Journal of Computational and Applied Mathematics 315 (2017) 228–239

Contents lists available at ScienceDirect

Journal of Computational and Applied

Mathematics
journal homepage: www.elsevier.com/locate/cam

Estimation of the inverse Weibull parameters under adaptive

type-II progressive hybrid censoring scheme
M. Nassar ∗ , O.E. Abo-Kasem
Department of Statistics, Faculty of Commerce, Zagazig University, Egypt

article info abstract

Article history: This paper describes the frequentist and Bayesian estimation for the scale parameter λ
Received 12 August 2016 and shape parameter β of the inverse Weibull (IW) distribution based on adaptive type-
Received in revised form 12 October 2016 II progressive hybrid censoring scheme (AT-II PHCS). We discuss the maximum likelihood
estimators (MLEs) and the approximate MLEs, where the MLEs cannot be obtained in closed
Keywords: forms. The Bayes estimates for the IW parameters are obtained based on squared error (SE)
Inverse Weibull distribution
loss function by using the approximation form of Lindley (1980). The optimal censoring
Adaptive type-II progressive hybrid
censoring scheme
scheme has been suggested using two different optimality criteria. A real life data set is used
Maximum likelihood estimation for illustration purpose. Finally, the different proposed estimators have been compared
Bayesian estimation through an extensive simulation studies.
Lindley approximation © 2016 Elsevier B.V. All rights reserved.

Contents

1. Introduction............................................................................................................................................................................................. 228
2. Maximum likelihood estimation............................................................................................................................................................ 230
3. Approximate maximum likelihood estimation (AMLE) ....................................................................................................................... 230
4. Bayesian estimation ................................................................................................................................................................................ 231
5. Data analysis and simulation results ..................................................................................................................................................... 233
5.1. Real data example....................................................................................................................................................................... 233
5.2. Simulation study ......................................................................................................................................................................... 235
6. Optimal censoring scheme ..................................................................................................................................................................... 237
7. Conclusion ............................................................................................................................................................................................... 239
Acknowledgment .................................................................................................................................................................................... 239
References................................................................................................................................................................................................ 239

1. Introduction

If the random variable Y has a Weibull distribution, then the random variable X = Y −1 has an IW distribution with
probability density function (pdf), cumulative distribution function (cdf) and hazard rate function respectively given by

f (x) = λβ x−(β+1) e−λx

−β
, x > 0, λ, β > 0, (1)

∗ Corresponding author.
E-mail address: mezo10011@gmail.com (M. Nassar).

http://dx.doi.org/10.1016/j.cam.2016.11.012
0377-0427/© 2016 Elsevier B.V. All rights reserved.
 M. Nassar, O.E. Abo-Kasem / Journal of Computational and Applied Mathematics 315 (2017) 228–239 229

(a) Experiment terminates before time T .

(b) Experiment terminates after time T .

Fig. 1. Schematic representation of AT-II PHCS.

and
−β
F (x) = e−λx , (2)
and
 −1
h(x) = λβ x−(β+1) eλx
−β
−1 .
The IW distribution is more appropriate model than the Weibull distribution because the Weibull distribution does not
provide a satisfactory parametric fit if the data indicate a non-monotone and unimodal hazard rate functions. The hazard
rate function of IW distribution can be decreasing or increasing depending on the value of the shape parameter. The IW
distribution is useful to model several data such as the time to breakdown of an insulating fluid subjected to the action of
a constant tension and degradation of mechanical components such as pistons and crankshafts of diesel engines. Extensive
work has been done on the IW distribution, see for example, [1–4] and for more details about the generalizations of IW
distribution see [5]. In addition, many articles have considered IW distribution under different censoring schemes. Among
others, Kundu and Howlader [6], Musleh and Helu [7], Sultan et al. [8] and Xiuyun and Zaizai [9].
Progressive hybrid censoring scheme in the context of life testing experiments was introduced by Kundu and Joarder [10].
They considered a type-I progressive hybrid censoring scheme, in which n identical units are placed on test with
predetermined progressive censoring scheme R1 , R2 , . . . , Rm . At the time of the first failure x1:m:n , R1 units are randomly
removed from the remaining n − 1 surviving units. Similarly, at the time of the second failure x2:m:n , R2 units of the remaining
n − 2 − R1 units are randomly removed and so on. The experiment is terminated at random time T ∗ = min {xm:m:n , T }
where xm:m:n is the mth failure and T ∈ (0, ∞) is a predetermined time. For more details see also [11]. The drawback of
the type-I progressive hybrid censoring scheme is that the effective sample size is random and it can turn out to be a very
small number, therefore, the statistical inference procedure may not be applicable or will have low efficiency. Ng et al. [12]
introduced an AT-II PHCS to increase the efficiency of statistical analysis and save the total test time and analyzed the data
under the assumption of exponential distribution. In AT-II PHCS the effective number of failures m is fixed in advance and
the experimental time is allowed to run over time T which is an ideal total test. In this case, the progressive censoring
scheme R1 , R2 , . . . , Rm is provided, but the values of some of the Ri may change accordingly during the experiment. If the
mth progressively censored observed failures occurs before time T (i.e. Xm:m:n < T ), the experiment stops at this time
Xm:m:n , and we will have a usual type-II progressive censoring scheme with the pre-fixed progressive censoring scheme
R1 , R2 , . . . , Rm (Fig. 1(a)). Otherwise, if XJ :m:n < T < XJ +1:m:n , where J + 1 < m and XJ :m:n is the Jth failure time occur before
time T , then we will not withdraw any items from the experiment by setting RJ +1 , RJ +2 , . . . , Rm−1 = 0 and at the time of
J
the mth failure all remaining surviving items are removed, i.e., Rm = n − m − i=1 Ri (Fig. 1(b)). The main advantage of
this scheme is to speed up the experiment when the experiment duration exceed the predetermined time T and assures us
to get the effective number of failures m.
Many authors have considered AT-II PHCS. Lin et al. [13], discussed the maximum likelihood and approximate maximum
likelihood estimators for the Weibull distribution. Hemmati and Khorram [14], studied the maximum likelihood and
approximate maximum likelihood estimators for the log-normal distribution. Mahmoud et al. [15], investigated the
maximum likelihood and Bayes estimates of the unknown parameters of Pareto distribution. Ashour and Nassar [16] showed
230 M. Nassar, O.E. Abo-Kasem / Journal of Computational and Applied Mathematics 315 (2017) 228–239

