Computational Statistics and Data Analysis 53 (2009) 3560–3570 Contents lists available at ScienceDirect Computational Statistics and Data Analysis journal homepage: www.elsevier.com/locate/csda Expected values of the number of failures for two populations under joint Type-II progressive censoring Safar Parsi a , Ismihan Bairamov b,∗ a Department of Statistics, University of Mohaghegh Ardabili, Ardabil, Iran b Department of Mathematics, Izmir University of Economics, Izmir, Turkey article info Article history: Received 27 November 2008 Received in revised form 19 March 2009 Accepted 20 March 2009 Available online 7 April 2009 a b s t r a c t The model in which joint Type-II progressive censoring is implemented on two samples from different populations in a combined manner is considered and the probabilities of failures are discussed. This model may have relevance for practical applications, when an experimenter need to know the expected values of the number of failures for each population. This knowledge plays an important role in the selection of appropriate sampling plans and the construction of efficient estimators for parameters. The formula allowing numerical computation of the expected value of the number of failures for each of the two populations is given. Also, a detailed numerical study of this expected value is carried out for different parametric families of distributions. © 2009 Elsevier B.V. All rights reserved. 1. Introduction Recently, Balakrishnan and Rasouli (2008) have introduced a joint censoring scheme which is important in performing comparative life-tests of two identical products manufactured in different units under the same conditions. To describe this scheme, suppose products are being manufactured by two lines under the same conditions and two samples of products of sizes m and n are selected from these two lines and are placed simultaneously on a life-testing experiment. Then, based on cost considerations and time restrictions for completion of the test, suppose the researcher chooses to terminate the lifetesting experiment as soon as a certain number of failures occur. In this situation, either point or interval estimation of the mean lifetimes of products manufactured by two lines may be of interest. Under joint Type-II censoring, specimens of two products under study are placed on a life-test simultaneously, the successive failure times and the corresponding product types are recorded, and the experiment is terminated as soon as a specified total number of failures (say, k) occurs. Suppose X1 , . . . , Xm , the lifetimes of m specimens of product A, are independent and identically distributed (i.i.d.) random variables, with cumulative distribution function (c.d.f.) F (x) and probability density function (p.d.f.) f (x). Let Y1 , . . . , Yn be i.i.d. random variables with c.d.f. G(x) and p.d.f. g (x), representing the lifetime of n specimens of product B. Further, suppose W1 ≤ · · · ≤ WN denote the order statistics of the N = m + n random variables {X1 , . . . , Xm ; Y1 , . . . , Yn }. Then, under the joint Type-II censoring scheme, the observable data consists of (Z, W), where W = (W1 , . . . , Wk ), with k (1 ≤ k < N) being a pre-fixed integer, and Z = (Z1 , . . . , Zk ) with Zi = 1 or 0 accordingly as Wi is from an X - or Y -failure. A joint censoring scheme has been previously considered in the literature. For example, Basu (1968), discussed a generalized Savage statistic, while Johnson and Mehrotra (1972) studied the locally most powerful rank tests. Bhattacharyya and Mehrotra (1981) considered the problem of testing the equality of two distributions, under the assumption of exponentiality. The developments under this sampling scheme have been reviewed by Bhattacharyya (1995) in Chapter 7 ∗ Corresponding author. Tel.: +90 232 4888131; fax: +90 232 4888339. E-mail addresses: [email protected] (S. Parsi), [email protected] (I. Bairamov). 0167-9473/$ – see front matter © 2009 Elsevier B.V. All rights reserved. doi:10.1016/j.csda.2009.03.023 S. Parsi, I. Bairamov / Computational Statistics and Data Analysis 53 (2009) 3560–3570 3561 of Balakrishnan and Basu (1995) and all research has focused on nonparametric and parametric tests of hypotheses. Balakrishnan and Rasouli (2008) have studied the exact likelihood inference for two exponential populations under joint Type-II censoring. If a researcher desires to remove live units at points other than the termination point of the life-test, the above described scheme will be of no value because the joint Type-II censoring does not allow for units to be lost or removed from the test at points other than the final termination point. More general censoring schemes are therefore