Fixed-Width Simultaneous Confidence Intervals for Multinormal Means in Several Intraclass Correlation Models

Transkript

1 Journal of Multivariate Analysis 66, 4663 (1998) Article No. MV Fixed-Width Siultaneous Confidence Intervals for Multinoral Means in Several Intraclass Correlation Models Makoto Aoshia Tokyo Gakugei University, Tokyo, Japan and Nitis Mukhopadhyay University of Connecticut, Storrs, Connecticut Received August 27, 1996; revised August 5, 1997 The proble of constructing fixed-width siultaneous confidence intervals for coparing ean vectors of k(2) independent ultivariate noral distributions is considered when those covariance atrices have the intraclass correlation structures. Two-stage procedures are developed for which the siultaneous confidence levels are shown to be at least 1&:, the preassigned noinal value, 0<:<1. Asyptotic efficiency properties are addressed. In the case of fixed and finite initial saple size, efficiency related issues are also discussed Acadeic Press AMS 1991 subject classifications: 62L05, 62H99, 62J15. Key words and phrases: Equal intraclass covariance atrix, unequal intraclass covariance atrices, siultaneous inference, two-stage procedures, exact consistency, efficiency. 1. INTRODUCTION Let us suppose that there are k(2) p-variate noral distributions having unknown ean vectors + i and unknown covariance atrices i, i=1,..., k. In order to proceed with ultiple coparisons on the ean vectors, we consider siultaneous confidence intervals (s.c.i) for the double linear cobination a$mb, for all a # R p &[0], and b # R k &[0] where M=(+ 1,..., + k ), a p_k atrix. The proble is to construct the fixed-width s.c.i., all having a preassigned length 2d (>0), while the siultaneous confidence coefficient is at least 1&:, where 0<:<1 is also preassigned. It is known that there is no fixed-saple procedure for this proble when i 's are unknown. Refer to Dantzig (1940) and Section 3.7 in Ghosh et al. (1997) for soe perspectives. In order for a solution to enjoy the X Copyright 1998 by Acadeic Press All rights of reproduction in any for reserved.

2 SIMULTANEOUS CONFIDENCE INTERVALS 47 exact consistency property, that is, the siultaneous confidence coefficient is at least 1&:, one ust adopt sapling in at least two-stages. Here, we intend to ephasize upon (ultivariate) two-stage procedures for the proble on hand. Aoshia (1991a, 1992) gave soe solutions for this proble when i 's are unequal and positive definite (p.d.) by iprovising upon Chatterjee's (1959a,b) ethodology. We recall that Chatterjee's ethodology was built upon Stein's (1945) type I schee. Healy (1956), on the other hand, produced his two-stage procedure built upon Stein's type II schee, and along these lines Aoshia (1994a) gave a two-stage schee for a siilar proble. It was shown in Aoshia (1994a,b) that the solution based on Stein's type II schee was ore econoical, in the sense of saller saple size, than the one based on Stein's type I schee. One is also referred to Chatterjee (1991) for a review. In the case when i 's are unknown, p.d., but have soe special structures, the proble has been addressed by Nagao (1996), aong others, only in the case when k=1. Also, when k=1 and =_ 2 [\11$+ (1&\) I p ] (= I, say) with _ 2 (>0) and &( p&1) &1 <\<1 unknown, 1$=(1,..., 1), Bhargava and Srivastava (1973) proposed a purely sequential sapling schee under the assuption that a$1=0. In this sae setup, Chaturvedi et al. (1992) focussed on second-order asyptotic properties of the purely sequential ethodology. Recently, Hyakutake et al. (1995) proposed a two-stage procedure in the case of this particularly structured I, along the lines of the type II schee without any assuption on the vector a. A purely sequential procedure and soe relevant coparisons were also provided in Hyakutake et al. (1995). Here and elsewhere, I p stands for the p_p identity atrix. In the present paper, we extend the two-stage procedure of Hyakutake et al. (1995) proposed for the case k=1 to the general case when k2. We suppose that i 's have the intraclass correlation structures. In Section 2, we deal with the situation when 1 =}}}= k = I that depends on unknown _ 2 and \, while two-stage procedures are provided generalizing further the ideas behind Healy's and Hyakutake et al.'s sapling schees. It is shown that these procedures satisfy both the exact consistency property and asyptotic ( first-order) efficiency property. In Section 3, we deal with the situation when i 's are unequal, and ore specifically when i has the intraclass correlation structure Ii that depends on unknown _ 2 i and \ i, i=1,..., k. Two-stage procedures are proposed in this case which parallel the developents given in Section 2. Exact consistency and asyptotic ( first-order) efficiency results are also provided in Section 3. It should be noted that the asyptotics are presented as d 0, and when the starting saple size grows to infinity at soe suitable rate as d 0. For each of these procedures, one needs expressions for certain percentage points of coplicated statistics arising fro the ultivariate structures,

