APPMATH ISRN Applied Mathematics 2090-5572 International Scholarly Research Network 250393 10.5402/2012/250393 250393 Research Article Bayesian Estimation for Burr Distribution Type III Based on Trimmed Samples Abd-Elfattah A. M. 1 Alharbey A. H. 2 Chyuan S.-W. Gong S. W. Kuhl D. Kyriacou G. Wen X. Yuen K.-V. 1 Department of Mathematical Statistics, Institute of Statistical Studies and Research Cairo University 5 Dr. Ahmed Zewail Street, Dokki, Giza 12613 Egypt kau.edu.sa 2 Faculty of Science King Abdulaziz University, 21589 Jeddah Saudi Arabia kau.edu.sa 2012 18 11 2012 2012 27 06 2012 02 09 2012 2012 Copyright © 2012 A. M. Abd-Elfattah and A. H. Alharbey. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Trimmed samples are widely employed in several areas of statistical practice, especially when some sample values at either or both extremes might have been contaminated. The problem of estimating the parameters of Burr distribution type III based on a trimmed samples and prior information will be considered. In this paper, we study the estimation of unknown parameters based on doubly censored type II. The problem discussed using maximum likelihood method and Bayesian approach to estimate the shape parameters of Burr type III distribution. The numerical illustration requires solving nonlinear equations, therefore, MathCAD (2001) statistical package used to asses these effects numerically.

1. Introduction

Burr  introduced a family of twelve cumulative distribution functions for modeling lifetime data. The two important members of the family are Burr types III and XII. The two important distributions, Burr type III and Burr type XII, are interrelated through simple transformation. Burr type III distribution allows for a wider region for the skewness and kurtosis plane, which covers several distributions including the log-logistic, and the Weibull and Burr type XII distributions.

However, outliers may occur in the data set. Trimmed samples are widely employed in several areas of statistical practice, especially when some sample values at either or both extremes might have been contaminated. The problem of estimating the parameters of Burr distribution type III based on a trimmed sample and prior information is considered.

Many authors discuss different methods of estimation for Burr type XII distribution. Al-Hussaini and Jaheen  and Al-Hussaini et al.  used different techniques for obtaining Bayes estimates of the shape parameters c and k, reliability and failure rate functions based on type II censored samples. Al-Hussaini et al.  obtained the maximum likelihood, uniformly minimum variance unbiased, Bayes and empirical Bayes estimators for the parameter k and reliability function when c is known. Wingo  developing the theory for the ML point estimation of the parameters of the Burr distribution when Type II singly censored sample is at hand. Ali-Mousa and Jaheen  obtained interval estimates of the parameter k and reliability function when c is known, using a Bayesian approach based on Type II censored data. Ali-Mousa  obtained empirical Bayes estimation of the parameter k and the reliability function based on accelerated Type II censored data. Gupta et al.  discusses analysis of failure time data by Burr distribution. Wang et al.  obtained the maximum likelihood estimation of the Burr XII distributions parameter with censored and uncensored data. Hossain and Nath  deal with unweighted least squares estimation of the parameters. They compared the results with the maximum likelihood and maximum product of spacing methods. Ali-Mousa and Jaheen  obtained the maximum likelihood and Bayes estimates for two parameters c,  k and the reliability function r(t) of the Burr Type XII distribution based on progressive type II censored samples.

The objective of this paper is to obtain the estimators of the unknown shape parameters of Burr type III based on doubly censored type II. The problem discussed using maximum likelihood method and Bayesian approach to estimate the shape parameters of Burr type III distribution requires solving nonlinear equations, therefore, numerical study is carried out to asses these effects using MathCAD (2001) statistical package.

2. Burr Type III Distribution

The Burr family of distributions has, in recent years, assumed an important position in the field of life testing because of its uses to fit business failure data, also it includes the well-known exponential and Weibull distributions as special cases. Burr  has suggested this family of distributions by solving the following differential equation: (2.1)dydx=y(1-y)g(x,y),y=F(x), where g(x,y)>0,  0y1,  x in the range over which the solution is being satisfied. By using different forms of g(x,y) and Pearson systems, Burr obtained twelve distribution functions which listed in Burr .

