Professional Documents
Culture Documents
Censored Data
Masoud Yarmohammadi
Hassan Pazira
Abstract
In this paper, we obtain Bayesian and non-Bayesian estimators for the shape
parameter, reliability and failure rate functions of the Generalized-Exponential
distribution in the case of complete and type-II censored samples. The Bayes
estimators are obtained using symmetric and asymmetric loss functions.
Comparisons are made between these estimators using Monte-Carlo simulation
study.
1. Introduction
Recently a new distribution, called Generalized-Exponential (GE)
distribution, has been introduced. This distribution can be used quite
effectively in situations where a skewed distribution is needed. It has been
studied extensively by Gupta and Kundu (1999, 2001a, 2001b, 2002, 2003a,
2003b, 2004), Raqab (2002), Raqab and Ahsanullah (2001), Zheng (2002) and
Kundu, Gupta and Manglick (2004). Singh et al. (2008) estimated the
parameters of this distribution under some symmetric and asymmetric loss
functions using Lindley's approximation technique. Note that the generalized
exponential distribution is a sub-model of the exponentiated Weibull
distribution introduced by Mudholkar and Srivastava (1993) and later studied
by Mudholkar, Srivastava and Freimer (1995) and Mudholkar and Hutson
(1996).
We say that the random variable T has a Generalized Exponential
distribution if T has the density f (t ;θ ) , i.e.,
θˆP2 = E (θ 2 | t ) . (1.5)
Then, we use the LINEX loss function which is asymmetric, was introduced
by Varian (1975). This loss function was widely used by several authors;
among of them Basu and Ebrahimi (1991), Pandey (1997), Soliman (2000) and
Nassar and Eissa (2004). The LINEX loss function may be expressed as
θˆ L = − ln (E [e −cθ | t ]),
1
(1.8)
c
provided that the expectation E [e −cθ | t ] exists and is finite, see Calabria and
Pulcini (1996).
We were interested with maximum likelihood estimation as a classical
approach among non-Bayesian methods. The maximum likelihood is based on
the information provided by empirical data. The invariant property was hold to
obtain maximum likelihood estimators (MLE’s) of reliability and failure rate
functions.
In this paper, a discussion of the MLE’s is considered in section 2. In
section 3, Bayesian estimators are obtained. Also, numerical illustration and
comparisons are presented in section 4. Finally, our conclusion is stated in
Section 5.
(
l(t ;θ ) ∝ θ r e −θT 1 − V θ )
n−r
(2.1)
where
( )
r
T = −∑ ln 1 − e −ti
−t ( r )
and V = 1 − e .
i =1
The logarithm of the LF is given by
L = ln l(t ;θ ) ∝ r ln θ − θT + (n − r ) ln (1 − V θ ). (2.2)
r
θˆM = −1
. (2.3)
T + (n − r )⎛⎜V −θ M − 1⎞⎟ ln V
^
⎝ ⎠
Classical and Bayesian estimations 1421
3. Bayesian Estimation
δ ν V −1 −δθ
g (θ ) = θ e , θ > 0,ν > 0, δ > 0. (3.1)
Γ(ν )
From which the prior mean and variance of θ are given, respectively, by
ν δ and ν δ 2 . Applying Bayes theorem, we obtain from Equations (2.1) and
(3.1), the posterior density of θ as
g (θ t ) =
(δ + T )r +ν r +ν −1 −θ (T +δ )
θ e (
1−V θ ) , θ
n−r
> 0,ν > 0, δ > 0, (3.2)
kΓ(r + ν )
where
n− r − (r +v )
⎛ jp ⎞
k = ∑ w( j )⎜1 − ⎟ , (3.3)
j =0 ⎝ δ +T ⎠
⎛n − r⎞ − t( r )
w( j ) = (−1) j ⎜⎜ ⎟⎟ , p = ln V and V =1− e (3.4)
⎝ j ⎠
Estimation of θ :
The Bayes estimate θˆS of θ relative to squared error loss function is given
by
r+v
θˆS = ξ1 , (3.5)
k (δ + T )
where
n−r − ( r + v +1)
⎛ jp ⎞
ξ1 = ∑ w( j )⎜1 − ⎟ .
