Exponential Selection Problems Under Progressive Type II Censoring By Lii-Yuh Leu Department of Industrial Engineering and management, St. John’s University, Tamsui, Taipei, Taiwan,25135, R.O.C.,E-mail address: leu@mail.sju.edu. Abstract This paper considers the selection problems of exponential distributions under progressive type II censoring mechanism. The selection problems of the largest location and scale parameters are proposed. The problems concerning with good populations and bad populations are considered. Simultaneous confidence intervals, that can be derived with the proposed selection procedure , are discussed. Simultaneous upper confidence bounds for all distances of the parameters from the largest one are proposed. Keywords: subset selection, exponential distribution, good population, bad population, multiple comparison with best. Let i , i 1,..., k parameter i be k exponential populations with location and scale parameter i and density function: 1 e x p (x (i i) x/ ) ,i i f ( x :i ,i )i 0 , o t h e r w i s e . We denote the distribution by Let Xij , j 1,..., n be i , i 1,..., k ,under , 0 E( i ,i ) . i.i.d. samples from population progressive type II censoring scheme ( R1 ,..., Rm ) ,we will let Rm ) Yij X ij( :Rm1 :,..., , j 1,...m, i 1,..., k . n We have Z1 n(Yi1 i ) / i , Z 2 (n R1 1)(Yi 2 Yi1 ) / i ,..., Z m (n R1 R2 ... Rm 1 m 1)(Yim Yi ( m 1) ) / i are i.i.d. E(0,1) We will consider the cases for known or unknown parameters. 1. known, unknown If that are known and 0 .For are unknown, we may assume the selection problem, let the ordered value of ' s. [1] [2] ... [ k ] Our goal is to select t (1 t k 1) best populations associated with the t largest scale parameters [i ] , i k t 1,..., k . 1.1 Indifference zone formulation Let ( ) { (1 ,..., k ) | [ k t ] / [ k t 1] }, 0 1 fixed be the preference zone. Under progressive type II censoring scheme ( R1 ,..., Rm ) , i ˆi is estimated by 1 m ( R j 1Y) m j 1 We have 2mˆi / i ~ 22 .m . i j be Let Yi mˆi ,then Yi ~ Gamma(m,i ) , the proposed selection procedure is R: select t populations corresponding to Y [ k ] , Y [ k 1] ,..., and Y [ k t 1] . Let Y( i ) be the unknown observed sample corresponding to the parameter [i ] , then P (CS | R) P (max(Y(1) ,..., Y( k t ) ) min(Y( k t 1) ,..., Y( k ) )) The least favorable configuration in ( ) is given by [1] ... [ k t ] ,[ k t 1] ... [ k ] and [ k t ] [ k t 1] . Thus inf P (CS | R) (k t ) Fmk t 1 (u )(1 Fm ( u ))t dFm (u ) ( ) 0 Where Fm (u ) For t=1, it is u 1 x m1e x dx . 0 ( m) 0 0 (k 1) Fmk 2 (u)(1 Fm ( u))dFm (u) Fmk 1 (u / )dFm (u) . 1.2 Subset selection formulation The proposed selection is R*:select i iff Yi cY[ k t 1] , 0 c 1. If S is the selected subset, then P (CS | R*) P ( [i ] S , i k t 1,..., k | R*) P (Y(i ) cY[ k t 1] , i k t 1,..., k | R*) P (c max(Y(1) ,..., Y( k t ) ) min(Y( k t 1) ,..., Y( k ) )) Hence i nP f C( S 0 u(m ) ( F1 t cm u(. d) F ) 0 0 u( ) (k 1) Fmk 2 (u)(1 Fm (cu))dFm (u) Fmk 1 (u / c)dFm (u) . For t=1, it is 2. unknown, known If that |R * ) k ( t mk )t1F are unknown and 1.For are known, we may