required. Rasouli and Balakrishnan (in press) consider the Joint Type-II Progressive Censoring (JPC) scheme. Under the JPC scheme, m units of product A with c.d.f. F (x) and p.d.f. f (x) and n units of product B with c.d.f. G(y) and p.d.f. g (y) are placed on test at time zero (thus, we have N = m + n units on test), and k failures from both product A and product B are going to be observed. Immediately following the first failure, r1 of the surviving units from both of product A and product B are randomly selected and removed (we can partition r1 into r10 and r100 such that r10 and r100 are the number of failures of product A and the number of failures of product B respectively (r1 = r10 + r100 )). Then, immediately following the second observed 0 failure, r2 of the surviving units from both product A and product B are removed (r2 = r2 + r200 ), and so on. This experiment terminates at the time when the kth failure is observed and the remaining rk = N − k − r1 − r2 − · · · − rk−1 surviving units from two populations are all removed. If r1 = r2 = · · · = rk−1 = 0, then rk = N − k, which corresponds to the joint Type-II censoring. If r1 = r2 = · · · = rk = 0, then N = k, which corresponds to the complete sample. Several authors have addressed inferential issues based on progressive Type-II censored samples (for example, see Gibbons and Vance (1983), Balakrishnan et al. (1987), Balakrishnan and Sandhu (1995) and Bairamov and Eryilmaz (2006)). Viveros and Balakrishnan (1994) discussed interval estimation of life parameters based on progressive Type-II censored data. Kuş and Kaya (2007) examined a simultaneous confidence interval for parameters of Pareto distribution based on progressive type-II censored data. (For more details see also Balakrishnan and Aggarwala (2000) and Balakrishnan (2007).) Rasouli and Balakrishnan (in press) developed exact inferential methods based on MLEs and compared their performance with those based on approximate, Bayesian and bootstrap methods, based on the JPC scheme under the assumption of exponentiality for both samples. For recent works on progressive censoring see also Balakrishnan and Rasouli (2008), Huang and Wu (2008), Wu (2008), Sanjel and Balakrishnan (2008), Soliman (2008) and Balakrishnan et al. (2008). In practical applications, a researcher may need to know the expected values of the number of failures for each of the two populations. This information is important for the selection of an appropriate sampling plan, because in a statistical inference for parameters, the number of failures for a population are directly related to the estimator efficiency. In this study, we consider a joint Type-II progressive censoring scheme for two populations when the censoring is implemented on the two samples in a combined manner. Section 2 presents the details of the proposed model. The structure of probability of failures is discussed in Section 3. The formula which is most valid for the computation of expected values of the number of failures for each of the two populations is given in Section 4. Finally, a detailed numerical study of the expected values are carried out for various classes of life distributions such as Gamma, Pareto and Weibull. It is important to examine how well the proposed method works for approximation of the expected values (A.E.) of the number of failures. To study the performance of this approach 1,000, JPC samples from different distributions and censoring schemes are simulated and the simulated expected values of the number of failures are obtained. The simulation results are also shown in Tables 1–3 in the Appendix. It can be seen that the approximate expected values of the number of failures are close to the simulated expected values of the number of failures for different distributions and censoring schemes. Hence, the proposed method is reliable for approximating the expected values of the number of failures. 2. The model Suppose m and n independent units are placed in a test with the corresponding lifetimes being identically distributed with p.d.f.’s f (x; θ 1 ) and g (y; θ 2 ) and c.d.f.’s F (x; θ 1 ) and G(y; θ 2 ), θ 1 = (θ11 , θ12 , . . . , θ1d1 ), θ 2 = (θ21 , θ22 , . . . , θ2d2 ), θ 1 ∈ Rd1 , θ 2 ∈ Rd2 , respectively. Under the JPC with censoring scheme R = (R1 = r1 , . . . , Rk = rk ), the observable data consists of 0 (WR , Z, R ) where WR = (W1R:k:N , W2R:k:N , . . . , WkR:k:N ) and Z = (Z1 , . . . , Zk ) with Zi = 1 if WiR is from X (product A) failure 0 0 and Zi = 0 if WiR is from Y (product B) failure and R0 = (R1 = r10 , . . . , R0k−1 = rk−1 ). Pk R R Let Mk = i=1 Zi denote the number of X -failures in W and Nk = k − Mk (i.e., the number of Y -failures in W ). The R 0 likelihood function of (W , Z, R ) is derived by Rasouli and Balakrishnan (in press) as: L(θ 1 , θ 2 ; wR , z, r0 ) = C k Y r 0 r 00 {f (wi ; θ 1 )}zi {g (wi ; θ 2 )}1−zi F̄ (wi ; θ 1 ) i Ḡ(wi ; θ 2 ) i , (1) i=1 where w1 < w2 < · · · < wk , and m0 ≡ n0 ≡ 0 and F̄ = 1 − F and Ḡ = 1 − G are the survival functions of the two 0 populations. Note that rk = N − k − r1 − r2 − · · · − rk−1 , rk0 = m − mk − r10 − r2 − · · · − rk0 −1 . For i = 1, 2, . . . , k, ri00 = ri − ri0 and C = D1 D2 with D1 = k Y j =1 " m− j −1 X i=1 zi − j −1 X i=1 ! 