See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/271749566

Estimating process capability index C pk : classical approach versus Bayesian

approach

Article in Journal of Statistical Computation and Simulation · December 2014

DOI: 10.1080/00949655.2014.914211

CITATIONS READS

28 242

3 authors, including:

Chin Chieh Wu Chia Huang Wu

National Central University National Yang Ming Chiao Tung University
49 PUBLICATIONS 689 CITATIONS 81 PUBLICATIONS 1,274 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Chin Chieh Wu on 22 June 2016.

The user has requested enhancement of the downloaded file.

 Journal of Statistical Computation and Simulation

ISSN: 0094-9655 (Print) 1563-5163 (Online) Journal homepage: http://www.tandfonline.com/loi/gscs20

Estimating process capability index Cpk: classical

approach versus Bayesian approach

Wen Lea Pearn, Chin Chieh Wu & Chia Huang Wu

To cite this article: Wen Lea Pearn, Chin Chieh Wu & Chia Huang Wu (2015) Estimating process
capability index Cpk: classical approach versus Bayesian approach, Journal of Statistical
Computation and Simulation, 85:10, 2007-2021, DOI: 10.1080/00949655.2014.914211

To link to this article: http://dx.doi.org/10.1080/00949655.2014.914211

Published online: 02 May 2014.

Submit your article to this journal

Article views: 182

View related articles

View Crossmark data

Citing articles: 1 View citing articles

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=gscs20

Download by: [42.73.75.84] Date: 18 April 2016, At: 19:09

 Journal of Statistical Computation and Simulation, 2015
Vol. 85, No. 10, 2007–2021, http://dx.doi.org/10.1080/00949655.2014.914211

Estimating process capability index Cpk : classical approach

versus Bayesian approach
Wen Lea Pearna , Chin Chieh Wub and Chia Huang Wua∗
a Department of Industrial Engineering & Management, National Chiao Tung University, 1001 University
Road, Hsinchu, Taiwan 300, ROC; b Institute of Statistics, National Chiao Tung University, 1001
University Road, Hsinchu, Taiwan 300, ROC
Downloaded by [42.73.75.84] at 19:09 18 April 2016

(Received 4 December 2013; accepted 8 April 2014)

Process capability index Cpk has been the most popular one used in the manufacturing industry dealing
with problems of measuring reproduction capability of processes to enhance product development with
very low fraction of defectives. In the manufacturing industry, lower confidence bound (LCB) estimates the
minimum process capability providing pivotal information for quality engineers to monitoring the process
and assessing process performance for quality assurance. The main objective of this paper is to compare
and contrast the LCBs on Cpk using two approaches, Classical method and Bayesian method.

Keywords: Bayesian method; credible interval; Classical method; lower confidence bound; process
capability index

1. Introduction

Process capability indices (PCIs), Cp , Cpk , Cpm and Cpmk have been proposed in the manufac-
turing industry and the service industry to provide numerical measures on whether a process is
capable of reproducing items within specification limits preset in the factory (see [1–4]). Those
indices essentially compare the predefined manufacturing specifications with the actual production
performance of the investigated quality characteristics and are defined as follows:
 
USL − LSL USL − μ μ − LSL
Cp = , Cpk = min , ,
6σ 3σ 3σ
 
USL − LSL USL − μ μ − LSL
Cpm =  , Cpmk = min  ,  ,
6 σ 2 + (μ − T )2 3 σ 2 + (μ − T )2 3 σ 2 + (μ − T )2

where USL and LSL are the upper and lower specification limits, respectively. μ is the process
mean, σ is the process standard deviation (process variation) and T is the target value. The index
Cp considers the process variability relative to the manufacturing tolerance, reflecting product
quality consistency and potential process capability. The index Cpk takes the magnitude of process
variance as well as process departure from target value and has been regarded as a yield-based
index since it provides lower bounds on process yield. The index Cpm emphasizes on measuring

∗ Corresponding author. Email: [email protected]

