c09.qxd 5/15/02 8:02 PM Page 277 RK UL 9 RK UL 9:Desktop Folder: 9 Tests of Hypotheses for a Single Sample CHAPTER OUTLINE 9-1 HYPOTHESIS TESTING 9-1.1 Statistical Hypotheses 9-1.2 Tests of Statistical Hypotheses 9-1.3 One-Sided and Two-Sided Hypotheses 9-1.4 General Procedure for Hypothesis Tests 9-2 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE KNOWN 9-2.1 Hypothesis Tests on the Mean 9-2.2 P-Values in Hypothesis Tests 9-2.3 Connection between Hypothesis Tests and Confidence Intervals 9-2.4 Type II Error and Choice of Sample Size 9-2.5 Large-Sample Test 9-2.6 Some Practical Comments on Hypothesis Tests 9-3 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE UNKNOWN 9-3.1 Hypothesis Tests on the Mean 9-3.3 Choice of Sample Size 9-3.4 Likelihood Ratio Approach to Development of Test Procedures (CD Only) 9-4 HYPOTHESIS TESTS ON THE VARIANCE AND STANDARD DEVIATION OF A NORMAL POPULATION 9-4.1 The Hypothesis Testing Procedures 9-4.2 -Error and Choice of Sample Size 9-5 TESTS ON A POPULATION PROPORTION 9-5.1 Large-Sample Tests on a Proportion 9-5.2 Small-Sample Tests on a Proportion (CD Only) 9-5.3 Type II Error and Choice of Sample Size 9-6 SUMMARY TABLE OF INFERENCE PROCEDURES FOR A SINGLE SAMPLE 9-7 TESTING FOR GOODNESS OF FIT 9-8 CONTINGENCY TABLE TESTS 9-3.2 P-Value for a t-Test 277 c09.qxd 5/15/02 8:02 PM Page 278 RK UL 9 RK UL 9:Desktop Folder: 278 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE LEARNING OBJECTIVES After careful study of this chapter, you should be able to do the following: 1. Structure engineering decision-making problems as hypothesis tests 2. Test hypotheses on the mean of a normal distribution using either a Z-test or a t-test procedure 3. Test hypotheses on the variance or standard deviation of a normal distribution 4. Test hypotheses on a population proportion 5. Use the P-value approach for making decisions in hypotheses tests 6. Compute power, type II error probability, and make sample size selection decisions for tests on means, variances, and proportions 7. Explain and use the relationship between confidence intervals and hypothesis tests 8. Use the chi-square goodness of fit test to check distributional assumptions 9. Use contingency table tests CD MATERIAL 10. Appreciate the likelihood ratio approach to construction of test statistics 11. Conduct small sample tests on a population proportion Answers for many odd numbered exercises are at the end of the book. Answers to exercises whose numbers are surrounded by a box can be accessed in the e-Text by clicking on the box. Complete worked solutions to certain exercises are also available in the e-Text. These are indicated in the Answers to Selected Exercises section by a box around the exercise number. Exercises are also available for some of the text sections that appear on CD only. These exercises may be found within the e-Text immediately following the section they accompany. 9-1 HYPOTHESIS TESTING 9-1.1 Statistical Hypotheses In the previous chapter we illustrated how to construct a confidence interval estimate of a parameter from sample data. However, many problems in engineering require that we decide whether to accept or reject a statement about some parameter. The statement is called a hypothesis, and the decision-making procedure about the hypothesis is called hypothesis testing. This is one of the most useful aspects of statistical inference, since many types of decision-making problems, tests, or experiments in the engineering world can be formulated as hypothesis-testing problems. Furthermore, as we will see, there is a very close connection between hypothesis testing and confidence intervals. Statistical hypothesis testing and confidence interval estimation of parameters are the fundamental methods used at the data analysis stage of a comparative experiment, in which the engineer is interested, for example, in comparing the mean of a population to a specified value. These simple comparative experiments are frequently encountered in practice and provide a good foundation for the more complex experimental design problems that we will discuss in Chapters 13 and 14. In this chapter we discuss comparative experiments involving a single population, and our focus is on testing hypotheses concerning the parameters of the population. We now give a formal definition of a statistical hypothesis. Definition A statistical hypothesis is a statement about the parameters of one or more populations. c09.qxd 5/16/02 4:15 PM Page 279 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark Files: 9-1 HYPOTHESIS TESTING 279 Since we use probability distributions to represent populations, a statistical hypothesis may also be thought of as a statement about the probability distribution of a random variable. The hypothesis will usually involve one or more parameters of this distribution. For example, suppose that we are interested in the burning rate of a solid propellant used to power aircrew escape systems. Now burning rate is a random variable that can be described by a probability distribution. Suppose that our interest focuses on the mean burning rate (a parameter of this distribution). Specifically, we are interested in deciding whether or not the mean burning rate is 50 centimeters per second. We may express this formally as H0: 50 centimeters per second H1: 50 centimeters per second (9-1) The statement H0: 50 centimeters per second in Equation 9-1 is called the null hypothesis, and the statement H1: 50 centimeters per second is called the alternative hypothesis. Since the alternative hypothesis specifies values of that could be either greater or less than 50 centimeters per second, it is called a two-sided alternative hypothesis. In some situations, we may wish to formulate a one-sided alternative hypothesis, as in H0: 50 centimeters per second H0: 50 centimeters per second or H1: 50 centimeters per second (9-2) H1: 50 centimeters per second It is important to remember that hypotheses are always statements about the population or distribution under study, not statements about the sample. The value of the population parameter specified in the null hypothesis (50 centimeters per second in the above example) is usually determined in one of three ways. First, it may result from past experience or knowledge of the process, or even from previous tests or experiments. The objective of hypothesis testing then is usually to determine whether the parameter value has changed. Second, this value may be determined from some theory or model regarding the process under study. Here the objective of hypothesis testing is to verify the theory or model. A third situation arises when the value of the population parameter results from external considerations, such as design or engineering specifications, or from contractual obligations. In this situation, the usual objective of hypothesis testing is conformance testing. A procedure leading to a decision about a particular hypothesis is called a test of a hypothesis. Hypothesis-testing procedures rely on using the information in a random sample from the population of interest. If this information is consistent with the hypothesis, we will conclude that the hypothesis is true; however, if this information is inconsistent with the hypothesis, we will conclude that the hypothesis is false. We emphasize that the truth or falsity of a particular hypothesis can never be known with certainty, unless we can examine the entire population. This is usually impossible in most practical situations. Therefore, a hypothesis-testing procedure should be developed with the probability of reaching a wrong conclusion in mind. The structure of hypothesis-testing problems is identical in all the applications that we will consider. The null hypothesis is the hypothesis we wish to test. Rejection of the null hypothesis always leads to accepting the alternative hypothesis. In our treatment of hypothesis testing, the null hypothesis will always be stated so that it specifies an exact value of the parameter (as in the statement H0: 50 centimeters per second in Equation 9-1). The alternate hypothesis will allow the parameter to take on several values (as in the statement H1: 50 centimeters per second in Equation 9-1). Testing the hypothesis involves taking a random sample, computing a test statistic from the sample data, and then using the test statistic to make a decision about the null hypothesis. c09.qxd 6/4/02 2:26 PM Page 280 RK UL 6 RK UL 6:Desktop Folder:montgo: 280 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE 9-1.2 Tests of Statistical Hypotheses To illustrate the general concepts, consider the propellant burning rate problem introduced earlier. The null hypothesis is that the mean burning rate is 50 centimeters per second, and the alternate is that it is not equal to 50 centimeters per second. That is, we wish to test H0: 50 centimeters per second H1: 50 centimeters per second Suppose that a sample of n 10 specimens is tested and that the sample mean burning rate x is observed. The sample mean is an estimate of the true population mean . A value of the sample mean x that falls close to the hypothesized value of 50 centimeters per second is evidence that the true mean is really 50 centimeters per second; that is, such evidence supports the null hypothesis H0. On the other hand, a sample mean that is considerably different from 50 centimeters per second is evidence in support of the alternative hypothesis H1. Thus, the sample mean is the test statistic in this case. The sample mean can take on many different values. Suppose that if 48.5 x 51.5, we will not reject the null hypothesis H0: 50 , and if either x 48.5 or x 51.5, we will reject the null hypothesis in favor of the alternative hypothesis H1: 50 . This is illustrated in Fig. 9-1. The values of x that are less than 48.5 and greater than 51.5 constitute the critical region for the test, while all values that are in the interval 48.5 x 51.5 form a region for which we will fail to reject the null hypothesis. By convention, this is usually called the acceptance region. The boundaries between the critical regions and the acceptance region are called the critical values. In our example the critical values are 48.5 and 51.5. It is customary to state conclusions relative to the null hypothesis H0. Therefore, we reject H0 in favor of H1 if the test statistic falls in the critical region and fail to reject H0 otherwise. This decision procedure can lead to either of two wrong conclusions. For example, the true mean burning rate of the propellant could be equal to 50 centimeters per second. However, for the randomly selected propellant specimens that are tested, we could observe a value of the test statistic x that falls into the critical region. We would then reject the null hypothesis H0 in favor of the alternate H1 when, in fact, H0 is really true. This type of wrong conclusion is called a type I error. Definition Rejecting the null hypothesis H0 when it is true is defined as a type I error. Now suppose that the true mean burning rate is different from 50 centimeters per second, yet the sample mean x falls in the acceptance region. In this case we would fail to reject H0 when it is false. This type of wrong conclusion is called a type II error. Definition Failing to reject the null hypothesis when it is false is defined as a type II error. Thus, in testing any statistical hypothesis, four different situations determine whether the final decision is correct or in error. These situations are presented in Table 9-1. c09.qxd 5/15/02 8:02 PM Page 281 RK UL 9 RK UL 9:Desktop Folder: 9-1 HYPOTHESIS TESTING Reject H0 Fail to Reject H0 Reject H0 µ ≠ 50 cm/s µ = 50 cm/s µ ≠ 50 cm/s 48.5 50 Table 9-1 Decisions in Hypothesis Testing x 51.5 281 Figure 9-1 Decision criteria for testing H0: 50 centimeters per second versus H1: 50 centimeters per second. Decision H0 Is True H0 Is False Fail to reject H0 Reject H0 no error type I error type II error no error Because our decision is based on random variables, probabilities can be associated with the type I and type II errors in Table 9-1. The probability of making a type I error is denoted by the Greek letter . That is, P(type I error) P(reject H0 when H0 is true) (9-3) Sometimes the type I error probability is called the significance level, or the -error, or the size of the test. In the propellant burning rate example, a type I error will occur when either x 51.5 or x 48.5 when the true mean burning rate is 50 centimeters per second. Suppose that the standard deviation of burning rate is 2.5 centimeters per second and that the burning rate has a distribution for which the conditions of the central limit theorem apply, so the distribution of the sample mean is approximately normal with mean 50 and standard deviation 1n 2.5 110 0.79. The probability of making a type I error (or the significance level of our test) is equal to the sum of the areas that have been shaded in the tails of the normal distribution in Fig. 9-2. We may find this probability as P1X 48.5 when 502 P1X 51.5 when 502 The z-values that correspond to the critical values 48.5 and 51.5 are z1 48.5 50 1.90 0.79 and z2 51.5 50 1.90 0.79 Therefore P1Z 1.902 P1Z 1.902 0.028717 0.028717 0.057434 This implies that 5.76% of all random samples