assume [1] [2] ... [ k ] the selection problem, let the ordered value of ' s. be Our goal is to select t (1 t k 1) best populations associated with the t largest location parameters [i ] , i k t 1,..., k . n(Yi1 i ) ~ E(0,1) . Let The estimation of i is Yi1 and Ti Yi1 , i 1,..., k. 2.1 Indifference zone formulation Let ( ) { ( 1 ,..., k ) | [ k t ] [ k t 1] }, 0 fixed be the preference zone. The proposed selection procedure is R: select t populations corresponding to T [ k ] , T [ k 1] ,..., and T [ k t 1] Then inf P (CS | R) P(max( Z1 ,..., Z k t ) n min( Z k t 1 ,..., Z k )) ( ) (1 e n ) k t (k t )e nt I (e n ; t 1, k t ) I (en ; t 1, k t ) where is an incomplete beta function. (1 e n )k 1 (k 1)ent I (en ; 2, k 1) . For t=1, it is 2.2 Subset selection formulation The proposed selection is R*:select i iff Ti T[ k t 1] c, c 0. and inf P (CS | R*) P(max( Z1 ,..., Z k t ) nc min( Z k t 1 ,..., Z k )) (1 e nc ) k t (k t )e ntc I (e nc ; t 1, k t ) (1 e nc )k 1 (k 1)entc I (enc ; 2, k 1) . For t=1, it is 2.3 Good Populations, Bad Populations, and Selection Problems. Let i ~ E ( i ,1) scheme and under progressive type II censoring ( R1 ,..., Rm ) , 1/( k 1) q=-log(1- ( P*) Theorem 2.1: Proof: let Ti Yi1 , i 1,..., k. ) and S={i| T T i [k ] G {i | i [ k ] } , If q / n }. Then P (G S ) P * n(Ti i ), i 1,..., k are i.i.d. E(0,1), we have P(max(n(Ti i )) min(n(Ti i )) q) P * If i [ k ] , then T[ k ] ( k ) Ti i q / n , so T[ k ] Ti T[ k ] Ti i [ k ] T[ k ] Ti i ( k ) q / n . Next, we let G {i | i [ k ] 1}, and B {i | i [ k ] 2 }, 1 2 be the set of good populations and bad populations. We want P(G S Bc ) P * . ( 2 1 )n 2q, Let For this problem we take n so that and take t so that S {i | Ti T[ k ] t / n} , n 2 q t q n1. then Theorem 2.2: P(G S B ) P * c Proof: i [ k ] 1 , and max(n(Ti i )) min(n(Ti i )) q , T[ k ] Ti 1 q / n t / n, and hence then iS . On the other hand, i [ k ] 2 , and max( n(Ti i )) min( n(Ti i )) q then Ti i T( k ) [ k ] q / n, so T( k ) [ k ] Ti i q / n Hence T[ k ] T( k ) Ti [ k ] i q / n Ti [ k ] i t / n 2 Ti t / n, and i S 2.4 Simultaneous Confidence Intervals Let i [ k ] i , i 1,..., k , our goal is to find two sided confidence intervals for i ' s with confidence coefficient at least P*. We have Theorem 2.3: P(G S , (T[ k ] Ti q / n) i (max(T j Ti ) q / n) , i 1,..., k ) P * j i Proof: max(n(Ti i )) min(n(Ti i )) q, and i G, then G S , Ti i T( k ) [ k ] q / n, then i T( k ) Ti q / n (max(T j Ti ) q / n) , i 1,..., k j i T[ k ] ( k ) Ti i q / n, so i ( k ) i T[ k ] Ti q / n (T[ k ] Ti q / n) , i 1,..., k Further, we will consider the pairwise comparison of the ranked parameters [i ] [ j ] , i j. We will use the following lemmas: Lemma 1: min(T[ j ] ( j ) ) T[i ] [i ] max(T[ j ] ( j ) ) . 1 j i i j k Proof: min(T[ j ] ( j ) ) min(T[i ] ( j ) ) T[i ] max [ j ] T[i ] [i ] 1 j i 1 j i 1 j i Similarly, max(T[ j ] ( j ) ) max(T[i ] ( j ) ) T[i ] min ( j ) T[i ] [i ] i j k i j k i j k Lemma 2: min(T[ i ] ( i ) ) min(T[ i ] [ i ] ), max(T[ i ] [ i ] ) max(T[ i ] ( i ) ) 1i k 1i k 1i k 1i k Proof: Let min(T[i ] [i ] ) T[l ] [l ] , then min(T[ i ] ( i ) ) T[ l ] [ l ] , by Lemma 1. 