23
This article was downloaded by: [New York University] On: 23 November 2014, At: 14:28 Publisher: Taylor & Francis Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK Communications in Statistics - Theory and Methods Publication details, including instructions for authors and subscription information: http://www.tandfonline.com/loi/lsta20 Whither jackknifing in stein-rule estimation Pranab Kumar Sen a a Department of Biostatistics , 20Th University of North Carolina , Chapel Hill, NC, 27514 Published online: 02 Nov 2010. To cite this article: Pranab Kumar Sen (1986) Whither jackknifing in stein-rule estimation, Communications in Statistics - Theory and Methods, 15:7, 2245-2266, DOI: 10.1080/03610928608829246 To link to this article: http://dx.doi.org/10.1080/03610928608829246 PLEASE SCROLL DOWN FOR ARTICLE Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and should be independently verified with primary sources of information. Taylor and Francis shall not be liable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of the Content. This article may be used for research, teaching, and private study purposes. Any substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions

Whither jackknifing in stein-rule estimation

  • Upload
    pranab

  • View
    222

  • Download
    1

Embed Size (px)

Citation preview

Page 1: Whither jackknifing in stein-rule estimation

This article was downloaded by: [New York University]On: 23 November 2014, At: 14:28Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registered office:Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Communications in Statistics - Theory andMethodsPublication details, including instructions for authors and subscriptioninformation:http://www.tandfonline.com/loi/lsta20

Whither jackknifing in stein-rule estimationPranab Kumar Sen aa Department of Biostatistics , 20Th University of North Carolina , Chapel Hill,NC, 27514Published online: 02 Nov 2010.

To cite this article: Pranab Kumar Sen (1986) Whither jackknifing in stein-rule estimation, Communications inStatistics - Theory and Methods, 15:7, 2245-2266, DOI: 10.1080/03610928608829246

To link to this article: http://dx.doi.org/10.1080/03610928608829246

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”)contained in the publications on our platform. However, Taylor & Francis, our agents, and ourlicensors make no representations or warranties whatsoever as to the accuracy, completeness, orsuitability for any purpose of the Content. Any opinions and views expressed in this publication arethe opinions and views of the authors, and are not the views of or endorsed by Taylor & Francis.The accuracy of the Content should not be relied upon and should be independently verified withprimary sources of information. Taylor and Francis shall not be liable for any losses, actions, claims,proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or howsoevercaused arising directly or indirectly in connection with, in relation to or arising out of the use of theContent.

This article may be used for research, teaching, and private study purposes. Any substantialor systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, ordistribution in any form to anyone is expressly forbidden. Terms & Conditions of access and use canbe found at http://www.tandfonline.com/page/terms-and-conditions

Page 2: Whither jackknifing in stein-rule estimation

COMMUN. STATIST.-THEOR. METH., 1 5 ( 7 ) , 2 2 4 5 - 2 2 6 6 ( 1 9 8 6 )

NHITHEE JACKKIdIFIMG I N STEIN-RULE ESTIMATION ?

Pranab Kumar Sen

Department o f B i o s t a t i s t i c s 201H U n i v e r s i t y o f l i o r t h C a r o l i n a

Chapel H i l l , i l C 27514

Key Words and Phrases : Asymptotic risk ; bias ; bias-reduction ; inahissibiZity ; jackknifing ; jackknifed estimator of dispersion matrices ; minimaxity ; Pitman-alternatives ; risk-estimation ; shrinkage estimators.

ABSTRACT

I n mu1 t i -pa ramete r ( mu1 t i v a r i a t e ) e s t i m a t i o n , t h e S t e i n r u l e

p rov ides minimax and admiss ib le e s t i m a t o r s , compromising g e n e r a l l y

on t h e i r unbiasedness. On the o t h e r hand, t h e p r i m a r y aim o f j a c k -

k n i f i n g i s t o reduce t h e b i a s o f an e s t i m a t o r ( w i t h o u t n e c e s s a r i l y

compromising on i t s e f f i c a c y ) , and, a t t h e same t ime, j a c k k n i f i n g

p rov ides an e s t i m a t o r o f t h e sampling v a r i a n c e o f t h e e s t i m a t o r as

w e l l . I n shr inkage e s t i m a t i o n ( where m i n i m i z a t i o n o f a s u i t a b l y

d e f i n e d r i s k f u n c t i o n i s t h e b a s i c goal ) , one may wonder how f a r

t h e b i a s - r e d u c t i o n o b j e c t i v e o f j a c k k n i f i n g i n c o r p o r a t e s t h e dual

o b j e c t i v e o f n i i n imax i t y ( o r a d m i s s i b i l i t y ) and e s t i m a t i n g t h e

r i s k o f t h e e s t i m a t o r ? A c r i t i c a l a p p r a i s a l o f t h i s b a s i c r o l e o f

j a c k k n i f i n g i n shr inkage e s t i m a t i o n i s nade here. R e s t r i c t e d , semi-

r e s t r i c t e d and t h e usual ve rs ions o f j a c k k n i f e d sh r inkage es t imates

a re considered and t h e i r performance c h a r a c t e r i s t i c s a r e s t u d i e d . It i s shown t h a t f o r Pitnian-type ( l o c a l ) a l t e r n a t i v e s , u s u a l l y ,

j a c k k n i f i n g f a i l s t o p r o v i d e a c o n s i s t e n t e s t i m a t o r o f t h e ( asymp-

t o t i c ) r i s k o f t h e shr inkage e s t i m a t o r , and a degenerate asympto-

t i c s i t u a t i o n a r i s e s f o r t h e usual f i x e d a l t e r n a t i v e case,

Copyright O 1986 by Marcel Dekker, Inc

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 3: Whither jackknifing in stein-rule estimation

SEN

1. INTRODUCTION

L e t X_, ,, .. ,X,, be n (> - 1 ) independent and i d e n t i c a l l y d i s t r i -

buted random v e c t o r s (i. i.d.r.v.) w i t h a p ( - > 1 ) - v a r i a t e d i s t r i b u -

t i o n f u n c t i o n Cd.f.1 F hav ing a f i n i t e mean v e c t o r -" 0 and a p o s i t i v e

d e f i n i t e (p.d.) d i s p e r s i o n m a t r i x X , where b o t h 0 and C a r e unkn- .d -" .d

own. When F i s assumed t o be mu l t i -no rma l , t h e maximum 1 i k e l i hood

e s t i m a t o r (m.1 .e.) o f i s t h e sample mean v e c t o r Tn = nn1yn i = l - i X a

and t h i s possesses some op t ima l p r o p e r t i e s too, However, f o r p 3,

S t e i n (1956) showed t h a t Tn i s n o t admissible, and es t in ia to rs o f e -"

dominat ing over Tn , i n r i s k , known as t h e Stein-ruZe estimators , have been considered by James and S t e i n (1961) and a hos t o f o t h e r

workers. A v e r y d e t a i l e d account o f t h e S t e i n - r u l e e s t i m a t i o n t h e -

o r y f o r mu1 t i - n o r m a l F i s g iven i n Berger (1980) as w e l l as i n some

o t h e r contemporary tex tbooks on m u l t i v a r i a t e s t a t i s t i c a l a n a l y s i s .

