1. Introduction
A system’s reliability function,
, represents the probability that an object survives longer than time
t, which is also called the survival function. Reliability analyses are mainly devoted to forecasting the probability of survival time for experimental systems, which can help administrators control costs, plan maintenance actions, and assess system availability. Based on the needs of practical applications, numerous scholars have investigated the reliability of different distributions, making use of classical estimators and Bayesian estimators under various loss functions and different priors. The authors of [
1] provided the maximum-likelihood and Bayesian estimates for the three-parameter exponential–Weibull distribution with type-II censored samples. The authors of [
2] analyzed the parameter estimates and reliability features of the one-parameter Lindley distribution using classical and Bayesian techniques with progressive type-II censored samples. A novel parameter estimation method named the E-Bayesian method was introduced in [
3], which was used to estimate the reliability under a binomial distribution. Under a hybrid model with two exponential distributions of certain mixing proportions, the authors of [
4] investigated the unbiased estimations of the reliability function. Furthermore, the equivalent unbiased estimators of the reliability were also derived under the scenario where the negative weighting of the components of the mixture distribution was allowed. The authors of [
5] considered an estimator of the reliability function of the power Lomax distribution in stress–strength models. The authors of [
6] obtained the classical reliability estimations of a multi-component system, in which the stress and strength variables followed the unit-generalized exponential distributions. The Bayesian estimations with Gamma and weighted Lindley priors were also discussed. On account of the situation where the data of torpedo loading tests were limited and included samples without failures, the authors of [
7] proposed a new Bayesian evaluation method based on fused information to estimate the reliability of torpedo loading, for which the lifetime distribution and failure rate prior distribution were assumed to follow exponential and Gamma distributions, respectively. The authors of [
8] calculated the estimators of the reliability for the adaptive, type-II progressive censored data of the unit-Lindley distribution using frequentist and Bayesian methods.
In contrast to the limitations of classical estimation methods, which exclude prior information, Bayesian estimates have been more abundantly researched by scholars in recent years due to their consideration of prior distributions, allowing for the inclusion of historical or subjective information. In Bayesian methods, estimates are obtained using different loss functions. Therefore, one of the main challenges for researchers is choosing the right loss functions according to the specific situation; that is, a well-designed loss function yields a more accurate Bayesian estimator. Many scholars have investigated new loss functions and compared the performance of Bayesian estimators under different loss functions. The authors of [
9] discussed the Bayesian parameter estimator of the Lindley distribution and their corresponding posterior risks separately using seven different loss functions. However, in response to the derivation error of the posterior risks under the entropy loss function in the paper, a correct derivation was given in [
10]. The authors of [
11] proposed a novel analysis method to evaluate potential loss in a process safety field. The potential loss analysis was discussed in detail along with case studies. The authors of [
12] suggested defining a new loss function for use in Bayesian methods to decrease the effects of outliers in crowd behavior analysis. The authors of [
13] proposed a novel loss function named the weighted composite Linex loss function and compared the performance of the estimators under it with other estimators under the Linex loss, weighted Linex loss, and composite Linex loss functions. The authors of [
14] introduced a novel weighted general entropy loss function and compared it with other loss functions. The authors of [
15] derived the parameter estimations of the exponential distribution with complete data, which were comparatively analyzed using maximum-likelihood and Bayesian estimations with the squared error loss, entropy loss, and composite Linex loss functions.
The Burr XII distribution is very flexible as a lifetime distribution, as its hazard rate function can be either monotonic or unimodal, which means that it is much more suitable for modeling different types of lifetime data. Thus, it is more compatible with the precise situations of system reliability analysis and has been widely studied by scholars and applied in practice.
The probability density function (PDF) and cumulative distribution functions are expressed as
where
and
represent the shape parameters.
The hazard rate and reliability functions are expressed as
As a vitally flexible distribution, the Burr XII distribution has many tight connections with other well-known distributions. When , , and , the Burr XII distribution is equivalent to the Pareto, log-logistic, and paralogistic distributions, respectively, with all their scale parameters equal to 1.
The PDF plots corresponding to the selected parameter values are shown in
Figure 1.
Based on
Figure 1, when
, the PDF plots of the Burr XII distribution have a single peak, which is skewed to the right. As
increases, the density plot becomes steeper. In contrast to the case of
, the function curve monotonically decreases when
.
The hazard rate plots with different values of
are shown in
Figure 2, indicating that the hazard rate decreases monotonically for
. However, when
, the hazard rate quickly reaches its peak and then decreases as the argument
x increases. As
x approaches
∞, the hazard rate value converges to a constant value. Moreover, the hazard rate has higher values when
is larger.
The reliability plots are shown in
Figure 3.
