This pllper was presented at the Journal 01 Quality Technology Session at the 36th Annual Fa" Technical Conference

01 the Chemical
and Process Industries Division and Statistics Division 01 the American Society lor Quality Control and the Section on Physical &
Engineering Sciences 01 the American Statistical Association In Philadelphia, Pennsylvania, October 8-9, 1992.

Recent Developments in Process Capability Analysis

ROBERT N. RODRIGUEZ
SAS Institute lnc., Cary, North Carolina 27513-8000

This is the introductory article in a special issue of the Journal of Qualify Technology devoted to the
topic of process capability analysis. Statistical issues generated by the current use of process capability
indices are outlined. The contributions of the articles in this issue are highlighted, and the role of
related statistical and graphical methods is discussed.

Background 4. How should normality be checked?

5. Are there appropriate capability indices for non-
f~r
§ pability havetechniques
T AT IST ICAL analyzing process ca-
been applied on a large scale by a
normal process data?
variety of industries since the early 1980's. Although 6. Why does pronounced nonnormality occur so often
a standard definition has not yet emerged for the term in process data? How can inherent nonnormality
"process capability analysis", there is widespread (occurring naturally) be distinguished from in-
duced nonnormality (due to causes such as strat-
agreement that its objective is to determine how well
ification) ?
the output from a process meets specification limits
set by engineering requirements or by the customer. The answers require statistical theory, but the ques-
There is also agreement-unfortunately not as wide- tions themselves are not academic. As corporate qual-
spread-that a process must be in statistical control ity improvement efforts become standardized, and as
before its capability can be assessed. In other words, evidence of capability becomes a requirement in sup-
a stable, predictable distribution for the output is a plier contracts, companies are reconsidering the value
prerequisite for capability analysis. and reliability of single-number summaries for the
To most process engineers, capability analysis behavior of complex processes.
means the use of histograms and capability indices Some organizations have required that tests of nor-
such as Cp and Cpl<' These continue to be the tools most mality accompany reported indices or have mandated
widely taught and required by industry for this pur- larger sample sizes (as large as 120 in one company).
pose. However, during the past three years there has Other organizations have introduced their own ca-
been a growing recognition that these tools are limited pability indices, while still others have abandoned or
and, in particular, that standard capability indices are heavily restricted the use of indices after discovering
appropriate only with measurements that are inde- that their misuse is an obstacle to improvement.
pendent and reasonably normally distributed. Thus a
number of statistical questions are being asked: Various authors have commented on the weak-
nesses of capability indices. Gunter (1989), for ex-
1. What are appropriate interpretations for capability ample, discusses the limitations of Cpk with nonnormal
indices-with or without normality? data and cautions that unless the process is in control
2. Assuming normality, what are the distributional and hence predictable, the use of Cpk "becomes a kind
and sampling properties of commonly used indices? of mindless effort that managers confuse with real
How do these depend on sample size? statistical process control efforts". Other writers have
3. Assuming normality, how can confidence bounds criticized standard capability indices as over-simpli-
be obtained for capability indices? fications (see, for instance, Kitska (1991)).
In spite of these misgivings, the popularity of ca-
Dr. Rodriguez is Manager of Statistical Quality Improvement pability indices has not diminished, largely because
Research and Development. He is a Member of ASQC. single-number summaries are irresistible to managers

Journal of Qualify Technology 176 Vol. 24, No.4, October J992

 RECENT DEVELOPMENTS IN PROCESS CAPABIlITY ANALYSIS 177

responsible for hundreds of processes running con- and

currently. Thus, it is fortunate that a number of theo-
retical studies of process capability indices have ap- Cpic -- mtn
. (USL -
3(1
IL IL -
,
LSL)
3(1
peared in recent years.
Recall that Cp simply compares the tolerance spread
In This Issue of JQT
(USL - LSL) to the natural spread of the process (as
This issue of the Journal of Quality Technology pre- measured by 6(1), whereas Cpk incorporates the loca-
sents four articles that advance our understanding of tion of the process as well as its spread.
capability indices and propose new methods for pro- Johnson (1992) provides additional interpretation
cess capability analysis. The four articles are as fol- for Cpm by showing that it is related in a straightfor-
lows: ward way to the expected relative loss L. for the pro-
• "The Relationship of Cpm to Squared Error Loss" by cess. Johnson defines L. as the ratio between the ex-
T. Johnson pected squared error loss (which is fundamental to
• "Distributional and Inferential Properties of Process the Taguchi approach to quality improvement) and
Capability Indices" by W. L. Pearn, S. Kotz, and the value that the product is worth when the process
N. L. Johnson mean IL is equal to its target T.
• "Bootstrap Lower Confidence Limits for Capability The PKJ article points out that the squared error
Indices" by L. A. Franklin and G. S. Wasserman loss function is almost always chosen because it sim-
• "Confidence Bounds for Capability Indices" by plifies mathematical derivations and not because it
R. H. Kushler and P. Hurley describes available loss data. These authors present
In the introduction that follows, the term "article" is an example demonstrating that the choice of loss
reserved for these four articles to distinguish them function can make a subtle but critical difference in
from other references. For brevity, the article by the assessment of capability; see their section titled
Pearn, Kotz, and Johnson is referred to as the PKJ "Loss Functions".
article. Since some of the articles cover more than one
topic and since the articles overlap in their coverage, Distributional Properties of Capability Indices
they are introduced here by topic. The PKJ article begins with a more traditional in-
terpretation of capability indices as surrogates for the
Loss Function Interpretations for Capability Indices proportion of conforming output (Le., the proportion
The article by Johnson (1992) deals with the ca- within the specification limits). The PKJ section titled
pability index "Varieties of PCl's" presents useful relationships
among standard capability indices, and it reviews the
relationship between capability indices and the pro-
portion of nonconforming output. A key point made
here-and one that is frequently misunderstood-is
Here USL and LSL denote the upper and lower speci- that Cpk does not uniquely determine the proportion
fication limits for the process, IL and (1 denote the mean nonconforming. This section also points out draw-
and standard deviation of the process distribution backs of some recently proposed modifications to
(assumed to be normal), and T denotes the target for standard indices.
the process. (Identical notation is used throughout
The PKJ section titled "Estimation of PCl's" dis-
this issue of JQT.)
cusses the expected values and variances of estimators
The index Cpm was proposed independently by for Cp , CpIc, and Cpm. Some of these results have been
Hsiang and Taguchi (1985) and Chan, Cheng, and derived elsewhere; see, for instance, Kotz, Pearn, and
Spiring (1988), and it has been studied by Beazley Johnson (1992) and Zhang, Stenback, and Wardrop
and Marcucci (1988), Boyles (1991), and Franklin (1990). Here, the presentation is exceptionally com-
and Wasserman ( 1992b). The fact that Cpm allows for prehensive, and the distributional techniques in the
differences between the mean IL and the target T dis- PKJ appendices will be of special interest to statistical
tinguishes this index from the two most widely used researchers.
indices,
One application of this work is that it provides the
C = USL - LSL basis for examining the performance of capability in-
p 6(1 dex estimators. To illustrate this point, consider the

