International Statistical Review (2008), 76, 3, 347-367 doi:10.111 l/j.l751-5823.2008.00060.

Basic Process Capability Indices: An

Expository Review
Mohammed . Anis
SQC and OR Unit, Indian Statistical Institute, Kolkata, India
E-mails: [email protected]; [email protected]

Summary
A review of the four basic process capability indices has been made. The interrelationship among
these indices has been highlighted. Attention has been drawn to their drawbacks. The relation of
these indices to the proportion nonconforming has been dwelt upon and the requirement of the
adequate sample size has been emphasized. Cautionary remarks on the use of these indices in the
case of nonnormal distributions, skewed distributions, and autocorrelated data are also presented.
The effect of measurement error on process capability indices has been dealt with in great detail.

Key words'. Autocorrelation; measurement error; nonnormality; sample size; skewed distributions.

1 Introduction

With the world becoming borderless, at least as far as business is concerned, there is intense
national and international competition amongst business organisations. This competition is
compelling business organisations to manufacture defect-free products. To achieve this objective,
companies have started adopting different strategies like Total Quality Management (TQM)
and Six Sigma throughout their organisations. A part of the philosophy of these strategies
requires the monitoring of the performance of the individual processes. These results are
then compared with those of industry leaders through competitive benchmarking. One metric
popularly used is the Process Capability Index (PCI) (Spiring, 1995). Essentially a PCI measures
the variability of a process relative to its specification limits. Being unitless, these indices permit
comparisons amongst hundreds of processes emanating from a whole range of production
processes, industries, and even countries. Many (large) companies have instituted programmes
that inherently make use of these indices to promote and drive quality improvement programmes
throughout their organisations (Barnett, 1990; Gill, 1990; McCoy, 1991).
Moreover, the incorporation of capability analysis into a company's Six Sigma programme
makes it a particularly important topic for management reporting. Briefly, Six Sigma is a quality
and business improvement methodology that makes heavy use of statistical methods. It began in
Motorola in the 1980s. While originating in manufacturing, it has expanded to financial services,
health care, and even nonprofit organisations. The rapid spread of Six Sigma is due to the fact
that it has delivered significant bottomline results. It is perhaps the largest and most important
statistically based initiative in history. The Six Sigma methodology includes five steps, namely,
definition of the measuring unit critical to quality (CTQ), measurement of the current process

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute. Published by Blackwell Publishing Ltd, 9600 Garsington Road,
Oxford 0X4 2DQ, UK and 350 Main Street, Maiden, MA 02148, USA.
 348 M. . Anis

performance, analysis of the root cause and identification of the solutions, the improvements of
the process quality, and control of the process quality. Recently, a session (Invited Paper Meet
74) was devoted to Six Sigma at the 56th session of the International Statistical Institute, held
in Lisbon during August 22-29, 2007.
Historically, Feigenbaum (1951) and Juran (1951) first proposed 6 as a measure of process
capability. This represented process capability as a measure of the inherent variability of a
process, but is divorced from customer specifications. Juran (1962) overcame this lacuna
by comparing 6 to the tolerance width as a method of determining the need for process
improvement activities. Nevertheless, capability itself was still interpreted separately from
specifications. Finally, Juran & Gryna (1980) proposed the first metric that directly compared
process variability to customer specifications. They proposed a capability ratio:
6 variation
Capability ratio =
Total tolerance
(1)
Like the capability ratio, all PCIs explicitly link process variability to customer specifications.
Thus, they emphasize the suppliers' responsibility to satisfy those specifications. However,
capability indices also have advantages over the capability ratio. They (generally) increase in
value as the process performance improves. Furthermore, they indicate the relative benefits of
improvement in both process location and variability.
However, it is also important to understand their limitations, especially because organisations
are placing a greater emphasis on quality-related measures. One essential prerequisite for the
process of improving quality via capability indices is that the process be in a state of statistical
control. Further, these measures can themselves be distorted and may not accurately indicate
the extent and type of improvements needed. A lack of understanding of process variability
has caused a significant amount of controversy over the use of these indices (Kotz & Johnson,
1993, pp. 1-2). At times, the underlying correlation in a process when coupled with outliers
can mask out-of-control points, and thereby make it appear to be in a state of statistical control.
Ignoring these interactions may make the process appear better or worse than it actually is. This,
in turn, may lead a manager to divert resources to improving processes where the returns are
minimal but ignore areas that can lead to large reductions in variability and significant quality
improvements. In the long run, this lack of understanding will frustrate the managers and affect
the overall profitability of the organisation.
There is a large body of literature dealing with PCIs. Mention may be made of the books by
Kotz & Johnson (1993), Kotz & Lovelace (1998), Wheeler (1999), Bothe (2001), and Pearn &
Kotz (2006). Papers relating to PCI have appeared in journals of statistics, management science,
industrial engineering, quality and TQM. Spiring et al (2003) give a bibliography, whereas
Kotz & Johnson (2002) provide a largely theoretical overview. However, most of these works
have a theoretical flavour directed towards researchers. There is a paucity of literature, which
unmasks the intricacies of these indices. This paper endeavours to fill this gap. In this paper, an
attempt has been made to study the effectiveness of the existing PCIs in relation to the decision
making process of the users and advocate caution. It is hoped that this expository paper will be
useful to the practicing engineer, the management personnel, and serious students of theoretical
statistics.
The rest of the paper is organised as follows. Section 2 briefly dwells on the need for indices.
Section 3 introduces the indices and their motivation; develops the interrelationships among
them and discusses their estimators. Section 4 reviews the basics regarding index interpretation
and process improvements and attempts to highlight the drawbacks. Section 5 discusses the
indices in relation to the number of nonconforming (NC) products produced and the sample size
required for any scientifically meaningful study. Section 6 considers the effect of nonnormality,

International Statistical Review (2008), 76, 3, 347-367

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 Basic Process Capability Indices 349

correlation, and asymmetry in the calculation of PCIs. Section 7 deals with the effect of
measurement error. Section 8 concludes the paper with a discussion.

2 The Need for an Index

The behaviour of a process is often described by a probability distribution. In order to

assess its adequacy, the hypothesised distribution has to be compared with the corresponding
specifications. A PCI attempts to summarise the process performance and hence is a function
of the process distribution and the corresponding specification.
Shewhart (1931), Juran (1951), and Gryna (1988) discussed the use of process capability
information to determine specification limits. They recommended that the tolerance width
should not be tighter than 6 . Feigenbaum (1951) and Juran (1951) referred to the use of process
capability information to assign jobs to machines. Kane (1986) described six application areas
for capability indices, which are as follows. The indices help in the prevention of NC products
by establishing a benchmark capability. Being dimensionless, they facilitate communication
between engineering and manufacturing departments and between manufacturers and suppliers.
They aid in establishing the priority areas for process improvement and continuous improvement.
The indices also provide information on the location and variability of a process and hence
suggest the road map for process improvement. Finally, the indices can be used in audits to help
establish the problem areas. Crain (1993), Chou (1994), and Schneider & Pruett (1995), among
others, discussed the use of these indices in customer-supplier settings. Bulba & Ho (2006) used
methods to obtain approximate confidence intervals for the variance of the output characteristic
to make inference on capability indices when the variable of interest is not an observable one
but is a function of a set of independent input variables. Bothe (2006) described a method for
assessing the capability of a process to locate hole centres within a circular tolerance zone.
These give a partial listing of the applications of a PCI.
The important objectives of a PCI have already been discussed by Tsui (1997). Suffice it
to say that a PCI should be informative enough to guide the users in their decision problems
adequately and unambiguously. Another desirable feature of a PCI is that its numerical value
should increase when the variability decreases.

