Statistical Methods & Applications (2001) 10:273-285

SMA

(~) Springer-Verlag2001

Statistical analysis of process capability indices with

measurement errors: the case of Cp
Silvano Bordignon 1, Michele Scagliarini 2 *
1 Dipartimento di Scienze Statistiche Via C. Battisti 241/243, 35121 Padova, Italy
(e-mail: bordignon@stat.unipd.it)
Dipartimento di Scienze Statistiche Via Belle Arti 41, 40126 Bologna, Italy
(e-mail: scagliar@stat.unibo.it)

Abstract. Process capability indices (PCIs) have been widely used in manufacturing industries to provide a quantitative measure of process potential and performance. While some efforts have been dedicated in the literature to the statistical
properties of PCIs estimators, scarce attention has been given to the evaluation
of these properties when sample data are affected by measurement errors. In this
work we deal with the problem o f measurement errors effects on the performance
of PCIs. The analysis is illustrated with reference to Cp, i.e. the simplest and most
c o m m o n measure suggested to evaluate process capability.
K e y w o r d s : Process capability indices, measurement errors, Cp, estimation

1. Introduction
Process capability indices (PCIs) are widely used in manufacturing industries to
measure the ability of a process to realize items that meet the specification limits.
The analytical formulation of these indices is generally easy to understand and
straightforward to apply. Further, they provide management with a single-number
summary of what is happening on the production floor. These are the main reasons
o f their extensive use in quality assurance and quality improvement activities of
many companies.

* The authors would like to thank two anonymous referees for their comments and suggestion that
were useful in the preparation and improvement of this paper. This work was partially supported by a
MURST research grant.

274

s. Bordignon, M. Scagliarini

However, since PCIs are calculated on sampling observations, a certain amount

of uncertainty, due to the sampling error, is necessarily present in the evaluation of
the process performances. In the operative context this fact is rather neglected and
conclusions about the capability of the process are often based only on the single
numerical value of the index provided by sampling data.
Clearly, this approach is not reliable, since sampling errors are ignored. This
uncritical use of PCIs and the scarce attention given to accuracy problems, have
caused an increasing interest toward the statistical properties of PCIs estimators in
the literature of the last years.
Evidence of this attention is confirmed by the recent book of Kotz and Lovelace
(1998) and other numerous articles as, for example, Bissel (1990), Chan et al (1990),
Chou and Owen (1989), Gunter (1989).
A further source of uncertainty in the evaluation of the PCIs is given by the
frequent presence of measurement errors in the sampling observations. The causes
of measurement errors are numerous as, for example, insufficient gauge calibration
and external influences on the measuring devices.
The analysis of the effects of measurement errors on PCIs has been received
scarce attention in the literature. An exception is represented by the paper of Mittag. This author quantified the percentage error on PCIs evaluation in presence of
measurement errors. In this work our goal is to extend the analysis of Mittag to the
inferential properties of the estimator of Cp. The paper is organized as follows. In
Sect. 2 the main results on the effects of measurement errors on some theoretical
capability indices are briefly reported. Section 3 extends the analysis considering
the statistical properties of the Cp estimator in the measurement error case. Some
concluding remarks and directions for future work are reported in Sect. 4.

2. Process capability indices and measurement errors: some general remarks

Let X N N ( # , a 2) denote the relevant quality characteristic of a manufacturing
process. Given upper and lower specification limits, L S L and U S L respectively,
two capability indices, frequently used to describe the performances of a process
relative to the specification limits, are
Cp -

USL - LSL
d
6a
- 3a

(1)

and
Cpk =

m i n ( U S L - #, p - L S L )

d - IV - ml

3a

3a

'

(2)

where d = ( U S L - L S L ) / 2 is the half-length of the specification interval and

m = ( U S L + L S L ) / 2 is the midpoint of the specification interval.
Cp compares the 6a spread of the process to the tolerance spread. It does not
require knowledge of the process location # and for this reason it can be seen as
a measure for the process capability of an optimally centered process. Cpk was
introduced to give/z some influence on the value of the index. Both Cp and Cpe

Capability indices and measurementerrors

275

are designed to reflect changes in the amount of product beyond the specification
limits.
Neither index considers the target value of the process which is a critical quantity
to assess the process performances. For this reason Chan, Cheng and Spiring (1988)
proposed the index