Burr family of distributions includes twelve types of cumulative distribution functions, which yield a variety of density shapes. Many standard theoretical distributions, including the Weibull, exponential logistic, generalized logistic, Gompertz, normal, extreme value, and uniform distributions are special cases or limiting cases of Burr family of distributions. The simple closed form of these distributions has been applied to studies in simulation. Burr  developed the family of Burr distributions as an outgrowth into methods for fitting cumulative frequency functions rather than probability density functions to frequency data. Some of the forms of Burr distributions are related by simple transformation For example, the Burr type III distribution can be obtained from Burr type II distribution by replacing X with ln(X), see Johnson et al. . Similarly, Burr type XII distribution can be derived from Burr type III distribution by replacing X with (1/X), see Burr and Cislak . The Burr type I family is more commonly known as the uniform distribution. The Burr III, IV, V, IX, and XII families have a variety of density shapes. Types III and XII are the simplest functionally and therefore, the two distributions are the most desirable for statistical modeling. The Burr type III distribution is much richer than Burr type XII distribution.

The twelve distributions of Burr are listed in Burr  and Johnson et al. .

Properties of Type III Burr <inline-formula><mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M20"><mml:mo stretchy="false">(</mml:mo><mml:mi>c</mml:mi><mml:mo>,</mml:mo><mml:mi>k</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:math></inline-formula>

The distribution of Burr (c,k) type III is (2.2)F(x)=(1+x-c)-k,x>0, where the parameters c>0 and k>0 are the shape parameters of the distribution. Its density function is (2.3)f(x)=ckx-(c+1)(1+x-c)-(k+1). The rth moment (2.4)E(xr)=μr=ck0xrx-(c+1)(1+x-c)k+1dx=kβ(1-rc,k+rc),r<c, where β(a,b) is the standard beta function.

The expectation of the distribution is obtained as follows.

If r=1 in (2.4) we have: (2.5)μ1=kβ(1-1c,k+1c)=λ1Γ(k),c>1, where λi=Γ(1-i/c)Γ(k+i/c),  i=1,2,,r.

The variance (2.6)var(x)=Γ(k)λ2-λ12Γ2(k),c>2.

The mode (2.7)x={c(k+1)c+1-1}1/c,ck1,c>0,k>0.

The median (2.8)x=1(21/k-1)1/c.

3. Bayesian and Non-Bayesian Estimation Methods

In this section we will obtain the estimation of Burr parameters based on trimmed samples using Bayesian and non Bayesian methods of estimation.

3.1. Non-Bayesian Estimation for Burr Parameters Based on Trimmed Sample

In this section, the estimation problem of Burr type III with two parameters (c,k) under type II double censored data (trimmed samples) are obtained. Some data may not be observed, a known number of observation in an ordered sample are missing at both ends in failure censored experiments, the observations are the smallest r and the largest rs are random then the data collected will be x(r+1)x(r+2)x(n-s) and the likelihood function in double censored type II takes the following form: (3.1)L(x,θ)=n!(r-1)!(n-s)!i=rsf(x(i),θ)[F(x(r),θ)]r-1[1-F(x(s),θ)]n-s,L(x,c,k)=n!(r-1)!(n-s)![1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s*i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]. It is usually easier to use maximize the natural logarithm of the likelihood function rather than the likelihood function itself. Therefore, the logarithm of the likelihood function is (3.2)lnL=lnw+(r-1)ln[1-(1+x(r)-c)-k]-(n-s)kln(1+x(s)-c)-i=rs[(c+1)lnx(i)+(k+1)ln(1+x(i)-c)]+(s-r+1)ln(ck), where w=n!/((r-1)!(n-s)!). The maximum likelihood estimators c^,k^ of c,k are the solutions of the system of equations obtained by letting the first partial derivatives of the total log likelihood with respect to c, k be zero. The systems of equations are as follows: (3.3)lnLk=(r-1)ln(1+x(r)-c)[1+x(r)-c]-k1-(1+x(r)-c)-k-(n-s)ln(1+x(s)-c)+(s-r+1)k-i=rsln(1+x(i)-c),(3.4)lnLc=-(r-1)k(1+x(r)-c)-k-1x(r)-cln(x(r))1-(1+x(r)-c)-k+(n-s)kx(s)-cln(x(s))1+x(s)-c+(s-r+1)c-i=rsln(x(i))-(k+1)i=rsx(i)-cln(x(i))1+x-c(i). From (3.3) the maximum likelihood estimator of k is expressed by (3.5)(n-s)ln(1+x(s)-c^)+i=1sln(1+x(i)-c^)=(s-r+1)k^+(r-1)ln(1+x(r)-c^)[1+x(r)-c^]k^-1. From (3.4) the maximum likelihood estimator of c is expressed by (3.6)(n-s)kx(s)-c^ln(x(s))1+x(s)-c^+(k+1)i=rsx(i)-c^ln(x(i))1+x(i)-c^=(r-1)k(1+x(r)-c^)-k-1x(r)-c^ln(x(r))1-(1+x(r)-c^)-k+(s-r+1)c^+i=rsln(x(i)). Since the closed form solution to nonlinear equations (3.5) and (3.6) is very hard to obtain, Newton-Raphson method is applied for solving the nonlinear equations simultaneously to obtain c^,k^.