© 2014 Taylor & Francis

 2008 W. Lea Pearn et al.

the ability of the process to cluster around the target, which therefore reflects the degrees of
process targeting. The index Cpmk which inherits the merits of both indices is constructed from
combining the modifications to Cp that produces Cpk and Cpm .
Process capability index Cpk has been the most popular one used in the manufacturing industry
dealing with problems of measuring reproduction capability of processes to enhance product
development with very low fraction of defectives. The index Cpk provides a lower bound on
the process yield, a widely recognized criterion for measuring process capability. In practice,
the process mean μ and process variance σ 2 are unknown. Sample data are collected to estimate the
value of Cpk and thus a great degree of uncertainty is introduced into capability analysis. This
is highly unreliable since the estimated index values are random variables with sampling errors.
A reliable approach is establishing the lower confidence bound (LCB) on Cpk to meet at least a
certain degree of the process capability. In other words, we can claim that the LCB contains the
true index value with a reasonable confidence level. Most research efforts on process capability
analysis have used the traditional Classical method. However, the sampling distribution of the
estimated index is usually complicated so that the interval estimation becomes very difficult.
Downloaded by [42.73.75.84] at 19:09 18 April 2016

Bayesian statistical techniques are alternative to the Classical method. The Bayesian approach
specifies a prior distribution for the investigated parameter, to obtain the posterior distribution for
the parameter, then infers about the parameter using its posterior distribution given the sample data.
Similarly, a lower credible bound analogous to the confidence interval on Cpk can be established on
the basis of the posterior distribution. It is not difficult to obtain the posterior distribution when a
prior distribution is set, even in the case where the form of the posterior distribution is complicated.
One can always use numerical computation techniques to obtain an accurate estimate.
In the manufacturing industry, LCB estimates the minimum process capability and provides
pivotal information to quality engineers on monitoring the process and assessing process perfor-
mance for quality assurance. The main objective of this paper is to compare and contrast the lower
confidence and credible bounds on Cpk using two approaches, Classical method and Bayesian
method. Throughout this paper, it is assumed that the process measurements are independent
and identically normally distributed and the process is under statistical control. The rest of the
paper is organized as follows. Section 2 briefly summarizes the Classical method. The Bayesian
method for lower credible bound on the process capability index Cpk is given in Section 3. The
comparisons of the LCB evaluated by Classical method and the lower credible bound obtained
by Bayesian method are in Section 4. Finally, some conclusions are given in Section 5.

2. Classical approach

In this section, we review the estimations of Cpk index and various Classical approaches for
establishing the LCB on Cpk .

2.1. Sampling distribution of Ĉpk

The Cpk index can be defined as

d − |μ − m|
Cpk = ,
3σ
where d = (USL − LSL)/2 is the half length of the specification interval, m = (USL + LSL)/2
is the midpoint between the lower and the upper specification limits. The natural estimator Ĉpk
defined as
d − |x̄ − m|
Ĉpk =
3s
 Journal of Statistical Computation and Simulation 2009

can be obtained by replacing the process mean μ and the process standard deviation σ by their
sample estimators
n 
n
i=1 (xi − x̄)
2
i=1 xi
x̄ = and s = .
n n−1

We note that the process must be under statistical control in order to produce a reliable estimate
of process capability. Under the assumption of normality, Kotz and Johnson [5] obtained the rth
moment including the mean and the variance of Ĉpk . Pearn and Shu [6] defined K = (n − 1)s2 /σ 2 ,
Z = n1/2 (X̄ − m)/σ and Y = |Z|. Then, the estimator Ĉpk can be rewritten as
√ √
n − 1(3Cp n − Y )
Ĉpk = √ .
3 nK
Downloaded by [42.73.75.84] at 19:09 18 April 2016

It is noted that K ∼ xn−1

2
, a chi-square distribution with n − 1 degrees of freedom. Since Z is
distributed as the normal distribution N(n1/2 ξ , 1) with ξ = (μ − m)/σ , Y is distributed as the
folded-normal distribution [7,8]. The folded-normal distribution is taking the absolute value of a
N(μ, σ 2 ) random variable with the probability density function (PDF) as

