Available online at www.sciencedirect.

com

Procedia Social and Behavioral Sciences 8 (2010) 663–669

International Conference on Mathematics Education Research 2010 (ICMER 2010)

Bayesian Estimator for Weibull Distribution with Censored Data

using Extension of Jeffrey Prior Information
Al Omari Mohammed Ahmeda,*, Noor Akma Ibrahimb
a
Department of Mathematics, Faculty of Science, Universiti Putra Malaysia
b
Institute for Mathematical Research, Universiti Putra Malaysia

Abstract

The Weibull distribution provides a statistical model which has a wide variety of applications in many areas, including life
testing and reliability theory. The main advantageous of this distribution is its ability to provide reasonably accurate failure
analysis and failure forecasts with extremely small samples. Bayesian approach has received much attention and in contention
with other estimation methods. In this study we explore and compare the performance of the Bayesian using Jeffrey prior and the
extension of Jeffrey prior information with maximum likelihood method for estimating the parameters of Weibull distribution
with censored data. Through the simulation study comparisons are made on the performance of these estimators with respect to
the Mean Square Error (MSE) and Mean Percentage Error (MPE). For all the varying sample size, several specific values of the
scale parameter of the Weibull distribution and for the values specify for the extension of Jeffrey prior, the estimators of Jeffrey
prior result in smaller MSE and MPE compared to Bayesian estimator using extension of Jeffrey prior in majority of the cases.
Nevertheless in all cases for both methods the MSE and MPE decrease as sample size increases.
© 2010 Elsevier Ltd. All rights reserved. Open access under CC BY-NC-ND license.

Keywords: Extension of Jeffrey prior information; Weibull distribution; Bayes method; Right Censoring

1. Introduction

The Weibull distribution is most useful for modelling and analyzing of life time data in the fields of medicine,
biology, engineering sciences and others. A great deal of research has been done on estimating the parameters of the
Weibull distribution using both classical and Bayesian techniques. A very good summary of this work can be found
in Johnson, Kotz and Balakrishnan (1994). Recently, Hossain and Zimmer (2003) have discussed some comparisons
of estimation methods for Weibull parameters using complete and censored samples. Many methods have been
proposed for estimating parameters of the Weibull distribution because of its useful applications. The maximum
likelihood and the moment estimation are the most-used methods and are considered to be traditional (see Cohen
and Whitten, 1982). The maximum likelihood estimation method is the most popular for its efficiency, and the
moment estimation method is computationally easy and provides explicit estimators of the parameters. Sinha and
Sloan (1988) obtained Bayes estimator of three parameters of the Weibull distribution and compared the posterior
standard deviation estimates counterparts and numerical example are given. Soliman (2006) compared the estimates

* Corresponding author. Tel.:+60176517234; fax: +6038942 3789.

E-mail address: mamo1050@hotmail.com (A.O.M. Ahmed).

1877-0428 © 2010 Published by Elsevier Ltd. Open access under CC BY-NC-ND license.
doi:10.1016/j.sbspro.2010.12.092
664 Al Omari Mohammed Ahmed and Noor Akma Ibrahim / Procedia Social and Behavioral Sciences 8 (2010) 663–669

of record statistics using Weibull model with Bayesian and non-Bayesian approaches. In this paper we propose
Bayes with Jeffrey prior and extension of Jeffrey prior information for the Weibull parameters estimation.
The rest of the paper is arranged as follows. In section 2 Bayes estimator with the Jeffrey prior and the new
extension of Jeffrey prior information are presented. Section 3 presented the simulation study and results and this is
followed by the conclusion in section 4.

2. Bayes Estimation

Let t1 , , tn be a random sample of size n with distribution function F (t ,T , p) and probability density
function f (t ,T , p) .
In the Weibull case, we assumed that the probability density function of the lifetime is given by:

p
p p 1 t
f (t ; T , p ) t exp(  )
T T

2.1 Jeffrey Prior Information

We find Jeffery prior by taking g (T ) k I (T ) , where

n
2n  ¦ G i
I (T ) i 1 , then
2
T

n
k 2n  ¦ G i
i 1
g (T )
T

We can found Bayes estimator with Jeffrey prior of parametric distribution by using conditional distribution, depend
on joint probability density function and marginal probability density function, so the posterior distribution is given
by:

H t1, } tn ; T , p 

– t1, } tn ; T , p  p(t1, } , tn )

The joint probability density function f (t1 , tn ;T , p) is given by:

n G 1G
H (t1,..., tn ; T , p) – [ f (ti ; T , p)] i [ S (ti ; T , p)] i g (T )
i 1

n
k 2n  ¦ G i
n
¦ Gi § ¦ n p
t ·
i 1 § p n p 1 ·i 1 ¨ i 1i ¸
T ©T
¦
¨ i 1 t ¸
¹
exp
¨ 
T ¸ dT
¨ ¸
© ¹
The marginal probability density function of (t1 , , tn ) is given by:
p(t1 ,..., tn ) ³ H (t ,..., t ) dT
1 n
 Al Omari Mohammed Ahmed and Noor Akma Ibrahim / Procedia Social and Behavioral Sciences 8 (2010) 663–669 665

n
§ n p 1 · ¦ Gi
p¦t i 1
n n ¨ i 1 ¸
k ( ¦ G i  1)! 2n  ¦ G i
i 1 i 1 ¨ n p ¸
¨ ¦t ¸
© i 1 ¹