3.1.1. Asymptotic Variances Covariance Matrix

The asymptotic variances covariance matrix of the parameters (c and k) is obtained by inverting the Fisher information matrix (3.7)Iij=E[-2Lθiθj],i,j=1,2, where θ1,θ2=c or k. Hence the approximate variance-covariance matrix is given by (3.8)[2lnLc2|c^,k^2lnLck|c^,k^2lnLkc|c^,k^2lnLk2|c^,k^]-1=[Var(c^)Cov(c^,k^)Cov(k^,c^)Var(k^)]. The elements of matrix I are the negative of second derivatives of the natural logarithm of likelihood function defined in (3.2).

The elements of matrix I are given as follows: (3.9)2lnLk2=-(s-r+1)k2-(r-1)[ln(1+x(r)-c)]2(1+x(r)-c)-k[1-(1+x(r)-c)-k]2,(3.10)2lnLc2=(r-1)kln(x(r))[𝒜x(r)cln(x(r))-x(r)-c(k+1)(1+x(r)-c)kx(r)-cln(x(r))-x(r)-cln(x(r))][(1+x(r)c)k+1-(1+x(r)c)]2-(n-s)klnx(s)x(s)-clnx(s)(1+x(s)c)2-(s-r+1)c2+(k+1)i=rsln(x(i))2(1+x(i)-c)2,(3.11)2lnLck=-i=rsx(i)-cln(x(i))1+x(i)-c+(n-s)x(s)-cln(x(s))1+x(s)-c-(r-1)x(r)-clnx(r)1+x(r)-c+(r-1)(1-(1+x(r)-c)-k)x(r)-cln(x(r))(1+x(r)-c)((1-(1+x(r)-c)-k))2-[ln(1+x(r)-c)]k(1+x(r)-c)-k-1+x(r)-clnx(r), where 𝒜 denotes [(1+x(r)-c)k+1-(1+x(r)-c)]. The maximum likelihood estimators c^,   k^ have asymptotic variance-covariance matrix defined by inverting the Fisher information matrix.

3.2. Numerical Illustration

In estimation problem, it is required to study the properties of the derived expressions for the estimators theoretically or analytically. Sometimes it seems very difficult to study the properties of the estimators theoretically because of the complicated formula of the estimators. Consequently, a simulation study will be set up, treating separately the sampling distribution of the estimators. Simulation studies have been performed using MathCAD (2001) for illustrating the theoretical results of the estimation problem. The simulation procedures will described below.

Step 1.

Generate a random sample of size 10, 20, 40, 60, 80, and 100 from Burr type III distribution. The generation of Burr type III distribution is very simple if U has a uniform (0,1) random number, then Y=[U(-1/k)-1](-1/c) follows a Burr type III distribution. In double censored type II, the true parameters selected values are (c=1.2,k=1.5), (c=1.2,k=2), (c=2.5,k=1.5), and (c=2.5,k=2).

Step 2.

Choose censored failure (r=0.6,s=6).

Step 3.

For each sample and for the four sets of parameters distribution were estimated under doubly censored type II.

Step 4.

Newten-Raphson method was used for solving the nonlinear equations for c and k defined in (3.4) and (3.6).

Step 5.

Equations (3.9) and (3.10) were used to obtain the variance covariance matrix of (c^,k^).

Results are tabulated in Table 1.