0 ri ! # j −1 j −1 X X 00 zj + n − (1 − zi ) − ri (1 − zj ) , i=1 i=1 3562 S. Parsi, I. Bairamov / Computational Statistics and Data Analysis 53 (2009) 3560–3570 j−1 jP −1 jP −1 0 jP −1 P m− zi − ri n− (1−zi )− ri00 i=1 i=1 i=1 i=1 0 0 rj rj −rj k−1 Y D2 = . ! jP −1 j=1 m+n−j− ri i=1 rj From (1), we have the joint density of (WR , Z, R0 ) as fWR ,Z,R0 (w, z, r0 ; θ 1 , θ 2 ) = C k Y {f (wi ; θ 1 )}zi {g (wi ; θ 2 )}1−zi ri0 F̄ (wi ; θ 1 ) ri00 Ḡ(wi ; θ 2 ) i=1 0 < w1 < w2 < · · · < wk < ∞. In the following theorem, the expressions for the probability of failures are given. Theorem 3.1. The joint density of (WR , Z, R0 ) is 0 fWR ,Z,R0 (w, z, r0 ; θ 1 , θ 2 ) = fWR |Z=z,R0 =r0 (w; θ 1 , θ 2 |Z = z, R0 = r0 )P (Z = z, R0 = r ), where 0 P (Z = z, R0 = r ) = αk k−1 Y αi βi , (2) i=1 and mz1 + n (1 − z1 ) α1 = P (Z1 = z1 ) = mp + nq 0 β1 = P (R1 = r1 |Z1 = z1 ) = 0 m−m1 0 r1 pz1 q1−z1 , n−n1 r1 −r10 N −1 r1 z1 = 0 or 1, (3) , r10 = max(0, r1 + n1 − n), . . . , min(r1 , m − m1 ), (4) and for i = 2, . . . , k, zi = 0 or 1, αi = P (Zi = zi |Z1 = z1 , . . . Zi−1 = zi−1 , R01 = r10 , . . . R0i−1 = ri0−1 ) ! ! iP −1 iP −1 0 00 m − mi−1 − rj zi + n − ni−1 − rj (1 − zi ) j =1 = j =1 i −1 m − mi−1 − P ! 0 rj i −1 p+ n − ni−1 − j=1 P pzi q1−zi . ! 00 rj (5) q j =1 For i = 2, . . . , k − 1, with r00 ≡ r000 ≡ 0, 0 βi = P (R0i = ri0 |Z1 = z1 , . . . Zi = zi , R01 = r10 , . . . , Ri−1 = ri0−1 ) i−1 i−1 m−mi − P 0 j=1 0 ri P 00 r n−ni − rj j=1 = N− iP −1 r j −i 0 ri −ri ! j , j=1 ri ri0 = max(0, ri + ni + i−1 X rj00 − n), . . . , min(ri , m − mi − ri0−1 − · · · − r10 ), j =1 where p = P (Z = 1) and q = 1 − p. (6) S. Parsi, I. Bairamov / Computational Statistics and Data Analysis 53 (2009) 3560–3570 3563 3. Expected value of Mk Pk R With the knowledge of the expected value of Mk = i=1 Zi , the number of X -failures in W presents data of interest. In this section, by using Theorem 3.1, we will be able to approximate the expected value of Mk . We will need the Taylor expansion of the function in several variables (see, Adams (2007)). For the d-dimensional function T (x1 , . . . , xd ) we have T (x1 , . . . , xd ) = T (a1 , . . . , ad ) + d X (xi − ai )Txi (a1 , . . . , ad ) + R2 , i =1 where Txi denotes the partial derivative with respect to xi and R2 is the remainder, which reduces to the tangent plane approximation ≈ T (x1 , . . . , xd ) ≈ T (x1 , . . . , xd ) = T (a1 , . . . , ad ) + d X (xi − ai )Txi (a1 , . . . , ad ), i =1 ≈ where T (x1 , . . . , xd ) is the Taylor polynomial of T (x1 , . . . , xd ) of degree 1 near the point (a1 , . . . , ad ). In the special case, when d = 2 and the function depends on two variables, x1 and x2 , the Taylor polynomial near the point (a1 , a2 ) is: T (x1 , x2 ) = T (a1 , a2 ) + (x1 − a1 )Tx1 (a1 , a2 ) + (x2 − a2 )Tx2 (a1 , a2 ) + R2 and ≈ T (x1 , x2 ) ≈ T (x1 , x2 ) = T (a1 , a2 ) + (x1 − a1 )Tx1 (a1 , a2 ) + (x2 − a2 )Tx2 (a1 , a2 ). (7) From (2)–(6), we have µZ1 ≡ E (Z 1) = m mp + nq p (8) 0 r (m − m ) 1 1 µR01 |Z1? ≡ E R1 |Z1 = z1 = . (N − 1) (9) Conditioning with respect to Z1 and Z2 gives µZ ? 0? 2 |Z1 ;R1 ≡ E Z2 |Z1 = z1 , R01 = r10 m − m1 − r10 = p, (10) m − m1 − r10 p + n − n1 − r100 q 0 µR0 |Z ? ;R0 ? ≡ E R02 |Z1 = z1 , Z2 = z2 , R1 = r10 2 2 1 r2 m − m2 − r10 . = N − r10 − 2 (11) Continuing this process we obtain µZ | Z ? i 0 ;R ? i−1 i−1 0 ≡ E Zi |Z1 = z1 , . . . , Zi−1 = zi−1 , R01 = r1 , . . . , R0i−1 = ri0−1 ! iP −1 0 m − mi−1 − rj j =1 = i−1 m − mi−1 − P ! 0 rj i−1 p+ n − ni−1 − j =1 P ! p, 00 rj i = 3, . . . , k (12) q j =1 and 0 µR0 |Z ? ;R0 ? ≡ E R0i |Z1 = z1 , . . . , Zi = zi , R01 = r10 , . . . , Ri−1 = ri0−1 i i i−1 ! iP −1 0 ri m − mi − rj j=1 = i −1 N− P j =1 0 ! , rj − i i = 3, . . . , k. (13) 3564 S. Parsi, I. Bairamov / Computational Statistics and Data Analysis 53 (2009) 3560–3570 From (8) and (9), it can be seen immediately that m µZ1 = E (Z 1) = p mp + nq h 0 µR01 = EZ1 ER1 R1 |Z1 = z1 i r1 m − = m p mp+nq . (N − 1) However, it can be observed from (10)–(13), that it is not easy to get exact expressions for µZi and µR0 for i = 2, 3, . . . , k. By i using the Taylor expansion of the function of several variables, we can approximate them. For approximating µZ2 , we define ≈ ≈ µZ1 = µZ1 , µR0 = µR0 and 1 1 Ψ (z1 , r10 ) = µZ 2 0 = E Z2 |Z1 = z1 , R01 = r1 0 |Z ? ; R ? 