Multivariate Theory and Applications

Quality Control
THEORY AND
APPLICATIONS
(ami1 Fuchs
Tel Aviv University Tel Aviv,Israel
Ron S, Kenett
KPA Ltd. and Tel Aviv University Raanana, Israel
M A R C E L
MARCELDEKKER, INC.
D E K K E R
NEWYORK BASEL HONG KONG
Library of Congress Cataloging-in-Fublication Fuchs, Camil Multivariate quality control: theory and applications / Camil Fuchs, Ron S. Kenett. p. cm. - (Quality and reliability : 54) Includes bibliographical references and index. ISBN 0-8247-9939-9 (alk. paper) 1. Qualitycontrol-Statisticalmethods.2.Multivariate analysis. I. Kenett, Ron 11. Series. TS 156.F817 1998 658.562015195354~21
98-2764 CIP
The publisher offers discounts on this book when ordered in bulk quantities. For more information, write to Special SaledProfessionalMarketing at the address below. This book is printed on acid-free paper. Copyright Q1998 by MARCEL DEKKER, INC. All Rights Reserved. Neither this book nor any part may be reproduced or transmitted in form any or by anymeans, electronic or mechanical, including photocopying, microfilming, and recording, or by any information storage and retrieval system, without permission in writing from the publisher. MARCEL DEKKER,INC. 270 Madison Avenue, New York, New York 10016 http:/hw.dekket:com Current printing (last digit): 10987654321 PRINTED IN THE UNITED STATES OF AMERICA
To Aliza, Amir and Efrat
"CF
To my mentors, DavidCox, Sam Karlin, George B ox and Shelley Zacks who shaped my interest and understanding of industrial statistics. "RSK
Data are naturally multivariate and consist of measurements on various characteristicsfor each observational unit. An early exampleof the requirements multivariate data is provided byNoahs Ark for which were: The length of the ark shall be three hundredcubits, the breadth of itJiJky cubits, and the height of it thirty cubits (Genesis 6:15). The ark was a single specimen producedby a craftsman, and statistical quality control procedures were unlikely to have been applied for ensuring the conformance to standards of the trivariatedata for that product. However, modern industrial processes with mass production are inconceivable without the use of proper statistical quality control procedures. Although multivariate methods are better suited for multivariate data, theiruse is not common practice in industrial settings. The implementation of multivariate quality control involves more complex calculations, creating a barrier to practitioners who typically the simpler prefer and less effective univariate methods. But this barrier can andbe should removed. Modern computers have created new opportunities for data collection and data analysis. Data are now collected en masse and stored in appropriate databases. Adequate statistical tools that transform data into information and knowledgeare needed to help the practitioner to take advantage of the new technologies. Multivariate Quality Control: Theory and Applications presents such tools in the context of industrial and organizational processes.
V
vi
Preface
The main objectiveof the book is to provide a practical introduction to multivariate quality control by focusing on typical quality control problem formulations and by relying on case studies for illustrating the tools. The book is aimed at practitioners and students alike and can be used in regular university courses, in industrial workshops, and as a reference. The typical audience for industrial workshops based the on book includes quality and process engineers, industrial statisticians, quality and production managers, industrial engineers, and technology managers. Advanced undergraduates, graduates, and postgraduate students will also be able to use the book for self-instruction in multivariate quality control and multivariate data analysis. One-semester courses such as those on statistical process control or industrial statistics can rely onthe book as a textbook. Our objective was to make the text comprehensive and modern in terms of the techniques covered, correct in the presentation of their mathematical properties, and practical with useful guidelines and detailed proceduresfor the implementionof the multivariate quality control techniques. Special attention was devoted to graphical procedures that can enhance the book's applicabilityin the industrial environment. The bookpresentsmacroswritten in the statisticallanguage MINITABTM, which can be translated to other languages such S-Plus, as SAS, SPSS or STATGRAPHICS. Access to MINITAB 11.0 or higher is advisable for running the macros attached to the text. Earlier versions of MINITAB, in which the calculations are performed with single precision, are inadequate for the calculations involved in multivariate quality control. The idea of writing a book on multivariate quality control has matured after several years of collaboration between the authors, The structure of the book reflects our understanding of how multivariate quality control should be presented and practiced. The book includes nonstandard solutions to real-life problems such as the use of tolerance regions, T2-decompositionand MP-charts. There is no doubt that with the expansion of automatic data collection and analysis, new problems will arise and new techniques will be proposed for their solution. We hope this book contributes to bridging the gap between theory and practice and expandingthe implementation of multivariate quality control. This book distills years of experience of many people. They in-
Preface
vii
clude engineers and operators from the metal, electronic, microelectronic, plastic, paper, medical, and food industries who were willing to pilot the implementation of multivariate quality control. We thank them all for their willingness to try new procedures that in many cases proved effective beyond expectations. Special thanks Mr. to Ori Sharon for his help in writing some of the MINITAB macros, toMr. Brian Vogel for running the star plots, and to Mr. Russell Dekkerfor ongoing editorial support and continuous encouragement. Camil Fuchs Ron S. Kenett
Contents
Preface
Multivariate with Control Quality 1. Data
1
9 27
2. The Multivariate Normal Distribution in Quality Control 3. Quality Control with Externally Assigned Targets
4. Quality Control with Internal Targets-Multivariate Process Capability 47

Targets from 5. Quality Control with Data with Multivariate Control Charts 6. Analyzing Characteristics 7. Detection of Out-of-Control Regions Approach 8. The Statistical Tolerance Quality Control with Units in Batches 9. Multivariate Principal 10. Applications ofComponents for Multivariate 11. Additional Graphical Techniques Quality Control Multivariate Control Quality 12. Implementing
a Reference Sample
63 75 81 91 101 113 121 135
ix
Contents 145
ontrol
Appendix 1. MINITABTM Macros for Multivariate Quality
Appendix 2. The Data from
the Case Studies
179
195 201
205
Appendix 3. Review of Matrix Algebra for Statistics with MINI TAB^ Computations
References
Index
Quality Control
l
Quality Control with Multivuriute Datu
Objectives:
The chapter introduces the basic issues o f multivariate quality control. The construction of multivariate control charts is illustrated using the data fiom Case Study l on physical dimensions o f an aluminum part. Several approachesto multivariate quality controlarc diswsedand thefour casestudiesused in the book are briefly described.
K e y Concepts
Management styles Quality control 0 Multivariate data collection 0 Multivariate scatterplot matrix 0 The Hotelling l!-distance 0 The T2-chart
0 0
Multivariate Data with Control Quality
Organizations can be managed in different ways. The reactive laison consumer complaints, liability suits, shrinksez faire approach relies ing marketshare or financial losses to trigger efforts to remedy current problems. In this mode of thinking, there are no dedicated efforts..to anticipate problems or improve current conditions. An alternative approach, that requires heavy investments in screening activities, relies on inspection of the product or service providedby the organizations customers. One set of statistical tools that apply to such a screening, or inspection-based management style, is acceptance sampling. Using such tools enables decision makers to determine what action to take on a batch of products. Decisions basedon samples, rather thanon 100% inspection, are usually more expedient and cost effective. Such decisions, however, only applyto the specific populationof products from which the sample was drawn. There are typically no lessons learned by inspection-based management and nothing is done to improvethe conditions of future products. Quality control requires a different style of management. This third management approach focusesthe on processes that make the products. Data collectedis used to track process performance over time, identify the presence of assignable causes and expose common causes that affect the process in an ongoingway. Assignable causes occur at specific points in time and usually require immediate are opportunities for improvecorrective action, while common causes ments with long-term payoffs. Feedback loops are the basic structure for process-focused management. For a comprehensive treatment of industrial statistics in the context of three management styles: inspection based, process focused quality control, and proactive quality by design, (1998). see Kenett and Zacks Quality control is based on data sequentially collected, displayed and analyzed. Most data are naturally multivariate. Physical dimensions of parts can be measured at several locations and various parameters of In practice, one will fresystems are typically derived simultaneously. quently use a critical dimensionas the one to record and track for purposes of quality control.An alternative approach to handle multivariate data is to aggregate several properties and transform data collected by several measurement devices into attribute data. Under this approach, a unit is labeled as passor fail. Trackingthe percentage of defective products or nonconformities is indeed a popular approach tostatistical
Chapter 1
quality control. However, the transformation of continuous quantitative data into qualitative attribute data leads to a loss of information, This reduces the ability to identify the occurrence of assignable causes and impairs the understanding of the process capability. In this book we present tools that provide effective and efficient analysis of multivariate data without suchloss of information. Modern computers and data collection devices allow for much data to be gathered automatically. Appropriate tools that transform data into information and knowledge are required for turning investments in automation into actual improvements in the quality of products and processes. The main objective of this book is to provide a practical introduction to such tools while demonstrating their application. Our approach is to de-emphasize theory and case use studies to illustrate how statistical findings can be interpreted in real life situations. We base the computations on the MINITABTM version11 .O computer package and present in some instances the code for specific computations. Previous versions of MINITAB did not havethe double precision capability and caused numerical errors in several of the computations used in this book. The use of earlier versions of MINITAB is therefore not recommended for these kind of multivariate computations. Case Studies 1 to 4 are analyzed in Chapters 3 to 1 1. They provide examples of statistical quality control and process capability studies. The data of Case Study 1 consist of six physical dimension measurements of aluminum pins carved from aluminum blocks by a numerically controlled machine. Case Study 2 is a complex cylindrical part with 17 variables measured on each part. Case Study 3 comes from the microare used with very stringent electronicsindustry were ceramic substrates requirements on critical dimensions. Case Study 4 consists of chemical properties of fruit juices that are tested to monitor possible adulteration in naturalfruit juices. We only deal indirectly with issues related to the complexities of multivariate data collection. In some cases major difficulties in collecting multivariate data are an impediment to statistical quality control initiatives (D. Marquarclt, personal communication). For example, a sample can be subjected to a variety of tests performed on various instruments located in several laboratories and requiring different lengths of time to
complete. Under such conditions, collecting the data in a coherent data base becomes a major achievement. Univariate quality control typically relies on control charts that track positions of new observations relative to control limits. A significantly high deviation from the process average signals an assignable cause affecting the process. Multivariate data are much more informative than a collection of one dimensional variables. Simultaneously accounting for variation in several variables requires both an overall measure of departure of the observation from the targets as well as an assessment of the data correlation structure.A first assessmentof the relationships among the variables inthe multivariate datais given by the analysisof their pairwise association, which can be evaluated graphically by dis1.1presents an illustration of such playing an array of scatterplots. Figure a scatterplot matrix, computed for the six variables fromCase Study 1 mentioned above. the variables. The data display a clear pattern of association among For example,we observe inFigure 1.1 that thereis a close relationship between Diameter1 and Diameter2 but a rather weak association be-
# + . & e@
993
Q @
% @
&@ +$@
@ &
& ,
@$
@ @
Figure 1 .l :Scatterplot matrixfor the 6 variables fromCase Study 1 (ungrouped).
Chapter 1
tween Diameter1 and Lengthl. A significantly largedistance of a p dimensionalobservation, X from a target m will signal that an assignable cause is affecting the process. The pattern of association among the variables, as measured by their covariance matrix has tobe accounted for when an overall measure of departure from targets is to be calculated. A measure of distance that takes into accountthe covariance structure was proposedby Harold Hotellingin 1931. Itis called HotellingsT2 in honor of its developer. Geometrically we can view T2 as proportional to the squared distance of a multivariate observation from the target where equidistant points form ellipsoids surrounding the target. The higher the T2-value, the more distant is the observation fromthe target. The target vector m can be derived either from a previouslycollected referencesample from the tested sample itself (i.e., as an internal target), or from externally set nominal values (i.e., as an external target). The T2-distance is presented in Chapter 2 and some of its theoretical properties are discussed there.The T2 distances lend themselves to graphical display and the T2-chart is the classical and most common among the multivariate control charts (Eisenhart et al., 1947; Alt, 1984). It should be noted that the chart, like T2 itself, does not indicate the nature of the problem, but rather that one exists. Figure 1.2 presents the T2-chart for Case Study 1. The vertical axis indicates the position of the 99.7-th and 95th percentiles of T2. We shall expand on issues related to that graph in Chapters 3 to 5 that discuss, respectively, targets that are derived externally, internally and from a reference sample. Multivariate control charts are further discussed in Chapter 6. The multivariate analysis assessesthe overall departureof the p variable observations from their target. However, by themselves, the analysis of the T2-statisticsand of the appropriate controlchart do not provide an answer to the important practical questionof the detection of the variables that caused the out-of-control signal. Chapter 7 presents some approaches and solutions to this issue. We present in Chapter 8 the statistical tolerance region approach to qualitycontrol and its related concept of natural process region. This approach is different from the classical Shewhart approach that relies shifts in primarily on three standard deviations control limits to detect the mean or the variance of a distribution. Statistical tolerance regions
" " " " " " " " " " "
"-
31
I 40
I 50
I 60
I 70
Observations
Figure 1.2: T2-chartfor the 6 variables from Case Study 1 (ungrouped).
consist of a rangeof values ofthe measured variables spanning a given percentage of the population corresponding to a process in control. This statement is made with a specific level of confidence. For example, one can construct an interval of values for which we can state, with a confidence of 0.95, that this range includes 90% of the population from a process under control. Such regions permit identification of shifts in distributions without the limitation of singling out shifts in means or variances as in the classical Shewhart control charts. In some cases, the analyzed data are presented in batches and the quality control procedure should take this feature into consideration. Methods of analysis for this situation are presented in Chapter 9 and applied tothe data of Case Study3. Another approach to multivariate quality control is to transform a p-dimensional set of data into a lower dimensional of setdata by identifying meaningful weighted linear combinations of the p-dimensions. In Chapter 1 0 we Those new variablesare called principal components.
Chapter 1
present the principal components approach and illustrate it using Case Study 1. Both the T 2distances and the principal components lend themselves to graphical displays. In Chapter 11 we compare two additional graphical displays that isolate changes in the different individual dimensions. Modern statistical software providemany options for displaying multivariate data, Most computerized graphical techniques are appropriate for exploratory data analysis and can greatly contribute to process capability studies. However, ongoing quality control requires graphical displays that possess several properties as such those presented and illus11 with the data from Case Study 1. For completeness trated in Chapter we also review in Chapter 11 modern dynamic graphical displays and illustrate them withthe data from Case Study 3. We conclude the text with some guidelines for implementing multivariate quality control. Chapter 12 provides specific milestones that can be used to generate practical implementation plans. The Appendices contain a descriptionof the case studies and theirdata and references (Appendix 2), a brief reviewof matrix algebra with applications to multivariate quality control, using the MINITAB computer package (Appendix 3), and aset of 30 MINITAB macros for the computation of statistics defined throughout the book (Appendix 1). The macros provide an extensive set of tools for numerical computations and graphical analysis in multivariate quality control. These macros complement the analytical tools presented in the book and support their practical implementation.
The Multivuriute Normal

Distribution in Quulity Control
Objectives:
The chapter reviews assumptions and inferential method in the multivariate normal case. Results thatprovide the theoreticalfouradationsfor multivariate qualiv control procedures are stated and demonstrated. Simukations are used to illustrate the kcy theoreticalproperties of the bivariate normal distribution.
K e y Concepts
The multivariate normal distribution Maximum Likelihood Estimators 0 Distributions of multivariate statistics 0 Base sample 0 Data transformations
0 0
Distribution TheNormal Multivariate
l1
A multivariate observation consists of simultaneous measurement of a numerically of several attributes. For example, a part coming out controlled metal cutting lathe can be measured on several critical dimensions. Multivariate observations are conveniently represented by vectors with p entries corresponding, to the p variables measured on each observation unit.This chapter discusses inferential methods applicable to multivariate normal data. In most of the chapter we assume multivariate normality of the data and present methods which enable us to produce estimates of population parameters and construct tests of hypotheses samabout multivariate population parameters using a simple random ple of multivariate observations. The analysis of multivariate data poses a challenge beyond the naive analysis of each dimension separately. One aspect of this challenge stems from the simultaneous consideration of several probability statements, such as p-tests of hypothesis. This creates a problem of multiple comparisons that requires adjustmentsof univariate significance levels so as to attain a meaningful overall significance level. A second component of the multivariate data analysis challenge is to account for the internal correlation structure among the p-dimensions. This again affects overall significance levels. We begin by describingthe multivariate normal distribution, proceed to coverthe distribution of various test statistics, and then briefly mention approaches to induce normality into non-normal The chapdata. ter provides a notation and a theoretical foundation for the following chapters.
The Multivariate Normal Distribution. A p-dimensional vector of random variables X = (X(), .. . , X()); 0 0 < X ( e ) .c 0 0 ,l = 1, , , . , p, is said to have a multivariate normal distribution if its density function,f(X) is of the form:
f(X) = f(X(, , * * ,x
exp
[ - $(X.
q = (ax)-flZl-f
p)Z:-l(x- p )
where p = ( p ( ) , , . . , p())is the vector of expected values p(e) = E(X()),l = 1, , , ., p and E = [(ne)] 4,U = 1, , .., p, is the variancecovariance matrix of (X(), , , , X ( P ) ) , neN = cov ( X ( e ) , X)) and Gel = 2 .
12
Chapter 2
We shall indicate that the density of X is of this form by
where N,,(-,denotes a p-dimensional normal distribution with the respective parametersof location and dispersion. When p = 1, the one dimensional vector X = (X('))' has a normal ; , i.e. distribution with meanp ( ' )and variance a
a)
or
X(2))' has a bivariate normal distribution When p = 2, X = (X('), p(2)) and a covariwith a two-dimensional vector of means p = (#U('),
ance matrix C = a ;
021
a ' ; ] where a ; and a ; are the variances of X(')

02
and X ( 2 ) ,respectively, and012 = a 2 1 is the covariance between X(')and X ( ' ) . Let p = be the correlation betweenX(')and X ( 2 ) , If X(') and X ( 2 )are independent ( p = 0)theirjoint bivariate normal distribution is the product of two univariate normal distributions, i.e.
In the general case (- 1 5 p 5 l), the bivariate normal distributionis given by
We can define a set of values which includes a specific proportion of multivariate distribution. Such a set of values forms a natural process
The Multivariate Normal Distribution
13
092
Simulated Length l
Simulated Length 2
Figure 2.la: Bivariate normal density with parameters of Length 1 and Length 2 from Case Study 1.
region. The values contained in the natural process region are characterized by the fact that their distances from the mean vector p do not exceed the critical values which define the specific proportion of the population. The classical 3-sigma ruleis equivalent, in this context, to testing whether an observation is contained in the central 99.73% of the reference population(P = ,9973). In the univariate case,the respective critical values are called natural process limits (e.g. ASQC, 1983). Figures 2.1 and 2.2 presents a bivariate normal distribution whose means, variances and correlation coefficient are equal to those found empirically in the bivariate distribution of Length1 and Length2 in the first 30 observations of Case Study1,where p(1)= 49.91, p(2)= 60.05, 01 = 0.037, 02 = 0.037, p = 0.723. Figure2.1 is a3-Drepresentation while Figure 2.2 displays contour plots for several values of f(Length1, Length2). Whenever the parameters of the distribution are unknown and have to be estimated fromthe sample, as is usually the case, we cannot stipu-
14
Chapter 2
60.1092
60.08256
v
60m264 59.976
Simulated
Lenath l
60.05592y
Figure 2.1 b: Bivariate normal density (rotated).
late with certaintythat the natural process region will contain the stated proportion of the population. Since we now have an additional source of uncertainty, any statement concerning the proportion of the population in a particular region can only be made with a certain level of confidence. Each combination of a stated proportion and a given confidence level define a region, The newly formed statistical tolerance region is used in this case instead of the natural process region. In the univariate case the natural process limits are replaced by statistical tolerance limits. The statistical tolerance regions approach to multivariate quality control is illustrated inChapter 8. The shape of the multivariate distribution depends on the mean vector p and covariance matrix E.When one or both of these parameters are unknown, as is usually the case, they have to be estimated empirically from the data. Let XI, , , ,X , be n p-dimensional vectors N,(p, E)and p 5.n - 1. of observations sampled independently from The observed mean vector and the sample covariance matrix S are given by
15
,9396
Figure 2.2: Contour plotof bivariate normal density with parameters of Length 1 and
Length 2 from Case Study1.
S=
C(Xi- X)(X, - X ) ' / ( n - 1) ,

n
i=l
areunbiasedestimates of p andrespectively. We note thatunderthenormaldistribution,theobservedmean vector is also the maximum likelihood estimator (MLE) of p, while the MLE for 'I=, is S. The ( l ,l')-th element of the S-matrix is the estimated covariance between the variables l and l',l,l' = 1,2, , p , i.e.
. ..
16
Chapter 2
The diagonal elements of S are the corresponding sample variances. When a covariance is standardized by the appropriate standard deviations, the resulting estimated value of the correlation coefficient assesses the association betweenthe two variables. If the sample is composed of k subgroups, each of size n, and if the mean andthe covariance matrixin the j -th subgroupare Xj and , , - S respectively, j = 1,2, , , , , k, then we can define the grand mean % and the pooled covariance matrix as follows:
We illustrate the computations with two simulated data sets. The first sample was generated under a bivariate normal distribution with ungrouped observations. The observations in the seconddata set have four variables and are grouped in pairs, i.e. subgroups of size two. The advantage of the use of simulated data for illustrating the theory is that the are known and wecan thus underlying distribution with its parameters observe the performance of test statistics under controlled situations. In the first data set, the parameters of the first 50 observations were identical with the values calculated empirically for Length l and Length 2 in the first 30 observations of Case Study 1, i.e.
and 01 = 0.037, a 2 = 0.037, p = 0.723. The population covariance matrix 10,0000 I :is thus
a
I:=[13.69 9.90 9.90 13.69
Normal Multivariate ibution The
17
Those first 50 observations simulate an in-control base sample (or in-control process capability study).The sample mean for those observations was - 49.9026
x = [60.04411
and the elements of the S-matrix are given by

50
811
= C(X!
i=l
50
- 49.9026)2/49= ,00122 - 60.0441)2/49= ,001367 ,

,001355
812
= 821 = C(X, - 49.9026)(X,2- 60.0441)/49 = .00114

822
= C(X,2
i31
i=l 50
i.e.
BO1123 S = [ .001204 .001123
.Appendix3 contains a brief reviewof matrix algebra with applications to multivariate quality control, as well as a review of the computations which can be performed usingthe MINITAB computer package.If the data are in columns 1 and 2 (or, in MINITAB notation,in Cl and C2), the MINITAB command which leads to the computation of the S-matrix is simply COVARIANCE Cl-C2, and we obtain the S-matrix presented above. We mention that throughout the book the computations can be performed both with MINITAB as well as with other statistical software. Since versions of MINITAB before version 1 1.Ostore the data with only six significant digits, the precision is lower than that achieved by software which uses double precision. The results of the computations presented in the book are those obtained by MINITAB version11 .O. From the entries in the S-matrix sample correlation coefficient between the variables is given by
r=
.001140 = .883 * 4.001220 * 001367
18
Chapter 2
For the second data set, the parameters for the first 50 subgroups were identical with the values calculated empiricallyfor the four diameters i.e. Diameter 1 - Diameter 4 in the first 30 observations of Case Study
and the covariance matrix 10,000 E is

9
1.7080 1.8437 1.8529 1.8460 1.8195 1.8529 2.1161 1.9575 L 1.8264 1.8460 1 .g575 2.3092
1.
The underlying distributionof the simulated data was multivariate normal with the above mentioned parameters. The data were groupedin 50 subgroups of size 2.Thus, the first group had the observations
X1= k.9976 9.9830 9.9804 14.98481

Xt2= k.9553 9.9574 9.9543 14.94921
Its mean vectorof the first subgroup is
X 1 = k.9765 9.9702 9.9673 14.96701

and the covariance matrix within the subgroup (~10,000)is
8.9558 5.4181 S 1 = 5.5307 7.5527
5.4181 3.2779 3.3460 4.5692
5,5307 3.3460 3.4156 4.6642
7.5527 4.5592 4.6642 6.3693
The computations are performed similarlyfor the 50 subgroups and we obtained
50
X = x x j / 5 0 = k.9864, 9.9793, 9.9752, 14.97681

jel
19
and 10,000 S p = 10,000 * C(50- 1)Sj/[50 * (2 - l)]

j=l
50
1.2151 1.1428 1.1600 1.2669
1.1428 1.3559 1.2854 1.2792
1.1600 1.3288 1.5728 1.4183
1.2669 1.3033 1.4183 1.6854
A MINITAB program which computes the S,,-matrix canbe easily designed if we define a macro which computes for each group its(n l) 5''-matrix and sums them up.In the macro presented in Appendix 1.1, the number of parameters ( p ) is stored inK2 (in ourcase K2 = 2). The group size (n)is stored inK3 and the number of groups(k)is stored in K4. The indices of the observations in the current group are from K5 to K6. The macro (named POOLED.MTE3) is given in Appendix 1.1. The program POOLED.MTB, can be run by typing
MTB> EXEC 'POOLED.MTE3' 1.
The theoretical properties of the estimates X,2, S and S,,, have been extensively studied and were determined to be optimal under a variety of criteria (sufficiency, consistency, and completeness). of those estimates for a random normal The distribution properties . . ,X,,are given bythe following theorems (e.g. Seber, 1984): sampleXI,
(ii) If X is distributed as in(i) above, then
n(X - p)'X"(X
- p ) x;
where x; is the Chi-square distribution with p degrees of freedom. (iii) (n - 1)s W,,@ 1, E) where W,,(., stands for the Wishart distribution which is considered to be the multivariate analogue of the Chi-square distribution. (iv) If Z and D are independent random variables distributed respectively as
m)
20
Chapter 2
then the quadratic form
is distributed as
T2
-f
fP -P+l
FP.f -p+]
where Fp,f-,,+l is theFisher F-distribution with ( p , f -p+1) deof Z, denoted by p,, differs grees of freedom. If the expected value from zero, thenT 2 (multiplied by the appropriate constant) has a non-central F-distribution with parameter of centrality p', E;lp,. (v) Let Z N,,(O, XL)and f D W,,( f,X,), f > p where f D can be decomposed as f D = (f - 1)Dl ZZ', where (f - 1)Dl W,,( f - 1, X), and Z is independent of Dl. Then the quadratic form T 2= Z'D"Z
is distributed as
T 27 f B(P, f - P) where B(p, f - p) is the central Beta distribution with (f, f - p) degrees of freedom, respectively. If the expected value of Z, p, differs from zero, then T 2 has a non-central Beta distribution with parameter of non-centrality
P', X + , .
Z is independent Note that unlike (iv) above, do not we assume that of D, but rather that it is independent of Dl. (vi) If the sample is composed of k subgroups of size n all originated from the distribution defined above with subgroup means j = 1, ,k and grand mean denoted by i.e.
...
x,
xj,
21
/&c-
Xj - X
Np(O, X)
(vii) Under the conditions of (ii) above, if YI, Y n are an extra subgroup from the same distribution, then
. ..
(viii) If the sample is composed of k subgroups of n identically distributed multivariate normal observations, and if S is the sample covariance matrix from the j-th subgroup, j = 1, .,k, then
..
X(n - 1)Sj
W,(k(n- l ) , X)
Note that this additive property of the sumof independent Wishart variables, is an extensionof that property which holds for univariate Chi-square variables. The distributional properties of S and T2provide the theoretical basis for deriving the distributions of the statistics used in the multivariawequality control procedures discussed in subsequent chapters. Those statistics assess the overall distance of a p-dimensional vecm= (,(l), m(2), . ,m(,,>, tor of observed means from the target values
z,
..
Let Y : ) , i = 1 , . , n; L = 1, . , p ben multivariate measurements and let F = (F(,), ,B@)) be observed means. Up to a constant, the appropriate Hotelling statistic T is computed by multiplying Y -m i.e. the transposed vector of deviations between
..
. ..
..
Y and m,by the inverse of a certain empirical covariance matrix (S) and then by the vector of derivations Y -m .The Hotelling T2-statistic (- )
has already been mentioned in the previous chapter where we presented the chart of the values obtained from the first 30 observations from Case Study 1. For the case when a sampleY 1, ,Y n of n independent normal p-dimensional observations is used for assessing the distance between
.. .
22
Chapter 2
the meany and the expected values p, the Hotelling T2-statistic,denoted by T i , is given by T i = n(v
- p)S(v -p) .
The T i statistic can be expressed in terms of the means of the : ) as measurements Y
where is the (a, a)-th element of the matrix S-1, A MINITAB program which computes the T2-statistics has to invert the S-matrix computed above and to compute the quadratic form by multiplying the appropriate matrices. In Appendices 1.2-1.3 we present several macros for computing T2-statistics for the various cases detailed in the next chapters. Under the null hypothesis that the dataare independent and normally distributed with the postulated mean vector p, it follows immediately from (iv) above that the Ti-statistic justdefined is distributed as
When the covariance matrix was calculated from n observations and we want to assess the distance of a single observation Y from the expected valuep, the Ti statistic is obtained asT i = (Y - 1)S-l (Y Another additional important special case is when p = 1. The T i of the mean of n observations statistic used for assessing the deviation from the expected value reduces in this case to the square of the t statistic t = which is the statistic used for testing hypotheses on the mean of a univariate normal population. If the expected value of Y is p, then the t-statistic has a Student t-distribution with n - 1degrees of freedom and since the coefficient of T$ reduces to1, then t 2 F1,n-l as expected.
