MINI PAPER

Process Capability Indices

By Robert H. Mitchell
3M Company

Process capability indices have been popular for over

20 years, since Joseph Juran12 popularized the Capability
Ratio (Cr) in his Quality Control Handbook. Eagerness to
establish a single index to measure process capability has
resulted in the proliferation of indices: Cp/Cpk, Vp/Vpk,
Pp/Ppk, Cpm, Tz, %OOL, PPM, and more recently
Generalized Cp/Cpk. Abuse of these indices is well
documented (Gunter)8. Though simple to compute, they
can lead to scorekeeping by management, incorrect interpretations and tampering, with little or no product or
process improvement. This article reviews the basics of
estimating process capability and introduces new process
capability indices.

Values of Cp exceeding 1.33 indicate that the process is adequate to

meet the specifications. Values of Cp between 1.33 and 1.00 indicate that
the process, while adequate to meet specifications, will require close
control. Values of Cp below 1.00 indicate the process is not capable of
meeting specifications.

Cp is called the Process Potential - it simply relates

the Process Capability (6) to the Spec Range; it does not
relate the location of the process with respect to the specs.
Consider the three distributions in fig. 1, all from processes having a Cp of 1.00 or better.
Figure 1

Cp, Cpk
The original Capability Index published in Jurans
Quality Control Handbook is defined as Tolerance Width
divided by Process Capability.
Cp = Process Potential
= (Tolerance Width) / (Process Capability).
Juran defines Process Capability as six standard deviations
for a process in statistical control. Process Capability =
6 where is the in-control process standard deviation.
The easiest method to determine process standard deviation is from the control chart of a stable process: Process
Std. Dev. = R-bar/d2, where the appropriate value of d2 is
read from a table for the subgroup size.
For the Cp/Cpk index to be valuable, the tolerance
width, or specification range, should have real meaning,
i.e. be based on functional limits per known end-user
requirements (Voice of the Customer). Occasionally, for
lack of end-user requirements, one is asked to set specifications (specs) based solely on the producers process
capability. The danger is that a process aim and natural
variation can have little relation to customer-perceived
quality. How does one choose specs: use arbitrarily wide
limits (+/- 6s) to ensure good Cpk values? Use arbitrarily
narrow limits in an effort to lock out competition (and
unwittingly, perhaps yourself)? Neither strategy focuses on
the customer. Furthermore, focusing on specs alone asks,
What is the worst we can get away with? while emphasis
on target alignment and variability reduction asks, What
is the best we can do?
According to Quality Assurance for the Chemical and
Process Industries A Manual of Good Practices 2nd
Edition 1:
16

If the process is centered within the specs, and is approximately normal then Cp = 1.00 results in a fraction nonconforming (f.n.c.) of 0.27%. Note that Cp=1.00 doesnt
guarantee that there will be only 0.27% non-conforming
product. What it does guarantee is that, assuming
normality, a stable process centered with respect to the
specs, and the correct value of , there will not be less
than 0.27% of non-conforming product.

Cpk = Process Capability Index.

CPK is used to summarize how a process is running
relative to its spec limits. As with Cp, this measure is
appropriate only when the process is stable (in-control).
Cpk = minimum of

X bar - LSL, USL - X bar

3

where Xbar and are the mean and standard deviation of

the process. So, Cpk measures how far the process mean
is from the nearer spec limit in terms of 3 distances. For
processes with a one-sided spec the term corresponding
to the missing limit is omitted.
ASQ STATISTICS DIVISION NEWSLETTER, VOL. 18, NO. 1

MINI-PAPER
Continued from page 16
Special consideration must be made for non-normal distributions. Cpk works well only for the bell-shaped normal (Gaussian) distribution. For others it is an approximation. Unlike Xbar control charts there is no Central Limit
Theorem effect in estimating process capability because
Cp and Cpk relate to the distribution of individual items.
We need to be concerned about non-normal distributions.
Table 1 gives the fraction non-conforming (PPM) for different distributions, all having Cpk=1.
Table
Distribution
Chi-Square (4.5df)
Heavy-tailed (B2>3)
Uniform
Normal

formance based on total incoming variability. This is illustrated in figure 2.

Figure 2. Within subgroup Capability versus weighted average of both
the within and between spread (Performance).

f.n.c.
14,000
4,000
0
2700

One approach to dealing with non-normal data is to

transform the data. Typical transformations include taking
the reciprocal, reciprocal square root, natural log, or
square root of the raw data. The corresponding spec limits
must be similarly transformed. Box, Hunter, and Hunter5
offer a relatively simple method for determining a transformation to give constant variance.

