DESIGN and ANALYSIS

of QUALITY of LIFE
STUDIES in CLINICAL
TRIALS
CHAPMAN & HALL/CRC
I n t e r d i s c i p l i n a r y S t a t i s t i c s S e r i e s
Series editors: N Keiding, B Morgan, T Speed, P van der Heijden.
AN INVARIANT APPROACH TO S. Lele and J. Richtsmeier
STATISTICAL ANALYSIS OF SHAPES
ASTROSTATISTICS G. Babu and E. Feigelson
CLINICAL TRIALS IN ONCOLOGY J. Crowley, S. Green,
and J. Benedetti
DYNAMICAL SEARCH L. Pronzato, H. Wynn, and
A. Zhigljavsky
GRAPHICAL ANALYSIS OF K. Basford and J. Tukey
MULTI-RESPONSE DATA
INTRODUCTION TO M. Waterman
COMPUTATIONAL BIOLOGY:
MAPS SEQUENCES AND GENOMES
MARKOV CHAIN MONTE CARLO W. Gilks, S. Richardson,
IN PRACTICE and D. Spiegelhalter
STATISTICS FOR ENVIRONMENTAL A. Bailer and W. Piegorsch
BIOLOGY AND TOXICOLOGY
DESIGN AND ANALYSIS OF QUALITY Diane L. Fairclough
LIFE STUDIES IN CLINICAL TRIALS
CHAPMAN & HALL/CRC
A CRC Press Company
Boca Raton London New York Washington, D.C.
DESIGN and ANALYSIS
of QUALITY of LIFE
STUDIES in CLINICAL
TRIALS
I n t e r d i s c i p l i n a r y S t a t i s t i c s
Diane L. Fairclough
c2638 C2638Texturesfm February 12, 2002 9:57
Library of Congress Cataloging-in-Publication Data
Fairclough, Diane Lynn.
Design and analysis of quality of life studies in clinical trials / Diane L.
Fairclough.
p. cm. (Interdisciplinary statistics series)
Includes bibliographical references and index.
ISBN 1-58488-263-8 (alk. paper)
1. Clinical trialsLongitudinal studies. 2. Clinical trialsStatistical
methods. 3. Quality of lifeResearchMethodology. I. Title.
II. Interdisciplinary statistics
R853.C55 .F355 2002
615.5

07

24dc21 2002017479
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 eorts have been made to publish reliable data and in-
formation, 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, microlming, 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. Specic 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 trade-
marks, and are used only for identication and explanation, without intent to infringe.
Visit the CRC Press Web site at www.crcpress.com
2002 by Chapman & Hall/CRC
No claim to original U.S. Government works
International Standard Book Number 1-58488-263-8
Library of Congress Card Number 2002017479
Printed in the United States of America 1 2 3 4 5 6 7 8 9 0
Printed on acid-free paper
c2638 C2638Texturesfm February 12, 2002 9:57
Contents
Preface xv
Acknowledgments xix
1 Introduction 1
1.1 Health-related quality of life 1
1.2 Measuring health-related quality of life 2
Characteristics of various measures 2
Health status vs. patient preferences 2
Objective vs. subjective 4
Generic vs. disease-specic instruments 4
Global index vs. prole of domain-specic measures 4
Response format 6
Period of recall 6
1.3 Example 1: Adjuvant breast cancer trial 7
Patient selection 7
Treatment 7
Quality of life measure 8
Timing of HRQoL assessments 8
Questionnaire completion/missing data 10
1.4 Example 2: Advanced non-small-cell lung cancer (NSCLC) 10
Treatment 12
Quality of life measure 12
Timing of assessments 13
Questionnaire completion/missing data 13
1.5 Example 3: Renal cell carcinoma trial 15
Patient selection 15
Treatment 15
Quality of life measure 15
Timing of HRQoL assessments 16
Questionnaire completion/missing data 17
1.6 Summary 18
2 Study Design and Protocol Development 19
2.1 Introduction 19
2.2 Background and rationale 19
2.3 Research objectives 20
Domains of interest 21
Pragmatic vs. explanatory inference 21
c2638 C2638Texturesfm February 12, 2002 9:57
vi CONTENTS
2.4 Selection of subjects 22
2.5 Longitudinal designs 23
Event- or condition-driven designs 23
Time-driven designs 24
Timing of the initial HRQoL assessment 24
Timing of the follow-up HRQoL assessments 24
Timing of HRQoL assessments when therapy
is cyclic 25
Trials with dierent schedules of therapy 25
Frequency of evaluations 25
Duration of HRQoL assessment 26
Assessment after discontinuation of therapy 27
2.6 Selection of a quality of life measure 27
Trial objectives 28
Validity and reliability 29
Appropriateness 30
2.7 Conduct 31
Mode of administration 31
Data collection and management 32
Avoiding missing data 32
Education 33
Forms 34
Explicit procedures for follow-up 34
Scoring instruments 35
Reverse coding 35
Scoring multi-item scales 35
Item nonresponse 37
SAS example 38
2.8 Summary 39
3 Models for Longitudinal Studies 41
3.1 Introduction 41
Repeated measures models 41
Growth curve models 42
Selection between models 42
Adjuvant breast cancer study (Example 1) 42
NSCLC study (Example 2) 43
Renal cell carcinoma study (Example 3) 43
3.2 Building the analytic models 43
Statistics guiding model reduction 44
Likelihood ratio tests 44
Other statistics 46
3.3 Building repeated measures models 46
Mean structure 46
SAS example of a cell means model 47
Covariance structures 48
c2638 C2638Texturesfm February 12, 2002 9:57
CONTENTS vii
Unstructured covariance 48
Structured covariance 49
SAS example of a repeated measures model 50
Hypothesis testing 52
3.4 Building growth curve models 54
Model for means 54
Polynomial models 54
Piecewise linear regression 54
Covariance structure 57
Variance of random eects 57
Variance of residual errors 58
SAS example of a polynomial growth curve model 59
Fully parameterized model for the means 59
Covariance structure 59
Model reduction 63
Estimation 63
Hypothesis testing 64
SAS example of a piecewise linear regression model 65
Fully parameterized model for the means 65
Covariance structure 65
Model reduction 66
Estimation 66
Testing 67
3.5 Summary 68
4 Missing Data 69
4.1 Introduction 69
Terminology 69
Why are missing data a problem? 69
How much data can be missing? 70
Similar patterns of dropout among intervention arms 71
Prevention 71
4.2 Patterns of missing data 71
Example: NSCLC study 72
4.3 Mechanisms of missing data 72
Notation 72
Example 73
The concept 74
MCAR: Missing completely at random 75
The concept 75
Covariate-dependent dropout 75
Identifying covariate-dependent missingness 76
Example: NSCLC study 76
Analytic methods 77
MAR: Missing at random 77
The concept 77
c2638 C2638Texturesfm February 12, 2002 9:57
viii CONTENTS
Identication of dependence on observed data (Y
obs
i
) 78
Analytic methods 80
A test of MCAR vs. MAR for multivariate normal data 81
Notation 81
Test statistic 81
NSCLC example 81
Implementing in SAS 82
MNAR: Missing not at random 84
The concept 84
Analytic methods 84
Identication of dependence on unobserved
data, Y
mis
i
85
Example: NSCLC study 85
4.4 Renal cell carcinoma study 87
Plotting outcome by dropout 89
4.5 Summary 90
5 Analytic Methods for Ignorable Missing Data 93
5.1 Introduction 93
Hypothetical example 93
5.2 Repeated univariate analyses 94
NSCLC example 96
5.3 Multivariate methods 96
Complete case analysis (MANOVA) 97
NSCLC example 98
Maximum likelihood estimation with all
available data 101
NSCLC example 102
Further comments 104
Exclusion of subjects 104
Exclusion of observations 105
5.4 Baseline assessment as a covariate 105
NSCLC example 107
5.5 Change from baseline 108
NSCLC example 109
5.6 Adding other baseline covariates 110
NSCLC example 111
5.7 Empirical Bayes estimates 112
5.8 Summary 114
6 Simple Imputation 115
6.1 Introduction 115
Limitations of simple imputation 116
NSCLC example 116
6.2 Mean value substitution 117
6.3 Explicit regression models 118
c2638 C2638Texturesfm February 12, 2002 9:57
CONTENTS ix
Identication of the imputation model 119
Simple univariate regression 120
Conditional predicted values 123
6.4 Last value carried forward 125
-Adjustments 126
Arbitrary high or low value 127
6.5 Underestimation of variance 128
6.6 Sensitivity analysis 130
6.7 Summary 130
7 Multiple Imputation 131
7.1 Introduction 131
7.2 Overview of multiple imputation 131
Step 1: Selection of the imputation procedure 131
Step 2: Generation of M imputed data sets 132
Step 3: Analysis of M data sets 132
Step 4: Combining results of M analyses 132
7.3 Explicit univariate regression 133
Identication of the imputation model 133
Computation of imputed values 134
Practical considerations 135
Extensions to longitudinal studies 135
Assumptions 135
NSCLC example 136
7.4 Closest neighbor and predictive mean matching 140
Closest neighbor 142
Predictive mean matching 142
7.5 Approximate Bayesian bootstrap 142
The basic procedure 143
Extensions to longitudinal studies 144
Propensity scores 144
Practical issues 144
The assumptions 145
Nonignorable missing data 145
7.6 Multivariate procedures for nonmonotone missing data 146
NSCLC example 146
7.7 Combining the M analyses 147
SAS example 148
7.8 Sensitivity analyses 150
7.9 Imputation vs. analytic models 151
7.10 Implications for design 152
7.11 Summary 152
8 Pattern Mixture Models 153
8.1 Introduction 153
NSCLC example 154
c2638 C2638Texturesfm February 12, 2002 9:57
x CONTENTS
8.2 Bivariate data (two repeated measures) 155
NSCLC example 156
Complete-case missing variable (CCMV) restriction 156
NSCLC example 158
Browns protective restrictions 158
NSCLC example 159
Sensitivity analyses with intermediate restrictions 160
Large-sample inferences for
2
160
NSCLC example 161
8.3 Monotone dropout 161
NSCLC study 162
Complete-case missing value restriction 162
Available case missing value restriction 164
Neighboring case missing value restriction 164
Comparison of CCMV, ACMV, and NCMV estimates 167
8.4 Parametric models 167
Linear trends 168
Variance estimation 172
SAS example of a pattern mixture model 174
Step 1: Estimates of
P
h
174
Step 2: Estimates of
P
h
174
Step 3: Pooling estimates and computing variance 175
Step 4: Hypothesis testing 177
8.5 Additional reading 178
Extensions of bivariate case 178
Extensions of the sensitivity analysis 178
Nonparametric analyses 178
8.6 Algebraic details 178
Simple linear regression of Y on X 178
Complete-case missing variable restriction 179
Equation 8.9 179
Equation 8.11 179
Browns protective restriction 179
Equation 8.17 179
Equation 8.19 180
Other 180
8.7 Summary 181
9 Random-Eects Mixture, Shared-Parameter,
and Selection Models 183
9.1 Introduction 183
Mixture models 183
Selection models 184
Overview 184
9.2 Conditional linear model 185
c2638 C2638Texturesfm February 12, 2002 9:57
CONTENTS xi
Testing MAR vs. MNAR under assumptions
of conditional linear model 187
NSCLC example 187
Estimation of the standard errors 192
Assumptions 192
Random-coecient mixture model 192
9.3 Joint mixed-eects and time to dropout 193
Testing MAR vs. MNAR under the assumptions
of the joint model 194
Selection or mixture model? 195
NSCLC example 195
Initial estimates 197
Extension to more complex mixed-eects models 198
Renal cell carcinoma example 198
9.4 Selection model for monotone dropout 198
Outcome-dependent selection model 201
NSCLC example 201
Oswald program 206
Longitudinal model 206
Dropout model 208
Oswald warnings 210
9.5 Advanced readings 210
Intermittent missing data 210
More selection models 210
Heckman probit stochastic dropout model 210
Wu and Carroll 210
Mori 210
Nonparametric analyses 211
9.6 Summary 211
10 Summary Measures 213
10.1 Introduction 213
Addressing multiplicity of endpoints 213
Summary measures vs. summary statistics 213
Strengths and weaknesses 215
Easier interpretation 215
Increased power 215
Weakness 215
10.2 Choosing a summary measure 215
10.3 Constructing summary measures 217
Notation 220
Missing data 220
Average rate of change (slopes) 222
NSCLC example 222
Missing data 222
Area under the curve 225
c2638 C2638Texturesfm February 12, 2002 9:57
xii CONTENTS
Missing data 225
Dierences at baseline 227
Average of ranks 227
Missing data 227
Univariate analysis of summary measures 228
Stratied analysis of summary measures 228
NSCLC example 229
10.4 Summary statistics across time 231
Notation 231
Area under the curve 231
Repeated measures 231
Growth curve models 232
10.5 Summarizing across HRQoL domains or subscales 235
Summary measures 235
Weighting proportional to the number
of questions 236
Factor analytic weights 236
Patient weights 237
Statistically derived weights: Inverse correlation 238
Summary statistics 239
10.6 Advanced notes 240
Nonparametric procedures 240
Combining HRQoL and time to event 240
Area under the curve 240
Latent variable models 241
10.7 Summary 241
11 Multiple Endpoints 243
11.1 Introduction 243
Limiting the number of conrmatory tests 243
Summary measures and statistics 244
Multiple comparison procedures 244
11.2 Background concepts and denitions 245
Univariate vs. multivariate test statistics 245
Familywise and experimentwise error rates 245
Global vs. individual tests 246
11.3 Multivariate statistics 246
Global tests 246
Closed multivariate testing procedures 246
Limitations 248
11.4 Univariate statistics 249
NSCLC example 249
Alpha adjustments for K univariate tests 249
Bonferroni adjustment 249
R ugers inequality 252
Simes global test 252
c2638 C2638Texturesfm February 12, 2002 9:57
CONTENTS xiii
Sequential rejective Bonferroni procedure 253
p-value adjustments 253
11.5 Resampling techniques 254
11.6 Summary 255
12 Design: Analysis Plans 257
12.1 Introduction 257
12.2 General analysis planWho is included? 258
12.3 Models for longitudinal data 259
Ignorable missing data 259
Nonignorable missing data 259
Event-driven designs and repeated
measures models 259
Time-driven designs and growth curve models 260
Modication of analysis plan 260
12.4 Multiplicity of endpoints 260
Primary vs. secondary endpoints 260
Summary measures 261
Multiple comparison procedures 261
12.5 Sample size and power 262
Simple linear combinations of 262
Basic assumptions 264
Incomplete designs 264
Example 1: Repeated measures 266
Example 2: Growth curve model 268
Other considerations 270
Intermittent missing data patterns
and time-varying covariates 270
Unequal allocations of subjects to treatment groups 270
Multivariate tests 271
Small sample size approximations 272
Restricted maximum likelihood estimation 272
12.6 Reporting results 272
12.7 Summary 274
Appendix I Abbreviations 275
Appendix II Notation 277
Appendix III Formal Denitions for Missing Data 281
References 285
Index 295
c2638 C2638Texturesfm February 12, 2002 9:57
c2638 C2638Texturesfm February 12, 2002 9:57
Preface
For almost 20 years I have been engaged by the challenges of design and analy-
sis of longitudinal assessment of health outcomes in children and adults. Most
of my research has been on the psychosocial outcomes for pediatric cancer
survivors and health-related quality of life (HRQoL) of adult cancer patients,
but it has included other diagnoses. As I have designed and analyzed these
studies, many of the same themes continued to occur. This was irrespective
of whether the studies were undertaken by the large cooperative groups, the
pharmaceutical industry, or single institutions. The most common issues con-
cerned missing or mistimed observations in longitudinal studies. The other
issue is that of multiple comparisons. Five years ago, I began giving work-
shops three or four times a year on the design and analysis of HRQoL studies.
My intent was to summarize that experience in this book; little did I realize
how much more I would learn during the 2 years that I have been writing.
Numerous books discuss the wide range of topics that need to be addressed
in the evaluation of quality of life. For example, Spilker [142] edited a com-
prehensive book published in 1996 with 1259 pages that includes chapters on
general concepts; developing, choosing, and administering instruments; special
populations; and cross-cultural aspects. Only 2 of the 127 chapters address
analysis. Two years later, Staquet et al. [146] edited another book focusing
on methods in clinical trials. Of the 18 chapters, 4 are devoted to design and
analysis. Fayers and Machin [45] more recently published a text that focuses
on assessment (questionnaire development and testing), analysis, and inter-
pretation. Three chapters are devoted to analysis of longitudinal studies and
three address design issues in clinical trials.
There still seemed to be a need for a book that addresses design and anal-
ysis in enough detail to enable readers to apply the methods to their own
studies. To achieve that goal, I have limited the focus of the book to longitu-
dinal studies of quality of life in clinical trials. Three clinical trials are used
throughout the book to illustrate practical implementation of strategies and
analytic methods. Finally, examples of SAS and S-Plus programs are included
in the book. The intent is that readers, by following the examples in the book,
will be able to follow the steps outlined with their own studies.
Intended Readers
My primary audience for this book is the researcher who is directly involved
in the design and analysis of HRQoL studies. However, the book will also
c2638 C2638Texturesfm February 12, 2002 9:57
xvi PREFACE
be useful to those who are expected to evaluate the design and interpret the
results of HRQoL research.
More than any other research that I have been involved with, HRQoL re-
search draws investigators from all elds of study with a wide range of train-
ing. This has included epidemiologists, psychologists, sociologists, behavioral
and health researchers, clinicians, and nurses from all specialties, as well as
statisticians. I expect that most readers will have had some graduate-level
training in statistical methods including multivariate regression and multi-
variate analysis of variance. However, that training may have been some time
ago and prior to some of the more recent advances in statistical methods.
With that in mind, I have organized most chapters so that the concepts are
discussed in the beginning of the chapters and sections. When possible, the
technical details appear later in the chapters. Finally, each chapter ends with
a summary of the important points.
Contents
Chapter 1 describes the general concept of HRQoL and characteristics of
instruments used to assess HRQoL. Then the three clinical trials that are
used throughout the book are described in detail.
I consider the careful and explicit denition of the research question to be
the most important and critical step in the entire research process. All other
decisions about the design and analysis follow. The principles of design will be
considered throughout the book. In addition, Chapter 2 focuses on protocol
development for longitudinal studies with HRQoL as an endpoint. Chapter 12
concludes the book with the development of analytic plans.
Because longitudinal studies of free-living subjects are subject to missing
data, a major proportion of this book is devoted to that topic. Measurement
of HRQoL and other patient-reported outcomes is particularly vulnerable to
missing data in studies of patients who are experiencing morbidity or mortality
due to their diseases or their treatments. Chapter 3 describes two general ap-
proaches to modeling longitudinal data, repeated measures and growth curve
models, with a strategy for choosing between them. This will be a review for
analysts who are regularly presented with longitudinal data. Chapter 4 de-
nes patterns and mechanisms for missing data and then illustrates how one
might determine which mechanisms apply in a particular clinical trial. This
guides the selection of the appropriate analytic methods that are discussed in
Chapters 5 through 9.
The greatest analytic challenge occurs when observations are missing be-
cause the patient is currently experiencing poorer (or better) HRQoL. If the
data are missing for reasons related to toxicity of therapy or progression of dis-
ease, clinical experience suggests that the HRQoL at those time points where
the patient was unable to complete the assessment is likely to be poorer than
that at the time points where the patient completes the assessment. It is likely
that no single approach is the best in all settings. Chapter 5 describes sev-
eral widely used methods of analysis of longitudinal studies including repeated
c2638 C2638Texturesfm February 12, 2002 9:57
PREFACE xvii
univariate analyses, MANOVA, and mixed eects models. Chapter 6 discusses
simple imputation of missing data. Most of the concepts in these two chap-
ters are basic, but I am motivated to include this basic discussion because of
common mistakes that I frequently observe. Chapters 7 through 9 discuss the
more challenging methods for studies with non-ignorable (nonrandom) miss-
ing data. Analytic strategies such as multiple imputation (Chapter 7), pat-
tern mixture models (Chapter 8), and shared parameter and selection models
(Chapter 9) allow us to examine the sensitivity of the results to dierent as-
sumptions about the missing data mechanism.
Typically, clinical trials measure multiple aspects of HRQoL or patient-
reported outcomes repeatedly over time. As a result there are multiple scales
with longitudinal assessment. To address the issues raised by this, Chapter 10
proposes the use of summary measures to facilitate interpretation and to
increase the power of studies to detect meaningful changes in HRQoL. In
Chapter 11, I discuss and illustrate the use of multiple comparison proce-
dures to reduce the Type I error rates associated with hypothesis testing of
multiple endpoints.
The Future
It was very dicult to stop writing this book. I would like to explore many
topics that are not covered, new developments, and yet undiscovered old
developments. One of the most obvious omissions is the development, trans-
lation, cultural adaptation, and validation of HRQoL instruments. If done
well, instrument development requires considerable time and resources. This
eort is often not appreciated by researchers who have never been involved
in the process. I have chosen not to address this subject, as there are a
large number of books already devoted to instrument development, including
Schuman and Presser [143], Converse and Presser [25], Floyd and Fowler [50],
DeVellis [33], Streiner and Norman [147], Juniper et al. [142, chapter 6],
Frank-Stromborg and Olsen [52], Staquet et al. [146, part IV], and Fayers
and Machin [45, chapters 2 through 7]. I am also sure that I will discover
new ideas even before this book is published, but that is for the future.
Diane L. Fairclough
c2638 C2638Texturesfm February 12, 2002 9:57
c2638 C2638Texturesfm February 12, 2002 9:57
Acknowledgments
I would like to thank all my friends and colleagues for their support and help.
I would specically like to thank Pam Wolfe, Mirium Dickenson, Melanie
Bell, and Kathe Bjork for their careful proofreading of the material and Doris
Borchert for nding all the journal articles that I requested.
c2638 C2638Texturesfm February 12, 2002 9:57
c2638 C2638TextureC01 February 8, 2002 15:39
CHAPTER 1
Introduction
1.1 Health-related quality of life
Traditionally, clinical trials have focused on endpoints that are physical or lab-
oratory measures of response. For example, therapies for cancer are evaluated
on the basis of disease progression and survival. The ecacy of a treatment for
anemia is evaluated by hemoglobin levels or number of transfusions required.
Although traditional biomedical measures are often the primary endpoints in
clinical trials, they do not reect how the patient feels and functions in daily
activities, yet these perceptions reect whether or not the patient believes
he or she has beneted from the treatment. In certain diseases, the patients
perception of his or her well-being may be the most important health out-
come [144]. More recently, clinical trials are including endpoints that reect
the patients perception of his or her well-being and satisfaction with therapy.
Sometimes clinical investigators assume that a certain change in therapy or
a traditional biomedical outcome will improve the patients quality of life.
While in many cases this may be true, sometimes surprising results are ob-
tained when the patient is asked directly. One classic example of this occurred
with a study by Sugarbaker et al. [148] comparing two therapeutic approaches
for soft-tissue sarcoma. The initial study compared two therapeutic options.
The rst was limb-sparing surgery followed by radiation therapy. The second
treatment approach was full amputation of the aected limb. The investigator
hypothesized, Sparing a limb, as opposed to amputating it, oers a qual-
ity of life advantage. As a result of the study, the hypothesis was rejected;
subjects receiving the limb-sparing procedures reported limitations in mobil-
ity and sexual functioning. These observations were conrmed with physical
assessments of mobility and endocrine function. Radiation therapy was mod-
ied and physical rehabilitation was added to the limb-sparing therapeutic
approach [65].
The World Health Organization (WHO) dened health in 1948 [164, 165]
as a state of complete physical, mental and social well-being and not merely
the absence of inrmity and disease. This denition reects the focus on a
broader picture of health. Wilson and Cleary [163] propose a conceptual model
of the relationships among health outcomes (Figure 1.1). Five levels of out-
comes progress from biomedical outcomes to quality of life. The biological
and physiological outcomes include the results of laboratory tests, radiologi-
cal scans, and physical examination as well as diagnoses. Symptom status is
dened as a patients perception of an abnormal physical, emotional or cogni-
tive state. Functional status includes four dimensions: physical, physiological,
c2638 C2638TextureC01 February 8, 2002 15:39
2 INTRODUCTION
Biological and
Physiological
Factors

