 5.6569 7-59. n  100 7-61. ˆ  x3

Section 8-4 8-53. (a) [0.02029, 0.06637] (b) [ , 0.0627] 8-55. (a) [0.501, 0.571] (b) [0.506, ) 8-57. (a) [0.225, 0.575] (b) 2305 (c) 2401 8-59. 666

CHAPTER 8

Section 8-6 8-61. [52131.1, 68148.3] 8-63. [1.068, 1.13] 8-65. [292.049, 679.551] 8-67. [263.7, 370.7] 8-69. [2193.5, 2326.5] 8-71. 90% PI  [2.71, 3.09] 90% CI  [2.85, 2.95] 99% CI  [2.81, 2.99] 8-73. [49555.54, 70723.86] 8-75. [1.06, 1.14] 8-77. TI  [237.18, 734.42] CI  [443.42, 528.08] 8-79. TI  [247.60, 386.60] CI  [301.06, 333.34] 8-81. TI  [2.49, 3.31] CI  [2.84, 2.96]

ˆ  21.86, ˆ 5000  109,300, 7-65.  pˆ  0.7

Section 8-1 8-1. (a) 96.76% (b) 98.72% (c) 93.56% 8-3. (a) 1.29 (b) 1.65 (c) 2.33 8-5. (a) 1st CI  50, 2nd CI  50 (b) higher confidence implies a wider CI 8-7. (a) 4 (b) 7 8-9. (a) Longer (b) No (c) Yes 8-11. [87.85, 93.11] 8-13. (a) [74.0353, 74.0367] (b) [74.035, ) 8-15. (a) [3232.11, 3267.89] (b) [3226.4, 3273.6] 8-17. 267 8-19. 22 8-21. (a) [13.383, 14.157] (b) [13.521, ) (c) 1 (d) 2 Section 8-2 8-23. (a) 2.179 (b) 2.064 (c) 3.012 (d) 4.073

Supplemental Exercises 8-85. (a) 0.0997 and 0.064 (b) 0.044 and 0.014 (c) 0.0051 and  0.001 8-87. (a) Normality (b) [16.99, ) (c) [16.99, 33.25]

8-89. 8-91.

8-93. 8-95.

8-97.

8-99.

737

(d) ( , 343.74] (e) [28.23, 343.74] (f ) 16.91    29.09 15.85  2  192.97 (g) mean: [16.88, 33.12], variance: [28.16, 342.94] (a) [13.74, 16.92] (b) [13.24, 17.42] (a) Yes (b) [197.84, 208.56] (c) [185.41, 220.99] (d) [171.21, 235.19] [0.0956, ) (a) Yes (b) [1.501, 1.557] (c) [1370, 1.688] (d) [1.339, 1.719] (a) [0.0004505, 0.009549] (b) 518 (c) 26002 (a) Normality (c) [18.478, 26.982] (e) [19.565, 123.289]

Mind-Expanding Exercises 8-101. (b) [28.62, 101.98] 8-103. (a) 46 (b) [10.19, 10.41], p  0.6004 8-105. 950 of CIs and 0.9963

CHAPTER 9 Section 9-1 9-1. (a) Yes (b) No (c) No (d) No (e) No 9-3. (a) H0:  20nm H1:  20nm (b) No 9-5. (a)   0.02275. (b)   0.15866 (c)   0.5 9-7. (a) 11.4175  Xc  11.42 (b) 11.5875  Xc  11.59 (c) 11.7087  Xc  11.71 (d) 11.7937  Xc  11.84 9-9. (a) P-value  0.0135 (b) P-value  0.000034 (c) P-value  0.158655 9-11. (a) 0.09296 (b)   0.04648 (c)   0.00005 9-13. (a)   0.005543 (b)   0.082264 (c) As n increases,  decreases 9-15. (a)   0.05705 (b)   0.5 (c)   0.05705.

JWCL232_AppB_731-748.qxd

1/23/10

10:06 AM

Page 738

738

APPENDIX B ANSWERS TO SELECTED EXERCISES

9-17. (a) (b) (c) (d) 9-19. (a) (b) (c) 9-21. (a) (b) (c)

  191.40   185.37   186.6   183.2 P-value  0.2148 P-value  0.008894 P-value  0.785236   0.0164 1   0.21186  will increase and the power will increase with increased sample size. P-value  0.238 P-value  0.0007 P-value  0.2585   0.29372   0.25721 0   0.99506

9-23. (a) (b) (c) 9-25. (a) (b) 9-27. (a) (b)

Section 9-2 9-29. (a) H0 :   10, H1 :   10 (b) H0 :   7, H1 :   7 (c) H0 :   5, H1 :   5 9-31. (a) a  z  2.33 (b) a  z  1.64 (c) a  z  1.29 9-33. (a) P-value  0.04 (b) P-value  0.066 (c) P-value  0.69 9-35. (a) P-value  0.98 (b) P-value  0.03 (c) P-value  0.65 9-37. (a) StDev  0.7495, z0  0.468, P-value  0.68, fail to reject H0 (b) one-sided (c) 19.42    20.35 (d) 0.640 9-39. (a) 0.6827 (b) one-sided (c) P-value  0.002, reject H0 (d) 98.8518   (e) P-value  0.15, fail to reject H0 9-41. (a) z0  0.95  1.96, fail to reject H0 (b)   0.80939 (c) n  16 9-43. (a) z0  1.26  1.65 fail to reject H0 (b) P-value  0.1038 (c)   0.000325 (d) n  1 (e) 39.85  

9-45. (a) z0  1.56  1.65, fail to reject H0 (b) P-value  0.94 (c) Power  0.97062 (d) n  5 (e)   104.53 9-47. (a) z0  1.77  1.65, reject H0 (b) P-value  0.04 (c) Power  1 (d) n  2 (e) 4.003  

9-65. (a) t0  0.15  1.753., fail to reject H0; P-value  0.40. (b) n  4. Yes 9-67. t0  3.46  1.833, reject H0 0.0025  P-value  0.005 9-69. (a) t0  14.69  1.6604, fail to reject H0; P-value  0.995 (b) Yes (c) power  1 (d) n  15

Section 9-3 9-49. (a) critical value  2.539 (b) critical value  1.796 (c) critical value  1.345 9-51. (a) 0.05  p  0.1 (b) 0.05  p  0.1 (c) 0.5  p  0.8 9-53. (a) 0.95  p  0.975 (b) 0.025  p  0.05 (c) 0.6  p  0.75 9-55. (a) 9 (b) .05  P-value   0.1, fail to reject H0 (c) two-sided (d) 11.89    13.23 (e) t0  1.905, reject H0 (f) reject H0 9-57. (a) t0  0.6665, fail to reject H0 (b) equal to P-value L 0.5 (c) 246.84    404.15, fail to reject H0 9-59. (a) 0 t0 0  3.48  2.064 reject H0 P-value  0.002 (b) Yes (c) power  1 (d) n  20 (e) 98.065    98.463 9-61. (a) 0 t0 0  1.456  2.064, fail to reject H0; 0.1  P-value  0.2. (b) Yes (c) power  0.80 (d) n  100 (e) 129.406    130.100 9-63. (a) 0 t0 0  1.55  2.861, fail to reject H0; 0.10  P-value  0.20 (b) Yes, see normal probability plot (c) power  0.30 (d) n  40 (e) 1.9    4.62

Section 9-4 9-71. (a) critical values 6.84 and 38.58 (b) critical values 3.82 and 21.92 (c) critical values 6.57 and 23.68 9-73. (a) 21 ,n 1  7.63 (b) 21 ,n 1  4.57 (c) 21 ,n 1  7.79 9-75. (a) 0.5  P-value  0.9 (b) 0.5  P-value  0.9 (c) 0.005  P-value  0.01 9-77. (a) 20  0.23  26.30, fail to reject H0; P-value  0.995 (b) 0.07  9-79. (a) 20  109.52  71.42, reject H0; P-value  0.01 (b) 0.31   0.46 9-81. (a) 20  12.46  7.26, fail to reject H0; 0.1  P-value  0.4 (b)  5240 9-83. (a) 20  11.52  19.02, fail to reject H0; 0.2  P-value (b) 0.45 (c) n  30 Section 9-5 9-85. (a) one-sided (b) appropriate (c) pˆ  0.3564, z0  1.1867 P-value  0.118, pˆ  0.6105 (d) 0.2354 9-87. (a) z0  1.31  1.65, fail to reject H0; P-value  0.095 (b) p  0.0303 9-89. (a) z0  2.06  1.65, reject H0; P-value  0.0196 (b) 0.7969  p 9-91. (a) z0  0.94  2.33 fail to reject H0; P-value  0.826 (b) 0.035  p

JWCL232_AppB_731-748.qxd

1/25/10

7:34 AM

Page 739

APPENDIX B

z0  1.58  2.33, fail to reject H0 9-95. (a) z0  0.54  1.65, fail to reject H0; P-value  0.295 (b)  0.639, n ⬵ 118 9-93.

Section 9-7 9-97. (a) 20  6.955  15.09, fail to reject H0 (b) P-value  0.2237 (from Minitab) 9-99. (a) 20  10.39  7.81 reject H0 (b) P-value  0.0155 (from Minitab) 9-101. (a) 20  769.57  36.42, reject H0 (b) P-value ⬵ 0 Section 9-8 9-103. 20  20.05,6, fail to reject H0; P-value  0.070 (from Minitab) 9-105. 20  20.01,9 reject H0; P-value  0.002 (from Minitab) 9-107. 20  20.01,3, fail to reject H0; P-value  0.013 9-109. (a) 20  20.05,1 (b) P-value  0.005 Section 9-9 9-111. (a) P-value  2P(R   7| p  0.5)  0.132 (b) 0.180 9-113. (a) P-value  2*P(R 3| p  0.5) ⬇ 1 (b) z0  0, P-value ⬇ 1 9-115. Ignore ties (a) w  9  w*0.05,n 9  5, fail to reject H0 (b) 0.110 9-117. Ignore ties (a) w  27  w*0.05,n17  41, reject H0 Supplemental Exercises 9-119. (a) 15 (b) SE Mean  1.1525, t0  1.449, 0.1  P-value < 0.2, 95.874    100.786 (c) fail to reject H0 (d) fail to reject H0 9-121. (a) t0  0.5694, 0.25  P-value  0.4 (b) 10.726    14.222

9-123. (a) n  25, 0.9783; n  100, 0.9554; n  400, 0.8599; n  2500, 0.2119 (b) n  25, 0.3783; n  100, 0.2643; n  400, 0.1056; n  2500, 0.0009; significant when n  2500. 9-125. (a) 0.4522 (b) 0.4404 (c) 0.3557 (d) 0.2981 9-127. (a) 0.6406 (b) 0.2877 (c) 0.0537 9-129. (a) n  100, 0.3632; n  150, 0.2119; n  300, 0.0352 (b) n  100, 0.6293; n  150, 0.4522; n  300, 0.1292 (c) 1 (d) 24, 5 9-131. (a) d  2,  0; d  3,  0 (b) 2 (c) d  1, ⬇ 0.1; d  1.5 ⬇ 0.04; n  4 9-133. 20  0.000009, reject H0 9-135. (a) Value

0

1

2

3

Observed 3 7 4 6 Expected 3.6484 6.2128 5.2846 4.6954

20  0.88971, fail to reject H0 (b) P-value  0.6409 (from Minitab) 9-137. (a) normal distribution used because sample size is large, z0  6.12, fail to reject H0 (b) P-value ⬇ 1 (c) Interval x  45.50 45.50  x  51.43 51.43  x  55.87 55.87  x  59.87 59.87  x  63.87 63.87  x  68.31 68.31  x  74.24 x 74.24

Obs. Exp. Frequency. Frequency. 9 7.5 5 7.5 7 7.5 11 7.5 4 7.5 9 7.5 8 7.5 6 7.5

20  5.06, fail to reject H0 9-139. (a) H0:   0.635 vs. H1:   0.635 (b) normal distribution used because sample size is large, z0  5.31, fail to reject H0 (c) P-value ⬇ 1 9-141. (a) t0  6.10, P-value < 0.001, reject H0

739

(b) d  4.54, Power ⬇ 1 (c) d  2.27, n 5 (e) From a normal probability plot, assumption is reasonable 9-143. (a) t0  0.37, fail to reject H0 (b) From a normal probability plot, assumption is reasonable (c) 0.25  P-value  0.4 9-145. (a) 20  58.81, reject H0 (b) P-value  0.01 9-147. (a) 20  0.509, fail to reject H0 (b) 20  0.509, P-value  0.01, reject H0

CHAPTER 10 Section 10-1 10-1. (a) 1.96  z0  0.9  1.96, do not reject H0; P-value  0.368 (b) 9.79  1  2  3.59 (c) Power  0.14 (d) n1  n2  180 10-3. (a) z0  0.937  2.325, do not reject H0; P-value  0.174 (b) 1  2 4.74 (c) Power  0.04 (d) Use n1  n2  339 10-5. (a) z0  5.84  1.645, do not reject H0; P-value  1 (b) 1  2 6.8 (c) Power  0.9988 (d) The sample size is adequate 10-7. (a) z0  7.25  1.645 reject H0; P-value ⬵ 0 (b) 3.684  1  2  2.116 (c) n1  n2  11 10-9. (a) 5.83  1  2  0.57; P-value  0.0173 (b) Yes (c) Power  0.9616; n ⬵ 10 (d) Normal Section 10-2 10-11. (a) df  26.45 ⬇ 26, 1  2  1.688, 0.0025  P-value  0.005, one-sided

JWCL232_AppB_731-748.qxd

740

10-13.

10-15.

10-17.

10-19.

10-21.

10-23.

10-25.

10-27.

10-29.

1/23/10

10:08 AM

Page 740

APPENDIX B ANSWERS TO SELECTED EXERCISES

(b) reject H0 (c) Yes (d) 0.005  P-value  0.01, reject H0 (a) t0  1.94  1.701 0.025  P-value  0.05 (b) 1 2  0.196 (c) Power  0.95 (d) n  n1  n2  21 (a) 2.042  t0  0.230  2.042, do not reject H0; P-value  0.80 (b) 0.394  1 2  0.494 (a) t0  3.11  2.485, reject H0 (b) 5.688  1 2  0.3122 (a) Assumptions verified (b) t0  2.83  2.101 reject H0; 0.010  P-value  0.020 (c) 0.7495  1 2  0.1105 (a) t0  5.498  2.021 reject H0; P-value  0.0010 (b) n1  n2  38 (a) t0  3.03  2.056 reject H0; 0.005  P-value  0.010 (b) t0  3.03  1.706, reject H0 (a) t0  7.0  2.048, reject H0; P-value  0 (b) 14.93  1 2  27.28 (c) n  8 (a) t0  2.82  2.326 reject H0; P-value  0.025 (b) 1 2  0.178 (a) Normal (b) t0  2.558  2.101 reject H0; P-value  0.02 (c) 1.86  1 2  18.94 (d) Power  0.05 (e) n  51

Section 10-3 10-31. (a) w2  75  w*0.025  51, fail to reject H0 (b) P-value  2[1 P(Z  0.58)]  0.562 10-33. (a) w1  77  w*0.01  78, reject H0 (b) P-value  2[1 P(Z  2.19)]  0.034 10-35. (a) Min (258, 207)  w*0.05  185, fail to reject H0 (b) P-value  0.0155

Section 10-4 10-37. (a) 0.1699  d  0.3776 (b) t-test is appropriate. 10-39. 727.46  d  2464.21 10-41. (a) t0  5.465  1.761 reject H0; P-value  0 (b) 18.20  d 10-43. (a) t0  8.387  1.833 reject H0 (b) t0  3.45  1.833 reject H0 (c) Yes 10-45. (a) Normal (b) 0.379  d  0.349 (c) 6  n 10-47. (a) P-value  P(R+  r +  14 | p  0.5)  0.0005, reject H0

10-65. (a) 0.248  f0  3.337  4.03, do not reject H0 (b) No

Section 10-5 10-49. (a) f0.25,5,10  1.59 (b) f0.10,24,9  2.28 (c) f0.05,8,15  2.64 (d) f0.75,5,10  0.529 (e) f0.90,24,9  0.525 (f ) f0.95,8,15  0.311 10-51. f0  0.805  0.166, fail to reject H0; 1  2  2.20 10-53. (a) f0  1.21  0.333, fail to reject H0; 21 0.403  2  3.63 2 (b) Power  0.65 (c) n  31 10-55. (a) f0  0.923  0.365, do not reject H0 21 (b) 0.3369  2  2.640 2 1 10-57. (a) 0.6004   1.428 2 1 (b) 0.5468   1.5710 2 1 (c) 0.661  2 10-59. 0.4058  f0  1.78  2.46, do not reject H0 10-61. 0.248  f0  0.640  4.04, do not reject H0; 21 0.159  2  2.579 2 10-63. 0.333  f0  1.35  3, do not reject H0; 21 0.45  2  4.05 2

Supplemental Exercises 10-73. (a) df  38, t0  0.988, 0.2  P-value  0.5, 2.478  1 2  0.758 (b) one-sided (c) fail to reject H0 10-75. (a) normality, equality of variance, and independence of the observations. (b) 1.40  1 2  8.36 (c) Yes 21 (d) 0.1582  2  5.157 2 (e) No 10-77. (a) t0  2.554  1.895, reject H0 (b) t0  2.554  2.998, do not reject H0 (c) t0  1.986  1.895, do not reject H0 (d) t0  1.986  2.998, do not reject H0 10-79. (a) z0  6.55  1.96, reject H0 (b) z0  6.55  2.58, reject H0 (c) z0 is so large 10-81. (a) 0.0335  p1 p2  0.0329 (b) 0.0282  p1 p2  0.0276 (c) 95% CI: 0.0238  p1 p2  0.0232 90% CI: 0.0201  p1 p2  0.0195 10-83. (a) Yes (b) Yes if similar populations

Section 10-6 10.67. (a) one-sided (b) z0  1.4012, P-value  0.0806, 0.0085  p1 p2 (c) reject H0 at   0.10, fail to reject H0 at   0.05 10-69. (a) z0  4.45  1.96 reject H0; P-value  0 (b) 0.039  p1 p2  0.1 10-71. (a) z0  5.36  2.58 reject H0; P-value  0

JWCL232_AppB_731-748.qxd

1/23/10

10:09 AM

Page 741

APPENDIX B

10-85. (a) 0.0987  1 2  0.2813 (b) 0.0812  1 2  0.299 (c) 1 2  0.2813 (d) z0  3.42  1.96, reject H0; P-value  0.00062 (e) n1  n2  9 10-87. (a) z0  5.36  2.58, reject H0 (b) conclusions are the same (c) n  60 10-89. (a) No (b) data appear normal with equal variances (c) It is more apparent the data follow normal distributions. 2V (d) 18.114  2  294.35 M (e) f0  72.78  4.03, reject H0 10-91. (a) Normality appears valid. (b) 0.50  P-value  0.80, do not reject H0 (c) n  30 10-93. (a) It may not be assumed that 21  22 (b) t0  2.74  2.131, reject H0 (c) Power  0.95 (d) n  26 Mind-Expanding Exercises 10-99. (c) 0.519    3.887

CHAPTER 11 Section 11-2 11-1. (a) ˆ 0  48.013, ˆ 1  2.330 (b) 37.99 (c) 39.39 (d) 6.71 11-3. (a) ˆ 0  14.195, ˆ 1  10.092, 2  27.24 (b) 89.95 (c) 10.092 (d) 0.99 (e) 7.98, 3.13 11-5. (a) ˆ 0  6.3355, ˆ 1  9.20836, 2  3.7746 (b) 500.124 (c) 9.20836 (d) 1.618 11-7. (a) ˆ 0  16.5093, ˆ 1  0.0693554, 2  7.3212 (b) 1.39592 (c) 49.38 11-9. (b) ˆ 0  234.071, ˆ 1  3.50856, 2  398.25

(c) (d) 11-11. (b) (c) 11-13. (a) (b) (d) 11-15. (b) 11-17. (a) (b) 11-19. (b)

128.814 156.833 and 15.1175 ˆ 0  2625.39, ˆ 1  36.962, 2  9811.2 1886.15 ˆ 0  0.658, ˆ 1  0.178, 2  0.083 3.328 (c) 0.534 1.726 and 0.174 ˆ 0  2.02, ˆ 1  0.0287, 2  0.0253 yˆ  39.2 0.0025x ˆ 1  0.0025 ˆ *0  2132.41, ˆ *1  36.9618

Section 11-4 11-21. (a) t0  12.4853, P-value  0.001; t1  20.387, P-value  0.001; MSE  2.194; f0  415.91, P-value L 0 (b) reject H0:   0 (c) 2.194 11-23. (a) f0  74.63, P-value  0.000002, reject H0 (b) ˆ 2  27.2, se1ˆ 1 2  0.2696 (c) se1ˆ 0 2  0.9043 11-25. (a) f0  61.41, P-value  0, reject H0 (b) se1ˆ 1 2  1.288 se1ˆ 0 2  9.059 (c) t0  0.072, fail to reject H0 11-27. (a) f0  74334.4, P-value  0, reject H0 (b) se1ˆ 0 2  1.66765, se1ˆ 1 2  0.0337744 (c) t0  23.37, P-value  0.000, reject H0 (d) t0  3.8, P-value  0.005, reject H0 11-29. (a) f0  44.0279, P-value  0.00004, reject H0 (b) se1ˆ 0 2  9.84346, se1ˆ 1 2  0.0104524 (c) t0  1.67718, P-value  0.12166, fail to reject H0 11-31. (a) f0  155.2, P-value  0.00001, reject H0 (b) se1ˆ 0 2  2.96681, se1ˆ 1 2  45.3468 (c) t0  2.3466, P-value  0.0306, fail to reject H0