P)
Normal bution Multivariate The
23
We illustrate the computations and the hypothesis with the first simulated data set presented above. We mentioned that the first 50 observations in the data set simulate data from an in-control process capability study. The underlying distribution was normal with
49.91
p = p o = [60.051
We computed the T i values for those 50 observations and compared them withthe appropriate critical values from the distribution of F p , n - p , which in this case (for p = 2,n = 50)are 13.97, 12.35 and 6.64for a! = .0027, a ! = ,005and a! = .05, respectively. The results are presented in the last column of Table 2.1.We mention that noneof the Ti-values in the sample exceeded even the critical value a! = at,05. This result could have been expected, since we tested the distances of of the distribution from the empirical data from the location parameters which the data were generated. The only source of observed differences is thus random error. the In next sections, we shall return to this example, this time using an extended version of the data set, which simulated both a base sample as well as some tested samples. When a sample of kn independent normal observations are grouped into k rational subgroupsof size n, then the distance betweenthe mean Y j of the j-th subgroup and the expected values p is computed by TL
is Note that unlike the ungrouped case, the estimated covariance matrix now pooled fromall the k subgroups. Again, under the null hypotheses the data are independent and normally distributed with the postulated mean vectorp1 . It follows from (iv) and (viii) above that
In the case of the second set of simulated data with p = 4, k = 50, n = 2,the critical valuesfor T i are 10.94, 18.17 and 20.17for a! = 0.5, a ! = ,005and a ! = ,0027, respectively. As we can see from Table 2.2, among the 50 subgroups, four of the Ti-values exceeded the critical
TABLE 2.1
Means with external values(49.91,60.05)-the parameters usedto generate the data. The S-matrix from the base sample (50 observations). The data are the base sample:
VARl
l 2 3 4 5 6 7 8 9 10 l1 12 13 14 15 16 17 18 l9 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50
1.6073 49.8585 4.2488 49.8768 49.8706 5.3767 49.91 17 4.5009 49.8470 0.4419 49.8883 0.2205 49.91 58 0.5629 49.91 52 2.2046 49.9055 1.0951 49.8969 4.9861 49.9137 2.9972 49.8586 1.9884 49.9514 0.0122 49.8988 5.5016 49.8894 1.5848 49.9403 1.2866 49.9132 2.6554 49.8546 2.6315 49.8815 5.3803 49.831 1 49.8816 2.5277 49.8501 7.0252 49.9778 49.8690 49.8779 1.0088 49.8680 49.9388 0.3785 49.9133 49.9120 0.7337 49.9250 49.9442 49.8386 54 2.4121 49.9492 49.9204 1.5246 49.8994 1.0678 49.8703 49.8846 0.2658 2.7294 49.9580 0.1939 49.8985 1.1778 49.9397 49.8741 7 0.8206 49.9140 49.9501 49.8865 0.6997 49.8912 0.2766 49.9252 0.8988 0.7513 49.9326 49.9680 4.4314 0.5856 49.9289 0.3528 49.9233
VAR2
60.0008 59.9865 60.0055 60.0126 60.0165 60.0216 60.0517 60.0673 60.0726 60.0208 60.0928 59.9823 60.0866 60.0402 60.0720 60.0681 60.0350 60.0145 59.9982 59.9963 60.0457 59.9860 60.0875 60.01 59 60.0055 60.0088 60.071l 60.0634 60.0560 60.0749 3.8446 60.1 100 3.81 59.9725 60.1014 1 60.0803 60.0625 60.0219 60.0271 60.0878 60.0329 60.0826 1.091 60.0061 60.0401 60.081 0 60.0169 60.0406 60.0532 60.0741 60.1219 80.0709 60.0632
T2#
1.0904
l.8579
0.9596 1.2957 1.2284 0.1053
S392
2.0330
TABLE 2.2
Means with external values (9.9863,9.9787,9.9743 and 14.9763)the parameters used to generate the data. The Spooled-matrix from the base sample (SO groups of 2 observations). The data are the base sample:
Group
1 2 3 4 5 6 7 8 S 10 11 12 13 14 15 16 l? VARl 9.9765 . . . 9.9936 9.9870 9.9962 9.9774 9.9710 9.9813 9.9817 9.9887 9.9861 9,9997 10.0076 9.9725 9,9902 9.9786 9.9903 9,9975 9.9878 9.9679 9.9928 9.9819 9,9894 9.9891 9.9884 10.0112 9.9776 9.9693 9.9833 9.9975 9.9843 10.0107 9.9689 10.0061 9.9755 gig928 9.9789 9,9645 10.0007 9,9844 9.9924 9.9909 9.9770 9.9923 9.9696 9.9861 9.9774 9.9757 9.9696 9,9809 9.9886 VAR2 9.9702 9.9851 9.9719 9.9866 9.9718 9.9683 9.9733 9.9774 9.9842 9.9789 9.9980 9.9969 9.9648 9,9814 9,9720 9.9893 9.9870 9.9881 9.9840 9.9839 9.9742 9.9813 9.9810 9.9783 10.0054 9.9686 9.9793 9.9761 9.9877 9.9760 10.0034 9.9619 9.9972 9.9673 9.9869 9.9735 9.9594 9.9934 9.9729 9.9830 9.9858 9.9684 9.9865 9.9651 9.9771 9.9714 9.9708 9.9621 9.9781 1.4853 9.980914.9796
~ ~ ~~
la
19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 38 37 38 39 40 41 42 43 44 45 46 47 4a 49 50
VAR3 9.9673 9.9893 9.9693 9.9777 9.9651 9.9585 9.9729 9.9731 9.9810 9,9812 10.0001 9.9939 9.9553 9.9770 9,9669 9.9788 9.9872 9.9841 9.9774 9.9801 9.9668 9.9741 9.9751 9.9765 10.0015 9.9665 9.9793 9.9688 9.9816 9.9796 10.0013 9.9582 9.9930 9.9613 9.9832 9.9651 9.9503 9.9880 9.9694 9.9788 9.9800 9.9659 9,9851 9.9587 9.9773 9.9678 9.9690 9.9602 9.9667 9.9728
~ ~~ ~
VAR4 14.9670 14.9813 14.9661 14.9856 14.9716 14.9638 14.9706 14.9723 14.9787 14.9761 14.9935 14.9985 14.9590 14.9781 14.9630 14.9900 14.9867 14.981 3 14.9750 14.9826 14.9737 14.9772 14.9782 14.9722 15.0117 14.9683 14.9604 14.9705 14.9883 14.9826 15.0057 14.9452 14.9919 14.9686 14.9870 14.9739 14.9525 14.9876 14.9733 14.9822 14.9749 14.9612 14.9888 14.9650 14.9802 14.9749 14.9701 14.9630 14.9683
~~~~
T2M
1.8495 8.5031 10.2566 3.5200 3.7810 7.571 l lB984 l.l 358 1.541 0 3.7382 10.9735 7.7374 4.8619 0.6139 3.3072 12.4277 3.8160 5.3670 2.41 78 0.7681 0.9384 0.7797 0.3178 4.0895 16.3026 1.a331 1.4995 0.8288 2.8201 6.2309 10.8461 19.6493 8.5064 4.2052 1.7318 5.2074 9.0034 4.7334 1.2992 0.7626 4.9135 4.4176 3.0087 7.9559 1.g193 5.5854 3.4430 5.9166 3.3282
26
Chapter 2
value atQ = .05 (for the 1lth, 16th, 25th and 32nd subgroups). The Ti for the 32nd subgroup exceeded the critical valuesQat = ,005 as well. The methods of analysis described throughout this text tacitly assume thatthe distribution of the analyzed datais fairly well represented by the multivariate normal distribution. Proceduresfor testing the hypothesis of multivariate normality include both analytical methods and graphical techniques (e.g., Mardia, 1975; Koziol, 1993). When the data present clear evidence of lack of normality, we may try to transform the data using, for example, a multivariate generalization of the Box and Cox (1964) transformation. In the univariate case, the Box and Cox transformation improvesthe adherence to normality of a variable X by finding an appropriate constanth and transformingX into X(A) where
On the method for finding h see Box and Cox (1964) or Box, Hunter and Hunter (1978). Andrews (1971) generalized this approach to the multivariate case by introducing a vectorh = (AI, . , .h,,) and the corresponding transformeddata vector
(X(A1)
I
...(
which is constructed so that X@) is approximately distributedas N p(p, E).In the univariate case we can find a transformationg(X) that will transform data to approximate normally up to any level of precision. One should note, however, that inthe multivariate case an appropriate transformation might not exist [Holland (1973)l. We thus continue to consider in the following chapters that either the original data are approximately normally distributed, or that the data have been preliminarily transformed and, on the new scale, the observations have, approximately, a multivariate normal distribution,
3
QuaLip ControL witb Externally Assigned Targets
Objectives:
The chapter provides methods of analysis f i r the cae in which the targets are assigned external4 and are not based on data analysis. Situations in which external targets are appropriate are discussed and the computations involved are presented in detail wing simulated data.
K e y Concepts
Target values Multivariate process control 0 Multivariate hypothesis testing 0 External d u e s 0 Tested sample 0 Data grouping 0 Decomposition of the Hotelling l?-distance
0
27
Quality Control with Externally Assigned Targets
29
We mentioned in Chapter 1that the statistics involved in multivariate process control depend on the origins of the target valuesm. The methodology and the applications presented in this chapter are for the case when the target valuesare assigned externally, i.e. from external an requirement or when a prespecified standard m0 has tobe met. It should be noted that comparisons with externally specified targets are not usually considered in statistical process control since such targets do not account properly for valid process characteristics. However, whenever multivariate prespecified standards are set, multivariate quality control procedures can be used to assess whether the multivariate meansof the products equal the external targets mo. The next two chapters deal with cases in which target values are derived and com/puted internally. We shall distinguish between targets derived from the tested sample itself, and targets which are derived from a reference or base sample. If we denote by p, the multivariate meanof the productsY,the statistical analysis in quality control with external targets is equivalent to testing hypotheses ofthe form H 0 : p = m0 against
H a :p#mo.
Given a sample of size n l from the population, we can compute Ti = n l - m0)S-I - mo). We have seen in the previous chapter that for normally distributed data, if m0 is the expected value of then F0 = = T i Fplnl-p. the statistic T i has a distribution given by Thus, under the null hypothesis mentioned above, F 0 Fp1,nI-p. Under Ha, F 0 has a non-central F-distribution. The critical value for T i is thus
(v
(v
- v,
where is the upper 100%percentile of the central F-distribution with ( p , nl - p ) degrees of freedom. If the value of Ti exceeds the critical value, we conclude that the derivations between P and the external targetsm0 cannot be explained by random error. In such cases an assignable cause affecting the process is to be suspected.
30
Chapter 3
As a first example we shall use an extended version of the first simulated data set presented in the previous chapter. Now, in addition to the base sample of 50 observations presented in Chapter 2, we generated 25 bivariate observations which simulate a series of tested samples, whose parameters are as follows: the parameters of observations 51 55 are identical to those of the base sample (i.e. the data are in-control); the population meanfor the first component in observations 56 - 65 has been shifted upward by two standard deviation i.e.
66 - 75, the population mean of the first compofinally, in observations nent was shifted downward and that of the second upward, both by one standard deviation, i.e.
We recall that the first 50 bivariate observations were generated from the normal distribution with the mean vector
49.91 = 60.05
We shall consider two sets of target values:the first set satisfiesmi) = po, i.e. the means used for generating the data are taken as target values; are in the second set, the target values
m f ) = 60,0]
We start by testingthe hypothesis that the generated data originate from the population with the prespecified targets, m i.e. = m!), If we further assume that the covariance matrixis known (as it is indeed the case in the simulated data), we can compute for each observation in the tested sample the statistic
[49.9
(YI - rnit))W1(YI -m!))
, i = 51,. . . ,75
Quality Control
with Externally Asslgned Targets
31
and comparethose values with the critical values from the chi-squared distribution with two degrees of freedom (see property (vi) in Chapter 2). Alternatively, the covariance matrix C can be estimated fromthe "base" sample by the empirical covariance matrix S. The resulting T i statistics have then to be compared withthe critical values based on the F-distribution which are presented in this chapter. We now perform for single observaeach of the 25 observations from the tested sample, (i.e, tions Y)both the multivariate T2-test based on the Ti-statistic as well as the univariate t-test for each of the two components.We recall that we performed in Chapter 2 the corresponding multivariate tests for the first 50 observations fromthe "base" sample. In the multivariate test, the null hypothesis is that the mean of the specific bivariate observation being tested satisfies:
m = , : ) = [
49.91 60.05
1,
F
while in theunivariate test, the hypotheses are stated for each component F2,48 separately. The critical values for T i from the distribution of are 13.693, 12.103and 6.515 for a! = .0027, a! = ,005and a! = .05, respectively. The corresponding critical valuesfor the two sided t-test are It1 = 3.16,2.94 and 2.01 for the values of a! as above. The results are presented in Table3.1, For each of the three subsets of the tested sample (of sizes 5 , 10 and 10, respectively), we summarize in Table 3.2 the number of observations for which the null hypothesis has been rejected by the three statistics. The first subset of five observations, which are in control, behave as expected andfor none of them, theTi exceeded the appropriate critical values. In the second subset of 10 observations, only the mean of the first component has been shifted (by two standard deviations). We see that the empirical power of the multivariate T i test exceeded that of the univariate test performed specifically on the affected component.The single observation for which the second component exceeded the critical I error sinceits mean has not value ata! = .05 is to be considered a type been shifted. Let us now consider the second set of target values (which are defined as the means used to generatethe data but this time with only one significantdigit), i.e.
32
Chapter 3
TABLE 3.1
Means with external values(49.91,60.05) which are the parametersused to generate the data. The S-matrixf r o m the base sample(50 observations). The data are the tested samples:
VARl 51 60.0417 49.8798 49.9208 52 49.9606 53 49.9498 14 55 49.8390 49.9284 56 49.9648 57 49.9780 58 50.0218 59 50.0606 60 50.0365 61 49.9756 62 49.9840 63 64 50.0028 65 49.9770 49.8579 66 49.8997 67 49.9156 68 49.9258 69 49.8384 70 49.8937 71 72 49.8631 72 49.9406 49.9046 74 75 49.8718 VAR2 60.0292 60.1172 60.0543 59.9665 60.0079 60.0482 60.0186 60.0854 60.1399 60.1005 60.0387 60.0857 60.0482 60.0278 60.0588 60.0820 60.1415 60.1132 60.0449 60.0893 60.0757 60.1298 60.0739 60.0676
t VARl
t VAR2
TZM
-0.8654 1.6092 -0.2258 -0.5635 0.3100 1 .a164 1.4490 0.1 153 388 1.l -2.2584 -2.0318 -1.1400 015272 -0.0483 1 S704 -0.8506 l.g465 0.9562 3.2024 4.3127 2.4314 1.3651 3.6221 -0.3067 l.a781 0.9646 2.1 186 2.6584 -0.0488 1.g174 -0.6002 0.2384 -1.4911 -0.2938 0.8653 2.4750 0.1600 0.4538 1.7082 -2.0487 -0.1381 1.0619 -0.4657 0.6960 -1.3417 0.8750 2.1 589 0.6463 -0.1554 -1.0931 9.831 0.4761
3.6568 3.9386 5.6359 4.4074 12.6205 12.8573 35.1 905 27.9776 29.3890 30.3644 22.3401 9,3746 34.8288 28.9178 12.2043 5.4627 24.2380 7.8701 15.6046 9.5333 16.8326 9.5522 2.5946 0
49.9 = L0.01
For the first 50 observations from the base sample as well as for the first five observations in the tested sample, the targets deviate from 0.27~1 the presumably unknown means of the bivariate distribution by and -1.35~2, respectively. The deviations are more substantial for the other tested observations. The results of the testingfor the 50 observations in the base sample with the covariance matrix being estimated by S, are presented
33
0 0 0
or40
o m 0
o t - m
TABLE3.3
Means with external values(49.9,60.0)'. The S-matrix from the base sample
(50 observations).The data are the base sample:
VARl -1 49.8585 l 60.0008 49.8768 2 49.8706 3 49.91 17 4 49.8470 5 49.8883 6 49.9158 7 49.9152 8 9 49.9055 49.8969 10 l1 49.9137 49.8586 12 13 49.9514 49.8988 14 49.8894 15 49.9403 16 49.9132 17 18 49.8546 49.8815 l9 49.8311 20 49.8816 21 22 49.8501 23 49.9778 49.8690 24 25 49.8779 49.8680 26 49.9388 27 49.9133 28 29 49.9120 30 49.9250 31 49.9442 32 49.8386 49.9492 33 49.9204 34 49.8994 35 49.8703 36 49.8846 37 49.9580 38 49.8985 39 49.9397 40 49.8741 41 49.9140 42 49.9501 43 49.8865 44 49.8912 45 49.9252 46 49.9326 47 49.9680 48 49.9289 49 60.0632 50 49.9233 VAR2
t VARI
t VAR2
TZM 1.6073 4.2488 1.0904 5.3767 4.5009 0.4419 0.2205 0.5629 2.2046 1.0951 4.9861 2.9972 1.g884 0.0122 5.5016 l .S840 1.2866 2.6554 2.631 5 5.3803 1.8579 2.5277 7.0252 0.9596 1.2957 1.0088 1.2284 0.3785 0.1 053 0.7337 3.8446 3.81 54 2.4121 1.5392 1.5246 1.061 8 0.2658 2.7294 0.1939 1.l 778 1.0917 0.8206 2.0330 0.6997 0.2766 0.8988 0.7513 4.4314 0.5856 0.3528
0.0221 .l 881 59.9865 -0.6646 60.0055 -0.841l 60.0126 0.3349 60.0165 -1 S169 60.0216 -0.3339 60.051 7 0.4536 60.0673 0.4354 60.0726 0.1576 60.0208 -0.0897 60.0928 0.3913 59.9823 -1.1861 60.0866 1.4729 60.0402 -0.0342 60.0720 -0.3022 1.1550 60.0681 60.0350 0.3788 -1,2992 60.0145 59.9982 -0.5303 59.9963 -1.9721 60.0457 -0.5279 59.9860 -1.4297 60.0875 2.2265 60.0159 -0.8881 60.0055 -0.6324 60.0088 -0.9156 60.071 1 1.1104 60.0634 0.3821 60.0560 0.3432 60.0749 0.7151 60.1100 1.2659 59.9725 -1.7595 60.1 01 4 1.4090 60.0803 0.5831 60.0625 -0.0177 60.0219 -0.8496 60.0271 -0.4414 60.0878 1B598 60.0329 -0.0428 60.0826 1.l 365 60.0061 -0.7428 60.0401 0.4020 60.0810 1.4337 60.0169 -0.3857 60.0406 -0.2515 60.0532 0.7221 60.0741 0.9335 60.1219 1 .g476 60.0709 0.8285
~
-0.3639 0.1482 0.3400 0.4473 0.5856 1.3974 1.8202 l.g625 0.5621 2.5109 -0.4778 2.3434 1.0866 1.9479 1.8434 0.9479 0.3922 -0.0485 -0.1007 1.2355 -0.3777 2.3657 0.4294 0.1493 0.2381 1.Q224 1.7144 1.5150 2.0251 2.9759 -0.7427 2.7436 2.1732 1.6898 0.5914 0.7323 2.3742 0.8896 2.2342 0.1651 1.0851 2.1907 0.4581 1.0989 l.4392 2.0037 3.2971 1.9185 1.7102
35
TABLE3.4
Means with external values(49.9,60.0). The S-matrix from the base sample (50 observations). The data are the tested samples:
51 52 63 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 VARI 49.8798 49.9208 49.9606 49.9498 49.8390 49.9264 49.9648 49.9780 50.0218 50.0606 50.0365 49.9756 49.9840 50.0028 49.9770 49.8579 49.8997 49.9156 49.9258 49.8384 49.8937 49.8631 49.9406 49.9046 49.8718 VARZ 60.0417 60.0292 60.1 172 60.0543 59.9665 60.0079 60.0482 60.0186 60.0854 60.1399 60.1005 60.0387 60.0857 60.0482 60.0278 60.0588 60.0820 60.1415 60.1132 60.0449 60.0893 60.0757 60.1298 60.0739 60.0676
t VARI
-0.5791 0.5962 1.7352 l.4250 -1.7455 0.81 35 l.8567 2.2327 3.4887 4.5989 3.9083 2.1644 2.4049 2.9447 2.2037 -1.2048 -0.0075 0.4463 0.7401 -1.7624 -0.1795 -1.0554 1.1613 0.1308 -0.8068
t VARZ
1.1267 0.7889 3.1689 1.4677 -0.9059 0.2125 l.3041 0.5018 2.3086 3.7838 2.7175 1.0458 2.3170 1.3036 0.7522 1S908 2.2177 3.8274 3.0606 1-2144 2.4143 2.0484 3.5113 1.Q988 lB285
T2M
1.6092 3.6568 3.9386 5.6359 4.4074 12.6205 12.8573 35.1905 27.9776 29.3890 30.3644 22.3401 9.3746 34.8288 28.9178 12.2043 5.4627 24.2380 7.8701 15.6046 9.5333 16.8326 9.5522 2.5946 9.8310
in Table 3.3. The corresponding resultsfor the tested sample canbe found in Table 3.4. The computations were performed by the macros presented in Appendix 1.2. Those macros compute the relevant statistics for the analyses of ungrouped data presented in Chapters3-5. The program withthe macros that compute the statistics for ungrouped data UU.MTB 1. can be run by typing MTB>EXEC The summary mble 3.5 presents the number of observations for which the null hypothesis has been rejected both for the base as well as for the various tested samples. We observe againthat the success of the univariate tests to detect deviations is much smaller than that of the multivariate test.
36
Chapter 3
o o m o
-0000
37
We mentioned in the introduction that the Ti statistic does not provide the important information onthe variable(s) which causedthe out-of-control signal. On the other hand, we have seen inthe example above that the power of the univariate t-test is lower than that of the multivariate test. Multiple univariate testings have a compounded problem when the significance level is assessed, due to the multiple comparisons for detecting the outlying issue. In Chapter 7 we present other methods variables. These methods are extensions of the multivariate testingand use the information from all the variables and their covariance structure. As a second example of external targets we use the real data from are dimensions of several lots of ceCase Study3, where the variables ramic substrates. Raw materials used in the manufacturing of hybrid microcircuits consist of components, dyes, pastes, and ceramic substrates. The ceramic substrate plates undergo a process of printing and firing through which layers of conductors, dielectric, resistors, and platinum or gold are added to the plates. Subsequent production steps consist of laser trimming, component mounting and reflow soldering,chip or enwire bonding. The last manufacturing stage is the packaging and sealing of the completed modules. The ceramic substrates are produced in lots of varying sizes.The first production batch (labeled Reference) proved to be of extremely good quality yielding an overall smooth production with no scrap and repairs. This first lot was therefore considered a 'standard' to be met by all following lots. Five dimensions are considered inCase Study 3, with labels (a, b, c, W , L ) . The first three are determined by the laser inscribing processthe last two are outer physical dimensions. In our next example we consider only the first set of three dimensions from the table from Appendix 2. The Reference sample has average dimension, in mm, for (a, b, c) ~ ( 1 9 9 550.615, , 550.923). The engineering nominal specifications are (200,550,550). By redefining the (a, b, c ) measurements as the deviations from the nominal specification target values, a natural null hypothesis to test on the three dimensional population mean is:
38
Chapter 3
TABLE 3.6
Measurements in reference sample (Case Study 3)
b Means S"-matrix
(v)
-1.000 1.081 0.100 -0.432
0.923 0.615 0.100 1.336 0.063
-0.432 0.063 2.620
against
The means (with respectthe to center of the nominal specifications) and the S" matrix of the Reference sampleare as shown in Table3.6. Since the Reference sample has13 units we have that n = 13 and
T i = 13(V - 0)'s"
1.081 0.100 -0.432 = 131:-1.00 0.615 0.9231 * 0.100 1.336 0.063 0.063 2,620
= 13[-1.427 0.793 2.8901
(V - 0)
p%]
0.923
1[
*
-1.000 0.615 0.9231
= 13 X 4.58 = 59.54
Now since = 27.23 we have to reject H 0 at a 1% level of significance. Therefore we conclude that, on the average, although the Reference sample wasof extremely good quality did it not meet the required nominal specifications. Kenett and Halevy(1984) investigated the multivariate aspects of military specifications documents publishedby the US Department of Defense. They show that a multivariate approach is indeed requiredfor setting the inspection criteria of such products, since most product characteristics tend to be correlated. Again, such standards yield externally assigned targetsthat are used to monitor actual production.
39
Another source of externally assigned targetsis the development process of a productor process. In such a development process one typically accounts for customer requirements, internal technical capabilities, and what the competition offers. The Quality Function Deployment (QFD) matrices mentioned in Chapter 4below are used to integrate these different sources of information yielding agreed upon targets that are set so as to create "selling points" for the new product or process (Juran, 1988). Such targets are set on several dimensions and T i can be used to determine if these targetsare met usingthe techniques presented here.
Grouping the Data. If the tested sample of size n 1 is grouped in k rational subgroups of
size n j , n 1 = C nj the empirical covariance matrix is a pooled estimate
j=l
k
from the k sample covariance matrices calculated from each subgroups,

k
Fuihermore, if the subgroup sizesare all equal, i.e.nl = kn then the pooled covariance matrixS, is the average of the k individual matrices,
When we now test the mean of a subgroup j of n observations, the test statistic is defined as
where Yj is the mean of the n observations in the subgroup. The critical value for testing the hypothesis thatthe j-th subgroup does not deviate is given by: from the targets m0 by more than random variation
The groupingof the data also enables us to compute a measure of internal variability within the 'subgroup: For the j-th subgroup the measure of variability is given by
40
Chapter 3
where Yij is the i-th observation in the j-th subgroup. When considering all the observations of the j-th subgroup, relative to the targets mo, one obtains a measureof overall variability T$ defined as
From basicalgebra we have that
The critical values for Ti, can be approximated by U C L = (n - 1)x,2(a) where x, (a) is the 100 a-th percentile of the chi-squared distribution with p-degrees of freedom (see, e.g. Johnson,l985). The approximation is based on the substitution of S , by E in the formulafor the computation of Ti,. To illustrate the use of the test statistics we use an extended version of the second simulated data set with grouped data presented in Chapter 2. In addition to the first 100observations (50 subgroups of size two), we generated 90 additional observations grouped in 45 subgroups of size two whose distributions were as follows: the parameters of observations 101-110are identical to those of the base sample; the population mean for the first component of observations 110-130 have been shifted upward by two standard deviations, i.e. 10.0133 14.9763
41
in observations 13 1-150, the population mean of the first component was shifted downward and that of the second component upward, both by one standard deviation, i.e.
c13=c10+
[+f] [
-61
9.9728
g.g743] 9.9923 14.9763
In the fourth tested sample (observations 151-170) the shift was by one standard deviation in the first component withthe sign of the deviation alternating in consecutive observations as follows: in the observations whose case number is odd, the mean of the first componentwas shifted downward, while in the observations whose case number is even, the shift was upward. Thus when thedata are grouped in pairs, the last 10 groups formed from observations 151-170 have average mean whose expected valuesare as inthe base sample, but the within group deviations are expected to be large. Finally, the population meansfor each of the four components in observations 171-190 were shifted upward by one standard deviation,i.e.
i] [
=
The target values were set at
9.9998 14.9915 9.98881 9.9923
which deviate from the population meansby
r - 0 . 4 7 ~1 ~
42
Chapter 3
We performed the tests for the 75 groups and compared the resulting Tis and Tis with the critical values atQ = .05, which are 10.94and 9.49respectively. The results are presented in Tables 3.7 and 3.8. The computations were performed by the macros presented in Appendix 1.3. Those macros compute the relevant statistics for the analyses of grouped data presented in Chapters 3-5. The program with the macros which compute the statistics for grouped data can be run by typing MTB>EXEC GG.MTB 1. We summarize in Table 3.9 the numberof groups for which the appropriate T i and Ti-values exceeded the critical values in the base sample as well as inthe four tested samples. We can see from the table that evenfor the relatively small deviations between the targets and the actual means (by less than 0 . 5 for ~ each variable), the power of the multivariate test is about 60% for the groups in which the population means were not shifted. The detection is 100%in the second and the third tested samples andthe power is again about 60%-70% in the last two tested samples. Note that in the last tested sample all the means were shifted by one standard deviation, but they were all shifted in the same direction. The probability of detection of such a shift is considerably smaller than inthe case in which only two out of the four components are shifted by one standard deviation, but the shifts are in opposite directions. Obviously, a shift opposite directions by one standard deviationof all four components in (say in pairs) would have yielded even larger Ti-values. The results in the column which summarizes the results for the Ti-tests illustrate the power of the test to detect within group variation in the fourth tested sample.
TABLE 3.7
Means with external values (9.98,9.98,9.98and 14.98)'.The Spooled-matrix from the base sample: base sample (50 groups of 2 observations). The data are the
Group
l 2
3 4 5 6 7
9 10 11 12 13 14 15 16 17 18 l9 20 21 22 23 24 25 26 27 2a 29
30
31
32
33 34
35 36 37 38
39
40 41 42 43 44 46 46 47 4a 49
60
VARl 9.9765 9.9936 9.9870 9.9962 9,9774 9.9710 9.9613 9.9817 9,9887 9.9861 9.9997 10.0076 9.9725 9.9902 9.9786 8.9903 9.9975 9,9878 9.9879 9,9928 9.9819 9,9894 9.9891 9.9884 10.0112 9.9776 9.9893 9.9833 9.9975 9.9643 10.0107 9.9669 10.0061 9.9755 9.9928 9.9789 9.9645 10.0007 9.9844 9,9924 9.9909 9.9770 9.9923 9.9696 9.9661 9.9774 9.6757 9.9696 9,9809 9.9886
VAW 9.9702 9.9651 9.9719
9.9666
9.9718 9.9663 9.9733 9.9774 9.9642 9.9789 9,9980 9,9969 9.9648 9.9814 9.9720 9.9893 9.9870 9.9681 9.9640 9.8839 9.9742 9.9813 9.9610 9.9783 10.0054 9.9686 9.9793 9.9761 9.9877 9.9760 10.0034 9.9619 9.9972 6.9673 9,9669 8.9735 9,9594 9.9934 9.9729 9.9630 9,9858 9.9684 9.9665 9.9651 9.9771 9.9714 9,9706 9.9621 9.9761 9.9809
VAR3 9.9673 9.9893 9.9693 9.9777 9.9651 9,9585 9.9729 9.9731 9.9810 9.9812 10.0001 9.9939 9,9553 9.9770 9.9669 9.9768 9,9672 9.9841 9.9774 9.8801 9.9688 9.9741 9.9751 9.9765 10.0015 9.9665 9.9793 9.9688 9.9816 9.9796 10.0013 9.9582 9.9930 9,9613 9.9832 9.9651 9.9503 9.9880 9.9694 9.9788 9,9800 9.9659 9.9851 9.9587 9.9773 9.9678 9.9690 9.9602 9.9667 9.9728
VAR4 14.9670 14.9813 14.9661 14.9856 14.9718 14.9636 14.9706 14.9723 14.9787 14.9761 14.9935 14.9985 14.6590 14.9781 14,9630 14.9900 14.9667 14.9813 14.9750 14.9826 14.9737 14.9772 14.9762 14.9722 15.0117 14.9663 14.9804 14.9705 14.9883 14.9826 15.0057 14.9452 14.9819 14.9686 14.9870 14.9739 14.9525 14.9876 14.9733 14.9822 14.9749 14.9612 14.9868 14.9650 14.9802 14.9749 14.9701 14.9630 14.9683 14.9796
T2M
6.9126 18.4056 35.1165 18.8737 6.7657 9.1575 9.3300 5.9162 7.8017 9.8200 11.3981 22.6423 16,5223 13.5629 14.9679 10.5040 16.6040 4.3620 11.8774 12.4014 8.4692 14.1076 12.0774 20.1141 19.5601 9.1897 10.8675 12.7171 16.2634 5.5541 19.4523 36.0729 26.2016 9.8866 6.8393 8.4442 15.3261 19.4869 14.3155 12.9165 18.3518 16.8827 5.0976 7.3916 7.1717 5.0016 3.4050 8.0880 11.2887 11.9922
T'D
11.5787 2.61 10 2.3787 7.6355 3.1150 5.1778 1.8340 4.2786 0.2000 1.6535 4.6827 6.6440 0.9796 5.4059 2.7758 6.1168 2.4932 0.1646 5.1691 3.6190 2.2632 0.4522 7.6379 4.6382 3.5733 2.2299 l.8949 2.4218 3.5760 0.9283 1.0866 1.1480 3.3385 5.2634 4.0766 7.7542 3.6448 14.6015 0.8168 7.6439 3.0003 8.5347 7.0051 1.1271 2.1927 7.9634 4.1985 5.0988 0.7487 2.4071
T'o
18.4917 21.0175 37.4961 26.5084 9.8803 14.3352 l1.l 650 10.1941 6.0003 11.4742 16.0805 29.2664 17.5028 18.9667 17.7426 16.6199 19.0972 4.5464 17.0457 16.0202 10.7336 14.5589 19.7158 24.7498 23.1 334 11.4188 12.7836 15.1385 19.8582 6.4625 20.5401 37.221 1 29.5403 15.1706 10.9169 16.1987 19.1714 34.0904 15.1326 20.561 1 21.3509 25.41 77 12.1022 8.5188 9.3647 12.9645 7.6042 13.1861 12.0375 14.3999
TABLE 3.8
Means with external values(9.98,9.98,9.98 and 14.98). The Spooled-matrix from the base sample (50 groups of 2 observations). The dataare the tested samples:
VAR4 Group VAR3 VAR2 VARl S1 9.9752 9.9948 52 63 9.9825 64 9.9808 9.9986 66 66 10.0136 10.0068 57 10.0189 68 69 10.0184 60 10.0124 61 10.0094 62 9.9974 63 10.0282 10.0127 64 66 10.0042 9.9977 g.9817 66 9.9650 67 S8 9.9839 69 9.9768 9.9711 70 71 8.9722 72 9.9784 73 9.9727 9.9583 74 9.9877 76 9.9769 76 B.9789 77 9.9901 78 79 9.9720 80 9.9736 81 9.9839 82 9.9819 83 9.9916 84 9.9851 86 9.9715 86 9.9962 87 10.0015 88 10.0056 89 10.0085 9.9945 90 91 9,9881 92 10.0039 10.0003 93 94 8,9878 95 9.9972
T2M
9,9631 g.9782 9.9780 9.9748 9.9819 9.9778 9.9729 9.9765 93872 9,9818 9.9789 9.9655 10.0023 9.9794 9.9657 9.9865 9.9832 9.9922 9.9946 9,9902 0.9907 9.9880 9.9834 10.0065 9.9687 9.9682 9.9798 g.9648 9.9625 9.9759 9.9740 9.9845 9.9818 9.9642 9.9875 9.9960 10.0059 10.0020 9.9838 9.9830 9.9915 10.0019 9.9830 9.9861 9.9629 9.9728 9.9706 9.9664 9.9791 9,9720 9.9661 9.9786 9.9854 9.9746 9.9695 9.9548 9.9959 9.9784 9.9550 9.9808 9.9723 9.9709 9,9809 93759 9.9738 9.9781 9.9719 9.9624 9.9825 9.9704 9.9590 9.9719 9.9620 9.9596 9.9705 9.9705 9.9804 9.9718 9.9610 9.9802 9.9925 9.9992 10.0011 9.9811 9.9783 9.9921 9.9945 9.9778 9.9894 14.9669 14.9831 14.9844 14.9717 14.9747 14.9673 14.9723 14.9802 14.9924 14.9797 14.9769 14.9616 14.9947 14.9758 14.9619 14.9881 14.9711 14.9621 14.9819 14.9797 14.9736 14.9767 14.9796 14.9646 14.9964 14.9653 14.9656 14.9830 14.9607 14.9666 14.9793 14.9731 14.9792 14.97~13 14.9553 14.9804 14.9925 15.0061 15.0012 14.9775 14.9786 14.9909 14.9986 14.9756 14.9893 13.4041 29.1965 10.2608 9.4387 49.8481 194.9893 134.1977 198.7654 110.2353 121.0134 118.8845 126.4785 130.0325 145.0159 180.6132 34.8933 43.3239 36.0596 23.3029 57.9523 34.7953 20.5829 33.2976 69.4106 63.4783 11.8044 19.2991 15.7072 10.3572 12.9881 8.4083 7.8902 12.9792 8.5208 17.5724 22.3014 13.7984 14.0653 18.2087 24.4965 7.5604 26.0642 10.5187 10.7594 15.1882
~ ~~~
T20
4.3124 1.1444 5.1232 9.0740 1.3336 2.4845 12.4554 15.6996 1.0674 2.5841 7.5874 5.2893 6.7462 2.4281 3.7681 5.4240 5.9636 9.1947 8.5072 7.0179 3.0990 3.6764 5.5188 3.3428 3.8257 36.0720 30.5845 43.8885 22.2830 14.8953 10.2363 18.0869 24.5291 28.8474 . . 21.9927 3.9807 5.4804 5.7841 9.0957 1.2341 5.6353 2.8166 1.3998 0.1619 17.4943
T20
17.7152 30.3390 15.3835 18.5131 51.1806 197.4676 146.6554 214.4651 111.3039 123.5975 126.4702 131.7633 138.7787 147.4440 184.3891 40.3167 49.2889 45.2506 29,6100 64.9679 37.8982 24.2601 38.8147 72.7517 67.3056 47.8761 49.8822 59.3950 32.6401 27.8834 18.6437 25.9764 37.5092 37.3676 .. 39.5664 26.2821 19.2793 19.8486 27.3036 25.7311 13.1958 28.8802 11.9183 10.9214 32.6826
45
TABLE3.9
Number of times that T i and Ti exceeded the critical values
TA
Groups 1-50 - "Base" sample Groups 5 1-55 - First "tested" sample Groups 56-65 - Second "tested" sample Groups 66-75 - Third "tested" sample Groups 76-85 -Fourth ''tested" sample Groups 86-95 Fifth "tested" sample
29 3
10 10
Ti
2
Number of groups
50 5
10 10 10 10
0
2 1 10 1
6 7
4
Quulity Control with Internal Targets-Multivariate Process Capubility Studies
Objectives:
The chapterprovides examples ofprocesscapability studies carried outon multivariate data. The dzferent steps in p e f i m i n g such studies me illustrated A m u hfor setting control limits as well as multivariate capability indices computedfiom the analyzed data arepresented The chapter also presentsguidelinesfir interpreting T 2-charts and capability indices.