Vol. 24, No.4, October J 992 Journal of Qualify Technology

178 ROBERT N. RODRIGUEZ

bias of Cllk, the estimator for Cpk' Given n independent

process measurements

with the same normal distribution (mean Il and stan-

dard deviation 00), Cpk is usually estimated as
~ . (U8L - % %- LSL)
C11k = mtn 38 ' 38 .

This is the estimator constructed by using the sample

mean
_ 1 n
X=-LXi
n i=1

in place of u, and the sample standard deviation 20.0

8 = V- 1
-
n - 1 i-I
± (Xi -%)2 FIGURE 2. Bias of Cpk for n = 100.

in place of a. Several authors have pointed out that

this estimator is slightly biased, that is, the expected
value of Cpk (denoted by E(C pk)) is not exactly equal
B( ~ {d
Cpk) = -1- - -
3bn - 1 00
yr;::;
-
lI'n
e->./2

to Cpk. Thus, natural questions to ask are how does

the bias depend on the sample size n, and for a given
n, what is the maximum bias over a reasonable range
of the unknown parameters Il and u? where
For any sample size and specification limitsLSLand
U8L, the bias

is a complicated function of Il and a. One can examine

this function with the help of equation (14) of PKJ, ). = n(1l - T)2100 2
which provides an exact expression for E(Cpk). Sub-
stituting this expression into the above equation gives d = (U8L - LSL)/2.
This expression is easily computed and plotted.
Figures 1 and 2 provide surface plots of the bias for
sample sizes n = 30 and n = 100, respectively. These
plots are fairly general since they display bias as a
function of the standardized differences diu and
III - Ml/u, where M = (U8L + LSL)/2. Clearly the
bias is much smaller in absolute value for n = 100
(although it is already quite small for n = 30). For
both sample sizes the bias is negative for small values
of diu and 11l-Mil a,
Likewise, one can build on the work of PKJ to ex-
amine the effect of sample size on the mean squared
error (M8E) of Cpk' Recall that M8E is the expected
value of the squared error loss in estimating Cpk by
CpI<, that is,
20.0
M8E(CpI<) = Var(CpI<) + [E(Cpk) - CpI<]2

FIGURE 1. Bias of Cpk for n = 30. = Var(CpI<) + [B(CpI<)]2.

Journal of Qualify Technology Vol. 24, No.4, October J 992

 RECENT DEVELOPMENTS IN PROCESS CAPABILITY ANALYSIS 179

a Taylor series approximation to construct 95% con-

fidence limits for CpIc as

C1!k[1 ± 1.96 V~
9nC pk + 2( n ~ 1)]'
Also see Porter and Oakland (1990, 1991). Zhang,
Stenback, and Wardrop (1990) propose similar,
slightly more complicated confidence limits that are
based on moment derivations.
A different approach is taken by Chou, Owen, and
Borrego (1990) who derive lower confidence limits
for Cpk by working with the exact distribution of
Cpk • This involves a bivariate noncentral t distribution
originally studied by Owen ( 1965); also see Chou and
2 0.0 Owen (1989) and the computer program by Guirguis
and Rodriguez (1992) that appears in this issue of
FIGURE 3. Mean Square Error of Cp k for n = 30. JQT. However, Table 5 of Chou, Owen, and Borrego
(1990) is based on the assumption that the quantities
Equation (21) of PKJ provides an exact expression
en
CPu = (USL - X)/3Sand = (X - LSL)/3S are the
same, and investigation by several authors, including
for the variance of Cpk that, together with the above
Franklin and Wasserman (1992a) and Kushler and
expression for B( Cpk ) , can be substituted into the for-
Hurley (1992), reveals that the tabulated intervals
mula for MSE. The resulting expression for MSE (not
are conservative when this is not the case. Here "con-
shown here) is also easily computed.
servative" means that the actual confidence is greater
Figures 3 and 4 provide surface plots of MSE for than the nominal 95% confidence level of the table
sample sizes n = 30 and n = 100, respectively. The and that the tabulated lower limits are lower than
surfaces are functions of the same parameters d/ (T necessary to achieve 95% confidence.
and I#L - M I/ (T displayed in Figures 1 and 2. The plots
The article by Kushler and Hurley (1992) is an in-
for MSE reveal that for fixed n, MSE increases with
depth study of various confidence intervals for Cpk and
d/ (T. Comparing the plots shows how MSE decreases
Cpm. These authors introduce additional methods for
as the sample size increases.
constructing interval estimates, and they compare the
The mean square error of Cpk is only one of several
performances of a variety of confidence intervals in-
ways in which the variability of the estimator can be
cluding those mentioned above. The comparison,
expressed. For example, one can construct displays
similar to Figures 3 and 4 for the quantities
VMSE(Cpk), Var(C pk ) , and VVar(Cpk).