Though such S t e i n - r u l e e s t i m a t o r s have been s t u d i e d f o r some

s p e c i f i c non-normal F, i n genera l , f o r an a r b i t r a r y F a t h e c o n s t r -

u c t i o n o f such S t e i n - r u l e e s t i m a t o r s and v e r f i c a t i o n o f t h e i r ( ex-

a c t ) dominance o v e r t h e co r respond ing m . 1 . e . ' ~ a r e y e t t o be made

i n f u l l g e n e r a l i t y . Considerable d i f f i c u l t i e s a r i s e i n t h e s tudy

o f t h e (exac t ) d i s t r i b u t i o n t h e o r y o f such e s t i m a t o r s ( f o r f i n i t e

n ), and these, i n t u r n , make i t d i f f i c u l t t o e s t a b l i s h t h e a n t i c i -

pated dominance o f t h e S t e i n - r u l e es t imato rs . On t h e t o p o f t h a t , f o r normal as w e l l as non-normal d i s t r i b u t i o n s , t h i s dominance o f

t h e S t e i n - r u l e e s t i m a t o r ho lds o n l y i n a s h r i n k i n g neighbourhood

o f t h e p i v o t ( i n t h e sense t h a t as t h e sample s i z e n increases,

f o r any f i x e d d e v i a t i o n o f t h e t r u e parameter p o i n t f rom t h e p i v o t ,

t h e r a t i o o f t h e r i s k s of t h e S t e i n - r u l e e s t i m a t o r and t h e c o r r e s -

ponding m.1.e. converges t o one, and t h e a n t i c i p a t e d dominance may

h o l d o n l y when t h e t r u e parameter 0 l i e s i n a b a l l w i t h t h e p i v o t

as t h e c e n t r e and hav ing a r a d i u s " o f t h e o r d e r n-' 1. Never the less,

f o r F n o t n e c e s s a r i l y mu1 t i -no rma l and f o r genera l estimable parum-

eters , under a p p r o p r i a t e moment-condit ions, t h e asympto t i c dominance

o f a S t e i n - r u l e e s t i m a t o r has been s t u d i e d i n a more genera l setup

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 4: Whither jackknifing in stein-rule estimation

WHITHER JACKKNIFING IN STEIN-RULE ESTIMATION? 2247

C of U-s t a t i s t i c s 1 by Sen (19841, where the usual Pitman-type (i.

e. , l o c a l ) a l t e r n a t i v e s C t o the p ivo t ) have been adapted t o pro-

vide a meaningful and c l e a r p ic ture of t h i s asymptotic dominance.

A very s imi la r p ic ture holds f o r the Ste in-rule est imators r e l a t ed to the m.1 . e O 1 s f o r an a r b i t r a r y F I v iz . , Sen(1906 )I as well as f o r the usual l i n e a r est imators i n l i n e a r models 1 v i z , , Saleh and

Sen (l96G)l.

The theory of Ste in-rule est imation has a l so been considered i n some nonparametric setups ( r e l a t ing t o possible non-linear e s t - imators ) by Sen and Saleh (1985a,b) and Saleh and Sen (1985). The

essent ia l c h a r a c t e r i s t i c s of the Stein-rule est imation theory a r e i n t a c t i n t h i s asymptotic nonparametric setup too, In nonparamet-

r i c inference , jackknifing plays an important ro l e , Often, t o re- duce the bias of an est imator o r t o est imate i t s sampling variance, jackknifing i s very successful ly employed. This brings us t o the natural question : Mhat i s the ro le of jackknifing in the Ste in-

ru l e est imation theory ? Our primary object ive i s t o address t h i s

issue.

A S te in-rule est imator i s genera l ly biased. This i s mainly due

t o the f a c t t h a t in the shrinkage f a c t o r , one incorporates a t e s t s t a t i s t i c (on the adequacy of the p ivot) which introduces non-lin-

e a r i t y , and t h i s in turn introduces bias. In a non-normal o r non- parametric setup, of ten , t h i s bias i s too complicated t o be evalu- ated i n an exact form . For an a r b i t r a r y ( and,possibly, biased )

es t imator , jackknifing not only reduces the order of magnitude of the b ias , but a l so provides an est imator of the sampling variance of the est imator ( o r the corresponding r i s k when i t i s convenient- l y defined i n t e rns of the sampling dispersion matrix of the estim- a t o r 1. An inherent reversed martingale s tmcture of jackknifing C explored sys temat ica l ly by Sen (1977)) provides a c l e a r p ic ture of the mechanism of t h i s dual ro l e of bias-reduction and estimation

of the r i sk . In the t r ad i t iona l case ( without having any pivot in

mind), f o r the mul t iva r i a t e location problem, the r i s k of the m.1.e.

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 5: Whither jackknifing in stein-rule estimation

i s t r a n s i a t i o n - i n v a r i a n t , and hence, i s a c o n s t a n t [depending on

t h e o t h e r nu isance parameters a s s o c i a t e d w i t h t h e u n d e r l y i n g d. f , ) .

T h i s cons tan t r i s k can be v e r y e f f e c t i v e l y e s t i m a t e d by t h e usua l

j a c k k n i f i n g technique, However, t h e p i c t u r e may be q u i t e d i f f e r e n t

i n t h e case of sh r inkage e s t i m a t i o n , as w i l l be d e a l t w i t h here.

F o r any f i x e d a1 t e r n a t i v e (i .e., t h e t r u e parameter be ing d i f f e r -

e n t f rom t h e p i v o t ) , a s y m p t o t i c a l l y , a S t e i n - r u l e e s t i m a t o r and i t s

j a c k k n i f e d v e r s i o n a r e b o t h e q u i v a l e n t , and n e i t h e r dominates over

t h e conven t iona l one. I n v iew o f t h e i n e f f e c t i v e n e s s o f t h e S t e i n -

r u l e e s t i m a t i o n t h e o r y f o r f i x e d - d l t e r n a t i v e asympto t i cs , as has

been nlent ioned e a r l i e r , t h i s o v e r a l l equ iva lence o f j a c k k n i f e d and

t h e usual ve rs ions o f t h e S t e i n - r u l e e s t i m a t o r s i s o f no p r a c t i c a l

s i g n i f i c a n c e . however, f o r P i tman-type ( l o c a l ) a1 t e r n a t i v e s , t h e

s i t u a t i o n i s q u i t e d i f f e r e n t . I n such a case, t h e conven t iona l es-

t i n a t o r Tn , t h e r e l a t e d S t e i n - r u l e e s t i m a t o r and i t s p l a u s i b l e Ja-

c k k n i f e d v e r s i o n s a r e g e n e r a l l y n o t a s y m p t o t i c a l l y e q u i v a l e n t , and

t h e i r r e l a t i v e r i s k p i c t u r e s c l e a r l y convey some meaningfu l asymp-

t o t i c dominance r e l a t i o n s , and these w i l l be s t u d i e d here.