would lead to rejection of the hypothesis H0: 50 centimeters per second when the true mean burning rate is really 50 centimeters per second. α /2 = 0.0287 α /2 = 0.0287 48.5 µ = 50 51.5 X Figure 9-2 The critical region for H0: 50 versus H1: 50 and n 10. c09.qxd 5/15/02 8:02 PM Page 282 RK UL 9 RK UL 9:Desktop Folder: 282 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE From inspection of Fig. 9-2, notice that we can reduce by widening the acceptance region. For example, if we make the critical values 48 and 52, the value of is 48 50 52 50 b P aZ b P1Z 2.532 P1Z 2.532 0.79 0.79 0.0057 0.0057 0.0114 P aZ We could also reduce by increasing the sample size. If n 16, 1n 2.5 116 0.625, and using the original critical region from Fig. 9-1, we find z1 48.5 50 2.40 0.625 and z2 51.5 50 2.40 0.625 Therefore P1Z 2.402 P1Z 2.402 0.0082 0.0082 0.0164 In evaluating a hypothesis-testing procedure, it is also important to examine the probability of a type II error, which we will denote by . That is, P(type II error) P(fail to reject H0 when H0 is false) (9-4) To calculate (sometimes called the -error), we must have a specific alternative hypothesis; that is, we must have a particular value of . For example, suppose that it is important to reject the null hypothesis H0: 50 whenever the mean burning rate is greater than 52 centimeters per second or less than 48 centimeters per second. We could calculate the probability of a type II error for the values 52 and 48 and use this result to tell us something about how the test procedure would perform. Specifically, how will the test procedure work if we wish to detect, that is, reject H0, for a mean value of 52 or 48? Because of symmetry, it is necessary only to evaluate one of the two cases—say, find the probability of accepting the null hypothesis H0: 50 centimeters per second when the true mean is 52 centimeters per second. Figure 9-3 will help us calculate the probability of type II error . The normal distribution on the left in Fig. 9-3 is the distribution of the test statistic X when the null hypothesis H0: 50 is true (this is what is meant by the expression “under H0: 50”), and the normal distribution on the right is the distribution of X when the alternative hypothesis is true and the value of the mean is 52 (or “under H1: 52”). Now a type II error will be committed if the sample mean X falls between 48.5 and 51.5 (the critical region boundaries) when 52. As seen in Fig. 9-3, this is just the probability that 48.5 X 51.5 when the true mean is 52, or the shaded area under the normal distribution on the right. Therefore, referring to Fig. 9-3, we find that P148.5 X 51.5 when 522 c09.qxd 5/15/02 8:02 PM Page 283 RK UL 9 RK UL 9:Desktop Folder: 283 9-1 HYPOTHESIS TESTING 0.6 0.6 0.5 Probability density Probability density 0.5 Under H1:µ = 52 Under H0: µ = 50 0.4 0.3 0.2 Under H1: µ = 50.5 0.4 0.3 0.2 0.1 0.1 0 46 Under H0: µ = 50 48 50 52 54 0 46 56 48 50 –x Figure 9-3 The probability of type II error when 52 and n 10. 52 54 56 x– Figure 9-4 The probability of type II error when 50.5 and n 10. The z-values corresponding to 48.5 and 51.5 when 52 are z1 48.5 52 4.43 0.79 and z2 51.5 52 0.63 0.79 Therefore P 1 4.43 Z 0.632 P 1Z 0.632 P 1Z 4.432 0.2643 0.0000 0.2643 Thus, if we are testing H0: 50 against H1: 50 with n 10, and the true value of the mean is 52, the probability that we will fail to reject the false null hypothesis is 0.2643. By symmetry, if the true value of the mean is 48, the value of will also be 0.2643. The probability of making a type II error increases rapidly as the true value of approaches the hypothesized value. For example, see Fig. 9-4, where the true value of the mean is 50.5 and the hypothesized value is H0: 50. The true value of is very close to 50, and the value for is P 148.5 X 51.5 when 50.52 As shown in Fig. 9-4, the z-values corresponding to 48.5 and 51.5 when 50.5 are z1 48.5 50.5 2.53 0.79 and z2 51.5 50.5 1.27 0.79 Therefore P1 2.53 Z 1.272 P1Z 1.272 P1Z 2.532 0.8980 0.0057 0.8923 Thus, the type II error probability is much higher for the case where the true mean is 50.5 centimeters per second than for the case where the mean is 52 centimeters per second. Of course, c09.qxd 5/15/02 8:02 PM Page 284 RK UL 9 RK UL 9:Desktop Folder: 284 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE 0.8 Under H1: µ = 52 Probability density Under H0: µ = 50 0.6 0.4 0.2 Figure 9-5 The probability of type II error when 52 and n 16. 0 46 48 50 52 54 56 x– in many practical situations we would not be as concerned with making a type II error if the mean were “close” to the hypothesized value. We would be much more interested in detecting large differences between the true mean and the value specified in the null hypothesis. The type II error probability also depends on the sample size n. Suppose that the null hypothesis is H0: 50 centimeters per second and that the true value of the mean is 52. If the sample size is increased from n 10 to n 16, the situation of Fig. 9-5 results. The normal distribution on the left is the distribution of X when the mean 50, and the normal distribution on the right is the distribution of X when 52. As shown in Fig. 9-5, the type II error probability is P 148.5 X 51.5 when 522 When n 16, the standard deviation of X is 1n 2.5 116 0.625, and the z-values corresponding to 48.5 and 51.5 when 52 are z1 48.5 52 51.5 52 5.60 and z2 0.80 0.625 0.625 Therefore P1 5.60 Z 0.802 P1Z 0.802 P1Z 5.602 0.2119 0.0000 0.2119 Recall that when n 10 and 52, we found that 0.2643; therefore, increasing the sample size results in a decrease in the probability of type II error. The results from this section and a few other similar calculations are summarized in the following table: Acceptance Region Sample Size 48.5 x 51.5 10 0.0576 0.2643 0.8923 10 0.0114 0.5000 0.9705 16 0.0164 0.2119 0.9445 16 0.0014 0.5000 0.9918 48 x 52 48.5 x 51.5 48 x 52 at 52 at 50.5 c09.qxd 5/15/02 8:02 PM Page 285 RK UL 9 RK UL 9:Desktop Folder: 9-1 HYPOTHESIS TESTING 285 The results in boxes were not calculated in the text but can easily be verified by the reader. This display and the discussion above reveal four important points: The size of the critical region, and consequently the probability of a type I error , can always be reduced by appropriate selection of the critical values. 2. Type I and type II errors are related. A decrease in the probability of one type of error always results in an increase in the probability of the other, provided that the sample size n does not change. 3. An increase in sample size will generally reduce both and , provided that the critical values are held constant. 4. When the null hypothesis is false, increases as the true value of the parameter approaches the value hypothesized in the null hypothesis. The value of decreases as the difference between the true mean and the hypothesized value increases. 1. Generally, the analyst controls the type I error probability when he or she selects the critical values. Thus, it is usually easy for the analyst to set the type I error probability at (or near) any desired value. Since the analyst can directly control the probability of wrongly rejecting H0, we always think of rejection of the null hypothesis H0 as a strong conclusion. On the other hand, the probability of type II error is not a constant, but depends on the true value of the parameter. It also depends on the sample size that we have selected. Because the type II error probability is a function of both the sample size and the extent to which the null hypothesis H0 is false, it is customary to think of the decision to accept H0 as a weak conclusion, unless we know that is acceptably small. Therefore, rather than saying we “accept H0”, we prefer the terminology “fail to reject H0”. Failing to reject H0 implies that we have not found sufficient evidence to reject H0, that is, to make a strong statement. Failing to reject H0 does not necessarily mean that there is a high probability that H0 is true. It may simply mean that more data are required to reach a strong conclusion. This can have important implications for the formulation of hypotheses. An important concept that we will make use of is the power of a statistical test. Definition The power of a statistical test is the probability of rejecting the null hypothesis H0 when the alternative hypothesis is true. The power is computed as 1 , and power can be interpreted as the probability of correctly rejecting a false null hypothesis. We often compare statistical tests by comparing their power properties. For example, consider the propellant burning rate problem when we are testing H0: 50 centimeters per second against H1: 50 centimeters per second. Suppose that the true value of the mean is 52. When n 10, we found that 0.2643, so the power of this test is 1 1 0.2643 0.7357 when 52. Power is a very descriptive and concise measure of the sensitivity of a statistical test, where by sensitivity we mean the ability of the test to detect differences. In this case, the sensitivity of the test for detecting the difference between a mean burning rate of 50 centimeters per second and 52 centimeters per second is 0.7357. That is, if the true mean is really 52 centimeters per second, this test will correctly reject H0: 50 and “detect” this difference 73.57% of the time. If this value of power is judged to be too low, the analyst can increase either or the sample size n. c09.qxd 5/15/02 8:02 PM Page 286 RK UL 9 RK UL 9:Desktop Folder: 286 9-1.3 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE One-Sided and Two-Sided Hypotheses A test of any hypothesis such as H0: 0 H1: 0 is called a two-sided test, because it is important to detect differences from the hypothesized value of the mean 0 that lie on either side of 0 . In such a test, the critical region is split into two parts, with (usually) equal probability placed in each tail of the distribution of the test statistic. Many hypothesis-testing problems naturally involve a one-sided alternative hypothesis, such as H0: 0 H1: 0 or H0: 0 H1: 0 If the alternative hypothesis is H1: 0 , the critical region should lie in the upper tail of the distribution of the test statistic, whereas if the alternative hypothesis is H1: 0, the critical region should lie in the lower tail of the distribution. Consequently, these tests are sometimes called one-tailed tests. The location of the critical region for one-sided tests is usually easy to determine. Simply visualize the behavior of the test statistic if the null hypothesis is true and place the critical region in the appropriate end or tail of the distribution. Generally, the inequality in the alternative hypothesis “points” in the direction of the critical region. In constructing hypotheses, we will always state the null hypothesis as an equality so that the probability of type I error can be controlled at a specific value. The alternative hypothesis might be either one-sided or two-sided, depending on the conclusion to be drawn if H0 is rejected. If the objective is to make a claim involving statements such as greater than, less than, superior to, exceeds, at least, and so forth, a one-sided alternative is appropriate. If no direction is implied by the claim, or if the claim not equal to is to be made, a two-sided alternative should be used. EXAMPLE 9-1 Consider the propellant burning rate problem. Suppose that if the burning rate is less than 50 centimeters per second, we wish to show this with a strong conclusion. The hypotheses should be stated as H0: 50 centimeters per second H1: 50 centimeters per second Here the critical region lies in the lower tail of the distribution of X . Since the rejection of H0 is always a strong conclusion, this statement of the hypotheses will produce the desired outcome if H0 is rejected. Notice that, although the null hypothesis is stated with an equal sign, it is understood to include any value of not specified by the alternative hypothesis. Therefore, failing to reject H0 does not mean that 50 centimeters per second exactly, but only that we do not have strong evidence in support of H1. In some real-world problems where one-sided test procedures are indicated, it is occasionally difficult to choose an appropriate formulation of the alternative hypothesis. For example, suppose that a soft-drink beverage bottler purchases 10-ounce bottles from a glass c09.qxd 5/16/02 4:15 PM Page 287 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark Files: 9-1 HYPOTHESIS TESTING 287 company. The bottler wants to be sure that the bottles meet the specification on mean internal pressure or bursting strength, which for 10-ounce bottles is a minimum strength of 200 psi. The bottler has decided to formulate the decision procedure for a specific lot of bottles as a hypothesis testing problem. There are two possible formulations for this problem, either H0: 200 psi H1: 200 psi (9-5) or H0: 200 psi H1: 200 psi (9-6) Consider the formulation in Equation 9-5. If the null hypothesis is rejected, the bottles will be judged satisfactory; if H0 is not rejected, the implication is that the bottles do not conform to specifications and should not be used. Because rejecting H0 is a strong conclusion, this formulation forces the bottle manufacturer to “demonstrate” that the mean bursting strength of the bottles exceeds the specification. Now consider the formulation in Equation 9-6. In this situation, the bottles will be judged satisfactory unless H0 is rejected. That is, we conclude that the bottles are satisfactory unless there is strong evidence to the contrary. Which formulation is correct, the one of Equation 9-5 or Equation 9-6? The answer is it depends. For Equation 9-5, there is some probability that H0 will not be rejected (i.e., we would decide that the bottles are not satisfactory), even though the true mean is slightly greater than 200 psi. This formulation implies that we want the bottle manufacturer to demonstrate that the product meets or exceeds our specifications. Such a formulation could be appropriate if the manufacturer has experienced difficulty in meeting specifications in the past or if product safety considerations force us to hold tightly to the 200 psi specification. On the other hand, for the formulation of Equation 9-6 there is some probability that H0 will be accepted and the bottles judged satisfactory, even though the true mean is slightly less than 200 psi. We would conclude that the bottles are unsatisfactory only when there is strong evidence that the mean does not exceed 200 psi, that is, when H0: 200 psi is rejected. This formulation assumes that we are relatively happy with the bottle manufacturer’s past performance and that small deviations from the specification of 200 psi are not harmful. In formulating one-sided alternative hypotheses, we should remember that rejecting H0 is always a strong conclusion. Consequently, we should put the statement about which it is important to make a strong conclusion in the alternative hypothesis. In real-world problems, this will often depend on our point of view and experience with the situation. 