Confidence Limits for Capability Indices Assuming

Nonnality
As an alternative to summarizing process capability
with a point estimator such as Cpk , one can report
confidence intervals for process capability indices.
Unfortunately, the statistical theory for constructing
such intervals is not completely developed except for
simple indices such as Cp (see Kane (1986)).
For more complex indices such as Cpk and Cpm, var-
ious authors have constructed approximate confidence
intervals by making various assumptions about the
distribution of the point estimator (in all of this work
it is assumed that the measurements themselves are 20.0

normally distributed). For instance, Bissell (1990)

assumes that the distribution of Cpk is normal and uses FIGURE 4. Mean Square Error of Cp k for n = 100.

Vol. 24, No.4, October J 992 Journal of Qualify Technology

180 ROBERT N. RODRIGUEZ

which is made by computing and plotting "miss rates" van Zwet (1992) that the narve bootstrap is valid when
based on a geometric argument, will be helpful to the statistic of interest is asymptotically normally dis-
readers attempting to decide which of the currently tributed. The asymptotic normality of Cpic has, in fact,
available interval estimates to use. been established by Chan, Xiong, and Zhang (1990),
but the relevance of this result to Cpk bootstrap appli-
Confidence Limits for Capability Indices Using the cations has not been explored.
Bootstrap
Robust Capability Indices
In contrast with the article by Kushler and Hurley
(1992), which assumes normality of the data, the ar- The article by PKJ discusses the problem of defining
ticle by Franklin and Wasserman (1992b) deals with a process capability index whose interpretation, unlike
bootstrap confidence limits for Cp , CpIc, and Cpm, which that of standard indices such as CpIc, is insensitive to
avoid the assumption of normality. For related work departures from normality. This article proposes in-
by the latter authors, see Franklin and Wasserman dices of the form
(1991), Wasserman, Mohsen, and Franklin (1991),
and Wasserman and Franklin (1992). C = USL -LSL
9
(Ju
The bootstrap is referred to as a resampling method
because it reuses the original sample of process mea- where (J is a constant chosen so that the probability
surements XI, X2, ... , s« to synthesize many additional
Pr[#l - (Ju < X < #l + Ou]
samples from which replicates of a statistic (such as
Cpd can be computed. These replicates are then used will be both close to one and as independent as possible
to estimate useful quantities such as confidence in- of the distribution of X. In particular, taking (J = 5.15
tervals or bias corrections that would otherwise re- results in a probability of 0.99 for a variety of distri-
quire knowledge of the sampling distribution of the butions including the normal.
statistic. See Diaconis and Efron (1983) for a very
readable, non-mathematical introduction to resam- Note that a "robust capability index" is not the same
pling ideas and Efron and Tibshirani (1986) for a re- as a robust estimator for a standard index. In order to
view of bootstrap applications to measures of statis- guard against outliers in process data for which nor-
tical accuracy. Also see the recent discussion by Gun- mality is otherwise a reasonable assumption, one can
ter (1991-1992). estimate Cpk as

The article by Franklin and Wasserman (1992b) USL - X X-LSL)

compares the performances of three bootstrap confi- min ( 4.45MAD' 4.45MAD
dence limits with normal theory confidence limits. where
Three kinds of data distributions are simulated: nor-
mal, symmetric with heavy tails, and skewed. Al- x = median (XI ,X 2 , ..• ,Xn )
though the results demonstrate the advantages of the MAD = medianf lZ, -XI, ... , IXn -XI).
bootstrap in working with non normal process distri-
butions, Franklin and Wasserman (1992b) are careful Here, the median is used as a robust estimator for #l
to point out that the capability index one is attempting (a trimmed mean or a Winsorized mean could be used
to estimate may not be meaningfulfor nonnormal process instead), and the median absolute deviation MAD (di-
data. Thus the practical significance of this study is vided by 0.6745) is used as a robust estimator for a.
not that it eliminates the need for assuming normality, Of course, a robust capability index can also be esti-
but rather that it provides an understanding of the mated robustly!
behavior of normal theory methods when mild non- It should also be noted that the robust capability
normality is encountered. indices proposed by PKJ are multiples of Cp and have
The version of the bootstrap applied by Franklin very similar interpretations when tne-data are normally
and Wasserman (1992b) is the so-called naIve boot- distributed. For example, since C5 .15 = 1.17Cp , one can
strap, meaning that the resample size is the same as paraphrase the interpretation of Cp given by Ekvall
the original sample size n. Readers interested in the and Juran (1974) to classify values of C5 .15 less than
technical aspects of the bootstrap should be aware 1.17 as "not adequate", values between 1.17 and 1.56
that this method is presently the subject of consid- as "adequate", and values greater than 1.56 as "more
erable research to determine general conditions under than adequate". On the other hand, the practical in-
which it fails or succeeds. It has been suggested by terpretation of C5 .15 is not apparent for data distribu-

Journal of Qualify Technology Vol. 24, No.4, October 1992

 RECENT DEVELOPMENTS IN PROCESS CAPABILITY ANALYSIS 181

tions that are highly skewed, and even for normal matrix A. Clearly, these are unrealistic assumptions,
distributions the interpretation suffers from the same but as indicated by PKJ, the distribution theory be-
limitations as C/ relative to Cpk and Cpm.
J comes complicated when these assumptions are re-
There does not yet exist (to my knowledge) a truly moved.
nonparametric index that characterizes process ca- The problem of summarizing multivariate capabil-
pability for any given distribution and can be esti- ity has practical appeal in situations where multivar-
mated accurately with practical sample sizes. Indeed, iate control charts are being applied. However, con-
this is not a reasonable goal for a single-number sum- sidering the known limitations of capability indices
mary. However, one step in this direction is the "flex- in the univariate situation, it appears impossible to
ible capability index" proposed by Johnson, Kotz, and characterize multivariate capability with a single
Pearn (1992) which is presently being investigated value. Also, as noted by PKJ, it is difficult to imagine
by these authors. that the specification limits will be provided as an
Extensions of Univariate Process Capability Indices ellipsoid even though this makes sense from a statis-
tical standpoint. It is much more likely that the limits
The article by PKJ proposes a process capability will be provided as a v-dimensional box defined by a
index Cpmk defined as series of inequalities

c.; = C p) VI + (IL : Tf . LSLi :s; Xi :s; USL i, i = 1, ... , v.

