A Contingency

Table Approach
to Nonparametric
Testing

J.C.W. Rayner
D.J. Best

CHAPMAN & HALL/CRC

Boca Raton London New York Washington, D.C.

©2001 CRC Press LLC

 Library of Congress Cataloging-in-Publication Data

Rayner, J.C.W.
A contingency table approach to nonparametric testing 1 J.C.W. Rayner, D.J. Best.
p. cm.
Includes bibliographical references and index.
ISBN 1-58488-161-5 (alk. paper)
I . Nonparametric statistics. 2. Contingency tables. I. Best, D.J., 11. Title.
QA278.8 .R39 2000
5 19.5-dc21 00-050443

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

Neither this book nor any part may be reproduced or transmitted in any form or by any means, electronic
or mechanical, including photocopying, microfilming, and recording, or by any information storage or
retrieval system, without prior permission in writing from the publisher.

The consent of CRC Press LLC does not extend to copying for general distribution, for promotion, for
creating new works, or for resale. Specific permission must be obtained in writing from CRC Press LLC
for such copying.
Direct all inquiries to CRC Press LLC, 2000 N.W. Corporate Blvd., Boca Raton, Florida 33431.
Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are
used only for identification and explanation, without intent to infringe.

Visit the CRC Press Web site at www.crcpress.com

O 2001 by Chapman & HalllCRC

No claim to original U.S. Government works

International Standard Book Number 1-58488-161-5
Library of Congress Card Number 00-050443
Printed in the United States of America 2 3 4 5 6 7 8 9 0
Printed on acid-free paper
Contents

Preface
1 Introduction
1.1 Parametric or Nonparametric?
1.2 Instructors Example
1.3 Quadratic Differences and Ranking
1.4 Outline and Scope
1.5 Applications of Nonparametric Methods to
Sensory Evaluation

2 Modelling Ties
2.1 Introduction
2.2 The Sign Test and Ties
2.3 Modelling Partitioned Ties in the Sign Test
2.4 Modelling Unpartitioned Ties in the Sign Test
2.5 McNemar’s Test
2.6 Partitioning into Components
2.7 Ties in a Multinomial Test
2.8 Ties When Testing for Independence

3 Tests on One-Way Layout Data: Extensions to the Median

and Kruskal-Wallis Tests
3.1 Introduction
3.2 A Model and Pearson’s X2 Test
3.3 Partitioning Pearson’s Statistic
3.4 The Kruskal-Wallis Test with No Ties
3.5 The Kruskal-Wallis Test with Ties
3.6 Generalised Median Tests

4 Tests Based on a Product Multinomial Model: Yates’ Test

and its Extensions
4.1 Introduction
4.2 One-Way Tables
4.3 Two-Way Tables
4.4 Partitioning X 2P Using Score Statistics
4.5 Other Methods for Ordered Data
4.6 Small Sample Size and Power Comparisons
4.7 Examples

©2001 CRC Press LLC

5 Further Tests Based on a Product Multinomial Model:
Order in the Sign Test and Ordinal Categorical
Data with a Factorial Response
5.1 Introduction
5.2 How Order Affects the Sign Test
5.3 The Sign Test and Gart’s Tests
5.4 A New Model and Score Test
5.5 Comparison of the Sign and Score Tests
5.6 Sports Drink Example
5.7 Recommendations
5.8 Nonparametric Analysis of Ordinal Categorical
Data with Factorial Response
5.9 Olives Data Example
5.10 Cross Cultural Study Example

6 Tests on Complete Randomised Blocks:

Extensions to the Friedman and Cochran Tests
6.1 Peach Example
6.2 Friedman’s Test and its Extensions
6.3 Derivations
6.4 Page’s Test and its Relationship to Friedman’s,
Anderson’s, and Pearson’s Tests
6.5 An Alternative Partition of the Anderson Statistic:
an Umbrella Test
6.6 Ties
6.7 Cochran’s Test
6.8 Stuart’s Test and its Extensions