the MLE and asymptotic confidence intervals in the presence of competing risks. Ismail [17], considered estimation problems
of Weibull distribution under step-stress partially accelerated life test model. Sobhi and Soliman [18], discussed the problem
of estimating parameters of the exponentiated Weibull distribution. Hemmati and Khorram [19], discussed the AT-II PHCS
in the presence of competing risks with unknown causes of failures.
This paper aims to investigate the classical and Bayesian estimation for the unknown parameters of the IW distribution
based on AT-II PHCS. It is observed that the MLEs cannot be obtained in explicit forms and we suggest to use the approximate
MLEs by expanding the normal equations using Taylor expansion. The Bayes estimates are obtained based on SE loss function
under the assumption of independent gamma priors. The rest of this paper is organized as follows: In Section 2, we discuss
the maximum likelihood estimation. In Section 3, we derive the approximate MLEs. Bayesian estimates under SE loss
functions using Lindley approximation [20] is provided in Section 3. In Section 4, Monte Carlo simulation results and the
analysis of real data set are presented. We provide the optimal censoring scheme in Section 6. The paper is concluded in
Section 7.

2. Maximum likelihood estimation

Let x1:m:n < · · · < xJ :m:n < T < xJ +1:m:n < · · · < xm:m:n be an adaptive type- II progressive censored sample of size
J
m from IW distribution with pdf (1) and cdf (2) with censoring scheme R1 , . . . , RJ , 0, . . . , 0, n − m − i=1 Ri , then, the
likelihood function ignoring the additive constant can be written as
m
 β m J
−λ xi β β
β+1
 
L(λ, β) = λ β e m m i=1 (1 − e−λxm )Rm xi (1 − e−λxi )Ri (3)
i=1 i =1

i:m:n for simplicity of notation. The logarithm of the likelihood function (ℓ) is
where xi = x− 1

m m J
 β
  β β
ℓ = m ln(λβ) − λ xi + (β + 1) ln(xi ) + Ri ln(1 − e−λxi ) + Rm ln(1 − e−λxm ). (4)
i=1 i =1 i=1

The MLEs of λ and β can be obtained by deriving (4) with respect to λ and β and equating the results to zero as follows
J
∂ℓ m
m
β β
Ri xi ui + Rm xβm um
 
= − xi + (5)
∂λ λ i=1 i=1

J
∂ℓ m
m
β
m
β
Ri xi ui ln(xi ) + λRm xβm um ln(xm )
  
= −λ xi ln(xi ) + ln(xi ) + λ (6)
∂β β i =1 i =1 i =1
β β β
where ui = e−λxi (1 − e−λxi )−1 = (eλxi − 1)−1 . It can be seen that Eqs. (5) and (6) cannot be solved explicitly, so, numerical
methods are used to solve (5) and (6).

3. Approximate maximum likelihood estimation (AMLE)

It is noted that from (5) and (6) the MLE of λ and β cannot be obtained in explicit form, we derive the AMLEs which have
explicit forms. Let Z = − ln X , then Z ∼ extreme value distribution with pdf and cdf given respectively by
z−µ z−µ
   
1
g ( x) = exp − exp , −∞ < z < ∞, (7)
σ σ σ
and
z−µ
  
G(x) = 1 − exp − exp , (8)
σ
where µ = − ln λ/β and σ = 1/β are the location and scale parameters respectively. From (7) and (8) and under AT-II
PHCS the log likelihood function can be written as follows
m
 m
 J

ℓ∗ = −m ln(σ ) + ti − e ti − R i e ti − R m e tm (9)
i=1 i =1 i =1

where ti = i σ , i = 1, . . . , m. To obtain the AMLEs we expand the function eti in Taylor series around the points
z −µ

νi = F −1 (1 − qi ) = ln(− ln(1 − qi )), where

m

m
j+ Rk
 k=m−j+1
qi = 1 − m
, i = 1, . . . , m.

j=m−i+1 j+1+ Rk
k=m−j+1
 M. Nassar, O.E. Abo-Kasem / Journal of Computational and Applied Mathematics 315 (2017) 228–239 231

Keeping only the first order derivatives, we have

eti ≃ αi + βi ti ,
where
αi = eνi (1 − νi ) and βi = eνi ≥ 0.
Using these linear approximations, we can obtain the approximate log likelihood equations as
 
J
∂ℓ∗ 1
m
 
≃− m− (αi + βi ti ) − Ri (αi + βi ti ) − Rm (αm + βm tm ) = 0,
∂µ σ i =1 i=1

and
 
J
∂ℓ∗ −1 m 
≃ m+ ti (1 − (αi + βi ti )) − Ri ti (αi + βi ti ) − R tm (αm + βm tm ) = 0.
∗
(10)
∂σ σ i=1 i=1

The solutions of Eqs. (10) are

ˆ = A + σ̂ˆ B
µ̂
and
√
D+ D2 + 4mE
σ̂ˆ =
2m
where
m J
βi xi + Ri βi xi + Rm βm xm
 
i =1 i=1
A=
J
,
m
βi + Ri βi + Rm βm
 
i =1 i =1
m J
αi + Ri αi + Rm αm − m
 
i=1 i =1
B=
J
,
m
βi + Ri βi + Rm βm
 
i=1 i=1
m
 J

D= (αi − 1)(xi − A) + Ri αi (xi − A) + Rm αm (xm − A),
i =1 i=1

and
m
 J

E= βi (xi − A)2 + Ri βi (xi − A)2 + Rm βm (xm − A)2 .
i=1 i=1

It is noted that the AMLEs have explicit solutions and they can be used as initial values to obtain the MLEs in (5) and (6).

