Open access peer-reviewed chapter

Robust Bayesian Estimation

Written By

Ahmed Saadoon Mannaa

Submitted: 11 February 2022 Reviewed: 01 March 2022 Published: 23 November 2022

DOI: 10.5772/intechopen.104090

From the Edited Volume

Bayesian Inference - Recent Advantages

Edited by Niansheng Tang

Chapter metrics overview

64 Chapter Downloads

View Full Metrics

Abstract

Bayes methods in statistical inference are one of the important methods, and most of the research and messages tend to use the Bayes method in the estimation process. The regular Bayes method does not meet this problem, so in this thesis it is possible to verify the existence of prior data conflict by modeling the parameters of the prior distribution and then comparing the standard deviation of the prior distribution with the standard deviation of the posterior distribution, if the value of the standard deviation of the prior distribution is greater than the deviation. The standard distribution for the posterior distribution, it means that there is a problem of prior data conflict. Then we used an approach to solve this problem through a set of prior distributions called this approach by the robust Bayesian method, to identify the behavior of the estimators, two types of failure models were used, the first Weibull distribution to match it with continuous data. The second is a (Binomial) distribution to match the discrete data, the regular Bayes method is compared with the robust Bayesian method by using integrated mean square error (IMSE). In the Weibull distribution, the scale parameter (θ) and the survival function were estimated for two simulation experiments, the first was in the case of prior data unconflict the second was in the case of prior data conflict, so the simulation results showed that the robust Bayes method is the best by using the comparison criterion integrated mean square error (IMSE). On the practical side, real data were collected from Al-Manathira Hospital of the Najaf Health Department for the deaths of heart attack patients for 2018, the time of admission of the patient to the hospital until death was recorded, which is the time Exit where a sample of (15) patients was collected and the test of goodness of fit showed that the data follow a Weibull distribution with two parameters, the robust Bayes method was used to estimate the scale parameter and the survival function. As for the Binomial distribution, the parameter (P) and survival function were estimated for two experiments from the first simulation, which was in the case of prior data unconflict, as for the second experiment, it was in the case of prior data conflict. The simulation results showed that the robust Bayes method is the best by using the comparison criterion (IMSE). On the practical side, real data were collected from Yarmouk Teaching Hospital on breast cancer patients’ mortality from 2010 to 2017, and the test of goodness of fit showed that the data follow a Binomial distribution, the robust Bayes method was used to estimate the parameter (P) and survival function.

Keywords

  • robust Bayesian
  • prior data conflict
  • survival function
  • iLuck model
  • regular Bayesian
  • Weibull distribution
  • binomial distribution

1. Introduction

Both the reliability function and the survival function have the same property, which is the measurement of the life span of a particular system or organism. In systems and equipment, it is called the reliability function, but for the organism, it is called the survival function.

Sometimes, especially in the analysis of survival functions, the failure events are few, so we need to include prior information, which can be used by the Bayes method. When merging the prior information with the observations to obtain the posterior distribution according to the Bayes rule, a problem may appear to us, which is the problem of prior data conflict with views, in regular Bayes method, this problem is not checked and is not addressed and thus unreal estimators are obtained, In this thesis, we will use an approach to address this problem, which is the prior data conflict problem, and thus this method is called the Robust Bayes Procedure.

What is meant by the prior data conflict is when the information of the prior distribution is combined with the distribution of observations, which may cause us this problem, meaning that the data under study are less homogeneous when the information of the prior distribution is combined with it, and thus we obtain unreal estimations without realizing.

In analyzing the problem, the researcher relied on two models of failure, the first model is the Weibull distribution with two parameters to match the continuous data, the second model is the Binomial distribution to match the discrete data to identify the behavior of the capabilities in these two types of data and the appropriateness of the robust methods to deal with the existence of the problem prior data conflict.

Advertisement

2. Robust Bayesian procedure

We also noted earlier that the concept of Bayes theory depends on prior information so that prior information is combined with the distribution of observations according to the Bayes rule, for the purpose of obtaining the posterior distribution, from here we may have a problem, which is a problem that prior data conflict. Whereas prior data they are the default values that are assumed for the parameters of the prior distribution, to find out this problem by updating the parameters of the prior distribution through two methods, namely Expected Conditional or Canonical Exponential Family and provided that the prior distribution is conjugate prior, after obtaining the prior distribution with the updated parameters, we extract the posterior distribution and then extract the standard deviation of the distribution if the value of the standard deviation of the prior distribution is greater than the standard deviation of the posterior distribution, then this means that there is a problem of prior data conflict, Thus, this problem can be addressed through the steps that we will explain later, This method is called the robust Bayesian method [1].

