Process Capability Indices

<

and Norman L. Johnson Unwersity of North Carolina, ChapelHill, North Carolina, USA

CHAPMAN' cStHALL London. Glasgow. New York. Tokyo. Melbourne. Madras

rm

("\ '-)

,'-, '..,,"" 'VI., 'I 'i '/

l,

q..

I'ublishl'd by (,h:II)II1:1II.~ lIall,. 2 (, Houlldary I{cn~, LOlldoll SI<:I !II IN

Chapman & Hall. 2-6 Boundary Row, London SEI 8HN, UK Blackie A\:ademk& ProfessioliaL WeslerCleddens Road, Bislwpbrig[!.s, Glasgow 064 2NZ. UK Chapman & HalrInc., 29 West 35th Street, New York NYlOOOl, USA Chapman & Hall Japan, Thomson Publishing Japan, Hirukawacho Nemoto Building,6F, 1-7-11 Hirakawa-cho, Chiyoda-ku, Tokyo .102, Japan Chapman & Hall Australia, Thomas Nelson Australia, 102 Dodds Strect, South Melbourne, Victoria 3205, Australia Chapman & Hall India, R. Seshadri, 32 Second Main Road, CIT East, Madras 600 035, India

Foreword

Preface 1 Introduction Facts needed from distribution theory Approximations with statistical differentials ('della mel hod') 1.3 Binomial and Poisson di.stributions ..1,4 Normal distributions 1.5 (Central) chi-squared distributions and non-central chi-squared distributions 1.6 Beta distributions 1.7 t-distribuljons and non-central t-distributions I.R Folded distributions J.~ Mixture distributions 1.10 Multinormal (multivariate normal) distributions I. JI Systems of distribu!ion:.; LI2 Facts from statistical methodology . Bibliography' 2 'rile basic process capability indices: C", C"k and their modifications 2.1 Introduction 2.2 The Cp index 2.3 Estimation of Cp 2.4 The Cpkindex
2.5 Estimation of' Cpk

Firs}-e"dilion993 1
\..,..~993 Chapman & Hall Typeset in l1/13pt Times by Interprint Limited, Malta. Printed in Great Britllin by St Edmundsbury Press, Bury $( Edmunds, Suffolk ISBN 0 412 54380 X

1.1 1.2

vii IX 1 3 10 14 16 18 22 24 26 27 29 30 32 36 , 37 37 38 43 51 55 74 76 : 78

Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted under the UK Copyright Designs and Patents Act, 1988, this publication may not be reproduced, stored, or trunsmitled. in any form 01' by any means, without the prior permission in writing of the publishers. or in the case of reprographk reproduction only in a\:wrdance with the terms of the li\:en\:es issued by thc ('l1pyrighl Licensing Agency in the UK, or in accordan\:e. with the terms of licences issued by the appropriate Reprodu\:tion Rights Organization outside the UK, Enquiries con\:erning reproduction omside the terms stated here should be sent to the publishers at the London address printed on this page. The publisher makes no representation, express or implied, with regard to the acwracy of the information contained in this book and cannot a\:cept any legal responsibility or liability for any errors or omissions that may be made, A catalogue Library record for this book is available (' , , . '.. - :.. "..~O: from the British Library :..- ~ :'..'.\..

of Congress

Process
1111111111111111111111111111111

capability

indices

168718

e
Q,

Katz, Samuel' P rmted on permau"", a u-"..."""" 1"'1"",U""IU""""""U ... ..wv.u,...~e with the proposed ANSI/NISO Z 39.48-199X and ANSI Z 39.48-1984
'

658.562

KOT/P

N93

CENTH

L'!.

!.i :t(~ ( :n,

2.6 2.7 2.8

Comparison of capabilities Correlated observations PCls for attributes

v

Lt

1'" r.i1"~18i7ti~)

..

VI

Contents

Appendix 2.A Appendix 2.B Bibliography . 3 The Cpmindex and related indices
.

79 8'1 83 87 87 gg 95 98 112 117 121 127 133 135 135 137 . 151 1M 170 172 175 179 179 180 .183 187 194 .197
. )()()

Foreword

3.1
3.2 3.3 3.4

Introduction

The Cpmindex Comparison of Cpmand Ctm with C;pand Cpk Estimation of Cp~ (and Ctm)
Boyles' modified Cpm index

i 3.5

3.6 3.7

The Cpmkindex Cpm(a)indices Appendix 3.A: A minimum variance unbiased estimator for W = (9C,~.;)- 1 Bibliography

4 Process capability indices under non-normality: . robustness properties 4..1 Introduction 4.2 Effects of non-normality 4.3 Construction of robust PCls 4.4 Flexible PCTs 4.5 Asymptotic properties Appendix 4.A Bibliography 5 Multivariate process capability indices 5.1 Introduction 5.2 Construction of multivariatePCIs 5.3 Multivariate analogs of Cp 5.4 Multivariate analogs of Cpm 5.5 Vector-valued PCls Appendix 5.A Bihliography Note on computet' programs Postscript Index

203 205 20t)

Me.asures of process capability known as process capability indices (PCls) have been popular for well over 20 years, since the capability ratio (CR) was popularized by Juran. Since that time we have seen the introduction of Cp, Cpk, Cpm,Ppk and a myriad of other measures. The lIse of tbese measures has sparked a significant amount of controversy and has had a major economic impact (in many cases, negative) on industry. The issue does n.ot generally lie in the validity of the mathematics Qf the indices, but in their application by those who believe the values are deterministic, rather than random va~iabIes. Once the variability is understood and the bias (if any) is known, the use of these PCls can be more constructive. As with statistics in general, it is imperative that the readcr or Ihis lext has a work ing knowlcdgc of sta tistica] methods and distributional theory. This basic understanding is what has been lacking in the application of PCls over ~~~. I It is hoped that this text will as~jst in moderating (as in this writer's case) some of the polarization that has occurred during the past decade over the use of PCls and that we can work on process improvements in general, rather thart focusing upon a single measure or index.' . There are those of us who feel that the use of PCls has been ineffective and should be discontinued. There are those who feel they have their use when used in conjunction with other measures and there are those who use PCls as absolute
VB

,
Vl11

Forword

measures. I would suggest that an umkrstanding l)!' the concepts in this text could bring the majority of viewpoints to a common ground. Robert A. Dovich Decemher. 1992

Preface

The use and abuse of process capability indices (PCls) have hecome topics of considerable controversy in the last few years. Although seemingly simple in formal use, experience has shown thatPCls lend themselves rather easily to i11~ based interpretations. It is our hopl that this monograph will provide background sufficient to allow for informal assessment of what PCls can and cannot be expected to do. We have tried to keep the exposition as elementary a:s possible, without obscuring its essential logical and mathematical components. Occasionally the more elaborate of the latter are moved to an appendix. We hope that Chapter '1 wiJl provide a reminder of the concepts and techniques needed to study the book with profit. We also hope, on the other hand, that more sophisticated researchers will find possible topics for further developmental work in the book. Beforc venluring into this (\cld of currcnt controversy we had somc misgivingsabout' our abilities in this Jldd..We have benefited substantially from the advice of practitioners (especially Mr. Robert A. Dovich, Quality Manager, Ingersoll Cutting Tool Company, Rockford, Illinois and Dr. K. Hung, Department of Finance, Western Washington University, Bellingham, Washington) but our approach to the subject is based primarily on our own experience of numerical applications in specific problems over many years and our prolonged study .of distributional-theoretic aspects of statistics. We hope our efforts will contribute to lessening the gap between theory and practice. ix

Pr~j'ace

We thank Dr. W.L. Pearn, of the National Chiao Tung University, Hsinchu, Taiwan, for his extensive collaboration in the early stages of this work, Mr. Robert A. Dovich for useful comments on our first version, and Ms. Nicki Dennis for her enthusiastic support of the project. Conversations and corn.:spondence with Mr. Eric lknson and Drs. Russell 1\. Boyles, Richard Burke, Smiley W. Cheng"LeRoy 1\. Franklin, A.K. Gupta, Dan Grove, Norma F. Hubele, S. Kochcrl,akota, Peter R. Nelson, Leslie 1. Porter, Barbara Price, Robert N. Rodriguez, Lynne Seymour and N.F. Zhang contributed to clarilkation of our ideas on th'e somewhat controversial subject matter of this monograph. To paraphrase George Berkeley (1685-1753) our task and aim in writing this monograph was 'to provide hints to thinking men' and women who have determination 'and curiosity to go to the bottom of things and pursue them in their own minds'. January 1q93 Samuel Kotz University of Maryland College Park, MD Norman L. Johnson University of North Carolina Chapel Hill. NC

Introduction

Process capability indices (PCls) have proliferated in' both use and varicty during the Jast decade, Their widespread and often u,ncritical use may, almost inadvertently, have led to some improvements in quality, but also, almost certainly,' have been the cause of many unjust decisions, which might have been avoided by better knowledge of their properties. These seemingly innocuous procedures for determining 'process capability' by a single index were propagated mainly by over-zealous customers who viewed them as a panacea for problems of quality improvement. Insistence on rigid adher-

cnce to rules for i;alculating the indices Cp and Cpk (see

Chapter 2) on a daily basis, with the goal of raising them above 1.333 as much as possible, caused a revolt among a number of inOuentialand open,-mindedquality control statisticians. An extreme reaction against these indices took the
form of accusations of 'statistical terrorism'

Dovich's letter

(1991 a), Dovich (1991b) and Burke et al. (1991) - and of unscrupulous manipulation and doctoring, and calls for their elimination - Kitska's letter (1991). There were also more moderate voices (Gunter, 1989) and more defenders (McCormick (in a letter, 1989 with reply by Gunter, 1989b); Steenburgh, 1991; McCoy, 1991). However, the heated and intense debate on the advisability of continuation of the practical application of these indices ,- which took place during the first four months of 1991 - indicates that something is, indeed, 1

..

J /1/

rod

11<'1 iO/1

Facts 11'eded .Ii'om distribution theory. (

wrong either with the indices themselves or the manner they had been 'sold' to Clualitycon.trol praditioners and the 'man on the shop floor;. Additional indices have been introduced in leading journals on quality control in the last few years (Chan et ai., 1988; Spiring, 1991; Boyles, 1991; Pearn et ai., 1992) in efforts to improve on Cp and Cpk by taking into account the possibility that the target (nominal) value may not correspond to the midpoint between fhe specification limits, and possible non-normality of the original process characteristic. The number of different capability indices has increased and so has confusion among practitioners, who have so far been denied an adequate and clear explanation of the meaning of the various indices, and more importantly the underlying assumptions. Their suspicion that a single number cannot fully capture the 'capability' of a process is well justified! The purpose of this monograph is to clarify the statistical. meaning of these indices by presenting the underlying theory, to investigate and compare, in some detail, the properties of various indices available in the literature, and finally to convince practitioners of the advisability of representing estimates of process capability by a pair of numbers rather by a single one, where the second (supplementary) number might provide an indicator of the sample variability (accuracy) of the primary index, which has commonly been, and sometimes still is, mistakenly accepted as a precise indicator of the process capability, without any reservations. This should defuse to some extent the tense atmosphere and the rigidity of opinions prevalent among quality controllers at present, and, perhaps, serve as an additional bridge of understanding between the consumer and the producer. It may even reduce the cost of produdion without compromising the qUillity. We consider this to be an important educillional goal. To benefit from studying this monograph, some basic knowledge of statistical methodology and in particular of the theory of a number of basic statistical distributions is essen-

tial. Although many books on quality control provide excellent descriptions of the normal distribution, the distributiona.l (SI;llislicaJ)propenies of Ihe process capability indiccs involvc . several othcr importallt but much less familiar distributions. In fact, these indices give rise to new interesting statistical distributions which may be of importance and promise in other branches of engineering and sciences in general. For this reason most of this chapter will be devoted to a condensed, and as elementary as possible, study of those statistical distributions and their properties which are essential for a deeper understanding of the distributions of the process capability indices. Further analytical treatment may be found in books and articles cited in the bibliography. Specifically, we shall cover basic properties and characteristics of random variables and estimators in general, and then survey the properties of the following: J. hinomial and Poisson distributions
2. 110 r111a distributions 1
,

.1. (central) chi-squared (and more generally gamma) distributions 4. 5. 6. 7. X. 9. non-central chi-squared distributions beta distributions t- and non-central t-distributions /(}Jdednormal distributions mixture distributions multinormal distributions

We also include, in section 1.12, a summary of techniques in statistical methodology which will be used in special applications.

1.1 FACTS NEEDED FROM DISTRIBUTION THEORY

,

We assllme that the reader possesses, at least, a basic knowledge of the elements of probability theory. We commence

..

Introduction

Faels needed from distrihution theory To represent quantities

with a condensed account of basic concepts to refresh the memory and establish not'ation, This is followed by a summary of results in distribution theory which wil1be used later in this monograph, We interpret 'probability' as representing long-run relative frequency of an event rather than a 'degree of belicf'. Probabilities are always in the interval [0, 1].. If an event al\vays (never) occurs the probability is J (0). We will be dealing with probabilities of various types of events. A random variable X represe.nts a real-valued varying quantity such that the event 'X ~ x' has a probability for each real number x. Symbolically
Pr[X ~ xJ = F x(x). (1.1)

such as length, weight, etc., which can (in principle) be thought of as taking any value over a rangc (tinite or in/inite interval) we u:-.econtinuous random variables for which the derivative

.!x(.X) = --dx

dF x(x)

= Fx(x)
I

(1.4)

density function(PDF) of the random variable X. For a Pr[a~X ~/J] = Pr[a<X ~/J] = Pr[a~X <fJJ =Pr[a<X<{3J
= F x({3) - F x(a)

of the CDF exists. The function J'x(x) is the probability

continuous random variable, X, for any real a and {3,({3 a): >

The symbol Fx(.x) stands for the cumulative distribution

function(COF) of X (or morc prccisely its valuc at X = x),

Note that it is a function (in the mathematical sense) of x, hot of X. It represents a property of the random variable X, namely its distribution. For a proper random variable
x"'"

I: fx(x) dx

(1.5)

lim Fx(x)=O
-00

X"'"

lim Fx(x)= 1
00

(1.2)

(For such a variablc, Pr[X ='aJ = Pr[X = 11J 0 for any a and =
{3,)We will use Xe to denote the (unique) solution of F x(xe)= 8.

The joint CDF of k random variables Xl, .., , X k is To represent quantities taking only integer values -- for example, number of nonconforming items - we use discrete random variables. Generally these are variables taking only a number (finite, or count ably infinite) of possible values {xd. In the cases we are dealing with the possible values are just non-negative integers and the distribution can be defined by the values Prf X =J1 J=O, L 2,... For such random variables the COF is fx(x)= fx, Fx(x) =
j<;.x

Vdx)

Vv "", ,.I'"(x J , ""

Xk)

Prr( X I

< X I) n

n(X2~x2)n... =pr

n(Xk~xk)J
'

[n

.1=1

(Xj~Xj)

(1.6)

For continuous random variables, the joint PDF is rlF x(x) Xk(XI,"', xd= ax! aX2'~..aXk (1.7)

L Pr[X

=.iJ

(1.3)

(the kth mixed derivative with respect to each variable). ,

..

IlllfOdL/CliOIl

F acts needed from distrihution theory

The variance of X is by definition

I~L( X

We denote the conditiona' probubility of an event Et given an event E2 by pr[E1IE2l By definition, El is independent of E2 if Pr[El \E2J = Pr[Ell Generally Pr[El and E2J, denoted by Pr[El n E2}, is equal to Pr[E2J Pr[El \E2J =

- E(X ))2J

-0.

E(X 2) -

LE(X )]2

(LIO a)

It is also written as
varIX) = f.l2= a2(X) = a2

Pr[E,J Pr[E2IE,l
Also pr[EI or El or bothJ, denoted by Pr[Et u El], is (1.10 b)

equal to Pr[E1J+Pr[E1J-Pr[ElnE2]. . If E 1 is independent of E2 and Pr[E;J > 0, then E2 IS

independent of E1 and we say that El. and E2 are mutually independent. Random variables X 1 and X 2 are (mutually) ind.ependent if the events (X1~X1) and (X 1~X2) are mutually independent for all Xl and Xl' so that
( 1.1)
FXl,X2(XI, X2)= F Xt(Xl )FX2(Xl)

The square root of the variance is the standard deviation, usually denoted by c(X) or simplya. This will be an important and frequently used concept in the sequel. More generally, the rth (crude) moment of X is E[Xr] = I/;.(X). Note that /1'1X) = Er Xl =~. ( .. The rth central moment is by definition: E[(X-E(X))']=llr(X) (LlI)

The expected value of a random variable represents the long-run average of a sequence of observed values to wh.ich it corresponds. The expected value of a continuous random variable, X is (1.8 a)

. The variance is the second central moment. The coefficient of variation is the ratio of standard deviation to the expected value (CV.(X) = a(X)/E[X]), but it is oftcn expressed as a percentage (lOO{a(x)/E[X]}of<)).
.

If X and Y are mutually independent, then E[X Y]=

r.=E[XJ= f~ooXfx(X)dX
also denoted as /1'1X). Often the term 'mean value' or simply ( 'mean' is used. More generally, if g(X) is a function of X (1.9 a)
E[g(X)]

E[X]E[Yl For any two random variables; whether independent or not. .

E[X + YJ=E[X]+E[Y]
The correlation coefficient between two random variables X and Y,is defined as

= J~

00

g(x)f x(x) dx

For a discrete random variable

~= E[XJ = '2>iPr[X i

=XiJ

( \.8 hI (1.9 b)

E[(X - E[X])(Y- E[YJ)] PXY = a(X)a(Y)

The numerator, EL(X- ELXJ)(Y- ELYJ)J is called the covariance of X and Y, cov (X, Y).

E[g(X)]= I,g(xJPr[X=xiJ i

..

Introduction Facts needed .Ii"omdistribution theory

Occasionally we need to use higher moments (r> 2), in order to compute conventioI].al measures of skewness (asymmetry) and kurtosis (flat- or peaked-'topness'). T,hese are
I J.l3(X) J.l3(X) (f3i (X2 = {J.l2 (X)} 1 = {a(X)p

(1.12 a)

(the symbols O:3(X)or A3(.r) are also used fa! this parameter), and J.l4(X) J.l4(X)

/1'2 cannot be less than 1. It is not possible to have {/2 ..~ /J1- J <0. (If /12- 111,-] =0 the distribution of X reduces 1 to the 'two-point' distribution with Pr[X = a]:::- - Pr[X = bJ . for.arbitrary a # h.) We state here without proof, results which we will lIse in later chapters. If X1,X2,...,Xn are mutually independent with common expected value, ~, variance, a2, and kurtosis shape factor, /32then for the arithmetic mean,
n

f32(X)= {/l2(X)]2 ='{a(X)}4

(1.12 b) X=n-i
,

i= 1

LXI

(the symbols O:4(X) or P'4(X)+ 3} arc. also used for this parameter). When there is no fear of confusion the '(X)' is often omitted, so we have for example

E[XJ=~

a2 var(X)=.:...

11

(1.14a)

For the sample variance

J /3 = a3
1

It J

/12 = It4 a4

. (1.12(')

S = --and C.V.(X)

.1" -2 (XI-X) n-11=1

= 1OOa//t't (X) (=

we have
I OOa/~o;;l)

(1.13)

.j7J; and f32 are referred to as moment ratios or shape factors (because they reflect the shape of the distribution rather than its location measured by c; or its scale c-

E[S2]=a2

(1.14 h)

measured by a). Neither nor /12 is changed if X is replaced by Y = aX + b, when a and b are constants and

j7i;

. ' n-3 var(S2)= IJ2- "n-l

a4

(1.14 c)

a> O. If a < 0, the sign of j{f; is reversed, but /12 remains unchanged. If J7E. <0 the distribution is said to be negatively skew, if j7f;> 0 the distribution is said to be posi. tively skew. For ,any normal distribution, j/II - 0 and f32= 3. If f32< 3 .the distribution is said to be platykurtic ('flat-topped'); if f32> 3 the distribution is said to b~ leptokurtic ('peak-topped').

We also not~ the simple but useful inequality E[IXI] ~ IE[XJI (1.14 d)

(Since E[IXIJ ~ E[X] and E[IXIJ = E[I- XI] ~ E[ - X] = - E[X].) Note also th~ identity min(a, b)=t(la+bl-la-bl) (1.14 e)

..

(0

lilt /'odui'I iOIl

1.2 APPROXIMATIONS WITH STATISTICAL DIFFERENTIA,LS ('DELTA METHOD')

Approximations

1vithstatistical d(lferentials

11

The deviations A - a, B - a of A and B from their expected Much of our analysis of properties of PCls will be based on the assumption that the process distribution is normal. In this case we can obtain exact forl11l1l;\s mOl11entsof sall1plc for estimators of the PCls we will discuss, In Chapter 4, however, we address problems arising from non-noimality, and in most cases we will not have exact formulas for distributions of estimators of PCls, or even for the moments. It is therefol'e useful to have some way of appmximating these values, one of which we will now describe. (Even in the case of normal process dIstributions, our approximations may be useful when they replace a cumbersome exact formula by a simpler, though only approximate formula. It is, of course, desirabk to reassure ourselves, 'so far as possible, with regard to the accuracy of the latter.) Suppose we have to deal with the distribution of the ratio
yalUes arc calJed statistical differentials. Clear]y

'"

E[A-aJ=O

E[(A-a)2J=0'~

E[(A-a){B-fJ)J=PABO'AO'/I

L'

'.
It
If '~. IC

JI ,II

(The notation A-a==i>(A); B-fJ=(5(B) is sometimes us~d, giving rise to the alternative name: delta method.) We now expand formalJy the second and third components of the right-hand side of (1.15).If r is an integer,the expansion of the second component terminates, but that of the third component does not. Thus we have (with r= 1) a

r, Ie
IC
10

G=fj 1+'-a-

A- a

){1-#+ (
,
'

B-

#2 -Ii )
B-

(1.,16)

of two random variables

'

G--B and

The crucial step in the approximation is' to suppose that we can terminate the last expansion and neglect the remaining terms. This is unlikely to lead to satisfactory results unless I(B - 13)/ fJ I is small. Indeed, if I(B -- fJ)/fJ I exceeds 1, the series does not even converge. Setting aside our misgivings for the present, we take expected values of both sides of (1.16), obtaining E[G] ~~

E[A] =a var(A)=O'~
cov(A,

E[B] ~# var(B)=O'~ fJ

(
2

1-

fJAI/(JA(J1/

(J~

afJ

+ 132

)
,
'

(1.l7a)

B) = PABO'AO'B

A similar analysis leads to

"' Vdi\

i.e. the correlation coefficient between A and B is flAil' To approximate the expected value of Gr we write G=
r

G) -'-

-\(Ja2

) (--0'~

2PAIIO'AO'I1

,afJ

+-

(J~

132

(1.17b)

()
B

a+(A-a)

{ 13+ (B -

13) }

( 13

)(

'

1+-

A-a

0:

)(

.1+-n-"'fJ
j1

)
\',\

"

Formulas (1.17a) and (1.17b) can be written conveniently in terms of the coeflicients of variation of A and B (see p. 7). E[GJ ~ fi [I - flAil': . V.(A )c. V.(B)}+ {C.V.(B)} 2J (1.17c) c

( 1.15)

...

12
var( G) ~

Introduction Approximations with statistical d!fferentials

13

(~) 2 [{c.v.(A)2
+{C.V.(B)}2]

- 2PAB{C.V.(A)C. V.(B)} (l..l7d)

We take

/)
It'

" L: (Xi--X)2 1=1 -' as our 'A', and g(A)= A 2. From (1.31) we will find that
II.

Another application of the method is to approximate the expected value of a function g(A) in terms of moments of A. Expanding in a Taylor series
g(A)

= g(a + A - a)

==

g(a)+ (A

- a)g'(a) + t(A

- a)2y"(a)

+ ... Hence

1= 1

L (XI-X)2
a

is distributed

as

X;-1(J"2

~ g(a) + (A - a)g'(a) -1- (A - aF?J"(a) ~

Taking expected values of both sides

= E [A ] = (n -

1)(J"2

E[g(A)] ~g(a) + t?/'(a)(J"~ Similar aI1alysis k:ads lo var(g(A)) ~ {g'(a)} 2 var (A) = {g'(a)} 2(J"~

(1.18a)

I)

and
. (J"~.:::2(n-l)(J"4

(1.18b)

I)

Also

g'(A)= -fA -~ g"(A)=;iA-~ Example (see Sections 1.4 and 1.5 for notation) Suppose we want to approximate the expected value and variance of
II ElA jJ~{(fl-l)(J"2} ~(n-1r-!{1+ .!+1{2(n-l)([4}{;i(n-l) 4(n~1)}(}'-1 1(J"-5} (1.19 a)

Hence

{ i=

f
-

(XI-X)2

.1

-!

and s var(A -!) ~ 2(n -1)(}'4 x (4~3 )

.

where X),...,X"

and

arc indcpcndent N(~,(J"2)random variables

X=- LXI,
n 1=1

.
=2(n-1

1
(}' x 4(n-1)
4

-6
3 (J"

..

14

I ntroductiol1

t
- 2(n--;-I)2a2

Binomial and Poisson distributions (1.I9h)

(Recall (n-I\+J).)that n!=1 x 2 x... x nand
n(kl=--=n

15

x (n-l)

x...

Remembering the formulas (1.14 b, c) for E[S2] and var(S2) we find that

weDenoting thaI Ihe quantities ohs~rve if - J--I' (a common notation in the literature),
Pr[X=O], Pr[X=l], Pr[X=2],..., Pr[X=I1] of

a(S-

n-3 1 1 )~2 /32- n-l

1 a-.

(1.19c)

for any distribution.

1.3 BINOMIAL AND POISSON DISTRIBUTIONS

are the (q+ p)". successive terms in the binomial expansion

The distribution defined by (1.20) is called a binomial distribution with parameters nand p, briefly denoted by llin~,~. . The expected valuc and variance of the distribution are
c:
11 1\ 11 II

Suppose that in a sequence of n observations we observe whether an event E (e.g. X =::;;x)occurs or not. We can represent the number of occasions on which E occurs by a random variable X. To establish an appropriate distribution for X, we need to make certain assumptions. The most common set of assumptions is: 1. for each observation the probability of occurrence of li has the same value, p, say - symbolically - Pr[E] = p: and 2. the set of results of n observations are mutually independen t. Under the above stated conditions (1.20)

j.l'l(X)=E[X]=np
112(X)= Var(X)

= npq

respectively. If n is increased indefinitely and p decreased, keeping the expected value np constant (= 0 say) then

L~ ~.

Pr[X = k] approaches to
I The distribution defined. by

Okexp(

k!

- 0)

pr[x=k]=G)pk(1--P)"-k where

for 1<=0,1,...,/1

))

Pr[X

= k] = OkeXp~~O) k=O, 1,

'"

is. a Poisson distribution with expected value 0, denoted symbolicaJly as:

( )= k!(n-k)!-kf
k

n!

n(k)

X", Poi(O)

(1.21)

For this di~tribution,E[X] = e and var (X) = e.

..

16

Introduction

N onnal distributions

17

The Il/lIllil/ol'll/lll dislrill/ll;o/l extends the binomial distrihu tion to represent that of numbers of observations N 1, ..., Nk

The CDF and PDF of X=~+aU 1

x

for a>O and ~:f=Oare

2

of mutually exclusive events E1, .,. , Ek in n independent trials with Pr[Ej] =Pj (j = 1,..., k) at eachtrial.Wehave

Fx(x) =

.
Pr[(Nl=ndn...n(Nk=nk)]=n!})l

a v 2n -

r:L

1 t.-t.
00

exp[ --

( )]
_.-:.
a

x-( dt=(}>-

0'

)
(1.24 a)

(~j)
k

p~j\

(1.22)
and
'

with
k
j

f (

: x x)= a~exp

- ~ 2 1 X- ~ [ -:2 --;- J =~(P --;--'

1

'

( )

( ) (1.24h)

j= 1

L L n = n, j= 1 Pj = 1

respectively.We write symbolicallyX '" N(~,a2).

The .expected value of X is E[XJ = ~ and the variance of X is a2. The standard deviation is 0'. The distribution of U is symmetric about :tero, i.e. fv( -u)=fv(u). The distribution of X is symmetric about its expected (mean) value ~, i.e. fx(~ +x)=fx(~ -x) (1.25)

1.4 NORMAL DISTRIBUTIONS

A random variable U has a standard (or unit) normal distribution if its CDF is 1
Fv(u)= r;;:
u
. 00

v 2n J -

exp(-~t2)dt=<b(u)

(1.23a)
I)

There are extensive tables of (}>(u)available which are widely used. Typical values are shown in T,lble 1.1.
Table 1.1 Values W(u), the standardized normal CDF

and its PDF is 1 fv(u)=-;= exp( -lu2) = (p(u) J2n .

(1.23 b)
~)

<1>( u)

<1>(u)

(We shall use the notation (}>(u), (p(u) throughout this book.) . . This distribution is also called a Gaussian distribution,and occasionally a bell-shaped distribution. The expected value and the variance of U are 0 and 1 respectively.

III.

hi D~

0 0.5 1.0 1.5 2.0 2.5 3.0

0.5000 0.6915 0.8413 0.9332 0.97725 0.9938 0.99865

0.6745 1.2816 1.6449 1.9600 2.3263 2.5758

3.0902.
3.2905

0.7500 0.9000 0.9500 0.9750 0.9900 0.9950 0.9990 0.9995

,,

.
18
Introduction

Central and non-centredchi-squared distributions

19

Note .that Ul = 1.645,and U2 = 1.96 correspond to $(ud= 0.95 and <I>(uz) 0.975 respectively, i.e. the area under the =

normal curve from

- ro up to 1.645is95%

and that from- Cf:)

up to 1.96 is 97,51<" Hence by symmetry the area betw~en .

-1.645 and 1.645 is 901Yt) nd between -1.96 and 1.96, it is a 95 (Yo,For applications in process capability indices we note the area under the normal curve between - 3 and 3 to be 99.73 (%, Thus the area outside these limits is 0.0027' (an extremely small proportion), If Xt,X2,...,X" arc indcpendent random variables witli

1 ~ ~

(Gamma function is a generalization of factorials; S1l1ce f(n+ l)=nl for any positive integer n.) This is a X2 (Chi-squared) distribution with v degrees of freedom. Symbolically we write L: Uf"'x~

II It

;= I

common N(~, (J2) distribution, thcn " 1 X=-(XI+X2+...+X")=-,,,N n

L: Xi
2

The expected value of a x; variable is v and its variance is 2v. If VI, V2,..., Vvare mutually independent normal variables each with expected value zero and standard deviation (J (variance (J2);then

i= 1

( )
~,(J

(1,26)

~)

w= Vi+ V~+.., + V~
has the PDF

1,5 (CENTRAL) CHI-SQUARED DISTRIBUTIONS AND NON-CENTRAL CHI-SQUARED DISTRIBUTIONS' 1.5.1 Central chi-squared distributions

fW(W)~[2jWW}!'-'exp(;~~)
Sym bolicalJy
W"'(J2X;

w>O

(1.29)

If U I, U 2, ..., V v are mutually independerit unit normal variables, the distribution of

II JY has expected value (J2V and variance 2(J4V, Generally

X~=UT+U~+... +U~ has the PDF

"

'

/-lrXv JI)

) -E [(

2 r

,Xv

)]

_2',rch+r) rch)

frt(X)=[2ivr(~)J-1xh-lexp( -i)
where l'(a) is the Gamma functioll ddilH.:d by
r(a)=

x>D

(1.27)

(= v(v+ 2).., (v+ 2 ;=1) if r is a positive integer). If r:E:;!v, then /-l;(X;)is infinite. Generally a distribution with PDF
lx(x) = fi:x;(a)XCC-l exp(~x) x> 0; a> 0, {3 0 > (1.30)

f '" y'x-Icxp(--y)dy
0

, '

(1.2R)

,'HI

is cal1ed a gamma (iX, ) distribution. P

.
20
Introduction Central and nor.-central chi-squared distributions
I, \I
\

21

-I

A chi-square random variable with v degrees of freedom (X~) has a gamma (tv, 2) distribution. If X1,X2,...,X" are independent N(~,(}2) random variables, then 'the sum of squared deviations from the mean' is distributed as X;-1 (}2.Symbolically (d. p. 13)
,

-x

x {2,v+Jr
,'-,

(J+"2) }

x,v+rlexp"2

)
(1.32)

=.i~O ~

(2-)fx~+,/x)
exp(~~:2)

iF

"
~ -2 L.. (Xi-X)
i= 1 '

2 "'X,,-I(}

(1.31 )

where

Also " I.tSj X)2 and

X " /I I L Xj
j= 1

p.i 2

() ()
2

c52

c52 j

~=Pr[Y=jJ .I.

i= 1

(which is distributed N(~, (}2/n)) are mutually independent.

1.5.2 Non-central chi-squared distributions

Following a similar line of development to that in Section 1.5.1,the distribution of

X~2(b2)=(U 1+ ~d2 +(U 2+ ~2)2 + .., +(U v+ ~\)2

if Y has a Poisson distribution with expected value tc52. ,This is a non.central chi-squared distribution with v degrees of freedom and non-centrality parameter c5 Note that 2. the distribution depends only on the sum c5 and not on the 2 individual values of the summands ~f. The distribution is a mixture (see section 1.9) of X~+:j distributions with weights Pj{tP) (j=O, 1,2, ...). Symbolically we write

x'v\P)", x~ I 2.! /\ POi

, .T

()
2
k

c52

(1.33)
I

where Vi'" N(O, 1) with

v

L a=b2 j= 1
has the PD F

The expected value and variance of the arc (v+b2) arid 2(v+2c52) respectivdy. Note that
k
h=1 k '" X~2

x~\P) distribution

.
ct:J c52 j 1

L X~'h(c5 ~)

fX~2(62)(X)= [(2' j~O

) Ji.J exp(':-2 )
I

c52

(L b )
~

,.

h=l

with v =

h=1
,

Vh

(1.34)

if the Xi'S in the summation are mutually independent.

22
1111 /'oi!l/cl iOIl

Beta distributions

23

1.6 BETA DISTRIBUTIONS

In particular, a
E [ YJ = J1't ( Y.\ =

A random variable Y has a standard heta (ex, n distrihution if I its PDF is

.

~+
rx(a

(1.38 a)
13

1
B(rx,/3) yOt-l(l - y)/J-l O~y~ 1 and IX, 3>0 1 (1.35 a)
E [Y 2]

'.

+ 1)

fy(y)

It

2(Y) = (a + 13)(:;+ (1+ 1)

(1.38'b)

and hence
where B(a, h) is the beta function, Thus
va r( Y )
=-=

= f( rx)r(fJ)/n

rx+ (1).
rxf3
E (Y J )

{E ( Y ) } 2

= (a +

(J) 2 ( IX + (J+ 1 )

(1.38c)

- r(a+{3) y"-l(1_y)P-l fyCv)- f(rx)r(fJ)

O~y~l

(1.35 h)

I)
t~
1'1

We note here the following definitions: I. the (complete) beta function;

'l

If a = fJ= 1, the PDF is constant and this special case is balled a standard uniform distribution. . . If Xl and X2 are mutually independent and Xj"'X~j (j= 1, 2) then
2

B(a, [3)= 0 y"-l(I--

y)P-l

dy

rx,/3>0

Xl+X2"'X\,1+\,l

and X1+X2

-~'"

Xl.

eta

(2' 2/

V V2 -1 - )

(1.36)

2. the incomplete beta function

111 )

m9reover Xl /(X 1+ X :2)and Xl + X 2 are mutually independent. (These results will be of importance when we shall derive distributions of various PCls.) . The rth (crude) moment of Y is
I/;.(n
nIX -1-r) f(IX + f/) 1':1y", I '(a)1 '(a I (i I /') IXrrJ (IX I {I)I,.I (U7j

111

IJ,,(I'i.,fJ)"' ,II" 0

"
I(I'-y)/I t dy O<h<]

I'\I

3. the incomplete beta function ratio

1,,((1.,

III

rn B"i,~.jn = J3(o:,If)'

If }' has the PDF (USa) then the CDF is where a[b]=a(a+ l)...(a+b-1) is the bth ascending factorial of a. If r~ -a, It~(Y) is infinite. .
till!

