• 검색 결과가 없습니다.

Reference Priors in a Two-Way Mixed-Effects Analysis of Variance Model

N/A
N/A
Protected

Academic year: 2021

Share "Reference Priors in a Two-Way Mixed-Effects Analysis of Variance Model"

Copied!
12
0
0

로드 중.... (전체 텍스트 보기)

전체 글

(1)

2 002 , V ol. 13, N o.2 p p . 31 7~328

R e f e re n c e P rior s in a T w o - W ay M ix e d - E ff e c t s A n aly s i s o f V ari an c e M o de l

In H on g Ch an g 1) an d B y un g H w e e K im 2)

A b s trac t

W e fir st deriv e g r ou p or derin g r efer en ce prior s in a t w o - w ay m ix ed - effect s an aly sis of v arian ce (A N OV A ) m odel. W e sh ow t h at post erior distribut ion s are proper an d pr ov ide m ar gin al p ost erior dist ribu t ion s un der r efer en ce prior s . W e also ex am in e w h eth er t h e r eferen ce prior s s at isfy t h e pr ob ability m at chin g crit erion . F in ally , t h e r eferen ce prior satisfy in g th e pr ob ability m at chin g crit erion is sh ow n t o b e g ood in t h e sen se of fr equ ent ist cov era g e prob abilit y of th e post erior qu an tile .

K e y w o rd s : Error V arian ce, F r equ ent ist cov era g e pr ob ab ilit y , J effr ey s ' prior , M at ch in g prior , R eferen ce prior , T w o- w ay m ix ed - effect s A N OVA m odel.

1 . In tro du c ti on

T h e problem of est im at in g v arian ce in an aly sis of v arian ce m odel h a s b een inv est ig at ed by m any st atist ician s . E specially in Bay esian poin t of v iew , th e infer en ce of v arian ce com pon en t s in r an dom effect s m odel h a s b een t reat ed b y m an y st at istician s for a lon g tim e. F ir st , con sider t h e b alan ced on e - w ay r an dom effect m odel : y

ij

= +

i

+

ij

, i = 1 , 2 , , I , j = 1, 2 , ,J , w h ere is an kn ow n con st ant s , an d t h e

i

an d

ij

ar e in depen dent n orm al v ariable s w ith 0 m ean s

1. Lect uring Pr ofes s or , Departm ent of Applied St atist ics , Konkuk Univ er sity , S eoul, 143- 701, Kor ea

E - m ail : ihchang @konkuk .ac.kr

2. Pr ofes s or , Departm ent of M athem at ics , Hany ang Univ er s ity , Seoul, 133- 791, Kor ea

(2)

an d v arian ce

2

an d

2

, r espectiv ely . B ox an d T iao (1973 ) ch ose a prior dist ribu tion ( ,

2

,

2

)

- 2

(

2

+ J

2

)

- 1

an d calcu lat ed t h e post erior dist ribu t ion s . T h ey ob s erv ed t h e r elat ion ship b et w een t h e post erior of = J

2

/

2

an d a fr equ ent ist r esu lt in a hy pot h esis t est in g problem . Y e (1994 ) dev elop ed r efer en ce prior s for , ex am in ed fr equ en tist cov er ag e pr ob abilities of post erior qu an tiles for v ariou s an d com p ar ed of th e Bay e s est im at or s for r efer en ce prior s . T h e or der ed g r ou p refer en ce prior alg orit hm of B er g er an d B ern ar d (1989 ) is applied t o t h e b alan ced on e - w ay r an dom effect m odel by Ber g er an d Bern ar do (1992a ). A lso, Ch un g an d Dey (1998 ) deriv ed r efer en ce prior s an d fir st order prob abilit y m at chin g prior s for int r acla s s corr elation =

2

/ (

2

+

2

) an d ex am in ed t h e fr equ ent ist cov er ag e pr ob ab ilit ies of post erior qu an tiles for v ariou s . Recent ly Kim , K an g , an d Lee (2001) deriv ed t h e secon d or der pr ob ability m at chin g crit erion for t h e r at io of th e v arian ce com pon en t s . T h ey sh ow s t h at am on g all of t h e r efer en ce prior s g iv en in Y e (1994 ), t h e on ly on e refer en ce prior s atisfie s t h e secon d or der m at ch in g crit erion .