741

(d) t0  57.8957, P-value  0.00001, reject H0 (e) t0  2.7651, P-value  0.0064, reject H0 11-33. (a) P-Value  0.0000, reject H0 (b) ˆ 2  0.083 se1ˆ 0 2  0.1657, se1ˆ 1 2  0.014 (c) Reject H0 11-35. (a) P-value  0.310, No (b) ˆ 2  30.69 (c) se1ˆ 0 2  9.141 11-37. 0.55 Sections 11-5 and 11-6 11-39. (a) ( 2.9713, 1.7423) (b) (46.7145, 49.3115) (c) (41.3293, 43.0477) (d) (39.1275, 45.2513) 11-41. (a) [10.02, 15.28] (b) [ 4.30, 32.69] (c) [85.59, 104.27] (d) [62.77, 127.10] 11-43. (a) (9.10130, 9.31543) (b) ( 11.6219, 1.04911) (c) (498.72024, 501.52776) (d) (495.57344, 504.67456) 11-45. (a) (0.03689, 0.10183) (b) ( 47.0877, 14.0691) (c) (44.0897, 49.1185) (d) (37.8298, 55.3784) 11-47. (a) (201.552, 266.590) (b) ( 4.67015, 2.34696) (c) (111.8339, 145.7941) 11-49. (a) ( 43.1964, 30.7272) (b) (2530.09, 2720.68) (c) (1823.7833, 1948.5247) (d) (1668.9013, 2103.4067) 11-51. (a) (0.1325, 0.2235) (b) (0.119, 1.196) (c) (1.87, 2.29) Section 11-7 11-53. (a) R2  67.2% 11-55. (a) R2  99.986% 11-57. (a) R2  87.94% 11-59. (a) R2  85.22% 11-61. (a) R2  89.6081% (c) R2  95.73% (d) ˆ 2 old  9811.21, ˆ 2 new  4022.93 11-65. (a) f0  207, reject H0

JWCL232_AppB_731-748.qxd

742

1/23/10

10:11 AM

Page 742

APPENDIX B ANSWERS TO SELECTED EXERCISES

Section 11-8 11-67. (a) t0  4.81, P-value  0.0005, reject H0 (b) z0  1.747, P-value  0.04, reject H0 (c)   2.26, reject H0 11-69. (a) t0  5.475, P-value  0.000, reject H0 (b) (0.3358, 0.8007) (c) Yes 11-71. (a) yˆ  0.0280411

0.990987x (b) f0  79.838, reject H0 (c) 0.903 (d) t0  8.9345, reject H0 (e) z0  3.879, reject H0 (f ) (0.7677, 0.9615) 11-73. (a) yˆ  5.50 6.73x (b) 0.948 (c) t0  8.425, reject H0 (d) (0.7898, 0.9879) 11-75. (a) r  0.82 (b) t0  7.85, reject H0, P-value  0.005 (c) (0.664, 0.908) (d) z0  1.56, fail to reject H0, P-value  0.119 Section 11-9 11-77. (a) Yes (b) No (c) Yes (d) Yes 11-79. (b) yˆ  0.8819 0.00385x (c) f0  122.03, reject H0 Section 11-10 11-81. (a) ˆ 0  5.340, ˆ 1  0.00155 (b) Test that all slopes zero: P-value  0 11-83. (a) ˆ 0  7.047, ˆ 1  0.00074, ˆ 2  0.9879 (b) Test that all slopes zero: P-value = 0.036 (c) 0.771 11-85. (b) yˆ  0.966824 1.54376x (c) f0  252263.9, P-value  0, reject H0 (d) [1.5367, 1.5509] (e) t0  199.34, reject H0 Supplemental Exercises 11-87. y*  1.2232 0.5075x where y*  1y

11-89. yˆ  0.7916x 11-91. (b) yˆ  0.6714 2964x (c) R2  21.5% 11-93. (b) yˆ  0.699 1.66x (c) f0  22.75, reject H0, P-value  0.001 (d) (3.399, 5.114) 11-95. (c) all data: (7741.74, 10956.26), outlier removed: (8345.22, 11272.79)

CHAPTER 12 Section 12-1 12-1. (b) yˆ  171.055 3.714x1 1.126x2 (c) 189.49 12-3. (b) 2 12-5. (a) Satisfaction  144 1.11 Age 0.585 Severity 1.30 Anxiety (b) 49.5 (c) 5.90, 0.13, 0.13, 1.06 (d) No, standard errors differ 12-7. (a) yˆ  49.90 0.01045x1 0.0012x2 0.00324x3

0.292x4 3.855x5

0.1897x6 (b) ˆ 2  4.965 se1ˆ 0 2  19.67, se1ˆ 1 2  0.02338, se1ˆ 2 2  0.01631, se1ˆ 3 2  0.0009459, se1ˆ 4 2  1.765, se1ˆ 5 2  1.329, se1ˆ 6 2  0.273 (c) 29.867 12-9. (a) yˆ  47.82 9.604x1

0.44152x2 18.294x3 (b) ˆ 2  12.3 (c) se1ˆ 0 2  49.94, se1ˆ 1 2  3.723, se1ˆ 2 2  0.2261, and se1ˆ 3 2  1.323 (d) 91.38 12-11. (a) yˆ  440.39 19.147x1

68.080x2 (b) ˆ 2  55563 se1ˆ 0 2  94.20, se1ˆ 1 2  3.460, and se1ˆ 2 2  5.241 (c) 186.675 12-13. (a) yˆ  0.1105 0.4072x1

2.108x2

(b) ˆ 2  0.00008 se1ˆ 0 2  0.2501, se1ˆ 1 2  0.1682, and se1ˆ 2 2  5.834 (c) 0.97074 12-15. (a) yˆ  238.56 0.3339x1 2.7167x2 (b) ˆ 2  1321 (c) se1ˆ 0 2  45.23, se1ˆ 1 2  0.6763, and se1ˆ 2 2  0.6887 (d) 61.5195 12-17. (a) yˆ  2.99 1.20x3

4.60x7 3.81x10 (b) ˆ 2  4.14 (c) se1ˆ 0 2  5.877, se1ˆ 3 2  0.974, se1ˆ 7 2  0.385, se1ˆ 10 2  0.486 (d) 81.96 12-19. (a) yˆ  383.80 3.6381x1 0.1119x2 (b) ˆ 2  153.0, se1ˆ 0 2  36.22, se1ˆ 1 2  0.5665, se1ˆ 2 2  0.04338 (c) 180.95 (d) yˆ  484.0 7.656 x1 0.222x2 0.0041x12 (e) ˆ 2  147.0, se1ˆ 0 2  101.3, se1ˆ 1 2  3.846, se1ˆ 2 2  0.113, se1ˆ 12 2  0.0039 (f ) 173.1 Section 12-2 12-21. (a) t0  53.0872, P-value  0, t1  15.02, P-value  0; t2  23.43, P-value  0; MSE  25.5833; f0  445.2899, P-value  0 (b) f0  445.2899, reject H0 (c) t-test for each regressor is significant 12-23. (a) f0 184.25, P-value  0.000, Reject H0 (b) t0 1 ˆ 1 2  16.21, P-value  0.001, reject H0 t0 1 ˆ 2 2  11.04, P-value  0.001, reject H0

JWCL232_AppB_731-748.qxd

1/23/10

10:12 AM

Page 743

APPENDIX B

12-25. (a) P-value  8.16 E-5, 3.82 E-8, and 0.3378 (b) t0  0.98, fail to reject H0 12-27. (a) f0  19.53, reject H0 12-29. (a) f0  828.31, reject H0 (b) t0  2.58, reject H0 t0  1.84, fail to reject H0 t0  13.82, reject H0 12-31. (a) f0  99.67, reject H0, P-value  0 (b) t0  5.539, reject H0 t0  12.99, reject H0 12-33. (a) f0  9.28, reject H0, P-value  0.015 (b) t0  0.49, fail to reject H0 t0  3.94, reject H0 12-35. (a) f0  36.59, reject H0 (b) t0  12.30, reject H0 t0  11.94, reject H0 t0  7.84, reject H0 (c) f0  142.66, reject H0 12-37. (a) f0  97.59, reject H0, P-value  0.002 (b) t0  6.42, reject H0 t0  2.57, fail to reject H0 (c) f0  6.629, fail to reject H0 (d) f0  7.714, fail to reject H0 (e) f0  1.11, fail to reject H0 (f ) 147.0 12-39. (a) f0  65.55, P-value  0,

reject H0 (b) Age: t1  8.40, Severity: t2 = 4.43, Anxiety: t3 = 1.23, not all necessary Sections 12-3 and 12-4 12-41. (a) (49.927  0  292.183) (0.033  1  7.393) ( 2.765  2  0.513) (b) (158.82, 220.13) (c) (126.06, 252.88) 12-43. (a) ( 20.477  1  1.269) ( 0.245  2  1.076) (14.428  3  22.159) (b) (77.582, 105.162) (c) (82.133, 100.611) 12-45. (a) ( 6.9467  1  0.3295) ( 0.3651  2  0.1417) (b) ( 45.8276  1  30.5156) ( 1.3426  2  0.8984) ( 0.03433  3  0.04251)

12-47. (a) (12.1363  1  26.1577) (57.4607  2  78.6993) (b) ( 233.4, 63.2) (c) ( 742.09, 571.89) 12-49. (a) (0.0943  1  0.7201) ( 8.743  2  12.959) (b) (0.861, 0.896) (c) (0.855, 0.903) 12-51. (a) ( 2.173  1  2.841) ( 5.270  2  0.164) (b) ( 36.7, 125.8) (c) ( 112.8, 202.0) (d) CI: (107.4, 267.2) PI: (30.7, 344.0) 12-53. (a) 10.18  0  16.62 (1.00  3  1.45) (3.85  7  5.00) ( 5.11  10  3.07) (b) 0.3877 (c) (81.3, 82.9) 12-55. (a) [ 9.052, 15.024], [0.999, 1.398], [3.807, 5.384], [ 4.808, 2.817] (b) 0.43 (c) [80.965, 82.725] Section 12-5 12-57. (a) 0.893 12-59. (a) 0.95 12-61. (a) 0.843 12-63. (a) 0.997 12-65. (a) 0.756 12-67. (a) 0.985 (b) 0.99 12-69. (b) 0.9937, 0.9925 12-71. (a) 0.12 (b) Yes Section 12-6 12-73. (a) yˆ  1.633 1.232x 1.495x2 (b) f0  1,858,613, reject H0 (c) t0  601.64, reject H0 12-75. (a) yˆ  1.769 0.421x1

0.222x2 0.128x3 0.02x12 0.009x13

0.003x23 0.019x21 0.007x22 0.001x23 (b) f0  19.628, reject H0 (d) f0  1.612, fail to reject H0 12-77. (b) yˆ  56.677 0.1457x1 0.00525x2 0.138x3 4.179x4

743

12-79. (a) Min Cp : x1, x2 Cp  3.0, MSE  55563.92 yˆ  440.39 19.147x1

68.080x2 Min MSE is same as Min Cp (b) Same as part (a) (c) Same as part (a) (d) Same as part (a) (e) All models are the same 12-81. (a) Min Cp: x1 Cp  1.1, MSE  0.0000705 yˆ  0.20052

0.467864x1 Min MSE is same as Min Cp (b) Same as part (a) (c) Same as part (a) (d) Same as part (a) (e) All models are the same 12-83. (a) Min Cp: x2 Cp  1.2, MSE  1178.55 yˆ  253.06 2.5453x2 Min MSE is same as Min Cp (b) Same as part (a) (c) Same as part (a) (d) Same as part (a) (e) All models are the same 12-85. (a) Min Cp: x1, x2 Cp  2.9, MSE  10.49 yˆ  50.4 0.671x1

1.30x2 Min MSE is same as Min Cp (b) Same as part (a) (c) Same as part (a) (d) Same as part (a) (e) All models are the same 12-87. (a) Min. MSE (MSE  0.01858) model Att, PctComp, Yds, YdsperAtt, TD, PctTD, PctInt; Min. Cp (Cp  5.3) model PctComp, YdsperAtt, PctTD, PctInt; (b) PctComp, YdsperAtt, PctTD, PctInt (c) PctComp, YdsperAtt, PctTD, PctInt (d) Att, PctComp, Yds, YdsperAtt, TD, PctInt 12-89. (a) Min Cp  1.3: yˆ  61.001 0.02076xcid 0.00354xetw 3.457xaxle Min MSE  4.0228 yˆ  49.5 0.017547xcid

JWCL232_AppB_731-748.qxd

744

1/23/10

10:14 AM

Page 744

APPENDIX B ANSWERS TO SELECTED EXERCISES

0.0034252xetw 1.29xcmp 3.184xaxle 0.0096xc02 (b) yˆ  63.31 0.0178xcid 0.00375xetw 3.3xaxle 0.0084xc02 (c) Same as Min MSE model in part (a) (d) yˆ  45.18 0.00321xetw 4.4xaxle 0.385xn/v (e) Min Cp model is preferred (f ) Min Cp  4.0, Min MSE  2.267 yˆ  10 0.0038023xetw

3.936xcmp 15.216xco 0.011118xc02 7.401xtrans

3.6131xdrv1 2.342xdrv2 Stepwise: yˆ  39.12 0.0044xetw

0.271xn/v 4.5xtrns

3.2xdrv1 1.7xdrv2 Forward selection: yˆ  41.12 0.00377xetw

0.336xn/v 2.1xaxle 3.4xtrans 2.1xdrv1

2xdrv2 Backward elimination is same as Min Cp and Min MSE 12-91. (a) yˆ  0.304 0.083x1 0.031x3 0.004x22 Cp  4.04 MSE  0.004 (b) yˆ  0.256 0.078x1

0.022x2 0.042x3

0.0008x23 Cp  4.66 MSE  0.004 (c) Prefer the model in part (a) Supplemental Exercises 12-95. (a) 2 (b) 0.0666, 0.0455 (c) 6.685 12-97. (a) f0  1321.39, reject H0, P-value  0.00001 (b) t0  1.45, fail to reject H0 t0  19.95, reject H0 t0  2.53, fail to reject H0 12-99. (a) yˆ  4.87 6.12x*1 6.53x*2 3.56x*3 1.44x*4 (b) f0  21.79, reject H0 t0  5.76, reject H0 t0  5.96, reject H0 t0  2.90, reject H0 t0  4.99, reject H0 12-101. (a) yˆ*  21.068 1.404x*3

0.0055x4 0.000418x5

(b) (c) (d) (e) 12-103. (a) (b) (c) (d) (e) 12-105. (a) (b) (c)

MSE  0.013156 Cp  4.0 Same as part (a) x4, x5 with Cp  4.1 and MSE  0.0134 The part (c) model is preferable Yes yˆ  300.0 0.85x1

10.4x2, yˆ  405.8 f0  55.37, reject H0 0.9022 MSE  10.65 f0  0.291, fail to reject H0 f0  18.28, reject H0 f0  2, do not reject H0 MSE 1reduced2  0.005 MSE 1Full2  0.004

Mind-Expanding Exercises 12-109. R2  0.449

CHAPTER 13 Section 13-2 13-1. (a) 4 (b) 5 (c) f0  1.58, 0.1  P-value  0.25 (d) fail to reject H0 13-3. (a) f0  14.76, reject H0 13-5. (a) f0  12.73, reject H0 (b) P-value  0 13-7. (a) f0  16.35, reject H0 (c) 95%: (140.71, 149.29) 99%: (7.36, 24.14) 13-9. (a) f0  1.86, fail to reject H0 (b) P-value  0.214 13-11. (a) f0  8.30, reject H0 (b) P-value  0.002 (d) (69.17, 81.83) (e) (8.42, 26.33) 13-13. (a) f0  0.72, fail to reject H0 (b) P-value  0.486, SSE  0.146 13-15. (a) f0  2.62, fail to reject H0 (b) P-value  0.083 13-27. n  3

Section 13-3 13-29. (a) f0  5.77, reject H0 (b) 0.01412 (c) 0.0148 13-31. (a) f0  0.75, fail to reject H0 (b) 0 (c) 24 13-33. (a) H0: 2  0 H1: 2  0 (c) Set equal sample sizes, 2  0.164, 2  0 Section 13-4 13-35. f0  147.35, P-value  0, reject H0 at   0.05 or   0.01 13-37. (a) f0  8.92, reject H0 13-39. (a) f0  3.00, fail to reject H0 13-41. (a) f0  1.61, fail to reject H0 13-43. (a) 3 (b) 3 (c) f0  23.15, P-value < 0.01 (d) reject H0 Supplemental Exercises 13-45. (a) f0  76.09, reject H0 (c) (132.97, 147.83) 13-47. (a) f0  7.84, reject H0 (b) P-value  0.007 13-49. (a) f0  6.23, reject H0 (c) Algorithm 5 13-51. (a) Power  0.2 (b) n  50

CHAPTER 14 Section 14-3 14-1. (a) 1. H0 : 1  2  0 H1 : at least one i  0 2. H0 : 1  2  3  0 H1 : at least one j  0 (b) f 1G 2  273.79, f 1P 2  8.84, f 1GP 2  1.26 reject H0 for only main effects 14-3. (a) f 1M 2  7.91, f 1T 2  28.97, f 1MT 2  3.56 reject H0 for both main effects and the interaction 14-5. (a) Yijk   i j 12 ij i  1, 2, 3

僆ijk • j  1, 2, 3, 4 k  1, 2, 3, 4, 5, 6 (b) f 1I 2  40.07, f 1T 2  0.32, f 1IT 2  1.70 reject H0 for only insulation

JWCL232_AppB_731-748.qxd

1/23/10

10:15 AM

Page 745

APPENDIX B

14-7. (a) f 1D 2  25.23, f 1A 2  543.52, f 1DA 2  3.54 reject H0 for both main effects and the interaction 14-9. ( 3.40, 7.64) Section 14-4 14-11. (a) f 1H 2  7.64, f 1C 2  31.66, f 1F 2  19.92, f 1HC 2  2.92, f 1HF 2  2.97, f 1CF 2  0.96 H, C, F, HF are significant at   0.05. The P-value for HC is 0.075. Section 14-5 14-13. Significant effects: A  17.00, C  10.875, D  8.375, AD  9.125 14-15. (a) Cleaning Method  5.93 Test Position  1.280 Clean*test  1.220 (b) Cleaning Method is the only significant factor 14-17. (a) Significant effects: A  11.8125, B  33.9375, C  9.6875, AB  7.9375 14-19. None of the effects are significant 14-21. (b) A, B and AB (c) yˆ  400 40.124x1 32.75x2 26.625x1x2 14-23. f0  5.11, do not reject H0 14-25. (a) For model with A, B, C, AB, s  2.92; s 1center pts2  2.86 (b) F0  192.71, curvature is significant 14-27. (a) Large effects: C  39.79, D  198.47, E  64.86 (b) y  546.90 39.79xC 198.47xD 64.86xE (d) none 14-29. (a) Large effects: B  0.000750 (b) with all effects in error except B the P-value  0.007 Section 14-6 14-31. Significant effects: A  15.625, C  10.625,

D  8.875, AD  8.875, CD  3.125, ACD  1.875 14-37. (a) Effect JED is confounded with blocks (b) Marginal significant effects: J  19.0, D  14.75, JD  18.0, ED  25.5, 14-39. (a) Blocking important, SSBlocks large relative to SSError (b) ABC, ABD, ACD, and BCD (c) Coefficient for AD  15.14, t-statistic for AD  1.525, dfBlocks  1, MS for interactions  1165.33 Section 14-7 14-41. (a) 25 1 (b) A  10.8750, B  33.6250, C  10.6250, D  0.6250, E  0.3750, AB CDE  7.1250, 14-43. (a) E  ABCD (b) Resolution V (c) E  0.4700, BE  0.4050, DE  0.3150 14-45. (c) A  1.435, B  1.465, D  4.545, AB  1.15, AD  1.23 14-47. (b) Design Generator: D  BE, E  AC Defining Relation: I  ACE  BDE  ABCDE Aliases A  CE  BCDE  ABDE B  DE  ACDE  ABCE C  AE  ABDE  BCDE D  BE  ABCE  ACDE E  AC  BD  ABCD (c) A  1.525, B  5.175, C  2.275, D  0.675, E  2.275 14-49. (a) Alias Structure: I ABD ACE BCF

DEF ABEF ACDF

BCDE A BD CE B AD CF

745

C AE BF D AB EF E AC DF F BC DE AF BE CD (b) Alias Structure: I ABCG

ABDH ABEF

ACDF ACEH

ADEG AFGH

BCDE BCFH

BDFG BEGH

CDGH CEFG

DEFH A B C D E F G H AB CG DH EF AC BG DF EH AD BH CF EG AE BF CH DG AF BE CD GH AG BC DE FH AH BD CE FG 14-51. (a) Generators are E  BCD, F  ACD, G  ABC, and H  ABD, I  BCDE  ACDF  ABEF  ABCG  ADEG  BDFG  CEFG  ABHD  ACEH  BCFH  DEFH  CDGH  BEGH  AFGH (b) Glassware  1.4497, Reagent  0.8624, Prep  0.6034, Tracer  0.6519, Dissolution  0.8052, Hood  1.3864, Chemistry  0.0591, Ashing  0.0129 Section 14-8 14-53. (b) yˆ  82.024 1.115x1 2.408x2 0.861x21 1.59x22 1.801x1 x2 14-55. Path of steepest ascent passes through the point (0, 0) and has a slope 0.81.5  0.533 14-57. (a) Central composite design, not rotatable (b) yˆ  150.04 58.47x1

3.35x2 6.53x21

JWCL232_AppB_731-748.qxd

746

1/23/10

10:16 AM

Page 746

APPENDIX B ANSWERS TO SELECTED EXERCISES

10.58x22 0.50x1 x2 The linear terms are significant ( p  0.001), while both the square terms and interaction terms are insignificant 14-59. (a) along the vector (1.2, 2.1, 1.6, 0.6) (b) (1.22, 2.13, 1.62, 0.61) Supplemental Exercises 14-61. (a) t 1 p2  2.54, t1 pc2  5.02 The main effect of pH and the interaction of pH and Catalyst Concentration (CC) are significant 14-63. (a) f 1L2  63.24, f 1S 2  39.75, f 1LS 2  5.29 L, S, LS are significant 14-65. (a) A  2.74, B  6.66, C  3.49, AB  8.71, AC  7.04, BC  11.46, ABC  6.49, not significant 14-67. (a) V  15.75, P  10.75, G  25.00, PG  19.25, (b) yˆ  102.75 7.87x1

5.37x3 12.50x4 9.63x34 14-69. V  3.25, F  10.25, P  10.25, G  23.75, PG  11.75 14-71. Design Generators: D  AB E  AC Alias Structure I ABD ACE BCDE A BD CE ABCDE B AD CDE ABCE C AE BDE ABCD D AB BCE ACDE E AC BCD ABDE BC DE ABE ACD BE CD ABC ADE 14-73. (a) 22 factorial with two replicates (b) Significant effects: x1  0.795, x2  1.160 (c) yˆ  0.1994 0.07688 x1 14-75. (a) H0 : 1  2  3  0 H0 : 1  2  3  0 H0 : 12 11  p  12 33  0

(b) f 1N 2  311.71, f 1P 2  119.17, f 1PN 2  92.94 (d) ˆ  1.22 14-77. (a) Generators are E  ABC, F  ABD, and G = ACD; I  ABCE  ABDF  CDEF  ACDG  BDEG  BCFG  AEFG (b) A  74.9, B  76.1, C  366.4, D  236.9, E  213.4, F  119.9, G  101.9 14-79. (a) Block 1: (1), bc, ac, ab; Block 2: a, b, c, abc (b) A  0.500, B  13.000, C  6.500, AB  2.000, AC  14.500, BC  5.000 14-81. (a) A: 4 levels, B: 3 levels (b) 1 (c) AB interaction not significant (d) dfError  6, SSB  34670882, MSE  29736583, f0  1.36

15-9.

