STATISTICAL QUALITY

CONTROL METHODS
 STATISTICS

Textbooks and Monographs

A SERIES EDITED BY

D. B . 0 WEN, Coordinating Editor

Department of Statistics
Southern Methodist Unioersitu
Dallas, Texn1

PAUL D. MINTON JOHN W . PRATT

Virginia Commonwealth U niversity Harvard University
Richmond, Virginia Boston, Massachusetts

VOLU;\·fE 1: The Generalized Jackknife Statistic

H. L. Gray and W. R. Schucany
VoLUME 2: Multivariate Analysis
Anant M . Kshirsagar
VoLUME 3: Statistics and Society
Walter T . Federer
VoLUME 4: Multivariate Analysis:
A Selected and Abstracted Bibliography, 1957-1972 (out of print)
Kocherlakota Subrahmaniam and Kathleen Subrahmaniam
VoLUME 5: Design of Experiments: A Realistic Approach
Virgil L . Anderson and Robert A. McLean
VoLUME 6: Statistical and Mathematical Aspects of Pollution Problems
edited by John W. Pratt
VoLUME 7: Introduction to Probability and Statistics (in two parts).
Part I: Probability. Part II: Statistics
Narayan C. Giri
VoLUME 8 : Statistical Theory of the Analysis of Experimental Designs
J. Ogawa
VOLUME 9: Statistical Techniques in Simulation (in two parts)
Jack P. C. Kleijnen
VoLUME 10: Data Quality Control and Editing (out of print)
Joseph I. Naus
VoLUME 11: Cost of Living Index Numbers: Practice, Precision,
and Theory
KaliS. Banerjee
VoLUME 12: Weighing Designs. For Chemistry, Medicine, Economics,
Operations Research, Statistics
KaliS . Banerjee
VOLU;\IE 1:3: The Search for Oil: Some Statistical Methods
and Techniques
edited by D. B. Owen
VoLU!\fE 14: Sample Size Choice: Charts for Experiments with
Linear Models
Robert E. Odeh and Martin Fox
VoLUME 15: Statistical Methods for Engineers and Scientists
Robert M . Bethea, Benjamin S. Duran, and Thomas L. Boullion
VOLU:\IE 16: Statistical Quality Control Methods
Irving W . Burr

OTHER VoLUMES IN PREPARATION

STATISTICAL
QUALITY
CONTROL
METHODS
Irving W. Burr
Professor Enze11'tus
Statistics Department
Purdue Uniuersity
West Lafayette, Indiana
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 1976 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed in the United States of America on acid-free paper
International Standard Book Number-13: 978-0-8247-6344-2 (Hardcover)
Library of Congress Catalog Number: 75-17457

This book contains information obtained from authentic and highly regarded sources. Reprinted material is quoted with
permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts have been made to publish
reliable data and information, but the author and the publisher cannot assume responsibility for the validity of all materials
or for the consequences of their use.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized
in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying,
microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copyright.com
(http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC) 222 Rosewood Drive, Danvers, MA
01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users. For
organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for
identification and explanation without intent to infringe.

Library of Congress Cataloging-in-Publication Data

Catalog record is available from the Library of Congress

Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
To Terry, Mary Kate, Peter, and John
 PREFACE

The basic aim of this book is to give a clear, sound and natural exposition

of statistical methods, especially useful in quality control in industry. The

criteria used in choosing t op ics and their presentation are usefulness and wide

appli cability, rather than elegance, The emphas i s is on data-analysis and dec i sion-

making . The statistical techniques discussed ar e also of great use in areas not

specifically quality control, such as labor& tory analyses and scientific research .

Methods of statistical quality control a:e becoming of greater importance

for many reasons: (a) the heightening of competition, both domest ic and inter-

national (quality being one of the key ingredi ents ) , (b) the increas ing n eed to

avoid loss of material and to save man-hours of time, (c) the intense profit

squeeze , (d) the rise of consumerism, (e) the rapid increase in number of l ega l

liability cases which emphasizes the need for greater reliabil i t y of the product

and adequate documentary records for legal defense, (f) the need to know one's

process capabilitie s , (g) ne w and stricter quality related l aws (e.g., on weights

and purity), (h) the proliferation of mandat ory industrial standards, and ( i) the

growth of international standards for intern ational trade. This book is therefore

of direct interes t to t h e over 20,000 membe rs of the American Society for Qua l ity

Cant rol.

The book is written for junior, sen ior an d graduate students in engi neering,

industrial man agement, statistics and science, for industri al courses and also fo r

se lf- st udy by those in indus try. It is assumed th at the student has at leas t a

l itt l e background in statistics . Ch apter 2 reviews the basics of probability and

statis tics for t hose with some background. l·ioreover, this ch apte r can almost

stan d on i t s own for the reade r with very l itt le or no previous statistical back-

v
vi PREFACE

ground . Some use is made of calculus, but not to a great extent as the book deals
mostly in concepts .
Nearly all of the problems and illustrative examples are from actual cases

in the author's extensive collection, gathered through the years from industry

and from research at Purdue University and elsewhere . The y thus represent actual

practical problems. The aim of the book is, however, unity and cohesiveness, to

emphasize principles rather than to present a host of case histories . The latter

can easily leave the student with a feeling of frustration and disunity. Presen-

tation of too many competing techniques and too many details and exceptions is

avoided.

Some attention is given to the growing tendency toward automation in industry .

Some calculational aids are presented. Very little is included in the direction of

"decision theory," because of the large amount of inputs required which are often

unobtainable, and the inability to codify such approaches so as to avoid starting

from "scratch" every time. Nevertheless, elements of this viewpoint often appear

in the book, subjectively guiding the choice of critical levels and assumed risks .

The present book is to some extent an outgrowth of the author ' s "Engineering

Statistics and Quality Control , " McGrmv-Hill , 1953 . Some paragraphs , data and

examples taken from this volume are used in the present book without specific

referencing .

For notations, the author has chosen to use Greek letters for population

charactersitics and ordinary (Latin) letters for sample characteristics for

measurement data, as is common practice in most statistical areas. Prime marks

are used for population characteristics for counted data. In connection with

notations, the reader will find the glossary to be of much use.

The author owes much to the pioneers in statistical quality control at the
Bell Telephone Laboratories: George D. Edwards, Walter A. Shewhart, Harry G. Romig,
PREFACE vii

and especially to Harold F. Dodge, with whom the author worked on the Standards
Committee of the American Society for Quality Control for over 20 years. The

author has had the privilege of working with many other people from industry, of

whom we may particularly mention Arthur Bender , Jr., Richard Bingham, Walter Breisch,

Matt Dalton, Richard Ede, Ernest Fay , Hugh Ferguson , Aldis Hayes, Carl Hoover,

John Malmstrom, Harry Marsh , George McDermott, Paul Peach, Carl Noble, Kort Pfabe ,

Keith Ross and Wade Weaver .

It has also been most rewarding to have taught in many intensive courses with

Professors Cecil Craig, John Henry, Charles Hicks, Lloyd Kno1~ler, Gayle ~1cElrath,

Edwin Olds, Mason Wescott and Holbrook Working , as well as with some of those

from industry mentioned previously.

The author would like to heartily thank Professor Shanti S. Gupta, Head of the

Statistics Department at Purdue University, for encouragement and generous use of

facilities. Warm thanks are also due to the department's secretaries for expert

typing of the manuscript: Norma Lucas, Edna Hicks, Sandy Emory and Rebecca Fagan .

Finally, the author must heart i ly thank his wife, Elsie, whose interest,

encouragement and enthusiasm have been so helpful .

Irving W. Burr
 CONTENTS

PREFACE v

1. INTRODUCTION

1.1 Hhat is Statistics? 1

1.2 Hhy Statistical Quality Control? 1
1.3 Aim of Book 3
1. 4 Prerequisites 3
1. 5 Suggestions to the Student 3

2. BRIEF REVIEW OF STATISTICAL BACKGROUHD 5

2.1 Population and Sample 5

2.2 Parameter and Statistic 5
2.3 Notations 6
2.4 Probability Definitions and Laws 9
2.5 Distribution of Sample Statistics 12
2.6 Statistical Estimation 16
2.7 Testing Hypotheses 18
References 21

3. CONTROL CHARTS IN GENERAL 23

3.1 Introduction 23
3 .2 Running Records of Performance 23
3.3 Random and Assignable Causes 26
3.4 The Control Chart for Interpreting Variations in Quality 29
3.5 Two Purposes of Control Charts 32
3.6 Economic Balance Between Two Errors 33
3.7 Meaning of a Process in Control and Potential Advantages 33
Problems 35
References 38

4. CONTROL CHARTS FOR MEASUREMENTS 39

4.1 Introduction 39
4.2 Control Charts for Averages, Ranges and Standard
Deviations, from Given Standards 39
4.3 Control Charts for Averages, Ranges and Standard
Deviations, Analyzing Past Data 49
4.4 Comparison of the Two Charts for Variability 53
4.5 Comparison of a Process with Specifications 54
4.6 Continuing the Charts 60
4.7 Illustrative Examples 61
4.8 Kinds of Assignable Causes 78
4.9 When to Set Standard Value s - Process Capability 83
4.10 Us e of Runs for Out of Controlness 85
4.11 Summary 86
Problems 89
References 96

ix
X CONTENTS

5. BACKGROUND OF CONTROL CHARTS FOR MEASUREMENTS 97

5.1 Introduction 97
5.2 The Empirical Approach to Sampling Distributions 97
5.3 Control Charts for Samples from Normal Populations of x's 98
Problems 106
References 106

6. CONTROL CHARTS FOR ATTRIBUTES 107

6. 1 Introduction 107
6.2 Control Charts for Defectives 108
6. 3 Control Charts for Defects 128
6.4 General Comments on Control Charts for Attributes 138
6. 5 A Method of Rating Product Quality 1lf0
6. 6 Summary 142
Problems 144
References 156

7. MISCELLANEOUS TOPICS IN CONTROL CHARTS 157

7. 1 Control Charts for Individual Measurements, Hoving Ranges 157

7. 2 Tool Wear and Other Gradual Process Changes 161
7.3 Charts for x and s, Large Samples 168
7. 4 Charts for x and R or s , with Varying Sample Size 169
7.5 Different Types of Samples 172
7.6 Relative Efficiency of Different Types of Charts 180
7. 7 Artificial "Specifications" for Greater Efficiency
in Attribute Charts 183
7. 8 Summary Charts for ~ and R 185
7. 9 Variations of Control Charts 186
7.10 Summary 190
Problems 190
References 196

B. APPLICATIONS OF CONTROL CHARTS 203

8. 1 ~lhat Manufacturing Problem to Work On 203

8. 2 Attacking an Industrial Problem 204
8. 3 List of Typical Applications of Control Charts 207
8. 4 Use in the Laboratory 211
8. 5 The Quality Control Person 212
8. 6 Bibliography on Applications 213

9. ACCEPTANCE SAHPLING BY ATTRIBUTES 225

9.1 Why Acceptance Sampling? 225

9.2 Characteristics of Sampling Palns 226
9.3 Acceptance Criteria for Attribute Acceptance Plans 228
9. 4 Operating Characteristic (OC) Curve. Defectives. 230
9.5 The Average Outgoing Quality Curve, AOQ Curve 239
9. 6 The Average Sample Number Curve, ASN Curve 243
9. 7 The Average Total Inspection Curve, ATI Curve 244
9.8 Characteristics of Double Sampling Plans 245
9.9 Further Notations on OC Curves 253
9. 10 Characteristics for Defects Plans 255
CONTENTS xi

9.11 Devising a Sampling Plan to Match Two Points on an OC

Curve 257
9.12 Use of Control Chart Record with Acceptance Sampling 259
9.13 Summary - General Principles of Sampling Acceptance
and Some Misconceptions 260
Problems 269
References 271

10. SOME STANDARD PLANS FOR ATTRIBUTES 273

10.1 Introduction 273

10.2 The Dodge-Romig Sampling Inspection Tables 275
10.3 The ABC or Hilitary Standard 105D Plan 281
10.4 What Next? 315
10.5 Summary 318
Problems 319
References 322

11. ACCEPTANCE SAMPLING BY HEASUREMENTS 323

11. 1 Potential Greater Efficiency by Variables

than by Attributes 323
11.2 Dependence Upon Knowledge of Distribution of
Individual x's 324
11.3 Use of Control Charts and Frequency Distributions
for Past Data 324
11.4 Decision-Making on Lot Means, Known o 325
11.5 Decision Making on Percent Beyond a S~ecification,
Kno~m o 331
11.6 Decisio~ Making on Percent Beyond a Specification,
Unknown a 332
11.7 Single Sa~pling for Variabi~ity 336
11.8 Military Standard 414 Plan 341
11.9 Checking a Process Setting and a Process Capability 345
11.10 Summary 347
Problems 348
References 350

12, SEQUENTIAL ANALYSIS 353

12.1 Introduction 353

12 . 2 The General Approach 354
12.3 The Sequential Probability Ratio Test 355
12.4 Six Special Cases of Sequential Analysis Summarized 356
12 . 5 Some Examples 363
12.6 Use in Checking a Process Setting 369
12.7 A Sequential Test for Process Capability, Using Ranges 371
12.8 Use in the Laboratory 371
12.9 Two Derivations of Acceptance-Rejection ConstanLs 372
12.10 Summary 375
Problems 375
References 379

13, SOME OTHER SAMPLING PLANS 381

13.1 Sampli ng Inspection of Continuous Production 381

13.2 Chain Sampling Plans, ChSP-1 390
xii CONTENTS

13 . 3 Skip-Lot Sampling Plan 392

13 . 4 Cusum Charts 395
13.5 Surmnary 399
Problems 400
References 401

14 . STATISTICS OF COHBINATIONS, TOLERANCES FOR HATING PARTS 403

14 . 1 Introduction 403
14 . 2 Statistics of Sums and Differences 404
14.3 Use of Distributions in Assemblies 40!3
14.4 Tolerancing Assemblies by Specifying Process Controls
and Acceptance Sampling 417
Products, Quotients , and Other Functions 422
Summary 430
Problems 431
References 435

15 . SOliE OTHER FREQUENCY DISTRIBUTIONS 437

15.1 Introduction 437

15.2 Homents of Distributions 437
15 . 3 Some Frequency Distribution Nodels 441
15.4 A Goodness of Fit Test for Frequency Data 454
15.5 Summary 460
Problems 461
References 463

ANS\.JERS TO ODD-NUHBERED PROBLm!S 465

APPENDIX 475

GLOSSARY OF SYNBOLS 507

AUTHOR INDEX 515

SUBJECT IliDEX 519

 CHAPTER 1

INTRODUCTION

1.1. lfuat is Statistics? Statistics is the science concerned with the

drawing of inferences and making decisions from data which are subject to

variation. Wherever results vary from trial to trial, time to time or

person to person, you have a statistical problem, whether you like it or

not, whether you know it or not. Since random variation is so universall y

present in industry, engineering and science in general, the question

becomes not whether you are going to perform the functions of a statistician;

you are, each time you make a decision or draw an inference! The only ques-

tion is as to how well you are going to perform that funct i on.

Sometimes, of course, especially where the random variation is rela-

ti vely minor, the dec is ion is perfectly "obvious." In such cases no statisti-

cal methods are needed. But often enough what appears obvious, proves to be

unjustifiabl e or false. Moreover there are many cases where the deci sion is

not at all clear, and statistical methods are invaluab le in estimat ing the

risks and providing objective decisions. Statistical methodology is also

concerned with the setting up of experiments, the drawing of samples and the

proper analysis of results, for maximum effi ciency and objectivity.

1.2. Why Statistical Quality Control? Quality control in various forms

has been practiced for thousands of years, certainly dating back to the building

of the pyramids in Egypt. But the use of statistical quality control, that

is, the use of statistical methods in the control and improvement of quality

in industrial production, is a quite recent development. The real beginnings

1
2 INTRODUCTION

were made in the 1920's by men in the Bell Telephone Laboratories, among
whom we may mention Walter A. Shewhart, Harold F. Dodge, Harry G. Romig,
George D. Edwards, Thornton C. Fry and E. C. Molina.
A basic reason for the need for statistical quality control is that
industry is continually trying to work to ever closer tolerances. It is

not difficult to make parts all "exactly" alike to some fairly gross degree
of precision. But when we are asked to have a certain dimension of a part
to always lie within an extremely narrow tolerance range (two limits), we

may be forced to the utmost with our production and measurement capabilities,
and statistical control becomes basic. For example, years ago, a diesel

engine plant had to make plunger rods for forcing fuel through small holes.

The diameter of the rods was to meet a total tolerance range of only .00004".
Even with guages they had constructed, measuring to the nearest .00001", the
problem was very difficult. They could of course produce rods all exactly

alike to the nearest . 001" or even . 0001" perhaps. But to the nearest . 00001"
there were varying diameters.
Another form of variation to be controlled is the incidence of non-

conforming parts in industrial production. These may have one or another

dimension out of tolerance, or have some defect of slight or substantial

importance. Production completely free of all such non-conforming parts is

the ideal of course. But it may be unnecessary, or uneconomic, or even

impossible. In the latter case is the production of millions of parts each

of which can have any one or more of 20 different defects of varying impor-
tance. Another common example in which we cannot have absolute assurance of

all parts meeting standards, is that in which the test destroys the part.
Some examples of such tests are tensile strengths of rivets, life tests of

electronic parts, and muzzle velocity of ammunition. If we test every

part and perhaps find that all met the standard in such a test, we have
none left to use!
1.5 Suggestions to the Student 3

In all these cases statistical quality control methods are of basic

importance.

1.3. Aim of Book. In the last, say, twenty years, there has been a

great proliferation of methods of statistical quality control. In this book

we cannot begin to cover this broad a field. What we can do is (a) to cover

a basic set of methods of statistical quality control of wide usage and

flexibility, (b) to emphasize general principles and philosophy, and (c)

to provide the reader with background upon which to grow in his use and

application of statistical quality control. To make full use in a company

of just the techniques herein presented could take fifty years.

Furthermore the aim is not to present statistical methods in general,

however interesting this would be, but instead to concentrate on those of

particular use in statistical quality control.

1.4. Prerequisites. It will be assumed that the student is somewhat

familiar with basic statistical methods, as covered in a beginning course.

For those without such background, however, as well as a refresher for those

with this background we present a review in Chapter 2. Also some knowledge

of elementary calculus is assumed. However, in order to make the book

readable to those who are weak in these two assumed prerequisites we shall

try to make the early parts of sections and chapters as readable as possible

and take up more technical aspects further along. The answers to problems

can also be a considerable aid to those having difficulties.

1.5. Suggestions to the Student. In statistics it seems to be

impossible to avoid some difficulties with notations. This is perhaps

especially true in statistical quality control. Hence the student is

urged to make full use of the glossary.

A great many of the problems in the book use data, in most cases,

actually obtained in industry. Care in calculation is important as always.

A slide rule is usually sufficient. (But please, not like some people's
4 INTRODUCTION

idea of "slide-rule accuracy" permitting about a 3% error!) Of course a

desk calculator, logarithms and tables of square roots are also useful.

In substantial calculations, retaining more places than fully justified,

and then rounding off at the end, is good practice.

In this book many different industrial parts and products occur in

examples and problems. If you are not fully (or at all) acquainted with

one, do not let this bother you, the author may not be either. The important

thing is that the part or product has one or more measurable characteristics

whose distribution we may study. Or there may be some properly defined

"defects" we wish to study the incidence of. In both cases we can use

statistical distributions without intimate knowledge of the product in ques-

tion.

The student is urged to try to think of examples and illustrations in

areas where he does have some background. This can considerably strengthen

his grasp of the methods and make them more interesting.

A healthy skepticism of the obvious conclusion and an active seeking for

all possible errors of tabulation, calculation, logic, nonapplicability of a

technique or interpretation should be practiced and developed. Until one has

made several serious blunders, it is difficult to realize the diabolical

variety of traps which lie in wait for the unwary statistical worker. Only by

eternal vigilance can they be avoided.

 CHAPTER 2

BRIEF REVIEW OF STATISTICAL BACKGROUND

2.1. Population and Sample. In statistical work we are concerned

with populations of objects, parts, people, trials or measurements. Such

a population may be finite as in the first three cases or conceptually

infinite as in the last two. Each individual in the population can be

characterized by (a) a number such as length, weight or strength, or (b)

can possess one or more attributes, such as, being defect-free or having

one or more defects. These two ways of characterizing the individuals of

a population gives rise to numeri cal or measurement data, and to attribute

data respectively. Rather different statistical population models are used

for the two different cases. The student should often be aski ng: Is this

collection measurement or attribute data?

Probably the most basic question in all of statistics is the relation

between a sample and the population from which it comes. How does the

population determine what to expect of samples from it, and what does the

sample tell about the population? Of course neither question has any answer

unless the samples are drawn in an approved, unbiased manner, usually and

hopefu lly at random, such as, by using a table of random numbers. (See

Table IV in the back of the book.)

In general, no matter what the population is like, it will determine, in

some manner , the behavior of the samples, whether or not we can fully charac-

terize this relation.

2 .2. Parameter and Statistic. In genera l a population is characterized

by one or more "parameters," and a mathemat ical model, particularized by the

5
6 BRIEF REVIEW OF STATISTICAL BACKGROUND

given parameters in question. Thus a stable industrial process producing

parts whose diameter is of interest, wi11 have a population of some form

with a mean diameter and standard deviation (typical departure from the mean).

The "form" may be "normal" or not. Or a stable industrial process may be

producing light bulbs, a fixed proportion of which fail to light when first

used. This proportion is the parameter characterizing this "binomial" popula-

tion. It is binomial because, from this viewpoint, there are but two kinds of

bulbs produced: those lighting and not lighting.

As distinct from population parameters, we also have the characteristics

of a sample described by a sample statistic or several. Thus for a sample of

diameters from the population, we may find the sample mean and the sample stan-

dard deviation. These statistics are supposed to be, and hopefully are, clos·e

to the respective population parameters. (How close? See Section 2.5.) Such

sample statistics are often used as estimates of the respective population

parameters.

In the other example cited, we might count the light bulbs in a sample

of 500 which fail to light initially. If this is five, then we would find

the sample proportion to be 5/500 = .010. This would be regarded as an

estimate of the population parameter: the long-run proportion not lighting,

under this process.

A sample containing a substantial collection of observations is often

studied by making a frequency tabulation (of the observations into numerical

classes), and then picturing the distribution by some frequency graph such as

a histogram (blocks of height proportional to the frequency in each numerical

class).

2.3. Notations. We now come to the troublesome problem of notations.

Unfortunately there is not just one notational system for statistical measures.

Fortunately there are now basically only two such systems. The most widely
2.3. Notations 7

used one reserves Greek letters for population parameters, and ordinary letters
for sample statistics. Corresponding letters in the two alphabets are used

insofar as convenient or feasible. On the other hand a notational system,

widely used in the quality control area [1]* and in industrial standards
groups around the world, is to use for a sample statistic any convenient

symbol, whether ordinary letter or Greek, and then to put a prime mark on it

to designate the corresponding population characteristic. This is a flexible

system and makes good sense. Moreover it introduces fewer Greek letters for

the layman to learn to use. But unfortunately it is not the system toward

which most fields of statistics have gradually evolved.

The author was greviously torn between using one or the other. Fortunately

at least some progress has been made toward uniformity, in the standard [1]

of the American Society for Quality Control. After considerable soul-searching,

the author decided to use the notation, ordinary letter- Greek letter,
for the sample statistics versus population characteristics, but use

the prime notation for attribute data. Parentheses will also be used to

include alternative notations.

We shall use the following notations. An observed measurement will be

denoted by x. Then for a finite population of N numbers x, the population

mean will bet

l\ (or x' ), (2 .1)

*Such bracketed numbers refer to references at the ends of chapters.

If the population is infinite, then the definition is strictly analogous, but

defined by an integral instead of a summation.

8 BRIEF REVIEW OF STATISTICAL BACKGROUND

The sample mean for n observed x values is analogous

X' (2. 2)

For the "variance" of a finite population of N x 1 s:

l:~=l (xi - IJ )2
(2. 3)
N

This is the mean squared deviation from the population mean. For a sample of

n - 2
l:i=l (xi - x)
(2.4)
n - 1

The n - 1 in the denominator (instead of n) is the number of degrees of freedom

for the sample. It is used so that if we were to take all possible samples
from the population, and for each, find s 2 , then the mean of all of these very
many s 21 s will be exactly cr 2 We may summarize this relation by using the
expectation symbol E, and say that
2
(orcr' ), (2. 5)

(This is true for all populations having a cr 2 , and all samples of two or more.)

Likewise it may be proven that

E(x) = p = Px (or x'), (2.6)

Next we have the respective standard deviations. They are merely the

square roots of (2.3) and (2.4):

yl:~=1 (xi - p )2
cr = cr X (or cr') 1 (2.7)
N
n - 2
l:i=l (x.l - x)
= s • (2.8)
n - 1
Each is a type of average departure of x's from their mean.
2.4. Probability Definitions and Laws 9

For discrete or counted data, we shall be concerned with two types: We

may have a sample of n parts and count how many of these n are non-conforming

or defective, that is, possess one or more "defects", suitably defined. In

this way, each part or piece is either a good one or a defective. If we let

d be the number of defective or non-conforming parts in a sample of n,then

we use for the sample "fraction defective"

p = d/n = sample fraction defective. (2.9)

We note that pis a sample statistic. For a finite lot of N parts, we suppose

that there are D defectives. Then for the lot or population fraction defective

p1 = D/N =population fraction defective. (2.10)

Here p' is a parameter. (In the Greek system n or o/ might be used. neither of

which is too happy a choice.) Another usage of p' is for a stable process, for

which the probability of a defective is constantly p', as each part is produced.

This is analogous to an infinite lot.

The second type of discrete data is that in which we count the number of

defects in a sample of one or more parts. Thus for example we can count the

number of "seeds" (small bubbles) on say, 24 glass bottles. Then

c = number of defects, (2 .11)

Here c is a statistic. The parameter is the long-run average number of

defects on such samples.

c' = population average number of defects. (2 .12)

(In the Greek system, letters used are:\, Y and p.)

2.4. Probability Definitions and Laws. In general we are concerned

with experiments or trials in which the outcome is uncertain. We can describe

the "sample space" of possible individual outcomes. For example,we may have

a lot of 100 parts and a trial consists of drawing two parts without replace-
10 BRIEF REVIEW OF STATISTICAL BACKGROUND

ment. Regarding the 100 parts as distinguishable somehow, perhaps by a serial

number, we then have a sample space consisting of all possible samples of two
from the lot. Counting unordered samples we have by combinations C(lOO, 2) =
100(99)/2 = 4950 possible outcomes or points in the sample space. If drawing

is at random, these are all equally likely and the probability for each is
1/4950. In general an "event" is a collection or set of possible outcomes.

For example if 10 of the 100 parts are defective, then an event might be that

the sample of two contain no defectives. How many such samples are there? The

answer is C(90, 2) = 90(89)/2 = 4005, and these too are equally likely.

With finite numbers of equally likely outcomes the definition of the

probability of an event A is

p (A)
no. of ways A can occur N (A)
(2. 13)
total no. of possible outcomes N (S)

where Sis the sample space (an event too). In this equally likely case we
see that

1. 0 .:_ P (A) < 1

2. P (S) =1
3. If events A and B have no outcome in common P(A or B) P(A) + P (B)
In our example, we have

P ( no
d f .
e ect1ves
) N(no
= N(S) defectives) = 4005 = . 809 .
4950

If we extend the last to a countable sequence of sets, these three

properties may be used as defining postulates for probabilities of events.

For finite sample spaces with equally likely individual outcomes, there
is much need for counting techniques such as permutations and combinations,

because the numbers are so huge.

In probability we often have use for "conditional probabilities." We

may ask what is the probability of event A, if it is known that event B did

occur, designated
2.4. Probability Definitions and Laws 11

P(AjB) =probability of A given B did occur. (2 .14)

For this we limit our space from S down to just those outcomes of S in which

B does occur. Then among these, how many are also favorable to A?

P(AjB) = N(A and B) (2. 15)

N(B)

In the example suppose that part number 1 is a good one and we define event B

to be the event that the sample of two contains part number 1. Now N(B) = 99,
because there are 99 other parts with which to complete the sample of two.

Next let A be the event that the sample contains no defective. Then

I
N (A B) = 89

because there are 89 good parts other than number 1. Hence

P(AJB) = 89/99 = .899.

Note that this is higher than the .809 figure for A by itself. P(AjB) can be

higher, lower or equal to P(A). If equal then we call events A and B

"independent," because the occurrence of B has no effect upon A.

Thus:
P(AjB) = P(A) A, B independent. (2.16)

Dividing numerator and denominator of the fraction in (2.15) by N(S) we

have
P(AJB) = P(A and B) (2 .17)
p (B) '

P(A and B) = P(B)P(AJB), (2. 18)

From (2.16) and (2.18) we have an alternative definition of independence:

P(A and B) = P(A)P(B) A, B independent (2.19)

We are now in a position to state four useful laws of probability:

l. Multiplicative laws

a. P(A and B) P(B)P(AJB) P(A)P(BJA) (2. 20)

b. P(A and B) P(A)P(B) A, B independent (2 .19)

12 BRIEF REVIEW OF STATISTICAL BACKGROUND

2. Additive laws

a. P(A and/or B) P(A) + P(B) - P(A and B) (2.21)

b. P(A and/or B) P (A) + P (B) A, B mutually exclusive (2. 2 2)

2.4.1 Interpretation of a Probability. Suppose that on trial after

trial the probability of an event remains constant at p'. The observed ratio

of occurrence of the event in question is p = d/n. How are p and p' related?

Intuition and logic both indicate that p approaches p' somehow, as n increases.

The manner of approach is not, however, absolute as in calculus. Instead it

is an approach "in probability.'' Suppose that p' = .5; then we might wish to

continue trials until p is between .48 and .52. There is however, no absolute

guarantee that p will ever lie between these limits no matter how large an n

is taken. But we can determine an n, sufficiently large so that we can have,

say, a .99 probability that p lies between .48 and .52. Such an approach is

called "stochastic," or we say p approaches p' in a statistical sense.

2.5. Distribution of Sample Statistics. We shall briefly review the

behavior of sample statistics for samples drawn at random from a more or less

completely specified population.

The population commonly assumed for measurements x is the normal dis-

tribution. It is often at least a reasonable first approximation to observed

distributions of data. Given that the mean is p and standard deviation a,

then the probability density function is

f(x) affiT exp [- (x~~~ 2 J _ oo < X < co (2.23)

In order to use tables of normal curve probabilities it is necessary to

standardize the variable. For this let

z=~ or x = )1 + za. (2. 24)

a

Then for this standardized variable, z, the density function becomes

1 2
H z ) = - exp [-(z /2)] (2. 25)
l2iT
2.5. Distribution of Sample Statistics 13

Tables of probabilities for (2.25) are areas under curves between two limits.
Many tables, such as our Table I, give the probability for the random variable

Z to lie between -~ and z in the table. This is a cumulative probability.

2.5.1. Samples from a Normal Population. Of particular interest for such

samples is the behavior of the sample mean, x, and the variance s 2 . If the

population has mean p~and standard deviation crx, then we have

Distribution of Sample Means from Popn. N(~x' cr~)

1. Mean x's = ECx) = Px (2. 26)

crx
2. Standard deviation x 1 s = cr-X = (2. 27)
rn
3. The x 1 s are normally distributed

Note that x's center at the same value ~X as the x's do, but that the

variability crx is less than that for the x's, crx' by a factor vn. Thus

to cut crx in half we need n = 4 observations for x, etc. This might be

called the ''inverse square-root law of statistics.''

Next consider the distribution of sample variances, s 2

Distribution of Sample Variances from Popn. N(~x' cr~)

1. (2.28)

2. Standard deviation of s 2's = cr 2 = cr 2 ~ (2.29)

s
2
3. The variable (n-l)s2 follows the chi-square distribution with
cr
n - 1 degrees of freedom.

The chi-square distribution is tabulated in Table II. The chi-square

distribution is extremely skewed, being J-shaped for one or two degrees of

freedom, and becomes less unsymmetrical as the degrees of freedom increase,

2
but the longer tail is always to the right side, that is,for high x values.

2.5.2. Non-normal Continuous Populations. If a population is non-norm~l,

but with mean~ and standard deviation cr, then (2.26), (2.27) and (2.28) still

hold exactly. Moreover the distribution of x's approaches normality as n

increases, by the Central Limit theorem. Formula (2.29) is approximately

14 BRIEF REVIEW OF STATISTICAL BACKGROUND

true, and we will commonly assume the chi-square distribution for

(n - l)s 2;cr 2 , knowing that it is only approximate.

2.5.3. Binomial Data. When we have the following conditions we have a

binomial sampling distribution:

Conditions for Binomial Distribution

1. Two possible outcomes per trial, e.g. good vs. non-conforming, or

success vs. failure.

2. ConstgD1 probability p' for one particular type of outcome, for

example, a non-conforming part.

3. Independence of trials: each trial independent of previous outcomes.

4. Taken trials, with interest in how many yield the type of occur-

renee in question.

Given these conditions, then

P(d occ. in n) d = 0, 1, 2, ... , n. (2.30)

This is essentially a sampling distribution consisting of a sample of n trials

for each of which we note whether we have an occurrence or non-occurrence.

For the binomial we can show

E (d) Pd = np' (2.31)

/np I ( 1 - pI} ' (2. 32)

The shape of the distribution is always at least somewhat unsymmetrical if

p' < .5, with the longer tail ford values above np'. As n increases, with

p' fixed, the distribution becomes "normal."

Likewise, for p d/n, we have

E (p) p' (2. 33)

{p'(l- p')/n. (2. 34)

The distribution shape is like that for d of course.

2.5 Distribution of Sample Statistics 15

2.5.4. Poisson Data. Another sampling distribution much used in

statistical quality control is the Poisson distribution. It is the typical

model for analyzing data on the number of defects on samples of material.

For example, we may count visual defects on 24 bottles, specks and other

blemishes on test areas of paper, typographical errors on a page, defects

in each of a series of sub-assemblies, or breakdowns of insulation in 1000

meters of insulated wire being tested.

Practical Conditions for Poisson Distribution

l. The number, c, of "defects" suitably defined to be counted over

"equal areas of opportunity,"

2. The defects occur independently, that is, the occurrence of one

defect does not make it more, or less, likely that another will

occur in the sample.

3. The maximum possible number of defects on a sample is vastly greater

than the average number c'.

Under these conditions (made more rigorous) we have

e-c' (c')c
P(c defects) c ~ 0, 1, 2, ... ' (2. 35)
c!

where e = 2.71828· . . .
Thus knowing the parameter c', we can find the probability for any number of

defects c to occur. Moreover we can prove

E(c) fie = c' (2. 36)

(J
c
0. (2. 3 7)

The Poisson distribution is also unsymmetrical with the long tail toward high

values of c, for all c' > 0, but becomes more and more symmetrical as c'

increases.
16 BRIEF REVIEW OF STATISTICAL BACKGROUND

2.6. Statistical Estimation. One of the main uses for drawing samples
and analyzing them is to estimate what the population is like. More specifi-

cally we commonly assume a~ of population, for example, normal, binomial

or Poisson, and then try to estimate the parameter(s) from the sample or

samples at hand. (In quality control, as we shall see, we often have a

series of samples.) There are two forms of estimation: point and interval.

2.6.1. Point Estimation. In point estimation the objective is to

obtain a one-number guess at the parameter, from the sample. In general

there are more than just one such estimator and we want to choose the "best."

Several criteria of "best" are: (a) The theoretical average value of the

estimator should be the parameter, We then have an unbiased estimator of

the parameter. (b) The estimator should be subject to as little error as

possible, that is, among unbiased estimators we would like that estimator

with minimum standard deviation. (c) The estimator should use every bit of

relevant information in the sample. The following are estimators, each pos-

sessing all three of these desirable properties:

x for Px (normal population of x's) (2.38)

2 2
sx for ox (normal population of x's) (2.39)

p for p' (binomial population) (2.40)

c for c' (Poisson population). (2.41)

Hence we use them.

2.6.2. Interval Estimation. Since it is not to be expected that we

can hit a parameter with a single guess or estimate we commonly use interval

estimation. For this we choose (preferably in advance of sample-drawing) a

desired degree of confidence or probability that the interval which we deter-

mine from the sample will contain the true parameter. This is often 90% or

95%. Then from the sample data we find two limits, and assert that the parame-

ter lies between, having, say, 90% confidence that we are telling the truth.
2.6. Statistical Estimation 17

Slightly less rigorously we say that we are 90% confident that the parameter
lies between the two limits from the sample. This seems entirely satisfactory

for practical purposes.

2
For a normal population of measurements we have Px and ax to be estimated.

In general,elementary statistics courses discuss limits giving 1- a confi-

dence. For Px• when ax is somehow known, we have the following limits
a
X+ Z ~ (2.42)
- (a/2 above) ;n
where z is the standard normal variable and the probability (area) above is

a/2. Such limits are very nearly correct, even when the x's are not normal

and n is small.

However, again assuming a normal population of x's, if we do not know

ax' then we are forced to use the sample standard deviation s, and the so-

called t distribution as follows:

x +- t
(n - 1 degr. of freedom, a/2 above)
s
10 (2.43)

Then we use Table III in the Appendix.

For 1 - a confidence limits for ax' we use the sample standard deviation

and the chi-square distribution of Table 1I. Limits for a, derivable from

the distribution of s 2 (Section 2.5.1) are

s. j x~n-1 n-l

d.f., 1-a/2 below) S•

JX~n-1 n-1
d.f., a/2 below)
(2.44)

For discrete data populations, there are many approaches possible, the

easiest of which is to use such a table as [2], and for given d and n merely

look up the limits. However, there are rather approximate limits available

which are useful. For binomial data, with p < .5 and np at least 10, say, the

following approximate limits may be used:

Limits for p'• p ± z( /Z b ).jp(l-p). (2.45)

· a a ove n
18 BRIEF REVIEI'l OF STATISTICAL BACKGROUND

These stem from (2.33), (2.34) and the approach of the distribution of p to
normality.
Likewise for a Poisson sample we observe c defects. Then if c is, say,
at least 10, then we may use the approximate limits:

Limits for c 1: c±z (a/2 above) •IC. (2.46)

These came from (2.36), (2.37) and the approach toward normality of the
Poisson distribution as c' increases.

Many other cases of interval estimation are usually given in beginning

statistics courses, for example, limits for function~of parameters for two

populations, such as p1 - p 2 .
2.7. Testing Hypotheses. Complementary to the technique of estimation

is that of testing hypotheses. The basic question is as to whether the observed

sample could reasonably have come from a population more or less completely
specified by the hypothesis. The hypothesis may be concerned with a parameter

or with the ~of population. For the latter we might for example test the

hypothesis that the population is normal. Tests of goodness of fit are used
for such hypotheses, such as the chi-square test. More often we are concerned
with one (or more) parameters, and compare the observed sample statistics with

the hypothesized parameter.

Elements of Testing a Hypothesis on One Parameter, for Example, p·

1. Basic assumptions are made which are assumed true and not open to

question in the test. Commonly the ~of population is assumed,

for example, that it is normal. The value of parameters not being

tested may also be assumed, for example, o = 5.

2. Although in a sense a null hypothesis is an assumption, it is not

like the assumptions in 1. Instead the hypothesis is under test

2.7. Testing Hypotheses 19

and may be rejected, whereas we never use our test to reject an

assumption as in 1. An example of a simple null hypothesis is

p = 100. (We could also have a composite hypothesis p~ 100.)

3. Rejecting a hypothesis when it is actually true is committing an

error of the first kind.

4. Because of variability it is in general impossible or infeasible

to make a test for which the probability a of an error of the first

kind is zero. Nevertheless we do want to keep the risk a at some

specified low value, perhaps .01.

5. A sample statistic (an estimator) for the parameter in question is

chosen. For p we could use the mean x or the median, commonly

choosing the former because of its smaller standard deviation. A

sample size n is usually decided upon also at this point.

6. An alternative hypothesis is next chosen, containing other values of

the parameter considered possible, or of economic or scientific

interest. For example, as alternatives top = 100 we could have

p < 100, p > 100 or p f 100.

7. The chosen risk a in 4, the type of alternative hypothesis in 6 and

knowledge of the distribution of the statistic in 5 enables us to set

a critical region or rejection region for the statistic. Then when-

ever the statistic lies in this region we reject the basic or null

hypothesis and accept the alternative hypothesis. But if the statis-

tic does not lie in the rejection region we "accept" the null hypothe-

sis, meaning that it perfectly well could be true. A critical region

-
can be stated for the sample statistic itself,e.g. x or s, or for a

standardized transformation of it, for example, z = X - ru 0 or

o;rn
 20 BRIEF REVIEW OF STATISTICAL BACKGROUND
2
2 (n - l)s The critical region will have two parts for alter-
X 2
cro
natives such as p * 100, and but one part for those such as p < 100

or p > 100. In the latter two cases the critical region is in the

lower tail or upper tail of the distribution of x around 100

respectively.

8. An ~of the second kind is committed when we accept the null

hypothesis when in fact it is not true. The probability of an error

of the second kind is called s. But note carefully that S depends

upon how far off from the null-hypothesis value the parameter

actually is. For the example, if the alternative is ~ > 100, and

cr =3 and n = 4, then if~ actually is 101, Swill be large, but if

~ 110, Swill be very small. Thus here S is a function of~.

9. In general it is well to draw an operating characteristic (OC)

~giving the probability of acceptance of the null hypothesis

for each value of the parameter. This tells how much discriminating

power the test has. If insufficient then a larger sample size is

needed. In statistical quality control, much use is made of OC

curves.

To make significance tests we simply use the sampling distributions of

the various sample statistics as given in the four subsections of Section 2.5.

We can set up a critical region for z or t for testing x vs. p, or for x2

for testing s 2 vs. cr 2 Then calculate an observed z, t or x2 and see if it

lies in the critical region. Or, if we have attribute data, we could

find a critical region for d or c. But usually it is easier to find the

probability of as great or a greater discrepancy from the expected than was

observed, If this is < a (or a/2 if a two-tail test) we reject the hypothesis.
References 21

For example if n = 100, p' = .05 and the alternative hypothesis is p' > .05,
and a = .01, then what decision is made if d = 12? Tables show

P(l2 or moreln = 100, p'= .05) = .00427. So the hypothesis pi = .05 is

rejected in favor of p' > .05.

Elementary books give tests for two samples as well. We shall be more

interested in single sample testing, double sampling (a second sample

contingent on the first) and upon a series of samples. Also we shall discuss

sequential sampling.

References

1. American Society for Quality Control,Standard Al-1971, Milwaukee, Wise.

2. D. Mainland, L. Herrera and M. I. Sutcliffe, "Statistical Tables for

Use with Binomial Samples-Contingency Tests, Confidence Limits, and

Sample Size Estimates." New York University College of Medicine,

NY, 1956.
 CHAPTER 3

CONTROL CHARTS IN GENERAL

3.1. Introduction. In this chapter we shall introduce Shewhart

control charts in general terms. Then in Chapters 4 through 8, various

types of control charts and their applications are discussed. Control

charts are of great use in the analysis and control of manufacturing

processes, so as to produce quality that is satisfactory, adequate,

dependable and economic.

3.2. Running Records of Performance . One often sees in industrial

plants, running record graphs of performance, plotted up each day or

whenever new data become available. The points may represent various

kinds of quality performance such as dimensions, strengths, per cent

rejections, or defects found at the final inspection. Or, the data may

not be directly quali t y associated, such as, production or sales figures,

absenteeism, number of accidents per week, power consumption, stock market

prices, etc.

In Figure 3. 1 we show four running records of quality performances.

These are the plottings of data for three products, as given in Tabl e 3.1.

The top two gr aphs are measurement data for samples of n =5 roll ers. The

top chart for x's is of course concerned with the process level, while the

second for ranges, R, is concerned with process variability. The latter

two charts are for attribute or counted data. The third shows the fraction

defective of a certain test on electrical equipment, the sample size being

about 1340 per day. The last is for misalignments, whi ch is just one of

23
 24 CONTROL CHARTS IN GENERAL

4720

4715

4710

Roller Diameters .0001"

0
Roller Diameters . 0001"

• OS
p

()
Fraction Defective B Test

IS

!0

c
5

()
Mj sa 1 ignment s

FIG. 3.1. Running record charts for data of Table 3.

many different categories of defects, for which a record was kept. Which

of the four running record charts seems to show the greatest stability or
control, that is, seems to have just a random pattern of points? Which
 3.2 Running Records of Performance 25

Table 3.1. Four Running Records of Quality Performance Over Time

~lean and range in .0001" of Fraction defective Number of misalignments

diameters of rollers for on "B" test, two c, in 50 aircraft
samples of n=S, specifications factories, August, sub-assemblies
.4710" + .0010" Sample size about 1340

Sample X R Date Frac. Def. Assembly c Assembly c

4714.0 4 2 .0374 1 7 26 7
2 4716.8 1 3 .0518 2 6 27 13
3 4716.6 2 4 .0427 3 6 28 4
4 4714.0 2 5 .0512 4 7 29 5
5 4714.0 4 6 .0323 5 4 30 9

6 4713.6 3 7 .0558 6 7 31 3
7 4714.4 3 9 .0325 7 8 32 4
8 4713.8 2 10 .0673 8 12 33 6
9 4713.2 1 11 .0106 9 9 34 7
10 4713.8 3 12 .0649 10 9 35 14

11 4713.6 5 13 .0265 11 8 36 18
12 4712.6 6 14 .0425 12 5 37 11
13 4714.6 2 16 .0426 13 5 38 11
14 4711.2 3 17 .0293 14 9 39 11
15 4713.4 3 18 .0487 15 8 40 8

16 4 715. 0 3 19 .0448 16 15 41 10
17 4715.2 5 20 .0413 17 6 42 8
18 4713.2 4 21 . 0357 18 4 43 7
19 4712.4 3 23 .0390 19 13 44 16
20 4713.0 4 24 .0204 20 7 45 13

25 . 0369 21 8 46 12
26 .0284 22 15 47 9
27 .0342 23 6 48 11
28 .0660 24 6 49 11
30 .0314 25 10 50 8

31 .0296

seems to show the poorest control or most unstable performance? Try to

guess.

Performance Varies. Certainly noticeable in all four graphs is

variation. Results do vary from time to time, material to material, etc.

26 CONTROL CHARTS IN GENERAL

Variation is universal, it is part of life, therefore it is the part of

wisdom to learn to live with it. In the first graph both high and low
points are undesirable, while in the other three, the lower the point is
the better we like it. A misconception we must guard against is that of

concluding that since the points vary, the underlying production conditions
are also varying. Such conditions may in fact be just as uniform as it is
possible for human beings to make them, and still the points will vary,
simply by chance.
Unusual Performance Calls for Action. In all four graphs, the purpose

in plotting them is to keep track of the respective processes so as to take

appropriate action whenever it is indicated. Thus in the first graph of

averages, x's, resetting of the process is called for if an x strays too

far away from the desired process average, in ~ direction. On the
other three graphs too high a point may well be regarded as a danger signal
and some action on the process may be needed to improve performance. Thus
a high range R, for example means that the process is not consistently
repeating itself.

What Performance is Unusual? This is the fundamental question in all

such analyses of varying performance, and where does performance not vary?

Commonsense simply is not a reliable guide for avoiding two types of

errors: (a) taking action when none is desirable, and (b) missing
opportunities for taking appropriate action or learning what makes the

process behave as it does. In order to minimize the chances of making

these two types of errors, we need to use the control chart and other
statistical tools. These risks are those called a and B in Chapter 2.

3.3. Random and Assignable Causes. In the preceding section we

have begun to consider two more or less distinct types of causes. By
3.3. Random and Assignable Causes 27

random or chance ~we mean the whole host of small influences lying
behind the particular measurement or result we happen to obtain. For

example, in machining an outside diameter of some part, even when we try

to hold all conditions as alike as possible, there are bound to be slight

differences in diameter from one angular direction to another, from one

position along the axis to another, and from one piece to another. The

hardness of the stock will vary slightly even within a single piece. The

rotational speed, the depth of cut and pressure of the tool will tend to

vary, as will the lubrication, coolant, floor vibration, and temperature.

Then after the piece is removed, the actually recorded measurement is ob-

tained by an inspector with a gage. Different places along the piece,

varying temperatures of gage and piece, presence of slight dust or varying

thickness of oil film, varying pressure of gage upon the piece, the position-

ing of the piece in the gage, and finally the inspector's reading of the

gage itself, can each cause results to vary. All of these and many others

not mentioned may each have a relatively minor effect, but all taken toge-

ther give variability to the resulting measurements. Note that the foregoing

causes are operative even if we assume that all conditions are held as con-

stant as is humanly possible. Thus there is bound to be some variation

natural to every process, if only we try to take fine enough measurements.

Such variation we call natural to the process, and say that it is due to

chance causes. Ideally the results behave as though we were drawing

numbered chips at random from a bowl containing a fixed population (repla-

cing each chip before drawing the next).

Although no absolutely hard and fast distinction can be made between

chance and assignable ~' they differ typically in the three character-

istic respects shown in the accompanying tabulation.

28 CONTROL CHARTS IN GENERAL

Effect Worth seeking?

Number present of each
Chance causes Very large number Slight Not worth looking for
Very few, perhaps only Well worth seeking out,
Assignable causes Marked
one or none and possible to do so

In the study of a running record chart, our problem therefore becomes

one of trying to tell when a high or low point on the graph (1) is an indica-

tion of the presence of an assignable cause of a change in the process, which

should be sought out,or (2) can so readily occur by chance causes alone that

there is no reliable evidence of the presence of an assignable cause. This

is then the fundamental problem: When shall we assume that an assignable

cause has been responsible for some apparently unusual performance, and when
shall we attribute such performance to chance?

Some people think that the problem of deciding when to take action is an
easy one. It is true that in some situations the assignable cause is so

obvious that anyone can pick it up just by a glance at the chart. Unfor-

tunately not all cases are so clear-cut, and also many samples that seem to
be perfectly reliable indicators of assignable causes actually have arisen

from chance causes and should therefore have been attributed to chance.

Before going further the reader is challenged to try to pick out from

among the poinrs in Figure 3.1, which ones are reliable indicators of the

presence of assignable causes. Without probability to guide you, it is a

difficult and hazardous job.

To further emphasize the difficulty, the author has often shown

audiences two running records of the same identical data plotted to two

different scales. Nearly all will pick the one with a small vertical

scale as showing the ·~est control:' Then it is told them that the two

are of the same data, and pointed out that if a decision depends at all

upon the scale used in plotting then it cannot be a very objective decision.
 3.4. Interpreting Variations in Samples 29

3.4. The Control Chart for Interpreting Variations in Samples. The

epoch-making discovery and development of the control chart were made in

1924 and the following years by a young physicist of the Bell Telephone

Laboratories, Walter A. Shewhart. Wrestling with a problem made complicated

by the presence of random variation, he came to realize that the problem was

statistical in nature. Some of the observed variation in performance was

natural to the process and unavoidable. But from time to time there would

be variations which could not be so explained.

He reached the brilliant conclusion that it would be desirable ~

possible to set limits upon the natural variation of any process, so that

fluctuations within these limits could be readily explained by chance

causes, but any variation outside this band would indicate a change in

the underlying process. There followed the development of charts for

measurements and both kinds of attributes, the concept of the rational

sub-group and its use in most effectively letting the process set the limits

of natural variability, and a large amount of practical experimentation in

trying out the new methods in actual plant problems. Then in 1931 came

Shewhart's great "Economic Control of Quality of Manufactured Product," *

in which the whole field was laid out including its theory, philosophy,

applications, and, most pertinently, its economic aspects. Few fields of

knowledge have ever been so completely explored and charted in the first

exposition. The industrial field of quality control gets its name from this

book. Control charts are often called "Shewhart control charts," Since all

control charts stem from Shewhart 1 s work, we shall use the shorter term,

recognizing that "Shewhart" is the implied prefix for all of them.

*D. Van Nostrand Company, New York.

30 CONTROL CHARTS IN GENERAL

The basis of all control charts is the following: Any varying quantity

(whether a measurement x, an average x, a range R, a standard deviation s, a

fraction defective p, or a count on the number of defects c) forms a distri-

bution if chance causes alone are at work. Any such distribution has a mean

and a standard deviation. Quite regardless of the shape of the distribution

(unless extremely badly behaved), there will be, by chance causes only, very

few points outside of the band between the mean minus three standard deviations

and the mean plus three standard deviations. Hence, having set such limits,

we have a band of normal variability for the statistical measure in question.

Now suppose a point does go outside this band of normal variability. It is

conceivable that such a point is just due to a rare "ganging up" of chance

causes, and that no assignable cause was at work. But since industrial processes

are so prone to contain assignable causes, it is a much better bet that the

point outside the band is due to some assignable cause. Hence when such a

point comes along, we assume there was some assignable cause at work and try

to see what process conditions might have changed while this sample was

produced and tested. Conversely, when a point lies inside the control

band, we do not say that no assignable cause was present while the sample

was produced and tested, but only that we have no reliable evidence for

supposing that there~ an assignable cause at work. Hence no action is

taken. We attribute such points within the band as being due to chance

causes only. Thus with control charts we save time looking for assignable

causes when none are present and use this saved time for more careful search

in those cases in which we do have reliable evidence of some nonchance factor

at work. As we shall see, the control chart is a most powerful tool for making

more fruitful our efforts in stabilizing and controlling our processes at

desired performances.

Now let us consider again the data given in Table 3.1. Figure 3.2 shows

the same running records as did Figure 3.1, but now we have added central lines
 3.4. Interpreting Variations in Samples 31

4720

4715
X

4710
Roller Diameters .0001"

0
Roller Diameters .0001"

.OS
p

0
Fraction Defective B Test

15

10
c
5

0
Misalignments

FIG. 3.2. Control charts for data of Table 3.1.

(solid) and control limit lines (dotted). These latter bound the respective

bands of normal variability. As we see the chart for the ranges R is in per-

feet control, that is, the within-sample variability is apparently homogeneous

The chart for misalignments shows only one point of the 50 marginally outside.
32 CONTROL CHARTS IN GENERAL

It is worth investigating to see why so many misalignments occurred on this

subassembly. Although there is perhaps some reason for each misalignment,
there are so very many opportunities for misalignments in each subassembly

that the total number per sub-assembly behaves in quite a chance fashion
(Poisson distribution).

The chart for xs

1 reveals evidence of an assignable cause in the second

and third points and also that for number 14, a low point. The B test results
show five points indicating causes to be sought. The high points are taken as

indications of "bad" assignable cause, while the two low points suggest

"good" assignable causes (unless due to lax inspection). How did you,the

reade~ make out in your guessing?

3.5. Two Purposes of Control Charts. Up to this point in the chapter

we have been primarily concerned with the first purpose of control charts,

namely to analyze past data for control. The idea is to try to see whether

the results are homogeneous. Could they readily have come from a single

population, that is, by chance variation from a stable process? Statistically,

we are testing the hypothesis that we have, say, k samples from a single
population with constant but unknown parameter(s). Such an approach is

commonly used until the process begins to show really good control. Note

that the control limits are set by the process itself in this case.

A second purpose of control charts is to compare observed results

against a population whose parameter or parameters are known. The hypothesis

being tested is that this population, completely specified, is the one from
which the samples are drawn. Such an approach is used after exploratory

work is done on the process, and reasonable standards can be set for it.

In one sense, comparing each single sample point with the control band is a

little significance test. Is the sample compatible or not? The a risk for

each point is very small, perhaps .005 or less. But having all of 20 sample
3.7. Meaning of a Process in Control 33

points lying inside the control band has a much larger a risk, perhaps

around .05 or .10. This too is a test of the hypothesis that the specified

population with the given parameters is the one from which the samples were

drawn, as well as being a test of homogeneity. This case of control charts

is often called "standards given." Note that here the standards set the

limits, not the current data, as was the case of "analysis of past data."

3.6. Economic Balance between Two Errors. In the preceding sections

we have been concerned with two types of errors which the industrial worker can

make. The first is to conclude that a sample differs significantly from the

other samples or the standards when in reality the discrepancy is entirely

due to chance causes. This type of error leads us to hunt for an assignable

cause 1vhen none is present. Many undesirable process changes and unjustified

blaming of personnel, tools or material stem from this error.

The other general type of error is to miss noting the presence of

some assignable cause of either poorer or better performance. Such errors

may well miss evidence of impending or actual trouble.

Much experimentation in the Bell System from 1924 onward seemed to indi-

cate that the use of limits at plus and minus three standard deviations

around the central line givesa good balance in industrial situations between

the risks of the two errors. Subsequent industry-wide experience has born

out the choice. If ± 2o limits were used we would tend to be looking too often

for non-existent or relatively weak assignable causes. Or if we were to use

± 4o limits we would too often miss indications of assignable causes and

opportunities for process improvement.

3.7. Meaning of a Process in Control and Potential Advantages. Basically

we say a process is "in statistical control" when results behave like samplings

from a single population. This can be thought of as drawings with replacement

34 CONTROL CHARTS IN GENERAL

from a bowl of numbered chips. Or if there are enough chips, replacement be-

comes unimportant.

Some of the advantages which may accrue when a process is brought into

good control are:

1. The process appears to be free from sources of variation worth identi-

fying, and hence it is doing about all that we can expect of it. If this is
not good enough, then we know that we shall have to make a more or less

fundamental change in the process. Just "tampering" with it cannot be

expected to help.

2. The act of getting a process into good statistical control ordi-

narily involves the identification and removal of undesirable assignable

causes and possibly the inclusion of some good ones such as new materials

or methods. Hence quality performance has commonly been much improved.

Many a process which had been unable to hold specifications proves easily

able to meet them when brought into control.

3. If a control chart for laboratory tests on uniform material shows

lack of control, results from the test are suspect, because the lack of

control shows that the laboratory conditions have not been satisfactorily

standardized. On the other hand, if a control chart of reasonable length

shows good control, we are justified in assuming homogeneity of laboratory

conditions.

4. A process in good statistical control is predictable. It can be

counted upon. Therefore if the chart for incoming material from a supplier
Problems 35

shows good control, we can justifiably decrease inspection, plottingeach addi-

tional point on the chart, and merely watching for any evidence of lack of

control. Thus we do not need to regard each submitted lot as a law unto it-

self. In the presence of good statistical control by the supplier all the

previous lots supply evidence on the present lot, which is not safely the case

if there has been lack of control.

5. If our own processes are in good statistical control, we can more

safely guarantee our product. We know what it will do.

6. When testing is destructive, a chart in control gives us confidence

in the quality of the untested product, but if the process is out of control,

we are in a far weaker position for predictability on the remainder .•

7. A chart in control in experimentation enables us to determine

soundly the experimental error, which is not the case if the chart shows

lack of control of variability.

8. The soundest way to cut inspection is through getting the process

into control.

PROBLEMS

3.1. Make a list of chance causes and a list of possible assignable

causes for some specific problem you are familiar with in your field of

interest.

The following tables list quality performance data. Plot the x- and R

in the first two, p in the third or c in the fourth, as assigned, and comment

on the running records.

 Table P3.3. Final Mooneys (Measure
Table P3.2. Pounds Per Square Inch Of Internal Viscosity) of GR-S Rubber
To Operate Hydraulic Stop-Light Samples of Four Successive VI
C7'
Switch, Four Switches Per Sample Measurements

Average x Range R Average x Range R

73 10
83 20 49.750 6.0
75 49.375 2.0
15
71 5 50.250 2.0
76 5 49.875 2.5
47.250 1.5
71 5
77 20 45.000 6.0
77 5 48.375 1.5
66 10 48.500 1.0
59 20 48.500 1.0
46.250 3.5
76 20
80 20 49.000 2.0
95 10 48.125 1.5
79 25 47.875 2.5
84 15 48.250 2.0
47.625 1.5
77 20
75 0 47.375 1.5 n
50.250 3.5 0
76 5 z
47.000 2.0 -l
76 10 ;N
47.000 1.0 0
80 15 r
49.625 2.5 n
72 20 ~
49.875 1.0 ;N
71 5 -l
47.625 2.5 Cfl
96 25
49.750 1.5 >-<
86 15 z
86 10 48.625 2.0 C)
tTl
z
tTl
76 5 ;N
)>
82 5 r
81 10
 ...,
...0
0"
.....
(I)
sUl
Table P3.5. Average Number of
Table P3.4. Fraction Defective in Defects Per Truck, Daily
Samples of 100 Auto Carburetors Production 225 Trucks

Sample Fraction p Date Average No. Defects, c

1 .10 Sept. 6 (Labor Day)
2 .04 7 .81
3 .06 8 .80
4 .12 9 .80
5 .06 10 .62
6 .08
7 .10 13 .67
8 .12 14 .67
9 .08 15 . 74
10 .07 16 .92
11 .03 17 .74
12 .04
13 .03 20 .86
14 .04 21 .79
15 .04 22 .71
16 .10 23 .72
17 .07 24 .86
18 .05
19 .03 27 .87
20 .06 28 1.21
21 .08 29 .87
22 .10 30 1.40
23 . 04
Oct. 1 .78

Vol
-...!
38 CONTROL CHARTS IN GENERAL

References

W. A. Shewhart, "Economic Control of Quality of Manufactured Product."

Van Nostrand-Reinhold, Princeton, New Jersey, 1931.
2 American Society for Quality Control, Standard B3-1970. Also American

National Standards Institute, 21.3-1970.

 CHAPTER 4

CONTROL CHARTS FOR MEASUREMENTS

4 . 1. Introduction . In the preceding chapter we have discussed the

general characteristics of control charts, and their usefulness, and empha-

sized two general purposes : (a) process control versus given standards ,

and (b) analysis of past data for control .

In the present chapter, we are concerned with the most frequently

used control charts for measurements : x charts for the process level, and R

charts and s charts for the process variability. In general one will use an

x chart for average s , and only one of the R and s charts. However, we shall

carry along both of the latter variability charts for comparison. There are

a number of other control charts for measurements, some of which are dis-

cussed in Chapter 7 .

4. 2. Contro l Charts for Averages, Ranges and Standard Dev i ations, from

Given Standards. In this section we consider the case i n which we have

given the two parameters p and cr . These may result from substantial past

history of good control at satisfactory performance , or possibly be i n the

nature of goals . In the latter case one might possibly use the mi d-point of

the specified tolerance band for the standard mean ~· But the standard a

should virtually always be the result of experience , for otherwise all proba-

bility interpretations of points outside the control limits and inside too

are lost.

In order to illustrate statistical principles in this book we often

resort to sampling experiments, or at least suggest some. In order to illus-

trate x, R and s charts with p and a given we shall use Population A of Table 4 . 1.

39
 40 CONTROL CHARTS FOR MEASUREMENTS

Table 4.1. Approximately Normal Populations for Sampling Experiments

Number x A B c D E F G

+11 1
+10 1 1
+ 9 1 1 1
+ 8 1 3 3
+ 7 1 3 5 10
+ 6 3 5 8 23
+ 5 1 10 8 12 39
+ 4 3 23 12 16 48
+ 3 10 39 16 20 39 1
+ 2 23 48 20 22 23 3
+ 1 39 39 22 23 10 10 1
0 48 23 23 22 3 23 3
- 1 39 10 22 20 1 39 10
- 2 23 3 20 16 48 23
- 3 10 1 16 12 39 39
- 4 3 12 8 23 48
- 5 1 8 5 10 39
- 6 5 3 3 23
- 7 3 1 10
- 8 l 1 3
- 9 1
-10 1
N 200 200 201 201 200 200 200
fl 0 +2 0 +1 +4 -2 -4
G 1. 715 1.715 3.47 3. 47 1. 715 1. 715 1. 715
X

In this table are listed several convenient populations with which to experi-

ment. These are the same, approximately normal distributions as were exten-

sively used in the famous War Production Board courses in quality control

by statistical methods, as developed by Working and 01ds [1). One of

their most valuable contributions to the presyntation of statistical quality

 4.2. Control Charts from Given Standards 41

control in particular and of statistical methods in general was the widespread

and ingenious use of various sampling experiments.

Beads or chips may be marked with the numbers listed in the first column

in appropriate frequency; Any one population, or a combination of several

may placed in a bowl, thoroughly mixed, and samples drawn.

Experimentation with beads or chips may possibly seem at first rather

childish to some readers. But it has been the author's experience that a

great many people find that such experiments illuminate the principles of

random variation and sampling better than the use of equations and formulas.

In fact it seems to prove valuable and interesting even for those with con-

siderable mathematical sophistication. Even if an experiment is not performed,

a description of how one might be designed to illustrate a technique, can

be valuable, after one has seen a few experiments.

It is helpful for the reader to learn to associate the sample drawings,

from populations like those in Table 4.1, with coded results from data from

problems in his own field. This merely requires a little imagination. For

example, we may consider Rockwell hardness on the C scale. If the specifi-

cations are 50 to 60, we may record our readings relative to the nominal 55.

Thus 57 is listed as a +2 and 51 as a -4. Other examples are the following:

Carbon in steel, .20 to .30%

Moisture content in stack of lumber, 6.5 ± .5%

*It is well to have different colors for the different populations, so that

they may be readily sorted out after mixing, should this happen intentionally

or by mistake. Chips may be dyed by Easter egg dye, or differing colored

India inks used. Beads 1 em in diameter are convenient. Circular chips,

from thick fiber, are sometimes waste output from stampings and can be

obtained free.
 42 CONTROL CHARTS FOR MEASUREMENTS

Compression of springs to specified length, 12.0 ± .5 lb

Outside diameter of a roller, 2.214 ± .005 in.
Weight of an ingot, 5,300 ± 50 lb

Root-mean-square finish measurement, nominal 25, maximum 30

Edge widths of piston rings ground to .2120 ± .0005 in,

Resistance of coils, 800 ± 5 ohms

Timing of relays, 30 ± 5 sec

Weights of castings, 4.15 ± .OS oz

On temperature at which a thermostatic switch operates,SS ±5°

Purity of oxygen to nearest .05 per cent, 99.50 to 100.00 per cent

Thermometers ± ,5° from standard,

4.2.1. A Sampling Experiment. Table 4.2 shows the results for drawings
of 50 samples, each of n = 5 observations, x. Population A of Table 4.1 was

used for all 250 drawings which were made with replacement after each indi-

vidual x, so as to have always the same population*.

Then for each sample x, R and s were found, as listed in the last three

columns of the table.

The three series of sample statistics are plotted in Figure 4.1. Note

that the units on the vertical scales differ. They were chosen as shown,to

give roughly the same variability on each graph. On examining the charts

for x's and R1 s it is seen that the two are not at all correlated. This is

as it should be because x is concerned with the process level, whereas, R is

a measure of the sample variability. On the other hand, the R and s charts

~ correlated, the high and low points correspond perfectly, and the entire

graphs would fairly well coincide if one were placed over the other. They

*Actually this is not necessary, since a sample of five does not deplete the

population of 200 appreciably. But after a sample of five, replacement should

be made.
 Table 4 . 2 . Random Samples of Five from Distribution
A of Table 4 . 1 , Drawings of One Number at a Time
and Replacing Before the Next Number was Drawn
Standard
Sample xl x2 x3 x4 x5 Ave!_'age Range Deviation
No . X R s
1 0 +1 0 0 -2 - .2 3 1.10
2 -1 +2 +1 +2 -1 + .6 3 1. 52
3 -1 +3 0 +1 -2 + .2 5 1. 92
4 -5 +1 0 0 -1 -1.0 6 2 . 35
5 0 +2 +1 +1 -2 + .4 4 1.52
6 +3 -2 +2 +1 0 + .8 5 1. 92
7 -1 +3 +1 +1 -1 + .6 4 1. 67
8 +1 -1 -2 0 -3 -1.0 4 1.58
9 -1 0 +1 -1 +2 + .2 3 1.30
10 -2 -5 -2 +2 +1 -1.2 7 2. 77
11 0 -2 +1 +1 +1 + .2 3 1. 30
12 +1 +2 0 +1 -1 + .6 3 1.14
13 0 -2 -5 0 -1 -1.6 5 2 . 07
14 -2 0 +1 0 -4 - 1.0 5 2 . 00
15 +3 -4 0 0 0 - .2 7 2. 49
16 +1 +1 +1 -1 +2 + .8 3 1.10
17 0 +1 -1 0 +1 + .2 2 . 84
18 -2 -2 +3 -2 -1 - .8 5 2.17
19 -2 +2 +2 -2 +2 + .4 4 2.19
20 0 0 +1 +1 -1 + .2 2 . 84
21 +2 +1 0 -1 +1 + .6 3 1.14
22 -2 -2 -1 -3 0 -1.6 3 1.14
23 -4 +4 -2 0 0 - .4 8 2 . 97
24 -2 -1 -1 0 0 - .8 2 .84
25 0 0 +1 -2 0 - .2
+ .4
3 1.10
26 -1 0 +1 +3 -1 4 1.67
27 -4 +1 0 +1 +3 + .2 7 2 . 59
28 +1 0 +4 0 -3 + .4 7 2.51
29 -1 +1 -2 -1 +1 - .4 3 1. 34
30 +2 +2 -2 -4 0 - .4 6 2 . 61
31 +2 -1 0 -3 0 - .4 5 1. 82
32 0 -2 +4 +1 0 + .6 6 2 . 19
33 +1 +1 0 +2 0 + .8 2 . 84
34 -1 -3 0 +1 +1 - .4 4 1.67
35 - 0 +2 -1 -2 +2 + .2 4 1. 79
36 0 -3 -1 +2 -3 -1.0 5 2.12
37 -1 +I +1 -1 -1 - .2 2 1.10
38 0 -2 +1 +1 -2 - .4 3 1. 52
39 -2 -3 +2 0 0 - .6 5 1. 95
40 0 -3 0 0 0 - .6 3 1.34
41 +1 +1 -3 -1 +3 + .2 6 2.28
42 -1 -1 +2 +1 -2 - .2 4 1.64
43
44
0
+2
+3
+1
+1
-1
-4
-1
-2
-1
- .4
.0
7
3
2.70
1.41
45 +1 0 -1 +1 -1 .0 2 1.00
46 +1 -1 +3 0 0 + .6 4 1.52
47 +2 0 +1 -3 -1 - .2 5 1.92
48 -1 -2 0 +1 +3 + .2 5 1.92
49 -1 -1 0 +1 0 - .2 2 . 84
so 0 -2 0 -4 +3 - .6 7 2.60

Totals -6 . 6 213 85 . 87

43
44 CONTROL CHARTS POR MEASUREMENTS

2. 50 ;L ~ - - - - - - - - - - - - - - - - - - - - - - - -

2.50

10
8.43

5
R

0~-----------------------------------------------

4
3.37

0~-----------------------------------------------
50
10 20 30 40

FIG. 4.1. Control charts for x,- R, and s for 50 samples of 5 each from
Distribution A of Table 4.1. Central lines and control limits set from
~ = 0, crx = 1.715 (Population A).
4.2. Control Charts from Given Standards 45

both tell basically the same story (in different units) about process vari-
ability.

In this experiment we are illustrating the control chart case of "standards

given." The standard values (parameters) for Population A of Table 4.1 are

~ = 0, ax= 1.715. Now with these, what central line and control limits shall

we set for the x chart? By (2.6) we have the central line

standards given. ( 4. 1)

t~ pointed out in Chapter 3, Section 3.6, limits for control charts are set

at ± three standard deviations for the characteristic in question, around the

mean. Thus we need~ ± 3ax. But by (2.27), ax is smaller than ax by a factor

of rn
a-=a/v'l1. ( 4. 2)
X X

Hence we have for ~ ± 3ax

Limits X
- = u ± (3/l:n)a X standards given
r
( 4. 3)

If we define

A = 3//:n l4. 4)

then (4.3) becomes

Limitsx = ~ ± Aax. (4.5)

The quantity A in (4.4) and (4.5) is a control chart constant. Values of

A for n = 2 to 25 are given in Table V, in the back of the book, along

with other control chart constants.

In our experiment, since A = 3/15 = 1.342, we have from ~ 1.715

{ x = o, Limits x =o ± 1.342·(1.715) =± 2.30.

All x- points lie well inside these limits, as was to be expected, since all

data were from the same bowl (population of known~ and ax).
46 CONTROL CHARTS FOR MEASUREMENTS

Next consider the plotting of the ranges, R. The appropriate expressions

involving control chart constants are given at the bottom of Table V, being

tR = d 2crx standard given (4. 6)

( 4. 7)

See Chapter S for further explanation of these expressions. Suffice it to

say that they give E(R) ± 3crR. For our experiment we find

{_R = 2.326·(1.71S) 3.99

LCLR = 0(1.71S) = 0 4.918·(1.71S) = 8.43.

The one largest range among the SO was 8,and hence again we have perfect control

relative to the given standard crx = 1. 71S. (Jl is not involved in an R chart).

When LCLR = 0, we do not bother to plot it on the chart.

Finally for the sample standard deviations, s, we have

standard given ( 4. 8)

Limits (4. 9)
s

Since a 1. 71S these give

ts = .9400·(1.715) 1.61

LCLs = 0·(1.71S) = 0 UCLs = 1.964•(1.715) = 3.37,

Again all sample standard deviations lie inside the control band indicating

compatibility with cr = 1.715.

On all three charts, the probability of the next point lying outside of

the control band is very small, and meanwhile the probability of all of 2S or

SO lying inside is still quite high, if the standards used in setting the

limits are correct.

Now what happens if we have an assignable cause which changes the

population, and our data are taken from the new population? For example, try

Population B of Table 4.1. We show in Figure 4.2 the central line and limits
 4.3. Control Charts from Given Standards 47

-2.5 -2-:-36' - - - - - - - - - - - - - - -

4
p] _________ _

0~------------------------------------
10 20 30 40 50

FIG, 4.2. Control limits for Population A, of Table 4.1 extended from Fig, 4,1,
and compared to x,
R, and s for Populations B and C, respectively samples 1-25,
26-50, Data from Table 4.3,

set from p = o, ax =1.715, extended to the new data, and 25 samples of n = 5

run off. The data are given in Table 4.3. The very first two x points are

above the upper control limit for xs ,

t and in all 25 there are nine above.
However, the first one, all Ill' itself, is sufficient cause for an investiga-
tion. It is evidence of a process change (population change) from p = 0,
48 CONTROL CHARTS FOR MEASUREMENTS

Table 4.3. Random Samples of Five, Numbers 1-25 from

Population B (p=+2, crx=l.715) and 26-50
from Population c (p=O,crx=3.47) of Table 4.1.
Standard
Sample x1 x2 x3 x4 x5 Ave:-age Range Deviation
No. X R s
1 +1 +3 0 +4 +4 +2.4 4 1. 82
2 +3 +2 +4 +3 +1 +2.6 3 1.14
3 +3 +4 +1 -2 +3 +1. 8 6 2.39
4 +1 +1 0 -1 +1 + .4 2 . 89
5 0 +2 -1 +4 +2 +1.4 5 1.95
6 +4 +4 +5 +3 +2 +3. 6 3 1.14
7 +2 +2 +3 +1 +3 +2.2 2 .84
8 +1 +3 +3 +3 -1 +1.8 4 1. 79
9 +5 +2 +2 +1 +1 +2.2 4 1.64
10 +2 +3 +2 +3 +4 +2.8 2 . 84
11 +1 +2 +4 +1 +3 +2.2 3 1. 30
12 +2 0 +1 +4 +2 +1.8 4 1.48
13 +3 +2 +3 0 +2 +2.0 3 1. 22
14 +2 +1 -1 +2 +1 +1.0 3 1. 22
15 -1 +4 +2 -2 +3 +1. 2 6 2.59
16 +2 +4 +2 +4 +2 +2.8 2 1.10
17 +4 +4 +2 +1 +2 +2.6 3 1. 34
18 +2 +2 +1 +1 +2 +1. 6 1 .55
19 +3 +1 +2 +4 +1 +2.2 3 1. 30
20 +5 +1 +2 +1 +3 +2.4 4 1. 67
21 +2 +3 +7 +1 +1 +2.8 6 2.49
22 +1 +1 +1 +2 +2 +1.4 1 .55
23 +3 +3 +3 +2 0 +2.2 3 1. 30
24 0 +1 -1 +1 +2 + .6 3 1.14
25 +4 +3 +4 0 +3 +2.8 4 l. 64
26 +2 +3 +6 +3 +5 +3.8 4 1. 64
27 -1 -2 -5 -3 +5 -1.2 10 3. 77
28 +3 0 0 +5 -2 +1.2 7 2. 77
29 -2 0 +2 +2 +1 + .6 4 1.67
30 +1 +6 +1 0 -1 +1.4 7 2.70
31 +1 +6 -1 -2 +3 +1.4 8 3.21
32 -2 +4 -4 -3 +5 .0 9 4.18
33 -1 -3 -1 -1 +3 - .6 6 2.19
34 -5 -3 -2 +1 0 -1.8 6 2. 39
35 +5 +6 +1 +2 -8 +1. 2 14 5.54
36 -1 +6 -3 -1 0 + .2 9 3.42
37 -1 -5 -2 +2 -8 -2.8 10 3.83
38 +1 -6 -3 -5 0 -2.6 7 3.05
39 -1 +1 -6 -5 0 -2.2 7 3.11
40 0 +8 -4 -5 0 - .2 13 5.12
41 -2 -2 -4 -1 -2 -2.2 3 1.10
42 +5 -5 -3 +3 -4 - .8 10 4.49
43 +1 0 0 +8 -2 +1.4 10 3.85
44 -2 +3 0 +1 +3 +1.0 5 2.12
45 +6 +7 +1 +2 -2 +2.8 9 3.70
46 +2 -3 0 +1 -1 - .2 5 1. 92
47 +1 -3 -3 +3 +8 +1.2 11 4.60
48 +2 +6 -1 +1 0 +1.6 7 2.70
49 -3 -2 -2 +3 0 - .8 6 2.39
50 +4 -1 -1 +2 +1 +1.0 5 2.12
4.3. Control Charts Analyzing Past Data 49

ax= 1.715. Since the x chart has the point out, we presume that a shift in
the true process level p has occurred. But as we shall shortly see, an

increase in variability ax can also cause a point out on the x chart, because

it would give more spread to the x's.

Now how are the range points, 1 to 25? They all lie inside the control

limits. This is to be expected, because ax is still at 1.715. The same is

also true of the standard deviations s.

Samples 26-50 in Table 4.3 were drawn from Population C of Table 4.1.

It has p at 0, but twice the variability with ax = 3.47. Looking first at

the R chart, we find that 10 of the 25 ranges lie above UCLR figured from

ax= 1.715. Also the point around which the ranges center has moved up to

about 8. The first signal for action occurred on sample 27.

Next look at the x's for the same samples. We see that on the very

first sample, number 26, a point went above the UCLx. One might erroneously

conclude that this was due to the process level p shifting up. Actually all

we can confidently say is that such a point out on the x chart indicates a

process change (population change). It can be from either a shift in p or an

increase in ax. Here we know it to be due to an increase in ax' because for

Population C, p = 0, ax= 3.47. However, usually a point outside the x band

does signal an assignable cause affecting the level P• rather than increasing

The chart for s's follows similarly to that for R's. Again with the

scale shown the R and s charts practically duplicate each other and certainly

give the same conclusions.

If control limits were to be set from p =+ 2, a= 1.715, which are the

parameters for Population B, the corresponding points would all be in control.

The same is true for the data of Population C.

4.3. Control Charts for Averages, Ranges and Standard Deviations,

Analyzing Past Data. Let us now turn to the case in which we do not have
so CONTROL CHARTS FOR MEASUREMENTS

reasonable standard values to use, and wish to analyze a run of samples. This
is the typical approach in the early analysis of a process. We may already

have data available, or decide to take a series of samples as rapidly as

convenient, and use them to test for process control. The hypothesis we are

testing is that the samples come from a single population. Or in more

applied terms, that the process variation is all due to chance causes, and

that no assignable causes are at work. Take note that a process may be in

perfect control and not meeting specifications at all well. All we are trying

to do is to determine whether the process is running consistently. Then we

compare what it is doing with what we should like it to do. "We first let the

process speak for itself."

Let us again consider the data of Table 4.2. In order to get started

in analyzing a process, a preliminary run of about 25 samples is taken. So

let us use the first 25 samples of Table 4.2 and set central lines and control

limits for x, R and s charts. We use the following formulas for the case

"analysis of past data," supposing we have k samples.

t-X
k
l:i=l x/k = X ( 4 .10)

k
<t:R zi=l R/k R (4.11)

= s-
k ( 4. 12)
= l:i=l s./k
~5 1

Limits -
X
X. -+A 2 ·R or X ± A3s (4.13)

LCLR o3R. UCLR D4R ( 4. 14)

LCL
s
s 3s UCL
s = 845.
(4.15)

wherein we meet some more control chart constants. Note that these all

require data from the observed samples, and make no use of the unknown

parameters or standard values of the population. But (4.10) to (4.15) supply

estimates of (4.1), (4.5)-(4.9), assuming control, for which we are testing.

4.3. Control Charts Analyzing Past Data 51

For the first 25 samples of Table 4.2 we find

X - 4.2/25 = - .17
R. 102/25 = 4.08
s- 40.98/25 = 1.64.

All three of these are relatively quite close to the central line values set

from p and a in the previous section. For the respective control limits,

using Table V:

Limits -
X
.17 ± . 577 (4. 08) = - .17 ± 2.35 = - 2.52, + 2.18 from R.

Limits -
X
- .17 ± 1. 427 (1. 64) = - .17 ± 2.34 - 2.51, + 2.17 from s

LCLR 0 UCLR 2.115(4.08) 8.63

LCL
s =0 UCL
s 2.089(1.64) = 3.43.

If we were to use these limits on the plottings of Fig. 4.1 we would find all

points in control just as before. We would now have a firm base (homogeneity

or in-controlness) to justify estimating the parameters. This is done as

follows:

Est. of )1 = J.l X ( 4.16)

Est. of a a R/d 2 , or ( 4 .17)

X X

Est. of a a s/c4. (4.18)

X X

We then have for the data

)1 - . 17 (vs . )1 = 0)
a 4.08/2.326 1. 75 (vs, a 1.715)
X X
A

aX 1.64/.9400 1. 74 (vs. a 1.715).

X

Finally since ~have control, we can say that nearly all of the individual

measurements x will lie between

52 CONTROL CHARTS FOR MEASUREMENTS

which are for our data

- .17 ± 3(1.75) =- 5.42, + 5.08.

These limits actually contain 100% of Population A. See Table 4.1.

Control lines set from the preliminary run of data may be extended
and new points plotted immediately after each sample is taken. All of the

next 25 samples of Table 4.2 show perfect control. It is good practice to

revise control lines periodically. Suppose we now revise them on the basis

of the first 50 samples

X =- 6.6/50 .13 R = 213/50 = 4.26 -s 85.87/50 = 1.72

Limits
x .13 ± .577(4.26) =- .13 ± 2.46 (from R)

Limits -
X
.13 ± 1.427(1. 72) =- .13 ± 2.45 (from s)

LCLR 0 UCLR 2.115(4.26) 9.01

LCLs 0 UCLs 2,089(1.72) 3.59.

All 50 samples showed good control relative to these limits hence we can, extend
them without revision to watch as new data come in. Also we could again esti-

mate cr :
X

cr R/d 2 4.26/2.326 1. 83
X

cr
X
= s/c4 1.72/.9400 1. 83.

(Instead of being closer to crx than the earlier estimates, as we might have

expected, these prove to be a bit further away, but still not bad.)
As a further example let us show the calculations for the x, R data of

Table 3.1. For x's we code from 4710 (.0001 11 ) and find

x 4710 + (78.4/20) = 4710 + 3.92 = 4713.92 (in .0001")

R 63/20 = 3.15 (in .0001").

4.4. Comparison of the Two Charts for Variability 53

Then for these samples of n = 5, the above are the center lines and

Limits X- X ± A2R = 4713.92 ± .577(3.15) 4712.10, 4715.74

2.115(3.15) = 6.66.

Since the x- chart is not in control we do not have but just a single

population, and hence we cannot set meaningful limits on the individual

x's. Since the R chart was in control, however, we can estimate ox by

0
X
R/d 2 = 3.15/2.326 = 1.35 (in .0001 11 ),

Thus if we can maintain control at the nominal (middle of the specification

range) . 4 710", we can meet

.4710" ± 3(.000135") = .47059", .47141".

Since these are so comfortably inside of the specification limits for x 1 s,

that is, . 4 700" to . 4 720", we can afford to let p move around somewhat,

perhaps permitting some wear in the cutting tool without resetting the

process.

4.4. Comparison of the Two Charts for Variability. Throughout most

of the two previous sections we carried both variability charts: Rands.

In actual practice this is never done; we always settle on one or the other.

For small samples up to, say, n =8 or 10, the two charts are almost exactly

equivalent. For samples of three or more the chart for s's is always just

a bit more reliable, but above 10 the range begins to lose out to the standard

deviation rather rapidly*. There are three uses of R's or s's we may study:

*One way to slow this deterioration is to use R for sub-samples instead of R

for all. For example, if n = 40, use five ranges for sub-samples of n = 8,
and use R of the five.
54 CONTROL CHARTS FOR MEASUREMENTS

1. Control charts for within-sample variability.

2. Use of R or 5 to help in setting x limits.
3. If the variability chart shows good control, to estimate cr
X

for the process by ax= R/d 2 or ax= s/c 4 .

As we have been seeing, R1 s and s's for the three purposes are almost exactly

equivalent. The author has seen dozens of such experiments, and results are

about this close every time.

Now which shall we use? If work is to be done by hand or by a desk

calculator ranges are much easier, so we use them. But if a digital computer

or automatic line calculator is to be used then we prefer to uses, because

it is a bit more accurate, and s is more easily programmed than R.

A final comment is that it is possible to use a control chart for s 2

instead of s. There are two objections to this: (a) When analyzing past

data we would need to use~. to test for homogeneity. But a single large

s
2
will have a more marked effect on s
2 than will the same samples on s.
Thus the center line is more distorted, if there is non-homogeneity and with

the inflated :Z we may miss the warning. (b) The distribution of s 2 is far

more unsymmetrical than that for s.

4.5. Comparison of a Process with Specifications. In general measurable

characteristics of all types such as dimensions, weights, surface finish,

strength, chemical composition, impurities, physical properties or electrical

characteristics, such as we might make x and R charts for, are subject to

specifications. These specifications are most often in the form of maximum

and minimum limits (U and L) for individuals x. Or more recently they might

be set by a maximum ox and limits on ~. See also (2] for another modern

approach. But for the purposes of this section let us be concerned with

specificationsL and U set for individual measurements x. All or at least

some acceptably high percentage are to lie between L and U.

4.5. Comparison of a Process with Specifications 55

There are a number of complications and pitfalls in this important

comparison. All it takes is clear thinking, but that is asking quite a

bit! The first is that in general specifications are written for individual

x 1 s, not for averages x's. (Exceptions do occur such as the content weight

for packages where the control may be on average content weights.) Then

next we have cases where the process is in control or not in control. Thus

there is the question of what process adjustments are feasible to make.

As a start let us say that there are four cases which occur and which

must be treated quite differently. Failure to distinguish between these cases

costs U.S. industry tens of millions of dollars annually at the very least,

probably much more. The cases are:

1. Process in control, and safely meeting specifications.

2. Process not in control, but "safely" meeting specifications.

3. Process in control, but not safely meeting specifications.

4. Process not in control, and not safely meeting specifications.

What approaches shall we use in the respective cases? Case l is the

happiest of the four? In this case we are doing well. We may, however,

give some consideration to (a) decreasing the frequency of taking samples

to save time and money, and/or (b) running at the most economical level for

p to save material or time. Case 2 is still satisfactory. Note the quotation

marks on "safely." The reason for them is that the lack of control is a

danger sign. One of the assignable causes present may suddenly come in much

more strongly and cause pieces out of specifications.

Case 3 is one in which the process is stable but its capabilities are

in question. Since we have control, we are justified in estimating the

limits for individual x's by

(4.19)

Since we are not meeting specifications, one or the other or both of these
56 CONTROL CHARTS FOR MEASUREMENTS

natural limits are outside of the range L to U. Perhaps the trouble is

entirely from the process level being off-center, and we can adjust ~·

Otherwise we are in trouble. Just "tampering" with the process wi 11 not

help. It will take some fundamental process change.

Finally in Case 4, the first step is to try to get the process into

control, seeking the assignable causes, meanwhile trying to get ~ to run

closer to the "nominal" (L + U)/2. Large amounts of dollars, yen, ruples,

pounds and francs are thrown away through not treating cases 3 and 4 differ-

ently. The common error in Case 3 is to try to reset frequently on insuffi-

cient evidence when the process spread 6&, exceeds the specified tolerance

U - L. The error in Case 4 is to fail to seek out the assignable causes so

as to gain the best possible performance out of the process.

A very common error, especially for the beginner, is to compare control

limits for x 1 s with ?P.ecification ~imits for x 1 s. If the process is in

control and the control limits for x, such as, x ± A2R lie between L and U,

this does not say we are meeting specifications for x's. All it does ~

is that the xJs are inside the specifications for x's. But where are the x 1 s

in such a case? They may well be outside L to U. This is because x~s spread

more widely than do x 1 s, in fact /i1 times as much. Why? See (2. 27).

Let us take an example from industry. In caps for sealing food within

glass jars, "gasket space" is the distance from top of rubber to top of cap.

For samples of n = 5, x= .1204", R = .0154". Among 24 samples only one

x point was out of control (which should be investigated) and noR points.

So control was quite good. For xs

1 we have

Limits x = x ± A2R = .1204" ± .577(.0154") = .1115", .1293".

Now suppose specifications for x• s are .1100" to .1300". The unwary might

conclude that the specifications are being met. But assuming control, all

we can say is that the great majority of averages x are between L = .1100"
 4.5. Comparison of a Process with Specifications 57

and U .1300". What about the x•s? For these we must estimate ax. This

is done by

.0154"/2.326 .0066".

Then limits for xJs set by the process are

x ± 3a X
.1204" ± 3(.0066") = .1204" ± .0198" = .1006", .1402 11

These are well outside of the specifications .1100" to .1300". See Figure

4.3 which pictures the relation. To estimate the percentage of x's outside

specifications, we may use (2.24)

u -
z =-A--=
x .1300" -- .1204"
1.45 Probab. above .073
ax .0066"

. 1400" UCL
X

~a
X
.1300"=U

38-
x

LCL-=.1115
X
.llOO'=L

.1000" LCL
X

FIG. 4.3. Comparison of an in-control process with specifications, x

= ,1204",
R = ,0154". Distributions and three sigma limits for x's and x's are shown.
Specifications for x' s are , 1100 11 to , 1300 11 , Although practically all x 1 s will be
within these limits, the shaded areas on the x distributions show the proportion out,
 58 CONTROL CHARTS FOR MEASUREMENTS

L - x .1100 11 .1204 11
z =- -=
-

.0066" - 1.58 Probab. below .057

ax Proportion out .130

Thus about 13% will be outside of these specifications.

In further discussion of process vs. specifications, consider Figure 4.4,
in which we assume the process in good statistical control with distributions
of x's as shown, estimated from x and R.

Liil
I raJ I

tA
I rbJ I

L (f) U

FIG. 4.4. Six cases of process distributions in relation to specification, Normal

distribution of x' s. Reproduced with permission from I. W. Burr, "Engineering
Statistics and Quality Control", McGraw-Hill, New York, 1953, p. 119.
4.5. Comparison of a Process with Specifications 59

In case (a) the process is well able to meet the specifications,

because the distribution is perfectly centered and the natural spread
of the process, namely 6crX is a fair amount less than the "tolerance"

U-L. In fact it is possible to let the process mean drift somewhat and

we shall still be meeting the specifications. There is some latitude.

In case (b), however, the natural process limits p± 3ax are just barely

within the range U to L, hence close control over p needs to be maintained.

In case (c), the shaded area shows the proportion of pieces above U. The

width of the distribution would be all right if p were moved down to the

nominal. Another more difficult possibility would be to reduce greatly

the process variability while maintaining the same average ~·

In case (d) the process spread 6cr X is much less than the specified

tolerance, thus permitting us to run close to the lower specification with

adequate protection against pieces being below the lower specification L.

We might want to use this approach if we were particularly concerned with a

minimum specification for strength, or for content weight in a container.

We also might wish to set a lathe working on an outside diameter at such a

level initially, so that, as the tool wears and diameters tend to increase,

we can let ~ increase to a m~ximum extent. In this way we can get a maximum

length of time before retooling or resetting.

In case (e), although the average is right on the nominal, there is still

a substantial proportion of pieces outside of specifications in both directions.

The only solution is a cut in process variability (unless the specifications

can be relaxed). Since we have assumed good control, we shall need to make

some rather drastic change in the process, such as going to another type of

machine, if we are to cut the variability sufficiently.

In case (f) we are not making any appreciable number of pieces below L

but have many above U. Better centering would help somewhat in reducing the

total percentage out of specifications. But it is possible, for example on

60 CONTROL CHARTS FOR MEASUREMENTS

an outside diameter, that those pieces below L are scrap, whereas those above

U can be reworked. Thus in (f) when we balance costs of inspection, reworking

and scrapping we might find that the p value shown minimizes the total cost.

This case is the old expensive one of sort 100 per cent, rework, sort again

100 per cent, and rework, etc. The only real solution is to make a radical

change in the process, or else to see whether the tolerance can be increased.

A final warning is that unless you have these concepts clear, one can

hardly trust you to practice statistical quality control in a plant. So

reread this section as often as necessary.

4.6. Continuing the Charts. In the initial stages of a control-chart

application, we are primarily concerned with letting the process do the talking,

that is, we collect the data, plot them, make the calculations, draw the lines,

and interpret the results. As a consequence of indicated lack of control, we

seek out the assignable causes responsible and may possibly find one or more

from the preliminary run of data. In any case we face the problem of contin-

uing the chart. Usually we merely extend the center line and control limits

on both the i and R (or s) charts. There arises the question,however, as to

whether to include the data corresponding to points which lay outside the

control band. Data produced while assignable causes were operative should be

eliminated and x and R revised if both of the following conditions are met:
(a) the assignable cause for such performance was found, and (b) it was

eliminated. Now if the assignable cause behind the unusual performance was

not found or, having been found, nothing has been done to remove it, .then

such data are still as typical of the process as any other and should be

retained. In fact whenever any significant change is made in the process,

all data produced before the change are no longer typical of the revised

process and should be discarded.

4.7. Illustrative Examples 61

Periodically the center lines and control limits should be revised,

perhaps after every 25 samples, and in any case after a process change, as

soon as 20 or 25 samples under the new conditions have accumulated. In this

way up-to-date, meaningful control lines are maintained. Periodically, also,

comparison with specifications should be made so that when the process becomes

stabilized with x ± 3R/d 2 lying within specifications we can substitute the

control chart for 100 per cent sorting.

After a good period of production in control at satisfactory quality

levels, it may well be possible to decrease inspection by lengthening the

time between samples. Then any out-of-control points or undesirable trends

are cause for taking more frequent samples and for investigation of the

reason for the off performance.

Many of the points of this and previous sections will be illustrated in

the following section.

4.7. Illustrative Examples. Example 1. Weaver (3] gives an inter-

esting example of the use of control charts in the steelmaking.industry. The

problem concerns the weights of ingots and the per cent of yield in rolling.

Molten metal from open-hearth furnaces is poured into ingot molds, which are

removed after a sufficient period of cooling. After being reheated, the ingots

are then rolled into bars, perhaps four, each 25ft long and 5 in,square. Each

such bar might then be rolled into six billets, 2 in. square and 2.4 ft

long. There is some unavoidable loss, but unless the weights of the

ingots are well controlled, there may be a considerable loss in the

process.
With a view to improving the yield, Weaver used control charts to

study the weights of the ingots for a certain type of steel. In order to

avoid as much of the multiple cut loss as possible, the ingot had been

designed to weigh 5,300 lb. Figure 4.5 shows some typical data for April
62 CONTROL CHARTS FOR MEASUREMENTS

1945, each point representing the weights of four ingots chosen at random
from a single heat from an open-hearth furnace. Only a portion of the month's
data is shown, but the center lines and control limits are for the whole month.
It is immediately apparent that the average weight, 5,620 lb, is 320 lb
above the desired 5,300 lb. Moreover, the average range was 190 lb, and one

range was over 440 lb, showing a large variation among the weights of ingots
within a single heat of steel. The charts showed lack of control and there-
fore gave promise that the situation could be improved if the assignable

causes could be found. Weaver continues (page 8):

"It was obvious that ingot weights were much too erratic and all too

heavy for the intended purpose. Naturally, the controlling factor in the
ingot weight must be the volume of the mold into which it is poured. These
molds were being obt.ained from two suppliers, and a check of their dimensions
showed that while one supplier was doing much better than the other, both
were making molds oversize in dimensions. Mold suppliers were contacted,

specifications and tolerances discussed, and three months later the situation

had improved as shown in [the middle of Figure 4.5].

"The average overweight had decreased from 320 pounds to 190 pounds,
the variation from lot to lot was reduced, and the variation within a given

lot was now only a 134 pound average [that is, R] compared to the original

190 pounds. We set our sights for a 5300 pound ingot with not over 100 pound
average variation from one to another, made further corrections in the process,
talked further with our mold suppliers, and by April, 1946, the results were
remarkable. [See the last part of Figure 4.5.]
"Our ingots now averaged 5296 pounds and had an average variation one to
another of only 82 pounds. All this consumed just one year's time, but it was

well worth the effort. ·This brings us to an extremely important part of any

quality control program: the evaluation of results, and, of course, the expres-
 4.7. Illustrative Examples 63

•
-;,-. -.
__ ..,!.~~--~..!
•
___._
------~-
•••• ••••• ••
-.-----..&-
500
400
R 300
•• •
-----'---
200 1=----::.:----r_.__ • • •
• ----;.--·-;:-
100
• • • • . • • • I -,au -
0 •
Aprii,IIJ45 April,1946

FIG. 4.5. Control charts for x

and R for weights of steel ingots, each
sample point containing four from a single heat. Progress is shown
toward the desired weight of 5300 lb and toward better control. Limits
set from data from entire month. Reproduced with permission from I. W.
Burr, "Engineering Statistics and Quality Control", McGraw-Hill, New York,
1953, p. 122.

sion which registers most readily with anyone, be he workman, foreman, super-

intendent, or president, is dollars. It was a simple matter to calculate the

savings in pounds per ingot and, from the number of ingots poured during the
year, the total saving in tons. The value of a ton of steel at this point

being well known, the savings for one year could be evaluated. This particular

application turned up a saving of slightly over $175,000.

"The savings from this improvement meant improved yield due to decreased

loss, and actually put more money in the pockets of the foremen participating

in the incentive bonus plan, for yield is one of the major items affecting this

plan. How much trouble do you suppose we had on the next project with these

foremen? Did we receive cooperation? Were they enthusiastic? No answer

required!"

The control charts helped in this work by forcefully picturing the high

average weight and the large average variability within ingots of a single lot.
64 CONTROL CHARTS FOR MEASUREMENTS

Also, the erratic performance as shown by the lack of control was a clear
indication that improvement was possible. The charts helped check improvement

during the progress of the work and gave an excellent idea of the results

obtained, as seen in Fig. 4.5.

Example 2. Our second example is concerned with the deflector for a

rear bearing. Figure 4.6 pictures the piece in question. The characteristic

causing the trouble was the concentricity (distance between centers) of the

9.254-in. outside diameter in relation to the 6.085-in. inside diameter. The

upper specification for this concentricity was .0020 in. In December almost

80 per cent of the pieces exceeded this upper specification. Accordingly,

average and range charts were run, revealing the condition shown in the

first part of Figure 4.7. Data are given in Table 4.4. The one point out

of control on the x chart was some indication of an assignable cause of trouble,

but since such a high percentage of defective pieces lay above the specifica-

·z: )
t<---- ------9.154"-- ---- --->j

FIG. 4.6. Deflector for a rear bearing. The dimension in question

is the concentricity of the 9.254" outside diameter in relation to
the 6.085" inside diameter. The upper specification is .0020".
Reproduced with permission from I. W. Burr, "Engineering Statistics
and Quality Control", McGraw-Hill, New York, 1953, p. 123.
 4.7. Illustrative Examples 65

o"
Before correction Dec.8 After correction Feb.4

0.004"

R 0.002''

0" fniiifrtt,trr
FIG. 4.7. Control charts for x- and R for the concentricity of the
deflector for a rear bearing, shown in Figure 4.6. The effect of
process changes is clearly shown in the two sets of charts and the
two distributions of individual measurements x. Data are given in
Table 4.4. Reproduced with permission from I. W. Burr, "Engineering
Statistics and Quality Control", McGraw-Hill, New York, 1953, p. 123.

tion, rather drastic steps were obviously necessary. Additional x and R chart

studies were made on the three operations affecting the concentricity, and it

was found that each of the operations contributed to the final condition.

Investigation determined the assignable causes, and corrective action

was instituted. The following steps were taken:

1. The machine repair department reground the pot fixture used for

locating on the 9.254-in. pilot OD (outside diameter).

2. A new type of holding fixture was designed and built for locating on

the rough-bore ID (inside diameter).

3. A compressed-air method was devised properly to remove chips during

the machining of the serrations in the 6.085-in. ID. It also served

as a coolant.

4. A new operator was assigned to the machine which cut the serrations

and maintained the 6.085-in. ID.

 66 CONTROL CHARTS FOR MEASUREMENTS

Table 4.4 Deflector - Rear Bearing, Concentricity of Diameters Shown in

Figure 4.6. Specification: .0020 In. (Maximum). Sample Size Four,

Before correction, Dec. 8, After correction, Feb. 4,

Average x, in. Range R, in. Average x, in. Range R, in.

.0018 .0012 .0009 .0014

.0039 .0033 .0013 .0009
.0024 .0025 .0014 .0009
.0027 .0015 .0018 .0013
. 0022 .0009 .0011 .0009
.0024 .0019 .0012 .0012
.0021 . 0011 .0010 .0014
.0032 .0004 .0012 .0006
.0029 .0006 .0006 .0008
. 0034 .0007 .0007 .0008
.0032 .0012 .0012 .0004
.0030 .0025 .0014 .0008
. 0028 .0019 .0010 .0005
.0022 .0006 ....QQ11_ .0005
.0025 .0032 .0159 .0124
.0019 .0015
.0026 .0012
.0022 .0022
..J2Qll .0015
. 0495 ~

X . 00261 in. R .00157 in. X .00114 in . R .00089 in.

UCL .00375 in. UCL .00358 in. UCL .00179 in. UCL .00203 in.
LCL .00147 in. LCL 0 LCL .00049 in, LCL =0

After the corrective action was accomplished, the process was restudied in

February, and the results are shown in the second part of Figure 4.7.

Subsequent investigation and study indicated that the process was in an

even better state of control and that only occasionally was there a piece

beyond the specification.

The two frequency graphs shown in Figure 4.7 indicate the distribution

of individual measurements in the respective cases. The first one is drawn

as a normal curve. Determining the percentage of pieces above the upper

specification is rather hazardous because of the lack of control. Treating

the case as though it were in control and from a normal population, however,

we first find ox R/d 2 = .00157/2.059 = .000763, then

 4.7. Illustrative Examples 67

specification - x .0020 in. - .00261 in. - .80

z
ax .000763 in.

From Table I we then find that about 79 per cent of the pieces can be expec-

ted to lie above .0020 in.

For the data after correction we cannot very well assume normality be-

cause the average is too close to zero, below which a concentricity cannot go.

Hence a skewed curve has been drawn and the percentage above .0020 in. estimated

as 4 per cent.

It may also be mentioned that we are showing in the graphs of these illus-

trative examples a number of different ways of plotting the control-chart

points. In Figure 4.7 two different kinds of point designation for the two

charts were used. The crosses for the average points and the vertical bars

with a large dot for the ranges contrast well so that, if used consistently in

a company, one can always tell at the first glance whether he is looking at an

x or an R chart. In cases where we have points so far from the control band

that the two charts tend to overlap, it is convenient to have two different

kinds of points plotted. The vertical bars are rather effective for ranges

because the length of the bar shows the distance between the smallest and

largest measurements in the sample.

Example 3. The third example is on the weight of charge of an insecti-

cide dispenser. Inventory checks showed that about $14,000 worth of fluid

was being given away free each month because of overfill of the dispensers.

Such average overfill may, however, have been necessary in order safely to

meet the minimum specification requirement. Upon his return from a short

course in statistical quality control, an engineer was given the problem to

work on. His first SO samples of data are shown in Table 4.5. The data

for Dec. 13 and Dec. 14 constituted the preliminary run. From them

x = 463.5 and R = 25.8 g

 68 CONTROL CHARTS FOR MEASUREMENTS

Table 4.5. Charge Weight of Insecticide Disperser, in Grams

Specifications: 454 ± 27g
Sample Observed charge Ave~age Range
No. Date weights, g. X R Remarks
1 Dec. 13 476 478 473 459 471.5 19
2 485 454 456 454 462.2 31
3 451 452 458 473 458.5 22
4 465 492 482 467 476.5 27
5 469 461 452 465 461.8 17
6 459 485 447 460 462.8 38
7 450 463 488 455 464.0 38
8 Lost 478 464 441 461.0 37 Sample of 3
9 456 458 439 448 450.2 19
10 459 462 495 500 479.0 41
11 Dec. 14 443 453 457 458 452.8 15
12 470 450 478 471 467.2 28
13 457 456 460 457 457.5 4
14 434 424 428 438 431.0 14
15 460 444 450 463 454.2 19
16 467 476 485 474 475.5 18
17 471 469 487 476 475.8 18
18 473 452 449 449 455.8 24
19 477 511 495 508 497.8 34
20 458 437 452 447 448.5 21
21 427 443 457 485 453.0 58
22 491 463 466 459 469.8 32
23 471 472 472 481 474.0 10
24 443 460 462 479 461.0 36
25 461 476 478 454 467.2 24 End of preliminary
data

The specifications were an average of 454 g (1 lb) and a minimum of 427 g.

A shop bogey of 481 g was also imposed as a maximum. Even though the average
charge weight was considerably above 454 g, there were some not meeting the
lower specification. Among those tested, there was a 424 in sample 14, and

doubtless more among those not tested. The control chart (Fig. 4.8) shows
two points out of control on the x chart and one point practically on the

upper control limit of the R chart. These signs of instability in the

process are indications that there are assignable causes present which if

found and removed will enable closer control of the weights. Investigations
4.7. Illustrative Examples 69

Table 4.5 (continued)

26 Dec. 17 450 441 444 443 444.5 9

27 454 451 455 460 455.0 9
28 456 463 Lost 445 454.7 18 Sample of 3
29 447 446 431 433 439.2 16
30 447 443 438 453 445.2 15
31 440 454 459 470 455.8 30
32 480 472 475 472 474.8 8
33 449 451 463 453 454.0 14
34 454 455 452 447 452.0 8
35 474 467 477 451 467.2 26
36 459 457 465 444 456.2 21
37 465 475 456 468 466.0 19
38 458 450 451 451 452.5 8
39 447 417 449 445 439.5 32
40 453 442 456 453 451.0 14
41 Dec. 18 471 467 461 455 463.5 16
42 462 454 462 468 461.5 14
43 474 471 471 463 469.8 11
44 461 454 468 452 458.8 16
45 473 453 465 475 466.5 22
46 474 455 486 490 476.2 35
47 466 471 482 474 473.2 16
48 447 454 476 486 465.8 39
49 473 488 482 475 479.5 15
50 460 450 461 445 454.0 16

were started and the center lines and control limits extended. The next 25

samples are shown. There are three points below the lower control limit for

x s,
1 giving some concern as to whether the lower specification is being met.

The ranges appear to have decreased somewhat, which is an encouraging sign.

Some of the assignable causes which the engineer found were obstruc-

tions in the tubes, defective cutoff mechanism, and lack of control in

timing and pressure. Within a month the rate of loss from overfill per

month was down to about $12,000, and inside of three months it was down

to about $2,000 per month, with much safer meeting of the minimum specifi-

cation.
 CONTROL CHARTS FOR MEASUREMENTS

0 0
415 oo 0 0 0
0
0 0
0 0 0 0 0 0
0 0 0
0 0
0 oo 0 0 0
0 0 0 0 00 0 0
450 LCL-=444.1
-----------~-~---------
°
0 0

0
425

UCL= 58.9
---------o--------------
50
0
00 0 0
0 0
0
0 0
R 25~~L-----~o_____o~----~o~----o---o--------------
0 o o o ooo o ooo o o~ o o o
0 0 0 000 0
0 0
00 0 0 0

FIG. 4.8. Control charts for x

and R for charge weights in grams of
insecticide dispensers. Specifications 427 to 481 g. Data from
Table 4.5. Sample size 4. The preliminary period covered the data
of Dec. 13 and 14. Reproduced with permission from I. W. Burr,
"Engineering Statistics and Quality Control", McGraw-Hill, New York,
1953, p. 126.

A few comments on the statistical aspects of the problem are in order.