Cpm

USL - LSL

6V/a2 + ( / z - M ) 2

3V/o.2 + ( # _ M ) 2

(3)

where M is the target value of the process, while the term V/a 2 + (# - M ) 2 is the
standard deviation about the target. It is most often assumed that M coincides with
the midpoint of the specification range.
Let us consider now the case of measurement errors. Following the notation of
Mittag (1997), we distinguish between two distinct types of measurement error:
bias, where the measured variable X ~ is consistently offset by a constant bias c
according to
X r = X + c,

(4)

where X is the true unobservable variable;

imprecision, or stochastic measurement errors. In particular we assume that this
error is described by a random variable V ~,, N(0, a 2 ) . It is further assumed
that X and V are additively linked according to
x e= x + v

(5)

and that X and V are stochastically independent. Thus, according to (5), one
observes instead of the N (#, a 2)-distributed variable X the N (#, cr~)-distributed empirical variable X ~, with a~ = a 2 + a 2. Therefore the observable PCIs
Cp, Cpk and C~m are obtained by formulae (1), (2) and (3) after substituting
ae for o-.
The effects of constant and random measurement errors on the performances of
theoretical capability indices are examined in Mittag (1997). In the case of random
measurement errors this author showed that the ratios Cp/Cp and Cpk/Cpk are
functions of the contamination degree T = a y / a , according to the following
relationship
c ~ _ c~k _
1
cp
cpk
v~+T2

(6)

It is clear from (6) that the ratios between observable and tree PCIs are decreasing
functions o f t (Fig. l) and consequently C~ and C~k, calculated using the observable
variable X e, systematically understate the true capability of the process.
The results provided by Mittag are quite interesting, since they emphasize that
the accuracy of a capability analysis could be significantly influenced by the accuracy of the gauges and, consequently, measurement errors should receive greater
attention. However, the analysis of Mittag is confined to considering the effects of

276

S. Bordignon,M. Scagliarini

1
0.9B
0.96
0.94
0.92
0.9
0.B8
0.86

II

II

I II

II

I 11

II

[ II

I I[

I[

I II

II

I ~111

I II

~1

III

,~- 0o c~ ~o t~ ~- Qo t~ co ~
oo

Fig. 1 C~
Cp
9

~1

I II

I II

II

I II

II

[111~1

,~ 0o t~ ~o to
o

c-~k
Cpk as functionsof T

measurement errors only on the behavior of theoretical capability indices, while

such effects, when PCIs are estimated from sample data, are not taken into account.
In the next section we deal with the stochastic measurement error problem with
reference to Cp.

3. Statistical analysis of Cp
Denoting with {Xj, j = 1, 2 , . . . , n } the random sample of size n from the quality
characteristics X , the most commonly employed estimator of Cp is
~p _ U S L - L S L

68
where ~2 -

'

(7)

n -1 1 ~ j =n l ( X j _ ~ ) 2 and X = g1 ~ j =n l X j . Kotz and Johnson (1993)

derived the following expressions for the expected value and the variance of Cp

E(O ) :

(8)

(9)
where

bf =

r ( 89 - 1 ) ) '

f = n - 1 and F ( . ) is the G a m m a function. Expression (8) shows that Cp is a

biased estimator of C v and, since bf < 1, the resulting bias is positive. However,
this bias goes to zero as n -+ <x~.As Var(Cp) goes to zero as n --+ o~, Cp is also
mean square consistent.

Capability indices and measurement errors

277

0.3
--0

0.2
0.1

0
-0.1

~.

~~ ~

~-

~ ~~ .~__::

-_ ~. ~ o.~
---+I

- . . . ~

-0.2

~*~''-"~

~
....

4"-'-4A-"

"~

--'~

....

-0 ....

41' . . . .