Symptom
Status

Functional
Status

General
Health
Perception

Overall
Quality
of Life
Figure 1.1. Five progressive levels of outcomes from Wilson and Clearys concep-
tual model of the relationship among health outcomes.
social, and role. General health perceptions include the patients evaluation of
past and current health, his or her future outlook, and concerns about health.
The term quality of life has been used in many ways. Although the exact
denitions vary among authors, there is general agreement that it is a multi-
dimensional concept that focuses on the impact of disease and its treatment
on the well-being of an individual.
In the broadest denition, the quality of our lives is inuenced by our phys-
ical and social environment as well as our emotional and existential reactions
to that environment. Kaplan and Bust [77] proposed the use of the term
health-related quality of life (HRQoL) to distinguish health eects from other
factors inuencing the subjects perceptions, including job satisfaction and
environmental factors. In most clinical research, we limit the scope to focus
on the measurement of HRQoL. Cella and Bonomi [16] state, Health-related
quality of life refers to the extent to which ones usual or expected physi-
cal, emotional and social well-being are aected by a medical condition or its
treatment. In some settings, we may also include other aspects like economic
and existential well-being. Patrick and Erickson [112] propose a more inclu-
sive denition, which combines quality and quantity: The value assigned to
duration of life as modied by the impairments, functional states, perceptions
and social opportunities that are inuenced by disease, injury, treatment or
policy. It is important to note that an individuals well-being or health sta-
tus cannot be directly measured. We are only able to make inferences from
measurable indicators of symptoms and reported perceptions.
Often the term quality of life is used when any patient-reported outcome
is measured. Side eects are not equivalent to quality of life. Solely assessing
symptoms is a simple, convenient way of avoiding the more complex task of as-
sessing HRQoL. Symptoms do impact HRQoL and are part of the assessment
of HRQoL.
1.2 Measuring health-related quality of life
Characteristics of various measures
Health status vs. patient preferences
There are two general types of HRQoL measures, health status assessment and
patient preference assessment [169]. These two forms of assessment have de-
veloped as a result of the dierences between the perspectives of two dierent
disciplines: psychometrics and econometrics. In the health status assessment
c2638 C2638TextureC01 February 8, 2002 15:39
MEASURING HEALTH-RELATED QUALITY OF LIFE 3
measures, multiple aspects of the patients perceived well-being are self-
assessed and a score is derived from the responses on a series of questions. This
score reects the patients relative HRQoL as compared to other patients and
to the HRQoL of the same patient at other times. The assessments range from
a single global question asking patients to rate their current quality of life to
a series of questions about specic aspects of their daily life during a recent
period of time. These instruments generally take 5 to 10 min to complete. Ex-
amples include the Functional Living IndexCancer (FLIC) [130], European
Organization for Research and Treatment of Cancer EORTC QLQ-C30 [1],
and the Functional Assessment of Cancer Therapy (FACT) [13]. Among these
health status measures, there is considerable range in the context of the ques-
tions, with some measures focusing more on the perceived impact of the dis-
ease and therapy (How much are you bothered by hair loss?) and other mea-
sures focusing on the frequency and severity of symptoms (How often do you
experience pain?). These measures are primarily designed to compare groups
of patients receiving dierent treatments or to identify change over time within
groups of patients. As a result, these measures have most often been used in
clinical trials to facilitate the comparisons of therapeutic regimens.
Quality of life measures in patient preference assessments are inuenced
strongly by the concept of utility, borrowed from econometrics, that reects
individual decision making under uncertainty. These preference assessment
measures are primarily used to evaluate the tradeo between the quality and
quantity of life. Values of utilities are always between 0 and 1, with 0 gener-
ally associated with death and 1 with perfect health. Examples include time
tradeos [98], standard gamble [150], and multiattribute assessment measures
such as the Health Utility Index (HUI) [46] and the Q-Tility index [159]. Time
tradeo utilities are measured by asking respondents how much time in their
current health state they would give up for a specied period of time in per-
fect health. If, for example, a patient responded that he would give up 1 year
in his current health for 5 years of perfect health, the resulting utility is 0.8.
Standard gamble utilities are measured by asking respondents to identify the
point at which they become indierent to the choices between two hypothet-
ical situations. For example, one option is a radical surgical procedure with
no chance of relapse but signicant impact on HRQoL and the other option
is watchful waiting, with a chance of relapse and death. The chance of relapse
and death is raised or lowered until the respondent considers the two options
to be equivalent. Assessment of time tradeo and standard gamble utilities re-
quires the presence of a trained interviewer or specialized computer program.
Because of the resource needs, these approaches are generally too time- and
resource-intensive to use in a large clinical trial. Multiattribute assessment
measures combine the advantages of self-assessment with the conceptual ad-
vantages of utility scores. Their use is limited by the need to develop and
validate the methods by which the multiattribute assessment scores are con-
verted to utility scores for each of the possible health states dened by the
multiattribute assessments. For example, if a measure assesses ve domains of
HRQoL, with three possible levels for each domain, there will be 243 possible
c2638 C2638TextureC01 February 8, 2002 15:39
4 INTRODUCTION
health states with a corresponding utility score. Utilities have traditionally
been used in the calculation of quality-adjusted life years for cost-eectiveness
analyses and in analytic approaches such as Q-TWiST [56, 58].
Objective vs. subjective
Health status measures dier in the extent to which they assess observ-
able phenomena or require the respondent to make inferences. On the one
hand, these measures will assess symptoms or functional benchmarks. In these
instruments, subjects are asked about the frequency and severity of symptoms
or whether they can perform certain tasks such as walking a mile. At the other
end of the scale, the impact of symptoms or conditions is assessed. In these
instruments, subjects are asked how much symptoms bother them or inter-
fere with their usual activities. Many instruments provide a combination. The
value of each will depend on the research objectives: Is the focus to identify
the intervention with the least severity of symptoms or to assess the impact
of the disease and its treatment?
There may exist a misconception that objective assessments are more valid.
This is generally based on the observation that patient ratings do not agree
with ratings of trained professionals. If we take objective assessments to be
more valid, we are assuming that the ratings of the professionals constitute the
gold standard, when in fact they have more limited information than the pa-
tient about how the patient views his or her health and quality of life. Further,
many of the biomedical endpoints that we consider objective include a high
degree of measurement error (e.g., blood pressure), are misclassied among
experts, or have poor predictive and prognostic validity (e.g., pulmonary func-
tion tests) [162].
Generic vs. disease-specic instruments
There are two basic types of instrumentsgeneric and disease-specic. The
generic instrument is designed to assess HRQoL in individuals with and with-
out active disease. The MOS SF-36 is a classic example of a generic instru-
ment. This broad basis of a generic test is an advantage when comparing vastly
dierent groups of subjects or following subjects for extended periods after
treatment has ended. Disease-specic instruments narrow the scope of assess-
ment and address in a more detailed manner the impact of a particular disease
or treatment. As a result, they are more sensitive to smaller but clinically sig-
nicant changes induced by treatment.
Global index vs. prole of domain-specic measures
Some HRQoL instruments are designed to provide a single global index of
HRQoL. Others are designed to provide a prole of the dimensions such as
the physical, emotional, functional, and social well-being of patients. Many
instruments attempt to assess both but often in dierent ways. The index
and the prole represent two dierent approaches to the use of the HRQoL
c2638 C2638TextureC01 February 8, 2002 15:39
MEASURING HEALTH-RELATED QUALITY OF LIFE 5
Table 1.1. Questions assessing global quality of life in the EORTC QLQ-30.
How would you rate your overall health during the past week?
1 2 3 4 5 6 7
Very poor Excellent
How would you rate your overall quality of life during the past week?
1 2 3 4 5 6 7
Very poor Excellent
measure. The advantage of a single index is that it provides a straightforward
approach to decision making. Indexes that are in the form of utilities are very
useful in cost-eectiveness analyses performed in pharmacoeconomic research.
Health proles are useful when the objective is to measure potentially dierent
eects on the multiple dimensions of HRQoL.
When a construct is quite simple, a single question may be adequate to
provide a reliable measure.* As the construct becomes more complex, more
aspects of the construct need to be incorporated into the measure, creating a
demand for multiple items. However, at some point the construct becomes so
complex that a single question is used, allowing the subject to integrate the
complex aspects into one response. At this point there is no way to identify
all of the various aspects that are relevant.
Two widely used instruments designed for patients with cancer, FACT
and EORTC QLQ-30, use summated scores from multiple questions to cre-
ate domain-specic measures, but they take very dierent approaches to ob-
tain global measures. For the FACT measures [14, 15], the overall score is
the summated score from all items in the questionnaire. In Version 4 of the
FACT, this consists of seven questions, each measuring physical, functional,
and social well-being; six questions measuring emotional well-being; and a
variable number addressing disease-specic concerns. In the EORTC QLQ-
30, Aaronson et al. [1] do not propose adding the subscale scores to form an
overall score, arguing that a prole of the various domains more accurately
reects the multidimensional character of quality of life. The global score is
instead constructed from two questions that are distinct from those used to
measure specic domains (Table 1.1).
A global measure of HRQoL has the obvious attraction of the simplicity of
a single indicator. Unfortunately, it is dicult, if not impossible, to construct
a single score that aggregates the multiple dimensions of HRQoL that is valid
in all contexts. There is also a potential for the loss of information in a single
measure. It is possible that a particular intervention may produce benets in
one dimension and decits in another which cancel each other and are thus
not observed.
*
In general, scales constructed from multiple items have better reliability than single items,
as the random measurement errors tend to cancel.
c2638 C2638TextureC01 February 8, 2002 15:39
6 INTRODUCTION
Table 1.2. Example of a Likert and visual analog scale.
Likert Scale
How bothered were you?
Not at all Slightly Moderately Quite a bit Greatly
0 1 2 3 4
Visual Analog Scale
How bothered were you?
Not at all Greatly
Response format
Questionnaires may also dier in their response format. Two of the most
widely used formats are the Likert and visual analog scales (VAS) (Table 1.2).
The Likert scale contains a limited number of ordered responses that have a
descriptive label associated with each level. Individuals can discriminate at
most seven to ten ordered categories [101, 147, p. 35], and reliability and
the ability to detect change drop o at ve or fewer levels. It is likely that
as patients become more ill, the number of response levels they can easily
respond to drops.
The VAS consists of a line with descriptive anchors at each end of the line.
The respondent is instructed to place a mark on the line. Often the length of
the line is 10 cm. The concept behind the VAS is that the measure is continu-
ous and can potentially discriminate more eectively than a Likert scale. This
has not generally been true in validation studies where both formats have
been used. The VAS format has several limitations. First, it requires a level
of eyehand coordination that may be unrealistic for anyone with a neurolog-
ical condition and for elderly people. Second, it requires an additional data
management step in which the position of the mark is measured. If forms are
copied, rather than printed, the full length of the line may vary, requiring two
measurements and additional calculations. Third, VAS precludes telephone
assessment and interview formats.
Period of recall
HRQoL scales often request subjects to base their evaluations on the last 7
days or 4 weeks. The selection of the appropriate time frame is a balance
between being able to detect dierences between treatments and to minimize
short-term uctuations (noise) that do not represent real change [102]. Scales
specic to diseases or treatments where there can be rapid changes will have
a shorter recall duration. Two widely used instruments for cancer patients,
the Functional Assessment of Cancer Therapy (FACT) and the European
Organization for Research and Treatment of Cancer scale (EORTC QLQ-C30)
c2638 C2638TextureC01 February 8, 2002 15:39
EXAMPLE 1: ADJUVANT BREAST CANCER TRIAL 7
use 7 days. HRQoL instruments designed for assessment of general populations
will often have a longer recall duration. For example, the Medical Outcomes
Study SF-36 uses a 4-week window for most of its questions.
1.3 Example 1: Adjuvant breast cancer trial
Increasing dose intensity and scheduling drugs more eectively are two strate-
gies for improving the eectiveness of breast cancer adjuvant chemotherapy.
Although these strategies have advanced cancer therapy, the increased toxic-
ity with more aggressive regimens is a concern, especially if there are modest
improvements in survival. If more aggressive regimens produce comparable
disease control and survival benets, then the selection of optimum therapy
should include not only the assessment of toxicity but also the impact of
treatment on HRQoL.
In this trial, a 16-week dose-intensive therapy was compared to a conven-
tional 24-week therapy (CAF*) in patients with breast cancer [47].** The
primary hypothesis of the study was that disease-free and overall survival is
superior on the more aggressive 16-week regimen. However, the 16-week reg-
imen is a more dose-intense therapy than the 24-week CAF regimen and, in
addition to potentially increasing physical symptoms, the inconvenience of the
treatment and expected fatigue may have a greater impact on the psychosocial
aspect of patients lives. Thus, it was hypothesized that HRQoL during CAF
therapy is superior to that during the 16-week regimen. This discrepancy in
expectations creates a need to reconcile competing outcomes if they indeed
occur.
Patient selection
The patients eligible for the treatment trial had hormone receptor-negative,
node-positive breast cancer. Enrollment in the HRQoL substudy started after
the initiation of the treatment trial. Patients registered on the treatment trial
who had not yet started therapy were eligible for the quality of life study.
Patients were also required to be able to read and understand English to
be eligible. Consent was obtained separately for the treatment trial and the
HRQoL substudy.
Treatment
Patients were randomized to receive either 24 weeks of a three-drug (CAF)
regimen or a shorter, more intense 16-week multidrug regimen [48]. The CAF
regimen consisted of 28-day cycles with 14 days of oral therapy and 2 days
of intravenous therapy (days 1 and 8). Thus, patients on this regimen had
* CAF = cyclophosphamide, doxirubicin, and 5-urouracil.
** Eastern Cooperative Oncology Group funded by the National Cancer Institute Grant
CA-23318.
c2638 C2638TextureC01 February 8, 2002 15:39
8 INTRODUCTION
a 2-week break every 4 weeks. In contrast, the 16-week regimen consisted of
weekly therapy. During odd-numbered weeks patients received 7 days of oral
therapy plus 2 days of intravenous therapy. During even-numbered weeks,
patients received 2 days of intravenous therapy.
Quality of life measure
The Breast Chemotherapy Questionnaire (BCQ) was selected among avail-
able validated HRQoL instruments that were suitable for cooperative group
trials [102] and also suited to the goals of this study. The BCQ was developed
to evaluate treatment-related problems identied by patients and clinicians as
most important to HRQoL during breast cancer adjuvant chemotherapy [86].
It is a self-administered questionnaire of 30 questions about the past 2 weeks,
answered on a 1- to 7-point scale (Table 1.3). Seven domains were identied:
consequences of hair loss, positive well-being, physical symptoms, trouble and
inconvenience, fatigue, emotional dysfunction, and nausea.
The overall raw score (R) is calculated as the mean of answers to the
30 questions; higher scores indicate better HRQoL, with a range of possi-
ble scores from 1 to 7. The raw score is then rescaled (S =10 (R1)/6) so
that the range of possible BCQ scores (S) is from 0 to 10 points [86]. Similarly,
the subscale score for each of the seven domains is the mean of the responses
to the questions for that domain rescaled in the same manner. If any of the
items are skipped, the overall and subscale scores are calculated using the
mean of the completed items.
Timing of HRQoL assessments
The BCQ assessments were limited to one assessment before, during, and af-
ter treatment. The assessment prior to therapy was scheduled to be within 14
days of start of chemotherapy. The assessment during treatment was sched-
uled on day 85 of treatment. This was halfway through the CAF therapy (day
1 of cycle 4) and three quarters of the way through the 16-week regimen (day
1 of week 13). By day 85 it was expected that patients would be experiencing
the cumulative eects of both regimens without yet experiencing the psycho-
logical lift that occurs at the end of treatment. The third assessment following
therapy was scheduled 4 months after the completion of therapy. Since the
duration of therapy diered between the two treatment regimens (16 vs. 24
weeks), the third assessment occurred at dierent points in time for patients
on the dierent regimens but at comparable periods relative to the completion
of therapy. Additional variability in the timing of the third assessment was
introduced for women who discontinued therapy earlier than planned or for
those whose times to complete treatment may have been extended or delayed
as a result of toxicity. The exact timing of the assessments is illustrated in
Figure 1.2.
c2638 C2638TextureC01 February 8, 2002 15:39
EXAMPLE 1: ADJUVANT BREAST CANCER TRIAL 9
Table 1.3. Sample questions from the Breast Chemotherapy Questionnaire.
1. How often during the past 2 weeks have you felt worried or upset
as a result of thinning or loss of your hair?
(1) All of the time
(2) Most of the time
(3) A good bit of the time
(4) Some of the time
(5) A little of the time
(6) Hardly any of the time
(7) None of the time
2. How often during the past 2 weeks have you felt optimistic or
positive regarding the future?
(1) None of the time
(2) A little of the time
(3) Some of the time
.
.
.
(7) All of the time
3. How often during the past 2 weeks have you felt your ngers were
numb or falling asleep?
(1) All of the time
(2) Most of the time
(3) A good bit of the time
.
.
.
(7) None of the time
4. How much trouble or inconvenience have you had during the last
2 weeks as a result of having to come to or stay at the clinic or
hospital for medical care?
(1) A great deal of trouble or inconvenience
(2) A lot of trouble or inconvenience
(3) A fair bit of trouble or inconvenience
.
.
.
(7) No trouble or inconvenience
5. How often during the past 2 weeks have you felt low in energy?
(1) All of the time
(2) Most of the time
(3) A good bit of the time
.
.
.
(7) None of the time
c2638 C2638TextureC01 February 8, 2002 15:39
10 INTRODUCTION
Figure 1.2. Timing of observations in a study with an event-driven design with as-
sessments before (B), during (D), and 4 months after (A) therapy. Data are from the
adjuvant breast cancer study [41]. Each row corresponds to a randomized subject.
Subjects randomized to the 16-week regimen appear in the upper half of the gure
and subjects randomized to the CAF regiment are in the lower half of the gure.
Questionnaire completion/missing data
There were 163 eligible patients registered on the HRQoL study. Assessments
were completed by 157 (96%) patients prior to therapy, by 149 (91%) of the pa-
tients during therapy, and by 141 patients (87%) following therapy (Table 1.4).
In an exploratory analysis (Table 1.5), women under 50 years of age were less
likely to complete the assessment prior to therapy. Missing assessments dur-
ing therapy were associated with early discontinuation of therapy (regardless
of reason), concurrent chronic disease, and four or more positive nodes at
diagnosis. Assessments were more likely to be missing following therapy in
women who went o therapy early or who had four or more positive nodes at
diagnosis. The association of missing assessments with factors that are also
associated with poorer outcomes suggests that these assessments may not be
missing by chance. (We will develop this argument in Chapter 4.)
1.4 Example 2: Advanced non-small-cell lung cancer (NSCLC)
To evaluate therapeutic eectiveness of new strategies for NSCLC, a multi-
center phase III clinical trial* was activated to compare two new paclitaxel
cisplatin regimens with a traditional etoposidecisplatin regimen for the
* Eastern Cooperative Oncology Group funded by the National Cancer Institute Grant
CA-23318.
c2638 C2638TextureC01 February 8, 2002 15:39
EXAMPLE 2: ADVANCED NON-SMALL-CELL LUNG CANCER (NSCLC) 11
Table 1.4. Documentation of HRQoL assessments in the adjuvant breast cancer
study (Example 1).
Before During After
therapy therapy therapy
Status/reason N (%) N (%) N (%)
Completed 157 (96) 149 (91) 141 (87)
Too late for baseline
a
4 (2)
No longer receiving therapy
b
2 (1)
Patient refused 1 (1) 1 (1) 5 (3)
Sta oversight 0 (0) 4 (2) 3 (2)
Patient too ill 0 (0) 0 (0) 4 (2)
Other, Early o therapy 0 (0) 7 (4) 5 (3)
Other, Not specied 1 (1) 0 (0) 4 (2)
Not documented 0 (0) 0 (0) 1 (1)
Total expected 163 163 163
a
First assessment occurred after therapy began, excluded from analysis.
b
Second assessment occurred after therapy ended, excluded from analysis.
Table 1.5. Patient characteristics associated with missing assessments (Example 1).
Time of assessment Characteristic
b
Odds ratio (95% CI)
Before therapy Age(<50 vs. 50) (93 vs. 100%)
a
During therapy O therapy early 202 (24, 999)
Concurrent disease 7.9 (1.3, 47)
4+ positive nodes 15.3 (1.4, 166)
Following therapy O therapy early 9.6 (2.8, 34)
4+ positive nodes 6.2 (1.2, 32)
a
Odds ratio could not be estimated because of division by zero. CI =
condence interval.
b
Other potential explanatory variables include minority race, initial per-
formance status, treatment arm, the presence and severity of selected
toxicity including vomiting, diarrhea, alopecia, and weight loss, and early
discontinuation of therapy specically related to toxicity, relapse, or pa-
tient preference.
treatment of stage IIIBIV non-small-cell lung cancer (NSCLC) [7]. In addi-
tion to traditional endpoints, such as time to disease progression and survival,
HRQoL was included in this trial. The stated objective of the HRQoL com-
ponent of this study was to compare quality of life for the three treatment
arms and correlate quality of life to toxicity.
c2638 C2638TextureC01 February 8, 2002 15:39
12 INTRODUCTION
Table 1.6. Sample questions from the Functional Assessment of Cancer Therapy
(FACT)-Lung (Version 2) Questionnaire.
Please indicate how true each statement has been for you during
the past 7 days
Not at A little Some- Quite Very
Additional concerns all bit what a bit much
34. I have been short of breath 0 1 2 3 4
35. I am losing weight 0 1 2 3 4
36. My thinking is clear 0 1 2 3 4
37. I have been coughing 0 1 2 3 4
38. I have been bothered by 0 1 2 3 4
hair loss
39. I have a good appetite 0 1 2 3 4
40. I feel tightness in my chest 0 1 2 3 4
41. Breathing is easy for me 0 1 2 3 4
If you ever smoked,
please answer 42:
42. I regret my smoking 0 1 2 3 4
Treatment
All three treatment arms included cisplatin at the same dose. The traditional
treatment arm contained VP-16 with the cisplatin. The other two arms con-
tained low-dose paclitaxel or high-dose paclitaxel with G-CSF. The planned
length of each cycle was 3 weeks. Treatment was continued until disease pro-
gression or excessive toxicity. In all, 332 patients (58%) started four or more
cycles of therapy and 210 (36%) started six or more cycles.
Quality of life measure
The Functional Assessment of Cancer Therapy, Lung Cancer Version2 (FACT-
L) was used to measure HRQoL in this trial [7]. It is a self-administered
43-item questionnaire, 33 items of which are general questions relevant to all
cancer patients and 10 of which are items relevant to lung cancer (Table 1.6).
It provides reliable scores for physical well-being, social/family well-being,
emotional well-being, functional well-being, and lung cancer symptoms [7,
15]. In addition to the ve major subscales, a number of summary scales can
be constructed, including the FACT-Lung Trial Outcome Index (FACT-Lung
TOI), which is the sum of the physical well-being, functional well-being, and
lung cancer symptom scores. Higher scores imply better quality of life. Details
of scoring the FACT are presented in Chapter 2. For easier interpretation of
the presentations in this book, all subscales were rescaled to have a possible
range of 0 to 100 with higher scores indicating better outcomes.
c2638 C2638TextureC01 February 8, 2002 15:39
EXAMPLE 2: ADVANCED NON-SMALL-CELL LUNG CANCER (NSCLC) 13
Timing of assessments
Four HRQoL assessments were scheduled for each patient. The assessment
times were carefully selected after balancing a number of considerations:
1. Concern for excessive burden to this often debilitated group of patients
2. The relatively short median time to both progressive growth of the tumor
and death expected in these patients
3. Practical matters related to feasibility of the study at multiple clinical
treatment sites.
These issues contributed to the decision to have relatively few assessments
(four) over a fairly brief period of time (6 months from initiation of treatment).
Standard community oncology practice for metastatic lung cancer very of-
ten involves the administration of two cycles of cytotoxic chemotherapy with
reevaluation at the end of the second cycle (6 weeks) before moving on to
further chemotherapy. Because the investigators were particularly interested
in obtaining data from this study that would be useful to the community
practitioner, they selected cycle 3, day 1 as the rst clinically relevant assess-
ment after baseline. The third assessment, which occurred at the end of the
fourth cycle (cycle 5, day 1), was selected because it represented the same
interval of time (6 weeks) and we believed it possible that patients experi-
ence with therapy would enable them to tolerate it better over time. The nal
assessment at 6 months was selected as the best long-term follow-up in this
population because it was several months before the expected median survival,
which ensured that a sucient number of patients could be studied. Thus, the
four assessments were scheduled to be prior to the start of treatment, before
the start of the third and fth courses of chemotherapy, and at 6 months
(week 26). If patients discontinued the protocol therapy, then the second and
third assessments were to be scheduled 6 and 12 weeks after the initiation of
therapy.
The actual timing of the HRQoL assessments showed much more variabil-
ity than the plan would suggest (Figure 1.3). Some of the variation was to
be expected. For example, when courses of therapy were delayed because of
toxicity, the second and third assessments were delayed. Some variation in the
6-month assessment was also expected, as the HRQoL assessment was linked
to clinic visits that might be more loosely scheduled at the convenience of the
patient and medical sta. There was also an allowed window of 2 weeks prior
to the start of therapy for the baseline assessment.
Questionnaire completion/missing data
A total of 1402 HRQoL assessments were obtained in this study. This rep-
resents 94, 72, 60, and 50% of the expected assessments in surviving pa-
tients at each of the four planned assessment times (Table 1.7). The most
commonly documented reasons for missing assessments are patient refusal
(8%) and sta oversight (8%). A detailed listing is presented in Table 1.7.
A more extensive exploration of the missing data will follow in Chapter 4.
c2638 C2638TextureC01 February 8, 2002 15:39
14 INTRODUCTION
Table 1.7. Documentation of FACT-Lung assessments (Example 2).
Baseline 6 weeks 12 weeks 6 months
Status/reason missing N (%) N (%) N (%) N (%)
Completed
a
538 (94) 382 (72) 290 (60) 192 (50)
Refusal 9 (2) 40 (8) 53 (11) 69 (18)
Language 0 (0) 2 (0) 0 (0) 0 (0)
Sta unavailable 0 (0) 5 (1) 2 (0) 1 (0)
Sta oversight 9 (2) 35 (7) 66 (14) 57 (15)
Patient feels too ill 0 (0) 4 (1) 3 (1) 3 (1)
Sta felt patient too ill 1 (0) 19 (4) 15 (3) 11 (3)
Other 9 (2) 12 (2) 23 (5) 19 (5)
Unknown 9 (2) 30 (6) 28 (6) 32 (8)
Total expected 575 529 480 384
Patients expired
b
0 (0) 46 (8) 95 (17) 191 (33)
Total patients 575 575 575 575
a
Percent of expected assessments (e.g., patient still alive).
b
Percent of all possible assessments.
Figure 1.3. Timing of observations in a time-driven design with four planned as-
sessments. Data are from the NSCLC study (Example 2). Each row corresponds
to a randomized patient in this study. Symbols represent the actual timing of the
HRQoL assessments relative to the date the patient was randomized.
c2638 C2638TextureC01 February 8, 2002 15:39
EXAMPLE 3: RENAL CELL CARCINOMA TRIAL 15
Both exploration and sensitivity analysis will suggest that dropout in this
study was likely to be nonrandom [42]. Consequently, methods of analysis
described in Chapters 7 through 9 must be considered for this study.
1.5 Example 3: Renal cell carcinoma trial
A multicenter randomized phase III trial was conducted to determine whether
combination therapy with 13-cis-retinoic acid (13-CRA) plus interferon alpha-
2a (IFN2a) was superior therapy to IFN2a alone in patients with advanced
renal cell carcinoma.* The inuence of cytokine treatment on quality of life
was considered to be an important aspect of the management of advanced
renal cell carcinoma. No dierence was found for the clinical endpoints [104].
This population experiences considerable morbidity and mortality, with a me-
dian survival of approximately 15 months.
Patient selection
Between April 1994 and July 1996, 284 patients were entered into the random-
ized trial. Eligibility requirements included histological conrmation of renal
cell carcinoma, Karnofsky performance status (KPS) 70, and estimated life
expectancy of more than 3 months. Patients enrolled after February 1995 were
asked to complete HRQoL assessments.
Treatment
Patients were randomized to receive either daily treatment with IFN2a plus
13-CRA or IFN2a alone. IFN2a was given as a single subcutaneous injec-
tion and the 13-CRA was given orally. Doses were adjusted in response to
symptoms of toxicity. Treatment was continued until progression of disease,
complete response, or development of excessive toxicity.
Quality of life measure
The trial utilized the Functional Assessment of Cancer Therapy Version 3
(FACT-G) as the tool for HRQoL assessment. Since the eect of biologic res-
ponse modiers (BRM) has not been adequately studied, 17 disease-specic
questions were appended to the core questionnaire to measure the expected
symptoms of 13-CRA and IFN2a (Table 1.8). Validation of the disease-
specic questions was undertaken to assess whether these questions are a
reliable measure of HRQoL. Using factor analysis and internal consistency
coecients on the 2-week and baseline assessments, the original pool of 17
questions was reduced to a 13-item measure [2]. The factor analysis suggested
that two denable dimensions were being measured by this set of questions.
Examination of item content of the rst dimension or subscale (factor 1) sug-
gests it could be labeled as a BRM-Physical component. The second factor
*
Memorial Sloan-Kettering Cancer Center and Eastern Cooperative Oncology Group
funded by the National Cancer Institute Grant CA-05826.
c2638 C2638TextureC01 February 8, 2002 15:39
16 INTRODUCTION
Table 1.8. Sample questions from the Functional Assessment of Cancer Therapy
Biologic Response Modiers including CRA (FACT-BRM/CRA Version 3) ques-
tionnaire.
Please indicate how true each statement has been for you during
the past 7 days
Not at A little Some- Quite Very
Additional concerns all bit what a bit much
35. I get tired easily 0 1 2 3 4
36. I feel weak all over 0 1 2 3 4
37. I have trouble concentrating 0 1 2 3 4
38. I have trouble remembering 0 1 2 3 4
things
39. My thinking is clear 0 1 2 3 4
40. I have a good appetite 0 1 2 3 4
41. I have pain in my joints 0 1 2 3 4
42. I am bothered by the chills 0 1 2 3 4
43. I am bothered by fevers 0 1 2 3 4
44. I am bothered by dry skin 0 1 2 3 4
45. I am bothered by dry mouth 0 1 2 3 4
46. I am bothered by dry eyes 0 1 2 3 4
47. I get depressed easily 0 1 2 3 4
48. I get annoyed easily 0 1 2 3 4
49. I have emotional ups and downs 0 1 2 3 4
50. I feel motivated to do things 0 1 2 3 4
51. I am bothered by sweating 0 1 2 3 4
included items related to mental or emotional symptoms/side eects (BRM-
Mental). The initial BRM-Physical subscale included three items that were
specic to dryness (skin, mouth, eyes), which were specic to the eects of
CRA. This subscale was subsequently modied to reect a more generic class
of biologic responsemodifying therapies. Coecients of reliability (internal
consistency) range from 0.61 to 0.92 for the baseline assessments and from
0.64 to 0.94 for the week 2 assessments. The Trial Outcome Index (FACT-
BRM TOI), calculated by adding the physical and functional well-being scores
and the two BRM subscale scores, was used as a summary measure of HRQoL.
Details for scoring the FACT scales are presented in Chapter 2. In the exam-
ples presented in this book, the original 0- to 108-point scale was rescaled to
have a 0- to 100-point range.
Timing of HRQoL assessments
At the time the study was designed, knowledge about the impact of therapy
on the HRQoL of patients with renal cell carcinoma was extremely limited. Six
assessments were scheduled. The rst was to be obtained in a 2-week period
c2638 C2638TextureC01 February 8, 2002 15:39
EXAMPLE 3: RENAL CELL CARCINOMA TRIAL 17
Table 1.9. Renal cell carcinoma trial: Summary of missing HRQoL assessments over
time among 213 patients with at least one HRQoL assessment; 17 patients did not
complete any assessments.
Schedule (weeks) 0 2 8 17 34 52
Surviving patients 213 212 201 176 142 118
TOI completed 190 163 123 75 55 27
% of surviving patients 89 77 61 43 39 23
% of total patients 89 77 57 35 26 13
Windows dened as <7 days, 1 to <6 weeks, 6 to <15 weeks, 15 to
<30 weeks, 30 to <48 weeks, 48 to 65 weeks.
Figure 1.4. Timing of observations in a time-driven design with increasing numbers
of mistimed observations. Data are from the renal cell carcinoma study [3]. Each
row represents a subject.
prior to initiating therapy. The second was scheduled 2 weeks after initiation
of therapy to assess acute toxicity. The remaining assessments were to occur
at 2, 4, 8, and 12 months (8, 17, 34, and 52 weeks) to assess chronic toxicity.
This schedule was to be continued for patients who discontinued the protocol
therapy as a result of toxicity or disease progression to assess the continued
impact of the disease and toxicity. The exact timing of the assessments became
more variable with time (Figure 1.4).
Questionnaire completion/missing data
At least one questionnaire was received from 213 patients, with a total of
735 questionnaires. The compliance rates among the 213 patients for the
six assessments were 89, 77, 61, 43, 39, and 23%, respectively (Table 1.9).
c2638 C2638TextureC01 February 8, 2002 15:39
18 INTRODUCTION
Early dropout from the HRQoL study was characterized by poorer prognosis
(Kendalls
b
= 0.25), lower initial FACT-Lung TOI scores (
b
= 0.27), lower
nal FACT-Lung TOI score (
b
= 0.22), and early death (
b
= 0.48), but not
with treatment (
b
= 0.01). A more extensive exploration of the missing
data will follow in Chapter 4. As in the NSCLC study, this suggests that
dropout in this study was likely to be nonrandom. Consequently, methods of
analysis described in Chapters 7 through 9 must be considered for this study.
1.6 Summary
Health-related quality of life
1. HRQoL represents the impact of disease and treatment on a patients
daily living; it does not represent aspiration or satisfaction.
2. HRQoL is multidimensional.
3. HRQoL is subjective.
Example 1: Adjuvant breast cancer study
1. Simple design with one assessment before, during, and after therapy.
2. Minimal (<5%) missing data.
Example 2: Advanced non-small-cell lung cancer study
1. Four planned assessments: one assessment before therapy and three dur-
ing or after therapy.
2. Extensive missing data as a result of death and other reasons probably
related to HRQoL. More than 50% missing at time of last assessment.
Example 3: Advanced renal cell carcinoma study
1. Six planned assessments: one assessment before therapy and ve during
or after therapy.
2. Extensive missing data as a result of both death and other reasons prob-
ably related to HRQoL. More than 50% missing at time of last assess-
ment.
3. Timing of assessments became more variable as the study progressed.
c2638 C2638Ch02 February 12, 2002 11:35
CHAPTER 2
Study Design and Protocol
Development
2.1 Introduction
Implicit in the use of measures of HRQoL, in clinical trials and in eectiveness
research, is the concept that clinical interventions such as pharmacologic thera-
pies, can aect parameters such as physical function, social function, or mental
health. (Wilson and Cleary, 1995 [163])
Development of a priori objectives and an analytic plan are essential to both
good trial design and subsequent scientic review. This chapter will focus on
design and protocol development, with the exception of the statistical consid-
erations such as sample size and analysis plans (see Chapter 12). Table 2.1 is a
checklist for protocol development. All principles of good clinical trial method-
ology are applicable [49, 53, 99, 113, 141], but there are additional require-
ments specic to HRQoL. These include selection of an appropriate measure
of HRQoL and the conduct of the assessment to minimize any bias. Because of
the multidimensional nature of HRQoL and repeated assessments over time,
objectives need to be explicitly specied and a strategy developed for handling
multiple endpoints.
2.2 Background and rationale
Providing sucient background and rationale to justify the resources required
for an investigation of HRQoL will contribute to the success of the investiga-
tion. The rationale should provide answers to such questions as: How exactly
might the results aect the clinical management of patients? How will the
HRQoL results be used when determining the eectiveness of the treatment
arms? The justication should include a motivation for the particular aspects
of HRQoL that will be measured in the trial (e.g., physical, functional, or
emotional) as they relate to the disease and its treatment.
The same demands for a rationale should be applied to all information
collected in any clinical trial. However, investigators already have a habit
of collecting laboratory and radiological tests so that minimal motivation is
required to implement this data collection. The same is not necessarily true
for the collection of data that requires patient self-assessment, and greater
motivation may be necessary. Writing a rationale for HRQoL assessments
also facilitates the development of well-dened research objectives.
c2638 C2638Ch02 February 12, 2002 11:35
20 STUDY DESIGN AND PROTOCOL DEVELOPMENT
Table 2.1. Checklist for protocol development.
Rationale for studying HRQoL
Explicit research objectives
Strategies for limiting the exclusion of subjects from the HRQoL study
Rationale for timing of assessments and o-study rules
Rationale for instrument selection
Details for conducting HRQoL assessments that minimize bias and missing
data
Analytic plan (Chapter 12)
Table 2.2. Checklist for developing well-dened objectives.
Which dimensions of HRQoL are relevant?
What is the population of interest (inference)?
What is the time frame of interest?
Is the intent conrmatory or exploratory?
Is the intent to demonstrate superiority or equivalence?
2.3 Research objectives
The most critical component of any clinical trial is a list of explicit research
objectives; Table 2.2 is a checklist. It is only when the research objectives
provide sucient detail to guide the design, conduct, and analysis of the study
that the success of the trial is ensured. If no further thought is given to the
specic research questions, it is likely that there will be neither rhyme nor
reason to the design, conduct, and analysis of the HRQoL assessments. Then,
because of either poor design or the lack of a denite question, the analyst
will ask questions such as: What is the question? Why did you collect the data
this way? Why didnt you collect data at this time? There is a real concern
that the lack of meaningful results from poorly focused and thus potentially
poorly designed and analyzed trials will create the impression that HRQoL
research is not worthwhile.
General objectives state the obvious. Regardless of whether HRQoL is con-
sidered to be a secondary or primary outcome, a stated HRQoL objective such
as To compare the quality of life between treatment A and B is too nonspecic.
When the outcome is a univariate measure such as survival, it is reasonable to
state the objective simply. Unlike the event of death, which occurs at a single
point in time, HRQoL is experienced over time. The assessment of HRQoL
also involves the various aspects of HRQoL, including physical, functional,
emotional, and social well-being. A vague objective is insucient.
c2638 C2638Ch02 February 12, 2002 11:35
RESEARCH OBJECTIVES 21
Further, the detail of the objectives includes identifying (1) relevant dimen-
sions of HRQoL, (2) population of interest, and (3) the time frame. Hypotheses
should be clear whether the intent is to demonstrate superiority or equivalence
and whether the objective is conrmatory or entirely exploratory.
Consider a hypothetical example in which two analgesics (Pain-Free and Re-
lief) are compared for a painful condition. Both drugs require a period of titra-
tion in which the dose is slowly increased to a maintenance level. Assessments
are obtained during the titration phase and the maintenance phase of the trial.
An objective such as To compare the quality of life in subjects receiving Pain-
Free vs. Relief does not provide any guidance. If the condition is permanent
and the drugs are intended for extended use, the objective could be restated
as To compare the quality of life while on a maintenance dose of Pain-Free vs.
Relief. Alternatively, if the pain associated with the condition was brief, the
drug that provided earlier relief would be preferable and the objective could
be stated as To compare the time to 20% improvement in quality of life
with Pain-Free vs. Relief. These objectives now provide more guidance. In the
rst situation, assessments during maintenance are important, and in the sec-
ond the earlier assessments are the basis of dening the outcome and should
be frequent enough to detect dierences between the two regimens. The same
logic should be applied to the dierent dimensions of HRQoL.
Domains of interest
If HRQoL is dened as a primary endpoint, it implies that the demonstration
of ecacy of intervention is based on some or all of the dimensions of HRQoL
and other primary endpoints. Secondary endpoints are supportive of primary
endpoints. In this case, HRQoL may measure a more distant outcome of ther-
apy. For example, in the treatment of anemia, the primary endpoints are the
expected biological changes (increased hemoglobin levels) and the secondary
endpoints are consequences of the physiological changes (reduction in trans-
fusions, less fatigue, improved HRQoL). Finally, when HRQoL is considered
to be a tertiary outcome, the analysis is intended to be descriptive. Tertiary
outcomes often address safety rather than ecacy measures. In HRQoL, ter-
tiary measures are typically subscales that are not expected to be aected by
the intervention.
Pragmatic vs. explanatory inference
The objectives should clarify to whom the inferences will apply: all study sub-
jects, subjects who complete therapy, or only subjects alive at the end of the
study? A useful terminology for characterizing the trial objectives is described
by Schwartz et al. [63, 136]. They distinguish between pragmatic and explana-
tory investigational aims. The investigational aims for HRQoL will typically
be pragmatic rather than explanatory. The distinction becomes important
when the investigators make decisions about the study design and analysis.
c2638 C2638Ch02 February 12, 2002 11:35
22 STUDY DESIGN AND PROTOCOL DEVELOPMENT
Pragmatic objectives are closely related to the intent-to-treat analytic strate-
gies. In these trials, the objective is to compare the relative impact of each
intervention on HRQoL under practical and realistic conditions. For example,
it may not be desirable for all subjects to complete the entire course of in-
tended therapy exactly as specied because of toxicity or lack of ecacy. Some
individuals may need to be given additional therapy. Others may be noncom-
pliant for a variety of reasons. In a study with a pragmatic aim, the intent
is to make inferences about all subjects for the entire period of assessment.
Thus, assessment of HRQoL should be continued regardless of whether the pa-
tient continues to receive the intervention as specied in the protocol. HRQoL
assessments should not stop when the patient goes o study.
In contrast, with an explanatory aim, we compare the HRQoL impact of
treatments given in a manner that is carefully specied in the protocol. This
is sometimes described as the analysis of the per-protocol subgroup. In this
setting, HRQoL assessments may be discontinued when subjects are no longer
compliant with the treatment plan. The analysis of these studies appears
simple on the surface, but there is a real chance of selection bias. The analyst
can no longer rely on the principle of randomization where unmeasured patient
characteristics that confound the results are unbalanced across the treatment
arms.
2.4 Selection of subjects
Ideally, the subjects who are involved in the HRQoL assessments are the same
subjects who are evaluated for all of the study endpoints. This is recommended
for several reasons [59]. The rst is practical. It is much easier to implement
a study if all patients require the same assessments. The second is scientic.
The credibility and interpretation of a study is increased when all subjects
are evaluable for all endpoints.
In practice, there may be some limitations. Physical, cognitive, or language
barriers may make self-assessment infeasible. If the exclusions constitute a
small proportion of the subjects, it is unlikely to compromise the validity of
the results. However, if the investigation involves a substantial proportion of
subjects either initially or eventually unable to complete self-assessments, a
strategy must be developed to deal with the issue. For example, in a study
where progressive cognitive impairment is expected, a design option is to
start with concurrent subject and proxy assessments, continuing the proxy
assessments when the patient is no longer able to provide assessments. The
value of this design will be discussed in later chapters.
In some cases, the number of subjects required to detect a clinically mean-
ingful dierence in the HRQoL endpoints is substantially smaller than that
required for the other endpoints. If the resource savings are substantial enough
to warrant the logistical diculties of obtaining HRQoL assessment on a sub-
set of patients, the selection must be done in a completely random manner.
It is not acceptable to allow selection of the subjects by the researchers or
c2638 C2638Ch02 February 12, 2002 11:35
LONGITUDINAL DESIGNS 23
even by the patients accessibility to telephone interview, as these methods
are likely to introduce a selection bias.
Care should be taken when dening which subjects are excluded from an
analysis. It is not uncommon to read a protocol where the analysis is limited
to subjects with at least two HRQoL assessments (baseline and one follow-
up). This criterion may change the population to which the results can be
generalized by excluding all patients who drop out of the trial early. If this
is a very small proportion of patients, it will not matter. But if a substantial
number of subjects on one or more arms of the trial drop out before the
second HRQoL assessment, this rule could have a substantial impact on the
results.
2.5 Longitudinal designs
Longitudinal data arise in most HRQoL investigations because we are inter-
ested in how a disease or an intervention aects an individuals well-being
over time. The optimal timing of HRQoL assessments depends on the disease,
its treatment, and the scientic question. The number and timing of HRQoL
assessments are inuenced by both the objectives and practical considerations
such as when the investigators have access to the subjects. Some studies have
event- or condition-driven designs. Other studies have periodic evaluations
based on the time elapsed since the beginning of treatment or the onset of
the disease. Some studies incorporate a mixture of these two designs. The
timing of assessments inuence how we approach the analysis of each study
(see Chapter 3).
Event- or condition-driven designs
When the objective of the study is to compare HRQoL in subjects experi-
encing the same condition, assessments are planned to occur at the time of
clinically relevant events or to correspond to specic phases of the interven-
tion. These designs are more common when the intervention or therapy is
of limited duration. For example, one might consider a simple design with
a pre- and postintervention assessment when comparing the HRQoL after
two surgical procedures. The adjuvant breast cancer study (Example 1) pro-
vides a second example in which therapy is of limited duration. The design
includes one assessment to measure HRQoL during each of the three phases:
prior to, during, and after therapy. Many variations are possible. For example,
there may be reason to believe that dierences in HRQoL may exist during
only the early or later periods of therapy. In this case, the design would in-
clude a minimum of two assessments during therapy. Studies with a limited
number of phases of interest in the scientic investigation, where HRQoL is
expected to be constant during those phases, are amenable to an event-driven
design.
c2638 C2638Ch02 February 12, 2002 11:35
24 STUDY DESIGN AND PROTOCOL DEVELOPMENT
Time-driven designs
When the scientic questions involve a more extended period of time or the
phases of the disease or its treatment are not distinct, the longitudinal de-
signs are based on time. These designs are appropriate for chronic conditions
where therapies are given over extended periods of time. In the NSCLC study
(Example 2) and the renal cell carcinoma study (Example 3), therapy was to
be given until there was evidence that it was ineective or that it occasioned
unacceptable toxicity. Thus, the duration of therapy was indeterminate at the
onset of the study. More obvious examples would include treatment of chronic
conditions such as diabetes and arthritis. Timing in these designs is based on
regularly spaced intervals (e.g., every 3 months), sometimes more frequent
initially (e.g., every month for 3 months and then every 3 months).
Timing of the initial HRQoL assessment
It is critical that the initial assessment occur prior to randomization. Because
the measurement of HRQoL is generally based on self-evaluation, there is a
potential that the knowledge of treatment assignment will inuence a sub-
jects responses [9]. This is especially true when the patient is aware that one
of the interventions is new, possibly more eective than standard therapy, and
exciting to his or her physician. The possible exception occurs if the interven-
tion is double-blinded and treatment starts within a reasonable period of time
of randomization. In this case, it is allowable to obtain initial assessment prior
to the beginning of treatment but after randomization.
Timing of the follow-up HRQoL assessments
Similar attention should be paid to the timing of follow-up assessments. As-
sessments should be made consistently across treatment arms. Attention
should be paid to the timing of diagnostic procedures. Especially with life-
threatening diseases, the choices are not particularly easy. Prior to the testing,
patients are likely to be experiencing stress in anticipation of the yet unknown
results. After the procedure, the patients will be experiencing either great re-
lief or anxiety.
The period of accurate recall is between 1 and 4 weeks, with better recall
of major events and more recent experiences. Schipper [131] notes that the
side eects of chemotherapy in patients with cancer may have a less adverse
eect on a patients HRQoL than similar side eects attributable to the dis-
ease. This observation may be true in other disease conditions as well. Testing
immediately after toxicity occurs will emphasize that experience and deem-
phasize the benets of treatment and disease symptoms. It is important not
to pick a particular timing that will automatically bias the results against
one treatment arm. In studies where the timing and length of treatment dier
across arms, this may be challenging if not impossible.
c2638 C2638Ch02 February 12, 2002 11:35
LONGITUDINAL DESIGNS 25
Timing of HRQoL assessments when therapy is cyclic
Treatment for some conditions is cyclic, with administration of therapy only
during the initial part of the cycle. This is characteristic of most chemother-
apy regimens for the treatment of cancer. Typically, intense and toxic therapy
is given for 1 to 2 weeks, with a hiatus for another 1 or 2 weeks to allow for
recovery from the toxic eects of the treatment. Thus, it may not be possible
to identify a time when the HRQoL is typical. For very practical reasons,
HRQoL assessments have traditionally been scheduled to occur just prior to
the beginning of the next cycle. Patients return to the clinic at this time
for laboratory tests and other clinical evaluations; thus, they are consistently
available for HRQoL assessment. As a result, there is potential for the mea-
surement obtained just prior to the next cycle to overestimate the HRQoL
over the entire cycle. More frequent assessment would solve this problem, but
it must be weighed against increased patient burden and logistical issues of
obtaining HRQoL assessments between patient visits.
Trials with dierent schedules of therapy
Another example is the timing of assessments when the impact of the ther-
apy may dier by treatment arm. The timing of HRQoL assessments may,
either by design or by chance, favor one of the treatment arms. The adjuvant
breast cancer study (Example 1) provides a good example of how dierent
treatment schedules can make the timing of HRQoL assessment a dicult
choice. In this study, patients on one of the regimens receive some type of
therapy on a weekly basis. On the other arm, therapy is cyclic with the pa-
tients receiving treatment during the rst 2 weeks and then being free from
treatment for the next 2 weeks. The HRQoL tool used in this trial asked
the patient to respond relative to the last 2 weeks. The HRQoL assessments
were scheduled to occur just prior to the beginning of a new cycle of therapy.
Thus, in one arm the period of evaluation included the time when the patient
was receiving therapy and in the other arm it did not. One possible solution
might have been to change the time of the HRQoL assessment to the middle of
the CAF treatment cycle, but this would have required an additional patient
visit.
Frequency of evaluations
The frequency of the assessments should correspond appropriately to the natu-
ral history of the disease and the likelihood of changes in HRQoL within that
period. Other considerations are a practical follow-up schedule and the tim-
ing of therapeutic and diagnostic interventions (Table 2.3). The assessments
should be frequent enough to capture meaningful change but not so frequent
as to be burdensome to the patient or incongruent with the assessment tool.
If rapid change is expected during the early part of the study, more frequent
early assessments are needed. In the renal cell carcinoma study (Example 3),
c2638 C2638Ch02 February 12, 2002 11:35
26 STUDY DESIGN AND PROTOCOL DEVELOPMENT
Table 2.3. Timing of HRQoL assessments relative to clinical and laboratory
assessments.
Design 1: HRQoL less frequent
Clinical/Labs X X X X X
HRQoL Q Q Q
Design 2: HRQoL linked with clinical
Clinical X X X X X
Labs L L L L L L L L L
HRQoL Q Q Q Q Q
Design 3: HRQoL linked with labs
Clinical X X X X X
Labs L L L L L L L L L
HRQoL Q Q Q Q Q Q Q Q Q
Design 4: HRQoL more frequent
Clinical/Labs X X X X X
HRQoL Q Q Q Q Q Q Q Q Q
assessments were planned more frequently during the early phase of the study
than toward the end. However, in retrospect, even more assessments during
the early phase of therapy would have been informative, as HRQoL changed
very rapidly during the early weeks of the treatment.
Assessments should not be more frequent than the period of recall dened
for the instrument. The quality of the patients life does not generally change
on an hourly or daily basis as one would expect for symptoms. Thus, HRQoL
scales often request the subjects to base their evaluations on the last 7 days
or 4 weeks. Scales specic to diseases or treatments where there can be rapid
changes generally have a shorter recall duration. Thus, if the HRQoL instru-
ment is based on recall over the previous month, assessments should not be
weekly or daily.
Duration of HRQoL assessment
First, it is important that the assessment is of sucient duration to observe
changes in HRQoL. Physiological responses to a therapy may occur more
rapidly than changes in HRQoL. This is especially true for chronic diseases
that have associated physical or functional disabilities.
For practical reasons it is wise to dene a specic limit to the duration
of assessment, specically avoiding statements such as and every 6 months
thereafter. This is another situation that illustrates the need to have a well-
dened objective. For example, are the investigators interested in the HRQoL
of subjects while on a therapy that is of limited duration or is it especially
c2638 C2638Ch02 February 12, 2002 11:35
SELECTION OF A QUALITY OF LIFE MEASURE 27
important to understand the long-term impact of the treatment on HRQoL?
If the former is the case, the additional information that can be obtained from
continued assessments may diminish after some point either because there is
little or no change or because the number of subjects with assessments is too
small to analyze. If none of the objectives of the study requires continued
assessment, then follow-up beyond that point is unwarranted.
Assessment after discontinuation of therapy
A very clear policy needs to be developed for following patients who cannot
follow the treatment protocol. There are two major considerations when devel-
oping this policy. The rst is scientic and depends on the research question.
For example, if the discontinuation of treatment limits any future therapy
to more intensive and toxic treatments or eliminates treatment options alto-
gether as the disease progresses, the failure to continue HRQoL assessment
can lead to overoptimistic bias. A treatment arm with a high rate of dropout
may appear articially benecial because only the healthiest of the patients
remain on the treatment.
On the other hand, discontinuation of assessment may make scientic sense
in other disease settings. If HRQoL assessment is continued, the o-therapy
assessments can be excluded if later deemed uninformative with respect to
the research question. The opposite is not true; one can never retrospecti-
vely obtain the o-therapy assessments at a later date if they are determined
to be of interest. The second aspect is practical; the o-treatment assessments
are often dicult to obtain, especially if the patient no longer remains under
the care of the same physician.
2.6 Selection of a quality of life measure
Guyatt et al. [62] dene an instrument to include the questionnaire, the
method of administration, instructions for administration, the method of scor-
ing, and analysis and interpretation for a health status measure. All these
aspects are important when evaluating the suitability of an instrument for a
clinical trial; Table 2.4 is a checklist. Many of the steps in the selection of
the HRQoL instruments for a trial can be accomplished by examining the
literature and the questions posed in the instrument. However, empirical data
are also valuable, especially with diseases or treatments that have not been
previously studied. Bouchet et al. [8] describe a pilot study in which they
evaluated three possible measures for a primary prevention trial. Included in
the evaluation were ceiling eects, convergent validity, known discrimination,
reproducibility among individuals reporting no change, and responsiveness
among individuals reporting change.
Selecting a previously validated instrument is vastly preferable to the use
of a new, unvalidated instrument. Developing a new instrument is expensive,
is time consuming, and should be undertaken only when existing instruments
c2638 C2638Ch02 February 12, 2002 11:35
28 STUDY DESIGN AND PROTOCOL DEVELOPMENT
Table 2.4. Checklist for the selection of HRQoL instruments for a clinical trial.
Does the instrument measure what it proposes to measure?
Is the information relevant to the research question?
How well does the instrument cover the important aspects of what is to
be measured?
Is a generic or disease-specic instrument more appropriate?
Heath prole (rating scale) vs. patient preference (utility)?
Will the instrument discriminate among subjects in the study and will it
detect change?
How well does the instrument predict related outcomes?
Are the questions appropriate for the subjects across time?
Are the format and mode of administration appropriate to the subjects and
the trial?
Has the instrument been previously validated in the target or similar pop-
ulation? If not, what are the plans to do so for current study?
If using new instrument or items, the rationale and reasons for new items
are indispensable.
are clearly unsuitable. Clinical researchers who have never undertaken this
task severely underestimate the time and eort required. Instrument develop-
ment is not limited to the generation of a list of questions and use in a single
clinical trial. Development involves numerous steps of item generation (ex-
pert panels, focus groups), cognitive interviews evaluating comprehension of
questionnaire wording (pilot tests with probing questions), data collection in
a wide range of patients receiving the entire range of treatment options, item
reduction, validation studies, translation, and cultural adaptation. Develop-
ment of a new instrument can easily require 3 years and often more. The
nal limitation is that new instruments or questions may not be accepted as
primary HRQoL endpoints in regulatory settings.
Trial objectives
Ware et al. [156] suggests:
When searching for measures of health status, one rst needs a clear understand-
ing of the reasons for studying health status. A second requirement is a clear
statement of the aspect of health being studied.
It is important to be clear on which of the dierent dimensions (physical,
cognitive, emotional, and social) are relevant to the specic research questions
and verify that all relevant domains are included. If the objective is to provide
data for economic evaluations, a utility measure will facilitate the calculation
of quality-adjusted life years. If the objective is to compare the impact of
c2638 C2638Ch02 February 12, 2002 11:35
SELECTION OF A QUALITY OF LIFE MEASURE 29
two therapies for the same condition on specic domains of disease-specic
HRQoL, a disease-specic instrument is the most suitable.
Validity and reliability
There are numerous aspects and terms for the various components of reli-
ability and validity of an instrument. For a formal presentation, the reader
is referred to one of the numerous textbooks presenting these denitions. A
partial list specic to HRQoL includes Streiner and Norman [147], McDowell
and Newell [95, chapters 1 and 2], Juniper et al. [in 142, chapter 6], Naughton
et al. [in 142, chapter 15], Frank-Stromborg and Olsen [52], Staquet et al. [146,
part IV], and Fayers and Machin [45, chapters 3 through 7].
The validity of a measure in a particular setting is the most important and
the most dicult aspect to establish. This is primarily because there are no
gold standards against which the empirical measures of validity can be com-
pared. In establishing the validity of measures of HRQoL, we compare the
measures to other potentially awed measures of HRQoL. Nonetheless, we
can learn a good deal about an instrument by examining the instrument itself
and the empirical information that has been collected.
The rst questions about the content of an instrumentDoes the instru-
ment measure what it proposes to measure? Are the questions comprehensi-
ble and without ambiguity?are referred to as face validity. Note that across
instruments the same label or descriptor is sometimes given to groups of ques-
tions that measure dierent constructs. The wording of the questions should
be examined to establish whether the content of the questions is relevant to
the population of subjects and the research question. Although expert opin-
ion (physicians, nurses) may make this evaluation, it is wise to check with
patients, as they may have a dierent perspective.
Criterion validity is the strength of a relationship between the scale and a
gold-standard measure of the same construct. As there is no gold standard
for the dimensions of quality of life, we rely on the demonstration of construct
validity. This is the evidence that the instrument behaves as expected and
shows similar relationships (convergent validity) and the lack of relationships
(divergent validity) with other reliable measures for related and unrelated
characteristics.
The next questionWould a subject give the same response at another
time if he was experiencing the same HRQoL?is referred to as reliability.
If there is a lot of variation (noise) in responses for subjects experiencing the
same level of HRQoL, then it is dicult to discriminate between subjects
who are experiencing dierent levels of HRQoL or change in HRQoL over
time. Finally, we ask: Does the instrument discriminate among subjects who
are experiencing dierent levels of HRQoL? Is the instrument sensitive to
changes that are considered important to the patient? These characteristics
are referred to as discriminant validity and responsiveness.
Although all the above characteristics are necessary, responsiveness is the
most important in a clinical trial, as it directly aects the ability to detect
changes that occur as the result of an intervention. One of the obvious factors
c2638 C2638Ch02 February 12, 2002 11:35
30 STUDY DESIGN AND PROTOCOL DEVELOPMENT
that can aect responsiveness is the oor and ceiling eect. If responses are
clustered at either end of the scale, it may not be possible to detect change
due to the intervention.
Appropriateness
Will the instrument discriminate among subjects in the study and will it
detect change in the target population? Ware et al. [156] suggest two general
principles:
1. When studying general populations, consider using positively dened mea-
sures. Only some 15% of general population samples will have chronic phys-
ical limitations and some 10 to 20% will have substantial psychiatric im-
pairment. Relying on negative denitions of health tells little or nothing
about the health of the remaining 70 to 80% of the general population.
2. By contrast, whenstudying severely ill populations, the best strategy may be
to emphasize measures of the negative end of the health status continuum.
Are the questions appropriate and unambiguous for subjects? One cannot
always assume that a questionnaire that works well in one setting will work
well in all settings. For example, questions about the ability to perform the
tasks of daily living, which make sense to individuals who are living in their
own homes, may be confusing when administered to a patient who has been in
the hospital for the past week. Similarly, questions about the amount of time
spent in bed provide excellent discrimination for nonhospitalized subjects but
not for a hospitalized patient.
Are the questions appropriate for the subjects across time? In cases where
the population is experiencing very dierent HRQoL over the length of the
study, very careful attention must be paid to the selection of the instrument
or instruments. Some studies will require dicult choices between the abil-
ity to discriminate among subjects during dierent phases of their diseases
and treatments. In the adjuvant breast cancer study, the subjects were free
of any symptoms or detectable disease. At the time of the pre- and post-
treatment assessments, they were much like the general population. During
therapy, they were likely to be feeling ill from the side eects of the treat-
ment. In the example, at the time that the study was planned there were
very few choices of HRQoL instruments. The compromise was the selection of
the Breast Chemotherapy Questionnaire (BCQ), which was very sensitive on
chemotherapy side eects but may have been insensitive to any post-treatment
dierences.
If an international trial, has the instrument been validated in other lan-
guages and cultures? Simple backward and forward translations are unlikely
to be adequate for HRQoL instruments. There are numerous examples where
investigators have found problems with certain questions as questionnaires
were validated in dierent languages and cultures. Cognitive testing with sub-
jects describing verbally what they are thinking as they form their responses
has been very valuable when adapting a questionnaire to a new language or
culture. This should be followed by formal validation studies.
c2638 C2638Ch02 February 12, 2002 11:35
CONDUCT 31
2.7 Conduct
In many clinical trials, the decision to include HRQoL assessments was made
at the end of the planning phase. Often, a questionnaire was added to the data
collection without any appreciation of the amount of sta time required and
with no allocation of additional resources. This generally resulted in overly
ambitious assessment schedules and large amounts of missing data, which
made analysis dicult and any results open to criticism. Although this be-
havior has decreased over time, there are still too many trials in which the
details of how the HRQoL assessments are to be obtained are missing from
the protocol and training materials.
Hopwood et al. [71] surveyed 29 centers participating in one or more of
three randomized trials for lung and head and neck cancer. They observed
that there was a very high proportion of preventable missing data. The three
most commonly reported problems were that sta were not available, ques-
tionnaires were not available, and the sta considered the patient to be too ill.
Preplanning and budgeting have the potential to x the rst two problems.
Education is needed to address the third.
Mode of administration
Methods of administration that have been used successfully in clinical trials
include paper-and-pencil self-report, in-clinic face-to-face assessments with
trained interviewers, and centralized telephone administration with trained
interviewers. Paper-and-pencil self-report is the most economical but requires
that the patients be available on a regular basis. Interviewer administration
is useful when the population has low literacy (children, immigrant popula-
tions), physical diculty exists with pen-and-pencil forms (advanced neuro-
logical conditions, hospice patients), or the questionnaire involves complex
skip patterns (standard gamble assessments).
Schipper [131] cautions against administration of an instrument designed
for self-administration by a third party, noting several issues. There is evi-
dence that self-report data dier from interview-generated data in ways that
cannot be predicted. This is particularly true when there is the potential for
the inuence of social desirability on the responses. Great care must be taken
to ensure that patients do not feel the need to please the clinical investiga-
tor with their answers. Similarly, patients may feel reluctant to answer the
questions honestly in the presence of family and friends. Answering the ques-
tionnaire at home may result in dierent responses than in a hospital/clinic
environment [162].
Regardless of what mode of administration is selected, it is preferable to
use the same mode throughout the study unless it has been shown that re-
sponses are not aected by mode of administration. This advice is balanced
by considerations of greater bias that can be introduced if this policy results
in more nonresponse among selected groups of patients (e.g., those who are
sicker). Often a compromise is required to balance feasibility and resources
c2638 C2638Ch02 February 12, 2002 11:35
32 STUDY DESIGN AND PROTOCOL DEVELOPMENT
with ideal research conditions. Whatever procedures are selected, they should
be carefully documented in the protocol and emphasized during training.
Data collection and management
First, it should be absolutely clear who is the key person at each clinical site
responsible for administering the HRQoL. This will include, in addition to the
usual responsibilities associated with the clinical trial, ensuring that there is
someone who knows when the patient will arrive, will make sure the patient
receives the questionnaire prior to undergoing diagnostic or therapeutic pro-
cedures and has a quiet place to complete the assessment, and is responsible
for implementing follow-up procedures when the patient is not available as
expected. At the time of the rst assessment, this key person should empha-
size the importance to the investigators of obtaining the patients perspective,
review the instructions with the subjects, emphasize that there are no correct
or incorrect responses, encourage the subjects to provide the best answer they
can to every question, and remind the subjects that they will be asked to re-
peat the assessment at later dates (if applicable). This person may have the
responsibility of reviewing the forms for missing responses, but care needs to
be taken to balance condentiality with the need to minimize missing data. If
the assessment consists of an interview, it requires sucient trained personnel
to schedule and conduct the interview.
Second, there needs to be a system that identies when patients are due for
assessments. This may include preprinted orders in the patients chart that
identify which HRQoL assessments should be administered at each clinic visit.
This process may be assisted by support from a central data management of-
ce where calendars and expectation notices are generated. Stickers on the
patients chart identifying him or her as part of a study may also be help-
ful. Other options include ow sheets, study calendars, and patient tracking
cards [102].
Avoiding missing data
Although analytic strategies exist for missing data, their use is much less
satisfactory than initial prevention. Some missing data, such as that due to
death, is not preventable. However, both primary and secondary prevention
are desirable. In terms of primary prevention, missing data should be min-
imized at both the design and implementation stages of a clinical trial [37,
171]. In most studies, a nurse or research coordinator is responsible for giv-
ing the HRQoL questionnaire to the patient. Among these individuals, the
reasons for missing data include lack of time and perceived lack of physician
support, inadequate protocols, lack of knowledge on justication and ratio-
nale for collecting HRQoL data, lack of reminders, and lack of adequate sites
for questionnaire completion [171]. Thus, clearly specied procedures in the
protocol for collecting HRQoL are the rst step in minimizing missing data.
This should include information of collection procedure if treatment sched-
ule is disrupted and procedures for completion of the questionnaire when the
c2638 C2638Ch02 February 12, 2002 11:35
CONDUCT 33
patient requires assistance. Provide a system for prompting nurses and re-
search personnel that a HRQoL assessment is due. Consider alternative meth-
ods to obtain follow-up data when patients do not complete questionnaires.
Educate patients, research assistants, and primary investigators about the
importance of collecting these assessments on all patients willing to complete
the questionnaire. Point out that reluctance to approach all patients on all
occasions will lead to selection bias. The timing and duration of assessment
should also be reasonable. Be practical about how often and how long you
can follow patients.
Secondary prevention consists of gathering information that is useful in the
analysis and interpretation of the results. This includes collection of data on
factors that may contribute to missing assessments and data that are likely
to predict the missing HRQoL measures. Thus, one should prospectively doc-
ument reasons for missing data. The classications that are used should be
specied in a manner that helps the analyst decide whether the reason is
related to the individuals HRQoL. For example, Patient refusal does not
clarify this, but reasons such as Patient refusal due to poor health and Pa-
tient refusal unrelated to health will be useful. Other strategies for secondary
prevention may include gathering concurrent data on toxicity, evaluations of
health status by the clinical sta, or HRQoL assessments from a caretaker.
Uses of this type of data are discussed in later chapters.
Education
Education can be an important part of minimizing missing data. It must start
at the investigator level and include research assistants (often nurses) as well
the patient. Vehicles for education include the protocol (with strong justica-
tions for the HRQoL assessments), symposia, videos, and written materials.
Videos may be valuable as training vehicles for both research sta and pa-
tients. Although there are often face-to-face training sessions at the initiation
of a study, research personnel can change over time. Training tapes directed
toward research personnel can deal with procedures in more detail than is
possible in the protocol. Examples include how to handle a patient who is not
able to ll in the questionnaire and not letting family or friends assist with the
completion of the questionnaire. Training tapes are especially useful for pro-
viding positive ways of approaching the patient. For example, instead of refer-
ring to participation as burdensome (e.g., We have a lot of forms that youll
need to ll out), the HRQoL assessment can be placed in a positive light [13]:
We want to know more about the quality of life of people as they go through this
treatment and the only way to know is to ask you. In order to do this, we ask
that you complete this brief questionnaire. It usually takes about (X) minutes ...
Hopwood et al. [71] noted that, in three trials for lung and head and neck
cancer, stas considering the patient to be too ill to complete the HRQoL
assessments was the most commonly cited problem aecting the distribution
of questionnaires. However, patient refusal was the least-cited problem. It is
understandable that study personnel are reluctant to approach patients when
c2638 C2638Ch02 February 12, 2002 11:35
34 STUDY DESIGN AND PROTOCOL DEVELOPMENT
they appear to be feeling particularly ill, but to minimize the bias from se-
lecting out these patients, all should be asked to complete the questionnaire.
There may be ways of encouraging ill patients, specically by providing condi-
tions that make it as easy as possible for them to complete the questionnaire.
When a patient refuses, of course, that refusal must be respected.
Patient information sheets, which explain to the patient the rationale be-
hind the HRQoL assessments, will minimize missing data. These sheets can
contain messages about the importance of the patients perspective, notes that
there are no correct answers to the questions, and the reasons it is important
to respond to every question and to complete the follow-up questionnaires. In
addition to the persuasive information, the fact that patients can refuse the
questionnaire without aecting their treatment or their relationship with their
doctor should be included.
Forms
The data collection forms should be attractive and professional in appearance,
using fonts that are large enough to ensure readability (e.g., 12-point char-
acters or greater). Do not use two-sided forms! Patients will often not look
at the back of the page, and if forms are copied at the sites, there is a high
probability that only the front side will be copied.
Explicit procedures for follow-up
A practical schedule with HRQoL assessments linked to planned treatment
or follow-up visits can decrease the number of missing HRQoL assessments.
When possible, it is wise to link HRQoL assessments with other clinical assess-
ments (Designs 2 and 3 in Table 2.3). This has several advantages. The avail-
ability of the patient increases the likelihood that the HRQoL assessment will
be completed. Sta may be more likely to remember when the HRQoL assess-
ment is scheduled if the timing is linked to clinical or laboratory follow-up. Fi-
nally, it is possible to link clinical events and laboratory values to the HRQoL
assessments. Less frequent assessment of HRQoL (Design 1) decreases patient
burden slightly but may introduce confusion about the schedule and thus lead
to missed assessments. If more frequent assessments of HRQoL (Design 4) are
specied in the design, strategies for obtaining the additional assessments
must be identied. If the duration of HRQoL assessment continues after ther-
apy is discontinued, this should be clearly stated and protocol owcharts for
treatment and assessment schedules should clearly reect the dierence.
The protocol and training materials should include specic procedures to
minimize missing data. It should be clear what the allowable windows are
for each assessment. The protocol should specify whether or not follow-up by
telephone or mail is acceptable. Finally, since missing data will occur, it is im-
portant to document the reasons for the missing assessment. When construct-
ing the set of possible reasons, the responses should dierentiate whether the
nonresponse was likely to be related to the patients HRQoL. Documentation
of the reasons for missing assessments can be combined with other questions
c2638 C2638Ch02 February 12, 2002 11:35
CONDUCT 35
Table 2.5. Reverse coding procedure of subtracting response from sum of the lowest
and highest possible responses to obtained reversed score.
Original responses 0 1 2 3 4
Calculation 40 41 42 43 44
Reversed score 4 3 2 1 0
Original responses 1 2 3 4 5
Calculation 61 62 63 64 65
Reversed score 5 4 3 2 1
When possible responses range from 0 to 4, sum is 4. When
responses range from 1 to 5, sum is 6.
about the conditions under which the HRQoL was administered. For example:
What was the site and mode of administration? Was any assistance given and,
if so, by whom?
Scoring instruments
Protocols should include explicit instructions or references to the instructions
for creating summary or subscale scores for the HRQoL instruments. Refer-
ences to methodology should be accessible and not in hard-to-obtain manu-
als or personal communications. If the scoring procedure does not include a
strategy for handling missing items, there is the potential for a substantial
number of subjects to have missing scores, even when they completed the as-
sessment. If there are any proposed variations in the scoring procedure, they
also should be specied in the protocol.
Reverse coding
The FACT scales are typical of many HRQoL health status measures. The
rst step is to reverse the scoring of negatively worded questions (Table 2.5).
Notice that, in contrast to most of the questions in the Social/Family Well-
being section of the FACT (Table 2.6), a higher score is associated with a
negative outcome for questions 9 (distant from friends) and 13 (communica-
tion is poor). A quick trick to reverse the coding is to add the lowest and
highest possible scores for the questions and subtract the response from that
sum (see Table 2.5). For example, when the range of scores is 0 to 4, the sum is
4. If the responses to questions 9 and 13 are 2 and 3, respectively, the reverse
scores are 4 2 = 2 and 4 3 = 1.
Scoring multi-item scales
Most HRQoL instruments combine the responses from multiple questions to
derive summated scores for the relevant dimensions of HRQoL. The motivation
is to increase the reliability of the scores and to facilitate interpretation. The
most widely used method of combining responses is the method of summated
c2638 C2638Ch02 February 12, 2002 11:35
36 STUDY DESIGN AND PROTOCOL DEVELOPMENT
Table 2.6. Sample questions from the Social/Family Well-being subscale of the Func-
tional Assessment of Cancer Therapy (FACT) questionnaire (version 3).
Below is a list of statements that other people with your illness have said are
important. By circling one number per line, please indicate how true
each statement has been for you during the past 7 days.
Not at A little Some- Quite Very
Social/family well-being all bit what a bit much
9. I feel distant from my friends 0 1 2 3 4
10. I get emotional support from my 0 1 2 3 4
family
11. I get support from my friends and 0 1 2 3 4
neighbors
12. My family has accepted my illness 0 1 2 3 4
13. Family communications about my 0 1 2 3 4
illness is poor
14. I feel close to my partner (or person 0 1 2 3 4
who is my main support)
15. Have you been sexually active 0 1 2 3 4
during the past year? No Yes
If yes: I am satised with my sex life
ratings or Likert summated scales [45]. The total score is obtained by adding
the responses from all items. For a scale containing k items, each scored from
0 to r, the summated score will range from 0 to k r. If the scale is scored
from 1 to r, the summated score will range from k to k r.
Because dierent scales (and even subscales within the same instrument)
have a dierent number of items and a dierent range of responses, the range of
these summated scores will have dierent ranges, making interpretation more
dicult. For example, the emotional well-being score of the FACT (version 2)
has a range of 0 to 20; the functional well-being, physical well-being, and
family social well-being scores have a range of 0 to 28; and the total score
has a range of 0 to 140. It is common practice to standardize the summated
scores to range from 0 to 100. This is done by subtracting the lowest possible
summated score (S
min
) from each score (S
i
) and then multiplying the result
by 100/(S
max
S
min
).
A much smaller number of HRQoL instruments use factor-analytic weights
to construct scales. The source of these weights should be very carefully ex-
amined before accepting them. The weights need to be established in very
large representative samples of subjects that include all levels of disease (or
lack of disease) and expected types of treatment. Fayers and Hand [44] point
out how sensitive the factor structure and resulting weights are to the pop-
ulation from which they were derived. Weights derived from another clinical
c2638 C2638Ch02 February 12, 2002 11:35
CONDUCT 37
Table 2.7. Selected physical function questions from the SF-36 Health Survey that
have a roughly hierarchical structure.
The following items are about activities you might do during a typical day.
Does your health now limit you in these activities? If so, how much?
Yes, Yes, No, not
limited limited limited
a lot a little at all
a. Vigorous activities, such as running, lifting 0 1 2
heavy objects, participating in strenuous
sports
b. Moderate activities, such as moving a table, 0 1 2
pushing a vacuum cleaner, bowling, or
playing golf
c. Lifting or carrying groceries 0 1 2
d. Climbing several ights of stairs 0 1 2
e. Climbing one ight of stairs 0 1 2
f. Bending, kneeling, or stooping 0 1 2
g. Walking more than a mile 0 1 2
h. Walking several blocks 0 1 2
i. Walking one block 0 1 2
j. Bathing or dressing yourself 0 1 2
trial or a selected population are likely to reect the side eects of the specic
treatments in that trial rather than the underlying construct.
Item nonresponse
It is not uncommon to have a small proportion (1 to 2%) of the items skipped.
The frequency of missing items has been observed to be higher among elderly
individuals [81] and those that are more ill. If the responses from selected
patients are excluded from analysis because of missing items, there is a danger
of biasing the results. Sometimes skipping an item is inadvertent, but it may
also occur because a question is not applicable for that subject. Question 13 of
the Social/Family Well-being is a good example of a question that is skipped
by some subjects. To avoid missing subscale scores because of one or two
missing responses, most scales specify exactly how to score the question in
the presence of missing items. When a strategy is specied for a particular
instrument, that strategy should be used for scoring when reporting the results
of a clinical trial unless the new strategy was specied in the protocol as an
alternative scoring system.
The most typical strategy, sometimes referred to as the half rule, is to
substitute the mean of the answered questions for that specic subscale for
c2638 C2638Ch02 February 12, 2002 11:35
38 STUDY DESIGN AND PROTOCOL DEVELOPMENT
the missing responses as long as at least half the questions were answered.
This strategy works well for scales where there is no particular ordering of
the diculty of the questions, as in the FACT [38], and the diculty of the
question does not have a wide range across the subscale items. It should be
used cautiously (or not at all) when the questions have a hierarchy,* as when
functional well-being is assessed through the ability to do certain activities
(Table 2.7). The physical functioning scale of the SF-36 is a good example of
a scale where the individual items have a wide range of diculty. If individuals
are limited a lot in moderate activities, it is clear that they are also limited a
lot in strenuous activities. Similarly, if subjects respond that they can walk a
mile without limitations, they will also be able to walk several blocks. Special
strategies may need to be specied for these types of questions [81].
SAS example
Both reverse coding and scoring can be accomplished in a few lines of code
using macros rather than separate statements for every item and scale. This
minimizes the possibility of typographical errors. The responses to the items
are ITEM1, ITEM2,..., ITEMK, etc. where missing values are coded as SAS
missing values. The rst step is to reverse-code questions that are nega-
tively worded. The second step is to create a scoring macro that (1) checks
whether the number of completed responses (N(of &ITEMS(*))) is greater
than half the possible responses (DIM(of &ITEMS(*))), (2) calculates the sum-
mated score in the presence of missing items, and (3) standardizes the sum-
mated score to have a range of 0 to 100. Arrays are then constructed with the
items specic to each subscale included, and the macro is invoked to calculate
the subscales.
DATA WORK.SCORES;
SET WORK.ITEMS;
*** Reverse Code Selected Items Coded 0-4 ***;
*** REV is the list of questions that are reverse coded ***;
ARRAY REV ITEM1 ITEM7 ITEM13 ... ITEM31;
DO I=1 to DIM{REV};
REV[I]=4-REV[I];
END;
*** Define Scoring Macro ***;
*** ITEM is the name of an array of items ***;
*** MIN is the lowest possible response to a question ***;
*** MAX is the highest possible response to a question ***;
*** SUMSCORE is the name of the summated score ***;
*** STDSCORE is the name of the standardized score ***;
%MACRO SCORE(ITEMS,MIN,MAX,SUMSCORE,STDSCORE);
N_SCALE=DIM(&ITEMS); * # items in subscale *;
* When questions have a strict hierarchy, the scale is referred to as a Guttman scale.
c2638 C2638Ch02 February 12, 2002 11:35
SUMMARY 39
IF N(of &ITEMS(*)) GE N_SCALE/2 THEN
&SUMSCORE=MEAN(of &ITEMS(*))*N_SCALE;
&STDSCORE=(&SUBSCORE-(&MIN*N_SCALE))
*100/((&MAX-&MIN)*N_SCALE);
%MEND SCORE;
*** Array items for each subscale ***;
ARRAY PWB ITEM1 ITEM2 ITEM3 ITEM4 ITEM5 ITEM6 ITEM7;
ARRAY EWB ITEM20 ITEM21 ITEM22 ITEM23 ITEM24;
*** Compute summated scores and standardized scores ***;
%SCORE(PWB,0,4,PWB_SUM,PWB_STD);
%SCORE(EWB,0,4,EWB_SUM,EWB_STD);
2.8 Summary
It is critical to establish explicit research objectives during protocol devel-
opment. Details should include (1) domains, (2) population, and (3) time
frame relevant to the research questions.
The HRQoL instruments should be selected carefully, ensuring that they are
appropriate to the research question and the population under study. New
instruments and questions should be considered only if all other options
have been eliminated.
The frequency and duration of HRQoL assessments should reect the re-
search objectives and the disease and treatment under study.
The protocol should include explicit instructions for the conduct of the
study to avoid bias and missing data. Educational materials for research
sta and patients will be useful.
c2638 C2638Ch02 February 12, 2002 11:35
c2638 C2638TextureC03 January 28, 2002 17:19
CHAPTER 3
Models for Longitudinal Studies
3.1 Introduction
This chapter addresses the analysis of longitudinal studies in the context of
HRQoL investigations. In the previous chapter, event- or condition-driven
and time-driven designs were described. These designs have corresponding
analytic models: a repeated measures model and a growth curve model. In a
repeated measures model, time is conceptualized as a categorical variable.
Each assessment must be assigned to one category. In the NSCLC study
(Example 2), these categories are Pre-therapy, 6 weeks, 12 weeks, and
6 months. In a growth curve model, time is conceptualized as a continuous
variable. In the NSCLC study, the variable is days since randomization.
In some trials, either analytic design is appropriate, and in other cases there
will be clear reasons for preferring one or the other of these approaches. Al-
though the schedule for HRQoL assessments is often constrained by practical
considerations such as the timing of patient visits, the research question and
the analytic methods appropriate for that question should be the primary
consideration.
Repeated measures models
Repeated measures models are used in longitudinal studies with event-driven
designs where there is a strong conceptual identity for each assessment. A
good example is the adjuvant breast cancer study (Example 1) where assess-
ments occurred during the three phases of the study: prior to therapy, while
on therapy, and 4 months after the end of therapy. Repeated measures mod-
els may also be useful in some studies with a limited number (two to four) of
assessments. If the timing of the longitudinal assessments occurs within pre-
dened windows of time, then it is feasible to classify each of the assessments
uniquely as one of the repeated measures.
Assignment of assessments to a planned assessment time (landmark) may
become dicult in studies with more frequent assessments. These studies
typically have more mistimed* observations with increasing rates of dropout.
The process of dening windows of time for each assessment becomes increas-
ingly dicult. If the windows are wide, an individual may have more than
one observation within the window. This will require the analyst to discard
all but one of the observations occurring within that window. Not only will
* The term mistimed does not necessarily refer to design (protocol) violations but rather
indicates that the interval between observations may vary across subjects.
c2638 C2638TextureC03 January 28, 2002 17:19
42 MODELS FOR LONGITUDINAL STUDIES
this reduce the statistical power of the study, but it may also lead to biased
estimates, depending on how the observations are selected for discard. For ex-
ample, bias may occur if discarded assessments were delayed when individuals
were experiencing toxicity. Using narrow intervals may lead to fewer subjects
with data at each landmark. This creates instability of the estimation of both
the means and the covariance. In summary, one of the restrictions associated
with the choice of a repeated measures model for the analysis of the trial is
the necessity of identifying each assessment with each landmark.
Growth curve models
The term growth curve models comes from early applications to describe
changes in height and weight as a function of age. Typically, these measure-
ments were taken at dierent ages for each child in a study. Curves were t
to the data to estimate the typical height or weight for a child as a func-
tion of age. The same concept is relevant when we want to describe changes
in HRQoL as a function of time. The curves can be dened in many ways,
but they are most often described by polynomial functions or piecewise lin-
ear regressions. The simplest example of a growth curve model is a straight
line.
Growth curve models are useful in several settings. The rst is when the
timing of assessments diers widely among individuals. Growth curve models
are also useful in studies with a large number of HRQoL assessments, where
it is feasible to model changes over time with a smaller number of parameters
than is required for a repeated measures model. Finally, there may be analytic
approaches where a mixed-eects model is necessary to model HRQoL in the
presence of dropout. Two of the models described in Chapter 9 will utilize a
mixed-eects model.
Selection between models
Adjuvant breast cancer study (Example 1)
In the adjuvant breast cancer study, the design requires the use of a repeated
measures model. Each of the three assessments was planned to measure HRQoL
at a dierent phase of the study. Because of dierent durations of therapy, the
nal assessments, scheduled 4 months after the completion of therapy, did not
occur as scheduled for all subjects but at approximately 32 and 40 weeks after
the beginning of the study (see Figure 1.2). Some post-therapy observations
also occurred earlier when a patient discontinued therapy early. However, the
intent of the design was to compare HRQoL when subjects had been o ther-
apy long enough to recover from the eects of acute toxicity. Thus, the exact
timing of the HRQoL assessment relative to the time they started therapy
was not relevant.
c2638 C2638TextureC03 January 28, 2002 17:19
BUILDING THE ANALYTIC MODELS 43
NSCLC study (Example 2)
With only four assessments that were widely spaced in time, either a repeated
measures or a growth curve model can be justied. Figure 1.3 displays the
actual timing of the assessments, which were clustered around the planned
time of the assessment but became more variable as the study progressed.
As previously mentioned, one restriction associated with a repeated measures
model is the necessity of identifying each assessment with a landmark time.
In this example, this was not too dicult because of the limited number of
assessments and the wide spacing between them. There were no cases where an
individual had more than four assessments, and there were only a few cases
where there was some question about whether an assessment was closer to
the 12- or 26-week target (see Figure 1.3). If this study had been designed to
continue past 26 weeks or if there had been more frequent assessments, such
as at 18 weeks, it would have been more dicult to classify each assessment
uniquely, as required in a repeated measures model.
A growth curve model could also be considered for the NSCLC study. Al-
though the assessments are based on the schedule of chemotherapy cycles,
the changes in HRQoL can be easily modeled as a function of calendar time.
Further, the assessments are not directly linked to an event or condition. The
chemotherapy was given in 3-week cycles, but cycles could be delayed as a
result of toxicity. With recognition of the potential for delays in therapy, the
second and third HRQoL assessments were specied to occur on the rst day
(prior to therapy) of the third and fth cycle of therapy, at approximately 6
and 12 weeks. The nal assessment was planned for 26 weeks.
Renal cell carcinoma study (Example 3)
In the renal cell carcinoma study, only a growth curve model is feasible. This
is because of the large number of potential observations where the timing rela-
tive to randomization becomes more varied as time progresses (see Figure 1.4).
Thus, it becomes increasingly dicult to assign each observation to a land-
mark time. Forcing this study into a repeated measures design will also pro-
duce unstable estimates of covariance parameters during the later follow-up
periods.
3.2 Building the analytic models
Consider a longitudinal study where n assessments of HRQoL are planned for
each subject over the course of the study. We will consistently use h to indicate
the hth group, i to designate each unique individual, and j to indicate the
jth assessment of the ith subject. Thus, Y
hij
indicates the jth observation of
HRQoL on the ith individual in the hth group. The indicator of group (h) will
often be dropped to simplify notation when it is not necessary to distinguish
between groups.
c2638 C2638TextureC03 January 28, 2002 17:19
44 MODELS FOR LONGITUDINAL STUDIES
A general linear model for the HRQoL outcomes can be expressed as
Y
ij
= X
ij
+
ij
(3.1)
or in matrix notation as
Y
i
= X
i
+
i
, (3.2)
where
Y
i
= the complete data vector of n planned observations of the outcome
for the ith individual, which includes both the observed data Y
obs
i
and missing data Y
mis
i
X
i
= the design matrix of xed covariates corresponding to the complete
data (Y
i
)
= the corresponding vector of xed-eects parameters