What average level would be safe for x so as to avoid having any charge

weights below the lower specification? As long as there is lack of control,

no level is really safe! But if we could get the process in control and R
were still as large as 25.8, then we would estimate the standard deviation

for individuals,

25.8 12
2. 059 = .5
4.7. Illustrative Examples 71

Hence, safely to meet the lower specification we should have x at 427 + 3ax

464.5. Figure 4.9 shows that if the process level is maintained at 464.5 g

and control is achieved, there will be very few individual weights x below

427 g. Here "very few" means the proportion of cases below a z of - 3 for a

normal curve, which we have assumed here. Thus the proportion is .0013 by

Table I. Such a proportion is, of course, arbitrary but is a commonly

used figure in industry. If, for example, we could afford to have about .5

per cent of the charge weights below 427 g, then we could let x be somewhat

lower than 464.5 g. From Table I, we find by interpolation that when

z =- 2.575, the area below is .0050. Hence we would have 427 + 2.575(12.5),

or 459.2 g, as the desired process level. It is obvious that control

must be obtained and the process variability decreased in order safely to

meet the lower specification and yet average 454 g, which is as low as one

is allowed to average.

Minimgm
safe x 464.5g

Lower Spec.= 427g

FIG. 4.9. Normal curve distribution of individual charge weights,

placed at such a level as just to meet the lower specification.
72 CONTROL CHARTS FOR MEASUREMENTS

There is also the question of the missing values. The two samples of
three were treated as though the x and R were for samples of four. This

is not technically correct, as such points should have different control

limits and even the center line for the range should be different. Since

neither sample was exceptional, the graphs were not complicated by changing

the lines. The correct way to handle such samples is given in Chapter 7.

Example 4. The final example to be given in this section is the

famous rheostat knob problem, which was extensively used in the wartime

quality control courses under Working and Olds [1], and which is still

being used because it is such an instructive example.

The knob consisted of a metal insert on which a plastic molding was

pressed. The fit of the knob in assembly was largely dependent upon the

dimension from the plane of the back of the plastic disk to the far side

of a pinhole. Both the metal insert, which was bought from an outside

supplier, and the molding operation influenced the dimension. The

engineering specifications for the dimension were .140 ~ .003 in. They

were being inspected 100 per cent against this specification with a go and

not-go gage.

Since the rate of rejection was very high, it was decided that a study

would be made with x and R charts. In order to obtain numerical

measurements, a right-triangular wedge-shaped gage was designed to slip

between the pin and plastic disk, it being so calibrated that the dimension

was read directly from a scale, according to how far in the wedge could be

pushed. Samples of five knobs every hour were drawn, and the gage used

to obtain the data as recorded in Table 4.6. Proceeding as usual

==
X 3.7974 in.
27 = , 14064 .
ln. -R = · 233 27in. = . 00863 ·
ln.

from which the control limits are

4.7. Illustrative Examples 73

x: x ~ A2 R .14064 in.~ .577(.00863 in.)= .1357 in., .1456 in.

R: o3R. = 0 o4R. = 2.115(.00863 in.) = .0183 in.

All points for Mar. 18 and 19 in Fig.4.10 are inside the control bands;

in fact, control is unusually good even if chance variation alone were at

work. The engineer was therefore in a good position to estimate the process

capabilities and compare them with the specifications. From the foregoing

x= .14064 in.
A

a
X
.00371 in.

so that the "natural specifications" for individual knob dimensions are

.14064 in. ~ 3(.00371 in.) = .1295 in. and .1518 in.

Since these lay way outside of the engineering specifications, there was

bound to be a large percentage of out-of-specification knobs. (Actual

count gave 14 below and 28 above among the 135 measured, or a total of

over 30 per cent.)

But this was already known; so what had the control chart done on

the measurements which could not already be told by the go and not-go

gage? There were two things. In the first place, the chart said that

the process was doing about all that could be expected of it without

a rather radical change in the process. Pressure from supervision could

not be expected to help. Above all, it indicated that process adjustments

downward when a knob above .143 in. was encountered and upward when one

below .137 in. was found would not help, but would serve only to give

even more knobs out of specifications. The second thing the control

chart did was to tell what sort of specifications the process could be

expected to meet, that is, .130 to .152 in.

In a circumstance such as this there are only three things to do,

(1) make a more or less fundamental change, such as a redesign or use

of a different machine, (2) ask for a revision of the specifications, or

 74 CONTROL CHARTS FOR MEASUREMENTS

Table 4. 6. Rheostat Knobs, Data on Dimension from Back of Knob to Far

Side of Pinhole in Thousandths of an Inch. Specification: .140 + .003 in.

Measurements on each of
Average Range
five items in series
Hour for for
Date
produced xl x2 x3 x4 x5 sample sample

Mar. 18' 1943 ....... 12-1 140 143 137 134 135 137.8 9
1-2 138 143 143 145 146 143.0 8
2-3 139 133 147 148 139 141.2 15
3-4 143 141 137 138 140 139. 8 6
5-6 142 142 145 135 136 140.0 10

6-7 136 144 143 136 137 139.2 8

7-8 142 147 137 142 138 141.2 10
8-9 143 137 145 137 138 140.0 8
9-10 141 142 147 140 140 142.0 7
10-11 142 137 145 140 132 139.2 13

Mar. 19, 1943 ....... 7-8 137 147 142 137 135 139.6 12
8-9 137 146 142 142 140 141.4 9
9-10 142 142 139 141 142 141.2 3
10-11 137 145 144 137 140 140.6 8
11-12 144 142 143 135 144 141.6 9

12-1 140 132 144 145 141 140.4 13

1-2 137 137 142 143 141 140.0 6
2-3 137 142 142 145 143 141. 8 8
3-4 142 142 143 140 135 140.4 8
4-5 136 142 140 139 137 138.8 6

5-6 142 144 140 138 143 141.4 6

6-7 139 146 143 140 139 141.4 7
7-8 140 145 142 139 137 140.6 8
8-9 134 147 143 141 142 141.4 13
9-10 138 145 141 137 141 140.4 8

10-11 140 145 143 144 138 142.0 7

11-12 145 145 137 138 140 141.0 8

Mar. 20, 1943 ....... 7-8 137 145 139 142 142 141.0 8
8-9 140 142 145 145 144 143.2 5
9-10 141 144 146 14 7 146 144.8 6
10-11 138 144 145 146 141 142.8 8
11-12 137 137 146 141 136 139.4 10

12-l. 142 144 147 137 141 142.2 10

1-2 137 146 139 144 138 140.8 9
2-3 142 140 140 145 141 141.6 5
3-4 141 136 146 141 145 141.8 10
 4.7. Illustrative Examples 75

Table 4.6 (continued)

Measurements on each of
five items in series Average Range
Date Hour for for
produced xl x2 x3 x4 x5 Sample Sample

Mar. 24, 1943 ....... 7-8 143 141 139 137 141 140.2 6
8-9 143 143 139 141 145 142.2 6
9-10 142 135 138 146 142 140.6 11
10-11 141 135 145 141 142 140.8 10
11-12 140 145 141 146 143 143.0 6
12-1 135 142 141 140 141 139.8 7
1-2 140 140 139 141 147 141.4 8
2-3 142 139 139 143 141 140.8 4

Apr. 28, 1943 ....... 5-6 146 148 148 144 147 146.6 4
6-7 146 146 148 149 149 147.6 3
7-8 145 144 146 148 149 146.4 5

8-9 146 144 144 145 149 145.6 5

9-10 146 145 148 147 149 147.0 4
10-11 146 148 147 146 149 147.2 3
Apr. 29, 1943 ....... 5-6 145 149 147 148 149 147.6 4
6-7 146 148 147 147 147 147.0 2

May 11, 1943 ....... 5-6 143 144 144 143 146 144.0 3
6-7 145 143 142 144 148 144.4 6
7-8 144 143 144 149 147 145.4 6
8-9 146 144 143 145 148 145.2 5
9-10 144 143 144 147 146 144.8 4

10-11 145 146 145 143 147 145.2 4

5-6 141 142 141 144 146 142.8 5
6-7 145 147 145 145 147 145.8 2
7-8 145 145 146 146 147 145.8 2
8-9 145 146 145 144 146 145.2 2

(3) sort the pieces 100 per cent. Since (2) would be least expensive, a

number of pieces outside of the specifications were tried in assembly and

seemed to work perfectly satisfactorily. Some were even found further from

.140 in. than any among the data of Mar. 18 and 19, and no trouble was

encountered. Accordingly the engineering department was asked to review

 76 CONTROL CHARTS FOR MEASUREMENTS

----------- ---------------'¥..::::-:~.~

------- -·- --------------------

o''
3/18 3/19 S/ II

FIG, 4,10. Charts on the rheostat knob data given in Table 4.6. The
preliminary run of data 1vas on March 18 and 19, Control limits were extended
from these through the rest of March (center lines solid, control limits
dashed), At the end of March these lines were revised and extended ready
for new data, In April both charts show out of control, the R chart by all
the R's being way below the center line: new process conditions. The new
data are, however, in control relative to their own limits (center lines by
long dashes, limits by dots). Samples of five taken each hour during production,
Reproduced with permission from I, IV, Burr, "Engineering Statistics and Quality
Control", HcGraw-Jiill, New York, 1953, p. 136,

the specifications in the light of these findings. After checking the fit

with various knobs, it was decided to change the specifications to .125 to

.150 in. (Such a large relaxing of tolerances shows courage in the engi-

neering department.)

Although the natural specifications of the process were .1295 to .1518

in., which do not lie wholly within the new specifications, the agreement

was regarded as sufficiently good to justify removing the 100 per cent sorting

and instead to continue the control chart as sufficient evidence for safety.

[Assuming normality we could use z = (.150 in. - .14064 in.)/.00371 in. = 2.52
and estimate that only about .6 per cent would be too high.]
4.7. Illustrative Examples 77

The chart is continued in Fig. 4.10 for Mar. 20 and 24, and good control
continues. At the end of this time production was stopped. Because of the

good control at satisfactory quality performance, it was possible at that

point to set standard values for the process and use these to set control

limits for future production whenever it occurred. Using all the data at

hand, and adopting Greek letters for population or standard values rather than

the circumflex for estimates, we have

~ x= 6 · 2 ~~ 8 in. = .14100 in .

. 362 in.
.00823 in.
44
R = .00823 in. = 00354 .
d2 2.326 · ln.

Using these together with formulas (4.5) to 0.7),

t-X ~ .14100 in. = .14100 in.

Limits x= ~ ± Ao
X
.1362 in., .1458 in.

These are shown in Fig. 4.10 as continuations after the date Mar. 24.

In the meantime work was done with the vendor to try to get him to

reduce the variability of the insert, which apparently was more responsi-

ble for variation in the dimension than was the plastic molding operation.

Results for Apr. 28 and 29 immediately showed out of control on both charts.

All but one of the x points was above the upper control limit, which is an

extremely clear-cut indication of a shift in process level. At the same

time there was an almost equally clear indication of a decrease in process

variability, which was most welcome. All points on the range chart were well

below the old ~ line. It was therefore perfectly obvious that there was

a new process condition. The big question was: Would it meet the specifica-
78 CONTROL CHARTS FOR MEASUREMENTS

tion of .125 to .150 in.? First, to check control, even though there are too
few samples to place great reliance on the results,

¢._ X
1.1750 in.
= .14688 in.
X 8
. 030 in .
<t_R R 8
. 00375 in .

Limits
X
"'
X ± A2R .1447 in., .1490 in.

Limits R" o3R = 0 o4R = .0079 in.

There is good control within the new process; hence the natural specifications

can be determined.

x ± dz3R = .14688 in. ± .00484 in. .1420 in., .1517 in.

There is again some question about the upper specification, with some

danger of a few knobs being beyond the upper specification. This danger

had apparently been remedied by the time the next production, on May 11,

took place. There is a shift downward toward a safer process level, and

the data for that date are in control with respect to themselves. It was

subsequently possible to decrease the amount of inspection to considerably

less often than one sample of five per hour.

4.8. Kinds of Assignable Causes. There are several different types of

assignable causes which may occur in a process or laboratory and which affect
our control charts. It is therefore desirable to summarize them. Basically

an assignable cause brings about a change in the population from which our
measurements are being drawn. The following are some of the different effects

which, in practice, assignable causes can have upon the population: (1) change

in process average or level; (2) change in process variability or spread; (3)

change in curve shape; (4) steady shift in one of the foregoing; (5) cyclic or
erratic shift in (1), (2), or (3); (6) mistakes. We shall be concerned with

the way in which these effects show on a control chart. It must be remembered
4.8. Kinds of Assignable Causes 79

that they do not commonly appear in a perfectly clear-cut manner and may be
quite obscure.

4.8.1. Change in Process Average. We are here concerned with rather

sudden changes in level. Such, for example, would occur from starting up a

machine cold in the morning, a change in source of material such as iron ore,

acid or brass stock, a change in method or machine, a chipped tool, sudden

presence of an impurity, etc, Grant [4] gives an interesting example on

steel castings, in which two x charts (on yield point and per cent elongation)

showed sudden changes simultaneously. The cause of the undesirable shift in

level was finally traced to the source of quench water. As soon as the

original source was restored, the charts came back into control.

Changes in process average can show only on the x chart and not on

the R chart.* (A change in mean~. with constant a, gives same size of

R's.) Probably the commonest way it appears is by a point going out of

control on either the high or the low side of the control band. The other

way for such an effect to make its presence known is by a run of about 10

consecutive points either above or below the center line, even though no

point goes out of the control band. Such a run is ample evidence of a

shift in level.

4.8.2. Changes in Process Variability. Whenever the process variability

increases significantly, we can expect svoner or later to have points above

the upper control limit for ranges. We may also have a run above the center

*This is an invariable rule, unless one draws his sample of n pieces from a

lot which contains pieces made both before and after a potential change such

as a resetting or retooling. We must always guard against this by taking our

sample from pieces produced under as homogeneous conditions as possible. If

a sample contains pieces produced both before and after a change in process

level, we are quite likely to have a high range, perhaps even out of control.
80 CONTROL CHARTS FOR MEASUREMENTS

line for ranges. On the other hand, if the sample size is not over six, we

have a lower control limit of zero, and hence there is no way in which a point

can lie below the lower control limit for ranges. Hence the only way in which

a decrease in process variability can show up is by a run below the center line

for ranges.

However, a change in process variability affects not only the range chart

but also affects the x chart. Thus, as we saw in Sec. 4.2, when the population

variability increases we may get points out of control on either one or

both sides of the band. On the other hand, a decrease in process varia-

bility may appear on the x chart as a series of points all exceptionally

close to the center line. Hence, although evidences of lack of control

on the R chart may be taken as indications of changesin process variability

(or lack of homogeneity in sampling) a point out of control on the x chart

may be from an increase in process variability £! from a shift in level

(or a combination of the two). Such a point should therefore be interpreted

in conjunction with the range chart.

4,8.3. Change in Curve Shape. The x and R control charts do not in general

detect changes in curve shape when such changes are shifts in the shape of

the underlying population. One would need large samples, perhaps of 200 or

more measurements each, for such a study. But what is akin to a change in

curve shape is an assignable cause which may give a considerably higher value

to about every tenth piece. Then if our sample of 5 is taken over the last

100 produced, we are in effect sampling from the old population with a "bump"

added onto it a long way from the center. A sample from this no longer

homogeneous population may give a point out on either the x or the R chart,

or both. If, for example, the mechanism which is designed to segregate the

first and last gaskets from each tube being cut into gaskets becomes defective,

such an effect can occur.

4.8. Kinds of Assignable Causes 81

4.8.4. Steady Shift in One of the Foregoing. Steady progressive changes are
among the commonest phenomena in industr~ and occur quite often in the lab-

oratory. Tool wear, deterioration of abrasives, chemical and physical changes

such as viscosity, furnace condition, and seasonal changes are familiar

examples. Although some may cause trends in variability, most of them cause

trends in process level. flence we usually see evidence of such influences

by slanting trends, either upward or downward, on the x chart. This subject

is taken up in more detail in Chapter 7.

4.8.5 Cyclic and Erratic Shifts. Many times the process level is

affected by cycles. These may arise from daily, weekly, or seasonal effects, or

for psychological, chemical, or mechanical reasons. Many process industries

have cycles due to raw-materials changes. Industrial accidents and hospital

calls often follow cycles. Industries in which there are extensive tooling

changes each week-end may show cycles. The general appearance on the x chart,
where almost all such cyclic phenomena show up, will have the aspect of an
-
ordinary x chart superimposed on a cyclic curve ·rather than along a straight

line for a center. It will in general show many points out of control, at

both the tops and bottoms of the cycles. A stimulating example in the machine

shop is given by McCoun [5]. One x chart showed beautiful control at

apparently very satisfactory quality, but it was found that there were many

pieces out of specification and far away from where the in-control chart

predicted. The measurements on each of 100 consecutive flywheels (outside

diameters) showed a perfectly clear-cut cycle. It was then found that all

the previous samples for the in-control x chart had been taken near the

bottom of each cycle. The period between consecutive samples had happened to

be precisely in step with this cycle. Hence the x chart was completely mis-

leading. This example points up the value which may accrue by taking a series

of consecutive individual measurements and the importance of not taking one's

samples of n pieces at too rigid and uniform a time interval.

82 CONTROL CHARTS FOR MEASUREMENTS

If the cycle is long enough in time so that we get about 10 or more xs

1

per cycle, the cyclic character will show up within two or three cycles, but

if we get only two or three samples per cycle, or only one as in the case

just discussed, then the cyclic appearance may not show up. If a cyclic

effect is to be expected for any reason, the solution is to take more fre-

quent samples or to measure every individual.

Assignable causes which act in an erratic manner are common in indus-

trial processes and research. They may be operative for several consecu-

tive samples or on only one and may affect either the x chart or the R chart,

or both. In any case they cause a temporary change in the population, and

of course the greater the change, i.e., the stronger the assignable cause,

the greater the probability of an indication on the charts.

One of the commonest types of erratic assignable causes is that of too

frequent resetting of a machine or process. Often, changes of setting are

done on insufficient evidence, perhaps on only one piece or two. When this

is the case, the adjustment is frequently in the wrong direction and seldom

by the right amount. All that is accomplished by such resetting is that the

product output is more variable than it otherwise would be. For example, in

one large company, as an experiment, a process whose level had been adjusted

68 times in one 8-hr shift was let run for 8 hr without any adjustment at all.

(The engineering department had agreed to pay for the losses in out-of-

specification pieces.) The result was less scrap produced than previously,

when all the frequent adjustments had been made!

4.8.6. Mistakes. These are just plain blunders, in using the equipment or

measuring devices, misrecordings, using wrong formulas, misplotting, and

errors in calculation. They are unpredictable, they should never occur, but

unfortunately they do. They must be guarded against with every possible

precaution and check. It can be very embarrassing for a point out of the
4.9. When to Set Standard Values 83

control band to turn out to be a mistake in calculation by the quality-

control man!

There is a valuable article by Olmstead [6], in which he discusses

the various statistical tools (control charts and others) which may be used

to investigate the different types of assignable causes.

4.9. When to Set Standard Values -Process Capability. In Example 4

of Section 4.6, standard values for the process were set, for the purpose

of watching future production. It was possible to set standard values only

because two conditions were met, that is, (1) there was a sizable history of

in-control production, which implies, first, that we are justified in assuming

that the data come from only one population and, second, that we have a

reasonably adequate amount of data from which to estimate the population

characteristics; and (2) the performance summarized by those population

characteristics is satisfactory with respect to the specifications.

The standard or population values p and a are set from all the data
X

at hand which were used in x and R, from which

x = p
R = ax or
5
- = ax
d2 c
4
The "natural process limits", ).1 ±)ax, set in this manner ~ the process,

must lie within the engineering specification limits L and U, or at least

not significantly outside. However, when such natural process limits

~ ± 3ax lie about on or barely inside the range L to U, then the process

average p must be tightly controlled to the nominal (L + U)/2. If such is

the actual situation, then we can use p and ax in (4.1), (4.5) and (4.6),

(4.7) or (4.8), (4.9) to set control lines for future production. These

are exactly the same as would be obtained from the case "analysis of past

dat~' using (4.10) to (4.15).

84 CONTROL CHARTS FOR MEASUREMENTS

Now there is still some controversey as to how we should define

"process capability." This is because each range R or standard deviation

s is calculated from a sample of material produced over a short period of

time during which the process average is likely to be virtually constant.

Thus if there is control on the R or s chart, as we are assuming,

( 4. 20)

On the other hand for the entire production of an eight-hour shift, say,

the standard deviation is likely to be larger than that in (4.20). This

is because of variation in level ~,from imperfect control on the x chart.

Long-term o~ = standard deviation of entire output (4.21)

Depending upon how closely the process average p is controlled by the x chart,
the long-term a~ may be 1.2ox to l.Sox. Therefore, in order to meet specifi-
cation limits L to U fairly comfortably, we should have, say,

U - L > Ba
X

rather than U - L ~ 6o .
X

In the latter case we would use (4.1), (4.5) and (4.6), (4.7) or (4.8),

(4.9) for control lines for future production. But if U - L > Box and we are

not interested in going further in discovering assignable causes on the x

chart, we can use the methods of Section 7. 2, employing "modified limits"

or slanting x limits.
Therefore many industrial people, when considering the capability of a

machine or other process to meet a tolerance of U - L, require of the short-

term a :
X

Ba < U- L (4.22)
x-

or of the anticipated long-term o~:

6o 1 < U - L (4. 23)

x-
4.10. Use of Runs for Out of Controlness 85

Thus many regard 8o x or 6o ~ as the "practical process capability", which must

not exceed the tolerance U - L.

4.10. Use of Runs for Out of Controlness. Points beyond the control

limits are the common way to obtain an indication of an assignable cause.

One other way, fairly often used, is that of a long run of points on just one

side of the centerline of a chart. Obviously a run of but three in succession

on one side of a center line will occur much too often to give a clear signal

of an assignable cause. Equally obvious is that 20 in succession on one side

would be an extremely clear indication of an assignable cause. Where do we

stop between? Let us see. Consider first an x chart. Commonly we assume that

the x 1 s are normally distributed even if the x 1 s are not. Then we would expect

about half of the x's on one side and half on the other (although there might

be a slight chance of an x right on the center line). What is the probability

of the next k all being below the center line? The probability for each would

be l/2, assuming no assignable cause. Thus by (2.19) P(next k lying Qelow x) =

(l/2)k. This is also P(next k lying ~ove x). Each event is to be considered

an indication of an assignable cause. Thus the P(next k all lying on the same

side) = 2(1/2) k = (I/2) k-1 . Now suppose, as some writers suggest, that we use

k = 7 for the number on one side to provide a clear signal. Then P(signal) =
(1/2) 6 = 1/64. Now in 25 consecutive Xpoints there would 19 potential x•s

to start such a run. This would seem to the author to give too great a chance

for such a run to occur with no assignable cause at work. Thus it would rather

greatly increase the risk a of thinking an assignable cause is at work when

none is. Therefore the present author recommends that runs of 8 to 10 on

one side be required for an indication on the x chart.

With R or s charts the situation is a bit different, because, owing to

the lack of symmetry P(R < R) > P(R > R), and similarly for s 1 s. Thus we
86 CONTROL CHARTS FOR MEASUREMENTS

might well use 7 in succession above R and 9 in succession below R. Or it

is possible, as some have done, to use the center line at the median range,

rather than at the mean range R, which would make the case of runs like that

for x's.

One final word is that some authors suggest a whole host of ways of

concluding that an assignable cause was at work. The more such warning

signals one uses (a) the greater the chance we have of finding an assignable

cause when one is at work, that is, the lower the S risk becomes, but (b) the

greater the a risk becomes, and (c) the more complicated the chart study and

interpretation becomes. If enough warning criteria are used practically all

data will be "out of control."

The present author recommends just the use of 8 to 10 (take your choice)

on one side of a center line in addition to use of three-sigma control limits.

4.11. Summary. The formulas for x, R and s charts are here included in

one single outline to show the unity of the formulas, in the hope that the

student by seeing them all together may be able to learn them by heart, for

in so doing he will also understand them better.

I. Purpose: analysis of past data

A. Using averages and ranges we have x, R

l. Averages: x-

a. Center line: x-

b. Limits: x± A2R

2. Ranges: R

a. Center line: R
b. Limits: o3R, o4 R
3. Individuals: x--for comparison with specifications

if process is in control

a. Center line: x

b. Natural limits: x± (3R/d 2 )

4.11. Summary 87

B. - s
Using averages and standard deviations, we have x,

1. Averages: x-

a. Center line: x

b. Limits: x ± A3s
2. Standard deviations: s

a. Center line: s-

b. Limits: s 35, s 45
3. Individuals: x--for comparison with specifications if

process is in control

a. Center line: X

b. Limits: X ± (3s/c 4 )

II. Purpose: control from standards, p, (J

X

A. Using averages and ranges

l. Averages: X

a. Center line: ).l

b. Limits: ).1 ± Aa
X

2. Ranges: R

a. Center line: d2a x

b. Limits: Dlax' D2ax
B. Using averages and standard deviations

1. Averages: x

a. Center line: p

b. Limits: p ± Aax

2. Standard deviations: s

a. Center line: c 4ax

b. Limits: B5ax' B6ax

C. Individuals: x

1. Center line: p
2. Limits: p ± 3ax
88 CONTROL CHARTS FOR MEASUREMENTS

As we have already pointed out in Chapter 3, each of the varying

quantities x, R, s or x forms a distribution if the samples are drawn

from the same population. The formulas in the foregoing outline are all

designed to give, as well as possible, the average of the distribution of

the statistic in question for a center line, and for the control limits

points which are three standard deviations of the statistic in question

above and beiow its average. Thus if the process is in control, which

means that the results behave as though the samples were randomly drawn

from the same (normal) population, than we shall not more than one time

in a hundred get a value of the sample statistic outside of these control

limits. On the other hand if the samples are not all from the same popula-

tion then there is a greater chance of a sample point being outside of the

control band. And if the assignable cause is strong we can expect a point

out in even a short run of samples while this cause is at work.

Thus we have from sampling theory formulas for the following limits or

else estimates of them.

E Cx) ± 3cr-
X

E (R) ± 3crR

E(s) ± 3cr s

E (x) ± 3cr X

In the next chapter more will be said about these distributions and the

control chart constants.

Chapter 8 will be devoted to applications of control charts. Meanwhile

the illustrative examples and the problems will serve to acquaint the reader

with different aspects of the job of applying the methods. It must always

be emphasized that success in the use of control charts depends directly

upon the care used in drawing samples. In so far as possible all measurements
Problems 89

included in a single sample should have arisen under as nearly identical con-
ditions as possible. In this way the variation within the sample reflects

random causes only. Then subsequent samples can be taken under different

conditions, careful records being kept of what those conditions were: time,

personnel, material, process resetting, gage sets, etc. Thus one can hope to

pin down the cause of an out-of-control point through noting what conditions

for this sample were different from those for others which were in control.

The subject of the effect of different kinds of sampling upon control charts

will be taken up in Section 7.5.

PROBLEMS

4.1. Briefly describe a potential application in some field of

interest to you. What might you hope to learn?

4.2. (a). For the data of Table 3.2, make x and R charts. (b). Draw

what conclusions you can. (c). If justified, estimate~ and/or a . State

X
justification. (Pressure to operate hydraulic switch.)

4.3. (a). For the data of Table 3.3, make x and R charts. (b). Draw

what conclusions you can. (c). If justified estimate p and/or ox. State

justification. (Final Mooneys for rubber).

4.4. The following data give the results for the presample Mooneys

[measure of internal viscosity of Government Reserve styrene (GR-S) for

rubber making]. These samples of n=4 correspond respectively to the 24

samples of Table 3.3 and Problem 4.3. (a). Make x and R charts for the

data and draw what conclusions you think justified. (b). To what precision

were the original measurements made? (c). li assigned, correlate these

24 xs
1 with those of Problem 4.3.
90 CONTROL CHARTS FOR MEASUREMENTS

Samples 1-8 Samples 9-16 Samples 17-24

-
X R x R -
X R
33.250 6.5 31.750 2.5 30.250 2.5
33.375 .5 35.125 1.5 29.000 .5
32.500 2.5 34.875 3.0 30.375 2.0
30.250 8.5 33.875 2.0 30.375 1.5
34.125 5.5 34.375 3.0 29.375 2.0
29.500 1.5 32.125 2.0 31.125 1.5
35.000 5.0 31.625 1.5 30.875 4.0
32.000 5.5 31.125 4.0 30.250 4.5
766.500 73.5

4.5. In [7] is given a set of thicknesses of polystyrene films

calculated from wave lengths and spacing of interference fringes between

3.6 and 5.0 microns. Thicknesses in mm x 104.

Film Repeated Measurement

Sample - R s
xl x2 x3 X

1 884 888 882 884.7 6 3.06

2 773 764 770 769.0 9 4.58
3 807 797 821 808.3 24 12.06
4 886 894 897 892.3 11 5.69
5 817 823 813 817.7 10 5.03
6 787 808 788 794.3 21 11.85
7 882 898 894 891.3 16 8.33
5857.6 97 50.60

(a). Make control charts for -x and R. (b). How is control on the

R chart? (c), Would you be justified in estimating cr? Why? If so, make

the estimate. (d). How is control on the x- chart? Is this surprising?

4.6. Same as Problem 4.5, but x and s charts.

4.7. The following data represent six readings for each of 10 groups
of cement-mortar briquettes. [8]

Draw the control charts for s, R and x,computing the limits for
the last in two ways. Can the within-group variabilities be regarded as

homogeneous, i.e., are the Rands charts in control? Can one readily assume

that all groups are from one homogeneous lot?

Problems 91

Group xl x2 x3 x4 x5 x6 X R s
No.
1 390 380 445 360 375 350 383.3 95 33.4
2 578 500 470 520 530 450 508.0 128 45.6
3 530 540 470 560 460 470 505.0 100 43.3
4 623 532 547 600 600 594 582.7 91 35.2
5 596 528 540 562 590 530 557.7 68 30.0
6 345 322 312 358 375 310 337.0 65 26.5
7 488 550 530 568 420 530 514.3 148 53.3
8 625 625 610 615 610 600 614.2 25 9.7
9 722 727 690 700 705 700 707.3 37 14.2
10 800 798 750 724 720 726 753.0 80 37.1

4.8. This problem is taken from an article by Ott [9]:

"During the manufacture of phonograph pick-up assemblies in a plant

of Sonotone Corporation, an electrical characteristic of the completed

assemblies was more variable than desired.

11 In an effort to manufacture the units with greater uniformity, attention

was directed toward a ceramic component of the assembly. These component pieces

were cut in the plant from ceramic strips purchased from an outside vendor;

some 25 individual pieces were obtained from each ceramic strip. The

production foreman and engineer agreed that it was worthwhile to determine how

the variability of assemblies containing pieces cut from different str1ps

would compare with the variability corresponding to different pieces from

a single strip.

''It was decided arbitrarily to cut 7 pieces from each of 11 different

strips. The eleven sets were used to complete eleven assemblies by putting

them through the regular production process ..... The 11 sets of ceramic

pieces were made into completed assemblies by one group of operators in

about an hour's production time. This was done deliberately since the

purpose of the production experiment was to determine possible effects

resulting from different strips of ceramics. If different sets had been

done by different shifts and on different days, it was entirely possible

that unintentional differences in processing might cause significant

92 CONTROL CHARTS FOR MEASUREMENTS

variations which would then be confused with the effects being considered

in this simple but effective experiment .... The following readings of the

electrical characteristics of the final assemblies were then read and

recorded.

Set - R
xl x2 x3 x4 x5 x6 x7 X

1 16.5 17.2 16.6 15.0 14.4 16.5 15.5 16.0 2.8

2 15.7 17.6 16.3 14.6 14.9 15.2 16.1 15.8 3.0
3 17.3 1S.8 16.8 17.2 16.2 16.9 14.9 16.4 2.4
4 16.9 15.8 16.9 16.8 16.6 16.0 16.6 16.5 1.1
5 15.5 16.6 15.9 16.5 16.1 16.2 15.7 16.1 1.1
6 13.5 14.5 16.0 15.9 13.7 15.2 15.9 15.0 2.5
7 16.5 14.3 16.9 14.6 17.5 15.5 16.1 15.9 3.2
8 16.5 16.9 16.8 16.1 16.9 15.0 16.1 16.3 1.9
9 14.5 14.9 15.6 16.8 12.9 16.6 10.9 14.6 5.9
10 16.9 ).6.5 17.1 15.8 15.7 13.0 15.0 15.7 4.1
11 16.5 16.7 16.3 14.0 14.9 15.6 16.8 15.8 2.8

"The range of Sample No. 9 was large and the average was low, essentially

from the influence of the 10.9 reading. However, the production engineer

recognized this low reading as resulting from an electrical short in the

assembly, and advised that this one reading be disregarded."

Show the control chart analysis you would use for these experimental data.

4.9. This problem is concerned with the bearing diameter for a blower

rotor shaft, for which the specification is .9841 in. + .0000 in., -.0002 in.,

.9839 to .9841 in,

"Preliminary study revealed that 35.1 per cent were below the low

specification, also that 2.6 per cent were above the upper limit. [See

first part of table.] Investigation showed that the operator was not

exercising enough caution to maintain this close tolerance. As a result

of the wide spread of individual pieces inspection was performing 100

per cent sorting of the part. In addition to salvage and rework,some 6

per cent were being scrapped.

"It was decided that the operator should be relieved of the responsibility

of backing the wheel away from the work,and an Arnold gage, which automaticall
Problems 93

backs the wheel off when the predetermined diameter is reached, was installed
on the job. The results of using this type of gage are shown in the second

part of the table.

"Inspection was reduced to 5 pieces out of each tote pan of 60. A point

is plotted on the x and R charts. If the points are in control the pan

is accepted; if out of control the pan js sorted 100 per cent."

Draw the charts from the data, and comment on the preceding analysis,

with n = 5. Can you set a standard value for a for the later data? What

standard ~ would you suggest for continuation? What precision of measurement

x was used?

Early Data in .0001" from .9839":

Sample -X R Sample X R Sample -

X R
1 .64 1.7 8 . 36 2.8 15 .56 1.7
2 .10 1.3 9 .so 1.6 16 - .02 2.2
3 -.04 1.9 10 .58 1.0 17 .60 2.0
4 -.36 3.2 11 .78 2.6 18 - .04 2.2
5 .68 1.4 12 .28 1.4 19 - .04 3.8
6 1. 06 1.0 13 . 86 1.2 20 - .54 2.9
7 .38 1.8 14 . 32 2.7 +6.66 40.4

Later Data in .0001" from .9839"

Sample X R Sample X R Sample -

X R
1 1. 42 .3 8 1. 00 .2 15 1. 52 .3
2 1. 22 .5 9 1. 54 .3 16 1.40 .5
3 1. 24 .3 10 1. 28 .3 17 1. 52 .4
4 1. 16 .5 11 1. 36 .2 18 1. 38 .4
5 1. 36 .6 12 1. 54 .1 19 1. 20 .6
6 1. 06 .7 13 1. 32 .5 20 .98 .8
7 1. 58 .2 14 1. 48 .2 26.56 7.9

4.10. The data of Table P4.7 give results for eccentricity (distance

between two center points) for a needle valve. (a) Plot the first 25 xs
1

and R' s ~nd calculate control lines. (b) Comment on results as to control

and meeting of maximum specification. (c) What control lines would you

use for the next 25 samples? (d) Plot them and comment. (e) Revise and
94 CONTROL CHARTS FOR MEASUREMENTS

Table P4.7. Eccentricity Between Conical Point and Pitch Diameter of a

Needle Valve, in .0001 in. (Maximum Specification: .0100 in.) Bore-matic.
Spindle 1.
-
Operator Time Date X R

M.D. 1 P.M. May 3 32 30 30 31 2

M.D. 7 A.M. May 5 37 18 37 31 19
9 so 35 36 40 15
11 57 24 75 52 51
1 P.M. 49 6 24 26 43
M.D. 6 A.M. May 6 67 25 25 39 42
8 52 56 53 54 4
10 18 39 47 35 29
1 P.M. 40 51 51 47 11
M.D. 7 A.M. May 7 31 61 28 40 33
9 15 10 35 20 25
11 27 49 19 32 30
1 P.M. 51 34 40 42 17
H.D. 3 19 32 10 20 22
M.D. 7 A.M. May 8 30 16 so 35 34
9 15 30 so 32 35
11 32 46 29 36 17
1 P.M. 39 19 34 31 20
H.D. 6 42 40 30 37 12
9 70 16 57 48 54
M.D. 7 A.M. May 9 12 19 23 18 11
9 34 14 40 29 26
11 58 36 41 45 22
1 P.M. 7 11 8 9 4
H. D. 3 P.M. May 13 40 12 12 21 28
6 66 58 19 48 47
M.D. 8 A.M. May 14 29 37 74 47 45
10 20 9 15 15 11
1 P.M. 30 38 88 52 58
H.D. 3 83 57 90 77 33
6 40 100 82 74 60
9 24 63 8 32 55
M.D. 7 A.M. May 15 9 41 5 18 36
9 44 80 66 63 36
11 102 10 41 51 92
1 P.M. 43 51 53 49 10
H.D. 6 21 27 49 32 28
M.D. 7 A.M. May 16 18 19 31 23 13
9 7 22 13 14 15
11 43 48 19 37 29
1 P.M. 67 35 5 36 62
A.W.G. 3 35 51 33 40 18
M.D. 7 A.M. May 19 39 35 49 41 14
9 66 18 56 47 48
11 39 13 43 32 30
1 P.M. 14 27 10 17 17
M.D. 7 A.M. May 20 23 71 27 40 48
9 17 24 60 34 43
11 55 31 32 39 24
M.D. 7 A.M. May 26 6 34 39 26 33
Problems 95

suggest ~ and a if reasonable. (f) The distribution of x's is unsymmetrical.

Would this affect the control chart constants used?

4.11. The following data are for the effects of minor alloy additions

on corrosion resistance of aluminum bronze [10]. Ten of 22 elements

were randomly chosen for this problem. Ten observations per sample.

Element Symbol X s
Cobalt Co 31.77 2.25
Nickel Ni 30.96 1.69
Tellurium Te 31.99 .76
Magnesium Mg 31.88 1. 22
Iron Fe 31.62 1.48
Manganese Mn 30.89 1.57
Titanium Ti 31.10 1. 76
Silver Ag 30.13 1. 44
Bismuth Bi 28.44 .97
Selenium Se 29.44 1. 23

Make x and s charts, analyze results and comment.

4.12. Show how one could tell what the sample size is from the

marked control limits and central lines on Figure 4.8.

4.13. For the data in Table 4.3, samples 1-25, without plotting points

find center lines and control limits for xs1 and R's analyzing as past data.

Is process in control? Do the same using standards given: p =2.00,a =1.715,

for Populations B.

4.14. Same as Problem 4.13 but for samples 26-50. (Standards given

for Population C: p = 0, a= 3.47.)

4.15. How would a decrease in ax with no change in ]J show up on

x and R (or s) charts?

4.16. If x limits lie outside one or both specification limits for

x 1 s, do you need to test for control to tell whether you are meeting

specifications? If they lie inside do you need to check control to tell?

96 CONTROL CHARTS FOR MEASUREMENTS

References

1. H. Working and E. G. Olds. "Manual for an Introduction to Statistical

Methods of Quality Control in Industry, Outline of a Course of Lectures

and Exercises," Office of Production Research and Development, War

Production Board, Washington, D. C., 1944.

2. I. W. Burr, Specifying the desired distribution rather than maximum

and minimum Limits. Indust. Quality Control 24 (No. 2), 94-101 (1967).

3. W. R. Weaver, The foreman's view of quality control. Indust. Quality

Control 5 (No. 2), 6-14 (1948).

4. E. L. Grant. "Statistical Quality Control." Mc-Graw Hill, N. Y., 1952.

Example 8, p. 117.

5. V. E. McCoun, The case of the perjured control chart. Indust. Quality

Control 5 (No. 6), 20-23 (1949).

6. P. S. Olmstead, How to detect the type of an assignable cause, Part I,

Clues for particular types of trouble, Indust. Quality Control 9 (No. 3),

32-38 (1952). Also Part II, Procedure when probable cause is unknown,

Indust. Quality Control 9 (No. 4), 22-32 (1953).

7. E. C. Copelin, "Attempts to Establish Performance Tests for Infrared

Spectrophotometers." Ph.D. Thesis, Purdue University, p. 45 (1958).

8. E. S. Pearson, The percentage limits for the distribution of range

in samples from a normal population. Biometrika 24, 404-417 (1932).

9. E. R. Ott, Variables control chart in production research, Practical

Aids Department of Indust. Quality Control 6 (No. 3), 30-31 (1949).

10. R. H. Hoefs, "The Effect of Minor Alloy Additions on the Corrosion

Resistance of Alpha Aluminum Bronze." Ph.D. Thesis, Purdue University

(1953).
 CHAPTER 5

BACKGROUND OF CONTROL CHARTS FOR MEASUREMENTS

5.1. Introduction. This chapter is designed t o explain further t he

control chart constants we have met in Chapter 4; to give some formulas and

some interrelations between them. Such study will strengthen the student's

grasp of the control chart constants and their role .

There are two me thods by which control chart constants may be found for

a given population. We may proceed empirically, through a large collection of

random samples, tabulated and analyzed. Such results are only approximate ,

but this method may be the only one avai l ab le. See Section 5.2.

The second method is to derive the constants by theoretical cons i de r ati ons.

This approach is used in Section 5 .3. It may yield exact or approximate resu lts.

5.2. The Empirical Approach to Sampling Distributions. A method which is

always available is to draw a large number of samples of n x's each with r ep lace-

ment i f a finite population, or by use of tables of random numbers if an

infinite popul ation. Th en one tabulates the distributions of the obtained

sampl e statistics x, R and s, and finds the respective means and standard devi-

ations of the three. These would be estimates of the true values E(i), E(R),

E(s) ; oi , oR, os. Us ing t he obtained estimates as in Section 5.3.1 . we would

obt ai n approximations to the control chart constants for the popul ation of x's

at hand. This approach is necessary if th eory is not available.

The approach was used by Lois Niemann [1]. She drew 4000 samples, each

of n 4, from a strongly unsymmetrical population (constructed by 1000 numb ered

be ads, skewness a 3 = 1.15). Using available known theory on i•s but finding

97
98 BACKGROUND OF CONTROL CHARTS FOR MEASUREMENTS

a she estimated the control chart constants

s
appropriate for samples from this population. They were found to be remark-

ably close to those for a normal population of x's, differing at most by 10%.

The calculations were done on desk calculators, this being before the flowering

of the electronic computer.

The foregoing approach is quite a bit more feasible with a modern elec-

tronic computer. But still one must fight the fact that errors in such

empirical statistical results vary inversely, only as the square root of the

number of values of the statistic which were obtained. The method is useful

where theory is as yet unavailable.

5.3. Control Charts for Samples from Normal Populations of x's. We next

use available normal curve sampling theory to derive the control chart

constants we need.

5.3.1. Formulas for Control Chart Constants. Suppose we have a

population of x's with given ~ and ax. Then whether normal or not we have

E(x) = ~x = ~x and a-X = a X!ln. (5.1)

Then just as in Section 4.2

(5.2)

Limits X
- = ~x- ± 3ax- = ~x ± 3a X!In=~· X ± Aa X , A= 3/ln, (5.3)

This formula for A is correct for ~ population having a finite ax. Check a

couple of cases in Table V.

In order to derive the other control chart constants we assume normality

of x's. If this is assumed then we can evaluate c 4 and c 5 as in Section 5.3.2

to give

E(s) ~ = c 4aX (5.4)

S
(5.5)
5.3. Samples from Normal Populations of x's 99

Note that c 4 is not 1 in Table V, but approaches 1 as n increases. Using

these we have

= c 40 X (5. 6)

Limits (5.7)
s

We define the limits by multipliers of ox:

Limits s = B5ox' B6ox, (5. 8)

from which

Next suppose that Px• ox are unknown. We could not then use (5.3) and

(5.8). But if we have a collection of samples yielding x and s,we assume control

tentatively and use them.

Since s's and therefore s's average C40 '

X
but not 0
X
' s is a biased estimate
of o .
X
For example, if n = 5, 5 averages .940o , that is, it underestimates
X
ox QD the average. But if we divides by .940, then s/.940 will average ox'

and thus provides an unbiased estimate. Hence we take

0 (5 .10)
X

to replace ox in (5.3), (5.6) and (5.7). Thus

Limits x- = ~~rx ± 3~ x;.;n = x± 3s/(c 4 v'ii) (5.11)

Therefore

Verify this relation for n = 4, say, in Table V. Further

What could be more natural than 5 for the center line? Continuing

the B3 , B4 constants being defined as appropriate multipliers for s. Therefore

100 BACKGROUND OF CONTROL CHARTS FOR MEASUREMENTS

(5.15)
Or, using (5.9)
(5.16)

Now for the constants associated with the range and R in particular we
proceed the same way exactly. Assume

Then an unbiased estimate of crx from R is

Using this to replace crx in (5.3) along with Px x we have

(5. 19)

(5. 20)

Now also we have for the case standard crx given

(5.21)

(5.22)

defining D1 , D2 as multipliers of crx for the limits. Thus

Now using ;x = R/d 2 to replace crx in (5.21)-(5.23) yields

4._ R = R

Limits (5.24)
s

So
(5.25)

Or, using (5.23)

5.3. Samples from Normal Populations of x's 101

(5.26)

It is easily seen from the foregoing that the basic constants needed are just

5.3.2.* Derivation for Standard Deviations. If samples are drawn randomly

and independently from a normal population with standard deviation ox a, then

it can readily be proved that

(n - l)s 2 2
X (5.27)
2
a

follows the i distribution with n - l degrees of freedom. A proof may be found

in any mathematical statistics book.

The density function for this x 2 variable is

c/) [(n-1)/2]-le-x 2/2 2

O<x <oo
2(n-l)/ 2 r[(n-l)/2]

The reader may not be familiar with the gamma function in the denominator. Its

purpose here is to make the integral of f(x 2 ) from 0 to oo to be 1. By definition

k-1 e-w dw
r (k) f w k > 0 (5. 28)
0

which depends upon the exponent k for its value. In particular

r (l) J e-w dw 1. ( 5. 29)

0

A convenient recursion relation

r(k+l) = kr(k) k > 0 (5.30)

is obtained by integration by parts:

*Sections designated by an asterisk may be omitted without destroying the

continuity.
102 BACKGROUND OF CONTROL CHARTS FOR MEASUREMENTS

"' k -w k
r (k+ 1) f0 w e dw u = w, dv = e-w dw
k-1 dw, v
du = kw = -e -w
r(k+l) + f kwk-l e-w dw
0

From these two relations we have,taking k as any positive integer n

r(n) = (n-1)1 (5. 31)

Thus the gamma function takes factorial values for positive integers and may

be regarded as an interpolation formula between them.

In the x2 distribution we need r(l/2) which proves to be liT. A sketch

of the proof follows:

r (1/2) = f"' w-1/2 e -w dw > 0

0
2
Let w X ' dw = 2xdx
2
r(l/2) =f 2e -x dx
0
This cannot be evaluated directly, so we use a trick and find its square
"' 2 "' 2
[r(l/2)] 2 = f 2e-x dx f 2e-y dy
0 0
2 2
= 4/ f e-(x +y ) dx dy
0 0
Then transforming to polar coordinates by

x = r sin Q, y = r cos Q, dxdy = r dr dQ, x2 + y 2

the integral over the first quadrant becomes

"' TT/2 2
[f(l/2)] 2 = 4/ f re -r dr dQ
0 0
"' -r 2 TT/2
2/ e 2r dr f dQ
0 0
2 "'
2 [-e -r ] • (71/2) = Tf
0
5.3. Samples from Normal Populations of x's 103

Since r(l/2) > 0

f(l/2) = Iii. (5. 32)

Now for the moments of s, we use the density function of x2 as follows,

using (5.27),

(n-l)s 2
(Cn-l~s 2 JCn-l)/2-l e (n-l~s 2
2o 2 d
f( / ) d/ = _ __;;o'--_,..___,..,.....,.=---------_..;:.o_
2(n-l)/ 2 r[(n-1)/2)

Then, distributing the 2(n-l)/ 2 as needed, the expectation of si is:

(n-1) s 2

s
i(Cn-l)s 21Cn-l)/2-l
2 J e
2i 2
d ...,(n_-_l_,),....s_
J
oo
2o 2i
0 r[(n-1)/2]

2
v';z,;;-
(n-1) s
Now let w s=o·
2i ' fi-T

i (n-l+i)/2-1 2i/2
J
oo
0 w
0 r[(n-l)/2](n-l)i/ 2 dw,

and using (5.28)

oi 2i 12 r[(n-l+i)/2].
(5. 33)
(n-l)i/ 2 r[(n-l)/2]

Taking i 1, yields

E (s)
f"2 r(n/2)
Ps =ovn:r r[Cn-1)/2]

Hence

!2 r(n/2) (5. 34)

c4 =vn:r r[(n-1)/2]

For example if n = 7, using (5.30)-(5.32)

104 BACKGROUND OF CONTROL CHARTS FOR MEASUREMENrs

12 r(7/2)= (5/2) (3/2) (1/2) liT.

c4 /7:1 r(3) 2113

If, however, i 2, (5.33) yields

r/2 r[ (n+l)/2) 2a 2 [(n-1)/2] r[(n-1)/2]

n-1 r[(n-1)/2] n-1 r[(n-1)/2]
2
r;. (5. 35)

(This is true also for non-normal populations.)

Making use of the theorem a2 = E(x 2) - [E(x)] 2 we have

X

2 2 2 2 2 2 2 2
r;
s = E(s )-[E(s)] = a -c 4a = (l-c 4)cr •

Hence

Thus

(5. 36)

We can therefore readily evaluate c 4 and c 5 .

5.3.3.* The Distribution of Range. This general problem requires

approximate integration for all but the very simplest cases, in fact double
or triple integration. Letting the density function for the x's be f(x), the

cumulative or distribution function F(x) is defined by

X
F(x) = P(X < x) = J f(v)dv.
- 0
Now take a sample of n x•s randomly and independently from the population.

Instead of calling the first one x1 , call the smallest x1 , the second smallest

x2 , up to the largest xn = x1 + R, since the range is xn- x 1 . Now the overall

probability, to within higher order infinitesimals, for the random variable x1
to lie in x1 , x 1 + dx 1 , random variable R' to lie in R, R+dR and the other n-2

between is given by the following:

5.3. Samples from Normal Populations of x's 105

(5.37)

The first part says that there are n choices among the n x's for the lowest

x 1 (it to be in the given interval). Then the second part has n-1 choices

for the x to be the largest one and in the interval x 1+R to x 1+R+dR. The

third part uses

and this is the probability of the all other x's lying between x 1 and x 1 + R.

Now to find the density function of R, g(R), we need to integrate (5.37) from

x1 =- 00 to x 1 = oo which eliminates x 1 . Then finally we need

E(R), E(R 2 ) by j Rg(R)dR, jooR 2g(R)dR.

0 0
2 2 2
Then we use aR = E(R ) - [E(R)] . In the case of the normal curve for f(x)

this requires approximate triple integration. However, despite the difficulties

E. S. Pearson [2] had quite full results available for a normal distribution

by 1932, building on earlier work by Tippett [3]. This work, facilitating the

use of ranges instead of standard deviations for small samples, was a great

boon to the users of control charts for measurements and helped a great deal in

their spread to applications.

From such work we have the numerical values of d 2 and d 3 and can find all

the other control chart constants.

Fairly recent work by the author [4], showed that for population distri-

butions departing considerably from the normal curve, the control chart con-

stants are not greatly affected, for small sample sizes (up to n = 10). Hence

we can use the control chart constants of Table V with much confidence.
106 BACKGROUND OF CONTROL CHARTS FOR MEASUREMENTS

PROBLEMS

5.1. Given n = 8, evaluate c 4 , c5 from (5.34), (5.36). Then take d2 and d 3

from Table V. Using these as needed in appropriate formulas verify

the other control chart constants in Table V, for n = 8.

5.2. Same as Problem 5.1. but use n = 9.
5.3. Use a; = E(s 2) - [E(s)] 2 to show that E(s) < a, that is, that c 4 < 1.
(Hint: We know E(s 2) = a2 • What would happen if E(s) =a?)

Refe'rences

1. Lois J. H. Niemann, "An Experiment in Sampling from a Pearson

Type III Distribution." M.S. Thesis, Purdue University (1949).
2. E. S. Pearson, The percentage limits for the distribution of
range in samples from a normal population. Biometrika 24, 404-417

(1932).
3. L. H. C. Tippett, On the extreme individuals and the range of
samples taken from a normal population. Biometrika 17, 364-387 (1925).
4. I. W. Burr. The effect of non-normality on constants for x and R
charts, Indust. Quality Control 23 (No. 11), 563-569 (1967).
 CHAPTER 6
CONTROL CHARTS FOR ATTRIBUTES

6.1. Introduction. This chapter is concerned with control charts

for attribute or countable data . Such data are of two basic kinds, but

both involve "defects", and this can easily be confusing. Let us make

the distinction clear .

Definition [1). A "defect" is an instance of a failure to meet a

requirement impos ed on a unit with respect to a single quality character-

istic. (The term defect thus covers a wide range of possible severit y ; on

the one hand it may be merely a flaw or a detectable devi ati on from some

minimum or maximum limiting value, or on the other, a fault sufficiently

severe to cause an untimely product failure.)

In inspec tion or t esting, each piece or unit is checked to see if it

does or does not contain any defects. Then we may describe the outcome of

the inspection in two distinct ways.

In the first place we may examine each piece and record it as a "good"

one if it contains~ defe cts, or record it as a "defective" one if it con-

tains any defects. In this scoring, every piece is ei ther a good one or a

defective one. Aft~r inspecting n pieces we will have found say d of t hem

to be defectives and n-d of them to be good ones. We emphasize that a

defective piece may have but one defect or several, but it still counts as

just one defective.

On t he other hand we may count and record the number of defects, c,

we find on a singl e piece. Thi s count may be 0,1, 2, .. . Such an approach

107
108 CONTROL CHARTS FOR ATTRIBUTES

of counting defects on a piece becomes especially useful if most of

the pieces contain one or more defects. Then the count tells not only

that the piece is a defective, but how defective it is. Examples of

defect counting might be (a) defects in a complicated electronic or

mechanical assembly, (b) typographical errors on a page, (c) accidents

in a month, (d) flaws on a sheet of paper or metal. We may count the

defects on a single unit or on a sample of n units, recording the count

as c defects. We thus count defects on a field of opportunity for defects

to occur.

Again it is emphasized that a defective piece may have only quite a

minor defect or a major or critical one; it all depends upon the piece in

question and the class of defects being inspected for.

We may always convert a measurable characteristic of a piece to an

attribute by setting limits, say, Land lJ for x. Then if x lies between,

the piece is a good one, or if outside, it is a defective one. But the

converse is not true. For example, an electric light bulb, when tested

either lights or it does not, giving respectively a good or a defective.

But there is no way to put a measurement x on this bulb. It simply works

or it does not.

6.2. Control Charts for Defectives. Let us first consider the two

kinds of charts available for defectives. We shall first take the simplest

case, namely, when the sample size n is constant. Thus we have in each

sample n pieces to inspect for a single type of defect, or for all of the

defects in some class. There might be 20 or more different kinds of defects

in the class. But still each piece is either a defective (contains one or

more defects), or a good one (contains no defects in the class). Then

n number of pieces inspected. (6. 1)

d number of defective pieces in the n (6.2)

6.2. Control Charts for Defectives 109

p d/n = fraction defective in the sample (6. 3)

np number of defective pieces in the n* (6.4)

q 1 - p = fraction effective or good in the sample. (6. 5)

Now also we have a true or population fraction defective for the production

process, for which we use the primed notation (rather than nor some other

Greek letter). This follows the notations in [1].

p' fraction defective in the population (6.6)

q' 1 - p' = fraction effective in the population, (6. 7)

Then we may plot on a control chart successive values of p, giving a

fraction defective chart, or of d = np, giving a chart for g~f~ctives. With

constant n's it makes little difference which we use, but perhaps more people

will understand a chart for p, especially using a percent scale.

Let us recall here the conditions for the binomial distribution to apply,

as given in Section 2.5.3:

1. Two possibilities: good and defective pieces

2. Constant fraction defective p' for each piece

3. Independence of pieces, that is, the occurrence of a defective or a

good one does not change p' or q' for the next one to be defective

or good. Thus the defectives do not tend to "come in bunches,"

These conditions are assumed whenever we make control charts for p or np. If

any of them are violated, the control limits and interpretation thereof do not

apply. Some such counter cases will be mentioned in Section 6·2.5, later on.

6.2.1. Control Charts for Fraction Defective, Standard Given. Let us

first consider the simple case of a p chart, when p' is given and n is

* We have here two notations for the number of defectives: np and d. Charts

for this number are most often called "np charts."

110 CONTROL CHARTS FOR ATTRIBUTES

constant. Then using Section 2.5.3 on the binomial distribution, E(p) = p',
a
p
= (p' (1-p')/n. Therefore we have the center line and three-sigma limits:

<l. p = p' (6.8)

Limits p = p' ~ 3 (p' q' /n. (6.9)
Now how do we know p'? Just as we have explained for measurement charts,
standards can be set when we have a recent period of good control at

satisfactory performance. Also in the case of p charts we may set a standard

p' as a goal to be desired. Or, at the end of a month we may review the

past month with those involved with the process, eliminating some out of

control results or no longer typical data, and set a standard p' for the

next month.

As an example of a p chart with a standard given, consider Table 6.1.

It lists the number of defective carburetors, that is, a defective had one
or more defects, excluding leaks. There were perhaps 20 different possible defects.

Each sample consisted of 100 consecutive carburetors inspected at the

end of the production line. This is in a real sense a sampling of

production conditions obtaining while these 100 were produced. The table
also gives the fraction defective p.

For a long time a goal of 2 percent had been set, but this had seemed

impossible of attainment. At the time these data were collected, all

samples being on the same day, the fraction defective had actually begun

to reach the 2 per cent level, through continuous follow-up of all assignable
causes in assembly and production. Figure 6.1 shows the p chart for the
data, Using the standard fraction defective p' = .02 and the sample size of
n = 100 in (6.8) and (6.9) gives

tp .02
3/'02(.98)
UCL
p
.02 +
100
. 02 + . 042 = .062
LCL
p .02 - .042 =-
6.2. Control Charts for Defectives 111

Table 6.1. Inspection Results on a Type of Carburetor at End of Assembly

Line, All Types of Defects Except Leaks, Jan. 27, Two Shifts. Sample Size

n = 100

No. Fraction No. Fraction

Sample defectives defective Sample defectives defective
No. d p No. d p

1 4 .04 19 4 .04

2 5 .OS 20 2 .02

3 1 .01 21 .01

4 0 .00 22 2 .02

5 3 .03 23 0 .00

6 2 .02 24 2 .02

7 .01 25 3 .03

8 6 .06 26 4 .04

9 0 .00 27 .01

10 6 .06 28 0 .00

11 2 .02 29 0 .00

12 0 .00 30 0 .00

13 2 .02 31 0 .00

14 3 .03 32 1 .01

15 4 .04 33 2 .02

16 1 .01 34 3 .03

17 3 .03 35 3 .03

18 2 .02
73

The lower control limit would be -.022 if we used the formula, so we

disregard 1t just as we disregarded zero lower control limits for ranges.

The interpretation is that when p' = .02 and n = 100, the event of a

sample with no defectives is not rare enough to warrant any action.

112 CONTROL CHARTS FOR ATTRIBUTES

0.10

FIG. 6,1. Fraction defective chart on final inspection results at end of

assembly line on a type of carburator, All types of defects except leaks.
Work of two shifts on Jan, 27. Sample size 100, Reproduced with permission
from I, W. Burr, "Engineering Statistics and Quality Control", McGraw-Hill,
New York, 1953, p. 217.

Thus we could not regard such a point as an indication of the presence of

some assignable cause which has temporarily lowered p'. The appropriate
binomial, n = 100, p' = .02, gives by (2.30) P(d=O) = .133, and hence by
chance causes alone about one sample in seven or eight will have nv

defectives. Such a point is then fairly common and is not a reliable

indication of an assignable cause. The binomial distribution is rather
unsymmetrical for such a case, having points considerably farther above

the expected p' = .02 than below.

As judged by Figure 6.1, the process at this stage can justifiably be

called stable, with a fraction defective of .02. The goal of 2 per cent
has clearly been attained this day, and the good control gives confidence

that it can continue to be met.

6.2.2. Control Chart for Number of Defectives, Standard Given. In this
case we are again given a standard value p', and a sample size of n,
assumed constant here for the present. But now we plot the number of
defectives, np or d. Such a chart is commonly called an "np chart,"

although the author would prefer the one-letter designation "d chart."
6.2. Control Charts for Defectives 113

Using the data of Table 6.1 as an example we plot the 35 values of

d, the number of carburetors with at least one defect, The chart would look
absolutely identical to Figure 6.1, except that the vertical scale would

be for np or d, and since n=lOO, .OS is replaced by 5, .10 by 10, and so on.
Also we have from the binomial distribution (2.30) that E(d) = np' and

ad= /np'(l-p') from (2.31) and (2.32). Therefore we have

{ np = np' (6.10)

Limitsnp = np' ~ 3/np'(l-p'). (6.11)

Now using p' = .02, n = 100 yields

Limits 2 + 3/100(.02) (.98) = 2 ~ 3(1.4) = -,6.2.

np
These control lines correspond exactly to those for p: .02, .062, if we

multiply by n = 100, for (6.10) and (6.11) are exactly n times (6.8) and

(6.9). The same conclusions of course apply as for the p chart.

Still a third vertical scale could be used namely a per cent scale,

obtained by multiplying the p scale by 100%. Then p = .OS corresponds to

5%, etc. If it is desired to use a E!::! ~t defective scale, it is best

to do all the calculations in fraction defective, and then convert to a

per cent scale for those who will view it.

6.2.3. Charts for p and np, Analysis of Past Data. In Section 6.2.4

we shall take up the cases of unequal sample sizes, but for the present

we continue with equal n's.

In the present case, for either type of chart, we do not know p'.

Since we are testing the hypothesis that all samples came from the same

binomial population, we assume p' has an unknown but constant value. The

best estimate of p', from k samples, is p, that is

l:~=l di total defectives (6 .12)

p = -k-- total of all samples
l:i=lni
114 CONTROL CHARTS FOR ATTRIBUTES

This formula is used whether or not the n's are equal. Assuming equal n's,

however, this may also be written

p- n's equal (6.13)

Now we substitute p- for p' in (6.8) to (6.11) obtaining

p (6.14)
Limits
p
p ~ 3/p(l-p)/n (6.15)

tnp np- (6.16)

Limits
np np ~ 3Vnii (1-p). (6.17)

For an example consider the data of Table 6.2. These are for k=30

samples, each of n=SO sheets of paper, inspected against a certain standard.

The p values are plotted in Figure 6.2.

Now for p we use (6.12)

72
p 30(50) = .0480.

We could also have used (6.13)

p 1364 = .0480.

Thus the center line for the p chart is .0480. For the control

limits we have by (6.15)

Limits = .0480 + 31.048(.952)/50 .0480 + .0907 .1387.

p

Drawing the two control lines we see that there are no significant runs

relative to the center line, but that sample number 12 is out of the

control band signaling an assignable cause even though it is but one

out among 30 samples. We should look for the cause of such performance.

What conditions were different then?

Now for an np chart, we would plot the whole-number counts, d, of

defectives. Then using p = .0480, n = 50 in (6.16) and (6.17) we find

 6.2. Control Charts for Defectives 115

Table 6.2. Samples of 50 Sheets of Book Paper, 25 by 38 In., Tested for

Sheet Formation Against a Carefully Selected Standard Sheet, so that if
the Sheet Tested Has Poorer Formation than the Standard, It Is Called a

Defective

No. Fraction No. Fraction

Day defectives defective Day defectives defective
No. d p No. d p

1 1 . 02 16 3 .06

2 0 .00 17 0 .00

3 1 .02 18 3 .06

4 4 .08 19 3 .06

5 1 .02 20 2 .04

6 5 .10 21 4 .08

7 2 .04 22 3 .06

8 1 .02 23 1 .02

9 3 .06 24 2 .04

10 3 .06 25 3 .06

11 3 .06 26 0 .00

12 7 .14 27 2 .04

13 3 .06 28 5 .10

14 2 .04 29 2 .04

15 2 . 04 30 1 .02

72 1.44

'i np 50 (. 0480) = 2.4

Limits
np 50(.0480) .:!:. 3v'so(.048) (.952) = 2.4.:!:. 4.53 =-,6.93.
Thus d = np = 7 is the only sample value showing out of control. Hence
we have identical results to those for the p chart, as indeed we must.
116 CONTROL CHARTS FOR ATTRIBUTES

To emphasize that the np chart is just a scale change from the p chart,
observe the right hand scale on Figure 6.2, for np's. The values are just

n=SO times as great as for the p values on the left scale.

If on investigation we find the assignable cause and remove it from the

process, then we should revise p to 65/1450, changing the control lines

accordingly and extend them for new production. But if we have not found

the cause, or having found it, do nothing to prevent its recurrence, then

this sample is still as typical of the process as any, and should be left

in. In this case then, our p will not be ~n estimate of but a single

p', but instead an estimate of an average (weighted) of two values of p',

one when the assignable cause is not at work and the other when it is.

Another point to be mentioned is that the definition of a defective

as used may not mean at all that 4.8 per cent of the sheets of book paper

are unusable. A rather high standard of sheet formation may have been

chosen to increase artificially the percentages of "defectives." This

kind of approach is often done so that smaller samples can be used to

look for the presence of assignable causes. If defectives are extremely

rare and the average number per sample is much less than 1, then the chart

will have little power to give a warning of an assignable cause. We can

.16 8
.14 - .7
.12 6
.10 5
p 4 np
.08
.06 3
.04 2
.02
.00

FIG. 6.2. Fraction defective chart for samples of 50 sheets of

book paper tested for "sheet formation" against a minimum
quality standard sheet. Data from Table 6.2. Also interpretable
as an np chart by using scale on right side.
6.2. Control Charts for Defectives 117

in such a case (a) greatly increase n, or instead (b) use an artificially

severe criterion of a "defective" with the~ sample size. In either

case we gain more power to pick up assignable causes when present.

As a further example, let us consider some data from a random

experiment, designed to illustrate some of the principles. The drawings

were made from a box of l em diameter beads, using a paddle with 50 1 holes.

(Paddles with various numbers of holes can be used effectively in such

experiments, and also to illustrate acceptance sampling by attributes,

which we take up later on.) In the present experiment, the box contained

1600 yellow beads, and enough red ones to provide desired population

fractions defective, such as, p' = 102/1702 = .060 for samples 1 to 75.

Theoretically the law governing the probability of d = 0,1,2 defectives

in n = 50 is the hypergeometric law, since in 50 consecutive drawings, each

of 50 beads, the probability p' varies slightly after each drawing of a

single bead, depending on the previous draws. However, since 50 is such

a small proportion of 1702, we can readily use the binomial law

(2.30) as an excellent approximation. Hence a p chart approach is applicable.

After each 50 the beads are returned, and remixed.

The results are shown in Table 6.3. Treating the first 25 samples of

50 as the preliminary data we find

p = 81/1250 = .0648
Limits
p
.0648 ~ 31.0648(.9352)/50 = .0648 ~ .1044 = -,.169,
Using these lines we find the first 25 samples to be in perfect control.

The fraction defective, may or may not be satisfactory from a practical

viewpoint. The process is stable (as we certainly expected). The

control lines are extended for the next 25 samples and good control continues.

There is, however, a run of nine points in succession below the center

line (samples 36 through 44). This might or might not be regarded as

 118 CONTROL CHARTS FOR ATTRIBUTES

Table 6.3. Defectives Data Drawn Randomly from Various

Populations, Simulating a Production Process, Analysis of
Past Data. True Fraction Defective Shown in Remarks Column

Sample Sample Number of Fraction Remarks

No Size Defectives Defective
n d p = d/n

43 50 1 .02
44 2 .04
45 4 .08
46 2 . 04
47 1 .02
48 6 .12
49 3 .06
50 50 4 . 08 Revised p= 155/2500 = .062
1250 ~ 1. 48 Limits -' . 164 n = 50
51 200 9 .045 Limitsp .011,.113 n = 200
52 200 6 .030 p
53 14 .070
54 15 .075
55 12 .060
56 11 .055
57 17 .085
58 6 .030
59 11 .055
60 8 .040
61 14 .070
62 14 .070
63 15 .075
64 14 .070
65 13 .065
66 14 .070
67 5 .025
68 14 .070
69 6 .030
70 6 .030
71 14 .070
72 10 .050
73 16 .080
74 12 .060
75 200 15 . 075 Revised p = 446/7500 = .0595
29T 1.455
76 50 9 .18 Limits = .0595 + .1004 =-,.160
77 5 .10 Eviden~e of assignable cause
78 6 .12 on sample 76
79 7 .14 p' .12 on samples 76-100
80 7 .14
81 7 .14
82 8 .16
83 4 .08
84 9 .18
85 so 4 .08
 6.2. Control Charts for Defectives 119

Table 6.3 (continued)

Sample Sample Number of Fraction

No. Size Defectives Defective Remarks
n d p = d/n

1 so 3 .06 p' .06

2 so 4 .08
3 5 .10
4 1 .02
s s .10
6 s .10
7 3 .06
8 3 .06
9 2 .04
10 1 .02
11 3 .06
12 7 .14
13 1 .02
14 3 .06
1S 4 .08
16 2 .04
17 s .10
18 4 .08
19 2 .04
20 1 .02
21 4 .08
22 6 .12
23 l .02
24 4 .08
2S so 2 .04 End of preliminary data
12SO 81 l. 62 p- = .0648 Limits =-, .169
p
26 so 1 .02
27 2 .04
28 3 .06
29 3 .06
30 8 .16
31 s .10
32 2 .04
33 2 .04
34 s .10
3S 6 .12
36 3 .06
37 3 .06
38 2 .04
39 3 .06
40 2 .04
41 1 .02
42 so 0 .00
 120 CONTROL CHARTS FOR ATTRIBUTES

Table 6.3 (continued)

Sample Sample Number of Fraction

No. Size Defectives Defective Remarks
n d p = d/n

86 50 8 .16
87 9 .18
88 7 .14
89 7 .14
90 5 .10
91 6 .12
92 2 .04
93 6 .12
94 7 .14
95 4 .08
96 6 .12
97 13 .26
98 7 .14
99 3 .06 For samples 76-100 is
100 50 4 .08 p= 160/1250 = .128
16/) 3.20

101 50 0 .00 Continue limits set from

102 2 .04 samples 1 to 75, that is,
103 1 .02 -,.160 for n =SO
104 1 .02 p'= .03 for samples 101-125
105 4 .08
106 1 .02
107 1 .02
108 2 .04
109 1 .02
110 1 .02
111 2 .04
112 2 .04
113 0 .00
114 2 .04
115 3 .06 Run of 10 in succession on
116 1 . 02 or below center line .
117 2 .04
118 1 .02
119 3 . 06
120 2 .04
121 0 .00
122 1 .02
123 4 .08
124 4 .08
125 50 2 . 04
---.r.r- . 86
6.2. Control Charts for Defectives 121

.26
.24
.22
.20
.18
.16
.14

J "ljJ v

60 70 80 90

FIG. 6.3. Samples 1-25 preliminary data: p'=.06, n=50, p=.0648. Limits
extended. Sample 26-50: p'=.06, n=50, revised p = .0620. Samples 51-75,
p'=.06, n=200, limits from p=.062. Revised p=.0595. Limits extended for
samples 76-100: p'=.l2, n=50. Out of control. Limits from p=.0595
extended through samples 101-125: p'=.03, n=50. Data from Table 6.3.

significant by the person in charge.* If so regarded

it would indicate a desirable assignable cause, unless it proves to be

lax inspection.

Suppose that at this point it is decided to use n = 200. We revise

the limits on the basis of the first 50 samples, obtaining
p = 155/2500 = .0620

Limits = .0620 + 31.0620(.9380)/200 = .0620 _+ .0512 = .0108,.1132.

p -
We see from the limits that a single point can now be outside the control

band on either side which was not so before. But none were outside, thereby

signaling the continuation of perfect control. The student should note how

* For a binomial distribution with p' = .06, n = 50 we may find P(3 or less

in SO) = .6473, that is, considerably above .5, thus enhancing the probability

of such a run as compared to that for a .5 probability.

122 CONTROL CHARTS FOR ATTRIBUTES

much less variation there was in plottings for n = 200 than for n = SO.

This of course follows from (2.34).

We now again revise the center line and limits and revert to n SO.

p = 446/7SOO = .OS9S
Limits
p
.OS9S ~ 3I.OS9S(.940S)/SO = .OS9S + .1004 = -,.160.
Immediately we have evidence of an assignable cause on sample number 76.

The defectives total nine, giving p = .18. Action should be taken at

once. However, we have'continued the drawings from sample 76 to 100 with p'

at .12. The next few samples after 76 are back in control, but 82 shows a

p value on the limit and other such warnings follow.

We now assume that the assignable cause has been found after sample

100, and appropriate action taken. The data 76-100 are eliminated as not

applicable, and the old limits from samples 1 to 7S are extended for sample

101 and onward. The points seem to be running low. At llS there have been

ten consecutive points on or below the center line, which indicates an

assignable cause of low fractions defective. Actually samples 101-125

were made with p' = .03.

This experiment illustrates how a p chart might run, and shows how

evidence occurs when p' is changed from .06 up or down, with samples of 50.

6.2.4. Varying Sample Sizes for p and np Charts. Quite often in

practice, the sample size for data on defectives varies. Commonly when this

is the case, it is because the sample represents results from sorting 100%

the entire production of a shift or a day, or possibly one hour. Since the

production is not exactly constant, n varies. How shall we handle this

problem, and which chart, p or np, shall we use?

Let us consider the data in Table 6.4. These results are for September

following those given in Figures 3.1 and 3.2 for results of applying the B

test to an electrical assembly in August. The fraction defective, p = d/n,

for each day is listed, and plotted in Figure 6.4.
 6.2. Control Charts for Defectives 123

Table 6.4. B-test Failures of Electrical Equipment for September.

Sample Size Variable. Fractions Defective Plotted in Figure 6.4.
The Data Follow those of Table 3.1.

Date Sample Size Defectives Fraction Defective

1 1600 41 .0256
2 1390 34 .0245
3 1510 34 .0225
4 1640 17 .0104
6 1780 34 .0191
7 1670 7 .0042
8 1750 15 .0086
9 2000 39 .0195
10 1640 30 .0183
11 1710 14 .0082
13 1780 15 .0084
14 1510 21 .0139
15 1780 16 .0090
16 1600 57 .0356
17 1670 31 .0186
18 1540 34 .0221
20 1670 30 .0180
21 1450 16 .OllO
29,690 485

.04

.03

p .02

.01

.oo~,~2~3~4~6~7~8~9~,~o~,~,~,~3~1~4~1~5~1~6~1~7~,a~2~0~2,
September

FIG. 6. 4. B-test failures for electrical equipment for September, Data from
Table 6,4, Variable sample size, p = ,0163,
124 CONTROL CHARTS FOR ATTRIBUTES

Using (6.12), (6.14) and (6.15), we find

cl p p = 485/29,690 =
= .0163

Limits
p
p -+ 3/pq/n.1 i = 1,2, ... ,18

.0163 + 31.0163(.9837)/n. = .0163 + .380/ln 1. ,

- 1

The last form is quite convenient and permits rapid repeated calculation

of the + values to find appropriate limits. In Figure 6.4 we have plotted

the individual control limits for each day's p. Control is much better

than for August with a much better p. But there is still evidence of an

undesirable assignable cause on the 16th, and of a good one on the 7th,

unless the latter proves to be lax testing. A desirable feature of this

p chart is that, by (6.14), the center line is always at p= .0163, regardless

of n. Thus runs on one side of the center line may be checked for, and

also the higher the point the worse the observed quality.

On the other hand, suppose that we make an np chart by plotting the

d np values. Now by (6.16) the center line is

i = 1,2, ... ,18

and it will vary with each sample size. The limits are
Limits = .0163n. ± .380·~
np 1 1

We note that the width of the control band for a p chart becomes narrower

as ni increases, but for an np chart it becomes wider as n increases

(varying as~). But the chief source of change lies in the center line
1

which varies as ni itself. No longer can we easily check for runs, and also

a point being below another does not necessarily mean a better p. For this

reason we almost always use a p chart if then's vary much at all.

Now let us reconsider Figure 6.4. The control limits are quite uniform.

How might we simplify our calculations? Can we make do with "average" limits?

The answer would seem to be yes. Now

ii. k
l:i=l n/k (6. 18)

which gives us here 29,690/18 = 1650. This yields limits of

6.2. Control Charts for Defectives 125

.0163 ± .380/~1650 = .0163 ± .0094 = .0069, .0257.

This one set of average limits may be used for all the p points. The decisions
are practically all obvious. But for points close to an average limit, either
just inside or just outside we may wish to use the actual n. For example for

Sept. 1 the point is inside, but close. Using the exact n = 1600 still leaves
the point inside. One might wish to check also the exact limits for Sept. 8, 11
and 13. As shown in the figure all three are in control and would be for the

average lower limit too.

Now on the other hand an np chart would have a large amount of variation
among the limits, much of which is due to the center line varying around, as
well as the width of the band. Using n for all would not be feasible here.

About how much variation from n can we tolerate? In our particular data
here the smallest n = 1390 which is 16% below n, while the largest n = 2000
is 21% above n. It would seem that we might go even further.
A rule of thumb is that if our n 1 s do not differ over 30% from n then for

a p chart we may use n for all in this grouping of n•s, checking any points very

close to the limit. (For an np chart, the author suggests not over 5% variation
of n from n.)
It is perhaps worth emphasizing that when the n 1 s vary, we do not find

p by Ep/k, as this would give equal weight to all sample p's whether n is

large or small. Instead we use (6.12). But note that (6.12) may also be
written
-=
p (6.19)

which clearly shows p- to be a weighted average of the p's.

6.2.5. Not All Per cent Defective Data is Analyzable by a p Chart.

Sometimes one meets per cent defective data which should not or cannot be

analyzed by a p chart. Three types involve violations of the conditions for

a binomial distribution to apply.
126 CONTROL CHARTS FOR ATTRIBUTES

1. If we have extremely large sample sizes, such as the entire output

of a plant, then the limits may be ridiculously narrow. For example, if the
total output of glass jar caps per week is 4,000,000 and p = .020 non-functional
defectives, then the limits would be

.020 ~ 31.02(.98)/4,000,000 = .020 ~ .0002 = .0198, .0202

It is virtually impossible for a single weekly p to lie within such limits. The
reason is that p' is not constant among such diversity of product. Such data
can be analyzed by charts for individuals x and moving range, see Section 7.1.

2. When a number of parts are made at once, such as small metal castings,

stampings, plastic pieces, rubber molded parts, etc., then the incidence of
defectives is likely not to be independent. If one is bad, those near it are
likely to be also, because of defective material or conditions. Sometimes
such proportions defective can be analyzed to advantage by using x and R charts.
That is, treat each sample proportion as a single x.

3. A moderately common case is that of per cent spoilage of bulk

product. For example we may have 5000 gallons of paint of which 40 gallons
are removed from the final product. The proportion defective is 40/5000 = .008.
Can we treat such results as p values? The answer is yes, but only if we
have a series of 5000 one gallon containers, each one of which is declared
good or defective. But such is not likely to be the case. The paint may well
be in a large vat,and the undesirable product removed. The chief trouble

is that there is no logical sample size. If gallons, n = 5000; if pints,

n = 40,000, and if cubic feet, n = 668. What limits should one use? Thus
the binomial distribution is inapplicable. Again one can call such a
fraction defective an x and use x and R charts to analyze.

6.2.6. Further Comments on Charts for Defectives. As mentioned

before, people from top management on down through foremen to the inspector
6.2. Control Charts for Defectives 127

and machine operator seem to understand percentages better than decimal

fractions defective. So it is well to make the calculations in decimal

fractions and plot them as such, but then to use the scale in per cents.

In calculations, especially those for p charts, it is easy to make a

decimal point error! Hence great care should be exercised in calculating

limits, or you are likely to have limits too wide by lfO or too narrow by

1/ v'IO. Either should "look" queer. One can also forget the multiplier 3

in front of the square root!

In choosing a sample size there are many practical considerations. One

is that n should be small enough so that production and testing conditions

can be presumed relatively uniform, in order that the sample can be called

a "rational subgroup." Otherwise the chart loses much power to tell when

production conditions changed. On the other hand the sample size needs to

be large enough so that we will average at least .5 defective for nP. If P

iS very small this will require a large n. The author knows of one case

where np was only 1 or 2 with n lO,OOO,and the np chart was not helping

much, because too many things could happen while 10,000 were made. But

after developing a way to obtain a measurement, x; x and R charts for n 4,

enabled them to find the causes and eliminate them, giving much further

improvement.

Usually the fraction defective or per cent defective chart is used,

especially if the sample size varies appreciably. However, if n is con-

stant, then use of an np =d chart enables one to plot the whole numbers d

instead of bothering to figure p = d/n and plotting decimals. If the n is

messy like 312 or 39 this is a saving. Such charts are especially useful

in conjunction with acceptance sampling by attributes, where a record is

kept of results on the first sample from each lot.

128 CONTROL CHARTS FOR ATTRIBUTES

If it is desired to compare the current month's performance against

a standard p' (or the old p) we can do the following in addition to a control
chart for the current month. Suppose there k samples of n each. Then we
may treat p for the month as though it were a p for a big sample of nk

pieces.

Limits-
p
= p' ± 3/p'q'/kn (6. 19)

If p is outside either limit we have clear evidence of either an improvement

or deterioration. Note, however, that this test is rather meaningless unless

there was reasonable control around p. (This is really a significance test

on p.)

Finally, we fairly often have quite low p's or p 11 s, in which case the

q or q 1 term is close to 1 and its square root even closer to 1. Hence we

may omit such a term from the calculations and have for example

Limits np = np ± 3~ p small (6.20)

6.3. Control Charts for Defects. The second main type of chart for

attributes is that for the number~ defects on a sample or unit of product.

In the charts for defects, the "c chart," we are concerned with a count of

the defects found on, say, a complicated sub-assembly. The average number of

defects of all types found in final inspection may be, say, 12 in such sub-

assemblies. For them it can easily be true that almost never does one come

along having no defects at all. Thus if we were to try to plot a fraction

defective chart,p would always be 1.00 n/n, since each sub-assembly contains
at 1east one defect and is therefore a "defective." Such a string of 1. 00

p values would tell us nothing! Instead when we count the number of defects

c, we have whole numbers centering at 12, but running both above and below,

and measuring quality performance in a meaningful way.

6.3. Control Charts for Defects 129

6.3.1. Control Charts for Defects, Standard Given. We take up first the case

where we have a given standard number of defects per "area of opportunity"

for defects.

c' = standard number of defects. (6. 21)

Such a standard value is commonly the result of experience after a period of

reasonably good control. Or it may be some requirement or goal, but such

values must be reasonable, or we shall lose most probability interpretations

of points out.

The Poisson distribution is the basic model used for a c chart, that is,

given c 1 , the probability for each c value to occur is given by (2. 35) . In

order for the Poisson distribution to apply we must have the conditions given

in Section 2.5.4., namely that (a) the fields of opportunity for defects to

occur must be equal, (b) the defects must occur independently of each other,

and (c) the maximum possible number of defects, (max c), must be far more

than c 1 • Assuming these conditions satisfied, then we are given c' and we

use (2.36) and (2.37) yielding

<fc (6. 22)

Limits
c
c 1 ± 3o
c
c1 ± 3W. (6.23)

We may now plot these three lines, or two if c' - 3/CI < 0, and are prepared

to take action on the very first sample, and of course subsequent ones. A

point outside the limits is an indication of the c value being incompatible

with c'. This indicates an assignable cause of the true average number of

defects being either above or below the standard c 1 , according to which

limit is transgressed.

Note that unless c' is realistic, it would be entirely possible for

the process to be homogeneous, that is, in control relative to some other

130 CONTROL CHARTS FOR ATTRIBUTES

average number of defects, and yet to show many points outside limits set
from the standard c 1 • For example, suppose c 1 = 2, then the upper control

limit is at 2 + 3/2 = 6.2. But if the actual process average is stable at

6 defects per unit, then many points will be above UCLc. On the other hand

all the c values might lie within the limits c' ± 3/CI = 6 ± 316 =-, 13.4,

and therefore be in control.

We could illustrate a c chart with given c' by drawing chips or beads

from a bowl, marked according some Poisson population. Then draw from a

different Poisson population, etc. Such populations are easily made up

from tables of Poisson distributions giving P(c), c = 0,1, ... About 500

chips should be used so that at least a few rare c's are in the collection

or bowl, which could give a point out of control.

6.3.2. Control Charts for Defects, Analysis of Past Data. We now

come to the more common case of analysis of past data. Here we do not yet

have a standard c'. We are, however, testing the hypothesis that there is

an unknown but constant c', and under this assumption, the best estimate of

c' from k sample cls is c:

(6. 24)

Then we use c and have the provisional or approximate lines from substituting

c for c' in (6.22) and (6.23).

-c (6.25)
<Lc
Limits -c ± 3vfc. (6. 26)
c

These are drawn through the points constituting the preliminary run of, say,

25 sample c values. Appropriate action is taken, c possibly revised, and

it and limits found, against which we can plot each new c value as it is

found.
 6.3. Control Charts for Defects 131

As an example, let us take the data of Table 6.5. These are for counts

of the total number of defects found per day on electrical switches. Several

different types of defects were inspected for, c being the total number of

all types of defects found by inspection of all of the 3000 switches produced

daily. Hence when c = 90 defects were found, this could make 90 of the switches

defectives, or, because some of the switches might contain more than one of

the defects, there might be less than 90 defective switches. Here either a

c chart, or an np or p chart, could be used. The quality control man used a

c chart.

Figure 6.5 shows the data plotted. For analysis we have from (6.25),

(6.26):

c = 1198/21 = 57.0

Limits -c ± 3f"i = 57.0 ± 3157.0 = 34.4, 79.6.

c

Considerable lack of control is indicated by the control limit lines, both

potentially good and bad assignable causes.

Careful records were kept on how many times each particular defect

occurred and action taken to improve the status of the most prevalent or

costly ones. It is possible to run a separate c chart on each of the more

common types of defects. It is interesting that December of the previous year

Table 6. 5. Total Defects of All Kinds on Day's Productions of 3000

Electrical Switches. December, 21 Working Days.

Day Number 1 2 3 4 5 6 7 8 9 10 11

Total Defects c 30 56 47 86 44 23 16 64 80 54 73

Day Number 12 13 14 15 16 17 18 19 20 21
Total
Total Defects c 65 76 69 53 58 30 91 90 36 57 1198
132 CONTROL CHARTS FOR ATTRIBUTES

100

o~------~5--------~,o~-------,~5--------~20~
Day

FIG. 6.5. Total defects of all kinds on day's production of

3000 electrical switches. December, 21 days. Data from Table 6.5.

had c- of 406, or about seven times as many defects per day. But there is

still work to be done as indicated by the points out of control.

It is perhaps worth noting that the control limits (6.26) for a c chart

are similar to the approximate limits (6.20) for an np chart. This is

because, when ij is near 1, ~ is therefore small and we are reaching the

conditions under which the Poisson distribution is an excellent approxi-

mation to the binomial.

6.3.3 Some Examples of c Chart Data. Many examples of possible appli-

cations of c charts occur in industry, not all of which are directly in

quality control. Some are: errors or defects of any assembly or completed

product, such as, automobiles, trucks, farm implements, radios, airplanes,

electronic panels or bicycles; pinholes in a test area of painted surface;

visual defects on enameled ware; surface defects in steel plate or tinplate;

slivers on lengths of copper wire; breakdowns of insulation in 10,000' lengths

of covered wire at (excess) test voltage; typographical errors per page of

print; bacterial colonies on an agar plate; accidents or hospital calls

per week; lost-time accidents per year; absenteeism; defects on yard goods,

shirts, etc.; bubbles or hard spots in a dozen glass jars; counts of instances

of any of 20 kinds of defects in jar caps in say 10,000; gas holes or other
6.3. Control Charts for Defects 133

defects in castings; foreign matter in an assembly. In using a c chart,

however, one should keep in mind the requirements justifying the Poisson

distribution, given in Section 2.5.4. The next section describes how to

avoid trouble in some cases with unequal "areas of opportunity" for defects.

6.3.4. Chart for Defects per Unit. We next discuss a type of chart

which is exactly equivalent to the c chart, but is in a more convenient

form sometimes, especially if the "area of opportunity" or sample size

varies. Moreover by means of this approach the overall picture can be

given without too much detail. Hence such a chart is often useful to

management.

Let us illustrate by an example. A plant making trucks began to

analyze the number of errors or defects, in the final inspection at the

end of the assembly line. The average number of defects per truck from

September 16 through November 1 was 1.95. Since daily production was 95

trucks and the period includes 35 working days, the total production was

about 3325, hence there were about 6484 errors altogether, so that the

average number per truck is 6484/3325 = 1.95.

Let us now introduce some notations.

n = number of units of product per sample (6. 27)

c = total number of defects in sample of n units (6. 2fl)

u c/n = average number of defects per unit (6. 29)

over n units
k
(Ei=l ui)/k = average of u's for k samples (6. 30)

u'= population or process average defects per unit. (6.31)

In the example being discussed, there would be 35 u values in the 35

days. September 16 might have had 190 errors in 95 trucks, or u = 190/95

2.00. The average of the 35 such u values was 1.95. After this past data,

the plant began to plot the u points daily. The data are given in Table 6.6.
 134 CONTROL CHARTS FOR ATTRIBUTES

Table 6.6. Average Number of Errors Per Truck, Assembly Line 1, Station 2.
Average Daily Production: Nov. 4 to 20, 95 Trucks; Nov. 21 to Jan. 10' 130

Trucks

Average no. Average no. Average no.

Date errors u Date errors u Date errors u

Nov. 4 1. 20 Nov. 25 2.25 Dec. 17 1. 85

5 1. 50 26 2.50 18 1.82

6 1. 54 27 2.05 19 2.07

7 2.70 29 1.46 20 2.32

8 1. 95 Dec. 2 1.54 23 1. 23

11 2.40 3 1. 42 24 2.91

12 3.44 4 1.57 26 1.77

13 2.83 5 1.40 27 1.61

14 1. 76 6 1. 51 30 1. 25

15 2.00 9 1. 08 Jan. 3 1.15

18 2.09 10 1. 27 6 1. 37

19 1. 89 11 1.18 7 1. 79

20 1. 80 12 1. 39 8 1.68

21 1. 25 13 1.42 9 1. 78

22 1. 58 16 2.08 10 1. 84

Thus on November 4, c = 114, u 114/95 1.20, etc. Figure 6.6 shows all

45 u values plotted.

Now what shall we do for center line and limits? There are as usual the

two cases.

Standard u' given:

A-
'l:u
u' (6.32)
6.3. Control Charts for Defects 135

•
• • •
2.3~--.a.------.£:3.~._.!---------- •
1.95 • • •1.88 •
-r.----.- --.-...
... -----------.-
• •158
1.51__. _ _ _ _ _ _ _ _'-+----r-.-.----- 1.55 •
• • • • • 1.22 •
• --~

FIG. 6.6. Control chart for average number of errors per truck.
Average daily production at station: 95 Nov. 4 to 20; 130 Nov. 21
to Jan. 10. Data in Table 6.6. Reproduced with permission from
I. W. Burr, "Engineering Statistic& and Quality Control", McGraw-
Hill, New York, 1953, p. 240.

But cru = crc/n = crc/n. But u 1 = c 1 /n where c 1 is the population average

total number of defects on the samples of n units. Thus c 1 = nu 1 • Hence crc
{;!= v'nu 1 , so that a
u
= v'nu 1 /n = v'u 1 /n. Therefore

Limitsu =u 1 ± 3v'u 1 /n. (6.33)

For the case analysis of past data, we would merely substitute u- for u 1 having

(6. 34)

Limitsu -
u ± 3fu/n. (6. 35)

Hence for the data at hand we use

II- 1. 95
't.u

Limitsu 1.95 ± 311.95/ 95 = 1.95 ± .43 = 1.52, 2.38.

These limits are shown on Figure 6.6 through November 20. Control is quite
erratic. This could easily be in some degree due to non-independence of

errors. (If an error occurs on a truck the next few may well have the same
136 CONTROL CHARTS FOR ATTRIBUTES

error.) On November 21, production was increased to 130, so that limits

were revised to

1.95 ± 311.95/130 = 1.58, 2.32.

Because of the obvious improvement in late November and early December, new

control limits were calculated for production after December 16, based upon

data of November 21 to December 13, as follows:

u = 24.87/16 = 1.55
Limits u = 1.55 ± 311.55/130 = 1.22, 1.88.

Considerable trouble was encountered for a while in meeting the new

control limits. Particularly note the high point for Christmas Eve. In

January the quality performance has settled fairly well into the lower

control band.

The average number of errors per truck was driven steadily downward

by the psychological effect of the charts and by following through on the

various kinds of trouble occurring. A sizable saving in man-hours, delays,

supervision and material was achieved.

It is interesting to compare the last limits for u = average defects

per truck with limits for defects on an individual truck. Such limits are

1.55 + 3YT:SS= -,5.3. Thus it takes 6 errors for a single truck to be out

of control, but only an average of 1.88 errors per truck for the day to be

out of control.

There is a variation on the foregoing which may seem different to the

reader but which uses in reality just the same principle. It is that case

in which we do not have n separate discrete units in a sample, but instead

some bulk product by volume or area. To use the standard c chart, the only

requirement would be that the area or volume be always the same, so that the
6.3. Control Charts for Defects 137

field of opportunity for defects is constant. When this is not the case, then
we must make some allowance for the changing conditions.

Suppose, for example, that in our original problem we are testing the

insulation on 5,000-, 10,000-, and 12,000-ft lengths. We first choose a

standard length, say 10,000 ft. Then the other cases are reduced to this

basis. The number of defects on any 5,000-ft length is doubled to give a

number of defects comparable with that for 10,000 ft, while any number of

defects observed on a 12,000-ft length would be divided by 1.2 for the same

reason. Then we plot the resulting data, which are thus defects per unit

(1 unit= 10,000 ft).

Now to get the center line, we want the total average defects per unit,

where each such figure is given the proper weight. Suppose that there are

18 breakdowns of insulation in ten 5,000-ft lengths; then this is 18 defects

in 50,000 ft, which is the equivalent of five 10,000-ft lengths. In five

12,000-ft lengths suppose that there are 15 breakdowns. Since we have

60,000 ft in the five, this is equivalent to 15 defects in six 10,000-ft

lengths. Finally suppose that there were 43 defects in twelve of the standard

lengths. Then altogether we have 18 + 15 + 43, or 76 defects in 5 + 6 + 12,

or 23 units of wire. Dividing gives the weighted average

u- = 3.30

To systematize the foregoing we may set up the accompanying tabulation.

No. Total Individual Length Number Units

pieces defects length, ft 10,000 ft (1) X (4)
(1) (2) (3) (4) (5)

10 18 5,000 .5 5
5 15 12,000 1.2 6
12 43 10,000 1.0 12
76 23
138 CONTROL CHARTS FOR ATTRIBUTES

Now for the control limits we are planning to plot defects per unit u,
that is, defects on the piece divided by the item in column (4) (or multiplied
by its reciprocal), and hence all points will have the same center line. The

width of the control band, however, will vary. The item in column (4) corre-
sponds ton in (6.35), and hence we have

5DOO ft: 3. 30 .t 3 /3."30 = - - ,

v-:s 11.0

12,000 ft: 3. 3D± 3 /3."30

VT:"Z = --, 8.3

10 ,ODD ft: 3.30±3M= --, 8.7

In this way each point has its own control limits. Of course, as with the

p chart, if the variation in area of opportunity for defects is not great

no adjustment need be made.
6.4. General Comments on Control Charts for Attributes. Both general

types of control charts for attributes involve inspection or testing for

defects. It is therefore vital that each defect be clearly defined. For

example a very small blemish may not be a defect at all, but above some size
it becomes a defect. In fact sometimes there may be three classes as to size

(a) not a defect, (b) minor defect (c) major defect. The same can be

true of a dimension. For visual defects a collection of limit samples at

the dividing line can be a great aid to inspection. Unless each defect is
unequivocally defined and agreed upon, inspection is quite ineffective and
charts meaningless. The author recalls an actual case in which 60 caps were
passed around to many key people in a plant and each one asked which caps

had functional defects. The numbers of caps so designated by the people

ran all the way from 3 to 6D! A similar case was observed on castings. If

key people cannot get together on defects, what hope is there for inspection?
6.4. General Comments on Control Charts for Attributes 139

Some of the steps to be taken in attribute charting are:

1. Decision on what is the purpose of the study. Is it worthwhile?

Is there enough at stake?

2. Where in the production process should the samples be drawn for

inspection? Usually it is desirable to be as close as possible to

the place where trouble is presumed to arise.

Often it' has been found that the first rough operation (for example

a grinding) is the key to the trouble.

3. What characteristics would it be desirable to inspect for, and which

can be so inspected? What characteristics should be included in a

class of defects or defectives to go on one chart?

4. Choice can often be made between charting defectives or fraction

defective vs. charting defects. Which should be used? Or would x,R

charts have a better chance of getting at the trouble or making

improvement?

5. Choice between p or np charts, c or u charts.

6. Selection of samples should be done so as to be a rational subgroup,

that is, all conditions relatively constant (men, material, machine,

measuring). This may require rather small samples.

7. Forms can be devised so that a record is made of which defects

occurred. A code can be used. Then the more prevalent or costly

ones can be followed up on. This form may include a log of the

production conditions and personnel involved, or this may be kept

separately. Potential assignable causes should be kept track of.

8. Past data may be available. But conditions surrounding such data

should be known, or the data can easily be meaningless.

140 CONTROL CHARTS FOR ATTRIBUTES

9. As soon as reliable data are available preliminary analysis by control

lines should be made and appropriate action taken.

10. A list of all the potential assignable causes which may affect a

process could well be made, contributed to by everyone involved.

The probability of a point out on a p or np chart for given p', or

on a c chart for given c' can vary quite a bit. This is because the plotted

data only go by jumps (whole numbers ford or c), and also because the

binomial or Poisson models give unsymmetrical distributions commonly.

Therefore the a risk of a point outside the control band varies quite a

little, although it is always small. See [2]. Sometimes an executive or

foreman wants to be more sure of catching an assignable cause when one comes

in, than is done by using three-sigma limits. Then, at the expense of an

increased a risk of a false signal, one can use two-sigma limits. This does

require careful explanation.

6.5. A Method of Rating Product Quality. There is a useful method for

describing quality performance, which was developed in the Bell Telephone

Laboratories (3], [ 4]. It is especially useful for complicated assemblies.

If one uses a c chart for all the defects in such assemblies, then each defect

will be given a weight of one. The system begun in the Bell System was to give

demerit points for each defect noted. They used four classes:

A - Very Serious (Demerit value 100 points)

B - Serious (Demerit value, SO points)

C - Moderately Serious (Demerit value 10 points)

D - Not Serious (Demerit value l point)

For a quite complete description of what defects to include in each such

class, see (4], which also covers the general subject thoroughly.

We may define the "demerits" for a unit of product by

( 6. 36)
6.5. A Method of Rating Product Quality 141

where the w's are the demerit values 100, SO, 10, and 1, and the c's are the

numbers of defects noted in each class. As such we then have a demerit

rating for each sample, which takes account of the relative seriousness of

each defect.

Probably the most common application is in a monthly or weekly rating

of all product inspected over that time period. For the month,say, then

we have
n
I Di = wAicA + w8 Ic 8 + wcicc + w0 Ic 0 (6.37)
i=l
where the I 1 s are over 1 to n also, and give the total defects found for the

month in each class.

There may be standard values cl···. ,cb, of defects per unit, or we may

use ~A'' ..• ~ 0 for the previous month or several months. In the latter case

we may eliminate some data before figuring the averages, if defects were

produced from known and eliminated assignable causes.

Since the number of units inspected may well vary each month, the gross

demerit sum in (6.37) is not as meaningful as the demerits· per unit, which

is found by
wAicA+w 8 Ic 8 +wCicc+w 0 Ic 0
u n
( 6. 38)

Now cA is a count of defects, distributed by the Poisson law, if the conditions

are fulfilled, and hence

ICJ:• then a- ./cl/n.

cA
Then

( 6. 39)

U=
CJ (6.40)
n
2 2 2 2 2
since aax+bY = a ax + b cry.
142 CONTROL CHARTS FOR ATTRIBUTES

Bence

<f. U E(U)

Limitsu = E(U) ± 3ou (6.41)

If n is small and we use weights for U of 100, SO, 10 and l, the chart

will be quite meaningless, because a single 100 or 50 point defect will

throw it out of control, otherwise it will be in control. But for sizable

n's the chart can be quite useful for giving the broad picture.

6.6. Summary. We hereby present the basic formulas for control charts

for attributes for convenience and to show the essential unity.

I. Charts involving defectives

A. p chart, fraction defective

l. Analysis of past data, use p

a. <tp = p = L:d/L:n
b. Limits
p
=p ± 3~q/n
2. Standard given, p'

a. { = p'
p
b. Limits = p' ± 3/p'q'/n
p
B. np or d chart, defectives. Change of scale from p chart

l. Analysis of past data, use p- = L:d/L:n

a. 4:np = np

b. Limits
np np ± 3~
2. Standard given, p'

a. t np = np'

b. Limi u 11P = np' ± 3vfiP'Ci'

6.6. Summary 143

II. Charts involving defects

A. c chart, defects on a unit or sample

1. Analysis of past data, k samples, use -c

a. 4_c = c= L:c./k
1

b. Limits c = c± 3~

2. Standard given, c'

a. <!;. = c'
c
b. Limits = c' ± 3!C'
c
B. u chart, average defects per unit = c/n

1. Analysis of past data, k samples, of n each,use

- k k
u = L:i=l ui/k Ei=l ci/nk

a. <{. u = u
b. Limits u = u~ ~
2. Standard given u' =population average defects per unit

a. tlu = ul

b. Limits = u' ± 31U'7n

u
C. Quality rating, demerits, see (6.39), (6.40).

III. A relation between np or d charts and c charts: If p' or p- is

small then 1-p' = q 1 or q is nearly 1, and may be neglected. Then

np ~ 3~ is like c+ 3./i. and np 1 ~ 3v'np' is like c 1 + 30.

IV. Unequal sample sizes

A. Defectives charts

1. d chart: may well use fixed control lines if n's do not

differ over about 5 per cent from n.

144 CONTROL CHARTS FOR ATTRIBUTES

2. p chart: may well use fixed control lines if n 1 s do not

differ over about 30 per cent from n.
3. Otherwise use different control lines for each grouping of

reasonably homogeneous n 1 s.

B. Defects charts - 'adjustment of cal cul at ion and limits

relative to some chosen sample size. Could use criteria

similar to Al for c chart and A2 for u chart.

Again we emphasize the need for sound workable definitions of defects,

independence of defects, and the use of rational subgroups, not letting

conditions vary too much during the production ard inspection of a sample.

Also a thorough log of conditions is helpful in tracking down causes.

PROBLEMS

6.1. Data from an optical company are given below, each article, 39 per box,

inspected being either good or broken. Make an appropriate control chart

for the data and comment. Why do you think the sample size of n = 39 was

used? Using a Poisson approximation and np as c', estimate the probability

of a box having less than 36 good articles.

Table P6,1

Sample No. Broken Sample No. Broken Sample No. Broken

1 2 16 1 31 6
2 1 17 0 32 1
3 1 18 1 33 2
4 0 19 0 34 2
5 2 20 1 35 2

6 3 21 0 36 1
7 2 22 2 37 1
8 1 23 0 38 0
9 1 24 1 39 2
10 2 25 1 40 2

11 1 26 0 41 2
12 1 27 42 5
13 1 28 0 43 2
14 2 29 7
15 1 30 10
 Problems 145

6.2. The following data give results on beer cans inspected for
visual defects, in lots of n = 312 cans. Use an appropriate control chart

analysis and comment. What would you suggest for a p for the next samples?

Table P6.2

Lot n Defectives Lot n Defectives Lot n Defectives

1 312 6 11 312 7 21 312 9

2 7 12 7 22 5
3 5 13 6 23 11
4 7 14 6 24 15
5 5 15 6 25 312 10
6 5 16 6 196
7 4 17 23
8 5 18 10
9 12 19 8
10 312 6 20 312 5

6.3. The following data are on 5-gal pails for paints and varnish.

Draw an appropriate control chart. What do you conclude? Is p Lp/11?

Why not?

Table P6,3

No. No. Fraction

inspected defective defective
Date n d p

June 3 1,024 30 .0293

5 2,056 48 .0233
6 16,835 86 . 0051
7 16,069 108 .0067
12 7,365 54 .0073
13 8,738 29 .0033
14 1,220 10 .0082
17 16,242 112 .0069
18 15,145 76 .0050
24 5,845 8 .0014
25 13,582 154 . 0113
104,121 715

6.4. Lots of bushing screws from a single supplier were inspected

100 per cent with the following results. Draw an appropriate control chart
 146 CONTROL CHARTS FOR ATTRIBUTES

for the data. What can you conclude? Under what circumstances would good
control be desirable here if it existed? If the sum of the p's (last column)
is divided by 8, do we get p = 2.3 per cent? Why not?

Table P6,4

Amount No. Percent

Lot No. in lot rejected defective

1 6,000 183 3.0

2 6,000 131 2.2
3 10,000 184 1.8
4 1,783 23 1.3
5 3,089 171 5.5
6 5,774 65 1.1
7 3,000 46 1.5
8 2' 724 85 3.1
38,370 888 2.3

6,5, The following data are from [1], Chapter 4. The data are on an
oxygen pressure switch, which has a bourdon tube. Depending on the pressure,
the tube opens or closes the electrical circuit through contact points. The
notes in the last column refer top charts on the separate types of defects.
Treating all the data as ~set of preliminary data, draw the control
chart. How many different sets of control limits are needed? What conclu-
sions can you draw? Is the simple average of the p's (their sum divided by
28) equal to p? Why not?

Table P6.5

Date No. items No. Fraction

produced inspected defective defective Other information

Feb. 1 2,500 300 .120 Points off and defective receptacle

2 2,314 214 .093
3 2,435 418 .172 Tube defective, close bars
4 1,217 117 .096
5 1,348 148 .110 Defective receptacle
6 1,229 142 .115 Miscellaneous trouble
8 1,384 134 .097
9 1,278 125 .098
Problems 147

Table P6.5 (continued)

10 1,197 100 .084
11 1,128 67 .059
12 1,135 85 .075
13 1 '391 105 . 076
15 1,212 85 .070
16 1,222 82 .067
17 1,167 80 .069
18 1,292 84 .065
19 1,082 71 .066
20 1,185 103 .087 Points off running high
22 1,260 110 .087 Points off running high
23 1,286 70 .054
24 1,204 54 .045
25 1,142 72 .063
26 1,216 66 .054
27 1,204 92 .076
Mar. 1 1,265 118 .093
2 844 124 .147 Defective receptacle, out of control
3 759 89 .117 Defective receptacle, out of control
4 745 75 .101 Defective receptacle, out of control
Total 36,641 3,330

6.6. Data on final inspection of carburetors on a line, preceding

those in Table 6. 1 by 12 days. (In the check-sheet form for inspectors,
there were listed 32 defects, each with three degrees of severity. Other
defects could also occur.) Two columns are shown for defectives d and
two for per cent defective. They represent (a) observed results and (b)
results with deletions for a discovered and removed assignable cause.
Plot an appropriate chart (a) for data as observed, and (b) after deletion
of defectives from the assignable causes. Comment.

Table P6,6

CARBURETOR NO. EVl LINE #7 DATE RUN Jan. 15

Sample Size No. of Defectives % Defective

* # * #
100 10 9 10% 9%
100 4 3 4 3
100 6 5 6 5
100 12 6 12 6
100 6 3 6 3
100 8 4 8 4
148 CONTROL CHARTS FOR ATTRIBUTES

Table P6.6 (continued)

100 10 3 10 3
100 12 6 12 6
100 8 2 8 2
100 7 3 7 3
100 3 0 3 0
100 4 1 4 1
100 3 1 3 1
100 4 1 4 1
100 4 0 4 0
100 10 3 10 3
100 7 2 7 2
100 5 1 5 1
100 3 0 3 0
100 6 1 6 1
100 8 1 8 1
100 10 6 10 6
100 4 3 4 3
52 2 2 3. 8% 3.8%
2352 156 66

*Before Removal of Assignable Causes

#After Removal of Assignable Causes

6.7. An interesting example from [1], Chapter 4 is the following.

With what factors might an assignable cause be associated? Make an

appropriate chart and offer a solution.

"Notes on Unprimed Case Visual Data"

(Small Arms Ammunition)

"These data represent visual defects overlooked by regular inspectors

in a 100% inspection operation at an intermediate stage in the manufacture

of cartridge cases. The purpose of this inspection operation was to get

rid of as much of the imperfect material as possible at an early stage,

rather than spend time and effort working it up only to throw it out at the

final inspection. The data cover the output of a single station, working
Problems 149

3 shifts a day, over a two-week period. There was a shift change at the
end of the first week.