3 48 AOSHIMA AND MUKHOPADHYAY while these depend on :, p, k and the initial saple size. In Section 4, large saple approxiations for such percentage points are provided for ease of ipleentation of the proposed double sapling schees. Finally, in Section 5, superiority of the two-stage procedure derived along the lines of Hyakutake et al. (1995) is established in the ultivariate case with k2 when the initial saple size is held fixed. 2. EQUAL COVARIANCE MATRIX WITH INTRACLASS CORRELATION STRUCTURE We assue that the i th population's distribution is given by N p (+ i, ) where = I =_ 2 [\11$+(1&\) I p ], with all the paraeters + i # R p, _ #(0,), and \ #(&(p&1) &1, 1) unknown, i=1,..., k. Let X ij, j=1,2,... be independent rando vectors fro the i th population, X n=(x 1n,..., X kn) be the p_k atrix, where X in=n &1 n j=1 X ij, i=1,..., k. The proble of constructing the fixed-width s.c.i. for a$mb can be viewed as the fulfillent of the following requireent: P[ a$(x n&m) b d for all a # R( p), b # R(k)]1&: (2.1) with given d (>0) and : # (0, 1), where R(s)=[c: &c&=1, c # R s ] for s= p, k. Had _ 2 and \ been known in the considered structure with 1 =}}}= k = I =_ 2 [\11$+(1&\) I p ], the ``optial'' fixed saple size associated with (2.1) would then be given by where n*= _%2 ( p, k, :) d 2 * &+1 (2.2) [x]=largest interger <x; % 2 ( p, k, :) is the upper : point of the distribution of the axiu latent root % 2 ( p, k) ofaw p (k, I p ) atrix; and * is the axiu latent root of I, that is, *=ax({ 1, { 2 ) with { 1 =_ 2 ( p\+1&\) and { 2 =_ 2 (1&\). This n* is optial in the sense that it is the sallest fixed saple size required to guarantee (2.1), had _ 2 and \ been known. Since * is unknown, we estiate it on the basis of the initial observations, each of size ([pk &1 ]+2), fro all populations. Let * be the axiu latent root of k S =& &1 : i=1 : j=1 (X ij &X i)(x ij &X i)$, &=k(&1).

4 SIMULTANEOUS CONFIDENCE INTERVALS 49 Extending along Healy's (1956) two-stage schee, one can define Q=ax _ %2 ( p, k, :) {, * d & +1 =, (2.3) 2 with % 2 ( p, k, :) being the upper : point of the distribution of the axiu root % 2 (=%2 ( p, k)) of the deterinantal equation U&% 2 &&1 V =0, (2.4) where U and V are, respectively, independent W p (k, I p ) and W p (&, I p ) atrices. Here and elsewhere, we write W p (r, I p ) for the central Wishart distribution with r degrees of freedo, in p diension, and its custoary atrix replaced by I p.ifq=, we do not saple any ore in the second stage fro any population. On the other hand, if Q>, then we saple the difference (Q&) fro each population in the second stage. Fro the cobined set of observations fro each population, one coputes the p_k saple ean atrix, X Q. Theore 2.1. Under the two-stage sapling schee (2.3), for fixed d (>0), we conclude that the stateent a$mb #[a$x Qb\d ] for all a # R( p), b # R(k) (2.5) provides a solution for the fixed-width s.c.i. proble which is exactly consistent. Under the additional condition that =(d) satisfies the following: (d), d 2 (d) 0, as d 0, we conclude that the two-stage procedure (2.3) is also asyptotically ( first-order) efficient, that is, with n* defined in (2.2). li E(Qn*)=1, (2.6) d 0 Proof. We have P[ a$(x Q&M) b d for all a # R( p), b # R(k)] E { a$(x Q&M)(X Q&M)$ a _P Qd 2 a$(s Q) a * for all a # R( p) S =& Q&M)(X Q&M)$ a P _a$(x % 2 a$(s Q) a ( p, k, :) for all a # R( p) & =1&:, (2.7)

