CHAPTER OUTLINE 9.1 The Estimation of Variances 290 9.2 Hypotheses Concerning One Variance 293 9.3 Hypotheses Concerning Two Variances 295 Review Exercises 299 Key Terms 300 Unbiased estimation of a population variance INFERENCES CONCERNING VARIANCES population mean, how to construct confidence intervals for means, and how to per form tests of hypotheses about the means of one and o methods apply to inferences about other population parameters In this chapter we shall concentrate on population variances, or standard deviations wiih are not only important in their own right, But which must sometimes be estimated before inferences about other parameters can be made. Section 9. is devoted to the ‘estimation of o and a, and Sections 9.2 and 9.3 deal with tests of hypotheses about | 1h Chapters 7 and 8 we learned how to judge the size of the error in estimating a ‘wo populations. Very similar these parameters. 9.1 The Estimation of Variances In the preceding chapters, there were several instances where we estimated a popu- lation standard deviation by means of a sample standard deviation—we substituted the sample standard deviation S for o in the large sample confidence interval for j con page 230, in the large sample test concerning 1 on page 252, and in the large sample test concerning the difference between two means on page 270, There are ‘many statistical procedures in which S is thus substituted for 6, or $* for a2, Let Yai-*? oi s be the sample variance, based on a random sample from any population, discrete or continuous, having variance 07. It follows from Example 34 of Chapter 5 tha the ‘mean of the sampling distribution of S? is E(S*) = 0? whatever the value if 07 ‘The sample variance Die? is an unbiased estimator of ? Although the sample variance is an unbiased estimator of 0, it does not follow that the sample standard deviation is also an unbiased estimator of «In fact, it isnot However, for large samples the bias is small and it is common practice to estimate fo with sEXAMPLE | Solution See 91 The Estinaon of Varances 29 Besides s, population standard deviations are sometimes estimated in terms of the sample range R, which we defined in Section 2.6 asthe largest value of a sample minus the smallest, Given a random sample of size n from a normal population, it can be shown that the sampling distribution of the range has the mean d; o and the standard deviation 3 ¢, where dz and ds are constants which depend on the size of the sample, Form = 1, 2,..., and 10, their values are as shown in the following table: 7 2 3 4 § 6 7 8 9 dz | 1128 1693 2059 2326 2534 2.704 2847 2970 3.078 ds | 0853. 0888 0.880 0.864 0848 0.833 0820 0.808 0.797 Thus, R/dy is an unbiased estimate of o, and for very small samples, n << 5, it provides nearly as good an estimate of o as does s. As the sample size increases, it becomes more efficient to use s instead of R/dz. Nowadays, the range is used to estimate o primarily in problems of industrial quality control, where sample sizes. are usually small and computational ease is of prime concern. This application will be discussed in Chapter 15, where we shall need the above values of the constant d3 Using the sample range to estimate o ‘With reference to Example 7, Chapter 8, use the range of the first sample to estimate « for the resiliency modulus of recycled materials from the first location. Since the smallest value is 604, the largest value is 707, and n = 6 so that dy 2.534, we get R _ 707-604 E> os = 18 Note that this is moderately close tothe sample standard deviation s = 34.7. In most practical applications, interval estimates of ¢ or 6? are based on the sample standard deviation or the sample variance. For random samples from normal populations, we make use of Theorem 6.5, according to which (n-1)S* is a random variable having the chi square distribution with nm — 1 degrees of freedom As illustrated in Figure 9.1 with 2 as defined on page 208, the two quantities 1 hgy 308 2 C0 off ara a /2 in the left and right al, respectively, We ean then assert that, whatever the value of 0”, n-1)S? 2 (den woe. xan) = Once the data have been obtained, we make the same assertion with (1 ~ a) 100% confidence, Solving each inequality for 22, we obtain the confidence interval292 Chapter erences Concerning Varances Figure 9.1 ‘Two percentiles of the 7? disteibution Confidence interval for o? Confidence interval for ¢ Gzunep Solution fa ee (maDF oo (nas z 2 Kaj? Kina? If we take the square root of each member of this inequality, we obtain a correspond- ing (1 — «)100% confidence interval foro. @ =e _ |@ . 1s? z a Ka jr Minas Note that confidence intervals for ¢ or o? obtained by taking equal tails, as in the above formulae, do not actually give the narrowest confidence intervals, be- cause the chi square distribution is not symmetrical. (Sec Exercise 7,26.) Neverthe- less, they are used in most applications in order to avoid fairly complicated calcula- tions. For moderate degrees of freedom, the choice of equal tails is inconsequential A.95% confidence interval for the standard deviation « of weight Referring to Example 8, Chapter 5, and the n = 80 measurements of the weight of cheese, construct a 95% confidence interval for the population standard deviation o- From Example 8, itis very reasonable to assume thatthe population is normal ‘With 80 — 1 = 79 degrees of freedom, we could interpolate in Table SW but computer calculations first give y3y5 = 56.309 and x3¢°= 105.473. and then 5 = 9.583. Substituting into the formula for the confidence interval for 6? yields (799.583 a _ (791 9.583)? “105475 36300 or 68.8 <0? < 1288 ‘and, taking the square root, 8.29 <6 < 1135Sec92 Hypotheses Concerning One Variance 293 ‘This means we are 95% confident that the interval from 8.29 to 11.35 pounds contains othe population standard deviation of the weight of cheese. 