j =0 ⎝ δ +T ⎠
Under Precautionary loss function, the Bayes estimate θˆP of θ using
Equation (1.5) can be obtained as
1422 M. Yarmohammadi and H. Pazira
⎛ (r + v + 1)(r + v ) ⎞
12
θˆP = ⎜⎜ ξ 2 ⎟⎟ (3.6)
⎝ k (δ + T )
2
⎠
where
n−r −(r +v + 2 )
⎛ jp ⎞
ξ 2 = ∑ w( j )⎜1 − ⎟ .
j =0 ⎝ δ +T ⎠
Under LINEX loss function, the Bayes estimate θˆL of θ using Equation
(1.8) can be obtained as
1
θˆL = − ln ξ 3 , c ≠ 0 (3.7)
c
where
− (r +v )
⎛ c − jp ⎞
n− r
ξ 3 = k ∑ w( j )⎜1 +
−1
⎟ .
j =0 ⎝ δ +T ⎠
Estimation of R(t ) :
Consider the reliability R = R(t ) is a parameter itself. Replacing θ in terms
of R by that of Equation (3.2), we obtain posterior density function of R as
Q r +ν
g (R t ) = k −1
Γ(r + ν )
(
[ϕ1 (R )]r +ν −1 e −(Q −1)ϕ1 (R ) 1 − V Zϕ1 (R ) )
n−r
, 0 < R < 1, (3.8)
where
Q = Z (δ + T ), ϕ1 (R ) = ln (1 − R )
−1 1
. and Z =
(3.9)
ln 1 − e −t
−1
( )
Assuming the quadratic loss is appropriate, the Bayes estimate of the
reliability function R is
n−r
Rˆ S = k −1 ∑ w( j )(ξ 4 − ξ 5 ), (3.10)
j =0
where
− ( r +v ) − (r +v )
⎛ jpZ ⎞ ⎛ jpz − 1 ⎞
ξ 4 = ⎜⎜1 − ⎟ , ξ 5 = ⎜⎜1 − ⎟ ,
⎝ Q ⎟⎠ ⎝ Q ⎟⎠
12
⎡ n− r ⎤
, Rˆ P = ⎢k −1 ∑ w( j )(ξ 4 − 2ξ 5 + ξ 6 )⎥ (3.11)
⎣ j =0 ⎦
Where
Classical and Bayesian estimations 1423
− (r +v )
⎛ jpZ − 2 ⎞
ξ 6 = ⎜⎜1 − ⎟ .
⎝ Q ⎟⎠
Under LINEX loss function, the Bayes estimate of R using Equation (1.8)
is
1
, Rˆ L = 1 − ln ξ 7 , c ≠ 0 (3.12)
c
where
− (r +v )
∞
c i n− r ⎛ i − jpZ ⎞
ξ 7 = k ∑ ∑ w( j )⎜⎜1 +
−1
⎟ .
i =0 i! j =0 ⎝ Q ⎟⎠
Estimation of H (t ) :
To derive the Bayes estimate of the cumulative failure rate function
H (t ) = − ln R(t ) , we first obtain the posterior density function of H = H (t ) ,
which can be given by
Q r +v e− H
(3.13) g (H t ) = k −1
Γ(r + v ) 1 − e −H
(
[ϕ 2 (H )] r +v−1 e −(Q )ϕ2 (H ) 1 − V Zϕ2 (H ) ) n−r
, H >0
where
ϕ 2 (H ) = ln (1 − e − H ) .
−1
Q r +ν
Hˆ S = k −1 G1 , (3.14)
Γ(r + ν )
where
∞
(
G1 = ∫ ln 1 − e − x )−1
(
x r +ν −1e −Qx 1 − V Zx ) n−r
dx.
0
n−r ∞
j ⎛n − r⎞ Γ(r + ν )
G1 = ∑∑ (− 1) ⎜⎜ ⎟⎟ .