4. Bayesian estimation

In this section, the SE loss function is used to obtain the Bayes estimates of the unknown parameters λ and β . The Bayes
estimates are considered under the assumption of independent gamma priors of λ and β with the following joint density

g (λ, β) ∝ λa−1 β c −1 e−[bλ+dβ] , a, b, c , d > 0. (11)

Based on the likelihood function (3) and the joint prior density (11), the joint posterior distribution of λ, β and the data is
 
m
β J
xi +b −dβ m

−λ β β
β+1
 
g λ, β y ∝ λm+a−1 β m+c −1 e (1 − e−λxm )Rm (1 − e−λxi )Ri
  
i=1
xi (12)
i=1 i =1

under SE loss function, the Bayes estimate of the function of the parameters λ and β , say U (λ, β) is

U (λ, β) eQ (λ,β) .dλ.dβ

∞∞ ∞∞
U (λ, β)L(λ, β)g (λ, β) dλ dβ
Ũ (λ, β) = 0 0
∞∞ = 0
0 ∞  ∞ (13)
0 0
L(λ, β)g (λ, β) dλ dβ 0 0
eQ (λ,β) dλ.dβ
232 M. Nassar, O.E. Abo-Kasem / Journal of Computational and Applied Mathematics 315 (2017) 228–239

where Q (λ, β) = ln L(λ, β) + ln g (λ, β) ≡ ℓ + ρ(a, b). It is not possible to compute the ratio of the two integrals given by
(13) in a closed form. Therefore, in such situation, we suggest using Lindley’s approximation to obtain the Bayes estimates
of the unknown parameters. For the two parameter case, (λ, β) = (θ1 , θ2 ), Lindley’s approximation of (13) takes the form
1  1 1
Ũ (θ ) = U (θ ) + uij σij + Uj ρj + L30 σ11 U1 + L21 (2σ12 U1 + σ11 U2 )
2 2 2
1 1
+ L12 (σ22 U1 + 2σ12 U2 ) + L03 σ22 U2 , (14)
2 2
where
∂ η+ζ L(θ ) ∂ U (θ ) ∂ 2 U (θ )
Lηζ = , η, ζ = 0, 1, 2, 3 and η + ζ = 3, ui = , uij = for i, j = 1, 2,
∂ η θ1 ∂ ζ θ2 ∂θi ∂θi ∂θj
∂ρ(θ)
ui σki , ρj = j = 1, 2 and σij is the (i, j)th element of the inverse of the Fisher information matrix.

Uk = i ∂θj
,
Evaluation of all functions in Eq. (14) at the MLE of (θ1 , θ2 ), produces the approximation Ũ (θ ) to (13). Now, to apply Lindley’s
approximation (14), we obtain the quantities
J
∂ 2ℓ m  2β̂ 2β̂
L20 = − = + Ri xi ϕi u2i + Rm xm ϕm u2m ,
∂λ2 λ̂2
i =1
J
∂ 2ℓ m m
β̂
 β̂

β̂

L02 = − 2 = + λ̂ xi ln(xi )2 − λ̂ Ri xi ln(xi )2 ui 1 − λ̂xi ϕi ui
∂β β̂ 2 i=
1  i =1
β̂
− λ̂Rm xm ln(xm )2 um 1 − λ̂xβ̂m ϕm um ,
J
∂ 2ℓ  m
β̂
 
2β̂ β̂

L11 = − = xi ln(xi ) + Ri λ̂xi ln(xi )ϕi ui + xi ui ln(xi )
∂λ∂β i =1 i=1
 
2β̂
+ Rm λ̂xm ln(xm )ϕm um + xβ̂m um ln(xm ) ,
J
∂ 3ℓ 2m 3β̂
Ri xi ϕi u2i [1 − 2ϕi ui ] − Rm x3mβ̂ ϕm u2m [1 − 2ϕm um ] ,

L30 = = −
∂λ3 λ̂ 3
i=1
J
∂ ℓ
3
2m m
β̂
 
β̂

β̂

L03 = = − λ̂ x ln (x i ) 3
+ λ̂ Ri ln(xi )2 xi ui ln(xi ) 1 − λ̂xi ϕi ui
∂β 3
β̂ 3 i=
i
i=1 
1
2β̂ β̂ β̂
− λ̂xi ϕi u2i ln(xi ) 2 + λ̂xi − 2λ̂xi ϕi ui + λ̂Rm ln(xm )2
    
× xβ̂m um ln(xm ) 1 − λ̂xβ̂m ϕm um − λ̂xm 2β̂
ϕm u2m ln(xm ) 2 + λ̂xβ̂m − 2λ̂xβ̂m ϕm um ,
J
∂ 3ℓ  
2β̂

β̂ β̂

2β̂

L21 = = − Ri x ln( xi )ϕi u 2
− 2 λ̂x ϕi u i + λ̂ x + 1 + x ln (x i )ϕ i u 2
∂λ2 ∂β i=1
i i i i i i
   
2β̂
− Rm xm ln(xm )ϕm u2m −2λ̂xβ̂m ϕm um + λ̂xβ̂m + 1 + x2mβ̂ ln(xm )ϕm u2m ,
and
J
∂ 3ℓ m
β̂
 
β̂

β̂

2β̂
L12 = = − x ln (x i )2
+ R i ln(x i )2
x ln( xi )u i 1 − λ̂ x ϕi u i − λ̂xi ϕi u2i
∂β 2 ∂λ i=1
i i i
 i=1 
β̂ β̂
   
× 2 + λ̂xi − 2λ̂xi ϕi ui + Rm ln(xm )2 xβ̂m ln(xm )um 1 − λ̂xβ̂m ϕm um − λ̂x2mβ̂ ϕi u2m
 