1i k 1i l Hence min(T[i ] (i ) ) min(T[i ] (i ) ) min(T[i ] [i ] ). 1i k 1i l 1i k Let max(T[i ] [i ] ) T[l ] [l ] , then T[l ] [l ] max(T[i ] (i ) ), by Lemma 1. 1i k l i k Hence max(T[i ] (i ) ) max(T[i ] (i ) ) T[l ] [l ] max(T[ i ] [ i ] ) 1i k l i k 1i k Theorem 2.4: P(T[i ] T[ j ] q / n [ i ] [ j ] T[ i] T[ j] q / n , i , j 1,..., k ) P * . Proof: If max(Ti i ) min(Ti i ) q / n, then, by Lemma 2 max(T[ j ] [ j ] ) max(T[ j ] ( j ) ) min(T[ i ] ( i ) ) q / n min(T[ i ] [ i ] ) q / n, 1 j k 1i k 1 j k 1i k Hence T[ j ] [ j ] T[ i ] [ i ] q / n, i j , and [ i ] [ j ] T[ i ] T[ j ] q / n, i j. Similarly, If max(Ti i ) min(Ti i ) q / n, then, by Lemma 2 max(T[ j ] [ j ] ) max(T[ j ] ( j ) ) min(T[ i ] ( i ) ) q / n min(T[ i ] [ i ] ) q / n, 1 j k 1i k 1 j k 1i k Hence T[i ] [i ] q / n T[ j ] [ j ] , i j , and T[ i ] T[ j ] q / n [ i ] [ j ] , i j. 3. unknown, unknown, is the parameter of interest Under progressive type II censoring scheme will let ( R1 ,..., Rm ) ,we Rm ) Yij X ij( :Rm1 :,..., , j 1,...m, i 1,..., k . n We have Z1 n(Yi1 i ) / i , Z 2 (n R1 1)(Yi 2 Yi1 ) / i ,..., Z m (n R1 R2 ... Rm 1 m 1)(Yim Yi ( m 1) ) / i are i.i.d. E(0,1) . Hence m U i Z j ~ Gamma(m 1,i ) j 2 3.1 Indifference zone formulation Let ( ) { (1 ,..., k ) | [ k t ] / [ k t 1] }, 0 1 fixed be the preference zone. As in section 1, the proposed selection procedure is R: select t populations corresponding to U [ k ] , U [ k 1] ,..., and U [ k t 1] . Then inf P (CS | R) (k t ) Fmk1t 1 (u )(1 Fm (1 u ))t dFm (u1 ) ( ) 0 3.2 Subset selection formulation The proposed selection is i R*:select iff U i cU[ k t 1] , 0 c 1. Then i nP f C( S 4. |R * ) k ( t mk1)t1F u(m 1) ( F1 0 t cm u1( d) F ) u( ) unknown, unknown, is the parameter of interest We will assume that i , i 1,..., k . II censoring scheme ( R1 ,..., Rm ) , Under progressive type let Rm ) Yij X ij( :Rm1 :,..., , j 1,...m, i 1,..., k . n We have Z1 n(Yi1 i ) / , Z 2 (n R1 1)(Yi 2 Yi1 ) / ,..., Z m (n R1 R2 ... Rm 1 m 1)(Yim Yi ( m 1) ) / are i.i.d. and Ti Yi1 E(0,1) . are and Yi1 ˆ The maximum likelihood estimators of and k m ˆ ( R j 1)(Yij Yi1 ) / k (m 1) . i 1 j 2 are independent. 4.1 Indifference zone formulation Let ( ) { ( 1 ,..., k ) | [ k t ] [ k t 1] }, 0 fixed be the i preference zone. The proposed selection procedure is R: select t populations corresponding to T [ k ] , T [ k 1] ,..., and T [ k t 1] Then inf P (CS | R) P(max( Z1 ,..., Z k t ) n min( Z k t 1 ,..., Z k )) ( ) (1 e n ) k t (k t )e nt I (e n ; t 1, k t ) 4.2 Subset selection formulation The proposed selection is R*:select i iff Ti T[ k t 1] cˆ / n, c 0. and inf P (CS | R*) P(max( Z1 ,..., Z k t ) cˆ / min( Z k t 1 ,..., Z k )) P(max( Z1 ,..., Z k t ) cx min( Z k t 1 ,..., Z k ))dG ( x) ((1 e cx )k t (k t )ecxt I (ecx ; t 1, k t ))dG ( x) where G(x) is the cdf of ˆ / , which is Gamma(k(m-1),k(m-1)). 