 Open Access
 Total Downloads : 410
 Authors : K. G. Potdar, D. T. Shirke
 Paper ID : IJERTV2IS80806
 Volume & Issue : Volume 02, Issue 08 (August 2013)
 Published (First Online): 29082013
 ISSN (Online) : 22780181
 Publisher Name : IJERT
 License: This work is licensed under a Creative Commons Attribution 4.0 International License
Reliability Estimation for the Distribution of a kUnit Parallel System with Rayleigh Distribution as the Component Life Distribution
Reliability Estimation for the Distribution of a kUnit Parallel System with Rayleigh Distribution as the Component Life Distribution
K. G. Potdar
Department of Statistics, Ajara Mahavidyalaya, Ajara, Dist Kolhapur, India416505,
D. T. Shirke
Department of Statistics, Shivaji University, Kolhapur, India416004,
Abstract A kunit parallel system having component lifetime distribution to be Rayleigh is considered. Based on the progressively TypeII censored sample, the maximum likelihood estimator (MLE) of the scale parameter of the Rayleigh distribution is derived and is used to estimate reliability function. EM algorithm is used to obtain MLE. Asymptotic confidence interval for the scale parameter and reliability function is constructed. Confidence interval based on the log transformed MLE is also constructed. Simulation study is conducted to investigate performance of estimates and confidence intervals. An example with real data is presented for illustration.
Index Terms Progressively TypeII censoring, Rayleigh distribution, EM algorithm, MLE, Reliability, confidence interval, coverage probability.
AMS subject classification 62N01, 62N02, 62N05.
ACRONYMS
CI Confidence Interval
CDF Cumulative Density Function EM Expectation Maximization MLE Maximum Likelihood Estimate MSE Mean Square Error
PDF Probability Density Function
X(i) Order Statistics from a progressively censored sample of size m.
Ri Number of surviving units withdrawn from the experiment after ith failure.
L Loglikelihood function of all n observations. L() Likelihood function of observed data.
Lc Loglikelihood function of complete data. Ix() Observed Fisher information.
Iw() Complete Fisher information. Iwx() Missing information.
R(t) Reliability of system at time t.