After the default values for the parameters of the prior distribution are chosen, the standard deviation of the prior distribution and the posterior distribution are extracted. If the value of the standard deviation of the prior distribution is greater than the standard deviation of the posterior distribution, this means that there is a problem of prior data conflict and provided that the posterior distribution is conjugate prior, this is the method that will be used in this chapter to verify the prior data conflict.

There are other ways to verify the prior data conflict that we did not used in this chapter, for example (Conflict checks based on relative belief, Connections between the relative belief and score checks and Other approaches to prior-data conflict checking) [2].

Then we move on to addressing the problem of prior data conflict through the proposal presented by (Walter and Augustin; 2009), this is for the purpose of generating a set of prior parameters, in short0=n¯0n¯0xy¯0y¯0, So that this model that generates a set of prior parameters is called generalized iLuck-model and therefore we will get a set of posterior distributions, And then a Bayes estimator is obtained according to the type of loss function used, and thus this method is called the robust Bayesian method [3].

Advertisement

3. Weibull distribution

The Weibull distribution was used in 1951 by researcher Waldi Weibull in many experiments related to the reliability in the mechanical aspect and survival in the human aspect.

The emergence of this distribution, especially in the Second World War, and its wide applications in the field of reliability and life tests, was the focus of the attention of a number of researchers in this field, great in theory and practice.

The probability density function (pdf) for a two-parameter Weibull distribution is in the following form [4]:

ft=βθtβ1etβθ;θ>0,t>0E1

Since:

β: shape parameter.

θ:Scale parameter.

The formula for the (CDF) cumulative function is:

Ft=1etβθE2

The formula for the survival function is:

St=etβθE3

and the formula for moment rth is:

Mr=θrβг1+rβE4

3.1 Bayesian estimation of scale parameter for Weibull distribution

Suppose we have a sample that follows the Weibull distribution shown in Eq. (1) and the appropriate prior distribution for the parameter (θ) is inverse gamma distribution according to the following formula [5]:

fθ/ab=baгaθa1ebθE5

By using the Bayes rule, we get the posterior distribution, as shown below [6]:

fθ/t=ft1t2tn/θβfθ/ab0ft1t2tn/θβfθ/abE6
=i=1ntiβ+ba+nгa+nθa+n1ei=1ntiβ+bθ

Since:

i=1ntiβ=τt
fθ/tInverse Gammaa+nτt+b

Eq. (6) represents the posterior distribution for the parameter (θ), and according to the squared loss function, the Bayesian estimator for the parameter (θ) is the mean of the posterior distribution, as in the following steps [7]:

Eθ/t=θ̂
θ̂=0θfθ/t
=0θτt+ba+nгa+nθa+n1eτt+bθ
=0τt+ba+nгa+nθa+n1+1eτt+bθ

By using the transformation:

Lety=τt+bθθ=τt+by,J=τt+by2
θ̂=0τt+ba+nгa+nτt+bya+neyτt+by2dy
θ̂=τt+ba+nan+1гa+n01ya+n+2eydy
θ̂=τt+bгa+n0ya+n2eydy
θ̂=τt+bгa+nгa+n1
θ̂=τt+ba+n1E7

3.2 Bayesian estimation of survival function for Weibull distribution

From Eq. (6), according to the squared loss function, the Bayesian estimator for the survival function is [8]:

Ŝt=0Stfθ/t
Ŝt=b+τtb+τt+tiβa+nE8

3.3 Checking for prior data conflict for Weibull distribution

Suppose we have a sample that follows a Weibull distribution according to the Eq. (1) and the prior distribution suitable for the scale parameter (θ) is Inverse Gamma because it is a conjugate prior and as in the Eq. (5) [9].

Then we need to update the parameters of the prior distribution so that it is n0>1,y0>0 instead of the parameters (a, b), through two methods we get the prior distribution as shown in the following steps [10]:

The first method: It is the Expected Conditional method, as shown in the following steps:

Eθ/ab=y0=ba1=bn0b=n0y0,n0=a1a=n0+1

Then the prior distribution with the updated parameters can be written in the following form:

fθ/n0y0n0y0n0+1гn0+1θn0+11en0y0θE9

Since:

y0: Pre-guessing the scale parameter.

n0: Pre- guessing the sample size.

