s(po,j) =po(r + j )
which is uniform for (0, a) and
a, r, 0 > O,
which is gamma inverted on 0, so that h(p, O)=
( fl')/(ar~ ) o-('+ 1)exp(-fl/O); a, r, 0 > O.
Iz(s)= ( I / r s )
f:
e x p ( _ z ) z ~- l dz.
The Bayes estimator T B of 0 is obtained as
TB = E(O IX)
Thus
~bh(p, 0), i f p #P0, 0 # 00 g(p, 0) = ((1 - b), i f p =P0, 0 = 00.
= f,~ffOL(X/p,O)g(p,O)dpdO/
(5)
The joint posterior o f p and 0, via the Bayes theorem, will be: ~(P,
(6)
(4)
Following Ref. [5] consider a prior g(p, O) o f p and 0, which places a weight (1 - b) on the prior or guess values P0 and 00 of the parameters p and 0, and distributes the rest of probability mass b over some suitable interval of p and 0 according to the joint density h (p, 0). Further it is assumed that p and 0 are independent with the marginals
h2(0 ) = (fl~)/(Fot) 0-t,+ 1)exp(-fl/O);
P.=
c2(po)
O<~xt<~x2<~'"<~x,<~O, p,O>O.
f
fo' L(X/p, O)g(p, 0) dp dO.
c , ( p ) = p" [-I x V 1
r x p ' O -rp H xpi- 1. i=l
hl(p)=l/a;
r
This, on simplification, reduces to:
2. BAYESIAN SHRINKAGE ESTIMATORS OF p, 0 AND R(t)
L(X /P, O) =
f
O) = L(X/p, O)g(p, 0)/
fff[L(X/p,O)g(p,O)dpdO.,
(8)
After simplification
/
fff[L(X/p,O)g(p,O)dpdO., To obtain the Bayesian shrinkage estimator of p, 0 and R(t), we first find Bayes estimators by considering a squared error loss function. The Bayes estimator PB of p for the above prior g(p,O) is given by
PB = E(p IX)
TB=
C~(p) ~. C(j)xP/bT3(j,p)dp ]=0
+
C,(po) ~', C(j)x~°(1-b)r4(j,p) 1=0
if:
C~(p) ~ C(j)x~JbT3(j,p)dp j=O
+ C2(po)n-r ~. C(j)xW(1 - b)T4(j, p) 1 .
(9)
j=O
= f~; f[pL(X/p,O)g(p,O)dpdO/
The reliability function R(t) can easily be obtained without explicit knowledge of the posterior of R
A failure time model because R(t) is an increasing function of 0. The Bayes estimator/~a(t) of R(t) is given by
R B ( t ) = [ f f f : R(t)L(x/O,p)g(p,O)dp dO]/ I f f f : L(x/O,p)g(p,O)dpdO] Rs(t) = 1 -
I[;
(10)
r
tPC~(p) ~ C(j)xP/bTs(j,p)dp j=O
+tP°C2(po) ~ C(j)x~e(1 - b)T6(j,p) C~(p) Y~ C(j)xfbT3(j,p) dp
2041
and E3 = MSE(RML(t))/MSE(RBs (t)), respectively. It is not feasible to determine El, E 2 and E 3 by analytical methods so Monte Carlo methods have been used to determine empirical efficiencies. These are based on 500 simulated samples from a finite range failure time distribution of size n = 10, censored at r = 5. Prior values of the parameters were chosen as 00 = 30.00 and P0 = 0.70. Efficiency computations are carried out for 0/00= 0.25, 1.00, 2.00, 4.00, p/po=0.50, 1.00, 2.00, k =0.25, 0.50, 0.75, b = 0.25, 0.50, 0.75, ~ = 12.00, fl = 3.00, t = 0.40 and a = 2.00.
]=0
4. NUMERICAL FINDINGS AND CONCLUSION
+C:(Po) ~ C(j)XPr°J(1 - b)T4(zp ) , j=O
(ll)
where
Ts(j,p) - [F(a + p(r,j) + p)I~/x, (~ + s(p,j) +p)]/ ( a r ~t3 ,¢pJ ~+ , )
T6(j,p) = l/0ficp°J~+p°. Thus the ordinary Bayes estimators of p, 0 and are given by (7), (9) and (ll). The re.I.e, of 0 given by Ref. [1] is
R(t)
TM,=
(12)
x~,~.
The m.l.e, values o f p and PML = --(2n -
R(t) are
r)/ ~ (log X i - log)(,), (13) /
i~l
RML(t) = 1 -- (t/O) ~.
(14)
The Bayesian shrinkage estimators of the shape parameter p scale parameter 0 and reliability R(t) are, therefore, given by
Pas = kpML + (1 -- k)pa,
(15)
Ts s = kTM, + (1 - k)Ta
(16)
kRML(t ) + (1 -- k)Rs(t),
(17)
respectively, where 0 ~
The Bayes shrinkage estimators PBs, Tas and
RBs(t) are compared with m.l.e, values of p, 0 and R(t) on the basis of their efficiencies. The efficiencies of Pns, TBs and RBs(t ) with respect to corresponding m.l.e, values /'ME, TML and RML(t ) are defined as E~ = MSE(PML)/MSE(PBs),
E: = MSE( TML)/MSE( Tss)
(i) For p
and Ras(t) =
The three tables for the efficiencies EL, E 2 and E 3 have been obtained for three values of P/Po where b and k are equal. The tables are not presented here because of space constraints. Our findings are as follows.
(i) For p
2042
M. PANDEYand V. P. SINGH
Acknowledgements--The authors wish to thank Dr S. K. Upadhyay, Mr Borhan Uddin, Mrs J. Ferdos and Mr Rajesh Singh (comprising the research group in the Statistics Department of the B.H.U.) for their useful suggestions to debug the computer program and in typing this paper.
5. 6.
REFERENCES 7. 1. S. M. Ali and A. Islam, Parameter estimation in a failure time distribution, J. I A P Q R 2, 35-37 (1977). 2. C. D. Lai and S. P. Mukherjee, A note on a finite range distribution of failure times, Microelectron. Reliab. 26, 183-189 (1986). 3. J. Alan Gross and A. Virginia Clark, Survival Distributions Reliability Applications in the Biomedical Science. Wiley, New York (1975). 4. M. Pandey and V. P. Singh, Bayesian shrinkage estimation of reliability from a censored sample from a
8. 9. 10.
finite range failure time model, Microelectron Reliab. 29, 955-958 (1989). H. H. Lemmar, From ordinary to Bayesian shrinkage estimators, S. Aft. statist. J. 15, 57-72 (1981). M. Pandey and S. K. Upadhyaya, Bayesian shrinkage estimation of reliability from censored sample with exponential failure model, S. Aft. statist, J. 19, 21--23 (1985). M. Pandey and S. K. Upadhyay, Bayes shrinkage estimators of the Weibull parameters, IEEE. Trans. Reliab. R34, 491-494 (1985). M. Pandey, A Bayesian approach to shrinkage estimation of Weibull scale parameter, S. Afr. Statist. J. 22, 1-13 (1988). S. K. Sinha and B. K. Kale, Life Testing and Reliability Estimation. Wiley Eastern Limited, Delhi (1980). J. R. Thompson, Some shrinkage techniques for estimating the mean, J. Am. statist. Assoc. 63, 113 122 (1968).