In this case, a reasonable approach is to estimate the
This is referred to as a "third generation" process ca- proportion nonconforming by estimating the mean
pability index since it is constructed from Cpk and Cpr" vector and covariance matrix and then integrating
in the same way as these so-called "second generation" the corresponding multinormal density over the com-
indices are constructed from Cp (the "first generation" plement of the box.
index). The motivation for Cprllk is increased sensitivity
to departure of the process mean IL from the target T. Other Methods for Process
The interpretation of Cpmk, like that of its ancestors, Capability Analysis
requires that the data be normally distributed, and Having set the stage for the articles in this issue of
the PKJ article derives the first two moments of Cpmk JQT, it is worth discussing a number of topics that
under this assumption. they do not cover. The purpose of this section is not
A similar index was discussed earlier and indepen- to provide a comprehensive survey of methods for
dently by Lupienski (1989) who refers to it as Cpt. process capability analysis, but rather to focus on sev-
This index is discussed by Steenburgh (1989, 1990) eral methods that are misused or underutilized.
without reference to its distributional properties. Capability Indices Based on Systems of Distributions
Capability Indices for Multivariate Data Recently there have been various attempts to ex-
The article by PKJ discusses the problem of con- tend the definitions of standard capability indices to
structing and estimating suitable process capability nonnormal distributions (see, for instance, Clements
indices for multivariate process data. As in Chan, (1989)). These extensions use the fact that for the
Cheng, and Spiring (1990), it is assumed that v char- normal distribution the quantity 30' is both the dis-
acteristics Xl, X2 , ••• , X. are measured for each of n tance from the median (mean) to the upper 99.87
items, and that the specification limits for these char- percentile and the distance from the lower 0.13 per-
acteristics are prescribed in the form centile to the median (mean). The idea is to generalize
the formulas for standard indices by replacing 30' with
(X - T)'A -leX - T):s; c2
one or the other of these distances as appropriate.
where X' = (XII' .. ,X.), T' = (Til' .. , T.), and A is In particular, writing Pa for the 100a percentile, the
a v X v positive definite matrix. For v = 2 the speci- formula for Cpk can be generalized as
fication region is an ellipse in the (Xl, X2 ) plane, and
. ( USL - PO.5
for v > 2 the region is a v-dimensional ellipsoid. , P LSL )
O.5 -
Cpk = mzn .
PO.9987 - P O.5 P O.5 - PO.OO I3
The PKJ article proposes multivariate analogues
.C~ and .C;'" of the indices c, and Cpm for the case con- This extension applies to skewed distributions (es-
sidered by Chan, Cheng, and Spiring (1990) where it pecially if they are unimodal and fairly smooth), and
is assumed that the mean of X is the known target it simplifies to the familiar expression for Cpk when
vector T, and the covariance matrix of X is the known the distribution is normal. Incidentally, some of the

Vol. 24, No.4, October 1992 Journal of Quality Technology

182 ROBERT N. RODRIGUEZ

proposed extensions retain the mean in the original • One is limited to working with chi squared goodness-
definition of C1'k rather than replacing it with the me- of-fit tests since more powerful tests are not gen-
dian Por,. This complicates the interpretation of the erally available for the Pearson and Johnson systems
extended index since the mean can be considerably (see the section on "Goodness-of-Fit Methods").
larger (or smaller) than the median for heavily • The black box approach tends to displace basic
skewed distributions. practices, such as plotting the data and applying
As noted by PKJ, the basic disadvantage of percen- simple normalizing transformations, that provide
tile-based extensions is that the extreme percentiles genuine understanding of the process.
PO,OOl 3 and PO,9987 cannot be estimated precisely unless Capability Indices Based on Families of Distributions
a massive number of process measurements are avail-
Equations for fitted Pearson or Johnson curves are
able. This problem can be circumvented if one as-
generally complicated and difficult to interpret. This
sumes a parametric distribution for the data and com-
is not necessarily a disadvantage if the sole objective
putes the percentiles from the fitted distribution. Fur-
is to estimate a process capability index. However,
thermore, by working with distributions from a large
simpler distributions chosen from a smaller family of
class such as the Pearson system or Johnson system,
distributions such as the gamma, lognormal, or Wei-
one can fit data distributions with a wide variety of
bull can provide useful models for the process distri-
shapes. See Rodriguez (1983) for a review of systems
bution as well as reasonable curve flexibility. (Note
of distributions.
that the gamma family is the Pearson Type III family,
This approach should be regarded with considerable whereas the lognormal family is the Johnson SL family;
caution, especially when it is implemented by a black the Weibull family does not belong to either system.)
box computer program used to analyze large numbers
Working with a family of distributions has the fol-
of processes. Some of the potential difficulties are as
lowing additional benefits.
follows.
• Parameter estimation is relatively stable and
• Within a system of distributions, some distributions
straightforward with the method of maximum like-
are more difficult to fit than others. The method of
lihood.
moments and the method of percentiles, which are
often used to fit Pearson and Johnson curves, can • The method of maximum likelihood yields asymp-
totic variances for parameter estimates, and it is
yield unstable or inefficient estimates for curve pa-
amenable to the construction of confidence limits
rameters in some cases. Unfortunately, these cases
for extended capability indices (although, as far as
depend on the unknown parameters themselves.
I know, this has not been attempted).
• Unless the estimation technique is applied skillfully,
• Goodness-of-fit tests based on the empirical distri-
one can obtain fitted curves that are meaningless
bution function (see the next section) are available
over certain ranges of the data. For example, with
for various families including the gamma, lognormal,
the method of moments, an easily made mistake is
and Weibull.
to fit a Pearson Type III distribution whose estimated
threshold is less than the lower bound for the process
output, thereby invalidating the estimates of PO.OO13
and c.; 15.0
-