⎝ j ⎠ i(Q + i − jZp )
r +ν
j = 0 i =1
12
⎛ k −1Q r +ν ⎞
Hˆ P = ⎜⎜ G2 ⎟⎟ , (3.15)
⎝ Γ(r + ν ) ⎠
where
( )x
∞
(
G2 = ∫ ln 1 − e − x )
−1 2 r +ν −1 −Qx
e (1 − V )Zx n − r
dx.
0
n−r ∞ ∞
j ⎛n − r⎞ Γ(r + ν )
G 2 = ∑∑∑ (− 1) ⎜⎜ ⎟⎟ .
⎝ j ⎠ il (i + l + Q − jZp )
r +ν
j = 0 i =1 l =1
1 ⎛ k −1Q r +ν ⎞
Hˆ L = − ln ⎜⎜ G3 ⎟⎟, c ≠ 0, (3.16)
c ⎝ Γ(r + ν ) ⎠
where
∞
(
G3 = ∫ (1 − e − x ) c x r +ν −1e −Qx 1 − V Zx )
n−r
dx.
0
n−r c
i+ j ⎛ n − r ⎞ ⎛ c ⎞ Γ(r + ν )
G3 = ∑∑ (− 1) ⎜⎜ ⎟⎟ ⎜⎜ ⎟⎟ r +ν .
j =0 i =0 ⎝ j ⎠ ⎝ i ⎠ (i + Q − jZp )
4. Simulation Study
We obtained, in the above Sections, Bayesian and non-Bayesian estimates
for the shape parameter θ , reliability, R(t ) , and failure rate, H (t ) , functions
of the GE (θ ) distribution. We adopted the squared error loss, Precautionary
and LINEX loss functions. The MLE’s are also obtained.
In order to assess the statistical performances of these estimates, a
simulation study is conducted. The mean square errors (MSE’s) using
generated random samples of different sizes are computed for each estimator.
The random samples are generated as follows:
computations are carried out for such sample sizes and censored samples of
sizes: 12, 20 and 40, respectively.
3. The MLE of the parameter θ , θˆM is obtained by iteratively solving the
Equation (2.3). The estimators Rˆ (t ) and Hˆ (t ) of the functions R(t ) and
M 0 M 0
Table 1. Bias and MSE of different estimators of θ , for different sample size.
Figure 1. MSE's of the estimators θˆM , θˆS , θˆP and θˆL in the case of complete samples.
Figure 2. MSE's of the estimators θˆM , θˆS , θˆP and θˆL in the case of type-II censored samples.
Table 2. Bias and MSE of different estimators of R (t ) , for different sample size.
n r Criteria R̂M R̂S R̂P R̂L
CC
1 5
Figure 3. MSE's of the estimators R̂M , R̂S , R̂P and R̂L in the case of complete samples.
Figure 4. MSE's of the estimators R̂M , R̂S , R̂P and R̂L in the case of type-II censored samples.
Table 3. Bias and MSE of different estimators of H (t ) , for different sample size.
n r Criteria Ĥ M Ĥ S Ĥ P Ĥ L
CC
1 5
Figure 6. MSE's of the estimators Ĥ M , Ĥ S , Ĥ P and Ĥ L in the case of type-II censored samples.
5. Conclusion
In this paper, we have presented the Bayesian and non-Bayesian estimates
of the shape parameter θ , reliability, R (t ) , and failure rate, H (t ) , functions of
the lifetimes follow the GE (θ ) distribution. The estimations are conducted on
the basis of complete and type-II censored samples. Bayes estimators, under
squared error, Precautionary and LINEX loss functions, are derived. The
MLE’s are also obtained. Our observations about the results are stated in the
following points:
1- Table 1 shows that the Bayes estimates under the LINEX loss function
have the smallest estimated MSE's as compared with the estimates under
Precautionary loss function, squared error loss function or MLE's. This is true
for both complete and censored samples, see figures 1 and 2. It is immediate to
note that MSE's decrease as sample size increases. On the other hand, for
Classical and Bayesian estimations 1429
complete sample, the Bayes estimates under the LINEX loss function are
underestimation, but MLE's and the Bayes estimates under Precautionary and
squared error loss function are overestimation. For censored sample, the Bayes
estimates under the LINEX loss function, for large loss parameter, are
underestimation, but the rest are overestimation.