× 2 + λ̂xβ̂m − 2λ̂xβ̂m ϕm um ,
β
where ϕi = eλxi . The elements of the variance covariance matrix σij , can be obtained as
L02 L20 −L11
σ11 = , σ22 = and σ12 = σ21 = .
L20 L02 − L211 L20 L02 − L211 L20 L02 − L211
Based on the joint prior function (11), we obtain
ρ(λ, β) = ln g (λ, β) ∝ (a − 1) ln λ + (c − 1) ln β − (bλ + dβ)
hence
∂ρ(λ, β) a−1 ∂ρ(λ, β) c−1
ρ1 = = − b and ρ2 = = − d.
∂λ λ ∂β β
 M. Nassar, O.E. Abo-Kasem / Journal of Computational and Applied Mathematics 315 (2017) 228–239 233

Table 1
MLEs, AMLEs and Bayes estimates of λ and β .
Scheme MLEs AMLEs Bayes estimates
λ β λ β λ β
Complete 2.157 1.392 282.762 4.277 2.149 1.387
1 2.335 0.965 68.277 4.086 2.325 0.957
2 2.288 1.128 33.14 2.799 2.279 1.121
3 1.984 1.353 102.357 3.550 1.975 1.348

If U (λ, β) = λ, u1 = 1, u2 = 0 and u11 = u12 = u21 = u22 = 0, then

L30 σ11
2
+ 3L21 σ11 σ12 + L12 (σ11 σ22 + 2σ12
2
) + L03 σ12 σ22
λ̃ = λ̂ + ρ1 σ11 + ρ2 σ12 + .
2
Similarly, when U (λ, β) = β , hence, u2 = 1, u1 = 0 and u11 = u12 = u21 = u22 = 0. Thus

L30 σ11 σ12 + L21 2σ12 + σ11 σ22 + 3L12 σ22 σ12 + L03 σ22

2
 2
β̃ = β̂ + ρ1 σ12 + ρ2 σ22 + .
2

5. Data analysis and simulation results

The purpose of this section is to compare the performance of the different methods of estimation discussed in the previous
sections. We analyze a real data set for illustrative purpose; also, a simulation study is carried out to check the behavior of the
proposed methods as well as to evaluate the statistical performances of the estimators under different sampling schemes.

5.1. Real data example

In this subsection, we reanalyze a real data set given by Blischke and Murthy [21]. The data set consists of 87 observations
and it represents the failure times of aircraft windshields. Blischke and Murthy [21] proved that the IW distribution gives a
good fit for it. From the original data we generate three adaptive progressively hybrid censored samples with the following
schemes
Scheme m T Censoring scheme
1 40 2.5 {0 ∗ 20, 1 ∗ 19, 28}
2 50 3 {1 ∗ 25, 0 ∗ 24, 12}
3 70 3.5 {17, 0 ∗ 69}

the generated adaptive progressively hybrid censored samples given in the following table

Scheme Censored data

1 0.301 0.309 0.557 0.943 1.070 1.124 1.248 1.281 1.281 1.303 1.480
1.505 1.506 1.568 1.615 1.619 1.652 1.652 1.757 1.795 1.866 1.899
1.911 1.981 2.010 2.085 2.089 2.097 2.154 2.194 2.223 2.300 2.349
2.481 2.625 2.632 2.661 2.823 2.902 2.934

2 0.301 0.309 0.557 0.943 1.124 1.248 1.281 1.281 1.432 1.480 1.505
1.568 1.615 1.652 1.652 1.795 1.866 1.876 1.899 1.911 1.912 1.981
2.038 2.085 2.135 2.154 2.194 2.224 2.229 2.324 2.385 2.610 2.632
2.646 2.688 2.890 2.934 2.962 3.000 3.114 3.117 3.344 3.376 3.385
3.467 3.478 3.578 3.595 3.699 3.779

3 0.301 0.309 0.557 0.943 1.070 1.248 1.281 1.281 1.432 1.480 1.505
1.506 1.615 1.619 1.652 1.652 1.757 1.795 1.866 1.876 1.899 1.911
1.912 1.914 1.981 2.038 2.085 2.089 2.097 2.135 2.154 2.190 2.194
2.223 2.229 2.324 2.349 2.481 2.610 2.625 2.632 2.646 2.661 2.688
2.823 2.890 2.934 2.962 3.000 3.103 3.117 3.344 3.376 3.385 3.467
3.478 3.578 3.595 3.779 4.035 4.121 4.240 4.255 4.278 4.305 4.449
4.485 4.570 4.602 4.694
234 M. Nassar, O.E. Abo-Kasem / Journal of Computational and Applied Mathematics 315 (2017) 228–239

Table 2
Different priors for the two sets of parameters of λ and β .
Parameters Prior

(λ = 0.5, β = 1) and (λ = 0.75, β = 2) Prior 0: a = b = c = d = 0.

(λ = 0.5, β = 1) Prior 1: a = 1, b = 2, c = 1, d = 1.
(λ = 0.75, β = 2) Prior 2: a = 3, b = 4, c = 4, d = 2.

Table 3
The average values and MSEs of the parameters (α, β) = (0.5, 1) under different censoring schemes and for T = 0.5.
(n, m) Scheme Parameter AML estimates ML estimates Bayes estimates
Prior 0 Prior 1

(30, 10) 1 λ 0.84120 0.62652 0.67608 0.63785

0.12074 0.08001 0.09109 0.06194
β 1.38931 1.01470 1.01639 1.00692
0.16330 0.10278 0.09627 0.06646
2 λ 0.80237 0.55475 0.59505 0.57089
0.09563 0.08011 0.08183 0.05796
β 1.43781 1.05662 1.02897 1.03481
0.19823 0.11566 0.10672 0.07282
3 λ 0.76593 0.50452 0.53115 0.51538
0.07328 0.07075 0.06965 0.04991
β 1.48584 1.12891 1.10470 1.09623
0.24096 0.15532 0.14096 0.09344

(50, 10) 1 λ 0.89938 0.72228 0.79207 0.72570

0.16010 0.12108 0.15587 0.09724
β 1.46719 1.03026 1.06119 1.01944
0.22143 0.08340 0.08097 0.05013
2 λ 0.88584 0.68010 0.74499 0.38939
0.14959 0.09758 0.12429 0.07874
β 1.49804 1.01656 1.04696 1.01312
0.25100 0.08649 0.08307 0.05179
3 λ 0.83352 0.55118 0.50708 0.51742
0.11231 0.06373 0.06171 0.04452
β 1.56630 1.18009 1.15314 1.12570
0.32316 0.18966 0.17083 0.09651