Theoretical Cp, Cpk

Cp* and Cpk* are calculated when the process is not
stable, yet one desires to estimate how good the process
might be if no Special Causes existed. Theoretical Cp,
Cpk use a best estimate of the true process standard
deviation (sigma-hat). Special Causes are excluded from
the data when appropriate, to estimate the potential
natural process variation. A theoretical process sigma-hat
is calculated and Cp*/Cpk* estimated.
In general, I advise against this methodology because
of potential confusion. Although meant as a measure of
how good the process might be, one cannot predict since
the process is not stable. These are not true Cp, Cpk
values.

Pp, Ppk
Pp and Ppk are measures of process performance from
a customer perspective. Process capability measures short
term ability of a process to meet specs. Process performance measures long term ability of a process to meet
specs. Process Performance should be distinguished from
Process Capability. Pp is similar in definition to Cp, and
Ppk is similar in definition to Cpk but in each we use the
overall standard deviation of the data, including any
Special cause variation, rather than the short term estimate
of standard deviation. The overall standard deviation is a
weighted average of both within-group and betweengroup variation. As a customer, one wants to measure perASQ STATISTICS DIVISION NEWSLETTER, VOL. 18, NO. 1

Cpm (Cp-Taguchi)
The Cpm index was introduced in 1988 (Chan6). The
principal difference between Cpm and Cpk is the relative
importance of Conformance to Specs versus Run to Target.
Cpk measures how well the process mean is centered
within the spec limits, and what percentage of product
will be within spec. Instead of focusing on spec limits
Cpm focuses on how well the process mean corresponds
to the process target, which may or may not be midway
between the spec limits. Cpm is motivated by Taguchis
Loss Function. The denominator of Cpm includes the
Root Mean Square deviation from the target.
Cpm =

USL LSL
6 { 2 + (Xbar Tgt)2 }1/2

Cpk is preferred to Cp because it measures both

process location and process standard deviation. Cpm is
often preferred to Cpk because the variability term used in
the index is more consistent with Run to Target
philosophy.

Cr
Capability ratio (Cr)12, 13 is the inverse of Cp. If Cp =
1.33 or more is considered a capable process, then a value
0.75 or less is desired for Cr.
Cr =

6
USL

LSL

Tz
Target Z is a measure of targeting and is similar to, but
simpler to estimate than Cpm.
Tz =

(Xbar Target)

Values of Tz between 0.5 and 0.5 are considered good.

17

MINI-PAPER
Continued from page 17

%OOL
Percent Outside of Limit uses the z-statistics to estimate the porportion of a population that lies beyond the
spec limits. This measure assumes a Normal distribution.
For a two-sided specification, the ZLower and ZUpper proportions are calcuated from the process measn, standard deviation, and spec limits.
ZLower =
ZUpper =

Xbar

LSL

USL

Xbar

Then look up the proportions from any Z-table. One-sided

or nonsymmetrical specifications usually correspond to
heavily skewed distributions and can yield exaggerated
%OOL values.

PPM
Parts Per Million (defective) is similar to %OOL. One
multiplies the ZL and ZU proportions by 1,000,000 each,
then sum together. Again, these are theoretical estimates
based on the assumption of a normal distribution. Table 2
illustrates the relation between Cpk and PPM. If receiving
inspection is performed, then it is possible to compare
PPM-Observed (from receiving inspection) and PPMCalculated (from the Normal Distribution).
Table 2
Cpk
.333
.667
1.0
1.33
1.50
1.67
2.00

PPM and Cpk fraction non-conforming (f.n.c.)

PPM
317,400
45,500
2,700
63
7
0.6
2.0 PPB

*Note: This comparison of Cpk and PPM does not

include the 1.5 shift included in the Motorola Six
Sigma program.

GCpk, GPpk
Traditional Cpk can only be estimated from a stable
process; i.e., no Special Causes. Joe Voelkel19, at the 1998
Fall Technical Conference, introduced Generalized
Cp/Cpk (GCp/GCpk). Joe noted that there are two distinct types of Assignable (Special) cause variation, as discussed by Brian Joiner11 - Erratic cause and Structural
cause. Examples of erratic special cause are an untrained
operator, raw material variability, or an unknown process
shift. Tool wear, and multiple cavity tools are examples of
structural variation. GCpk is calculated from the fraction

18

non-conforming (f.n.c.) of a given process. Traditional

Cpk should only be estimated for stable, predictable
processes; Generalized Cpk is promoted for use in cases
where structural type of Special Causes is present, and
Ppk should be calculated for processes affected by erratic
Special Cause variation. Again, without stability there is no
prediction. Software to estimate Generalized Cp/Cpk is not
yet commercially available.