7 Further Tests on Randomised Blocks:

Extensions to Durbin’s Test
7.1 Introduction
7.2 Durbin’s Test and its Extensions
7.3 Derivations
7.4 A Page-Type Test
7.5 Paired Comparisons with a 2n Factorial Structure

8 Extensions to a Nonparametric Correlation Test:

Spearman’s Test
8.1 Introduction
8.2 A Smooth Model and Tests for Independence
8.3 Smooth Extensions
8.4 Interpretation of the Components
8.5 Discussion
8.6 Multi-Way Tables

©2001 CRC Press LLC

9 One and S-Sample Smooth Tests of Goodness of Fit
9.1 Introduction
9.2 One-Sample Testing for Uncategorised Distributions
9.3 One-Sample Testing for Categorised Distributions
9.4 S-Sample Testing
9.5 Derivations and Simulation Study

10 Conclusion
10.1 Introduction
10.2 Partitioning Pearson’s Chi-Squared Statistic for at
Least Partially Ordered Three-Way Contingency Tables
10.3 Generalised Correlations
10.4 Partially Parametric Testing
10.5 Concluding Comments

Appendices
A.1 Statistical Prerequisites
A.2 Orthogonal Matrices
A.3 Orthonormal Functions
A.4 Direct Sums and Products
A.5 Likelihood Ratio, Score, and Wald Tests
A.6 Assessing Univariate Normality
A.7 Multivariate Normality
A.8 Confidence Circles and Correspondence Analysis Plots
A.9 Permutation and Bootstrap Methods

References

©2001 CRC Press LLC

Preface

The prime objective of this book is to provide a unification a n d

extension of some popular nonparametric statistical tests by linking
them to tests based on models for data that can be presented in
contingency tables.
We asked several colleagues, “What, in your experience, are the
most frequently used nonparametric techniques?” The answers
focused on the sign, median, Wilcoxon, Kruskal-Wallis, Page, Friedman,
Durbin, Cochran, Spearman and Kendall’s tau tests. These tests all use
ranks as data. The treatment of ties is a major concern.
For almost all of these tests, and for some other important tests
also, we are able to present the data in contingency tables. We can
then calculate a Pearson-type X2 statistic, and its components. In the
case of univariate data, the initial tests based on these components
detect mean differences between treatments, and in the case of
bivariate data, they detect correlations. The later components provide
tests that detect variance, skewness and higher moment differences
between treatments with univariate data, and higher bivariate moment
differences with bivariate data. This approach provides a unification
of much popular nonparametric statistical inference, and makes the
traditional, most commonly performed nonparametric analyses much
more complete and informative. In addition, the contingency-table
approach means tied data are easily handled, and almost exact Monte
Carlo p-values can be obtained. Modern statistical packages such as
StatXact (1995) calculate p-values this way.
Sprent (1993, section 9.3) gives a description of the StatXact
(1995) approach and also notes that many of the common
nonparametric location tests can be given via a contingency table
formulation. Neither StatXact (1995) nor Sprent (1993) looks at
extensions based on higher order moments or at the extensions
involving three-way contingency tables that we give. We also do not
consider all marginal totals to be fixed as they do. Further, we consider
tests for data from incomplete block designs, which neither StatXact
(1995) nor Sprent (1993) considers.
Categorical data as well as ranked data can be presented in
contingency tables, and we also look at nonparametric tests for certain