15-11.

CHAPTER 15 Section 15-3 15-1. (a) x chart: UCL  242.78, CL  223, LCL  203.22, R chart: UCL  72.51, CL  34.286, LCL  0 ˆ  223 ˆ  14.74 (b)  15-3. (a) x chart: UCL  245.11, CL  223, LCL  200.89 S chart: UCL  30.77, CL  13.58, LCL  0 ˆ  223, ˆ  14.74 (b)  15-5. (a) x chart: UCL  4.930, CL  4.668, LCL  4.406, R chart: UCL  0.961, CL  0.454, LCL  0 (b) x chart: UCL  4.910, CL  4.668, LCL  4.425, S chart: UCL  0.355, CL  0.17, LCL = 0 15-7. (a) x chart: UCL  37.5789, CL  34.32, LCL  31.0611 R chart: UCL  11.9461, CL  5.65, LCL  0 (b) x chart: UCL  37.4038, CL  34.0947, LCL  30.7857 R chart: UCL  12.1297,

15-13. 15-15.

CL  5.73684, LCL  0 ˆ  2.4664 (a) x chart: UCL  17.4, CL  15.09, LCL  12.79 R chart: UCL  5.792, CL  2.25, LCL  0 (b) x chart: UCL  17.96, CL  15.78, LCL  13.62 R chart: UCL  5.453, CL  2.118, LCL  0 (c) x chart: UCL  17.42, CL  15.09, LCL  12.77 S chart: UCL  3.051, CL  1.188, LCL  0 Revised x chart: UCL  17.95, CL  15.78, LCL  13.62 S chart: UCL  2.848, CL  1.109, LCL  0 (a) x chart: UCL  0.0635, CL  0.0629, LCL  0.0624 R chart: UCL  0.0020, CL  0.0009, LCL  0 (b) x chart: UCL  0.0634, CL  0.0630, LCL  0.0626 R chart: UCL  0.0014, CL  0.0007, LCL  0 (c) x chart: UCL  0.0634, CL  0.0630, LCL  0.0626 S chart: UCL  0.00058, CL  0.00027, LCL  0 Revised: x chart: UCL  0.0633, CL  0.0630, LCL  0.0626 S chart: UCL  0.00058, CL  0.00028, LCL  0 (a) 2.73 (b) no s  2.956, rd2  1.251

Section 15-4 15-17. (a) Individual chart: UCL  60.889, CL  53.05, LCL  45.211 MR chart: UCL  9.634, CL  2.94737, LCL  0 The process appears to be in control. ˆ  53.05, ˆ  2.613 (b) 

JWCL232_AppB_731-748.qxd

1/23/10

10:18 AM

Page 747

APPENDIX B

15-19. (a) Individual chart: UCL  10.5358, CL  10.0272, LCL  9.5186 MR chart: UCL  0.625123, CL  0.19125, LCL  0 The process appears to be in control. ˆ  10.0272, ˆ  0.1696 (b)  15-21. (a) Initial study: Individual chart: UCL  130.5, CL  100.78, LCL  71.06 MR chart: UCL  36.51, CL  11.18, LCL  0 Revised: Individual chart UCL  127.08, CL  99.48, LCL  71.88 MR chart: UCL  33.91, CL  10.38, LCL  0 ˆ  99.4792, ˆ  9.20059 (b)  15-23. (a) X chart: UCL  116.43, CL  82.0, LCL  47.57, R chart: UCL  42.30, CL  12.95, LCL  0; in control (b) 82.0, 11.48 Section 15-5 15-25. (a) 1.3333 (b) 26 15-27. (a) PC  PCRK  1.5 (b) 0 15-29. Proportion nonconforming is 0.00779 PCR  0.905 PCRK  0.837 15-31. 0.0009 PCR  1.13 PCRK  1.06 15-33. (a) PCR  1.35 PCRK  1.217 (b) 0.00013 15-37. (a) Fraction defective  0.002, PCR  1.03, PCRK  1.03 (b) Fraction defective  0.057, PCR  1.03, PCRK  0.526 15-39. (a) 0.03 (b) PCR  1.11, PCRK  0.778 Section 15-6 15-41. (a) not in control UCL  0.0835, CL  0.0645, LCL  0.0455

(b) Revised P-chart: UCL  0.1252, CL  0.0561, LCL  0 15-43. (a) P chart: UCL  0.1986, CL  0.1506, LCL  0.1026 (b) Revised P chart: UCL  0.2062, CL  0.1573, LCL  0.1085 15-45. (a) The limits need to be revised. UCL  3.811, CL  1.942, LCL  0.072, sample 5 and 24 exceed limits (b) U chart: UCL  3.463, CL  1.709, LCL  0 15-47. (a) UCL  0.3528, CL  0.2598, LCL  0.1667, not in control (b) UCL  0.6694, CL  0.5195, LCL  0.3696, not in control, points 4, 5, 10, 12, 18, 20 exceed the control limits Section 15-7 15-49. (a) 4 (b) 0.0228. (c) 43.8596 15-51. (a) 0.2177 (b) ARL  4.6 15-53. (a) 0.1515 (b) ARL  6.6 15-55. (a) 0.1020 (b) ARL  9.8 15-57. (a) 0.2877 (b) ARL  3.48 15-59. (a) 5.196 (b) 0.01 (c) 102.04 Section 15-8 15-61. (a) h  4, k  0.5, UCL  3.875, LCL  3.875 Yes, this process is in-control. (b) Observation 20 is out of control, CUSUM  6.08 15-63. (a) ˆ  0.1736 (b) h  4, k  0.5, UCL  0.0678, LCL  0.0678, out of control at the specified target level 15-65. (a) ARL  38.0 (b) ARL  10.4 15-67. (a) 0.169548 (b) The process appears to be in control. UCL  10.17, CL  10, LCL  9.83

747

(c) Out of control at observation 13, UCL  10.29, CL  10, LCL  9.71 15-69. (a) prefer ␭  0.1 and L  2.81 (b) prefer ␭  0.5 and L  3.07 (c) 9 15-71. (a) UCL  12, LCL  12, not in control (b) shift from 70 to 80: 2.01  ARL  2.57 Supplemental Exercises 15-73. (a) x chart: UCL  64.019, CL  64, LCL  63.982 R chart: UCL  0.046, CL  0.018, LCL  0 ˆ  64, ˆ  0.011 (b)  (c) PCR  0.63 (d) PCRk  0.63 (e) 2  10.00332 2  0.000011 (f ) ARL  12.9 15-75. (a) The process appears to be in control. (b) P chart: UCL  0.1717, CL  0.1063, LCL  0.04093 15-77. (a) Individual chart: UCL  60.327362, CL  60.32641, LCL  60.325458 MR chart: UCL  0.001169, CL  0.000358, LCL  0 (b) Individual chart: UCL  0.001362, CL  00.00041, LCL  0.000542 MR chart: UCL  0.001169, CL  0.000358, LCL  0 (c) ˆ  60.3264 ˆ  0.0003173 PCR  1.0505 PCRk  0.9455 15-79. (a) Trial control limits S chart: UCL  170.25, CL  86.42, LCL  2.59 x chart: UCL  670.00, CL  558.77, LCL  447.53 Revised S chart: UCL  158.93, CL  80.68, LCL  2.42

JWCL232_AppB_731-748.qxd

748

1/23/10

10:18 AM

Page 748

APPENDIX B ANSWERS TO SELECTED EXERCISES

(b) (c) (d) 15-89. (a) (b) (c) 15-91. (a)

(b)

x chart: UCL  655.79, CL  551.95, LCL  448.10 PCRK  0.8725 ˆ  36.9917 ARL  18.6 ARL  43.9 ARL  6.30 ARL  2.00 x chart: UCL  140.168, CL  139.49, LCL  138.812 R chart: UCL  2.48437, CL  1.175, LCL  0 Revised: x chart: UCL  140.417, CL  139.709, LCL  139.001

R chart: UCL  2.596, CL  1.227, LCL  0 ˆ  0.5276 (c) PCR  1.26 PCRk  1.08 (d) 2  0.081 (e) ARL  5.55 15-93. P1X  USL2  0.00135 15-95. (a) P1U  12.24 when ␭  16)  0.96995 (b) P1U  10.68 when ␭  16)  1 15-97. (a) ARL  17.5 (b) ARL  3.63 15-101. (a) ˆ  3.0411, probability  0.03 (b) 31.95

Mind-Expanding Exercises 15-103. 0.125, 0.004 15-105. (b) ARL  1p where p  1 1k 1n2

1 k 1n2 (c) ARL  22.0 for k  2 (d) ARL  4.47 k2 11 p2 15-107. n  p 15-111. (a) C chart CL  8 UCL  16.49 LCL  0 (b) Yes

JWCL232_AppC_749-750.qxd

1/22/10

9:50 PM

Page 749

Appendix C Bibliography INTRODUCTORY WORKS AND GRAPHICAL METHODS Chambers, J., Cleveland, W., Kleiner, B., and P. Tukey (1983), Graphical Methods for Data Analysis, Wadsworth & Brooks/Cole, Pacific Grove, CA. A very well-written presentation of graphical methods in statistics. Freedman, D., Pisani, R., Purves R., and A. Adbikari (1991), Statistics, 2nd ed., Norton, New York. An excellent introduction to statistical thinking, requiring minimal mathematical background. Hoaglin, D., Mosteller, F., and J. Tukey (1983), Understanding Robust and Exploratory Data Analysis, John Wiley & Sons, New York. Good discussion and illustration of techniques such as stem-and-leaf displays and box plots. Tanur, J., et al. (eds.) (1989), Statistics: A Guide to the Unknown, 3rd edition, Wadsworth & Brooks/Cole, Pacific Grove, CA. Contains a collection of short nonmathematical articles describing different applications of statistics. Tukey, J. (1977), Exploratory Data Analysis, Addison-Wesley, Reading, MA. Introduces many new descriptive and analytical methods. Not extremely easy to read.

PROBABILITY Hoel, P. G., Port, S. C., and C. J. Stone (1971), Introduction to Probability Theory, Houghton Mifflin, Boston. A wellwritten and comprehensive treatment of probability theory and the standard discrete and continuous distributions. Olkin, I., Derman, C., and L. Gleser (1994), Probability Models and Applications, 2nd ed., Macmillan, New York. A comprehensive treatment of probability at a higher mathematical level than this book. Mosteller, F., Rourke, R., and G. Thomas (1970), Probability with Statistical Applications, 2nd ed., Addison-Wesley,

Reading, MA. A precalculus introduction to probability with many excellent examples. Ross, S. (2005), A First Course in Probability, 7th ed., Prentice-Hall, Englewood Cliffs, NJ. More mathematically sophisticated than this book, but has many excellent examples and exercises.

MATHEMATICAL STATISTICS Efron, B., and R. Tibshirani (1993), An Introduction to the Bootstrap, Chapman and Hall, New York. An important reference on this useful but computer-intensive technique. Hoel, P. G. (1984), Introduction to Mathematical Statistics, 5th ed., John Wiley & Sons, New York. An outstanding introductory book, well written, and generally easy to understand. Hogg, R., and A. Craig (2004), Introduction to Mathematical Statistics, 6th ed., Prentice-Hall, Englewood Cliffs, NJ. Another classic work on the mathematical principles of statistics; higher level than the Hoel book, but contains excellent discussions of estimation and hypothesis testing. Larsen, R., and M. Marx (1986), Introduction to Mathematical Statistics, 2nd ed., Prentice-Hall, Englewood Cliffs, NJ. Written at a relatively low mathematical level, very readable. Larson, H. J. (1982), Introduction to Probability Theory and Statistical Inference, 3rd ed., John Wiley & Sons, New York. An extremely well-written book that gives broad coverage to many aspects of probability and mathematical statistics.

ENGINEERING STATISTICS Devore, J. L. (2008), Probability and Statistics for Engineering and the Sciences, 7th ed., Duxbury & Brooks/Cole, Pacific

749

JWCL232_AppC_749-750.qxd

750

1/22/10

9:50 PM

Page 750

APPENDIX C BIBLIOGRAPHY

Grove, CA. Covers many of the same topics as this text, but at a slightly higher mathematical level. Many of the examples and exercises involve applications to biological and life sciences. Hines, W. W., and D. C. Montgomery (1990), Probability and Statistics in Engineering and Management Science, 3rd ed., John Wiley & Sons, New York. Covers many of the same topics as this book. More emphasis on probability and a higher mathematical level. Ross, S. (1987), Introduction to Probability and Statistics for Engineers and Scientists, John Wiley & Sons, New York. More tightly written and mathematically oriented than this book, but contains some good examples. Walpole, R. E., Myers, R. H., and S. L. Myers (2002), Probability and Statistics for Engineers and Scientists, 7th ed., Prentice-Hall, Inc., Upper Saddle River, New Jersey. A very well-written book at about the same level as this one.

REGRESSION ANALYSIS Daniel, C., and F. Wood (1980), Fitting Equations to Data, 2nd ed., John Wiley & Sons, New York. An excellent reference containing many insights on data analysis. Draper, N., and H. Smith (1998), Applied Regression Analysis, 3rd ed., John Wiley & Sons, New York. A comprehensive book on regression written for statistically oriented readers. Kutner, Nachtsheim, Neter, Li (2005), Applied Linear Statistical Models, 4th ed., McGraw-Hill/Irwin, Columbus, OH. The first part of the book is an introduction to simple and multiple linear regression. The orientation is to business and economics. Montgomery, D. C., Peck, E. A., and G. G. Vining (2006), Introduction to Linear Regression Analysis, 4th ed., John Wiley & Sons, New York. A comprehensive book on regression written for engineers and physical scientists. Myers, R. H. (1990), Classical and Modern Regression with Applications, 2nd ed., PWS-Kent, Boston. Contains many examples with annotated SAS output. Very well written.

DESIGN OF EXPERIMENTS Box, G. E. P., Hunter, W. G., and J. S. Hunter (2005), Statistics for Experimenters, 2nd ed., John Wiley & Sons, New York. An excellent introduction to the subject for those readers desiring a statistically oriented treatment. Contains many useful suggestions for data analysis. Mason, R. L., Gunst, R. F., and J. F. Hess (2003), Statistical Design and Analysis of Experiments, 2nd ed., John Wiley & Sons, New York. A comprehensive book covering basic statistics, hypothesis testing and confidence intervals, elementary aspects of experimental design, and regression analysis.

Montgomery, D. C. (2009), Design and Analysis of Experiments, 7th ed., John Wiley & Sons, New York. Written at the same level as the Box, Hunter, and Hunter book, but focused on engineering applications.

NONPARAMETRIC STATISTICS Conover, W. J. (1998), Practical Nonparametric Statistics, 3rd ed., John Wiley & Sons, New York. An excellent exposition of the methods of nonparametric statistics; many good examples and exercises. Hollander, M., and D. Wolfe (1999), Nonparametric Statistical Methods, 2nd ed., John Wiley & Sons, New York. A good reference book, with a very useful set of tables.

STATISTICAL QUALITY CONTROL AND RELATED METHODS Duncan, A. J. (1986), Quality Control and Industrial Statistics, 5th ed., Richard D. Irwin, Homewood, Illinois. A classic book on the subject. Grant, E. L., and R. S. Leavenworth (1996), Statistical Quality Control, 7th ed., McGraw-Hill, New York. One of the first books on the subject; contains many good examples. John, P. W. M. (1990), Statistical Methods in Engineering and Quality Improvement, John Wiley & Sons, New York. Not a methods book, but a well-written presentation of statistical methodology for quality improvement. Montgomery, D. C. (2009), Introduction to Statistical Quality Control, 6th ed., John Wiley & Sons, New York. A modern comprehensive treatment of the subject written at the same level as this book. Nelson, W. (2003), Applied Life Data Analysis, John Wiley & Sons, New York. Contains many examples of using statistical methods for the study of failure data; a good reference for the statistical aspects of reliability engineering and the special probability distributions used in that field. Ryan, T. P. (2000), Statistical Methods for Quality Improvement, 2nd ed., John Wiley & Sons, New York. Gives broad coverage of the field, with some emphasis on newer techniques. Wadsworth, H. M., Stephens, K. S., and A. B. Godfrey (2001), Modern Methods for Quality Control and Improvement, 2nd ed., John Wiley & Sons, New York. A comprehensive treatment of statistical methods for quality improvement at a somewhat higher level than this book. Western Electric Company (1956), Statistical Quality Control Handbook, Western Electric Company, Inc., Indianapolis, Indiana. An oldie but a goodie.

JWCL232_Glo_751-761.qxd

1/22/10

11:33 PM

Page 751

Glossary 2k factorial experiment. A full factorial experiment with k factors and all factors tested at only two levels (settings) each. 2k-p factorial experiment. A fractional factorial experiment with k factors tested in a 2p fraction with all factors tested at only two levels (settings) each. Acceptance region. In hypothesis testing, a region in the sample space of the test statistic such that if the test statistic falls within it, the null hypothesis cannot be rejected. This terminology is used because rejection of H0 is always a strong conclusion and acceptance of H0 is generally a weak conclusion. Addition rule. A formula used to determine the probability of the union of two (or more) events from the probabilities of the events and their intersection(s). Additivity property of 2. If two independent random variables X1 and X2 are distributed as chi-square with v1 and v2 degrees of freedom, respectively, Y  X1  X2 is a chi-square random variable with u  v1  v2 degrees of freedom. This generalizes to any number of independent chi-square random variables. Adjusted R2. A variation of the R2 statistic that compensates for the number of parameters in a regression model. Essentially, the adjustment is a penalty for increasing the number of parameters in the model. Alias. In a fractional factorial experiment when certain factor effects cannot be estimated uniquely, they are said to be aliased. All possible (subsets) regressions. A method of variable selection in regression that examines all possible subsets of the candidate regressor variables. Efficient computer algorithms have been developed for implementing all possible regressions. Alternative hypothesis. In statistical hypothesis testing, this is a hypothesis other than the one that is being tested. The alternative hypothesis contains feasible conditions, whereas the null hypothesis specifies conditions that are under test. Analysis of variance (ANOVA). A method of decomposing the total variability in a set of observations, as measured by the sum of the squares of these observations from their average, into component sums of squares that are associated with specific defined sources of variation. Analytic study. A study in which a sample from a population is used to make inference to a future population. Stability needs to be assumed. See Enumerative study.