i
= the vector of residual errors

i
= the covariance of the complete data (Y
i
), which is a known function
of the vector of unknown variance parameters, :
i
= f()
This general linear model can be formed as either a repeated measures model
or a growth curve model.
The recommended model building process [33, 150] starts by dening a fully
parameterized model for the means (X
i
), then identifying the structure of

i
, and nally simplifying the mean structure. This procedure is described in
the remainder of this chapter.
Statistics guiding model reduction
Likelihood ratio tests
Two models, one nested within the other, can be compared with a maximum
likelihood (ML) ratio test [75] or restricted maximum likelihood (REML) ratio
test [87, p. 278]. The statistics are constructed by subtracting the values of 2
times the log likelihood and comparing the statistic with a
2
distribution
with degrees of freedom equal to the dierence in the number of parameters
in the two covariance structures. Tests based on the REML are valid as long
as the xed eects are the same. Thus, either ML or REML ratio tests are
valid when comparing nested covariance structures. However, likelihood tests
for nested xed-eects models must be limited to the use of ML (METHOD=ML)
because the restricted likelihood adjustment depends upon the xed-eects
design matrix [87, pp. 298, 502].
To identify nesting, consider whether a set of restrictions on the parameters
in one model can be used to dene the other model. Typical restrictions are
constraining a parameter to be zero or two parameters to be equal to each
other. For example, the restriction
1
=
2
=
3
on the Toplitz structure
denes the Compound Symmetry structure (see Table 3.1). The homogeneous
structures are all nested within their respective heterogeneous structure,
1
=

2
=
3
=
4
. Most structured covariance matrices are nested within the
c2638 C2638TextureC03 January 28, 2002 17:19
BUILDING THE ANALYTIC MODELS 45
Table 3.1. Examples of covariance structures for three repeated measures. Examples
progress from the least structured at the top of the table to the most structured at
the bottom of the table.
No. of
parameters Structure
Unstructured 10
2
1

12

13

14
TYPE=UN
2
2

23

24

2
3

34

2
4
Heterogeneous 7
2
1

1

1

1

2

1

3
Toplitz
2
2

2

1

2

2
TYPE=TOEPH
2
3

3

2
4
Heterogeneous 5
2
1

1

2

1

3

1

Compound
2
2

2

3

2

Symmetry
2
3

3

TYPE=CSH
2
4
Heterogeneous 5
2
1

1

2

1

3
Autoregressive
2
2

2

3

2

2
[AR(1)]
2
3

3

TYPE=ARH(1)
2
4
Toplitz 4
2

1

2

2

2

3
TYPE=TOEP
2

1

2

2
First-order 3
2

2

2

2

2
Autoregressive moving
2

2

2

Average [ARMA(1,1)]
2

TYPE=ARMA(1,1)
2
Compound 2
2

2

2

2

Symmetry
2

2

2

TYPE=CS
2

2
First-order 2
2

2

2

3
Autoregressive
2

2

2

2
[AR(1)]
2

TYPE=AR(1)
2
c2638 C2638TextureC03 January 28, 2002 17:19
46 MODELS FOR LONGITUDINAL STUDIES
unstructured matrix, but structures such as the Toplitz and AR(1) models
cannot be directly compared.
Other statistics
Criteria such as Akaikes Information Criterion (AIC) and the Bayesian Infor-
mation Criterion (BIC) are also useful. These are the 2 log likelihood values
penalized for the number of parameters estimated. The BIC imposes a greater
penalty than the AIC.
3.3 Building repeated measures models
Mean structure
Repeated measures models for the HRQoL can be conceptualized several ways
that are analytically equivalent. The most straightforward model is referred
to as a cell means model. The parameters (usually means) from the repeated
measures models are interpreted in the same way that parameters are inter-
preted in simple regression and analysis of variance (ANOVA) models. If we
think of two-factor design with treatment arm and time as the factors, each
cell represents one point in time within a specic treatment arm (Table 3.2).
The equation for the model is
Y
hij
=
hj
+
hij
, (3.3)
where
hj
is the average HRQoL score for the jth measurement of the hth
group.
For example, in the adjuvant breast cancer study we have two treatments
and three assessments over time. The simplest model for this study is a two-
factor (treatment and time) design with the mean HRQoL score estimated for
Table 3.2. Adjuvant breast cancer example: Repeated measures cell means model
with two factors, treatment (h) and time (j).
Model Y
hij
=
hj
+
hij
Before During After
Treatment arm therapy therapy therapy
Cell Means (
hj
)
CAF
11

12

13
16-week
21

22

23
Estimates
CAF (
1j
) 7.6 6.8 8.0
16-week (
2j
) 7.7 6.3 8.1
Dierence (
2j

1j
) 0.18 0.51 0.09
c2638 C2638TextureC03 January 28, 2002 17:19
BUILDING REPEATED MEASURES MODELS 47
Table 3.3. Adjuvant breast cancer example: Repeated measures model with two
discrete factors (treatment and time) and a continuous covariate (age).
Model Y
hij
=
hj
+
age
X
i
(age) +
hij
Treatment Before During After
arm therapy therapy therapy
Cell Means
CAF
11
+
age
X
i
(age)
12
+
age
X
i
(age)
13
+
age
X
i
(age)
16-week
21
+
age
X
i
(age)
22
+
age
X
i
(age)
23
+
age
X
i
(age)
each of the two treatments prior to therapy, during therapy, and after therapy
(Table 3.2).
Covariates can be added to the model when needed. For example, if the
eect of age is independent of treatment and time, we can add age at the time
of diagnosis (without interactions) to the model (Table 3.3). The equation for
the model is
Y
hij
=
hj
+
age
X
i
(age) +
hij
, (3.4)
where X
i
(age) = age in years at diagnosis 50. The interpretation of the
means (
hj
) will change with the addition of these covariates. To facilitate the
interpretation, it is advisable to center continuous covariates at a meaningful
point. In this example, the average age at diagnosis is approximately 50 years,
so the covariate is centered at that point. The interpretation of
hj
is the
expected (or average) HRQoL score for the jth measurement of the hth group
for a woman who is 50 years of age. Without this centering process, estimates
of
hj
may be outside the range of possible values and the interpretation of
each parameter is dicult. Using the age example without centering, we are
estimating the mean HRQoL for a woman who is 0 years of age.
More covariates can be added to the model as needed, including those in
which there are interactions with treatment, time, or both. Indiscriminate ad-
dition of covariates with missing values can create analytic problems because
of the deletion of cases from the study.
SAS example of a cell means model
In the adjuvant breast cancer study, the datale EXAMPLE1 has a unique record
for each HRQoL assessment. CASEID identies the patient. FUNO identies the
time of the assessment with values of 1, 2, and 3. TRTM identies the treatment
arm. The rst part of the MIXED procedure for a cell means model might appear
as
PROC MIXED DATA=EXAMPLE1;
* Cell Means Model *;
CLASS FUNO TRTM;
MODEL BCQ=FUNO*TRTM/NOINT SOLUTION;
c2638 C2638TextureC03 January 28, 2002 17:19
48 MODELS FOR LONGITUDINAL STUDIES
The CLASS statement identies TRTM and FUNO as categorical variables with
two levels of treatment and three levels of time. The term FUNO*TRTM in the
MODEL statement creates a design matrix corresponding to the cell means
model previously displayed in Table 3.2. The intercept term is suppressed by
the NOINT option. The estimates of the means for each treatment by time
combination are generated by the SOLUTION option in the MODEL statement:
Solution for Fixed Effects
Effect FUNO TRTM Estimate Std Error
FUNO*TRTM Pre- Treatment A 7.56837973 0.12785128
FUNO*TRTM Pre- Treatment B 7.75274306 0.13857029
FUNO*TRTM During Treatment A 6.82574053 0.16542708
FUNO*TRTM During Treatment B 6.31521676 0.17716578
FUNO*TRTM 4 months Post-Tx A 8.00584072 0.12788503
FUNO*TRTM 4 months Post-Tx B 8.09419911 0.13935828
With addition of the age covariate AGE EVAL, the MIXED statements become
PROC MIXED DATA=EXAMPLE1;
* Cell Means Model *;
CLASS FUNO TRTM;
MODEL BCQ=FUNO*TRTM AGE_EVAL/NOINT SOLUTION;
In this example, none of the baseline covariates was associated with the out-
come and none is included in the remainder of the analysis.
Covariance structures
The major dierence between a univariate regression model for independent
observations and a multivariate model for repeated measures is that the assess-
ments for each individual are assumed to be correlated over time, Var[Y
i
] =
i
.
Examples of covariance structures include unstructured, Toplitz, compound
symmetry (exchangeable), and autoregressive (Table 3.1).
Unstructured covariance
The unstructured covariance is the least restrictive and generally the best
choice when the number of repeated measures is small. With only three as-
sessments, this structure has six variance parameters. First, the variance of
the HRQoL measure at each time point (
2
1
,
2
2
,
2
3
) is allowed to be dierent.
The need for this type of exible structure is illustrated in the adjuvant breast
cancer study (see Table 3.4) where there is more variation in the HRQoL mea-
sure while the subjects are on therapy (
2
2
= 2.28) than before or after therapy
(
2
1
= 1.43,
2
3
= 1.32). Further, the covariance of each pair of HRQoL mea-
sures (
12
,
13
,
23
) is allowed to be dierent. It is not uncommon for HRQoL
assessments occurring during therapy to be more strongly correlated with each
other than with the initial pretherapy assessment.
c2638 C2638TextureC03 January 28, 2002 17:19
BUILDING REPEATED MEASURES MODELS 49
Table 3.4. Estimated covariance and correlation for the adjuvant breast cancer study
(Example 1).
Structure r AIC BIC Covariance Correlation
Unstructured 6 1370.9 1408.0 1.43 1.11 0.75 1.00 0.62 0.54
TYPE=UN 2.28 1.12 1.00 0.64
1.32 1.00
Heterogeneous 5 1369.1 1403.1 1.44 1.14 0.75 1.00 0.63 0.54
Toplitz 2.28 1.09 1.00 0.63
TYPE=TOEPH 1.31 1.00
Heterogeneous 4 1370.1 1401.0 1.45 1.08 0.84 1.00 0.60 0.60
Compound 2.23 1.04 1.00 0.60
Symmetry 1.34 1.00
TYPE=CSH
Heterogeneous 4 1375.1 1406.0 1.44 1.13 0.53 1.00 0.63 0.39
Autoregressive 2.28 1.07 1.00 0.63
TYPE=ARH(1) 1.29 1.00
Homogeneous 4 1386.4 1414.2 1.64 0.98 0.95 1.00 0.59 0.58
Toplitz 1.64 0.98 1.00 0.59
TYPE=TOEP 1.64 1.00
r =Number of unique parameters in covariance structure; AIC=Akaikes Information Cri-
terion, smaller values are better; BIC=Bayesian Information Criterion, smaller values are
better.
When the number of repeated measures increases, the number of parameters
in the unstructured covariance matrix increases dramatically. For example,
as the number increases from three to six repeated measures, the number
of parameters increases from 6 to 21 (Table 3.5). In large data sets with
nearly complete follow-up, estimation of the covariance parameters is not a
problem. As the dropout increases, especially in smaller studies, it becomes
more dicult to obtain stable estimates of the covariance parameters. In these
settings it may be advisable to place restrictions on the covariance structure.
Structured covariance
If the correlation between measures is roughly equal regardless of how far apart
in time the observations are taken, then compound symmetry is a possible
candidate. This covariance structure is typically observed when the prole of
each subjects HRQoL is roughly parallel over time. This structure is typical
of measures that are unaected by the disease or its treatment during the
course of the study or when all subjects are equally aected.
When HRQoL observations are taken closely in time, the correlation of
the residual errors is likely to be strongest for observations that are close in
c2638 C2638TextureC03 January 28, 2002 17:19
50 MODELS FOR LONGITUDINAL STUDIES
Table 3.5. Number of parameters required to model covariance for two to seven
repeated measures.
No. of repeated measures
Covariance structure 2 3 4 5 6 7
Unstructured 3 6 10 15 21 28
Heterogeneous Toplitz 3 5 7 9 11 13
Heterogeneous autoregressive 3 4 5 6 7 8
Heterogeneous compound 3 4 5 6 7 8
Toplitz 3 3 4 5 6 7
ARMA(1,1) 2 3 3 3 3 3
Compound symmetry 2 2 2 2 2 2
Autoregressive 2 2 2 2 2 2
Note that the number of parameters increases more rapidly for
the least-structured covariance matrices.
time and weakest for observations that are the furthest apart. The Toplitz,
autoregressive moving average (ARMA(1,1)) and autoregressive structures
(AR(1)) allow observations that are further apart to be less strongly correlated.
The autoregressive is the most restrictive of these structures. The correlation
between two observations is a function of the separation between the obser-
vations.* The covariance decays exponentially as a function of the time lag.
It is rare that the correlation of HRQoL measures declines this rapidly, even
when the observations are far apart in time. The Toplitz and (ARMA(1,1))
provide more exible structures that more closely t the correlation typical
of HRQoL measures.
There are numerous other possible covariance structures that are variations
of these. One option is to allow the variance to be heterogeneous but retain
the correlation structures. Additional options for covariance structures are
described in detail in other sources, including Jennrich and Schluchter [75],
chapters 3 and 6 of Jones [76], chapter 3 of Verbeke and Molenberghs [154],
and in the SAS Proc Mixed documentation [115].
SAS example of a repeated measures model
In the adjuvant breast cancer example, there are only three repeated mea-
sures and enough observations to reliably estimate an unstructured covariance
(Table 3.4). The unstructured covariance of the repeated measures is dened
by the REPEATED statement with TYPE=UN. METHOD=ML is added to the PROC
statement to facilitate comparisons between dierent models. The expanded
* The autoregressive structure shown in Table 3.1 assumes the observations are equally
spaced. Further discussion of covariance structures for unequally spaced observations can
be found in Jones [76, chapters 3 and 6].
c2638 C2638TextureC03 January 28, 2002 17:19
BUILDING REPEATED MEASURES MODELS 51
MIXED procedure assuming an unstructured covariance would appear as
PROC MIXED DATA=EXAMPLE1 METHOD=ML;
* Cell Means Model *;
CLASS FUNO TRTM;
MODEL BCQ=FUNO*TRTM/NOINT;
* Unstructured Covariance *;
REPEATED FUNO/SUBJECT=CASEID TYPE=UN R RCORR;
Other TYPE= options specifying the covariance structure are displayed in
Table 3.4. The estimates of the covariance parameters appear as follows:
Covariance Parameter Estimates
Cov Parm Subject Estimate
UN(1,1) CASE 1.4112
UN(2,1) CASE 1.0963
UN(2,2) CASE 2.2496
UN(3,1) CASE 0.7389
UN(3,2) CASE 1.1037
UN(3,3) CASE 1.3052
The R and RCORR options request the output of the covariance and correlation
matrices for the rst subject in the more familiar format. (R=2 would request
the covariance matrix for the second subject.)
Estimated R Matrix for Subject 1
Row Col1 Col2 Col3
1 1.4112 1.0963 0.7389
2 1.0963 2.2496 1.1037
3 0.7389 1.1037 1.3052
Estimated R Correlation Matrix for Subject 1
Row Col1 Col2 Col3
1 1.0000 0.6153 0.5445
2 0.6153 1.0000 0.6441
3 0.5445 0.6441 1.0000
The procedure generates statistics that can be used to guide the model
selection.
Fit Statistics
-2 Log Likelihood 1346.9
AIC (smaller is better) 1370.9
AICC (smaller is better) 1371.6
BIC (smaller is better) 1408.0
The correlations estimated for the unstructured covariance range from 0.54
to 0.64, with the correlation the lowest between the two observations furthest
c2638 C2638TextureC03 January 28, 2002 17:19
52 MODELS FOR LONGITUDINAL STUDIES
apart in time. This suggests that the correlation would possibly t the Toplitz
structure. The increase in variation during the second assessment (during
chemotherapy) suggests heterogeneity along the main diagonal.
More than one covariance structure is likely to provide a reasonable t to the
data. Unless the sample size for the study is very large, it is dicult to choose
denitively among the various covariance structures. The information criteria
can be used to provide guidance. For example, although the heterogeneous
Toplitz structure provides the best t as indicated by AIC, this structure
only reduces the number of parameters by one. Similarly, the heterogeneous
compound symmetry is suggested by the BIC. More critically, the estimates of
all three covariance structures are very similar and the dierences are unlikely
to aect the results of the primary analyses.
Hypothesis testing
Although the cell means model generates estimates of the means for each
treatment, we are usually interested in comparisons of these means. Any hy-
pothesis that can be expressed as a linear combination of the estimated pa-
rameters (means) can be tested. For example, to test for dierences at each
time point the null hypotheses are
H
0
:
21
=
11
vs. H
A
:
21
=
11
,
H
0
:
22
=
12
vs. H
A
:
22
=
12
,
H
0
:
23
=
13
vs. H
A
:
23
=
13
.
These equations can be rewritten, placing all parameters on one side of the
equality:
H
0
:
21

11
= 0 vs. H
A
:
21

11
= 0,
H
0
:
22

12
= 0 vs. H
A
:
22

12
= 0,
H
0
:
23

13
= 0 vs. H
A
:
23

13
= 0.
Putting the parameters in exactly the same order as they appear in the
Solution for Fixed Effects displayed above,
H
0
: 1
11
+ 1
21
+ 0
12
+ 0
22
+ 0
13
+ 0
23
= 0,
H
0
: 0
11
+ 0
21
1
12
+ 1
22
+ 0
13
+ 0
23
= 0,
H
0
: 0
11
+ 0
21
+ 0
12
+ 0
22
1
13
+ 1
23
= 0.
The estimates of these dierences and the corresponding tests are generated
by the ESTIMATE statements with the coecients from the above equations
following the FUNO*TRTM term:
* Estimates of Treatment Differences *;
ESTIMATE Pre-Therapy Diff FUNO*TRTM -1 1 0 0 0 0;
ESTIMATE During Therapy Diff FUNO*TRTM 0 0 -1 1 0 0;
ESTIMATE Post-Therapy Diff FUNO*TRTM 0 0 0 0 -1 1;
c2638 C2638TextureC03 January 28, 2002 17:19
BUILDING REPEATED MEASURES MODELS 53
The results appear in the output as
ESTIMATE Statement Results
Parameter Estimate Std Err DF t Pr > |t|
Pre-Therapy Diff -0.184 0.189 161 -0.98 0.3296
During Therapy Diff 0.511 0.242 151 2.11 0.0368
Post-Therapy Diff -0.088 0.189 143 -0.47 0.6411
One might consider an alternative hypothesis that patients in both treat-
ments experienced a signicant decline in HRQoL during therapy. The null
hypothesis is
H
0
:
12

11
= 0 and
22

21
= 0.
The hypotheses can be tested separately with ESTIMATE statements or jointly
with the CONTRAST statement.
* Estimates of Change from Pre-therapy to Post-therapy *;
ESTIMATE Change in CAF pts FUNO*TRTM -1 0 1 0 0 0;
ESTIMATE Change in 16wk pts FUNO*TRTM 0 -1 0 1 0 0;
* Test Change from from Pre-therapy to Post-therapy *;
CONTRAST During-Pre Change FUNO*TRTM -1 0 1 0 0 0,
FUNO*TRTM 0 -1 0 1 0 0;
Notice that the CONTRAST statement contains the same syntax for identifying
the terms in the hypothesis as the two ESTIMATE statements, but the two
elements are separated by a comma.
The estimates of change in the BCQ scores while on therapy in the CAF
and 16-week arms represent declines of approximately 0.6 and 1.1 standard
deviations*; for the BCQ measure, these are moderate and strong eects,
respectively [18].
ESTIMATE Statement Results
Parameter Estimate Std Err DF t Pr > |t|
Change in CAF pts -0.743 0.136 145 -5.46 0.0001
Change in 16 wk pts -1.438 0.145 144 9.92 0.0001
CONTRAST Statement Results
Source NDF DDF F Pr > F
During-Pre Change 2 145 64.10 0.0001
In this example, the option DDFM=SATTERTH was added to the MODEL state-
ment. This option calculates the degrees of freedom of test statistics based
on the actual design rather than relying on the variable names in the MODEL
statement. (See appendix A.4 of Verbeke and Molenberghs [154] and the SAS
Proc Mixed documentation [115] for further explanation.)
* 0.743/1.29 and 1.438/1.29, where 1.29 =

2
and
2
is the average baseline variance.
c2638 C2638TextureC03 January 28, 2002 17:19
54 MODELS FOR LONGITUDINAL STUDIES
3.4 Building growth curve models
Model for means
Polynomial models
The most commonly used method of tting growth curves is to use a polyno-
mial function to describe change:
Y
hij
(t) =
h0
+
h1
t
hij
+
h2
t
2
hij
+
hij
.
The purpose of the polynomial model is to approximate a curve that ts the
observed data. There is no biological or theoretical reason that growth of chil-
dren or HRQoL would follow a such a model. However, polynomial functions
can provide a good approximation of the average HRQoL. Figure 3.1A shows
how a polynomial model might appear in the renal cell carcinoma example.
Without CRA:

Y
1ij
(t) = 87.0 11.2t + 2.92t
2
0.272t
3
+ 0.0101t
4
0.000130t
5
. (3.5)
With CRA:

Y
2ij
(t) = 85.7 14.8t + 3.38t
2
0.279t
3
+ 0.0095t
4
0.000114t
5
. (3.6)
The higher-order terms, such as quadratic (t
2
) and cubic (t
3
) terms, allow the
curves to depart from linearity. However, when the model contains higher-
order terms, interpretation of the corresponding coecients is dicult. In
theory, the maximum number of terms (including the intercept) is equal to
the number of assessments. In the example, six is the maximum number of
terms that can be included. But with extensive dropout and mistiming of the
later observations, more parsimonious models may be necessary.
An overall test of postrandomization dierences is straightforward; we com-
pare the parameters across the H groups. But this test does not provide in-
formation about when the groups may be experiencing dierences in HRQoL.
Comparisons of the time-specic estimates are more useful. If, for example, in
the renal cell carcinoma study, we were interested in comparisons of HRQoL
shortly after beginning treatment (2 weeks or 0.46 months), well into treat-
ment (2 months), and 6 months after treatment began, we might construct
the estimates displayed in Table 3.6.
Piecewise linear regression
A useful alternative is a piecewise linear regression model. The change in
HRQoL is modeled as a linear function over short intervals of time. As in the
polynomial model, the model is not based on theory but the need to get a
reasonable t to the data. Thus, although we do not expect changes in HRQoL
to be strictly linear, it is reasonable to assume that changes are approximately
linear over short intervals of time. Figure 3.1B illustrates the use of a piecewise
linear model for the renal cell carcinoma study. In this study, we could consider
c2638 C2638TextureC03 January 28, 2002 17:19
BUILDING GROWTH CURVE MODELS 55
Figure 3.1. Growth curve models for the average FACT-TOI score among patients
treated on the renal cell carcinoma study [3]. (A) Polynomial model, (B) piecewise
linear regression.
c2638 C2638TextureC03 January 28, 2002 17:19
56 MODELS FOR LONGITUDINAL STUDIES
Table 3.6. Renal cell carcinoma example: Polynomial growth curve model.
Treatment Baseline 2-week mean 2-month mean 6-month mean
Model
CRA
10

5
0

1k
0.46
k

5
0

1k
2
k

5
0

1k
6
k
+CRA
20

5
0

2k
0.46
k

5
0

2k
2
k

5
0

2k
6
k
Estimates
CRA 90.3 78.3 72.3 81.3
+CRA 88.7 74.9 67.2 67.9
Dierence 1.6 3.3 5.0 13.4
Table 3.7. Renal cell carcinoma example: Piecewise linear regression model with
selected time-specic means.
Baseline 2-week 2-month 6-month
Treatment mean mean mean mean
Model
CRA
10

10
+
11
2
10
+
11
8.67
10
+
11
26
+
12
6.67 +
12
24
+
13
0.67 +
13
18
+CRA
20

20
+
21
2
20
+
21
8.67
20
+
21
26
+
22
6.67 +
22
24
+
32
0.67 +
23
18
Estimates
CRA 90.9 76.3 74.5 77.9
+CRA 89.0 73.3 66.0 72.0
Dierence 1.8 2.8 8.5 5.9
terms that allow the slope to change at 2, 8, 17, and 34 weeks (Table 3.7).
Y
hij
(t) =
h0
+
h1
t
hij
+
h2
t
[2]
hij
+
h3
t
[3]
hij
+
h4
t
[4]
hij
+
h5
t
[5]
hij
+
hij
,
t
[c]
hij
= max(t
hij
T
[c]
, 0),
T
[1]
= 0, T
[2]
= 2, T
[3]
= 8, T
[4]
= 17, T
[5]
= 34.
The selected points of change should correspond to the times when changes
in HRQoL might occur as the result of treatment or some other clinically
relevant process. In practice, the points in time where observations are clus-
tered will provide the most ecient estimates of change. Ideally, the study is
designed so that the points of clinical interest correspond to data collection.
Note that as with the polynomial model, we start with a maximum of six
unknown parameters in each group.
c2638 C2638TextureC03 January 28, 2002 17:19
BUILDING GROWTH CURVE MODELS 57
After a model-tting procedure, described in more detail in the next section,
the piecewise regression model for the renal cell study is
without CRA:

Y
1ij
(t) = 90.9 7.30t + 7.03t
[2]
+ 0.44t
[8]
(3.7)
and with CRA:

Y
2ij
(t) = 89.0 7.85t + 6.76t
[2]
+ 1.38t
[8]
. (3.8)
This can be interpreted as follows: The two groups start with average scores
of 91 and 89, respectively. There is a rapid decline during the rst 2 weeks,
which is slightly more rapid in the patients receiving CRA (7.85 vs. 7.30
points/week). This initial rapid decline ceases around week 2, with estimated
rates of decline between weeks 2 and 8 equal to 0.27* and 1.09** points per
week in the two arms, respectively. After approximately 8 weeks, HRQoL
begins to improve slightly in both arms (+0.17*** and +0.29 points per
week, respectively).
Covariance structure
Because there are multiple assessments on each subject, we expect the as-
sessments for each individual to be correlated over time. The most typical
approach for modeling the covariance structure uses a mixed-eects model.
The term mixed refers to the mixture of xed and random eects:
Y
hi
= X
hi

h
. .
Fixed eects
+ Z
hi
d
hi
. .
Random eects
+ e
hi
..
Residual error
. (3.9)
In the context of growth curve models, the xed eects (X
i
) model the
average HRQoL, whereas the random eects (Z
i
d
i
) model the variation among
individuals relative to the average. The xed eects are also referred to as
the mean response, the marginal expectation of the response [35], and the
average evolution [154, 155]. The covariance of a mixed-eects model for Y
hi
has the general structure:

hi
= Var(Y
hi
) = Var(Z
hi
d
hi
+ e
hi
)
= Var(Z
hi
d
hi
) + Var(e
hi
)
= Z
hi
D
h
Z

hi
+ R
hi
, (3.10)
where D
h
and R
hi
can have a number of structures.
Variance of random eects
The simplest random-eects model (Z
hi
d
hi
) has a single random eect (d
hi1
).
This model has a random intercept for each individual (Y
hi
= X
hi

h
+d
hi1
+
* Slope between weeks 2 and 8 for patient without CRA=
11
+
12
=7.30+7.030.27.
** Slope between weeks 2 and 8 for patient with CRA=
21
+
22
=7.85 + 6.76 1.09.
*** Slope after 8 weeks for patient without CRA=
11
+
12
+
13
=7.30 + 7.03 + 0.44.
Slope after 8 weeks for patient without CRA=
21
+
22
+
23
=7.85 + 6.76 + 1.38.
c2638 C2638TextureC03 January 28, 2002 17:19
58 MODELS FOR LONGITUDINAL STUDIES
e
hi
), where d
hi1
can be interpreted roughly as the average dierence between
the individuals response and the mean response. This implies the variation
between individuals does not change over time. Basically, the curves for each
individual are parallel:
Z
hi
= [1], D
h
= Var[d
hi1
] =
_

2
h1

. (3.11)
A more typical model for longitudinal studies has two random eects. The
second random eect (d
hi2
) allows variation in the rate of change over time
among individuals. This model has a random intercept and random slope for
each individual (Y
hi
= X
hi

h
+ d
hi1
+ d
hi2
t
hi
+ e
hi
). In most examples, the
two random eects are correlated:
Z
hi
= [1 t
hi
] , D
h
= Var
_
d
hi1
d
hi2
_
=
_

2
h1

h12

h21

2
h2
_
. (3.12)
In theory we can keep adding random eects, but in most cases the addition
of these two random eects is sucient to obtain a good approximation of the
covariance structure.
Variance of residual errors
The second part of the variance structure, R
hi
, can be as simple as
2
h
I or quite
complex (Table 3.8). The most common alternative to the simple structure is
a rst-order autoregressive error structure (AR(1)). Although this structure
is rarely suitable for the repeated measures setting, it can be useful when used
to model the residual errors in combination with random eects [76].
Table 3.8. Useful covariance structures for R
hi
in a mixed-eects model.
No. of parameters Structure
Simple (
2
I) 1
2
0 0 0
TYPE=SIMPLE
2
0 0

2
0

2
Autoregressive 2
2

2

2

3
Equal spacing
2

2

2

2
TYPE=AR(1)
2

2
Autoregressive 2
2

12

13

14
Unequal Spacing
2

23

24

rc
= |t
r
t
c
|
2

34
TYPE=SP(POW)(time var)
2
Toplitz 4
2

1

2

2

2

3
Equal Spacing
2

1

2

2
TYPE=TOEP
2

2
c2638 C2638TextureC03 January 28, 2002 17:19
BUILDING GROWTH CURVE MODELS 59
Note that certain structures, such as compound symmetry, described in
Table 3.1, cannot be used for modeling R
hi
. In the case of compound symme-
try, one of the parameters in D can be rewritten as a linear combination of
parameters in R. In fact, when Z
hi
= [1] and R =
2
I, the covariance struc-
ture is the same as compound symmetry when the random eect is omitted.
Finally, the covariance structures may dier among the treatment arms.
For example, there may be more variation in the rate of change of HRQoL
among patients receiving an active treatment than a placebo treatment.
SAS example of a polynomial growth curve model
Fully parameterized model for the means
In the renal cell carcinoma study, the datale EXAMPLE3 has a unique record for
each HRQoL assessment. CASEID identies the patient, TOI is the score for the
FACT-TOI, MNTH identies the time of the assessment relative to randomiza-
tion, and TREAT identies the treatment arm.* The CLASS statement identies
two levels of treatment. The terms in the MODEL statement create a design ma-
trix corresponding to the polynomial model displayed in Table 3.6, with the
default option for an intercept term suppressed by the NOINT option. As pre-
viously mentioned, the rst step is to specify a fully parameterized model for
the mean. The rst part of the SAS program for this model might appear as:
PROC MIXED DATA=EXAMPLE3 COVTEST IC METHOD=ML;
CLASS TREAT CASEID;
* Polynomial Model *;
MODEL TOI=TREAT
TREAT*MNTH
TREAT*MNTH*MNTH
TREAT*MNTH*MNTH*MNTH
TREAT*MNTH*MNTH*MNTH*MNTH
TREAT*MNTH*MNTH*MNTH*MNTH*MNTH
/NOINT SOLUTION PREDMEANS DDFM=SATTERTH;
Covariance structure
In the next step, we examine possible covariance structures. Table 3.9 illus-
trates some possible combinations of covariance structures for the random
eects and residual errors. The random eects contribution to the covariance
structure is dened by the RANDOM statement. For a model with a single ran-
dom eect that is homogeneous across groups (Models 1 and 2, Table 3.9),
the following statement denes Z
hi
and D
h
:
* Single random effect (Equal across groups) *;
RANDOM INTERCEPT /SUBJECT=CASEID TYPE=UN;
* The choice of months rather than weeks as a measure of time is a practical consideration.
By using the larger unit of time, we minimize the dierences in the magnitude of the
coecients for higher-order terms. This does not aect any of the statistics but facilitates
construction of tables reporting coecients. In this example, we avoid reporting coecients
with seven zeros after the decimal place.
c2638 C2638TextureC03 January 28, 2002 17:19
60 MODELS FOR LONGITUDINAL STUDIES
Table 3.9. Variance structures tested for growth curve models of the renal carcinoma
study (
hi
= Z
hi
D
h
Z

hi
+R
hi
).
Random effects (D
h
) Residual error (R
hi
)
Variance Total var.
model No. of effects Groups Structure Groups parameters
a
1 1 Equal
1
Simple Equal
3
1 +1 =2
2 1 Equal
1
AR(1)
b
Equal 1 +2 =3
3 1 Unequal
2
Simple Equal 2 +1 =3
4 1 Unequal
2
Simple Unequal
4
2 +2 =4
5 2 Equal
1
Simple Equal 3 +1 =4
6 2 Unequal
2
Simple Equal 6 +1 =7
1
D
1
= D
2
;
2
D
1
= D
2
;
3
R
1
= R
2
;
4
R
1
= R
2
.
a
# in D
h
+ # in R
h
= Total.
b
Autoregressive for unequal spacing of observations.
The term INTERCEPT denes Z
hi
= [1]. Since D
h
is a scalar, it is not necessary
to specify the structure. With the TYPE=UN option, the estimate of
2
h1
is
displayed as UN(1,1). In Model 1, the estimate of
2
h1
is 185.23.
The residual error contribution to the covariance structure is dened by the
REPEATED statement. When there is no RANDOM statement, as in Model 1, a sim-
ple homogeneous structure is assumed (R =
2
I) and displayed as Residual.
This is the specication for Models 1, 3, and 5 (see Table 3.9). In Model 1, the
estimate of the diagonal elements of R is 116.13 and the o-diagonal elements
are 0.
Covariance Parameter Estimates - Model 1
Standard Z
Cov Parm Subject Estimate Error Value Pr Z
UN(1,1) CASEID 185.23 23.9908 7.72 <.0001
Residual 116.13 8.1905 14.18 <.0001
An autoregressive error structure for unequal spacing of observations (Model
2) is specied using the term TYPE=SP(POW)(QSDAY), where QSDAY is the num-
ber of days since randomization. Note that QSDAY has a unique value for each
HRQoL assessment within a particular patient.
* Autoregressive structure for unequal spacing of Observations *;
REPEATED /SUBJECT=CASEID TYPE=SP(POW)(QSDAY);
The estimate of the additional parameter is displayed as SP(POW). Thus, the
estimate of the diagonal elements of R are 122.13 and the o-diagonal elements
(r, c) are
2

rc
or 122.13 0.8979

rc
, where
rc
is the absolute dierence in
the number of days QSDAY between the rth and cth assessment.
c2638 C2638TextureC03 January 28, 2002 17:19
BUILDING GROWTH CURVE MODELS 61
Table 3.10. Covariance structures tested for the renal carcinoma study.
Variance Polynomial Piecewise
Model Parameters 2 log L AIC 2 log L AIC
1 2 5162.4 5190.4 5154.2 5182.2
2 3 5159.7 5189.7 5148.3 5178.3
3 3 5162.3 5193.0 5154.1 5184.1
4 4 5162.1 5194.9 5154.0 5186.0
5 4 5160.4 5192.4 5152.2 5184.2
6 7 5157.9 5193.9 5150.5 5188.5
Covariance Parameter Estimates - Model 2
Standard Z
Cov Parm Subject Estimate Error Value Pr Z
UN(1,1) CASEID 179.12 24.6260 7.27 <.0001
SP(POW) CASEID 0.8978 0.03726 24.10 <.0001
Residual 122.13 10.0471 12.16 <.0001
By adding GROUP=TREAT to the RANDOM statement, we request dierent es-
timates of D
h
in each treatment arm (Models 3 and 4, Table 3.9).
* Single random effect (Unequal across groups) *;
RANDOM INTERCEPT/SUBJECT=CASEID TYPE=UN GROUP=TREAT;
In Model 3 we have separate estimates for
2
h1
in the two treatment groups,
195.85 and 176.90, respectively. If this model represented a signicant im-
provement over Model 1, it would suggest that there is more variation among
subjects in the IFN+CRA group than in the IFN-only group. The statis-
tics displayed in Table 3.10 suggest this is not the case.
Covariance Parameter Estimates - Model 3
Standard Z
Cov Parm Subject Group Estimate Error Value Pr Z
UN(1,1) CASEID TREAT IFN + CRA 194.85 36.2256 5.38 <.0001
UN(1,1) CASEID TREAT IFN only 176.90 31.2410 5.66 <.0001
Residual 116.03 8.1815 14.18 <.0001
Addition of GROUP=TREAT to the REPEATED statement species dierent es-
timates of the variance parameter in each treatment group (R
hi
=
2
h
I), as in
Model 4.
* Simple uncorrelated random errors (Unequal across groups) *;
REPEATED/SUBJECT=CASEID TYPE=SIMPLE GROUP=TREAT;
The separate estimates of
2
wh
in the two groups are 119.80 and 112.41, re-
spectively. Again, there is no evidence that there is more or less variation
between the two groups (see Table 3.10).
c2638 C2638TextureC03 January 28, 2002 17:19
62 MODELS FOR LONGITUDINAL STUDIES
Covariance Parameter Estimates - Model 4
Standard Z
Cov Parm Subject Group Estimate Error Value Pr Z
UN(1,1) CASEID TREAT IFN + CRA 192.35 36.6248 5.25 <.0001
UN(1,1) CASEID TREAT IFN only 178.64 31.5004 5.67 <.0001
Residual CASEID TREAT IFN + CRA 119.80 12.0797 9.92 <.0001
Residual CASEID TREAT IFN only 112.41 11.0939 10.13 <.0001
A second random eect, allowing variation in the rate of change over time
among subjects, can be added to the random-eects model. The following
statement denes a model with two correlated random eects (Models 5 and 6,
Table 3.9) as dened in Equation 3.12.
* Two correlated random effects *;
RANDOM INTERCEPT MNTH/SUBJECT=CASEID TYPE=UN;
The second random eect is added to the RANDOM statement, where MNTH de-
nes Z
hi2
= t
hi
. TYPE=UN species an unstructured covariance of the random
eects with three covariance parameters, UN(1,1), UN(1,2), and UN(2,2),
which correspond to
2
1
,
1,2
, and
2
2
.
Covariance Parameter Estimates - Model 5
Standard Z
Cov Parm Subject Estimate Error Value Pr Z
UN(1,1) CASEID 181.89 25.1413 7.23 <.0001
UN(2,1) CASEID 1.2952 3.4955 0.37 0.7110
UN(2,2) CASEID 0.3870 0.4216 0.92 0.1793
Residual 111.64 8.7304 12.79 <.0001
Finally, we attempt to allow these random eects to vary across the treat-
ment groups by adding a GROUP=TREAT to the random statement. Note that
the estimate of the variation in the slopes in the IFN-only group UN(2,2) is
close to zero and the algorithm fails to obtain an estimate.*
Covariance Parameter Estimates - Model 6
Standard
Cov Parm Subject Group Estimate Error
UN(1,1) CASEID TREAT IFN + CRA 203.18 39.0561
UN(2,1) CASEID TREAT IFN + CRA -3.5354 5.3937
UN(2,2) CASEID TREAT IFN + CRA 0.5752 0.5891
UN(1,1) CASEID TREAT IFN only 158.36 31.0568
UN(2,1) CASEID TREAT IFN only 6.9927 4.5872
UN(2,2) CASEID TREAT IFN only 0 .
Residual 112.71 8.2840
* An alternative parameterization TYPE=FA0(2) ensures that the estimates of the variance
of the random eects is non-negative denite. This is particularly helpful when there is
the possibility that one of the variance parameters is close to zero.
c2638 C2638TextureC03 January 28, 2002 17:19
BUILDING GROWTH CURVE MODELS 63
Table 3.10 summarizes the results of testing the six covariance structures.
Based on the AIC, Model 2 is the best model. However, there is very little
dierence between the AIC values for Models 1 and 2, and a formal test of
the two additional parameters in Model 2 is nonsignicant. Either model will
provide a good approximation of
i
.
Model reduction
The next step is to consider simplifying the model for the means. A por-
tion of the SAS output from the fully parameterized model follows. Neither
the estimates nor the test statistics suggest a strategy for model reduction.
Elimination of the highest-order terms does not improve the model t as mea-
sured by the AIC (5781.3 vs. 5770.6), so we will continue with the 5th degree
model.
Solution for Fixed Effects
Effect TREAT Estimate Std Err t Pr > |t|
TREAT 0 89.77 1.91 46.94 0.0001
TREAT 1 87.95 1.93 45.39 0.0001
MNTH*TREAT 0 -28.44 3.78 -7.51 0.0001
MNTH*TREAT 1 -30.93 3.75 -8.24 0.0001
MNTH*MNTH*TREAT 0 13.85 2.58 5.36 0.0001
MNTH*MNTH*TREAT 1 14.73 2.62 5.62 0.0001
MNTH*MNTH*MNTH*TREAT 0 -2.50 0.64 -3.89 0.0001
MNTH*MNTH*MNTH*TREAT 1 -2.90 0.66 -4.40 0.0001
MNTH*MNT*MNT*MNT*TREAT 0 0.18 0.066 2.85 0.0047
MNTH*MNT*MNT*MNT*TREAT 1 0.25 0.068 3.70 0.0003
MNTH*MNT*MNT*MN*MN*TRE 0 -0.005 0.002 -2.11 0.0357
MNTH*MNT*MNT*MN*MN*TRE 1 -0.008 0.002 -3.25 0.0013
Tests of Fixed Effects
Source NDF DDF Type III F Pr > F
TREAT 2 280 2131.99 0.0001
MNTH*TREAT 2 280 62.18 0.0001
MNTH*MNTH*TREAT 2 280 30.14 0.0001
MNTH*MNTH*MNTH*TREAT 2 280 17.23 0.0001
MNTH*MNTH*MNTH*MNTH*TREAT 2 280 10.90 0.0001
MNTH*MNTH*MNTH*MNTH*MNTH*TREAT 2 280 7.50 0.0007
Estimation
As previously mentioned, it is dicult to interpret the polynomial terms di-
rectly. Estimating the means at specic points in time, as shown in Table 3.6,
is often more informative.