--.e---

-0.3
5

10

15

1
Fig. 2. Ky

20

25

30

35

40

45

50

D.6

~ o . 7

55

~
D,@

60

1 as a functionof n for different values of ~-

3.1. Bias o f the estimator in the measurement error case

In the measurement error case we can observe X e = X + V instead of the true

variable X. Denoting with {X~, j = 1, 2 , . . . n}, the random sample of size n
from X ~, Cp is estimated by the quantity
~

~_ U S L - L S L

68e
with ~2
1 ELI
O'e -- n--1

'

(10)

( / ; _ ~ e ) 2 and ~ e : ~1 E 2 =n I X~.
e

By the same arguments of Kotz and Johnson (1993) we obtain

1

and, consequently, the bias of C~ with respect to the true process capability index
Cp is given by

.(o;) : u(e;)- c , : co
-c,:
P bj
: Cp

1
v~1 + T 2

,)

(12)

From the analysis of B(C~) we can find some interesting results. The bias arises
from two sources, the sampling error, which leads to overestimate Cp, and the
measurement error, which leads to underestimate Cp. From the combination of
these two effects we have that B (C~) can assume either positive or negative values.
1
1
1) as a function of n is shown
In particular, the behavior of the term ( t~
- 1), as
in Figure 2 for different values of r, while the 3-D graph of ( 1 ~ 1
a function of n and r, is reported in Figure 3. It can be seen that when r and n
increase the bias becomes negative.

278

S. Bordignon, M. Scagliarini

!
Fig. 3. ~:
a

1 as a functionof n (x -- axis) and T (y -- axis)

Let r > 0 be fixed. Then, the asymptotic behavior of B(C~) is given by

B(C~) -+ C p

(1 +
~1

72

'

which implies that Cp is asymptotically biased, while it is worth reminding that Cp

is an asymptotically unbiased estimator.
The asymptotic bias of C~ is always negative and grows in absolute value with
7. However, for small contamination degrees it remains small and very near to
zero. For example, when 7 = 0.1 the asymptotic bias is only -0.0049Cp, while
for 7 = 0.8 it assumes the more significant value -0.2191Cp.
It is interesting to note that there are some values of 7- for which the two effects
(i.e. overestimation and underestimation of C v) exactly compensate each other and
therefore/3(C~) = 0. More precisely, this happens when
7=T',

(13)

where

T !

--

1~

(14)

while

B(r

>

Ofor ~

< 7',

B ( ~ ; ) < Ofor 7 > 7'.

Some numerical values of 7 pcorresponding to different sample sizes n are reported
in the second column of Table 1. These values represent thresholds for 7: under

Capability indices and measurement errors

279

Table 1. Values of r', r" and rot. Relevant relationships: a) T _< r' : 0 < B(C~) _< B(Cp); b)

T' ( r < T": --B(C~) < B(0p); T < rol : MSE(O~) (_ MSE(Cp)
n

r /

T//

701

10

0.44426
0.29201
0.23318
0.19977
0.17754
0.16139
0.12398
0.08713

0.67786
0.42654
0.33660
0.28680
0.25408
0.23049
0.17635
0.12357

1.01406
0.58219
0.44971
0.37954
0.33444
0.30234
0.22979
0.16026

20
30

40
50

60
100
200

these thresholds the overestimation effect due to the sampling error prevails, while
the negative effect of bias due to measurement errors becomes more prominent for
values o f 7- greater than the thresholds.
For example, if n = 50, then r' = 0.17754. For this value of r , corresponding
to av = (17.754%)a, B ( C ~ ) = 0. Only values of r in the interval 0 < 7- <
0.17754 can provide the positive bias 0 < B ( C ~ ) < B(Cv) = 0.01523Cp. For
7- > 0.17754 the underestimating effect due to measurement errors definitively
prevails and becomes more prominent as 7- increases.
In order m complete the picture on the behavior of B ( C ~ ) , a formal comparison
between B(C~) and the bias in the measurement error free case,
through the following inequalities

IB(@~) I < B(Op)for

IB(C;)I > B(0p) for

B(Cp), is provided

7- < 7-",

(15)

7- > 7-",

(16)

where

2b f - - 1

(17)

- 1.

Some numerical values of T" are shown in the third column of Table 1, while the
graph o f 7-" as a function of n is shown in Fig. 4.

1,5
1

0,5
0
5

10

15

20

25

30

35

40

Fig. 4. r " as a function of n

45

50

55

60

280

S. Bordignon, M. Scagliarini

As an example we report in Table 2 some numerical values of B ( C p ) for

n = 30, 50,100,200, ~- = 0, 0.1, 0.2, 0.3, 0.4 and different values of Cp. Table 2
should be examined jointly with the third column of Table 1. Picking out from
Table 1 the values o f T " for n = 30 and n = 50, 0.33660 and 0.25408 respectively,
one can observe that for values of T not greater than these values IB (C~) ] < B (Cp)
according to (15) and (16).

