Research Journal of Applied Sciences, Engineering and Technology 10(1): 22-28,... ISSN: 2040-7459; e-ISSN: 2040-7467

advertisement
Research Journal of Applied Sciences, Engineering and Technology 10(1): 22-28, 2015
ISSN: 2040-7459; e-ISSN: 2040-7467
© Maxwell Scientific Organization, 2015
Submitted: ‎
November ‎10, ‎2014
Accepted: February ‎8, ‎2015
Published: May 10, 2015
Bayesian Estimation of a Mixture Model
Ilhem Merah and Assia Chadli
Probability and Statistics Laboratory, Badji Mokhtar University, BP12, Annaba 23000, Algeria
Abstract: We present the properties of a bathtub curve reliability model having both a sufficient adaptability and a
minimal number of parameters introduced by Idée and Pierrat (2010). This one is a mixture of a Gamma distribution
G(2, (1/θ)) and a new distribution L(θ). We are interesting by Bayesian estimation of the parameters and survival
function of this model with a squared-error loss function and non-informative prior using the approximations of
Lindley (1980) and Tierney and Kadane (1986). Using a statistical sample of 60 failure data relative to a technical
device, we illustrate the results derived. Based on a simulation study, comparisons are made between these two
methods and the maximum likelihood method of this two parameters model.
Keywords: Approximations, bayes estimator, mixture distribution, reliability, weibull distribution
INTRODUCTION
Pierrat (2010). These authors showed that this model of
two parameters is equivalent to a mixture of two
Weibull distributions. Equivalence was shown using
Kolmogorov Smirnov distance. We propose to use a
Bayesian approach, with a non-informative prior and a
squared-error loss function.
The main objective of this study is to estimate the
two parameters and the survival function of the new
model. The methods under consideration are:
Maximum likelihood Estimation, Bayesian Methods
with two types of approximation: Lindley's
approximation and Tierney-Kadane's approximation.
The methods are compared using a real data analysis
and simulation. Conclusions based on the study are
given.
Mixture models play a vital role in many
applications. For example, direct applications of finite
mixture models are in economics, botany, medicine,
psychology, agriculture, zoology, life testing and
reliability engineering. However, a modelling by a
finished mixture of Weibull distributions can be more
relevant and physically significant. Several authors are
interesting by Bayesian estimation in mixture models
like Tsionas (2002) and Gelman et al. (2003).
Modelling by a Weibull mixture distributions was often
occulted by the survival analysis experts to the profile
of a simple model of Weibull, because of the high
number of parameters to be estimated and absence of
an effective estimation method of the parameters.
Several methods are used for estimating the
parameters of the mixture. The graphic approach was
used by Jiang and Murty (1995) and Jiang and
Kececioglu (1992) initially to adjust data with a
mixture of two Weibull distributions, then to estimate
its parameters. The two methods of the moments and
the maximum likelihood were largely used while being
based on the use of EM algorithm (ExpectationMaximization), this algorithm is not only one digital
technique, it gives also useful statistical data, for more
detail to see Dempster et al. (1977). A modification of
this algorithm was introduced by Wei and Tanner
(1990), where in the step E, the expectation is
calculated by Monte Carlo simulations, this algorithm
noted MCEM was used by Levine and Casella (2001)
and more recently by Elmahdy and Aboutahoun (2013).
The model in which, we are interested is a mixture
model of two distributions introduced by Idée and
PROPERTIES OF THE MODEL
The model Probability Density Function (PDF) is
given as:
π‘₯π‘₯
π‘₯π‘₯