In a life testing experiment, censoring is common practice because of various restrictions on data collection such as time limit, cost etc. In such a situation we remove some observations in the experiment. Such data is called censored data. Censoring is broadly classified into two types; TypeI and TypeII censoring.
TypeI censoring is related with time. In this type, an experiment continues up to a predetermined time T. Units having failure time after time T are not observed. Here, lifetime will be known exactly only if it is less than T. For example, suppose n units are put on test, but decision is made to terminate the test after time
T. In this experiment life times will be known exactly only for those units that fail before time T. In TypeI censoring, the number of exact life times observed is random.
NOTATIONS
n Sample size.
m Number of observed failures in a TypeII censored sample.
Scale parameter of the Rayleigh distribution.
MLE of .
g(y) PDF of the Rayleigh distribution. G(y) CDF of the Rayleigh distribution.
f(x) PDF of the life distribution of k unit parallel system.
F(x) CDF of the life distribution of k unit parallel system.
TypeII censoring is related with number of failures, that is, experiment continues up to the pre determined number of failures. For example, in life testing experiment, suppose n units are put on test, but instead of continuing until all n units have failed; the test is terminated at the time of the mth (m n) unit failure. In case of TypeII censoring, the number of exact life times observed is fixed.
Based on the time epochs for removals of units from the experiment, the censoring is further classified in to two types, such as conventional censoring and progressive censoring. In conventional TypeI and Type II censoring, units may not be removed before terminal point. In progressive censoring scheme, units may be
removed at different stages rather than at terminal point. Progressive censoring scheme is applied in both TypeI and TypeII censoring schemes.
=
=
In progressive TypeI censoring scheme, m censoring times T1, T2,, Tm and R1, R2,, Rm are fixed in advance. Since 1 = . At time T1 remove R1 units, at time T2 remove R2 units and so on. The experiment terminates at time Tm with Rm units still surviving. In progressive TypeII censoring scheme, suppose n units are put on test. The number m and R1, R2,.,Rm are fixed prior to the test. At the first failure R1 units are removed randomly from remaining n1 units.
At the second failure R2 units are removed randomly from remaining n2R1 units and so on. At the mth failure all remaining Rm units are removed. Here, we observe failure time of m units and remaining nm units
are removed at different stages of experiment. In conventional TypeII censoring scheme R1=R2== Rm1 = 0 and Rm= nm. In this paper, we consider the progressive TypeII censoring scheme.
Progressive TypeII censoring scheme for various lifetime distributions has been discussed by number of researchers. Cohen [1] studied MLE of the parameters of exponential and normal distribution for progressively TypeII censored samples. Mann ([2], [3]) considered Weibull distribution with progressive censoring. Balkrishnan ([4], [5] and [6]) discussed inference for the scaled halflogistic, Gaussian and extreme value distributions under progressive TypeII censoring scheme respectively. Ng [7] studied parameter estimation for modified Weibull distribution under progressively TypeII censored samples. Balkrishnan and Aggarwala [8] gave details about progressive TypeII censoring scheme.
Balkrishnan [9] studied various distributions and inferential methods for progressively censored data. Pradhan [10] considered point and interval estimation of a kunit parallel system based on progressive TypeII censoring scheme with exponential distribution as the component life distribution. Chein and Balkrishnan [11] discussed consistency and asymptotic normality of MLE based on progressive TypeII censored samples.
Iliopoulos and Balkrishnan [12] studied likelihood
Dempster et al. [17] introduced EM algorithm. They presented maximum likelihood estimation for incomplete data. Mclachlan and Krishnan [18] gave more details about EM algorithm. Little and Rubin [19] discussed EM algorithm for exponential family of distributions. Pradhan and Kundu [20] used EM algorithm to estimate parameters of generalized exponential distribution under TypeII censoring scheme. Ng et al. [21] used EM algorithm to estimate parameters of lognormal and Weibull distributions under TypeII censoring scheme. In this article, we used EM algorithm for estimation of the parameters of a kunit parallel system based on progressive TypeII censoring scheme when component lifetime follows Rayleigh distribution with scale parameter .
=
=
Parameter estimation is based on the life times of the system. We assumed that n items put on test and failure times of 1 = items are censored. Lifetimes of these censored items are unknown. We consider this data as missing and used EM algorithm to
compute MLE. Louis [22] presented technique for computing observed Fisher information within EM algorithm framework. We used this technique to obtain observed Fisher information. Asymptotic normal distribution of MLE is used to construct CI for the parameter.
Maximum likelihood method is used to estimate the parameter and reliability function of the Rayleigh distribution. This estimation procedure is discussed in Sectio 2. By using asymptotic normality of the MLE, we derived CI and coverage probability is computed in Section 3. To investigate performance of procedure, simulation study has been made in Section 4. Results of simulation study are discussed in Section 5. In Section 6 proposed estimation methods are illustrated for real data set. Conclusions are presented in Section 7.
Consider Rayleigh distribution with scale parameter . The PDF g y and CDF G y are respectively given by,
inference for Laplace distribution based on progressive TypeII censored samples. Krishna and Malik [13] discussed reliability estimation in Maxwell distribution
g y =
2y e
2
y 2
y 2
y 0, > 0
based on progressively TypeII censord data. Recently,
G y = 1 e
y 0, > 0
Potdar and Shirke [14] studied inference for the scale parameter of lifetime distribution for kunit parallel system based on progressively censored data. Potdar and Shirke [15] discussed estimation for the distribution of a kunit parallel system with exponential distribution as the component life distribution based on TypeII
Consider k unit parallel system with independent and identically distributed components. Let Y1, Y2, , Yk be the lifetimes, where Yi is the lifetime of the ith component with Yi ~ g y . Life time of the system X= max(Y1, Y2, . , Yk). The CDF of X is
x 2 k
progressively censored data. Potdar and Shirke [16] studied Inference for the parameters generalized inverted family of distributions.
F x = 1 e
x 0, > 0. (1)
and the PDF of X is
2
m
2 k1 2 0
2kx x
x
+ 3 Ri a xi, k,
f x =
2 e
1 e
x 0, > 0. (2)
i=1
Suppose n systems of kunit parallel systems are under test and we observe lifetimes of m systems under progressive TypeII censoring. Let (R , R ,.., R )
2 k 1 m
3 Ri b xi, k, 0 . 6
i=1
where a xi, k, 0 = E z2 zij > xi
1 2 m ij
is a progressive censoring scheme. The likelihood
2kz z 2
z 2
k1
function for the observed data is
2 e
1 e
m = z2 dz .
L = C f(x
) 1 F x
Ri , (3)
x 2 k
i
i
i=1
i
m1
i
j
xi 1 1 e
zij 2
z2e
where C = n n j Ri .
and b xi, k, 0 = E ij
zij 2
zij > xi
m
2kx
j=1
i
i
x 2
i=1
1 e
L = C
i=1
i e
2
2 z 2
2kz z 2
2 e
z 2
1 e
k1
x i 2
Ã— 1 e
k1
= z e
z 2
x 2 k
dz.
x 2 k Ri
xi 1 e
1 1 e i
dL
(i)
We have to solve equation c = 0 to obtain 1
Ã— 1 1 e
. (4)
d
as the solution. But this equation does not have solution
Suppose x(1), x(2),., x(m) is the observed data and z , z , . , z is the censored data. We note that z
in the closed form. Therefore we used NewtonRaphson method and compute 1. By using 1, we
1 2 m
th i computed a x , k, 1 and b x , k, 1 . This ends the M
is a vector with Ri elements. That is at the i failure, we i i
remove Ri (i=1,2,.,m) systems. Observations on the removed systems are not available. The censored data Z= (z1, z2, . , zm) can be considered as missing data. X= (x(1), x(2), , x(m)) is observed data. W=(X, Z) is the complete data set. Then complete log likelihood function is
Lc = nlog 2 + nlog k 2nlog
step. We continue this procedure until convergence took place.
In Newton Raphsonmethod, we need to choose initial value of . We used least square estimate. Ng [7] discussed estimation of model parameters of modified Weibull distribution based on progressively TypeII censored data where the empirical distribution function is
m
+ log xi
m
2
2
1 x2
computed as (see Meeker and Escober [23])
i
i=1
i=1
m
i
xi 2
F x
i
= 1 (1 p j).
j=1
+ k 1 log 1 e
i=1
with
m Ri
m Ri
1
p = 1
, for j = 1,2, , m.
+ log zij
n
n
i=1 j=1
2 zi2j
i=1 j=1
j j
k=2
Rk1
j + 1
m Ri
zij 2
The estimate of the parameters can be obtained by least
+ k 1 log 1 e . (5)
i=1 j=1
square fit of simple linear regression.
yi=x(i) with = 1/.
1 1
In order to obtain MLE of , we use EM algorithm due to Dempster et al. [17]. For the E step in EM algorithm we take Expectation of Zij. The derivative of Lc with respect to is taken for the M step, where
y = ln 1 F k x i1 +F k x i , for i=1,2,..,m.
i
i
2
F x 0 = 0.