3.2. Mean square error of the estimator in the measurement error case
Examining now the variance of C~, by the same arguments of Kotz and Johnson
(1993), we can obtain the following expression

f
b72) (c;) 2 =
Var(gp) = f--2
-- ( f f
bf2) (USL- LSL)2 =
2
36(cr2 + a~/)
_( f
b}-2) 2 1
f-- 2

(18)

Cp 1 +T 2"

Since "r > 0, it follows that

Va,,(ep) >_Var(~;).
Let us consider now the mean square errors
(10) respectively

MSE(Cp)

(MSEs) of the two estimators (7) and

B(Cp) 2 -[- Var(Cp),

(19)

MSE(Cp) = B(C~)
^~ 2 +Va~(C~).
-e

(20)

Comparing (19) and (20), we obtain

MSE(Cp) > MSE(Cp) for T < ~-ol,

MSE(Cp) = MSE(Cp) for T = To,,
MSE(Cp)

<

MSE(Cp) for

(21)

T > TO,,

where

To1=

i{' }2
f

f--2

f--2

-- 1.

_~

bf

In the fourth column of Table 1 are reported some values of 1-0.11 and Fig.5 shows
the graph of T01 as a function of n, while the ratio MSE(Cp)/MSE(C~) as a
function of both T and n is shown in Fig. 6.

Capability indices and measurement errors

281

Table 2. Numerical values of B(C~) a n d S )

cp

B(Cp)

B(C~)

for different values of n, T, and Cp

B(C$)

T -~ 0.1

0.666667
1
1.333333
1.666667
2

0.666667
1

1.333333
1.666667
2

0.666667
1
1.333333
1.666667
2

0.666667
1

1.333333
1.666667
2

0.002523
(0.033887)
0.003784
(0.050830)
0.005046
(0.067774)
0.006307
(0.084717)
0.007569

T = 0.2
n =30
0.014490
0.004594
(0.094209)
(0.091850)
0.021735
0.006891
(0.141313)
(0.137775)
0.028980
0.009188
(0.188418)
(0.183700)
0.036225
0.011485
(0.235522)
(0.229626)
0.043470
0.013781
(0.282626)
(0.275551)
n = 50
0.007067
--0.002721
(0.070016)
(0.068700)
0.010600
--0.004082
(0.105024)
(0.103050)
0.014134
-0.005443
(0.140032)
(0.137400)
0.017667
--0.006804
(0.175040)
(0.171751)
0.021201
-0.008164
(0.210048)
(0.206101)
n = 100
0.001772
--0.007940
(0.048068)
(0.0479985)
0.002657
-0.011910
(0.072102)
(0.071998)
0.003543
-0.015880
(0.096135)
(0.095997)
0.004429
--0.019850
(0.120169)
(0.119996)
0.005315
--0.023820
(0.144203)
(0.143995)
n = 200
--0.000798
--0.010472
(0.033634)
(0.034752)
-0.001197
--0.015709
(0.050452)
(0.052128)
--0.001596
-0.020945
(0.067269)
(0.069504)
-0.001996
-0.026181
(0.084087)
(0.086881)
--0.002395
--0.031417

(0.101661)

(0.100904)

0.017887
0.0952468)
0.026831
(0.142870)
0.035775
(0.190494)
0.044718
(0.238117)
0.053662
(0.285740)
0.010427
(0.070778)
0.015641
(0.106167)
0.020854
(0.141556)
0.026068
(0.176945)
0.031282
(0.212334)
0.005105
(0.048544)
0.007658
(0.072816)
0.010211
(0.097088)
0.012764
(0.121360)
0.015316
(0.145632)

(0.104257)

B(C;)

B(C;)

T = 0.3

~- = 0.4

-0.010983
(0.090277)
-0.016474
(0.135416)
--0.021966
(0.180554)

-0.031074
(0.092252)
-0.046611
(0.138378)
--0.062148
(0.184504)

-0.027457

-0.077685

(0.225693)
-0.032948
(0.270836)

(0.230630)
-0.093223
(0.276756)

--0.018128
(0.069461)
--0.027192
(0.104191)
-0.036257
(0.138921)
--0.045321
(0.173652)
-0.54385
(0.208382)