π‘₯π‘₯ 𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½−οΏ½ οΏ½οΏ½
π‘₯π‘₯
πœƒπœƒ
π‘₯π‘₯
πœƒπœƒ
οΏ½−1+ οΏ½ ln οΏ½ �𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½−οΏ½ οΏ½οΏ½
πœƒπœƒ
πœƒπœƒ
+ πœ€πœ€
�𝑓𝑓(π‘₯π‘₯, πœƒπœƒ, πœ€πœ€) = (1 − πœ€πœ€) πœƒπœƒ
πœƒπœƒ
π‘₯π‘₯ > 0, πœƒπœƒ > 0, 0 ≤ πœ€πœ€ ≤ 1
πœƒπœƒ
(1)
The survival function is:
οΏ½
𝑑𝑑
𝑑𝑑
𝑑𝑑
𝑑𝑑
𝑅𝑅(𝑑𝑑, πœƒπœƒ) = οΏ½1 + (1 − πœ€πœ€) οΏ½ οΏ½ + πœ€πœ€ οΏ½ οΏ½ ln οΏ½ οΏ½οΏ½ 𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½− οΏ½ οΏ½οΏ½
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
𝑑𝑑 > 0, πœƒπœƒ > 0, 0 ≤ πœ€πœ€ ≤ 1
(2)
The failure rate is also given as:
Corresponding Author: Ilhem Merah, Probability and Statistics Laboratory, Badji Mokhtar University, BP12, Annaba
23000, Algeria
22
Res. J. Appl. Sci. Eng. Technol., 10(1): 22-28, 2015
𝑑𝑑
1+πœ€πœ€ ln οΏ½ οΏ½
1
πœƒπœƒ
β„Ž(𝑑𝑑, πœ€πœ€, πœƒπœƒ) = οΏ½1 −
𝑑𝑑
𝑑𝑑
𝑑𝑑 οΏ½
πœƒπœƒ
οΏ½
1+(1−πœ€πœ€)οΏ½ οΏ½+πœ€πœ€οΏ½ οΏ½ ln οΏ½ οΏ½
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
𝑑𝑑 > 0, πœƒπœƒ > 0, 0 ≤ πœ€πœ€ ≤ 1
(3)
𝑙𝑙(πœƒπœƒ, πœ€πœ€/π‘’π‘’π‘’π‘’β„Ž) =
πœƒπœƒ
πœƒπœƒ 𝑛𝑛
π‘₯π‘₯ 𝑖𝑖
π‘₯π‘₯ 𝑖𝑖
∏𝑛𝑛𝑖𝑖=1
π‘₯π‘₯ 𝑖𝑖
οΏ½(1 − πœ€πœ€) οΏ½ οΏ½ + πœ€πœ€ οΏ½ − 1οΏ½ ln οΏ½ οΏ½οΏ½
πœƒπœƒ
We pose:
πœƒπœƒ
πœ‹πœ‹(πœ€πœ€, πœƒπœƒ) =
𝑆𝑆
𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½−οΏ½ οΏ½οΏ½
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
πœ€πœ€Μ‚ =
Then:
𝑙𝑙(πœƒπœƒ, πœ€πœ€/π‘’π‘’π‘’π‘’β„Ž) =
𝑆𝑆
𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½− ∑𝑛𝑛𝑖𝑖=1 οΏ½
πœƒπœƒ
πœƒπœƒ 𝑛𝑛
The log- likelihood is:
∏𝑛𝑛𝑖𝑖=1�𝑒𝑒𝑖𝑖 (πœ€πœ€, πœƒπœƒ)οΏ½
𝑆𝑆
𝐿𝐿(πœƒπœƒ, πœ€πœ€/π‘’π‘’π‘’π‘’β„Ž) = −𝑛𝑛 ln(πœƒπœƒ) − + ∑𝑛𝑛𝑖𝑖=1 ln(𝑒𝑒𝑖𝑖 )
πœƒπœƒ
πœ•πœ•πœ•πœ•
πœ•πœ•πœ•πœ•
πœ•πœ•πœ•πœ•
𝑆𝑆
= − 𝑛𝑛 + ∑𝑛𝑛𝑖𝑖=1 οΏ½
πœƒπœƒ
=
−π‘₯π‘₯
∑𝑛𝑛𝑖𝑖=1 οΏ½ 𝑖𝑖
πœƒπœƒ
−π‘₯π‘₯ 𝑖𝑖 +πœ€πœ€πœ€πœ€
π‘₯π‘₯ 𝑖𝑖
πœƒπœƒ
+οΏ½ −
πœƒπœƒ
π‘₯π‘₯
From where, we obtain this system:
𝑆𝑆
πœƒπœƒ
οΏ½
− 𝑛𝑛 + ∑𝑛𝑛𝑖𝑖=1 οΏ½
∑𝑛𝑛𝑖𝑖=1 οΏ½
−π‘₯π‘₯ 𝑖𝑖
πœƒπœƒ
−π‘₯π‘₯ 𝑖𝑖 +πœ€πœ€πœ€πœ€
πœƒπœƒ
π‘₯π‘₯
π‘₯π‘₯
=0
π‘₯π‘₯
∞ 1
πœƒπœƒ
𝑆𝑆
∞ 1 𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½−οΏ½πœƒπœƒ οΏ½οΏ½ 𝑛𝑛
∏𝑖𝑖=1 �𝑒𝑒 𝑖𝑖 (πœ€πœ€,πœƒπœƒ)�𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
∫0
πœ€πœ€πœƒπœƒ 𝑛𝑛
∞ 1
∫0 ∫0 𝑙𝑙(πœƒπœƒ,πœ€πœ€/π‘’π‘’π‘’π‘’β„Ž)πœ‹πœ‹(πœ€πœ€,πœƒπœƒ)𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
𝑑𝑑
𝑑𝑑
𝑆𝑆
𝑑𝑑 𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½−οΏ½πœƒπœƒ οΏ½οΏ½ 𝑛𝑛
∏𝑖𝑖=1 �𝑒𝑒 𝑖𝑖 (πœ€πœ€,πœƒπœƒ)�𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
πœ€πœ€πœƒπœƒ 𝑛𝑛 +1
𝑆𝑆
∞ 1 𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½−οΏ½πœƒπœƒ οΏ½οΏ½ 𝑛𝑛
∏𝑖𝑖=1�𝑒𝑒 𝑖𝑖 (πœ€πœ€,πœƒπœƒ)�𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
πœ€πœ€πœƒπœƒ 𝑛𝑛 +1
To solve this ratio of integrals explicitly, we will
use approximation methods.
(7)
Lindley’s procedure: Lindley (1980) developed
approximate procedures for the evaluation of the ratio
of integrals of the form:
π‘₯π‘₯
πœƒπœƒ
∫ 𝑀𝑀(πœƒπœƒ)𝑒𝑒𝑒𝑒𝑒𝑒 {𝐿𝐿(πœƒπœƒ)}𝑑𝑑𝑑𝑑
∫ 𝑣𝑣(πœƒπœƒ)𝑒𝑒𝑒𝑒𝑒𝑒 {𝐿𝐿(πœƒπœƒ)}𝑑𝑑𝑑𝑑
(9)
In our case, π‘šπ‘š = 2 and πœ†πœ† = (πœ€πœ€, πœƒπœƒ)
𝜌𝜌(πœ€πœ€, πœƒπœƒ) = log{πœ‹πœ‹(πœ€πœ€, πœƒπœƒ)}
ʌ(πœ€πœ€, πœƒπœƒ) = log{πœ‹πœ‹(πœ€πœ€, πœƒπœƒ|π‘₯π‘₯)} = 𝐿𝐿(πœ€πœ€, πœƒπœƒ) + 𝜌𝜌(πœ€πœ€, πœƒπœƒ) =
𝑆𝑆
= − ln(πœ€πœ€πœ€πœ€) − 𝑛𝑛 ln(πœƒπœƒ) − + ∑𝑛𝑛𝑖𝑖=1 ln�𝑒𝑒𝑖𝑖 (πœ€πœ€, πœƒπœƒ)οΏ½
πœƒπœƒ
(10)
Denoting 𝛷𝛷(𝑑𝑑) = 𝑅𝑅𝑑𝑑 and using Lindley's method,
expanding about the posterior mode, the Bayesian
estimator of 𝑅𝑅𝑑𝑑 in Eq. (2) becomes:
The resolution of the system is done using the
iterative methods and provides us estimators of the
maximum of likelihood of πœƒπœƒ and πœ€πœ€ noted respectively
by πœƒπœƒπ‘€π‘€π‘€π‘€ and πœ€πœ€π‘€π‘€π‘€π‘€ . To have an estimator of the survival
function, it is enough to replace (πœ€πœ€, πœƒπœƒ) in the expression
(2) by (πœƒπœƒπ‘€π‘€π‘€π‘€ , πœ€πœ€π‘€π‘€π‘€π‘€ ), we obtain an estimator of the
survival function noted 𝑅𝑅𝑀𝑀𝑀𝑀 (𝑑𝑑).