5 50 AOSHIMA AND MUKHOPADHYAY in view of the choice of % 2 ( p, k, :) via (2.4). The last inequality holds in view of (2.3), since Qd 2 * % 2 ( p, k, :) w.p. 1. In order to verify (2.6), one writes the basic inequality along the lines of Mukhopadhyay (1980) and notes that % 2 ( p, k, :) %2 ( p, k, :) as well as * * w.p. 1 as d 0. Further details are oitted for brevity. K Next, we consider another two-stage schee by extending the ideas fro Hyakutake et al. (1995). We start with an initial saple of size ([pk &1 ]+2) fro all populations. Define {^ 1= p &1 1$S 1 and {^ 2= ( p&1) &1 [tr(s )&{^ 1] respectively for unbiased estiators of { 1 and { 2, while * is then estiated by eans of ax({^ 1, {^ 2). Let U be a W p (k, I p ) atrix, & i V i be independent / 2 rando variables with & i degrees of freedo, i=1, 2, where & 1 =k(&1), & 2 =k( p&1)(&1), while U and (V 1, V 2 ) are also assued independent. Define $ 2 ( p, k, :) as the upper : point of the distribution of the axiu root $ 2 (=$2 ( p, k)) of the deterinantal equation, Now, let U&$ 2 diag(v 1, V 2,..., V 2 ) =0. (2.8) N=ax _ $2 ( p, k, :) {, ax({^ 1, {^ 2) d = 2 &+1, (2.9) and this two-stage procedure is ipleented in a siilar fashion as it was in the case of (2.3). Finally, based on N observations fro each population, we obtain the p_k saple ean atrix, X N. Theore 2.2. Under the two-stage sapling schee (2.9), for fixed d (>0), we conclude that the stateent a$mb #[a$x Nb\d] for all a # R( p), b # R(k) (2.10) provides a solution for the fixed-width s.c.i. proble which is exactly consistent. Under the additional condition that =(d) satisfies the following: (d), d 2 (d) 0, as d 0, we conclude that the two-stage procedure (2.9) is also asyptotically ( first-order) efficient, that is, with n* defined in (2.2). li E(Nn*)=1, (2.11) d 0

6 SIMULTANEOUS CONFIDENCE INTERVALS 51 Proof. Let H be a p_p orthogonal atrix whose first colun is p &12 1. Then, H$ I H=diag({ 1, { 2,..., { 2 ) and write I=H diag({^ 1, {^ 2,..., {^ 2) H$. Also, define Y in=h$(x in&+ i ), i=1,..., k, and we further write Y N=(Y 1N,..., Y kn) for the corresponding p_k ``saple ean'' atrix. Let h=h$a for a # R( p). Fro (2.9), we observe that Nd 2 $ 2 ( p, k, :) ax({^ 1, {^ 2) w.p. 1, and thus P[ a$(x N&M) b d for all a # R( p), b # R(k)] h$y NY $ =E { N h _P h$h$( IN) Hh Nd 2 h$y NY $ E { N h _P h$h$( IN) Hh Nd 2 ax({^ 1, {^ 2) h$h$ I Hh for all h # R( p) I=& for all P { h$y NY $ N h h$h$( IN) Hh $2 ( p, k, :) for all h # R( p) = h # R( p) I=& =1&:, (2.12) since & 1 {^ 1{ 1 and & 2 {^ 2{ 2 are distributed as / 2 & 1 and / 2 & 2 respectively, while they are also independent. The last equality in (2.12) follows in view of the choice of $ 2 ( p, k, :) via (2.8). The proof of (2.11) is siilar to the proof of (2.6). Details are oitted. K It is clear fro (2.6) and (2.11) that both the two-stage procedures (2.3) and (2.9) are going to take close to n* observations fro all the populations asyptotically, when =(d), but d 2 (d) 0 as d 0. In Section 5, we reark on the superiority of the two-stage schee (2.9) over the two-stage schee (2.3) for all p and k, when the initial saple size is kept finite, in the sense that N<Q with probability one. 3. UNEQUAL COVARIANCE MATRICES WITH INTRACLASS CORRELATION STRUCTURES We assue that the i th population's distribution is given by N p (+ i, i ) where i = Ii =_ 2 i [\ i11$+(1&\ i ) I p ], with all the paraeters + i # R p, _ i #(0,), and \ i #(&(p&1) &1, 1) unknown, i=1,..., k. Let X ij, j = 1, 2,... be independent rando vectors fro the i th population,

7 52 AOSHIMA AND MUKHOPADHYAY X n =(X 1n 1,..., X kn k ) be the p_k saple ean atrix, where X in i = n &1 i n i X j=1 ij, i=1,..., k and n=(n 1,..., n k ). Now, the proble of constructing the fixed-width s.c.i. for a$mb can be viewed as the fulfillent of the following requireent: P[ a$(x n &M) b d for all a # R( p), b # R(k)]1&: (3.1) with preassigned d (>0) and : #(0,1). Let * i denote the largest latent root of Ii, that is * i =ax({, 1 {) with 2 { 1 =_2( p\ i i+1&\ i ) and { 2 =_2(1&\ i i), i=1,..., k. Also, let us write / 2 (:) for the upper : point of the kp /2 kp distribution, and D=diag (n &1 1 a$ I1 a,..., n &1 k a$ Ik a). Now, (3.1) leads to P[ a$(x n &M) b d for all a # R( p), b # R(k)] P {b$(x n &M)$ aa$(x n &M) b b$db for all a # R( p), b # R(k) = ax 1ik d 2 (* i n &1 i ) =P {a$(x n &M) D &1 (X n &M)$ a ax 1ik d 2 (* i n &1 i ) for all a # R( p) = P { : k i=1 (X in i &+ i )$ (n &1 i i ) &1 (X in i &+ i ) ax 1ik d 2 (* i n &1 i )= Hence, the optial fixed-saple sizes, when _ i 's and \ i 's are assued known, are. n i *= (:) kp * _/2 d i& +1, (3.2) 2 i=1,..., k. In order to estiate * i, we take initial observations of size (p+1) fro each population and let * be the axiu latent root of S =&&1 : j=1 (X ij &X i)(x ij &X i)$, &=&1, i=1,..., k.