9-1.4 General Procedure for Hypothesis Tests This chapter develops hypothesis-testing procedures for many practical problems. Use of the following sequence of steps in applying hypothesis-testing methodology is recommended. 1. 2. From the problem context, identify the parameter of interest. State the null hypothesis, H0. 3. Specify an appropriate alternative hypothesis, H1. 4. Choose a significance level . c09.qxd 5/15/02 8:02 PM Page 288 RK UL 9 RK UL 9:Desktop Folder: 288 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE 5. 6. 7. 8. Determine an appropriate test statistic. State the rejection region for the statistic. Compute any necessary sample quantities, substitute these into the equation for the test statistic, and compute that value. Decide whether or not H0 should be rejected and report that in the problem context. Steps 1–4 should be completed prior to examination of the sample data. This sequence of steps will be illustrated in subsequent sections. EXERCISES FOR SECTION 9-1 9-1. In each of the following situations, state whether it is a correctly stated hypothesis testing problem and why. (a) H0: 25, H1: 25 (b) H0: 10, H1: 10 (c) H0: x 50, H1: x 50 (d) H0: p 0.1, H1: p 0.5 (e) H0: s 30, H1: s 30 9-2. A textile fiber manufacturer is investigating a new drapery yarn, which the company claims has a mean thread elongation of 12 kilograms with a standard deviation of 0.5 kilograms. The company wishes to test the hypothesis H0: 12 against H1: 12, using a random sample of four specimens. (a) What is the type I error probability if the critical region is defined as x 11.5 kilograms? (b) Find for the case where the true mean elongation is 11.25 kilograms. 9-3. Repeat Exercise 9-2 using a sample size of n = 16 and the same critical region. 9-4. In Exercise 9-2, find the boundary of the critical region if the type I error probability is specified to be 0.01. 9-5. In Exercise 9-2, find the boundary of the critical region if the type I error probability is specified to be 0.05. 9-6. The heat evolved in calories per gram of a cement mixture is approximately normally distributed. The mean is thought to be 100 and the standard deviation is 2. We wish to test H0: 100 versus H1: 100 with a sample of n = 9 specimens. (a) If the acceptance region is defined as 98.5 x 101.5 , find the type I error probability . (b) Find for the case where the true mean heat evolved is 103. (c) Find for the case where the true mean heat evolved is 105. This value of is smaller than the one found in part (b) above. Why? 9-7. Repeat Exercise 9-6 using a sample size of n 5 and the same acceptance region. 9-8. A consumer products company is formulating a new shampoo and is interested in foam height (in millimeters). Foam height is approximately normally distributed and has a standard deviation of 20 millimeters. The company wishes to test H0: 175 millimeters versus H1: 175 millimeters, using the results of n 10 samples. (a) Find the type I error probability if the critical region is x 185 . (b) What is the probability of type II error if the true mean foam height is 195 millimeters? 9-9. In Exercise 9-8, suppose that the sample data result in x 190 millimeters. (a) What conclusion would you reach? (b) How “unusual” is the sample value x 190 millimeters if the true mean is really 175 millimeters? That is, what is the probability that you would observe a sample average as large as 190 millimeters (or larger), if the true mean foam height was really 175 millimeters? 9-10. Repeat Exercise 9-8 assuming that the sample size is n 16 and the boundary of the critical region is the same. 9-11. Consider Exercise 9-8, and suppose that the sample size is increased to n 16. (a) Where would the boundary of the critical region be placed if the type I error probability were to remain equal to the value that it took on when n 10? (b) Using n 16 and the new critical region found in part (a), find the type II error probability if the true mean foam height is 195 millimeters. (c) Compare the value of obtained in part (b) with the value from Exercise 9-8 (b). What conclusions can you draw? 9-12. A manufacturer is interested in the output voltage of a power supply used in a PC. Output voltage is assumed to be normally distributed, with standard deviation 0.25 Volts, and the manufacturer wishes to test H0: 5 Volts against H1: 5 Volts, using n 8 units. (a) The acceptance region is 4.85 x 5.15. Find the value of . (b) Find the power of the test for detecting a true mean output voltage of 5.1 Volts. 9-13. Rework Exercise 9-12 when the sample size is 16 and the boundaries of the acceptance region do not change. 9-14. Consider Exercise 9-12, and suppose that the manufacturer wants the type I error probability for the test to be 0.05. Where should the acceptance region be located? c09.qxd 23/7/02 9:25 M Page 289 9-2 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE KNOWN 9-15. If we plot the probability of accepting H0: 0 versus various values of and connect the points with a smooth curve, we obtain the operating characteristic curve (or the OC curve) of the test procedure. These curves are used extensively in industrial applications of hypothesis testing to display the sensitivity and relative performance of the test. When the true mean is really equal to 0, the probability of accepting H0 is 1 . Construct an OC curve for Exercise 9-8, using values of the true mean of 178, 181, 184, 187, 190, 193, 196, and 199. 9-16. Convert the OC curve in Exercise 9-15 into a plot of the power function of the test. 9-17. 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 400 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? Hint: use the normal approximation to the binomial. 289 9-18. The proportion of residents in Phoenix favoring the building of toll roads to complete the freeway system is believed to be p 0.3. If a random sample of 10 residents shows that 1 or fewer favor this proposal, we will conclude that p 0.3. (a) Find the probability of type I error if the true proportion is p 0.3. (b) Find the probability of committing a type II error with this procedure if p 0.2. (c) What is the power of this procedure if the true proportion is p 0.2? 9-19. The proportion of adults living in Tempe, Arizona, who are college graduates is estimated to be p 0.4. To test this hypothesis, a random sample of 15 Tempe adults is selected. If the number of college graduates is between 4 and 8, the hypothesis will be accepted; otherwise, we will conclude that p 0.4 . (a) Find the type I error probability for this procedure, assuming that p 0.4. (b) Find the probability of committing a type II error if the true proportion is really p 0.2. 9-2 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE KNOWN In this section, we consider hypothesis testing about the mean of a single, normal population where the variance of the population 2 is known. We will assume that a random sample X1, X2, p , Xn has been taken from the population. Based on our previous discussion, the sample mean X is an unbiased point estimator of with variance 2 n. 9-2.1 Hypothesis Tests on the Mean Suppose that we wish to test the hypotheses H0: 0 H1: 0 (9-7) where 0 is a specified constant. We have a random sample X1, X2, p , Xn from a normal population. Since X has a normal distribution (i.e., the sampling distribution of X is normal) with mean 0 and standard deviation 1n if the null hypothesis is true, we could construct a critical region based on the computed value of the sample mean X , as in Section 9-1.2. It is usually more convenient to standardize the sample mean and use a test statistic based on the standard normal distribution. That is, the test procedure for H0: 0 uses the test statistic Z0 X 0 1n (9-8) c09.qxd 5/15/02 8:02 PM Page 290 RK UL 9 RK UL 9:Desktop Folder: 290 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE N(0,1) Critical region N(0,1) N(0,1) Critical region α /2 α /2 Acceptance region – zα /2 0 zα /2 Critical region α Acceptance region Z0 0 (a) zα α Z0 Acceptance region –zα (b) Z0 0 (c) Figure 9-6 The distribution of Z0 when H0: 0 is true, with critical region for (a) the two-sided alternative H1 : 0, (b) the one-sided alternative H1 : 0, and (c) the one-sided alternative H1 : 0. If the null hypothesis H0: 0 is true, E1X 2 0 , and it follows that the distribution of Z0 is the standard normal distribution [denoted N(0, 1)]. Consequently, if H0: 0 is true, the probability is 1 that the test statistic Z0 falls between z2 and z2 , where z2 is the 1002 percentage point of the standard normal distribution. The regions associated with z2 and z2 are illustrated in Fig. 9-6(a). Note that the probability is that the test statistic Z0 will fall in the region Z0 z2 or Z0 z2 when H0: 0 is true. Clearly, a sample producing a value of the test statistic that falls in the tails of the distribution of Z0 would be unusual if H0: 0 is true; therefore, it is an indication that H0 is false. Thus, we should reject H0 if the observed value of the test statistic z0 is either z0 z2 or z0 z2 (9-9) and we should fail to reject H0 if z2 z0 z2 (9-10) The inequalities in Equation 9-10 defines the acceptance region for H0, and the two inequalities in Equation 9-9 define the critical region or rejection region. The type I error probability for this test procedure is . It is easier to understand the critical region and the test procedure, in general, when the test statistic is Z0 rather than X . However, the same critical region can always be written in terms of the computed value of the sample mean x. A procedure identical to the above is as follows: Reject H0: 0 if either x a or x b where a 0 z2 1n and EXAMPLE 9-2 b 0 z2 1n Aircrew escape systems are powered by a solid propellant. The burning rate of this propellant is an important product characteristic. Specifications require that the mean burning rate must be 50 centimeters per second. We know that the standard deviation of burning rate is 2 centimeters per second. The experimenter decides to specify a type I error probability or significance level of 0.05 and selects a random sample of n 25 and obtains a sample average burning rate of x 51.3 centimeters per second. What conclusions should be drawn? c09.qxd 5/15/02 8:02 PM Page 291 RK UL 9 RK UL 9:Desktop Folder: 9-2 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE KNOWN 291 We may solve this problem by following the eight-step procedure outlined in Section 9-1.4. This results in 1. The parameter of interest is , the mean burning rate. 2. H0: 50 centimeters per second 3. H1: 50 centimeters per second 4. 0.05 5. The test statistic is z0 x 0 1n 6. Reject H0 if z0 1.96 or if z0 1.96. Note that this results from step 4, where we specified 0.05, and so the boundaries of the critical region are at z0.025 1.96 and z0.025 1.96. 7. Computations: Since x 51.3 and 2, z0 8. 