Additional Considerations
Structural Variation
Structure occurs when there are consistent, repeating
patterns in the data. The patterns can occur over time (e.g.
cycles, process deterioration), or within a subgroup (e.g.
multiple cavity tooling, fixed crossweb differences, etc.).
The effect of structural variation is exaggerated (wide)
control limits. Structural variation is usually fairly easy to
identify; points plotted on the X-bar or I chart cluster
around the centerline. The ideal solution to structural variation is to eliminate the structure; however, this requires a
process change, often difficult to achieve. If the structure
occurs within the subgroups the 3-Chart method
(Wheeler20) will limit the effects to the within-subgroup
Range chart. The preceding paragraph introduced erratic
and structural types of Special Cause variation. Stu Janis10
elaborated on the topic of structural variation for Moving
Web, Injection Molding, and batch processes. Stu
explained that the standard deviation used to determine
control limits should be based on the random portion of
variability. It should not include biases such as fixed differences between cavities in mold tooling or crossweb differences in a web. The fixed bias only comes to play in
determining the central line of a chart to control variability
across cavities or across a web. Fixed differences often
result in exaggerated control limits.
Multiple Sources of Random Variation
In the same paper Janis also explained the Space vs.
Time concept of random variability. The sources of variability affecting within-subgroups (i.e. cavity-to-cavity or
crossweb) are different than those affecting between subgroups (shot-to-shot or jumbo). Attempts to use withinsubgroup variation (space) to set control limits for
between subgroup averages (time) often result in limits
that are too narrow. 3-Way Control Charts (Don
Wheeler20) monitor within-subgroup variability (space)
using a Range chart, use a Moving Range chart to monitor
short-term between-subgroup variability (time), and an
Individuals charts to monitor differences between subgroup Means (time - longterm).
These same considerations may apply when estimating
the standard deviation of individuals for Cp, Cpk indices
of 3-Chart processes. If no bias exists, then the for
Continued on page 19
ASQ STATISTICS DIVISION NEWSLETTER, VOL. 18, NO. 1

MINI-PAPER
Continued from page 18
Individuals can be estimated using:
where d2k = the value for d2 using k number of sub groups

in the Moving Range, and

d2n = the value for d2 using n number of individuals within
a subgroup

Roadmap

Sample Size
Because process capability indices are determined from
estimates of standard deviation, they are affected by sample size (degrees of freedom). As expected, the stability of
estimates of the standard deviation increases with sample
size. We can show, using Chi-square tables and bootstrap
techniques7 , that a sample size (n) of 10 does not provide
a very stable estimate of process capability. Even when n
is 40 there is still substantial uncertainty in the estimator
of Cpk. Tables 3 and 4 provide estimates of 95%
Confidence Bounds for Cpk (lower bound) and Ppk (twosided interval), assuming normality:

Cpk
1.00
1.10
1.50
1.667

n=30
0.72
0.80
1.12
1.25

n=50
0.79
0.87
1.21
1.35

n=75
0.83
0.91
1.26
1.40

Table 415 95% Confidence Interval for Ppk

Ppk
n=30
n=60
1.00
.76-1.31
.83-1.21
1.33
1.02-1.76
1.11-1.61
1.67
1.29-2.19
1.49-2.01

n=120
.88-1.14
1.17-1.52
1.47-1.90

Formulas (and further discussion) to compute the 95%

Confidence Intervals for process capability indices are
shown in Montgomerys16 Introduction to Statistical
Quality Control.
Practitioners often forget that process capability indices
are merely point estimates. To avoid the pitfalls of making decisions using point estimates the concepts of
Statistical Thinking should be employed (variability exists),
and Cpk values plotted on control charts. A plot of Cpk
values assists in the detection of process deterioration (or
improvement).
Capability Studies
Process capability refers to the uniformity of the
process. Montgomery16 defines process capability analysis
as an engineering study to estimate process capability.
The AT&T Statistical Quality Control Handbook 3 defines
the process capability study as a Scientific systematic
procedure for determining the capability of a process
and defines capability as the predictable series of
effects produced by a process when allowed to operate
without interference from outside causes. The estimate
of process capability may be in the form of a probability
distribution having a specified shape, center, and spread.
For this definition a process capability analysis may
be performed without specs. i.e. Process Capability = 6.
Or, process capability may be expressed as a percent of
product outside spec limits. This type of capability study
ususally measures product functional performance, not the
process itself. When the engineer can directly observe the
process and can control the data collection methods this
study is a true process capability study (Montgomery).
When historical data is used and direct observation of the
process is not possible, Montgomery refers to this as a
product characterization study. In a product characterization study we can only estimate the distribution of the
product quality characteristics; we can say nothing about
the statistical stability of the process.