The maximum likelihood estimator, the standard deviation, and covariance of the Burr distribution with two parameter under doubly censored sample when (a) (c=1.2, k=1.5), (b) (c=1.2, k=2), (c) (c=2.5, k=1.5), (d) (c=2.5, k=2).

n s r c ^ k ^ Co var      (c^,k^)
MLE Standard deviation MLE Standard deviation
5 4.619 3.689 7.841 4.341 58.814
10 7 2 4.534 3.524 4.672 4.172 50.042
9 4.338 3.308 4.722 3.202 22.784

10 6.874 3.874 6.654 4.154 66.589
20 15 2 6.483 2.483 6.439 3.939 64.752
18 6.147 3.147 6.307 3.807 41.058

20 6.498 3.498 6.802 4.302 74.941
40 30 3 5.384 3.384 6.471 3.971 62.603
38 5.287 2.287 5.629 3.129 39.913

30 5.692 2.692 6.794 4.294 62.054
60 45 3 5.542 2.542 6.678 4.178 47.829
58 5.201 2.201 5.489 2.989 16.475

40 5.448 2.448 6.824 4.324 78.475
80 65 5 5.707 2.707 6.671 4.171 66.942
78 5.334 2.334 5.391 2.891 11.313

50 6.689 3.689 6.841 4.341 58.814
100 75 5 6.524 3.524 6.672 4.172 50.042
98 6.308 3.308 5.702 3.202 30.784

n s r c ^ k ^ Co var      (c^,k^)
MLE Standard deviation MLE Standard deviation
5 6.689 3.689 6.841 4.341 58.814
10 7 2 6.524 3.524 6.672 4.172 50.042
9 6.308 3.308 5.702 3.202 30.784

10 6.874 3.874 6.654 4.154 66.589
20 15 2 6.483 3.483 6.439 3.939 64.752
18 6.147 3.147 6.307 3.807 41.058

20 6.498 3.498 6.802 4.302 74.941
40 30 3 6.384 3.384 6.471 3.971 62.603
38 5.287 2.287 5.629 3.129 39.913

30 5.692 2.692 6.794 4.294 62.054
60 45 3 5.542 2.542 6.678 4.178 47.829
58 5.201 2.201 5.489 2.989 16.475

40 5.448 2.448 6.824 4.324 78.475
80 65 5 5.707 2.707 6.671 4.171 66.942
78 5.334 2.334 5.391 2.891 11.313

50 6.689 3.689 6.841 4.341 58.814
100 75 5 6.524 3.524 6.672 4.172 50.042
98 6.308 3.308 5.702 3.202 30.784

n s r c ^ k ^ Co var      (c^,k^)
MLE Standard deviation MLE Standard deviation
5 5.952 2.952 5.159 3.159 69.043
10 7 2 5.712 2.712 4.914 2.914 55.876
9 5.819 2.819 4.249 2.249 38.874

10 5.984 2.984 5.147 3.147 68.497
20 15 2 5.147 2.147 4.271 2.271 24.984
18 5.078 2.078 4.014 2.014 14.872

20 5.618 2.618 5.081 3.081 69.187
40 30 3 5.491 2.491 4.941 2.941 57.078
38 5.107 2.107 4.028 2.028 23.174

30 5.347 2.347 4.627 2.627 70.842
60 45 3 5.148 2.148 4.489 2.489 58.247
58 5.004 2.004 4.309 2.309 41.872

40 5.334 2.334 4.658 2.658 74.194
80 65 5 5.238 2.238 4.748 2.748 60.827
78 4.997 1.997 4.102 2.102 47.145

50 0.447 0.553 1.39 0.19 1.987
100 75 5 0.474 0.526 1.317 0.117 1.854
98 1.218 0.218 1.234 0.034 1.823

n s r c ^ k ^ Co var      (c^,k^)
MLE Standard deviation MLE Standard deviation
5 1.395 0.395 2.717 0.717 42.175
10 7 2 1.501 0.501 2.585 0.585 35.214
9 1.014 0.014 2.461 0.461 18.258

10 1.284 0.284 2.648 0.648 40.954
20 15 2 1.481 0.481 2.441 0.441 33.254
18 1.008 0.008 2.314 0.314 16.418

20 1.658 0.658 2.995 0.995 68.284
40 30 3 1.018 0.018 2.147 0.147 50.258
38 1.001 0.001 1.654 0.346 52.547