8 SIMULTANEOUS CONFIDENCE INTERVALS 53 Extending along Healy's (1956) two-stage schee, we define T i =ax _ t2 {, ( p, k, :) d 2 * & +1 =, (3.3) where t 2 ( p, k, :) is the upper : point of the distribution of the su of k independent Hotelling's T 2 ( p, &) rando variables, i=1,..., k. We write T=(T 1,..., T k ). If T i =, we do not saple any ore fro the i th population, however, if T i >, then we saple the difference (T i &) fro the i th population in the second stage, i=1,..., k. Fro the cobined set of observations fro all populations, we copute the p_k saple ean atrix, X T. Theore 3.1. Under the two-stage sapling schee (3.3), for fixed d (>0), we conclude that the stateent a$mb #[a$x Tb\d] for all a # R( p), b # R(k) (3.4) provides a solution for the fixed-width s.c.i. proble which is exactly consistent. Under the additional condition that = (d) satisfies the following: (d), d 2 (d) 0 as d 0, we conclude that the two-stage procedure (3.3) is also asyptotically ( first-order) efficient, that is, with n i * defined in (3.2), i=1,..., k. li E(T i n i *)=1, (3.5) d 0 Proof. Fro (3.3), we observe that T i d 2 t 2 ( p, k, :) * w.p. 1, i= 1,..., k, and thus with D*=diag((a$S (1) at 1),..., (a$s (k) at k)), we have P[ a$(x T &M) b d for all a # R( p), b # R(k)] E _P { a$(x T &M) D* &1 (X T &M)$ a P { : k i=1 for all a # R( p) S (1),..., S (k) =& ax 1ik d 2 (T &1 i * ) T i (X it i &+ i )$ (S )&1 (X it i &+ i )t 2 ( p, k, :) = =1&:, (3.6)

9 54 AOSHIMA AND MUKHOPADHYAY in view of the choice of t 2 ( p, k, :) in (3.3). The result in (3.5) follows iediately, since t 2 ( p, k, :) /2 (:) and * * kp i w.p. 1, i=1,..., k as. K Next, we consider another two-stage schee by extending the ideas fro Hyakutake et al. (1995). We start with (p+1) observations fro each population. Define {^ 1= p &1 1$S 1 and {^ 2=(p&1) &1 [tr(s )&{^ ] respectively for unbiased estiators of { 1 and { while * 2 i is then estiated by ax({^ 1, {^ 2), i=1,..., k. Let & 1 =&1 and & 2 =(&1)(p&1), and we consider independent rando variables Y i =U 1 +(p&1) U, 2 i=1,..., k, where U 1 and U 2 are independently distributed as F 1, &1 and F p&1, &2 respectively. We write f 2 ( p, k, :) for the upper : point of the distribution of k Y i=1 i and define N i =ax _ f 2 ( p, k, :) {, ax({^ d, {^ ) & +1 =, (3.7) i=1,..., k. Let N=(N 1,..., N k ). If N i =, we do not take any ore observations fro the i th population, however, if N i >, then we saple the difference in the second stage fro the i th population, i=1,..., k. Finally, based on all the observations fro all populations, we obtain the p_k saple ean atrix, X N. Theore 3.2. Under the two-stage sapling schee (3.7), for fixed d (>0), we conclude that the stateent a$mb #[a$x Nb\d] for all a # R( p), b # R(k) (3.8) provides a solution for the fixed-width s.c.i. proble which is exactly consistent. Under the additional condition that = (d) satisfies the following: (d), d 2 (d) 0 as d 0, we conclude that the two-stage procedure (3.7) is also asyptotically ( first-order) efficient, that is, with n i * defined in (3.2), i=1,..., k. li E(N i n i *)=1, (3.9) d 0 Proof. Let H be a p_p orthogonal atrix whose first colun is p &12 1. Then, H$ Ii H=diag({, 1 { 2,..., { ), i=1,..., k. Let D 2 * = diag(n &1 1 a$ I1 a,..., N &1 k a$ Ik a) where Ii=H diag({^ 1, {^ 2,..., {^ 2) H$, i=1,..., k. Fro (3.7), we observe that N i d 2 f 2 ( p, k, :) ax({^ 1, {^ 2) w.p. 1, i=1,..., k, and thus