3 Process Capability Indices

It is assumed that there is only one quality characteristic (say X) of interest. Let U and L be the
upper and lower specification limits, and let be the "target value" and define M = (U + L)/2,
and D = (U ? L)/2. Let the underlying process mean and standard deviation be denoted by
and , respectively. Unless otherwise stated, we shall assume that the quality characteristic is
normally distributed.
Depending upon the situation, the specification forX can be one of the following types:

(a) Unilateral (one-sided, with target not specified)

(i) Only U
(ii) Only I
(b) Bilateral (two-sided, with target specified)
(i) Centred target, that is, = M
(ii) Off-centred target, that is, M

International Statistical Review (2008), 76, 3, 347-367

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 350 M. . Anis

3.1 Definition of the Indices and Their Motivations

Assuming the quality characteristic to be normally distributed with mean and variance 2,
the process yield is, in general, given by

% Yield = 100

where denotes the standard normal cumulative distribution function. The index Cp measures
only the distribution spread (process consistency/precision), which only reflects the consistency
of the product quality characteristic. The yield-based index Cpk provides lower bounds on process
yield by taking the process location into consideration, which offsets some of the weaknesses in
Cp, but can fail to distinguish between on-target and off-target processes. The index Cpm takes
the proximity of process mean from the target value into account, which is more sensitive to
process departure than Cpk. The index Cpmk provides a greater level of quality assurance with
respect to process yield and process loss to the customers than the Cpk and Cpm indices.

3.1.1 The most basic indices

We begin with the most basic index.

(a) Unilateral with only U:

U
^Cpu ?
?
3 ' (2)
provided that < U
(b) Unilateral with only L:

Cpi = 3 (3)
provided that L <

(c) Bilateral with = M:

U-L
Cp = 6 ' (4)

(d) Bilateral with M:

U-T T-L
C* = min
3 3 (5)

These indices are from Kane (1986). Cp is the most basic capability index and is said to be a
first-generation index. The observant reader will immediately notice that Cp is the reciprocal of
Juran & Gryna's capability ratio (defined in Section 1).

International Statistical Review (2008), 76, 3, 347-367

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 Basic Process Capability Indices 351

USL- ,
( 2. 2)

( 5, 2)
( 2)

LSL 1 ( 3' 2)

Figure 1. Distributions associated with five populations.

Conceptually, Cp compares the allowable process spread to the actual process spread, and
can be thought of as indicating the potential of the process to produce conforming material.
Consider a process that has both an on-target mean and a high Cp value (C* value). By sufficiently
shifting the process mean in the direction of any one of the specification limits, we can still
obtain any proportion of items outside the specification limits and still maintain a high Cp
value (C*p value). For example, in Figure 1, samples from any of the five normal distributions
( ( , 2), = 1,..., 5) will produce estimates of the Cp{C"p) index that are almost the same.
This is due to the fact that these five distributions have the same variance. As the actual process
spreads are smaller than the allowable process spread, the process capability index (C* ) will
be greater than 1, suggesting that the processes are capable. Only processes from distribution
1 are on target. It may be argued that processes from distributions 2 and 3 are still within the
specification limits and hence should be judged capable, even though they are not centred at
the target. However, the possibility of necessary adjustments could be costly. Finally, processes
from distributions 4 and 5 are incapable of meeting the specifications required as both have
substantial proportions of production beyond the specification limits. This example shows that
the Cp and C*p indices simply relate the process spread to the specification limits and do not
take into account the possible shifts of the process mean away from the target value.
Next we shall consider indices that take into consideration both the process mean and the
process dispersion.

3.1.2 The Cpk index

To deal with violations of the centring assumptions, the following pair of indices was developed
for the case = M:

!
U ? ? L
Cpk = min 3 3 (6)

?-TO
The index k represents a measure of the distance that the process lies o
the reduction in process capability caused by the lack of centring.
generation index. Using the algebraic identity min{a, b} ? [(a + b) -
of the Cpk can alternatively be written as:
-\ - \
Cpk =- -'
where D = (U - L)/2.
International Statistical Review (2008), 76, 3, 347-367
? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 352 M. . Anis

It should be noted that both Sullivan (1984) and Kane (1986) describe k as an absolute value;
however, Palmer & Tsui (1999) feel it is useful for k to retain its sign.
When the target value for the process mean is not necessarily equal to the mid-point M of
the specification limits, the analogous indices are defined by

C;k = mm(CPL*, CPU*), (8)

where
0 if \T - \ > -L
CPL* =
3 LV
( _ \?_?^ otherwise (9)

if \ - \ > -
cnr- - , \ -*\ ( )
U ?
)
** =-?-. (11)
min(T-L,U-T) J
Interestingly Boyles (1991) shows that Cpk is essentially a measure of process yield only and
can fail to distinguish between off-target and on-target process.

3.1.3 The Cpm index

The Cp and Cpk indices are appropriate measures of progress for quality improvement in
which reduction of variability is the guiding principle and process yield is the primary measure
of success. Taguchi (1986) has suggested a different approach to quality improvement in which
reduction of variation from the target value is the guiding principle. In fact, Taguchi (1988) was
the first author to propound the concept that there is a loss to society associated with missing
the target. This concept of societal loss is difficult, if not impossible, to quantify. However, to
be useful from a business perspective, a tool must be well defined, must be easy to use, and
must have a quantifiable financial impact so that results can be attributed to the success of the
business. Taguchi realised that just being within specification is not sufficient, so he developed
the concept of the quadratic loss function to address the deficiency of the "goal post" approach
to specification limits.
In this approach, any measured value of a product characteristic X entails a monetary loss
L{x) to the customer as well as to the society in general. The loss function L is usually assumed
to be well approximated by the symmetric squared error loss function,

L{x) = k(x - Tf,

for some positive constant k, so that L(T) = 0; and any deviation from the ideal value entails
some positive loss to the consumer or to the society. The capability of the process is represented
by the expected loss

E(L) = kE{(x - )2}.

This is a measure of process variation in terms of deviation of the characteristic X from the
target T. The appeal of expected loss is that it expresses process capability in monetary units,
and therefore enters naturally into management decision-making process.

international Statistical Review (2008), 76, 3, 347-367

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 Basic Process Capability Indices 353

Chan et al. (1988) introduced the so-called Taguchi capability index Cpm that is measurable
and directly related to the quadratic loss of the measured feature. It is defined by
U-L
(12)
Cpm - Irruir )2
assuming that the target value is equal to the mid-point M of the specification limits. The Cpm
is another second-generation index.
When M, the corresponding index is given by
U-T
C*pm = min (13)
3^( 2 + ( - )2' 37( 2 + ( - )2
Boyles (1991) presents the general statistical methodology for the capability index Cpm
without the restrictive assumption = T. Johnson (1992) exploits the relationship between the
capability index Cpm and the expected squared error loss to provide an intuitive interpretation
of Cpm in terms of the percentage loss. He shows that it is related to the expected relative loss
Le for the process, where the expected relative loss is defined as the ratio between the expected
squared error loss and the value that the product is worth when the process mean is equal to
its target T. In an interesting paper, Denniston (2006) provides the motivation for using Cpm.
He shows that Cpm can indicate the probability of meeting the customer's product specification.
It can be used to provide a better estimate of the cost of poor quality, and hence can be used to
better manage product quality to the customer.