30 1.228 0.228 2.954 0.954 68.952
60 45 3 1.109 0.109 2.756 0.756 53.218
58 1.019 0.019 2.341 0.341 40.021

40 1.721 0.721 2.481 0.481 77.248
80 65 5 1.856 0.856 2.441 0.441 80.514
78 1.007 0.007 1.761 0.239 54.951

50 1.395 0.395 2.717 0.717 42.175
100 75 5 1.501 0.501 2.585 0.585 35.214
98 1.014 0.014 2.461 0.461 18.258

From Table 1, we note that the standard deviation decreases when n is increasing. Also, we note that, when c and k increase the standard deviation of the estimators decrease. Similarly, the maximum likelihood estimator of the parameters has the same behaviors when the sample size becomes large and the properties of two parameters c and k at (1.2 and 1.5), respectively is better than the other values.

3.3. Bayesian Analysis for Burr Distribution Type III Based on Trimmed Samples

The Bayesian approach allows both sample and prior information to be incorporated into analysis, which will improve the quality of the inferences. In this section, Bayesian estimators and posterior variance of the shapes parameters are obtained in the case of double censored type II. The prior distribution could be “none informative” such as a flat, uniform distribution, assuming equal probability for the parameter value within a realistic range, and when the prior could be more informative such as a normal distribution or other possible distribution that represents an initial assessment of what is known about the parameter before the collection of data and the analysis. Moreover, a numerical examples are given for illustration study.

3.3.1. Bayesian Estimation for Burr Distribution Type III Parameters Based on Trimmed Sample in Case of Noninformative Prior

The likelihood function takes the form in (3.1). Assumed that the parameters c, k have independent prior distribution and let the noninformative prior (NIP) for c  and k are, respectively, given by: (3.12)u1(c)αc-1c>0,u2(k)αk-1k>0. Consequently, the joint (NIP) will be defined as follows: (3.13)u(c,k)=(ck)-1c,k>0. The joint posterior density functions of c, k under double censored sample type II will be (3.14)g(c,kx)=u(c,k)L(xc,k)λc>0,k>0,g(c,kx)=(ck)-1[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]λ, where λ is the normalized constant defined as follows: (3.15)0(ck)-1[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]dcdk. Now, the marginal posterior of one parameter is obtained by integrating the joint posterior distribution with respect to the other parameter, hence the marginal posterior probability density function of c will be (3.16)g1(ck,x)=c-10k-1[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]dkλ. Similarly integrating the joint posterior (3.14) with respect to k. Then, the marginal posterior of k is (3.17)g2(kc,x)=k-10c-1[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c-1)(1+x(i)-c)-(k+1)]dcλ.

It is well known that under a squared error loss function, the Bayes estimator of the parameter will be its posterior expectation. To obtain the posterior mean and posterior variance, a numerical integration is required. Then, the posterior mean and posterior variance of the shape parameters (c,k) are expressed as follows: (3.18)E(ck,x)=c~,c~=0k-1[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]dcdk,var(c~k,x)=E(c~-c)2,var(c~k,x)=0(c~-c)2×(ck)-1[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]dcdk,E(kc,x)=k~,k~=0c-1[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]dcdk,var(k~c,x)=E(k~-k)2,var(k~c,x)=0(k~-k)2×(ck)-1[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]dcdk. Equations (3.18) are very difficult to be solved exactly so that an iterative procedure is needed to solve these equations numerically. Using the statistical package, MathCAD (2001), the posterior mean and posterior variance of the shapes parameter (c,k) will be obtained.

The posterior mean and the posterior variance of c, k are obtained numerically as described below.

Repeat Steps 1 and 2 in Section 3.2.

Step 3.

Solve the nonlinear equations to obtain posterior variance of the shape parameter (c,k) using (3.18).

Step 4.

The posterior mean and the posterior variance of the estimators for the shape parameters (c,k) for all sample size and for sets of parameter values were obtained.

Numerical results are summarized in Table 2.

