797 Internat. J. Math. & Math. Sci. VOL. 12 NO. 4 (1989) 797-804 BOUNDS FOR THE MEAN SQUARE ERROR OF RELIABILITY ESTIMATION FROM GAMMA DISTRIBUTION IN PRESENCE OF AN OUTLIER OBSERVATION M.E. GHITANY and W.H. LAVERTY Department of Mathematics University of Saskatchewan Saskatoon, Saskatchewan Canada STN OW0 (Received August 6, 1987 and in revised form December 5, [988) In this paper we discuss the behavlor of the statistic ABSTRACT. minimum variance unbiased (UMVU) with unknown scale R(t). as a the uniformly estimate for the reliability of gamma distribution parameter a when an outlier observation is outlier effect on c, (t), Given the present. we determine bounds for the mean and mean square error (MSE) of A semi-Bayeslan approach is discussed when the outlier effect on prior distribution of random variable having a beta is treated of Results type. the exponential distribution (Sinha [I]) are given as particular cases of our results. KEY WORDS AND PHRASES. UMVU estimation, gamma distribution, reliability function, outlier observation, confluent hypergeometric series. 1980 AMS SUBJECT CLASSIFICATION CODE. I. 62F33. INTRODUCTION. be such that n-I of them are (XI,X2,...,X n) Let the independent random variables distributed as f(x:o) [(N I)’ q(’]-I e .-’x/ ( x N-I x > 0, o" > (1.1) 0, where N is a natural number, and one of these random variables Is distributed as f(x; (;/(*) while each X.i has [(N I)’ (o/ooN] -I a priori probability e x N-I x > 0, 0 < < I, I/n of being distributed as (1.2). context of outlier studies the model (I.I) is known as the "homogeneous case". (1.2) In the M. E. GHITANY AND W. H. LAVERTY 798 The reliability probability law f(x; a)Is f R(t) time" "mission at of t a system llfe whose follows the given by f(x); o) dx t e -t/e (t/) k (1.3) k! k=O Basu [2] and Nath [3], considering different approaches, obtained the unique UMVU R(t), namely, estimate of the reliability function N-I (t) J--0 where Aj I(I (t/s) N-j t/s) (n-l)N+j, (nN-1) (N-j-l)! [n-l]N+J)! and X s 0,I,...,N-I, having p.d.f i [(nN-l)’ fl(s;)= The (1.4) s, t nN]-1 e -s/o s nN-l, s > (I.6) 0. finding UMVU estimate for the reliability function from the gamma problem of distribution r() f(x;;,) #’]-I e -x/ x -I x > O, > 0, o > 0 with unknown parameters ,has not yet been solved. 2. VARIANCE OF (t), HOMOGENEOUS CASE. Since the second moment around the origin of R(t) is E[(t) 12 t we find that I E[(t) 12 where for any v L (t) v [(t) ]2 fl(S; c0 ds, 2 A.Ak 3 0j ekN- ( J+k)/2 (t) 0 [(nN-l)! onN] -I 2(N-v-I) t f e -s/ Os-(nN-l) (s t) 2[(n-l)N+v] ds. (2.2) t The integral I (t) can be simplified as follows v Iv(t) where l(1)(t) v (2) (t) v I . (1)(t) v + nN-1 r=O B . r: 1 (2)(t) v o)u v(t) 0f e-(t/ (l+u)-(nN-r-1) du 2[ (n- l) N+v rffinN B r:v (t) f e -(t/ )U(l+u)-(nN-r-1)du, 0 (2.3) (2.4) BOUNDS FOR THE MEAN SQUARE ERROR OR RELIABILITY ESTIMATION 799 with B r’v (t) (2[ (2[(n-l)N + v])) (n-|)N + v r)! (n--f)-!