51.3 50 2 225 3.25 Conclusion: Since z0 3.25 1.96, we reject H0: 50 at the 0.05 level of significance. Stated more completely, we conclude that the mean burning rate differs from 50 centimeters per second, based on a sample of 25 measurements. In fact, there is strong evidence that the mean burning rate exceeds 50 centimeters per second. We may also develop procedures for testing hypotheses on the mean where the alternative hypothesis is one-sided. Suppose that we specify the hypotheses as H0: 0 H1: 0 (9-11) In defining the critical region for this test, we observe that a negative value of the test statistic Z0 would never lead us to conclude that H0: 0 is false. Therefore, we would place the critical region in the upper tail of the standard normal distribution and reject H0 if the computed value of z0 is too large. That is, we would reject H0 if z0 z (9-12) as shown in Figure 9-6(b). Similarly, to test H0: 0 H1: 0 (9-13) we would calculate the test statistic Z0 and reject H0 if the value of z0 is too small. That is, the critical region is in the lower tail of the standard normal distribution as shown in Figure 9-6(c), and we reject H0 if z0 z (9-14) c09.qxd 5/15/02 8:02 PM Page 292 RK UL 9 RK UL 9:Desktop Folder: 292 9-2.2 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE P-Values in Hypothesis Tests One way to report the results of a hypothesis test is to state that the null hypothesis was or was not rejected at a specified -value or level of significance. For example, in the propellant problem above, we can say that H0: 50 was rejected at the 0.05 level of significance. This statement of conclusions is often inadequate because it gives the decision maker no idea about whether the computed value of the test statistic was just barely in the rejection region or whether it was very far into this region. Furthermore, stating the results this way imposes the predefined level of significance on other users of the information. This approach may be unsatisfactory because some decision makers might be uncomfortable with the risks implied by 0.05. To avoid these difficulties the P-value approach has been adopted widely in practice. The P-value is the probability that the test statistic will take on a value that is at least as extreme as the observed value of the statistic when the null hypothesis H0 is true. Thus, a P-value conveys much information about the weight of evidence against H0, and so a decision maker can draw a conclusion at any specified level of significance. We now give a formal definition of a P-value. Definition The P-value is the smallest level of significance that would lead to rejection of the null hypothesis H0 with the given data. It is customary to call the test statistic (and the data) significant when the null hypothesis H0 is rejected; therefore, we may think of the P-value as the smallest level at which the data are significant. Once the P-value is known, the decision maker can determine how significant the data are without the data analyst formally imposing a preselected level of significance. For the foregoing normal distribution tests it is relatively easy to compute the P-value. If z0 is the computed value of the test statistic, the P-value is 2 31 1|z0|2 4 P • 1 1z0 2 1z0 2 for a two-tailed test: H0: 0 for a upper-tailed test: H0: 0 for a lower-tailed test: H0: 0 H1: 0 H1: 0 H1: 0 (9-15) Here, 1z2 is the standard normal cumulative distribution function defined in Chapter 4. Recall that 1z2 P1Z z2 , where Z is N(0, 1). To illustrate this, consider the propellant problem in Example 9-2. The computed value of the test statistic is z0 3.25 and since the alternative hypothesis is two-tailed, the P-value is P-value 231 13.252 4 0.0012 Thus, H0: 50 would be rejected at any level of significance P-value 0.0012. For example, H0 would be rejected if 0.01, but it would not be rejected if 0.001. It is not always easy to compute the exact P-value for a test. However, most modern computer programs for statistical analysis report P-values, and they can be obtained on some hand-held calculators. We will also show how to approximate the P-value. Finally, if the c09.qxd 5/15/02 8:02 PM Page 293 RK UL 9 RK UL 9:Desktop Folder: 9-2 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE KNOWN 293 P-value approach is used, step 6 of the hypothesis-testing procedure can be modified. Specifically, it is not necessary to state explicitly the critical region. 9-2.3 Connection between Hypothesis Tests and Confidence Intervals There is a close relationship between the test of a hypothesis about any parameter, say , and the confidence interval for . If [l, u] is a 10011 2 % confidence interval for the parameter , the test of size of the hypothesis H0: 0 H1: 0 will lead to rejection of H0 if and only if 0 is not in the 100 11 2 % CI [l, u]. As an illustration, consider the escape system propellant problem discussed above. The null hypothesis H0: 50 was rejected, using 0.05. The 95% two-sided CI on can be calculated using Equation 8-7. This CI is 50.52 52.08. Because the value 0 50 is not included in this interval, the null hypothesis H0: 50 is rejected. Although hypothesis tests and CIs are equivalent procedures insofar as decision making or inference about is concerned, each provides somewhat different insights. For instance, the confidence interval provides a range of likely values for at a stated confidence level, whereas hypothesis testing is an easy framework for displaying the risk levels such as the P-value associated with a specific decision. We will continue to illustrate the connection between the two procedures throughout the text. 9-2.4 Type II Error and Choice of Sample Size In testing hypotheses, the analyst directly selects the type I error probability. However, the probability of type II error depends on the choice of sample size. In this section, we will show how to calculate the probability of type II error . We will also show how to select the sample size to obtain a specified value of . Finding the Probability of Type II Error Consider the two-sided hypothesis H0: 0 H1: 0 Suppose that the null hypothesis is false and that the true value of the mean is 0 , say, where 0. The test statistic Z0 is Z0 X 10 2 X 0 1n 1n 1n Therefore, the distribution of Z0 when H1 is true is 1n Z0 N a , 1b (9-16) c09.qxd 5/15/02 8:02 PM Page 294 RK UL 9 RK UL 9:Desktop Folder: 294 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE Under H0 : µ = µ 0 Under H1: µ ≠ µ 0 N(0,1) N (δ σ√n , 1( β Figure 9-7 The distribution of Z0 under H0 and H1. –zα /2 0 zα /2 δ √n σ Z0 The distribution of the test statistic Z0 under both the null hypothesis H0 and the alternate hypothesis H1 is shown in Fig. 9-7. From examining this figure, we note that if H1 is true, a type II error will be made only if z2 Z0 z2 where Z0 N11n , 12 . That is, the probability of the type II error is the probability that Z0 falls between z2 and z2 given that H1 is true. This probability is shown as the shaded portion of Fig. 9-7. Expressed mathematically, this probability is az2 1n 1n b a z2 b (9-17) where 1z2 denotes the probability to the left of z in the standard normal distribution. Note that Equation 9-17 was obtained by evaluating the probability that Z0 falls in the interval 3 z2, z2 4 when H1 is true. Furthermore, note that Equation 9-17 also holds if 0, due to the symmetry of the normal distribution. It is also possible to derive an equation similar to Equation 9-17 for a one-sided alternative hypothesis. Sample Size Formulas One may easily obtain formulas that determine the appropriate sample size to obtain a particular value of for a given and . For the two-sided alternative hypothesis, we know from Equation 9-17 that az2 1n 1n b a z2 b or if 0, az2 1n b (9-18) since 1 z2 1n2 0 when is positive. Let z be the 100 upper percentile of the standard normal distribution. Then, 1 z 2 . From Equation 9-18 z z2 or 1n c09.qxd 5/15/02 8:02 PM Page 295 RK UL 9 RK UL 9:Desktop Folder: 295 9-2 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE KNOWN 1z2 z 2 2 2 n 2 (9-19) where 0 This approximation is good when 1 z2 1n 2 is small compared to . For either of the one-sided alternative hypotheses the sample size required to produce a specified type II error with probability given and is n 1z z 2 2 2 2 (9-20) where 0 EXAMPLE 9-3 Consider the rocket propellant problem of Example 9-2. Suppose that the analyst wishes to design the test so that if the true mean burning rate differs from 50 centimeters per second by as much as 1 centimeter per second, the test will detect this (i.e., reject H0: 50) with a high probability, say 0.90. Now, we note that 2, 51 50 1, 0.05, and 0.10. Since z2 z0.025 1.96 and z z0.10 1.28, the sample size required to detect this departure from H0: 50 is found by Equation 9-19 as n 1z2 z 2 2 2 2 11.96 1.282 2 22 42 112 2 The approximation is good here, since 1 z2 1n 2 1 1.96 112 142 22 1 5.202 0, which is small relative to . Using Operating Characteristic Curves When performing sample size or type II error calculations, it is sometimes more convenient to use the operating characteristic curves in Appendix Charts VIa and VIb. These curves plot as calculated from Equation 9-17 against a parameter d for various sample sizes n. Curves are provided for both 0.05 and 0.01. The parameter d is defined as d 0 0 0 0 0 (9-21) c09.qxd 5/16/02 4:15 PM Page 296 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark Files: 296 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE so one set of operating characteristic curves can be used for all problems regardless of the values of 0 and . From examining the operating characteristic curves or Equation 9-17 and Fig. 9-7, we note that The further the true value of the mean is from 0, the smaller the probability of type II error for a given n and . That is, we see that for a specified sample size and , large differences in the mean are easier to detect than small ones. 2. For a given and , the probability of type II error decreases as n increases. That is, to detect a specified difference in the mean, we may make the test more powerful by increasing the sample size. 1. EXAMPLE 9-4 Consider the propellant problem in Example 9-2. Suppose that the analyst is concerned about the probability of type II error if the true mean burning rate is 51 centimeters per second. We may use the operating characteristic curves to find . Note that 51 50 1, n 25, 2, and 0.05. Then using Equation 9-21 gives d 0 0 0 00 1 2 and from Appendix Chart VIa, with n 25, we find that 0.30. That is, if the true mean burning rate is 51 centimeters per second, there is approximately a 30% chance that this will not be detected by the test with n 25. EXAMPLE 9-5 Once again, consider the propellant problem in Example 9-2. Suppose that the analyst would like to design the test so that if the true mean burning rate differs from 50 centimeters per second by as much as 1 centimeter per second, the test will detect this (i.e., reject H0: 50) with a high probability, say, 0.90. This is exactly the same requirement as in Example 9-3, where we used Equation 9-19 to find the required sample size to be n 42. The operating characteristic curves can also be used to find the sample size for this test. Since d 0 0 0 12, 0.05, and 0.10, we find from Appendix Chart VIa that the required sample size is approximately n 40. This closely agrees with the sample size calculated from Equation 9-19. In general, the operating characteristic curves involve three parameters: , d, and n. Given any two of these parameters, the value of the third can be determined. There are two typical applications of these curves: 1. 2. For a given n and d, find (as illustrated in Example 9-3). This kind of problem is often encountered when the analyst is concerned about the sensitivity of an experiment already performed, or when sample size is restricted by economic or other factors. For a given and d, find n. This was illustrated in Example 9-4. This kind of problem is usually encountered when the analyst has the opportunity to select the sample size at the outset of the experiment. Operating characteristic curves are given in Appendix Charts VIc and VId for the onesided alternatives. If the alternative hypothesis is either H1: 0 or H1: 0, the abscissa scale on these charts is d 0 0 0 (9-22) c09.qxd 5/15/02 8:02 PM Page 297 RK UL 9 RK UL 9:Desktop Folder: 9-2 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE KNOWN 297 Using the Computer Many statistics software packages will calculate sample sizes and type II error probabilities. To illustrate, here are some computations from Minitab for the propellant burning rate problem. Power and Sample Size 1-Sample Z Test Testing mean null (versus not null) Calculating power for mean null + difference Alpha 0.05 Sigma 2 Sample Target Actual Difference Size Power Power 1 43 0.9000 0.9064 Power and Sample Size 1-Sample Z Test Testing mean null (versus not null) Calculating power for mean null difference Alpha 0.05 Sigma 2 Sample Target Actual Difference Size Power Power 1 28 0.7500 0.7536 Power and Sample Size 1-Sample Z Test Testing mean null (versus not null) Calculating power for mean null difference Alpha 0.05 Sigma 2 Sample Difference Size Power 1 25 0.7054 In the first part of the boxed display, we asked Minitab to work Example 9-3, that is, to find the sample size n that would allow detection of a difference from 0 50 of 1 centimeter per second with power of 0.9 and 0.05. The answer, n 43, agrees closely with the calculated value from Equation 9-19 in Example 9-3, which was n 42. The difference is due to Minitab using a value of z that has more than two decimal places. The second part of the computer output relaxes the power requirement to 0.75. Note that the effect is to reduce the required sample size to n 28. The third part of the output is the solution to Example 9-4, where we wish to determine the type II error probability of () or the power 1 for the sample size n 25. Note that Minitab computes the power to be 0.7054, which agrees closely with the answer obtained from the O.C. curve in Example 9-4. Generally, however, the computer calculations will be more accurate than visually reading values from an O.C. curve. 