Download the full version of the ebook now at ebooksecure.

com

Exact Statistical Inference for Categorical

Data 1st Edition Shan - eBook PDF

https://ebooksecure.com/download/exact-
statistical-inference-for-categorical-data-ebook-
pdf/

Explore and download more ebook at https://ebooksecure.com

 Recommended digital products (PDF, EPUB, MOBI) that
you can download immediately if you are interested.

(eBook PDF) Applied Statistical Inference with MINITAB,

Second Edition 2nd Edition

https://ebooksecure.com/product/ebook-pdf-applied-statistical-
inference-with-minitab-second-edition-2nd-edition/

ebooksecure.com

(eBook PDF) An Introduction to Probability and Statistical

Inference 2nd Edition

https://ebooksecure.com/product/ebook-pdf-an-introduction-to-
probability-and-statistical-inference-2nd-edition/

ebooksecure.com

(eBook PDF) Probability and Statistical Inference 10th

Edition by Robert V. Hogg

https://ebooksecure.com/product/ebook-pdf-probability-and-statistical-
inference-10th-edition-by-robert-v-hogg/

ebooksecure.com

(eBook PDF) Fire and Emergency Services Instructor:

Principles and Practice 3rd Edition

https://ebooksecure.com/product/ebook-pdf-fire-and-emergency-services-
instructor-principles-and-practice-3rd-edition/

ebooksecure.com
(eBook PDF) Supervision Today! 9th Edition by Stephen P.
Robbins

https://ebooksecure.com/product/ebook-pdf-supervision-today-9th-
edition-by-stephen-p-robbins/

ebooksecure.com

Business in Action 8th Edition (eBook PDF)

https://ebooksecure.com/product/business-in-action-8th-edition-ebook-
pdf/

ebooksecure.com

Writing about Writing: A College Reader Third Edition by

Elizabeth Wardle (eBook PDF)

https://ebooksecure.com/product/writing-about-writing-a-college-
reader-third-edition-by-elizabeth-wardle-ebook-pdf/

ebooksecure.com

Human Anatomy & Physiology Laboratory Manual, Main Version

12th Edition by Elaine N. Marieb (eBook PDF)

https://ebooksecure.com/product/human-anatomy-physiology-laboratory-
manual-main-version-12th-edition-by-elaine-n-marieb-ebook-pdf/

ebooksecure.com

Harrison’s Hematology and Oncology 3rd Edition Dan L.

Longo - eBook PDF

https://ebooksecure.com/download/harrisons-hematology-and-oncology-
ebook-pdf/

ebooksecure.com
Fluid Mechanics (9th Edition) Frank M. White - eBook PDF

https://ebooksecure.com/download/fluid-mechanics-9th-edition-ebook-
pdf/

ebooksecure.com
Exact Statistical Inference
for Categorical Data
Exact Statistical Inference
for Categorical Data

Guogen Shan
University of Nevada, Las Vegas, NV, USA

AMSTERDAM • BOSTON • HEIDELBERG • LONDON

NEW YORK • OXFORD • PARIS • SAN DIEGO
SAN FRANCISCO • SINGAPORE • SYDNEY • TOKYO
Academic Press is an imprint of Elsevier
Academic Press is an imprint of Elsevier
125 London Wall, London, EC2Y 5AS, UK
525 B Street, Suite 1800, San Diego, CA 92101-4495, USA
225 Wyman Street, Waltham, MA 02451, USA
The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, UK

Copyright © 2016 Elsevier Ltd. All rights reserved.

No part of this publication may be reproduced or transmitted in any form or by any means, electronic or
mechanical, including photocopying, recording, or any information storage and retrieval system,
without permission in writing from the publisher. Details on how to seek permission, further
information about the Publisher’s permissions policies and our arrangements with organizations such as
the Copyright Clearance Center and the Copyright Licensing Agency, can be found at our website:
www.elsevier.com/permissions.
This book and the individual contributions contained in it are protected under copyright by the
Publisher (other than as may be noted herein).

Notices
Knowledge and best practice in this field are constantly changing. As new research and experience
broaden our understanding, changes in research methods, professional practices, or medical treatment
may become necessary.
Practitioners and researchers must always rely on their own experience and knowledge in evaluating
and using any information, methods, compounds, or experiments described herein. In using such
information or methods they should be mindful of their own safety and the safety of others, including
parties for whom they have a professional responsibility.
To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors, assume
any liability for any injury and/or damage to persons or property as a matter of products liability,
negligence or otherwise, or from any use or operation of any methods, products, instructions, or ideas
contained in the material herein.