Table 37 Approximate 95% lower bound for Cpk.

Continued on page 20
ASQ STATISTICS DIVISION NEWSLETTER, VOL. 18, NO. 1

19

MINI-PAPER
Continued from page 19
There are three primary techniques used to estimate
process capability: histograms and probability plots,
control charts, and hierarchical (nested) designed
experiments.
Histograms (or stem-and-leaf plots) require at least 100
observations. If the data sequence is preserved, Mean
Square of Successive Differences (MSSD) can be used to
estimate the Short Term Standard Deviation (STSD). Or, an
estimate of process standard deviation can be obtained
from -hat=Rbar/d2.
The probability plot has an advantage over histograms
since it produces reasonable results for small sample sizes
(Montgomery). However, other statistical methods are
often needed to supplement the probability plots.
The control chart method is a simple, effective tool for
process capability analysis. The control chart is the preferred technique for process capability analysis because it
displays the potential capability of a process: patterns,
trends, and other Special Cause signals.
An important consideration with use of the control
chart method is selection of the proper rational subgroup
to estimate the common cause process variability, against
which special cause variability estimates are compared.
The hierarchical experimental design is a systematic
approach to document and quantify the sources of variability in a process, and aids in identifying variation reduction opportunities. The fully balance, nested model is generally preferred due to its ease of statistical analysis,
though unbalanced, staggered, and mixed models can be
more accurate.

Final Thoughts
I have just spent several pages introducing various
process capability indices. Nonetheless, I prefer to monitor
progress of continuous improvement efforts with control
charts rather than columns of process capability indices.
Bert Gunter8 eloquently listed problems associated with
focus on Cpk values:
1. Cpk cannot be used with one-sided specs or when the
process is not normal.
2. Because the sampling distribution of the Cpk statistic is
so variable it should not be used unless relatively large
sample sizes are obtained (100-200).
3. Cpk goals can be impossible to meet when measurement error is large. Recall that
2
2
2
Total = Product + Test error.
Reducing test error improves the Cpk value but does
not really improve the product. Conversely, any
process variability reduction without test method
improvements may not result in much larger Cpk
values.

20

4. Widening the product specs will result in a better

(bigger) Cpk value but do nothing to improve the
product or satisfy the customer.
5. Most importantly, Cpk is a meaningless measure of
process capability unless your process is in a state of
statistical control. Without statistical control a process is
not predictable.
Unfortunately, a simple to understand, easy to calculate
alternative to Cpk does not exist. Paraphrasing Bert
Gunter, we must exercise caution to prevent process
capability index scorekeeping from being confused with
real improvement.

Bibliography
1. ASQ Chemical & Process Industries Division,
Specifications for the Chemical and process Industries:
A manual for development and use, (ASQ Quality
Press), 1996.
2. ASQ Chemical & Process Industries Division, Quality
Assurance for the Chemical and Process Industries,
(ASQ Quality Press), 1997
3. AT&T Statistical Quality Control Handbook, (Delmar
Printing Co.), 1956.
4. Bailey, S. P., How to Deal with the process Capability
Catch-22, ASQ AQC Transactions, 1992.
5. Box, G. P., Hunter J. S., Hunter W. G., Statistics for
Experimenters, (Wiley) 1978.
6. Chan, L. K., Cheng, S.W. and Spiring, F.A. A New
Measure of Process Capability Cpm, JQT Vol. 20 No.
3., 1988.
7. Franklin, L. A. and Wasserman, G. S., Bootstrap Lower
Confidence Limits for Capability Indices, JQT Vol. 24
No. 4, October 1992.
8. Gunter, Bert, The Use and Abuse of Cpk, Quality
Progress magazine, Jan., Mar., May, July 1989.
9. Herman, J. T. Capability Index Enough for Process
Industries?
43rd ASQ AQC Transactions, 1988.
10. Janis, Stuart J., Is Your Process Too Good for Its
Control Limits?, 43rd ASQ AQC, 1988.
11. Joiner, B. L., Fourth Generation Management,
(McGraw-Hill) 1994.
12. Juran, J. M., Quality Control Handbook 4th Edition,
(McGraw-Hill) 1988.
13. Juran, J. M. and Gryna, F. M., Quality Planning and
Analysis 3rd Edition (McGraw-Hill) 1998
Continued on page 21