Arithmetic mean. The arithmetic mean of a set of numbers x1, x2, …, xn is their sum divided by the number of observations, or n 11 n2 g i1xi. The arithmetic mean is usually denoted by x, and is often called the average. Assignable cause. The portion of the variability in a set of observations that can be traced to specific causes, such as operators, materials, or equipment. Also called a special cause. Asymptotic relative efficiency (ARE). Used to compare hypothesis tests. The ARE of one test relative to another is the limiting ratio of the sample sizes necessary to obtain identical error probabilities for the two procedures. Attribute. A qualitative characteristic of an item or unit, usually arising in quality control. For example, classifying production units as defective or nondefective results in attributes data. Attribute control chart. Any control chart for a discrete random variable. See Variables control chart. Average. See Arithmetic mean. Average run length, or ARL. The average number of samples taken in a process monitoring or inspection scheme until the scheme signals that the process is operating at a level different from the level in which it began. Axioms of probability. A set of rules that probabilities defined on a sample space must follow. See Probability. Backward elimination. A method of variable selection in regression that begins with all of the candidate regressor variables in the model and eliminates the insignificant regressors one at a time until only significant regressors remain. Bayes estimator. An estimator for a parameter obtained from a Bayesian method that uses a prior distribution for the parameter along with the conditional distribution of the data given the parameter to obtain the posterior distribution of the parameter. The estimator is obtained from the posterior distribution. Bayes’ theorem. An equation for a conditional probability such as P1A 0 B2 in terms of the reverse conditional probability P1B 0 A2 . Bernoulli trials. Sequences of independent trials with only two outcomes, generally called “success” and “failure,” in which the probability of success remains constant. Bias. An effect that systematically distorts a statistical result or estimate, preventing it from representing the true quantity of interest. Biased estimator. See Unbiased estimator.

751

JWCL232_Glo_751-761.qxd

752

1/22/10

11:33 PM

Page 752

GLOSSARY

Bimodal distribution. A distribution with two modes. Binomial random variable. A discrete random variable that equals the number of successes in a fixed number of Bernoulli trials.

distribution and (2) testing goodness of fit of a theoretical distribution to sample data. Coefficient of determination. See R2.

Bivariate distribution. The joint probability distribution of two random variables.

Combination. A subset selected without replacement from a set used to determine the number of outcomes in events and sample spaces.

Bivariate normal distribution. The joint distribution of two normal random variables.

Comparative experiment. An experiment in which the treatments (experimental conditions) that are to be studied are included in the experiment. The data from the experiment are used to evaluate the treatments.

Block. In experimental design, a group of experimental units or material that is relatively homogeneous. The purpose of dividing experimental units into blocks is to produce an experimental design wherein variability within blocks is smaller than variability between blocks. This allows the factors of interest to be compared in a environment that has less variability than in an unblocked experiment.

Completely randomized design (or experiment). A type of experimental design in which the treatments or design factors are assigned to the experimental units in a random manner. In designed experiments, a completely randomized design results from running all of the treatment combinations in random order.

Box plot (or box and whisker plot). A graphical display of data in which the box contains the middle 50% of the data (the interquartile range) with the median dividing it, and the whiskers extend to the smallest and largest values (or some defined lower and upper limits).

Components of variance. The individual components of the total variance that are attributable to specific sources. This usually refers to the individual variance components arising from a random or mixed model analysis of variance.

C chart. An attribute control chart that plots the total number of defects per unit in a subgroup. Similar to a defects-per-unit or U chart.

Conditional mean. The mean of the conditional probability distribution of a random variable.

Categorical data. Data consisting of counts or observations that can be classified into categories. The categories may be descriptive.

Conditional probability. The probability of an event given that the random experiment produces an outcome in another event.

Causal variable. When y  f(x) and y is considered to be caused by x, x is sometimes called a causal variable. Cause-and-effect diagram. A chart used to organize the various potential causes of a problem. Also called a fishbone diagram. Center line. A horizontal line on a control chart at the value that estimates the mean of the statistic plotted on the chart. See Control chart. Central composite design (CCD). A second-order response surface design in k variables consisting of a two-level factorial, 2k axial runs, and one or more center points. The two-level factorial portion of a CCD can be a fractional factorial design when k is large. The CCD is the most widely used design for fitting a second-order model. Central limit theorem. The simplest form of the central limit theorem states that the sum of n independently distributed random variables will tend to be normally distributed as n becomes large. It is a necessary and sufficient condition that none of the variances of the individual random variables are large in comparison to their sum. There are more general forms of the central theorem that allow infinite variances and correlated random variables, and there is a multivariate version of the theorem. Central tendency. The tendency of data to cluster around some value. Central tendency is usually expressed by a measure of location such as the mean, median, or mode. Chance cause. The portion of the variability in a set of observations that is due to only random forces and which cannot be traced to specific sources, such as operators, materials, or equipment. Also called a common cause. Chi-square (or chi-squared) random variable. A continuous random variable that results from the sum of squares of independent standard normal random variables. It is a special case of a gamma random variable. Chi-square test. Any test of significance based on the chi-square distribution. The most common chi-square tests are (1) testing hypotheses about the variance or standard deviation of a normal

Conditional probability density function. The probability density function of the conditional probability distribution of a continuous random variable. Conditional probability distribution. The distribution of a random variable given that the random experiment produces an outcome in an event. The given event might specify values for one or more other random variables. Conditional probability mass function. The probability mass function of the conditional probability distribution of a discrete random variable. Conditional variance. The variance of the conditional probability distribution of a random variable. Confidence coefficient. The probability 1   associated with a confidence interval expressing the probability that the stated interval will contain the true parameter value. Confidence interval. If it is possible to write a probability statement of the form P1L  U 2  1   where L and U are functions of only the sample data and  is a parameter, then the interval between L and U is called a confidence interval (or a 100(1  )% confidence interval). The interpretation is that a statement that the parameter  lies in this interval will be true 100(1  )% of the times that such a statement is made. Confidence level. Another term for the confidence coefficient. Confounding. When a factorial experiment is run in blocks and the blocks are too small to contain a complete replicate of the experiment, one can run a fraction of the replicate in each block, but this results in losing information on some effects. These effects are linked with or confounded with the blocks. In general, when two factors are varied such that their individual effects cannot be determined separately, their effects are said to be confounded.

JWCL232_Glo_751-761.qxd

1/22/10

11:33 PM

Page 753

GLOSSARY

753

Consistent estimator. An estimator that converges in probability to the true value of the estimated parameter as the sample size increases.

Counting techniques. Formulas used to determine the number of elements in sample spaces and events.

Contingency table. A tabular arrangement expressing the assignment of members of a data set according to two or more categories or classification criteria.

Covariance. A measure of association between two random variables obtained as the expected value of the product of the two random variables around their means; that is, Cov(X, Y )  E[(X  X)(Y  Y)].

Continuity correction. A correction factor used to improve the approximation to binomial probabilities from a normal distribution.

Covariance matrix. A square matrix that contains the variances and covariances among a set of random variables, say, X1, X2, p , Xk. The main diagonal elements of the matrix are the variances of the random variables and the off-diagonal elements are the covariances between Xi and Xj. Also called the variance-covariance matrix. When the random variables are standardized to have unit variances, the covariance matrix becomes the correlation matrix.

Continuous distribution. A probability distribution for a continuous random variable. Continuous random variable. A random variable with an interval (either finite or infinite) of real numbers for its range. Continuous uniform random variable. A continuous random variable with range of a finite interval and a constant probability density function. Contour plot. A two-dimensional graphic used for a bivariate probability density function that displays curves for which the probability density function is constant. Contrast. A linear function of treatment means with coefficients that total zero. A contrast is a summary of treatment means that is of interest in an experiment. Control chart. A graphical display used to monitor a process. It usually consists of a horizontal center line corresponding to the incontrol value of the parameter that is being monitored and lower and upper control limits. The control limits are determined by statistical criteria and are not arbitrary, nor are they related to specification limits. If sample points fall within the control limits, the process is said to be in-control, or free from assignable causes. Points beyond the control limits indicate an out-of-control process; that is, assignable causes are likely present. This signals the need to find and remove the assignable causes. Control limits. See Control chart. Convolution. A method to derive the probability density function of the sum of two independent random variables from an integral (or sum) of probability density (or mass) functions. Cook’s distance. In regression, Cook’s distance is a measure of the influence of each individual observation on the estimates of the regression model parameters. It expresses the distance that the vector of model parameter estimates with the ith observation removed lies from the vector of model parameter estimates based on all observations. Large values of Cook’s distance indicate that the observation is influential.

Correction factor. A term used for the quantity 11 n2 1 g i1 xi 2 2 that is n subtracted from g i1 x2i to give the corrected sum of squares defined as n 2 11n2 g i1 1xi  x2 . The correction factor can also be written as nx 2. n

Correlation. In the most general usage, a measure of the interdependence among data. The concept may include more than two variables. The term is most commonly used in a narrow sense to express the relationship between quantitative variables or ranks. Correlation coefficient. A dimensionless measure of the linear association between two variables, usually lying in the interval from 1 to 1, with zero indicating the absence of correlation (but not necessarily the independence of the two variables). Correlation matrix. A square matrix that contains the correlations among a set of random variables, say, X1, X2, p , Xk. The main diagonal elements of the matrix are unity and the off-diagonal elements rij are the correlations between Xi and Xj.

Critical region. In hypothesis testing, this is the portion of the sample space of a test statistic that will lead to rejection of the null hypothesis. Critical value(s). The value of a statistic corresponding to a stated significance level as determined from the sampling distribution. For example, if P(Z  z0.05)  P(Z  1.96)  0.05, then z0.05 = 1.96 is the critical value of z at the 0.05 level of significance. Crossed factors. Another name for factors that are arranged in a factorial experiment. Cumulative distribution function. For a random variable X, the function of X defined as P(X x) that is used to specify the probability distribution. Cumulative normal distribution function. The cumulative distribution of the standard normal distribution, often denoted as (x) and tabulated in Appendix Table II. Cumulative sum control chart (CUSUM). A control chart in which the point plotted at time t is the sum of the measured deviations from target for all statistics up to time t. Curvilinear regression. An expression sometimes used for nonlinear regression models or polynomial regression models. Decision interval. A parameter in a tabular CUSUM algorithm that is determined from a trade-off between false alarms and the detection of assignable causes. Defect. Used in statistical quality control, a defect is a particular type of nonconformance to specifications or requirements. Sometimes defects are classified into types, such as appearance defects and functional defects. Defect concentration diagram. A quality tool that graphically shows the location of defects on a part or in a process. Defects-per-unit control chart. See U chart. Defining relation. A subset of effects in a fractional factorial design that define the aliases in the design. Degrees of freedom. The number of independent comparisons that can be made among the elements of a sample. The term is analogous to the number of degrees of freedom for an object in a dynamic system, which is the number of independent coordinates required to determine the motion of the object. Deming. W. Edwards Deming (1900–1993) was a leader in the use of statistical quality control. Deming’s 14 points. A management philosophy promoted by W. Edwards Deming that emphasizes the importance of change and quality.

JWCL232_Glo_751-761.qxd

754

1/22/10

11:33 PM

Page 754

GLOSSARY

Density function. Another name for a probability density function.

Error variance. The variance of an error term or component in a model.

Dependent variable. The response variable in regression or a designed experiment.

Estimate (or point estimate). The numerical value of a point estimator.

Design matrix. A matrix that provides the tests that are to be conducted in an experiment.

Estimator (or point estimator). A procedure for producing an estimate of a parameter of interest. An estimator is usually a function of only sample data values, and when these data values are available, it results in an estimate of the parameter of interest.

Designed experiment. An experiment in which the tests are planned in advance and the plans usually incorporate statistical models. See Experiment.

Event. A subset of a sample space.

Discrete distribution. A probability distribution for a discrete random variable.

Exhaustive. A property of a collection of events that indicates that their union equals the sample space.

Discrete random variable. A random variable with a finite (or countably infinite) range.

Expected value. The expected value of a random variable X is its long-term average or mean value. In the continuous case, the expected  xf 1x2 dx where f (x) is the density function of value of X is E1X 2  兰 the random variable X.

Discrete uniform random variable. A discrete random variable with a finite range and constant probability mass function. Dispersion. The amount of variability exhibited by data. Distribution free method(s). Any method of inference (hypothesis testing or confidence interval construction) that does not depend on the form of the underlying distribution of the observations. Sometimes called nonparametric method(s).

Experiment. A series of tests in which changes are made to the system under study. Exponential random variable. A continuous random variable that is the time between events in a Poisson process.

Distribution function. Another name for a cumulative distribution function.

Extra sum of squares method. A method used in regression analysis to conduct a hypothesis test for the additional contribution of one or more variables to a model.

Efficiency. A concept in parameter estimation that uses the variances of different estimators; essentially, an estimator is more efficient than another estimator if it has smaller variance. When estimators are biased, the concept requires modification.

Factorial experiment. A type of experimental design in which every level of one factor is tested in combination with every level of another factor. In general, in a factorial experiment, all possible combinations of factor levels are tested.

Empirical model. A model to relate a response to one or more regressors or factors that is developed from data obtained from the system.

False alarm. A signal from a control chart when no assignable causes are present.

Enumerative study. A study in which a sample from a population is used to make inference to the population. See Analytic study.

F distribution. The distribution of the random variable defined as the ratio of two independent chi-square random variables, each divided by its number of degrees of freedom.

Erlang random variable. A continuous random variable that is the sum of a fixed number of independent, exponential random variables.

Finite population correction factor. A term in the formula for the variance of a hypergeometric random variable.

␤-error (or ␤-risk). In hypothesis testing, an error incurred by failing to reject a null hypothesis when it is actually false (also called a type II error).

First-order model. A model that contains only first-order terms. For example, the first-order response surface model in two variables is y  0  1x1  2x2  . A first-order model is also called a main effects model.

␣-error (or ␣-risk). In hypothesis testing, an error incurred by rejecting a null hypothesis when it is actually true (also called a type I error). Error mean square. The error sum of squares divided by its number of degrees of freedom. Error of estimation. The difference between an estimated value and the true value. Error propagation. An analysis of how the variance of the random variable that represents that output of a system depends on the variances of the inputs. A formula exists when the output is a linear function of the inputs and the formula is simplified if the inputs are assumed to be independent. Error sum of squares. In analysis of variance, this is the portion of total variability that is due to the random component in the data. It is usually based on replication of observations at certain treatment combinations in the experiment. It is sometimes called the residual sum of squares, although this is really a better term to use only when the sum of squares is based on the remnants of a model-fitting process and not on replication.

Fisher’s least significant difference (LSD) method. A series of pair-wise hypothesis tests of treatment means in an experiment to determine which means differ. Fixed factor (or fixed effect). In analysis of variance, a factor or effect is considered fixed if all the levels of interest for that factor are included in the experiment. Conclusions are then valid about this set of levels only, although when the factor is quantitative, it is customary to fit a model to the data for interpolating between these levels. Forward selection. A method of variable selection in regression, where variables are inserted one at a time into the model until no other variables that contribute significantly to the model can be found. Fraction defective control chart. See P chart. Fraction defective. In statistical quality control, that portion of a number of units or the output of a process that is defective. Fractional factorial experiment. A type of factorial experiment in which not all possible treatment combinations are run. This is usually done to reduce the size of an experiment with several factors.

JWCL232_Glo_751-761.qxd

1/22/10

11:33 PM

Page 755

GLOSSARY Frequency distribution. An arrangement of the frequencies of observations in a sample or population according to the values that the observations take on. F-test. Any test of significance involving the F distribution. The most common F-tests are (1) testing hypotheses about the variances or standard deviations of two independent normal distributions, (2) testing hypotheses about treatment means or variance components in the analysis of variance, and (3) testing significance of regression or tests on subsets of parameters in a regression model. Gamma function. A function used in the probability density function of a gamma random variable that can be considered to extend factorials. Gamma random variable. A random variable that generalizes an Erlang random variable to noninteger values of the parameter r. Gaussian distribution. Another name for the normal distribution, based on the strong connection of Karl F. Gauss to the normal distribution; often used in physics and electrical engineering applications. Generating function. A function that is used to determine properties of the probability distribution of a random variable. See Moment generating function. Generator. Effects in a fractional factorial experiment that are used to construct the experimental tests used in the experiment. The generators also define the aliases. Geometric mean. The geometric mean of a set of n positive data values is the nth root of the product of the data values; that is, n g  1 w i1 xi 2 1n. Geometric random variable. A discrete random variable that is the number of Bernoulli trials until a success occurs. Goodness of fit. In general, the agreement of a set of observed values and a set of theoretical values that depend on some hypothesis. The term is often used in fitting a theoretical distribution to a set of observations. Harmonic mean. The harmonic mean of a set of data values is the reciprocal of the arithmetic mean of the reciprocals of the data 1 n 1 1 values; that is, h  a g i1 b . n xi Hat matrix. In multiple regression, the matrix H  X1X¿X2 1X¿ . This a projection matrix that maps the vector of observed response values into a vector of fitted values by yˆ  X1X¿X2 1X¿y  Hy. Hidden extrapolation. An extrapolation is a prediction in a regression analysis that is made at point (x1, x2, …, xk) that is remote from the data used to generate the model. Hidden extrapolation occurs when it is not obvious that the point is remote. This can occur when multicollinearity is present in the data used to construct the model. Histogram. A univariate data display that uses rectangles proportional in area to class frequencies to visually exhibit features of data such as location, variability, and shape. Homogeneity test. In a two-way (r by c) contingency table, this tests if the proportions in the c categories are the same for all r populations. Hypergeometric random variable. A discrete random variable that is the number of success obtained from a sample drawn without replacement from a finite populations. Hypothesis (as in statistical hypothesis). A statement about the parameters of a probability distribution or a model, or a statement about the form of a probability distribution.

755

Hypothesis testing. Any procedure used to test a statistical hypothesis. Independence. A property of a probability model and two (or more) events that allows the probability of the intersection to be calculated as the product of the probabilities. Independence test. In a two-way (r by c) contingency table, this tests if the row and column categories are independent. Independent random variables. Random variables for which P(X A, Y B)  P(X A)P(Y B) for any sets A and B in the range of X and Y, respectively. There are several equivalent descriptions of independent random variables. Independent variable. The predictor or regressor variables in a regression model. Inference. Conclusion from a statistical analysis. It usually refers to the conclusion from a hypothesis test or an interval estimate. Indicator variable(s). Variables that are assigned numerical values to identify the levels of a qualitative or categorical response. For example, a response with two categorical levels ( yes and no) could be represented with an indicator variable taking on the values 0 and 1. Individuals control chart. A Shewhart control chart in which each plotted point is an individual measurement, rather than a summary statistic. See Control chart, Shewhart control chart. Influential observation. An observation in a regression analysis that has a large effect on estimated parameters in the model. Influence is measured by the change in parameters when the influential observation is included and excluded in the analysis. Interaction. In factorial experiments, two factors are said to interact if the effect of one variable is different at different levels of the other variables. In general, when variables operate independently of each other, they do not exhibit interaction. Intercept. The constant term in a regression model. Interquartile range. The difference between the third and first quartiles in a sample of data. The interquartile range is less sensitive to extreme data values than the usual sample range. Interval estimation. The estimation of a parameter by a range of values between lower and upper limits, in contrast to point estimation, where the parameter is estimated by a single numerical value. A confidence interval is a typical interval estimation procedure. Intrinsically linear model. In regression analysis, a nonlinear function that can be expressed as a linear function after a suitable transformation is called intrinsically linear. Jacobian. A matrix of partial derivatives that is used to determine the distribution of transformed random variables. Joint probability density function. A function used to calculate probabilities for two or more continuous random variables. Joint probability distribution. The probability distribution for two or more random variables in a random experiment. See Joint probability mass function and Joint probability density function. Joint probability mass function. A function used to calculate probabilities for two or more discrete random variables. Kurtosis. A measure of the degree to which a unimodal distribution is peaked.

JWCL232_Glo_751-761.qxd

756

1/22/10

11:33 PM

Page 756

GLOSSARY

Lack of memory property. A property of a Poisson process. The probability of a count in an interval depends only on the length of the interval (and not on the starting point of the interval). A similar property holds for a series of Bernoulli trials. The probability of a success in a specified number of trials depends only on the number of trials (and not on the starting trial). Least significance difference test (or Fisher’s LSD test). An application of the t-test to compare pairs of means following rejection of the null hypothesis in an analysis of variance. The error rate is difficult to calculate exactly because the comparisons are not all independent.

Logistic regression. A regression model that is used to model a categorical response. For a binary (0, 1) response, the model assumes that the logarithm of the ratio of probabilities (for zero and one) is linearly related to the regressor variables. Lognormal random variable. A continuous random variable with probability distribution equal to that of exp(W ) for a normal random variable W. Main effect. An estimate of the effect of a factor (or variable) that independently expresses the change in response due to a change in that factor, regardless of other factors that may be present in the system.

Least squares (method of). A method of parameter estimation in which the parameters of a system are estimated by minimizing the sum of the squares of the differences between the observed values and the fitted or predicted values from the system.

Marginal probability density function. The probability density function of a continuous random variable obtained from the joint probability distribution of two or more random variables.

Least squares estimator. Any estimator obtained by the method of least squares.