1 1 m − m1 − r10 = p. 0 m − m1 − r1 p + n − n1 − r100 q The function Ψ (z1 , r10 ) depends on two variables, z1 and r10 . We use the Taylor expansion for the function Ψ (z1 , r10 ) ≈ ≈ substituting (a1 , a2 ) = (µZ1 , µR0 ) in (7). Thus, we have 1 ≈ ≈ µ µ m − Z1 − R0 1 ≈ p Ψ (z1 , r10 ) = ≈ ≈ ≈ ≈ m − µZ1 − µR0 p + n − (1 − µZ1 ) − (r1 − µR0 ) q 1 1 ≈ ≈ ≈ ≈ ≈ ≈ + (z1 − µZ1 )Ψz1 (µZ1 , µR0 ) + (r10 − µR01 )Ψr10 (µZ1 , µR01 ). 1 By applying a repeated conditional expectation formula, we have ≈ ≈ µZ2 = EZ1 [E 0 (Ψ (z1 , r10 ))] R 1 ≈ ≈ m − µZ 1 − µR 0 1 = ≈ ≈ m − µZ1 − µR0 p+ 1 ≈ p. ≈ n − (1 − µZ1 ) − (r1 − µR0 ) q 1 Since the expected values after 2 terms vanish, when approximating µR0 this method yields 2 2 ≈ r2 m− ≈ P j =1 µR 0 = 2 ! ≈ µZ j − µR 0 1 . ≈ µ N − R0 − 2 1 Thus, by using similar considerations, we get the approximate values of µZi and µR0 for i = 3, 4, . . . , k as follows: i m− ≈ µZi = m− i−1 ≈ i−1 ≈ j =1 j =1 P µZ j − P µ i −1 ≈ i −1 ≈ j=1 j =1 P µZj − ! i −1 p+ 0 Rj P n− P µ ! 0 Rj ≈ (1 − µZj ) − j=1 i−1 P ≈ ! p i = 3, . . . , k. (rj − µR0j ) q j =1 And ri ≈ i ≈ iP −1 ≈ P µZj − µ 0 m− R j=1 µR 0 = i j =1 i −1 ≈ N− P j =1 ! j ! i = 3, . . . , k. µR 0 − i j Therefore, the approximate value for the expected value of the number of failures from X ’s is E (Mk ) ≈ k X ≈ µZ i . i =1 (14) S. Parsi, I. Bairamov / Computational Statistics and Data Analysis 53 (2009) 3560–3570 3565 3.1. Some special distributions To evaluate the expected value of Mk for special distributions such as Gamma, Pareto and Weibull, we need to compute the value of p given in (15) for these distributions. Gamma distribution. Let X and Y be independent random variables with Gamma p.d.f.’s Gamma(α1 , β1 ) and Gamma(α2 , β2 ), respectively. The p.d.f.’s of X and Y are f (x; α1 , β1 ) = xα1 −1 x I (x > 0) α exp − α x exp − I (x > 0), β2 Γ (α1 )β1 1 β1 and g (x; α2 , β2 ) = xα2 −1 Γ (α2 )β2 2 where I is the indicator function. It is not difficult to verify that, in this case, I Pg = P (Z = 1) = P (X < Y ) = β2 β1 +β2 (α1 , α2 ) B(α1 , α2 ) , where B(a, b) is the beta function and Ix (a, b) is the incomplete beta function. The expected value of Mk , in the case where X ’s have Gamma(α1 , β1 ) and Y ’s have Gamma(α2 , β2 ) distributions, respectively, under the setup of the JPC scheme, can be calculated using (14) by substituting p = Pg . Pareto distribution. Let X and Y be independent random variables with Pareto p.d.f.’s Pareto(λ1 , µ1 ), and Pareto(λ2 , µ2 ), respectively. The p.d.f.’s of X and Y are λ f (x; λ1 , µ1 ) = λ1 µ1 1 I (x > µ1 ) xλ1 +1 and λ g (x; λ2 , µ2 ) = λ2 µ2 2 xλ2 +1 I (x > µ2 ). In this case, we have λ1 µ1 λ2 , λ1 +λ2 µ2 Pp = P (Z = 1) = P (X < Y ) = λ1 µ2 λ2 , λ1 + λ2 µ1 1 − if µ1 < µ2 if µ1 > µ2 . The expected value of Mk for two Pareto distributions under the setup of the JPC scheme can be evaluated from (14) by substituting p = Pp . Weibull distribution. Let X and Y be independent random variables with corresponding Weibull p.d.f.’s, Weibull(β1 , α1 ) and Weibull(β2 , α2 ), where αi , i = 1, 2, are the scale parameters and βi , i = 1, 2, are the shape parameters. The p.d.f.’s of X and Y are given as β1 f (x; β1 , α1 ) = α1 β2 g (x; β2 , α2 ) = α2 x β 1 −1 " # β1 x exp − I (x > 0) α1 β2 −1 " # β2 x exp − I (x > 0). α2 α1 and x α2 In this case, we have ∞ X (−1)j α2 jβ1 β1 1 − Γ j + 1 , j! α1 β2 j=0 Pw = P (Z = 1) = P (X < Y ) = ∞ X (−1)j α2 jβ1 β1 Γ j+1 , j! α1 β2 j=0 if β1 < β2 if β1 > β2 . The expected value of Mk for these two Weibull distributions under the JPC scheme can be found by substituting p = Pw in (14). 3566 S. Parsi, I. Bairamov / Computational Statistics and Data Analysis 53 (2009) 3560–3570 Fig. 1. The approximate expected values of the number of failures. 4. Computation results and simulation study In this section, some numerical results on calculation of the expected value of Mk are presented. We have considered different sample sizes for the two populations from different distributions and censoring schemes. The expected values for Mk , calculated from the formula (14) for Gamma, Weibull and Pareto distributions mentioned above, are presented in Tables 1–3 in the Appendix. Table 1 consists of the approximate values of E (Mk )− the expected number of failures from X ’s in the JPC scheme, for observations X1 , X2 , . . . , Xm and Y1 , Y2 , . . . , Yn from populations having Gamma(α1 , β1 ) and Gamma(α2 , β2 ) distributions respectively, with the censoring scheme R = (r1 , . . . , rk ). Similarly, in Table 2, the numerical values of Mk for two Pareto(λ1 , µ1 ) and Pareto(λ2 , µ2 ) distributions and in Table 3 for two Weibull(β1 , α1 ) and Weibull(β2 , α2 ) distributions are given. It is