Fy(y)= ly(rx, J) O<y<l f

24

Introduction 1.7 (-DISTRIBUTIONS AND NON-CENTRAL (-DISTRIBUTIONS tribution

t-distributions of

and non-central t-distributions

25

If V,,-,N(O,a2) and S"-'Xva/~ are mutua!ly independent then T= V/S= Ua/S has the PDF f (1)=r(i(V+I))
y}

U+~ T=~
Xv

(1.43)

(vn)!r(1v)

-~yl(V"l)

4 v)

(.1.39)

~
is called a non-central i-distribution with v degrees of freedom and non-centrality parameter Na, The expected value of T' (the rth moment of T about the origin or the rth crude moment) is

This is known as a i-distribution with v degrees of freedom, The t-dis,tribution is symmetrical about zero; thus E[T] =(), Note that V = Ua, where U is a unit normal variable, Hence E[T'] = E [vlrurx~rJ = vlrE [U rJ E [(X~) - I"J
(which is also the rth central moment E ['1'] = 0), The variance of l' is th us (I AO)

E[Tr]

= v~rE[

( + ~)]
U
V

E[(X~ )-!r]

(1.44)

in this c;\se since

In particular
'

var(T)= E[T2] = v.1.(v-2)-1 =.

v (1A1)

v-2

A common and widely used statistic giving rise to a t-distribution is

Erlj
and

(2 r(}) a
,

( ) )r 2 ~
J

\'-1

(1.45 a)

__~(X-iL{ n-1

-,-

i=1

L (X;-X)2

II_

(1.42)

E[T2]=v
If the XI'in the denominator x',().) the resulting variable'

( )
' (j U +-

1 1+ a2 v-2
(52

(1.45 b)

}
of T is replaced by a noncentral
II . I

where X is 'sample mean' as defined in section 1.2 and. XI,.." XII are mutually independent N(~, a2) random vari- ' abies. This statistic ha$ a t-distribution with (n- I) degrees of freedom. If T is defined as above, but expected value of V is not 6 but b, so that V= (U + 6/a)a with U "-' N(O,1) then the dis-

T'=~ II X~(A) ~

26

Introduction and

Mixture distributions

27

has a double non-central t distribution with parameters ()/(J and )" and v degrees of freedom.

114

= (4
"

+ 6Co.L +-J(T + -r::-IIJ

""

. " )

X(J\
.../2rc ,4

exp -- \

()2

1.8 FOLDED DISTRIBUTIONS Generally the distribution of {O+ IX - OH is caned the distribution of X folded (upwards) about 0; 'upwards' is usually omitted. It is used if it is necessary to distinguish it from the folded (downwards) about 0 distt,"ibution which is the dis-. tribution of {O-IX -Ol}. The foldednormal distributionformed by folding N(~, (J2) upwards ahout () has the POF
'I

'

+ 2( C - 3 (J- )Jlt - 3J!1

(I:landl The (1l)(>I)), rolded (upwards)

"

)
.

,2

(1.46 e)

beta (0(,In about

~ has

beta distribution

formed

by folding

the POF

1
B( IX Ii) {x~" '(1 - X)/1- 1 + xll- \(1 -- x )U- \ +
1'<x<1
:

(1.47)

~~-J3~'I exp\ -~,

(--

X-sV
()

1
\

+ exr \

- '2 -~ -";-_c..)

211

\'

(\2\

\( IA() II) x> II

I)

The corresponding folded (downwards) beta distribution has a PDF with the same mathematical form but valid over the interval O::S; x.:(! rather than -t::;;;x~ 1.

If fJ= 0, the variable is IXI = (T Iv + c)I with expected value 1.9 MIXTURE DISTRIBUTIONS J!'l-.:; ,'~ exp - 2'
where [) = ~/(Tand

() ( )
222 J!2 = ~ + (J

2' 1

C)2

(1.46b)

.- ~ll-- 2(\}(- 6))

.)

If 17,(x), F~('(),

Fdx) are CDFs then

k k

the v,i.riance is - J/t

(1.46c)
.I

Fx(.x)=

j=\

L w)<)x)

Wj>O,

j=l

L wj=l

(1.48)

Also

is the COF of a mixture distribution, with k components F\(x), Fdx),ancj weights W\,...,COk respectively. If the PbFs Jj(x) = Fj(x) exist, then the mixture distribution has POF
k

J!3=2

,3

2'

J!1 -~

J!1-

".2 v,.

~2

)J exp - -2- -

(1.46d)

II

fx(x)

L w,Jj(x)
j= 1

(1.49)

2~

/111 willie!

ion

Afullinomwl

(multivariate

normal) distributions

29

The rth crude moment or X is

. k

More generally. e might have a continuous distribution with PDF .I~")(t). Thcn the CDf of the mixture distribution is
(L50)
(1.520)

!-1~(X)= L: Wj!-l),. j= 1

where Il~il's the I;th crude moment of the jth component i distribution. In particular
k

F(x) = J-~<x/~")(I)G(\". 1) dl
If dG(~ dx

E[X]:;::: L:
j= 1

()Jj.~j

(1,51 a)

= g(x, t)

and
k E[x2J=

is the PDF corresponding to G(x, t), the PDF of the mixture distribution is

I: (J)j(~T+O07) j= 1
k 2

(1.51 b)
.

f(x) =

~(D?}(x, t).f'e(t) dt

(1.52 b)

so that
"

var(X)=
-

j~l (J)j(~T+(Jnk '\ L. U).(J2+ .1.1 j= 1 k

j~l (J)j~j

)
(1.51 c)

We denote 'mixture' symbolically by !\ e where 0 is called the mixing parameter. For example the distribution constructed hy mixing N(~,(J2) distributions with (J"'2 having a gamma (0';, distribution will be represented symbolicalJy as Ii)
N(~, (J2)

'\' (J).((.~Z;)2 L. .1-.1 j= 1

1\ gamma (0:,f3)
(f"2

(1.53)

where ~j' (Jj arc the cxpected value and standard deviation, respectively, of the jth component distribution and
k

1.10 MULTINORMAL (MULTIVARIATE NORMAL) DISTRIBUTIONS The [J random variables X = (XI"", xl'J' have a joint multivariate normal, or multi normal distributionif their joint PDF is Qf the form

~'=

j= 1

L wl~.I

(=E[XJ)
function

If Fi(X)

= G(x,

OJ),where G(x, OJ)is a mathematical.

or x tlnd OJ (e.g. a' fUI)t:Uon corn.:sponding to thc N(O,frY) distribution), we may regard a parameter e as having the discrete distribution
Pr[0==Oj]=w.I j=I,...,k

~
II .

Ix(x)=(2n)-lfiIVol-l
wIICI'C' HLX;ISj

exp{ -!(x-~)'VO1(x-~))

(1.54)

(.ice-I,...",)

the, variance-.covariance

i.e. qXJ-'S, ,Inti Vo is matrix of XI"", X fl' The diagonal

30

1ntroduction of X\>...,XI'

SYSTemsf distributions o 1.11.1 Pearson system

31

elements of Yo, (J11>...,(J1'1'arc thc varianccs

respectively, and the off diagonal element (Ji}is the covariance of Xi and X}, with
(Jj} -'- {lij((Jij(Ji.J

This is defined as distributions with PDF U(x)) satisfying the dilkrential equation d(logf(x)) -(a+x) -~~~ - co+Cjx+czxz
(1.58)

( 1.55)

Of course, is the standard deviation ,of Xi~ {lij is the correlation between Xi and X}. For i=1,...,p, Xi",N(~i,On. The quadratic form (x-~)'Vo1(x._;) generalizes the quantity (x-~)z/(Jz= (x -- ~)(J- 2(.\-~) which appcars in .lhc PDl,' of lhc lIn~variatc

normal distribution. The ellipsoids

(x - ;)'V O~I(X-cJ = C2

This system was developt:d by Karl Pearson (1857-1936), in the decade 1890 -1900. The values of the parameters Co, Cj, ('z determine the shape of the graph of f(x). This can vary quite considerably, and depends on the roots of the quadratic equation
CO+C1X+CZXZ=0

(1.56 ) (1.59)

(where CZ is some constant) define contours of equal probability density (I,(x)=cOI1SI.). The probabilily lhat X rails within this contour is Pr[xi, < cZJ; the volume inside the

Among the families (commonly cal1ed Pearson type curves) are

ellipsoid is

'

. (:0>0; Cl = C2 = 0 - normal distribution (section 1.4)

(1.57)

nil' lV Ii a cP { 1(1 +~p) }

. .

(Type III) Cz:":: 0; C1 :;6 0 - Gamma distribution (section 1.5) (Type 1) Roots of (1.59) real but opposite sign -- Beta

distribution (scclion 1.6)

(TVI1' VII) CI=O, co,cz>O this includes thc 1 distribu-

IVai is called the generalized variance of X.

1.11 SYSTEMS OF DISTRIBUTIONS

tion (scdion., 1.7) Thcse families, we have already encountered, and further I
I

details are available for example,in JohnsoQand Kotz (1970,

pp. 9-15).

In addition to specific individual families of distributions, of the kind discussed in sections 1.3-1.10, there are systems including a wid? variety of families of distributions, but based on some simple defining concept. Among these we will describe first the Pearson system and later the Edgeworth (Gram-Charlier) system. Both are systems of continuous distributions.

1.11.2 Edgeworth (Gram-Charlier) distributions These will be encountered in Chapter 4, and will be described there, but a brief account is available in Johnson and Kotz (1970, pp. 15-22).

32

Introduction 1.12 FACTS FROM STAT1ST1CAL METHODOLOGY

Facts from statistical methodology

33

This section contains informat'ion on more advanced topics of statistical theory. This information is needed for full appreciation of a few sections later in the book, but it is not essential to master (or even read) this material to understand and use the results.

1.12.1 Likelihood

are callcd sufficient for the parameters in the PDFs of the Xs. It follows that maximum likelihood estimators of 0 must he functions of the suflieient statistics (TJ,...', li,)::::: Also the T. joint distribution of the Xs, given values of T does not depend 011O.This is sometimes described, informally, hy the phrase, T contains all the information 0:1 ().' If it is not possible to lind all)' function 11('1') f T with expecled value zero, except o when Pr[h(T) :::::OJ 1, then'(T is called a complete sufficient ::::: statistic (set).

For independent continuous variables Xl"", X" possessing PDFs the likelihood function is simply the product of their PDFs. Thus the likelihood of (X 1,..., X,,), L(X)is given by
L(X)=fxJX 1)f:dX
2)'"

1.12.4 Minimum variance unbiased estimators

fx.,(X,,)
( I.60)

"

= rlfdX;)
j= 1

It is also tr'ue that among unbiased estimators of a parameter I an estimator determined by the complete set of sufficient statistics Ti>..., Th will have the minimum possible variance. Furthermore, given any unbiased estimator of y, G say, such a minimum variance unbiased estimator (MVUE) can,be obtained by deriving Go(T)= E[GITJ
i.e. the expcctcd valuc of G, given T. This is based on the Blackwell-Rao 1947; Rao, 1945).

(For discrete variables the f..dXj) would' be replaced by the probability function J>Xj(Xj) here J>,dx;)=PrlXj=xiJ) w 1.12.2 Maximum likelihood estimators If the distributions of X J,"', X" depend ('In values of par{)Z, ... , e ( = 0) then the likelihood function, L, is a s

(1.61)

theorem (Blackwell,

ameters e 1,

function of 9. The values ()1, Oz, ..., Osl = 9) that maximize L are, called maximum likelihood estimators of e J, .,. , Osrespectively. If XI>"',

Example
X" are independent
N(~, (Tz) random

variables then

1.12.3 Sullicicnt stutistks

,

the likelihood function is

. /I

If the likelihood function can be expressed entirely in terms qf functions of X ('statistics') Tl(X), ..., T,,(X)the set lTl>..., Th)

L(X)= I1.--c-exp
;=1 (T,/br.

t X 1: 2 - j -- S 2, { (T } ]

34

Introduct

ion
/I

F acts from statistical methodology 1.12.5 Likelihood ratio tests

35

- (0" iW exp Y L-Iq 1

l-

~2I ,
I

1 (X i-/I

l;)'~

SlippOSC WL: wanl

10 choOSL: hdwL:en

hypotheses

Ilj (i

0, I)

- (aJ2n)"exp - 2a' { J, (X,-X)'+n(X -~)'}J

The maximum likelihood estimator of

(1.62)

specifying the values of parameters ()= (01)..., Os) to ,be 9j= (elj,..., es;) (j=O,1). The two likelihood functions, are
.
/I

~ is
(1.63 a)

L(XIHj)=

':,=X.
i?

i'" ,

TIfx,(xiIOj)

j=O,l

(1.65)

The maximum likelihood ,estimator of a is

/I

The likelihood ratio test procedure is of the following form

1
1

6= { n-l.L,:

(Xi-X)2 }

(1.63b)

If L~~JHd
L(XIHo)

<..c,

choose

Ho

(~, 62) are sufficient statistics for (~,0").they are also, in fact complete sufficient statistic E[4J =~, so ~ is an unbiased estimator of ~. Since E[X IX, 62J = X, 4(= X) is also a minimum variance unbiased estimator of ~. But since

If f~(XIII d
L(X!Ho)

.
;?-:c, Cfl00SC lI,

( 1.66)

E[6J =

rez(n-l))
.

;(

~I~-

(~.)

2 1 ai=a

(1.64)

it is not an unbiased estimator of a. However \ J'(in-I)) 6 :: 2 rein)

()

is not only unbiased, but is a MVUE of a. (There are more elaborate applications. of the Blackwell-Rao theorem in Appendices2A and 3A.)

The constant c can be chosen arbitrarily. If sufficient informa" tion on costs and frequencies with which H 0, H 1 respectively occur is available, it. is possible to choose c to produce oplim:d results in terms of expectedeost. Sorndinlcs (' is chosen so as to make the probability of not choosing Ho, when it'is, indeed, valid i.e. 'making an erroneous density equal to some predetermined value, e, say. T,he procedure is then termed a 'test of Ho against the alternative hypothesis H 1, at significance level e'. Ho is often termed the 'null' hypothesis, even when there is no special meaning attachec,1to the word 'null' in the particular circumstances. .The significance level, I:, is also caned tht: probability of

error of'the 'first kind'. The other kind of error, choQsingH 0

when the alternative 'second kind'. H l' is valid is called an error of the

36

I ntr()duction

BIBLIOGRAPHY Blackwell, D. (1947) Conditional expectation and unbiased sequential estiIl1ation, Ann. Math. Statist., 18, 105-110. Boyles, R.A. (1991) The Taguchi capability index. .TournaiofQuality 'f.'ellllololl.1',23(I), 17 2:\. Burke, R..J.,Davis, R.D. and Kaminsky, F.C. (1991) Responding to statistical terrorism in quality control, Proc. 47th Ann. CO/1f ArneI'.Soc. Quat. Control, Rochester Section, March 12. Chan, L.K., Cheng, C,W., and Spiring, FA (1988) A new measure of process capahility. ./0111'11111 (jlllllily .ft.dllloloilY. 211\:1). (!( 162--175. Dovich, R.A. (1991a) in ASQC Statistical Dipision Newsletter, Spring,S. Dovich, R.A. (1991 b) Statistical Terrorists II - its not sq(e yet, C pk is out there,' MS, Ingersoll Cutting Tools Co., Roc.kford, Illinois, E1andt, R.C. (1961) The folded normal distribution: two methods of estimating parameters from moments, Technometrics, 3, 551-62. Gunter, B.H. (1989a) in Quality Progress, January, 72. Gunter, B:H. (1989 b) in Quality Progress, April. Johnson, N,L. and Kotz, S. (1970) Distributions in Statistics: Continuous Univariate Distribution - 1, Wiley: New York. Kilska, D. (1991) in Qlllllity Prowess, March, R. McCormick, C. (1989) in Quality Progress, April, 9-.10. McCoy, P.F. (1991) in QIUllity Prowess, fo\:hruary, 49 55. Peam, W.L., Kot:.!;, . und Johnson, N.L. (1992) Distl'ibutiomd and S inferential properties of process capability indices, .T.Qual. Technol., 24, 216-231. Ran. C.R. (1945) Information and the accuracy attainahle in (he estimation ofstatislical parameters. Hli/I.('tI/c1l1/1I llih. .)0('" J7. M . 81-91. Spiring, FA (1991) in Quality Progress, February, 57- 61. Steenburgh,T. (1991)in Quality Progress,January, 90.

The basic process capability

indices: Cp', Cpk and their modifications

2.1 INTRODUCTION

MlIch of the material presented in this chapter has also appeared in several books on statistical quality control, published in the last decade. These books usually provide clear explanations and motivation for the use of the pels they describe, and often (though not always) emphasize their limitations, and provide warnings intended to assist recognition of situations wherein extreme caution is needed. In the last few years, also, many issues of Quality Progress contain letters to the Edit<)r in which the use of PCls is criticized (and only occasionally defended), often combined with passionate appeals for their discontinuance on the grounds that their inherent weaknesses guarantee widespread abuse. Some referees for this book have even warned us that we may cause serious damage by further propagating the inadequate and dangerous concepts implicit in use of PCls. . Our opinion is that we hope (even believe) that we may provide background for rational use of PCls, based, indeed, on knowledge of their weaknesses. We do not, of course, recommend uncritical or exclusive use of these indices. The 37

38

Tl1e h([sic !Irocess C([fli/hilit.\' indices

quirement index.

The Cp index (LSL < X < USL)

IS

39

backlash against PCls seems to CO.11erom two broad, 'and f opposed, sources: (i) a lack of appreciation of statistical theory and its applications; and (ii) a demand for precise statistical analysis, to exclusion of other constraints. So.urce (i) does not fully comprehend the meaning of the indices, and resents pressure from above to include 'meaningless' numbers in their reports - which may, in,deed, contradict the actual state of the process as seen by experienced operators. They regard this as part of the 'tyranny of numbers' which has its roots in the proliferation of statistical (i.e. numerical) information now becoming available, and tending to dominate our existence. Often, this resentment underlies the unpopularity of, and lack of respect for 'statistics' among otherwise enlightened engineers, technicians, sales managers and other professionals. Source (ii), on the other hand, arises from those who are well aware that assumptions, on which P('ls :Ire h:lsed. :II'C often violated in practice. 'l'his is also true of many olhl:r statistical procedures (such as ANOV A) which, however, are better tolerated, and more widely accepted. albeit somewhat grudgingly. After these introductory remarks, we now proceed to define and examine the earliest form of PCI, generally denoted by CpoThe background information in Chapter 1 will assist in understanding the basis for many of our results especially in regard to estimators of PCls - but they are not . essential for general comprehension.

the

process, capability

USL-- LSL C -----

- p --

60'

(2.1 a)

I I I
.

where (Jdenotes the standard deviation of X. Clearly, large values of Cp are desirable and small values uudesirable (because a large standard deviation is undesirable). The motivation for the multiplier '6' in the denominator seems to be as follows: If the distribution of X is normal, and if ~, the expected value of X, is equal to !(LSL+ USL) - t.he mid-point of the specification interval - then the ex-pected proportion of NC product is 2<1>(diu) where d = -!(USL-- LSL) - the halflength of the specification interval. From (2.1 a) we see that d Cp= 30'
(2.1 b)

~ .
I' ~I
~

so the expected proportion (p) of NC product (assuming ~= t(LSL + USL)) is

2<1)( 3 Cp)

t'
~I

(2.2)

2.2 THE Cp INDEX

Consider a situation wherein there arc lower alld lippeI' specification limits (LSL, USLrespectively) for the valu.e of a measured characteristic X. Values of X outside these limits will be termed 'nonconforming' (NC). An indirect measure of potential ability ('capability') to meet the re-

II III'

h II t

If Cp= 1 this expected proportion is 0.27%, which is regarded by some as 'acceptably small'. In fact, it is often required that for acceptance we should have Cp::::; with c = 1,1.33,or 1.5 c corresponding to USL - LSL = 60',So' or 90'. It is important to note that Cp= 1 does not guarantee that there will be only 0:271<) NC product. In fact, all that it of does,guarantee is that, with the assumption of normality and the relevantvalueof u. there will neverbe lessthan 0.27'10 .expected proportion of NC product! (It is only when ~=1(LSL+USL) that the expected proportion is as small as ,

40

Till'

/Josi(' fJI'()(,I'SS (,OfW/Jility

indi('('s

The Cp index

4)

0.27;().) What the value Cp = I does indicate is that it is . possible to have the expected proportion of NC product as small as 0.27%, provided the process mean ~ is pn~ciscly' controlled at ~(LSL + USL). . More recently, the view has been put forward that expected proportion of NC woduct is not (or, perhaps, should not he) the primary motivation in use of PCls, but rather that loss functionconsider~ltions should prevail. Although there arc some attractions in this attitude, it does not explain the multiplier '6' in the denominator of (2.1 a). Also, uncertainty in actual proportion NC is balanced (perhaps more than balanced) by uncertainty in knowledge of the true loss function. The form of loss function by far the most favoured is a quadratic function of ~, though little evidence for this choice has appeared. Also, if one is really interested in a loss function, why not just estimate its average value, and not some unnecessarily complicated function (e.g. reciprocal) of it'! Indeed, there would seem to be no need for LSL and USL to appear in the PCI formula at all (except in as far. as they may define the loss function). The title and content of a recent paper (Carr (1991)) seem to imply, not only that the original motives underlying definition of PCI have been forgotten, but' that they are now returning. Constable and Hobbs (1992) also define 'capable' as referring to percentage of output within specification. It has even been suggested that 2<l>( 3Cp) itself be used as a PCI, since it is the minimum expected (or the potentially attainable) proportion of Nt items provided that normality is a valid assumption. Lam and Littig (1992) on the other hand. suggest defining a PCl
tpp=.\<)-I(1(r> I I))

based on an estimator of p from the observed values of X. Wierda (1992) suggests using --t<f)- 1(p). Appendix 2.A contains discussion of estimators of p, based on the assumption that the process distribution is normal. Herman (1989) provides a thought-provoking criticism of the PCl concept, based on engineering considerations. He distinguishes between the 'mechanical industries' (e.g. automotive, wherein Cp first arose) and the 'process industries', In the former, 'emphasis is on the making of an assembly of parts with little if any measurement error of consequence', while this is not generally true for the process industries. The (Jin the denominator is intended to represer,it process variability when production is 'in control', but thd quantity estimated by
S2=-

L (Xi-X)2 n-l}=1
.,>,for lot"

II

(see (2.3) below) is (assuming no bias in measureme~t error) equal to (J2+(variance of meas~rement erro~). \
Herman further notes that If allowance IS mad

to-lot variability, both (J2 and the second term ave two components - from within-Jots and among-lots ariation. 'd ~ Again, the (T in the denominator or Cp is intcn dC t< rerer to within-lot process variatiol,1. This can be consider lbly less than the overall standard dcvip.tion -- (Jtotabsay. \ Herman ,suggests that a different index, th~ 'process perf9rmance index' (PPI).
, ,

USL-LSL p =---.
P

6 (Jtotal

( ,'

and using the estimator Cpp=!CD-1(!( p+ 1))

i,

might 'have more value to a customer than Cp'. . Table. 2.1 gives values of the mipimum possible e~pected

I
proportIon (2<1>( 3C,)) of NC -

l(ems correspo (Og.

to

6'
0 '"
"""I'>' r-

Estimation

oj C p

43

0..
0..

;':::::0 0
'"

'-":r'" ..... rr,

S
NIl'>

0..

selected values of Cp, 011the assumption of normal variation. One can see why small values of Cp are had signs! But large values of Cp do not 'guarantee' acceptability, in thc absence of information about the value of the process mean. Montgomery (1985) cites recommended minimum values for Cr, as follows: . for an. existing process, 1.33 . for a new process, 1.50 For characteristics related to essential safety, strength or performance features - I'or example in manufacturing of bolts used in bridge construction - minimum values of 1.50 for existing, and 1.67 for new processes, are recommended. General1y, we prefer to emphasize the relations of values of PCls to expected proportions of NC items, whenever this is possible.

'.:!(o.. 0 0

~ i7\
..,fV)
""" II

s
Or-O ONO . r. -ON
II

";~

2:

'"

~
-II'>

.
0..

E
.~

~
25

Z
\
I

0 II

2.3 ESTIMATION OF Cp

>

'0
i:: 0

r- 2:
NIl'> V) r""!
II

The C'i1ly parameter in (2.1a) which may need to be estimated is (J"the standard deviation of X. A natural estimator of (Jis
(J'

.-

~ - 'b 0 .0
g. ....
0..

-r

'S

] .s i ~ 2: ~ O~ ~ E ~ ~ 25 NOo II E --;:I

In-I
~
'"2

\.
/I

L., j= I (Xi

x )".r

(2.:1)

where

.8
::= S ... g,
'C' i:: e<j 0II

X =-

I Xj n j= 1

/I

~
QJ
:;s

If the distribution of X is normal, then 82 is distributed as

-(J2 11- 1 x (chi-squared

""

U I"'i",
I

Os
00i:...J

-..;....

f-o

-~

""oS< UN

with (n-I)

degrees of freedom), (2.4 a)

\
\

44

The hasic process capahility indices Estimation of Cp (2.4) and (2.7), we have
,

45

or, symbolically
_(J2

n-l

Xn-1

(2.4 b) pr C\,> C

[ c"

] = Pr[x~ - 1 <

(n -- I)c - 2 J

(2.8)

(see section 1.5). Estimation of (J using random samples from a normal population has been very thoroughly studied. If only

2 S. ,Inee I r.Xr7--1 ~X"-1.1: ] =i:, we h ave [

.

'

-1 6 3 Cp = USL-LSL (J=d(J

(2.5)

(J 2 PI' n-lX"-1."/2~a

-2

(J ,2 ~n-lXn-l,I-"/2

J -I-a

The interval had been us~d as a PCI, existing theory could have been applied directly to

(
6

J~)82
2

X,,-I,

1-,,/2

'

~'I)cr~
2

Xn-I,"/2

)
(n-l)l -

(2.9 a),

-Cp

6
=USL-LSL

38

l1=d

(2.6)

is a lOO(l-ap,-;) confidence interval for 0 2; and so the int-.:rval

which is distributed as

(.iJSL'='-LSLX~~~.~~~~' USL- LSL X:,:~a )

1{Xn-l/J(n-l)}

(n-l);.,

(2.9 h)

C;
Analysis for

is a 1OO( - a)% confidence interval for C; 1

1;

and

(
C =USL-,~,SL=_~=:,:C " 68 38 8 "
(2.7)

~JSL-=-~~!:: X"':J..",/2 USL- 'JSL Xn-1, 1-,,/2 . 68 (11-1)1' 6cr (n-l)! = Xn-~c" (n I)'

)
(2.9c)

Xn-l,l.-~/2C (n-l)'

(distributcd as C"~I";x,, I) is a hit morc complicated, hut not substantially so. Nevcrtheless, there are several papCl:S studying this topic, including Kane (19X6),Chou ('( (/1.(I <)90), Chou and Owe'n (1989) and Li, Owen and Borrego (1990). \V-.: give here a succinct statement of results for the distribution of Cp, assuming X has a normal distribution. Fig. 2.1 (see p. 65) / exhibits PDFs of Cp for various vah.;es of d/(J= 3Cp. From.

is a 1O0(1-iX)% confidence interval for C". Lower and upper 100(1-a)iI'~1confidence limits for CI' arc, of course,
Xn ~C- 1" and (11-1)~ 'p (
I

X,-1,1-,,-C " !(11--1) p

.

I
I

respecti vely. ...

46

The basic process capability indices

Estimation

of Cp

4.7

lI' t:lhks of percen[age points of [he chi-sqll:lred distrihlltion arc not availabk, it i.s necessary [0 use approximate formulae. In the absence of such tables, two commonly usecl approximations are X,.a ~(" - ~)! + -_:'F ( Fisher)

. III
.Il'

Assuming normality he introduces the approximations

tt!

f
e LS I J~ (I

I 41. I)} n1

1.2.! I lIj

-)2

,I)
~

var(S - 1 )~ 2(n - 3)0-2

(cf. section 1.2) From (2.11 b) we would obtain the approximation

(2.11 b)

and
X".a~"\

l-9,,+za

( ))
9~'

\ J (Wilson-Hilfel'ty)

tv)

where (l)(za)=ex. (cf. section 1.4). With these approximations we would have (from (2.9 c)) the

1 C~ var(Cp)~2 (11--3)
I

(2.11c)

100(l-ex)It) confidence interval for Cp:

Ilhl:
'

(j

However, Heavlin (19'88),presumably using terms of higher order than n-l in (2.11b) obtains

(n~l)\Cp((n-iY1

Zjr}
3 \

((n-2

var(Cp)~2(n-3)

c2
p

1+n-1'

(,

(2.11 d)

\,

(n-1)jC,
(using Fisher) or

ZI-~2\

fi)
2

(2.10a)

1, II

The interval

Cp ,1- 9(n-1j-Zl-a/2

9(11-1)

))

(\{I

C(n~3j.(1

+~~-i)YZ1'a()}'

Cp{1+(2(n~3) Cp 1-9(m-1)+ZI-a/2

(1+ n~)YZ1-a/2}

\ (2.12) I

'

9(n-1)

))

~ ; (2.10 b)

IS 11'1
11

(using Wilson-Hilferty) Heavlin (1988), in an unpublished t.echnical report available from the author, attacks the distribution of S-1 directly.

ill filii

all r I
,

is suggested as a 100(1- ex)(Yt) confidence interval for Cpo Additional generality may be gained by assuming solely that 8 is a statist'ic, independent of X, distributed (possibly approximately) as (xflJI)(J. (The above case corresponds to f= 11- 1.) This includes situations in which ()' is estimated as a multiple of sa.mple range (or of sample mean deviation).

48

The basic process capability indices Estimation of Cp

4~

From (2.7), Cp is distributed as

where (2.13)
II

USL-LSLy:[=y:[Cp,
60" Xf Xf

bf= The estimator

()

I'(ln

l rcfU-l))
(2.17a)

so G=(Cp/Cp)2 is distributed as flx}l, and the probability density function of G is

'

,
c~ = hfCp
has expected value Cp and so is an unbiased estimator of Cpo Its variance is

, .,

f ( If ) =..
('

,,'"

f If

---, i f

\ r"

eXl)-'

:2 '/1 'dI)'

f If

'

~'I

0 <'
'

If
,

(2:l4;
.Ii

.', '

The rth moment of Cp about zero is E[C~,] =f \rC~E[X;rJ =f lrC~E[(X,7r !rJ

. (C
,
A

- ffb}
-If-

Vdl

I (' 2 2- } p

(2.17 b)

()
'.1'

\" rnU-r))
J'( 1,n

cO' I' .

(2.1))

Table 2.2 gives a few values of the unbiasingfactor, bf' For f greater than 14, an accurate approximate' formula is b ~l-~

In particular

4'

f '-l

(2.18)

f-l E [. CPJ= I.

L)' r(2'" (2

'

~ r

(cr (1.19a)). With this approximation

CP= 'j;-Cp f

(2.16a)
,Jj . . " ~' ,((XI 1,_9) 2 ('I~ Vd I( ( p) '- (I":' 2)(4.1 - 3)

(2.1 ()

1I)

2,
f
2

E [C p ] =.1'- 2 C p ,
and

(2.16h)
/'1

8/+9 var(C~)~ 16f(f-2)

C2 p

(2.19 b)

var( C1~) =

(f! 2 -bi

)c~

(2.1()c) ,I

Table 2.3 gives numerical expet:ted values and standard devi~tions (S.D,s) of ,Cp(not C~) for the same values of f as in Table 2,2, and d/O" 2(1)6. (Note that d=t(USL-LSL)). = Since E[CpJ/Cp and var(Cp)/C; do not depend on Cp, one could deduce all values in Table 2.3 from those for Cp=1(d/0"=3), for example.

The
-r-0\ 00 lI) 0\

Cpk index

51

The population ("true') values of Cp are thc expected values

forI'/',

0 0 0

\0 '<too lI) 0\

vi

0'<t~0r<) r--00r<)00\0 -o\r<)~""'OOlr\r<)~""'o\o\ r<) V> '<t r<) r<) ~ ~ ~ ~ ~ "";00000000000

~.

. I i
0 o.=:>

lI) 0\ 00 '<to\

IJJ 0
vi

r--0\~r<)V>'<t~0\V>':"'00~0 000 00~V>'<tr<)r<)r<)~~0 V> 0 0'0 0 0 0 0 NNNNNNNNr--iNr--i.NN

('I 0\ ~ '<t 0 '<t -.:I" 00 ~ 'n

rr, '<t00 '<t0\

0 0\ ~ 0 ~ r<) ('I 0\ 00 r-- ~ '.I'I '<tr<)r<)('I r--I - - '""" "";6000000::;;OCO

00

0 0 0

v>1
.....

O\'<tr<)~--V>O~r<)ooor-1.!.1

OON~r<)~-OOO\O\O\OO~ Ooor--r--r--r--r--r--~~~~~
N"";"";"";"";"";"";"";"";"";"";"";""; '<t~-OO\r---O\ooOr<)r-r-- 0\ 0\ '<t 0 co r-- "1

I'
t 1;-

0\ 00 M 0\

00 '<tt-M 0\

'V '-'

0. V

0.
,

0. V r<)
'-'

0
'<tl

vi uJ

00r<)~~~-------

"'" '<1" (""'. N

a
0
0 \D

II

~
t I

'<t
0\ t--

vi II vi
;;:;'
<7i' <U .... '"
"" "-

- r<) ~3:::;:;:;:J:;:;:;:J:;:
0000000000::;;0
0\ 0 0\ r-- 0\ '<t 0 r--v>-oor--~\C~lr\V>V>II')r<)

r-- '<t ~

The distribution of ('p when the distribution of X is Ilul normal will be considered in Chapter 4. We note that whell f = 24, for example (with n = 25, if /=n-I) the S.D. of Cp is very roughly 14--15;;1of the expected value (i.e. coef-I-icient f variation is 14--15 o %). This represents a quite considerable amount of dispersion. As a further example take Montgomery's (1985) suggested minimum value of Cp for 'existing processes', which is 1.33 (see the end of section 2.2). The corresponding value of d/a is 4. If a sample of size 20 is available, we see (from f = 19 in Table 2.3) that if Cp= 1.33, the expected value of Cp would be 1.389 and its standard deviation is 0.24. Clearly, the results of using Cp will be unduly favourable in over 50% of cases actually we have from (2.8), with c = 1,
Pr[C;p> 1.33J= Pr[XI9 < 19J = 53.3% Un is as small as 5

NO;
0

"-, -Irl

~ --I

'00
'<t \0 N 0\

'v l

,.-, 0. "-I II IJJ

J

0
r<)1

VJ

'n 0\ - 00 ".1', r 1- - 0 0 0\ --t 000000000000

1.0<"I<"I~

'n

r-- 00

- - - - ".1', --0 00 0\

"n

u= 4) and

Cp = 1.33, then the standard

;
t:'

d
1

r<)-.:I"oo~r<)r--r<)Or--~'<tr<)O

X
s;;

0 0\\0 0\

uJ

0.

iq~8~q~~~CiCiCi$~ ------------r-- 00 II') 0 r<) 0\ -.:I" ~ -.:1"

0 :a ::I

N 'i
.lS
l

'V '0
.:!J I:: 0

'<t~
0\

";J

> <1) i3
...:-

vi

0000::;;0000000

'<t -.:I" ~ "', "'<t 0 r-- (""', 0 0\ 00 r-- r-- r-- \0.\0 I 0000000

C3
'" 0 ::1 (; >0\

d
'<t
..... 0\

"tf')
II' ::

~
I"!