Marginal probability distribution. The probability distribution of a random variable obtained from the joint probability distribution of two or more random variables.

Level of significance. If Z is the test statistic for a hypothesis, and the distribution of Z when the hypothesis is true are known, then we can find the probabilities P(Z zL) and P(Z  zU). Rejection of the hypothesis is usually expressed in terms of the observed value of Z falling outside the interval from zL to zU. The probabilities P(Z zL) and P(Z  zU) are usually chosen to have small values, such as 0.01, 0.025, 0.05, or 0.10, and are called levels of significance. The actual levels chosen are somewhat arbitrary and are often expressed in percentages, such as a 5% level of significance. Levels of a factor. The settings (or conditions) used for a factor in an experiment. Likelihood function. Suppose that the random variables X1, X2, p , Xn have a joint distribution given by f(x1, x2, p , xn; 1, 2, p , p ) where the s are unknown parameters. This joint distribution, considered as a function of the s for fixed x’s, is called the likelihood function. Likelihood principle. This principle states that the information about a model given by a set of data is completely contained in the likelihood. Likelihood ratio. Let x1, x2, p , xn be a random sample from the population f (x; ). The likelihood function for this sample is n L  w i1 f 1xi; 2. We wish to test the hypothesis H0:  , where  is a subset of the possible values  for . Let the maximum value of L with respect to  over the entire set of values that the parameter can ˆ 2 , and let the maximum value of L with  retake on be denoted by L1 stricted to the set of values given by  be L( ˆ ). The null hypothesis is ˆ 2 , or a simple ˆ 2  L1 tested by using the likelihood ratio ␭  L1 function of it. Large values of the likelihood ratio are consistent with the null hypothesis. Likelihood ratio test. A test of a null hypothesis versus an alternative hypothesis using a test statistic derived from a likelihood ratio. Linear function of random variables. A random variable that is defined as a linear function of several random variables. Linear model. A model in which the observations are expressed as a linear function of the unknown parameters. For example, y  0 

1x   and y  0  1x  2 x2   are linear models. Location parameter. A parameter that defines a central value in a sample or a probability distribution. The mean and the median are location parameters.

Marginal probability mass function. The probability mass function of a discrete random variable obtained from the joint probability distribution of two or more random variables. Maximum likelihood estimation. A method of parameter estimation that maximizes the likelihood function of a sample. Mean. The mean usually refers either to the expected value of a random variable or to the arithmetic average of a set of data. Mean square. In general, a mean square is determined by dividing a sum of squares by the number of degrees of freedom associated with the sum of squares. Mean square(d) error. The expected squared deviation of an estimator from the true value of the parameter it estimates. The mean square error can be decomposed into the variance of the estimator ˆ 2  E1 ˆ  2 2  plus the square of the bias; that is, MSE1 2 ˆ ˆ V1 2  3E1 2   4 . Mechanistic model. A model developed from theoretical knowledge or experience in contrast to a model developed from data. See Empirical model. Median. The median of a set of data is that value that divides the data into two equal halves. When the number of observations is even, say 2n, it is customary to define the median as the average of the nth and (n  1)st rank-ordered values. The median can also be defined for a random variable. For example, in the case of a continuous random variM able X, the median M can be defined as 兰 f 1x2 dx  兰M f 1x2 dx  1  2. Method of steepest ascent. A technique that allows an experimenter to move efficiently toward a set of optimal operating conditions by following the gradient direction. The method of steepest ascent is usually employed in conjunction with fitting a first-order response surface and deciding that the current region of operation is inappropriate. Mixed model. In an analysis of variance context, a mixed model contains both random and fixed factors. Mode. The mode of a sample is that observed value that occurs most frequently. In a probability distribution f (x) with continuous first derivative, the mode is a value of x for which df (x)兾dx  0 and d2f (x)兾dx2  0. There may be more than one mode of either a sample or a distribution.

JWCL232_Glo_751-761.qxd

1/22/10

11:33 PM

Page 757

GLOSSARY

757

Moment (or population moment). The expected value of a function of a random variable such as E(X  c)r for constants c and r. When c  0, it is said that the moment is about the origin. See Moment generating function.

Nuisance factor. A factor that probably influences the response variable, but which is of no interest in the current study. When the levels of the nuisance factor can be controlled, blocking is the design technique that is customarily used to remove its effect.

Moment estimator. A method of estimating parameters by equating sample moments to population moments. Since the population moments will be functions of the unknown parameters, this results in equations that may be solved for estimates of the parameters.

Null distribution. In a hypothesis test, the distribution of the test statistic when the null hypothesized is assumed to be true.

Moment generating function. A function that is used to determine properties (such as moments) of the probability distribution of a random variable. It is the expected value of exp(tX). See Generating function and Moment. Moving range. The absolute value of the difference between successive observations in time-ordered data. Used to estimate chance variation in an individual control chart. Multicollinearity. A condition occurring in multiple regression where some of the predictor or regressor variables are nearly linearly dependent. This condition can lead to instability in the estimates of the regression model parameters. Multinomial distribution. The joint probability distribution of the random variables that count the number of results in each of k classes in a random experiment with a series of independent trials with constant probability of each class on each trial. It generalizes a binomial distribution. Multiplication rule. For probability, a formula used to determine the probability of the intersection of two (or more) events. For counting techniques, a formula used to determine the number of ways to complete an operation from the number of ways to complete successive steps. Mutually exclusive events. A collection of events whose intersections are empty. Natural tolerance limits. A set of symmetric limits that are three times the process standard deviation from the process mean. Negative binomial random variable. A discrete random variable that is the number of trials until a specified number of successes occur in Bernoulli trials. Nonlinear regression model. A regression model that is nonlinear in the parameters. It is sometimes applied to regression models that are nonlinear in the regressors or predictors, but this is an incorrect usage. Nonparametric method(s).

statistical

method(s). See

Distribution

free

Normal approximation. A method to approximate probabilities for binomial and Poisson random variables. Normal equations. The set of simultaneous linear equations arrived at in parameter estimation using the method of least squares. Normal probability plot. A specially constructed plot for a variable x (usually on the abscissa) in which y (usually on the ordinate) is scaled so that the graph of the normal cumulative distribution is a straight line. Normal random variable. A continuous random variable that is the most important one in statistics because it results from the central limit theorem. See Central limit theorem. NP chart. An attribute control chart that plots the total of defective units in a subgroup. Similar to a fraction-defective chart or P chart.

Null hypothesis. This term generally relates to a particular hypothesis that is under test, as distinct from the alternative hypothesis (which defines other conditions that are feasible but not being tested). The null hypothesis determines the probability of type I error for the test procedure. Observational study. A system is observed and data might be collected, but changes are not made to the system. See Experiment. Odds ratio. The odds equals the ratio of two probabilities. In logistic regression, the logarithm of the odds is modeled as a linear function of the regressors. Given values for the regressors at a point, the odds can be calculated. The odds ratio is the odds at one point divided by the odds at another. One-way model. In an analysis of variance context, this involves a single variable or factor with a different levels. Operating characteristic curves (OC curves). A plot of the probability of type II error versus some measure of the extent to which the null hypothesis is false. Typically, one OC curve is used to represent each sample size of interest. Optimization experiment. A experiment conducted to improve (or optimize) a system or process. It is assumed that the important factors are known. Orthogonal. There are several related meanings, including the mathematical sense of perpendicular, two variables being said to be orthogonal if they are statistically independent, or in experimental design where a design is orthogonal if it admits statistically independent estimates of effects. Orthogonal design. See Orthogonal. Outcome. An element of a sample space. Outlier(s). One or more observations in a sample that are so far from the main body of data that they give rise to the question that they may be from another population. Overcontrol. Unnecessary adjustments made to processes that increase the deviations from target. Overfitting. Adding more parameters to a model than is necessary. P chart. An attribute control chart that plots the proportion of defective units in a subgroup. Also called a fraction-defective control chart. Similar to an NP chart. Parameter estimation. The process of estimating the parameters of a population or probability distribution. Parameter estimation, along with hypothesis testing, is one of the two major techniques of statistical inference. Parameter. An unknown quantity that may vary over a set of values. Parameters occur in probability distributions and in statistical models, such as regression models. Pareto chart. A bar chart used to rank the causes of a problem. PCR. A process capability ratio with numerator equal to the difference between the product specification limits and denominator equal to six times the process standard deviation. Said to measure the potential

JWCL232_Glo_751-761.qxd

758

1/22/10

11:33 PM

Page 758

GLOSSARY

capability of the process because the process mean is not considered. See Process capability, Process capability ratio, Process capability study, and PCRk . Sometimes denoted as Cp in other references. PCRk . A process capability ratio with numerator equal to the difference between the product target and the nearest specification limit and denominator equal to three times the process standard deviation. Said to measure the actual capability of the process because the process mean is considered. See process capability, process capability ratio, process capability study, and PCR. Sometimes denoted as Cpk in other references. Percentage point. A particular value of a random variable determined from a probability (expressed as a percentage). For example, the upper 5 percentage point of the standard normal random variable is Z0.05  1.645. Percentile. The set of values that divide the sample into 100 equal parts. Permutation. An ordered sequence of the elements in a set used to determine the number of outcomes in events and sample spaces. Point estimator. See Estimator. Poisson process. A random experiment with events that occur in an interval and satisfy the following assumptions. The interval can be partitioned into subintervals such that the probability of more than one event in a subinterval is zero, the probability of an event in a subinterval is proportional to the length of the subinterval, and the event in each subinterval is independent of other subintervals. Poisson random variable. A discrete random variable that is the number of events that occur in a Poisson process. Pooled t-test. A hypothesis to compare the means of two populations with the variances assumed to be equal. Pooling. When several sets of data can be thought of as having been generated from the same model, it is possible to combine them, usually for purposes of estimating one or more parameters. Combining the samples for this purpose is usually called pooling and it is commonly used to estimate a variance.

PRESS statistic. In regression analysis, the predicted residual sum of squares. Delete each point and estimate the parameters of the model from the data that remain. Estimate the deleted point from this model. Restore the point and then delete the next point. Each point is estimated once and the sum of squares of these errors is calculated. Prior distribution. The initial probability distribution assumed for a parameter in a Bayesian analysis. Probability. A numerical measure between 0 and 1 assigned to events in a sample space. Higher numbers indicate the event is more likely to occur. See Axioms of probability. Probability density function. A function used to calculate probabilities and to specify the probability distribution of a continuous random variable. Probability distribution. For a sample space, a description of the set of possible outcomes along with a method to determine probabilities. For a random variable, a probability distribution is a description of the range along with a method to determine probabilities. Probability mass function. A function that provides probabilities for the values in the range of a discrete random variable. Probability plot. A scatter plot used to judge if data can reasonably be assumed to follow a particular probability distribution. A normal probability plot is often used to evaluate the normality assumption of data or residuals. Process capability. The capability of a process to produce product within specification limits. See Process capability ratio, Process capability study, PCR, and PCRk. Process capability ratio. A ratio that relates the width of the product specification limits to measures of process performance. Used to quantify the capability of the process to produce product within specifications. See Process capability, Process capability study, PCR, and PCRk. Process capability study. A study that collects data to estimate process capability. See Process capability, Process capability ratio, PCR, and PCRk.

Population variance. See Variance.

P-Value. The exact significance level of a statistical test; that is, the probability of obtaining a value of the test statistic that is at least as extreme as that observed when the null hypothesis is true.

Population. Any finite or infinite collection of individual units or objects.

Qualitative (data). Data derived from nonnumeric attributes, such as sex, ethnic origin or nationality, or other classification variables.

Posterior distribution. The probability distribution for a parameter in a Bayesian analysis calculated from the prior distribution and the conditional distribution of the data given the parameter.

Quality control. Systems and procedures used by an organization to assure that the outputs from processes satisfy customers.

Population standard deviation. See Standard deviation.

Power. The power of a statistical test is the probability that the test rejects the null hypothesis when the null hypothesis is indeed false. Thus, the power is equal to one minus the probability of type II error. Prediction. The process of determining the value of one or more statistical quantities at some future point in time. In a regression model, predicting the response y for some specified set of regressors or predictor variables also leads to a predicted value, although there may be no temporal element to the problem. Prediction interval. The interval between a set of upper and lower limits associated with a predicted value designed to show on a probability basis the range of error associated with the prediction. Predictor variable(s). The independent or regressor variable(s) in a regression model.

Quantiles. The set of n  1 values of a variable that partition it into a number n of equal proportions. For example, n  1  3 values partition data into four quantiles, with the central value usually called the median and the lower and upper values usually called the lower and upper quartiles, respectively. Quantitative (data). Data in the form of numerical measurements or counts. Quartiles. The three values of a variable that partition it into four equal parts. The central value is usually called the median and the lower and upper values are usually called the lower and upper quartiles, respectively. See Quantiles. R2. A quantity used in regression models to measure the proportion of total variability in the response accounted for by the model. Computationally, R2  SSRegression 兾SSTotal, and large values of R2 (near

JWCL232_Glo_751-761.qxd

1/22/10

11:33 PM

Page 759

GLOSSARY unity) are considered good. However, it is possible to have large values of R2 and find that the model is unsatisfactory. R2 is also called the coefficient of determination (or the coefficient of multiple determination in multiple regression).

759

Regression diagnostics. Techniques for examining a fitted regression model to investigate the adequacy of the fit and to determine if any of the underlying assumptions have been violated.

R chart. A control chart that plots the range of the measurements in a subgroup that is used to monitor the variance of the process.

Regression line (or curve). A graphical display of a regression model, usually with the response y on the ordinate and the regressor x on the abscissa.

Random. Nondeterministic, occurring purely by chance, or independent of the occurrence of other events.

Regression sum of squares. The portion of the total sum of squares attributable to the model that has been fit to the data.

Random effects model. In an analysis of variance context, this refers to a model that involves only random factors.

Regressor variable. The independent or predictor variable in a regression model.

Random error. An error (usually a term in a statistical model) that behaves as if it were drawn at random from a particular probability distribution.

Rejection region. In hypothesis testing, this is the region in the sample space of the test statistic that leads to rejection of the null hypothesis when the test statistic falls in this region.

Random experiment. An experiment that can result in different outcomes, even though it is repeated in the same manner each time.

Relative frequency. The relative frequency of an event is the proportion of times the event occurred in a series of trials of a random experiment.

Random factor. In analysis of variance, a factor whose levels are chosen at random from some population of factor levels. Random order. A sequence or order for a set of objects that is carried out in such a way that every possible ordering is equally likely. In experimental design, the runs of the experiment are typically arranged and carried out in random order. Random sample. A sample is said to be random if it is selected in such a way so that every possible sample has the same probability of being selected. Random variable. A function that assigns a real number to each outcome in the sample space of a random experiment. Randomization. Randomly assign treatments to experimental units or conditions in an experiment. This is done to reduce the opportunity for a treatment to be favored or disfavored (biased) by test conditions. Randomized complete block design. A type of experimental design in which treatment or factor levels are assigned to blocks in a random manner. Range. The largest minus the smallest of a set of data values. The range is a simple measure of variability and is widely used in quality control.

Reliability. The probability that a specified mission will be completed. It usually refers to the probability that a lifetime of a continuous random variable exceeds a specified time limit. Replicates. One of the independent repetitions of one or more treatment combinations in an experiment. Replication. The independent execution of an experiment more than once. Reproductive property of the normal distribution. A linear combination of independent, normal random variables is a normal random variable. Residual. Generally this is the difference between the observed and the predicted value of some variable. For example, in regression a residual is the difference between the observed value of the response and the corresponding predicted value obtained from the regression model. Residual analysis (and plots). Any technique that uses the residuals, usually to investigate the adequacy of the model that was used to generate the residuals. Residual sum of squares. See Error sum of squares.

Range (control) chart. A control chart used to monitor the variability (dispersion) in a process. See Control chart.

Resolution. A measure of severity of aliasing in a fractional factorial design. We commonly consider resolution III, IV, and V designs.

Rank. In the context of data, the rank of a single observation is its ordinal number when all data values are ordered according to some criterion, such as their magnitude.

Response (variable). The dependent variable in a regression model or the observed output variable in a designed experiment.

Rational subgroup. A sample of data selected in a manner to include chance sources of variation and to exclude assignable sources of variation to the extent possible. Reference distribution. The distribution of a test statistic when the null hypothesis is true. Sometimes a reference distribution is called the null distribution of the test statistic. Reference value. A parameter set in a tabular CUSUM algorithm that is determined from the magnitude of the process shift that should be detected. Regression. The statistical methods used to investigate the relationship between a dependent or response variable y and one or more independent variables x. The independent variables are usually called regressor variables or predictor variables. Regression coefficient(s). The parameter(s) in a regression model.

Response surface. When a response y depends on a function of k quantitative variables x1, x2, p , xk, the values of the response may be viewed as a surface in k  1 dimensions. This surface is called a response surface. Response surface methodology is a subset of experimental design concerned with approximating this surface with a model and using the resulting model to optimize the system or process. Response surface designs. Experimental designs that have been developed to work well in fitting response surfaces. These are usually designs for fitting a first- or second-order model. The central composite design is a widely used second-order response surface design. Ridge regression. A method for fitting a regression model that is intended to overcome the problems associated with using standard (or ordinary) least squares when there is a problem with multicollinearity in the data.

JWCL232_Glo_751-761.qxd

760

1/22/10

11:33 PM

Page 760

GLOSSARY

Rotatable design. In a rotatable design, the variance of the predicted response is the same at all points that are the same distance from the center of the design. Run rules. A set of rules applied to the points plotted on a Shewhart control chart that are used to make the chart more sensitized to assignable causes. See Control chart, Shewhart control chart. Runs test. A nonparametric test to compare two distributions or check for independence of measurements. S chart. A control chart that plots the standard deviation of the measurements in a subgroup that is used to monitor the variance of the process. Sample. Any subset of the elements of a population. Sample mean. The arithmetic average or mean of the observations in a sample. If the observations are x1, x2, p , xn, then the sample mean is n 11n2 g i1 xi. The sample mean is usually denoted by x.

Sample moment. The quantity 11n2 g moment.

n k i1 xi

is called the kth sample

Sample range. See Range. Sample size. The number of observations in a sample. Sample space. The set of all possible outcomes of a random experiment. Sample standard deviation. The positive square root of the sample variance. The sample standard deviation is the most widely used measure of variability of sample data. Sample variance. A measure of variability of sample data, defined as n s2  31 1n  12 4 g i1 1xi  x2 2, where x is the sample mean.

Six-sigma process. Originally used to describe a process with the mean at least six standard deviations from the nearest specification limits. It has now been used to describe any process with a defect rate of 3.4 parts per million. Skewness. A term for asymmetry usually employed with respect to a histogram of data or a probability distribution. Specification limits. Numbers that define the region of measurement for acceptable product. Usually there is an upper and lower limit, but one-sided limits can also be used. Standard deviation. The positive square root of the variance. The standard deviation is the most widely used measure of variability. Standard error. The standard deviation of the estimator of a parameter. The standard error is also the standard deviation of the sampling distribution of the estimator of a parameter. Standard normal random variable. A normal random variable with mean zero and variance one that has its cumulative distribution function tabulated in Appendix Table II. Standardize. The transformation of a normal random variable that subtracts its mean and divides by its standard deviation to generate a standard normal random variable. Standardized residual. In regression, the standardized residual is computed by dividing the ordinary residual by the square root of the residual mean square. This produces scaled residuals that have, approximately, a unit variance. Statistic. A summary value calculated from a sample of observations. Usually, a statistic is an estimator of some population parameter. Statistical inference. See Inference.

Sampling distribution. The probability distribution of a statistic. For example, the sampling distribution of the sample mean X is the normal distribution.

Statistical Process Control (SPC). A set of problem-solving tools based on data that are used to improve a process.

Scatter diagram. A diagram displaying observations on two variables, x and y. Each observation is represented by a point showing its x-y coordinates. The scatter diagram can be very effective in revealing the joint variability of x and y or the nature of the relationship between them.

Statistics. The science of collecting, analyzing, interpreting, and drawing conclusions from data.

Screening experiment. An experiment designed and conducted for the purpose of screening out or isolating a promising set of factors for future experimentation. Many screening experiments are fractional factorials, such as two-level fractional factorial designs. Second-order model. A model that contains second-order terms. For example, the second-order response surface model in two variables is y  0  1x1  2x2  12x1x2  11x 21  22x 22  . The second order terms in this model are 12x1x2, 11x 21, and 22x22. Shewhart control chart. A specific type of control chart developed by Walter A. Shewhart. Typically, each plotted point is a summary statistic calculated from the data in a rational subgroup. See Control chart. Sign test. A statistical test based on the signs of certain functions of the observations and not their magnitudes. Signed-rank test. A statistical test based on the differences within a set of paired observations. Each difference has a sign and a rank, and the test uses the sum of the differences with regard to sign. Significance. In hypothesis testing, an effect is said to be significant if the value of the test statistic lies in the critical region. Significance level. See Level of significance.

Statistical quality control. Statistical and engineering methods used to measure, monitor, control, and improve quality.

