Abstract
Process capability indices (PCIs) aim to quantify the capability of a process of quality characteristic (X) to meet some specifications that are related to a measurable characteristic of its produced items. One such quality characteristic is life time of items. The specifications are determined through the lower specification limit (L), the upper specification limit (U), and the target value (T). Maiti et al. (2010) have proposed a generalized process capability index that is the ratio of proportion of specification conformance to proportion of desired conformance. Bayesian estimation of the index has been considered under squared error loss function. Normal, exponential (nonnormal), and Poisson (discrete) processes have been taken into account. Bayes estimates of the index have been compared with the frequentist counterparts. Data sets have been analyzed.
1. Introduction
The purpose of process capability index (PCI) is to provide a numerical measure on whether a production process is capable of producing items within the specification limits or not. It becomes very popular in assessing the capability of manufacturing process in practice during the past decade. More and more efforts have been devoted to studies and applications of each PCIs. For example, the and indices have been used in Japan and in the US automotive industry. For more information on PCIs, see Hsiang and Taguchi [1], Choi and Owen [2], Pearn et al. [3], Pearn and Kotz [4], Pearn and Chen [5], Mukherjee [6], Yeh and Bhattacharya [7], Borges and Ho [8], Perakis and Xekalaki [9, 10], and Maiti et al. [11].
The usual practice is to estimate these PCIs from data and then judge the capability of the process by these estimates. Most studies on PCIs are based on the traditional frequentist point of view. The main objective of this note is to provide both point and interval estimators of the PCIs given by Maiti et al. [11] from the Bayesian point of view. We believe this effort is well justified since the Bayesian estimation has become one of popular approaches in estimation. In addition, the Bayesian approach has one great advantage over the traditional frequentist approach: the posterior distribution is sometimes very easy to derive, and credible intervals, which are the Bayesian analogue of the classical confidence interval, can be easily obtained either by theoretical derivation or Monte Carlo methods. Lower credible limits (lcls) are constructed. Upper credible limits can also be obtained in a similar manner. However, only the case of lcls is considered as these are of greater interest (due to the fact that large values of PCIs are desirable).
The paper is organized as follows. We give a brief review on the PCIs, , , , and in Section 2. In Sections 3, 4, and 5, we derive the Bayes estimators for (with process median being the process center) with respect to some chosen priors under the assumption of normal, exponential (nonnormal), and Poisson (discrete) distribution, respectively. Simulation results have been reported and discussed in Section 6. In Section 7, data sets have been analyzed to demonstrate the application of the proposed Bayesian procedure. Section 8 concludes.
2. Review of Some Process Capability Indices
The most popular PCIs are , , and . The index is defined as where and are the lower and upper specification limits, respectively, and is the process standard deviation. Note that does not depend on the process mean. The is then introduced to reflect the impact of on the process capability indices. The index is defined as The index was introduced by Chan et al. [12]. This index takes into account the influence of the departure of the process mean from the process target . The index is defined as Maiti et al. [11] suggested a more generalized measure which is directly or indirectly associated with all the previously defined capability indices. The measure is as follows: where is the process yield that is, , is the cumulative distribution function of X, and is the desirable yield that is, , and be the lower and upper desirable limit, respectively. When the process is off centered, then but the proportion of desired conformance is achieved. In that case, the index is as follows: where with being the median of the distribution and the process center is to be located such that that is, , , and . It generally happens that process target is such that ; if , the situation may be described as “generalized asymmetric tolerances” have been described by the term “asymmetric tolerances” when . Under this circumstance, the index is defined as follows:
3. Bayes Estimate of for Normal Process
Let , , , …, be observations from normal distribution with parameter and . Then, the joint distribution of , , , …, is Regarding selection of the the prior distributions, it is advisable to choose conjugate prior, since in this situation, even if prior parameters are unknown in practice, these may be estimated approximately from the likelihood functions as discussed in subsequent sections. When there is no information about the parameter(s) of the distribution, noninformative prior choice is good one.
3.1. Conjugate Prior Distributions
Here we assume that the prior distribution of is of the following form where the given follows normal distribution with mean and variance and follows an inverted gamma distribution of the form: Hence the posterior distribution of is given by where and .