K e y Concepts
Refkrence sample 0 Process capability study 0 Statistical Control Charts 0 Multivariate process capability indices 0 Internal targets with the Leaveone out approach 0 Data grouping 0 Base Sample
0
Studies Capability Multivariate Process
49
During an ongoing industrial process, quality control is typically performed with target values derived from a standard reference sample whose units have been determinedbe to of acceptable quality on all the analyzed variables. However, at the initial stage of a new or changed process, a thorough investigation of its characteristics and capabilities is required. Ina process capability study, there typically are no preexisting data on the process characteristics. In particular, no target values on based prior information fromthe production of the component are available, and the target values therefore have to be calculated internally. It is important to distinguish between product specification limits which are derived from customer needs, and process quality characteristics which depend upon the processes involved the in production and delivery of the product. Univariate process characteristics are typically givenin terms of an average, estimating the process mean, and a standard deviation, estimating the process variability. Inthe multivariate case the vector of means replaces the univariate process mean the andcorrelations between variables are added to the variables standard deviationsor variances. The Western Electric Statistical Quality Control Handbook(1956) defines the term process capability study as ...the systematic study of a process by means of statistical control charts in order to discover whether it is behaving naturally or unnaturally; plus investigation of any unnatural behavior to determine its cause, plus actionto eliminate any of the unnatural disturbance. Process capability studiesare therefore much more than simple a exercise of just estimating the means andthe covariance matrix. Some of the steps in amultivariateprocess capability study include:
1. Determination of the boundaries of the process selected be to stud-
ied and of the variables characterizing the process outputs. This step is typically the responsibility of management. 2. Determination of representative time frames for data collection to define the sampling process and the rational subgroups. The design of the data collection system hasbeto tailored to the variables characterizing the process outputs.This step requires in-depth knowledge of the process as it is actually being operated andof future needs and plans that might affect the process.
50
Chapter 4
3. Performance of a cause and effect analysis linking process outputs characteristicsto internal parameters and control factors. The tools diused inthis phase can include the simple and efficient fishbone agram or the more comprehensive Quality Function Deployment (QFO)matrices. The analysis may be further refined and confirmed with statistically designed experiments. At this stage one can conduct experiments using concepts of robust designin order to make the best use of control factors as countermeasuresto noise factors affecting the process. For more detailson robust designs see, for example, Phadke (1989), and Kenett and Zacks (1997). 4 . Data collection and analysis using univariate control chart on the individual variables, multivariate control charts on a combination of variables, and various statistical and graphical methods to investigate the structure displayedby the data. 5. Elimination of special causes of variability. This step requires evaluation of stability of the process and an identification of sources of variability such as measurement devices, work shifts, operators, batches or individual item. 6. Evaluation of the underlying probability model for the process, including checking for multivariate normality. When needed, transformations may be usedto induce normality. 7. Computation of process performance indices or process capability indices. Process performance indices represent a measure of hisare calculated using all of the data without torical performance and considering stability over time. They are used to summarizepast performance. Process capability indices are computed whendata is collected over a short time frame (say 30 observations) and are used as predictors of process capability. Such indices require process stability,iepresentative samples, normality of the process distribution and independence between the collected observations. Recently several authors suggested using multivariate process capability indices (e.g. Chan et al., 1991; Wierda, 1993).
The multivariate process capability study is a classical situation in which quality control is performed with internal targets. The vector
Multivariate ProcessStudies Capability
51
of target values m is computed from the data (after the exclusion of outlying subgroups), and each observation or mean of subgroup in the process is compared withthose target values. empirical covariance matrix When the data are ungrouped and the S is based onthe entire sample of nl observations, in the T i statistic for the i-th observation given by
i -Y and S are not independently distributed. However the statistics Y it can be shown that (n - 1)s can be decomposed as (n - 1)s = (n- 2) S1+ ( Y i-7)( Y i such that (n- 2)S1 has a Wishart distribution and S 1 is independent of ( Y i Y ) .See e.g. Wierda (1994).Therefore, the distribution of Ti does not follow anymore(up to a constant) a Fisher F distribution but rather a beta distribution (see property (v) in Chapter 2). The appropriate UCL is given by
( 3
-v)'
where .) is the upper 100 a-th percentile of the beta distribution with the appropriate numberof degrees of freedom. If we ignore the dependence between( Y i - and S,we obtain (as suggestedfor example by Jackson (1959),Ryan (1988),Tracy et al. (1992)) an approximated UCL given by
&(a,
However, lately Wierda (1994)compared the values of UCL and UCL* for a! = ,005; p = 2,5,10 and for various values of n1. He found that the differences between UCL and UCL* are very substantial and the use of the approximated critical value is clearly unwarranted. It is however possible to define for the capability study T2-type statisticsfor which the statistics which measure the distance of the tested observation from the mean are independent of the estimated covariance
52
Chapter 4
matrix and whosecritical values are based on the percentiles of the Fdistribution. W O main alternatives were suggested in the literature: The first approach suggested by Wierda (1994) who considers
where only the covariance matrix is based on the nl - 1 observations but 7 is based on the entire sample of size n 1. The critical value for this
The second approach was suggested byBruyns (1992) who considers the "Leave One Out" method and definesthe statistic
where y(+, &i) are the vector of the means and the covariance matrix, respectively, calculated from all but the i-th observation. The critical value for that statisticis
It can be shown that for each observation i, TZ can be obtained as a function of T i as follows:
This relationship can be very useful since it enables us to compute TZ without recomputing the covariance matrixfor each observation separately. The critical values for T i are based onthe F-distributions which is more readily available than the beta distribution. Moreover, since the function which relates T i i with T i i is monotonic, the critical values can be obtained directly by the relationship. Thus
Multivariate Process Capability Studies
53
Wierda (1994) presents the three statistics and mentions that he finds the regular T i preferable, since on one hand it resembles more closely the Ti-statistics computed in other situations and onthe other hand it avoids the need to compute several timeseither the covariance matrix (for T i ) or both the covariance matrix andthe mean vector (for TZ). The disadvantage is obviously the need to appeal tothe critical values from the beta distribution. For methodological practical reasons we recommenda different approach than that suggested by Wierda.Since, we use T i with internal targets in the process capability study, in our opinion it is advisable to compare each observation with statistic a on whichthat observation had no effect, i.e. theLeave One Out approach. The relatively cumbersome computation of y(-il and S(-i)for each i, can easily be implemented in this age of fast computers. The use of the critical values based on the F-distribution can be considered in thiscase an extra bonus but certainly nota crucial consideration. While as mentioned, the Leave One Out approach is preferable from the methodological point of view, in practical conditions the differences between the two methods is seldom noticeable. Unless the number of observations is very small, the effect of a single observation on the computed statisticsis rarely such that outlier the is detected by the Leave One Out approach and not detected by TA. For an illustration of the methods, let us now return to the first simulated data set with ungrouped data presented in Chapter 2. We consider only the first 50 observations as simulating data derived from a process capability study with unknown parameters. We recall that the first 55 observations were generated from the distribution with the same (PO,X) while the means of the next 20 observations were shifted. We of course assume that at this testing stage,the underlying parameters are unknown to the investigator, and proceed totest each observation sep-
54
Chapter 4
arately versus the empirical overall mean of the first 50 observations. The covariance matrixis also estimated fromthe base sample. We present in Table 4 . 1the values of the statistics Ti,TZ and T2 M respectively. The statistics were computed by the program whose main macro is U U . M T B from Appendix 1.2. Notethat we have in fact only two distinct methods since there is a one-to-one correspondence between Ti and Ti.The critical values for a! = .05 are 5.76,6.40 and 6.66and for a! = .005 they are 9.69, 11.90 and 12.39, respectively. The critical values for T i are based onthe beta distribution, while those for Ti and T,$ are based on percentiles of the F-distribution. We observe that only one out of the 50 Ti-statistics exceeded the critical values at a! = .05 (observation 23) and that none of them exceededthe critical values for smaller significance levels. The T2-value for observation 23 exceeded the appropriate critical values ata! = .05 for all the methods. If this was indeed a process capability study, it is doubtful if this observation would have been labeled as outlier. It is more likely that the entire sample would have been accepted and used as a reference sample for future testings. The observation 23 would have been considered in thiscase a random error (one out of 50, at a! = .OS). From the data generation process, we know that this was indeed a random error.
Grouping the Data. We mentioned that in the process capability stage,the estimation of the internal variability within the subgroups maybe particularly important. The data has of course to be divided at this stage into rational subgroups, much the same as it is likely to happen in future stages of the production process. Thus, we test each of the k subgroup means yj, j = 1, . ,, k against the overall mean The statistic
T.
is compared tothe UCL of

F;kn-k-p+l
TABLE 4.1
Means from the base sample(49.9026,60.0441)'.The S-matrix from the base sample (50 observations). The data are the base samples:
VARI 1.6073 1 60.0008 49.8585 2 49.8768 49.8706 3 4 49.91 17 5 49.8470 49.8883 6 49.9158 7 49.9152 8 49.9055 9 49.8969 10 49.9137 11 49.8586 12 49.9514 13 49.6988 14 49.8894 15 49.9403 16 49.9132 17 18 49.8546 19 49.8815 49.831 l 20 49.8816 21 22 49.8501 49.9778 23 49.8690 24 25 49.8779 49.8680 26 49.9388 27 49.9133 28 49.9120 29 49.9250 30 49.9442 31 49.8386 32 49.9492 33 49.9204 34 49.8994 35 49.8703 36 49.8846 37 49.9580 38 49.8985 39 49.9397 40 49.8741 41 49.9140 42 49.9501 43 49.8865 44 49.8912 45 49.9252 46 49.9326 47 49.9680 48 49.9289 49 0.3528 60.0632 49.9233 50 VAR2 59.9865 60.0055 60.0126 60.0165 60.0216 60.0517 60.0673 60.0726 60.0208 60.0928 59.9823 60.0866 60.0402 60.0720 60.0681 60.0350 60.0145 59.9982 59.9963 60.0457 59,9860 60.0875 60.0159 60.0055 60.0088 60.071 1 60.0634 60.0560 60.0749 60.1100 59.9725 60.1014 60.0803 60.0625 60.0219 60.0271 60.0878 60.0329 60.0826 60.0061 60.0401 60.0810 60.0169 60.0406 60.0532 60.0741 60.1219 60.0709
T'M
T ' 'M 1.6559 4.9924 1.0968 6.721 1 5.3606 0.431 8 0.2134 0.5529 2.3363 1.1018 6.0980 3.301 8 2.0856 0.01 17 6.9263 1.6310 1.3059 2.8763 2.8470 6.7271 1.g366 2.7209 9.6859 0.9596 1.3157 1.0110 1.2436 0.3688 0.1014 0.7262 4.4224 4.3822 2.581 9 1.5806 1.5647 1.0667 0.2577 2.9671 0.1874 1.1895 1.0983 0.8155 2.1368 0.691 5 0.2683 0.8963 0.7443 5.2581 0.5758
. I
4.2488 1.0904 5.3767 4.5009 0.4419 0.2205 0.5629 2.2046 1.0951 4.9861 2.9972 1.9884 0.0122 5.5016 1.5848 1.2866 2.6554 2.631 5 5.3803 1.8579 2.5277 7.0252 0.9596 1.2957 1.0088 l.2284 0.3785 0.1053 0.7337 3.8446 3.8154 2.4121 1S392 1.5246 1.061 8 0.2658 2.7294 0.1939 1.1778 1.091 7 0.8206 2.0330 0.6997 0.2766 0.8988 0.7513 4.4314 0.5856
1B962 4.7536 1.1379 6.1752 5.0657 0.4549 0.2259 0.5813 2.3570 1.1426 5.6755 3.2602 2.1156 0.0124 6.3380 1B716 l.3483 2.8674 2.8393 6.1804 1.971Q 2.7213 8.3933 0.9988 l .3581 1.0510 1.2859 0.3894 0.1077 0.7600 4.2628 4.2274 2.5905 1B220 1B064 1.1078 0.2726 2.9514 0.1986 1.2315 1.1393 0.8514 2.1652 0.7241 0.2839 0.9341 0.7786 4.9798 0.6047
56
Chapter 4
in orcfer to detect outlying subgroups (e.g., Ryan, 1988). This test procedure can also be used whena base sample is being calibrated.If we identify l subgroups withTi >UCL and determine an assignable - cause that justifies their removal fromthe data set we can recompute ff and S, using the remaining k l subgroups. The Upper Control Limits (UCL) for a future subgroup of n observations is then set to be:
UCL =
p (k - l l ) ( n - 1) (k-1)n-k+l-p+l
F;,(k-l)n-k+l-p+l
However, even after removing obvious outliers, the testing of each of the subgroup means againstthe overall mean should be performed with caution since possible remaining outlying observations or trends the in sample, may affect both the overall mean andthe empirical covariance matrix. The subsequent tests for the individual subgroups could thus be biased. It is noteworthy that the fact that the data are grouped does not necessarily imply that the estimated covariance matrix has to be pooled over all the groups. Specifically, both in thischapter as well as in the other chapters in this book, we focus on two main cases with respect to the grouping issue:either (a) the data are ungrouped, and then in the TA we consider the distance of each observation from computation of Y,i.e. (Y, - Y),and use the empirical covariance matrix S, -or (b) the data are grouped andfor the j-th group we consider - y ) with the pooled covariance matrix S,. However, those are not the only possible methods of analysis. Indeed, even if the data are grouped in k groups of size n, if we believe that all the data obtained duringthe process capability study originate from the same distribution, the covariance matrixS (with kn 1 degrees of freedom) is a more efficient estimator of E than Sp with k(n - 1 ) degrees of freedom (e.g. Alt (1982), Wierda (1994)). Some authors therefore suggest the use of the statistic
e,
57
The distribution of this statistic is similar to that presented for the ungrouped case, and we have
statisticwhich has, under In this case, we can also compute an alternative the null hypothesis, an F-type distribution. The statistic is:
The definition of is more complicated than the in ungrouped case, = S(-t), i.e. the covariance matrix could have been obtained There, directly fromall but the i-th observation. In the grouped case we have
N
F(-,)
F(-,) = [(n - l)k$ + n(k - l)ST-,)]/(kn - 2)

where ST-,) is the covariance matrix computed from the k - 1 group means for all but the j-th group. The formula for ST-,, is thus
UCL' =
(kn - 2)(k - 1)p k(kn - p - 1) F;,,kn-p-l
As in the ungrouped case there is no need to actuallycompute the T g statistic directly. From the value of T i j we can compute
T$ =
The use of the more efficient estimator of E (S instead of Sp) can be important during a process capability study when the number
58
Chapter 4
observations is limited. It is, however, obvious that the advantage gained by the use of a more efficient estimatorof the covariance matrix may be counterbalanced by the bias induced if the assumption that the entire data set has the same distribution, is invalid. The values of the parameters are obviously unknown and many in cases slight shifts may occurthe in process capability study whichcan effect the dispersion as well as the location. Therefore, we believe that in manycases it is unwarranted to , is more appropriate in assume the risk and we suggest that the use of S practical situation as an estimatorof the covariance matrixfor grouped data, and weuse it in our computations. In the multivariate process capability studies, where we assess the characteristics of the process, the estimation of the internal variability within the subgroups may be especially important. The T i measure defined in Chapter 3 for the case when the targetsare from a reference sample, can be similarly calculated for the j-th subgroup as
i=l
where Y , , ,i = 1, , . ,n is the i-th observation in the j-th subgroup. The overall measure of variability T o 2 = T i T i provides the sum - of the weighteddistances of the n observations from the grand mean P. We illustrate in Table 4.2 the results of the computations with the second simulated data set with four variables. The 100 observations from the simulated - process capability study were grouped in50 pairs. The grand mean and the pooled are - empirical covariance -matrix presented in Chapter 2. (Note that from this chapter equals from Chapter 2). For each of the 50 subgroups, we computeT i and T i and compare them withthe critical values.The statistics were computed by the program for grouped data whose main macro is GG.MTB from Appendix 1.3. The critical values for T i are 18.96, 17.08 and 10.28 for a = .0027, a = .005and a = .05, respectively. The statistic T i which assesses the within group variability is compared to the critical values approximated by Xi(a),that in this case are 16.25,14.86and 9.488 for the a-values as before. Among the 50 subgroups, we find five whose Ti-values exceeded 10.28, and one group whose Ti-value exceeded
TABLE4.2
Means from the base sample(9.9864,9.97932,9.97525 and 14.9768)'. The Spooled- matrix from thebase sample (50 Groups of 2 Observations). The data are the tested samples:
Group
51 62 63 54 55
56
57 58 69 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79
80 81
82 83 84 85 86 87 88 89 90 91 92 93 94
95
VARl 9.9752 9.9948 9.9825 9.9808 9.9986 10.0136 10.0068 10.0189 10.0184 10.0124 10.0094 9.9974 10.0282 10.0127 10.0042 9.9817 9.9650 9.9639 9.9768 9.9711 9.9722 9.9784 9.9727 9.9583 9.9877 9.9769 9.9789 9.9901 9.9720 9.9736 9.9839 9,9819 9.9916 9.9851 9,9715 9.9962 10.0015 10.0056 10.0085 9.9945 9.9881 10,0039 10.0003 9.9878 9.9972
VAR2 9.9631 9.9782 9.9780 9.9748 9.9819 9.9778 9.9729 9.9765 9.9872 9.9818 9.9789 9.9655 10.0023 9.9794 9.9657 9.9977 9,9865 9.9832 9,9922 9.9946 9.9902 9.9907 9.9880 9.9834 10.0065 9.9687 9.9682 9.9798 9.9648 9.9625 9.9759 9.9740 9.9845 9.9818 9.9642 9.9875 9.9960 10.0059 10.0020 9.9838 9.9830 9.9915 10.0019 9,9830 9.9861
VAR3 9.9629 9,9728 9.9706 9.9664 9.9791 9.9720 9,9661 9.9786 9.9854 9.9746 9.9695 9,9548 9.9959 9.9784 9,9550 9.9808 9,9723 9.9709 9.9809 9.9759 9.9738 9.9761 9.9719 9.9624 9.9825 9.9704 9,9590 9.9719 9.9620 9.9596 9.9705 9.9705 9.9804 9.9718 9.9610 9.9802
VAR4 14.9669 14.9831 14.9844 14.9717 14.9747 14.9673 14.9723 14.9802 14.9924 14.9797 14.9769 14.9616 14.9947 14.9758 14.9619 14.9881 14.9711 14.9621 14.9819 14.9797 14.9736 14.9767 14.9796 14.9646 14.9964 14.9653 14.9656 14.9830 14.9607 14.9666 14.9793 14.9731 14.9792 14.9793 14.9553
T2M
4.3124 1.1444 5.1232 9.0740 1.3336 2.4845 12.4554 15.6996 l.0674 2.5841 7.5874 5.2893 6.7462 2.4281 3.7681 5.4240 5:9636 9.1947 6.5072 7.0179 3.0990 3.6764 5.5188 3.3426 3.8257 36.0720 30.5845 43.6885 22.2830 14.8953 10,2363 18.0869 24.5291 28.8474 21.9927 3.9807 5.4804 5.7841 9.0957 1.2341 5.6353 2.8166 1.3998 0.1619 17.4943
T'D
5.6006 8.8938 13.2905 2.0769 19.4155 125.5682 77.3265 131.9750 62.2530 67.2113 65.1 534 70.8581 76.6345 87.6476 112.6605 55.8683 72.9030 58.2018 45.6482 88.6448 55.3878 33.0686 56.2431 101.1653 85.0237 4.6509 4.5457 4.7208 3.7056 6.3224 3.5303 0.5604 1.2320 4.7872 6.7332 4.301 5 4.4330 15.5645 9.0404 5.7927 0.3996 8.6158 12.2313 1.3657 5.3729
T20
8T9325
9.9992 10.0011 9.9811 9,9783 9,9921 9.9945 9.9778 9.9894
14.9925 15.0061 15.0012 14.9775 14.9786 14,9909 14.9988 14.9756 14.9893
9.9132 10.0373 18.4131 11.1502 20.7486 128.0478 89.7838 147.6752 63.3211 69.7954 72.7396 76.1438 83.3807 90.0757 116.4350 61.2914 78.8694 67.3922 52.1554 95.6601 58.4912 36.7462 61.7600 104.5051 88.8508 40.7232 35.1300 48.4088 25.9887 21.21 81 13.7670 18.6473 25.7615 33.6342 28.7265 8.2821 9.9134 21.3489 18.1360 7.0271 6.0351 l 1.4324 13.6305 1.5276 22.8668
60
T ' ( 0 . 9 9 5 )
103
Chapter 4
- """"""""_" " " " " " " " " " -
80? ' (S ,)
Flgure 4.1 : T2-chart for the 17 variables fromCase Study 2 (ungrouped).
18.98. In an actual capability study, such a subgroup should be exclud as a precaution.We know inthis case that the data were simulated, and that the underlying distribution was identical to that of the other cases. However, in a real situation, such information is not available. As for the within group variability in our data, we observe from the 9.49 and one table that there were two Ti-values whose values exceeded which exceeded 14.86. There was no Ti-value higher than 16.25. We mention that the seemingly outlying groups fromthe point of view of location (i.e. differ fromthose which have high internal variability. This result gives us a (weak) indication that there is a possibility that the outliers are in fact random errors.
2'')
Control Chartsin Multivariate Process Capability Studies. After the deletion of the detected outliers, the process capability study just presented provides a "base" sample for future testing of the production process.We now illustrate the display for the Ti-statistic in a multivariate controlchart for the ungrouped data from Case Study 2. The data have been collected in a process capability study of a complex mechanical part on which the 17 recorded variables include11 diameters, 4 length and2 wall-width measurements.The internal targets were
61
obtained from the 36 observations and the resulting T2-chart is presented inFigure 4.1. The method of analysis is similar tothat described in the section on analysis of control charts in the previous chapter, We only mention that thereis no evidence of outlying observations inthis particular data set. While it is quite possible that in this particular case the process was in control throughout the process capability study special a feature of this data set deserves special attention. The large numberof variables (17) relative tothe number of observations (36) tends to favorthe null hypothesis and to decrease the power of the test. It is thus recommended in similar situations to continue the data analysis on particular multivariate subsetsof variables (possibly physically related) and investigate potentially assignable causes which may effect the studied process.
5
Quahy Control with Tdrgets fiom a Reference Sample
Objectives:
The chapter presents quality control proceduresfor comparing netu4 collected observations with targets determinedfiom a ref erence sample.When the tatgets are based on a rference sample, the objective o f multivariate qualily control is to reach decisions on an on-going basis regarding thepetjirmance ofaprocess over time as compared to the valuesfiom reference sample. This decision makingprocessis illustrated with real li@exampks.
K e y Concepts
Reference sample Process capability study o Tested sample e Univariate control limits 0 Multivariate control limits 0 Data grouping e Decomposition of the Hotelling T2-distance
e e
63
from Targets
65
In the first phase of the process, i.e. the process capability study a reference or base sample is typically selected and based on its values, targets are set as reference for future testing of observations, Moreover, the data from the reference sampleare used to assess the covariance pattern of the analyzed characteristics as well. In a second phase of the process the newly tested samples are routinely compared against the target values determined from the reference samples. We note that comparison between a testedsample and the target values from a reference sample can also occur inthe advanced stages of a process capability study. The selection of the reference sample out of the entire set of available data takes place as an initial stage of the process capability study. Following that stage, there may still be additional available observations which can be tested againstthe setup targets. The statistical analysis for comparing a tested sample (whose observations are denoted byY ) with a reference sample (denoted byX) is based on the procedures for testing the equality of mean vectors of two parent populations. The sequential comparison of single observations or of subgroups of observations from the tested sample with the reference sampleis an extension of the same procedures. the tested sample which includes n 1 independent Let us assume that observations Y 1, ,Y,, ,n 1 2 1 distributed asY N,, (p1, E l ) .Furthermore, let us assume that here we also have a reference sample of n2 independent observationsX I , .. ,X,, distributed as X N,,(pz,E 2 ) . We assume throughout this chapter that E1 = E2 = E. We start with the case where the samples ungrouped. Let?? a n d x be the vectors of averages for these two multivariate samples, respectively. We want to test the hypothesis H 0 : p1 = p2. From the results in Chapter 2, we have that T? N,,(pl,+E) and Np the independence of and H we also have that
. ..
-x
x-
If E is known, we can calculate the statistic
66
Chapter 5
which follows a noncentral chi-square distribution with parameter of noncentrality
When the null hypothesis H 0 : p1 = p2 holds, the parameter of noncentrality t 2is zero, and
Ti
Xi
However, X is typically unknown and has to be estimated from the samples. Let SX be the sample covariance matrix fromthe reference sample. If n1 2 2, we can also calculate from the tested sample its covariance matrixS y Based on SX and S y ,a pooled estimatorof can in principle be derived as
If the quality control procedure calls for the comparison of consecutive single observations fromthe tested sample with the reference sample (i.e. nl = l), l2 can obviously be estimated only from SX.However, even when n1 2 2, the information from the tested sample is rarely combined with that of the reference sample inthe estimation of the covariance matrix. Since the comparison with the reference sample is performed routinely,it is much more convenient to keep the estimate of the covariance matrix constant when we compare the various tested samples with the inference samples. Furthermore, if the tested data contains outliers, they may differ from the parent population not only in their means, but also in their covariance structure. Thus, the test statistic for H 0 :p 1 = p2
is compared tothe Upper Control Limit (UCL) which is in thiscase
Targets from a Reference Sample
67
Fpm-P
UCL =
Pb2
- l>(nl + n2)
n2(n2
-P)
The control limits are used to determine departures from statistical control. A value of Ti > U C L , or trends, or other patterns in the Ti-values, point out to a possible assignable cause affecting the process under observation. When a departure from statistical control is detected, one typically initiates an investigation to look for the specific characteristic that could be affected by that cause. Individual control charts are a logical next step in that direction. However, one should keep in mind that an aberrant Ti value mightbe a result of observations whose correlation structures differ from that of the parent population and those featuresdo not show up the oncontrol chart for individual characteristics. Lately, Mason et al. (1995) have shown that the interpretation of a signal from a Ti-statistic is aided if the corresponding valueis partitioned into independent parts. Information on which characteristic contributes significantly to the signal is readily available from that decomposition. We shall expand onthat in the next chapter. Let us now illustrate the methods of analysis with the simulated data set presented in Chapter 2 which includes both a base or reference sample of 50 observations as well as a tested sample with 25 observations. Unlikethe analyses performed on thatdata set in the previous chapters, we now assume that both the vector of means as well as the covariance matrixare estimated fromthe reference sample.As mentioned in Chapter 2, we foundin the reference sample
and
x = [60.0441
49.9026
sx = [ .001220 .001440
.001140 .001367
Using the macro W.MTB from Appendix 3.2 with the properly defined parameters, we can compute the relevant statistics for each of the 25 single observations in the tested sample (i.e., n1 = 1). We test the statistics obtained versusthe UCL-values computedfor various as. The results are presented in Table 5.1.
60
Chapter 5
TABLE 5.1
Means from the base sample(49.9026,60.0441). The S-matrix from the base sample (50 observations). The data are the tested samples:
VARl 51 52 53 54 55 66 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 49.8798 49.9208 49.9606 49.9498 49.8390 49.9284 49.9648 49.9780 50.0218 50.0606 50.0365 49.9756 49.9840 50.0028 49.9770 49.8579 49.8997 49.9156 49.9258 0.0221 49.8384 -1.8365 49.8937 49.8631 49.9406 49.9046 49.8718 VAR2 60.0417 60.0292 60.1172 60.0543 59.9665 60.0079 60.0482 60.0186 60.0854 60.1399 60.1005 60.0387 60.0857 60.0482 60.0278 60.0588 60.0820 60.1415 60.1132 60.0449 60.0893 60.0757 60.1298 60.0739 60.0676
,
t VARl
-0.6532 0.5222 1.6612 1.3510 -1.8195 0.7394 l.7826 2.1587 3.4146 4.5249 3.8343 2.0903 2.3308 2.8706 2.1296 -1.2789 -0.0816 0.3722 0.6860 -0.2535 -1.1295 1.0872 0.0568 -0.8809
t VAR2
-0.0656 -0.4033 1.Q766 0.2754 -2.0982 -0.9798 0.1 ll 9 -0.6905 1.1163 2.5915 1.5253 -0.1465 1.1247 0.1113 -0.4401 0.3985 1.0254 2.6351 l.8683
T2M
1.6092 3.6568 3.9386 5.6359 4.4074 12.6205 12.8573 35.1905 27.9776 29.3890 30.3644 22.3401 9.3746 34.8288 28.9178 12.2043 5.4627 24.2380 7.8701 15.6046 9.5333 16.8326 9.5522 2.5946 9.8310
1.2220 0.8561 2.3190 0.8065 0.6362
We also summarize in Table 5.2 the number of observations for which the statistics exceeded the appropriate critical values. The numof the tested ber of rejections are presented for each ofthe three subsets becompared sample (of sizes 5,lOand 10,respectively).The results can with a similar table presented in Chapter The 3. difference betweenthe tables is that in Chapter 3the null hypothesis specified the target value of m0 while here we comparethe Yi S with the empirical obtained from the reference example. The critical values are changed accordingly, andfor the multivariate test they are 15.202,12.346 and 6.645for Q = .0027, Q = .005and Q = .OS, respectively.
69
o+e4
o m 0
70
Chapter 5
Figure 5.1 :Scatterplot of ten bivariate observations from Case Study 1 with univariate and multivariate control limits.