N ow , con sider a t w o- w ay m ix ed - effect s an aly sis of v arian ce (A N OV A ) m odel:

y

ij

=

i

+

j

+

ij

, i = 1 , 2 , , p ( > 1) , j = 1 , 2 , , q( > 1) , ( 1 . 1) w h er e

i

is t h e it h m ean effect of fix ed effect s , r an dom effect

j

' s in depen dent an d iden tically distribut ed a s N ( 0 ,

2

) ,

ij

' s ar e a s su m ed t o b e in depen den t an d ident ically dist ribu t ed a s N ( 0 ,

2

) . F urt h er , t h e

j

' s ar e also a s su m ed t o b e in dep en den t of

ij

' s . V oun at su an d S m ith (1997) stu died Bay sian appr oach for v arian ce com pon ent m odel an d h ier ar ch ical m odel w it h 2- v arian ce com pon en t for b alan ced an d u nb alan ced ca se in t his m odel. Recent ly Ch an g an d Kim (2002) con sider t h e pr oblem of e st im at in g

2

in th is m odel (1.1) u sin g J effrey s ' prior , r efer en ce prior , an d m at chin g prior s . T h ey t h en com p ar e qu an tiles of m arg in al p ost erior den sit ies of

2

in t w o r eal dat a s et .

In t his paper , w e con sider a Bay e sian an aly sis of err or v arian ce in th e m odel

(1.1) u sin g r efer en ce prior s . S in ce ou r focu s is fu lly B ay esian , ch oice of prior s is

v ery im p ort ant . T h e det erm in at ion of r ea s on able n oninform at iv e prior s in

m u lt ip ar am et er pr oblem is n ot ea sy ; com m on n oninform at iv e prior s , su ch a s

J effery s ' prior can h av e featu r es th at h av e dr am at ic effect on t h e post erior . M or e

specifically , B ern ar do (1979 ) poin t ed ou t t h at if w e ar e int er est ed in a su b set of

p ar am et er s , t h e r est b ein g n uisan ce par am et er s , t h en J effrey s ' prior m ay b e

in appr opriat e for r epr es ent in g v a gu e of litt le prior inform at ion . In or der t o

ov er com e t his pr oblem , B ern ar do (1979 ) pr opos ed th e r efer en ce prior appr oach for

t h e dev elopm en t of t h e n oninform at iv e prior . Ber g er an d Bern ardo (1989, 1992b )

ex t en ded t h eir alg orit hm t o m ultipar am et er pr oblem .

(3)

T h e pu rpose in th is pap er is t o obt ain refer en ce prior s for (

2

,

2

, ) w h er e

2

is t h e p ar am et er of in t ere st . T h e paper is arr an g ed a s follow s . S ect ion 2 deriv es t h e r eferen ce prior s . In S ection 3, w e sh ow th at post erior dist ribu tion s ar e pr oper an d pr ov ide m ar g in al p ost erior dist ribu t ion s u n der r eferen ce prior s . In S ection 4, w e ex am in e w h et h er t h e r efer en ce prior s sat isfy th e pr ob ability m at ch in g crit erion . F in ally , S ection 5 pr ov ides fr equ en tist cov er ag e pr ob abilit ie s of t h e p ost erior cr edible set s u sin g t h e r efer en ce prior s .

2 . T h e Grou pe d Orde rin g Re f e ren c e P ri ors

In Ber g er an d Bern ardo ' s (1992a ) r efer en ce prior appr oach t o t h e on e - w ay r an dom - effect m odel, t h e or der ed g roup is v ery im port ant . T h at is , t h e form of r efer en ce prior s can b e ch an g ed b y t h e order ed gr ou pin g . T his div ides t h e p ar am et er s int o t w o sub g r ou p s , called par am et er of in t er est an d nu is an ce p ar am et er s . N ot ation su ch a s {

2

, ( ,

2

) } w ill b e u sed t o specify t h e gr ou p an d t h e im port an ce of p ar am et er s ; {

2

, ( ,

2

) } m ean s t h at th er e ar e t w o g r ou p s , w ith

2

b ein g m ore im port an t th an t h e g r ou p ( ,

2

) . A g r ou p su ch a s {

2

, ,

2

}

m ean s th at

2

is t h e m ost im port an t par am et er an d

2

is t h e lea st im port ant . U sin g Ber g er an d Bern ar do ' s alg orit hm of com put in g t h e r eferen ce prior s , th e r efer en ce prior distribut ion s for differ en t g roup s of or derin g of (

2

, ,

2

) ar e obt ain ed a s follow s :

F or m odel (1.1), th e lik elih ood fun ction of par am et er s (

2

, ,

2

) is giv en b y l(

2

,

2

, | y ) (

2

)

-

q( p - 1)

2

(

2

+ p

2

)

-

q 2

ex p [ - 1 2 (

p

q

j = 1

( y

.j

- y

. .