If the process quality characteristic follows normal distribution with mean and variance , then the generalized process capability index is given by Then, the Bayes estimate of under squared error loss is given by Now, and hence
It is to be noted that the Bayes estimate of and the variance depend on the parameters of the prior distribution of and . These parameters could be estimated by means of an empirical Bayes procedure, see Lindley [13] and Awad and Gharraf [14]. Given the random samples , the likelihood function of given is normal density with mean and the likelihood function of is inverted gamma with and . Hence it is proposed to estimate the prior parameters , and and from the samples by , and and , respectively. The variances of these estimators are , and 0, respectively. The expressions of and will be and , respectively.
3.2. Noninformative Prior Distributions
Here we assume that the prior distribution of is noninformative of the form Hence the posterior distribution of is of the form Estimates are to be found out in the same way as in Section 3.1.
4. Bayes Estimate of for Exponential Process
Let , , , …, be observations from exponential distribution with parameter . Then, the joint distribution of , , , …, is
4.1. Conjugate Prior Distributions
Here we assume that the prior distribution of is gamma with parameter that is, the distribution of is given as Hence the posterior distribution of is given as As a process whose distribution can be regarded to be the exponential distribution, the generalized process capability index is given by Then, the Bayes estimate of under squared error loss is given by Now, Again, Thus, If we put that is, if only upper specification limit is given, then with posterior distribution and the Bayes estimate is given by Similarly, if that is, if only lower specification is given, then with posterior distribution and the Bayes estimate is given by
The Bayes estimate of and the variance depend on the parameters of the prior distribution of . Given the random samples , the likelihood function of is gamma density with parameters . Hence it is proposed to estimate the prior parameters and from the samples by and with variances 0 and , respectively. Hence
4.2. Noninformative Prior Distributions
In this subsection, we obtain the Bayes estimator of under the assumption that the parameter is random variable having noninformative prior .
Hence, the Bayes estimator with respect to squared error loss function will be When only upper (lower) specification is to be given, then we will get the expressions substituting .
5. Bayes Estimate of for Poisson Process
Let , , , …, be observations from Poisson distribution with parameter . Then, the joint distribution of , , , …, is
5.1. Conjugate Prior Distributions
Let the prior distribution of is assumed to be gamma with parameter . Then the distribution of is given as Now, the posterior distribution of is given as Now, the process yield is Then, Again,
Here, the Bayes estimate of and the variance depend on the parameters of the prior distribution of . Given the random samples , the likelihood function of is gamma density with parameters . Hence it is proposed to estimate the prior parameters and from the samples by and with variances and 0, respectively. Substituting these in the above expressions, we will have the empirical Bayes estimates.
5.2. Noninformative Prior Distributions
In this subsection, we obtain the Bayes estimator of under the assumption that the parameter is random variable having noninformative prior .
Hence, the Bayes estimator with respect to squared error loss function will be When only upper (lower) specification is to be given, then we will get the expressions substituting .
6. Simulation and Discussion
In this section, we present some results based on the Monte Carlo Simulations to compare the performance of frequentist (maximum likelihood and minimum variance unbiased estimators) as well as the Bayesian method of estimation. All the computations were performed using R-software and Mathematica, and these are available on request from the corresponding author. The maximum likelihood estimator (MLE) and uniformly minimum variance unbiased estimator (UMVUE) and their mean square errors (MSEs) were shown in Maiti et al. [11]. We have performed the Bayes estimators and their MSEs in Tables 1–6. All the results are based on replications.
We represented the average value and the MSE for normal process in Tables 1 and 2. We take the same set up of Maiti et al. [11] to make comparable with the Bayesian approach. We take , for two choices of (L, U) as and , and for sample of sizes . We generate observations from normal distributions with choices of , and . First column of Tables 1 and 2 shows the values of and the corresponding . Remaining columns show average , and its MSE, for the above-mentioned sample sizes. It is observed that in almost all the cases, MSEs of in the Bayesian set up using the empirical Bayes procedure of the prior parameters are larger than those obtained in the frequentist approach. overestimates the true in general. Therefore, this empirical Bayes estimate is not so encouraging compared to maximum likelihood estimator or minimum variance unbiased estimator.
We represented the average value and the MSE for exponential process in Tables 3 and 4. We simulate observations from the exponential distribution with rate . We take , and 1.0. From Tables 3 and 4, we find that for , the empirical Bayes estimate of gives better result than the ML estimate of in MSE sence, but for , it reverses. As soon as the mean quality characteristic gets larger (when ), the empirical Bayes estimate becomes better in MSE sense and hence, it is recommended to use it. For smaller mean quality characteristic, the use of UMVUE of is fair even though it is, to some extent, computation intensive.