3.1.4 The Cpmk index

It is easy to observe that Cpk is derived from Cp by modifying the numerator, whereas Cpm
is obtained by modifying the denominator. If the two modifications are combined, a new index
Cpmk, first proposed by Pearn et al. (1992), is obtained. It is defined by

Cpmk = min U-? ?'L 1 (14)

3 / 2 + ( - )2' 37 2 + ( - Tf
Observe that a manufacturing process satisfying the capability requireme
satisfy the capability condition "Cpm > c". On the other hand, a process sat
requirement "Cpm > c" may not satisfy the capability requirement "
manufacturing process does satisfy both capability requirements "Cpk > c"
process satisfies the capability requirement "Cpmk > c" as Cpmk < Cpk and
the index Cpmk provides a greater level of quality assurance with respe
process loss to the customers than the other two indices. This is a desired p
today's modern quality theory, as a reduction of process loss (variation fro
as important as increasing process yield (meeting the specifications). A
the more popular and widely used index, Cpmk is arguably the most u
processes with two-sided specification limits. The index alerts the user
increases and/or the process mean deviates from its target and is designed
near-normal processes. For semiconductor and microelectronics manufa
Cpmk is an appropriate index for capability measurement due to the high s
requirements on product quality and reliability.
The Cpmk index has a value of 0 when is at either specification limits
indicates closeness of to the specification limits. This is a desirable ch

International Statistical Review (2008), 76, 3, 347-367

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 354 M. . Anis

not shared by Cpm. When the target value is not the specification mid-point, the maximum value
for Cpmk is not the same as the maximum for Cpk and Cpm (recall that this maximum value is
the Cp value). We can observe that this index explicitly takes into account that the process mean
may not be midway between the specification limits and incorporates a penalty when deviates
from the target T. The index is constructed so that the larger the index, the more capable the
process. The Cpmk index is said to be a third-generation index.
In current practice, a process is called "inadequate" if Cpmk < 1.00, "marginally capable" if
1.00 < Cpmk < 1.33, "satisfactory" if 1.33 < Cpmk < 1.50, "excellent" if 1.50 < Cpmk < 2.00,
and "super" if 2.00 < Cpmk 2 .00 (Hsu et al, 2007).

3.2 Relationship among the Indices

We shall now show the relationship among the different indices and get the bounds for their
values.
It is easy to see that

Cpk = (l- \k\)Cp, (15)

and

Cpm = , Cp (16)
7"! + ^
Clearly,
Cpmk < Cpk < Cp, (17)
and
Cpmk S Cpm < Cp. (18)
The relationship between Cpk and Cpm is less obvious. Using equations (15) and (16), we

have Cpk = (1 ? \k\)Cpm^\ + ^ ~ 2 ? In the special case when = M, it can be shown that
Cpk < Cpm, if |^| < 9?2, where
U-L
D = ? (19)
Moreover, Parlar & Wesolowsky (1999) have observed that if T= M, then

c^ = cp-^I\7^) - (2?)
or equivalently
C
Cpm = p -. (21)
pm yf\+%Cp-Cpkf
lfT=M, we can also write
?
Cpt = -1 + {J\+?2) Cpm, (22)
where

? = (23)
International Statistical Review (2008), 76, 3, 347-367
? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 Basic Process Capability Indices 355

Boyles (1991) studied both Cpk and Cpm extensively, including the comparison of Cpk and
Cpm as functions of and , without assuming = T. He used a graphical technique to show
that Cpk fails to address adequately the problem of process centring.
From the definition of Cpmk, it follows that

Cpmk ? / ? (24)

or equivalently

Cpmk = Cpm* Cpk (using equation (16)). (25)

We note that

( \ - \\
:C
^ + {^fPk' (26)
:Cp

If ? ? , then clearly Cp = Cpk = Cpm = Cpmk> They differ in behaviour when T.

By plotting the four indices as surfaces, we can get a feeling for the sensitivity with regard to
departures of the process mean from the target value T, assuming that = M. It is easy to
see that, for fixed , when moves away from , then Cp does not change; Cpk changes, but
slowly; Cpm changes somewhat more rapidly than Cpk\ but Cpmk is the one that changes most
rapidly.
Proofs of some of these relationships, together with the conditions under which Cpm is greater
than or less than Cpk, can be found in Kotz & Johnson (1999).

3.3 Estimates of the Indices

For using these indices meaningfully, they have to be estimated based on sample data. The
estimate will depend upon how the statistics and are estimated. We shall assume that we
have a random sample of size n, given as {X\ ,X2,...,Xn} from a stable process. On the basis of
this assumption, we shall obtain a point estimate and confidence interval for these indices using
the natural estimators for and .
However, for applications where routine-based data collection plans are in usage, the
parameters and may be estimated using control charts. Again, from a practical perspective,
manufacturing characteristics information about the process may be obtained by estimating the
process capability using the past in-control data that includes multiple samples rather than a
single sample. In such cases, the distribution of the estimated process capability index based on
subsamples ought to be available. For brevity, we shall not discuss these two cases and refer the
reader to Pearn & Kotz (2006) and the references therein.
Because of the sampling variation introduced by estimation, it is important and relevant to
construct a confidence interval providing a range of values that includes the true index with a
high probability.

International Statistical Review (2008), 76, 3, 347-367

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 356 M. . Anis

3.3.1 Estimate of Cp Index

A natural estimator of Cp is given by
U-L
Cp = 6S '
where S = E?=i ^ ? X)2/(n ? A)]1/2 is the usual estimator of the process standard devia
, obtained on the basis of a random sample of size from a stable process.
A 100(1 ? a)% confidence internal for Cp may be expressed as

f y?
A/1-1,1-a/2 A ' Xn-l,ot/2
1

where 2_1a/2 and ^_ !_a/2 are the upper a/2 and 1 ? a/2 quantit?s of a chi-squared
distribution with ? 1 degrees of freedom, respectively.
It should be noted that if the standard deviation is estimated based on control charts, then
the appropriate sampling distribution should be used to get the confidence interval. For details,
see Pearn & Kotz (2006).

3.3.2 Estimate of Cpk Index

The natural estimator Cpk is obtained by replacing the process mean and process standard
deviation by their estimators X and S, respectively. Thus,

=-M| = I 1 - J- " \ CP'

The construction of the exact confidence intervals for Cpk is difficult because the distribution
of Cpk involves the joint distribution of two noncentral ^-distributed random variables.
Nagata & Nagahata (1992) proposed the following two-sided confidence interval for Cpk'.

Nagata & Nagahata (1994) provide a thorough treatment of the construction of approximate
confidence intervals for Cpk.

3.3.3 Estimate of Cpm Index

The index Cpm involves the unknown parameters and , which needs to be estimated from
a sample. Boyles (1991) proposed the following estimator of Cpm:
d
C ?