The Burr XII distribution has a broad range of applications in lifetime analysis due to its excellent flexibility and practicality as a lifetime distribution. The Burr XII distribution was adopted in [
16] to investigate the reliability of a transit traffic control system with medium loads. Using a progressive type-II censored scheme, the authors of [
17] considered estimates for the Burr XII distribution using E-Bayesian methods. The authors of [
18] discussed the Bayesian estimators of the parameters and reliability under the Burr XII distribution, which were compared with classical maximum-likelihood estimation. For a mixture system of three Burr XII distributions, the authors of [
19] analyzed the Bayesian and maximum-likelihood estimators, along with the elegant closed-form algebraic expressions of Bayesian estimators, as well as the posterior risks. Under the Burr XII distribution, the authors of [
20] introduced an improved adaptive type-II progressive censoring scheme, in which the frequentist and Bayesian inferences were established.
Concentrating on the Burr XII distribution, this study covers the maximum-likelihood estimation (MLE), uniformly minimum-variance unbiased estimation (UMVUE), and Bayesian estimations of the shape parameter and the reliability function. For Bayesian estimations, the squared error loss (SEL), general entropy loss (GEL), and Q-symmetric entropy loss (QEL) functions, together with the new loss function proposed in this paper named the weighted Q-symmetric entropy loss (WQEL), were applied. Our innovation lies in providing UMVUE and Bayesian estimations under the new loss function for the reliability of the Burr XII distribution. The evaluation of the given estimators, especially the Bayesian estimators under the new WQEL function, is also discussed.
The remainder of this paper is structured as follows. In
Section 2, the frequentist methods of the MLE and UMVUE under the Burr XII distribution are investigated.
Section 3 deals with the Bayesian estimators under the three pre-existing loss functions, namely the SEL, GEL, and QEL functions. In
Section 4, a novel loss function named the weighted Q-symmetric entropy loss function is proposed and applied. The Monte Carlo simulation is established in
Section 5, which is conducted to assess the performance of the given estimators. In
Section 6, a remission time data set is used to demonstrate the utility of the discussed estimators.
5. Simulation
Monte Carlo simulation is widely employed in the literature to assess the performance of different models. For example, the authors of [
22] employed it to estimate the reliability of logistics and supply chain networks. This section concentrates on a graphical demonstration of the performance under the applied reliability estimators for comparison and assessment purposes using Monte Carlo simulation. The simulation program was operated using the statistical software R, version 4.2.2. We conducted the simulation using a Windows 10 system with an 11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.42 GHz CPU.
Different estimation methods were evaluated by comparing the mean squared error (MSE), which is defined as
where
N represents the number of iterations.
Given that the
acts on a given value
t for
, we then adopted the Mean Integrated Squared Error (
) to measure the performance of the discussed estimators for the reliability function globally. The
is defined as
where
refers to the particular values of
t selected from the reliability function
and T is the number of the selected values denoted by
.
The simulation involved the following procedures:
- (1)
Before proceeding with the simulation, we initially determined the values of each known parameter in the model. The parameters and appeared in the Burr XII distribution and were set as and . We employed sample sizes of and . In the Bayesian estimators, the hyperparameters of the informative Gamma prior were , and the values of m in the quasi-density prior were . As for the value p, which occurred in the GEL function, it was set as and . We considered for the QEL function, as well as and for the WQEL function. For the choice of , we aimed to verify that the WQEL estimator for was equivalent to the estimator under the SEL function. For the latter, it represented a choice of parameters for the WQEL estimator that exhibited the best performance, which we determined after many attempts.
- (2)
We obtained n samples randomly selected from the Burr XII distribution using the inverse transforming sampling method with the given values of and .
- (3)
We repeated step (2) times and calculated the values. When N was large enough, the effects of random errors occurring each time we obtained the random sample in step (2) diminished, resulting in more informative simulation results.
As observed in
Figure 1, significant variations emerged with small values of
t in the probability density plots. Thus, it should be noted that when computing the
, we took the argument of the reliability function as
. The simulation records are presented in
Table 1,
Table 2,
Table 3,
Table 4 and
Table 5.
The estimates were evaluated with varying values of , n, m, p, q, and z.
From the Monte Carlo simulation results, we can conclude the following:
- (1)
By comparing
Table 2,
Table 3,
Table 4 and
Table 5, it is evident the estimators under the WQEL function for the parameter
exceeded the estimators under other loss functions from a comprehensive point of view due to their smaller MISE values. This result is better visualized in
Figure 5.