For s i m p l i c i t y o f p r e s e n t a t i o n s , we c o n s i d e r here t h e case o f

t h e m u l t i v a r i a t e l o c a t i o n problem a l t h o u g h a s i m i l a r t h e o r y works

o u t w e l l f o r genera l e s t i m a b l e parameters. S ince t h e S t e i n - r u l e es-

t i m a t o r , g e n e r a l l y , i n v o l v e s a t e s t s t a t i s t i c ( f o r t h e a p p r o p r i a -

teness o f t h e p i v o t ) i n a d d i t i o n t o t h e sample mean v e c t o r and co-

va r iance m a t r i x , i n t h i s setup, j a c k k n i f i n g may o n l y be a p p l i e d t o

t h e mean v e c t o r , o r t o t h e t e s t s t a t i s t i c and mean vec to r , o r t o

a l l t h e t h r e e s e t s o f random elements. These r e l a t e t o t h e restric-

!., I, ::c'rni-r~cst-rictecl and unrestricted v e r s i o n s o f j a c k k n i f e d s h r i n -

kage es t imato rs . These v e r s i o n s a r e a l l f o r m a l l y i n t r o d u c e d i n Sec-

t i o n 2, S e c t i o n 3 dea ls w i t h t h e i r asympto t i c r i s k p r o p e r t i e s f o r

f i x e d a l t e r n a t i v e s , and t h e genera l equ iva lence r e s u l t s a r e presen-

t e d the re . The main r e s u l t s on t h e i r asympto t i c r e l a t i v e r i s k p r o -

p e r t i e s f o r P i tman- type ( l o c a l ) a l t e r n a t i v e s a r e p resen ted i n Sec-

ti011 4 . it1 t h i s con tex t , t h e n o t i o n o f usymptotic 2istribv.tionaZ

r'i::ii i'A:)l?i i s i n c o r p o r a t e d i n t h e d e f i n i t i o n and e s t i m a t i o n o f t h e

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 6: Whither jackknifing in stein-rule estimation

WHITHER JACKKNIFING IN STEIN-RULE ESTIMATION? 2249

of asympto t i c r i s k , and t h e r o l e of j a c k k n i f i n g i n p r o v i d i n g con-

s i s t e n t e s t i m a t o r s o f t h i s asympto t i c r i s k i s c r i t i c a l l y examined.

I t i s shown t h a t i n t h i s asympto t i c setup, g e n e r a l l y , j a c k k n i f i n g

f a i l s t o p r o v i d e a c o n s i s t e n t e s t i m a t o r o f t h e ADR of a S t e i n - r u l e

e s t i m a t o r of 0 . T h i s f i n d i n g a lone niay r a i s e t h e genera l i s s u e .. o f implement ing j a c k k n i f i n g i n S t e i n - r u l e e s t i m a t i o n , and t h e con-

c l u d i n g s e c t i o n i s devoted t o some genera l d i s c u s s i o n s ( i n t h i s

v e i n ) on t h e r o l e o f j a c k k n i f i n g i n sh r inkage e s t i m a t i o n .

2. JACKKNIFED STEIN-RULE ESTIEIATORS OF 0 .. For an e s t i n ~ a t o r Ln = I(;, ,'" ,En) o f 8 , c o n s i d e r a quadratic ..

Loss function

where Q i s some g i v e n p.d. m a t r i x . Genera l l y , Tn may be chosen as - .. a t r a n s l a t i o n - i n v a r i a n t e s t i m a t o r , and hence, t h e c h o i c e o f 4 may

be made independent l y o f any i n f o r m a t i o n one niay have on 0 ( such

as t h e p i v o t which w i l l be i n t r o d u c e d l a t e r on) . The r i sk ( i .e . ,

t h e expected l o s s ) o f 3 , i s t h e n g i v e n by

P,,( 1, 8 = E & ( 1,; 8 ) = Tr(QZn) , ( 2 . 2 ) .. where En = nEg{(Tn - !)(I, - 0 ) l i i s t h e mean-product moment m a t r i x

o f nk ( I n - 0 ) : I n a s t e i n - r u l e e s t i m a t i o n problem, we a r e g i v e n a

p i v o t f o , and shr inkage o f in i s made towards !o . Because o f

t h e t r a n s l a t i o n - i n v a r i a n c e o f t h e sample mean v e c t o r Tn , we may

take, w i t h o u t any l o s s o f g e n e r a l i t y , f o = 0. Then, f o l l o w i n g t h e - genera l p r e s c r i p t i o n of Berger e t a l . (1977) ( though n o t necessa-

r i l y assuming t h a t t h e u n d e r l y i n g d . f . F i s m u l t i - n o r m a l ) , we may

c o n s i d e r t h e f o l l o w i n g S t e i n - r u l e e s t i m a t o r :

where I i s t h e i d e n t i t y m a t r i x of o r d e r p , cn i s a p o s i t i v e number

( converg ing t o a l i m i t c : 0 < c < 2 ( p - 2 ) ) , p i s assumed t o be p o t 2 -1 -1-

l e s s than 3, Tn = n!:,SQ Xn) i s t h e t e s t s t a t i s t i c ( f o r t h e n u l l

hypo thes is Ho: Bo= 0 i.e., t h e adequacy o f t h e p i v o t ! , dn = cl; (QS ) , w p ..-n

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 7: Whither jackknifing in stein-rule estimation

2250 SEN

i s t he smal les t c h a r a c t e r i s t i c root of QS and the sample covari- -1 n W-n ,

ance matrix Sn = n XiG1 (=Xi - F n ) tzi - Tn) i s an unbiased estim-

a t o r of 5 . For mu1 ti-normal F and s u i t a b l e I cn ) , the minimaxity " S of en has been es tabl ished i n Cerger e t a l . (1977). Irl our study,

we do not assume t h a t F i s necessar i ly multi-normal. Under sui ta t ; le " S moment-conditions on F , t he dominance of !n over Tn ( i n an asym-

p to t i c se tup) follows from the r e s u l t s of Sen(1984) on more gener- al i l - s t a t i s t i c s . Ele rewri te (2.3) a s

"S -2 -1 -1 = ( : - T n bn )Tn ; ;n = cndn9 :n a -n (2.4)

-2 and note t h a t the th ree s e t s of random elements (v iz , , I n , fin and

?n "S a r e a l l based on z1 ,..., E n and appear in el, . !$low, j a c k k n i f i n ~

m y be used on a pa r t of these randor!! elements C o r a l l ) . Keeping t h i s in mind, we nay consider the following versions of jackknifed Ste in-rule es t imators of 8 .

*

( I ) Restricted jackknifed estimator. Since Fn i s the most re- levant pa r t f o r the est imation of 8 , we nay incorporate jackknif- ... ing only on t h i s p a r t , and , thereby, obtain the following jackkni-

fed est imator corresponding t o the S te in - ru le es t imator in (2.4) :

;SRJ = n-l z v S R -n 1=1 'n,i , (2.5)

where ?R. = ^S "SR

~ n . 1 n !n - ( n - l ) ~ n - i j i ) , i = 1 ,..., n, and

@R - -2 nyn -; . + n - ~ ( i ) - Q - Tn !n)Tn-l(i) ; ?!n-l(i) = .mj 9

(2-6 )

f o r i = l , , . . , n . I t i s easy t o ve r i fy t h a t

eSRJ = 2 , f o r every n 2 2. - n - n Thus, r e s t r i c t e d jackknifing, being based on l i n e a r opera t ions , do-

es not induce any change over the o r ig ina l Ste in-rule es t imator in (2.4).

(11 ) Semi-restricted jackknifed estimator, tiere, jackknifing applied t o both the c l a s s i c a l es t imator Tn and t h e t e s t s t a t i s t i c

2 Pfi ( b u t not t o the o ther f a c t o r A,) , so t h a t D

ownl

oade

d by

[N

ew Y

ork

Uni

vers

ity]

at 1

4:28

23

Nov

embe

r 20

14

Page 8: Whither jackknifing in stein-rule estimation

WHITHER JACKKNIFING IN STEIN-RULE ESTIMATION?