"In the inspection operation the cases were fed mechanically to a

track, along which they were moved by two screws. Mirrors were placed

parallel to the track, affording the operator a view of the inside and

outside, including the head. The chief source of illumination was a Dazor

unit with two tubes; during the day there was additional light from large

windows nearby. It was the operator's duty to pick off defective cases

and drop them down a chute. Unless so picked off, the cases rolled to

the end of the track and fell off into tote boxes; these were dumped on

the conveyor from time to time by a spot checker.

"At the time these data were obtained much trouble was being encoun-

tered due to scale, a condition caused by faulty brass. Sometimes the scale

was conspicuous, with metal actually peeling away from the case; often,

however, it appeared only as a brownish speck the size of a pinhead, or

sometimes as a rippled or watery appearance of the brass.

"Twice during each shift, at times varying from day to day, a sample

of 200 cases was taken from the accepted material off each twin screw

machine. This sample was visualled by a specially trained inspector, who

recorded the number of defectives found. (If the initial inspection had

been perfect there would of course have been no defectives in the sample).

"The accompanying data are a typical set, and have the additional

interest that from them and the information contained in these remarks it

is possible to deduce an assignable cause of variation which was later

removed."
 150 CONTROL CHARTS FOR ATTRIBUTES

Table P6,7

CONTROL CHARTS FOR ATTRIBUTES

Operation: Unprimed Case Visual TSM No. 6

Data by SDB Sample size 200

Week Ending May 15

Mon Tue Wed Thu Fri Sat

A shift
Foreman: Pat
lst sample 3 3 5 3 5 2
2nd sample 2 0 0 3 1 3
B shift
Foreman: Rex
lst sample 2 3 5 5 3 9
2nd sample 3 6 5 10 6 3
c shift
Foreman: George
lst sample 4 5 3 5 2 5
2nd sample 1 3 2 1 5 3
Week Ending May 22
(shift change over week end)
Mon Tue Wed Thu Fri Sat

A shift
Foreman: Rex
lst sample 2 2 4 2 4 0
2nd sample 5 5 2 3 2 6
B shift
Foreman: George
lst sample 4 8 5 4 2 11
2nd sample 6 7 5 6 3 2
c shift
Foreman: Pat
1st sample 2 5 5 2 3 6
2nd sample 4 4 3 3 3 1

Total number of samples (2 weeks), 72

Total defectives found, 270
A shift 11 P.M. to 7 A.M.
B shift 7 A.M. to 3 P.M.
C shift 3 P.M. to 11 P.M.

6.8. The data below cover categories of defects found on inspection

of large aircraft. Using all data in a category of defects (as

assigned) as preliminary data run a c chart and comment. What c might you

suggest for future production and why?

 0
""'"'
Table P6.8 c::r
......
(1)
:3
til
50 51 52 53 54 55 56 57 58 59 60 61 L2 63
Alignment 7 6 6 7 4 7 8 12 9 9 8 5 5 9
Adjust 2 1 1 1 4 3 2 2 4 2
Workmanship 111 145 123 131 147 158 127 157 263 272 118 272 310 145
Tighten 1b9 155 146 139 134 160 138 161 69 62 135 41 79 149
Cello-Seals 16 23 19 11 16 25 20 25 7 25 17 10 ll 10
i'Iis1ocations 4 6 3 2 3 1 4 3 6 2 1
Incomplete 299 313 317 360 300 396 378 304 263 266 342 235 289 388
Clearance 108 91 115 116 93 111 108 86 97 125 101 123 110 127
Damage-I·Iuti1ations 16 35 23 25 23 37 36 48 18 28 47 25 28 28
Foreign Natter 13 19 18 18 17 16 22 8 9 8 17 8 10 14
Missing Rivets 8 16 14 19 11 15 8 11 21 12 23 16 9 25

Defective Rivets 4 3 3 1 1 1 5 2 2
Oxygen Leaks 1
Hyd. System Leaks 18 12 32 5 17 20 15 19 13 30 18 15 15 20
Incorrect Air Pressure

Replace 2 11 12 13 ll 19 11 10 18 2l 18 23 20 7
.....
U1
.....
 Table P6.8 (continued)
1-'
(J1
N
Plug Holes 2 3 5 0 5 6 5 4 6 9 9 6 5 9

Drawing Changes
Not Incorporated 1 3 1
TOTAL 780 632 839 854 780 975 879 853 798 873 863 791 897 932

64 65 66 o7 68 69 70 71 72 75 77 78 79 80

Alignment 8 15 b 4 13 7 0 15 6 6 lJ 7 13 4
Adjust l 1 3 2 11 5 1 5 2 6 4
Workmanship 176 295 114 247 125 239 161 234 141 268 261 95 96 ll6
Tighten 150 127 98 96 171 108 184 189 192 141 134 161 120 81
Cello Seals ll 14 15 4 6 10 8 9 7 9 14 5 10 l
n
0
Mislocation 3 2 1 4 4 6 2 4 l 3 6 2 z
-l
;o
0
Incomplete 369 265 312 245 363 243 249 267 264 219 160 296 304 237 t-
n
Clearance 106 109 75 88 107 76 96 119 103 94 71 79 80 64 5::
;o
-l
20 en
Damage-Hutilation 44 17 23 30 18 20 10 14 22 18 25 22 23 "'1
0
;o
Foreign Matter 18 10 17 14 10 14 15 2J 19 10 12 ll 4 25 ;,;.
-l
-l
Missing Rivets 15 9 9 14 ll 9 10 22 7 28 9 12 8 14 ;o
>--<
o:l
Defective Rivets l 2 c:
l l 4 1 2 -l
m
en
 >i
""
Oxygen Leaks 1 2 0
1 u
.....
(1)

Hyd. Syste~ Leaks 29 24 18 7 14 34 5 30 12 17 9 17 16 16 sVI

Incorrect Air Pressure

Replace 14 23 10 25 13 32 27 40 12 24 13 14 18 5
Plug Holes 9 8 7 6 ll 12 12 g 8 7 5 5 4 7
Drawing Changes
Not Incorporated 2 1 ll 1
TOTAL 955 920 706 782 881 819 809 972 788 851 722 734 706 599

......
V1
..,.,
 154 CONTROL CHARTS FOR ATTRIBUTES

6.9. The following data may be appropriately analyzed by a c chart,

if the conditions are met. Review the conditions needed and proceed
to chart if you think they might be met. Comment on control.

From each of thirty bales of dry bleached hemlock pulp, a

9" x 12" sheet was taken. The following table gives the
number of dirt specks greater than 0.1 sq. m.m. found on
these sheets.
Table P6.9

No. of No. of No. of

Sample Dirt S;eecks Sam;ele Dirt S;eecks Sam;e1e Dirt S;eecks

1 12 11 32 21 15
2 24 12 24 22 32
3 21 13 14 23 10
4 30 14 33 24 28
5 16 15 22 25 13

6 22 16 14 26 18
7 12 17 16 27 12
8 15 18 21 28 21
9 31 19 17 29 16
10 18 20 13 30 8

580

6.10. The following data are from Burns [5]. The data are on

outlet leaks in auto radiators found in the first assembling together

of the two radiator parts. Make a u chart, using as few different

classes of control lines as possible for the varying sample sizes.

What conclusions can you draw?

Table P6,10

No, tested No. outlet Average number per

Date n leaks c radiator u

June 3 39 14 .36
4 45 4 .09
5 46 5 . 11
6 48 13 .27
7 40 6 .15
10 58 2 .03
11 50 4 .08
12 50 11 .22
Problems 155

Table P6.10 (continued)

13 so 8 .16
14 so 10 .20
17 32 3 .09.
18 so ll .22
19 33 1 .03
20 so 3 .06
24 so 6 .12
2S so 8 .16
26 so s .10
27 so 2 .04
841 116

6.11. The data shown below are on three consecutive months for daily

samples of 30 farm implements. All types of defects included in final inspection

of implements. For example May 1, 84 defects on 30 implements gave 2.80 for u

per implement. Plot the May data and comment. What u might you carry on?

Table P6.11

May u June u July u

l 2.80 2 3.13 1 2.85

2 4.S7 3 3.55 2 3.10
s 4.13 4 4.30 3 1. 66
6 2.96 s 3.52 7 2.91
7 2.65 6 3.63 8 **2.65
8 4.54 9 3.38 9 3.17
9 3.3S 10 3.83 10 3.01
12 3.95 11 3.95 11 4.15
13 4.42 12 2.82 14 3.58
14 4.34 13 Down 15 2.49
IS 4.02 16 3.78 16 2.92
16 4.39 17 3.35 17 3.80
19 4.02 18 2.76 18 No work
20 4.03 19 2.86 21 No work
21 2.86 20 3.59 22 3.03
22 3.37 23 2.96 23 2.51
23 3.08 24 3.17 24 2.82
26 3.19 2S 3.75 25 2.44
27 t4.96 26 3.41
28 2.88 27 *4.56
29 2. 94 30 3.43
Total 77.45 Total 69.73 Total 47.09

t Section IV out of control

* Production reduced, jobs changed.
** Altered production
156 CONTROL CHARTS FOR ATTRIBUTES

6.12. Similarly to 6.11, but using the June data as preliminary data.
6.13. Similarly to 6.11, but using the July data as preliminary data.
6.14. If for a c chart c' = 2.5, is the probability of a point above the

center line equal to that for a point below? Does this have any bearing on

interpreting runs?

6.15. If on a p chart p- is indicated as .150 and the upper control limit

as .180, what is the approximate sample size?

6.16. How might you show an engineer that a sample having p .06,

provides no reliable evidence that p' > .03?

References
1. American Society for Quality Control, Standard Al-1971 (ANSI

Standard Zl.5-1971). Milwaukee, WI.

2. L. A. Aroian.1 What makes a control chart tick? Indust. Quality

Control 10 (No. 6), 38-43 (1954).

3. H. F. Dodge, A method of rating manufactured product. Bell System

Tech.~· 7, 350-368 (1928). Bell Telephone Laboratories Reprint B315.

4. H. F. Dodge and M. N. Torrey, A check inspection and demerit rating

plan. Indust. Quality Control 13 (No. 1}, 5-12 (1956).

5. W. L. Burns, Quality control proves itself in assembly. Indust.

Quality Control 4 (No. 4}, 12-17 (1948).

 CHAPTER 7

MISCELLANEOUS TOPICS IN CONTROL CHARTS

7.1. Control Charts for Individual Measurements, Moving Ranges.

In over-all data for management's use, a single measurement x may represent

one whole shift, day, week or month, or even a whole year. Consequently

we can have but few such values available, and moreover we do not want

to have to wait until a sample of four or five such results has accumu-

lated before plotting an x point . Often a decision is wanted "yesterday"

on the new result. In research too, each measurement obtained may be so

expensive or time-consuming that we cannot have many of them in our study .

Hence the need for the x chart, which we now describe.

The foregoing must not be allowed to give the reader the idea that

with 20 measurements and an x chart we can get as good results as we can

with 100 measurements, in 20 samples of n=S, using x and R charts. The x

chart is merely one way to try to get at least "something" from quite

meager data .

Naturally, limits for x's require knowledge of what mean ~ and crx to

use . If such standards are available we may easily proceed to set up

control charts, but if not available, as in the case analysis of past

data then we must use estimates . This is usually done through the use

of "moving ranges . " For the first moving range we use R1 = I x 1-x 2 1 , that

is, an ordinary range for a sample of n=2. Then for the second moving range,

we again use x2 but now in association with x 3 , thusly R2 = lx 2 -x 3 1.

157
158 MISCELLANEOUS TOPICS IN CONTROL CHARTS

Proceeding in this manner k observations x, provide k-1 R's.* Then in

the analysis of past data we will use x fork x's and R for k-1 RJs.

Let us consider an example. Table 7.1 gives 36 monthly per cents of

rejection of steel. Assuming as we do provisionally that all data come

from the same (normal) population, we then estimate p by

- =
X (7. 1)

Then we also compute R for as rational subgroups as are available, n=2:

(7.2)

and use this to estimate ax by

(7.3)

using d 2 for n=2.

We now come to the question of what limits to use. Recalling that we

are proceeding on a minimum of data, it seems to the author desirable to

use "two-sigma" limits for the limits for x 1 s. This is because, with points

occurring rather far apart in point of time, we can afford to increase the

risk of looking for an assignable cause when none is present. This will

then give us less risk of failing to look for an assignable cause when

one is present. Hence the author suggests use of the following from (7.1)

and (7.2)
-
X (7.4)

Limits
X
X ± 2· CR/1.128). (7. 5)

*Some writers obtain k R's by using Rk = /xk-x / but x ,xk is not in any
1 1
sense a rational subgroup. It can cause serious inflation of R if there

is something of a trend up or down in the x's.

TABLE 7 . 1.Monthly Per Cents of Rejection of a Grade of Steel . "Stee l"
Defects Found by Customer after Cold Drawing of Bars .

Per Cent Per Cent

Month Range R Month Range R
rejection x rejection x

January 4.4 July 4.1

1.5 .2
February 2. 9 August 4.3
1.3 1.0
March 4. 2 September 5.3
1.2 .0
April 5.4 October 5.3
1.2 4. 7
May 6.6 November .6
3.0 .2
June 3. 6 December .8
2. 6 .1
July 1.0 January .9
.3 2.2
August 1.3 February 3.1
.1 1.5
September 1. 2 March 1. 6
.0 .4
Octooer 1.2 April 2. 0
.2 1.5
November 1.0 May .5
1.8 2. 0
December 2. 8 June 2. 5
.8 .8
January 2. 0 July 3.3
1.0 1.1
February 3. 0 August 2. 2
1.2 2.4
March 1.8 s~ptember 4.6
.5 .s
April 2. 3 October 4. 1
1.0 .3
May 1.3 November 3. 8
3.9 .4
June 5.2 December 3.4
1.1
103 . 6 42 . 0
160 MISCELLANEOUS TOPICS IN CONTROL CHARTS

Then we use, as usual, for the R chart:

tR = R, UCLR = 3.267 R, (7. 6)

04 being 3.267 for n=2. (One could use two-sigma limits her~ too, if

desired. How?)

Now from Table 7.1 we find

{x x= 103.6/36 = 2.88

R 42.0/35 = 1.20

Limits 2.88 ± 2(1.20/1.128) .75, 5.01

X

UCLR = 3.267(1.20) = 3.92

The charts are plotted in Figure 7.1. Considerable lack of control shows

up on the x chart and even some on the moving range chart. There the

highest point comes from the sharp drop from October to November.

Some disadvantages of the x chart relative to the x chart are the

following: (1) The x chart is less efficient than the x chart, in the

sense that it takes more individual measurements on the average to get an

indication of the presence of an assignable cause, unless the shift in p

is extremely large, say, 3o. (See Section 7.7.) (2) The x value does not

summarize a set of conditions as well as an x does, it being an average of

several x's. (3) If the distribution of xls is nonnormal, this will tend

to distort the interpretation of the control limits, on the x chart. On

the other hand by the central limit theorem the distribution of x's is

likely to be quite normal even for small n 1 s.

The x chart does have the advantages (1) of being up-to-date at the

time the last x is plotted, (2) of having no arithmetic needed before

plotting each x, and (3) of being very easy for shopmen to understand.

Also comparison of an x chart with specifications, if any, is direct.

7.2. Tool Wear and Other Gradual Process Changes 161

PIG. 7.1. Control charts for individuals x's and for moving
ranges R. Monthly per cent of rejection of a grade of steel.
Data of Table 7.1. Note that the point for the moving range is
plotted between the two months involved. Reproduced with permission
from I. W. Burr, "Engineering Statistics and Quality Control",
McGraw-Hill, New York, 1953, p. 266.

7.2. Tool Wear and Other Gradual Process Changes. Let us consider

the following adaptation of x control charts. When we take a sample of

product, periodically from the most recent production, these x's may be

expected to contain random variation only (unless there has been a very

recent resetting of the process average~). Hence ranges R from such

samples and their average, R, reflect chance causes only. When the R

chart shows good control we may estimate the short term o by R/d 2 . Then
------ X

x± 3(R/d 2) = x± 30 x are approximate limits to the x distribution for the

most recent production. Now unless the process average p is quite tightly

controlled, x± 3(R/d 2 ) will not be the x limits for the total production

over, say, a shift. There will be in fact some o(l ong term ) which exceeds

o The former of these will include shifts in the process mean

(short term)"
as well as random variation. The whole set of x's for the shift probably

will lie within x± 3o(long term)"

162 MISCELLANEOUS TOPICS I N CONTROL CHARTS

It is the last set of limits which must meet specifications for x's .

Now,_again, if6a specified tolerance ~ U-L, then the process

(short term)
mean must be tightly controlled and conventional x and R charts are in order.

However, cases do occur where U-L > 8a(short term)' and it would be possib le

to let the process mean~ vary around a bit. In fact if U-L~ Sa( short term) ,

then as long as ~ lies between U-3a(short term) and L+3a (short term) ~u-S a (short term)
we will still be having at l eas t 99.9 % of pieces between the limits . Thus ther e

is a band 2a(short term) wide within which p may be permitted to vary, in this
case. See Figure 7.2.

Let us in the following understand that a or a R/d 2 refer to short

term variation, as usual. Now when the above situation of U-L > 8a is

present, how may we proceed, if we do not wish to maintain tight control

on the x chart? Let us proceed to illustrate by an example.

Consider the data given in Table 7.2. Looking at the x column we

can s ee quite a definite trend of x's with a sharp break downward after

12:30 July 27.

la

U-3a

U -Scr ~L+3cr

U-Ba=L

FIG. 7.2. Two normal curves drawn to show the permissible variation
in process mean~. when a(short term) ~ cr is (U-L)/8.
 7. 2. Tool Wear and Other Gradual Process Changes 163

Table 7.2. Outside Diameters, of Spacers, Samples of Five Every 30 Min.

From an Automatic Machine.Specifications: .1250 In., + .0000 In., - .0015

In. Readings in . 0001 In. Below .1250 In .

Date Time xl x2 x3 x4 x5 X R

July 27 8:00 - 8 - 9 -7 - 8 -8 - 8.0 2

8:30
9:00
-8
- 6
-
-
9
6
-
-
7
8
-
-
7
7
-
-
6
7
-- 6.8
7.4 3
2
9:30 - 4 - 6 - 5 - 6 - 6 - 5.4 2
10:00 - 7 - 5 - 4 - 5 - 7 - 5.6 3

11:00 - 5 -6 - 2 - 3 - 4 - 4.0 4
12:30
1:15
-
-
2
10
-4
- 11
-
-
2
12
- 4
- 11
-
-
5
10
--10.8
3.4 3
2
2:00 - 10 - 12 - 13 - 12 - 11 -11.6 3
2:45 - 11 - 11 - 10 - 11 - 9 -10.4 2

3:15 -8 - 10 - 9 - 10 - 11 - 9.6 3
4:00
4:30
-
-
9
10
-
-
10
11
- 11
- 7
-
- 11
10 -9
- 8
-- 9.8
9.4
2
4

July 30 9:00 - 7 - 8 - 8 - 5 - 9 - 7.4 4

9:30 - 9 - 10 - 7 - 6 - 7 - 7.8 4
10:00
11:00
-
-
8
5
-
-
6
4
-
-
8
5
-
-
7
6
-
-
8
7
-
-
7.4
5.4
2
3
12:15 -7 - 8 - 6 - 7 - 7 - 7.0 2

12:45 - 7 - 6 - 7 - 4 - 5 - 5.6. 3
1:15 - 3 - 3 - 4 -- 57 - 4 - 3.8 2
1:45 - 3 - 4 - 3 - 6 - 4.6 4
2:15 - 6 - 3 - 4 -2 - 6 - 4.2 4
2:45 - 1 - 3 - 2 - 4 - 4 - 2.8 3

3:30 -9 - 6 - 5 -4 -7 -- 5.6
6.2 5
4:00
4:30
- 4
-5
-- 77 -
-
8
6
- 4
- 7
-5
- 6 - 6.2
4
2

July 31 8:00 - 9 - 10 -7 - 11 - 10 - 9.4 4

- 185.8 81

The quality control man, having been recently trained in a short

course was quite surprised when he ran an ordinary -

X chart and an R chart.

These had

x =-185.8/27 = -6.88 R = 81/27 = 3.00

Limits- = -6.88 ± .577(3.00) = -8.61, -5. 15, UCLR 2.115(3.00) 6.34.

X
164 MISCELLANEOUS TOPICS IN CONTROL Cr~RTS

He was gratified to find the R chart well in control, but the i chart had

13 points out of control. There must be assignable causes. At first he

floundered around quite a bit, but inquiry and study revealed that it was

due to tool wear. (As a tool, cutting an outside diameter, gradually wears,

its point grows farther from the center line, giving slightly larger

diameters.) Thus the assignable cause was known. Was it necessary to

eliminate this assignable cause by very frequent resettings of the tool?

This last is the basic question. Let us see. Short term variation is

estimated by

ox R/d 2 = 3.00/2.326 = 1.29

But in the coded units U=O, L=-15, so the tolerance is U-L=l5. Thus

(U-L)/cr 11.6, so we have a lot of latitude for~ to varf within, in fact

X

at least from -15 + 3(1.29)= - 11.13 up to 0-3(1.29)= - 3.87. (As long as

~ lies between, there will be a very small fraction defective.) Therefore

quite a bit of tool wear was permissible in this example, because specifi-

cations were lenient.

7.2.1 Slanting Control Lines for x's. An adaptation for ~'s which

is of much use in such a situation is the use of slanting control lines.

We plot the i's and R's as in Fig. 7.~, analyzing the R's as usual. If

they show good control we are justified in estimating crx. Next we look

at each trend of x points and try to draw the best straight line through

them.~' This represents our guess as to the way the true ~ runs with the

i•s varying around this trend line. To test as to whether the tool wear is

proceeding along in control, we put in control limits at 3&x = A2Rabove

and below our slanting trend line, measured vertically. If the x points

~'Fitting such a line by the objective least-squares method is hardly

justified.
7.2. Tool Wear and Other Gradual Process Changes 165

lie between these control limits then the tool is "wearing in control."
This is the case in Fig. 7.3. There, 3cr-X =A 2R = .577(3.00)= 1.73. (Com-
pare this with 30x= 3R/d 2= 3.87.)

Now at what a level for ~ should we aim to start a run? This is

L + 3&X minimum safe process average (7. 7)

Then as x points accumulate we can draw in the trend line. At what level

should we plan to reset the process mean back down again? This should be

done just as soon as the trend line hits

U - 3& = maximum safe process average (7. 8)

X

In Figure 7.3, the trends seem to have been handled quite well, perhaps

~a ___________________ _
R
•
3.00 • ••• •• • •
• •• • • • • • • •
0~----------------------------------
FIG. 7.3. Control charts for outside diameter of a spacer. Samples of
five from an automatic. Measurements in .0001 in. from .1250 in.
Specifications .1235 in. to .1250 in. Trend control lines for x's.
Reproduced with permission from I. W. Burr, "Engineering Statistics and
Quality Control", McGraw-Hill, New York, 1953, p. 269.
166 MISCELLANEOUS TOPICS IN CONTROL CHARTS

being let run a bit longer than to 3a X

from U. (But it must be remembered
that keeping ~ at 30x away from the specification limit is being really

quite conservative. Perhaps one can run within 2.50 or even 2& of U
X X

occasionally.)

For comparison the author calculated s~2.64 for the 135 x's in the

whole Table 7.2. Thus we have

&
(long term) 2 · 64 • 0 (short te~- 1 · 29

the latt~r being only a half of the former. We see that the tolerance of

15 is about 6&( 1ong t erm)'

Careful following up of x points outside of the slanting control limits

can be fruitful in finding assignable causes in the process.

Besides resetting when the trend line hits the maximum safe process

average, we should also reset if a single x goes out of specifications.

Trends can also be downward, as for example for an inside diameter.

Other examples of trends are (1) die wear, (2) grinder or polisher wear,

(3) human fatigue, (4) room temperature or humidity, whether or not air-

conditioned, (5) viscosity in a chemical process, (6) moisture content

of lumber or a plastic. The author recalls hearing one executive claim

that such trend charts had doubled the tool and die life in his division!

7.2.2.* Modified Limits for ~'s. A practice associated with the

approach just described is the use of "modified limits" for ~~ s, The general

idea is to place modified limits for ~!s on the chart instead of ordinary

control limits, when the tolerance U-L is substantially greater than

6cr(short term)' that is, when we can afford to let r vary around a bit

and still meet the tolerance limits for x's, Land U. Such usage gives up

on trying to find assignable causes affecting ~·

Let us reconsider the example of Sections 7.2 and 7.2.1. When we are

near the upper specification limit U,we have to be concerned with the "max-
 7.2. Tool Wear and Other Gradual Process Changes 167

imum safe process average"=- 3.87. This is as high as iJ. can be permitted

to go. Now if iJ.=- 3.87, where will the x's lie? They will be between

- 3.87 +A
- 2
R =- 3.87 --~ 1.73 =- 5.60,- 2.14. Suppose now that an x point

goes above -2.14. See Figure 7.4. Then surely ~ is above -3.87. On the

other hand, if an; point lies below -2.14 does this give assurance that

1-L is at or below -3.87? Not at all. In fact ~might be at -2.14 or even

as high as -1 and still yield such an x point. Thus the -2.14 line, which

we call "modified limit A", gives rather unsymmetrical protection against

the two errors: stopping and resetting the process when we should not, and

not stopping to reset when we should.

The author recommends that a better balance is obtained by "modified

limit B" found by again coming down from U by 3;x (to -3.87}, then back up

by only 2~- (from -3.87 to -2.72}. We then have

-- X

Mod. limit B
-2.)2

-3.87
Max. &afe process
aver~ge

-7.74 ,_..,....,...,.....,---..,.....,.,.-.,-.L-...,..,.-7:""':'--......,,....,....,--..~.-..,.,....,.,....,____,,--,.-=-'
X cAistribution X ~istr. X distr. Xdistr. X distr. X distr-:
(al <bl (cl

riG. 7.4. Three process levels to illustrate the operating characteristics

of two different modified limits. This is in relation to the data of Table
7. 2. Reproduced with permission from I. W. Burr, "Engineering Statistics
and Quality Control", McGraw-Hill, New York, 1953, p. 272.
168 MISCELLANEOUS TOPICS IN CONTROL CHARTS

Modified limit A U- 3R/d 2 + 3[R/(dz"'';)] (7. 9)

Modified limit B U- )R/d 2 + 2[R/(d 2/r;')] (7. 10)

Figure 7.4 shows in section (a) the x distribution if~ is at -3.87 and the x-

distribution centered there too. Almost no x's will be above limit A and

only about 2% above limit B, in each case giving a false warning that ~

is too high. Now in section (b) both distributions center at -2.58 giving

2.28% of x's outside U. Under this condition much less than half of the x's

will exceed limit A,providing a warning; but over half are above limit B.

Finally in section (c) where nearly 16% of x's are beyond U, about 7% of

the time x will not be above limit A under this alarming situation, while

only .7% of the time will an x nQ! be above limit B. Thus limit B provides

much better protection than limit A, at little expense on the a risk. See Fig. 7.5.

Modified limit B is much like those used in England, whereas limit A

is often used in the USA. Neither is safe unless the R chart shows control,

Limit B can be useful in short runs for an immediate check of setup.

Reference [1] gives quite a complete discussion of this subject,

along with operating characteristic curves for various lines.

7.3. Charts for x and s, Large Samples. If it is desired to run

control charts for x and s for sample sizes larger than those given in Table

VI, then we can treat c 4 as 1' which it is approaching. Also

o
S
~ o
X
;;2; or c 5 = l/l2n (7 .11)

Then using these in (5. 9) and (5. 16) gives

(7.12)

Also by (5 .11)

A ,; 3/rn. (7.13)
3
Then we may use these results in the formulas in Section 4.10 for the

charts.
7.4. Charts for x- and R or s 169

1.00 '
''
''
'
''
.80 \
I
I
I
\
I
\
.60 \
\
\
Pace
I
'B
I
:40 \
\
\
I
\
\
I
\
.20 \
\
\

''
''
.00
-4 -3 -2 -1 O•Uf!.
,003 ,010 .026 .061 .122 .219 p'•Pix>Ul

FIG. 7.5. Operating characteristic curves for Type A and Type B modified
limits. Probability of acceptance versus process average~ and fraction
defective p 1 •

As was true for p charts, we may let the n's vary from their average

n, by about 30% of ii. Then for any points close to a limit we can use

the actual n for more accurate comparison.

7 .4. Charts for x and R or s, with Varying Sample Sizes. Sometimes

in laboratory or research data we wish to test for homogeneity of level and

of variability by control charts. This is the equivalent of testing the

hypothesis that all samples come from the same (normal) population with

unknown ~and ox. In analyzing such data for control we seek the best

estimates of ~ and ox' assuming the hypothesis is true. For this we may

proceed as shown in f2]. For~ we use a simple weighted average

~ X (7 .14)

For ox we use the weighted average

 170 MISCELLANEOUS TOPICS IN CONTROL CHARTS

elR1 + e2R2 + ... + ek\_

aX
(7 .15)
f1 + f2 + ... + fk

where ei' fi are for Ri 1 s sample size ni' in Table 7.3. Or, if the sample
variabilities are measured by s, we use

gl sl + g2s2 +. · .+ gksk
& (7 .16)
X hl + h2 + ... + ~

Table 7.3. Factors for Best Linear Estimates of Population Standard Deviation

o for Sample R's and s's. [2)

Sample For Ranges For Standard Deviations

Size Equation (7 .15) Equation (7.16)
n e f g h

2 1.55 1. 75 2.20 1. 75
3 2.14 3.63 4.13 3.66
4 2.66 5.48 6.09 5.61
5 3.12 7.25 8.07 7.59
6 3.52 8.93 10.06 9.57
7 3.90 10.53 12.05 11.56
8 4.24 12.06 14.04 13.55
9 4.55 13.52 16.04 15.55
10 4.84 14.91 18.03 17.54
11 20.03 19.54
12 With sample sizes 22.03 21.53
13 above 10 divide 24.03 23.53
14 into subgroups. 26.02 25.53
15 28.02 27.53
16 30.02 29.52
17 32.02 31.52
18 34.02 33.52
19 36.02 35.52
20 38.02 37.52
 7.4. Charts for x and R or s 171

Then the next step is to treat our ~ and aX

as though they were given

standards~ and ax' and use appropriate formulas (4.5) to (4.9). In this

latter work we can use an fi if the n's are close together. For example

use n=5 if fi is near 5, and the nts are 4, 5 or 6.

Let us give an example from (2], see Table 7.4, for the x's, x's, R's

for five sets of lamps.

x = 158.82/16 9.93 = ~

Then by (7 .15)

a 3.52(.47) + 2.14(.48) + ••• + 1.55(.92) = 4.9032 = 249

X 8.93 + 3.63 + ••• + 1.75 19.69 .

Now the center line for all x's is i = 9.93, the limits vary as n varies

using x ± Acrx. Meanwhile both the center line and limits for R vary with n,

using d2&x and D2°x respectively. (D1 =0, up through n=6.) We then have

n=2 n=3 n=6

Center { x: 9.93 9.93 9.93

Lines R: .281 .422 .631

Table 7.4. Samples of Lamps for Luminous Flux Measurements, in Lumens

per Watt. From (2]. Various Companies.

Company Measurements n x R e f

1 9.47, 9.00, 9.12, 9.27, 9.27, 9.25 6 9.23 .47 3.52 8.93
2 10.80, 11.28, 11.15 3 11.08 .48 2.14 3.63
3 10.37, 10.42, 10.28 3 10.36 .14 2.14 3.63
4 10.65, 10.33 2 10.49 .32 1.55 1. 75
11< 9.54, 8.62 ..1.. 9.08 .92 1.55 1. 75
Total 158.82 16

1< Held in stock for one year.

 172 MISCELLANEOUS TOPICS IN CONTROL CHARTS

Control{ x: 9.40,10.46 9.50, 10.36 9.63,10.23

Limits R: --,.918 --,1.085 --,1.264

Looking at Figure 7.6, we see all i points on a control limit or outside the

limits indicating marked differences between companies. There is also one

high range point indicating excessive variability for the lamps held for one

year.

7.5. Different Types of Samples. The success of anyone who uses statistics

in science or industry will depend in large measure upon how he collects his

samples. This section describes some types of samples.

12

--~--------

xlo r--------------------------------------------------------
.
8

l
l\ ---------<~>---

0 •
2 3 4 l(after
company year)

FIG. 7.6. Control charts for laboratory tests on luminous flux (lumens per
watt) for samples of lamps from various companies. Data in Table 7.4.
7.5. Different Types of Samples 173

7.5.1. Rational Sample. The ideal way to sample is to take a rational

sample or subgroup. By this we mean that all of the pieces or measurements

have been taken under the same conditions, so that what variation there is,

within the sample, is all attributable to chance causes only. Thus the pieces

or tests were made on the same machine (under one setting), by the same people,

with the same methods, and at approximately the same time. Succeeding samples

will be taken under differing sets of conditions. Whenever a point goes out

of the control band, we try to find what condition was different for the out-of-

control sample than for the other in-control samples. In this way one can hope

to traek down the assignable cause of the off performance. The success of the

program depends in part upon careful records of conditions in the production

process and measuring techniques.

An illustrative experiment often used in short courses in statistical quality

control, is to use five bowls, wh5.ch represent five machines, or five spindles,

heads, stations or orifices for one machine. Suppose that specifications are

-5 to +5. Then distributions of Table 4.1 are used in five bowls as follows:

three of A(~=O, cr=l.7), one of B(~=2, cr=l.7) and one of C(~=O, U=3.5). A

sample of n=S is taken from each bowl at the first "time," a record being

kept of which bowl is used for each sample, and x's and R1 s plotted, in such

a way as to dist:i.nguish the sources of the data. Then at the next "time"

five more samples, each of five are taken. The procedure continues until five

or six sets of samples are taken, all being plotted on the same charts. Then

x, R and control limits are found for all 25 or 30 samples taken together. If

in control, the lines are extended and new samples drawn. But usually either
distribution B or C or both will provide at least one point out of control. Then

appropriate "actionnis taken to "reset the process" or to "overhaul" the process.

Data from the out of control bowl are eliminated,center lines recalculated, etc.
174 MISCELLANEOUS TOPICS IN CONTROL CHARTS

Such an experiment can be quite realistic and a stimulus to thinking, illustrating

many of the principles of control chart operation.

7.5.2. Random Sample from Mixed Product. A very common type of sampling in

industry is to take a sample of n at random from the total output of a machine or

process. Such a sample is often taken from a conveyor belt or a tote pan of parts.

These parts have been made from different heads, spindles, mold cavities, etc.

But when such a sample is drawn we cannot tell from which head we have each part.

We have simply sampled the overall output which includes not only chance variation

but also 3·7 acsignable-cauoe diffcrencecbetween heads.

Now if such samples show good control and the calculation of x± 3R/d 2 shows
the natural process limits to lie about on or within specification limits for

x 1 s, then we are on quite good ground and do not need to dig more deeply into

the process. But if the natural process limits from the x's do not meet

specifications, then we need to analyze more deeply into the process and to

try to compare the output from the different heads or spindles, etc.

The author recalls being asked by a man from a can company what the prob-

ability is for a random sample of 12 cans, from the total output of a can-

making machine with four heads, to contain at least one can from each of the

four heads. He said many answers were arrived at by his associates, but all

were different! The probability proves to be .87345; rather lower than was

hoped for.

7.5.3. Stratified Sample. Another type of sample often used is the

stratified sample. In it we include one part from each head or spindle on the

machine. For example, on a five-spindle automatic machine, we can take the

next five pieces to be produced. Then we are sure to have one part from each,

but unless we are very careful, we cannot tell which part was made by spindle

number one, which by number two, etc. Now such a sample does even better at

estimating the overall process mean, than does the random sample from mixed
 7.5. Different Types of Samples 175

product. But it has the same weakness of lumping into one sample, parts and there-
fore x's from potentially different populations (each spindle). Thus again the
ranges, R, will include potential assignable causes as well as chance variation.

Hence, again Rwill be inflated. Moreover such sampling does not permit com-

parison of spindles.

But now if we do make a record of which spindle produced each part, then we can

plot rational samples for each spindle. Then using such rational samples we have from
Ra valid estimate of chance variation, and can moreover distinguish between spindles.

Let us illustrate some of these principles with experiments. In Table 7.5

are shown the results of drawing 30 stratified samples. The x 1 measurement

is of a part from spindle 1, x 2 from spindle 2, etc. Then at, say, 8:15, each

spindle contributes an x to the sample. Thus we have as ~ in the table, 30

stratified samples. Figure 7.7(a) shows the x and R charts. For these samples

we have

i = -3.4/30 = -.11 i = 263/30 = 8.77 UCta= 2.115(8.77) = 18.5

Limits -
X
= -.11 ± .577(8.77) = -.11 ± 5.06 = -5.17, + 4.95

As seen in the figure, control is perfect. But in point of fact it is entirely

too good. On the x chart there is no x closer than a third of the way from

center line to limit, that is, all x's lie within+ lcr-. This results from
- X
the ranges having been substantially inflated by including spindle-to-spindle

differences as well as chance causes. Also the R chart exhibits rather too

perfect control. This is because we always have an x from the highest average

spindle and one from the lowest in the same sample thus forcing rather unnatural

uniformity in the R's.

Next in Figure 7.7(b) we see the same data plotted in rational samples.

These samples are shown in columns of Table 7.5 with the x and R given. Thus

we find 6 x's and 6 R's in the x 1 column for spindle 1, etc. Now

i = -.11 R = 119/30 = 3.97 UCLR = 2.115(3.97) = 8.40

 176 MISCELLANEOUS TOPICS IN CONTROL CHARTS

Table 7.5. Samples of n = 5. Five Populations of Table 4.1, Used for x1 to x5

Respectively. Stratified Samples Horizontally, and Rational Samples Vertically

in Each Block. "Specifications" are ± 5.

Time Date xl x2 x3 x4 x5 x R
8:15 2/11/73 + 2 - 51 ++ 11 0 + 2 o.o 7
17 0 - 5 - 1 + 6 +1.0 7
19 + 1 - + 01 - 41 + 6 -0.4 11
21
23
+ 2
+ 2
+ 1
--3.26 +0.2
-
- 2 +-0.82
+ 5
+ 4
+1.6
o.o
6
10
x +1.4 +4.6
2 7 3 6 4
-
R
9:20 + 2 5 - 3 0 + 4 -0.4 9
22 + 1 - - 2 -3 + 4 -1.0 9
--
5
24 + 2 4 0 0 + 3 +0.2 7
26
28
x
+ 1
0
+1.2
--4.64
5 + 3 + 1
- 1
-0.6
-
1
-0,6
+ 4
+ 6
+4.2
+0.8
o.o
9
10

R z 1 6 4 3

-- --
10:09 + 2 5 + 1 1 + 5 +0.4 10
11 0 5 + 1 2 + 2 -0.8 7

--
13 + 4 2 + 2 - 2 + 5 +1.4 7
15
17
+ 2
+ 2
6
4 -
+ 1
+ 2
--
3
5
+ 7
+ 2
+0.2
-0.6
13
7
~ +2.0 -4.4 +1.4 -2.6 +4.2
R 4 4 1 4 5
10:57 + 3
+ 4
7 -- + 1
+ 2
-
2 + 6
0 + 2
+0.2
+0.8
13
8
-
59 4
11:01 + 2 5 - 1 -
2 + 6 o.o 11
03 + 2 - 0 - + 2 -1.2 7
- -
5 5
05 0 5 0 5 + 4 -1.2 9
X +2.2 -5.2 +0.4 -2.8 +4.0
R 4 3 3 5 4
11:47 0 7 - + 2 -
2 + 5 -0.4 12
49 + 3
+ 2
3
4
-
-
0
0
-
3 + 1
+ 2 + 6
-0.4
+1.2
6
10
-
51

--
53 0 2 4 + 1 + 1 -0.8 5
55
x
+ 2
+1.4
6
-4.4
-
-0.6
1 -
1 + 3
-0.6 +3.2
-0.6 9

3 6
-- -
R 5 5 5
1:03 2/11/73 0 8 1 -
3 + 6 -1.2 14

---
+ 2 + 2 2 + 1 -0.2 6

---
05 4
07 0 7 + 2 4 + 2 -1.4 9
+ 2 4 + 3 4 0 -0.6 7
-
09
11 + 1 4 1 0 + 4 o.o _J!
x +1.0 -5.4 +1.0 -2.6 +2.6 263
R 2 4 4 4 6
~x +9.2 -27.2 +1.8 -10,0 +22.8 -3.4
L:R li 24 23 28 27 119
 7.5. Different Types of Samples 177

+5 UCL~-- - - - - - ------ ---------

LCL-
-5 - ~ .. - - - - - - - - - - - - - - - - - - - - - - - - - - -

QL--------------------------------------
(a l Stratified samples

+5
• • • •
iO
ucL-
tx
-~--------~----~--------
0 0
+
0
•
0
+ ·-
+
+A +A
LCLx
------- -6---------c-
•
• • • •
-5
•

.
10 !;!C_LR_
R
t
• +
• •
0
+ • ... 0
0
0~--------~0~·~--------~+ ______________________________
(b) Rational samples

FIG. 7.7. Two analyses of data of Table 7.5. In (b) the rational samples
have each spindle pictured by a different symbol.

Limits-·""
X
-.11 +
-
.577(3.97) = -.11 ± 2.29 = -2.40, + 2.18.

The x and R points are plotted with different point symbols in rotation, a set

of five at each basic time. The R chart shows perfect control with typical

appearance. But the x chart shows bad control, with points outside the control
178 MISCELLANEOUS TOPICS IN CONTROL CHARTS

band in both directions. Examining more closely, we find all six x2 •s below
the LCLx and also three of x4 •s. Also all six x 5 •s are above UCLx. Thus

we can readily perceive substantial differences in spindles as to ~vel (but

the variabilities are apparently homogeneous). Action can be taken to adjust

the spindles so as to meet specifications of± 5. But rather tight control of

~must be exercised, since a is close to 1.7 and thus is about a sixth of

the total tolerance of 10.

Actually the x 1 1 s were from population B; x 2 1 s: G; x 3)s: A; x4 1 s: F and

x 5 1 s: E, of Table 4.1.
The author ran a companion experiment in which all five populations were

dumped into a single bowl and thoroughly mixed. Then random samples were taken

from this "conveyor belt of mixed product." The naive person might expect to

have one from each spindle in such a random sample. Actually none of the 30

samples had such a balanced composition. One even had four of the five from

the same spindle. We shall not list the data, but did have:

X = +6.4/30 = +.21 R = 221/30 = 7.37 UCLR = 15.6

Limits -
X
= -4.04, + 4.46,

The x's ran -2.6 to +3.4 and were well in control as were the R's, 3 to 12.

This would in practice justify estimation of the process population:

+.21 8 : 7.37/2.326 3.17

X

yielding x limits of

+.21 ± 3(3.17) = -9.30, + 9.72

If one combines the five populations A, B, E, F and G of Table 4.1, the

overall distribution has ~-o, cr=3.32,and it runs from -9 to+ 9. If these

had been the specifications we would not have to analyze the individual spindles

and adjust them. But with specifications of ± 5, we do need to analyze the

individual spindles.
7.5. Different Types of Samples 179

Other plottings of the rational samples of Table 7.5 could be made.

We could plot all six samples of x 1 's first, then the six of x 2 's etc. This

gives a fine picture and is quite feasible when analyzing past data. But

plotting in rotation may be better when extending into the future.

Or one can make a "group control chart, '' For it one plots for a given

time the highest and lowest~ points and highest and lowest R points. So

instead of plotting five x 1 s and fiveR's at, say, 8:19a.m. we only plot

two x's and two R's, the highest and lowest. Then near the highest x point

at 8:19 we place a little number to indicate which spindle had the highest x,

and similarly for the lowest. In Table 7.5 we would plot x 5 = -~.6 and x2 = -3.2

and put respectively a "5" and "2" near the two points. Similarly the largest

R was R2 and lowest R1 . Only these two R's are plotted, with the "2" and "1".

Now for the x's there would be a string of 51 s by the high points indicating

that spindle 5 is set high, and simila~ly a string of 2's by the low x's.

But for the R's, the spindles "take turns" being high and low indicating no

particular differences in variabilities. Group control charts are often used

with greater numbers of spindles, heads, etc. and n=2 at each.

7. 5. 4. * Stratification Control Charts. There is another approach

which can be used on stratified samples, if a record is kept of which

observation goes with which mold cavity or head, etc. The basic approach

is to find the average per cavity, over, say, 20 or 25 stratified samples.

Then the next step is to substract from all of the cavity 1 results, the

amount xl - x, where xl is the average of all cavity 1 x's i.e. the xl's.

The adjusted x 1 values then average x, (the whole set of adjusted x 1 's),

not x1 . Adjust all the x2 1 s by substracting x2-x, etc. Then we use an

ordinary x and R chart set-up for the adjusted data. Since the sum of the

adjustments is 0, the row means (means at each time) are unaffected. But
 180 MISCELLANEOUS TOPICS IN CONTROL CHARTS

the ranges will in general be much smaller, since they are no longer inflated
by cavity-to-cavity differences, these having been removed by the adjustments.

This approach was developed in (3] and (4].

7.6.* Relative Efficiency of Different Types of Charts. One of the

reasons for using an x chart rather than an x chart is that the former will

average fewer measurements before giving an indication of a moderate shift

in the process average. (But if the shift is extremely large, the x chart

will be quicker.) Suppose, for example, that a process which has been in

control is suddenly affected by an assignable cause which increases ~ by lox

How soon can we expect an indication of this on an x chart, and how soon on

an x chart with n=4? See Figure 7.8. In the left hand section showing the

original process and limits for x and x, we see that in either case the

UCLy_
- - - - - n=l-

UCL;:
n·4
1-----+----h~--'-'-''-:::;o.c---- ,)shift in iJ.

Original process Process 01fte,. a shift

of lcrx in 1-L

FIG. 7.8. Distributions of x and x

(n=4), with given~ and cr, and 3cr upper
control limits, in left hand section. Right hand section shows distributions
after ~ increases by lcr . Probabilities of a point out of the respective
upper control limits ar~ .0228 and .1587. Reproduced with permission from
I. W. Burr, "Engineering Statistics and Quality Control", McGraw-Hill,
New York, 1953, p. 280.
7.6. Different Types of Charts 181

probability of a point outside is .0026 (3 sigma limits, standards given).

In the right hand section, we see the distributions after p is increased by lcrx,

with the original upper control limit lines extended. The probability of an

x point lying above the original UCLx is now .0228, while that for an x point

above the original UCLx is .1587. These are from respectively

IJ + 3(crx/14) - (IJ + ax)

z = +2 z = +1
cr X /14

Thus there is a much greater chance for an x point to show out of control,

than for an x point. But ,remember that the x chart has four times as many

opportunities to have a point out, than does the x chart, because the
-
latter requires four x's to make one x. How can we obtain a fair comparison?
The solution is to find the "average run length" or ARL, in terms of

the average number of x 1 s before a warning. Let us consider the xls first,

and find the average number till the first point goes outside the limits.

Let the probability be p' and its complement q'. Then the probability of

the first warning on the first point is

P(l) = p',

and the probability of the first point out occurring on the second point is

P(2) = q'p',
and in general the probability of the ~~ po~~~ out occurring on the i'th

point is
i-1
P(i) = (q') p' (7 .17)

(This distribution of number of trials is called the "geometric distribution.")

Hence to find the ~verage number till the first point goes out we need

E(n) = lP(l) + 2P(2) + 3P(3) + ... + iP(i) + ... , (7 .18)

or, say,

~ = lp' + 2q'p' + 3q' 2p• + 4q 13 p• + .... (7 .19)

n
Multiplying (7.19) by q' gives
182 MISCELLANEOUS TOPICS IN CONTROL CHARTS

(7.20)
Subtracting (7. 20) from (7. 19) yields
2
(1-q')~n"' p' + q'p' + q' p' + q'3p' + ..•
,2
p'(l+q' + q + q 13 + .•. )

But by division 1/(1-q') =1 + q' + q' 2 + q• 3 + ... ,and hence

\1-q')~
lc
= p'[1/(l-q')J.
Since (1-q') = p' we have

p'~ p'/p'.
n
Thus
J.l = 1/p' (7. 21)
n

Hence the average number of x 1 s to receive a warning for the case in Figure 7. 8, is

~n = ARLx = 1/.0228 = 44.

On the other hand the average number of xs
1 to receive a warning on the x chart

is

ARLx = 1/.1587 = 6.3.

But this is 6.3 sets of 4 x's each or

4ARLx = 4(6.3) = 25
Thus on the average, it takes about 25 x's made up into x's to receive a

warning point USing xJs, but 44 x's for the X chart.

Note that this is for x vs. x charts for the case "standards given" and

when n=4 and the shift is lcr

X

If the jump is kcrx and the sample size is n, then generalization of the

foregoing gives
n
nARL- (7.22)
x
r3-klil Hz)dz
for the average number of x 1 s till the first point out.

Note that use of (7.22) with k=O, gives ARL 1 s to a false warning, when
process is actually in control. P(point out of ± limits) = .0026 in either

case. Then ARLx = 38S, while ARLx = 4/.0026 = 1540.

7.7. Artificial "Specifications" 183

Comparison may also be made between x- and p charts when the criterion
for defectives is a measurable characteristic, normally distributed. Then if

some jump in p of an amount kox occurs, we may use (7.22) for the average

number of x 1 s till a warning is obtained. Moreover, the jump of kax in p

involves some increased fraction defective relative to the upper specification

U, say from p' top". Then we can find by the binomial distribution (2.30)

the probability of a point above the upper control limit for a p chart of

sample size n. Finally, (7.21) is used by n/P(p point above UCL).

7. 7*. Artificial "Specifications" for Greater Efficiency in Attribute

Charts. This section is basically a continuation of the preceding section.

The idea is to increase artificially the fraction "defective" by using

artificial "specification limits." Consider a somewhat extreme case, where

ax of the process is precisely a sixth of the tolerance U-1. Then assuming

normality and p right at (U+L)/2, there will be a fraction defective relative

to L,U of only .0026. As shown in Section 7.6, if we use x's for n=4, the

probability of a point outside if p increases by lox is .1587. See Figure 7.8.

Hence the average run length to the first x point out is 4/.1587, or about

25 xis.

Now what if we are using a snap gage or go -- not -- go* gage instead of a

measurement gage, just to tell whether or not each piece is within L,U? (This

costs much less time per piece than measuring.) Take n=200, say. As shown above,

p'=.0026, np'=.52. For an np chart we may neglect q'=.9974, and use limits like

a c chart as follows

np' ± 3/np' = .52 ± 31:52 =-, 2.68.

Thus 3 defectives in 200 would be out of control. Using the Poisson Table VII,

we find by interpolation P(2 or less I .52) = .984 or P(point outl.52) .016.

Hence the average number of attribute inspections: good or defective, is

*Two gages: the part should enter the "go" gage and not enter the "not-go" gage.
184 MISCELLANEOUS TOPICS IN CONTROL CHARTS

200/.016 = 12,500. This ARL compares favorably with ARLx = 380 and ARLx = 1540
for x and x charts respectively under this case of controlled conditions. The
point is that we like long ARL's when the process is in control, but short ARL's

when the process goes out of control by some specified amount.

Next what is the effect on our assumed p chart if p were to jump from
(U+L)/2 to (U+L)/2 + cr ?
X

Now the fraction defective p' jumps from .0026 to .0228. Hence now

np' = 200 (.0228) = 4.56, giving, by Table VII, a probability of ,833

for an occurrence of three or more defectives in the 200, thus signaling

lack of control. Then ARL
np
200/.833 = 240. This compares with

ARLx = 44, ARLx = 25; in fact, unfavorably.

Now suppose that we artificially use "specifications" at L' =L + . 25 (U-L) =

(3L + U)/4 and U' = U-.25(U-L) = (L + 3U)/4. This gives a centered "tolerance
range" L' to U', just half as wide as the actual tolerance range. Also under

our assumptions L' is at p - 1.5crx and U' at ~ + 1.5crx. Thus by Table I the
fraction "defective" has been artificially raised to p' = 2 (. 0668) = • 1336.
Also suppose we try n=20 for an np chart:
np' = 20(.1336) 2.67

Limits = 2.67 ± 312.67(.8664) 7. 23.

np

Now what is the probability of a warning signal, that is, of finding eight or

more defectives in the sample, if p' = .1336. From appropriate binomial tables,
this proves to be only about .0029, giving an ARL of 20/.0029 = 690 under con-

ditions of control. Again this is compared to ARLx = 385, ARLx = 1540. Now
suppose as before that ).l jumps by lcr X to (U+L)/2 + cr X. Then the probability

of a piece above U' = (U+L)/2 + 3crx/2 is found by z = [(U+L)/2 + 3o/2 -

(U+L)/2 - cr]/cr = .5, yielding .3085. Therefore ARL

np
= 20/.3085 = 65.
This is quite comparable to the ARL's after a lcr jump of 44 for x's and 25
X

for x's.
7.8. Summary Charts for x and R 185

We must point out that to make a really fair comparison in these cases
of efficiency of charts, the ARL's should be approximately equal when the

process is in control at the satisfactory condition. Moreover we should

compare the charts with various conditions such as varying k in a kcrx jump.

Also we could vary the n's, and the multiple of sigma for control limits,

and so on. See [5] - [7].

Such an approach can be useful. Other terms for this type of artificial

limits are tests of "increased severity" or "narrow limit gaging,"

7.8*. Summary Charts for x- and R. From time to time the author hears some-

one mention an "R chart" and he immediately listens intently. Nearly always it

is someone speaking loosely about an R chart. (True: the centerline is at R.)

But occasionally it is a real "R chart."

For summarization purposes, we may sometimes wish to summarize a collection

of xs1 and R's taken during a shift. Suppose, for example, that eight samples

of five pieces each are measured about* every hour during a shift.
Then we can let k=8, n=5 and find a pair of points to represent the whole shift:

These are x and R points for charts so named.

To set centerlines and limits let us suppose we have 15 shifts in a week.

Then

~ X= Ix/15=~ <tR= I R/15 = R. (7. 23)

Now R is the average of 15(8)=120 R's for the week, each of a sample of n=S.

Hence we have

*It is rather desirable not to make the interval exactly the same, for several

reasons.
186 MISCELLANEOUS TOPICS IN CONTROL CHARTS

the d2 value being for n=S. Then we have

crR = d3ox = .864crx

and since there are eight R1 s per shift

aR = . 864;/ 18.
Likewise for x from 8(5) 40 X 1 S (or 8 x's):

Then these are used for "three-sigma" limits:

(7. 24)

(7,25)
LimitsR R ± 3d 3ox/vK.
- h AC

Such limits are not especially good at pointing out specific assignable
causes, but are good for giving general trends and summarizing, A single
point may be summarizing a considerable variety of conditions. It is true,
however, that an operator or a gage may be "off" level for one whole shift,
giving an ~ point out of control,

7.9.* Variations of Control Charts. Let us now discuss quite briefly

some other adaptations and related literature on control charts. In the next

chapter, we present many references on applications along with appropriate

discussion.

1. Tool Wear and Trends. A quite simple approach to control of processes

subject to consistent change in level has already been given in Section 7.2.

An early paper is by Tippett [8) presenting the general approach, while an

applicational one is by Manuele (9]. Somewhat more sophisticated approaches
using mathematically fitted trend lines are Mandel [10] and Behnken [11].

Gibra [12) gives a least cost approach to such problems using necessary input.
2. Trend Charts and Moving Averages. We presented control charts for
individual measurements x and moving ranges R, in Section 7.1. It is possible

to use this same approach with moving averages, that is, R is used for
estimating ox, then we may plot x for the last two, last three or the last
7.9. Variations of Control Charts 187

four x's, as each new x comes along. Such an approach has the highly

desirable element of simplicity. It can be used along with x's for expensive

and/or rather infrequent measurements, such as, octane numbers for gasoline.

A variation of considerable merit is to give the most recent x the maximum

weight, and earlier x 1 s successively less weight. These weights can be on a

geometrically decreasing basis. Amplification on the approach is given by

Hammer [13], Beall [14] and Roberts [15]. Divers [16] discusses this

approach and time-related control charts. Freund [17] gives a useful

presentation of several approaches. Two other recent papers are Wortham and

Heinrich [18] and Turner [19].

3. Limits Other than Three-sigma. Some workers have experimented with

control limit lines other than those set at plus and minus three standard

deviations (estimated) of the characteristic being plotted. As a matter of

fact Shewart [3.1] and his associates did a great amount of such experimentation

before settling upon three-sigma limits. See Section 3.6. Two variations

are (1) to use some other multiple than three, of the estimated or standard

sigma for the characteristic, and (2) to provide "warning limits" in addition

to three-sigma "action limits."

Freund in [1] developes the former variation by setting an alpha (a)

risk of signaling an assignable cause of change of level when none is present

and setting a beta (S) risk of failing to call for action when the level is

off by a specified amount. This leads to a requisite sample size and the

needed limits. We shall elaborate on this type of significance test in

Chapter 11.
Another related approach is that of modified limits, which may be used

when tight control of process level p is not needed and cr is known, so that

we may set modified limits for x's inside of specifications for x's. See

Section 7.2.2. Also see Hill [21].

188 MISCELLANEOUS TOPICS IN CONTROL CHARTS

Often with warning limits, two x points in succession outside of warning

limits is cause for action. Some workers use two-sigma limits for warning.

Early work on this general approach was done by Page [22], [23]. See

also Mitten and Sanoh [24] and Shahani [25]. Other multiple warning

criteria are given in Moore [26], Satterthwaite [27], Weiler [28],

Goetz [29] and Gibra [30]. These approaches buy a more favorable operating

characteristic curve for the test on the process for a given n, but at a cost

of considerable loss in simplicity and/or in required input, such as, known a

and a desire to guard against a particular size of process jump. Hence the

author is not inclined to urge their use, although they do have their place.

4. Multiple Measurements. An applicational paper on group control charts,

described here in Section 7.5.3, is by Boyd [31]. Multiple measurements on a

single piece are discussed by Hammer [32].

5. Miscellaneous Graphical Procedures. A variety of graphical procedures

are discussed in Freund [33) and Noble [34]. Further analyses of means are

presented by Ott [35) and Small [36). Useful approaches are given by

Seder [37], [38]. Traver and Davis [39], describe how to determine process

capabilities from short runs. Rhodes [40] discusses the sources of variation

in lots of bulk material, and how they may be estimated. When a process is in

good control but with too much spread to fi.t between the specifications L to U,

one may find the most economical level to minimize scrap plus rework costs,

Springer [41].

6, Adjustments for Small Initial Run. When but few measurements are

initially available, the control limits can be adjusted slightly to take this

into account, so as to preserve the usual probabilities. See Proschan and

Savage [42] and Yang and Hillier [43].

7. Power Characteristics of Control Charts. The operating characteristic

curve (probability of acceptance) or its inverse, the power curve (probability

of rejection) are of importance in yielding risks of wrong decisions. For the

7.9. Variations of Control Charts 189

case of standards given (especially a), this is not a difficult problem. But
when analyzing past data it is more difficult. Several papers on the subject

are Olds [44], Page [45], [46], Roberts [47) (in which he gives average

run lengths), and Enrick [48]. Hilliard and Lasater [49] consider the

rather difficult problem of risks when more than one test is used for rejection

criteria for measurement charts.

8. Adaptations Using Other Sample Statistics. In the cases of ranges

and standard deviations when a is known, we use as center lines d 2a and c 4a,

that is, the expected values respectively. Or if a standard a is unknown

we use R or 5. But the distributions of R and s are not symmetrical (the

amount of asymmetry depending upon n), and there is a greater probability of

a value below the center line than above. As a consequence a long run of

points below the center line is much more likely than a long run above it.

To equalize the probability of long runs by chance we may use the median of

R's or s's instead of their mean, Eisenhart [50].

Some workers suggest the use of medians rather than x's. Or mid-ranges,

[x(l) + x(n)]/2, where x(i) is the i'th smallest, may be used. Such statis-

tics can be plotted on a plotting of x's. See Ferrell [51], [52],

Clifford [53], von Osinski [54] and Shahani [55]. When the sample size
-
varies we may plot a standardized variable z =~ if a is known, with ~ set,
a/In
perhaps at (L+U)/2. See Duncan [56].

Another adaptation is to plot x 1i - x 2i yi i = l, ... ,n, where each

pair is on a single part or individual, Grubbs [57] and Eichelberger [58].

Or, one may plot the maximum x(n) and minimum x(l) in each sample, and

set an upper limit on the former and a lower limit for the latter, Howell [59]

and Madison and Ostle [60).

9. Use of Computers. An optimum procedure for adjusting the process

level, readily adapted to machine control is given by Grubbs [61]. A

variety of papers on computer applications follows in chronological order:

190 MISCELLANEOUS TOPICS IN CONTROL CHARTS

Amber and Amber (62], Roberts (63], Williams and Johnson [64] a resume
for process industries, Kallet [65] and Lieberman [66].
7.10. Summary. It seems pointless to try to summarize here the diversity

of problems described in the preceding sections. These adaptations and

special cases can be usefully applied when occasions occur. Moreover careful

study of them will surely aid the student in consolidating his understanding

of control charting principles, the distinctions between attribute and variable

charts and the interplay of charts with specifications.

PROBLEMS

7.1. The following data give the number of .01 per cent of moisture in a

polymer (Polysar S). The maximum desired is .50 per cent. Measurements

once every 2 hr., starting at 4 p.m., Sept. 25: 36, 20, 16, 21, 32, 34,

32, 34, 23, 25, 12, 31, 25, 31, 34, 38, 26, 29, 45, 27, 29, 26, 33, 33, 38,

45, 42, 47, 45, 35, 44, 37. Draw the x and moving range charts. How

many x's would have been outside three standard deviation x limits?

7. 2. The following data represent one analysis per heat on each of 30

consecutive heats of 1045 steel. The specifications are .70 to .90

per cent for manganese. Analyses in .01 per cent of manganese: 74,

79, 77, 81, 72, 66, 75, 80, 76, 86, 84, 70, 80, 62, 74, 71, 68, 79, 81,

76, 79, 79, 84, 78, 74, 88, 71, 80, 79, 74. Draw the x and moving

range charts, and comment. How many x's would have been outside three

standard deviation x limits?

7.3. Below are given for 40 consecutive reels the average basis weight x

(pounds per 1000 sq. ft.) of paperboard before the start of a quality

control system. Make an x and moving range chart and draw any conclu-

sions you can. [ 6 7] .

Problems 191

Table P7,3

Reel X Reel X Reel X Reel X Reel X

1 53.6 9 52.2 17 49.8 25 50.0 33 52.8

2 51.6 10 50.6 18 50.6 26 49.8 34 51.8
3 51.6 11 49.6 19 50.0 27 48.8 35 50.6
4 54.4 12 49.0 20 49.8 28 50.8 36 50.6
5 51.2 13 48.2 21 50.2 29 52.0 37 50.0
6 50.8 14 49.2 22 50.0 30 52.8 38 52.6
7 50.6 15 47.6 23 50.0 31 53.4 39 50.4
8 51.6 16 49.6 24 49.4 32 55.2 40 49.2
2032.0

7 .4. For the data in Problem 7 .3, form ten samples of n=l• from t1•c

consecutive x's, and run x and R charts, comparing with the x and moving

range chart results.

7.5. Make an x and moving range chart for the following data taken

while the author was listening to a lecture. Time in seconds for a large

"pill bug" to pass across consecutive 93,;" tiles: 35, 17, 18, 19, 23, 20, 25,

25, 24, 25, 20, 29, 19, 23, 26. Comment,

7.6. A description of a job by a quality-control man follows (see Table

7.6):

"Three different vendors were supplying forgings. These forgings were

machined on a Gridley automatic. There were constant tool changes and tool

resets. As a result of these changes tool life was short and pieces were

often out of specification. It was the opinion that the difference among

vendors was responsible for pieces being out of specification. However, a

study, such as attached, was made on each vendor's pieces and there was no

significant difference in the trend of any of them. A chart was placed on

the machine and the operator was instructed to set the machine so that the
192 ~1ISCELLANEOUS TOPICS IN CONTROL CHARTS

average of 5 pieces would approximate the Lower Modified Control Limit and
then to let the machine run until the Upper Modified Control Limit was

approached. At this time he was to reset the tool. This chart indicated

that approximately 75 pieces could be produced before a tool needed to be

reset,

"By using this type of chart, what used to be a difficult operation

with short tool life is now running without defects, and 100% inspection

is reduced to 5 pieces each 30 minutes (checked at the machine). Tool

life has increased considerably."

Table P7.6. Outside Diameter at Base of Stem of an Exhaust-Valve Bridge.

Specifications: 1.156In.,+ . 000 In., - .001 In . Samples of Five, all
Pieces Measured. Data in .0001 In. Above 1.154 In.

xl x2 x3 x4 x5
-
X R

12 13 13 14 13 13.0 2

13 14 14 15 14 14.0 2

14 13 15 15 15 14.4 2

15 14 16 14 14 14.6 2

15 15 16 14 15 15.0 2

17 15 16 15 15 15.6 2

17 17 17 17 16 16,8 1

17 17 18 16 17 17.0 2
Problems 193

Table P7.6 (continued)

18 18 18 17 17 17.6 1

17 18 17 18 19 17.8 2

18 17 19 18 18 18.0 2

18 19 19 18 19 18.6 1

19 20 19 18 19 19.0 2

12 13 11 11 12 11.8 2

12 13 12 12 13 12.4 1

11 13 13 13 13 12.6 2

13 13 12 13 13 12.8 1

14 12 13 12 14 13.0 2

13 13 14 13 14 13.4 1

13 13 15 14 13 13.6 2

The "modified limits" used were analogous to what we called modified

limit A in Sec. 7-.2~2. Draw the charts, slanting control limits, and modified

limits, and comment on this job.

7.7. For the data of Table 7.7, first draw the R chart for all 44 points.

Is there evidence of an assignable cause? What is it? If you were in charge

of this operation, could you remove it? Let us eliminate the two points from

the data and recompute Rand the limits. Is there control? Next analyze the

data by slanting control lines on the x chart. What can you conclude?
 194 MISCELLANEOUS TOPICS IN CONTROL CHARTS

Table 7.7. Outside Diameter of a Spacer. .0001 In. below .1250 In,
Specifications: .1235 to .1250 In. Another Run of Data Subsequent
to Table 7.2.
Sample
- R
xl x2 x3 x4 x5 X

1 - 1012 - 133 - 139 - 112 -- 103 - 11.8

5.4 8
2 -- 12 - - - - 11.4 3
3 - 8 - 13 - 1312 - 11 - 5
4 - 11 -- 1010 - 8 - 10 -- 1112 - 10.4 4
5 - 11 - 13 - - 11.2 3
6 - 11 - 9 - 1311 - 11 - 13 -- 11.4 4
- 1212 -- 11
--- 108
7 - 10 10 - 9 10.4 3
8 -- 12 - - 1111 - 10.8 4
9 8 -- 1010 -- 89 - 8 - 9,6 3
10 - 11 - 9 - 9.2 3
11 - 10 - 8 - 109 - 89 - 109 9.2 2
12 - 9 - 10 - 8 - - 8 - 9,2 2
13 - 10 - 10 - 9 -- 99 - 8 - 9.0 2
14 - 8 - 7 -- 6 - 10 -
8.2 2
15 - 7 - 7 - 8 7.6 4
16 - 7 - 6 - 9 -- 87 - 7
- 8
7.4 3
17 - 6 - 6 - 76 -- 43
- 6.8 2
18 - 4 - 5 - 4 - 7 - 5.2 3
19
20
-
- 11
4 -
-
2
2
- 10
-
- 3
- 9 - 10
-- 3.2
8.4
2
9
21 - 1011 -- 109 - 118 - 11 - 9 - 9.4 3
22 -- 9 - 11 - 119 9 - 10,0 2
23 - 77 -- 9 - 9 - 9.4 4
24 - 9 -- 10 - 10 - 8 - 8.6 3
25 - 8 - 9 - 8 - 7 - 8.4 3
26 - 10 - 11 - 8 - 8 - 7 - 8.8 4
27 - 7 - 8 - 8 - 7 6 - 7.2 2
28 - 6 - 8 - 7 - 9 - 9 7.8 3
29 - 6 - 7 - 7 - 8 8 - 7.2 2
30 - 7 - 9 - 5 - 7 - 7 - 7.0 4
- - -
--
31 5 7 - 7 6 5 6.0 2
32 5 -- 8 - 7 - 6 - 7 - 6.6 3
33 - 7 6 - 8 - 8 - 7 - 7.2 2
34 - 8 -- 8 - 7 9 -- 7 - 7.8 2
35 - 7 8 - 7 - 6 7 - 7.0 2
36
37
-- 6
5
-
-
5
8
-- 6
5
-
-
5
6
-
-
5
4
-
-
5.4
5.6
1
4
38 - 5 - 6 - 4 - 6 6 - 5.4 2
39 - 4 -- 4 - 5 -- 11 5 -- 4 - 4.4 1
40 - 11 12 - 10 12 - 11.2 2
41 -- 99 -- 8 -- 118 - 10 - 108 - 9.6 3
42 9 - 10 - 8.8 2
43 - 78 - 8 -- 89 - 10 -- 108 - 8.4 3
44 - - 9 - 7 - 8.4 3
Problems 195

7.8. The following data are for edge widths (thicknesses) of samples of
piston ring~ following the preliminary disc grind. Specifications .1070 ±
.0007 in. Data in table are for the number of .0001 in. from .1070 in.

Day 1 2 3 4 5 6
-
X + 6.6 + 3.2 + 4.6 + 3.8 + 7.5 + 4.7

s 10.8 5.0 6.1 4.5 8.4 5.1

n 100 120 135 105 125 105

What conclusions can you draw?

7.9. The following data are for edge widths (thicknesses) of samples of

piston rings after "clear disc grind." Specifications .0957 ± .0005 in. Data

in table are for the number of .0001 in. from .0957 ± .0005 in. Same part
number as in Problem 7.8. Data are in .0001 in. from .0957 in.
Day 1 2 3 4 5 6

+ .6 + .4 + .07 + .2 + .2 - .05
s 3.0 1.6 1.8 1.6 1.5 2.5

n 140 90 135 105 135 100

What conclusions can you draw?

7.10. The data below are from [68} Use control charts to test the

data for any differences between types of packing. Chemical engineering

research data.

Type of Packing Pressure Drops for Individual Runs

Random B: 46, 29, 48

Random C 37, 35, 50, 39

Random D 29, 16, 5, 24

Dense E 36, 27, 25, 28, 23

Dense F 16, 23, 29

Random H 21, 29
196 MISCELLANEOUS TOPICS IN CONTROL CHARTS

7.11. For a grinding operation, samples of n=5 are taken about every

half hour, 16 per shift, with three shifts per day, five days per week.

At the end of a week, we find x = .22510 in., and R = .00071 in. Set

control limits for a shift x point and R point. Compare to specifications

of .2250 + .0010 in., making and stating what assumptions you need.

7.12. Compare the average r~n length ARLx before the f~r~t indi-

cation of out-of-controlness for the following two charts for the case

p,ox given, when p goes to~ + 1.5ox:

(a). x chart, n=5

(b). x chart with+ 2ox limits.

REFERENCES

1. R. A. Freund, Acceptance control charts. Indust. Quality Control 14

(No. 4), 13-23 (1957).

2. I. W. Burr, Control charts for measurements with varying sample sizes.

J. Quality Techn. 1, 163-167 (1969).

3, A. E. R. Westman and B. H. Lloyd, Quality control charts for -x and R

adjusted for within - subgroup pattern. Indust. Quality Control 5

(No. 5), 5-10 (1949).

4. I. W. Burr and W. R. Weaver, Stratification control charts. Indust.

Quality Control 5 (No. 5), 10-15 (1949).

5. W. L. Stevens, Control by gauging. ~· Roy. Statist. Soc. ~· 10,

54-98 (1948).

6, A. E. Mace, The use of limit gages in process control. Indust.

Quality Control 8 (No. 4), 24, 28-31 (1952).

7. E. R. Ott and A. B. Mundel, Narrow-limit gaging. Indust. Quality

Control 10 (No. 5), 21-28 (1954).

References 197

8. L. H. C. Tippett, The control of industrial processes subject to

trends in quality. Biometrika 33, 163-172 (1944).
9. J. Manuele, Control chart for determining tool wear. Indust. Quality
Control 1 (No. 6}. 7-10 (1945).
10. B. J. Mandel, The regression control chart. J. Quality Techn. 1,
1-9 (1969}.

11. D. W. Behnken, Moving trend control charts. Indust. Quality Control

14 (No. 5), 7-11 (1957).

12. I. N. Gibra, Optimal control of processes subject to linear trends.

J. Indust. Engin. 18, 35-41 (1967).

13. P. C. Hammer, Limits for control charts. Indust. Quality Control 2

(No. 3), 9-11 (1945).

14. G. Beall, Control of basis weight in the machine direction. TAPPI 39,

26-29 (1956).

15. S. W. Roberts, Control charts based on geometric moving averages.

Technometrics 1, 239-250 (1959).

16. C. K. Divers, Some unique but highly successful control charts. Annual
Tech. Conf. Trans. Amer. Soc. Quality Control, 589-597 (1966).

17. R. A. Freund, Control charts eliminate disturbance factors. Chern. Engin.

73 (No. 3), 70-76 (1966).

18. A. W. Wortham and G. F. Heinrich, A computer program for plotting

exponentially smoothed average control charts. ~· Quality Techn. 5,

84-90 (1973).

19. T. F. Turner, Smoothing data trend ratios. Quality Progress 4

(No. 4), 20-22 (1971).

20. R. A. Freund, A reconsideration of the variables control chart,

with special reference to the chemical industries. Indust. Quality Control 16,

(No. 11), 35-41 (1960).

198 MISCELLANEOUS TOPICS IN CONTROL CHARTS

21. D. Hill, Modified control limits. ~· States. 5, 12-19 (1956).

22. E. S. Page, Control charts with warning lines. Biometrika 42,

243-257 (1955).

23. E. S. Page, A modified control chart with warning lines. Biometrika

49, 171-176 (1962).

24. L. G. Mitten and A. Sanoh, The x warning limit chart. Indust. Quality

Control 18 (No. 2), 15-19 (1961).

25. A. K. Shahani, A three-zone control chart. Qual. Engin. 37 (No. 4),

95-99 (1973).

26. P. G. Moore, Some properties of runs in quality control procedures.

Biometrika 45, 89-95 (1958).

27. F. E. Satterthwaite, Pre-control for supervisors. Quality Progress 6

(No. 2), 26-28 (1973).

28. H. Weiler, A new type of control chart limits for means, ranges and

sequential runs. {· Amer. Statist. Assoc. 49, 298-314 (1954).

29. B. E. Goetz, 3cr? Adv. !:::1.8!_. 23, 23 (1958).

30. I. N. Gibra, Economically optimal determination of the parameters of

X-control chart. Management Science 17, 635-646 (1971).

31. D. F. Boyd, Applying the group chart for X and R. Indust. Quality

Control 7 (No. 3), 22-25 (1950).

32. P. C. Hammer, Limits for control charts, second article. Indust.

Quality Control 2 (No. 6), 20-22 (1946).

33. R. A. Freund, Graphical process control. Indust. Quality Control 18

(No.7), 15-22 (1962).

34. C. E. Noble, Variations in conventional control charts. Indust.

Quality Control 8 (No. 3), 17-22 (1951).

35. E. R. Ott, Analysis of means - a graphical procedure. Indust. Quality

Control 24 (No. 2), 101-109 (1967).

References 199

36. B. B. Small, Simplified method of performing an analysis of variance

with control charts. Proc. Fourteenth Midwest Qual. Control Conf.,
223-231 (1959).

37. L. A. Seder, Diagnosis with diagrams. Indust. Quality Control,

Part I 6 (No. 4), 11-19, Part II 6 (No. 5), 7-10 (1950).

38. L. A. Seder, The technique of preventing defects. Indust. Quality

Control 9 (No. 6), 23-30 (1953).

39. R. W. Traver and J. M. Davis, How to determine process capabilities

in a developmental shop. Indust. Quality Control 18 (No. 9), 26-29 (1962).

40. R. C. Rhodes, Lot structure and sampling bulk materials. Quality

Progress 5 (No. 11), 24-27 (1972).

41. C. H. Springer, A method for determining the most economic position of

a process mean. Indust. Quality Control 8 (No. 1), 36-39 (1951).

42. F. Proschan and I. R. Savage, Starting a control chart. Indust.

Quality Control 17 (No. 3), 12-13 (1960).

43. C. H. Yang and F. S. Hillier, Mean and variance control chart limits

based on a sm?.ll number of subgroups. ~· Quality Techn. 2, 9-16 (1970).

44. E. G. Olds, Power characteristics of control charts. Indust. Quality

Control 18 (No. 1), 4-10 (1961).

45. E. S. Page, Control charts for the mean of a normal population. J.

Roy Statist. Soc.~· 16, 131-135 (1954).

46. E. S. Page, Comparison of process inspection schemes. Indust.

Quality Control 21, 245-249 (1964).

47. S. W. Roberts, A comparison of some control chart procedures.

Technometrics 8, 411-430 (1966).

48. N. L. Enrick, Operating characteristics of reject limits for measure-

ments. Indust. Quality Control, 2 (No. 2), 9-10 (1945).

200 MISCELLANEOUS TOPICS IN CONTROL CHARTS
49. J, E. Hilliard and H. A. Lasater, Type I risks when several tests
are used together on control charts for means and ranges, no standard

given. Indust. Quality Control 23 (No. 2), 56-61 (1966).

50. C. Eisenhart, Probability center lines for standard deviation and

range charts. Indust. Quality Control 6 (No.1), 24-26 (1949).

51. E. B. Ferrell, Control charts using midranges and medians. Indust.

Quality Control 9 (No. 5), 30-34 (1953).

52. E. B. Ferrell, A median, mid-range chart using run-size subgroups.

Indust. Quality Control 20 (No. 10), 22-25 (1964).

53. P. C. Clifford, Control charts without calculations: some modifications

and some extensions. Indust. Quality Control 15 (No. 11), 40-44 (1959).

54. R. von Osinski, Use of median charts in the rubber industry. Indust.

Quality Control 19 (No. 2), 5-8 (1962).

55. A. K. Shahani, A control chart based on sample median. Qual. Engin.

35 (No. 7), 7-9 (1971).

56. A. J. Duncan, Detection of non-random variation when sample size varies.

Indust. Quality Control 4 (No. 4), 9-12 (1948).

57. F. E. Grubbs, The difference control chart with an example of its use.

Indust. Quality Control 3 (No. 1) 22-25 (1946).

58. L. S. Eichelberger, Statistical quality control in a press shop. Indust.

Quality Control 13 (No. 2), 12-17 (1956).

59. J. M. Howell, Control chart for largest and smallest values. Ann.

Math. Statist. 20, 305-309 (1949). Errata: 21, 615-616 (1950).

60. R. L. Madison and B. Ostle, An evaluation of the Maytag double-line

chart. Indust. Quality Control 14 (No. 3), 11-13 (1957).

61. F. E. Grubbs, An optimum procedure for setting machines or adjusting

processes. Indust. Quality Control 11 (No. 1), 6-9 (1954).

References 201

62. G. H. Amber and P. S; Amber, QC computers for machine control.

Electrical Mfg. I 58 (No. 1), 80-88, II 58 (No. 2), 78-85, 298 (1956).

63. T. W. Roberts, Jr., Automatic quality control: pipe dream or reality?

Indust. Quality Control 14 (No. 6), 14-18 (1957).

64. T. J. Williams and R. C. Johnson, Computers in the process industries.

Instruments and Automation 31, 90-94 (1958).

65. F. T. Kallet, The computer-another tool for quality control. Indust.

Quality Control 22 (No. 2), 89-90 (1965).

66. G. J. Lieberman, Statistical process control and the impact of

automatic process control. Technometrics 7, 283-292 (1965).

67. C. A. Bicking, The fundamentals of control charts. TAPPI 36, 544-549

(1953).

68. W. B. Seefeldt, Pressure drop in packed columns. M.S. Thesis, Purdue

University (1948).
 CHAPTER 8

APPLICATIONS OF CONTROL CHARTS

8.1. What Manufacturing Problem to Work On. The time of a Quality

Control Department, or an individual person, is divided between (1) routine

controls, (2) "putting out fires" on an emergency basis, and (3) problem-

solving. Also some time should be available for long-range planning. Too

often nearly all of the time is spent on routine matters and in meeting

emergencies as they arise. Keeping in mind the profit and loss statement,

it is of high importance to reserve a generous portion of time beyond day-

to-day matters, to seek improvements in quality and costs. For this we need

to choose processes and operations to work on.

Each part number has many characteristics, each of which is determined

by one or more processes or operations. Ideally a study should be made,

perhaps by Accounting, on the money costs of bad quality for each operation.

Such costs may be classified under (1) prevention (planning, problem solving,

improvement, process control), (2) detection (inspection, after the fact) and

(3) failure (scrap, rework, salvage, field failure, warranties and claims).

Often too little effort is put on the prevention of defective or non-conforming

material.

Studies have shown that if one finds the costs for each operation, a very

few will be found to contribute a completely disproprotionate share of the

total. (Such a maldistribution is often said to follow a "Pareto law.")

Accordingly we ought to try to seek out those operations which are the most

costly, and bend our efforts to markedly improve them, that is, to work where

there is the potential for big financial gains. Not a few companies have obtaine•

203
204 APPLICATIONS OF CONTROL CHARTS

reliable quality costs on each of their various operations, and placed strong
effort on the most costly, to their great advantage. Even without such a
complete study, this philosophy can be used with subjective judgement and such

objective information as may be available.

In summary then we may say "Seek out the operations where the biggest
losses for bad quality occur, and place substantial effort and problem-solving
talent on them." Do not let day-to-day matters prevent such effort.
8.2. Attacking an Industrial Problem. "Now go out and apply these methods."

Easier said than done! A practical problem often looks entirely different from

a book exercise, with the data all collected and neatly tabulated. Hence we

give the following suggestions:

In the first place, finding a solution to an industrial problem is a team

~- Secondly, appropriate action must be taken at the proper time. This is

usually the function of management at some level. Finally, follow-through

is necessary in order to avoid having the original trouble return, and also to

be on guard against future trouble.

With these general observations in mind let us proceed with the following
outline of steps:
1. The problem naturally involves some product or material which is not

satisfactory. Determine what quality characteristic is causing the most trouble.

2. If for the characteristic in question, an appropriate measuring device

is available, test it to make sure that it gives repeatable results. This can
be tested by using an R chart as follows: choose several (say 8 or 10) samples

of material or individual pieces, which may well differ from each other, but

each such sample, by itself, is as homogeneous as possible. Present to the

technician the several samples or pieces to measure or analyze. Then give him

the same samples or pieces in a different order to be measured again,

preferably not telling him they are the same. Repeat until we have n measurements
8.2. Attacking an Industrial Problem 205

on each sample or piece, where n may be 3, 4 or 5. This gives 8 or 10 samples,

each of n repeated measurements. Naturally an x chart would show poor control,

but the R(or an s) chart should show good control, if the measurement technique

is repeatable. If good control is demonstrated then R/d 2 (or s/c 4 ) is an

estimate of the measurement error a . If control is poor take steps to

e
improve the technique.

3. If a measuring device for the characteristic is not available, it may

well be desirable to devise one, since measurement control charts are more

efficient at finding trouble than are attribute charts.

4. If attribute charts are to be used (or must be), make sure that

the definition of a defect or a defective is clear and specific. This can be

checked in the same sort of way as in 2. A larger number of samples or pieces

should be offered for attribute inspection than for repeated measurements.

5. Go back in the plant operations on the product to where it is believed

that the trouble begins. (In machining or grinding operations, for example,

this is surprisingly often the first rough work.) Samples should be drawn as

near to that operation as possible, and the testing or measurement done as

soon after the operation was performed on the pieces or material as possible.

6. Make a list of the potential assignable causes which we wish to correct.

Everyone associated with the operation should be consulted in this: the operator,

setup man, inspector, foreman, engineer, designer, chemist, or whoever may

have some ideas on it.

7. There may be some useful data already available. Usually this is

not the case, because the results may not be in a useful sample size, condi-

tions were not adequately controlled, or a record was not made of the

conditions under which the pieces were made and tested. Attribute data

more often are of use, however. In any case it is worth looking for past

data, at least for a comparison.

206 APPLICATIONS OF CONTROL CHARTS

B. Arrange for the collection of new data. This involves the choice of
sample size. For measurements, four or five are standard, while for a p chart
the sample size should be small enough so as to make it as easy as possible
to keep conditions uniform, but not so small that np is much below 1. As
far as possible, the testing or measuring should be done by the same person

on the same test set, so that any out-of-control indications are due to
product variation and not to the testing or personnel.
9. Careful records should be made of the conditions under which each

sample was made. Ideally the situation with regard to all the potential

assignable causes in (6) should be recorded. But it is obviously impossible

in any practical problem to record them all.

10. If past data are not available, then a new run of preliminary data
must be collected, This is ordinarily done by taking samples about as often
as they can be handled, rather than taking them as far apart as they will

routinely be taken. In this way a quicker start can be made. Also, it is

often advantageous to see how a run of most, if not all, of the pieces does
behave.
11. Analyze the preliminary data. Show them to all the people involved,
and explain with great care and simplicity the interpretation of the charts.

12. The hunt for assignable causes is now on, and this is where teamwork

counts. Basically the approach is to find what conditions or potential

assignable causes were different for the out-of-control samples. Was there
any one condition present for all of the points out on the high side and
not present on many, if any, of the other points? Were there any trends?
(Be sure that any resettings or retoolings are marked on the chart.) If
the process is in control and still not meeting specifications, this will

require a rather drastic change in the process.

13. Corrective action may be tried on any suspected assignable cause.
If no further points go out for a while, it might even be worth while to
8.3. Typical Applications of Control Charts 207

bring in the apparent assignable cause again to see whether or not any points
go out. Finding and proving an assignable cause is an experimental job.

14. Corrective action on a proved assignable cause may require action

by management and possibly the expenditure of money. One of the surest

ways to kill off a promising quality-control man (or department) is for

management repeatedly to fail to take action when the need is clearly proved

and when it is also economically sound.

15. A brief, clear, and factual report should present the case to the

appropriate persons in the organization. Interesting graphs help.

Money saved is always of interest to management. Also improved

quality assurance.

16. Follow-through is essential. Control charts should usually be

maintained even after the problem is completely solved. The samples may be

taken at rather long time intervals, and they can still give warning when the

process is getting out of hand. This is especially true when 100 per cent

sorting has been replaced by sampling. Operations have a tendency to get

out of hand when not kept track of.

8.3. List of Typical Applications of Control Charts. The following list

is given to stimulate your interest and imagination, and as a check list of

ideas for you to consider in your own organization. They are all based on

actual work in a wide variety of organizations. Many applications are

written up in the references in Section 8.6.

1. Saving time in looking for assignable causes when none is present.

In particular avoiding resetting or tampering with the process average when

it is actually just about right. Often such excessive resetting is done

when the process is simply incapable of holding the specified tolerance

(maximum minus minimum). When this is the case, we are simply increasing

variability and making matters worse.

208 APPLICATIONS OF CONTROL CHARTS

2. Knowing when a machine or process is doing the best that can be

expected of it. If in control, but the variability is too great, a fundamental

process change is needed.

3. Help in finding causes of trouble by telling when to look, and by

rational sampling, where. Careful records of conditions under which each sample

was taken can be of much help in tracking down the causes.

4. Decreasing product variability. Careful watch of R or s chart and

improving control of these charts and the x chart usually greatly decreases

variability.

5. Determining repeatability of a measuring or analytical technique and

its error. First step is to gain control on R or s chart for homogeneous

material. Then R/d 2 or s/c 4 estimates ae. See Burr [26].

6. Saving on scrap and rework costs. Control charts can give aid in running

at a safe level from both specification limits to minimize scrap relative to

one limit and rework relative to the other. Also review 4 in this outline.

7. Increasing tool and die life, determining when to retool or reset.

See Section 7.2 again.

8. Decreasing inspection for processes in control at satisfactory level.

Process control samples can then be spaced farther apart and/or final inspection

or customer's receiving inspection reduced. Example 4 of Section 4.7 gives a

case where the sampling was reduced to once in eight hours.

9. Safer guaranteeing of product, reducing customer complaints. Processes

in control are the basic ingredient. Enables Sales to know what to promise

customers.

10. Obtaining warning of impending trouble. In a woolen processing line

a control chart would have given a warning right at the time of process change,

whereas without such a chart, trouble was not noticed for six weeks resulting

in fumbling for six months.

8.3. Typical Applications of Control Charts 209

11. Improving engineering-production relations. By production obtaining

maximum economic performance of processes and reporting capabilities to

engineering, more realistic specifications may be set and costs minimized.

12. Improving producer-consumer relations, by more factual knowledge of

process capabilities, and by making measurements and defect definitions

compatible.

13. Decreasing defects on sub-assemblies. Many cases could be cited of

improvement through p, c and u charts.

14. Comparison of several inspectors, machines, processes, heads,

spindles, orifices and so forth, by comparing one sample from each. See also

Section 7.9, 4, on group control charts.

15. More efficient use of present equipment, knowing when new equipment

is needed, checking it for stability and capability when received or even in

the builder's plant.

16. Stabilization of metallurgical or chemical processes.

17. Justifying a "pat on the back." Charts give an objective way of

telling who is doing an excellent job and when. Can be a great stimulus to

quality performance.

18. As an objective, fact-finding tool for deciding between courses of

action or proposals.

19. Determining stability and quality level of producer. May be able to

safely decrease inspection (100 percent sorting to sampling, or sampling to

reduced sampling). Can compare performance of different producers.

20. Meaningful reports from consumer to producer, and vice versa. In

latter, producer's records may well justify greatly reduced receiving inspection

or even just spot checking, as performance is improved and confidence built

up.
210 APPLICATIONS OF CONTROL CHARTS

21. Convenient, meaningful and compact records of quality performance.

Standardized forms help. Records can be absolutely vital in event of a liability

suit.

22. Interpretation of the many management figures such as, production,

absenteeism, unit cost, budget performance, man-hours per ton, returns, sales,

complaints, and so forth.

23. Exploratory work in research and development, comparison of experimental

conditions.

24. Savings in loading or filling operations. As better control is

obtained it becomes possible to run the average fill closer to the minimum

requirement, and to do so safely. A very commonly occurring opportunity. May

be able to cut "factors of safety" in similar manner.

25. Improvement of visual inspection by balancing errors of (a) missing

defectives, and (b) rejecting good pieces. See Hayes [65].

26. Saving loss from too many meals loaded on an airplane.

27. Decreasing clerical errors, for example, in a mail-order house.

28. Getting an organization "quality minded."

29. Decreasing accidents. Out-of-controlness in hospital calls may

foreshadow a lost-time accident, and suggest a safety campaign.

30. Classifying lumber stacks by moisture content for less spoilage and

more efficient use of kilns (for drying).

31. Improving packaging of product.

32. Improving tin-plate and printing.

33. Facilitating random assembly so as to avoid the necessity of

selective assembly.

34. Control of inventory, alerting as to when to restock by studying

demands on inventory.
8.4. Use in the Laboratory 211

35. Getting at the causes of trouble rather than expecting the non-
conforming pieces to be sorted out and not missed. The objective is to "make

it right the first place," and thus to save on inspection, reworking and

scrapping.

8.4. Use in the Laboratory. Is there any evidence of nonrandomness in our

sample? In the light of our observed data, do we have a right to assume that

laboratory conditions were held constant? How much is the repeatability error

in our measurements or analyses? Control charts can supply answers to these

basic questions in the laboratory.

Most statistical methods for analyzing one or more samples of data such

as those in Chapters 9-12, assume randomly chosen samples. When feasible, the

randomness should be tested, at least until proper sampling techniques are being

used. A simple and useful test of randomness is to form the measurements into

a series of small samples of three, four or five each, in~~~~

the results occurred.*

The same approach can be used in attribute cases, but the opportunities

are less common. Thus from a sample of 1000 pieces or tests, 10 samples of

100 each (or 20 of 50) could be used to form an np chart. Likewise, if the counts

or colonies of bacteria are taken from 15 slides or plates, a c chart with 15

points could be run. In these cases, as with measurements, the average performance

for the whole sample is meaningful and sound only if there is some positive

evidence that the underlying conditions were substantially constant throughout,

and that thus we can assume a single population.

*What will happen if we mix our large sample and keep drawing small samples
from it until it is exhausted? The charts will show good control even if the
large sample was obtained with conditions badly controlled. See Section 7.5.2.
212 APPLICATIONS OF CONTROL CHARTS

The control chart can be used as a tool of research by taking a sample

under homogeneous conditions, then varying the conditions in some way and

taking another sample, etc. In this way a series of samples may be taken and

tested by control charts. This approach is particularly helpful in the exploratory

phases of research where it is not yet known which factors are most influential

on the variable under study, or, if they are known, at approximately what level

they should be held. When the important factors are known, and approximately

the best levels (for desired performance, for example production or yield),

then it is possible to use a statistical technique known as analysis of variance

in efficient experimental designs for answering a number of questions on a

minimum of data. It is possible to use control charts to analyze a rectangular

table of data for both row-to-row and column-to-column effects, as explained in

Section 7.5.4 Although a similar approach can be used for more complex designs,

analysis of variance supplies a more direct method.

Whenever a new measuring device or analytical technique is under study, a

control chart is a useful tool. If charts are run on a series of measurements

and they show control, then we have some evidence of the repeatability of the

measurements, and can also estimate the error of measurement by R/d 2 or s/c 4 .

This will help us in interpreting results using this technique.

The following are helpful articles on this subject: A. J. Duncan [41),

M.D. Gross [59), R. H. Noel and M.A. Brumbaugh [87], E. R. Ott (94), [95),

L. A. Seder [104), G. Wernimont [124), [126).

8.5. The Quality Control Person. Many of the personal qualities which

make for success in other activities are important to those in quality control.

Since quality control depends so largely upon teamwork, the good quality-control

worker cannot be a credit grabber. For the same reason he should like people

and be able to get along well with them. They should have confidence in him.

The more familiar he is with the processes and products in the plant, the better.
8.6. Bibliography on Applications 213

He must be practical and down-to-earth, and able to explain statistical tools

and interpretations to those in the plant who have come up "the hard way," and
in fact most of the people he works with will have had scarcely any statistical

background. Thoroughness and common sense are essential too. Of course some

mathematical and analytical ability and background are necessary. Since much

of the time, those who are untrained in statistical matters need to trust the

quality-control person's word on problems, he should be of complete honesty

and integrity. Keeping on studying the quality control literature, past and

present, and attending related society meetings is basic too. We have left for

the last that most necessary characteristic--imagination. It is the priceless

ingredient.

Now, naturally no one is going to be fully equipped in all of the foregoing

respects. But they are all worth developing and keeping in mind as ideals toward

which to strive. Quality control is an extremely varied, interesting and import-

ant work for those who qualify.

8.6. Bibliography on Applications. In this section we shall list a sub-

stantial number of references to control charting in particular, and to quality

control in general. Many of these are in the journal, Industrial Quality Control

of the American Society for Quality Control. This journal was split into two

in 1968: Quality Progress, and the Journal of Quality Technology. Abstracts of

quality control articles are available in two early volumes*, and in the

abstracting periodicals: International Journal of Abstracts on Statistical

Methods in Industry, 1954-1963, and Quality Control and Applied Statistics,

Interscience Publishers, May, 1956--. All of these sources are well indexed.

You can seek out from the following, titles of articles that appear of

interest in your work.

*G. I. Butterbaugh, "A Bibliography of Statistical Quality Control"

1946 and supplement 1951. University of Washington Press, Seattle, Washington.
214 APPLICATIONS OF CONTROL Cf~RTS

1. Standard Group of the Standards Committee of Amer. Soc. for Quality

Control, "Definitions, Symbols, Formulas, and Tables for Control Charts."
Standard Al-1971 (Also Zl.5-1971 of Amer. Natl. Standards Inst.)

Milwaukee, Wisconsin (1971).

2. Standards Group of the Standards Committee of Amer. Soc. for Quality

Control, "Glossary of General Terms Used in Quality Control." Standard

A3-1971 (Also Zl.7-1971 of Amer. Natl. Standards Inst.) Milwaukee,

Wisconsin (1971).

3. Standards Committee of Amer. Soc. for Quality Control, "Guide for Quality

Control." Standard Bl (Also Zl.l of Amer. Natl. Standards Inst.) Milwaukee,

Wise. (1969).
4. Standards Committee of Amer. Soc. for Quality Control, "Control Chart

Method of Analyzing Data." Standard B2 (Also Zl. 2 of Arner. Natl. Standards

tnst .) Milwaukee, Wise. (1969).

5. Standards Committee of Amer. Soc. for Quality Control, "Control Chart

Method of Controlling Quality During Production." St1.ndard B3 (Also

Zl.3 of Amer. Natl. Standards Inst.) Milwaukee, Wise. (1969).

6. Statistics Technical Committee of Amer. Soc. for Quality Control,
"Glossary and Tables for Statistical Quality Control." Milwaukee,
Wisconsin (1973).

7. Amer. Soc. for Testing Materials, "Manual on Quality Control of Materials."

ASTM, Philadelphia, Pennsylvania, 1951.

B. Industry's Use of Quality Control. Indust. Quality Control 2 (No. 4A),

2-19 (1946). Twenty-one short articles on applications.

9. W. H. Abbott, Which standard deviation? Tooling and Production 22,
86-88 (1956).

10. P. D. M. Anderson and H. Herne, Application of computers in quality

control. Qual. Engin. 33 (No. 5), 16-25 (1969).

8.6. Bibliography on Applications 215

11. G. R. Armstrong and P. C. Clarke, Frequency distribution vs. acceptance

table. Indust. Quality Control 3 (No. 2), 22-27 (1946).

12. W. M. Armstrong, Foundry applications of quality control. Indust.

Quality Control 2 (No. 6), 12-16 (1946).

13. C. A. Bicking, How statistics contributes to quality control in the

chemical industry. Indust. Quality Control 3 (No; 4), 17-20 (1947).

14. C. A. Bicking, Statistical techniques for production control and plant

scale experimentation. Indust. Quality Control 3 (No. 6), 11-14 (1947).

15. C. A. Bicking, Quality control as an administrative aid. Indust.

Quality Control 14 (No. 11), 36-43 (1958).

16. C. A. Bicking, New angles on old problems of measurement and data analysis.

Indust. Quality Control 22 (No. 10), 510-515 (1966).

17. C. A. Bicking and R. T. Trelfa, Some influences of paper machine design

and operation on variability of paperboard. Indust. Quality Control 10

(No. 2), 17-19 (1953).

18. W. Bidlack, E. R. Close and J. C. Warren, Applications of quality control

at Bausch and Lomb Optical Company. Indust. Quality Control, 2 (No. 1),

3-9 (1945).

19. R. S. Bingham, Jr., Practical chemical process control. Indust. Quality

Control 13 (No. 11), 46-54 (1957).

20. R. S. Bingham, Jr., Tolerance limits and process capability studies.

Indust. Quality Control 19 (No. 1), 36-40 (1962).

21. R. S. Bingham, Jr., Quality control applications in the coated abrasives

industry. Indust. Quality Control 19 (No. 5), 5-12 (1962).

22. H. R. Bolton, The use of average and range charts to control material

consumption. Indust. Quality Control 10 (No. 2), 50-52 (1953).

23. H. L. Breunig, Some uses of statistical control charts in the pharmaceutical

industry. Indust. Quality Control 21 (No. 2), 79-86 (1964).

216 APPLICATIONS OF CONTROL CHARTS

24. D. G. Browne, Quality control in the forging industry. Indust. Quality

Control 7 (No. 3), 25-30 (1950).
25. M. A. Brumbaugh, A report on pre-war quality. Indust. Quality Control

2 (No. 2), 11-12 (1945).

26. I. W. Burr, Confidence limits on process capabilities. Indust. Quality

Control 10 (No. 2), 52-55 (1953).

27. I. W. Burr, Short runs. Indust. Quality Control 11 (No. 2), 16-22 (1954).

28. R. Chateauneuf, Modern QC pays off in woodwork. Indust. Quality Control

17 (No.3), 19-25 (1960).

29. Editors, Chern. Week, How to cash in on process control computers. Chern.

Week 102. (Apr. 20), 72-82 (1968).

30. V. J. Clancey, Statistical methods in chemical analysis. Nature 159,

339-340 (1947).

31. B. L. Clark, Statistical methods in the chemical industry. Chern. and

Engin. News 27, 1426-1428 (1949).

32. P. C. Clifford, Integrated control of product quality--what it means. Qual.

Engin. 28, 90-92 (1964).

33. P. C. Clifford, A process capability study usin~ control charts. J. Quality

Techn. 3, 107-111 (1971).

34. A. G. Dalton, Some engineering aspects of quality control. Mech. Engin.

37, 305-307 (1948).

35. J. A. Davies, Quality control in radio tube manufacture. Proc. of Inst.

of Radio Engins. 37, 548-556 (1949).

36. W. E. Deming, Some principles of the Shewhart methods of quality control.

Mech. Engin. 66, 173-177 (1944).

37. W. E. Deming, Report to management. Quality Progress 5 (No. 7), 2, 3, 41

(1972).

38. H. F. Dodge, Keep it simple. Quality Progress 1 (No.8), 11-12 (1973).

8.6. Bibliography on Applications 217

39. J. W. Dudley, Inventory policy by control chart. Indust. Quality Control

16 (No. 7), 4-8 (1960).

40. D. J. Duffy, A control chart approach to manuf:Kturing expense. J. Indust.

Engin. 11, 451-458 (1960).

41. A. J. Duncan, Confidence limits for X' when kn is less than 100.

Indust. Quality Control 10 (No. 4), 31 (1954).

42. A. R. Eagle, A method for handling errors in testing and measuring.

Indust. Quality Control 10 (No. 5), 10-14 (1954).

43. N. M. Ebeoglu and S. P. Phocas--Cosmetatos, How to promote statistical

control charts without really trying. Quality, EOQC Jour, 10, 13-17

(1966).

44. L. S. Eichelberger, Statistical quality control in a press shop. Indust.

Quality Control 13 (No. 2), 12-16 (1956).

45. S. Eilon and J. R. King, On total quality control. Qual. Engin. 29,

8-15 (1965).

46. N. L. Enrick, Variations flow analysis for process improvement. Indust.

Quality Control 19 (No. 1), 23-29 (1962).

47. E. J. Erdman, L. E. Bailey and W. C. Knowler, The production line SQC

helped. Quality Progress 2 (No. 8), 20-22 (1969).

48, A. A. Evans, Foundry quality control. Indust. Quality Control. 9 (No. 4),

12-15 (1953).

49. A. V. Feigenbaum, Quality programs for the 1970's. Quality Progress

1 (No. 12), 9-16 (1968).

SO. V. Filimon, R. Maggass, D. Frazier and A. Klingel, Some applications of

quality control techniques in motor oil can filling. Indust. Quality

Control 12 (No. 2), 22-25 (1955).

51. W. C. Frey and W. M. Spencer, Automatic statistical filling control.

Indust. Quality Control 14 (No.8), 13-16 (1958).

218 APPLICATIONS OF CONTROL CHARTS

52. L, G. Ghering, Refined method of control of cordiness and workability of

glass during production. J. Amer. Ceram. Soc. 27, 373-387, (1944).

53. E. F. Gibian, Quality control by statistical methods applied to line

production. Indust. Quality Control 3 (No. 2), 7-12 (1946).

54. P. E. Gnaedinger, Quality control of factory mixed batches of rubber.

~ ~ 64, 711 (1949).

55. P. E. Gnaedinger, Attaining uniformity in processing rubber batches.
Indust. Quality Control 11 (No. 7), 5-10 (1955).

56. W. P. Goepfert, Some quality control applications in the Aluminum

Company of America. Indust. Quality Control 11 (No. 3), 17-25 (1954).
57. R. W. Good, Quality control in relation to time-study. Indust. Quality
Control 4(No. 2), 13-16 (1947).
58. W. L. Gore, Statistical methods in plastics research and development.
Indust. Quality Control 4 (No. 2), 5-8 (1947).
59. M. D. Gross, Quality control in the synthetic rubber industry. Indust.
Quality Control 4 (No. 3), 11-16 (1947).
60. F. E. Grubbs, On estimating precision of measuring instruments and
product variability. J. Amer. Statist. Assoc. 43, 243-264 (1948).
61. F. E. Grubbs and H. J. Coon, On setting test limits relative to specifi-
cation limits. Indust. Quality Control 10 (No. 5), 15-20 (1954).
62. R. L. Grunwald and C. Williams, Quality control system fits short-run
jobs. Mill and Factory 61, 120-123 (1957).
63. C. K. Hamlin, Statistical quality control methods at Rome Cable Corpora-

tion. Indust. Quality Control 14 (No. 2), 8-16 (1957).

64. H. R. Harrison, Statistical quality control will work on short-run jobs.
Indust. Quality Control 13 (No.3), 8-11 (1956).

65. A. S. Hayes, Control of visual inspection. Indust. Quality Control 6

(No. 6), 73-76 (1950).

8.6. Bibliography on Applications 219

66. B. Hecht, Process control methods. Indust. Quality Control 4 (No. 1), 7-11
(1947).

67. G. V. Herrold, The introduction of quality control at Colonial Radio

Corporation. Indust. Quality Control 1 (No. 1), 4-9 (1944).

68. D. Hill,Modified control limits. Appl. States. 5 (No. 1) 12-19 (1956).

69. C. E. Hoover, SQC in the manufacture of piston rings. Indust. Quality

Control 6 (No. 3), 21-22 (1949).

70. B. W. Jenney, General cost advantages of a quality control system.

Qual. Engin. 36 (No. 7), 7-9 (1972).

71. J. M. Juran, Cultural patterns and quality control. Indust. Quality Control

14 (No. 4), 8-13 (1957).

72. W. A. Kerr, Some axioms of quality control. Indust. Quality Control

1 (No. 1), 3, 14-15 (1944).

73. E. P. King, Probability limits for the average chart when process standards

are unspecified. Indust. Quality Control 10 (No. 6), 62-64 (1954).

74. J. R. King and R. Miller, Automation and quality control. Indust.

Quality Control 17 (No. 4), 13-18 (1960).

75. A. G. Klock and C. W. Carter, Woolen carding meets quality control.

Indust. Quality Control 8 (No. 6), 35-38 (1952).

76. G. J. Lieberman, Statistical process control and the impact of automatic

process control. Technometrics 7, 283-292 (1965).

77. D. L. Lobsinger, Air transportation finds new and lucrative uses for SQC.

Indust. Quality Control 6 (No. 6), 76-78 (1950).

78. J. Mandel, Efficient statistical methods in chemistry. Ind. and Engin.

Chern. 17, 201-206 (1945).

79. V. E. McCoun, The case of the perjured control chart. Indust. Quality

Control 5 (No. 6), 20-23 (1949).

220 APPLICATIONS OF CONTROL CHARTS

80. R. A. McLean, V. L. Anderson, H. S. Bayer and G. W. McElrath, A scientific

approach to experimentation for consulting statisticians. J. Quality
Techn. 5, 1-6 (1973).
81. T. R. Meyer, J. H. Zambone and F. L. Curcio, Application of statistical
quality control in glass fabrication. Glass Indust. 34, 539-541, 564

(1953).
82. D. Mills, The influence of materials and processes on quality control
costs. Qual, Engin. 36 (No. 7), 11-17 (1972).

83. J. A. Mitchell, Quality control in raw materials acceptance and chemical

specifications. Indust. Quality Control 4 (No. 3), 16-20 (1947).
84. V. F. Moore, Control charts to cut costs--a case study. J. Indust. Engin.
16, 288-290 (1965).
85. E. J. Newchurch, J. S. Anderson and E. H. Spencer, Quality control in
petroleum research laboratory. Anal. Chern. 28, 154-157 (1956).

86. R. H. Noel, Statistical quality control in the manufacture of pharmaceuticals.

Indust. Quality Control 7 (No. 2), 14-18 (1950).

87. R. H. Noel and M.A. Brumbaugh, Applications of statistics to drug

manufacture. Indust. Quality Control 7 (No. 2), 7-14 (1950).

88. J. F. Occasione, Quality control as applied to continuous processes.

Indust. Quality Control 13 (No. 4), 9-13 (1956).

89. A. Oladko, Modern quality control applied to compounding. Rubber Age

65, 44 (1949).
90. E. G. Olds and L. A. Knowler, Teaching statistical quality control for
town and gown. J. Amer. Statist. Assoc. 44, 213-230 (1949).

91. P. S. Olmstead, How to detect the type of an assignable cause. Indust.

Quality Control Part I, Clues for particular types of trouble, 8 (No. 3),

32-38 (1952), Part II, Procedure when the probable cause is unknown, 8

(No. 4), 22-32 (1953).