Bayesian estimation: Given a random sample,
(π‘₯π‘₯1 , π‘₯π‘₯2 , … , π‘₯π‘₯𝑛𝑛 ) of size n, from a mixture distribution
and ⨱ the vector of the observed data. The likelihood
function can be constructed as follows:
𝑑𝑑
∫0 ∫0
+ οΏ½ 𝑖𝑖 − 1οΏ½ ln οΏ½ 𝑖𝑖 οΏ½οΏ½ 𝑒𝑒𝑖𝑖−1 = 0
πœƒπœƒ
𝑆𝑆
∞ 1 𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½−οΏ½πœƒπœƒ οΏ½οΏ½ 𝑛𝑛
∏𝑖𝑖=1�𝑒𝑒 𝑖𝑖 (πœ€πœ€,πœƒπœƒ)�𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
𝑛𝑛
+1
πœƒπœƒ
∞ 1
πœ€πœ€
∫0 ∫0 π‘™π‘™οΏ½πœƒπœƒ,𝑒𝑒𝑒𝑒 β„Ž οΏ½πœ‹πœ‹(πœ€πœ€,πœƒπœƒ)𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
∫0 ∫0 οΏ½1+(1−πœ€πœ€)πœƒπœƒ +πœ€πœ€ πœƒπœƒ ln οΏ½πœƒπœƒ ��𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½−πœƒπœƒ οΏ½
− πœ€πœ€ οΏ½ 𝑖𝑖 οΏ½ ln οΏ½ 𝑖𝑖 οΏ½οΏ½ 𝑒𝑒𝑖𝑖−1 = 0,
πœƒπœƒ
(8)
∫0 ∫0
𝑅𝑅𝑑𝑑∗ = 𝐸𝐸(𝑅𝑅𝑑𝑑 |π‘₯π‘₯ )
π‘₯π‘₯
1οΏ½ ln οΏ½ οΏ½οΏ½ 𝑒𝑒𝑖𝑖−1
πœƒπœƒ
∏𝑛𝑛𝑖𝑖=1�𝑒𝑒𝑖𝑖 (πœ€πœ€, πœƒπœƒ)οΏ½
∝
The Bayesian estimator of 𝑅𝑅𝑑𝑑 under the loss
function is the posterior mean:
(5)
− πœ€πœ€ οΏ½ 𝑖𝑖 οΏ½ ln οΏ½ 𝑖𝑖 οΏ½οΏ½ 𝑒𝑒𝑖𝑖−1 = 0
πœƒπœƒ
πœƒπœƒ
(6)
π‘₯π‘₯ 𝑖𝑖
𝑙𝑙(πœƒπœƒ,πœ€πœ€/π‘’π‘’π‘’π‘’β„Ž)πœ‹πœ‹(πœ€πœ€,πœƒπœƒ)
∫
πœƒπœƒοΏ½ = 0
Differentiating (5) with respect to πœ€πœ€ and πœƒπœƒ and
equating to zero, we have:
πœ•πœ•πœ•πœ•
∞ 1
∫0 ∫0 𝑙𝑙(πœƒπœƒ,πœ€πœ€/π‘’π‘’π‘’π‘’β„Ž)πœ‹πœ‹(πœ€πœ€,πœƒπœƒ)𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑
The Bayesian estimates of parameters under the
loss function are:
π‘₯π‘₯ 𝑖𝑖
𝑒𝑒𝑖𝑖 (πœ€πœ€, πœƒπœƒ) = οΏ½(1 − πœ€πœ€) οΏ½ οΏ½ + πœ€πœ€ οΏ½ − 1οΏ½ ln οΏ½ οΏ½οΏ½ ;
𝑆𝑆 = ∑𝑛𝑛𝑖𝑖=1 π‘₯π‘₯𝑖𝑖
1
πœ€πœ€πœ€πœ€
πœ‹πœ‹(πœ€πœ€, πœƒπœƒ|⨱) =
πœ€πœ€πœƒπœƒ 𝑛𝑛 +1
π‘₯π‘₯ 𝑖𝑖
∏𝑛𝑛𝑖𝑖=1�𝑒𝑒𝑖𝑖 (πœ€πœ€, πœƒπœƒ)οΏ½
The posterior distribution is given by:
(4)
πœƒπœƒ
π‘₯π‘₯ 𝑖𝑖
πœƒπœƒ 𝑛𝑛
Using the non-informative prior:
Maximum likelihood: We lay out of a sample of n
independent and complete observations of the same
distribution given in (1) noted (π‘₯π‘₯1 , π‘₯π‘₯2 , … , π‘₯π‘₯𝑛𝑛 ). The
likelihood of the sample is written:
π‘₯π‘₯
𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½− ∑𝑛𝑛𝑖𝑖=1οΏ½ 𝑖𝑖 οΏ½οΏ½
𝑆𝑆
𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½−οΏ½ οΏ½οΏ½
πœƒπœƒ
𝑙𝑙(πœƒπœƒ, πœ€πœ€/π‘’π‘’π‘’π‘’β„Ž) =
23
1
1
𝑅𝑅𝑑𝑑∗ = π›·π›·οΏ½πœ†πœ†Μ‚οΏ½ + ∑ 𝛷𝛷𝑖𝑖𝑖𝑖 οΏ½πœ†πœ†Μ‚οΏ½πœπœπ‘–π‘–π‘–π‘– + ∑ ΚŒπ‘–π‘–π‘–π‘–π‘–π‘– 𝛷𝛷𝑙𝑙 οΏ½πœ†πœ†Μ‚οΏ½πœπœπ‘–π‘–π‘–π‘– πœπœπ‘˜π‘˜π‘˜π‘˜ (11)
2
2
Res. J. Appl. Sci. Eng. Technol., 10(1): 22-28, 2015
where,
πœ•πœ• 2 𝛷𝛷
𝛷𝛷𝑖𝑖𝑖𝑖 =
πœ•πœ•πœ†πœ† 𝑖𝑖 πœ•πœ•πœ†πœ† 𝑗𝑗
, ΚŒπ‘–π‘–π‘–π‘–π‘–π‘– =
πœ•πœ• 3 ʌ
πœ•πœ•πœ†πœ† 𝑖𝑖 πœ•πœ•πœ†πœ† 𝑗𝑗 πœ•πœ•πœ†πœ† π‘˜π‘˜
, 𝑒𝑒𝑒𝑒𝑒𝑒
ʌ222 =
(12)
The πœπœπ‘–π‘–π‘–π‘– values are defined as the (𝑖𝑖, 𝑗𝑗)th element of
the negative of the inverse of the Hessian matrix of
−1
of ʌ: οΏ½πœπœπ‘–π‘–π‘–π‘– οΏ½ = −οΏ½ΚŒπ‘–π‘–π‘–π‘– οΏ½ .