)

2

2

+ p

2

+

p i = 1

q

j = 1

( y

ij

- y

i .

- y

.j

+ y

. .

)

2

2

+ q

p

i = 1

( y

i .

- y

. .

- (

i

-

.

) )

2

2

+ pq( y

. .

-

.

)

2

2

+ p

2

) ] , ( 2 . 1)

w h er e y

i .

= 1 q

q

j = 1

y

ij

, y

.j

= 1 p

p

i = 1

y

ij

, y

. .

= 1 pq

p i = 1

q

j = 1

y

ij

, an d

.

= 1

p

p i = 1 i

.

N ow u sin g n ot at ion s y = ( y

11

, , y

pq

)

T

an d = (

1

, ,

p

)

T

, (2.1) b ecom es

(4)

l(

2

,

2

, | y ) (

2

)

- q( p - 1)

2

(

2

+ p

2

)

- q 2

ex p [ - 1 2 (

p

q

j = 1

( y

.j

- y

. .

)

2

2

+ p

2

+

p i = 1

q

j = 1

( y

ij

- y

i .

- y

.j

+ y

. .

)

2

2

) ]

ex p [ - q

2

2

( ( - y )

'

( I

p

-

2

2

+ p

2

J

p

) ( - y ) ) ] . ( 2 .2)

F r om (2.2), F ish er in form ation m atrix is

I

1

(

2

,

2

, ) =

q ( p - 1)

2

4

+ q

2 (

2

+ p

2

)

2

pq

2 (

2

+ p

2

)

2

0

'

p q 2 (

2

+ p

2

)

2

p

2

q

2 (

2

+ p

2

)

2

0

'

0 0 q

2

I

p

- q

2

2

( p

2

+

2

) J

p

, ( 2 . 3)

w h er e I

p

is t h e p p iden tity m at rix an d J

p

is th e p p m at rix w it h 1 a s all elem en t s .

F r om (2.3 ), w e h av e th e follow in g t h eor em of th e r efer en ce prior of th e or derin g g r ou p (

2

, ,

2

) .

T h e o re m 2 .1 . F or t h e b alan ced t w o - w ay m ix ed - effect s A N OV A m odel if

2

is t h e p ar am et er of int er est , t h en t h e r efer en ce prior dist ribu t ion s for differ ent gr oup s of or derin g for (

2

, ,

2

) are ;

Gr oup ord e ring R ef er en ce p r ior

{

2

, ( ,

2

) }

1

(

2

)

-

1

2

(

2

+ p

2

)

-

3 2

{ (

2

,

2

) , }{

2

,

2

, }{

2

, ,

2

}

2

(

2

)

- 1

(

2

+ p

2

)

- 1

{ (

2

, ) ,

2

}

3

(

2

)

-

p 2 - 3

4

(

2

+ p

2

)

- 1

{ (

2

, ,

2

) }

4

(

2

)

-

p + 1

2

(

2

+ p

2

)

-

3 2

P ro o f . W e apply an alg orith m by Ber g er an d Bern ar do (1989 ) t o (2.3 ). S in ce t h e deriv at ion s of ot h er refer en ce prior s ar e sim ilar , w e con sider only th e r efer en ce prior for t h e g roup { (

2

, ) ,

2

} .

S t ep 1 : T h e u su al r efer en ce prior for

2

w it h (

2

, ) g iv en b ecom es ,

(

2

| (

2

, ) ) = ( 2 (

2

p + p

2

q

2

)

2

)

1 2

S t ep 2 : Ch oose an in cr ea sin g s equ en ce K

1

K

2

of com p act sub set s of t h e p ar am et er space for ( (

2

, ) ,

2

) su ch t h at

i

K

i

= . T ak e K

i

= [ 1

i , i] [ - i , i]

p

[ 1

i , i] . T h en n orm alize (

2

| (

2

, ) ) on

i , ( 2, )

(5)

= {

2

: ( (

2

, ) ,

2

) K

i

} obt ain in g p

i

(

2

|(

2

, ) ) = C

i

( (

2

, ) ) (

2

| (

2

, ) ) w h er e

C

i- 1

( (

2

, ) ) =

i 1 i

( p

2

q 2 )