We simulate observations from Poisson distribution with mean . We take , and 10. From Tables 5 and 6, we find that for and 4, the UMVUE of gives better result than the empirical Bayes estimate of in MSE sense, but for and 10, it is opposite. Here also if mean quality is getting larger and larger, like exponential process, the empirical Bayes estimate is estimated efficiently. So, it is advisable to use the empirical Bayes estimate of when mean quality characteristic is large, but for smaller mean, use of UMVUE of is a fair one.
It is expected that when there is prior information regarding parameters, the performance of the Bayes estimates would be better than their traditional frequentist counterpart. But here we choose empirical estimate of parameters following the approach of Lindley [13] and Awad and Gharraf [14]. Since it is an empirical approach, it may not perform uniformly better than the frequentist approach that has been reflected in simulation study. The performance is less encouraging in case of normally distributed quality characteristic whereas it performs better when the underlying distribution is exponential with larger mean and also performs better when the quality characteristic distribution is Poisson with a larger mean.
7. Data Analysis
This section is devoted for demonstrating inferential aspect of , by analyzing some data sets. We choose two data sets fit approximately exponential and Poisson distribution, respectively.
(a) For demonstration purpose, we consider here the data that represent the number of miles to first and succeeding major motor failures of 191 buses (cf. Davis [15]) operated by a large city bus company. Failures were either abrupt, in which some part broke and the motor would not run or, by definition, when the maximum power produced, as measured by a dynamo meter, fell below a fixed percentage of the normal rated value. Failures of motor accessories which could be easily replaced were not included in these data. The bus motor failures are compared with exponential distribution, and observed chi-square index has been calculated as 3.40 with value 0.32.
Here, we assume that the upper specification limit (U) and lower specification limit (L) are 75 and 15, respectively. Sample size, , sample mean .
Then, we find out the MLE, MVUE and the Bayes estimate of the index as , , and , respectively. And lower confidence limit (lcl) of the Bayes estimate is 0.563874661.
Now, if we consider the case that only upper specification limit (UCL) has been given, then the MLE and MVUE of the index are and , respectively. And lower confidence limit (lcl) for the index is given as 0.872705746 (cf. Maiti and Saha [16]). Here, we also find out the Bayes estimate of the index as and the corresponding lcl, given as 1.039383292
On the other hand, if we consider the case that only lower specification limit has been given, then the MLE and MVUE of are and , respectively. And lower confidence limit (lcl) for the index is given by 0.635114903. In this case, the Bayes estimate and corresponding lcl of the index are and 0.629801423, respectively.
(b) Data on dates of repair calls on 15 hand electric drill motors are taken from Davis [15]. Mean number of days between failures for each drill was used as a milepost and frequency distribution compared with the theoretical Poisson distribution, and observed chi-square index has been calculated as 38 with value 0.16. Here, we assume that the upper specification limit (U) and lower specification limit (L) are 3 and 1, respectively. Here sample size, and sample mean . Then, the MLE, MVUE, and Bayes estimate of are , , and , respectively. And lower confidence limit (lcl) of the Bayes estimate is 0.659472.
Now, if we consider the case in which only upper specification limit (UCL) has been given, then the MLE of the index is and the MVUE of the index is . 95% lower confidence limit (lcl) for the index Cpyu is given as 1.024565558. Here the Bayes estimate and corresponding 95% lcl are and 1.077885192, respectively.
On the other hand, if we consider the case that only lower specification limit has been given, then the MLE of the index and the MVUE is . 95% lower confidence limit (lcl) for the index is given as 0.448874947. In this case, we also find out the Bayes estimate and 95% lcl of the corresponding index, which are given as and 0.626240727, respectively.
8. Concluding Remark
In this paper, the Bayesian inference aspects of generalized PCI (cf. Maiti et al. [11]) have been presented. The Bayes estimates of the generalized process capability index have been studied for normal, exponential (nonnormal), and Poisson (discrete) processes. The empirical Bayes estimation procedure has been discussed when parameters of the prior distribution are unknown. The Bayes estimates have been compared with their most frequent counterpart, and situations have been mentioned when the Bayes estimates are better through simulation study. Because of its appealing features, examining its potential use in other types of processes often arising in connection with applications would be of practical importance. Other loss functions can be used to find out the estimates in similar fashion.
Acknowledgment
The authors would like to thank the referee for a very careful reading of the paper and making a number of nice suggestions, which improved the earlier version.