important to examine the effectiveness of the proposed method in approximating the expected values of the number of failures. To study the performance of this approach, we also simulate 1,000 JPC samples from different distributions and different combinations of N, k, m, n, p and censoring schemes R and obtain the simulated expected values of the number of failures. The simulation results are also shown in Tables 1–3. The approximate expected values of the number of failures are close to the simulated expected values of the number of failures, for different distributions and censoring cases. Hence, this method is reliable where it is necessary to obtain the approximate values of the expected values of the number of failures. In practical applications, it is very important for the researcher to have information about the expected values of the number of failures for the two populations before the experiment. Because in statistical inference procedures for parameters the number of failures for a population is directly related to estimators efficiency, information about the expected values of the number of failures for the two populations before the experiment is extremely important in the selection of a sampling plan. From Table 1, for example, for X and Y having Gamma distributions with parameters (α1 , β1 ) = (4, 12), (α2 , β2 ) = (6, 4), respectively, and p = 0.166, k = 10, m = 15, n = 15, if we choose the censoring scheme (0, . . . , 0, 20), then we expect that approximately 2 failure times will be from the X sample. In the case of the censoring scheme (4, 0, 4, 0, 4, 0, 4, 0, 4, 0), we will expect approximately 4 observations of failure times from the X sample. From Table 2, for X and Y with Pareto distributions with parameters (λ1 , µ1 ) = (3, 6) and (λ2 , µ2 ) = (4, 4), respectively, for m = 50, n = 50, p = 0.085, k = 50, if the ordinary joint Type-II scheme is selected as (0, . . . , 0, 50), then approximately 8 observations of failure times are expected from the X sample, whereas for the JPC scheme (50, 0, . . . , 0), this number increases to 25. Therefore, for inference procedures for the parameters of X sample, we should prefer this scheme. From Table 3, with X having Weibull distribution with parameters (β1 , α1 ) = (5, 3) and Y having Weibull distribution with parameters (β2 , α2 ) = (5, 5) we observe that for m = 10, n = 20, k = 10, p = 0.928 in an ordinary joint Type-II censoring scheme (0, 0, . . . , 0, 20), approximately 8 observations are expected from the X sample and 2 from the Y sample. However, if we choose the JPC scheme (10, 10, 0, . . . , 0) we would expect 5 observations of failure times from the X sample and 5 from the Y sample. This means that if m and n are small and q less than p, and we employ an ordinary Type-II censoring scheme, then it is quite likely that we will observe very few failures from the second population in our censored sample, thus resulting in poor inference for the parameters of the second population. A JPC scheme balances this situation and results in failures from both populations, thus yielding more accurate inference for both parameters of the two populations. For example, Fig. 1, below, shows the approximate expected values of the number of failures with respect to p ∈ (0, 1) for S. Parsi, I. Bairamov / Computational Statistics and Data Analysis 53 (2009) 3560–3570 3567 Table 1 The Approximation of the Expected (A.E.) and Simulated Expected (S.E.) values of Mk for some small, moderate and large values of m, n, k and Pg for two Gamma populations when joint Type-II progressive censoring is implemented. Scheme no. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 526 53 54 55 56 57 N 30 k Scheme(R) (m, n) (α1 , α2 , β1 , β2 ) = (4, 6, 12, 4), p = 0.166 A.E . Mk S .E . Mk (α1 , α2 , β1 , β2 ) = (2, 5, 7, 3), p = 0.580 A.E . Mk S .E . Mk (α1 , α2 , β1 , β2 ) = (3, 5, 2, 3), p = 0.904 A.E . Mk S .E . Mk 10 (0, . . ., 0, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (30, 20) (25, 25) (20, 30) (30, 20) (25, 25) (20, 30) (30, 20) (25, 25) (20, 30) (30, 20) (25, 25) (20, 30) (70, 30) (50, 50) (30, 70) (70, 30) (50, 50) (30, 70) (70, 30) (50, 50) (30, 70) (70, 30) (50, 50) (30, 70) 3.568 2.038 1.083 6.389 4.710 3.023 7.246 5.711 3.940 3.567 2.037 1.083 5.917 4.861 3.419 5.413 3.550 2.287 6.316 3.584 1.865 9.802 7.313 4.813 7.898 4.805 2.427 10.184 5.965 3.018 13.202 9.882 6.560 6.143 4.358 3.014 11.773 9.792 7.813 8.530 6.030 4.137 14.811 12.329 9.849 7.235 3.737 1.741 13.691 9.727 5.763 23.637 12.142 5.433 34.807 24.831 14.888 7.238 5.665 3.944 6.794 5.125 3.432 7.860 6.256 4.375 7.237 5.664 3.944 8.267 6.706 4.771 7.831 6.440 4.846 10.747 8.356 5.774 10.115 7.607 5.080 10.778 8.378 5.780 14.144 10.910 7.471 