1

2

(

2

+ p

2

)

- 1

d

2

= ( p

2

q 2 )

1

2

1

p [ log (

2

+ p

i

) - log (

2

+ p i ) ]

S t ep 3 : T h e m ar gin al r eferen ce prior for (

2

, ) w it h r esp ect t o p

i

(

2

| (

2

, ) ) is

i

( (

2

, ) )

= ex p

{ 2 1

1ii

p 1 [ log ( (

2

+ p

2

+ p

i

) - log (

2

)

- 1 2

+ p i ) ]

log q

p + 1

( p - 1)

2 (

2

)

- p - 1

(

2

+ p

2

)

- 1

d

2

}

= [ q

( p + 1)

( p - 1)

2 ]

1 2

(

2

)

-

p + 1

2

e

-

1

4[ log ( 2+ pi) + log ( 2+ p i) ]

S t ep 4 : T h e r efer en ce prior for ( (

2

, ) ,

2

) is for fix ed ( (

20

,

0

) ,

2

0

) , ( (

2

, ) ,

2

) = lim

i

C

i

(

2

, )

i

(

2

, )

C

i

(

20

,

0

)

i

(

20

,

0

) (

2

|(

2

, ) )

= (

2

)

- p + 1 2

(

20

)

-

p + 1 2

e

- 1

4 log

2 2

0

( p

2

q

2 (

2

+ p

2

)

2

)

1 2

(

2

)

-

p + 1

2 - 1

4

(

2

+ p

2

)

- 1

= (

2

)

- p 2 - 3

4

(

2

+ p

2

)

- 1

.

N ot e th at t h e prior dist ribu tion s ar e all fr ee of th e locat ion p ar am et er . A ll t h e refer en ce prior distribut ion s ar e pr oport ion al t o a n eg at iv e p ow er s of

2

an d

(

2

+ p

2

) . T h er efor e, a g en er al form of t h e prior can b e w rit t en a s

( ,

2

,

2

) (

2

)

- a

(

2

+ p

2

)

- b

, ( 2 .4 ) w h er e a an d b ar e p osit iv e n um b er s .

A lso, Jeffr ey s ' prior w hich is t h e s qu ar e root of t h e det erm in ant of t h e ex p ect ed F ish er in form ation m atrix is giv en b y

J

( ,

2

,

2

) (

2

)

-

p + 1

2

(

2

+ p

2

)

-

3

2

. ( 2 . 5)

T h er efor e, t h e J effr ey s ' prior is s am e a s t h e r efer en ce prior

4

for (

2

, ,

2

) .

It h a s b een arg u ed t h e r eferen ce prior dist ribu t ion depen d on th e s am ple size

an d design an d t h er efore v iolat e th e lik elih ood prin ciple. H ow ev er , t h er e is n o

s at isfact ory m et h od of obt ain in g a n on in form ativ e prior in th is s cen ario.

(6)

3 . M arg in al P o s t e rior D i s tribu tion s

A ccor din g t o Bay e s t h eor em , t h e p ost erior dist ribu tion of (

2

, ,

2

) w it h r espect t o t h e prior s in (2.4 ) is g iv en by

(

2

, ,

2

| y ) (

2

)

- a - q ( p - 1)

2

(

2

+ p

2

)

- b - q 2

ex p [ - 1 2 (

p

q

j = 1

( y

. j

- y

. .

)

2

2

+ p

2

+

p i = 1

q

j = 1

( y

ij

- y

i .

- y

. j

+ y

. .

)

2

2

+ q

p

i = 1

( y

i .

- y

. .

- (

i

-

.

) )

2

2

+ p q( y

. .

-

.

)

2

2

+ p

2

) ] . ( 3 . 1) In t egr at in g out an d

2

, th e r esu lt in g m ar gin al post erior dist ribu tion of

2

is

(

2

| y ) (

2

)

- p ( q - 1)

2 - a - b + 1

e

- S

2 2

(

1

0

w

q - 1 2 + b - 2

e

- S S 2 2 w

d w ) , ( 3 . 2) w h er e w =

2

2

+ p

2

. W e can see post erior distribut ion is pr oper fr om t h e follow in g t h eor em .

T h e o re m 3 .1 . T h e post erior dist rib ut ion s (3.1) is pr oper if a > 1 - ( p - 1)( q - 1) 2 an d b > 3 - q

2 .