ISBN: 978-0-08-100681-8

British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library

Library of Congress Cataloging-in-Publication Data

A catalog record for this book is available from the Library of Congress

For information on all Academic Press publications

visit our website at http://store.elsevier.com/
LIST OF FIGURES

1.1 Tail probability plots for a binomial comparative study based

on three test statistics. ....................................................... 12
1.2 Tail probability plots for a binomial comparative study based
on partial maximization by using the Z-pooled test statistic.
The two green lines are the exact 0.999 confidence interval
for the nuisance parameter, and the purple dot is the value for
the BB p-value. .............................................................. 14
1.3 Two-sided and one-sided type I error rate comparisons among
the five exact approaches with a balanced sample size of
100 per group at the significance level of α = 0.05, based
on the test statistics TPD , TZuP , and TZP from the first row to
the third row. .................................................................. 18
1.4 Two-sided and one-sided type I error rate comparisons among
the five exact approaches with unbalanced sample size,
nt = 100 and nc = 50 at the significance level of α = 0.05,
based on the test statistics TPD , TZuP , and TZP from the first
row to the third row. ......................................................... 21
1.5 Power comparisons for two-sided and one-sided problems
among the five exact approaches with balanced sample size,
nt = 100 and nc = 100 at the significance level of α = 0.05,
under the alternative pc = 0.3 and pt = pc + θ, based on
the test statistics TPD , TZuP , and TZP from the first row to
the third row. .................................................................. 23
1.6 Power comparisons for two-sided and one-sided problems
among the five exact approaches with unbalanced sample size,
nt = 100 and nc = 50 at the significance level of α = 0.05,
under the alternative pc = 0.3 and pt = pc + θ, based on
the test statistics TPD , TZuP , and TZP from the first row to
the third row. .................................................................. 25
2.1 Plots of the tail probability as a function of the nuisance
parameter for the animal carcinogenicity study with K = 4
and a sample size of 10 per group. ....................................... 33

vii
viii List of Figures

2.2 Type I error rate for the five exact approaches when K = 3
and a sample size of 30 per group at the significance level of
α = 0.05, the left plot with the score value d = (0, 1, 2), and
the right plot with the score value d = (0, 1, 4). ...................... 34
2.3 Power comparison among the five exact approaches when
K = 4, a sample size of 30 per group, and the score value
d = (0, 1, 2, 3). ................................................................ 36
2.4 Power comparisons among the C approach, the M approach,
and the C+M approach, using the χ 2 test with total
sample sizes of 25, 50, 100, and 300 from the first row to
the fourth row.................................................................. 40
LIST OF TABLES

1.1 A 2 × 2 Contingency Table ................................................ 2

1.2 A Randomized Placebo-Control Two-Arm Study for Patients
with Chronic Noncancer-Related Pain .................................. 2
1.3 Association Between Smoking and UADT Cancer .................. 4
1.4 Airway Hyper-Responsiveness (AHR) Status Before and After
Stem Cell Transplantation (SCT)......................................... 5
2.1 A 2 × K Contingency Table ............................................... 30
2.2 Data From the Animal Carcinogenicity Study with K = 4 ........ 33

ix
PREFACE

With the development of computational techniques (e.g., super-computers,

parallel computing) and statistical software packages (e.g., SAS, R, Stata,
StatXact, SPSS, Matlab, PASS), exact statistical inference for categorical
data analysis is increasingly available for use in practice. In the cases that
traditional asymptotic approaches do not have satisfactory performance with
regards to type I error control and accurate sample size determination,
exact approaches should be utilized. This book provides an overview of
exact approaches, including Fisher’s exact approach, which is also known
as the exact conditional approach, and several efficient exact unconditional
approaches. Real examples are provided to illustrate the application of these
exact approaches, and these approaches are also comprehensively compared
in many important statistical problems.