10 SIMULTANEOUS CONFIDENCE INTERVALS 55 P[ a$(x N &M) b d for all a # R( p), b # R(k)] E { _P a$(x N &M) D &1 (X * N &M)$ a ax for all a # R( p) S (1),..., S (k) =& d 2 1ik{ ax({^, {^ ) 1 2 N i = P[a$(X N &M) D &1 (X * N &M)$ a f 2 ( p, k, :) for all a # R( p)] P { k : &1 N i (X in i &+ i )$ Ii (X in i &+ i ) f 2 ( p, k, :) = i=1 =P _ k : 1 V 1 {^ + { 2 {^ i=1{ { 1 2 V 2 = f 2 ( p, k, :) &, (3.10) where V 1 and V, i=1,..., k, are independent 2 /2 and 1 /2 rando p&1 variables respectively, and (V, V ) is also independent of ({^, {^ ), i=1,..., k. The last line in (3.10) can be clearly written as P[ k Y i=1 i f 2 ( p, k, :)], which is exactly 1&:, in view of the choice of f 2 ( p, k, :). The result in (3.9) follows iediately since f 2 ( p, k, :) /2 kp (:) and ax({^, {^ ) * 1 2 i w.p.1 for i=1,..., k, asd 0. K It is clear fro (3.5) and (3.9) that both the two-stage procedures (3.3) and (3.7) are going to take close to n* i observations fro the i th population asyptotically, when =(d), but d 2 (d) 0asd0, i=1,..., k. In Section 5, we reark on the superiority of the two-stage schee (3.7) over the two-stage schee (3.3) for all p and k, when the initial saple size is kept finite, in the sense that N i <T i with probability one, i=1,..., k. 4. LARGE SAMPLE APPROXIMATIONS FOR THE UPPER : POINTS Each two-stage procedure developed in Sections 2 and 3 requires evaluation of the upper : point of soe appropriate statistic's distribution. Now, we provide soe useful approxiating forulas that can be used instead of the exact upper : point under consideration, while ipleenting the relevant two-stage procedure Equal Intraclass Correlation Matrix Here, we obtain approxiating forulas for % 2 ( p, k, :) and $2 ( p, k, :) which were respectively utilized in defining the two-stage schees (2.3) and (2.9). The distributions of the largest latent root of deterinantal equations, such as (2.4) and (2.8), are very coplicated even for fixed, but large, values of p, k and. One ay refer to Aoshia (1991b) for other relevant exaples.

11 56 AOSHIMA AND MUKHOPADHYAY First, observe that the probability expression given in the last but one line in (2.7) is bounded fro below by k P { : (X i&+ i )$ S &1 (X i&+ i )% 2 ( p, k, :) = i=1 =P[& tr(uv &1 )%2 ( p, k, :)] (4.1) ) is the standard LawleyHotelling T2 0 statistic. Utilizing a large saple approxiation for the distribution of tr(uv &1 ), given in where tr(uv &1 Siotani et al. (1985, p. 252) along the lines of Lea 55 in Aoshia (1994a), we can derive the result stated as Theore 4.1. In a sense, the asyptotic expansion forulas given in this section can be viewed as descendants of the well-known approaches that capitalize expansions of a distribution function to obtain expansions of percentiles of the distribution. See, for exaple, Hill and Davis (1968). Also, we keep p, k, and : fixed throughout the anuscript, and hence the asyptotic expansions are provided up to certain negative powers of, the starting saple size. Theore 4.1. The design constant % 2 ( p, k, :) defined via (2.4) has an upper bound c~ which can be expanded as / kp(:) (k+ p+1) 2 kp (:)+/2 / 2& { 2 kp (kp+2) (:)&k+ p+1 = + /2 (:) kp 24& { 6(k+2)(k&1)(p+2)(p&1) / 6 2 (kp+6)(kp+4)(kp+2) (:) 2 kp + (4p3 +6p 2 +34p+36) k+(6p 2 +6p+20) k 2 +(4k 3 p+20p 2 +36p+8) (kp+4)(kp+2) 2 _/ 4 kp(:)& (11k2 &13p 2 &24p&7) / 2 (kp+2) kp(:)&12( p+1) k +7k 2 +7p 2 +12p+1 =+O(&3 ), (4.2) where &=k(&1) and / 2i (:) stands for kp [/2 kp (:)]i with i=2, 3. Next, we turn our attention to derive a siilar expansion for a suitable upper bound of $ 2 ( p, k, :). Observe that the probability expression given in the last but one line in (2.12) is bounded fro below by k P { : &1 Y $ i diag({^, {^ &1,..., {^ &1 ) Y i$ 2 ( p, k, :) = i=1 =P { { 1 U 1 + { 2 U 2 $ 2 {^ 1 {^ ( p, k, :) 2 =, (4.3)