©2001 CRC Press LLC

categorical data. Moreover ranks are but one choice of category
scores. Our formulation permits other user-defined scores.
The methods advanced in this book are somewhat traditional in
that practitioners trained 20 or more years ago will be comfortable with
them. The power of modern computers is used to obtain accurate p-
values using resampling and Monte Carlo methods that at least
augment and sometimes replace p-values based on asymptotic
distribution theory. When sample sizes are small, or the approach to
the asymptotic limit is slow, the computer intensive methods are more
reliable and meaningful than the older approaches. So the approach we
recommend, while familiar, is more complete and more valid than was
previously possible. However we have not considered modern
methods in robustness and Bayesian analysis.
Some readers who are familiar with our published goodness of
fit work may be surprised by this excursion into nonparametrics. We
see it as totally consistent with our view of data analysis. If a
parametric analysis is available and valid it will often, but not always, be
more powerful and more efficient than the competitor nonparametric
test. If so it should be used. Assessing validity will usually require,
amongst other things, an assessment of the distributional assumptions.
Smooth tests of goodness of fit, discussed in Rayner and Best (1989a),
may be helpful in that assessment. If the parametric analysis is not
valid, then nonparametric procedures, such as outlined here, may be
appropriate.
Finally our thanks to
• Geoff Robinson and those other reviewers whose names we do not
know, for all your constructive and insightful suggestions;
• James Chipperfield, Robert Denham, Don Findley, Jo Lea, Nell
Stetner-Houweling, and Chris Valentine, students who worked
through some of the material as part of their honours year studies at
the University of Wollongong;
• Eric Beh who extended some of our work as part of his Ph.D.
studies;
• Anthony Carolan, who shed light on several matters during his
Ph.D. studies;
• Pam Davy, who assisted both Jo and Eric; and
• Chin Diew Lai and Geoff Jones, colleagues from Massey University
in Palmerston North, New Zealand, for interesting discussions on
various aspects of this and other work.
The book has benefited from their positive and insightful advice.

J.C.W. Rayner
D.J. Best

©2001 CRC Press LLC

1
Introduction

1.1 Parametric or Nonparametric?

Parametric tests are based on parametric models. One such test is the
traditional t-test for assessing if the mean of a population, assumed to
come from a normal distribution, takes some specified value. A
nonparametric alternative that has been traditionally employed in this
situation is the Wilcoxon test. This book is about tests that have
traditionally been labelled nonparametric. We avoid a formal definition of
nonparametric, but such a definition wouldn’t necessarily be helpful.
Perhaps the essential point is that nonparametric tests are based on fewer
assumptions than parametric tests. For an enlightening discussion see, for
example, Sprent (1993, section 1.1). Our problem is that we can present
some nonparametric tests in a way that is inconsistent with some of the
favoured definitions of nonparametric. There is no ambiguity in saying
this book extends and unifies some of the tests at the core of
nonparametrics as it has been known and developed over the 20t h
century, including the sign, Mann-Whitney/Wilcoxon, Kruskal-Wallis,
Page, Friedman, Durbin, Cochran, Stuart, and Spearman tests.
Sometimes the data analyst has no choice but to employ
nonparametric methods; for example when the data are available only on
the nominal or ordinal scales of measurement. (See A.1.6 for brief
descriptions of nominal, ordinal, interval, and ratio scales.) Often there is
a choice, as with, for example, in the application of either the Kruskal-
Wallis or the one factor analysis of variance ANOVA F test. As is well
known, under certain assumptions the ANOVA F test will be more
powerful. However the Kruskal-Wallis test is available when these
assumptions do not hold. Some statisticians retreat to nonparametric
methods when there is the slightest doubt about the parametric