m 2 xi 2
i
i
dLc = 2n + 2 x2 2 k 1 xi e
The least square estimates of is given by
3
3
3
3
d
i=1
i=1 1 e
xi 2
m x2

xi 2
= i=1 (i) . (7)
6 k 1
x2 e
m
i=1
x 2
x 2
i
i
x(i)yi
E
4
i=1
i . (9) 1 e
We used as an initial value of in Newton
Raphson method. Reliability function at time t is
R t = P X > = 1 F t .
Missing information is given by
m
t 2 k
I = R I i
R t = 1 1 e
t 0, > 0.
WX
i=1
m
i WX
Ri
d2log f Z X ,
ij i
The Maximum likelihood estimate of R(t) is
= EZX
d2 .
(10)
R t = 1 1 e
t 2 k
t 0, > 0.
Consider
i=1 j=1
We compute observed Fisher information using
f zij ;
f Z X , =
the idea of missing information principle of Louis [22].
zX
ij i
1 F xi;
2kz
zij 2
zij 2
k1
The observed information
ij e
2
1 e
= complete information missing information
Ix() = Iw() Iwx(), (8)
where,
Let,
=
1 1 e
.
xi 2 k
Complete information = Iw
= E d2 L ,
2
2
d
log f = log fzX Zij Xi, .
where, L is the log likelihood function of all n observations.
Therefore,
zij 2
logf = log 2 + log k + log z 2 log
Now,
ij
zij 2
xi 2 k
L = nlog 2 + nlog k
n n
+ k 1 log 1 e
log 1 1 e .
2nlog + log x 1 x2
zij 2
i 2 i
dlogf
2 2z2
2 k 1 z2e
n i=1
xi 2
i=1 = +
d
ij
3
ij
zij 2
+ k 1 log 1 e .
i=1
3 1 e
x 2
x 2 k1
and
2 i
i
n
dL 2n 2
2 k 1 n
x2e
xi 2
2kxi e
1 e
.
d = +
x2
i
i
3
i .
xi 2
xi 2
3
xi 2 k
3 1 1 e
i=1
i=1 1 e
d2 L
n
2n 6 2
4 k 1 n
x4 e
xi 2
and
2
d2 = 2 4 xi
6

2 2
zij
2 4
i
i
i=1
i=1 1 e i
d2logf
2 6zij
4 k 1 zij e
i
i
n
n
x 2
2
=
d2 2 4
6
zij 2 2
+ 6 k 1 xi e .
1 e
i
i
4 x 2
zij 2
i=1 1 e
6 k 1 z2e
ij
ij
+
Complete information is given by,
zij 2
2n Iw = 2 +
n
i
i
6 E x2
4
4 1 e
xi 2

2 k2
xi 2
4 k 1
i=1
n
xi 2
x4e
4kx4 e
i
i
1 e
x
1 ke
2 k
+ E i
6 i
6 xi 2 2
1 1 e
i=1 1 e
xi 2
i
i
4k2x4 e2
xi 2 2(k1)
1 e
2 = 1
n
n
I n
is the estimated variance of n .
+ 2
6
6
xi 2 k
1 1 e
Therefore, 100(1)% asymptotic CI for is given by
n /2 2 n , n + /2 2 n , (12)
2 xi 2
xi 2 k1
i
i
6kx2 e 1 e
+ .
where /2 is the upper 100(/2)th percentile of
i
i
x 2 k
6 1 1 e
standard normal distribution.
Meeker and Escober [23] reported that the
Hence Missing information is
m
asymptotic CI for based on ln( n ) has better coverage probability. An approximate 100(1)% CI for ln() is
WX
WX
IWX = Ri I i
ln n 2 ln n ,
i=1
R
2 , (13)
m i d2log f Zij Xi ,
ln(
) + 2 ln( )
= EZX i=1 j=1
d2 ,
n /2 n
m Ri
where 2 ln n is the estimated variance of
= 2 n m + 6 E z2
ln n which is approximately obtained by
2 4
i
i
m R
i=1 j=1
ij
zij 2
2 ln n
2
n
n
n 2
Hence, an approximate 100(1
4 k 1
z4e
)% CI for is
+ 6
E ij
zij 2 2
2
2
i=1 j=1
1 e
/2 n
/2 n
6 k 1 m Ri
zij 2
z2e
n e
n
, n e
n
. (14)
4
E
i=1 j=1
ij
1 e
zij 2
4. Simulation Study
R xi 2
xi 2 k2
xi 2
m i x4 e
1 e
1 ke
+ 4k i
6
i=1 j=1
xi 2 k
1 1 e
A simulation study is carried out to study the performance of MLE by considering bias and MSE for various progressively TypeII censoring scheme.
m Ri
4 2 xi 2
x 2 2 k1
i
i
Approximate CIs based on MLE and logtransformed
MLE are compared through their coverage probability.
4k2
xi e
1 e
Balkrishnan and Sandhu [24] presented algorithm for
6 2 k 2
sample generation from progressively TypeII censored
i=1 j=1
1 1 e
xi
scheme. Using this algorithm, we generate samples from
m Ri
2 xi 2
x 2 k1
i
i
the distribution of a kunit parallel system with Rayleigh
distribution as the component life distribution.
6k
4
xi e 1 e
x 2 k
. (11)
Algorithm

Generate i.i.d. observations (W1, W2, , Wm) from
i=1 j=1
1 1 e
i
U(0, 1).