h
(t) =
5

hk
t
k
. (3.13)
The estimates of the means for each treatment by time combination generated
by the ESTIMATE statements is illustrated below:
c2638 C2638TextureC03 January 28, 2002 17:19
64 MODELS FOR LONGITUDINAL STUDIES
* Group Specific Estimates at selected times *;
ESTIMATE 2 weeks -CRA
TREAT 1 0
TREAT*MNTH 0.462 0
TREAT*MNTH*MNTH 0.213 0
TREAT*MNTH*MNTH*MNTH 0.0983 0
TREAT*MNTH*MNTH*MNTH*MNTH 0.0454 0
TREAT*MNTH*MNTH*MNTH*MNTH*MNTH 0.0209 0
TREAT*MNTH*MNTH*MNTH*MNTH*MNTH*MNTH 0.00967 0;
Alternatives include the change from baseline

h
(t)
h
(t = 0) =
5

hk
t
k

h0
=
5

hk
t
k
. (3.14)
* Group Specific Estimates at selected times *;
ESTIMATE 2 weeks change from baseline -CRA
TREAT*MNTH 0.462 0
TREAT*MNTH*MNTH 0.213 0
TREAT*MNTH*MNTH*MNTH 0.0983 0
TREAT*MNTH*MNTH*MNTH*MNTH 0.0454 0
TREAT*MNTH*MNTH*MNTH*MNTH*MNTH 0.0209 0
TREAT*MNTH*MNTH*MNTH*MNTH*MNTH*MNTH 0.00967 0;
Hypothesis testing
An overall test of postrandomization treatment dierences can be constructed
using the CONTRAST statement illustrated below.
H
0
:
1k
=
2k
, k = 1, . . . , 5. (3.15)
*** Overall Post Baseline Treatment Effect ***;
CONTRAST Overall Diff
TREAT*MNTH 1 -1,
TREAT*MNTH*MNTH 1 -1,
TREAT*MNTH*MNTH*MNTH 1 -1,
TREAT*MNTH*MNTH*MNTH*MNTH 1 -1,
TREAT*MNTH*MNTH*MNTH*MNTH*MNTH 1 -1,
TREAT*MNTH*MNTH*MNTH*MNTH*MNTH*MNTH 1 -1;
The results appear as follows:
CONTRAST Statement Results
Source NDF DDF F Pr > F
Overall Diff 6 485 1.94 0.0732
Tests of dierences between the groups at a specic time (t) can be constructed
using an ESTIMATE statement:
H
0
:
5

1k
t
k
=
5

2k
t
k
. (3.16)
*** Test Differences at Specific Times ***;
ESTIMATE Baseline Diff TREAT 1 -1;
ESTIMATE 2 weeks Diff TREAT 1 -1
c2638 C2638TextureC03 January 28, 2002 17:19
BUILDING GROWTH CURVE MODELS 65
TREAT*MNTH 0.462 -0.462
TREAT*MNTH*MNTH 0.213 -0.213
TREAT*MNTH*MNTH*MNTH 0.0983 -0.0983
TREAT*MNTH*MNTH*MNTH*MNTH 0.0454 -0.0454
TREAT*MNTH*MNTH*MNTH*MNTH*MNTH 0.0209 -0.0209
TREAT*MNTH*MNTH*MNTH*MNTH*MNTH*MNTH 0.00967 -0.00967;
Finally, contrasts of changes from baseline are tested. The intercept terms
cancel (see Equation 3.14) and are dropped:
H
0
:
5

1k
t
k
=
5

2k
t
k
. (3.17)
*** Test Differences in Change from Baseline at Specific Times ***;
ESTIMATE 2 weeks change Diff
TREAT*MNTH 0.462 -0.462
TREAT*MNTH*MNTH 0.213 -0.213
TREAT*MNTH*MNTH*MNTH 0.0983 -0.0983
TREAT*MNTH*MNTH*MNTH*MNTH 0.0454 -0.0454
TREAT*MNTH*MNTH*MNTH*MNTH*MNTH 0.0209 -0.0209
TREAT*MNTH*MNTH*MNTH*MNTH*MNTH*MNTH 0.00967 -0.00967;
SAS example of a piecewise linear regression model
Fully parameterized model for the means
In the same renal cell carcinoma study, we construct ve new variables to
model possible changes in the slope at 2, 8, 17, and 34 weeks.
DATA EXAMPLE3;
SET EXAMPLE3;
* Piecewise regression terms *;
WEEKS2=MAX(WEEKS-2,0);
WEEKS8=MAX(WEEKS-8,0);
WEEKS17=MAX(WEEKS-17,0);
WEEKS34=MAX(WEEKS-34,0);
RUN;
The rst part of the SAS program for this model might appear as
PROC MIXED DATA=EXAMPLE3 COVTEST IC;
CLASS TREAT CASEID;
* Piecewise Regression Model *;
MODEL TOI=TREAT TREAT*WEEKS TREAT*WEEKS2 TREAT*WEEKS8
TREAT*WEEKS17 TREAT*WEEKS34
/NOINT SOLUTION DDFM=SATTERTH;
Covariance structure
The procedure for examining the covariance structure is exactly the same as
outlined for the polynomial model. The results for the piecewise regression
model are summarized in Table 3.10 and are very similar to those found for
the polynomial model.
c2638 C2638TextureC03 January 28, 2002 17:19
66 MODELS FOR LONGITUDINAL STUDIES
Model reduction
In contrast to the polynomial model, the strategy for simplifying the piecewise
regression model is straightforward. The following SAS output comes from the
fully parameterized model. The results indicate that there is no evidence of
a change in the slopes after 34 weeks (
k5
= 0) and that this term can be
dropped. Notice that this model reduction is performed in a stepwise manner,
as we are relying on Type III test statistics. Although the terms involving
WEEKS17 appear to be nonsignicant, the test of the WEEKS17*TREAT is con-
ditional on inclusion of WEEKS34*TREAT and the result may be quite dierent
when these terms are removed.
Tests of Fixed Effects
Source NDF DDF Type III F Pr > F
TREAT 2 280 2088.76 0.0001
WEEKS*TREAT 2 280 61.28 0.0001
WEEKS2*TREAT 2 280 31.82 0.0001
WEEKS8*TREAT 2 280 3.37 0.0358
WEEKS17*TREAT 2 280 0.78 0.4585
WEEKS34*TREAT 2 280 0.11 0.8948
After retting the model without the WEEKS34*TREAT terms, it is obvious
that the WEEKS17*TREAT terms can also be dropped (AIC=5757.5). The
nal model (See Figure 3.1B) allows the slope to change at 2 and 8 weeks
(AIC=5172.1).
Tests of Fixed Effects
Source NDF DDF Type III F Pr > F
TREAT 2 284 2086.56 0.0001
WEEKS*TREAT 2 284 61.00 0.0001
WEEKS2*TREAT 2 284 33.45 0.0001
WEEKS8*TREAT 2 284 6.58 0.0016
Estimation
The terms in the MODEL statement create a design matrix corresponding to the
polynomial model displayed in Table 3.7. The estimates of the means for each
treatment by time combination are generated by the ESTIMATE statements.
PROC MIXED DATA=WORK.EXAMPLE3;
* Group Specific Estimates at selected times *;
ESTIMATE 2 weeks -CRA TREAT 1 0
TREAT*WEEKS 2 0;
ESTIMATE 2 months -CRA TREAT 1 0
TREAT*WEEKS 8.67 0
TREAT*WEEKS2 6.67 0
TREAT*WEEKS8 0.67 0;
c2638 C2638TextureC03 January 28, 2002 17:19
BUILDING GROWTH CURVE MODELS 67
Testing
An overall test of postrandomization treatment dierences can be constructed
using a CONTRAST statement illustrated above.
H
0
:
1k
=
2k
, k = 1, . . . , 3. (3.18)
*** Overall Post Baseline Treatment Effect ***;
CONTRAST Overall Diff TREAT*WEEKS 1 -1,
TREAT*WEEKS2 1 -1,
TREAT*WEEKS8 1 -1;
CONTRAST Statement Results
Source NDF DDF F Pr > F
Overall Diff 4 517 1.45 0.2161
Tests of dierences between the groups at specic times (t) can be con-
structed using the ESTIMATE statement:
H
0
:
4

1k
t
[k]
=
4

2k
t
[k]
. (3.19)
*** Test Differences at Specific Times ***;
ESTIMATE Baseline Diff TREAT 1 -1;
ESTIMATE 2 weeks Diff TREAT 1 -1
TREAT*WEEKS 2 -2;
ESTIMATE 2 month Diff TREAT 1 -1
TREAT*WEEKS 8.67 -8.67
TREAT*WEEKS2 6.67 -6.67
TREAT*WEEKS8 0.67 -0.87;
Tests for dierences in the change from baseline can also be constructed by
dropping the intercept term TREAT from the ESTIMATE statement:
H
0
:
4

1k
t
[k]

10
=
4

2k
t
[k]

20
. (3.20)
ESTIMATE 2-week change Diff TREAT*WEEKS 2 -2 ;
ESTIMATE 2-month change Diff TREAT*WEEKS 8.67 -8.67
TREAT*WEEKS2 6.67 -6.67
TREAT*WEEKS8 0.67 -0.67;
ESTIMATE 6-month change Diff TREAT*WEEKS 26 -26
TREAT*WEEKS2 24 -24
TREAT*WEEKS8 18 -18;
c2638 C2638TextureC03 January 28, 2002 17:19
68 MODELS FOR LONGITUDINAL STUDIES
3.5 Summary
Event-driven designs are generally associated with repeated measures
models.
Time-driven designs are associated with growth curve models.
The recommended model building process starts by dening a fully param-
eterized model for the means (X
i
), then identifying the structure of
i
,
and nally simplifying the mean structure.
c2638 C2638TextureC04 February 12, 2002 11:36
CHAPTER 4
Missing Data
4.1 Introduction
Missing self-assessments of HRQoL occur in most longitudinal studies. Be-
cause the patients participating in these trials are free-living individuals and
often experience disease- and treatment-related morbidity and mortality,
missing assessments are inevitable. Missing HRQoL data may occur for rea-
sons totally unrelated to the subjects current quality of life (a missed ap-
pointment due to a snowstorm) or may be intimately related (severe episodes
of nausea or vomiting). The term missing is sometimes used to include only
assessments expected according to the guideline of the protocol; another term
used for this narrow denition is noncompliance. The term missing often has
a broader meaning to include missing for any reason, including unavoidable
attrition due to death or termination of follow-up as specied in the protocol.
Terminology
This book uses the term missing to include both attrition and noncompliance.
The term dropout will refer to the discontinuation of the outcome measures,
Y
i
. The term censoring will refer to incomplete follow-up of the time to an
event associated with dropout, T
D
i
; see Table 4.1.
Why are missing data a problem?
The rst potential problem associated with missing data is loss of power to de-
tect clinically meaningful dierences as a result of a reduced number of obser-
vations. However, in many large (Phase III/IV) clinical studies, the sample size
is often based on dierences in a binary outcome such as survival. These sam-
ple sizes are generally more than sucient to detect clinically meaningful dif-
ferences in the HRQoL endpoints. The loss of power may be an issue in small
studies, but the problem can be addressed by increasing the planned sample
size of the trial. If the objectives associated with the HRQoL assessment are
not of sucient importance to warrant this increase in sample size, then the in-
vestigators should consider omitting the HRQoL assessments altogether rather
than burdening patients when the design is inadequate in terms of sample size.
Why are missing data a problem?
Fewer Observations Loss of Power
Nonrandomly Missing Data Bias
c2638 C2638TextureC04 February 12, 2002 11:36
70 MISSING DATA
Table 4.1. Summary of terms.
Attrition No assessment due to termination of HRQoL as-
sessment as specied in protocol
Noncompliance No assessment but expected per protocol
Missing No assessment when information at time of the
scheduled assessment was not obtained, regard-
less of whether the assessment was required per
protocol
Dropout Discontinuation of the observation of the outcome
Censoring Discontinuation of observation of the time to an
event associated with dropout
The second problem is the potential bias of estimates due to missing data.
For example, patients who are experiencing poorer HRQoL because of in-
creased morbidity or mortality may be less likely to complete HRQoL assess-
ments as a result of treatment-related toxicity, progressing disease, or death.
If we ignore the presence of missing data and the analyses are based only on
the observed data of patients who are doing well, HRQoL is overestimated.
Alternatively, it is possible that individuals will drop out of a study when they
are no longer experiencing the signs or symptoms of their diseases, in which
case HRQoL is underestimated.
Why subjects fail to complete HRQoL assessments
Patient misses a specied appointment
Sta forgets to administer questionnaire
Translation not available in patients language
Patient drops out from study because no longer ill
Patient refuses to complete questionnaire
Patient states he or she is too ill to complete questionnaire
Sta does not approach patient because of health status
Patient dies
How much data can be missing?
There are no magic rules about how much missing data is acceptable in a clin-
ical trial. When the proportion of missing assessments is very small (<5%),
the potential impact on power or bias may be very minor. In some cases, 10
to 20% missing data will have little or no eect on the results of the study.
In other studies, 10 to 20% may matter. As the proportion of missing data
increases to 50%, the conclusions one is willing to draw will be restricted.
The seriousness of the problem depends on the reasons for the missing data,
the objectives of the study, and the intended use of the results. For example,
c2638 C2638TextureC04 February 12, 2002 11:36
PATTERNS OF MISSING DATA 71
results from trials inuencing regulatory issues of drug approval or health
policy decisions may require more stringent criteria than those used to design
future studies. The challenge for the analyst is to provide either a convinc-
ing argument that conclusions drawn from the analysis are insensitive to the
missing data or clear limits for interpretation of the results. The concepts and
tools to do that are the focus of this and the following chapters.
Similar patterns of dropout among intervention arms
A common question is whether missing data can be ignored if there are sim-
ilar proportions of missing data across all study arms. Although experience
indicates that comparisons between treatment arms are often less sensitive
than the estimates of change over time to missing data, there is always the
possibility that this is not a safe assumption.
Prevention
Although analytic strategies exist for missing data, their use is much less satis-
factory than initial prevention. Some missing data, such as that due to death,
is not preventable. However, both primary and secondary prevention are de-
sirable. In terms of primary prevention, missing data should be minimized at
both the design and implementation stages of a clinical trial [37, 171]. Strate-
gies are discussed in Chapter 2. Secondary prevention consists of gathering
information that is useful in the analysis and interpretation of the results.
This includes collection of data on factors that may contribute to missing as-
sessments and data that are likely to predict the missing HRQoL measures.
Thus, one should prospectively document reasons for missing data. The clas-
sications that are used should be specied in a manner that helps the analyst
decide whether the reason is related to the individuals HRQoL. For example,
Patient refusal does not clarify this, but reasons such as Patient refusal due
to poor health and Patient refusal unrelated to health will be useful. Other
strategies for secondary prevention may include gathering concurrent data on
toxicity, evaluations of health status by the clinical sta, or HRQoL assess-
ments from a caretaker. Use of this type of data is discussed in later chapters.
4.2 Patterns of missing data
Patterns of missing data can be considered either terminal dropout (mono-
tone) or intermittent (nonmonotone). Examples of both patterns are shown
in Table 4.2. In the rst case, an observation on a patient may be missing, but
a subsequent observation is observed. In the second case, no observations are
made on a patient after a certain point in time. Within these two patterns of
missing data, the causes of missing data are either related or unrelated to the
patients HRQoL. Similarly, the causes are planned (by design) or unplanned.
In patients with chronic diseases, terminal dropout may occur because
a treatment works or the patient is no longer experiencing the signs and
c2638 C2638TextureC04 February 12, 2002 11:36
72 MISSING DATA
Table 4.2. Terminal dropout and intermittent missing data.
Terminal dropout Intermittent
(Monotone) (Nonmonotone)
X X X X X X X X X X
X X X X X X X X
X X X X X X X
X X
.
.
.
X X X X
symptoms that prompted him or her to obtain treatment. In patients with
progressive or fatal diseases, such as advanced cancer and AIDS, there are con-
siderable missing HRQoL data or dropout due to death. Intermittent missing
data might include missed clinical visits because of weather or episodes of se-
vere acute toxicity. Causes such as the weather are unrelated to the subjects
HRQoL, whereas causes such as toxicity are related to the subjects QoL.
Example: NSCLC study
In most studies both patterns exist, possibly within the same patient. In
the NSCLC study (Table 4.3), the majority of missing data ts the terminal
dropout pattern (76% of patients, including complete cases and those with no
data). A smaller proportion of subjects (14%) have an intermittent or mixed
pattern (intermittent followed by dropout).
4.3 Mechanisms of missing data
Understanding the missing data mechanism is critical to the analysis and in-
terpretation of HRQoL studies. All possible analytic methods for longitudinal
data depend on assumptions about the missing data. The robustness of ana-
lytic methods to violations of the assumptions varies. If the assumptions are
incorrect, then the resulting estimates will be biased. Thus, examination of
available evidence for or against these assumptions is the rst major task in
the analysis of HRQoL studies.
Notation
Consider a longitudinal study where n
i
assessments of HRQoL are planned
for each subject over the course of the study. As dened in Chapter 3, Y
ij
indicates the jth observation of HRQoL on the ith individual.
Y
i
= the complete data vector of n
i
planned observations of the out-
come for the ith individual, which includes both the observed
data (Y
obs
i
) and missing observations (Y
mis
i
) of HRQoL.
c2638 C2638TextureC04 February 12, 2002 11:36
MECHANISMS OF MISSING DATA 73
Table 4.3. Patterns of observations in the NSCLC study.
Assessment No.
Pattern 1 2 3 4 N %
Complete case X X X X 146 25.4
Monotone dropout X X X 98 17.0
X X 94 16.4
X 133 23.1
No data 26 4.5
Intermittent X X X 15 2.6
X X X 11 1.9
X X 11 1.9
X X X 5 0.9
X X 2 0.4
X 1 0.2
Mixed X X 18 3.1
X X 3 0.5
X 11 1.9
X 1 0.2
X indicates the HRQoL assessment completed.
R
i
= a vector of indicators of the missing data pattern for the ith
individual, where R
ij
= 1 if Y
ij
is observed and R
ij
= 0 if Y
ij
is missing.
Note that the term complete data is dened as the set of responses that one
would have observed if all subjects completed all possible assessments. This is
in contrast to the term complete cases, which is dened as the set of responses
on only those subjects who completed all possible assessments (R
ij
= 1 for all
possible Y
ij
from the ith subject). In some of the following chapters, we will
dierentiate data from these complete cases (Y
C
i
) and data from incomplete
cases (Y
I
i
). Table 4.4 is a summary of terms.
Example
If a study were planned to have four HRQoL assessments at 0, 4, 13, and 26
weeks and the HRQoLs of the ith subject were missing at 4 and 26 weeks,
then that subjects data might look like
Y
i
=

78
NA
58
NA

and R
i
=

1
0
1
0

or
Y
obs
i
=

78
58

Y
mis
i
=

NA
NA

,
c2638 C2638TextureC04 February 12, 2002 11:36
74 MISSING DATA
Table 4.4. Summary of terms.
Complete data Y All responses that one would have ob-
served if all subjects had completed all
possible assessments
Observed data Y
obs
All responses that were observed
Missing data Y
mis
All responses that were not observed
Complete cases Y
C
Responses from cases (or subjects) where
all possible responses were observed
Incomplete cases Y
I
Responses from cases (or subjects) where
one or more possible responses are
missing
where NA indicates a missing observation. Notice that the vector of the ob-
served data (Y
obs
i
) contains the rst (78) and third (58) observations and the
vector of the missing data (Y
mis
i
) contains the second and fourth observations.
The corresponding design matrix for this subject in a model estimating an
intercept and a linear trend over time would appear as
X
i
=

1 0
1 4
1 13
1 26

and
X
obs
i
=

1 0
1 13

X
mis
i
=

1 4
1 26

.
Even though the dependent variable Y
i
is missing at some time points, X
i
is
fully observed if one assumes that the time of the observation is the planned
time of the second and fourth HRQoL assessments.
The concept
There are three major classes of missing data, diering by whether the rea-
sons for missingness are related to the subjects quality of life. For example, if
HRQoL data are missing because the patient moved out of town or the sta
forgot to administer the assessment, the data are Missing Completely at Ran-
dom (MCAR). At the other extreme, if data are missing because of increased
toxicity, progressive disease, and death, they are nonignorable or Missing Not
at Random (MNAR). Intermediate cases are referred to as Missing at Random
(MAR). In this case, missingness depends on the subjects observed HRQoL,
generally the most recently observed HRQoL. Table 4.5 is an overview.
The remainder of this chapter presents the general concepts of these three
mechanisms in more detail and suggests methods for distinguishing among
them. Formal statistical denitions are presented in Appendix III. The sub-
sequent chapters describe methods of analysis that can be used under the
various assumptions.
c2638 C2638TextureC04 February 12, 2002 11:36
MECHANISMS OF MISSING DATA 75
Table 4.5. Simple overview of missing data mechanisms.
Dependent on. . . Independent of . . .
MCAR
a
Covariates Observed HRQoL
Missing HRQoL
MAR Covariates Missing HRQoL
Observed HRQoL
MNAR Missing HRQoL
a
And covariate-dependent dropout.
MCAR: Missing completely at random
The concept
The strongest assumption about the missing data mechanism is that the data
are missing completely at random (MCAR). The basic assumption is that
the reason for a missing HRQoL assessment is entirely unrelated to HRQoL.
Examples might include a patient moving or sta forgetting to provide the
assessment. (Note that this accounted for about 2% of the missing data in Ex-
ample 1 and 8.5% in Example 2.) Assessments might be changed as a result of
appointments delayed because of scheduling problems. Finally, some patients
may not be able to participate because translations of the questionnaire are
not available in their languages.
Covariate-dependent dropout
Dropout solely dependent on treatment, but constant within each treatment
group irrespective of the outcome (Y
i
), is also MCAR [64] conditional on treat-
ment. If this assumption is true, then the patients with missing data are a
random sample of all patients in each treatment group. Thus, there is no bias
attributable to missing data. In the context of HRQoL studies, it is dicult
to imagine scenarios where this is true. Dropout related to side eects or lack
of eectiveness is likely to aect HRQoL and would not t this condition. The
missingness may also be associated with covariates (X
i
) measured prior to
randomization or treatment assignment. An example of covariate-dependent
dropout is the cultural inuences on follow-up observed in an international
trial of breast cancer patients [116], where the proportion of missing assess-
ments diered among the cultural groups. When these covariates are included
in the analysis, the resulting estimates are adjusted for the covariate-dependent
dropout.
In summary, while some missingness of HRQoL assessments is completely
random in clinical trials, it is rare that this assumption holds for the majority
of the missing data.
c2638 C2638TextureC04 February 12, 2002 11:36
76 MISSING DATA
Identifying covariate-dependent missingness
The rst step in the process of exploring the missing data process is to identify
covariates that predict missing observations. One method is to examine the cor-
relation between an indicator of missing observations (r
ijk
) and possible co-
variates. Kendals
b
is useful when one variable is dichotomous, such as the
missing data indicator, and the other variables are a mixture of dichotomous,
ordered categorical, and continuous variables. A second approach is to perform
a series of logistic regression analyses to model the probability of a missing
assessment. From a practical standpoint, patient characteristics (covariates)
that are likely to be also related to the subjects HRQoL should be given
priority. The reasons for this will become more obvious in the later sections
of this chapter and later chapters.
Example: NSCLC study
In the NSCLC study, missing assessments were weakly associated with phys-
ical characteristics such as older age, male gender, performance status, and
disease symptoms at the time of randomization. Similar results were obtained
using either bivariate correlation (Table 4.6) or multivariate logistic regression
(Table 4.7).
In most studies, because prerandomization covariates are not directly re-
lated to the reasons for dropout, they are likely to be only weakly predic-
tive of missing data. We see this in the NSCLC study. Although statistically
signicant, none of these covariates individually explains more than 2% of
the variation in the likelihood of a missing assessment as indicated by the
squared value of the correlation coecients displayed in Table 4.6. It should
be noted that this procedure is an exploratory rst step in understanding
the missing data mechanism. Because of the number of comparisons, there is
Table 4.6. Correlation (Kendall
b
) of missing assessments with baseline covariates.
Positive correlations indicate increased likelihood of missing assessment.
Correlation coecients
Assessment no.
1 2 3 4
Older age at randomization 0.00 0.09

0.04 0.07

Poorer performance status 0.01 0.12

0.13

0.09

Metastatic disease symptoms 0.01 0.00 0.06 0.10

Systemic disease symptoms 0.01 0.09

0.08

0.04
Randomized to Taxol regimen 0.01 0.13

0.07 0.06
Nonsignicant correlations: gender, weight loss in last 6 months, primary disease
symptoms, associated chronic diseases, prior radiation therapy, assigned to G-CSF
arm.

p < 0.05,

p < 0.01,

p < 0.001.
c2638 C2638TextureC04 February 12, 2002 11:36
MECHANISMS OF MISSING DATA 77
Table 4.7. Prerandomization patient characteristics associated with missing HRQoL
assessments.
Assessment Characteristic OR (95% CI)
Baseline 5% weight loss prior 6 months 0.72 (0.52, 1.00)
6 weeks Older age 1.24
a
(1.02, 1.50)
Poorer performance status 1.72 (1.16, 2.53)
Taxol therapy 0.55 (0.38, 0.79)
12 weeks Poorer performance status 1.71 (1.20, 2.44)
26 weeks Older age (decades) 1.25 (1.04, 1.50)
Systemic symptoms 1.53 (1.08, 2.19)
Notes: Odds ratios (OR) and 95% condence intervals (CI) from logistic re-
gression models of the probability of a missing observation at each of the four
planned assessments. Associations identied using stepwise logistic regression
with a selection criteria of 0.05.
Potential covariates include age, gender, 5% weight loss in 6 months prior
to randomization, primary disease symptoms, metastatic disease symptoms, sys-
temic disease symptoms, associated chronic diseases at the time of randomiza-
tion, and prior radiation therapy.
a
Age in years at randomization with OR estimated for 10-year dierence.
the possibility of spurious correlations. This may be the explanation for the
counterintuitive association of less missing data at baseline with prior weight
loss.
Analytic methods
In general, analytic methods that assume the data are MCAR either exclude
data for some individuals, such as methods that require complete cases for
analysis (MANOVA), or ignore data obtained at other points in time, such
as repeated univariate tests [105]. Detailed examples of these are presented in
the following chapter.
MAR: Missing at random
The concept
The assumptions about the missing data mechanism are relaxed slightly when
we assume the data are missing at random (MAR). Specically, the missing-
ness depends on observed HRQoL after conditioning on covariates. In most
cases this is the initial (baseline) or previous HRQoL assessment, but it could
possibly include subsequent HRQoL assessments. It is not uncommon for miss-
ing HRQoL data to depend on covariates and observed HRQoL scores. In
most settings, individuals who are experiencing better health and HRQoL
will generally be more likely to be available for follow-up and more likely to
c2638 C2638TextureC04 February 12, 2002 11:36
78 MISSING DATA
complete the self-assessments required for the evaluation of HRQoL. However,
there are exceptions, such as individuals with an acute condition who drop
out once they become well.
It is possible that data are MAR by design. Murray and Findlay [105] de-
scribe this in hypertension trials, where a subject is withdrawn from a trial by
design because of inadequate blood pressure control. In this example, dropout
is a function of a previously observed value of the outcome.
Identication of dependence on observed data (Y
obs
i
)
Dierences between MCAR and MAR can be tested by examining the as-
sociation of missing data with observed HRQoL scores. Various methods in-
clude graphical presentations, formal tests, correlational analyses, and logistic
regression.
A graphical approach is displayed in Figure 4.1 for the NSCLC study. This
plot displays the average observed FACT-Lung TOI in groups of patients
dened by their pattern of missing data. Because the total number of pat-
terns is large and the number of subjects with intermittent or mixed patterns
is small (see Table 4.3), this plot was simplied by grouping subjects by the
time of their last HRQoL assessment. The resulting gure suggests two things.
First, subjects who dropped out earlier had poorer FACT-Lung TOI scores
at baseline. Second, FACT-Lung TOI scores were lower at the time of the
assessment just prior to dropout (the last observation). Since missingness
depends on previously observed FACT-Lung TOI scores, the data are not
MCAR. Fayers and Machin [45] describe an alternative graphical approach
where the origin of the horizontal axis is the date of the last assessment. The
HRQoL is then plotted backward in time.
Figure 4.1. Average FACT-Lung TOI scores stratied by time of dropout.
c2638 C2638TextureC04 February 12, 2002 11:36
MECHANISMS OF MISSING DATA 79
Little [89] proposed a single test statistic for testing the assumption of
MCAR vs. MAR. The basic idea is that if the data are MCAR, the means
of the observed data should be the same for each pattern. If the data are
not MCAR, then the means will vary across the patterns (as is observed in
Figure 4.1). This test statistic is particularly useful when there is a large
number of comparisons, either as a result of a large number of patterns or
dierences in missing data patterns across multiple outcome variables. The
computational details of the statistic are shown in Section 4.3. In the NSCLC
study, there is considerable evidence for rejecting the hypothesis of MCAR
(
2
= 108.8, df = 26, p < 0.001).
The next approach examines the association (correlation) between an indi-
cator of missingness or dropout and the observed data. In the NSCLC study,
we observe a moderate association between dropout and both the baseline
and the most recent HRQoL scores (Table 4.8). The association is strongest
for the scales measuring the physical or functional aspects of HRQoL such as
functional well-being, physical well-being, and lung cancer symptoms, all of
which are part of the FACT-Lung TOI.
We also wish to conrm that the missingness depends on the observed data
after adjusting for the dependence on covariates. This can be tested by rst
forcing the baseline covariates identied in the previous step into a logistic
model and then testing the baseline or previous measures of HRQoL. In the
rst set of models (Table 4.9), all subjects with baseline data are included in
the logistic regression model. Age, performance status, and the presence of
Table 4.8. Correlation of missing assessments with baseline or previous HRQoL.
Kendall
b
correlation coecients
Assessment no.
1 2 3 4
HRQoL at baseline assessment
Physical well-being NA 0.18

0.21

0.20

Functional well-being 0.12

0.15

0.18

Lung cancer subscale 0.14

0.12

0.14

Emotional well-being 0.03 0.10

0.03
Social/family well-being 0.03 0.03 0.05
HRQoL at previous assessment
Physical well-being 0.18

0.19

0.17

Functional well-being 0.12

0.18

0.21

Lung cancer subscale 0.14

0.17

0.23

Emotional well-being 0.04 0.16

0.10
Social/family well-being 0.03 0.06 0.03
Note: Negative correlations indicate increased likelihood of missing as-
sessment with poorer HRQoL scores.

p < 0.05,

p < 0.01,

p < 0.001.
c2638 C2638TextureC04 February 12, 2002 11:36
80 MISSING DATA
Table 4.9. Association (odds ratios) between baseline characteristics and either base-
line or previous FCT-TOI assessment.
Time of assessment
6 weeks 12 weeks 6 months
Age
a
1.48

1.17 1.40

Performance status 1.18 1.22 1.02

Systemic symptoms 1.13 1.06 0.82
Baseline FACT-L TOI
b
0.76

0.74

0.66

Age
a
1.48

1.03 1.46

Performance status 1.18 1.18 0.72

Systemic symptoms 1.13 0.91 0.75
Previous FACT-L TOI
b
0.76

0.69

0.64

Note: First set of OR is estimated from a logistic regression model

for a missing observation regardless of patient status or missing data
pattern. Second set of OR is estimated from a logistic regression model
for dropout, given subject has completed previous assessment.
a
Age in years at randomization with OR estimated for 10-year
dierence.
b
OR estimated for 10-point dierence in FACT-Lung TOI.

p < 0.05,

p < 0.01,

p < 0.001.
systemic symptoms are forced into the model. For all follow-up assessments,
poorer baseline HRQoL, as measured by the FACT-Lung TOI score, is highly
predictive of missing data. Age is sometimes predictive, but baseline per-
formance status and symptoms are no longer predictive in this multivariate
model. In the second set of models, only those subjects who have completed
the previous assessment are included. This is a model of the probability of
missing observations given that an assessment occurred at the previous follow-
up. Again, for all follow-up assessments, poorer HRQoL, as measured by the
FACT-Lung TOI score, is highly predictive of missing data. These analyses
reinforce the evidence that the missingness is dependent on observed HRQoL
scores (Y
obs
i
) and that in this NCSLC study we cannot assume that the ob-
servations are MCAR.
Analytic methods
In general, analytic methods that assume the missingness is MAR utilize all
available HRQoL assessments. When the missing data are ignorable,* unbiased
estimates can be obtained from likelihood based methods using all observed
data (Y
obs
i
) and covariates X
i
that explain the missing data mechanism. Note
that exclusion of part of the observed data or omission of information on
covariates in the analysis may result in biased estimates. For example, if miss-
ingness is dependent on baseline scores and some individuals are excluded
* Formal denition of ignorable is presented in Appendix III.
c2638 C2638TextureC04 February 12, 2002 11:36
MECHANISMS OF MISSING DATA 81
from an analysis of change from baseline dierences because they do not have
follow-up assessments, the estimates may not be unbiased. Similarly, if miss-
ingness is strongly dependent on a covariate and that covariate is ignored, the
estimates are biased. Examples of methods assuming ignorable missing data
are presented in Chapter 5 in more detail.
A test of MCAR vs. MAR for multivariate normal data
Notation
Consider a study designed to obtain J measurements. Let P be the number
of distinct missing data patterns (R
i
), where J
{p}
is the number of observed
variables. n
{p}
is the number of cases with the pth pattern, and

n
{p}
= N.
Let M
{p}
be a J
{p}
J matrix of indicators of the observed variables in pattern
P. The matrix has one row for each measure present consisting of (J1) 0s and
one 1 identifying the observed measure. For example, in the NSCLC example,
if the rst and third observation were obtained in the 6th pattern, then
M
{6}
=

1 0 0 0
0 0 1 0

Y
{p}
is the J
{p}
1 vector of means of the observed variables for pattern p.
and

are the ML estimates of the mean and covariance of Y
i
, assuming that
the missing data mechanism is ignorable.
{p}
= M
{p}
is the J
{p}
1 vector
of ML estimates corresponding to the pth pattern, and

{p}
= [N/(N 1)]
M
{p}

M
{p}
is the corresponding J
{p}
J
{p}
covariance matrix with a cor-
rection for degrees of freedom.
Test statistic
Littles [89] proposed test statistic, when is unknown, takes the form

2
=
P

p=1
n
{p}
(

Y
{p}

{p}
)

{p}1
(

Y
{p}

{p}
). (4.1)
Little shows that this test statistic is asymptotically
2
distributed with

J
{p}
J degrees of freedom.
NSCLC example
The rst step is to obtain the maximum likelihood estimates of the mean and
covariance of the available data:
=

65.63
63.60
62.37
60.81

,

=

258.6 144.2 121.8 113.4

144.2 262.3 178.1 142.2
121.8 178.1 280.0 142.2
113.4 142.2 142.2 261.1

.
The second step is to split the data into the P patterns and obtain J
{p}
and

Y
{p}

{p}
= M
{p}
and

{p}
= [N/(N1)]M
{p}

M
{p}
are then calculated
c2638 C2638TextureC04 February 12, 2002 11:36
82 MISSING DATA
Table 4.10. Summary of J
{p}
and

Y
{p}
by pattern.
p J
{p}
y
{p}
1
y
{p}
2
y
{p}
3
y
{p}
4
n
{p}
1 4 71.83 69.63 68.99 65.39 146
2 3 66.63 65.24 60.02 . 98
3 3 69.25 66.06 . 59.88 15
4 2 63.27 57.50 . . 94
5 3 67.47 . 72.61 63.10 11
6 2 66.63 . 53.21 . 18
7 2 73.03 . . 69.85 11
8 1 58.26 . . . 133
9 3 . 70.21 78.68 69.85 5
10 2 . 71.48 75.49 . 3
11 2 . 51.12 . 62.54 2
12 1 . 61.86 . . 11
13 1 . . 73.81 . 1
14 1 . . . 47.32 1
for each of the patterns (Table 4.10). For example, in pattern 6:

{6}
= M
{6}
=

1 0 0 0
0 0 1 0

65.63
63.60
62.37
60.81

65.63
62.37

{6}
=
N
N 1
M
{6}

M
{6}
=
18
17

258.6 121.8
121.8 280.0

.
Combining all terms in Equation 4.1 yields
2
26
= 51.6, p = 0.002.
Implementing in SAS
By combining output options from Proc MI* with the PROC IML capacity for
matrix manipulation, we can save ourselves some tedious hand calculations.
The rst step is to create one record per subject, with the four repeated mea-
sures appearing as separate variables (TOI1 TOI2 TOI3 TOI4).
*** Create one record per subject ***;
PROC TRANSPOSE DATA=SASDATA.FACTL2(KEEP=CASEID FUNO FACT_T2)
OUT=WORK.TRANS PREFIX=TOI;
BY CASEID;
VAR FACT_T2;
ID FUNO;
RUN;
The second step is to obtain the pattern-specic means

Y
{p}
and the pooled
estimates
obs
and

obs
. Proc MI calculates all of these quantities when the
* Experimental procedure available in SAS 8.1 for Multiple Imputation.
c2638 C2638TextureC04 February 12, 2002 11:36
MECHANISMS OF MISSING DATA 83
default (MCMC) technique is used. The ODS statement creates temporary data
sets containing the desired information.
*** Obtain means for each pattern ***;
ODS OUTPUT MISSINGPATN=PATTERN INITPARM=ESTIMATE;
PROC MI DATA=WORK.TRANS IMPUTE=1;
VAR TOI1 TOI2 TOI3 TOI4;
RUN;
Next, the information from the two data sets is read in Proc IML to create
the matrices containing the pattern-specic means (PATMEAN) and the pooled
estimates
obs
(MEAN) and

obs
(SIGMA).
*** Calculate Statistic ***;
PROC IML;
*** Read data into PROC IML ***;
USE WORK.ESTIMATE;
READ ALL VAR{TOI1 TOI2 TOI3 TOI4}
WHERE((TYPE=MEAN)&(IMPUTATION=1))
INTO MEAN[COLNAME=COLNAME];
READ ALL VAR{TOI1 TOI2 TOI3 TOI4}
WHERE((TYPE=COV)&(IMPUTATION=1))
INTO COVAR[ROWNAME=NAME COLNAME=COLNAME];
PRINT Pooled Mean, MEAN [FORMAT=5.2];
PRINT Pooled Covariance, COVAR [FORMAT=5.2];
USE WORK.PATTERN;
READ ALL VAR{TOI12 TOI22 TOI32 TOI42} INTO PATMEAN;
READ ALL VAR{GROUP} INTO M;
PRINT Observed Means for Pattern, M PATMEAN [FORMAT=5.2] ;
Finally, the test statistic is calculated.
*** Calculate statistic for test ***;
STAT=0; DF=0; * Initializes STAT and DF *;
DO P=1 TO NROW(PATMEAN);
PATP=((PATMEAN[P,])>0); * Indicator of non-missing value *;
LOCP=LOC(PATP>0); * Location of non-missing value *;
JP=SUM(PATP); * # of non-missing values *;
IF JP^=0 THEN DO;
DP=DIAG(PATP)[LOCP,];
VP=DP*COVAR*DP; * Covariance for pattern *;
RP=MEAN[,LOCP]-PATMEAN[P,LOCP]; * Difference in means *;
STAT=STAT+M[P]*(RP*INV(VP)*RP); * Sums statistic *;
DF=DF+JP; * Sums degrees of freedom*;
END;
END;
DF=DF-NCOL(MEAN); * Adjusts df *;
PVALUE=1-PROBCHI(STAT,DF); * Calculates p-value *;
PRINT STAT DF PVALUE;
QUIT;
c2638 C2638TextureC04 February 12, 2002 11:36
84 MISSING DATA
MNAR: Missing not at random
The concept
The nal classication for missing data mechanisms, MNAR, has the least-
restrictive assumptions. In this classication, the missingness is dependent on
the HRQoL at the time the assessment is missing as well as on covariates
and observed data. This might occur because assessments are more likely to
be missing when an individual is experiencing side eects of therapy, such as
nausea/vomiting or mental confusion. Alternatively, subjects might be less
willing to return for follow-up (and be more likely to have missing assess-
ments) if their HRQoL has improved as a result of the disappearance of their
symptoms.
Analytic methods
In the past few years, there has been an increase in the amount of theoretical
research on the analysis of nonrandom missing data [21, 23, 26, 30, 57, 66,
92, 103, 129, 133, 166168, 170]. Little [92] has published a comprehensive
review of most of the methods published prior to 1995. More recent reviews
include Hogan and Laird [67]. There are three widely studied approaches to
the analysis of longitudinal studies with nonrandom missing data: selection
models, mixture models, and joint shared-parameter models. In all, the joint
distribution of the longitudinally measured outcome Y
i
and either the dropout
time T
D
i
or the missing data indicators R
i
are modeled.
Selection models factor the joint distribution into the product of the com-
plete data, f(Y), and the missing data mechanism, f(T
D
|Y) or f(R|Y).
Because the complete data are not fully observed, the analyst must make
untestable assumptions about the form of f(T
D
|Y).
Mixture models factor the joint distribution into the product of f(Y|M)
and f(M), where M is either dropout time T
D
i
, the pattern of missing data
R
i
, or a random coecient
i
. The complete data, f(Y ), is characterized as a
mixture (weighted average) of the conditional distribution f(Y |M) over the
distribution of dropout times, the patterns of missing data, or the random
coecients. Pattern mixture models are a special case of the mixture models,
where the complete data are a mixture over the missing data patterns (R).
In mixture models, because the complete data are not fully observed, some of
the parameters characterizing f(Y|M) cannot be estimated without imposing
untestable restrictions.
An important feature of these models is that all of the proposed meth-
ods are based on assumptions that, while reasonable, cannot be tested or
proved. Because we do not observe the missing outcome, all of these models
are underidentied. This means it is impossible to estimate all parameters in
the response and dropout models without making assumptions that are not
testable with the available data. However, even though these models rely on
assumptions that often cannot be veried, they can be useful. Examples of
their usefulness are presented in more detail in Chapters 8 and 9.
c2638 C2638TextureC04 February 12, 2002 11:36
MECHANISMS OF MISSING DATA 85
Identication of dependence on unobserved data, Y
mis
i
Because we have not observed the missing HRQoL score, it is not possible to
test formally a hypothesis that missingness does not depend on HRQoL at the
time the assessment is missing. The data that we need to test the hypothesis
are missing. However, it is possible to gather evidence that suggests the data
are MNAR. Specically, when there are other available measures of the disease
or outcomes of treatment that are strongly associated with HRQoL, it may
be possible to test for an association between missing observations and these
measures. For example, if missingness is associated with more severe side
eects and/or disease progression, then there is a reasonable likelihood that
the data are MNAR.
Example: NSCLC study
In the NSCLC study, we have already shown that the missingness depends
on both the initial and the most recent HRQoL scores. The question that
remains is whether, given the observed HRQoL scores (and possibly the base-
line covariates), there is additional evidence that missing assessments are more
frequent in individuals who are experiencing events likely to impact HRQoL.
The rst step is to identify these events on a theoretical, a clinical, or an
observational basis. Ideally, we want to identify variables associated with the
missing HRQoL scores, but we only have access to observed data. Thus, we
have to assume that the relationship between the observed HRQoL measures
and these events is similar to the relationship between the missing HRQoL
measures and these events.
In the NSCLC study, we might expect toxicity, disease progression, and
nearness to death to impact both HRQoL and missingness. A plot of the
FACT-Lung TOI during the months prior to death (Figure 4.2) illustrates
the strong relationship between this measure and the proximity to death.
Because one would expect that increased toxicity is associated with poorer
HRQoL, indicators of toxicity were included in this exploratory analysis. We
did not observe as strong a relationship between toxicity and observed HRQoL
scores. However, this may have occurred because toxicity was recorded as the
maximum toxicity score any time the patient was on treatment and not as
the grade of toxicity experienced during the same time frame as the HRQoL
assessment.
The nal step is to determine if these events or outcomes are associated
with missingness after adjusting for the covariates and observed HRQoL iden-
tied as associated with missingness. Obviously, death is a perfect predictor of
missingness. After forcing age and prior FACT-Lung TOI score into a logistic
regression model for missing assessments among survivors, progressive disease
during the rst six cycles of therapy and death within 2 months of the planned
assessment are strong predictors of missing assessments (Table 4.11). This sug-
gests that the data are MNAR. Unexpectedly, missing assessments were less
likely among individuals with more toxicity. One possible explanation is that
c2638 C2638TextureC04 February 12, 2002 11:36
86 MISSING DATA
Table 4.11. Patient outcomes associated with missing HRQoL.
Assessment Characteristic Odds ratio (95% CI)
6 weeks Age
a
1.44 (1.13, 1.84)
Prior FACT-L TOI
b
0.86 (0.75, 0.99)
Neurotoxicity
c
0.74 (0.61, 0.91)
PD
d
within 6 cycles 1.23 (1.23, 3.07)
Death within 2 months 4.40 (2.45, 7.93)
12 weeks Age
a
1.19 (0.90, 1.57)
Prior FACT-L TOI
b
0.80 (0.68, 0.95)
PD
d
within 6 cycles 2.84 (1.66, 4.87)
Death within 2 months 5.34 (2.20, 12.9)
6 months Age
a
1.31 (0.96, 1.77)
Prior FACT-L TOI
b
0.76 (0.63, 0.92)
Other Toxicity
c
0.74 (0.61, 0.91)
Death within 2 months 2.70 (1.25, 5.84)
Note: Logistic regression model of missing assessment among survivors with
age and prior FACT-Lung TOI were forced into the model; potential covariates
included maximum grade of hematological, neurological, and other toxicity; best
response of complete or partial remission; progression of disease within the rst
six cycles of therapy and death within 2 months of planned HRQoL assessment.
a
Age in years at randomization, with OR estimated for 10-year dierence.
b
OR estimated for 10-point dierence in FACT-L TOI.
c
Maximum NCI Common Toxicity Criteria (CTC) grade.
d
Progressive disease.
Figure 4.2. Change in FACT-Lung TOI prior to death in NSCLC study.
c2638 C2638TextureC04 February 12, 2002 11:36
RENAL CELL CARCINOMA STUDY 87
these patients are more likely to have follow-up visits and thus more likely to
be available for HRQoL assessments.
A caveat is necessary. The lack of an identied relation between these ad-
ditional variables and missingness is not proof that the data are not MNAR.
Although it may increase the analysts comfort with that assumption, there is
always the possibility that important indicators of the dropout process were
not measured in the trial or the analyst failed to identify them. Often, evidence
for the potential for nonrandom missing data will come from sources outside
the data. Clinicians may provide the most useful anecdotal information that
suggests the presence of nonrandom missing data. A knowledge of the disease
and its treatment is critical. Some of the analytic models for MNAR data [30,
133] will also provide evidence that dropout is MAR. But the absence of
evidence from one particular model does not prove the null hypothesis of ran-
domness. There is an innite number of possible models for the nonrandom
mechanism. Just because one model has been tested and there is no evidence
of a nonrandom mechanism under that set of assumptions, we cannot rule out
other mechanisms.
4.4 Renal cell carcinoma study
Examining the missing data mechanism in the renal cell carcinoma study is a
bit more dicult than in the NSCLC study because of the greater variation
in timing of the observations. Even when dening windows of time for each
assessment, we see a complicated pattern with a large number of nonmono-
tone patterns (Table 4.12). Although the majority (65%) of patients have a
monotone pattern of dropout, a substantial number have intermittent missing
data prior to dropout.
The test MCAR vs. MAR proposed by Little [89] and described in Sec-
tion 4.3 yields strong evidence against the MCAR assumption (
2
118
= 1320,
p < 0.001). This is conrmed by examining a plot of estimates as a function
of the time of dropout (Figure 4.3). The three dropout groups were dened
so that approximately a third of the patients were in each group. There is a
general trend for subjects who remain on the study longer to both start with
higher FACT-BRM TOI scores and maintain higher FACT-BRM TOI scores
over time.
Simplifying the remaining exploratory analysis, dropout during the early
and middle periods was examined using logistic regression. In both periods,
the predictors of dropout are the baseline FACT-BRM TOI scores and the du-
ration of the patients survival (Table 4.13). The odds of dropout are reduced
by about a third in both periods, with a 10-point increase in the baseline
FACT-BRM TOI scores. The odds are reduced by about half in the early
period for every doubling of the survival time.*
* Solely for the purposes of this exploratory analysis, censoring of the survival times was
ignored. Note that the majority had died (75%) and the minimum follow-up was 2 years;
thus, the missing information would have only a moderate eect when the length of
survival was expressed on a log scale.
c2638 C2638TextureC04 February 12, 2002 11:36
88 MISSING DATA
Table 4.12. Summary of patterns of missing data in renal cell carcinoma trial.
Last Assessment no.
measure Monotone 1 2 3 4 5 6 Frequency Percent
Complete Yes X X X X X X 8 3.8
1 year No . X X X X X 1 0.5
No X . X X X X 2 0.9
No X X . X X X 2 0.9
No X . . X X X 1 0.5
No . . . X X X 1 0.5
No X X X . X X 1 0.5
No . X X . X X 1 0.5
No X X . . X X 1 0.5
No X X X X . X 3 1.4
No X . X X . X 1 0.5
No X X . X . X 1 0.5
No X X X . . X 1 0.5
No . X X . . X 1 0.5
No X X . . . X 1 0.5
No . X . . . X 1 0.5
8 months Yes X X X X X . 12 5.7
No . X X X X . 3 1.4
No X . X X X . 3 1.4
No X X . X X . 2 0.9
No X X X . X . 6 2.8
No . X X . X . 1 0.5
No X X . . X . 6 2.8
No X . . . X . 3 1.4
No . . . . X . 1 0.5
4 months Yes X X X X . . 19 9.0
No . X X X . . 2 0.9
No X . X X . . 5 2.4
No . . X X . . 1 0.5
No X X . X . . 4 1.9
No X . . X . . 4 1.9
8 weeks X X X . . . 44 20.8
No . X X . . . 3 1.4
No X . X . . . 5 2.4
2 weeks Yes X X . . . . 33 15.6
No . X . . . . 6 2.8
Baseline Yes X . . . . . 22 10.4
c2638 C2638TextureC04 February 12, 2002 11:36
RENAL CELL CARCINOMA STUDY 89
Figure 4.3. Change in FACT-BRM TOI displayed by time of last assessment.
Plotting outcome by dropout
The plots displayed in Figures 4.1 and 4.3 are easily generated. Using the
model developed in Chapter 3, we add a variable dening groups in terms of
their missing data patterns (DROP GRP). In this example, we have chosen to
dene the groups by the time of their last HRQoL assessment. The rst step is
to generate a data set WORK.PLOTDATA containing the predicted values of the
outcome. The variable dening the groups is included in the CLASS statement
and as an interaction with every term in the MODEL statement.
PROC MIXED DATA=EXAMPLE3;
CLASS TREAT DROP_GRP;
c2638 C2638TextureC04 February 12, 2002 11:36
90 MISSING DATA
Table 4.13. Predictors of dropout in renal cell carcinoma trial. Results of explo-
ratory multivariate logistic regression models.
Period Characteristic OR (95% C.I.)
<5 weeks Baseline FACT-BRM TOI
a
0.66 (0.50, 0.87)
ln survival
b
0.46 (0.34, 0.62)
5 to <25 weeks Baseline FACT-BRM TOI
a
0.67 (0.47, 0.94)
ln survival
b
0.27 (0.16, 0.45)
a
Odds ratio estimated for 10-point dierence in FACT-BRM TOI.
b
Odds ratio estimated for doubling of the length of survival (0.693 units on
the natural log scale.
* Piecewise Regression Model *;
MODEL TOI=TREAT*DROP_GRP TREAT*WEEKS*DROP_GRP
TREAT*WEEKS2*DROP_GRP TREAT*WEEKS8*DROP_GRP
/NOINT SOLUTION OUTPM=WORK.PLOTDATA;
The data set is then sorted by treatment group, dropout group, and time.
Two plots are then generated, one for each treatment group, by specify-
ing treatment in the BY statement. Predicted values are plotted over time
(PRED*WEEKS) within each dropout group (=DROP GRP).
PROC SORT DATA=WORK.PLOTDATA;
BY TREAT DROP_GRP WEEKS;
PROC GPLOT DATA=WORK.PLOTDATA;
WHERE TOI NE .; * Excludes BLUPS *;
BY TREAT;
PLOT PRED*WEEKS=DROP_GRP;
SYMBOL1 I=JOIN L=1 V=NONE C=BLACK;
SYMBOL2 I=JOIN L=2 V=NONE C=BLACK;
SYMBOL3 I=JOIN L=20 V=NONE C=BLACK;
If the rst group is dened as having only a single (baseline) assessment,
as in the NSCLC example, then a symbol (STAR) is substituted for the line.
SYMBOL1 V=STAR C=BLACK;
4.5 Summary
It is important to understand the missing data patterns and mechanisms
in HRQoL studies. This understanding comes from a knowledge of the
disease or condition under study and its treatment as well as statistical
information.
MCAR vs. MAR can be tested when the data have a repeated measures
structure using the test described by Little [89] (see Section 4.3).
Graphical techniques are useful when examining the assumption of MCAR
for studies with mistimed assessments (Figures 4.1 and 4.3).
c2638 C2638TextureC04 February 12, 2002 11:36
SUMMARY 91
It is not possible to test MAR vs. MNAR without assuming a specic model
(see Chapter 9); however, it may be useful to examine the relationship
between missing data and other observed outcomes, such as indicators of
toxicity and response, that are expected to be related to HRQoL.
The lack of statistical evidence of MNAR when there is clinical information
(based on data or anecdote) that suggests that missing data are related to
the subjects HRQoL should not be interpreted as sucient to ignore the
missing data.
A general strategy for missing data is as follows:
1. Plan the study in a manner that avoids missing data.
2. Collect covariates that explain variability in the outcome and missing
data patterns.
3. Document the reasons for missing data.
4. Perform sensitivity analyses.
c2638 C2638TextureC04 February 12, 2002 11:36
c2638 C2638texC05 January 28, 2002 18:14
CHAPTER 5
Analytic Methods for Ignorable
Missing Data
5.1 Introduction
There are several methods that are often misused for the analysis of HRQoL
studies when the data are not MCAR. The objective of this chapter is to con-
vey the relationship between the assumptions of MCAR or MAR and com-
monly used analytic procedures. When missingness depends on observed data
(Y
obs
i
) and the analytic procedure does not include all available data, the re-
sults are biased. If dropout is dependent on a covariate (X
i
) and the covariate
is not included in the analytic model, then the results are biased. Popular
analytic methods such as repeated univariate analyses and MANOVA of com-
plete cases are examples of procedures that exclude observed data; results
from these analyses are unbiased only if the missing data are MCAR [105].
The magnitude of the bias depends on the amount of missing data and the
strength of the association between the missing data (Y
mis
i
) and the observed
data (Y
obs
i
) or the covariates (X
i
). The safest approach is to use methods
that make the least restrictive assumptions about the missing data. These
include multivariate likelihood methods with all the available data, such as
mixed-eects models or repeated measures for incomplete data.
Hypothetical example
This example illustrates how the various methods presented in this chapter
perform when we know the underlying missing data mechanism. Because all
values are known and the example is simple, it is possible to show how the
analytic results are aected by the missing data assumptions.
Assume there are 100 subjects with two assessments. The scores are gen-
erated from a standard normal distribution ( = 0, = 1). Three sets of
data are generated such that the correlations between the assessments (
12
)
are 0.0, 0.5, and 0.9. Correlations of HRQoL assessments over time in clinical
trials are generally in the range of 0.4 to 0.7. The extremes of 0.0 and 0.9 are
outside this range but will serve to illustrate the concepts.
All subjects have the rst assessment (T1), but 50% are missing at the
second assessment (T2). This is a larger proportion of missing data than
one would desire in a clinical trial, but setting it this high magnies the eect
we are illustrating. Finally, missing values are generated in two ways. In the
rst, observations are deleted in a completely random manner.* Thus, the
* Uniform selection: Pr[r
i2
= 1] = 0.5.
c2638 C2638texC05 January 28, 2002 18:14
94 ANALYTIC METHODS FOR IGNORABLE MISSING DATA
Table 5.1. Expected means for hypothetical examples with 50% missing data at T2.
= 0.0 = 0.5 = 0.9
MCAR MAR MCAR MAR MCAR MAR
Complete cases N = 50
T1 y
C
1
0.00 0.57 0.00 0.57 0.00 0.57
T2 y
C
2
0.00 0.00 0.00 0.28 0.00 0.51
Incomplete cases N = 50
T1 y
I
1
0.00 0.57 0.00 0.57 0.00 0.57
T2 y
I
2
(0.00) (0.00) (0.00) (0.28) (0.00) (0.51)
( ) indicates the mean of the deleted observations.
data are missing completely at random, or MCAR. In the second, the proba-
bility of a missing assessment at the second time point depends on the observed
values at the rst time point.* Thus, the data are missing at random (MAR),
conditional on the observed baseline data. The means for these hypothetical
data are summarized in Table 5.1. Data from the complete cases are noted
as Y
C
and data from the cases with any missing data are noted as Y
I
. For
subjects with both the T1 and T2 assessments, y
C
1
and y
C
2
are the averages
of the T1 and T2 scores. For subjects with only the T1 assessment, y
I
1
is the
average of the observed T1 scores. y
I
2
is the average of the deleted T2 scores,
which is known only because this is a simulated example.
5.2 Repeated univariate analyses
One of the most commonly used analytic approaches for data from a study
with a repeated measures design is repeated univariate analysis at each time
point using test procedures such as the t-test, ANOVA, or Wilcoxon rank
sum test. Although simple to implement, this approach has several disadvan-
tages [31, 97, 117]. The rst is the restrictive assumption of MCAR. If miss-
ingness is associated with previous or future observations of HRQoL, then
estimates based on only the HRQoL assessments at one point in time will be
biased. This is illustrated in the hypothetical example for repeated univariate
analysis (Table 5.2, rst row), where the estimates of
2
utilize only the data
available at T2 from the complete cases ( y
C
2
).
_