signe second derivatives
Expanding Eq. (7), we obtain:
1
𝑅𝑅𝐡𝐡∗ (𝑑𝑑) = π›·π›·οΏ½πœ†πœ†Μ‚οΏ½ + 𝛷𝛷11 οΏ½πœ†πœ†Μ‚οΏ½πœπœ11 + 𝛷𝛷12 οΏ½πœ†πœ†Μ‚οΏ½πœπœ12 +
1
2
1
2
1
2
𝛷𝛷22 οΏ½πœ†πœ†Μ‚οΏ½πœπœ22 + ʌ111 �𝛷𝛷1 οΏ½πœ†πœ†Μ‚οΏ½πœπœ11
+ 𝛷𝛷2 οΏ½πœ†πœ†Μ‚οΏ½πœπœ11 𝜏𝜏12 οΏ½ +
2
ʌ112 =
2 )οΏ½
ʌ οΏ½3𝛷𝛷1 οΏ½πœ†πœ†Μ‚οΏ½πœπœ11 𝜏𝜏12 + 𝛷𝛷2 οΏ½πœ†πœ†Μ‚οΏ½(𝜏𝜏11 𝜏𝜏22 + 2𝜏𝜏12
+
2 112
1
2
1
2
2 )οΏ½
ʌ122 οΏ½3𝛷𝛷2 οΏ½πœ†πœ†Μ‚οΏ½πœπœ22 𝜏𝜏12 + 𝛷𝛷1 οΏ½πœ†πœ†Μ‚οΏ½(𝜏𝜏11 𝜏𝜏22 + 2𝜏𝜏12
+
2
ʌ222 �𝛷𝛷2 οΏ½πœ†πœ†Μ‚οΏ½πœπœ22
+ 𝛷𝛷1 οΏ½πœ†πœ†Μ‚οΏ½πœπœ22 𝜏𝜏12 οΏ½
(13)
ʌ122
In our case, with π‘šπ‘š = 2 and πœ†πœ† = (πœ€πœ€, πœƒπœƒ), we obtain
from Eq. (2), with 𝛷𝛷(𝑑𝑑) = 𝑅𝑅𝑑𝑑 :
𝛷𝛷1 =
πœ•πœ•π›·π›·
πœ•πœ•πœ•πœ•
𝛷𝛷12 =
πœ•πœ•π›·π›·
πœ•πœ•πœ•πœ•
=οΏ½
𝑑𝑑
𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 οΏ½−οΏ½ οΏ½οΏ½
πœƒπœƒ
=
πœ•πœ• 2 𝛷𝛷
=
πœ•πœ•πœ•πœ•πœ•πœ•πœ•πœ•
𝑑𝑑𝑒𝑒 𝑑𝑑
𝛷𝛷22 =
πœƒπœƒ 2
οΏ½ 𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½− οΏ½ οΏ½οΏ½,
πœ•πœ• 2 𝛷𝛷
πœ•πœ•πœƒπœƒ 2
=
πœƒπœƒ
𝑑𝑑
πœƒπœƒ
πœƒπœƒ 2
𝑑𝑑
𝑑𝑑
οΏ½−1 + ln οΏ½ οΏ½οΏ½ , 𝛷𝛷11 =
πœƒπœƒ
𝑑𝑑
𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 οΏ½−οΏ½ οΏ½οΏ½
πœƒπœƒ
𝑑𝑑
𝑑𝑑 2 𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½−οΏ½ οΏ½οΏ½
𝑑𝑑
πœ•πœ• 2 𝛷𝛷
𝑑𝑑
πœ•πœ•πœ€πœ€ 2
= 0,
οΏ½− + οΏ½ − 1οΏ½ ln οΏ½ οΏ½οΏ½ , 𝛷𝛷2 =
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
πœ€πœ€
οΏ½
𝑆𝑆
−(1 + 𝑛𝑛) + + ∑𝑛𝑛𝑖𝑖=1 𝑣𝑣𝑖𝑖 𝑒𝑒𝑖𝑖−1 = 0,
πœƒπœƒ
The partial derivatives of the log-posterior density
function are:
ʌ22 =
�𝑒𝑒𝑖𝑖−1
ʌ12 =
πœ•πœ• 2 ʌ
πœ•πœ•πœƒπœƒ 2
=
1
πœ€πœ€ 2
𝑙𝑙 =
So that:
+ ∑𝑛𝑛𝑖𝑖=1(−𝑠𝑠𝑖𝑖2 𝑒𝑒𝑖𝑖−2 ),
1+𝑛𝑛
πœƒπœƒ 2
−
2𝑆𝑆
πœƒπœƒ
3 +
∑𝑛𝑛𝑖𝑖=1
2
πœ•πœ•πœ€πœ€ 3
=−
2
πœ€πœ€ 3
+ 2 ∑𝑛𝑛𝑖𝑖=1 𝑠𝑠𝑖𝑖3 𝑒𝑒𝑖𝑖−3 ,
πœƒπœƒ
log 𝑣𝑣(πœ†πœ†)+𝐿𝐿(πœ†πœ†|π‘₯π‘₯ )
𝑛𝑛
, 𝑙𝑙 ∗ =
πœƒπœƒ
π‘₯π‘₯ 𝑖𝑖
πœƒπœƒ
+
log 𝛷𝛷(πœ†πœ†)+log 𝑣𝑣(πœ†πœ†)+𝐿𝐿(πœ†πœ†|π‘₯π‘₯ )
𝑛𝑛
∫ 𝑒𝑒𝑒𝑒𝑒𝑒 (𝑛𝑛𝑙𝑙 ∗ )𝑑𝑑𝑑𝑑
.