-
• The method of moments and the method of percen- Sur1wMIy
N 1225
tiles do not yield estimates of the variability or bias 12.5
f- 1.415
in the estimated indices. Likewise these methods do r--
811:10.. 0.41
10.0 "Obo < LSL 2.2lI
not yield confidence intervals for indices. "Obo > USL 0.1

• Not every data distribution can be described ade- - .--

quately with a Pearson or Johnson curve. Figure 5 f-
r-t-
shows a histogram for the effective channel length -
5.0
of 2500 field effect transistors. Although the first four f- -
moments of this distribution can be matched exactly 2.5

by those of a Pearson or Johnson curve, neither

curve will capture the three peaks that are evident 0.0 r ILl
0.48 O.llll 1.44 1.l12
in the data distribution (this example is discussed - - a.nn.I lMlgIh (rNcrone)
further in the section on "Kernel Density Esti-
mates"). FIGURE 5. Histogram for Effective Channel Length.

Journal of Quality Technology Vol. 24, No.4, October 1992

 RECENT DEVELOPMENTS IN PROCESS CAPABILITY ANALYSIS 183

Am Lognormal An8lysls Goodness-of-Fit Methods

50
- l.ogr1onMI Curve Whether a process capability analysis is based on a
A ~Cpt 0.478 normal distribution or some other parametric distri-
40
/ Pr > 011- Squore 0.04
bution, a goodness-of-fit test should be used to check

1\
p, > A-aqu.. 0.34
p, > 0
p, > W- Squore
0.12
0.28
the adequacy of the distributional model. The statis-
30
tical tests recommended for this purpose include those
J
I \ based on the empirical distribution function (EDF):

-.
20
the Kolmogorov-Smirnov test, the Cramer-von Mises
test, and the Anderson-Darling test. For a thorough
10
<, discussion see d'Agostino and Stephens ( 1986).

0
I ) '--t-- In addition to greater statistical power, one advan-
tage of EDF tests over the chi squared goodness-of-fit
-7.5 7.5 22.5 ~.5 52.5 87.5 S2.5
-'~ilmm test is that they do not depend on how the data are
binned. Figures 6 and 7 illustrate this point with dif-
FIGURE 6. Lognormal Fit with First Set of Midpoints. ferent histograms using the same set of n = 30 process
measurements for which a lognormal model is being
assessed. In Figure 6, the P-value for the chi squared
• Estimated quantities such as percentiles, extended test is 0.04, so on this evidence the lognormal model
indices, and the proportion nonconforming are is rejected at the 0.05 level of significance. In Figure
straightforward to express in equation form and to 7, with a slightly different set of midpoints for the
compute. bins, the P-value is 0.28, so the lognormal model is not
rejected. On the other hand, regardless of how the
To illustrate this last point, consider Figure 6 which midpoints are chosen, the EDF tests indicate that the
displays a histogram for n = 30 measurements super- data support the lognormal model. The P-values are
imposed with a fitted two-parameter lognormal den- 0.12 for the Kolmogorov-Smirnov test, 0.26 for the
sity curve. The density function is Cramer-von Mises test, and 0.34 for the Anderson-
Darling test.
1
j(x)= ,rn-exp(-[log(x)-rJ2/2u2), x>O A capability analysis based on the lognormal model
xv21l"u
is given in the previous section. Alternatively, one
where U denotes the shape parameter and denotes r can apply the EDF tests for normality to the loga-
the scale parameter. The fitted curve corresponds to rithms of the data and then compute standard capa-
the maximum likelihood estimates u = 0.491 and bility indices from the transformed measurements.
t = 3.36. An advantage is that one can apply the normal theory
methods discussed in the first section of this article.
The percentiles of the lognormal distribution are

Second Lognormal An8IysIs

where 4>-1 denotes the inverse standard normal cu-
50
mulative distribution function. It follows that PO.OO I3

I~
-~Curv8

= e t - 3a , PO.5 = e t , and Po 9987 = e t +3a • Consequently, the GenenlIDd Cpt 0.478

40 p, > QlI- Squar8 0.28
extended capability index for the lognormal distri- Pr > A- SqI.8rs 0.34
Pr> 0 0.12
bution function can be written as p, > W- Squar8 0.28
30

J
I
. (USL e- t - 1 1 - LSL e- t )
_
Cpk - mm e3 a - l ' 1-e -a3 '
Substituting LSL = 15, USL = 75, and the parameter
20

10
\ ""
estimates into this expression yields = 0.478. The c; ......
estimated proportion nonconforming is 0
~
0 15 30 45 eo 75
4>((log(LSL) - t)/u) +1- 4>((log(USL) - t)/u) - . 0lI8et iI mm

which is equal to 0.118. FIGURE 7. Lognormal Fit with Second Set of Midpoints.