P ro o f . W e prov e th e r esu lt for r eferen ce prior (2.4). F ull post erior dist ribu t ion s is g iv en by

R

(

2

, ,

2

| y ) [ C( y ) ](

2

)

- a - q ( p - 1)

2

(

2

+ p

2

)

- b - q 2

ex p [ - 1 2 (

p

q

j = 1

( y

.j

- y

. .

)

2

2

+ p

2

+

p i = 1

q

j = 1

( y

ij

- y

i .

- y

.j

+ y

. .

)

2

2

+ q

p

i = 1

( y

i .

- y

. .

- (

i

-

.

) )

2

2

+ pq( y

. .

-

.

)

2

2

+ p

2

) ] ,

w h er e

(7)

[ C( y ) ]

- 1

=

0 0 Rp

(

2

)

- a - q ( p - 1)

2

(

2

+ p

2

)

- b - q 2

ex p [ - 1 2 (

p

q

j = 1

( y

.j

- y

. .

)

2

2

+ p

2

+

p i = 1

q

j = 1

( y

ij

- y

i .

- y

.j

+ y

. .

)

2

2

+ q

p

i = 1

( y

i .

- y

. .

- (

i

-

.

) )

2

2

+ p q( y

. .

-

.

)

2

2

+ p

2

) ] d d

2

d

2

. ( 3 .3 ) In t egr at in g out in (3.3 ), w e obt ain follow in g r esult

Rp

ex p [ - q 2

2

(

p

i = 1

( y

i .

- y

. .

- (

i

-

.

) )

2

+ p

2

2

+ p

2

( y

. .

-

.

)

2

) ] d

=

Rp

ex p [ - q

2

2

( ( - y )

'

( I

p

-

2

2

+ p

2

J

p

) ( - y ) ) ] d

= ( 2 )

p 2

q

- p 2

(

2

)

p - 1

2

(

2

+ p

2

)

1

2

. ( 3 .4 )

T h en , u sin g th e re sult of (3.4 ), (3.3) is giv en by ( 3 . 3) = ( 2 )

p 2

q

-

p 2

0 0

(

2

)

-

q ( p - 1) - ( p - 1)

2 - a

(

2

+ p

2

)

-

q - 1

2 - b

ex p [ - 1

2

2

( S +

2

2

+ p

2

S S ) ] d

2

d

2

, ( 3 . 5)

w h er e S =

p i = 1

q

j = 1

( y

ij

- y

i .

- y

.j

+ y

. .

)

2

is err or su m of s qu ar e,

S S = p

q

j = 1

( y

.j

- y

. .

)

2

is sum of squ ar es du e t o r an dom effect s . Let r = 1

2

, w =

2

2

+ p

2

, t h en

2

= 1 - w

p r w an d | J | = 1

p r

3

w

2

. T h er efor e, ( 3 . 5)

0 1

0

r

p q - p

2 + a + b - 3

w

q - 1 2 + b - 2

ex p [ - r

2 ( S + w S S ) ] d w d r

<

0

r

p q - p

2 + a + b - 3

e

- r 2 s

0

w

q - 1 2 + b - 2

e

- rS S

2 w

d w

0

r

( p - 1) ( q - 1 )

2 + a - 2

e

-

r 2 s

d r

< if ( p - 1) ( q - 1)

2 + a - 1>0 an d q - 3

2 + b >0 .

R e m ark 3 .1 . T h e prior distribut ion w it h a = 1/ 2 in (2.4 ) pr odu ces an im pr oper p ost erior dist rib ut ion w h en p = 2 an d q = 2 . T his m ean s th at t h e prior

i

for t h e g r ou ped or derin g {

2

, ( ,

2

) } pr odu ces an im pr oper post erior dist ribu t ion w h en

p = 2 an d q = 2 .

(8)

4 . P ro b ability M at c hin g P rior

N ow w e ob t ain t h e pr ob abilit y m at ch in g prior s an d s ee w h et h er t h ey in clu de s am e of t h e r efer en ce prior s dev elop ed in S ect ion 2. F ir st , w e briefly rev iew t h e pr ob ab ilit y m at ch in g prior a s follow s .