"SS - "S ^SS where - - mn - (n-1 N n - l ( i ) , i = l , , , , , n , and - n , l

where T:-, ( i ) i s based on (X1 ,. . . ,X ,;i+l ,* * . ,:n) , f o r i = 1 ,... , -i-1 "S n . This i s a nonlinear adapta t ion, and, i n genera l , O n and e:SJ a re

not the same. (111) Unres t r i c t ed jackkni fed e s t ima tor . In t h i s case , the

jackknifing i s incorporated i n t he usual way o f recomputing the

(shrinkage) es t imators from each of the n sub-samples of s i z e n -1,

so t h a t

S U "S "S U where !n,i = n$, - Cn-1 ) in - l ( i ) , i= l , . . . ,n, and

;su - 2 - - ) - l i n l i ( i 3 1 (2.11)

with the A,,,-l ( i , defined on the subset (El ,. . . ,Xi-l , X i + l ,.. . ,tn 1, . . f o r i = l , , . . , n . Again we have a n o n l i ~ e a r adapta t ion, and hence, in

'v genera l , o3 and gSUJ a r e not the same.

-I1 -I I

Side by s i d e , we introduce the jackkni fed d i s p e r s i o n ma t r i x

"SRJ - -1 n "SR "SRJ "SR ;SRJ) I - n - n !n,i - B n ) ( !n,i - In a

^SSJ "SS - SSSJ)' iSS - e ) ( on = i -n - n 9

lCn "SU - gSUJ SU GSUJ)' i= l ( !n , i -n I D n - .. n

9ur main i n t e r e s t l i e s in the study of the a s p p t o t i c p rope r t i e s of tliese jacknifed Ste in-rule es t imators and t h e i r est imated disper-

sion matrices. The ro le of these l a t t e r es t imators i s examined in

the context of the est imation of the ADF! of the proposed jackknifed versions of the Ste in-rule es t imator i n (2.41.

3. FIXED ALTERNATIYE ASYMPTOTICS ,FOR JACKKNIFLNG

i t f o l l o ~ s from Sen (1984) t h a t f o r any (fixed) e # 0 , t he es- OS ... -

t i~r ia tors En and !n a r e asymptotically r i sk-equivalent in the sense

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 9: Whither jackknifing in stein-rule estimation

2 2 5 2 SEN

2 tlie convergence of the f i r s t order negative moment of Tn ensures

t h a t the r a t i o of the two r i s k s asymptotically converges t o l , i . e . ,

E~T; ' B + 1 + 0 as n j m

^S => limn., {pn[e -., , e , ) / pnC? , !)I = 1 . (3 .1)

,iote t h a t t h i s negative moment condit ion i s needed not f o r the r i s k -

cor,il~:~tatiori of the c l a s s i ca l es t imator X , but f o r t9e s !~r inkage -. n one . For o ther modes of asymptotic equivalence of these two e s t i -

iliators [ such as almost sure [a.s.) equivalence o r s tocl !as t ic one ) ,

t h i s condition nay be dispensed with. In f a c t , when tlie underlying

d . f . F has f i n i t e 4th order moments, using the f a c t t h a t f o r 8 # 0, - , -2 n converges a.s. t o 0 (as n -+ a) , we obta in from (2.3) and some

elementary inequal i ty t h a t

6y v i r tue of (2 .7 ) , we conclude t h a t (3 .1) and (3.2) a l so hold f o r

ii:RJI . llote t h a t f o r the computation of the RDR r e s u l t s ( as will

be made in the next s e c t i o n ) , we do not need t h e negative moment-

conuit ion in (2 .1) .

ilext, we note t h a t Fn = n - ' x n - X i=l-.n-1 ( i ) . Hence, by (2 .3) , (2 .8)

and (2.9), we have

Let c,, be the sigma-field generated by the unordered col l ec t ion

,... ,/; ) and ;; , j > 1 , f o r n > 1 , so t h a t en i s rlonotone -11 ,n+j - - -

nonincreasing. Also, l e t pn = p( " S . k > n) be the sigma-field /'kY , k Y - gerierateti by K S - k > n , f o r n > 2, so t h a t F, ccn and F, i s ,kY,I<' - - a l so nionotone nonincreasing. Then

Thus, we have

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 10: Whither jackknifing in stein-rule estimation

WHITHER J A C K K N I F I N G I N STEIN-RULE ESTIMATION?

2 -1 - n 2 - 2 lx (n-l)T, {An(n ZiZl{l - TnTn-l(i) S n - l ( i ) ) 1

= (n-

+ A,E - Now, f o r every 8 # 0, - *

( n- 1

while, independently of 8, -

where A = ch (41). Further, note t h a t P - -

2 T n = T ( y n , S n ) , n ? 2 (3 .8 )

where Xn and Sn - a r e both U - s t a t i s t i c s and T( . ) i s a smooth funct ion.

Hence, proceeding as in Sen (1977) [See h i s (3 .20) ] , i t follows

f o r every E > 0 (and we l e t O<E<!). Simi la r ly ,

2 -2 - - ! c ) = O ( n - 1 + ~ ) E{(1 - T , , T n 4 ) ( 5 , 4 - E n ) , ,, a . s . , a s n + -.

(3.10)

Thus, the r i g h t hand s i d e ( r h s ) of (3 .5) i s 0(n-"') a . s . , a s

n - -, and t h i s ensures t h a t f o r every ( f ixed) O b f 0, - nil 6isJ - $ % I I + 0 a . s . , a s n -+ (3.11)

Final ly , note t h a t d n = ch ( Q - I S - ' ) i s a ismoath' (nonnegative ) p - -n

function of ( t h e s tochas t i c matrix ) Sn , so t h a t by appeal t o Sen

(1977), we have

n l a x l < i < n I 1 % - I l i ) - A n I = 0(n-'2 ) a , s . , a s n -. , (3.12) --

a n d , t he re fo re , proceeding as in (3.4) through (3.10) and using

(2.4), (2.13) and (2 ,11) , we obtain t h a t f o r every (fixed ) R f 0 , - - 1 ^SUJ "S n Z I / - O n / / -. 0 a . s . , as n -. . (3.13)

*

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 11: Whither jackknifing in stein-rule estimation

2 2 5 4 SEN

I t may be noted t h a t f o r the asymptotic a.s. equivalence res-

u l t s in (3.2), (3.11) and (3.13), t he a . s , convergence r e s u l t i n

(3.6) plays a basic role. I f f3 = 0 ( o r i s c lose t o 0 in the usual - w

Pitman-sense ), then the r i g h t hand s ide of (3.6) blows o f f , and , as a r e s u l t , (3.11) and (3.13) may not hold. This c a l l s f o r the

need to study these asymptotic r e s u l t s from a d i f f e r e n t perspect i -

ve when the pivot and the t r u e parameter point a r e c lose t o each

o the r , and we shal l take up t h i s i ssue in the next sec t ion,

From (3 .2) , (3-11) and (3.13), we note t h a t i f the t rue para-

rtieler (point ) 0 d i f f e r s from the p ivot (here , 0) by any fixed -. amount , jackknifing may not lead t o any s i g n i f i c a n t change t o the