1 (−x − μ)2 1 (x − μ)2

f (x) = √ exp − + √ exp − , x ≥ 0.
σ 2π 2σ 2 σ 2π 2σ 2

It should be noted that the sampling distribution of Y = |Z| is not a half-normal distribution
since the mean of E(x̄ − m) = μ − m is not always equal to zero. Consequently, substituting
Cp = d/3σ and |ξ | = 3(Cp − Cpk ) into the PDF, the PDF of Y is
√ √ √ √
fY (y) = φ(y − ξ n) + φ(y + ξ n) = φ(y − |ξ | n) + φ(y + |ξ | n)
√ √
= φ[y − 3(Cp − Cpk ) n] + φ[y + 3(Cp − Cpk ) n], y ≥ 0,

where φ(·) is the PDF of the standard normal distribution. Using the integration technique similar
to that presented in Vännman,[9] Pearn and Shu [6] obtained an exact and explicit form of the
cumulative distribution function (CDF) of the natural estimator Ĉpk under the assumption of
normality. The CDFs of Ĉpk are expressed in terms of a mixture of the χ 2 distribution and the
standard normal distribution as following:
√ √
b n
(n − 1)(b n − t)2 √ √
FĈpk (x) = 1 − G [ϕ(t + ξ n) + ϕ(t − ξ n)] dt, x > 0,
0 9nx 2

where b = d/σ , ξ = (μ − m)/σ , G(·) is the CDF of the χ 2 distribution with n − 1 degrees of
freedom.

2.2. Lower confidence bounds on Cpk

On the basis of the CDF of Ĉpk presented above, a 100 γ % LCB on Cpk can be obtained when
the sample size n, the confidence level γ and the value of estimator Ĉpk calculated from sample
data are given. It should be noted that b = d/σ can be expressed as b = d/σ = (d − |μ − m| +
|μ − m|)/σ = 3Cpk + |ξ |. Therefore, to obtain a 100 γ % LCB on Cpk (expressed as Cpk L
), we can
 2010 W. Lea Pearn et al.

solve the following equation:

√ √
L
(3Cpk +|ξ |) n (n − 1)[(3Cpk
L
+ |ξ |) n − t]2 √ √
1− G 2
[φ(t + ξ̂ n) + φ(t − ξ̂ n)] dt = γ .
0 9nĈpk
(1)
Because the parameter ξ in Equation (1) is unknown, Boyles [10] and Spiring [11] recommended
to directly replace ξ by ξ̂ = (x̄ − m)/s observed from the sample data. A LCB on Cpk (expressed
L(BS)
as Cpk ) can be obtained by solving
√ √
L
(3Cpk +|ξ̂ |) n (n − 1)[(3Cpk
L
+ |ξ̂ |) n − t]2 √ √
1− G 2
[ϕ(t + ξ̂ n) + ϕ(t − ξ̂ n)] dt = γ .
0 9nĈpk

Later, Pearn and Shu [6] pointed that such approach introduces additional sampling errors because
of the estimation of ξ . The sampling error makes this approach less reliable in finding the LCBs.
Downloaded by [42.73.75.84] at 19:09 18 April 2016

To eliminate the sampling error of estimating the distribution characteristic parameter ξ , Pearn
and Shu [6] examine the behaviour of the LCB value against the value of parameter ξ . They
proposed that the LCB attains its minimal value at ξ = 1 in all the cases. Hence, for practical
L(PS)
purpose, we may solve Equation (2) to obtain a consecutive LCB on Cpk (expressed as Cpk )
for given Ĉpk , n, and γ by setting ξ = 1 without further estimating the parameter ξ .
√ L(PS) √
(n − 1)[(3Cpk + 1) n − t]2
L(PS)
(3Cpk +1) n √ √
1− G 2
[ϕ(t + n) + φ(t − n)] dt = γ .
0 9nĈpk
(2)
This approach can maintain the obtained LCB satisfying that the risk of making a wrong decision
of overestimating the real process capability no greater than the given Type I error.