12 SIMULTANEOUS CONFIDENCE INTERVALS 57 where 12 Y i tn p (0, diag({ 1, { 2,..., { 2 )), and U 1 and U 2 are independent rando variables, (U 1, U 2 ) is independent of ({^ 1, {^ 2), while U 1 t/ 2 and k U 2 t/ 2 k( p&1). Thus, the last expression in (4.3) siplifies to P[kZ 1 +k( p&1) Z 2 $ 2 ( p, k, :)] (4.4) where Z 1, Z 2 are independent rando variables and Z 1 tf k, &1, Z 2 tf k( p&1), &2, with & 1 =k(&1), & 2 =k(&1)(p&1). Utilizing the asyptotic expansion of the distribution function of qz with ZtF q, r as given in Siotani (1957), we obtain the following fro (4.4) by inversely transforing the characteristic function: For x>0, we write P[kZ 1 +k( p&1) Z 2 x] =G kp (x)+ k 4& [(kp&4) G kp(x)&2kpg kp+2 (x)+(kp+4) G kp+4 (x)] k + 96( p&1) & 2 [[(3k2 p 2 &32kp+96)(p&1) k&16p] G kp (x) &12(kp&4)(p&1) k 2 pg kp+2 (x)+6(3kp+8)(p&1) k 2 pg kp+4 (x) &4[(3k 2 p 2 +28kp+96)(p&1) k+32p] G kp+6 (x) +3[(k 2 p 2 +16kp+96)(p&1) k+48p] G kp+8 (x)] +0( &3 ), (4.5) where G s (x)=p(/ 2 s x) and &=k(&1). Now, one cobines (4.5) with Lea 55 in Aoshia (1994a), to derive the result in Theore 4.2. Theore 4.2. The design constant $ 2 ( p, k, :) defined via (2.8) has an upper bound c which can be expanded as / 2 kp (:)+/2 kp (:) 2p& { (kp+4) (kp+2) /2 kp (:)&kp+4 = + /2 kp (:) 24( p&1) p 2 & 2 _ {24(3p2 &10p+10) kp+12(p&1) k 2 p 2 +12(12p 2 &48p+48) (kp+6)(kp+4)(kp+2) 2 / 6 kp (:) + 4(kp+11)(p&1) k2 p 2 +8(p 2 +18p&18) kp+4(4p 2 +48p&48) (kp+4)(kp+2) 2 _/ 4 kp (:)&11( p&1) k2 p 2 &8p 2 &144p+144 (kp+2) / 2 kp (:) +(7kp&48)(p&1) kp+8p 2 +48p&48 =+0(&3 ), (4.6) where &=k(&1) and / 2i kp (:) stands for [/2 kp (:)]i with i=2, 3.

13 58 AOSHIMA AND MUKHOPADHYAY Reark 4.1. Equations (4.5)(4.6), when specialized for k=1, lead to Theore 2 of Hyakutake et al. (1995). Also, under the condition that =(d), but d 2 (d) 0asd0, it is easy to see that both c~ and c converge to / 2 kp (:) asd 0. Note also that the two-stage procedures (2.3) and (2.9) with % 2 ( p, k, :) and $2 ( p, k, :) respectively replaced by c~ =c~ ( p, k, :) and c=c ( p, k, :) certainly enjoy the exact consistency property as well. Table I provides the values of c~ and c given by the expressions in (4.2) and (4.6) respectively, for k=2(1)5, p=2(1)5, =10, 15, 20, 25, 100 and :=0.05. It is seen that for fixed p, k, and, the agnitude of c is saller than that of c~. We observe this tendency for larger p and k also. That is particularly in agreeent with our expectations in view of (4.2) and (4.6). Such a feature is noticeable even when is sall, but p is large. The referee asked us to check the closeness between c~ and % 2 ( p, k, :), as well as between c and $ 2 ( p, k, :). Exact nuerical evaluations of % 2 ( p, k, :) and $2 ( p, k, :) are very hard and extreely tie consuing. TABLE I Values of c~ and c for :=0.05 p=2 p=3 p=4 p=5 k c~ c c~ c c~ c c~ c