(60, 15) 1 λ 0.87597 0.74573 0.78807 0.74330

0.14202 0.10340 0.12475 0.09172
β 1.45344 1.08966 1.11037 1.07632
0.20808 0.05249 0.05396 0.04199
2 λ 0.85861 0.65411 0.69224 0.66041
0.12949 0.06174 0.07405 0.05486
β 1.49069 1.04743 1.06676 1.04078
0.24306 0.04983 0.05005 0.03940
3 λ 0.80155 0.50084 0.52877 0.51678
0.09195 0.04940 0.04966 0.03961
β 1.54896 1.07197 1.05670 1.05788
0.30328 0.09662 0.09179 0.06853

(60, 20) 1 λ 0.83662 0.66112 0.68658 0.66317

0.11570 0.05742 0.06507 0.05247
β 1.41143 1.07117 1.08558 1.06426
0.17351 0.04028 0.04134 0.03535
2 λ 0.79318 0.57892 0.59918 0.58438
0.08829 0.05319 0.05519 0.04592
β 1.45456 1.02938 1.04213 1.02974
0.20926 0.05373 0.05295 0.04678
3 λ 0.75834 0.52652 0.54495 0.53430
0.06810 0.04541 0.04580 0.03829
β 1.50145 1.02493 1.01455 1.02196
0.25351 0.05805 0.05638 0.04933

The MLEs, AMLEs and Bayes estimates using Lindley’s approximation are reported in Table 1. The Bayes estimates of λ and
β are computed based on a non-informative prior, i.e., a = b = c = d = 0. From Table 1, it is noted that the MLEs and Bayes
estimates of the unknown parameters are quite close to each other. Also, the estimates based on adaptive progressively
 M. Nassar, O.E. Abo-Kasem / Journal of Computational and Applied Mathematics 315 (2017) 228–239 235

Table 4
The average values and MSEs of the parameters (α, β) = (0.5, 1) under different censoring schemes and for T = 1.
(n, m) Scheme Parameter AML estimates ML estimates Bayes estimates
Prior 0 Prior 1

(30, 10) 1 λ 0.84926 0.73272 0.78493 0.72916

0.12332 0.11794 0.14242 0.09645
β 1.39117 1.04086 1.07073 1.02981
0.15957 0.08473 0.08296 0.05671
2 λ 0.82409 0.60422 0.65013 0.61638
0.10653 0.05799 0.06810 0.04571
β 1.44180 1.02692 1.00187 1.02223
0.20065 0.08288 0.07749 0.05450
3 λ 0.78353 0.51580 0.58024 0.58026
0.08306 0.06699 0.06029 0.04532
β 1.48678 1.31944 1.29266 1.23237
0.24150 0.32484 0.29855 0.16178

(50, 10) 1 λ 0.89989 0.71648 0.78650 0.72134

0.16050 0.11728 0.15136 0.09437
β 1.46482 1.02301 1.05425 1.01192
0.21908 0.07936 0.07642 0.04864
2 λ 0.88556 0.69129 0.75653 0.69904
0.14932 0.10087 0.12912 0.08178
β 1.49425 1.01951 1.04985 1.01299
0.24722 0.08411 0.08086 0.04973
3 λ 0.84124 0.63707 0.59261 0.58789
0.11734 0.05350 0.04442 0.03295
β 1.56069 1.28768 1.25882 1.21168
0.31697 0.22532 0.20335 0.11488

(60, 15) 1 λ 0.87473 0.77594 0.81770 0.76998

0.14104 0.12545 0.14904 0.11080
β 1.44579 1.09747 1.11672 1.07996
0.20162 0.05273 0.05500 0.04232
2 λ 0.85548 0.72880 0.76672 0.72682
0.12704 0.09645 0.11421 0.08560
β 1.48228 1.08661 1.10513 1.07331
0.23530 0.05005 0.05186 0.04064
3 λ 0.81379 0.60824 0.58022 0.58037
0.09948 0.04596 0.04095 0.03381
β 1.54321 1.22298 1.20485 1.18671
0.29723 0.15309 0.14204 0.10617

(60, 20) 1 λ 0.83568 0.81355 0.83929 0.80397

0.11345 0.13249 0.14835 0.12113
β 1.40612 1.14773 1.16119 1.13289
0.16856 0.04968 0.05294 0.04320
2 λ 0.80755 0.69775 0.72027 0.69518
0.09546 0.06476 0.07353 0.05949
β 1.45627 1.10287 1.11557 1.09319
0.21124 0.03802 0.03999 0.03346
3 λ 0.76361 0.54139 0.52242 0.52791
0.07080 0.03354 0.03203 0.02736
β 1.50307 1.11271 1.10111 1.10137
0.25555 0.08441 0.08013 0.06772

hybrid type-II samples are very close to those of complete data and this demonstrate the importance and usefulness of AT-II
PHCS.

5.2. Simulation study

This subsection aims to conduct a Monte Carlo simulation study to compare the performance of the different methods of
estimation proposed in the previous sections. The performance of all the estimates is compared in terms of their mean
square errors (MSEs). The average values and MSEs of MLEs, AMLEs and Bayes estimates of λ and β are evaluated for
(λ, β) = (0.5, 1) and (λ, β) = (0.75, 2) and by considering different values n, m, T and three censoring schemes as
follows
236 M. Nassar, O.E. Abo-Kasem / Journal of Computational and Applied Mathematics 315 (2017) 228–239

Table 5
The average values and MSEs of the parameters (λ, β) = (0.75, 2) under different censoring schemes and for T = 1.
(n, m) Scheme Parameter AML estimates ML estimates Bayes estimates
Prior 0 Prior 2