3. Bayesian approach

Bayesian statistical techniques are an alternative to the Classical approach. The Bayesian approach
specifies a prior distribution for the interested parameter and obtains the posterior distribution for
the parameter, then infers about the parameter using its posterior distribution given the sample
data. It is not difficult to obtain the posterior distribution when a prior distribution is set, even in
the case where the form of the posterior distribution is complicated. One can always use numerical
computation methods [12,13] to obtain an accurate interval estimate.
Cheng and Spiring [14] proposed a Bayesian procedure for assessing process capability index
Cp . Shiau et al. [15] applied Bayesian approach to process capability index Cpm and the testing of
the process capability index Cpk but under the restriction that the process mean equals the midpoint
of the two specification limits, i.e. μ = m. This is rather an impractical assumption for most factory
applications, since in this case Cpk reduces to Cp . Consequently, Pearn and Wu [16] considered
a Bayesian procedure for the process capability index Cpk without this restriction. The results
are more general and practical for realistic applications. Based on the Bayesian approach under
a non-informative index, Lin et al. [17] developed a procedure to obtain the interval estimation

for the generalized process incapability index Cpp . A Bayesian procedure for judging whether
the process capability is satisfactory was also proposed. Wu and Lin [18] considered the problem
of estimating and testing process capability based on the third-generation capability index Cpmk
from the Bayesian point of view. Recently, Lin et al. [19] extended the results of Wu and Lin [18]
to the case of asymmetric tolerances. Miao et al. chose the conjugate Bayesian approach to
 Journal of Statistical Computation and Simulation 2011

obtain the Bayes estimators for measuring the process capability for multi-batch and low volume
production. Ali and Riaz [20] considered two non-informative priors and conjugate priors to study
the generalized capability indices under various symmetric and asymmetric loss functions. More
researches on PCI from the Bayesian view point can refer to [21–24].

3.1. Posterior distribution of Cpk

The most important problem in Bayesian approach is how to specify an appropriate prior dis-
tribution. If we have no information, we want a prior with minimal influence on the inference.
There are mainly two types of priors, informative and non-informative. The informative prior
distributions summarize the evidence about the parameters concerned from many sources and
often have a considerable impact on the results. On the other hand, with non-informative prior,
Bayesian analysis often leads to the procedures with approximate validity of classical method
while retaining the Bayesian flavor. It allows certain amount of reconciliation between the two
Downloaded by [42.73.75.84] at 19:09 18 April 2016

conflicting paradigms of statistics and provides some mutual justifications.

Cheng and Spiring,[14] Shiau et al.,[15] and Pearn and Wu [16] adopted the following non-
informative reference prior for their researches based on the Bayesian approach.

1
π(μ, σ ) = , −∞ < μ < ∞, 0 < σ < ∞.
σ
Using above prior distribution, the posterior probability density function of (μ, σ ) given the
sample x can be expressed as
√ n
2 n (xi − μ)2
f (μ, σ |x) = √ × σ −(n+1) × exp − i=1 2 ,
2π (α)β α 2σ
∞
where α = (n − 1)/2, β = 2/[(n − 1)s2 ] and (α) = 0 t α−1 e−t dt is the gamma function of α.
Pearn and Wu [16] have derived the posterior probability Pr{Cpk > w|x} in the following:
∞
1 1
Pr(Cpk > w|x) = exp − × { [b1 (y)] + [b2 (y)] − 1} dy, (3)
0 (α)yα+1 y

where
   
√ 1 √ 2 1
b1 (y) = 3 n Ĉpk × −w , b2 (y) = 3 n Ĉpk + δ × −w ,
αy 3 αy

δ = |x̄ − m|/s and (·) is the CDF of the standard normal distribution.
Actually, we note that the derivation of the above posterior probability in Pearn and Wu [16]
is somewhat incomplete. Based on the integration region displayed in Figure 1, the revised and
complete form is presented in the following:

d − |μ − m| 
Pr(Cpk > w|x) = Pr > w x = Pr(0 ≤ |μ − m| < d − 3σ w|x)
3σ

d 
= Pr m − d + 3σ w < μ < m + d − 3σ w, 0 ≤ σ < x
3w 
d/3w m+d−3σ w
= f (μ, σ |x) dμ dσ . (4)
0 m−d+3σ w
 2012 W. Lea Pearn et al.
Downloaded by [42.73.75.84] at 19:09 18 April 2016

Figure 1. The region corresponding to the integration in Equation (4).