" s usual Ste in-rule es t imator !n . Further , i n the same setup , i: - and 1: a r e asymptotically r i sk-equivalent , so t h a t shrinkage does -n not lead toany asylvptotic improvement over the c l a s s i ca l e s t i ~ a t o r

SRJ too, Noting t h a t a l l the th ree d ispers ion matrix es t imators V_., , vSSJ and !.':UJ involve (jackknifed) U - s t a t i s t i c s , we may proceed as , n in Sen (1977) and claim t h a t f o r any ( f ixed) Fi # 0 , as n -) m , - -

SUJ vSRJ i , vSSJ L and Y n -n -n 5 9 (3.14)

so t h a t T r ( Q l n ) converges t o Tr(Qz), in p robab i l i ty ( a s n -t ) in w-

each of the three cases of jackknifing considered e a r l i e r , Fur ther ,

note t h a t even under the f i n i t e n e s s of the second order moments of

F, S -f X a , s . (as n -t ) , and hence, Tr(4(ln) + T r ( 0 ~ ) a , s . , a s -11 - w..,

n + a . tlence, t he re i s no gain in the asymptotic r i sk-eff ic iency

~ u e t o the usual shrinkage estimation o r jackknifing on the top of

t h a t , and the r i s k est imation in the jackknifed s i t u a t i o n s may de- mand more s t r inqen t moment condit ions than in the usual case. Thus,

thcrc sceriis t o be very l i t t l e ground in aCvocating jackknifing in si~rinkage e s t i ~ ~ l a t i o n in the large sample case when the t r u e param-

e t e r 0 i s d i f f e r e n t from the pivot. This f ixed-a l t e rna t ive asymp- u

t o t l c p ic ture a l so provides f i n i t e sample j u s t i f i c a t i o n s when the pardmeter point O i s away from the pivot . !+owever, f o r e c lose t o

.., *

tre pivot , we may derive more meaningful r e s u l t s through consider-

& t i o m of Pitman-type a l t e r n a t i v e s , a s will be done i n Section 4.

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 12: Whither jackknifing in stein-rule estimation

WHITHER JACKKNIFING IN STEIN-RULE ESTIMATION?

4. ASYNPTOTICS FOR PITEAN-TYPE ALTERNATIVES IKnl

For Pitman-type (local ) a1 t e r n a t i v e s , t he dominance of t h e Ste in-rule es t imator over the c l a s s i c a l one follows d i r e c t l y from

the general r e s u l t s on shrinkage U-s t a t i s t i c s in Sen (1984). I t i s

therefore of i n t e r e s t t o study the performance c h a r a c t e r i s t i c s of the jackknifed versions under the same asymptotic se tup, and, the-

reby, t o examine the ro le of jackknifing in Stein-rule est imation

theory ( t h i s cons t i tu t e s the main theme of t h i s s tudy) . We conceive of a t r i angu la r a r r ay ; n - > 11) of

row-wise i . i .d.r.v. ' s , where f o r each n( 2 1 ) , X n i has the d , f .

F(? - nm4h ) , f o r some ( f ixed) h E E ~ , p 2 3 , and F has the mean

vector O and a p.d, d ispers ion matrix C (both unknown). We denote .. .. t h i s sequence ( of Pitman-type a l t e r n a t i v e s ) by {Knl . I t may be noted t h a t the X n i - n-'h , i 2 1 , a r e i . i ,d.r .v. with mean 0 and - dispersion matrix ; , and hence, by (2.2), under Kn , the r i s k of X i s given by l'r(QZ), f o r every n 1. We denote,for l a t e r use, -n -.. t h i s constant r i s k by

-9 P p(h) = Tr(QZ) ..- ( = limn, pn(Tn;n 'A ) ) , - X E E . (4.1)

For the computation of the asymptotic r i s k of the Ste in-rule e s t i - "S mator sn in (2 .3) , we may note t h a t under IKn} , ii2 converges in

d i s t r i b u t i o n to X-2 where x2 stands f o r a pos i t ive r,v. hav- P , A P ? A

ing the noncentral chi-square d i s t r i b u t i o n with p degrees of f ree-

dom (DF) and noncentral i t y parameter A = ?'C-'h - - , while dn conver- ges a.s. t o 6 = ch (QC); however, these a r e not enough t o ensure

P -- the ~i~ornent-convergence r e s u l t needed to show t h a t (2.2) converges ( under {I(,} ) t o a l i m i t as n increases inde f in i t e ly . I f we assume t h a t under {Kn) ,

d n ~ i 2 converges in L1-norm t o -2 ' - X ~ , A ' (4.2)

then, by appeal to the general r e s u l t s i n Sen(1984) I s e e a l so Sen

and Saleh (1985, Sec. 3)], we obtain t h a t under lKnl , whenever F possesses f i n i t e moments of order r , f o r some r > 2,

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 13: Whither jackknifing in stein-rule estimation

SEN

= Tr(Qr) ..,- - 2 c ( p - 2 ) 6 ~ ( ~ ; ? ~ ) + 2 2

c a A*E( $4,A ) + T ~ P - -1 - -1 )E(x~:~,~) I , (4.3)

where A* = X ' C -1 -1 -1

* - C A and c = l imn- cn ( i t i s assumed t h a t

t h i s l i m i t e x i s t s and c E (0, 2(p-2) ) ) . Even w i t h o u t t h e L1-conv-

ergence r e s u l t i n (4.2), one may c o n s i d e r t h e asympto t i c d i i t r i b u -

t i o n o f n"( i: - n'% ) (under i K n I ) and compute t h e r i s k f rom

t h i s asympto t i c d i s t r i b u t i o n . T h i s i s termed t h e ADR o f t h e S t e i n - "S r u l e e s t i m a t o r fin . Then, as i n Sen (1984) w i t h a d a p t a t i o n s f rom

Saleh and Sen (1985), we conclude t h a t (4.3) ho lds f o r t h e ADR o f

;S , w i t h o u t r e q u i r i n g (4.2); f i n i t e n e s s o f t h e moments o f o r d e r r -. n o f F, f o r some r > 2 s u f f i c e s f o r t h i s . Now, t h e r i g h t hand s i d e

o f (4.3) i s l e s s t h a n T r ( Q C ) = p(X) f o r every f i n i t e A E E~ and c: -.., O < c < 2(p-2), and as A moves away from t h e o r i g i n ( i.e,, A o r

A* i nc reases ) , pS(X) converges t o p (A) .

Given t h i s asympto t i c p i c t u r e , i t i s o f i n t e r e s t t o s tudy t h e

e f f e c t o f j a c k k n i f i n g 011 t h e ADR ( o r t h e asympto t i c r i s k i t s e l f )

o f t h e S t e i n - r u l e e s t i m a t o r . A lso, t h e r e i s a more i m p o r t a n t ques- S t i o n : Can p (A) ( o r t h e o t h e r r i s k ) be es t imated c o n s i s t e n t l y by -

any o f t h e t h r e e j a c k k n i f i n g procedures ?

F i r s t , c o n s i d e r $,RJ. By ( 2 . 7 ) , we conclude t h a t (4.3) h o l d s

f o r t h i s r e s t r i c t e d j a c k k n i f e d v e r s i o n as w e l l . F u r t h e r , by (2 .5 ) ,

( 2 .6 ) and (2.12) , we have

Note t h a t under {Kn I ,

There fo re , by (4.4) and (4 .5 )

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 14: Whither jackknifing in stein-rule estimation

WHITHER JACKKNIFING IN STEIN-RULE ESTIMATION? 2257

and, as a r e s u l t ,

and t h e r h s o f (4 .7 ) i s a nondegenerate r . v . f o r every A: 11 < w.