The results in the table above are similar to those presented in Chapter 2 but the powers of the tests are higher in this case. We see again that the empirical power of Ti exceeds thatof the univariate test performed specifically on the effected component. An interesting pattern can be observed in the last subset of 10 observations from the tested sample (observations 66-75). The population meansin those observations were all shifted by one standard deviations, but in opposite directions for the two components. In this case, the power of the multivariate test is substantial, while the univariate test detects almost no deviation from the standard. To further investigate the phenomenon, we present in Figure 5.1 the scatterplot of the data with the appropriate control limits. The interior andthe boundaries of the ellipse is the area for which
The rectangular lines represent the control limits for each of the two components separately. The control limitsare set for the two sidedtest at a! = .05. The difference in the percent of rejections between the multivariate and univariate procedures is due to the fact that the multivariate test accounts for the dependence betweenthe variable whilethe univariate test does not. Let us consider for example observations 70 and 72.The first component incase 70 is below its population mean by 1.845~1 and the second component by 0.025~2. In observation 72 the
from Targets
aSample Reference
71
first component is below the mean by1.13S1, and the second component is above its mean by 0.8682.None of these marginal deviations exceeds the univariate critical limits, but given the correlation of r = 383,the probability of the events represented bythe bivariate values wouldbe very unlikely if the process were in control.
Grouping the Data. Let us now turn to the case in which the samples - are grouped in rational subgroups of size n. We denote in this case by the grand mean of the reference samples, and by SxP the pooled estimate covariance matrix basedon k subgroups of size n. When we now test the mean ofa new group of n observations, (i.e. n1 = n),the test statistic
is compared to the Upper Control Limit (UCL), which in this case is
For a new subgroup of n observations Yi,i = 1.. ,n, with mean Y we can also compute the internal measureof variability T i , i.e.
When we consider the deviations of all the observations in the new subgroupfrom the mean of the referencesample,weobtain the measure of overall variability T:,
with
T i = T i 4- T i
72
Chapter 5
We illustrate the computation of statistics presented for grouped data with a further analysis of the data from the capability study of aluminum pins (Case Study 1). We mentioned in the previous chapter that the 70 observations in that study were composed of two subsets: the first 30 observations were considered a reference sample against which we compare the other 40 observations. Furthermore,the data in both the reference and the tested sample, are grouped in subgroups of two consecutive observations (15 and 20 subgroups, respectively). hble 5.3 presents both the relevant statistics fromthe reference sampleas Y 1and Y2) of subgroup 16, which is the well as the two observations( first subgroup in the tested sample. Thoseare the observations Y 3 1 and Y32 in the original data set. The vector of means is also presented. The statistics from the reference sample are the vector of means 2 and the covariance matrix SX,,(pooled over the 15 subgroups) and its inverse The Ti - statistics from subgroup16 is given by
(v)
which in our case is:

TL = 2*[0.0137
0.0113 0.0157 0.0187 0,01771 0.0177 0.0137 0.0113
75824.7-14447.7-29073.3-29977.6-4940.95301.9 709.5 856.4 -14447.7 63781.5 -31818.8 -4551.4 3508.6 -3419.4 -29073.3 -31818.8 61607.7 -253.9 -29977.6 -4551.4 -253.9 32532.6 1937.6 -2630.0 3508.6 1937.6 3172.4 -3029.8 -4940.9 709.5 5301.9 856.4 -3419.4 -2630.0 -3029.8 4158.8
= 2*[-323.488
=2*9.24 = 18.48
- 60.6431
324.8947 222.7156 141.2242 147.4971*
0.0137 0.0113
0.0177 -0.0177,
:::f::
Note that unlike the calculations in the previous chapters, the targets are now the means of the reference sample. In Chapter 3, the targets were the external values m0 while in Chapter 4 they were, at first, the
73
TABLE 5.3
Computation of the Ti-statistics - case study 1 Diameter1Diameter2Diameter3Diameter4 Lengthl Length2 49.92 49.93 49.925 60.03 60.03 60.03 60.048
y2
Y X
Y I
10 10 10 9.986
9.99 9.99 9.99 9.979
9.99 9.99 9.99 9.974
14.99 15 14.995 49.907 14.976
.10 0.83 0.97 1.13
Sp-matrix(**lO,OOO) Diameterl 0.87 Diameter2 1.10 0.87 Diameter3 1.030.90 Diameter4 12.33 -1.20 -1.10 -1.30 Lengthl -0.97 -0.83 8.70 -0.63 -0.80 Length2 Sil-matrix
1.so 8.83
Diameterl 75824.7 63781.5 -14447.7 Diameter2 61607.7 -31818.8 Diameter3 -29073.3 32532.6 -253.9 -4551.4 Diameter4 -29977.6 Lengthl 3172.4 1937.6 3508.6709.5 -4940.9 8.8 029.8 -2630.0 -3419.4 856.4 5301.9 Length2
internal values based on the entire sample of 70 observations. Also, the S i ; matrix presented above differs from the Si1matrix usedin the previous chapters since its calculation is based on 15 (rather than 35) subgroups. The T i for the same subgroup in Chapter 3 was 68.43. The internal measure of variability T i = is given in thiscase by T i = [0 0 0 - 0.005 - 0.005 *[O 0 0 - 0.005 - 0.005
{=l
c (Yi-y)Sij (Y, -v)

e L
74
Chapter 5
+[0 0 0 0.0050.005O I S . ; ,
*[O 0 0 0.005 0.00501

= 2 0.99 = 1.98
and the value of T =
iel
c (Xi- v)Sij(Xi- v) is
2
Ti = [0.0137 0.0113 0.0157 0.0137 0.0127 - 0.01771S;i

*[0.0137 0.0113 0.0157 0.0137 0.0127 - 0.01771 +[0.0137 0.0113 0.0157 0.0237 0.0227 - 0.01771S;: *[0.0137 0.0113 0.0157 0.0237 0.0227 - 0.01771 = 6.59 13.87 = 20.46
It can easily be checked that T = Ti Ti.The decomposition of the 7-value into the components which measure the distance from target (Ti) and the internal variability (Ti), respectively, clearly indicates that the distance from target accounts for the mainpart of the overall subvariability (18.48 out of 20.46). The two observations which form group 16 (observations 31 and 32), differ only slightly from each other, by .01 in the fourth and the fifth variables (see Table A2.1 in Appendix 2). The small Ti value reflects this homogeneity withinthe 16th subgroup. We repeat the conclusion presentedin Chapter 3 on the relative magnitude of Ti and Ti in this subgroup. The relatively large value of indicates a substantial deviation fromthe mean of the reference sample, while the small Ti-value indicate that the two observations in subgroup 16 are internally homogeneous.
G
Analyzing Datuwith Multivariate Control Cbarts
Objectives:
The chapter shows how to practically analyze data with multivariate control charts, The multivariate control chart is one o f the main practical tooh in multivariate quality control. The chart disphys the T 2-distance with the upper control limits The control limits are setat the critical vahes determined by the nature o f the targets, af presented in theprevious three chapters.
K e y Concepts
0
Graphical displayof data Univariate controllimits 0 Critical values 0 Multivariate control limits 0 Data grouping 0 The T2-chart
Analyzing Data with Multivariate Control
Charts
77
The graphical display of data plays an important role in the detection of changes in the level of a variable characterizing a process. In fact, in many practical instances with a single attribute monitored,statistical quality control consists mainly of plotting the data on a control chart assessing levels and/or trends in the variable based on the graphwhen it is determined that an assignable ical display, andtaking action cause is present. Furthermore, even in the cases when several attributes are monitored, the most common practice is to monitor them on several separatedisplays. Thus, although we focus on multivariate observations, it is important to review briefly the univariate control charts which in many practical cases will at least supplement the multivariate charts. The univariate control charts introduced by Shewhart in 1924 and described in his classic book of 1931 are considered a cornerstone in the development of statistical quality control in industry. Trends and changes in the levels of a variable are naturally and quickly perceived of tabulated numbers. from a graphical display rather than from a listing This general observation is particularly valid within an industrial setting. The basic univariate Shewhart control charts successively display values of individual observations, or averages of subgroups of observations, andare termed respectivelyx-charts and F-charts. When the data are grouped, the F charts are frequently accompanied by charts which display the value of a statistic which measures the dispersion within the groups (such as ranges or standard deviations). Examplesof those are the R-charts and the SD-charts. The paired g, R) charts are usually displayed and maintained on the production floor in industry to help monitor key process characteristics. We illustrate in Figure 6.1 the ( Z , R) chart with the data set on the sixth variable (overall length with cap or Length2) from Case Study 1 (35 subgroups of size 2). From the chart we can observe that the overall length is significantly below target in subgroups 24 and 26 and so are,to alesser extent, the two adjacent neighboring subgroups (23 and 27). The R-chart indicates that there are large dispersions between the two observations which form subgroups25 and 3 1. Various additional charts, for monitoring both process location and process dispersion, have been developed and used in industry. These include median charts, midrange, CUSUM,moving average, exponentially weighted moving average, etc. A detailed description of those univariate
78
60.12
Chapter 6
3.0SL=60.09
3
60.02
Y V
I I I I I I
X=60.03
-3.OSL-59.96
I
59.92 . I I subgroup 0 0.15 0.10

0.05
0.00
10
15
20
25
30
35
3.0SL=0.1120
R=0.03429
-3.OSL=O.000
Figure 6.1 : Shewartchart of Zbar andrange of Length 2 from Case Study 1 (2 observations per group).
control charts can be found in Wadsworth et al. (1986)and Kenett and Zacks (1997). However, when several variables have be to monitored simultaneously, even if the univariate charts pertaining to all the relevant variables are created, these separate displays provide only partial information about the process. This is especially true whenthe quality characteristics are mildly or highly correlated. The pairwise scatterplots of the four diameter and two length measurements in Figure1.1 indicate that this is indeed the case for the aluminum pins data. There are also obvious difficulties in simultaneously analyzing multiple charts. In particular, when tracking several dimensions simultaneously, the decision that an observation is out-of-control is subject to a higher rate of false alarms than for a single chart due to the multiple comparisons effect. One approach to overcome these difficulties is to extend the univariate analysis by plotting a statistic which measures the overall deviations of the (multivariate) observations from the target. The most commonly used among such statistic is the Hotelling Ti presented in the previous sections. The multivariate charts which displaythe Hotelling Ti-statistic are sometimes called Shewhart charts as well (e.g. Crosier, 1988), although Shewhart has not been involved in their development As in the univariate case, when the data are grouped, the chart which displays the distance of the group means from the targets canbe accompanied by a chart depicting a measure of dispersion within the subgroupsfor
Analyzing Data with Multivariate Control Charts
79
""_
10
Groups
20
30
Figure6.2a: Ti-chart for the 6 variables from Case Study 1 (2 observations per group).
all the analyzed characteristics such as the ?';-statistic presented above, The combined T i and Ti charts thus provide a multivariate counterparts of the univariate (Z,R) or ( F ,SD)charts and display for the multivariate data the overall measures of distance from the target and dispersion within the group. We present in Figure 6.2a the T i and in Figure 6.2b the T i charts for the data on all six variables from the capability study of aluminum pins. The vertical position of each plotted symbol corresponds of observations. to the value of T i (or T i ) for the respective subgroup We observe fromthe multivariate control chart that starting the at 18-th subgroup and up to the 26-th subgroup the process shifts out of control (with the exception of the 23-rd subgroup). At the 26-th group the deviations fromthe target are large enough(at least on some variables) such that the overall T i value exceeds the critical 0.997 value. The process then seems to return to be in control (with the possible exception after the of group 33). We are not aware of any corrective action taken it out. The Ti-chart shows evidence 26-th group, although we do not rule of large dispersions betweenthe two observations within subgroups 31 and 33. As in the analysis of the Ti-chart, we cannot assess, fromthe overall Ti-values, which dimensions display large variations. The examination of the two T2-charts is simpler than the exam-
80
Chapter 6
ination of the six individual univariate Shewhart chartsfor each variable. Nevertheless, the multivariate Shewhart charts are considerably less popular than their univariate counterparts. Besides the relative difficulty involved in their computation, there are some additional features which diminish their appeal. First, unlike the univariate case, the scale of the values displayed on the chart is not related to the scales of any of the monitored variables. Secondly, when the T2-statisticexceeds the critical value marked on the graph, the signal conveyed to the user lacks the relevant information on what might have caused it. Further investigation is required to determine which particular variables are responsible for the overall outlying value. This information is clearly essential in any attempt to detect the causes of the deviations and possibly correct 7 some recently developed analytical the process. We present in Chapter methods designedto address this problem and in Chapter 11 we present some graphical methods designed to detect the causes of deviations. The usefulness of the graphical methods presented in Chapter 11 is illustrated by reinvestigating the data which form the above mentioned outlying groups.
-1
10
d
oroupe
3ci
Figure 6.2b: Ti-chart for the 6 variables from Case Study 1 (2 observations per group),
7
Detection of Out-of-ControL Characteristics
Objectives:
The chapterprovides several multivariate methodsfor detecting the variables that cause the out-ofcontrol signal. Special emphasis is devoted to the step-down procedure which ispresented in detail.
K e y Concepts
Subsets of variables 0 Step-down procedures 0 The T2 decomposition 0 Regression adjusted variables
0
81
Detection of Out-Of-Control Characteristics
83
In the previous chapters, we presented the Ti-statistics and control chartsfor testing the hypotheses that new a multivariate observation (or groupof observations) originates from the same distribution asthat which generatedthe target values. When the 7-value signals that an observation is out of control, an important practical issue is to detect the variables which caused that signal.While it is common in industry to monitor individual process characteristics, theuse of such procedures alone, is far from being satisfactory, since those procedures do not account for the correlations amongthe variables. In recent years a series of multivariate control procedures was developed for detecting the variables which are out of control. In general the methods use various functions of decompositions of 7;-statistics. The Ti-statistics are either the overall statistics obtained the from initial analysis, or those obtained f r o m the analysis of a subset of the p characteristics, Several methods are presented in this chapter. For the sake of simplicity, we restrict ourselves tocase thewhen the target valuesare assessed from a reference sample of size nz and the tested sample is a new p-dimensional observation (i.e. nl = l), but the extension to grouped data is straightforward. As mentioned in Chapter 5 , in this case the overallTi-statistic is:
7 ; = (Y- X)S(Y - X)
and under the null hypothesis,
When the Ti-value exceeds the critical value, we attempt to detect the variables responsible for that signal. The appropriate methods of analysis can be broadly divided into two groups: (1) procedures which assume that there is an a priori ordering among subsets of the variables. The tests are performed successively. At each stage (except the first), we test conditional hypotheses for a specific subset. The conditioning is on that the means of the variables in the previous subsets are as in the reference population (step-down procedures),
84
Chapter 7
(2) procedures which assume no specific ordering among the variables.
The step=down procedures. The new observation Y , whose T2-valuetested significant, has a distribution with mean py. Its covariance matrixX is assumed to be as in the reference population. We mentioned thatthe step-down procedure assumes an a priori ordering among the subsets of the variables. This ordering also governs the stages of the testings as well as the partitioning of the pr-vector and of the %matrix. Thus py is partitioned into q subsets denoted as pyl , . . , pyq ,respectively, i.e.
The mean vector px of the reference population is partitioned similarly into pxl, , . . , pxq. Let us denote further subsets of py and px as follows: p; includes all the variables belonging to all the original subsets from 1 to j (including). Similarlyfor p:]. N o t e that unlike the FYI,p , subsets ( j = 1, ,q ) , the p!j 'S (and the pEj 'S) are not mutually exclusive. Let pj be the number of variables in the j-th subset, satisfying
.. .
j-l
C pj
= p , and let qj =
C pj be the number of variables in p,

k=l
.i
(and
in PXj The step-down procedure tests sequentially

H p
My1
>.
= PXl
versus b y , # cc%, ;
9
then
Hp
Hdy)
Pyl = P X 2
versus Pyl # Px2 given P;l = P;l versus Py, # bx, given
and so on, until

CLy,
= cc,
Out-Of-Control Characteristics Detection of
85
The test statistic calls for the computation of a series of statistics q2 defined as
Ti = (Y- X)(j)SG$Y -
X)(j),
j =l,*..,q
which are the unconditional T,&statistics for the first (preordered) qj variables. Now, for testing thej-th hypothesis we compute the statistics
GI = 1
+ q:,/(n2 - 1)
TJ - TJ-1
j =l,,.,,q
where T o 2 = 0. The statistics which test the hypotheses in the step-down proceH 0 (Roy 1958, Subbaiah and dures are independently distributed under Mudholkar, 1978). So, if we set up a control chart for each test statistic, the overall levelof the procedure that simultaneously uses those charts can be easily obtained. This is a major advantage over both the procedures presented in Section 7.2 as well as over the procedures based on p univariate control charts. Another advantage the over p univariate charts is that inthe test for each subhypothesis, the procedure uses the earlier variables as covariates, and so it properly accounts for the correlation structure. Under the null hypothesis, the distribution of G; (conditional on G:,. . , G;-l) is
The q subhypothesis may be tested at different significance levels. Thus, the critical value for the j-th subhypothesis is
where olj are not necessarily equal.
86
Chapter 7
The step-down procedure does not reject H 0 : py = px if and only if Gj 5 U C Lj, for all j = 1, . ,q. The procedure canbe considered as an alternative to the regular T i chart and not only as its supplement. The crux of the method is the plotof the q statistics G;, . . , G: instead of the Ti-control chart. If at least one of the Gj values exceedsthe corresponding UCLj, the process is declared statistically out of control, and a special cause of variation mustbe sought. The search should concern the subset corresponding to the first G$ that signalled that the process is out of control. Note, that the control charts corresponding to later subsets in the a priori orderinglose their meaning since they test p, =pi assuming pyj-l which has been rejected. We mentioned that the statisticsG$ are independently distributed. Therefore the overall levelof the step-down procedure is
..
=,u~,-~,
j=l
The step-down procedure was first formulated by Roy (1958) for the particular case in which the vector p Y is partitioned into p subsets, each consisting of exactly one element. Subbaiah and Mudholkar (1 97 developed the general case where eachof the q subsets contains at least one element. For further references see Dempster (1963), Subbaiah and Mudholkar (1978b), Mudholkar and Subbaiah (1980a, 1980b, 1988), Kabe (l 984), Marden and Perlman (1980,1989,1990), Wierda (1994). Wierda (1994) analyzed the performances of the T i - control chart, the step-down procedure, and the univariate procedure. He considers two performance characteristics:the power to detect a certainshift and the ability to indicatethe type of shift. Basedon a simulation study Wierda (1994) concludes that in the great majority of the cases the power of the step-down procedure is very close to that of the Ti-control chart. Furthermore, in general,the power of the univariate procedureis lower than those of the multivariate procedures.The exception is in the case when the shifts in meanare of a medium magnitude relative to the standard deviations, and the shift structure is in accordance to the correlation structure.
a7
As for the ability to indicatethe type of shift, the analysis has to be restricted to the step-down and the univariate procedures since the Ti-control chart provides no indication of that issue. The step-down procedure gives indicationof the earliest among the q ordered subsets that has a mean shifted. Due to the conditional testing, nothing can be said about the means in later subsets the on ordering. The univariate control chart may provide extra information concerningthe characteristics contained in those subsets. For the step-down procedure, Wierda concluded that the highest probability of correctly detecting the characteristic (or subset) in which the shift occured, is obtained when the shift occurs inlater subsets. In such cases,the step-down procedure takes advantage of the correlation structure, having the persons characteristicsas covariates. As a result, he suggests that in practical settings, means which have a lower a priori probability of being shifted, should be included earlier in the orderings. ajS for those subsets. Also, he suggests setting smaller We performed a stepdown analysis on the tested ungrouped data (40 observations) from Case Study 1, as a two-step analysis: the first subset of variables contained the four diameter variables (i.e. variables 1-4); in the second step, we tested the significance of the contribution of the two length variables, conditional on the first subset. Ata = .05, the critical value for the G;-variable is 13.05.We found that only in one case (case 49) the value was statistically significant. Note that in the case with the overall largest T i (case 66 with T A = 82.63)there was no contribution due to the length variables and all the deviations from standard originate from the diameter variables. The test presented above is the likelihood ratio test for testing H 0 : p,,] = p, versus pyl # pxj,given that = p ~ j - Several I. other tests have been suggested, such as the tests based onW j = q2- q : , or on q2! = 1, . , ,q. Basing the test on q2alone is not recommended since It ignores the informationthat = p:]-,, andthustreats the variates andthe covariates alike. Marden and Perlman (1990)compare the tests and conclude that for a large of range alternatives,the test based on G; has a higher power and is thus preferable. The question on whether the assumption of an a priori ordering is too strong for practical settings is debatable. It can be argued that
88
Chapter 7
in many multivariate processes, not all quality characteristicsare likely to show a shift in the mean, or at least not as likely as the others. The ordering andthe levels of the step-down procedureare a quantification of those differences. The step-down procedure cannotbe used in cases in whichthere is a total lack of information on the a priori ordering. For such cases, Wierda recommends the case of the traditional Ti-control chart, supplemented by the univariate charts for detecting the responsible characteristics whenever T i signals for an out-of-control situation. He claims that ...for the time being thereis no alternative. However, the statement, as such, is far from being accurateas we can find in the literature several multivariate methods which have been proposed for the detectionof the out-of-control characteristics. However, the critical values suggested in the appropriate articles do not control the overall level of the test. We elaborate on this issue in the next section.
Other procedures for identifying out-of-control characteristics. Mason, Tracy and Young (1995) present a unified approach for the procedures basedon the decomposition of T i . the procedureis presented herealong with other methods whichare briefly mentioned.The advantage of the approaches presentedin this section is that no a priori ordering hasto be assumed. Let us start with the T i statistic
T i = (Y- x)S,(Y
-K)
and let us decomposeit into independent components, each reflecting the (conditional)contribution of anindividual variable. Assume that we wish to group the first p - 1 variables together and isolate the pth variable. Let us decompose the vectors Y, and the matrix S,accordingly. The Ti-statistic based only on the first p - 1 variables is given by:
Now, let us regress the pth variables on the first ( p - 1)-variables in the riz observations of the base sample. Let us denote by S ~ . ~ , . . .the ,~-I residual standard deviation from the regression. The predicted valueof the p-th variable of Y,given its values on the first (p - 1) variables, is

"
89
Yp= x,
+ C b j ( Y j - x,>,
,=l
and its standardized valueis
It can be shown that
Ti = Tj-l + T i 1 . 2 ,...,p-
I '
Since Tj-lis a regular Ti-statistic on p further partition it into
- 1 variables, we can
Ti-l = Tj-2 + Tj-1.1 ,...,p2 '

and so on. Continuing to iterate and partition, we obtain
The first variable, Tf is the Hotelling T 2for the first variable, whichis the square of the univariate t-statistic,
In the decomposition presented above there is an inherent ordering and as such, the assumption is not different in principle that from required in the step-down procedure. However, that decomposition is not the only one. There are p ! such decompositions and while the p terms within are independent of each other, they are not any particular decomposition independent acrossthe different decompositions. Under the appropriate null hypothesis, each of the terms is distributed as follows:
90
Chapter 7
Obviously, if one wants to ensure that the probability of at least one false rejection of the null hypothesis isa,the individual tests have to be performed at significance levels governed by appropriate multiple comparison procedures. Mason, Tracy and Young (1995) ignorethe problem and suggest performing all the tests at thea-level. overallGiven the large number of individual tests for any particular observation, we We suggest that it is important cannot agree with their recommendation. for the investigators to be aware of the possibility thatif they perform all the tests at level a,the actual overall levelof the test can be much larger thana. As mentioned at the beginning of this section, other methods of decomposition have been suggested as well. The test proposed by Murphy (1987) for each particular subsetof variables is based on the statistic This statistic contains a portion of the conditional terms in the decomposition of the T 2statistic, since if we arrangethe j variables as the first ones, we can write
e= j
Hawkins (1991,1993) defined a set of regression-adjusted variables in which he regresses each variable all the on others. His test statistics are the p-adjusted values, which can be shown to be related to the statistics presented in Mason, Tracy and Youngs (1995) decompositions by:
(1) zj
-Tiel
- s j . 1 , ...,j - I , j + I
g...
9.
j-l,j+l
I...,
,....p
An additional approach presented by Hawkins is based on the decomposition of the q2statistics, using the standardized residuals from the j-th variable onthe first j - 1 variables, i.e
(2)
Ti.l,*..,j-l
*
s j - l , .. j. , j - 1
The Statistical Tolerance Regions Approach
Objectives:
This chapter presentsa non-standzrd approach to multivariate quality control that relies on the construction o f statistical tolerance regions. The approach isparticularly appealing to engineers for whom, in many cases, the concepto f tokrance regions is vey natural The method is illustrated with Case Study 4 derived j h m an ongoingprocedureto monitor the quality o f c i mjuice marketed as naturalhit juice.
Key Concepts
Statistical tolerance region Level of confidence 0 Proportion of multivariate observations 0 Critical values 0 Cross validation
0 0
91
Statistical Approach Reglons Tolerance
93
A different approach can be taken to determine whether an observation, or a group of observations, come from a given multivariate normal distribution. Instead of setting up a null hypothesis, onecan identify a region which, under statistical control, contains a certain proportion of the multivariate observations. When the parameters p and E of the distribution are known and the resulting region is called a normal process region, already mentioned in Chapter 1. Each new observation Y is tested to determine whether it belongs to the prespecifiedbyregion computing P ( Y ) = (Y- p)C(Y - p ) which is compared tothe appropriate percentiles of the chi-squareddistribution with p degrees of freedom. We illustrate the normal process regions withthe data from Case Study1. Assuming a multivariate normal distribution with parameters equal to the empirical estimates from the reference sample of size 30, we usethe ungrouped estimate for the covariance matrix to compute the valueof T2(Y) for the 66th observation whose measurements on the six dimensions are (9.90,9.89,9.91, 14.88,49.99,60.14)
,
The value of T2(Y) = 82.63 is larger than the 99.9994 percentile of the chi-squared distribution with 6 degreesof freedom. We thus conclude that the new observation does not belong to the mid 99.999% of the population defined by the reference sample. When the multivariate distribution is bivariate, the testing may also be performed graphically by plotting the new observation on the graph with the normal process regions. Figures 8.1 and 8.2 present the 50%, 90%, 95%, 99% and 99.73% bivariate regions corresponding (respectively) to the parameters of the two pairs of variables (Diameterl, Length2) and (Lengthl, Length2) on thefirst 30 observations. The regions were calculated under the same assumptions asin the calculation of T2(Y). On each plot we also marked the position of the 66-th observation with respect to the plotted variables. We observe that the tested observation does not belong to the central 99.7% of the bivariate distribution defined by the first pair of variables. (The same is true for all the 14 bivariate distributions whose pairs of variables contain at least one of
Chapter 8
4 "
0.0
9.92
9.M
9.W
0.00
10
1D.02
1D.M
1o.m
Figure 8.1: Bivariate normal density ellipses of Length2 versus Diameter 1 with parameters from the first 30 observations fromCase Study 1 with 99.73%, 99%,95% and 50% contours and observation66.
the four diameters). However, on the second plot,the tested observation is well withinthe 90% normal process region. As mentioned, the computation of the normal process regions requires knowledge of the parameters of the underlying distribution. When we cannot assume that the parameters are know (as is usually the case) Stuone can still identify the distribution under statistical control using tistical Tolerance Regions (Fuchs and Kenett,1987,1988). A region of S) statistical tolerance regionif observations is said to constitute a (P, we can assert with a levelof confidence S that the proportion of multivariate observations contained in that region is greater or equal to P. In one dimensional analysis, the statistical tolerance regions are called statistical tolerance intervals (IS0 3207,1975) or statistical tolerance limits (ASQC,1983).A new observation,Y,is tested to determine whetherit belongs to a specific region by computing
T2(Y) = (Y- Z)'S"(Y
- Z) ,
where and S are the empirical vector of means and the covariance matrix, respectively, as computed from the reference sample.
Statistical Approach Regions Tolerance
95
69.116 49.7
.I
49.76
49.11
49.116
49.9
49.96
60.06
W.?
Figure 8.2: Length2 versus Lengthl.
When T2(Y) 5 K , the confidence levelin rejecting the statistical statement that Y belongs to the lOO.p% central part of the distribution is given by S . The practical implication of not rejecting the statistical statement is that the new observation is consistent with the process under statistical control. Fuchs and Kenett (1987) show that a good approximation of the critical valueK is given by
K =
g (P, g (1 - 6,(n
- 1) P - 1) P )
where P,(x; 5 g ( 8 , r ) ) = 8, i.e. g ( 8 , r ) is the lower Oth percentile of a chi-square distribution with r degrees of freedom. We illustratethe use of multivariate tolerance regions methodology in the context of quality control, with the data from Case Study 4 derived from an ongoing procedure to monitor the of citrus quality juice marketed as natural fruit juice. The data used belowconsist of a referencesample The of pure juice (n = 80) and of an additional specimen to be tested. attributes analyzed are labeled BX, AC, SUG, K, FORM,PECT (see Table 8.1).The means of the six attributes in the reference sample and in the tested specimen are presented inthe first two rows of the table. Table
96
Chapter 8
8.1 also contains the S the S " matrices computed from the reference sample. For example, the sample varianceof BX (firstelement in the S matrix) is:
z(X11 R n
h=l
XI)^/(^ - 1) = 0.38
where = 10.4is the mean of BX in the reference sample. Also from the S matrix, the sample covariance between BX and AC is:
x 1
C(X1p- Xl)(X,
80
- X 2 ) / ( n - 1) = 0.035
h=l
and so on. The quadratic form of Hotelling's T 2which measures the distance between the values of the tested specimen (Y)and those of the reference sample is: (Y - X)'S-' (Y - X)= 6.407, If we assumedfor a momentthat the parameters were known and equal to those found empirically in the reference sample,the value of T2(Y)= 6.407 would have corresponded to the 62.08-th percentile of xi, namely, the observation is not part of the 60% normal process region but it is contained in, say the 65% region. Let us now relax the assumption of known parameters and use the statistical tolerance regions methodology. From the formula presented above for the approximation of the tolerance region for Y , we derived critical valuesfor several valuesof P and S with n = 80,p = 6.These are:
S = 0.7 S = 0.9 S = 0.95 P = 0.5 5.539 5.825 5.969 P = 0.6 6.465 6.799 6.967 P = 0.9 11.003 11.571 11.858
By comparing the T 2 value of 6.407to above critical values, we conclude that the tested specimen is not part of the central 50% of the
Approach Regions Tolerance Statistical
97
TABLE 8.1
Fruit juice example (CaseStudy 4) BX AC
1.3
0.0
SUG
7.7 1.2
K
1180.0 277.1
FORM
22.2 2.3
PECT
451.0 84.4
35.4 24.5 1457.1 Y 8.9
1.3
11.9 10.4 1.5 0.38 0.04
,Y
-x
S-matrix BX AC SUG
0.04
0.29 0.26 0.01 70.04 6.22 K 38.83 43590 0.470.38 FORM 6.53 207 -0.22 19.79 PECT 16554 65.53 5026 12.05 -1.93
S" -matrix
BX AC SUG
36.61 -25.12 60.39 -35.01 20.96 38.31 0.02 0.00 -0.00 0.00 0.29
-0.00
-0.02 0.00 -0.67 0.35 FORM 2.45 PECT 0.01 0.01-0.01
0.00
reference population butis indeed contained in the 60% region. Obviously the specimen is also included in regions with coverage larger than 60%. The statement canbe asserted for confidence levels above 0.7. We note that we assumed the parameters to be known,the same T2(Y) did not correspond a tonormal process regionof 60%.The additional source of uncertainty createdby the estimation of the parameters the 60% statistical caused an expansion of the bordersof the limits and tolerance limits now include the new observation. Since the tested specimen was not found be to an extremecase of the reference population it is likely to be accepted as unadulterated. Alternatively, ifthe data analyzed above refer to a quality control test of an ongoing manufacturing
98
Chapter 8
TABLE8.2
Cross-validation results for juice example (Table entries are tolerance limits K and in brackets, the percentageof cases with T2distance below the tolerancelimits.)
loop
95% 90% 680% 70% 50% 11.00 8.89 7.52 5.54 0.70 (50.6%)(67.9%)(77.8%)(86.4%)(91.4%) 11.24 9.08 7.68 5.66 0.80 (50.6%)(67.9%)(77.8%)(87.7%)(93.8%) 13.73 11.57 9.35 7.91 5.83 0.90 (54.3%)(67.9%)(79.0%)(88.9%)(93.8%) 14.07 11.86 9.58 8.11 5.97 0.95 (55.6%)(71.6%)(80.2%)(91.4%)(93.8%) 13.05
13.33
process, the results indicate that there is no reason to suspect that the process is out of control. We note that unlike the normal process regions, the statistical toland erance regions are not as easily adapted to graphical representation computer programs are not readily available even in the bivariate case. The decision making should thus rely on the analytical computations. We proceed to evaluate the empirical performance of the tolerance region procedure underthe null hypothesis, i.e.the hypothesis that the process is in control. Empirical assessments are necessary since specific data sets usually depart from the theoretical assumptions and the behavior of statistical methods may be sensitive to such departures, The performance of the tolerance region methodology will be evaluated subsequently by cross-validation. In the technique of cross-validation each observation is successively compared to the reference sample made up of the remaining (n- 1)observations. The percentage of observations inside the tolerance region is then estimated empirically the by proportion of cases (among n comparisons) that the tested observation was included in the tolerance region determined by the other n - 1 observations,
Statistical Approach Regions Tolerance
99
The empirical results for our data set with n = 80 and p = 6 are summarized in Table 8.2. We observe that the empiricalpercentage of coverage is close to the postulated P and quite insensitive to the value of 6 . The method was thus found appropriate for routine testing of natural fruit juice authenticity. For more on the topic of multivariate tolerance regions andthe citrus fruit juice case study, see Fuchs and Kenett (1987).