S 0

[.l.J

~ 0\ II') '<1" ,00 ".I', ~ 0 00 r-- ~ ..-, r-r<) ('1 0 0\ 00 00 00 00 ['--. r-- r-- (,- ~ oor--r--\.O~~~~~~~~~

deviation of Cp is O.~ - over half the value of Cp! 'Estima7 tion' in this situation is of very little use, but if n exceeds 50, fHirlyreasonable accuracy is attained. Another disturbing faclor possible elrccls of nOll-normality - will be the topic of Chapter 4. 'Bayesian estimation of Cp has been studied by Bian and Saw (1993). r'

oddoo~:i.doddcoo

d
00 0\ '<t I'c:

N
<:J

",<;; ,.c;:

(j)
::

~
C':I f-<

2.4 THE Cpk INDEX

'-., '<t 0\ '<to\","O'\-.:I"O\-.:I"O\",,"""'- - ~ r~ r<),.-,"""-q' II')V> g

II

..!::::J'-,

'0

The Cp index does not reyuire knowledge of the value of the process mean, ~, for its evaluation. As we have already noted, there is a consequent lack of direcl relatioll between Cp and probability of production of NC product. The Cpkindex was introduced to give the value of ~ some influence on the value

52
The basic process capability indices The Cpk index
...

53

of the PCl. This index is defined as

Cpk=

min(USL-~, ~- LSL)
3a

(2.20 a)

= d -I ~-!(LSL + USL)I (using (1.14'e)) 3a

'I

(If ~ were outside the specification range, Cpk would be I1cgalive, and the process would c1earlybe inadequate for controlling values of X.) If the distribution of X is normal, then the expected proportion of NC product is

= 1

1~-!(LSL+USL)I
d

c" }

(2.2017)

<I)

LSL-~

--;;-

+ {1-

<I)

(-~

USL--(

)}

(2.21 a)

Since C,,=dj(3<T),we must have C"k~C", with equality if and only if ~= t(LSL + USL). Kane (1986) also defines ,'C"k' for a general target value T by replacing i(LSL + USL) by T in (2.20b). This index shares, and indeed enhances, the feature of Cp' that small values of the index correspond to worse . quality. The numerator of Cpk is the (signed) distance of ~ from the nearer specification limit. Some authors have used the erroneous formula min(I~-LSLI, IUSL-~I) 3a
(2.20 c)

From (2.20a) we see that, if t(USL + LSL) ~ ~~ USL then CPk=USL-~ 3a and

LSL-~ _. (USL-- ~)-(~L~ 3a

L_~~= Cpk-2Cp~ - Cpk

(since Cp~ Cpk)' The exact expected proportion of NC pr'oduct can be expressed in terms of the two PCls Cp and Cpb as follows:
<1>(- 3(2C"-C,,k))! <(--3Cpk)

for Cpk' This does, indeed, give the same numerical value as formula (2.20a), provided'; is in the specification range, LSL to USL. However, if, for example, ~=LSL-d or ~=USL+d

(2.21 ")

lienee thc expected proportion or NC product is less than

2<1>(

formula (2.20c) will give the value dj(3a) for Cpk - the same value as it would for ~= !(LSL + USL), at the middle of the specificationrange. The correct valuc, as givcn by (2.20f/) is ~dj(3<T). We will assume, in the following discussion, that LSL~~~USL

- 3Cpd
- 3Cpk)

(2.22 a)

,
.
1*

but it is greater than

<I>(

(2.22 b)

The case LSL~~~!(LSL+ USL) can be treated similarly. Altho~lgh the preceding formulae enable information about expected proportion NC to be expressed in terms of Cpk(and

54

The basic process capability indices

Estimation

of

Cpk

55

Cp), it can also (and more simply) be expressed in terms of the process mean, ~, and standard deviation (J. Porter and Oakland (1990) discuss the relationships be~ tween Cpkand the probability of obtaining a sample arithme~ tic mean, based on a random sample of size n, outside control chart limits. Although it is useful to be aware of such relationships, it is much simplcr to cxprcss thcm in terms or process mean and standard deviation, than in terms of PCls. Relationships between Cp and Cpk are also discussed by Gensidy (1985) Barnett (1988)and Coleman (1991). .

the reduction to a single index had to be paid for by losing separ:lte informatio(J on ]ocatio!1 (~) and dispersion (rr). Another way of approaching Cpkis to introduce the 'lower' and 'upper' PCls.

Cpkl--- ~-LSL 3(J . USL-~

C pk
II

(2.25 a)

-=---:,-;

(2.25 b)

2.4.1 Historical note

and to define
Cpk = min(Cpkh
Cpkll)

It seems that the original way of compensating for the lack of input from ~ in the calculation of Cp, was to use as an additional statistic,
k= I~ - i(LSL+

(2.25 c) than Cpkt. . T

Kane (1986) uses the notation CpktJ. He also introduces indices

Cpt. CPIi rather USLTCpktJ=~

LJSL)I

(2.23)

TCpkl=3(J

T -.-LSL

and

(Kane (1986) defines kJ~-TI but assumes T=i(LSL+

.

(our notation).

A similar PPI to Pp (seesection 2.2)can be constructed by

modifying Cpk to

d P pk'-- min(~ -;-LSL, USL -- 0

3(J total

USL) in the sl~quel.)

This was equivalent to using ~ (through Id and (J(through

2.5 ESTIMATION OF Cpk

Cp) separately. While it might be said that one might as weB use ~ and (Jthemselves, or their straightforward estimators, s: and 8, respectively, k and Cp do have the advantage of being dimensionless, and related to the specification limits. With their combination into the index Cpk=(I-k)Cp (2.24)

A natural estimator

of Cpk is

Cpk= d-IX-i(~;L+ where 6-is an estimator of (J.

USLD

(2.26)

56

The hasic process capahility indices

Estimation of Cpk

57

If the distribution of X is normal then not only do we have a di~ributed (perhaps approximately) as XJ(J ',,;1]', ut also: / b (a) X is 'normally distributed with expected value ~ and standard deviation (J/Jn; and (b) X and a arc mutually independent.
From (2.26), using (a) and (h). we find

t I II

cxr5) - n{~.=l(~~~_+~L~~21~1 !5-~(LS~+USL)1 2(J2 (! L )(

t
~

x {1- 2<1> (

ftl

~-t~LSL+ USL)I)}J

(2.29)

ECC~'kJ=;rE[a-rJ .

()

(-l)j

()
~

.I

dr-jE[IX~W~SL+USL)lil

and
A

(~ )
3(J

rE[Xf'J

j=O

f
9(f-2)

(-l)j'

()
~

.l

var(Cpk)

[(~ -

) (~)[( )
-2

Jm

xexp x

n[~-'hLSL+
2(J2

lISL)J2 }

Cft)' [I-f'-lX E

(2.27)
-

1~-t(LSL+ lISL) I +--(J

~(~SL + USL)}

IJ

II

The statistic JIll j( -l(LSL + USL)I/(J has a 'foldcd normal' distribution (see s'ection 1.8), as defined by Leone, :Nelson and Nottingham (1961). It is distributed as

x {1-24>( -ftl~-!~SL+
I ,

lISL)I)}J
1'"1
'
,

'.
I

+ -~~

.f ~ .,.-}(LSL ,+ tJSL)1, 2
(J 2

~.;":_.,-"" L- + "'

;1! I

. " .' .' >,1,....I

P ( ("
"

'1 ), ,1~

(2.30)

IU +61
where U is a standard normal variable and

(see also Zhang et aZ.(1990)) If Wc usc

<5=';;/- {(-t(LSL+ lISL)} (J

Taking r= 1,2 we obtain
,

(2.28)

a=[---=-i" i= 1 (Xi-X? J , L n
I,

L"

~,

then f = n- 1. Some numerical values of E[CpkJ and

'-'I

()

I'

(f 2
,

) ~_ 2 )
', ,
,

S.D.(Cpk)are shown in Table 2.4 and corresponding values of Cpkarc shown in Table 2.5.

'

Note that
(a)

Cpk

is a biased estimator of Cpk'The bias arises

from two sources:

Eca-1J=bj:l(J-li=(J-l,

E[C,.] ~3 2

rG)

nn,

This bias is positive because hr< 1 -, see Table 2.2 ~

5~ (b) E ftIX
,

Th(' !Jusic I1I'0c(,sSC(/flu/lilily

il/dic('s

~t(LSL+ U~L)I - ftl~-J(LSL+

a

USL)I O. ~
II

This leads to negative bias, because ftlX' ~-!'(LSL+ , USL)I has a negative sign in the numerator of Cpk. The resultant bias is positive for all cases shown in Table 2.4 for which ~=I=t(LSL+ USL)=m. Wheh ~=m, the bias is positive for /1= 10, but becomesnegativefor larger numbers. Of course, as /1-HX),the bias tends to zero. Table 2.6 exhibits the behaviour of E[CpkJ, when d/a=3 (Cp=Cpk=I), as n increases, in some detail. Studying the values of S.D. (Cpk) we note that the standard deviation increases as d/a increases, but decreases as 1~-t(USL+LSL)1 increases. As one would expect, the standard deviation decreases as sample size (Il) increases. For intcrprclalion of cakul:1led valucs of (1'1' il is especi;lIly importun,t to bear in mind the standard deviation (S,U.)of the estimator. It can be seen that a sample size as small as 10 cannot be relied upon to give results of much practical value. For example, when d/a = 3 (C p = 1) the standard deviation of Cpkfor various values ()fa-ll~--t(LSL+ USL) is shown in Table 2.7, together with the true value, (;pk' Even when n is as big as 40 there is still substantial uncertainty in the estimator of Cpk>and it is umvise to use

J
(~ ~

a
C/i

00000

O\V)-<O<:t" -<V)I")N"""-NI")<:t"

I") 00 I") 0\-<1"--1")0 0

I"-- V) N I")

00000
0\ V)

O\O\I")\oN I"-- 0\ <:t" 0\ 00,""""""

00000

or, N

1.

0 V)0<:t" 0 \0 1"), 0\ 01")1"--0<:t"

000"";"";
~.

O"",v-,ooCV)OV)...... OI")I"--O<:t"
000"';"";
Of""'.<:t"

Ol"--V)NO\ 070\'</"00 0 I") 'oJ:) 0 000"";"';

I")

I'
V,?

a
C/i

O\-<V)I"--N NO'\I"--\O\O -< -< N I") '</"

00000

0<:t"01"--<:t" N N

o
I")

00000

"<t00\'</"...... ooN\ONOO 0 -< N

00000

N.

I.:U

('01 I"-- ('01 I"-- ...... oo7"""1"--<:t" V) 0\ N '.0 000"";"';

'.0 0\ N <:t" I"-I"-- "I 00 f""', 00 'II 00 I"~ 'II 000"";"";

00 I") I"--V) 1"--1")0 V)

""2'"",,00V)1") 1"--1"1\0 -V)ooN'n

\0

000"";""; 0\""2'\01")0 O\<:t" 0\ V)...... 0 < N I")

-'"

's:,c-I-?,

a
C/i

"'-0'</"0 V) I") N -NI")<:t"

...... V)

00000
'r, 0\ 7 oJ:) 1"1 0\

I"-- N I") I") 00000

00000

vj II "V'
J,

I
r:

a c.ri

I.:U

I") r- c <:t" 00 do"";,"';"";

0\ '</" V) N

cr, or, 00 0 V) 0 V) -< M r--O<:t" 00"";"';""; -t'~0\0\

I") \0 1"--,

!;;~~&j~
I") \0

I")

I"--

00"";"";"';

II

,:-, .. c'\.) I-..J U.J II i.U

u:i
or.

;;t?2~~~
-<

00000
("I ",

I")

<:t" V)

00000

<:t"O\OM N N

I")

00 cr. G\

00000
0
("I

0\0\<:t"-0 -\0 N """-NNM

001")

0 I.:U

\0

pk,

as a peremptory

guide to action, unless large sample

sizes (which may bc regarded as uneconomic or even, infeasible) are available. Chou and Owen (1989) utilize formula (2.20a) as their starting point for deriving the distribution of Cpk' We have
Cpk'=min(Cpkl> CpkU)

"'" ~,

.7.~~1~8
00"";
00 00 ('I 00 N

.-

,...; ('oi

I"-- 0 1"1 'r. (" 1 C/j cr, 00 'r, 00 (I 'r, 00"";"";"";

I"-- 'n

\0 00 (" J 00"";"";"";

'r.

N \0

'r. 0\

0'\ 0

'0 '"
Q)

Q C/i
0

00000

00 \0 "1' 1- 1- I"-I") <:t" 'n

("'I '"T

00000
I") V) I") 00 \00\1")\00

00 'r, 1"1 r 1 c')

- - (')'I">
("

O\\O'</"I")N .I"-- f""'. 0\

- - 1"1 cr. ('" II">' 00000

f"')Or-:'<:t"N

II

s c
Q)

(2.31)

I
II

:?: ""2'

I.:U

00 NI"--N \0 (')0\01")0\ '.0 0 I") I"-- 0 o"";"";"";('.j

0001") I") O\<:t"

0
V)

0"";"";

('.j
01;:,

I") 00 N I"-\00\1")\00 00"";"';('.j

~i

where
~
'

:is
<.':S

c -

X
and

Cpkl=-=L~L 38

cpku--USL-X 38-

r-o

::::

II ~

I")

<:t"

V)

\0

-<--, II ~
I;:,

N ---::::

('01

I")

<:t"

V)

\0

::::

II ~

I")

<:t"

V)

\0

Table 2.4 (Cont.) Imlla 0 E

11= 25

0.5 S.D. E S.D. E

1 S.D. E

1.5 S.D. E

2 S.D.

d(J
")

, .'
-+

0.634

.;; 6 11=30 da
") -

0.978 1.322 1.666 2.010

0.105 0.154 0.205 0.256 0.307

0.516 0.860 1.204 1.549 1.893

0.104 0.148 0.195 0.245 0.295

0.344 0.688 1.032 1.31" 1.721

0.087 0.126 0.171 0.220 0.270

0.172 0.516 0.861 1.205 1.549

0.074 0.105 0.148 0.195 0.245

0.000 0.344 0.688 1.033 1.377

0.070 0.087 0.126 0.171 0.220

, .)

-+

:5 6
11=

0.635 0.977 1.319 1.662 2.004

0.095 0.139 0.184 0.230 0.277

0.513 0.856 1.198 1.540 1.882

0.094 0.133 0.175 0.220 0.265

0.342 0.68:5 1.027 1.369 1.71]

0.078 0.113 0.154 0.198 0.242

0.171 0.513 0.856 0.198 1.540

0.067 Q.094 0.133 0.176 0.220

0.000 0.342 0.685 1.027 1.369

0.063 0.079 0.113 0.154 0.198

35
0.636
.

da

")

, .)

0.977

0.087 0.127

0.511 0.852

0.086 0.122

0.341 0.682

0.072 0.103

0.171 0.511

0.062 0.087

0.000 0.341

0.058 0.072

-t :5
(1

1.318 1.659

0.169 0211

1.193 1.534

0.161 0.201

2.000 0.253 1.875 0.242

1.023 0.141 1.354 0.r81 1.70:5 0.222

0.852 1.193 1.534

0.122 0.161 0.201

0.682 1.023 1.364

0.103 0.141 0.181

-"'

--

- -- 0.170 0.510 0.850 1.190 1.530

-.

n=40 dla 2 3 4 5 6 11=45 dla 2 3 4 5 6 11=50 dla 2 3 4 5 6

0.637 0.977 1.317 1.657 1.991

0.081 0.119 0.157 0.196 0.235

0.500 0.850 1.190 1.530 1.870

0.080 0.113 0.149 0.186 0.225

0.340 0.680 1.020 1.360 1.700

0.067 0.096 0.131 0.168 0.200

0.058 0.080 0.t12 0.149 0.186

0.000 0.340 0.680 1.020 1.360

0.054 0.067 0.096 0.131 0.168

0.638 0.977 1.316 1.655 1.995

0.076 O.l1t 0.147 0.184 0.220

0.509 0.848 1.187 1.526 1.865

0.075 0.106 0.139 0.175 0.210

0.339 0.678 1.018 1.357 1.696

0.063 0.090 0.122 0.157 0.192

0.170 0.509 0.848 1.187 1.526

0.054 0.075 0.106 0.139 0.175

0.000 0.339 0.678 1.017 1.357

0.051 0.063 0.090 0.122 0.157

0.639 0.977 1.316 1.655 1.993

0.072 0.105 0.139 0.174 0.208

0.508 0.846 1.185 1.524 1.862

0.071 0.100 0.132 0.165 0.198

0.339 0.677 1.016 1.354 1.693

0.060 0.085 0.116 0.148 0.182

0.169 0.508 0.846 1.185 1.523

0.051 0.071 0.100 0.132 00165

0.000 0.339 0.677 1.016 1.354

0.048 0.060 0.08 0.116 0.148

Estimation Table 2.5 Values of Cpk

(?('Cpk

63

q C/)
I
\.IJ

'Ot---o7 <noo-"<t
000-00000
000'O"<t~ O<'">t---<n 0<'">'00<'">

"<t "<t

000-00000
0

"<t 'n

t-t--

<n 0

"<t ,')

\Eml
2 3
4

0.0 2
Il 3

0.5 1

1.0
1 3 2 :1

1.5
l 6 l 2 6 '.

2.0 0
1 :1 2 J

000-:-:

8~~8~
t-- 'r, "<t'O0\-"<t

00 <n ~, \-<

2 .5. 6

000-:-:
- 0\ 0"0 0 0 -' 00000

q C/)
<n -:
I,

0\ 00 <n <n '0 "<t'0 0\ ~ <n

5 6
m=

1! 2 (USL + LSL)

000-00000

If, 1 Ii

I It

.5.

Ii

It

11

\.IJ

O\t--<n<'">'O0"<t00~ -<noo-<n

000-:-:
<n00-7t-00--t--

$cg:;t&3S:: -<noo-<n
000-:-:
"<tt--<n"<t"<t <n r-'-0 <'">'0'

'('I

!::

I'

00000

00--......

Table 2.6 Values of E[CpkJ for =m and d/(J=3 corresponding to Cp= Cpk= 1 for a series of increasing values of n

00000

Samplesize
00 '0 "<t('~0 <'">t---<n0\ <'">'00<'">'0

".J' I

00-:-:-:

00 'n r-r, - 00 <') t-- - <n 00 <')'00<'">'0

n 10 20 30 60 80 100 200 400 600 2200 3200 5400 10800 30500 79500
are

E[CpkJ 1.002 0.980 0.977 0.978 0.980 0.981 0.985

O.9W)

00-:-:-:

q C/)

00 <n <n '0 00 'O0\~<n00

00--00000

<n-O\O\O 'OO\-"<too

00--00000

t--<n<'">-O\

0 "<t 00 '"'~ n < <n00-<n00

00-:-:-:

'0 "<t r~ 0\ O"<too-<n 1f")00-<noo

t--

00-:-:-:
0\

-..; .
~
~ M

O\O~<noo

0---'0 0 ~,

'0

'0 '0 '0 t-- 0\ '0 0\ ("~ <n 00

00000
\.IJ

00--00000
-00'O~,-

000'O"<t~ "<tt---<n0\ '00\<'">'00\

00-:-:-:

"<t t-- - or, 0\ '00\<'">'00\

00-:-:-:

:c c::I
f-<

<1J

<nb <n
II

b ~. '0
'n \0 II"""
:;:

0.999 0.995 0.996 0.997 0.998 0.999 1.000

'

"""'"'I ,') "T

rl

,r, "T

'r. ";:;

:::

natural estimators of Cpkl and Cpku respectively. The

distributions of 3v/'l x Cpkl and 3vr,;- x CpkUare noncentral I with I degrees of freedom and noncentrality parameters J1t-(~-LSL)/O" and y'~(USL-~)/O" respectively. Chou and

64

The basic process capability indices Table 2.7 Variability of Cpk(Cp= 1) 1~-t(LSL+USL)I/O" Cpk . S.D. (Cpk)(n= 10) 0.0 0.5 1.00 0.83 0.28 0.27 1.0 0.67 0.23 0.095 1.5 0.50 0.19 0.08 2.0 0.33 0.155 0.07

Estimation of C pk

65

Tn(2.33),we see that, as n--HD,the term xd In becomes negligiblecompared with 3cX,,-I/(n -- ])J (which tcnds to 3c),
and the distribution of Cpk tends to that of a multiple of a

S.D.(Cpk)(n=40)

0.12 0.11

'reciprocal X' -- in fact, as n-+00 Cpk~ Cp and so has the

approximate distribution (2.13) with f = n --1. . I
,

Owen (1989) use Owen's (1965) formulae for the joint distribution of these two dependent noncentral t variables to derive the distribution of Cpk' The formula is too complicated to give here, but it is the basis for a computer program presented by Guirguis and Rodriguez (1992). This paper also contains graphs of the probability density function Cpk for selected values of the parameters, and sample sizes of 30 and 100. The following approach may be 9f interest, and provides some insight into the structure of the distribution of Cpk' The inequality
Cpk < C

I
, I

Figures 2.2--5 ~how the shape factors (J(31 and (32 -- see section 1.1) for the distribution of Cpk' In each figure, the dotted lines represent the locus of points (.J (31' (32) for with D= I~- m I/O"varying, or for fixed D (Figs 2.3 and 2.5) with d* =d/O" varying. Figures 2.2 and 2.3 are for sample size (n) equal to 30; Figs 2.4 and 2.5 are for sample size 100. The other lines on the figures represent loci for other standard distributions, to assist in assessing the shape

the distribution of

Cpk

for fixed d*=d/O"(Figs 2.2 and 2.4)

I
~

11;,

of the

Cpk

distribution

relative to well-established dis"- -- -

tributions.

is cquivalent to

1Or81

d-IX -ml <3cd i.e. IX-ml+3c8>d

Under the stated assumptions,
()

I I

~-,

Curve

30 ~amPI:-SiZe(n)

[I

l~~~=-- ~go'-.-J

(2.32)

;g:

ftiX

-ml is distributedas

XIO", n-l)t8 is distributed as Xn-IO"and these two statistics ( are mutually independent (see end of section 1.5). Hence
~

Pr[Cpk <c]

= Pr

3c + (n-1)jXn-1

d >-;;:J

(2.33)
~

0.0

0.5

1.0

1.5 X

2.0

2.5

3.0

ftXI

where Xl and Xn- 1 are mutually independent.

Fig. 2.1 Probability density functions of Cp for various sample sizes and. values of d/cr.

66
2.6 2.8 3.0

The /Josic fJl'Oces.'.; cOfJo/Jility indices Estimation

/,,"_._,,~ . d* ,".0.5','.

of' Cpk

67

2.6 2.6
'..d*=1 """"",
,\,- .0 --l

3.2

#.

I;

,,/

'

3.0
eJ' I!,
~

3.4

.~ 3.6 0

'-<./1' 0'/ Jf.

<l-'/
t.""

if

t::9

,/ .,1 , I

';(

~;

! /

"

'

3.21
.

~ 3.6
4.0 4.2 4.41
4.6j

,if /t

,;;",."'~\,

d* ; 2
.

'-<./ o' /%<:> ~!*7 c! ! '6. j

0'/ ;PI

RHN /f.

/ 07c I

III
-

~~d*=2.51

~ 3.61
:2 3.6

3.41

'\1 IC
ReflectedType III
Reflected Lognormal

J IC
Reflected Reflected Reflected llolloctod Type III Lognormal Type V Invorso Chi

4.0
4.2
4.4
.J

/
I ~

//1

ReflectedType V
.

Reflected Inverse Chi

I

. ---

~-~~~-~._-

-1.2 ,

-1.0

- 0.6

- 0.6

- 0.4 - 0.2 0.0 Skewness (0';;)

0.2

0.4

0.6

0.6

4.6., -1.2 :-1.0 - 0.6 - 0.6 - 0.4 - 0.2 0.0 Skewness (v(i1) 0.2 0.4 0.6 0.6

Fig. 2.2 Moment ratios of Cpk for n = 30 and levels of d* = d/~. Lab.elled points identify special distributions: N (normal), IC (inverse chi with df= 29), RIC (reflected inverse chi with df = 29),RHN (rcllcctcd half normal).

Fig. 2.3 Moment ratios of Cpk for n = 30 and levelsof c5 I~ - TI/a. = Labelled points identify special distributions: N (normal), IC (inverse chi with df=29), RIC (reflected inverse chi with df=99), RHN (reflected half normal).
.

Comparison of Figs 2.2-3 with Figs 2.4-5 shows a greater (N:::(0,3)) for the larger sample size (n = 100). Bissell (1990) uses the modified estimator

concentration of

(J131, f)2)

values near the normal point

and could simply use

C~k

= min(C~kl' C~kU)

(2.35)

where
~, C'pk-

X - LSL 38

USL:X 30' -

for ~~ LSL+" USL .

~

(2.34)
I

for ~~.LSL+ USL 2

CtkU= USL-c; 38

and

C* _~-LSL pkl--- 38

The distribution of C~k>as defined by.(2.34), is that of This differs from (ipk only in the use of ~ in place of X in the inequalities defining which expression is to be used. Note that C~k cannot be calculated unless ~ is known. However, if ~ is known, one would not need to estimate it
I

3.jn

x (noncentral t with f deg~es of freedom and noncentrality parameter ~{d-I~-1(LSL+ USL)I}/O')

(see section 1.7.)

68
2.6
2.8 3.0 3.2 -;;; 3.4 .3
'ijj 3.6 '0'

The basic process capability indices Estimation of Cpk

..Q /0* '" ;

69

'~

'--"'.

'"

d *105 .

2.6
d*~1 d*~1.5 d* ~ 2 . .: d* 2.5

l/'
.;p /
cP".

//

,.\~

2.8
S.O

-R/ 1.(1
I

/ ., /" ' ., /ftiC

'

.'//,::0

//:::="-,

--'-"

-~.~~;:'o'-~--:~~:--~---~--l ,,/ '- 01

' '

r
I

I I

".

I! / //
:JJ.~/ b <l! ,~
"

IC
~

1.2j
.3

",

f! :J

::.::

3.8 4.0 4.2 4.4 4.6"

RN

/ 1( d
cY/

f/
/

CJ::

,/I
,

//

-R/,'/ . / ,,;//'.,.,' , ",

V
.'

I'

U-,

"

u~O,25

016.
,

/ .. ..

. .,,-.-

(\~ 0.4

/'O":\)'
.,/

,!?
IC

" I
I

if

,/1 /f

/1 /{
Reflected Reflected Reflected , Reflected
'

.!!J 3.4 1 . (/) 3 ' 6

0 'I::

"'CJ:: ,"

.il

Y'9j/
"
,/,'

"

IC /j

. ,.,/

.',il 11 .

,I

~
Type III Lognormal Type V Inverse Chi

3.8
4.01

/f /(
/ (,

1.
,

. ..,./f ,,{ '~ ,/~'. . .' ,if'

,

'-'-' -"'-

" /1'

4,2 t

'

,!

'"

r
0.2 0.4 0.6 .
=:0

4.41 4.61, -1.2

,iI 1/ /11

';I

Reflected Type III .-. -. Reflected Log normal --- ReflectedType V

-1.2

-1.0

-0.8

-0.6 ,-0.4 -0.2 0.0 Skewness (v%)

of Cpk for /I =

0.8

,._.~_-,-~.~~eflect~~~~..L~--~-~---e -1.0 -0.8 -0.6 -0.4 -0.2 0.0 (v:61) 0.2 0.4 0.6 0.8 Skewness

Fig. 2.4 M omel1t

ratio~

100 and levels or d~ d/(f.

I

Labelled points identify special distributions: N (normal), IC (inverse chi with df=99), RIC (reflected inverse chi with df=99), RI-IN (reflected half normal).

Although C~k is not of practical use, it will not differ greatly from Cpb except when ~ is close to ~(LSL + USL). This is because, the probability that X --!-(LSL + USL) has the same sign as ~-!(LSL+USL) is

Fig. 2.5 Moment ratios of Cpk for n = 100 and levels of ()= I~ - Tla. Labelled points identify special distributions: N (normal), IC (inverse chi with df=99), RIC (reflected inverse chi with df=99), RHN (reflected half normalj.

1-W( - ftl ~--!-(~SL+USL~l)

z(CSL + USL). As the distribution (noncentnil t) of C;,k is a well-established OIlC,il call bc lIscd as all approxImalioll 10 ihal or ('Pk lIIHkr suitable conditions-- namely that ~ differs from !(LSL+ USL)
' , '

and this will be close to I

1 for n large, unless ~=

difficult, because two parameters (~,0') are involved. It is tempting to obtain separate confidence intervals for the two paralllctcrs, and thcn construct a confidcnce interval for Cpk' with limits equal to the least and gretest possible values for Cpk corresponding to pairs of values (~,0') within the rectangular region defined by the separate confidence intervals. Although this may give a useful indication of the accuracy of estimation of Cpk it is necessary to remember the following two points. (a) If eaeh of the separate intervals has confidence coefficient 100(1- iX)%the confidence coefficient of the derived interval for Cpk will not in general be 100(1-iX)%. For . example, if the statistics used in calculating the separate intervals were mutually independent, the probability that each will include its appropriate parameter value would

by a substantial amount, and n is not too small. Construction of a confidemce interval for Cpk is more

70

The basic process capability indices l'

Estimation

of Cpk

71

not be 100(1-0:)%, but 100(1-0:)2%. This would mean, for example, with 0:= 0.05, that we would have 90.25%, rather than 95(Yoconfidence coelTIcient. Generally, the combined probability will be less than the confidence coefficient of each component confidence interval. (b) Another effect, arising from the fact that only part of the rectangular region contributes to the interval for Cpk, and that other pairs of values (~,a) could give Cpk values in the same interval, tends to counterbalance effect (a).
In the case of Cpb the 100(1-0:)(1<) separate intervals for

t
I.

ChOll, Owen and Borrego (1990) provided tables of 'apc proximate 95% lower confidence limits' for Cpk (inter alia) under the assumption of normality. A few of their values are shown below in Table 2.8.
Table 2.8 Approximate fidence limits for Cpk
Cpk

95% lower conn=75 0.83 0.91 1.26 1.40

,/

"

n=30 0.72 0.80 1.12 1.25

n=50 0.79 0.87 1.21 1.35

,/

and a are

x-

tr. I - al 2 a ~ ~~ X + t f. I - ~I2 a
XI,I-aI2

(2.36a)
,

1.0 1.1 1.5 1.667

t
II

~~a~~lL

(2.36b)
X.r.rJ.12

Although

X and a are mutually independent, the events defined by (2.36a) and (2.36b) are not independent. However, it is still true that the probability that the interval for Cpk, calculated as described above, really includes the true value will be less than 100(1- 0:)%. The extent of the counterbalancing effect of (b) is difficult to ascertain, but it is not likely

Franklin and Wasserman (1992) carried Qut simulations to assess the properties, of these limits, and discovered that they are conservative. They found that the actual coverage for the limits in Table 2.8, is about 96-7%. Guirguis and Rodriguez (1992, p. 238) explain why these approximate limits are conservative. On the other hand, Franklin and Wasserman's studies showed that the formula
~2

to be comparable to the effectof (a).

.

. I
CPk-ZI-a(9n

'

Zhang et al. (1990),Kushler and Hurley (1992) and Nagata arid'Nilgahilf'til (199~) provide a thorough treatment of construction or conlidenel.: intervals 1'01' pk which will bl.: relaC ' . tively simple to calculate. Heavlin (1988) in the report already referred to, suggests 6 L! n 1 2 1 { I Cpl,-ZI-rJ.12 9n(n-3)+CPk2(n-3)
~ n-1 ~2 1

Cpk
+2(n-'-l)

)
2

(2.38) (cf(2,37))

1+ n-1
6

)I'

'
2

Cpk+Zl-rJ.12 { 9n(n-3)+CPk2(n-3)

(1+ n-1 )} )

(2.37)

as a 100(1-0:)% confidence interval for Cpk'

produces (for n ~ 30) remarkably accurate lower 100(1- 0:)% confidence limits for Cpk' Nagata and Nagahata (1993) suggest modifying (2.38) by adding 1/(30';;;). They obtain very good results from simulation experiments for the two-sided confidence intervals with this modification. The parameter values employed were C"k=0.4(0.3)2.5; n= 10(10)50, 100; 1-0:=0.90, 0.95. The actual coverage was never greater than 0.902 for 1- 0:= 0.90, 0.954 for 1- 0:= 0.95; and never less than 1- 0:.

72

The basic process capability indices

Kushler and

11urley

(Il)l) I) suggest the simpler formula

Cpk

1-Z1-cx

(2n

- 2)1 J

(2.39)

.u

and Dovich (1992) reports th,at the corresponding approximate 100(1-0:)% formula for confidence jntervallimits:-

Cpk

[ (2n -

1 +Z1-CX/,2
2)2

(2.40)

""I

gives good results. *The interested reader may find the following brief discussion illuminating. Difficulties in the construction of confidence intervals for Cpk arise from the rather complicated way in which the paramctcrs ~ and rr appcar jn thc cxpression for Cpk' We first consider the much simpler problem arising when the value of (1is known. We then calculate C* - min(X -LSL, USL-X)
p,k3(1

.><:

--!:: ~ 0 !::

~
"'I
... '" \,.)

....

::0

~I

--~
0

c; ;;.
.... Q)

..8 en

~
I

.S
I

I c;i I ::>I I I;:J l;g E l,j I~ Iii I '""I I I I I I I I I I I

as our estimate of Cpk' The 100(1-0:)(% confidence interval for ~,

I I I I I I I

8 !::
"0

Q)

<.j:1 !:: 0 U \C N

~ 1;

f(X)~ ~~ [(X)
with f(X) = X- ZI-a/2(1/ Ii, j [(X) = X + Z1-a/2(1/jIi

(2.41)
can be

~I

used as a b,asisfor constructing a confidenceregion for

"
.

I I I

C k = min( - LS L,~~~!:.::~~) ~
p 3(1

(2.42)

,JI'

J
'.-'

,I

J ::

:;;

-"-

I -":;;

The confidence region is obtained by replacing the numerator by the range of values it can take for (2.41).A little care is

31

74

The basic process capability indices

Comparison of capabilities

75

needed; this range is not, in general, just the interval with end points

particular, if process variation is normal, the index ..

Cpl =! 3 ~ ,- CT LSL --

min(~(X)- LSL,USL- ~(X));min(;;(X)-LSL, USL- ;;(X)). For example,if LSL<~(.X)<!(LSL+USL) and ;;(XUSL
(see Fig. 2.6 c) the.n the-confidence interval for Cpkis
1 ~-(USL-s(X)),
30"
..

..

11 -'z(USL-LSL)=30"

d
30"

(2.43)

I Some possibilities are shown in' Table 2.9 (with 111 =

!(LSL + USL)). (See also Figs. 2.6 a-d.) , When 0"is not known, but is estimated by S, the problem is more complicated. If we replace the values ~(X), ;;(X) by

I I

x - tll-1.al 2S and X + tll-l .l-a

l 2S

fi

fi

is directly linked to the expected proportion of output which is below the LSL, by the formula <1>( 3Cpd.Comparisonof estimators Cplt, Cpl2 of Cpl values (Cplt, Cp12)for two processes is therefore equivalent to comparing proportions of output for the two processes with values falling below their respective LSLs. With sufficient data, of course, this could be effected by the simple procedure of comparing observed proportions of values falling below these limits. However, provided the assumption of normality is valid, the use of Cpll, Cpl2might be expected to give a more powerful test procedure. Chou and Owen (1991)have developed such tests, which can
be regarded as a test of the hypothesis Cplt = Cpl2 against either two-sided (Cplt =CpI2) or one-sided (Cplt <CpI2, or Cplt > Cp12) alternative hypotheses, for the case when the
.

((2.36a) with f = n -1) we would still get 100(1- ex) confid% ence limits from Table 2.9, but computation of these limits, would need the correct value of CT evaluate the multiplier to 1/(30". )
Table 2.9 Confidence interval for Cpk(O' nown) k Valueof f(X') ~(X)
:;;;'m :;;;'m

sample sizes from which the estimators Cpl1 , Cpl2 are calculated a're the same - n, say. In an obvious notation the statistics variaqles with (n-1) degrees of freedom, and noncentrality
Tt

= 3fiCPlt, 1'2 3JnCp12 are independent noncentral t =

Confidence limits (x 30')

~ I

pnrnnleters JIi(~1 LSL1)/(fI,jn(~r-LSL2)/(f2 respectively. Applying a generalized likelihood ratio test technique (section 1.12),Chou and Owen obtain the test statistic

:;;;'m (~(X)-LSL,f(X)-LSL)

>m

>m (min(~(g)-LSL, USL-~(X)), t(USL-LSL)) >m (USL:'~(X), USL-f(X))

u = [{T'y + 2(n-l)
..

}{1'~+ 2(n -I)} J1- Tt 1'2

(2.44)

Bayesian estimation of Cpkhas been studied by Bian and Saw (1993). 2.6 COM PARlSON OF CAPABILITIES

To establish approximate significance limits for U they carried out simulation experiments as a result of which they suggest the approximdtion, for the distribution of U, when Ho is valid: .'all IQg {!U/(n - I)} approximately distributed as chi-squared
with
VII

PCls are measures of 'capability' of a process. They may also be used to compare capabilities of two (or more) processes. In

degrees offreedom.' Table 2.10 gives values of alland Vi"

estimated by Chotl and Oweil from their simulations.