2- Table 2 shows that the Bayes estimates under the Precautionary loss
function have the smallest estimated MSE's as compared with the estimates
under LINEX loss function, squared error loss function or MLE's. This is true
for complete samples, see figures 3. For censored samples, the Bayes estimates
under the Squared error loss function have the smallest estimated MSE's as
compared with the rest, however the Bayes estimates under Precautionary and
Squared error loss functions are almost equal, see figures 4. It is immediate to
note that MSE's decrease as sample size increases. On the other hand, for
complete sample, the Bayes estimates under the LINEX, Precautionary and
squared error loss functions are underestimation, but MLE's are overestimation.
For censored sample, all the estimates are overestimation.
3- Table 3 shows that the Bayes estimates under the LINEX loss function
have the smallest estimated MSE's as compared with the estimates under
Precautionary loss function, squared error loss function or MLE's. This is true
for complete samples, see figures 5. For censored samples, the Bayes estimates
under the Precautionary loss function have the smallest estimated MSE's as
compared with the rest, see figures 6. It is immediate to note that MSE's
decrease as sample size increases. On the other hand, for censored sample, the
Bayes estimates under the LINEX, Precautionary and squared error loss
functions are underestimation, but MLE's are overestimation. For complete
sample, all the estimates are overestimation.
References
[1] A. P., Basu and N., Ebrahimi, Bayesian approach to life testing and
reliability estimation using asymmetric loss function. J. Statist. Plann. Infer.,
(1991) 29, 21-31.
[6] R.D., Gupta and D., Kundu, Closeness between the gamma and
generalized exponential distributions. Communications in statistics-theory and
methods, (2003a) 32, 705-722.
[7] R.D., Gupta and D., Kundu, Discriminating between the weibull and
generalized exponential distributions. Computational statistics and data
analysis, (2003b) 43, 179-196.
[8] R.D., Gupta and D., Kundu, Discriminating between the gamma and
generalized exponential distributions. Journal of statistical Computation and
simulation, (2004) 74, 107-122.
[9] D., Kundu, R.D., Gupta. and A., Manglick, Discriminating between the
log-normal and generalized exponential distributions. Journal of statistical
planning and inference (2004) (to appear).
[10] G.S., Mudholkar and A.D., Hutson, The exponentiated weibull family,
some properties and a flood data applications. Communications in statistics-
theory and methods, (1996) 25, 3059-3083.
[11] G.S., Mudholkar and D.K., Srivastava, exponentiated weibull family for
analyzing bathtub failure-rate data. IEEE Transactions on reliability, (1993)
42, 299-302.
[12] G.S., Mudholkar, D.K., Srivastava, and M., Freimer, The exponentiated
weibull family, a reanalysis of the bus-motor failure data. Technometrics,
(1995) 37, 436-445.
[13] M. M., Nassar and F. H., Eissa, Bayesian estimation for the exponentiated
Weibull model. Commun. Statist. Theory Meth, (2004) 33(10), 2343-2362.
[14] J.G., Norstrom, The use of precautionary loss function in risk analysis.
IEEE Trans. Reliab., (1996) 45(3), 400-403.
[17] M.Z., Raqab and M., Ahsanullah, Estimation of the location and scale
parameters of the generalized exponential distributions based on order
statistics. Journal of statistical Computation and simulation, (2001) 69, 109-
124.
[18] R.,Singh, S.K., Singh, U., Singh and G.P., Singh, Bayes Estimator of
Generalized-Exponential Parameters under LINEX loss function using
Lindley's Approximation. Data Scie. Journal, (2008) 7, 65-75.
[19] A. A., Soliman, Inference for reliability and stress-length for a scaled
Burr type X distribution. Commun. Statist. Theory Meth, (2000) 29, 95-107.
[20] G., Zheng, On the fisher information matrix in type-II censored data from
the exponentiated exponential family. Biometrical journal, (2002) 44, 353-
357.