14 SIMULTANEOUS CONFIDENCE INTERVALS 59 We proceed via siulations to generate the distribution functions of % 2 ( p, k) and $2 ( p, k) in order to finally obtain the values of %2 ( p, k, :) and $ 2 ( p, k, :) within of accuracy. Table II provides those values. (It took us nearly 170 hours of coputer tie to generate Table II!) Fro Table II, one finds that % 2 ( p, k, :) is always larger than $ 2 ( p, k, :), as expected. For saller k and, the agreeents between % 2 ( p, k, :) and c~ as well as between $2 ( p, k, :) and c are clearly visible Unequal Intraclass Correlation Matrices Now, we address approxiating forulas for t 2 ( p, k, :) and f 2 ( p, k, :) which were respectively utilized in defining two-stage schees (3.3) and (3.7). In this vein, t 2 ( p, k, :) was earlier handled by Aoshia (1994a), and thus, we focus our attention only on f 2 ( p, k, :). Recall that f 2 ( p, k, :) was TABLE II Values of % 2 ( p, k, :) and $2 ( p, k, :) for :=0.05 p=2 p=3 p=4 p=5 k % 2 $ 2 % 2 $ 2 % 2 $ 2 % 2 $

15 60 AOSHIMA AND MUKHOPADHYAY defined as the upper : point of the distribution of k Y i=1 i where Y i =U 1 +(p&1) U 2, i=1,..., k, are independent, U tf 1 1, & 1, U t 2 F p&1, &2 and U, U 1 2 are also independent, with & 1 =&1, & 2 = (&1)(p&1). Next, in a fashion analogous to (4.5), by inversely transforing the characteristic function for k Y i=1 i, we obtain the following expansion: For x>0, we write P { : k i=1 Y i x = =G kp (x)+ k 4& [( p&4) G kp(x)&2pg kp+2 (x)+(p+4) G kp+4 (x)] k + 96k( p&1) & [[3( 2 p&1)(p&4)2 k&8p 2 +40p&48]G kp (x) &12(p&1)(p&4) kpg kp+2 (x)+6(3kp 2 &16k+8p+16)(p&1) _G kp+4 (x)&4[3( p+4)(p&1) kp+16p p&96]G kp+6 (x) +3[( p+4) 2 ( p&1) k+8p p&80]G kp+8 (x)] +0( &3 ), (4.7) with &=&1 and G s (x)=p(/ 2 s x), as before. Then, one cobines (4.7) with Lea 55 in Aoshia (1994a) to derive the following result. Theore 4.3. The design constant f 2 ( p, k, :), that was used in (3.7), can be expanded as / 2 kp (:)+/2 kp (:) 2p& { ( p+4) (kp+2) /2 kp (:)&p+4 = + /2 kp (:) 24( p&1) p 2 & 2 _ {24(4p2 &9p+8) kp&12(p 3 &9p 2 +44p&48) (kp+6)(kp+4)(kp+2) 2 / 6 kp (:) + 4( p2 +7p&6) kp 2 +4(5p 3 +35p 2 +12p&48) (kp+4)(kp+2) 2 / 4 kp (:) & (11p3 &19p 2 &144p+144) / 2 (kp+2) kp(:)+7p 3 &47p 2 +96p&48 = +0( &3 ), (4.8) with &=&1 and / 2i (:) standing for kp [/2 kp (:)]i, for i=2, 3. Reark 4.2. Equations (4.7)(4.8), when specialized for k=1, lead to Theore 2 of Hyakutake et al. (1995). In case =(d), but d 2 (d) 0, it is easy to see that f 2 ( p, k, :) /2 kp (:) asd0.

16 SIMULTANEOUS CONFIDENCE INTERVALS 61 TABLE III Values of t 2 ( p, k, :) and f 2 ( p, k, :) for :=0.05 p=2 p=3 p=4 p=5 k t 2 f 2 t 2 f 2 t 2 f 2 t 2 f Table III provides the values of t 2 ( p, k, :) and f 2 ( p, k, :) for k=2(1)5, p=2(1)5, =10, 15, 20, 25, 100, and :=0.05. The values of t 2 ( p, k, :) are coputed utilizing the table fro Aoshia (1994a), whereas the values of f 2 ( p, k, :) are coputed using the expression given in (4.8). Again, we note the feature that f 2 ( p, k, :) value is saller than t2 ( p, k, :), even when is sall, but p is large. 5. EFFICIENCY CONSIDERATIONS In the context of (2.3) and (2.9), one can verify that * ax({^ 1, {^ 2) for all 2, in view of Lea 2 of Hyakutake et al. (1995). Hence, when one replaces % 2 ( p, k, :) and $2 ( p, k, :) in (2.3) and (2.9) by c~ and c respectively, one will then conclude that the saple size associated with the two-stage schee (2.9) would be saller, with probability one, than that associated with the two-stage schee (2.3). This substantiates the superiority of the two-stage sapling schee derived along the lines of