Consequently, following the procedure similar to Pearn and Wu,[16] Equation (4) implies
β(d/3w)
1 1
Pr(Cpk > w|x) = exp − × { [b1 (y)] + [b2 (y)] − 1} dy. (5)
0 (α)yα+1 y
L(PW )
From Equation (4), an exact 100 γ % lower credible bound on Cpk (express as Cpk ) based on
the Bayesian method, analogous to confidence interval in classical statistics, can be established
by solving
L(PW ) 2
β(d/3Cpk )
L(PW ) 1 1
γ = Pr(Cpk > Cpk |x) = α+1
exp −
0 (α)y y
× { [b1 (y)] + [b2 (y)] − 1} dy. (6)
L(PW )
It should be note that the lower credible bound Cpk depends on the sample size n, the estimator
δ = |x̄ − m|/s and the estimate Ĉpk .

4. A comparison of two lower bounds

For Cpk index, the procedures to establish the LCB by Classical method and to obtain the lower
credible bound by Bayesian method are presented. In order to compare the performance of the
constructed lower bounds, some discussions are exhibited as follows.
In Section 2, Pearn and Shu [6] proposed that the LCB attains its minimal value at ξ = 1 in all the
L(PS)
cases then the LCB Cpk can be obtained by Equation (2) without further estimating the parameter
ξ . Then, In Section 3, we have proposed that the LCB can be obtained by Equation (6) with the
L(PS) L(PW )
given value of δ. In the following, under various values of lower bounds Cpk = Cpk = w, we
consider the integrations in Equations (2) and (6) with ξ = 1 and δ = 0. That is, for one specific
value of lower bound w, we calculate and plot the confidence levels for these two methods.
Figures 2–5 plot the curves of confidence levels versus the parameter w for n = 10, 20, 30, 40
with the estimate Ĉpk = 2.00.
Figures 2– 5 show that the curves of the Classical method are always right-hand side of the ones
of the Bayesian method. That is, the confidence level of the Classical method is always greater
 Journal of Statistical Computation and Simulation 2013
Downloaded by [42.73.75.84] at 19:09 18 April 2016

Figure 2. Confidence levels for n = 10, Ĉpk = 2.00.

Figure 3. Confidence levels for n = 20, Ĉpk = 2.00.

 2014 W. Lea Pearn et al.
Downloaded by [42.73.75.84] at 19:09 18 April 2016

Figure 4. Confidence levels for n = 30, Ĉpk = 2.00.

Figure 5. Confidence levels for n = 40, Ĉpk = 2.00.

than the one of the Bayesian method for any parameter w. It can be concluded that for the same
confidence level the LCB of the Classical method are always larger than the lower credible bound
of the Bayesian method.
 Journal of Statistical Computation and Simulation 2015

Table 1. Lower confidence bounds

for n = 10, Ĉpk = 2.00, γ = 0.95 and
δ = 0(0.05)1.

Bayesian Classical
δ method method

0.00 1.1235 1.1946

0.05 1.1332 1.1946
0.10 1.1419 1.1946
0.15 1.1496 1.1946
0.20 1.1564 1.1946
0.25 1.1623 1.1946
0.30 1.1675 1.1946
0.35 1.1719 1.1946
0.40 1.1757 1.1946
0.45 1.1789 1.1946
0.50 1.1816 1.1946
0.55 1.1839 1.1946
Downloaded by [42.73.75.84] at 19:09 18 April 2016

0.60 1.1858 1.1946

0.65 1.1873 1.1946
0.70 1.1887 1.1946
0.75 1.1897 1.1946
0.80 1.1906 1.1946
0.85 1.1914 1.1946
0.90 1.1920 1.1946
0.95 1.1925 1.1946
1.00 1.1925 1.1946

Table 2. Lower confidence bounds

for n = 20, Ĉpk = 2.00, γ = 0.95 and
δ = 0(0.05)1.

Bayesian Classical
δ method method

0.00 1.3942 1.4446

0.05 1.4055 1.4446
0.10 1.4147 1.4446
0.15 1.4222 1.4446
0.20 1.4281 1.4446
0.25 1.4327 1.4446
0.30 1.4361 1.4446
0.35 1.4386 1.4446
0.40 1.4404 1.4446
0.45 1.4417 1.4446
0.50 1.4427 1.4446
0.55 1.4433 1.4446
0.60 1.4437 1.4446
0.65 1.4440 1.4446
0.70 1.4442 1.4446
0.75 1.4443 1.4446
0.80 1.4444 1.4446
0.85 1.4445 1.4446
0.90 1.4445 1.4446
0.95 1.4446 1.4446
1.00 1.4446 1.4446

Tables 1– 4 tabulate the LCBs obtained by the Classical method and the lower credible bounds
by the Bayesian methods for the estimate Ĉpk = 2.00, the confidence level γ = 0.95, sample sizes
n = 10, 20,30,40, the values of parameter δ = 0(0.05)1. Since the Classical method sets ξ = 1
 2016 W. Lea Pearn et al.