For (R1, R2, . , Rm) censoring scheme, set Ei= 1/
Using expressions in equations (10) and (11), we
(i + Rm+Rm1+. + Rmi+1) for i=1,2,.,m.

Set Vi= W Ei for i=1,2, .. , m.
obtained Fisher information.

Set Ui=
i
1 Vm
Vm1
.. V
mi+1
) for i=1,2, . , m.
3. Confidence Intervals
Then (U1, U2, , Um) is the progressively TypeII censored sample from U(0,1).

For given values of the parameter , set
i
i
Using property of asymptotic normality of MLE we construct CI for . Let is the MLE of and
= 2 log 1 (U )1/k 1/2, for i=1, , m. (15) Then (x(1), x(2), , x(m)) is the required
progressively TypeII censored sample from the
distribution of a kunit parallel system with Rayleigh distribution as the component life distribution. For
Table 1.
Progressively TypeII censored schemes used for simulation study
n
m
Scheme no.
Scheme
5
2
[1] (3,0)
[2] (0,3)
[3] (1,2)
[4] (2,1)
15
5
[5] (10, 4*0)
[6] (4*0, 10)
[7] (2,2,2,2,2)
15
10
[8] (5,9*0)
[9] (9*0,5)
[10] (3,2, 8*0)
20
10
[11] (10,9*0)
[12] (9*0,10)
25
10
[13] (15,9*0)
[14] (9*0,15)
[15] (5,5,5,7*0)
25
15
[16] (10,14*0)
[17] (14*0,10)
30
10
[18] (20, 9*0)
[19] (9*0,20)
30
15
[20] (15, 14*0)
[21] (14*0,15)
[22] (5,5,5,12*0)
30
20
[23] (10, 19*0)
[24] (19*0,10)
[25] (0,5,5,17*0)
50
20
[26] (30,19*0)
[27] (19*0,30)
50
35
[28] (15,34*0)
[29] (34*0,15)
[30] (5,5,5,32*0)
In Table 1, scheme (a, b) stands for R1= a and R2=b. Similar meaning holds for schemes described through completely specified vector, while scheme (10, 4*0) means R1=10 and rest four Ris are zero. i.e. R2=R3=R4=R5=0.
Simulation was carried out for 3unit parallel system and 5unit parallel system (i.e. k=3, k=5) with =1. EM algorithm and NewtonRaphson method are used to compute MLE. For each particular progressive censoring scheme, 10,000 sets of observations were generated. The bias, the MSE and the coverage probability for the corresponding approximate CIs for are displayed in the Tables 2 and 3 for k=3 and k=5 respectively.Further, reliability estimate for different time period t= 1 given in Tables 4 and 5 for k=3 and k=5 respectively.
simulation study we consider 30 different progressively TypeII censored schemes as mentioned in Table I.
Table 2.
Bias, MSE and Coverage Probability for k=3 and =1
Schem e No.
Bas
MSE
Confidence level (MLE)
90% 95%
Confidence level (log MLE)
90% 95%
[1] 0.0195
0.0406
0.8653
0.9075
0.8806
0.9323
[2] 0.0196
0.0351
0.8683
0.9118
0.8865
0.9376
[3] 0.0210
0.0359
0.8675
0.9080
0.8831
0.9335
[4] 0.0222
0.0377
0.8651
0.9092
0.8832
0.9350
[5] 0.0082
0.0174
0.8861
0.9327
0.8919
0.9432
[6] 0.0079
0.0133
0.8829
0.9324
0.8893
0.9414
[7] 0.0083
0.0143
0.8847
0.9297
0.8919
0.9414
[8] 0.00274
0.0096
0.8951
0.9425
0.8961
0.9482
[9] 0.0023
0.0082
0.8945
0.9421
0.8944
0.9497
[10] 0.0036
0.0096
0.8929
0.9416
0.8966
0.9457
[11] 0.0039
0.0093
0.8927
0.9422
0.8977
0.9513
[12] 0.0027
0.0076
0.8908
0.9407
0.8949
0.9462
[13] 0.0051
0.0093
0.8897
0.9408
0.8959
0.9455
[14] 0.0035
0.0067
0.8990
0.9460
0.9052
0.9505
[15] 0.0037
0.0087
0.8945
0.9412
0.8990
0.9464
[16] 0.0024
0.0063
0.9017
0.9456
0.9025
0.9515
[17] 0.00157
0.0053
0.8943
0.9430
0.8963
0.9471
[18] 0.0038
0.0091
0.8927
0.9431
0.8990
0.9492
[19] 0.0045
0.0066
0.8919
0.9437
0.8977
0.9453
[20] 0.00314
0.0064
0.8952
0.9433
0.8963
0.9468
[21] 0.0039
0.0048
0.9022
0.9496
0.9039
0.9534
[22] 0.0031
0.0061
0.8952
0.9425
0.8981
0.9471
[23] 0.0018
0.0049
0.8939
0.9423
0.8951
0.9457
[24] 0.00096
0.0041
0.8988
0.9463
0.9008
0.9490
[25] 0.0016
0.0047
0.8985
0.9451
0.9026
0.9482
[26] 0.0013
0.0048
0.8960
0.9451
0.8966
0.9489
[27] 0.0016
0.0033
0.8981
0.9449
0.8999
0.9498
[28] 0.00001
0.0028
0.9002
0.9498
0.9021
0.9509
[29] 0.00034
0.0024
0.8971
0.9475
0.8961
0.9485
[30] 0.0004
0.0027
0.9022
0.9525
0.9019
0.9537
From Table II and III we observe the following for the MLE of the scale parameter.