13.436 10.089 6.731 13.178 11.247 9.217 12.120 10.111 8.096 16.342 13.913 11.371 15.110 12.600 10.086 15.150 11.440 7.273 14.131 10.127 6.092 37.307 27.806 17.402 35.109 25.092 15.061 9.369 8.725 7.456 7.020 5.414 3.728 8.309 6.743 4.828 9.368 8.724 7.456 9.565 8.083 6.329 9.495 8.919 7.123 13.863 12.532 9.652 10.272 7.720 5.265 13.556 11.178 7.843 17.972 14.884 9.999 13.510 10.160 6.836 18.159 17.175 15.662 12.232 10.269 8.289 22.364 20.822 18.354 15.187 12.712 10.236 19.014 17.747 15.074 14.270 10.347 6.406 46.723 41.546 29.080 35.130 25.204 15.253 (20, 0, . . ., 0) (10, 10, 0, . . ., 0) (0, . . ., 0, 10, 10) (0, 0, 0, 0, 10, 10, 0, 0, 0, 0) (4, 0, 4, 0, 4, 0, 4, 0, 4, 0) 15 (0, . . ., 0, 15) (15, 0, . . ., 0) (5, 0, . . ., 0, 5, 0, . . . 0, 5) 20 (0, . . ., 0, 10) (10, 0, . . ., 0) 50 20 (0, . . ., 0, 30) (0, . . ., 0, 30) 25 (0, . . ., 0, 25) (25, 0, . . ., 0) 100 20 (0, . . ., 0, 80) (80, 0, . . ., 0) 50 (0, . . ., 0, 50) (50, 0, . . ., 0) 3.562 2.020 1.080 6.433 4.753 3.123 6.161 4.567 3.001 3.586 2.103 1.095 4.958 3.534 2.189 4.529 2.928 1.664 6.432 3.716 1.828 9.870 7.346 4.886 7.603 4.697 2.354 10.312 6.146 3.063 13.133 9.854 6.537 6.178 4.371 3.044 11.861 9.840 7.940 8.619 6.130 4.143 14.774 12.302 9.906 7.174 3.736 1.754 13.617 9.673 5.897 23.629 12.149 5.440 34.849 24.741 14.796 7.246 5.676 3.903 6.639 5.071 3.357 6.724 5.031 3.437 7.228 5.607 3.944 6.911 5.332 3.701 7.069 5.490 3.754 10.764 8.381 5.784 10.013 7.632 5.063 10.488 8.122 5.579 14.104 10.885 7.470 13.364 10.063 6.745 13.215 11.250 9.320 12.151 10.009 8.061 16.273 13.840 11.417 15.014 12.474 10.138 15.133 11.491 7.318 13.969 10.027 6.205 37.252 27.684 17.349 35.104 24.981 15.038 9.365 8.719 7.438 6.953 5.219 3.715 7.036 5.454 3.950 9.308 8.722 7.374 8.029 6.930 5.560 8.713 7.668 6.072 13.849 12.490 9.467 10.124 7.702 5.270 13.187 10.729 7.684 17.941 14.647 9.953 13.392 10.109 6.854 18.132 17.229 15.637 12.245 10.342 8.341 22.378 20.741 18.225 15.157 12.769 10.192 18.970 17.711 15.068 14.225 10.348 6.397 46.748 41.423 28.902 35.225 25.092 15.260 N = 30, k = 10, m = 20, n = 10 for two cases R = (0, . . . , 0, 20) (joint Type-II censoring) and R = (20, 0, . . . , 0) (JPC) schemes. It is easy to see that the JPC scheme balances the situation. 3568 S. Parsi, I. Bairamov / Computational Statistics and Data Analysis 53 (2009) 3560–3570 Table 2 The Approximation of the Expected (A.E.) and Simulated Expected (S.E.) values of Mk for some small, moderate and large values of m, n, k and Pp for two Pareto populations when joint Type-II progressive censoring is implemented. Scheme no. N k Scheme(R) (m, n) (λ1 , λ2 , µ1 , µ2 ) = (3, 4, 6, 4), p = 0.085 A.E . Mk S .E . Mk (λ1 , λ2 , µ1 , µ2 ) = (4, 2, 7, 5), p = 0.340 A.E . Mk S .E . Mk (λ1 , λ2 , µ1 , µ2 ) = (3, 7, 3, 5), p = 0.849 A.E . Mk S .E . Mk 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 30 10 (0, . . ., 0, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (30, 20) (25, 25) (20, 30) (30, 20) (25, 25) (20, 30) (30, 20) (25, 25) (20, 30) (30, 20) (25, 25) (20, 30) (70, 30) (50, 50) (30, 70) (70, 30) (50, 50) (30, 70) (70, 30) (50, 50) (30, 70) (70, 30) (50, 50) (30, 70) 2.352 1.145 0.559 6.293 4.553 2.922 7.044 5.510 3.902 2.352 1.145 0.559 5.140 4.328 2.966 4.643 2.929 1.729 5.212 2.266 1.016 9.722 7.218 4.682 7.594 4.423 1.598 10.000 5.035 1.852 13.156 9.845 6.463 4.005 2.557 1.646 11.705 9.727 7.720 6.317 3.843 2.378 14.755 12.282 9.794 4.498 2.009 0.870 13.577 9.641 5.755 20.583 7.785 2.924 34.736 24.790 14.865 5.387 3.656 2.218 6.536 4.821 3.104 7.512 5.880 4.005 5.386 3.656 2.218 7.142 5.639 3.839 6.467 4.823 3.354 8.404 5.763 3.517 9.907 7.375 4.819 8.979 6.307 3.911 11.727 8.160 5.025 13.272 9.916 6.529 9.492 7.473 5.655 11.887 9.873 7.850 12.191 9.632 7.308 14.907 12.397 9.878 11.291 7.100 3.794 13.870 9.841 5.791 30.042 19.304 10.432 34.921 24.911 14.889 9.013 8.118 6.629 6.962 5.298 3.632 8.075 6.511 4.664 9.013 8.118 6.629 9.334 8.100 6.105 9.214 8.427 6.817 13.286 11.636 8.891 10.215 7.698 5.211 12.997 10.838 7.656 17.180 14.233 9.892 13.438 10.124 6.806 17.229 15.950 14.209 12.200 10.218 8.239 21.167 19.336 16.845 15.152 12.679 10.199 18.424 16.572 13.212 14.227 10.286 6.329 45.018 38.585 26.970 35.117 25.176 15.205 (20, 0, . . ., 0) (10, 10, 0, . . ., 0) (0, . . ., 0, 10, 10) (0, 0, 0, 0, 10, 10, 0, 0, 0, 0) (4, 0, 4, 0, 4, 0, 4, 0, 4, 0) 15 (0, . . ., 0, 15) (15, 0, . . ., 0) (5, 0, . . ., 0, 5, 0, . . . 