3Js? + (X-t/
where X = "= X?/n; S2 = "= (x? ~ ?/ - Note that and S2n are the maximum
likelihood estimations (MLEs) of and 2, respectively. Hence, the estimator Cpm is also
^ pm
theMLE of Cpm.
Several authors have suggested approaches for constructing approximate lower confidence
bounds for Cpm. Marcucci & Beazley (1988) propose using the ordinary chi-squared distribution
International Statistical Review (2008), 76, 3, 347-367
? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 Basic Process Capability Indices 357

to approximate the noncentral chi-squared distribution, resulting in the following approximate

lower confidence bound:

cPm]?^, 0<?<1,
where 2 x_a is the (1 ? a)-th percentile of the ordinary central chi-squared variable with
degrees of freedom. When the process is on target, that is, = , this provides an exact bound;
otherwise the bound is conservative.

3.3.4 Estimate of Cpmk Index

For a normally distributed process under statistical control, Pearn et al. (1992) suggested
using the natural estimator of Cpmk given by

U-X X-L
Cpmk = min
^S2n+{X-T)2 ^S2 + {X-T)2
D ? \X ? T\
+ (X-T)2
where X = "= Xi I and S2 = __^=i (X? ? X)2/n are the MLEs of and 2, respectively.
Chen & Hsu (1995) have shown that Cpmk is consistent and asymptotically unbiased. They
have also proved that if E(X4) < oo, then Cpmk is asymptotically normal. They have derived an
asymptotic 100(1 ? a)% confidence interval for Cpmk as
?~pmk
Cpmk ~F za/2

where
1
+
28 72 2 + ( - 1) 2
pmk Cpmk "T" ^pmk'
72(1 + 2)2
9(1 + ?52) 3(l+?52)3/2j
where a2mk is the asymptotic estimator of Var^C^); za? is the upper a/2 quantile of
the standard normal distribution; and m4 = __^=1 (Xi ? X)4/n, = (X ? T)/Sn, and S2 =
Y?i=l{Xi-X?/n.

4 Index Interpretation and Drawbacks

It must be noted that a state of statistical control is required of the process for the capability
index to have any long-term meaningful interpretation.
The index Cp only indicates the potential proportion conforming. A minimum value of Cp ?
1.33 is generally used for an ongoing process (Montgomery, 2001, p. 361). If the Cp value is 1,
and the process characteristic X is normally distributed and properly centred at the mid-point,
that is, = M = then the proportion (p) of NC items produced is rather small (0.27%).
In spite of a high value (>1) of Cp, can be more than 0.27% if the process is not properly
centred.
This drawback of the Cp value index is taken care of by the Cpk index. Recall from
equation (15) that Cpk = (1 ? \k\)Cp. This shows that Cpk is bounded above by Cp. The
International Statistical Review (2008), 76, 3, 347-367
? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 358 M. . Anis

Cpk index will achieve its maximum value (equal to the Cp value) when the process mean is at
the specification mid-point. As the process mean drifts away from the specification mid-point,
the Cpk value decreases linearly until it reaches a value of 0, when the process mean is equal to
one of the specification limits.
The indices Cp, Cpk, and k can be used for process improvement. Adapting from Palmer &
Tsui (1999), the steps for process improvement can take the following direction:

Step I: Obtain the estimates of Cp and Cpk.

Step II: If Cpk < Cp and Cp > 1, then evaluate k.

(a) If k > 0, then adjust the process location to decrease the process mean until Cpk = Cp.
(b) If k < 0, then adjust the location to increase the process mean until Cpk = Cp.

Step III: If Cp < 1.0, then identify and remove sources of variability and go to Step II.

It is also interesting to note that the two indices Cp and Cpk can be associated with the stepwise
loss function. Gunter (1989) has advocated caution in the use of the Cpk index.
The third index Cpm attains its maximum value when = T, and decreases in value
symmetrically, in a bell-shaped pattern, as the process mean shifts away from the target value.
If the process mean is at the target value, that is, = , then from equation (16), we have
Cpm = Cp. Unlike Cpk, the value of Cpm does not decrease to 0 as the process mean approaches
the specification limits. The value of the Cpm index is independent of the closeness of process
mean to the specification limits. Only the distance between and the target is considered.
The entire curve for Cpm shifts with the target value, regardless of the actual locations of the
specification limits. It should be noted that the target value is not necessarily the specification
mid-point value.
As Cpm indicates the reduction in process capability due to shifts in the process mean away
from the target, the pair of indices Cp and Cpm can be used to direct process improvement
activities. Palmer & Tsui (1999) suggest one such road map. It is worth noting that Cpm was the
first index to be developed that explicitly used a quadratic loss function. In view of the fact that
Cpm is bounded above by Cp, it is clear that Cp index enjoys membership in both categories.
The term ( ? )2 in the denominator of Cpmk in equation (14) may be viewed as an additional
penalty to the process quality for the departure of the process mean from the target. This penalty
ensures that Cpmk will be more sensitive to departure than Cpk, and therefore, is better able to
distinguish between off-target and on-target processes.
It should be clearly understood that process capability cannot be adequately characterised
by a single index. Bothe (2002) has shown that Cpk can be misleading, and is inappropriate
for product features with asymmetric tolerances. He advises the reporting of Cp, Cpk, Plsl
(percentage nonconforming below the LSL), and pusL (percentage nonconforming above the
USL) to have a very good idea of what is happening regarding the process output and what
actions are necessary to improve it.
Deleryd & Vannman (1999) and Vannman (2001, 2005) have introduced the concept of
process capability plots as powerful tools to monitor and improve the capability of industrial
processes. An advantage of process capability plots is that they instantly provide information
about the location and spread of the characteristic under study. When the process is noncapable,
the plots are helpful in trying to understand if it is the variability, the deviation from the target,
or both that need to be reduced to improve the capability.
Daniels et al. (2005) address the problem of comparing two capability indices (they consider
Cp, Cpk, and Cpm ) for two different processes or the same process before and after an adjustment.
They provide recommendations for selecting an appropriate method based on power and test size

International Statistical Review (2008), 76, 3, 347-367

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 Basic Process Capability Indices 359

computations, whereas Melloy & Chandra (1992) demonstrate that the casual estimation of the
proportion of NC items using the Cpk index can be very misleading when items are screened.

5 Indices, NCs, and Sample Size

In this section, we shall comment on the relationship of the indices to the number of the NC
products and the sample size required for scientifically valid comparisons.

