J ournal of K or ean
D a ta & I nf orma t ion S cience S ocie ty 2 002 , Vol. 13, N o.1 p p . 139~ 145
B ay e s i an P re dic t io n A n aly s i s
f o r t h e E x p o n e n t i a l M o d e l U n d e r t h e C e n s o re d S am p le w it h In c o m p le t e In f o rm at io n
Y e un g - H oon Kim , Je on g - H w an K o 1)
A b s t ra c t
T his paper deals w ith the problem of obt aining the Bayesian predictiv e density function and the prediction int erv als for a future observ ation an d the p - th or der st atistics of n ' fut ure observ ations for the exponential m odel under the censor ed sampling w it h incomplet e inform ation .
K ey W or d s a nd P hra s es : Bay esian Approach , Prediction , Incomplet e Censored Sample, Exponent ial distribut ion
1. In trodu c tion
T h e pr oblem of pr edict in g a fut u r e ob ser v at ion h a s r eceiv ed m u ch at t ent ion an d h a s b een dealt m ain ly in t w o appr oach es . On e is th e u su al cla s sical appr oach an d t h e ot h er is B ay esian appr oach . B a sed on t h e Bay esian appr oach , Ch hik ar a an d Gut t m an (1982), S inh a (1989 ), Up adhy ay an d P an dey (1989 ), an d Nigm an d A L - W ah ab (1996 ) su g g e st ed t h e B ay esian infer en ce ab out pr edict ion for Gau s sian , log n or m al, ex pon en tial an d Bur r dist r ib ut ion , r esp ect iv ely .
T h e ex pon en t ial dist r ibu t ion play s an im p or t ant r ole in m an y pr act ical r eliabilit y an aly s es . It is th e fir st m odel for w hich st at istical m eth od s w er e ex t en siv ely dev eloped an d is w idely u sed a s a r eliability m odel. E lper in an d Gert sb akh (1988 ) in v est ig at ed th e Bay e s in t er v al e st im at ion for ex pon ent ial p ar am et er in a r an dom cen sor ed sam plin g w ith in com plet e infor m at ion w h ich in clu des , a s part icular ca ses , b ot h t h e r an dom cen s or in g m odel an d t h e qu ant al - r e spon s e m odel. Calabr ia an d P u lcini (1990) pr opos ed t h e Bay esian pr ocedu r e for estim at in g t h e ex pon en tial m ean lifetim e an d t h e r eliability fun ct ion in a t im e cen sor in g m odel w it h in com plet e infor m at ion u sin g t h e s qu ar ed er r or 1. Department of St atistics, Andong National University.
E - m ail : jhko@anu .andong .ac.kr
los s . Kim (1995 ) pr op os ed th e B ay esian pr ocedu r e for est im at in g th e R ay leigh r eliabilit y fu n ction un der t h e cen sor ed s am ple w it h in com plet e in for m at ion .
In th is paper , w e w ill st u dy t h e B ay esian pr edict iv e den sity fu n ct ion an d t h e pr edict ion int er v als for a fu tu r e ob s er v at ion an d th e p - th or der st at is tic s of n ' fu tu r e ob s erv ation s , at a specified m is s ion tim e t , b a s ed on a tim e c en s or ed s am ple w it h in com plet e infor m at ion ob s er v ed fr om th e ex p on ent ial m odel.
In s ect ion 2, w e deriv e t h e pr ediction an a ly sis of a fut ur e ob s er v a tion fr om t h e ob s er v at ion . In s ect ion 3, w e pr opos e th e pr e dict ion a n aly sis of th e p - th or der s t at ist ic s y ( p ) .