9
MaLtivuriute QuuL i t y ControL with Units in Butcbes
Objectives: The chapterproviaks solutiom to situationsin which production are basedon the redts o f samplesfiom complete batches.Curtailing the inspection process is ah0 illustratedwith the objective to reach decisions faster and at lower costs. Indices that identiJj,various sourceso f variability are defined and used to analyze real l$ data fiom a ceramic substrate manuficturer (Case Study 3).
is organized in batches and decisions
K e y Concepts
Batch production 0 Out of control process 0 Corrective action 0 Curtailing of inspection 0 Overall variability decomposition
0
Units in Batches
103
This chapter presents a quality control procedure for production batches which relieson the multivariate statistical tolerance region approach discussed in Chapter8. Quality control of production lots typically relies on sampling and analyzing samples of units in order to determine if the production process is in statistical control. We consider here the situation in which the production is organized in batches and control is achieved at time intervals corresponding the to completion of entire batches. When the production process is determined to be out of control an alarm signalis given to trigger corrective action. The quality control procedure appropriatefor these conditionsis illustrated withdata from Case Study 3. A set of identical modules are manufactured simultaneously one on ceramic substrate plate. The modules are separated before final packaging along lines inscribed on the substrate with a laser inscribing process. Exact positioning and dimensions of these lines determine, to a large extent, the success of the printing and component mounting operations. The ceramic substrates comein production batchesof about 100 items. The quality control procedure used by the manufacturer is to draw a sample, whose size varies from 4 to 14, and carry out on these sampled units various measurements which are subsequently recorded. Because of the variety of modules manufactured, specifications of the substrate dimensions vary from batch to batch in nominal target values but not in specifications width. In this case study, weuse deviations of the actual measurements fromthe center of the specifications as the analyzed observations, This normalization enablestous treat all batches on a uniform basis.The assumption made is that thereare no location effects. A preliminary study validated this assumption. All five dimensions (a, b, c) and (W, L ) from Case Study 3 are considered. As mentioned in Chapter3, the (a, b, c) measurements are determined by the laser inscribing process while the (W, L ) are outer physical dimensions. These dimensions are measured on a different instrument than the first three. The two sets of measurements have in common the batch affiliation. The data presented in the Appendix on Case Study 3 detail the measurements for four (out of 31) such batches. There are two possible outcomes to the decision process:1) determine the production process to be in statistical control and, in practical
104
Chapter 9
terms, accept the batch, 2) stop the production flow as a result of a signal that the production process is out of control and start a 100% screening of substrates. During screening, we have to decide,for each item, if it meets our standard or if it has to be scrapped. We want to reach a decision (reject and stop testing entire batch) as soon as possible. The straightforward applicationof the multivariate quality control methods requires the testing of the entire sample. However, a complete test may be wasteful when results fiom the first few items tested suffices to stop the process. Therefore a curtailed decision scheme, that does not necessarily requireall units to be tested,may be preferable in this case. From the reference sample we determine standard achievable dimensions including the correlation structure between the measured variables. A stopping rule for curtailing the testing procedureis presented below. Let Y 1, , , , ,Y , ,be the tested sample batch.The curtailed testing technique is based onthe sequential analysisof the n observations. Define % = l if the T 2distance of the i-th tested observation is greater than K and V i = 0 otherwise. We assume that the observations are independent. We recall thatthe critical valueK given in Chapter 7 defines P of the standard population(with a region which covers a proportion confidence S ) . For the i-th observation, let us definethe null hypothesis state as: that the process is in statistical control to
Pr(% = 0) >, P.
The curtailing of the inspection can obviously occur only at an observation g for which V, = 1. Given that the inspection was not curtailed prior to the g-th observation, we define byN ( g ) the total number of individual rejectionsof the null hypothesisup to and including,the g-th observation, i.e. N ( g ) =
C x.To illustrate the notation, let us

i=1
assume that we test sequentially 13 observations and at the tenth observation we havethe second rejectionof the null hypothesis, withthe first rejection being at the second observation. Then we haveg = 10, v2 = 1 , v10 = 1 and vi = 0 for j = 1, . . , 9;j # 2. Also, N(10)= 2, but, say N(9) = 1 and so on. The number g, of observations upto the N(g)-th rejection, is compared to a critical valueR ( N ( g ) ) defined as follows: For each number of rejections U (out of n), R(U)+ is the min-
Units in Batches
105
imal number of observations allowed upto the U-th individual rejection, without rejectingthe overall null hypothesis. Thus, if g 5 R ( N ( g ) ) ,the N(g)-th rejection came too early and we curtail the inspection and declare that the proportion of coverage is less than P . To continue the previous example, if for a proportion of coverage P = .95, the corresponding values for U = 2 is R(U)= 7,then the second rejection at the tenth observation,does not warrantthe decision thatthe proportion of coverage is less than 95%. Let a be the prespecified significance level for the overall null hypothesis. The R(U)s are the minimal integer values such that under the null hypothesis
For fixed R ( U ) , the random variable
R(W
V, has a negative binomial
i=1
distribution and for a given a,we can compute R ( U ) from the inverse of the negative binomial distribution. Although a table of R(U)can be computed foreach 1 5 U 5 n, in practice, we need only compute the function for U 5 U 0 5 n, where U 0 is the minimal valuefor which
The value U0 defines the minimal number of individual rejections (out of n) for which we rejectthe null hypothesis. We would thus certainly rejectthe null hypothesis if we observe more rejections and/or if 0 rejection is less than n. Table the number of inspected items up to U the 9.1 presents the R(U)values for n = 13, P = 0.95, anda = 0.05,0.01, and 0.001. We observe that for the three values of a,U0 = 3,4and 5 , respectively. If for example, we adopt a = 0.01, we rejectthe null hypothesis at or before the third observation when C V, = 2 up to that stage. We 0 at or before the 9th observation if C V, = 3 (provided that reject H
106
Critical R (U) values for curtailed inspection
c11
Chapter Q
TABLE9.1
c11
u=2 U=3
= 0.05 7 13
u=4
= 0.01 3 9 13
c11
= 0.001
4 9
we did not rejectit already fromthe results of the first twoor three obwe obtain the fourth valueof V = 1, we servations). Finally, whenever curtail the testing process. We nowreturn to our example related the toanalysis of batches of ceramic substrates. The decision process in that analysis has two possible outcomes: to accept the batch as is or reject the batch and start 100%screening of substrates. The need for this decision process comes from our lack of confidence in the stability of the substate manufacturing process. During screening,we have to decide for each item whetherit meets our standard or has to be scrapped. In the sampling mode we want to reach a decision (accept or screen) as soon as possible. Therefore a curtailed inspection scheme should be used. The standard used in our analysis consists of measurements on a prototype batch of substrates which, when used in the subsequent manufacturing stages, proved to be of good quality in that, that later no problems related to physical dimension surfaced. From the reference sample we determined standard achievable dimensions includingthe correlation structure between the measured variables. An extended data set from Case Study 3 included the results of the analyses performedon the Reference batch and on 30 other batches of ceramic substrate measurements. The results of testing the Reference and the other 30 batches of ceramic substrates are presented in Tables 9.2 and 9.3. Table 9.2 contains the group means of the (a,b,c) measurements and additional information related to the adherance of the sample tothe meastandard. Table 9.3 contains the same information for the (W,L) surements. Based on the results on small sample approximations for K , the Mahalanobis distances are compared to the computed K - values. Any
Units in Batches
107
TABLE 9.2
Analysis of the batchesfor (a, b, c)-measurements means
SMPL m
REF 13 1 13 2 13 3 13 4 13 5 13 6 16 7 13 8 13 9 14 1 0 12
c RH
C . 1
T o 2
Ti
T; I1
I 2
11 13 12 12 13 13 14 13 16 13 17 13 18 13 19 13 20 13 2114 22 6 23 5 24 4 25 13 26 4 27 6 29 13 30 4
-1 .6 -9 -2.7 .4 . 4 .2 2 . 5 -.4 10 3 1 . . 7 7 .5 0 . 0 -.2 -.l -.l -2 . 3 -2.7 1 . 8 2.0 14.7 2.2 1 . 6 .4 -3.2 -4.3 .4 1.1 .5 . 7 1,3 .3 1 . 2 2 . 2 1 . 0 -.2 1 . 6 0.6 -.7 -4.8 1.3 .3 1 . 4 . 7 .7 3.0 1 3 . 7 1.5 0.9 .S 1.1 0.2 1 . 7 . 6 -.l .O .3 0 1 . 6 1 . . 6 6 -.3 .2 .3 0 0 0 -.5 0 0 0 .7 .7 .S .S .3 0 .8 .8 0 . 2 .2 .3 -12.3 -12.3
0 13 36.0 0 . 0 36.0 0.0 1 . 0 3 2 468.2 172.0 296.2 0.37 0.63 2 326.6 235.8 90.8 0.73 0.27 1 0 2 489.3 373.2 116.1 0.76 0.24 1 13 89.2 68.8 20.4 0.77 0.23 0 13 59.3 34.3 25.0 0.57 0.43 13 2 438.1 154.0 284.1 0.35 0.65 10 3 12045.1 3466.2 8578.9 0.29 0.71 13 2 3784.4 1315.2 2469.6 0.35 0.65 0 14 6 8 . 1 32.2 35.9 0.47 0.53 2 12 177.4 72.0 105.4 0.41 0.59 6 2 133.8 113.5 20.3 0.84 0.16 6 5 345.6 51.4 294.3 0.15 0.85 13 2 687.3 500.8 186.5 0.73 0.27 5 8 135.8 41.9 93.8 0.31 0.69 10 2 197.3 164.7 32.6 0.83 0.17 3 8 97.9 59.1 38.7 0.60 0.40 1 13 77.4 66.6 10.8 0.86 0.14 10 13 277.1 185.6 91.4 0.67 0.33 0 13 61.1 54.7 6.3 0.90 0.10 0 14 89.4 62.7 26.6 0.70 0.30 0 6 2 0 . 5 11.8 8.7 0.58 0.42 0 5 22.8 22.8 0.0 1-00 0.00 0 4 0.54 16.9 7.7 0.46 9.2 0 13 29.8 18.7 11.1 0.63 0.37 6.7 0.61 0.39 0 4 17.2 10.4 0 6 1 5 . 5 7.7 7 . 8 0.50 0.50 89.0 72.0 17.0 0.81 0.19 2 13 2 4 9913.6 2861.9 7071.7 0.29 0.71
observation witha T2value beyond the 95%tolerance region leads us also include information on to rejectthe production process. The tables the rejections(REJ), as wellas the number of items (C.1) to be inspected if we adopted a curtailed inspection technique. Eighteen batches were rejected on the basis of ( U , b, c)-measure-
l08
Chapter 9
ments; seventeen batches were rejected on theof basis (W, L)-measurements.Allrejectedbatches are subjected to 100% screening.Only batches 20, 22-27 were determined as not statistically different form the reference batch, implying that for these batches the process is in control and that there is no need to screen the production batch. We note from the results that in most the cases non-standard batches could have been detected the in very early stages of the testing process. In several of the groups in which thereare considerable numberof rejections, the first two tested items yielded T2 distances larger thanK and the testing process could thus have been curtailed. We mention in this context that someof the batches deviate considerably fromthe reference samples. On the other hand, the 13 (W, L)observations in batch 18 and the 5 (a,b, c)-observations in batch 23 were at their respective centers of specifications. The Tt-values in the two cases differ, however, since inthe reference samplesthe means of the (W, ,?,)-measurements were at their centers of specification while the means of the ( U , b, c)-measurements deviate from the centers of specification. The rejection strategy defined above refers to the analysis of the individual observations within the sample and can lead to a decision before the whole sample is tested onall dimensions. Once testing is completed, we can proceed to analyze the decomposition of the overall variability computed from the entire sample batch (Tt) into the two measurementsof variability: T i and Ti. The decomposition of Tt can be perceived as a descriptive tool rather thanas an exact testing hypotheses procedure. Based on the.fact that Tf, T i and T i have asymptotic x 2 distribution with np, p and (n - 1)p degrees of freedom, respectively (e.g. Jackson (1985)), we compute two indices I 1 and I 2 for detecting whether the overall variability is mainly due to the distances between means of the tested sample batch and those of the reference sample, or to variability withinthe tested sample batch.The indices are defined as normalized versionsof the the two components of . T : Since the expected value of a chi-squared variable equals its number of degrees of freedom we first divide T i and T i by p and (n - l)p, respectively, i.e.
Units in Batches
109
TABLE9.3
Analysis of the batchesfor the (U:L)-measurements means
SMPL m
REF 1 2 3 4 5 6 7 8 9 1 0 11 12 13 14 1 5 1 6 1 7 1 8 1 9 20 2 1 22 2 3 2 4 25 26 27 28 2 9 30
L RH ( 2 . 1
T o 2
Ti
Ti -
11
12
13 0 0 13 13.5 4 . 5 13 13.8 0.6 13 1 6 . 1 4 . 5 4.4 4 13 . 4 5 . 5 4 . 8 13 13 1 6 . 8 7 . 5 0 . 3 0 . 3 14 1 13 13,s 0 . 7 3.2 3 . 2 13 8 0 . 6 0 . 6 5 0 . 4 1.2 13 11.3 -6.4 13 -4.0 15.8 9.2 -5.6 13 4 . 1 3.7 13 0 . 7 0.0 13 13 9 . 1 -1.5 13 0 . 0 0.0 13 11.1 3.1 -0.7 13 -0.3 13 1 0 . 2 1 13 -2.0 -2.6 13 1.1 1 . 2 13 -0.2 0.5 13 0.9 0.8 13 1.1 0.8 6 1.5 1.3 13 -35.9 -43. 13 -0.2 -0. 13 0 . 1 -1.1
013 2 4 . 0 0.0 24.0 0.00 1.00 13 2 1082.1 821.9 260.2 0.76 0.24 13 2 11 13.0 842.5 270.5 0.76 0.24 13 2 1300.9 1156.1 144.8 0.89 0.11 1 2 2 1 8 7 . 0 182.5 4 . 5 0.98 0.02 1 1 5 276.3 239.8 3 6 . 5 0 . 8 7 0.13 14 2 2666.3 1491.8 1174.5 0 . 5 6 0.44 1 2 3 11 12.2 475.2 637.0 0.43 0.57 0 3 . 1 522.2 0.61 0.39 13 2 1325.3 8 1 1 3 . 6 6 3 96.1 1 7 . 5 0.85 0.15 0 8 4.5 3 . 4 0.24 0.76 1.1 12.8 0 5 8 . 9 0.31 0.69 3 . 9 13 2 1529.3 1062.1 467.2 0.69 0 . 3 1 13 2 2438.7 2095.8 343.9 0 . 8 6 0.14 12 2 1079.4 749.4 330.0 0.69 0.31 8 2 181.7 141.6 40.1 0.78 0.22 1 0 . 1 0 13 7 . 9 0.22 0.78 2 . 2 498.6 422.8 75.8 0.85 0 1 2 3 . 1 5 0 13 0.0 0.0 0.0 11 2 1426.6 550.2 876.4 0.39 0.61 43.1 0 13 3.5 3 9 . 6 0.080.92 1 13 31.4 11.1 2 0 . 3 0.350.65 6 2 . 1 0 13 56.0 6 . 1 0.90 0.10 17.4 12.6 4.8 0.72 0.28 0 13 1 3 . 7 2 . 7 0 13 11.8 0.20 0.80 1 4 . 5 7 . 3 6.8 0 . 5 3 0.47 0 13 0 13 13.8 8 . 3 5.5 0.60 0.39 0 6 10.4 8 . 5 1 . 9 0.82 0.18 . 3 9 0 . 6 1 1 2 2 38288.2 15025.8 23262.4 0 0 13 9 . 0 7 . 9 0.12 0.88 1.1 25.2 1 6 . 8 0.33 0.67 0 13 8 . 4
110
Chapter 9
I; = Tj/(n- 1)p
and thendivide those values by their sum, thus obtaining the indices I1 and 12 which sum to 1,
It can be shown that in termsof the original T2-statisticswe have

I 1 = (n
- l)Ti/[(n
- 1)Ti + Ti]
and
I2
=Ti/[(.
- 1)Ti + 2-21.
When nosource of variability significantly supersedes the other, the two indices should be approximately of the same order of magnitude. The normalized I 1 and 12 are an extended version of the indices proposed by Fuchs and Kenett (1988). The indices Z1 and I2 for the data in our example are presented in Tables 9.2 and 9.3. of variabilThe indices I 1 and I2 are helpful in identifying sources ity. We illustrate the assertion with two examples based the on (W, L)measurements. In batch 4 we conclude from the very low value of I2 (0,001) that there is almost no variability within the sample.The large number of rejections (12 out of 13) indicates an almost constant bias of the magnitude of the sample means with respectthe toreference sample (and, in this case, also with respect to the centers of specification). On the other hand, in batch 20, a significant part of the variability is due to the variability within the sample (I2 = .49), with the (W, L)-means being close to'those in the reference sample. The variability of the (W, L)-measurements in batch 20 did not suffice to yield rejections of the individual substrates. In other cases, such as in the (a, b, c)-measurements in batch 12, the deviations were large enough to generate six rejections. Further analyses in a similar manner can be performed on the indices of the other samples. We note that several measurements differed so drastically from the standard, that errors in 28 recording are to be suspected. The13 ( W , L)- observations in batch
Units in Batches
111
yielded a very large T&value due to large deviations from the standard. The substrates were found to be smooth and not inscribed and thus no (a,b, c)-measurements are available. The batch has obviously been rejected. Data on the (a, b, c)-measurements were also not available for batch 15.
10
Applications of Principal Components
Objectives:
The chapterprovides proceduresfir constructingprincipalcomponents with an empbusis on their interpretation in multivariate In some cases,principalcomponents quality contvolapplications. have provento be an effective approachin reducing dimensions of multivariate duta.Tbe aluminumparts data (Cae Study I ) is used to illustrate the approach.
K e y Concepts
Eigenvalues Eigenvectors 0 Principal component charts
0
0
113
115
A quite common approach in multivariate quality control is to attempt to reduce the dimensionality of the data by transforming the origidata set obtained by identifying nal p variables into a lower dimensional several meaningful linear combinations of the p dimensions. Theconstruction of specific linear combinations, calledprincipal components,is described here. Case Study 1 will be used to illustrated in the application of principal componentsin multivariate quality control. X?), . , X,") be the i-th p-dimensional obserLet X = (Xi'), vation in the original data. We create a new p-dimensional observation (Z?, . , , Zy)), such thatthe t-th variable inZ, Zf) is a linear combinations of the deviations of the original p dimensions from their targets, i.e.:
..
U-l
Let us now present in moredetail the rationale andthe construction of the principal components. For a process with multidimensional data, its overall variance is defined as the sum of the variances of the pvariables, i.e. the sum of the diagonal elementsthe ofcovariance matrix X = [ ( ~ e ~ ) ] .If we denote by 81,82, ,e, the characteristic roots (eigenvalues) of the covariance matrix X, it is well known that their sum equals the overall variance of the process defined above, X& i.e. = Zgee. Furthermore, if we order the characteristic roots such that 8 12 O2 2 2e , the matrix C whose rows are the corresponding characteristic vectors (eigenvectors)Cl,. . ., C, satisfies CC' = X,C ' C = SZ where SZ is the matrix havingthe 8's on the diagonal and zero for all the off-diagonalelements. Now it can be shown that if we make the linear transformations
. ..
the new Z,?) variables will be uncorrelated, and if the covariance matrix X is known their respective variances are &,t = 1, , , , p. For the i-th observation, the linear combinations Z(') = Cl'(X - m) has the maximal variance among all possible linear combinations of the X's. The proportion 81/ X& is sometimes termed the "variance explained" by
116
Chapter 10
the first principalcomponent. Similarly, for &/ X& and so on. Parsimony the ratio e,/ E& is achieved if for the several last principal components is small enoughto justify ignoring the corresponding 2-variables from the analysis. When the covariance matrixX is not known, but rather estimated, the 2-variables will still be orthogonal, and 51 is now their estimated rather than exact covariance matrix.
Analyzing Data with Principal Component Charts. For each multivariate observation, the principal components are linear combinations of the deviationsof the p variables from their respective targets. Principal components have two important features are which useful in quality control:
(a) the new variables are uncorrelated(or nearly so), and (b) in many cases, a few principal components may capture most of the variability in thedata so that we do not need to use all the p principal componentsfor control.
Let us illustrate the feature of parsimony mentioned in (b) above with the data from Case Study 1.The first columnof Table 10.1contains the eigenvalues of the inverse of the covariance matrixfor the first 30 ungrouped data points. We notice thatno more than three out of the six principal components are worth analyzing. The first principal component explains 67% of the overall variation,the next one is responsible for additional 21.5% and the third explains another 10%. The variance accountable for the last three componentsis ignorable (less than 2%). The advantages of principal components charts over individual Shewhart control charts (uncorrelated variables and possible parsimony) are obtained by losing the original identity of the variables. However, in some cases, the specific linear combinations corresponding to the principal components with the largest eigenvalues may yield meaningful measurement units. Inour example, the columns labeled Veclto Vec6 of Table 6 display the eigenvectors corresponding theto six eigenvalues presented in the first column.These are the coefficients of the original variables used in the definition of the principal components. The first three principal components, which explain almost all the variance, are
117
TABLE 10.1
Eigenvalues and eigenvectorsof covariance matrix first 30 observations of case study 1
Vet 1 -0.04
-0.05
99.69
100.00
-0.04 -0.03 0.70 0 . 7 1
-0.26 0.07 -0.84 -0.40 0.60 0.46 0.84 0.01 0.06

0.05
-0.03
11
-0.01 0.01 -0.02 0.00
clearly interpretable in terms of the physical dimensions analyzed, The first component is defined almost exactly as an average of variables 5 and 6, i.e. the two length variables.The second componentis basically an average of the other four variables which are diameter measurements at various locations on the pin, althoughthe contribution of the length measurements is not totally ignorable. Finally, the third componentis interpretableas the difference between variables 6 and 5, which is in our case the length of the main partof the pin (without the cap). Since the principal components are linear combinations of the deviations of the variables from the targets, when the process is under control their values should be around zero with variances determined by the eigenvalues. Thus,if we divide allthe principal components by their respective standard deviations, (i.e. by the square roots of the respective eigenvalues) we can plot the resulting standardized principal components on charts with mean value set at zero and constant UCL and LCL of f 3 corresponding to 3-standard deviations control limits. Each principal componentcan be plotted on a separate chart and pairs can be plotted on a scatter of plot two (usually the first two) components. Figure 10.la and Figure 10.lb present the principal components charts for the first two principal components and Figure 10.2 displays the corresponding scatter plot. We observe thatthe first principal component has significant negative value at the 48-th observation andsignificant positive value for observations 49 and 61.The second principal component hasa significant outlier at the 66th observation. Giventhe
118
4 ,
Chapter 10
._
MearrO
-3
-4
--- 40
I
LCL
I 10
20
50
80
70
Obeervatlon
Figure 10.la: Principal components chart for the first component of data from Case Study 1 (ungrouped).
coefficients of those components, those outlying values should correspond to outliers in the length and diameter measurements, respectively. We shall confirm this observation later in Chapter11 where we present methods for simultaneous displayof univariate and multivariate statistics. Meaningful interpretation of the first principal components are quite common in practice and have been reported in the literature. In a frequently quoted example of ballistic missile data (Jackson 1959, 1985), the coefficients inthe first principal component had the all same sign and were roughly of the same magnitude. The resulting variable was thus related to the average of the original variables and could be considered to measure the product variability. Further, as in the ballistic missile example, when the coefficients for the next one (or more) principal components differ in signs andare pairwise of the same order of magnitude, the resulting principal components represent differences between the means of the respective subsets of variables. Obviously, the coefficients of the variables in the principal components do not have to obey the quite simplistic patterns mentioned above. Sometimes, as in our example, the coefficients corresponding to several
Applications of Components Principal
119
"""""_
I 10
I 20
30
I 40
ObWstiC#l
I 50
I 80
Flgure 10.1b: Principal components chart for the second component of data from Case Study 1 (ungrouped).
"4
" " " " " " "
I 1"I P48
m
P17
-1 -2
I I I I
=P81
-5
-4q
-4
- -1I
I -2 I -1
I
I
-I
-3
I l
Flrst Principal Component
Figure 10.2: Scatterplot of the first two principal components of data from Case Study 1 (ungrouped).
120
Chapter 10
attributes are close to zero. The resulting principal components canthen be considered to represent linear combinations of only a subset of the p variables. Moreover, whenthe coefficients for the various attributes differ in their order of magnitude, other types of interpretations (if at all) can be attached to the resulting principal components, such as the comparison of one variable with the average of several others and so on. Since the principal components do not provide an overall measure of departure of the multivariate data from the norms, those charts may complement but not replace the multivariate T2-chartspresented inthe previous section. The analysis of the several charts simultaneously suffer from the problems mentioned in connection withthe assessment of an out-of-control situation from the control charts of the original variables (although the number of relevant charts is likely to be somewhat reduced in the principal components analysis). Moreover, as the inTi-chart, the individual attributes most prominently responsible for an out-of-control observation are unidentifiable.
l1
Additional Graphical Techniques for Multivariate Quality Control
Objectives:
The chapterprovides additionalgraphical techniques usingcomputer technoloafir analyzing multivariate data in quality control applications. The star plots and the MP-charts areparticuhrly mphaked. Key Concepts
Computerized graphical displays Symbol plots 0 Three dimensions plots 0 Scatterplot matrix 0 Start plots 0 MP-charts
0
121
ues Graphical Additional
123
Background on Modern Computerized Graphical Displays

The display of data is a topicof substantial contemporary interest and one that has occupied the thoughts of many scholars for almost 200 years. A pioneer in this area was W. Playfair who in his 1786 Commercial and Political Atlas displayed differences between English imports and exports in informative and resourceful ways (see e.g. Costigan-Eaves and MacDonalds-Ross, 1990).A vision for data analysis in the 80s was a perspective focused outlined by John Tukey (1962), who has reiterated on data-based graphics(Wey, 1986). Tukeys vision has largely materialized inthe Exploratory Data Analysis approach whose foundations was established in the books by Tukey (1977), Mosteller and Tukey (1977), and Velleman and Hoaglin (1981). Modern computerized data analysis builds on advances in computer technology and statistical methodology. Computing environments specifically designed for data analysis have evolved sincethe 60s on computing platforms that reflect the state of the art in computing technology: first mainframes, then personal computers, and now workstations. Such environments incorporate aspects of data analysis that include model identification, diagnostic checking, transformation the and usual inferential procedures (misted, 1986; Huber, 1987). Becker et al. (1987) specify requirements for computerized dynamic graphics methods that include identification, deletion, linking, brushing, scaling, and rotation. These techniques are briefly defined below and some of them are then illustrated in Figures 11.1a-l 1.IC with data from Case Study 2. 1. Identification: Labelling and locatingof specific observations on graphical displays. 2. Deletion: Deleting specific points from a graph and redrawing the graph. 3. Linking: Linking methods that make effective use of color andor plotting symbols for indicating the value of a dimension not explicitly present in the graphical display. 4. Brushing: Conditioningthe analysis on the different valuesof any variable. 5. Scaling: The ability to refocus, move and resize graphs.
124
232.7
Chapter 11
+
+ +
0
*
0* 0 . 0
*
0
+
+ +
+
232.6
-+
+
+
*
+
+
29.08 29.08
+
+
+
+
232.5
*
0
I I
30
10
20 Obaesnration
Figure 11.l a: Time sequence plot of variable 15 from Case Study 2 with symbols indicating levels of variable 14.
28.98 4- 29.00
0
46
8
29.04 29.08
40
Figure 11.l b: 3D-Plot of variable 15 from Case Study 2 with symbols indicating levels of variable 14.
125
lo+
* .I
'
6,
b 9
126
Chapter 11
6. Rotation: The ability to rotate three dimensional scatterplots in in a either an automatic continuous motion mode ormanual mode.
Several software packages for personal computers, that include such features, are now available (e.g. MINITABTM, EXECUSTAT? J@, DataDeskQand STATGRAPHICSq. Case Study 2 consists of data collected in a process capability study of a complex mechanical part on which 11 diameters, 4 length and 2 wall-width measurements were recorded. In our analysis we focus on the overall length of the part (Variable 15) and use MINITAB for the analysis. Figure 11,l a presents a time sequence of plot Variable15 using symbols to indicate the levels of Variable 14, an internal measurement of the part. The shortest part is the 22nd (232.495 mm). Its value on Variable 14 (29.06 mm) is one of the largest in the sample. However, such large values of Variable 14 occurred also on longer parts. Figure 1 l . l b presents a three dimensional scatterplot of Variable 15 versus the 1. Again, the points are labelled observation index and versus Variable with symbols indicating the level of Variable 14. The figure indicates that the group of the first 10parts are low on both Variable1 and Variable 14. We observe an outlying point with a very low value for Variable 14, and an unusually high value for Variable 1. Figure 1l.lc is a matrix of scatterplots for Variables 1,2, 3,4 and 15, again, with the labellingof the observations to indicate the level of Variable 14.The two diameters - Variable 1 and Variable2 are clearly linearly related.The same type of relationship appears to hold between Variable 3 and Variable 4. It is also worth noting that the symbols for Variable 14 tend to cluster in the Variable 2 versus Variable 3 scatterplot, thus indicating a further correlation with Variable 14. The graphical displays presented above provide the reader with a glimpse at the capabilities of modern interactivedata analysis. Indeed, computerized data analysis has become a new discipline, not merely an extensionof Playfair's graphical innovations. Industrial statisticians are now developing comprehensive approaches to computerized data analysis. Comprehensive example of a strategy for on line multivariate exploratory graphical analysis was developed by the CIBA-GEIGY Mathematical Applications group in Base1 (Weihs andSchmidi, 1990). For another example of dynamic graphics for exploratory analysis of
127
spatial data using a"map view" see Haslett et al., 1991. For an algorithmic approach to multivariate quality control see Chua and Montgomery, 1991.
Graphical Techniques for Quality Characteristics in Multivariate Quality Control. In the previous chapters, we mentioned that when T2-value a or the value of a principal component corresponding to an observation (orto a group of observations) exceeds the critical limit marked on the graph, one cannot identify the attributes primarily responsiblefor the overall outlying value. This has been illustrated in our example where several as outliers, but without the observations (or subgroups) were identified indication on the behavior of the particular characteristics. Additionally, on the chart is unlike the univariate case, the scale of the values displayed not related to the scales of any of the monitored variables. Both pieces of information are relevant and important when one monitors the process. On the other hand, we have already described the shortcomings of relying only on univariate charts when the data are intrinsically multivariate. Thus, in the multivariate case, neither the univariate charts for each variable, nor the multivariate chart provide by themselves a complete picture of the process. The problems raised above can be considerably alleviated if all the information about the individual characteristics as well as the overall measure of distance from targets are displayed on the same chart. Several graphical methods for coping with these problems have recently in chapter: the been proposed.Tho such methods will be presentedthis starplot (STATGRAPHICS 3.0,1988) and the MPchart (Fuchs and Benjamini 1991). These plotting techniques are illustrated with the six dimensional data from the process capability study for turning aluminium the previous sectionwe start by anapins (Case Study 1). As in most of lyzing the data grouped in pairs. Along the presentation of the graphical display we will return and address someof the findings mentioned in chapters 5 and 9, thus complementing the analyses performed with multivariate quality control and principal, component charts. The quality characteristics which are to be displayed simultaneously on the charts are on different scales. Therefore, proper scaling of the data is required. Furthermore, if we want to relate each variable to
128
Chapter 11
its reference value, the statistics displayed have to be (scaled) deviations from those targets. Let us define for the 8-th variable in the j-th subgroup of size n, the standardized deviation of
x:,
(4
where X:; is the i-th observation in subgroup j on the 8-th variable, Xj - C Xb/n,m(.!) is the target value and v.! is the scale factor for the 8-th variable, estimated from the base sample. As an alternative, one can consider the scaled data rather thanthe scaled deviations,i.e.