The first chapter reviews the three sampling methods to generate data, then
presents the five exact approaches. Data that can be organized in a 2 × 2
contingency table is considered in this chapter. Among the five approaches,
one is the exact conditional approach, and the remaining four are uncondi-
tional. This book is the first to comprehensively compare the performance
of the five exact approaches for data from a binomial comparative study.
Chapter 2 deals with data from a 2 × K table by applying the exact
approaches. Such data is commonly obtained from a dose-response study,
and a genetic study. The last chapter is given to sample size determination
based on exact approaches. Power analysis is an essential part of a research
proposal, and accurate sample size determination would increase the chance
of the proposal being funded and finished in a timely manner.

I thank those who provided valuable comments and helpful suggestions

about the book, including anonymous reviewers for this book. I am grateful
to Prof. Chris Lloyd from Melbourne Business School for developing the
most recent exact approach for categorical data analysis. I also thank Profs.
Daniel Young, Weizhen Wang, Changxing Ma, Gregory Wilding, and Alan
Hutson for their valuable comments. Finally, I would like to thank my wife,
Yanjuan, for her continued support of my academic career, and our parents
for taking care of us and our children.

xi
CHAPTER 1
Exact Statistical Inference for a 2 × 2 Table

A 2 × 2 contingency table commonly arises in a comparative study

(e.g., a study to compare the response rate between two treatments) or
a cross-sectional study (e.g., a study to test the association between two
dichotomous variables). Barnard [1] was the first to describe three distinct
sampling methods to generate a 2 × 2 contingency table; see Table 1.1. They
are often termed as a 2 × 2 independent study, a 2 × 2 comparative study,
and a double dichotomy study with both marginal totals fixed, one marginal
fixed, and no marginal fixed, respectively. In a double dichotomy study
with no marginal fixed, the total sample size of a study, N, is considered
fixed.

Independent Study
In the first sampling method for a 2 × 2 independent study, both marginal
totals of a 2×2 table are considered as fixed, that is, sample sizes (n1 , n2 ) for
the factor A and (m1 , m2 ) for the factor B, are known before the study. One
classical example is the experiment described by Fisher [2]: a lady claimed
that she could tell whether milk was poured before or after tea in a cup. In
this interesting experiment, eight cups were prepared with four of each kind.
The lady was informed that among these eight cups, four were prepared with
tea first and the remaining four with milk first. After the lady tasted all eight
cups, she reported which four cups she thought had milk added first. It is
obvious in this experiment that both marginal totals are fixed, with four for
each kind on both marginal totals from the truth and the lady’s answer. Such
studies are relatively rare in practice due to the fact that subjects in the study
were informed about the number of each kind, and for this reason, a 2 × 2
independent study will not be further discussed in detail here.

Comparative Study
A 2 × 2 comparative study involves two independent groups with sample
sizes n1 and n2 . At the end of the study, the associated number of events
(e.g., response, survival) x1 and x2 are observed from the first group and the
second group, respectively. It is often interesting to compare the response
Exact Statistical Inference for Categorical Data. http://dx.doi.org/10.1016/B978-0-08-100681-8.00001-4
Copyright © 2016 Elsevier Ltd. All rights reserved. 1
2 Exact Statistical Inference for Categorical Data

Table 1.1 A 2 × 2 Contingency

Table
Factor A
Yes No Total
Yes n11 n12 m1
Factor B
No n21 n22 m2
Total n1 n2 N

rate between the two groups. The following Phase III study is used to
illustrate the setting of a 2 × 2 comparative study.

Example 1.1. A randomized, placebo-controlled two-arm Phase III

clinical trial was conducted to evaluate oral lubiprostone for constipation
associated with non-methadone opioids in patients with chronic noncancer-
related pain [3]. Patients were randomized into either the treatment group
treated with lubiprostone or the placebo group, and they were followed for
12 weeks in the study. The data is displayed in Table 1.2. At the end of
the study, x1 = 58 responders out of n1 = 214 were recorded from the
treatment group, while x2 = 41 out of n2 = 217 patients were observed
from the placebo group.