76

The basic rwocess capability indices

I

Correlated observations

77

Table 2.10 Values of (/11 ami VII

I
I

n 5 10 15 20

all

VII

8.98 19.05 26.47 36.58

1.28 1.22 1.07 1.06

however, that in this case we cannot rely on independence of J( and S. Correlation can arise naturally when 0"2 is estimated from data classified into several subgroups. Suppose that we have k subgrou.ps, each containing m items and denote by Xuv the value of vth item in the uth subgroup. If the model XUV=Yu+Zuv u= 1, .." k; v= 1, ..., m (2.45)

The progression of values is rather irregular, possibly

reflecting random variation from the simulations (though. there were 10000 samples in each'case).[If the valucs of (IIO and a15 were about 18.2 and 27.3 respectively, progression would be smooth.]' .
2.7 CORRELATED OBSERVATIONS

applies with all Ys and Z mutually independent random variables, each with expected value zero, and with
var(Yu)=O"~ var(Zuv)=O"~

The above theory based on mutually independent ,observations of X can be extended to cases in which the observations of the process characteristic are correlated, using results presented by Yang and Hancock (1990).They show that E[S2]=(1-p)0"2

thenw~ have n = km,

0"2=0"~+(1i

and correlation between

. . .

o
{ (O"J10"~)

if u#u'
if u=u' (and v 7"" vi)

wherep = averageof the 1n(n-1) correlationsbetweenpairs of Xs. '(SeeAppendix 2.A.) This implies that in the estimator
so that 1d

XII" :\/1<J XII"" =

CP=:3S

P
of Cp,the denominator willtend to underestimate 0" and so
C\,
~

) n(n- 1)km(m-1)- (O"y+O"z km(m-1) O"~

0"2 y

..

will tend to overestimate CpoThis reinforces the effect of the bias in l/S as an estimate of 1/0",which also tends to produce . overestimation of Cp (see Table 2.3). There will be a similar effect for Cpk' We warn the reader,

n(n-1)~0"2

m-1 <1~
=n-10"2

rooI

78

The hasic process capahility indices

l

Appendix 2.A
:I

79

In the case of equal correlation we have pij = j5 for all

. i,j(i=j).ln this case S2 is distributed as (1- p)X;-la2j(n-I).

I

'shrinking'

R towards

the value 1.0, to produce a PCI estimator

(,
2.8 PCTs FOR ATTRIBUTES

- Ii I~ I

(I

k)

(2.48)

An interesting proposal for PCI for attributes (characteristics : taking values 0 and 1 only - e.g., conforming (C) or (NC)), was made by Marcucci and Beazley (1988). This has not as yet received attention in the literat.ure. We suppose that an upper limit w, for proportion of NC product is specified. It is proposed ,to use the odds ratio R=w(l-wd (I-W)Wl (2.46)

t
"

for some k. (The subscript 'a' stands for 'attribute.') They try to choose k to minimize the mean square error E[(C~--R)2J, and suggest taking

I
where
! I
aJ?!

1-1P k=aj+l-R2

(2.49)

= R2(X

+ ~)- t +(11--X + ~)- t].

(see, e.g., Fleiss (1973, pp. 61 et seq.) or Elandt-Johnsonand Johnson (1980, p. 37)) as a PCI, where W is the actual proportion of NC product. The ratio w(l- (I)) 1 measures the odds for the characteristic being NC, and w 1(1- w d - 1 isthe which is just acceptable, while R = 2 reflects an unfortunate situation, in which the odds for getting a NC item are twice the maximum acceptable odds. Unlike Cp, Cpk' etc., large values of this, PCI indicate bad situations. Marcucci and Beazley (1988) study the properties ,of es- . timators of R. Given a random sample of size /1,with X items . found to be NC, the estimator
rhaximally acceptable odds. Thus R

III

APPENDIX 2.A

= 1 corrcsponds

to product

r
r
I'

If X 1~X 2"", Xn are independent normal variables with common expected value

~ and

standard

deviation a then

l-/J= Pr[LSL<X, <USL] ~",(.tJs~-:- ~)_",(LS~ - ~)

(2.50)I ,
A natural estimator of p (the expected proportion of NC items) is
A

RA

X +! - WI 1n-X +1 WI

'(2.47)

p= 1-<1> I II
..

USL-X
S

) (
.

LSL-X

+<I>-S--

(2.51)

is well-known in' statistical practice. The !8 arc so-callcd 'Sheppard corrections' for continuity.'Unfortunately, the variance of R is large, and so is the (positive) bias, especially in small

samples. Marcucci and Beazley (1988), therefore, suggest

This estimator is biased. A minimum variance unbiased estimator of p is derived below. It is quite compIipated in form and. for most occasions (2.51) will be more convenient, and also quite adequate.

80

The basic process capability indices

.1

Appendix 2.B II.

81

We will apply the Blackwell-Rao theorem (section 1.12.4). The statistics (X, S) are sufficient for ( and (j, and (1- Y), where

where Io(a,b) denotes incomplete beta function ratio (section 1.6) and

y=

1 if LSL<X1 <USL { 0 otherwise

.,j

8,(X,SJ=KI + US~-X {(n~l)' +(US~-X)rJ e'(X,SJ=Kl~ LS~-X {(n~l),+(LS~-X)'} -'J

is an unbiased estimator of p. Hence the conditional expected value of 1 - Y, given (X, S), will be a minimum variance

unbiased estimator of p.

.
II,

The conditional distribution of X I' given X and S, is derived from the fact that

Calculation of p is tables of the function quite heavy, but can be facilitated by

J(n, c)= I![I+C{~+cZ)-jlt(n -1), :t(n-1)) for various values of nand c. Then (2.53 a)

Xl-X
S

( )
n-1
~II tll

is distributed as 1/1-1 (l with (n-l) degrees of freedom) (section 1.7). Hence the conditional distribution of X I' given X and S, is that of
-

X + -;;-

( )
[

n-1

p= I-J(n, ~S~-X)+J(n,~SI~-X)
St/l-I
I
.

(2.53 b)

(In this expression, X and S are to be regarded as constants.) Hence

~I

tion.

It must be emphasized that the appropriateness of both of the estimators fi and p depends on the correctness of the assumption of normality for the process distribu-

p=E[pIX, SJ = 1-pr <t/l-1<

USL-X
S

LSL-X
S. l

( )
n-1

APPENDIX 2.B

( )J
.

.
,

Xl, X2, '" , XII

are dependent random variableswith common

f/- J

,
II

expected value ( and variance (j2; the correlation between Xi and )fj wi1lbe denoted by Pij. We have
.

= 1-Iodx,s)(!(n-1), !(n-1)) +Ioz(x,s)(!(n-1), !(n-l))

(2.52)
11\

var(X)=n-2(j2

n+n(n-1)

L Pij }=-{1+(n-l)p}(j2 ii'j n

.

82
where

The basic process capability indices

Bibliography BIBLIOGRAPHY

83

'L LPij - i~j P=n(n-1) If

/I 5 =(n-1) 2 -1" L., (Xj-X)j= 1 -~ .

then E[52J=

~ 1 j=
1

E[(Xj-X)2]
2

=-

n -1

j= 1

~ L.,E[(Xj-~)
j=l

]-2(X-~)(Xj-~)+(X-~)]

-;

I
2
"

=~

n-1

lt

E[(Xj-~)2]-nE(X-~)2

since ~(X j - ~)= n(X -~) so

E[52]=~- 1 {n var(Xj)-/1 var(X)} /11 =-[n- {l +(n-1)p} ](j2 n-1 =(1-,o)(j2 We emphasize that (2.54)
,III ,

I
1I(n-1)
i~j

fJ='--~--- L LPiJ
is the average of the correlation coefficients among the !n(n-1) differentpairs of Xs,

t
.111

Uarnctt, N.S. (Ing) Process control (/nd product quulity: The.Cp und Cpk revisited. Footscray Inst. Techno!., Victoria, Australia. Bian, G. and Saw, S.L.e. (1993) Bayes estimates for two functions of the normal mean and variance with a quality control application; . Tpch. Rep., National University of Singapore. Bissell, A.F. (1990) How reliable is your capability index? Appl. Statist., 39, 331-40. Bunks, J. (1989) Principles of Quality Control, Wiley: New York. Carr, W.E. (1991) A new process capability index: parts per million, Quality Progress, 24, (2), 152. Chen, L.K., Xiong, Z. and Zhang, D. (1990) On asymptotic distributions of some process capability indices, Commun. Statist. TheO/'. Meth., 19, 11-18. Chou, Y.M. and Owen, D.E. (1989) On the distribution of the estimated process capability indices, Commun. Statist. - TheaI'. Meth., 18,4549-60. Chou, Y.M. and Owen, D.B. (1991) A likelihood ratio test for the equality of proportions of two normal populations, Commun. Stat'ist. - TheO/'.Alt!th., 20, 2357-2374. Chou, Y.M., Owen, D.B. and Borrego, A.S.A. (1990) Lower confidence limits on process capability indices, ./. Qual. Technol, 22, 223-9. Clements, R.R. (1988) Statistical Process Control, Kriger: Malabar, Florida. Colcman, D.E. (1991) Relationships between Loss and Capability Indices, Api. Math. Camp. Techn., ALCOA Techn. Center, PA. Constable, O.K. and Hobbs, J.R. (1992) Small Samples and Non~ normal Capability, Trans. ASQC Quality Congress, 1-7. Dovich, R.A. (1992) Private communications. Elandt-Johnson, R.C. and Johnson, N.L. (1980) Sul't,ivalModels and Data Analysis, Wiley: New York. Fleiss, J.L. (1973) Statistical Methods for Rates and Proportions, . . Wiley: New York. Franklin, L.A. and Wasserm~n, 0.5. (1992). A note on the conservative nature of the tables of lower confidence limits for Cpkwith a suggestedcorrection, Commun.Statist. Simul. Compo Gensidy, A. (1985) Cp and Cpk, Qual. Progress, 18(4),7-8.

84

The basic process capability indices

I
,."

Bihliography

85

Grant, E.L. and Leavenworth, R.S. (1988) Statistical Quality Control (6th edn.), McGraw-Hill, New York. . Guirguis, G. and Rodriguez, R.N. (1992) Computation of Owen's Q function applied to process capability analysis, J. Qual. Technol., 24, 236-246. Heavlin, W.D. (1988) Statistical properties of capability iadices, Technical Report No. 320, Tech. Library, Advanced Micro Devices, Inc., Sunnyvale, California. . Herman, J.T. (1989) Capability index - enough for process industries? Trans. ASQC Congress, Toronto, 670--5. John, P.W.M. (1990) Statistical Methods in Engineering and Quality Assurance, Wiley: New York. Johnson, M. (1992) Statistics simplified. Qual. Progress, 25(1), 10--11. . Kane, V.E. (1986) Process capability indices, J. Qual. Tech/1ol., 18, 41-52. '. Kirmani, S.N.U.A., Kocherlakota, K. and Kocherlakota, S. (1991) Estimation of if and the process capability index based on
subsamples, Cammun. Statilit.

..,

Nagata, Y. and Nagahata, H. (1993) Approximation formulas for the confidenceintervals of processcapability indices, Tech. Rep. Okayama University, .lapan. (submitted for publication). . Owen, D.H. (1965) A special ease of a bivariate noncentral tdistribution, Biometrika, 53, 437-46. Owen. M. Berlin, Germany. and Continuous Improvement, Springer Verlag, (cd.) (1989) SPC Pearn, W.L., Kotz, S. and Johnson, N.L. (1992) Distributional and, inferential24, 216-231. of process capability indices, J. Qual. Technol., properties Porter, LJ. and Oakland, 1.8. (1990) Measuring process capability nat., 6,indices -- some new considerations, Qual. ReI. Eng. Interusing 19-27.

~I

~
tll

TheaI'. Meth., 20, 275---291 (Cor-

~II

rection, 20, 4083). Kotz, S., Pearn, W.L. and Johnson, N.L. (1992) Some process capability indices are more reliable than one might think, Appl. Statist., 42, 55-62. Kushler, R. and Hurley, P. (1992) Confidence bounds for capability indices, J. Qual. Technol., 24, 188-195. Lam, C.T. and Littig, S.J. (1992) A new standard in process capability measurement. Tech. Rep. 92-93, Dept. Industrial and Operations Engineering, University of Michigan, Ann Arbor. Leone, F.C., Nelson, L.S., and Nottingham, R.R. (1961) The folded normal distribution, Technometrics, 3, 543-50. Li, H., Owen, D.B. and Borrego, A.S.A. (1990) Lower confidence limits op process capability indices based on the range, Commun.
Statist.

~. tl ~\

Wadsworth, H.M., Stephens, K.S. and Godfrey, A.B. (1988) Modern Methods for Quality Control and Improvement,. Wiley: New York. Wierda, S.J. (1992) A multivariate process capability index, Proc. . 9th Internat. Conj. Israel Soc. Qual. Assur. (Y. Bester, H. Horowitz and A. Lewis, eds.) pp. 517-522.[ Yang. K. and Hancock, W.M. (1990) Stntistical quality c'ltrol for correlated samples, inter". J. Product. Res., 28, 595-608.1 Zhang, N.F., Stenback, G.A. and Wardrop, D.M. (1990) ,~nterval estimation of process capability index Cpk' Commun. S,tatist. Theor. Meth.. 19,4455-4470. \ I I (

/" ~'I

! !

Simul. Camp., 19, 1-24.

~~

Marcucci, M.O. and Beazley, cc. (1988) Capability indices: Process performance measures, Trans. ASQC Conwess, 516 21 Montgomery, D.C. (J985) Introductio/1 to Statistical Quality Control, Wiley:New York.

.<

t'
~,

Nagata, Y. (1991) Interval estimation for the process capability indices,J. Japan. Soc. Qual. Control,21, 109-114.

(
.. -'

3 T.he Cpmindex and related

indices

3.1 INTRODUCTION

'I'
p'
dl ~I
l~ t~

,. ~I

,
I. ..
1'1 W

I \
1

The Cpm index of process capability was introduced in the published literature by Chan et al. (1988 a). An elementary discussion of the index, written for practitioners, was presented by Spiring (1991 a), who also (Spiring, 1991 b) extended the application of the index to measure process capability in the presence of 'assignable causes'. He had previously (Spiring, 1989) described the application of Cpmto a toolwear problem, as an example of a process in which there was variation due to assignable causes. R(;ct;nlpapers (lInpllhlished at the time of writing) in which thl: usc of Cpmis discussed indude Subbaiah and Taam (I SI<) I) and Kushler and Hurley (1992).(All four of these authors were in the Department of Mathematical Sciences, Oakland University, Rochester, Minnesota.) These papers study estimation of Cpm'in some detail; graphical techniques for estimating Cpmare discussed briefly by Chan et al. (1988b). The use of the Cpm index was proposed by Hsiang and Taguchi (1985) at the American Statistical Association's Annual Meeting in Las Vegas, Nevada (though they did not use the symbol Cprn)'Their approach was motivated exclusively by loss function considerations, but Chan et al. (1988a, b) .w~re more concerned with comparisons with Cpb discussed
87

:88

The Cprn index and related indices

The Cpmindex

89

in Chapter 2. Marcucci and Beazley (1988), apparently independently, but quite possibly influenced by Hsiang and Taguchi, proposed the equivalent index, also defined explicitly by Chan et al. (1988 a) Cpg=C;,; (proportional to average quadratic loss). Johnson (1991) see section 3.2 - also reaches 'Le', a multiple of C;rn2from loss - or, conversely, 'worth' - funct.ion considerations.. Chan et al. (1988 b) defirie a slightly dif1crent index, C~I;1 (see (3.2) below), which is more closely related, in spirit and form, to Cpk than is Cprn' Finally an index Cpmb combining features of Cpmand Cpk is introduced, and studied in section 3.5.

,
.

,tin.
'111

~'I!.

If T=m (as is often the case), then C;m=Cpm' If T=I=m, ie d same type of pitfalls beset uncritical use of C;m as with Cpm' The denominator in each,of (3.1) and (3.2)
3{a2 +(~- T)2}f= 3{E[(X - T)2]}!

(3:3),

~
"
~IIII

that is, three times the root-mean-square deviation of X from the target value, T. It is clear that, since
(12+ (~ - T)2 ~ (12

II

""

"'"

we :tpust have
I

, ,k

Cpm~ Cp

. (3.4)1

.
3.2 THE CPIllINDEX

The relationship between Cpmand Cpk is discussed below; see (110) et seq. In the approach of Hsiang and Taguchi (1985) it is supposed ,that the loss function is proportional to (X - 1')2, and so (12+ (~- 1')2 is a measure of average loss. From this point of view, there is no need to include USL or LSL in the' numerator, or, indeed, the multiplier 6 in the denominator, of the formula for Cpm,except to maintain comparability with the earlier index Cpo which was introduced with quite a different background, although for the same purpose. Proponents of Hsiang and Taguchi's approach point out that "it does allow for differing costs arising from different values of X in the range LSL to USL, though in a necessarily arbitrary way. (A discussion between Mirabella (1991) and Spiring (1991 b) is relevant to this point.) It is suggested that a re,asonable loss function might be constant outside this range, i perhaps giving a function of form

The original definition of Cprn is

,lJSL-LSL d Cpm= 6{(,.2+(~-T)2}f = 3{a2+(~-T)2}!

(3.1)

where ~ is the expected value and a is the standard deviation of the measured characteristic, X, l' is the target value and d = t(USL - LSL). The target value is commonly the midpoint of the specification' interval, m = 1(USL + LSL). This need not be the case, but if it is not so, there ,can be serious pitfalls in the uncritical use of Cpm, as we shall see later. The modified index, C~rn'is defined using (J. J4 e) by

(
"

, ..,

C(X- T)2
* _min(USL-T,T-LSL)d-IT-ml ' Ii I

for LSL~ X ~ USL

for X ~ LSL

\
I

c(LSL -

1')2

Cpm-

3{a2+(~-T)2}f

- 3{a2+(~-T)2}f

(3.2)

{c(USL-.T)2

. (3.5)

for X ~ USL

90

The Cpm index and related indices The Cpm index

/I .,f
r: I
!

One is, of course, free to use one's intuition ad libitum in construction of a loss function. However,if this approach is accepted, there seems to be little justification, in respect of 'practicalvalue,ease of calculation,or simplicityofmathematical theory in using the cumbersomeCpm C;m)form rather (or
than the loss function itself (as in, for example, Cpgor Le - see 'sections 3.1 and 3.4). .", As already remarked, we take the position that expected proportion of NC product is of primary importance, though we acknowledge that other factors can be of importance in specific cases. Essentially, thc usc of any singlc'indcx cannot be entirely satisfactory, though it can be of value in rapid, informal communication or preliminary, cursory assessment of a situation. The following interrelations should be noted. We have
'

91

(:~l(UsL+LSL)=m) so that d=t(65--35)=15, and con- , siders three processes with the following characteristics: I

t
I
I

~= 50(= T); =5 B: ~= 57.5; a = 2.5

A:
(j

C: ~= 61.25;

(j

= 1.25

t:

The values of the indices Cp, Cpk and Cpm are

I
I

I
, ,
Jill

Process A B C

Cp 1 2 4

Cpk 1 1 1

Cpm 1 0.63 0.44

I d Cpm= 3"a'

(ctm = Cpm, since T= m). (3.6) Note that, in this example, C" and Cpm move in opposite directions. According to Cp, the processes are, in order of desirability - C:B:A - while for CPOI the order is reversed. According to the index Cpkothere is nothing to choose among the processes. At this point we repeat our warning that CPOl'as defined in (3.1), is not a satisfactory m~asure of process capability unless
t/w larget value is equal to the midpoint of the specification interval (i.e. T:'::1II=1(LJSL+LSL)). The reason for this is that, whether the expected value of X is T - (5or T + (5,we obtain the same value of CpnlO namely

wherea'= {E[(X T)2J}! measures the 'total variability' about the target value. Recalling the definition (in Chapter 2)
C
p

, I,
r
t
J'I

= USL

- LSL = ! ~ 6a 3a

(3.7)

we have
(i Cpm= ,Cp=Cp(1 (J

1
+,2)-!

(3.8)

3 d(a2 + (52) - J,

(3.9) ,

where'
i

= (~- T)/a. So Cpm~ Cp with equality if~

= T.

"
..

i I
I

Boyles (1991) provides an instructive comparison be-, tween Cp and Cpm'He takes USL=65, LSL=35, T=50

although, unless T = m, these processes can have markedly different expected proportions of NC items (items with X outside specification limits). For example, if T=iuSL+!LSL

,,'

I I

92 and

The Cpm index and related indices

It

The Cpm index

93

I.e. as

c5=.t(USL-LSL)=!d
then with E[X] = T + D= USL, we would expect at least 50% NC items if X has a symmetrical distribution, while when E[X] = T - (j = m, this proportion would, in general, be much less. In the light of this, it is, in our opinion, pointless to discuss properties of Cpm (as defined in 0.1)) except when T=m. These comments also apply to C~m' Generally, it is not possible to calculate Cpmfrom Cpk,or conversely, without knowledge of values of the ratios d:0": (- mI. I If T= m then from (3.1), and (2.20b) I
~

(1-I(~ml){1+(~~7r;,
If 1n'=T, then (3.12) becomes

or

:::;; 1

(3.12)

(l-U){l +(~y U2 ~ r
(where

or

:::;; 1

\
. ~!,

u= I~-ml d I.e.

= I(~- TI .
d'

~:: ={I_I(~ml} x [1+{(~mrJ

If, further Cp = 1 (Le. d = 30")then

(3.10)

1+
(3.11)

()
-;;

U2~

or

~(1_-U)-2

9(1-CPk)2=((~mr =C;~-l
Since (d. (2.20 c)) C
p k 2

(~)
0"
1

or

~(1-u)-2-1
U2

= l-(l-u)~=
u2(1-u)2

2-U
u(1-u)2
,

(3.13)

=
,

I~-ml C
d

C
p pm

1+

i!I_T -=--)
G'.

-! }

C
P

Usually u< 1 (or else ~ is outside specification limits). Inequalities (3.13) can be written 2-u

we have
CI' ~ max (Cpk>C,",,) Also, Cpk ~ or :::;; Cpm according I~-ml 1 - ---yas

C; ~
,

or
'

:::;; 9u(1-u)2

Here are some values of this function.

or

:::;; 1+ --;-{

~-

)}

-!

u -- 2-u 9u(1-u)2

0 co

0.2 1.56

0.4 1.23

0.5 1.33

0.6 1.62

0.8 4.17

1
CIJ

94
This function

Th('

lP1l1 il/d("\" 1/111//"1'11/(('(/ indic('s

has a minimum

valuc

of 1.23 at /I = OJX

(appro x.). If Cp is unsatisfactory (Cp~ 1) then we always have Cpk~ Cpm (if T= m), but we can have Cpk> Cpm when C~> 1.23(Cp> 1.11), provided u is not too close to 0 or 1. If u=O (i.e. ~=1'=1J1.)then Cp=Cpk=Cpm' . Johnson (1991) also approached the Cpm index from the point of view of loss functions, using the converse concept of 'worth'. He supposed that the maximum worth, W1', of an item corresponds to the characteristic X having the target value, T. As the deviation of X from T increases, the worth becomes less, . eventually becoming zero, and then negative. If the worth function is of the simple form
WT-k(X ~ 1')2 for W1'~k(X - 1')2

It

IVI;

IS If

Comparison of Cpm and C~m with Cp and Cpk Recalling that

95

IIiCII

C;m

= (~Y {E[X

-- 1'fJ}-1

.'he Ipl
11'".
1110

we see that,

",.. lid Illu

Le=(3~)2C~ Hence Le is effectivelyequivalent to Cpm,although it has a differentbackground. A natural unbiased estimator of Le is

Le
~

t
~I) ~ ,r ,

1 = n~ -2

L. (Xii=1

1')

it becomes zero when IX-1'I=JWT/k. In Johnson's (1991) approach the values 1':f:.JWT/k play similar roles to those of the specification limits in defining Cpm and we may define ~=.J~';k. . The ratio (worth)/(maximum worth)

The distribution

of n~ 2Ie is therefore that of

2 '2
(J

,'I~
II
I~ II
.1

Xn

n(~-- 1')2
(J 2

= 1- k(X - 1')2
Wr =1-- (X - 1')2 ~2

discussed in section 1.5. The resulting analysis for construction of confidence intervals follows lines to be developed in scction direct. 3.5. Extcnsion to nonsymmetric loss functions is

'is termed rhe relative worth,and (X -

1')2 2 /~

is termed the

relative loss. The expected relative loss is then (in Johnson's notation) E[(X

3.3 COMPARISON OF CpmAND qm WITH Cp AND Cpk We first note that if 1'=~, then Cptn is the same as CpoIf, also, ~= rn,then both Cpm and C~mare the SaIne as Cpk' If, further, the process characteristic has a 'stable normal distribution' . thcn a vaillc of ! (for all four indiccs) mcans that thc expected proportion of product with values of X within specification

~2

- 1')2J - L e

96

The Cpm index and related indices

~llr
11111 I'll

limits is a little over 99.7%. Negative values are impossible for Cp and Cpm'They are impossible for C;m if T falls within specification limits (and for Cpk if ~ falls within specification limits). . The principal distinction between Cpm and Cpk is, as l'I1irabella (1991) implicitly indicates, and Kushler and Hurley (1992) state explicitly, in the relative importance of the specification limits (USL and LSL)' versus the target value, T. As explained in Chapter 2, the main function of Cpk is to indicate the degree to which the process is within specification limits. For LSL< ~< USL, Cpk---+ (X) as 0"---+0, but large values of Cpk do not provide information about discrepancy between

Comparison
If Cp

of Cpm and Ctm with Cp and Cpk

97

= c, we have ((,0") points on the line

II ,ml
'

1d 0"=-3c parallel to the (-axis, and tangential to the semi-circle corre= sponding to we have at the point (T,idle). If Cpk= c Cpm c /(-m/ +3cO"=d

,r

,'I 'II III I ..

"

~ and

T. The index Cpn"

I':; t,

however, purports to measure the degree of 'on-targetness' of the process, and Cpm-HX) if ~---+T,as well as 0"---+0. With Cpm, specification limits are used solely in order to scale the loss function (squared deviation) in the denominator. Kushler and Hurley (1992, p.2) emphasize that, if Tis not equal to m, the following paradox can arise'... moving the process mean towards the target (which will increase the Cpkt and Cpmindices)' can reduce the fraction of the distribution which is within' specification limits.'
,

II
"I

lines 0");0). This corresponds to ((,0") points on the pair of , (and ( =m+d-3CO" and (=m-d + 3CO"

III
Ii
'II

II

Finally, we compare plots of values of (~, 0")corresponding to fixed values of Cp, Cpk and Cpm' If Cpm= c, then

.~

0"2+(~-T)2=(} ~r
These (~,0")points lie on a semi-circle with centre at (T,O), radius *d/c, and restricted to non-negative 0".
tThe Cpk index referred to here is the modified index described by Kane (1986), namely Hd-I~-TI}/a.

m,; 50,2d= 30)withc= -.t,1, 1and l

general value of c, exceeding
Cpk

These lines intersect at the point (m, Ad/c) and meet the (-axis al Ihe points (m-d,o) and (m+d,o). Boyles (1991, Figs. 4 and 5) provides diagrams representirtg these loci from the example already described in section 3.2 (USL = 65,LSL= 35,
'

Figure 3.1 is a simplified version, giving ((,0") loci for' a meet the (-axis between its points of intersec- . tion ((=m:ttd/c) with the Cpm=C semi-circle.) If T=I=m,he t lines Cpk= c and Cp= c would be unchanged but the Cpm c = semi-circle would be moved horizontally to centre at the' point (T, 0), with points of intersection with the (-axis at ~= T:ttdlc.
""
, "-"':"'~';'~~';'>\

= c would

t with T = m. (If c < t the

lin~s

"">'~\

98

The Cpm index and related indices

(J

Estimationof Cpm
~

(and C1~m)

99

Tbe corresponding estimator of C;m is

'* C!"n (a)

!~/--=E:~--::_!!1l! 3

{n

}-

(Xi i~ I

-j

T)2

(3.16)

.(a)

.1(11) II ~

This,also,is a biasedestimator.The biasesof both Cpm and

C;m wiII general1y be positive (because the bias of a' is negative). For the case when each of Xl"", Xn has a normal distribution with expected value ( and standard deviation (J, the

i
m-d m'-:!d/c m m-dd/c m+d

r- ~
11'1

exact distribution of distribution of

Cpm

(and of C;m) can be derived. The

n

Fig. 3.1 Loci of points (~,a) such that: (a) Cp=c; (b) Cpk=C; (c) Cpm=c; when T=m, C>5. (If c<t m.-d/3c<m-d<m-t-d<in+ dj3c.) If c=} then a=d.

na'2 =

2::
i~ I

(Xi

- 1')2

I" I j

is that of (J2x (noncentral X2 witb n degrees of free,dom and noncentrality parameter A=n((- Tf(J-2), symbolically'
(J2X~2(A)
I
lit

3.4 ESTIMATION

OF Cpm (AND C~n,)

(3.17)

3.4.1 Point estimation A natural estimator of Cpm is obtained E[(X - Tf] by its unbiased estimator
(;'2=

This isa mixture of (central) chi-squared distributions with (n+ 2j) degrees of freedom and corresponding Poissod, weights
PltA)=exp(
'

by replacing

, ,"~

.(l.A)j -:tA)-~
J.

, j=O, 1,...
.'

j
I

(3.1~)

~to; 1 (Xj-T)2 I n

(3.14) ~I"I
-'

(See section 1.5.2). The zero) isexpected value of

C~m

(the rth moment of (1A)j

Cpm

about
I

Of course, a' is not an unbiased estimator of (/, but has a negativebias, since E[a'2]> {E[a']}2.Nor is t,dja' an unbias. ed estimator of Cpm' We define
d I n Cpm = -.:.. = d - '\' (X.-I L., 3(J 3 {n
I

,I

Jl~(Cpm)=

..

(-~ ) exp(-:tA) J- 0 ~E[(X;+2j)-!jr I .~ .1

r
,

dvhz

00

I
~

- dv,r;; r

1' )2

-i

(3 )
(J

_1
exp(

IX)

(JA)j r(hn-r)l
.1. .,
)

j)
!

2A) J,2:: ~o

2Jrrn n -~.1 ')

, (3.19)

i~l

(3.15)

')
(Fo, r;;'n,I4(C~rn) infinite.) is

I I

100

The Cpmindex and related indices

1"11

Estimation of Cpm (and C;m)

101

The rth moment about zero of C~m is obtained from (3.19) by replacing d by d-IT-ml. Taking r=l and r=2 we 'obtain
/11(c pm ) = E[ Cpm ]
I

Since r(y-l)jY/ny 1 for all y~ 1 the bias is positive, as expected see the remarks following (3.16). Some authors, for example, Spiring (1989) use the estimator
h.

= d~

' ( _lJe ) ~ U!)}q~(n-l)+J) L., ., 2\r( 1 + .1) 3(1 exp 2 j=() .I, 2n

Cpm=

(3.20 u)

t II

3{n~l,t, (X,- r)'

r ": lYC,rn (3.23aJ

~

and var(Cpm)= in place of Cpm, (replacing the divisor n by (n-1), in the deno'minator). This is presumably inspired by analogy with . the unbiased estimator of <12,
S2
I-I ,

( )\XP(-~Je) ,~ T
00

d~
3(1

(lJe)}

)-0

n-2+2j (3.20b)

- {E[CpmJ}2

=~ f. (Xi-X)2 n-1i=1
.

Table 3.1 gives values of E[Cpm] and S.D. (Cpm)= (var(C"",))! for sclel:ted values of /I, tlla and I~" Tlla( ~j, (See a~so the note on calculation of these quantities, at the end ofithis section.) Taqle 3.2 gives corresponding values of Cpm'
.

Since
.

Cpm

< Cpm,the bias of Cpmmust be less than that of

the bias of Cprn is positive, however, that of

I,

C\'m. Although
Cpm

can be negative,though it cannot be less than

When

~= T

we have ),=0 and

'

-Cpm( 1- (n: 1Y)

(3.21 a)

(3.23b)

d
Cpm

3<1( = Cp)

\'I
II

(Suhbaiah and Taam, 1991). (This follows from the fact that
~ .

E[C

pm
2

] = d~ n!(n-l))
3<1 2!n!n)

(321 b)
"

E[CpmJ-Cpm= -nn-l

var(CpnJ= ~
I

I
,

( )[
pm

d~

1 n!(n-1))

n-2 - '2{ n!n) } ]

.

(3.21c)
",I

> -n-

( ) E[CpmJ-Cpm ( )
!

n~

I !

Cpm-<;pm

cL (2.16a) with f= n, and the bias of Cpm,as an estimator ~)f

Cpm,\is' \ ' ' . , . .
'

because E[CpmJ > Cpm') If thl1 bias of Cpm (as well as that of Cpm) is positive, and

E ,[ C

pm

] - C =!

3<1

[( 2)

~ \ n!(n-1))
n!n)

-l

J
:,

(3 "2 )

also ~= T, then the mean square error of

I

.-

Cpm is less than that of Gpm,because both the variance and the squared bias of Cpm are less ~han those of Cpm' It should be noted, however, that this

168718

. .0 .f;-g . 5tS~
'.,J./' JI..

Prnr"""Q

"',..""..,~"

J_-"J_..

~.'.

Table 3.1 Moments of Cpm:E= E[CpmJ;S.D.=S.D. (Cpm)

I(- TI/G
0 Q5 S.D. E S.D. E 1

1.5 S.D. E S.D. E

2 S.D.

E n= 10
ciG -

.:; 6
11=

0.722 1.084 1.445 1.806 .2.167

0.183 0.275 0.366 0.458 0.550

0.644 0.967 1.289 1.611 1.933

0.16] 0.241 0.322 0.402 0.482

0.501 0.752 1.002 1.253 1.503

0.110 0.165 0.220 0.275 0.330

0.386 0.578 0.771 0.964 1.157

0.069 0.103 0.137 '0.171 0.206

0.307 0.460 0.613 0.767 0.920

0:044 0.066 0.088 0.110 0.132

15

"G
~
.:;

0.702 1.054 1.405 1.756 2.107

0.139 0.209 0.278 0.348 0.417

0.627 0.122 0.941 1).183 1.254 0.244 1.568 0.305 1.881 . 0.366

0.490 0.736 0.981 1.226 L471

0.084 0.126 0.168 0.210 0.252

0.380 0.570 0.760 0.950 1.140

0.053 0.080 0.106 0.133 0.160

0.304 0.456 0.607 0.759 0.911

0.035 0.052 0.070 0.087 0.104

'-

...

-.-"
---

.-'-

..-. ---_.

. .

n=20 dJG

2 3 4 5 6 n=25 diG 2 3 4 5 6 n=30 diG' 2 3 4 5 6

0.693 1.040 1.386 1.733 2.079

0.fl6 0.174 0.233 0.291 0.349

0.619 0.928 1.238 1.547 1.857

0.102 0.153 0.204 0.255 0.306

0.485 0.728 0.971 1.214 1.456

0.070 0.106 0.141 0.176 0.211

0.377 0.566 0.755 0.943 1.132

0.045 0068 0.090 0.113 0.135

0.302 0.453 0.605 0.756 0.907

0.030 0.044 0.059 0.074 0.089

0.688 1.031 1.375 1.719 2.063

0.102 0.153 0.204 0.255 0.306

0.614 0.921 1.228 1.536 1.843

0.089 0.134 0.179 0.223 0.268

0.482 0.724 0.965 1.206 1.447

0.062 0.093 0.124 0.155 0.186

0.376 0.564 0.752 0.939 1.127

0.040 0.060 0.079 0.099 0.119

0.30J 0.452 0.603 0.754 0.904

0.026 0.039 0.052 0.066 0.079

0.684 1.026 1.368 1.710 2.052

0.092 0.138 0.184 0.229 0.215

0.611 0.911 1.222 0.528 1.833

0.080 0.121 0.161 0.201 0.241

0.481 0.721 0.961 1.201 1.442

0.056 0.084 0.112 0.140 0.167

0.375 0.562 0.750 0.937 1.124

0.036 0.054 0.072 0.090 0.108

0.301 0.451 0.602 0.752 0.903

0.024 0.036 0.048 0.060 0.071

Table 3.1 (Cont.)

1(- TI/o0

0.5 S.D. E S.D. E

1.5 S.D. E S.D. E

2 S.D.

E n=35 d/o2 3 4 5 6 n=40 d/o2 3 4 5 6

0.681 1.022 1.363 1.703 2.044

0.084 0.126 0.168 0.210 0.253

0.609 0.913 1.218 1.522 1.827

0.074 0.111 0.148 0.184 0.221

0.479 0.719 0.958 1.198 1.438

0.051 0.077 0.102 0.128 0.154

0.374 0.561 0.748 0.935 1.122

0.033 0.050 0.066 0.083 0.099

0.300 0.451 0.601 0.751 0.901

0.022 0.033 0.044 0.055 0.066

0.680 1.019 1.359 1.699 2.039

0.078 0.117 0.156 0.195 0.235

0.607 0.911 1.215 1.518 1.822

0.069 0.103 0.137 0.t71 0.206

0.478 0.717 0.956 1.196 1.435

0.048 0.071 0.095 0.119 0.143

0.373 0.560 0.747 0:934 1.120

0.031 0.046 0.062 0.077 0.092

0.300 0.450 0.600 .0.750 0.901

0.020 0.031 0.041 0.051 0.061

--- '

---.