(30, 10) 1 λ 0.80937 0.97243 1.01723 0.96578

0.00670 0.12133 0.13846 0.07229
β 2.40940 2.02838 2.08998 2.09384
0.27804 0.32448 0.31099 0.09764
2 λ 0.77352 0.82462 0.86329 0.85689
0.00463 0.06817 0.07100 0.03292
β 2.56589 2.10313 2.04529 2.05653
0.39789 0.37125 0.34155 0.09090
3 λ 0.77801 0.81926 0.78693 0.75430
0.00422 0.09084 0.08408 0.03356
β 2.70167 2.40655 2.35887 2.02725
0.54953 0.92012 0.85149 0.19757

(50, 10) 1 λ 0.87252 1.05224 1.11549 1.02722

0.01585 0.18545 0.22284 0.10679
β 2.63341 2.08213 2.14326 2.10348
0.43911 0.34465 0.33693 0.09149
2 λ 0.85626 1.02066 1.07942 1.00374
0.01219 0.16279 0.19344 0.09451
β 2.72912 2.08233 2.14169 2.11278
0.56563 0.32825 0.32144 0.09292
3 λ 0.80559 0.86882 0.82355 0.75794
0.00435 0.08934 0.07901 0.03493
β 2.94120 2.44654 2.38578 2.01157
0.91059 0.79588 0.72915 0.78184

(60, 15) 1 λ 0.84135 1.07890 1.13491 1.06367

0.00912 0.17779 0.20186 0.12897
β 2.56649 2.19988 2.23831 2.15770
0.35271 0.20610 0.21579 0.11154
2 λ 0.81974 1.04386 1.07635 1.01778
0.00568 0.13741 0.15549 0.09975
β 2.68623 2.18700 2.22380 2.16041
0.49808 0.19983 0.20813 0.11244
3 λ 0.76626 0.79554 0.76933 0.75069
0.00144 0.06345 0.06002 0.03192
β 2.88341 2.22991 2.19805 2.07146
0.79959 0.39298 0.36808 0.12119

(60, 20) 1 λ 0.79880 1.05816 1.07993 1.03358

0.00368 0.13076 0.14337 0.10414
β 2.46771 2.22537 2.25353 2.19499
0.25553 0.15954 0.16944 0.11358
2 λ 0.76764 0.86429 0.88375 0.86827
0.00222 0.04151 0.04522 0.03185
β 2.62354 2.05494 2.08244 2.07294
0.41512 0.13237 0.13210 0.08740
3 λ 0.78881 0.77950 0.76286 0.75235
0.00281 0.03809 0.03636 0.02327
β 2.74665 2.15901 2.13672 2.08142
0.57809 0.21259 0.20164 0.10755

Scheme 1: R1 = · · · = Rm−1 = 0 and Rm = n − m,

Scheme 2: R1 = · · · = Rm−1 = 1 and Rm = n − 2m + 1, and
Scheme 3: R1 = · · · = Rm−1 = Rm = (n − m)/m.
We replicate the process 1000 times. The average values and MSEs are tabulated in Tables 3–6. The Bayes estimates are
all computed under non-informative and informative priors. The priors for the two sets of parameters are given in Table 2.
We use prior 0 for the two sets of parameter values, Prior 1 for (λ, β) = (0.5, 1) and prior 2 for (λ, β) = (0.75, 2). From
Table 2, it may be noted that prior 1 describes the case of non-informative prior and prior 1 and prior 2 are chosen in such
way that prior means equal to the original means.
From Tables 3–6, it can be seen that the Bayes estimates using prior 1 perform better than those by using AMLEs, MLEs
and Bayes estimates using prior 0 in terms of minimum MSEs for (λ, β) = (0.5, 1), while the AMLEs has the minimum MSEs
among all other methods for (λ, β) = (0.75, 2). It is observed that the Bayes estimates with respect to the non-informative
 M. Nassar, O.E. Abo-Kasem / Journal of Computational and Applied Mathematics 315 (2017) 228–239 237

Table 6
The average values and MSEs of the parameters (λ, β) = (0.75, 2) under different censoring schemes and for T = 1.5.
(n, m) Scheme Parameter AML estimates ML estimates Bayes estimates
Prior 0 Prior 2

(30, 10) 1 λ 0.80063 1.08155 1.12358 1.04211

0.00484 0.19673 0.22080 0.12138
β 2.35464 2.11599 2.17462 2.13630
0.24384 0.35206 0.34848 0.11575
2 λ 0.76925 0.95569 0.99240 0.94513
0.00269 0.11513 0.12717 0.06750
β 2.52477 2.02937 2.08430 2.09950
0.36646 0.32898 0.31716 0.10404
3 λ 0.76182 0.92067 0.88585 0.80097
0.00339 0.10518 0.09175 0.03047
β 2.69557 2.67367 2.61847 2.04240
0.53999 1.29705 1.19010 0.18736

(50, 10) 1 λ 0.87179 1.06107 1.12404 1.03381

0.01571 0.19045 0.22894 0.11103
β 2.62097 2.07749 2.13861 2.10085
0.42686 0.34702 0.33905 0.09573
2 λ 0.80555 1.02722 1.08566 1.00775
0.01207 0.16575 0.19734 0.09785
β 2.72037 2.07678 2.13629 2.10681
0.55510 0.33217 0.32462 0.11298
3 λ 0.80381 0.86408 0.81913 0.75530
0.00396 0.06718 0.05780 0.01844
β 2.93084 2.42988 2.37646 2.04042
0.89270 0.65334 0.59091 0.06849