©2001 CRC Press LLC

 assumptions. Others rarely use nonparametric procedures. It is known,
for example, that the ANOVA F test is robust to some of its assumptions,
and this vague information is often taken as licence to apply the test when
the assumptions are so obviously false that the analysis can have little
validity. Like most statisticians, our view is between these extremes.
There is often a considerable parametric structure that can be used
to one’s advantage even when the usual parametric model clearly fails (see
section 10.4 on Partially Parametric Testing). The idea is that normality
may fail, but an alternative parametric model is appropriate, and this
alternative model is the basis for a far more powerful analysis of the data
than either the original parametric analysis or a nonparametric analysis.
The alternative parametric analysis described in section 10.4 tends to be
highly computer intensive, and its development is as yet far from
complete.
Parametric model failure produces several problems, not all of which
can be remedied by nonparametric analysis. Nonparametric analysis can
assist if the originally proposed model is not, in fact, the correct one. We
note that:
• the p-value or significance level based on the parametric model may be
quite different under the correct model;
• a test that has high power under the parametric model may be quite
poor under the correct model.
Moreover the proposed parametric model may be an important part in the
analysis of the data. For example, in testing if lifetimes have a given mean,
the assumption of exponentiality is an important part of the modelling.
Failure of the exponential model may be more important than what the
mean is. Retreating to a nonparametric analysis without careful
consideration of the underlying model could be at best unfortunate.
The nonparametric tests of initial interest in this book detect, for
univariate data, mean differences between treatments, or, in the case of
bivariate data, correlations. Our approach produces additional
component-based tests that detect variance, skewness and higher moment
differences between treatments with univariate data, and higher bivariate
moment differences with bivariate data. These additional tests make the
traditional, most commonly performed nonparametric analyses much more
informative. Our additional tests may all be derived by presenting the

©2001 CRC Press LLC

data in contingency tables; they are all components of Pearson’s chi-squared
test statistic, X 2P (see A.1.3). This provides the rationale for the title of this
book.
As we present the data in contingency tables, it is reasonable to ask
why we do not advocate using tests based on log-linear models. As with
the ANOVA test above, these models make strong assumptions and so do
not necessarily apply as widely as nonparametric tests. Further, log-linear
models need iterative methods which can present numerical problems if
almost exact p-values are required. This is especially so with sparse data
sets. It is nevertheless true that log-linear models are currently in
widespread use; so in several cases we provide comparisons of our
approach with that based on log-linear models. In all cases the p-values
produced are very similar.
Other important features of our methods are that they easily cope
with tied data and that exact or almost exact p-values can easily be
obtained. Our nonparametric analyses also give least significant difference
(LSD) assessments, contrasts and graphical presentations to assist the
statistician to explain the analysis to those with little statistical background.

1.2 Instructors Example

We will now consider an example which illustrates the nature of our
refinements to standard nonparametric tests. Our interest in this data set,
and others we give, is more to illustrate points concerning the statistical
methodology than to assess the application. In statistical practice, of
course, the reverse is usually the case.

Instructors Example. Conover (1998, p.293) gave the following data of

three instructors who assign grades in five categories according to Table
1.1. We assign the scores 1 to 5 to the categories, with A scored as 1, B as
2, and so on.

©2001 CRC Press LLC

Table 1.1 Grades assigned by three instructors

Grade
A B C D E Total

Instructor 1 4 14 17 6 2 43
Instructor 2 10 6 9 7 6 38
Instructor 3 6 7 8 6 1 28

Total 20 27 34 19 9 109

Assume that a standard computer package is available, and is used

to explore these data. In testing for normality, the Kolmogorov-Smirnov
normality test yields a p-value of greater than 0.20 for each of the three
instructors, while the Shapiro-Wilk normality test yields p-values of at least
0.10 for all three distributions. These tests cannot reject the null
hypothesis of normality, perhaps in part because of the small sample sizes.

©2001 CRC Press LLC

 Figure 1.1: Instructors' Data
5.25

4.5

3.75
Score

3.0

2.25

1.5

0.75

1 2 3

Instructors

A one-way analysis of variance is in any case robust to normality

and so may be confidently applied, yielding a p-value of 0.78. No
assessment has yet been made of homogeneity of variance. In fact
inspection of the data, via box plots, suggests that the variability differs
between the instructors (see Figure 1.1). The failure of this assumption
decreases our confidence in the p-value from the analysis of variance.
The Kruskal-Wallis test adjusted for ties has a p-value of 0.85. The
less powerful median test has p-value 0.77, while the usual Pearson’s X 2P
test for homogeneity of rows has p-value 0.15, all consistent with the
conclusion that there seems to be no difference between the instructors in