The second method: In this method, the prior distribution can be determined with parameters n0y0 through the following steps:

The first step: writing the model in the form of canonical exponential family and as shown below [10]:

fx/θ=axexp(ψ.τxnbψE10
ft/θβ=βni=1ntiβ1ei=1ntiβθ1nlnθE11
ax=βni=1ntiβ1,ψ=1θ,τt=i=1ntiβ,bψ=lnθ

The second step: constructing the prior distribution with parameters n0y0 through the following form:

fψ/n0y0expn0y0.ψbψE12
fψ/n0y0expn0y01θlnθ
=1θ2
fθ/n0y0=fψ/n0y0expn0y0θn0lnθ1θ2
fθ/n0y0θn0+11en0y0θE13

For the purpose of testing whether or not there is a problem of prior data conflict, we extract the standard deviation of the prior distribution and the standard deviation of the posterior distribution.

After we get the prior distribution with the updated parameters, we extract the standard deviation, as in the following steps:

Mr=0θrfθ/n0y0
Mr=n0y0n0+1гn0+10θn0+11+ren0y0θ

By using the transformation:

letz=n0y0θθ=n0y0z,J=n0y0z2
Mr=n0y0n0+1гn0+10n0y0zn011+rezn0y0z2dz
Mr=n0y0n0+1n02+r+1гn0+10zn0+2r2ezdz
Mr=n0y0rгn0+1гn0r+1E14

After that, we extract the posterior distribution through the Bayes rule, as shown in the following steps:

fθ\t=n0y0n0+1гn0+1θn0+11en0y0θβθni=1ntiβ1eτtθ0n0y0n0+1гn0+1θn0+11en0y0θβθni=1ntiβ1eτtθ
=n0y0n0+1гn0+1βni=1ntiβ1θn0+n+11en0y0+τtθn0y0n0+1гn0+1βni=1ntiβ10θn0+n+11en0y0+τtθ
=θn0+n+11en0y0+τtθ0θn0+n+11en0y0+τtθ=θn0+n+11en0y0+τtθгn0+n+1n0y0+τtn0+n+10n0y0+τtn0+n+1гn0+n+1θn0+n+11en0y0+τtθ
=n0y0+τtn0+n+1гn0+n+1θn0+n+11en0y0+τtθE15
fθ\tIGn0+n+1n0y0+τt

After we get the posterior distribution and according to the above equation, we extract the standard deviation, as in the following steps:

Mr=n0y0+τtn0+n+1гn0+n+10θn0+n+11+ren0y0+τtθ

By using the transformation:

letz=n0y0+τtθθ=n0y0+τtz,J=n0y0+τtz2
Mr=n0y0+τtn0+n+1гn0+n+10n0y0+τtzn0n2+rezn0y0+τtz2dz
Mr=n0y0+τtn0+n+1n0n2+r+1гn0+n+10zn0+nrezdz
Mr=n0y0+τtrгn0+n+1гn0+nr+1E16
s.dposterior=n0y0+τt2n0+n2n0+n1E17

The above equation represents the standard deviation of the posterior distribution, after that the comparison is made between the value of the standard deviation of the prior distribution with the standard deviation of the posterior distribution. If the value of the standard deviation of the prior distribution is greater than the value of the standard deviation of the posterior distribution, this means that there is a problem of prior data conflict.

A second way to get the standard deviation of the posterior distribution is through the following steps:

Through the following form which represents the posterior distribution:

fθ/t=n0y0+τtn0+n+1гn0+n+1θn0+n+11en0y0+τtθ

Compensation for:

yn=n0y0+τtn0+n,nn=n0+n

The posterior distribution becomes as follows:

fθ/nnyn=nnynnn+1гnn+1θnn+11ennynθE18

From the above, we conclude that fθ/t=fθ/nnyn this means that the standard deviation of the prior distribution and the standard deviation of the posterior distribution will be according to the following formula:

s.dprior=y02n01E19
s.dposterior=yn2nn1E20

3.4 Standard error for mean

The indicator was employed for the purpose of testing the problem of prior data conflict. If the value of the standard error of the mean of the prior distribution is greater than the standard error of the mean of the posterior distribution, then this means that there is a problem of prior data conflict, and its formula is in the following form:

SEx¯.prior=s.dpriornE21
SEx¯.posterior=s.dposteriornE22

3.5 Address the problem of prior data conflict for Weibull distribution

Although this problem is represented by the prior data conflict problem, we can use a model to address the prior data conflict problem, This is done through the use of a set of prior parameters and according to the proposal presented by (Quaeghebeur and Cooman; 2005) [11], In short 0=n0xy¯0y¯0, Another proposal was submitted (Walter and Augustin; 2009) In order to obtain a set of prior parameters, in brief 0=n¯0n¯0xy¯0y¯0, In general, the model presented to obtain a set of prior parameters is called (Generalized iLuck-Model), after that we get a set of posterior distributions, as shown below [3]:

f1θ/n¯0y¯0=n¯0y¯0n¯0+1гn¯0+1θn¯0+11en¯0y¯0θE23
f2θ/n¯0y¯0=n¯0y¯0n0+1гn¯0+1θn¯0+11en¯0y¯0θE24
f3θ/n¯0y¯0=n¯0y¯0n¯0+1гn¯0+1θn¯0+11en¯0y¯0θE25
f4θ/n¯0y¯0=n¯0y¯0n¯0+1гn¯0+1θn¯0+11en¯0y¯0θE26

Since:

n¯0:Minimum.

n¯0: Maximum.

y¯0: Minimum.

y¯0: Maximum.