The bias and MSE of the MLE decrease with increase in sample size n as well as with increase in the effective sample size m.

The bias and MSE of the MLE decrease as k (no. of units in parallel system) increases.
Table 3.
Bias, MSE and Coverage Probability for k=5 and =1
Table 4.
Bias, MSE and Coverage Probability of reliability estimate for k=3, t=1 and =1, R(t)=0.7474
Scheme No.
Bias
MSE
Confidence level (MLE)
90% 95%
Confidence level (log MLE)
90% 95%
[1] 0.0160
0.0258
0.8733
0.9171
0.8831
0.9362
[2] 0.0091
0.0217
0.8807
0.9290
0.8909
0.9407
[3] 0.0122
0.0224
0.8788
0.9248
0.8879
0.9403
[4] 0.0113
0.0237
0.8792
0.9256
0.8917
0.9367
[5] 0.0046
0.0115
0.8879
0.9397
0.8941
0.9449
[6] 0.0042
0.0078
0.8925
0.9418
0.8969
0.9482
[7] 0.0044
0.0087
0.8918
0.9412
0.8932
0.9458
[8] 0.0009
0.0065
0.8985
0.9472
0.9002
0.9492
[9] 0.0014
0.0053
0.8997
0.9460
0.8992
0.9492
[10] 0.00014
0.0064
0.8958
0.9442
0.8957
0.9475
[11] 0.0019
0.0063
0.8962
0.9458
0.8978
0.9479
[12] 0.0023
0.0046
0.8970
0.9450
0.8979
0.9477
[13] 0.0017
0.0061
0.8971
0.9473
0.8991
0.9495
[14] 0.0026
0.0042
0.8977
0.9456
0.8994
0.9504
[15] 0.0017
0.0055
0.9009
0.9463
0.9006
0.9502
[16] 0.0005
0.0043
0.8993
0.9465
0.8998
0.9491
[17] 0.0007
0.0034
0.8958
0.9481
0.8978
0.9481
[18] 0.0016
0.0061
0.8963
0.9438
0.8965
0.9490
[19] 0.0016
0.0039
0.8944
0.9452
0.8969
0.9454
[20] 0.0016
0.0042
0.9013
0.9490
0.9029
0.9507
[21] 0.0013
0.0031
0.8962
0.9461
0.8965
0.9484
[22] 0.0018
0.0040
0.8970
0.9440
0.8997
0.9478
[23] 0.0001
0.0033
0.8968
0.9464
0.8978
0.9484
[24] 0.0006
0.0027
0.9026
0.9502
0.9045
0.9512
[25] 0.00002
0.0031
0.9005
0.9524
0.9020
0.9529
[26] 0.0014
0.0032
0.8973
0.9435
0.8958
0.9446
[27] 0.0011
0.0021
0.8971
0.9448
0.8968
0.9460
[28] 0.0008
0.0019
0.8980
0.9499
0.8989
0.9500
[29] 0.0004
0.0015
0.9047
0.9488
0.9032
0.9497
[30] 0.0003
0.0019
0.9004
0.9463
0.9009
0.9489
Scheme No.
Bias
MSE
Confidence level (MLE) 90% 95%
[1] 0.0648
0.0427
0.8258
0.8716
[2] 0.0627
0.0374
0.8424
0.8867
[3] 0.0620
0.0378
0.8390
0.8881
[4] 0.0662
0.0395
0.8414
0.8880
[5] 0.0302
0.0164
0.8702
0.9193
[6] 0.0265
0.0125
0.8774
0.9299
[7] 0.0267
0.0132
0.8780
0.9246
[8] 0.0166
0.0085
0.8856
0.9342
[9] 0.0143
0.0071
0.8878
0.9356
[10] 0.0157
0.0083
0.8811
0.9310
[11] 0.0174
0.0082
0.8867
0.9360
[12] 0.0146
0.0066
0.8820
0.9335
[13] 0.0168
0.0082
0.8819
0.9323
[14] 0.0141
0.0060
0.8853
0.9356
[15] 0.0170
0.0078
0.8872
0.9336
[16] 0.0107
0.0054
0.8892
0.9361
[17] 0.0095
0.0043
0.8937
0.9421
[18] 0.0155
0.0077
0.8918
0.9422
[19] 0.0116
0.0055
0.8873
0.9384
[20] 0.0113
0.0055
0.8869
0.9348
[21] 0.0084
0.0040
0.8922
0.9403
[22] 0.0104
0.0051
0.8916
0.9407
[23] 0.0086
0.0041
0.8868
0.9412
[24] 0.0064
0.0034
0.8884
0.9386
[25] 0.0084
0.0039
0.8901
0.9415
[26] 0.0073
0.0039
0.8890
0.9408
[27] 0.0058
0.0028
0.8929
0.9448
[28] 0.0049
0.0022
0.8984
0.9472
[29] 0.0040
0.0019
0.8964
0.9465
[30] 0.0044
0.0022
0.8986
0.9453
Scheme No.
Bias
MSE
Confidence level (MLE)
90% 95%
Confidence level (log MLE)
90% 95%
[1] 0.0160
0.0258
0.8733
0.9171
0.8831
0.9362
[2] 0.0091
0.0217
0.8807
0.9290
0.8909
0.9407
[3] 0.0122
0.0224
0.8788
0.9248
0.8879
0.9403
[4] 0.0113
0.0237
0.8792
0.9256
0.8917
0.9367
[5] 0.0046
0.0115
0.8879
0.9397
0.8941
0.9449
[6] 0.0042
0.0078
0.8925
0.9418
0.8969
0.9482
[7] 0.0044
0.0087