(60, 15) 1 λ 0.84097 1.09847 1.08445 1.06455

0.00917 0.18361 0.20742 0.13316
β 2.56394 2.19213 2.23079 2.15411
0.35049 0.20527 0.21443 0.11037
2 λ 0.84087 1.10032 1.09626 1.06585
0.00914 0.18444 0.20836 0.13353
β 2.56455 2.19536 2.23396 2.15630
0.35102 0.20795 0.21721 0.11316
3 λ 0.77329 0.86897 0.84140 0.80016
0.00159 0.05540 0.04869 0.02330
β 2.87405 2.38978 2.35430 2.17473
0.78316 0.45160 0.41630 0.12152
(60, 20) 1 λ 0.79119 1.08908 1.14962 1.09439
0.00280 0.18348 0.19805 0.14591
β 2.41656 2.27255 2.29980 2.23131
0.21571 0.18525 0.19727 0.13355
2 λ 0.75871 1.01324 1.03098 0.99195
0.00121 0.10422 0.11268 0.08199
β 2.58340 2.21404 2.23927 2.19336
0.37295 0.16030 0.16840 0.11776
3 λ 0.78150 0.83505 0.81766 0.79505
0.00246 0.04505 0.04146 0.02512
β 2.74098 2.30239 2.27811 2.17909
0.57189 0.34530 0.32513 0.15389

prior (prior 0) are quite close to the MLEs. For fixed n and T , when m increases the MSEs of MLEs, AMLEs and Bayes estimates
decrease except come cases. For fixed n and m, when T increases we do not observe any specific trend in the MSEs. In most
cases, for the parameter λ, the MSEs of scheme 3 are always smaller than other schemes, while, for the parameter β the
scheme 1 has the minimum MSEs, except some cases.

6. Optimal censoring scheme

In practice, it is very important to choose the optimum censoring scheme to obtain the highest amount of information
about the unknown parameters. In the case of AT-II PHCS, for fixed n and m, the practitioner might be interested to choose
the ideal time T and the progressive censoring scheme R1 , . . . , Rm , where n = m + i=1 Ri and RJ +1 , RJ +2 , . . . , Rm−1 = 0
m
if XJ :m:n < T < XJ +1:m:n , which provides more information about the unknown parameters under consideration. Here, we
238 M. Nassar, O.E. Abo-Kasem / Journal of Computational and Applied Mathematics 315 (2017) 228–239

Table 7
Different censoring planes for m = 30 and T = 2.

Scheme λ̂ β̂ V (Q̂0.05 ) V (Q̂0.95 ) φ

{0 ∗ 29, 57} 2.3080 0.9320 0.6870 48.3870 0.0820
{0 ∗ 20, 1 ∗ 9, 48} 2.3170 0.9220 0.7404 52.4120 0.0821
{1 ∗ 29, 28} 2.4009 0.8718 1.1089 80.881 0.0894
{2 ∗ 10, 1 ∗ 10, 0 ∗ 9, 27} 2.3683 0.8963 0.8726 63.136 0.0896
{1 ∗ 20, 2 ∗ 9, 19} 2.3599 0.9122 0.8193 56.501 0.0868
{57, 0 ∗ 29} 1.3827 1.2417 0.2051 4.2986 0.0806
{0 ∗ 10, 1 ∗ 10, 0 ∗ 9, 47} 2.3654 0.9044 0.8865 61.6079 0.0851

Table 8
Different censoring planes for m = 30 and T = 2.5.

Scheme λ̂ β̂ V (Q̂0.05 ) V (Q̂0.95 ) φ

{0 ∗ 29, 57} 2.3080 0.9320 0.6870 48.3870 0.0820
{0 ∗ 20, 1 ∗ 9, 48} 2.3890 0.9020 0.9680 64.0130 0.0870
{1 ∗ 29, 28} 2.4390 0.8490 1.3820 101.8880 0.0920
{2 ∗ 10, 1 ∗ 10, 0 ∗ 9, 27} 2.5140 0.8120 1.8590 148.9720 0.1000
{1 ∗ 20, 2 ∗ 9, 19} 2.4920 0.8150 1.8820 147.4520 0.0960
{57, 0 ∗ 29} 1.9755 1.3453 0.2105 5.3845 0.1522
{0 ∗ 10, 1 ∗ 10, 0 ∗ 9, 47} 2.3450 0.9140 0.8060 56.4180 0.0840

Table 9
Different censoring planes for m = 40 and T = 2.

Scheme λ̂ β̂ V (Q̂0.05 ) V (Q̂0.95 ) φ

{0 ∗ 39, 47} 2.2342 1.0629 0.3977 15.1367 0.0739
{0 ∗ 20, 1 ∗ 19, 28} 2.1421 1.1307 0.2224 9.8835 0.0714
{1 ∗ 39, 8} 2.3354 1.018 0.5452 20.4027 0.0821
{2 ∗ 15, 1 ∗ 10, 0 ∗ 14, 7} 2.5324 0.9137 1.1716 46.9090 0.0958
{1 ∗ 9, 2 ∗ 19, 0 ∗ 12} 2.4130 1.0077 0.6174 22.4141 0.0873
{47, 0 ∗ 39} 1.5517 1.2985 0.0443 3.3231 0.0759
{0 ∗ 15, 1 ∗ 15, 0 ∗ 9, 32} 2.2898 1.0057 0.5575 21.9797 0.0770

Table 10
Different censoring planes for m = 40 and T = 2.5.

Scheme λ̂ β̂ V (Q̂0.05 ) V (Q̂0.95 ) φ

{0 ∗ 39, 47} 2.2342 1.0629 0.3977 15.1367 0.0739
{0 ∗ 20, 1 ∗ 19, 28} 2.2612 1.0248 0.4949 19.2707 0.0753
{1 ∗ 39, 8} 2.2750 1.0316 0.4437 18.1368 0.0785
{2 ∗ 15, 1 ∗ 10, 0 ∗ 14, 7} 2.3690 1.0263 0.4804 19.2229 0.0870
{1 ∗ 9, 2 ∗ 19, 0 ∗ 12} 2.3535 1.0211 0.4702 19.9133 0.0836
{47, 0 ∗ 39} 1.6393 1.3048 0.0068 3.4906 0.0827
{0 ∗ 15, 1 ∗ 15, 0 ∗ 9, 32} 2.2893 0.9886 0.6055 24.6367 0.0769

investigate the optimal censoring schemes in terms of (a) Criteria 1: minimum trace of the variance–covariance matrix of
the MLEs (φ ) and (b) Criteria 2: minimizing the variance of the estimate of the pth quantile given by
 2
V (Q̂p ) = V11 (θ̂ ) + F(−p)1 V22 (θ̂ ) + 2F(−1−
1
p) V12 (θ̂ )

where Q̂p is the MLE of the pth quintile and F(−p)1 is the inverse CDF of the IW distribution defined as
−1/β
ln(p)