∫ 𝑒𝑒𝑒𝑒𝑒𝑒 (𝑛𝑛𝑛𝑛 )𝑑𝑑𝑑𝑑
∗
𝑛𝑛
πœ•πœ• 3 ʌ
π‘₯π‘₯ 𝑖𝑖
|𝛯𝛯 |
𝐸𝐸� (𝛷𝛷(πœ†πœ†)|π‘₯π‘₯ ) = οΏ½ |𝛯𝛯| οΏ½
πœ•πœ• 2 ʌ
1
π‘₯π‘₯𝑖𝑖
π‘₯π‘₯𝑖𝑖
= οΏ½ �𝑒𝑒𝑖𝑖−1 οΏ½1 − ln οΏ½ οΏ½οΏ½ − 𝑣𝑣𝑖𝑖 𝑠𝑠𝑖𝑖 𝑒𝑒𝑖𝑖−2 οΏ½,
πœƒπœƒ
πœƒπœƒ
πœ•πœ•πœ•πœ•πœ•πœ•πœ•πœ•
πœƒπœƒ
ʌ111 =
π‘₯π‘₯
πœƒπœƒ
, (14)
(15)
The Bayesian estimator of 𝛷𝛷(πœ†πœ†) takes the form:
2πœ€πœ€π‘₯π‘₯𝑖𝑖
π‘₯π‘₯𝑖𝑖
2π‘₯π‘₯𝑖𝑖 + πœ€πœ€π‘₯π‘₯𝑖𝑖 − πœ€πœ€πœ€πœ€
οΏ½
ln οΏ½ οΏ½ +
οΏ½ − 𝑣𝑣𝑖𝑖2 𝑒𝑒𝑖𝑖−2 οΏ½
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
𝑖𝑖=1
π‘₯π‘₯
οΏ½ − πœ€πœ€ οΏ½ 𝑖𝑖 οΏ½ ln οΏ½ 𝑖𝑖 οΏ½;𝑠𝑠𝑖𝑖 = −
𝐸𝐸(𝛷𝛷(πœ†πœ†)|π‘₯π‘₯ ) =
1
πœƒπœƒ
πœƒπœƒ
Tierney
and
Kadane’s
method:
Lindley's
approximation requires the evaluation of the third
derivatives of the likelihood function or the posterior
density which can be very tedious and requires great
computational precision. Tierney and Kadane (1986)
gave an alternative method of evaluation of the ratio of
integrals, by writing the two expressions:
1
=
−π‘₯π‘₯ 𝑖𝑖 +πœ€πœ€πœ€πœ€
οΏ½ − 1οΏ½ ln οΏ½ οΏ½
− + ∑𝑛𝑛𝑖𝑖=1 𝑠𝑠𝑖𝑖 𝑒𝑒𝑖𝑖−1 = 0,
πœ•πœ• 2 ʌ
𝑛𝑛
π‘₯π‘₯ 𝑖𝑖
With 𝛷𝛷(𝑑𝑑) = 𝑅𝑅𝑑𝑑 . The joint posterior mode of Eq.
(10) is obtained by solving the system of equations:
πœ•πœ•πœ€πœ€ 2
𝑖𝑖=1
+ 𝑣𝑣𝑖𝑖 𝑠𝑠𝑖𝑖−2 𝑒𝑒𝑖𝑖−3 οΏ½ ,
πœ•πœ• 3 ʌ
1
π‘₯π‘₯𝑖𝑖 − πœƒπœƒ 2π‘₯π‘₯𝑖𝑖
π‘₯π‘₯𝑖𝑖
=
= 2 οΏ½ �𝑒𝑒𝑖𝑖−1 οΏ½
+
ln οΏ½ οΏ½οΏ½
2
πœ•πœ•πœ•πœ•πœ•πœ•πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
𝑖𝑖=1
2πœ€πœ€π‘₯π‘₯
π‘₯π‘₯
𝑖𝑖
𝑖𝑖
− 𝑠𝑠𝑖𝑖 𝑒𝑒𝑖𝑖−2 οΏ½
ln οΏ½ οΏ½
πœƒπœƒ
πœƒπœƒ
2π‘₯π‘₯𝑖𝑖 + πœ€πœ€π‘₯π‘₯𝑖𝑖 − πœ€πœ€πœ€πœ€
οΏ½
+
πœƒπœƒ
π‘₯π‘₯𝑖𝑖
π‘₯π‘₯𝑖𝑖
−2
+ 2𝑣𝑣𝑖𝑖 𝑒𝑒𝑖𝑖 οΏ½1 − ln οΏ½ οΏ½οΏ½
πœƒπœƒ
πœƒπœƒ
2 −3 οΏ½
+ 2𝑠𝑠𝑖𝑖 𝑣𝑣𝑖𝑖 𝑒𝑒𝑖𝑖
𝑣𝑣𝑖𝑖 = οΏ½
πœƒπœƒπœƒπœƒ − πœƒπœƒ − 4𝑑𝑑 + 3πœ€πœ€πœ€πœ€
2πœƒπœƒπœƒπœƒ
𝑑𝑑
𝑑𝑑
�𝑒𝑒𝑑𝑑 +
+οΏ½
− 4πœ€πœ€οΏ½ ln οΏ½ οΏ½οΏ½ + 3 𝛷𝛷(𝑑𝑑)
𝑑𝑑
𝑑𝑑
πœƒπœƒ
πœƒπœƒ
ʌ11 =
𝑛𝑛
πœ•πœ• 3 ʌ
2
π‘₯π‘₯𝑖𝑖
π‘₯π‘₯𝑖𝑖
= οΏ½ �𝑠𝑠𝑖𝑖 𝑒𝑒𝑖𝑖−2 οΏ½1 − ln οΏ½ οΏ½οΏ½
πœ•πœ•πœ€πœ€ 2 πœ•πœ•πœ•πœ• πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
where,
πœƒπœƒ
πœƒπœƒ 4
−2(𝑛𝑛 + 1) 6𝑆𝑆
+ 4
πœƒπœƒ 3
πœƒπœƒ 𝑛𝑛
1
−6π‘₯π‘₯𝑖𝑖 + 2πœ€πœ€πœ€πœ€ − 5πœ€πœ€π‘₯π‘₯𝑖𝑖
+ 3 οΏ½ �𝑒𝑒𝑖𝑖−1 οΏ½
πœƒπœƒ
πœƒπœƒ
𝑖𝑖=1
π‘₯π‘₯𝑖𝑖
π‘₯π‘₯𝑖𝑖
− 6πœ€πœ€ ln οΏ½ οΏ½οΏ½
πœƒπœƒ
πœƒπœƒ
2πœ€πœ€π‘₯π‘₯𝑖𝑖
π‘₯π‘₯𝑖𝑖
−2
− 3𝑣𝑣𝑖𝑖 𝑒𝑒𝑖𝑖 οΏ½
ln οΏ½ οΏ½
πœƒπœƒ
πœƒπœƒ
2π‘₯π‘₯𝑖𝑖 + πœ€πœ€π‘₯π‘₯𝑖𝑖 − πœ€πœ€πœ€πœ€
οΏ½ + 2𝑣𝑣𝑖𝑖3 𝑒𝑒𝑖𝑖−3 οΏ½
+
πœƒπœƒ
24