1

2
_
=
_
y
C
1
+ (1 ) y
I
1
y
C
2
_
=
_
y
1
y
C
2
_
, (5.1)
where is the proportion of subjects with complete data. Note that when
dropout depends on the measure of HRQoL at T1 and responses are correlated
* Probit selection: Pr[r
i2
= 1|Y ] = (y
i1
).
c2638 C2638texC05 January 28, 2002 18:14
REPEATED UNIVARIATE ANALYSES 95
Table 5.2. Estimates of means at T2 (
2
) in hypothetical example.
= 0.0 = 0.5 = 0.9
MCAR MAR MCAR MAR MCAR MAR
Repeated univariate 0.01 0.00 0.01 0.28 0.01 0.50
Complete case 0.01 0.00 0.01 0.28 0.01 0.50
All available data 0.00 0.00 0.00 0.02 0.01 0.02
Baseline (naive)
a
0.01 0.00 0.01 0.28 0.01 0.50
Baseline (correct)
b
0.00 0.00 0.00 0.02 0.01 0.02
Note: Average of 100 simulations. The true mean at T2 is 0. Bold estimates indicate
bias larger than the average standard error of the mean.
a
Simple regression with baseline as a covariate; analysis limited to subjects with both
assessments.
b
Means estimated at average baseline for all subjects.
( = 0), there is signicant bias in the estimated T2 mean. The bias increases
as the correlation increases.
There are other disadvantages to the repeated univariate approach. First,
the pool of subjects is changing over time. Thus, the inferences associated with
each comparison are relevant to a dierent set of patients. Second, the analyses
produce a large number of comparisons that often fail to answer the clinical
question but rather present a confusing picture. Further, the probability of
concluding that there are signicant dierences in HRQoL when none exists
(the Type I error) increases as the number of comparisons increases. Finally,
these univariate methods can be dicult to implement if measurements are
mistimed as a result of delays in therapy or other factors.
Summary Repeated univariate analyses
Examples Two-sample t-tests
ANOVA
Wilcoxon rank sum test
Assumption MCAR (R
i
does not depend on Y
obs
i
) because anal-
ysis ignores information in Y
obs
i
from other time
points
Advantages Easy to run
Easy to describe
Disadvantages Biased if data are not MCAR
Compares dierent groups of subjects at each time
Large number of comparisons that are dicult to
interpret
Increasing Type I error rate with more repeated
measures
c2638 C2638texC05 January 28, 2002 18:14
96 ANALYTIC METHODS FOR IGNORABLE MISSING DATA
Table 5.3. Results of a repeated univariate analysis of the NSCLC trial.
Follow-up no. Taxol N(%) Mean (S.E.) t df p
1 No 177(91%) 64.7 (1.21)
Yes 349(91%) 66.0 (0.87)
Di 1.34 (1.49) 0.90 524 0.37
2 No 109(56%) 63.7 (1.51)
Yes 265(70%) 65.5 (0.99)
Di 1.77 (1.82) 0.98 372 0.33
3 No 85(44%) 64.6 (1.95)
Yes 197(52%) 65.6 (1.09)
Di 0.99 (2.10) 0.47 280 0.64
4 No 57(29%) 65.0 (1.99)
Yes 134(35%) 64.8 (1.36)
Di 0.22 (2.45) 0.09 189 0.93
NSCLC example
Consider the NSCLC study. Repeated univariate analyses can be generated
as simply as this:
PROC SORT DATA=EXAMPLE2 OUT=WORK.SORTED;
BY FUNO;
RUN;
PROC TTEST DATA=WORK.SORTED;
BY FUNO;
CLASS TAXOL;
VAR FACT_T2;
RUN;
The results are summarized in Table 5.3. First, note that the means are
almost constant over time, suggesting no change in the outcome. However, the
pool of subjects that is being compared is changing across time, with a higher
proportion of the original sample being retained in the arms randomized to
the Taxol therapy. Basically, we are comparing smaller and smaller subgroups
of patients over time. It is likely that we are not evaluating the impact of
the treatment on the HRQoL of all patients, but rather on a select group of
patients in each treatment group.
5.3 Multivariate methods
The term multivariate can be used to indicate both models that include more
than one explanatory (independent) variable (e.g., multivariate logistic regres-
sion) and models that include more than one outcome (dependent) variable.
In this chapter, the term refers to repeated multiple outcome measures, either
c2638 C2638texC05 January 28, 2002 18:14
MULTIVARIATE METHODS 97
the same HRQoL scale measured over time or dierent HRQoL scales. Mul-
tivariate methods include multivariate analysis of variance (MANOVA) and
maximum likelihood estimation (MLE) for mixed-eects models or repeated
measures models. In these models, the multiple observations on each subject
are assumed to be correlated and all observations are used simultaneously in
the estimation of the parameters. In contrast, univariate methods use only the
observations occurring at one point in time. Univariate methods are usually
repeated for each of the scheduled follow-ups.
Among multivariate methods, the major distinction that is relevant to the
missing data problem is whether or not the analytic approach uses all available
data. For example, traditional software for MANOVA restricts the analysis to
those cases that are observed on all occasions. This is often referred to as com-
plete case analysis. In contrast, modern methods, developed almost 20 years
ago, are available for mixed-eects and repeated measures models for incom-
plete data. These methods, described in detail in Chapter 3, produce maxi-
mum likelihood or restricted maximum likelihood estimates (MLE/REML).
They have the important advantage of allowing us to use all available out-
come data (Y
obs
i
) when the covariates (X
i
) are fully observed. MANOVA is
a special case of the repeated measures model, where the design matrix X
i
is the same for all subjects and the analysis is restricted to subjects with no
missing data.
Complete case analysis (MANOVA)
Limiting the analysis to only those subjects who have completed all assess-
ments is another widely used method of analysis. This is also referred to as
casewise or listwise deletion. Traditional software for MANOVA automatically
limits the analysis to the complete cases. When data are MCAR, the only con-
sequence of choosing this approach is a loss of power due to the reduced num-
ber of subjects. However, when the missingness depends on observed HRQoL,
the results of these analyses are biased. This is illustrated in the hypothetical
example (see Table 5.2, second row). Casewise deletion results in an estimate
of
1
that is the average of only the T1 assessments from the complete cases
( y
c
1
) and an estimate of
2
that is the average of the T2 assessments ( y
c
2
) from
the complete cases.
_

1

2
_
=
_
y
C
1
y
C
2
_
. (5.2)
When the data are MCAR, these estimates are unbiased. However, when the
data are MAR, both
1
and
2
are likely to be biased. This is illustrated
in the hypothetical example where
2
is biased when the observations are
correlated (see Table 5.2). In addition,
1
is always biased and this will bias
the estimated change from T1 to T2 (Table 5.4).
[
2

1
] =
_
y
C
2
y
C
1

. (5.3)
c2638 C2638texC05 January 28, 2002 18:14
98 ANALYTIC METHODS FOR IGNORABLE MISSING DATA
Table 5.4. Estimates of change (
2

1
) in hypothetical example.
= 0.0 = 0.5 = 0.9
MCAR MAR MCAR MAR MCAR MAR
Complete cases 0.01 0.56 0.00 0.28 0.01 0.07
All available data 0.01 0.01 0.01 0.01 0.01 0.01
Change (naive)
a
0.01 0.56 0.00 0.28 0.01 0.07
Change (correct)
b
0.01 0.01 0.01 0.01 0.01 0.01
Note: Average of 100 simulations. The expected change is 0. Bold estimates indicate
bias larger than the average standard error of the mean.
a
Change from baseline analysis limited to subjects with two assessments.
b
Includes baseline as a covariate, means estimated at average baseline for all subjects.
Summary Multivariate analysis of complete cases
Example MANOVA
Assumptions MCAR (R
i
does not depend on Y
obs
i
); complete case
analysis ignores information in Y
obs
i
and X
i
for
individuals with missing data
Advantage Familiar method
Disadvantages Biased unless data are MCAR
Loss of information from subjects with partial data
MANOVA does not accommodate mistimed obser-
vations or time-varying covariates
NSCLC example
In most settings, an analysis based on only the complete cases will overstate
the benet of treatment because only patients for whom treatment is success-
ful are included. In the NSCLC study, the analysis is limited to survivors who
complete the 6-month follow-up assessment.
The rst step in the MANOVA analysis is to create a data set with one
record per subject using the TRANSPOSE procedure of SAS. The new data set
contains the case identier (CASEID), an indicator variable for assignment to
one of the Taxol arms (TAXOL), and the four possible FACT-Lung TOI scores
(TOI1, TOI2, TOI3, TOI4).
*** Transform data to one record per subject ***;
PROC TRANSPOSE DATA=BOOK.EXAMPLE2 OUT=WORK.ONEREC PREFIX=TOI;
BY CASEID TAXOL;
ID FUNO;
VAR FACT_T2;
RUN;
c2638 C2638texC05 January 28, 2002 18:14
MULTIVARIATE METHODS 99
The MANOVA analysis can be generated as follows:
*** Complete Case MANOVA ***;
PROC GLM DATA=WORK.ONEREC;
MODEL TOI1 TOI2 TOI3 TOI4 = TAXOL;
*** Univariate Estimates ***;
ESTIMATE TAXOL - NO INTERCEPT 1;
ESTIMATE TAXOL - YES INTERCEPT 1 TAXOL 1;
ESTIMATE TAXOL - DIFF TAXOL 1;
*** Multivariate Tests ***;
*** Differences at T2, T3 and T4 ***;
MANOVA H=TAXOL M=TOI2,TOI3,TOI4;
*** Differences in change from Baseline ***;
MANOVA H=TAXOL M=TOI2-TOI1,TOI3-TOI1,TOI4-TOI1;
RUN;
The GLM procedure rst selects the complete cases and then runs both
univariate and multivariate analyses. In this example, only 146 cases from the
original 575 subjects are included, with 20 and 28% of the subjects randomized
to the Taxol and No Taxol arms, respectively. This is noted in the output:
The GLM Procedure
Number of observations 575
NOTE: Observations with missing values will not be included in this
analysis. Thus, only 146 observations can be used in this analysis.
The results of the multivariate analyses are summarized in Table 5.5. The
estimated means from the complete cases are generally higher than those
Table 5.5. Results of a multivariate analysis of only complete cases from the NSCLC
trial.
Follow-up no. Taxol N(%) Mean (S.E.) t df p
1 No 39(20%) 72.7 (2.19)
Yes 107(28%) 71.5 (1.32)
Dierence 1.23 (2.55) 0.48 144 0.63
2 No 39(20%) 69.3 (2.21)
Yes 107(28%) 69.8 (1.33)
Dierence 0.49 (2.58) 0.19 144 0.85
3 No 39(20%) 69.1 (2.29)
Yes 107(28%) 69.0 (1.38)
Dierence 0.12 (2.67) 0.04 144 0.97
4 No 39(20%) 63.6 (2.38)
Yes 107(28%) 66.0 (1.44)
Dierence 2.40 (2.78) 0.86 144 0.39
c2638 C2638texC05 January 28, 2002 18:14
100 ANALYTIC METHODS FOR IGNORABLE MISSING DATA
estimated from the repeated univariate measures. The biggest dierences oc-
cur prior to therapy (follow-up no. 1).
Finally, multivariate tests are generated. Note that TAXOL is an indicator
variable (0 if no Taxol, 1 if Taxol), and it is included in the model without
the CLASS statement. The group not assigned to Taxol becomes the reference
group and the model parameter for the intercept (INTERCEPT) is the mean
for that group. The model parameter TAXOL is an estimate of the dierence
between the two groups. The mean for the group receiving Taxol must be cal-
culated using an ESTIMATE statement. The MANOVA statement is used to test
the two multivariate hypotheses. The rst MANOVA statement (see p. 99) is
a test of treatment dierences across all follow-up assessments. TAXOL esti-
mates the dierence between the two arms, and the M= option allows us to
select the post-treatment observations.
Multivariate Analysis of Variance
M Matrix Describing Transformed Variables
toi1 toi2 toi3 toi4
MVAR1 0 1 0 0
MVAR2 0 0 1 0
MVAR3 0 0 0 1
MANOVA Test Criteria and Exact F Statistics for the Hypothesis of
No Overall TAXOL Effect on the Variables
Defined by the M Matrix Transformation
Statistic Value F Value Num DF Den DF Pr > F
Wilks Lambda 0.99351043 0.31 3 142 0.8187
Pillais Trace 0.00648957 0.31 3 142 0.8187
Hotelling-Lawley Trace 0.00653196 0.31 3 142 0.8187
Roys Greatest Root 0.00653196 0.31 3 142 0.8187
The second MANOVA statement (see p. 99) is a test of treatment dierences
in the change from baseline. Again, TAXOL estimates the dierence between
the two arms and the M= option allows us to specify the change from base-
line.
Multivariate Analysis of Variance
M Matrix Describing Transformed Variables
toi1 toi2 toi3 toi4
MVAR1 -1 1 0 0
MVAR2 -1 0 1 0
MVAR3 -1 0 0 1
MANOVA Test Criteria and Exact F Statistics for the Hypothesis of
No Overall TAXOL Effect on the Variables
c2638 C2638texC05 January 28, 2002 18:14
MULTIVARIATE METHODS 101
Defined by the M Matrix Transformation
Statistic Value F Value Num DF Den DF Pr > F
Wilks Lambda 0.98936428 0.51 3 142 0.6768
Pillais Trace 0.01063572 0.51 3 142 0.6768
Hotelling-Lawley Trace 0.01075006 0.51 3 142 0.6768
Roys Greatest Root 0.01075006 0.51 3 142 0.6768
The standard MANOVA statistics for main eects and interactions can be
generated automatically with a REPEATED TIME statement. With the excep-
tion of the TAXOL*TIME interaction, the statement does not generate the tests
of interest and was not used.
Maximum likelihood estimation with all available data
If the data are missing at random, conditional on the observed data Y
obs
i
, then
the method used for the analysis should include all the information from Y
obs
i
.
In the hypothetical example, this is accomplished by a repeated measures
model for the two possible assessments,
_
y
i1
y
i2
_
. .
Y
i
=
_
1 0
0 1
_
. .
X
i
_

2
_
. .

+
_
e
i1
e
i2
_
. .
e
i
,
where the observations on each subject are allowed to be correlated,
Var
_
y
i1
y
i2
_
=
_

11

12

12

22
_
.
The maximum likelihood estimates of the means (
1
,
2
) are
_

1

2
_
=
_

1
i
X
i
_
1
X

1
i
Y
i
.
It is dicult to see how the estimates are calculated when written in matrix
notation. Consider a special case where all subjects have their rst assess-
ment but only a proportion, , have their second assessment. The resulting
estimates of the means are
_

1

2
_
=
_
_
y
C
1
+ (1 ) y
I
1
y
C
2
+ (1 )

12

11
_
y
I
1
y
C
1
_
_
_
=
_
_
y
1
y
C
2
+ (1 )

12

11
_
y
I
1
y
C
1
_
_
_
. (5.4)
If the variances of Y
1
and Y
2
are equal,
_

1

2
_
=
_
y
1
y
C
2
+ (1 )
_
y
I
1
y
C
1
_
_
. (5.5)
c2638 C2638texC05 January 28, 2002 18:14
102 ANALYTIC METHODS FOR IGNORABLE MISSING DATA
The estimate of the mean at T1 is the simple average of all T1 scores ( y
1
),
because there are no missing data for the rst assessment. With missing data
at the second time point (T2),
2
is not the simple mean of all observed T2
assessments ( y
C
2
) but is a function of the T1 and T2 scores as well as the cor-
relation between them. The second term in Equation 5.5 ((1 ) ( y
I
1
y
C
1
))
is the adjustment of the simple mean ( y
C
2
). If we use the information in
Table 5.1 for = 0.5,

2
= y
C
2
+ .5
_
y
I
1
y
C
1
_
= 0.28 + (0.5)(0.5)(0.57 0.57) = 0.0.
(The discrepancy is due to rounding of values in Table 5.1).
The estimates of the T2 means, using MLE for all available data, are dis-
played in the third row of Table 5.2. Estimates of change from T1 to T2 are
displayed in Table 5.4. Note that the ML estimates are unbiased for both
simulated data sets.
Summary Multivariate analysis of all available data
Examples Mixed-eects models
Repeated measures models
Assumptions MAR (R
i
depends on Y
obs
i
and X
i
but not on Y
mis
i
);
thus, estimates unbiased if model includes Y
obs
i
and
X
i
Advantages Uses all available data (all subjects with at least one
observation)
Huge class of models
Allows mistimed observations (mixed-eects models)
Allows time-varying covariates (X
i
)
Software: SAS Proc Mixed, BMDP 5V, S-plus Oswald
NSCLC example
Chapter 3 describes the implementation of repeated measures and growth
curve models in detail. In that chapter we illustrate a cell means model for
repeated measures for the adjuvant breast cancer study. In this chapter we
show an alternative reference cell model. The test statistics generated from
both models are identical, but the parameters have dierent interpretations.
As in the MANOVA analysis, the group not assigned to receive Taxol therapy
is the reference group. In combination with the NOINT option and the CLASS
statement, the term FUNO in the model statement has four levels correspond-
ing to the reference group mean at each of the four time points. The term
TAXOL*FUNO also has four levels corresponding to the dierence between the
two treatment groups at each of the four time points.
PROC MIXED DATA=BOOK.EXAMPLE2;
CLASS FUNO;
*** Model for the Fixed Effects ***;
c2638 C2638texC05 January 28, 2002 18:14
MULTIVARIATE METHODS 103
MODEL FACT_T2=FUNO TAXOL*FUNO /NOINT SOLUTION;
*** Covariance Structure ***;
REPEATED /SUBJECT=CASEID TYPE=UN;
The resulting parameter estimates for the xed eects appear as:
The Mixed Procedure
Solution for Fixed Effects
Standard
Effect FUNO Estimate Error t Value Pr > |t|
FUNO 1 64.6443 1.2085 53.49 <.0001
FUNO 2 61.8049 1.4401 42.92 <.0001
FUNO 3 61.7782 1.6540 37.35 <.0001
FUNO 4 60.4076 1.9290 31.32 <.0001
TAXOL*FUNO 1 1.4781 1.4831 1.00 0.3194
TAXOL*FUNO 2 2.5843 1.7259 1.50 0.1349
TAXOL*FUNO 3 0.8862 1.9857 0.45 0.6556
TAXOL*FUNO 4 0.6083 2.3080 0.26 0.7922
To estimate the means for the group receiving Taxol, ESTIMATE statements
are added:
*** Estimates of Means ***;
ESTIMATE BASELINE NOTX FUNO 1 0 0 0;
ESTIMATE WEEK 6 NOTX FUNO 0 1 0 0;
ESTIMATE WEEK 12 NOTX FUNO 0 0 1 0;
ESTIMATE WEEK 26 NOTX FUNO 0 0 0 1;
ESTIMATE BASELINE TX FUNO 1 0 0 0 FUNO*TAXOL 1 0 0 0;
ESTIMATE WEEK 6 TX FUNO 0 1 0 0 FUNO*TAXOL 0 1 0 0;
ESTIMATE WEEK 12 TX FUNO 0 0 1 0 FUNO*TAXOL 0 0 1 0;
ESTIMATE WEEK 26 TX FUNO 0 0 0 1 FUNO*TAXOL 0 0 0 1;
The estimates from the analysis of all available data are summarized in
Table 5.6. Note that the estimated means are consistently lower than those
obtained when the analysis was limited to the complete cases. This analysis
is also statistically more ecient than either of the two previous approaches,
as reected by the smaller standard errors.
*** Change from Baseline ***;
ESTIMATE WEEK 6-0 NOTX FUNO -1 1 0 0 ;
ESTIMATE WEEK 12-0 NOTX FUNO -1 0 1 0 ;
ESTIMATE WEEK 26-0 NOTX FUNO -1 0 0 1 ;
ESTIMATE WEEK 6-0 TX FUNO -1 1 0 0 FUNO*TAXOL -1 1 0 0;
ESTIMATE WEEK 12-0 TX FUNO -1 0 1 0 FUNO*TAXOL -1 0 1 0;
ESTIMATE WEEK 26-0 TX FUNO -1 0 0 1 FUNO*TAXOL -1 0 0 1;
The same multivariate tests can also be constructed:
*** Multivariate Tests (Contrasts) ***;
CONTRAST FU DIFFERENCES FUNO*TAXOL 0 1 0 0,
FUNO*TAXOL 0 0 1 0,
c2638 C2638texC05 January 28, 2002 18:14
104 ANALYTIC METHODS FOR IGNORABLE MISSING DATA
Table 5.6. Results of the multivariate analysis of all available data from the NSCLC
study.
Follow-up no. Taxol Mean (S.E.) t df p
1 No 64.6 (1.21)
Yes 66.1 (0.86)
Dierence 1.48 (1.48) 1.00 549 0.32
2 No 61.8 (1.44)
Yes 64.4 (0.95)
Dierence 2.58 (1.73) 1.50 549 0.13
3 No 61.8 (1.65)
Yes 62.7 (1.10)
Dierence 0.89 (1.99) 0.45 549 0.66
4 No 60.4 (1.93)
Yes 61.0 (1.27)
Dierence 0.61 (2.31) 0.26 549 0.79
FUNO*TAXOL 0 0 0 1;
CONTRAST CHANGE FROM T1 FUNO*TAXOL -1 1 0 0,
FUNO*TAXOL -1 0 1 0,
FUNO*TAXOL -1 0 0 1;
In the NSCLC example there are no dierences in the conclusions between
the results of the analyses of the complete cases and all available data, but
that will not always be true unless the data are missing completely at random.
Contrasts
Num Den
Label DF DF F Value Pr > F
FU differences 3 549 0.82 0.4843
Change from T1 3 549 0.42 0.7358
Further comments
Exclusion of subjects
Exclusion of subjects, and thus observed data, from the analysis also occurs
in subtler ways. It is not uncommon to include only those subjects with a
baseline and at least one follow-up assessment in the analysis. This strategy
automatically excludes some subjects with observed data because they have
only one observation. The assumption is that the reason they do not have
at least two assessments is completely random. In practice this is rarely true.
Often these are the sickest patients with the poorest HRQoL scores at baseline.
In the NSCLC study, only 5% of the subjects are missing all four FACT-Lung
TOI scores (Table 5.7). However, another 27% of the subjects are excluded
c2638 C2638texC05 January 28, 2002 18:14
BASELINE ASSESSMENT AS A COVARIATE 105
Table 5.7. Potential for exclusion of subjects in NSCLC study.
No Taxol Taxol Total
N (%) N (%) N (%)
No FACT-L TOI data 10 (5) 16 (4) 26 (5)
No baseline FACT-L TOI 7 (4) 16 (4) 23 (4)
No follow-up FACT-L TOI 56 (29) 77 (20) 133 (23)
Baseline + follow-up 121 (62) 272 (71) 393 (68)
Total subjects 194 (100) 381 (100) 575 (100)
if a baseline and at least one follow-up assessment are required for inclusion;
4% because of the requirement for a baseline assessment; and 23% because of
the requirement for a follow-up assessment.
Taking this idea a bit further, patients with no HRQoL scores should not be
completely excluded from analyses. If there are covariates that are strongly
associated with both missingness and observed HRQoL, the values of the
covariates from all subjects can be used to improve the estimates; this is
illustrated later in this chapter.
Exclusion of observations
Exclusion of observations from the analysis should be performed very cau-
tiously. In some settings, there is a valid conceptual reason to do so. For
example, in the adjuvant breast cancer study, four assessments were excluded
from the pretherapy assessments because they occurred after therapy started
and two assessments were excluded from the on-therapy assessments because
they occurred after therapy had been stopped. In other settings, the articial
attempt to force observations into a repeated measures model may result in
the exclusion of observations. If the exclusion is not random or missingness at
other occasions depends on the the value of these observations, the resulting
analyses are biased.
5.4 Baseline assessment as a covariate
Another popular strategy for the analysis of repeated measures is to include
the baseline assessment in the model as a covariate (X
i
) rather than using it
as one of the repeated measures. In this model,
y
ij
= X
ij

j
+ y
i1

j
+
ij
, (5.6)
where y
i1
is the baseline measure for the ith subject. This strategy partially
addresses the issue of ignoring the observed data by including the information
in the baseline assessment. However, it should be used cautiously if there are
missing data in either the baseline or early follow-up assessments. Most an-
alytic procedures will, by default, exclude individuals with missing baseline
c2638 C2638texC05 January 28, 2002 18:14
106 ANALYTIC METHODS FOR IGNORABLE MISSING DATA
HRQoL or no follow-up assessments from all aspects of the analysis. The po-
tential result of excluding these subjects from the analysis is to bias the esti-
mates of HRQoL.
Consider the hypothetical example where the model is
y
i2
= + y
i1
+
i2
. (5.7)
If the mean of the T2 scores is estimated using only the baseline data from
the subjects with follow-up assessments (y
C
i1
), then the mean HRQoL scores
at T2 will be overestimated because y
C
1
overestimates the baseline mean,
1
.
This is illustrated in Table 5.2 in the row identied as Baseline (naive).

2
=

+ y
C
1
. (5.8)
In this simple case, the estimates of
2
are identical to those obtained for
the complete case and the univariate analyses. In contrast, if all observed
T1 assessments are included in the estimation of
2
, the estimates are no
longer biased. This is illustrated in Table 5.2 in the row identied as Baseline
(correct).

2
=

+ y
1
, y
1
= y
C
1
+ (1 ) y
I
1
. (5.9)
When there are multiple follow-up assessments, the same issues of mul-
tivariate vs. univariate analyses apply. If missingness is related to both the
baseline and previously observed assessments of HRQoL, then the appropriate
analysis is a multivariate analysis of the follow-up assessments. With extended
follow-up this will become more important, as there is a stronger relationship
between dropout and the more recent assessments of HRQoL than with the
initial baseline measure. Repeated univariate analyses will ignore information
from the previous assessments and will be biased if dropout depends on the
previously observed HRQoL.
Summary Baseline as a covariate
Assumption MAR, if means of all baseline data are included
MCAR, if limited to subjects with baseline and
follow-up data (many programs ignore informa-
tion in the covariates X
i
for subjects with miss-
ing data; must force E[Y
i
] to be estimated at the
mean of X
i
for all subjects)
Advantages Easy to use
Easy to describe
Disadvantage Cannot estimate change from baseline
SAS note To obtain unbiased estimates, use LSMEANS .../AT
MEANS E; and data set must include rows for
missing observations
c2638 C2638texC05 January 28, 2002 18:14
BASELINE ASSESSMENT AS A COVARIATE 107
NSCLC example
As a rst step, the baseline FACT-Lung TOI scores (TOI1) from WORK.ONEREC
are merged with the follow-up scores from BOOK.EXAMPLE2. The SQL procedure
allows us to do this in one step:
PROC SQL;
*** Merges Baseline Score and change ***;
CREATE TABLE WORK.FOLLOWUP
AS SELECT *,
(FACT_T2-TOI1) AS CHANGEBL /* Change */
FROM BOOK.EXAMPLE2 AS L
LEFT JOIN WORK.ONEREC AS R
ON L.CASEID=R.CASEID
HAVING L.FUNO>1 /* Selects FU 2, 3, 4*/
ORDER BY L.CASEID,L.FUNO;
The repeated follow-up assessments are estimated using the MIXED
procedure:
PROC MIXED DATA=WORK.FOLLOWUP NOCLPRINT;
TITLE Baseline FACT-Lung TOI as a covariate - With AT MEANS;
CLASS FUNO TAXOL;
MODEL FACT_T2 = TAXOL*FUNO TOI1/NOINT SOLUTION DDFM=SATTERTH;
REPEATED /SUBJECT=CASEID TYPE=UN;
LSMEANS TAXOL*FUNO/AT MEANS;
RUN;
A portion of the LSMEANS output appears as follows:
The Mixed Procedure
Least Squares Means
Standard
Effect FUNO TAXOL TOI1 Estimate Error
FUNO*TAXOL 2 No 65.58 62.7618 1.3248
FUNO*TAXOL 2 Yes 65.58 64.0660 0.8600
FUNO*TAXOL 3 No 65.58 61.8029 1.5891
FUNO*TAXOL 3 Yes 65.58 61.8082 1.0564
FUNO*TAXOL 4 No 65.58 59.6353 1.9326
FUNO*TAXOL 4 Yes 65.58 60.2863 1.2579
In the NSCLC study, analysis of the follow-up observations with the base-
line scores as covariates limits the analysis to the 383 subjects with both
baseline and at least one follow-up observation, just over two thirds of the
randomized subjects (see Table 5.7). When the AT MEANS option is omit-
ted from the LSMEANS statement, the means are estimated at a value that
is the mean of the covariate across all observations with a nonmissing value of
the outcome (FACT T2). In this example, the value was 69.3 (Table 5.8, naive
estimates).
c2638 C2638texC05 January 28, 2002 18:14
108 ANALYTIC METHODS FOR IGNORABLE MISSING DATA
Table 5.8. Results of multivariate analyses of three repeated follow-up measures
with the baseline scores as a covariate compared with four repeated measures with
baseline as a dependent variable (Table 5.6).
Covariate
Four repeated
Naive estimates Correct estimates measures
Follow-up no. Taxol Mean (S.E.) Mean (S.E.) Mean (S.E.)
2 No 64.7 (1.32) 62.8 (1.32) 61.8 (1.44)
Yes 66.0 (0.86) 64.1 (0.86) 64.4 (0.95)
3 No 63.7 (1.59) 61.8 (1.59) 61.8 (1.65)
Yes 63.7 (1.05) 61.8 (1.05) 62.7 (1.10)
4 No 61.6 (1.92) 59.6 (1.92) 60.4 (1.93)
Yes 62.2 (1.26) 60.3 (1.26) 61.0 (1.27)
Baseline
a
69.3 65.6
a
Baseline scores used to estimate LSMEANS.
We can add back some information from the subjects who had only a
baseline assessment by including their baseline values when estimating the
LSMEANS. When the AT MEANS option is included and records exist for all
subjects at all possible times, the LSMEANS are estimated at a value that
is the mean of the covariate across all subjects with a nonmissing value of
the covariate (TOI1). In this example, this value is 65.6 (Table 5.8, naive esti-
mates). The parameter estimates obtained from the naive analysis (with the AT
MEANS option omitted) are contrasted with the corrected analysis in Table 5.8.
The estimates from the naive procedure are consistently higher than from
the correct procedure, although the dierences between the groups are not
aected. Note also that the estimates from the correct procedure are very
similar to those obtained with the multivariate analysis of all available data
(Table 5.6). The dierences occur because of the small dierence of 64.6 vs.
66.1 in the baseline scores between the two groups.
5.5 Change from baseline
Another popular method is the analysis of change from baseline. The strategy
is to subtract the baseline scores from the follow-up scores:
y
ij
y
i1
= X
ij

j
+
ij
. (5.10)
The analysis is then limited to subjects with both baseline and follow-up
assessments. This is another example of how an analysis that is not carefully
implemented will result in biased estimates. Because the outcome is now the
change from baseline, the data are MAR only if the missingness depends on the
observed change. In our hypothetical example, this is not a valid assumption,
as the missingness also depends on the change in those subjects where we have
c2638 C2638texC05 January 28, 2002 18:14
CHANGE FROM BASELINE 109
not observed the change. The biased estimates of change resulting from this
analysis are summarized in the row labeled Change (naive) in Table 5.4.
Note that the estimates of change in this setting are the same as those obtained
from the complete case analysis.
In this simple example, we can x the problem by adding baseline as a co-
variate. Now the data are MAR, conditional on the observed baseline response.
y
ij
y
i1
= X
ij

j
+ y
i1

j
+
ij
. (5.11)
The unbiased estimates of change resulting from this analysis are summarized
in the row labeled Change (correct) in Table 5.4. Again, it is important to
use the mean of all observed data at T1 ( y
1
), and not just the mean at T1 for
those subjects who have both T1 and T2 assessments ( y
C
1
), when estimating
the change.
Again, when there are multiple follow-up assessments, the arguments in
favor of multivariate over univariate analyses apply. With extended follow-up,
the use of a multivariate analysis will become more important, as there is
a stronger relationship between dropout and the more recent assessments of
HRQoL. These multivariate models may also become very complex, as the
relationship () between the baseline assessment and change is moderated by
either time or the intervention. Specically, the relationship may be dierent
for each of the J time points or H groups.
NSCLC example
Analysis of the change from baseline scores limits the analysis to the same
67% of the randomized subjects with both baseline and at least one follow-up
observation (Table 5.7). When missingness depends on the baseline scores,
these estimates are biased because the scores of subjects with no follow-up
are ignored in the analysis (Table 5.9). By adding baseline as a covariate
and estimating the change at the mean for all subjects, we minimize that
bias.
PROC MIXED DATA=WORK.FOLLOWUP;
TITLE Change from Baseline with Baseline as a Covariate;
CLASS FUNO TAXOL;
MODEL CHANGEBL = TAXOL*FUNO TOI1/NOINT SOLUTION DDFM=SATTERTH;
REPEATED /SUBJECT=CASEID TYPE=UN;
LSMEANS TAXOL*FUNO/AT MEANS;
RUN;
Least Squares Means
Standard
Effect FUNO TAXOL TOI1 Estimate Error
FUNO*TAXOL 2 No 65.58 -2.8200 1.3248
FUNO*TAXOL 2 Yes 65.58 -1.5158 0.8600
c2638 C2638texC05 January 28, 2002 18:14
110 ANALYTIC METHODS FOR IGNORABLE MISSING DATA
Table 5.9. Results of a multivariate analysis of the three changes from baseline
scores compared with estimates of change obtained from model with four repeated
measures (Table 5.6).
Change from baseline scores
Four repeated
Without covariate With covariate measures
Follow-up no. Taxol Mean (S.E.) Mean (S.E.) Mean (S.E.)
2 No 4.03 (1.45) 2.82 (1.32) 2.84 (1.40)
Yes 2.64 (0.94) 1.52 (0.86) 1.73 (0.92)
3 No 5.17 (1.75) 3.77 (1.59) 2.87 (1.69)
Yes 5.38 (1.16) 3.77 (1.06) 3.46 (1.13)
4 No 8.41 (2.09) 5.95 (1.93) 4.24 (1.99)
Yes 7.56 (1.36) 5.30 (1.26) 5.10 (1.31)
FUNO*TAXOL 3 No 65.58 -3.7788 1.5891
FUNO*TAXOL 3 Yes 65.58 -3.7736 1.0564
FUNO*TAXOL 4 No 65.58 -5.9464 1.9326
FUNO*TAXOL 4 Yes 65.58 -5.2955 1.2579
5.6 Adding other baseline covariates
If dropout depends on certain patient characteristics, the previous discus-
sion would suggest that it will be important to include corresponding covari-
ates in all analysis models. However, in practice, the inclusion or omission
of these baseline covariates rarely makes a dierence when the intent is to
compare treatment arms in a randomized clinical trial. First, if the character-
istic is not correlated with the HRQoL measure, then its inclusion/exclusion
will not aect the results. Second, if the characteristic is correlated with the
HRQoL measure, much of the covariance with postrandomization measures
is explained by the prerandomization baseline measure. However, it is advis-
able to perform a sensitivity analysis by repeating the analysis with these
covariates.
In theory, the treatment arms should be balanced with respect to these pa-
tient characteristics. When there are apparent dierences detected in baseline
characteristics, two possibilities exist: (1) the dierences are real or (2) there is
no dierence between the groups and a Type I error has occurred. In the rst
case, it is appropriate to adjust the estimates by inclusion of a covariate in
the analysis, but in the second case the dierences are ignored. Unfortunately,
it is not possible to distinguish the two cases. Again, a sensitivity analysis is
advisable.
c2638 C2638texC05 January 28, 2002 18:14
ADDING OTHER BASELINE COVARIATES 111
Summary Change from baseline
Examples Baseline subtracted from all FU observations
Assumptions MCAR (R
i
cannot depend on the dierence,
Y
ij
Y
i1
)
Possibly MAR with baseline as a covariate
Advantages Familiar method
Possible strategy when there are initial
(baseline) dierences in Y
Disadvantage Cannot estimate means and some summary
measures
NSCLC example
Based on the results presented in Chapter 4, a sensitivity analysis with age,
performance status, and symptoms of systemic disease appears warranted.
There were no identied dierences between the two treatment groups, so no
additional variables were selected for this sensitivity analysis. As an alter-
native to using the LSMEANS option to obtain the adjusted estimates, in this
example each of the variables is centered at its mean. The SQL procedure can
be used to create three new variables (CNT AGE, CNT PS and CNT SXSYS) in a
single step:
PROC SQL;
CREATE TABLE WORK.CENTERED AS
SELECT *,
L.AGE_TX-MEAN(L.AGE_TX) AS CNT_AGE,
L.ECOGPS-MEAN(L.ECOGPS) AS CNT_PS,
L.SX_SYS-MEAN(L.SX_SYS) AS CNT_SXSYS
FROM BOOK.PATIENT2 AS L
LEFT JOIN BOOK.FACTL2 AS R
ON L.CASEID=R.CASEID
ORDER BY CASEID;
or in a simple two-step process:
PROC SQL;
CREATE TABLE WORK.TEMP AS
SELECT CASEID,
AGE_TX-MEAN(AGE_TX) AS CNT_AGE,
ECOGPS-MEAN(ECOGPS) AS CNT_PS,
SX_SYS-MEAN(SX_SYS) AS CNT_SXSYS
FROM BOOK.PATIENT2
ORDER BY CASEID;
DATA WORK.CENTERED;
MERGE BOOK.FACTL2 WORK.TEMP;
BY CASEID;
RUN;
c2638 C2638texC05 January 28, 2002 18:14
112 ANALYTIC METHODS FOR IGNORABLE MISSING DATA
Table 5.10. Results of the multivariate analysis of all available data with and without
baseline covariates associated with dropout (age, performance status, symptoms of
systemic disease) from the NSCLC study.
Without covariates With covariates
Follow-up no. Taxol Mean (S.E.) Mean (S.E.)
1 No 64.6 (1.21) 64.8 (1.17)
Yes 66.1 (0.86) 66.2 (1.83)
Dierence 1.48 (1.48) 1.39 (1.43)
2 No 61.8 (1.44) 61.7 (1.38)
Yes 64.4 (0.95) 64.0 (0.91)
Dierence 2.58 (1.73) 2.23 (1.66)
3 No 61.8 (1.65) 61.8 (1.59)
Yes 62.7 (1.10) 62.6 (1.05)
Dierence 0.89 (1.99) 0.85 (1.90)
4 No 60.4 (1.93) 59.9 (1.85)
Yes 61.0 (1.27) 60.2 (1.22)
Dierence 0.61 (2.31) 0.33 (2.22)
The MIXED procedure statements are identical to those used for the multi-
variate maximum likelihood estimation with all available data except for the
inclusion of the covariates in the MODEL statement.
PROC MIXED DATA=WORK.CENTERED;
TITLE ML WITH COVARIATES;
CLASS FUNO;
MODEL FACT_T2=FUNO TAXOL*FUNO CNT_AGE CNT_SXSYS CNT_PS
/NOINT SOLUTION ;
REPEATED /SUBJECT=CASEID TYPE=UN;
Because we have used centered covariates, none of the other statements needs
to be changed.
Coecients for performance status (CNT PS) and symptoms of systemic
disease (CNT SXSYS) were signicant (p < 0.001), and age was marginally
nonsignicant (p = 0.09). The resulting estimates of the parameters, with
and without the covariates, are displayed in Table 5.10. The results from the
two analyses were quite similar.
5.7 Empirical Bayes estimates
In the mixed-eects model, the expected value (unconditional expectation) of
an observation is
E[Y
i
] = E[X
i
+ Z
i
d
i
+ e
i
] = X
i
. (5.12)
c2638 C2638texC05 January 28, 2002 18:14
EMPIRICAL BAYES ESTIMATES 113
Linear functions of the xed eects are called estimable functions. X
i
is an
estimable function if is estimable, and X
i

is called the best linear unbiased

estimate (BLUE) of X
i
.
The conditional expectation of Y given the random eects is
E[Y
i
|d
i
] = E[X
i
+ Z
i
d
i
+ e
i
] = X
i
+ Z
i
d
i
. (5.13)
Linear functions of the xed and random eects are called predictable func-
tions. The solutions for estimates of and d
i
provide the best linear unbiased
predictor (BLUP) of X
i
+ Z
i
d
i
.
Some insight into the estimates obtained from the mixed-eects model is
provided by further examination of the BLUPs of the responses for a particular
individual,

Y
i
. The empirical Bayes estimates of the random eects,

d
i
, are
the expectation of the random eect, d
i
, conditional on the observed data
Y
obs
i
, where
E
_
Y
i
d
i
_
=
_
X
i

0
_
, (5.14)
Var
_
Y
i
d
i
_
=
_

i
Z
i
D
DZ
i
D
_
, (5.15)
E [d
i
|Y
i
] = 0 +DZ
i

1
i
(Y
i
X
i
). (5.16)
With a bit of algebraic manipulation, we see that the predicted values are
a weighted combination of average population estimates, X
i

, and observed
data, Y
i
[154, p. 119].