.,-,,""'"

~_..

n=45 d/o2 3 4 5 6 n=50 d/o2 3 4 5 6

0.678 1.017 1.356 1.695 2.034

0.073 0.110 0.147 0.183 0.220

0.606 0.909 1.212 1.515 1.818

0.064 0.096 0.129 0.161 0.193

0.477 0.716 0.955 1.194 1.432

0.045 0.067 0.089 0.112 0.134

0.373 0.560 0.746 0.933 1.119

0.029 0.043 0.058 0.072 0.087

0.300 0.450 0.600 0.750 0.900

0.019 0.029 0.038 0.048 0.058

0.677 1.015 1.354 1.692 2.031

0.069 0.104 0.139 0.173 0.208

0.605 0.908 1.210 1.513 1.815

0.061 0.091 0.121 0.152 0.182

0.477 0.715 0.954 1.192 1.430

0.042 0.063 0.084 0.106 0.127

0.373 0.559 0.745 0.932 1.118

0.027 0.041 0.055 0.068 0.082

0.300 0.450 0.600 0.749 0.899

0.018 0.027 0.036 0.046 0.055

106

The Cpm index and related indices

I -I'

I
Table 3.2 Values of Cpm . I-TI
(J 1

Estimation of Cpm (and C~m)

107

where 1" 2)"

J has a Poisson distribution with expected value

d (J

0.0 0.667 1.000 1.333 1.667

2.000

0.5 0.596 0.894 1.193 1.491 1.789

1.0 0.471 0.707 0.943 1.179 1.414

1.5 0.370 0.555 0.740 0.925 1.109

2.0 0.298 0.447 0.596

To evaluate EJ [(n - 2 + 2J) - 11, we resort tCJthe method of statistical differentials (or 'delta method'), obtaining

2 3 4

E[n--2~+2J~ n-2:2E[JJ J dr(n-2+ 2J)-1

+
r= 1

5 6

.
.

0.745 .
0.894

!lr(J)
1

dF

.T=E[J]=J.j2'7

is established only under these rather [estrictive conditions. If minimization of mean square error were a primary objective, neither n nor (n 1) would be the appropriate divIsor in the denominator. The effect of correlation among the X s may be less marked for Cpm than for Cp (sec section 2.6), because
~

(3.25) The quantities Ilr(.I) are the central moments of the Poisson distribution with expected value p, namely
E[.lJ = 1,1; !/2 (J) = 1A.;!/3 (J) = 1,1; !/4(J) = 1,1 + iA.2; ...
t"

~. '
I
t
~I

n - 1E f,(X r- T)2 = 0-2 + (~- T)2

[ j= 1

I ',.

(Higher moments are given by Haight (1967, p. 7) or can be evaluated using the recurrence formula - for moments !l~(j) about zero -a!l~(J) 2[ t (.1)+ ~1(A/2)_r!/~.J
.Ie

whatever the values of Pi}= corr(Xj,X). Note on calculation of E[C;'mJ

/I~ 11 (.1)=

The series for E[C~mJ, obtained by putting 1'=2 in (3.19), is often quite slowly convergent. The following alternative approach was found to give good results. Since the expected value of (X;+2j)-1 is (n-2+2))-1, the expected value of C;m is (dJ;;/3(J)2 times the weighted mean of thcsc quantitics with Poisson wcights Pj( !A.).We can wri1c

- see e.g. Johnson et al. (1993, Ch. 3).)

.~

3.4.2 Confidence interval estimation ,r

Since
Cpm=td{ (J2+(~ - T)2}-j

E[C;mJ=

(~ )

dJ;;

1
EJ n-2+2J

(324)
II III

construction of a confidence interval for

construction of a confidence interval

Cpm

is equivalent to

!n

for ()'2+(~- T)2. If

108

The Cpm illdex alld related illdices

r..",

Estimation of cpm (and C:m)

109

statistics A(X),B(X),based on Xl, X 2, ... , X /I' could be found,

such that Pr[A(X) ~ 62 + (c;- T)2 ~ B(X)]

for large n we obtain the approximate 1O0(1-a)% interval

I I
= 1-a
(3.26)
:rt)

I'

L (Xin i=1

/I

T)2 :tZl-aI26[2n-l{62+2(~- T)2}]! (3.28)

,

(i.e. (A(X), B(X)) is a 1OO( - a);;) confidence interval for 1

62 + (c;

T)2) then

~t'll'
I

1'*

(3(B~X))i' 3(A~~))i)

'I b a
If I Ilr"f
W

would be a 100(1-a)!% interval for ('.1111' However, it is not possible to tind statistics satisfying (3.26) exactly, on the usual assumption of normality of distribution of X. We therefore try to construct approximate
100(1 - oc)% confidence intervals. Recalling that ~i'=1 (X i '1')1 is distributed
'I'

"

i
:1

where <I>(Zt-aI2)=1-a/2, for 62+(c;_T)2. However since ~ and 6 are not known (otherwise we would know the value' of Cpmand not need to estimate it), it is necessary to replace them by .estimates. The resulting interval for Cpm is of quite complicated form. Boyles (1991) and Subbaiah and Taam (1991) use a better approximation to the distribution of X~2(A),namely that introduced by Patnaik (1949). This is the distribution of Cx] where (T2+2(~-T)2 c=- '62+(c;-T)2 {a2+(~-T)2}2 J= 62+2(~-T)2 n 62

as 62X;.\~.) with
.

1"1

A=n{(c;- T)/6}2, we have

and

var(itl (Xi- T)2) =2ncr2{ 2+ 2(~- T)2} 6

(as well as

(3.27)

1'1

r It

(The values of c and J give the cQrrect expected value and variance.) It is still necessary to use estimates c and j, of c and f respectively. Now 1
I"~, II

ELtl (Xi-T)2]=62+(c;-T)2
i

II

which is valid if E[X]

= c; and

var(X)= 62, whatever the

.

distributionof X).
/I

( )

cpm 2 -.L (Xi- T)2 ..,.,.- - n.=1

Cpm' 62+(c;-T)2
II

-6-2 "- I (Xi- T)2 L n 2

If we use a normal approximation to the distribution of

1+ -

6 {~-T } .

L (Xii= 1

6-2

T)2
'.
I

-, --

(Xi- T)2 i=1 cf

(3.29j

0'

110

The Cpm index and related indices

~~

Estimation of Cpm (and Ctm)

111

and hence is approximately distributed as a2ex}

a2cf

and (3.34)

=T
'

X}

(3.30) Cpm(1--9~~ tZl-rx/2(9}

I~

and P r,[( I
'I

YY

XI.a/2 ) L-pm<

1?1

C pm<,

-I (I

XI. 1"a/2

. )1?1

1'" t L'pm =~ -a

I
II

OJI)

where xL is definedby Prexy< xL,] = /:

We construct an approximate 10O(1-a)(Yc) confidence interval

. ~

For practical use, we recommend (3,33). Several other approximate formulas can be obtained for 100(1-a)% confidence intervals for Cpm,using approximate distributions of various functions of Cpm(e.g. log Cpm,.)C;:, etc.). Subbaiah and Taam (1991)present results of a compara-

tive Monte Carlo study of sixteensuch formulas. . . , From (3.17)it followsthat (Cpm/Cpm)2 distributed as is
1
1+
.

n {.

( a )}
~

'

-1
x [noncentral X2 with n degrees of

(( J

lc

. x;"1 a/2 ~C pm' f pm

) )

(3.32)
.~

freedom and noncentrality parameter n{(- T)/a}2J

Writing (~- T)/rr=(), we have

for Cpmby using the estimate " n{82+(~-T)2}2 .I = 82{82 + 2(~- T)2}
i \

'i

p{ X;'\/2(n(52) n(1 + 152)( ~::Y < X~~ <

whence.;

I _rx/2(nc52)

='

1-

IX

where

~ 8-=- 1 i= 1 (Xi-X) -2 t n-1

?

and

~=A

A;?

p{Co ~ 02)YX~.rx/2(nc52)Cpm < Cpm <

<

Further approximating the distribution of Xl leads to the approximate 100(1-IX)% confidence intervals

(
I

11(1-1- ()2)

)
A

!
X:,. I . adno2 )Cpm

]= 1

IX

It is tempting to say that 1

Cpm 1:\:

Zl-rx/2

J2! )

(3.33)

((n(1+c52)

XII.o</2(nc5Cpm, )

n(1+c52)

X~.1-rx/2(nc5 )Cpm

-I
i

112

The Cpm index and related illdice~ .

111

Boyles' modified Cpm index

!1

113

provides a 100(1-a)% confidence interval for Cpm' Unfortunately, calculation of these confidence limits involves knowledge of the value of b. If the natural estimator

as a modification
!

of Cpm for use

when

~# T. Here
(3.36a)

,1

Ex.-:'['[(X - T)2]

= E[(X ..- 1')21 X < T]Pr[X < T]

i
lit

b=X-T S
is used, the resulting confidence interval will not have 100(1- oc)%coverage,but something rather less. If ~= T, then Cpm= Cp(see 3.21 a),.and Cpmis distributed as

and EX>T[(X - 1')zJ =E[(X - T)ZIX> TJPr[X> TJ (3.36b) These are sometimes called semivariances. (The possibility that X might equal T is ignored, since we suppose that we are dealing with a continuous variable, and so Pr[X = T] =0.) Boylesintroduced this index from a loss function point of, view, to allow for an asymmetric loss function. However, it has some features in common with the index Cjkp, to be introduced in Section 4.4 which is intended to allow for some asymmetry in the distribution of X (the process distribution). Note that, if T=!(LSL+ USL). then C:m=Cpm, so C:m is not affected by asymmetry of the process distribution in that case. A natural estimator of C:m is C:m=
ill

u.)

f
X...a/Z

f'"",~ 1".

II
Iw II'i.

Thedistribution f Cpm o

for sample size n is thus the same as that of C"for sample size (n + 1).The unit increase in effective si:le of sample comes from the additional knowledge that ~= T. In particular, we obtain 100(1-a)% confidence limits

'1IIt.~ .\ 11.1\ Q

(-fitc

. X..,l-aI2 C

pm

-fit

pm

)
",I ,Ii
\I (

for Cpm( ,Cp)in this case. = Chan et ai. (1988a) discuss the use of Bayesian methods (with a 'noninformative' prior distribution for C1) this situation, and in provide some tables. See also Cheng (1992) for a nontechnical description (with further tables) of this approach,
3.5 BOYLES' MODIFIED CpmINDEX

! [! {(T-LSL)-2 3 n
+(USL-T)-z'I
.

(Xi-T)2 XI<1'
-1

(Xi-T)2
Xi>T

}]

(3.37)

In general, the distribution of C:m is complicated, even when the process distribution is normal. We note that

Boyles (1992) has proposed the PCI

w=
(3.35)
II

n
~+2 9C =(T-LSL)-2
'pm

X,<1'

(Xi-T)2 (3.38)

t:m =t[(T-LSL)-2Ex<T[(X - T)2] +(USL- T)-ZEx>T[(X T)Z]r! -

+(USL'"'

T)-Z

XI>T

L (Xi-

T)2 .

~,:

114

The Cpm index and related indices

We now derive expressions for the expected value and variance of W when Xi is distributed normally with expected value ~ equal to T, and variance (Jz. (Of course, ~ and (J are not in fact known; otherwise we woud not need to estimate

')

i ,

Boyles' mod(fied Cpm index Hence the overall expected value of W is

115

them. The analysis studies what would happen if ~= T == m

and the formula for C:m is used.) Conditionally on there being K Xs less than T and (n-K) greater thall T, W is distributed as
'

E[W]

= E[E[W

IK]] =:tn{(T -- LSL)-z +(USL-

T)-Z}(Jz (3.42a)

anu the overall variance is var(W)= E[E[WZIK]J - {E[Tf']P

~"

'l

f 'J'

-,

LSL)

"

2 2

XK T

(lJSL ,- I ) ,

' '

'

2 XII

I, 2 K j (J

nJ9)

=E[2{(1'-LSL)-4K +(USLT)-4(n-K)} + {(1'- LSL)-zK

the two X2S being mutually independent. The conditional expected value and variance of W, given K are

~i

E[WIK] = {(T- LSL)-z K +(USL- T)-Z(n- K)}(Jz

(:1.40a) and var[WIK] = 2{(T -LSL)-4K +(USLT)-4(n-K)}(J4 (3.40h)
"

+(USL-- 1')-z(n- K)}Z] -,tnZ{(T - LSL)-z +(U'SL- T)-Zp(J4

= n[(1'-LSL)-4+(USL-(USLT)-Zp](J4
(JZ)

T)-4+i{(1'-LSL)-z (3.42b) we have

Since X is distributed as N(T,

~= T. Hence

Since Pr[X = T] = 0, the distribution of K is binomialwith parameters (n,~), hence

I
I

('

+Z
pill

1 ------ <)x 1[(T- LSL) 2 +(USL-

T)-Z}(Jz

E[K] = E[n-K] =1n

var(K) = var(n - K) = tn

(3.41a)
'1
I

and so

(3.41h)
'I
I

E[KZJ = var(K) + {E[K]}Z = ,tn(n + 1)= E[(n - K)Z]

(1141')

c:; W, (3.43) c:; = tn{(T-LSL)-z +(USL- 1')=2}(Jz

whence

and E[K(n- K)] =nE[KJ - E[KZ] =tn(n-l) (3.41d)

E[(~YJ=1

(3.44 a)

116
and

The CPI1lindex ilml relil/ed indices

I

The Cpmk index

117

var -

c+ pm

2
'

(( 6;'"

))

. - [tn{(T- LSL}-2+(USL- T)-2}(J"2Y 2}2

(3.44 b)
"11)

var(W)

3.6 THE CpmkINDEX

I

- 4{(T-LSL)-4+(USLT)-4} + {(T-LSL)'-:! -(USL-T) {(T-LSL}-2+(USL-T)-:-2}2

I
III LH

j. I

Cpk is obtained from Cp by modifying the numerator; Cpm is obtained from Cp by modifying the denominator. If the two modifications are combined, we obtain the index

d-I~-ml
Cpmk= 3{E[(X - T)2J}i

d-I~-ml
= 3{(J2 +(~ - T)2}i

(3.46)

Boyles (1992) suggests that the distribution of (C:'1;IC;m)2 .might be well approximated by Ih:11 a 'mean chi-squared' of X~Iv. This would give the correct first two moments, if v has the value
n{(T-LSL)
2{(T

introduced by Pearn et al. (1992). We note that

\
Cpmk=

2+(USL

T) 2}2

- LSL)-4 + (\}SLnO +R2f 2(1+R4)+t(l-R2f

T)-4} +t{(T - LSL)- 2- (USL -- T)- 2}2

1-~

={1

) ~+ ( ).
Cpm
2

I~-ml

1'

-! Cp k

(J

with

~(1_1~

:ml){1+ (~: Tn c,

(3.47)

R= USL-T

T - LSL

(3.45)

")
II II.
"(J
110 IlK

I
I

[If T=-!(LSL+ USL), then R= 1.J Although C;,; is not an unbiased estimator of C;~ (nor is C;m, of C;m), the statistic W=nl(9C;~) is an unbiased estimator of (9C;~)-1. It is not, however, a minimum variance unbiased estimator. Details of calculation of such an estimator, based on the method outlined in section 1.12.4, are shown in the Appendix to the present chapter. \. The index Cjkp, to be described in section 4.4 also ust:Sthe semivariances, and, in its estimator, the partial sums of squares
X;<T

,
I
I

not know that ~=T=m.

Hence Cpmk~ Cpm, and also Cpmk~ Cpk' In fact Cpmk= CpkCpm/Cp. If ~=m=T then Cp=Cpk=Cpm=Cpmk' However, if we were estimating Cpmk the other PCls) we would (or

A natural estimator for Cpmkis I

C~k=

d-IX -ml
I
n-

(3.48)

3{;; i~l (Xi- 1')2}

The distribution of Cpmkis quite complicated, even if m = T. We first suppose, in addition to normality, that ~=m=T and obtain case. in that an expression for the rth moment about zero of Cpmk

III

".

(Xi-T)2

and

Xi>T

(Xi-T)2.

.118

The Cpm index and related indices

The Cpmk index

119

If ~= T=m, we have,

=~= crpmk n

:\ In.l

()

(-l)j

( )(
~

~
,

r-.j

.I

(nj"2

)
(3.50a)

nir(d-IX -ml)'
-m)2 }

!r

3' { i~! (Xi-X)2+n(X

rct(j+ l))[(l(n-r+j)) .
[(1(n+ 1))

=~t (-1)j ( )
~

dr-jn!(r-j),

3 j=O

r
jj

If we allow for the possibility that

~ is

not equal to m

n(X -m)z.

(though still assuming T = m), then (a) remains unchanged, but Yz is distributed as
(}"2 x(noncentral XZ with one degree of freedom and noncentrality parameter A=(}"-Z(~-m)2)=i(}"ZX'!\A)

{ i= !

i (X1-X)z+n(X
1

-m)Z

}
j(r

- j)
(3.49)
(~ 'J)
O'

Recalling from (1.33) that x't (A) is distributed as a mix-

X { it! (X/-X)Z+n(~

-m)Z

ture of central XI 21 distributions with Poisson weights +

, '

We now use the following well known results:

n

bl ~
(g\
'

eX:p(-1A)(H)i/i! (i=0,1,...), by..

(c) and .1)must be replaced

(a) Y! =

L (X/i= !

X)2 is distributed as (}"2 !; X;,

(b) Yz =n(X -m)Z is distributed as (}"2Xt; (c) Yz/(Y!+ Yz) has a BetaCLi(n -1)) distribution (see section 1.6); (d) YZ/(Y1 Yz) and (Y! + Yz) are mutually independent (see + section 1.6.) Then, from (3.49) ~md (aHd)
\,

. (c,d)' thejoint distribution of Yz/(Y! + Yz) and (Y1+ Yz) is that of a . mixture of pairs of independent variables . . . Z " I (1 . 1 WII, d Istn b utlOns Bctaz+l, z(11- 1)), XII+Zi espectIver I I ly and the Poisson weights given above. (Remember r that X:,2, ;(A)+ X~ I is distributed as X::,z;(A) if the 2 two variates are independent.)
I

I I
.1

'

For this more general case we find

to.

I~

E[C~mk]=rreXP(-A/2).f 1=0 (A~~)/i (-1)j l. j=O

x

(1)( )
~
()"

dJn

r-j

-r E[Cpmk]=-;:

1 3

L (-1)J j=O
r

.)(--

dJn
()"

r-j

E[ {Beta(i+i,

i(n-1))}jjZ]E[(Xf

+2;)-(r-j)fZ]

x E[{Beta('ti(n-1))}!J]

E[(X;)-!(r-j)]

= 3 -r

exp(- A/2)

1=0

(Ae)i
I!

120
x

The Cpm index and related indices

Cpm(a) indices
HI d~1 I

121
Cpmk

L
j:O

(- 1).1,

( )( ~-)
J

1'

1
(J

1 r-.i

From (3.50c) and (3.50d), the variance of

derived. If A = 0 we have

can be

.I

1(1(1 + j) + i)r(1(n - I'+ j) + i) x f(!+i)f(!(n+j)+i) Taking

1'=

(3.50b)
tll/l)

E[C~,nkJ=!
9

1 we obtain
]. ==1

[ n-2

-~

( )
(J

~j;I

- ~(n-1)~

( )
(J

dj;I + !

(A/2)i ~ n f(!(n-1)+i) l . ( E[ L-pmk 3exp -,1 /2) L...,., [ (J 2 r( 2n+1 ) i:() I.

.:>I

\
I
\

11 (3.501)

dl
HI

Table 3.3 gives values of expected values and standard

f(1+i)f(1n+i)

f(1+i)r(1(n+1)+i)
00

J
[ (J
1

I I
\

1111

[,

deviations of Cpmb from (3.50) for a few values of 11and dl(J. Remember that these values apply for ~= T = m. Just as Cpmk~ Cpm, with equality if ~= m, a similar argument shows that Cpmk Cpm' However var(Cpmd~ var( Cpm)' ~ If T;:m we have from (3.46)
1--I

= iexp( - ,1/2)L
2f(1

i=O

(A/2); d :;- L

- Rj
2

~.-m
(J
l

+ i)

- (n-1 + 2i)f(1+ i) RiJ

(3,50 c)
I~ICo)

~
d
,

(3.51)

C,mk~ 30{l+(,:m)'r with Ri=r(1(n-1)+ i)/r(1n+ i). Taking ,1.=0we find

d 1 E[Cpmk]=- 3 (J [
'

,.
.

~2R-L
00

2
(11-1)

1
1t .,ficRJ

tI,l 11,1)

,(3.50d)

Table 3.4 gives values of Cpmk for representative dj(J and 1~-mll(J,

values of

with R = nHn-1))/nin). I Now taking 1'=2, we obtain from (3o50b) after some simplification
'

3.7 Cpm(a) INDICES

'lMltno
ro

E[CpmkJ hxp( -!A) =

i'cO

(!W d./h -"T

I.

[( ,(J

)-

11-

1 ' 2 2

r.

I~ l'Ii
ell
I

The index Cpm attempts to represent a combination of ef. fects of greater (or less) variability and of greater (or less) relative deviation of ~ from the target value. The class of indices now to be described attempts to pro,vide a flexible choice of the relative importance to be ascribed to these two e,ffects. From (3.8), we note that if '" = I~- TI/(J is small, then
Cpm ~(l-1(2)Cp

- 2-fi(
\

dj;t
(J

+ i) ( 1 n!+i) \n-1+2i
[(1

+ 1+ 2i

n+2iJ
(:\.~O e)

'I
I')

(3.52)

Table 3.3

Moments of Cpmk: E=E[CpmkJ;

S.D.=S.D.
I(-ml/a

(Cpmd

0.0

0.5

1.0
S.D.

1.5 S.D. E S.D. E

2.0 S.D.

E
11=10 d/a 2 3 4 5 6 II= 20 d/a 2 3 4 5 6

S.D.

0.636 0.997 1.359 1.720 2.081

0.193 0.281 0.371 0.462 0.553

0.491 0.813 1.126 0.458 1.780

0.192 0.268 0.347 0.426 0.505

0.264 0.515

0.135 0.187 0.765 0.241 1.016 0.295 1.266 0.350

0.106 0.299 0.492 0.684 0.877

0.081 0.114 0.148 0.182 0.216

0.006 0.160 0.313 0.467 0;620

0.051 0.072 0.094 0.116 0.138

0.633 0.979 1.326 1.672 2.019

0.124 0.180 .0.237 0.294. 0.352

0.470 0.779 1.089 1.398 L708

0.129 0.176 0.225 0.275 0.325

0.249 0.491 0.734 0.977 1.220

----

0.087 0111 0.155 0.190 0.225

0.099 0.188 0.476 0.665 0.854

0.054 0.076 0.098 0.120 0.142

-'

0.003 0.154 0.305 OA57 0.608

"

0.035 0.049 0.063 0.078 0.093

...... ..
n=30 d/a 2 3 4 5 6 n=40 d/a 2 3 4 5 6 n=50 d/a 2 3 4 5 6

..

111!118!

".""'-:::

-~.

. t..;

._-

". c

,-

'"-

.-.~---,--

0.635 0.977 1.319 1.661 2.003

0.099 0.142 0.187 0.232 0.278

0.462 0.768 1.074 U79 1.685

0.103 0.141 0.179 0.218 0.258

0.244 0.48:' 0.725 0.965 1206

0.070 0.096 0.123 0.151 0.178

0.097 0.284 0.471 0.659 0.846

0.043 0.060 0.078 0.096 0.114

0.002 0.152 0.303 0.453 0.604

0.028 0.039 0.051 0.063 0.075

0.637 0.977 1.317 1.656 1.996

0.084 0.459 0.121 0.762 0.160 . 1.066 0.198 1.370 0.237 1.673

0.088 0.120 0.153 0.186 0.220

0.242 0.481 0.720 0.959 1.199

0.060 0.082 0.105 0.129 0.152

0.096 0.282 0.469 0.656 0.843

0.037 0.052 0.067 0.082 0.097

0.002 0.152 0.302 0.452 0.602

0.024 0.034 0.044 0.054 0.064

0.639 0.978 1.316 1.654 1.993

0.075 0.108 0.141 0.175 0.210

0.456 0.759 1.061 1.364 1.666

0.079 0.107 0.136 0.165 0.195

0.241 OA 79 0.718 0.956 1.194

0.053 0.073 0.093 0.114 0.135

0.095 0.281 0.468 0.654 0.840

0.033 0.046 0.059 0.073 0.087

0.001 0.151 0.301 0.451 0.601

0.021 0.030 0.039 0.048 0.057

Table 3.3 (Cont1 I~-ml/(J 0.0 E n=60

dl(J

0.5 S.D. E S.D. E

1.0 S.D. E

1.5 S.D. E

2.0 S.D.

2 3 4 5 6 n=70
di(J

0.641 0.978 1.316 1.653 1.991

0.068 0.098 0.128 0.159 0.190

0.455 0.756 1.058 1.360 1.662

0.071 0.097 0.123 0.150 0.177

0.240 0.478 0.716 0.954 1.192

0.048 0.066 0.085 0.104 0.123

0.095 0.281 0.467 0.653 0.830

0.030 0.042 0.054 0.066 0.079

0.001 0.151 0.301 6.450 0.600

0.019 0.027 0.036 0.044 0.052

2 3 4 5 6
.-----------------.-----iii ~

0.642 0.979 1.316 1.653 1.990

0.063 0.090 0.118 0.147 0.175

0.454 0.755 1.056 1.357 1.659

0.066 0.089 0.114 0.138 0.163

0.239 0.477 0.715 0.952 1.190

0.044 0.061 0.078 0.096 0.113

0.094 0.280 0.466 0.652 0.838

0.028 0.039 0.050 0.061 0.073

0.001 0.151 0.300 0.450 0.599

0.018 0.025 0.033 0.040 0.048

---

"---

L~

"
=-~.~ ~~-~ ~---_..-

n=80 dl(J 2 3 4 5 6 n=90 dl(J 2 3 4 5 6 n= 100 dl(J 2 3 4 5 6

0.643 0.980 . 1.316 1.653 1.989

0.058 0.084 0.110 0.137 0.163

0.453 0.754 1.055 1.355 1.656

0.061 0.083 0.106 0.129 0.152

0.239 0.476 0.714 0.951 1.188

0.041 0.057 0.073 0.089 0.105

0.094 0.280 0.466 0.651 0.837

0.026 0.036 0.047 0.057 0.068

0.001 0.150 0.300 0.449 0.599

0.017 0.024 0.031 0.038 0.045

0.644 0.980 1.316 1.653 1.989

0.055 0.079 0.104 0.129 0.154

0.452 0.753 1.053 1.354 1.654

0.058 0.078 0.100 0.121 0.143

0.239 0.476 0.713 0.950 1.187

0.039 0.054 0.069 0.084 0.099

0.094 0.280 0.465 0.651 0.837

0.024 0.034 0.044 0.054 0.064

0.001 0.150 0.300 0.449 0.599

0.016 0.022 0.029 0.036 0.042

0.645 0.981 1.317 1.653 1.988

0.052 0.075 0.098 0.122 0.146

0.452 0.752 1.052 1.353 1.653

0.055 0.074 0.094 0.115 0.135

0.238 0.475 0.712 0.949 1.186

0.037 0.051 0.065 0.079 0.094

0.094 0.279 0.465 0.651 0.836

0.023 0.032 0.042 0.051 0.060

0.001 0.015 0.150 0.021 0.300 0.027 0.44-9 0.034 0.599 0.040

126

The Cpm index and related indices

I I Appendix 3,A 127

Table 3.4 Values of Cpmkwhen T=m I-ml

(J"

d (J

=(n-l)'/2. 1.0 1.5

0.0925 0.277 9.462 0.647 1.832

(-IV
()

0.0
0.667 1.000 1.333 1.667 2.000

0.5
0.447 0.745 1.043 1.342 1.640 .

2.0
0.000 0.149 0.29H 0.447 0.596

1~

( ){
~

a(nn-l) } j -E[(U

+~.\/;;)2j]E[Xn-:i'2j]
(3.56 a)

2 3 4 5 6

0.236 0.471 0.707 1.943 1.179

In particular using (2.16a, b)

E[Cpm(a)]

= { 1-

and It is suggested that we consider the class of PCls defined hy j

Cpm(a)=(1-a(2)Cp (3.53)

a(n -1) 2 n(n-4) (1 + n( )} E[Cp]

(3.56 b)

E [{C
'

pm

2 2a(n-l) } (a) ]= { 1-(I+n(Z)+

.
'

n(n-5)

'.

where a is a positive constant, chosen with regard tq the desired balance between the effects of variability and relative departure from target value. A natural estimator of Cpm(a)is

a2(11-1)2

~Z(n 5)(n -7) (3 + 6n(2 + n2(4) } E[C~J

(3.56c)
It must be remembered that the Cpm(a)indices suffer from the same drawback as that noted for Cpmin section 3.2, namely, that if T is not equal to m=:hLSL+ USL), the same value of
t,

Cpm(a)={ 1-a (

X- T
s-

d X- T s} CP=3S { 1-a

) } (3.54)
(3.55)

Cpm(a),obtained

when c;=T-~ or c;=T+~, can correspond

Assuming normal (N(~, (JZvariation, Cpm(a)is distributed as (n-1)1

Xn-l {

1- a(n; 1)(U+ '~)Z

nXn-l

Cp

to markedly different expected proportions of NC items. It is possible for a C"m(a)indcx to bc negative(unlikeC" or Cpn" but like Cpd. Sincc this occurs. only for large values of "I, the relative drawback. . would not be a departure of c; from the target value T, this GUpla and Kolz (1993) have labulated values of E[Cpm(1)J and S.D.(Cpm(!). ,
APPENDIX 3.A: A MINIMUM VARIANCE UNBIASED ESTIMATOR FOR W=(9C:,;r1

where U and Xn- 1 are mutually independent and U is a unit normal variable. From (3.55),the r-th moment of Cpm(a)is
E[{ Cpm(a)}'] =

This appendix is for more mathematically-inclined readers. Hthe independent random variables X1,...,Xn each have the normal distribution with expected value c; and variance

128
(f2, then

The Cpm index and related indices i

I

Appendix 3.A

129

Now, the inequality X,< T is equivalent to'

{( n
n -1
j= I

2
I

) (Xi. X)}
1

i (Xj-X)2

.t
WI

h X= -

n j= 1

~ Xj
I I
I

I I ,
I
'

'

r Vi<_-=L~ -S n-l X-T

Hence
EXi<I.[(Xi- T)2IX, S] =

(3.59)

, has a beta distribution with parameters !.!(n - 2); as defined , in section 1.6 (see e.g., David et al. (1954).)

= {Bet !n-l)} -1

The PDF of y:.=

, is thus

f-1

- Cn(X - T)fS
(X- T+c;;1Sv)2(I-v2)!n-2 dv (3.60)

'(n-l)2

( )
S

Xi-X

=(~Ygn(Cn(~~- T))
with CII ; j;;/(n =

wIth (n-l)S2 = j~1 (Xj-X)2

- t) and
f~:(Z+V)2(I-V2)!n-2dv (3.61a)

fYi(y)={B(1.,!n-1)}-ly-!(1_y)in-2 and the PD F of

O<y<1

(3.57)

gll(z)={B(ttn-l)}-1

j;; S V'=n-l Xi-X

is

=fi

Since !(V/+ 1) has a beta distribution with parameters (tn--l, tn-I), an alternative form for gn(z) can be obtained in terms of incomplete beta function ratios (see section 1.6). We find sgn(X,-X)) '

YII(z)=4{h2hOn-t, !n,-I)-hhOn, n + 2n-l Ihnn+ 1,tn-I)}

fVI(v)={B(-~, tn-l)}-1(I-v2)!n-2 -1<v<1 (3;58)

tn-I) (3.61b)

The conditional distribution of Xi, given the sufficient statistics X and S, is that of the linear function of Vj:
,

where h= 1(1- z). Since E[~]=O, E[VrJ=(n-l)-:-t and X/-T=(X-T)+ (n-l/~)SV,

EXi<T[(x,-T) 2iT 2IA,S]+ExI>T[(Xj-T)'IX,S] (3.62)

n-l

X+

j;; SVi

= E[(Xi- T)2IX, S] =(X - T)2 + n-l S2 n

---.

---

Table 3.5-Function gn(x)

12

10 x -1 -0.95 -0.9 -0.85 -0.8 -0.75 -0.7 -0.65 -0.6 -0.55 -0.5 -0.45 -0.4 -0.35 -0.3 -0.25 -0.2 -0.15 1.1111 1.0136 0.9211 0.8336 0.7511 0.6736 0.6010 0.5334 0.4707 0.4118 0.3597 0.3112 0.2672 0.1276 0.1922 0.1608 0.1331 001 090

20 1.0526 0.9551 0.8626 0.7751 0.6926 0.6151 0.5426 0.4751 0.4126 0.3551 0.3025 0.2549 0.2122 0.1743 0.1410 0.1122 0.0877 0.0672

30 1.0345 0.9370 0.8445 0.7570 0.6745 0.5970 0.5245 0.4570 0.3945 0.3370 0.2845 0.2370 0.1944 0.1568 0.1240 0.0960 0.0724 0.0532

40 1.0256 0.9281 0.8356 0.7481 0.6656 0.5881 0.5156 0.4481 0.3856 0.3281 0.2756 0.2281 0.1856 0.1481 0.1155 0.0877 0.0646 0.0460

50 1.0204 0.9229 0.8304 0.7429 0.6604 0.5829 0.5104 0.4429 0.3804 0.3229 0.2704 0.2229 0.1804 0.1429 0.1103 0.0827 0.0599 0.0416

100 1.0101 0.9126 0.8201 0:7326 0.6501 0.5726 0.5001 0.4326 0.3701 0.3126 0.2601 0.2126 0.1701 0.1326 0.1001 0.0726 0.0500 0.0324

200 1.0050 0.9075 0.8150 0.7275 0.6450 0.5675 0.4950 0.4275 '0.3650 0.3075 0.2550 0.2075 0.1650 0.1275 0.0950 0.0675 0.0450 0.0275
--._-

600 1.0017 0.9042 0.8117 0.7242 0.6417 0.5642 0.4917 0.4242 0.3617 0.3042 0.2517 . 0.2042 0.1617 0.1242 0.0917 0.0642 0.0417 0.0242
--------

1000 1.0010 0.9035 0.8110 0.7235 0.6410 0.5635 0.4910 0.4235 0.3610 0.3035 0.2510 0.2035 0.1610 0.1235 0.0910 0.0635 0.0410 0.0235
---_.-

-------------1.."'-

--------._'-__n-

-"'-"-'-

-." ..'

;."""

",""r''"

, ',:::;:-

------'-'---0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 0.0883 0.0705 0.0556 0.0431 0.0328 0.0246 0.0180 0.0129 0.0089 0.0060 0.0039 0.0024 0.0014 0.0008 0,0004 0.0002 0.0001 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0504 0.0369 0.0263 0.0182 0.0122 0.0079 0.0049 0.0029 0.0016 0.0009 0.0004 0.0002 0.0001 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.03'79 0.061 0.0172 0.0109 0.0066 0.0038 0.0010 0.0010 0.0005 0.0002 0.0001 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0315 0.0206 0.0128 0.0075 0.0042 0.0021 0.0010 0.0004 0.0002 0.0001 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 Om76 0.0173 0.0102 0.0056 0.0029 0.0013 0.0006 0.0002 0.0001 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0193 0.0105 0.0051 0.0021 0.0008 0.0002 0.0001 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0149 0.0068 0.0025 0.0007 .0.0001 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0117 0.0041 0.0008 0.0001 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 00110 0.0035 0.0005 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 O. 000 0 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

Bibliography
132

133

The Cpm index alld related indices

~II N. Iw

and so

E[WIX, S] =n[(T- LSL)-ZEx;<T[(X/- T)ZIX, S] Z Z +(USL- T)- EXI>T[(X/- T) IX,S]] =n(~)l{(T-LSL)-Z -(USL.

giving values of gll(z),will be of some assistance in this regard. Note that 9n(0)= -Hn-1) -1 and gn(-1) = 1+ (n-1)'- 1. It can be seen that gn(x) is small for x> 0, while for x ~ -0.25 gn(x)~xz+(n-l)-l (3.64)

H
M' 01 c'

T)-Z}gn(Cn(Xs- T)) ;T CIIY ~~ I}] + (3.63)

+(USL-n-Z{(X

pr III
W' YI

The advantage (reduction of variance)attained by use of the MVUE has not been evaluated as yet. It may be worth the extra effort involved in its calculation and it also provides an example.of relevance of 'advanced' statistical methodology to problems connected with quality control. It should be noted that the Blackwell-Rao theorem (see section 1.12.4), on which the derivation of the MVUE is based, is almost 50 years old!
BIBLIOGRAPHY

The statistic W is, of course, more easily evaluated than would be the MVUE, E[WjX, S]. Tablc 3.5, and Fig. 3.2
1.2 1.0j\\ \\ ,\,
0.8

lit

,, ..,

C:urve SampleiSlze(n) 'I ~.~~= ,-,-,

, "

III
I

\.\ \,\ '..,

'. , '\\
'\.\ Ii..,.\ '. Ii\.\ \ .. Ii..\ \ -.. '\ "'-;~':'"
~';""'" ~""'>'"
"""

l
-

~~ 30
40

50

' "

~ 0.6 01
0.4

0.2

0.0 -1.0

-0.5

""""' ""-:::;:.-, ~ --.-0.0 x

--r0.5

J
1.0
I

II
"

I
'I

Fig. 3.2 The function gn(x) for various sample sizes.