5.1 Indices and NC

Capability indices are popular because they provide single-number summaries to managers
responsible for many processes running simultaneously. However, it should be noted that it
is very difficult, if not altogether impossible, for a single index to capture the dynamics of a
process.
If the process is normally distributed and centred at the mid-point, then a Cp value of 1
indicates that the proportion of NC products is 0.27%. It is easy to show that the probability of
obtaining an NC value is

20(-3C/?). (27)
where (.) denotes the standard normal cumulative distribution function.
The index Cpk alone does not determine the proportion (p) of NCs; but provides an upper
bound, given by

< 2d>(-3CM). (28)

However, Cp and Cpk together determine by the equation

<b(-3(2Cp-Cpk)) + <S>(-3Cpk). (29)

If the expected proportion NC is regarded as the most important criterion, then Cpm is
unreliable, because the same value of Cpm can be associated with a wide range of values of
the expected proportion NC. However, it should be noted that the motivation for Cpm does not
arise from examining the number of NC products in a process but rather from requiring the
ability of the process to be in the neighbourhood of the target. This motivation has little to do
with the number of NC parts. However, the index provides an upper bound to the proportion
of NC products, < 2 (?3Cpm). Moreover, it can be shown that | ? T\ < D/(3Cpm). This
inequality can be interpreted as a Cpm?value of 1 implies that the process mean lies within
the middle-third of the specification range. As the design of Cpm is based on the average process
loss relative to the manufacturing tolerance, the index Cpm provides an upper bound on the
average process loss.
Similarly, the Cpmk index provides an upper bound to the proportion of NC products, <
2<&(?3Cpmk). However, Cpmk is much more sensitive than other capability indices to movement
in the process average relative to M. Further, Cpmk reveals the maximum information about
the location of the process average. It can be shown that the distance between and M is less
than D/(l + 3Cpmk). This can be interpreted as a Cpmk?value of 1 implies that the process
mean lies within the middle-fourth of the specification range. This is appreciably a very small
interval.
Ranking the four common indices in an increasing order of sensitivity to departures of the
process mean from the target value, we obtain (1) Cp (2) Cpk (3) Cpm (4) Cpmk. If the proportion

International Statistical Review (2008), 76, 3, 347-367

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
360 M. . Anis

of NC units ( ) is of primary importance, then Cp and Cpk should be used as they are more
closely concerned with the percentage of NC products.

5.2 Sample Size

Kane (1986) proposed the use of Operating Characteristic (OC) curve to analyse the sampling
variation of Cp. He tests the hypothesis:

Hq . Cp < co (process is not capable)

versus

H\\Cp> co (process is capable)

as a test for process capability (Cp >co). He has given a table for critical value deter
for testing Cp for different sample sizes. Using this table, Kane (1986) points out
automotive industry machine qualification practices that use sample size of =
for the most part, inadequate. He points out that even though current automotive
machine qualification use a Cp value of 1.33, accounting for nonsampling problems
qualification runs different from production runs, these practices may not adequate
for sampling variability. Though both Charbonneau & Webster (1978, p. 112) and M
(2001, p. 361) recommend that new equipment qualifications should use Cp = 1.5,
them recommends a sample size.
Chan et al. (1988) proposed an analogous approach to analyse the sampling va
Cpm. They also present a table for determining sample size and critical values for testin
However, the same criticisms presented above are applicable to the OC curve appr
analysing the stochastic properties of Cpm.
Chou et al. (1990) show that fairly large sample sizes are needed to determine Cpk
For example, to be 95% certain that the true Cpk is not more than 10% below the
sample estimated Cpk when the sample estimated Cpk is 1.33 (a fairly used target),
have a sample size of about 350, assuming that the underlying quality characteristic has
distribution.
Shore (1997) has studied the effect of autocorrelation on process capability anal
correlated observations contain less information than noncorrelated ones, the sample siz
to arrive at an accurate estimate of the capability index is larger. In fact, Zhang (1998) h
that for < 100, the estimated values Cp and Cpk are quite variable, whereas the var
Cp and Cpk decrease slowly when is greater than 100. Hence, he recommends using
size > 100 to avoid large variations in Cp and Cpk.

6 Effect of Nonnormality, Asymmetry, and Autocorrelation

The theory of PCIs discussed so far is based on the assumption that the distributi
underlying process characteristic is normally distributed and the observations are indepe
However, there are situations when such assumptions are violated. The distribution
ovality, concentricity, run-out, etc., are essentially nonnormal and skewed. The ex
nonnormality include impurity content in chemicals, measures of acidity/alkalinity, par
powder, and measures of squareness, parallelism, and flatness in machined componen
Large Scale Integration (VLSI) processing technology, the distributions of boron-i
atomic concentration in silicon are more and more negatively skewed from norm
International Statistical Review (2008), 76, 3, 347-367
? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 Basic Process Capability Indices 361

implantation energy increases (Sze, 1988). Such a process displays natural nonnormal behaviour.
Distributions of most measured critical features sizes of photomasks deviate significantly from
normality. Nonnormality can also arise from truncation of data out of specification. As pointed
out by Ryan (2000) and illustrated therein, normal distributions do not exist in practice. Gunter
(1989) has expressed the view that in most real industrial processes, nonnormality is the norm.
Similarly, in process industries, it is very common to expect and encounter autocorrelated data.
We shall now consider how the PCIs are affected in such cases.

6.1 Nonnormal Data

Much work on these four popular indices has been done on the assumption that the
measured quality characteristic is normally distributed (at least approximately). However, it
is difficult to believe that a good industrial process must result in a normal distribution for
every measured characteristic (Pyzdek, 1995). Munechika (1992) details several examples of
machining processes that are inherently nonnormal. If the normality-based PCIs are used to
deal with nonnormal processes, the results are generally incorrect, as expected. Somerville &
Montgomery (1996-97) have observed that normality-based PCI cannot calculate the process
fallout accurately when the underlying distribution is nonnormal. Gunter (1989) has shown three
different distributions with identical values of Cp and Cpk but different proportions of NC parts.
Kokcherlakota et al. (1992) give additional information on the effects of nonnormality on PCIs.
A large body of literature has appeared on dealing with nonnormal data. The available methods
include the empirical percentile method (Clements, 1989; McCormark et al, 2000), and Monte
Carlo simulation (English & Taylor, 1993). Chou et al. (1998) and Polansky et al. (1998) have
suggested using Johnson's system of distribution curves to transform the nonnormal data into
normality. Tang & Than (1999) review seven methods for performance comparison in their
ability to handle nonnormal data. They also suggest applicable methods for each defined range
of skewness and kurtosis under mild and severe departures from normality. In a recent paper, Wu
& Swain (2001) proposed a method based on weighted variance to deal with nonnormal data.
On the basis of simulation, they conclude that the weighted variance-based estimators perform
best in both accuracy and efficiency and the recommended sample size for better estimating the
nominal values would be at least 100. Ding (2004) presents a method to evaluate the PCI for
a set of nonnormal data from its first four moments. Compared with some existing methods,
his method gives a more accurate PCI estimation and shows less sensitivity to sample size. Pal
(2005) suggests using the generalised lambda distribution (GLD) to evaluate nonnormal process
capability indices as the GLD has an edge over the other family of distributions while modelling
a process data.