-(-l)r e -t/o (t/ ) nN for every r --0, l,...,2[(n-l)N+v]. A direct simplificat[on of the expressions in (2.3) and (2.4) gives us I e t/a nN-r-2 nN-3 (If(t) v [ [. r=O k--I (-t/ l)nN-r-2 B r: v . [ r=nN whe re -Ei(-Y) exponent [al e -Z z nN- r-k-2 BnN_l:v (t) (t/ )-I (t/o) -( r-nN- k+2) k! (2.6) (2.7) Idz, (2.8) Now, function. integral Ei(-t/ o) + (r-nN+ Br:v(t) k=O [(k-l)! (-t/) BnN_2:v (t) 2[(n-I)N+v] r-nN+l l(2)v (t) the (nN- r-2) Ei(-t/ )} a nd is (t) var[(t)] Eft(t)] 2 R2(t) can be computed. 3. MSE((t) ), BOUNDS FOR NONHOMOGENEOUS CASE. For the nonhomogeneous case [t can be shown that the p.d.f of s[n this case is given by N-I -n)N e- s/o s nN-I (a fct(S: ) r= 0 whe re D and (N-l-r)! r! ([n-l]N-l)! [(n-l)N+r] r IF1(.;.;.) (The notations [ (;m:z) k=O (B)k (-l)r s>o, (3.1) (3.2) r=0,1,...,N-l, Kummer’s confluent hypergeometrlc series, i.e. is the IFI and Dr F1 (l;(n_l)N+r+l;(l_ct)_s, (U)k k z (m) k!" k (3.3) are shifted factorials defined by ()k ( + I)...( + k I) (B)O=I) In particular s implies N-I [ and we get D r=O fl (s: q) as given by (1.6). f (s;o), the final r F (l:(n-1)N+r+l;O) (nN-l) .N..]-I e (nN-l) s nN-I s> O, Although MSE(R(t) a) can now be found explicitly by using result is determine bounds for MSE(R(t) ). not of practical form. Therefore, our For this purpose we consider the c.d.f aim is to 800 M. E. GHITANY AND W. H. LAVERTY F (n;o) Pr(s>,11;c where s is distributed as in (3.1). Fa o) Fl(n; It can be shown that q > 0. (3.4) s > O, (3.5) (; o), It follows that f (s;), f (s;o) and consequently E JR(t) (3.6) R(t) . At the same time we have m [(t)] N--1 -n)N (a D r=0 -n (o r (t) N- )N D r-O r -s/ o S nN-I IFI(I ;(n-|)N+r+l;(l-) )ds (l;(n-l)N+r+l;(l- ).._ (t) e- s/ osnN- Ids t F e t (3.7) L(a,t) R(t), where N-I L(,t)--N D (nN- l)! r=O IFIII;Cn-L)N+r+L;(I-). r (3.8) Using (3.6)and (3.7), we obtain E JR(t)] L(,t) R(t) "RCt). (3.9) By similar arguments as before, it can be shown that Since MSE(R(t)Ja) E ^ LCa,t) E[R(t)] El(t)]2 E[(t)]2 L(a, t) [(t)]2 2 R(t) E RZ(t) MSE(R(t)Ia 2 E[(t)]2. [(t)] + (3.10) R2Ct)- E[R(t)I we finally obtain 2 [2 L(a,t) II R2(t) (3.11) 2 ^ where R(t), E[R(t)] ,L(a,t)are given by (1.3), (2.1) and (3.8), respectively. Note and each of the bounds of (3.9) becomes the implies that L(l,t) that a-variance of (t). E Since [(t)] m a [(t)] where J r: (t) . . (t) t it follows that f (s; o) ds, N-I D r=O r J Ne-t/ (t/)nN (3.12) r: (t) N- J=O ([n-I]N + j)’ A (1)k [(l-o) t/ Jk=O ([n-l]N+j+l)k V( [n-l]N+J+l k k! [n-I N+J+k+2 t/ (3.13) BOUNDS FOR THE MEAN SQUARE ERROR OF RELIABILITY ESTIMATION 801 and / e-PV v -I o -<m(see Erdelyi [4]). m r:a (t) --(nN-l)! r<.- ) ) o< [,o + r(p) 2FI(.,.;.;.) m -2 ), m > (3. 4) 2, Using (3.14) in (3.13), it can be shown that N-l N e (z [(t/ o) N-j-I j=0 (N-J-l)’. [n-l]N+r+1" (I-,)) + (t/ o) where Idv +v) m-- 2 F I(I,[n-I]N+j+I; (3.15) [n-l]N+r+l;(l-a)t/ o)} IFI(I the Gauss’ hypereometrlc series, i.e. [s 2 F[(I’ 2; (l)k(P2)k k=O" re;z) --)k k z k’. (3.16) Further simplification leads to the approximation E [R(t) e a -t/o (t/o) r=0 for large n and small t, i.e. r r! the presence of a sngle outlier has little effect on the estimation of the reliability function R(t) of gamma distribution if there [s a large number n of items testing over a short period of time t. (Similar result is proved by Sinha [l] for the exponential distribution). 