The multivariate measure of distance from target displayed on the three types of charts is the T2-statistics. When the data are grouped, we can also construct simultaneous charts for dispersion. In those cases, the multivariate measuresof dispersion displayedare the values of T;f as presented in the previous section. On the T2-charts, we can also display some common univariate statistics (i.e. the ranges or standard deviations), again after proper scaling. The simultaneous chartsfor dispersion have been developed only for the MP-charts but obviously they can be constructed as companions to the starplot. The multivariate charts presented here are symbolic scatter plots. The starplot for groups is shown in Figure 11.2 and the MP-charts is displayed in Figure 11.3. Theoretical considerations and empirial evidence showed the MP-chart to press desirable properties (Fuchs and Benjamini, 1991). The starplot has already gained relative popularity and has been implemented in other software packages (e.g. SAS). The stars in the starplot and the symbolsin the MP-chart are positioned vertically at the group T2-value. Thus, those charts resemble the multivariate Shewhart charts, except that for each group we display We describe at its T2-value a more complex symbol than a single dot. below the two types of charts and illustrate their use with the data from Case Study 1.
Additional Graphical Techniques
129
Flgure 11.2: Starplot of the 6 variables from Case Study 1.
Starplots The starplot was apparently first developed at the SCS corporation as an enhancementto the Multivariate Control Charts available through the STATGRAPHICS statistical software package. The starplots consist 11.2). of stars positioned verticallyat the group T2-value (see Figure Each multivariate observation is portrayed in the starplot bystar a with the rays proportional tothe deviations of the individual variables from the minimal values across all observations (or groups of observations). Thus, the effect of displaying the hy)-values rather than the df)'s is that when whenthe process is under control the stars are of medium size and when a variable is out of control with a value considerably below the target, the ray shrinks toa minimal value, attracting minimal visual attention. Moreover, that starplots give no indication of the statistical significance of the ray's length.
MP-charts. The MP-charts were proposedby Fuchs and Benjamini (1991) as an improvement over the starplot and other methods proposed in the literature.
130
Chapter 11
1' 234
5 67891011121314151617181920
Groups
..........................................................
..
84
123
4 5 6 7 8 91011121314151617181920
Groups
Figure 11.3: Mp-charts for means and dispersion for the 20 tested groups from Case Study1,
131
The symbol in the MP-chart is an adaptation of a profile plot which encodes visually the size and deviation of the group average of each variable from its target. After the horizontal base line of a symbol is drawn, a bar is constructed sequentially for each variable,The bar extendseither above or below the base line according to the signof the deviation, The bars are proportional to the absolute value of dy up to a value of 4, after which they remain at a constant length of 4. The bar darkens (or changes color) as the level of significance of the deviations increases (see Figure 11.3). Minor adjustments are necessary on the vertical axis to account for the size of the symbol. The critical value corresponding to T2-value a of ,997 (equivalent to the 3 standard deviations rule in the univariate chart) is at one symbol-size distance fromthe top, and that corresponding to 0 is placed half a symbol size distance fromthe bottom. As mentioned, the MP-chart for dispersion is constructed along the same lines as the MP-chart for means and together theyenable the investigator to study both the level andthe variation in the process.
Reexaminationof Case Study1. We conclude this chapter with a renewed analysis of CaseStudy 1. The analysis provides a comparison of graphical methodsas well as an extension to the analysis derived earlier by the use of multivariate control charts and principal components. Let us now reiterate the main findings from the aluminium pins process capability study mentioned in the analysis of the Shewhart charts, the T2-charts and the principal components:
(a) From the Shewhart charts on the sixth variable (Lengths), we observed that the overall length is significantly below target in groups 24 and 26 and that that there are large dispersions between the two observation which form group 25 and 31 (Figure 6.1). (b) In the analysis of the 7;-charts we mentioned that at the 26-th subgroup the deviationsfrom the target are large enough (at least on some variables) such that the overall Ti-value exceeds the critical 0.997 value. Also, we noticed that the Tj-chart shows evidence of large dispersions between the two observations which form the subgroups 3 1 and 33 and, in a lesser extent, subgroup 27 (Figures 6.2a and 6.2b).
132
Chapter 11
(c) The assessment of the principal components (for ungrouped data) a significant negarevealed that the first principal component had tive value the at 48-th observation and significant positive for value observations 49 and 61. The second principal component had sig- a nificant outlier at the 66-th observation (Figures 10.la and 10.lb). We mentioned that the the coefficients of the first two components are such that the components roughly correspond to the mean of the two length measurements andthe mean of the four diameter measurements, respectively. The Shewhart controlchart in Figure 6.1 flags subgroups24,26, and 27 as having a significantly low valuefor variable 6. Only subgroup 26 stands up as significantly different onthe T2-chart. The starplot indicates only implicitly the decrease in variable 6 from subgroup 10 to subgroup 24 and that the measurements are significantly below target. See forexample subgroup 33 in Figure 11.2.As mentioned above, another drawback of starplots is that the rays carry no indication about the significance level of the departures from targets. We nowturn to examine Figure 11.3 which contains the MP-charts for means and dispersion for the last 20 tested subgroups. First we mention that the findings fromthe Shewhart charts about the sixth variable (Length2), are also directly observable from the MP-charts. Secondly, about the findings from the T2-charts,we observe that indeedthe in26-th subgroup (i.e. the 1l-th tested subgroup) the value of the sixth variable is significantly belowits target. The second andthe third variables exceed 2 to 3 standard deviations. On the other hand, their respective targets by subgroup 33 (i.e. the 18-th tested subgroup) has three highly significant deviations from target(at variables 1,2, and 4) but its overall T2-value is smaller than in group 26. The apparent inconsistency is easily explained, asit it clearly seen from the graph, by checking the correlations among the variables. In group 26 (the 1 l-th tested subgroup) a very large deviation in the sixth variableis accompanied by a much smaller deviation in the fifth one. This pattern does not correspond to the the high positive correlation between those variables in the base sample, On the other hand, the pattern in subgroup 33 (the 18-th tested subgroup) is as expected, i.e.,the four diameter measurements deviate together (in this case considerably belowtheir targets) and the two length measurements
Additional Graphical Techniques
133
134
Chapter 11
are again together (less than 2 standard deviations above their respective targets). The large within group dispersions which yielded the large T i values for groups 31 and 33 (the 1 6 4 and the 18-th tested subgroups, respectively), are caused by two different sets ofvariables.We observe from the MP-chart for dispersion that in group 31 the two length measurements vary between the observations (61 and 62) which form group 31. In group 33 the diameter measurements are responsible for the large Ti-value. Tbo interesting patternsare observed for groups 25 and 27 (the 10-th andthe 12-th tested subgroups, respectively). In group 27, although none of the within group variances are mentionable, the Ti-value is large. The reason lies in a quite intricate configuration of correlations among variables 1-4 in the two observations 53 and 54 which form that group.On the other hand, in group 25 although large within group variances are observed for the length measurements, the overall Ti-value is quite small. This is due to the unusual small variances inthe first three variables, and to the adherence to the postulated correlation structure. The pattern of the variables for the observations within the subgroups are best observed from theungroupedMP-chart for means. Figure 11.5presents that chart for the 40 tested observations. From chart that we can also reexamine the findings from the principal components chart. The significant negative value of the first principal component at the 48-th observation (i.e. the 1 8 4 tested observation) is immediately related to the significant negative deviationsthe intwo last variables which are the length measurements. Similarly, significant positive deviations in those variables in observations 49 and 61 (i.e. the 1 9 4 and the 3 1 4 tested observations) yielded significant value in the first principal component. The interesting pattern in observation 66, (i.e. the 36-th tested observation) where the diameter measurements (first four variables) are significantly below their targets, while of both the length measurements exceed by 2 to 3 standard deviations their respective targets, are reflected in the significant negative value of the first principal component and in the borderly significant value of the second principal component.
12
Implementing Multivariate Quality Control
Objectives:
The chapter provides a smtegy anda generic plan for impkmmting the multivariate qualitycontrol method presented in the boob The proposed plan spec$es steps and milestones that have been proved to be effective in successfil implementation of quality control.
K e y Concepts
Quality control process Control subjects 0 Common causes 0 Chronic problems 0 Special causes 0 Sporadic spikes 0 Strategic plan for implementing Multivariate Quality Control
0 0
135
Implementing Multivariate Control Quality
137
Maintaining controlis a basic activity that can be observed wherin its ever there is change. For example, the human body exerts control efforts to maintain body temperature, blood count,other and physiological functions. Industrially,it takes the form of meeting goals: delivery according to schedule, using funds according to budget, and producing quality accordingto specification. In thischapter we discuss the basic elements of multivariate quality control and then present a planning strategyfor developing such a quality control process. The chapter provides its to reader an integrated review of the topics presented inthe above nine chapters. J u m (1988) specifies a series of 7 steps which, when applied to problems of quality, effectively implement a quality control process. These steps create a feedback loop which produces quality control. First, a goal, target or standard is established as the intended value of the control subject, Then, when it is determined that the difference between actual performance and target warrants an operator intervention a control response is being activated. In multivariate quality control these steps require special consideration. We elaborate on these special issues by commenting on Jurans 7 steps in the context of multivariate quality control. We label the steps S147. A flowchart of these stepsis presented in Figure12.1.
S1.Determination of the control subjects

to be used for quality control. In general, a production process has many sources or causes of variation. Complex interactions between material, tool, machine, work methods, operators, and the environment combine to create variability in the process. Factors that are permanent, as a natural part of the are called commoncauses of variprocess, cause chronic problems and ation. The combined effect of common causes can be described using probability distributions. The only way to reduce the negative effects of chronic, common causes of variability, is to modify the process. Special causes, assignable causes or sporadic spikes arise from external temporary sources that are not inherent to the process. In order to signal the occurrence of special causes we need a central mechanism. The statistical approach to process control allows us to distinguish between chronic problems andsporadic spikes.
138
Chapter 12
S2
Determine Variables Determine Process to be Measured
(Re)Establish Goals Collection orand Targets
Mecbism
Evaluate Performance
Interpret Difference between Actual Performance and Goals
S7 Take Actionon the Difference

Flgure 12.1:Establishing and maintaining quality control.
In the univariate case there is a clear correspondence betweenthe control subject and the monitored variable used for control. Each control subject is being measured thus producing the actual performance data on the monitored variable. In multivariate quality control such a correspondence is not as obvious. If we decide to monitor variables in p dimensions, say, not all measurements must come from the same control subjects. This is particularly relevant when measurements are carried out at several locations. For example some tests mightbe performed in a remote special laboratory while others are easy to perform
139
on site and produce readily available results. In such circumstances the observation vector can consist of data collected at different times, by different measurement devices and on different control subjects. Choosing the control subjects for multivariate quality control therefore has to take into consideration the needs of the process being monitoredas well as the measurement system and logistic capabilities of the process environment.
S2. Determination of the variables to be measured for quality control. The process capability study described in Chapter 4 produces an understanding of the structure of the data collected on the process being studied. P a r t of this understanding leads to an identification of key variables to be used in the quality control feedback loop.
S3. Establishment of a goal for the control subject through internally or externally assigned targets for the measured variables. Chapters 3-5cover three alternatives to setting targets for the monitored variables.If a previously collected sampleis used as a reference sample then quality control determinesif the observation being tested is consistent with the process that generated the reference sample. Internally computed targets are used in the context of process capability studies for signalling unusual, outlying observations. Externally assigned targetsare derived f r o m considerations outsidethe process such as standards or competitive benchmarks. Here quality control is designed to establish adherence to such standards orto meet goals set by other competitive processes.
S4. Creation of a sensor or data collection mechanism to evaluate actual performance. Referring backto the comments onstep 1 it is worth emphasizing the complex issues of collecting multivariate data in industrial environments. The data collection mechanism has to include the logistics of getting the control subjectsto the measurement devices, scheduling the tests and providing the results back to the operator in charge of controlling the process.
140
Chapter 12
S 5 Evaluation of actual performance.

Multivariate quality control involves, in come cases; a complex scheduling of tests so that results can be synchronized to arriveat the right place, at the right time. Evaluation of actual performance is dependent on such a synchronized system. Univariate quality control is much simpler since once a variable has been measured and the results have been analyzed the control operationis possible.
S6. Interpretation of the difference between actual performance and goal. Once a T 2chart indicates a condition out of control an investigation using either univariate control or simultaneous univariate and multivariate charts, is triggered. This is so because theT 2chart does not pointout the actual cause of discrepancy between actual performance and goal. Chapter 11 discusses several graphical techniquesthat can be used in such investigations. S7. Taking action on the difference using prespecitled criteria and responses. The multivariate case entails a much more elaborate set of criteria and responses than the univariate quality control process.A good unis necessary for the quality derstanding of cause and effect relationships control response to deviation of actual performance from targets to be well founded and systematic. Results from designed experiments mentioned inthe process capability study section of chapter 4 can establish such cause and effect relationships.The possible responses have to be determined in the context of the organization structure of the process being monitored. This should include a description of the levelof operator enpowerment and what are the recommended actions to be initiated as a result of out-of-control signals. Milestones of a Strategic Plan for Implementing Multivariate Quality Control
In order to implement the 7 step multivariate quality controlprocess described above we propose aplan including 6 milestones listed as M1M6. Specific organizations canuse this plan by adding a time table and assign individuals responsible to meet these milestones.
141
M1. Perform a multivariate process capability study and set targets using a process capability study team. M2. Design the data collection and feedback system to meet the needs of the process operators and their management. M3. Develop criteria for action on the differences between actual performanceand goal by specifying who is entitled to take what action under what conditions. M4. Establish a review system to monitor the quality control process using inputs from the process internal customers and process audits. M5. Pilot the quality control process and provide training to operators and management. M6. Address resistanceto change by planning the transfer to on-going operations ofthe quality control process.
This plan willenable an organization to implement a multivariate to steps 1-7 listed above. quality control process that operates according A sample plan is presented as Figure 12.2. We conclude with a highlight of main points presented in the text that are specific to multivariate quality control: 1. Hotelling's T 2is a basictool for multivariate quality control. With grouped data it can be decomposed into an overall measure of distance of the group means from the target, T i , and an internal measure of variability, T i (Chapters 1-5). 2. The analysis of Tz-charts needs to be complemented with addi6 and 11). tional graphical displays (Chapters 3. The detection of the variables responsible for a multivariate out of control signalcan be performed eitherby graphical univariate methods or by multivariate decompositionof the (Chapters 7 and 11). 4 . A large number of variables (dimensions) relative tothe number of observations will produce stable Tz-charts (Chapter 4 ) . 5. We distinguish between multivariate quality control with targets from a previously collected references sample (Chapter 5 ) from
7 ' '
142
Chapter 12
M6
and Training Transfer to Operation
Ongoing Operations
Figure 12.2: A sample plan for institutionalizing quality control.
internally assigned targets (Chapter 4) or from externally assigned targets (Chapter3). 6. Statistical tolerance regions offer an alternative to the Shewhart control chart approach to quality control. Statistical tolerance regions are particularly useful in the multivariate case (Chapter 8). 7. Curtailing inspection of samples from batches can carry many advantages in multivariate quality control. Data collected from -.batches also allows for a comparisonof variability withinand between batches with the 11and ZZ indices (Chapter9).
lmplementlng Multivariate Control Quality
143
8. Principal components can be useful for multivariate quality control, especially when they leadnew to dimensions that can be naturally interpreted (Chapter10). 9. Graphical techniques available through modern computer technology provide efficientand effective environmentsfor multivariate quality control (Chapter11). 10. Implementing multivariate quality control requires a plan. Chapter 12 offers some basic elements of such a plan including milestones that have proven essential in successful implementations.
Appendix I MINITAB T M Macros f i r Multivariate Quality ControL
APPENDIX 1.1: POOLED.MTB. note this is the macro pooled.mtb for grouped data note the number of variables in K2; note group size in K3; note the data are originally in cl-ck2; note from the original data set note the first 100 observations are copied into note c101-ck110. note number of observations in data set is in K10 note number of groups in data set is in K11 let k2=4 let k3=2 let kllO=lOO+k2 let k130=120+k2 let k6 10=6OO+k2 let k3 1=l let k32=k31+k3-1 COPY cl-ck2 c101-ck110; use 1:100.
145
146
Appendix 1
count cl01 k10 let k l l=klO/k3 set c600 WO) end let k601=601 store let ck601=c600 let k601=k601+1 end exec k2 copy c601-ck610 m2 store COPY c101-ck110 ~121-ck130; use k31:k32. covariance c121-ck130m1 print m1 add m1m2 m2 let k31=k31+k3 let k32=k31+k3-1 end exec kl1 let k12=l/kl1 mult m2 k12 m3 print m3
APPENDIX 1.2: Macro for Analysis of Ungrouped Multivariate Data UU.MTB The Main Macro note this is the W.Mtb program for ungrouped data note Tsqbtugr.MtbperformsT-squares on UNGROUPED base and tested sample note data in cl-c9 note base sample in c41-c59 note tested sample in c21-c39 note overall means of base in m50 note the number of variables in K2
MINITABTM Macros for Multivariate Quality Control
147
note note ENTER HERE NUMBER OF VARIABLES (INTO K2): let k2=6 note ENTERHEREROW NUMBER FOR FIRST OBS. I N BASE note (INTO K97 1): let k971=1 note ENTER HERE ROW NUMBER FOR LAST OBS.IN BASE note ( I N T O K972): let k972=30 note ENTERHERE ROW NUMBER FOR FIRST OBS. INTESTED note ( I N T OK973): let k973=31 note ENTERHERE ROW NUMBER FOR LASTOBS. IN TESTED note (INTO K974): let k974=705 let k7 10=700+k2 note ENTER HERE ZEROSINTO C700 I F THE BASE GRAND note MEANS ARE note TO BE COMPUTED FROM SAMPLE note OTHERWISE - ENTER THE EXTERNAL VALUES: set c700 WO) end exec 'initmmtb' 1 store copy c700 ck706; use k806. let c717=abs(ck706) let c718=sum(c717) let k706=k706+1 let k806=k806+1 end exec k2 let c719=(c718 gt 0) copy c719 k719 print k7 19 COPY Cl-cklO ~41-ck30;
148
Appendix 1
use k971:k972. count c41 k4 set c80 1:k4 end COPY Cl-cklO ~21-ck20; use k973:k974. count c21 k500 set c60 1 :k500 end let k706=706 note c61-ck40 and m50 has the base mean x-bar-bar Store mean ck21 ck31 let ck31=(1-k719)*ck31+k719*ck706 std ck21 ck61 let ck71=(ck21-ck31)/ck61 let k21=k21+1 let k31=k31+1 let k61=k61+1 let k71=k7l+l let k706=k706+1 end exec k2 copy c61-ck40m50 print c61-ck40 note now testsfor base exec 'tsbsugl .mtb' 1 let c180=c160 erase c 160 let c181=(~00-2)*c180)/((k500-1)-~500/(k500-1))*~180) note now testsfor Tested let kl1=21 let k31=61 let k61=121 let k81=161
MINITABTMMacros for Multivariate Quality Control
149
store let ck81=(ckl l-ck31)/ck61 let kll=kl1+1 let k3 1=k31+l let k61=k61+1 let k81=k8l+l end exec k2 exec 'tstsugl .mtb' 1 let c183=c160 erase c160 note now performs T-sq with Leave-one-out from c21 -ck30 note means and covariance with "left-out" with data in cl0l-ckl10 note the T-square results inc93 let kll=k4 exec 'tsqugloo.mtb' k4 let c 182=cl60 note: The X-bar-bar, the S matrix, and the S (-1)- matrix for Base print c61-ck40 print m18 m15 note: Univariate t-statistics for each variable and observation in Base print cl41-ck80 note: Multivariate'Q?M ,V ' M and Tj"M (LOO) for each observation note: in Base print c180-cl82 note: Univariate t-statistics for each variable and observation in Tested print c161-ck90 note: Multivariate VM for each observation in Tested print c60 c l 83 exec 'uclbtung.mtb' 1
INITUU.MTB
note this is the inituu.Mtb programfor ungrouped data let k991=10 let k993=1 store let kk991=(k993>12)*20+(k993-1)*20+k2
150
Appendix 1
let k991=k991+10 let k993=k993+1 end exec 21 let k992=21 let k993=2 store let kk992=k993*20+1 let k992=k992+10 let k993=k993+1 end exec 14 let k701=701 let k32=0 let k706=706 let k806=1
MSCOVUG.MTB note Mscovug.Mtb is a routine of TsbsugLMtb, Tstsugl.Mtb note and of Tsquglool .Mtb in UU.Mtb note count c201 k300 let k101=201 let kl11=221 let k121=241 let k110=20O+k2 let k120=220+k2 let k130=240+k2 let k140=280+k2 let k141=281 store mean cklOl ckl11 stack ckl11 ck141 ck141 let ckl21=ckl0l-ckl11 let k101=k101+1 let kl1 l=kl11+1 let k121=k121+1
151
let k141=k141+1 end exec k2 copy c221-ck120 m17 copy c241-ck130 m11 trans m11 m12 mult m12 m11 m13 let k299=l/(k300-1) mult m13 k299 m18 invert m18 m15
TSBSUG.MTB note This is Tsbsug.Mtb, a routine of UU.Mtb note COPY ~41-ck30 ~2Ol-ckllO exec 'mscovug.mtb' l let kll=k4 COPY ~41-ck30 ~20l-ck110 exec 'tsql.mtb' k4 TSQl.MTB note Tsql.Mtb is a routine of Tsbsug1.Mtb in UU.Mtb note COPY ~201-ck110 ~421-ck210; use c80=kl1, let k51=101 let k211421 let kl11=221 store let cWl=ck21 l-ckl11 let k51=Wl+l let k211=k211+1 let kl1 l=kl11+1 end exec k2 101-ck60 m23 copy c trans m23 m24
152
Appendix 1
mult m23 m15 m25 mult m25 m24 m26 copy m26cl00 let k l l = k l l - l stack cl00 c160 c160
TSQL001.MTB note Tsqloo1.Mtb is a routine of Tsqugloo.Mtb in UU.Mtb note let k51=101 let k21l 4 2 1 let kl11=221 store let cWl=ck211-cklll let k51=k51+1 let k211=k211+1 let kl1 l=kl11+1 end exec k2 copy c101-ck60 m23 trans m23 m24 mult m23 m15 m25 mult m25 m24 m26 copy m26 100 c let k l l = k l l - l stack cl00 c160 c160 TSQTSLMTB note Tsqts1.Mtb is a routine of Tstsug1.Mtb in UU.Mtb note COPY ~201-ck110 ~421-ck210; use c60=kl1. let k51=101 let k211421 let kl11=221 store let cWl=ck21 l-ckl11

let k51=W1+1 let k211=k211+1 let kl1 l=kl11+1 end exec k2 copy c101-ck60 m23 trans m23 m24 mult m23 m15 m25 mult m25 m24 m26 copy m26 100 c let kl l=kll-1 stack cl00 c160 c160
153
TSQUGLOO.MTB note Tsqugloo.Mtb is a routine of UU,Mtb note COPY ~41-ck30 ~201-~kllO; omit c 8 k k l I . exec 'mscovug.mtb' 1 mult m15 1 m33 COPY ~41-ck30 ~421-ck210; use c80=kl1. exec 'tsqlool.mtb' 1 TSTSUG1.MTB note Tstsugl .Mtb is a routine of UU,Mtb COPY c41-ck30 ~201-ckllO exec 'mscovug.mtb' 1 let kll=k500 COPY ~21-ck20 ~201-ckllO exec 'tsqtsl.mtb' W O O UCLBTUNG.MTB note Uclbtung.Mtb is a routine of UU.Mtb note set c400 ,975 .g975 .g9865
154
Appendix 1
end let k 2 0 1 = ( k 4 1 ) 6 0 1; InvCDFC400 c T K201, set c 4 0 0 . 9 5. g 9 5 , 9 9 7 3 end let k 2 0 1 = ( k 4 l ) * ( k 4 l ) / k 4 l e t k202=k2/2 l e tk 2 0 3 = ( k 4 k 2 1 ) / 2 InvCDFC400 c404; B e t a K202 K203. let c 6 2 1 = c 4 0 4 * k 2 0 1 let k 2 1l=(k4-1)*(k4-2>*kU(k4*(k4-k2-1)) let k212=k2 let k 2 1 3 = ( k 4 k 2 1 ) InvCDF C400 c414; F K212 K213. l e tc 6 4 1 = c 4 1 4 * k 2 1 1 let k221=k4*(k4-2)*k2/((k4-l)*(k4-k2-1)) let k222=k2 l e tk 2 2 3 = ( k 4 k 2 ' 1 ) InvCDF C400 c444; F K222 K223. l e tc 6 6 1 = c 4 4 4 * k 2 2 1 l e t k231=(k4+l)*(k4-1)*kU(k4*(k4-k2)) l e t k232=k2 let k 2 3 3 = ( k 4 k 2 ) InvCDF C400 c464; F K232 K233. l e tc 6 81=c464*k23 1 print c 4 0 0 c 6 c621 0 1 c 6 4 1 c 6 6 1 ; format ( l x , f 9 . 4 , 4 f 1 0 . 3 ) . print c 4 0 0c681; format ( l x , f 9 . 4 , f 1 0 . 3 ) .
155
APPENDIX 1.3: MACRO FOR ANALYSIS OF GROUPED MULTIVARIATE DATA GG.MTB The Main Macro note this is the gg.mtb for grouped data note the tsqbsgrp macro performs T-squares on GROUPED base and note tested sample in c l 1-c19; note data in cl-c9 base sample in c21-c29; tested sample note overall meansof base in m50 note the number of variables in K2 ; group size in K3; # of observ in tested is in note number of observ in base is in K10; note K20; is in is in K11; # of groups in tested note number of groups in base note K21 ; note ENTER HERE NUMBER OF VARIABLES (INTO K2): let k2=6 note ENTER HERE GROUP SIZE (INTO K3): let k3=2 note ENTER HERE ROW NUMBER FORFIRST OBS.IN BASE note (INTO K971): let k971=1 I N BASE note ENTER HERE ROW NUMBER FOR LAST OBS. note (INTO K972): let k972=30 note ENTER HEREROW NUMBER FORFIRST OBS.IN TESTED note (INTO K973): let k973=3 1 note ENTER HERE ROW NUMBER FORLAST OBS.I N TESTED note (INTO K974): let k974=70 let k9=0+k2 let k110=100+k2 let k210=200+k2 COPY Cl-Ck9 clOl-ck110; use k971:k972. count cl01 k10 COPY C 1-ck9 c20 1-ck2 10;
l56
Appendix 1
use k973:k974. count c201k20 exec 'initgg.mtb' 1 Store mean cklOl ck131 let ckl51=ckl0l-ckl31 let ck251=ck201-ck131 let k101=k101+1 let k201=k201+1 let k131=k131+1 let k151=k151+1 let k251=k251+1 end exec k2 note c131-ck140has the overall mean x-bar-bar and m50 the S(-l) copy c151-ck160 m11 trans m1 1 m12 mult m12 m1 1 m13 let k299=l/(k10-1) mult m13 k299 m18 invert m18 m50 let kll=klO/k3 let k21 =k2Ok3 exec 'groupbs.mtb' 1 erase c401 -c499 erase c80 of (n-l)*S(i) 'S and m52 has Sp(-l) note m51 has the sum count c201 k20 let k21=k20/k3 set c12 1:klO end set cl 3 1:kll end set c22 1:k20
MINITABTM Macrosfor Multivariate Quality Control
157
end set c23 1x21 end exec 'groupts.mtb' 1 erase c401-c499 erase c80 exec 'tsqbsgrp.mtb' 1 let ~506=((kl1-2)*~505)/(kl1-1-c505*kl l/(kll-1)) erase c401-c499 erase c80 exec 'tsqtsgrp.mtb' 1 erase c401-c499 erase c80 note now performs T-sq with Leave-one-out from cl0l-ckl10 note means and covariance with "left-out" with data in cl0l-ckl10 note the T-square results in c507 let k406=k10 let k405=klO-k3+1 let k407=kl l exec 'tsqgrloo.mtb' kl1 let c507=k3*c80 exec 'uclsbtgr.mtb' 1 note c200 alphas-c301 c311c321 c331 UCL's with S matrix note (beta+two F's for Base+one for Tested) print c200 c301 c311 c321 c331; format (lx,f9.4,4f10.3). exec 'uclspbtg.mtb' 1 note c200 alphas- Ucls's- c311for Base, 321 for tested withSp - matrix print c200 c311 c321; format (lx,f9.4,2f10.3). let k506=k10 let k505=k506-k3+1 store copy c501 c502 c511 c512 c901 c902 c911 c912; use k505:k506. sum c901 k903
158
Appendix 1
sum c902 kg04 sum c91 1 k913 sum c912 k914 stack k903c503 c503 stack k904 c504 c504 stack k913 c513 c513 stack k914 c514 c514 let k505=k505-k3 let k506=k506-k3 end exec kl1 let k506=k20 let k505=kS06-k3+1 store copy c601 c602 c611 c612 c901 c902 c911 c912; use k505:W06. sum c901 k903 sum c902 k904 sum c91 1 k913 sum c912 k914 stack k903c603 c603 stack k904 c604 c604 stack k913 c613 c613 stack k914 c614 c614 let k505=k505-k3 let k506=k506-k3 end exec k21 note: X-bar-bar, the S-matrix, S(-1)-matrix, sum of(n-l)S(i) & Sp(-l) for Base print c131-ck140 print m18 m50 m51 m52 note For Base with S:T20(groups)in c503 T2D(groups)in c504 T2M (groups)in c505 note F M (groups) in c506 T2M (groups - LOO) in c507 print c503-c507 note For Base with Sp:T20(groups)in c513 T2D(groups)in c514 TzM
MINITABTMMacros for Multivariate Quality
Control
159
note (groups)in c515 print c513415 note For Tested with S:T20(groups)in c603 T2D(groups)in c604 T2M note (groups)in c605 print c603-c605 let c617=c614/k3 note ForTested with Sp:pO(groups)in c613 pD(groups)in c614 T2M note (groups)in c615 note ForTested with Sp:T2D(groups) divided obs. by per group in c617 print c613-c615 c617 note c200 alphas-c301 c311 c321 c331 UCL's with S - matrix note (beta+two F's for Base+one for Tested) print c200 c301 c311 c321 c331; format (lx,f9.4,4f10.3). InvCDF c200 c327; Chisquare k3. note c200 alphas c327 are the Chi-square values UCL asfor T2D print c200 c327; format (lx,f9.4,f10.3).
