2018, 29 ( 3), 807–814
Estimating the parameter of an exponential distribution under multiply type II hybrid censoring †
Wonhee Lee 1 · Kyeongjun Lee 2
1 Department of Statistics, Daegu University
2 Division of Mathematics and Big Data Science, Daegu University
Received 3 April 2018, revised 10 May 2018, accepted 12 May 2018
Abstract
In this paper, we propose a multiply type II hybrid censoring scheme. And, we derive statistical inference on the scale parameter for the exponential distribution under proposed multiply type II hybrid censoring scheme. The scale parameter is estimated by maximum likelihood estimator (MLE) and approximate MLE using Taylor series expansion. We also obtain the asymptotic confidence interval (CI) for scale parameter under multiply type II hybrid censoring scheme. We compare the MLE and approximate MLE in the terms of the mean square error (MSE) and bias. The simulation procedure is repeated 10,000 times for the sample size n = 20, 40 and various multiply type II hybrid censoring. Consequently, approximate MLE is generally more efficient than the MLE. The average CL of approximate MLE is longer than the corresponding average CL of MLE.
Keywords: Approximate maximum likelihood estimator, confidence interval, exponen- tial distribution, multiply type II hybrid censoring, Taylor series expansion.
1. Introduction
For a random variable X with probability density function (pdf) f (x) and cumulative density function (cdf) F (x), its exponential distribution has the cdf
F (x; θ) = 1 − exp
− x θ
, x > 0, θ > 0, (1.1) and pdf
f (x; θ) = 1 θ exp
− x θ
, x > 0, θ > 0. (1.2) In life-time test, there are many situation in which units are lost or removed from test before failure. The loss may occur unconsciously or carelessly. For example, unconsciously
† This research was supported by the Daegu University Research Grant, 2017.
1
Graduate Student, Department of Statistics, Daegu University, Gyeongsan 68453, Korea.
2
Corresponding author: Assistant professor, Division of Mathematics and Big Data Science, Daegu
University, Gyeongsan 68453, Korea. E-mail : indra [email protected]
II hybrid censoring scheme. Type II hybrid censoring scheme may arise in a situation when the tester determines that at least r failures must be observed, and has prepaid for the use of the testing facility for T units of time.
However, some units can be failed between two points of observation with exact times of failure of these units unobserved under type II hybrid censored sample. For example, loss may arise in life-testing experiments when the failure times of some units were not observed due to mechanical or experimental difficulties. Therefore, the new censoring method is needed.
This study has two aims. The first is to propose a multiply type II hybrid censoring scheme. The second aim is to consider the maximum likelihood estimator (MLE) of the θ when the data are multiply type II hybrid censored samples. However, MLE cannot be obtained in a closed form. We use the approximate MLE as an approximate estimator of θ.
The rest of the paper is organized as follows. In Section 2, we propose the multiply type II hybrid censoring scheme. Also, we derive MLE of the θ for exponential distribution under the multiply type II hybrid censored samples. Moreover, the asymptotic confidence interval (CI) for θ is presented. In Section 3, we derive some approximate MLE of the θ for the exponential distribution under the multiply type II hybrid censored samples. The θ is estimated by using Taylor series expansion method. Also, the asymptotic CI for θ is presented. In Section 4, the description of different estimators that are compared by performing the Monte Carlo simulation is presented. In Section 5, we conclude the paper.
2. Multiply type II hybrid censoring
Under the type II hybrid censoring scheme, suppose the tester fails to observe the middle observations. Then, for known and , we have the following two cases under multiply type II hybrid censoring scheme;
Case I: X a
1:n < X a
2:n < · · · < X a
d:n < T, if X a
d:n < T, (2.1) Case II: X a
1:n < X a
2:n < · · · < X a
d:n < T < X a
d+1:n < · · · < X a
r:n ,
if d < r and X a
d:n < T < X a
r:n , (2.2)
where X a
i:n denote a i -th observed failure time, and R i denote the number of censoring be-
tween X a
i:n and X a
i−1:n . A schematic representation of the multiply type II hybrid censoring
scheme is presented in Figure 2.1.
Figure 2.1 The multiply type II hybrid censoring scheme
Then, the likelihood functions based on (2.1) and (2.2) are as follows.
Case I: L ∝
d
Y
i=1
f (x a
i:n )
d−1
Y
i=1
F x a
i+1:n − F (x a
i:n ) R
i+1F (x a
1:n ) R
1[1 − F (x a
d:n )] n−a
d,
Case II: L ∝
r
Y
i=1
f (x a
i:n )
r−1
Y
i=1
F x a
i+1:n − F (x a
i:n ) R
i+1F (x a
1:n ) R
1[1 − F (x a
r:n )] n−a
r.