9-2.5 Large-Sample Test We have developed the test procedure for the null hypothesis H0: 0 assuming that the population is normally distributed and that 2 is known. In many if not most practical situations 2 c09.qxd 5/15/02 8:02 PM Page 298 RK UL 9 RK UL 9:Desktop Folder: 298 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE will be unknown. Furthermore, we may not be certain that the population is well modeled by a normal distribution. In these situations if n is large (say n 40) the sample standard deviation s can be substituted for in the test procedures with little effect. Thus, while we have given a test for the mean of a normal distribution with known 2, it can be easily converted into a largesample test procedure for unknown 2 that is valid regardless of the form of the distribution of the population. This large-sample test relies on the central limit theorem just as the largesample confidence interval on that was presented in the previous chapter did. Exact treatment of the case where the population is normal, 2 is unknown, and n is small involves use of the t distribution and will be deferred until Section 9-3. 9-2.6 Some Practical Comments on Hypothesis Tests The Eight-Step Procedure In Section 9-1.4 we described an eight-step procedure for statistical hypothesis testing. This procedure was illustrated in Example 9-2 and will be encountered many times in both this chapter and Chapter 10. In practice, such a formal and (seemingly) rigid procedure is not always necessary. Generally, once the experimenter (or decision maker) has decided on the question of interest and has determined the design of the experiment (that is, how the data are to be collected, how the measurements are to be made, and how many observations are required), only three steps are really required: 1. Specify the test statistic to be used (such as Z0). 2. Specify the location of the critical region (two-tailed, upper-tailed, or lower-tailed). 3. Specify the criteria for rejection (typically, the value of , or the P-value at which rejection should occur). These steps are often completed almost simultaneously in solving real-world problems, although we emphasize that it is important to think carefully about each step. That is why we present and use the eight-step process: it seems to reinforce the essentials of the correct approach. While you may not use it every time in solving real problems, it is a helpful framework when you are first learning about hypothesis testing. Statistical versus Practical Significance We noted previously that reporting the results of a hypothesis test in terms of a P-value is very useful because it conveys more information than just the simple statement “reject H0” or “fail to reject H0”. That is, rejection of H0 at the 0.05 level of significance is much more meaningful if the value of the test statistic is well into the critical region, greatly exceeding the 5% critical value, than if it barely exceeds that value. Even a very small P-value can be difficult to interpret from a practical viewpoint when we are making decisions because, while a small P-value indicates statistical significance in the sense that H0 should be rejected in favor of H1, the actual departure from H0 that has been detected may have little (if any) practical significance (engineers like to say “engineering significance”). This is particularly true when the sample size n is large. For example, consider the propellant burning rate problem of Example 9-3 where we are testing H0: 50 centimeters per second versus H1: 50 centimeters per second with 2. If we suppose that the mean rate is really 50.5 centimeters per second, this is not a serious departure from H0: 50 centimeters per second in the sense that if the mean really is 50.5 centimeters per second there is no practical observable effect on the performance of the aircrew escape system. In other words, concluding that 50 centimeters per second when it is really 50.5 centimeters per second is an inexpensive error and has no practical significance. For a reasonably large sample size, a true value of 50.5 will lead to a sample x that c09.qxd 5/15/02 8:02 PM Page 299 RK UL 9 RK UL 9:Desktop Folder: 9-2 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE KNOWN 299 is close to 50.5 centimeters per second, and we would not want this value of x from the sample to result in rejection of H0. The following display shows the P-value for testing H0: 50 when we observe x 50.5 centimeters per second and the power of the test at 0.05 when the true mean is 50.5 for various sample sizes n: Sample Size n P-value When x 50.5 Power (at 0.05) When True 50.5 10 25 50 100 400 1000 0.4295 0.2113 0.0767 0.0124 5.73 10 7 2.57 10 15 0.1241 0.2396 0.4239 0.7054 0.9988 1.0000 The P-value column in this display indicates that for large sample sizes the observed sample value of x 50.5 would strongly suggest that H0: 50 should be rejected, even though the observed sample results imply that from a practical viewpoint the true mean does not differ much at all from the hypothesized value 0 50. The power column indicates that if we test a hypothesis at a fixed significance level and even if there is little practical difference between the true mean and the hypothesized value, a large sample size will almost always lead to rejection of H0. The moral of this demonstration is clear: Be careful when interpreting the results from hypothesis testing when the sample size is large, because any small departure from the hypothesized value 0 will probably be detected, even when the difference is of little or no practical significance. EXERCISES FOR SECTION 9-2 9-20. 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) Should the water temperature be judged acceptable with 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? 9-21. Reconsider the chemical process yield data from Exercise 8-9. Recall that 3, yield is normally distributed and that n 5 observations on yield are 91.6%, 88.75%, 90.8%, 89.95%, and 91.3%. Use 0.05. (a) Is there evidence that the mean yield is not 90%? (b) What is the P-value for this test? (c) What sample size would be required to detect a true mean yield of 85% with probability 0.95? (d) What is the type II error probability if the true mean yield is 92%? (e) Compare the decision you made in part (c) with the 95% CI on mean yield that you constructed in Exercise 8-7. 9-22. 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 H0: 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? 9-23. 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 melting point is normally distributed with 1.5 F. (a) Test H0: 155 versus H0: 155 using 0.01. (b) What is the P-value for this test? c09.qxd 5/15/02 8:02 PM Page 300 RK UL 9 RK UL 9:Desktop Folder: 300 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE (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. 9-24. 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. 9-25. An engineer who is studying the tensile strength of a steel alloy intended for use in golf club shafts knows that tensile strength is approximately normally distributed with 60 psi. A random sample of 12 specimens has a mean tensile strength of x 3250 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) Explain how you could answer the question in part (a) with a two-sided confidence interval on mean tensile strength. 9-26. Suppose that in Exercise 9-25 we wanted to reject the null hypothesis with probability at least 0.8 if mean strength 3500. What sample size should be used? 9-27. 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 hypotheses H0: 100 versus H1: 100 using 0.05. (b) Compute the power of the test if the true mean speed is as low as 95 meters per second. (c) 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? (d) Explain how the question in part (a) could be answered by constructing a one-sided confidence bound on the mean speed. 9-28. A bearing used in an automotive application is suppose 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 hypotheses H0: 1.5 versus H1: 1.5 using 0.01. (b) Compute the power of the test if the true mean diameter is 1.495 inches. (c) 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? (d) Explain how the question in part (a) could be answered by constructing a two-sided confidence interval on the mean diameter. 9-29. Medical researchers have developed a new artificial heart constructed primarily of titanium and plastic. The heart 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) Compute the power of the test if the true mean battery life is 4.5 hours. (c) 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? (d) 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 8-3, 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 c09.qxd 5/15/02 8:02 PM Page 301 RK UL 9 RK UL 9:Desktop Folder: 301 9-3 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE UNKNOWN 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 8-3 to devise the t-confidence interval for . Now consider testing the hypotheses H0: 0 H1: 0 We will use the test statistic T0 X 0 S 1n (9-23) 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 locate the critical region to control the type I error probability at the desired level. In this case we would use the t percentage points t2,n 1 and t2,n 1 as the boundaries of the critical region so that we would reject H0: 0 if t0 t2,n 1 or if t0 t2,n 1 where t0 is the observed value of the test statistic T0. The test procedure is very similar to the test on the mean with known variance described in Section 9-2, except that T0 is used as the test statistic instead of Z0 and the tn 1 distribution is used to define the critical region instead of the standard normal distribution. A summary of the test procedures for both two- and onesided alternative hypotheses follows: The OneSample t-Test Null hypothesis: Test statistic: H0: 0 T0 X 0 S 1n Alternative hypothesis Rejection criteria H1: Z 0 H1: 0 H1: 0 t0 t/2,n 1 or t0 t/2,n 1 t0 t,n 1 t0 t,n 1 Figure 9-8 shows the location of the critical region for these situations. 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 9-8 The reference distribution for H0: 0 with critical region for (a) H1: Z 0, (b) H1: 0, and (c) H1: 0. T0 c09.qxd 5/15/02 8:02 PM Page 302 RK UL 9 RK UL 9:Desktop Folder: 302 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE EXAMPLE 9-6 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 “spring-like 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. 9-9 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 eight-step procedure for hypothesis testing is as follows: 1. 2. 3. 4. 5. The parameter of interest is the mean coefficient of restitution, . H0: 0.82 H1: 0.82 . We want to reject H0 if the mean coefficient of restitution exceeds 0.82. 0.05 The test statistic is t0 6. x 0 s 1n Reject H0 if t0 t0.05,14 1.761 99 95 Percentage 90 80 70 60 50 40 30 20 10 Figure 9-9. Normal probability plot of the coefficient of restitution data from Example 9-6. 5 1 0.78 0.83 Coefficient of restitution 0.88 c09.qxd 5/16/02 4:15 PM Page 303 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark Files: 9-3 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE UNKNOWN 7. Computations: Since x 0.83725, s 0.02456, 0 0.82, and n 15, we have t0 8. 303 0.83725 0.82 2.72 0.02456 115 Conclusions: Since t0 2.72 1.761, we reject H0 and conclude at the 0.05 level of significance 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: 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 T 2.72 SE Mean 0.00634 P 0.008 Notice that Minitab computes both the test statistic T0 and a 95% lower confidence bound for the coefficient of restitution. 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. Minitab also calculates a P-value for the test statistic T0. In the next section we explain how this is done. 9-3.2 P-Value for a t-Test The P-value for a t-test is just the smallest level of significance at which the null hypothesis would be rejected. That is, it is the tail area beyond the value of the test statistic t0 for a onesided test or twice this area for a two-sided test. Because the t-table in Appendix Table IV contains only 10 critical values for each t distribution, computation of the exact P-value directly from the table is usually impossible. However, it is easy to find upper and lower bounds on the P-value from this table. To illustrate, consider the t-test based on 14 degrees of freedom in Example 9-6. The relevant critical values from Appendix Table IV 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 Notice that t0 2.72 in Example 9-6, and that this is between two tabulated values, 2.624 and 2.977. Therefore, the P-value must be between 0.01 and 0.005. These are effectively the upper and lower bounds on the P-value. Example 9-6 is an upper-tailed test. If the test is lower-tailed, just change the sign of t0 and proceed as above. Remember that for a 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 2.72 for a two-tailed alternate based on 14 degrees of freedom. The value t0 2.624 (corresponding to 0.02) and t0 2.977 (corresponding to 0.01), so the lower and upper bounds on the P-value would be 0.01 P 0.02 for this case. c09.qxd 5/15/02 8:02 PM Page 304 RK UL 9 RK UL 9:Desktop Folder: 304 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE Finally, most computer programs report P-values along with the computed value of the test statistic. Some hand-held calculators also have this capability. In Example 9-6, Minitab gave the P-value for the value t0 2.72 in Example 9-6 as 0.008. 