Table 3. Lower confidence bounds

for n = 30, Ĉpk = 2.00, γ = 0.95,
and δ = 0(0.05)1.

Bayesian Classical
δ method method

0.00 1.5097 1.5508

0.05 1.5213 1.5508
0.10 1.5304 1.5508
0.15 1.5371 1.5508
0.20 1.5419 1.5508
0.25 1.5452 1.5508
0.30 1.5473 1.5508
0.35 1.5487 1.5508
0.40 1.5496 1.5508
0.45 1.5501 1.5508
0.50 1.5504 1.5508
0.55 1.5506 1.5508
Downloaded by [42.73.75.84] at 19:09 18 April 2016

0.60 1.5507 1.5508

0.65 1.5507 1.5508
0.70 1.5508 1.5508
0.75 1.5508 1.5508
0.80 1.5508 1.5508
0.85 1.5508 1.5508
0.90 1.5508 1.5508
0.95 1.5508 1.5508
1.00 1.5508 1.5508

Table 4. Lower confidence bounds

for n = 40, Ĉpk = 2.00, γ = 0.95,
and δ = 0(0.05)1.

Bayesian Classical
δ method method

0.00 1.5773 1.6130

0.05 1.5892 1.6130
0.10 1.5979 1.6130
0.15 1.6038 1.6130
0.20 1.6077 1.6130
0.25 1.6100 1.6130
0.30 1.6114 1.6130
0.35 1.6122 1.6130
0.40 1.6126 1.6130
0.45 1.6128 1.6130
0.50 1.6129 1.6130
0.55 1.6129 1.6130
0.60 1.6130 1.6130
0.65 1.6130 1.6130
0.70 1.6130 1.6130
0.75 1.6130 1.6130
0.80 1.6130 1.6130
0.85 1.6130 1.6130
0.90 1.6130 1.6130
0.95 1.6130 1.6130
1.00 1.6130 1.6130

without further estimating the parameter ξ = |μ − m|/σ , the LCB will be the same for any values
L(PS)
of ξ . For example, if Ĉpk = 2.00 for n = 30, then the LCB Cpk = 1.5508 and the lower credible
L(PW )
bound is Cpk = 1.5504 for parameter δ = 0.50, so we can conclude that Cpk ≥ 1.5508 with
 Journal of Statistical Computation and Simulation 2017
Downloaded by [42.73.75.84] at 19:09 18 April 2016

Figure 6. Plots of the lower confidence bound versus δ for n = 10, Ĉpk = 2.00, γ = 0.95.

Figure 7. Plots of the lower confidence bound versus δ for n = 20, Ĉpk = 2.00, γ = 0.95.
 2018 W. Lea Pearn et al.
Downloaded by [42.73.75.84] at 19:09 18 April 2016

Figure 8. Plots of the lower confidence bound versus δ for n = 30, Ĉpk = 2.00, γ = 0.95.

Figure 9. Plots of the lower confidence bound versus δ for n = 40, Ĉpk = 2.00, γ = 0.95.
 Journal of Statistical Computation and Simulation 2019

Table 5. Various LCB values and the corresponding nonconformities.

Bayesian method Classical method

Ĉpk n LCB NCPPM LCB NCPPM

1.80 10 1.0010 2673.3243 1.0707 1317.6850

1.82 10 1.0133 2366.5671 1.0832 1155.6743
1.84 10 1.0255 2094.4630 1.0956 1013.3277
1.86 10 1.0378 1849.4517 1.1079 888.3170
1.88 10 1.0500 1632.7046 1.1203 776.8894
1.90 10 1.0623 1438.0648 1.1327 678.5561
1.92 10 1.0745 1266.3429 1.1451 591.8981
1.94 10 1.0868 1112.5519 1.1575 515.6350
1.96 10 1.0990 977.2350 1.1698 449.1189
1.98 10 1.1113 856.3744 1.1822 390.2366
2.00 10 1.1235 750.3190 1.1946 338.6314
Downloaded by [42.73.75.84] at 19:09 18 April 2016

Table 6. Various LCB values and the corresponding nonconformities.

