Académique Documents
Professionnel Documents
Culture Documents
<
and Norman L. Johnson Unwersity of North Carolina, ChapelHill, North Carolina, USA
rm
("\ '-)
l,
q..
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'!.
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
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-
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
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.
We assllme that the reader possesses, at least, a basic knowledge of the elements of probability theory. We commence
..
Introduction
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)
continuous random variable, X, for any real a and {3,({3 a): >
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)
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)
..
IlllfOdL/CliOIl
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)
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<)).
.
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 + 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
~= E[XJ = '2>iPr[X i
=XiJ
( \.8 hI (1.9 b)
E[g(X)]= I,g(xJPr[X=xiJ i
..
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
(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)
J /3 = a3
1
It J
/12 = It4 a4
. (1.12(')
S = --and C.V.(X)
= 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;
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
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)
'
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
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) ~
13
(~) 2 [{c.v.(A)2
+{C.V.(B)}2]
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
= 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
-!
where X),...,X"
and
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
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-
1 a-.
(1.19c)
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,
'"
( )= k!(n-k)!-kf
k
n!
n(k)
X", Poi(O)
(1.21)
..
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
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
'
( )
( ) (1.24h)
j= 1
L L n = n, j= 1 Pj = 1
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
<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~
3.0902.
3.2905
,,
.
18
Introduction
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 =
- ro up to 1.645is95%
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
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)
'
/-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
.
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
p.i 2
() ()
2
c52
c52 j
~=Pr[Y=jJ .I.
i= 1
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
()
2
k
c52
(1.33)
I
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 ~)
) Ji.J exp(':-2 )
I
c52
(L b )
~
,.
h=l
with v =
h=1
,
Vh
(1.34)
22
1111 /'oi!l/cl iOIl
Beta distributions
23
In particular, a
E [ YJ = J1't ( Y.\ =
~+
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
(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)
O~y~l
(1.35 h)
I)
t~
1'1
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
y)P-l
dy
rx,/3>0
Xl+X2"'X\,1+\,l
and X1+X2
-~'"
Xl.
eta
(2' 2/
V V2 -1 - )
(1.36)
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
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!
24
t-distributions of
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 particular
'
v (1A1)
v-2
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
"
. " )
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
(1.46 e)
~ 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)
(--
X-sV
()
1
\
+ exr \
- '2 -~ -";-_c..)
211
\'
(\2\
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)
.)
If 17,(x), F~('(),
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
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~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)
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,
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
30
1ntroduction of X\>...,XI'
31
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
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
ellipsoid is
'
. .
(Type III) Cz:":: 0; C1 :;6 0 - Gamma distribution (section 1.5) (Type 1) Roots of (1.59) real but opposite sign -- Beta
tion (scdion., 1.7) Thcse families, we have already encountered, and further I
I
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
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)'"
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
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
35
l-
~2I ,
I
1 (X i-/I
l;)'~
10 choOSL: hdwL:en
hypotheses
Ilj (i
0, I)
(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)
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)
()
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
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.
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
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
- 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
~
t'
~I
(2.2)
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'
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
( ,'
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 -
to
6'
0 '"
"""I'>' r-
Estimation
oj C p
43
0..
0..
;':::::0 0
'"
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
""
U I"'i",
I
Os
00i:...J
-..;....
f-o
-~
""oS< UN
with (n-I)
\
\
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
'
-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)
which is distributed as
(n-l);.,
(2.9 h)
C;
Analysis for
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
I
I
46
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'
tt!
f
e LS I J~ (I
I 41. I)} n1
1.2.! I lIj
-)2
,I)
~
(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)
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
4~
where (2.13)
II
USL-LSLy:[=y:[Cp,
60" Xf Xf
()
I'(ln
l rcfU-l))
(2.17a)
,
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
.', '
. (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
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
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
0 0 0
\0 '<too lI) 0\
vi
~.
. I i
0 o.=:>
lI) 0\ 00 '<to\
IJJ 0
vi
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<)~~~-------
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
r<)1
VJ
'n
r-- 00
- - - - ".1', --0 00 0\
"n
u= 4) and
;
t:'
d
1
r<)-.:I"oo~r<)r--r<)Or--~'<tr<)O
X
s;;
0 0\\0 0\
uJ
0.
