Likelihood Ratio Tests 439 ‘the mean of a normal distribution with unknown variance was derived in Example 6.5.1. We next derive the two-sample t-test. Example 8.3.1. Let the independent random variables X and Y have distributions that are N (01,03) and 1V (02,05), where the means 6; and 82 and common variance 03 are unknown. ‘Then 9 = {(61,62,03) : -00 < 81 < 00,00 < Op < 00,0.< Bs < 00}. Let X1,X2,..-Xn and Yi,Yo,.»-,¥%m denote independent random samples from these distributions. ‘The hypothesis Ho : &, = 2, unspecified, and 8 unspecified, is to be tested against all alternatives. ‘Then w = {(81,62,03) : 00 < 6; = @2 < 00,0 < 83 < 00). Here Xi, X2,-..,Xns¥is¥oy---4¥m are n-+m > 2 mutually independent random variables having the likelihood functions (cool fern} 1 and (orm » um) = (Ze) * oo {-3, [Se a+ S3m-094}. 1 S542) and 29K He) are equated to zero, then (Exercise 8.3.2) Yea) + tua) ; ; é [de ~098 Sa a ‘The solutions for 0, and Oy are, respectively, w= @+m {S=+3n} (nt+m)* {Se -uP+ Liu = ot}. Further, w and w maximize L(a)- The maximum is ety orma (=) Alog L(2) DlogL(Q) log L(M) Oh” 0, " = i 3 (83.3) 8 " L(@) Ina like manner, if440 Optimal Tests of Hypotheses are equated to zero, then (Exercise 8.3.3) (83) ‘he solution for 8, Band 8 ae, respectively, —— wD = Sw w= (nem [two + Sonn, and, further, a, 12, and w" maximize L(9). ‘The maximum is Aleiyes osm daseees¥en) = A= a = (ema 1) = (5 so that orn “The random viable dined by 42/0 i FP + ow -¥)? DO = (OR + mPa tm} + SG OOF + MP ln-+ my]? t T Now EMR) =F [emer SEY] = Seep an (e-MERELikelihood Ratio Tests at Se Ee) Elo n( _ nk +m¥ and (Sr) = weep Hence, the random variable defined by A?/("+™) may be written Yo-7+ Dm - 7? DO% XP + OF)? + [nmr + my YP _ 1 — [nm/(n +m) Ty loan my" ame me yy If the hypothesis H : 0, = 8 is true, the random variable 1p T a ie o mF {+m-2) [sm -xP+ yu -7| } (83.5) thas, in accordance with Section 3.6, a t-distribution with n-+m— 2 degrees of freedom, Thus the random variable defined by A2/(™+™) ig ntm-2 em-D+T™ ‘The test of Hy against all alternatives may then be based on a t-distribution with nm ~2 degrees of freedom. The likelihood ratio principle calls for the rejection of Ho if and only if A < Xo < 1. Thus the significance level of the testis = Pizg|A(X1y 02+) Xny Yay Yon) S ole40 Optimal Tests of Hypotheses However, A(Xi,..-)Xm Yis-s++¥m) £ Ao is equivalent to [7] > o, and s0 a= P(|T| > ¢; Ho). For given values of rand m, the number c is determined from Table TV in Appendix. B (with n+m—2 degrees of freedom) to yield a desired a. Then Ho is rejected at ‘significance level a if and only if jt} > ¢, where t is the observed value of 7. If, for instance, n= 10, m = 6, and a = 0.05, then c= 2.145, m For this last example as well as the one-sample t-test derived in Example 6.5.1, it was found that the likelihood ratio test could be based on a statistic which, when the hypothesis Ha is true, has a t-distribution, ‘To help us compute the powers of these tests at parameter points other than those described by the hypothesis Ho, ‘we turn to the following definition Definition 8.3.1. Let the random variable W te N(6,1); let the random variable V bex*(r), and W and V be independent. The quotient Ww. Wir is said to have a noncentral t-listribution with r degrees of freedom and noncen- irality parameter 5. If 5 =0, we say that T has a central t-distribution. In the light of this definition, let us reexamine the t-statistics of Examples 6.5.1 ‘nd 83.1. In Example 6.5.1 we had HX). Xa) = Lew - PF /n-1) Vink) DG —X)*/lo?(m—0)] Here, where is the mean of the normal distribution, Wi = VACX/o is N(fi/o;1), Vi = DQ — X)/e? is x2(n— 1), and MW and Vs are independent, ‘Thus, if 4 #0, we s0e, in accordance with the definition, that t(X1,...,X,) has a non- central t-isribution with n — 1 degrees of freedom and noncentrality parameter 4, = vib; /o. In Example 8.3.1 we had We Vilaem ay We \ee-7/« Tr8.3. Likelihood Ratio Tests 443 and Yo FP + 06 - PP a Here Ws is N[y/nm](n + m)(0; ~ 02)/0, 1], Vo is x2(n+m—2), and W and V2 are independent. Accordingly, if@: # 8, T has a noncentral t-distribution with n-+m—2 degrees of freedom and noncentrality parameter 62 = /nm/(n + m)(0; —62)/0. It is interesting to note that 51 = y/nidi/o measures the deviation of 8; from 0; = 0 in units of the standard deviation a//n of X. The noncentrality parameter 62 = /nm] (nm + m)(0; — 62)/o is equal to the deviation of 0; — 02 from 9; ~ 62 = 0 in units of the standard deviation o/ (+ m)/mn of X —Y. ‘The packages Rand S-PLUS contain functions which evaluate noncentral t- distributional quantities. For example, to obtain the value P(T’< t) when T has ‘tlistribution with a degrees of freedom and noncentrality paraineter b, use the command pt(t, a, ncp=d). For the value of the associated pdf at t, use the command dt(t, a, ncp=d). There are also various tables of the noncentral t- distribution, but they are much too cumbersome to be included in this book. Y= Remark 8.3.1. The one and two sample tests for normal ineans, presented in Examples 6.5.1 and 8.3.1 are the tests for normal means presented in most elemen- tary statistics books. ‘They are based on the assumption of normality. What if the underlying distributions are not normal? In that case, with finite variances, the test statistics for these situations are asymptotically correct. For example, consider the one sample t-test. Suppose X1,..-,%q are iid with a common nonnormal pdf which has mean 0; and finite variance 72, The hypotheses remain the same, ie, 4, = 0, versus Hy : 0, £05. The t-test statistic, Ta, is given by n= VIR 8), Sa where Sq is the sample standard deviation, Our critical region is Cx = {[Tal > ta/an—1}+ Recall that Sn +o in probability. Hence, by the Central Limit Theorem, under Ho ar o VaCK-H) b gas, where Z has a standard normal distribution. Hence, the asymptotic test would use the critical region Cy = {|Tq| 2 Zay2}- By (8.3.7) the critical region C would have approximate size a. In practice, we would use C;. Because t critieal values are generally larger that z critical values, the use of C; would be conservative; ie, the size of C would be slightly smaller than that of G2. In terms of robustness, we would say the t-test possesses robustness of validity. But the t-test. does not possess robustness of power. For nonnorinal situations, there are more powerful tests than the t-test; see Chapter 10 for discussion. ‘As Exercise 8.3.4 shows, the two-sample t-test is also asymptotically correct, provided the underlying distributions have the same variance. m (8.3.6) T. (83.7)