Vol. 24, No.4, October 1992 Journal of Quality Technology

184 ROBERT N. RODRIGUEZ

However, one must be prepared to report the results

in the log scale, and this is not always practical. 2.50.r---------------,

Quantile-Quantile Plots
2.25
Distributional adequacy can also be checked graph- S1Iu1up Sample
ically using probability plots and theoretical quantile- N 40
quantile (Q-Q) plots (see d'Agostino and Stephens [ aoe Mean 1.51
SId D8v 0.41
(1986)). Both types of plots can be constructed for - 1.75 CpA: 0.58
nonnormal distributions as well as normal distribu-
tions. Quantile-quantile plots are not as well known j 1.50

~
to process engineers as probability plots, but they offer
advantages that merit their application in capability 1.25
analysis. Furthermore, there is a useful connection
between the normal Q-Qplot and CpIc that is highlighted 1.00
below (also see Rodriguez ( 1986) ) . o 0
The first step in constructing a Q-Q plot is to sort 0.75"'0-----'-'......--,---,.--......,..--..---...,.
the measurements Xl, ... , X" into increasing order, -0.75 -0.50 -0.25 0.00 0.25 0.50 0.75
resulting in the ordered values Cpt Scale

XCI) :s; X(2) :s; ... :s; xc"),

FIGURE 9. Normal Q-Q Plot Rescaled for Estimation of
Each point on the plot corresponds to a measurement. Cpl<'
The y-coordinate of the i t h point (viewed from left to
right) is simply XCi), the i th largest value. The z-co-
ordinate of the i t h point is computed as ct>-I(i/n). ear pattern is evidence that a normal distribution with
Figure 8 illustrates a normal Q-Q plot for n = 40 some unknown mean p. and standard deviation u ad-
values that were simulated from a normal distribution. equately describes the data. The slope of the line es-
Here, the point pattern is linear. Except for sampling timates a, and the intercept of the line estimates p.. It
variation, the quantiles of the data distribution (which is straightforward to estimate these quantities from
are the ordered values themselves) vary linearly with the plot since both axes are scaled linearly (this is not
the theoretical normal quantiles. With real data, a lin- so easy with a probability plot).
In a capability analysis, this interpretation can be
taken one step further. The horizontal axis (which is
scaled from -3 to 3) indicates the distance between
a measurement and the process average in standard
2.50..-----------------,
deviation units (multiples of u). This applies to spec-
o ification limits as well as measurements. For instance,
2.25 0
S1Iu1up Sample 0 an upper specification limit that is two standard de-

I ~oo
N
Mean
SId D8V
1.51
0.41
40

tSJ
iii
afl
cfI
viations above the mean will intersect with the linear
pattern at a point whose z-coordinate is 2. This dis-
tance (divided by three) is, in fact, the value of the
- 1.75 one-sided capability index
0

j 1.50 rI CPU = (USL - p.)/3u.

i 1.25

0
~
/ Likewise, the standardized distance ( divided by three)
between a lower specification limit and the mean is
1.00 0
00 CPL = (p. - LSL)/3u.
0
0.75 "ty----.--..----,-----,,....--.,-----.,J Consequently, provided that one rescales the hori-
-3 -2 -1 0 1 2 3 zontal axis of a normal Q-Q plot by a factor of three,
NormII au.1IIIee visual estimates of CPU and CPL can be obtained di-
rectly from the plot. The Q-Q plot in Figure 9 is a
FIGURE 8. Normal Q-Q Plot. rescaled version of the plot in Figure 8. By dropping

Journal of Quality Technology Vol. 24, No.4, October J992

 RECENT DEVELOPMENTS IN PROCESS CAPABILITY ANALYSIS 185

perpendicular lines from the points of intersection estimate does depend on the choice of a bandwidth
between the linear point pattern and the specification parameter. To a large extent this choice can be au-
limits, one reads estimates of CPU and CPL from the tomated, but it is good practice to experiment with
horizontal axis as 0.73 and 0.58, respectively (the neg- several bandwidth selections when one is exploring
ative sign for -0.58 is ignored). Then since a set of data. This is why density estimates for three
different bandwidth parameters (denoted as c) are
Cpk = min(CPU, CPL)
illustrated in Figure 10. Second, one might ask
it follows that a visual estimate of Cpk is 0.58. whether the density estimate can be integrated outside
Rescaled normal Q-Q plots provide accurate visual the specification limits to estimate the proportion
estimates of Cpk only if the point pattern is highly lin- nonconforming. The answer is that such an estimate
ear, which in turn is evidence of normality that jus- will be biased and offers no advantage over reporting
tifies the interpretation of the index. Of course, such the observed proportion of nonconforming measure-
displays cannot replace statistical goodness-of-fit ments (in Figure 5 this proportion is 0.0229 + 0.0016
measures and computed indices, and their value is = 0.0245).

that they enable the user to see the relationship be- Comparative Histograms
tween Cpk and the assumption of normality. Con-
Every process capability study should attempt to
versely, however, nonlinear point patterns should not
understand the sources of variability in the process.
be ignored simply because they do not support the
A basic but highly effective graphical tool for this
use of capability indices. Some nonlinear patterns in-
purpose is the comparative histogram.
dicate the presence of outliers, while others provide
clues leading to more adequate distributional models. In the analysis of channel length displayed in Figure
Helpful diagnostic rules are given by Chambers et al. 10, the discovery of the peaks in the data distribution
(1983) and Fowlkes (1987). prompted a process engineer to ask whether they were
due to different tools, operators, or raw materials. In-
Kernel Density Estimates vestigation indicated that the transistors had come
The graphical method most commonly used for ca- from three different lots, each produced by a different
pability analysis is the histogram. Yet, many users are tool, and subsequently individual histograms were
unaware that the interpretation of a histogram de- constructed for the measurements from each lot.
pends on how the data are binned, even when the It is helpful to arrange these histograms in a single
bins have been determined with a good statistical al- display with a uniform axis. This display, shown in
gorithm (for instance, the method of Terrell and Scott Figure 11 and referred to as a comparative histogram,
(1985)). This problem can be overcome by super- reveals that the tool producing Lot 2 is the most ca-
imposing (or replacing) the histogram with a smooth pable, and that the data for each lot are normally dis-
nonparametric estimate of the probability density tributed. Thus, it is reasonable to report a capability
such as a kernel density estimate or a spline estimate. index for each lot.
Very readable introductions to these techniques are
given by Silverman (1986) and Eubank (1988). Here,
the point to be made is that nonparametric density
estimation can playa valuable role in process capa- 15.0
DonIIIIy ElIIIrn-.
rr:
bility analysis, particularly when one encounters
nonnormal data.
12.5
If/~ - - - I e ) 0.4
....· - - l e ) 0.7
----let 1
~,\
r~t~#
Figure 10 shows three kernel density estimates for 10.0 '