8.6. Bibliography on Applications 221

92. R. von Osinski, Use of median charts in the rubber industry. Indust.
Quality Control 19 (No. 2), 5-8 (1962).
93. R. von Osinski, Computers and your job. Quality Assur. 5 (No. 3),

20-25 (1966).
94. E. R. Ott, Indirect calibration of an electronic test-set. Indust.

Quality Control 3 (No. 4), 11-14 (1947).

95. E. R. Ott, Variables control chart in production research. Indust.

Quality Control 6 (No. 3), 30-31 (1949).

96. R. G. Pippitt, More than cost reduction. Quality Progress 2 (No. 6),

18-20 (1969).
97. J. B. Pringle, SQC methods in telephone transmission maintenance. Indust.
Quality Control 19 (No. 1), 18-22 (1962).
98. H. W. Reece, Optimising quality costs. Qual. Engin. 36 (No. 8), 15-19 (1972).

99. T. W. Roberts, Jr., Automatic quality control, pipe dream or reality?

Indust. Quality Control 14 (No. 6), 14-18 (1957).

100. W. T. Rogers, Quality control of tubular steel products. Indust. Quality

Control 13 (No. 2), 6-11 (1956).

101. G. A. T. Saywell, Practical problems in quality control. Qual. Engin.

28, 58-62 (1964).

102. R. Schin, Quality control engineering in process and product control.

Indust. Quality Control 16 (No. 10), 11-17 (1960).

103. E. M. Schrock, Matters of misconception concerning the quality control

chart. J. Amer. Statist. Assoc. 39, 325-334 (1944).

104. L. A. Seder, The technique of experimenting in the factory. Indust.

Quality Control 4 (No. 5), 6-14 (1948).

105. W. A. Shewhart, Contribution of statistics to the science of engineering.

~Progress 41, 854-858 (1942). Also Bell Telephone System Monograph

B 1319.
222 APPLICATIONS OF CONTROL CHARTS

106. J. Sittig, Defining quality costs. Qual. Engin. 28, 105-113 (1964).
107. B. B. Small, Use of control chart techniques in making a quality audit.

Indust. Quality Control Part I, 6 (No. 1), 15-19, Part II, 6 (No. 2),

11-15 (1949).

108. H. M. Smallwood, Quality control in the manufacture of small-arms ammunition.

Mech. Engin. 66, 179-182 (1944).

109. H. Smith, Jr., Testing non-randomness. Indust. Quality Control 22 (No. 3),

691-692 (1966).

110. K. M. Smith, Twenty-one ways to fail in quality control. Foundry 92

(No. 8), 46-50 (1964).

111. H. E. Thompson, A talk with the foreman about quality control. Indust.

Quality Control 6 (No. 6), 29-34 (1950).

112. W. W. Thompson Jr., Some decision rules for establishing an optimal process

mean. AIIE Transactions 2 (No. 2), 118-121 (1970).

113. W. S. Traylor, Use of statistical methods for time study of batch processes.

Indust. Quality Control 4 (No.4), 18-22 (1948).

114. G. Ver Beke, Statistical quality control in the foundry. Indust. Quality

Control 7 (No. 6), 82-88 (1951).

115. J. F. Verigan, Quality control at the Weatherhead Company. Indust. Quality

Control 6 (No. 6), 36-40 (1950).

116. G. R. Vils and M. D. Gross, Application of statistical methods to the

production of synthetic rubber. Rubber Age 57, 551-558 (1945).

117. H. A. Vorhees and J. E. Culbertson, Control charts and automation applied

to analysis of field failure data. Proc., Second Natl. Sympos, on Qual,

Control and Reliab. in Electronics 18-45 (1956).

8.6. Bibliography on Applications 223

118. J. T. Walter, Continuous process control in a petroleum refinery.

Indust. Quality Control 12 (No. 6), 5-7 (1955).

119. J. T. Walter, How reliable are lab analyses? Petrol. Refiner. 35,

106-108 (1956).

120. R. V. Ward, SQC applications in the chemical industry. Indust. Quality

Control 20 (No. 1), 4-8 (1963).

121. C. B. Way, Statistical quality control applications in the food industry.

Indust. Quality Control 17 (No. 11), 30-34 (1961).

122. W. R. Weaver, The foreman's view of quality control. Indust. Quality

Control Part I, 5 (No. 2), 6-14, Part II, 5 (No. 3), 19-22 (1948).

123. G. Wernimont, Use of control charts in the analytical laboratory.

Indust. and Engin. Chern. 18, 587-592 (1946).

124. G. Wernimont, Statistical quality control in the chemical laboratory.

Indust. Quality Control 3 (No.6), 5-11, (1947).

125. G. Wernimont, Statistical control in chemical production. Chern. Eng.

55, 272 (1948).

126. G. Wernimont, Statistics applied to analysis. Anal. Chern. 21, 115-120 (1949).
 CHAPTER 9

ACCEPTANCE SAMPLING FOR ATTRIBUTES

9.1 . Why Acceptance Sampling? In the relationship between producer and

consumer, whether two separate business concerns, or two departments within

a single organization, there is always the problem of acceptance of material

(units, pieces, parts, articles or bulk product). Should the consumer accept

or reject the lot of material? In the latter case, should it be returned to

the producer for rectification or should the lot be 100% sorted or otherwise

rectified at the consumer's inspection station?

If each lot received is to be sorted 100 per cent for all characteristics

of importance, there are only two problems : arranging for sound inspection

and paying for all of it. Such an approach is necessary, if most lots are not

of satisfactory quality when submitted. If, however, many of the lots would

be satisfactory as submitted, then acceptance sampling becomes feasible.

Acceptance sampling is the use of a sample (or samples) from a lot so as to

make a decision to accept or reject the lot.

Why acceptance sampling? There are two basic reasons, both good . First,

it saves money, and second, in many cases it is absolutely necessary . As to

money, one company using a great many small parts cut its receiving inspection

costs from $7.00 per 1000 parts to $.28 by using scientific acceptance sampling.

Furthermore no more trouble was encountered in assembly than previously. In

this way only lots needing such attention were sorted 100 per cent, the other

lots being accepted about as rece1ved.

225
226 ACCEPTANCE SAMPLING FOR ATTRIBUTES

The second reason for sampling is that it is often impossible to inspect

100 per cent. Thus, if the test is destructive (as for example in tensile

strength or life tests), we must use sampling inspection if we are to have

any pieces left to use! Furthermore, even a moderately simple part may have

forty or fifty characteristics which are written into the description of the

piece. We cannot possibly inspect for each of them on every piece. We can

concentrate only on the most important attributes of the part. The less

important characteristics of all parts are only sampled, if indeed checked

at all.

Since acceptance sampling is desirable and often essential, we ought to

make sure that we choose the right plan for our purposes and that we use it

properly. Under the latter comes sound definition of a defect, sound method

of inspection or testing, careful usage of that method, and random sampling

from the lot. The present chapter discusses how to analyze the characteristics

of attribute sampling plans for defectives or defects, while Chapter 10 describes

some standard sampling plans, and Chapter 11 discusses acceptance sampling

using measurements.

9.2. Characteristics of Sampling Plans. In general we need a way of

describing lot quality so that we can be objective about it. When lot quality

is to be described in terms of defectives, then for the class of defectives

in question, each piece is a "good" or a "defective" one. We let

p' = lot or process fraction defective. (9.1)

For a process or a series of lots, this is the proportion of pieces which are

classified as defectives. Or for one particular lot

N number of pieces in the lot (9. 2)

D number of defective pieces in the lot. (9. 3)

Then for this one lot

p' = D/N. (9.4)

9.2. Characteristics of Sampling Plans 227

Currently, lot quality is often described as defectives per hundred

pieces. Let us use P' for this, that is

P' lot or process defectives per 100 pieces (9. 5)

per cent defective.

Also when interest is in defects rather than defectives and several "defects"

can occur on one piece, then we can use for lots or processes:
c' = average number of defects per piece (9. 6)

or

C' = average number of defects per 100 pieces. (9. 7)

Meanwhile for measurable characteristics of product we have the lot or

process average p and standard deviation a, which together are usually a

sufficient description, unless the distribution of x's is markedly non-normal.

Now when sampling a lot or process of any given quality, and using
objective acceptance-rejection criteria, there will be some calculable

probability of acceptance, Pa.

Pa = probability of acceptance. (9. 8)

This is always a conditional probability, depending upon the quality level

submitted. Now if by some definition the quality submitted is "undesirable"

or "bad," then we would like to reject the lot. But we cannot be perfectly

sure, so there is always some risk of erroneous acceptance. We call such a

risk by the Greek letter S, and also call it a consumer's risk:

P(acc.irejectable quality) = 13 = -::onsumer's risk = CR. (9. 9)

Moreover, unless the lot is perfect, we cannot be absolutely sure of the

sampling plan accepting a lot of acceptable quality. Thus there is almost

always some risk of erroneous rejection. We call such a risk by the Greek

letter a, and also call it a producer's risk:

P(rej.lacceptable quality)= a= producer's risk= PR. (9.10)

For most sampling plans it is easy to draw a curve showing Pa versus

submitted quality. We call this the Operating Characteristic or OC curve

228 ACCEPTANCE SAMPLING FOR ATTRIBUTES

for the sampling plan. It vividly shows the risks under the plan. See

Section 9.4.
Another characteristic of a sampling plan is the co'st of using it.

Such costs are directly related to the sample size needed for a decision,

and the cost of inspecting or testing each piece and making the decision.

We are interested not only in the risks of wrong decisions and cost

of inspection for just one lot, but also on these characteristics for a

series of lots: protection vs. cost. Since a high degree of discriminating

power is purchased at relatively high costs, we need to seek a balance

between (a) low protection at low cost and (b) high protection at high cost.

9. 3. Acceptance Criteria for Attribute Acceptance Plans. We shall

give here criteria for decisions based upon single and double sampling.

In single sampling, just one sample of !2_ pieces is drawn at random and in-

spected, from which a decision is always made. In double sampling, after

the first sample is inspected 1ve may accept, reject, or require a second

sample. After the latter is inspected then a decision is always rendered.

The following notations will be used in addition to those in the previous

section.

Single sampling:

n = number of pieces in the single sample (9. 11)

d number of defectives or defects found in the sample(9.12)

Ac acceptance number for d (9.13)

maximum allowable number of defectives or defects for acceptance

c (in some systems)

Re rejection number for d (9 .14)

minimum number of defectives or defects for rejection

Ac + 1
9.3. Acceptance Criteria 229

Double sampling:

nl number of pieces in first sample (9. 15)

n2 number of pieces in second sample (9.16)

dl number of defectives or defects in first sample (9 .17)

d2 number of defectives or defects in second sample (9 .18)

Ac 1 acceptance number for dl (9 .19)

cl (in some systems)

Re 1 rejection number for dl (9. 20)

Re 1 > Ac 1 + 1 C9. 21)

Ac 2 acceptance number for d 1 + d2 (9. 22)

c 2 (in some systems)

Re 2 rejection number for d 1 + d2 (9. 23)

Re 2 Ac 2 + 1 (9.24)

Flow Chart, Single Sampling:

n from lot yielding
d defectives or defects

Ace. Lot + d < Ac d > Re Reject Lot

Flow Chart, Double Sampling:

n 1 from lot yielding

d 1 defectives or defects

Ace. Lot + Reject Lot

n 2 more from lot yielding

d 2 defectives or defects

Ace. Lot + Reject Lot

230 ACCEPTANCE SAMPLING FOR ATTRIBUTES

9.4. The Operating Characteristic (OC) Curve. Defectives.

In analyzing the characteristics of sampling plans, the one most basic is
the OC curve. It is concerned with the protection provided by the plan:

protection against acceptance of "bad" lots if offered, and protection

against rejection of "good" lots when offered for acceptance. Or, if used

on a process, we have "bad" or "good" process performance. What is the

probability of acceptance, Pa, as a function of submitted quality?

For sampling decisions on defectives we have two different types of OC

curves, designated as Type A and Type B. [1] .

(a). Type~ OC Curve. Account is taken of the lot size. The

probabilities, Pa, are calculated for a finite lot of size ~.

which contains D defectives. Historically such probabilities

were of direct interest to the consumer, especially regarding lots

of undesirable quality. For example, if a consumer regards 1% as

a satisfactory P' lot quality, but would regard P' = 4% as

distinctly undesirable, then he might ask of a sampling plan:

Suppose a lot of 500 should come along containing .04(500) = 20

defectives, then what is the probability it will be accepted?

The hypergeometric distribution is exact for such Type A

probabilities. See Section 9.4.2 for discussion of this distribution.

(b). Type ~ OC Curve. No account is taken of the lot size.

The probabilities, Pa, are calculated for lots considered to be

random samples from a process having a constant fraction defective,

p'. In this case the number of defectives Q_ per lot of ~ will

vary from lot to lot, according to the binomial distribution (2.30).

Horeover, the probability Pa for acceptance of a lot by the

sampling plan is calculated from the binomial distribution, because

9.4. The Operating Characteristic Curve 231

it can be proven that in this case random samples of n from

such a series of lots behave like random samples directly from

a process at p', even though they are samples of~ chosen from

samples of ~· (Further it can be shown that in this case of

a controlled process at p', the defectives din nand D-d in N-n

are independent.) Historically, such probabilities, Pa, were

of direct interest to the producer. Thus if the consumer is

going to use a sampling plan n = 100, Ac = 3, he may ask at

what level p' must he run his process, so as to have 95% of his

lots accepted. The binomial distribution is exact for such Type B

probabilities. See Section 2.5.3

We thus see that Type A probabilities make practical sense from the

consumer's viewpoint when considering bad quality lots, and that Type B

probabilities make practical sense to the producer when considering how good

a quality level, at which he must run, to have a high proportion of his lots

accepted. Thus the two types were historically associated with CR and PR

respectively. As we shall see, however, it makes rather little difference

which type of calculation is used. This is fortunate indeed because

calculations using the hypergeometric distribution are extremely tedious.

9.4.1. Type B Operating Characteristic Curves. Let us now consider

the calculation of Type B OC curves for single sampling plans. First

consider a single sampling plan:

n = 80, Ac = 3, Re = 4.
Then Pa = P(3 defectives or less in 80). The exact calculation in this

case involves the binomial (2. 30), namely, P (d) = C(80 ,d) p •d q • 80 -d, and

thus

Pa l:3 C( 8 0 d) ,d q'80-d
d=O ' p
232 ACCEPTANCE SAMPLING FOR ATTRIBUTES

Direct use of this formula can be quite a chore, but can be simplified
a bit by first finding P(d=O) q• 80 , then using it to find the other needed
probabilities as follows:
P(l) = P(0)(80/l)·p'/q 1 , P(2) = P(l)(79/2)·p'/q 1 , etc.
But there are available tables, I2J - f6], which provide for most cases

in practice. (In fact it is not too great an undertaking to program a

computer for a specific job.)

Now an available further simplification is to use the Poisson

distribution (2.35) as an approximation to the binomial. This begins to be

a useful apprxoimation when n ~ 10 and p' ~ .10. To visualize the approach,

see Figure 9.1 from [7]. If we make use of the Poisson approximation we

may use our Table VII, in the back of the book, interpolating as necessary.
The Pa values for the Type B OC curve for the given single sampling

plan are shown in Table 9.1. The exact binomial entries were found in {3]

using Pa = P(3 or less) = l-P(4 or more), the last P being entries in the

table used. In order to use the Poisson approximation we form an np'

column in Table 9.1 providing the expected number of defectives in the
sample. Then these np' are used in the left column of Table VII, and

c = 3 for the column heading, and so the entry in the body of the table
gives Pa. These are shown in the last column.
Careful inspection of Table 9.1 reveals that for very low p', .00 to .03,

the exact Pa is greater than or equal to the approximate Pa, whereas, for
worse p''s, .04 on up, the exact Pais lower than the approximate Pa. Thus
the exact Pa gives a slightly more "square-shouldered" OC curve than does

the Poisson approximation. Or to put it another way the exact OC curve

provides slightly smaller risks, both PR and CR than does the approximated
OC curve. This is in general true. Hence using the Poisson approximation
overstates the risks of wrong decisions in general.
 <0
-1:-

;i
(1)

~
(1)
'1
P>
r+
.....
::f
()Q
A.- Distributions with pn =I B.- Distributions with pn = 2
n
LEGEND FOR :::r
RELATIVE FREQUENCY e;
P>
~ (l
r+
.z e (1)
.41 - '1
1-'·
0 ~:~-· Vl
r+
.....
(l

n
~
<
(1)

FIG. 9.1. Binomial approach to Poisson distribution [7]. Reproduced by

kind permission, from "Holbrook Working, A Guide to Utilization of the
Binomial and Poisson Distributions in Industrial Quality Control".
Stanford University Press, Stanford University, California, 1943.

N
{;;!
{;;!
234 ACCEPTANCE SAMPLING FOR ATTRIBUTES

Table 9.1. Type B OC Curve for Single Sampling Plan:

n = 80, Ac = 3, Re = 4

Process Exact Pa[3] Expected Defectives Poisson Approximation

p' = P(3 or less) in Sample 80·p' P(3 or less). Table VII

.00 1.000 .0 1.000

.01 .991 .8 .991
.02 .923 1.6 .921
.03 .781 2.4 .779
.04 .602 3.2 .603
.OS .428 4.0 .433
.06 .286 4.8 .294
.07 .181 S.G .191
.08 .109 6.4 .119
.09 .063 7.2 .072
.10 .035 8.0 .042

Figure 9.2 shows the OC curve for the plan together with that for a
double sampling plan with quite similar characteristics. We shall shortly
be discussing analyses of double sampling plans. Section 9.8.

9.4.2~ The Hypergeometric Distribution. For calculations for Type A OC

curves, we need the hypergeometric distribution. Since the distribution may

well not be included in elementary statistics courses, we find it desirable to

discuss it here.

Consider the problem of drawing randomly four motors from a lot of ten

motors, of which three are defective somehow (not necessarily unusable). We

might now ask what is the probability that the sample contains precisely one

defective motor (and three good ones)? To answer this we treat the ten motors
as distinguishable, for example, by a serial number.

Now how many distinct samples are possible, drawing at random four out of

the ten, without replacement? The answer is the combination

C(l0,4) = 10·9·8·7
4 . 3 . 2 . 1 = 210.
9.4. The Operating Characteristic Curve 235

1.00

0,80

0,60

Po

0.40

0.20

0.00 ,___ _ _ _ _-:-::-:---::-::--------~

0.00 0.02 0,04 0.06 0.08
p'

FIG. 9.2. Two Type B OC curves. (a) n = 80 Ac = 3, Re 4.

(b) n 1 = n 2 = so, Ac 1 = 1, Re 1 = 4, Ac 2 = 4, Re 2 = 5.

Because of random drawing, all of these 210 are equally likely. Next, of the

210, how many will contain exactly one of the defectives and exactly three of

the good ones? We can choose the former in

C(3,1) =3
distinct ways, and the latter three out of the seven good ones in the lot in

C(7 3) = 7 · 6 ' 5 = 35
' 3·2·1
distinct ways. Since each of the 3 ways can be paired with each of the 35 ways,

this gives 3(35) = 105 samples with just one defective. Thus
P(l) = lOS = .500.
210
Hence half of the time the sample of four will contain exactly one defective.

The other probabilities prove to be:

P(O) = .1667, P(2) = .3000, P(3) .0333.

 236 ACCEPTANCE SAMPLING FOR ATTRIBUTES

Now let us set down the usual symbols for the problem and obtain the general
form. We have in the above

N = 10, D = 3, n = 4, d

or in general

N lot or population size (9.25)

D number of "defectives" in lot (9.26)

n = number of pieces drawn at random from lot (9. 27)

d number of "defectives" in sample (9.28)

With this notation, then, the number of possible samples of n from the lot is

C(N,n). Meanwhile we can draw d defectives from the D in the lot in C(D,d)

ways, and n-d good from the N-0 good in the lot in C(N-D,n-d) ways. Multiplying

the last two combinations gives the number of possible samples with d defectives

in the n. Thus with all specified samples being equally likely we have:

P(d!N,D,n) C(D,d) C(N-D,n-d)

C(N,n)
(9.29)
D! (N-D) !n! (N-n)!
d! (D-d)! (n-d)! (N-D-n+d) !N!

For calculational purposes we can best use the latter form in terms of factorials,

by using Table IX, which gives logs of factorials. Since factorials only exist

for zero or positive integers, inspection of the denominator shows the range of

sample defectives to be

Max(O,D+n-N) < d < Min(D,n), (9.30)

In quality control problems, since both D and n are usually much smaller than N,

the lower limit for d is almost always 0, not positive.

Calculation of hypergeometric distribution terms is quite a chore. The

best approach is to try to find the desired probabilities in a table. About

the only one of substantial size is [8]. This is because we must make a
9.4. The Operating Characteristic Curve 237

quadruple entry table with probabilities listed for d, N, D, n. Hence any

semblance of completeness requires a huge number of entries. Fortunately, as

pointed out in [8), there are many symmetries, which reduce the number of entries

needed. For example, in (9.29) we can interchange D and n without affecting the

probability.

It may readily be shown that the average and standard deviation for sample

defectives d are

E (d) (9.31)

where the second expression uses p' = D/N, the lot fraction defective in

order to show the resemblance to the binomial characteristics (2.31), (2.32).

E(d) is identical to (2.31) and if n is small relative toN, then ad is nearly

identical to (2.32).

These results suggest that the binomial is a good approximation for the

hypergeometric, if n is, say, N/10 or less. Thus by (2.30)

P(dJN,D,n) ~ C(n,d) (p')d(l-p')n-d p' = D/N. (9. 32)

Or, if D is smaller than n, a better approximation is usually given by

P(d IN,D,n)
d D-d
~ C(D,d) (n/N) [1-(n/N)) . (9.33)

See [9) on these two approximations. This reference also gives formulas for

improving these approximations by provision of a correction term.

Furthermore if n is, say, at least 20, and D/N, say, .05 or less, we can

go even further and use the Poisson distribution (2.35) as an approximation

as follows:

c' = nD/N. (9. 34)

Thus we often can use Table VII to approximate required probabilities, even for

Type A operating characteristic curves.

238 ACCEPTANCE SAMPLING FOR ATTRIBUTES

But if we do need to calculate hypergeometric terms from direct use of

(9.29), the usual procedure is to find one term, using logarithms of factorials,
and then find subsequent terms recursively. Thus from (9.29)

P(d+liN,D,n)
P(djN,D,n)

D!(N-D)!n!(N-n)ldi(D-d)!(n-d)!(N-D-n+d)IN!
(d+l)I(D-d-l)l(n-d-l)!(N-D-n+d+l)!N!D!(N-D)!n! (N-n)!

(D-d)(n-d)
(d+l)(N-D-n+d+l)

Therefore

P(d+ljN,D,n) (D-d)(n-d) I
(d+l) (N-D-n+d+l) P(d N,D,n), (9.35)

It is often easier to use the actual integers of N, D, n, d rather than treating

(9.35) as a formula to be substituted into.
For purposes of comparison we give Table 9.2, wherein N=800, n=lOO and D=l6.

It is easily noted that, of the two binomial approximations, the second which

treats D as n (and n as D) does much better. This is because D of 16 was so

Table 9.2. Comparison of Probabilities from Hypergeometric Distribution

with Two Binomials and a Poisson Approximation. N=800, n=lOO, D=l6.

No. defectives Hyper geometric Binomial A£Eroximations Poisson

d, in Sample Distribution n';foo J PI"', 0"2 n=l6,p'=.l25 c'=2

0 .11553 .13262 .11807 .13534

.26985 .27065 .26986 .27067
2 .29207 .27341 .28915 .27067
3 .19443 .18227 .19276 .18045
4 .08909 .09021 .08950 .09022
5 .02979 .03535 .03068 .03609
6 .00752 . 01142 .00804 .01203
7 .00146 .00313 .00164 .00344
8 .00022 .00074 .00026 .00086
9 .00003 .00015 .00004 .00019
10 .00000 .00003 .00004
11 .00001 .00001
9.5. The Average Outgoing Quality Curve 239

much smaller than n=lOO. So in general use as n in (2.30) the smaller of D

and n. Also the Poisson approximation is a little further off than either

binomial.

9.4.3. Type A Operating Characteristic Curves. For Type A operating

characteristic curves, we are interested in the probability of acceptance of

a lot of N pieces, which contains exactly D "defectives." This may well be the

situation of interest when we are to make a decision on one or two isolated

lots, instead of a decision on each of a series of lots. For exact treatment

this requires us to take into account the lot size, and thus to make "finite-

lot" calculations, using the hypergeometric distribution. A characteristic of

such sampling is that the probabilities on the draws does not remain constant.

Thus in the example of Table 9.2, if on the first draw a good piece is obtained

then on the second draw

p' = 16/799

while if the first piece was a defective

p' = 15/799.
Always p' at any draw depends on what happened before this draw. The hypergeometric

distribution takes this dependency into account.

We therefore have for a single sampling plan defined by n and Ac, for

given N:

Ac
P(acciD inN) L PCdiN,D,n) Type A OC, (9.36)
d=O

The OC curve is then drawn for Pa versus fraction defective D/N. Usually it

is entirely satisfactory to use one of the three approximations given in

Section 9.4.2. Or one could use a second order approximation from [9], unless

[8] gives the needed entries.

9. 5. The Average Outgoing Quality Curve, AOQ Curve. Tlw is a

second type of curve which pictures another protection characteristic of

 240 ACCEPTANCE SAMPLING FOR ATTRIBUTES

a sampling plan. When it is feasible to sort 100% the rejected lots,

screening them of all defective pieces, we may ask what the average frac-
tion defective is, including both accepted lots and those rejected and
screened. This average is called the "average outgoing quality", AOQ.

Obviously, the AOQ will depend upon the average submitted quality p',

because if p', is relatively very good, then nearly all lots will be ac-

cepted and the AOQ fraction defective will be approximately p', (but

actually a trifle less). But if p' is relatively poor, many lots will be

rejected and screened, and the AOQ will be well below p'. Thus

AOQ average outgoing quality (9.37)

average fraction defective in ~ lots, after the rejected lots

have been sorted and cleared of defectives; an average quality for

practically perfect lots (those sorted) and lots still approximately

at p' (those accepted).

Also an important number, descriptive of a plan's protection is:

AOQL average outgoing-quality limit (9.38)

maximum value of the AOQ's for all values of p'.

This is the worst long-run ~ge quality with which we will have to work,

no matter what p' or p''s come in for acceptance.

We shall assume Type B calculation, that is, a series of lots produced

from a process at p', so that the binomial distribution is exact.

There are several slightly different interpretations possible for AOQ's,

depending upon what one does with defectives which have been found. Three are:

1. The defectives found in the samples may be retained with the accepted
lots and not replaced with good ones. They are presumably reworked or

salvaged. Also all the defectives in the rejected lots are found and

replaced by good ones. This is the simplest assumption. It may be

practical in some vendor-certification programs.
9.5. The Average Outgoing Quality Cune 241

2. The defectives found in the samples and the sorting of rejected lots

are all removed, but not replaced by good pieces. This is the usual assump-

tion where sampling inspection is done at some considerable distance from

the place where the pieces were produced. Hence it is not very feasible

to replace the defectives. This is typical of much receiving inspection.

3. The defectives found in the samples and sorting are all removed and

replaced by good pieces. This is the usual assumption where sampling in-

spection is done close to the place where the pieces are produced, so that

defectives may readily be replaced. This is typical of acceptance sampling

of lots at final inspection.

In all three cases, as we do throughout the chapter, we assume that the

inspectors find all the defectives in the samples and in the sorting of the

lots, and that they do not reject any pieces which are actually good ones.

Although this is never quite fulfilled in practice, it does provide a reason-

able first approximation. Adjustments can be made if one knows in what way

an inspector's work is inaccurate, whether too lenient or harsh, or inaccurate

in both types of error.

Suppose for the moment we assume interpretation 1. Then we have two

kinds of lots with which to work: (a) accepted lots still at an average of

just about p', and (b) perfect lots. Moreover all lot sizes are still at N.

Now suppose that 1000 lots each of 2000 pieces with an average fraction

defective of p' = .02 are submitted for inspection under the single plan:

n=80, Ac=3. We can then expect that 921 will be accepted, Table 9.1, and 79

will be rejected. Thus AOQ is

921(.02)2000 + 79(.00)2000
AOQ .921(.02) .0184
921(2000) + 79(2000)

In general it is seen that the proper formula for interpretation 1 is

242 ACCEPTANCE SAMPLING FOR ATTRIBUTES

p'Pa (interpretation 1). (9.39)

This is also very nearly correct for the other interpretations.

Under interpretation 3 when using single sampling, the calculation of

AOQ may be slightly modified to again attain full accuracy. In 3, we note

that we have in the accepted lots n pieces which are all good (following

any needed replacements), and N-n still at about p'. Hence the average

fraction defective in the accepted lots in p'(N-n)/N, so that under inter-

pretation 3:

AOQ,
p
= p'Pa[l-(n/N)] (interpretation 3). (9.40)

This formula may also be used in double sampling by using the average

number of pieces used for a decision instead of the single n.*

Interpretation 2 gives an AOQ somewhere between those of the other two

interpretations, because it does not have the improving benefit of the removed

defectives being replaced by good ones.

Table 9.3 lists the AOQ's for both plans, along with other characteristics

yet to be discussed. Figure 9.3 shows the two AOQ curves for the plans we

are analyzing, using (9.39). The AOQL for the single sampling plan

may be read off from the AOQ curve, Figure 9.3 or from inspection of the

AOQ entries as in Table 9.3. It is about .024. The AOQL for the slightly more

lenient double sampling plan is about .027. Such AOQL 1 s will only occur in

practice, however, if p' remains at just about .035. If the incoming fraction

defective is either lower or higher than .035 then the outgoing fraction

defective AOQ will be lower than the AOQL. Thus in practice the actual average

outgoing quality is likely to be considerably better than the AOQL of a sampling

plan.

*Since n/N is usually small, this correction is likely to be of little

practical importance.
 9.6. The Average Sample Number Curve 243

Table 9.3. Comparison of Characteristics (Type B) of Two Sampling Plans

with N=2000: (a) Single n=80, Ac=3, Re=4; (b) Double n 1=n 2=SO, Ac 1=1, Re 1=4,
Ac 2=4, Re 2=5. Calculations by Poisson Distribution. Interpretation 1 for AOQ's.

Process
Fraction Single Double

Defective p' Pa AOQ ASN ATI Pa- AOQ ASN ATI

.00 1.000 .0000 80 80 1.000 .0000 50.0 so
.01 .991 .0099 80 97 .996 .0100 54.4 62
.02 .921 .0184 80 232 .950 .0190 62.2 168
.03 .779 .0234 80 504 .831 .0249 68.8 393
.04 .603 .0241 80 842 .662 .0265 72.6 722
.05 .433 .0216 80 1169 .488 .0244 73.6 1058
.06 .294 .0176 80 1436 .339 .0203 72.4 1346
.07 .191 .0134 80 1633 .224 .0157 70.0 1567
.08 .119 .0095 80 1772 .145 .0116 67.0 1720
.09 .072 .0065 80 1862 .091 .0082 64.0 1824
.10 .042 .0042 80 1899 .057 .0057 61.2 1890

0.03

0.02

AOQ

QOI

0.00 ~---------------------
0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07
I
p'

FIG. 9,3, AOQ curves for (a) single sampling plan: n=80, Ac~3, (b) double
plan: n 1=n 2 =SO, Ac 1 =1, Re 1=4, Ac2 =4, Re 2=S,

9.6. The Average Sample Number Curve, ASN Curve. The average sample

number for a sampling plan is the average number of pieces needed to be inspected

from a lot, in order to reach a decision, whether for acceptance or rejection.

In single sampling, after inspection of n pieces we always reach a decision.

244 ACCEPTANCE SAMPLING FOR ATTRIBUTES

Now it could well occur for a lot of relatively poor quality that we reach the
rejection number Re, well before inspection of n pieces. Since by then d=Re,

we could reject at once without completing the sampling. Indeed this is what

we would likely do if we have an isolated lot. But if we are sample inspecting

a series of lots we will in general complete the inspection of all n in each lot

for the record, so as to study the performance of the producer as to quality

level and consistency, via a p or an np chart.

Thus for single sampling the ASN curve is in fact a straight line at

height n, for every incoming p'. (As a matter of practical expediency however,

if most pieces or all proved defective, by some mischance, we should probably

not complete the entire sample.)

For double sampling, however, the ASN curve starts off at n 1 , if p'=O,

because all pieces in the sample are good and acceptance with d 1=0 always occurs.

But as p' increases from zero the ASN curve rises, reaches a maximum and then

begins to decline, eventually dropping to n 1 again. See Table 9.3.

9.7. The Average Total Inspection Curve, ATI Curve. A second charac-

teristic directly related to inspection cost is the average total inspection,

ATI, as a function of submitted quality. This is the average number of pieces

inspected per lot, including those required for the decision, plus those addi-

tionally required in 100% inspection of rejected lots.

For single sampling a formula for the ATI is easily obtained. First, we

must always inspect n pieces to reach a decision. Whenever this decision is

for rejection, inspection of the remainder of the lot is called for, namely

N-n pieces. This additional inspection occurs in a proportion of lots, 1-Pa,

which of course is dependent upon the incoming fraction defective p'. Thus

we have the formula

ATI 1 = n + (N-n)(l-Pa) Single sampling. (9.41)

p
9.8. Double Sampling Plans 24S

For our example, the ATI values are given in Table 9.3. There we see that

the ATI starts at n for p'=O, and steadily rises to N, when p' is high enough

that virtually all lots are rejected.

The calculations for the AOQ, ASN and ATI are done in general only for

Type B probabilities of acceptance Pa, since they are averages and only mean-

ingful over a series of lots. Type A probabilities Pa are usually only used for

one or a few isolated lots, if used at all.

9.8. Characteristics of Double Sampling Plans. We now consider the

analysis for the characteristics of a double sampling plan. We have already

presented the decision making process in Section 9.3. In Figure 9.2 and Table

9.3 we have some of the characteristics of a double sampling plan in comparison

to the single sampling plan exemplified. It had n 1 =SO, n 2 =SO, Ac 1 =1, Re 1=4,

Ac 2=4, Re 2=S. Such decision criteria are often placed in a box like the following:

n In Ace. No. Rej. No.

so so 1 4

so 100 4 s

Thus if d 1 ~ 1 there is acce?tance on the first sample, or if d 1 ~ 4, there

is rejection on the first sample; while if d 1=2 or 3, we must draw a second

random sample of SO. Inspection of the second sample yields d 2 defectives.

Then if d 1 + d 2 ~ 4, the lot is accepted, while if d 1 + d 2 25 the lot is rejected.

Therefore, for Pa we must calculate the probabilities for the various

"avenues" by which acceptance can occur:

d1 0 or 1

d1 2 followed by d 2 ~ 2, so that d 1+d 2 < 4

d1 3 followed by d 2 ~ 1, so that d 1+d 2 < 4,

The calculations are made for each avenue of acceptance and then added, for

Pa, since these avenues are mutually exclusive.

246 ACCEPTANCE SAMPLING FOR ATTRIBUTES

We could make the calculations for Type A OC curves, if necessary. But

this would be very tedious, because of the necessity of using hypergeometric

probabilities, and because the composition of the remainder of the lot, after

the first sample is drawn, varies with the results on that sample. If the lot

size is 100 or less, one can use Lieberman and Owen [8]. Otherwise log(n!)

tables are necessary for (9.29). The usual recourse, however, is to use the

binomial or Poisson distributions for an approximation.

More often, however, we are interested in Type B calculations of Pa, for

which the binomial is exactly correct (for lots from a process with an assumed

constant fraction defective p', and independence piece to piece). Then

binomial tables [2] to [6], are exact and may be used. However, since p' is
usually small in practical applications and n ~ 20, we can use the Poisson

distribution as an approximation, and thus Table VII.

Let us now illustrate the calculation of Pa for our double sampling

plan, using Table VII which gives

P(c or lesslc')

that is, cumulative probabilities. We take the case of p' .04. A con-

venient tabular form follows:

p' = .04 Contrib.

to Pa
Expectation:

P(l or lessl2)=.406 .406

P(212)=.271 P(2 or lessj2)=.677 .677 (. 271)=.183
P(2 or lessl 2)=.677
P(312)=.180 P(l or lessj2)=.406 .406(.180)=.073
P(3 or lessj2)=.857
Pa=.662

The rationale is quite simple as discussed previously. Out of say 1000

lots incoming from a process at p' = .04, 406 of these are expected to

yield none or one defective in the first sample of 50, and thus all 406 are

accepted at once. Likewise 271 of these lots are expected to yield 2 defectives
9.8. Double Sampling Plans 247

on the first sample because P(2) = P(0-2) - P(O,l). Judgment is withheld

till a second sample is inspected. P(0-2) = .677, so 67.7% of

these 271 lots can be expected to be passed by having 2 or less defectives for d 2 ,

so that d 1 + d 2 ~ Ac 2 = 4. This gives 183 lots of 1000 being accepted via

this route. Similarly for those with d 1 = 3, of which we expect 180, 40.6%

of them are expected to be accepted. Then we total the probabilities for Pa.

For the exact binomial calculation at p' = .04 we use f3] obtaining

the following

Contr. to Pa

P(O,l) .40048 . . . . . . . .40048

P(2) .27623 P(0-2} .67671 .67671(.27623} .18693
P(0-2} · 67671 P(3} .18416 P(O,l) .40048 .40048(.18416} .0737S
p (0-3} .86087 Pa .66116

This result turns out to be very close to that for the Poisson approximation.

Let us do one more, using the Poisson approximation, for comparison.

Take p' .08.

p' .08 nl = so n2 so Contrib.

to Pa
Expectation: n 1p 1 =4.0 n2p' 4.0

P(!) ,1) .092. .092

P(Z)_ .146 P(0-2) .238 .238(.146) .03S
P(0-2) .238
p (3) .19S p (0 ,1) .092 . 092 ( .19S) .018
P(0-3) .433 Pa --:T4S
Both of these points on the OC curve are somewhat above the

corresponding points for the single sampling plan, indicating the double

plan to be more lenient at p' = .04 and .08. Other points are shown in

Table 9.3. Upon comparing entries in Table 9.3 we see that the double

sampling plan has a higher Pa at each p'(except at 0}, and it is therefore

somewhat more lenient. The OC curves are both shown in Figure 9.2, that for

the double plan lying above the single plan throughout.

 248 ACCEPTANCE SAMPLING FOR ATTRIBUTES

Next consider the average outgoing quality, AOQ, curve. We again might
consider the three alternate interpretations given in Section 9.5, and could
work with either Type A or Type B probabilities. Interpretation 1 leads
to the same equation
AO~, = p'Pa (interpretation 1, double sampling) (9.42)
using the same reasoning as for single sampling.

If we use interpretation 3, then in the accepted lots we have either

n 1 pieces which we know are all good, or if a second sample is drawn resulting
in acceptance, then n 1+n 2 pieces are known to be good. Taking account of
the probability of acceptance on (a) the first sample and (b) the second
sample, we have an average number of pieces known to be good as being the

ASN of the plan. See below. Thus we have the following

AOQ
p' = p'Pa[l (ASN/N)] (interpretation 3, double sampling). (9.43)
This formula is not quite exact. Once again, (9.42) and (9.43) are likely
to give very similar results, and also interpretation 2 will give results lying
between. See Table 9.3 and Figure 9.3 for the example.
For the average sample number, ASN, curve, under double sampling we
really do have a "curve." In order to reach a decision we will sometimes

require only n 1 pieces in a sample and other times n 1+n 2 pieces. The
probability of the former is
P(decision on first sample) = P(d 1SAc 1) + P(d 1 ~e 1 )

while the probability of the second case is

P(decision on second sample) = P(Ac 1<d 1 <Re 1).
Since we always look at n1 pieces, we can use the latter as the probability of
having to look at n 1+n 2 . Hence we have

(9. 44)

In order to calculate ASN's we can easily use our OC calculation tables as

illustrated earlier in this section. For example if p'=.04, then

9.8. Double Sampling Plans 249

AT! = 50 + (.857-.406)50 50 + . 451 (50) 72.6

or if p'=.08
ATI =50+ (.433-.092)50 =50+ .341(50) = 67.0.
The ASN curve for this double sampling plan is shown in Figure 9.4 from

Table 9.3. Note that the average amount of inspection is everywhere less

than that for the single plan, even though they have quite similar OC curves.

(Of course part of the time the sample size required for the double sampling
plan to reach a decision is 100, exceeding n=80 for the single plan.)

Finally let us take up the calculation for the average total inspection,

AT! for double sampling. The overall inspection load is compounded of

three possibilities: n 1 and n 1+n 2 under acceptance and N under rejection, since
in the last case all pieces must be inspected. We have but to multiply these

three amounts of inspection by the corresponding probabilities to obtain the

average amount. The respective probabilities are seen to be

P(acceptance on first sample) = P(d 12Ac 1)

P(acceptance on second sample) = Pa- P(d 12Ac 1)

P(rejection) = 1-Pa.

Therefore
(9.45)

100

0 L--------------------------------------
0,00 0.02 0.04 0.06 0.08 0.10
p'

FIG. 9.4. ASN curves for (a) single plan: n=SO, Ac=3 and (b) double plan:
n1=n 2=SO, Ac 1=1, Re 1=4, Ac 2=4, Re 2=S.
 250 ACCEPTANCE SAMPLING FOR ATTRIBUTES

For double sampling plan when p'=.04 we thus have

AT! 50P(d 1 ~1) + lOO[Pa - P(d 1 ~1)) + 2000[1-Pa)

50(.406) + 100(.662-.406) + 2000(1-.662) 722

Note that the sum of the weights is one.
The ATI curves are drawn in Figure 9.5 from the results given in Table
9.3. The curves start at the minimum sample sizes, and steadily rise to
N=2000 "asymptotically."
9.8.1.* Truncation of Inspection. When using single sampling on a

series of lots, ordinarily the full sample of n pieces is inspected on each

lot so as to complete the record. (An exception is made, however, if the great
majority of the pieces should prove to be defectives.) By completing the entire
sample we thereby make it easy to run an np chart (or a c chart if concern is
with defects), and also we can easily give each lot the same weight in determ-

ining p for the producer's process.

On the other hand, in double sampling we always complete the first sample
for the record as in single sampling, but since a second sample is not always
required for a decision, we are at liberty to curtail the inspection as soon

2000

1500

1000

ATI

500

0.01 002 0,03 I 0,04 0.05 0.06 0.07

p

FIG. 9.5. ATI curves for (a) single sampling plan: n=80, Ac 1=3, (b) double
plan: n 1=n 2=50, Ac 1=1, Re 1=4, Ac 2=4, Re 2=5.
9.8. Double Sampling Plans 251

as a decision is reached. Thus in our example of a double sampling plan,

suppose that we find d 1=2 on the first sample. We are then allowed two addi-

tional defectives on the second sample to permit acceptance. Suppose we find

none in the first 48. We would know then that we will have an acceptance and

could stop inspection. Why? Or, if we had one additional defective in the first

49 of the second sample we would again have a sure acceptance, even if the last

one were a defective. But these possible gains of Ac 2-d 1 are so relatively

small that we commonly never curtail or truncate sampling in the case of an

acceptance decision.

But in the case of a rejection decision there can easily be a consid-

erable gain. Again consider, the same double sampling plan, and suppose

that d 1=2 which means that we go into the second sample as favorably as

possible. But even so, three defectives in the second sample could occur on

anywhere from three to 50 pieces. Thus the possible truncated second sample

sizes run from Re 2-d 1=S-2=3 to n 2=SO, or the decrease in amount of inspection

anywhere from 47 on down to zero. Hence, it is often well to truncate

inspection as soon as rejection becomes certain (on the second sample).

Curtailing the second sample only when the decision is for rejection is

called "semi-truncation."

Let us now consider the effect of "semi-truncation" on the four charac-

teristics of a double sampling plan. Quite obviously the probability of

acceptance Pa is unaffected, because we only truncate when the decision is

already assured. Horeover, for the average outgoing quality, AOQ, using

interpretation 1 of Section 9.5, we again have very closely AOQ = p'Pa,

as in (9.42). The other interpretations, if worked exactly,give complicated

formulas, but the results are very nearly the same as for 1.

The main difference occurs for the average sample number, ASN. For the

calculation of the ASN we need to use a formula available in [10). Suppose

252 ACCEPTANCE SAMPLING FOR ATTRIBUTES

that in a sample of n pieces, k defectives calls for rejection, but k-1 or

fewer yields acceptance. (For single sampling k would be Re, while for

double sampling on the second sample, k would be Re 2-d 1). We are assuming the

binomial distribution (Type B calculations). Then the average sample number

to a decision, whether acceptance or rejection for given p' is

ASN (k/p')·P(k+l or more defectives in n+l) +

+ nP(k-1 or fewer defectives inn). (9.46)

The first term on the right is the contribution under rejection, and the

second term is obviously the contribution under acceptance, the full sample

being completed.

We thus have for our example, the calculation

SO+PC2ISO,p')[(3/p')P(4 or moreiSl,p') + SOP(2 or lesslso,p')] +

+ P(3!SO,p')[(2/p')P(3 or moreiSl, p') + SOP(O or liSO,p')]

where the first 50 is for the first sample, the inspection of which is

assumed to be always completed. Then the other two terms are the respective

relative additional contributions from going into the second sample with 2

and with 3 defectives, so that the k's are 3 and 2. Such calculations are

something of a chore, but tables can be useful. The calculations can be readily

programmed on a computer.

A little reflection will make it clear to the reader that semi-truncation

has little effect on the ASN curve if p' is relatively low, but for higher p',

when the probability of a second sample is substantial it can have a rather

marked effect.

As for the average total inspection ATI, semi-truncation produces no

effect, because when the eventual decision is acceptance either n 1 or n 1+n 2

pieces are inspected just as for the non-truncation case; and whenever the

decision is rejection we must look at all N pieces whether under truncation

or not, the probability of such being 1-Pa.

9.9. Further Notations on OC Curves 253

9.9. Further Notations on OC Curves. In this Section we introduce a

few additional notations, related to the OC curve. First we may consider

the inverse job of finding p' for a given Pa. Consider Table 9.4 for

the single sampling plan n=lOO, Ac=5. To find Pa, one begins at the first

column, proceeds to the next and then to the third column. To solve the

inverse problem, finding p' for given Pa, we start at the third column with

our Pa, then look in Table VII, in the c=5 column for the given Pa=.500, say.

We find P(5 or less\c'=5.6) = .512 and P(5 or less\c'=5.8) = .478. Interpolation

gives P(5 or less\c'=5.67) = .500. Then we divide this np' value by n to

find p'. Such a p' is designated p• 50 . Two other such special points of

interest, p• 95 and p• 10 , are also given in the table.

Table 9.4. OC Curve, Type B calculations: binomial (exact), Poisson

(approximate) for n=lOO, Ac=5. Additional points p• 95 , p• 50 , p• 10

also calculated.

Process Expectation Poisson Binomial

Fraction p' np' Pa = P(d < 5) Pa P(d ::_ 5)

.01 1 .999 .999

.02 2 .983 .985
.03 3 .916 .919
.04 4 .785 .788
.05 5 .616 .616
.06 6 .446 .441
.07 7 .301 .291
.08 8 .191 .180
.09 9 .116 .105
.10 10 .067 .058
.0261 p'9s 2.61 .950
.0567 P'5o 5.67 .500
.0930 p'lO 9.30 .100
 254 ACCEPTANCE SAMPLING FOR ATTRIBUTES

The fraction defective p• 50 , is commonly called "indifference quality."

Quite a few writers and practitioners call p'gs acceptable quality level,
and p• 10 lot tolerance proportion defective. But this practice does not

follow other common practice, for example the military standard, MIL-STD-1050
[11]. There the acceptable quality level AQL is a nominal process per cent

defective, such that, if a producer runs at this level, the great majority of

his lots will be accepted. This type of definition is given in [12]:

"Acceptable Quality Level IAQL] The maximum percent defective (or

the maximum number of defects per hundred units) that, for purposes of

acceptance sampling, can be considered satisfactory as a process avt;rage.

"Note: When a consumer designates some specific value of AQL for
a certain characteristic or group of characteristics, he indicates to the

supplier that his (the consumer's) acceptance sampling plan will accept

the great majority of the lots that the supplier submits, provided that the

process average level of percent defective in these lots be no greater than

the designated value of AQL. Thus the AQL is a designated value of percent
defective (or of defects per hundred units) that the consumer indicates
will be accepted a great majority of the time by the acceptance sampling

procedure to be used."

Therefore, although often an AQL does equal p• 95 , this is by no

means always the case. When it is true then the producer's risk, PR = a .OS.

Similarly we find the following definition in [12]:

"Lot Tolerance Percent Defective [LTPD] Expressed in percent defective,

the poorest quality in an individual lot that should be accepted."

Here again LTPD is not tied to a definite risk value. However, many

think of an LTPD in terms of a consumer's risk of .10. If this is the case

then we have in various segments of the literature the following notations:

LTPD = pt = p• 10
The last is "limiting quality" in the Military Standard MIL-STD lOSD [11].
9.10. Characteristics for Defects Plans 255

We have shown in Figure 9.6 some of these notations. The regions along
the p' axis aTe not specifically bounded by p• 95 and p• 10 , but often would
be so thought of in practice. In Table 9.4 are also given the exact
probabilities (to the best three places) from the binomial. Note that
these exact Pa's lie above the Pa's from the Poisson approximation, for p'
up to .04, and below from p' :::_ .06. Thus, as is typical, the binomial gives
somewhat more square-shouldered or discriminating OC curves than does the

Poi·sson approximation. If we were to use a Type A OC curve calculation

with the hypergeometric distribution, the OC "curve" (actually a series of

discrete points for the possible D values) would be even more square-shouldered,
depending upon N. Thus usage of approximations tends to overstate the risks,

both PR and CR. The OC curves cross at about p' = Ac/n.

9.10. Characteristics for Defects Plans. As has been mentioned earlier

acceptance sampling plans might also be given in terms of defects per hundred

units rather than the per cent defective. Defects plans can be adapted to

0.75

Po 0.50

0.25

0.00 ,___ _ _ _ _ _ _ _ _ _ _ _ _ _ ___,..___ _

0.00 0.02 p~5 0.04 P50 o.os 0.08 Po.IO 0.10 P

Acceptable Marginal Rejectable

FIG. 9.6. An OC curve for n=lOO, Ac=5, showing various notations.

Probabilities from Table 9.4.
256 ACCEPTANCE SAMPLING FOR ATTRIBUTES

opportunities for defects on bulk product, for example areas of paper or

tinplate. But, commonly, defects plans are used for counts of defects on

samples of n, or n 1 and n 2 discrete pieces.

When the average number of defects per hundred pieces is small, such

as 1.0 or 1.5, it makes little difference which basis is used. For

example, if we have a lot of 10,000 pieces, and the acceptable quality

level AQL is 1 defect per 100 pieces, then we might think of randomly

"throwing" 100 defects at the lot of 10,000. They could well all "hit"

different pieces, thus giving 100 defective pieces. But more than one

defect could conceivably reach the same piece, thus giving slightly fewer

than 100 defective pieces in the 10,000. But we would only expect about

one piece to contain two defects. So C' = 1.0 defects per hundred pieces

is about equivalent to .99 per cent defective.

On the other hand if the AQL = 10 defects per hundred pieces then there

is likely to be sufficient duplication to make this equivalent to 9

defectives per hundred units. Above AQL's of 10, the usual thing is to use

defects per hundred units rather than per cent defective, P', for AQL. In

fact we can sometimes regard as many as 1000 defects per hundred units as

satisfactory quality. It all depends on the type of product and the

"defects" in question.

As to calculations for an OC curve, the Type B calculations, where we

are concerned with a process average of C' defects per hundred units, the

Poisson distribution is not approximate but is exact, if the usual Poisson

assumptions are met. When Type A calculations are to be made, we have a

finite lot to draw from, instead of a process. Defects on the pieces in

the finite ~cannot be exactly Poisson, but for most all practical purposes

we can assume a Poisson distribution.

9.11. Devising a Sampling Plan 2S7

The OC and ASN calculations are unchanged essentially. For AOQ and ATI,
in which 100% sorting of the lot occurs, such sorting must be possible,
which is not always the case, for example with some kinds of bulk product.

9.11. Devising a Sampling Plan to Match Two Points on an OC Curve.

There is considerable intuitive appeal to the following approach: Decide

that we approve of a fraction defective of, say, p• 1 = .01, or less, but frown
upon one of p• 2 = .040, or more. Then we may decide upon risks respectively
of .OS and .10. Thus

AQL .010 PR .OS

RQL .040 CR .10.

Now suppose we seek a single sampling plan to satisfy these two conditions.
We can only choose whole numbers for Ac: 0, 1, 2, etc. Therefore we cannot

expect to hit ~ctly both points desired. We may choose to match either

point as well as possible, using the Poisson approximation, and then for the
plan so determined, the risk at the other point will be somewhat decreased,
if anything. It leads to smaller sample sizes if we choose to match the

(p• 2 ,S) point on the OC curve as closely as the Poisson approximation permits,

instead of seeking to match the (p• 1 , 1-a) point. But if we are anxious to

cut the S risk rather than the a risk we can match the (p 1 1 ,1-a) point, at

a cost of a somewhat larger n.

Table 9. 5 considerably ;facilitates the finding of a desired plan. ·For

it we have assumed a= .OS, S = .10 as is widely used. Then we find the

"operating ratio" [13] :

R0 = p' 2/p' 1 . (9.47)

This ratio measures the relative strength of discriminating power, smaller

ratios being :related to greater power.

To use the table, we need to agree upon a= .OS, S = .10, then we divide

our specified p• 2 by p• 1 and have the desired R0 . This will likely not be
 258 ACCEPTANCE SAMPLING FOR ATTRIBUTES

Table 9.5. Operating Ratios Ro for Single Sampling Plans, Using Poisson
Approximation, For Fitting Two Points on OC Curve: p 1 1 , PR=a= .05; p 1 2 ,
CR=B= .10. Acceptance Numbers Ac, and Expected numbers of Defectives per
Sample also Given.

R0* c = Ac np'2

44.7 0 2.30
10.9 1 3.89
6.51 2 5.32
4.89 3 6.68
4.06 4 7.99
3.55 5 9.27
3.21 6 10.53
2.96 7 11.77
2. 77 8 12.99
2.62 9 14.21
2.50 10 15.41
2.40 11 16.60
2.31 12 17.78
2.24 13 18.96
2.18 14 20.13
2.12 15 21.29
2.07 16 22.45
2.03 17 23.61
1.99 18 24.76

*Since sums of terms of the Poisson may be obtained from the chi-square
2
distribution, R =X (ZAc + 2 • B above) Also np• 2 = .5 iC2Ac + 2, B above).
O x2 (2Ac + 2, 1-a above)

an entry in the table. If not, we take the next smaller entry. In the row

so determined is the acceptance number Ac. Also in this row we find the

Poisson parameter np• 2 . Given such an entry and p 1 2 specified we find n by

division. This gives usn and Ac, thus determining the plan. The a risk

will be less than .OS, and as a matter of fact, if we make the probability
9.12. Use of Control Chart Record 259

calculation using the binomial distribution for p• 2 and n, the Pa at p' 2 ,

that is,

P(d ~ Acln,p' 2)

will be less than 8 also, because the binomial gives a slightly more discrim-

inating OC curve than does the Poisson approximation.

Let us consider the example at the beginning of the section: p• 1 = .010,

PR =a = .05; p• 2 = .040, CR =8 = .10. Then the desired R0 = 4.00. This

says use Ac 5. Also np• 2 = 9.27. Therefore n(.040) = 9.27 or n = 232.

So the plan is n = 232, Ac = 5. Now how does this plan do on p• 1 = .01?

np• 1 2.32, and

Pa = P(d ~ sJnp• 1 2.32} .968

and hence a = .032.

A set of plans of this nature is given in [14]. If one wishes a double

sampling plan to match two given points on the OC curve, the problem is not

at all easy. But one can seek a single sampling plan in [11] with the

desired OC curve, and then this standard provides a double sampling plan with

similar characteristics.

9.12. Use of Control Chart Record with Acceptance Sampling. It is

good practice to keep a control chart record of each producer on each part

number which is sample inspected. Each lot is represented by the one single

sample or by the first sample in double sampling. Since the sample sizes

are equal, in general, an np or a c ~ha~t can be used very easily. In this

way a history of the ~ and stability of performance of the producer can

be obtained. This is of much help in dealing with the producer.

In one case the chief inspector of a manufacturer of hot water heaters

announced that they wanted to use a sampling plan with an AOQL of .0025 or

.25% for the temperature control which, while inexpensive, costs considerably

to be removed and replaced. The manufacturer, was immediately descended

260 ACCEPTANCE SAMPLING FOR ATTRIBUTES

upon by the producer's president, production manager and chief inspector,

all of whom said in various words "We cannot live with this." Then the man-

ufacturer's chief inspector said "What do you mean you cannot? Look at these

records. For months you have been running at .001 or .101~." The visitors

were gratified that they were shown to be so good, but a bit chagrined that

they did not know how good they were.

If a producer's record on a part is sufficiently good, the consumer

may be able to safely reduce the sample size used. This is the philosophy

behind reduced sampling in the Military Standard [11]. It is simply that if the

probability of a lot coming in at say 2% or worse is extremely low, then

we may well use a sampling plan with a S risk at p' = .02, of as much as

.SO. This is because past history strongly indicates that such a "bad" lot

just is not going to be coming along!

On the other hand if p is relatively high and/or performance is erratic,

then we may wish to use a tighter sampling plan 1 usually in the form of lower

acceptance numbers, but it could be with a larger sample size. Thus we

might change from a plan with an acceptable quality level AQL of 1% to one

with an average outgoing quality limit AOQL of 1%. The former is lenient

toward lots at 1% and fairly lenient with lots somewhat worse, while the

AOQL plan will be tighter and will make sure the outgoing quality averages

no worse than 1%.

If the control chart record indicates poor enough quality we may even

go off from sampling inspection altogether.

9.13. Summary - General Principles of Sampling Acceptance and Some Mis-

conceptions. We now wish to emphasize certain basics of acceptance sampling.

9.13.1. Prerequisites to Sound Decision Makiug. We again emphasize

four prerequisites. We first need a clear, unequivocal definition for each

defect or characteristic to be inspected for. Secondly, we need an objective

9.13. Summary 261

way to inspect. In the case of visual defects this may take the form of

limit samples, which provide the dividing line between a defect and a non-

defect. For a measurable characteristic we need a good measuring technique,

or at least a good go-not-go gage. Thirdly, we need to obtain by proper

training and supervision, careful, accurate inspection. Finally, we need

good representative samples from each lot. The pieces for the sample can

be drawn purely at random from small lots. This can be done by use of ran-

dom sampling numbers if each piece can be assigned a number. At the very

least the sampler should draw pieces from all around the lot. If a lot is

large, then such pure random sampling may be infeasible. A stratified

approach can be used by drawing the same size small sample from each container or

carton. Or, if there are too many such cartons, we might chaos e a random

sampling of cartons by random numbe~s, then from each one chosen, a small

sample is drawn to make up the required total sample. In double sampling,

many people draw enough pieces for the two samples, should both be needed.

The whole aim is to obtain as representative and unbiased a sample as possible.

In one actual industrial case a sample of 100 piston ring castings was

drawn from a box containing 3000. The sample contained 25 defective ones or

25% defective. They thought that the uninspected 2900 might well have over

700 defectives. So the 2900 were sorted 100% to salvage the good ones. But

in the 2900 only 4 more defectives were found! Was the sample of 100 randomly

chosen?

9 ..13.2.conditional Character of Probabilities of Acceptance. We have earlier

mentioned that the probability of acceptance Pa P(acclp'), always depends

upon the incoming quality, for example p' or c'. Thus if, say, we let

p' = .06 be p• 10 , then Pa = .10. This leads some unwary individuals to think

that the probability of having a lot at 6% to work on is .10, that is, one

lot in ten of those passed will be at 6%. But such a situation ~ only
262 ACCEPTANCE SAMPLING FOR ATTRIBUTES

arise if all incoming lots are at 6%. Meanwhile nine lots in ten will have
been rejected, and the sensible consumer would have quickly tightened up

inspection or have begun sorting each lot 100%.

In actuality, if there can somehow be ascertained an~ priori dis-

tribution of incoming p' qualities for the producer, it is then possible

to use this distribution and the OC curve to find the ~ posteriori distri-

bution of outgoing quality levels by use of

P(p')P(acc\p')

for a set of frequency classes of p'. Occasionally it is possible from past

records to obtain such an a priori distribution of lot parameters p', but

probably not very commonly.
It is therefore best to think of the OC curve as giving the prob-

ability of acceptance Pa of a lot at p', should such a one be offered. We

may then perhaps from control chart records, have some subjective idea of

the likelihood of a bad quality lot coming along for acceptance. From this

we may choose an appropriate OC curve. Or probably better, we may use a

sampling acceptance system such as [11], as described'in the next chapter.

9.13.3. What Is a Defective? This question has become more and more

important especially as greater protection is being provided for the con-

suming public. Judges, attorneys and juries may well take an unfriendly

attitude toward the presence of a "defective" in any product being used by

the public. For this reason many organizations have begun speaking of "non-

conforming" pieces rather than "defectives." We have in this book been using

the terms "defective" and "defect" frequently and consistently. But it must

be emphasized that there is an enormous range in the seriousness of a "defect"

from something quite dangerous, to the mildest sort of blemish. For example,

we may have categories of defects: critical, major A, major B, minor and

incidental. A critical defect may be so important or dangerous that we must

9.13. Summary 263

inspect each piece in the lot not once, but perhaps many times. Only after

a .long history of perfect production may we permit some sampling. It is for

major and minor defects that we can well use sampling acceptance plans suit-

ably chosen. For incidental defects, if sampling, then only a ve~y small

sample size need be used. Or perhaps a spot check only need be given.

Of all the characteristics of industrial .parts, some deliberate choice

should be made between (a) inspecting and testing each piece many times, (b)

inspecting 100%, (c) sampling inspection, (d) spot checking, and (e) no

inspection at all for some characteristic. Things influencing the choice

are (a) the seriousness of the defect, (b) the difficulty and cost of in-

specting for it, (c) the likelihood of its occurrence, and (d) whether or

not the test destroys the part.*

The author sometimes wonders what proportion of characteristics can

actually be carefully patrolled. For example, of 20 pages of specifications

for insulated wire, how many characteristics are actually checked regularly?

9.13.4. The Proportion of Sample Size to Lot Size is Quite Unimportant.

We have been giving major emphasis to Type B calculations, namely where we

consider lots to have been taken from a process in control at p' or c', or

where we assume infinite 1Qt size, so that the binomial and Poisson apply

exactly. Now if we are concerned with defectives, how much difference does

it make to take into account the lot size, and thus to use Type A calculations,

which call for the use of the hypergeometric distribution? As previously

mentioned, if n/N ~ .1, it makes very little difference. We may even go

furthe~Consider Figure 9.7, where all the OC curves are for the one single

plan: n=lO, Ac=O. However, the lot sizes vary: 20, 40, 100, 300 and in-

* Sometimes the test may weaken the part or increase its likelihood of fail-

ure. We may well treat such cases like destructive testing.

 264 ACCEPTANCE SAMPLING FOR ATTRIBUTES

FIG. 9.7. OC curves for the plan n=lO, Ac=O. The lot sizes are 20 for the
bottom curve, 40 for the next, 100, 300, finally infinity for the top curve.
Last is a Type B curve, the others are Type A curves. For most practical
purposes the top four curves may be considered alike. Reproduced with
permission from I. W. Burr, "Engineering Statistics and Quality Control",
McGraw-Hill, New York, 1953, p. 316.

finity. The last is a Type B OC curve, the others Type A. The lowest curve

is for N=20, in which we are sampling 50% of the lot. This OC curve is not

greatly below that for N~40, in which we are sampling 25% of the lot. All

five of the curves are really quite similar and scarcely differ from a

practical viewpoint.

Actually the bottom four curves were drawn to guide the eye, because

they really consist only of a series of discrete points. For example, if

N~20, then the only possible p' values are .00, .OS, .10, .15 etc., for

D=O, 1, 2, etc.

From such studies we can say that a sampling plan of say n=lO, Ac=O

will do just about the same job on N=l,OOO,OOO as it will on N=lOO, or even

T\!::40.
Now there are two practical considerations which suggest that somewhat

larger sample sizes be used with larger N's than smaller N's. First, it is
 9.13. Summary 265

more difficult to sample randomly a very large lot. Secondly, it is a more

serious error, perhaps, to erroneously reject a very large lot and thus call

for 100% sorting or salvaging. But the amount of increase inn is not pro-

portional toN. This may be seen in MIL-STD lOSD [11].

It is not the proportion that the sample is of the lot that determines

the discriminating power of the sample, but is the gross size of the sample

and the acceptance-rejection criteria.

9.13.5. Misconception 1: "We are being scientific and fair by sampling

the same percentage of each lot." As we have just seen in Section 9.13.4, the

percentage of the lot sampled (n/N)lOO% is relatively unimportant, certainly

so in comparison with n itself. But we still find people taking, say, 10%

of each lot for the sample. This malady is called "percentitis."

To further consolidate this viewpoint consider Figure 9.8, in which are

shown a number of OC curves, in each of which the sample is 10% of the lot.

"'u
J'" 0.60
0.
"'uu
t;
4--
0

0.12 0.14

FIG. 9.8. OC curves for four 10% sampling plans, each with Pa = .90 at
p' = .01. Reproduced with permission from I. W. Burr, "Engineering Statistics
and Quality Control", McGraw-Hill, New York, 1953, p. 317.
266 ACCEPTANCE SAMPLING FOR ATTRIBUTES

Each curve has about the same Pa of .90 at p'=.Ol. But there the resemblence

ends. For the N=lOO, n=lO, Ac=O plan, Pa decreases very slowly. This is quite

an undiscriminating plan. As N increases we use appropriate Ac numbers to go

along with n=.lN, to give P(alp'=.Ol)=.90, and the curves become more and

more sharply discriminating. The plan for N=lOO is very likely much too

undiscriminating, suggesting a greater sampling than 10%, whereas that for

N=25,000 is likely to be far more sharply discriminating than is needed.

Thus we need much less sampling than 10%, when N is very large.

Moral: pay attention to the OC curve much more than to the proportion

n/N sampled.

9.13.6, Even 100% Inspection Does Not Give Complete Assurance. The

author has seen many instances of the inefficiency of 100% sorting. In one

case an inspector was using a dial gage with a pointer which was supposed

to stop in a green sector, if the piece was between limits. The assistant

plant manager noted that the inspector seemed to be just going through the

motions of passing the pieces. So he covered up the dial with his hand, and

the inspector passed the next 40 pieces, without being able to see the dial

at all! In another case, a lot of washing machine motors, supposedly checked

100% electrically, all were packed up and shipped without any rotors (the rotating
shaft and core).

Of course not all 100% inspection is this bad. But studies have shown

that seldom is 100% inspection more than 80% efficient. Thus only 80% of

the defectives present 2£ ~. will actually be found, Meanwhile many

good pieces may well be rejected!

Efficiency of such sorting tends to be highest where there is but one

objective defect possible per piece and the inspection is not too hurried.

But efficiency can be quite low if there are many ~ defects to be looked

for and the pieces move along rather rapidly on a conveyor, say. Finally

efficiency tends to be lower when defects are quite rare.

9.13. Summary 267

9.13.7. Misconception 2: "We can only use perfect lots." As has

just been mentioned, 100% inspection does not give complete assurance of

perfect lots. In fact even 300% or 400% inspection may not weed out all

of the defectives present. How then can one hope to have perfect lots of

articles to work with? The biggest hope of success is to concentrate on

improving the process to make virtually all product right in the first

place. It is still true, as has often been said "Quality is made at the

machine, not at the inspection station."

But still there simply is no way to provide absolute guarantees of per-

fect lots. Commonly we must settle for something slightly less than perfec-

tion. In fact we are often forced into trade-offs between extremely high

quality at very high cost and less high quality at lower cost. Fortunately

a great many types of defects will show up in assembly or at final inspection

and be removed or fixed.

9.15.8. Misconception 3: "We have a strong sampling plan because we do

not tolerate any defectives in the sample." Many make such a boast. But

the true picture shows up in the OC curve. Using Ac=O, we of course find a

lazy OC curve if n is small, say, 10 or 20. There is a good chance for in-

ferior lots to get by, i f offered. See Figure 9. 7 again. On the other hand

if n~200, say, then for p'~.Ol, Pa=.l334 from a binomial table, which is very

low. But if p'= .01 would be satisfactory, such a Pa would be most unfortunate.

Or if p'=.OOl, Pa is still only up to .8186. If this is what you like and

need, fine. But if so you must spell out your needs carefully to the pro-

ducer, otherwise you will have trouble with him, for you are forcing him to

make virtually perfect lots to have a good chance of their passing. However,

perfection always comes only at a premium price.

The principle here is that OC curves for Ac=O always rise relatively

steeply to Pa=l at p'=O, being concave upwards throughout. The OC curves

for Ac > 0 all have points of inflection having the concave side down for
268 ACCEPTANCE SAMPLING FOR ATTRIBUTES

low p' and up for high p', which gives an OC curve commonly making more

sense.

9.13.9. Misconception 4: "If a sampling plan is any good it ought

to give the same decision every time on a given lot." This bit of nonsense

only comes from those who have not thought through the implications of the

OC curve. Only if p'=O are we sure of the same decision (acceptance) every-

time. When p' > 0, there is always some chance for each of the two decisions

to occur. Only when Pa is very close to 1, or to 0, will the decision be

the same nearly all the time.

9.13.10. Misconception 5: "Acceptance sampling can only be csed i f the

process is in control." Such an idea comes about perhaps in our study of

OC curve calculations for the Type B case. But that is only to find out how

the plan operates on lots from a process at p'. There is nothing that says

p' remains fixed. It can vary wildly, and Type A OC calculations tell us

what Pa is for a lot of any given fraction defective p':D/N.

9.13.11. Summary. We have the following characteristics in the form

of four curves:

A. Protection

1. OC curve, Pa vs. p' or c'

2. AOQ curve, AOQ vs. p' or c'

B. Cost of inspection

1. ASN curve, ASN vs. p' or c' for a decision

2. ATI curve, ATI vs. p' or c' for total inspection.

These may be based on p' or c' for the underlying process at these levels:

Type B; or for an individual lot at these levels: Type A. Quality is often

defined in terms of defectives per hundred units P', or defects per hundred

units C'.
Problems 269

PROBLEMS

For the following single sampling plans find an appropriate four or

five points and sketch the following curves, labeling axes: (a) OC, (b) AOQ, (c) ASN,
(d) ATI. (e) What is the AOQL for the plan?

9 .1. N 4000, n = 200, Ac 1' Re 2 defectives [11]

9.2. N 4000, n = 200, Ac 3, Re 4 defectives [11]
9.3. N 4000, n = 200, Ac 5, Re 6 defectives [11]
9.4. N 1000, n = 80, Ac 2, Re 3 defectives [11]
9. 5' N 1000, n 80, Ac 5' Re 6 defectives [11]
9.6, N 500, n 50, Ac 7, Re = 8 defects [11]
9. 7. N 500, n = 50, Ac 10, Re 11 defects [11]
9.8.N 4000 n 95, Ac 3, Re 4 defectives [1]
9.9.N 1000 n 65, Ac 2, Re 3 defectives [1]
9.10.N=500n 39, Ac 1' Re 2 defectives [1]

The following double sampling plans correspond respectively to those above,

9.1 to 9.10, wi~hin the cited references. Follow the same instructions for
parts (a) through (e). Hint: Use p''s for the OC curve which give simple
n1p''s to avoid interpolation in Table VII.

9. 11. N=4000, n 1=n 2=125, Ac 1=0, Re 1=2, Ac 2=1, Re 2=2

9.12. N=4000, n 1=n 2=125, Ac 1=1,Re 1=4, Ac 2=4, Re 2=5
9.13. N=4000, n 1=n 2=125, Aci=2, Re 1=5, Ac 2=6, Re 2=7
9.14. N=lOOO, n 1=n 2=50, Ac 1=0, Re 1=3, Ac 2=3, Re 2=4
9.15. N=lOOO, n 1=n 2=50, Ac 1=2, Re 1=5, Ac 2=6, Re 2=7
9.16. N=500, n 1=n 2=32, Ac 1=3, Re 1=7, Ac 2=8, Re 2=9
9.17. N=SOO, n 1=n 2=32, Ac 1=5, Re 1=9, Ac 2=12, Re 2=13
9.18. N=4000,n 1=80, n 2=140, Ac 1=1, Re 1=8, Ac 2=7, Re 2=8
9.19. N=lOOO, n 1=38, n 2=87,Ac 1=0, Re 1=5, Ac 2=4, Re 2=5
9.20. N=500, n 1=34, n 2=56, Ac 1=0, Re 1=4, Ac 2=3, Re 2=4

9.21. For the plan N=5000, n=300, Ac=4, Re=5, find p• 95 ,p• 50 ,p• 10 .
9.22. For the plan N=4000, n=200, Ac=2, Re=3, find p• 95 ,p• 50 ,p• 10 .
 270 ACCEPTANCE SAMPLING FOR ATTRIBUTES

Devise a single sampling plan to meet (approximately) as well as possible

the following pairs of points on the OC curve:
9.23. pl=.Ol, a=.OS; Pz=.04, S=.lO
9.24. pl=.OOS, a=.OS; Pz=.06, S=.lO
9.25. pl=.02, a=.OS; Pz=.07, S=.lO
9.26. pl=.OlS, a=.OS; Pz=.08, 6=.10
9.27. How might you go about finding Pio for a double sampling plan?

9.28. Why is it in the nature of a double sampling plan that rejection is

never made on the occurrence of but one "defective" piece in the one or two

samples drawn?

9.29. The following plan was devised for a part that was extremely

expensive to test, lot size being 100: n 1=2, Ac 1=0, Re 1=2, n 2=1, Ac 2=1, Re 2=2.

Sketch the OC curve and comment. (Hint: It is easy to derive an equation

for Pa) .

9.30. What form would a triple sampling plan take? Give an example of

one to illustrate. What advantage might it have?

9.31. In Problems 9.12 to 9.17, usually Re 1 < Re 2 , whereas in the older

9.18 to 9.20 Re 1=Re 2 . What advantage is there to setting Re 1 < Re 2? (Consider

9.18, suppose d 1=7, how good is the conditional probability of acceptance now?)

9.32. Compare single and double sampling as to the relative practical

advantages of each.

9.33. When is the Poisson distribution a reasonable approximation to the

binomial distribution? The binomial for the hypergeometric? The Poisson for

the hypergeometric?

9.34. What might be wrong with using two sampling plans, single and

double with similar OC curves as follows: Start out on the double plan, then

if one takes a second sample, only continue on the second sample till we reach

a total sample equivalent to n of the single plan and now use Ac of the single plan?

Does this give the same OC curve?

 References 271

9.35. Verify the other three probabilities for the example in Section
9.4.2.
Find for the hypergeometric distribution the following terms, either

using direct calculation and Table IX, or [8].

9.36. P(3J50,10,8)
9.37. P(lj40,6,4)
9.38. p (4 J60' 12 ,6)
9.39. P(2j30,5,4)

9.40. One of the helpful symmetries used in [8] is that n and D may be

interchanged without changing the value. Thus P(2l30,5,4) = P(2J30,4,5) in

(9.29). Prove this is true.

9.41. Note the recursion relation (9.35) for the hypergeometric distribution

Use the same approach to derive simpler recursion relations for the binomial

(2.30) and the Poisson (2.35).

9.42. Find the ASN without and with truncation of second sample (under

rejection), for the double sampling plan: N=2000, n 1=n 2=40, Ac 1=1, Re 1=4, Ac 2=4,

Re 2=5 at p'=.05. Use binomial calculation, and appropriate tables if available.

References

1. H. F. Dodge and H. G. Romig, "Sampling Inspection Tables." Wiley, New

York, 1959.

2. U. S. Dept. of the Army, "Tables of the Binomial Distribution." Natl. Bur.

Standards, Appl. Math. Ser. 6 U. S. Govt. Printing Office, Washington, DC,
1950.

3. Harvard Uni v., Comput. Lab., "Tables of the Cumulative Binomial Probability
Distribution." Harvard Univ. Press, Cambridge, Massachusetts, 1955.

4. W. H. Robertson, "Tables of the Binomial Distribution Function for Small

Values of p." Tech. Services, Dept. of Commerce, Washington, DC, 1960.

5. H. G. Romig, "Fifty to 100 Binomial Tables." Wiley, New York, 194 7.

6. S. Weintraub, "Cumulative Binomial Probability Distribution for Small

Values of p." Free Press of Glencoe, Collier-Macmillan, London, 1963.
272 ACCEPTANCE SAMPLING FOR ATTRIBUTES

7. H. Working, "A Guide to Utilization of the Binomial and Poisson Distributions."

Stanford Univ. Press, Stanford, California, 1943.

8. G. J. Lieberman and D. B. Owen, "Tables of the Hypergeometric Probability

Distribution." Stanford Univ. Press, Stanford, California, 1961.

9. I. W. Burr, Some approximate relations between terms of the hypergeometric,

binomial and Poisson distributions. Communications in Stat. 1, 297-301
(1973).

10. Statistical Research Group, Columbia Univ., "Sampling Inspection." McGraw-

Hill, New York, 1948.

11. Military Standard MIL-lOSD (ABC), "Sampling Procedures and Tables for
Inspection by Attributes." Dept. of Defense, 1963. Also Amer. Nat.
Standards Inst. 21.4, and available from Amer. Soc. for Quality Control,
Milwaukee, Wisconsin.

12. ASQC Standard A2-1971, "Definitions and Symbols for Acceptance Sampling
by Attributes," Amer. Soc. for Quality Control, Milwaukee, Wisconsin.

13. P. Peach and S. B. Littauer, A note on sampling inspection by attributes.

Ann. Math. Statist. 17, 81-84 (1946).

14. R. L. Kirkpatrick, Binomial sampling plans indexed by AQL and LTPD.

Indust. Quality Control 22 (No. 6), 290-292 (1965).
 CHAPTER 10

SOME STANDARD PLANS FOR ATTRIBUTES

10.1. Introduction. Chapter 9 discussed fundamental characteristics of

acceptance sampling plans for attributes. In particular two of the

characteristics were related to protection, namely the OC and AOQ curves,

while the other two were related to cost, that is, the ASN and ATI curves.

Calculations of these various curves for single and double sampling plans were

discussed and illustrated. In Section 9.11 there was presented a method for

finding a plan matching two points on the OC curve, a table being given for
I
the case when p 95 and p 10 are both specified.

In the foregoing we have been concerned with individual sampling plans,

and their analysis. In the present chapter we consider collections of sampling

plans or systems, integrated to satisfy appropriate goals. Different aims

give rise to differing sets of plans. It is a great convenience to have a set

of plans all calculated out and available. But to make sound use of them one

should be very familiar with the basic elements as discussed in Chapter 9.

Also we take up multiple sampling, a natural extension of single and

double sampling.

The early work in acceptance sampling by attributes was largely concentrated

in the Bell Telephone Laboratories, beginning about 1923. Among the concepts

originated by engineers and mathematicians in that organization were acceptable

quality level (AQL), lot tolerance per cent defective (LTPD), consumer's risk

(CR or S), producer's risk (PR or a), average outgoing quality (AOQ), average

outgoing-quality limit (AOQL), average total inspection (ATI), reduced inspection,

tightened inspection, and double and multiple sampling. Among the men

273
274 SOME STANDARD PLANS FOR ATTRIBUTES

responsible for the developments were (alphabetically) H. F. Dodge, G. D. Edwards,

T. C. Fry, E. C. Molina, P. S. Olmstead, H. G. Romig, and W. A. Shewhart.
W. Bartky working with Western Electric Co., did important work in multiple
sampling. The famous Dodge-Romig tables were one outgrowth of this Bell
Laboratories work.
In order to meet the enormous problems of acceptance of material during
World War II, the Army Ordnance Department carried on further work in acceptance

sampling procedures. Several key people from the Bell Laboratories worked on
the project with Army Ordnance personnel. Among the latter were H. R. Bellinson,
G. R. Gause, L. W. Shaw, L. E. Simon, and A. Stein. The ultimate result

was the development of a set of 12 integrated tables of sampling plans. These

were commonly called the Army Ordnance Tables.
Early in 1945, the Statistical Research Group of Columbia University,

which had been working on statistical problems arising in national defense,

was asked by the United States Navy to prepare a manual of tables, procedures,
and principles for sampling inspection. The Navy manual, or Navy Tables, for
sampling inspection by attributes was thus produced. Similar work for
variables was discontinued in August, 1945. Subsequently the attribute tables

were adopted for use by both the United States Army and United States Navy, and
they became known as the Joint Army-Navy Service Forces Tables, or JAN Tables.
The men chiefly responsible for the preparation of the Navy Tables were H. A.
Freeman, Milton Friedman, Frederick Mosteller, L. J. Savage, D. H. Schwartz,
and W. A. Wallis. See [1].
The history of acceptance sampling by attributes is of much interest. An
excellent and authoritative presentation of this history is given in [2], which

covers the evolution through the currently most widely used system ABC-STD-105,
which is also called MIL-STD-105D. It is also an American National Standards
10.2. The Dodge-Romig Tables 275

Institute ANSI-Zl.4. We will discuss this system in Section 10.3. It is

presently in world-wide use.
10.2. The Dodge-Romig Sampling Inspection Tables. These tables were

originally published in 1929, coming out in a book in 1944. An enlarged and

extended second edition came out in 1959 [3]. The tables aim at providing

consumer protection on a minimum inspection cost. Basically and specifically

Aim of Dodge-Romig Sampling Inspection Tables.

1. To provide specified consumer protection, either a specified LTPD

' ), or a specified AOQL value.
(p 10

2. Among the many sampling plans which could supply the specified

kind and amount of consumer protection, the tables give that one
plan which gives this protection on a minimum average total
inspection ATI for given lot size N and quality level p.

Let us illustrate this aim in reference to one of the Dodge-Romig tables

here reproduced as Table 10.1. This table has six column sections corresponding
to classes of process average p. Then too the 19 rows are for classes of lot

size N. If one wants a double sampling plan providing an AOQL protection of

1%, he uses this table and enters it via his N and his estimated p ' , that is p.
(If no p is yet available the rightmost section for p= .81 to 1.00% can well

be used, tillsome history is built up.) Thus N and p- lead to a block in the

table. There one finds sample sizes n 1 and n 2 , the acceptance number c 1 for

the first sample (Ac 1 in Chapter 9), the acceptance number c 2 for the combined
sample of n 1 + n 2 (Ac 2 in Chapter 9). In this system c 2 + 1 is the rejection
number for both the first and for the combined samples (that is, both Re 1 and

Now suppose that we have N 1500. In the appropriate rows we find six

plans
 TABLE 10.1. A double sampling, average quality Dodge-Romig table.
Double Sampling Table for
Average Outgoing Quality Limit (AOQL) 1.0%

Process Average Process Average Process Average

0 to 0.02% 0.03 to 0.20% 0.21 to 0.40%
Lot Size Trial 2 Trial 1
Trial 1 Trial 2 Trial1 Trial 2
--- ·- Pt
%
Pt
%
Pt
%
"' Ct
"'
-
nt+nz cz

- - -
"' Ct
"'
-
n,+nz c,

- - -
---
"' ·"'
Ct nt+nz c,
1- 25 All 0 All 0 All 0 - - - -
26- 50 22 0 - - - 7.7 22 0 - - - 7.7 22 0 - - - 7 .7
51- 100 33 0 17 50 1 6.9 33 0 17 50 1 6.9 33 0 17 50 1 6.9
101-200 43 0 22 65 1 5.8 43 0 22 65 1 5.8 43 0 22 65 1 5.8
201- 300 47 0 28 75 1 5.5 47 0 28 75 1 5.5 47 0 28 75 1 5.5
301- 400 49 0 31 80 1 5.4 49 0 31 80 1 5.4 55 0 60 115 2 4 .8
401- 500 50 0 30 80 1 5.4 50 0 30 80 1 5.4 55 0 65 120 2 4.7
501- 600 50 0 30 80 1 5.4 50 0 30 80 1 5.4 60 0 65 125 2 4.6
601-800 50 0 35 85 1 5.3 60 0 70 130 2 4.5 60 0 70 130 2 4.5
801-1000 55 0 30 85 1 5.2 60 0 75 135 2 4.4 60 0 75 135 2 4.4
1001- 2000 55 0 35 90 1 5.1 65 0 75 140 2 4.3 75 0 120 195 3 3.8
2001-3000 65 0 80 145 2 4.2 65 0 80 145 2 4.2 75 0 125 200 3 3.7
3001-4000 70 0 80 150 2 4.1 70 0 80 150 2 4 .1 80 0 175 255 4 3.5
4001- 5000 70 0 80 150 2 4.1 70 0 80 150 2 4.1 80 0 180 260 4 3.4
5001-7000 70 0 80 150 2 4.1 75 0 125 200 3 3.7 80 0 180 260 4 3.4
7001- 10,000 70 0 80 150 2 4.1 80 0 125 205 3 3.6 85 0 180 265 4 3.3
10,001- 20,000 70 0 80 150 2 4.1 80 0 130 210 3 3.6 90 0 230 320 5 3.2
20,001- 50,000 75 0 80 155 2 4.0 80 0 135 215 3 3.6 95 0 300 395 6 2.9
50,001- 100,000 75 0 80 155 2 4.0 85 0 180 265 4 3.3 170 1 380 550 8 2.6

Process Average Process Average Process Average

0.41 to 0.60% 0.61 to 0.80 % 0.81 to 1.00 %
Lot Size -------·------
Trial 1 Trial 2 Trial 1 Trial 2 Trial 1. Trial 2
Pt
%
--- PI
%
Pt
%
n, n1+n2 c, n, n':! n1+n2 c, n'.! n1+n2
"'
Ct Ct
"' Ct C:!

1- 25 All 0 - - - - All 0 - - - - All 0 - - -· -

26- 50 22 0 - - - 7.7 22 0 - - - 7.7 22 0 - - -

1
7.7
6.9
51- 100 33 0 17 50 1 6.9 33 0 17 50 1 6.9 33 0 17 50
101-200 43 0 22 65 1 5.8 43 0 22 65 1 5.8 47 0 43 90 2 5.4
201- 300 55 0 50 105 2 4.9 55 0 50 105 2 4.9 55 0 50 105 2 4.9
301- 400 55 0 60 115 2 4.8 55 0 60 115 2 4.8 60 0 80 140 3 4.5
401 - 500 55 0 65 120 2 4. 7 60 0 95 155 3 4.3 60 0 95 155 3 4.3
501- 600 60 0 65 125 2 4.6 65 0 100 165 3 4.2 65 0 100 165 3 4.2
601- 800 65 0 105 170 3 4.1 65 0 105 170 3 4.1 70 0 140 210 4 3.9
801-1000 65 0 110 175 3 4.0 70 0 150 220 4 3.8 125 1 180 305 6 3.5
1001- 2000 80 0 165 245 4 3.7 135 1 200 335 6 3.3 140 1 245 385 7 3.2
2001- 3000 80 0 170 250 4 3.6 150 1 265 415 7 3.0 215 2 355 570 10 2.8
3001- 4000 85 0 220 305 5 3.3 160 1 330 490 8 2.8 225 2 455 680 12 2.7
4001- 5000 145 1 225 370 6 3.1 225 2 375 600 10 2.7 240 2 595 835 14 2.5
5001-7000 155 1 285 440 7 2.9 235 2 440 675 11 2.6 310 3 665 975 16 2.4
7001- 10,000 165 1 355 520 8 2 .7 250 2 585 835 13 2.4 385 4 785 1170 19 2.3
10,001 - 20,000 175 1 415 590 9 2 .6 325 3 655 980 15 2.3 520 6 980 1500 24 2.2
20,001- 50,000 250 2 490 740 11 2.4 340 3 910 1250 19 2.2 610 7 1410 2020 32 2 .1
50,001 - 100,000 275 2 700 975 14 2.2 420 4 1050 1470 22 2.1 770 9 1850 2620 41 2 .0

Trial 1: n 1 = first sample size; c,

= acceptance number for first sample
"All" indicates that each piece in the lot is to be inspected
Trial 2: n 2 = second sample size; c2 = acceptance number for first and second samples combined
p 1 =lot tolerance per cent defective with a Consumer's Risk (Pc) of 0.10

Reproduced with permission from H. F. Dodge and H. G. Romig, "Sampling

Inspection Tables," Wiley, New York, 1959, p. 210.
276
10.2. The Dodge-Romig Tables 277
n1 c1 n2 nl+n2 c2

55 0 35 90 1
65 0 75 140 2
75 0 120 195 3
80 0 165 245 4
135 1 200 335 6
140 1 245 385 7

Each one of these six plans has approximately the same AOQL, namely 1%, but

they have differing ATI's at each incoming quality. Now if p has been

running at p = .5%, then we should use the 80, 165 plan to obtain the lowest
of the six ATI's when p = .5%. But if p is say ,9% we obtain the lowest ATI
at .9% by using the 140,245 plan. Thus, knowing Nand how p has been running
we can find a plan providing protection at an AOQL of 1%, and which does so
on a minimum ATI, for this p.

We note also that in each block there is listed the lot tolerance per
I
cent defective pt, for which the consumer's risk is .10, that is, p 10 • These
run from 5.1% down to 3.2%. Therefore we obtain this degree of LTPD protection,

along with our 1% AOQL protection. So in reality we have two kinds of

protection for the consumer.

The Dodge-Romig tables necessarily involve some compromises or approximations.

It is essential that classes be used for N's and p's to avoid having infinitely
many cases to handle. Thus with only six sections (classes of p) we cannot

absolutely guarantee that we have minimized the ATI for each and every

combination of p and N leading to that block. But the ATI will be very nearly

a minimum. With whole numbers for c 1 and c 2 we also may not hit the desired
AOQL exactly. Finally, n 1+n 2 is always made divisible by 5, and the Poisson

distribution is used as an approximation to the hypergeometric, whenever feasible.

278 SOME STANDARD PLANS FOR ATTRIBUTES

We here give also as Table 10.2 an LTPD table for double sampling plans,
with LTPD = pt = 4.0%. In this table all plans have pt = 4.0% closely, and

the AOQL-'s vary, increasing in general as p and n's increase.

Further, we reproduce in Table 10.3 a single sample,average quality table,

analogous to Table 10.1.

The available LTPD's in the older tables for pt protection are .5, 1.0,

2.0, 3.0, 4.0, 5.0, 7.0 and 10.0%, while for the newer ones for average

quality protection, the AOQL's are .1, .25, .50, ,75, 1.0, 1.5, 2.0, 2.5, 3.0, 4.0,

5.0, 7.0 and 10.0%.

In the book there is provided a great collection of Type B (binomial)

OC curves. All OC curves for the AOQL plans, both single and double are given.

Then in another appendix there are single sample OC curves for c = 0, 1, 2, 3

and n < 500. These can be very useful.

The introductory chapters are also of much interest. The tables are

still widely useful.

We also mention that by replacing per cents defective by defects ~

hundred units the tables may be used for defects instead of defectives, Thus

for acceptance one would count defects in samples rather than defectives to

compare with acceptance and rejection numbers.

In summary we give the decisions and requirements needed to find a

desired Dodge-Romig sampling plan:

1. Decision on whether to use single or double sampling.

2. Decision on whether to specify AOQL or LTPD protection and the desired

percent for that type chosen. Decisions in 1 and 2 lead to the

appropriate table.

3. Decide whether to use defectives or defects.

4. Specify the lot size N.

 TABLE 10.2. A double sampling, lot quality Dodge-Romig Table.
Double Sampling Table for
Lot Tolerance Per Cent Defective (LTPD) = 4.0%

Process Average Process Average Process Average

Oto0.04% 0.05 to 0.40% 0.41 to 0.80%
Lot Size Trial1 Trial2 AOQL Trial1 Trial2 AOQL Trial1 Trial 2 AOQL
---
n, in o/o n, c, in% in%
c, n, nt+n:! C;!

-
C! n, nt+n 2

- -
n, c,
--- n, n,+nz c,

1- 35 All 0 - - 0 All 0 - 0 All 0 - - - 0

36-50 34 0 - - - 0.35 34 0 - - - 0.35 34 0 - - - 0.35
51-75 40 0 ·- - - 0.43 40 0 - - - 0 .43 40 0 - - - 0.43
76- 100 50 0 25 751 0.46 50 0 25 75 1 0.46 50 0 25 75 1 0.46
101- 150 55 0 30 851 0.55 55 0 30 85 1 0.55 55 0 30 85 1 0.55
151-200 60 0 30 90 1 0.64 60 0 30 90 1 0.64 60 0 30 90 1 0.64
201 -300 60 0 35 95 1 0.70 60 0 35 95 1 0.70 60 0 65 125 2 0.75
301-400 65 0 35 100 1 0.71 65 0 35 100 1 0.71 65 0 65 130 2 0.80
401- 500 65 0 40 105 1 0.73 65 0 70 135 2 0.83 65 0 70 135 2 0.83
501-600 65 0 40 105 1 0.74 65 0 75 140 2 0.85 65 0 100 165 3 0.93
601- 800 65 0 40 105 1 0.75 65 0 75 140 2 0.87 65 0 110 175 3 0.97
801- 1000 70 0 40 110 1 0.76 70 0 75 145 2 0.90 70 0 105 175 3 0 .98
1001- 2000 70 0 40 110 1 0.78 70 0 80 150 2 0.94 70 0 145 215 4 1.1
2001- 3000 70 0 80 150 2 0.95 70 0 115 185 3 1.1 70 0 180 250 5 1.2
3001-4000 70 0 80 150 2 0.96 70 0 115 185 3 1.1 110 1 175 285 6 1.3
4001 - 5000 70 0 80 150 2 0.97 70 0 115 185 3 1.1 115 1 170 285 6 1.3
5001- 7000 70 0 80 150 2 0.98 70 0 115 185" 3 1.1 115 1 205 320 7 1.4
7001 - 10,000 70 0 80 150 2 0.98 70 0 150 220 4 1.2 115 1 205 320 7 1.4
10,001- 20,000 70 0 80 150 2 0.98 70 0 150 220 4 1.2 115 1 235 350 8 1.5
20,001-50,000 70 0 80 150 2 0.99 70 0 150 220 4 1 .2 115 1 270 385 9 1.6
50,001- 100,000 70 0 80 150 2 0.99 70 0 185 255 5 1.3 115 1 300 415 10 1.7

Process Average Process Average Process Average

0.81 to 1.20% 1.21 to 1.60% 1.61 to 2.00%
Lot Size
Trial1 Trial2 AOQL Trial1 Trial2 AOQL Trial1 Trial2 AOQL
n, c, n'!. nt+n:? C'!.
in o/o n, Ct ll'! n 1+n2 C'!
in% n, Ct nz n,+nz c, in%

1-35 All 0 - - - 0 All 0 - - - 0 All 0 - - - 0

36- 50 34 0 - - - 0.35 34 0 - - - 0.35 34 0 - - - 0.35
51-75 40 0 - - - 0.43 40 0 - - - 0.43 40 0 - - - 0.43
76-100 50 0 25 75 1 0.46 50 0 25 75 1 0.46 50 0 25 75 1 0.46
101-150 55 0 30 85 1 0.55 55 0 30 85 1 0.55 55 0 30 85 1 0.55
151-200 60 0 55 115 2 0.68 60 0 55 115 2 0.68 60 0 55 115 2 0.68
201-300 60 0 65 125 2 0.75 60 0 90 150 3 0.84 60 0 90 150 3 0.84
301-400 65 0 95 160 3 0.86 65 0 95 160 3 0.86 65 0 120 185 4 0.92
401-500
501- 600
-65
65 0
0
100 165
135 200
3
4
0.92
1.0
65 0
105 1
130 195
145 250
4
6
0 .96
1.1
105 1
105 1
140 245
175 280 7
6 1.0
1.1
601- 800 65 0 140 205 4 1 .0 105 1 185 290 7 1.2 105 1 210 315 8 1.2
801-1000 110 1 155 265 6 1.2 110 1 210 320 8 1.2 145 2 230 375 10 1.3
1001- 2000 110 1 195 305 7 1.3 150 2 240 390 10 1.5 180 3 295 475 13 1.6
2001-3000 110 1 260 370 9 1.4 185 3 305 490 13 1.6 220 4 410 630 18 1.7
3001- 4000 150 2 255 405 10 1.5 185 3 340 525 14 1.6 285 6 465 750 22 1.8
I
4001-5000 150 2 285 435 11 1.6 185 3 395 580 16 1.7 285 6 520 805 24 1.9
5001-7000 150 2 320 470 12 1.6 185 3 435 620 17 1.7 320 7 585 905 27 2 .0
7001-10,000 150 2 325 475 12 1.7 220 4 460 680 19 1.9 320 7 645 965 29 2.1
10,001- 20,000 150 2 355 505 13 1.7 220 4 495 715 20 1.9 350 8 790 1140 35 2 .2
20,001-50,000 150 2 420 570 15 1.7 255 5 575 830 24 2.0 385 9 895 1280 40 2.3
50,001- 100,000 150 2 450 600 16 1.8 255 5 665 920 27 2.1 415 10 985 1400 44 2.4

Trial 1: n, = first sample size; c, = acceptance number for first sample

"All" indic~;~tes that each piece in the lot is to be inspected
Trial 2: n, = second sample size; c, = acceptance number for first and second samples combined
AOQL = Average Outgoing Quality Limit

Reproduced with permission from H. F. Dodge and H. G. Romig, "Sampling

Inspection Tables," Wiley, New York, 1959, p. 192.
 280 SOME STA~DARD PLANS FOR ATTRIBUTES

TABLE 10.3. Single sampling, a·verage quality Dodge-Romig tables.

Single Sampling Table for
Average Outgoing Quality Limit (AOQL) 1.0%

Process Average Process Average Process Average Process Average Process Average Process Average
Lot Size 0 to 0.02% 0.03 to 0.20 % 0.21 to 0.40% 0.41 to 0.60 % 0.61 to 0.80 % 0.81 to 1.00% .

1- 25
n

All
-
0
----
c Pt%

-
---
n

All 0 .-
-- - - - -- --- - -
c P t%
All 0
n

All 0
c Pt%

- All -
n c Pt% n c P t%

0 -
n
--
All
---
c Pt %

0 -
26-50 22 0 7 .7 22 0 7 .7 22 0 7.7 22 0 7.7 22 0 7 .7 22 0 7 .7
51-100 27 0 7.1 27 0 7.1 27 0 7 .1 27 0 7.1 27 0 7 .1 27 0 7 .1
101-200 32 0 6.4 32 0 6.4 32 0 6 .4 32 0 6.4 32 0 6.4 32 0 6.4
201-300 33 0 6.3 33 0 6 .3 33 0 6 .3 33 0 6.3 33 0 6 .3 65 1 5 .0
301-400 34 0 6.1 34 0 6.1 34 0 6.1 70 1 4.6 70 1 4.6 70 1 4 .6
401-500 35 0 6.1 35 0 6.1 35 0 6 .1 70 1 4.7 70 1 4.7 70 1 4 .7
501-600 35 0 6.1 35 0 6 .1 75 1 4.4 75 1 4.4 75 1 4 .4 75 1 4 .4
601-800 35 0 6.2 35 0 6 .2 75 1 4 .4 75 1 4.4 75 1 4 .4 120 2 4.2
801-1000 35 0 6.3 35 0 6.3 80 1 4.4 80 1 4.4 120 2 4.3 120 2 4 .3
1001-2000 36 0 6.2 80 1 4.5 80 1 4 .5 130 2 4.0 130 2 4.0 180 3 3 .7
2001- 3000 36 0 6.2 80 1 4 .6 80 1 4 .6 130 2 4 .0 185 3 3 .6 235 4 3 .3
3001- 4000 36 0 6 .2 80 1 4.7 135 2 3.9 135 2 3 .9 185 3 3 .6 295 5 3 .1
4001- 5000 36 0 6 .2 85 1 4 .6 135 2 3 .9 190 3 3.5 245 4 3 .2 300 5 3 .1
5001- 7000 37 0 6.1 85 1 4 .6 135 2 3.9 190 3 3 .5 305 5 3 .0 420 7 2.8
7001-10,000 37 0 6.2 85 1 4 .6 135 2 3.9 245 4 3.2 310 5 3.0 430 7 2 .7
10,001-20,000 85 1 4.6 135 2 3 .9 195 3 3.4 250 4 3.2 435 7 2.7 635 10 2. 4
20,001- 50,000 85 1 4.6 135 2 3 .9 255 4 3 .1 380 6 2.8 575 9 2.5 990 15 2 .1
50,001-100,000 85 1 4.6 135 2 3 .9 255 4 3.1 445 7 2.6 790 12 2 .3 1520 22 1 .9

Single Sampling Table for

Average Outgoing Quality Limit (AOQL) 1.5%

Process Average Process Average Process Average Process Average Process Average Process Average
Lot Size 0 to 0.03% 0.04 to 0.30 % 0.31 to 0.60 % 0.61 to 0.90 % 0.91 to 1.20 % 1.21 to 1.50 %

--- - - --- - ---- -- ---

c P t%
n
---
c Pt %
--
n c P t%
--
n
---
c P t% n c Pt % n n P t%

-
c

1-15 All 0 -All 0 All - - All 0

0 All 0 - All -
0
16- 50 16 0 11. 6 16 0 11 .6 16 0 11 .6 16 0 11 .6 16 0 11 .6 16 0 11.6
51- 100 20 0 9. 8 20 0 9 .8 20 0 9 .8 20 0 9.8 20 0 9 .8 20 0 9 .8
101- 200 22 0 9.5 22 0 9 .5 22 0 9. 5 22 0 9 .5 22 0 9 .5 44 1 8 .2
201- 300 23 0 9.2 23 0 9 .2 23 0 9.2 47 1 7.9 47 1 7.9 47 1 7. 9
301- 400 23 0 9.3 23 0 9 .3 49 1 7.8 49 1 7 .8 49 1 7 .8 49 1 7 .8
401- 500 23 0 9.4 23 0 9 .4 50 1 7.7 50 1 7 .7 50 1 7 .7 50 1 7.7
501- 600 24 0 9 .0 24 0 9 .0 50 1 7 .7 50 1 7.7 50 1 7 .7 50 1 7 .7
601- 800 24 0 9. 1 24 0 9 .1 50 1 7 .8 50 1 7 .8 80 2 6 .4 80 2 6 .4
801-1000 24 0 9 .1 5E 1 7 .0 55 1 7 .0 85 2 6. 2 85 2 6 .2 85 2 6 .2
1001- 2000 24 0 9. 1 55 1 7 .0 55 1 7 .0 85 2 6 .2 120 3 5 .4 155 4 5 .0
2001- 3000 24 0 9. 2 55 1 7 .1 90 2 5 .9 125 3 5 .3 160 4 4 .9 200 5 4 .6
3001- 4000 24 0 9. 2 55 1 7 .1 90 2 5 .9 125 3 5 .3 165 4 4.8 240 6 4. 4
4001- 5000 24 0 9. 2 55 1 7 .1 90 2 5.9 125 3 5. 3 205 5 4 .6 280 7 4. 2
5001-7000 24 0 9.2 55 1 7 .1 90 2 5 .9 165 4 4 .8 205 5 4 .6 325 8 4 .0
7001- 10,000 24 0 9 .2 55 1 7 .1 130 3 5.2 165 4 4.8 250 6 4. 2 3'75 9 3 .8
10,001- 20,000 55 1 7.1 90 2 5. 9 130 3 5. 2 210 5 4. 4 340 8 3 .8 515 12 3 .4
20,001- 50,000 55 1 7 .1 90 2 5 .9 170 4 4.7 295 7 4 .0 480 11 3 .5 860 19 3.0
50,001- 100,000 55 1 7 .1 130 3· 5 .2 210 5 4.4 340 8 3.8 625 14 3.3 1120 24 2. 8

n = sample size; c = acceptance number

"All" indicates t hat each piece in t he lot is t o be inspected
p 1 = lot t olerance per cent defect ive with a Consumer's R isk (Pc) of 0.10

Reproduced with permission from H. F. Dodge and H. G. Romig, "Sampling

Inspection Tables," Wiley, New York, 1959, p. 200.
10.3. The ABC Plan 281

5. Use p in percent. If P is unknown the conservative approach is to

use the section corresponding to the largest p's. For defects basis,
given -c as the average number of defects per individual piece or

unit, then c·lOO, the average defects per 100 units, corresponds to

p- in %.
6. N and p in 4 and 5 lead to the specific plan in the appropriate

table. The amount of the opposite consumer protection is shown.

7. As information on p accumulates from single samples or first

samples in double sampling, different sections of the table may be

used.

10.3. The ABC or Military Standard 105D Plan. This sampling

system is the outgrowth of the Army Service Forces and the Joint Army-Navy

tables. The present system is the third revision of the Military Standard
MIL-STD 105A. It was developed by an international committee from the United

States, Great Britain and Canada, thus the current designation "ABC." This

evolution is detailed in Part III of [2]. It was the work of many people!
We are here reproducing the introductory material, most of the summary
tables, special descriptive tables, and an example table for one of the

"code letters," We shall describe various aspects of the plans in relation

to this material.

10.3.1. Basic Aim. The ABC Standard 105D aims to insure that, on the

average, the consumer will be using material at the specified acceptable quality
level AQL, or better. It is designed so that if a producer runs his process

at the AQL exactly, the great majority of his lots will be accepted. This

probability of acceptance at the AQL runs from about .88 for small lots and

relatively tight AQL's up to about .995 for large lots and/or relatively loose
282 SOME STANDARD PLANS FOR ATTRIBUTES

AQL 1 s. [4]. Of course if his process is better than the AQL, then the Pa will
be correspondingly higher.

But if the producer runs at a worse p than the AQL, he may well have

some lots accepted, but sooner or later the sampling plan will go onto

"tightened inspection," and then a smaller proportion of lots will be accepted.

Moreover if the producer does not soon improve his process, so as to justify

a return to "normal inspection," then 100% inspection will be instituted and

stern action taken. But if he improves his process sufficiently the plan will

reinstitute "normal inspection."

Thus the basic aim of the ABC Standard is to enable the outgoing quality

to be maintained at the specified AQL or better.

10.3.2. General Description. The sampling plans are contained in nine

summary tables: Single, Table II; Double, Table III, and Multiple, Table IV.

As conditions warrant, one may specify any one of the three kinds of sampling.

See below for "multiple sampling." Then each of the three tables is divided

into three parts for "normal," "tightened" and "reduced inspection," the last being

for small sample sizes when quality is consistently excellent relative to the

AQL. Institution of tightened inspection is essential when called for,

but use of reduced inspection, when available is optional.

The AQL's are to be specified in terms of defectives per hundred units,

that is, percent defective, or in terms of defects per hundred units. The

latter may be used for all AQL's whereas the former is only used for AQL 1 s

of 10.0 or lower. See Sections 4.2 to 4.6 in the tables.

There are various "inspection levels," providing differing sample sizes

to give relatively lower or higher discriminating power to the OC curve.

The standard levels are I, II and I II, in which II is in general assumed, or

used if none is specified. Level I gives less discrimination, III greater.

 10.3. The ABC Plan 283

SAMPLING PROCEDURES AND TABLES

FOR INSPECTION BY ATTRIBUTES

1. SCOPE

1.1 PURPOSE. This publication estab- The plans may also be used for the inspection
lishes sampling plans and procedures for of isolated lots or batches, but, in this latter
inspection by attributes. When specified by case, the user is cautioned to consult the
the responsible authority, this publication operating characteristic curves to find a plan
shall be referenced in the specification, con- which will yield the desired protection (see
tract, inspection instructions, or other docu- 11.6).
ments and the provisions set forth herein
shall govern. The "responsible authority" 1.3 INSPECTION. Inspection is the proc-
shall be designated in one of the above ess of measuring, examining, testing, or
documents. otherwise comparing the unit of product (see
1.5) with the requirements.
1.2 APPLICATION. Sampling plans des-
ignated in this publication are applicable, but 1.4 INSPECTION BY ATTRIBUTES. In-
not limited, to inspection of the following: spection by attributes is inspection whereby
either the unit of product is classified simply
a. End items.
as defective or nondefective, or the number
b. Components and raw materials. of defects in the unit of product is counted,
with respect to a given requirement or set
c. Operations. of requirements.

d. Materials in process.
1.5 UNIT OF PRODUCT. The unit of
e. Supplies in storage. product is the thing inspected in order to
determine its classification as defective or
f. Maintenance operations. nondefective or to count the number of de-
fects. It may be a single article, a pair, a set,
g. Data or records. a length, an area, an operation, a volume, a
component of an end product, or the end
h. Administrative procedures.
product itself. The unit of product may or
These plans are intended primarily to be may not be the same as the unit of purchase.
used for a continuing series of lots or batches. supply, production, or shipment.