0.8918
0.9412
0.8932
0.9458
[8] 0.0009
0.0065
0.8985
0.9472
0.9002
0.9492
[9] 0.0014
0.0053
0.8997
0.9460
0.8992
0.9492
[10] 0.00014
0.0064
0.8958
0.9442
0.8957
0.9475
[11] 0.0019
0.0063
0.8962
0.9458
0.8978
0.9479
[12] 0.0023
0.0046
0.8970
0.9450
0.8979
0.9477
[13] 0.0017
0.0061
0.8971
0.9473
0.8991
0.9495
[14] 0.0026
0.0042
0.8977
0.9456
0.8994
0.9504
[15] 0.0017
0.0055
0.9009
0.9463
0.9006
0.9502
[16] 0.0005
0.0043
0.8993
0.9465
0.8998
0.9491
[17] 0.0007
0.0034
0.8958
0.9481
0.8978
0.9481
[18] 0.0016
0.0061
0.8963
0.9438
0.8965
0.9490
[19] 0.0016
0.0039
0.8944
0.9452
0.8969
0.9454
[20] 0.0016
0.0042
0.9013
0.9490
0.9029
0.9507
[21] 0.0013
0.0031
0.8962
0.9461
0.8965
0.9484
[22] 0.0018
0.0040
0.8970
0.9440
0.8997
0.9478
[23] 0.0001
0.0033
0.8968
0.9464
0.8978
0.9484
[24] 0.0006
0.0027
0.9026
0.9502
0.9045
0.9512
[25] 0.00002
0.0031
0.9005
0.9524
0.9020
0.9529
[26] 0.0014
0.0032
0.8973
0.9435
0.8958
0.9446
[27] 0.0011
0.0021
0.8971
0.9448
0.8968
0.9460
[28] 0.0008
0.0019
0.8980
0.9499
0.8989
0.9500
[29] 0.0004
0.0015
0.9047
0.9488
0.9032
0.9497
[30] 0.0003
0.0019
0.9004
0.9463
0.9009
0.9489
Scheme No.
Bias
MSE
Confidence level (MLE) 90% 95%
[1] 0.0648
0.0427
0.8258
0.8716
[2] 0.0627
0.0374
0.8424
0.8867
[3] 0.0620
0.0378
0.8390
0.8881
[4] 0.0662
0.0395
0.8414
0.8880
[5] 0.0302
0.0164
0.8702
0.9193
[6] 0.0265
0.0125
0.8774
0.9299
[7] 0.0267
0.0132
0.8780
0.9246
[8] 0.0166
0.0085
0.8856
0.9342
[9] 0.0143
0.0071
0.8878
0.9356
[10] 0.0157
0.0083
0.8811
0.9310
[11] 0.0174
0.0082
0.8867
0.9360
[12] 0.0146
0.0066
0.8820
0.9335
[13] 0.0168
0.0082
0.8819
0.9323
[14] 0.0141
0.0060
0.8853
0.9356
[15] 0.0170
0.0078
0.8872
0.9336
[16] 0.0107
0.0054
0.8892
0.9361
[17] 0.0095
0.0043
0.8937
0.9421
[18] 0.0155
0.0077
0.8918
0.9422
[19] 0.0116
0.0055
0.8873
0.9384
[20] 0.0113
0.0055
0.8869
0.9348
[21] 0.0084
0.0040
0.8922
0.9403
[22] 0.0104
0.0051
0.8916
0.9407
[23] 0.0086
0.0041
0.8868
0.9412
[24] 0.0064
0.0034
0.8884
0.9386
[25] 0.0084
0.0039
0.8901
0.9415
[26] 0.0073
0.0039
0.8890
0.9408
[27] 0.0058
0.0028
0.8929
0.9448
[28] 0.0049
0.0022
0.8984
0.9472
[29] 0.0040
0.0019
0.8964
0.9465
[30] 0.0044
0.0022
0.8986
0.9453
Table 5. Bias, MSE and Coverage Probability of reliability estimate for k=5, t=1 and =1, R(t)=0.8991
Scheme No.
Bias
MSE
Confidence level (MLE)
90% 95%
[1] 0.0512
0.0203
0.8512
0.8869
[2] 0.0455
0.0169
0.8506
0.8911
[3] 0.0447
0.0168
0.8550
0.8920
[4] 0.0471
0.0185
0.8511
0.8920
[5] 0.0226
0.0062
0.8791
0.9179
[6] 0.0164
0.0040
0.8852
0.9279
[7] 0.0169
0.0045
0.8826
0.9256
[8] 0.0113
0.0029
0.8840
0.9260
[9] 0.0095
0.0023
0.8917
0.9341
[10] 0.024
0.0031
0.8800
0.9239
[11] 0.0118
0.0028
0.8878
0.9287
[12] 0.0082
0.0019
0.8895
0.9318
[13] 0.0107
0.0028
0.8782
0.9237
[14] 0.0086
0.0019
0.8870
0.9329
[15] 0.0107
0.0025
0.8840
0.9281
[16] 0.0078
0.0017
0.8933
0.9346
[17] 0.0063
0.0013
0.8925
0.9366
[18] 0.0112
0.0027
0.8936
0.9341
[19] 0.0073
0.0016
0.8940
0.9370
[20] 0.0072
0.0017
0.8894
0.9333
[21] 0.0058
0.0013
0.8939
0.9380
[22] 0.0080
0.0017
0.8900
0.9365
[23] 0.0061
0.0013
0.8936
0.9399
[24] 0.0047
0.0011
0.8875
0.9383
[25] 0.0060
0.0013
0.8939
0.9404
[26] 0.0056
0.0012
0.8939
0.9369
[27] 0.0043
0.0008
0.8976
0.9457
[28] 0.0031
0.0007
0.8970
0.9424
[29] 0.0028
0.0006
0.8914
0.9415
[30] 0.0039
0.0007
0.8961
0.9444