1οΏ½
2
𝑒𝑒𝑒𝑒𝑒𝑒�𝑛𝑛�𝑙𝑙 ∗ οΏ½πœ†πœ†Μ‚∗ οΏ½ − π‘™π‘™οΏ½πœ†πœ†Μ‚οΏ½οΏ½οΏ½ (16)
where, πœ†πœ†Μ‚∗ and πœ†πœ†Μ‚ maximize 𝑙𝑙 ∗ and 𝑙𝑙 respectively
𝛯𝛯 ∗ And 𝛯𝛯 are negatives of the inverse Hessians of 𝑙𝑙 ∗
and 𝑙𝑙 at πœ†πœ†Μ‚∗ and πœ†πœ†Μ‚, respectively. The matrix 𝛯𝛯 takes the
form:
Res. J. Appl. Sci. Eng. Technol., 10(1): 22-28, 2015
𝛯𝛯 = οΏ½
−
−
πœ•πœ• 2 𝑙𝑙
−
πœ•πœ•πœ€πœ€ 2
πœ•πœ• 2 𝑙𝑙
πœ•πœ•πœ€πœ€ 2
πœ•πœ•πœ•πœ•πœ•πœ•πœ•πœ•
οΏ½
πœ•πœ• 2 𝑙𝑙
−
πœ•πœ•πœ•πœ•πœ•πœ•πœ•πœ•
πœ•πœ• 2 𝑙𝑙
πœ•πœ• 2 𝑙𝑙
− ln (πœ€πœ€)
1
𝑛𝑛
−
(1+𝑛𝑛) ln (πœƒπœƒ)
𝑛𝑛
−
𝑑𝑑
𝑆𝑆
𝑛𝑛𝑛𝑛
1
+ ∑𝑛𝑛𝑖𝑖=1 ln(𝑒𝑒𝑖𝑖 ):
𝑑𝑑
𝑛𝑛
𝑑𝑑
:𝑙𝑙 ∗ = ln οΏ½1 + (1 − πœ€πœ€) + πœ€πœ€ ln οΏ½ οΏ½οΏ½ −
ln (πœ€πœ€)
Here:
𝑛𝑛
𝑛𝑛
−
(1+𝑛𝑛) ln (πœƒπœƒ)
𝑛𝑛
−
1
πœ€πœ€πœ€πœ€
οΏ½ (1+𝑛𝑛)
−
+
Also:
𝑛𝑛𝑛𝑛
1
+
πœƒπœƒ
πœƒπœƒ
1 𝑛𝑛
∑𝑖𝑖=1 ln(𝑒𝑒𝑖𝑖 )
𝑛𝑛
πœƒπœƒ
(𝑑𝑑+𝑆𝑆)
𝑛𝑛𝑛𝑛
1
π‘›π‘›πœ€πœ€ 2
1
+ ∑𝑛𝑛𝑖𝑖=1(−𝑠𝑠𝑖𝑖2 𝑒𝑒𝑖𝑖−2 ),
𝑛𝑛
𝑛𝑛
πœ•πœ• 2 𝑙𝑙
1
π‘₯π‘₯𝑖𝑖
π‘₯π‘₯𝑖𝑖
=
οΏ½ �𝑒𝑒𝑖𝑖−1 οΏ½1 − ln οΏ½ οΏ½οΏ½ − 𝑠𝑠𝑖𝑖 𝑣𝑣𝑖𝑖 𝑒𝑒𝑖𝑖−2 οΏ½,
πœ•πœ•πœ•πœ•πœ•πœ•πœ•πœ• 𝑛𝑛𝑛𝑛
πœƒπœƒ
πœƒπœƒ
πœ•πœ•πœƒπœƒ 2
To apply the method, we need to maximize:
𝑙𝑙 =
=
πœ•πœ• 2 𝑙𝑙
πœ•πœ•πœƒπœƒ 2
=
(1+𝑛𝑛)
π‘›π‘›πœƒπœƒ 2
𝑖𝑖=1
−
2𝑆𝑆
π‘›π‘›πœƒπœƒ 3
+
1
π‘›π‘›πœƒπœƒ 2
∑𝑛𝑛𝑖𝑖=1
2πœ€πœ€π‘₯π‘₯𝑖𝑖
π‘₯π‘₯𝑖𝑖
2π‘₯π‘₯𝑖𝑖 + πœ€πœ€π‘₯π‘₯𝑖𝑖 − πœ€πœ€πœ€πœ€
�𝑒𝑒𝑖𝑖−1 οΏ½
ln οΏ½ οΏ½ +
οΏ½ − 𝑣𝑣𝑖𝑖2 𝑒𝑒𝑖𝑖−2 οΏ½
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
−
Partial derivatives of 𝑙𝑙 produce the system of equations:
+ ∑𝑛𝑛𝑖𝑖=1 𝑠𝑠𝑖𝑖 𝑒𝑒𝑖𝑖−1 = 0,
𝑆𝑆
𝑛𝑛
π‘›π‘›πœƒπœƒ 2
+
1
𝑛𝑛𝑛𝑛
∑𝑛𝑛𝑖𝑖=1 𝑣𝑣𝑖𝑖 𝑒𝑒𝑖𝑖−1 = 0
2
𝑑𝑑
𝑑𝑑 𝑑𝑑
− + ln οΏ½ οΏ½
πœ•πœ• 2 𝑙𝑙 ∗ πœ•πœ• 2 𝑙𝑙 1
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
=
− οΏ½
οΏ½ ,
πœ•πœ•πœ€πœ€ 2 πœ•πœ•πœ€πœ€ 2 𝑛𝑛 1 + (1 − πœ€πœ€) 𝑑𝑑 + πœ€πœ€ 𝑑𝑑 ln οΏ½ 𝑑𝑑 οΏ½
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
𝑑𝑑 𝑑𝑑
𝑑𝑑
𝑑𝑑
𝑑𝑑
ln οΏ½ οΏ½
πœ•πœ• 2 𝑙𝑙 ∗
πœ•πœ• 2 𝑙𝑙
𝑑𝑑
𝑑𝑑 οΏ½− πœƒπœƒ + πœƒπœƒ ln οΏ½πœƒπœƒοΏ½οΏ½ οΏ½1 + πœ€πœ€ ln οΏ½πœƒπœƒοΏ½οΏ½
πœƒπœƒ
=
−
οΏ½
οΏ½+ 2
,
πœ•πœ•πœ•πœ•πœ•πœ•πœ•πœ• πœ•πœ•πœ•πœ•πœ•πœ•πœ•πœ• π‘›π‘›πœƒπœƒ 2 1 + (1 − πœ€πœ€) 𝑑𝑑 + πœ€πœ€ 𝑑𝑑 ln οΏ½ 𝑑𝑑 οΏ½
π‘›π‘›πœƒπœƒ
𝑑𝑑
𝑑𝑑 2
𝑑𝑑
(1
+
πœ€πœ€
ln
οΏ½
οΏ½οΏ½
οΏ½1
+
−
πœ€πœ€)
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
2
𝑑𝑑
𝑑𝑑
2
2 + πœ€πœ€ + 2πœ€πœ€ ln οΏ½ οΏ½
1 + πœ€πœ€ ln οΏ½ οΏ½
πœ•πœ• 2 𝑙𝑙 ∗ πœ•πœ• 2 𝑙𝑙
2𝑑𝑑
𝑑𝑑
𝑑𝑑
πœƒπœƒ
πœƒπœƒ
=
−
+
οΏ½
οΏ½+ 4οΏ½
οΏ½ .