the distribution of channel length displayed in Figure

5. The density estimates help the user visualize the
three peaks in the distribution. Features such as these
I 7.5
~/ ~ ~\~rI ~~.\
5.0
are especially difficult to discern from a histogram
based on a small sample because peaks and valleys f.~
are prone to be artifacts of sampling variation or the
choice of bins.
2.5

0.0
0.12
!J
,
~,~#

0.84 1.ll8
Fi
2.2lI
"
3.110
Two technical points regarding kernel density es- ~ a.moI l.8ngIh (mk:larw)
timates are worth mentioning. First, the term "non-
parametric" is slightly misleading because the density FIGURE 10. Kernel Density Estimates for Channel Length.

Vol. 24, No.4, October 1992 Journal of Qualify Technology

186 ROBERT N. RODRIGUEZ

CHOU, Y. and OWEN, D. B. (1989). "On the Distribution of the

~i~U1J:Iik 'C~-Tl
Estimated Process Capability Indices". Communications in Sta-
tistics-Theory and Methods 18, pp. 4549-4560.
CHOU, Y.; OWEN, D. B.; and BORREGO, S. A. (1990). "Lower Con-
fidence Limits on Process Capability Indices". JournalofQuality
~

A rrfl
0

i~i~U
Technology 22, pp. 223-229.
CLEMENTS, J. A. (1989). "Process Capability Calculations for
Non-Normal Distributions". QualityProgress 22, pp. 98-100.
D'AGOSTINO, R. B. and STEPHENS, M. A. (editors) (1986). Goodness-

~i~M ~
of-Fit Techniques. Marcel Dekker, Inc., New York, NY.
DIACONIS, P. and EFRON, B. (1983). "Computer-Intensive Methods

oU
in Statistics". Scientific American 248(5), pp. 116-130.

0.6
i
0.76 0.92 1.08 1.24 1.4
i-dilUlhlJ
1.56 1.72 1.66 2.04 2.2 2.36
EFRON, B. and TIBSHIRANI, R. (1986). "Bootstrap Methods for
Standard Errors, Confidence Intervals, and Other Measures
EII8cIIve Ch8moI I.a1gIh (m1aono) of Statistical Accuracy". StatisticalScience 1, pp. 54-77.
EKVALL, D. N. and JURAN, J. M. (1974). "Manufacturing Plan-
FIGURE 11. Comparative Histogram for Channel Length. ning" in Quality Control Handbook, Third Edition, (J. M. Juran,
ed.). McGraw-Hill, New York, NY.
EUBANK, R. L. (1988). SplineSmoothing and Nonparametric Regres-
More importantly, focusing on tool differences rep- sion. Marcel Dekker, Inc., New York, NY.
resents an opportunity for process improvement that FOWLKES, E. B. (1987). A Folio of Distributions: A Collection of Theo-
might have been missed if the analysis had been based retical Quantile-Quantile Plots. Marcel Dekker, Inc., New York,
on one overall capability index accompanied by a NY.
simple histogram. Comparative histograms provide FRANKLIN, L. A. and WASSERMAN, G. (1991). "Bootstrap Confi-
visual clarity and are readily understood by operators, dence Interval Estimates of Cpk: An Introduction". Communi-
cationsin Statistics-Simulation and Computation 20 ( 1), pp. 231-
process engineers, and managers. They should be used
242.
whenever stratification is suspected and to examine
FRANKLIN, L. A. and WASSERMAN, G. S. (1992a). "A Note on the
the results of improvements in process capability.
Conservative Nature of the Tables of Lower Confidence Limits
Acknowledgments for Cpk With a Suggested Correction". Communications in Sta-
tistics-Simulation and Computation, to appear.
I would like to thank Jason Brown and J. Stephen FRANKLIN, L. A. and WASSERMAN, G. S. (1992b). "Bootstrap Lower
Marron of the Department of Statistics, University of Confidence Limits for Capability Indices", Journal of Quality
North Carolina, for helpful discussions concerning the Technology 24, pp. 196-210.
bootstrap and kernel density estimation, respectively. GUIRGUIS, G. H. and RODRIGUEZ, R. N. (1992). "Computation of
Owen's Q Function With Applications to Process Capability
References Analysis", Journal of QualityTechnology 24, pp. 236-246.
BEAZLEY, C. E. and MARCUCCI, M. D. (1988). "Capability Indices: GUNTER, B. (1989). "The Use and Abuse of Cpk: Parts 1-4".
Process Performance Measures", ASQC QualityCongress Trans- Quality Progress, January 1989, pp. 72-73; March 1989, pp. 108-
actions, pp. 516-523. 109; May 1989, pp. 79-80; July 1989, pp. 86-87. Also see "The
BISSELL, A. F. (1990). "How Reliable is Your Capability Index?". Use and Abuse of Cpk Revisited", January 1991, pp. 90-94.
AppliedStatistics 39, pp. 331-340. GUNTER, B. (1991-1992). "Bootstrapping: How to Make Some-
BOYLES, R. A. (1991). "The Taguchi Capability Index". Journal thing From Almost Nothing and Get Statistically Valid An-
of Quality Technology 23, pp. 107-126. swers". QualityProgress, Decem ber 1991, pp. 97-103; February
CHAMBERS, J. M.; CLEVELAND, W. S.; KLEINER, B.; and TUKEY, 1992, pp. 83-86; April 1992, pp. 119-122; June 1992, pp. 79-
P. A. (1983). Graphical Methodsfor Data Analysis. Duxbury Press. 83.
Boston, MA. HSIANG, T. C. and TAGUCHI, G. (1985). "A Tutorial on Quality
CHAN, L. K.; CHENG, S. W.; and SPIRING, F. A. (1988). "A New Control and Assurance-The Taguchi Methods". Unpublished
Measure of Process Capability: Cpm". Journal of Quality Tech- presentation given at the Annual Meetings of the American
nology 30, pp. 162-175. Statistical Association, Las Vegas, NV.
CHAN, L. K.; CHENG, S. W.; and SPIRING, F. A. (1990). "A Mul- JOHNSON, N. L.; KoTZ, S.; and PEARN, W. L. (1992). "Flexible
tivariate Measure of Process Capability". Journal of Modeling Process Capability Indices". Institute of StatisticsMimeo Series,
and Simulation 11, pp. 1-6. No. 2072. Department of Statistics, University of North Car-
CHAN, L. K.; XIONG, Z.;and ZHANG, D. (1990). "On the Asymptotic olina, Chapel Hill, NC.
Distributions of Some Process Capability Indices". Commu- JOHNSON, T. (1992). "The Relationship of Cpm to Squared Error
nicationsin Statistics-Theory and Methods. 19, pp. 11-18. Loss", Journal of Quality Technology 24, pp. 211-215.