F or a prior , let

11 -

( ; y ) den ot e th e ( 1 - ) th per cen tile of th e p ost erior dist ribu tion of

1

, t h at is

P [

1 11 -

( ; y ) | y ] = 1 - , (4 . 1)

w h er e = (

1 , 2

,

3

)

T

an d

1

is t h e param et er of in t ere st an d

2

an d

3

ar e n uisan ce param et er s . W e w ant t o fin d prior s for w hich

P [

1 11 -

( ; y ) | ] = 1 - + o( n

- u

) ( 4 .2 ) for som e u >0 , a s n g oes t o infinity . P rior s s atisfy in g (4.2) ar e called pr ob ab ilit y m at chin g prior s . If u = 1

2 , th en is r eferr ed t o a s a fir st or der m at ch in g prior , w hile if u = 1 , is r eferr ed t o a s a secon d or der m at chin g prior . In or der t o fin d su ch m at chin g prior s , it is con v en ien t t o in tr odu ce orth og on al p ar am etrization . T o th is en d , let

2

= ,

2

= 1

p ( - ) , = . ( 4 . 3)

W it h th is par am et rizat ion , b a s ed on t h e lik elih ood fun ction (2.2) an d F ish er inform at ion m at rix (2.3 ) of par am et er s (

2

,

2

, ) , t h e F ish er in form ation m atrix of p ar am et er s ( , , ) is g iv en by

I

2

( , , ) =

q( p - 1)

2

2

0 0

'

0 q

2

2

0

'

0 0 q

I

p

- q( - )

p J

p

. (4 .4 )

T hu s is orth og on al t o an d in th e sen se of Cox an d Reid (1987). By T ib shirani (1989), t h e cla s s of fir st or der pr ob abilit y m at ch in g prior is ch ara ct erized b y

( 1)

M

( , , )

- 1

g ( , )

an d th e fir st or der pr ob ability m at chin g prior of origin al param et er s (

2

,

2

, ) is g iv en by

( 1)

M

(

2

,

2

, ) (

2

)

- 1

g (

2

+ p

2

, ) , (4 . 5)

w h er e g (

2

+ p

2

, ) is arbit r ary differ en t iable fun ction .

(9)

Clearly th e cla s s of prior g iv en in (4.5 ) is qu it e lar g e, an d it is im port ant t o n arr ow dow n t his cla s s of prior s . T h er efor e w e deriv e th e cla s s of s econ d or der pr ob ab ilit y m at ch in g prior s by M u k erj ee an d Gh osh (1997 ).

T h e o re m 4 .1 . T h e secon d order pr ob ability m at chin g prior s ar e giv en b y

( 2)

M

(

2

,

2

, ) (

2

)

- 1

g (

2

+ p

2

, ) , (4 . 6)

w h er e g (

2

+ p

2

, ) is arbit r ary differ en t iable fun ction .

P ro o f . Let =

1

, =

2

,

i - 2

=

i

, i = 3 , , p + 2 , t h en secon d or der pr ob ability m at ch in g prior is of t h e form (4.5), an d also g m u st s atisfy an addition al differ en tial equ at ion a s follow s :

(

M( 1)

, ) =

2 2 1

( I

-

1 2

11

g (

2

, ,

p + 2

) )

-

p + 2 r = 2

p + 2

s = 2 s

( L

11 r

I

11

I

rs

I

1 2

11

g (

2

, ,

p + 2

) ) - 1

3

1

( L

111

( I

11

)

2

I

1 2

11

g (

2

, ,

p + 2

) ) = 0 , (4 . 7)

w h er e

L

111

= E (

3

L

3 1

) = - q( p - 1)

3 1

+ 3 ( p - 1) ( q - 1)

3 1

- 3 ( p - 1)

3 1

,

L

112

= L

113

= = L

11( p + 2 )

= 0 ,

an d I

ij

is ( i , j ) elem en t of

I

2- 1

=

2

21

q( p - 1) 0 0

'

0 2

22

q 0

'

0 0

1

q I

p

+

2

q J

p

.

T h en ev ery differ en t iable fun ction g s atisfie s (4.7 ). T h at is , all th e fir st or der pr ob ab ilit y m at chin g prior s ar e t h e secon d or der pr ob ab ilit y m at chin g prior s . T h u s t h e r esultin g secon d or der pr ob abilit y m at chin g prior s for ar e giv en by

( 2)

M

( , , )

- 1

g ( , )

an d th e secon d or der pr ob abilit y m at chin g prior s for

2

ar e

( 2 )

M

(

2

,

2

, ) (

2

)

- 1

g (

2

+ p

2

, )

T his com plet es t h e proof.