Boyles, R.A.(1991)The Taguchi capability index, J. Qua/. Techno/.,23. Boyles, R.A. (1992) Cpm for asymmetrical tolerances, Tech., Rep. Precision Castparts Corp., Portland, Oregon. Chan, L.K., Cheng, S.W. and Spiring, FA (1988a) A new measure of process capability, CPIn,J. Qua/. Techno/., 20, 160-75. Chan, L.K., Cheng, S.W. and Spiring, FA (1988b) A graphical IcchniqtH~for process capahilily, Trans. ASQC Conyress, Dallas, Texas, 268-75. , Chan, LX., Xiong, Z. and Zhang, D. (1990) On the asymptotic distributions of some process capability indices, Commun. Statist. - Them'. Meth., 19,11-18. Cheng; S.W. (1992) Is the process capable? Tables and graphs in 'assessing Cpn" Quality Engineering 4, 563-76. Chou, Y.-M., Owen, D.B. and Borrego, S.A. (1990) Lower confidence limits on process capability indices, Quality Progress, 23(7), 231-236. David, B.A., Hartley, H.O. and Pearson, .S. (1954) Distribution of the ratio, in a single normal sample, of range to sta.ndard deviation Biometrika, 41, 482-93.

.,...134
The Cpm index and related indices

Grant, E.L. and Leavenworth, R.S. (\988) Statistical Quality Control, (6th edn), McGraw-Hili: New York. Gupta, A.K. and Kotz, S. (1993) Tech. Rep. Bowling Green State University, Bowling Green Ohio. Haight, F.A. (1967) Handbook of the Poisson Distribution, Wiley: New York. Hsiang, T.C. and Taguchi, G. (1985) A Tutorial on Quality, Control and Assurance - The Taguchi Methods. ArneI'. Statist. Assoc. Annual Meeting, Las Vegas, Nevada (188 pp.). . Johnson, N.L.. Kotz. S. and Kemp. A.W. (1993) Distrilmtio/ls
in Statistics

Process capability indices

under non-normality: ro bustness properties

Discrete

Distributions

(2nd edn),

Wiley:

New

York. Johnson, T. (\ 991) A new measure of process capability related to Cp"" MS. Dept. Agric. Resource Econ.. North Carolina State University, Raleigh, North Carolina. Kane, V.E. (1986) Process capability indices, J. Qual. Technol., 18, 41-52. Kushler, R. and Hurley, P. (1992) Confidence bounds for capability , iridices. J. Qual. Tee/mo/.. 24, 188-195. Marcucci, M.O. and Beazley, C.F. (1988) Capability indices: Process performance measures, Trans. ASQC Congress, 516-23. Mirabella, J. (1991) Determining which capability index to use, Quality Prouress, 24(8), 10. Patnaik, P.B. (1949) The non-central X2- and F-distributions and their applications, Biometrika, 36, 202-32. Peam, W.L., Kotz, S. and Johnson, N.L. (1992) Distributional and inferential, properties of process capability indices, J. Qual. Technol., 24, 216-31. Spiring, FA (1989) An application of Cpmto the tool wear problem, Trans. ASQC Conoress. oront'o, 123-8. T Spiring. 1".1\. 1991a) Assessing prm;ess capability in the presence or ( systematic assignable Cll se, J. Qual. T eclmol., 23, 125-34. u Spiring, FA (1991b) The Cpmindex, Quality Progress, 24(2), 57-61. Subbaiah, P. and Taam, W. (1991) Inference on the capability index: Cpm, MS, Dcpt. Math. ScL, Oakland University, Rochester, Minnesota.
.

4.1 INTRODUCTION

I U "8'
I' 1'8"
I

1-.1

The efTects non-normality of the distribution of the measof ured characteristic, X, on properties of PCls have not been a IP..ajor research item until quite recently, although soine practitioners have been well aware of possible problems in this respect. In his seminal paper, Kane (1986) devoted only a short paragraph to these problems, in a section dealing with 'drawbacks'. 'A variety of processes result in a non-normal distribution Jor a characteristic. It is probably reasonable to expect that capability indices are somewlJ.at sensitive to departures from normality. Alas, it is possible to estimate the percentage of parts outside the specification limits, either directly or with a fitted distribution. This percentage can be related to' an equivalent capability for a process having a normal distribution.' A more alarmist and pessimistic assessment of the 'hopelessness' of meaningful interpretation of PCls (in particular, of Cpk) when the process is not normal is presented in Parts 2 and 3 of Gunter's (1989) series of four papers. Gunter emphasizes the difference between 'perfect' (precisely normal) 135

.,

I-,

136

Process

capability

illdices under 1IOII-IIOrmality

Effects of non-normality
I 10I \ Iii

137

and 'occasionally erratic' processes, e.g., contaminated processes with probability density functions of form
,

I, l'i I

will provide such an insight into several features of a distribution. Indeed one of us (NLJ) has some sympathy with the
.

pcp(x;~1, ad

+ (1.- p)cp(x;~2' a2)

(4.1)

viewpoint f M. Johnson(1992), o whoasserts that '... noneof

the indices Cp, Cph Cpu, Cpmand Cpk adds any knowledge or understanding beyond that contained in the equivalent basic parameters J1,a, target value and the specification limits'.. McCoy (1991, p. 50) also notes: 'In fact, the number of reallife production runs that are normal enough to provide truly accurate estimates of population distribution are more likely than the exceptions', although (p. 51) '... a well-executed study on a stable process will probably result in a normal distribution'.
4.2 EFFECTS OF NON-NORMALITY

", '\ .1)

where 0 <p < 1;a1, a2 >0; (~1' ad:(~2' a2) and

I'
lill

cp(x;~, ) = (j21t<T)-.lexp{-~( x~~y} a

.,,1

'II Iii,

Usually p is close to 1 and (1- p) is small; the second component in (4.1) represents 'contamination' of the basic distribution, which corresponds to the first componenL It is often hard to distinguish such types of non-normality, in practice, by means of standard, conventional testing procedures, though they can cause the behaviour of Cpk to vary

i
11

II. ntl Ilc iln (JO.

1111

I
I 'I

quite drastically.

ny

The discussion of non-normality falls into two main parts. The first, and easier of the two, is investigation of the properties of PCls and their estimators when the distribution of X has specific non-normal forms. The second, and more difficult, is development of methods of allowing for non-normality and consideration of use of new PCls specially designed to be robust (i.e. not too sensitive) to non-normality. There is a point of view (with which we disagree, in so far as we can understand it) which would regard our discussion in the next three sections as being of little importance. McCoy (1991) regards the manner in which PCls are used, rather than the effect of non-normality upon them, as of primary importance, and writes: 'All that is necessary are statistically generated control limits to separate residual noise from a statistical signal indicating something unexpected or undesirable is occurring'. He also says that PCls (especially Cpk)are dimensionless and thus'... become(s) a handy tool for simultaneously looking at all characteristics of conclusions relative to each other'. We are not clear what this means, but it is not likely that a single index

ull tll nos

J IIS Il .Is 2 Itnd
,

Gunter (1989) has studied the interpretation of Cpk under three different non-normal distributions. These distributions all have the same mean (~) and standard deviation (a), they therefore all have the same values for Cpk, as well as CpoThe three distributions are: 1. a skew distribution with finite lower boundary

- namely

3 ,'"SI

xL distribution (Chi-squarewith 4.5 degreesof freedom) 2. a heavy-tailed([32>3)distribution- namelya t8distribution 3. a uniform distribution. In each case, the distribution is standardized (shifted and scaled to produce common values (~=0, 0' = 1)formeanand standard deviation). The proportions (ppm) of NC items outside limits:!: 3a are:(approximately)
.

nfl'lce, led

lI'nl
~Ol~~ .

For (1.) 14000 (all above 3a). . For (2.)4000 (half above 3a~half below
For (3.) zero!

- 3a).
- 3a).

A ,at ue

1-

dex

Arid for normal, 2700 (half above 3a, half below Figure 4.1 shows the four distribution curves.

138,

Process capahility indices under non-normality

Effects of non-normality Table 4.1 Values of Pr[CpcJ

'Ill

139

0.48 t(2) 0.32 J, I \ \ .. (iulIssiulI (normul)

from simulation (Cp= 1) Triangular 1.000 0.881 0.561 0.320 0.193 0.079 0.039 Un!form 1.000 0.924 0.522 0.267 0.152 0.058 0.025 Exponential 0.969 0.846 0.683 0.527 0.404 0.240 0.150

1/ 'I 'VI .t \

Distrihution of X 11= c= 5 0.50 0.75 1.00 1.25 1.50 2.00 2.50

Normal 0.996 0.873 0.600 0.373 0.223 0.089 0.041

0.16t-

I'

\\1

---

(3)

Ii)

10

0.00t-5.0

.../' I I
-2.5 0.0

I
'

2.5

.
7.5

0.50 0.75 1.00 1.25 1.50 2.00

5.0

I
I

--r
20
I' 0), I

2.50
0.50 0.75 1.00 1.25 1.50
') I\(\

1.000 0.933 0.568 0.243 0.091 0.013 0.002

1.000 0.979 0.548 0.129 0.017
fI flfI1

1.000 0.959 0.529 0.189 0.064 0.011 0.003

1.000 0.994 0.523 0.087 0.011
fI flfIfI

1.000 0.989 0.501 0.127 0.035 0.005 0.001 1.000

1.0pO 0:508 0.041 0.002
fI flAA

0.985 0.865 0.644 0.423 0.259 0.097 0.039 0.996

0.900 0.617 0.315 0.135
A A,.".,

Fig.4.1 Four different shaped process distributions with the same c; and 0', and hence the same Cpk'(Redrawn from Gunter, RH. (1989), Quality Progress, 22, 108-109.)

I 20 7.5
I

A number of investigationsinto effectsof non-normality on estimators of PCls have been published. English and Taylor (1990)carried out extensiveMonte Carlo (simulation)studies
of the distribution

I
i
I
ell ",v

..11 ~,r

triangular, uniform and exponential distribution"',with

sample sizes n = 5,10 (10)30, 50. The results are summarized
in part in Table 4.1. From these values of Pr[Cpc]
inter alia,

of Cp (with Cp= 1) for normal, symmetrical'

L~N 1'1

50

I I
I
J

we note,

'Iill II ..h II .Id

Ille,

that for n less than 20, there can be substantial,

departures from the true Cp value (as well as differencesin

interpretations of these values for non-normal distributions).
.. All four distributions have the same expected value (~=O), and standard deviations (0'= 1).The specification range was LSL= -J.to USL=3, so the value of Cp was t for all four distributions.

--~~n
Shll

II "Inl

I
,

AI., ~J).
dill he ('II ru

"

j
I
,

'I

Also, the values for the (skew) exponential distributions differ sharply from the values for the three other (symmetrical) distributions especially for larger values of c. On the more theoretical side, Kocherlakota et af. (1992)

140

Process capability indices under non-normality

Effects of non-normality

141

'have established the distribution of (;p = idj8 in two cases: when the process distribution is (i) contaminated normal with'
(11

t'r
11',lh 1)

= (12

= (1, see (4.1); and (ii) Edgeworth

series

The essential clue, in deriving expressions for the moments of (;" for this model. is the following simple observation. A random sample of size n from a population with PDF (4.4) can be regarded as a mixture of random samples of sizes N 1>"', N k from k populations with PDFs CP(X;~1,(1),...,cp(x;~k>(1),where N=(N1)...,Nk) have a joint multinomial distribution with parameters (n; P1,..., pd so that

fx(x)

= [1-iA3D3

+~A4D4 +-h-A~D6Jcp(x; 0,1)

.

(4.2)

where Dj signifies jth derivative with respect to X,

aJ:1,d

A3

{Ji2(X)}i=JP;

Ji3(X)

(4.3a)

~a)

and A4= Ji4(X) {Ji2(X))2 -3=/32-3 (4.3 b)

rr[N=nJ=r{jOl
-~ nl

(Nj=nj)]
k nP
"J
j

10)

Onj!j=t j= 1 k

are standardized measures of skewness and kurtosis respectively. (See section 1.1 for definitions of the central moments Jir(X) and the shape factors A3,A4,$, /32')

,I.

~I
I I

j=1 i I

Lnj=n

j=1

LPj=l

n=(n1",.,nk)
(4.5)

The conditional distribution of the statistic

4.2.1 Contaminatednorm~l We will refer in more detail, later, to the results of Kochlerlakota et ai. (1992). First, we will describe an approach used by ourselves (Kotz and Johnson (1993)) to derive moments of Cp for a more. general contamination model, with k components (but with each component having the same variance, (12), with PDF
k

" ,.

(wl.r
(71., nll.,
I~

"

j= t

L (XJ-~O)2

(4.6)

(where ~o is an arbitrary, fixed number), given N = n, is that of (12x [noncentral chi-squared with n degrees of freedom and

noncentrality
k

J=1

PjCP(x; (1) ~j,

(4.4)

fl

parameter

A~

L ni~j-'o)2]-symbolically,
j= 1
","-'-"." . . ::, ',112 (",", ' '~~> . '.\,.: - '. ;, . ;' . ... '. ~.,. . ,/''," "'" ~,.,

(12X;,\.t~)
(4.7)

1-

142

Process capahility indices under non-normality

1\

The conditional distribution of

n ,- 1

Effects of non-normality
8~

143

and in particular E[C1,IN= n] = C,,(n--1)2 exp( -'2)'n)

CfJ

(n-1)82=

L (Xj-X)2 j= 1
(4.8) , J.8)

is that of
0"2X~2-1 (An)

,~o

b::()

n
L'

)
1-/-

(4.11a)

'

J2r(~;l
E[C~IN =nJ = C~(n-l) exp(-.tAn)
x CfJ(Hn)l

+)
.

with
\~

An=

j~ 1

L nj((j-~1)2

where

~I=n j

L nj(j
I

1=0---:-;- n- 3 + 2. l. 1

(4.qI b)

Expressions for moments of noncentral chi-squared distributions were presented in section 1.5 and have been used .in Chapters 2 and 3. We will now use these expressions to derive formulas for the moments of d CP=38 Conditional on N = n, Cp is distributed as
I

A';~-

i 10
~ve
,

Averaging over the distribution of N, we obtain

I!I/'

E[C~J=L
n

~
( j= 1
TInj!

.iI P?

)(

J=1

)E[C~lnJ

1'=1,2,...

(4.12)

, l~

The summation is over all nj~O, constrained by

k

!. d(n -1)% !x~3

0"

1(An) = Cp(n -1)i /x~ - 1(An)

(4.9)

t()

j~ 1

L nj=n.

Recalling the formula for moments of noncentral chi-squared from Chapter 1, we have

~~rcd
I'd

E[C~IN =nJ = C~(n-1)ir exp(-lAn)

'

In the special symmetric case of k=3, with ~dO"=-'a, ~2= 0, ~3/0" a, (a> 0) and PI = P3= P, say, the noncentrality = parameter is
1
An=

, i= 0

I {(t~n)i} rC; z!

J+i-~ n-1 2 2r/2 r( -2+i

{ nl +n3--~(n3-nd

}a

(4.13)

(4.10)
'I"~ h I

LIO) -.
14I

Some values of E[CpJ/Cp and S.D.(C\)/Cp, calculated from (4.11a) and (4.11 b), are shown in Table 4.2 (based on Kotz

Effects of non-normality

145

0
q
<

.cr.)

\0 r-"<t M\ON OOMr-"<tM""" N N"""""""""-

r "<too\OM OO\Or-Ir\MN NC"I--"""-

000,000
V)~;:,"<tO\or"<to\O\"<tOOM 0000\000000 "":00000
"<tOooO\MN O\OOr-\OO\M NN--"""-

000000
O\\Ooo "<tr-O'IMoo00\0'10\0\0\ "":00000

" II

Iu.J
0

"<tOO

(I

"<toooo"<tO\\O O\oor-r-MM

000000 II", 0 -Ivi I u

. u.J
--O\N"<tOO ooMNOO-\o 0000\00\ "":"":"":0"":0 \OMO\r-Ooo 0\0\ r-r-"<t M NN ..................

NN---900000
N-OO',...\o

II,

t:: <1) ..c

<U'"

OO\oM""""""o\ 000000\ "":"":"":"":"":0

\0 V) 0\ O\O\r-r-"<tM NN-""""""""" 0'1 0 0\

and Johnson (1992)) for selected values of a and p. (Hung and Hagen (1992) have constructed a computer program using GAUSS for calculation of these quantities.) , The table is subdivided into two parts. In both parts Cp(=1d/0')=1. In Table 4.2a the cases k=3, n= nt +n2 +n3= 10,20, and 30; Pt =P3 =0.05, and 0.25; a = 0.005, 0.01, 0.125, 0.25, 0.5, and 1.0 are presented. Table 4.2 b covers situations when k = 2 for the same values of n and a, with p=O.l, and 0.5.. From the tables we note that
,
I

'0

Iq
V)

000000

000000
"""\OO\"<t"<tO\ O\ooMMN000000

cr.) '-' ;::: .8 ..... e<j ';;

<1) "0 "0
"0 ..... CIJ
"0

0'

1u.J

qqqqqq

Noooo\O"<t...... O\r-MNN......
......

V)I N
......

000000

r-\OoO\oo 0\0\00r-"<t"<t N N""""""""""""

r-r-OO\OO 0\0\00r-"<t"<t N N""""""""""""

000000
MN"""O\O1r\ O\O\"<t"<tNN 000000
: : : : : : : : :

MO"""oo\OM O\O'I"<tMNN

000000
: :

1. for given n, the bias and S.D. both decrease as a increases; 2. for given a, the bias and S.D. both decrease as n increases; 3. for given a and n, the bias and S.D. do not vary greatly with Pt and P2 (when k=3) or with P (when k=2). The bias is numerically greater when the 'contaminating' par. ameters are smaller; 4. on the other hand, the greater the value of a, the more marked is the variation (such as it is) with respect to Pt and P3 (or p). Observe, also, that the bias is positive when a is small, and decreases as the proportion of contamination increases. HQwever, the bias is negative for larger values of a and p, and it becomes more pronounced as n increases. This is
contrary
.I

;; ;;
;>

r-r-oooo

G3' 0 '-' <1) 0 ;:j

10 ...... cr.)
1u.J

0\0\0000"<t"<t C"IN........................

r-r-OOOO 0\0\ 00 oo"<t"<t

N N

000000

- - -'...... 000000
"<t"<tNNr-r: : : : :

(I' ~-t II,

"<t"<tNNr-r: : :

~~~~~~
: :

~~~~~~
: :

nll~

w4

to the situation

when p = 0 (no contamination)

in

13 t)

<1)

~
;:j .....

c..

0
I

. ~
II
,.,

0'I0\0000"<t"<t N N """""""""

r-r-oooo

W
r-r-OOOO .0'I0\0000"<t"<t 000000

~ I-<

~
0

V)

000000
000000
: :

C'!f'j~~~~

/I

w
,

>< '8 ~

~ I 'V. ....
"J'
II

"<t"<tNNOOOO O\Cl\"<t"<tNN
: : : :

qqqqqq -......-.
b ~
I~ II

"<t"<tC"lC"lr-rO\O\"<t"<tNN

th'l"

1'1,,' (I II

Z
~ .!! .c =
Eo-<

II
I

...

II

>lJ>,.,ONONON 11::>..000000
II

V)

Ir\ Ir\ Ir\ V). or,

>lJ>1::>.._V)"""V)-1r\ I 000000
.... ."J>

mill

I ,1

which the bias of Cp is always positive, though decreasing as n increases. Gunter (1989) also observed a negative bias, when the contaminating variable has a larger vflriance than' the main one. . Kocherlakola et al. (1992) provide more detailed tables for the case k = 2. They also derived the moments of Cpu= (USL - X)/(38). The distribution of Cpu is essentially a mixture of doubly noncentral t distributions; see section 1.7.

>lJ> r'1"

... I::>..

~ ,.

II
0

N
0 N
0 M

= :.:-

II ~ .. 0 .c:.:-

0 N

0 M

146

Process capahility indices under non-normality

Effects of non-normality
III!

147

4.2.2 Edgeworth series distributions The use of Edgeworth expansions to represent mild deviations from normality has been quite I~lshionable in recent years. There is a voluminous literature on this topic, including Subrahmaniam* (1966, 1968a, b) who has been a pioneer in this field. . It has to be kept in mind that there are quite severe limitations on the allowable values of .j7f;and fJ2to ensure that the PDF (4.2) remains positive for all x; see, e.g. Johnson and Kotz (1970, p.l8), and for a more detailed analysis, Balitskaya and Zolotuhina (1988)). Kocherlakota et al. (t 992) show that for the process dislribution (4.2) the expt:<.;(t:d va/lit:
of Cp is

KocherIakota et al. (1992) also show that the variance of Cp isi

n -1 2 n(n'-3)Cp

_I.
lint ~I. IfQI'

I 2 1 1 { { [ 1+~ 1+g"<fiz-3)-9h,,fJu ] - E[Cp]}

~

(4.14 c)

Table 4.3 is based on Table 8 of Kocherlakota et al. (1992).

As is to be expected from section 1.1, for given fJz the values~ E[CpJjCp do not vary with j7f;, and the values of of S.D.(Cp)jCp vary only slig~htly with ft. Also, for

given J7I!, the values of S.D.(Cp)jCp increase with fJ2'

Kocherlakota et at. (1992) carried out a similar investigation for the distribution of the natural estimator of Cpu (=!(lISL-~)j(j),
iI!

(n-1)Jr(~(n-2))
fir(i(n-I)j

C,,{t +~y"(fJ2-3)--~h,,fil}
(4.14")
~(/)

Cpu= I
,
.

USL.-X 3S

(4.15)

E[C.]

where

Table 4.3 Expeltcd vailicand standard deviation of ,Cp/Cp for Edgeworth process distributions

n-1
g,,= n(n+1)

n=
10

;
0.0
:1:0.4

/32

E[Cp]/Cp

S.D. (Cp)/Cp

and

3.0 4.0 5.0

3.0

h-

(n-2)
30
Ia-

"-: n(n + l)(n + 3) The premultiplier of Cp will be recognized as the bias correlation factor h"- I, defined in (2.16), so (4.t 4 a) can be written

0.0 :1:0.4

3 J I E[C,,] = E[Cplnormal] {J + /J?!,J/J23)-, Hh,,/J1 (4./4h) -i

...

I
'1"1 Ib) II 1"'f1he

4.0 5.0 3.0 4.0 5.0 3.0 4.0 5.0

1.094 1.128 1.161 1.094 1.128 1.161 1.027 1.039 1.051 1.027 1.039 1.051

0.297 0.347 O.3R9 0.300 0.349 0.392 0.140 0.169 0.193 0.141 0.170 0.194

K. Subrahmaniam was the (pen) name used by K. Kocherlakota in the 1960s.

If the process is symmetrical (so that fJl =0), E[CpuJ is proportional to Cpu' Table 4.4 (based on Kocherlakota et al.

148

Process capability indices under nOI1-normalit y

Effects of non-normality

149 Skewness (J f31 ) 0 0 0.506 -0.506 0.667 1.000 1.333 2.006 2.309 2.828 3.163 3.651 4.000

Table 4.4 Expected value and standard deviation of Cpu for Edgeworth process distributions
n 10

II. for,

Jpl

E[Cpu]
P2 4.0 5.0 3.0 4.0 5.0 3.0 4.0 5.0 3.0 4.0 5.0 3.0 4.0 5.0 3.0 4.0 5.0 Cpu= 1.0 1.086 1.119 1.153 1.094 1.128 1.161 1.100 1.134 1.107 1.024 1.036 1.048 1.027 1.039 1.051 1.029 1.041 1.053 Cpu= 1.5 1.632 1.682 1.733 1.641 1.592 1.742 1.647 1.697 1.748 1.538 1.556 1.574 1.540 1.55H 1.576 1.542 1.560 1.578

- 0.4 3.0
0.0

0.4

30

-0.4

0.0

0.4

S.D.(Cpu) Cpu= 1.0 Cpu= 1.5 0.287 0.426 0.337 0.502 0:381 0.569 0.320 0.433 0.366 0.533 0.407 0.596 0.345 0.486 0.388 0.554 0.427 0.615 0.136 0.201 0.166 0.246 0.190 0.283 0.154 0.220 O.IHO 0.261 0.203 0.297 0.168 0.235 0.193 0.274 0.214 0.308

I"~
. I

r' II 1\ II II,
III;
'11

Distribution number and name [I] Normal (50,1) [2J Uniform (48.268,51.732) [3J 10 x Beta (4.4375,13.3125)+47.5 [4J 10 x Beta (13.3125,4.4375)+42.5 [5J Gamma (9,3)+47 [6] Gamma (4,2)+48 [7J (xl. 5) Gamma (2.25, 1.5)+ 48.5 [8J (Exponential) Gamma (1,1)+49 [9] Gamma (0.75,0.867)+49.1340 [10] Gamma (0.5,0.707)+49.2929 [l1J Gamma (0.4,0.6325) + 49.3675 [12J Gamma (0.3,0.5477) + 49.4523 [13J Gamma (0.25,0.5)+49.5

I'll
111

WI

(1992)) shows that even when /31is not zero, E[CpuJ is very nearly proportional to CpuoOf course, if /31is large, it wil1not be possible to represent the distribution in the Edgeworth form (4.2). Table 4.3 exhibits reassuring evidence of robustness of Cp to mild skewness.

(II h ~ry
~Ol

In all cases; the expected value of the process distribution is 50, and the standard deviation is 1. The values shown in the tables below are based on computer output kindly provided by Drs Barbara and Kelly Price of Wayne State University in 1992. The specification limits are shown in these tables, together with the appropriate values of Cp and Cpk' The symbols (M), (L) and (R) indicate that T =--=, >, < 1(LSL+ USL) respectively. In all cases, T = 50. (M) (L) (R) (M). LSL 48.5 45.5 48.5 47.0 USL 51.5 51.5 54.5 53.0 Cp 0.5 0.5 0.5 1.0 Cpk 0.5 1.0 1.0 1.0 (L) (R) (M) (L) (R) LSL 44.0 47.0 44.0 41.0 44.0 USL 53.0 56.0 56.0 56.0 59.0 Cp 1.0 1.0 2.0 2.0 2.0 Cpk 1.5 1.5 2.0 2.5 2.5

itlh I'llIt

I~

,
.

4.2.3 Misccllancous Price and Price (1992) present values, estimated by simulation, of the expected values of Cp and CPk> from the foHowing process distributions, numbered [lJ to [13J

II, g
Ihl

The value of E[CpjCp] does not depend on ~, so we do 11Ot need to distinguish (M), (L) and (R), nor does it depend on Cpo Hence the simple Table 4.5 suffices.In this table, the distributions are arranged in increasing order of I~I.

150

Process capahility indices under non-normality

O'

Construction of robust PCls

151

Table 4.5 Simulated values of E[Cp/CpJ Distribution [lJ [2J [3J [4J [5J [6J [7J [8J [9J [10] [11J [12J [13] (normal) (Uniform) (Beta) (Beta)} (Gamma) (Gamma) (Gamma) (Gamma) (Gamma) (Gamma) (Gamma) (Gamma) (Gamma) ' n= 10 1.1183 1.0420 1.1171 1.1044 1.1155 1.1527 1.2714 1.3478 1.5795 1.6220 1.8792 2.2152 n=30 1.0318 1.0070 1.0377 1.027 ' 1.0371 1.04"/4 1.0801 1.1155 1.1715 1.2051 1.2595 1.2869 n= 100 1.0128 1.0017 1.0137 1.0091 1.0143 1.0146 1.0242 1.0449 1.0449 1.0664 1.0850 1.0966

Hn
HI'

al~ HIiJ a~.

fn
0

8h till
PI', III'
111 10 IOns
11.1

These values do not depend on '/: The values for the two beta distributions have been combined, as they should be the same.

I.hl

Comparing the estimates for the normal distribution ([lJ) with the correct values (1.0942, 1.0268, 1.0077 for 11= 10,30, 100 respectively) we see that the estimates are in excess

st~11J ) (/J ~== llilficss Ya, 0111.of

by about 2% for n = 10,t% for n= 30or 100.

00

Sampling variation is also evident in the progression of values for the nine gamma distributions, especially in the n = 100 values for distributions [6J and [7J, and [9J and [10]. As skewness increases, so does E[CpjCp], reaching quite remarkable values for high values of y1J;. These howeyer, correspond to quite remarkable (and, we hope and surmise, rare) forms of process distributions, having even higher values of y/f3; than does the exponeI2:tialdistribution. For moderate skewness, the values of E[CpjCpJ are quite close to those of the normal distribution (d. Table 4.3). Table 4.6 presents values of E[CpkJ estimated from simulation for a number of cases, selected from those presented by
Price and Price (1992). 0

Illhe MlllfOJ. III,lIe Ih,~w1I\\.,nd hll~cn ton. fllte

of Cpkas an estimator of Cpk- fordistribution[13] whenn= 10, and only relatively smaller biases when n=30 and n= 100. For the exponential distribution [8J there is a quite substantial positive bias in Cpk' The bias is larger when ( is greater. than l(LSL+ USL) - case (L) - than when it is smaller"-~ case oCR).The greater among these biases for exponential, in Table 4.6, are of order 25-35% (when n= 10), falling to 21-5% when n= 100. As for Cp, the results for the extreme distribution [13J are sharply discrepant from those for normal, and mildly skew distributions. Of course, [13J is included only to exhibit the possibility of such remarkable biases, not to imply that they are al1ything like everyday occurrences. Coming to the variability of the estimators Cp and Cpk>we note that the standard deviation of Cp might be expected to be approximately proportional to J({32-1) where f32( J14/(J4)is the kurtosis shape factor (see section 1.1) for = the process distribution. We would therefore expect lower standard deviations for uniform process distributions (f32 = 1.8) than for normal (f32 = 3) and higher standard deviations when f32> 3 (e.g. for the exponential [8J with f32= 9). Values of J(f32 -1) are included in Table 4.7, and the estimated standard deviations support these conjectures. It should be realized that Tables 4.5-7, being based on simulations, can give only a broad global picture of variation in bias and standard de:viation of Cp and Cpkwith changes in the shape of the process distribution. More precise values await completion of relevant mathematical analyses which we
hope interested readers might undertake. 0'

4.3 CONSTRUCTION

OF ROBUST PCls

UaThtfby
ll'llt nl1l4ll11e

We again note the extremly high positive bias - this time

The PCls described in this section are not completely distribution-free, but are intended to reduce the effects of nonnormality.

-~-

-~~~-

---

,--_.

Table 4.6 Values of E[Cpk] from simulation (Price and Price (1992
Cpk

min(-LSL, d 1

USL-'

Process distribution* [1] [2] [3][4] [7] [8] [13] [l]LR [2]LR [3]L[ 4]R/[3]R[ 4]L [7] L/[7] R [8]L/[8]R [13]L/[13]R [1] [2] [3] [4] [7] [8]
[13]

n=lO 0.464 0.424 0.464 0.480 0.527 0.904 0.559 0.521 . 0.569/0.548 0.597/0.555 0.671/0.600 1.259/0.956 1.023 0.945 1.022
-

n=30

n= 100 0.480 0.474 0.481 0.480 0.485. 0.519 0.506 0.501 0.513/0.504 0.510/0.504 0.515/0.509 0.557/0.540 0.986 0.975 0.988 0.987
. 0.997-

0.5

1/2

1.0

0.468 0.453 0.471 0.474 0.487 0.582 0.516 0.504 0.525/0.509 0.528/0.519 0.548/0.532 0.670/0.617 0.985 0.957 0.989 0.998 1.057 . 1.028 '1.163
2.012 1.226 ~
1.226

1.067 -"'-' i I ! i

"iIiio

-..;

- ""'1"

;:'

.::c.-""

[13]

- -----.
-

'!a... ~-~~

"" ,j

"""'"

..

~.

~,,"~;. ~

"" JijJ

z~
-- -

1.007

---

--

-,

.--

. _.

..

--="

'-' .J

2/3 -

[l]LR

2.0

4/5

[2]LR [3]L[ 4]R/[3]R[ 4]L [7]L/[7]R [8]L/[8]R [13]L/[13]R [1] [2] [3][4] [7J [8J [13J [lJLR [2JLR [3JL[ 4JR/[3JR[ 4JL [7] L/[7J R [8JL/[8JR [13JL/[13JR
-

H18 1.042 1.129/1.111 1.174/1.132 1.307/1236 2.337/2.064 2.141 1.987 2139 2209 2.435 4.227 2.237 2.084 2.245/2.233 2.326/2.284 2.578/2.507 4.582/4279

1.032 1.007 1.043/1.032 1.052/1.043 1.088/1.072 1.J13/1.261 2.016 1.964 2.027 2.045 2.108 2.513 2.064 2.014 2.081/2.070 2.099/2.090 2.168/2.153 2.600/2.548

1.013 1.002 1.017/1.011 1.017/1.012 1.027/1.021 1.105/1.088 1.999 1.976 1.998 2.002 2.021 2.164 2.026 2.003 2.030/2.024 2.032/2.026 2.052/2.046 2.201/2.185

*Notes:

(a) When min(~-LSL, USL-Q/d=l, we have ~=1(LSL+USL1 so only (M) applies. (b) Since [3] and [4] are mirror images, results can be merged as shown. (c) For symmetrical-distributions (normal [1] and uniform [2]), the L and R values should be the same and so they have been.averaged.

----

Table 4.7 Estimates of S.D. (Cp) and S.D. (Cpd from simulation (Price and Price (1992 Distribution Normal [1J Uniform [2J
xi.s [7J

(Jh-1)! . 1.414 0.894

2.160

sgng-1(LSL+USL)} O(M) 1(L), -l(R) } O(M) l(L), -l(R) }

O(M) l(L)

n=30 n=l00 S.D.(CpfCp) S.D.(Cpk) S.D.(CpfCp) S.D.(CpJ 0.077 0.148 0075 0148 } { { 0.160} . . 0.082 0.047 0045 0089 SO.088 } { } . lO.110 .' 0.056 0.109 0.192 0,199

-l(R) Exponential [8J 2.828 O(M) l(L) l(R)

to process distributions

}
0.273

{ 0.245 } 0.~66
0.258
0.237

0.111

{ 0.137}
0.092 0.139

}
with Cpk = 1.

{ 0.255'}

0.142

{ 0.169}
0.119

Values for S.D.(Cpi;) correspond

-I sgn(h)=

for h<O 0 for h=O

{ ,... I

for h> 0

--. . Table

-4.8 0.135% and

99..8.65D/. poiD1s ol:

- ~'~F

~'-

---

p~

~~

--1.6

--;- ,>'

Table 4.8 0.135% and 99.865% points of standardized Pearson curves with positive skewness (vip v7f;. <0, interchange 0.135% and 99.865% points and reverse signs. {Clements (1989

> 0). If

Ji;
/32

0.0 -1.727 1.727 -2.210 2.210 - 3.000 3.000 -3.000 3.000 - 3.261 3.261 - 3.458 3.458 - 3.611 3.611 -3.731 3.731 - 3.828 3.828

0.2 -1.496 1.871 - 1.912 2.400 - 2.535 2.869 - 2689 3.224 - 2.952 3.484 -3.118 3.678 -3.218 3.724 - 3.282 3.942 - 3.325 4.034

0.4 -1.230 1.896 -1.555 2.454 -1.930 2.969 - 2.289 3.358 - 2.589 3.639 -2.821 3.844 -2.983 3.997 - 3.092 4.115 -3.167 4.208

0.6 -0.975 1.803 -1.212 2.349 -1.496 2.926 -1.817 3.385 - 2.127 3.175 - 2.396 3.951 -2.616 4.124 -2.787 4.253 -2.91 4.354
-

0.8 -0.747 1.636 -0.927 2.108 - 1.125 - 2.699 -1.356 3.259 -1.619 3.681 - 1.887 3.981 - 2.132 4.194 - 2.345 4.351 -02.524 4.468

1.0

1.2

1.4

1.8

1.8

2.2 2.6 3.0 3.4 3.8 4.2 4.6 5.0

- 0.692 1.822 -0.841 2.314 - 1.000 2.914 - 1.178 3.468 -1.381 3.883 -1.602 4.177 -1.821 4.386 - 2.023 4.539 -0.616 1.928 -0.739 2.405 -0.865 2.993 -1.000 3.861 -1.149 3.496 -1.316 4.311 - 1.494 4.532 -0.531 1.960 -0.634 2.398 -0.736 2.945 -0.840 3.529 -0.950 4.015 - 1.068 4.372 ~0.533 '2.322 -0.617 2.798 - 0.701 3.364 -0.785 3.907 -0.510 2.609 - 0.580 3.095

For each (J Ph P2) combination, the upper rmv.contains-9:-!-3-5%,-puiub(O;T1fmillierower, 99:8'65)'0' points (0.).