©2001 CRC Press LLC

regard to location differences.
The techniques we develop in this book enable us to decompose
Pearson’s X 2P into a location detecting component with p-value 0.85, a
dispersion detecting component with p-value 0.01, and a residual with p-
value 0.75. An LSD analysis of the dispersion detecting component shows
that the first instructor is less variable than the third who is less variable
than the second. The nonsignificant location component confirms the
analyses of the location tests, while the nonsignificant residual indicates
that there are no further effects in the data. This rather complete analysis
makes no assumptions. Conover’s analysis, like the analysis of variance
and Kruskal-Wallis tests, did not identify the dispersion effect. Our
analysis is very effectively presented graphically in what we call a product
map. For an introduction to product maps see Appendix A.8. One axis
gives information about linear effects, that we interpret as location
differences between treatments, while a second gives information about
quadratic effects, that we interpret as dispersion differences (see Figure
1.2). This example will be revisited later at the end of section 3.3.

©2001 CRC Press LLC

 Figure 1.2: Instructors' Map
4.3

2
2.15
Quadratic Effect

0.0 3

1
-2.15

-4.3

-4.3 -2.15 0.0 2.15 4.3

Linear Effect

1.3 Quadratic Differences and Ranking

The previous section gave an example in which a linear or location effect
was not significant but a quadratic or dispersion effect was. Differences in
quadratic effects are often associated with scale or dispersion effects. If
the linear effect had been significant this may have masked the detection
of the quadratic effect. To help see this consider the following data sets A
and B:

©2001 CRC Press LLC

data set A: -2.019, -0.802, -0.153, -1.151, -1.247, 0.293, 1.848, 1.430, 1.875,
-0.192,
data set B: 12.838, 29.986, 16.284, 18.131, 23.386, 14.810, 41.251, 13.601,
10.940, 20.448.

Data set A consists of 10 random N(0, 1) values, while data set B consists
of 10 random N(20, 10) values. Suppose we rank the data from smallest
to largest.

Data Value -2. 019 ... 1.875 10.940 ... 41.251

Rank 1 ... 10 11 ... 20

A 1 ... 1 0 ... 0
B 0 ... 0 1 ... 1

The A’s occupy the first 10 ranks and B’s the last 10. If the analysis of the
previous section is used, the linear effect is very highly significant (p-value
< 0.001) but the quadratic component is zero (p-value one). Clearly the
ranking transformation has destroyed the quadratic differences between
A and B. To compensate for this problem with ranking, a common
procedure is to align the data sets by subtracting the A mean from each A
value, and the B mean from each B value. See, for example, Sprent (1993,
p. 124).
However when the raw data are available we suggest using these
raw data as scores rather than using the ranks. If this is done here the
analysis in Table 1.2 is obtained. Details will be given later. The quadratic
effect is now significant at the 5% level, although a smaller p-value may
have been obtained with an aligned ranks scale test.
In this text we will use the partition of X 2P method with raw data as
scores in preference to alignment or standardisation. We discuss this
point again in the Boos example at the conclusion of section 3.4. The
essential point here is that each choice of scores gives a different test, and
every conclusion is conditional on the test chosen. Using ranks, natural
scores, and the data as scores all may or may not result in similar

©2001 CRC Press LLC

conclusions. The user should be aware that each choice casts the data in
possibly different lights. Examples of the use of different scores with the
same data set are given at the end of section 4.7.