6.2 Skewed Population

For many quality characteristics, such as circularity, cylindricity, straightness, and flatness,
positive skewness in the inspection data is the norm and in fact is desirable. To deal with such
data, Wright (1995) proposed an index C5, which is basically an adaptation of the Cpmk, proposed
by Pearn et al. (1992). Wright included a skewness term in the denomination to define

^ =.min
Cs / ?.U - -L
?=============
\3 / 2 + ( - )2 + | 3/ | 3>2 + ( - )2 +
The absolute value of the skewness parameter, added to the
reducing the value of the index when asymmetry exists. This in

International Statistical Review (2008), 76, 3, 347-367

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
362 M. . Anis

handle situations where worsening capability is characterised not only by an increase in variance
and/or deviation of the mean from the target but also an increase in skewness. He advocates
the use of the Cs index for monitoring near-normal processes where loss of capability leads to
asymmetry.
Nahar et al. (2001) modified Wright's index by simply omitting the absolute value sign in the
denominator and subtracting the skeweness value. Thus, they propose

n ? I U - -L
Csm = min ? -, ? =
\3^ 2 + ( - -( 3/ ) 3 / 2 +
Hence, a positive skewness value (in which "sk
characteristic) would cause the index to increase and
characteristic, would have the effect of decreasing the
Chang et al. (2002) proposed a new method of con
skewed population based on a weighted standard de
standard deviation of a quality characteristic into up
factors in computing the deviation above and below
the index using decomposed deviations in accordanc
data. For symmetric populations, the proposed PCIs

6.3 Autocorrelated Data

Traditionally, a process must be considered free from

its capability can be determined. However, there are
inherently present and it is very difficult, if not altog
example, the degradation of a cutting tool. This essenti
in the data. During the last decade, the determination o
has been discussed by a few authors. We shall attempt t
In those situations where a variation due to an as
traditional process capability measures cannot be m
capability to be considered dynamic, Spiring (1991
process capability in such situations. That is, for proces
assignable cause, the capability of the process is con
process ages. The changing ability of the process can
index that considers both process variation and proxim
and Cpm are appropriate indices, which take both the t
author has developed the theory for the Cpm index (be
by Chan et al, 1988 has a pdf that can be used to d
the sampling results). The total variation 2 is decom
due to assignable causes ( 2) and variation due t
2. In assessing process capability, 2 is only conside
a constantly changing capability essentially requires
be made by examining the process capability over sm
proposed capability measure is that it can be used to
influence of systematic assignable cause. A similar ex
Hsu (2007), whereas the present author is investigating
Jagadeesh & Babu (1994) have investigated the pro
in the presence of tool wear. They have used four diffe
variability 2, and concluded that the results based on th

International Statistical Review (2008), 76, 3, 347-367

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 Basic Process Capability Indices 363

In an influential paper, Shore (1997) dealt with the problem of process capability analysis
when data are autocorrelated. He showed that autocorrelation affects the variance of the sample
mean and hence the confidence interval associated with the sample mean. The author studied
four approaches of estimating capability for autocorrelated data and concluded that when
both performance and convenience in application are important, the model-free approach is
superior.
Zhang et al. (1990) have shown that the sample variations in estimates of capability indices
cannot be ignored and hence interval estimation should be considered. Assuming {Xt} is a
discrete Gaussian process, Zhang (1998) has shown that the variance of Cp(Cpk)are functions
of Cp(Cpk), the sample size and the process autocorrelation p? (from a lag of 1 to n).
Scholz & Vangel (1998) are concerned with the construction of tolerance bounds for Cpk
when samples come in batches and the intrabatch correlation reduces the amount of independent
information. They reduce the problem to the independent and identically distributed case by the
simple device of effective sample size.
Noorossana (2002) has shown the variance estimate obtained from the original data is no
longer an appropriate estimate to consider for process capability studies when observations
are autocorrelated. He suggests using a combined procedure based on multiple regression and
time-series modelling to remove the autocorrelation patterns that may be present in the data and
also estimate model parameters effectively.

7 Measurement Error and PCIs

In spite of the large volume of work done on various aspects of process capability indices, the
effects of measurement errors on these indices has received comparatively very little attention.
McNeese & Klein (1991-92) were perhaps the first one to point out that the variability
inherent in the measurement systems and sampling techniques adds variability to the output
from a process and hence affects the process capability. Hence it is necessary to have a capable
measurement system, which they define as a system in statistical control with respect to the
average and variation, whose average value is equal to the true value, and which is responsible
for less than 10% of the total process variance. To decrease the total variability, it is necessary to
determine where the greatest opportunity for improvement exists. This can be done by examining
the components of variation. Usually, the components of variation are due to sampling, variation
due to measurement system (gauge R & R), and error due to manufacturing process. Porter &
Oakland (1991) also emphasize that process capability assessment is dependent upon the test
or measurement method. They suggest that the capability of the test method should be at least
two-and-a-half times the observed capability.
Persijn & Nuland (1996-97) also dwell upon the relationship between measurement system
capability and process capability. They argue that process capability analysis is meaningful
only if the measurement system is capable. For this purpose, they introduce the concept of
measurement index (MI) and define it as the ratio of the process standard deviation ( )
to the standard deviation (am) of the measurement method. Thus, MI = ap/om. A capable
measurement system has MI > 3, that is, the characteristics of the process can be measured
and hence production can be assessed. In fact, the greater is the MI, the less important is the
measurement error.
In a significant work, Mittag (1997) discusses the effects of measurement error on the
performance of the four most basic process capability indices. When the error is a constant
measurement error, the Cp index is unaffected, whereas the other three indices (Cpk, Cpm, and
Cpmk) may be affected in either direction.

International Statistical Review (2008), 76, 3, 347-367

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 364 M. . Anis

When the measurement error is random, instead of the true variable X, we observe the
empirical variable Xe, where Xe = X + V, with V being the random error component. It is
further assumed that the random variablesX and V are stochastically independent mdE(V) = 0.
Clearly, the process variance is increased. Mittag (1997) has shown that stochastic measurement
error always implies a decrease in the indices. For the indices Cp and Cpk, the extent of downward
distortion only depends on the value of the error contamination degree = / . In contrast,
for the indices Cpm and Cpmk-> the effect of a random measurement error is determined by as
well as by = ( ? )/ . Whereas for Cp and Cpk the measurement error effects remains
unchanged at M, the distortion of the Cpm and Cpmk weakens (up to being negligible) with
increasing departure from the target. This implies that the index Cpk is inferior to the indices
Cpm and Cpmk with respect to the robustness against normally distributed measurement errors.
Mittag 's work is important as it shows that random and constant measurement errors can
considerably falsify the results of process capability analyses. They emphasize that the accuracy
of a capability analysis could be significantly influenced by the accuracy of the gauges. This
fact underlines the importance of ensuring gauge capability before evaluating process capability
and, consequently, measurement errors should receive greater attention.
However, although the analysis of Mittag (1997) is confined to considering the effects of
measurement errors only on the behaviour of theoretical capability indices, such effects are not
taken into account when PCIs are estimated for sample data. Bordignon & Scagliarini (2002)
extend the analysis of Mittag (1997) to the inferential properties of the estimators of Cp and
Cpk by considering the effects of measurement errors on the properties of capability indices
estimated from sample data. They have shown that the Cp estimator obtained, from the sample
data contaminated by random measurement errors, is biased, tending towards steady negative
value as the sample size -> oo, and increases with the contamination degree r. In contrast,
the usual estimator of Cp in the measurement error-free case always has a positive bias going to
zero as ?> oo. They have also shown that Var(Cp is never greater than Va^C^). Therefore,
when comparing the mean squared errors (MSEs) of the two estimators, the bias component
plays a more important role. Similar results have been obtained for Cpk index. Scagliarini (2002)
has analysed the properties of the estimator of Cp for autocorrelated data in the presence of
measurement error. Later, Bordignon & Scagliarini (2006) study the behaviour of the estimator
of Cpm in the presence of measurement error.
Pearn & Liao (2006) consider the estimation and testing of Cpu and CpL in the presence
of measurement error to obtain adjusted lower confidence bounds and critical values for true
process capability. These can be used to determine whether the factory processes meet the
capability requirement when the measurement errors are unavoidable.
Hsu et al. (2007) conduct a sensitivity study for the Cpmk index in the presence of gauge
measurement errors. They consider the use of capability testing of Cpmk as a method for
obtaining lower confidence bounds and critical values for true process capability when gauge
measurement errors are unavoidable. Their research shows that using the estimator with sample
data contaminated by measurement errors severely underestimates the true capability, resulting
in an imperceptibly smaller test power.