Construction of robust PCl s

159

,v
~ ..0
l

-~

Moment esti'mators for (4.17)and (4.18) are obtained hy

.

replacing (Jby an appropriate estimator.

g '-'
1

88~:2~
1

d<""i<""i..f'l-;

.~ R ::;j
'"
1-<
c.8

MO-'<tV)

f6

0...

~~~~~ OMM'<tV)
~"<I:~~~
r-OO'<tLn

II
<:I:>I M ~I'"

'b'

OMM'<tv)

HJ'

OOO\Mv)
'<t 1.0 00

VI IOI N ~ VI
b
I

1.0

d<""i<""i'<iV'i
M 0 00 V)00001.O1

V)

V)

~ - d<""i<""i'<iV'i I 'v' I r- 0 r- N V) '--' ("0. 1.0 0 00 \0 I-< ,....

~
I
I

i,

'<tM'<t

V)

...... cod ...c::

1

...c::

OOV)NV) ooNoo\Od..f<""i..fV'i 00'<t-V) OV)oo\O""<'<i<""i'<iV'i MONO\O 1"1000\0""<v-i<""i'<iv-i

u ::;j
'"

<:1:>100 ...... 0

~ > oi:i 11)

I-<

:::I 1 \0

I ,

g
'r .::: .U ~
.0'

.g i!)
1:1

i!)

V) 00

~
~
I-<

0\0\0\ d d d
0...
I-<

0\

I
I

.
d
"I

II II II
0...
I-<

I I

'<t

~ ~

c...
I-<

I I
I I L

~ ",c.8c.8c.8 <::0.<::0.<:1:><:1:><:1:>

, "

At this point we reiterate the difference between Clements' method and the Johnson-Kotz-Pearn method. In the first method there is an attempt to make a direct allowance for the values of the skewness and kurtosis coefficients, while the second method aims at getting limits which are insensitive to these values. In the second method we no longer have guaranteed equal tail probabilities, but we do not have to estimate .jlf; and {32which it may be difficult to achieve with accuracy, because of bse of third and fourth sample moments, which are subject to large fluctuations. Both methods rely on the assumption that the population distribution has a unimodal shape close to a Pearson distribution for Clements' method,and more restrictively, close to a gamma distribution for Johnson-;Kotz-Pearn method. Another approach, also aimed at enhancing connection between PCI values and expected proportions NC, tries to 'correct' the PCI, so that the corrected value corresponds (at least approximately) to what would be the value for a normal process distribution with the same expected proportion NC. Munechika (1986, 1992) utilized an approximate relation het\vccncorresponding percentiles of standardized normal and Cham-Charlier (Edgeworth) distribution (see Johnson and,Kotz (1970, p. 34)) to obtain an approximate relationship between the process PCI and the corrected (equiv~Jent normal) pel values. He appHed this only to the case where there is only an upper specification limit (110LSL), with the CpkU index (see equation (2.25b)).The approximation is quite good for gamma process distributions which are not too skew. From the relationship (in an obvious notation) XGt~ Ua+ i(U; -1)J7f; he obtained j(process index) ~ 3(corrected index) + i {(corrected index)2-1 } leading to (corrected index~(1/3..JP~)[{j31 + 18(process index)~ + 9}1/2- 3J. The inverse (also approximate) relationship UGt~XGti(x;: 1)J7f; would give the somewhat simpler formula
~

160

Process capahility indices under non-normality

-,I
.

Construction of robust PCI s

161

(corrected index)~ (processindex)-

fa{9(process index)2~ 1~

The bases for their choices are that for a normal distribution the interval (~--2(1, ~+2(1) oflength 4(1contains 95.46% of
..values, and

(This formula was not used in Munechika (1986, 1992).) Of course, use of this correction requires a value for -Jf31' Unless this is well established, it is necessary to estimate it Such estimation may well be subject to quite large sampling

.
I 1
~

variability.
4.3.3 'Di,stribution-free' PCls

the interval (~- (1,~ + (1) of length 2(1 contains 68.26% of values (and, of course, 6(1=1 x 4(1= 3 x 2(1).

It is here that we must part company with them, as it seemsunreasonable to use relationships based on normal distributions to estimate values which are supposed to be distribution-free!

Chan et at. (19~H~) proposed the following method or Qbtaining 'distribution-free' PCI s. They note that the estimator, 8, in the denominator of Cp is used solely to cstimtltc the length (6(1)of the interval covering 99.73% of the values of X (on the twin assumptions of normality and perfect centring, at ~=~(LSL + USL)). They propose using distribution-free tolerance intervals, as defined, for example, in Guenther (1985) (not to be confused with Gunter (1989)1) to estimate this length. These tolerance intervals (widely used in statistical methodology for the last 50 years) are. designed' to include at least 100f3% of a distribution with preassigned probability 100(1-1J()%, for given f3 (usually close to 1) and IJ((usually close to zero). In the PCI framework, a natural choice for f3 would be f3= 0.9973,with, perhaps, 1J(=0.05. Unfortunately construction of such intervals would require prohibitively large samples (of size 1000 or more). Chan et at. (1988) suggest that this difficulty can be overcome in the following way. Construction of tolerance intervals
with smaller values of f3(but still with IJ( 0.05) requires smaller =

4.3.4 Bootstrap methods Franklin and Wasserman (t 991), together with Price and Price (1992) should be regarded as the pioneers of application of bootstrap methodology in estimation of Cpk' The bootstrap method was introduced some twelve years ago (see Efron, 1982) and has achieved remarkably rapid acceptance among statistical practitioners since then. (Over 600 papers on the bootstrap method were published in the period 1979, 90!) It is not until very recently, however, that its application in the field of PCls has been developed. . The bootstrap method is a technique whereby an estimate of

I t

.
\
I. I, ~

,. i

~ . .

samples (lessthan 300).They recommend taking

.
I

the.distribution functionofa statisticbased on a samplesizen, say, is obtained from data in a random sample,of sizem (~n) say, by 're-sampling' samples of size n - with replacement from these m values and calculatingthe correspondingvalues of the statistic in question. Usually m= n,but thisneednot be the case. Here is a brief formal description of the method. Given a sample of size m with sample values XI,X2,"',X", we choose (with replacement) a random sample ([1], say)
.x[IJ1>"" X[IJII of size n, and calculate
C(1JPk

. f3= 0.9546 and place of 68, or . f3= 0.6826 and

place of 68.

from this new

using! x (length of tolerance interval) in

I

'sample'. This is repeated many (g) times and we obtain a set

~
1

using 3 x (length of tolerance interval) in

of values C[I]Pk,C[2]pko""C[9Jpk' which we regard as approximating the distribution of Cpk in samples of size n - this estimate is the bootstrap distribution. (The theoretical basis

162

Process capahility indices under non-normality

Flexible PCls 4.3.4c The bias-corrected percemile confidence interval

163

of this method is that we use the empirical cumulative distribution from the first sample - assigning probability m - 1 to each va:lue - as an approximation to the true CDF of X.) Practice has indicated that a minimum of 1000 bootstrap samples are needed for a reliable cakulation of bootstrap confidence intervals for Cpk'

itIJ

According to Hall (1992) a leading expert on bootstrap

~

methodology - difficulties 'in applying the bootstrap to process capability indices is that these indices are ratios of random variables, with a significant amount of variability in the variable in the denominator'. (Similar situations exist in regard to estimation of a correlation coefficient, and of the ratio of two expected values. The bootstrap performs quite poorly in these two (better-known) contexts.) Franklin and Wasserman (1991) distinguish three types of bootstrap confidence intervals for Cpk, discussed below.

This is intended to produce a shorter confidence interval by allowing for the skewness of the distribution of Cpk' Guenther (1985) pointed out the possibility of doing this in the general case, and Efron (1982) developed a method applicable in bootstrapping situations. The first step is to locate the observed Cpkin the bootstrap' order statistics Cpk(l)~... ~ Cpk(lOOO).For example, if we bootstrapped values we find

have

Cpk = 1.42 from

the original data, and among the,

Cpk(366)= 1.43

Cpk(365)= 1.41 and

\

I!

t '
: i
,

then we estimate
A

365
1,000= 0.365 = Po, say

- Pr[C11k 1.41] as ~
4.3.4a The 'standard' confidence interval

One thousand bbotstrap samples are obtained by re-sampling from the observed values XI, X 2"", X n (with replacement). The arithmetic mean C~k and standard deviation S*(Cpk) of the CpkSare calculated. The 100(1-a)% confidence interval ~~~~ (C~k - ZI -a/2S*(Cpk)' C~k + Zl-a/2S *(Cpk)) where <1>(ZI-a/2)= 1-a/2. (4.19)

"
""

We then calculate <1>-I(po)=zo (i.e. <1>(zo)=Po)' our In

example, Zo= <1> 1 (0.365) = - 0.345. We next calculate PI =<1>(2z0-Z1-a/2) and

r
u'

Pu = <1>(2zo

+ZI-a/2)

-Ii

I'
II

i ,

(in our case, with a=0.05; ZI-a/2 =<1>-1(0.975)=1.960, PI= 11>( 2.650)= 0.004; Pu= <1>( 1.270)== 0.898, and form the confidence interval (Cpd1000PI)' Cpk(1000pu))'

It, 4.3.4b The percentile confidence interval

Ii

The 1000 CpkS are ordered as Cpk(1)~Cpd2)~...~ Cpk(1000). The confidence interval is then (Cpk500a), Cpk500(1-cx))), where < ) denotes 'nearcst integer to'.

1,,1 MI-I

II

In our example, this is (Cpk(4),Cpk(898)). The rationale of this method is that since the observed Cpk is not at the median of the bootstrap distribution (in our case, below ~he median) the confidence limits should be adjusted approximately (in our case, lowered, because Zo<0). Evidently, the method can be applied to other PCls.

164

Process capability indices under non-normality

Flexible PCl s (J2 and expected value l' we would have EX>T[(X - 1')2J =EX<T[(X - 1')2J =1(J2
~'

165

Schenkler (1985) has presented results casting doubt on the efficiency of the 'percentile" and 'bias-percentile' methods above. Hall (1992) suggests that the method of bootstrap iteration described by Hall and Martin (1988) might be more useful in this case. Hall, et al. (1989) describe an application of this method to estimation of correlation coefficients, but, to the best of our knowledge, it has not as y~t been tested. for estimation of PCls.

(4.22)

Finally they define

Cjkp=min (CUjkp, CLjkp)= 3)2 mID {EX>T[(X -1')2]}!' 1'-LSL

USL-

l'

4.4 FLEXIBLE PCls

i.

{EX<T[(X

1')2]}!

Johnson et ai. (1992) have introduced a 'flexible' pct, taking into account possible differences in variability above and below the target value, 1'. They define one-sided PCls 1
I

(4.23)

Note that if we have 1'=!(USL+LSL)=m, 1'-LSL=d, then d

so that USL- 1'=

USL - l' (4.20a)

CUjkp= 3)2 {EX>T[(X -1')2]}! and

1
CLjkp

Cjkp= 3)2max([Ex>T[(X

-1')2], EX<T[(X - 1'f]r! (4.24)

t.
l'

- LSL
1')2]}!

= 3)2 {Ex<T[(X -

(4.20b)
I

4.4.1 Estimation of C]kp A natural estimator

~

of Cjkp is
1.

where (as in section 3.5)

EX>T[(X - 1')2] =E[(X and Ex< T[(X - 1')2] = E[(X - 1')21X< 1']Pr[X < 1'] (4.21b)
I

1')2IX> 1']Pr[X> 1'] (4.21a)

Cjkp

= 3)2

mID

USL- l' 1'-LSL (8 + In)! ' (8- In)!

(4.25a)

where
8+= Note that
.

Xi>T

(X/-1'f

and 8_=

X,<T

(X1-1')2

(It is assumed that Pr[X

==

T] =0). Note that the radon;

'.

Pr[X> 1'], Pr[X < 1') make allowance for how often positive 'and negative deviations from l' occur. The multiplier 1/(3)2), while the earlier PCls use 1. aris~s from the fact that for a symmetricaldistribution with variance

E[S +]

= nE[(X .:..7')21X > 1']Pr[X> 1']

so that S+ In (notS+ I[number of XIs greater than T]) is a.n

166
I,

Process

capability indices under non-normality

Flexible PCls

167

unbiased estimator of (4.21a) (and S- jn) is an unbiased

estimator of (4.21b). The es.timator CjkP can be calculated quite straightforwardly - the only special procedure needed is separate calculation of sums of squares for Xi> T and Xi< T, as for Boyles' modified Cpm index, described in section 3.5. Our analysis will be based on the very reasonable condition LSL< T<USL. We can express CjkPas
A

distribution, consideration of this case can provide an initial point of reference. Later we will indicate ways in which the results can be extended to somewhat broader situations though not as broad as we would wish. Table 4.10 presents :71Umerical values of E[CjkPjCjkP] and S.D. (CjkPjCjkp) (assuming normal process distribution) for several values of (USL-T)jd, and n=1O(5)50. Note that if (USL-T)jd=1, then, T = (LSL + USL)j2. It is instructive to compare these values with similar quantities for CpkjCpk and CpmjCpnu in Kotz and Johnson (1992) and Pearn et al. (1992) respectively (see also Tables 2.4 and 3J). In general the estimator CjkPis biased. The bias is negative when T=1:(USL+ LSL) but increases as (USL- T)jd decrea.ses.It is quite substantial when (USL- T)jd is as small as 0.4. As might be expected the variance of CjkPdecreases as ,t increases .- it increasesas (USL- T)jd decreases,as the target value gets nearer to th6 upper specification limit.,The bias, alsb, decreases asrdncreases; fliiseffeci'Is particulariy noticeable for snulller values of (USL - T)jd. For n ~ 25, and (USL - T)jd ~ 0.60 thc stability of both E and S.D. is noteworthy. Since the index Cjkp is intended to make allowance for' asymmetry in the distribution of the process characteristic X it is of interest to learn something of the distribution of the
I

Cjkp=

d~

.
mill

USL-T

;;::;

!'

T-LSL
"

(3y 2)0'

d ~

(S(fz

d~-='

(fz

))
(4.26)

(4.25 b)
I
I

I
I

where 0' is an arbitrary constant. To study the distribution of CjkPit will be convenient to consider the statistic

t
..

'J
"

:Po I t,

'I

,I

n d z -Z -Z ' -Z D= 18;;: Cjkp=max(alS+(f ,azS-O' )

A

()

:,

"

where
I

4 al= -- d

USL~ T

-z

T -LSL and a2= d

-Z

I' I I.
I

'11

II

Note that ai"!+ai"!=2, and

I '1 !

1.1 14' III

estimator CjkP undersuchconditions. he analysisin AppenT

dix 4A can be extended to certain kinds of asymmetry of distribution of X. We note two ways in which this may be done. 1. If the population density function of X is

i
1

E[CjkP]= Ln8 (~Yrr E[J5-!r]

The distribution of jj will be, in general, quite complicated. In Appendix 4A we discuss a special case in which the distribution of X is, indeed, normal with expected value T and variance 0'2.Although this is not, in fact, an asymmetrical

l'

Hg(x; T, O'd+g( -x: - 'T,0'2)]

(4.27)

_..~ -.~

-~-

0\

'"

en :.a

;:j ..0 .~ ......

.8 ......
h 1\\ ~
H

~ 8 H

h V
~

c.
0 I-.

>. ..0 "0 <U 0

CIj

v
d'

;>.

'08"0"0

8 CIj.
s::

~ <U s:: "E~;:j~

0 0

~
<U

o~
0

+->.8

8
-

..0
:;::
......

-00
N

~ --.E
CIj
s:: CIj
CIj"O""';

oS

~
-N
I ><. I

~-~Q ~

~ 0

CIj

oS
00 0 ..... 0

s::

8 I-<
.... E:;'

0 s::
N

.8

.8

"..., -;:j ~ ..0 ..c:: 'i: ...... en

.8 ......

~:e~o
;.::::: t> 0\

..0
CIj."",

~
V
0

CIj <1>;:j

s::o- s:: .- E-<.!..

bJ)OO

s::

':i--. N

v Q.,. k:

.......
<:,,)

">

:z

..c:: II) ..c:: ...... ..... 0

<.!..

~
"--'"

-\0 N
I

:.a
0

~ s::
'"
en CIj

en 0

en

8 0..
0 ..c:: .....

-I~
NN

-I
b

.- ~ V 1) -- 6\1:t:: 0 ~ b ~ ..2 8:;vo s:: ~0000:> +-> ~ O\:I: ;> .- h :;: 2 ~ ;> ;:joo
8 ~

0'0S::q:: s::~ s::

CIj~ ..c::

s::
0 ......

~ gOE~
0
I!)

<I> CIj

b h~

eno~ <U 00 > ,",,~CIj 01)c.L. Or' <U 0 o-s::

.~

I ~
N

"0 IMN b H
"0 CIj

>:

oE:: ..... 0 0

.~ ~

~ 0 p..
II) ..c:: ......
o~

-IN

I '-y-..J 0.. X
0

~ 0 ~ 0 ..c:: . .~

b ~~

NI~ ~

II b h" >f ~

+->bo S::b"-:I: <U, "..., '=: ~-

..0 ~..c:: 'i: .-; +-> ~ II.~ - O""'N "O':<b 0--0 ~ ..c::

..... 0 (~ 0 0 0 s:: CIj s:: 'c 0 CIj .~ > ;:j ,-. 0

I --

N-

+
b

o::CIj.8~"O:: "'" ". ""'S::b::",0I)0~ 0 :: "'o~oX+->

<I>~-O""'~OO

"0 -g

8
H

O'.S
+-> 0

~ "0 -~ ~ ~
I!) ~ +->
~

CI)

..e~CIj~

s:: 0 0

-IN
N

g= ~~8 .~ ..0 ~ C ~ ~ e

II

""'c;;~+->S::os::

:.a ... ~ oo-:e

0
0

:e2r::'0~0 1= ~ ~ ';::+-> ,...s <I>01) ;:j ~ -

=-0

. s::o~'- g, ..0

-g .S 8 0 0
0 ~

+
-<

-0 ~ CIj..o s:: E <U I-<

0 o 'i: <u--Clj+-> 0..r-- 8 o~ ~C'!~oo <u~S::d
I-<.S::

~ C'! "0

.8

W~I
s:: I

0.. en <U I-<

:.: I ""0: >< "0 s::

~ I-<

h>f
15; c:..

-b

~.~

'0 ~en L.

<I>_o-fl =,.goClj:es:::e

s:: "'."0

o~

E o~ . C'i =;:j;:j; ..... -

0..0

--

CIj

...... CIj"O t$: INb ..... s:: en w~ ~..2<O 8 >( 0 ; <U 0 H-<t: _0-:: I-<~

_0

~ s::-o +-> <I><.!..s:: 0 CIj.C'i '".P ~ 0-::

h-

CIj

:.: ><

,,

0 ~

+-> ~ ,,~ CIj+->~""

.~ t:.~ "0,,00
C'i

-s:: CIj

-0

-0

I-<en=+->
s::

CIj"<f" ~ x.

~ 0 X I-< +->-0'I!) .~ ~. 8.~ ~.8 CIj ~ ~ho~~

~:.a s:: s:: 0 0 0..

EO -<t:

> 0..

----.----.---.-----.

----

'O"S

---:"=i!"""

.------

- -6LI"0 rvoOl 6LI"0 Iv(n 6LY'0 ltO'r

6LYoO IPOOI
"]-'. .OO"S "]
O'S "] O"S "] ooS "] O"S O] "O"S ']

6010 P60

zro

091"0

ooor zOor

vLI"O 60'r v81"0

VO"[

8LI"0 Ovo"r 061"0

~POOI VOZ'O r~oOI ZZZOO 6~0"1 VPZ"O 6901

6LooI
lVOOI 161"0 ~vo.t
LOZoO Z~OOI 9Zz00 0901 Z ~zoO ILOOI L8Z'0 L8001

os

8ro 891"0 ~Iro Z60 ooor vool

Z6rO Z6rO ~VO'I ~vooI

LOZ.O Z~O'I

Z61"0 z~ro ~VO'I . ~VOOI LOZ'O Z~ooY 8Zz00 0900y

~P
Ov
~ O ~Z
OZ ~Y 01 =u

Izro ~vro 06'0 ZOOOI 6zro ~ro

L'l60

9Lro 901 L81"0

961"0 8vOOl II Zoo v~O"1 !IZZ'O Z900r o~Zoo ZLO'I

LOZoO LOZ'O Z~Ooy zso'r 8'lZ'0 O9O'r

azoo
0900y

100'1 60'1

8Zz00 090'Y

6ro V9ro 66 ro PZ60 0001 vOOI "Z~ro ZZ6'0 OLro 616'0 86ro 8160

~~Z'O ~~Z'O ZLO'Y 'liO'r 6Z'0 880'Y Z~oO ~In 96Z00 6801

%Z'O 9~ZoO ZLOor 'liO'Y L6z00 6800r 89'0 Lln IJ~'O 86z00 6801 vLoO Lln 9S~'0 ~Ln

8Lro 6660 L6ro 6660

~rz'o 8Pool:
LZoO p~oor

a"o 801 'lI "0

8aoo 9801

pon

90 'lIn 'lIP"O ~n

19'0 LIn 08v'0 69rI

L'lZ'O oaoo QOOol 90'1 Z.o Z80'1

080

OZoo 69'0 ~n LOn ~6'0

6VP'0 9n

OS
O] O'S O]

Un

O~Z'O 8z00 IZ60 800'1 001 060

vn
OCO

99poO ~KoO Z'l"I 6n

1)90 O~OO

190 ZCO 9'l"I 68Z'1

OV'O O'O

OL8'0 6~n ~vn vln

0,'0 01"0

=p/lL -7Sn)

:L=~ u::>qh\d'lf.:)/(d'lf;;)O'S="O'S ptrn d'lf.:?/[d'lf.:)JH=HJOS::>I!.(LfH:t31q8.L '

170

Proccss cafJahility indices /llIdl'/' nOll-normality 4.5 ASYMPTOTIC PROPERTIES. /32= 3, so that

Asymptotic

properties

171

There are some large-sample properties of PCls which apply to a wide range of process distributions and so contribute to our knowledge of behaviour of PCls under non-normal conditions.' Utilization of some of these properties calls for knowledge of the shape factors A.l and ..14,and the need to estimate the values of these parameters can introduce substantialerrors, as we have already noted. Nevertheless, asymptotic properties, used with clear understanding of their limitations, can provide valuable insight into the nature of indices.

2 .1 .l 2 0'pk = 9 + 2 C pk

(4.31)

In the very special case when ~= m, the limiting distribution of ~(Cpk - Cpd is that of
- HI U1/+!Cpk(,82 -1)!U 2}

(4.32) I

I I

Large-scaleapproximations to the distributions of Cp,

and
Cpm

Cpk

have been studied by Chan et ai. (1990).

Since

(note that Cp= Cpk in this case) where Uland U2 are standardized bivariate normal variables, with correlation coefficient .j7J~/(,82-1)t. If the process distribution is normal the distribution is that of

C~ = (~)8
~

-!
3{ j
~

IU11+ ~C

J2

kU2
P

(4.33) normal

the asymptotic distribution of Cp can easily be approached by considering the asymptotic distribution of 8. Chan et ai. (1990) provided a rigorous proof that ~(Cp - Cp) has an asymptotic normal distribution with jmean zero and variance
I I
O'~ =

where Uland variables.

U2 are independent standardized

For Cpm, we have ..j';;(Cpm Cpm)asymptotically normally distributed with expected value zero and variance

t(/32 -1) C~

(4.29)
I

If /32=3 (as for normal process distribution), O'~=~C~. For Cpk, Chan et ai.jI990) found that if ~:m the asymptotic distribution

value zero and variance

2

of vi n(Cpk - Cpk) is normal

with expected .~

6;m=

(~-m) +jp, (7 )+:[' + (~ )

0'

~-m

1(/3 -1) C;m (4.34)

.

{1

6m

<p

O'pk= '9

+ 2Y PI Cpk+
<p

ro

'4(fJ2-1)Cpk

(4 .10)

'f

If the process distribution is normal, .j7f; = 0 and /32= 3, hence

where

<p

= 1 if ~> m, = ,-I if ~< m.

If the process distribution is normal, then

.jji;=.0

0'2 pm

--

(-a ) +!
{I+ e:m)';'

~-m

')

and

(4.35)
C;m

172

Process capability indices under non-normality

.'III

Appendix 4.A

173

and if, also, ~= m ".2

(In this case, also, Cp=Cpm)' " pm -

_1C pm Z
2

From (4.26) S
~

(4.36)
I III " I ale

'-2
+(f

D=

for 8 + 81 8-

> az al

{
APPENDIX 4.A So 15 is distribu,ted as
Z Q1HXn

az8-

(f-Z

8+ az for - < al

(4.38)

As noted in the text, it is assumed that the distribution of X is normal, with expected value T and standard deviation (f.

t' 10r

H
(1 - H

az . -l.e. al

H>- az al+aZ

With the stated assumptions, we know that:

I

Z t' az ( { az 1-H)x1l lor H<- al + az

1. the distribution of (Xi- T)2 is that of XT(f2; 2. this is also the conditional distribution of (Xc- T)z,
whether Xi>T or Xi<T;
.

The overall distribution of 15can be represented as

I'

3. the number, K, of Xs which exceed T has a binomial distribution with parameters n, 1-- denoted Bin(n, hand hence 4. given K, the conditional distributions of 8 + (f - Z and 8 - (j-z are those of xi:, X;'-K respectively, and 8+ (f-z and 8 - (f- Z are mutually independent; and also 5. the distribution of H = 8 + /(8 + + 8-) is that of xV(xi+ X;-K) which is Beta(1-K,l(n-K)) so that the density function of H is
, fH(h)={B(lK,l(n-K))}-lhlK-l(1-h)l(n-KJ-l O<h<1 (4.37)

fJ "' alH
I I

{ az(1-II)

z for H>h for H < b} XII

1\ BetaO:K,1-(n-K 1\ Bin(n,1) Ii K
[I \I I

(4.39)

t'
~

where the symbol /\ y means 'mixed with respect to Y' having the distribution that follows, and h=az/(al +a2). Without loss of generality, we assume that

,-

III

USL- T:::; - LSL. Then, for a symmetrical distribution T with mean T and variance (fz,
C

!.
, for K=I,2,...,n-l, and 11 and 8++S- arc mutually I independent; and further 6. the conditional distributions of 8 + (f - Z and 8 - (f - z, given K and H, are those of H X; and (1- H)X;, respectively.
.11
I

jkp

-~ USL-T=~~ - 3(f d 3(f r;; y al

1 C+ 1 ;:--=2

(4.40)

.and

Val

V Qz

174
Process capability indices under non-normality Bibliography

175

Hence (from (4.26) and 4.40))

CjkP

where
na1 ! 15-!

Cjkp

( )
2

(4.41)

I. i ,
,

I"
Bv(Ul,U2)=JoyUI-I(1--y)"2-ldy
and B(UloUZ)=B1(1l1,llz) (see section 1.6). In particular

and

i"

E[(~~::YJ ~ (n~1 E[15~!r] yr

Now, from (4.37), (4.38) and (4.39), noting that
E[(X;)-!r]

(4.42)

I I
\

CjkP

[,Cjkp ]

= rO:(n.-1~
2"+1r(-!n)
,,-1

[( ) +
az \

at

= {2!rr(tn)}-1r(t(n - r))
~-!r - r(t(n-r))

I
I

(~)
{ B1-b(t(n-k),1(k-1))

+ k~l B(tk, t(n-k))

+

E[D

]-

2\rnin) 2" az
x {ai\rB1-b(!(n-k),

-!r

,,-1

(~)

+ k~l 8(1k, l(n-k))

t(k-r))

I
I
1 I

(::yr Bb(1k, (n-k-r))} 1 ]

BIBLIOGRAPHY

+ a2!rBb(tk,!(n-k-r))}
whence E

+al!rJ

[( Cjkp)
CjkP

n!r:l~(n;r))
2 r(zn)

a1

Balitskaya, B.O. and Zolotuhina, L.A. (1988) On the representation of a density by an Edgeworth series, Biometrika, 75, 185187. B,\rnard, G.A. (1989) Sophisticated theory and practice in quality improvement, Phil. Trans. R. Soc., London A327, 581-9. Boyles, R.A. (1991). The Taguchi capability index, J. Qual. Technol.,

[( az )]
(~)

!r+ 1

23, 17-26.

,,-I

II.

+ k~l B(tk..Hn-k { 81-1,(z(n-k)02(k-r))

r
I I

+ (::)!rBb(1k,!(n-k-r))}]

(4.43)
.1.

Chan, L.K., Cheng, S.W. and Spiring, FA (1988).The robustness of process capability index Cp to departures from normality. In Statistical Theory and Data Analysis, II (K. Matusita, ed.), NorthHolland, Amsterdam, 223-9. Chan, L.K., Xiong, Z. and Zhang, D. (1990) On the asymptotic distributions of some process capability indices, Commun. Statist. - Theor. Meth., 19, 1-18. . Clements, J.A. (1989). Process capability calculations for non-

normal calculations,Qual.Progress,22(2),49-55.

176

Process capahility indices under non-norl'l'!ality

Bibliography

177

Efron,B. (1982) he Jackknife, the Bootstrap and Other Re-sampling T

Plans, SIAM, CBMS-NSF Monograph, 38, SIAM: Philadelphia, Pennsylvania.
'. .'

English,J.R. and Taylor, G.D. (1990)ProcessCapability AnalysisA Robustness Study, MS, Dept. Industr. Eng., University of

Arkansas, ayetteville. F

Fechner, G.T. (1897) Kollektivmasslehre, Engelmann, Leipzig. Franklin, L.A. and Wasserman, G. (1992). BQotstrap confidence interval estimates of Cpk: An Introduction, Commun. Statist.Simul. Comp.,20, 231-44' Gruska, G.F., Mirkhani, K.' and Lamberson, L.R. (1989) Nonnormal Data Analysis, Applied Computer Solutions, Inc.; St. Clare Shores, Michigan. Guenther, W.H. (1985). Two-sided distribution-free tolerance intervals and accompanying sample size problems, J. Qual. Teclmol., 17, 40-3. . Gunst, R.F. and Webster, J.T. (1978). Density functions of the bivariate chi-squared distribution, J. Statist. Compo Simul., 2, 275"';88. Gunter, B.H. (1989) The use and abuse of Cpk' 2/3, Qual. Progress, 22(3), 108-109; (5), 79-80. . Hall, P. (1992) Private communication, Hall, P. and Martin, M.A. (1988) On bootstrap resampling and iteration, Biometrika, 75, 661-71. . Hall, P., Martin, M.A. and Schucany, W.R. (1989) Better nonparametric bootstrap confidence intervals for the correlation

coefficient,J. Statist. Compo Simul,33, 161-72.

,

Hsiang, T.e. and Taguchi, G. (1985) A Tutorial on Quality Control and Assurance - The Taguchi Methods, ASA Annual Meeting,

Kocherlakota, S., Kocherlakota, K. and Kirmani, S.N.U.A. (1992) Process capability indices under non-normality. Internat. J. Math. Statist. 1. Kotz, S. and Johnson, N.L. (1993) Process capability indices for non-normal populations, Internat. J. Math. Statist. 2. Marcucci, M.a. and Beazley, C.e. (1988) Capability indices: Process .performance measures, Trans. ASQC Tech. Conf., Dallas, . Texas, 516-22. McCby, P.F. (1991) Using performance indexes to monitor production processes,Qual.Progress,24(2),49-55. Munechika, M. (1986) Evaluation of process capability for skew distributions, 30th EOQC Conf. Stockholm. Sweden. Munechika, M. (1992) Studies on process capability in matching processes, Mem. School Set. Eng., Waseda Univ. (Japan), 56, 109~124. Pearn, W.L. and Kotz, S. (1992) Application of Clements' method for calculation second and third generation PCls from nonnormal Pearsonian populations. (Submitted for publication). Pearn, W.L., Kotz, S. and Johnson, N.L. (1992) Distributional and inferentialproperties of process capability indices,J. Qual. Tech1'101. 216-31. 24, Pearson, E.S. and Tukey, J.W. (1965) Approximate means and standard deviations based on differences between percentage points of frequency curves, Biometrika, 52, 533-46. Price, B. and Price, K. (1992) Sampling variability in capability indices, Tech. Rep. Wayne State University, Detroit, Michigan. Rudo1t\~E. and Holfman, L. (1990) Bicomponare Verteilung - eine erweiterte asymmetrische form der Gaul3schen Normalverteilung,

Las Vegas,Nevada.

Hung, K. and Hagen D. (1992) Statistical Computation Using GAUSS: Examples in Process Capability Research, Tech. Rep. , Western Washington University, Bellingham. Johnson, M. (1992)Statisticssimplified,Qual.Progress,25(1),10-11.

I
I.

Textiltechnik,40, 49-500.

..

Johnson. N.L., and Korz, S. (1970) Distributions in Statistics:

Continuous Univariate Distributions, John Wiley, New York.

Schenker, N. (1985) Qualms about bootstrap confidence intervals, J. Amer. Statist. Assoc., 80, 360-1. .' Subrahmaniam, K. (1966, 1968a, 1968b). Some contributions to the theqry of non-normality, I; II; III. Sankhya, 28A, 389-406; 30A, 411...:32; 0B, 383-408. 3

Johnson, N.L., Kotz, S. and Pearn, W.L. (1992)Flexible process capability indices(Submittedfor publication). Kane, V.E. (1986)Process capability indices,J. Qual. Technol.,18, 41-52.

l .1

Multivariate process capability indices

5.1 INTRODUCTION

I !
I
;1

"
I~ II II ill
"

,. Ii

Fr~quently - indeed usually - manufactured items need values of several different characteristics for adequate description of their quality. Each of a number of these characteristics must satisfy certain specifications. The assessed quality of the product depends, inter alia, on the combined effect of these characteristics, rather than on their individual values. With modern monitoring devices, simultaneous recording of several characteristics is becoming more feasible, and the utilizationof such measurements is an important issue. Multivariate control methods (see All (1985) for a comprehensive review), although originating with the work of Hotelling (1947), have only recently become an active and fruitful field or research. The use of PCls ih connection with multivariate measurements is hedged arohnd with even more cautions and drawbacks than is the case for univariate measurements. IIi particular, the intrinsic difficulties arising from use of a single index as a quality measure are increased when the single index has to summarize measurements on several character~ istics rather than just one. In fact, most of the multivariate PCls which have been proposed, and will be discussed in this chapter, can be best understood as being univariate PCls,

\ I I I

180

Multivariate

process capability indices "

I'
I

Construction of multivariate PCl s

181

based on a particular function of the variable.s representing the characteristics. Further, so far as we are aware, multivariate PCls are, as yet, used' very rarely (if at all) in many

industries.

'

The contents of this chapter should, therefore, be regarded mainly as theoretical background, indicating some interesting possibilities, and only an initial guide to practice. Nevertheless, we believe that study of this chapter will not prove to be a barren exercise. The reader should gain insight into the nature and consequences 'of multiple variation, and also to understand pitfalls in the simultanepus use of separate PCls, one for each measured variate.

;1
I

JI ill

']

I I

5.2 CONSTRUCTION

OF MULTIVARIATE PCls

l'

Just as in the univariate case, we ne6d to consider the inherent structure of variation of the measured characteristics - but, this time, not only variances but also correlations need to be taken into account. Deviations from target vector (T) also need to be considered. The structure of variation has to be related to the specification region (R) for the measured variates Xl"", X v' In principle this region might be of any form - later we will discuss a favourite theoretical form though in practice it is usually in the form of a rectangular paralellopiped(a vdimensional 'box').defined by

'.1 1m

this is that a very bad (i.e. small) Cpmvalue for one variate can be compensated by sufficiently large Cpmvalues for the other values, giving an unduly favourable value for the multivariate PCI. (If one component is only rarely within specification limits, it is small consolation if all the others are never nonconforming!) A more promising line of attack may be possible using results of Kocherlakota and Kocherlakota (1991), .who have derived the joint distribution of Cpms calculated for two characteristics, with a joint bivariate normal distribution. We will return to this work in section 5.5. As with univariate PCls, there are two possibe main lines of approach - one based on expected proportion of NC items, the otper based primarily on loss functions. Both are based on assumptions: about distributional form in the first approach; and about fo~'mof loss function in the second. These more or less arbitrary assumptions are, of necessity, more extensiv,ein the multivariate case, just because there are more variates involved. Note that the approaches of Lam and' LittIg (1992) and Wierda (1992a) (see chapter 2, section 2) can

It
11

be extended straightforwardlyto multivariate situations.

Corresponding to the assumption of normality in the univariate case, we have, in the multivariate case, the assumption of multinormality, with

LSL/~X/~USL/ i=1,...,v

(5.1)

i I'

1' I
I "1

E[XiJ = ~/ var(Xd= at corr(Xj,Xr)=p// i, if= 1,..., v i.e. expected value vector ~=(~1'~2""'~v)
:~

If the specification region is of this form, Chan et al. (1990) suggest using the product of the v univariate Cpmvalues as a multivariate PCI. A moments' reflection however, shows that apart from the serious defects of the univariate Cpm,described in Chapter 3, this could lead to absurd situations, even if the v measured variates are mutually independent. The reason for

covariance

matrix

i
I

I
.

means that the joint PDF of,X = (X b ... , X v) is

V0 =(p/i'Q"/al)

with

PII

= 1 for

and varianceall i. This

fx(x)=(2n)-v/2IV 0I-~ exp{ -!(x -~)'Vo l(X-~)} (see section 1.8.)

I

(5.2)

I~

182

Multivariate

process capability indices Multivariate analogs of Cp

183

If a loss-,function approach is employed, a natural generalization of the univariate ~oss function k(X - T)2 is the quadratic form L(X)=(X-T)'A -l(X-T) (5.3)

I I
I

control charts. Nevertheless, a vector-valued multivariate PCI has been proposed. This will be discussed, briefly, in section 5.6.