The bias and MSE of the MLE decrease with increase in sample size n as well as with increase in effective sample size m. The bias and MSE of the MLE decrease as k (no. of units in parallel system) increases. The MSE of the MLE is relatively small for conventional TypeII censoring scheme as compared with progressive TypeII censoring scheme. There is negligible difference between bias in case of conventional TypeII censoring scheme and progressive TypeII censoring scheme. Coverage probability in case of progressive TypeII censoring scheme are better than confidence levels in case of conventional TypeII censoring scheme. Same trend is observed in confidence levels for the logtransformed MLE. Coverage probability for k=5 is better than coverage probability for k=3 in case of small sample size. Same trend is observed in coverage probability for the logtransformed MLE. There is negligible difference in coverage probability for different k for large sample size
in both MLE and logtransformed MLE case. Coverage probability using logtransformed MLE are close to nominal levels as compared to the coverage probability of MLE for small size, while for large sample size both are similar.
The bias and MSE of the MLE of Reliability decrease with increase in sample size n as well as with increase in the effective sample size m. The bias and MSE of the MLE of Reliability decrease as k (no. of units in parallel system) increases. There is negligible difference between the bias and MSE of the MLE of Reliability in case of conventional TypeII censoring scheme and progressive TypeII censoring scheme. For small sample size, coverage probability in case of conventional TypeII censoring are better than coverage probability in case of progressive TypeII censoring, whereas for large sample size coverage probability in case of progressive TypeII censoring are better.

Consider the following data which represent the number of revolutions to failure (in hundreds of millions) for each of 23 ball bearings given by Lieblein and Zelen
[25] (also given by Lawless [26]).0.1788, 0.2892, 0.33, 0.4152, 0.4212, 0.4560, 0.4840,
0.5184, 0.5196, 0.5412, 0.5556, 0.6780, 0.6864, 0.6864,
0.6888, 0.8412, 0.9312, 0.9864, 105.12, 105.84, 127.92,
128.04, 173.40
According to Raqab and Madi [27], scale parameter Rayleigh distribution satisfactory fit to this data. We consider this data as outcome for life time of two unit parallel systems and three unit parallel systems. Reliability estimate for k=2, 3 and different time periods (t=0.5, 1, 1.5, 2) is given in Tables 6.
k
n
m
Scheme
MLE
t=0.5
R(t)^
t=1 t=1.5
t=2
2
11
7
(1,1,1,1,3*0)
0.6141
0.7651
0.1361
0.0051
0
(4,6*0)
0.7332
0.8617
0.2871
0.0302
0.0012
(6*0,4)
0.7155
0.8507
0.2635
0.0245
0.0008
(3*0,1,1,1,1)
0.7606
0.8769
0.3236
0.0405
0.0020
11
(11*0)
0.7380
0.8645
0.2935
0.0319
0.0013
3
7
4
(1,1,1,0)
0.7064
0.9388
0.3523
0.0327
0.0010
(3,3*0)
0.8906
0.9802
0.6321
0.1657
0.0192
(3*0,3)
0.7451
0.9523
0.418
0.0512
0.0022
7
(7*0)
0.7894
0.9639
0.4898
0.0789
0.0049
k
n
m
Scheme
MLE
t=0.5
R(t)^
t=1 t=1.5
t=2
2
11
7
(1,1,1,1,3*0)
0.6141
0.7651
0.1361
0.0051
0
(4,6*0)
0.7332
0.8617
0.2871
0.0302
0.0012
(6*0,4)
0.7155
0.8507
0.2635
0.0245
0.0008
(3*0,1,1,1,1)
0.7606
0.8769
0.3236
0.0405
0.0020
11
(11*0)
0.7380
0.8645
0.2935
0.0319
0.0013
3
7
4
(1,1,1,0)
0.7064
0.9388
0.3523
0.0327
0.0010
(33*0)
0.8906
0.9802
0.6321
0.1657
0.0192
(3*0,3)
0.7451
0.9523
0.418
0.0512
0.0022
7
(7*0)
0.7894
0.9639
0.4898
0.0789
0.0049
Table 6. Reliability Estimate for k=2, 3
From Table 6 we observe the following.
Reliability estimate increase with increase in effective sample size m. Reliability estimate increase as k (no. of units in parallel system) increases. Reliability estimate in case of progressive TypeII censoring scheme are better than reliability estimate in case of conventional TypeII censoring scheme.

Conclusion
The study reveals that for small sample size n and the smallest effective sample size m, EM algorithm method works well. Overall both conventional TypeII censoring scheme and progressive TypeII censoring scheme give better results. According to MSE, conventional TypeII censoring method are superior to the progressive TypeII censoring method, while CIs perform well in case of progressive TypeII censoring methods. In this study both conventional and progressive censoring methods give better performance. In many situations units are removed or lost from the experiment before the completion. For example individuals from clinical trial may drop from the experiment. In such situations we have
no alternative but to use progressive censoring method and analyze the data accordingly.