Bayesian method Classical method

Ĉpk n LCB NCPPM LCB NCPPM

1.80 20 1.2477 181.7638 1.2971 99.7086

1.82 20 1.2624 152.3545 1.3119 82.9545
1.84 20 1.2770 127.6235 1.3267 68.8863
1.86 20 1.2917 106.5809 1.3414 57.1693
1.88 20 1.3064 88.8433 1.3562 47.2965
1.90 20 1.3210 74.0138 1.3709 39.1058
1.92 20 1.3357 61.4688 1.3857 32.2313
1.94 20 1.3503 51.0210 1.4004 26.5504
1.96 20 1.3650 42.2168 1.4152 21.8009
1.98 20 1.3796 34.9127 1.4299 17.8915
2.00 20 1.3943 28.7814 1.4446 14.6557

Table 7. Various LCB values and the corresponding nonconformities.

Bayesian method Classical method

Ĉpk n LCB NCPPM LCB NCPPM

1.80 30 1.3529 49.3467 1.3932 29.2022

1.82 30 1.3686 40.2918 1.4091 23.6512
1.84 30 1.3842 32.8719 1.4248 19.1653
1.86 30 1.3999 26.7269 1.4406 15.4761
1.88 30 1.4157 21.6845 1.4563 12.4874
1.90 30 1.4313 17.5562 1.4721 10.0405
1.92 30 1.4470 14.1836 1.4878 8.0669
1.94 30 1.4626 11.4503 1.5036 6.4584
1.96 30 1.4783 9.2115 1.5194 5.1594
1.98 30 1.4941 7.3843 1.5351 4.1187
2.00 30 1.5097 5.9235 1.5508 3.2809

95% confidence for Classical method and Cpk ≥ 1.5504 with 95% confidence for the Bayesian
method.
From Tables 1–4, for the lower credible bound based on the Bayesian method, we can observe
that the lower credible bound increases when the value of parameter δ increases. On the other hand,
for fixed value of Cpk , the minimum lower credible bound occurs when the departure |μ − m| ≈
|x̄ − m| = 0 implies the maximization of variance σ 2 . The smaller the value of δ is, the bigger the
 2020 W. Lea Pearn et al.

difference between two lower bounds obtained by Classical method and Bayesian method is. As
the parameter δ increases, the lower credible bound converges to the lower bounds. For example,
in Table 2, the lower bounds are the same for the two methods as δ ≥ 0.95. Furthermore, as the
sample size n becomes larger, smaller value of δ is needed to reach this phenomenon.
Figures 6–9 plot the curves of the 95% lower credible and the confidence bound on Cpk index
versus the estimated parameter δ = |x̄ − m|/s for Ĉpk = 2.00 and n = 10(10)40, respectively. It
is noted that the lower credible bound obtained by Bayesian method is always less than the LCB
established by the Classical method. When the value of parameter δ increases, the lower credible
bound by the Bayesian method converges to the LCB by the Classical method. In practice, if we
have only the value of estimator Ĉpk but without any information about δ, the conservative lower
credible bound by setting δ = 0 should be considered. From Figures 6–9, it is observed that using
the Classical method to calculate the LCB yields a higher result when a fixed value of estimate
Ĉpk is given. That is, the LCB obtained by the Classical method provides better estimation than
the one by the Bayesian method under a fixed confidence level.
In the following, for convenience, with sample sizes of n = 10(10)30 and the values of Ĉpk =
Downloaded by [42.73.75.84] at 19:09 18 April 2016

1.80(0.02)2.00, Tables 5–7 tabulates various values of LCB calculated by Classical method, the
conservative lower credible bound obtained by Bayesian method with parameter δ = 0 and the
corresponding nonconformities in ppm (NCPPM).

