Estimating the parameters of the Weibull distribution under generalized type II hybrid censoring
Kyeongjun Lee 1
1 Division of Mathematics and Big Data Science, Daegu University
Received 30 June 2021, revised 12 July 2021, accepted 14 July 2021
Abstract
In hybrid censoring both the time and the number of failures are considered for the life testing of the product. The combination of type I and type II censoring is called the hybrid censoring. Though the type II hybrid censored scheme guarantees a pre-fixed number of failures, it might take a long time to complete the test. In order to provide a guarantee in terms of the time to complete the test, generalized type II hybrid censoring scheme was introduced. In this paper, we consider the MLEs of the parameters and reliability when the data are generalized type II hybrid censored Weibull data. However, the MLEs cannot be obtained in a closed form. We use the approximate MLEs using Taylor series expansion. Also, we consider the Bayes estimation for the parameters and reliability when the data are generalized type II hybrid censored Weibull data. In Bayes estimation, Lindley’s approximation is used to obtain the Bayes estimators. Simulation experiments are performed to see the effectiveness of the different estimators. Finally, a real data set has been analysed for illustrative purposes.
Keywords: Bayes estimation, generalized type II hybrid censoring, Linex loss, squared error loss, Weibull distribution.
1. Introduction
Consider a life testing experiment in which n units are put on test. Assume that the life times of n units are independent and identically distributed (i.i.d) as Weibull distribution with the cumulative density function (cdf)
F (x; α, λ) = 1 − exp (−λx α ) , x > 0, α > 0, λ > 0, (1.1) the probability density function (pdf)
f (x; α, λ) = αλx α−1 exp (−λx α ) , x > 0, α > 0, λ > 0 (1.2) and the reliability function
R(t; α, λ) = exp (−λt α ) . (1.3)
1
Assistant professor, Division of Mathematics and Big Data Science, Daegu University, Gyeongsan
38453, Korea. E-mail: indra [email protected]
The estimation for the parameters of Weibull distribution based on the censored sample has been investigated by many authors such as Almetwally et al. (2020), Lee et al. (2020a, 2020b), Okasha and Mustafa (2020), Wang (2018), Wang et al. (2020). Almetwally et al.
(2020) considered a maximum product spacing estimation of Weibull distribution under adaptive type-II progressive censoring. Lee et al. (2020a, 2020b) considered Bayes estima- tion of Weibull distribution based on generalized adaptive progressive hybrid censoring.
Okasha and Mustafa (2020) considered E-Bayesian estimation for the Weibull distribution under adaptive type I progressive hybrid censored competing risks data. Wang (2018) and Wang et al. (2020) considered inference for Weibull competing risks data under generalized progressive hybrid censoring.
Hybrid censoring scheme is commonly used in the life testing situations which was intro- duced by Epstein (1954) for the exponential distribution as the life time distribution (Aslam et al., 2017). In hybrid censoring both the time and the number of failures are considered for the life testing of the product (Epstein, 1954). There are situations in which only the time is fixed for the life testing, which is known as the Type-I censoring. When the number of observed failures is fixed in the life testing, it is called the Type-II censoring. The combina- tion of these two censoring schemes is called the hybrid censoring (Balakrishnan and Kundu, 2013). Following Childs et al. (2003) introduced a type I hybrid censoring scheme and type II hybrid censoring scheme. In type II hybrid censoring scheme, the termination point is max {X r:n , T}. Here, the ith ordered life times of items be denoted by X i:n , r ∈ {1, 2, · · · , n}
and T ∈ (0, ∞) are pre-fixed. Though the type II hybrid censored scheme guarantees a pre- fixed number of failures, it might take a long time to observe r failures. In order to provide a guarantee in terms of the time to complete the test, Chandrasekar et al. (2004) introduced generalized type II hybrid censoring scheme. The detail description of the generalized type II hybrid censoring scheme is described as follows. If the rth failure occurs before time T 1 , terminate the experiment at T 1 ; if the rth failure occurs between T 1 and T 2 , terminate at X r:n ; if the rth failures occur after T 2 , terminate at T 2 (see Chandrasekar et al., 2004). Let d 1 and d 2 denote the number of observed failures up to time T 1 and T 2 , respectively.
This study is to consider the MLEs of the parameters and reliability when the data are generalized type II hybrid censored data. However, the MLEs cannot be obtained in a closed form. We use the approximate MLEs using Taylor series expansion. Also, we consider the Bayes estimation for the parameters and reliability when the data are generalized type II hybrid censored data.
2. Maximum likelihood estimation
Assume that the failure times of the units are the Weibull distribution with cdf (1.1) and pdf (1.2). The likelihood functions for three different cases are as follows.
Case 1 : L 1 (α, λ) = n!
(n − d 1 )! α d
1λ d
1d
1Y
i=1
x α−1 i:n exp
"
−λ
d
1X
i=1
x α i:n + (n − d 1 )T 1 α
!#
,
Case 2 : L 2 (α, λ) = n!