1NITGG.MTB note thisis the 1nitgg.mtb routine for grouped data let k991=120 let k993=1 store let kk991=k991-10+k2 let k991=k991+10 let k993=k993+1 end exec 18 let k410=400+k2 let k992=101 let k993=1 store let Mt992=k992 let k992=k992+10 let k993=k993+1
160
Appendix 1
end exec 20
GMEANSTS.MTB
~701-~k710 ~401-~k410; use k405:k406. let k407=k406-k405+1 let k100=110 store let k999=klOO+lO let k998=k100-9 let kk999=k999-0+k2 1 let kk998=k998 let k100=k100+10 end exec 88 let k521=521 let k53 1=53 1 let k54 1 =54 1 let k551=551 let k561=561 let k530=521+k2 let k540=531+k2 let k550=541+k2 let k560=55 1 +k2 let k570=561+k2 store mean ck401 ck411 std ck401 c528 stack ck411 ck441 ck441 stack c528 ck521 ck521 let ck421=ck401-ck411 let ck461=ck411-ck131 let ck551=sqrt(k407)*(ck41 l-ckl31)/ck541 stack ck421 ck45 1 ck45 1 stack ck461 ck471 ck471 stack ck551 ck561 ck561
COPY
161
let k401=k401+1 let k421=k421+1 let k131=k131+1 let k441=k441+1 let k451=k451+1 let k461=k461+1 let k471=k471+1 let k521=k521+1 let k531=k531+1 let k541=k541+1 let k551=k551+1 let k561=k561+1 end exec k2 let k405=k405-k3 let k406=k406-k3 let k521=521 let k531=531 let k541=541 let k551=551 let k561=561 let k57 1=57 l let k581=581 let k591=591
GROUPBS.MTB let k403=kl1 let k406=k10 let k405=k406-k3+1 let k401401 let k411=411 let k421421 let k131=131 let k441=441 let k45 1 4 51 let k461461 let k541=541
162
Appendix 1
let k410=400+k2 let k420=410+k2 let k430=420+k2 let k440=430+k2 let k450=440+k2 let k460=450+k2 let k470=460+k2 let k480=470+k2 let k710=700+k2 let k610=600+k2 let k550=540+k2 set c600 end let k601=601 store let ck601=c600 let k601=k601+1 end exec k2 copy c601-ck610 m14 erase c601 -ck610 COPY c101-ck110 ~701-~k710 exec 'mscov.mtb' k403 invert m18 m15 COPY ~441-ck450 c l 11-ck120 COPY ~451-ck460 ~161-ck170 COPY ~471-ck480 ~17l-ck180 copy m14 m51 diag m18 c558 let c559=sqrt(c558) copy c559 m60 trans m60 m61 copy m61 c541-ck550 copy m15 m52 copy m13 m53
163
GROUPTS.MTB let k403=k21 let k406=k20 let k405=k406-k3+1 let k100=110 store let k999=k100+10 let k998=k100-9 let kk999=k999-10+k2 let kk998=k998 let k100=k100+10 end exec 88 COPY c201-ck210 c701-ck710 exec 'gmeansts.mtb'k403 let k521=521 let k531=531 let k571=571 let k581=581 let W91=591 store mean ck531 ck581 std ck531 ck591 let ck571=(ck521-ck5 let k521=k521+1 let k531=k531+1 let k571=k571+1 let k581=k581+1 let k591=k591+1 end exec k2 COPY ~441-ck450 c21 1-ck220 COPY ~45l-ck460 ~261-ck270 COPY c471 -ck480 1c27 -ck280 let W31=531 let k571=571
164
Appendix 1
let k581=581 let k591=591
MSCOV.MTB ~701-~k710 ~401-~k410; use k405:k406. count c401 k400 let k401=401 let k411411 let k421421 let k131=131 let k441=441 let k45 1=45 1 let k461461 let k47 1 4 71 let k531=531 store mean ck401 ck411 std ck401 c529 stack ck411 ck441 ck441 stack c529 ck531 ck531 let ck421=ck401-ck411 let ck461=ck411-ck131 stack ck421 ck45 1 ck45 l stack ck461 ck471 ck471 let k401=k401+1 let k411=k411+1 let k421=k421+1 let k131=k131+1 let k441=k441+1 let k451=k451+1 let k461=k461+1 let k471=k471+1 let k531=k531+1 end exec k2 let k531=531
COPY
165
copy c421-ck430 m11 trans m11 m12 mult m12 m11 m13 add m13 m14 m14 let k299=1/(k403*(k400-1)) mult m14 k299 m18 let k405=k405-k3 let k406=k406-k3
TSQBSGRP.MTB erase c401 -c499 erase c80 COPY ~151-ck160 ~401-~k410 copy m50 m60 let k600=k10 let c70=c12 exec 'tsqnew.mtb' k10 let c501=c80 erase c401-c499 erase c80 let k600=k10 let c70=c12 COPY ~161-ck170 ~401-~k410 copy m50 m60 let k600=k10 let ~ 7 0 x 1 2 exec 'tsqnew.mtb' k10 let c502=c80 erase c401-c499 erase c80 COPY ~171-ck180 ~401-~k410 copy m50 m60 let k600=kl1 let c70=c13 exec 'tsqnew-mtb' kl1 let c505=k3*c80 erase c401 -c499
166
Appendix 1
erase c80 COPY ~151-ck160 ~401-~k410 copy m52 m60 let k600=k10 let c70=c12 exec 'tsqnew.mtb' k10 let c51 1=c80 erase c401-c499 erase c80 COPY ~16l-ck170 ~401-ck410 copy m52 m60 let k600=k10 let c70=c12 exec 'tsqnew.mtb' k10 let c5 12=c80 erase c401-c499 erase c80 COPY C 17 -&l 1 80 c401-C, 10 copy m52 m60 let k600=kl1 let c70=c13 exec 'tsqnew.mtb' kl1 let c515=k3*c80 erase c401 -c499 erase c80
TSQBSNEW.MTB copy c4 1 1 -ck420 m23 trans m23 m24 mult m23 m60 m25 mult m25m24 m26 copy m26 c50 TSQGRLOO.MTB let k801=801 let k811=811 let k821=821
167
let k831=831 let k841=841 let k851=851 let k810=800+k2 let k820=810+k2 let k420=410+k2 let k830=820+k2 let k840=830+k2 let k850=840+k2 let k860=850+k2 COPY c101-ck110~801-ck810; use c12=k405:k406. count c801 k400 store mean ck8Ol ck811 let k801=k801+1 let k811=k8ll+l end exec k2 COPY clOl-ckl10~821-ck830; omit c12=k405:k406. let k801=801 let k811=811 let k821=821 let k831=831 let k841=841 let k85 1=85 1 store mean ck821 ck831 let ck841=ck821-ck831 let ck851=ck81 1 -ck831 let k811=k811+1 let k821=k821+1 let k831=k831+1 let k841=k841+1 let k851=k851+1 end
168
Appendix 1
exec k2 copy c841-ck850m1 1 count c841 k400 trans m11 m12 mult m12 m1 1 m13 let k299=l/(k400-1) mult m13 k299 m18 invert m18 m15 COPY ~851-ck860 ~41l-ck420 copy m15m60 exec 'tsqbsnew.mtb' 1 stack c50 c80 c80 let k405=k405-k3 let k406=k406-k3 let k407=k407-1
TSQNEW.MTB ~401-~k410 c411-ck420; use c70=k600. copy c411-ck420 m23 trans m23 m24 mult m23 m60 m25 mult m25m24 m26 copy m26c50 let k600=k600-1 stack c50 c80 c80
COPY
UCLSBTGR.MTB set c200 .95 ,995 .g973 end let k701=(kl l-l)*(kl l*k3-l)/(kll) let k702=k2/2 let k703=(kl1*k3-k2- 1)/2 InvCDF C200 c704; Beta K702 K703.
169
let c301=c704*k701 let k71 l=(kl l-l)*(kl l*k3-2)*kU(kl l*(kl l*k3-k2-1)) let k7 12=k2 let k713=(kll*k3-k2-1)
InvCDF C200 c7 14; F K712 K713. let c311=c714*k711 let k721=kll*(kll*k3-2)*k2/((kl1-l)*(k11*k3-k2-1)) let k722=k2 let k723=(kl1 *k3-k2-1) InvCDF C200 c724; F K722 K723. let c321=c724*k721 let k73l=(kll+l)*(kll*k3-l)*k2/(kll*(kll*k3-k2)) let k732=k2 let k733=(kll*k3-k2) InvCDF C200 c734; F K732 K733. let c331=c734*k731
UCLSPBTG.MTB set c200 .95 .995 .9973 end let k711=(kll-l)*kll*(k3-l)*k2/(kll*(kll*(k3-1)-k2+1)) let k7 12=k2 let k713=(kll*(k3-1)-k2+1) InvCDF C200 c7 14; F K712 K713. let c311=c714*k711 let k721=(kll+l)*kl l*(k3-l)*k2/(kl l*(kl l*(k3-1)-k2+1)) let k722=k2 let k723=(kl l*(k3-1)-k2+1) InvCDF C200 c724; F K722 K723. let c321=c724*k721
170
Appendix 1
APPENDIX 1.4: MACRO FOR CREATING MP-CHART FOR UNGROUPED DATA MPGRAPH.MTB THE MAIN MACRO note This is the main MpGraph program note Base calculations are in file UU.Mtb note erase c245-c250 erase c801-c806 l e t c801=(~81>4)*4+(~81<-4)*(-4)+(~81 ge -4and c81 le 4)*c81 let c802=(~82>4)*4+(~82<-4)*(-4)+(~82 ge -4and c82 l e 4)*c82 l e t c803=(~83>4)*4+(~83<-4)*(-4)+(~83 ge -4and c83 l e 4)*c83 l e t c804=(~84>4)*4+(~84<-4)*(-4)+(~84 ge -4and c84 l e 4)*c84 l e t c805=(~85>4)*4+(~85<-4)*(-4)+(~85 ge -4and c85 le 4)*c85 l e t c806=(~86>4)*4+(~86<-4)*(-4)+(~86 ge -4and c86 l e 4)*c86 l e t c903=(~93>~341(3))*(~341(3)+4)+(~93 le c341(3))*c93 count c801 k20 exec 'mpdatl .mtb' k20 l e t c248=c247+c245 let c25O=(abs(c247)>3)* l+(abs(c247) 3 le and abs(c247)>2)*4+(abs(c247) l e 2 and abs(c247)>1)*13 l e t k%4=c341( 1) l e t k245=c341(2) l e t k246=c341(3) chart sum(c248)*c246; cluster c249; bar; base c245; type 1 color c250; tick 2 k244 k245 k246; TSize 1; W o n t 1; label 'T290''T295' 'T29973'; tick 1; T S i z e 1; GRID 2; Case Study 1'; t i t l e 'MP- Chart for Variables:
171
Wont 1; axis 2; Wont 1; label 'Univariate and Multivariate Statistics'; axis 1; Wont 1; label 'Groups'. end
MPDATLMTB note MpDatl is a routineof MpGraph.Mtb note let k l l = 6 exec 'mpdat.mtb' kl1 let k20=k20-1 end MPDAT.MTB note MpDatis a loop routine of MpDatl .Mtb in MpGraph.Mtb note let k13=kl1+800 copy ck13 c7; use k20. stack c7 c247 c247 copy c903 c7; use k20. stack c7 c245 c245 copy k l 1 c7 stack c7 c249 c249 copy k20 c7 stack c7 c246 c246 let k l l = k l l - l
172
Appendix 1
APPENDIX 1.5: MACRO FOR CREATING MP-CHART FOR GROUPEDDATA MPGROUP.MTB-THE MAIN MACRO FOR MP CHART FOR MEANS note Thisis the main MpGraph program note It displays the MpGraph for 6 variables note Datais in file Case1 sm6.Mtw in the Data folder are in file GG.Mtb note Base calculations note (Calculated data is in file Caselsm6-Tmp.Mtw) note note ENTER HERE NUMBER OF VARIABLES (INTOK2): let k2=6 note ENTER HERE NUMBER OF FIRST COLUMN IN THE SERIES note OF K2 COLUMNS OF DATA ( I N T O K973): let k973=561 note ENTER HERE NUMBER OF COLUMN CONTAINING THE T2 note VALUES (INTOK974): let k974=605 note ENTER HERE NUMBER OF COLUMN CONTAININGTHE 3 note CRITICALVALUES IN ASCENDING ORDER (INTOK975): LET K975=321 note ENTER HERE NUMBER OF FIRST EMPTY COLUMN IN A note SERIESOF K2 EMPTY COLUMNS ( I N T O K972): let k972=801 OF FIRST EMPTY COLUMN IN A note ENTER HERE NUMBER note SERIESOF 6 EMPTY COLUMNS (INTO K991): let k991=245 note ENTER HERE NUMBER OF FIRST EMPTY COLUMN IN A note SERIES OF 2 EMPTY COLUMNS (INTO K976): let k976=903 note let k977=k976+1 let k992=k991+1 let k993=k991+2 let k994=k991+3 let k995=k991+4 let k996=k991+5
MiNITABTM Macros for Multivariate
Quallty Control
173
let k982=k972+k2-1 let k983=k973+k2-1 erase ck991 -ck996 erase ck972-ck982 store let ck972=(ck973>4)*4+(ck973~-4)*(-4)+(ck973 ge -4and ck973 le 4)*ck973 let k972=k972+1 let k973=k973+1 end exec k2 let k972=k972-k2 let k973=k973-k2 letck976=(ck974>ck975(3))*(ck975(3)+4)+(ck974 leck975(3))*ck974 count ck972 k20 exec 'mpdat1 . m t b 'k20 let ck994=ck993+ck991 let ck996=(abs(ck993)>3)*l+(abs(ck993) le 3 and abs(ck993)>2)*4+(abs(ck993) le 2 and abs(ck993)>1)*13 let k244=ck975(1) let k245=ck975(2) let k246=ck975(3) chart sum(ck994)*ck992; cluster ck995; bar; base ck991 ; type 1; color ck996; t i c k 2 k244 k245 k246; TSize 1; Wont 1; label '"95' 'T2995' T9973'; t i c k 1; TSize 1; GRID 2; axis 2; Wont 1;
174
Appendix 1
label 'Univariate and Multivariate Statistics '; axis 1; TFont 1; label 'Groups'. end
MPGDISP.MTB-THE MAIN MACRO FOR MP CHART FOR DISPERSION note This is the main MpGrDisp program note It displays the MpGraph for Dispersion k2 variables note Base calculations are in file GG.Mtb note OF VARIABLES (INTO K2): note ENTER HERE NUMBER let k2=6 note ENTER HERE NUMBER OF FIRST COLUMN IN THE SERIES note OF K2 COLUMNS OF DATA (INTO K973): let k973=571 note ENTER HERE NUMBER OF COLUMN CONTAINING note THE (T2DK3) VALUES (INTO K974): let k974=617 note ENTER HERE NUMBER OF COLUMN CONTAINING note THE 3 CRITICAL VALUES IN ASCENDING ORDER note (INTO K975): LET K975=327 note ENTER HERE NUMBER OF FIRST EMPTY COLUMN IN note A SERIES note OF K2 EMPTY COLUMNS (INTO K972): let k972=801 note ENTER HERE NUMBER OF FIRST EMPTY COLUMN IN note A SERIES K991): note OF 6 EMPTY COLUMNS (INTO let k991=245 note ENTER HERE NUMBER OF FIRST EMPTY COLUMN IN note A SERIES K976): note OF 2 EMPTY COLUMNS (INTO let k976=903
175
note let k977=k976+1 let k992=k991+1 let k993=k991+2 let k994=k991+3 let k995=k991+4 let k996=k991+5 let k982=k972+k2-1 let k983=k973+k2-1 erase ck991 -ck996 erase ck972-ck982 store let ck972=(ck973>4)*4+(ck973<-4)*(-4)+(ck973 ge -4and ck973 le 4)*ck973 let k972=k972+1 let k973=k973+1 end exec k2 let k972=k972-k2 let k973=k973-k2 let ck976=(ck974>ck975(3))*(ck975(3)+4)+(ck974 leck975(3))*ck974 count ck972 k20 exec mpdatl.mtb k20 let ck994=ck993+ck991 let ck996=(abs(ck993)>3)*l+(abs(ck993) le 3 and abs(ck993)>2)*4+(abs(ck993) le 2 and abs(ck993)>1)*13 let kW=ck975( 1) let k245=ck975(2) let k246=ck975(3) chart sum(ck994)*ck992; cluster ck995;
bar; base ck991 ;

type 1; color ck996;
176
Appendix 1
tick 2k244 k245 k246; TSize 1; "Font 1; label 'Chi295' 'Chi2995' 'Chi29973'; tick 1; TSize 1; GRID 2; note title 'FIGURE 11.4b'; note "Font 1; note title 'Case Study 1'; note "Font 1; axis 2; "Font 1; label 'Univariate and Multivariate Statistics '; axis 1; "Font 1; label 'Groups'. end
MPDAT1G.MTB note MpDatl is a routineof MpGraph.Mtb note let k l l = k2 exec 'mpdat.mtb' k2 let k20=k20- 1 end MPDATG.MTB note MpDat is a loop routine of MpDatl.Mtw in MpGraph-Mtw note let k13=kll+k972-1 copy ck13 ck977; use k20. stack ck977 ck993 ck993 copy ck976 ck977; use k20. stack ck977 ck991 ck991
177
copy kl1 ck977 stack ck977 ck995 ck995 copy k20 ck977 stack ck977 ck992 ck992 let kl l=kl1-1
Appendix 2 The Dataporn the Case Stadies
179
The DataCase from the
Studies
181
CASE STUDY l. The data in Case Study 1 originate from a process capability study for turning aluminium pins. Six diameter and length measurements were recorded for each of the 70 observations. The first three variables are diameter measurements on three different locations on the main part of the pin, the forth is a diameter measurement at the cap, andthe last two are the lengths measurements, without and with the cap, respectively.In the analyses, the first 30 out of the 70 observations were selectedas a base sample against which we compare all the observations.
Figure A1 :Six critical dimensionsof aluminum pins.
182
Appendix 2
TABLE A2.1
The data from Case Study 1
The Datafrom the Case Studies
183
TABLE A2.1 Continued.
184
Appendix 2
CASE STUDY 2. The data in Case Study 2 originate from another process capability study - of a quite complex mechanical part on which 11 diameter, 4 length and 2 wall width measurements were recorded. The first four variables are inner diameter measurements at four locations of a cylindrical segment of the part, and the next four are the outer diameter measurements at the same locations. Variables9 and 10 are again two diameter measurements with equal reference value (on another cylindrical segat a ment ofthe part), while the 1lth variable is a diameter measurement are of various portions different location. The four length measurements of the part, with the last among them being the overall length.The external reference values for the two wall width measurements (variables 16 and 17) had only anupper critical value (of 0.20). In various analyses we used 13 variables (by deleting variables 3 , 4and 7,8) or 11 variables (by deleting the last two width variables in addition to variables 3 , 4and 7,8).
Flgure A2: 17 variables of a complex cylindrical part.
185
TABLE A2.2
186
Appendix 2
187
188
Appendix 2
CASE STUDY 3. Raw materials used inthe manufacturing of hybrid microcircuits consist of components, dyes, pastes and ceramic substrates. The substrates plates undergo a process of printing and firing during which layers of conductors, dielectric, resistors and platinum or gold are added to the plates. Subsequent production steps consist of laser trimming, mounting and reflow soldering or chip bonding. Thelast manufacturing stage is the packaging and sealing of complete modules. Five dimensions of substrate plates are considered here, with labels (a, b, c), (W, L). The first treeare determined by the laser inscribing process. The last two are outer physical dimensions. These outer dimension are measured on a different instrument than thefirst three, using a differentsample of substrates. The two setsof measurements havein common the batch affiliation. The data in the Case Study3 are the values of the (a, b, c) dimensions of four lots of ceramic substrates. The first production lot of 13 units (labelledReference) proved tobe of extremely good quality yielding an overall smooth production with no scrap and repairs. This first lot was therefore considered a "standard" to be met by all following lots. Any evidence of lots not meeting this standard should give a signal for action.
\
["p"""
t
f
.. ._.
iI
L
I I
I
I
Ceramic substrate layout.
Flgure A3: Ceramic substrates used in the microelectronics industry with 5 critical dimensions.
189
TABLE A2.3
The data from Case Study3
a
~~
c
~
Samule ~ ~
~ ~~
_ _ _ _
200 199 197 198 199 199 200 200 200 200 198 198 199 200 201 201 201 200 192 192 199 196 199 202 202 203 202 199 201 196 191
552 550 551 552 552 550 550 551 550 550 550 550 550 550 501 550 550 551 551 553 553 554 552 553 554 552 552 552 552 553 553
550 550 551 551 551 551 551 552 552 551 551 550 551 550 550 550 501 552 552 552 550 551 552 553 551 550 551 551 552 551 552
REFERENCE REFERENCE REFERENCE REFERENCE REFERENCE REFERENCE REFERENCE REFERENCE REFERENCE REFERENCE REFERENCE REFERENCE REFERENCE SAMPLE 30 SAMPLE 30 SAMPLE 30 SAMPLE 30 SAMPLE 7 SAMPLE 7 SAMPLE 7 SAMPLE 7 SAMPLE 7 SAMPLE 7 SAMPLE 7 SAMPLE 7 SAMPLE 7 SAMPLE 12 SAMPLE 1 2 SAMPLE 12 SAMPLE 12 SAMPLE 12
4001 4001 4003 3999 3997 4002 3998 3999 3998 4000 4001 4000 4001 4002 4000 4001 4001 4005 4010 4005 4005 4005 4016 4020 4010 4000 4009 4012 4020 4016 4017
4002 4000 4000 3998 4000 4001 4000 4000 4000 4002 4000 3997 4000 3998 4001 3998 3999 4000 3985 3995 3985 3999 3995 4000 4005 4010 3993 3993 3998 4003 4002
(continued)
190
Appendix 2

a 201 201 200 203 200 201 202 201
b 551 552 552 550 551 552 551 550
C -
S ample
W
4016 4012 4011 4003 4003 4008 4008 4012
551 551 551 551 551 548 543 550
SAMPLE 12 SAMPLE 12 SAMPLE 12 SAMPLE 12 SAMPLE 12 SAMPLE 12 SAMPLE 12 SAMPLE 12
4002 3998 3995 3979 3979 3989 3987 3997
Data The
from Studies the Case
191
CASE STUDY 4. The quality control in fruit juice is usually performedby comparing the values of several constituents in tested samples against standards derived from the analysis of a base sample composed of presumably pure products. Both the issues of the selection of the base sample and that of the proper methodsof identification of possible adulteration have been debated in the professional literature. The choice of the base sample is critical to the quality control process. Previous studies have shown that the values of the attributes in fruits differ considerably between varieties.The values also vary due to environmental and climatic factors. Proposed universal base samplesthe with appropriate test procedures have been shown to reject far too many pure samples of natural juice. We do not address here the the practical issues related to selecthe - before adultertion of the base sample. Rather we assume that a priori ation, if any took place- the base sampleand the test samples come from the same population. This somewhat idealized case corresponds to a base sample created from the fruitjuice used in the industrial process of juice extraction. The data set in Case Study 4 is consistent with the assumption presented above. The data consist of the concentrations in miug per standard volume of 11amino-acids (LYS=Lysine, ARG=Arginine, ASP=Aspartic acid, SER=Serine,GLU=Glutamine acid, PRO=Proline, GLY=Glycine, ALA=Alanine, VAL=Valine, PHA=Phenyl alanine, GABA=Gammaamino butricacid) in 69 specimens of a purejuice. The process of dilution and adulteration has been simulated as follows: a base sample of predermined size 36 has been randomly selected from the data set. For the remaining second sampleof size 33 tested specimens,the measured values of the concentration variables were multiplied by a,representing a variables were adulterated by adding lOO(1- a)% of the means of those variables. This represents an ideal adulterationsince the original meansof these components were recovered (although not the original values). The statisticalmethods are applied to the resulting samples and we assess their ability to detect the adulteration. In the following data the base sampleis followed by the test sample.
192
Appendix 2
TABLE A2.4
LYS ARG ASP SER GLU PRO GLY ALA VAL PHA GABA
0.48 0.47 0.42 0.35 0.43 0.40 0.35 0.34 0.27 0.39 0.37 0.26 0.24 0.20 0.26 0.52 0.42 0.45 0.47 0.44 0.48 0.49 0.37 0.36 0.46 0.34 0.34 0.31 0.30 0.30 0.30 0.30 0.33 0.32
5.81 2.12 5.25 2.75 4.98 2.79 4.79 2.79 4.92 2.88 5.61 2.26 4.54 2.96 3.82 2.86 3.42 2.27 3.60 2,99 3.39 2.78 2.72 3.82 3.13 3.35 2.15 3.28 2.89 3.67 5.53 '2.97 5.07 3.06 5.46 3.06 5.79 2.91 2.52 2.40 5.14 2.66 4.77 2.42 4.35 3.04 4.01 2.37 4.26 2.51 3.46 2.20 4.13 2.72 3.70 2.77 3.18 2.54 3.57 2.45 3.31 2.53 3.13 2.82 3.10 3.01 3.84 3.79
4.68 4.42 3 . 8 5 3.39 3.53 3.39 3 . 8 9 3.63 4.81 5.03 5.96 6.03 5.76 5.80 6.34 3.37 4.32 4.68 4.44 4.09 4.04 5.92 5.07 3 . 9 3 7.29 3.80 6 . 0 1 5.29 5.04 5.70 5 . 2 1 5.85 7.15 6.08
0.7812.41 0.88 14.72 0.7512.13 0.8112.77 0.78 13.11 0.69 12.69 0.88 14.01 0.8615.73 0.90 8.99 0.9213.71 0.84 12.92 1 . 1 7 7 . 1 8 0.96 6.75 1.04 5.34 1.22 5.87 0.78 10.74 0.9115.37 0.84 16.52 0.8016.21 0.7212.81 0.94 16.77 1.00 15.62 0.87 15.81 0.76 11.28 1.0718.57 0.9311.73 0.9513.96 0.8510.80 0.95. 11.25 1 . 0 61 2 . 2 8 0.88 9.10 1.0010.31 1.0412.71 1 . 0 1 10,13
0.31 0.30 0.32 0.25 0.25 0.20 0.24 0.22 0.23 0.28 0.24 0.15 0 . 2 1 0.22 0.18 0.24 0.47 0.39 0.35 0.28 0.33 0.34 0.31 0.22 0.37 0.26 0.34 0.22 0.21 0.26 0.23 0.21 0.23 0.18
0.96 1 . 0 4 0.99 0.75 0.91 1 . 0 6 0.86 1.34 1.43 1.99 1.76 1.30 1.14 1 . 0 6 1.10 0.96 1.32 1.35 1 . 2 0 0.86 0.97 1 . 9 3 2.08 0.75 2.67 1.40 2.30 1.68 1.84 1.53 1 . 3 7 1.55 1.79 1.30
0.18 0.19 0.15 0.16 0.16 0.16 0.16 0.14 0.10 0.13 0.12 0.11 0 . 11 0.12 0.14 0.10 0.16 0.14 0.20 0.18 0.22 0.50 0.19 0.12 0.19 0.18 0.10 0.10 0.10 0.10 0.08 0.10 0.09 0.09
0.20 0.22 0.20 0.15 0 . 1 5 0.18 0.12 0.12 0.10 0.10 0.14 0.07 0.08 0.08 0.12 0.16 0.20 0.18 0.18 0.23 0.23 0.15 0.10 0.12 0.10 0.10 0.08 0 . 0 1 0 . 0 1 0.10 0.01 0.08 0.10 0.01
4.73 3.96 3.94 3.69 4.23 3.76 3.92 2.88 2.68 2.88 3.01 3.40 2.43 2.41 2.40 3.40 3.63 3.89 4.52 4.43 4.90 4.05 4.17 3.27 2.95 3.06 3.06 2 . 6 1 2.48 2.46 2.55 2.69 3.52 3.67
Data The
from Studies the Case
193

LYS ARG ASP SER GLU PRO GLY ALA VAL PHA GABA
0.30 3.75 2.83 6.24 0.26 3.34 3.46 7 . 0 1 0.43 5.84 2.84 3.54 0.50 4.61 2.08 5.70 0.51 6.19 3.55 4.29 0.43 5.44 2.71 4.38 0.38 5.22 2.54 3.97 0.50 5.19 3.13 4.32 0.40 4.68 2.38 3.47 0.43 4.99 2.03 3.52 0.41 5.33 2.64 4.22 0.45 5.42 2.96 4.80 0.36 4.83 2.72 3.32 0.40 4.34 1 . 9 2 4.57 0.36 4.41 2.88 3.76 0.30 4.14 2.50 5.26 . 9 1 2.32 5.14 0.38 3 0.42 3.90 2.45 5 . 2 6 0 . 3 1 3.56 2 . 6 1 5-40 0.32 4.18 3.76 5.53 0.32 3.05 3.24 6.87 0.23 3.13 3.43 6.30 0.24 2.85 3.18 4.64 0.36 4.31 2.25 3.15 0.35 4.62 2.40 2.94 0.39 4 . 5 1 2.82 4.00 0.41 4.12 2.38 5.14 0.33 3.60 2.36 5.07 0.43 4.11 2.22 6.86 0.31 3.70 2.77 5.44 0.36 3.64 2.21 6.56 0.27 3.25 2.82 4.92 0.28 2.91 3.21 6.41 0.30 3.64 2.73 5.76 0.28 2.68 3.61 6.38 1.06 6.94 0.22 1.22 0.11 0.11 2.71
0.71 6.20 1.02 6.68 0.80 11.90 0.7118.46 1.16 19.01 0.79 13.59 0.73 14.47 0.90 16.74 0.6812.01 0.63 9.84 0.8113.66 0.9115.73 0.75 12.28 0.74 11.13 0.89 14.32 0.86 15.48 0.82 14.27 0.94 18.14 0.97 12.29 0.98 10.81 1 . 4 31 3 . 0 1 1.15 10.67 0.86 6.91 0.65 11.32 0 . 7 1 1 0 . 1 8 0.8713.76 0.8311.36 0.94 13.93 1.12 14-35 1.02 12.68 1.02 15.53 0.91 8.43 1.35 9.42 0.73 5.55
0.16 0.20 0.30 0.42 0.40 0.35 0.30 0.34 0.26 0.24 0.30 0.30 0.23 0.28 0.25 0-35 0.29 0.29 0.22 0.22 0.24 0.26 0.21 0.22 0.19 0.27 0.26 0.30 0.27 0.32 0.39 0.20 0.22 0.20
1.20 1.52 0.86 1.91 1.20 1.23 0.98 1.09 0.92 0.71 0.86 1-09 0.71 1.63 0.89 2.34 1 . 8 7 2 . 0 3 1.59 1.32 1 . 8 1 1.67 1.08 0.83 0.89 0.88 1 . 7 1 1 . 6 2 1 . 6 8 1 . 7 5 1 . 9 6 1.53 1.80 0.94
0 . 0 5 0.10 0.20 0.18 0.15 0.14 0.15 0.20 0.16 0.19 0.17 0.19 0.13 0.14 0.14 0 . 2 1 0.22 0.16 0.08 0.10 0.10 0.12 0.01 0.19 0.19 0.17 0.16 0.10 0.10 0.10 0.10 0.10 0.12 0.05
0.08 0.08 0.18 0.18 0.18 0.20 0.20 0.22 0.18 0.20 0,22 0.20 0.12 0.10 0.12 0.12 0.10 0.12 0.10 0.14 0.10 0.16 0.12 0.20 0.20 0.12 0.08 0.08 0.12 0.10 0.10 0.01 0.01 0.05
3 . 0 1 2.18 3.88 6.14 4.72 4.08 4.18 4.85 3.95 4.06 4.52 3.80 3.63 3.34 3.35 3.02 3.98 3.65 2.82 2 . 91 2 . 91 2.86 2.75 3.66 3.01 3.56 3.65 3.51 3.96 3.47 3.07 2.32 2.85 3.14
Appendix 3
Review of Matrix Algesrdfor Statistics with MINITAB T M Computations
We summarize in the present appendix some properties and techare useful in statistical analysis. Basic niques of matrix operations which results are presented but not proven. The interested reader can study the subject from one of the many textbooks on matrix algebrafor statistics (such as Graybill, 1965). We use MINITAB to perform matrixcalculations. An array of n rows and1column is called an n-dimensional vector. We denote a vector with low-case bold type letter a = (a,,a2, . . ,an) where ai is the i* component of vector a. The transpose ofan ndimensional vector a is an array of 1 row and n columns. We denote the transpose of a by a'. We can add (subtract) only vectors of the same dimension. Thus a+b is a vector c whose i-th component is ci = ai +bt i = 1, , , , n. The vector 0 is one whose components are all equal to zero. Obviously, a (-a) = 0. If a and b are vectors of the same di-
mension, n, then the inner product of a and b is a'b =
C aibi. The
i=l
Euclidean norm (length) of a vector a is \la11 = (a'a)'I2. If ar is a scalar

195
196
Appendix 3
(real number) and a a vector then aa is a vector whose i-th component is aut, i = 1, . . ,n. If a1,a2, . , a k are n-dimensional vectors and
al,. ., , a k are scalars then C aiai is an n-dimensional vector.
..
. ., ,ak if there exist no scalars a], .. . ,ak such that b = C aiai k vectors al, . . ., ak
A vector b is linearly independent of vectors ai,
k
1 4
i=l
are mutually independentif no oneis a linear combination of the other (k - 1) vectors. A matrix of order n x k is an array of n rows andk columns, i.e., A = (al, 8 2 , . .. , ak). An n x n matrix is called a squared matrix. An n x k matrix A will be denoted also as
A=(aij,i=l,
...,n , j = l , ...,k).