4. SEMI-BAYESI EN APPROACH. Consider a as a random variable having prior distribut[on of beta type with non- negative parameters p and q: _r_(p+q) g(a) ap-i r(p) r(q) (l-a) q-l, 0 < a < I. (4.1) The marginal p.d.f of s is given by h P’q f f = (s; o) (s; o) . g() dr 0 N-I M(p,q) fl(s;o) =0 Dr 2F2 (l,q’[n-l]N+r+l,N+p+q" s/ ), (4.2) whe re M(p q) )___r(N+p)r(nN) 2F2(I, z;ml is the generalized corresponding to p (4.3) F(p) F(N+p+q) and m2;z) hypergeometrlc and q-- I, (l)k (D2)k k z Z k=O (-m)k (m)k 2 k series. For we have the (4.4) homogeneous case, which is 802 M. E. GHITANY AND W. H. LAVERTY 2F2 I, (l,l;(n-l)N+r+l, ; s) and r=O,l,...,N-l, N-I M((R),I) which implies that h ,I Denote by E P,q f (s; o). (s; o) [R(t)] the epectatlon of R(t) when is distributed as in (4.1). Uslng (3.5), we get fl (s; a) f (s;) gCa)da a 0 that is h p,l (s;a) g(a)da, 0 (4 5) f (s;o), Consequently E P,q [(t)] Also, we have E P,q [(tll f R(t) t ) (4.6) (R(t). M(p,q) . hp ’q (s: o)ds D rffiO F (I q;[n-l]N+r+l N+p+q;tlo) r 2 2 t (t) fl(s;o)ds (4.7) L (p,q,t)R(t) where , L (p,q,t) N-I M(p,q) D r=0 r 2F2 (l,q;[n-1]N+r+l,N+p+q;t/ a). (4.8) R(t). (4.9) Using (4.6) and (4.7), we obtain L (p,q,t) R(t) E P,q [(t)] Similar ly L (p,q,t) E[(t)l 2 Ep, q [(t)l 2 Eta(t)12 (4.10) Finally, we have L (p,q,t) Et(t)]Z-RZ(t) MSE [(t) p,q] ErR(t)] z {2L*(p,q,t) -l}R2(t). It is easy to verify that for the homogeneous case, i.e. bounds in (4.11) becomes the variance of R(t). p=(R) (4.11) and q=l, each of the BOUNDS FOR THE MEAN SQUARE ERROR OF RELIABILITY ESTIMATION 5. 803 EXPONENTIAL DISTRIBUTION AS A PARTICULAR CASE. N=I, When i.e. we have an exponential distribution with scale parameter o, we find that R(t) e (t) (1 (5.1) s- )n-I t f (s O) a a r(n) on t e ( -s/O s n-I F (I ;n; (l-,)s/o) IFl(l;n;(l-a))[R(t)]2-e -2t/ -{2 -] p+q (5.2) s, MSE((t) 4 4 1Fl(1;n;(1-e)t/o)-1} (I q;n,p+q+l;t/o) E[R(t)] 2 -e 2t/o 2F2 ) E[(t)] 2 {2 (5 3) E[(t)] e -2t/O (MSE((t) 2F2(l,q;n,p+l+l;t/s) (5.4) p,q) i} e -2t/ (5.5) ,whe re E[(t)] 2 . with Il)(t)+ $2)(t) n-3 n-r-2 (t) r=O k=1 Br:0(t) (n-r-2) {(k-l)’. (-t/o) n-r-k-2 -et/(-t/o)n-r-2 Ei(-t/o) + (t/o) Bn_2:o(t) 2(n-l) r-n+l (t) r =n and B r:O (t) " Br:0 k =0 (t) (5.6) (r-n+l)Wk! Bn_l:0(t) -I (t/O)-(r-n-k+2) (2(n-l)) (-1) r e -t/o (t/o) n, r! (2(n-1)-r)! (n-l)! Ei -t / o) (5.7) (5.8) (5.9) The results in this section are those of Sinha’s [I]. ACKNOWLEDGEMENT The authors are grateful to the editor Dr.Lokenath Debnath and the referee for their helpful suggestions for improving the presentation of this paper. 804 M. E. GHITANY AND W.H. LAVERTY REFERENCES [. SINHA, 2. BASU, S.K., Outlier Observat[on, A.P., Estimation Reliability Estimates per. of in Life Testing In the Presence of an Res. 20, (1972), 888-894. Reliability for some Distributions Useful in Life testing, Technoetrlcs 6, (1964), 215-219. 3. 4. NATH, G.B., Unbiased Estimates of Reliabillty Distribution, Scand. Actuarial J.,(1975), 181-186. ERDELYI, A., Higher Transendental Funct[ons, Vol. I, McGraw-Hill, New York, 1953.