2 . P re dic t i on of a F u t u re Ob s e rv at io n
T h e pr ob abilit y den sit y fun ct ion of th e ex p on ent ial dist r ib ut ion , den ot ed by E ( ) , is g iv en by
f ( x | ) = - 1 e - x / , >0 , 0 < x < . (2.1) T h en t h e r eliab ilit y at a s pecified tim e t > 0 is giv en by
P ( X > t ) = e - t/ . (2.2)
L et t b e t h e fix ed tim e t o in spe ction an d let x i b e t h e ex pon en tial lifetim e of it em i ( i = 1, 2 , , n ) . T h en th r ee pos sibilitie s ca n b e occu r r e d in t e st in g it em i : F ir s t , t h e it em fails at th e in st ant x i ( x i < t ) an d th e fa ilur e is n ot sign alled . T h e it em is foun d failed on t h e in sp ect ion t im e t . In th is ca s e , th e failu r e t im e is b efor e t h e in spe ction t im e . S econ d , t h e it em fails at th e in s t an t x i ( x i < t ) an d th e failu r e is im m edia t ely sig n a lled. In t his ca s e , t h e failur e t im e is ex a ctly k n ow n . T hir d , t h e it em is foun d un fa iled at tim e t . In t his ca s e , th e failur e tim e w ou ld b e b ey on d t h e in s pect ion t im e .
W h en n it em s a r e t e s t ed , th e cor r e s pon din g lik elih ood fun ct ion is giv en by
L ( | x ) - n
2e - u / ( 1 - e - t/ ) n
1e - tn
3/ , (2.3)
w h er e n 1 is t h e n um b er of elem ent s in t h e s et of n on c en s or ed a n d n on sign alled ob s erv ation s , n 2 is t h e n um b er of elem en t s in t h e s et of n on cen s or ed an d s ig n alle d ob s er v at ion s , n 3 is t h e n um b er of elem ent s in t h e s et of cen s or e d ob s erv ation s , an d u is t h e su m of squ ar es of th e failu r e t im e of it em s in t h e set of n on cen sor ed an d sign alled ob s er v at ion s . T h en n = n 1 + n 2 + n 3 an d th e v alu e of n 1 an d n 2 depen d on th e v alu e of t h e pr ob abilit y of failu r e - t o - sig n al p . On t h e av er ag e n 2 / ( n 1 + n 2 ) = p .
Un for t un at ely , sin ce t h e lik elih ood fu n ction is n ot ex pon en tial, a n at u r al
conju g at e pr ior can n ot b e fou n d . T hu s a pr ior den sit y g en er ally inv olv e s
n um er ical int egr at ion s . But u sin g t h e bin om ial for m u la , th e lik elih ood fun ct ion of can b e w rit t en a s
L ( | x ) - n
2n
1j = 0 c j e - {u + ( n
3+ j ) t }/ , (2.4)
w h er e c
j= n
1( ) j ( - 1)j, j = 0 , 1, , n
1. H en ce th e com m on ly u sed in for m ativ e pr ior for y ield s som e e st im at or s of w hich ar e in an an aly t ical for m .
N ow w e con sider t h e J effr ey s (1961) ' n oninform at iv e pr ior for . W h en t h e t im e t is sp ecified, t h e n on in for m ativ e pr ior for is
( ) - 1 , >0 (2.5)
F r om (2.4 ) an d (2.5 ), t h e post er ior den sity of is g iv en by
( | x ) =
- ( n
2+ 1) n
1j = 0 c j e - {u + ( n
3+ j ) t }/
( n 2 )
n
1j = 0 c j {u + ( n 3 + j ) t } - n
2, >0 , (2.6)
w h er e ( n ) is th e g am m a fun ction defin ed by ( n ) = 0 z n - 1 e - z dz . T h e dist r ib ut ion of a fut ur e ob ser v ation y g iv en is
f ( y | ) = - 1 e - y / , >0 , 0 <y < . (2.7) T h en t h e pr edict iv e den sit y fu n ct ion for y can b e obt ain ed an d is g iv en in t h e follow in g t h eor em .