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-<
..!::::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
Cpk=
min(USL-~, ~- LSL)
3a
(2.20 a)
'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
(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 ")
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*
(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
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.
(2.25 a)
-=---:,-;
(2.25 b)
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+
LJSL)I
(2.23)
TCpkl=3(J
T -.-LSL
and
(our notation).
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
USLD
(2.26)
56
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
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)
-
~(~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
P ( ("
"
'1 ), ,1~
(2.30)
IU +61
where U is a standard normal variable and
(2.28)
a=[---=-i" i= 1 (Xi-X? J , L n
I,
L"
~,
()
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
E[C,.] ~3 2
rG)
nn,
5~ (b) E ftIX
,
il/dic('s
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"-- V) N I")
00000
0\ V)
00000
or, N
1.
000"";"";
~.
O"",v-,ooCV)OV)...... OI")I"--O<:t"
000"';"";
Of""'.<:t"
I")
I'
V,?
a
C/i
00000
0<:t"01"--<:t" N N
o
I")
00000
00000
N.
I.:U
\0
-'"
's:,c-I-?,
a
C/i
...... V)
00000
'r, 0\ 7 oJ:) 1"1 0\
00000
vj II "V'
J,
I
r:
a c.ri
I.:U
0\ '</" V) N
I") \0 1"--,
!;;~~&j~
I") \0
I")
I"--
00"";"";"';
II
u:i
or.
;;t?2~~~
-<
00000
("I ",
I")
<:t" V)
00000
<:t"O\OM N N
I")
00 cr. G\
00000
0
("I
001")
0 I.:U
\0
pk,
as a peremptory
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"-- 'n
\0 00 (" J 00"";"";"";
'r.
N \0
'r. 0\
0'\ 0
'0 '"
Q)
Q C/i
0
00000
("'I '"T
00000
I") V) I") 00 \00\1")\00
- - (')'I">
("
II
s c
Q)
(2.31)
I
II
:?: ""2'
I.:U
0
V)
0"";"";
('.j
01;:,
~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
1 S.D. E
1.5 S.D. E
2 S.D.
d(J
")
, .'
-+
0.634
.;; 6 11=30 da
") -
, .)
-+
:5 6
11=
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
-"'
--
-.
(?('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,
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-:-:-:
n 10 20 30 60 80 100 200 400 600 2200 3200 5400 10800 30500 79500
are
00-:-:-:
q C/)
00--00000
<n-O\O\O 'OO\-"<too
00--00000
t--<n<'">-O\
00-:-:-:
t--
00-:-:-:
0\
-..; .
~
~ M
O\O~<noo
0---'0 0 ~,
'0
00000
\.IJ
00--00000
-00'O~,-
00-:-:-:
00-:-:-:
:c c::I
f-<
<1J
<nb <n
II
b ~. '0
'n \0 II"""
:;:
'
rl
,r, "T
'r. ";:;
:::
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
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
I
~
11;,
of the
Cpk
distribution
tributions.
is cquivalent to
1Or81
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
Fig. 2.1 Probability density functions of Cp for various sample sizes and. values of d/cr.
66
2.6 2.8 3.0
of' Cpk
67
2.6 2.6
'..d*=1 """"",
,\,- .0 --l
3.2
#.
I;
,,/
'
3.0
eJ' I!,
~
3.4
.~ 3.6 0
if
t::9
,/ .,1 , I
';(
~;
! /
"
'
3.21
.
~ 3.6
4.0 4.2 4.41
4.6j
,if /t
,;;",."'~\,
d* ; 2
.
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
.
. ---
~-~~~-~._-
-1.2 ,
-1.0
- 0.6
- 0.6
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)
= min(C~kl' C~kU)
(2.35)
where
~, C'pk-
X - LSL 38
USL:X 30' -
(2.34)
I
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
68
2.6
2.8 3.0 3.2 -;;; 3.4 .3
'ijj 3.6 '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
.'//,::0
//:::="-,
--'-"
r
I
I I
".
I! / //
:JJ.~/ b <l! ,~
"
IC
~
1.2j
.3
",
f! :J
::.::
RN
/ 1( d
cY/
f/
/
CJ::
,/I
,
//
V
.'