Table 1.2 Partition of X 2P for random normal data

Statistic Degrees of Freedom Value Monte Carlo

p-value

Linear 1 14.40 < 0.001

Quadratic 1 4.41 0.036
Residual 17 1.19 -

Total (X 2P) 19 20.000 -

1.4 Outline and Scope

For almost all of the tests considered in this book, and for some other
important tests also, we are able to present the data in contingency tables.
This provides a unification of these nonparametric tests. Order is usually
important for at least one of the categorisations. For the sign test and
one-sample goodness of fit tests for categorised data, one-way
contingency tables suffice. Two-way contingency tables with fixed
marginal totals are appropriate for the median and Kruskal-Wallis tests.
Data for the relatively little known Yates’ test can be presented in two-
way contingency tables with one set of marginal totals fixed and one set
random. Special two-way contingency tables that we call Anderson tables
are appropriate for the Friedman and Durbin tests. If no decisions about
ties and the category totals are made before collecting the data, then data
for Spearman’s test may be presented in two-way contingency tables with
no marginal totals fixed. Marginal distributions of categorical data from
matched pairs or from blocks can be given in contingency tables and
homogeneity tested using statistics due to Cochran and Stuart, and

©2001 CRC Press LLC

extensions of these.
Chapter 2 focuses on ties: how they have been traditionally treated
and how different models lead to different treatments. The context is
mainly what is for many the first nonparametric test, the sign test. There
is also an introduction to the idea of decomposing a test statistic into its
components to give a more detailed scrutiny of the data.
Chapters 3 to 8 focus on tests for which the data may be given in
two-way contingency tables. The standard tests presented assess mean
or correlation departures from the null. Since the traditional correlation is
the (1, 1)th bivariate moment, all are essentially first order moment tests.
We extend all these tests to detect higher moment departures from the
null. This gives a detailed and informative scrutiny of the data.
Chapter 9 concerns one- and S-sample smooth tests of goodness of
fit. One-sample goodness of fit was covered in detail in Rayner and Best
(1989a), and this is an overview and update of that material. The S-sample
material is relatively new, having been published only recently.
Chapter 10 includes a discussion of recent work on partially
parametric testing. This has grown out of the work on one sample smooth
tests of goodness of fit. The probability (density) function of the
distribution tested for is nested in a rich family of probability (density)
functions. If, for example, normality is rejected, a more complicated
distribution from this family provides a better fit to the data. It seems
reasonable to base inference on such a distribution. This distribution isn’t
that assumed by the classical parametric methods. However the methods
are not nonparametric: they do assume considerable parametric structure.
Also included in this chapter are some extensions of some current two-
way table work to multiway tables.
The reader interested in methods only will need to refer carefully
to the index, as some traditional topics appear in more than one chapter.
Thus the sign test is treated in the chapter dealing with ties, and also later,
where it is thought that the order in which the treatments are presented
may be important. The appropriate model is then a two-way contingency
table with fixed marginal totals, and so this is presented in a chapter
dealing with the product multinomial model.
Note that while we do not often use traditional parametric models
like the normal, we do use models. Since tables involve count data, we

©2001 CRC Press LLC

extensively use the multinomial, and in our smooth testing for goodness of
fit, quite complicated nested families are employed.
For almost all of the location detecting tests, Page-type tests for
ordered alternatives can be constructed, and we usually do consider this.
For almost all tests considered, normal and other scores may be used.
However, we usually do not consider scores other than natural (the row
or column number), mid-rank, and the raw data. Other options may be
substituted if the user wishes.
The contingency-table approach means tied data are easily handled,
and almost exact Monte Carlo p-values can be obtained. Modern
statistical packages such as StatXact (1995) calculate p-values this way.
For data presented in contingency tables we can calculate Pearson’s
2
X P statistics for testing goodness of fit, independence or homogeneity as
appropriate, and their components. As mentioned above, these
components detect moment departures from the null, and are in many
ways like contrasts used in the analysis of variance. Components are:
• additive, in that they have sum, or sum of squares, the original omnibus
statistic;
• easy to use, as they are asymptotically mutually independent and
asymptotically have the standard normal distribution;
• easy to interpret, detecting such important alternatives as a mean
difference between treatments, or lesser variability between treatments
than expected;
• the basis for either the familiar tests mentioned above, or extensions of
them; these extensions are often new tests, not yet in the statistical
literature; and
• the basis for new informative graphical displays.
We have made a major effort to keep the presentation of the
technical material at an accessible level. Two main approaches are used.
In one a vector of counts is constructed. The Pearson statistic is a
quadratic form in this vector with matrix the inverse of the covariance
matrix. Obtaining the components and the extensions of the traditional
tests requires the diagonalisation of the covariance matrix. The covariance
matrices of interest often involve direct or Kronecker products. These are
described in the Appendix that gives some of the background