*This page and those through page 308 are reproduced by permission from the
Department of Defense from Military Standard MIL-STD lOSD, Sampling Proce-
dures and Tables for Inspection by Attributes, 1963.
 284 SOME STANDARD PLANS FOR ATTRIBUTES

2. CLASSIFICATION OF DEFECTS AND DEFECTIVES

2.1 METHOD OF CLASSIFYING DEFECTS. 2.1.3 MINOR DEFECT. A minor defect
A classification of defects is the enumeration is a defect that is not likely to reduce ma-
of possible defects of the unit of product terially the usability of the unit of product
classified according to their seriousness. A for its intended purpose, or is a departure
defect is any nonconformance of the unit of from established standards having little bear-
product with specified requirements. Defects ing on the effective use or operation of the
will normally be grouped into one or more unit.
of the following classes; however, defects
may be grouped into other classes, or into 2.2 METHOD OF CLASSIFYING DEFEC·
subclasses within these classes. TIVES. A defective is a unit of product which
contains one or more defects. Defectives will
2.1.1 CRITICAL DEFECT. A critical de- usually be classified as follows:
fect is a defect that judgment and experience
indicate is likely to result in hazardous or 2.2.1 CRITICAL DEFECTIVE. A critical
unsafe conditions f o r individuals using, defective contains one or more critical de-
maintaining, or depending upon the product; fects and may also contain major and or
or a defect that judgment and experience minor defects. NOTE: For a special provi-
indicate is likely to prevent performance of sion relating to critical defectives, see 6.3.
the tactical function of a major end item such
as a ship, aircraft, tank, missile or space
vehicle. NOTE: For a special provision re- 2.2.2 MAJOR DEFECTIVE. A major de-
lating to critical defects, see 6.3. fective contains one or more major defects,
and may also contain minor defects but con-
2.1.2 MAJOR DEFECT. A major defect tains no critical defect.
is a defect, other than critical, that is likely
to result in failure, or to reduce materially 2.2.3 MINOR DEFECTIVE. A minor de-
the usability of the unit of product for its fective contains one or more minor defects
intended purpose. but contains no critical or major defect.

3. PERCENT DEFECTIVE AND DEFECTS PER HUNDRED UNITS

3.1 EXPRESSION OF NONCONFORM· Percent defective
Number of defectives
= Number of units inspected X 100
ANCE. The extent of nonconformance of
product shall be expressed either in terms 3.3 DEFECTS PER HUNDRED UNITS. The
of percent defective or in terms of defects per number of defects per hundred units of Any
hundred units. given quantity of units of product is one
3.2 PERCENT DEFECTIVE. The percent hundred times the number of defects con-
defective of any given quantity of units of tained therein (one or more defects being
product is one hunderd times the number of possible in any unit of product) divided by
defective units of product contained therein the total number of units of product, i.e.:
divided by the total number of units of prod- Defects per Number of defects X 100
uct, i.e.: hundred units = Number of units inspected
 10.3. The ABC Plan 285

4. ACCEPTABLE QUALITY LEVEL {AQL)

4.1 USE. The AQL, together with the describe the protection to the consumer for
Sample Size Code Letter, is used for index- individual lots or batches but more directly
ing the sampling plans provided herein. relates to what might be expected from a
series of lots or batches, provided the steps
4.2 DEFINITION. The AQL is the max- indicated in this publication are taken. It is
imum percent defective (or the maximum necessary to refer to the operating character-
number of defects per hundred units) that, istic curve of the plan, to determine what
for purposes of sampling inspection, can be protection the consumer will have.
considered satisfactory as a process average
(see 11.2). 4.4 LIMITATION. The designation of an
AQL shall not imply that the supplier has
4.3 NOTE ON THE MEANING OF AQL. the right to supply knowingly any defective
When a consumer designates some specific unit of product.
value of AQL for a certain defect or group
of defects, he indicates to the supplier that 4.5 SPECIFYING AQLs. The AQL to be
his (the consumer's) acceptance sampling used will be designated in the contract or by
plan will accept the great majority of the lots the responsible authority. Different AQLs
or batches that the supplier submits, pro- may be designated for groups of defects con-
vided the process average level of percent sidered collectively, or for individual defects.
defective (or defects per hundred units) in An AQL for a group of defects may be des-
these lots or batches be no greater than the ignated in addition to AQLs for individual
designated value of AQL. Thus, the AQL defects, or subgroups, within that group.
is a designated value of percent defective (or AQL values of 10.0 or less may be expressed
defects per hundred units) that the consumer either in percent defective or in defects per
indicates will be accepted most of the time hundred units; those over 10.0 shall be ex-
by the acceptance sampling procedure to be pressed in defects per hundred units only.
used. The sampling plans provided herein
are so arranged that the probability of ac- 4.6 PREFERRED AQLs. The values of
ceptance at the designated AQL value de- AQLs given in these tables are known as
pends upon the sample size, being generally preferred AQLs. If, for any product, an AQL
higher for large samples than for small ones, be designated other than a preferred AQL,
for a given AQL. The AQL alone does not these tables are not applicable.

S. SUBMISSION OF PRODUCT

5.1 LOT OR BATCH. The term lot or for other purposes (e.g., production, ship-
batch shall mean "inspection lot" or "inspec- ment, etc.).
tion batch," i.e., a collection of units of prod- 5.2 FORMATION OF LOTS OR BATCHES.
uct from which a sample is to be drawn and The product shall be assembled into identi-
inspected to determine conformance with the fiable lots, sublots, batches, or in such other
acceptability criteria, and may differ from a manner as may be prescribed (see 5.4). Each
collection of units designated as a lot or batch lot or batch shall, as far as is practicable,
 286 SOME STANDARD PLANS FOR ATTRIBUTES

5. SUBMISSION OF PRODUCT (Continued)

consist of units of product of a single type, batches, lot or batch size, and the manner
grade, class, size, and composition, manu- in which each lot or batch is to be presented
factured under essentially the same condi- and identified by the supplier shall be des-
tions, and at essentially the same time. ignated or approved by the responsible au-
thority. As necessary, the supplier shall
5.3 LOT OR BATCH SIZE. The lot or
provide adequate and suitable storage space
batch size is the number of units of product
for each lot or batch, equipment needed for
in a lot or batch.
proper identification and presentation, and
5.4 P R ESE N TAT I 0 N OF LOTS OR personnel for all handling of product re-
BATCHES. The formation of the lots or quired for drawing of samples.

6. ACCEPTANCE AND REJECTION

6.1 A C C EP TA B IL I T Y OF LOTS OR critical defects. The right is reserved to in-

BATCHES. Acceptability of a lot or batch spect every unit submitted by the supplier for
will be determined by the use of a sampling critical defects, and to reject the lot or batch
plan or plans associated with the desi~nated immediately, when a critical defect is found.
AQL or AQLs. The right is reserved also to sample, for crit-
ical defects, every lot or batch submitted by
6.2 DEFECTIVE UNITS. The right is re- the supplier and to reject any lot or batch
served to reject any unit of product found if a sample drawn therefrom is found to con-
defective during inspection whether that tain one or more critical defects.
unit of product forms part of a sample or
not, and whether the lot or batch as a whole 6.4 RESUBMinED LOTS OR BATCHES.
is accepted or rejected. Rejected units may Lots or batches found unacceptable shall be
be repaired or corrected and resubmitted for resubmitted for reinspection only after all
inspection with the approval of, and in the units are re-examined or retested and all de-
manner specified by, the responsible au- fective units are removed or defects cor-
thority. rected. The responsible authority shall deter-
mine whether normal or tightened inspection
6.3 SPECIAL RESERVATION FOR CRITI- shall be used, and whether reinspection shall
CAL DEFECTS. The supplier may be required include all types or classes of defects or for
at the discretion of the responsible authority the particular types or classes of defects
to inspect every unit of the lot or batch for which caused initial rejection.

7. DRAWING OF SAMPLES
7.1 SAMPLE. A sample consists of one 7.2 REPRESENTATIVE SAMPLING. When
or more units of product drawn from a lot or appropriate, the number of units in the sam-
batch, th~ units of the sample being selected
ple shall be selected in proportion to the size
at random without regard to their quality.
The number of units of product in the sample of sublots or subbatches, or parts of the lot or
is the sample size. batch, identified by some rational criterion.
 10.3. The ABC Plan 287

7. DRAWING OF SAMPLES (Continued)

When representative sampling is used, the pies may be drawn during assembly of the
units from each part of the lot or batch shall lot or batch.
be selected at random.
7.4 DOUBLE OR MULTIPLE SAMPLING.
7.3 TIME OF SAMPLING. Samples may When double or multiple sampling is to be
be drawn after all the units comprising the used, each sample shall be selected over the
lot or batch have been assembled, or sam- entire lot or batch.

8. NORMAL, TIGHTENED AND REDUCED INSPECTION

8.1 INITIATION OF INSPECTION. Nor- a. The preceding 10 lots or batches (or
mal inspection will be used at the start of more, as indicated by the note to Table VIII)
inspection unless otherwise directed by the have been on normal inspection and none
has been rejected on original inspection; and
responsible authority.
b. The total number of defE>ctives (or de-
8.2 CONTINUATION OF INSPECTION. fects) in the samples from the preceding 10
Normal, tightened or reduced inspection lots or batches (or such other number as was
used for condition "a" above) is equal to or
shall continue unchanged for each class of less than the applicable number given in
defects or defectives on successive lots or Table VIII. If double or multiple sampling
batchs except where the switching proce- is in use, all samples inspected should be in-
dures given below require change. The cluded, not ''first" samples only; and
switching procedures given below require a
c. Production is at a steady rate; and
change. The switching procedures shall be
applied to each class of defects or defectives d. Reduced inspection is considered de-
independently. sirable by the responsible authority.
8.3.4 REDUCED TO NORMAL. When re-
8.3 SWITCHING PROCEDURES.
duced inspection is in effect, normal inspec-
8.3.1 NORMAL TO TIGHTENED. When tion shall be instituted if any of the following
occur on original inspection:
normal inspection is in effect, tightened in-
spection shall be instituted when 2 out of 5 a. A lot or batch is rejected; or
consecutive lots or batches have been re- b. A lot or batch is considered acceptable
jected on original inspection (i.e., ignoring undE>r the procedures of 10.1.4; or
resubmitted lots or batches for this proce- c. Production becomes irregular or de-
dure). layed; or
8.3.2 TIGHTENED TO NORMAL When d Other conditions warrant that normal
tightened inspection is in effect, normal in- inspection shall be instituted.
spection shall be instituted when 5 consecu- 8.4 DISCONTINUATION OF INSPECTION.
tive lots or batches have been considered In tHe event that 10 consecutive lots or
acceptable on original im.pection. batches remain on tightened inspection (or
such other humber as may be designated by
8.3.3 NORMAL TO R-EDUCED. W h en the responsible authority), inspection under
normal inspection is in effect, reduced inspec- the provisions of this document should be
tion shall be instituted providing that all of discontinued pending action to improve the
the following conditions are satisfied: quality of submitted m:~terial.
 288 SOME STANDARD PLANS FOR ATTRIBUTES

9. SAMPLING PLANS

9.1 SAMPLING PLAN. A sampling plan tain the sampling plan from Tables II, III or
indicates the number of units of product IV. When no sampling plan is available for a
from each lot or batch which are to be in- given combination of AQL and code letter,
spected (sample size or series of sample the tables direct the user to a different letter.
sizes) and the criteria for determining the The sample size to be used is given by the
acceptability of the lot or batch (acceptance new code letter not by the original letter. If
and rejection numbers). this procedure leads to different sample sizes
for different classes of defects, the code letter
9.2 INSPECTION LEVEL. The inspection corresponding to the largest sample size de-
level determines the relationship between rived may be used for all classes of defects
the lot or batch size and the sample size. The when designated or approved by the respon-
inspection level to be used for any particular sible authority. As an alternative to a single
requirement will be prescribed by the re- sampling plan with an acceptance number
sponsible authority. Three inspection levels: of 0, the plan with an acceptance number of 1
I, II, and III, are given in Table I for general with its correspondingly larger sample size
use. Unless otherwise specified, Inspection for a designated AQL (where available), may
Level II will be used. However, Inspection be used when designated or approved by the
Level I may be specified when less discrimi- responsible authority.
nation is needed, or Level III may be speci-
fied for greater discrimination. Four addi- 9.5 TYPES OF SAMPLING PLANS. Three
tional special levels: s-1, s-2, s-a and S-4, types of sampling plans: Single, Double and
are given in the same table and may be used
Multiple, are given in Tables II, III and IV,
where relatively small sample sizes are neces-
respectively. When several types of plans are
sary and large sampling risks can or must be
available for a given AQL and code letter,
tolerated.
any one may be used. A decision as to type
NOTE: In the designation of inspection of plan, either single, double, or multiple,
levels s-i to S-4, care must be exercised to when available for a given AQL and code
avoid AQLs inconsistent with these inspec- letter, will usually be based upon the com-
tion levels. parison between the administrative difficulty
and the average sample sizes of the available
9.3 CODE LETTERS. Sample sizes are plans. The average sample size of multiple
designated by code letters. Table I shall be plans is less than for double (except in the
used to find the applicable code letter for the case corresponding to single acceptance num-
particular lot or batch size and the prescribed ber 1) and both of these are always less than
inspection level. a single sample size. Usually the administra-
tive difficulty for single sampling and the
9.4 OBTAINING SAMPLING PLAN. The cost per unit of the sample are less than for
AQL and the code letter shall be used to ob- double or multiple.
 10.3. The ABC Plan 289

10. DETERMINATION OF ACCEPTABILITY

10.1 PERCENT DEFECTIVE INSPECTION. number of defectives found in the first and
To determine acceptability of a lot or batch second samples shall be accumulated. If the
under percent defective inspection, the ap- cumulative number of defectives is equal to
plicable sampling plan shall be used in or less than the second acceptance number,
accordance with 10.1.1, 10.1.2, 10.1.3, 10.1.4, the lot or batch shall be considered accept-
and 10.1.5. able. If the cumulative number of defectives
is equal to or greater than the second rejec-
10.1.1 SINGLE SAMPLING PLAN. The tion number, the lot or batch shall be rejected.
number of sample units inspected shall be
equal to the sample size given by the plan. 10.1 .3 MULTIPLE SAMPLE PLAN. Under
If the number of defectives found in the multiple sampling, the procedure shall be
sample is equal to or less than the acceptance similar to that specified in 10.1.2, except that
number, the lot or batch shall be considered the number of successive samples required
acceptable. If the number of defectives is to reach a decision may be more than two.
equal to or greater than the rejection num-
10. 1.4 SPECIAL PROCEDURE FOR RE-
ber, the lot or batch shall be rejected.
DUCED INSPECTION. Under reduced in-
spection, the sampling procedure may termi-
10.1.2 DOUBLE SAMPLING PLAN. The nate without either acceptance or rejection
number of sample units inspected shall be criteria having been met. In these circum-
equal to the first sample size given by the stances, the lot or batch will be considered
plan. If the number of defectives found in acceptable, but normal inspection will be
the first sample is equal to or less than the reinstated starting with the next lot or
first acceptance number, the lot or batch batch (see 8.3.4 (b)).
shall be considered acceptable. If the num-
ber of defectives found in the first sample is 10.2 DEFECTS PER HUNDRED UNITS IN-
equal to or greater than the first rejection SPECTION. To determine the acceptability
number, the lot or batch shall be rejected. of a lot or batch under Defects per Hundred
If the number of defectives found in the first Units inspection, the procedure specified for
sample is between the first acceptance and Percent Defective inspection above shall be
rejection numbers, a second sample of the used, except that the word "defects" shall be
size given by the plan shall be inspected. The substituted for "defectives."

11. SUPPLEMENTARY INFORMATION

11.1 0 P E R A T I N G CHARACTERISnC and multiple sampling are matched as closely

CURVES. The operating characteristic curves as practicable. The 0. C. curves shown for
for normal inspection, shown in Table X AQLs greater than 10.0 are based on the
(pages 30-62), indicate the percentage of Poisson distribution and are applicable for
lots or batches which may be expected to be defects per hundred units inspection; those
accepted under the various sampling plans for AQLs of 10.0 or less and sample sizes of
for a given process quality. The curves shown 80 or less are based on the binomial distri-
are for single sampling; curves for double bution and are applicable for percent defec-
 290 SOME STANDARD PLANS FOR ATTRIBUTES

11. SUPPLEMENTARY INFORMATION (Continued)

tive inspection; those for AQLs of 10.0 or 11.5 AVERAGE SAMPLE SIZE CURVES.
less and sample sizes larger then 80 are based Average sample size curves for double and
on the Poisson distribution and are applica- multiple sampling are in Table IX. These
ble either for defects per hund1·ed units in- show the average sample sizes which may be
spection, or for percent defective inspection expected to occur under the various sampling
(the Poisson distribution being an adequate plans for a given process quality. The curves
approximation to the binomial distribution assume no curtailment of inspection and are
under these conditions). Tabulated values, approximate to the extent that they are
corresponding to selected values of probabil- based upon the Poisson distribution, and that
ities of acceptance (Pa• in percent) are given the sample sizes for double and multiple
for each of the curves shown, and, in addi- sampling art! assumed to be 0.631n and 0.25n
tion, for tightened inspection, and for defects respectively, where n is the equivalent single
per hundred units for AQLs of 10.0 or less sample size.
and sample sizes of 80 or less.
11.6 LIMITING QUALITY PROTECTION.
11.2 PROCESS AVERAGE. The process The sampling plans and associated proce-
average is the average percent defective or dures given in this publication were designed
average number of defects per hundred units for use where the units of product are pro-
(whichever is applicable) of product sub-
duced in a continuing series of lots or batches
mitted by the supplier for original inspec-
over a period of time. However, if the lot
tion. Original inspection is the first inspec-
or batch is of an isolated nature, it is desira-
tion of a particular quantity of product as
ble to limit the selection of sampling plans
distinguished from the inspection of product
to those, associated with a designated AQL
which has been resubmitted after prior
rejection. value, that provide not less than a specified
limiting quality protection. Sampling plans
11.3 AVERAGE OUTGOING QUALITY for this purpose can be selected by choosing
(AOQ). The AOQ is the average quality of a Limiting Quality (LQ) and a consumer's
outgoing product including all accepted lots risk to be associated with it. Tables VI and
or batches, plus all rejected lots or batches VII give values of LQ for the commonly used
after the rejected lots or batches have been consumer's risks of 10 percent and 5 percent
effectively 100 percent inspected and all de- respectively. If a different value of con-
fectives replaced by nondefectives. sumer's risk is required, the O.C. curves and
their tabulated values may be used. The
11.4 AVERAGE OUTGOING QUALITY
concept of LQ may also be useful in specify-
LIMIT (AOQL). The AOQL is the maximum
ing the AQL and Inspection Levels for a
of the AOQs for all possible incoming quali-
ties for a given acceptance sampling plan. series of lots or batches, thus fixing minimum
AOQL values are given in Table V-A for sample size where there is some reason for
each of the single sampling plans for normal avoiding (with more than a given consumer's
inspection and in Table V-B for each of the risk) more than a limiting proportion of de-
single sampling plans for tightened inspec- fectives (or defects) in any single lot or
tion. batch.
 ......
0

TABLE l - Sample size code letters 0-l

-3
::>"
(!)

)>
Special inspection levels General inspection levels
n'""
Lot or batch size
""......§
S-1 S-2 S-3 S-4 I II Ill

2 to 8 A A A A A A B
9 to 15 A A A A A B c
16 to 25 A A B B B c 0

26 to 50 A B B c c 0 E
51 to 90 B B c c c E F
91 to 150 B B c 0 0 F G

151 to 280 B c 0 E E G H
281 to 500 B c 0 E F H J
501 to 1200 c c E F G J K

1201 to 3200 c 0 E G H K L
3:201 to 10000 c 0 F G J L M
10001 to 35000 c 0 F H K M N

35001 to 150000 0 E G J L N p
150001 to 500000 0 E G J M p Q
500001 and over 0 E H K N Q
N
~-
---------·-··- - ------ --- --- ----- - -------
-R~j tD
......
 ....,
<0
....,
TABLE II-A -Single sampling plans for normal inspection (Master table)

Acceptable Quality levels (normal inspection)

Sample
size S.a.ple 1000
0.010 I 0.015 I 0.025 I O.IHO I 0.065 I 0.10 O.IS 0.25 0.40 0.65 1.0 l.S 2.S 4.0 6.S 10 15 25 40 6S 100 ISO 250 400 650
code size
,..... Ac ReiA.c ReiAc ReiA.c ReiAe ReiAc ReiAc ReiAc ReiAc ReiAc ReiAc ReiAc ReiAc ReiAc ReiAc ReiAc ReiAc ReiAc ReiAc ReiAc Rei.A.c ReiAc ReiAc ReiAc ReiAc ReiAc Re

s 6 7 8 10 II 14 IS 21 22130 31
A 2 01..0. 122334S6 7 8 10 II 14 IS 21 22 30 31 « 45
B 3
c
I 5
R.R RlL ,[]. ~ '41- .[7.
oi.O~t22334S678 10 II 14 IS 21 22 30 31
' ' ' ' ' •
« 45

0 !...:;:>~ 1 2 2 3 3 4 s 6 7 81011 14 IS 21 22 30 31 « 45 ..:(;>..

D 8 , ot.O~t22334S67810llt4tS 21 22 30 31 45 ~
E 13 «
F
I 20 , 0 I ~V 1 212 3 3 4 S 6 7 8 10 II 14 IS 21 22
~ ~
G 32 0
-<::7 1 1 2 2 3 3 ' s 6 7 8 10 ll,l. ISI21 22
H 50 0 <>
I..(? ...(?'li"'l:r ""!:?- I 2 2 3 3 4 56 7 8 10 ll IS 21 22
14 ~
J 80 ~~~ 0 I ""'i:J'-iv 1 2 2 3 3 4 5 6 7 8 10 II 14 IS 21 22 ~
Cfl
K 125 1 2 2 31 3 <I s 61 7 0
~ 0 1 "'i?- ~ BI!O llll< !SI21 22~ :;;::
L 200 0 I 2 2 3 3 • 5 6 7 8 10 II 14 !SI21 22 ,.. m
iv "'i?- ~I
~ 315 I I 2 2 3 5 6 7 8 10 II 14 IS 2122~1 Cfl
I~ o ...:;:>. ~ 3 '

N 500 v 0 1 1 2 2 3 3 ' 5 6 7 8 10 ll 14 15 21 22 ~
...:;:::;:. .,0.. 0
p 800 I 2 4 5 6 7 8 10 II 14 ~ 21 22 );>
-0>- ~ 2 3 3 ;.:>
Q 1250 -.(7 1 2 2 3 s 6 7 8 10 II 14 IS 21 22 0
3 '
l-- "0
§;::
R 2000 1 21 2 31 3 41 5 61 7 alto 111 14 tSI 21 z
Cfl
"'1
0
;:tr zzju ~ i~ ;.:>

~ Use fint u.pliBB plu. below mvw. I( saMple size eqaals, or ezceeds, lot or batcb size. do 100 percent inspection.
~
0 -l
;.:>
~ • Use first salllpli111 plan abowe art'Q'If. 1-<
Ac = Acceplaoce ....... o:l
c:
Re = Rejectioa a...ber. -l
m
Cfl
 ......
0
TABLE li-B -Single sampling plans for tightened inspection (Master table) VJ

-3
~
(l)

Acceptable Quality Levels bisbteaed iaapectioe) ;I>

Sample
size Sample
""n
...od. 0.010 0.015 0.025 0.040 0.065 0.10 0.15 0.25 0..0 0.£0 1.0 1.5 2.5 4.0 6.5 10 15 25 40 (0 100 150 250 400 650 1000 '1::1
size
letter ......
ll'
;:J
Ac Re Ac Re A< R Ac Re A.c Re Ac Re Ac Re A< R Ac Re Ac Re Ac Re Ac Re Ac Re Ac Re Ac Re Ac Re Ac Re Ac Re Ac Re A' Re Ac Re Ac Re Ae Re Ac Re A< Re Ac Re

A 2 .0 ..!> 1 2 2 3 3 5 6 a 9 12 13 18 19 <128
B 3
..[l 0 1 ..[l 1 2 2 3 3 4 5
•
6 8 9 12 13 18 19 27 28 41 42
c 5 0 1 1 2 2 3 3 5 6 8 9 12 13 18 19 27'28 41 42
R R • I"" i'"

D
E
a
13
j
] gR 0 1
0 1
1 2
1
2
2 2
3 3
3
5
3 4
6
5
a
6 a 9 12 13 18 19 27 28 41 42
9 12 13 1a 19 27 28 41 42 . ,..
~

F 20 0 1 ~ 1 2 2 3 3 • 5
•
6 a 9 12 13 18 19 I~ ,.. , .
I"' r-
G 32 1.. 0 1 1 2 2 3 3 • 5 6 a 9 12 13 18 19
H
J
50
80 0 1
0 1
!n
J]r ~ 1 2
1
2 3
2 2
3
3
• 5
3 • 5
6 8
6 a
9 12 13 1a 19
9 12 13 18 19
j.<
,..
K 125 I"' 7 0 1 1 2 2 3 3 • 5 6 8 9 12 13 1a 19
n I"'
L 200 0 1 ~ 1 2 2 3 3 • 5 6 8 Q 12 13 18 19
M 315 0 1 1 2 2 3 3 4 5 6 8 9 12 13 1a 19

N 500 I"' 7 0 1
g JJ 1 2 2 3 3 • 5 6 8 9 12 13 18 19
p 800 7 0 1 ~ 1 2 2 3 3 • 5 6 8 9 12 13 18 19
Q 1250 0 1 1 2 2 3 3 • 5 6 8 9 12 13 18 19
"' ~ ~
R 2000 0 1 i]'" n 1 2 2 3 3 • 5 6 8 9 12 13 18 19 "'ii 1J 1r i~
s 3150 1 2
jI
-- -- - L ____ - -- -- '------ -

...().. = Use first sampling plan below arrow. If sample size eqllels or es.ceeds lot or batch size, do 100 percent inspection.
~ = Use first sampling plan above arrow.
Ac = Acceptance number.
Re = Rejection number.
N
'-D
VJ
 N
<.0
~
TABLE 11-C- Single sampling plam for reduced inspection (Master table)

Accepcable Qaality Le.ela ~- i..,..,t;..)t

Saqtle
size Sample
0.010 0.015 0.025 0.040 0.065 0.10 D.l5 0.25 D.40 0.65 1.0 1.5 2.5 4.0 6.5 10 15 25 40 65 100 150 250 400 650 1000
code size
leiter

Ac Re Ac Re Ac Re A.c Re Ac Re Ac Re AcRe AcRe AcRe Ac Re AcRe AcRe AcRe Ac lie Ac Ro AeRo AcRe AcRe AcRe AcRe AcRe AcRe AcRe AcRe AcRe AcRe
A 2 0 1 1 2 2 3 3 4 5 6 1 8 10 11 14 15 2122 30 31
..!J.- ..0.
8 2 ~ 0 1 0 2 1 3 2 4 3 5 5 6 1 8 10 11 1~ 15 2122 30 31
c 2 ~ 0 1 ~ zn 0 2 1 3 I 4 2 5 3 6 5 8 1 10 10 13 14 17 2124

D 3 R 0 1 ~ ~ 0 2 1 3 1 4 2 .s 3 6 5 • 1 10 10 13 14 17 21 24
E 5 0 1 ir ..!J.- 0 2 1 3 1 4 2 5 3 6 5 8 7 10 10 13 14 IT 2124
f 8 0 I -.!). 0 2 1 3 1 4 2 5 3 6 s 8 1 10 10 13 I~
~ ~ 'fr
G 13 I"' ~ o I if .JJ. 0 2 I 3 1 4 2 5 3 6 s 8 7 10 10 13 I• ~
H 20 ~ 0
•-"lf- ...[):.- 0 2 1 3 1 4 2 5 3 6 5 8 7 10 10 13 1 . . . . I
J 32 ~ 0 •'"lJ'- -& 0 2 1 3 1 4 2 s 3 6 5 8 7 10 10 13
Cll
K 50 if -.!}. 0 2 I 3 1 4 2 5 3 6 5 8 7 10 10 13 0
'"' " 0 1 :;,::
L 80 I"' ~ o 1"1)'- -.(7 0 2 1 3 1 4 2 s 3 6 5 8 1 10 10 13 . [T:1

M 125 0 0 2 1 3 I 4 2 5 3 6 5 8 7 10 10 1 Cll

N 200
...
""
0 1 if
•'"lJ'- .JJ.
..[!.- 0 2 1 3 I 4 2 5 3 6 5 8 7 10 10 13
~ I
-l
s;
p liS 1.... " 0 ..[!:. 0 2 I 3 I 4 2 5 3 6 5 8 7 10 10 13 0
•if )>
5. 3 5 8 7 10 10 13 :;t:l
Q 500 0 I ..!J.- 0 2 1 3 I 4 2 6
0
-o
R BOO IJ 0 2 I 3 I 4 2 5 3 6 5 8 7 10 10 13 11 s;:
z
v 1f 1f vi I
I
Cll
"r1
- I 0
:;t:l

Use rust ....,.i.., plu below UIOW. If SU~ple size equals or esceeds lot or bMCia size. do 100 pen:eal iaspectioa.
~ Use first sampliag plaa above IUIOW.
:;t:l
Ac
Re
Acceptuce aUIIIber.
Rejectioo a....bet-.
s
,_,
t:X:I
t u the acceptace • ..-lou- eaceeded. bot the ..;ectioo •.-lou ... bees taehed. - dtelao, ............. _ . .............. ( - 10.1.41.
~
[T:1
Cll
 TABLE Ill-A-Double sampling plans for normal inspection (Master table)
....
0
Vl

Ac-ceptable Qulily Levels (nonn.t iaspectiotd

Saoople c.--
sia .....I. l•li~
;i
Sample .;~ 0.010 0.015 0.025 0.040 0.065 0.10 0.15 0.25 0.<0 0.65 1.0 1.5 2.5 4.0 6.5 10 15 25 6S 100 150 2SO 400 650 1000
code ....,.le
.. m
lf'llft' . ~

4.c: Ue Ae lte Ac He Ac He Ac . . . . Re He .-c lte Ac He A.c lle Ae Re Ae He Ac Re ""lie Ac lie "" Re A.c ne Ae Re AcRe AcRe eAcReAc Re Ae He Ae IW >
"" ... 0::0
"" C"l
A .
..0. .,JJ.. . . . . """
II t"irst 2 2 0 2 0 3 I 2 5 3 7 5 9 7 11 11 16 17 22 25 Jl
""....
Second l 1 2 3 5 6 7 8 9 12 13 18 19 26 27 37 38 56 57
§
• R R 1J •••
c first; 3 3 0 2 0 3 1 2 5 3 7 5 9 71 1 11 16 11 22 25 3
3 6 i.i ..[!. 1 2 3 1- 5 6 7 8 • 12 13 18 1926 2737 38 56 5 :.
0 2 0 3 1 •
••
II First 5 5 2 5 3 7 s 9 7 11 1116 17 22 2S 31

I~
-·
~ond

Firsl
5
8
10
8
n iJ ..[!...
0 2
I
0
2
3
3
1
• • s
2 s
6
3
7 8 9 12 13 18 19 26 27 37
7 s 9 711 11 16 17 22 25
38 56 57
31 ,..
... "[J ..[].. 1 2 3

F
Sec~d

to"irsl
.. ~
8

13
13
16

13
26
"'
.
i 0 2 0 3 1
• •
ll ..[].. 1 2 3 • • 5• 62 57 83 7
• s 6 7 8

s
9 12 13 18 19 26 27 37 38 56

7 11 11 16
•
9 12 13 18 19 2627
,..
57

G t~irsl 20 20
. 0 2 0 3 i 7 11 11 16
Second 20 .. i.i ..[!... 1 2 3 • • •5 62 75 83 97 12s 139 18 19 2627
0 2 0 3 1 2 5 3 7 s 9 7 11 11 16 ,.
"' .
H First 32 32 •
Second 32 .. if" .JJ,. 1 2 3 • • s 6 7 9 12 I3 18 19 26 27
•
J First so so . 0 2 0 3 1 2 s 3 7 s 9 1 11 11 16
Secood 50 100 1J .[!.. 1 2 3 • •
•
5 6 7 8 9 12 13 18 19 26 27

K t"irst 80 so 0 2 0 3 1 2 5 3 7 5 9 7 11 II 16 ,..
Second 80 160 LJ ..[].. 1 2 3• •
•
5 6 7 8 9 12 13 18 19 26 27

L fo"int 125 125 . 0 2 0 3 1 2 5 3 7 5 9 7 11 11 16

1 ...
Secood 125 2SO >- "iJ ..[].. 1 2 3 •
•• 5 6 7 8 9 12 13 18 19 26 27
First 200 200 0 2 0 3 1 • 2 5 3 7 5 9 711 1! 16 ,..
" s.c-1 200 400 I• . iJ..[].. 1 2 3 • s 6 7 8 9 12 13 18 19 2627
•
First 315 315 0 2 0 3 1 • 2 5 3 7 5 9 711 II 16
' ~cond 315 63<1 1 2 3 • • 5 6 7 8 9 12 13
• >-
. "[J.JJ,. 18 19 2627
r 1-"irst ;oo 500 0 2 0 J 1 • 2 s 3 7 5 9 7 11 u 16
Secuood 500 lOCO I"' I 2 3 s 6 7 8 9 12 13 18 19 26 27
. ll .[!..
••
<} First 800 800 0 2 0 3 I 2 s 3 7 5 9 1 II 11 ltl
~ 800 1600 I 2 3 5 6 7 8 9 12 13 18 19 26 27
•••
II Fim: 1250 1250 0 3 I 2 s 3 7 s 9 711 ll 16
Second 1250 2500 I 2 7 9 12 13 j
i.i 1J:D: 3
••• 5 6 8 18 19 26V iJ 1f i~ -- ---
~ J

....(!... a Use first umplins plan bot low arrow. lr sample sb:e ~·Is ar eace~s lo1 nr Latclo si:~:e-. Jo 100 percent inspeclion.
N
~ • Use finu s.-nrli111 plat~ ebove •no• 1.0
Ac • Ac«rt•nce ~tull'lber U1
11e - ne,ecuon numbet-
u. correspomfiar si"Bie u•pli111 plan (ot •heJDallvi!'J,, usr doul.lle u•npling pia"' l>elow, .. be~ •~•il~le).
 TABLE lll-B-Double sampling plans for tightened inspe&tion (Master table)
N
<.0

An-rpi.W,~ l}llalil)" l.o....tOo hif:ha-t ID"'II"Ciioft) "'

- c-
= ~ s::-- U.OlO tllll3 tltl.:"• tt1lkl t't"'-; till! lll!'"o tl..'!l U.W t1r6 I.U 1.5 ;!.5 40 65 IG 15 25 40 65 100 150 2"'ill iOO 650 1000
lit.., siu
= -\e !lot \c II \o.· lio· \, 1{.• \.· ~;,. \o llo· \o J;,. \o llo· \o· llo· Ao 11•· A.o llo· 1\o llo· tV 11.. A•· It Ac lie Ac lie 1\c liP Ac fl., 4c H -\c H,. 4c H~ 4.c 11,. \r: k \c H.. ,,. II \c It

8 F"n'St 2: 2 r-:-
i!Jl~!R~ 0 2 0 3 I 4 2 5 3 7 1i 10 <, I .t I) 1.•! .!J 29
SKo-d 2 4 I 2 3 4 4 5 6 7 ll 12 15 Hi .!1 .!J ;J li i.! .iJ
,
C Fiat;
~
3
3
.J
6
! 0
1
2
2
0
3
3
4
I
4
4
5
2
6
.. . . .
.i 3 7 6 10 ~ li ll 2() ~J
7 11 12 13 16 2l ~' ;.t H >2
2~
ll
0 Fi.- 5 S 0 2 0 3 l 4 2 .i 3 7 6 10 9 U 15 20 .!3 ::,9
Seco.d 5 10 1 2 3 4 4 5 6 7 ll 12 IS 16 2J .!.J U 15 5.:: 53

f. Fifst
s.ec..t
8
8
8
16
n o
1
2
2
0
3
3
4
1
4
4
5
2
6
5 3 7 6 10 9 U 15 20 .!3 J.~t
7 11 12 15 16 13 .Z; 3J 35 5.? SJ
?-

•. Fi.s~ ll ll 0203142537bl0914 ?'-

__. 13 26 12344567111215162324

.,; Finl 20 20 •
!! 0 2 0 3 1 4 2 5 3 7 6 10 9 14
...... 20.. ...> 12344567111215162324

II Finl 32 l2
ll 0 2 0 3 I 4 2 5 3 7 6 10 9 141 ~ ?"-
Seeaad 32 " 1 2 3 4 4 5 6 1 u 12 ts16 21 2c

J First SO SO
I"'
! 0 2 0 3 I 4 2 5 3 1 6 10 9 14 ~
Seco.l SO 100
! I 2 3 4 4 5 6 1 11 12 IS 16 23 24 en
K First 80 80 0 2 0 3 1 4 2 5 l 7 6 10 9 14 ~
Seco.d 80160 12344S671112ISI62324
R en
L Ft.st 125 125 • 0 2 0 3 1 4 2 5 3 1
10 9~ 141""' ?"
Seco.d 125 250 I"" 1 2 3 4 4 s 6 7 11 12 IS 16 23 24

II First 200 200

! 0203142537610914·
~
0
--.. 200 400 12344567lll215162324 ;J>
R ;;o
0
'FU ..
Seco.i 315
315 315
630 • ! 02031<253761091<~
12344567111215162324

P Firsr. 500 SOO 020Jl42Sl1610914 .

t""
""
See-d SOO 1000 ~ 1 2 3 4 4 S 6 1 II 12 lS 16 23 24
I ~
en
'Tl
0
;;o
. : : . ==,.
R Fihl 1!!50 1250 • -i"?'- 0203142537610914~
! n::::::::.: ~ .: =; ~ 1r ~
5ecoeMI 1250 2500 U I 2 3 4 4 S 6 7111215162324 U ;J>
o-j
S Fitsl: 2000 2000 0 2 o-j
s..-t 2000 <000 I 2
;;o
.....
o:l
..0 Use first sompli.~~tr plaa below t~now. lr sa.-ple size eqals OC" exa::eeds l01 ar batda si:JC. do 100 petlletll iupection. c:
o-j
.0 Use first sall'lplinlf pt.. abowe t~now. m
•e Acceptat~Ceaulftbeto
en
. Rejectiollaudlu
Use c:-spcadi"' single s-w:Ji•B plaa («, ahnntiwely, ~;~se- doubR s.mrlifll plnloelow, wbeft, •••illlble).
 TABLE lll-C- Double sampling plans for reJueeJ inspection (Master table) ,_.
0

Acceptable QulilJ ~.ewe~. (~educed iupec:rioa)t ""'

!ii&lllf"le c.-
--,]
o.~l 0.10 0.15 0.251 0.40 I o.~ 1.0 1.5 2.5 •.o 6.5 10 15 25 .. 65 100 ISO I 250 I 050
... I I ,... I ;:r
..... (\)
,;,. Is•..,,. 1.......
,;,. ....,,. 1
1...;.. 0.010 II 0.0151I 0.025 II o.o.o II I I I I I I I I I I I I I I I I I
lellf'l' saze
;I>
Ac Rei A< ReiAc ReiAc ReiAc ReiAe ReiAe ReiAe Rei A< ReiAc ReiAe Rei A< ReiAc ReiAe ReiAe ReiAc Rei A< Rei A< ReiAc Rei A< ReiAe ReiAe ReiAc ReiAe ReiAc Rei A< Ro O:l
n ~. n
A n ,.o.. • • • • • • • •
R • ~ ..:.? '"0
,_.
c ~~
~R!L ...u.-. . .• •. .• .• .• .• .• .• .• i" Ill
::s
0 First ··-<;>.on o2o o•o•l5273a5lo7uun
Secoad U ""V'" 0 2 0 I 53 6 • 7 6 9 8121216182226:11

FiDI • l"'fr n o 2 o 3 o o • , 5 2 1 3 • 5 ,. 1 12 u 11 0
Secood Iu ~ 0 2 0 ' 1 5 3 ' ' 7 6 9 I 12 12 16 11 2Z 26 3D

Firat 5
5ecood 10 • '<";>.1
u n1 o 21 o 31 o
""(....70204153647698121216
'I o 'I' 512 713 '15 tol""l"r-l~l~
G ..... 8
5ecood 16
• '-<?>lnlo 2lo 1lo 'lo 512 713 '15 toK"r-
u ~020,1536,7698121216 'I'
H Fi,. 13 13 • 2 0 3 0 ' 0 ' 1 5 2 7 3 • 5 101<
....... i3 26 0 2 0 ' 1 5 3 6 ' 7 6 9 8 12 12 16

..... 20 20 0203040.152731510.,.-;:..
Secood 20 .. ,... 0 2 0 • I 5 3 6 ' 7 6 9 8 12 12 16
..... 32 32
5ecood 32
-<?>lnlo s 1
U ~ o z2loo ,1lot 'lo 'I'
' , t , 12 16
512 ,713a tz'1510J"'ii"-
I~
L Fin1 50
"
50 • ..r,:.. n o 2 o 3 o , o , , 5 2 1 1 •I 5 1o,
....... 50 ioo ~ U "'<..>' o 2 o • 1 5 J 6 • 1 6 9 a 12112 161

.... 80 80 ..1· ~n ozoJo•o•ts2tllsto r-

5ecood 80 160 I~I U ~ o z o , t s 1 , , 1 , , a tz tz 16

firM 125 125 ·l~n o2o3o•o••527,as1o

Secood 125 250 u ~0104 1536 47698121216

p .... 200 200

o 2l o 3l o o 1 2 7l J • 1 5 'ollT
Secood 200 ... ~I 'lir!~ 1~
0 3
v
0 2
.'.. '''I ...'I' '. .......
0 •

R 0 3 0 •
0 • I 5

Ua !'ina u..,tiac pin belo• - · · If .ample Iiiii! eq.aJ• at efl:'Hd1 lot Of' bi!K"h size, do 100 ~- i ..pectto..
U• fint HlllpliftC plea aboft - · N
• Aceepcuc:e aunGer. <0
? .....,
Re • Rejectioa au..mer.
• U11111 cone•pondiftl •U.,Ie HIIIPii"' plu (or •lleftl•tiwely, •• ~le •11111rli.. pl•• below, •Wn •••ilable.)
t ~ lr. llfw u.e .ecad ....,1c, &be~ ......... beell eaeeecled, HI dre rejectloe a-t:.H •II• _,.ben fHCMd, .:cep~ die IGt. .... llliuiiMe _ . . iupact:iol (... tO. loA).
 TABLE IV-A-Multiple sampling plans for normal inspection (Master table)
N
<D
00
, .. ,.,.,,.,,J.I.- •.tu..lu, 1..--1~ !........,I in~t>MI)

'>:a"'Jll<" r........
S,.""'l.. lah•f' 0.!110 O.OIS UIJ25 u\llh llllh·, llltl u_t·, o::S uw 1.0 1.5 !.:i
Sampl"' 6.5 10 15 ,. (6 100 ISO :150 (60 1000
,"'""
..... ..~ ............. "·'"
l ..u .. ~ . ~
'" .. ...
-'c H" -\c 1\,. !J.c 1;,. \c ~: .. \r 1: \• ltr \< \!r \r- 1:.- \, lk- \r 1/1' \, 11.- M llf' .\c It Ac I~
" Ac ''"' ltc R"' A< H. .\clkA.:~Ar-ReAcfWA cRe"<::Re.\c:RlM.IkAcRe

\ .. . . . . . .
II
....... .... +-+ ......
.[1,. ~ -0-- .[1,. v
c
ll iF ...[!.. ...... ....... - ~ ......... ++++ ---
- - .......... ..... ;::
Fin~ 2 • 2 • 2 • 3 •4.04.0517 29ll2616
" 2 I • 2 0 j 0 3 1 5 1 6 3 8 4 1 0 7 - •• 11 19 11 Z1
Thi..i
""'""" 2 0 2 0 3 I 4 2 6 3 8 6 1 0 1 1 3 1 jJ919Z72939
fourtb 2
•
8 0 3
1 ' 375108131211192SZ134.4049
Fifth 2 10 I J 2 '
3 6 S 8 7 II ll IS 17 20 25 2936405358
Si:ub 2 2 '
12 I 3 7 9 10 12 14 17 21 Z3 3 I ll 6 47 C6 61
2 2 j
J ' ' 6
6 7 9 10 13 14 II 19 25 26 31 l8SJ547778
~..... ·"· ' 5
> Fi""'t J
" j • 2 • 2 • j I 7 2 • • 12 6 16
~oad J • 2 ' 0
0 '
Third J
6
i ij 0 2
0
0
J
3
0
I
3
•
..
1
2
I 6 3 8 4 1 0 7 1 4 1 I 19 17 27
'
6
] 8 6 1 0 8 1 3 1 3 1 9 1 9272939
Founh 3 12
• 0 3 1 ' 2 5 3 7 8 13 12 17 19 25 2734.4049
r.rth J 15 1 3 2 • 3 6 ' 10
7Jilll517201529 36405.158
'>inh 18 1 3 ' 8
7 • 1012141721 Zllll) 45 47 (6 68
5-ev.. l\lk ' ' 6
2 J •
3 5' 6 7 9 IV l31418192S2637.J85J >47778
j
'
First " 5 ::! • 2 • 3 0
051729 •
' "'-I
Tbir<l
foonth
3
5
5 10
J5
20
l 0
0
2
2
0
0
3
J
0
1
2
J
•
s
..
I
2
j
>
6
7
38410714
1
J
6108131319
6
8
3 ' 8 u 12 11 19 25 '"' '"'
Fiftb :!5 1 3 z1 4 3 6 ll ll1517202S29
s 8 '7 lO
Su:~h '3 JO 1 7 • 10 12 14 17 :!I 23 Jl l3
3 j ' ' 6
Selr~nt 5 .IS 2 • 7 • 10 11 u. ta 19 :5 Z6 n JS
1"7 ] ' ' Cll
G Finl 8 8 • 2 • 2 • l • • 0 • 1 7 2 •
0
'i«on" 8 16
1J ll • 2 0 3 0 3 1
' 1
6 0j 8' .. 10 7 u
:::
tn
Thirwl 8 ,, • 2 0 3 1 • 2 6 3 8 6 lO 8 lJ 13 19
Fnu"h 8 J2 0 3 1 J 7 5 10 • lJ 12 17 19 25 Cll
• 2 '
hftll 8 I 3 2 s 8 7 ll II IS l7 20 .!5 .!9
' 3 6
'i•uh 8 "'
. I 3 j 5 4 6 7 • 10 12 14 17 Zl Z3 Jl JJ
Y~cnth 8
,. 2 3 4 5 6 7 • 10 13 14 18 19 :s 26 J7 J8 ~
0
" .... Fint
~,
IJ
Jj
"'
Jj
26
1J I
0
2
3
•
0
3
3
•
J
4
5
0

1
•

'
0
3 8
5 1 7
• 10
'
7 l4
. ;J>
;:o
ThiM IJ j<} 0 3 1 4 2 6 J 8 • 10 8 Jj u 19 0
Fourth 13 52 1 • 2 s 3 7 s 10
8 Jj 1:! 17 19 ::s
f'iftl!. 13 3 6 5 8
7 ll lt IS J7:'0 ;3:.'9
. 2 '
Siuh 13 "78 J s 4 6 7 9 10 12 14 17 :!1 ::l Jl jJ
~- .., 13 9J • s 6 7 9 10 18 19
u '"
J7J8
s;:""
r·;.-u
1Jti • j
""' z
Cll
J . •
0 ' 0 ' 1 7 2 •
20 ."' • 2 0 3 0 3 1 5
1 6 j 8 • 10 - 7 It
Ttn...t
""""" "' 60 0 2 0 J I < 2 6
3 8 6 10 8 13 13 19
.,
0
~-"""" 0 3 I 4 3 7
5 10 8 lJ 1:! 17 19 15
Fihh
""' 20 100 I J 2 4
z3 '6 s 8
7 11 II IS ,,. ;:o
'<i•tb
"' I J 3 5 • 6 7 9 10 12 14 11 :!1 ""' ZJ ll jJ
~ ;J>
s._,.,.,..., "' 20 :: ,~
1Jl'"' z 3 4 5 6 7 • 10 IJ It 18 19
"" 37 Jtl i - :l
;:o
H
~ := u~ first U"'9Ji. pl•ll hoe-Jo- - - (~fet to cooti-iOOI o!t.bfe- roJiowi-s polp', . . Ilea"...,.). If .....leo aize e•~...b or~ Jnc or bu;o;h &izr, dn \00 ~ i~. o:>
~ = U..,. finn s.unpliq pl•o .t.oYe *"'-· c
"c- A.cc""f'l~"~· -'1
R., = Rot1r.-ctiocon .....t>n. tn
= u., com.,pontfi"'l' ~i"'l"' ,.rnplins pl .. (or •ltotmall....,ly. IUO! ,.,ahip~ nlllpli"f fl•• bf.k.w, wlw:n, •••ilabld. Cll
+-t- = U-"! cortespoading <!<>~~hi<' .....,plin« pl.., (Qr ollt';nlet!.-ely,- mwhipko N111pli1111 below, ~ ••eil.bk-1.
pi••
Acc..,uoe« 1101 pen!riued •t dai• siz.
'"ll'lpl"'
 TABLE IV-A-Multiple sampling plans for normal inspection (Master table)
(Continued) .....
0

Accqltable (Juility Lnds (nonnal insJM'clinn) "'

s.....,..... cu,u-
site -s......... lative 0.010 0.015 0.025 0.040 0.065 0.10 0.15 0.25 U.-10 0.65 1.0 1.5 2.5 4.0 6.5 10 15 25 40 65 100 ISO 150 400 650 1000 -l
,..,. Samplr size sample ::T
.. u (1)
'"It""
-\c He Ac lte Ac H Ac IWAc Re <\c He Ac Re <\c He Ac Ar Ac He Ac He Ac AeAcfleAcHe..\cH.. AcReA c Ae Ac He Ac Re Ac Re Ac R" Ac fl,. Ac A Pkift.AcH.. AcH"
;t>
Fine J2 J2 • 2 • J •• 0 • 0 5 I .7 2 9 -"" 00
• <;~and J2 . •
.' 2 0 J 0 J I 5 I 6 J a • 10 7 I< I-<! I" ?" (")
Third J2 0 2 0 3 I 4 2 6 3 8 6 10 8 13 13 19
fo"ounh J2 I :>I
" 0 J I 4 2 5 3 7 5 10 8 13 12 17 19 25 '"0
.-·,rth 160 I 3 2 4 3 6 5 8 7 II II IS 17 20 25 29 .....
Si•th .i!
" 192 I 3 J 5 4 6 7 9 1012141721 2.J Jl Jj §
S,.,...ndt J2 224 2 J 4 5 6 7 9 I0 1314181925 26 3738

L l'"irst so 50 • 2 • 2 • 3 • 4
0 4 0 5 I 7 2 9
S«ond
11oird
Fourth
so
so
50
100
ISO
.!00
n~n •
0
0
2
:!.
3
0
0
I
3 0 3
3 1 4
4 2 5
I
2
J
5
I 6 3 8 ..
6
3 8 6 10 8
7
5 I 0 8 13 12
10
lJ
17
7 14
13 19
19 25
Fifth 50 250 I 3 2 4 3 6 5 8 7 1I II IS 17 20 25 29
Si11.th 50 300 I 3 3 5 4 6 7 9. 10 12 14 17 21 2J 31 33
~YPI'Ith 50 350 TI R 2 3 • 5 6 7 9 10 13 I .. 18 19 25 26 J7 38

First ·.., • 2 • 2 • 3 0 4 0 5 l 7 2 9
• 2 0 J 0 3 I 5 I 6 3 8 .. 10 7 14
• ,.,.... "'
Third
"' 160
240 0 2 0 J I 4
.'
2 6 3 8 6 I0 8 13 lJ 19
F'ounh
"' 80 320 0 3 I 4 2 5 3 7 5 10 8 IJ 12 11 19 2S
Filth 400 I 3 2 4 3 6 5 8 711 II I5 17 20 25 29
Si11th
"' 80 I J 3 5 4 6 7 9 10 12 14 I 7 21 23 .II JJ.
Se-~nth 80 560
"" 2 3 4 5 6 7 9 10 13 14 18 l 9 25 26 37 .38
First 1:5 125 • 2 • 2 • J • 4 0 • 0 5 I
' 1:!5 ::so
~n • 2 0 J 0 3 I 5 I 6 3 8 4 I0 7 II
7 2 '
Tloi~
'"'""" 125 J75 0 2 0 3 I 4 2 6 3 a 6 I0 8 I J 13 19
1-'ourth 1.!5 ;oo 0 3 I 4 2 5 J 7 5 10 8 13 12 I 7 19 25
Fif!l\ I.:S 623 I J 2 • 3 6 5 8 7 II II 15 17 2025:9
"iuh 1::5 7SO I 3 J 5 4 6 7 9 10 12 14 17 21 2J Jl JJ I
S..v,.nth 1.25 81S 2 j 4 5 6 7 9 10 13 14 18 19 25 26 37 J8
,. firs! !00 :00 • 2 • 2 • 3 • 4 0 4 0 5 I 1 2 9
SKonoi
Thin!
~·ounh
"''
'""
400
600
..,.
in •
0
0
2
2
3
0
0
I
3
3
4
0
I
2
3
4
5
I 5 I 6
2 6 3 8
3 7 5 10
3 8 4 I0
6 10 II 13
8 13 I~ 17
7 14
13 19
19 25
Fifth ""' 1000 I 3 2 • 3 6 5 8 7 II II IS 17 20 25 29
o;;;,,h 000
""' 1200 I 3 3 5 4 6 7 9 10 12 It l1 21 2J Jl lJ
~"''fllh _'QIJ 1400 2 3 4 5 6 7 • 10 I 3 14 18 19 25 26 37 JB

f•rst lj:j JIS • 2 • 2 • j 0 4 0 5 I 7 2 9

'• V.o:on•l
TloiN
fourth
Jl:i
Jl5
Jl5
...
945
1260 ~R
iR •
0
0
2
2
3
0
0
I
J
3
4
0
I
2
j
4
' 6
..
I 5 I 6 J 8 4
J 8 6 10 8
5 l 7 5 10 8 13 12
10 7 14
13 13 19
17 19 25
t-"ifth JJS 1575 I 3 2 4 3 6 5 8 7 II ll IS
Si"h 315 1890 I 3 J 5 4 6 7 • 10 12 14 t7 21 2J
""' 2529Jl ll
S..v.. nth liS z~:; z J • 5 6 7 • 10 IJ 14 18 19 2S 26 37 38
I< First 500 2 2 • J • 4 0 • 0 5 I 7 ! 9
500 1000 2 0 J 0 3 I 5 I 6 j 8 4 10 714
Tnifll
'"'""" 500 1500 0 2 0 3 I 4 2 6 3 8 ' 10 8 13 13 19
Fow.rth SilO 2000 0 3 I 4 2 5 J 7 5 10 8 13 12 17 19 25
Fiflh
St11th
500
~i I 3 2 4 3
4
6 5 8 7 II II IS 1720
6 7 • 10 12 14 l7 21 23 31 J3
500 I l 3 5 ""'
~Y .. flf 500 3500 2 J 4 5 6 7 • 10 1314 18 19 25 26 37 38
:..........._ .~
___J __ -
= V • Us.e r;,_, H"'Pii., piN belo• lln'OJIW.
~
1£ W"'Pi"' •ne eq>~Sls ar ucl'edl lot or b.1o:h siu-. On 100 ~n;enr 1nspeocrion.
N
..:0- : Uw fil"ll ••"'91i"' plan above •wo• (..ferro ~<!'dins pate, when n«USV"f).
.4.: = Accerr•"« n111ftbH.
<D
tt~ ~ Reje<:liOflaiOI'I!Oer.
<D
:::o: Uw c-spotltlii!S si11fJie u~~~phr~S plan (or ahemari"ly, ue "'"hir!e rlaro bel ow, wh.eft audahlel
= A.cxeplaiCe- pennirled at rhia .."""le ai:e.
 TABLE IV-8-Multiple sampling planS for tightened inspection (Master table)
0
""
0

Malpi.W.!Juljty ~-b...... i~J

s-,.~~ c-

........ aiu
.:.: I_.. I'::' I=- 1-+-+-+-lf-+--+-+-+-l--+--+-+-1--f--+---+---~--4-+-+-+---lf-+-+-+~
n~Jln...o-·
fl . . ~ ~-------
· ····
-- ------
fu,q
W.-.!
ill • •
•
2• 2•
2 0 30
3•
.Jl
4010 618310615
52 73 961210171615
nu..t 6
' 0 2 0 3 I 4 2 6 4 9 7 12 11 17 17 2tf' 36
F.-.L 0 3142 53 761110JSI62221ll716
fiflb
"'•nlo
••• ! ~ ~ : ! : ~ : 1: ~ :: ~ ~ :pz ~~ :
;.'""''" "..
font l l a •2•2•J•404Q6J8JJQ6JS
;., ... l • 20 3 0 3152 71961210171625
Tloood
,..... l
l 12
••
~ t . "".""....~ .'.
f,flb l 1 3 2 3 6 S I 9 12 lt 17 22 25 l2 37 9 5S
. : ~ ~ !t ~ : : ~ : .~ t! ~ :: ~ ~ ~~ :
Sill.1t. l 18 l 33 s. 6 7 92 ltl8202729.4361 6t
....... l 21 2 34 s 6 79104 152l2232llt84972 73
I"'
Fi,.. 5 5
,.S.C.-I 5 10 •
• 2
20• 320• 331•5 2
' 7 '[
0 3 96 61l 28 l"'ii'-1"'
Tloiool 5 IS 02031426&97121117
fouoV 5 ,. 0 ll 42 53761110151622
Fifllo 5 25 I 32 43 6 5 89124172225
,.... 5 311 1 l l 54 6 1 9 1 2 1 4 8 202729
::;.,..., s l5
en
2345679101415 223233

''"'
. • • • 2 • l •• ,.
~
m
8 • 2 0 l 0 j 1 s 2 7 l ' 9 6 12
'""-'
n1..t
,_, 8
8 ,.,.•• 0
•
2
l
0
l
..
l
•
l
2
4 2 6 4 • 7
5 3 7 6 11 10
12 l
.6~
IS 16
17
8
22
en
FifO 8 I l 2
' 3 12 14 17 25
Sid 8 l 3 3
5 ' •• 18 20 29
8
....
56
: ~ :1.: ~
....... ~ I"' 2 l • 5 • 7 9 10114 1521 22 33
§,:
First lJ lJ ::<'
...,_, ll ,. : :: ~ : ! : :: ; : : ! .:14 0
T\ifll ll I i 02031426497l2UJ1
,_ ll 52
" 0 31 42 53 761110151622
..,
Fire• ll 65 132 43 65 891%14172225
Sillth ll 78 l 3 3 5 4 6 7 912 14 18 20 Z7 29
§;:
s...,_.,t. ll 91 z
2 3 4 5 6 7 9 1014 15 21 22 32 lJ
t'itsl ,. en
""'-' ,."' .. • 2 0 3 0 3 1 5 2 739612 'TI
0 2 0 l l 4 2 6 4 7 12 11 17 0
t'o.nh ... ·····~···~··8~
0Jl42 37611!0151622 ::<'
Firo.
""" ,."' 100
"' 13243 589114172225
Si..
u. 13354 79J21182DXT29
20
....... "' ::: 1-.!..!r 1.. 1..
j :!J456 910141223233
~
::<'
.....
...(> - Uae fin1 sa.p~;.,..,a.. a.low • - c,.,r.riO~aa.ort.w.. rouo.;.,,.., ..___,).rr....-.-.,......~ lol.oriiMcla.a.. dDtoo,.._. ........
-<> - -.u.rn~UMf'li.-pl•._-.
"'Si
Ac- Ac:~~
m
He - ~~---- en
u-~.................. taraJ.....awel,. -.urq.~e ...,u.. pl• ..._, ..._...UIIble).
U.~......ao:..-pl.i._pl•(orel~.ue-.lliple .....illrpl•be._,..._........a.J,.
~- ........... diU ..........
 TABLE lV-B- Multiple sampling plans for tightened inspection (Master table)
(Continued) ......
0

-\ccC'puloiP ()ualitv Lr•e]$ (ught~:~~...t •nspeeti<~nl

Sam pi~ "'
Sa•eJilf'
s•....,,l!' C"m"-k---r-...--r--.--.----r-r-r::lrt.-;-r::-~~T:T:-r-;;;T.;J-;;-r.;~-:;;r:~
co•k- I l 0.010
I I I I Io.o1s
I I Io.ozs
I I I 0.040
I I o.065 o 11>
25 I 40 I 65 I 100 I o as o.25
I 250 I 400 I 650 I 1000 o 40 o.65 1.0 1.5 2.s ~-o 6.5 10 l5 ISO
IPUPr
>-l
:::;;.1 1 1 1 1 1 1 1 1 1 1 I I I I -I I I I I I I I I I I I :7
A.o:- Hll'IAc Hel .. c fh•J.I.c He)4.c H•dA.c llr(.\c IIP)Ac lle)Ac HeJAc HejAc fl._.JAc HPJ..\c llejAc RrJAc HP)Ac Hrl.\c lleiAc lleiAc: Hehc ReiA.c Reh.c _Rrlo\c Re)Ac Rel-'c Hoe!A.c R" Cll

Fint 32 32 2 2 • 3 0. >-
s.-~d 2 0 3 0 3 I 5 2 7 o:J
.. 30 611
9 6 128
Thu-d 32
" ••
96 2 0 3 I 4 2 • 7 12 ll 17 n
Fourth 32 128 3 I • 2 5 3 7 6 II
. ' 10 IS 16 22
Fifrh 32 160 3 2 • 3 • 5 8 1417 22 25
~~i•rh 32 192 3 J 5 •• 12' 12
14 18 20 27 29
~Vef'ltk 3 • 5 • 7 7 ' ""......
32 '10 14 15 21 22 3Z 33
5Q
"'
50
ll ll! • 2 2 l 0 • 0 • I 8
§
~COIId
It· 50 100 • 2 0 3 0 l I 5 2 7 • 12
Third 5Q ISQ 0 2 0 3 I • 2. 712
l ' II 17
Fo.. nh 5Q 0 3 I • 2 5 3 7 6 II 10 IS 1622
.. .'
Fifth 50 25Q I 3 2 • 3 • 5 8 14 17 2225
Si•th 5Q
""
300 I 3 3 5 • ' 12
• 7 ' 12 u. 18 20 2729
Srventh 50 350 2 3 . 5 • 1 9 10 14 15 21 22 32 33