The response rates, the primary endpoint, are estimated to be 27.1%

and 18.9% for the treatment group and the placebo group, respectively.
To compare the response difference between two groups, Pearson’s
χ 2 test statistic
(n11 n22 − n12 n21 )2 N
Tχ 2 =
n 1 n 2 m1 m2
is used for testing the null hypothesis
H0 : p 1 = p 2
against the alternative hypothesis

Table 1.2 A Randomized Placebo-Control

Two-Arm Study for Patients with Chronic
Noncancer-Related Pain
Treatment Group Placebo Group
Yes 58 41
Response
No 156 176
Total 214 217
Source: From Jamal et al. [3], with permission.
 Exact Statistical Inference for a 2 × 2 Table 3

Ha : p1 = p2 .
This test can be found in the function prop.test from statistical software R,
and the freq procedure from SAS to compare two independent proportions.
It should be noted that the χ 2 test statistic is equivalent to the Z test statistic
with a pooled variance estimate, which is given as
p̂1 − p̂2
 ,
p̂(1 − p̂)(1/n1 + 1/n2 )
where p̂1 = x1 /n1 , p̂2 = x2 /n2 , and p̂ = (x1 + x2 )/(n1 + n2 ). It is obvious
that the Z test statistic can be applied to a one-sided problem, but the χ 2 test
statistic Tχ 2 is only used for a two-sided problem.

The asymptotic limiting distributions of the χ 2 test and the Z test are
often used for statistical inference, and they are appropriate for use in
practice only when cell frequencies are large enough. The χ 2 test is not
recommended for use when the lowest expected frequencies from the four
cells is less than 5 [4, 5]. However, Cochran [6] argued that the cut point
value 5 is chosen arbitrarily, and this cut point may be modified when new
evidence from data becomes available. For data with small cell frequencies,
exact approaches (e.g., Fisher’s exact conditional approach) are generally
recommended [2, 7, 8]. Several exact approaches [2, 8–15] will be discussed
later in Section 1.1.

Double Dichotomy Study

In a double dichotomy study, only the total sum is fixed, which is common in
a cross-sectional study for testing an association between two dichotomous
variables. A sample of size N is drawn from a population, and each member
of the sample is classified according to the two dichotomous variables, A
and B. For such studies, the row and column totals are not fixed in advance;
only the total sum is fixed. One typical example is a cross-sectional study to
test the association between smoking and cancer.

Example 1.2. Krishnatreya et al. [16] reported a retrospective study

of upper aero digestive tract (UADT) cancer patients from 2010 to 2011
from the hospital cancer registry. For the N = 56 patients documented
with the occurrence or presence of synchronous primaries, each patient was
asked about his/her smoking status, and was tested whether or not UADT
appears at both index and synchronous. Data from this study is presented
in Table 1.3. One of the main research questions from this study was to
test the association between the smoking history and the presence of UADT
synchronous cancers.
Visit https://testbankfan.com
now to explore a rich
collection of testbank or
solution manual and enjoy
exciting offers!
4 Exact Statistical Inference for Categorical Data

Table 1.3 Association Between

Smoking and UADT Cancer
UADT tumors
Yes No
Yes 45 1
Smoking history
No 5 5
Source: From Krishnatreya et al. [16], with
permission.

The Pearson χ 2 test was used for testing the association, and the p-value
was found to be much less than 0.05. Then, the authors concluded that
smoking was significantly associated with the occurrence of synchronous
primaries in UADT. They also mentioned that the Yates’ correction was
used in the χ 2 test statistic as small expected frequencies were observed
from the table.

In addition to the commonly used Pearson χ 2 test statistic, the likelihood

ratio χ 2 test, often referred to as the G test, is an alternative to test the
hypothesis of independence. Based on the standard maximum likelihood
method, the likelihood ratio χ 2 test will be close in results to the Pearson’s
χ 2 test for large samples. The likelihood ratio χ 2 test statistic has the form as


2 
2  
nij N
TLR = 2 nij log ,
mi n j
i=1 j=1

where 0 log 0 = 0. Although these asymptotic tests perform well in the

presence of large sample sizes, they could perform poorly in a small sample
setting [17].

In addition to the cross-sectional study, a matched-pairs design is another

important application of a double dichotomy study. With the total N subjects
enrolled in a study, each subject is measured twice, before and after an
intervention.

Example 1.3. Bentur et al. [18, p. 847] conducted a study on airway

hyper-responsiveness (AHR) status before and after stem cell transplan-
tation (SCT) on 21 patients; see Table 1.4 for the data. The AHR status
for each patient is assessed using a methacholine challenge test (MCT)
twice, before and after SCT. A positive MCT (AHR yes) is defined by
PC20 < 8 mg/ml.
 Exact Statistical Inference for a 2 × 2 Table 5

Table 1.4 Airway

Hyper-Responsiveness (AHR) Status
Before and After Stem Cell
Transplantation (SCT)
Before SCT
AHR yes AHR no
AHR yes 1 7
After SCT
AHR no 1 12
Source: From Bentur et al. [18], with permission.