(n − r)! α r λ r
r
Y
i=1
x α−1 i:n exp
"
−λ
r
X
i=1
x α i:n + (n − r)x α r:n
!#
,
Case 3 : L 3 (α, λ) = n!
(n − d 2 )! α d
1λ d
1d
1Y
i=1
x α−1 i:n exp
"
−λ
d
1X
i=1
x α i:n + (n − d 2 )T 2 α
!#
.
Here, Cases 1, 2, and 3 can be combined, and be represented as
L (α, λ) = n!
(n − d)! α d λ d
d
Y
i=1
x α−1 i:n exp
"
−λ
d
X
i=1
x α i:n + (n − d)T α
!#
. (2.1)
Here, d means the number of failures. d = d 1 if x r:n < T 1 , D = r if T 1 < x r:n < T 2 , and d = d 2 if T 2 < x r:n . T means the terminated time. T = T 1 if x r:n < T 1 , T = x r:n if T 1 < x r:n < T 2 , and T = T 2 if T 2 < x r:n . Ignoring the constant, from (2.1), corresponding log likelihood functions are
lnL (α, λ) = dlnαλ + (α − 1)
d
X
i=1
lnx i:n − λ
d
X
i=1
x α i:n − λ (n − d) T α .
On differentiating the log-likelihood functions with respect to α and λ and equating to zero, we obtain the estimating equations:
∂lnL
∂α = d α +
d
X
i=1
lnx i:n − λ
d
X
i=1
x α i:n lnx i:n − λ (n − d) T α lnT = 0, (2.2)
∂lnL
∂λ = d λ −
d
X
i=1
x α i:n − (n − d) T α = 0.
Then, we get the MLE of λ, say ˆ λ(α), as a function of the MLE of λ as
λ(α) = ˆ d
P d
i=1 x α i:n + (n − d) T α (2.3)
and putting the value of ˆ λ(α) into equation (2.2), we obtain d
α +
d
X
i=1
lnx i:n − ˆ λ(α)
" d X
i=1
x α i:n lnx i:n + (n − d) T α lnT
#
= 0. (2.4)
Then, we obtain
k(α) = α, (2.5)
where k(α) = d/{ˆ λ(α) h P d
i=1 x α i:n lnx i:n + (n − d) T α lnT i
− P d
i=1 lnx i:n }.
From (2.5), we introduce a simple iterative scheme to solve for α. It has been proposed in the literature by Lee (2020a). Start with an initial guess of α, say α (0) , then obtain α (1) = k(α (0) ) and proceed in this way iteratively to obtain α (n+1) = k(α (n) ). Stop the iterative procedure, when |α (n+1) − α (n) | < , some pre-assigned tolerance limit. Once we obtain the MLE of α, say ˆ α, then the MLE of λ can be obtain as ˆ λ = ˆ λ( ˆ α) from (2.3).
With α and λ replaced by the ˆ α and ˆ λ, in Eq (1.3), the reliability estimator of the Weibull distribution based on generalized type II hybrid censoring scheme are obtained as:
R(t) = exp ˆ
−ˆ λt α ˆ
. (2.6)
2.1. Approximate MLE
Because the log-likelihood equations cannot be solved explicitly, it will be desirable to consider an approximation to the likelihood equations that will provide explicit estimators of α and λ. In this section, therefore, we provide the approximate MLE which have explicit form. First of all, let y i:n = lnx i:n . Also, let z i:n = (y i:n − µ)/σ and T z = (T − µ)/σ where σ = 1/α and µ = −1/αlnλ. Then, the Eq (2.1) can be obtained as
L(µ, σ) = n!
(n − d)!
1 σ d
d
Y
i=1
f (z i:n )[1 − F (T z )] n−d , (2.7)
where F (z) = 1−exp(−e z ), f (z) = exp(z −e z ). On differentiating the log-likelihood function with respect to µ and σ, we obtain the two estimating equations
∂lnL
∂µ =
d
X
i=1
f 0 (z i:n )
f (z i:n ) − (n − d) f (T z )
1 − F (T z ) = 0, (2.8)
∂lnL
∂σ = d +
d
X
i=1
f 0 (z i:n )
f (z i:n ) z i:n − (n − d) f (T z )
1 − F (T z ) T z = 0. (2.9) Because the Eqs (2.8) and (2.9) cannot be solved explicitly, we expand the function f 0 (z i:n )/f (z i:n ) and f (T z )/[1−F (T z )] in Taylor series expansion as follows. Let p i = i/(n+1) and q i = 1 − p i for i = 1, 2, · · · , n. Also, let p d
∗1
= (p d
1+ p d
1+1 )/2 and q d
∗1
= 1 − p d
∗1
for x r:n < T 1 . And, let p d
∗2
= (p d
2+ p d
2+1 )/2 and q d
∗2
= 1 − p d
∗2