I'
U-,
"
u~O,25
016.
,
/ .. ..
. .,,-.-
(\~ 0.4
/'O":\)'
.,/
,!?
IC
" I
I
if
,/1 /f
/1 /{
Reflected Reflected Reflected , Reflected
'
"'CJ:: ,"
.il
Y'9j/
"
,/,'
"
IC /j
. ,.,/
.',il 11 .
,I
~
Type III Lognormal Type V Inverse Chi
3.8
4.01
/f /(
/ (,
1.
,
'-'-' -"'-
" /1'
4,2 t
'
,!
'"
r
0.2 0.4 0.6 .
=:0
,iI 1/ /11
';I
-1.2
-1.0
-0.8
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
ratio~
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.
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
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
"
,/
and a are
x-
tr. I - al 2 a ~ ~~ X + t f. I - ~I2 a
XI,I-aI2
(2.36a)
,
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
. 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)
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
Kushler and
11urley
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 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
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
Comparison of capabilities
75
needed; this range is not, in general, just the interval with end points
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)
!(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
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
~ 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
u = [{T'y + 2(n-l)
..
(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
76
Correlated observations
77
I
I
n 5 10 15 20
all
VII
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)
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
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
CP=:3S
P
of Cp,the denominator willtend to underestimate 0" and so
C\,
~
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
Appendix 2.A
:I
79
'shrinking'
R towards
(,
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
(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
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
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
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=
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
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)
~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-
LSL-X
S. l
( )
n-1
APPENDIX 2.B
( )J
.
.
,
f/- J
,
II
expected value ( and variance (j2; the correlation between Xi and )fj wi1lbe denoted by Pij. We have
.
var(X)=n-2(j2
n+n(n-1)
82
where
Bibliography BIBLIOGRAPHY
83
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
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
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
~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
! !
~~
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.
(
.. -'
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 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
(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
\
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
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
C: ~= 61.25;
(j
= 1.25
t:
I
I
I
, ,
Jill
Process A B C
Cp 1 2 4
Cpk 1 1 1
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.
"
..
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
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'
(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
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('
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
95
IIiCII
C;m
= (~Y {E[X
-- 1'fJ}-1
.'he Ipl
11'".
1110
we see that,
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
,'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
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
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
97
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
~ and
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.
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
lin~s
"">'~\
98
Estimationof Cpm
~
(and C1~m)
99
!~/--=E:~--::_!!1l! 3
{n
}-
(Xi i~ I
-j
T)2
(3.16)
.(a)
.1(11) II ~
i
m-d m'-:!d/c m m-dd/c m+d
r- ~
11'1
Cpm
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
(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
-'
C~m
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)=
..
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
i~l
(3.15)
')
(Fo, r;;'n,I4(C~rn) infinite.) is
I I
100
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
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
I,
C\'m. Although
Cpm
When
~= T
'
(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))
(3.21c)
",I
> -n-
( ) E[CpmJ-Cpm ( )
!
n~
I !
Cpm-<;pm
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 )
.-
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_..
~.'.
I(- TI/G
0 Q5 S.D. E S.D. E 1
2 S.D.
E n= 10
ciG -
.:; 6
11=
15
"G
~
.:;
0.627 0.122 0.941 1).183 1.254 0.244 1.568 0.305 1.881 . 0.366
'-
...
-.-"
---
.-'-
..-. ---_.
. .
n=20 dJG
2 S.D.
--- '
---.
.,-,,""'"
~_..
106
I -I'
I
Table 3.2 Values of Cpm . I-TI
(J 1
107
d (J
To evaluate EJ [(n - 2 + 2J) - 11, we resort tCJthe method of statistical differentials (or 'delta method'), obtaining
2 3 4
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
[ 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
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
Since
Cpm=td{ (J2+(~ - T)2}-j
E[C;mJ=
(~ )
dJ;;
1
EJ n-2+2J
(324)
II III
Cpm
is equivalent to
!n
108
109
I I
= 1-a
(3.26)
:rt)
I'
L (Xin i=1
/I
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
and
(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
= c; and
distributionof X).
/I
( )
1+ -
6 {~-T } .