Y
i
= X
i

+Z
i

d
i
= X
i

+Z
i
DZ

1
i
_
Y
obs
i
X
i

_
=
_
I Z
i
DZ

1
i
_
X
i

+ Z
i
DZ

1
i
Y
obs
i
=
_

1
i
Z
i
DZ

1
i
_
X
i

+ Z
i
DZ

1
i
Y
obs
i
=
2
w
I
..
Within

1
i
X
i

+ Z
i
DZ

i
. .
Between

1
i
Y
obs
i
. (5.17)
Recall that
i
= Z
i
DZ

i
+
2
I and note that the numerator of the weight
on X
i

,
2
I, is the within-subject residual covariance and the denominator
is the total variance,
i
. Likewise, the numerator of the weight on Y
obs
i
is
the between subject variance, Z
i
DZ

i
. As the portion of the variance that
is attributed to between subject dierences (D) increases, the weight on the
observed data increases. Although it is not so obvious, the number and timing
of the observations aect the weights. When there are two random eects, the
second term will increase both as the number of observations increase and as
they spread over time. For a subject with two observations at the beginning
and end of the study, this term is greater than for a subject with only two
early observations.
c2638 C2638texC05 January 28, 2002 18:14
114 ANALYTIC METHODS FOR IGNORABLE MISSING DATA
Note that the sucient statistics in the EM algorithm [32], which is often
used to to estimate the mixed-eects parameters, are constructed from these
conditional expectations.
5.8 Summary
In the presence of missing or mistimed data, it is advisable to employ ana-
lytic methods that use all of the observed data such as repeated measures
models for incomplete data or mixed-eects models (Chapter 3).
Methods that exclude subjects or observations from the analysis should
be avoided unless one is convinced that the data are missing completely
at random (MCAR). These methods include repeated univariate analyses,
which ignore observations that occur at dierent times, and MANOVA,
which deletes cases with any missing assessments.
c2638 C2638TextureC06 February 7, 2002 12:19
CHAPTER 6
Simple Imputation
6.1 Introduction
Traditional use of imputation was motivated by the desire to avoid bias due to
missing data and the unavailability of analytic methods for incomplete data.
Prior to the mid-1980s, statistical software for longitudinal studies or repeated
measures, such as MANOVA or MANCOVA, required the exclusion of all in-
dividuals with any missing data. This disadvantage has disappeared with the
accessibility of software for incomplete data such as maximum likelihood esti-
mation (MLE) for longitudinal studies with ignorable missing data [109, 115,
132] described in Chapter 5. When it is reasonable to assume that the missing
data in a HRQoL study are either MCAR or MAR, these MLE procedures
are appropriate and the need to impute missing observations diminishes.
There remain two situations where imputation may be useful in studies in-
volving HRQoL measurement. The rst situation occurs when there are miss-
ing explanatory (independent) variables or covariates in the analysis [153].
Missing covariates will result in the deletion of individuals with missing data
from any analysis. The second situation occurs when HRQoL is the outcome
(dependent) measure and there is a concern that the missing data are not
ignorable. Imputation methods are useful as part of sensitivity analyses to ex-
amine the dependence of the results on specic assumptions about the HRQoL
of individuals with missing observations [153]. However, imputation should
not be considered either an easy x or a denitive solution to the problem of
missing data; in all cases imputation requires the analyst to make untestable
assumptions concerning the missing data.
This chapter covers simple imputation techniques. Simple imputation is
the process of substituting a single reasonable value for a missing observation.
This procedure, if thoughtfully performed, may have a limited usefulness in
examining the sensitivity of estimates of the means to various assumptions.
Common simple imputation methods (Table 6.1) include techniques such as
last value carried forward (LVCF) and substitution of predicted values from
regression-based techniques. The choice among methods should be made only
after careful consideration of why the observations are missing, the general
patterns of change in HRQoL over time, and the research questions being
addressed. Without this careful consideration, imputation may increase the
bias of the estimates. For example, if missing HRQoL observations are more
likely in individuals who are experiencing toxicity as a result of the treatment,
the average value from individuals with observations (who are less likely to be
experiencing toxicity) will likely overestimate the HRQoL of individuals with
missing data.
c2638 C2638TextureC06 February 7, 2002 12:19
116 SIMPLE IMPUTATION
Table 6.1. Examples of simple imputation methods.
Means Mean of observed data
Predicted values Predicted values from univariate or multivariate
linear regression models
Hot deck Randomly sampled observation
LVCF (LOCF) Last value (observation) carried forward
MVCF Minimum value carried forward
Low/high value A theoretically justied low or high value; possibly
the minimum or maximum observed value
Table 6.2. Selected patients from the NSCLC study.
6 12 26 Months
No. Taxol Baseline weeks weeks weeks CR/PR survival
12 No 60.9 M X X No 2.5
13 No 82.1 M X X No 2.7
60 Yes 57.2 40.6 M X No 3.6
527 Yes 63.5 52.4 M M No 7.2
277 Yes 67.5 71.6 M M No 7.8
Observed data are displayed in bold, M indicates missing and alive, X indicates patient
has expired. CR/PR = complete or partial response.
Limitations of simple imputation
The most common criticism of simple imputation methods is that most anal-
yses treat imputed values as if they were observed values. This results in
underestimation of the variance of the parameter estimates. Without proper
adjustments, simple imputation is inappropriate when the goal is to construct
test statistics and condence intervals. Naive implementation of these tech-
niques may produce a false sense of condence that we have solved the missing
data problem when, in fact, we have just obscured it.
NSCLC example
Some of the problems with simple imputation cited above are illustrated by
examining the results of simple imputation in the NSCLC study. Consider ve
patients selected from this trial (Table 6.2). All failed to respond to therapy.
The rst two (subjects 12 and 13) have a similar short survival and could only
be assessed prior to therapy and at 6 weeks. The major dierence between
these two subjects is that one had a higher initial score than the other. The
third subject is missing an assessment at 12 weeks and died prior to the time
of the nal assessment. The fourth and fth subjects have similar survival and
c2638 C2638TextureC06 February 7, 2002 12:19
MEAN VALUE SUBSTITUTION 117
are both missing two assessments. Subject 277 reported very little change in
his HRQoL, whereas subjects 60 and 527 both report a signicant decline over
the rst 6 weeks of therapy. As dierent methods of simple imputation are
presented in the next section, the imputed values are illustrated for these ve
subjects.
6.2 Mean value substitution
A common simple imputation technique is the substitution of the average
value from the individuals who completed the measure of HRQoL for the
missing observations. This mean may be computed across all individuals or
may be specic to treatment groups or to individuals with similar characteris-
tics. The method assumes that the reasons some individuals did not complete
their HRQoL assessments is completely unrelated to their HRQoL. This may
be valid if the missing data are due to administrative problems, but it is hard
to justify in most situations. For example, even in the case where a subject
with a missing assessment was documented to be experiencing toxicity and
the mean HRQoL score was computed from responses among other individu-
als experiencing toxicity, there is still the assumption that the impact of the
toxicity on the subjects HRQoL is equivalent among those who completed
the HRQoL assessment and those who did not.
Table 6.3 displays mean imputation for ve cases in the NSCLC study.
Means were imputed separately for each treatment group, ignoring all other
information about the patient. Thus, the rst two subjects have the same
imputed value at 6 weeks, even though subject 12 had a much lower base-
line score than subject 13. The last three subjects have the same value at
12 weeks, irrespective of the dierences in their survival. However, in the
NSCLC study, we have observed that subjects with lower baseline HRQoL
and those who are closer to death are more likely to have missing assess-
ments; simple mean imputation is likely to result in biased estimates of the
overall mean.
Table 6.3. Simple imputation using mean of observed data illustrated for selected
patients in the NSCLC study.
No. Taxol Baseline 6 weeks 12 weeks 26 weeks Months survival
12 No 60.9 63.7 X X 2.5
13 No 82.1 63.7 X X 2.7
60 Yes 57.2 40.6 65.6 X 3.6
527 Yes 63.5 52.4 65.6 64.8 7.2
277 Yes 67.5 71.6 65.6 64.8 7.8
Observed data are displayed in bold and imputed data in italics. X indicates patient has
expired.
c2638 C2638TextureC06 February 7, 2002 12:19
118 SIMPLE IMPUTATION
Figure 6.1. Distribution of observed and imputed values using group means.
The distributions of observed scores and imputed scores are displayed in
Figure 6.1. The imputed values are centered relative to the observed scores,
emphasizing the assumption that the missing values are MCAR.
6.3 Explicit regression models
In the explicit regression model approach, we identify a regression model to
predict the missing observation. The explicit regression approach has the ad-
vantage that the regression model used to impute missing observations can
readily use auxiliary information about the subjects HRQoL that could not be
c2638 C2638TextureC06 February 7, 2002 12:19
EXPLICIT REGRESSION MODELS 119
included in the model used for the ultimate analysis of the clinical trial. This
additional information might include indicators of the side eects of treat-
ment and the clinical course of the disease when the therapy succeeds or fails.
Alternatively, it could include measures of HRQoL provided by a caregiver
when the subject is no longer able to complete self-assessments.
Analytic model: Y
i
= X
i
+
i
Imputation model: Y

i
= X

i
B

i
Identication of the imputation model
Recall that one of the objectives of imputation is to obtain unbiased estimates
of the missing observations. In the context of the explicit regression approach,
we are looking for a model where the missingness in the imputation model de-
pends only on the observed data (Y

i
) and covariates in the imputation model
(X

i
). Basically, we are attempting to augment the analytic model with aux-
iliary outcome data (Y

i
) or covariates (X

i
), so that the MAR assumption
in the imputation model seems reasonable. The primary candidates for co-
variates include measures that are strongly correlated with both the measure
being imputed and the probability that it is missing. In the NSCLC example,
this suggests adding covariates associated with toxicity, disease progression,
and the proximity of the patient to death. Note that

is used to distinguish
the covariates and corresponding parameters of the imputation model from
those included in the analytic model.
Assuming that the objective of the analysis is the comparison of treatment
groups, the following strategy is recommended for identication of the impu-
tation model:
1. Identify patient characteristics, measures, and clinical outcomes that are
strong predictors of the HRQoL outcome to be imputed.
2. Identify patient characteristics, measures, and clinical outcomes that pre-
dict missingness of the HRQoL outcome to be imputed.
3. Remove variables from the above lists if they are frequently missing. If
X
mis
i
includes covariates that are missing, then we will be unable to impute
values for Y
mis
i
.
4. Select and test potential covariates in the imputation models, giving the
highest priority to variables that are in both lists.
5. If the sample size is large, develop separate imputation models for each
treatment group. Otherwise, force variables identifying treatment groups
into the model. If models are not being developed separately for each treat-
ment group, evaluate interactions between treatment and potential covari-
ates. Failure to do this will bias the treatment comparisons toward the null
hypothesis.
When the intent of the analysis extends beyond treatment comparisons, all
important explanatory variables on which inference is planned should be in-
cluded as explanatory variables in the imputation model to avoid biasing the
evidence toward the null hypothesis.
c2638 C2638TextureC06 February 7, 2002 12:19
120 SIMPLE IMPUTATION
Simple univariate regression
The parameters of the imputation model for the jth assessment,

B

j
, are esti-
mated using the observed data with the model
Y
obs
ij
= X
obs
ij
B

j
+

ij
. (6.1)
The predicted values are then computed for the missing values:
Y
mis
ij
= X
mis
ij

B

j
. (6.2)
This approach will result in unbiased estimates if the regression model sat-
ises the assumptions. Specically, the missingness should depend only on
covariates that were included in the imputation model. In practice, this ap-
proach will require the careful documentation of the reasons for missing data
and the luck or foresight to measure the patient characteristics and outcomes
(covariates) that explain the missing data mechanism.
Table 6.4 displays the imputed values for the same ve cases from the
NSCLC study for simple unconditional imputation. Imputed values are the
predicted values from simple linear regression models. For the rst two cases,
the predicted value at 6 weeks for subjects on the Etoposide arm is
Y
mis
ij1
= X
mis
ij

B

j
= 63.4 + 6.6 X

ij1
12.4 X

ij2
, j = 2, (6.3)
where X

1j
= 1 if the best response is a complete or partial response or 0
otherwise and X

ij2
= 1 if the patient will expire within 2 months of the jth
assessment or 0 otherwise. For both cases, X
i21
= 0 and X
i22
= 1; therefore,
the predicted value is 51.0. At 12 weeks, the predicted value for subjects on
the Taxol arm is
Y
mis
ij
= X
mis
ij

B

j
= 63.8 + 6.7 X

ij1
20.2 X

ij2
, j = 3. (6.4)
Because subject 60 is within 2 months of death (X

i32
= 1), the imputed value
is much lower than for subjects 277 and 527. These values are much more
intuitively appealing than the simple means, especially for subjects closer
to death. The assumptions about the missing data have been relaxed slightly.
Table 6.4. Simple imputation using linear regression models with covariates.
Weeks postrandomization
No. Taxol 0 6 12 26 Months survival CR/PR
12 No 60.9 51.0 X X 2.5 No
13 No 82.1 51.0 X X 2.7 No
60 Yes 57.2 40.6 43.6 X 3.6 No
527 Yes 63.5 52.4 63.8 50.9 7.2 No
277 Yes 67.5 71.6 63.8 50.9 7.8 No
Observed data are displayed in bold and imputed data in italics. NSCLC example.
CR/PR=complete or partial response.
c2638 C2638TextureC06 February 7, 2002 12:19
EXPLICIT REGRESSION MODELS 121
Figure 6.2. Distribution of observed and imputed values using simple regression
with baseline characteristics (treatment group, performance status, prior weight loss,
symptoms of primary and systemic disease).
We are still assuming that relationships between HRQoL and clinical outcomes
such as response and survival are the same for subjects with missing data as
for those with observed HRQoL.
The distributions of observed scores and imputed scores, predicted by simple
linear regression, are displayed in Figures 6.2 and 6.3. In the rst set of plots
(Figure 6.2), the explanatory variables in the regression models are limited to
baseline characteristics. The imputed values remain centered relative to the
c2638 C2638TextureC06 February 7, 2002 12:19
122 SIMPLE IMPUTATION
Figure 6.3. Distribution of observed and imputed values using simple regression
with baseline characteristics (treatment group, performance status, prior weight loss,
symptoms of primary and systemic disease) and outcome (best response and death
within 2 months).
observed scores. In the second set of plots, two covariates are added: complete
or partial response and death within 2 months. The distribution of imputed
values is shifting down slightly, especially for those subjects who are within
the last 2 months of their lives. Note that the range of imputed values is
much smaller than the range of observed values. The implications of this are
discussed later in this chapter.
c2638 C2638TextureC06 February 7, 2002 12:19
EXPLICIT REGRESSION MODELS 123
Conditional predicted values
In the longitudinal setting, the likelihood of missing data often depends on
such observed data (Y
obs
i
) as previous HRQoL assessments. If we wish to
obtain unbiased estimates, it is critical that we include the observed data
in the imputation procedure. Specically, we want to predict values of the
missing HRQoL scores using the previously observed HRQoL scores from that
individual. These conditional estimates are referred to as either empirical best
linear unbiased predictors (EBLUP) or Bucks conditional means [11, 88]. (See
Section 5.6 for more details on EBLUPs.) For both the repeated measures and
random-eects models, the estimates take the form
E
_
Y
mis
i

Y
obs
i
,

B

= X
mis
i

B

mo

1
oo
_
Y
obs
i
X
obs
i

B

_
, (6.5)
Var
_
Y
obs
i
Y
mis
i
_
=
_

oo

om

mo

mm
_
. (6.6)
For a mixed-eects model (Equation 3.9),the estimates are a special case:
E
_
Y
mis
i

Y
obs
i
,

B

,

d

= X
mis
i

B

+ Z
mis
i

d

i
= X
mis
i

B

+ Z
mis
i

DZ
obs
i

1
oo
_
Y
obs
i
X
obs
i

B

_
. (6.7)
Note that the observed data (Y
obs
i
) are now included in the equation used
to predict the missing values for each individual. When the observed HRQoL
scores for the ith subject are higher (or lower) than the average scores for
similar subjects, the dierence (Y
obs
i
X
obs
i

) is positive (or negative). As

a result, the imputed conditional predicted value is larger (or smaller) than the
imputed unconditional value (X
mis
i

). Further, when HRQoL scores within

the same individual are more strongly correlated, the second term in these
equations,

mo

1
oo
(Y
obs
i
X
obs
i

) or Z
mis
i

i
, is larger in magnitude and
the dierence between the conditional and unconditional imputed values will
increase.
Table 6.5 displays the imputed values for the same ve cases from the
NSCLC study for conditional predicted. Imputed values are computed using
the results of a repeated measures model of all available data. For the subjects
Table 6.5. Simple imputation using (EBLUPs) conditional predicted values.
No. Taxol Baseline 6 weeks 12 weeks 26 weeks Months survival
12 No 60.9 51.0 X X 2.5
13 No 82.1 61.8 X X 2.7
60 Yes 57.2 40.6 34.4 X 3.6
527 Yes 63.5 52.4 54.9 43.0 7.2
277 Yes 67.5 71.6 65.9 51.2 7.8
Observed data are displayed in bold and imputed data in italics. NSCLC example.
c2638 C2638TextureC06 February 7, 2002 12:19
124 SIMPLE IMPUTATION
on the Etoposide arm who will survive less than 2 months after the scheduled
6-week assessment, the unconditional predicted values at baseline and 6 weeks
are
Y
mis
ij
= X
mis
ij1

1
= 65.1 + 5.4X
mis
ij1
10.4X
mis
ij2
= 65.1, j = 1, (6.8)
Y
mis
ij
= X
mis
ij2

2
= 62.1 + 6.7X
mis
ij1
9.0X
mis
ij2
= 53.1, j = 2. (6.9)
The conditional predicted value for the 6-week assessment given the baseline
assessment is
Y
mis
i2
|Y
obs
i1
= X
mis
i2

2
+
21

1
11
_
Y
obs
i1
X
obs
i1

1
_
. (6.10)
Specically, for subjects 12 and 13,
Y
mis
i2

Y
obs
i1
= 53.1 +
123.9
243.4
_
Y
obs
i1
65.1
_
= 53.1 + 0.51(60.9 65.1) = 51.0 for subject 12
= 53.1 + 0.51(82.1 65.1) = 61.8 for subject 13.
Note that because subject 13 had a much higher than average baseline value,
the imputed value at 6 weeks is also higher. The equations for the remain-
ing three subjects are messy and involve the inversion and multiplication of
matrices. For subject 60:
_
Y
mis
i3

Y
obs
i1
Y
obs
i2
_
= X

i3

B

3
+ [
31

32
]
_

11

12

21

22
_
1
__
Y
obs
i1
Y
obs
i2
_

_
X
obs
i1
X
obs
i2
__

B

2
_

_
and for subjects 277 and 527:
_
Y
mis
i3
Y
mis
i4

Y
obs
i1
Y
obs
i2
_
=
_
X
obs
i3
X
obs
i4
__

B

4
_

+
_

31

32

41

42
__

11

12

21

22
_
1

__
Y
obs
i1
Y
obs
i2
_

_
X
obs
i1
X
obs
i2
__

B

2
_

_
.
The imputed values are easily obtained using existing software (e.g., SAS
Proc Mixed). Note that in Table 6.5 the imputed values at 12 and 26 weeks
reect both the earlier death of subject 60 and the previously observed scores
for all three subjects.
The entire distribution of observed scores and imputed scores, predicted
by multivariate linear regression, is displayed in Figure 6.4. The same set of
explanatory variables used to generate the data in Figure 6.3 was used to
generate these values. However, the imputed values now include information
about the patients previous HRQoL. As a result, the distribution of the im-
puted values appears to be shifted down and the range of values has greatly
expanded, but is still smaller than for the observed scores.
c2638 C2638TextureC06 February 7, 2002 12:19
LAST VALUE CARRIED FORWARD 125
Figure 6.4. Distribution of observed and imputed values using BLUPs from a re-
peated measures model with baseline characteristics (treatment group, performance
status, prior weight loss, symptoms of primary and systemic disease) and outcome
(best response and death within 2 months).
6.4 Last value carried forward
Another popular approach is to use the last value (or observation) carried
forward (LVCF or LOCF), where the patients last available assessment is
substituted for each of the subsequent missing assessments. Although analy-
ses using LVCF are often reported, the authors rarely justify the use of LVCF
or examine the pattern of change in HRQoL as a function of dropout [120].
c2638 C2638TextureC06 February 7, 2002 12:19
126 SIMPLE IMPUTATION
Table 6.6. Simple imputationLVCF.
No. Taxol Baseline 6 weeks 12 weeks 26 weeks Months survival
12 No 60.9 60.9 X X 2.5
13 No 82.1 82.1 X X 2.7
60 Yes 57.2 40.6 40.6 X 3.6
527 Yes 63.5 52.4 52.4 52.4 7.2
277 Yes 67.5 71.6 71.6 71.6 7.8
Observed data are displayed in bold and imputed data in italics. NSCLC example.
This approach assumes that the last available response of a patient withdraw-
ing from a study is the response that the patient would have if the patient
had remained in the trial. The assumption that HRQoL does not change after
dropout seems inappropriate in most studies. Thus, this approach has limited
utility [60, 64, 94] and should be employed with great caution.
Consider a situation where treatment may be associated with toxicity, such
as the study of adjuvant therapy for breast cancer previously described. In this
study, most women reported poorer HRQoL during therapy than they were
experiencing prior to therapy. Thus, carrying forward their previous base-
line assessment would create an overly optimistic picture of the HRQoL of
subjects on that treatment arm, giving a possible benet to the therapy with
more dropout. As a second example of possible misuse, consider a study where
HRQoL is decreasing over time. In this situation, LVCF would make a treat-
ment with early dropout appear better than a treatment where tolerance to
the therapy was higher.
Table 6.6 displays the imputed values for the same ve cases from the
NSCLC study with the last value carried forward. In each of these cases, this
value is greater than that imputed using information about response, survival,
and prior HRQoL. The entire distribution of observed scores and imputed
scores, using the LVCF, is displayed in Figure 6.5. Imputed scores cover the
entire range of possible scores, with the distribution of imputed scores much
closer to the observed scores than observed in Figure 6.4. In this trial, we
are assuming that a patients HRQoL is the same after dropout as when the
patient completed the last assessment. In this particular setting, that does
not seem to be a reasonable assumption.
-Adjustments
Diehr et al. [34] describe a variation on LVCF to address the problem of dif-
ferences in HRQoL between individuals who were able to complete HRQoL
assessments and those who were not. In the proposed procedure, a value k
is subtracted from (or added to) the last observed value. If this value can
be justied, then this approach is a useful option in a sensitivity analysis.
In Diehrs example, a value of 15 points on the SF-36 physical function scale
was proposed, where 15 points is justied as the dierence in scores between
c2638 C2638TextureC06 February 7, 2002 12:19
LAST VALUE CARRIED FORWARD 127
Figure 6.5. Distribution of observed and imputed values using LVCF.
individuals reporting their health as unchanged vs. those reporting worsening
[157].
Arbitrary high or low value
Another example of simple imputation is the substitution of an arbitrary high
or low value for the missing assessments. This is most commonly used when
the missing HRQoL data are the result of an adverse event such as death [43,
120]. In some cases a value of 0 is used; in others a value just below the
minimum of all observed scores is substituted for the missing data. Both ap-
proaches can be partially justied, but neither can be completely defended.
c2638 C2638TextureC06 February 7, 2002 12:19
128 SIMPLE IMPUTATION
Table 6.7. Strategy for simple imputation of ranks. Assumes that all subjects who
completed the study have an observed response (Y
obs
i
) and all others are missing
data.
Outcome N Assumed value Rank value
1 Died n
1
Y
mis
i
= 0 Lowest tied
rank
2 Withdrawal due to n
2
0 < Y
mis
i
< min(Y
obs
i
) Next lowest
lack of ecacy or tied rank
toxicity
3 Completed study n
3
Y
obs
i
R
obs
i
4 Withdrawal with n
4
max(Y
obs
i
) < Y
mis
i
Highest tied
cure rank
5 Withdrawal not n
5
Omitted
related to toxicity
or ecacy
Lowest tied rank = avg(1, . . . , n
1
). Next lowest tied rank = avg(n
1
+1, . . . , n
1
+n
2
).
R
obs
i
= n
1
+n
2
+Rank of Y
obs
i
among those who completed the study. Highest tied
rank = avg(n
1
+n
2
+n
3
+1, . . . , n
1
+n
2
+n
3
+n
4
).
One approach that avoids some of the controversy is an analysis based on the
ranked HRQoL scores [43, 60, 120], for example, the Wilcoxon Rank Sum test.
Gould [60] suggests classifying the subjects into four groups. The idea is ex-
tended to ve groups in Table 6.7.
The strategy for imputing missing data has the advantage that one only has
to assume a relative ordering of the HRQoL. For example, we are assuming
that the HRQoL of subjects who died is poorer than of those that remain alive.
However, even with this seemingly straightforward procedure, it is important
to consider the assumptions very carefully. In practice it will not be easy to
classify all dropouts into one of the three groups (2, 4, and 5) dened in
Table 6.7. Heyting [64] and Pledger and Hall [114] describe situations where
this strategy may not be appropriate. Also note that when a large proportion
of the subjects expire, this approach becomes an approximation to the analysis
of survival rather than an analysis of HRQoL.
6.5 Underestimation of variance
The major disadvantage of simple imputation techniques is the underestima-
tion of the variance of the observations. There are several reasons for the un-
derestimation. First, when we use means or predicted values, we are assuming
that there is no variability among the missing observations (i.e., Var[
hij
] = 0).
Second, we assume that we know the true mean or predicted value (i.e.,
Var[ ] = 0 or Var[X

] = 0), when, in fact, we have estimated these values.
Finally, we assume that we have observed the HRQoL of all subjects rather
c2638 C2638TextureC06 February 7, 2002 12:19
UNDERESTIMATION OF VARIANCE 129
Table 6.8. Naive estimation of standard deviation (S.D.) and standard error (S.E.)
in the NSCLC study; 26-week estimates for survivors.
Observed S.D. S.E. Ratio
a
of
Imputation Taxol (n) Mean ( )
_

n1
_
S.E.
None No 57 65.0 15.0 1.99 1.00
Yes 134 64.8 15.7 1.36 1.00
Simple mean No 125 65.0 10.1 0.90 0.45
Yes 265 64.8 11.1 0.68 0.63
Covariates No 125 62.8 11.3 1.01 0.51
Yes 265 63.6 11.9 0.73 0.54
Conditional No 125 61.1 13.1 1.18 0.59
Yes 265 62.4 13.5 0.83 0.61
LVCF No 121 63.9 15.9 1.45 0.73
Yes 259 64.3 15.6 0.97 0.73
a
Reference (denominator) is the standard error with no imputation.
than just a subset of the subjects. As a result, test statistics and condence
intervals based on a naive analysis of the observed and imputed data will not
be valid.
This is illustrated in the NCSLC study. Consider the scores at 6 months
(26 weeks); the estimated standard deviation of the observed data is roughly
15 to 16 points.
=

_
n

i=1
(Y
hij


Y
hj
)
2
/(n 1). (6.11)
When we impute values for the missing data using the mean value of the ob-
served data and use a naive estimate of the variance, we add nothing to the
squared terms because Y
mis
hij

Y
hj
= 0, but we do increase the apparent number
of observations. The eect of this is illustrated in Table 6.8, where the naive
estimate of the standard deviation decreases by almost one third as we in-
crease the apparent number of observations by about twofold. In the simple
univariate case, the underestimation of the variance is roughly proportional
to the amount of missing data. The naive estimate of the variance of y
i
is

2
=

n
obs
(y
obs
i
y
obs
i
)
2
+ n
mis
( y
obs
i
y
obs
i
)
2
n
obs
+ n
mis
1
=
(n
obs
1)
2
obs
+ 0
(n 1)
=
n
obs
1
n 1

2
obs
.
When
2
obs

2
, then E[
2
] n
obs
/n
2
. The standard deviation is underes-
timated by a factor proportional to the square root of the amount of missing
data. While it is straightforward to adjust the estimate of the variance using
c2638 C2638TextureC06 February 7, 2002 12:19
130 SIMPLE IMPUTATION
mean imputation, it becomes a much more dicult task for other imputation
procedures.
The problem with the underestimation of the variance of the observations
is compounded when we attempt to estimate the standard errors of means
or regression parameters. This in turn aects test statistics and condence
intervals. For example, the naive estimate of the standard error of the mean
(S.E.( ) =
_

2
/n) assumes that we have information on all n individuals
rather than the n
obs
individuals who completed the HRQoL assessments. In
the 6-month estimates for survivors of the NSCLC study, we analyze a data
set with 389 observations when only 191 subjects were observed. The eect is
illustrated in Table 6.8, where the naive standard errors are roughly half the
true standard errors. This makes a substantial dierence in the test statistics.
Consider a small dierence of 3 points (
1
5
S.D.) in the means of the two groups.
With no imputation, the t-statistic for a test of dierences is 1.24 (p = 0.22).
With mean imputation, the t-statistic is 2.65 (p = 0.008), a highly signicant
dierence. For the other simple imputation methods displayed in Table 6.8,
the estimates of the standard deviations are biased for all approaches but
LVCF, and the standard errors are biased toward zero for all approaches.
6.6 Sensitivity analysis
Despite the limitations of simple imputation, these methods are useful as
part of a limited analysis of the sensitivity of results. Visual comparison of the
means or medians is helpful, but comparisons of test statistics are inappropri-
ate. Several authors have used simple imputation combined with nonparamet-
ric analysis to examine the sensitivity of their results to the assumptions made
about the missing data. Raboud et al. [120] used this approach in a study of
antimicrobial therapy in patients with HIV to demonstrate the sensitivity
of estimates of treatment eects to dierent rates of dropout and survival.
Fairclough et al. [43] demonstrated the consistency of their conclusions, in a
clinical trial of adjuvant therapy for breast cancer, to the assumptions made
about the missing data. The means displayed in Table 6.8 illustrate an ex-
ample of such an analysis. The sensitivity of the results to these dierent
methods of simple imputation should not be unexpected for a study with ex-
tensive dropout. Estimates of the means vary from 61.1 to 65.0 for patients
not receiving Taxol and from 62.4 to 64.8 for those receiving Taxol. These
dierences are approximately
1
4
of the standard deviation.
6.7 Summary
The primary limitation of simple imputation is the underestimation of the
variance of any estimate and the corresponding eect on any test statistic.
Last value carried forward (LVCF), if used, should be well justied and any
underlying assumptions veried. This approach will not be conservative in
all cases and may, in some settings, bias the results in favor of a treatment
that results in more dropout associated with morbidity.
c2638 2638TexturesC07 February 7, 2002 12:21
CHAPTER 7
Multiple Imputation
7.1 Introduction
The major criticism of simple imputation methods is the underestimation of
the variance (see previous chapter). Multiple imputation [125, 126] retains
many of the advantages of single imputation but recties this problem. Mul-
tiple imputation of missing values will be worth the eort only if there is a
substantial benet that cannot be obtained using maximum likelihood meth-
ods for the analysis of incomplete data (Chapter 5). Two quotes summarize
the issues:
Multiple imputation is not a panacea. Although it is a powerful and useful tool
applicable to many missing data settings, if not used carefully it is potentially
dangerous. The existence of software that facilitates its use requires the analyst
to be careful about the verication of assumptions, the robustness of imputation
models, and the appropriateness of inferences. For more complicated models (e.g.,
longitudinal or clustered data) this is even more important. (Norton and Lipsitz
[107])
It is clear that if the imputation model is seriously awed in terms of capturing
the missing data mechanism, then so is any analysis based on such imputation.
This problem can be avoided by carefully investigating each specic application,
by making the best use of knowledge and data about the missing-data mechanism.
(Barnard and Meng [4])
7.2 Overview of multiple imputation
The basic strategy of multiple imputation is to impute 3 to 20 sets of values for
the missing data that incorporate both the variability of the HRQoL measure
and the uncertainty about the missing observations. Each set of data is then
analyzed using complete data methods and the results of the analyses are then
combined. The general strategy is summarized in four steps:
Step 1: Selection of the imputation procedure
Although the least technical, selection of an appropriate imputation procedure
is the most dicult step. There is a variety of implicit and explicit methods
that can be used for multiple imputation [125]. Explicit methods generally uti-
lize regression models, whereas implicit methods utilize sampling techniques.
Four specic examples of these strategies are described in the subsequent
sections.
c2638 2638TexturesC07 February 7, 2002 12:21
132 MULTIPLE IMPUTATION
There are three desirable properties of an imputation procedure. The rst
objective is that it produce unbiased estimates. This is particularly dicult
when there is a suspicion that the missingness depends on the HRQoL of
the individual at the time the observation was to be made (MNAR). Sec-
ond, the procedure should incorporate the appropriate variation to reect the
randomness of the observations, the loss of information due to the missing ob-
servations, and the uncertainty about the reason for nonresponse. Finally, the
procedure should not distort the covariance structure of the repeated measure-
ments in longitudinal studies, especially when the analyses involve estimation
of changes in HRQoL over time. The challenge is to select an appropriate
model and fully understand the assumptions that are being made when that
model is implemented. This is the focus of the subsequent sections.
Step 2: Generation of M imputed data sets
Once the model and procedure for imputing the missing observations are se-
lected, the details for implementing the procedure are generally well worked
out (see Sections 7.3 through 7.5). One of the questions often asked is, How
many datasets should be imputed? There is no straightforward answer. The
recommendations generally range from 3 to 10. As the proportion of missing
data increases, the between imputation variance will increase (see Section 7.7)
and will become more important in the calculation of the overall variance. In-
creasing the number of imputed data sets (M) will improve the precision of
this estimate. Imputing more rather than fewer (10 or 20) data sets is recom-
mended, as the additional time required is a trivial part of the entire analytic
eort with modern computing power.
Step 3: Analysis of M data sets
Each of the M data sets is then analyzed using either complete data or
maximum likelihood methods such as MANOVA or mixed-eects models.
In this step, estimates of the primary parameters (

(m)
) and their variance
(Var[

(m)
]) are obtained for each of the M complete data sets. Step 3 is often
expanded to include hypothesis testing (H
0
: = C = 0) with the estimation
of secondary parameters (

(m)
= C

(m)
) and their variance (Var[

(m)
]).
Step 4: Combining results of M analyses
The results of the M analyses are then combined by averaging the parameter
estimates from each of the M data sets. Finally, the variance of these estimates
is computed by combining within and between components of the variance of
the estimates from the M data sets. As is shown later, it is the combination of
the imputation procedures and the computation of the variance estimates that
results in the appropriate estimates of variance. Full details for this procedure
are presented in Section 7.7.
c2638 2638TexturesC07 February 7, 2002 12:21
EXPLICIT UNIVARIATE REGRESSION 133
7.3 Explicit univariate regression
In the explicit regression model approach to multiple imputation, we identify
a regression model to predict the missing observation. This is basically the
same model that is used for the simple imputation procedure described in the
previous chapter. However, the multiple imputation procedure diers in two
ways. First, because the true parameters of the imputation model are un-
known, random error is added to the estimated parameters (

B

).* These new

values of the parameters (
(m)
) are then used to predict the average HRQoL
for a subject with specic characteristics dened by the covariates (X
(mis)
).
Then, additional random error is added to these values to reect the natural
variability of the individual outcome measures (Var[Y
i
]).
Analytic model: Y
i
= X
i
+
i
Imputation model: Y
obs
i
= X
obs
i
B

i
Identication of the imputation model
The explicit regression approach has the advantage that the regression model
used to impute missing observations can readily use auxiliary information
about the subjects HRQoL that could not be included in the model used for
the ultimate analysis of the clinical trial. This additional information might
include indicators of the side eects of treatment and the clinical course of
the disease when the therapy succeeds or fails. Alternatively, it could include
measures of HRQoL provided by a caregiver when the subject is no longer
able to complete self-assessments.
Recall that one of the objectives of imputation is to obtain unbiased es-
timates of the missing observations. In the context of the explicit regression
approach, we are looking for a model where the missingness in the imputation
model depends only on the observed data included in the imputation model
(Y

i
) and the covariates in the imputation model (X

i
). Basically, we are at-
tempting to augment the analytic model with auxiliary outcome data (Y

i
) or
covariates (X

i
), so that the MAR assumption in the imputation model seems
reasonable. The primary candidates for covariates include measures that are
strongly correlated with both the measure being imputed and the probability
that it is missing. In the NSCLC example, this suggests adding covariates
associated with toxicity, disease progression, and the proximity of the patient
to death. Assuming that the objective of the analysis is the comparison of
treatment groups, the same strategy outlined in Section 6.2 is recommended
for identication of the imputation model. To avoid biasing the comparison of
treatment groups toward the null hypothesis, it is critical that imputation of
* We use the superscript asterisk to dierentiate the imputation model from the analytic
model.
c2638 2638TexturesC07 February 7, 2002 12:21
134 MULTIPLE IMPUTATION
postrandomization data either be done separately for each treatment group
or include treatment as a covariate.
Computation of imputed values
The general procedure for imputing M sets of missing values for univariate
data is as follows [21, 94, 126]:
1. Estimate the parameters of the regression model (Y
obs
i
= X
obs
i
B

i
)
using the observed data.

=
_

X
obs
i
X
obs
i
_
1
X
obs
i
Y
obs
i
, (7.1)

2
=

_
Y
obs
i


B

X
obs
i
_
2
/(n
obs
p

), (7.2)
where p

is the number of unknown parameters in B

.
2. Generate M sets of model parameters (
(m)
and
2(m)
) by adding random
error to the estimates reecting the uncertainty of the estimates.

2(m)
=
2
(n
obs
p

)/K
(m)
, (7.3)
U

= Var[

B

] =
2(m)
_

X
obs
i
X
obs
i
_
1
, (7.4)

(m)
=

+U

Z
m

, (7.5)
where
K
(m)
=a randomly drawn number from a -square distribution
with n
obs
p

degrees of freedom
U

=the upper triangular matrix of the Cholesky decomposition of

the variance of

B

Z
(m)

=a vector of p

random numbers drawn from a standard

normal distribution
The Cholesky decomposition is also referred to as the square root method.
In this procedure, we have assumed that the parameters have a normal
distribution with variance approximately equal to
2
(

X
obs
i
X
obs
i
)
1
.
This is a reasonable assumption for large studies but may not be valid
for smaller studies. When there is concern about these assumptions, more
complex methods of generating these parameters are required [129].
3. Generate the imputed values of Y
mis
i
, for the mth imputation, by adding
random error corresponding to the between- and within-subject variability
of the outcome (Var[Y
i
] =
2
).
Y
mis(m)
i
= X
mis
i

(m)
+
(m)
Z
(m)
Y
, (7.6)
where Z
(m)
Y
is a random number drawn from a standard normal distribu-
tion.
4. The second and third steps are repeated for each of the M data sets.
c2638 2638TexturesC07 February 7, 2002 12:21
EXPLICIT UNIVARIATE REGRESSION 135
Practical considerations
There are two major barriers to implementing this procedure. The rst is a
small sample size. The second is the lack of strong predictors of the HRQoL
measure. When the sample size is small, the analyst is unable to obtain pre-
cise estimates of the imputation model parameters (

B

). This will result in a

wide range of values of
(m)
. Thus, the imputed values may only add noise
to the observed data. The lack of strong predictors will have similar conse-
quences. If the variation in the outcome explained by the imputation model
is small (the R
2
is small), the values of
2(m)
will be large and, again, the
imputed values may only add noise to the observed data. An additional nui-
sance associated with both problems is that some of the predicted values may
lie outside the possible range of values of the HRQoL scale.
The above description of the procedure for identifying the imputation model
does not specify a cuto for the signicance of potential covariates. This aspect
of the model-tting procedure is as much art as it is science. Adding more
covariates to the imputation model will improve the procedure by increas-
ing the R
2
and reducing
2(m)
, thus decreasing the statistical noise added in
Equation 7.6. However, this is balanced by adding parameters with large vari-
ance and increasing the statistical noise added in Equation 7.5. Thus, adding
covariates that result in small increases in the R
2
is unlikely to improve the
imputation procedure, even when the statistical signicance of a particular
parameter is large.
Extensions to longitudinal studies
The above procedure was developed for cross-sectional studies. However, in
most clinical trials HRQoL is measured longitudinally. Little and Yau [94]
suggest a sequential procedure for a monotone missing data pattern. The
procedure is rst to ll in the missing values of the rst observation (Y
i1
),
generating M sets of data for the rst observation (Equation 7.7). The second
step is to ll in the missing values of the second observation (Y
i2
), given the
observed (Y
i1
) or imputed values (Y
(m)
i1
) of the rst observation (Y
i1
). Note
that in this step, only one set of new values is generated for each of the M data
sets. Subsequent missing values are imputed using all previously observed and
imputed data (Equations 7.8 and 7.9). The imputation models at each step are
Y
obs
i1
= X
obs
i1

1
+

i1
, (7.7)
Y
obs
i2
= X
obs
i2

(m)
2
+Y
(m)
i1

(m)
2|1
+

i2
, (7.8)
Y
obs
i3
= X
obs
i3

(m)
3
+Y
(m)
i1

(m)
3|1
+Y
(m)
i2

(m)
3|2
+

i3
. (7.9)
Assumptions
When we impute the missing observations using this model, we are making at
least two assumptions that are not testable. First, we are assuming that the
relationship between the explanatory variables and HRQoL is the same when
c2638 2638TexturesC07 February 7, 2002 12:21
136 MULTIPLE IMPUTATION
individuals complete the HRQoL assessments and when they do not complete
the assessments. Basically, we are assuming that Y
obs
i
= X
obs
i
B

+
i
and
Y
mis
i
= X
mis
i
B

+
i
are both true. Second, we are assuming that we have
identied all the important relevant covariates such that the missingness no
longer depends on the missing HRQoL value (Y
mis
i
). As previously mentioned,
we have no way formally to test the assumption that the missing data are
MAR, given the observed data (Y
obs
i
) and covariates in the imputation model
(X
obs
i
).
Less critical assumptions in this procedure are that the residual errors (

i
)
and the parameter estimates (

B

) of the imputation model are normally dis-

tributed. The rst assumption can be assessed by examining the residual errors
(

i
), and the second is true for studies with moderate sample sizes.
NSCLC example
Although the procedure described above sounds straightforward, one quickly
realizes that there are many choices to be made. One issue is how to handle the
small proportion of subjects with nonmonotone missing data patterns. In the
following example, we have ignored later observations rather than excluded
these individuals from the analysis. A second choice is whether to run the im-
putation procedures separately for each treatment group or to include treat-
ment as a covariate in all imputation models. By taking the second option,
we run the risk of missing a potential interaction between one of the covari-
ates and treatment but gain precision in the estimation of the parameters
when the relationships are the same in each treatment group. This later ad-
vantage is more important when the sample size is small. Another decision
is how to handle imputed values that are out of the range of the scale. As
will be demonstrated later, this is a very minor issue and does not aect the
results.
The covariate (X

ij
) selection process had three steps. The rst was to iden-
tify potential variables. This selection was made on a conceptual basis and
summarized in Table 7.1. For example, treatment assignment should not have
any impact on the baseline scores if patients did not know their treatment as-
signment prior to completing the FACT-Lung questionnaire. Similarly, treat-
ment outcomes were not considered as predictors of the baseline scores but
were considered for all follow-up assessments. Potential covariates with more
than 5% missing data were not considered. The second step was to eliminate
variables where the correlation with the TOI scores was weaker (p > 0.001).
Finally, a backward elimination procedure was used with multivariate regres-
sion where the selection criterion was set at = 0.01. Because the intent of
the nal analysis is to compare the treatment groups, the treatment indicator
(TAXOL) was forced into all imputation models even though it was nonsigni-
cant. Previously imputed values from the rst set of imputed values (m = 1)
were included in these regression analyses. The variables excluded in this
third step would have explained only 1 to 2% more variation. The resulting
c2638 2638TexturesC07 February 7, 2002 12:21
EXPLICIT UNIVARIATE REGRESSION 137
Table 7.1. Covariate selection for multiple imputation with sequential univariate
regression models.
Assessment no.
Potential
covariate 1 2 3 4
Taxol X X X
Age
Gender
Performance status +
Weight loss +
Prior radiotherapy
Primary disease symptoms +
Metastatic disease symptoms
Systemic symptoms +
Associated chronic diseases
Baseline TOI score + +
6-week TOI score +
12-week TOI score +
Cycles of therapy
Complete or partial response +
Hemotological toxicity
Neurological toxicity
Withdrawal due to progressive disease
Survival (log transformation) + + +
R
2
0.19 0.32 0.35 0.34
X indicates forced into model; indicates tested but not included; + indi-
cates included.
R
2
values, summarized in Table 7.1, explain a moderate amount of the vari-
ation in FACT-Lung TOI scores.
SAS Versions 8.1 and 8.2 have an experimental procedure (Proc MI) that
can be used to implement this series of imputations. Because the dropout pat-
tern is not strictly monotone and the covariates dier over time, a series of im-
putations will be necessary. The rst step is to create a data set (WORK.ONEREC)
with one record per subject containing the four possible FACT-Lung TOI
scores (TOI1, TOI2, TOI3, TOI4) and the covariates to be used in the MI
procedure. Next, the missing values for the baseline assessment (TOI) are im-
puted with the selected baseline covariates (WT LOSS ECOGPS SX SYS SX PRI).
The MI procedure with a MONOTONE METHOD=REGRESSION; statement sequen-
tially imputes missing data using
Y
j
=
0
+
1
Y
1
+ +
j1
Y
j1
for the Y
1
, . . . , Y
j
identied in the VAR statement, with the restriction that
c2638 2638TexturesC07 February 7, 2002 12:21
138 MULTIPLE IMPUTATION
the missing data pattern is monotone among the variables for the order
specied.
PROC MI DATA=WORK.ONEREC OUT=WORK.MIDATA1 NIMPUTE=20;
MONOTONE METHOD=REGRESSION;
VAR WT_LOSS ECOGPS SX_SYS SX_PRI TOI1;
RUN;
This creates a new data set with 20 sets of imputed values for missing base-
line assessments. In the following steps we are going to use separate models
for each of the two treatment groups. We will use imputed missing baseline
values (TOI1) in the imputation model for the 6-week values (TOI2), imputed
6-week values (TOI2) in the imputation model for the 12-week data (TOI3),
etc. To do this we are going to rename the variable that identies these 20
sets from IMPUTATION to M and sort the data by imputation number and
treatment group.
PROC SORT DATA=WORK.MIDATA1
OUT=WORK.MIDATA2(RENAME=(_IMPUTATION_=_M_);
BY _IMPUTATION_ TAXOL;
RUN;
From this point on, we are going to impute only one set of values for each of the
20 sets generated in the rst step using the log of the survival times (LN SURV),
best response (CRPR), and previous FACT-Lung TOI scores as explanatory
variables.
PROC MI DATA=WORK.MIDATA2 OUT=WORK.MIDATA3 NIMPUTE=1;
BY _M_ TAXOL;
MONOTONE METHOD=REGRESSION;
VAR LN_SURV TOI1 TOI2;
RUN;
PROC MI DATA=WORK.MIDATA3 OUT=WORK.MIDATA4 NIMPUTE=1;
BY _M_ TAXOL;
MONOTONE METHOD=REGRESSION;
VAR LN_SURV CRPR TOI1 TOI2 TOI3;
RUN;
PROC MI DATA=WORK.MIDATA4 OUT=WORK.MIDATA5 NIMPUTE=1;
BY _M_ TAXOL;
MONOTONE METHOD=REGRESSION;
VAR LN_SURV TOI1 TOI2 TOI3 TOI4;
RUN;
Figure 7.1 displays the observed and imputed scores at the third and fourth
assessment. As expected in this study, the distribution of the imputed scores is
lower than the observed scores. A very small proportion, less than 0.1%, of the
scores lie outside of the possible range of 0 to 100. Low values (<0) represent
0.01, 0.03, 0.16, and 1.84% of the scores at each of the four assessments. High
values (>100) represent 0.14, 0.22, 0.19, and 0.26% of the scores, respectively.
Even when these out-of-range scores are replaced with 0 and 100, there is
little impact on the overall distribution of scores (Figure 7.1) or on the results
of the analysis of the imputed data sets (Table 7.2).
c2638 2638TexturesC07 February 7, 2002 12:21
EXPLICIT UNIVARIATE REGRESSION 139
Figure 7.1. Distribution of observed and multiply imputed values generated with
sequential univariate regression.
A more dicult decision is how to handle the observations that would have
occurred after the death of the patients. There is considerable controversy
about imputing HRQoL scores after the patient has died. I am not going to
attempt to resolve the existential question about rating the quality of life after
death. I would like to point out the statistical implications of not using im-
puted values after death in the analyses. When observations occurring after
death are missing, we assume that the data are MAR and that HRQoL in the
c2638 2638TexturesC07 February 7, 2002 12:21
140 MULTIPLE IMPUTATION
Table 7.2. Results of the sequential univariate MI procedures (M = 20).
Out-of-range Out-of-range After death
not replaced replaced excluded
Estimate Taxol Mean (S.E.) Mean (S.E.) Mean (S.E.)
Baseline No 64.8 (1.2) 64.8 (1.2) 64.8 (1.2)
Yes 66.1 (0.9) 66.1 (0.9) 66.1 (0.9)
6 weeks No 59.7 (1.6) 59.7 (1.6) 60.7 (1.6)
Yes 62.8 (1.1) 62.8 (1.1) 63.7 (1.0)
12 weeks No 58.0 (1.6) 58.0 (1.7) 60.1 (1.6)
Yes 59.9 (1.5) 59.9 (1.4) 61.8 (1.2)
26 weeks No 50.9 (2.9) 51.0 (2.8) 57.2 (2.0)
Yes 54.0 (2.3) 54.2 (2.0) 59.8 (1.3)
6-week change No 5.1 (1.6) 5.1 (1.6) 4.1 (1.6)
Yes 3.3 (1.1) 3.3 (1.0) 2.3 (1.0)
Di 1.8 (1.7) 1.8 (1.7) 1.8 (1.7)
12-week change No 6.8 (1.8) 6.8 (1.8) 4.7 (1.6)
Yes 6.2 (1.6) 6.1 (1.5) 4.2 (1.3)
Di 0.7 (2.1) 0.7 (2.1) 0.5 (2.0)
26-week change No 13.9 (3.0) 13.8 (2.9) 7.6 (2.1)
Yes 12.0 (2.3) 11.8 (2.1) 6.3 (1.3)
Di 1.8 (2.7) 1.9 (2.6) 1.3 (2.5)
patient who has died is similar to HRQoL in the survivors (conditional on
the observed scores). As discussed in Chapter 5, values used in the E-step of
the EM algorithm are imputed using the BLUPs. In Figure 7.2A, the aver-
age estimates of the observed and imputed data are displayed by the time of
death (<6 weeks, 6 to 12 weeks, 12 to 24 weeks, >24 weeks). The averages de-
cline over time, more quickly in patients who die earlier. Figure 7.2B displays
the corresponding estimates where observations after death are replaced by
the BLUPs generated by the MIXED procedure, with FACT-Lung TOI scores
after death set to missing. Regardless of whether the decision is to include or
exclude the imputed values after death, plots similar to Figures 7.2A and B
are essential to the interpretation of the results.
7.4 Closest neighbor and predictive mean matching
There are many variations on the explicit regression model approach to multi-
ple imputation. Both closest neighbor [126, 153] and predictive mean matching
[63, 125] have the advantage that it is not possible to impute a value that is
out of range for the HRQoL scale. They are also more likely to be robust to
deviations from the normality assumptions. But the more critical assumption
is that of equivalent models for subjects with observed data and subjects with
missing data.
c2638 2638TexturesC07 February 7, 2002 12:21
CLOSEST NEIGHBOR AND PREDICTIVE MEAN MATCHING 141
Figure 7.2. Average FACT TOI scores by time of death. (A) Estimates include
imputed values after death; (B) estimates include BLUP for values after death.
c2638 2638TexturesC07 February 7, 2002 12:21
142 MULTIPLE IMPUTATION
Closest neighbor
The initial steps for the closest neighbor procedure are the same as those for
the explicit regression procedure. M sets of parameter estimates for the im-
putation model are generated as previously described (Equations 7.1 through
7.5). The dierence is in the third step.
3. Predicted values are generated for subjects with both observed and missing
data:
E
_
Y
obs(m)
i