R e m ark 4 .1 . T h er e are in finit ely m an y m at chin g prior s for

2

up t o o( n

- 1

) .

A m on g th e r eferen ce prior s dev eloped in S ection 2

2

is th e on ly m at chin g prior .

(10)

5 . S im u l ation an d D i s cu s s ion

W elch an d P eer s (1963) pr ov ed t h at , u sin g J effr ey s ' prior th e differ en ce b et w een t h e fr equ en t ist cov er ag e of t h e t h post erior qu ant ile an d is th e or der of

o( 1/ n ) w h en n g et s larg e. A popu lar pr opert y is t h at t h e frequ en tist cov era g e pr ob ab ilit y of ( 1 - ) th p ost erior qu an tile sh ow ed b e close t o 1 - .

T ab le 5.1 pr ov ides fr equ ent ist cov er ag e pr ob abilit ies of 0.05 (0.95 ) post erior qu ant ile for

2

u n der four r eferen ce prior s for differ en t v alu e of p an d q . T h e com pu t ation of th ose n um erical v alu es is b a sed on t h e follow in g alg orit hm for any fix ed tru e

2

an d pr esp ecified pr ob abilt y v alu e . H er e is 0.05 (0.95 ). Let (

2

) ( | y ) b e t h e p ost erior - qu an tile of

2

g iv en y . T h at is t o s ay , F ( (

2

) ( | y ) | y ) = , w h er e F ( | y ) is t h e m arg in al post erior distribut ion of

2

. T h en t h e fr equ en tist cov era g e pr ob ability of th is on e sided cr edib le in t erv al of

2

is

P

2

( :

2

) = P

2

( 0 <

2

(

2

) ( | y ) ) .

T a b l e 5 .1 . F re qu e n t i s t c o v e ra g e pro b ab ilit i e s o f 0 . 05 ( 0 . 95) p o s t e ri o r qu an t il e s f o r

2

( p , q)

1 2 3 4

(2, 8 ) 0.02 (0.92) 0.03 (0.93 ) 0.02 (0.92) 0.01(0.75 ) (3, 4 ) 0.03 (0.94 ) 0.05 (0.95 ) 0.05 (0.93 ) 0.03 (0.89 ) (4, 3 ) 0.05 (0.97 ) 0.04 (0.97 ) 0.01(0.80) 0.01(0.76 ) (5, 5 ) 0.05 (0.96 ) 0.05 (0.95 ) 0.02 (0.85 ) 0.01(0.73 ) (5, 6 ) 0.03 (0.93 ) 0.04 (0.94 ) 0.03 (0.90) 0.01(0.74 ) (6, 5 ) 0.04 (0.95 ) 0.05 (0.95 ) 0.03 (0.87 ) 0.01(0.77 ) (5, 10) 0.03 (0.90) 0.05 (0.95 ) 0.02 (0.85 ) 0.01(0.78 ) (8, 2) 0.03 (0.93 ) 0.06 (0.98 ) 0.02 (0.80) 0.01(0.70) (10,10) 0.03 (0.95 ) 0.04 (0.95 ) 0.01(0.85 ) 0.02 (0.85 )

T h e est im at ed P

2

( :

2

) w h en = 0 . 05 ( 0 . 95 ) is sh ow n in T able 5.1. A ctu ally

T able 5.1 w a s com pu t ed in t h e follow in g w ay . F or fix ed

2

, w e t ak e 10.000

in dep en den t r an dom sam ples of y fr om t h e m odel (1.1). N ot e th at u n der t h e prior

(11)

, for fix ed y ,

2

(

2

) ( | y ) if an d on ly if F ( (

2

) ( | y ) | y ) . U n der t h e prior , P

2

( :

2

) can b e est im at ed by th e relat iv e fr equ en cy of F ( (

2

) | y ) . F r om t h e T able 5.1,

2

is th e m ost appealin g r eferen ce prior dist ribu tion in th e s en s e of th e a sy m pt otic fr equ ent ist cov er ag e pr ob ability .

R ef e re n c e s

1. Ber g er , J . O. an d B ern ar do, J . M .(1989 ). E st im atin g a P r odu ct of M ean s ; Bayesian Analy sis with Reference Priors. J . A m er. S tatis t. A ss oc., 84, 200- 207.

2. B er g er , J . O. an d Bern ar do, J . M .(1992a ). Refer en ce P rior s in a V arian ce Com p on ent ' s P r oblem , in P . Goel, E d ., P r oc. of t h e In do- U S A W ork sh op on Bayesian Analy sis in St atistics and Econometrics. Springer, New York, 323 - 340.