The posterior mean and posterior variance of the Burr distribution with two parameter distribution under doubly censored sample when (c=1.2, k=1.5) in case of noninformative prior.

n r S c k
Posterior mean Posterior variance Posterior mean Posterior variance
5 1.017 0.685 0.671 0.765
10 2 7 0.701 0.514 0.569 0.648
9 0.608 0.338 0.408 0.491

10 1.010 0.742 0.948 0.785
20 2 15 0.871 0.537 0.653 0.676
18 0.691 0.338 0.527 0.527

20 0.437 0.674 0.741 0.576
40 3 30 0.288 0.582 0.687 0.477
38 0.119 0.449 0.354 0.345

30 0.609 0.762 0.671 0.758
60 3 45 0.547 0.684 0.507 0.862
58 0.489 0.449 0.407 0.647

40 0.827 0.761 0.718 0.901
80 5 65 0.667 0.895 0.651 0.763
78 0.524 0.612 0.581 0.522

50 0.718 0.761 0.718 0.901
100 5 75 0.651 0.895 0.651 0.763
98 0.581 0.612 0.581 0.522
3.3.2. Bayes Estimator under LINEX Loss Function

Under the assumption that minimal loss function occurs θ~~=θ, the LINEX loss function for θ=θ(c,k) can be expressed as defined by Zellner , the Bayes estimator θ~~ of θ under LINEX loss function is (3.19)θ~~=-1alog(Eθ(exp(-aθ))),θ~~-1alne-aθf(θx)dθ.

Now, in (3.19) put θ=c, then the Bayes estimate c~~ of parameter c relative to the LINEX loss function is (3.20)c~~=-1alne-acg1(ck,x)dc. Set θ=k in (3.19), then the Bayes estimate k~~ of parameter k relative to the LINEX loss function is (3.21)k~~=-1alne-akg2(kc,x)dk. The equations (3.20) to (3.21) are very difficult to obtain their solutions exactly, an iterative procedure is needed to solve these equations numerically using MathCAD (2001) statistical package to obtain posterior variance of shapes parameters (c,k). The numerical procedures will describe as follow.

Repeat Steps 1 and 2 in Section 3.2.

Step 3.

Solve the nonlinear equations to obtain the posterior variance of the shape parameter (c,k) in (3.20) and (3.21).

Step 4.

The posterior mean and the posterior variance of the estimators for the shape parameter (c,k) for all sample size and for sets of parameters were obtained.

Numerical results are summarized in Table 3.

The expectation mean and variance of the Burr distribution with two parameter distribution under doubly censored sample when (c=1.2, k=1.5) under LINEX loss function.

n r S c k
Expectation mean Variance Expectation mean Variance
5 1.698 0.396 2.596 1.631
10 2 7 1.353 0.205 1.700 0.810
9 1.189 0.292 1.663 0.726

10 1.468 0.267 2.552 1.819
20 2 15 1.238 0.166 2.032 1.619
18 1.237 0.178 1.773 1.152

20 1.382 0.294 2.343 1.198
40 3 30 1.092 0.200 1.411 0.848
38 1.022 0.196 0.936 0.847

30 1.075 0.197 1.543 1.237
60 3 45 1.013 0.174 1.018 1.062
58 0.988 0.170 0.935 1.138

40 1.275 0.162 1.007 1.188
80 5 65 0.986 0.182 0.340 1.193
78 0.906 0.151 1.762 2.269

50 0.992 0.761 0.718 0.901
100 5 75 0.745 0.895 0.651 0.763
98 0.604 0.612 0.581 0.522
3.3.3. Bayes Estimator under General Entropy (GE) Loss Function

The Bayes estimator θ~^ under GE loss function is (3.22)θ~^=(Eθ(θ-v))-1/v,θ~^=[θ-vf(θx)dθ]-1/vv>0. Now, in (3.22) put θ~^=c, then the Bayes estimate c~^ of parameter c relative to the GE loss function is (3.23)c~^=[c-vg1(ck,x)dc]-1/v. In (3.22), set θ~^=k, then the Bayes estimate k~^ of parameter k relative to the GE loss function is (3.24)k~^=[k-vg2(kc,x)dk]-1/v.

The equations (3.23) to (3.24) cannot have exact solutions, an iterative procedure is needed to solve these equations numerically using MathCAD (2001) statistical package to obtain posterior variance of shapes parameter. The numerical procedures will be described as follow.

Repeat Steps 1 and 2 in Section 3.2.

Step 3.

Solving the nonlinear Bayesian for posterior variance of the shape parameter (c,k) in (3.23) and (3.24).