ASQ STATISTICS DIVISION NEWSLETTER, VOL. 18, NO. 1

NEW FTC
AGREEMENT
The three sponsors of the Fall
Technical Conference (Stat and CPI
Divisions of ASQ and SPES Section of
ASA) recently signed a new sponsorship agreement covering the next five
FTCs. There are a number of changes
to the agreement, changes intended
to improve the FTC, and thus provide
a better experience for members.
Changes have been made in site
selection procedures, vendor displays, Youden speaker selection, and
short course availability.
Some of the changes:
Site selection; CPID will select the
site, but will now do so following
input from SPES and STAT.
Vendors: A Vendor Chair will be
appointed, and vendors will be
allowed to exhibit at future FTCs.
The Technical Program Committee
will continue to have complete
authority to fill the entire technical
program, with three tracks: 1)
Statistics, 2) Quality Control, and 3)
Tutorials and Case Studies.

A committee of past-chairs of the

three sponsors will select the Youden
Speaker. The STAT past-chair will
chair this committee, and the STAT
Division will be responsible for
administrative detail for this speaker.
The Conference Registration fee
will be kept as low as possible. This
fee will be waived for one speaker
for each technical talk, and for the
two luncheon speakers, the Youden
Speaker, and the Hunter Award
recipient.
Short Courses; There will be as
many as four short courses, and
these will be scheduled either on
Tuesday/Wednesday preceding the
conference or on Saturday following
the conference. STAT will own two
courses, and CPID and SPES one
each. STAT will be fully responsible
for planning and executing the entire
short course program. Short course
fees will be kept to a minimum,
and be consistent across sponsors.

MINI-PAPER
Continued from page 20
14. Kane, V. E., Process Capability Indices, JQT Vol. 18 No. 1, 1986.
15. Kotz, Samuel and Johnson, Norman, Process Capability Indices, (Chapman
& Hall) 1993.
16. Montgomery, Douglas C., Introduction to Statistical Quality Control 3rd
Edition, (Wiley) 1997.
17. Pignatiello, J. J., and Ramberg J. S., Process Capability Indices: Just Say
No!, ASQ AQC Transactions, 1993.
18. Rodriguez, R. N., Recent Developments in Process Capability Analysis, JQT
Vol. 24 No. 4, October 1992.
19. Voelkel, J. O., Process Indices for Certain Non-Stable Processes, ASQ 43rd
Fall Technical Conference, Corning, NY, 1998.
20. Wheeler, D. J. and Chambers, D. S., Understanding Statistical Process
Control 2nd edition, (SPC Press) 1992.

WEB
SITES
The Statistics Division is responsible for developing and maintaining
three web sites. The URL for each of
these sites, and a brief discussion of
the purpose/content of each site
follows.
1. Statistics Division web page
(www.asq.org/statdiv) is the official
division electronic home. The viewer
will find a list of officers, committee
chairs and other volunteers; information on upcoming conferences (such
as AQC, FTC and Applied Statistics); a
list of division products and how to
obtain them; the Virtual Academy, a
page of links to statistics on-line
tutorials; minutes of Stat Division
meetings, etc.
2. The Statistical Clearinghouse.
This page will primarily provide links
to various Statistical resources on the
web. For example, there will be links
to major software vendors, major publishers of statistics materials, major statistics journals, and other statistics
societies. There will also be reviews of
statistics texts, software, etc. We
expect this to be one of the first
places someone would look in trying
to answer the question I wonder
if statistics. Temporarily, the
URL for this page is internet.roadrunner. com/~webstar/
3. Previous copies of the Statistics
Division Newsletter, beginning with
the 1980 issues, are being posted at
www.cba.bgsu.edu/asor/asqnews/
letter.html. We hope to eventually
have all previous newsletters posted,
with separate pages for past Youden
Addresses, past minipapers and
tutorials, past lists of officers and
committee people, etc.

Bob Mitchell is a Sr. Quality Specialist with 3M Personal Care & Related
Products Division. Bob is a senior member of the American Society for Quality, a
Certified Quality Engineer, a Certified Quality Manager, and Chair-elect of the
ASQ Statistics Division.

ASQ STATISTICS DIVISION NEWSLETTER, VOL. 18, NO. 1

21