T h e o re m 2 .1 . W it h a n on in for m ativ e pr ior for , th e pr edict iv e den sit y fun ct ion of y g iv en X = x is
(y | x ) = n 2
n
1j = 0 c j ( w j + y ) - ( n
2+ 1)
n
1j = 0 c j w j - n
2, (2.8)
w h er e w j = u + ( n 3 + j ) t.
N ow w e w an t t o con str u ct th e pr edict ion int er v als for a fu tu r e ob ser v at ion . If a n on in for m ativ e pr ior for is u sed , t h e 100 ( 1 - ) % equ al - t ail pr ediction int er v al ( C N L , C N U ) for a fut u r e ob serv ation y ar e t h e s olut ion s of th e equ at ion s
2 =
n
1j = 0 c j ( 1 - w j
w j + C N L ) n
2an d 2 =
n
1j = 0 c j ( w j
w j + C N U ) n
2. (2.9)
T h e 100 ( 1 - ) % m ost plau sible pr edict ion int er v al ( M N L , M N U ) for y can b e obt ain ed by solv in g sim u lt an eou sly t h e follow in g s :
n
1j = 0 c j ( 1 + M N L w j ) - n
2-
n
1j = 0 c j ( 1 + M N U
w j ) - n
2= 1 - (2.10)
an d
n
1j = 0 c j ( w j + M N L ) - ( n
2+ 1) =
n
1j = 0 c j ( w j + M N U ) - ( n
2+ 1) . (2.11)
A s a pr ior distr ibut ion for , w e con sider an inv ert ed g am m a pr ior dist ribu tion I G( , ) w it h t h e pr ob abilit y den sity fun ct ion
( | , ) = ( )
- ( + 1)
e - / , , >0 , >0 . (2.12) T h en on e can obt ain ea sily th e post er ior den sity of giv en X = x w hich is g iv en by
( | x ) =
- ( n
2+ + 1) n
1j = 0 c j e - {u + ( n
3+ j ) t + }/
( n 2 + )
n
1j = 0 c j {u + ( n 3 + j ) t + } - n
2, , >0 , >0 . (2.13)
H en ce th e follow in g th eor em can b e obt ain ed .
T h e o re m 2 .2 . W it h an in v er t ed g am m a pr ior for , th e pr edict iv e den sit y fu n ct ion of y giv en X = x is
(y | x ) =
( n 2 + )
n
1j = 0 c j ( v j + y ) - ( n
2+ + 1)
n
1j = 0 c j v j - ( n
2+ )
, (2.14)
w h er e v j = u + ( n 3 + j ) t + .
Un der t h e inv er t ed g am m a prior distr ibut ion for , t h e 100 ( 1 - ) % equ al - t ail pr ediction in t erv al ( C G L , C G U ) for y ar e th e solut ion s of th e equ at ion s
2 =
n
1j = 0 c j ( 1 - v j
v j + C G L ) n
2+ an d 2 =
n
1j = 0 c j ( v j
v j + C G U ) n
2+ . (2.15)
A lso t h e 100 ( 1 - ) % m ost plau sible pr ediction in t er v al ( M G L , M G U ) for y can
b e obt ain ed by solv in g sim u lt an eou sly t h e follow in g s :
n
1j = 0 c j ( 1 + M G L
v j ) - ( n
2+ ) -
n
1j = 0 c j ( 1 + M G U
v j ) - ( n
2+ ) = 1 - (2.16)
an d
n
1j = 0 c j ( v j + M G L ) - ( n
2+ + 1) =
n
1j = 0 c j ( v j + M G U ) - ( n
2+ + 1) . (2.17)
3 . P re di c tion of th e Orde re d Ob s erv ation
In th is sect ion w e con sider t h e pr edict iv e den sity fu n ct ion of th e p - th or der st at ist ic y ( p ) of n ' fut ur e ob s er v at ion s . T h e pr ob ability den sit y fun ction of
y ( p ) is a s follow s :
(y ( p ) | ) =
p - 1
i = 0 d i e - {( n
'