L (Xii= 1
6-2
T)2
'.
I
-, --
(3.29j
0'
110
111
and (3.34)
=T
'
X}
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)
. ~
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
) )
(3.32)
.~
for Cpmby using the estimate " n{82+(~-T)2}2 .I = 82{82 + 2(~- T)2}
i \
'i
I _rx/2(nc52)
='
1-
IX
where
and
~=A
A;?
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
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
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
!
when
~# T. Here
(3.36a)
,1
Ex.-:'['[(X - T)2]
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
w=
(3.35)
II
n
~+2 9C =(T-LSL)-2
'pm
X,<1'
(Xi-T)2 (3.38)
+(USL'"'
T)-Z
XI>T
L (Xi-
T)2 .
~,:
114
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 ,
115
E[W]
= E[E[W
T)-Z}(Jz (3.42a)
'l
f 'J'
-,
LSL)
"
2 2
XK T
(lJSL ,- I ) ,
' '
'
2 XII
I, 2 K j (J
nJ9)
the two X2S being mutually independent. The conditional expected value and variance of W, given K are
~i
= n[(1'-LSL)-4+(USL-(USLT)-Zp](J4
(JZ)
~= T. Hence
I
I
('
+Z
pill
T)-Z}(Jz
(3.41a)
'1
I
and so
(3.41h)
'I
I
E[(~YJ=1
(3.44 a)
116
and
117
var -
c+ pm
2
'
(( 6;'"
))
var(W)
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
\
Cpmk=
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
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
C~k=
d-IX -ml
I
n-
(3.48)
III
".
(Xi-T)2
and
Xi>T
(Xi-T)2.
.118
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
rct(j+ l))[(l(n-r+j)) .
[(1(n+ 1))
=~t (-1)j ( )
~
dr-jn!(r-j),
3 j=O
r
jj
~ 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'
X { it! (X/-X)Z+n(~
-m)Z
bl ~
(g\
'
(a) Y! =
L (X/i= !
(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
'
to.
I~
(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
Cpm(a) indices
HI d~1 I
121
Cpmk
L
j:O
(- 1).1,
( )( ~-)
J
1'
1
(J
1 r-.i
can be
.I
(3.50b)
tll/l)
E[C~,nkJ=!
9
1 we obtain
]. ==1
[ n-2
-~
( )
(J
~j;I
- ~(n-1)~
( )
(J
dj;I + !
\
I
\
11 (3.501)
dl
HI
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)
(3,50 c)
I~ICo)
~
d
,
(3.51)
,.
.
~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
'
'lMltno
ro
i'cO
[( ,(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
S.D.=S.D.
I(-ml/a
(Cpmd
0.0
0.5
1.0
S.D.
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.
0.264 0.515
...... ..
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
,-
'"-
.-.~---,--
0.084 0.459 0.121 0.762 0.160 . 1.066 0.198 1.370 0.237 1.673
1.0 S.D. E
1.5 S.D. E
2.0 S.D.
2 3 4 5 6 n=70
di(J
2 3 4 5 6
.-----------------.-----iii ~
---
"---
L~
"
=-~.~ ~~-~ ~---_..-
0.001 0.015 0.150 0.021 0.300 0.027 0.44-9 0.034 0.599 0.040
126
d (J
(-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
E[Cpm(a)]
= { 1-
(3.56 b)
E [{C
'
pm
.
'
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
(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
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
Appendix 3.A
129
{( 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
'
'
(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
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 =
- t) and
f~:(Z+V)2(I-V2)!n-2dv (3.61a)
O<y<1
(3.57)
gll(z)={B(ttn-l)}-1
=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)) '
tn-I) (3.61b)
The conditional distribution of Xi, given the sufficient statistics X and S, is that of the linear function of Vj:
,
n-l
X+
j;; SVi
---.
---
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
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'
+(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
,, ..,
III
I
'. , '\\
'\.\ Ii..,.\ '. Ii\.\ \ .. Ii..\ \ -.. '\ "'-;~':'"
~';""'" ~""'>'"
"""
l
-
~~ 30
40
50
' "
~ 0.6 01
0.4
0.2
0.0 -1.0
-0.5
--r0.5
J
1.0
I
II
"
I
'I
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
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
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
(4.1)
I'
lill
'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
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
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,
139
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 \
0.16t-
I'
\\1
---
(3)
Ii)
10
0.00t-5.0
.../' I I
-2.5 0.0
I
'
2.5
.