with a specification region L(X)::::; (A- 1 is called.. the C2. generating matrix of this quadratic form). T. Johnson (1992) interprets C2 as a maximum 'worth', attained when X = T i.e. all characteristics attain their respective target values and L(X) is a loss of 'worth', so that .zero worth is reached at the boundary of R. Note that A does not have any necessary relationship with the vari~nce covariance matrix Vo. It is, indeed, very unlikely

5.3 MULTIVARIATE ANALOGS OF Cp

If we accept the assumption of multinormality (5.2), it would be natural to use, as a basis for construction of PCls, the quadratic form

w =(X-~)Vo l(X-~)
,

(5.4)

that. A would be identical with V0, (though tempting for

theoreticians to explore what would be the consequences if this were the case, as we will see later). In the following se<:lions we describe ~I few proposed multivariate PCls and provide some critical assessment of their properties. , Before conclu'ding this section however, we stress that any inClexof process capability, based on multivariate characteristics, that is a single number, has an even higher risk of being misused and misinterpreted than is the case for univariate PCls. The time-honoured measures and tech-' niques of classical statistical multivariate analysis - e.g. Ho,teiling's T2, Mahalanobis' distance, principal component analysis - provide ways of obtaining more detailed assessment of process variation. It would be inefficient, not to utilize appropriate forms of these well-established measures and techniques. Reduction to a single statistic, however ingeniously constructed, is equivalent to replacing a multivariate problem by a univariate one, with attendant loss of information. However, if a process is stable, it is much easier to monitor a single index than, for example to use many mean and range
'

The statistic W would have a (central) chi-squared dis-

:11 III
I

tribution with v degrees of freedom. If, also (see remarks I near the end of section 5.2), the specification region R were of I form
2 2 W::::;XV,0.9973=Cv

"

/1'

l
I

III I

,I I

11

:;1

111

Ii;

v
III

.' i

tll 11'1

"'1

the expected proportion of NC product would be 0.27%. We also note that it has been suggested that an estimate (p, say) of the expected proportion of NC items. based on obsetved data, and exploiting the assumed form of process distribution (usually multinormality), might be, itself, used as a PCI. We have already noted this suggestion for the univariate case, in section 2.1 It is important to realize that the dependence on correct form of process distribution is even heavier in the multivariate than in the univariate case. Littig et al. (1992) suggest using $-1(1(p+l)) as a PCI (as in the univariate case). See also end of section 5.5. Recall that the univariate index Cp was defined as length of specification interval Cp = 6 x (standard deviation of X) (5.5)

I-'
184
Multivariate process capability indices

I
I I

the factor 6 being used because, if C p = 1 and variation is

Multivariate

analogs of Cp

185

normal, it isjust possible- by making ~:;::t(USL + LSL)- to

have the expected proportion of NC product as small as 0.27%. Taam et al. (1991), regarding the denominator in (5.5) as 'length of central interval containing 99.73% of values of X', propose the natural generalization C
P
I

I I I I

not satisfied if the 'box' region LSLi~Xi~USLi (i= 1,..., v)is used. The volume of the rectangular paralellopiped is
i= 1

I
j

n (USLi-LSLi)

(5.10)

volume of regIOncontaInIng 99.731Yo values of X. of

vol.ume of s~e~ification region

but Pr[LSLi ~ Xi ~ USLi for all i = 1,... , v] is not necessarily equal to 99.73%, even if
v .

(5.6) The denominator is the volume of the ellipsoid

(x - ~)'VO'1 (x -~) ~ which is
.

1=1
(5.7)

.n (USLi-LSLi)=

(nx~.9973)hIVol ntv+l)

(5.11)

X~. 0.9973

The property would hold if R were of form (X-~)'VO' 1(X-~)~K2 (5.12a)

'v

(nX~.0.9973)2 IVoil

ntv+ 1)

(5.8)
~

but, as we noted in section 5.2, this is unlikely to occur. Even if an ellipsoidal R, of form (X-~)'A -l(X -~)~K2
, (5.12b)

[Of course, this would be the volume of the ellipsoid (5.7), whatever the value of ~. The ellipsoidwould contain 99.73% of values of X however, only with ~ equal to the expected value vector.] This leads us to the d'efinition

.
~ ~

were specified, it is unlikely that A would equal Vo. Taam et al. (1991) propose to avoid this difficulty by using

c
P
I

a 'modified specification region' -. R*, say - defined as the

volume of R
Qv

IVoll

(5.9)

greatest volume ellipsoid with generating matrix VO'l, contained within the actual specification region, R. If this ellipsoid is given by (5.12a), then
.

where a:=(nx~.0.9973)!v/r(tV+ 1). However, some modification of this definition is needed in order to provicie a genuine generalization of the coverage property of Cp(v= 1),which would be that, if ~=T and T is the centre of the specification region, then Cp= 1 should ensure that 99.73% of values of X fall within R. This property is clearly

II I
I'

volume of R* of(X-~)'VO1(X-~)~X;.0.9973

I
I
/1
I~
/

C~= volume

= X;.0.997;

K2

V/2

(5.13)
.

for any ~ and any v.

The most one might reasonably hope for, in practice, is

that the region R is defined by (5.12b), and it is known that

.I

186

Multivariate

process capability indices

Multivariate

analogs of Cpm

187 is an unbiased

v 0 is proportional to A, i.e. V0= 02A, though the value of the multiplier, 02, is not known.. Total rejection of this possibility (an opinion held in some circles) seems to be an unduly pessimistic outlook, assuming that practitioners have so little knowledge of the process as to be unwilling to be somewhat flexible in adjusting engineering specifications to the actual behaviour of the process in order to benefit from available theory. We consider such an attitude to be unhealthy, and encourage practitioners to, take advantage of knowledge of likely magnitude of correlations among characteristics in

is distributed as e2X~-l)v, and S/{(n-1)v} estimator of (J2. A natural estimator of vCp is

vC p~

K
X",O.9973

( )
=
((n-1)v)j
X(n-l)v

(n-1 )v 1 S

(5.16a)

which is distributed as

setting specification,if possible.

If V 0

= ()2A

then

Xv,O.9973 X(n-l)v

((n-1)v)!

"Cp

Cp

= voumeo f x v 0 1X"'::Xv.O.9973 - --- 2 I

I

volume of x'A - lX ~ K2

and a 100(1- iX)%confidence interval for vCp is

I I

= volume of x'V()1 IX~(K/()f ' - --- 2 vo Iume 0f x V0 X"'::Xv.O.9973

(
(cf (2.9 c)).

X(n-l)V,iXI2

{(n-1)v}!

C
v p,

X(n-l)v,l

~
v

{(n-1)v}7

(5.16 b)

K2
2 2 Xv,O.9973

V/2

(5.14)
,

Pearn et at. (1992) reach a similar PCl hy regarding X;,O.9973 as a generalized 'length' corresponding to a proportion 0.0027 of NC items, and K2/h2 as a generalized 'allowable length'. They define

I j tI

5.4 MULTIVARIATE ANALOGS OF Cpm

Taam et al. (1991) proceed to define a mutivariate analog of Cpmby the formula C*
pm

volume of R*

.mm

...

..

vCp=

Xv,O.9973

= C~/v

(5.15)

I I~

volume of ellipsoid x'Vi lX<C2

(5.17)

(cf. (5.13)) where

Estimation of Cp (or vCp) is equivalent to estimation of O.If Vo={)2A, and valucs X/=(xlj, ...,x"j) arc availahle for 11 individuals (j= 1, .", n), the statistic,
11

VT= Vo+(~-T)(~-T)'

~,l..

Since IVTI IVol{1 +(~- T)'VC;l(~- T)}j =

C:m = Ct{ 1 + (~- T)'VC;l(~ - T)} -j
(cf. (3.8))(5.18)

s=

L (Xj-X)'A -l(Xj-X) j= 1

I
I

I
,j

,J_1

188

Multivariate

process capability indices

Multivariate analogs of Cpm

189

(Taam et at. (1991)usethe notationMCPfor C:m). Thisindex

has the property that when the process mean vector ~ equals the target vector T, and the index has the value 1, .then 99.73% of the process values lie within the modified specification region R *. This is analogous to the property of the univariate Cpmindex. The matrix VT can be estimated unbiasedly as

I
I

'
I .

Cpm as =[E[(X- T)';(j l(X- T)]T = {I + V-1( - T)V(j 1( - T)} -i

(5.20)

Cpm

\\= and V0 by

n j= 1

(Xj-T)(Xj-T)

If V0 is known, an unbiased estimator of the denominator of 2 . C pm is

(5.19a)

! j=1 (Xj-T)'V(j1(X-T) n
and a natural estimator of Cpmis

V= where

n-1

j=1

~ (Xj-X)(Xj-X)'L.

(5.19b)

Cpm =
I I
I
I' 0

'
n

nv
(Xj-T)'V()1(Xj-T)

(5.21)

[ j~1

X=Note that

L Xj n j= 1

~T=
,

n--1 - n V +(X-T)(X-T)'
~

.j A 'I

(5.19c)

~
'1"
.

Chan et at. (1990) assume that the specification region R is

(X-T)'V() I(X-T)~ K

At this point we note that Cpm is subject to the same drawback as Cpm,noted in section 3.2. Identical values of Cpm can correspond to dramatically different expected proportions of NC items, if the specification region is not centered at T. Also, calculation of C"", requires knowledge of the correct Vo. If all arbitrary matrix is used in place of V0 ill defining Cpm, we can encounter a similar effect of ambiguity in meaning of the value of Cpm' (See Chen (1992) and also Appendix 5.A). The numerator v is introduced because if ~= T, then
E[(X - T)'V (j 1(X- T)] = v

of form

, lit,

If ~= T, then C;.; is distributed as (nv)-1 X;v. Hence, in

thes~ special circumstances E[ Cpm] =

i.e. the generating matrix of the specification region. R. is

proportional to V0, the variance-covariance matrix of

the measured variates. They define a multivariate analog of

()
2

nv ! r( t(nv - 1))
l( tnv)

(5.22a)

190

Multivariate process capability indices

Multivariate analogs of Cpm

191

ap.d
var(Cpm) =Jnv
-

r(!nV)r(tnv-t)-

{r(!(nv-t)Y
] (5.22 b)

{r(tnv)}2' -

-I.

when nv= 100, we find Go.o2ss:f0.8802(correct value 0.8785) and Go.os~0.8978 (correct value 0.8968): and when nv = 200, we find GO.O25 ~0.9118 (correct value 0.9109)

and Go.os ~0.9251 (correct value 0.9245)].

(Chan et ai. (1990)). Percentage points for Cpm(on the assumption that are easily obtained since Pr[Cpm~ GaJ = Pr[C;'; ~ G;2J = Pr[(nv) - Ix;v ~ G; 2J =Pr[X~v~nvG;2]
Hence, to make Pr[Cpm ~ GaJ = Ctwe take

~= T)
I

Peam et al. (1992) felt that Cpmis not a true generalization of the univariate Cpm,and proposed the PCI 1 vCpm=vCp{ 1+ ;(~-T)'VOl(~-T)
=vCp X Cpm
II I I

-i }
(5.25)

(see (5.15) and (5.20)).

\ I
I

Note that the univariate

indices Cpm and Cp a!e related by

Ga-- (nv)! Xnv.l-IX Except when nv is small the approximation

GIX ~ (2nv)! zl-IX+(2nv-1)1

(5.23)
\

Cpm=Cp{l+(~~TY}

-1'

t
(5.24 a)

where ~ = E[XJ, (J2 = var(X) and T is the target value for X. A

natural estimator of vCpm is

-

- 0--' I K(nv) "Cpm=

~

Xv,O.9973

x (S+n(X-l)Vo

-,

- 1 if

(A--T)f

1 -!

(5.26)

where

<I>(z - IX) 1- Ct,or even 1 =

, , Zl-IX +1-GIX= C2nV)!

4nv-

-1

In all the previous work, distribution theory has leaned heavily on the assumption that: (5.24b)
I~

1. R is of form (X t)' A-I (X- T)~ K2 and.

2. 1\,= V0, or, at least, A is proportional as in (5.14). to V0, i.e. (12 = V0, A

give values which are sufficientlyaccurate for practical purposes. [For example,using (5.24a) when nv=40 we find Go.o25s:f0.8245(correct value 0.8210) and Go.os ~0.8492 (correct value 0.8470):

In this second case, even with ~#T, it is possible to evaluate, the distribution of
n ()2

j= 1

L (Xj-T)'Vo

l(Xj-T)

192 as that of

Multivariate

process capability indices regIOn IS

Multivariate

analogs of Cpm

193

02 X(noncentral X2 with nv degrees of freedom and noncentrality parameter n(~ - T)'Vo l(~ - T))

h( x -- T) < r 0

Table 5.1 Scalar indices for multivariate data

Source

However,if A is not proportional to V0, the distribution of

(X-T)'A -l(X-T), although known, is complicated (Johnson and Kotz (1968)). (Appendix SA outlines some details.) Other approaches to measuring 'capability' from multivari-

Symbol Cp

Definition vol. speen. region vol. of (x-~)'VOI(x-~)~X;.O,9973 As Cp, with modified speen. region Cl/v , p As C~, but denominator =(vol. of

Taam et al. (1991)

ate data include:

1. Use of the PCI

Taam et al. (1991) C: Pearn.et al. (1992) vCp Taam et al. (1991) qm
Chan et al. (1990) Pearn et al. (1992) Proposall. Proposal 2.
characteristics, and

vCR= IVoA-II +(~- T)'A -l(~-T)

(5.27)

(x'Vi 1X<C2))

Cpm

As but denominator E[(X-T)'Vo l(X- T)]

. The motivation for this PCI is that it is expressed in terms of the matrix A used in the specification, rather than the process variance-covariance matrix - on the lines of Taam et al. (1991). If v=l and A=(1d)2 we have ICR=C;n~' Note that small values of vCRare 'good', and large values are 'bad'. 2. Proceeding along the lines of Hsiang and Taguchi (1985) and Johnson (1992) (see section 3.2) we may introdu.ce L(x)=(x-T)'A -l(x-T) (5.28a)

, I

vCpm vCp X Cpm iVoA-ll+(~-T)'A -l(~-T) vCn E[L(X)] tr(A-1Vo)+(~-T)'A -l(~-T)

matrix of the joint distribution of the process

v 0 is the variance-covariance

V 1'= V 0 +(~

--T)(~ -T)',

where h(') is a nonnegative scalar function, satisfying the

condition h(tx)= th(x) for all t > O. Then a multivariate'

PCI can be defined as MCp=r/ro

as a loss function (generalizing L(x) = k(x- T)2).We.have E[L(X)] =tr(A -IV o)+(~- T)'A'-l(~-T)

where

(5.28 b)

Pr[h(X-Tr]

=a

where tr (M) denotes 'trace of M', which is the sum of diagonal elements of M. Large values of E[L(X)] are, of course, 'bad', and small values, 'good'. For readers' convenience we summ<Jrize,in Table 5.1,

the scalar indicesintroduced in this chapter.

This ensures that if MCp= 1 the ~xpected proportion of NC items is a. This definition includes, for example, the rectangular specification region IXj-Tjl<rj (j=l,...,v)

3. Chen (1992) has noted that a broad class of PCls can be defined in the following way. Suppose the specification

194

Multivariate

process capability indices

J4

Vector-valued PCls

195

by taking h(X-T)=
1 t;it; v

n"iaxrj-lIXj-

Tjl

111 I

and 1'0= 1. In this case, and several others, actual estimation of MC" (Le. of r) can entail quite heavy compulation. Some details of possible methods of calculation are provided by Chen (1992)~ 4.' Wierda's (1992a, b) suggestion that -1<1>-1 (expected proportion (p) ofNC items) be used as.a PCI, is applicable in ,multivariate, as well as 'univariate situations. It can then 'be regarded as a multivariate generalization of Cpk' The same comments - that one might as well use p, itself, and that p is estimated on the basis of an assumed form of process distribution, accuracy will depend on correctness , of this assumption - apply, as mentioned in section 2.

joint distribution of Cps for v = 2, under assumed multinormality. A different type of vector PCI has been proposed by Hubele et al. (1991),in which they suggest using a vector with I three components. 1. The ratio ar~a of specification rectangular paralellopiped + area of smallest rectangle containing the 99.73% ellipsoid, (x- ;)'V 0 l(X-;) ~ X~, ,9973 0 ~ ]
,

,I

= 'specification rectangte'

'process rectangle'

(5.29a)

The numerator is (see (5.10)

5.5 VECTOR-VALUED PCIs

As we have noted (several times) it is a risky undertaking to represent variation of even a univariate characteristic by A single index. The possibility of hiding important information ,is much greater when multivariate characteris.tics are
I

[/~ (USL/- LSL/)J+

Hubele et al. (1991) show that the denominator is

under consideration, and the desirability of using vector

valued PCls arises quite naturally. One such vector would

I I

I consist of the v PCls, one for each of the v variables. These I might be Cps, CpkSor Cpms,and would probably be related I to .the concept of rectangular parallelopiped specification "
regIOns.
I
'

I I
2, 1

2Xv.O.9973(

tl1lVOill}+ Ilvo

11)!'

(5.29b)

where, Voil is obtained from Vo 1 by deleting the ith row and the ith column. 2. The P-vatue of the Hotelling T2 statistic

We have already mentioned (in section 5.1) a suggestion to

use thc product of Cpms, and havc pointcd out thc dcli~icncics

I i
,

w2 =n(X- T)'~-l(X- T)

(5.30)

I of this

proposal. Interpretation of the set of estimated PCls would be assisted by knowledge of the joint distribution. The

work of Kocherlakota and Kocherlakota (1991)provides the,

where ~ is the s@!p~an~~:QYariance matrix. This component includes information about the relativelocation of the process and specificationvalues.If ~'=T - i.e.

196
,

Multivariate

process capability indices

thc proccss is accuratcly ccntrcd

distribution of

thcn WZ has the

!
I

Appendix

5.A

197

v(n-l) {Fv.n-v variable} n-v When X is close to T the P-value of WZ is nearly I;thc further X is from T, the nearer is the P-value to zero. 3. The third compone:nt measures location and length of sides of the 'process rectangle' relative to those of the 'specification rectangle'. For the case v= 2, and using the symbols UPRi, LPRi to denote upper and lower limits for variable Xi in the process rectangle, the value of the component is max l. IUPR1-LSLd, ILPR1- USLll, IUPRz-LSLzl,

,
J I ;
I

evant distribution theory. Nevertheless, (1.), (2.) and (3.), comhincd will give a useful idca of thc process capability especiaIIy indicating in what respects it is, and is not, likely to be satisfactory.

";

APPENDIX 5.A

We investigate the distribution of

~

W=(X-T)'A

-l(X-T)

USLt - LSL1 USL1- LSL1 USLz- LSLz

I I
I

where X has a v-dimensionfll multinormal distribution with expected value vector ~ and variance-covariance matrix V0, and A is a positive definite v x v matrix. From Anderson (1984, p. 589, Theorem A2.2) we find that there exists a 110nsingular v x v matrix F such that FVoF'=I (5.31a)

ILPRi - USLzl
USLz - LSLz J

I
A value greater than 1 indicates that part or whole, of the process rectangle falls outside the specification rectangle. Taam et al. (1991) point out that components (1.) and (2.) of : this triplet are similar to the numerator (Cp) and the estimator of the denominator,
I

and i
\

FAF' = diag (),)

(5.31b)

Ii'

I
I

i I

where diag (),) is a IIx IIdiagonal matrix in which the diagonal elements A1, ..., Av are roots of the determinantal equation

IA-AVO/=O and 1= diag (1) is a v x v identity matrix. Note that (5.31a) can be rewritten F'F=VOl and (531 b) as (F- l)'A -1 F-1 = diag(1/),)

(5.32)

[1 +(X- T)'V-1(X- T)]i = [1+n-l WZJi

of C~III sec (5.1R, rcspectively. ( Although these three components do measure dill'erent aspects of process capability, they are by no mearis exhaus, Hve. Also, when v> 2, calculation of component (3.) is cumbersome, and more importantly, there is no practically rel-

(5.31 c)

(5.31d)

198

Multivariate

proces capability indices

Appendix 5.A (W is the weighted sum of n independent

199

If Y = F(X - ;) then Y has a v-dimensionalmultinormal distribution with expected value 0 and variance-covariance matrix

xi variables.)

The general distribution of W in this case has been studied by

Johnson and Kotz (1968)inter alia. They provide tables of per-

E[YY'] = FE[(X- ;)(X - ;)']F' = FVoF'= I

from (5.31a). Then since X-;=F-Iy,

centage points of the distribution for v= 4 and v= 5 for various combinations of values of the ASsubject to the condition

W=(F-Iy + ;-T)'A -1(F-Iy +;-T) "';(Y'+(~-T)'F'}(F:-1)' A-1F-1(Y +F(~-T)) =(Y +F(;-T))' diag(1/l)(Y+F(;-T))
v

L -=v j= 1 Aj
These lOO(l-IX)% percentage points are values of c~ such that Pre W < c~] = 1 -IX. From the tables one can see how c~ varies with variation in the values of the AS. Some values for v = 4 are shown in Table 5.2.
Table 5.2 Percentage points of W A-I I. 2.5 2.0 1.5 1.5 1.2 *1.0
..1.-1 2

j= 1

Aj-1(Yj+<5Y

(5.33).

where 0' =51>"" <5v)=(;- T)'F'. We note that in the particular case when A = ()2V 0 we have from (5.32), A1=A2="'=Av=82 so
W=O--2
v (Yj+<5j)2

rl3 0.4 0.7

OJ

Ail 0.4 0.3 0.2 0.5 0.4 1.0

97.5% 14.3 12.8 12.4 12.3 11.7 11.1

99% 18.3 16.0 15.2 15.0 14.1 13.3

99.5% 21.4 18.4 17.3 17.1 15.9 14.9

99.73% 24.1 20.7 19.2 19.0 17.5 16.25

chisquared

J= I

0.7 1.0 1.5 1.5 1.2 1.0

0.5 1.2 1.0

(5.34) In all cases !:j=IAjl=4.

*Thc values in this row are percentage points of X

and is distributed as 8-2 x (noncentral chi-squared with v degrees of freedom and noncentrality parameter
/I

with four degrees of freedom.

t
c II

j= 1

<5} =o'o=(;-T)'F'F(~-T)=(~-T)'Vo

1(~-T)).

If we have T =;, i.e. the process is value, ~hen ,)= 0 and W reduces to
n

centred at the target

~. II II

The more discrepant the values of the AS, the greater the correct value of c~. This will, of course, decrease the values of the PCls defined in sections 5.3 and 5.4, implying a greater proportion of NC items than would be the case if all ASwere equal to 1. Of course, one must keep in mind that this conclusion is valid with regard to situations restricted by the requirement of a fixed value of th~ sum
n 1

w= I

-L

Y2
(5.~5)

j= I Aj

j= 1AJ

I-

200

M /III iIJi/rii/1 t'

l'nJ('('SS

(,ill'ilhi fil.l' iI/ii jet's

If all the AjS are lar'ger in the same proportion, one would obtain proportionutely smaller values of c;.
BIBLIOGRAPHY II Alt, F.B. (1985) Multivariate qliality control. In Ellcyclopedia (!I' Statistical Sciences, (S. Kotz, N.L. Johnson and.e,B. Read, eds.) 6, Wiley: New York, 110-22. . Anderson, T.W. (1984) An Introduction to Multivariate Statistical Analysis (2nd edn) Wiley, New York. . Chan, L.K. (1992), Cheng, S.W. and Spiring, FA (1990) A .multivariate measure of process capability, J. Modeling Simul., 11,1--6. [Abridged version in Advances in Reliability and Quality Control (1988) (M. Hamza, ed.) ACTA Press, Anaheim, CA; 195-9]. Chen, H.F. A multivariate process index over a rectangular paralellopiped, Tech. Rep. Math. Statist. Dept, Bowling Green State
I

Bibliography

201

Peam, W.L., Katz, S. and Johnson, N.L. (1992) Distributional and inferential properties of process capability indices, old and new, J. Qttal. Technol. 24, 215-31. Taam, W., Subbaiah, P. and Liddy, J.W. (1991)A note on multivariate capability indices. Working paper, Dept. Mathematical Sciences, Oakland University, Rochester MI. Wierda, S.l. (1992a) Multivariate quality control: estimation of the percentage good products. Research memo, No. 482, Institute of Economic Research, Groningen, Netherlands.Faculty of Economics, University of
.

.~

'

Wierda; S.l (1992b) A multivariate process capability index, P"oc. 9th Int(!l'Ilat.Conf. Israel. Soc. Qual. Assur. Jerusalem, Israel.

University,BowlingGreen, Ohio.

Hotelling, H. (1947)in Techniques of Statistical Analysis (e. Eisenhart, H. Hastay and W.A. Wallis, eds.) McGraw-Hill: New York, 111-84. Hsiang, T.e. and Taguchi, G. (1985) A Tutorial on Quality Control and Assurance - The Tayuchi Me/hods, Amer. Statist. Assoc. Meeting Las Vegas, Nevada (188pp.). 'Hubele, N.F., Shahriari, H. and Cheng, e.-S. (1991) A bivariate process capability vector, in Statistical Process Control in Manu, facturing (lB. Keats and D.e. Montgomery, eds.) M. Dekker: New York, 299-310. . Johnson, N.L. and Kotz, S. (1968) Tables of the distribution of quadratic forms in central normal variables, Sankhyii Ser. B, 30, 303-314. Johnson, T. (1992) The relationship of Cpmto squared error loss, J. Qual. Technol., 24, 211-15. Kocherlakota, S. and Kocherlakota, K. (1991) Process capability indices: Bivariate normal distribution, Commull. Statist. - Theor. Meth., 20, 2529-47. Littig, J.L., Lam, C.T. and Pollock, S.M. (1992) Process capability . measurements for a bivariate characteristic over an elliptical tolerance zone, Tech. Rep. 92-42, Dcpt. Industrial and Operations Engineering, University of Michigan, Ann Arbor.

I
I

I I

r '/ , I

Note on computer progralns

,
~

('

The following SAS programs, written by Dr Robert N. Rodriguez of SAS Institute Inc., SAS Campus Drive, Cary NC 27513-8000, USA, will be available in the SAS Sample Library provided with SAS/QC software (Release 6.08, expected in early 1993). A directory of these programs is given in the Sample Library program named CAPDIST. Chapter 2 Expected value and standard deviation of Cp Surface plot for bias of Cp . ' Surface plot for mean square error of Cp Expected value and standard deviation of Cpk Surfm:e plot for bias of Cpk Surface plot for mean square error of Cl'k Plot of probahility dcnsity function of Cp

I
I

Plot of probability density function of Cpk

i f I
.

.Confidence

J
.

intervals for Cp Approximate lower confidence bound for Cpk using the approach of Chou et al. (1990)(reproducestheir Table 5) Approximate lower confidence bound for Cpk using the approach cf Chou et al. (1990) and the numerical method of Guirguis and Rodngucz (1992) (generalizes Table 5 of Chou
et al.) .

1
.

Approximate confidence intervals for Cpk using the methods' of Bissell (1990) and Zhang et al. (1990) 203

204

Note 011computer programs

Chapter 3 Expected value and standard deviation of Cpm Expected value and standard deviation of Cpmk Surface plot for bias of Cpm Surface plot for mean square errr of Cpm Approximate confidence limits for Cpmusing the method of Boyles (1991) Chapter 4 Expected value and standard deviation of CjkP/Cjlil~ Surface plot for bias of CjkP Surface plot for mean square error of (\.\
,

Postscript

'i!!

Our journey down the bumpy road of process capability indices has come to an end, for the moment. We hope that those who have patiently studied this b90k, or even those who have browsed through it in less detail, have been motivated to seek for answers to the following questions. I. Are the resistance and objections to the use of PCls in 'practice justified? 2. is there truth in the statement that 'a manager would rather use a wrong tool which (s)he understands than a correct one which (s)he does no1'? 3. Are PCls too advanced concepts for average technical workers on the shop floor to comprehend? Let us state immediately that our answer to the last question should be an emphatic 'No'. In our opinion it is definitely not too much to expect from workers a basic understanding of elementary statistical and, prohahilistic concepts such '~s me,an and standard deviation (and target value and variab,ility in general). We also sincerely hope that the answer to, the second question is an equally emphatic 'No'. In spite of rather disturbing catch phrases and slogans, such as 'KISS (keep it simple, statisticians)' and even 'Statisticians keep out' (ascribed to Taguchi advocates at a meeting in London), we strongly believe that a great majority of managers and engineers are sufficiently enlightened and motivated to explore new avenues, and are not lazy, or afraid o~ the 'unknown'. Deming's philosophy and a shrinking world (from development of 'instant' communications) are factors

'D

-J

205

206

Postscript

Postscript

207

contributing to an atmosphere favourable to development of innovative techniques C!n the shop floor, and improvements in educational methods may be expected to produce many more open-minded individuals. Having explained our reactions - we hope, satisfactorily to the second and third questions we now attempt to reply to
.

the first, and most difficult,one. We rephrase the question should the use of PCls be encouraged? As we indicated in the introduction to Chapter 1, PCls can be used with advantage provided we understand their limitations - in particular, that a PCI is only a single value and should not usually be expected to provide an adequate on its own, and PCls are subject to sampling I measure .
, variation,

as yet; provide a sufficiently clearcut picture for theoreticians to guess the exact form of the sampling distribution, on the lines of W.S. Gosset ('Student')'s classical work on the distribution of XIS in 1908 wherein hand calculation coupled with perceptive intuition led to a correct conjecture, deriving the t-distribution.

as are other statistics. In the case of sample mean

(X), for example, the sampling distribution of X is usually approximately normal even for moderate sample sizes (n> 6, say), even if the process distribution is not normal. We therefore use a normal distribution for X with some confidence. On t4e other hand none of the sampling distributions of PCls in this book is normal, even when the process . distribution is perfectly normal. , Tables of the kind included in this book provide adequate

~ackground . for interpreting estimates of PCls when the

process distribution is normal. In our opinion, objections, ioic~d by some opponents of PCls, that are based on noting that the distribution of estimates of PCls when the process distribution is non-normal is unknown and may hide unpleasant surprises are somewhat exaggerated. The results of studies, like those of Price and Price (1992) and Franklin and Wasserman (1992) summarized in Chapter 4 provide some reassurance, at the same time indicating where real danger maylurk.In particularthe effects kurtosis(fJ2< 3, or more of

importantly {J2 3) are notable.

We appreciate greatly the efforts of those involved in these enquiries, although we have to remark - with all due respectthat their methods (of simulation and bootstrapping) do not, ,

Index

Approximations 10, 12, 46, 109 Asymmetry 113 Asymptotic distributions 170 Attributes 78 Bayesian methods 112, 151 Beta distributions 22, 118, '121, 128, 148, 172 functions 22 incomplete 23 ratio, incomplete 23, 80, 129 Bias 33, 57, 98 Binomial distributions 15, 114, 172 Bivariate normal distributions 179
.

nO~1central 99, 111, 20, 119, 141, 198 Clements' method 156 Coefficient of variation 7 Computer program 203 Confidence interval 45, 70, 107

for Cp 45, 47 .
for Cpk 69 for Cpm 107 for Cpm 187 Contaminated distribution 136, 140 Correlated observations 76, 106, 179 Correlation 7, 29, 78, 82, 106, 181 Cumulative distribution function 4 Delta method, see Statistical differential method Density function, see Probability density function Distribution function cumulative 4 Oistribution-free PCls 160

Blackwell-Rao theorem 33, . 80, 133 Bootstrap 161

Capability indices, process, see Process capability indices Chi-squared distributions 18,45,75, 137, 172, 181

210

Index
Index

211

Edgeworth series 140, 146 Ellipsoid of concentration 18,4 Estimation 43, 98 Estimator 32 maximum likelihood 32 moment 32 unbiased 79 unbiased minimum variance (MVUE) 79, 127 Expected proportion NC 40, 53, 92, 183
. value 6

Kurtosis 8 Likelihood 32 maximum 32 ratio test 35, 75 Loss function 91, 181, 192 Maximum likelihood 32 Mean deviation 47 Mean square error 92, 101 Mid-point 39 Mixture distributions 27, 99, 121, 169 Moments 7,59, 102, 122, 168 Multinomial distributions 16, . 141 Multinormal distributions 29, lXI, 19X Noncentral chi-squared distributions, see Chisquared distributions,. noncentral t-distributions, see [distributions, noncentral Nonconforming (NC) product 38 Non-normality 135 Normal distributions 16, 53, 161 bivariate, see Bivariate normal distributions folded, see Folded distributions, normal Odds ratio 78

Parallelepiped 178, 194 Pearson- Tukey (stabilizing) results 157 Percentile (percentage) point . 18, 190, 199
Performance indices, process, see Process performance indices

Exponential distribution 19, 138, 149 Folded distributions 26 beta 27, 174 norl1lul26, 56, 17] Gamma distributions 18, 149 function 18 incomplete 19 ratio 19 incomplete 20 Gram-Charlier series, see Edgeworth series History 1, 54, 89 Hotelling's T2 195 Hypothesis .35 Independence 7 Interpercentile distance 164 Johnson-Kotz-Pearn method 156

'I

Poisson distributions 15, 21, 99, 107 Probability density function 5 Process capahility indices Quadratic form 30 (PC Is) Cjkp 165 Ratio 47 Cp 38 Robustness 151 C p 186 Cpg 90 Semivariance 116, 153 C pk 51 Shape factor 8, 67, 151 Cpk 67 kurtosis 8 Cpk 66 skewness 8, 149 Cpkl 55 Significance test 35 Cpku 55 Simulation 75, 138 C pk(8)157 Skewness 8, 149 Cpm 90, 187 Specification limit 38 <;in 112 lower (LSL) 38 Cpm 187 upper (USL) 38 Cpmk 117 region 178 Cpm(a) 121 Stabilization 157 Cpp 40 Standard deviation 7 TCpkl 55 Statistical differential methOd 'TCpku 55 10, 107 vC p 187 Sufficient statistic 32 vC R 192 Systems of di~tributions 30 CLjkp 164 Edgeworth 31 Gram-Charlier 31 CUjkp 164 Le 90 Pearson 31, 156
'

MCp 193 MCP 188 comparison of 74, 95 tlexible 164 . for attributes 78 for multivariate data 177 vector-valued 194 Process performance indices (PPls) Pp 41 Ppk 55

212
t-distributions 24; 137

Index Variance 7 Vector-valued PCls 194

Worth 9.4

ill"

noncentral 25, 63 Total variance (Otolal) 41 Triangular distribution 138

Uniform distribution 22, 136, 149

,
!
*~ :ti, II 1 ~ ~
'"'till 'I

~ .:J

II-- t')

!::". /'
I -:> 0..,

I')
.? ",r'

I
.~ 'I '\

\< t; ~r) f~
~J C'?I.3
"",""""" :\.f:'~ ),"'-" ~. .~.,,'$ ~~ ' '~;:]I;'/

'(
., ...,". t,...
~ Ct

'" Process capability

indice

\\\\1\\111\\1\111\111111\\111\\

~
m It
[!

168718 KOT/P N93 658.562 Kotz, Samuel

!l
1~

lG871R

:~
g