3. B er g er , J . O. an d Bern ar do, J . M .(1992b ). On th e Dev elopm en t of R eferen ce Prior s (with discu ssion ). B ay es ian S tatis tics 4 (B ernardo, J . M . et al., eds .) . Ox for d Univ . P r es s , Oxf ord, 35 - 60.

4. Bern ardo, J . M .(1979). Referen ce P ost erior Distribution s for Bay esian In fer en ce (w it h dis cu s sion ). J . R oy al S ta tis t. S oc. (S er. B ) , 41, 113- 147.

5. Box , G. E . P . an d T iao, G. C.(1973 ). B ay es ian I nf er en ce in s ta tis tical A naly s is . J oh n W iley an d S on s , In c., N ew Y ork .

6. Ch an g , I. H . an d Kim , B . H .(2002). Bay esian A n aly sis for t h e E rr or Varian ce in a T w o- W ay Mix ed- Effect s ANOVA M odel U sing N on in form at iv e P rior s . T h e K orean J ournal of A p p lied S ta t is t ics , V ol. 15, N o. 2, 405 - 414.

7. Chun g , Y . an d Dey , D. K.(1998). Bay esian Appr oach t o E stim ation of Intraclass Correlation Using Reference Prior. Comm unications in S ta t is t ics - T he ory and M e thods , 27, 2241- 2255.

8. Cox , D . R . an d Reid, N .(1987 ). Ort h og on al P ar am et er s an d A ppr ox im at e Conditional Inference (with discu ssion ). J . R oyal S tatis t. S oc. (S er. B ) , 49, 1- 39.

9. Kim , D . H ., K an g , S . G., an d Lee, W . D .(2001). On S econ d Or der P r ob ability M at chin g Crit erion in th e On e - W ay Ran dom Effect M odel. T he K orean Com m un ica tions in S ta tis tics , V ol. 8, N o. 2, 29- 37.

10. M u k erj ee, R . an d Gh osh , M .(1997 ). S econ d Or der P r ob abilit y M at ch in g P rior s . B iom e tr ika, 84, 970- 975.

11. T ib sh iriani, R .(1989 ). N on in form at iv e P rior s for on e P ar am et er of M an y . B iom e tr ika 76, 604 - 608.

12. V ou n at sou , P . an d S m it h , A . F . M .(1997 ). S im u lation - Ba sed Bay esian Infer ence for T w o- Variance Compon ent s Lin ear Models. J ournal of S ta t is t ical P lann ing an d I nf e ren ce , 59, 139- 161.

13. W elch , B . L. an d P eer s , H . W .(1963 ). On F orm u la e for Confiden ce P oint s

B a sed on Int eg r als of W eight ed Lik elih ood s . J . R oy al S ta tis t. S oc. (S er. B ) ,

(12)

25, 318 - 329.

14. Y e, K .(1994 ). B ay esian R eferen ce P rior A n aly sis on t h e Ratio of V arian ces for t h e Balan ced On e - W ay Ran dom Effect M odel. J ournal of S ta t is t ical P lann ing and I nf eren ce, 41, 267 - 280.

[ 2002년 9월 접수, 2002년 10월 채택 ]

참조

관련 문서

mandibular canal (10mm back from the reference point); B, diameter of the mental canal; C, diameter of the incisive canal (5mm forward of the reference point); D, distance

t-test and analysis of variance were carried out to find out the differences along general characteristics and it was analyzed by setting correlation

3.4 The estimating stage of relative coordinates : (d) coordinates transfer based on estimated reference node (e) Get the azimuth of mobile node (f) Get.. the final

After first field tests, we expect electric passenger drones or eVTOL aircraft (short for electric vertical take-off and landing) to start providing commercial mobility

First, the general characteristics of the two groups were analyzed by frequency analysis and correlation analysis, which is technical statistics, and the

Evaluating parameters of osseointegrated dental implants using finite element analysis-a two- dimentional comparative study examining the effects of implant

Data were analyzed using independent-sample t tests for checking the equivalency between the experimental and control groups at pretest and three-way analysis

(구체적으로 어느 집단의 평균이 다른 것인지 판단하기 위한 것이 다음에 설명할 사후 검정이다.) 두 통계량의 p-값에 차이가 있다면, Welch 검정법이 더 강력하다고