Step 4.

The posterior mean and the posterior variance of the estimators for the shape parameter (c,k) for all sample size and for sets of parameters were obtained.

Numerical results are summarized in Table 4.

The expectation mean and variance of the Burr distribution with two parameter distribution under doubly censored sample when (c=1.2, k=1.5) general entropy loss function.

n r S c k
Expectation mean Variance Expectation mean Variance
5 1.641 0.685 3.825 0.765
10 2 7 1.375 0.514 2.614 0.648
9 1.255 0.338 1.661 0.491

10 1.474 0.742 3.423 0.785
20 2 15 1.299 0.537 2.490 0.676
18 1.147 0.338 1.550 0.527

20 1.294 0.674 3.076 0.576
40 3 30 1.087 0.582 1.378 0.477
38 1.016 0.449 1.263 0.345

30 1.032 0.762 1.672 0.758
60 3 45 0.977 0.684 1.173 0.862
58 0.924 0.449 1.038 0.647

40 1.162 0.761 0.860 0.901
80 5 65 1.034 0.895 0.396 0.763
78 0.888 0.612 0.185 0.522

50 0.992 0.761 0.718 0.901
100 5 75 0.745 0.895 0.651 0.763
98 0.604 0.612 0.581 0.522
3.4. Bayesian Estimation for Burr Distribution Type III Parameters Based on Trimmed Sample in Case of Informative Prior (IP)

Assume that the parameters c, k have independent prior distribution and let the informative prior (IP) for the parameter k is given by the gamma density as follow: (3.25)π1(kc)=cα+1Γα+1kαe-kck>0,c>0,α+1>0, and the prior density function of c is the exponential density as follow: (3.26)π2(c)=1βe-c/βc>0,β>0. Then the joint prior density of c and k, given by (3.27)π(c,k)=π1(kc)π2(c),π(c,k)=cα+1βΓα+1kαe-(k+1/β)cc>0,k>0. The joint posterior density function of c, k using double censored sample type II will be (3.28)f(c,kx)=π(c,k)L(xc,k)Ec>0,k>0=cα+1kαe-(k+1/β)c[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]EβΓα+1, where E is the normalized constant equal to (3.29)E=0cα+βΓα+1kαe-(k+1/β)c[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]dcdk. Now, the marginal posterior of one parameter is obtained by integrating the joint posterior distribution with respect to the other parameter, hence the marginal posterior probability density function of c is (3.30)f1(ck,x)=cα+1βΓα+10kαe-(k+1/β)c[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]dkλ. Similarly integrating the joint posterior (3.28) with respect to k the marginal posterior of k is (3.31)f2(kc,x)=kαβΓα+10cα+1e-(k+1/β)c[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]dcλ.

Hence, under a squared error loss function, the Bayes estimator of the parameter will be its posterior expectation. To obtain the posterior mean and posterior variance a numerical integration is required. Then, the posterior mean and posterior variance of the shape parameters (c,k) are expressed as follows: (3.32)E(ck,x)=c~˘,c~˘=1βΓα+10cα+2kαe-(k+1/β)c[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]dcdk,var(c~˘k,x)=E(c~˘-c)2,var(c~˘k,x)=1βΓα+10(c~˘-c)20cα+1kαe-(k+1/β)c[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]dcdk,E(kc,x)=k~˘,k~˘=1βΓα+10cα+1kα+1e-(k+1/β)c[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]dcdk,var(K~˘c,x)=E(K~˘-k)2,var(K~˘c,x)=1βΓα+10(k~˘-k)20cα+1kαe-(k+1/β)c[1-(1+x(r)-c)-k]r-1[(1+x(s)-c)-k]n-s×i=rs[ckx(i)-(c+1)(1+x(i)-c)-(k+1)]dcdk. Nonlinear equations (3.32) are very difficult to be solved exactly an iterative procedure is needed to solve these equations numerically. By using the statistical package, MathCAD (2001), the posterior mean and posterior variance of the shape parameters (c,k) will be obtained.

The posterior mean and the posterior variance of c, k are obtained numerically in the following procedures as follow.

Repeat Steps 1 and 2 in Section 3.2.

Step 3.

Solve the nonlinear Bayesian for posterior variance of the shape parameter (c,k) in (3.32).

Step 4.