Then the posterior probability density function of T given the data (t1 , , tn ) by:


H t1, } tn ; T , p 

– t1, } tn ; T , p  p(t1, } , tn )
n

 
n p ¦ Gi
¦t i 1
i 1i
§ ¦n p·
t
¨ i 1i ¸
exp 
n ¨ T ¸¸
¦ Gi 1 n ¨
Ti 1 ( ¦ G i  1)! © ¹
i 1

By using squared error loss function (Tˆ  T ) c(Tˆ  T )2 , the Risk function is

R(Tˆ  T ) E (Tˆ  T )
f
³ (Tˆ  T )– (t1,..., tn ; T , p ) dT
0
n

2 n
cTˆ  2cTˆ ³ n
 
i
n p ¦ Gi
¦ ti i 1
1
§ ¦
¨
n p·

exp  i 1
ti
¸
0 ¦ Gi 1 n ¨ T ¸ dT  \ (T )
¨ ¸
Ti 1 ( ¦ G i  1)! © ¹
i 1
wR(Tˆ  T )
Let 0 , then the Bayes estimator is
wTˆ
n

TˆBJ
 
n p ¦ Gi
¦t i 1 n
i 1i 1
§ ¦
¨
n p·

exp  i 1
ti
¸
n ³ n ¨ T ¸ dT ,
( ¦ G i  1)! 0 ¦ Gi ¨ ¸
i 1 Ti 1 © ¹

the scale parameter for Jeffrey prior is

n p
¦ ti
TˆBJ n
i 1 (1)
¦ Gi  1
i 1
666 Al Omari Mohammed Ahmed and Noor Akma Ibrahim / Procedia Social and Behavioral Sciences 8 (2010) 663–669

2.2 Extension of Jeffrey prior information

c 
The extension of Jeffrey prior is taking g (T ) v [( I (T )] , c  R

 
n c
k 2n  ¦ G i
i 1
then g (T ) , k is a constant.
2c
T

The same way above, we can find the Bayes estimator with extension of Jeffrey prior depend on posterior
probability density function as following: The joint probability density function is given by:

n G 1G
H (t1,..., tn,T , p ) – [ f (ti ;T , p )] i [S (ti ;T , p )] i g (T )
i 1

  §¨
n c
k 2n  ¦ G i
n
¦ Gi § ¦ n p
t ·
i 1 p n p 1 ·i 1 ¨ i 1i ¸
2c
T
¦t
©T i 1
¸
¹
exp
¨ 
T ¸ dT
¨ ¸
© ¹
The marginal probability density function of (t1 , , tn ) is given by:

p(t1 ,..., tn ) ³ H (t ,..., t ) dT

1 n
n

n n
k ( ¦ G i  2c  2)! 2n  ¦ G i
c
    n p 1 ¦ Gi
p¦ t
i 1
n
i 1

  n p ¦ Gi  2c 1
i 1 i 1
¦t i 1
i 1

Then the conditional probability density function of T given the data (t1 , , tn ) by:

H  t1 ,}tn ;T , p 
–  t ,}t ;T , p 
1 n
p(t1 , }, tn )
n

¦Gi  2c 1
§ n
p ·i 1 § n p ·
¨ ¦ ti ¸ ¨ ¦ ti ¸
©i1 ¹ exp ¨  i 1 ¸
¨ T ¸
n

¦Gi  2 c n
T i 1
(¦ G i  2c  2)! ¨
©
¸
¹
i 1

By using squared error loss function (Tˆ  T ) c(Tˆ  T )2 , the Risk function is

R(Tˆ T ) E (Tˆ T )
 Al Omari Mohammed Ahmed and Noor Akma Ibrahim / Procedia Social and Behavioral Sciences 8 (2010) 663–669 667

³ (Tˆ  T )– (t1 ,..., tn ;T , p) dT

0
n

¦Gi  2c 1
§ n
p ·i 1 § n p ·
n ¨ ¦ ti ¸ ¨ ¦ ti ¸
cTˆ 2  2cTˆ ³ ©i1 ¹ exp ¨  i 1 ¸ dT \ (T )
¨ T ¸
n

¦Gi  2c 1 n
(¦ G i  2c  2)! ¨ ¸
0
T i 1
© ¹
i 1

wR(Tˆ  T )
Let 0 , the scale parameter for Jeffrey prior is
wTˆ
n

¦t i
p

TˆBE n
i 1
(2)
¦G
i 1
i  2c  2

3. Simulation and Results

In this simulation study, we have chosen n=25, 50, 100 to represent small moderate and large sample size where the
percentage of censoring is 30% for each samples, several values of parameter T = 0.5, 1.5 and p=0.8, 1.2 and two
values of Jeffery extension c= 0.4, 0.8. The number of replication used was R= 1000. After the parameter was
estimated , mean square error (MSE) and mean percentage error (MPE) were calculated to compare the methods of
estimation, where

1000 1000
| Tˆi  T |
¦ (Tˆi  T )2 ¦ T
MSE (T ) i 1
MPE (T ) i 1

R R

The results of the simulation study are summarized and tabulated in tables 1-2 for the MSE and the MPE of the two
estimators for all sample size and T , p values respectively.