First 60 80 • 2 • 2 3 0 • I 8
SPcond 60 160 • 2 0 l 0 3 ; 5' 02 '7 • 12
0 2 0 J I 2 3 '
712
Third 80 240 lll7
Founh 80 320 0 3 I 4 2 5 3 7 6 11 10 IS 16 22
Filth I 3 6 5 8 9 12
80 2 ' 3
. .' ' I" 17 2::! 25
Sinh 80 480
"" I l 3 5 • • 7 9 12 14 1820 27 29
Sevenlh 80 560 2 3 • 5 • 7 9 10 1415 21 22 JZ 33
Firs! 125 125 2 • 2 • 3 ' 0 • I 8
,.,~, 2 0 3 0 3 5'
; 02 • 12
125 250 7 3 '
Thn•l 125 375 ; 2 0 3 I 4 2 9 712 ll 11
Four1h 125 500 0 3 I 4 2 5 3 7 6
. ' II 10 IS 16 22
Fifth 125 625 I 3 2 • 3 • 5 8 9 12 14 17 22 25
Suttn l2j 75<J I 3 3 5 • • l 9 12 14 18 :!0 27 29
~~~~nlil 1.Z5 875 2 3 '5 • l 14 IS 2122 12 ]]
' 10
fit.sl 200 200 2 2 • 3 • ' 0 ' 0 I 8
So<~d 200 2 0 l 0 l I 5 2 7 3 • 12
Third 200 600
"" ; 2 0 3 I 4 2. ' Y 7 I It 17
Fo10nh 200 800 0 3 I 4 2 5 3 7 6 II 10 I 16 Z:!
Fiflb 200 1000 I 3 2 l. 5 8 '12 2:? 25
Si11h 200 1200 I 3 3
•
5 ' • 7 ' 12 14 !: ~ 27 29
Snen1h 200 1400 2 3 • 5 • 7 l4 15 21 z. 32 J3
' 1U
F1rs1 liS 315 2 • 2 • 3 0 • 0 I
Second 315 630 2 0 j 0 3 I 5 2 7 3
•'
Thud 315 ; 2 0 J I ' 2 • T IZ II
Founh 315 1260
"' 0 3 I 4 2 5 j l 6 II 10 IS 16
'. . . '
Firth 315 1575 I 3 2 ' j
• 58 912 u. 17 2:l :!5
Si•th 315 1890 I 3 3 6 7 'I 1:? \4 18 :?tl 27 ZCJ
SeVPOih 315 ZlOS 2 l.
55 •' l 9 I0 14 IS 21 2~ 32]
Firs! 500 500 ' 2 ~ 2 •
Second 500 1000 • 2 0 3 0 33 o
I 45 02 47 03 69 16 8
12 1 J
Thud 500 1500 0 2 0 J I 4 2 6 4 9 7 1.? ll 17
Formh 500 2000 I • 2 5 J 7 6 II 10 IS 16 :22
Fthh 500 2500 2 • 3 6 5 8 9 12 14 !1 12 25
Siuh 500 I 5. 6 7 9 I~ 14 18 .!U 20 29
3000
"' ''
Sev .. ruh 500 3500 1J L 'I '' T 9 10 14 IS 21 22 32 3.1
' .
Fm.t 800 800 L
S.. ranrl 800 1600 L
Th•n.!. 800 2400 0 2
four1h 800 3200
Frfth 800 1000
Srulo 800 4800
5<!'Y .. IIIh 600 5600

- ('$<!' (rM!I '~mpl1nll' tol~n 1-·J.,,. ,lf1'<l• If S~"'pl<!' SIZ.. ~~~~Js or ,.r("f:f:d, Jo1 or bll<:h $IU'. do JOQ Pf'I'CO'III n•5pt'("llon.
2: = 11M" frrsl Suopl.n..: rol3n ~l~w~ ItT""' {,..,f~r I<>JI"'r<!'dtn~t Plll:t". wh<!'n nK~S$~1'y) 0
,, AccO'pianc,.nunol,..r
=
"'......
Ho = llO'IO'CUon numL<"r
"" l'"" con-es~ndtnll s•n,:l~: sa..,phnc l'lan lor ah~:rnauv .. Jy. uM' ., .. JupJ,. sanol'hnll' plan b!Pio•. wlof:f!" auolabl<!')
- ~- IDOt ~nH 81 U.•• a&Dpte 11&e.
 TABLE V-8-Average Outgoing Quality Limit Fa&tors for Tightened lnspe&tion (Single sampling)

A~ Qulity Left!

Code So.ple
...... m.
0.010 0.015 0.025 0.0411 0.065 0.10 0.15 0.25 0.40 0.65 1.0 1.5 2.5 4.0 6.5 10 15 25 40 65 100 ISO 2SO 400 650 1000

A 2 42 69 91 160 2160 400 6211 910

B 3 12 21 46 65 liD 110 210 410 650 1100
c 5 7.4 17 27 39 63 100 160 2SO 390 610

t---
D 8 4.6 II 17 24 40 64 99 160 240 380
E 13 2.8 6.5 II 15 24 40 61 9S ISO 240
I' 211 1.8 4.2 6.9 9.7 16 2111 40 62
(J.l
0
N ~
I
G 32 1.2 2.6 4.3 6.1 9.9 16 25 ])
H 50 0.74 1.7 2.7 3.9 6.3 10 16 25
J 110 0.46 1.1 1.7 2.4 4.0 6.4 9.9 16
!---
K 125 0.29 0.67 1.1 1.6 2.5 u u 9.9
L 200 0.18 0.42 0.69 0.91 1.6 2.6 u 6.2
II 315 0.12 0.27 0.44 0.62 1.0 1.6 2.5 3.9
r--
N 500 0.074 0.17 0.27 0.39 0.63 1.0 1.6 2.5
p 800 0.046 0.11 0.17 0.24 0.40 0.64 0.99 1.6
Q 12SO 0.029 0.067 0.11 0.16 0.25 0.41 0.64 0.99
1--
I( 2000 0.018 0.042 0.069 0.097 0.16 0.2111 0.40 0.62
3150 0.027
'

Notol For the exac:t AOQL, the aboYe nlu.. muat be muldplled b,. ( 1 • Sample ab< ) ( ••• 11.,)
Lot or Batch abe
 TABLE VI-A- Limiting Quality (in percent defective) for which Pa = 10 Percent
(for Normal Inspection, Single sampling) .....
0
(N

Acceptable Quality Level

Code Sample ~
(!)
letter size
>-
0.010 0.015 0.025 0.040 0.065 0.10 0.15 0.25 0.40 0.65 1.0 1.5 2.5 4.0 6.5 10 n""'
..,.....
i P'
A 2 68 ;::!

B 3 54
c 5 37 58
i--- 1---
D 8 25 41 54
E 13 16 27 36 44
F 20 ll 18 25 30 42

G 32 6.9 12 16 20 27 34
H 50 4.5 7.6 10 13 18 22 29
J 80 2.8 4.8 6.5 8.2 11 14 19 24

K 125 1.8 3.1 4.3 5.4 7.4 9.4 12 16 23

L 200 1.2 2.0 2.7 3.3 4.6 5.9 7.7 10 14
M 315 0.73 1.2 1.7 2.1 2.9 3.7 4.9 6.4 9.0

N 500 0.46 0.78 1.1 1.3 1.9 2.4 3.1 4.0 5.6
p 800 0.29 0.49 0.67 0.84 1.2 1.5 1.9 2.5 3.5
Q 1250 0.18 0.31 0.43 0.53 0.74 0.94 1.2 1.6 2.3

(N
R 2000 0.20 0.27 0.33 0.46 0.59 0.77 1.0 1.4 0
(N

- I'
 tH
0
.j>.
TABLE VI-B- Limiting Quality (in defects per hunJreJ units) for which P 11 = 10 Percent
(/or Normal lnspectwn, Single sampling)

Acceptable Quality Level

Code Sample
leuer size
0.010 0.015 0.025 0.040 0.065 0.10 0.15 0.25 0.40 0.65 1.0 1.5 2.5 4.0 6.5 10 15 25 40 65 100 150 250 400 650 1000

A 2 120 200 210 330 460 590 770 1000 1400 1900

8 3 77 130 180 220 310 390 510 670 940 1300 1800

c 5 46 78 llO 130 190 240 310 400 560 770 1100

-
0 8 29 49 67 84 120 150 190 250 350 480 670
E 13 18 30 41 51 71 91 120 160 220 300 410
F 20 12 20 21 33 46 59 77 100 140
'--
G 32 7.2 12 17 21 29 37 48 63 88
H 50 4.6 7.8 11 13 19 24 31 40 56 Cll

J 80 2.9 4.9 6.7 8.4 12 15 19 25 35

~
m
- Cll
K 125 1.8 3.1 4.3 5.4 7.4 9.4 12 16 23
L 200 1.2 2.0 2.7 3.3 4.6 5.9 7.7 10 u ~
0
II 315 0.73 1.2 1.7 2.1 2.9 3.7 4.9 6.4 9.0
>-
:::0
0
f--
:II 500 0.46 0.78 1.1 1.3 1.9 2.4 3.1 4.0 5.6 "'
p 800 0.29 0.49 0.67 0.84 1.2 1.5 1.9 2.5 3.5 ~
Cll
(.) 1250 0.18 0.31 0.43 0.53 0.74 0.94 1.2 1.6 2.3 'T1
0
:::0

II 2000 0.20 0.27 0.33 0.46 0.59 0.77 1.0 1.4 ~

-l
:::0
1-<
L .. ·- -
o:l
c:
-l
m
Cll
 TABLE VIII-Limit Numbers for Reduced Inspection

Accepooble Q11111i1y Level

Number of
sample uaita
r-1-10
lo&a or bau:hea
0.010 0.015 0.025 0.040 0.065 0.10 0.15 0.25 0.40 0.65 1.0 1.5 2.5 4.0 6.5 10 15 25 40 65 100 150 250 600 650 1000

211-29 0 0 2 4 8 14 22 40 68 115 181

30-49 0 0 I 3 7 13 22
.. .. .. .. .. .. .. .. 36 63 165 178 :rtr
.. .. .. .. .
50-79 . . 0 0 2 3 7 14 25 63
...
40
... . . 110 181 301
. . . . . . .
80-129 0 0 2 7 14 24 42 68 105 181 297
130- 199 0 0 2 4 7
.. .. .. .. .. .. .. . ' 13 25 42 72 115 177 301 490
200-319 . . . . . . 0 0 2 4 8 14 22 40 68 115 181 277 471
... ... ... ...
w
0
U'1 320-499 0 0 I 4 8 14 24 39 68 113 189
soc -799 0 0 2 3 7 14 25 40 63
.. .. .. .. .. .. .. . 110 181
800- 1249 . . . . 0 0 2 4 7 14 24 42 68 105 181
... . .
1250-1999 . 0 0 2 4 7 13 24 40 69 110 169
2000- 3149 0 0 2 4 8 14 22 40 68 115 Ill
.. .. .. ..
3150- 1999 . 0 0 I 4 8 14 24 31 61 lll 186
...
. .
5000-7999 0 0 2 3 7 14 25 40 63 liD 181
8000- 12499 .. . 0 0 2 4 7 14 24 42 68 105 181
12500 - 19999 0 0 2 4 7 13 24 40 69 118 169
...
20000- 31499 0 0 2 4 8 14 22 40 68 115 181
31500 - 49999 0 1 4 8 14 24 31 67 lll 186
sooooao... 2 3 7 14 25 40 63 110 181 301

Deaotea that the number of hllllple oilS &a. De laat tea loa. or baldaea is aot nmcieat for recluc:ed ialpeClioa lor daia AQL. Ia dais t.auace •re dlu ' - Iota or batches . .,. be uMCI
r..e
ct~lcvlatioa, provided dt.c the lola or ...cchea uMd ate die IIIOal receat oaea ia aqua.ce. d.at daey ••ve all beta • aOIIDal iupectio., aM diu ...ae Ua beea n:jected while oa oripall ~.
 TABLE IX- Average sample size &urves for Jouble antl multiple sampling
(1'10t'1'114l antl tighteneJ inspedion) tN
0
"'

3/h

112•

1/h

n x pr..-uon dlfldlve

3/h
Avenge
Sample U2a

Size ~
,,..... I I I I/ I I./ I
""' "'" en
0
3::
m
en
n x prq!Ortlon defldive ~
z
0
;J>
;:c
0
'"1:1
3/h L / - '-! L / I ,_ ........... I /,...,.,1 I /r::-.. '-J 1:'""
""" s;en
112• L / l L / --.'1 L / l I / I I I I
'Tl
0
;:c
.,.............. I L.._......' I L---"" I I ~ I l ___../ I
;J>
-l
-l
;:c
u 39 52 0 18 Zl J6 . S& .....
"' • txl
n x prqJortlon defective c:
-l
• = Equiweleel •iqle ••~~pie aize m
c: = Siafle •Uiple acc:ept•ce a...tter
en
t = AOL lcw •IHIUI ia.sp:c~i.-.
 TABLE X-}-Tables for sample size code letter: }
......
0
CHART J OPERATING CHARACTERISTIC CURVES FOR SINGLE SAMPLING PLANS V>
PERCENT OF' LOTS
~XPECTED TO Bt: (Curves for double and multiple sampling are matched u closely as practicable)
AC.CtPTEO tP_,l

;- . t-..• -l
~ ::r
(1)
!";;;
:H i t I+ + t
)>
80 " -r t>:l
\ ft+ . . f- I n
,, ~- +
' ··t+ t t !- -o
60
' \. ......
I i ~ ~r- ll'
so ;=j
- jt- _,
+a+ ~ -i r·
...
" '1. ' 'I..
;. i ·,
30
0. 0.65 1.0 -,;;;_l.5 2.5 6.5 10 IS i1+#1i
20 '·'
"'..i ',.,.,_ f'.l ">: ~.

10
~ >t-i_t-;Jt
' ,... ~
.,_ ' +tt T ~.::!-
10 ll 12 1J 14 IS 16 17 18 19 2() 21 22 23 24 2S 26 'l7 28 29 30 31 32 JJ 34 35 36 37 38 J'l ~0

QUALITY Of SUB\11TTF:O LOTS (p, in percent defective for A!.)L's ~ 10; in defects per hundrt>d twits ror .\()L's >IOl
Nole: figure3 011 c:UO"ea are Acceptable Qualily Le,.els IAQL"s) lor 11onnal inapeetioa

TABLE X-J-1 - TABULATED VALUES FOR OPERATING CHARACTERISTIC CURVES FOR SINGLE SAMPLING PLANS

.<\ccertable Quality Lf"vels (normal inspection)

Po o.l5 I o.65 1.0 1.5 2.5 4.0 X 6.5 10 0.15 0.65 1.0 1.5 2.5 4.0 X 6.5 X 10 X 15
><
p (in percent defective) p (in defect!> per hundred units)

99.0 0.013 0.188 0.550 1.05 2.30 3.72 4.50 6.13 7.88 9.75 0.013 0 180 0.545 1.03 2 23 3.63 438 5.% 7.62 9.35 12.9 15.7

95.0 0.064 0.444 1.03 1.73 3.32 5.06 5.98 7.91 9 89 11.9 0.064 0.444 1.02 1.71 3.27 4.9B 5.87 7.71 9.61 11.6 15.6 18.6

90.0 0.132 0.666 1.38 2.20 3.98 591 6.91 8.95 ll.O 13.2 0.131 0.665 1.38 2.18 3.94 5.82 6.79 8 78 10.8 12.9 17 l 20.3

75.0 0.359 1.202 2.16 3.18 5.30 7.50 8.62 10.9 13.2 15.5 0.360 l 20 2.16 3.17 5.27 7.45 8.55 lOB 13.0 15 3 19 9 234

30.0 0.863 2.09 3.33 4.57 7.06 9.55 10.8 13.3 15.8 18.3 0.866 2.10 3.34 4.59 7.09 9.59 10.8 13.3 15.8 18.3 23.3 27. l I

25.0 1.72 3.33 4.84 6.31 9.14 11.9 13.3 16.0 18.6 21.3 1.73 3.37 4.90 6.39 9.28 12.1 13.5 16.3 19 0 21.8 27.':! 31.2

10.0 2.84 4.78 6 52 8.16 11.3 14.2 15.7 18.6 21.4 24.2 2.88 4.86 665 8 35 11.6 147 16.2 19.3 22.2 25.2 30 9 35.2

5.0 3.68 5.80 7.66 9.39 12.7 15.8 17.3 20.3 23.2 26.0 3.75 5.93 7.87 9.69 13.1 16.4 180 21.2 24.3 27 4 334 37.8

1.0 5.59 8.00 10.1 12.0 15.6 18.9 20.5 23.6 26.5 29.5 5.76 8.30 10.5 12.6 16.4 20.0 21.8 25 2 :28.5 31.8 38 2 42 9

0.25 1.0 l.S 2.5 4.0 6.5 10 0.25 1.0 l.S 2.5 4.0 6.5 10 15 V>
X X X X X X X 0
A.cc~ptable Quality Levels (tightened inspection)
"
"-'-' AU...!- P-Ia •-•.W• _.._ Pot-4!~ • • ~ .. .._ aa-£aL
 TABLE X-J-2 - SAMPLING PLANS FOR SAMPLE SIZE CODE LETTER: J
0
"'
00

Acceptable Quality Levels (nonnal iaspectioo)

Typeol Cuma- Cumu-
lati..e Leaathaa Higher lative
-pli.. 0.15 0.25 0.40 0.65 1.0 1.5 2.5 4.0 6.5 10 15 thao
plu sample 0.15 X X X X sample
llize
15
size
Ac fte Ac Re Ac Rc Ac Re Ac Re Ac Re Ac Re Ac Re Ac Re Ac Re Ac Re Ac Re Ac fte Ac Re Ac Re Ac Re Ac R

SiOjJie 80 v 0 1 1 2 2 3 3 4 5 6 7 8 8 9 10 11 12 13 14 15 18 19 21 22 .6. 80

Use Use Use

50 v . 0 2 0 3 1 4 2 5 3 7 3 7 5 9 6 10 7 11 9 14 11 16 l::.. 50
Doable i Letter Letter Leuer
I 2 3 4 4 5 6 7 8 9 11 12 12 13 15 16 18 19 23 24 26 27 100
100
H L K

20 v . • 2 • 2 • 3 • 4 0 4 0 4 0 5 0 6 1 7 1 8 2 9 .6. 20

40 • 2 0 3 0 3 1 s 1 6 2 7 3 8 3 9 4 10 6 12 7 14 40

60 0 2 0 3 1 4 2 6 3 8 4 9 6 10 7 12 L 13 11 17 13 19 60
Multiple (/)
80 0 3 I 4 2 5 3 7 s 10 6 11 8 13 10 IS 12 17 16 22 19 25 80
~
t!1
100 I 3 2 4 3 6 s 8 7 11 9 12 11 15 14 17 17 2022 2525 29 100
(/)

120 I 3 3 s 4 6 7 9 10 12 2 14 14 17 18 20 21 2327 29 31 33 120

~
0
140 2 3 4 s 6 7 9 10 13 14 14 IS 18 19 21 2225 2632 33 37 38 140 ;I>
:;c
0
Less thaa Higher '1::l
0.25 0.25 X 0.40 0.65 1.0 l.S 2.5 4.0 6.S 10 IS than
X X X X 15 s;:
z
(/)
Acceptable Quality Levels (tightened inspection)
-- I ...,
0
:;c

~
b Use aen preceding sample size code letter for which acceptance aad rejectioo numbers are available. -l
:;c
Use aexr .subaequent .sample size code letter for which acceptance and rejection numbers are available. H
v Acceptuce number
00
Ac c
-l
fte Rejection number t!1
(/)
Use single sanpliq plan above (or ahematively use letter M)
a = Acceptaace DOt permitted at this sample size.
10.3. The ABC Plan 309

There are other special levels S-1 to S-4 for special conditions. See Section

9. 2.

10.3.3. Entering the Standard for a Plan. The following steps are taken

to obtain a sampling plan:

1. Decide on the size of lot N, which will be sample inspected. This

need not be a production lot size.

2. Decide upon an inspection level, in general II.

3. Using 1 and 2, enter Table I and find the corresponding sample-size

code letter A, B, to R, the last calling for largest sample sizes.

4. Decide upon single, double or multiple sampling.

5. Decide whether to start with normal (almost invariably), tightened or

reduced sampling.

6. Table IIA, IliA or IVA will thus be determined by 4 and 5 (if

normal inspection is used.)

7. Decide upon the inspection basis of defectives or defects.

8. Decide upon the desired AQL: for defectives (10.0 or less), or

defects per 100 units; using only what is available in tables.

Note that ".015" is .015%, or decimally .00015.

9. Enter the table determined in 6, using the AQL for the column from

8 and the row for code letter in 3, This commonly gives the

acceptance-rejection numbers, the sample size(s) being given to the

left of this block.

10. Following the above, we may reach an asterisk or an arrow. An asterisk

means to use the single sampling plan for the desired AQL and code

letter, instead of double or multiple sampling. If an arrow is

encountered, follow it to the first block with acceptance-rejection

numbers, using sample sizes to the left of this block, not to the

left of the original block.

310 SOME STANDARD PLANS FOR ATTRIBUTES

10.3.4. Further Explanatory Comments. Section 2 in the Standard gives

a classification of defects. Three categories of defects are given: critical,
major and minor, as defined. For critical defects 100% inspection or testing

is commonly used, although sampling may be used after adequate excellent history
is built up and then only for very low AQL's. Major and minor defects differ
according to seriousness and thus carry lower and higher AQL 1 s respectively.

Commonly in any product, there will be a list of major defects and another

list of minor defects. It must be born in mind that when there are ten or
fifteen different possible major defects, say, it is quite difficult to have

the AQL for defectives to be extremely small, because d will be the total

number of all major defectives, that is, pieces with one or more major defects.
Note the emphasis in Section 7 on proper choosing of samples. This is
essential to sound decisions.

Rules for "switching" are given in Section 8. As regards normal to

tightened inspection and vice versa, these have been greatly simplified from

earlier Standards, by adopting suggestions of Dodge (5]. It has proven rather

difficult in practical use to insure that those in charge of the inspection

go to tightened inspection when they should. If tightened inspection is not

invoked when called for, then the intended consumer protection provided by

the Standard is greatly weakened. This is because the plans are set up with

OC curves intentionally lenient toward producers who are cooperating by having

p 2 AQL. But to do this the shape of the OC curves shows that Pa for p ' 's a
bit above the AQL still give quite high Pa's. Hence unless tightened

inspection is invoked a producer may fairly comfortably run p above the AQL

for quite a long time. The newer rule calls for tightened inspection whenever

2 out of any 5 consecutive lots are rejected on initial submission.

In order to get back onto normal inspection, 5 consecutive lots must be

passed while under tightened inspection. Moreover this transition must be

10.3. The ABC Plan 311

accomplished within 10 consecutive lots, or else sampling inspection is

abandoned altogether.

Switching rules for going to reduced inspection are four in number,

~ of which must be fulfilled to justify the switch. Between them they quite
I
well insure that there will be a low probability of any lot with p above the

AQL being submitted. Thus a lenient plan with small n is feasible. See

Section 8.3.3. On the other hand there are four conditions for remaining

on reduced inspection. If any ~ fails, inspection is to revert to the

normal plan, See Section 8.3.4. One of these is a special procedure. For

example, see Table IIC, AQL 1.5, code letter J. For this single plan n 32,

Ac = 1, Re 4, which leaves a gap. If d ~ 4, of course we reject, or if d ~ 1

we accept. But if d is 2 or 3 we still accept that lot under some suspicion,

because quality has been excellent, but normal sampling is reinstated.

As pointed out in Section 11.1, the OC curves are calculated using the

binomial distribution when both AQL ~ 10.0 and n . 1 < 80. Thus these are
s1ng e -
exact Type B curves for percent defective. In the other cases the Poisson

distribution was used, and they are thus exact for defects per hundred units,

and good approximations for percent defective.

In Section 11.2 it is emphasized that the process average is to be

figured on results from lots on original submission only (not after any

rectifying 100% inspection). Why?

Section 11.4 indicates the inclusion of AOQL factors in Tables VA normal

and VB tightened plans. We do not include the former table, because in order

for the AOQL to become meaningful, p ' would have to be much ~ than the

AQL, and under these conditions tightened inspection would have been quickly

invoked.

An interesting inclusion is that of relative average sample number ASN

curves, Section 11.5. Read all of this Section carefully. The reason that so
312 SOME STANDARD PLANS FOR ATTRIBUTES

few ASN curves as in Table IX may be used to cover all cases is that both
the sequences of sample sizes and AQL's go up in arithmetic progression (very
nearly). Thus we have n's 200, 315, 500, 800, 1250, 2000. Recalling that
~ = 1.5849 = r, say, then 315 ~ 200r, 500 ~ 200r2 , ... ,2000 ~ 200r5
The same is true for the progression of AQL's. This permits the same

acceptance-rejection numbers down a diagonal in tables such as IIA. Then

all of the ASN curves for double and multiple plans corresponding to the

acceptance numbers c = 1, 2, ... , 44, for single sampling will be nearly

exactly alike, if we use a vertical scale in terms of n . 1 , and as horizontal

s1ng e
I
scale np /(100%), that is, the expected number of defectives per single

sample. The corresponding double sample size is (n . 1 )/1.5849 ~ (n . 1 ).631 and

s1ng e s1ng e
multiple sample size is (nsingle)/1.5849 ~ (nsingle).25. For all normal
3

single sampling plans, indexed by c numbers, the AQL is shown on the horizontal

scale by vertical arrows.

Take very special note that these relative ASN curves are only comparable

because they have OC curves very nearly alike. Thus the same power of

discrimination between lots of differing quality is available by single,

double or multiple plans for given AQL and code letter. Hence the relative

ASN's can be soundly compared. Table IX shows that the double sampling plans

have uniformly more favorable ASN curves than the corresponding single

sampling plans, especially so for very good or poor quality levels. Moreover,

except for the c =1 case, the ASN curve for the corresponding multiple
sampling plan lies uniformly below that of the double sampling plan. Of

course such more favorable ASN curves, for comparable protection, are bought

at the price of greater difficulty in teaching and supervision and slower

accumulation of history on a producer.

Another point is that Table IX is for non-truncated inspection, that

is, any sample started is assumed completely inspected, even if an earlier

10.3. The ABC Plan 313

decision could be reached. See Section 11.5. But in double and multiple

inspection, any inspection after completion of the first sample is commonly

ended as soon as a rejection decision is reached, if one is. This has little
I
effect if p ~ AQL or even near it, but has increasing effect as p increases.

Finally we may take note of sample pages 307 and 308 for plans under

code letter J. On page 307 are drawn all of the OC curves, the horizontal

scale being applicable to either defects or defectives. Then Table X-J-1

gives precise values of lot qualities for which Pa = .99, .95, ••• , .01

for each given AQL. This table is entered from the top for normal inspection

or the bottom for tightened inspection. The sarr. e method of entry is used

in Table X-J-2, which gives the acceptance-rejection criteria for all

three kinds of plans, and the AQL's.

10.3.5. Multiple Sampling. Let us now consider multiple sampling which

has been already involved in our description of the Standard. For AQL = 1. 0,

and code letter J, we have n 20, and the following decision criteria:

l:n 20 40 60 80 100 120 140

Ac * 0 0 1 2 3 4

Re 2 3 3 4 4 5 5
*Cannot accept on the first sample of 20

For this plan, if we find two or more defectives among the first 20 we reject

at once. But if there are zero or one, we draw a new sample of 20 (from all

over the lot) and inspect these. Suppose that d 1 was 0, then if d 2 is also 0

we can accept. Similarly if d 1 =0 and d 2 ~ 3 we reject. Otherwise we continue.

Or if d 1 =1 and d 2 ~ 2 we reject, while if d 2 =0 or 1 we continue. Note that

if d 1 =1 the earliest possible acceptance is after 60 more pieces, all of which

are good for then we would meet the Ac number for 80. We observe that after 140

a decision is always reached since there is no integer gap between 4 and 5.

314 SOME STANDARD PLANS FOR ATTRIBUTES

10.3.6.* Calculations for Multiple Sampling. If occasion should arise

for calculating the OC or ASN curve for multiple sampling, the following approach
has been found useful. Consider the plan given in Section 10.3.5 for code letter
J and AQL 1.0%. Suppose we figure out from Poisson or binomial distribution
I
the probability of 0, 1, 2, and so on, defectives in n = 20 for given p ,
p 0, pl' p 2... Then let Pi (j) be the probability of having j defectives after i
samples, without a decision having been made. Thus P1 (O) = p 0, P1 (1) p1 •
These are the only cases wherein the decision is still doubtful, since 2 would give

rejection,
Now we have after the second sample the incomplete cases where d 1 + d 2 1 or 2.

We find the probabilities of such being the case by using P1 (j) and pj:

These are the only possibilities for incomplete cases at the end of the

second sample. After 60 pieces we have

proceeding recurs ely to find P.1 (j) 1 s from P.1- 1 (j) 1 s. The P1. (j) 1 s are then
used for contributions to Pa. The probabilities of acceptance can only occur
after 40, 80, 100, 120, 140 pieces inspected. (Why?) Thus

Also the ASN is obtained by multiplying these five probabilities by 40, 80,
100, 120, 140, for the contribution under acceptance. One must also

calculate the probabilities of rejection at 20, 40, and so forth and obtain
the weighted sum to add to the above. This approach can also be used to find

the ASN under truncation, via a digital computer.

10.3,7. Isolated Lot Protection. Section 11.6 discusses the case of

sampling inspection for an isolated lot. Then we are not in any position to
10.4. What Next? 315

maintain the average outgoing quality and our problem becomes that of making
a decision on this one lot. We want to accept it if p' is sufficiently low

and to reject it if p' is sufficiently high. There will be some gap in these

p''s where it is somewhat immaterial whether we accept or reject and sort;

an indifference band. For such a decision we need to use an appropriate OC

curve. Convenient information is given in Tables VIA and VIB, which provide

the values of p' or c'lOO% for which Pa .10. Thus, choosing the column for

a desired AQL value of p' in Table VIA we can look down it until we find a

sufficiently low plO for our purposes. Then this determines the code letter

and we can use a single, double or multiple plan, as desired.

10.4. What Next? Since its issuance in 1963 the MIL-STD 1050 has been

studied and many possible items of revision considered. In 1967 Stephens and
Larson [6] published an evaluation of the standard as regards its

operating characteristics, treated as ~system. The idea is to take account

of the proportion of the lots for which sampling is on the normal plan, the

tightened plan and on the reduced plan, for a given process quality p'.

Then one works out the "composite" operating characteristic curve and the

composite average sample size curve. This gives insight into the working of

the system chosen (code letter and AQL) when p' is fixed. It is also possible

to let p' vary according to some dynamic model. Two cases were considered:

Case I, using normal, tightened and reduced inspection and Case II, using

only the first two (since reduced is not mandatory). The composite OC curve,

as one would expect, is more discriminating than the OC curves for either the

normal or tightened plans. Moreover the inclusion of reduced inspection

with its very lenient OC curve has virtually no eroding effect on the composite

OC curve.
316 SOME STANDARD PLANS FOR ATTRIBUTES

A distinguished Japanese Group made substantial studies of l05D and

fostered adoption of the standard by the Japanese Standards Association, with
some modifications of switching rules [7,8]. The Group's chief concern was
with a calculated "considerable" switching back and forth between normal and

tightened inspection (and even dropping of inspection altogether) when a

producer is running exactly at p' = AQL. In [8] a letter from the Standards Group
of the Standards Committee of the American Society for Quality Control, made

several points. The first point of difference is in the definition of AQL, where

reference [12] of Chapter 9, defines it as the maximum satisfactory process

average, not as a perfectly satisfactory quality level at which the producer
can run and expect virtually no trouble with going onto tightened inspection.
If running right at the AQL, he should expect to have some such trouble, since
that is natural in a plan designed to provide the consumer with protection
a~ainst p''s somewhat worse than the AQL.

The second point is that much of the calculated frequency of switching

was for single sampling plans with Ac = 1. But for these plans Pa(AQL) = .910,
whereas for those with Ac > 1, Pa(AQL) = .953 and above. For the latter
plans switching while p' = AQL would be much less frequent than if Ac = 1.
The third point is that the proposed use of both the 2 out of 5 criterion
in lOSD Section 8,3,1 ~the earlier (lOSB, lOSC) use of an upper limit top

(p out of control relative to the specified AQL) requiring frequent calculation

of p, makes the tightened inspection criterion quite complicated. This has
the calculable effect of giving the consumer considerably less protection
against p' > AQL, and gives the producer of such quality too much leniency.
Moreover, it would have the practical effect of discouraging the use of

tightened inspection at all, which was often avoided under the earlier rule,

but avoided much less often under the 2 out 5 rule. To require both
criteria to occur before going onto tightened inspection is a substantial
10.4. What Next? 317

stumbling block to the use of tightened inspection. But unless tightened

inspection is invoked when called for, the consumer loses much of his protection.
At the time this is being written the Standards Group of the ASQC

Standards Committee is working toward a revision of the American National

Standards Institute Zl.4 which is currently the same as the 1050. One
objective is to make the standard more.oriented to industrial production of
consumer goods, MIL-STO 1050 being worded more to military procurement. Changes

under study are

1. To make the probability of acceptance at the AQL as nearly uniform

at .95 as feasible for all Ac numbers 1, 2, .•. This would have the
effect, to quite an extent, of removing the objection of the
Japanese Group toward frequency of switching. Currently these Pa's
go from about .91 at Ac = 1, .95 at Ac = 2, uniformly up to about
.995 at Ac = 22.
2. It is difficult to fit Ac = 0 plans into the picture in 1, because
the OC curve for such plans is always concave upwards. Hence in

1050, Pa(AQL) is only about .88. To bolster this Pa,consideration

is being given to using Chain Sampling as an alternative to

Ac =0 plans. See Chapter 13.

3. By making the acceptance number for tightened inspection one less
than that for normal, except when Ac(normal) = 0, one obtains tightened
inspection plans with AOQL's fairly close to the AQL's specified. This
gives desired consumer protection, and at least in some places in
the system, less of a penalty in tightened inspection.

4. Adjustment of sample size in Ac =0 toward lower n's will raise the

Pa(AQL) from .88.

5. Corresponding adjustments can be made for double and multiple sampling

to match the OC curves for single plans.

318 SOME STANDARD PLANS FOR ATTRIBUTES

6. Provision of relative ASN curves for double and multiple sampling

under truncation after the initial samples,when rejection occurs.
7, Elimination of Table VA, AOQL Factors for normal inspection, since
as pointed out earlier, when the level of p' is poor enough to

approach that needed in normal sampling to make the AOQ rise to the
AOQL, the sampling would have quickly gone onto tightened inspection,
Thus AOQL has no meaning for normal inspection (unless one violates
105D completely and never invokes tightened inspection).

10.5. Summary. In this chapter we have discussed two sampling systems

with differing aims. These are collections of sampling plan~with instructions
as to when to go from one plan to another. The Dodge-Romig has single and
double plans to provide specified consumer protection, either AOQL or LTPD,
and to give this protection on a minimum ATI. Thus it is specific protection
on minimum inspection cost.

The ABC-105D system aims to maintain material at a quality level of the

specified AQL or better, being lenient toward quality at this AQL or better,
but tightening up when there is evidence of quality deterioration. Single,
double and multiple plans are provided.

Both plans are consumer oriented. Although they can be used at the

end of the producer's production line, a more natural and desirable approach
for a producer would seem to be process control, for example, control charts.
In this way, assignable causes of bad product may best be sought out and
eliminated and the product made to conform to requirements.
Other approaches are possible. For example, H. C. Hamaker and associates,

published in the Philips Technical Review (Eindhoven, Netherlands) in 1949 and

1950 three papers leading to a simple sampling system, indexed on the point of

indifference, p~ 0 . The plans were developed to equalize the producer and

consumer risks in a sense by keeping nearly constant
Problems 319

I
P5o

Then the producer's risk PR decreases from 50% as p' decreases from p~ 0 ,

at about same rate as the consumer's risk CR decreases from 50% as p'

increases from P~o·

PROBLEMS

Find Dodge-Romig sampling plans for the following conditions and

list the amount of the opposite kind of consumer protection:
10.1. Double sampling, AOQL=l.O%, N=3500, p=.S%, defectives

10.2. Double sampling, AOQL=l.O%, N=600, p=.9%, defectives

10.3. Single sampling, AOQL=l.O%, N=3500, p=.5%, defectives

10.4. Single sampling, AOQL=l.O%, N=600, p=.9%, defectives

10.5. Double sampling, LTPD=4.0%, N=3500, p=.S% defectives

10.6. Double sampling, LTPD=4.0%, N=600, p=.9% defectives

10.7. Find a few points to check the AOQL in 10.3, Type B OC curve.
10.8. Find a few points to check the AOQL in 10.4, Type B OC curve.
10.9. Check Pa at p'=4.0% in 10.5, Type B OC curve.

10.10. Check Pa at p'=4.0% in 10.6, Type B OC curve.

10.11. In Tables 10.1 and 10.3, for a fixed lot size, what happens to the
pt protection as p- increases? Why does this occur? Similarly what
happens to pt as p- remains fixed and N increases? Why does this
occur?

10.12. In Table 10.2, for a fixed p- and increasing lot size, what happens

to the AOQL? Why does this occur?

10.13. Comparing Tables 10.1 and 10.3, one sees that in the AOQL table the

p classes go up to the AOQL, whereas in the LTPD table the p classes

only go up to pt/2. Why does this make sense?
320 SOME STANDARD PLANS FOR ATTRIBUTES

10.14. In the Dodge-Romig double sampling plans Re 1 = c 2+1 and Re 2 = c 2+I

also. This sometimes leaves quite a gap between Ac 1=c 1 and Re 1 .
What are the implications of this, without and with truncation of
inspection under rejection on the second sample?

Find the MIL-STD105D sampling plans for the following sets of conditions,

giving cumulative sample sizes and corresponding acceptance and rejection

numbers. Also give approximate LTPD at Pa=lO% for normal plans, and, for

tightened plans, the AOQL:

Lot size Inspection Number of Severity AQL

N Level samples

10.15. 1000 II Single Normal .65

10.16. 1000 II Single Tightened .65

10.17. 1000 II Single Reduced .65

10.18. 1000 II Double Normal .65

10.19. 1000 II Multiple Normal .65

10.20. 1000 II Multiple Tightened .65
10.21. 1000 r Single Normal .65
10.22. 400 II Double Normal .10

10.23. 400 II Multiple Normal .10

10.24. 400 II Double Tightened .65

10.25. 4000 II Double Normal 40.

10.26. 4000 II Multiple Normal 40.

10.27. 25 II Multiple Normal 15.

10.28. Which of the plans in 10.15-10.27 are available for defectives?

Which for defects?

10.29. On the same axes show the OC curves for 10.15, 10.16, 10.17. Comment

10.30. On the same axes show the OC curves for 10.15, plus those under
the same conditions, except N=200, N=lO,OOO.
Problems 321

10.31. For 10.16, find the AOQL and compare it with the plan's AQL. What
is the point of this comparison?

10.32. Why is reduced inspection an optional feature of 105D, whereas

tightened inspection is an obligatory feature?

10.33. Why can such relatively lenient OC curves be permitted under

reduced inspection?

10.34. Why does Section 11.2 stipulate the omission of results from

resubmitted lots when figuring process averages?

10.35. How would a control chart be used for acceptance sampling results

under 105D, and how would it be useful?

10.36. What are some of the important considerations in choosing between

single, double and multiple sampling?

10.37. The AQL's in 105D are usable for defectives, only up to 10.0,

while for defects, all the way up to 1000. Why is it desirable to

have some such cut-off?

10.38. Why do you think that the single sample size n is not directly

proportional to the lot size N in either Dodge-Romig or 105D?

10.39. Find Pa for 10.15 and 10.18 at p 1 =.02=2%, and at .04 or 4%.

10.40. Find Pa for 10.19 (multiple) at p'=.02. Use Section 10.3.6.

10.41. Suppose that c'=.lO defect per unit, that is, 10 defects per 100

units. Then in a sample of 100 units we expect 10 defects on the

average. How many defectives can we expect in the 100? For this,

use Table VII to find q'=P(O defectslc'=.lO) then p' and 100 p'.

Do similarly for c' .40 per unit and c' = 1.00 per unit? Does

this help on 10.37?

10.42. In the double sampling Dodge-Romig table for AOQL=l.O% find a

plan having pt=4.0%. Compare this plan with the corresponding

one in the double sampling table having pt=4.0%. (p classes do not

correspond exactly.)
322 SOME STANDARD PLANS FOR ATTRIBUTES

References

1. Statistical Research Group, Columbia University, "Sampling Inspection."

McGraw-Hill, New York, 1948.

2. H. F. Dodge, Notes on the evolution of acceptance sampling plans, Parts I to

IV. ~· Quality Techn., 1, 77-88, 155-162, 225-232 and 2, 1-8, 19 (1969,19 70) .
3. H. F. Dodge and H. G. Romig, "Sampling Inspection Tables." Wiley, New York,
1959.

4. W. R. Pabst, Jr., MIL-STD 105D. Indust. Quality Control, 20 (No. 5), 4-9

(1963).

5. H. F. Dodge, A general procedure for sampling inspection by attributes -

based on the AQL concept. Technical Report No. 10, The Statistics Center,
Rutgers the State University, New Brunswick, New Jersey (1959).

6. K. S. Stephens and K. E. Larson, An evaluation of the MIL-STD-105D system of

sampling plans. Indust. Quality Control, 23 (No. 7) 311-319 (1967).

7. T. Koyama, Y. Ohmae, R. Suga, T. Yamamoto and T. Yokoh, assisted by W. R. Pabst, Jr. ,

MIL-STD-105D and the Japanese Modified Standard. J. Quality Techn., 2, 99-108

(1970).
8. Review and letters on reference 7 and letters relative to JIS Z 9015, Sampling

Inspection Plans by Attributes with Severity Adjustment. J. Quality Techn.,

3, 87-94 (1971).
 CHAPTER 11

ACCEPTANCE SAMPLING BY MEASUREMENTS

11.1. Potential Greater Efficiency by Variables than by Attributes.

In general the measurement for any characteristic for a part or an assembly

can be converted to an attribute. This is accomplished by choosing

specification limits for the characteristic. Then the part or assembly is

a "good" or a "defective" according to whether the measurement lies

between the limits or outside. For example, one often sees a green sector

on a dial gage, such that if the pointer lies in this sector the part is

a good one, otherwise it is a defective. *

The reverse is not true, since there are many attributes or defects

which a part or assembly can have, to which we can give no measurement.

In this case we must use "inspection by attributes" to describe lot or

process quality performance.

When either type of inspection may be used, we have a choice to

make. In general a measurement provides more information about the part

or assembly, than does a mere determination that it lies within or outside

limits. It is thus often possible to obtain the same discriminating

power on a smaller sample size n, when using ·measurements than when using

attributes. But, in order to make full use of measurements and thus obtain

this greater efficiency we commonly need to know the ~ of distribution

of the individual x's, and the standard deviation ox. Somewhat less

power is still obtainable if ox is unknom1. But there are other factors

which may sway the balance in favor of the method of attributes, namely

*We may also have a band just outside of L to U, such that x's within are
minor defectives, and those further out major defectives.

323
324 ACCEPTANCE SAMPLING BY MEASUREMENTS

(a) the cost of a single measurement vs. that of finding whether it is

in limits or not, (b) the cost of recording, and (c) the cost of analysis.

So use of a larger n with attributes might still possibly be cheaper.

11.2. Dependence Upon Knowledge of Distribution of Individual xls,

If our objective in sampling a lot is to determine whether or not the

percentage outside of specification limits is excessive, then we must

know what~ of distribution the x's have. Is it normal or is it

somewhat unsymmetrical or skewed? Is it one homogeneous distribution,

even if not normal, or is it a composite of several distributions? The

latter often occurs when a process has been let run at one setting for a

while then reset at a new level, then perhaps at a third level, etc.

And then too, there is the question as to whether the distribution of x's

has been multilated by 100% sorting. This often chops off one or both

natural tails of the distribution if it original;y lay outside of the

limits. Sometimes this is quite a drastic truncation. And there is even

the case where we receive the outer tails of a distribution, the central

part having been sorted out and sent to another customer with tighter

limits than ours!

Unless we can build up evidence on the distribution of measurements,

we had better stay with the method of attributes.

On the other hand if we can be reasonably sure of an approximately

normal distribution of x's then we are in a strong position, even if we

do not know ax. Our position is further enhanced if ax is known.

11.3. Use of Control Charts and Frequency Distributions for Past

Data. When a series of lots is being inspected, x and R or s charts form

an invaluable record. They provide a picture of the producer's quality

level and stability of performance. And, as we have seen, comparison can

be made to specifications. Frequency tabulation of the x's obtained should

also be made. If performance is consistently satisfactory we may well

11.4. Decision-Making on Lot Means 325

consider reducing the sample size per lot. Use may also be made of process
control chart data obtained by the producer, if he is willing to send them

along.

However, when lots are only sent from time to time our past records

are helpful, but the process may well have changed in the interval between

lots, and larger samples with a frequency distribution may be called for.

We shall now study what can be done in acceptance sampling by

measurements, when we can assume a normal distribution of x 1 s.

11.4. Decision-Making on Lot Means, Known ox. We assume that the

x's are normally distributed. Also in this chapter we shall only consider

single sampling. "Sequential sampling", in which the maximum sample size

is not fixed in advance, is taken up in the next chapter.

The present problem is basically significance testing on the mean,

with a known. We may have one-way protection (e.g. high ~ is good, low ~
X

is bad), or two-way protection. There are also two approaches: (a) set n

and an a risk of erroneous rejection if 11 is actually "good" and find the

criterion of decision, and sketch an OC curve, and (b) set an a risk as

above and a i3 risk of erroneous acceptance if ~ is actually at some "bad"

level and find n and the acceptance-rejection criterion. We shall adopt the

latter approach as being perhaps a more useful and direct approach in

industry. If it leads to too large an n,modifications can be made.

11.4.1. One-way Protection on the Mean, Known ox. Let us consider

an example to illustrate this method. For tensile strength of a certain

casting, ox was known to be about 3500 pounds per square inch. The lower

specification was 58,000 psi. Let us set a safe lot mean and an unsafe one.

These depend upon the importance of the specification limit. Suppose that

2% of the castings lying below 58,000 psi would be undesirable. Then we

might take p 1 = 58,000 + 2.05(3500) ; 65,000 as an undesirable lot mean.

On the other hand if only .2% are below, this might be called fully
326 ACCEPTANCE SAMPLING BY MEASUREMENTS

satisfactory. So p2 = 58,000 + 2.88(3500) ~ 68,000. In a sense we still

have said nothing until we decide upon respective risks. Again this
will depend upon engineering considerations. How likely is it that a lot

comes along with p around 58,000, etc? Suppose that we take the following:

P(reji~J 2 ) =a .10 (11.1)

We are thus specifying two points on the operating characteristic

curve. See Figure ll.l(a). Now the problem is to find a sample size nand
a critical value, say, K such that if x> K we accept, or if x< K we

reject. Thus

(11.2)
We need n large enough so as to cut down a-X = aX;10 sufficiently that we can
have the situation shown in Figure ll.l(b).
Suppose we define za for the normal curve (2.25) by

rz
a
.p(v)dv =a (11.3)
that is, za cuts off a tail area of a. Then we have from (2.24) and
(2.27)

--= za and -z a • (11.4)

aXI Iii a
X
!Iii

Substituting in what we know, gives equations

K-65,000 +1. 645 K-68,000

-1.282.
3500/10 3500/10

Subtracting the second from the first gives

300010
3550 = 2.927 10 = 3.415 n = 11.7 or 12.

If a sample size of 12 tensile tests seems reasonable in view of the size

of a lot and the risks involved we can proceed to find K. But if n = 12
appears excessive, then we must increase one or both risks, and/or widen

the gap between p1 and p2 . Either will give a less discriminating or

"lazier" OC curve. Suppose that n = 12 is deemed economically feasible.
11.4. Decision-Making on Lot Means 327

( 0)

65000 K 68000

65000 K X

K 68000

FIG, 11.1. One-way protection on mean. OC curve shown in (a), two x

distributions in (b).

Then we may find K from either equation. Substituting into the first will

preserve S and slightly decrease a, and vice versa, if we substitute into

the second. Let us preserve S.

K = 651 000 + 1.645(3500)/ill = 66,662.

Then z = (66, 662 - 68,000) I (3500/ m) = -1. 324
which gives a = .0928.
Thus the plan is:

Take 12 tensile tests and find x

Then x< 66,662 reject lot, x> 66,662 accept lot*.

*The only advantage to retaining so much precision on K is to make x- K

unlikely.
328 ACCEPTANCE SAMPLING BY MEASUREMENTS

The OC curve may be easily sketched by using the above two Pa's and the
third point

P(acclp=66,662) = .SO.

If desired we may easily find a few more points on the OC curve by using

different p's and proceeding as for the actual a risk.

Summarizing the procedure we have

One-way Protection on Mean, ax Known

1. Given AQL = p2 > RQL = p1 , and risks P(accJp 1) a.

2. Find sample size n, rounding up to an integer

n = [ax(za + zs)J2 (11.5)

J.l2 - Jll
3. Using the integer n, find K by either

K = p 2 - ( z aa/ In) or K = J.ll + ( z 6a/ In) (11.6)

4. Plan: x < K reject, x > K accept

n n-
5. If AQL = J.ll < RQL = p2 , that is, J.ll is still the lower one, but

now low p's are preferred, then use a at J.l 1 , Sat p 2 . (11.5)

still gives n. But for K use

or K = JJl + (z a /vn), ( 11. 7)

ax

The above procedure is quite satisfactory even if we do ~ have a

normal distribution but it i§ homogeneous, perhaps skewed, provided we

can settle on acceptable and rejectable p values. This is because

x's are quite normally distributed even if the x's are not, in line

with the central li~it theorem. This is also justified by (2.26) and

(2.27) still holding under non-normality.

11.4.2. Two-way Protection on the Mean, Known ax. Let us illustrate

this case with an actual example, which saved considerable money. The

application- was to the doubled thickness of rubber gaskets, cut from

tubes which were extruded continuously and cut off at 30" lengths and
 11.4. Decision-Making on Lot Means 329

cured. The gaskets, somewhat resembling rubber bands, were for caps

for food jars. Standard practice had been to measure the doubled

thickness of each tube in about five places, form a mental average

and then put the tube into one of three piles. The three piles of tubes

were supposed to yield gaskets when cut running (a) .096" to .100",

(b) .100" to .106" and (c) .106" to .110". But because of the considerable

variability within a single tube due to the curing, there was large over-

lapping. It was finally decided to ask the producers to run an entire

lot at one process setting. If they thought that it was running too far

off from the desired level, then to stop that lot, and after re-setting

to begin a new lot number. Then a decision as to which specification to

use the rubber for would be made on the basis of a sample of tubes.

This program brought three benefits: (a) the distribution of gaskets

became normal, (b) the standard deviation in a lot was considerably

decreased, and it was known to be .002", and (c) far less inspection time

was needed.

Now let us see the statistics involved. This was a case of o

X

known, namely . 002". To test for specification (b) the AQL = p 0 was

set at .103". Then the alternative lot averagelor RQL's were set at

.098" and .108", that is, symmetrically placed around .103". Taking an

a risk of .001 and a Brisk of .002, what would be the plan? It is of

the form

x between . 1 03" - k and . 1 03" + k, accept for (b)

x below . 103" - k, accept for (a)

x above .103" + k, accept for (c).

(There was no trouble with rubber too thin for (a) nor too thick for (c).)

Let us therefore seek n and k. Four equations could be set up, but

owing to symmetry, only two are independent:

330 ACCEPTANCE SAMPLING BY MEASUREMENTS

.103" + k- .103" .103" + k- .108"

. 002"/lrl
z.ooos . oo2"1 rn
-z.002

Note particularly that a =.001 is split two ways, because we can reject
for (b) usage by either too high or low an x. Also notice the sign of

z.002' We then have

klrl Ck - • oo5") rn _
. 002" = 3.30 .002" - -2.88

giving by subtraction
. 005"/ri 6.18 rn =
. 002" = 2.47 n = 7.

Then using the second equation we find k .005"- 2.88(.002")/17 = .00282"

Thus for samples of 7 tubes
x- between .10018" and .10582" accept for (b)
x- below .10018" accept for (a)
x above .10582" accept for (c).

Actually the company decided to make the decision for lots of several
hundred tubes by a random sample of 20 or 25 tubes which is being very

conservative, but still brought a large saving.

Let us summarize as follows:

Two-way Protection on the Mean, Known crx

1. Set an AQL = p0 and two symmetrically placed RQL's on opposite

sides of ~0, say, p1, p2 so that p0 - p1 = p2 - Po·

2. Set risks P(rejjp 0) =a and P(accjp 1) = P(accjp 2) S

llo + k - )Jo llo + k - P2

3. Solving
cr
X
;rn cr X I lrl

gives

n = ( 11. 8)
11.5. Percent Beyond Specification, Known ax 331

4. Using the integer value of n, find k by

(11.9)

5. Then for x for n x's:

x between ~O - k, ~O + k accept

Otherwise reject.

Again, the normality assumption is not too important as long as a is

X

known.

11.5. Decision Making on Percent Beyond Specification, Known ax. This

problem is very straight forward for the one-way protection case. Suppose

first that we are interested in an upper specification U. Then in terms of

fraction defective p 1 , let

AQL = Pi < Pz = RQL

and the respective risks be a and S. The assumption of normality now

becomes much more crucial, because we are using x distributions, not

x distributions, for the set-up. Now we set ~ 1 so as to have fraction

defective Pi beyond U and ~2 so as to have Pz beyond U. Then

j.l
1
=u - a z
x pi ~2 u- a Z I • (11.10)
X p2

We may now substitute these into (11.5) obtaining

(11.11)

Then forK some care must be exercised. In (11.10) ~l is the AQL and

p2 the RQL. To reach K we must go ~ from ~1 by an amount zaax/(11.

This gives

K=U-a(z -z/lri\ (11.12)

x p 11 a

Or we must go down from p2 by an amount z 6ax/ln:

K =U- a (Z 1 + Z0 / fri\ (11.13)

X p2 "
332 ACCEPTANCE SAMPLING BY MEASUREMENTS

At a lower limit L, formula (11.11) still follows, but now

K = L + o (z
x Pi
- z I vn) or K=L + o (z , + Z0 ) (11.14)
a X p2 "

Probably the best approach is not to use these formulas (except as

a check) and instead to think through the distributions of x and x involved

and perform the necessary arithmetic.

The student should compare the formulas (11.12) and (11.13) with

those for modified limits (7.9), (7.10).

For two-way protection on the mean, the first step is to see whether

ox is sufficiently small to permit as little as Pi outside when ~ = (U+L)/2,

that is, perfectly centered. Thus

U - (U+L)/2 > o z , 12
- X pl

or

(11.15)

Now one possibility is to run a one-way test at whichever specification

x is nearest. This approach is recommended if ox is considerably less

than the maximum given by the left side of (11.15), for then the lot mean

~ has some latitude to move around in, and x must be some little distance

away from the nominal (U+L)/2 toward, say U, before we might reject. But

then the fraction defective below L will surely be negligible. So we

can use P} (not Pi/2) and P2 in (11.11) and (11.12).

If, however, ox is close to the maximum given in (11.15), then the

method given in the Military Standard 414 [1) seems preferable. There,

the percent beyond U and that beyond L are estimated using x, and the

sum of these estimated percents compared with a predetermined maximum

allowable estimate. If below, the lot is accepted, if above, it is

rejected.

11.6. Decision Making on Percent Beyond a Specification, Unknown ox.

This case is a bit more complicated than that when crx is known, and also
11.6. Percent Beyond a Specification, Unknown ax 333

places more demands upon the assumption of normality. We shall follow the

method given in (2].

The test is quite similar to that given in Section 11.5 for one specifi-

cation limit. There, for example at an upper specification U, we reject if

X > U-a (z
X
I
pl
- z
Cl
Jrn) or X + a (z
X
I
pl
- z
C1
Jrn) > u'

Now we do not know ax in the present case, so, taking an analogy of the latter

inequality we use the criterion

x + ks > U reject; x + ks S. U accept,

The latter means that we are apparently safely below U. The problem then is

to find n and k from Pi• e~; Pz• ~. We shall shortly derive the formulas

(11.16) and (11.17) below. But first we give the general plan and illustrate

with an example.

Percent Beyond One Specification, Test on x, Unknown ax

1. Case of protection against an upper specification U

a. Set fraction defectives relative to U: AQL = p 1 < p 1 RQL.

1 2
and respective risks Cl = P(rej \Pi> and ~ -.. P(acc \Pz)
b. Find the two constants k and n determining the plan:

Z Z I + ZQ Z I
Cl P2 .., Pl
k (11.16)
z + z
Cl ~

k2 + 2 (11.17)
n "" 2

If n is non-integral, round~ to the next integer [3].

c. Then the plan is to take n measurements, finding x and s:

x+ ks ~ U accept (11.18)

x + ks > U reject (11.19)

334 ACCEPTANCE SAMPLING BY MEASUREMENTS

2. Case of pvotection against a Lower specifiaat:Lon L

a. Identical to Section la

b. Use (11.16) and (11.17) just as in 1b

c. Then the plan is to take n measurements, finding x and s:

x - ks > L accept (11.20)

-
x - ks < L reject (11.21)

Comparing (11.17) with (11.11), it is easily seen that the ratio of the

required n's, with a unknown and known is (k 2 + 2)/2, which is always above
X

1, and increasesas we try to discriminate between smaller pl and p; fractions.

Example 11.6. Suppose that the maximum time of blow of fuses is 150

seconds under a certain marginal current, and that 1% is considered sat-

isfactory, but 6% not. Call these Pi and Pz· Let us also take corresponding

risks a= .OS and 8 = .10. Then we have from Table I,

1. 645, z~ 1.282

Thus
1.645(1.555) + 1.282(2.326)
k 1.893
1. 645 + 1. 282

1.893 2 + 2 • ( 1.645 + 1.282 )~ 2

n -=
2 ~ 2.326 1.555 40.23 or 41

Since the test is destructive, this is likely to be too many to be practical,

so some compromise is needed. If ox were known, we could use (11.11), giving

n ' 15. Also Table 9.5 would give Ac = 3, npz = 6.68 yielding n = 112 for a
pure attribute plan, that is, blowing time below or above 150 seconds.

11.6.1"'· Derivation of (11.16), (11.17). The following facts are

available in most mathematical statistics texts:

1. If the population of x's is normal, then x and s are independent and

a. x has mean ~and standard deviation a- =a ;);, and the x's are
X X
normally distributed.
11.6. Percent Beyond a Specification, Unknown crx 335

b. s has mean approximately ax (actually c 4 ax) and standard deviation

as approximately ax//2(n-1) (actually c 5 a), and the s's are fairly

normally distributed, increasingly so as n increases.

2. If two variables u and v are independent, then ~au+bv = a~u + b~v

2
and aau + bv a 2·a 2 + b 2·a 2 and au+ bv is more normal than the
u v
less normal of au and bv.

Now consider the variable

w : :it + ks.
From 1 and 2 above we have that

1. w has mean approximately~+ kax

2. w has standard deviation

-a
ax+ks- x
. /1 k2 -
n
+2(n-l)
---

3. w is approximately normally distributed.

Let us consider the problem of distinguishing between the two critical

means of x 1 s (givenfraction defectives Pi• Pz)

lJ. = U- z 1 a = AQL and [.1. 2 = U- zp 1 a RQL.
1 pl X 2 X

These give respective means for the variable w ~ x + ks of

U- z 1 a + k a "' U - a (z 1 - k) lJ. ( ) U- a (z
x 1
p2
k).
pl X X X pl ' W 2

Recalling that U is the cut-off point for the criterion of acceptance or

rejection, we have in direct analogy to (11.4):

U- [u- a (z 1 - k)]
X pl
z (11.22)
0!

U- [U- a (z
X P2 - k)]
(11. 23)
- Z~•

1 k2
-
n
+2(n-l)
---
336 ACCEPTANCE SAMPLING BY MEASUREMENTS

Simplifying and further approximating by changing the n-1 to n gives

z - k
p'
1
z (11.24)
a

(11. 25)

Dividing (11.24) by (11.25) gives

z
_a,' (11.26)
Z t - k z~
P2

Solving (11.26) fork gives (11.16). On the other hand, if we subtract

(11.25) from (11.24) we have

which when solved for n gives (11.17).

11.7. Single Sampling for Variability. A moderately common problem

is to test the variability of a process or measurement technique. In the

former case we are concerned with process capability, while in the latter

case the interest is in measurement error. Typically low population standard

deviations ax are good and higher ax poor. Accordingly a one-way test is

needed.

We may set up the test by deciding upon

AQL = cr 1 < a 2 ~ RQL

We might proceed in either of two ways:

1. Decide upon risks of wrong decisions

and then find n and the acceptance - rejection criterion.

11.7. Single-Sampling for Variability 337

2. Decide upon the sample size nand one of the two risks, then find the
acceptance - rejection criterion and evaluate the other risk.

The former approach seems particularly natural and by use of Table VIII,

proves to be very easy. In either approach assuming the x's are normally

distributed, s 2 is the best test statistic and we would use the criteria

accept, reject, (11.27)

Also in either case it is wise to at least sketch the OC curve.

In order to use Table VII~ we need to use a= S set at one of .10, .05,

.02 or .01. If one should desire unequal a, S or other risk values, this

can be done, but Table VIII, is not applicable.

Example 11.7. Suppose that we are concerned with the consistency in

muzzle velocity of ammunition. cr 1 = 30 ft/sec is regarded as satisfactory

and crz = 60 ft/sec as unsatisfactory. (The test is unooneerned with averag~ muzzle

velocity.) Let us settle upon risks a =B= .OS. Then we first find a 2;a 1 = 2.00.
Now entering Table VIII, in the left side at the .05 column, we seek 2.00.

We find for n = 13, 2.01 and for n = 14, 1.95. So we take n = 14 (unless

we permit the risks to slightly exceed .OS). Then looking at the right side

of the table under .OS, opposite n = 14 we find the multiplier for cr~ to give

K. It is 1.72, giving

K = 1.72•(30 ft/sec) 2 = 1548 (ft/sec) 2

Then this K value is used to make the decision as in (11.27).

Thus if lots have a = 30 ft/sec, they will be accepted 95% of the time,
X

but if 60 ft/sec they will be rejected 95% of the time. Probabilities for

other a's appear on the OC curve.

Protection Against Excessive Variablity

1. Assume normal population

2. Choose an acceptable lot or process standard deviation cr 1 , and a re-

jectable standard deviation cr2·

338 ACCEPTANCE SAMPLING BY MEASUREMENTS

3. Decide upon the risk a that a lot or process having a = a 1 will be erron-

eously rejected, and a risk S that one with a = a 2 will be erroneously

accepted. If these are made equal and at one of .10, .OS, .02 or .01,

Table VIII can be used.

4. To use the table, form a 2 !a 1 and in whichever column (2) to (5) corres-

ponds to a = S, seek the desired ratio. If it does not occur, take the

next lower entry (interpolating if n > 31). This gives n.

5. Now find K by finding the entry in the appropriate column (6) to (9)
2
opposite n. Multiply this entry by a 1 , for K.
2 .
6. Then plan is s ~ K accept lot or process, otherwise reject.

7. This plan will match a, but slightly cut B.

8. An economic balance should be struck between sample size and the dis-

criminating power of the test. This is easily done by using the table.

9. An OC curve can be sketched as below.

OC Cu1·ve. The probabilities for other a 1 s than a 1 and a 2 are easily found

as follows. A standard theorem from mathematical statistics is that if a

random sample of n x's is chosen from a normal population having standard

deviation a, then

2
(n-l)s
(11. 28)
2
a
follows the chi-square distribution with n-1 degrees of freedom. Now from

(11.27)

Pa = P(s 2 :S K\a). (11.29)

In order to build up (11.28) within the parenthesis multiply both sides of

the inequality by (n-l)/a 2 , obtaining

2 2 2
Pa = P[(n-l)s ja ~ (n-l)Kia ].
11.7. Single Sampling for Variability 339

But by (11.28) we have

2
Pa = P[Xn-l ::; (n-l)K/a 2] . (11.30)

Hence

2 2
X(Pa below) (n-l)K/a , (11.31)

To use (11.31), we might substitute desired a's and then try to find Pa by

interpolation in Table II, at the row corresponding to the degrees of

2
freedom. But this is a nuisance. Instead let us find X(Pa below) for

values of Pa available in the table. Then

2 (n-1) • K
(j
2 (11.32)
X(Pa below)

Thus in Example 11.7, n ~ 14, df 13 and we have

2 13 (1548)
cr
2
X(Pa below)

From Table II, with df 13 we have

2 2
Pa .25 x.25 9.299 (j 2164 (ft/ sec) 2 cr = 46.5 ft/sec
2 2
Pa . 75 X. 75 15.984 cr 1259 (ft/sec) 2 a = 35.5 ft/sec

These supplement the cr 1 , cr 2 points. The full curve is shown in Figure 11.2.

11.7.1*. Derivation of Method. Making use of (11.28) we have the

following two equations for the plan which is to have (cr 1 , Pa = 1 - a),

(cr 2 , Pa = B).

1 - 0! (11.33)

2
Now as before, form Xn-l on the left sides:
 340 ACCEPTANCE SAMPLING BY MEASUREMENTS

Pa 0.50

20 40 60
CT It/sec

2 2
FIG. 11.2. OC curve for plan n=l4, s .::_ 1548(ft/sec) accept, otherwise reject.

2
P[(n-1)s 2 /cri :S (n -1) K/ cr 1 J 1 - 0! (11.34)
2 2
P[(n-1)s /cr 2 :S (n-l)K/cr~] ~. (11.35)

Thus we have

xn-1
2
df, 1-0! be 1ow
(n-1)K/cr 12 (11. 36)

X2n- 1 d f, ;.;o b e 1ow (n-1)K/cr22 (11.37)

Now we would like to solve for n and K, but this cannot be done directly,

and of course again we shall go to an integer n for sample size. Dividing

(11.37) into (11.36) and taking the square root yields

x2
n-1 df, l-0! below (11. 38)
x2
n-1 df, ~ below

This gives the entries in columns (2) to (5) in Table VIII, tabulated against

n. Risks 0! and ~ were taken equal in order to cut down the multiplicity

of cases. Then we solve (11.36) forK giving

x2
n-1 df, 1-0! below 2
K = n-1 •crl, (11. 39)
11.8. Military Standard 414 Plan 341
2
The multiplier of cr 1 is what is given in columns (6) to (9) in Table VIII ,

tabulated against n.

11.8. Military Standard 414 Plan. We will now describe MIL-STD 414,

[1], which is a collection of sampling plans to use measurement

statistics, such as, x, to control the percentage of pieces beyond

specifications. It is thus concerned with the types of problems we have

considered in Sections 11.5 and 11.6. As this is being written there is

considerable discussion of Standard 414 in many quarters around the world.

Because of this we shall content ourselves with a description of the plan.

Many cases can be handled adequately by the methods already discussed here.

Early work on such techniques was done by Romig [4]. Military

Standard 414 is an outgrowth of [2], [3], [5]- [7]. All of these

plans, whenever they are concerned with the percentage beyond specification

limits, make strong use of the assumption of normality, because lack of

normality tends to affect the tails of distributions most strongly. The

normality assumption is often insufficiently emphasized in practice.

Now let us proceed with a description of the 414 plans.

A. All are single sampling plans; making decisions on the basis of

n measurements, rather than upon counts of the number outside of

specifications. All assume the distribution of measurements x

in the lot to be normal.

B. The plans are nevertheless concerned with controlling the

percentage outside of limits L to U, or L alone, or U alone.

These are two-way or one-way plans.

C. Plans are available for two general cases

1. a known. Section D, using a.

2. a unknown. Section B, using sample s, or Section C using

sample ranges, Rts.

342 ACCEPTANCE SAMPLING BY MEASUREMENTS

D. Plans are to control fraction defective. They include normal,

tightened and reduced plans.
E. There are two forms, 1 and 2 (unfortunately, for it makes the

standard more bulky and complicated). Decisions on a lot would

be the same under either form. Since Form 2 is used whenever a

two-way decision (both L and U given) is made, we here

recommend using Form 2 throughout, for one-way decisions too.

F. OC curves for Pa vs. lot fraction defective are provided for

normal and tightened inspection plans. They are in Table A-3

of the Standard, a page for each sample size code letter, listed

for normal inspection; but also correct for tightened inspection

by finding what normal inspection plan corresponds to the

desired tightened plan.

G. To start to find a plan, specify an AQL (acceptable quality level)

in percent defective. Table A-1 converts this to one of the index

AQL 1 s, for example, .65%. Then for inspection levels I to V

(smallest to largest n's), and lot size N, Table A-2 gives a

sample size code letter B, C, ... , Q. Use level IV if none is

specified. Thus we have a code letter and an AQL in ~ ~·

H. Now choose Section B, C or D according to desire and knowledge

as in our Section C above.

I. Form 2. Using sample standard deviation s. Section B.

1. One-way, say, U

a. For the AQL and code letter from G, find in Table B-3

the required sample size n. Then for a random sample

of n measurements x, find x and s. Also note the

maximum allowed estimated percent defective M in

Table B-3. (This M will be larger than the AQL, so that

lots with p' = AQL will have a high Pa,)

11.8. Military Standard 414 Plan 343

b. Form the "quality index"

QU = (U- x)/s- (11.40)

(This is much like (2.24), z = (x-~)/a for obtaining

the percentage beyond x, but uses instead the sample

x and s.) From Q0 and n, enter Table B-5 to find estimated

percentage above U, say Pu·

c. Decision:

Pu _:::. M accept Pu > M reject, (11.41)

d. If protection is relative to a lower specification L,

use Section a as it is, but in Section b use

QL = ex- L)/s. (11. 42)

Then use Table B-5 as above to find pL and make

decision by

pL _:::. M accept pL > M reject. (11.43)

e. Tightened plans are found in Table B-3 by entering from

the AQL's at the bottom rather than those at the top as in

normal plans. Reduced plans are found in Table B-4.

2. Two-way, L and U both given

a. Same as Section la.

b. Form QL by (11.42) and QU by (11.40) and use Table B-5

to find estimates pL and Pu·

c. Decision
Pu + pL _:::. M accept lot Pu + PL > M reject lot (11.44)

d. Same as 1e.

3. Switching rules are based in considerable part upon the estimated

fraction defective from a series of lots, that is, upon

p 0 , or pL' or Pu + pL' usually for ten lots. The rules are

rather similar to those in MIL-STD 105D.

344 ACCEPTANCE SAMPLING BY MEASUREME~l'TS

J. Form 2. Using sample ranges, R. Section C.

1. If n 3,4,5 or 7, useR. Or if n = 10,15, etc. break
up the sample into subsamples of 5 each, finding the
range for each and use R.
2. Proceed as in II and I2, except using R or R and Tables C-3
to C-5, and

QU = c(U - x)/R, QL = c(x - L)/R. (11. 45)

which provide estimates Pu and p 1 , from Table C-5. Tables

C-3 and C-4 provide the c values for each n; they are like d 2

values for control charts.

K. Form 2. Using known population cr. Section D.

1. Steps very similar to Section I, but use Tables D-3 to

D-5, the first two providing n, M and a v-quantity. Then use

QU = v(U - x)/a Q1 = v(x - L)/cr. (11.46)

and find estimates p 1 and Pu from Table D-5.

L. If there is real doubt as to the normality of the distribution, then

one possibility is to accept only by MIL-STD 414. But if the
variables approach 414 would call for rejection, then do not

reject yet, but continue onward, counting the number of pieces

outside of specifications, until the n for an appropriate

attribute plan is completed. Such a plan might be from MIL~STD-1050.

This approach is called "variables-attributes sampling.'' It particular!:

protects a producer who may have a process running outside of

L to U, but has sorted his product carefully to these limits.

M. If the shape of the distribution of x's is not normal, but the

curve shape is known, then one might use [8]. See also [9]

and [10]. If the long tail in a skewed curve is toward U,one

can use a smaller n than for the normal, but if the short tail is
toward U a larger n is needed.
11.9. Checking a Process Setting 345

11.9. Checking a Process Setting and a Process Capability. In

general it is recommended that process control be maintained by measurement

control charts, x and R or s. This is because they supply the dual role

of helping find assignable causes thereby improving and stabilizing the

process, and secondly of telling when to adjust the process, for example

resetting the level. If the object is only that of periodically

checking the process level p, in relation to specifications, and ax

is known from past records, then we may wish to make a one-way or a

two-way check of the process level p.

Two-way Check of Level. Suppose that the specifications are rather

tight so that the tolerance T = U - L is only six or seven axts and we must

therefore maintain p close to the nominal (L+U)/2 = p 0 , say. Then a good,

practical plan is the following one, with risks below 10%:

"Safe process level" )1 0 : P(approval) = .942 ~ .95

"Unsafe process levels" Jloflax: P(rej) = .897;; .90

Take n = 10, and find x. Then

x between p0 - .6ax and Po+ .6ax' approve setting

x outside, reject and reset process.

This plan follows from (11.8), using a= .10 = 8, but then using n = 10

instead of n = 9 for simplicity in finding xJs, and also making the critical

value simpler.

One-way Check of Level. If, contrary to the above, the tolerance

T = U - L is eight or more times ax, then there is some latitute for the

process mean p to move around in. This is a case like Section 7.2. A

practical one-way check is the following one with approximate 10% risks:

1. Level nearing upper specifica~ion U

-
Take n = 7 measurements and find x. Then

x < U-2.5a approve setting x > U-2.5ax reject setting

X

P(approvallp=U-3.0ax) = .907 p' = .0013 above U

346 ACCEPTANCE SAMPLING BY MEASUREMENTS

P(approval/~=U-2.5a)
X
.500 p' .0062 above U

P(approval/~=U-2.0ax) . 093 p' .0228 above U

2. Level nearing lower specification L.

Take n = 7 measurements and find x. Then

x> L + 2.5ax approve setting x < L + 2.5ax reject setting.

These two tests may be combined by using n = 7 and approving setting

if x lies between L + 2.5ax and U- 2.5ax.

Checking Process Capability. When buying machine tools for a particular

production requirement, or testing the capabilities of present equipment

for a job, we want to test the variability. In general, as in Section 7.2

we think of ax as the short-term process standard deviation. Now if

the needed tolerance T = U - L is only 6ax, then ~ must be held very

closely to the nominal (U+L)/2. Thus in order to permit any degree of

drift of ).! (for tool wear for example) or inaccuracy of set-up, we need

T to be well above 6ax. One possibility for checking the process capability

is to call

a 1 = T/10 and a 2 = T/6.

Then we may proceed as in Section 11.7. We may well use a .05

for the test. Then using Table VIII, we find

and then since a 1 = .lT,

K = 1.54(.1T) 2 = .0154T 2

Thus approve process if s 2 ~ .0154T 2, etc. This plan, however, will

reject the process about half of the time when T Sax, Hence one might

alternatively let

a1 = T/8. 33, a 2 = T/5

with the same risks. Then still use n 23, but instead use

K 1.54(T/8.33) 2 = ,0222T 2

Then T 8.33ax will only be rejected about 5% of the time.

11.10. Summary 347

11.10. Summary. This chapter has been concerned with sampling

decision-making on the basis of measurement data. The decision can be on

a lot or a process. In the latter case, the use of control charts on the

process is recommended in order to seek out assignable causes as well as

to take action on the process. The use of sampling acceptance tends to

concentrate only on the second, taking action, such as adjusting the level

up or down.

For decision-making on a particular lot from a produce~ or a series

of lots, control charts can still be of much use, but the primary job

is to take action on this one lot. Then acceptance sampling is a natural

decision-making tool.

If we cannot tell what kind of distribution of measurements we have

in the lot, then we are much safer to use a random sample of parts or

pieces and measure each, comparing each x with specification limits L

and U, and taking action on the basis of some attribute plan. This will

call for a substantial sample size in general. The measurements can well

be recorded too, and thus build up knowledge on the producer's frequency

distribution. If this proves to be normal or nearly so we may begin to

try the normal curve methods described in this chapter. These would be

unknown-ox plans, Section 11.6. In a series of lots, plotting of R

or s charts can quickly lead to a good estimate of ox, if there is

control. Then we can use this in known-ox plans, Sections 11.4, 11.5,

and thereby further decrease the required sample size.

Military Standard 414 is an effective system of integrated plans,

providing normal, tightened and reduced sampling, for protection against

one or two specification limits. It is concerned with fraction or

per cent defective. The plans assume normality, but one hedge, if

this is not an available assumption, is to use variables - attributes

348 ACCEPTANCE SAMPLING BY MEASUREMENTS

inspection, permitting quick acceptance via measurements, but only

rejecting via attributes.

Checkson process level and process capability were also given, for

use when normality of xJs can be assumed.

PROBLEMS

11.1. Verify the risks given in the two-way check of level in Sec-

tion 11.9, for Pa vs. ~·

11.2. Verify the risks given in the one-way check of level in Sec-

tion 11.9, for Pa vs. ~· Also verify the three corresponding fractions

defective for Pa vs. p'.

11.3. A minimum specification for individual tensile strength tests

is 90,000 pounds per square inch (psi) and a is known to be about 4000
X

psi. The distribution of tests is approximately normal. For lot average

~. AQL = 102,000 psi, and RQL = 98,000 psi, respective risks are set a=

B = .OS. Determine an appropriate sampling plan for a decision on a lot

and sketch the OC curve Pa vs. ~. labeling axes.

11.4. For "acid number" for lead battery material the minimum specif-

ication is 8, while ax is known to be .39. There is approximate normality.

Set risks as follows P(accl~ ~ 8.8) ~ .OS and P(rej\~ = 9.2) = .OS. Set up

an appropriate sampling plan, cutting B if necessary and sketch the OC curve

Pa vs. ~.

11. S. For an inner axle carrier, specifications of 12.742" to 12. 74S"

are uncomfortably tight, so control of the mean~ essential. The standard

deviation is .000,834".Taking RQL's at 12.743" and 12.744',' and risks a=

.10 = B determine a two-way protection plan for testing ~·

11.6. For a hydraulic stop light switch,specifications for operation

are 4S to 100 pounds per square inch (psi). ax is known to be just about
Problems 349

6 psi. Set up two one-way checks of the process level as in Section 11.9.

Why not use one two-~vay test?

11.7. For a distributor valve bushing, carrying specifications of 1.1760"

to 1.1765", CJ is known to be about .000,040". Set up two one-way process

X

setting checks as in Section 11.9. Why not use one two-way test?

11.8. Suppose that the maximum time of blow of a fuse under a certain

current is 150 sees. (a) If CJx is unknown, set up a looser plan than that

in Section 11.6, with Pi = .005, Pz = .080 and respective risks a .05,

~ = .10. (Assume normality.) (b) If CJ is known to be 25.0 sees. what

X

does the plan become? (c) What attribute plan would provide the protection

desired?

11.9. Weight of fill of a container of insecticide is subject to a

minimum of 427 g. Suppose that the proportions below 427 g. are taken as

Pi ~ . 005 and as Pz = . 050. Respective risks are a = . 05, ~ ~ .10. (Assume

normality.) (a) Set up an appropriate test for x with ox unknown.

(b) Set up an appropriate test on x, i f CJ

X
is known to be 8. 7 g. (c)

What is the attribute plan for the desired protection?

11.10. For the "Scott value" for a material for a battery the maximum

specification is 26. CJx = .56, and assume normality. Set up a sampling

plan i f ~ = 25 is to be rejected 98/, of the time and ~ ~ 24.5 accepted 90%

of the time.

11.11. An analytical technique is supposed to have a standard error

of measurement of 2 parts per million. Set 0 1 = 1.5 ppm and 02 = 2.5 ppm

and risks both .05, and determine an appropriate test for the technique.

11.12. A standard gage has had a measurement error of o 1 = .000,025".

A less expensive one is being considered. Set o2 = .000,035" and risks of .OS.

Find an appropriate test for the gage.

350 ACCEPTANCE SAMPLING BY MEASUREMENTS

11.13. For a lot of thermometers, a standard deviation of .2•c is

considered satisfactory, while .3•c is unsatisfactory. Using risks of .10

respectively on each, set up an appropriate sampling plan. (Thermometers

are to be placed at constant temperature and read.) What is assumed?

11.14. Given cr 1 -, 1.715, cr 2 "'3.43 as AQL and RQL respectively, with

a= S = .10, find a single sample test on variability. Draw a sample of

the required size from distribution A of Table 4.1 and make the test. Do

the same for a sample from distribution C or D.

11.15. Specifications for overall length of an inner roller bearing

race are 1.2475" ± .0025 11 • Set up a process capability test, T/10 vs. T/6

for the process. How would you choose the races to be measured? a=S=.05.

11.16. For the inside diameter of a transmission main shaft bearing

retainer the specifications are 2.8341"- 2.8351". Set up a process cap-

ability test, T/10 vs. T/6 case. How would you choose the required retainers

for the sample?

References

1. Military Standard MIL-STD 414, "Sampling Procedures and Tables for

Inspection by Variables for Percent Defective." U. S. Govt. Printing

Office, Washington, D.C., 1957. Also Amer. Nat. Standards Inst. 21.8.

2. Statistical Research Group, Columbia Univ., "Techniques of Statistical

Analysis." McGraw-Hill, New York, 194 7.

3. W. A. Wallis, Lot quality measured by proportion defective. In "Accept-

ance Sampling, A Symposium," Arner. Statist. Assoc., 1950, 117-122.

4. H. G. Romig, "Allowable Average in Sampling Inspection." Ph.D. Thesis

Columbia Univ., New York, 1939.

5. Naval Ordnance Standard OSTD 80(1952). "Sampling Procedures and Tables

for Inspection by Variables."

References 351

6. A. H. Bowker and H. P. Goode, "Sampling Inspection by Variables."

McGraw-Hill, New York, 1952.

7. Ordnance Inspection Handbook, ORD-M608-10, "Sampling Inspection by

Variables." 1954.

8. W. J. Zimmer and I. W. Burr, Variables sampling plans based on non-normal

populations. Indust. Quality Control 20 (No. 1), 18-26 (1963).

9. D. B. Owen, Summary of recent work on variables acceptance sampling

with emphasis on non-normality. Technometrics 11, 631-637 (1969).

10. J. N. K. Rao, J. Subrahmaniam and D. B. Owen, Effect of non-normality

on tolerance limits which control percentages in both tails of a normal

distribution. Technometrics 14, 571-575 (1972).

 CHAPTER 12

SEQUENTIAL ANALYSIS

12.1. Introduction. Work by Harold Dodge and Harry Romig, and Walter

Bartky on double and multiple sampling respectively supplied background to

the development of sequential analysis. Experimentation in successive stages

was also suggested and used by Harold Hotelling and P. C. Mahal anobis . The

problem of sequential analysis first arose in the Statistical Research Group

of Columbia University in some comments of Captain G. L. Schuyler of the

Bureau of Ordnance. Following this idea up, Milton Friedman and W. Allen

Wallis conjectured that sequential procedures might be deve loped having as

good control of errors of miscl assifying lots as single or double plans,

but which would require l es s average inspection. This was brought to the

attention of the l ate Abraham Wald, who then developed the probability ratio

sequential test, and worked out the general theory of the test and its

application to many important problems. Hence, sequential analysis is in

l arge part a contribution of Abraham Wald . The fore going account is l argely

from reference [1) .

Fundamentally the aim of sequential analysis is to obtain decisions of

requisite se curity on the basis of a smaller average number of tests or

determination s than is possible even with the very best single sampling plans

possible. The saving is often as much as SO% at critical levels of quality

and considerably more at very good or poor quality . In this chapter we will

discuss briefly the fundamental principles of sequential analysis and how

these work out in the various cases in practice. Further, we want to show

how unif ied and comparatively simple the sequential plans are .
354 SEQUENTIAL ANALYSIS

In general, a sequential test is one in which, after each measurement

or determination, we may either accept a hypothesis (or lot or process) re-

ject the hypothesis, or request additional evidence. As such, the sample

size is, in general, a variable; sometimes it is very small, while at other

times it is quite large. Single and double sampling are special cases of

sequential analysis, in which neither acceptance nor rejection is possible

until n (or n 1) pieces are measured or tested. Thus the only decision on

the earlier pieces is a request for more evidence. The sequential test te~

minates according to definite rules, that is, when sufficient evidence for

a decision has accumulated.

12.2. The General Approach. For either sequential or single sampling

(as we have been seeing in the latter case) we may proceed as follows. Let

q stand for any population quality characteristic (for example~· a or p'),

X

and let the AQL be q 1 , RQL (rejectable quality level) be q 2 . Then we

consider the "null" hypothesis H1 : q = q1 , and alternate hypothesis

H2 : q = q 2. Corresponding to these quality levels we set risks a and S,

that is, a= P(rej H1 iq = q1) and S = P(acc H1 iq = q 2), just as usual.

Commonly rejection of H1 means rejection of a lot or adjustment of the

production process, and acceptance of H1 means acceptance of a lot or

approval of a process.

Given the four quantities q1 , a; q2 , S it is possible to find single

and sequential plans, and in ,some cases double or multiple plans, to meet

these conditions. One must always keep in mind that to compare two sampling

plans as to cost and usefulness, we must first make sure that they supply

comparable protection, that is, have similar operating characteristic,OC

curves. If the four quantities just given, which determine two points on

the OC curve,are the same, then the curves will be sufficiently alike for

sound comparison. Knowing that the OC curves are alike, we can compare the
12.3. The Sequential Probability Ratio Test 355

average sample number ASN curves for single and sequential plans, to see
whether the gain in using the latter is worth the increasing complexity of

the sequential plan. The gain tends to be greater for very good or very

poor q values and least for q values somewhere between q1 and q2 • (ASN's

for double plans commonly lie between those for single and sequential, but

not always.)

Sequential sampling plans are not really very complicated at all, once

one gets acquainted with the approach.

12.3. The Sequential Probability Ratio Test. This approach seems

quite straightforward and makes excellent practical sense. Let us form the

following relative probabilities for each sample size n.

P2n relative probability of observed sample, i f H2 is true (12 .1)

Pln relative probability of observed sample, if H1 is tru~. (12.2)

Now form the probability ratio criterion P2n/Pln' Then

If P2n/Pln.:::; B accept Hl (12.3)

If p2 /Pl >A
n n- reject Hl (12 .4)

If B < P2n/Pln < A continue sampling, (12.5)

Criterion (12.3) makes sense, because in this case we have continued our

sampling until P2n is a great deal smaller than Pln' Therefore the explana-

tion H2 is so much poorer as an explanation of the observed sample than is

H1 that we are justified in accepting H1 • (B is of course considerably

smaller than 1 in any practical case.) On the other hand in (12.4) we have
continued sampling until H2 has become a sufficiently better explanation

than H1,and we are justified in taking the action indicated. (Likewise A is

in general much above 1.)

We now need to be able to set A and B to develop a specific plan. These

are largely determined by the risks ~ and ~. but also to a slight extent by
356 SEQUENTIAL ANALYSIS

the type of case in question. In practice, however, a and B alone would

determine A and B very closely. In fact we set, in general

A = (1-B) Ia B B/ (1-a) ( 12. 6)

Sequential plans with A and B so chosen actually will not have precisely

the requested risks a and B, but instead, say, risks a' and B' according

to the case. But it can be shown [1] that:

a'< a/(1-B), B' < B/(1-a) ( 12. 7)

a' + 8' .:::; a+ B. (12. 8)

It could still happen that one of a', B' obtained by using A and B from

(12.6) could be a very small amount above the respective a or B, but their

sum is always less than or equal to the sum of the desired risks a+ B.
So we use (12.6) to set A and B in practice.

In the special cases we shall see that the use of criteria (12.3) to

(12.5) can be codified into quite simple rules, providing acceptance and

rejection numbers for each n. At any given n, the test statistic for the

case in question will lie in one of three intervals (a) accept, (b) re-

ject, and (c) continue.

In many cases we can find points on the OC curve besides q~q 1 , Pa=l-a; q=q 2 ,

Pa = B. And thus we would have a spectrum of conditional probabilities

P(accJq). In most cases the average sample number ASN curve is avail-

able too, providing ASNq as a function of q. This is the average number of

observations needed to reach a decision, whatever it may be.

One other point which bothers some persons is the possibility of a

sequential test going on indefinitely and never terminating. The prob-

ability of this can be shown to be zero. In practice we do set a maximum

sample size and curtail the sampling there.

12.4. Six Special Cases of Sequential Analysis Summarized. We can

easily handle the following cases by sequential analysis:

12.4. Six Special Cases 357

(a). Fraction defective, hypotheses: p' = Pi = AQL and p' p'

2
RQL.

Binomial population for defectives d, assumed.

(b), Number of defects per unit, hypotheses: c' c' AQL and c' c' RQL.
1 2
Poisson population for defects c, assumed.

(c). Mean of measurements, one-way, a known, hypotheses: JJ = JJ 1 = AQL,

]J = JJ 2 = RQL. Normal population of x's assumed. Here we assume JJl < ]J 2 and

that low values are good, with a risk at JJ 1 , 8 risk at JJ 2 . See later on,

the modification if high values are good, subsection (7).

(d), Mean of measurements, two-way, a known, hypotheses: ~ = J.J.0 = AQL

and ~ ~ ~O ± d = RQL. Normal population of x's, assumed. Here there is

one AQL and two symmetrically placed RQL's with respective risks a and B.

(e). Standard deviation, one-way, ~ known, hypotheses: a = a 1 = AQL and

a= a 2 = RQL, a 1 < cr 2 . Normal population of x's, assumed.

(f). Standard deviation, one-way, ~unknown, hypotheses: a

a 1 < a2. Normal population of x's, as~umed.

12.4.1. Particulars of the Six Cases. The following criteria and

formulae contrast and compare the various cases, and show their essential

unity.

(1), Constants a and b from A and B of (12.6). For case (a), defectives,

we use base 10 logarithms, "log":

a = log[(l-8)/a) = log A b = log[(l-a)/S] = log (1/B). (12.9)

But for cases (b) through (f) we use base e logarithms, "ln":

a = ln[(l-8)/a] = ln A b = 1n[(1-a)/S] = 1n (1/B). (12.10)

Thus a and b reflect the specified risks at AQL and RQL. Note that a and

b are both positive, in general.

(2), The sequential test for each n is a cumulated function of the n ob-

servations in the sample, as follows:

 358 SEQUENTIAL ANALYSIS

(a\ Total defectives, d, in n pieces or parts

(b~ Total defects, c, in n units

(c), Total measurements, L:x, or often coded from k and using L:(x-k) for

n measurements

(d). Total deviations from ~-'o• summed algebraically, then absolute value

taken, IL:<x-1-' 0 )1 for n measurements

2
(e). Total variation from known mean f.L, L:(x-f.L) for n measurements

(f). Total variation from sample mean x, L:(x-x) 2 ~ L:x 2 - [(L:x) 2 /n] for n

measurements

(3), Criteria of acceptance, rejection and continuation at sample size n

are all based upon acceptance number -h 1 + sn and rejection number h 2 + sn,

where -h 1 and h 2 are intercepts and s is the slope of two parallel lines for the

test criterion plotted against n. If the test criterion lies above the

upper line we reject, if below the lower line we accept, if between we con-

tinue. See Figure 12.1. (However, in case (d) there is a slight curving

in the lower line for small n's.)

(4), Formulas for h 1 , h 2 and s for the cases. Be sure to use the correct

formula for a and b, that is, (12. 9) for case (a), (12 .10) for the others.

b a
(a), hl (12.11)
p'(l-p')
2 1 Pz (1-pi>
log log
p'(l-p') pl (1-pz)
1 2

log [ (1-pi' I (1-p~)]

s ~ (12.12)
p' (1-p')
2 1
log
p'(l-p')
1 2

b a
(b), hl (12.13)
ln C I - ln cl
2
 12.4. Six Special Cases 359

Test
criterion

FIG. 12.1. General sequential test criteria regions.

c ' - c'
s = --=.,2--'1~-- (12 .14)

(12 .15)

(12.16)

(d). h 1 = (u 2 I d) (b-. 693) (u 2 I d) (a + . 693) (12.17)

s = d/2 (12 .18)

Technically, for small n the acceptance criterion is

(12.19)

But it usually suffices to plot the intercept

(12.20)

and draw a curve to the acceptance straight line, -h 1 + sn.

360 SEQUENTIAL ANALYSIS
2b 2a
(12.21)
(1/cr~)-(1/cr~)

ln (cr 2 2
2 ;cr 1)
s = (12. 22)
(1/cr~)-(1/cr~)

(f), Use same h 1 , h 2 and s as in (e) but instead of plotting -h 1 + sn and

h 2 + sn vs. n, plot -h 1 + s(n-1), h 2 + s(n-1) vs. n.

(5), Plotting of three points and using Pa = 0 and Pa = 1 as asymptotic

lines is usually enough for a sketch. These are called "five point OC

curves," By appropriate formulas they may be supplemented [ 2].

(a), p' = 0 Pa 1
p' ·= p' Pa 1 - 01
1
p' = s Pa = h2/(hl + h2)
p' p' Pa ,. 13
2
p' 1 Pa 0

(b), c' 0 Pa 1
c' = c' Pa 1 - 01
1
c' = s Pa = h2/ (hl + h2)
c' c' Pa 13
2
c' large Pa ·= 0

(c), I.L low Pa = 1

I.L = I.Ll Pa 1 - 01
I.L ... s Pa ., h2/(hl + h2)
I.L = I.L2 Pa = 13
I.L high Pa ,;, 0

(d), I.L low Pa :., 0

I.L ... I.Lo - d Pa 13
I.L .., I.Lo Pa 1 - Ct
I.L "' I.Lo + d Pa = 13
I.L high Pa ~ 0
12.4. Six Special Cases 361

(e). a = 0 Pa 1
a =a Pa 1 - 01
1
a = rs lslope Pa "' h2/ (hl + h2)
a = a2 Pa 13
a high Pa = 0

(f). Same as (e) .

(6). Likewise we may draw the average sample number ASN curves. These

too are often sketched by five points, but intermediate points may be

found [2].

(a). PI 0 ASN h 1 /s round up to whole number

PI PI (1-a) ·h 1 - a·h 2
1 ASN

pI ~ S ASN h 1h/[s(l-s)]

= (l-f3)h2 - f3 hl
PI PI ASN
2
Pz- s
PI 1 ASN h 2 /(l-s) round up to whole number

(b), C I 0 ASN h 1 /s round up to whole number

(1-0I)hl - 01 h2
cl cl ASN
1

c1 =s ASN h1h 2/s

(l-i3)h - f3 h
2 1
c1 = c2 1 ASN
c1 - s
2
cl ASN 0 (we reach rejection number before completing inspection
of even one unit)

(c), IJ. low

IJ. =s
362 SEQUENTIAL ANALYSIS

(1-B)h 2 - ~ h1
IJ. ~ IJ.2 ASN
IJ.2 - s

IJ. high ASN ;. 1

(d), ASN curve not available

(e), cr =0 ASN ~ h /s round up to whole number

2
(l-a)h 1 - a h 2
ASN

2
cr = Is /slope ASN h 1h 2 /2s

(l-~)h 2 ~ h1
cr ASN
(]2 - s
2
(] high ASN !, 1

(f~ ASN precisely 1 greater than for case (e).

(7), Comments on particular cases.

(a), A great many particular cases are worked out in [2] providing

h 1 ,h 2 , sand the five ASN points, for pl' Pz and with a= .OS and S

either .10 or .SO. The high S's give "reduced" sampling plans.

(b), Note that test criterion c is the total defects on n units. It

might be thought of as c 1 + c 2 + ... +en if the c''s are above 1 (or

even below) .

(c), If high values of x are good and low values bad, then it is

convenient to retain ~J. 1 as the low mean but now call it an RQL; ~J. 2 as

the high mean but now call it an AQL. Also we retain Cl' as the ris~ at

J.L 1 , namely for acceptance, and ~ as the risk at ~J. 2 for rejection (just

contrary to common practice). This is done so that all formulas in

subsections (4), (S) and (6) follow through, except that Pain (S) is

now interpreted as P(rej).

Another point is that coding around some number near IJ.l and ~J. 2 is

desirable, so as to avoid huge s's which make the acceptance-rejection

 12.5. Some Examples 363

graph unmanageable. In particular if we use coding from the nominal,

~O = (~ 1 + ~ 2 )/2, then s becomes zero and the test criterion ~(x-~ 0 )

may be compared to two horizontal straight lines. That is, the accep-

tance and rejection numbers are constant.

(d), Great care must be used with signs in accumulating I~Cx-~ 0)1.
2
(e), (f). It is fairly easy to accumulate the test criteria ~(x-~) , but it

is quite a chore to make direct use of ~(x-x)

- 2 , so the form ~ x
2 -

[(~ x) 2 /n] is recommended. See also the sequential use of ranges in

Section 12.6.

12.5 Some Examples. Let us now illustrate the preceding material

with a few examples.

12.5.1. Defectives. One company in the author 1 s experience changed

over from large amounts of 100% sorting of purchased piece parts to the use

of sequential analysis. This dropped the cost per 1000 parts received,

from $7.00 to $.28.that is, a 96% saving. Moreover there was no more

trouble with bad parts than usual. The few rejected lots could be more

carefully sorted than the former routine sorting, and still obtain the

96% saving. The plan used was pl = .01, a = .05; Pz = .02, 8 = .15, and

truncation was used.

Another company making carburetors developed and made wide use of

three sequential sampling plans in receiving inspection. They used the rather

tight risks of a .OS, 8 = .01 on three plans:

(a) Critical pl . 01' Pz .03

(b) Major pl" .01, Pz .OS

(c) Minor Pl .02, Pz .10

To use the plans they made three tables of acceptance and rejection numbers,

by classes of n's going up till n was about equal to (a) 1.5 (ASN at

Pi), (b) 2.3(ASN at Pi), and (c) 3(ASN at Pi). If no decision was reached

by then, they sorted the lot 100%.

 364 SEQUENTIAL ANALYSIS

Let us illustrate with a sequential sampling plan for

P} = .01, a = .OS; P2 = .OS, B = .10.

A corresponding single sampling plan for these requirements may be found

from Table 9.5 to be Ac = 3, Re = 4, n = 134.

From (12.9), (or we could use Table XI),we find for this case of a= .OS,

B = .10: a= 1.25527 b = .97772. Then we use the formulas for case (a),

(12.11) and (12.12)

. 97772 1. 25527
1. 3638, h2 1. 7510
1 .05(.99) same
og .01(.95)

1 . 99
og -:95
s = .02500,
same

Next we use subsections (Sa) and (6a):

1. 7510
Pa(O)=l, Pa(.Ol)=.95, Pa(.025)= 3 . 1148 .562, P(.OS)=.lO, Pa(l)=O.

ASN(O)=l.3638/.02500=54.6 or 55

.95(1.3638) - .05(1.7510)
ASN(.OlO)= 80.5
.02500 - .010

ASN (. 025) = 1.3638(1.7Sl0) = 98.0

.02500(.975)
ASN(.OSO)= .9(1.7510) - .1(1.3638) = 57 . 6
.050 - .02500

ASN(l)= 1.25527/(1-.02500)=1.29 or 2.

Thus even when the ASN is at its approximate maximum there is a considerable

gain over the single n 134. Much more gain is obtained as p' moves away

from the value p' = .025 = s,

The acceptance-rejection regions are shown in Figure 12.2. These

are drawn up to three times the larger of ASN(p{) and ASN(pz), as it is

12.S. Some Examples 36S

2
+h2

"I
-h,
-2
0 40 80 120 160 200 240

FIG. 12.2. Acceptance-rejection regions for defectives

sequential plan: Pi = .01, ~ = .os; Pz = .os, S = .10.

common and quite sound prac1tice to truncate at such an n and make the decision
according to which region the test criterion is closest to. Such a procedure

does not do much violence to the a and s risks. Thus we go up ton = 240.
Then if at n = 240 we have 7 defectives in 240, we would reject, or if S or
6 we accept. In fact just as soon as a seventh defective is found we could

stop and reject. This could occur as early as n = 210.

Although a figure such as 12.2 could be used to plot observed results
and to make decisions, it is commonly easier to make up a table of acceptance

and rejection numbers from

Ace. No. = -h 1 + sn Rej. No. (12.23)

Thus here we have

366 SEQUENTIAL ANALYSIS

Ace. No. = -1.3638 + .02500n Rej. No. = 1.7510 + .02500n.

A desk calculator can easily be used to show that for the integral numbers
of defectives we have:
Ace. No. Class of n's Rej. No. Class of n 1 s
(Negative) 1 -54 2 2 - 9
0 55 - 94 3 10 - 49
1 95 - 134 4 so - 89
2 135 - 174 5 90 - 129
3 175 - 214 6 130 - 169
4 215 - 254 7 170 - 209
8 210 - 249
For this we find the Ace. No. by (12.23) to be negative up through n 54

where it is still -.0138. But for n = 55, it becomes +.0112, whence d =0

is acceptable. But then the Ace. No. remains less than 1 till n = 94 when
it is .9862, but at n = 95, it is 1.0112, whence d =1 is now acceptable.
And so on. On the other hand continuous calculation of the Rej. No. by

(12.23) gives 1.8010 at n = 2, so d = 2 is rejectable (if the first two

pieces are both defectives). The Rej. No. remains below 2 till n = 9.
But at n = 10 it becomes 2.0010 so that it now takes d =3 to exceed the Rej.
No. The last remains below 3 through n = 49 (2.9760). But at n =SO the

Rej. No. becomes 3.0010 and it then takes d =4 to reject, etc. Such classes
are obtained continuously without clearing the machine. They can now be

put together as follows:

Class of n Ace. No. Rej. No.

2 - 9 * 2
10 - 49 * 3
so - 54 * 4
55 - 89 0 4
90 - 94 0 5
95 - 129 1 5
130 - 134 1 6
12.S. Some Examples 367

13S - 169 2 6
170 - 174 2 7
17S - 209 3 7
210 - 214 3 8
21S - 249 4 8
* Cannot accept on such small n's.

12.S.2. Test on Means, Lower Specification. Let us take as an illustrati\

example, that of Section 11.4.1 for tensile strength. Assuming it is feasible

to prepare and test samples one at a time sequentially, quite a bit of sample-
size saving is available using a sequential test. We here use the comments

under Section 12.4.1, subsection (7c). We take ~l = 6S,OOO psi, ~2 = 68,000

psi (the higher average). Also a= 3SOO psi. Now in Section 11.4.1, since
~l is rejectable, the probability of erroneously accepting such a lot was
called ~ = .OS, and of erroneously rejecting a lot with p = p2 was called

a= .10. Now in order to make direct use of the formulas in Section 12.4.1,
subsections (1), (2c), (4c), (~c) and (~c), we rename the risk at ~1 = 6S, 000

psi to be a = .OS, and that at p 2 68,000 psi to be a = .10. We are now

ready to use the formulas for a, b, h 1 , h 2 and s. We may find a and b from

Table XI, or use (12.10) obtaining

a = 2.890 b = 2.2Sl.
Next for s, if we use the formula in (4c) we find

s = (6S,OOO +
268,000) psi= 66 ,soo psi

We would find it awkward to use such a large slope, so suppose we code in

100 psi units since the original measurements were to the nearest 100 psi.

Now in line with (2c) we could choose u = (x - k)/100 psi where k is a

convenient number of psi. k could well be ~l = 6S,OOO psi. But the most

convenient one is ~0 = (~ 1 + ~ 2 )/2 = 66,SOO psi. Then we would measure

the tensile strengths in terms of u = the number of 100 psi from 66,SOO psi,
368 SEQUENTIAL ANALYSIS

and p1 = -15, p2 +15 and s becomes zero! Meanwhile a X = 35. Next let us
find h 1 , h 2 :
2.251(35 2)/30 91.92

2.890(35 2)/30 118.01

Hence for the test criterion

~ui = ~(xi - 66,500)/100

we have the very simple rule
"Ace. No." -h 1 -91.92

"Rej. No'." +h 2 +118.01.

The quotation marks were used because in reality if

~ui < - 91.92 Reject lot
~ui > +118.01 Accept lot.

Now how much sampling do we do on the average (ASN)? We have from (6c):

ASN(11 1) = (l-a)h 1-ah 2 .95(91.92)-.05(118.01)

= 5.43
S-)11 0-(-15)

(1-8)h 2-Sh 1 .9(118.01)-.1(91.92)

ASN ( 11z) 6.47.
11 2-s +15-0

Meanwhile the approximate maximum ASN occurs at p = 66,500, giving

2
ASN[(l1 1+11 2)/2] = h 1h 2/a = 91.92(118.01)/35 2 = 8.86
These ASN's compare favorably with the single n = 12.
12.5.3. Test on Means, Two Specification Limits. Let us consider

a test on hardness Rockwell C. Suppose that a= 1.7 points and p0 = 65

is desired (specifications perhaps being 60 and 70). Let d 1.7 also
matching a. Then suppose that we set risks a = ~ = .10. Then by the appropriate
Section 12.4.1, subsections (1), (2d), (4d) and (5d) we can find the desired

criteria and sketch the OC curve:

a = 2.197 = b
h 1 =(a 2 /d)(b-.693) (1.7 2/1.7)(2.197-.693) 2.56
12.6. Use in Checking a Process Setting 369

h2 = (a 2/d)(b+.693) = (1.7 2/1.7)(2.197+.693) 4.91

s = d/2 = . 85.

These suffice to set the acceptance-rejection regions, except for the

short curved section for acceptance. The intercept of this curve on the n

axis is at 2bo 2/d 2 = 2(2.197) = 4.4. See Figure 12.3, which shows a curve

with vertical slope at n = 4.4, rounding into the acceptance line as·an
asymptote. The OC curve can only be sketched from the a and 8 points in
this case, and the ASN curve is not available at all.

The quality manager or engineer in charge would have to decide whether

risks of .10 are too large. A single sampling plan giving comparable protection

has an n = 9.
12.6. Use in Checking a Process Setting. A frequently occurring problem
in industry, is that of determining when to adjust or reset the process
level and when to leave it alone. Of course the short-term process variation

II x-65)1

16

12

0~------~~--~-------------------------------

2 4 6 8 10 12 14

FIG. 12.3. Acceptance and rejection regions for two-way test on

means. Test criterion Icx-65). ~ 0 = 65, d = 1.7, o = 1.7, ~ = .10,
8 = .10.
370 SEQUENTIAL ANALYSIS

a is what complicates this decision. In the early stages of using any

X

process, we recommend the use of x and R or s charts to seek out and eliminate
undesirable assignable causes. Eventually the variability chart, at least,
stabilizes, and we may estimate the short-term process standard deviation, ox

It then becomes a question as to whether to continue use of an x, R chart

set-up, or to go to periodically checking the process level. In any case,
some way of checking the process average level at the beginning of a production
run is standard practice. We have already presented a pair of single

sampling checks of process average in Section 11.9, for one-way and two-way
cases.
Two sequential checks of process average were developed in [3] for this
problem, and tables provided so that ~ calculation is needed prior to making
the process check, as long as ox is known well enough to place it in one of the
given ranges of ox. See Tables XIII and XIV, given in the back of this book.
Knowing a from past data (R or s chart in control), we find which of the 12
X

class-intervals contains our ox' coding the x's if necessary.

Consider first a one-way check. Table XIII, gives the acceptance and
n
rejection numbers for l:i=l(xi-p 1), where p1 is safely below an upper
specification, U. This is for protection against a level p 2 = p1 + ox
which is unsafely close to U. Respective risks are a = S = .10. If at a
lower specification, L, we still call PI
the lower mean which is now "unsafe'!
n
and p2 = p1 + ox now "safe'~ and make the test on l:i=l(xi-~ 1 ). But now the
numbers called "Ace." in the table are rejection numbers. That is,
l:(xi-pl) ~"Ace." then reject the setting as being too low, etc.
For a two-way check of u vs. p + a with risks both at .10, use
•o o - x
Table XIV, and use the test criterion !l:~= 1 (xi-p 0 )!. Signs must be
carefully watched, then the numerical value of the sum taken, to compare
against the appropriate pairs of acceptance and rejection numbers in the table.
12.7. Sequential Test for Process Capability 371

These sequential checks of the process setting reach a decision much

more quickly on the average, than do single sampling plans with the same
discriminating power, but they are harder to teach and to administer. This
requires a decision!

12.7. A Sequential Test for Process Capability, Using Ranges.

Corresponding to the process capability check as given in Section 11.9,
there is a very simple check involving ranges [4]. As pointed out in
Section 11.9, if the tolerance T = U - L must be met, then cr 1 = T/10
provides comfortable meeting of this tolerance, but cr 2 = T/6 makes it

difficult to maintain the tolerance T. (Both cr's are for the short-term

variability.) Two sequential plans are given with risks (a) a = S = .OS
and (b) a = S = .01. To use the plan one has but to multiply the entries

in columns (3) and (4), or (S) and (6) in Table XII by the required T. Then
take a random sample of n = 8 pieces from production (over a short period
of time) and find the range R1 . I~ for example,with risks .OS, P1 ~ .19T,

accept the process as being capable. But if R1 ~ .S4T, reject it. But
if R1 lies between, take another sample of n =8 and find Rz· Then if &1

+ R2 ~ .SST accept, if ~ .90T r~ject, otherwise take an additional sample,

etc. A maximum of 8 samples of n =8 might be required for a decision, if

a = B = .05. This plan is simple, and yet it has a much more favorable

ASN curve than does the comparable single sampling plan. In fact its ASN
-2
curve is very nearly as good as the sequential plan for cr's using ~(xi- x),
~ unknown, which is quite a chore to use, with the necessity of refiguring

~x~-
1
[(~x.)
1
2 /n) after each observation.

12.8. Use in the Laboratory. Sequential analysis for measurements

can be quite an aid to laboratory experiments. This is especially true

where observations or outcomes, such as runs, naturally occur in succession

rather than in parallel. The latter would suggest use of single sampling.
 372 SEQUENTIAL ANALYSIS

One can run an effective significance of difference test on two materials,

two formulations, two methods, etc., if one can estimate the respective cr 1 and cr 2

(or a reasonable upper limit). All we need do is to start taking pairs of

observations (x 11 , x 21 ), (x 12 , x 22 ), (x 13 , x 23 ), Then form

Y1 = xll - x21' Yz = xl2 - x22' Y3 = xl3 - x23'

Now one reasonable test is as to whether the two materials differ or are
reasonably alike. We can set

p0 for an AQL for y's

and
~0 ~ d for two RQL's for y's.

Using appropriate risks a and S and cry = ~i + a~, we may then use case (d)
of Section 12.4.1 for a two-way test on Py· Decisions would be to take
action appropriate if (a) ~l ~ ~ 2 + d, or (b) ~l ~ ~ 2 - d, or if (c) ~l

~2 (no appreciable difference).

It is easy to modify the foregoing for a one-way check on ~ by deciding

y
upon two critical levels, so that there is no alternative such as (c) above.
12.9*. Two Derivations of Acceptance-Rejection Constants. In order
to give some of the flavor of sequential analysis proofs we derive formulas

for h 1 , h 2 and s for cases (a) and (c) of Section 12.4.1, using (12.1)

through (12.6).
For the case of defectives, we have observed some ordered sample

of good (G) and defective (D) pieces, which has not yet lead to a decision.
For example,for n pieces we might have GG ... GOG ... GDG ... G where
there are two D's inn pieces. The probability for this particular ordered

sequence of outcomes is p~ 2 (1- p')n- 2 . Now form P2n/Pln' We have ford

D's, n - d G1 s:
")n-d
- P2 > A gives rejection. (12.24)
')n-d
- pl
12.7. Sequential Test for Process Capability 373

The criterion is to be solved for the rejection numbers for d, that is,

for each n, what d would reject. Taking logs to the base 10 of (12.24) gives

d log P:2 + (n-d) ·log (1-p:Z) - d log Pi- (n-d) ·log (1-pi) ~log A.

Collecting terms and using the laws of logs gives

rej. (12.25)

where log A a, as defined in (12.9). Solving (12.25) for an inequality

for d gives

l-p1
log -1 ,
-p2
+ n. -----"--- rej. (12.26)

since the coefficient of dis always positive with P:2 >Pi· Defining (12.26)

to be
rej. (12.27)

we find the formulas for h 2 and s given in Section 12.4.1, subsection (4a).

For h 1 we first recall that from (12.9) b = log(l/B) or log B = -b,

and then everything goes through as for the rejection case. The quantities

h2 + sn and -h 1 + sn are called "rejection" and "acceptance numbers,"

respectively.

For the case of a one-way test on the mean, say with AQL = p1 < ~2

RQL and with a known, we assume a normal distribution. We call P1n and

P2n relative probabilities because we actually use density functions.*

Now for the first observation we have

*They can be made approximate ordinary probabilities by multiplying by products

of ~x's, which then proceed to cancel out.

374 SEQUENTIAL ANALYSIS
2 2
e- (xl-ll2 ) /2 /ofu ° e-(x 1-ll 2 ) 2/2o 2 + (x 1-ll 1) 2/2o.
2
(12.28)
2 2
e-(xl-lll) / 20 /ofu

Now consider the exponent of e. We have

2 2 2 2 2 2
-(xi- 2xlll2 + ll2) + (xl- 2xllll +Ill) = 2xl(ll2- Ill)- (ll2- Ill)
2o 2 2o 2

where "exp" means the exponential function. Now i f P21 /P 11 lies between

A and B we continue with a new x 2 . Then

because we merely add in the new exponent. In general then reject

as soon as

Now take base e logarithms "ln," giving

2 2
2(Exi)(ll2- Ill) - n(ll2- Ill) > lnA =a
2o 2

by (12.10). Solving for Exi as test criterion yields

if we recall that ll 2 > )J 1. Thus, since this is h 2 + ns we have the formulas

for h 2 and s as given in Section 12.4.1, subsection (4c). The formula for

h 1 follows similarly.

Derivations for other cases and for OC and ASN curve points are given

in [1].
12.10. Problems 375

12.10. Summary. Sequential analysis, as we have seen, can handle

the various cases of significance testing or decision-making normally

done by single sampling (or double sampling). Moreover, sequential

sampling is the usual starting point for developing a multiple sampling

plan for defectives, with a desired OC curve.

The various cases are compared in Section 12.4. Two other cases can

be handled: (a) differences between two proportions, for example, two

processes, [1], and (b) a one-way test on the mean with crx unknown.

The latter requires use of a new t-test calculation after each new

observation [5].

Use of sequential analysis is most natural when measurements are relat-

ively expensive and can best be done in sequence, rather than several all

at once. This may frequently be the case in laboratories and in research.

Another opportunity, which is quite feasible, occurs when we can work out

the test for routine usage, as in Sections 12.6 and 12.7, so that we do not

need to start calculating the criteria all over again for each new decision

situation.

Again it is emphasized that average sample number curves are only to

be compared if the protection (OC) curves are quite similar.

PROBLEMS

12,1, Set up a sequential sampling plan for defectives: Pi = .005,

a = ,05; Pz = ,030, 8 = .10, Find h 1 ,h 2 ,s and sketch OC and ASN curves,

labeling axes, Make up a partial table of acceptance and rejection numbers

for classes of n,

12,2, Set up a sequential sampling plan for defectives: Pl = ,010,

a = ,05; Pz = ,080, 8 = ,10, Find h 1 ,h 2 ,s and sketch the OC and ASN curves,

labeling axes, Make up a partial table of acceptance and rejection numbers

for classes of n,
376 SEQUENTIAL ANALYSIS

12,3, Set up a sequential sampling plan for defectives as in problem 11,9

(c), Find h 1 ,h 2 ,s for the plan, sketch acceptance-rejection regions, and the

OC and ASN curves, labeling axes,

12.4, Set up a sequential sampling plan for defectives for the case in

Problem 11,8(c), Find h 1 ,h 2 ,s, sketch acceptance-rejection regions and the OC

and ASN curves, labeling axes,

12,5, Corresponding to Problem 11,3, use ox = 4000 psi, AQL = 102,000 psi,

a = ,05; RQL = 98,000 psi, S = ,05, and set up a sequential test assuming
normality, Code the x's in convenient fashion, find h 1 ,h 2 ,s and sketch