0, 5) 20 (0, . . ., 0, 10) (10, 0, . . ., 0) 50 20 (0, . . ., 0, 30) (0, . . ., 0, 30) 25 (0, . . ., 0, 25) (25, 0, . . ., 0) 100 20 (0, . . ., 0, 80) (80, 0, . . ., 0) 50 (0, . . ., 0, 50) (50, 0, . . ., 0) 2.366 1.117 0.547 6.337 4.698 3.158 6.025 4.546 2.992 2.454 1.157 0.551 4.360 2.993 1.991 3.836 2.267 1.266 5.411 2.349 0.966 9.714 7.292 4.884 7.224 4.130 1.610 10.025 5.247 1.869 13.090 9.871 6.563 4.064 2.563 1.635 11.880 9.756 7.842 6.514 3.908 2.376 14.812 12.297 9.878 4.434 2.051 0.851 13.500 9.640 5.891 20.709 7.849 2.901 34.829 24.708 14.803 5.345 3.672 2.228 6.558 4.918 3.204 6.521 4.808 3.190 5.451 3.672 2.186 6.041 4.319 2.737 5.670 4.061 2.494 8.371 5.762 3.585 9.858 7.508 4.949 8.864 6.236 3.788 11.756 8.256 5.057 13.242 9.941 6.716 9.572 7.490 5.653 11.909 9.907 7.885 12.079 9.709 7.344 14.903 12.441 9.888 11.259 7.143 3.847 13.913 9.916 5.897 30.001 19.353 10.417 34.937 25.015 14.904 9.009 8.084 6.572 6.792 5.247 3.635 7.047 5.427 3.855 9.001 8.118 6.578 7.843 6.608 5.076 8.462 7.273 5.560 13.275 11.636 8.777 10.204 7.649 5.217 12.660 10.412 7.430 17.093 14.096 9.770 13.451 10.129 6.771 17.258 15.937 14.233 12.288 10.274 8.393 21.097 19.249 16.800 15.180 12.628 10.258 18.456 16.599 13.283 14.116 10.323 6.349 44.993 38.529 26.857 35.106 25.120 15.324 Acknowledgements The authors are grateful to two anonymous referees, the Co-Editor and Associate editor for the careful reading and constructive suggestions, which were helpful in improving the presentation. S. Parsi, I. Bairamov / Computational Statistics and Data Analysis 53 (2009) 3560–3570 3569 Table 3 The Approximation of the Expected (A.E.) and Simulated Expected (S.E.) values of Mk for some small, moderate and large values of m, n, k and Pw for two Weibull populations when joint Type-II progressive censoring is implemented. Scheme no. N k Scheme(R) (m, n) (β1 , β2 , α1 , α2 ) = (4, 5, 3, 2), p = 0.160 A.E . Mk S .E . Mk (β1 , β2 , α1 , α2 ) = (6, 2, 3, 3), p = 0.431 A.E . Mk S .E . Mk (β1 , β2 , α1 , α2 ) = (5, 5, 3, 5), p = 0.928 A.E . Mk S .E . Mk 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 30 10 (0, . . ., 0, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (20, 10) (15, 15) (10, 20) (30, 20) (25, 25) (20, 30) (30, 20) (25, 25) (20, 30) (30, 20) (25, 25) (20, 30) (30, 20) (25, 25) (20, 30) (70, 30) (50, 50) (30, 70) (70, 30) (50, 50) (30, 70) (70, 30) (50, 50) (30, 70) (70, 30) (50, 50) (30, 70) 3.490 1.976 1.045 6.379 4.704 3.025 7.236 5.692 3.919 3.490 1.975 1.044 5.866 4.823 3.387 5.382 3.503 2.247 6.234 3.497 1.805 9.797 7.306 4.799 7.875 4.761 2.370 10.151 5.886 2.939 13.198 9.882 6.571 6.004 4.236 2.917 11.768 9.791 7.805 8.382 5.885 4.016 14.807 12.326 9.849 7.059 3.614 1.675 13.683 9.724 5.762 23.394 11.855 5.253 34.802 24.828 14.886 6.143 4.427 2.839 6.578 4.890 3.224 7.605 5.962 4.107 6.142 4.427 2.838 7.651 6.006 4.126 7.008 5.442 3.901 9.335 6.762 4.355 9.927 7.405 4.900 9.656 7.095 4.625 12.640 9.216 5.968 13.274 9.919 6.577 10.954 8.925 6.983 11.912 9.900 7.894 13.821 11.283 8.842 14.921 12.409 9.902 12.911 8.759 5.002 13.917 9.887 5.886 32.936 22.583 13.007 34.942 24.915 14.904 9.525 9.012 7.891 7.147 5.450 3.762 8.423 6.844 4.857 9.525 9.011 7.890 9.669 8.285 6.633 9.619 9.146 6.940 14.128 12.985 9.919 10.335 7.818 5.316 13.829 11.285 7.942 18.365 15.006 10.000 13.476 10.196 6.856 18.588 17.773 16.414 12.268 10.307 8.319 22.938 21.575 19.078 15.210 12.725 10.258 19.265 18.283 16.038 14.349 10.386 6.444 47.499 43.050 29.728 35.141 25.216 15.276 (20, 0, . . ., 0) (10, 10, 0, . . ., 0) (0, . . ., 0, 10, 10) (0, 0, 0, 0, 10, 10, 0, 0, 0, 0) (4, 0, 4, 0, 4, 0, 4, 0, 4, 0) 15 (0, . . ., 0, 15) (15, 0, . . ., 0) (5, 0, . . ., 0, 5, 0, . . . 0, 5) 20 (0, . . ., 0, 10) (10, 0, . . ., 0) 50 20 (0, . . ., 0, 30) (0, . . ., 0, 30) 25 (0, . . ., 0, 25) (25, 0, . . ., 0) 100 20 (0, . . ., 0, 80) (80, 0, . . ., 0) 50 (0, . . ., 0, 50) (50, 0, . . ., 0) 3.543 1.996 1.057 6.393 4.788 3.150 6.171 4.597 3.021 3.549 2.020 1.048 4.992 3.449 2.186 4.466 2.816 1.607 6.371 3.563 1.854 9.790 7.327 4.859 7.721 4.665 2.307 10.263 6.035 2.960 13.214 9.945 6.656 6.067 4.313 3.006 11.642 9.696 7.803 8.473 5.949 4.010 14.800 12.363 9.863 6.959 3.500 1.645 13.771 9.682 5.820 23.529 11.884 5.291 34.686 24.787 14.862 6.179 4.458 2.794 6.707 4.923 3.396 6.583 4.916 3.225 6.157 4.483 2.790 6.340 4.675 3.065 6.284 4.527 2.963 9.305 6.736 4.348 9.950 7.442 4.941 9.469 7.000 4.508 12.602 9.202 5.968 13.324 9.943 6.567 10.921 8.885 7.013 11.988 9.909 7.944 13.803 11.208 8.830 14.970 12.524 9.924 12.874 8.771 4.960 13.982 9.996 5.942 32.920 22.490 13.057 35.071 24.928 14.946 9.544 9.033 7.873 6.846 5.316 3.676 7.043 5.523 3.942 9.571 9.019 7.737 8.120 7.039 5.785 8.910 7.889 6.302 14.070 13.015 9.739 10.122 7.698 5.294 13.396 10.940 7.871 18.302 14.856 9.986 13.446 10.171 6.830 18.599 17.704 16.307 12.216 10.247 8.206 22.935 21.588 18.896 15.172 12.682 10.214 19.254 18.327 16.058 14.053 10.350 6.476 47.552 42.966 29.564 35.126 25.110 15.289 Appendix Proof of Theorem 3.1. We can obtain αi from the fact that in a