9-3.3 Choice of Sample Size The type II error probability for tests on the mean of a normal distribution with unknown variance depends on the distribution of the test statistic in Equation 9-23 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 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 two-sided 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 VIe, VIf, VIg, and VIh that plot for the t-test against a parameter d for various sample sizes n. Curves are provided for two-sided alternatives on Charts VIe and VIf. The abscissa scale factor d on these charts is defined as d 0 0 0 00 (9-24) For the one-sided alternative 0 or 0 , we use charts VIg and VIh with d 00 0 0 0 (9-25) 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 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 0.020.02456 0.81. By referring to the operating characteristic curves in Appendix Chart VIg (for 0.05) with d 0.81 and n 15, we find that 0.10, c09.qxd 5/15/02 8:02 PM Page 305 RK UL 9 RK UL 9:Desktop Folder: 9-3 TESTS ON THE MEAN OF A NORMAL DISTRIBUTION, VARIANCE UNKNOWN 305 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: 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 Sample Difference Size Power 0.02 15 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 Sample Difference Size Power 0.01 15 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 Sample Target Actual Difference Size Power Power 0.01 39 0.8000 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 asked Minitab to compute the power if the difference in and 0 0.82 we wanted to detect was 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 to give 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. 9-3.4 Likelihood Ratio Approach to Development of Test Procedures (CD Only) EXERCISES FOR SECTION 9-3 9-30. 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 temperature (°C) reported was as follows: 23.01, 22.22, 22.04, 22.62, and 22.59. c09.qxd 4/6/02 2:55 PM Page 306 306 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE (a) Test the hypotheses H0: 22.5 versus H1: 22.5, using 0.05. Find the P-value. (b) Is there evidence to support 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? (e) Explain how the question in part (a) could be answered by constructing a two-sided confidence interval on the mean interior temperature. 9-31. 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 Wundrlich”) 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 hypotheses H0: 98.6 versus H1: 98.6, using 0.05. Find the P-value. (b) Compute the power of the test if the true mean female body temperature is as low as 98.0. (c) 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? (d) Explain how the question in part (a) could be answered by constructing a two-sided confidence interval on the mean female body temperature. (e) Is there evidence to support the assumption that female body temperature is normally distributed? 9-32. 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 by Simpson, Alsen, and Eden, “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. (b) Is there evidence 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. 9-33. The sodium content of thirty 300-gram boxes of organic corn flakes 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, 131.15, 130.69, 130.91, 129.54, 129.64, 128.77, 130.72, 128.33, 128.24, and 129.65. (a) Can you support a claim that mean sodium content of this brand of cornflakes is 130 milligrams? Use 0.05. (b) Is there evidence that sodium content is normally distributed? (c) Compute the power of the test if the true mean sodium content is 130.5 miligrams. (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. 9-34. Reconsider the tire testing experiment described in Exercise 8-22. (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? Use the sample standard deviation s as an estimate of in reaching your decision. 9-35. Reconsider the Izod impact test on PVC pipe described in Exercise 8-23. Suppose that you want to use the data from this experiment to support a claim that the mean impact strength exceeds the ASTM standard (foot-pounds per inch). Formulate and test the appropriate hypotheses using 0.05. 9-36. Reconsider the television tube brightness experiment in Exercise 8-24. Suppose that the design engineer believes that this tube will require 300 microamps of current to produce the desired brightness level. Formulate and test an appropriate hypothesis using 0.05. Find the P-value for this test. State any necessary assumptions about the underlying distribution of the data. 9-37. Consider the baseball coefficient of restitution data first presented in Exercise 8-79. (a) Does the data support the claim that the mean coefficient of restitution of baseballs exceeds 0.635? Use 0.05. (b) What is the P-value of the test statistic computed in part (a)? (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? 9-38. Consider the dissolved oxygen concentration at TVA dams first presented in Exercise 8-81. (a) Test the hypotheses H0: 4 versus H1: 4 . Use 0.01. c09.qxd 5/15/02 8:02 PM Page 307 RK UL 9 RK UL 9:Desktop Folder: 307 9-4 HYPOTHESIS TESTS ON THE VARIANCE AND STANDARD DEVIATION OF A NORMAL POPULATION (b) What is the P-value of the test statistic computed in part (a)? (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? 9-39. Consider the cigar tar content data first presented in Exercise 8-82. (a) Can you support a claim that mean tar content exceeds 1.5? Use 0.05 (b) What is the P-value of the test statistic computed in part (a)? (c) Compute the power of the test if the true mean tar content is 1.6. (d) What sample size would be required to detect a true mean tar content of 1.6 if we wanted the power of the test to be at least 0.8? 9-40. Exercise 6-22 gave data on the heights of female engineering students at ASU. (a) Can you support a claim that mean height of female engineering students at ASU is 65 inches? Use 0.05 (b) What is the P-value of the test statistic computed in part (a)? (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? 9-41. Exercise 6-24 presented data on the concentration of suspended solids in lake water. (a) Test the hypotheses H0: 55 versus H1: 55, use 0.05. (b) What is the P-value of the test statistic computed in part (a)? (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? 9-42. Exercise 6-25 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. (b) What is the P-value of the test statistic computed in part (a)? (c) Compute the power of the test if the true mean distance is 290 yards. (e) 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? 9-4 HYPOTHESIS TESTS ON THE VARIANCE AND STANDARD DEVIATION OF A NORMAL POPULATION 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 The Hypothesis Testing Procedures 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 (9-26) we will use the test statistic: X 20 1n 12S 2 02 (9-27) If the null hypothesis H0: 2 02 is true, the test statistic X 20 defined in Equation 9-27 follows the chi-square distribution with n 1 degrees of freedom. This is the reference c09.qxd 5/16/02 4:15 PM Page 308 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark Files: 308 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 x α α 2α , n – 1 0 (a) x 0 21 – α , n – 1 x (b) Figure 9-10 Reference distribution for the test of H0: (b) H1: 2 20 , and (c) H1: 2 20 . 2 20 (c) with critical region values for (a) H1: 20 , 2 distribution for this test procedure. Therefore, we calculate 20 , the value of the test statistic X 02, and the null hypothesis H0: 2 20 would be rejected if 20 22, n1 or if 20 212,n1 where 22,n1 and 212,n1 are the upper and lower 1002 percentage points of the chisquare distribution with n 1 degrees of freedom, respectively. Figure 9-10(a) shows the critical region. The same test statistic is used for one-sided alternative hypotheses. For the one-sided hypothesis H0: 2 20 H1: 2 20 (9-28) we would reject H0 if 20 2,n1, whereas for the other one-sided hypothesis H0: 2 20 H1: 2 20 (9-29) we would reject H0 if 20 21,n1. The one-sided critical regions are shown in Figure 9-10(b) and (c). EXAMPLE 9-8 An automatic 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 eight-step procedure results in the following: 1. 2. 3. 4. 5. The parameter of interest is the population variance 2. H0: 2 0.01 H1: 2 0.01 0.05 The test statistic is 20 1n 12s2 20 c09.qxd 5/15/02 8:02 PM Page 309 RK UL 9 RK UL 9:Desktop Folder: 10-4 HYPOTHESIS TESTS ON THE VARIANCE AND STANDARD DEVIATION OF A NORMAL POPULATION 309 6. Reject H0 if 20 20.05,19 30.14. 7. Computations: 20 8. 1910.01532 29.07 0.01 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. 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 0.10. The actual P-value is P 0.0649. (This value was obtained from a calculator.) 9-4.2 -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 9-26, Charts VIi and VIj plot against an abscissa parameter 0 (9-30) 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 VIm and VIn 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 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 VIk. 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. c09.qxd 5/15/02 8:02 PM Page 310 RK UL 9 RK UL 9:Desktop Folder: 310 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE EXERCISES FOR SECTION 9-4 9-43. Consider the rivet holes from Exercise 8-35. If the standard deviation of hole diameter exceeds 0.01 millimeters, there is an unacceptably high probability that the rivet will not fit. Recall that n 15 and s 0.008 millimeters. (a) Is there strong evidence to indicate that the standard deviation of hole diameter exceeds 0.01 millimeters? Use 0.01. State any necessary assumptions about the underlying distribution of the data. (b) Find the P-value for this test. (c) If is really as large as 0.0125 millimeters, what sample size will be required to defect this with power of at least 0.8? 9-44. Recall the sugar content of the syrup in canned peaches from Exercise 8-36. Suppose that the variance is thought to be 2 18 (milligrams)2. 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. (b) What is the P-value for this test? (c) Discuss how part (a) could be answered by constructing a 95% two-sided confidence interval for . 9-45. Consider the tire life data in Exercise 8-22. (a) Can you conclude, using 0.05, that the standard deviation of tire life exceeds 200 kilometers? State any necessary assumptions about the underlying distribution of the data. (b) Find the P-value for this test. 9-5 9-46. Consider the Izod impact test data in Exercise 8-23. (a) Test the hypothesis that 0.10 against an alternative specifying that 0.10, using 0.01, and draw a 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? 9-47. Reconsider the percentage of titanium in an alloy used in aerospace castings from Exercise 8-39. 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. (b) Explain how you could answer the question in part (a) by constructing a 95% two-sided confidence interval for . 9-48. Consider the hole diameter data in Exercise 8-35. 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 Exercise 9-43? 9-49. Consider the sugar content in Exercise 9-44. 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? 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 P̂ 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 P̂ is approximately normal with mean p and variance p(1 p)n, if 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 large-sample 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 include hypothesis testing about p. We will consider testing H0: p p0 H1: p p0 (9-31) c09.qxd 5/16/02 4:15 PM Page 311 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark Files: 9-5 TESTS ON A POPULATION PROPORTION 311 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 Z0 X np0 1np0 11 p0 2 (9-32) and reject H0: p p0 if z0 z 2 or z0 z 2 Note that the standard normal distribution is the reference distribution for this test statistic. Critical regions for the one-sided alternative hypotheses would be constructed in the usual manner. EXAMPLE 9-10 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 eight-step hypothesis-testing procedure as follows: 1. The parameter of interest is the process fraction defective p. 2. H0: p 0.05 3. H1: p 0.05 This formulation of the problem will allow the manufacturer to make a strong claim about process capability if the null hypothesis H0: p 0.05 is rejected. 4. 0.05 5. The test statistic is (from Equation 9-32) z0 x np0 1np0 11 p0 2 where x 4, n 200, and p0 0.05. 6. Reject H0: p 0.05 if z0 z0.05 1.645 7. Computations: The test statistic is z0 8. 