7 ‘The method which we have discussed applies only to random samples from normal populations (or at least to random samples from populations which can be approximated closely with normal distributions). Exercises 9.1 Use the data of Exercise 7.14 to estimate @ forthe key (a) the sample standard deviation: performance indicator in terms of (a) the sample standard deviation: (6) the sample range. ‘Compare the two estimates by expressing their dffer- (©) the sample range. cence asa percentage of the fist. ‘Compare the two estimates by expressing ther differ- 9.4 With reference to Exercise 7.56, construct a 95% con- ence as a percentage of the first fidence interval forthe variance of the yield 9.2 With reference to Example 7, Chapter 8, use the range 9.5 With reference to Exercise 7.63, construct a 99% of the second sample to estimate o for the resiliency confidence interval forthe Vatiance of the population ‘modulus of recycled materials from the second loca~ sampled tion, Compare the result with the standard deviation 9,6, Use the value s obtained in Exercise 9.3 to construct of the second sample, 298% confidence interval for ¢, measuring the actual 9.3. Use the data of pat (a) of Exercise 8.13 to estimate variability in the hardness of Alloy 1 «6 for the Brinell hardness of Alloy 1 in terms of 9.2. Hypotheses Concerning One Variance In this section we shall consider the problem of testing the mull hypothesis that a population variance equals a specified constant against a suitable one-sided or two- sided alternative; that is, we shall test the null hypothesis o? = oj against one of the alematives o? < of,02 > of, oro? # of. Tests lke these are important whenever lis desired 6 contol the unifoctaty ot product or an operation, For Example, suppose that a silicon disk or wafer is tobe out info small squares, or dic, to be used in the manufacture ofa semiconductor device. Sine certain clectical charactersis of the finshed device may depend onthe thickness of the die, ts important that all dice eut from a wafer have approximately the same thickness. This, not only must the mean thickness of a wafer be kept within specifications, but also the variation in thickness from location to location onthe wafer, ‘Using the same sampling theory as on page 291, we base such tests on the fact that fora random sample from anormal popoation withthe vatiance of Statistic for test concerning variance (normal population) is a random variable having the chi square distribution with n—1 degrees of freedom, ‘The critical regions for such tests are as shown in the following table:294 Chapter berences Concerning Varanees Eze Solution Figure 9.2 The rejection region and P-value for Example 3 ers Cee) ‘Alternative hypothesis Fed o> ap Pod tag P< an a Pox, where 12 is as defined on page 208. Note that equal tails are used forthe 1wo- Sided stematve, hough thsi actusly not the best procedure since th chi square tistbutin snot symmetical. For moderate degrees of freedom. the io tess ate neatly the same Testing hypotheses concerning a standard deviation ‘The lapping process which is used to grind certain silicon wafers to the proper thickness is acceptable only if , the population standard deviation of the thick- ness of dice cut from the wafers, is at most 0.50 mil. Use the 0.05 level of signif- icance to test the null hypothesis o = 0.50 against the alternative hypothesis © > 0.50, if the thicknesses of 15 dice cut from such wafers have a standard deviation 0f 0.64 mil, 1. Null hypothesis: 6 = 0.50 Alternative hypothesis: 6 > 0.50 2. Level of significance: a = 0.05 3. Criterion: Reject the null hypothesis if x2 > 23.685, the value of x25 for 14 degrees of freedom, where 4. Calculations: 1)(0.64 (0.50)? | x 22.94 S$. Decision: Since x2 = 22.94 does not exceed 23.685, the null hypothesis cannot be rejected; even though the sample standard deviation exceeds 0.50, there is not suflicient evidence to conclude that the lapping process is unsatisfactory. ‘The rejection region and P-value are shown in Figure 9.2. 0.061 0 2368 . Statistical software is readily available to obtain other values of x3. See Exer- cise 9,23 for the MINITAB and Example 8, Chapter 6 for the R commands,Stat jc for test of equality of two variances (normal popul ons) Sec Hypotheses Concering Two Variances 295, 9.3 Hypotheses Concerning Two Variances Tho two sample f test described in Section 8.2 requires that the variances of the two populations sampled are equal. Tn this section we describe a test ofthe null hy- pothesis 0? — 02, which applies to independent random samples from two normal populations. It mast be used with some discretion as itis very sensitive to departures from this assumption If independent random samples of size my and np are taken from normal popt- lations having the same variance, it follows from Theorem 6.6 that is a random variable having the F distribution with nj —1 and mp —1 degrees Of freedom, Thus, ifthe nll hypothesis of =o2 is true, the ratio of the sample variances $? and 52 provides a statistic on which tests of the null hypothesis can be based ‘The critical region for testing the null hypothesis 0? tive hypothesis o? > of is F > Fe, where Fy is as defined on page 209. Similaly the critical region for esting the all hypothesis agains the altemative hypothesis G7 Fy, where Fy isthe appropriate critical value of F for ny — Land nj — 1 degrees of freedom. For the two-sided alternative 07 4.02, the critical region is F< F\_a/2 oF F > Fyyp, where F = $2/S3 and the degrees of freedom are ny — 1 and ny ~ 1. In practice, we modify ths test as inthe preceding paragraph, so that we can again use the table of F values corresponding to right-hand tals of « = 0.05 and a = 0.01. To this end we let S3,represent the larger of the two sample variances, $3 the smaller, and we write the corresponding sample sizes as nyg and my. Thus, he test statistic becomes F = Sj,/S%, and the critical region is as shown in the following table: pre ne rec es Cee) ‘Alternative Test hypothesis _ statistic Reject null hypothesis if op Falm—lm—1) of od F> Palm —lm—1) of tot F > Fyne 1)