8 Conclusion

This paper attempts to survey the four most popular process capability indices. In addition to
introducing the indices, an attempt has been made to provide interpretation and point out the
drawbacks of these indices. The popular misunderstanding with respect to the requisite sample
size has been pointed out. The neglected aspects of nonnormal distribution, skewed distribution,

International Statistical Review (2008), 76, 3, 347-367

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 Basic Process Capability Indices 365

and autocorrelated data have been studied. The impact of measurement error on the PCI has
been highlighted. It is hoped that this paper will help in the correct and proper understanding and
appreciation of the process capability indices and their correct applications. Incorrectly applied
and/or interpreted, these indices can generate an abundance of misinformation that will confuse
the shop floor personnel and the management alike, waste resources, and lead to incorrect
decision making.

Acknowledgements

Thanks are due to the Referee, the Associate Editor, and the Editor whose comments and
constructive suggestions improved the presentation of the paper.

References
Barnett, N.S. (1990). Process control and product quality. Int. J. Qual. Reliab. Manage., 7, 34-43.
Bordignon, S. & Scagliarmi, M. (2002). Statistical analysis of process capability indices with measurement errors.
Qual. Reliab. Eng. Int., 18, 321-332.
Bordignon, S. & Scagliarmi, M. (2006). Estimation of Cpm when measurement error is present. Qual. Reliab. Eng.
Int., 22, 787-801.
Bothe, D.R. (2001). Measuring Process Capability. Cedarburg, WI: Landmark Publishing.
Bothe, D.R. (2002). Discussion. J. Qual. Technol, 34, 32-37.
Bothe, D.R. (2006). Assessing capability for hole location. Qual. Eng., 18, 325-331.
Boyles, R.A. (1991). The Taguchi capability index. J. Qual. Technol, 23, 17-26.
Bulba, E.A. & Ho, L.L. (2006). Capability index for relational functions: inferential procedures. Qual. Eng., 18,69-77.
Chan, L.K., Cheng, S.W. & Spiring, EA. (1988). A new measure of process capability, Cpm.J. Qual. Technol, 20(3),
162-175.
Chang, Y.S., Choi, LS. & Bai, D.S. (2002). Process capability for skewed populations. Qual. Reliab. Eng. Int., 18,
383-393.
Charbonneau, H.C. & Webster, G.L. (1978). Industrial Quality Control. Englewood Cliffs, NJ: Prentice Hall, 112.
Chen, S.M. & Hsu, N.F. (1995). The asymptotic distribution of the process capability index Cpmk. Commun. Stat. -
Theory. Methods., 24(5), 1279-1291.
Chou, Y.M. (1994). Selecting a better supplier by testing process capability indices. Qual. Eng., 6, 427^38.
Chou, Y.M., Owen, D.B. & Borrego, A.S.A. (1990). Lower confidence limits on process capability indices. J. Qual.
Technol, 22(3), 223-229.
Chou, Y.M., Polansky, A.K. & Mason, R.L. (1998). Transforming non-normal data to normality in statistical process
control. J. Qual. Technol, 30(2), 133-141.
Clements, J.A. (1989). Process capability calculations for non-normal distributions. Qual. Prog., 22(9), 95-100.
Crain, CR. (1993). Process capability indices and customer-supplier relations. ASA Proc. Sect. Qual. Product, 156
158.
Daniels, L., Edgar, B., Burdick, R.K. & Hubele, N.F. (2005). Using confidence intervals to compare process capability
indices. Qual. Eng., 17, 23-32.
Deleryd, M. & V?nnman, . (1999). Process capability plots?a quality improvement tool. Qual. Reliab. Eng. Int.,
15,213-227.
Denniston, . (2006). Capability indices and conformance to specification: The motivation for using Cpm. Qual Eng.,
18, 79-88.
Ding, J. (2004). A method of estimating the process capability index from the first four moments of non-normal data.
Qual. Reliab. Eng. Int., 20, 787-805.
English, J.R. & Taylor, G.D. (1993). Process capability analysis?a robustness study. Int. J. Prod. Res., 31,1621-1635.
Feigenbaum, A.V (1951). Quality Control. New York: McGraw Hill.
Gill, M.S. (1990). Stalking six sigma. Business Month., January 1990, 42^6.
Gryna, EM. (1988). Quality Control Handbook, 4th ed., Ed. J.M. Juran, pp. 14-16. New York: McGraw Hill.
Gunter, B.H. (1989). The use and abuse of Cpk. Qual Prog., 22(3), 108-109.
Hsu, B.M., Shu, M.H. & Pearn, W.L. (2007). Measuring process capability based on Cpmk with gauge measurement
errors. Qual. Reliab. Eng. Int., 23(5), 597-614.
Jagadeesh, R. & Babu, A.S. (1994). Process capability assessment with tool wear: an investigative study. Int. J. Qual
Reliab. Manage. ,11,51 -62.