Cohen, A. C., 1963. Progressively censored samples in life testing, Technometrics, 5, 327329.

Mann, N. R.,1969. Exact threeorderstatistic confidence bounds on reliable life for a Weibull model with progressive censoring, Journal of the American Statistical Association, 64, 306315.

Mann, N. R., 1971. Best linear invariant estimation for Weibull parameters under progressive censoring, Technometrics, 13, 521533.

Balkrishnan N. and Asgharzadeh A. , 2005. Inference for the Scaled HalfLogistic Distribution based on progressively TypeII censored samples , Communication in Statistics Theory and Methods, 34, 7387.

Balakrishnan, N., Kannan, N., Lin, C. T. and Ng, H.
K. T., 2003. Point and interval estimation for Gaussian distribution, based on Progressively Type II Censored samples, IEEE Transactions on Reliability, 52(1), 9095. (2)

Balakrishnan, N., Kannan, N., Lin, C. T. and Wu, S.

S. , 2004. Inference for the extreme value distribution under Progressive TypeII Censoring, Journal of Statistical Computation and Simulation, 74, 2545.


Ng, H. K. T., 2005. Parameter estimation for a modified Weibull distribution, for progressively TypeII censored samples, IEEE transactions on Reliability, 54, 374380.

Balakrishnan, N. and Aggarwala, R., 2000. Progressive Censoring: Theory, Methods and Applications, Birkhauser, Boston.

Balakrishnan, N., 2007. Progressive Censoring Methodology: An Appraisal. Test, 16(2), 211259.

Pradhan, B., 2007. Point and Interval Estimation for the Lifetime Distribution of a kUnit Parallel System Based on Progressively TypeII Censored Data, Economic Quality Control, 22, 2, 175186.

Chein, T. L. and Balkrishnan N., 2011. Asymptotic properties of maximum likelihood estimators based on progressively TypeII censoring, Metrika, 74, 3, 349360

Iliopoulous, G. and Balkrishnan L., 2011. Exact Likelihood Inference for Laplace Distribution based on TypeII censored samples, Journal of Statistical Planning and Inference , 141, 12241239.

Krishna H. and Malik M., 2012. Reliability estimation in Maxwell distribution with progressively censored data. J. Stat. Comput. Simul. 82: 623641.

Potdar K. G. and Shirke D. T., 2012. Inference for the scale parameter of lifetime distribution of kunit parallel system based on progressively censored data. J. Stat. Comput. Simul. Online published http://dx.doi.org/10.1080/00949655. 2012. 700314:115

Potdar K. G. and Shirke D. T., 2012. Inference for the distribution of a kunit parallel system with exponential distribution as the component life distribution based on Type II progressively censored sample. Int. J. Agricult. Stat. Sci. 8(2): 503517.

Potdar K. G. and Shirke D. T., 2013. Inference for the parameters generalized inverted family of distributions. ProbStat Forum. 6: 1828.

Dempster, A. P. , Laird, N. M. and Rubin, D. B., 1977. Maximum Likelihood from Incomplete Data via the EM algorithm, Journal of the Royal Statistical Society, B, 39, 138.

Mclachlan, G. J. and Krishnan, T., 1997. The EM Algorithm and Extensions, John Wiley and sons.

Little, Roderick J. A. and Rubin, Donald B., 2002. Statistical analysis with missing data, Wiley Series in Probability and Statistics.

Pradhan, B. and Kundu, D., 2009. On Progressively Censored Generalized Exponential Distribution, Test, 18, 497515.

Ng, H. K. T., Chan, P.S. and Balkrishnan N., 2002. Estimation of parameters from progressively censored data using EM algorithm, Computational Statistics and Data Analysis, 39, 371386.

Louis, T.A., 1982. Finding the observed information matrix using the EM algorithm, Journal of Royal Statistical Society, B, 44, 226233

Meeker, W. Q. and Escober, L.A., 1998. Statistical Methods for Reliability Data, John Wiley & Sons, New York.

Balakrishnan, N. and Sandhu, R. A., 1995. A simple simulation algorithm for generating progressive TypeII censored samples, The American Statistician, 49, 229230.

Leibein, J. and Zelen, M. (1956). Statistical Investigation of the Fatigue Life of DeepGroove Ball Bearings, Journal of Research of the National Bureau of Standards, 57, 273316.

Lawless, J. F. (1982). Statistical Models and Methods of Lifetime Data, John Wiley and Sons.

Raqab, M. Z. and Madi, M. T. (2002). Bayesian Prediction of the Total Time on Test Using Doubly Censored Rayleigh data, Journal of Statistical Computation and Simulation, 72, 781789.

G. Potdar is an Assistant Professor of Statistics at the Ajara Mahavidyalaya, Ajara. He received the B.Sc. in 1990, M.Sc. in 1992 and M.Phil. in 2008 in Statistics from Shivaji University, Kolhapur, Maharashtra. His area of interest includes reliability theory and inference.
D. T. Shirke is Professor of Statistics at the Shivaji University, Kolhapur. He received the B.Sc. in 1985, M.Sc. in 1987 and Ph.D. in Statistics in 1993 from Shivaji University, Kolhapur, Maharashtra. His areas of interest include reliability theory, inference, industrial statistics, survival analysis and data mining.