Comparing (4 .3 ) and (4 .7 ) , we conclude t h a t whereas t h e r e s t r i c t e d - S j a c k k n i f e d e s t i m a t o r !:RJ agrees w i t h ", t h e a l l i e d j a c k k n i f e d

d i s p e r s i o n m a t r i x , v : ~ ~ , f a i l s t o p r o v i d e a c o n s i s t e n t e s t i m a t o r S o f t h e asympto t i c r i s k ( o r ADR) p ( A ) = pSRJ(h) . I n pass ing , we

may remark t h a t f o r Q = z-I, (6 = 1); we have ;he r h s o f (4 .7 )

equal t o p ( l - - 2 T2, and t h i s may even exceed p whenever - 2 C X ~ ,A

ck . 2 (wh ich has a p o s i t i v e p r o b a b i l i t y ) . Reca l l t h a t i n t h i s P,*

case o f r e s t r i c t e d j a c k k n i f i n g , we have a l i n e a r a d a p t a t i o n , so

t h a t t h e p i c t u r e w i t h t h e b i a s of t h e S t e i n - r u l e e s t i m a t o r remains

t h e same, b u t t h e o t h e r dual purpose o f e s t i m a t i n g t h e asympto t i c

r i s k ( o r t h e ADK) i s n o t served . Hence, t h e r e seems t o be n o t

n~uch o f impor tance i n t h e use o f t h i s r e s t r i c t e d j a c k k n i f i n g i n t h e

S t e i n - r u l e e s t i m a t i o n theory .

Consider n e x t t h e case o f a:SJ. We r e w r i t e ( 3 .3 ) as

t h a t

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 15: Whither jackknifing in stein-rule estimation

2258 SEN

A t t h i s stage, we n o t e t h a t as n +

I Itn\ I = 0 (1 ) a.s.,

max l r i r n 1 l!i - !nl

(where we assume t h e f i n i t e n e s s o f 4 t h o r d e r moments o f F ) , and

under i K n ) (as w e l l as Ho: h - = O), - n f l I* , - 1 = 0 1 There fo re , -P

we o b t a i n ( t h r o u g h some s tandard s teps) t h a t

2 - 2 - - x 1 I =,,I = Z T ; ~ % + Ep(n-'t). E{ U i - 1 - Tn ) Un -n (4.15)

B Y ( 4 . 8 ) , (4.14) and ( 4 . 1 5 ) ~ we o b t a i n t h a t under {Kn l and t h e ass-

enled f i n i t e n e s s o f 4 th o r d e r moments o f F Censuring (4,1211, f o r

t n e s e n i i - r e s t r i c t e d j a c k k n i f i n g

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 16: Whither jackknifing in stein-rule estimation

WHITHER JACKKNIFING IN STEIN-RULE ESTIMATION?

"S TIIUS, i:SJ and !n a r e n o t a s y m p t o t i c a l l y e q u i v a l e n t under it$,},

though a s y m p t o t i c a l l y i:SJ a c t s l i k e a sh r inkage e s t i m a t o r ( w i t h

~ ~ 1 ; ~ b e i n g r e p l a c e d by ( p - 2 ) b n ~ i 4 ). T h i s suggests t h a t t h e usual

tecnn iques f o r e s t a b l i s h i n g t h e (asympto t i c ) dominance o f s h r i n k -

age e s t i m a t o r s o v e r t h e c l a s s i c a l ones may be used here t o d e r i v e

a p a r a l l e l r e s u l t f o r g;SJ. F o r t n e asympto t i c r i s k , we need t o

s t r e n g t h e n (4.2) t o

d n ~ i 4 converges (under Kn) i n L1-norm t o 6x-4 P,A '

(4.17)

Proceeding then as i n Sen (1984), we have under { K n l ,

where c = limn, cn , 6 = ch p (QL), -.- W - -. X ( w , I ) , u = , T - 1 2 ~ and A - - - .. -. = )>-'LQ-l y%

- - - - . F o r t h e p a r a l l e l ADR r e s u l t , we do n o t need (4,17);

f i n i t e n e s s o f t h e 4 t h o r d e r moments o f F ( e n s u r i n g (4 .12) ) and t h e

asympto t i c d i s t r i b u t i o n a l r e s u l t s on t h e component s t a t i s t i c s l e a d

us t o t h e same express ion ,

LfJe i n t e n d t o show t h a t f o r s u i t a b l e range o f c and o f p , t h e

rhs o f (4.18) i s s m a l l e r t h a n Tr(QC) ,ensur ing t h e asympto t i c domi- -- nance o f g:SJ o v e r Tn. Towards t h i s , we make use o f t h e S t e i n i d e -

n t i t y [ v iz . , Appendix B o f Judge and Bock (1978)J and o b t a i n t h a t

W ~ I W -.. -.-. 1 = AE( 1

= E( - ( P - ~ ) E ( ) , (4.19) - 6 -8

*E( xp:4,A = E( xp,* - PE( xpt2,* i (4.20) -1 -1 E~(W'W)-~W'AWI -. -. -. -.- = Tr(4 - - Z ) E C W , , ~ ~ , ~ ) + A*ECX~;~,*), (4-21 )

1 1 1 where A * = A ' C - 4- E- h , and i n o r d e r t h a t (4,21) i s w e l l de f ined , - - .-. - - "S we need t h a t p 2 7 ( compared t o p > 3 f o r !n ). I n t h i s c o n t e x t , -

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 17: Whither jackknifing in stein-rule estimation

2260 SEN

we may note t h a t E ~ ( w ' u ) - ~ w ' A w ~ _.. _ _ _ - > ch ( A ) E ~ ( w ' w ) - 3 ~ , and hence, f o r P - .. -

the f in i t eness of the l e f t hand s ide of (4,21), p > 6 i s necessary

( s ince t h i s negative 4th moment does not a r i s e with the or ig inal

Ste in-rule es t imator , t h i s condit ion i s not necessary t h e r e . ) We

may fu r the r observe t h a t 6A*/6 - < O C I ~ ~ ( Q - ~ L - ~ ) - - = 1 and ~ T ~ ( Q - ' Z - ' ) - - -1 -1 -1 -1

= Tr(Q -. - T: ) c 1 ) 2 p. Thus, from (4.18) through (4 .21) , we

obtain t h a t

If we have t o have the asymptotic dominance of the jackknifed e s t -

imator gSSJ over Tn , we need t o have the inequal i ty sign in the -n l a s t formula holding f o r a l l A - z0, and , t h i s leads t o

p - > 7 and 0 < c < 2(p-4)(p-6)/(p-2)