= X
obs
i

(m)
, (7.10)
E
_
Y
mis(m)
i

= X
mis
i

(m)
. (7.11)
4. For each subject (i

) with a missing observation, the subject with the closest

predicted value, E[Y
obs(m)
i
], is selected and that subjects actual observed
value, Y
obs
i
, is the imputed value for subject i

.
5. Repeat steps 2 to 4 for M data sets.
Predictive mean matching
1. In the initial step of predictive mean matching [63, 125], M bootstrap
samples are generated from the subjects with observed data. This consists
of sampling N
obs
subjects at random with replacement.
2. The parameters of the imputation model (

(m)
) are then estimated within
each of the M samples.
3. Predicted values are generated for both the cases with observed data and
the cases with missing data (Equations 7.10 and 7.11).
4. The ve nearest matches among the subjects with observed data are iden-
tied for each subject with missing data and one of the ve is selected at
random. The observed value of the HRQoL measure from that subject is
the imputed value.
7.5 Approximate Bayesian bootstrap
An alternative to explicit model-based imputation is implicit or sampling-
based imputation. One commonly used method is approximate Bayesian boot-
strap (ABB). This was proposed initially for ignorable missing data from
simple random samples [125127]. The basic concept is rst to draw a set
of potential responses, at random with replacement, from a subset of sub-
jects with the same characteristics (X

). In the second step, the imputed

values are drawn from the set of potential responses. This double sampling
ensures proper variation of the responses [127], assuming large enough sets of
subjects with the same set of covariates.*
* The alternative procedure, simple hot-deck imputation, of simply drawing the imputed
values from the set of observed values does not properly reect sampling variability [125].
c2638 2638TexturesC07 February 7, 2002 12:21
APPROXIMATE BAYESIAN BOOTSTRAP 143
Table 7.3. ABB imputation for a bivariate response (Y) with a dichotomous covari-
ate (X) with a monotone missing data pattern.
Subject(i) X
i
Y
i1
Y
i2
1 A 1 1
2 A 1 0
3 A 0 0
4 B 1 1
5 B 1 (1, 1)
6 B 2 1
7 B 2 2
8 B (1, 2) (1, 2)
The two sets (M = 2) of imputed values are
indicated as (*,*).
The basic procedure
To illustrate the procedure, consider the following simple example with a bi-
variate response (Y
1
, Y
2
), a dichotomous covariate and monotone missing data
pattern (Table 7.3). Two subjects have missing observations. Both outcomes
are missing for subject 8 and the second is missing for subject 5. For purposes
of illustration, we will let M = 2.
1. Impute M values of the rst response variable for all subjects with missing
values (Y
81
in our example) matching on the covariate X.
a. Draw n
obs
values from the n
obs
subjects with observed data for the rst
response and matching covariates. There is a pool of four potential sub-
jects (4, 5, 6, and 7) with matching values of the covariate for subject
8. For the rst imputation, draw four values of Y
1
at random with re-
placement from the responses for those four subjects (e.g., 1, 1, 1, 2).
b. Draw a single value (e.g., 1) randomly from that set of potential values.
c. Repeat the same procedure M times. In our example, the second poten-
tial set of values may consist of 1, 2, 2, 1 and the value 2 is drawn from
that set.
2. Impute values for subjects who are missing the second response variable
(y
52
and y
82
), using the original data set with y
81
= 1 for the rst imputa-
tion and then with y
81
= 2 for the second imputation. In the rst imputa-
tion, subject 4 is the only possible subject with X = B, Y
1
= 1 to impute
the values of y
52
and y
82
. In the second imputation, subject 4 is still the only
possible subject with X = B, Y
1
= 1 to impute y
52
, whereas subjects 6 and
7 are possible donors for y
82
.
More formally, the procedure is as follows. Consider a set of n subjects with
the same values of X, where a univariate outcome was observed for n
obs
and
missing for n
mis
subjects (n = n
obs
+n
mis
). For m = 1, . . . , M, rst randomly
select a set of n
obs
possible values with replacement from the n
obs
values.
c2638 2638TexturesC07 February 7, 2002 12:21
144 MULTIPLE IMPUTATION
Then choose n
mis
observations at random with replacement from each of the
M potential samples.
Extensions to longitudinal studies
This procedure can be extended to situations where there are multiple mea-
surements (Y
1
, . . . , Y
J
) with a monotone missing data pattern. The assump-
tion is that the values of Y
j
are missing at random conditional on X and
Y
1
, . . . , Y
j1
. After imputing M sets of values for the rst measurement (Y
1
),
subjects are grouped by both X and Y
1
. Within these groups, a possible
sample of n
obs
observations of Y
2
is randomly selected from which the n
mis
imputed values are randomly selected. This procedure is repeated for each
subsequent measure for each of the M sets of imputations. Although this
extension is theoretically possible, the practical limitations quickly become
apparent. With each step to impute missing values of Y
k
, the possible combi-
nations of X and Y
1
, . . . , Y
j1
increase so that ultimately there are insucient
subjects with a particular combination to provide an adequate sample. When
Y
1
is continuous, the possible combinations of X and Y
1
can be extremely
large, and it will not be possible to impute values of Y
2
using this procedure
without some modications.
Propensity scores
In the previously described procedure, the rst step was to identify patient
strata with increasing likelihood for missing observations. These strata could
be dened by one or two patient characteristics as well as the previously
observed HRQoL measures. In the NSCLC study, tumor responses, length of
survival, cycles completed, and progressive disease during therapy are possible
candidates. As noted above, the number of strata becomes unmanageable
when the number of characteristics used to dene the strata increases or if
the characteristics are continuous.
To solve this problem, Lavori et al. [84] propose a strategy based on a
propensity score for the probability of dropout. The idea is to nd a univari-
ate score that summarizes the multiple variables dening the history of each
patient, stratify patients based on that score, and then use ABB to impute the
missing values within each stratum. In this approach, a logistic model is used
to estimate the probability of a missing assessment, which can be a function
of multiple covariates and previously observed and imputed HRQoL scores.
Strata are then formed on the basis of quintiles or quartiles of the predicted
probability of a missing observation.
Practical issues
Although the procedure sounds straightforward, there are a number of chal-
lenges in the implementation of this procedure for HRQoL in longitudinal
studies. Again, the sample size needs to be large enough to implement the
c2638 2638TexturesC07 February 7, 2002 12:21
APPROXIMATE BAYESIAN BOOTSTRAP 145
procedure. To guard against biasing the results toward the null hypothesis,
the imputation procedures must be done within treatment groups when the
objective of the subsequent analysis is to compare treatment groups. Within
the treatment groups, it may become necessary to reduce the number of strata
when the stratum with the highest propensity for missing data has too few
observed values.
In many studies, the missing data pattern will not be strictly monotone as
was the case for the study described by Lavori et al. [84]. In the NSCLC ex-
ample, approximately 15% of the cases have a missing assessment followed by
an observed assessment. One option is not to impute values for these missing
data. If documentation of the reasons for missing observations indicates that
this is a random event, such as an administrative error, then this is a reason-
able strategy. Because we now have readily available software for longitudinal
analyses that will not exclude these subjects, it is not necessary to impute
every value. The other option is to ignore the subsequent observations when
imputing the missing data. This is a preferable strategy if these intermittent
missing values tend to be associated with episodes of acute toxicity or other
conditions likely to impact HRQoL.
The assumptions
The critical assumption of the ABB procedure is that the data are missing at
random conditional on a set of covariates (X) and previously observed values
of the outcome. However, when the propensity scores are used to dene strata,
the missing data are assumed to occur completely at random within the strata
dened by the propensity score. This is a strong assumption that is dicult
to justify for HRQoL studies where there is a strong association between
missing data and clinical outcomes. Increasing the number of strata dened
by the propensity scores would improve the likelihood that this assumption is
correct. But this may also result in strata with too few subjects available for
drawing the imputed values.
Another limitation occurs when estimates of longitudinal change are of
interest. The ABB procedure with propensity scores uses only the information
about the association of patient characteristics and previous HRQoL with
whether the HRQoL values are missing. It does not use any information about
the correlation among the variables. Thus, it is useful when the intent of the
analysis is to make cross-sectional comparisons among groups, but it is not
appropriate for the estimation of change over time. Similarly, if missing values
of covariates are imputed with this approach, then the relationship between
the covariate and the outcome may be underestimated.
Nonignorable missing data
To deal with nonignorable missing data, Rubin and Schenker [127] suggest
independent draws of the n
obs
possible values, where the probability of sam-
pling an observation is proportional to some function of Y (e.g., Y
2
or

Y ) to
c2638 2638TexturesC07 February 7, 2002 12:21
146 MULTIPLE IMPUTATION
increase the proportion of large or small values of Y . Another approach that
would increase the probability of drawing a smaller (or larger) value to replace
the missing observation is to randomly draw k (usually 2 or 3) observations
and take the minimum (or maximum) of the k draws. VanBuuren et al. [153]
suggest a delta-adjustment where a predetermined constant is subtracted from
the imputed value. Unfortunately, the dicult choice among these possible ap-
proaches cannot be made on the basis of the observed data or any statistical
tests. In practice, testing a range of choices as part of a sensitivity analysis is
appropriate.
7.6 Multivariate procedures for nonmonotone missing data
To this point, we have discussed methods for monotone missing data patterns.
The Markov Chain Monte Carlo (MCMC) method was adapted to handle the
more complex problem of nonmonotone patterns. Extending the univariate
regression methods described in Section 7.3 to the multivariate setting requires
the generation of random numbers for multidimensional and often intractable
probability distributions. The MCMC method simulates the joint posterior
distribution, assuming the data are from a multivariate normal distribution.
A detailed description of this method is beyond the scope of this book, and
the interested reader is referred to Schafer [129].
NSCLC example
Following the same logic as in Section 7.3, we will use separate imputation
models for the missing baseline and follow-up data. The imputation of the
missing baseline data is exactly the same as before. In contrast to the previous
procedure, where the follow-up assessments were imputed in three steps, one
single step is used to impute missing scores for TOI2, TOI3, TOI4 with the
MCMC procedure, including the log transform of the survival time (LN SURV),
an indicator of a complete or partial response (CRPR), and the baseline scores
(TOI1). One of the variables in this application is dichotomous and thus does
not have a normal distribution. However, in moderately large data sets when
dichotomous (and other non-normal) variables are included as an explanatory
variable rather than one of the variables to be imputed, the consequences of
violating the multivariate normal assumptions will be minimal. When trans-
formations are possible (such as for the survival times), they should be utilized.
PROC MI DATA=WORK.MIDATA2 OUT=WORK.MIDATA5 NIMPUTE=1;
BY _M_ TAXOL;
MCMC;
VAR LN_SURV CRPR TOI1 TOI2 TOI3 TOI4;
RUN;
The distribution of the resulting data and the estimates are almost identical
to those displayed in Figures 7.1 and 7.2.
c2638 2638TexturesC07 February 7, 2002 12:21
COMBINING THE M ANALYSES 147
7.7 Combining the M analyses
For each of the imputed data sets, the parameter estimates and their variance
are computed using an appropriate analytic method (Table 7.4). In addition,
linear combinations of the parameters (

(m)
= C

(m)
), which will be used to
test specic hypotheses (for example, H
0
:

(m)
= 0) and the corresponding
variance (U
(m)

= Var(

(m)
)), may be estimated using each of the M data
sets. The combined parameter estimates (

and

) are obtained by averag-
ing the M within-imputation estimates (

(m)
and

(m)
). The total variance
of the parameter estimates incorporates both the average within imputation
variance of the estimates (

and

U

) and the between imputation variabil-

ity of the M estimates (B

and B

). The total variance is computed by the

sum of the within-imputation component (

U) and the between-imputation

component (B) weighted by a correction for a nite number of imputations
(1 +M
1
).
The fraction of information about the parameters that is missing due to
nonresponse () for each parameter can be estimated from the ratio of the
between-imputation variance to the total variance. As the amount of missing
data increases, increases. In contrast, as the amount of information about the
missing data in the covariates of the imputation model increases, decreases.
Table 7.4. Computations for combining the analyses of the M data sets.
Within imputation
Estimates

(m)
Variance of estimates U
(m)

= Var(

(m)
)
Between imputation
Variance between estimates B

=
1
M1

M
m=1
(

(m)

)
2
Pooled analyses
Pooled estimates

=

M
m=1

(m)
/M
Pooled within imputation variance

U

=

M
m=1
U
(m)

)/M
Total variance V

=

U

+ (1 +
1
M
)B

Miscellaneous
Between to within variance ratio r

= (1 +M
1
)B

/U

Fraction of missing information

a

= (1 +M
1
)B

/V

Univariate test statistics and condence intervals

Degrees of freedom
b

= (M 1)(1 +r
1

)
2
t-statistic t

=

/V
1/2

100(1 )% interval

t

,(1/2)
V
1/2

a
Fraction of information about (or ) that is missing due to nonresponse
( = r/(1 + r)).
c2638 2638TexturesC07 February 7, 2002 12:21
148 MULTIPLE IMPUTATION
For tests with a single degree of freedom ( is scalar), condence interval
estimates and signicance levels can be obtained using a t-distribution with
degrees of freedom [125, 126]. For tests with multiple degrees of freedom (
is a k 1 vector of linear contrasts), signicance levels can be obtained using
the multivariate analogue [127]. The statistic
F = (
0

U
1

(
0

)/[(1 +r)k] (7.12)

has an approximate F distribution with
1
and
2
degrees of freedom.
1
= k
and
r

=
_
1 +
1
M
_
trace
_
B

U
1

_
/k.

2
= 4 + [k(M 1) 4](1 +a/r)
2
,
a = 1 2/[k(M 1)] if k(M 1) > 4, or

2
= (k + 1)(M 1)(1 +r)
2
/2 if k(M 1) 4.
Both procedures assume the data set is large enough that if there were no
missing values, the degrees of freedom for standard errors and denominators
of F statistics would be eectively innity. Barnard and Rubin [5] suggest an
adjustment for the degrees of freedom for small sample cases. However, in most
cases, if the data set is large enough to use multiple imputation techniques,
this adjustment will not be necessary.
SAS example
In the following example, PROC MIXED is used to generate estimates of the
primary and secondary parameters. Now that the data are complete, other es-
timation procedures could also be used. In this example, the M imputations
are identied by the variable IMPUTATION . For the model, FACT-T2=FUNO,
the primary parameters are the means at each time point. These are stored in
a table labeled SOLUTIONF and exported to a SAS data set using the ODS state-
ment below. Secondary parameters are generated by the ESTIMATE statement
and also exported to a SAS data set. In this example, these are the estimates
of change from baseline. The resulting WORK.MIXPARMS and WORK.MIXESTS
data sets will each contain M sets of parameter estimates. They are merged
and the results are combined using a macro labeled MI SUM.
ODS LISTING CLOSE; * Suppresses listing of Proc Mixed output *;
*** Run analysis of multiply imputed data ***;
PROC MIXED DATA=WORK.TOI3;
BY _IMPUTATION_;
CLASS FUNO;
MODEL FACT_T2=FUNO/NOINT SOLUTION COVB;
REPEATED FUNO/SUBJECT=CASEID TYPE=UN;
ESTIMATE 6 WK CHANGE FUNO -1 1 0 0;
ESTIMATE 12 WK CHANGE FUNO -1 0 1 0;
c2638 2638TexturesC07 February 7, 2002 12:21
COMBINING THE M ANALYSES 149
ESTIMATE 6 MO CHANGE FUNO -1 0 0 1;
ODS OUTPUT SOLUTIONF=WORK.MIXPARMS
ESTIMATES=WORK.MIXESTS;
RUN;
ODS LISTING; * Turns listing back on *;
*** Runs macro that combines the results ***;
%MI_SUM(_IMPUTATION_,WORK.MIXPARMS,WORK.MIXESTS,WORK.SUMDATA2,FUNO);
Details of the SAS Macro MI SUM are as follows. The code assumes that
primary and secondary parameter estimates are exported from Proc MIXED
(V8.1) using SOLUTIONF and ESTIMATES options.
%MACRO MI_SUM(M_VAR,BDATA,TDATA,SUMDATA,FIXED);
****************************************************************;
*** Macro designed to be used when data are analyzed ***;
*** using PROC MIXED ***;
*** USER SUPPLIED ARGUMENTS: ***:
*** M_Var - Variable that identifies imputation # ***;
*** BDATA - SAS dataset created by ODS SOLUTIONF statement ***;
*** TDATA - SAS dataset created by ODS ESTIMATES statement ***;
*** SUMDATA - Output dataset ***;
*** FIXED - List of covariates in the MODEL statement ***;
****************************************************************;
*** Combines primary and secondary parameter estimates ***;
DATA WORK.POOLED;
SET &BDATA &TDATA;
BY &M_VAR;
*** Creates Identifier of Parameter ***;
IF FIRST.&M_VAR THEN ESTNO=0;
ESTNO=ESTNO+1;
RETAIN ESTNO;
*** Compute Variance ***;
VAR=STDERR*STDERR;
RUN;
*** Between and Within Imputation Calculations ***;
PROC SUMMARY DATA=WORK.POOLED NWAY MISSING;
CLASS ESTNO EFFECT &FIXED LABEL;
VAR ESTIMATE VAR;
OUTPUT OUT=WORK.SUMSTATS MEAN=THETA WITHIN VAR=BETWEEN DUMMY;
RUN;
*** Compute pooled estimates and statistics ***;
DATA &SUMDATA;
SET WORK.SUMSTATS(DROP=_TYPE_ DUMMY RENAME=(_FREQ_=M));
TOTAL=WITHIN+(1+(1/M))*BETWEEN;
TOT_SE=SQRT(TOTAL);
c2638 2638TexturesC07 February 7, 2002 12:21
150 MULTIPLE IMPUTATION
RATIO=((1+(1/M))*BETWEEN)/WITHIN;
GAMMA=((1+(1/M))*BETWEEN)/TOTAL;
T_STAT=THETA/TOT_SE;
DF=(M-1)*(1+(1/RATIO))**2;
P_VALUE=(1-PROBT(ABS(T_STAT),DF))*2;
LABEL M=# Imputations
THETA=Estimate
WITHIN=Within- Imputation Var
BETWEEN=Between- Imputation Var
TOTAL=Total Variance
TOT_SE=Total S.E.
RATIO=Between-to-within ratio
GAMMA=Fraction Missing Information
T_STAT=T Statistic (Ho: EST=0)
P_VALUE=P (2-sided);
FORMAT THETA WITHIN BETWEEN TOTAL TOT_SE RATIO
GAMMA T_STAT P_VALUE 7.4 DF 6.;
RUN;
*** Print results ***;
PROC PRINT DATA=&SUMDATA LABEL;
TITLE5 MULTIPLE-IMPUTATION PARAMETER SUMMARY;
ID ESTNO;
VAR EFFECT &FIXED LABEL M;
RUN;
PROC PRINT DATA=&SUMDATA LABEL;
TITLE5 MULTIPLE-IMPUTATION VARIANCE INFORMATION;
ID ESTNO;
VAR BETWEEN WITHIN TOTAL RATIO GAMMA;
RUN;
PROC PRINT DATA=&SUMDATA LABEL;
TITLE5 MULTIPLE-IMPUTATION PARAMETER ESTIMATES;
ID ESTNO;
VAR THETA TOT_SE T_STAT P_VALUE;
RUN;
TITLE5 ;
%MEND MI_SUM;
7.8 Sensitivity analyses
Many of the articles describing multiple imputation incorporate sensitivity
analyses as part of the presentation. These sensitivity analyses include both
comparisons of multiple imputation with simple methods [19, 21, 84] and
explorations of the eects of assumptions that have been made in the process
of imputing the missing observations [63, 94, 153]. There are several outcomes
of these sensitivity analyses. The analyses may demonstrate that the estimates
and inference are insensitive to the results. Or, if the results do depend on
c2638 2638TexturesC07 February 7, 2002 12:21
IMPUTATION VS. ANALYTIC MODELS 151
the model used for imputation, precise statements can be made about the
conditions under which the conclusions are valid.
7.9 Imputation vs. analytic models
If the multiple imputation models use only the same data that is used in the
analytic models (the observed response data and baseline covariates), the re-
sulting estimates are almost identical to those obtained using the maximum
likelihood method with the same information. This is illustrated in Table 7.5,
where the estimates obtained using a repeated measures analysis of incomplete
data are almost identical to those obtained after multiple imputation using
the four FACT-Lung TOI scores and treatment group (TAXOL). In contrast,
when additional information about the patient, which is potentially related to
his or her HRQoL when the observation is missing, is added to the imputa-
tion procedure, the resulting estimates may be dramatically dierent. This is
illustrated in Table 7.5, where the length of survival and number of cycles of
therapy started were included in the imputation procedure and the estimated
rate of decline in the FACT-Lung TOI scores more than doubled. Reiterating
the points: (1) multiple imputation is the hard way to analyze data where
missingness is MAR and (2) it will only provide a benet when the analyst
has additional information (data) that is related to HRQoL both when the
response is observed and when it is missing.
Table 7.5. Comparison of estimates obtained using maximum likelihood estimation
of all available FACT-Lung TOI data (MLE), multiple imputation using all available
data (MI Obs), and multiple imputation with additional outcome data (MI Plus).
MLE Obs
a
MI Obs
b
MI Plus
c
Parameter Est (S.E.) Est (S.E.) Est (S.E.)
Mean 0 weeks 65.69 (0.70) 65.70 (0.70) 65.74 (0.69)
Mean 6 weeks 63.63 (0.79) 63.65 (0.78) 61.72 (0.85)
Mean 12 weeks 62.40 (0.91) 62.44 (0.94) 58.92 (1.16)
Mean 6 months 60.84 (1.06) 60.83 (1.11) 53.25 (1.29)
Change 6 weeks 2.05 (0.77) 2.05 (0.85) 4.02 (0.78)
Change 12 weeks 3.29 (0.94) 3.26 (0.98) 6.82 (1.16)
Change 6 months 4.85 (1.09) 4.87 (1.14) 12.49 (1.23)
a
MLE Obs: REML with repeated measures design of all available data (SAS Proc
MIXED).
b
MI Obs: Multiple imputation using MCMC method limited to the four repeated
measures of the FACT-Lung TOI, M = 20 (SAS Proc MI).
c
MI Plus: Multiple imputation using MCMC method including the four repeated
measures of the FACT-Lung TOI plus log transformation of survival and number of
cycles of therapy started, M = 20 (SAS Proc MI).
Includes 548 patients with at least one observation and no missing covariate data.
c2638 2638TexturesC07 February 7, 2002 12:21
152 MULTIPLE IMPUTATION
7.10 Implications for design
If multiple imputation is proposed as an analysis strategy for a clinical trial, a
number of factors should be considered. As mentioned above, these procedures
all require large data sets. Initially, multiple imputation was developed for
large surveys. Application of multiple imputation to smaller clinical trials
should be proposed cautiously. How large is necessary will depend on how
much auxiliary information (strong covariates) is available and how much
missing data is expected.
Anticipating the use of multiple imputation, it is wise to ensure the col-
lection of auxiliary information that is required. When, as a result of careful
planning of a study, there is adequate documentation of the reasons for miss-
ing data along with information that will link these reasons to changes in
HRQoL, it is possible to develop a good imputation model. For example, a
possible strategy is to use surrogate assessments of HRQoL or related mea-
sures obtained from the health-care providers or caregivers. These measures
can be obtained both concurrently with the patients assessments of HRQoL
and also when the patients assessments are missing. Although there has been
considerable documentation of the discrepancies between the patients self-
evaluations of HRQoL and those of others, these assessments may be stronger
predictors than other measures. This strategy is particularly useful in settings
where subjects are expected to be unable to provide their own assessment of
HRQoL over the course of the study, for example, patients with brain tumors
or Alzheimers disease.
7.11 Summary
Multiple imputation provides a exible way of handling missing data from
multiple unrelated causes or when the mechanism changes over time. For
example, separate models can be used for early and late dropout.
Multiple imputation will provide a benet only when the analyst has addi-
tional information (data) that is related to HRQoL both when the response
is observed and missing.
Multiple imputation will be dicult to implement in studies with mistimed
observations and where the sample size is small.
Sequential univariate regression will be useful when the missing data pat-
terns are predominately monotone and the variables explaining dropout
are very dierent over time.
MCMC will be useful when the missing data pattern is not strictly mono-
tone and the same set of variables explain dropout over time.
ABB is not recommended for longitudinal studies, as it underrepresents the
correlation of observations over time. It also makes stronger assumptions
about the missing data mechanism than the explicit procedures (MCMC
and regression).
c2638 C2638texC08 February 12, 2002 11:38
CHAPTER 8
Pattern Mixture Models
8.1 Introduction
Non-ignorable missing data are the common type of missing data in HRQoL
studies where there is dropout as a result of toxicity, disease progression, or
even therapeutic eectiveness. Studies with this type of missing data are also
the most dicult to analyze. The primary reasons are (1) there are a large
number of possible models and (2) it is impossible to verify statistically the
correctness of any model because the data required to distinguish between
models are missing. In this chapter, we describe one group of models, pattern
mixture models, that has been proposed for this problem. In pattern mix-
ture models, the portion of the model specifying the missing data mechanism
(f[R|X, ]) does not depend on the missing values. Thus, for pattern mixture
models, we only need to know the proportion of subjects with each pattern of
missing data and we do not need to specify how missingness depends on Y
mis
i
.
This advantage is balanced by (1) the large number of potential patterns
of missing data and (2) the diculties of estimating all parameters in each
pattern.
The basic concept [9092] is that there are P dierent missing data patterns
(R
i
M
{p}
). The distribution of the responses, Y
i
, may dier across the P
patterns with dierent parameters,
{p}
, and variance,
{p}
.
Y
i
|M
{p}
N
_
X
i

{p}
,
{p}
i
_
, p = 1, . . . , P.
For example, change in HRQoL among subjects in each pattern is described
using a dierent intercept (
{p}
0
) and slope (
{p}
1
). This would allow patients
who drop out earlier to have lower HRQoL scores initially and to decline more
rapidly over time. The patients who drop out earlier also may have more or
less variability in their scores (dierent variance) than patients who remain
in the study.
The true distribution of the measures of HRQoL for the entire group of
patients is a mixture of the distributions from each of the P groups of patients.
The quantities of interest are the expected values of the parameters averaged
over the missing data patterns:
E[] =
P

{p}

{p}
,
where
{p}
is the proportion of subjects observed with the pth pattern.
The general method is to stratify the patients by the missing data pattern.
Then, the parameters (
{p}
,
{p}
) are estimated within each of the strata.
c2638 C2638texC08 February 12, 2002 11:38
154 PATTERN MIXTURE MODELS
Weights are determined by the number of subjects within each of the P missing
data patterns (
{p}
= n
{p}
/N). The population estimates are the weighted
average of the estimates from the P missing data patterns (

=


{p}

{p}
).
The procedure sounds easy and straightforward, but there are some practi-
cal problems. The rst diculty is the large number of missing data patterns
that occur where there may be only a few subjects with some patterns. In a
study with J assessments over time, there are theoretically 2
J
possible pat-
terns. Thus, with only four assessments, there are 16 possible patterns.
The second diculty is that, for most of the patterns, the model (X
i

{p}
)
is underidentied. This means that all the parameters cannot be estimated in
each pattern unless additional assumptions are made. Daniels and Hogan [26]
cite three approaches to deal with the lack of information. Some options for
solving these problems are presented in the following sections.
1. Impose explicit restrictions such as complete case restrictions (see Sections
8.2 and 8.3) or missing data assumptions (MAR).
2. Impose model restrictions such as selection model restrictions that assume
a nonignorable mechanism (see Section 8.2).
3. Carry out a sensitivity analysis across a range of choices (see Section 8.2).
Option 1 is not viable when we believe the missing data are nonignorable.
Other options apply only to a few limited situations. Further, the assump-
tions that allow us to estimate all of the parameters are often dicult to
communicate and cannot be tested because they depend on the values of the
missing data. This chapter focuses on the various restrictions and assumptions
required to estimate parameters in these underidentied models. Sections 8.2
and 8.3 describe restrictions for bivariate data and monotone dropout, re-
spectively. In Section 8.4, the use of parametric models to extrapolate growth
curves is described.
NSCLC example
The existence of a large number of patterns is clearly illustrated in the NSCLC
study. Of the 16 possible patterns, 15 were observed (see Table 4.2). Of the
15 patterns, 10 have less then 5% of the observations. If we simplify the
problem by dening the patterns by the time of dropout, we still have ve
patterns. Average HRQoL scores are displayed for four of the patterns with
any observations in Figure 8.1 (one pattern has no observations).
As mentioned before, the second diculty is that, for most of the patterns,
the unknown parameters of the model ( and
i
) are underidentied. If we
were using a cell means model in the NSCLC example, we would need to es-
timate a mean at each time in each pattern (
{p}
1
,
{p}
2
,
{p}
3
,
{p}
4
) and the
ten covariance parameters (
{p}
11
,
{p}
12
, . . . ,
{p}
44
). Without making additional
assumptions, we could do that in only the one pattern with all four assess-
ments. Even if we were considering a simple linear growth curve, we could
only estimate both the intercept and slope in the patterns where there are at
least two observations.
c2638 C2638texC08 February 12, 2002 11:38
BIVARIATE DATA (TWO REPEATED MEASURES) 155
Table 8.1. All possible missing data patterns for two repeated measures.
Pattern Y
1
Y
2
R
1 Y
{1}
1
Y
{1}
2
(1,1)
2 Y
{2}
1
? (1,0)
3 ? Y
{3}
2
(0,1)
4 ? ? (0,0)
Figure 8.1. Patterns dened by timing of last assessment.
8.2 Bivariate data (two repeated measures)
The simplest case for longitudinal studies consists of two repeated measures.
There are four possible patterns of missing data (Table 8.1). The rst pattern,
in which all responses are observed, contains the complete cases. The second
and third patterns have one observation each. In the fourth pattern, none of
the responses is observed.
In each of the four patterns, there are ve possible parameters to be esti-
mated: two means (
{p}
1
,
{p}
2
) and three parameters for the covariance
(
{p}
11
,
{p}
12
,
{p}
22
). We can estimate 9 of the 20 total parameters from the data:
all ve parameters from pattern 1 (
{1}
1
,
{1}
2
,
{1}
11
,
{1}
12
,
{1}
22
) and two each
from patterns 2 and 3 (
{2}
1
,
{2}
11
and
{3}
2
,
{3}
22
). Thus, the model is under-
identied and some type of restriction (assumption) is required to estimate
the remaining parameters.
c2638 C2638texC08 February 12, 2002 11:38
156 PATTERN MIXTURE MODELS
Table 8.2. Number of subjects with each missing data pattern for the initial and
12-week assessments in the NSCLC study.
Pattern Y
1
Y
2
No Taxol Taxol
1 Y
{1}
1
Y
{1}
2
83 190
2 Y
{2}
1
? 94 159
3 ? Y
{3}
2
2 7
4 ? ? 15 25
Table 8.3. Identied parameters in each pattern. Covariance parameters were re-
stricted to be equal across pattern but allowed to dier across treatment groups.
Means Covariance
Pattern
{p}

{p}
1

{p}
2

{p}
11

{p}
12

{p}
22
No Taxol
1 0.43 68.4 64.1 249.5 118.3 320.7
2 0.48 61.4 ? 249.5 ? ?
3 0.01 ? 84.2 ? ? 320.7
4 0.08 ? ? ? ? ?
Taxol
1 0.50 69.9 65.2 245.7 101.2 242.0
2 0.42 61.4 ? 245.7 ? ?
3 0.02 ? 75.0 ? ? 242.0
4 0.07 ? ? ? ? ?
NSCLC example
To illustrate the problem of identifying all parameters in the model, consider
the initial and 12-week assessments in NSCLC study. The majority of subjects
are in either the rst or second pattern (Table 8.2). The estimable parameters
for these rst two patterns are summarized in Table 8.3.
Complete-case missing variable (CCMV) restriction
One possible set of restrictions is based on the complete cases [90]. The
assumption is that the missing value distributions are equal to the complete
case distributions. For the bivariate case shown in Table 8.1, the restrictions
are

{2}
[21]
=
{1}
[21]
,
{3}
[12]
=
{1}
[12]
, and
{4}
=
{1}
, (8.1)
where
{1}
[21]
denotes the parameters from the regression of Y
2
on Y
1
(or the
conditional distribution of Y
2
given Y
1
) using the complete cases in pattern 1.
c2638 C2638texC08 February 12, 2002 11:38
BIVARIATE DATA (TWO REPEATED MEASURES) 157
For a moment, consider an example where we have only the rst two pat-
terns. All subjects were observed initially, but some are missing at the second
assessment. We estimate
{1}
1
,
{1}
2
, and
{2}
1
directly from the data.

{1}
1
=

Y
{1}
1
=
1
n
{1}

Y
{1}
i1
. (8.2)

{1}
2
=

Y
{1}
2
=
1
n
{1}

Y
{1}
i2
. (8.3)

{2}
1
=

Y
{2}
1
=
1
n
{2}

Y
{2}
i1
. (8.4)
We then apply this complete-case missing variable restriction to estimate
{2}
2
.
Consider the regression of Y
2
on Y
1
:
Y
{1}
i2
=
{1}
0[21]
+
{1}
1[21]
Y
{1}
i1
+ e
i
, (8.5)
Y
{2}
i2
=
{2}
0[21]
+
{2}
1[21]
Y
{2}
i1
+ e
i
. (8.6)
Because we cannot estimate the parameters in the second equation due to
missing data, we assume the same relationship holds in pattern 2 as in pat-
tern 1.

{2}
0[21]
=
{1}
0[21]
, (8.7)

{2}
1[21]
=
{1}
1[21]
. (8.8)
The procedure is to estimate the intercept
{1}
0[21]
and slope
{1}
1[21]
using the
complete cases in pattern 1; then, applying the restrictions, the estimated
mean for the second assessment in pattern 2 is

{2}
2
=

{1}
0[21]
+

{1}
1[21]

Y
{2}
1
=

Y
{1}
2


{1}
12

{1}
11
_

Y
{1}
1


Y
{2}
1
_
. (8.9)
The algebraic details for Equation 8.9 are at the end of this chapter. Combin-
ing the estimates for the two patterns, the overall estimates of the two means
are

1
=
{1}

{1}
1
+
{2}

{2}
1
=
{1}

Y
{1}
1
+
{2}

Y
{2}
1
=

Y
1
, (8.10)

2
=
{1}

{1}
2
+
{2}

{2}
2
=

Y
{1}
2
+

{1}
12

{1}
11
_

1

Y
{1}
1
_
. (8.11)
The algebraic details for Equation 8.11 are at the end of this chapter. Pro-
cedures for estimating the covariance parameters are outlined in detail by
Little [90, 91].
Note that for this special case these estimates are the same as those ob-
tained using MLE with all available data (see Chapter 5, Equation 5.4). In the
example described above, the restrictions are equivalent to assuming that the
missingness of Y
2
depends only on Y
1
. Recall that this is exactly the denition
c2638 C2638texC08 February 12, 2002 11:38
158 PATTERN MIXTURE MODELS
Table 8.4. Baseline and 12-week FACT-Lung TOI means estimated under the
CCMV restrictions. Example limited to subjects with baseline assessments.
No Taxol Taxol
Pattern
{p}

{p}
1

{p}
2

{p}

{p}
1

{p}
2
1 0.47 68.4 64.1 0.54 69.9 65.2
2 0.53 61.4 60.8 0.46 61.4 61.6
Pooled 64.7 62.3 66.0 63.6
of MAR. Thus, this restriction is helpful when trying to understand pattern
mixture models, but it is not of much practical use when we believe that the
data are MNAR.
NSCLC example
To illustrate the CCMV restriction, consider the baseline and 12-week as-
sessments for subjects who had a baseline assessment. (This limitation is
not required under the CCMV restriction, but it will simplify the illustra-
tion and allow comparison with the restriction presented in the next section.)
The estimated means at 12 weeks for the subjects in pattern 2 follow from
Equation 8.9 (Table 8.4). Notice that the 12-week estimates for pattern 2 are
lower than those for pattern 1 (60.8 vs. 64.1 for patients not receiving Taxol,
and 61.6 vs. 65.2 for patients receiving Taxol). This is because the baseline
observations are lower in pattern 2 than in pattern 1 and the two measures
are positively correlated.
Browns protective restrictions
A second set of restrictions was proposed by Brown [10] for monotone dropout
where patterns 1 and 2 are observed. The assumption is that the missingness
of Y
2
depends on Y
2
. This is the MNAR assumption. For the bivariate case, the
restriction is
{2}
[12]
=
{1}
[12]
. (Note that this is the mirror image of the complete-
case restriction.)
Again, consider the bivariate example where we have only the rst two pat-
terns. We need to estimate the second mean in the second pattern,
{2}
2
. This
time, we will regress Y
1
on Y
2
.
Y
{1}
i1
=
{1}
0[12]
+
{1}
1[12]
Y
{1}
i2
+ e
i
, (8.12)
Y
{2}
i1
=
{1}
0[12]
+
{1}
1[12]
Y
{2}
i2
+ e
i
. (8.13)
The assumption is that the same relationship holds in pattern 2 as in pattern 1:

{2}
0[12]
=
{1}
0[12]
, (8.14)
c2638 C2638texC08 February 12, 2002 11:38
BIVARIATE DATA (TWO REPEATED MEASURES) 159

{2}
1[12]
=
{1}
1[12]
. (8.15)
The procedure is to estimate the intercept
{1}
0[12]
and slope
{1}
1[12]
using the
complete cases in pattern1. To estimate
{2}
2
, we assume

Y
{2}
1
=
{1}
0[12]
+
{1}
1[12]

{2}
2
(8.16)
and then solve for
{2}
2
:

{2}
2
=
_

Y
{2}
1

{1}
0[12]
__

{1}
1[12]
=

Y
{1}
2


{1}
22

{1}
12
_

Y
{1}
1


Y
{2}
1
_
. (8.17)
Combining the estimates for the two patterns, the overall estimates of the two
means are

1
=
{1}

{1}
1
+
{2}

{2}
1
=
{1}

Y
{1}
1
+
{2}

Y
{2}
1
=

Y
1
, (8.18)

2
=
{1}

{1}
2
+
{2}

{2}
2
=

Y
{1}
2
+

{1}
22

{1}
12
_

1

Y
{1}
1
_
. (8.19)
The algebraic details for Equations 8.17 and 8.19 are at the end of the chapter.
NSCLC example
To illustrate, again consider the subjects in patterns 1 and 2. Applying Equa-
tion 8.17, we compute the 12-week means in the second pattern (
{2}
2
) to
be 44.9 and 45.0 in the two treatment arms. Comparing these estimates of

{2}
2
with those made under the CCMV restrictions (see Table 8.4.), we see
that they are approximately 20 points lower (Table 8.5). With roughly half
of the cases in each pattern, the pooled estimates of the 12-week means are
about 10 points lower under Browns protective restrictions than those with
the CCMV restrictions. The discrepancy between the estimates with the two
dierent sets of restrictions is not surprising, as the rst set of estimates was
made under an assumption that the data were MAR and the later under an as-
sumption that the data were MNAR. Three factors inuence the magnitude of
Table 8.5. Baseline and 12-week FACT-Lung TOI means estimated under Browns
protective restrictions. Example limited to subjects with baseline assessments.
No Taxol Taxol
Pattern
{p}

{p}
1

{p}
2

{p}

{p}
1

{p}
2
1 0.47 68.4 64.1 0.54 69.9 65.2
2 0.53 61.4 44.9 0.46 61.4 45.0
Pooled 64.7 53.9 66.0 56.0
c2638 C2638texC08 February 12, 2002 11:38
160 PATTERN MIXTURE MODELS
the discrepancy. The dierences between the estimates will increase as (1) the
proportion of missing data increases, (2) the magnitude of the dierence
between the baseline means (

Y
{1}
and

Y
{2}
) increases, and (3) the corre-
lation between the baseline and 12-week scores decreases.* In this example,
approximately 50% of the subjects are missing an assessment at 12 weeks,
there is roughly a 7-point dierence between the baseline means (0.5 S.D.),
and the correlation between the baseline and 12-week score is 0.4.
Sensitivity analyses with intermediate restrictions
Although both approaches use the complete cases to dene the restrictions,
the rst approach assumes that dropout is completely dependent on the initial
measure of the outcome, Y
1
, and the second approach assumes that dropout
is completely dependent on the nal measure of the outcome, Y
2
. Reality is
likely to be somewhere in between. Little [91] proposed a mixture of these two
approaches in which missingness depends on Y
1
+ Y
2
.

2
=

Y
{1}
2
+b
()
_

1

Y
{1}
1
_
, b
()
=

{1}
22
+
{1}
12

{1}
12
+
{1}
11
. (8.20)
If = 0, this is equivalent to the CCMV restriction.
b
()
=

{1}
22
+
{1}
12

{1}
12
+
{1}
11
=

{1}
12

{1}
11
. (8.21)
As , by LH opitals rule, we have
lim

b
()
= lim

{1}
22
+
{1}
12

{1}
12
+
{1}
11
=

{1}
22

{1}
21
. (8.22)
This is equivalent to Browns protective restriction.
Unfortunately, we are unable to estimate . We can, however, use a range of
values of to perform a sensitivity analysis. Sensitivity to will decrease as
the correlation between Y
1
and Y
2
increases. For example, when the variance
of Y is constant over time (
{1}
11

{1}
22
) and the correlation approaches 1,
b
()
will vary only slightly from a value of 1 regardless of the value of .
Large-sample inferences for
2
The large-sample variance of
2
is derived using Taylor series approxima-
tions [91]:
var(
2
) =

22
N
+
_

Y
1
{1}
_
2
V
b
+
n
2
n
1
N
_

{1}
22
2b
()

{1}
12
+ b
()2

{1}
11

, (8.23)
* Note that the quantity
1

Y
{1}
1
in Equation 8.19 is a function of both
{2}
and

Y
{1}

Y
{2}
(see Section 8.6).
c2638 C2638texC08 February 12, 2002 11:38
MONOTONE DROPOUT 161
Table 8.6. Sensitivity analysis.
No Taxol Taxol Dierence

11

12

1

21

22

2

22

12
(S.E.)
0 64.7 62.3 2.3 66.0 63.6 2.4 1.31 (2.22)
1 64.7 59.6 5.1 66.0 61.4 4.6 1.79 (2.52)
2 64.7 58.2 6.5 66.0 60.2 5.8 1.96 (2.89)
4 64.7 56.8 7.9 66.0 58.9 7.1 2.07 (3.40)
64.7 53.9 10.8 66.0 56.0 10.0 2.40 (4.82)
Example limited to NSCLC subjects with baseline assessments.
hk
is the
mean at the kth assessment for the hth treatment group.
h
is the change
between the initial and 12-week assessments for the hth treatment group.
where
{1}
11
,
{1}
12
,
{1}
22
are the parameters of the covariance of Y
1
and Y
2
esti-
mated from the complete cases in pattern 1,

11
=
1
N

(Y
i1

1
)
2
,

22
=
{1}
22
+ b
()2
_

11

{1}
11
_
,
V
b
= var(b
()
) =
_

{1}
11

{1}
22

{1}2
12
__

2

{1}
22
+ 2
{1}
12
+
{1}
11
_
2
n
1
_

{1}
12
+
{1}
11
_
4
.
NSCLC example
Application of this type of sensitivity analysis is illustrated in Table 8.6. Al-
though the estimates of the 12-week means (
h2
) and of the change over time
(
h2

h1
) are very sensitive to the value of , the dierences between the
two treatments are very insensitive. Thus, one would feel much more con-
dent making inferences about treatment dierences at 12 weeks than about
the presence or absence of change over time within each group. However, with
approximately 50% of the data missing, all conclusions about the dierences in
the FACT-Lung TOI should be made cautiously. It is also reassuring that the
estimated variance of the parameters increases as the dependence shifts from
the observed data (Y
i1
) to the missing data (Y
i2
). For the estimated group dif-
ferences at 12 weeks, the standard error more than doubles from 2.22 to 4.82.
8.3 Monotone dropout
Next we extend the possibilities for missing data patterns to longitudinal
studies with monotone dropout. Three sets of restrictions have been proposed
for monotone missing data patterns. The rst is an extension of Littles CCMV
restriction described in the previous section. The second is the available case
missing value (ACMV) restriction. The third is the neighboring case missing
value (NCMV) restriction.
c2638 C2638texC08 February 12, 2002 11:38
162 PATTERN MIXTURE MODELS
Table 8.7. NSCLC patients with monotone dropout patterns by treatment.
Observation No Taxol Taxol
Pattern 1 2 3 4 N % N %
1 X X X X 56 30.6 131 36.1
2 X X X 42 23.0 80 22.0
3 X X 30 16.4 75 20.7
4 X 55 30.0 77 21.2
NSCLC study
In the NSCLC study, patients with one of the four patterns of data conform-
ing exactly to a monotone dropout pattern account for 82% of the patients. In
practice, one is reluctant to omit 18% of the subjects from the analysis. How-
ever, for the purpose of illustration, we will use only the patients with a mono-
tone dropout pattern. Note that each pattern has approximately one fourth of
the patients, so that no pattern contains fewer than 30 subjects (Table 8.7).
Complete-case missing value restriction
In the CCMV restriction, the data from the subjects in pattern 1 are used to
impute the means for the missing observations in the remaining patterns.

{2}
[4123]
=
{1}
[4123]
, (8.24)

{3}
[3412]
=
{1}
[3412]
, (8.25)

{4}
[2341]
=
{1}
[2341]
, (8.26)
where
{1}
[3412]
denotes the parameters from the regression of Y
3
on Y
1
and Y
2
and the regression of Y
4
on Y
1
and Y
2
using the complete cases in pattern 1.
As a practical note, this restriction is feasible only when the number of cases
in pattern 1 is sucient to estimate these parameters reliably.
The CCMV restrictions result in six equations that must be solved for
the unknown means and variance parameters. Although this is burdensome,
solving the equations for the unknown parameters is straightforward. However,
deriving the appropriate variance for the pooled estimates is very complex.
Curren [23] suggests an analytic technique using multiple imputation to avoid
this problem. The procedure is as follows:
1. Pattern 2 (Equation 8.24): Impute missing values at T4 using cases in
pattern 1.
2. Pattern 3 (Equation 8.25): Impute missing values at T3 and T4 using cases
in pattern 1.
3. Pattern 4 (Equation 8.26): Impute missing values at T2, T3, and T4 using
cases in pattern 1.
c2638 C2638texC08 February 12, 2002 11:38
MONOTONE DROPOUT 163
Figure 8.2. Observed and imputed means under CCMV restrictions displayed by
pattern of dropout. Observed means indicated by solid line. Imputed means indi-
cated by dashed lines.
4. Analyze each set of imputed data and combine estimates as previously
described in Chapter 7.
The resulting estimates for the NSCLC patients are displayed in Figure 8.2.
Note that the imputed values under the CCMV restriction tend to increase
initially after the last observed FACT-Lung TOI score, especially for patients
who had only the initial assessments. This illustrates the consequences of the
implicit assumption with the CCMV restriction. Obviously, in the setting of
the NSCLC study, we do not believe that the HRQoL of the individuals who
c2638 C2638texC08 February 12, 2002 11:38
164 PATTERN MIXTURE MODELS
drop out early is likely to be similar to that of individuals who have all four
assessments.
Available case missing value restriction
In the ACMV restriction, available data from subjects in all the patterns
are used to impute the means for the missing observations in the remaining
patterns. The restrictions for the patterns in Table 8.7 are

{2}
[4123]
=
{1}
[4123]
, (8.27)

{3}
[312]
=
{1,2}
[312]
,
{3}
[4123]
=
{1}
[4123]
, (8.28)

{4}
[21]
=
{1,2,3}
[21]
,
{4}
[312]
=
{1,2}
[312]
,
{4}
[4123]
=
{1}
[4123]
, (8.29)
where
{1,2}
[312]
denotes the parameters from the regression of Y
3
on Y
1
and Y
2
using the cases in patterns 1 and 2. As a practical note, this restriction is a
bit more feasible than the CCMV restriction, as more observations are used
to estimate some of these parameters.
Using the same multiple imputation approach, the procedure is as follows:
1. Pattern 2 (Equation 8.27): Impute missing values at T4 using cases in
pattern 1.
2. Pattern 3 (Equation 8.28): Impute missing values at T3 using cases in
patterns 1 and 2. Impute missing values at T4 using cases in pattern 1 and
the imputed values at T3 in pattern 3.
3. Pattern 4 (Equation 8.29): Impute missing values at T2 using cases in
patterns 1, 2, and 3. Impute missing values at T3 using cases in patterns 1
and 2 and the imputed values at T2 in pattern 4. Impute missing values
at T4 using cases in pattern 1 and the imputed values at T2 and T3 in
pattern 4.
4. Analyze each set of imputed data and combine estimates as previously
described in Chapter 7.
The resulting estimates for the NSCLC patients are displayed in Figure 8.3.
Note that the imputed values under the ACMV restriction no longer tend to
increase but rather tend to remain at the same level as the last observed
FACT-Lung TOI measure. This is a slight improvement over the CCMV but
still seems to overestimate the HRQoL of patients who drop out.
Neighboring case missing value restriction
In the NCMV restriction, available data from subjects in the neighboring
pattern are used to impute the means for the missing observations.