πœ•πœ•πœƒπœƒ 2 πœ•πœ•πœƒπœƒ 2 π‘›π‘›πœƒπœƒ 3 π‘›π‘›πœƒπœƒ 3 1 + (1 − πœ€πœ€) 𝑑𝑑 + πœ€πœ€ 𝑑𝑑 ln οΏ½ 𝑑𝑑 οΏ½
π‘›π‘›πœƒπœƒ 1 + (1 − πœ€πœ€) 𝑑𝑑 + πœ€πœ€ 𝑑𝑑 ln οΏ½ 𝑑𝑑 οΏ½
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
Partial derivatives of 𝑙𝑙 ∗ produce the system of equations:
𝑑𝑑
𝑑𝑑 𝑑𝑑
𝑛𝑛
− + ln οΏ½ οΏ½
1
1 1
πœƒπœƒ
πœƒπœƒ πœƒπœƒ
οΏ½
οΏ½
−
+
οΏ½
𝑠𝑠𝑖𝑖 𝑒𝑒𝑖𝑖−1 = 0,
𝑛𝑛 1 + (1 − πœ€πœ€) 𝑑𝑑 + πœ€πœ€ 𝑑𝑑 ln οΏ½ 𝑑𝑑 οΏ½
πœ€πœ€πœ€πœ€ 𝑛𝑛
𝑖𝑖=1
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
𝑑𝑑
𝑛𝑛
⎨ (1 + 𝑛𝑛) 𝑑𝑑 + 𝑆𝑆
1 + πœ€πœ€ ln οΏ½ οΏ½
1
𝑑𝑑
πœƒπœƒ
−1
βŽͺ
βŽͺ− 𝑛𝑛𝑛𝑛 + π‘›π‘›πœƒπœƒ 2 + 𝑛𝑛𝑛𝑛 οΏ½(𝑣𝑣𝑖𝑖 𝑒𝑒𝑖𝑖 ) + πœƒπœƒ 2 οΏ½
𝑑𝑑
𝑑𝑑 οΏ½ = 0.
𝑑𝑑
1 + (1 − πœ€πœ€) + πœ€πœ€ ln οΏ½ οΏ½
𝑖𝑖=1
⎩
πœƒπœƒ
πœƒπœƒ
πœƒπœƒ
⎧
βŽͺ
βŽͺ
Data analysis: We take the data π‘₯π‘₯𝑖𝑖 used by Lawless (2003, 2002):
14
464
1174
2001
3248
5267
34
479
1270
2161
3715
5299
59
556
1275
2292
3790
5583
61
574
1355
2326
3857
6065
69
839
1397
2337
3912
9701
80
917
1477
2628
4100
123
969
1578
2785
4106
142
991
1649
2811
4116
165
1064
1702
2886
4315
210
1088
1893
2993
4510
381
1091
1932
2993
4584
These data correspond to an empirical failure rate "out of bath-tub" that the author proposed to identify a
balanced additive mixture of two Weibull distributions having this survival function:
𝑑𝑑
𝛽𝛽 1
𝑑𝑑
𝛽𝛽 2
𝑅𝑅(𝑑𝑑, 𝛼𝛼1 , 𝛽𝛽1 , 𝛼𝛼2 , 𝛽𝛽2 , 𝑝𝑝) = 𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 οΏ½− �𝛼𝛼 οΏ½ οΏ½ + (1 − 𝑝𝑝)𝑒𝑒𝑒𝑒𝑒𝑒 οΏ½− �𝛼𝛼 οΏ½ οΏ½
1
2
25
Res. J. Appl. Sci. Eng. Technol., 10(1): 22-28, 2015
By maximization of likelihood:
𝛽𝛽1𝑀𝑀𝑀𝑀 = 1.66(0.49), 𝛼𝛼1𝑀𝑀𝑀𝑀 = 95.4(25.8), 𝛽𝛽2𝑀𝑀𝑀𝑀 = 1.40(0.18), 𝛼𝛼2𝑀𝑀𝑀𝑀 = 2774.5(314.2),
𝑝𝑝𝑀𝑀𝑀𝑀 = 0.137(0.051)
Fig. 1: Empirical reliability (black solid line), 𝑅𝑅𝑀𝑀𝑀𝑀 (red dotted lines), 𝑅𝑅𝐿𝐿 (green dotted lines)
Fig. 2: 𝑅𝑅𝑀𝑀𝑀𝑀 (𝑑𝑑) (black solid line), 𝑅𝑅𝐿𝐿 (red dotted lines)
πœ€πœ€π‘€π‘€π‘€π‘€ = 0.291(0.087), πœƒπœƒπ‘€π‘€π‘€π‘€ = 1190.4(78.7)
(The values between brackets are the standard
deviations associated to the estimators). The distance of
Kolmogorov-Smirnov between empirical reliability and
estimated reliability by this model of mixture with 5
parameters is 𝐷𝐷𝐾𝐾 = 0.0417. As alternative, we propose
a model of mixture with 2 parameters of the type (2) by
noticing that the first term of the density (1) translated
the fact that a small number of components of the
population presents defects known as of "youth" or with
"infant mortality". The maximization of likelihood
provides the following estimators:
Also Kolmogorov-Smirnov distances between
empirical reliability and the reliability estimated by
maximum likelihood and Lindley's method of the new
model are, respectively: 𝐷𝐷𝐾𝐾𝐾𝐾𝐾𝐾 = 0.0460 and 𝐷𝐷𝐾𝐾𝐾𝐾,𝐿𝐿 =
0.0490. The three distances of Kolmogorov-Smirnov
are very close what translates the great proximity of the
survival functions estimated starting from the
traditional model of mixture and of the new model and
the fact that they are practically indistinguishable. This
26
Res. J. Appl. Sci. Eng. Technol., 10(1): 22-28, 2015
is illustrated in the Fig. 1 and 2. The Kolmogorov
Smirnov distances are presented in Table 1 and 2.
this can be explained by the use of the non-informative
prior.
Simulation study: In trying to illustrate and compare
the methods as described above, a random sample of
size, 𝑛𝑛 = 25,50,100 and 1000 were generated from (1)
with πœ€πœ€ = 0.5 and πœƒπœƒ = 1.5. These were replicated 1000
times (Table 3 and 4).