In addition to a matched-pairs study where each subject is measured

twice, it could be a study in which each subject is matched with an equiva-
lent from another study. In practice, data from another experiment is already
exist or easy to obtain. Such designs can be used to reduce the influence of
possible confounding factors. Traditionally, the χ 2 test and the likelihood
ratio test are used for testing the association between two dichotomous
variables.

Let pij = nij /N be probability for the i-th level of the factor A and
j-th level of the factor B. Suppose p1 = p11 + p21 and p2 = p11 + p12 are
the marginal probabilities. The difference between these two proportions is
often the parameter of interest, p1 − p2 , or equivalently p21 − p12 . To make a
statistical inference for this parameter, the most commonly used test statistic
is the McNemar test [19]:
(n21 − n12 )2
TMC = .
n21 + n12
It should be noted that only the off-diagonal numbers, n12 and n21 , from a
2 × 2 table are used in the test statistic, and the diagonal values, n11 and n22 ,
have no influence on computing the test statistic and the p-value calculation.
There has been a long-term debate over whether all values should be used
in the test statistic.

1.1 EXACT TESTING PROCEDURES

When sample size in a study is increased from small to large, asymptotic
approaches are traditionally used for data analysis. However, the signifi-
cance value they provide is only an approximation, because the sampling
distribution of the test statistic is only approximately equal to the theoretical
6 Exact Statistical Inference for Categorical Data

limiting distribution, for example, a χ 2 distribution, a standard normal

distribution. The approximation is inadequate in cases where the total
sample size is small, or the expected values for cells in the table are low.

In discrete data analysis, unsatisfied type I error control from traditionally

used asymptotic approaches has been observed in many statistical problems.
In a comparative binomial study, Pearson’s χ 2 test is often associated
with an inflated type I error rate, while the χ 2 test based on Yates’
correction [20] is always conservative, with actual type I error rate being
less than the nominal level, and often less than half of the nominal level
[7, 11, 21–23]. Uncontrolled type I error rate in a study could lead to either
under- or overestimated sample size calculation. Several modified χ 2 test
statistics were proposed to increase the performance of the Pearson’s χ 2
test, for example, the uncorrected χ 2 test [24] and re-corrected χ 2 test [25].
Uncontrolled type I error occurs not only in a 2 × 2 table, but also in other
types of data. For example, a dose-response study to test a trend for data in
a 2 × K table, the traditionally used test statistic, the Cochran-Armitage test
[4, 26] always has an inflated type I error rate as the total sample size goes
to infinity [27].

In light of the problems of type I error control, procedures based on exact

probability calculations may be considered in order to preserve the nominal
level of a test. Two basic exact approaches, the conditional approach and
the unconditional approach, will be introduced first, followed by another
three exact unconditional approaches. To avoid too many mathematical
notations and symbols, we use a comparative binomial study to explain the
computation of these five exact approaches.

1.1.1 Conditional Approach

For the cases where asymptotic approaches are not adequate (e.g., the total
sample size is small, the expected sample size for some cells is too small),
exact approaches should be considered to make proper statistical inference.
Fisher [2] was among the first to propose an exact approach by fixing both
marginal totals to control for any nuisance parameter in the tail probability.
This is an exact conditional approach and is referred to be as the C approach.
This approach was originally developed for analyzing a 2 × 2 table, but it
can be applied to a general R × C contingency table. Mehta and Patel [28]
developed a network-based algorithm to implement Fisher’s exact approach
for different types of categorical data. But, the main application of Fisher’s
approach lies in a simple 2 × 2 table.
 Exact Statistical Inference for a 2 × 2 Table 7

Generally speaking, a test statistic should be used in conjunction with

the conditional approach to determine the tail area. A test statistic is
used to order all possible tables in the sample space where they all have
with the same marginal totals as the observed table. As we all know, the
limiting distribution of a test statistic and its property are very important in
asymptotic approaches for p-value calculation and efficiency comparison,
but not in exact approaches. It is only used for the purpose of ordering
all tables from the sample space. The probability of each table can be
calculated exactly from a hyper-geometric distribution whose probability
density function is only based on the values from a table.