F(−p)1 = − .
λ
Illustrative example:
Using the real data set of Blischke and Murthy [21], we generate an adaptive progressively hybrid censored samples
considering different values of m, T and censoring schemes. The comparison results of different optimal censoring schemes
are presented in Tables 7–10 by obtaining φ and the variance of 5th quantiles V (Q̂0.05 ) and 95th quantiles V (Q̂0.95 ).
From Tables 7–10, it is noted that the censoring scheme number six has the minimum variance of the estimate of the pth
quantile, i.e. V (Q̂0.05 ) and V (Q̂0.95 ).
 M. Nassar, O.E. Abo-Kasem / Journal of Computational and Applied Mathematics 315 (2017) 228–239 239

7. Conclusion

In this paper, we discussed the estimation problem of the unknown parameters of the IW distribution based on adaptive
type-II progressively hybrid censored data. We used classical and Bayesian estimation methods to estimate the unknown
parameters. It is observed that the MLEs cannot be obtained in explicit forms, therefore, the AMLEs have been suggested. We
obtained the Bayes estimates based on SE loss function under the assumption of independent gamma priors using Lindley’s
approximation. A real data set is used to show how the scheme works in practice. The performance of the different estimators
is compared based on simulation study in terms of their MSEs. It is observed that the Bayes estimates with non-informative
priors work very well in terms of MSEs, also, the Bayes estimates using the informative prior have the smallest MSEs among
all other estimates proposed. Finally, we have proposed the optimal censoring scheme based on two different criteria and
an illustrative example is provided. As a future work, the inferential results of some life time models under adaptive type-II
progressively hybrid censored data with random removals have not been developed yet. Also, one sample prediction based
on AT-II PHCS remains open.

Acknowledgment

The authors thank the anonymous referee for a careful reading of the article.

References

[1] A.Z. Keller, A.R.R. Kamath, Alternative reliability models for mechanical systems, in: Proceeding of the Third International Conference on Reliability
and Maintainability, 1982, pp. 411–415.
[2] P. Erto, M. Rapone, Non-informative and practical Bayesian confidence bounds for reliable life in the Weibull model, Reliab. Eng. 7 (1984) 181–191.
[3] R. Calabria, G. Pulcini, Bayesian 2-sample prediction for the inverse Weibull distribution, Commun. Stat.-Theory Methods 23 (1994) 1811–1824.
[4] M. Maswadah, Conditional confidence interval estimation for the inverse Weibull distribution based on censored generalized order statistics, J. Stat.
Comput. Simul. 73 (12) (2003) 887–898.
[5] B.O. Oluyede, T. Yang, Generalizations of the inverse Weibull and related distributions with applications, Electron. J. Appl. Stat. Anal. 7 (2014) 94–116.
[6] D. Kundu, H. Howlader, Bayesian inference and prediction of the inverse Weibull distribution for type-II censored data, Comput. Statist. Data Anal. 54
(2010) 1547–1558.
[7] R. Musleh, A. Helu, Estimation of the inverse Weibull distribution based on progressively censored data: Comparative study, Reliab. Eng. Syst. Saf. 131
(2014) 216–227.
[8] K.S. Sultan, N.H. Alsadat, D. Kundu, Bayesian and maximum likelihood estimations of the inverse Weibull parameters under progressive type-II
censoring, J. Stat. Comput. Simul. 84 (10) (2014) 2248–2265.
[9] P. Xiuyun, Y. Zaizai, Bayesian estimation and prediction for the inverse Weibull distribution under general progressive censoring, Commun. Stat.-
Theory Methods 45 (2016) 624–635.
[10] D. Kundu, A. Joarder, Analysis of type-II progressively hybrid censored data, Comput. Statist. Data Anal. 50 (2006) 2509–2528.
[11] N. Balakrishnan, D. Kundu, Hybrid censoring: models, inferential results and applications, Comput. Statist. Data Anal. 57 (1) (2013) 166–209.
[12] H.K.T. Ng, D. Kundu, P.S. Chan, Statistical analysis of exponential lifetimes under an adaptive Type-II progressively censoring scheme, Nav. Res. Logist.
56 (2009) 687–698.
[13] C.T. Lin, H.K.T. Ng, P.S. Chan, Statistical inference of type-II progressively hybrid censored data with Weibull lifetimes, Commun. Stat.-Theory Methods
38 (2009) 1710–1729.
[14] F. Hemmati, E. Khorram, Statistical analysis of the log-normal distribution under type- II progressive hybrid censoring schemes, Commun. Stat.-Simul.
Comput. 42 (2013) 52–75.
[15] M.A.W. Mahmoud, A.A. Soliman, A.H. Abd Ellah, R.M. El-Sagheer, Estimation of generalized Pareto under an adaptive type-II progressive censoring,
Intell. Inf. Manag. 5 (2013) 73–83.
[16] S.K. Ashour, M.M.A. Nassar, Analysis of generalized exponential distribution under adaptive Type-II progressive hybrid censored competing risks data,
Int. J. Adv. Stat. Probab. 2 (2014) 108–113.
[17] A.A. Ismail, Inference for a step-stress partially accelerated life test model with an adaptive Type-II progressively hybrid censored data from Weibull
distribution, J. Comput. Appl. Math. 260 (2014) 553–542.
[18] M.M. AL Sobhi, A.A. Soliman, Estimation for the exponentiated Weibull model with adaptive Type-II progressive censored schemes, Appl. Math. Model.
40 (2) (2015) 1180–1192.
[19] F. Hemmati, E. Khorram, On adaptive progressively type-II censored competing risks data, Commun. Stat.-Simul. Comput. (2015) http://dx.doi.org/
10.1080/03610918.2015.1129408, in press.
[20] D.V. Lindley, Approximate Bayesian methods, Trab. Estadistica. 31 (1980) 223–237.
[21] W.R. Blischke, D.N.P. Murthy, Reliability: Modeling, Prediction, and Optimization, Wiley, New York, 2000.