population containing two members (one X and one Y ), the probability of observing X first is p = P (Z = 1) = P (X < Y ) = Z 0 ∞ Ḡ(x; θ 2 )dF (x; θ 1 ). (15) 3570 S. Parsi, I. Bairamov / Computational Statistics and Data Analysis 53 (2009) 3560–3570 Now for the m observations of X ’s, n observations of Y ’s and k times censoring using Bayes’ theorem with prior probabilities p/(p + q) for z = 1 and q/(p + q) for z = 0 and with mass probabilities a/(a + b) for z = 1 and b/(a + b) for z = 0 for each step, we have the posterior mass probabilities ap/(ap + bq) for z = 1 and bp/(ap + bq) for z = 0. Here in the jth step for derivation of probability 0 P (Zj = zj |Z1 = z1 , . . . Zj−1 = zj−1 , R1 = r10 , . . . , R0j−1 = rj0−1 ), the numbers a and b are taken as aj = m − mj−1 − Pj−1 0 i=1 rj , bj = n − nj−1 − Pj−1 00 i =1 r j . Also, we can obtain βi from the fact that in the ith step, R0i has a hypergeometric distribution with parameters m − mi − Pi−1 0 j =1 r j , n − ni − Pi−1 00 j =1 r j and ri . Note that in the literature the probability p = P (X < Y ) is called stress–strength reliability. In recent years the problem of estimating p has been given much attention in reliability and related fields. For a comprehensive review of results on estimating stress–strength reliability we refer Kotz et al. (2003). References Adams, R.A., 2007. Calculus. Addison Wesley, Longman. Bairamov, I., Eryilmaz, S., 2006. Spacings, exceedances and concomitants in progressive type censoring scheme. J. Statist. Planning Inference 136, 527–536. Balakrishnan, N., 2007. Progressive censoring methodology: An appraisal. Test 16, 211–259. Balakrishnan, N., Aggarwala, R., 2000. Progressive Censoring: Theory, Methods and Applications. Birkhauser, Boston. Balakrishnan, N., Basu, A.P. (Eds.), 1995. The Exponential Distribution: Theory, Methods and Applications. Gordon and Breach, Newark, NJ. Balakrishnan, N., Kannan, R.A., Lin, C.T., Ng, H.K.T., 1987. A point and interval estimation for Gaussian distribution based on progressively type-II censored samples. IEEE Trans. Reliab. 52 (1), 90–95. Balakrishnan, N., Rasouli, A., 2008. Exact likelihood inference for two exponential populations under joint Type-II censoring. Comput. Statist. Data Anal. 52, 2725–2738. Balakrishnan, N., Sandhu, R.A., 1995. A simple simulation algorithm for generating progressively type-II censored sample. Amer. Statistician 49, 229–230. Balakrishnan, N., Burkschat, M., Cramer, E., Hofmann, G., 2008. Fisher information based progressive censoring plans. Comput. Statist. and Data Anal. 53 (2), 366–380. Basu, A.P., 1968. On a generalized Savage statistic with applications to life testing. Ann. Math. Statist. 39, 1591–1604. Bhattacharyya, G.K., 1995. Inferences under two-sample and multi-sample situations. In: Balakrishnan, N., Basu, A.P. (Eds.), The Exponential Distribution: Theory, Methods and Applications. Gordon and Breach, Newark, NJ, pp. 93–118 (Chapter 7). Bhattacharyya, G.K., Mehrotra, K.G., 1981. On testing equality of two exponential distributions under combined Type-II censoring. J. Amer. Statist. Assoc. 76, 886–894. Gibbons, D.I., Vance, L.C., 1983. Estimators for the 2-parameter Weibull distribution with progressively censored samples. IEEE Trans. Reliab. R-32, 95–99. Huang, S.R., Wu, S.J., 2008. Reliability sampling plans under progressive type-I interval censoring using cost functions. IEEE Trans. Reliab. 57 (3), 445–451. Johnson, R.A., Mehrotra, K.G., 1972. Locally most powerful rank tests for the two-sample problem with censored data. Ann. Math. Statist. 43, 823–831. Kuş, C., Kaya, M.F., 2007. Estimation for the parameters of the Pareto distribution under progressive censoring. Commun. Statist. Theor. Meth. 36, 1359–1365. Kotz, S., Lumelskii, Y., Pensky, M., 2003. The Stress-Strength Model and its Generalizations. Theory and Applications. World Scientific. Rasouli, A., Balakrishnan, N., 2009. Exact likelihood inference for two exponential populations under joint progressive type-II censoring. Commun. Statist. Theory Methods (in press). Sanjel, D., Balakrishnan, N., 2008. A Laguerre polynomial approximation for a goodness-of-fit test for exponential distribution based on progressively censored data. J. Stat. Comput. and Simul. 78 (6), 503–513. Soliman, A.A., 2008. Estimations for pareto model using general progressive censored data and asymmetric loss. Commun. Statist. Theory Methods 37 (9), 1353–1370. Viveros, R., Balakrishnan, N., 1994. Interval estimation of parameters of life from progressively censored data. Technometrics 36, 84–91. Wu, S.J., 2008. Estimation of the two-parameter bathtub-shaped lifetime distribution with progressive censoring. J. Appl. Statist. 35 (10), 1139–1150.
© Copyright 2024 Paperzz