You are on page 1of 12

COMMUN. STATIST. -THEORY METH.

, 1 8 ( 1 2 ) , 4549-4560 ( 1 9 8 9 )

O N THE DISTRIBUTIONS OF THE ESTIMATED


PROCESS CAPABILITY INDICES

Youn-Min Chou
Division of Mathematics, Computer Science, and Statistics
The University of Texas at San Antonio
San Antonio, Texas 78285
Downloaded by [Michigan State University] at 02:59 29 January 2015

D. B. Owen
Department of Statistical Science
Southern Methodist University
Dallas, Texas 75275

ABSTRACT

The exact distributions of the estimated process capability indices are pre-
sented and their means, variances, and mean-squared errors are given. The basic
assumption is that the process measurements are taken from a normal distribution.
The results in this article are useful in evaluating process capability.

Key Words: Process Capability; Noncentral t-distribution; bivariate noncentral


t-distribution.

Copyright O 1990 by Marcel Dekker, Inc.


CHOU AND OWEN

1. INTRODUCTION

Suppose that quality requirements are given in the form of specification limits
U and L on a quality characteristic X of individual items of products. A random
sample of n measurements X1,Xa, ...,X, are made to evaluate the process in re!a-
tionship to the specification limits. Assume that X has a normal distribution with
mean p and standard deviation u and the process is properly controlled. Process
capability indices are measures for the capability of the process in meeting estab-
lished specification limits. The commonly used measures of process capability in
terms of specification limits are Cp,CPU,CPL, and Cpk. They have been used by
many companies. For example, the Cp and Cpk indices have been used in Japan and
in the US. automotive industry such as Ford Motor Company (see Kane (1986)).
Downloaded by [Michigan State University] at 02:59 29 January 2015

The capability ratio l/Cp has been used by General Motors (1986). The 'Six Sigma
Qualityn is Motorola'a 1992 target for the quality of its products. Motorola, Inc.
defines the six-sigma quality by Cp = 2.0,Cpk = 1.5, and no more than 3.4 defects
per million parts (see Harry (1986)). In a recent article, Rado (1989) discusses how
Imprimis Technology, Inc. uses the process capability indices for program planning
and growth.

The process capability indices are defined as follows:

and

These indices involve unknown parameters p and u. Typically, their estimates


are used in process evaluation. The estimates are obtained by using for p and S
for u in the above indices, where and S are the sample mean and sample standard
deviation of the sample of n measurements. Therefore, the sampling distributions
THE ESTIMATED PROCESS CAPABILITY INDICES 4551

are of importance in evaluating process capability. The purpose of this paper is to


derive the exact sampling distributions of the estimated indices. In the following
discussion, we adopt the standard notations for estimators and density functions.
That is, a "hatn (A) will denote an estimated index and g,(.) stands for the pdf of
a variate Y.

2. THE DISTRIBUTION OF eP
Since ep= (U-L)/6S and Cp = (U- L)/6a,it is clear that ( n - l ) ( ~ ~ has
/ e ~ ) ~
a chi-square distribution with (n- 1) degrees of freedom. That is, the distribution of
ePis inversely proportional to a chi-distribution. This result has been used by Kane
(1986) and Chan et al. (1988) in calculating type I and type I1 error probabilities
Downloaded by [Michigan State University] at 02:59 29 January 2015

when assessing process capability.

It can be shown that the pdf of ePis

where f = n - 1 .

The mean of ePis given by Chan et al. (1988):


E ( & ) = C I I . C + , where C I I = f i . l " [ ~ -( lf) ] / r ( i ) . (0.1)

C l l is tabulated by Hogben et al. (1961) for 45 values of f from 2 to 1000.

It can be shown by direct integration that

Thus, the variance of epis given by

The coefficient (A- c : ~is)denoted by C12and is also tabulated in Hogben et


al. (1961). It follows that the mean-squared error of ePis
CHOU AND OWEN

3. THE DISTRIBUTIONS OF C?U AND C ~ L

The inequality C?U 5 y, or equivalently, 5 y can be written as

or simply

where Z has a standard normal distribution and 61 = - f i ( U -p)/u = -3fi.CPU.


The variate is denoted by Tf (61) which has a noncentral t distribution with f
degrees of freedom and the noncentrality parameter is 6i. Therefore,
Downloaded by [Michigan State University] at 02:59 29 January 2015

CFU
5 y if and only if Tf (6i) >_ -3fiy.

Similarly, we can show that the inequality C?L 5 y is equivalent to Tf(bz) 5


3,fiy, where 62 = f i ( p - L)/u = 3 f i . CPL. That is, both C?U and C?L have
distributions which are proportional to noncentral t distributions. Hence, the cdf of
CFUis given by

The moments of a noncentral t variate Tf(6) have been given by Hogben et


al. (1961), and are polynomial functions of 6 whose coefficients depend on f. The
mean and variance of Tf (6) are as follows:

where is defined in equation (2.1), Czz = & - cfl,and


THE ESTIMATED PROCESS CAPABILITY INDICES

Since C ~ is
U distributed as Tf ( & ) / ( - 3 A , we have

and

) Pt{Tf (62) 5
Similarly, it can be shown that the cdf of C ~ isLgiven by F C ~ L ( y=
3 6 y ) and the pdf of C ~ isLgcTL(y) = 3 f i . gT,(ao)(3fiy) for -00 < y < +m.
The mean and variance of C?L are as follows:
Downloaded by [Michigan State University] at 02:59 29 January 2015

and

~ ( c F L ) = C 2 z . ( ~ ~ ~czo z.
) 2 +

They have the same form as those of C ~ U Therefore,


.