Under the conditional framework with both marginal totals fixed, the
value at the (1, 1) cell, n11 , determines the other three values in a 2 × 2
contingency table. Therefore, we use n11 to represent the complete data
(n11 , n12 , n21 , n22 ) for simplicity. The probability of a 2 × 2 table as in Table
1.1 under the conditional approach is computed as
m1 !m2 !n1 !n2 !
PC (n11 ) = . (1.1)
N!n11 !n12 !n21 !n22 !
Then, the p-value is computed by adding the probabilities of the given table
and other more extreme tables. For example, in a one-sided hypothesis
problem that rejects the null hypothesis with a large test statistic, all the
tables with the test statistic values being larger than or equal to that of
the given table’s are in the rejection region and their probabilities are
added together to compute the p-value. Although the assumption for the
limiting distribution of a test statistic is not needed in exact approaches,
some assumptions related to a study itself must be satisfied, for example,
the independence assumption of participants. These assumptions can be
checked from the study.

Fisher’s exact approach has been applied to many statistical problems,

such as an association test between two categorical variables, a trend test
with binary endpoints [14], and proportion comparison for binary clustered
data [29]. While some theorists argued that Fisher’s exact test can only be
applied to a study that was originally designed with both marginal totals
fixed, actually, Fisher’s idea is a general approach, and it has been applied
to many studies whose marginal totals are not fixed. As aforementioned,
a study with both marginal totals fixed is rarely found in practice. One
frequently used area of Fisher’s exact approach is where the traditionally
used χ 2 test can not be applied due to a relatively small expected frequency.
When this assumption of the χ 2 test is not satisfied, the exact conditional
8 Exact Statistical Inference for Categorical Data

approach is the alternative that should be used to guarantee the type I error
rate. The exact conditional approach is widely available in the majority
of statistical software (SAS, R, StatXact, Stata, SPSS, etc.). For a simple
2 × 2 table, the p-value calculation should not take a long time even for a
large sample size.

The data from Example 1.1 is used to illustrate the application of the
conditional approach. This is a randomized placebo-control two-arm study
for patients with chronic non-cancer-related pain [3]. If this study assumes
that the treatment response rate should be higher than that from the placebo
group in advance, then a one-sided hypothesis would be appropriate with the
alternative hypothesis as Ha : pt > pc , where pt and pc are the response rate
for the treatment group and the placebo group, respectively. In this study, the
response rates for the treatment group and the placebo group are estimated as
pt = 27.1% and pc = 18.9%, respectively, from the observed data in Table
1.2. When both marginal totals are fixed as in the conditional approach,
the value, n11 , determines the other three values, and the sample space can
be simplified as a collection of all possible n11 values. Given the marginal
totals, n1 and n2 for column totals, m1 and m2 for row totals as in Table
1.1, the range of the possible values for n11 is from max(0, n1 + m1 − N)
to min(n1 , m1 ). For this particular example, this range is from 0 to 99,
therefore the size of the sample space for the conditional approach is 100.
For each data point in the sample space, its probability can be computed
from Equation (1.1). The null hypothesis99 will be rejected when n11 ≥ 58,
therefore, the p-value is calculated as n11 =58 PC (n11 ) = 0.028.

When a hypothesis testing problem is two-sided, multiple approaches

have been proposed to compute its p-value. Among these approaches, the
most commonly used one is from Irwin [30] who proposed computing the
two-sided p-value by adding all data whose probability is smaller than or
equal to that of the observed data. In other words, the table probability
in Equation (1.1) is used as a test statistic and the tail area includes all
tables with smaller probabilities. The second approach uses the test statistic:
|n11 − n1 m1 /N| to determine the tail area that is defined as a collection of
tables whose test statistics are at least as large as that of the given data.
Another widely used approach doubles the one-sided p-value, which may
be larger than 1 in some cases [5, 31]. It is obvious that different approaches
for a two-sided problem generally do not have the same p-value, thus they
may lead to different conclusions. It should be noted that the two-sided
exact conditional p-value would be strictly defined when the test statistic for
ordering the sample space is only used to compute a two-sided asymptotic
p-value, such as the χ 2 test statistic.
Exploring the Variety of Random
Documents with Different Content
back
back
back
back
back
back
back
back
back
back
back
back
back
back
back