7.5
5.0
I
I
--r
20
I' 0), I
2.50
0.50 0.75 1.00 1.25 1.50
') I\(\
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
L~N 1'1
50
I I
I
J
we note,
--~~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
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
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
(4.2)
A3
{Ji2(X)}i=JP;
Ji3(X)
(4.3a)
~a)
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)
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
(4.4)
fl
parameter
A~
L ni~j-'o)2]-symbolically,
j= 1
","-'-"." . . ::, ',112 (",", ' '~~> . '.\,.: - '. ;, . ;' . ... '. ~.,. . ,/''," "'" ~,.,
(12X;,\.t~)
(4.7)
1-
142
1\
Effects of non-normality
8~
143
(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
,
I!I/'
E[C~J=L
n
~
( j= 1
TInj!
.iI P?
)(
J=1
)E[C~lnJ
1'=1,2,...
(4.12)
, l~
(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
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!
{ 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.)
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
NN---900000
N-OO',...\o
II,
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
0'
1u.J
qqqqqq
Noooo\O"<t...... O\r-MNN......
......
V)I N
......
000000
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
10 ...... cr.)
1u.J
0\0\0000"<t"<t C"IN........................
000000
- - -'...... 000000
"<t"<tNNr-r: : : : :
"<t"<tNNr-r: : :
~~~~~~
: :
~~~~~~
: :
nll~
w4
to the situation
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)
>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
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
_I.
lint ~I. IfQI'
(4.14 c)
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
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
I
'1"1 Ib) II 1"'f1he
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
If the process is symmetrical (so that fJl =0), E[CpuJ is proportional to Cpu' Table 4.4 (based on Kocherlakota et al.
148
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
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'
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
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
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
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
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
}
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
-I sgn(h)=
{ ,... I
for h> 0
--. . Table
- ~'~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
- 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.).
159
,v
~ ..0
l
-~
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)
i,
'<tM'<t
V)
...c::
u ::;j
'"
<:1:>100 ...... 0
:::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
-,I
.
161
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
~ . .
.
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
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
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
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
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)
"
""
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))'
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
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)
USL-
l'
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)
Cjkp= 3)2max([Ex>T[(X
t.
l'
- LSL
1')2]}!
= 3)2 {Ex<T[(X -
(4.20b)
I
of Cjkp is
1.
Cjkp
= 3)2
mID
(4.25a)
where
8+= Note that
.
Xi>T
(X/-1'f
and 8_=
X,<T
(X1-1')2
==
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 +]
166
I,
Process
Flexible PCls
167
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
()
:,
"
where
I
4 al= -- d
USL~ T
-z
-Z
I' I I.
I
'11
II
i
1
l'
(4.27)
_..~ -.~
-~-
0\
'"
en :.a
.8 ......
h 1\\ ~
H
~ 8 H
h V
~
c.
0 I-.
v
d'
;>.
'08"0"0
8 CIj.
s::
~
<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
.8 ......
~:e~o
;.::::: t> 0\
..0
CIj."",
~
V
0
CIj <1>;:j
bJ)OO
s::
':i--. N
v Q.,. k:
.......
<:,,)
">
:z
<.!..
~
"--'"
-\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 ~
CIj~ ..c::
s::
0 ......
~ gOE~
0
I!)
<I> CIj
b h~
.~
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 ~
..0 ~..c:: 'i: .-; +-> ~ II.~ - O""'N "O':<b 0--0 ~ ..c::
I --
N-
+
b
"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::
=-0
. s::o~'- g, ..0
-g .S 8 0 0
0 ~
+
-<
~ C'! "0
.8
W~I
s:: I
~ I-<
h>f
15; c:..
-b
~.~
'0 ~en L.
<I>_o-fl =,.goClj:es:::e
s:: "'."0
o~
0..0
--
CIj
...... CIj"O t$: INb ..... s:: en w~ ~..2<O 8 >( 0 ; <U 0 H-<t: _0-:: I-<~
_0
h-
CIj
:.: ><
,,
0 ~
.~ t:.~ "0,,00
C'i
-s:: CIj
-0
-0
I-<en=+->
s::
CIj"<f" ~ x.