Steepest ascent (or descent). A strategy for a series of tests to optimize a response used along with response surface models. Stem-and-leaf diagram. A method of displaying data in which the stem corresponds to a range of data values and the leaf represents the next digit. It is an alternative to the histogram but displays the individual observations rather than sorting them into bins. Stepwise regression. A method of selecting variables for inclusion in a regression model. It operates by introducing the candidate variables one at a time (as in forward selection) and then attempting to remove variables following each forward step. Studentized range. The range of a sample divided by the sample standard deviation. Studentized residual. In regression, the studentized residual is calculated by dividing the ordinary residual by its exact standard deviation, producing a set of scaled residuals that have, exactly, unit standard deviation. Sufficient statistic. An estimator is said to be a sufficient statistic for an unknown parameter if the distribution of the sample given the statistic does not depend on the unknown parameter. This means that the distribution of the estimator contains all of the useful information about the unknown parameter.

JWCL232_Glo_751-761.qxd

1/23/10

10:58 AM

Page 761

GLOSSARY Tabular CUSUM. A numerical algorithm used to detect assignable causes on a cumulative sum control chart. See V mask. Tampering. Another name for overcontrol. t Distribution. The distribution of the random variable defined as the ratio of two independent random variables. The numerator is a standard normal random variable and the denominator is the square root of a chi-square random variable divided by its number of degrees of freedom.

761

Universe. Another name for population. V mask. A geometrical figure used to detect assignable causes on a cumulative sum control chart. With appropriate values for parameters, identical conclusions can be made from a V mask and a tabular CUSUM. Variable selection. The problem of selecting a subset of variables for a model from a candidate list that contains all or most of the useful information about the response in the data.

Test statistic. A function of a sample of observations that provides the basis for testing a statistical hypothesis.

Variables control chart. Any control chart for a continuous random variable. See Attribute control chart.

Time series. A set of ordered observations taken at points in time.

Variance. A measure of variability defined as the expected value of the square of the random variable around its mean.

Tolerance interval. An interval that contains a specified proportion of a population with a stated level of confidence. Tolerance limits. A set of limits between which some stated proportion of the values of a population must fall with a specified level of confidence. Total probability rule. Given a collection of mutually exclusive events whose union is the sample space, the probability of an event can be written as the sum of the probabilities of the intersections of the event with the members of this collection. Treatment. In experimental design, a treatment is a specific level of a factor of interest. Thus, if the factor is temperature, the treatments are the specific temperature levels used in the experiment. Treatment effect. The mean change to the response due to the presence of the treatment. Treatment sum of squares. In analysis of variance, this is the sum of squares that accounts for the variability in the response variable due to the different treatments that have been applied. t-test. Any test of significance based on the t distribution. The most common t-tests are (1) testing hypotheses about the mean of a normal distribution with unknown variance, (2) testing hypotheses about the means of two normal distributions, and (3) testing hypotheses about individual regression coefficients. Two-level factorial experiment. A full or fractional factorial experiment with all factors tested at only two levels (settings) each. See 2k factorial experiment.

Variance component. In analysis of variance models involving random effects, one of the objectives is to determine how much variability can be associated with each of the potential sources of variability defined by the experimenters. It is customary to define a variance associated with each of these sources. These variances in some sense sum to the total variance of the response, and are usually called variance components. Variance inflation factors. Quantities used in multiple regression to assess the extent of multicollinearity (or near linear dependence) in the regressors. The variance inflation factor for the ith regressor VIFi can be defined as VIFi = [1兾(1 ⫺ R 2i )], where R 2i is the coefficient of determination obtained when xi is regressed on the other regressor variables. Thus, when xi is nearly linearly dependent on a subset of the other regressors, R 2i will be close to unity and the value of the corresponding variance inflation factor will be large. Values of the variance inflation factors that exceed 10 are usually taken as a signal that multicollinearity is present. Warning limits. Horizontal lines added to a control chart (in addition to the control limits) that are used to make the chart more sensitive to assignable causes. Weibull random variable. A continuous random variable that is often used to model the time until failure of a physical system. The parameters of the distribution are flexible enough that the probability density function can assume many different shapes.

Type I error. In hypothesis testing, an error incurred by rejecting a null hypothesis when it is actually true (also called an ␣-error).

Western Electric rules. A specific set of run rules that were developed at Western Electric Corporation. See Run rules.

Type II error. In hypothesis testing, an error incurred by failing to reject a null hypothesis when it is actually false (also called a ␤-error).

Wilcoxon rank-sum test. A nonparametric test for the equality of means in two populations. This is sometimes called the MannWhitney test.

U chart. An attribute control chart that plots the average number of defects per unit in a subgroup. Also called a defects-per-unit control chart. Similar to a C chart. Unbiased estimator. An estimator that has its expected value equal to the parameter that is being estimated is said to be unbiased. Uniform random variable. Refers to either a discrete or continuous uniform random variable. Uniqueness property of moment generating function. Refers to the fact that random variables with the same moment generating function have the same distribution.

Wilcoxon signed-rank test. A distribution-free test of the equality of the location parameters of two otherwise identical distributions. It is an alternative to the two-sample t-test for nonnormal populations. With replacement. A method to select samples in which items are replaced between successive selections. Without replacement. A method to select samples in which items are not replaced between successive selections.

X chart. A control chart that plots the average of the measurements in a subgroup that is used to monitor the process mean.

JWCL232_ndx_762-765.qxd

1/22/10

9:53 PM

Page 762

Index 22 factorial design, 571 23 factorial design, 577 2k designs-mask on a cumulative sum control chart, 678 2k factorial designs, 571 2k-1 design, 602 2k-p fractional factorial designs, 608

A Acceptance region, 286, also see critical region Actual capability of a process, 664 Addition of center points to a 2k design, 588 Addition rules of probability, 37 Adjusted R2, 472 Aliases in a fractional factorial design, 603 All possible regression, 495 Alternate fraction, 604 Alternative hypothesis, 285, 292 Analysis of a second-order response surface model, 622 Analysis of variance (ANOVA), 417, 470, 520, 523, 534, 539, 540, 559 Analytic study, 12 ANOVA sum of squares identity for a randomized complete block design, 539 ANOVA sum of squares identity for a single-factor experiment, 518 ANOVA sum of squares identity for a two-factor factorial experiment, 560 Approximate sampling distribution of a difference in two sample means, 229 Assignable causes, 640 Attributes control charts, 643, 668 Average run length, 674 Axioms of probability, 34

B Backward variable elimination in regression, 500 Bayes’ theorem, 55, 56 Bayesian estimation of parameters, 244 Bayesian networks, 56 Bernoulli trial, 80, 86, 440

762

Beta distribution, 146, 147, 704 Beta random variable, 146 Bias of an estimator, 231 Binomial distribution, 79, 81, 127, 704 Binomial expansion, 81 Binomial probability table, 705 Binomial random variable, 81 Bivariate normal distribution, 177 Bivariate probability distributions, 154 Blocking, 538, 595 Bounds on a P-value, 311 Box plots, 208

C Calculating normal probabilities, 123 Calculation of interaction effects in a 2k factorial, 572, 578, 580 Categorical regressor variables, 492 Cause-and-effect relationships, 6, 352, 405 Center line on a control chart, 11 Center points and checking for curvature, 588 Central composite design, 622, 625 Central limit theorem, 118, 227 Chance causes, 640 Chi-square distribution, 266 Chi-square distribution percentage points, 710 Chi-square goodness of fit test, 330 Combinations, 26, 27 Comparative experiments, 7, 285, 351 Complement of an event, 23 Completely randomized design, 517 Completely randomized experiment, 315, 515 Components of variance, 518, 534 Conditional distribution of bivariate normal random variables, 179 Conditional probability, 41, 42 Conditional probability distributions, 158, 159 Confidence coefficient, 254 Confidence interval, 252, 253, 254, 295, 358, 368, 369, 379

Confidence interval on a correlation coefficient, 433 Confidence interval on the difference in means of two normal distributions, variances known, 358 Confidence interval on the difference in means of two normal distributions, variances unknown, 368, 369 Confidence interval on the difference in two proportions, 392 Confidence interval on the mean response in regression, 421, 480 Confidence interval on the ratio of the variances of two normal distributions, 387 Confidence intervals on a proportion, 271, 272 Confidence intervals on means in ANOVA, 521, 522, 523 Confidence intervals on regression coefficients, 421, 479 Confidence intervals on the mean of a normal distribution, variance known, 253, 254, 257 Confidence intervals on the mean of a normal distribution, variance unknown, 261, 263 Confidence intervals on the variance and standard deviation of a normal distribution, 266, 268 Confounding, 595 Connection between hypothesis testing and confidence intervals, 295 Contingency tables, 333 Continuity correction, 128 Continuous random variable, 58, 107, 108, 110, 144 Continuous sample space, 20 Continuous uniform distribution, 116, 704 Continuous uniform random variable, 116 Contrasts and effect estimates in a 2k factorial, 573 Contrasts in a 2k factorial, 573 Control chart, 11, 639 Control chart constants, 728

JWCL232_ndx_762-765.qxd

1/22/10

9:53 PM

Page 763

INDEX Control chart for defects per unit, 670 Control chart model, 642 Control chart performance, 673 Control charts for individual measurements, 658 Cook’s distance, 487 Correlation, 170, 173 Correlation and independent random variables, 174 Correlation coefficient, 432 Counting techniques, 24, 25, 26 Covariance, 170, 171 Covariance matrix, 461 Cp statistic in regression, 495 Critical region, 286 Critical values, 286 Cumulative distribution function, 71, 72, 111, 112 Cumulative frequency plot, 206 Cumulative sum control chart, 676

D Data collection, 5 Data versus information, 6 Decision interval, 679 Defect concentration diagram, 689 Defining contrast, 596 Defining relation for a fractional factorial design, 602, 609 Deming’s 14 points, 690 Descriptive statistics, 191 Design generator, 602 Design matrix, 577 Design of a control chart, 644 Design resolution, 608 Designed experiment, 5, 6, 296, 405, 513, 514, 552 Diagonal elements of the hat matrix, 486 Digidot plot, 211 Discrete random variables, 58, 66, 67 Discrete sample space, 20, 31 Discrete uniform distribution, 77, 704 Distribution of a subset of random variables, 165 Dot diagram, 4

E Effect of sample size on type II error probability, 290 Empirical model, 12, 13, 402 Engineering method, 2 Enumerative study, 12 Equally likely outcomes, 32 Erlang distribution, 138, 704 Estimation of parameters, 195 Events, 18, 22, 23 Expected value, 74, 114 Expected value of a function of a continuous random variable, 114

Expected value of a function of a discreet random variable, 76 Expected value of a function of two random variables, 171 Exponential distribution, 132, 704 Exponential random variable, 132 Exponentially weighted moving average (EWMA) control chart, 682 Extra sum of squares method, 474 Extrapolation in regression, 482

F Factor levels, 7 Factorial experiment, 7, 552, 555, 557, 558, 568, 571 Factorial experiments and interaction, 557 F-distribution, 382 F-distribution percentage points, 712 Finite population correction factor, 95 First quartile, 201 First-order response surface model, 619 Fisher least significance difference method of comparing means following ANOVA, 524 Fixed-effects model, 517, 559 Fixed significance level testing, 294, 312, 324 Formulating one-sided hypotheses, 293 Forward variable selection in regression, 500 Fraction defective control chart, 668 Fractional factorial experiment, 8, 9, 602 Frequency distributions, 203 Fundamental theorem of calculus, 112

G Gamma distribution, 138, 139, 704 Gamma function, 139 Gamma random variable, 139 Gaussian distribution, 118, also see normal distribution General factorial experiments, 568 General functions of random variables, 185, 186 General regression significance test, 474 Generalized interaction, 599 Geometric distribution, 86, 704 Geometric random variable, 86 Goodness-of-fit tests, 330 Graphical comparison of means following ANOVA, 525

H Hat matrix in regression, 486 Histogram, 109, 203, 204 Hypergeometric distribution, 92, 93, 704 Hypergeometric random variable, 93 Hypothesis, 284 Hypothesis testing, 7, 223, 284, 286, 295, 296, 299, 310, 319, 330, 333, 337, 342, 358, 362, 373, 384

763

Hypothesis testing on a population proportion, 323 Hypothesis testing on the difference in means of two normal distributions, variances unknown, 361, 362, 365 Hypothesis testing on the difference in two proportions, 389 Hypothesis testing on the mean of a normal distribution, variance known, 299 Hypothesis testing on the mean of a normal distribution, variance unknown, 310 Hypothesis testing on the ratio of the variances of two normal distributions, 384 Hypothesis testing on the variance and standard deviation of a normal distribution, 319

I Independence, 50, 161 Independent events, 51 Independent random variables, 161, 162, 166 Indicator variables in regression, 492 Influential observations in regression, 487 Interaction, 8, 452, 555 Interpreting a confidence interval, 255 Interquartile range, 201 Intersection of events, 23

J Jacobian, 186 Joint probability density function, 155 Joint probability distributions, 152, 153, 155, 163, 176 Joint probability mass function, 154

L Lack of memory property of the exponential random variable, 134, 135 Lack of memory property of the geometric random variable, 88 Large sample confidence interval for a parameter, 269 Large sample hypothesis testing on the mean of a normal distribution, 307 Large sample test for difference in two means, 355 Large-sample confidence interval for the mean, 258 Least squares estimators, 407, 408, 454, 457 Least squares normal equations for multiple linear regression, 454, 456 Least squares normal equations for simple linear regression, 407 Likelihood function, 239 Linear combinations of random variables, 181 Logistic regression, 440 Logit response function, 441 Lognormal distribution, 144, 704

JWCL232_ndx_762-765.qxd

764

1/22/10

9:53 PM

Page 764

INDEX

Lognormal random variable, 144 Lower confidence limit, 254 Lower control limit, 11, 645

M Main effect, 555, 572, 578, 579 Marginal distributions of bivariate normal random variables, 178 Marginal probability distribution, 156, 157, 164 Maximum likelihood estimator, 239, 240 Mean and variance of a continuous random variable, 114 Mean and variance of a discrete random variable, 74 Mean and variance of an average, 183 Mean of a linear combination of random variables, 182 Mean squared error of an estimator, 234 Mean squares, 519 Mechanistic model, 12, 13 Median, 199, 337 Method of least squares, 406, 452 Method of maximum likelihood, 239 Method of moments, 237 Method of steepest ascent, 620 Minimum variance unbiased estimator, 233 Model adequacy checking, 426, 484, 526, 527, 544, 564 Moment estimators, 238 Moments, 237 Moving range control chart, 658 Multicollinearity in regression, 502 Multinomial probability distribution, 176 Multiple comparisons in ANOVA, 524, 542 Multiple linear regression model, 450 Multiplication rule for counting techniques, 25 Multiplication rule for probability, 47 Mutually exclusive events, 23, 39

N Negative binomial distribution, 86, 88, 89 Noncentral t distribution, 315 Nonparametric statistical tests, 337, 342, 344, 373 Normal approximation for the sign test, 340 Normal approximation to Poisson distribution, 127, 130 Normal approximation to the binomial distribution, 127, 128, 270 Normal distribution, 118, 704 Normal probability plot, 215, 216 Normal probability plot of effects, 606, 611 Normal probability plotting of effects, 585 Normal probability plotting of residuals, 426, 485, 526, 544, 565, 576 Normal random variable, 119 Nuisance variables in an experiment, 538 Null hypothesis, 285

O Observational study, 5, 6, 352 One observation per cell in a factorial, 566 One-half fraction, 9 One-sided and two-sided hypotheses, 292, 300, 311, 320 One-sided confidence bounds, 257, 264, 268, 272 Operating characteristic curves, 305, 315, 322, 527, 717, 718, 719, 720.721, 722, 723, 724, 725 Optimal estimator, 235 Optimization experiment, 553, 619 Ordered stem-and-leaf diagram, 199 Orthogonal design, 581 Outlier, 208 Outliers, 486 Overcontrol, 10 Overfitting, 472

P Paired t-test, 376 Parameter estimation, 223 Pareto chart, 207, 688 Partial F-test, 475 Partial regression coefficients, 451 Patterns on control charts, 647 P-chart, 668 Percentile, 201 Permutations of similar objects, 26 Permutations, 25 Point estimate, 224 Point estimation, 224, 231, 237 Point estimator, 225 Poisson distribution, 97, 98, 127, 130, 704 Poisson process, 98 Poisson random variable, 98 Polynomial regression models, 451, 490 Pooled estimate of variance, 362 Pooled t-test, 362 Population standard deviation, 195 Population variance, 195 Population, 4 Posterior distribution, 244 Potential capability of a process, 664 Power of a statistical test, 292 Prediction interval for a future observation, 274, 275 Prediction interval in regression, 423, 481 PRESS statistic in regression, 496 Principal block, 596 Principal fraction, 604 Prior distribution, 244 Probability, 15, 17, 31, 33 Probability as degree of belief, 32 Probability density function, 108, 109 Probability distribution, 68, 107 Probability mass function, 69 Probability model, 16

Probability plots, 214, 330 Process capability determination, 662 Process capability ratios, 663, 665 Projection of 2k designs, 584 Projection of a fractional factorial design, 607 Properties of the maximum likelihood estimator, 241 P-value, 294, 300, 311, 321, 324, 386

Q Quality control and improvement, 638 Quartiles, 200

R R chart, 651 R2, 428, 472 Raleigh distribution, 142 Random effects model in ANOVA, 534 Random experiments, 18, 19 Random factors in an experiment, 533 Random sample, 43, 44, 226 Random variable, 4, 57, 58 Randomization, 6, 526, 558 Randomized complete block design (RCBD), 538 Randomized complete block design ANOVA, 540 Rational subgroups, 646 Reference distribution, 301, 354 Reference value, 679 Regression coefficients in a 2k factorial, 575 Regression model, 14, 401, 402, 404, 449, 450, 574, 583 Regressor variable, 405, 451 Relative efficiency of an estimator, 235 Relative frequency distribution, 204 Relative frequency, 32 Replication, 515 Reproductive property of the normal distribution, 183 Residual, 407, 457, 526, 544, 564, 574 Residual plots, 426, 484, 485, 526, 544, 564, 576, 577, 583 Resolution of designs, 608 Resolution III designs, 608 Resolution IV designs, 608 Response surface, 619 Response variable, 405 Retrospective study, 5 Rotatable second-order design, 625 Runs rules for control charts, 649

S Sample, 4 Sample correlation coefficient, 432 Sample mean, 192 Sample range, 195 Sample sizes for confidence intervals, 256, 272, 359

JWCL232_ndx_762-765.qxd

1/22/10

9:53 PM

Page 765

INDEX Sample sizes in ANOVA, 527 Sample sizes in hypothesis tests, 303, 304, 305. 314, 322, 326, 356, 357, 367, 387, 391 Sample space, 18, 19, 20 Sample standard deviation, 193 Sample variance, 193 Sampling distribution, 224, 225, 226 Sampling with replacement, 27 Sampling without replacement, 27 Saturated fractional factorial design, 615 Scatter diagram, 403 Scatter, see variability Science of data, see statistics Scientific method, 2 Screening experiment, 553 Second-order response surface model, 619, 622 Sensitivity of a statistical test, 292 Shewhart control charts, 642 Sign test, 337, 340, 341, 344 Sign test critical values, 726 Simple linear regression model, 404, 405 Single replicate of a 2k design, 585 Single-sample hypothesis testing, 7, 310, 319, 323 Six-sigma process, 666 Sources of variability, 3 Sparsity of effects principle, 585 Standard deviation, 74, 114 Standard error of a point estimator, 233 Standard error of a regression coefficient in a 2k factorial, 575 Standard error of regression coefficients, 415, 461 Standard normal distribution, 120, 144, 301 Standard normal distribution cumulative probability table, 708 Standard normal random variable, 120, 144 Standardized residuals, 484 Standardizing a normal random variable, 122 Statistic, 224, 226 Statistical inference, 4, 223, 225 Statistical process control (SPC), 12 Statistical process control, 640 Statistical quality control, 638, 639 Statistical thinking, 4

Statistical versus practical significance, 296, 297 Statistics, 3 Stem-and-leaf diagrams, 197 Stepwise regression, 499, 500 Strong versus weak conclusions in hypothesis testing, 291, 292 Studentized residuals, 430, 486 Summary table and guidelines for one-sample inference procedures, 274 Summary table and guidelines for two-sample inference procedures, 395 Sums of squares in a 2k factorial, 573

T Tabular cumulative sum control chart, 679 t-distribution, 262, 310, 315 t-distribution percentage points, 711 Test for significance of regression, 416, 417, 470 Test for zero correlation, 433 Test of homogeneity in a contingency table, 335 Test of independence in a contingency table, 333 Test statistic, 100, 310, 319, 324 Tests on individual regression coefficients in multiple regression, 472 Third quartile, 201 Three-sigma control limits, 673 Tier chart, 662 Ties in the sign test, 340 Time series data, 9, 210 Time series plot, 9, 210 Time weighted control charts, 676 Tolerance chart, 662 Tolerance interval for a normal distribution, 276 Total probability, 47, 48 Transformation of variables, 437 Treatments, 351, 515 Tree diagram, 21 t-test statistic for a regression coefficient in a 2k factorial, 575 t-tests in multiple regression, 473 t-tests in simple linear regression, 415 t-tests, 310, 362, 365, 376, 415, 473, 575

765

Two-factor factorial experiment, 558 Two-factor interaction, 556 Two-sample hypothesis testing, 7 Type I error, 287, 288 Type II error, 287, 289, 290, 292, 303, 367, 459, 460

U U-chart, 670 Unbalanced single-factor design ANOVA, 523 Unbiased estimator, 231, 409, 414 Union of events, 23 Upper confidence limit, 254 Upper control limit, 11, 645

V Variability, 3 Variable selection in regression, 472, 494, 499 Variables control charts, 643 Variance component estimation, 535 Variance inflation factors, 502 Variance of a linear combination of random variables, 182 Variance of a point estimator, 232 Venn diagram, 23 Verifying assumptions, 214

W Warning limits on control charts, 649 Weibull distribution, 141, 704 Weibull random variable, 141 Western Electric rules on control charts, 648 Wilcoxon rank-sum test, 373 Wilcoxon rank-sum test critical values, 727 Wilcoxon signed-rank test, 342, 344 Wilcoxon signed-rank test critical values, 736 Word in a defining relation, 609

X X and R charts, 649 X and S chart, 651, 652

Z Z-tests, 299, 323, 358

H0: p ⫽ p0

4. z0 ⫽

x ⫺ np0

␴ 20

1n ⫺ 12s2

2np0 11 ⫺ p0 2

x 20 ⫽

x x s2 pˆ

Mean ␮ of a normal distribution, variance ␴2 unknown Variance ␴2 of a normal distribution Proportion or parameter of a binomial distribution p

2.