If A andB are matrices of the same order thenA f B = C,where
cij=aijfbij, i = 1 ,
An n x n matrix A, such that
..., n ,
j=l,
..., k .
is called a diagonal matrix. The diagonal matrix In, such that its diagonal elements are all equal to 1 and all the off-diagonal elements are equal to 0 is called the identity matrix of order n. The vector l, will denote an n-dimensional vectorof ls. The matrix Jn will denotean n x n matrix whose elements are all equal to 1. Notice that l ,1 , = n. The product of A and B is defined if the number of columns of A is equal to the number of rows ofB.The product of an (n x k) matrix A by an (k x l ) matrix B is an (n x l ) matrix C such that
Review Statistics ofAlgebra Matrix for
197
Notice that two squared matricesof the same order can always be multiplied. Also Jn = n,ll',
We illustrate these operations by using MINITAB. In MINITAB,matrices are denoted by M1,M2,etc. Calculations with matricescan be found inthe Menu, under 'Calc'. Matrices can be read from a text file, or from the keyboard. For example, to read in the matrix
A=
we use the command MTB>Read33M1. DATA> 13 7 25 DATA> -4 15 5 DATA> 6
rL
-4 13
15 7 -2
2 55 1 201
- 2 20
The data is entered one row at a time. One can also copy several columns of equal size into a matrix, e.g.
MTB > COPY Cl
- C3 M5.
The transpose of an n x k matrix A, is a k x n matrix denoted by A' in which the i-th row in A is the i-th column in A'. T o obtain in MINITAB the transpose of A we write
MTB > transpose M1 M2
A' resides in M2.We can now compute B = A'A, by

MTB > multiply M2 M1 M3 . The matrix B is
198
Appendix 3
221 1 9 425 1 9 278 210 425 210 1050 Notice that the matrix B is symmetric, i.e., bfl = b,i for all i # j . Suppose that A is an n x k matrix, andk 5 n. The rank of A is the maximal numberof mutually independent column vectors of A. An n x n matrix of rank n is called nonsingular. Every nonsingular matrix A has an inverse A satisfying
1
.
AA- = A - ~ A I
To obtain the inverse of a matrix usingMINITAB we write
t 4 3i t 4 4 MTB > inverse i

The inverse ofthe matrix B above is
B =
0.03277 0.00916 -0.01510 0.00916 0.00680 -0.00507 -0.01510 -0.00507 0.00808
The Hotellings T 2 for a single multivariate observation = (X,, Xp, ,Xp)relative to a target E = (ml, mp, , mp)and with a covariance matrix S was defined as T 2 = (x - mJS (x - m J , In order to compute it, we need to determine:
...
...
a)S-m) b) (x -
a = transpose (x -J m
c) S = Inverse (S) The MINITAB commands transpose and inverse can be used to perform these operations that produce the required multivariate distance T2: Forexample, let us compute T 2 for = (23,28,35) B = (30,30,30) and 60 20 25 20 30 -151 60 -15 20
Review of Matrix Algebra for Statistics
199
1) MTB>READ 1 3 M11 DATA> 23 45 15 2) MTB>READ 1 3 M12 DATAs30 30 30 3) MTB>READ 3 3 M13 DATA> 25 20 60 DATA> 20 30-5 DATA> 60 -15 20 4) MTB>SUBTRACT M11 M12 M21 5) MTB >TRANSPOSE M21 M22
6 ) MTB>INVERSE M13 M23
7) MTB>MULTIPLY M22 M23 M24

8)' MTB>MULTIPLY M24M21 M25
ANSWER=27.6473 A squared matrix H is called orthogonal if H" = H ' . An example of orthogonal matrixis
H=
h h
-75
h' O -33 si h <

1
Use MINITAB to verify that indeed H" = H'. For every real symmetric matrix A, there existsan orthogonal matrix H such that, HAH' = A ,
where A is a diagonal matrix.The elements hii(i = 1, ,n) of A are Aii there called the eigenvalues of A. Corresponding to the eigenvalue is an eigenvector xi, satisfying the linear equations
...
200
Appendix 3
The eigenvalues and vectorsof a symmetric matrixare obtained by the command MTB > eigen A43 Cl M8
Here, the eigenvalues of M3 are stored in Cl and the eigenvectors are put in the matrix M8.The eigenvaluesof B are thus
1269.98, 255.73, 23.29

and the matrixof eigenvectors is
0.371082 0.334716 -9.866177 0.199177 -0.939762 -0.277847 0.907001 0.069367 0.415377

Notice that P is an orthogonal matrix.
References
Alt, F.(1984), Multivariate Quality Control, in the Encyclopedia of Statistical Science, eds. S. Kots, N.L. Johnson, and C.R. Read, John Wiley and Sons, Inc., New York,NY. Andrews, D.R. (1971), A Note of the Selection of DataTransformations, Biometriku, 58, pp. 249-254. ASQC (1983),Glossary and Tables for Statistical Quality Control, American Society of Quality Control, Milwaukee, Wisconsin, Becker, R., Cleveland, W., and Wilks,W. (1987), Dynamic Graphics for Data Analysis, Statistical Science, 2, pp. 355-395. Box, G.E.P. and Cox, D.R. (1964), An Analysis of Transformations, Journal of the Royal Statistical Society, B,26, pp. 21 1-252. Box, G.E.P., Hunter, W.G. and Hunter, J.S.(1978), Statistics for ExperiSons,New York, NY. menters, John Wiley and Chan, L.K., Cheng, S.W. and Spiring,EA. (1991), A Multivariate Measure of Process Capability,International Journal of Modelling and Simulation, 11, pp. 1-6. Chua, M.K., and Montgomery, DC. (1991),AMultivariateQuality Control Scheme, International Journalof Quality & Reliability Management, 6, pp. 2946. Dempster, A.P. (1963), Multivariate Theory for General StepwiseMethods, The Annals of Mathematical Statistics, 34, pp. 873-883. Eisenhart, C., Hastay, M. and Wallis, W. (1947), eds., Techniques of Statistical Analysis of Scientific and Industrial Research and Pro201
202
References
duction and Management Engineering, McGraw-Hill, New York,
NY.
Execustat 3.0@(1993),Student Edition, Duxbury Press, Belmont, California. Fuchs, C. and Benjamini, Y. (1991), Multivariate Profile Charts of SPC, Series in Applied Statistics Technical Report 91-01, Department of Statistics, Tel Aviv University. Fuchs, C. and Kenett, R.S. (1987), Multivariate Tolerance Regions and F-Tests, Journal of Quality Technology,19, pp. 122-131. Fuchs, C. and Kenett, R.S. (1988), Appraisal of Ceramic Substrates in Multivariate Tolerance Regions, The Statistician, 37, pp. 401-41 1. Haslett, J., Bradley, R., Craig, P . ,Unwin, A. and Wills, G (1991), Dynamic Graphicsfor Exploratory Spatial Data with Application to The American Statistician, Locating Global and Local Anomalies, 3, pp. 234-242. Hawkins, D.M. (1991), MultivariateControlBased on RegressionAdjusted Variables, Technometrics,33, pp. 61-75. Hawkins, D.M. (1993), Regression Adjustment for Variables in Multivariate Quality Control,Journal of Quality Technology, 26, pp. 197-208. .(1973), Covariance StabilizingTransformations, The Annals Holland, P of Statistics, 1, pp. 84-92. Hotelling, H.(1931), The Generalization of Students Ratio,Annuls of Mathemtical Statistics, 2, pp. 360-378. Hotelling, H.(1947), Multivariate quality control illustrated by air testing of sample bombsights, Selected Techniques of StatisticalAnalysis, C. Eisenhart, et al., Editors, McGraw-Hill,New York. Huber, P. (1987), Experiences with Three-Dimensional Scatter Plots, Journal of the American Statistical Association, 82, pp. 448453. IS0 3207-1975, Statistical Interpretation of Data-Determination of a Statistical Tolerance Interval, International Organization for Standardization, Geneva. Jackson, J. (1959), Quality Control Methodsfor Several Related Variables, Technometrics, 1, pp. 359-377. Jackson, J.E. (1985), Multivariate Quality Control, Communications in Statistics, Theory andMethods, 14, pp. 2656-2688.
References
203
JMP IN@(1989), Statistical Visualization Software for the Macintosh? Duxbury Press, Belmont, California. Juran, J.M. (1988), Quality Control Handbook, 4th edition, McGrawHill, New York. of MANOVAStepDown Kabe, D.G. (1984), On the Maximal Invariance Procedure Statistics, Communications in statistics - Theory and Methods, 13, pp, 2571-2581. Kenett, R.S. and Halevy, A. (1984), Some Statistical Aspects of Quality Conformance Inspection in Military Specifications Documents, Proceedings of the 5th International Conference of the IsraeliSociety for Quality Assurance, Tel Aviv 6.2.1, pp. 1-10, Kenett, R.S. and Zacks, S. (1998), Modern Industrial Statistics: Design and Control of Quality and Reliability, Duxbury Press, Belmont, California. Koziol, J.A. (1993), Probability Plots for Assessing Multivariate Normality, The Statistician, 42, pp. 161-173. Marden, J.I. and Perlman,M.D. (1980), Invariant Tests for Means and Covariates, Annals of Statistics, 8, pp. 25-63. Marden, J.I. and Perlman, M.D. (1989), On the Inadmissibility of the Modified Step-Down Test Based on Fisher's Methodfor Combining p-Values, in Glaser et al. eds., Contributions to Probability and Statistics: Essays in Honor of Ingram Olkin, Springer, New-York, 472485. Marden, J.I. and Perlman, M.D. (1990), On the Inadmissibility of StepDown Proceduresfor the Hotelling T 2Problems, Annals of Statistics, 18, pp. 172-190. Mardia, K.V. (1975), Assessment of Multinormality and the Robustness of Hotelling T Test, Applied Statistics, 24, pp. 163-1 7 1, Mosteller,F .and Tbkey, (1977), J. Data Analysis and Regression, AddisonWesley. Mudholkar, G.S. and Subbaiah, P . (1980a), A Review of Step-Down Procedures for Multivariate Analysis of Variance, in Gupta, R.P. ed, Multivariate Statistical Analysis, North-Holland, Amsterdam. Mudholkar, G.S. and Subbaiah, P. (1980b), Testing Significance of a Mean Vector- A Possible Alternative to Hotelling T 2 ,Annals of the Instituteof Mathematical Statistics, 32, pp. 43-52. Mudholkar, G.S. and Subbaiah, l? (1988), On a Fisherian Detour the of
204
References
Step-Down Procedure of MANOVA,Communications in Statistics
- Theory andMethods, 17, pp. 599-611.

Murphy, B.J. (1987), Selecting Out of Control Variables with the T 2 Multivariate Quality Control Procedure, The Statistician, 36,571583. Phadhe, MS. (1989), Quality Engineering Using Robust Design, Prentice Hall, Englewood Cliffs, New Jersey. Ryan, T.P. (1988), Statistical Methods for Quality Improvement, Wiley, New York. Seber, G.A.F. (1984), Multivariate Observations, Wiley, New York. Shewhart, S.A. (1931), Economic Controlof Quality of Manufactured Product, D. Van Nostrand Company Inc., New York. STATGRAPHICS 3.OQ(1988), Statistical Graphics System, STSC, Rockville, Md. Thisted, R. (1986), Computing Environments for Data Analysis, Statistical Science, 2, pp. 259-275. Tukey, J. (1962), The Future of Data Analysis, Annals o f Mathematical Statistics, 33, pp. 1-67. Corrected pp. 812. Tukey, J. (1977), Exploratory Data Analysis, Addison-Wesley. 'hkey, J. (1986), Data-Based Graphics: Visual Display inthe Decades to Come, Statistical Science, 5, pp. 327-339. Velleman, P .(1989),Learning Datahalysis with DataDeskq W.H. Freeman, New York. Velleman, P. and Hoaglin, D. (1981), Basics and Computing of Exploratory Data Analysis, Duxbury Press. Wadsworth, H.M., Stephens, K.S. and Godfrey, A.B. (1986), Modern Methods for Quality Control and Improvement, J. Wiley, New York. Weihs, C. and Schmidt, H. (1990), OMEGA (Online Multivariate Exploratory Graphical Analysis): Routing Searching for Structure with Discussion,Statistical Science, 2, pp. 175-226. Western Electric (1956), Statistical Quality Control Handbook, Western Electric Corporation, Indianapolis, Indiana. Wierda, S.J. (1993), A multivariate Process Capability Index, ASQL Quality Congress Transactions, Boston. Wierda, S.J. (1994), Multivariate Statistical Process Control, WoltersNoordhoff, Groningen.
Index
Acceptance sampling, 3 Alt, F., 6, 56,201 Analytical method, 26 Andrews, D.R. ,26,20 1 A-priori ordering, 83,86, 87, 88 ASQC, 13,94,201 Assignable cause, 3,4,5,29,56,61,
77, 137
Bradley, R., 202 Bruyns, A.F., 52,201 Capability index, 50 Capability study,23,49,50,51,53,
54,56,57,58,60,61,65,72, 126, 127, 131, 139, 140, 141, 181 Case Study 1, 1,4,5,7, 8, 13, 16, 18,21, 72, 73, 77, 91, 93, 115, 116, 117, 127, 128, 131 Case Study 2,4,60, 123, 126 Case Study 3,4,7, 8,37,38, 102, 106,189 Case Study 4,4,95, 192 Case studies, 4,8,99, 102 adulteration, 4, 19 1 aluminum pinyP4, 72, 78, 79 ballistic missile data, 1 18 ceramic substrate, 4,37, 102, 106, 188 cylindrical part, 4 fruit juice,,4, 95, 99, 191 hybrid microcircuit, 37, 188
Attribute data, 4 Base sample, 17,23,30,32,40, 54,

88,128,132,181,191
Batch, 3, 7, 37, 50, 101, 102, 103, 106, 107, 108, 109, 110, 111,
142,188
accept the batch, 103, 106 non-standard batches, 108 production batches, 37, 102, 108 reject the batch, 106 Becker, R.,123,201 Bivariate distribution, 13,32,93 normal distribution, 12, 13, 16 region, 93 value, 71 Box, G.E.P., 26,201
Cause and effect analysis, 50

205
206
Index [Control limit]

39,40,51,52, 53,54,56,57, 66,67,71, 85, 86, 117 Control subject, 137, 138, 139 Corrective action, 3, 79, 102
Chan, L.K., 50,201 Characteristic vector, 115 Chart median [see Median chart] midrange [see Midrange chart] MP [see MP-chart, MP-chart for dispersion, MP-chart for means] multivariate [see Multivariate control chart] principal components [see Principal components chart] R [see R-chart SD [see SD-chart] Shewhart [see Shewhart control chart] simultaneous [see Simultaneous chart] T2[see P-chart] univariates [see univariate chart, x-chart] Cheng, S.W., 201 Chronic problem, 137 Chua, M.K., 127,201 Cleveland, W., 201 Common cause, 3, 137 Competitive benchmark, 139 Completeness, 8,19 Component, 7, 11,30,31,37,40,41,
42,49,69,70,71,74, 88, 102, 108, 113, 115, 116, 117, 118, 120, 127, 131, 132, 134, 143,188,191, 195,196 Computer technology, 121, 123, 143 Computerized dynamic graphic, 123 Conditional hypothesis, 83 Consistency, 19 Contour plot, 13 Contractual requirement, 106 Control factor, 50 Control limit, 5,6, 67, 70, 117 Lower Control Limit (LCL), 117 Upper Control Limit (UCL), 29,
Correlation coefficient, 13,16,17 structure, 5, 11,67,85, 86, 87,

103,106,134
Covariance empirical, 2 1,31,39,56,58 estimated, 15,23, 51, 56 matrix,6, 11,12,14,16,18,21, 22,23,30,31,32,39,49,51,

52, 53,54,56,57,58,66,67, 72, 84,93,94, 115, 116, 198 pooled covariance, 16,39,56 population covariance, 16 structure, 6,37,66 Cox, D.R., 26,201 Craig, P., 202 Critical dimension, 3,4, 11 Critical value, 13,23,26,29,3 1, 39, 40,42,45,51,52,53,54, 57, 58,68,75,80, 83, 85, 87,88, 95,96, 103,131, 184 CUSUM, 77
DataDeskB, 126 Decision process, 102, 106 Decomposition of Tz[see P decomposition] Defective product, 3 Degrees of freedom, 19,20,22,29,
31,40,51,56,93,95, 108
Dempster, A.P., 86,201 Density fbnction, 11 Detection, 6,42,77, 88, 141 Diameter measurement, 117, 118,
132,134,181, 184
Dispersion, 12,58,77,78,79, 128,

131,132,134
Distributions, 6,7, 11, 12, 13, 14, 15,
Index [Distributions] 16, 18, 19,20,21,22,23,26, 29,30,31,32,40,50, 51,52, 53, 54, 56, 57, 58,60,66, 83, 84,85,93,94,95, 104, 108, 137 beta distribution, 20,5 1,53,54 binomial distribution, 104 bivariate distribution, 13,32,93 bivariate normal distribution, 12, 13,16 chi-squared distribution, 3 1,40, 93 F-distribution, 20, 29, 3 1, 52, 53, 54 multivariate distribution, 12, 14, 93 multivariate normal distribution, 11,26,93 negative binomial, 104 Dynamic graphical display, 8
207
Feedback system, 14 1 Fishbone diagram, 50 Fuchs, C., 94,95,99, 110, 127, 128, 129,202 Future subgroup, 56 Godfiey, A.B., 204 Grand mean, 16,20,58,71 Graphical display, 6, 8,77, 123, 126, 127, 141 Graphical technique, 8,26, 121, 140, 143 Group mean, 57,78, 106,141 Group of observation, 93, 127 Group variation, 42 Grouped data, 40,42,58,72,83, 141
Halevy, A., 38,203 Haslett, J., 127,202 Hastay, M,, 201 Hawkins, D.M., 90,202 Efficient estimator, 57, 58 Hoaglin, D., 123,204 Eigenvalue, 115, 116, 117, 199,200 Holland, P.,26,202 Eigenvector, 1 15, 116, 1 17, 199,200, Hotelling, H., 6,21,22,78, 89,96, 141, 198,202. [see also Characteristic Hotelling p , 21,22,78, 89 vector] Eisenhart, C., 6,201,202 Huber, P,,123,202 Ellipse, 70 Hunter, J.S., 26, 201 Hunter, W.G., 201 Ellipsoid, 6 Empirical percentage of coverage, 99 Implementationplan, 8 Empirical power, 3 1,70 Implementing multivariate quality EXECUSTAT@, 126,202 control, 8, 143 Expected value, 20,22,23,29,41, Improvement, 129,204 108 Individual characteristic, 127 Exponentially weighted moving Inferential method, 9, 11 average, 77 Inherent ordering, 89 External requirement, 29 External target, 6,27,29,37,38,39, Inner product, 195 Inspection, 3,38, 103, 104, 106, 107, 139,142 142,203 curtailing, 103,142 False rejection, 90 Internal correlation, 11 Feedback loop, 3,137,139
208
Index
Internal measure of variability, 71, [Matrix] inverse of a matrix, 198 73,141 matrix algebra, 195 Internal measurement, 126 matrix operation, 195 Internal target, 6,50,53 non-singular matrix, 198 Internal value, 73 Internal variability, 39, 54, 58, 60, 74 orthogonal matrix, 199,200 Matrix of scatterplots, 126 ISO, 94,202 Maximum likelihood estimator, 15 JMP IN, 203 Mean vector, 13, 14, 15, 18,22,30, Juran, J.M., 39, 137,203 53,65,84 Median chart,77 Kabe, D.G., 203 Midrange chart, 77 Measure of dispersion, 78, 128 Kenett, R.S., 3,38,50,78,94,95, 99,110,202,203 Measure of variability, 58,71,73, 141 Koziol, J.A., 26,203 Milestone, 8, 135, 140, 143 Laissez faire approach, 3 MINITAB@, 4, 17, 19,22, 126,195, Length measurement, 78, 117, 132, 197, 198,199 134,181,184 macros, 22,35,42 Level of confidence,14,94 Monitored variable, 80, 127, 138, Linear combination, 1 15, 116, 117, 139 120,196 Montgomery, D.C., 127,201 Linear equation, 199 Mosteller, F.,123,203 Linearly independent, 196 Moving average, 77 Location, 3, 12,23, 58,60,77, 102, MP-chart, 127, 128, 129,131 117, 138, 181, 184 MP-chart for dispersion, 131, 132, Lower dimensional set,7 134 MP-chart for means, 131, 132, 134 Mahalanobis distance, 107 Mudholkar, G.S., 85, 86,203,204 Management style,3 Multiple comparison, 11,37,78,90 Marden, J.I,86,87,203 Multivariate Mardia, K.V., 26, 203 computation, 4 Marginal deviation, 71 control chart, 1,6,50,60,75,79, Mason, R.L.,67,88,90,203,204 131 Matrix, 5,6, 11, 12, 14, 15, 16, 17, data, 3,4, 5, 8, 11,47, 79, 113, 18, 19,21,22,23,30,31,32, 120,121, 139 38,39,49,50, 51,52,53,54, data collection, 4 56,57, 58,66,67,71,72,73, distance, 198 84, 88,93,94,96,97, 115, distribution, 12, 14,93 116, 117, 126, 195, 196, 197, methods, 81,88 198, 199,200 normal distribution, 11,26,93 algebra, 17,40 observation, 6, 11, 77, 83, 93, 94, diagonal matrix, 199 116,129
Index
[Multivariate] population parameter, 11 procedure, 86 process capabilitystudy, 49, 50, 58, 141 quality control, 1,7,8, 14,29, 75, 91, 103, 113, 115, 127, 137, 138, 139, 140, 141, 142, 143, 202 statistical tolerance, 102 statistics, 1 18 test, 3 1,37,42, 70 Murphy, B.J., 90,204 Mutually independent, 198 Natural process limit, 13, 14 Natural process region,6, 13, 14 Neighboring subgroup, 77 Noise factor, 50 Nominal value,6 Nonconformities, 3 Normal process, 93,94, 96,97, 98 Null hypothesis, 22, 31, 35, 61,66, 68,83,85, 89,90,93,98, 103,104 Numerical error, 4 One dimensional variable,5 Ongoing quality control,8 Out of control, 79, 83, 86,98, 102, 103,129, 140 Outlier, 53,54,56,60,66, 117, 118, 127, 132 Outlying group,60,80 Outlying observation,6 1, 139 Outlying variable,37 Overall variability, 40,71,74, 108 Pairwise association, 5 Pairwise scatterplot, 78 Parameter of non-centrality, 66 Parent population, 66,67 Parsimony, 116
209 Pattern of association, 5,6 Percentiles, 6,29,40,51,52, 54,93, 95,96 Perlman, M.D., 86,87,203 Phadke, M.S., 50,204 Physical dimension, 3,4, 37, 102, 106, 117,188 Planning strategy, 137 Pooled estimate,39,7 1 Population, 3,6,7, 11, 13, 14, 16, 22,29,30,37,40,41,42,65, 66,67,70, 83, 84,93,97, 103,191 covariance, 16 mean, 30,40,41,42,70 parameter, 11 Power, 3 1,37,42,61,70, 86, 87 Prespecified standard,29 Principal component, 7, 113, 115, 116, 117, 118, 120, 127, 131, 132,134,143 chart, 116, 117, 127, 134 first component, 30,31,40,41, 70,71, 117 second component,11,3 1,4 1,7 1, 117 Prior information,49 Probability statement, 11 Process, 3,4,5,6,7, 8, 12, 13, 14, 17,23,29,37,39,47,49,50, 51,53, 54,56,57,58,60,61, 65, 67, 71, 77, 78, 79,80, 83, 86, 88,93,94,95,96,97,98, 102, 103, 106, 107, 108, 115, 117, 126, 127, 129, 131, 137, 139, 140, 141, 181, 184, 188, 191, 201,202,204 average, 5 capability study,4,23,47,49,50, 53, 54,57,58,60,65, 126, 127,131,139,140,141,181 characteristic, 29, 49, 77, 83 dispersion, 77
210
index Robust design, 50 Roy, J., 85,86,204 Ryan, T.P., 5 1,56,204
[Process] location, 77 output, 49,50 quality characteristic, 49 Production batch, 37, 102, 108 process, 54,60, 102, 103, Profile plot, 13 1 Proportion of coverage, 104 Proportion of multivariate observations, 94 Pure product, 1 91
Sample mean, 17, 110 Sample variance, 16,96 Scaled deviation, 128 107, 137 Scatterplot matrix, 5 Schmidli, H., 126,204 Screening, 3, 103, 106, 108 SD-chart, 77 Seber, G.A.F., 19,204 Sequential analysis, 103 Sequential comparison, 65 Quadratic form, 20,22, 96 Quality characteristic, 49,78,88, Shewhart, 6,7,77,78,80, 116, 128, 131,132,142,204 127 Quality control, 1,3,4,5,6,7,8,9, Shewhart control chart, 7,77, 116, 132, 142 14,29,49,50,63,66,75,77, 91,95,97, 102, 113, 103, Shewhart, SA., 204 115, 116, 121, 135, 127, 137, Significance level, 11,37,54, 85,90, 138, 139, 140, 141, 142, 143, 104,132 191,202 Simple random sample, 1 1 Quality control process, 137, 140, Simulated data, 16,23, 53, 30,58,67 141,191 Simulation, 86,201 Quality Function Deployment, 39 Simulation study, 86 Quantitative data, 4 Simultaneous chart, 128 Random error, 23,29,54,60 display, 1 18 Random variation, 39 measurement, 1 1 Rational subgroup, [see Subgroup] Small sample approximation, 106 R-chart, 77 Sources of variability, 50, 101 Rectangular lines, 70 Special cause, 50,86, 137 Reference population, 13,84,97 Sporadic spike, 137 Reference sample, 6,54,58,63,65, Spring, 201 83,93,94,95,96,98, 103, Spring, F.A., 201 106, 108, 110, 1.39 Squared distance, 6 Regression, 88, 90,202,203 Squared matrix, 196,197,199 Regression-adjustedvariables, 90, Standard deviation, 6,30,40,41,42, 202 49,70,77, 88, 86, 117, 128, Rejection strat,egy, 108 131,132,134 Residual standard deviation, 88 Standards, 29,38,139, 191 Resistance to change, 141 Starplot, 121, 127, 128, 129, 132 Review system, 141 Statistical control chart, 49
Index
Statistical quality control,3,4,77 Statistical software, 8, 129 Statistical tolerance region [see Tolerance region] Statistical tools,3 Statistically designed experiment, 50 Stephens, K.S., 204 Stopping rule, 103 Subbaiah, P,, 85, 86,203,204 Subgroup, 16, 18,20,21,23,26,39, 40,49,51,54, 56,58,65,71, 72,73,74,77,78,79, 127, 128,131,132,134 Subset of variables, 87,90 Sufficiency, 19 Symbol, 79, 123,126,128,131 T2,6, 7,20,21,22,31,42,47,51, 60,79,80, 84, 85, 87, 88, 89, 90,95,96,97, 107, 108, 110, 120, 127, 128, 131, 132, 140, 141,203,204 ?"'-chart, 6,47,60,79, 120, 128, 131, 132,140,141 T2decomposition, 74,83, 88, 89,90, 108,141 T'distance, 6, 7,97, 103, 108 Pstatistic,6,22,31,51,80,90,110,128 rM, 22,23,26,29,31,35,38,39, 40,42,45, 51,52,53,54,56, 57,58,60,66,69,70,71,72, 73,74,75,79, 83,85, 86,87, 88,89,90, 107, 108, 109, 110,120,131, 141 T'", 42,52, 53, 54, 57 P'h, 52,53,54 P D , 40,42,45,58,60,71,73,74, 79, 107, 108, 109, 110, 128, 131,134, 141 P o , 40,58,71,74, 85, 108, 111 Target value, 29,30,3 1,37,41,49, 51,65,68, 83, 102, 128 Test statistic, 1 1, 16,39,40,66,71, 85,90
21l
Tested observation,51,93,94,98, 103,134 Tested sample,6,23,29, 30,31, 39, 65,66,67,68,70,72, 83, 103,108,191 Tested specimen,95,96,97 Theoretical properties, 6, 19 Thisted, R.,123,204 Time sequence plot, 126 Tolerance, 6, 14,91,94,95,96,97, 98,99, 102,107,142,202 Tolerance interval, 94 Tolerance limit, 14,94,97,98 Tolerance region,6, 14,91,94,95, 96,98,99, 102, 107, 142 Tracy, N.D., 51,88,90,203,204 Transformation, 3,26,50, 115, 123, 201,202 Tukey, J., 123,203,204 Unbiased estimate, 15 Underlying distribution, 18,23, 60 Ungrouped case, 23,57 data, 35, 53,60, 87, 116 observation, 16 Univariate chart, 78,85,88, 127,131 methods, 141 significance level, 11 test, 31, 35, 37,70 Unwin, A., 202 Upper Control Limit [see Control limit] Velleman, P.,123,204 Wadsworth, H.M., 78,204 Wallis, W., 201 Weak association, 5 Weihs, C., 126,204 Width measurement, 60, 126, 184 Wilks, W., 201
Index
Wishart distribution, 19,5 1 Within group variability, 60 x-chart, 77 Young, J.C., 88, 90,203,204
Zacks, S., 3,50,78,203

Multivariate Theory and Applications

Uploaded by

Copyright:

Available Formats

Multivariate Theory and Applications

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Multivariate Theory and Applications

Uploaded by

Copyright:

Available Formats

Quality Control

NEWYORK BASEL HONG KONG

To Aliza, Amir and Efrat

4. Quality Control with Internal Targets-Multivariate Process Capability 47

63 75 81 91 101 113 121 135

Appendix 1. MINITABTM Macros for Multivariate Quality

Appendix 2. The Data from

the Case Studies

Multivariate Data with Control Quality

Multivariate Data with Control Quality

Figure 1 .l :Scatterplot matrixfor the 6 variables fromCase Study 1 (ungrouped).

Multivariate Data with Control Quality

Figure 1.2: T2-chartfor the 6 variables from Case Study 1 (ungrouped).

The Multivuriute Normal

Distribution TheNormal Multivariate

We shall indicate that the density of X is of this form by

a ' ; ] where a ; and a ; are the variances of X(')

In the general case (- 1 5 p 5 l), the bivariate normal distributionis given by

The Multivariate Normal Distribution

Figure 2.1 b: Bivariate normal density (rotated).

The Multivariate Normal Distribution

Length 2 from Case Study1.

C(Xi- X)(X, - X ) ' / ( n - 1) ,

I:=[13.69 9.90 9.90 13.69

Normal Multivariate ibution The

and the elements of the S-matrix are given by

- 49.9026)2/49= ,00122 - 60.0441)2/49= ,001367 ,

= 821 = C(X, - 49.9026)(X,2- 60.0441)/49 = .00114

BO1123 S = [ .001204 .001123

.001140 = .883 * 4.001220 * 001367

and the covariance matrix 10,000 E is

X1= k.9976 9.9830 9.9804 14.98481

X 1 = k.9765 9.9702 9.9673 14.96701

8.9558 5.4181 S 1 = 5.5307 7.5527

5.4181 3.2779 3.3460 4.5692

5,5307 3.3460 3.4156 4.6642

7.5527 4.5592 4.6642 6.3693

The computations are performed similarlyfor the 50 subgroups and we obtained

X = x x j / 5 0 = k.9864, 9.9793, 9.9752, 14.97681

The Multivariate Normal Distribution

and 10,000 S p = 10,000 * C(50- 1)Sj/[50 * (2 - l)]

1.2151 1.1428 1.1600 1.2669

1.1428 1.3559 1.2854 1.2792

1.1600 1.3288 1.5728 1.4183

1.2669 1.3033 1.4183 1.6854

MTB> EXEC 'POOLED.MTE3' 1.

(ii) If X is distributed as in(i) above, then

then the quadratic form

The Multivariate Normal Distribution

The T i statistic can be expressed in terms of the means of the : ) as measurements Y

Normal bution Multivariate The

Quality Control with Externally Assigned Targets

(YI - rnit))W1(YI -m!))

with Externally Asslgned Targets

Quality Control with Externally Assigned Targets

Quality Control with Externally Assigned Targets

Quality Control with Externally Assigned Targets

-1.000 1.081 0.100 -0.432

0.923 0.615 0.100 1.336 0.063

-0.432 0.063 2.620