Remark: Estimators of the survival function obtained
by maximum likelihood and Lindley's approximation
for any value of t; are practically indistinguishable and
very close to the true values of the survival function;
Table 1: Kolmogorov-Smirnov distances
𝐷𝐷𝐾𝐾 z
0.042
𝐷𝐷𝐾𝐾𝐾𝐾𝐾𝐾
0.046
Table 2: The estimators "MLE and Lindley of 𝑅𝑅𝑑𝑑 "
Method
𝑑𝑑 = 10
𝑑𝑑 = 20
MLE
0.9860
0.9754
Lindley
0.9862
0.9759
𝑑𝑑 = 40
0.9579
0.9587
Table 3: Estimates of parameters
MLE
--------------------------------------------------------------N
πœƒπœƒοΏ½
πœ€πœ€Μ‚
25
0.5951
1.2436
(9.044616×10⁻⁢)
(7.967312×10⁻⁡)
50
0.5000
1.2950
(2.456328×10⁻¹²)
(4.202017×10⁻⁡)
100
0.4965
1.3287
(1.238528×10⁻⁸)
(2.9349×10⁻⁡)
1000
0.4655
1.4917
(1.193768×10⁻⁢)
(6.869766×10⁻⁸)
Table 4: Survival function estimation
πœƒπœƒ
------------------------------0.5
𝑑𝑑
0.7048
𝑅𝑅𝑑𝑑
MLE
0.6320
LINDLEY
0.6367
𝑛𝑛 = 25
T-K
0.6601
MLE
0.6861
LINDLEY
0.6946
𝑛𝑛 = 50
T-K
0.2706
MLE
0.6911
LINDLEY
0.6945
𝑛𝑛 = 100
T-K
0.1703
MLE
0.7214
LINDLEY
0.7232
𝑛𝑛 = 1000
T-K
0.1123
𝑑𝑑 = 60
0.9432
0.9442
𝑑𝑑 = 80
0.9302
0.9315
𝑑𝑑 = 100
0.9185
0.9200
LINDLEY
---------------------------------------------------------------πœ€πœ€Μ‚
πœƒπœƒοΏ½
0.5291
1.1766
(8.461261×10⁻⁷)
(4.635779×10⁻⁡)
0.4675
1.2949
(1.058368×10⁻⁢)
(4.20737×10⁻⁡)
0.4832
1.3243
(2.836454×10⁻⁷)
(3.088453×10⁻⁡)
0.4632
1.5040
(1.35127×10⁻⁢)
(1.63151×10⁻⁸)
two Weibull distributions with a Kolmogorov Smirnov
distance which is the same result found by Idée and
Pierrat (2010); Idée (2006) in addition, the estimators of
parameters and the survival function of this model
obtained by the maximum of likelihood and Bayesian
methods with a non-informative prior and a quadratic
loss function are equivalent. Tierney-Kadane gives bad
results for the large samples (𝑛𝑛 ≥ 50).
θ = 1.5
-----------------------------1.5
2.5
0.5518
0.4267
0.4858
0.3549
0.4923
0.3600
0.4875
0.3266
0.5226
0.3772
0.5319
0.3870
0.1986
0.1317
0.5292
0.3867
0.5327
0.3901
0.1261
0.0860
0.5634
0.4302
0.5658
0.4334
0.0757
0.0367
REFERENCES
Dempster, A.P., N.M. Laird and D. Rubin, 1977.
Maximum likelihood from incomplete data via the
EM algorithm. J. Roy. Stat. Soc. B Met., 39(1):
1-38.
The comparisons were based on values from Mean
Square Error (MSE).
Where,
Elmahdy, E.E. and A.W. Aboutahoun, 2013. A new
approach for parameter estimation of finite
Weibull mixture distributions for reliability
modelling. Appl. Math. Model., 37: 1800-1810.
Gelman, A., J.B. Carlin, H.S. Stern and D.B. Rubin,
2003. Bayesian Data Analysis. 2nd Edn., Chapman
and Hall, New York.
Idée, E., 2006. Loi de probabilité obtenue par
perturbation aléatoire simple de l'un des
paramètres de la loi de référence. Prépublication
du LAMA-Université de Savoie, N° 07-06a.
Idée, E. and L. Pierrat, 2010. Genèse et estimation d'un
modèle de fiabilité en Baignoire à Taux de
Défaillance Borné. 42ème Journées de Statistique,
Marseille, France.
2
οΏ½
π‘€π‘€π‘€π‘€π‘€π‘€οΏ½πœƒπœƒοΏ½ οΏ½ = (1000)−1 ∑1000
𝑖𝑖=1 οΏ½πœƒπœƒπ‘–π‘– − πœƒπœƒοΏ½ .
The results of the survival function are given in the
Table 4.
CONCLUSION
The model used in this study, is a mixture of a
1
𝐺𝐺 οΏ½2, οΏ½ and 𝐿𝐿(πœƒπœƒ) distribution which depends on πœƒπœƒ. The
πœƒπœƒ
mixture parameter is πœ€πœ€. We have shown the equivalence
between this model and a classical mixture model of
𝐷𝐷𝐾𝐾𝐾𝐾,𝐿𝐿
0.049
27
Res. J. Appl. Sci. Eng. Technol., 10(1): 22-28, 2015
Jiang, R. and D.N.P. Murthy, 1995. Modeling failuredata by mixture of 2 Weibull distributions: A
graphical approach. IEEE T. Reliab., 44(3):
477-488.
Jiang, S.Y. and D. Kececioglu, 1992. Maximum
likelihood estimates from censored data for mixedWeibull distribution. IEEE
T. Reliab., 41:
248-255.
Lawless, J.F., 2002. Statistical Models and Methods for
Lifetime Data. 2nd Edn., A John Wiley and Sons,
New York.
Lawless, J.F., 2003. Statistical Models and Methods for
Lifetime Data. Wiley Series in Probability and
Statistics, John Wiley and Sons, Hoboken, NJ.
Levine, R.A. and G. Casella, 2001. Implementation of
the Monte-Carlo EM algorithm. J. Comput. Graph.
Stat., 10: 422-439.
Lindley, D.V., 1980. Approximate Bayesian methods.
Trabajos Estadist. Invest. Oper., 31: 232-245.
Tierney, L. and J.B. Kadane, 1986. Accurate
approximations for posterior moments and
marginal densities. J. Am. Stat. Assoc., 81: 82-86.
Tsionas, E.G., 2002. Bayesian analysis of finite
mixtures of Weibull distributions. Commun. Stat.
A-Theor., 31(1): 37-48.
Wei, G.C.G. and M.A. Tanner, 1990. A monte carlo
implementation of the EM Al-gorithm and the
poor man’s data augmentation algorithms. J. Am.
Stat. Assoc., 85: 699-704.
28
Download