3.

4.

z0 ⬍ ⫺z␣

H1: p ⬍ p0

Point Estimate

冟z0冟 ⬎ z␣/2 z0 ⬎ z␣

H1: p ⫽ p0 H1: p ⬎ p0

Mean ␮, variance ␴2 known

Problem Type

H1: ␴2 ⬎ ␴20 H1: ␴2 ⬍ ␴20

P ⫽ 2 31 ⫺ ⌽1z0 2 4 Probability above z0 P ⫽ 1⫺ ⌽(z0) Probability below z0 P ⫽ ⌽(z0)

P-value

O.C. Curve Parameter

g, h g, h

d ⫽ 1␮ ⫺ ␮0 2 Ⲑ␴ d ⫽ 1␮0 ⫺ ␮2 Ⲑ␴

ⱕ ␴2 ⱕ

␹ 21⫺␣/2,n⫺1

1n ⫺ 12s2

pˆ 11 ⫺ pˆ 2 pˆ 11 ⫺ pˆ 2 ⱕ p ⱕ pˆ ⫹ z␣Ⲑ2 n n B B

␹ 2␣/2,n⫺1

1n ⫺ 12s2

x ⫺ t␣/2, n⫺1 sⲐ 1n ⱕ ␮ ⱕ x ⫹ t␣/2,n⫺1 sⲐ 1n

x ⫺ z␣/2␴Ⲑ 1n ⱕ ␮ ⱕ x ⫹ z␣/2␴Ⲑ 1n

—

— —

␭ ⫽ ␴Ⲑ␴0 ␭ ⫽ ␴ Ⲑ␴0

—

— —

k, l m, n

i, j

e, f

d ⫽ 冟␮ ⫺ ␮0冟 Ⲑ␴

␭ ⫽ ␴Ⲑ␴0

c, d

a, b c, d

O.C. Curve Appendix Chart VII

d ⫽ 1␮0 ⫺ ␮2 Ⲑ␴

d ⫽ 冟␮ ⫺ ␮0冟 Ⲑ␴ d ⫽ 1␮ ⫺ ␮0 2 Ⲑ␴

Two-sided 10011 ⫺ ␣2 Percent Confidence Interval

P ⫽ 2 3 1 ⫺ ⌽1z0 2 4 Probability above z0 P ⫽ 1⫺ ⌽(z0) Probability below z0 P ⫽ ⌽(z0)

See text Section 9.4.

Sum of the probability above 冟t0冟 and below ⫺冟t0冟 Probability above t0 Probability below t0

pˆ ⫺ z␣Ⲑ2

2 ␹ 20 ⬎ ␹␣/2,n⫺1 or ␹ 20 ⬍ ␹ 21⫺␣/2,n⫺1 2 ␹ 20 ⬎ ␹␣,n⫺1 2 ␹ 0 ⬍ ␹ 21⫺␣,n⫺1

t0 ⬎ t␣,n⫺1 t0 ⬍ ⫺t␣,n⫺1

H1: ␮ ⬎ ␮0 H1: ␮ ⬍ ␮0 H1: ␴2 ⫽ ␴20

冟t0冟 ⬎ t␣/2, n⫺1

z0 ⬍ ⫺z␣

H1: ␮ ⬍ ␮0 H1: ␮ ⫽ ␮0

冟z0冟 ⬎ z␣/2 z0 ⬎ z␣

H1: ␮ ⫽ ␮0 H1: ␮ ⬎ ␮0

1.

Case

Summary of One-Sample Confidence Interval Procedures

H0: ␴2 ⫽ ␴20

3.

t0 ⫽

H0: ␮ ⫽ ␮0 ␴2 unknown x ⫺ ␮0 sⲐ 1n

x ⫺ ␮0 ␴Ⲑ 1n

z0 ⫽

H0: ␮ ⫽ ␮0 ␴2 known

Fixed Significance Level Criteria for Rejection

7:31 PM

2.

Test Statistic

Alternative Hypothesis

1/21/10

1.

Case

Null Hypothesis

Summary of One-Sample Hypothesis-Testing Procedures

JWCL232_IBC.qxd Page 1

t0 ⫽

H0: ␮1 ⫺ ␮2 ⫽ ⌬0 ␴21 ⫽ ␴22 unknown

Paired data H0: ␮D ⫽ 0

H0: ␴21 ⫽ ␴22

H0: p1 ⫽ p2

3.

4.

5.

6. z0 ⫽

a

s21

n2 ⫺ 1

d sd Ⲑ 1n

⫹

b

1s22/n2 2 2

n2

1 1 pˆ 11 ⫺ pˆ 2 c ⫹ d n1 n2 B

pˆ 1 ⫺ pˆ 2

f0 ⫽ s21 Ⲑs22

t0 ⫽

n1 ⫺ 1

⫹

s22 2

s21 s22 ⫹ n2 B n1

x1 ⫺ x2 ⫺ ⌬0

1 1 sp n ⫹ n 2 B 1

x1 ⫺ x2 ⫺ ⌬0

n1 1s21/n1 2 2

t0 ⫽

H0: ␮1 ⫺ ␮2 ⫽ ⌬0 ␴21 ⫽ ␴22 unknown

2.

␴21 ␴22 ⫹ n2 B n1

x1 ⫺ x2 ⫺ ⌬0

冟z0冟 ⬎ z␣/2 z0 ⬎ z␣ z0 ⬍ ⫺z␣

H1: p1 ⫽ p2 H1: p1 ⬎ p2 H1: p1 ⬍ p2

H1: ␴21 ⬎ ␴22

f0 ⬎ f␣Ⲑ2,n1⫺1,n2⫺1 or f0 ⬍ f1⫺␣Ⲑ2,n1⫺1,n2⫺1 f0 ⬎ f␣,n1⫺1,n2⫺1

t0 ⬎ t␣,n⫺1 t0 ⬍ ⫺t␣,n⫺1

H1: ␮d ⬎ 0 H1: ␮d ⬍ 0 H1: ␴21 ⫽ ␴22

冟t0冟 ⬎ t␣Ⲑ2,n⫺1

H1: ␮d ⫽ 0

t0 ⬎ t␣,v t0 ⬍ ⫺t␣,v

冟t0冟 ⬎ t␣Ⲑ2,v

H1: ␮1 ⫺ ␮2 ⫽ ⌬0 H1: ␮1 ⫺ ␮2 ⬎ ⌬0 H1: ␮1 ⫺ ␮2 ⬍ ⌬0

t0 ⬎ t␣,n1 ⫹n2⫺2 t0 ⬍ ⫺t␣,n1 ⫹n2⫺2

H1: ␮1 ⫺ ␮2 ⬎ ⌬0 H1: ␮1 ⫺ ␮2 ⬍ ⌬0

冟t0冟 ⬎ t␣Ⲑ2,n1 ⫹n2⫺2

z0 ⬍ ⫺z␣

H1: ␮1 ⫺ ␮2 ⬍ ⌬0

H1: ␮1 ⫺ ␮2 ⫽ ⌬0

z0 ⬎ z␣

冟z0冟 ⬎ z␣ Ⲑ 2

H1: ␮1 ⫺ ␮2 ⬎ ⌬0

H1: ␮1 ⫺ ␮2 ⫽ ⌬0

Fixed Significance Level Criteria for Rejection

P ⫽ 231 ⫺ ⌽1z0 2 4 Probability above z0 P ⫽ 1 ⫺ ⌽(z0) Probability below z0 P ⫽ ⌽(z0)

See text Section 10-5.2.

Sum of the probability above 冟t0冟 and below ⫺冟t0冟 Probability above t0 Probability below t0

Sum of the probability above 冟t0冟 and below ⫺冟t0冟 Probability above t0 Probability below t0

Sum of the probability above 冟t0冟 and below ⫺冟t0冟 Probability above t0 Probability below t0

Probability below z0 P ⫽ ⌽(z0)

Probability above z0 P ⫽ 1 ⫺ ⌽(z0)

P ⫽ 231 ⫺ ⌽1z0 2 4

P-value

q, r

␭ ⫽ ␴1 Ⲑ␴2

—

—

— —

o, p

␭ ⫽ ␴1 Ⲑ␴2

— —

— —

—

— —

—

— —

—

— —

—

g, h g, h

c, d

d ⫽ 1⌬ ⫺ ⌬0 2 Ⲑ2␴ d ⫽ 1⌬0 ⫺ ⌬2 Ⲑ2␴ where ⌬ ⫽ ␮1 ⫺ ␮2

⫹

␴22

c, d

a, b

e, f

2␴21

␮2 ⫺ ␮1 ⫺ ⌬0

2␴21 ⫹ ␴22

␮1 ⫺ ␮2 ⫺ ⌬0

2␴21 ⫹ ␴22

冟␮1 ⫺ ␮2 ⫺ ⌬0冟

O.C. Curve Appendix Chart VII

d ⫽ 冟⌬ ⫺ ⌬0冟 Ⲑ2␴

d⫽

d⫽

d⫽

O.C. Curve Parameter

9:06 PM

v⫽

z0 ⫽

H0: ␮1 ⫺ ␮2 ⫽ ⌬0 ␴21 and ␴22 known

Test Statistic

Alternative Hypothesis

1/21/10

1.

Case

Null Hypothesis

Summary of Two-Sample Hypothesis-Testing Procedures

JWCL232_IBC.qxd Page 2

Ratio of the variances ␴21 Ⲑ␴22 of two normal distributions

6.

Difference in two proportions of two binominal parameters p1 ⫺ p2

s21 s22

Difference in means of two normal distributions for paired samples ␮0 ⫽ ␮1 ⫺ ␮2

4.

5.

d

Difference in means of two normal distributions ␮1 ⫺ ␮2, variances ␴21 ⫽ ␴22 and unknown

3.

pˆ 1 ⫺ pˆ 2

x1 ⫺ x2

B

n1⫺1

1s21/n1 2 2

⫹

n2⫺1

1s22/n2 2 2

1s21/n1 ⫹ s22/n2 2 2

s21 s2 ⫹ 2 n2 B n1

B

1 f␣/2,n1⫺1,n2⫺1

B

pˆ 1 11 ⫺ pˆ 1 2 pˆ 2 11 ⫺ pˆ 2 2 ⫹ n1 n2

pˆ 1 11 ⫺ pˆ 1 2 pˆ 2 11 ⫺ pˆ 2 2 ⫹ n1 n2 ⱕ p1 ⫺ p2 ⱕ pˆ 1 ⫺ pˆ 2 ⫹ z␣/2

pˆ 1 ⫺ pˆ 2 ⫺ z␣/2

where f1⫺␣/2,n2⫺1,n1⫺1 ⫽

s21 ␴21 s21 2 f1⫺␣/2,n2⫺1,n1⫺1 ⱕ 2 ⱕ 2 f␣Ⲑ2,n2⫺1,n1⫺1 s2 ␴2 s2

d ⫺ t␣/2,n⫺1sd Ⲑ 1n ⱕ ␮D ⱕ d ⫹ t␣/2,n⫺1sd Ⲑ 1n

where v ⫽

ⱕ x1 ⫺ x2 ⫹ t␣/2,v

s21 s2 ⫹ 2 ⱕ ␮1 ⫺ ␮2 n2 B n1

n1 ⫹ n2 ⫺ 2

1n1 ⫺ 12s21 ⫹ 1n2 ⫺ 12s22

x1 ⫺ x2 ⫺ t␣/2,v

where sp ⫽

1 1 ⫹ n2 B n1

12:17 PM

ⱕ x1 ⫺ x2 ⫹ t␣/2,n1⫹n2⫺2sp

1 1 ⫹ ⱕ ␮1 ⫺ ␮2 n2 B n1

␴21 ␴22 ⫹ n n2 B 1

␴22 ␴21 ⫹ ⱕ ␮1 ⫺ ␮2 n2 B n1

x1 ⫺ x2 ⫺ t␣/2,n1⫹n2⫺2sp

ⱕ x1 ⫺ x2 ⫹ z␣/2

x1 ⫺ x2 ⫺ z␣/2

Two-Sided 10011 ⫺ ␣2 Percent Confidence Interval

1/18/10

x1 ⫺ x2

Difference in means of two normal distributions ␮1 ⫺ ␮2, variances ␴21 ⫽ ␴22 and unknown

2.

x1 ⫺ x2

Point Estimate

Difference in two means ␮1 and ␮2, variances ␴21 and ␴22 known

Problem Type

1.

Case

Summary of Two-Sample Confidence Interval Procedures

JWCL232_IBC.qxd Page 3

JWCL232_IFC.qxd

12/23/09

8:55 PM

Page 2

Index of Applications in Examples and Exercises BIOLOGICAL Amino acid composition of soybean meal Anaerobic respiration Blood Cholesterol level Glucose level Hypertension Body mass index (BMI) Body temperature Cellular replication Circumference of orange trees Deceased beetles under autolysis and putrefaction Diet and weight loss Disease in plants Dugongs (sea cows) length Fatty acid in margarine Gene expression Gene occurrence Gene sequences Grain quality Height of plants Height or weight of people Insect fragments in chocolate bars IQ for monozygotic twins Leaf transmutation Leg strength Light-dependent photosynthesis Nisin recovery Pesticides and grape infestation Potato spoilage Protein in Livestock feed in Milk from Peanut milk Protopectin content in tomatoes Rat muscle Rat tumors

Exercise 8-52 Exercise 2-144 Exercise 15-10 Exercises 13-25, 14-37 Exercises 4-143, 8-31, 11-8, 11-30, 11-46 Exercise 11-35 Exercise 9-59 Exercises 2-193, 3-100 Exercise 10-46 Exercise 2-92 Exercises 10-43, 10-77, 15-35 Exercise 14-76 Exercise 11-15 Exercises 8-36, 8-66, 8-76, 9-147, 9-113 Exercises 6-65, 13-50, 15-42 Exercises 2-195, 3-11 Exercises 2-25, 2-192, 3-13, 3-147 Exercise 8-21 Exercises 4-170, 4-171 Exercises 4-44, 4-66, 5-64, 6-30, 6-37, 6-46, 6-63, 6-73, 9-68 Exercises 3-134, 4-101 Exercise 10-45 Exercises 2-88, 3-123 Exercises 8-30, 9-64 Exercise 2-24 Exercises 12-14, 12-32, 12-50, 12-64, 12-84, 14-83 Exercise 10-94 Exercise 13-14

Rat weight Rejuvenated mitochondria Root vole population Sodium content of cornflakes Soil Splitting cell St John’s Wort Stork sightings Sugar content Synapses in the granule cell layer Tar content in tobacco Taste evaluation Tissues from an ivy plant Visual accommodation Weight of swine or guinea pigs Wheat grain drying

CHEMICAL Acid-base titration Alloys Contamination Cooking oil Etching Infrared focal plane arrays Melting point of a binder Metallic material transition Moisture content in raw material Mole fraction solubility Mole ratio of sebacic acid Pitch carbon analysis Plasma etching Polymers

Exercise 14-75 Exercises 13-13, 13-25, 13-33 Exercise 9-143 Exercises 13-40, 15-40 Exercise 6-15 Exercise 8-50

Exercise 8-57 Exercises 2-96, 3-88 Exercise 14-16 Exercise 9-61 Exercises 3-24, 12-1, 12-2, 12-23, 12-24, 12-41, 12-42 Exercise 4-155 Example 10-14 Exercises 4-100, 11-96 Exercises 8-46, 9-83, 9-114 Exercise 9-145 Exercise 8-95 Exercises 14-13, 14-31, 14-34, 14-50, 14-54 Exercise 2-130 Exercises 6-11, 6-16, 6-75 Exercises 9-142, 13-48 Exercises 13-47, 15-41

Propellant Bond shear strength Burning rate

Exercises 2-60, 2-132, 3-12, 5-48 Examples 6-4, 8-5 Exercises 10-21, 10-44, 10-59, 13-38, 15-17 Exercise 2-128, 4-113 Exercise 2-79 Exercises 10-19, 10-65, 10-34 Exercise 9-146 Exercise 9-42 Examples 8-1, 8-2 Exercise 3-6 Exercises 12-75, 12-91 Exercise 11-91 Exercises 12-10, 12-36, 12-50, 12-60, 12-68 Examples 14-5, 14-8 Exercise 7-32 Exercises 7-15, 10-8, 13-12, 13-24 Examples 15-1, 15-2, 15-4 Exercises 11-11, 11-31, 11-49, 15-32 Examples 9-1, 9-2, 9-3, 9-4, 9-5 Exercise 10-6

JWCL232_IFC.qxd

12/23/09

8:55 PM

Page 3

Purity Thermal barrier coatings

Exercise 15-42 Exercise 10-75

CHEMICAL ENGINEERING Aluminum smelting Automobile basecoat Blow molding Catalyst usage Concentration Conversion Cooling system in a nuclear submarine Copper content of a plating bath Dispensed syrup in soda machine Dry ash value of paper pulp Fill volume and capability

Filtration rate Fish preparation Flow metering devices Foam expanding agents Green liquor Hardwood concentration Impurity level in chemical product Injection molding Laboratory analysis of chemical process samples Maximum heat of a hot tub Na2S concentration NbOCl3 Oxygen purity pH and Catalyst concentration of Plating bath of a Solution of a Water sample Product color Product solution strength in recirculation unit Pulp brightness Reaction Time Redox reaction experiments Shampoo foam height Stack loss of ammonia Temperature Firing Furnace of Hall cell solution Vapor deposition Vapor phase oxidation of naphthalene Viscosity

Exercise 10-92 Exercises 14-56, 14-68 Exercise 16-59 Exercise 10-17 Examples 16-2, 16-6 Exercises 5-46, 6-68, 6-84, 10-9, 10-54, 15-64 Exercise 12-3 Exercise 9-130 Exercises 15-8, 15-34, 15-58 Exercises 8-29, 8-63, 8-75 Exercise 14-57 Examples 5-35, 8-6, 9-8, 9-9 Exercises 2-180, 3-146, 3-151, 4-62, 4-63, 5-62, 9-100, 10-4, 10-85, 10-90, 14-43, 15-38 Exercise 14-44 Exercise 13-46 Examples 15-3, 15-5 Exercises 9-126, 9-127 Exercises 10-16, 10-56, 10-88 Exercise 12-100 Example 13-2 Exercise 14-11 Exercises 15-3, 15-15 Example 14-9 Exercises 2-15, 2-137, 10-70 Exercise 2-43 Exercise 10-33 Exercises 11-7, 11-29, 11-41, 11-62 Exercise 6-36 Examples 11-1, 11-2, 11-3, 11-4, 11-5, 11-6, 11-7 Exercise 14-61 Exercises 15-1, 15-13 Exercise 6-17 Exercise 2-11 Exercise 14-45 Exercise 14-38 Exercise 13-31 Example 4-5 Exercises 2-13, 2-33, 4-56 Exercise 2-65 Exercises 8-91, 9-15, 9-16, 9-17, 9-18, 9-19, 9-128 Exercises 12-16, 12-34, 12-52, 12-66, 12-85 Exercise 13-15 Exercises 6-55, 6-109 Exercise 11-92 Exercises 13-28, 13-32 Exercise 6-54 Exercises 6-66, 6-88, 6-90, 6-96, 12-73, 12-103, 14-64, 15-20, 15-36, 15-86

Water temperature from power plant cooling tower Water vapor pressure Wine

Exercise 9-40 Exercise 11-78 Examples 12-14, 12-15 Exercises 6-35, 6-51

CIVIL ENGINEERING Cement and Concrete Hydration Mixture heat Mortar briquettes Strength Tensile strength Compressive strength

Intrinsic permeability Highway pavement cracks Pavement deflection Retained strength of asphalt Speed limits Traffic Wearing seat belts

Example 10-8 Exercises 9-10, 9-11, 9-12, 9-13, 9-14 Exercise 15-79 Exercises 4-57, 15-24 Exercise 15-25 Exercises 13-3, 13-9, 13-19, 14-14, 14-24, 14-48, 7-7, 7-8, 8-13, 8-18, 8-37, 8-69, 8-80, 8-87, 8-90, 15-5 Exercises 11-1, 11-23, 11-39, 11-52 Exercise 3-138, 4-102 Exercises 11-2, 11-16, 11-24, 11-40 Exercises 13-11,13-23 Exercises 8-59, 10-60 Exercises 3-87, 3-149, 3-153, 9-190 Exercises 10-82, 10-83