4 20010.052 1.95 120010.05210.952 Conclusions: Since z0 1.95 z0.05 1.645, we reject H0 and conclude that the process fraction defective p is less than 0.05. The P-value for this value of the test statistic z0 is P 0.0256, which is less than 0.05. We conclude that the process is capable. c09.qxd 5/16/02 4:15 PM Page 312 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark Files: 312 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE Another form of the test statistic Z0 in Equation 9-32 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 P̂ Xn is the sample proportion that belongs to that class. Now divide both numerator and denominator of Z0 in Equation 9-32 by n, giving Z0 Xn p0 1p0 11 p0 2 n Z0 P̂ p0 1p0 11 p0 2 n or (9-33) 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. Statistical software packages usually provide the one sample Z-test for a proportion. The Minitab output for Example 9-10 follows. Test and CI for One Proportion Test of p 0.05 vs p 0.05 Sample X N Sample p 95.0% 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. Notice that both the test statistic (and accompanying P-value) and the 95% one-sided upper confidence bound are displayed. The 95% upper confidence bound is 0.036283, which is less than 0.05. This is consistent with rejection of the null hypothesis Ho: p 0.05. 9-5.2 Small-Sample Tests on a Proportion (CD Only) 9-5.3 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 1p11 p2 n b a p0 p z 2 1p0 11 p0 2 n 1p11 p2 n b (9-34) If the alternative is H1: p p0, 1 a p0 p z 1p0 11 p0 2 n b 1p11 p2 n (9-35) c09.qxd 5/15/02 8:02 PM Page 313 RK UL 9 RK UL 9:Desktop Folder: 313 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 (9-36) 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 n c z2 1p0 11 p0 2 z 1p11 p2 2 d p p0 (9-37) for the two-sided alternative and n c z 1p0 11 p0 2 z 1p11 p2 2 d p p0 (9-38) for a one-sided alternative. EXAMPLE 9-11 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 9-35 as follows: 1 c 0.05 0.03 11.6452 10.0510.952 200 d 1 1 0.442 0.67 10.0311 0.032 200 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. 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 9-38 as follows: n c 1.64510.0510.952 1.2810.0310.972 2 d 0.03 0.05 832 where we have used p 0.03 in Equation 9-38. 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. c09.qxd 5/15/02 8:02 PM Page 314 RK UL 9 RK UL 9:Desktop Folder: 314 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE 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. 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 9-35 in Example 9-11. The second part of the output computes the sample size necessary to give a power of 0.9 ( 0.1) if p 0.03. Again, the results agree closely with those obtained from Equation 9-38. 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 9-50. 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? State and test the appropriate hypotheses using 0.05. 9-51. Continuation of Exercise 9-50. If it is really the situation that p 0.15, how likely is it that the test procedure in Exercise 9-50 will not reject the null hypothesis? 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? 9-52. Reconsider the integrated circuits described in Exercise 8-48. (a) Use the data to test H0: p 0.05 versus H1: p 0.05. Use 0.05. (b) Find the P-value for the test. c09.qxd 5/16/02 4:15 PM Page 315 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark Files: 9-7 TESTING FOR GOODNESS OF FIT 9-53. Consider the defective circuit data in Exercise 8-48. (a) Do the data support the claim that the fraction of defective units produced is less than 0.05, using 0.05? (b) Find the P-value for the test. 9-54. 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. (b) Find the P-value for this test. (c) Discuss how you could have answered the question in part (a) by constructing a two-sided confidence interval on p. 9-55. A manufacturer of interocular lenses is qualifying a new grinding machine and will qualify the machine if 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. (b) Find the P-value for the test in part (a). 315 9-56. 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. (a) Does this finding support the researcher’s claim? Use 0.01. (b) Find the P-value for this test. 9-57. 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? 9-58. The advertized 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 advertized time period, with proper charging procedures? Use a hypothesis-testing procedure with 0.01. 9-6 SUMMARY TABLE OF INFERENCE PROCEDURES FOR A SINGLE SAMPLE The table in the end papers of this book (inside front 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 for rejecting H0, and the formulas for constructing the 100(1 )% two-sided confidence interval. 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. c09.qxd 5/15/02 8:02 PM Page 316 RK UL 9 RK UL 9:Desktop Folder: 316 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE 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 k X 02 a i1 1Oi Ei 2 2 Ei (9-39) 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 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. EXAMPLE 9-12 A 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, (32 0 15 1 9 2 4 3)60 0.75. From the Poisson c09.qxd 5/15/02 8:02 PM Page 317 RK UL 9 RK UL 9:Desktop Folder: 9-7 TESTING FOR GOODNESS OF FIT 317 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: p1 P1X 02 p2 P1X 12 e 0.75 10.752 0 0.472 0! e 0.75 10.752 1 0.354 1! e 0.75 10.752 2 0.133 2! p4 P1X 32 1 1p1 p2 p3 2 0.041 p3 P1X 22 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 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 9-39 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 eight-step hypothesis-testing procedure may now be applied, using 0.05, as follows: 1. 2. The variable of interest is the form of the distribution of defects in printed circuit boards. H0: The form of the distribution of defects is Poisson. 3. 4. 5. H1: The form of the distribution of defects is not Poisson. 0.05 The test statistic is k 20 a i1 1oi Ei 2 2 Ei c09.qxd 5/15/02 8:02 PM Page 318 RK UL 9 RK UL 9:Desktop Folder: 318 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE 6. Reject H0 if 20 20.05,1 3.84. 7. Computations: 20 132 28.322 2 115 21.242 2 113 10.442 2 2.94 28.32 21.24 10.44 2 3.84, we are unable to reject the null hypothesis 8. Conclusions: Since 02 2.94 0.05,1 that the distribution of defects in printed circuit boards is Poisson. The P-value for the test is P 0.0864. (This value was computed using an HP-48 calculator.) EXAMPLE 9-13 A 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 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. A common practice in constructing the class intervals for the frequency distribution used in the chi-square goodness-of-fit 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 pi P1ai 1 X ai 2 ai f 1x2 dx ai 1 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 18 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 Totals 100 100 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 forth. We may apply the eight-step hypothesis-testing procedure to this problem. 1. 2. The variable of interest is the form of the distribution of power supply voltage. H0: The form of the distribution is normal. c09.qxd 5/15/02 8:02 PM Page 319 RK UL 9 RK UL 9:Desktop Folder: 319 9-7 TESTING FOR GOODNESS OF FIT 3. H1: The form of the distribution is nonnormal. 4. 0.05 5. The test statistic is k 20 a i1 6. 7. 1oi Ei 2 2 Ei 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. Therefore, we will reject H0 if 20 20.05,5 11.07. Computations: 1oi Ei 2 2 Ei i1 114 12.52 2 114 12.52 2 112 12.52 2 p 12.5 12.5 12.5 0.64 8 20 a 8. 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 9-59. Consider the following frequency table of observations on the random variable X. defined as the number of calls during that one-hour period. The relative frequency of calls was recorded and reported as Values Observed Frequency Value Relative Frequency Value Relative 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. 9-60. Let X denote the number of flaws observed on a large coil of galvanized steel. Seventy-five coils are inspected and the following data were observed for the values of X: Values Observed Frequency 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 this data? Use 0.01. (b) Calculate the P-value for this test. 9-61. The number of calls arriving at a switchboard from noon to 1 PM during the business days Monday through Friday is monitored for six weeks (i.e., 30 days). Let X be 5 6 8 9 10 0.133 14 0.200 15 0.067 11 0.067 12 0.100 13 0.133 0.133 0.067 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. 9-62. 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. c09.qxd 5/16/02 4:15 PM Page 320 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark Files: 320 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE 9-63. Define X as the number of underfilled bottles from a filling operation in a carton of 24 bottles. Sixty 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. 9-64. 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. 9-8 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 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. 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, 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 c09.qxd 5/16/02 4:15 PM Page 321 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark Files: 9-8 CONTINGENCY TABLE TESTS 321 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 (9-40) Therefore, the expected frequency of each cell is r 1 c Eij nûiv̂j n a Oij a Oij j1 i1 (9-41) Then, for large n, the statistic r 20 c a a i1 j1 1Oij Eij 2 2 (9-42) Eij has an approximate chi-square distribution with (r 1)(c 1) degrees of freedom if the null hypothesis is true. Therefore, we would reject the hypothesis of independence if the observed value of the test statistic 20 exceeded 2,(r1)(c1). EXAMPLE 9-14 A company has to choose among three pension 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. To find the expected frequencies, we must first compute û1 (340500) 0.68, û2 (160500) 0.32, v̂1 (200500) 0.40, v̂2 (200500) 0.40, and v̂3 (100500) 0.20. The expected frequencies may now be computed from Equation 9-41. For example, the expected number of salaried workers favoring pension plan 1 is E11 nû1v̂1 50010.68210.402 136 The expected frequencies are shown in Table 9-4. The eight-step hypothesis-testing procedure may now be applied to this problem. 1. 2. Table 9-3 The variable of interest is employee preference among pension plans. H0: Preference is independent of salaried versus hourly job classification. Table 9-4 Observed Data for Example 9-14 Expected Frequencies for Example 9-14 Pension Plan Pension Plan 1 2 3 Totals 1 2 3 Totals Salaried workers Hourly workers 160 40 140 60 40 60 340 160 Salaried workers Hourly workers 136 64 136 64 68 32 340 160 Totals 200 200 100 500 Totals 200 200 100 500 Job Classification Job Classification c09.qxd 5/15/02 8:02 PM Page 322 RK UL 9 RK UL 9:Desktop Folder: 322 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE 3. H1: Preference is not independent of salaried versus hourly job classification. 4. 0.05 5. The test statistic is r 20 c a a 1oij Eij 2 2 Eij i1 j1 6. 7. Since r 2 and c 3, the degrees of freedom for chi-square are (r 1)(c 1) 2 5.99. (1)(2) 2, and we would reject H0 if 02 0.05,2 Computations: 2 3 20 a a i1 j1 8. 1oij Eij 2 2 Eij 1160 1362 2 1140 1362 2 140 682 2 140 642 2 136 136 68 64 160 642 2 160 322 2 49.63 64 32 Conclusions: Since 20 49.63 20.05,2 5.99, we reject the hypothesis of independence and conclude that the preference for pension plans is not independent of job classification. The P-value for 20 49.63 is P 1.671 10 11. (This value was computed using a hand-held calculator.) 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. EXERCISES FOR SECTION 9-8 9-65. A company operates four machines 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 c09.qxd 5/15/02 8:02 PM Page 323 RK UL 9 RK UL 9:Desktop Folder: 9-8 CONTINGENCY TABLE TESTS Test the hypothesis (using 0.05) that breakdowns are independent of the shift. Find the P-value for this test. 9-66. 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: Surgical Medical Yes No 46 36 52 43 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. 9-70. A random sample of students is asked their opinions on a proposed core curriculum change. The results are as follows. Patient Category Medicare 323 Opinion 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. 9-67. Grades in a statistics course and an operations research course taken simultaneously were as follows for a group of students. 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? 9-68. 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? Lateral Deflection 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 9-69. 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: Class Freshman Sophomore Junior Senior Favoring Opposing 120 70 60 40 80 130 70 60 Test the hypothesis that opinion on the change is independent of class standing. Use 0.05. What is the P-value for this test? Supplemental Exercises 9-71. 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 p̂i X1 X2 p Xn n What are the sampling distribution, the sample mean, and sample variance estimates of p̂ 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. 9-72. Consider the situation of Exercise 9-76. 