Journal of Qualify Technology Vol. 24, No.4, October J992

 RECENT DEVELOPMENTS IN PROCESS CAPABILITY ANALYSIS 187

KANE, V. E. (1986). "Process Capability Indices". Journal oj of the Japanese Societyfor Quality Control 16, pp. 55-65. In Jap-
QualityTechnology 18, pp. 41-52. Corrigenda p. 265. anese.
KITSKA, D. (1991). Letter to the Editor, QualityProgress, March SILVERMAN, B. W. ( 1986). DensityBstimaiionfor Statisticsand Data
1991, p. 8. Analysis. Chapman and Hall, New York, NY.
KOTZ, S.; PEARN, W. L.; and JOHNSON, N. L. (1993). "Some Process STEENBURGH, T. J. (1989). "A Method to Study Process Capability
Capability Indicators Are More Reliable Than One Might Think for Manufacturing Strike Teams". Unpublished Xerox Cor-
(A Comment on Bissell (1990))". Apptied Statistics, to appear. poration technical report.
KUSHLER, R. H. and HURLEY, P. (1992). "Confidence Bounds for STEENBURGH, T. J. (1990). "Process Capability Indices for Tar-
Capability Indices". Journal of Quality Technology 24, pp. 188- gets". Unpublished Xerox Corporation technical report.
195. TERRELL, G. R. and SCOTT, D. W. (1985). "Oversmoothed Non-
LUPIENSKI, J. A. (1989). "Taguchi Quality Loss Function and Its parametric Density Estimation". Journal oj the AmericanSta-
Relationship to Process Capability Centering (C".)", Forty-fiJth tisticalAssociation 80, pp. 209-214.
Annual ASQC Quality Control Conference Transactions-Rochester VAN ZWET, W. (1992). "Resarnpling". Wald Memorial Lectures
Section. presented at the Annual Meeting of the Institute of Mathe-
OWEN, D. B. (1965). "A special case of a bivariate non-central matical Statistics. Annals oj Statistics, to appear.
t-distribution". Biometrika 52, pp. 437-446. WASSERMAN, G. S.; MOHSEN, H. A.; and FRANKLIN, L. A. (1991).
PEARN, W. L.; KOTZ, S.; and JOHNSON, N. L. (1992). "Distributional "A Program to Calculate Bootstrap Confidence Intervals for
and Inferential Properties of Process Capability Indices". the Process Capability Index, C". ''. Communications in Statistics-
Journal oJQuality Technology 24, pp. 216-231. Simulation and Computation 20, pp. 497-510.
PORTER, L. J. and OAKLAND, J. S. (1990). "Measuring Process WASSERMAN, G. S. and FRANKLIN, L. A. (1992). "Standard Boot-
Capability Using Indices-Some New Considerations". Quality strap Confidence Interval Estimates of C".. Computers and In-
and Reliability Engineering International 6, pp. 19-27. dustrialEngineering 22(2), pp. 171-176.
PORTER, L. J. and OAKLAND, J. S. (1991). "Process Capability ZHANG, N. F.; STENBACK, G. A.; and WARDROP, D. M. (1990). "In-
Indices-An Overview of Theory and Practice". Quality and terval Estimation of Process Capability Index C".". Commu-
Reliability Engineering International 7, pp. 437-448. nicationsin Statistics-Theory and Methods 19, pp. 4455-4470.
RODRIGUEZ, R. N. (1983). "Systems of Frequency Curves" in
Encyclopedia oj Statistical Sciences, Vol. 3. (S. Kotz and N. L. Key Words: Bootstrap Methods, Capability Indices, Con-
Johnson, eds.) pp. 212-225. fidence Intervals, Goodness-ofFitMethods, Graphical Methods,
RODRIGUEZ, R. N. (1986). "Some Graphical Methods for Analyz- Kernel Density Estimates, Process Capability Analysis, Sta-
ing Data in Statistical Quality Control Applications". Journal tistical Process Control.

----- -., -----

Vol. 24, No.4, October 1992 Journal of Qualify Technology