The above equations represent a set of the prior distributions obtained through the iLuck-Model, after that we extract the posterior set of distributions according to the following steps:

The first posterior distribution: From Eq. (23) and by using the Bayes rule, we get the first posterior distribution, as in the following equation:

f1θ\t=n¯0y¯0+τtn¯0+n+1гn¯0+n+1θn¯0+n+11en¯0y¯0+τtθE27

The above equation represents the first posterior distribution which is the Inverse Gamma distribution and by taking advantage of the properties of the Inverse Gamma distribution we get the central moments as in the following equation:

Mr=n¯0y¯0+τtrгn¯0+n+1гn¯0+nr+1E28

The second posterior distribution: From Eq. (24) and by using the Bayes rule we get the second posterior distribution as in the following equation:

f2θ\t=n¯0y¯0+τtn¯0+n+1гn¯0+n+1θn¯0+n+11en¯0y¯0+τtθE29

In short:

f2θ\tIGn¯0+n+1n¯0y¯0+τt

The above equation represents the second posterior distribution, which is the Inverse Gamma distribution, and by taking advantage of the properties of the Inverse Gamma distribution we get the central moments as in the following equation:

Mr=n¯0y¯0+τtrгn¯0+n+1гn¯0+nr+1E30

The third posterior distribution: From Eq. (25) and by using the Bayes rule, we get the third posterior distribution as in the following equation:

f3θ\t=n¯0y¯0+τtn¯0+n+1гn¯0+n+1θn¯0+n+11en¯0y¯0+τtθE31
f3θ\tIGn¯0+n+1n¯0y¯0+τt

The above equation represents the third posterior distribution which is the Inverse Gamma distribution and by taking advantage of the properties of the Inverse Gamma distribution we get the central moments as in the following equation:

Mr=n¯0y¯0+τtrгn¯0+n+1гn¯0+nr+1E32

The fourth posterior distribution: From Eq. (26) and by using the Bayes rule we get the fourth posterior distribution as in the following equation:

f4θ\t=n¯0y¯0+τtn¯0+n+1гn¯0+n+1θn¯0+n+11en¯0y¯0+τtθE33
f4θ\tIGn¯0+n+1n¯0y¯0+τt

The above equation represents the fourth posterior distribution which is the Inverse Gamma distribution and by taking advantage of the properties of the Inverse Gamma distribution we get the central moments as in the following equation:

Mr=n¯0y¯0+τtrгn¯0+n+1гn¯0+nr+1E34

After taking the arithmetic mean of the posterior distributions we get the iLuck-Model [10]:

y¯n=loweryn=n¯0y¯0+τtn¯0+nifτ¯ty¯0n¯0y¯0+τtn¯0+nifτ¯t<y¯0E35
y¯n=uperyn=n¯0y¯0+τtn¯0+nifτ¯ty¯0n¯0y¯0+τtn¯0+nifτ¯t>y¯0E36

Eqs. (35), (36) represent a generalized iLuck-Model, a model that represents the lower bound and the model that represents the upper bound is chosen based on the value of τ¯t the estimator we obtain will be in the form of an interval. Therefore we will take the average for that period and from the above the posterior distribution will be in the following form:

fθ/nmym=nmymnm+1гnm+1θnm+11enmymθE37
nm=lowernn+upernn2,ym=loweryn+uperyn2

3.6 Robust Bayesian estimation for scale parameter for Weibull distribution

From Eq. (37) and by using the squared loss function, we get a Bayes estimator for the scale parameter as follows [12]:

Eθ/nmym=θ̂Rob
θ̂Rob=ymE38

3.7 Robust Bayesian estimation for survival function for Weibull distribution

From Eq. (37) and by using the squared loss function we get a Bayes estimator for the survival function as follows [12]:

Ŝt=0Stfθ/t
Ŝt=nmymnm+1гnm+10etβθθnm+11enmymθ
Ŝt=nmymnm+1гnm+10θnm+11enmym+tβθ

By using the transformation:

letz=nmym+tβθθ=nmym+tβz;J=nmym+tβz2
Ŝt=nmymnm+1гnm+10nmym+tβznm+11eznmym+tβz2dz
Ŝt=nmymnm+1гnm+1nmym+tβnm+10znmezdz
ŜRobt=nmymnmym+tβnm+1E39
Advertisement

4. The experimental side of the Weibull distribution

In this section, Weibull distribution data will be generated by using the R program to estimate the scale parameter and the survival function in the prior data conflict, as shown in the following Tables 14:

Modelβ = 2n0Best
y0LowerUpper
LowerUpper25
θnpdf̂pdf̂rob
11.521.5100.0087380.004018Robust Bayesian
200.0045640.002699
400.0021840.001634
2232100.0061280.002333
200.0034210.001845
400.0015390.001094
32.542.5100.0048010.001671
200.0028110.001428
400.0012950.000895

Table 1.

Integrated mean square error (IMSE) of the probability density function (pdf) for the Weibull distribution in the case of prior data conflict.

Modelβ=2n0
Lowerupper
25
y0θnŜtŜrobtBest
LowerUpper
11.521.5100.0053320.002705Robust Bayesian
200.0028940.001948
400.0015700.001263
2232100.0048910.002070
200.0029160.001729
400.0014620.001120
32.542.5100.0053100.002075
200.0028020.001591
400.0014120.001050

Table 2.

Integrated mean square error (IMSE) of the survival function for the Weibull distribution in the case of prior data conflict.

B = 2n0
y0LowerUpper
68
lowerupperθ̂robStandard deviation priorStandard deviation posterior
162016.10352280.4339242.4159

Table 3.

Estimation of the scale parameter of the Weibull distribution in the case of prior data conflict.

B = 2n0
y0LowerUpper
68
loweruppertŜrobt
162010.937227322
20.772418759
30.561586411
70.050817125
100.003284835

Table 4.

Estimation of the survival function of the Weibull distribution in the case of prior data conflict.

Through Tables 1 and 2 the simulation results showed that the robust Bayesian estimator is the best through the comparison standard (IMSE) in the case of prior data conflict. From the above, the robust Bayesian estimator will be applied to the real data to estimate the scale parameter and the survival function.

Advertisement

5. Real data description for Weibull distribution

Real data of a size of (15) for heart attack patients were collected from Al-Manathira General Hospital of the Najaf Health Department for the year 2018, as the time of admission of the patient to the hospital until discharge was recorded and that all of them were in a state of death upon discharge. This data is complete data, ti = (2,1,1,1,1,2,1,3,7,1,2,10,7,1,1), as these times are in days.

Advertisement

6. Binomial distribution

It is one of the discrete distributions in which the experiment can be repeated for (n) times so that the probability of success is (p) and the probability of failure is (1-p), so that the probability density function has the following formula [13]:

fx=nxpx1pnx,x=0,1,,nE40

and the average is:

Ex=npE41

The variance is:

σ2=np1pE42

6.1 Bayesian estimation for parameter p for binomial distribution

Suppose we have a sample that follows the Binomial distribution shown in the Eq. (40) and the appropriate prior distribution is the Beta distribution according to the following formula [13]:

fp/ab=1βαβpα11pβ1;0<p<1E43

To obtain the posterior distribution according to the Bayes rule as follows:

fp/x=fx1x2xn/pfp/αβ01fx1x2xn/pfp/αβdp
=ni=1nxpi=1nx1pni=1nx1βαβpα11pβ101ni=1nxpi=1nx1pni=1nx1βαβpα11pβ1dp
=pi=1nx+α11pni=1nx+β101pi=1nx+α11pni=1nx+β1dp

Since:

i=1nx=s
=1βα+sn+βspα+s11pn+βs1E44

From Eq. (44) and by using the squared loss function we get a Bayes estimator for parameter (p) as follows:

Ep/s=p̂
p̂=α+sn+α+βE45

6.2 Bayesian estimation for survival function for binomial distribution

Through Eq. (44) and by using the squared loss function we get a Bayes estimator for the survival function as follows [11] (Figures 1 and 2):

Figure 1.

It shows the behavior of the survival function by using the robust Bayes estimator, which is decreasing as the value of (t) increases, and this is consistent with the statistical theory [11].

Figure 2.

It shows the behavior of the survival function of a binomial distribution which is decreasing and this is consistent with the statistical theory [13].