In each row of tables 1-2 we have two values of estimators that is the Jeffrey prior and extension of Jeffrey prior.
The better method is the method that gives the smaller value of (MSE) and (MPE).
668 Al Omari Mohammed Ahmed and Noor Akma Ibrahim / Procedia Social and Behavioral Sciences 8 (2010) 663–669

Table 1. MSE estimated parameters of Weibull distribution.

Size T C P Bayes Extension

25(8) 0.5 0.4 0.8 0.1954 0.3096

1.2 0.0530 0.0976
0.8 0.8 0.1954 0.3813
1.2 0.0530 0.1276
1.5 0.4 0.8 0.6466 1.1460
1.2 1.3830 2.2450
0.8 0.8 0.6466 1.4752
1.2 1.3830 2.7913
50(15) 0.5 0.4 0.8 0.1354 0.1723
1.2 0.0287 0.0412
0.8 0.8 0.1354 0.1916
1.2 0.0287 0.0481
1.5 0.4 0.8 0.3778 0.5240
1.2 0.9283 1.2014
0.8 0.8 0.3778 0.6039
1.2 0.9283 1.3459
100(30) 0.5 0.4 0.8 0.1171 0.1328
1.2 0.0210 0.0259
0.8 0.8 0.1171 0.1404
1.2 0.0210 0.0283
1.5 0.4 0.8 0.2934 0.3525
1.2 0.7888 0.9040
0.8 0.8 0.2934 0.3818
1.2 0.7888 0.9599

Table 2. MPE estimated parameters of Weibull distribution.

Size T C P Bayes Extension

25(8) 0.5 0.4 0.8 0.8065 1.0349

1.2 0.3840 0.5462
0.8 0.8 0.8065 1.1563
1.2 0.3840 0.6361
1.5 0.4 0.8 0.4581 0.6354
1.2 0.7061 0.9213
0.8 0.8 0.4581 0.7319
1.2 0.7061 1.0358
50(15) 0.5 0.4 0.8 0.6949 0.7896
1.2 0.2905 0.3587
0.8 0.8 0.6949 0.8352
1.2 0.2905 0.3925
1.5 0.4 0.8 0.3629 0.4375
1.2 0.6002 0.6895
0.8 0.8 0.3629 0.4737
1.2 0.6002 0.7326
100(30) 0.5 0.4 0.8 0.6645 0.7091
1.2 0.2626 0.2959
0.8 0.8 0.6645 0.7297
1.2 0.2626 0.3114
1.5 0.4 0.8 0.3365 0.3722
1.2 0.5715 0.6135
0.8 0.8 0.3365 0.3886
 Al Omari Mohammed Ahmed and Noor Akma Ibrahim / Procedia Social and Behavioral Sciences 8 (2010) 663–669 669

1.2 0.5715 0.6330

4. Discussion

In Table 1, when we compare parametric estimators of Weibull distribution in Bayes using Jeffery prior and
extension of Jeffery prior by Mean Square Error (MSE) we find the Bayes using Jeffery prior is better than
extension of Jeffery prior by 100%.

In Table 2, when we compare parametric estimators of Weibull distribution in Bayes using Jeffery prior and
extension of Jeffery prior by Mean Square Error (MPE) we find the Bayes using Jeffe ry prior is better than
extension of Jeffery prior by 100%.

5. Conclusion

The estimated parameters of Weibull distribution obtained from the. Bayes using Jeffrey prior gives better result
than Bayes using extension of Jeffrey prior. When the number of sample size increases the Mean Square Error
(MSE) and mean percentage error (MPE) decrease in all cases

References
Cohen, A.C., & Whitten, B. (1982). Modified Maximum Likelihood and Modified Moment Estimators for the Three-parameter Weibull
Distribution. Communications in Statistics Theory and Methods, 11 (23), 2631–2656.

Hossain, A. M.,& Zimmer, W. J. (2003). Comparison of Estimation Methods for Weibull Parameters: Complete and Censored Samples. Journal
of Statistical Computation Simulation, 73 (2), 145–153.

Johnson, N. L., Kotz, S., & Balakrishnan, N. (1994). Continuous Univariate Distributions, second edition, vol. 1. John Wiley. New York.

Sinha, S. K., & Sloan, J. A. (1988). Bayes Estimation of the Parameters and Reliability Function of the 3-parameter Weibull distribution. IEEE
Transactions Reliability, 37:364–356.

Soliman, A. A., Abd Ellah, A. H., & Sultan, K. S. (2006). Comparison of Estimates Using Record Statistics from Weibull model: Bayesian and
non-Bayesian Approaches. Computational Statistics and Data Analysis, 51(3), 2065-2077.