--0.038001
(0.075292)
--0.057001
(0.112938)
-0.076002
(0.150584)
--0.095002
(0.188230)
-0.114002
(0.225876)

-0.023226
(0.051744)
-0.034838
(0.077616)
-0.046451
(0.103488)
-0.058064
(0.129360)
-0.069677
(0.155232)

-0.042942
(0.062072)
-0.064413
(0.093109)
-0.085884
(0.124145)
-0.107355
(0.155182)
-0.128826
(0.186218)

-0.025699
(0.041330)
--0.038549
(0.061994)
0.051399
(0.082659)
-0.064248
(0.103324)
--0.077098
(0.123989)

-0.045340
(0.055138)
-0.068010
(0.082706)
-0.090680
(0.110275)
-0.113350
(0.137844)
-0.136019
(0.165413)

282

S. Bordignon, M. Scagliarini

.2
2

c;
cl
20

40

60

80

1O0

120

140

160

180

200

Fig. 5. to1 as a function o f n

Fig. 6. M S E ( Ce ) / M S E ( C~ ) as a function of n (x - axis) and T (y -- axis)

] Values of r01 exist only if f > 2 (n > 3). This because the condition

___f
f-2
f
2

f--2

bf

must be satisfied.
For f = 1 and f = 2 TO1 is not defined.
For f = 3 (n = 4) condition (22) holds if
2
----<0.
bf--

>1

(22)

Capability indices and measurement errors

283

From the previous results, one can observe that there exists a considerable
range o f values o f 7-, for which MSE(C~) < MSE(Cp), specially for small
and moderate sample sizes. As n increases, TOt goes to zero. Thus, for large n
MSE(C~) tends to be definitively greater than MSE(Cp), for almost every value
o f 7- > 0 and the difference between the two MSEs increases with 7-. However,
this difference is practically indistinguishable if 7- is very small.
As ~ example we report in Table 2 some numerical values of the square roots of
MSE(Cp) for n = 30, 50, 100,200 and 7- = 0, 0.2, 0.3, 0.4. Reading from Table
1 the values o f 7-ol for n = 30, 50, 100,200, 7-oi -= 0.44071, 0.33444, 0.22979,
0.16026 respectively one can observe that for values o f 7- not greater than these
values MSE(Cp) > MSE(C~) according to (21).

3.3. Inference on 7From the previous analysis it is evident that the contamination degree r plays a
relevant role in evaluating the statistical properties o f the Cp estimator. Therefore
it is important to evaluate this quantity in order to compare it with the discriminant
values T~, 7 ~tand TOl.
The contamination degree is defined as the ratio between a and a v . The measurement error variability cr~ is often known since, it is relative to the characteristics
of the measurement device, or can be estimated by a suitable experiment. The total
variability a~2 = a 2 + a 2, which includes both product and gage variability, can
be easily estimated using a random sample. Since the contamination degree can be
rewritten as,
~rv
2 " 2
the condition
T_<r*
holds when
0"2 ( 1 - / - ( T * ) 2)
2>
O" e

--

where 7-* may be one o f the discriminant values r ' , r " or to1. Thus, a test on 7- can
be formulated as a classical test on variances
Since by > 0 then condition (22) is satisfied.
For f > 3 (n > 4) condition (22) holds if
f - 2
2 (f -- 2)
--_<by<_-f
f
Since ~
J
f>2.

<

by < ~
J

is always satisfied for f > 3, we conclude that values of vol exist when

284

S. Bordignon, M. Scagliarini

a ~ (1 + (7-.)2)

no : ae2 >

(7-*)2

2
a~(l+(r*)2)
H i : ae <

i.e.(7- < 7-*),

(23)

i.e. (7- > 7-*).

The evaluation of 7- is useful also for giving an approximate probabilistic assessment of how to locate a biased estimate of the index with respect to the true

valueCp. Theevents{ Cp-Cp >x}and{(C~-Cp)2>_z2}areequivalent.