4. THE:DISTRIBUTION OFr,'pe

Since ck= rnin(CT~,@L), the inequality

&pk 1 Y

is equivdent to

Tf(6i) I- 3 6 1 1 and Tf (62) Z 3 6 y .

Thus the cdf of ePkis given by


F e p , ( ~ )= 1 - Pr(Tf(61) 5 -3d%Tf(61) > I&).
4554 CHOU AND OWEN

(Tf (6i), Tf (62)) has a bivariate noncentral t distribution with correlation coefficient
one. A discussion of this distribution, and its relations t o two-sided sampling plans
and tolerance intervals, is given by Owen (1965). Since U > L, we have 62 > 6i. It
fidlows that the cdf can be written as

where t = 3fiy, R = vfl,61 and 62are defined as before, and the Qffunction
ie discussed by Owen (1965).

It is shown in the Appendix that the pdf of ePkis given by


Downloaded by [Michigan State University] at 02:59 29 January 2015

where tl = -t2 = -3JFiy.

Attempts to determine the mean and variance of kPkdirectly using its pdf may
lead t o complexities. They can, however, be found explicitly by using the following
formula:

The derivation for ~ ( 6and~~ ~( 6is) given


~ in~ the) Appendix. E(ePk)and v(dpr)
can be written in terms of Cp and Cp - Cpk. Let d = Cp - Cpkand @(a) be the
standard normal cdf. Then

and

where Eo is defined in Theorem 3 and Equation (5.1).

The mean-squared error of ePkcan be obtained by the formula


THE ESTIMATED PROCESS CAPABILITY INDICES 4555

Since the expression for M S E ( & ~ ~is) very lengthy, we will not show it explicitly.

THEOREM 1

If the n process measurements are from a N ( C c , ~ distribution,


Z) then the pdf
of ePkis given by
Downloaded by [Michigan State University] at 02:59 29 January 2015

where t l = -tl = -3fiy, f = n - 1,61 = - 3 f i . CPU,62 = 3 6 . CPL, R =


h-t,Jf,Q,(t,6;~,
k-bl / (5-
R)= ~ ( f ) Q
o
R
6) zf-'.Q1(z)dz,c(f) =
6
r(f) . 2 ( f - W '
and @ is the standard normal cdf.

Proof.

The cdf of kPkcan be written as

Differentiating both sides with respect to y gives

Let t = 3 f i y, then tl = -t,t2 = t, and R = VJf.Note that R is a


function o f t , not a constant. It suffices to show that

for t > 0.
CHOU A N D OWEN

It can be shown that

Using integration by parts, the above integral can be written aa

We notice that the integrals on the right-hand side of the equation are of the
Downloaded by [Michigan State University] at 02:59 29 January 2015

same form as the Q function. Therefore,

Under the same assumptions and uaing the same notations as in Theorem 1,
we have the following results:

THEOREM 2
THE ESTIMATED PROCESS CAPABILITY INDICES

Proof.

Since ePk= ep- where T = y,and E(&,) .


= G1 Cp, we have
~ ( 6 =~ C114
~ ) - ? .E (q). When the process measurements are from a
T -XI and f are independent and E (f) = Cll/o. Thus,
normal distribution, I

It can be shown that

(-)~ I f +i ( P - T I . (-)u l f i
Downloaded by [Michigan State University] at 02:59 29 January 2015

0 Y-p Y-CI
/(Y - T)!JZ(Y)dY = --@ @
fi
We have

It is clear that the right-hand side of the above equation remains the same if T - p
is replaced by (T- PI. Let d = Cp - Cpk = v, then

This shows that

is given in equation (5.1), G1 in equation (2.1), and Czo


and Ca2 in equation (3.1).
CHOU AND OWEN

Proof.

Since e P k = ep- 9we


, have

It is shown in Section 2 that v(&) = C2zCj. Since IT - X


I and l/SZare indepen-
dent, E(1/s2) = C20/u2, we have
Downloaded by [Michigan State University] at 02:59 29 January 2015

Now,

Therefore,

If we use equation (5.1) for E ,then the right-hand side of the above equa-
tion becomes very lengthy and can not be simplified. So we will not use equation
(5.1) in v(ePk). Instead, we let 4=E
Then

Q.E.D.
THE ESTIMATED PROCESS CAPABILITY INDICES

COROLLARY.

I f p = T , then

and

are in equation (3.1), and f = n- 1.


where Cll is given in equation (2.1), Czo and Cz2
Downloaded by [Michigan State University] at 02:59 29 January 2015

By Theorem 3,

Since Cz2 = CZo- C!, and Cp = Cpk, we have

Q.E.D.

BIBLIOGRAPHY

Chan, L. K., Cheng, S. W., Spiring, F. A. (1988). A new measure of process


capability: Cp,. Journal of Quality Technology, 22,162-175.

General Motors (1986). Statieticd proceee control manual. ASQC Automotive


Division.
4560 CHOU AND OWEN

Harry, M. J. (1986). The nature of six sigma quality. Government electronics


group, Motorola, Inc.

Hogben, D., Pinkham, R. S., and Wilk, M. B. (1961). The moments of the non-
central t-distribution. Biometrika, 48,465-468.

Krme, V. E. (1986). Process capability indices. Journal of Quality Technology, u,


41-52.

Owen, D. B. (1965). Aspecialcase of a bivariate noncentral t-distribution. Biometrika,


m,437-446.
Rado, L. E. (1989). Enhance product development by using capability indexes.
Quality Progrcea, 22,38-41.
Downloaded by [Michigan State University] at 02:59 29 January 2015

You might also like