= pSSJ(?) < p(?) , f o r every h E E'. - (4.23)

ye niay note f u r t h e r t h a t (p-4)(p-6) / (p-2) < (p-6) (p -2 ) , so t h a t

f o r the asyn~ptot ic dominance of t l ~ e semi-res t r ic ted jackknifed es-

t imator , we not only need p to be g rea te r than 6 , but a l s o c has

t o be r e s t r i c t e d t o a smaller domain ( which genera l ly leads t o a

s n ~ a l l e r reduction in the r i s k due t o shr inkage) . Also, in (4 .3 ) ,

i dea l ly , we take c = p-2 , while, in (4.18) o r (4 .22) , we would

have a smaller value f o r the optimal c. The question may na tu ra l ly

dt - ise ill t h i s context whether i:SJ dominates over $ in the l i g h t

oi' irle Airf? ( o r ttie asymptotic r i s k i t s e l f ) ? TO address t h i s que-

s l i u i ~ , we note t h a t by (4.3) ail0 (4 .18) ,

I t i s easy to show t h a t the f i r s t term on the rhs o f (4.24) i s non-

nevative , f o r a l l A ( i , e . , A > 0 ) . A t the null point ( i . e . , X = - S - d ) , the second termwis a l so nonnegative. Thus, pSSJ(0) p (c ) ,

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 18: Whither jackknifing in stein-rule estimation

WHITHER JACKKNIFING IN STEIN-RULE ESTIMATION? 2261

f o r a l l admissible values of c , and hence, t he asymptotic domina-

nce i n question does not hold, For h away from 0 , the second term -, ,"

on the rhs of (4.24) i s negative though i s small compared t o the

f i r s t term, so t h a t (4.24) i s genera l ly pos i t ive and q u i t e small , indica t ing t h a t the re i s no material change i n the A D R due to the

semi-res t r ic ted jackknifing over the Ste in-rule es t imator when the t rue parameter point i s d i f f e r e n t from the pivot (even i n t he local sense). Having t n i s naive p ic tu re on the ADR of the semi-res t r ic ted jackknifed Ste in-rule es t imator , we now proceed on t o examine tne

S otner important question : Can p (A) o r pSSJ(h) be estimated in a consis tent manner i n t h i s semi-restr icted jackknifing ? In t h i s se tup, X , being unknown, i s t r ea t ed as a nuisance parameter. ..

2 2 - Let us denote by "i = T i - T , G n i = - X n .

-1 n n i = l , ..., n , and l e t ! = n z i = l n n i . Note t h a t C i , l $ n i = O . Then, n using (4.11) through (4 .15) , i t follows by some standard s t eps

t h a t under {Knl and the assumed regu la r i ty condi t ions ,

Therefore, using (2 .13) , ( 2 . 8 ) , ( 2 . 9 ) , ( 3 . 3 ) , and (4 .13) - (4 .15) ,

along with (4.25) and (4 .26) , we obtain by some rout ine s t eps t h a t

under CKnl and the assumed regu la r i ty condi t ions ,

As a r e s u l t , we have

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 19: Whither jackknifing in stein-rule estimation

SEN

Again, t h e r h s o f (4.28) i s a non-degenerate r . v . , and hence, h e r e

a lso , j a c k k n i f i n g f a i l s t o p r o v i d e a c o n s i s t e n t e s t i m a t o r o f t h e

asympto t i c r i s k p S S J ( ~ ) . +SUJ F i n a l l y , l e t us c o n s i d e r t h e case o f 8, . By (2.8) th rough

(2.11) , we have on d e f i n i n g nni, 6ni as i n be fo re ,

C = c n n o b t a i n

I

[ A t t h i s s tage, we assume t h a t IC, -~ - c n l = o(n- ') ; i n v iew o f

t h e f a c t t h a t limn-cn=c e x i s t s , we may always choose t h e sequence

tcn1 i n such a way t h a t t h i s ho lds. Suppose t h a t i n ( 2 . 4 ) , we

o n l y r e p l a c e cn by cnql and d e f i n e t h e r e s u l t i n g q u a n t i t y by R:. Then I \An - 1 = o ( n - i ) a.s., and t h i s rep lacement w i l l make no asympto t i c d i f fe rence . As such, f o r s i m p l i c i t y o f p r o o f , we l e t

- 1 = c.] Then, by an appeal t o [ (3 .13) o f ] Sen (1977) , we

t h a t as n -+ m ,

= I l n 1 - in I $ 1 1 I = ~ ( n - ~ ) ass., (4 .30)

where we make use o f t h e f a c t t h a t t h e elements of Sn a r e a l l

U - s t a t i s t i c s and d n = C h (QS ) i s a smooth f u n c t i o n o f 2,. F u r t h e r , P --n

n o t e t h a t (n-l)Ani = - (Xi - X,), l < i < n , so t h a t - -

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 20: Whither jackknifing in stein-rule estimation

WHITHER JACKKNIFING IN STEIN-RULE ESTIMATION? 2263

n i a ~ { \ ( n - , 1 ) 6 ~ ~ / : l < i < n - - 1 = o(ng ) a .s . , a s n - -. Further , by using (3.7) and C3.0) of Sen Cl977), we obtain t h a t

t , , Also, under {Knl , n 2 1 (x,,( / = 0 (1) . F ina l ly , using (4.13), we ob-

P t a i n t h a t

ma x -2 - 2 -% 1 T - T I 1 = oin ) a . s . , as n + . (4.12)

Consequently, from (4.29) through (4.32), we have under iKn) , ?- 6SSJ 1 + 0, in p robab i l i ty , a s n -+ . (4.33) n 2 1 1 $, --,,

As a r e s u l t , (4.16) through (4.23) a l so per ta in t o $UJ . Again , we may use (4.25) through (4.28) and conclude t h a t f o r t h e unrest- r i c t ed jackknifed version too, (4.27) and (4.28) hold. Thus, l i k e

the case of the semi-res t r ic ted jackknifing, here a l so , the dual

ro l e of minimaxity ( in the l i g h t of the ADR) and e s t i m a b i l i t y of

the ADR i s not s a t i s f a c t o r i l y served . 5. SOME GENERAL REMARKS

We have observed in Section 3 t h a t in the conventional f ixed d l t e rna t ive case, l i k e the Ste in-rule es t imator , i t s jackknifed

versions a r e of no special use i f n i s large. This asymptotic pic-

tu re a l so ind ica te s the performance c h a r a c t e r i s t i c s , f o r f i n i t e 11 , when 0 i s not very c lose t o the pivot (here, 0) . I f F i s i t s e l f

w .-. a rnultinormal d . f . , then, of course, such a f inite-sample study car1 be nade r e l a t i v e l y more p rec i se ly , but a t the cos t of tremend- ous mathematical complications. The asymptotic p ic ture w i t h the versions of jackknifing considered here a l so per ta ins t o the case of multi-normal F, and hence, suggests the lack of u t i l i t y of these

jackkriifeu versions f o r l a rge sample s i z e s when the p ivot and the t rue parameter points a r e not the same.

In the case of Pitman-type a l t e r n a t i v e s ( which i s q u i t e app- ropr i a t e i n t he asymptotic case ) , i t appears t h a t whereas $:RJ and '5 0 a r e in agreement, t he o ther two jackknifed versions a r e not so , , n

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 21: Whither jackknifing in stein-rule estimation

SEN

;Yoreover, t h e o t h e r two v e r s i o n s have

f o r va lues o f p 2 7 C compared t o p > - c < 2(p-4) (p-6)/(p-2) << 2[p-2). From

asympto t i c dominance o v e r zn "S 3 f o r !n ) and f o r va lues o f

t h i s p e r s p e c t i v e , j a c k k n i f i n g

i s n o t ve ry a t t r a c t i v e . On t h e t o p o f t h a t , f o r e i t h e r of t h e

t h r e e j a c k k n i f e d ve rs ions , t h e co r respond ing j a c k k n i f e d e s t i m a t o r

o f t h e d i s p e r s i o n m a t r i c e s f a i l s t o p r o v i d e a c o n s i s t e n t e s t i m a t o r

o f t h e ADR. T h i s asympto t i c p i c t u r e a l s o p r o v i d e s a good i n d i c a t i -

on o f t h e f i n i t e sample s i z e s i t u a t i o n when i s ' c l o s e t o ' t h e

p i v o t . Thus, i n e i t h e r case, we have i n s u f f i c i e n t ground t o advoc-

a t e t h e use o f j a c k k n i f i n q i n S t e i n - r u l e e s t i m a t i o n .