International Statistical Review (2008), 76, 3, 347-367

? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 366 M. . Anis

Johnson, T. (1992). The relationship of Cpm to squared error loss. J. Qual. Technol, 24, 211-215.
Juran, J.M. (1951). Quality Control Handbook, 1st ed., pp. 253-279; 404-409. New York: McGraw Hill.
Juran, J.M. (1962). Quality Control Handbook, 2nd ed., pp. 11-14; 11-27. New York: McGraw Hill.
Juran, J.M. & Gryna, F.M. (1980). Quality Planning and Analysis, 2nd ed., pp. 283-295. New York: McGraw Hill.
Kane, YE. (1986). Process capability indices. J. Qual Technol, 18(1), 41-52.
Kokcherlakota, S., Kocherlakota, K. & Kirmani, S.N.A.U. (1992). Process capability indices under non-normality. Int.
J. Math. Stai. Sci., 1(2), 175-210.
Kotz, S. & Johnson, N.L. (1993). Process Capability Indices. New York: Chapman and Hall.
Kotz, S. & Johnson, N.L. (1999). Delicate relations among the basic Process capability indices Cp, Cpk and Cpm and
their modifications. Commun. Stai. - Comput. Simul, 28(3), 849-866.
Kotz, S. & Johnson, N.L. (2002). Process capability indices?a review, 1992-2000 (with discussions). J. Qual. Technol.,
34, 2-53.
Kotz, S. & Lovelace, CR. (1998). Process Capability Indices in Theory and Practice. London: Arnold.
McCormark, D.W. Jr., Harris, LR., Hurwitz, A.M. & Spagon, P.D. (2000). Capability indices for non-normal data.
Qual. Eng., 12, 489^95.
McCoy, P.F. (1991). Using performance indexes to monitor production processes. Qual Prog., 24(2), 49-55.
McNeese, W.H. & Klein, R.A. (1991-92). Measurement systems, sampling and process capability. Qual. Eng., 4(1),
21-39.
Marcucci, M.O. & Beazley, C.C. (1988). Capability indices: Process performance measures. ASQC Qual. Congr.
Trans., 42, 516-523.
Melloy, B. J. & Chandra, M. J. (1992). Consequences of using the Cpk index to estimate the proportion of non-conforming
items. Int. J. Prod. Res., 30, 2253-2263.
Mittag, H.J. (1997). Measurement error effects on the performance of process capability indices. Frontiers in Statistical
Quality Control, Eds. H.J. Lenz & P.T. Wilrich, pp. 195-206. Heidelberg: Physica-Verlag.
Montgomery, D.C. (2001). Introduction to Statistical Quality Control, 4th ed. New York: John Wiley.
Munechika, M. (1992). Studies on process capability in machining processes. Rep. Stat. Appi Res. JUSE, 39, 14-29.
Nagata, Y. & Nagahata, H. (1992). Approximation formulas for the confidence intervals of process capability indices.
Rep. Stat. Appi Res. JUSE, 39(3), 15-29.
Nagata, Y. & Nagahata, H. (1994). Appreciation formulas for the lower confidence limits of process capability indices.
Okayama Econ. Rev., 25, 301-314.
Nahar, P.C., Hubele, N.F. & Zimmer, L.S. (2001). Assessment of a capability index sensitive to skewness. Qual. Rehab.
Eng. Int., 17, 233-241.
Noorossana, R. (2002). Process capability analysis in the presence of autocorrelation. Qual Reliab. Eng. Int., 18,
75-77.
Pal, S. (2005). Evaluation of non-normal process capability indices using generalized lambda distribution. Qual Eng.,
17, 77-85.
Palmer, K. & Tsui, K.L. (1999). A review and interpretations of process capability indices. Ann. Oper. Res., 87, 31?47.
Parlar, M. & Wesolowsky, G.O. (1999). Specification limits, capability indices and process centering in assembly
manufacture. J. Qual. Technol, 31(3), 317-325.
Pearn, WL. & Hsu, Y.C. (2007). Optimal tool replacement for processes with low fraction defective. Eur. J. Oper. Res.,
180, 1116-1129.
Pearn, W.L. & Kotz, S. (2006). Encyclopedia and Handbook of Process Capability Indices. Singapore: World Scientific.
Pearn, W.L. & Liao, M.Y. (2006). One-sided process capability assessment in the presence of measurement errors.
Qual Reliab. Eng. Int., 22, 771-785.
Pearn, W.L., Kotz, S. & Johnson, N.L. (1992). Distributional and inferential properties of process control indices. J.
Qual. Technol, 24(4), 216-231.
Persijn, M. & Nuland, Y.V (1996-97). Relation between measurement system capability and process capability. Qual.
Eng., 9(1), 95-98.
Polansky, A.K., Chou, Y.M. & Mason, R.L. (1998). Estimating process capability indices for a truncated distribution.
Qual. Eng., 11(2), 257-265.
Porter, L.J. & Oakland, J.S. (1991). Process capability indices?an overview of theory and practice. Qual. Reliab. Eng.
Int., 7, 437-448.
Pyzdek, T. (1995). Why normal distributions aren't [all that normal]. Qual. Eng., 7, 769-777.
Ryan, T. (2000). Discussion on "Controversies and contradictions in statistical process control". J. Qual. Technol, 32,
367-369.
Scagliarini, M. (2002). Estimation of Cp for auto-correlated data and measurement errors. Commun. Stat. - Theory.
Methods.,31, 1647-1664.
Schneider, H. & Pruett, J. (1995). Use of process capability indices in the supplier certification process. Qual. Eng.,
8, 225-235.
International Statistical Review (2008), 76, 3, 347-367
? 2008 The Author. Journal compilation ? 2008 International Statistical Institute
 Basic Process Capability Indices 367

Scholz, F. & Vangel, M. (1998). Tolerance bounds and Cpk confidence bounds under batch effects. Advances in
Stochastic Models for Reliability, Quality and Safety, Eds. W. Kahle, E. von Collani, J. Franz & U. Jensen,
pp. 361-379. Boston: Birkh?user.
Shewhart, W.A. (1931). Economic Control of Quality of Manufactured Products. New York: D. Van Nostrand.
Shore, H. (1997). Process capability analysis when data are autocorrelated. Qual. Eng.,9(4), 615-626.
Somerville, S.E. & Montgomery, D.C. (1996-97). Process capability indices and non-normal distributions. Qual. Eng.,
9, 305-316.
Spiring, F.A. (1991). Assessing process capability in the presence of systematic assignable cause. J. Qual. Technol.,
23(2), 125-134.
Spiring, F.A. (1995). Process capability: A total quality management tool. Total Qual. Manage., 6(1), 21-33.
Spiring, F, Leung, B., Cheng, S.W. & Yeung, A. (2003). A bibliography of process capability papers. Qual. Reliab.
Eng. Int., 19, 445^60.
Sullivan, L.P. (1984). Reducing variability: A new approach to quality. Qual. Prog., 17(7), 15-21.
Sze, S.M. (1988). VLSI Technology, 2nd ed. New York: McGraw Hill.
Taguchi, G. (1986). Introduction to Quality Engineering. Tokyo: Asian Productivity Organization.
Taguchi, G. (1988). Introduction to Quality Engineering, Designing Quality into Products and Processes. Tokyo:
Asian Productivity Organization.
Tang, L.C. & Than, S.E. (1999). Computing process capability indices for non-normal data: A review and comparative
study. Qual. Reliab. Eng. Int., 15, 339-353.
Tsui, K.L. (1997). Interpretation of process capability indices and some alternatives. Qual. Eng., 9(4), 587-596.
V?nnman, . (2001). A graphical method to control process capability. Frontiers in Statistical Quality Control,
Eds. HJ. Lenz & P.T.H. Wilrich, pp. 290-311. Physica: Heidelberg.
V?nnman, . (2005). The circular safety region: A useful graphical tool in capability analysis. Qual. Reliab. Eng. Int.,
21, 529-538.
Wheeler, D.J. (1999). Beyond Capability Confusion. Knoxville: SPC Press.
Wright, P.A. (1995). A process capability index sensitive to skewness. J. Stat. Comput. Simul., 52 , 195-203.
Wu, H.H. & Swain, J.J. (2001). A Monte Carlo comparison of capability indices when processes are non-normally
distributed. Qual. Reliab. Eng. Int., 17, 219-231.
Zhang, N.F. (1998). Estimating process capability indexes for autocorrelated data. J. Appi. Stat., 25(4), 559-574.
Zhang, N.F, Stenback, G.A. & Wardrop, D.M. (1990). Interval estimation of process capability index Cpk. Commun.
Stat. - Theory. Methods., 19, 4455-4470.

R?sum?

Une revue des quatre index de capacit? de proc?d? fondamentaux a ?t? faite. L'inter-relation entre ces index a ?t?
soulign?e. L'attention a ?t? dessin?e ? leurs inconv?nients. La relation de ces index au nonconformer de proportion a
?t? demeur?e sur et la condition de la taille d'?chantillon suffisante a ?t? soulign?e. Les remarques d'avertissement
sur l'usage de ces index dans le cas de distributions nonnormaux, les distributions d?form?s et les donn?es d'auto
correspondu sont aussi pr?sent?s. L'effet d'erreur de mesure sur les index de capacit? de proc?d? a ?t? trait? dans les
moindres d?tails.

[Received January 2008, accepted September 2008]

International Statistical Review (2008), 76, 3, 347-367