Therefore, using the Markov's inequality, we can obtain

Pr(g;-Cp

~ x) <_

MSE(Cp)
x2

(24)

4. Concluding remarks
In this work we have been concerned with the problem of measurement errors
when dealing with process capability indices. In particular, the pioneering analysis
of Mittag (1997) has been extended by turning our attention to the inferential
side of the problem, i.e. by considering the effects of measurement errors on the
properties of capability indices estimated from sample data. Our analysis has been
concentrated by way of example on the index Cp; however it seems that similar
developments could be worked out for Cpk and other more recently introduced
indices, as suggested by very preliminary findings from our work in progress.
In the following we summarize the main results obtained in this paper. First,
the Cp estimator, that one obtained from sample data contaminated by random
measurement errors, is generally biased. The bias has two sources, the sampling
error whose effect is overestimating Cp and the measurement error which causes
the underestimating effect on Cp. The combination of these two effects involves
that the overall bias can be either positive or negative, according to which effect is
predominant.Further, if the two effects offset each other, a null bias could occur.
The bias of C~ tends towards steady negative values as n --+ c~ and increases with
the contamination degree 7-. It should be noted that this behavior is rather different
from that displayed by the usual estimator of Cp in the measurement error free case,
where the bias is always positive and goes to zero as n --+ oo. Thus, for suitable
combinations of not very large n and small 7- a relatively small absolute bias can
occur, even smaller than the bias of Cp. But, as soon as n becomes sufficiently
large, C~ underestimates the true Cp.
As far as the variability of C~ is concerned, we have shown that

Var(C~) is

never greater than Var(Cp). Therefore, when comparing the mean square errors
of the two estimators, the biascomponent plays a more prominent role. In particular, for increasing n, MSE(C~) converges to a finite quantity greater than zero,

Capability indices and measurement errors

285

while MSE(Cp) goes to zero. Further, for small and moderate sample sizes, it
could happen that MSE(Cp) < MSE(Cp), but for large n MSE(Cp) tends to
be definitively greater than MSE(Cp). However the difference between the two
MSEs is practically insignificant if 7- remains very small.
It seems to us that the results obtained in this paper have some interest for the
following reasons: (a) they focus one's attention on an often neglected problem in
industrial practice, that one represented by the presence of errors when measuring a
quality characteristics; (b) they give a first picture about the effects of measurement
errors when dealing with process capability indices; (c) they provide the practitioner
with useful tools to evaluate the reliability of his capability index, taking into
account both the sampling error and the measurement error. In this connection such
results state that if n is not too large and 7 is sufficiently small, measurement errors
do not represent a serious problem in the evaluation of a process capability index,
while in the remaining situations it is reasonable to take in account its effects, since
they lead to a systematical undervaluation of the index. Clearly, the previous analysis
represents only a first step towards supplying a more useful and complete picture.
In this connection further extensions are needed and we are currently undertaking
this task turning our efforts towards two directions, concerning the investigation of
other process capability indices, as Cpk and Cpm, and different specifications of
the measurement error.

References
1. Bissel AF (1990) How reliable is your capability index?. Applied Statistics 39, 331-340
2. Chan LK, Cheng SW, Spiring FA 0988) A new measure of process capability, Cpra.Journal of
Quality Technology 20, 160-175
3. Chart LK, Xiong Z, Zhang D (1990) On the asymptotic distributions of some process capability
indices. Communication in Statistics, Theory and Methods 19, 11-18
4. Chou, YM, Owen DB (1989) On the distributions of the estimated process capability indices.
Communication in Statistics, Theory and Methods 15, 4549-4560
5. Chou YM, Owen DB, Borrego SA (1990) Lower confidence limits on process capability indices.
Journal of Quality Technology 22, 223-229
6. Gunter BH (1989) The use and abuses of Cpk: Parts 1-4. Quality Progress 22(1), 72-73; 22 (3),
108-109; 22 (5), 79-80; 22 (7), 86-87
7. Kane VE (1986) Process capability indices. Journal of Quality Technology 18, 41-52
8. Kotz S, Johnson NL (1993) Process capability indices. Chapman and Hall, London
9. Kotz S, Lovelace CR (1998) Process capability indices in theory and practice. Arnold, London
10. Mittag HJ (1997) Measurement error effects on the performance of process capability indices. In:
Lenz HJ, Wilrich PTh (eds.) Frontiers in statistical quality control 5, pp. 195-206. Physica-Verlag,
Heidelberg
11. Montgomery DC (1996) Introduction to statistical quality control, 3rd edn. John Wiley and Sons,
New York