For improved e s t i m a t i o n o f t h e niean v e c t o r o f a m u l t i n o r m a l

d i s t r i b u t i o n w i t h a known (p.d,) covar iance m a t r i x C , from c o n s i - - t i e r a t i o n s o f m i n i m a x i t y , Berger(1975) and Hudson(1974) b o t h sugge-

s t e d t h e use o f t h e f o l l o w i n g shr inkage e s t i m a t o r ( where a lso , we

use t h e p i v o t 0 ) : -

where

and r ( . ) = { r ( z ) , z E [O,m)} i s a s u i t a b l e nondecreasing and d i f f -

e r e n t i a b l e f u n c t i o n on [O,..); t h e y advocated t h e use o f t h e upper

bound o f 2(p-2) f o r r ( z ) . For p o s s i b l y unknown F and/or unknown Z,

one may r e p l a c e i n (5.1) and (5.2) Z by 2, , and cons ider a para-

l l e l e s t i m a t o r ; however, t h i s may n o t have t h e ( e x a c t ) m i n i m a x i t y

p r o p e r t y . The asympto t i c m i n i n l a x i t y o f such an e s t i m a t o r f o r Pitman

type l o c a l a l t e r n a t i v e s may be proved a long t h e same l i n e as i n

Sen (1984). As such, one may be i n t e r e s t e d i n i n c o r p o r a t i n g e i t h e r

o f the proposed j a c k k n i f e d v e r s i o n s i n t h i s S t e i n - r u l e e s t i m a t i o n .

By m a n i p u l a t i o n s v e r y s i m i l a r t o those i n S e c t i o n s 3 and 4, i t can

shown t h a t t h e genera l c r i t i c i s m s made on j a c k k n i f i n g i n e a r l i e r

s e c t i o n s a l s o p e r t a i n t o t h i s case. Thus, even i n such a somewhat

more genera l setup, we may n o t have s u f f i c i e n t grounds t o advocate

t h e use of j a c k k n i f i n g i n t h e S t e i n - r u l e e s t i m a t i o n theory. I f o u r S goal i s t o e s t i m a t e p (A), Sn may be used f o r L and we may a l s o

* .-" .-"

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 22: Whither jackknifing in stein-rule estimation

WHITHER JACKKNIFING IN STEIN-RULE ESTIMATION?

use the following es t imators of A and A* :

1 1 while K and Tr(Q- Y- ) may be cons i s t en t ly est ivatec! by d n and - - ~ r ( ( j - ' ? ; ~ ) , respect ively . These may then be used in (4.3) t o pro- viut! the desired answer, Ilowever, i n t h i s context, we may note t h a t

under {1$,1 , the es t imators in (5.3) and (5.4) a re not cons i s t en t

( but have 1 iniiting non-degenerate d i s t r i b u t i o n s ) , and hence, t h i s

s i r ~ ~ p l e procedure may not serve the purpose wel l , Perhaps, jackkni-

fed est imators of O and A* wil l have b e t t e r performance in t h i s

context.

ACKNOWLEDGEMENTS

This work was p a r t i a l l y supported by the liational Heart , L u n g

and Blood I n s t i t u t e , Contract IJIH-IIHLRI-71-2243-L from the National

I n s t i t u t e s of tlealth. Thanks a re due t o the r e fe ree f o r h i s very

c r i t i c a l reading of the manuscript and useful comments.

Berger, J.O. (1975). Hinimax estimation of locat ion vectors f o r a wide c l a s s of d e n s i t i e s . Ann. S t a t i s t , 3, 1318-1328.

Cercjer, J.O. (1930). f , robust generalized Eayes e s t i ~ : a t o r snc! con- f idence region f o r a mu1 t i v a r i a t e normal me?.n. Ann. S t a t i s t . 8, 716-761.

Berger, J.O,, Bock, h.E., Brown, L . D . , Casella, G. and Gleser, L. (1077). Minimax estimation of a normal mean vector f o r a r b i t - rary quadrat ic lo s s and unknown covariance matrix. Ann. S ta t - i s t . 5 , 763-771.

Hudson, j;. (1974). Er~lpirical Bayes Estimation. 2 c h , Report iv'o,Z8, i ~ t . of S t a t i s t i c s , Stanford Univ.

James, In;. and S te in , C , (1961). Estimation with quadrat ic lo s s . I roc. 4 th Berkelctl Symp. Math. S t a t i s t . ProbabiZity 2,361 -379.

Judye, 6.G. arld Bock, M.E. (1 978). 2he S ta t i s t i caZ Tmplications of ' r ~ i c s i and Stein-rule estimators i n ~conometr ics . liorth tlol- 1 ana Pub. , Amsterdam.

Saleh, A.K.Md.E. and Sen, P. K. (1 985). On shrinkage il-estimators of locat ion parameters, Comun. S t a t i s t . theo or. Heth, ~ 2 4 , 231 3- 2329.

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014

Page 23: Whither jackknifing in stein-rule estimation

2266 SEN

Saleh, A.K.Md.E. and Sen, P.K.(1986). On shrinkage l e a s t squares est imation i n a para1 l e l ism problem. C m n . S t a t i s t . Theor. Meth. AZ5, t o appear.

Sen, P.K. (1977). Some invariance p r inc ip le s r e l a t i n g to jackknif- ing and t h e i r ro l e i n sequential analys is . Ann. S t a t i s t . 5, 31 6-329.

Sen, P.K. (1984). A Janies-Stein detour of U-s t a t i s t i c s . C o m n . S t a t i s t . Theor. Meth. A 23, 2725-2748.

Sen, P.K. (1986). On the asymptotic d i s t r ibu t iona l r i sks of shr in- kage and preliminary t e s t versions of maximum l ikel ihood e s t i - mators. Sankhya, 5er.A 46, t o appear.

Sen, P.K. and Saleh, A.K.Md.E. (1985a). On some shrinkage estimat- o r s of mu1 t i v a r i a t e locat ion. Ann, S t a t i s t . 23, 272-281.

Sen, P .K. and Saleh, A.K.Md.E. (1 9856). !Ionparametric shrinkage estimators of locat ion in a mul t iva r i a t e simple regression model . Proc. Fourth Pannonian Symp, Math. S t a t i s t . (eds. Id. Wertz e t a1 . )

Ste in , C. (1956). Ir iadmissibil i ty of the usual es t imator f o r the mean of a mul t iva r i a t e normal d i s t r i b u t i o n . Proc. Third. Berk- e l ey Synp. Math. S t a t i s t . Probabi l i ty 2, 197-206.

Dow

nloa

ded

by [

New

Yor

k U

nive

rsity

] at

14:

28 2

3 N

ovem

ber

2014