©2001 CRC Press LLC

mathematics, statistics and computing used. Apart from direct or
Kronecker sums and products, undergraduate algebra and statistics is
largely sufficient. Multivariate normality underpins this approach, and a
largely theoretical development of this topic is given in the Appendix.
The other approach we use is specifying a model and finding at
least one of the asymptotically optimal tests, usually the score or Wald
test. These are asymptotically equivalent to the likelihood ratio test, and,
again, are described in the Appendix. The advantage of this approach is
the tests so constructed are weakly optimal for specifiable models.
For some of the situations we consider the covariance matrix is
prohibitively difficult to calculate. For others we have been unable to
construct an appropriate model. Readers will usually find the covariance
matrix approach easier to understand but mechanically more cumbersome.
Our first preference usually is to use the asymptotically optimal tests
because of their optimality properties. Every situation is different and
flexibility is often rewarded.
By this point it will be apparent that our focus is on testing rather
than estimation. This is more a matter of how we have chosen to present
our work. Our methods, especially the smooth models, do permit
consideration of estimation questions. However we leave such matters
for others to develop. In many situations this will be routine.
There is an important caveat we should place on our approach, or
more particularly, on how we interpret our components. There is a
growing literature, mainly in the area of goodness of fit, on how to
interpret the components. The evidence is that caution should be
exercised, that a significant first order component may not always indicate
a mean shift. But it is highly suggestive, a reasonable working hypothesis
that perhaps should be examined before any others. We shall discuss this
issue in greater detail later, but we note here that caution in the
interpretation of components should be exercised for all tests we examine
here.
Much of the material in this book has appeared in print in the
statistical literature in the recent past, or is about to. See the references
and in particular the papers of Beh, Best, Carolan, Davy, and Rayner. We
hope that collecting this material here will enable the reader to better

©2001 CRC Press LLC

appreciate the totality of the contingency table approach, and its power,
simplicity, and applicability.

1.5 Applications of Nonparametric Methods to Sensory

Evaluation

Popular texts for nonparametric methods such as Conover (1998),

Hollander and Wolfe (1999), Daniel (1990), Sprent (1993, 1998) and Siegel
and Castellan (1988) give many examples of applications of nonparametric
methods. In this book we make special mention of important applications
to sensory evaluation of foods. One of us (John Best) has been involved
with the analysis of sensory data related to food science for over a quarter
of a century. Many of the examples used in this book reflect that
experience.
Important nonparametric tests used in sensory evaluation are the
sign and triangle tests based on the binomial distribution and the Friedman
ranking test. The triangle test, involving the ability to choose an odd
sample from three, is commonly used in training expert taste-test panels,
while the Friedman test is often used in consumer or market research to
objectively compare competing brands.
We will discuss both the binomial and Friedman tests in greater
detail here than is often done in other texts, and we will also give
extensions of these tests that are important in the sensory evaluation of
food.
Categorical data are often obtained from sensory experiments or
consumer market research. Analysis of variance would usually be
employed for the analysis of such data, and while this may be satisfactory
for seven- or nine-point hedonic or liking scales, its use for three- or five-
point just right or likely to purchase data is more questionable. We will
discuss alternative analyses for such data with very few categories. These
analyses also have the advantage of allowing comparisons of distributions,
not just means, which is all the analysis of variance gives. Comparison of
distributions is important in identifying market segmentation in consumer
research applications.

©2001 CRC Press LLC