17 62 AOSHIMA AND MUKHOPADHYAY Hyakutake et al.'s (1995) ethodology in the case when the coon unknown covariance atrix has the intraclass structure, for all k2. In the context of (3.3) and (3.7), again one observes as before that * ax[{^, {^ 1 2 ], i=1,..., k, for all 2. Hence, we conclude that the saple sizes associated with the two-stage schee (3.7) would be correspondingly saller, with probability one, than those associated with the two-stage schee (3.3), for i=1,..., k, and for all 2. This again substantiates the superiority of the two-stage sapling schee derived along the lines of Hyakutake et al.'s (1995) ethodology in the case when the unknown covariance atrices are unequal, but they have the intraclass structures, for all k2. In either scenario, we find that the two-stage sapling schee based on the generalization of Hyakutake et al.'s (1995) ethodology beats the corresponding two-stage sapling schee based on the analogous generalization of Healy's (1956) ethodology, when the associated saple sizes are copared. The latter ethodology ends up taking ore observations, with probability one, fro the populations under consideration. Hyakutake et al. (1995) proved this result only when k=1, 2 being arbitrary. We have now concluded that the sae result holds when both k (2) and (2) are arbitrary. ACKNOWLEDGMENTS Research of the first author was supported by the Grant-in-Aid for Encourageent of Young Scientists fro the Ministry of Education, Science, and Culture, Japan, under contract nuber We thank the referee for aking several helpful rearks. REFERENCES 1. M. Aoshia, Interval estiation on the prespecified coparisons of ean vectors aong heteroscedastic populations, Sci. Univ. Tokyo J. Math. 27 (1991a), M. Aoshia, Asyptotic distributions of quasi Wishart atrix and its latent roots in the heteroscedastic case, Sci. Univ. Tokyo J. Math. 27 (1991b), M. Aoshia, ``Multiple Coparisons aong Mean Vectors in the Heteroscedastic Case,'' Ph.D. Thesis, Faculty of Science, Sci. Univ. of Tokyo, M. Aoshia, In heteroscedastic siultaneous inference, the heteroscedastic ethod II (based on Healy's two-stage sapling) is ore econoical than the original HM (which is based on Chatterjee's two-stage sapling), Aer. J. Math. Manageent Sci. 14 (1994a), M. Aoshia, Healy's saple size of two-stage procedure in heteroscedastic siultaneous inference, Co. Statist. Theory Methods 23 (1994b), R. P. Bhargava and M. S. Srivastava, On Tukey's confidence intervals for the contrasts in the eans of the intraclass correlation odel, J. Roy. Statist. Soc. Ser. B 35 (1973),

18 SIMULTANEOUS CONFIDENCE INTERVALS S. K. Chatterjee, On an extension of Stein's two-saple procedure to the ultinoral proble, Calcutta Statist. Assoc. Bull. 8 (1959a), S. K. Chatterjee, Soe further results on the ultinoral extension of Stein's two-saple procedure, Calcutta Statist. Assoc. Bull. 9 (1959b), S. K. Chatterjee, Two-stage and ultistage procedures, in ``Handbook of Sequential Analysis'' (B. K. Ghosh and P. K. Sen, Eds.), Chapter 2, pp. 2145, Dekker, New York, A. Chaturvedi, N. D. Shukla, and P. S. Shukla, Fixed-width confidence intervals for contrasts in the eans, Ann. Inst. Statist. Math. 44 (1992), G. B. Dantzig, On the non-existence of tests of Student's hypothesis having power functions independent of _, Ann. Math. Statist. 11 (1940), M. Ghosh, N. Mukhopadhyay, and P. K. Sen, ``Sequential Estiation,'' Wiley, New York, W. C. Healy, Jr., Two-saple procedure in siultaneous estiation, Ann. Math. Statist. 27 (1956), G. W. Hill and A. W. Davis, Generalized asyptotic expansions of Cornish-Fisher type, Ann. Math. Statist. 39 (1968), H. Hyakutake, Y. Takada, and M. Aoshia, Fixed-size confidence regions for the ultinoral ean in an intraclass correlation odel, Aer. J. Math. Manageent Sci. 15 (1995), N. Mukhopadhyay, A consistent and asyptotically efficient two-stage procedure to construct fixed width confidence intervals for the ean, Metrika 27 (1980), H. Nagao, On fixed width confidence regions for ultivariate noral ean when the covariance atrix has soe structure, Sequential Anal. 15 (1996), M. Siotani, Note on the utilization of the generalized Student ratio in the analysis of variance or dispersion, Ann. Inst. Statist. Math. 9 (1957), M. Siotani, T. Hayakawa, and Y. Fujikoshi, ``Modern Multivariate Statistical Analysis: A Graduate Course and Handbook,'' A. Sci. Press, Syracuse, NY, C. Stein, A two-saple test for a linear hypothesis whose power is independent of the variance, Ann. Math. Statist. 16 (1945),