5. Conclusions

The index Cpk provides a lower bound on the process yield, a widely recognized criterion for
measuring process capability. In practice, the process mean μ and process variance σ 2 are unknown
to calculate the index value. Thus, sample data are collected to assessing the value of the estimated
index, and a great degree of uncertainty is introduced into capability analysis. For dealing with
this problem, an LCB will be employed to present a measure on the minimum process capability.
In this paper, we compared the LCB on Cpk calculated based on the Classical method proposed
by Pearn and Shu [6] and the lower credible bound based on Cpk Bayesian method proposed by
Pearn and Wu.[16] The results showed that the LCB by the Classical method is always higher
than the one by Bayesian method. Particularly, for the conservative case without the value of
the estimate δ = |x̄ − m|/s, the results of the Classical method are significantly better than the
ones of Bayesian method. Therefore, we suggest the engineering adopting the Classical method
to evaluate the process capability.

References

[1] Kane VE. Process capability indices. J Qual Technol. 1986;18(1):41–52.

[2] Chan LK, Cheng SW, Spiring FA. A new measure of process capability: Cpm . J. Qual Technol. 1998;20:162–175.
[3] Pearn WL, Kotz S, Johnson NL. Distributional and inferential properties of process capability indices. J Qual Technol.
1992;24(4):216–231.
[4] Kotz S, Lovelace CR. Process capability indices in theory and practice. London, UK: Arnold; 1998.
[5] Kotz S, Johnson NL. Process capability indices. London: Chapman & Hall; 1993.
[6] Pearn WL, Shu MH. Manufacturing capability control for multiple power distribution switch processes based on
modified Cpk MPPAC. Microelectron Reliab. 2003;43(6):963–975.
[7] Leone FC, Nelson LS, Nottingham RB. The folded normal distribution. Technometrics. 1961;3:543–550.
[8] Johnson NL. The folded normal distribution: Accuracy of estimation by maximum likelihood. Technometrics.
1962;4(2):249–256.
[9] Vännman K. Distribution and moments in simplified form for a general class of capability indices. Commun Stat-
Theory Methods. 1997;26:159–179.
[10] Boyles RA. The Taguchi capability index. J Qual Technol. 1991;23:17–26.
[11] Spiring FA. A unifying approach to process capability indices. J Qual Technol. 1997;29:49–58.
[12] Berger JO. Statistical decision theory: Foundations, concepts, and methods. New York: Springer; 1980.
 Journal of Statistical Computation and Simulation 2021

[13] Kalos MH, Whitlock PA. Monte Carlo methods. New York: Wiley; 1986.
[14] Cheng SW., Spiring FA. Assessing process capability: a Bayesian approach. IIE Trans. 1989;21: 97—98.
[15] Shiau JH, Chiang CT, Hung HN. A Bayesian procedure for process capability assessment. Qual Reliab Eng Int.
1999;15:369–378.
[16] Pearn WL, Wu CW. Process capability assessment for index Cpk based on Bayesian approach. Metrika.
2005;61(2):221–234.
[17] Lin GH, Jehug WK, Hsieh KH, Wang LW, Lai ST. A Bayesian approach on monitoring process capability. J Stat
Manage Syst. 2007;10(2):241–252.
[18] Wu CW, Lin TY. A Bayesian procedure for assessing process performance based on the third-generation capability
index. J Appl Stat. 2009;36(11):1205–1223.
[19] Lin TY, Wu CW, Chen JC, Chiou YH. Applying Bayesian approach to assess process capability for asymmetric

tolerances based on Cpmk index. Appl Math Model. 2011;35:4473–4489.
[20] Ali S, Riaz M. On the generalized process capability under simple and mixture models. J Appl Stat. 2014;41(4):832–
852.
[21] Lin GH. A Bayesian approach based on multiple samples for measuring process performance with incapability
index. Int J Prod Econ. 2007;106:506–512.
[22] Wu CW. Assessing process capability based on Bayesian approach with subsamples. Eur J Oper Res. 2008;184:207–
228.
[23] Wu CW, Shu MH, Pearn WL, Cheng FT. A comparison of methods forcing loss-based capability index. J Stat Comput
Downloaded by [42.73.75.84] at 19:09 18 April 2016

Sim. 2009;79(9):1129–1141.
[24] Miao R, Zhang XY, Yang D, Zhao YZ, Jiang ZB. A conjugate Bayesian approach for calculating process capability
indices. Expert Syst Appl. 2011;38:8099–8104.

View publication stats