Therefore, case I and case II can be combined, and can be written as
L ∝
m
Y
i=1
f (x a
i:n )
m+1
Y
i=1
F (x a
i:n ) − F x a
i−1:n R
i, (2.3)
where m denotes the number of failures, F (x a
0:n ) = 0, F x a
m+1:n = 1, R i = a i − a i−1 − 1, a 0 = 0.
On differentiating the log-likelihood functions with respect to θ of (2.3) and equation to zero, we obtain the estimating equation as
∂lnL
∂θ = − 1 θ 2
"
mθ −
m
X
i=1
x a
i:n −
m+1
X
i=1
R i
x a
i−1:n e −
xai−1θ− x a
i:n e −
xaiθe −
xai−1θ− e −
xaiθ#
= 0. (2.4)
From (2.4), we get the MLE of as
θ = h (θ) , where
h (θ) = 1 m
" m X
i=1
x a
i:n +
m+1
X
i=1
R i
x a
i−1:n e −
xai−1θ− x a
i:n e −
xaiθe −
xai−1θ− e −
xaiθ#
. (2.5)
From (2.5), we propose a simple iterative scheme to solve for θ. This has been proposed
in the literature by Lee and Cho (2017). Start with an initial guess of θ (0) , then obtain
θ (1) = h θ (0) and proceed in this way iteratively to obtain θ (n) = h θ (n−1) . Stop the
∂θ 2 = − θ 3
i=1
R i
H 0 (θ) + θ 4
i=1
R i
H 0 (θ) 2 , (2.6)
where H j (θ) = x j a
i:n e −
xai:nθ− x j a
i−1:n e −
xai−1:nθ.
Let I (θ) denotes the Fisher information matrix of the θ. Then, the Fisher information matrix is obtained by taking expectations of minus Equation (2.6). Under some mild regu- larity conditions, ˆ θ is approximately normal with mean θ and variance I −1 (θ). In practice, we usually estimate I −1 (θ) by I −1 ˆ θ
. A simpler and equally valid procedure is to use the approximation ˆ θ ∼ N
θ, I −1 ˆ θ
. Therefore, 100(1 − α)% CI for θ is
θ ± Z ˆ α/2 r
I −1 ˆ θ ,
where Z α/2 is the percentile of the standard normal distribution with right-tail probability α/2.
3. Approximate maximum likelihood estimators
Let Z a
i:n= X a
i:n/θ and then the variables Z a
i:n have a standard exponential distribution with pdf f (z a
i:n ) = e −z
ai:n, and cdf F (z a
i:n ) = 1 − e −z
ai:n, z a
i:n > 0, we may rewrite the (2.3) as
L ∝ θ −m
m
Y
i=1
f (z a
i:n )
m+1
Y
i=1
F (z a
i:n ) − F z a
i−1:n R
i. (3.1)
On differentiating the log-likelihood functions with respect to and equation to zero, we obtain the estimating equation as
∂lnL
∂θ = − 1 θ [m −
m
X
i=1
z a
i:n + R 1
f (z a
1:n )
F (z a
1:n ) z a
1:n − (n − a m )z a
m:n
+
m−1
X
i=1
R i+1
f (z a
i+1:n )z a
i+1:n − f (z a
i:n )z a
i:n
F (z a
i+1:n ) − F (z a
i:n ) ]. (3.2)
Since (3.2) cannot be solved explicitly, it will be desirable to consider an approximation
to the likelihood equation which provide us with explicit estimators for θ. See for example
the work of Lee et al. (2011), Lee et al. (2013) and Gwag and Lee (2018). Let ξ a
i:n = F −1 (p a
i:n ) = −ln (1 − p a
i:n ), where p a
i:n = a i /n + 1, q a
i:n = 1 − p a
i:n , i = 1, · · · , m. Here, we expand the function F (z f (z
a1:n)
a1:n
) z a
1:n and f (z
ai:nF (z )z
ai:n−f (z
ai−1:n)z
ai−1:nai:n
)−F (z
ai−1:n) in Taylor series around the points ξ a
i:n . Then, we can approximate the function by;
f (z a
1:n )
F (z a
1:n ) z a
1:n ' α 1 + β 1 z a
1:n , (3.3) f (z a
i:n )z a
i:n − f (z z
i−1:n )z a
i−1:n
F (z a
i:n ) − F (z a
i−1:n ) ' α 2i + β 2i z a
i:n + γ 2i z a
i−1:n , (3.4) where
α 1 = q a
1:n ξ a 2
1