- (2)
Based on the results in
Table 2 and
Table 5, a Wilcoxon signed-rank test was conducted for the corresponding estimates under different priors in order to determine whether the estimates under the SEL and the WQEL with
were significantly different. The null hypothesis for the test was that there was no significant difference between the two sets of data. The significance level was set at
. Based on the results in
Table 2 and
Table 5, the
p-value of the Wilcoxon signed-rank test was calculated as
, indicating that the estimates under the SEL and WQEL function with
had similar values. Thus, when
, the estimators under the WQEL function were equivalent to those under the SEL function. This demonstrates the flexibility and advantages of the new WQEL function.
- (3)
As seen in
Table 1, the classical estimators under the Burr XII distribution with parameter
were more efficient than the estimators when
.
- (4)
As seen in
Table 2,
Table 3,
Table 4 and
Table 5, regardless of the loss function used, the estimates under the informative Gamma prior were more efficient than those under the non-informative quasi-density prior, which demonstrates the benefit of an appropriate informative prior distribution for Bayesian estimators.
- (5)
Meanwhile, as seen in
Table 3 and
Table 4, the estimates under the GEL and QEL functions with
yielded better results than those with
, which implies that the GEL and QEL values are sensitive to the values of the parameter
q.
- (6)
In order to provide an intuitive overview of the estimated reliability function under different estimators,
Figure 5 and
Figure 6 illustrate the differences between the Bayesian and classical estimators of the Burr XII distribution with the parameters
and
, assuming a sample size of
. In
Figure 5, the Gamma prior is set as
, while the parameter in the quasi-density prior is set to
. Meanwhile, the estimates under the SEL, GEL, QEL, and WQEL functions are compared for
and
.
In
Figure 5, it can be observed that the reliability functions estimated using the WQEL function are closer to the real value of
, which indicates the superiority of the estimators under the WQEL function when compared with other estimators.
Figure 6 shows that as
t increases, the estimated
progressively approaches the true value of
, whereas the UMVUE of
progressively deviates.
6. Real Data Analysis
A real data set was employed in this section to illustrate the utility of the Burr XII distribution and its corresponding estimators. The data set comprises 128 observations recording the remission time (in years) of bladder cancer patients, as studied in [
23]. An effective and accurate estimator of the reliability functions can greatly assist doctors in understanding the remission time of bladder patients globally and in managing the therapeutic interventions for these patients.
Before estimating the reliability of this data set, we initially calculated the Bayesian information criterion (BIC) for this data set for various commonly used lifetime distributions, including the Burr XII, Gamma, Weibull, Lindley, log-logistic, and linear–exponential distributions, to determine whether it is plausible to fit this data set with the Burr XII distribution.
The PDFs of the considered candidate distributions are structured as follows:
The goodness-of-fit results are shown in
Table 6.
The results in
Table 6 show that the Burr XII distribution should be selected over the other candidate distributions due to its smaller BIC values.
In order to enhance the persuasive power of this study, a Kolmogorov–Smirnov test was conducted as a validation supplement. Let
,
, and
denote the underlying distribution of the data, the estimated Burr XII distribution, and the empirical distribution, respectively. The test statistic for the null hypothesis
versus the alternative hypothesis
is shown below:
The Kolmogorov–Smirnov distance between the estimated Burr XII distribution and the empirical distribution was calculated as . This suggests that under a significance level of , there are no significant discrepancies between the estimated Burr XII distribution and the empirical distribution of the data. Moreover, the p-value of the test was calculated as . A high p-value indicates a high probability of obtaining the observed results, thus we cannot reject the null hypothesis, which verifies the correctness of choosing the Burr XII distribution as the best-fitting distribution.
In order to visualize this data set more intuitively and compare it with the fitted Burr XII distribution, the histogram of the remission time data and the corresponding Burr XII distribution are shown in
Figure 7. This suggests that both the histogram of the real data and the PDF of the underlying distribution exhibit similar patterns.
Table 6 indicates that for the Burr XII distribution, the MLE of
is
. In order to verify that the real data are consistent with the case of a fixed
and unknown
as studied in this paper, we carried out a likelihood ratio test for the integrity of our studies using the Burr XII distribution. Let the null and alternative hypotheses be
and
, respectively. The test statistic is
Considering that approximately obeys a chi-square distribution with 1 degree of freedom, the p-value was calculated as , which indicates that under a significance level of , we cannot reject the null hypothesis. Therefore, we determined that the real data set obeys the Burr XII distribution with and unknown.
Consequently, the classical and Bayesian estimates for selected points of the reliability function of the remission time data are presented in
Table 7. For the Bayesian estimates, we adopted the non-informative quasi-density prior with
, and the parameters in the GEL, QEL, and WQEL functions were considered as
, and
, respectively.
In
Table 7, it is evident that the running time of the estimator under the WQEL function was slightly longer than that of the other estimators, but the difference was not significant. Considering that the accuracy of the estimator under the WQEL function was better than that of the other estimators, such a time cost is acceptable.