COMMUNICATIONS, COMPUTERS, AND NETWORKS Cell phone signal bars Cellular neural network speed Code for a wireless garage door Computer clock cycles Computer networks Corporate Web site errors Digital channel Electronic messages Email routes Encryption-decryption system Errors in a communications channel

Passwords Programming design languages Response time in computer operation system Software development cost Telecommunication prefixes Telecommunications

Transaction processing performance and OLTP benchmark Viruses Web browsing

Examples 5-1, 5-3 Exercise 8-39 Exercise 2-34 Exercise 3-8 Example 4-21 Exercises 2-10, 2-64, 2-164, 3-148, 3-175, 4-65, 4-94 Exercise 4-84 Examples 2-3, 3-4, 3-6, 3-9, 3-12, 3-16, 3-24, 4-15, 5-7, 5-9, 5-10 Exercises 3-158, 4-98, 4-115 Exercise 2-184 Exercise 2-181 Examples 3-22, 4-17, 4-20 Exercises 2-2, 2-4, 2-46, 3-40, 4-116, 5-5, 5-12, 6-94, 9-135 Exercises 2-81, 2-97, 2-194, 3-91, 3-108 Exercise 10-40 Exercise 8-82 Exercise 13-49 Exercise 2-45 Examples 3-1, 3-14 Exercises 2-17, 3-2, 3-85, 3-105, 3-132, 3-155, 4-95, 4-105, 4-111, 4-117, 4-160, 5-78, 9-98, 15-9 Exercises 2-68, 2-175, 5-10, 5-34, 10-7 Exercise 3-75 Examples 3-25, 5-12, 5-13 Exercises 2-32, 2-191, 3-159, 4-87, 4-140, 5-6

JWCL232_IFC.qxd

12/23/09

8:55 PM

Page 4

ELECTRONICS Automobile engine controller Bipolar transistor current Calculator circuit response Circuits Conductivity Current Drain and leakage current Electromagnetic energy absorption Error recovery procedures Inverter transient point Magnetic tape Nickel charge Parallel circuits Power consumption Power supply Printed circuit cards Redundant disk array Resistors Solder connections Strands of copper wire Surface charge Surface mount technology (SMT) Transistor life Voltage measurement errors

Examples 9-10, 9-11 Exercise 14-7 Exercises 13-6, 13-18 Examples 2-35, 7-3 Exercises 2-135, 2-136, 2-170, 2-177, 2-190 Exercise 12-105 Examples 4-1, 4-5, 4-8, 4-9, 4-12, 16-3 Exercises 10-31, 15-30 Exercises 13-41, 11-85 Exercise 10-26 Exercises 2-18, 2-166 Exercises 12-98, 12-99, 12-102 Exercises 2-189, 3-125 Exercises 2-61, 3-48 Example 2-34 Exercises 6-89, 11-79, 12-6, 12-26, 12-44, 12-58, 12-80 Example 9-13 Exercises 2-3, 9-20, 9-21, 9-22, 9-23, 9-24, 9-28 Example 2-10 Exercises 2-42, 3-122 Exercise 2-127 Example 7-1 Exercise 6-86 Exercises 3-1, 15-43, 15-45 Exercise 2-77 Exercise 14-15 Example 16-5 Exercise 7-51 Exercise 4-48N

ENERGY Consumption in Asia Enrichment percentage of reactor fuel rods Fuel octane ratings Gasoline cost by month Gasoline mileage

Heating rate index Petroleum imports Released from cells Renewable energy consumption Steam usage Wind power

Exercises 6-29, 6-45, 6-59 Exercises 8-41, 8-71, 8-88 Exercises 6-22, 6-26, 6-38, 6-42, 6-58, 6-78, 10-7 Exercise 15-98 Exercises 10-89, 11-6, 11-17, 11-28, 11-44, 11-56, 12-27, 12-55, 12-57, 12-77, 12-89, 15-37 Exercise 14-46 Exercise 6-72 Exercise 2-168 Exercise 15-78 Exercises 11-5, 11-27, 11-43, 11-55 Exercises 4-132, 11-9

Asbestos Biochemical oxygen demand (BOD) Calcium concentration in lake water Carbon dioxide in the atmosphere Chloride in surface streams Cloud seeding Earthquakes

Exercises 2-28, 15-34 Exercises 6-83, 11-74 Exercise 10-93 Example 8-4 Exercise 4-181 Example 3-18 Exercises 8-94, 9-63, 9-140 Exercises 2-9, 11-90 Exercises 13-8, 13-20 Exercises 8-33, 8-65, 8-77 Exercises 6-32, 6-48, 6-60, 6-80, 9-70 Exercise 8-49 Exercises 6-92, 6-97 Exercises 9-27, 9-94 Exercise 2-37 Exercises 4-68, 9-137 Exercise 11-70

Temperature in Phoenix, AZ Temperature of sewage discharge Voters and air pollution Waste water treatment tank Water demand and quality Watershed yield

MATERIALS Baked density of carbon anodes Ceramic substrate Coating temperature Coating weight and surface roughness Compressive strength Flow rate on silicon wafers Insulation ability Insulation fluid breakdown time Izod impact test Luminescent ink Paint drying time Particle size Photoresist thickness Plastic breaking strength Polycarbonate plastic Rockwell hardness Temperature of concrete Tensile strength of Aluminum Fiber Steel Paper Titanium content Tube brightness in TV sets

Exercise 14-4 Example 16-4 Exercises 10-24, 10-60 Exercise 2-90 Exercises 7-56, 11-60 Exercises 13-2, 13-16, 15-28 Exercise 14-5 Exercises 6-8, 6-74 Exercises 8-28, 8-62, 8-74, 9-66, 9-80 Exercise 5-28 Examples 10-1, 10-2, 10-3 Exercises 14-2, 14-19, 15-8, 15-16 Exercises 4-33, 16-17 Exercise 5-63 Exercises 10-5, 10-20, 10-55 Example 2-8 Exercises 2-66, 2-76 Exercises 10-91, 9-115, 15-17 Exercise 9-58 Example 10-4 Exercises 7-3, 7-4, 13-3, 13-17 Example 10-9 Exercise 9-44 Example 13-1 Exercises 4-154, 11-86 Exercises 8-47, 9-79, 15-2, 15-12 Exercises 7-12, 8-35, 8-67, 8-79, 9-148, 9-67, 14-1

MECHANICAL

ENVIRONMENTAL Arsenic

Emissions and fluoride emissions Global temperature Hydrophobic organic substances Mercury contamination Ocean wave height Organic pollution Oxygen concentration Ozone levels Radon release Rainfall in Australia Suspended solids in lake water

Example 10-6 Exercises 12-12, 12-30, 12-48, 12-62, 12-76, 12-88, 13-39 Exercises 4-85, 4-169 Exercises 11-13, 11-33, 11-51 Exercise 8-9 Exercise 3-58 Exercises 11-10, 11-32, 11-48, 11-59 Exercise 9-60 Exercises 6-63, 9-102, 11-15, 15-46

Aircraft manufacturing

Artillery shells Beam delamination Bearings Diameter Wear

Examples 6-6, 12-12, 14-1, 15-6, 16-1 Exercises 6-8, 8-97, 10-42, 15-31, 15-13, 15-74 Exercise 9-106 Exercises 8-32, 8-64 Examples 8-7, 8-8 Exercise 9-95 Exercises 4-181, 9-42, 15-6, 15-14 Example 4-25 Exercises 5-22, 4-127, 12-19, 12-39, 12-45, 12-67 (Text continued at the back of book.)

JWCL232_Endpaper.qxd

1/22/10

9:52 PM

Page 766

Applications in Examples and Exercises, continued

MEDICAL

Bolts in bearing cap and plate Brake contact resistance Casing for a gear housing Cast aluminum parts Circular tubes yield strength Cold start ignition time Connector pull-off force

ACL reconstruction surgery Antirheumatoid medication Artificial hearts Bacteria exposure Basal metabolic rate (BMR) Cholesterol level Completing a blood sample test Diagnostic kit revenue Diagnostic Heart failure Hemoglobin level Knee injuries Lung cancer Meniscal tear Noise exposure and hypertension Pain medication Patient satisfaction Plasma antioxidants from chocolate Radiation dose in X-ray

Copper plate warping Cycles to failure Deflection temperature for plastic pipe Dot diameter Drag coefficient Electromechanical product F-117A mission duration Fatigue crack growth Flatness distortion Fretting wear Gap width of a magnetic recording head Glass bottle thickness Height of leaf springs Hole diameter

Jet-turbine or rocket thrust Machined dimensions Mechanical assembly Missile miss distance Molded parts Nonconforming coil springs Nozzle velocity Particleboard deflection Precision measuring instruments Robotic insertion tool Shaft and washer assemblies Shear strengths of Rubber of Spot weld Sheet metal operation Space shuttle flight control system Spindle saw processes Surface roughness

Suspension helmets impact test Suspension rod, piston rings, PVC pipe, and other diameters

Temperature of joint for O-rings Tire life Tool life Wear in auto parts Wire bond strength

766

Exercises 2-177, 3-170, 5-77 Exercises 13-45 Example 2-9 Exercise 2-27 Exercise 6-10 Exercises 6-53, 6-64, 6-98 Exercises 6-67, 7-28, 9-4, 9-131 Exercises 14-6, 14-9 Exercises 6-23, 6-27, 6-39, 6-43, 6-79 Exercise 10-18 Exercises 4-70, 4-174 Exercises 6-18, 6-56, 6-105 Exercise 2-44 Exercise 6-13 Exercise 14-8 Exercise 14-62 Exercises 11-9,11-47, 11-57, 11-61, 11-63 Exercise 4-23 Exercises 8-40, 8-70, 8-72, 8-82 Exercise 14-73 Examples 4-2, 4-4, 4-8, 4-31, 8-39 Exercises 8-10, 9-74, 15-21, 15-63 Exercises 12-74,12-96, 12-97, 12-101 Examples 5-12, 5-13 Example 3-27 Example 14-7 Example 2-1 Exercises 6-91, 9-134 Exercise 13-37 Exercises 11-14,11-34 Exercise 9-132 Exercise 2-184 Exercise 15-41 Exercise 10-37 Exercise 12-72 Exercises 6-31, 6-47, 6-61, 9-136, 11-69 Exercise 2-40 Exercises 3-22, 3-57 Exercise 10-25 Examples 14-2, 14-4, 10-14, 12-13 Exercises 2-78, 2-109, 2-158, 5-19, 13-42, 14-22, 14-67, 14-69 Exercise 8-58 Example 4-16 Exercises 6-7, 6-73, 7-3, 8-13, 8-34, 8-68, 8-78, 10-15, 10-57, 15-7, 15-33, 16-77 Exercises 6-19, 6-57, 6-77 Exercises 8-27, 8-61, 8-73, 9-65, 9-81, 10-39 Exercises 14-12, 14-30, 14-56 Exercises 9-41, 10-23 Examples 11-8, 12-1, 12-2, 12-3, 12-4, 12-5, 12-6, 12-7, 12-8, 12-9, 12-10, 12-11 Exercises 12-8, 12-28, 12-46, 12-82, 15-9, 15-15, 15-35, 12-76, 16-7, 16-11, 16-29

Recovering from an illness Salk polio vaccine trials Sick-leave time Skin desquamation (peeling) Success rates in kidney stone removals Surgical versus medical patients Syringe length Therapeutic drugs Tissue assay by liver manganese Treatment of renal calculi by operative surgery Weight and systolic blood pressure

Exercises 4-59, 5-61 Exercise 5-76 Exercise 9-47 Exercise 4-159 Exercise 8-100 Exercises 4-60, 10-41 Exercise 4-154 Exercise 2-200 Example 2-37 Exercises 2-112, 3-90, 3-107 Exercises 8-86, 8-89 Exercises 2-76, 3-19 Exercise 8-56 Exercises 8-56, 10-68 Exercise 11-58 Exercises 3-162, 10-84 Exercises 12-5, 12-106 Exercises 13-2, 13-25 Exercises 12-11, 12-31, 12-44, 12-61, 12-79 Exercise 3-161 Exercise 10-79 Exercise 4-158 Exercise 2-88 Exercise 2-115 Exercise 9-104 Exercises 15-10, 15-16 Exercise 9-133 Exercise 9-144 Exercises 9-89, 9-99 Exercises 11-72, 11-89

PHYSICS Alignment accuracy of optical chips Atomic clock Current draw in magnetic core Density measurement error Density of the earth Geiger counts Laser diode samples Laser failure time Number of stars Optical correlator modulation Oxygen in a superconducting compound Porosity of ultrafine powder Shortened latencies in dry atmosphere Silver particles in photographic emulsion Solar intensity Supercavitation for undersea vehicles Thermal conductivity Transducer calibration Tube conductivity Velocity of a particle in a gas Velocity of light Voids in a ferrite slab Wavelength of radiation

Exercise 14-64 Exercise 11-94 Exercise 11-73 Exercises 12-13, 12-49, 12-63, 12-81 Exercise 6-102 Example 4-23 Exercise 4-93 Example 2-15 Exercises 4-115, 4-160 Exercise 3-133 Exercises 12-15, 12-33, 12-51, 12-65, 12-83 Exercises 13-10, 13-22 Exercises 11-12, 11-50 Exercise 8-99 Exercise 4-164 Exercises 6-12, 6-69, 6-76, 8-34 Exercise 9-45 Example 7-5 Exercise 3-150 Exercises 13-7, 13-21 Exercise 5-72 Exercise 6-101 Exercises 2-12, 6-30, 6-46 Exercise 3-69

SEMICONDUCTOR MANUFACTURING Examples 2-17, 2-19, 2-20, 2-27, 2-28, 2-33, 3-2, 3-3, 3-5, 3-21, 4-26, 10-11, 10-13, 14-3 Exercises 2-41, 2-92, 2-95, 2-168, 3-21, 3-84, 4-61, 4-80, 5-83, 6-36, 6-50, 6-62, 7-29, 7-44, 9-88, 9-124, 12-9, 12-29, 12-43, 12-69, 14-17, 14-25, 14-41, 14-58, 14-65, 15-18, 15-84

JWCL232_Endpaper.qxd

1/22/10

9:52 PM

Page 767

SPORTS Baseball coefficient of restitution Electromyostimulation Football helmets Golf

Hockey player performance Liveliness of baseballs Major League Baseball National Hockey League Quarterback ratings

Prices of houses Exercises 9-62, 9-139 Exercise 8-48 Exercise 9-91 Examples 9-6, 9-7 Exercises 6-9, 6-33, 6-49, 6-95, 6-99, 9-69, 9-138, 10-29, 10-30, 10-61, 10-62 Exercise 9-77 Exercises 8-92, 8-93 Exercise 12-104 Exercises 12-18, 12-38, 12-56, 12-70, 12-86 Exercises 11-3, 11-25, 11-41, 11-53, 11-75, 12-17, 12-35, 12-53, 12-59, 12-87

TEXTILE MANUFACTURING Examples 13-4, 13-5 Breaking strength of yarn Thread elongation of a drapery yarn

Exercises 6-24, 6-28, 6-40, 6-44, 13-29, 14-10, 14-18, 14-32, 14-63 Exercises 8-10,10-78, 10-80 Exercises 9-5, 9-6, 9-7, 9-8, 9-9

INDUSTRIAL ENGINEERING AND ECONOMICS Airlines Overbooking Arrival and waiting times Passenger airline miles flown in UK Automobile features Bayesian network Breakdowns by shift Buying tickets by telephone Calculator owners College graduates in Tempe Credit card numbers Customer design evaluation Customer sampling Cytogenic revenue Diamond prices Disabled population Engineering education Fraud detection in phone cards Impact of quality on schedules and costs Inspection of shipments from suppliers Inspection Installation technician service License numbers Lottery Machine schedules Monthly champagne sales in France Multiple choice exam Optical inspection Orders for computer systems Parallel parking Pension plan preference Presidential elections Price of an electronic storage device

Exercises 3-93, 4-180 Exercises 4-22, 15-15, 15-25, 7-12 Exercise 6-70 Examples 2-4, 2-5 Exercise 2-14 Example 2-38 Exercise 9-103 Exercise 2-179 Exercise 7-33 Exercise 9-25 Exercises 2-62, 2-135 Exercise 2-147 Example 3-29 Exercises 2-8, 2-34, 3-173, 9-86 Exercise 3-25 Exercise 11-95 Exercise 4-81 Exercises 5-9, 8-105, 8-92, 9-105, 9-108, 11-71 Exercise 2-144 Exercise 3-94 Exercise 3-182 Exercise 9-87 Exercise 3-172 Exercise 2-63 Exercise 3-124 Examples 2-11, 2-36 Exercise 6-91 Exercise 3-88 Exercise 3-20 Exercises 2-16, 2-35 Example 10-11 Exercise 10-38 Example 9-14 Exercises 2-110, 2-146, 8-55, 10-69 Exercise 3-23

Printer orders Product and bar codes Repeatability in component assembly Revenue potential Risk analysis Shipment of customers’ orders Soldiers killed by horse kicks Survey favoring toll roads Time between arrivals Time to Fill an electronic form Locate a part Make pottery Prepare a micro-array slide Recharge a battery Unemployment data Unlisted phone numbers

Exercises 11-4, 11-26, 11-42, 11-54 Exercise 5-94 Examples 2-12, 3-13 Exercise 3-67 Exercise 10-55 Example 3-10 Exercise 5-93 Exercise 5-95 Exercise 2-174 Exercise 3-135 Exercise 9-26 Exercises 4-45, 4-99, 4-104, 4-119, 4-162, 5-20, 5-21 Exercise 4-42 Exercise 5-87 Exercise 5-58 Example 4-24 Exercise 4-58 Exercise 6-85 Exercise 10-81

DEFECTIVES, FLAWS, ERRORS Automobile front lights Bearings Calculators Computers and fans Connector Cotton rolls Electronic components Contamination Integrated circuits Lenses Machining stages Optical or magnetic disks Optical alignment Orange juice containers Ovens Oxygen containers Pareto chart for automobile doors Printed circuit boards Printing Response surface design for yield Surface flaws in parts, steel, and panels

Textbook Water meter Wires Yield

Exercise 2-81 Example 10-16 Exercise 8-96 Exercises 2-111, 3-109, 4-97 Exercise 2-107 Exercise 2-108 Exercises 4-8, 6-111, 8-101, 9-107, 11-68 Exercise 3-140 Exercises 7-59, 8-53 Exercises 9-90, 10-71, 15-87 Example 2-26 Example 3-33 Exercises 3-137, 3-171 Exercise 3-103 Exercise 2-125 Exercise 5-45 Exercise 2-161 Exercise 6-52 Example 9-12 Exercises 4-186, 5-4 Example 14-12 Examples 2-23, 2-30 Exercises 3-7, 3-139, 3-176, 3-184, 4-106, 9-97, 15-76 Exercise 3-160 Exercise 4-82 Examples 3-31, 3-32 Exercise 15-44 Examples 6-5, 10-5, 10-7, 14-6 Exercises 5-26, 6-25, 6-41, 8-11, 13-30, 14-20, 14-42, 14-47, 14-53, 14-72

LIFE AND FAILURE TIMES Assembly and components Batteries

Example 5-14 Exercises 4-103, 4-83, 5-82 Example 6-7 Exercises 9-43, 9-93, 14-3

767

JWCL232_Endpaper.qxd

1/22/10

9:52 PM

Censored components CPU Door latch Electronic amplifier Light bulb Machine Magnetic resonance imaging machine Packaged magnetic disk Recirculating pump Semiconductor laser Voltage regulators Yarn

Page 768

Exercise 7-76 Example 4-23 Exercises 4-129, 4-167 Exercise 8-8 Exercise 4-184 Exercises 8-14, 8-16, 8-17 Exercise 7-50 Exercise 4-149 Exercise 4-148 Exercise 4-163 Exercises 4-69, 4-143, 4-176 Exercise 4-96 Exercise 6-100

THICKNESS OF Aluminum cylinders Coatings Flange Halves Ink layers Laminated covering Layers and error propagation Parts Photoconductor and photo resist film Plastic connector Wood paneling

Exercise 4-57 Exercises 3-66, 4-35 Exercise 4-41 Exercise Example 5-18 Exercises 5-49, 5-59 Exercise 4-173 Examples 5-31, 5-91, 5-92 Exercises 5-66, 15-11, 15-57 Exercises 4-43, 10-22, 10-64 Examples 2-2, 2-7 Exercises 3-10, 3-42

LENGTH OF Computer cable Door casing Hinge Injection-molded plastic case Keyway depth Metal rod Panels Plate glass parts Punched part

768

Exercise 4-34 Exercise 5-60 Exercise 4-8 Exercise 4-157 Exercise 16-91 Exercises 4-11, 4-21 Exercise 5-85 Exercise 3-68 Exercise 5-56

WEIGHT OF Adobe bricks Candy Chemical herbicide Components Measured by a digital scale Paper Running shoes Sample and measurement error

Exercises 5-27, 5-84 Exercises 5-86, 7-49 Exercises 4-9, 4-36, 4-40 Exercises 5-65, 5-88 Exercise 2-21 Exercises 8-51, 9-78 Exercise 4-71 Exercise 4-72