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, p̂ 0.095 (b) n 100, p̂ 0.095 (c) n 500, p̂ 0.095 (d) n 1000, p̂ 0.095 (e) Comment on the effect of sample size on the observed P-value of the test. c09.qxd 8/6/02 2:16 PM Page 324 324 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE 9-73. 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. 9-74. Suppose that in Exercise 9-73, 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 9-73, does the probability of a type II error increase or decrease with the increase in standard deviation? Explain your answer. 9-75. 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. 9-76. 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. 9-77. 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. 9-78. 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. (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. 9-79. Consider the television picture tube brightness experiment described in Exercise 8-24. (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 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). 9-80. Consider the fatty acid measurements for the diet margarine described in Exercise 8-25. (a) For the sample size n 6, using a two-sided alternative hypothesis and 0.01, test H0: 2 1.0. (b) Suppose 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). 9-81. 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. c09.qxd 5/15/02 8:02 PM Page 325 RK UL 9 RK UL 9:Desktop Folder: 9-8 CONTINGENCY TABLE TESTS (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. 9-82. 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 9-83. Consider the 40 observations collected on the number of nonconforming coil springs in production batches of size 50 given in Exercise 6-79. (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. 9-84. Consider the 20 observations collected on the number of errors in a string of 1000 bits of a communication channel given in Exercise 6-80. (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. 9-85. Consider the spot weld shear strength data in Exercise 6-23. Does the normal distribution seem to be a reasonable model for these data? Perform an appropriate goodness-of-fit test to answer this question. 9-86. Consider the water quality data in Exercise 6-24. (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 goodnessof-fit test to answer this question. 9-87. Consider the golf ball overall distance data in Exercise 6-25. (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)? 325 (c) Do these data appear to be well modeled by a normal distribution? Use a formal goodness-of-fit test in answering this question. 9-88. Consider the baseball coefficient of restitution data in Exercise 8-79. 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 using the data in Exercise 8-79 and draw conclusions, using 0.01. (c) Find the P-value for this test. (d) In Exercise 8-79(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. 9-89. Consider the dissolved oxygen data in Exercise 8-81. 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, using 0.05, and the data from Exercise 8-81. (c) Find the P-value for this test. (d) In Exercise 8-81(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. 9-90. 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? c09.qxd 8/6/02 2:18 PM Page 326 326 CHAPTER 9 TESTS OF HYPOTHESES FOR A SINGLE SAMPLE MIND-EXPANDING EXERCISES 9-91. 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. 9-92. Derive an expression for for the test on the variance of a normal distribution. Assume that the twosided alternative is specified. 9-93. 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 1n 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 9-2.5, which would use S 1n in the denominator, because it is designed just for the 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? 9-94. 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 0 220 12n2 Use this result to test H0: 10 versus H1: 10 for the golf ball overall distance data in Exercise 6-25. 9-95. Continuation of Exercise 9-94. Using the results of the previous exercise, 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 . How would you estimate the standard error? 9-96. Continuation of Exercises 9-94 and 9-95. Consider the golf ball overall distance data in Exercise 6-25. We wish to investigate a claim that 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 6-25. What are your conclusions? 9-97. Let X1, X2, p , Xn be a sample from an exponential distribution with parameter . It can be shown that n 2 i1 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 In the E-book, click on any term or concept below to go to that subject. Connection between hypothesis tests and confidence intervals Critical region for a test statistic Null hypothesis One- and two-sided alternative hypotheses Operating characteristic curves Power of the test P-value Reference distribution for a test statistic Sample size determination for hypothesis tests Significance level of a test Statistical hypotheses Statistical versus practical significance Test for goodness of fit Test for homogeneity Test for independence Test statistic Type I and type II errors CD MATERIAL Likelihood ratio test PQ220 6234F.CD(09) 5/15/02 8:21 PM Page 1 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark F 9-3.4 Likelihood Ratio Approach to Development of Test Procedures (CD Only) Hypothesis testing is one of the most important techniques of statistical inference. Throughout this book we present many applications of hypothesis testing. While we have emphasized a heuristic development, many of these hypothesis-testing procedures can be developed using a general principle called the likelihood ratio principle. Tests developed by this method often turn out to be “best” test procedures in the sense that they minimize the type II error probability among all tests that have the same type I error probability . The likelihood ratio principle is easy to illustrate. Suppose that the random variable X has a probability distribution that is described by an unknown parameter , say, f (x, ). We wish to test the hypothesis H0: is in 0 versus H1: is in 1, where 0 and 1 are disjoint sets of values (such as H0: 0 versus H1: 0). Let X1, X2, p , Xn be the observations in a random sample. The joint distribution of these sample observations is f 1x1, x2, p , xn, 2 f 1x1, 2 f 1x2, 2 p f 1xn, 2 Recall from our discussion of maximum likelihood estimation in Chapter 7 that the likelihood function, say L(), is just this joint distribution considered as a function of the parameter . The likelihood ratio principle for test construction consists of the following steps: 1. Find the largest value of the likelihood for any in 0. This is done by finding the maximum likelihood estimator of restricted to values within 0 and by substituting this value of back into the likelihood function. This results in a value of the likelihood function that we will call L(0). 2. Find the largest value of the likelihood for any in 1. Call this the value of the likelihood function L(1). 3. Form the ratio L10 2 L11 2 This ratio is called the likelihood ratio test statistic. The test procedure calls for rejecting the null hypothesis H0 when the value of this ratio is small, say, whenever k, where k is a constant. Thus, the likelihood ratio principle requires rejecting H0 when L(1) is much larger than L(0), which would indicate that the sample data are more compatible with the alternative hypothesis H1 than with the null hypothesis H0. Usually, the constant k would be selected to give a specified value for , the type I error probability. These ideas can be illustrated by a hypothesis-testing problem that we have studied before—that of testing whether the mean of a normal population has a specified value 0. This is the one-sample t-test of Section 9-3. Suppose that we have a sample of n observations from a normal population with unknown mean and unknown variance 2, say, X1, X2, p , Xn. We wish to test the hypothesis H0: 0 versus H1: 0. The likelihood function of the sample is La n n 1 2 2 b eg i1 1xi 2 12 2 12 9-1 PQ220 6234F.CD(09) 5/15/02 8:21 PM Page 2 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark F 9-2 and the values of 0 and 1 are 0 0 and 1 {: }, respectively. The values of and 2 that maximize L in 1 are the usual maximum likelihood estimates for and 2: 1 n ˆ a xi x n i1 1 n ˆ 2 n a 1xi x2 2 i1 Substituting these values in L, we have L11 2 c n2 1 d e 1n22 12 n2 g 1xi x2 2 To maximize L in 0 we simply set 0 and then find the value of 2 that maximizes L. This value is found to be 1 n ˆ 2 n a 1xi 0 2 2 i1 which gives L10 2 c n2 1 e 1n22 2d 12n2 g 1xi 0 2 The likelihood ratio is L10 2 g 1xi x2 2 n2 c d L11 2 g 1xi 0 2 2 Now since n n i1 i1 2 2 2 a 1xi 0 2 a 1xi x2 n1x 0 2 we may write as n2 1 2 d 1 n1x 0 2 t g 1xi x2 2 n2 1 n 1x 0 2 2 n 1 t d b 1 c da g 1xi x2 2 n 1 1 c 1a n2 1x 0 2 2 s 1 bc d n1 s2 n PQ220 6234F.CD(09) 5/15/02 8:21 PM Page 3 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark F 9-3 1x 0 2 2 d t 2 is the square of the value of a random variable that has the s 2n t-distribution with n 1 degrees of freedom when the null hypothesis H0 : 0 is true. So we may write the value of the likelihood ratio as Notice that c n 2 1 c s 1 3t 2 1n 12 4 It is easy to find the value for the constant k that would lead to rejection of the null hypothesis H0. Since we reject H0 if k, this implies that small values of support the alternative hypothesis. Clearly, will be small when t 2 is large. So instead of specifying k we can specify a constant c and reject H0: 0 if t 2 c. The critical values of t would be the extreme values, either positive or negative, and if we wish to control the type I error probability at , the critical region in terms of t would be t t 2,n1 and t t 2,n1 or, equivalently, we would reject H0: 0 if t 2 c t2 2,n1. Therefore, the likelihood ratio test for H0: 0 versus H1: 0 is the familiar single-sample t-test. The procedure employed in this example to find the critical region for the likelihood ratio is used often. That is, typically, we can manipulate to produce a condition that is equivalent to k, but one that is simpler to use. The likelihood ratio principle is a very general procedure. Most of the tests presented in this book that utilize the t, chi-square, and F-distributions for testing means and variances of normal distributions are likelihood ratio tests. The principle can also be used in cases where the observations are dependent, or even in cases where their distributions are different. However, the likelihood function can be very complicated in some of these situations. To use the likelihood principle we must specify the form of the distribution. Without such a specification, it is impossible to write the likelihood function, and so if we are unwilling to assume a particular probability distribution, the likelihood ratio principle cannot be used. This could lead to the use of the nonparametric test procedures discussed in Chapter 15. 9-5.2 Small-Sample Tests on a Proportion (CD Only) 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 H1: p p0 Let X be the number of successes in the sample. A lower-tail rejection region would be used. That is, we would reject H0 if x c, where c is the critical value. When H0 is true, X has a binomial distribution with parameters n and p0; therefore, P 1Type I error2 P 1reject H0 when H0 is true2 P 3X c when X is Bin 1n, p0 2 4 B 1c; n, p0 2 PQ220 6234F.CD(09) 5/15/02 8:21 PM Page 4 RK UL 6 RK UL 6:Desktop Folder:TEMP WORK:MONTGOMERY:REVISES UPLO D CH 1 14 FIN L:Quark F 9-4 where B(c; n1, p0) is the cumulative binomial distribution. To find the critical value for a given , we would select the largest c satisfying B(c; n1, p0) . The type II error calculation is straightforward. Let p1 be an alternative value of p, with p1 p0. If p p1, X is Bin (n, p1). Therefore P 1Type II error when p p1 2 P 3X c when X is Bin 1n, p1 2 4 1 B 1c; n, p1 2 where B(c; n, p1) is the cumulative binomial distribution. Test procedures for the other one-sided alternative H1: p p0 and the two-sided alternative H0: p p0 are constructed in a similar fashion. For H1: p p0 the critical region has the form x c, where we would choose the smallest value of c satisfying 1 B(c 1, n, p0) . For the two-sided case, the critical region consists of both large and small values. Because c is an integer, it usually isn’t possible to define the critical region to obtain exactly the desired value of . To illustrate the procedure, let’s reconsider the situation of Example 9-10, where we wish to test H0: p 0.05 verses H1: p 0.05. Suppose now that the sample size is n 100 and we wish to use 0.05. Now from the cumulative binomial distribution with n 50 and p 0.05, we find that B(0; 100, 0.05) 0.0059, B(1; 100, 0.05) 0.0371, and B(2; 100, 0.05) 0.1183 (Minitab will generate these cumulative binomial probabilities). Since B(1; 100, 0.05) 0.0371 0.05 and B(2; 100, 0.05) 0.1183 0.05, we would select c 1. Therefore the null hypothesis will be rejected if x 1. The exact significance level for this test is 0.0371. To calculate the power of the test, suppose that p1 0.03. Now 1 B 1c; n, p1 2 1 B 11; 100, 0.032 1 0.1946 0.8054 and the power of the test is only 0.1946. This is a fairly small power because p1 is close to p0.