4.3 Good Populations, Bad Populations, and Selection Problems. Theorem 4.1: If k 2 k 1 v G {i | i [ k ] } , 1 (1)i (1 q '(i 1) / v) P*, v k (m 1), i 0 i 1 and S={i| T T i [k ] q 'ˆ / n }. Then P (G S ) P * Proof: If i [ k ] , and max(n(Ti i ) / ˆ) min(n(Ti i ) / ˆ) q ' ,then T[ k ] ( k ) Ti i q 'ˆ / n , so T[ k ] Ti T[ k ] Ti i [ k ] T[ k ] Ti i ( k ) q 'ˆ / n . P(max(n(Ti i ) / ˆ) min(n(Ti i ) / ˆ) q ') P(max(n(T ) / ) min(n(T ) / ) q 'ˆ / ) i i i i P(max(n(Ti i ) / ) min(n(Ti i ) / ) q ' x)dG ( x) (1 e q ' x ) k 1 dG ( x) v v 1 vx k 1 k 1 k 1 k 1 e i iq ' x v x i v ( 1) e dx (1) (1 iq '/ v) i i (v ) i 0 i 0 k 2 k 1 v 1 (1)i (1 q '(i 1) / v) P * i 0 i 1 Next, we let G {i | i [ k ] 1}, and B {i | i [ k ] 2 }, 1 2 be the set of good populations and bad populations. We want P(G S Bc ) P * . ( 2 1 )n / ˆ 2q ', and take t so that Theorem 4.2: Let Proof: For this problem we take n so that n 2 / ˆ q ' t q ' n1 / ˆ. S {i | Ti T[ k ] tˆ / n} , then P(G S Bc ) P * i [ k ] 1 , and max(n(Ti i ) / ˆ) min(n(Ti i ) / ˆ) q ' , T[ k ] Ti 1 q 'ˆ / n tˆ / n, and hence iS . On the other hand, i [ k ] 2 , and max(n(Ti i ) / ˆ) min(n(Ti i ) / ˆ) q ' then then Ti i T( k ) [ k ] q 'ˆ / n, so T( k ) [ k ] Ti i q 'ˆ / n Hence T[ k ] T( k ) Ti [ k ] i q 'ˆ / n Ti [ k ] i tˆ / n 2 Ti tˆ / n, and i S 4.4 Simultaneous Confidence Intervals i [ k ] i , i 1,..., k , Let our goal is to find two sided confidence intervals for i ' s with confidence coefficient at least P*. We have Theorem 4.3: P(G S , (T[ k ] Ti q 'ˆ / n) i (max(T j Ti ) q 'ˆ / n) , i 1,..., k ) P * j i Proof: max(n(Ti i ) / ˆ) min(n(Ti i ) / ˆ) q ', and i G, then G S , T T q 'ˆ / n, then T T q 'ˆ / n (max(T T ) q 'ˆ / n) , i 1,..., k i i (k ) [k ] i (k ) i j i j i T[ k ] ( k ) Ti i q 'ˆ / n, so i ( k ) i T[ k ] Ti q 'ˆ / n (T[ k ] Ti q 'ˆ / n) , i 1,..., k Further, we will consider the pairwise comparison of the ranked parameters [i ] [ j ] , i j. We will use the following lemmas: Lemma 1: min(T[ j ] ( j ) ) T[i ] [i ] max(T[ j ] ( j ) ) . 1 j i i j k Proof: min(T[ j ] ( j ) ) min(T[i ] ( j ) ) T[i ] max [ j ] T[i ] [i ] 1 j i Similarly, 1 j i 1 j i max(T[ j ] ( j ) ) max(T[i ] ( j ) ) T[i ] min ( j ) T[i ] [i ] i j k i j k i j k Lemma 2: min(T[ i ] ( i ) ) min(T[ i ] [ i ] ), max(T[ i ] [ i ] ) max(T[ i ] ( i ) ) 1i k 1i k 1i k 1i k Proof: Let min(T[i ] [i ] ) T[l ] [l ] , then min(T[ i ] ( i ) ) T[ l ] [ l ] , by Lemma 1. 1i k 1i l Hence min(T[i ] (i ) ) min(T[i ] (i ) ) min(T[i ] [i ] ). 1i k 1i l 1i k Let max(T[i ] [i ] ) T[l ] [l ] , then T[l ] [l ] max(T[i ] (i ) ), by Lemma 1. 1i k l i k Hence max(T[i ] (i ) ) max(T[i ] (i ) ) T[l ] [l ] max(T[ i ] [ i ] ) 1i k l i k 1i k Theorem 4.4: P(T[i ] T[ j ] q 'ˆ / n [ i] [ j] T[ i] T[ j] q 'ˆ / n, i, j 1,..., k ) P * . Proof: If max(Ti i ) min(Ti i ) q 'ˆ / n, then, by Lemma 2 max(T[ j ] [ j ] ) max(T[ j ] ( j ) ) min(T[i ] (i ) ) q 'ˆ / n min(T[i ] [i ] ) q 'ˆ / n, 1 j k 1 j k 1i k 1i k Hence T[ j ] [ j ] T[i ] [i ] q 'ˆ / n, i j , and [i ] [ j ] T[i ] T[ j ] q 'ˆ / n, i j. Similarly, If max(Ti i ) min(Ti i ) q 'ˆ / n, then, by Lemma 2 max(T[ j ] [ j ] ) max(T[ j ] ( j ) ) min(T[i ] (i ) ) q 'ˆ / n min(T[i ] [i ] ) q 'ˆ / n, 1 j k 1 j k 1i k 1i k Hence T[ i ] [ i ] q 'ˆ / n T[ j ] [ j ] , i j , and T[ i ] T[ j ] q 'ˆ / n [i ] [ j ] , i j.