> 0..
----.----.---.-----.
----
'O"S
---:"=i!"""
.------
6010 P60
zro
091"0
ooor zOor
6LooI
lVOOI 161"0 ~vo.t
LOZoO Z~OOI 9Zz00 0901 Z ~zoO ILOOI L8Z'0 L8001
os
~P
Ov
~ O ~Z
OZ ~Y 01 =u
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
~rz'o 8Pool:
LZoO p~oor
8aoo 9801
pon
90 'lIn 'lIP"O ~n
6VP'0 9n
OS
O] O'S O]
Un
vn
OCO
=p/lL -7Sn)
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
Cpk
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'~ =
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
with expected .~
6;m=
~-m
{1
6m
<p
O'pk= '9
+ 2Y PI Cpk+
<p
ro
'4(fJ2-1)Cpk
(4 .10)
'f
where
<p
.jji;=.0
0'2 pm
--
(-a ) +!
{I+ e:m)';'
~-m
')
and
(4.35)
C;m
172
Appendix 4.A
173
_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
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;
.
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)
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
(4.40)
.and
Val
V Qz
174
Process capability indices under non-normality Bibliography
175
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"
(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))
E[D
]-
2\rnin) 2" az
x {ai\rB1-b(!(n-k),
-!r
,,-1
(~)
I
I
1 I
+ 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.
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
Bibliography
177
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
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.
..
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
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
I'
I
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
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
.
V0 =(p/i'Q"/al)
with
PII
= 1 for
(5.2)
I~
182
Multivariate
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
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)
: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
Multivariate
analogs of Cp
185
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)
but Pr[LSLi ~ Xi ~ USLi for all i = 1,... , v] is not necessarily equal to 99.73%, even if
v .
1=1
(5.7)
.n (USLi-LSLi)=
(nx~.9973)hIVol ntv+l)
(5.11)
X~. 0.9973
'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
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)
.
.I
186
Multivariate
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
K
X",O.9973
( )
=
((n-1)v)j
X(n-l)v
(n-1 )v 1 S
(5.16a)
which is distributed as
= ()2A
then
Xv,O.9973 X(n-l)v
((n-1)v)!
"Cp
Cp
volume of x'A - lX ~ K2
(
(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
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~
(5.17)
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..
s=
L (Xj-X)'A -l(Xj-X) j= 1
I
I
I
,j
,J_1
188
Multivariate
189
I
I
'
I .
Cpm
\\= and V0 by
n j= 1
(Xj-T)(Xj-T)
(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"
.
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,
()
2
nv ! r( t(nv - 1))
l( tnv)
(5.22a)
190
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)
(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)
(5.23)
\
Cpm=Cp{l+(~~TY}
-1'
t
(5.24 a)
Xv,O.9973
x (S+n(X-l)Vo
-,
- 1 if
(A--T)f
1 -!
(5.26)
where
4nv-
-1
In all the previous work, distribution theory has leaned heavily on the assumption that: (5.24b)
I~
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
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
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) C: Pearn.et al. (1992) vCp Taam et al. (1991) qm
Chan et al. (1990) Pearn et al. (1992) Proposall. Proposal 2.
characteristics, and
(5.27)
(x'Vi 1X<C2))
Cpm
. 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
v 0 is the variance-covariance
V 1'= V 0 +(~
--T)(~ -T)',
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,
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
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)
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
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
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
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
!
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
W=(X-T)'A
-l(X-T)
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
\
(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)
(5.31 c)
(5.31d)
198
Multivariate
199
If Y = F(X - ;) then Y has a v-dimensionalmultinormal distribution with expected value 0 and variance-covariance matrix
xi variables.)
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
J= I
and is distributed as 8-2 x (noncentral chi-squared with v degrees of freedom and noncentrality parameter
/I
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
~. 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
l'nJ('('SS
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
,
~
('
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
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
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.
(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
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
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
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
ill"
,
!
*~ :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
indice
\\\\1\\111\\1\111\111111\\111\\
~
m It
[!
!l
1~
lG871R
:~
g