acceptance-rejection regions, OC and ASN curves, labeling axes,

12,6, Corresponding to Problem 11,4, set up a one-way sequential test of

means, Code the x's in convenient fashion, if desired, find h 1 ,h 2 ,s and sketch

acceptance-rejection regions, OC and ASN curves, labeling axes,

12,7, Set up a sequential test, as in Problem 11,9(b), first finding

the two critical mean weights of fill having Pi = ,005 and P2 = ,050, Then

find h 1 ,h 2 ,s using suitable coding, sketch acceptance-rejection regions, and

the OC and ASN curves, labeling axes,

12,8. Set up a sequential plan of process level for blowing time of fuses

as in Problem 11,8(b), finding h 1 ,h 2 ,s sketching acceptance-rejection regions

and the OC and ASN curves, labeling axes,

12,9, Corresponding to Problem 11,7, set up a sequential check of process

level for the lower specification 1,1760", Use 11 1 = 1,17608" = RQL, 11 2 =

Problems 377

1.17612 11 = AQL, a = [3 = ,10, ax = ,000,040", Find h 1 ,h 2 ,s using a convenient

coding, sketch acceptance-rejection regions, and the OC and ASN curves, labeling

axes,

12.10, Set up a one-way test on means, corresponding to Problem 11,10,

Find h 1 ,h 2 ,s and sketch acceptance-rejection regions, and the OC and ASN

curves, labeling axes,

12,11, Corresponding to Problem 11,5, set up a two-way sequential test

of means, Use coded variable z = (x-12,7435")/.0001", and find h 1 ,h 2 ,s,

sketch acceptance and rejection regions, and OC curves, labeling axes.

12,12, Corresponding to Problem 11,6 for pressure on hydraulic stop

light switch, set up a one-way sequential test at the upper specification 100

pounds per square inch, using a = [3 = ,10 at ~1 = 82 vs. ~2 = 88 psi, Find

h 1 ,h 2 ,s, sketch acceptance-rejection regions, and the OC and ASN curves,

labeling axes,

12,13, Find a sequential sampling plan to correspond with the single plan

on variability, ~unknown, in Problem 11.11, Find h 1 ,h 2 ,s and sketch acceptance-

rejection regions, and the OC and ASN curves, labeling axes.

12,14. Find a sequential sampling plan to correspond with the single plan

on variability, ~ unknown, in Problem 11.12, Find h 1 ,h 2 ,s and sketch acceptance-

rejection regions and the OC and ASN curves, labeling axes,

12,15, Corresponding to Problem 11,13, find a sequential plan to make the

test a1 = ,2°C, o 2 = ,3°C, a= [3 = ,10, Find h 1 ,h 2 ,s, sketch acceptance-

378 SEQUENTIAL ANALYSIS

rejection regions, and the OC and ASN curves, labeling axes, Use ~ known
case, Note that (a) ~ known as the exact temperature, then we are testing
"accuracy", which includes bias and repeatability, and (b) ~ unknown, then
we are testing repeatability or "precision".

12.16. Corresponding to Problem 11,14, set up a sequential test for the

case cr 1 = 1.715, a 2 = 3,43, a= S = .10. Sketch the acceptance-rejection
regions, and the OC curve and ASN curve, labeling axes, for the ~-known

case, Test once each with districutions A and C of Table 4,1, using given

~·s. (Class results, decisions and n's may be pooled,)

12,17, Corresponding to Problem 11,15, set up a sequential test of

process capability for specifications 1. 2450" to 1. 2500" or T = • 0050 11 for
cr 1 = T/10 vs, a2 = T/6, a= a= ,05, Find h 1 ,h 2 ,s, sketch acceptance-
rejection regions, and the OC and ASN curves, labeling axes for the case ~

known. Also give scales for ~ unknown,

12.18. Corresponding to Problem 11,16, set up a sequential test of process

capability for specifications 2,8341 11 to 2.8351 11 or T = ,0010 11 for cr 1 = T/10,

a2 = T/6, a = a = ,05, Find h 1 ,h 2 ,s and sketch the acceptance-rejection

regions, the OC and ASN curves, labeling axes, for the case~ known. Also
give sca.les for ~ unknown.

12,19, Set up a sequential range test of process capability with T=17, as

in Section 12,7, with a= S = ,05, Test once with population A orB of Table 4,1,
Test also once with population C or D, What would you guess Pa is for the

latter distributions?
References 379

12.20. Set up a sequential sampling plan for defects: cl = 4, a = .OS;

cz = 6, S = ,10. Find h 1 ,h 2 ,s and sketch the acceptance-rejection regions,

and the OC and ASN curves, labeling axes.

12.21. Set up a sequential sampling plan for defects: c1 = 6 a = OS·

1 ' • '
.10. Find h 1 ,h 2 ,s and sketch the acceptance-rejection regions

and the OC and ASN curves, labeling axes.

12.22. For a Poisson distribution, derive the formulas for h 1 ,h 2 and s

from criteria on P /P 1 •
2n n

12.23. In what respect is a sequential sampling plan more efficient than

a corresponding single sampling plan?

12,24. Suppose that a sequential sampling plan is worked out for a one-

way test on process means, for a given ox. If the resulting plan is used when

0 is actually larger than assumed, what does this do to the OC curve for
X

12,2S. Why must we always have a + S < 1? Explain in relation to the OC

curve for q 1 vs. q 2 •

References

1. A. Wald, "Sequential Analysis." Wiley, New York, 1947.

2. Statistical Research Group, Columbia Univ., "Sequential Analysis of

Statistical Data: Applications." Columbia Univ. Press, New York, 194S.

3. I. W. Burr, A new method for approving a machine or process setting,

Parts I to III. Indust. Quality ControlS (No.4), 12-18, 6 (No.2),
15-19, 6 (No. 3), 13-16 (1949).
380 SEQUENTIAL ANALYSIS

4. H. M. Wies, Jr. and I. W. Burr, A simple capability acceptance test, a

sequential test on ranges. Indust. Quality Control 21 (No. 5),266-268,
(1964).
5. National Bureau of Standards, Tables to Facilitate Sequential t-tests.
Dept. of Commerce, Applied Mathes. Series 7 (1951).
 CHAPTER 13

SOME OTHER SAMPLING PLANS

13.1. Sampling Inspection of Continuous Production. When it is

desired to use sampling-inspection control of a continuous flow of discrete

units, parts, items or assemblies, the situation is different from that in

which product appears in separate lots, such as was the case in Chapters 9

and 10 .

Of course one approach is to use a control chart set-up on such

product, with a view to finding causes of sub-standard material and

improving the production processes. Control charts could be for p, c,

or for x and R if a measurable characteristic is the criterion. Moreover

the chart could be for 100% of the units, segregated into samples of n

in succession or could be samples of n chosen unbiasedly from larger

segments of production. The objective would be to improve the processes

and thereby obtain adequate and dependable quality. The chart could be

thus used for action on the process, and if sampling is being employed, for

action on the present product, such as, sorting 100% if danger is signaled.

An alternative to this approach is to use a sampling inspection plan

for continuous production, some of which plans we will describe. These

plans can well be used during production or at a final inspection station,

especially under the following conditions.

Conditions for Sampling Inspection of Continuous Production

1. There are no natural lots submitted for inspection and decision.

Instead there is a continuous stream of product as on a

conveyor belt.

381
382 SOME OTHER SAMPLING PLANS

2. There are discrete items of product each of which is either a good

one or a defective, that is, has none or has one or more defects.

3. Production is not so rapid but what 100% inspection is possible.

4. The quality of production has an acceptably low fraction defective

for at least much of the time.

Continuous sampling plans for a particular defect, or for a class of

defects, call for alternating periods of 100% sorting and of sampling

inspection. The relative amount time spent on each depends upon the

quality level of the defect in question: if quality is excellent, most

of the time is spent under sampling inspection, and so forth.

13.1.1. The Original CSP-1 Plan. This plan was developed by Dodge [1].

The procedure is as follows:

1. At the outset, inspect, for the defect(s) in question, 100% of

the units consecutively as produced, and continue such inspection

until i units in succession are found clear of the defect(s).

2. When i units in succession are found clear of defects, discontinue

100% inspection, and inspect only a fraction f of the units,

selecting the sample units one at a time from the flow of units,

in such a way as to insure an unbiased sample.

3. Whenever, under sampling, a defective is found, revert immediately

to 100% inspection of succeeding units, and continue until again

i units in succession are found clear of defects as in section 1,

when sampling inspection is resumed.

4. Correct or replace with good units, all defective units found.

The protection provided by the plan is determined by the two chosen

constants i and f. The larger i is, the more difficult it is to qualify,

or requalify, for sampling inspection, and thus there is more protection

13.1. Sampling Inspection of Continuous Production 383

against relatively poor quality. Likewise the larger the proportion f we

sample, the more quickly we will find a defective and thus be returned to

100% inspection, especially when quality is relatively poor. Suppose that

the units being inspected come from a controlled process at some fraction

defective p'. Then the constants i and f will determine the long-~

proportion of the time (a) the inspection is on sampling with the outgoing

fraction defective still at about p', and (b) the inspection is 100% with

the outgoing fraction defective zero. Thus the outgoing fraction defective

or AOQ , is a weighted average of p' and zero, the weights for conditions
p
(a) and (b) depending upon p' and of course i and f.

Now over the whole spectrum of possible p' fractions defective there

will be some maximum AOQ, called the AOQL (average outgoing quality limit),

just as in lot by lot acceptance sampling. This will not be a function of

p' like the AOQ is, but will depend upon i and f only. Of course it will

occur at some critical p'. This p' will lie between relatively good p''s

for which there will be mostly sampling inspection with the AOQ close to

p', and relatively poor p',s with most of the inspection 100% thus also

giving low AOQ's. The AOQ curves for such plans would be much like those

for typical lot by lot plans. See Figure 9.3. Thus specified

constants i and f provide an AOQL protection, whenever the process is

controlled.

But what if the process is not controlled, which is the more likely

case? Here the AOQL still applies, if the shifts of p' are not directly

correlated with whether the plan is on sampling or 100% inspection. Since

such would seem to be very unlikely the AOQL is still meaningful. Moreover,

the actual average AOQ is likely to be considerably below the AOQL,

because the latter can only occur when p' is right at the critical p'.

Now how are the AOQL, i and f related? This relationship is shown

graphically in Figure 13.1, which uses logarithmic scales. One might

 w
00
.t>.
• tbe 'ftlua of per cent 4e-
fect1ve. 1n a conaecut1ve
run of N c 1000 product
units tor which the prob-
abilit7 ot acceptance.
is 0.10 tor a sample size
ot

(/)
0
:s:
m
0
:tm
(/)
""
, I, NUMBER or UNITS ~
'"0
!:""""
.....
z
Cl
FIG. 13.1. Curves for determining values of i and f for a given value of
'"0
AOQL [1]. Reproduced from reference [1] with the kind permission of the !:""""
;J>
Editor of the Annals of Mathematical Statistics. z
(/)
13.1. Sampling Inspection of Continuous Production 385

specify any two of the three and read the third from the graph. The
usual thing, however, is to decide upon the desired AOQL and a sampling

proportion f, for example 10%. Then read out the required number i to

qualify for sampling inspection. For example, iff= 12.5%,or one out of

eight, and the AOQL 2% we find i = 48. Or if f = S%, or one out of 20,

with the same AOQL 2%, then we find i=76. That is, if we are going to

sample inspect fewer pieces once onto sampling, we must make it harder

to qualify, if we are to have the same AOQL protection. Thus for fixed

AOQL, i and f are somewhat inversely related. Take note that great care

must be used in reading the logarithmic scales in Fig. 13.1.

With the CSP-1 plan there is really no operating characteristic

curve in the sense of those in Chapters 9 and 10. Thus we cannot speak of

an "acceptable quality level," AQL for the plan. But we can concern our-

selves with the proportion of the time we are on sampling inspection, that

is, with the percent of the total production accepted on a sampling basis.

Also we can describe the probability of remaining on sampling through-

out a production run of 1000 units. See the note in the upper right corner

of Figure 13.1 and the right hand scale. This pt is not analogous to pt in

lot by lot acceptance sampling, Section 10.2. Here pt is that

fraction defective for which the probability is only .10 that the next

1000 units produced will all be accepted while under sampling. Of these

1000, we shall only inspect lOOOf/(100%) K of them. Then the binomial

probability of finding no defectives among the K and thus of not going

back to 100% inspection is

(13.1)

Equating this to .10,

we may find pt by
.lO = (l-p )lOOOf/(100%) (13.2)
t
386 SOME OTHER SAMPLING PLANS

Note that pt is a function of f only, not i nor AOQL. If f = 5% we solve

.10 = (1-pt) so , using logarithms, for pt = .0450 or 4.5%.
Reference [1) gives appropriate formulas for finding the proportion of
the time on sampling and 100% inspection from which the AOQ's and AOQL's
may be obtained.
A practical problem arises as to how to sample at a rate of f%. Using
the reciprocal of f, we may have to sample one out of five, eight, ten or 20.
Then it is a good practice to choose which of each five, eight, ten or 20
consecutive pieces is to be inspected, by using a table of random numbers

such as Table IV. This avoids bias.

Another practical problem lies in the shifting inspection load. If
for example f = 10%, changing from sampling to 100% inspection involves
quite a jump in load. One possibility which is used is to have the
Production department take over the 100% inspection whenever it is called for.
Yet another practical matter is that of using CSP-1 on more than one

class of defects. For example, we may wish to have AOQL's of .5% for
major defects and 2.0% for minor defects. For this it seems good
practice to use the same value of f for both classes of defects, and let
the difference in treatment come via i, the qualification number. It is

then quite easy to use the same sample of units to inspect for both
classes of defects. For example, if f is 10% in the above we would find
the i values of 220 and 55. Once qualified on one or both, we take one unit
randomly out of each ten, for inspection for the defects qualified for

sampling.
Properly used, CSP-1 can be utilized for outgoing quality control,

and process control as well as by bringing attention to defect-producing

conditions when they occur.

13.1. Sampling Inspection of Continuous Production 387

As proven in [1], if p' is the process fraction defective, then

for given i and f, one has
u = average number of units inspected following a defective, till qualifying
i
1-(1-p')
i (13.3)
p' (1-p')

v = average number of units passed while under sampling

1/fp'. (13.4)
It can then be shown that the average fraction of the units which are

passed without inspection (that is, while on sampling inspection at a rate f)

is

1-F = 1 - u+fv . (13.5)

u+v
This proportion passed, for fixed i and f, is thus a function of the
process p'. It of course decreases asp' increases. It would seem
that the use of CSP-1 would tend to force process correction until p' is

sufficiently low to give 1-F above some critical value, say, 80% to 90%.

That is, p' should be improved until the great majority of the time
the units are on sampling inspection. (The AOQ will then be well below
the AOQL of the plan.)

Further discussion of CSP-1 is given in Dodge [2].

13.1.2. Modifications of CSP-1. Two variations of CSP-1 were
developed to reduce the frequency of alternation between 100% sorting and

sampling, of continuous production. These are given in Dodge and Torrey [3].
In the first of these, CSP-2, the qualification procedure is the same as
in CSP-1, that is, i consecutive good units must be observed under 100%
inspection to go onto sampling. However, when a defective unit is found

under sampling, 100% inspection is not reinstituted at once. Instead,

careful note is made of the ensuing sample-inspected units (still at the

388 SOME OTHER SAMPLING PLANS

rate f). As soon as any of the next i sampled units proves defective,

100% sorting is reinstituted. But if no defective occurs within the

next i sampled units, then the product is again "invulnerable" to a single

defective, having requalified, so to speak, through the i consecutive good

sampled units. If at any time after this run of i units, a new defective

occurs, a careful watch of the next i sampled units is made. A second

defective within these i causes the plan to revert immediately to 100%

inspection, and so on.

Now is not this a more lenient plan for given i and f, than is CSP-1?

Yes it is, because it will stay on sampl1ng inspection for longer periods,

thus reducing the protection and giving a higher AOQL. Hence in order to

obtain the same AOQL for the same frequency of sampling f, we must increase

i.

For example from [ 3] :

AOQL • 5~0 l. 0% 2.0%

f 5% 10% 20% 5% 10% 20% 5% 10% 20%

i CSP-1 305 220 142 151 108 71 76 54 36

i CSP-2 390 293 200 194 147 100 96 72 50

This brief table shows the price we pay for greater freedom from alternation

between sampling and 100% inspection. In these cases, it is roughly a

third more pieces required to qualify for sampling.

Another variation from CSP-1 is called CSP-3 (3]. Its objective is

to provide additional protection against a sudden run, perhaps brief, of

bad quality, that is, "spottiness" of production. It is probabilistically

identical to CSP-2, when inspecting units from a controlled process. It

starts off with the same qualifying number i as in CSP-2. Then it goes onto
13.1. Sampling Inspection of Continuous Production 389

sampling at the specified rate f. Whenever the first defective occurs,

the very next four units produced are all inspected. If any of these four

are defective, 100% inspection is reinstituted. If none of the. four is

defective, then these count as the first four of the i units all of which

must be good ones to remain on sampling inspection. (The choice of units

to inspect after the four proceeds as usual at a rate f.) Then if the i

units are all good ones, a new defective found does not cause reversion to

100% inspection. This occurs only if one is found in the next four units

produced or in the i-4 succeeding sampled units. And so on.

Thus we use the same i and f for a given AOQL in CSP-3 as in CSP-2,

but in usage have the additional protection against spottiness of production.

13.1.3. Other Plans. One variation on continuous sampling plans is

to use more than one sampling frequency. Such a plan is aptly called multi-

level continuous sampling. It proceeds as follows: After qualifying by

i consecutive good units, then sampling is instituted at a rate f. Then

if the next i units inspected under sampling are all good, the rate of

sampling goes to f 2 . If the next i inspected at this rate are all good the

rate goes to f 3 , and so on. But if there occur conditions while sampling

at a rate fk, which under CSP-1, CSP-2 or CSP-3 would call for a return
k-1
to 100% inspection, then under the multi-level plan the rate becomes f ,

that is, one level of a greater proportion sampled. Or if at f already,

then 100% inspection is reinstituted, requiring requalification. Such

plans may be designated CSP-M.

Another variation making use of "production intervals," operates like

CSP-1 (or CSP-2), but has a maximum number~ of defectives to be found in

any one production interval, including both 100% and sampling inspections.
390 SOME OTHER SAMPLING PLANS

There is a military standard [4) which makes use of CSP-1, CSP-2,

CSP-A and CSP-M plans. In the first two there is a maximum number of

units L, before the completion of which the plan must have permitted sampling

inspection, otherwise inspection is terminated pending improvement by the

producer. The standard has specified AOQL 1 s. But in an attempt to work

closely with MIL-STD-1050 it lists corresponding AQL 1 s as index values.

This seems to be a misnomer, since there really is no p' which can properly

be called an AQL. If one finds the p' at which, say, 95% of the pieces

will be accepted on a·sampling basis, this might be a useful thing to list

for the AOQL plan. But such a p' can hardly be called an AQL in the sense

of lot sampling (Chapters 9 and 10).

The standard is currently being revised as this is being written.

It is hoped that the above defect will be remedied, and also that the

standard will be much simplified and streamlined. Moreover, as it currently

stands, there is inadequate guidance for the user in the many needed

decisions. Supervision of usage of the standard would seem to be very

difficult.
The author is not greatly impressed by the practicability of multilevel

plans. They require low levels of f (1/2 or 1/3) in order to avoid

extremely low fractions on higher powers of f. This makes the saving

2
small at for even f , and the multi-levels make scheduling of work loads

difficult. But undeniably multi-level plans are interesting and mathematically

pretty. Multi-level plans are the outgrowth of work by Lieberman and

Solomon [5].

13.2. Chain Sampling Plan, ChSP-1. In the case of lot by lot sampling,

where small samples are called for and the acceptance number is c = 0, we

obtain rather undiscriminating OC curves. Moreover, with an acceptance

number of zero, the curve shape is always concave up like those in Figure
13.2. Chain Sampling Plan 391

9.7, instead of having an inflection point such as the curves in Figures

9.2 and 9.6. The c=O curves thus make it difficult to pass consistently,

even quite high quality (low p') lots. The chain sampling plan was devised

to obtain a more discriminating OC curve by cumulating the results from

several lots. It was developed by Dodge [6]. Mostly the plan gives

greater probability of acceptance of lots at relatively low p' than does

the c=O plan, yet giving about the same protection against accepting lots

at relatively poor p''s.

The conditions and procedure are given in [6] for using ChSP-1:

"1. Conditions for Application

(a) Interest centers on an individual quality characteristic that

involves destructive or costly tests, such that normally only

a small number of tests per lot can be justified.

(b) The product to be inspected comprises a series of successive

lots (of material or of individual units) produced by an essentially

continuing process.

(c) Under normal conditions the lots are expected to be of essentially

the same quality (expressed in percent defective).

(d) The product comes from a source in which the consumer has confidence.

"2. Procedure

(a) For each lot, select a sample of n units (or specimens) and test

each unit for conformance to the specified requirement.

(b) Acceptance number of defects, c=O; except c=l if no defects are

found in the immediately preceding i samples of n. (i=l,2,3 ... )"

Hence a lot is accepted if the sample of n units shows no defects,

or rejected if two or more defects. If there is only one defect, then

the lot can still be accepted if no other defect has occurred within the

samples for the immediately i preceding lots. Thus under the good history
392 SOME OTHER SAMPLING PLANS

of no defects in the preceding i small samples, these count somewhat as

though they were from the current lot. But if one of these samples contained
a defect, then the enlarged sample (of i n's) contains two defects and calls
for rejection of the current lot.

Our Figure 13.2 is reproduced from [6]. All of the ChSP-1 plans have

the elongated S shape characteristic of sampling plans with c > 0, showing

a point 'of inflection. Figure 13.2 is useful in picking out an appropriate
OC curve. In the revision of MIL-STD lOSD, the provision of ChSP-1 as an

alternative to ordinary single sampling plans with c=O, deserves consideration

and is a current possibility.
The OC curve has slightly different meaning than in lot by lot
sampling, wherein we can talk of the probability of acceptance of just the
one present lot. This is without regard to what has gone before. Under
ChSP-1, however, the vertical scale, as seen in Figure 13.2 is the expected

proportion of lots to be accepted, given p'. Under this plan if there is

no defect in ~he previous i lot-samples then for the current lot Pa = P(d~lln,p'),
while if there was one defect among the i samples then Pa = P(d=Oin,p').
(There could not be two defects in the i samples, why?)

Dodge suggests that the i=l plan is not a preferred choice. It gives

too little previous history to draw upon, and may well be too lenient.
The author regrets the confusion in the foregoing between "defects"

and "defectives," In the preceding, reading "defectives" instead of "defects"

is likely to be more what is meant, but the original source used "defects."
13.3. Skip-lot Sampling Plan, SkSP-1. The basic idea of this plan
is to authorize omission of receiving inspection altogether on some fraction
of the lots being submitted in a series, when quality in the past has been

sufficiently good and consistent. It is an application of continuous

sampling plan CSP-1 to lots (rather than to individual units), Dodge [7].
 13.3. Skip-Lot Sampling Plan 393

100 100
Ch5P-TPio.n. ,:)ample- !>i 3"'' 4 Ch !IP·I Plo.n :'lo.o-nple !Jjj.,: 5
l::. t l•i.
90
:t _""'"'*"'"""" Cri-l<.ri<>m
c=O, exc.epi: ~,.1
90
2.
Ac..c.ep~onc.c. Cri~c.ri"n:
c.:~ o, e"'-ept- c:. ~ I
3
80 i.f no in
d.efe~!-s.
80 if d.efec:.l-.& in.
110
+ immedia~el't pre•
4-
immedia+el'l pre·
'_5 5 c.eding f. :xtmples
c.ed.ini ~ 60.tt'tple:a
70 of 4-. 70 of 5.

60

50 ~0

4-0 !I !I Pfo.."-
n.=51 c.:O

.30

., 20

10

10 40 00 10 zo .30 -4-0

~-J~I -1=t:
'•-t Ch5P·l Plan '
o+ -~l'f"r I= '
I !>ample- Si3e: ~
100
~ -l H-H++-
lt C.h ~P-1 Pia." jf
_L
!l'""rle. ~;je: 10
c.. ~ *''rion :
A~;.c.eph..nc.c.. C..ri~eriot\: 90
~~t\l!~ l=t2, Ac.c.epto.nc.e.
o, OOX<.<Z.p~ c.= I
0 + ~~~
~ c.;01 cz."cepi- c.-=1 I ..,
<.=
r-/- 3 if 1\0 defec.h in . ' .. H:.3 if no cl~fe.c.t~ in.
..J_ :I
eo
i*' "
± ""5 ;m!'"cdia.hd'1 pre~
<.edin.!! i t~oomp(es.
'' .!++
·s
imm<.dia.+el'l pre·
c.cdi11g L sa.mriC~!o
0 ~ Qf ... 70 Qf 10.
-
0
I ' I
60
'
I

50 50
!I~Pia.n.
" ...,,.o 4-0
~~Pl.,--.
,_ n.=IO,c.•O
H- ..
-
t.:l=r r--
I
''
,-i 'I

:+-
30
r~-- 30
f-1 I'

-,- . --!- - !
:
1-' 20
20 -~+l "1 -'1
I
I·; t+-:-,
h 1-H1'-.- _, I

.$!
10 '
-~h
I
f- .
't+ + t=lt
10 20 30 40 co 15 40
PI=IOOUc.r QUALITY IN

FIG. 13.2. OC curves for ChSP-1 Plans with values of i from 1 to 5, For
comparison single sampling plans (n = 4, 5, 6 and 10, with c = 0) are shown.
Reproduced from reference [6) by kind permission of the American Society
for Quality Control.
394 SOME OTHER SAMPLING PLANS

It can be applied to either attribute inspection of a lot, or to measurement

characteristics as, for example, chemical or physical analyses of bulk

product. Under the system each lot or batch is examined or tested against

a specific requirement. This could be, for example, in attributes inspection,

relative to some single sampling plan n, Ac-Re; or for a measurement

characteristic the value lying within, or outside, of a specified range of

approved values. The latter could also be on the basis of an x, or x and

s, decision.

A series of lots or batches is thus examined until i in succession

have been passed. One then goes onto sampling, examining a fraction f of

the lots. As soon as a lot is rejected, the plan reverts to examination

of every lot until again i in succession are all passed.

Under SkSP-1, the meaning of the AOQL is somewhat changed (7]. In

the continuous plans, AOQL referred to the average fraction defective of

individual units, including sampled and sorted units. This AOQ was a

function of the incoming p' for units. In the skip-lot plan the AOQL refers

to the proportion of lots conforming to requirements in the outgoing lots,

after rejection (and possibly correction) of some of the lots, that is,

those examined and judged to be non-conforming. Now the AOQ refers to the

proportion of lots, which when examined, would be passed. In either plan

CSP-1 or SkSP-1, the AOQL is the maximum of all AOQ's for whatever the

incoming quality level might be.

There are two cases [7]. Procedure 1 calls for reprocessing or

correcting each rejected lot until it is conforming, or replacing the

rejected lot with a conforming one, then entering it into the stream of

outgoing lots. For this we use the value of i in a CSP-1 table (Figure

13.1). In Procedure 2, however, rejected lots are removed and not

13.4. Cusum Charts 395

corrected nor replaced. Then to qualify we must have i+l accepted lots

in succession to yield the AOQL protection specified.

Some plans given in [7] are:

i
AOQL f Proc. 1 Proc. 2

Standard Plan

2% 1/2 14 15

Other Plans

3% 1/2 9 10

5% 1/2 5 6

5% 1/3 9 10

5% 1/4 12 13

Careful consideration should be given to the legal and quality control

implications of making no test on a characteristic on some lots. If the.

characteristic can prove at all hazardous, if a lot is non-conforming,

use of the plan seems doubtful or dangerous. But when the characteristic

is not so critical, production is steady and the quality level consistently

high the plan can well be useful.

A number of plans, SkSP-2 by Dodge and Perry [8], are worked out for

the case where inspection is by attributes, for example, n, Ac-Re, and the

probability of lot acceptance is given for i and f. Use of f=2/5 provides

a possible alternative to reduced inspection in MIL-STD-1050, using the

normal plan, but only sample-inspecting two lots out of five when

qualified for reduced inspection.

13.4. Cusum Charts. This is the commonly used name for cumulative

sum charts. The idea is to follow any quality characteristic, say Q, in

relation to some standard value Q0 , by cumulating the deviations from Q0 .

396 SOME OTHER SAMPLING PLANS

Q could be x, x, R, s 2 , d defectives or c defects. We analyze the current

situation through plotting
k
cusum(Q) = I (Q. - Qo). (13. 6)
i=l 1

Then if Q0 is actually the average value of the characteristic Q, the

cusum will tend to remain relatively close to zero, or only drift slowly
away. But if the average value of Q is quite a bit different from q0 ,
the cusum will rather rapidly rise from zero (or go below i f E (Q) < q0 ).

Now, one could put upper and lower limits to the total amount of drifting

from zero. But it is also important to know how long (number of points)

it took to reach such a limit. Thus the steepness of ascent is important

too. This led those who developed the technique to use a V-shaped mask
to make a test after each new point arose. See Figure 13.3. This is the

usual test presently used. See Johnson and Leone [9] and Duncan [10] for

expositions of the approach, and references to the early papers.

The dimensions of the V-shaped mask, specifically 8 and d, are determined

from the a risk of a decision that E(Q) r Q0 when E(Q) is in fact Q0 , and

the S risk of not concluding E(Q) r Q0 when in fact E(Q) = Q0 + D (or Q0 -D).
Also d and 8 depend on the quality statistic in question. The setting up

of the mask and chart is something of a chore and must be done for each new

application.
The cusum chart is in reality a type of sequential analysis, since

it relies upon past data for each decision, which can be (a) conclude
E(Q) > Q0 , (b) conclude E(Q) < Q0 , or (c) continue with new data. Cusum
charts perform the same general test as do Shewhart control charts for the

case "standards given." Thus Q0 is a standard reference value, and for

cusum charts for x or x, cr is needed, or a close estimate of it. Cusum

charts do lliU correspond to Shewhart control charts for the case "analysis
13.4. Cusum Charts 397

8
6
4
2
X
0 X
~QI-Qol
I
-2
-4 X

-6
X
-8
-10
-12
-14

FIG. 13.3. A typical V-mask, showing d and the angles e. After each
point is plotted the mask is placed so that the left end of the d-segment
coincides with the point. If any previous points lie outside the V, a
decision is reached that E(Q) # Q0 at this point. In this figure
E (Q) < Qo·

of past data,'' wherein the aim is to watch for statistical control or

homogeneity, and for assignable causes disrupting such control.

The objective of cusum charts is to obtain an indication of departure

from the standard Q0 , and to reach such an indication quickly. Specifically

we can design a cusum chart which has a similar operating characteristic

(OC) curve to the standards-given control chart in the following sense:

If E(Q) = Q0 then there should be a long "average run length,'' ARL, of

398 SOME OTHER SAMPLING PLANS

points before a false warning is given. For example, on an x chart with

+ 3cr- limits, the probability of a point out is .0027. And thus the ARL
- X
is 1/.0027 = 370 to a false warning. (See Section 7.6.) But if the sample

size for x 1 s is n=4, say, and ~ should shift up lox' that is, 2crx, the

probability of an x point lying above the upper control limit is found

from Figure 13.4 to be .1587. Then the ARL under such a process level shift

is short, namely 1/.1587 = 6.3. Now if we set up a cusum chart to have an

ARL = 370, when E(x) = ~ 0 , the ARL, under a shift in p of +lcrx will be less

than 6.3. Thus the cusum chart is more powerful in this sense. We note

LCLa

FIG. 13.4. Two distributions of x

for n=4. One is centered at ~O
with control limits of ~~3crx. The upper one is centered lcrx = 2crx
higher. It has .1587 of its x's above the upper control limit for x's
from ~ 0 .
13.5. Summary 399

that this comparison is made for the cases (a) where the process stays at
the standard and (b) where the average level jumps to some specified new

level and remains there till apprehended.

For practical application, the worker must decide whether the greater

effort in setting up a cusum chart is worth the gain in power. The present

author is not convinced that it is very often worth the additional

complications.

13.5. Summary. In this chapter we have introduced additional sampling

plans. The first group (CSP) was for sampling control of continuous

production. The plans are applicable to a continuous flow of ~~its or

pieces which can be inspected 100% when necessary. But if quality is

sufficiently good then much of the time sampling inspection can be used,

meanwhile providing protection if quality should deteriorate. The savings

in inspection time can be considerable if,as should be the case,quality is

good enough to stay on sampling 80% to 90% of the time.

Next there were presented two plans for inspection of a continuous

flow Q£ lots. The chain sampling plan (ChSP) has as its objective the

attainment of a more discriminating OC curve than can be obtained by c;O

plans. The plan is to not reject on just one defective if there have

been no other recent defectives (in the last i lots). The application is

to conditions under which very small samples are needed, for example

destructive or very expensive testing.

The other plan for inspection of a flow of lots is the skip-lot plan

(SkSP). It makes use of the principles of CSP-1, but applies them to lots

rather than single units. Under the plan, after a sufficient run, i, of

lots, all of which were accepted, the plan goes to omission of all inspec-

tion in some proportion of the lots. Bolstering conditions would seem to

make the plan usable, unless the characteristic(s) are of critical nature.
 400 SOME OTHER SAMPLING PLANS

The fourth sampling plan was the cusum chart. It has resemblance to
Shewhart control charts for the case, standards given. Also it has close

relations to acceptance sampling for attributes or variables,and to sequential

analysis. The typical application is to a flow of units or continuous

material, inspecting or testing periodically for evidence of a process

change. The objective is to achieve desired discriminating power on ninimum

average run lengths (ARL).

PROBLEMS

13.1. Find a CSP-1 plan for the following cases, filling blanks:

i f AOQL

10% 3%

50 5%

2% 7%

60 25%

13.2. Find a CSP-1 plan for the following cases, filling blanks:

i f AOQL

10% 1. 5%

100 2%

200 • s~0

• 3~0 4%

13.3. Find a CSP-1 plan for f=lO% AOQL=l.S%. What would you expect

i to be for the corresponding CSP-2 or CSP-3 plan?

13.4. Find a CSP-1 plan for f=20%, AOQL=.4%. What would you expect

i to be for the corresponding CSP-2 or CSP-3 plan?

13.5. Explain why i and fare inversely related for fixed AOQL, in

practical terms.

13.6. Why is pt only related to f and not to i nor the AOQL in CSP-1?
References 401

13.7. If two classes of defects for the same piece are to be put
under CSP-1 plans, with f=l2.5%, find the respective i's for AOQL=l% and
3 ~0. Why not use the same i instead of the same f?

13.8. What administrative difficulties would you foresee in using a

multi-level continuous sampling plan? What is the potential gain?

13.9. Consider chain sampling plans ChSP-1 for n=4. What is the

percent of lots expected to be accepted, using i=2 and i=5, p'=5%?

Compare these with the probability for the single plan n=4, c=O. Do the

same at p'=20%.
13.10. Consider chain sampling plans ChSP-1 for n=lO. What is the

percent of lots expected to be accepted, using i=2 and i=5, at p'=2.5%.

Compare these with the probability for the single plan n=lO, c=O. Do the

same at p'=lO%.
13.11. In what sense does ChSP-1 give a more favorable OC curve?

13.12. Consider a skip-lot sampling plan for AOQL=2%, for f=l/2.

Explain the meaning of the two i's and of the AOQL.

13.13. Find a SkSP-1 for AOQL=2~ and f=l/3. For f=l/4.

13.14. Find a SkSP-1 for AOQL=l% and f=l/2. For f=l/3.

References

1. H. F. Dodge, A sampling inspection plan for continuous production.

Ann. Math. Statist. 14, 264-279 (1943).

2. H. F. Dodge, Sampling plans for continuous production. lndust.

quality Control 4 (No.3), 5-9 (1947).

3. H. F. Dodge and M. N. Torrey, Additional continuous sampling inspection

plans. Indust. Quality Control 7 (No. 5), 7-12 (1951).

4. Military Standard MIL-STD-1235 (ORD), "Single and Multi-level Contin-

uous Sampling Procedures and Tables for Inspection by Attributes."
Department of the Army, 1962.

5. G. J. Lieberman and H. Solomon, Multi-level continuous sampling plans.

Ann. Math. Statist. 26, 686-704 (1955).
402 SOME OTHER SAMPLING PLANS

6. H. F. Dodge, Chain sampling inspection plan. Indust. Quality Control

11 (No. 4), 10-13 (1955).
7. H. F. Dodge, Skip-lot sampling plan. Indust. Quality Control 11
(No. 5), 3-5 (1955).
8. H. F. Dodge and R. L. Perry, A system of skip-lot plans for lot by
lot inspection. Amer. Soc. Quality Control, Ann. Tech. Conf. Trans.
1971, 469-477.

9. N. L. Johnson and F. C. Leone, "Statistics and Experimental Design."

Vol I. Wiley, New York, 1964.

10. A. J. Duncan; "Quality Control and Industrial Statistics." Irwin,

Homewood, Illinois, 1965.

Additional References

H. F. Dodge and M. N. Torrey, A check inspection and demerit rating plan.

Indust. Quality Control 13 (No. 1), 5-12 (1956).
E. G. Schilling and H. F. Dodge, Procedures and tables for evaluating
dependent mixed acceptance sampling plans. Technometrics 11, 341-372
(1969).
K. S. Stephens, A Multi-level AOQL sampling plan with weighted defect
classification. The Engineer, Western Electric 6 (No. 4), 39-49 (1962).
 CHAPTER 14

STATISTICS OF COMBINATIONS, TOLERANCES FOR MATING PARTS

14.1. Introduction . Virtually all individual piece parts are made to

be combined with other parts. The s ame is basically true of processed

materials, such as, chemicals, plastics, rubber, glass, pharmaceuticals and

many food products. The main thing of interest is the distribution of the

important characteristics of the assembly or combination of the components,

not so much the distributions of the characteristics of the components. You

see we are here talking in terms of distributions rather than individual

values, just as we have been stressing throughout this book . Of course the

raw materials, operators and production processes must combine in such a way

as to provide a s atisfactory distribution to ea ch characteristic of the

components. We must deal in terms of distributions, since we cannot produce

components all exactly alike. There will always be some variation, if only we

measure precis e ly enough.

In a sens e the traditional design engineer is determining the distribution

of the characteristic of a component when he sets minimum and maximum

specifications, 1 and U, on each such characteristic. This is still usually

being done in a quite conservative manner, so that if each component part is

anywhere between its s pecified limits, the assembly chara cteristic will be

satis factory . For example, the " s tack-up" or sum of f our dimensions a dded

together might be the assembly characteristic in question. Suppose this sum

is to lie be tween 1 and U. Then the traditiona l approach says set 1 1 , ... ,14

so that i f all your components are at the se respective lower limits then

403
404 STATISTICS OF COMBINATIONS

1 1 + 1 2 + 1 3 + 14 = L, and similarly, for u1 + u2 + u3 + u4 = U. Now what

could be more logical or safe? What is wrong with this approach?

The main thing wrong with such an approach is that it completely ignores

probabilities and distributions. What is the probability of the four

components being right at the four lower limits 1 1 , 1 2 , 1 3 , 1 4 so that

their sum will be right at 1? It is extremely remote! (In fact theoretically

zero.) Let us consider that the first dimension x 1 lies somewhere in the

bottom 107o of the interval 1 1 to u1 and similarly for x 2 , x 3 , x4. Then y

xl + x2 + x3 + x4 lies in the bottom 10% of the interval L to u. Also

suppose that, very conservativel~we assume that the distributions between L•

1

and U. are rectangular, that is, the probability "mass" is uniformly distributed
1.

between the limits. Then there is only one chance in 10,000 of all four

components being simultaneously in their respective bottom 10%'s, which gives

,.,
the sum y in the bottom 10% of L to U. If the components' distributions

meet the specifications Li to Ui in a + 3a sense, with normal distributions,

then the probability for y in the bottom 10% of L to U is far rarer than

.0001 (actually about ,000,016). This is due to compensating variation: if

one component is low, others may well be high.

The objective of this chapter is to explain some of the theory of

combining distributions and to show how to make use of these laws in practic~.

14.2. Statistics of Sums and Differences. We may prove the following

facts for simple sums and differences. Let

y = x1 ± x2 ±···± ~· (14.1)
fork component dimensions or other characteristics. y is the assembly or

combination characteristic we wish to control.

"}(
There are other ways .Cor y to be in the bottom 10% of L to U. If the

intervals Li to Ui are all equal, the exact calculation yields a probability

of about .0011.
14.2. Statistics of Sums and Differences 405

Thm. 1. If each xi component has finite mean ~i' than for (14.1)
(14. 2)

This theorem holds whether or not the xi's are independent, or normally distributed.

For variabilities, however, we find it desirable to assume independence, that is,

the random variable (component) xi is independent of all of the other xj's;

knowledge of the latter tells nothing about the former.

Thm. 2. If each xi component has finite standard deviation cri and the

x 1 , ... ,~ are all mutually independent then

(14.3)

This theorem says that a is in general much less than the sum of the separate
y
ai's, but this requires independence of the xi1 s. If the xi's are not

independent then (14.3) is affected by the so-called linear correlation

coefficients. We then have as the very extreme limits:

(14.4)

Also to be noted in (14.3) is that although (14.1) has ~·s, (14.3) has +'s

only.

Next, we have, besides the mean and standard deviation of y, its

distribution shape. Theorems 1 and 2 made no assumption of normality of x's.

We have the following theorem.

Thm. 3. For y defined as in (14.1), and the x.'s normally distributed,

1

then y is also normally distributed.

Here the xi's need not be independent.

Now suppose that instead of (14.1) we have

(14. 5)

Then under the same corresponding hypotheses as in Theorems 1 to 3, we have

(14.6)
406 STATISTICS OF COMBINATIONS

(14. 7)

y is normally distributed.
If the xi 1 s are not normally distributed in (14.1) or (14.5), then the

distribution of y is not normal, but tends toward normality as k increases,

2
provided none of the contributions to cr in (14.3) or (14.7) predominates.
y
This is from a more general central limit theorem than that for component x 1 s

from the same distribution.

14.2.1. ~' Derivation of (14.6) and (14.7). Consider first the case of

k =2 in (14.5). Then we must have been given the joint density function of

x 1 and x 2 , that is, f(xl'x 2). The "marginal" density functions are found by

integration:

r
_co
(14. 8)

Then the respective means are given by

,co

1-11
-co

(14.9)

where E means "expected value of" or "theoretical average of." Also

Then too

]11-]111=0. (14.10)

Furthermore the definition of population variance is

14.2. Statistics of Sums and Differences 407

(14 .11)

and similarly for x 2 . Finally consider

E[xl- ~~)·(x2- ~2)] f -00

f -00
(xl- ~l)(x2- ~2)f(xl,x2)dxldx2 (14.12)

This is by definition the "covariance" of x 1 and x2 It is not in general

equal to zero. But if x1 and x 2 are independent then by definition

(14.13)

and we have

roo 00

[J (x1 - !J.1 ) f 1 <x 1 )dx1 J[j (x2 - !J.2 ) f 2 (x 2)dx2 ] = 0·0=0 (14.14)
-ro -oo

Now with this background consider (14.5) - (14.7) with k=2:

as in (14.6). Then if x 1 and x 2 are independent we have

408 STATISTICS OF COMBINATIONS
Breaking up into three separate double integrals we have respectively

yielding (14.7) for the independent case. Now if x 1 and x 2 are not independent,

the first and third terms are the same but the middle term will not in general

be zero, but instead 2a 1a 2 covar (x 1 ,x 2). As usual we may define the

"correlation coefficient" by
covar (x 1 ,x 2)
(14 .15)
0 1°2

which lies between -1 and +1, and which is zero if x 1 and x 2 are independent.

p can be zero when x 1 ,x 2 are not independent, in which case we call them

"uncorrelated." Thus we have for the generalization of (14.7) for k=2:

(14 .16)

Fork= 3, we derive (14.6) and (14.7) by regarding y as the sum of the

two random variables (a 1x 1 + a 2x 2) and a 3x 3 , then use the results fork= 2.

Higher k 1 s are similar .

Note that (14.1) through (14.3) are corollaries with ai = ±1.

14.3. Use of Distributions in Assemblies. Let us consider some examples

to illustrate the formulas of the previous section.

Example 1. Bearing and Shaft. Suppose that we are producing bearings and

shafts to be assembled. In order to insure that the shafts will all be capable

of assembly at random into a bearing the following non-overlapping specification

limits are set:

Inside diameter (ID) of bearing: x 1 .4022" ± .0012" or .4010", .4034".

Outside diameter (OD) of shaft: x 2 .4000" ± .0009" or .3991", .4009".

Taking the appropriate extreme limit~ the diametral clearance

(14.17)
14.3. Use of Distributions in Assemblies 409

would then lie between .0001" and .0043". Now suppose that our production
processes are capable of meeting the respective limits for x 1 and x 2 in a

±3cr sense and with approximately normal distributions. What then can we say

about the distribution of y? Using (14.2) and (14.3) we find

.4022" - .4000" .0022"

.0012"/3 = .0004" .0009"/3 .0003" so that

cr =l<.oo04") 2 + (.0003") 2 = .0005"

y
y approximately normally distributed.

Thus practically all (about 99.7%) of the diametral clearanceswill lie within

.0022" ± .0015" = .0007"' .0037".

Note carefully how much closer together these limits are than the extreme

possible limits .0001" to .0043".

What can then be done? One possibility is to run ~l and ~2 closer

together, so as to decrease the maximum clearance if this is desirable.

Thus if a minimum of .0001" is deemed satisfactory then we might use

.0016" ± .0015" = .0001"' .0031".

and

.4016" ± .0012" = .4004"' .4028" .4000" ± .0009" or .3991", .4009".

These limits must be met in a ±3cr sense and with approximately normal

distributions. Notice that they overlap somewhat but that taking into

account the distributions, 99. 7"/, of t:he pairs of bearing and shaft being

assembled would have clearance between .0001" and .0031".

Another alternative is to use the full limits .0001" to .0043", if these

limits are indeed permissible. To meet them in a ±3cr sense we can say that

for clearance y we can let

~
y
= .0022 11

and since the limits are .0022" ± .0021", we can let

410 STATISTICS OF COMBINATIONS

3o
y
= .0021" oro
y
= .0007".
Now suppose we decide to allocate o 1 (4/3)o 2 . Then using

(14.3)

.0007" = j(4/3) 2 o~ + o; (5/3)02

or

0 = .00042" 01 ~ .00056".
2

Then we could use specification limits for the components (if we can make

sure of their being met in a ±3o, normal curve sense) of

. 40220" ± . 00168" .40052"' .40388"

. 40000" ± .00126" . 39874"' . 40126" .

Again note that there is overlapping in these two sets of tolerance. Further

we probably should not be setting specifications to such high precision

(.00001").

Gains such as this of permitting 40% greater tolerance can mean substantial

savings in production, inspection and rework costs, not to mention scrapping

costs. But how can we safely obtain this gain? The answer is by controlling

the distributions of dimensions through use of measurement control charts

and/or acceptance sampling by measurements. See Section 14.4 for one recent

systematic approach. The watchword then is: Engineering to permit substantially

greater process variability in exchange for guarantees by Production of

adequate control of process means. Quality Control must work closely with both.

Example 2. Three resistors are to be connected in series so that their

resistances add together for the total resistance. One is a 150 ohm resistor

and the other two are of 100 ohms, carrying the same part number. Suppose the

respective specifications are

150 ± 7.5 ohms and 100 ± 6 ohms.

Let us take, similar to (14.1)

14.3. Use of Distributions in Assemblies 411

Then we set ~l = 150, ~2 = ~3 = 100 ohms. Now suppose that the specifications
are being met in a ±3cr sense by normal distributions. Then cr 1 = 2.5, cr 2 = u 3 =

2.0 ohms. Now what can we say about y? Surely if all three resistors are

inside their specifications, then y will lie inside 350 ±(7.5 + 6 + 6)

350 ± 19.5 ohms. But what of y's distribution? We have

~y = 350, cr
y
=~2.5 2 + 2.0 2 + 2.0 2 = 3.77, y is normal.

Thus very nearly all of the total resistances y will lie inside

350 ± 3(3.77) = 338.69, 361.31.

Note that these are far narrower than 350 ± 19.5 = 330.5, 369.5. In fact these

latter limits are "· + 5.17a . It is therefore apparent that if 350 _+ 19.5 is
~- y
all that is required on the total resistances y, then the specification limits

for the component resistors can be considerably widened. In fact we could use

150 ± 7.5(5.17/3) = 150 ± 12.9 and 100 ± 6.0(5.17/3) 100 ± 10.3.

But this is true only if we will meet such limits with approximately

normal distributions and in a ±3a sense.

In practice what often happens is that when the assembly characteristic

requires specifications like the ±19.5 limits, then this 19.5 is split up

into pieces whose~ is 19.5, for example, here into 7.5, 6.0 and 6.0.

Limits so set, which are unnecessarily tight, may well be beyond the process

capabilities, requiring expensive 100% sorting, reworking and scrapping,to be

met. On the other hand if we use the tolerance of T = 39 for the sum, y, and

set

(14.18)

then the component tolerances, Ti, can be much larger than the 2(7.5) = 15,
2(6.0) = 12 and 12. In fact as we have seen they could be 2(12.9) = 25.8,
2(10.3) = 20.6 and 20.6. It could well be that the production process could

have the capabilities of producing to these tolerances in a ±3a sense, with

412 STATISTICS OF COMBINATIONS

reasonable normality; thus by suitable process control all 100% inspection,

rework and scrap are eliminated. At least it would be much easier for

Production to meet these limits which are 72% wider.

But to make use of (14.18) the Engineering Department must be given

assurance that the Production Department will exercise control of the process

means of the components to be close to the nominal and that the distribution

will be reasonably normal. The latter tends to mean that the limits are not

to be met only by heavy truncation through 100% inspection. The next section

gives a set of process controls and acceptance sampling plans which, if used,

will permit the use of formulas like (14.18) to give maximum tolerances to the

various components.

Another point we can well make here is that it might be tempting to use

y = x 1 + 2x 2

here, as in (14.5) rather than

Y = xl + x2 + x3.
This first formula would give~
y
= 150 + 2(100) 350 all right, by (14.6).

But for cry we would have from (14.7)

a y2 = 2.5 2 + 22 (2.0) 2 = 22.25 or ay = 4.72.

This is larger than the 3.77 we found previously. The reason it is larger is

that it basically assumes that if one "100 ohm resistor" is 105 the other one

being assembled in the circuit will also be exactly 105 too. That is, it

assumes perfect correlat:i.on in the two. There might in fact be some

correlation leading to 0 a bit above 3.77, but not as high as 4.72.

y
Example 3. In the author's consulting experience a problem of additive

tolerances arose in which 10 dimensions (inside and outside diameters, pitch

diameters and eccentricities, that is, distances between center-points) added

and subtracted, determining whether the parts could be assembled. The

company had available frequency distributions of 300 observations on components.

14.3. Use of Distributions in Assemblies 413

These gave a clear picture of the process capabilities, since control was quite

good, as was normality. The desire was to double the tolerance on a pipe

thread pitch diameter, so as to eliminate a second, finer threading operation.

By direct use of (14.2) and (14.3) the author showed that there was not one

chance in 10,000 of a collection of parts being unable to be assembled. This

would only occur with all ten dimensions being at their very extreme wrong

directions for the parts to fail to assemble. When these facts were presented

to the government agency, the approval for the relaxed specification came back

in four days. This application saved 8¢ on each of an order of 4,000,000 assemblies.

Example 4. An engineer in the automotive industry made a thorough study

of a 12-volt motor for starting an automobile. He found that 38 dimensions

added and subtracted along the rotor shaft determining the amount of

"end-play" the rotor would have after assembly. Just for illustration let us

suppose that a total tolerance of .038" could be permitted, and that each of

the 38 dimensions would be equally hard to hold. Then the purely additive

approach would say "Give each dimension a total tolerance of .001" and set

the nominal mean dimensions so that they combine to give the desired average

to end-play." The nominal mean is (U+L) /2. Now suppose that the production

processes were each capable of meeting ±.0005 in. in a ±3o sense, with

reasonable normality. Then for each,

3cri = .0005 in. .000167 in.

Thus (14.3) gives

cr d = A8(.000167 in.) 2 .00103 in.

en -p 1ay
or

T = 6o = .00618 in.
end-play end-play
Compare this with the permissible tolerance of .038 in! It is only a sixth

as great. So although we can permit .038 in. tolerance in the assembly

characteristic we have forced a far smaller variation by our use of pure

additivity of tolerances.
414 STATISTICS OF COMBINATIONS

What size of equal tolerances could we use? They could be J38 ~ 6.16,
that is, about six times as great each. Then each individual tolerance Ti

would be .006 in. or ai ~ .001 in. Then

a d = v1s(.001 in.) 2 = .00616 in.

en -p 1ay
or

T 6a d
en -p 1ay
= .037 in.
end-play
But, and this is extremely important, the averages of the production distributions

would have to be held close to the various nominals (Li + Ui)/2. Normality should

be easier to attain, since the excessively tight tolerance might well be beyond

the production capabilities, thereby forcing 100% inspection.

Now in practice, undoubtedly some component part characteristics would be

more difficult to hold than others, and therefore should be given greater

tolerances. But this can be done subject to (14.2) and (14.3). Also some

mechanism is needed to insure that the means are properly controlled,

otherwise the whole approach is not very safe. Such an approach is given in

Section 14.4.

By using such an approach we can often enlarge the tolerance on those

dimensions which are the most difficult to hold. In this way it is often

possible to (1) eliminate 100% sorting thereby saving inspection and parts,

(2) use a cheaper production process, or (3) avoid the necessity of an extra

operation. In these ways huge savings are possible in industry.

14.3.1. The Rectangular Distribution. It is sometimes useful to assume

that the x's are uniformly distributed between the specification limits, as a

rather extreme alternative to the normal curve with ±3a just meeting the limits.

The rectangular distribution is defined by

 14.3. Use of Distributions in Assemblies 415

f(x) 1-
= -- L < x < U Rectangular distribution (14.19)
U-L
= 0 elsewhere.

For this distribution we have

_ U+L
I-LX - 2
Rectangular distribution (14.20)

U-1
crx =- Rectangular distribution (14. 21)
2/3

Thus for the rectangular distribution filling the interval from L to U

uniformly, the range U-1 is only 2/Jcrx = 3.46crx' instead of 6crx as in the

normal distribution (in a -+3crX sense). And so for component distributions

meeting limits 1 to U and centered at the nominal (U+L)/2, the range U-L might

be thought of as running between 3.46 and 6 cr 's.

X

Now suppose we have four components, perhaps resistors with their

resistances each meeting 100 ± 6 ohms, and connected in series so they add.

,.,
For the density function (14.19) we have

1-Lx expectation of x E(x)

ru U-Ldx
= J1
X
=
x2
2 (U-L)
JuL ~ u22(U-L)
-L 2 U+L
= --2--

2 2
Expectation of x = E(x )

Since it is easily shown that

cr 2
X
= E(x 2) - [E(x)] 2 . (14.22)

we have
2
Ci (U-1) 2 /12
X

or

Ci
= U-1
X
2/3
416 STATISTICS OF COMBINATIONS

If they meet the limits with a rectangular distribution, each would have

aX = 12/2/3 = 213 ohms.

6.93

So that y is meeting

400 ± 3(6.93) = 400 ± 20.8.

Note that this is in a ±3a sense with good normality by the Central Limit

theorem. For just four components it is already closer than the purely

additive tolerance

4(100) ± 4(6) = 400 ± 24,

even though we used very conservative distributions for meeting limits. The

gain is greater if k is larger, even for these assumed very conservative

distributions.

Consider once again Example 4, on the automobile motor. Again suppose

each of the 38 dimensions is given a tolerance of .001 in., but that now the

range Li to Ui is met no better than with a rectangular distribution; a very

conservative assumption. Now, what limits is the end-play y meeting? By (14.21)

ai = .001 in./213 = .000289 in.

so that by (14. 3)

a d
en -p 1ay
= ~8( .000289 in.)
2
.00178 in.

or

T = 6(.00178 in.) = .0107 in.

end-play
This is less than a third of the .038 in. tolerance the design engineer was

assumed to need, despite the conservatism of the assumption.

Reversing the problem as before

6cr = 6J'38cri2 = .038 in.

end-play
yielding
14.4. Tolerancing Assemblies 417

ai = .038 in./(6/38) = .00103 in.

Then using Ti = Ui - Li = z/3ai = .00357 in.

or about 3.6 times as great a tolerance as the purely additive tolerances yield.

Again we must emphasize that the gain comes about in considerable part only

by controlling all distributions to average reasonably close to their nominals,

(Ui + Li)/2. Also assumed is independence of the xi for (14.3), which is

quite likely to be the case in industrial production, certainly so if the

parts are produced under statistical control.

14.4. Tolerancing Assemblies by Specifying Process Controls and

Acceptance Sampling. This section is the outgrowth of a series of works

[1]-[6]. These assuredly are not to be thought of as the only potential

references. G. Bennett [7] lists 187 references, as of 1964, most of which

bear on the general subject of statistics of combinations.

Suppose that virtually all of the distribution of the assembly characteristic

is to lie within ~O ± (T/2) = L, U, that is, the tolerance is T. This can be

accomplished if L to U is ~O ± 3a and the distribution is reasonably normal.

Now there are several sources of variation in the dimension or other

characteristic of any part: (1) the variation of an individual part

characteristic around the lot mean or a short-term segment of production,

(2) the variation of the ~ of the lots or segments of production as in

(1), and (3) the mean of the means after applying applicable process controls

or acceptance sampling tests. This combined variation is what we wish to

control so that the total variation for the assembly shall be T or less and

centered close to ~o·

Appropriate process controls and acceptance sampling tests were developed

in [3] which led to simple approaches published in [4]. These were further

studied by James in [5]. In this his aim was to obtain sufficiently

discriminating operating characteristics so as to avoid, quite largely, the

418 STATISTICS OF COMBINATIONS

rejection and sorting of satisfactory lots or segments of production while

still having quite low probability of passing material needing screening.

He also showed that the approach did well on fairly non-normal distributions.

In [6] there is a compromise reached between the simplicity and somewhat

smaller sample sizes of [3] and [4] and the greater complications and

somewhat larger sample sizes of [5]. The author recommends the use of

[6], which is herewith summarized.

The general approach is as follows:

1. Set up the equation for the manner in which the characteristics of the

components combine to determine the characteristic for the assembly.

It is usually of the form (14.1) with some combination of signs:

(14 .1)

2. Determine the desired limits for y, the assembly characteristic, 1, U,

with tolerance T = U- 1 and with nominal mean (L+U)/2 = ~ .

y
3. Allocate the tolerances, Ti' for the component characteristics entering

into (14.1) by
r---------------
T = lri + T~ + ... + T~ (14.23)
··k
according to their supposed difficulty of manufacture.

,.,Limits for the assembly 1 and U are to be met in a ±3a normal curve sense,

so that a = T/6. Then in general we have (14.3) for independent components:

a = /oi +a~ +... +a~.

Multiplying both sides by 6 gives

6a = j(6cr1 ) 2 + (6cr 2) 2 +... + (6crk) 2

Defining

6cri = Ti
we have (14.23). (14.24)
 14.4. Tolerancing Assemblies 419

4. Set desired component means ~i so that, following (14.1), we have (14.2)

~y = ~1 ± ~2 ± .. ·± 11k (14. 2)
5. Thus for the i'th component we have the desired nominal ~i and tolerance

T .. These are then used to set up process controls or acceptance sampling

~

plans as below (instead of being used directly as limits in the

conventional sense). By following these process controls or sampling

plans thus determined for the components, all three of the sources of

variation given above will be adequately controlled, so that the

production system will yield assemblies meeting ~

y-
+ 3cr
y
with quite

normal distributions.

6. If production or processing of the component is done nearby, it is

recommended that one of the two process controls be used (which are

much like x and R charts). But if production is done at a distance,

as for example, for purchased parts, then the acceptance sampling

program is recommended.

Acceptance Sampling Program for Components

Program

1. Initially on an order, use Plan A, until five consecutive lots have

been passed, then use Plan B. Continue using Plan B until two lots

out of any five consecutive lots have been rejected for screening,

when Plan A is reverted to for qualification.

2. If a producer has had a satisfactory history of production on this or

similar parts, may start with Plan B.

A. Acceptance Sampling Plan A (Sample Size 30)

1. From the lot draw at random three samples, each of 10 parts, and

measure the characteristic xi for each of the 30 parts. Find the

average x of the 30. Also find three ranges, R, one for each sample

of 10.
420 STATISTICS OF COMBINATIONS

2. The tests for this, the i'th part are determined from the desired mean

~i and tolerance Ti' as follows:

a. Accept the lot if both of the following are met:

i. ~i - .13Ti ~X~ ~i + .13Ti (14.25)

ii. Rl + R2 + R3 ~ 1.45 Ti (14.26)

b. Reject the lot i.f either or both requirements in Section a are not

met, and screen the lot 100% to limits of ~i ±(3Ti/8). (Some parts

may be reworkable.)

B. Acceptance Sampling Plan B (Sample Size 10)

1. From the lot, draw at random a sample of 10 parts, aud measure the

characteristic xi for each of the 10. Find the average x and range R.

2. The tests for this the i'th part are determined from the desired mean

~i and tolerance Ti' as follows:

a. Accept the lot if both of the following are met:

i. ~i- .174Ti ~X~ ~i + .174Ti (14.27)

ii. R ~ .521Ti (14.28)

b. Reject the lot if either or both requirements in Section a are not

met, and screen the lot 100% to limits of ~i ±(3Ti/8). (Some parts

may be reworkable.)

This sampling acceptance program will adequately control the distribution

of outgoing product (including accepted and screened product) regardless of

incoming distributions. If used on each of the parts in an assembly, the

requirements for the assembly characteristic will be met.

Process Controls - Introduction

The process control procedures are provided to assure that the

variability of outgoing parts is not too great relative to Ti' and that the

average is relatively near the desired ~i' Commonly the control takes the

form of adjusting the process average as indicated, provided that the natural
14.4. Tolerancing Assemblies 421

variability in the process is not excessive. Two classes of controls are for

the cases of the production process (1) subject to the average level

consistently changing in one direction, as for example, as a tool or grinder

wears, and (2) subject to erratic jumps in average level. Also two methods

of adjusting the process level are provided for (1) automatic control and

(2) hand control.

Process Control Procedures

Given for the i'th component part characteristic a tolerance T. from

l.

(14.23) and a nominal mean ~i from (14.2). Then use the following rules to

control the distribution of the i'th part:

A. Take samples of n =5 parts periodically from the most recent production,

measure and find mean x and range R for the x's.

B. Approve the process and let it continue if both of the following are met:

1. R _::: .55Ti (14.29)

2. ~i - .16T.l . -
<X<~.
- l.
+ .16Ti (14.30)

C. If R > .55 Ti sort the product produced since the previous sample of

five, l00% 1 to the limits ~i ±(Ti/2), if feasible, and in any case take

steps to reduce the process variability.

D. If x lies outside the limits ~.

l.-
+.16T.,
l.
the process level is to be

reset according to the following cases:

1. Consistent process change, as in tool wear.

a. Resetting fixed amount, as in automatic control:

i. downward by .3Ti if dimension increases, or

ii. upward by .3T. if dimension decreases.

l.

b. Resetting to fixed level, as in hand control:

i. to ~·l. - (Ti/8) if dimension increases, or

ii. to ~i + (Ti/8) i f dimension decreases.

c. 100% inspection is not necessary in this case of Section 1.

422 STATISTICS OF COMBINATIONS

2. Erratic changes in process level

a. Resetting by amount depending upon x, as in automatic control:

i. downward by .S(X-~i) if x> ~i + .16Ti

ii. upward by . 8 (~Li -xi) if x < ~i - .16Ti

b. Resetting to desired level, as in hand control: to ~i as

nearly as possible.

c. Product should be sorted 100% to limits ~i ±(Ti/2) in this case

of Section 2.

Summary of Approach

By use of the foregoing acceptance sampling and process control program,

the design engineer can safely use (14.23) T = ~i + T; + ... + T~ instead of

the very conservative T = T1 + T2 + ... + Tk' because he can have full

confidence that the component means will be sufficiently well held that the

assembly characteristic will meet the specified limits L, U very well. This

will in general mean that with such assurance, he can greatly loosen up on

the variability, anywhere from 40% to 200% or more. This can bring great

savings in inspection, rework and scrap costs, and make relations between

Production and Engineering much more realistic.

14.5 Products, Quotients, and Other Functions. Although functions like

(14.1) and (14.5) are extremely common in industry one might have

(14.31)

(14.32)

Or, there may be several variables multiplied and divided, or any functions

of almost unlimited variety. We now describe ways of handling non-linear

functions for characteristic y.

As we shall see later on, the important point in functions such as (14.31)

and (14.32) is the relative variability within each component, not the absolute
14.5. Products, Quotients, and Other Functions 423

variability, as was the case for additive and subtractive functions. We define

as is usual the coefficient of. relative variation by

V
X
= aX /~X = coefficient of variation (14. 33)

where V may be expressed decimally or in percent.

X

We may show that for two independent random variables x 1 and x 2 which are

always positive we have the approximate relations

v2 - v2 + v2 (14.34)
xlx2 xl x2

v2 v2 + v2 (14.35)
xl/x2 xl x2
Thus if V .001 and V .030, which mean respectively that typical
xl x2
percentage variations from the mean are .1% and 3%, we have approximately:

v2 = .001 2 + .030 2 = .000,901 or

xlx2
v = .03001
xlx2
Therefore the product is subject to the same relative variation as is x 2 alone.

Decreasing a or V on down to zero will have negligible effect: x 2 will

xl ltl
still impose its 3% variation upon the product, x 1x 2 • Further decreasing of

vx will not help at all. Precisely the same thing can be said for x 1!x 2 .
1
Next we may inquire about the means of products and quotients. We have

the following

~
xlx2
= IJ.x1 1-Lx2 (14.36)

iJ. ;, ~ 1~-t (14.37)

X/x 2 x1 x2
Note that the first is exact, the second approximate. As a consequence of

(14.34) to (14.37) we have, for example, for always positive independent

components
= xlx2
y (14.38)
x3x4
. IJ.liJ-2
IJ.y =IJ.31J.4
-- (14.39)
424 STATISTICS OF COMBINATIONS

(14 .40)

Example 1. Consider the volume y of steel plates. Suppose we have the

following

xl: !J.l 300 em. vl .001 .crl .3 em.

x2: !J.2 120 em. v2 .001 .cr 2 .12 em.

x3: !J.3 1 em. v3 .020 .cr3 .020 em.

Then for the volume y = x1x 2x 3 we have approximately

!J.y = 36,000 cc

vy2 ~ .001 2 + .001 2 + .020 2 = .ooo4o2 vy ~ .02005

which shows that the small cr 3 , contributes nearly all of the variation to y

because !J.3 is so small that V3 is far greater than v1 or V2 . If volume is of

basic importance, further refinement of control of length and width will have

negligible effect on the volume. Effort on further control of thickness is

where real improvement can be made. As to distribution, the component with

the largest V has the most influence upon the distribution of y. Here, if

x 3 is "normal," so will y be, or if of "rectangular" distribution so will y be.

When the V's are more nearly equal the distribution tends toward normality,

much like the Central Limit theorem.

Example 2. How can we best describe the result of the calculation with

approximate numbers given below?

37.2(462)
y = .0402(45,700)

Now just what does 37.2 mean? If a measurement result, it should mean that the

result is accurate to the nearest .1 of a unit, that is, that we believe the

true value to lie somewhere between 37.15 and 37.25, for otherwise some other

tenth unit would be more appropriate. Moreover, it seems reasonable to

suppose that the true value is equally likely to lie anywhere between the
14.5. Products, Quotients, and Other Functions 425

limits, that is, to be rectangularly distributed between. If such is not the

case, it should be so stated!

Assuming independence, which is usually good, we use the model of (14.38),

and assume x 1 averages 37.2 and so on. Then for the mean result we have
37.2(462) -
~y = .0402(45,700) - 9 · 355
Now to how much error is this subject? We need the four coefficients of

variation to answer the question. We assume a rectangular distribution of the

true xi's, and thus for the 37.2 measurement having L1 = 37.15, u1 = 37.25,
U-L = .1, which yields by (14.21) cr 1 = .1/2/3, and thus v 1 = .1/[2/3(37.2)].

Hence
.1
vl .000, 776
2./3(37. 2)
1
v2 .000,625
2/3(462)
.0001 .000, 718
v3
2/3( .0402)
100
v4 .000,632.
zv'3(45, 700)

Therefore we use (14.40) finding

V
y
~ .001382 = cry /~y .
Substituting ~
y
= 9.355 gives cr
y
.0129. Now what of the distribution?

With four factors all more or less equal in relative variability we can

count on the distribution being quite normal. Hence the best way to express

the result y is to treat it as a distribution giving

~y 9.355, cry .0129, approximately normal distribution.

Examining this we see that the probability of a true value lying

inside 9.355 ±.013 = 9.342, 9.368 is about .68 and that the deviation from

9.355 can quite possibly be as much as .03. Thus to write the value of y

as 9.35 with the implication that it is accurate to the nearest .01, just

because the four factors were each good for "three significant figures" is

thoroughly misleading.
426 STATISTICS OF COMBINATIONS

Significant Figures

Example 2 gives us reason to doubt the usefulness of such descriptions as

"three significant figure accuracy." Let us consider this a bit further.

Compare

10.2) 0'1 = .1/2/3 = .0289 .00283

99.7) 0'
2.
= .1/2/3 = .0289 .000290

and we see that although both x 1 and x 2 carry "three significant figure

accuracy," one has about ten times the relative accuracy of the other.

They are like a foot rule which may be anywhere from 12 in. to 120 in. long!

Such a description of accuracy should be forever laid to rest. But statistical

sophistication being by no means universal, the practice will doubtless linger

on.
Example 3. In order to give a little light on the distribution of products

and quotients of rectangularly distributed components, consider the distributions

of y = x 1x 2 and z = x 1 /x 2 where x 1 has limits 8 to 12 and x 2 17 to 23. Figure

14.1 shows the~ distributions of the two functions, found by integration

over appropriate regions. The two distributions are surprisingly similar in

shape, with their single straight sections and two curved sections.

14.5.1. ~' The Delta Technique. This is the name given to a technique

for finding moments of the distribution of a function of random variables.

Define the deviation

IJ,x = x-f,LX (14 .41)

then the expections (E) are as follows

(14.42)

(14 .43)

Moreover by (14.12) using ~y

E(AxAy) = covar (x,y) (14.44)

so that if x andy are independent (or at least uncorrelated) (14.14) gives

14.5. Products, Quotients, and Other Functions 427

.015

FIG. Distributions of y = x 1x 2 and z = x 1/x 2 where x 1 is

14.1.
rectangularly distributed 8 to 12, and x 2 likewise, but 17 to 23, with
x 1 , x 2 being independent.

E(~x~y) = 0 x,y independent. (14.45)

We now consider x,y independent and each ~ 0

z = xy /:,z = z-1-1z . (14 .46)

Then using (14.13)

1-1 = E(z) = E(xy) = sm sm xy f(x,y)dx dy = Jm Jm xy £1 (x)f 2 (y) dx dy

z 0 0 0 0

(14.47)

And so using (14.41)

6z xy- ~X~Y = (~X+ ~X)·(~y + ~y) - ~X~Y

1-1 ~x + 1-1 ~y + ~x~y.

y X
428 STATISTICS OF COMBINATIONS

Squaring yields
2 2 2 2 2
(6z) = ~ (6x) + ~ (6y) +
y X
2
+ 21-1X 6x(6y) ,
2
Taking the expectation of each side gives a on the left. The first three
z
2 2
terms on the right have expectations a , j.1X2 ay2 , aX2 ay2 by independence,
y X
~

which also gives a zero expectation for each of the last three terms

proceeding as for (14.47). Thus

2 2 2 2 2 2 2 2
a Z = axy = 1-1y aX + j.1X a y + crX cr y (14 .48)
Now dividing through by the respective sides of (14.47) squared and using

(14.33) gives the exact version of (14.34)

v2 = v2 + v2 + v2v2. (14.49)
xy X y X y
The last term on the right is nearly always negligible.
For z = x/y we may again use the ~ technique, but now things do not go

quite as smoothly. For example

E(z) = E(x/y) = E(x)•E(l/y)

if x and y are independent. Then

1/y = 1/(iJ.y + ~y) = 1/[!J.~(l

, + ~y/1-1
y )]
~ ( ~·· \ 2 ( ~·· \ 3
= ...1 [1- ~+ ~-~+ ... ].
!J.y y l-1y l-1y

Taking the expectation of both sides we find

2 3
E(1/y) = (1/~Y)[1 + vy- vy a 3 :y+ ... ]

where a 3 = E[(z-~ y ) 3 ]/cr y3 is the skewness of y.

Thus for E(l/y) to even exist we need a type of convergence. Since V is
y
likely to be very small in industry we usually assume

E(l/y) ~ (1 + V2)/!J. (14 .50)

y y
or even merely 1/~ .
y
14.5. Products, Quotients, and Other Functions 429

An extension of the foregoing enables us to handle functions of more

general type, if the functions are "well behaved." For example, for

z = g(x,y) (14.51)
we may expand into a Taylor series around ~x'~y:

From this 1 subject to rapidity of convergence, we have

(14. 52)

2 2
a + a . (14.53)
X y

Actually these approximations are surprisingly good in most practical situations.

Example 4. Using extensions of (14.51)-(14.53)

we have

~y =
\ru:
~1 ~2/~3 z
2
cr -
y - - 2 - 'crl2 + {-~--r
r~i~z}
~3
2
- 2 cr2, +
21~2~3
r.f'"'f'
3
~3
2
cr3

~1hich may be greatly simplified by use of V 1 s to

2
Vy = 9V21 + V22 /4 + 2
4V 3 .

Note how the exponents whether plus or minus enter into equation squared.

In particular we have
V(xp) = pV X and V(r IX) = lr V
X
(14.54)
A final example of a function to which we may apply (14.52) and (14.53)

is the resultant eccentricity (distance between two center points) coming from

eccentricity x 1 : A to B, eccentricity x 2 : B to C, with random angle 8 =<ABC. The

distribution of e is usually taken as rectangular 0° to 180°. But the

distributions of x 1 and x 2 might be

430 STATISTICS OF COMBINATIONS

more or less normal or else skewed toward the few higher values, because of
the impossibility of negative eccentricities. Thus we seek information on
the distribution of

12 2
y = vx 1 + x 2 - 2x1x2cos9.
This can be handled by (14.52) and (14.53) but the shape of yJs

distribution is more difficult to determine. The distribution of y may be

approximated by random drawings of x 1 ,x2 ,e.

14.6. Summary. In this chapter we have been concerned with the

distribution of a function of random variables, given the distributions of

these component variables. In particular we had the simple case of additive

and subtractive functions, the mean and standard deviation of which are
easily obtained. Such functions are extremely common in industry wherever

there are assemblies of component parts. By taking advantage of the

compensating errors or variation in the part dimensions it is possible to

considerably loosen up tolerances of the parts, provided assurance is

obtained of reasonable control of the mean dimensions. A system to ensure

this was given in Section 14.4.
While we have been illustrating with more or less mechanical or

electrical cases, the methods apply quite broadly, for example, to chemical

mixtures.
The less frequently encountered non-linear functions can also be

handled, though not quite as easily. Pure products and quotients, perhaps

involving powers or roots give little difficulty through use of approximation

methods and the coefficients of variation. Applications were made to

calculations with approximate measurement data .

Problems 431

PROBLEMS

14.1. A brass washer and a mica washer are to be assembled together,

one on top of the other. For the former ~l = .1155 in., a 1 = .00045 in.,

while for the latter ~2 = .0832 in. and a 2 = .00180 in. (a). Find ~.a

for the total thickness. (b). Make what assumptions you need, and stating

them, find the percentage of pairs in random assembly which will nQ! meet

specifications of .200 ± .005 in. (c). Would it help much if a 1 were to

be cut in half? (d). How in practice would you determine the ~'sand a's?

14.2. A 1000 ohm resistor and two 200 ohm resistors are to be connected
in series so that their resistances add. The former is from a process

yielding i = 998, R = 12 ohms for samples of 5 each, while the latter two are

drawn at random from a process with x = 202, R = 3.1 ohms for samples of 4

each. (a). What percent can be expected to meet specifications of 1400 ± 15

ohms? (b). What did you assume in your solution?

14.3. The outside diameter of a pin showed good control with =

x1 = .53600,

Rl .00080 em, and the inside diameter of a mating collar also, with x2 = .53733,
R2 .00100 em, sample size being five in each case. In choosing for random

assembly a pin and a collar, what percentage of pairs will have less diametral

clearance than .00025 em?

14.4. For electric light bulbs the distance from the rim at the base of

the thread to the glass had x1 = 2.6667, s 1 = .1087 em. For the inner liner

of a socket the distance from rim to top had i 2 = 2.3061, s2 = .0305 em.

Both samples were substantial. What can you say about the distribution of

x 1-x 2 , stating any assumptions used?

432 STATISTICS OF COMBINATIONS

14.5. A problem from a letter was the following: Three parts were

assembled as shown. A and Care flush (that is the same point), Band F

flush. Knowing the distributions of AB, CD and EF, how can one find the

mean and standard deviation of the distance DE?

A B

c D E F

14.6. The true dimension x has added to it a measurement error e,

the total being the observed dimension y. (a). Suppose that in production

ax : .OS mm and ae : .01 mm, what is the standard deviation of the observed
dimension y? (b). Would it help to measure each piece twice and average

the two? (c) . Same as (a) and (b) except a

x
= .02, a
e
= .04 mm.

14.7. Eight different parts are assembled along a shaft, additively.

Each is given specification limits of ±.020 em. If the resulting minimum

and maximum limits are merely added, what would the specification limits be

on the sum? (a). If each part is produced to just meet its limits in a

±3a normal curve sense, what specification limits could be set for the sum?
(b) . If each part is produced to just meet its limits with a rectangular

distribution, what specifications can the sum meet with ±3a assurance?

14.8. In a plant manufacturing soaps and cosmetics, a question arose

as to whether to use a lot of caps for cologne bottles, which was on hand.

Accordingly tests were made on the distribution of cap strength (that is,

torque which would break the cap), and the distribution of the torque which

the capping machine would apply to a cap. They found in inch-pounds

respectively

Xl = 11.1, s 1 : 2.80; x 2 = 8.0, s 2 = 1.51.

Assuming independence (very safe here) and normality, estimate the mean and

standard deviation of x 1-x 2 (margin of safety), and estimate the percentage

of caps which can be expected to be broken,by negative x 1-x 2 .

Problems 433

14.9. Two parts are to fit one inside the other. The clearance between

them is to meet specifications of .0015±.0005 in., that is, difference of

diameters. (a). If the parts can be produced under conditions of good

statistical control, but you want to give 25% more tolerance to the inside

diameter of the outer piece than to the outside diameter of the inside piece,

what ± specifications would you recommend for each? (b). What problems are

presented if the ± specifications so set are only ±20 of the processes?

(c). If ±4o of the processes? (d). One ±2o, one ±40?

14.10. A bearing and shaft assembly is as shown:

- - - - - - - - 5 . 3 8 1 ± . 0 1 0 - - - - - - - -....
Shaft

Now first suppose that each part meets its specifications precisely, in a

normal curve ±30 sense. (a). Then what ±30 limits is y able to meet

(give 1-1 ±30)? (b) . Suppose on the other hand that each part is only

meeting its specifications with a rectangular distribution between the

limits. Then what limits can be met?

14.11. In manufacturing rectangular copper wire, the width has

1-1 1 .365 and o 1 ~ .0100 in., and thickness 1-12 ~ .152, 0 2 ~ .0048 in.

(a). Assuming independence, find the mean and standard deviation of the

cross-sectional area. (b). Is independence a very safe assumption?

14.12. Suppose that for sheets of metal, nominally 14 x 20 inches, the

respective dimensions x 1 and x 2 have 1-11 ~ 14.01, 0 1 ~ .030 in., 1-1 2 ~ 20.02,

02 ~ .011 in., and x 1 and x 2 are independent. What can you say about the

distribution of the area y ~ x 1x 2 ? (Assume the sheets are always perfectly

rectangular.)

14.13. For Example 3 of Section 14.5, find for the product y:

1-J., V and 0. Do the same for the quotient z.

434 STATISTICS OF COMBINATIONS

14.14. In a certain packaging operation, the net weight is estimated

from the total weight w, cover weight c, and weight of coverless empty

container e. Thus if x is the net weight, then x = w - c - e. Suppose

examination yields w= 107.5, sw = .4, c = 2.0, sc = .1, e = 5.0, se = .2

in pounds, for substantial samples. What can you say about the distribution

of x? (Careful with this one l)

14.15. Three independent variables xl'x 2 and x 3 are to be combined to

give y = xl + (x2x3) . If ~1 = 200, cr 1 = 2, ~2 = 12, <J

2 = .OS, ~3 : 20,

a
3
= .1. (a) . Find ~
y
and a
y
approximately. (Hint: first find ~ cr
x2x3 ' XzX3
) .
(b), If the three xias follow approximately normal distributions what

would you expect for y?

14.16. Derive (14.54).

14.17. For the combined resistance for two resistors connected in

1
parallel we have y = l/xl + l/x 2 ' where x 1 and x 2 are the separate

resistances. Knowing ~l' a 1, ~2 • a 2 , what can you say about the distribution

of y?

14.1fl. If x has ~ = 10, a= .1, what would you expect to have for
2
(a) . y = X (b). y = 1/x (c) , y = ./x?
14.19. Assuming rectangular distributions to the precisions indicated,

approximate ~Y and cry for

25.2 )( 10.2
y = 87.9 " .933

14.20. Same as 14.19 but y 12.8 X 93.77 ~ .146.

14. 21. Same as 14.19 but y 36.7/35.8.

14.22. Bottles are, we shall suppose, manufactured perfectly cylindrical

inside. Let the radius be r, with~

r
= 1.00, cr
r
= .02 in., and normal
References 435

distribution. The volume v of fill of liquid is also normally distributed,

with ~v = 15.10, crv = .OS cu. in. Find the mean height of liquid in a bottle

~y and its standard deviation cry' approximately.

14.23. Five pieces on a shaft have been carrying specifications as follows:

Two with 2.0140 ± .0040 in.

Two with 3.2060 ± .0035 in.
One with 3.8400 ± .0060 in.
They have been set to meet specifications of 14.2800 ~ .0210 in., for the sum

but difficulty has been encountered meeting the component limits. Preserving

the relative ratios of the tolerances, set Ti's by T = T12 2 2

+ T2+ .. ,+T 5 . Then

using Section 14.4, set up (a) an acceptance sampling program for the first

kind of part, and (b) set up a machine-reset process control of the first kind

of part if the average tends to rise, as is commonly the case on outer

dimensions of machined parts.

References

1. I. W. Burr, Some theoretical and practical aspects of tolerances for

mating parts. Indust. Quality Control 15 (No.3), 18-22 (1958).

2. I. W. Burr, On the distribution of products and quotients of random

variables. ~·Quality Control 18 (No.3), 16-18 (1961).

3. I. W. Burr, Unpublished research done while on sabbatical leave at

University of Washington, 1966-1967, supported by the National Science

Foundation.

4. I. W. Burr, Specifying the desired distribution rather than maximum and

minimum limits. Indust. Quality Control 24 (No.2), 94-101 (1967).

5. R. E. James, Statistical process controls and acceptance sampling

methods to control the distribution of combined output characteristics.

Ph.D. Thesis, Purdue Univ., Lafayette, Indiana, 1970.

436 STATISTICS OF COMBINATIONS

6. I. W. Burr and R. E. James, Specifying the desired distribution in lieu

of tolerance limits - a systems approach. Paper presenred to the Design

Engineering Division of the Amer. Soc. of Mech. Engrs.; Conference,

Chicago, Apr. 2, 1974.

7. G. Bennett, The application of probability theory to the allocation of

engineering tolerances. Ph.D. Thesis, Univ. of New South Wales, 1964.

 CHAPTER 15

SOME OTHER FREQUENCY DISTRIBUTIONS

15.1 Introduction. It is desirable for the quality control worker

in industry to know something about distributions other than those we have

been working with so much, namely, the binomial, Poisson and hypergeometric

discrete distributions, and the normal curve continuous one. We have also

used the rectangular distribution in Chapter 14 as one type of extreme

alternative to the normal curve.

In this chapter we first take up moments of frequency distributions

for descriptive purposes, and then present several more frequency distri-

butions, both discrete and continuous. Then finally a standard test for

comparing observed and calculated frequencies is presented.

15.2. Moments of Distributions. Population and sample means and

variances are examples of moments. The mean is the first moment figured.

from the origin of the x axis. The variance is a second moment around the

mean, the square root being the standard deviation. Thus the student already

has some background in moments. Let us now add to this. We first consider

population moments. Discrete and continuous populations lead to different

formulas, which are basically the same. We define the expected or theoreti-

cal average values of powers of x by

pk_ = E(xk ). (15.1)

The weights applied to the xk values are determined by the theoretical

distribution in question. The two cases are

pk_ = L p(x)xk, discrete I f(x)x

k
dx, continuous. (15.2)
X

These are quite analogous to moments in mechanics, We note that

437
438 SOME OTHER FREQUENCY DISTRIBUTIONS

Po = 1 and Pi = p = population mean (15.3)

In the foregoing the prime to the p's meant "moments about the origin."

We next take up central moments, that is, moments around the mean instead

of around the origin. Thus by definition

(15.4)

and so

k k
I p(x) (x-p) J f (x) (x-p) dx" (15.5)
X

We note that as in (14.10) and (14.11)

Po = 1 • P1 = o, P2 (15.6)

The ~1s may be found from the ~ 1 's as follows, by expanding out the

binomials in (15.5)' distributing the summation or integral symbols along,

and using the definitions (15.2) and (15.3):

2 2
l12 ~I
fl a (15. 7)
·2

}J3 ~I
.3 3jlzfJ' + 2p3 (15.8)
2 4
)14 P4 4 P31-L + 6jlzJl - 3)1 • (15. 9)

No11 what good are the various moments? They aid in characterizing or

describing a theoretical distribution. The central moments p3 and p4 give

major importance to the numerically large deviations x - P• since we are

raising them to the third or fourth power. In p3 the third power of

negative deviations are still negative, so that p3 is an algebraic sum

and there is a contest as to which sign will win! But in p 4 all fourth

powers are positive. These two moments are related to the curve shape of

the distribution. However, they really reach full value in description

only when "standardized," that is, when the variability is taken into
15.1. Moments of Distributions 439

account. This is done as follows:

}13
a3 312 = skewness (15.10)
)12

)14
2= kurtosis or contact. (15 .11)
P2

Such measures are free of units and of the variability, and thus comparable

from one distribution to another.

In all of these population moments considered, Greek letters were

used. The sample counterparts or equivalents use Roman letters corresponding,

except that for the sample mean we use x rather than m. Thus for observed

frequency distributions with mid-values xi and corresponding frequencies fi

we define

-
X (15.12)

H.x~
1 1
m'k ---w.--
1
(15.13)

(15.14)

Some workers define m2 as the "sample variance." But if we use (2.4)

we have

(15.15)

Also parallel to (15.7) - (15.9) are

-2 (15.16)
m' X
2
m' (15.17)
3
m' (15.18)
4

The measures of curve shape are

440 SOME OTHER FREQUENCY DISTRIBUTIONS

(15.19)

(15.20)

Most of the time, when one has an observed frequency distribution, the

calculations will be done through coding via, say

x-a
U = c' or X a + cu, c > 0 (15.21)

It can readily be shown that the values of a 3 and a 4 are unchanged by such

a general linear transformation, that is, that

(15.22)

TI1e objective in fitting an observed distribution by a theoretical

one is to try to match up a 3 with the observed a 3 , or a 3 and a 4 with a 3

and a 4 . Thus we try to match the curve shapes of the two distributions,

observed and theoretical.

As a rough rule of thumb, if Ja 3 J ~ .3 the distribution will appear

quite symmetrical, a 3 = 0 indicates perfect symmetry by this criterion.

Ja 3 J .3 to .7 is moderate asymmetry. Skewness of± 1.0 is substantial

asymmetry. As pointed out, a 4 coming from ~4 is always positive, in fact

at least one. For the normal curve a 4 = 3,0, and this is a good basis for

judging contact, that is, the relative rapidity with which the frequency or

the curve approaches zero. For the rectangular distribution a 4 = 1.8,

It can be shown that if one takes random independent samples of

n x's from the same population having a 3 and a 4 finite, then we have for

the distribution of i 1 s

a
3:x- (15.23)
15.3. Some Frequency Distribution Models 441

a4:x- 3
3 + (15.24)
a
4:x- n

Note what happens in (15. 23) and (15. 24) when n increases, namely.

a3 for the x's approaches zero and a4 for the x's approaches three, the

respective normal curve values. This is in line with the Central Limit

theorem which states that, under very general conditions, the distribution

of the x's becomes normal as n increases. Also we may remark that (15.23)

and (15.24) hold for sums of x's as well as xJs, since the former are only

a multiple of the latter, using (15.22).

15.3. Some Frequency Distribution Models. Let us first give moments

for the distributions already included in this book:

0, 3 normal distribution (15.25)

0, 1.8 rectangular distribution. (15.26)

These are constant for all examples for each distribution, since there are

no "shape parameters" and all examples of each distribution are merely

linear transformations of each other. This follows from the fact that

a 3 and a 4 are unaffected by any linear transformation (as are a 3 and a4 for

observed data).

On the other hand for the Poisson, binomial and hypergeometric distri-

butions the curve-shape moments, a 3 and a 4 , vary according to the various

parameters. We have for the Poisson probability as defined in (2.35)

a4 =3 + 1/c', Poisson distribution. (15.27)

As c' becomes very small, the skewness a 3 becomes very large, with the

great bulk of probability concentrated at c = 0, namely, e-c' Also

a4 is large. On the other hand as c' increases, a 3 approaches zero and

a4 approaches 3. In fact by the time c' reaches 16, a 3 is already down

to .25 and a 4 3.06. a 3 is always positive for the Poisson.

442 SOME OTHER FREQUENCY DISTRIBUTIONS

For the binomial (2.30) with parameters p' and n, we have

l-2p 1 l-6p' Cl-E')

3 + np'(l-p'J
C!3 = binomial distribution(l5.28)
lnp' (1-p')
Here we see that if p' = .5, a3 = 0. For all n's such binomials are

perfectly symmetrical around n/2. If we use the conditions for a Poisson

as the limit of the binomial as n-+oo, p'-+0, and np'-+c 1 , it is easily seen

that (15.28) approaches (15.27) as a limit, as is of course expected. On

the other hand if p' is fixed and n--, then the limits of (15.28) are

a3 =0 and a 4 = 3, namely the normal curve values. This is in line with

the Central Limit theorem which would in the binomial case say that if we

drawn from the "binomial" for n = 1:

p(O) 1-p'
(15.29)
p(l) p'

we approach a normal distribution function as n-+oo, It is also worth

mentioning that if in (15.29) we use p' = .5, we have a distribution with

a4 = 1, which is the absolute minimum for the moment.

The formulas for the hypergeometric distribution (9.29) with N, D and

n are the following complicated ones, in which we define D/N as p' [1):

l-2p'
lnp' (1-p')
I N-l N-2n
N-n N-2 hypergeometric (15,30)

2 2
(N-1) [N(N+l)-6n(N-n)+3p' (1-E'){N (n-2)-Nn +6n(N-n)}), (l 5 • 31 )
np' (1-p') (N-n) (N-2) (N-3)

If N increases with D/N = p' held constant and n constant these moments

approach those for the binomial (15.28), as we would expect.

15.3.1. The Geometric Distribution. This discrete distribution occurs

frequently in applications, expecially in 1 ife testing and rel iabi 1i ty.

It is concerned with consecutive trials of an event each of which may be

either a "success" or a "failure." The random variable is the number of

15.3. Some Frequency Distribution Models 443

trials i taken until the first "failure." Thus i can be 1, 2,.. . Let
p 1 be the probability of a failure, q 1 the probability of a success. Then
if the first failure occurs on the ith trial or test, there must have been

i-1 successes followed by a failure. Therefore if we assume independence

of trials and constant p', as in the binomial distribution, we have

p(i)=P(SS· ··SF) = P(S)P(S)· · ·P(S)P(F) = q 1i-1 p 1 geometric distribution, (15.32)
Fror:J the form of (15.32), we see that the probabilities p(i) start
at a maximum of p 1 and gradually decrease, giving rise to what is often

called a "J-shaped" distribution. By use of the geometric progression sum

we may show that

!: p(i) 1 (15.33)
i=1

as indeed it must be. For the various desired moments we need to use
techniques for summing infinite series. For example

!: i(ql)i-lpl
i=l
]J p 1 + p 12q 1 + p 13q 12 + ... =p 1(1+2q 1 + 3q' 2+ ... )
Call
S = 1 + 2q 1 + 3q 12 +, .. +Jq
• 1j-l +•••

then multiply both sides by q' giving

• 1) q I j -1 + .••
q IS = q I + 2q I 2+,,,+ ( ]-

and subtract yielding

(l-q 1)S = 1 + q 1 + q 1 2+, ,.+q I j -1 +,.

Again multiply by q' obtaining

q 1(1-q 1 )S q I + q 1 2+,,,+q ,j-1 +,,,

and subtract giving

444 SOME OTI~ER FREQUENCY DISTRIBUTIONS
2
(1-q 1 ) s = 1.

But since 1 - q' = p1 we have

s l/p 12 .

Substituting this S into the expression for ~ gives

~ = 1/pl geometric distribution. (15.34)

This formula for the average number of trials is just what one might

expect, since, for example, if p 1 = • 10, it seems reasonable to suppose

that we should take 10 trials on the average before observing a one-chance-

in-ten occurrence.

We may also derive

a = /f.:P'!p I geometric distribution (15.35)

(2-pl)jn.::pl (15.36)
(9- 9p 1 + p 12 )/(l-p1), (15. 37)

A typical application is to subject a part or an assembly to a series

of severe tests and require that the first failure shall not occur before

some minimum, for example, 1000 tests. Or one may test each of a series of

parts once each, and prescribe that the first failure observed be not before

the 200th part tested.

15.3.2. The Negative Binomial Distribution. This distribution is

concerned with the same kind of drawing as are the binomial and the

geometric distributions, namely independent, with a constant probability

at each drawing, for the two kinds of objects or outcomes.

The problem solved by the negative binomial is that of asking for

the probability that the )$th failure occurs at the j.th trial. (Note that

if k = 1, then this specializes into the geometric distribution.)

The problem appears difficult at first, but a little thought soon

indicates that the probability p(i) is m~le up from two separate independent

events whose probabilities are easily found. For the tth failure to
15.3. Some Frequency Distribution Models 445

occur on the ith trial, we must have had precisely k-1 failures on the

first i-1 trials, and then this must be followed by a failure on the Jth

trial. The first is an ordinary binomial probability, and .the latter is

merely p'. Thus

p(i) = C(i-l,k-l)p'k-l(l-p')i-k·P'
or

negative binomial distribution. (15.38)

Just as in the geometric distribution, it is the number of trials i

that is the random variable. That is, (15.38) gives the probability that

the kth failure occurred at the ith trial, where i k, k+l, ••.•

The negative binomial is in reality a logical extension of the

geometric distribution. For suppose k = 2. Then let i = i 1 + i 2 , where

i 1 is the number of trials till the first failure occurs, and i 2 the

number of additional trials till the second failure occurs. Or in general

Then, using (14.5) through (14.7) with the ai's all +1, we find

k/p' negative binomial for k (15. 39)

0 = (1/p')lk(l-p') (15.40)

Moreover, using (15.23), (15.24) which are true of sums of x's as well

as x's we find from (15.36) and (15.37)

(2-p')/lk(l-p') (15.41)

2
a = 3 +
6(1-p')+p'
4 k(l-p') (15.42)

In general for k > 1, the probabilities p(i) increase for a while

and then decrease.*

* The maximum p(i) has i satisfying

~ < i k-1 1 . 1 1
p'-
1
.::.pr-+ ' or p - pr .::_ 1 .::_ p -pr + •

There may be two modes, within one, of each other.

446 SOME OTHER FREQUENCY DISTRIBUTIONS

Individual probabilities and cumulative probabilities for an interval

for the negative binomial may be found from appropriate tables of the

binomial. This rests upon the equivalence of two events:

1. The 1th failure occurs at the nth trial or earlier.

There are ~ or more failures in n trials.

This is because in 1, there must be at least k failures in the n trials,

if we complete the n trials. But if so, then surely in the n trials there

are k, k+l, ... ,n failures. But this is 2.

Now the probability of event 1 is the sum of negative binomial terms

from i = k to i = n, whereas the probability of event 2 is the sum of

binomial terms from d =k to n. Thus

n n
I p(ilk,p')neg. binom. I d=k p(dln,p')binomial. (15.43)
i=k

15.3.3. The Discrete Uniform Distribution. This is a simple distribution,

where we have consecutive whole numbers for the random variable i, say,

0, 1, 2, ... ,M-1, the probability of each being the same. Since there are

M of them and the sum of the probabilities is 1 we have

p (i) 1/M i = 0,1, ... ,M-1 uniform distribution (15.44)

0 otherwise.

One can use formulas for sums of powers of consecutive integers, along with

(15.2), (15,3), (15. 7)- (15.9) to prove

(M-1)/2 uniform distribution 0, 1, ... ,M-1 (15.45)

a lcM 2-l)/12 (15.46)

0 (15.47)

.6(3M 2-7)/(M 2-l). (15.48)

This is the discrete analog of the continuous rectangular distribution of

(14.19) - (14. 21).

Probably the most important application of the uniform distribution
is in random digits, especially where M = 10, that is, 0, 1, 2, ... ,9.
15.3. Some Frequency Distribution Models 447

These are to be drawn with constant probability .1, and independently.

For them p 4.5, cr = 133/4 = 2.87, and a 4 = 1.78. Two-digit random
numbers 00 to 99 have M = 100 and thus p = 49. 5, cr = 28.87 and a4 = 1.·80.

We finally note that whenever we have discrete values of a random

variable, j, of the form a, a+d, a+2d, ... ,a + nd, we can always transform

them, to use (15.44) by letting n = M-1 and (j-a)/d =i from which we may

find moments of j from (15.45)-- (15.48), the last two being unaffected

by the linear transformation.

15.3.4. The Exponential Distribution. This is a continuous

distribution of wide use in testing for length of life to failure, with

the time to failure being a continuous variable. In fact in life testing,

the exponential distribution is about as standard a model as the normal

distribution is in other fields of application. The idea is that there

may well be early failures, but that the probability of failure occurring

in uniform lengths of time decreases as time goes on.

The density function is defined as follows:

0 < X
f(x) (15. 49)
X > 0

We may easily show that~ in (15.49) is in fact the population mean E(x),

as indeed it should be if we are to use the letter ~ for the constant.

Thus we have by (15.2)*

p exponential distribution. (15. 50)

*Using the transformation x/p won the integral in (15.50) gives

""
pf 0 we
-w
dw

which by the definition of the gamma function (5.28) becomes

~r(2) = wl·r(l) = p
448 SOME OTHER FREQUENCY DISTRIBUTIONS

Using properties of the gamma function (5.28) we may find

a = P• a3 = 2, a4 =9 exponential distribution, (15.51)
Here we see that the mean and standard deviation have both the same value

p. The curve-shape characteristics indicate quite extreme asymmetry and

long tailing out toward the x axis. Figure 15.1 shows the curve shape for

the exponential distribution for two values of~. one and two. All
exponential distributions are merely changes of scale from each other,

and therefore a 3 and a 4 are constant.

Since ~ governs both mean and variance, there is little freedom

of action. It is possible to use

-1 -(x-6)/p
f(x) = { ~ e e < x
(15.52)
0 x < e

So that the mean is e + ~ and variance p, which gives more freedom.

But, at least for length of life, it makes little sense, because it

1.0

.8

.6

f(x)

FIG. 15.1. Exponential distributions with means ~ =1 and ~ 2.

15.3. Some Frequency Distribution ~1odels 449

implies no failures till x = 8, at which point, suddenly maximum

incidence of failures occurs.

This distribution is the continuous analogy of the discrete geometric

distribution (15.32).

15.3.5. The Gamma Distribution. This distribution is also called

the Type III distribution in the system of distributions developed by

Karl Pearson. A definition is

0 < x, p > 0
(15.53)
gamma distribution

X~ 0,

where we have defined the gamma function in (5.28). One can then obtain

the following moments by use of definitions (15.2), (15.3) and formulas

(15. 7) through (15.11), by use of the properties of the gamma function

(5.29) - (5.32).

~X p gamma distribution (15.54)

a
X
rp (15.55)

a3 2//p (15.56)
2
a4 3 + (6/p) 3 + 1.5a3 . (15.57)

As defined in (15.5S) we have but a one-parameter family of

distributions, which seems quite restricted for applications, since mean,

standard deviation and curve shape are all tied together being functions

of p. But we can generalize to make a three-parameter family if we

first standardize x by

z = (x-p)/VP (15.58)

whence Pz
Then if our desired mean is p and standard deviation a for a variable,
say, y, we can let
450 SOME OTHER FREQUENCY DISTRIBUTIONS

z = (y-JJ)/cr. (15. 59)

We could equate the two expressions for z, and solve for x in terms of

y. But in practice in fitting data by the gamma distribution it is easiest

to use tables by L. R. Salvosa [2), [3), i f they are available.

First one finds y, sy and a 3 , as usual, the last by (15.19). Then

the first two are called ~ and a if n is, say, at least 100. Next for

desired limits, such as class boundaries of y, we convert y to z by

(15.59). We are then in a position to find the area or probability below

y, enter Salvosa's tables with the corresponding z values (called t there)

and a 3 (called "skewness" there). Interpolation is usually necessary on

z and a 3 , but the intervals are small enough to permit ordinary inter-

polation.

It is not necessary that p in (15.53) be integral, as it can be

any positive real number. But if p is a positive integer then there is an

intimate relation with the exponential distribution. For, if we let JJ be

1 in (15. 49), and then let x = x 1 + x 2+••·+xp with the xi 1 s independently

distributed by (15.49), their sum x will follow (15.53). Hence the gamma

distribution bears the same relation to the exponential distribution

(15.49) that the negative binomial distribution (15.38) does to the

r,eometric distribution (J.5. 32). Thus we may think of the ganuna and

negative binonial as being analogous to each other.

Figure 15.2 shO\'IS several galllr.la distributions for comparison. They

have p = 1, 2, 4 which by (15.54) are also the means. We also have

respectively by (15.55) and (15.56), a= 1, 1.41, 2 and a,= 2, 1.41,

,)

1. on. Another comparison could have been dra1m with the horizontal scale

z in (15.59) while arranging the vertical scales to make the total area

unity. This would coupare the curve shapes on a uniform basis.

15.3. Some Frequency Distribution Models 451

1.00

FIG. 15.2. Three gamma distributions with parameters p = 1, 2, 4, showing

varying degrees of skewness. For p=l, this is an exponential distribution.

As p increases the curve approaches normality. It is quite normal at

p = 100 with a 3 = .2, a 4 = 3,06. This of course follows from the Central

Limit theorem since for example, x for p = 100, is x 1 + x 2 + ••• +x 100 where

x. follows an exponential distribution.

1

Regarding the normal curve as a "first approximation" to continuous

data, we can think of the gamma distribution as a "second approximation"

to continuous data.

We may add that the chi-square distribution is a gamma distribution.

15.3.6*. A General Sy3tem of Distributions. A flexible system of

frequency distributions may be defined as follows. If a distribution has a

density function f(y), then we can let the cumulative probability F(y) be

defined by the probability that the random variable Y be less than or equal
to y. That is, y
F(y) = P(Y2J) =f f(v)dv. (15.60)

Then we have the probability of Y lying in an interval a to b by

b b a
P(a<Y2_b) f f(v)dv f f(v)dv- f f(v)dv F(b) - F(a). (15.61)
a -""
452 SOME OTHER FREQUENCY DISTRIBUTIONS

We now define the cumulative probability for the system by: [4]
c -k
1-(l+y) y > 0 c,k > 0
F(y) (15.62)
0 y < o.

At first this looks like a two-parameter family, but in reality it can be

used to fit four moments, typically the observed x,sx' a 3 and a4 • - In fitting
observed data a 3 and a4 are used to determine c and k, which are in essence

curve-shape parameters, for they yield a 3 , cx 4 combinations. They also

yield ~ , cr . Then we convert observed x values to y's by

y y

The x values are often class boundaries, from which we want the probabilities

between. Are you still with me? Probably not, So we need an example.
Consider Table 15.1, which is actually of discrete data but is fitted

just the same as though continuous data. The y' s in column (1) are treated