The posterior mean and the posterior variance of the estimators for the shape parameter (c,k) for all sample size and for sets of parameters were obtained.

Numerical results are summarized in Table 5.

The posterior mean and posterior variance of the Burr distribution with two parameter under doubly censored sample when (c=1.2, k=1.5) in case of informative prior.

n r s c k
Posterior mean Posterior variance Posteriormean Posterior variance
5 0.495 0.204 1.017 0.719
10 2 7 0.384 0.214 0.701 0.645
9 0.258 0.217 0.608 0.201

10 0.234 0.391 1.010 0.604
20 2 15 0.128 8.215 0.871 0.556
18 0.087 0.319 0.691 0.493

20 0.482 0.339 0.437 0.911
40 3 30 0.352 0.337 0.288 0.425
38 0.241 0.190 0.119 0.307

30 0.852 0.199 0.609 0.827
60 3 45 0.729 0.160 0.547 0.647
58 0.635 0.139 0.489 0.558

40 0.957 1.171 0.827 0.728
80 5 65 0.721 0.206 0.667 0.661
78 0.408 0.078 0.524 0.337

50 0.992 0.761 0.718 0.901
100 5 75 0.745 0.895 0.651 0.763
98 0.604 0.612 0.581 0.522

It is noted that the posterior mean decreases when n is increasing. Similarly the posterior variance of the parameters has the same behaviors when the sample size becomes large.

Acknowledgment

This project was founded by the deanship of scientific research (DSR), King Abdulaziz University, Jeddah under Grant no. (124/130/1432). The authors, therefore, acknowledge with thanks DSR technical and financial support.

Burr I. W. Cumulative frequency functions Annals of Mathematical Statistics 1942 13 215 232 0006644 Al-Hussaini E. K. Jaheen Z. F. Bayesian estimation of the parameters, reliability and failure rate functions of the Burr type XII failure model Journal of Statistical Computation and Simulation 1992 41 1-2 31 40 10.1080/00949659208811389 1276181 AL-Hussani E. K. Ali-Mousa M. A. M. Jaheen Z. E. Approximate Bayes estimators applied to tile Burr model Communications in Statistics 1994 23 1 99 121 Al-Hussaini E. K. Ali-Mousa M. A. M. Jaheen Z. F. Estimation under the Burr type failure model based on censored data: a comparative study Test 1992 1 1 47 60 10.1007/BF02562661 1266128 Wingo D. R. Maximum likelihood estimation of Burr XII distribution parameters under type II censoring Microelectronics Reliability 1993 33 9 1251 1257 Ali Mousa M. A. M. Jaheen Z. F. Interval estimation under the Burr type XII failure model based on censored data Journal of the Egyptian Mathematical Society 1994 2 67 73 1319068 Ali-Mousa M. A. M. Empirical Bayes estimators for the Burr Type XII accelerated life testing model based on Type-2 censored data The Journal of Statistical Computation and Simulation 1995 52 95 103 Gupta P. L. Gupta R. C. Lvin S. J. Analysis of failure time data by Burr distribution Communications in Statistics 1996 25 9 2013 2024 10.1080/03610929608831817 1410146 Wang F. K. Keats J. B. Zimmer W. J. Maximum likelihood estimation of the burr XII parameters with censored and uncensored data Microelectronics Reliability 1996 36 3 359 362 2-s2.0-0030106841 10.1016/0026-2714(95)00077-1 Hossain A. M. Nath S. K. Estimation of parameters in the presence of outliers for a Burr XII distribution Communications in Statistics 1997 26 3 637 652 2-s2.0-0031084390 Ali-Mousa M. A. M. Jaheen Z. F. Statistical inference for the Burr model based on progressively censored data Computers & Mathematics with Applications 2002 43 10-11 1441 1449 10.1016/S0898-1221(02)00110-4 1906365 Johnson N. L. Kotz S. Balakrishnan N. Continuous Univariate Distributions 1995 2 2nd New York, NY, USA John Wiley & Sons 1326603 Burr I. W. Cislak P. J. On a general system of distributions: I. Its curve-shape characteristics. II The sample median Journal of the American Statistical Association 1968 63 627 635 Zellner A. Bayesian estimation and prediction using asymmetric loss functions Journal of the American Statistical Association 1986 81 394 446 451 845882