Ŝt=Stfp/sdp
fp/s=1βα+sn+βspα+s11pn+βs1
St=j=tnnjpj1pnj
Ŝt=1βα+sn+βsj=tnn!j!nj!01pα+s+j11pn+βs1+njdp
Ŝt=1βα+sn+βsj=tnn!j!nj!01pα+s+j11p2n+βsj1dp
Ŝt=1βα+sn+βsj=tnn!j!nj!01pα+s+j11p2n+βsj1dp

Multiply and divide by:

βα+s+j2n+βsj
Ŝt=1βα+sn+βsj=tnn!j!nj!βα+s+j2n+βsjβα+s+j2n+βsj01pα+s+j1(1p)2n+βsj1dp
Ŝt=1βα+sn+βsj=tnn!j!nj!βα+s+j2n+βsjE46

6.3 Checking of prior data conflict for binomial distribution

Suppose that we have a sample that follows the binomial distribution and in short x ∼ bin (n,p) and as shown in the Eq. (40) then we need to determine the prior distribution through two methods [14]:

The first method: It is the Expected Conditional method we have previously explained this method so we go through the following steps [9]:

Ep/αβ=y0=αα+β
y0=αn0,wheren0=α+β
y0=αn0α=n0y0,n0=n0y0+ββ=n01y0

Then we substitute the parameters α=n0y0, β=n01y0 with the prior distribution to get the prior distribution with the updated parameters:

fp/n0y0=1βn0y0n01y0pn0y011pn01y01E47

The second method: In this method the prior distribution can be determined with the updated parameters through two steps, which are as follows [10]:

The first step: If the model can be written in the form of canonical exponential family, as shown below:

fx/p=axexp(ψ.τxnbψ)fs/p=nsps1pns=nsexplnp1psnln1pax=ns,ψ=lnp1p,τx=s,bψ=ln1pE48

The second step: the prior distribution can be built by using the following model:

fψ/n0y0expn0y0.ψbψ
fψ/n0y0expn0y0lnp1p+ln1p
dp=1p1p
fp/n0y0dp=fψ/n0y0dpdpexpn0y0lnp+n0n0y0)ln(1p1p1pdp

Then the prior distribution with the updated parameters is:

fp/n0y0=1βn0y0n01y0pn0y011pn01y01E49

From the above equation, we extract the standard deviation of the prior distribution, as follows:

Mr=1βn0y0n01y001pn0y0+r11pn01y01dp
=1βn0y0n01y0βn0y0+rn01y0
=гn0y0+n01y0гn0y0гn01y0гn0y0+rгn01y0гn0y0+r+n01y0
=гn0гn0y0+rгn0y0гn0+rE50
s.dprior=y01y0n0+1E51

From Eq. (47) and by using the Bayes rule, we extract the posterior distribution as shown in the following steps:

fp/s=nsps1pns1βn0y0n01y0pn0y011pn01y0101nsps1pns1βn0y0n01y0pn0y011pn01y01dp
=pn0y0+s11pn01y0+ns101pn0y0+s11pn01y0+ns1dp
=1βn0y0+sn01y0+nspn0y0+s11pn01y0+ns1E52

Eq. (52) represents the posterior distribution, after that we extract the standard deviation of the posterior distribution as follows:

Mr=1βn0y0+sn01y0+ns01pn0y0+s+r11pn01y0+ns1dp
=1βn0y0+sn01y0+nsβn0y0+s+rn01y0+ns
=гn0y0+s+n01y0+nsгn0y0+sгn01y0+nsгn0y0+s+rгn01y0+nsгn0y0+s+r+n01y0+ns
=гn0+nгn0y0+s+rгn0+n+rгn0y0+sE53

Since

fp/s=fp/nnyn

Then the standard deviation of the posterior distribution is according to the following formula:

s.dposterior=yn1ynnn+1E54

6.4 Address the problem of prior data conflict for binomial distribution

In the part on the distribution of Weibull, how to solve this problem was explained, so we will enter the following steps [3]:

f1p/s=1βn¯0y¯0+sn¯01y¯0+nspn¯0y¯0+s11pn¯01y¯0+ns1E55
f2p/s=1βn¯0y¯0+sn¯01y¯0+nspn¯0y¯0+s11pn¯01y¯0+ns1E56
f3p/s=1βn¯0y¯0+sn¯01y¯0+nspn¯0y¯0+s11pn¯01y¯0+ns1E57
f4p/s=1βn¯0y¯0+sn¯01y¯0+nspn¯0y¯0+s11pn¯01y¯0+ns1E58

The first posterior distribution: From Eq. (55) and by using the Bayes rule, we get the posterior distribution, as in the following equation:

f1p/s=1βn¯0y¯0+sn¯01y¯0+nspn¯0y¯0+s11pn¯01y¯0+ns1E59

The above equation represents the first posterior distribution which is the Beta distribution and by using the properties of the Beta distribution we get the central moments, as in the following equation:

Mr=гn¯0+nгn¯0y¯0+s+rгn¯0+n+rгn¯0y¯0+sE60

The second posterior distribution: From Eq. (56) and by using the Bayes rule, we get the second posterior distribution, as in the following equation:

f2p/s=1βn¯0y¯0+sn¯01y¯0+nspn¯0y¯0+s11pn¯01y¯0+ns1E61

The above equation represents the second posterior distribution, which is the Beta distribution, and by using the properties of the Beta distribution, we get the central moments, as in the following equation:

Mr=гn¯0+nгn¯0y¯0+s+rгn¯0+n+rгn¯0y¯0+sE62

The third posterior distribution: From Eq. (57) and by using the Bayes rule, we get the third posterior distribution, as in the following equation:

f3p/s=1βn¯0y¯0+sn¯01y¯0+nspn¯0y¯0+s11pn¯01y¯0+ns1E63

The above equation represents the third posterior distribution, which is the Beta distribution, and by using the properties of the Beta distribution, we get the central moments, as in the following equation:

Mr=гn¯0+nгn¯0y¯0+s+rгn¯0+n+rгn¯0y¯0+sE64

The fourth posterior distribution: From Eq. (58) and by using the Bayes rule, we get the fourth posterior distribution, as in the following equation:

f4p/s=1βn¯0y¯0+sn¯01y¯0+nspn¯0y¯0+s11pn¯01y¯0+ns1E65

The above equation represents the fourth posterior distribution, which is the Beta distribution, and by using the properties of the Beta distribution, we get the central moments, as in the following equation:

Mr=гn¯0+nгn¯0y¯0+s+rгn¯0+n+rгn¯0y¯0+sE66

After taking the average of the posterior distributions, we get the iLuck-Model [10]:

y¯n=loweryn=n¯0y¯0+τxn¯0+nifτ¯xy¯0n¯0y¯0+τxn¯0+nifτ¯x<y¯0E67
y¯n=uperyn=n¯0y¯0+τxn¯0+nifτ¯xy¯0n¯0y¯0+τxn¯0+nifτ¯x>y¯0E68

Eqs. (67), (68) represent a generalized iLuck-Model, a model that represents the lower bound and the model that represents the upper bound is chosen based on the value of τ¯x, the estimator we obtain will be in the form of an interval Therefore, we will take the average for that period, and from the above, the posterior distribution model will be in the final form and as in the following equation:

fp/nmym=1βnmymnm1ympnmym11pnm1ym1E69
nm=lowernn+upernn2,ym=loweryn+uperyn2

6.5 Robust Bayesian estimation for parameter for binomial distribution

From Eq. (69) and by using the squared loss function, we get the robust Bayesian estimator for parameter (P) as follows [15]:

Ep/nmym=p̂Rob
p̂Rob=nmymnmym+nm1ymE70

6.6 Robust Bayesian estimation for survival function for binomial distribution

From Eq. (69) and by using the squared loss function, we get a Bayesian estimator for the survival function, as follows [15]:

Ŝx=01Sxfp/sdp
Ŝrobx=1β(nmym,nm1ymj=xnn!j!nj!01pj1pnjpnmym11pnm1ym1dp
Ŝrobx=1β(nmym,nm1ymj=xnn!j!nj!01pnmym+j11pnm1ym+nj1dp

Multiply and divide the equation by:

βnmym+jnm1ym+nj
Ŝrobx=1β(nmym,nm1ymj=xnn!j!nj!βnmym+jnm1ym+njβnmym+jnm1ym+nj01pnmym+j11pnm1ym+nj1dp
ŜRobx=1βnmymnm1ymj=xnn!j!nj!βnmym+jnm1ym+njE71
Advertisement

7. The experimental side of the binomial distribution

In this section, Binomial distribution data will be generated by using the R program to estimate the (P) parameter and the survival function in the prior data conflict, as shown in the following tables:

Through Tables 5 and 6 the simulation results showed that the robust Bayesian estimator is the best through the comparison standard (IMSE) in the case of prior data conflict. From the above, the robust Bayesian estimator will be applied to the real data to estimate the (P) parameter and the survival function for the Binomial distribution.

Modeln0
lowerupper
24
y0PKnpmf̂pmf̂robBest
Lowerupper
10.20.40.45100.0063400.005353Robust Bayesian Estimator
10200.0023340.002143
20400.0008300.000799
20.30.50.55100.0061210.005225
10200.0022070.002041
20400.0008320.000802
30.40.60.65100.0057280.004898
10200.0023690.002206
20400.0007700.000743

Table 5.