as mid-values with corresponding frequencies in column (2). For calculational

purposes we code by v = y-12, and cumulate the sums shown at the bottom of
the table. These yield the m! moments for v's as shown next. Then the
J
central moments m.'s for the v's are found next, from (15.16) - (15.18),
J
and finally the curve shape characteristics a 3 and a 4 .

The closest entry in Table XV to a 3 = .3804, a4 = 3.240 is a 3 = .40,

a4 = 3.30, which has c = 3.341, k = 5.369. Also given in Table XV for this

c,k are p = .5638, cr .2039.

The class boundaries for the v scale, column (4), are next converted
to x 1 s for (15.62) by equating standardized variables for x and v as shown.
 15.3. Some Frequency Distribution Models 453

TABLE 15,1
Total Number of Letters in Alphabet in People's First and Last Names,
Fitted by (15.62)

(1) (2) (4) (5) CgJ (7) (8) (9)

No. of Letters Observed Code Bound- x in 1+x = 1-F(x)= Class Calc.
in names,y Freq.,f v ary,v (15,62) l+x3.341 (l+xC)-5.369 Probab, Freq., F

8 3 -4 .01935 2.1
-3.5 .18626 1.00364 .98065
9 5 -3 .06105 6.6
-2.5 .28866 1.01575 .91960
10 11 -2 .12364 13.2
-1.5 .39106 1.04342 .79596
11 23 -1 .17995 19.3
- .5 .49346 1.09444 .61601
12 22 0 .19973 21.4
+ .5 .59586 1.17731 .41628
13 17 +1 .17303 18.5
+1.5 .69826 1.30121 .24325
14 13 +2 .11955 12.8
+2,5 .80066 1.47581 .12370
15 7 +3 .06780 7.3
+3.5 .90306 1.71132 .05590
16 4 +4 ,03286 3.5
+4,5 1.00546 2.01840 .02304
17 0 +5 .01411 1.5
+5.5 1.10786 2.40813 .00893
18 2 +6 ,00893 1.0
107 157.T

l:vf = +20

For v's mi 20/107 = .186,916 mz = 4.000,000 m3 = 5.233,645 m4 54.018,692

m2 = m2 - Cmp 2 = 3.965,062 m3 = m3 - 3mzmi + 2Cmp 3 = 3.003,713

4
- 3(mi) = 50.940,533.
2
a 4 = m4 /(m 2) = 3.240

Take a 3=.4 a 4 =3.3 c=3,341 k=5.369 px=,5638 crx=.2039 Table XV

v ~. .18692, ,;m-; = 1.9912. Use v~:~~~~ 2 x~2~~~ 8 x = .1024v + .54466

 454 SOME OTHER FREQUENCY DISTRIBUTIONS

Columns (6) and (7) are steps on the way to finding probabilities for
classes by use of differences of 1- F(xi). Thus P(xi-l < x <xi) =
c -k
[l-F(xi_ 1)] - [1-F(xi)], 1-F(xi) being (l+xi) from (15.62). These
differences are given in column (8). Multiplying column (8) by n = 107
gives the calculated class frequencies Fin column (9). Note the good
agreement between columns (2) and (9), Also note that the first entry
in column (8) is for x = 0 to x = .18626 and the last entry is for
x = 1.10786 to~.

Much more complete tables are given in [4], with a 3 = ,0(.05)1,0(.1)2,0,

and a good variety of a 4 values by intervals of ,l or .2 to 9.0 maximum.
But since the a 3 , a 4 estimates of a 3 , a 4 are subject to considerable error
unless n is large, use of the nearest entry in Table XV is at least
reasonable. If a 3 is negative one can pick the entry with -a 3 nearest to
-a 3 , then reverse the scale direction by

v-m' x-~
___
1 =----X (15.63)
;m; ax

This causes our negative a3 to be positive in relation to the x variable

of (15,62).

15,4. A Goodness of Fit Test for Frequency Data, It is sometimes

desirable to compare a set of observed frequencies with those from some

theoretical distribution, for example normal or Poisson. Could the observed
data reasonably have come from the proposed theoretical distribution? In
general we have numerical class intervals, within each of which there is an

observed frequency, fi' out of a sample of n observations. The theoretical

distribution being compared provides theoretical probabilities, p., for each
1

such class, and thus expected frequencies, npi = Fi. Now we must expect
that the observed f.'s
1
will differ from their corresponding F.'s
1
even when
we draw all the sample results at random from the proposed theoretical
15.4. A Goodness of Fit Test 455

distribution. We therefore need a measure of the departures, taken as

a whole. This is supplied by the statistic named "chi-square." Now let

us be more specific:

1. The hypothesis being tested is that the population is of some

given type, for example, normal, binomial or Poisson.

2. The parameters (for example p, a, p' or c') may be specified, in

which case the parameters are part of the hypothesis. Or they

may have to be estimated from the data at hand, in which case the

hypothesis is only on the ~ of population.

3. A series of n observations of measurements x, or counts d or c,

is tabulated into numerical classes giving a frequency table. Let

fi be the observed frequency in the ith class.

4. Class probabilities pi's are calculated from the more or less

completely specified population. Then the expected class

frequencies are Fi = npi. Comparison is now to be made between

the fi and Fi frequencies to see whether or not the discrepancies

are readily explainable by chance variation. This comparison is

made on the set of discrepancies taken as a whole.

5. The test criterion is Galled chi-square (the same as that used for

variances from a normal curve population). It is

k
L (15.64)
i=l

where the subscript, nu, is the "degrees of freedom," and there

are k class frequencies to compare.

6. If any "tail" frequency Fi is less than 1.0, combine it with

the next adjacent Fi so as to make the lumped Fi, 1.0 at least.

456 SOME OTHER FREQUENCY DISTRIBUTIONS

This could occur at either or at each tail.* Then k is the number

of classes after any such lumping of frequencies is done.

7. Further, we let p stand for the number of independent parameters

estimated from the data, such as, ~ by x, a by s, c' by c or

p' by d/n. Then:

Degrees of freedom = v = k-1-p. (15.65)

8. The larger the observed chi-square (15.64) is, the relatively

poorer the fit. Chi-squares values of v are about average, by

chance variation. But if our observed chi-square is above the

1-a point the fit is judged unsatisfactory, where a is the risk

of erroneously rejecting il.n hypothesis when true. That is:

Observed 2 > x2
X v, 1-a below fit unsatisfactory
(15.66)
Observed X 2 -< x2v, 1-a below fit satisfactory.

In the former case we may try some other theoretical distribution.

In the latter case, there is no proof that the sample did come

from the theoretical distribution, but only that it perfectly

well could have come from it.

15.4.1. A Normal Curve Example. In Table 15.2 are shown observed

frequencies for an electrical contact, showing a "normal" looking distri-

bution, We may then wish to test whether the data could readily have

arisen from ~ normal population. We therefore find frequencies Fi to

compare with the observed fi. The first step is to find the mean x and
standard deviation s.

*This rule is given in [5]. It is a more liberal rule than those

most often given in textbooks.

15.4. A Goodness of Fit Test 457

TABLE 15.2

Observed Distribution of Dimensions of 1000 Electrical Contacts. Class

Frequencies Calculated Under Hypothesis of Normality, and Goodness of Fit
Tested by Chi-square.

(1) (2) (3) (4) (5) (6) (7) (8)

Mid-value, Observed Class Standard Area below Area Calcul. Contr.

x in. freq., f.]_ bounds,x var.z z in (4) between freq. F.]_ to x2

.412 6 .0152 15.2 5.5

.4145 -2.166 .0152
.417 34 .0409 40.8 1.1
.4195 -1.588 .0561
.422 132 .1001 100.1 10.2
.4245 -1.010 .1562
.427 179 .1767 176.7 .o
.4295 - .432 .3329
.432 218 .2251 225.1 .2
.4345 +.146 .5580
.437 183 .2074 207.4 2.9
.4395 +. 724 .7654
.442 146 .1381 138.1 .5
.4445 +1,302 .9035
.447 69 .0664 66.4 .1
.4495 +1. 880 .9699
.452 30 .0232 23.2 2.0
.4545 +2.458 .9931
.457 3 .0069 6.9 2.2
2
1000 X =24.7

Proceeding by usual methods for coding in a frequency distribution we

find

x- = .43324 in., s = .008651 in.

Next we convert the class boundaries in column (3) to the standard

variable z:

z = (x-x)/s = (x-.43324)/.008651 = ll5.594x -50.0798

(retaining more places than strictly justified to avoid calculational

errors on the differences). This gives column (4). The area or cumulative

probabilities below the class boundaries are then found from Table I.
458 SOME OTHER FREQUENCY DISTRIBUTIONS

interpolating, to yield column (5). The area or probabilities between

class boundaries are found by subtraction of entries in (5) giving (6).
Note that the first class includes all x's from -oo (pure fiction in

practice) to .4145 in., 1~hile the upper class is .4545 in. to oo(l.OOOO-
.9931), The class probabilities are then multiplied by n to give the

expected frequencies Fi of column (7).

To test the goodness of fit we form column (8). For example, in the

first class the contribution to x2 is

(6-15.2) 2/15.2 5.5

2
The total is x = 24.7. Is this indicative of a good or poor fit? It
depends upon~ and v. Suppose we take~= .01, and v = 10 - 1 - 2 = 7, the
"two" being for using x for 11 and s for cr. Then looking up the required

entry in Table II we find x~,. 99 below= 18.475. Since the observed chi-
square value of 24.7 is above this we regard the fit as unsatisfactory, and
conclude that the population is not normal. Not once in 100 times would
we get as poor a fit by chance, from a normal population. The large
contributions to x2 show the trouble spots.

15.4.2. A Poisson Example. For an example of the Poisson distri-

bution consider Table 15.3, wherein counts c. of a defect

1

called "plug holes" were tabulated for 54 large aircraft assemblies,

yielding frequencies fi as in column (2). Do these frequencies reasonably
well follow a Poisson distribution? In order to check this, we must
estimate c' for (2.35), that is, just the one parameter which fully

determines the Poisson distribution. The usual estimate is

c 308/54 = 5.70.
15.4. A Goodness of Fit Test 459

TABLE 15.3

Defects Called "Plug Holes" in Large Aircraft Assemblies. Fitting of F.

1
Done by Poisson Distribution.

(1) (2) (3) (4) (5) (6)

2
Plug Holes, c. Frequency, f. f.c. p.=P(cl5.70) Fi=pin Contr. to X
1 1 1 1 1

12 24 .0141
.8} 1.7 1.0
11 :}3 11 .0172 .9

10 1 10 .0334 1.8 .4

9 5 45 .0586 3.2 1.0

8 4 32 .0925 5.0 .2

7 7 49 .1298 7.0 .o
6 8 48 .1594 8.6 .0

5 9 45 .1678 9.1 .0

4 5 20 .1472 7.9 1.1

5 15 .1033 5.6 .1

2 3 6 .0543 2.9 .0

1 3 .0191
1.0}1.2 2.7
0 :}4 0 .0033 .2
2
54 308 1.0000 X9 = 6.9
-
c 308/54 = 5.70

Then we can either use (2.35) for P(clc'=5.70), or use a table [6) or [7].

Then multiplying column (4) entries by n = 54 gives the expected frequencies

of column (5):

To study the discrepancies between the fi and Fi, we work out the contri-
.
b ut1ons to x2
460 SOME OTHER FREQUENCY DISTRIBUTIONS

as given in column (6). The sum is x2 6.9. Note that we have combined

the frequencies fi and Fi for c = 11 and 12, and also for c =0 and 1,

in line with instruction 6 of Section 15.4 so as to have all Fi's at least

1. o.
2
To interpret this observed x we need the a risk and degrees of free-

Jom. Suppose we set a = ,OS. Then for v we use (15.65)

v = k - 1 - p = 11 - 1 - 1 = 9
p being one for estimating c' from c, and k being 11 classes after lumping.

Then by Table II,

2 16.919
x9,.95 below

so that the discrepancies between the fi and Fi are readily explainable by

chance, The true distribution of defects could perfectly well be a Poisson

distribution,

15.5. Summary. In this chapter we have presented some methods of

describing the curve-shape characteristics of frequency distributions,

theoretical and observed, by use of moments. The curve-shape characteristics

were given for distributions previously studied. Then we presented discrete

distributions: geometric, negative binomial and uniform, and continuous

distributions: exponential and gamma and a general system. Interrelations

between them were pointed out. Finally a test of goodness of fit by use of

chi-square was described and illustrated. Many other distributions could

have been included such as the beta and Weibull distributions.

Problems 461

PROBLEMS

15.1. Choose a random starting spot in the random digits of Table IV.

Then tabulate this digit and continue down the column, then the next

column to the right and so on for 200 random digits, to find observed

frequencies f 0 , f 1 , ... ,f9 . The expected frequencies Fi are each 20. Test

goo dness o f agreemen t by Xz Wl'th a = . 05 an d 1n

. t erpre t your resu lt .

15.2. Random digits one to six may be generated by rolling a single

perfectly balanced die. For this uniform distribution find P• a, a 4 •

What do these parameters become for the total on a roll of two balanced

dice? For three dice?

15.3. To illustrate the Central Limit theorem, one may delete the

face cards from two 52-card decks then draw five cards randomly from the

80 cards without replacement (with replacement is better, but five does not

unduly deplete the 80). These five form a sample. Regard aces as x=l, to

lOJs as X = 10. Find llj(• ax, a4, for the distribution. Would it be

quite normal appearing?

15.4. Sampling as in 15.3, each class member draws 10 samples, and

records his 10 x's. These are pooled, tabulating into appropriate classes,

and fitted by a normal curve, using p,a from 15.3 and tested for goodness of

fit by i
15.5. Derive (15.7) and (15.8) for the continuous case.

15.6. Derive (15.7) and (15.8) for the discrete case.

15.7. Show that {15.19) is unchanged under any linear transformation

for which the coefficient, a, of the variable is positive, in y = ax + b.

15.8. Using analogous definitions to (15.10), what are a 0 , a 1 , a 2 ?

 462 SOME OTHER FREQUENCY DISTRIBUTIONS

15.9. Data on the number of "Mislocations" on each of 54 large

aircraft assemblies follow:
No. defects, c 0 1 2 3 4 5 6 7 8 9

Frequency, f 12 9 11 8 7 0 4 0 2

Fit by a Poisson distribution, using Table VII, or [6] or [7], and test
goodness of fit. (The discrepancies are of a form indicating a "contagious"

tendency, that is, non-independence of defects.) Use a = .01.

15.10. Test for normality the following data on ladle carbon in

open-hearth heats of steel using a = .05.

%C .69 .70 .71 .72 .73 .74 .75 .76 .77 .78 .79 .80
f.1 2 21 32 63 89 106 97 51 33 4 1 1
15 .11. Test for normality the following data on density of glass at
20°C [8], using a = .05.
Density g/cc 2.5012 2.5022 2.5032 2.5042 2.5052 2.5062 2.5072

Freq.fi 2 6 25 33 19 10 4

15.12. Test the goodness of fit for Table 15.1, that is, columns (2)
vs. (9), using a = • OS. (Note that here p=4.)
15.13. From a box having two colors of beads, with a small proportion

of one, say, red colored, scoop up a sample of n=SO. Count the red ones,
return all beads, mix and scoop another 50. Continue for 100 samples. If

the number N in the box is large relative to n=SO, the counts should be
close to being binomially distributed. Use the known p' and a binomial
distribution for finding the theoretical probabilities F., and test for
1

goodness of fit.
15.14. P~ove (15.51), that a ~by using properties of the gamma

function and (15.7).

15 .15. Consider the following data on the "eccentricity" distance

between a conical point and the pitch diameter, of a needle value in .0001 in.
References 463

Fit a gamma distribution to the data and test the goodness of fit, with

Cl = .05.
Eccentricity 4.5 14.5 24.5 44.5 54.5 64.5 74.5 84.5 94.5

Frequency,fi 75 195 180 164 119 77 52 19 10 4

104.5 114.5 124.5 134.5

2 1 1 1

15.16. Fit the data of Problem 15.15 by (15.62) and test the goodness

of fit, with a .OS. (From [4], if a 3 = .95, a 4 = 4.4, then c = 1.993,536,

k = 8.293,100, ~ = .321,572, a= .181,784.)

References

1. M. G. Kendall and A. Stuart, "The Advanced Theory of Statistics."

Charles Griffin, London, 1969.

2. L. R. Salvosa, Tables of Pearson's Type III function, Ann. Math.

Statist. 1, 191-198 (1930).

3. H. C. Carver, "Statistical Tables." Edwards, Ann Arbor, 11ichigan,

1970.

4. I. IV. Burr, Parameters for a general system of distributions to match

a grid of a 3 and a 4 . Commun. in Stat. 2, 1-21 (1973).

5. W. G. Cochran, Some methods for strengthening the common x2 tests.

Biometrics 10, 417-451 (1954).

6. T. Kitagawa, "Tables of Poisson Distribution." Baifukan, Tokyo, 1952.

7. E. C. 11ol ina, "Poisson's Exponential Binomial Limit." Van Nostrand-

Reinhold, Princeton, New Jersey, 1947.
8. L. G. Ghering, Refined method of control of cordiness and workability
of glass during production. ~ Amer. Ceram. Soc, 27, 373-387 (1944).
 ANSWERS TO ODD-NUMBERED PROBLEMS
Numerical answers rounded to about slide-rule accuracy. Discussions could
well be made more complete than given here.

4.3 (a) x control lines: 46.74, 48.38, 50.02 R control lines: 0, 2.25, 5.13
(b) Both charts show lack of control. Erratic control of process.
(c) Not justified in estimating either 1.1 nor (J
X

4. 5 (a) x control lines: 822.6, 836.8, 851.0 R control lines: 0, 13 .9 , 35.7

(b) R Chart is in perfect control
(c) Can estimate ox by R/d 2 at 8.19, because R chart in control.
(d) Nox points within control limits. Not suprising, since samples
can be expected to give differing sets of repeated measurements.

4.7 From x, R x- control lines: 505.8, 546.2; 586.6 R control lines: 0, 83. 7,
167.7
From x, s x control lines: 504.0, 546.2, 588.4 s control lines: 1.0, 32.8,
64.6
Can regard within-group variabilities as homogeneous.
Cannot assume all data are from one homogeneous lot. 1.1's vary.

4.9 Early data x control lines: -. e33 , .333, 1.499 R control lines: 0,
2.02, 4.27. Both charts in control, need a fundamental change in
process to meet specifications. Estimate, ax= .868 by R/d2. Later
data x control lines: 1.100, 1.328, 1.556 R control lines: 0, . 395,
.835. Four x's outside, all R's inside. Estimate aX
= .170
by R/ d 2 .
For early dat a : tolerance/; = 2.30 which is bad. Can estimate percent
- X
out by (limit - x)/a X to check percents given. For later data:
tolerance/a
X
= 11.8, leaving room for 1.1 to vary.

4.11 x control lines: 29.4, 30.8, 32.2 s control lines: .41, 1.44, 2.47
Variabilities in control, but Bismuth and Selenium significantly low.

465
466 ANSWERS TO ODD-NUMBERED PROBLEMS

4.13 From x, x
ii. control lines: .09, 2.03, 3.97 R control lines: 0, 3.36,
7.11. From ll• ax
control lines: -.30, 2.00, 4.30 R control lines:
0, 3.99, 8.43. Control is perfect with respect to either set of lines.

4.15 x points closer to center line. R or s points running low (often below
center line)

CHAPTER 5

5.1 Need to use f(4) 31, f(3.5) (2.5) (1.5) (.5)1if

5.3 Continuing, since s's vary cr 2 > 0 and cr 2

s s

CHAPTER 6

6.1 Concern with defectives. np chart simplest. np = 1.72 UCL = 5.57.

Might use (6.20). Three consecutive points out. Probably to provide
36 good articles nearly always, P . 096 .

6.3 p chart. p -
• 00687. Use .00687 ::_ .2477//ri. Groupings could be n:
15,575, 7316, 2056, 1122. Five points out. Use weighted average of
p's, not LP/k. Use (6.12).

6.5 Defectives, p chart. P= .0909. Limits :

p
.0909 ~ .8623/v'n with
n = 2416, 1229, 783. Three high at beginning and end,. seven low
between. Out of control. Use (6.12) for weighted average of p's.

6.7 Assignable causes might be associated with (a) time of shift A, B or

C, (b) foreman and his crew, (c) day of week (d) first versus second
sample, or (e) the week. Use np chart, lines 0, 3.75, 9.5. Two points
out, for which the only thing in common is (a). High p's during the
day shift. Reason: excess light and eye strain. (Could make a three-
point or six-point chart by shifts.)
ANSWERS TO ODD-NUMBERED PROBLEMS 467

6.9 Control lines for c: 6.1, 19.3, 32.5. One above UCL, four very close.

6.11 Control lines for u: 2.64, 3.69, 4.74 One high.

6.13 Control lines for u: 2.00, 2.94, 3.88 Two out .

6.15 Set up formula for UCL with p . 150 and set to .180 and solve for n.
p
Gives 1275.

CHAPTER 7

7.1 -
X 32.0, R = 6.94, UCLR = 22.7, limitsx = 19.7, 44.3. R's in control,
six x's outside. One low x for 3s limits.

7.3 x = 50.8, R = 1.06, UCLR = 3.45, limitsx 48.9, 52.7. Variability

steady, but process average drifts.

7.5 x = 23.2, R = 4.64, UCLR = 15.2, limitsx = 15.0, 31.4. Some assignable
cause for high initial point. Ask the bug!

7.7 R = 3.02, UCLR = 6.4. Two R points out. Cause: before resetting and
after resetting pieces included in same sample. Can catch this. Hence
reviseR to 2.76. Slanting control limits at~ 1.59 from trend line.
The 3.2 should probably be called out of control depending on sketch of
trend.

7.9 Weighted averages x= .245, R = 2.02, Limitss = 1.63, 2.42, using

n = 117.5, badly out of control. Limitsx = -.316, +.805, all within.

7.11 Limitsx = .22500, .22520 in. LimitsR = .00051, .00091 in. Using x
aX ' estimate about .002 of pieces will be out of specifications.
Assume good control and normality of x's.
468 ANSWERS TO ODD-NUMBERED PROBLEMS

CHAPTER 9
Suggested Typical approx.
increment, p' value of p' Pa AOQ ASN ATI AOQL
9.1 . 004 .008 .525 .0042 200 2005 .0043

9.3 .008 .024 .651 .0156 200 1526 .016

9.5 .020 .060 .651 .0391 80 401 .040

9.7 .050 .200 .583 .116 50 238 .13

9.9 .020 . 040 .518 .0207 65 516 . 021

9.11 .004 .008 .503 .0040 171 2068 .0042

9.13 . 008 . 024 . 639 .0153 174 1551 .016

9.15 .020 .060 .639 .0384 70 404 .04

9.17 .050 . 200 .553 .111 45 247 .13

9. 19 .020 • 040 .653 .0261 75 381 .02

9.21 p' 95 .oo6s7 P'so = 0.62 p' 10 .0267

9,23 Ac = 5, n 232. Might use Ac 4, n 200, giving a a bit

over .OS.

9.25 Ac = 6, n 150. Might use Ac 5, n 132, giving a a bit

over .OS.

9.27 Find p' values, one giving Pa < .10, and the other Pa > .10, and
interpolate.

2
9.29 Pa q' (1+2p'). Then p' .25 Pa . 84; p' .SO, Pa .so; p' .75,
Pa .16.

9.31 It cuts off second sampling in the more or less "hopeless" cases.
ANSWERS TO ODD-NUMBERED PROBLEMS 469

9.33 Arbitrary rules of thumb: (1) p' < .OS and n ~ 20, (2) N ~ Sn,
(3) both (1) and (2) met.

9.37 .3929 9.39 .1095

(n-d)p'
9.41 P(d + 1 inn) (d+ 1 )q' P(d inn) P(c+1) = (c+c'1 ) P(c)

CHAPTER 10

10.1 n 1 = 85 c1 =0 n2 220 c2 5 3.3%

10.3 n = 135 c =2 pt 3.9%

10.5 n 1 = 110 c 1 = 1 n2 = 175 c2 =6 AOQL 1.3%

10.7 Try neighborhood of p' .02. AOQL about 1.01%

10.9 Pa = .108 by Poisson

-
10.11 pt decreases as p increases. Latter requires increasing samples which
give more discriminating OC curves which drop to the pt point for
lower p' values. (The AOQL occurs at p''s for which Pais about .55.)
Also pt decreases as N increases for same reason.

10.13 We might well still be using sampling inspection when p' is at or just
below the AOQL, but would have discontinued sampling if Pa were to
drop to anywhere near .10.

10.15 n = 80 Ac =1 Re = 2

10.17 n 32 Ac 0 Re 2

10.19 n. 20 Ac # # 0 0 1 1 2
1
Re 2 2 2 3 3 3 3
10.21 n = 20 Ac = 0 Re 1

10.23 n 125 Ac 0 Re 1
470 ANSWERS TO ODD-NUMBERED PROBLEMS

10.25 n1 16 In = 40 Ac 2 = 26 Re 2 = 27

10.29 Use large enough scale on p' to enable reduced to show a curve.

10.31 AOQL = .0067. So that when on tightened inspection we can be sure that
the average quality is at the acceptable level.

10.33 Because of the bolstering conditions which make receipt of bad quality
lots quite unlikely.

10.35 Yes. Useful in analyzing producer's process and in dealing with him.

10.37 Up to 10.0 defects or defectives per 100 units, the two are very
nearly equal. Above 10.0 they become less so.

10.39 Single: .525, .171. Double: .503, .172.

10.41 q' = P(no defect on unitlc' = .1) = .905 p' = .095 np' = 100(.095)
9.5 vs. 10 q' = P(no defect on unitlc' = .4) = .670 p' = .330 np' 33
vs. 40

CHAPTER 11

11.3 n = 11. Preserving a, K = 100,016 psi. Accept if x> K. Sketch,

using points at, say, 98,000, 100,016, and 102,000 psi .

11.5 n = 24. Preserving a, k • 00028 in. Accept if xis between 12.7435 + ~

in. 13 = .098.

11.7 n = 7. Accept if 1.1761 < x

< 1.1764 in. Use one-way tests, because
U-L 12.5C! • Two-way tests unnecessarily tight.
X

11.9 (a) n = 31. Accept Hx - 2.052s ~ 427g

(b) n = 10. Accept ifx ~ 444.88g
(c) n = 107, Ac =2
ANSWERS TO ODD-NUMBERED PROBLEMS 471

11.11 Test on homogeneous material, n = 23. Accept if s 2 2 3.46(ppm) 2

11.13 n = 22. Accept if s 2 < .0564(°C) 2 . Assume homogeneity, normality.

11.15 T = .0050 in. cr 1 = T/10, cr 2 = T/6, n = 23. Accept if s 2 ~ 3.85(10 -7 )sq.in.

From the same lot close together in time of production.

CHAPTER 12

12.1 h
1
= 1.239 h2 = 1.591 s = .0139 at p' s, Pa .562
p' 0 • 005 .0139 . 030 1
ASN 89 122 143 82 2

n 2-29 30-88 89-100 101-159 160-172

Ace. * * 0 0 1
Rej. 2 3 3 4 4

12.3 h1 = .958 h 2 = 1. 231 s = .0197 at p' s, Pa .562

p' 0 .oos .0197 .050 1
ASN 49 58 61 33 2

12.5 Most convenient is to work from 100,000 psi as zero, in, say, 1000
psi units. Then h 1 = 11.78 h2 = 11.78 s = 0, \11 = -2, \12 +2
Ace. no. = +11. 78, rej. no. = -11.78 for all n's. Pa(at O) = . s.
\l -2 0 +2
ASN 5.3 8.7 5.3

12.7 Take \ll = 441.31, \1 2 = 449.40, cr = 8.7. h1 27.04, h2 21.06,

s = 445.355
Pa(\1 1) = .10 Pa(s) = .562 Pa(\1 2) = .95
ASN(\1 1) = 5.50 ASN(s) = 7.52 ASN(\1 2) = 4.61

12.9 For lower specification \ll 1.17608", \1 2 = 1.17612". Code in

=
.0001" units from 1.17610". Then s = 0, h 1 = h 2 = .879 from cr .4
Pa(\1 1) = .1 Pa(s) = .s Pa(\1 2) = .9
ASN(\1 1) = 3.52 ASN(s) = 4.83 ASN(\1 2) 3.52
472 ANSWERS TO ODD-NUMBERED PROBLEMS

12.11 For coded variable z, ~ 0 = 0, ~ 0 ~ d = ~ 5, a = 8.34. h 1 = 20,93,

h 2 = 40.21 s = 2.5. Ace. intercept 12.2. Pa(~ 0 ) = .90 Pa(~~d) ,10

12.13 h1 = 20.70 = h 2 s = 3.592 Pa(l.S) = .9 Pa(IS) = .5 Pa(2.5) = .1

ASN(l.S) = 14.9 ASN(/S) = 17.6 ASN(2,5) = 8.0

12.15 h 1 = 3.16 = h 2 , s = .0584 o1 = .2 o2 = .3 Pa(o 1) = .90 Pa(IS) = .5

Pa(o 2) = .10 For~ known: ASN(O) = 6 ASN(o 1) = 13.8 ASN(/5) = 14.7
ASN(o 2) = 8.0

12.17 Coding in .0001": o1 = 5, o2 = 8.33 h 1 = 230.0 = h 2 s = 39.91

Pa(o 1) = .95 Pa(IS) = .so Pa(o 2) = .OS
~known: ASN(O) = 6 ASN(o 1) = 13.9 ASN(IS) = 16.6 ASN(o 2) = 7.0
For ~ unknown add 1 to above ASN's.

12.19 o1 = T/10 = 1.7 o2 = T/6 = 2.83

No. of R's of 8 1 2 3
Ace. for ~R 3.2 9.3 15.6 etc.
Rej. for ~R 9.2 15.3 21.4

12.21 h 1 = 4.41 h 2 5.66 s = 7,83 Pa(O) = 1 Pa(6) .95 Pa(s) = .562

Pa(lO) = .10 ASN(O) 1 ASN(6) = 2.13 ASN(s) 3,18 ASN(lO) = 2.14

12.23 The sequential plan has a lower ASN curve than the single plan
having comparable OC curves.

12.25 If a+ S > 1 then Pa(q 1) = 1 -a< S = Pa(q 2), that is, the Pa for
acceptable quality would be less than for rejectable quality.

CHAPTER 13
13.1 i = 37, pt = 2.2%; f = 1%, pt = 20.6% by (13.1); i 92, f = 3%;
AOQL = 1%, pt = .8%

13.3 il = 72, i2 = 96
ANSWERS TO ODD-NUMBERED PROBLEMS 473

13.5 Increasing either i or f gives greater protection, that is, lower AOQL.
So for fixed AOQL, if one increases, the other must decrease.

13.7 Use of same f permits use of the same sample of pieces for inspection,
after qualifying on both classes of defects. Use of same i does not
help at all.

13.9 93%, 87% compared to 81%. 48%, 41% compared to 40.5%.

13.11 Gives a higher Pa for lots under relatively good p' process levels.

13.13 For procedure 1: i 22, i 29

CHAPTER 14

14.1 (a) ~ = .1987, a = .00186 in. (b) Assuming normality and independence,
2. 28% will be outside. (c) Very little. (d) From control charts in
control.

14.3 ~ - .00133, a - .00055 em. 2.5% below.

14.5 ~ ~l - ~ 2 - ~ 3, a = I a 21 + a 22 + a 32 , d'1str1'b ut1on

. .
approx1mately normal.

14.7 If purely additive approach, assembly meeting~ .160 em. (a) + .0566 em.
(b) + , 0980 em.

14.9 (a) Specifications + .00039, ~ .00031 in. or in practice probably

~ .0004, ~ .0003 in. (b) Need better process, or ask relaxation
of tolerance or sort 100% (c) Use different allocation of tolerances.

14.11 (a) ~ = .0555, a = .00232 sq. in. (b) Could well be related, for
example inversely.

14.13 ~ 200, a 28.9; ~ .SO, a .0722

474 ANSWERS TO ODD-NUMBERED PROBLEMS

14.15 ~ ~ 440, cr ~ 2.54. Expect approximate normality.

y

14,17 ~y =. (1/~ 1 + 1I ~ 2 ) -1 , cry Cannot tell much about

shape of distribution.

14.19 ~ ~ 3.134, cr ~ ,00967

14.21 ~ = 1.025, cr ~ .0115

14.23 T1 = .0175, T2 = .0153, T3 = .0262 in.

x
Ace. Plan A: Ace. if 30 between 2.01172, 2.01628 in. and R1+R 2+R 3 ~ .0254 in,
Process control for rising by tool wear: ace, if R5 ~ .00963 and
x
5 between 2.0112 and 2.0168 in. Reset by D1ai or D1bi.

CHAPTER 15

15.3 ~x = 5.5, crx = 1.28, a 4 = 2.76, yes,

15.5 Simply expand (x-~)j under the integral and break into separate
integrals.

15.7 The cube of the coefficient a is in both numerator and denominator,

thus cancelling out.

15.9 c= 2.4630 Fe= 4.6, 11.3, 14.0, 11.5, 7.1, 3.5, 1.4, .5, .2, .1.
X~ 28.1, not a Poisson distribution.

2
15.11 Fi = 2.2, 8,6, 21.4, 3o.o, 23.5, 10.4, 3.0 x 4 2.9 Excellent fit
by normal.

-
15.15 Using u = (x-34.5)/10, u = -.19444, su = 1.9835, a 3 = ,956, a 4 = 4.396.
Use class boundaries for u variable, convert to z by (u-u)/s , Then
u
use [2) or [3], interpolating on z to .001 for a 3 = .9, 1.0, then on
a 3 for .96, giving cumulative probabilities. Then find differences and
multiply by n = 900 for F .• Gives F. 82.9, 170.4, 198.0, 168.0, 118.6,
1 1 . 2
74.1, 42.6, 22.8, 11.7, 5,8, 2.8, 1.3, .6, .5. G1ves x 9 = 10.7.
Satisfactory fit.
APPENDIX
 476 APPENDIX

TABLE I

Cumulative Probability, - 00 to z for the Standardized Normal Distribution,

<P(z) =!~co ¢(w)dw = P(Z.::_z) in Body of Table •

z .00 .01 .02 . 03 .04 .05 .06 .07 .08 .09

-- -- -- -- -- -- -- ---- --
.0002 .0002 .0002 .0002 .0002 .0002
-3.5 .0002 .0002 .0002 .0002
-3.4 .0003 .0003 .0003 .0003 .0003 .0003 .0003 .0003 .0003 .0002
-3.3 .0005 .0005 .0005 .0004 .0004 .0004 .0004 .0004 .0004 .oom
-3.2 .0007 .0007 .0006 .0006 .0006 .0006 .0006 .0005 .0005 .0005
-3.1 .0010 .0009 .0009 .0009 .0008 .0008 .0008 .0008 .0007 .0007

-3.0 .0013 .0013 .0013 .0012 .0012 .0011 .0011 .0011 .0010 .0010
-2.9 .0019 .0018 .0018 .0017 .0016 .0016 .0015 .0015 .0014 .0014
-2.8 .0026 .0025 .0024 .0023 .0023 .0022 .0021 .0021 .0020 .0019
-2.7 .0035 .0034 .0033 .0032 .0031 .0030 .0029 .0028 .0027 .0026
-2.6 .0047 .0045 .0044 .0043 .0041 .0040 .0039 .0038 .0037 .0036

-2.5 .0062 .0060 .0059 .0057 .0055 .0054 .0052 .0051 .0049 .0048
-2.4 .0082 .0080 .0078 .0075 .0073 .0071 .0069 .0068 .0066 .0064
-2.3 .0107 .0104 .0102 .0099 .0096 .0094 .0091 .0089 .0087 .0084
-2.2 .0139 .0136 .0132 .0129 .0125 .0122 .0119 .0116 .0113 .0110
-2.1 .0179 .0174 .0170 .0166 .0162 .0158 .0154 .0150 .0146 .0143

-2.0 .0228 .0222 .0217 .0212 .0207 .0202 .0197 .0192 .0188 .0183
-1.9 .0287 .0281 .0274 .0268 .0262 .0256 .0250 .0244 .0239 .0233
-1.8 .0359 .0351 .0344 .0336 .0329 .0322 .0314 .0307 .0301 .0294
-1.7 .0446 .0436 .0427 .0418 .0409 .0401 .0392 .0384 .0375 .0367
-1.6 .0548 .0537 .0526 .0516 .0505 .0495 .0485 .0475 .0465 .0455

-1.5 .0668 .0655 .0643 .0630 .0618 .0606 .0594 .0582 .0571 .0559
-1.4 .0808 .0793 .0778 .0764 .0749 .0735 .0721 .0708 .0694 .0681
-1.3 .0968 .0951 .0934 .0918 .0901 .0885 .0869 .0853 .0838 .0823
-1.2 .1151 .1131 .1112 .1093 .1075 .1056 .1038 .1020 .1003 .0985
-1.1 .1357 .1335 .1314 .1292 .1271 .1251 .1230 .1210 .1190 .1170

-1.0 .1587 .1562 .1539 .1515 .1492 .1469 .1446 .1423 .1401 .1379
-0.9 .1841 .1814 .1788 .1762 .1736 .1711 .1685 .1660 .1635 .1611
~0.8 .2119 .2090 .2061 .2033 .2005 .1977 .1949 .1922 .1894 .1867
-0.7 .2420 .2389 .2358 .2327 .2296 .2266 .2236 .2206 .2177 .2148
-0.6 .2743 .2709 .2676 .2643 .2611 .2578 .2546 .2514 .2483 .2451

-0.5 .3085 .3050 .3015 .2981 .2946 .2912 .2877 .2843 .2810 .2776
-0.4 .3446 .3409 .3372 .3336 .3300 .3264 .3228 .3192 .3156 .3121
-0.3 .3821 .3783 .3745 .3707 .3669 .3632 .3594 .3557 .3520 .3483
-0.2 .4207 .4168 .4129 .4090 .4052 .4013 .3974 .3936 .3897 .3859
-0.1 .4602 .4562 .4522 .4483 .4443 .4404 .4364 .4325 .4286 .4247
-0.0 .5000 .4960 .4920 .4880 .4840 .4801 .4761 .4721 .4681 .4641

Reproduced with permission from I. W. Burr, "Engineering Statistics and Quality

Control," McGraw-Hill, New York, 1953, pp. 404, 405.
APPENDIX 477

TABLE I (continued)

z .00 .01 .02 .03 .04 .05 .06 .07 .08 .09
-- -- -- -- ---- -- -- -- -~--

+O.O .5000 .5040 .5080 .5120 .5160 .5199 .5239 .5279 .5319 .5359
+0.1 .5398 .5438 .5478 .5517 .5557 .5596 .5636 .5675 .5714 .5753
+0.2 .5793 .5832 .5871 .5910 .5948 .5987 .6026 .6064 .6103 .6141
+0.3 .6179 .6217 .6255 .6293 .6331 .6368 .6406 .6443 .6480 .6517
+0.4 .6554 .6591 .6628 .6664 .6700 .6736 .6772 .6808 .6844 .6879
+0.5 .6915 .6950 .6985 .7019 .7054 .7088 .7123 .7157 . 7190 .7224

+0.6 .7257 .7291 .7324 .7357 . 7389 .7422 .7454 .7486 .7517 .7549
+0.7 .7580 .7611 .7642 .7673 .7704 . 7734 .7764 .7794 .7823 .7852
+0.8 .7881 .7910 .7939 .7967 .7995 .8023 .8051 .8078 .8106 .8133
+0.9 .8159 .8186 .8212 .8238 .8264 .8289 .8315 .8340 .8365 .8389
+1.0 .8413 .8438 .8461 .8485 .8508 .8531 .8554 .8577 .8599 .8621

+1.1 .8643 .8665 .8686 .8708 .8729 .8749 .8770 .8790 .8810 .8830
+1.2 .8849 .8869 .8888 .8907 .8925 .8944 .8962 .8980 .8997 .9015
+1.3 .9032 .9049 .9066 .9082 .9099 .9115 .9131 .9147 .9162 .9177
+1.4 .9192 .9207 .9222 .9236 .9251 .9265 .9279 .9292 .9306 .9319
+1.5 .9332 .9345 .9357 .9370 .9382 .9394 .9406 .9418 .9429 .9441

+1.6 .9452 .9463 .9474 .9484 .9495 .9505 .9515 .9525 .9535 .9545
+1.7 .9554 .9564 .9573 .9582 .9591 .9599 .9608 .9616 .9625 .9633
+1.8 .9641 .9649 .9656 .9664 .9671 .9678 .9686 .9693 .9699 .9706
+1.9 . 971:3 .9719 .9726 .9732 .9738 .9744 .9750 .9756 .9761 .9767
+2.0 .9772 .9778 .9783 .9788 .9793 .9798 .9803 .9808 .9812 .9817

+2.1 .9821 .9826 .9830 .9834 .9838 .9842 .9846 .9850 .9854 .9857
+2.2 .9861 .9864 .9868 .9871 .9875 .9878 .9881 .9884 .9887 .9800
+2.3 .9893 .9896 .9898 .9901 .9904 .9906 .9909 .9911 .9913 .9916
+2.4 .9918 .9920 .9922 .9925 .9927 .9929 .9931 .9932 .9934 .9936
+2.5 .9938 .9940 .9941 .9943 .9945 .9946 .9948 .9949 .9951 .9952

+2.6 .9953 .9955 .9956 .9957 .9959 .9960 .9961 .9962 .9963 .9964
+2.7 .9965 .9966 .9967 .9968 .9969 .9970 .9971 .9972 .9973 .9974
+2.8 .9974 .9975 .9976 .9977 .9977 .9978 .9979 .9979 .9980 .9981
+2.9 .9981 .9982 .9982 .9983 .9984 .9984 .9985 .9985 .9986 .9986
+3.0 .\J987 .9987 .9987 .9988 .9988 .9989 .9989 .9989 .9990 .9990

+3.1 .9990 .9991 .9991 .9991 .9992 .9992 .9992 .9992 .9993 .9993
+3.2 .9993 .9993 .9994 .9994 .9994 .9994 .9994 .9995 .9995 .9995
+3.3 .9995 .9995 .9995 .9996 .9996 .9996 .9996 .9996 .9996 .9997
+3.4 .9997 .9997 .9997 .9997 .9997 .9997 .9997 .9997 .9997 .9998
+3.5 .9998 .9998 .9998 .9998 .9998 .9998 .9998 .9998 .9998 .9998
 478 APPENDIX
TABLE II
Cumulative Probability Points for the Chi-Square Distribution, Zero to x2
for Degrees of Freedom \! in Row Heading. P = P(random variable x2 ~tabled
value).
p
~ Q.[Q~ 0 101 Q.~2~ ~.g5 D:tli o.i~
1 0.001 0.004
0.016 0.102
2 0.010 0.020 0.051 0.103
0.211 0.575
3 0.072 0.115 0.216 0.352
0.584 1.213
4 0.207 0.297 0.484 0.711
1.064 1.923
5 0.412 0.554 0.831 1.145
1.610 2.675
6 0.676 0.872 1.237 1.635
2.204 3.455
7 0.989 1.239 1.690 2.167
2.833 4.255
8 1.344 1.646 2.180 3.490
2.733 5.071
9 1.735 2.088 2.700 3.325
4.168 5.899
10 2.156 2.558 3.247 3.940
4.865 6.737
11 2.603 3.053 3.816 5.578
4.575 7.584
12 3.074 3.571 4.404 5.226
6.304 8.438
13 3.565 4.107 5.009 5.892
7.042 9.299
14 4.075 4.660 5.629 6.571
7.790 10.165
15 4.601 5.229 6.262 7.261
8.547 11.037
16 5.142 5.812 6.908 7.962
9.312 11.912
17 5.697 6.408 7.564 8.672
10.085 12.792
18 6.265 7.015 8.231 9.390
10.865 13.675
19 6.844 7.633 8.907 11.651
10.117 14.562
20 7.434 8.260 9.591 10.851
12.443 15.452
21 8.034 8.897 10.283 11.591
13.240 16.344
22 8.643 9,542 10.982 12.338
14.042 17.240
23 9.260 10.196 11.689 13.091
14.848 18.137
24 9.886 10.856 12.401 13.848 15.659 19.037
25 10.520 11.524 13.120 14.611 16.473 19.939
26 11.160 12.198 13.844 15.379 17.292 20.843
27 11.808 12.879 14.573 16.151 18.114 21.749
28 12.461 13.565 15.308 16.928 18.939 22.657
29 13.121 14.257 16.047 17.708 19.768 23.567
30 13.787 14.954 16.791 18.493 20.599 24.478
31 14.458 15.655 17.539 19.281 21.434 25.390
32 15.134 16.362 18.291 20.072 22.271 26.304
33 15.815 17.074 19.047 20.867 23.110 27.219
34 16.501 17.789 19.806 21.664 23.952 28.136
35 17.192 18.509 20.569 22.465 24.797 29.054
36 l7 ,887 19.233 21.336 23.269 25.643 29.973
37 18.586 19.960 22.106 24.075 26.492 30.893
38 19.289 20,691 22.878 24.884 27.343 31.815
39 19.996 21.426 23.654 25.695 28.196 32.737
40 20.707 22.164 24.433 26.509 29.051 33.660
41 21.421 22.906 25.215 27.326 29.907 34.585
42 22.138 23.650 25.999 28.144 30.765 35.510
43 22.859 24.398 26.785 28.965 31.625 36.436
44 23.584 25.148 27.575 29.787 32.487 37.363
45 24.311 25.901 28.366 30.612 33.350 38.291
Reproduced with permission from D. B. Owen, "Handbook of Statistical Tables,"
Addison-Wesley, Reading, Massachusetts, 1962, pp. 50-53,
APPENDIX 479

TABLE II (continued)

\) 0 1 7~ 2·20 0:9~ 0.975 0.99 0.995

1 1.323 2.706 3.841 5.024 6.635 7.879
2 2.773 4.605 5.991 7.378 9.210 10.597
3 4.108 6.251 7.815 9.348 11.345 12.838
4 5.385 7.779 9.488 11.143 13.277 14.860
5 6.626 9.236 11.071 12.833 15.086 16.750
6 7.841 10.645 12.592 14.449 16.812 18.548
7 9.037 12.017 14.067 16.013 18.475 20.278
8 10.219 13.362 15.507 17.535 20.090 21.955
9 11.389 14.684 16.919 19.023 21.666 23.589
10 12.549 15.987 18.307 20.483 23.209 25.188
11 13.701 17.275 19.675 21.920 24.725 26.757
12 14.845 18.549 21.026 23.337 26.217 28.299
13 15.984 19.812 22.362 24.736 27.688 29.819
14 17.117 21.064 23.685 26.119 29.141 31.319
15 18.245 22.307 24.996 27.488 30.578 32.801
16 19.369 23.542 26.296 28.845 32.000 34.267
17 20.489 24.769 27.587 30.191 33.409 35.718
18 21.605 25.989 28.869 31..526 34.80.5 37.156
19 22.718 27.204 30.144 32.852 36.191 38.582
20 23.828 28.412 31.410 34.170 37.566 39.997
21 24.935 29.615 32.671 35.479 38.932 41.401
22 26.039 30.813 33.924 36.781 40.289 42.796
23 27.141 32.007 35.172 38.076 41.638 44.181
24 28.241 33.196 36.415 39.364 42.980 45.559
25 29.339 34.382 37.652 40.646 44.314 46.928
26 30.435 35.563 38.885 41.923 45.642 48.290
27 31.528 36.741 40.113 43.194 46.963 49.645
28 32.620 37.916 41.337 44.461 48.278 50.993
29 33.711 39.087 42 •.557 45.722 49 •.588 52.336
30 34.800 40.2.56 43.773 46.979 50.892 53.672
31 35.887 41.422 44.985 48.232 52.191 55.003
32 36.973 42.585 46.194 49.480 53.486 56.328
33 38.058 43.745 47.400 50.725 54.776 57.648
34 39.141 44.903 48.602 51.966 56.061 58.964
35 40.223 46.059 49.802 53.203 57.342 60.275
36 41.304 47.212 50.998 54.437 58.619 61.581
37 42.383 48.363 52.192 55.668 59.892 62.883
38 43.462 49.513 53.384 56.896 61.162 64.181
39 44.539 50.660 54.572 58.120 62.428 65.476
40 45.616 51.805 55.758 59.342 63.691 66.766
41 46.692 52.949 56.942 60.561 64.950 68.053
42 47.766 54.090 58.124 61.777 66.206 69.336
43 48.840 55.230 59.304 62.990 67.459 70.616
44 49.913 56.369 60.481 64.201 68.710 71.893
45 50.985 57.505 61.656 65.410 69.957 73.166
480 APPENDIX

TABLE II (continued)
p
\) o 1 oo~ 2101 0:0~~ 2:0~ 0110 o.~

46 25.041 26.657 29.160 31.439 34.215 39.220

47 25.775 27.416 29.956 32.268 35.081 40.149
48 26.511 28.177 30.755 33.098 35.949 41.079
49 27.249 28.941 31.555 33.930 36.818 42.010
50 27.991 29.707 32.357 34.764 37.689 42.942
51 28.735 30.475 33.162 35.600 38.560 43.874
52 29.481 31.246 33.968 36.437 39.433 44.808
53 30.230 32.018 34.776 37.276 40.308 45.741
54 30.981 32.793 35.586 38.116 41.183 46.676
55 31.735 33.570 36.398 38.958 42.060 47.610
56 32.490 34.350 37.212 39.801 42.937 48.546
57 33.248 35.131 38.027 40.646 43.816 49.482
58 34.008 35.913 38.844 41.492 44.696 50.419
59 34.770 36.698 39.662 42.339 45.577 51.356
60 35.534 37.485 40.482 43.188 46.459 52.294
61 36.300 38.273 41.303 44.038 47.342 53.232
62 37.068 39.063 42.126 44.889 48.226 54.171
63 37.838 39.855 42.950 45.741 49.111 55.110
64 38.610 40.649 43.776 46.595 49.996 56.050
65 39.383 41.444 44.603 47.450 50.883 56.990
66 40.158 42.240 45.431 48.305 51.770 57.931
67 40.935 43.038 46.261 49.162 52.659 58.872
68 41.713 43.838 47.092 50.020 53.548 59.814
69 42.494 44.639 47.924 50.879 54.438 60.756
70 43.275 45.442 48.758 51.739 55.329 61.698
7l 44.058 46.246 49.592 52.600 56.221 62.641
72 44.843 47.051 50.428 53.462 57.113 63.585
73 45.629 47.858 51.265 54.325 58.006 64.528
74 46.417 48.666 52.103 55.189 58.900 65.472
75 47.206 49.475 52.942 56.054 59.795 66.417
76 47.997 50.286 53.782 56.920 60.690 67.362
77 48.788 51.097 54.623 57.786 61.586 68.307
78 49.582 51.910 55.466 58.654 62.483 69.252
79 50.376 52.725 56.309 59.522 63.380 70.198
80 51.172 53.540 57.153 60.391 64.278 71.145
81 51.969 54.357 57.998 61.261 65.176 72.091
82 52.767 55.174 58.845 62.132 66.076 73.038
83 53.567 55.993 59.692 63.004 66.976 73.985
84 54.368 56.813 60.540 63.876 67-.876 74.933
85 55.170 57.634 61.389 64.749 68.777 75.881
86 55.973 58.456 62.23-9 65.623 69.679 76.829
87 56.777 59.279 63.089 66.498 70.581 77.777
88 57.582 60.103 63.941 67.373 71.484 78.726
89 58.389 60.928 64.793 68.249 72.387 79.675
90 59.196 61.754 65.647 69.126 73.291 80.625
APPENDIX 481

TABLE II (continued)

v o.zs 0.90 0.95 o.2z~ 0.99 0.995

46 52.056 58.641 62.8~0 66.617 71.201 74.437
47 53.127 59.774 64.001 67.821 72.443 75.704
48 54.196 60.907 65.171 69.023 73.683 76.969
49 55.265 62.038 66.339 70.222 74.919 78.231
50 56.334 63.167 67.505 71.420 76.154 79.490
51 57.401 64.295 68.669 72.616 77.386 80.747
52 58.468 65.422 69.832 73.810 78.616 82.001
53 59.534 66.548 70.993 75.002 79.843 83.253
54 60.600 67.673 72.153 76.192 81.069 84.502
55 61.665 68.796 73.311 77.380 82.292 85.749
56 62.729 69.919 74.468 78.567 83.513 86.994
57 63.793 71.040 75.624 79.752 84.733 88.236
58 64.857 72.160 76.778 80.936 85.950 89.477
59 65.919 73.279 77.931 82.117 87.166 90.715
60 66.981 74.397 79.082 83.298 88.379 91.952
61 68.043 75.514 80.232 84.476 89.591 93.186
62 69.104 76.630 81.381 85.654 90.802 94.419
63 70.165 77.745 82.529 86.830 92.010 95.649
64 71.225 78.860 83.6 75 88.004 93.217 96.878
65 72.285 79.973 84.821 89.177 94.422 98.105
66 73.344 81.085 85.96'5 90.349 95.626 99.330
67 74.403 82.197 87.108 91.519 96.828 100.554
68 75.461 83.308 88.250 92.689 98.028 101.776
69 76.519 84.418 89.391 93.856 99.228 102.996
70 77.577 85.527 90.531 95.023 100.425 104.215
71 78.634 86.635 91.670 96.189 101.621 105.432
72 79.690 87.743 92.808 97.353 102.816 106.648
73 80.747 88.850 93.945 98.516 104.010 107.862
74 81.803 89.956 95.081 99.678 105.202 109.074
75 82.858 91.061 96.217 100.839 106.393 110.286
76 83.913 92.166 97.351 101.999 107.583 111.495
77 84.968 93.270 98.484 103.158 108.771 112.704
78 86.022 94.374 99.617 104.316 109.958 113.911
79 87.077 95.476 100.749 105.473 111.144 115.117
80 88.130 96.578 101.879 106.629 112.329 116.321
81 89.184 97.680 103.010 107.783 113.512 117.524
82 90.237 98.780 104.13!! 108.937 114.695 118.726
83 91.289 99.880 105.267 110.090 115.876 119.927
84 92.342 100.980 106.395 111.242 117.057 121.126
85 93.394 102.079 107.522 112.393 118.236 122.325
86 94.446 103.177 108.648 113.544 119.414 123.522
87 95.497 104.275 109.773 114.693 120.591 124.718
88 96.548 105.372 110.898 115.841 121.767 125.913
89 97.599 106.469 112.022 116.989 122.942 127.106
90 98.650 107.565 113.145 118.136 124.116 128.299
 482 APPENDIX

TABLE III

Cumulative Probabilities of Students t Distribution. p = P(t 2 tabled value)

for Given Degrees of Freedom v in Row.
p
\) 0.75 0:22 !2s2~ 0 1 97~ 0,92 2199~

1 1.0000 3.0777 6,3138 12.7062 31.8205 63,6567

2 0.8165 1,8856 2.9200 4.3027 6.9646 9,9248
3 0.7649 1,6377 2.3534 3,1824 4.5407 5.8409
4 0,7407 1,5332 2,1318 2.7764 3.7469 4.6041
5 0.7267 1.4759 2.0150 2,5706 3.3649 4,0322
6 0.7176 1.4398 1,9432 2.4469 3,1427 3.7074
7 0.7111 1.4149 1.8946 2.3646 2.9980 3.4995
8 0,7(1.64 1.3968 1,8595 2.3060 2.8965 3.3554
9 0.7027 1.3830 1.8331 2.2622 2.8214 3.2498
10 0,6998 1.3722 1.8125 2,2281 2.7638 3.1693
11 0,6974 1.3634 1.7959 2,2010 2.7181 3,1058
12 0.6955 1,3562 1.7823 2.1788 2.6810 3.0545
13 0.6938 1.3502 1.7709 2.1604 2.6503 3.0123
14 0,6924 1.3450 1.7613 2.1448 2.6245 2.9768
15 0.6912 1.3406 1.7531 2,1315 2.6025 2.9467
16 0,6901 1.3368 1. 7459 2,1199 2,5835 2,9208
17 0,6892 1.3334 1.7396 2.1098 2.5669 2,8982
18 0.6884 1.3304 1.7341 2.1009 2.5524 2.8784
19 0,6876 1.3277 1.7291 2.0930 2.5395 2.8609
20 0,6870 1.3253 1,7247 2,0860 2.5280 2.8453
21 0.6864 1.3232 1. 7207 2.0796 2.5177 2.8314
22 0,6858 1.3212 l. 7171 2.0739 2.5083 2.8188
23 0.6853 1.3195 1.7139 2.0687 2,4999 2,8073
24 0,6848 1.3178 1. 7109 2.0639 2.4922 2.7969
25 0,6844 1.3163 1.7081 2.0595 2.4851 2,7874
26 0,6840 1.3150 1.7056 2,0555 2.4786 2.7787
27 0,6837 1,3137 1.7033 2.0518 2.4727 2. 7707
28 0,6834 1.3125 1.7011 2.0484 2.4671 2,7633
29 0,6830 1.3114 1.6991 2.0452 2.4620 2.7564
30 0.6828 1,3104 1.6973 2.0423 2,4573 2.7500
31 0.6825 1.3095 1.6955 2.0395 2.4528 2. 7440
32 0.6822 1.3086 1.6939 2.0369 2,4487 2,7385
33 0,6820 1.3077 1.6924 2,0345 2.4448 2.7333
34 0.6818 1.3070 1.6909 2.0322 2.4411 2.7284
35 0,6816 1.3062 1.6896 2,0301 2.4377 2,7238
36 0.6814 1.3055 1,6883 2.0281 2.4345 2. 7195
37 0.6812 1.3049 1,6871 2.0262 2.4314 2.7154
38 0,6810 1.3042 1.6860 2.0244 2.4286 2. 7116
39 0,6808 1.3036 1,6849 2.0227 2.4258 2,7079
40 0.6807 1.3031 1,6839 2.0211 2.4233 2.7045
41 0.6805 1.3025 1.6829 2,0195 2.4208 2,7012
42 0,6804 1.3020 1.6820 2.0181 2,4185 2.6981
43 0,6802 1.3016 1.6811 2,0167 2,4163 2,6951
44 0,6801 1.3011 1.6802 2,0154 2,4141 2.6923
45 0.6800 1.3006 1,6794 2.0141 2.4121 2,6896

Reproduced with permission from D. B. Owen, "Handbook of Statistical Tables,"

Addison-Wesley, Reading, Massachusetts, 1962, pp. 28, 29.
APPENDIX 483

TABLE III (continued)

p
\)
olz5 Oa20 0 1 25 0:975 0,99 0.995
46 0,6799 1.3002 1.6787 2.0129 2.4102 2.6870
47 0,6797 1.2998 1,6779 2.0117 2.4083 2,6846
4a 0.6796 1.2994 1.6772 2.0106 2,4066 2,6822
49 0.6795 1.2991 1,6766 2.0096 2,4049 2.6800
so 0.6794 1.2987 1.6759 2.0086 2,4033 2.6778
51 0,6793 1.2984 1.6753 2.0076 2,4017 2.6757
52 0.6792 1.2980 1.6747 2.0066 2.4002 2.6737
53 0.6791 1.2977 1.6741 2,0057 2.3988 2.6718
54 0.6791 1.2974 1.6736 2.0049 2.3974 2.6700
55 0.6790 1.2971 1.6730 2.0040 2,3961 2,6682
56 0.6789 1.2969 1.6725 2.0032 2.394$ 2.6665
57 0.6788 1.2966 1.6720 2.0025 2,3936 2,6649
58 0.6787 1.2963 1.6716 2.0017 2.3924 2,6633
59 0,6787 1.2961 1,6711 2,0010 2,3912 2.6618
60 0.6786 1.2958 1.6706 2.0003 2.3901 2.6603
61 0.6785 1.2956 1.6702 1.9996 2.3890 2.6589
62 0.6785 1.2954 1,6698 1.9990 2.3880 2.6575
63 0.6784 1,2951 1.6694 1,9983 2.3870 2.6561
64 0,6783 1,2949 1.6690 1.9977 2.3860 2.6549
65 0.6783 1.2947 1.6686 1.9971 2.3851 2.6536
66 0.6782 1.2945 1.6683 1.9966 2.3842 2.6524
67 0.6782 1.2943 1.667.9 1.9960 2.3833 2.6512
68 0.6781 1.2941 1.6676 1.9955 2.3824 2.6501
69 0.6781 1.2939 1.6672 1,9949 2.3816 2,6490
70 0.6780 1.2938 1.6669 1.9944 2,3808 2.6479
71 0.6780 1.2936 1.6666 1.9939 2.3800 2.6469
72 0.6779 1.2934 1.6663 1.9935 2.3793 2.6459
73 0.6779 1.2933 1.6660 1.9930 2.3785 2,6449
74 0.6778 1,2931 1.6657 1.9925 2.3778 2.6439
75 0.6778 1.2929 1.6654 1.9921 2.3771 2,6430
76 0.6777 1.2928 1.6652 1.9917 2.3764 2.6421
77 0.6777 1,2926 1.6649 1.9913 2.3758 2,6412
78 0.6776 1.2925 1.6646 1.9908 2.3751 2,6403
79 0.6776 1.2924 1,6644 1.9905 2.3745 2,6395
80 0,6776 1,2922 1.6641 1.9901 2.3739 2,6387
81 0.6775 1.2921 1.6639 1.9897 2.3733 2,6379
82 0.6775 1.2920 1.6636 1.9893 2,3727 2,6371
83 0.6775 1,2918 1.6634 1.9890 2.3721 2.6364
84 0.6774 1,2917 1.6632 1,9886 2.3716 2,6356
85 0.6774 1.2916 1.6630 1.9883 2.3710 2,6349
86 0.6774 1.2915 1,6628 1. 9879 2.3705 2.6342
87 0,6773 1.2914 1.6626 1.9876 2.3700 2,6335
88 0,6773 1,2912 1.6624 1,9873 2,3695 2.6329
89 0,6773 1.2911 1.6622 1.9870 2.3690 2,6322
90 0.6772 1.2910 1,6620 1.9867 2.3685 2,6316
 484 APPENDIX

TABLE IV
Random Numbers
1368 9621 9151 2066 1208 2664 9822 6599 6911 5112
5953 5936 2541 4011 04.08 3593 3679 1378 5936 2651
7226 9466 9553 7671 8599 2119 533 7 5953 6355 6889
8883 3454 6773 8207 5576 6386 7487 0190 0867 1298
7022 5281 1168 4099 8069 8721 8353 9952 8006 9045

4576 1853 7884 2451 3488 1286 4842 7719 5795 3953
8715 1416 7028 4616 3470 9938 5703 0196 3465 0034
4011 0408 2224 7626 0643 1149 8834 6429 8691 0143
1400 3694 4482 3608 1238 8221 5129 6105 5314 8385
6370 1884 0820 4854 9161 6509 7123 4070 6759 6113

4522 5749 8084 3932 7678 3549 0051 6761 6952 7041
7195 6234 6426 7148 9945 0358 3242 0519 6550 1327
0054 0810 2937 2040 2299 4198 0846 3937 3986 1019
5166 5433 0381 9686 5670 5129 2103 1125 3404 8785
1247 3793 7415 7819 1783 0506 4878 7673 9840 6629

8529 7842 7203 1844 8619 7404 4215 9969 6948 5643
8973 3440 4366 9242 2151 0244 0922 5887 4883 1177
9307 2959 5904 9012 4951 3695 4529 7197 7179 3239
2923 4276 9467 9868 2257 1925 3382 7244 1781 8037
6372 2808 1238 8098 5509 4617 4099 6705 2386 2830

6922 1807 4900 5306 0411 1828 8634 2331 7247 3230
9862 8336 6453 0545 6127 2741 5967 8447 3017 5709
3371 1530 5104 3076 5506 3101 4143 5845 2095 6127
6712 9402 9588 7019 9248 9192 4223 6555 7947 2474
3071 8782 7157 5941 8830 8563 2252 8109 5880 9912

4022 9734 7852 9096 0051 7387 7056 9331 1317 7833
9682 8892 3577 0326 5306 0050 8517 4376 0788 5443
6705 2175 9904 3743 1902 5393 3032 8432 0612 7972
1872 8292 2366 8603 4288 6809 4357 1072 6822 5611
2559 7534 2281 7351 2064 0611 9613 2000 0327 6145

4399 3751 9783 5399 5175 8894 0296 9483 0400 2272
6074 8827 2195 2532 7680 4288 6807 3101 6850 6410
5155 7186 4722 6721 0838 3632 5355 9369 2006 7681
3193 2800 6184 7891 9838 6123 9397 4019 8389 9508
8610 1880 7423 3384 4625 6653 2900 6290 9286 2396

4778 !:1818 2992 6300 4239 9595 4384 0611 7687 208!1
3987 1619 4164 2542 4042 7799 9084 0278 8422 4330
2977 0248 2793 3351 4922 8878 5703 7421 2054 4391
1312 2919 8220 7285 5902 7882 1403 5354 9913 7109
3890 7193 7799 9190 3275 7840 1872 6232 5295 3148

0793 3468 8762 2492 5854 8430 8472 2264 9279 2128
2139 4552 3444 6462 2524 8601 3372 1848 1472 9667
8277 9153 2880 9053 6880 4284 5044 8931 0861 1517
2236 4778 6639 0862 9509 2141 0208 1450 1222 5281
8837 7686 1771 3374 2894 7314 6856 0440 3766 6047

6605 6380 4599 3333 0713 8401 7146 8940 2629 2006
8399 8175 3525 1646 4019 8390 4344 8975 4489 3423
8053 3046 9102 4515 2944 9763 3003 3408 1199 2791
9837 9378 3237 7016 7593 5958 0068 3114 0456 6840
2557 6395 9496 1884 0612 8102 4402 5498 0422 3335
Reproduced with permission from D. B. Owen, "Handbook of Statistical Tables,"
Addison-Wesley, Reading, Massachusetts, 1962, pp. 519, 520.
APPENDIX 485

TABLE IV (continued)
2671 4690 1550 2262 2597 8034 0785 2978 4409 0237
9111 0250 3275 7519 9740 4577 2064 0286 3398 1348
0391 6035 9230 4999 3332 0608 6113 0391 5789 9926
2475 2144 1886 2079 3004 9686 5669 4367 9306 2 595
5336 5845 2095 6446 5694 3641 1085 8705 5416 9066

6808 0423 0155 1652 7897 4335 3567 7109 9690 3739
8525 0577 8940 9451 6726 0876 3818 7607 8854 3566
0398 0741 8787 3043 5063 0617 1770 5048 7721 7032
3623 9636 3638 1406 5731 3978 8068 7238 9715 3363
0739 2644 4917 8866 3632 5399 5175 7422 2476 2607

6713 3041 8133 8749 8835 6745 3597 3476 3816 3455
7775 9315 0432 8327 0861 1515 2297 3375 3713 9174
8599 2122 6842 9202 0810 2936 1514 2090 3067 3574
7955 3759 5254 1126 5553 4713 9605 7909 1658 5490
4766 0070 7260 6033 7997 0109 5993 7592 5436 1727

5165 1670 2534 8811 8231 3721 7947 5719 2640 1394
9111 0513 2751 8256 2931 7783 1281 6531 7259 6993
1667 1084 7889 8963 7018 8617 6381 0723 4926 4551
2145 4587 8585 2412 5431 4667 1942 7238 9613 2212
2739 5528 1481 7528 9368 1823 6979 2547 7268 2467

8769 5480 9160 5354 9700 1362 2774 7980 9157 8788
6531 9435 3422 2474 1475 0159 3414 5224 8399 5820
2937 4134 7120 2206 5084 9473 3958 7320 9878 8609
1581 3285 3727 8924 6204 0797 0882 5945 9375 9153
6268 1045 7076 1436 4165 0143 0293 4190 7171 7932

4293 0523 8625 1961 1039 2856 4889 4358 1492 3804
6936 4213 3212 7229 1230 0019 5998 9206 6753 3762
5334 7641 3258 3769 1362 2771 6124 9813 7915 8960
9373 1158 4418 8826 5665 5896 0358 4717 8232 4859
6968 9428 8950 5346 1741 2348 8143 5377 7695 0685

4229 0587 8794 4009 9691 4579 3302 7673 9629 5246
3807 1785 7097 5701 6639 0723 4819 0900 2713 7650
4891 8829 1642 2155 0796 0466 2946 2970 9143 6590
1055 2968 7911 7479 8199 9735 8271 5339 7058 2964
2983 2345 0568 4125 0894 8302 0506 6761 7706 4310

4026 3129 2968 8053 2797 4022 9838 9611 0975 ?.437
4075 0260 4256 0337 2355 9371 2954 6021 5783 2827
8488 5450 1327 7358 2034 8060 1788 6913 6123 9405
1976 1749 5742 4098 5887 4567 6064 2777 7830 5668
2793 4701 9466 9554 8294 2160 7486 1557 4769 2781

0916 6272 6825 718fl 9611 1181 2301 5516 5451 6832
5961 1149 7946 1950 2010 0600 5655 0796 0569 4365
3222 4189 1891 8172 8731 4769 2782 1325 4238 9279
1176 7834 4600 9992 9449 5824 5344 1008 6678 1921
2369 8971 2314 4806 5071 8908 8274 4936 3357 4441

0041 4329 9265 0352 4764 9070 7527 7791 1094 2008
0803 8302 6814 2422 6351 0637 0514 0246 1845 8594
9965 7804 3930 8803 0268 1426 3130 3613 3947 8086
0011 2387 3148 7559 4216 2946 2865 6333 1916 2 259
1767 9871 3914 5790 5287 7915 8959 1346 5482 9251
 TABLE V

Control Chart Constants for Averages x,

Standard Deviations s and Ranges R; from Normal
Populations. Factors for Computing Central Lines and Three Sigma Control Limits
Factors for Factors for
Sample Factors for Control Standard dev·control limits Factors for Factors for control
Size n limits for x
iations, s for s ranges R limits for R

n A A2 A3 c4 c 83 84 s5 86 d2 d3 01 02 03 04
5
2 2.121 1.880 2.659 .798 .603 0 3.267 0 2.606 1.128 .853 0 3.686 0 3.267
3 1.732 1.023 1.954 .886 .463 0 2.568 0 2.276 1.693 .888 0 4.358 0 2.575
4 1.500 .729 1.628 .921 .389 0 2.266 0 2.088 2.059 .880 0 4.698 0 2.282
5 1.342 .577 1.427 .940 .341 0 2.089 0 1.964 2.326 .864 0 4.918 0 2.115

6 1.225 .483 1. 287 .952 .308 .030 1. 970 .029 1.874 2.534 .848 0 5.078 0 2.004
7 1.134 .419 1.182 .959 .282 .118 1.882 .113 1. 806 2.704 .833 .205 5.203 .076 1. 924
8 1.061 .373 1.099 .965 .262 .185 1. 815 .179 1.751 2.847 .820 .387 5.307 .136 1.864
9 1.000 .337 1.032 .969 . 246 .2391.761 .232 1.707 2.970 .808 .546 5.394 .184 1. 816
10 .949 .308 .975 .973 .232 . 284 1. 716 .276 1. 669 3. 078 .797 .687 5.469 .223 1. 777
11 .905 .285 .927 .975 .221 .321 1.679 .313 1.637 3.173 .787 .812 5.534 .256 1. 744
12 .866 .266 .886 .978 .211 .354 1.646 .346 1.610 3.258 .778 .924 5.592 .284 1.716
13 .832 .249 .850 .979 .202 .382 1.618 .374 1.585 3.336 .770 1. 026 5.646 .308 1. 692
14 .802 .235 .817 .981 .194 .406 1.594 .399 1.563 3.407 .762 1.121 5.693 .329 1. 671
15 .775 .223 .789 .982 .187 .428 1. 572 .421 1.544 3.472 .755 1. 207 5.737 .348 1.652

16 .750 .212 .763 .983 .181 .448 1. 552 .440 1.526 3.532 .749 1.285 5. 779 . 364 1. 636
17 .728 .203 .739 .985 .175 .466 1.534 .458 1.511 3.588 .743 1. 359 5.817 .379 1.621
18 .707 .194 .718 .985 .170 .482 1. 518 .475 1.496 3,640 ,738 1.426 5.854 . 392 1. 608
19 .688 .187 .698 .986 .165 .497 1.503 .490 1. 483 3. 689 .733 1.490 5.888 . 404 1. 596
20 .671 .180 .680 .987 .161 .510 1.490 .504 1.470 3, 735 .729 1. 548 5.922 .414 1.586

21 .655 .173 .663 .988 .157 .523 1.477 ,516 1.459 3. 778 .724 1.606 5.950 .425 1. 575
22 .640 .167 .647 .9€8 .153 .534 1.466 .528 1.448 3.819 .720 1.659 5.979 .434 1. 566
23 .626 .162 ,633 .989 .150 .545 1.455 ,539 1.438 3.858 .716 1. 710 6.006 .443 1. 557
24 .612 ,157 .619 .989 .147 .555 1.445 .549 1. 429 3. 895 .7i2 1. 759 6.031 .452 1.548
25 .600 .153 .606 .990 .144 .565 1.435 .559 1.420 3,931 .709 1.8 04 6.058 .459 1. 541

Formulas for Control Charts for Variables, s, R x,

Purpose of Chart Chart for Central Line 3-si~a Co~trol Limits
No Standard Given - used for Averages, i x x ± A2R, or
a~alrzi~g past data for control.
(x, R, s are average values x t A3s
for data being analyzed.) Ranges, R R o3R, D4R
Std. Devs., s s s3 s, s 4s
Standards Given - Used for Averages, x IJ + Acr
Dt4R~
1J
controlling quality with respect Ranges, R d 2cr or o1o,
to standards given IJ,cr. R' R' D 3 R~,
d2cr for n.
n Std. devs., s cn4 a s 5cr, B6cr

E(s) " c 4cr, (j c 5cr, E(R) d 2cr, crR " d 3cr.

.s
APPENDIX 487

TABLE VI

Moment Constants for Distribution of Sample Standard Deviation s*

n
'• Cs ~a:s n
'• '• o:a:'

2 .7979 .6028 .995 32 .9920 .1265 .130

3 .8862 .4633 .631 34 .9925 .1226 .125
4 .9213 .3888 .486 36 .9929 .1191 .122
5 .9400 .3412 .406 38 .9933 .1158 .118
6 .9515 .3075 .354 40 .9936 .1129 .115
7 .9594 .2822 .318 42 .9939 .1101 .112
8 .9650 .2621 .291 44 .9942 .1075 .109
9 .9693 .2458 .269 46 .9945 .1051 .107
10 .9727 .2322 .252 48 .9947 .1029 .105
11 .9754 .2207 .237 50 .9949 .1008 .102
12 .9776 .2107 .225 52 .9951 .0988 .100
13 .9794 .2019 .215 54 .9953 .0969 .098
14 .9810 .1942 .205 56 .9955 .0951 .096
15 .9823 .1872 .197 58 .9956 .0935 .095
16 .9835 .1810 .190 60 .9958 .0919 .093
17 .9845 .1753 .184 62 .9959 .0903 .091
18 .9854 .1702 .178 64 .9960 .0889 .090
19 .9862 .1655 .172 66 .9962 .0875 .089
20 .9869 .1611 .168 68 .9963 .0862 .087
21 .9876 .1571 .163 70 .9964 .0850 .086
22 .9882 .1534 .159 72 .9965 .0838 .085
23 .9887 .1499 .155 74 .9966 .0826 .083
24 .9892 .1466 .151 76 .9967 .0815 .082
25 .9896 .1436 .148 78 .9968 .0805 .081
26 .9901 .1407 .145 80 .9968 .0794 .080
27 .9904 .1380 .142 84 .9970 .0775 .078
28 .9908 .1354 .139 88 .9971 .0757 .076
29 .9911 .1330 .137 92 .9973 .0740 .075
30 .9914 .1307 .134 96 .9974 .0725 .073
31 .9917 .1286 .132 100 .9975 .0710 .072

a E(s) = c.u; a, = CsUi "'a:o.

Reproduced with permission from I. w. Burr, "Applied Statistical Methods,"

Academic Press, New York, 1974, p. 437.
 488 APPENDIX

TABLE VII

Poisson Distribution. Probabilities of c or less, given c', appears

in body of table multiplied by 1000.

~ - - -- - - - - - - --
0.02
0 1

980 1,000
2 3 4 5 6 7
------
8 9
--
0.04 961 999 1,000
0.06 942 998 1,000
0.08 923 997 1,000
0.10 905 995 1,000
0.15 861 990 999 1,000
0.20 819 982 999 1,000
0.25 779 974 998 1,000
0.30 741 963 996 1,000

0.35 705 951 994 1,000

0.40 670 938 992 999 1,000
0.45 638 925 989 999 1,000
0.50 607 910 986 998 1,000

0.55 577 894 982 998 1,000

0.60 549 878 977 997 1,000
0.65 522 861 972 996 999 1,000
0.70 497 844 966 994 999 1,000
0.75 472 827 959 993 999 1,000
0.80 449 809 953 991 999 1,000
0.85 427 791 945 989 998 1,000
0.90 407 772 937 987 998 1,000
0.95 387 754 929 984 997 1,000
1.00 368 736 920 981 996 999 1,000
1.1 333 699 900 974 995 999 1,000
1.2 301 663 879 966 992 998 1,000
1.3 273 627 857 957 989 998 1,000
1.4 247 592 833 946 986 997 999 1,000
1.5 223 558 809 934 981 996 999 1,000
1.6 202 525 783 921 976 994 999 1,000
1.7 183 493 .757 907 970 992 998 1,000
1.8 165 463 731 891 964 990 997 999 1,000
1.9 150 434 704 875 956 987 997 999 1,000
2.0 135 406 677 857 947 983 995 999 1,000

Reproduced with permission from I. W. Burr, "Engineering Statistics and Quality

Control," McGraw-Hill, New York, 1953, pp. 417-421.
APPENDIX 489

TABLE VII (continued)

~ 2.2
0

111
1

- --
-355
2 3 4 5

---- - - - - - - - - --
6

623 819 928 975 993 998 1,000

7 8 9

2.4 091 308 570 779 904 964 988 997 999 1,000
2.6 074 267 518 736 877 951 983 995 999 1,000
2.8 061 231 469 692 848 935 976 992 998 999
3.0 050 199 423 647 815 916 966 988 996 999
3.2 041 171 380 603 781 895 955 983 994 998
3.4 033 147 340 558 744 871 942 977 992 997
3.6 027 126 303 515 706 844 927 969 988 996
3.8 022 107 269 473 668 816 909 960 984 994
4.0 018 092 238 433 629 785 889 949 979 992
4.2 015 078 210 395 590 753 867 936 972 989
4.4 012 066 185 359 551 720 844 921 964 985
4.6 010 056 163 326 513 686 818 905 955 980
4.8 008 048 143 294 476 651 791 887 944 975
5.0 007 040 125 265 440 616 762 867 932 968
5.2 006 034 109 238 406 581 732 845 918 960
5.4 005 029 095 213 373 546 702 822 903 951
5.6 004 024 082 191 342 512 670 797 886 941
5.8 003 021 072 170 313 478 638 771 867 929
6.0 002 017 062 151 446 847 916
-- ·-- 285 606 744
10 11 12 13 14 15 16
---- -- -- -- --
2.8 1,000
3.0 1,000
3.2 1,000
3.4 999 1,000
3.6 999 1,000
3.8 998 999 1,000
4.0 997 999 1,000
4.2 996 999 1,000
4.4 994 998 999 1,000
4.6 992 997 999 1,000
4.8 990 996 999 1,000
5.0 986 995 998 999 1,000
5.2 982 993 997 999 1,000
5.4 977 990 996 999 1,000
5.6 972 988 995 998 999 1,000
5.8 965 984 993 997 999 1,000
6.0 957 980 991 996 999 999 1,000
490 APPENDIX

TABLE VII (continued)

~ 0 1 2 3 4 5 6 7 8 g

- -- - -- - - - - -- --
6.2 002 015 054 134 259 414 574 716 826 902
6.4 002 012 046 119 235 384 542 687 803 886
6.6 001 010 040 105 213 355 511 658 780 869
6.8 001 009 034 093 192 327 480 628 755 850
7.0 001 007 030 082 173 301 450 599 729 830
7.2 001 006 025 072 156 276 420 569 703 810
7.4 001 005 022 063 140 253 392 539 676 788
7.6 001 004 019 055 125 231 365 510 648 765
7.8 000 004 016 048 112 210 338 481 620 741
8.0 000 003 014 042 100 191 313 453 593 717
8.5 000 002 009 030 074 150 256 386 523 653
9.0 000 001 006 021 055 116 207 324 456 587
9.5 000 001 004 015 040 089 165 269 392 522
10.0 000 000 003 010 029 067 130 220 333 458
------ - - ------ - - -- -
10 11 12 13 14 15 16 17 18 19
--- - - - - - -- -- - - ----
6.2 949 975 989999 1,000
995 998
6.4 939 969 986 994 997 999 1,000
6.6 927 963 982 992 997 999 999 1,000
6.8 915 955 978 990 996 998 999 1,000
7.0 901 947 973 987 994 998 999 1,000
7.2 887 937 967 984 993 997 999 999 1,000
7.4 871 926 961 980 991 996 998 999 1,000
7.6 854 915 954 976 989 995 998 999 1 ,00,9
7.8 835 902 945 971 986 993 997 999 1,000
8.0 816 888 936 966 983 992 996 998 999 1,000
8.5 763 849 909 949 973 986 993 997 999 999
9.0 706 803 876 926 959 978 989 995 998 999
9.5 645 752 836 898 940 967 982 991 996 998
10.0 583 697 792 864 917 951 973 986 993 997
--
20 21 22
- -- -
8.5 1,000
9.0 1,000
9.5 999 1,(100
10.0 998 999 1,000
APPENDIX 491

TABLE VII (continued)

~ 0 1 2 3 4 5
- - - - - - - - - - - -- -- - --
002 007 021 050
6

102
7

179
8

279
9

397
10.5 000 000
11.0 000 000 001 005 015 038 079 143 232 341
11.5 000 000 001 003 011 028 060 114 191 289
12.0 000 000 001 002 008 020 046 090 155 24 2
12.5 000 000 000 002 005 015 035 070 125 201

13.0 000 000 000 001 004 011 026 054 100 166
13.5 000 000 000 001 003 008 019 041 079 135
14.0 000 000 000 000 002 006 014 032 062 109
14.5 000 000 000 000 001 004 010 024 048 088
15.0 000 000 000 000 001 003 008 018 037 070
-- - -- - - -- - - - - -- -- - - -
10 11 12 13 14 15 16 17 18 19
- - - - - - - - - -- - - - - - - - - -
10.5 521 639 742 825 888 932 960 978 988 994
11.0 460 579 689 781 854 907 944 968 982 991
11.5 402 520 633 733 815 878 924 954 974 986
12.0 347 462 576 682 772 844 899 937 963 979
12.5 297 406 519 628 725 806 869 916 948 969

13.0 252 353 463 573 575 764 835 890 930 957
13.5 211 304 409 518 623 718 798 861 908 942
14.0 176 260 358 464 570 669 756 827 883 923
14.5 145 220 311 413 518 619 711 790 853 901
15.0 118 185 268 363 466 568 664 749 819 875
---- -- - - - - -- - - ---- --
20 21 22 23 24 25 26 27 28 29
- - - - - - - - - - - -- - - - - -
10.5 997 999 999 1,000
11.0 995 998 999 1,000
11.5 992 996 998 999 1,000
12.0 988 994 997 999 999 1,000
12.5 983 991 995 998 999 999 1,000

13.0 975 986 992 996 998 999 1,000

13.5 965 980 989 994 997 998 999 1,000
14.0 952 971 983 991 995 997 999 999 1,000
14.5 936 960 976 986 992 996 998 999 999 1,000
15.0 917 947 967 981 989 994 997 998 999 1,000
492 APPENDIX

TABLE VII (continued)

~ --
16
4 5 6 7
- -- -
8 9 10 11

000 001 004 010 022 043 077 127 193 275
12 13

17 000 001 002 005 013 026 049 085 135 201
18 000 000 001 003 007 015 030 055 092 143
19 000 000 001 002 004 009 018 035 061 098
20 000 000 000 001 002 005 011 021 039 066
21 000 000 000 000 001 003 006 013 025 043
22 000 000 000 000 001 002 004 008 015 028
23 000 000 000 000 000 001 002 004 009 017
24 000 000 000 000 000 000 001 003 005 011
25
---- -- --
000 000 000 000 000 000 001 001 003 006

16
14
-- -- 15
-16- 17 18 -19
--
368 467 566 659 742 812
--
- 20 -21- -22- 23
868 911 942 963
--
17 281 371 468 564 655 736 805 861 905 937
18 208 287 375 469 562 651 731 799 855 899
19 150 215 292 378 469 561 647 725 793 849
20 105 157 221 297 381 470 559 644 721 787
21 072 111 163 227 302 384 471 558 640 716
22 048 077 117 169 232 306 387 472 556 637
23 031 052 082 123 175 238 310 389 472 555
24 020 034 056 087 128 180 243 314 392 473
25
-- --
012 022 038 060 092 134 185 247 318 394
-- -- -- --
24 25
-- ~
26 27 28
-- -- -29- -30- -31- -32- 33 --
16 978 987 993 996 998 999 999 1,000
17 959 975 985 991 995 997 999 999 1,000
18 932 955 972 983 990 994 997 998 999 1,000
19 893 927 951 969 980 988 993 996 998 999
20 843 888 922 948 966 978 987 992 995 997
21 782 838 883 917 944 963 976 985 991 994
22 712 777 832 877 913 940 959 973 983 989
23 635 708 772 827 873 908 936 956 971 981
24 554 632 704 768 823 868 904 932 953 969
25
-- -- -- --
473 553 629 700 763 818 863 900 929 950
--
19
34
- - -- -- -- -- --
35
- - 1,000
999
36 37 38 39 40
-41- 42 43- -- -
20 999 999 1,000
21 997 998 999 999 1,000
22 994 996 998 999 999 1,000
23 988 993 996 997 999 999 1,000
24 979 987 992 995 997 998 999 999 1,000
25 966 978 985 991 994 997 998 999 999 1,000
 TABLE VIII

Single Sample Tests for a= a 1 versus a= a 2 > a 1 , with Risks a= S.

Ratio of a 2 ;a 1 for Multiplier for a~ to
get K.
a =S a =S

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Sample
size .10 .05 .02 .01 .10 .05 .02 .01

2 13.1 31.3 92.6 206. 2.71 3.84 5.41 6.64

3 4.67 7.63 13.9 21.4 2.30 3.00 3.91 4.60
4 3.27 4.71 7.29 9.93 2.08 2.60 3.28 3.78
5 2.70 3.65 5.22 6.69 1.94 2.37 2.92 3.32
6 2.40 3.11 4.22 5.22 1.85 2.21 2.68 3.02
7 2.20 2.76 3.64 4.39 1.77 2.10 2.51 2.80
8 2.06 2.55 3.26 3.86 1.72 2.01 2.37 2.64
9 1.96 2.38 2.99 3.49 1.67 1.94 2.27 2.51
10 1.88 2.26 2.79 3.22 1.63 1.88 2.19 2.41
11 1.81 2.16 2.63 3.01 1.60 1.83 2.12 2.32
12 1.76 2.07 2.50 2.85 1.57 1.79 2.06 2.25
13 1.72 2.01 2.40 2.71 1.55 1.75 2.00 2.18
14 1.68 1.95 2.31 2.60 1.52 1.72 1.96 2.13
15 1.64 1.90 2.24 2.50 1.50 1.69 1.92 2.08
16 1.61 1.86 2.17 2.42 1.49 1.67 1.88 2.04
17 1.59 1.82 2.12 2.35 1.47 1.64 1.85 2.00
18 1.57 1.78 2.07 2.28 1.46 1.62 1.82 1.97
19 1.55 1.75 2.02 2.23 1.44 1.60 1.80 1.93
20 1.53 1.73 1.98 2.18 1.43 1.59 1.77 1.90
21 1.51 1.70 1.95 2.13 1.42 1.57 1.75 1.88
22 1.50 1.68 1.91 2.09 1.41 1.56 1.73 1.85
23 1.48 1.66 1.88 2.05 1.40 1.54 1.71 1.83
24 1.47 1.64 1.86 2.02 1.39 1.53 1.69 1.81
25 1.46 1.62 1.83 '1.99 1.38 1.52 1.68 1.79
26 1.44 1.61 1.81 1.96 1.38 1.51 1.66 1.77
27 1.43 1.59 1.79 1.93 1.37 1.50 1.65 1.76
28 1.42 1.58 1.77 1.91 1.36 1.49 1.63 1.74
29 1.41 1.56 1.75 1.89 1.35 1.48 1.62 1.72
30 1.41 1.55 1.73 1.87 1.35 1.47 1.61 1.71
31 1.40 1.54 1.72 1.84 1.34 1.46 1.60 1.70
40 1.34 1.46 1.60 1.71 1.30 1.40 1.52 1.60
50 1.30 1.40 1.52 1.61 1.27 1.35 1.46 1.53
60 1.27 1.36 1.46 1.54 1.24 1.32 1.41 1.48
70 1.25 1.33 1.42 1.49 1.22 1.30 1.38 1.44
80 1.23 1.30 1.39 1.45 1.21 1.28 1.36 1.41
90 1.21 1.28 1.36 1.42 1.20 1.26 1.33 1.38
100 1.20 1.26 1.34 1.39 1.19 1.24 1.31 1.36

For n, seek entry in columns (2) - (5), ~ a2 ;a 1 , giving n. Then for

this n and a= S, find in columns (6) - (9), the multiplier for o21 to

give K. Then s 2 < K, accept a = a 1 , s 2 > K, reject a = a 1 , and

conclude a > a 1 •
Reproduced with permission from I. W. Burr, "Applied Statistical Methods,"
Academic Press, New York, 1974, p. 448.
 ..,.
\D
..,.
TABLE IX

Logarithms of Factorials-Tens to Left, Units in Columns

0 1 2 3 4 5 6 7 8 9

00 0.0000 0.0000 0.3010 0.7782 1.3802 2.0792 2.8573 3.7024 4.6055 5.5598
10 6.5598 7.6012 8.6803 9.7943 10.9404 12.1165 13.3206 14.5511 15.8063 17.0851
20 18.3861 19.7083 21.0508 22.4125 23.7927 25.1906 26.6056 28.0370 29.4841 30.9465
30 32.4237 33.9150 35.4202 36.9387 38.4702 40.0142 41.5705 43.1387 44.7185 46.3096
40 47.9116 49.5244 51.1477 52.7811 54.4246 56.0778 57.7406 59.4127 61.0939 62.7841

50 64.4831 66.1906 67.9066 69.6309 71.3633 73.1037 74.8519 76.6077 78.3712 80.1420
60 81.9202 83.7055 85.4979 87.2972 89.1034 90.9163 92.7359 94.5619 96.3945 98.2333
70 100.0784 101.9297 103.7870 105.6503 107.5196 109.3946 111.2754 113.1619 115.0540 116.9516
80 118.8547 120.7632 122.6770 124.5961 126.5204 128.4498 130.3843 132.3238 134.2683 136.2177
90 138.1719 140.1310 142.0948 144.0632 146.0364 148.0141 149.9964 151.9831 153.9744 155.9700

100 157.9700 159.9743 161.9829 163.9958 166.0128 168.0340 170.0593 172.0887 174.1221 176.1595
110 178.2009 180.2462 182.2955 184.3485 186.4054 188.4661 190.5306 192.5988 194.6707 196.7462
120 198.8254 200.9082 202.9945 205.0844 207.1779 209.2748 211.3751 213.4790 215.5862 217.6967
130 219.8107 221.9280 224.0485 226.1724 228.2995 230.4298 232.5634 234.7001 236.8400 238.9830
140 241.1291 243.2783 245.4306 247.5860 249.7443 251.9057 254.0700 256.2374 258.4076 260.5808

150 262.7569 264.9359 267.1177 269.3024 271.4899 273.6803 275.8734 278.0693 280.2679 282.4693
160 284.6735 286.8803 289.0898 291.3020 293.5168 295.7343 297.9544 300.1771 302.4024 304.6303
170 306.8608 309.0938 311.3293 313.5674 315.8079 318.0509 320.2965 322.5444 324.7948 327.0477
180 329.3030 331.5606 333.8207 336.0832 338.3480 340.6152 342.8847 345.1565 347.4307 349.7071
190 351.9859 354.2669 356.5502 358.8358 361.1236 363.4136 365.7059 368.0003 370.2970 372.5959
---·--~ .. ---

Reproduced with permission from I. W. Burr, "Engineering Statistics and Quality >
'"Cl
'"Cl
Control," McGraw-Hill, New York, 1953, pp. 412-416. m
z0
><
-
 ?;
""mz
0
.....
><

0 1 2 3 4 5 6 7 8 9

200 374.8969 377.2001 379.5054 381.8129 384.1226 386.4343 388.7482 391.0642 393.3822 395.7024
210 398.0246 400.3489 402.6752 405.0036 407.3340 409.6664 412.0009 414.3373 416.6758 419.0162
220 421.3587 423.7031 426.0494 428.3977 430.7480 433.1002 435.4543 437.8103 440.1682 442.5281
230 444.8898 447.. 2534 449.6189 451.9862 454.3555 456.7265 459.0994 461.4742 463.8508 466.2292
240 468.6094 470.9914 473.3752 475.7608 478.1482 480.5374 482.9283 483.3210 487.7154 490.1116

250 492.5096 494.9093 497.3107 499.7138 502.1186 504.5252 506.9334 509.3433 511.7549 514.1682
260 516.5832 518.9999 521.4182 523.8381 526.2597 528.6830 531.1078 533.5344 535.9625 538.3922
270 540.8236 543.2566 545.6912 548.1273 550.5651 553.0044 555.4453 557.8878 560.3318 562.7774
280 565.2246 567.6733 570.1235 572.5753 575.0287 577.4835 579.9399 582.3977 584.8571 587.3180
290 589.7804 592.2443 594.7097 597.1766 599.6449 602.1147 604.5860 607.0588 609.5330 612.0087

300 614.4858 616.9644 619.4444 621.9258 624.4087 626.8930 629.3787 631.8659 634.3544 636.8444
310 639.3357 641.8285 644.3226 646.8182 649.3151 651.8134 654.3131 656.8142 659.3166 661.8204
320 664.3255 666.8320 669.3399 671.8491 674.3596 676.8715 679.3847 681.8993 684.4152 686.9324
330 689.4509 691.9707 694.4918 697.0143 699.5380 702.0631 704.5894 707.1170 709.6460 712.1762
340 714.7076 717.2404 719.7744 722.3097 724.8463 727.3841 729.9232 732.4635 735.0051 737.5479

350 740.0920 742.6373 745.1838 747.7316 750.2806 752.8308 755.3823 757.9349 760.4888 763.0439
360 765.6002 768.1577 770.7164 773.2764 775.8375 778.3997 780.9632 783.5279 786.0937 788.6608
370 791.2290 793.7983 796.3689 798.9406 801.5135 804.0875 806.6627 809.2390 811.8165 814.3952
380 816.9749 819.5559 822.1379 824.7211 827.3055 829.8909 832.4775 835.0652 837.6540 840.2440
390 842.8351 845.4272 848.0205 850.6149 853.2104 855.8070 858.4047 861.0035 863.6034 866.2044
------------ - . - -.

-!'>
<0
(/1
 .,.
tD
a-

TABLE IX (continued)

0 1 2 3 4 5 6 7 8 9

400 868.8064 871.4096 874.0138 876.6191 879.2255 881.8329 884.4415 887.0510 889.6617 892.2734
410 894.8862 897.5001 900.1150 902.7309 905.3479 907.9660 910.5850 913.2052 915.8264 918.4486
420 921.0718 923.6961 926.3214 928.9478 931.5751 934.2035 936.8329 939.4633 942.0948 944.7272
430 947.3607 949.9952 952.6307 955.2672 957.9047 960.5431 963.1826 965.8231 968.4646 971.1071
440 973.7505 976.3949 979.0404 981.6868 984.3342 986.9825 989.6318 992.2822 994.9334 997.5857

450 1000.2389 1002.8931 1005.5482 1008.2043 1010.8614 1013.5194 1016.1783 1018.8383 1021.4991 1024.1609
460 1026.8237 1029.4874 1032.1520 1034.8176 1037.4841 1040.1516 1042.8200 1045.4893 1048.1595 1050.8307
470 1053.5028 1056.1758 1058.8498 1061.5246 1064.2004 1066.8771 1069.5547 1072.2332 1074.9127 1077.5930
480 1080.2742 1082.9564 1085.6394 1088.3234 1091.0082 1093.6940 1096.3806 1099.0681 1101.7565 1104.4458
490 1107.1360 1109.8271 1112.5191 1115.2119 1117.9057 1120.6003 1123.2958 1125.9921 1128.6893 1131.3874

500 1134.0864 1136.7862 1139.4869 1142.1885 1144.8909 1147.5942 1150.2984 1153.0034 1155.7093 1158.4160
510 1161.1236 1163.8320 1166.5412 1169.2514 1171.9623 1174.6741 1177.3868 1180.1003 1182.8146 1185.5298
520 1188.2458 1190.9626 1193.6803 1196.3988 1199.1181 1201.8383 1204.5593 1207.2811 1210.0037 1212.7272
530 1215.4514 1218.1765 1220.9024 1223.6292 1226.3567 1229.0851 1231.8142 1234.5442 1237.2750 1240.0066
540 1242.7390 1245.4722 1248.2062 1250.9410 1253.6766 1256.4130 1259.1501 1261.8881 1264.6269 1267.3665

550 1270.1069 1272.8480 1275.5899 1278.3327 1281.0762 1283.8205 1286.5655 1289.3114 1292.0580 1294.8054
560 1297.5536 1300.3026 1303.0523 1305.8028 1308.5541 1311.3062 1314.0590 1316.8126 1319.5669 1322.3220
570 1325.0779 1327.8345 1330.5919 1333.3501 1336.1090 1338.8687 1341.6291 1344.3903 1347.1522 1349.9149
580 1352.6783 1355.4425 1358.2074 1360.9731 1363.7395 1366.5066 1369.2745 1372.0432 1374.8126 1377.5827
590 1380.3535 1383.1251 1385.8974 1388.6705 1391.4443 1394.2188 1396.9940 1399.7700 1402.5467 1405.3241
---- - ------------- ------------- ------ --· · - · · · - ----------------~--- -- ---~-
>-
"0
"0
m
z
Cl
......
X
 ;,:.
"'0
"'0
tTl
z
0
><
-
0 1 2 3 4 5 6 7 8 9

600 1408.1023 1410.8812 1413.6608 1416.4411 1419.2221 1422.0039 1424.7863 1427.5695 1430.3534 1433.1380
610 1435.9234 1438.7094 1441.4962 1444.2836 1447.0718 1449.8607 1452.6503 1455.4405 1458.2315 1461.0232
620 1463.8156 1466.6087 1469.4025 1472.1970 1474.9922 1477.7880 1480.5846 1483.3819 1486.1798 1488.9785
630 1491.7778 1494.5779 1497.3786 1500.1800 1502.9821 1505.7849 1508.5883 1511.3924 1514.1973 1517.0028
640 1519.8090 1522.6158 1525.4233 1528.2316 1531.0404 1533.8500 1536.6602 1539.4711 1542.2827 1545.0950
650 1547.9079 1550.7215 1553.5357 1556.3506 1559.1662 1561.9824 1564.7993 1567.6169 1570.4351 1573.2540
660 1576.0736 1578.8938 1581.7146 1584.5361 1587.3583 1590.1811 1593.0046 1595.8287 1598.6535 1601.4789
670 1604.3050 1607.1317 1609.9591 1612.7871 1615.6158 1618.4451 1621.2750 1624.1056 1626.9368 1629.7687
680 1632.6012 1635.4344 1638.2681 1641.1026 1643.9376 1646.7733 1649.6096 1652.4466 1655.2842 1658.1224
690 1660.9612 1663.8007 1666.6408 1669.4816 1672.3229 1675.1649 1678.0075 1680.8508 1683.6946 1686.5391
700 1689.3842 1692.2299 1695.0762 1697.9232 1700.7708 1703.6190 1706.4678 1709.3172 1712.1672 1715.0179
710 1717.8691 1720.7210 1723.5735 1726.4266 1729.2803 1732.1346 1734.9895 1737.8450 1740.7011 1743.5578
720 1746.4152 1749.2731 1752.1316 1754.9908 1757.8505 1760.7109 1763.5718 1766.4333 1769.2955 1772.1582
730 1775.0215 1777.8854 1780.7499 1783.6150 1786.4807 1789.3470 1792.2139 1795.0814 1797.9494 1800.8181
740 1803.6873 1806.5571 1809.4275 1812.2985 1815.1701 1818.0423 1820.9150 1823.7883 1826.6622 1829.5367
750 1832.4118 1835.2874 1838.1636 1841.0404 1843.9178 1846.7957 1849.6742 1852.5533 1855.4330 1858.3133
760 1861.1941 1864.0755 1866.9574 1869.8399 1872.7230 1875.6067 1878.4909 1881.3757 1884.2611 1887.1470
770 1890.0335 1892.9205 1895.8082 1898.6963 1901.5851 1904.4744 1907.3642 1910.2547 1913.1456 1916.0372
780 1918.9293 1921.8219 1924.7151 1927.6089 1930.5032 1933.3981 1936.2935 1939.1895 1942.0860 1944.9831
790 1947.8807 1950.7789 1953.6776 1956.5769 1959.~767 1962.3771 1965.2780 1968.179411971.081411973.9840

.1>-
<.0
-.J
 -""
!.0
00

TABLE IX (continued)
~~---

0 1 2 .3 4 5 6 7 8 9
--
800 1976.8871 1979.7907 1982.6949 1985.5996 1988.5049 1991.4107 1994.3170 1997.2239 2000.1313 2003.0392
810 2005.9477 2008.8567 2011.7663 2014.6764 2017.5870 2020.4982 2023.4099 2026.3221 2029.2348 2032.1481
820 2035.0619 2037.9763 2040.8911 2043.8065 2046.7225 2049.6389 2052.5559 2055.4734 2058.3914 2061.3100
830 2064.2291 2067.1487 2070.0688 2072.9894 2075.9106 2078.8323 2081.7545 2084.6772 2087.6005 2090.5242
840 2093.4485 2096.3733 2099.2986 2102.2244 2105.1508 2108.0776 2111.0050 2113.9329 2116.8613 2119.7902

850 2122.7196 2125.6495 2128.5800 2131.5109 2134.4424 2137.3744 2140.3068 2143.2398 2146.1733 2149.1073
860 2152.0418 2154.9768 2157.9123 2160.8483 2163.7848 2166.7218 2169.6594 2172.5974 2175.5359 2178.4749
870 2181.4144 2184.3545 2187.2950 2190.2360 2193.1775 2196.1195 2199.0620 2202.0050 2204.9485 2207.8925
880 2210.8370 2213.7820 2216.7274 2219.6734 2222.6198 2225.5668 2228.5142 2231.4621 2234.4106 2237.3595
890 2240.3088 2243.2587 2246.2091 2249.1599 2252.1113 2255.0631 2258.0154 2260.9682 2263.9215 2266.8752

900 2269.8295 2272.7842 2275.7394 2278.6951 2281.6513 2284.6079 2287.5650 2290.5226 2293.4807 2296.4393
910 2299.3983 2302.3579 2305.3179 2308.2783 2311.2393 2314.2007 2317.1626 2320.1250 2323.0878 2326.0511
920 2329.0149 2331.9792 2334.9439 2337.9091 2340.8748 2343.8409 2346.8075 2349.7746 2352.7421 2355.7102
930 2358.6786 2361.6476 2364.6170 2367.5869 2370.5572 2373.5281 2376.4993 2379.4711 2382.4433 2385.4159
940 2388.3891 2391.3627 2394.3367 2397.3112 2400.2862 2403.2616 2406.2375 2409.2139 2412.1907 2415.1679

950 2418.1457 2421.1238 2424.1025 2427.0816 2430.0611 2433.0411 2436.0216 2439.0025 2441.9839 2444.9657
960 2447.9479 2450.9307 2453.9138 2456.8975 2459.8815 2462.8661 2465.8511 2468.8365 2471.8224 2474.8087
970 2477.7954 2480.7827 2483.7703 2486.7584 2489.7470 2492.7360 2495.7255 2498.7154 2501.7057 2504.6965
980 2507.6877 2510.6794 2513.6715 2516.6640 2519.6570 2522.6505 2525.6443 2528.6387 2531.6334 2534.6286
990 2537.6242 2540.6203 2543.6168 2546.6138 2549.6112 2552.6090 2555.6073 2558.6059 2561.6051 2564.6046
2594.6241 ;J>
1,000 2567.6046 2570.6051 2573.6059 2576.6072 2579.6090 2582.6111 2585.6137 2588.6168 2591.6202 -o
-~ ------- --------- --- -o
m
z
0
......
><
APPENDIX 499

TABLE X

Common Logarithms of Numbers to Four Decimal Places.

Proportional Parta

-- --
6 8
-- -- -- -- -- --
N L. 0 1 2 3 4 5 7 9
--
1 2 3 4 5
-10 0000 0043
-- --
0086 0128 0170 0212 0253 0294 0334 0374
-4 - -
8 12 17 21
11 0414 0453 0492 0531 0569 0607 0645 0682 0719 0755 4 8 11 15 19
12 0792 0828 0864 0899 0934 0969 1004 1038 1072 1106 3 7 10 14 17
13 1139 1173 1206 1239 1271 1303 1335 1367 1399 1430 3 6 10 13 16
14 1461 1492 1523 1553 1584 1614 1644 1673 1703 1732 3 6 9 12 15

15 1761 1790 1818 1847 1875 1903 1931 1959 1987 2014 3 6 8 11 14
16 2041 2068 2095 2122 2148 2175 2201 2227 2253 2279 3 5 8 11 13
17 2304 2330 2355 2380 2405 2430 2455 2480 2504 2529 2 5 7 10 12
18 2553 2577 2601 2625 2648 2672 2695 2718 2742 2765 2 5 7 9 12
19 2788 2810 2833 2856 2878 2900 2923 2945 2967 2989 2 4 7 9 11

2!0 3010 3032 3054 3075 3096 3118 3139 3160 3181 3201 2 4 6 8 11
21 3222 3243 3263 3284 3304 3324 3345 3365 3385 3404 2 4 6 8 10
22 3424 3444 3464 3483 3502 3522 3541 3560 3579 3598 2 4 6 8 10
23 3617 3636 3655 3674 3692 3711 3729 3747 3766 3784 2 4 6 7 9
24 3802 3820 3838 3856 3874 3892 3909 3927 3945 3962 2 4 5 7 9

25 3979 3997 4014 4031 4048 4065 4082 4099 4116 4133 2 4 5 7 9
26 4150 4166 4183 4200 4216 4232 4249 4265 4281 4298 2 3 5 7 8
27 4314 4330 4346 4362 4378 4393 4409 4425 4440 4456 2 3 5 6 8
28 4472 4487 4502 4518 4533 4548 4564 4579 4594 4609 2 3 5 6 8
29 4624 4639 4654 4669 4683 4698 4713 4728 4742 4757 1 3 4 6 7

30 4771 4786 4800 4814 4829 4843 4857 4871 4886 4900 1 3 4 6 7
31 4914 4928 4942 4955 4969 4983 4997 5011 5024 5038 1 3 4 5 7
32 5051 5065 5079 5092 5105 5119 5132 5145 5159 5172 1 3 4 5 7
33 5185 5198 5211 5224 5237 5250 5263 5276 5289 5302 1 3 4 5 7
34 5315 5328 5340 5353 5366 5378 5391 5403 5416 5428 1 2 4 5 6

35 5441 5453 5465 5478 5490 5502 5514 5527 5539 5551 1 2 4 5 6
36 5563 5575 5587 5599 5611 5623 5635 5647 5658 5670 1 2 4 5 6
37 5682 5694 5705 5717 5729 5740 5752 5763 5775 5786 1 2 4 5 6
38 5798 5809 5821 5832 5843 5855 5866 5877 5888 5899 1 2 3 5 6
39 5911 5922 5933 5944 5955 5966 5977 5988 5999 6010 1 2 3 4 5

40 6021 6031 6042 6053 6064 6075 6085 6096 6107 6117 1 2 3 4 5
41 6128 6138 6149 6160 6170 6180 6191 6201 6212 6222 1 2 3 4 5
42 6232 6243 6253 6263 6274 t\284 6294 6304 6314 6325 1 2 3 4 5
43 6335 6345 6355 6365 6375 6365 6395 6405 6415 6425 1 2 3 4 5
44 6435 6444 6454 6464 6474 6484 6493 6503 6513 6522 1 2 3 4 5

45 6532 6542 6551 6561 6571 6580 6590 6599 6609 6618 1 2 3 4 5
46 6628 6637 6646 6656 6665 6675 6684 6693 6702 6712 1 2 3 4 5
47 6721 6730 6739 6749 6758 6767 6776 6785 6794 6803 1 2 3 4 5
48 6812 6821 6830 6839 6848 6657 6666 68i5 6884 6893 1 2 3 4 5
49 6902 6911 6920 6928 6937 6946 6955 6964 6972 6981 1 2 3 4 4

50 6990 6998 7007 7016 7024 7033 7042 7050 7059 7067 1 2 3 3 4
51 7076 7084 7093 7101 7110 7118 7126 7135 7143 7152 1 2 3 3 4
52 7160 7168 7177 7185 7193 no2 7210 7218 7226 7235 1 2 3 3 4
53 7243 7251 7259 7267 7275 7284 7292 7300 7308 7316 1 2 2 3 4
3 4
-N -- -- -- --5 --6 --7 --8 --9
7372 7380 7388 7396 1 2 2
--4
7356 7364
54 7324 7332 7340
2
7348
--
3
-1 -2 -3 -4 -5
L. 0 1
500 APPENDIX

TABLE X (continued)

Proportional parts
L. 1 3 4 5 6 7 8 9
-- -- -- -- -- -- -- -- -- --
N 0 2 1 2 3 4 5
-55 7404 7412 7419 7427 7435 7443 7451 7459 7466 7474
-1 -2 -2 -3 -
4
56 7482 7490 7497 7505 7513 7520 7528 7536 7543 7551 1 2 2 3 4
57 7559 7566 7574 7582 7589 7597 7604 7612 7619 7627 1 1 2 3 4
58 7634 7642 7649 7657 7664 7672 7679 7686 7694 7701 1 1 2 3 4
59 7709 7716 7723 7731 7738 7745 7752 7760 7767 7774 1 1 2 3 4

60 7782 7789 7796 7803 7810 7818 7825 7832 7839 7846 1 1 2 3 4
61 7853 7860 7868 7875 7882 7889 789-6 7903 7910 7917 1 1 2 3 3
62 7924 7931 7938 7945 7952 7959 7966 7973 7980 7987 1 1 2 3 3
63 7993 8000 8007 8014 8021 8028 8035 8041 8048 8055 1 1 2 3 3
64 8062 8069 8075 8082 8089 8096 8102 8109 8116 8122 1 1 2 3 3

65 8129 8136 8142 3149 8156 8162 8169 8176 8182 8189 1 1 2 3 3
66 8195 8202 8209 8215 8222 8228 8235 8241 8248 8254 1 1 2 3 3
67 8261 8267 8274 8280 8287 8293 8299 8306 8312 8319 1 1 2 3 3
68 8325 8331 8338 8344 8351 8357 8363 8370 8376 8382 1 1 2 3 3
69 8388 8395 8401 8407 8414 8420 8426 8432 8439 6445 1 1 2 3 3
70 8451 8457 8463 8470 8476 8482 8488 8494 8500 8506 1 1 2 3 3
71 8513 8519 8525 8531 8537 8543 8549 8555 8561 8567 1 1 2 3 3
72 8573 8579 8585 8591 8597 8603 8609 81315 8621 8627 1 1 2 3 3
73 8633 8639 8645 8651 8657 8663 8669 8675 8681 8686 1 1 2 2 3
74 8692 8698 8704 8710 8716 8722 8727 8733 8739 8745 1 1 2 2 3

75 8751 8756 8762 8768 8774 8779 8785 8791 8797 8802 1 1 2 2 3
76 8808 8814 8820 8825 8831 8837 8842 8848 8854 8859 1 1 2 2 3
77 8865 8871 8876 8882 8887 8893 8899 8904 8910 8915 1 1 2 2 3
78 8921 8927 8932 8938 8943 8949 8954 8960 8965 8971 1 1 2 2 3
79 6976 8982 8987 8993 8998 9004 9009 9015 9020 9025 1 1 2 2 3
80 9031 9036 9042 9047 9053 9058 9063 9069 9074 9079 1 1 2 2 3
81 9085 9090 9096 9101 9106 9112 9117 9122 9128 9133 1 1 2 2 3
82 9138 9143 9149 9154 9159 9165 9170 9175 9180 9186 1 1 2 2 3
83 9191 9196 9201 9206 9212 9217 9222 9227 9232 9238 1 1 2 2 3
84 9243 9248 9253 9258 9263 9269 9274 9279 9284 9289 1 1 2 2 3
8& 9294 9299 9304 9309 9315 9320 9325 9330 9335 9340 1 1 2 2 3
86 9345 9350 9355 9360 9365 9370 9375 9380 9385 9390 1 1 2 2 3
87 9395 9400 9405 9410 9415 9420 9425 9430 9435 9440 1 1 2 2 3
88 9445 9450 9455 9460 9465 9469 9474 9479 9484 9489 0 1 1 2 2
89 9494 9499 9504 9509 9513 9518 9523 9528 9533 9538 0 1 1 2 2
90 9542 9547 9552 9557 9562 9565 9571 9575 9581 9586 0 1 1 2 2
91 9590 9595 9600 9605 9609 9614 9619 9624 9628 9633 0 1 1 2 2
92 9638 9643 9647 9652 9657 9661 9666 9671 9675 9680 0 1 1 2 2
93 9685 9689 9694 9699 9703 9708 9713 9717 9722 9727 0 1 1 2 2
94 9731 9736 9741 9745 9750 9754 9759 9763 9768 9773 0 1 1 2 2
95 9777 9782 9786 9791 9795 9800 9805 9809 9814 9818 0 1 1 2 2
96 9823 9827 9832 9836 9841 9845 9850 9854 9859 9863 0 1 1 2 2
97 9868 9872 9877 9881 9886 9890 9894 9e99 9903 9908 0 1 1 2 2
98 ~912 9917 9921 9926 9930 9934 9939 9943 9948 9952 0 1 1 2 2

- -- -- -- -- -- -- -- -- -- -- - -
99
N L.
9956
0
9961
1
9965
2
9969
3
9974
4
9978
5
9983
6
9987
7
9991
8
9996
9
0
1 2
1 1
-3 -4 -5
2 2
 APPENDIX SOl

TABLE XI

Sequential Constants a and b from a and S

Base 10 logarithms
a if finding a, 8 if finding b

. 001 .01 .02 .03 .OS .10

. 001 3.000 2.000 1.699 1.522 1. 301 1.000
.01 2.996 1.996 1.69S l.Sl9 1.297 .996
8 i f finding a . 02 2.991 1. 991 1.690 1. Sl4 1. 292 . 991
a i f finding b .03 2.987 1. 987 1.686 1. SlO 1. 288 .987
.OS 2.978 1.978 1.677 1. 501 1. 279 .978
.10 2.9S4 1.9S4 1.6S3 1.477 1. 255 .954

1-8 1-a
a = log(-)
a b = log(T)

Base e logarithms
a if finding a, S i f finding b

. 001 . 01 . 02 . 03 . OS .10

• 001 6.907 4.604 3.911 3.506 2.995 2.302

.01 6.898 4.S9S 3.902 3.497 2.986 2.293
8 i f finding a .02 6.888 4.58S 3.892 3.486 2.976 2.282
a if finding b .03 6. 877 4.S7S 3.882 3.476 2.96S 2.272
.OS 6.8S6 4.SS4 3.861 3.4SS 2.944 2.2S1
.10 6.802 4.500 3.807 3.401 2.890 2.197

a = 1ne~ 8 ) b = ln (1-a)
- 8-
502 APPENDIX

TABLE XII
Sequential Range Test for Process Capability. To Meet Total Tolerance
ofT. Acceptable a = T/10, Rejectable a = T/6, with Corresponding Risks
a and S of Wrong Decisions. Ranges for Samples of n = 8. Multiply Given
Acceptance and Rejection Numbers by Specified T.*

Multiples of Tolerance T
No. of Cumulated for Acc.-Rej. Criteria
Ranges Sample Risks a = ,05 Risks a = •01
in Sum Size Ace, No. Rej, No, Ace. No. Rej. No.
1 8 0.19 0.54 0.09 0.63
2 16 .55 .90 .46 1. 00
3 24 .92 1.26 .82 1.36
4 32 1.28 1.63 1.18 1. 73
5 40 1.64 1. 99 1. 55 2.09
6 48 2.01 2.36 1. 91 2.45
7 56 2.37 2.72 2.27 2.82
8 64 2.91 2.91 2.64 3.18
9 72 3.00 3.54
10 38 3.36 3.91
11 88 3.73 4.27
12 96 4.36 4.36

* From H. M. Wies and I. W. Burr, Simple capability acceptance test,

a sequential test on ranges. Indust. Quality Control 21 (No. 5), 266-268
(1964), with kind permission of American Society for Quality Control.
 ;...
't:l
't:l
m
z
0
H
><

TABLE XIII

Acceptance and Rejection Numbers for One-way Sequential Check of Process Setting.
Maximum Risks of Wrong Decisions at pl and ~ 2 , a= S = .10. Given for Classes of a
X
Code x's So that 1 < aX<- 10.

Protection Against
Proc.Aver.High by: l. 21 1.46 1.77 2.15 2.61 3.16
Proc.Std.Dev.a 1. 00-1.21 1.22-1.46 1.47-1.77 1.78-2.15 2.16-2.61 2.62-3.16
X
Ace. Rej. Ace. Rej. Ace. Rej. Ace. Rej. Ace. Rej. Ace. Rej.
Sample Size 1 -2.1 +3.3 -2.5 +3.9 -3.1 +4.8 -3.7 +5.8 -4.4 +7.1 -5.4 +8.5
2 -1.5 3.9 -1.8 4.7 -2.1 5.7 -2.6 6.9 -3.1 8.3 -3.8 10.1
3 - .8 4.5 -1.1 5.4 -1.2 6.6 -1.5 7.9 -1.8 9.7 -2.2 11.7
4 - .2 5.1 - .3 6.2 - .4 7.5 - .4 9.1 - .5 10.9 - .6 13.3
5 + .4 5.7 + .4 6.9 + .5 8.4 + .7 10.1 + .8 12.3 + .9 14.9
6 .9 6.3 1.2 7.6 1.4 9.2 1.7 11.2 2.1 13.6 2.5 16.4
7 1.6 6.9 1.9 8.4 2.3 10.1 2.8 12.3 3.4 14.9 4.1 18.1
8 2.2 7.5 2.6 9.1 3.2 11.1 3.9 13.4 4.7 16.2 5.7 19.6
9 2.8 8.1 3.4 9.8 4.1 11.9 4.9 14.4 6.1 17.5 7.3 21.2
10 3.4 8.7 4.1 10.6 4.9 12.8 6.1 15.5 7.3 18.8 8.9 22.8
11 6.0 6.1 8.0 8.1 9.8 9.9 ll.8 11.9 14.3 14.4 17.3 17.4

(J1
0
(.N
 <Jl
0
-1:>

TABLE XIII (continued)

Protection
Margin, lax 3.83 4.64 5.62 6.81 8.25 10.00
Std. Dev.a 3.17-3.83 3.84-4.64 4.65-5.62 5.63-6.81 6.82-8.25 8.26-10.0
X
Ace. Rej . Ace. Rej. Ace. Rej. Ace. Rej. Ace. Rej .. Ace. Rej.
Sample Size 1 - 6.5 +10.3- 7.9 +12.5 - 9.5 +15.2 -11.6 +18.4 -14.1 +22.3 -16.9 +26.9
2 - 4.6 12.2 - 5.6 14.8 - 6.7 17.9- 8.2 21.8 - 9.9 26.4 -11.9 31.9
3 - 2.7 14.2 - 3.2 17.2 - 3.9 20.8 - 4.8 25.2 - 5.8 30.5 - 6.9 36.9
4 - .8 16.1 - .9 19.5 - 1.1 23.6 - l. 3 28.6 - l. 6 34.6 - l. 9 41.9
5 + l. 2 17.9 + 1.4 21.8 + 1.7 26.4 + 2.1 32.1 + 2.5 38.8 + 3.1 46.9
6 3.1 19.9 3.7 24.1 4.5 29.2 5.5 35.4 6.6 42.9 8.1 51.9
7 4.9 21.8 6.1 26.4 7.3 32.1 8.9 38.8 10.8 47.1 13.1 56.9
8 6.9 23.7 8.4 28.8 10.1 34.8 12.3 42.2 14.9 51.2 18.1 61.9
9 8.8 25.7 10.7 31.1 12.9 37.7 15.7 45.6 19.1 55.3 23.1 66.9
10 10.7 27.6 13.1 33.4 15.8 40.5 19.1 49.1 23.1 59.4 28.1 71.9
11 21.1 21.2 25.5 25.6 30.9 31.0 37.4 37.5 45.4 45.5 55.0 55.1

>-
tTl
"'"'
z
0
....
X
 TABLE XIV

Accept ance and Rejection Numb ers for Two-way Sequential Check of Process Setting.
~1aximum Ri s ks of Wrong Decisions at~ . and at~.:':. ox are a= .10, 13 = .10. Given
for Classes of ox Code x 1 s so that l ~ox~ 10.

Protection Against
"'P-;.r.::.oc::..:. :A.:..:v..;;e,;-r-'
. .;O::,f;;.:f::_::b:.t.Y..:.:-------,,....,-"1:-'-.""2~1 l. 4 6 l. 77 2. l 5 2 . 61 3. 16
Class, Std.Dev . Ox 1 . 00-1 . 21 1.22-1.46 1.47-1.77 1. 78-2.15 2.16-2.61 2.62-3.16
Ace. Rej . Ace. Rej . Ace. Rej. Ace. Rej. Ace . Rej. Ace . Rej.
Sample Size 1 4.1 4.9 6. 1 7.3 8.8 10.7
2 4. 7 5.7 6.9 8.4 10.2 12.3
3 5. 3 6.4 7.8 9.5 11.5 13. 9
4 5.9 7. 2 8.7 10.5 12.8 15.5
5 . 9 6 . 5 1.2 7.9 1.5 9.6 1.8 11.6 2.1 14. 1 2.6 17 . 1
6 1.7 7 . 1 2 . 1 8 . 6 2.6 10.5 3 . 1 12.7 3 . 8 15 . 4 4.6 18 . 6
7 2 . 4 7.7 2 . 9 9.4 3.5 11.4 4 . 3 13.8 5.2 16.7 6.3 20.2
8 3.1 8.3 3.7 10.1 4.4 12.3 5.4 14.8 6.5 17 . 9 7 . 9 21.8
9 3.6 8 . 9 4.4 10.8 5.3 13.1 6.4 15.9 7.8 19.3 9.5 23 4
10 4.2 9.6 5.1 11.6 6 . 2 14.1 7.5 16.9 9 . 1 20.6 11. 1 24.9
11 4 . 8 10.2 5.9 12 . 3 7.1 14.9 8.6 18.1 10.4 21.9 12.6 26.5
12 5 . 4 10 . 8 6.6 13 . 1 7.9 15.8 9.7 19 . 1 11.7 23 . 2 14.2 28 . 1
13 6.1 11.4 7.3 13.8 8.9 16 . 7 10.8 20.2 13.1 24.5 15.8 29 . 7
14 6.7 11.9 8.1 14 . 5 9.8 17.6 11.8 21.3 14.3 25 . 8 17 . 4 31.3
15 7 . 3 12.6 8.8 15.3 10.7 18.5 12.9 22.4 15.7 27 .l 18.9 32.9
16 7.9 13.2 9.5 15 . 9 11.6 19 . 4 13 . 9 23.5 16.9 28.4 20 . 5 34.4
17 8 . 5 13.8 10.3 16.7 12.4 20.3 15 . 1 24.5 18.3 29 . 7 22 . 1 36 . 1
18 9.1 14.2 11.1 17.2 13.3 20.9 16.1 25.3 19.6 30.6 23.7 37.1
19 9.7 14.2 11.7 17.2 14.2 20.9 17.2 25.3 20.9 30.6 25.3 37.1
20 10.3 14.2 12.5 17 . 2 15.1 20.9 18.3 25.3 22.2 30 . 6 26.9 37.1
21 10.9 14.2 13.2 17.2 15.9 20 . 9 19.4 25.3 23.5 30.6 28.4 37.1
22 14 . 1 14.2 17 . 1 17.2 20.8 20.9 25.2 25 . 3 30.5 30.6 37 . 0 37.1
Protection Margin, l ox 3.83 4.64 s.62 6 . 81 8 . 25 10 . 00
Class, Std. Dev. ox 3.17-3.83 3 . 84-4.64 4.65-5 . 62 5.63-6.81 6.82-8.25 8 . 26-10.00
Ace. Rej. Ace. Rej. Ace. Rej. Ace. Rej . Ace . Rej . Ace . Rej.
1 12 . 9 15.7 19. 1 23. 1 27 . 9 33.9
2 14.9 18.1 21.9 26.5 32 . 1 38.9
3 16.8 20.4 24. 7 29. 9 - 36. 2 43.9
4 18 . 7 22.7 27.5 33.3 40.4 48 . 9
5 3.1 20 . 7 3 . 8 25.1 4 . 6 30.3 5.6 36.7 6 . 8 44.5 8.2 53 . 9
6 5.5 22.6 6.7 27 . 3 8.1 33.1 9.8 40 . 1 11 . 9 48.6 14.4 58.9
7 7. 6 24 . 5 9 . 2 29.7 11. l 35.9 13.5 43.5 16 . 3 52.7 19.8 63.9
8 9.5 26.4 11.6 31.9 13 . 9 38 . 7 16.9 46.9 20.5 56.9 24.9 68.9
9 11.5 28.3 13.9 34 . 3 16 . 8 41.6 20.4 50.4 24.7 61.1 29.9 73.9
10 13.4 30.2 16.2 36 . 6 19.7 44.4 23.8 53.8 28.8 65.1 34 . 9 78.9
11 15.3 32.1 18.5 38 . 9 22.5 47.2 27.2 57.2 32.9 69.3 39.9 83.9
12 17.2 34.1 20 . 9 4 I. 3 25.3 49.9 30.6 60.6 37.1 73.4 44.9 88 . 9
13 19.1 35 . 9 23.2 43 . 6 28.1 52.8 34.1 63.9 41.2 77.5 49.9 93 . 9
14 21.1 37 . 9 25 . 5 45.9 30.9 55.6 37 . 4 67.4 45.4 81.6 54.9 98 . 9
15 22.9 39 . 8 27.8 48.2 33 . 7 58 . 4 40.8 70 . 8 49.5 85 .8 59.9 103 . 9
16 24. 9 41.7 30.2 50 . 5 36.5 61 . 2 44.3 74.2 53.6 89 . 9 64.9 108 . 9
17 26 . 8 43.6 32.5 52.9 39.3 64.1 47 . 7 77 . 6 57 . 7 94 . 1 69.9 113 . 9
18 28.7 44.9 34.8 54 . 4 42 . 2 65.8 51.1 79.7 61.9 96.6 74.9 117 . 0
19 30.6 44.9 37.1 54.4 44.9 65.8 54.5 79.7 65 . 9 96.6 79 . 9 117.0
20 32.6 44 . 9 39.4 54.4 47 . 8 65 . 8 57.9 79.7 70.1 96.6 84.9 11 7. 0
21 34 . 5 44 . 9 41.8 54.4 50 . 6 65.8 61.3 79.7 74.3 96.6 89.9 117.0
22 44 . 8 44 . 9 54.3 54.4 65.7 65.8 79 . 6 79.7 96.5 96 . 6 116.9 117 . 0
 TABLE XV

Parameters for a General System of Distributions (15.62)' for Selected

Curve Shape Characteristics Cl.3 and a 4 .

"'a "'4 c k 1-' a

.00 2.8 3.939 19.865 .4275 .1243

3.0 4.874 6. 158 .6447 . 1620
3.2 6.065 3.745 .7673 .1644
3.4 7.696 2.701 .8507 .1519
3.6 10.182 2.090 .9111 .1300
.10 2.8 3.520 19.606 .3901 .1257
3.0 4.297 6.283 .6077 .1714
3.3 5.832 3.235 .7830 .1776
3.6 8.298 2.164 .8896 .1539
3.9 13.716 1.549 .9595 .1097

.20 2.8 3.071 30.547 .2957 .1069

3.0 3.705 7.244 .5420 . 1737
3.3 4.895 3.548 .7347 .1944
3.6 6.640 2.353 .8538 .1803
3.9 9.858 1.694 .9352 .1447
.30 3.0 3.140 9.872 .4411 .1617
3.3 4.058 4.168 .6636 .2054
3.6 5.289 2.684 .8000 .2054
3.9 7.227 1.933 .8951 .1818
4.2 11.492 1.410 .9648 .1353
.40 3.0 2.625 19.800 .2887 .1214
3.3 3.341 5.369 .5638 .2039
3.6 4.226 3.223 .7240 .2235
3.9 5.449 2.289 .8353 .2152
4.2 7.492 1.706 .9190 . 1863
.50 3.2 2.538 12.523 .3353 .1477
3.5 3.162 4.927 .5653 .2172
3.8 3.917 3. 156 .7134 .2378
4.1 4.925 2.306 .8211 .2332
4.4 6.514 I. 752 .9054 .2093
.60 3.2 2.066 190.007 .D700 .0356
3.6 2.738 5.983 .4836 .2086
4.0 3.549 3.253 .6859 .2503
4.4 4.673 2.219 .8232 .2484
4.8 6.748 1.574 .9287 .2141
.70 3.5 2.065 20.670 .2080 .1087
3.9 2.656 5.246 .5017 .2258
4.3 3.354 3.155 .6815 .2641
4.7 4.292 2.238 .8105 .2656
5. 1 5.896 1.635 .9141 .2386
.80 3.8 2.005 14.088 .2434 .1325
4.3 2.653 4.406 .5413 .2494
4.8 3.448 2.717 .7260 .2823
5.j 4.651 1.897 .8616 .2727
5.8 8.262 1.204 .9858 .2044
 GLOSSARY OF SYMBOLS

A multiplier of a for ± limits for x, around ]..1

A an event (in probability)

multiplier of R for ± 1imi ts for x, around X

multiplier of s for - around =X

± limits for x,

Ac acceptance number for variable, single sampling

Ac 1 acceptance number on first sample for defectives or defects

Ac 2 acceptance number after two samples for total number of defectives

or defects

AOQ average outgoing quality

average fraction defective of all lots, after screening of rejected

lots

AO~,= AOQ if incoming quality is at p'

AOQL average outgoing quality limit

maximum of AOQ's for all p 1 's for a plan

AQL acceptable quality level

a nominal lot or process level considered satisfactory by consumer

ARL average run length

average number of trials to a decision

ASN average sample number

average number of inspections or tests till a decision is reached

ATI average total inspection

average number of inspections per lot, including accepted and rejected

lots

507
 508 GLOSSARY OF SYMBOLS

B an event (in probability)

B3 multiplier of 5 to give lower limit for s's

B4 multiplier of s- to give upper limit for s's

Bs multiplier of o to give lower limit for s's

B6 multiplier of o to give upper limit for s's

Cf= symbol for center line

c = number of defects on a sample of product

c = acceptable number of defectives, in single sampling

c1 acceptance number on first sample for defectives or defects

c2 acceptance number after two samples for total number of defectives

or defects (like Ac 2 )

c4 multiplier of o to give theoretical average of s's

c5 multiplier of o to give theoretical standard deviation of s's

c' true or theoretical average number of defects per sample of

specified size

c- average number of defects per sample

C' true or theoretical average number of defects per 100 units or pieces

CR consumer's risk, or probability of erroneous acceptance of a lot

at some given unacceptable quality, if offered

ChSP chain sampling plan

CSP continuous sampling plan

C (n, d) number of different possible combinations of d objects taken from n

distinct ones

d number of defective or non-conforming pieces in a sample of n pieces

d1 number of defective or non-conforming pieces in the first sample

d2 multiplier of o to give theoretical average of R's

d2 number of defective or non-conforming pieces in a second sample

GLOSSARY OF SYMBOLS 509

d3 multiplier of a to give theoretical standard deviation of R's

D number of defective or non-conforming pieces in lot

D total number of demerits on a sample or samples

Dl multiplier of a to give lower limit for R's

D2 multiplier of a to give upper limit for R's

D3 multiplier of R to give lower limit for R's

D4 multiplier of R to give upper limit for R's

e = natural logarithmic base, 2.71828 ...

E expected value or theoretical average of whatever variable is

written in parentheses after E

f proportion of units sampled in continuous plans

fi observed frequency of cases in i'th class

f(x) probability density function of x

Fi theoretical or calculated frequency of cases in i'th class

F(y) probability that variable x 2 y

H1 basic or null hypothesis, often that quality level is satisfactory

H2 alternate hypothesis, often that the quality characteristic is at

some unsatisfactory level

i number of a frequency class or variable

i number of units, all good in succession, to qualify for sampling in

continuous sampling plans CSP

i number of previous lot-samples to be free of defectives to tolerate

a defective in present lot in ChSP

i required number of consecutive lots, all accepted, to qualify for

skipping in SkSP

k number of samples

k an amount of departure from ~O considered rejectable, for ~

k a multiple for s, so that if x+ ks 2 U, accept

510 GLOSSARY OF SYMBOLS

K acceptance-rejection limit for x or for s 2

K number of units inspected per 1000 if always on sampling, in CSP
L lower tolerance or specification limit for individual x's

LCL lower control limit

LQ 10 that fraction defective having Pa of .10

LTPD lot tolerance proportion (or percent) defective

a nominal fraction defective to be accepted only a small proportion

of times offered

n = sample size, number of observations in sample

nl number of pieces in first sample

n2 number of pieces in second sample

n-1 number of degrees of freedom in a sample of n

n average of a set of sample sizes n

np number of defectives in a sample of n

N number of pieces or numbers in a lot or population

N(A) number of distinct ways event A can occur

N(A and B) number of distinct ways in which both A and B can

simultaneously occur

N(AjB) number of distinct ways in which event A can occur, given that

event B did occur

p sample fraction defective (= d/n)

p average fraction defective over several samples (= Ed/En)

p' true or theoretical fraction defective

D/N in finite lot

Pi an AQL fraction defective (often Pgs)

P2 an LTPD fraction defective (often Pio)
Pia that fraction defective having a 10% chance of acceptance

whenever offered to a plan

GLOSSARY OF SYMBOLS 511

Pso = that fraction defective having a SO% ch~ce of accept~ce

whenever offered to a plan

Pgs that fraction defective having a 95% chance of accept~ce

whenever offered to a pl~

estimated fraction defective below L

lot toler~ce per cent defective in Dodge-Romig (= Pial

estimated fraction defective above U

theoretical average of defectives per 100 units

Pa = probability of acceptance for a given plan on some given quality

level

PR producer's risk, or probability of erroneous rejection of a

lot at acceptable quality, if offered

P(A} probability of event A occurring

P(AjB) probability that event A occur, given that event B did occur

P(A ~d B) probability that both of events A and B occur on the trial

q sample fraction good or conforming (= 1-p)

q a quality characteristic in general

q' lot or process fraction good or conforming (= 1-p')

QL quality index at L in MIL-STD 414 (like -z)

Qu quality index at U in MIL-STD 414 (like z)

R r~ge, maximum minus minimum values in sample

R average of a set of ranges

Re rejection number for a single sample

Re 1 rejection number for first sample

Re 2 rejection number for total defectives or defects on first two

samples

RQL rejectable quality level on some characteristic

512 GLOSSARY OF SYMBOLS

s = sample standard deviation, square root of variance

sx sample standard deviation of x, (~
average of a set of s's
2
s sample variance
n 2
sample variance, ~ (x-x) /(n-1)
1
s the space of all possible outcomes of an experiment
SkSP skip lot sampling plan

t a standardized variable

(X-IJ)/(s/lii)

T tolerance (= U-L)

u = defects per unit

(total defects on n units)/n

u' true or theoretical average of u's for a process

-
u average of a set of u's

0 average demerits per unit

U =upper tolerance or specification limit for individual x's

UCL upper control limit

Vx coefficient of variation (crx/!Jx)

wA demerit points for each class A defect, w8 ,wC,wD similar

x observed measurement

xi the i'th x in a sample

n
an average of a sample of x's = ~ x/n
k 1
X average of a set of x•s = ~ x/k
1
y a measurable characteristic of an assembly

z standardized variable = (x-IJ)/cr

X

za normal curve value with a probability above J $(u)du = a

z
a
GLOSSARY OF SYMBOLS 513

Z random variable
a = probability of rejecting hypothesis H1 when true

probability of rejecting lot or process when it is "acceptable"

producer's risk

a3 a measure of lack of symmetry or skewness of a frequency distribution

E[ (a-11) 3 ]/cr 3

a4 similar, measuring kurtosis or contact

S probability of accepting hypothesis H1 when false by some amount

probability of accepting lot or process when it is rejectable

consumer's risk

r (k) gamma function of k J xk e -X dx

0
x-11x
11 population mean or average of some random variable

population or process mean of x's

a desired or nominal mean of x's

a particular value of 11, often an acceptable value

a particular value of 11, often a rejectable value

E[(X-11) 2 )
k
E[ (x-11) ] , special cases 11 3 ,11 4

E(xk), 113,114 special cases

11 estimate of 11, often x

\) = degrees of freedom
2
<P(z) e -z /2 j/2; = normal curve density function
z 2
<li (z) J (eu 12 ;!2;) du = cumulative normal function
P (z or less)

cr = population, lot or process standard deviation

~[(x-11) 2 ], for x's

514 GLOSSARY OF SYMBOLS
- 2 ,s, R
population standard devi ation of x,x,s
ax,ax,as2' as,aR
2
a population, lot or process variance

E[(x-~) 2 ], for x's

ax an estimate of ax (might use s, 5;c 4 or R/d 2)

L summation of whatever follows the sigma

x2 chi square random var iabl e

(n-l)s 2 /a 2 for normal curve samples

2
k (f.-F . )
i' 1 1
L for testing goodness of fit
i=l Fi
 AUTHOR INDEX

Deming, W. E., 216

Divers, C. K., 187, 197
Abbott, IV. H., 214 Dodge, H. F., 2, 216,271, 274, 275,
Amber, G. H., 190, 201 310, 322, 382, 387, 391, 392, 395,
Amber, P. S., 190, 201 401, 402
Ande~son, J. S., 220 Dudley, J. W., 217
Anderson, P. S. M., 214 Duffy, D. J., 217
Anderson, V. L., 220 Duncan, A. J., 189, 200, 217, 396
Armstrong, G. R., 215
Armstrong, W. M., 215
Eagle, A. R.' 217
Ebeoglu, N. M., 217
Bailey, L. E., 217 Edwards, G. D.' 2
Bayer, H. S., 220 Eichelberger, L. s.' 189, 200, 217
Beall, G., 187, 197 Eilon, s.' 217
Behnken, D. W., 186, 197 Eisenhart, c.' 189, 200
Bennett, G., 417, 436
Enrick, N. L.' 189, 199, 217
Bicking, C. A., 190, 201, 215
Erdman, E. J.' 217
Bidlack, W., 215 Evans, A. A., 217
Bingham, R. S., Jr., 215
Bolton, H. R., 215
Bowker, A. H., 351
Boyd, D. F., 188, 198 Feigenbaum, A. V., 217
Breunig, H. L., 215 Ferrell, E. B., 189, 200
Browne, D. G., 216 Filimon, V., 217
Brumbaugh, M. A., 216, 220 Frazier, D., 217
Burr, I. W., 96, 106, 196, 216, 272, 351, Freund, R. A., 187, 188, 196, 197, 198
379, 380, 435, 436, 463 Frey, W. C., 217
Fry, T. C., 2

Carter, C. W., 219

Carver, H. C., 463 Ghering, L. G., 218, 463
Chateauneuf, R., 216 Gibian, E. F., 218
Clancey, V. J., 216 Gibra, I. N., 186, 188, 197, 198
Clark, B. L., 216 Gnaedinger, P. E., 218
Clarke, P. C., 215 Goepfert, W. P., 218
Clifford, P. C., 189, 200, 216 Goetz, B. E., 188, 198
Close, E. R., 215 Good, R. W., 218
Cochran, W. G., 463 Goode, H. P., 351
Coon, H. J., 218 Gore, IV. L., 218
Copelin, E. C., 96 Grant, E. L., 79, 96
Culbertson, J. E. Gross, H. D., 218, 222
Curcio, F. L., 220 Grubbs, F. E., 189, 200, 218
Grunwald, R. L., 218

Dalton, A. G., 21 6
Davies, J. A., 216 Hamaker, H. C., 318
Davis, J. W., 188, 199 Hamlin, C. K., 218

515
516 AUTHOR INDEX

Hammer, P. C., 187, 188, 197, 198 Mills, D., 220

Harrison, H. R., 218 Mitchell, J. A., 220
Hayes, A. S., 218 Mitten, L. G., 188, 198
Hecht, B., 219 Molina, E. C., 2, 463
Heinrich, G. F., 187, 197 Moore, P. G., 188, 198
Herne, H., 214 Moore, V. F., 220
HerroLd, G. V., 219 Mundel, A. B., 196
Hill, D., 187, 198, 219
Hilliard, J. E., 189, 200
Hillier, F. S., 188, 199 Newchurch, E. J., 220
Hoefs, R. H., 96 Nieman, L. J. H., 97, 106
Hoover, C. E., 219 Noble, C. E., 188, 198
Howell, J. M., 189, 200 Noel, R. H., 220

James, R. E., 417, 435, 436 Occasione, J. F., 220

Jenney, B. W., 219 Ohmae, Y., 322
Johnson, N. L., 396, 402 Oladko, A., 220
Johnson, R. C., 190, 201 Olds, E. G., 40, 72, 96, 189, 199, 220
Juran, J. M., 219 Olmstead, P. S., 96, 220
von Osinski, R., 189, 200, 221
Ostle, B., 189, 200
Kallet, F. T., 190, 201 Ott, E. R., 96, 188, 196, 198, 221
Kendall, M. G., 463 Owen, D. B., 272, 351
Kerr, W. A., 219
King, E. P., 219
King, J. R., 217, 219 Pabst, W. R., Jr., 322
Kirkpatrick, R. L., 272 Page, E. S., 188, 189, 198, 199
Kitagawa, T., 463 Peach, P., 272
Klingel, A., 217 Pearson, E. S., 96, 105, 106
Klock, A. G. , 219 Perry, R. L., 395, 402
Knowler, L.A., 220 Phocas, S. P., 217
Knowler, W. C., 217 Pippitt, R. G., 221
Koyama, T., 322 Pringle, J. B., 221
Proschan, F., 188, 199

Larson, K. E., 315, 322

Lasater, H. A., 189, 200 Rao, J. N. K., 351
Leone, F. C., 396, 402 Reece, H. W., 221
Lieberman, G. J., 190, 201, 219, 272, Rhodes, R. C., 188, 199
390, 401 Roberts, S. W., 187, 189, 197, 199
Littauer, S, B., 272 Roberts, T. W., Jr., 190, 201, 221
Lloyd, B. H., 196 Robertson, W. H., 271
Lobsinger, D. L., 219 Rogers, W. T., 221
Romig, H. G., 2, 271, 275, 322, 341, 350

Mace, A. E., 196

Madison, R. L., 189, 200 Salvosa, L. R., 450, 463
Maggass, R., 217 Sanoh, A., 188, 198
Mandel, B. J., 186, 197 Satterthwaite, F. E., 188, 198
Mandel, J., 219 Savage, I. R., 188, 199
Hanue1e, J,, 197 Saywell, G. A. T., 221
McCoun, V. E., 96, 219 Schilling, E. G., 402
McElrath, G. W., 220 Schin, R., 221
McLean, R. A., 220 Schrock, E. M., 221
Meyer, T. R., 220 Seder, L. A., 188, 199, 221
Miller, R., 219 Seefeldt, W. B., 196, 201
AUTHOR INDEX 517

Shahani, A. K., 188, 189, 198, 200 Vils, G. R., 222

Shewhart, W. A., 2, 29, 221 Vorhees, H. A., 222
Sittig, J., 222
Small, B. B., 188, 199, 222
Smallwood, H. M., 222 Wald, A., 353, 379
Smith, H., Jr., 222 Wallis, W. A., 350
Smith, K. M., 222 Walter, J. T., 223
Solomon, H., 390, 401 Ward, R. V., 223
Spencer, E. H., 220 Warren, J. C., 215
Spencer, W. M., 217 Way, C. B., 223
Springer, C. H., 188, 199 Weaver, W. R., 61, 96, 196, 223
Stephens, K. S., 315, 322, 402 Weiler, H., 188, 198
Stevens, W. L., 196 Weintraub, S., 271
Stuart, A., 463 Wernimont, G., 223
Subrahmaniarn, J., 351 Westman, A. E. R., 196
Suga, R., 322 Wies, H. M., 380
Williams, C., 218
Williams, T. J., 190, 201
Thompson, H. E., 222 Working, H., 40, 72, 96, 272
Thompson, W. W., Jr., 222 Wortham, A. W., 187, 197
Tippett, L. H. C., 105, 106, 186, 197
Torrey, M. N., 387, 401
Traver, R. W., 188, 199 Yamamoto, T., 322
Traylor, W. S., 222 Yang, C. H., 188, 199
Trelfa, R. T., 215 Yokoh, T., 322
Turner, T. F., 187, 197

VerBeke, G., 222 Zambone, J. H., 220

Verigan, J. F., 222 Zimmer, W. J., 351
 SUBJECT INDEX

average range, 185, 186

Acceptable quality level, 254, 256, 273 defectives, 108
Acceptance sampling (see also Sampling) analysis of past data, 113-122
attributes, 225 --- ---- standard given, 112, 113
by variables, 331-336 defects, 128-133
criteria, 228, 229 analysis of past data, 130-132
basics, 260 standard given, 129, 130
characteristics of, 226-228 defects per unit, 133-138
misconceptions, 265, 267, 268 efficiency of, 180-185
variables fraction defective, 109-122
mean, 325-336 analysis of past data, 113-122
non-normal population, 328, 344 standard given, 109-112
variability, 336-341 gradual process changes, 161-168
Assumption, 18 group, 179
Average outgoing quality curve, 239, increased severity, 185
248, 273, 383 individuals, 157-161
Average outgoing quality limit, 240, measurements, 39-89
273, 275, 383 analysis of past data, 49-53
Average run length standards given, 39, 45-49
control charts, 181-184, 398 narrow limit gaging, 185
cusum charts, 397-399 per cent defective, 113, 127
Average sample number curve, 243, 244, purposes, 32
248, 249, 252, 273, 361, 362 analysis of past data, 32
Average total inspection curve, 244, standards given, 32, 33
245, 249, 250, 252, 273, 275 standard deviation large samples, 168,
169
stratification, 179
Binomial distribution (~ Distribution, trends, 161-168
binomial) tool wear, 161-168
variability, comparison R and s, 53, 54
varying sample sizes
Capability, process, 84 attributes, 122-125, 143, 144
Causes measurements, 169-172
assignable, 27, 28, 173, 174 with acceptance sampling, 259
types of, 78-83 Control chart constants, 98-101
chance, 27, 28, 173, 174 Control limits, revision of, 60, 61
random, 27, 28, 173, 174 Control lines
Chain sampling plan, 390-393 slanting, 164- 166
Coding, 41, 440 Costs of bad quality, 203
Combinations, 10 Curtailed inspection, 250-252
Combinations, statistics of, 403-430 Curve shape, changes in, 80
(see also Tolerances for mating parts) Cusum charts, 395
products, quotients, 422
Contact, 439
Continuous production plans, 381-390 Data
Control charts, 29-34, 39-89, 107-190 attribute, 5
artificial specifications, 183-5 discrete, 9
&verage large samples, 168, 169 measurement, 5

5H
520 SUBJECT INDEX

Defect, 9, 107 Histogram, 6

Defective, 9, 107 Hypergeometric distribution, 117
Defects per 100 units, 227 Hypothesis
Degrees of freedom, 8 alternate, 19
Distribution composite, 19
binomial, 14, 109 null, 18
conditions for, 14 simple, 19
moments, 14 testing, 18
chi-square, 17
discrete uniform, 446, 447
exponential, 447-449 Indifference quality, 254, 318
gamma, 449-451
general system, a, 451
geometric, 442-444 Kurtosis, 439
hypergeometric, 234-238
moments of a, 437
negative binomial, 444-446 Limits, modified, 166-168
normal, 12 Lot tolerance per cent defective,
Poisson, 15, 129, 130, 141 254, 273
conditions for, 15, 129
moments, 15 Mean
rectangular, 414 population, 7
sample mean, 13 sample, 8
sample statistics, 12 Mean, sample
sample variance, 13 moments of, 13
t' 17 Military standard sampling plan
Dodge-Romig Sampling inspection tables, attributes, 281-315
275-281 finding a plan, 309
finding a plan, 278 variables, 341-344
Military standard
1050, 281-315
Error 414, 341-344
of first kind, 19 Modified limits, 166-168
of second kind, 20 Moments
Estimate, 6 binomial, 442
Estimation, 16 discrete uniform, 446
interval, 16 exponential, 448
point, 16 gamma, 449
Event, 10 geometric, 444
independent, 11 hypergeometric, 442
Experiments, sampling, 41 negative binomial, 445
normal, 441
population, 437-442
Fraction defective, 9 Poisson, 441
population, 9 rectangular, 441
sample, 9 sample, 439
Function Moving ranges, 157-161
density, 12

Nominal, 83
Goodness of fit test, 454-456
examples, 456
SUBJECT INDEX 521

Operating characteristic curve, 20, 227, multilevel, 389

230-234, 253-255, 273, 307, 311, 340, double, 229, 245
460, 361 for defects, 225
binomial approximation, 237 mixed product, 174
Poisson approximation, 232, 237 multiple, 313
type A, 230, 239 calculations for, 314
type B, 230-234, 245-247 rational, 173
sequential, 353-380
single, 228, 230-234, 239-245
Parameter, 5 skip-lot, 392-395
Pareto law, 203 stratified, 174
Per cent defective, 227 types of, 172-180
Per cent defective, not p data, 125, 12( Sampling distributions
Poisson distribution (~ Distribution, empirical derivation, 97
Poisson) theoretical derivation, 101-104
Population, 5 Sequential analysis, 353-380
binomial, 6 probability ratio test, 355
normal, 13 process capability, 371
Probability, 10 significance of differences, 372
conditional, 10 special cases, 357-363
cumulative, 13 Shift
laws of, 11 cyclic, 81
Probability of acceptance, 227 steady, 81
Probability density function, 12 Skewness, 439
Process capability, che.ck of Skip-lot sampling, 392-395
sequential, 371 Slanting control lines, 164-166
single sample, 336-340, 371 Specification limits, 54
Process in control, 33 Specifications, comparison with, 54-60
advantages, 34, 35 Standard deviation, 8
Process level, check of long term, 161
single sample, 345, 346, 369-371 population, 8
sample, 8
short term, 161, 346
Ranges Standard values, setting of, 83, 84, 110
moving, 157-161 Statistic, 6
Rating Product Quality, 140 sample, 19
Statistics of combinations
Region
critical, 19 delta technique, 426-429
rejection, 19 Stochastic, 12
Rejectable quality level, 257
Risk Testing hypotheses, 18, 325-341
alpha, 19, 86, 140, 227, 273, 325, mean
354, 396 sigma known, 325-332
beta, 20, 86, 127, 273, 325, 354, 396 sigma unknown, 332-336
consumer's, 227, 254, 255, 257, 273 variability, 336-340
producer's, 227, 254, 255, 257, 273 Tolerance, 56, 162
Runs, 79, 85, 86 Tolerances for mating parts, 403
assembly as system, 417-422
examples, 408-414
Sample, 5 principles, 404-406
Sample space, 9 theory, 406-408
Sampling (~ also Acceptance Sampling) Truncation of inspection, 250-252
continuous production, 381-390 semi-truncation, 251, 252
522 SUBJECT INDEX

population, 8, 437
Variable sample, 8, 437
standardized, 12 Variance, sample
Variance moments of, 13