The integrated mean square error (IMSE) of the probability mass function (pmf) for the binomial distribution in the case of prior data conflict.

Modeln0
Lowerupper
24
y0pknŜxŜrobxBest
LowerUpper
10.20.40.45100.0343870.027765Robust Bayesian Estimator
10200.0243670.022128
20400.0182680.017509
20.30.50.55100.0345400.028961
10200.0245350.022408
20400.0165490.015884
30.40.60.65100.0320640.026952
10200.0228170.021020
20400.0167450.016075

Table 6.

The integrated mean square error (IMSE) of the survival function for the binomial distribution in the case of prior data conflict.

Advertisement

8. Real data description for the binomial distribution

Mortality data for patients with breast cancer were collected from Yarmouk Teaching Hospital for the period from 2010 to 2017, and the data collected are as follows (Tables 79):

Year20102011201220132014201520162017
Xi34232410

Table 7.

The real data for the binomial distribution.

y0n0
Lowerupper
46
p̂robStandard deviation priorStandard deviation posterior
LowerUpper
0.30.60.6035950.0420.006646

Table 8.

Robust Bayesian estimator for parameter (P) of the binomial distribution.

y0n0
Lowerupper
46
xŜrobx
LowerUpper
0.30.601
10.999
20.996
30.994
40.986

Table 9.

Robust Bayesian estimator for survival function of the binomial distribution.

References

  1. 1. Quaehebeur E, Decoman G. Imprecise Probability Models for Inference in Exponential Symposium on Imprecise Probabilities and their Applications. Pittsburgh, Pennsylvania; 2005
  2. 2. David JN, Michael E. Using Prior Expansions for Prior-Data Conflict Checking. 2020. Available from: https://arxiv.org/pdf/1902.10393
  3. 3. Walter G, Augustin T. Imprecision and prior-data conflict in generalized Bayesian inference. Journal of Statistical Theory and Practice. 2009;3:255-271
  4. 4. Abdulrahman S. Comparing different estimators of three parameters for transmuted Weibull distribution. Global Journal of Pure and Applied Mathematics. 2017;13:5115-5128
  5. 5. Abdulabaas F, Al-Mayali Y, Neama I. A comparison between the Bayesian and the classical estimators of Weibull distribution. Journal of Kufa for Mathematics and Computer. 2013;1(8):21-28
  6. 6. Erlandson F, Adriano K. Bayesian computational methods for estimation of two-parameters Weibull distribution in presence of right-censored data. Chilean Journal of Statistics. 2017;8(2):25-43
  7. 7. Chris B, Ibrahim A, Al OM. Bayesian estimation of two-parameter Weibull distribution using extension of Jeffrey’s prior information with three loss functions. Mathematical Problems in Engineering. 2012;2:13-28
  8. 8. Guure C, Ibrahim N, Adam M. Bayesian parameter and reliability estimate of Weibull failure time distribution. Bulletin of the Malaysian mathematical sciences. Society. 2014;37(3):27
  9. 9. Walter G, Coolen F. Robust Bayesian reliability for complex systems under prior-data conflict. ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems. 2018;4. DOI: 10.1061/AJRUA6.0000974
  10. 10. Walter G. Generalized Bayesian Inference with Sets of Conjugate Priors for Dealing with Prior-Data Conflict Course at Lund University. Lund, Sweden: Lund University; 2015
  11. 11. Al-Nasser AM. An Introduction to Statistical Reliability. Amman, AL-Sharjha, Al-Khabor: UB Group, Ithraa Publishing and Distribution, University Book Shop, Elmia Book Stores; 2009
  12. 12. Al-Doori EA, Mannaa AS. Robust Bayesian estimators for survival function under prior data conflict with practical application in the health side. Indian Journal of Forensic Medicine & Toxicology. 2020;14(2):778-789
  13. 13. Razzaghi M. On the estimation of binomial success probability with zero occurrence in sample. The Journal of Modern Applied Statistical Methods. 2002;17(1):28-40
  14. 14. Walter G, Coolen P. Sets of Priors Reflecting Prior-Data Conflict and Agreement. Switzerland: Springer International Publishing; 2016. DOI: 10.1007/978-3-319-40596-4_14
  15. 15. Al-Doori EA, Mannaa AS. Robust Bayesian estimators for binomial distribution under prior data conflict. Periodicals of Engineering and Natural Sciences. 2020;8(1):284-297

Written By

Ahmed Saadoon Mannaa

Submitted: 11 February 2022 Reviewed: 01 March 2022 Published: 23 November 2022