{2}
[4123]
=
{1}
[4123]
, (8.30)

{3}
[312]
=
{2}
[312]
,
{4}
[21]
=
{3}
[21]
, (8.31)

{4}
[4123]
=
{1}
[4123]
,
{4}
[312]
=
{2}
[312]
,
{3}
[4123]
=
{1}
[4123]
. (8.32)
c2638 C2638texC08 February 12, 2002 11:38
MONOTONE DROPOUT 165
Figure 8.3. Observed and imputed means under ACMV restrictions displayed by
pattern of dropout. Observed means indicated by solid line. Imputed means indi-
cated by dashed lines.
Using the same multiple imputation approach, the procedure is as follows:
1. Pattern 2 (Equation 8.30): Impute missing values at T4 using cases in
pattern 1.
2. Pattern 3 (Equation 8.31): Impute missing values at T3 using cases in
pattern 2. Impute missing values at T4 using cases in pattern 1 and the
imputed values at T3 in pattern 3.
3. Pattern 4 (Equation 8.32): Impute missing values at T2 using cases in
pattern 3. Impute missing values at T3 using cases in pattern 2 and the
c2638 C2638texC08 February 12, 2002 11:38
166 PATTERN MIXTURE MODELS
Figure 8.4. Observed and imputed means under NCMV restrictions displayed by
pattern of dropout. Observed means indicated by solid line. Imputed means indi-
cated by dashed lines.
imputed values at T2 in pattern 4. Impute missing values at T4 using cases
in patterns 1 and the imputed values at T2 and T3 in pattern 4.
4. Analyze each set of imputed data and combine estimates as previously
described in Chapter 7.
The resulting estimates for the NSCLC patients are displayed in Figure 8.4.
Note that the imputed values under the NCMV restriction tend to fall initially
c2638 C2638texC08 February 12, 2002 11:38
PARAMETRIC MODELS 167
Table 8.8. Sensitivity of estimates of change between the initial and 12-week as-
sessments for the hth treatment group with CCMV, ACMV, NCMV, and REML of
available data.
No Taxol Taxol Dierence
Est (S.E.) Est (S.E.) Est (S.E.)
CCMV 1.7 (2.7) 1.5 (1.1) 0.2 (3.0)
ACMV 3.9 (1.8) 2.9 (1.1) 1.0 (2.0)
NCMV 7.3 (2.3) 4.6 (1.3) 2.7 (2.4)
MLE 2.9 (1.8) 2.9 (1.2) 0.2 (2.1)
over time, especially for subjects who dropped out early. This would seem to
be the most appropriate of the three restrictions for the NSCLC study, but it
may still overestimate the HRQoL of subjects who drop out of the study.
Comparison of CCMV, ACMV, and NCMV estimates
Figure 8.5 and Table 8.8 displays the overall estimates for the three restric-
tions. Table 8.9 summarizes the three sets of restrictions used to generate
estimates. Notice the sensitivity of the estimates to the assumptions, particu-
larly the decreasing estimates for the 12-week scores. The upward trend in the
patients not receiving Taxol is clinically counterintuitive but not surprising,
as these estimates rely on the 26-week data from patients who completed all
assessments.
8.4 Parametric models
Another approach is to t growth curve models to the data in each pattern,
extrapolating past the point of dropout [23, 66]. Polynomial models can be
t within each of the P patterns where there are sucient observations over
time to estimate the parameters.
Y
{p}
hij
=
{p}
X
{p}
hij
+
{p}
hij
. (8.33)
Note that the xed eects and the covariance parameters may dier across
the patterns. The pooled parameter estimates are

k
=
P

{p}

{p}
. (8.34)
In practice, it may be necessary to pool some of the patterns with a small num-
ber of subjects. It may also be necessary to make some additional restrictions
to estimate all the parameters.
c2638 C2638texC08 February 12, 2002 11:38
168 PATTERN MIXTURE MODELS
Figure 8.5. Estimated means under CCMV, ACMV, and NCMV restrictions. Cur-
ves, from highest to lowest, correspond to CCMV, ACMV, and NCMV.
Linear trends
In a limited number of studies, it is reasonable to believe that the change in
HRQoL is linear over time (t).
Y
{p}
hij
=
{p}
0
+
{p}
1
t
{p}
hij
+
{p}
hij
. (8.35)
However, even with this simple model, we can immediately see the challenges.
The two xed-eects parameters (intercept and slope) can be estimated only
in the patterns with at least two observations per subject. Thus, we need to
c2638 C2638texC08 February 12, 2002 11:38
PARAMETRIC MODELS 169
Table 8.9. Contrast of CCMV, ACMV, and NCMV restrictions.
Pattern Week 0 (Y
1
) Week 6 (Y
2
) Week 12 (Y
3
) Week 26 (Y
4
)
CCMV
1
2

{1}
[4123]
Y
1,2,3
3

{1}
[312]
Y
1,2

{1}
[412]
Y
1,2
4

{1}
[21]
Y
1

{1}
[31]
Y
1

{1}
[41]
Y
1
ACMV
1
2

{1}
[4123]
Y
1,2,3
3

(1,2)
[312]
Y
1,2

{1}
[4123]
Y
1,2
, Y
(m)
3
4

(1,2,3)
[21]
Y
1

(1,2)
[312]
Y
1
, Y
(m)
2

{1}
[4123]
Y
1
, Y
(m)
2,3
NCMV
1
2

{1}
[4123]
Y
1,2,3
3

{2}
[312]
Y
1,2

{1}
[4123]
Y
1,2
, Y
(m)
3
4

{3}
[21]
Y
1

{2}
[312]
Y
1
, Y
(m)
2

{1}
[4123]
Y
1
, Y
(m)
2,3
For example,

{1}
[4123]
Y
1
, Y
(m)
2,3
in the last column of the fourth row of the AMCV or
NCMV sections indicates that data from pattern 1 will be used to estimate the regression
parameters and the observed data (Y
1
) and the imputed data (Y
(m)
2,3
) from pattern 4 will
be used to estimate Y
4
in pattern 4.
impose an additional restriction to estimate the parameters for the remaining
patterns. Consider the four dropout patterns observed in the NSCLC study
after collapsing the 15 observed patterns by time of dropout (see Table 8.7).
The slope is not estimable in pattern 4 without additional assumptions. One
reasonable restriction is to assume that the slope for subjects in pattern 4 is
the same as the slope for subjects in pattern 3:

{4}
1
=

{3}
1
. (8.36)
The plots in the upper half of Figure 8.6 illustrate this for the NSCLC
example.
An alternative approach is to place parametric assumptions on the parame-
ters using the time of dropout as a covariate [23, 100]. For example, one might
assume that there was an interaction between the time of dropout (d
{p}
) and
c2638 C2638texC08 February 12, 2002 11:38
170 PATTERN MIXTURE MODELS
Figure 8.6. Extrapolation of linear trends in patterns dened by the time of the
last observation. Dashed lines in upper plots are based on Equations 8.35 and 8.36.
Dashed lines in the lower plots are based on Equation 8.37. Solid lines are based on
MLE of available data using a piecewise linear regression model.
the intercept and slope in each pattern. With both linear and quadratic terms
for the time of dropout, the model might appear as
Y
{p}
hij
=
{p}
0
+
{p}
1
d
{p}
+
{p}
2
d
{p}2
. .
Intercept
+
{p}
3
t
hij
+
{p}
4
t
hij
d
{p}
+
{p}
5
t
hij
d
{p}2
. .
Slope
+
{p}
hij
. (8.37)
c2638 C2638texC08 February 12, 2002 11:38
PARAMETRIC MODELS 171
Figure 8.6. Continued
Pattern-specic estimates using this approach are illustrated in the plots
in the lower half of Figure 8.6. Examination of all plots suggests that both
approaches reasonably approximate the observed data. This is indicated by
the close approximation of the curves to estimates of change obtained using
MLE with a piecewise linear regression model. Table 8.10 summarizes both the
pattern-specic estimates and the pooled parameter estimates. Not surpris-
ingly, the major dierence between the two methods of extrapolation occurs
for subjects who have only a baseline observation (pattern 4). The estimates
of the slopes for the smaller treatment group not receiving Taxol are also
unstable, especially for the second approach. Note that there are almost 200
c2638 C2638texC08 February 12, 2002 11:38
172 PATTERN MIXTURE MODELS
Table 8.10. Pattern-specic estimates for two approaches for the estimation of a
linear model for changes in the FACT-Lung TOI. Model 1 is described by Equa-
tions 8.35 and 8.36. Model 2 is described by Equation 8.37.
Model 1 Model 2
Pattern Group Intercept Slope Intercept Slope
1 No Taxol 71.6 1.01 71.1 0.97
Taxol 71.3 1.01 71.3 1.02
2 No Taxol 65.7 1.80 66.7 0.26
Taxol 67.3 2.68 67.3 2.58
3 No Taxol 62.9 6.33 62.6 1.73
Taxol 63.3 2.67 63.7 2.98
4 No Taxol 57.2 6.33 57.4 11.3
Taxol 59.2 2.67 58.9 3.13
Pooled No Taxol 64.6 3.62 64.6 5.02
Taxol 66.2 2.06 66.0 2.15
subjects in that group. This is a relatively large sample size for many clini-
cal trials, but when we start dividing subjects into patterns, the number of
subjects per pattern may decline rapidly. The sensitivity of the results to dif-
ferent, but reasonable, assumptions is further illustrated in Table 8.8. The dif-
ferences in the 12-week means range from 0.5 to 6.0 points, bracketing the es-
timates obtained from the CCMV, ACMV, and NCMV restrictions. Although
the dierences remain statistically nonsignicant, the wide range is of concern.
Variance estimation
To describe the estimation of the variance of the pooled parameter esti-
mates, we need to introduce some matrix notation [23]. In this notation,
Equation 8.34 for the pooled parameter estimates is

k
=
P

{p}

{p}
= (
P
)

A
k

P
. (8.38)

P
and

P
are the stacked vectors of the r estimates of
{p}
and the c esti-
mates of
{p}
from all P patterns. A
k
is an r c matrix of known constants
(usually 0s and 1s) derived from the partial derivatives of Equation 8.34 with
respect to
P
and
P
:
A
k
=
_

P

{p}

{p}

{p}

{p}
_
. (8.39)
For example, consider the rst approach for the linear patterns where the
slopes of the third and fourth patterns are constrained to be equal,
{4}
h1
=
{3}
h1
.
c2638 C2638texC08 February 12, 2002 11:38
PARAMETRIC MODELS 173
Thus,

h1
=
{1}
h

{1}
h1
+
{2}
h

{2}
h1
+
{3}
h

{3}
h1
+
{4}
h

{4}
h1
=
{1}
h

{1}
h1
+
{2}
h

{2}
h1
+
{3}
h

{3}
h1
+
{4}
h

{3}
h1
. (8.40)
If
{p}
= [
{1}
h0
,
{1}
h1
, . . . ,
{4}
h0
,
{4}
h1
], the rst derivative with respect to
h1
is

h1

{p}
=
_

h1

{1}
h0
.
.
.

h1

{4}
h1
_

_
=
_
0,
{1}
h
, 0,
{2}
h
, 0,
{3}
h
+
{4}
h
, 0, 0
_
.
Notice that the columns that correspond to the intercept parameters are all
zero, as these parameters are not used in the estimate of the slope. Then, if

{p}
= [
{1}
h
,
{2}
h
,
{3}
h
,
{4}
h
], the second derivative with respect to is

h1

{p}

{p}
=
_

_
_
0,
{1}
h
, 0,
{2}
h
, 0,
{3}
h
+
{4}
h
, 0, 0
__

{1}
h
_
0,
{1}
h
, 0,
{2}
h
, 0,
{3}
h
+
{4}
h
, 0, 0
__

{2}
h
_
0,
{1}
h
, 0,
{2}
h
, 0,
{3}
h
+
{4}
h
, 0, 0
__

{3}
h
_
0,
{1}
h
, 0,
{2}
h
, 0,
{3}
h
+
{4}
h
, 0, 0
__

{4}
h
_

_
=
_

_
0 1 0 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0 0 0 0 1 0 0
0 0 0 0 0 1 0 0
_

_
.
This is more work than one needs to do to derive the pooled estimates, but
it greatly facilitates the estimation of the variance that follows.
When the proportions are known (and not estimated), the variance of

k
is
solely a function of the variance of the parameter estimates:
Var[

k
] =
P
A
k
Var[

P
]A

k

P
. (8.41)
However, because the proportions are also estimates, we use the delta method
(Taylor series approximation) to approximate the variance of the pooled esti-
mates:
Var[

k
] =
P
A
k
Var[

P
]A

k

P
+

P
A

k
Var[
P
]A
k

P
, (8.42)
where Var[
P
h
] = [diag(
P
h
)
P
h

P
h
]/N
h
and
Var[
P
] =
_

_
Var[
P
1
]
.
.
.
Var
_

P
H

_
c2638 C2638texC08 February 12, 2002 11:38
174 PATTERN MIXTURE MODELS
for the H independent groups (treatment arms). This approximation is ap-
propriate for large samples, but with small samples a bootstrap procedure
may be necessary. In very large samples, Var(
{p}
h
) is very small and, for all
practical purposes, ignorable.
SAS example of a pattern mixture model
In the NSCLC study, the datale EXAMPLE2 has a unique record for each qual-
ity of life assessment and the datale PATIENT2 has one record for each patient.
CASEID identies the patient, TAXOL identies the treatment arm, and MONTHS
identies the time of the assessment. The pattern is dened by LAST QoL,
which is the follow-up number of the last HRQoL assessment. For example,
pattern 1 contains subjects with only the rst HRQoL assessment. Thus, the
order of the patterns in the example is reversed from that shown above.
Step 1: Estimates of
P
h
The rst step is to obtain estimates for proportions in each pattern by treat-
ment group. This is done using the SAS FREQ procedure. We will need both the
proportions and the number in each treatment group, so we use SAS macro
variables &p1, &p2, &n1, and &n2 to store the information.
* Row=Treatment Group, Column=Pattern Identifier *;
PROC FREQ DATA=PATIENT2;
TABLE TAXOL*LAST_QoL/NOPERCENT NOCOL;
RUN;
%LET P1=.3043, .1576, .2283, .3098; %LET N1=184;
%LET P2=.2110, .2082, .2137, .3671; %LET N2=365;
Step 2: Estimates of
P
h
The second step is to obtain the parameter estimates for the linear regression
equations in each pattern. The CLASS statement denes TAXOL and LAST QoL
as categorical variables so that the term TAXOL*LAST QoL in the MODEL state-
ment, combined with the NOINT option, denes a design matrix for unique in-
tercepts within each pattern and treatment group. The term TAXOL*LAST QoL*
MONTHS denes a design matrix for unique slopes within each pattern and
treatment group. The SOLUTION and COVB options will allow us to create new
data sets containing estimates of
P
h
and its variance.
PROC MIXED DATA=WORK.MERGED NOCLPRINT COVTEST;
CLASS TAXOL LAST_QoL;
MODEL FACT_T2=TAXOL*LAST_QoL TAXOL*LAST_QoL*MONTHS
/NOINT SOLUTION COVB;
RANDOM INTERCEPT MONTHS/SUBJECT=CASEID TYPE=UN;
c2638 C2638texC08 February 12, 2002 11:38
PARAMETRIC MODELS 175
In this example, we assume a mixed-eects model with two random eects,
INTERCEPT and MONTHS, similar to that described in Chapter 3. We also assume
that the covariance structure is the same across all patterns and treatments.
In theory, the covariance of each pattern may be dierent. However, in prac-
tice some restrictions are required for estimation of the variance parameters
in patterns with a limited number of repeated assessments. Estimates of vari-
ance parameters in small samples lack the precision of estimates from larger
samples. Thus, pooling across treatments and patterns will also stabilize the
estimates of the variance parameters.
The addition of two statements creates data sets containing the estimates
of the parameters and their variances. The statements would appear as
ODS OUTPUT SOLUTIONF=WORK.BETA COVB=WORK.VARBETA;
We also create another macro variable, &NBETA, equal to the number of esti-
mated parameters for use later in the program.
%LET NBETA=16;
Note that if we needed only estimates and not the variance of the pooled in-
tercept (I) and slope (S) parameters, these could be obtained with ESTIMATE
statements using the estimated proportion of subjects in each stratum. How-
ever, because the proportions are also estimates, the standard errors and other
test statistics reect an underestimation of the true variance of the pooled es-
timates.
ESTIMATE No I TAXOL*LAST_QoL .30 .16 .23 .31
.00 .00 .00 .00;
ESTIMATE Yes I TAXOL*LAST_QoL .00 .00 .00 .00
.21 .21 .21 .37;
ESTIMATE No S TAXOL*LAST_QoL*MONTHS .00 .46 .23 .31
.00 .00 .00 .00;
ESTIMATE Yes S LAST_QoL*TAXOL*MONTHS .00 .00 .00 .00
.00 .42 .21 .37;
Step 3: Pooling estimates and computing variance
The next step is to pool the estimates and calculate the variance of the pooled
estimates. To accomplish this, the SAS IML procedure is used for matrix
manipulation. First, we import the estimates and variance of
P
obtained
from the second step. Then, we create the matrices P BETA and P COVB:
PROC IML;
USE WORK.BETA; * Input Estimates of Beta *;
READ ALL VAR{ESTIMATE} INTO P_BETA[ROWNAME=EFFECT];
USE WORK.VARBETA; * Input Variance of Beta *;
READ ALL VAR("COL1":"COL&NBETA") INTO P_COVB[ROWNAME=EFFECT];
Then the variances of the proportions are estimated using the SAS macro
variables created in the rst step. Note that the variance of the proportion
c2638 C2638texC08 February 12, 2002 11:38
176 PATTERN MIXTURE MODELS
is calculated for each treatment group separately and then combined into a
block diagonal matrix. The matrix VAR PI is created as follows:
P_PATTRN={&P1}//{&P2}; * Vector of proportions *;
VAR1=(DIAG({&P1})-{&P1}*{&P1})/&N1; * No Taxol *;
VAR2=(DIAG({&P2})-{&P2}*{&P2})/&N2; * Taxol *;
VAR_PI=BLOCK(VAR1,VAR2);
It is easier to identify A
k
and then compute the quantities

k
,
P
A
k
, and

P
A

k
separately for each
k
; thus, we use a short macro. Note that the
number of rows of A
k
should equal the number of patterns times the number of
groups (treatment arms), which is also the dimension of VAR PI. The number
of columns should equal the number of elements in P BETA.
%MACRO POOL(A,A_PI,A_BETA,BETA);
&BETA=P_PATTRN*&A*P_BETA;
&A_PI=(&A*P_BETA);
&A_BETA=P_PATTRN*&A;
%MEND POOL;
* No Taxol - Itercept *;
C1={1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0};
%pool(C1,AP1,AB1,Beta1);
* No Taxol - Slope *;
C2={0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0};
%pool(C2,AP2,AB2,Beta2);
* Taxol - Itercept *;
C3={0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0};
%pool(C3,AP3,AB3,Beta3);
c2638 C2638texC08 February 12, 2002 11:38
PARAMETRIC MODELS 177
* Taxol - Slope *;
C4={ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0,
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1};
%pool(C4,AP4,AB4,Beta4);
The estimates

k
,
P
A
k
, and

P
A

k
are concatenated and Var(

) is cal-
culated.
BETA=BETA1//BETA2//BETA3//BETA4;
A_PI=AP1//AP2//AP3//AP4;
A_BETA=AB1//AB2//AB3//AB4;
VAR_BETA=A_BETA*P_COVB*A_BETA+A_PI*VAR_PI*A_PI;
SE_BETA=SQRT(VECDIAG(NEW_VAR));
PRINT Estimates BETA[FORMAT=6.3] Std Err [FORMAT=5.3];
Step 4: Hypothesis testing
A SAS macro (TEST) is created to facilitate multiple hypothesis tests:
START TEST(BETA,VARBETA,C,C_NAME);
THETA=C*BETA;
VAR_TH=C*VARBETA*C;
SE_TH=SQRT(VECDIAG(VAR_TH));
Z=THETA/SE_TH;
P_VAL=(1-PROBNORM(ABS(Z)))*2;
CHISQ=THETA*INV(VAR_TH)*THETA;
DF=NROW(C);
P_VAL2=(1-PROBCHI(CHISQ,DF));
PRINT Estimates of Theta and Univariate Tests (2-sided),
C_NAME THETA[FORMAT=5.2] SE_TH[FORMAT=5.2]
Z[FORMAT=5.2] P_VAL[FORMAT=6.4];
PRINT Wald Chi2 Test of Theta=0,
CHISQ[FORMAT=5.2] DF[FORMAT=2.] P_VAL2[FORMAT=6.4];
FINISH;
We then estimate the means at 12 weeks (or 2.77 months) and test the
dierence.
PRINT TWELVE WEEK ESTIMATES;
C={1 2.77 0 0, 0 0 1 2.77};
C_NAME={NO TAXOL, TAXOL};
RUN TEST(NEW_BETA,NEW_VAR,C,C_NAME);
C={-1 -2.77 1 2.77};
C_NAME={NO VS YES};
RUN TEST(NEW_BETA,NEW_VAR,C,C_NAME);
c2638 C2638texC08 February 12, 2002 11:38
178 PATTERN MIXTURE MODELS
8.5 Additional reading
Extensions of bivariate case
Little and Wang [93] extend the bivariate case for normal measures to a more
general case for multivariate measures with covariates. However, their exten-
sion is still limited to cases where there are only two patterns of missing data,
one of which consists of complete cases. For the nonignorable missing data, the
multivariate analogue of the restriction is used (
{2}
[1.2]
=
{1}
[1.2]
). The model
is just identied when the number of missing observations exactly equals the
number of nonmissing observations in the second pattern. Explicit expres-
sions can be derived for the maximum likelihood estimates. When the number
of nonmissing observations exceeds the number of missing observations, the
model is overidentied. Little and Wang recommend an EM algorithm to ob-
tain estimates. Finally, if the number of missing observations is greater than
the number of nonmissing observations, additional restrictions are required.
Extensions of the sensitivity analysis
Small-sample Bayesian inference is described by Little [91] for the bivariate
case described in Section 8.2. Daniels and Hogan [26] describe an extension of
the sensitivity analysis to a more general missing-data pattern. They illustrate
the extension with a study that has three repeated measures and monotone
dropout. The sensitivity analysis is no longer a function of a single parameter,
, but of multiple parameters. The attraction of their approach is that by
reparameterizing the restriction, the unknown parameters are interpretable
as the between-pattern dierences in means and variances.
Nonparametric analyses
Yao et al. [170] describe an extension of a nonparametric analysis of longitudi-
nal data with nonignorable dropout. The advantage of the proposed method
is that it can be applied to settings where the HRQoL assessments do not fol-
low the planned timing of the assessments and the distribution of the HRQoL
scores is non-normal. There is, however, a strong assumption made in this
extension. Specically, this approach assumes that the dierences in HRQoL
between the two treatment arms do not vary over time. This assumption
severely limits the settings where the method is useful.
8.6 Algebraic details
Simple linear regression of Y on X

1[Y X]
=

(X
i

X)(Y
i

Y )

(X
i

X)
2
=

XY

2
X
,

0[Y X]
=

Y

1[Y X]

X =

Y

XY

2
X

X,
where
2
X
=

(X
i

X)
2
/(n 1) and
XY
=

(X
i

X)(Y
i

Y )/(n 1).
c2638 C2638texC08 February 12, 2002 11:38
ALGEBRAIC DETAILS 179
Complete-case missing variable restriction
Equation 8.9

{2}
2
=

{1}
0[21]
+

{1}
1[21]

Y
{2}
1
=

Y
{1}
2


{1}
12

{1}
11

Y
{1}
1
. .

{1}
0[21]
+

{1}
12

{1}
11
. .

{1}
1[21]

Y
{2}
1
=

Y
{1}
2


{1}
12

{1}
11
_

Y
{1}
1


Y
{2}
1
_
.
Equation 8.11

2
=
{1}

{1}
2
+
{2}

{2}
2
=
{1}

Y
{1}
2
+
{2}
_

Y
{1}
2


{1}
12

{1}
11
_

Y
{1}
1


Y
{2}
1
_
_
. .
Equation 8.9
= (
{1}
+
{2}
)
. .
1

Y
{1}
2
+

{1}
12

{1}
11
_

{2}

Y
{1}
1
+
{2}

Y
{2}
1
_
=

Y
{1}
2
+

{1}
12

{1}
11
_
_
_

{2}

Y
{1}
1

{1}

Y
{1}
1
. .
Subtracted
+
{1}

Y
{1}
1
. .
Added
+
{2}

Y
{2}
1
_
_
_
=

Y
{1}
2
+

{1}
12

{1}
11
_
_
_
_

{2}

Y
{1}
1

{1}

Y
{1}
1
. .

Y
{1}
1
+
{2}

Y
{2}
1
+
{1}

Y
{1}
1
. .

Y
1
or
1
_
_
_
_
=

Y
{1}
2
+

{1}
12

{1}
11
_

1

Y
{1}
1
_
.
Browns protective restriction
Equation 8.17

{2}
2
=
_

Y
{2}
1

{1}
0[12]
__

{1}
1[12]
c2638 C2638texC08 February 12, 2002 11:38
180 PATTERN MIXTURE MODELS
=
_
_
_
_
_
_
_
_

Y
{2}
1
[

Y
{1}
1


{1}
12

{1}
22

Y
{1}
2
]
. .

{1}
0[12]
_
_
_
_
_
_
_
_
_
_

{1}
12

{1}
22
. .

{1}
1[12]
=

{1}
22

{1}
12

Y
{2}
1


{1}
22

{1}
12

Y
{1}
1
+

Y
{1}
2
=

Y
{1}
2


{1}
22

{1}
12
_

Y
{1}
1


Y
{2}
1
_
.
Equation 8.19

2
=
{1}

{1}
2
+
{2}

{2}
2
=
{1}

Y
{1}
2
+
{2}
_

Y
{1}
2


{1}
22

{1}
12
_

Y
{1}
1


Y
{2}
1
_
_
. .
Equation 8.17
=
{1}

Y
{1}
2
+
{2}

Y
{1}
2
. .

Y
{1}
2
+

{1}
22

{1}
12
_

{2}

Y
{2}
1

{2}

Y
{1}
1
_
=

Y
{1}
2
+

{1}
22

{1}
12
_
_
_

{2}

Y
{2}
1
+
{1}

Y
{1}
1
. .
Added

{1}

Y
{1}
1
. .
Subtracted

{2}

Y
{1}
1
_
_
_
=

Y
{1}
2
+

{1}
22

{1}
12
_
_
_
_

{2}

Y
{2}
1
+
{1}

Y
{1}
1
. .

Y
1
or
1

{1}

Y
{1}
1

{2}

Y
{1}
1
. .

Y
{1}
1
_
_
_
_
=

Y
{1}
2
+

{1}
22

{1}
12
_

1

Y
{1}
1
_
.
Other

1

Y
{1}
1
=
_

{1}

Y
{1}
1
+
{2}

Y
{2}
1
_
. .

1

Y
{1}
1
= (
{1}
1)
. .

{2}

Y
{1}
1
+
{2}

Y
{2}
1
=
{2}

Y
{1}
1
+
{2}

Y
{2}
1
=
{2}
_

Y
{2}
1


Y
{1}
1
_
.
c2638 C2638texC08 February 12, 2002 11:38
SUMMARY 181
8.7 Summary
Pattern mixture models have the advantage that a model for the dropout
mechanism does not need to be specied.
This advantage is balanced by the need to specify a set of restrictions to
estimate all of the parameters.
The validity of these restrictions cannot be tested and the results may be
sensitive to the choice of restrictions.
It is, however, easy to display the underlying assumptions by plotting the
trajectories for each pattern.
A large number of patterns and small sample sizes will limit the application
of pattern mixture models.
c2638 C2638texC08 February 12, 2002 11:38
c2638 C2638TexturesC09 February 12, 2002 11:42
CHAPTER 9
Random-Eects Mixture,
Shared-Parameter, and Selection
Models
9.1 Introduction
In this chapter, we examine three more models for nonignorable missing data.
They represent a small subset of all those that have been proposed. However,
these three models clearly illustrate the following:
1. All models for nonignorable data require the analyst to make strong as-
sumptions.
2. These assumptions cannot be formally tested. Defense of the assumptions
must be made on a clinical basis rather than a statistical basis.
3. Lack of evidence of nonignorable missing data for any particular model
does not prove that the missing data are ignorable.
4. Estimates are not robust to model misspecication.
Mixture models
The term mixture model originates from the concept of mixing the distribu-
tions f(Y|M) or, more formally, integrating over f(M) to obtain the marginal
distribution of the longitudinal outcome. M
i
denotes a vector of variables
representing the missing data process. Often M
i
= R
i
, where R
i
denotes a
vector of binary indicators of whether a particular HRQoL score was observed
(r
ij
= 1) or missing (r
ij
= 0), but it can also denote the time to dropout, an
associated event (T
D
i
), or a random coecient (
i
). The specic form of M
distinguishes the various mixture models. When M
i
= R
i
, the missingness can
be classied into patterns (see Chapter 8). In the special case where missing
data are due to dropout, the distributions may be a function of the time to
dropout or an associated event.
Pattern mixture M
i
= R
i
f(Y
i
|R
i
). (9.1)
Time-to-event mixture M
i
= T
D
i
f
_
Y
i
|T
D
i
_
. (9.2)
Random-eects mixture M
i
=
i
f(Y
i
|
i
). (9.3)
Finally, in the random-eects mixture, f(Y
i
|
i
) is a mixed-eects model where
the random-eects model includes the dropout time T
D
i
. Both the conditional
linear model proposed by Wu and Bailey [167] and the joint model proposed by
Schluchter [133] are random-eects mixture models (see Sections 9.2 and 9.3).
c2638 C2638TexturesC09 February 12, 2002 11:42
184 RANDOM-EFFECTS MIXTURE
Selection models
The term selection model was originally used to classify models with a univari-
ate response y
i
, where the probability of being selected into a sample depended
on the response.
In the above sections we have used the notation f(M|Y
obs
, Y
mis
, X). We
can expand this denition to dierentiate among selection models. Specically,
in outcome-dependent [67] selection models the mechanism depends directly on
the elements of Y, and in random-eects [67] or random-coecient [92] selec-
tion models missingness depends on Y through the subject-specic random
eects
i
or d
i
.
Outcome-dependent selection f
_
M
i

Y
obs
i
, Y
mis
i
_
(9.4)
Random-eects-dependent selection f(M
i
|
i
) (9.5)
An example of an outcome-dependent selection model is the model proposed
by Diggle and Kenward [36] (see Section 9.4), and an example of a random-
eects selection model is the joint model proposed by Schluchter [133] and
DeGruttola and Tu [30] (see Section 9.3).
In addition to adjusting the estimates for the missing data, these models
allow the investigator to explore possible relationships between HRQoL and
explanatory factors causing missing observations. This is particularly interest-
ing, for example, if death or disease progression was the cause of dropout. The
criticism of selection models is that the validity of the models for the missing
data mechanism is untestable because the model includes the missing values
(Y
mis
i
) as an explanatory variable. The major concern is that the estimates of
the primary parameters () describing changes in HRQoL are very sensitive
to misspecication of the missing data model.
Overview
In some trials, it is reasonable to believe that the rate of change in HRQoL
is associated with the length of time a subject remains on the study. This is
typical of patients with rapidly progressing disease, where more rapid decline
in HRQoL is associated with earlier termination of the outcome measurement.
The rst two models are based on that principle. The rst model is the condi-
tional linear model proposed by Wu and Bailey [167] (see Section 9.2). Each
individuals rate of change in HRQoL is assumed to depend on covariates and
the time to an event associated with dropout. The second model is the joint
mixed-eects and time to dropout-event model proposed by Schluchter [133]
and DeGruttola and Tu [30]. These rst two models are appropriate to set-
tings where simple growth curve models describe the changes in the outcome
and there is variation in the rates of change among the individual subjects
(Table 9.1). Further distinctions between these two models are discussed in
more detail later in the chapter.
The third model is a logistic selection model proposed by Diggle and
Kenward [36] for a repeated measures design. This model is appropriate to
c2638 C2638TexturesC09 February 12, 2002 11:42
CONDITIONAL LINEAR MODEL 185
Table 9.1. General requirements of the conditional linear model, joint mixed-eects
model, and logistic selection model.
Conditional Joint mixed- Logistic
Model characteristic linear eects selection
Repeated measures No No Yes
Growth curve model Linear
a
Yes Possible
Slope random-eect Required
b
Required
c
No
d
Baseline missing Not allowed
e
Allowed Not allowed
Mistimed observations Allowed Allowed Not allowed
Monotone dropout Yes Yes Required
Intermittent pattern Yes if MAR Yes if MAR Not allowed
Censoring of T
D
Not allowed Yes
a
Higher-order polynomials possible but challenging.
b
Random intercept is unrelated to dropout.
c
Variation of slope is required.
d
Random intercept only.
e
Not allowed if baseline value included as a covariate in the model.
settings where there are equally spaced repeated measures and the missing
data patterns are strictly monotone. Other assumptions are described later in
the chapter.
9.2 Conditional linear model
Wu and Bailey [167] proposed a conditional linear model for studies where
the primary interest focuses on estimating and comparing the average rate
of change of an outcome over time (slope). The basis of their models is the
assumption that the slope depends on covariates, the baseline value of the
outcome, and the time of dropout T
D
i
. Consider a model where each individ-
uals HRQoL can be described by a linear function of time:
Y
i
= X
i

i
+
i
, X

i
=
_
1 1 1
t
i1
t
i2
t
iJ
_
,
i
= [
i1

i2
] (9.6)
with random variation of the intercept
i1
and rate of change (slope)
i2
.*

i
N(
h
, D)

h
= [
h1

h2
] (9.7)

i
N(0,
2
I) (9.8)
Each individuals slope may depend on M covariates (V
mi
), the initial
value of the outcome (Y
i1
), as well as a polynomial function of the time of
dropout (T
D
i
). The form of the relationship is allowed to vary across the h
* Note that the model is equivalent to a mixed-eects model, X
i

i
= X
i
+ Z
i
d
i
, when
X
i
= Z
i
and
i
= +d
i
.
c2638 C2638TexturesC09 February 12, 2002 11:42
186 RANDOM-EFFECTS MIXTURE
treatment groups. The expected slope for the ith individual is
E
_

i2

T
D
i
, V
mi
, Y
i1

=
L

l=0

hl
_
T
D
i
_
l
+
M

m=1

h(L+m)
V
mi
+
h(L+M+1)
Y
i1
, (9.9)
where
h0
, . . . ,
h(L+M+1)
are the coecients for the hth group and L is the
degree of the polynomial. The intercept for the ith individual,
i1
, is not
dependent on covariates or the time of dropout; thus, the expected intercept is
E[
i1
] =
h1
. (9.10)
By substituting Equations 9.9 and 9.10 into Equation 9.6,
Y
ij
= E[
i1
] +E[
i2
]t
ij
+
ij
=
h1
+
_
L

l=0

hl
_
T
D
i
_
l
+
M

m=1

h(L+m)
V
mi
+
h(L+M+1)
Y
i1
_
t
ij
+
ij
.
(9.11)
The mean slope in the hth group is the expected value of the individual slopes
of the subjects in the hth group (i h):

h2
= E
ih
_

i2

T
D
i
, V
mi
, Y
i1

. (9.12)
One of the estimates proposed by Wu and Bailey is referred to as the linear
minimum variance unbiased estimator (LMVUB):

h2
=
L

l=0

hl
_

T
D
h
_
l
+
M

m=1

h(L+m)

V
hm
+
h(L+M+1)

Y
h1
, (9.13)
where

T
D
h
is the mean dropout time in the hth group,

V
hm
is the mean of the
mth covariate in the hth group, and

Y
h1
is the mean of the baseline measure in
the hth group. In a randomized trial, prerandomization characteristics are the-
oretically the same in all treatment groups. To avoid introducing dierences
that are the result of random dierences in these baseline characteristics,

V
hm
and

Y
h1
are estimated using all randomized subjects (

V
m
and

Y
1
).
There are several practical consequences and assumptions of this model:
1. Changes over time are assumed to be roughly linear.
2. There is enough variation in the rate of change among subjects to allow
modeling of the variation.
3. All subjects must have a baseline measurement and complete data for the
selected covariates (V ) or they will be excluded.
4. The time of dropout is known for all subjects. All subjects have been fol-
lowed until they either dropped out or completed the nal assessment.
5. Subjects with an assessment at the last follow-up behave as if they dropped
out immediately after that point in time. Basically, there is no way to
distinguish between subjects who would have continued to have assessments
and those who would have dropped out before the next assessment if the
follow-up extended.
c2638 C2638TexturesC09 February 12, 2002 11:42
CONDITIONAL LINEAR MODEL 187
Testing MAR vs. MNAR under assumptions of conditional linear model
Note that if the slopes were dependent only on covariates, we would consider
the data to be MCAR and Equation 9.9 would reduce to
E[
i2
|V
mi
] =
h0
+
M

m=1

hm
V
mi
. (9.14)
If the slopes were dependent on the baseline measure of the outcome but not
the time of dropout, we would consider the data to be MAR. In this case,
Equation 9.9 would reduce to
E[
i2
|V
mi
, Y
i1
] =
h0
+
M

m=1

hm
V
mi
+
h(M+1)
Y
i1
. (9.15)
When the terms involving functions of the time of dropout T
D
i
remain in the
model, there is evidence that the data are MNAR. Thus, testing the hypothesis

hl
= 0 vs.
hl
= 0, l = 1, . . . , L is a test of the MAR assumption against
MNAR. If it is rejected, there is evidence that missingness is nonignorable.
However, if the test is not rejected, one can only conclude that the missingness
is ignorable only if this particular model is correct. Unfortunately, we cannot
test or prove that the model is correctly specied. Thus, failing to reject the
hypothesis cannot be considered proof that the data are MAR.
NSCLC example
Previous examination of the NSCLC data suggests that the change in the
FACT-Lung TOI scores over time is approximately linear. Assuming that the
endpoint of interest is comparison of the slopes of patients not receiving Taxol
to those receiving Taxol, one might proceed as follows:
1. Estimate the slopes in a simple mixed-eects model.
Model 1: Y
ij
=
hi1
+
h0
t
ij
+
ij
. (9.16)
The SAS statements are as follows:
PROC MIXED DATA=WORK.MERGED METHOD=ML COVTEST;
CLASS TAXOL;
MODEL FACT_T2=TAXOL TAXOL*MONTHS/NOINT SOLUTION;
RANDOM INTERCEPT MONTHS/SUBJECT=CASEID TYPE=UN;
ESTIMATE NO - SLOPE TAXOL*MONTHS 1 0;
ESTIMATE YES - SLOPE TAXOL*MONTHS 0 1;
ESTIMATE DIFF - SLOPE TAXOL*MONTHS -1 1;
RUN;
Check the variance of the random slope eects. If this value is small (i.e., not
signicantly dierent from zero), seriously reconsider using a conditional
linear model as an analysis strategy. First, it does not make sense to develop
a model to explain variation when there is none. Second, models will be very
slow to converge (if they do at all) when they include a variance parameter
c2638 C2638TexturesC09 February 12, 2002 11:42
188 RANDOM-EFFECTS MIXTURE
Table 9.2. Summary of slope estimates based on the conditional linear model.
Model No Taxol Taxol Dierence Variance 2 log L
2
2
1 0.74 (0.32) 0.92 (0.21) 0.18 (0.38) 1.74 (0.88) 11226.8
2 0.85 (0.32) 0.95 (0.21) 0.95 (0.38) 1.76 (0.88) 11215.6 11.2
3 0.97 (0.32) 0.89 (0.21) 0.07 (0.38) 0.55 (0.70) 10901.1 314.5
4 1.77 (0.59) 1.87 (0.35) 0.10 (0.68) 0.67 (0.75) 10886.4 14.7
5 3.35 (0.89) 2.64 (0.51) 0.77 (1.02) 0.46 (0.78) 10876.3 10.1
that is very close to zero. In this example, the variance of the slopes is
signicant but only moderately dierent from zero (Table 9.2).
2. Identify baseline covariates that might predict change and test the signi-
cance of their interaction with time in a mixed-eects model:
Model 2: Y
ij
=
hi1
+
h0
t
ij
+
M

m=1

hm
V
mi
t
ij
. .
Addition to Model 1
+
ij
. (9.17)
Symptoms of metastatic disease (SX MET) were the strongest predictor of the
slope among the available demographic* and disease measures prior to
treatment.** Note that the term TAXOL*MONTHS*SX MET has been added to
the MODEL statement.
PROC MIXED DATA=WORK.MERGED NOCLPRINT METHOD=ML;
CLASS TAXOL CASEID;
MODEL FACT_T2=TAXOL TAXOL*MONTHS TAXOL*SX_MET*MONTHS
/NOINT SOLUTION;
RANDOM INTERCEPT MONTHS /SUBJECT=CASEID TYPE=UN;
ESTIMATE NO - SLOPE TAXOL*MONTHS 1 0
TAXOL*SX_MET*MONTHS 0.55 0;
ESTIMATE YES - SLOPE TAXOL*MONTHS 0 1
TAXOL*SX_MET*MONTHS 0 0.55;
ESTIMATE DIFF - SLOPE TAXOL*MONTHS -1 1
TAXOL*SX_MET*MONTHS -0.55 0.55;
RUN;
Prior to running the procedure to estimate the parameters of the mixed-
eects model, the mean of SX MET was estimated in each of the two groups.
Because SX MET was coded as an indicator variable,*** the mean is also
the proportion of patients with symptoms of metastatic disease at diag-
nosis in each group, 58 and 54%, respectively. The means were inserted
into the ESTIMATE statement to estimate the group slopes (
h2
). Addition of
symptoms of metastatic disease prior to treatment explained a signicant
* Age and gender.
** Stage of disease, Eastern Cooperative Oncology Group (ECOG) performance status,
>5% weight loss in last 6 months, primary disease symptoms, metastatic disease symp-
toms, systemic disease symptoms.
*** 0 = No, 1 = Yes.
c2638 C2638TexturesC09 February 12, 2002 11:42
CONDITIONAL LINEAR MODEL 189
proportion of the variability of the slope (
2
2
= 35.3, p < 0.001), but
it did not aect the estimates of the slopes in the two treatment groups
(Table 9.2).
3. Add the baseline measure of the outcome variable to the model and test
for an interaction with time.
Model 3: Y
ij
=
hi1
+
h0
t
ij
+
M

m=1

hm
V
mi
t
ij
+
h(M+1)
Y
i1
t
ij
. .
Addition to Model 2
+
ij
.
(9.18)
The interaction term TAXOL*BASELINE*MONTHS is added to the MODEL state-
ment. To facilitate the interpretation of the coecients, BASELINE was
calculated as the initial FACT-Lung TOI score minus 65, which was ap-
proximately the mean of the baseline scores.
PROC MIXED DATA=WORK.MERGED NOCLPRINT METHOD=ML;
CLASS TAXOL CASEID;
MODEL FACT_T2=TAXOL TAXOL*MONTHS TAXOL*SX_MET*MONTHS
TAXOL*BASELINE*MONTHS /NOINT SOLUTION;
RANDOM INTERCEPT MONTHS /SUBJECT=CASEID TYPE=UN;
ESTIMATE NO - SLOPE TAXOL*MONTHS 1 0
TAXOL*SX_MET*MONTHS 0.55 0
TAXOL*BASELINE*MONTHS 0.582 0;
ESTIMATE YES - SLOPE TAXOL*MONTHS 0 1
TAXOL*SX_MET*MONTHS 0 0.55
TAXOL*BASELINE*MONTHS 0 0.582;
ESTIMATE DIFF - SLOPE TAXOL*MONTHS -1 1
TAXOL*SX_MET*MONTHS -0.55 0.55
TAXOL*BASELINE*MONTHS -0.582 0.582;
RUN;
As in the previous case, the average BASELINE value in each group is calcu-
lated prior to tting the mixed-eects model and inserted into the ESTIMATE
statement. Addition of the baseline value of the outcome also explained a
signicant proportion of the variability of the slope (
2
2
= 314.2, p < 0.001),
but again it did not aect the estimates of the slopes in the two treatment
groups (see Table 9.2). This should not be surprising, since Model 2 is
unbiased if missingness depends only on the observed baseline data.
4. Finally, the interaction of time of dropout with the slope is added to the
model.
Models 4 and 5: Y
ij
=
hi1
+
h0
t
ij
+
L

l=1

hl
_
T
D
i
_
l
t
ij
. .
Addition to Model 3
+
M

m=1

h(L+m)
V
mi
t
ij
+
h(L+M+1)
Y
i1
t
ij
+
ij
.
(9.19)
c2638 C2638TexturesC09 February 12, 2002 11:42
190 RANDOM-EFFECTS MIXTURE
There are several possibilities at this point for dening T
D
i
, including the
planned time of the last assessment and the observed time of the last
assessment. In the following illustration, T
D
i
was dened as the observed
time in months from randomization to the last HRQoL assessment
(LAST MO). Other possibilities are the time to disease progression, termi-
nation of therapy, or death. In the NSCLC example, a linear (Model 4:
L = 1) and a quadratic (Model 5: L = 2) model for the time of dropout
were tested. The quadratic model provided the best t:
PROC MIXED DATA=WORK.MERGED NOCLPRINT METHOD=ML;
CLASS TAXOL CASEID;
MODEL FACT_T2=TAXOL TAXOL*MONTHS TAXOL*SX_MET*MONTHS
TAXOL*BASELINE*MONTHS TAXOL*MONTHS*LAST_MO
TAXOL*MONTHS*LAST_MO*LAST_MO
/NOINT SOLUTION;
RANDOM INTERCEPT MONTHS/SUBJECT=CASEID TYPE=UN;
ESTIMATE NO TAXOL TAXOL*MONTHS 1 0
TAXOL*SX_MET*MONTHS 0.55 0
TAXOL*BASELINE*MONTHS 0.582 0
TAXOL*MONTHS*LAST_MO 2.62 0
TAXOL*MONTHS*LAST_MO*LAST_MO 13.62 0;
ESTIMATE TAXOL TAXOL*MONTHS 0 1
TAXOL*SX_MET*MONTHS 0 0.55
TAXOL*BASELINE*MONTHS 0 0.582
TAXOL*MONTHS*LAST_MO 0 3.04
TAXOL*MONTHS*LAST_MO*LAST_MO 0 15.74;
ESTIMATE DIFF TAXOL*MONTHS -1 1
TAXOL*SX_MET*MONTHS -0.55 0.55
TAXOL*BASELINE*MONTHS -0.582 0.582
TAXOL*MONTHS*LAST_MO -2.62 3.04
TAXOL*MONTHS*LAST_MO*LAST_MO 13.62 -15.74;
RUN;
Addition of the time of dropout also explained a signicant proportion of
the variability of the slope (Model 3 vs. 4:
2
2
= 14.7, p < 0.001; Model 4
vs. 5:
2
2
= 10.1, p = 0.005). In contrast with the previous models, there
was a dramatic eect on the estimates of the slopes, with a doubling of the
estimated rate of decline in both treatment groups when the linear interac-
tion was added and a tripling of the rates when the quadratic interaction
was added (see Table 9.2).
Examination of the coecients of Model 5 (Table 9.3) suggests that the
HRQoL measure decreases about 1 point per month faster in subjects with
symptoms of metastatic disease than in those without symptoms prior to
randomization (1.27 and 0.71 in the two groups). It also predicts a faster
decline of the same magnitude in a subject who scored 16 points lower at
baseline than another subject (1/0.061 and 1/0.064). The rate of decline is
slower in those who stay on therapy longer; initially, the relative increase
will range from 1 to 1.6 points per month for each additional month of
follow-up (positive linear terms). However, the benet will diminish with
c2638 C2638TexturesC09 February 12, 2002 11:42
CONDITIONAL LINEAR MODEL 191
Table 9.3. Summary of coecients from the conditional linear model 5.
Eect Taxol Estimate S.E.
Intercept No 64.9 1.17
Yes 66.3 0.82
Slope No 4.18 2.05
Yes 3.79 1.29
Slope metastic disease
a
No 1.23 0.61
Yes 0.71 0.38
Slope baseline
b
No 0.064 0.021
Yes 0.064 0.013
Slope time of dropout
a
No 1.64 0.49
Yes 1.02 0.34
Slope time of dropout
c
No 0.116 0.047
Yes 0.051 0.024
a
0 = No, 1 = Yes.
b
Score = 65.
c
Last HRQoL assessment (months).
Table 9.4. Comparison of estimates of change in FACT-Lung TOI per month from
selected models where change over time is assumed to be linear.
Method No Taxol Taxol Dierence
Ignorable (MLE) 0.75 (0.32) 0.91 (0.21) 0.16 (0.39)
Conditional linear model 4 1.91 (0.91) 1.98 (0.38) 0.06 (0.91)
Conditional linear model 5 3.72 (1.14) 2.83 (0.60) 0.90 (1.23)
Pattern-mixture 1 3.63 (0.96) 2.06 (0.53) 1.57 (1.10)
Pattern-mixture 2 5.03 (1.75) 2.21 (0.85) 2.82 (1.94)
time (negative quadratic terms). For example, a subject on the control arm
with no metastatic disease and a baseline score of 65 is expected to decline
at a rate of 4.2 points per month if the subjects last HRQoL assessment
was at baseline at a rate of 2.7 points per month if the subject dropped
out at 1 month,* and almost not at all if the subjects last HRQoL assess-
ment was at 3 months.** It is reassuring that all of these trends are in the
expected directions.
Table 9.4 contrasts the results of the conditional linear model with the
two pattern-mixture models described in the previous chapter. The esti-
mates of change are between those estimated under the MAR assumptions
and the pattern-mixture models.
* 1 month: 4.18 + 1.64 1 0.12 1
2
= 2.7.
** 3 months: 4.18 + 1.64 3 0.12 3
2
= 0.3.
c2638 C2638TexturesC09 February 12, 2002 11:42
192 RANDOM-EFFECTS MIXTURE
Table 9.5. Comparison of naive and bootstrap estimates of the variance.
Model Method No Taxol Taxol Dierence
Linear Naive 1.82 (0.61) 1.93 (0.35) 0.11 (0.70)
Bootstrap 1.91 (0.91) 1.98 (0.38) 0.06 (0.91)
Quadratic Naive 3.50 (0.91) 2.72 (0.52) 0.77 (1.05)
Bootstrap 3.72 (1.14) 2.83 (0.60) 0.90 (1.23)
Naive estimate assumes that means of covariates, baseline, and time to dropout
are xed and have no variance.
Estimation of the standard errors
Computations of the standard errors reported in the output associated with
ESTIMATE statements are made with the assumption that the means of the
covariates, baseline scores, and time to dropout are xed and known. Because
they are estimates, the standard errors will underestimate the true variance of
the slopes. Wu and Bailey [167] provide corrected estimates of the variance
for a special case. Bootstrapping is a more useful tool to handle all models,
especially with small or moderately sized samples (Table 9.5).
Assumptions
There are two major conditions that must be true to implement the condi-
tional linear model. The rst condition is a linear pattern of change in the
HRQoL measures over time. Although it is theoretically possible to extend
this model to more complex growth curves, the models are very complex and
dicult to estimate unless the size of the sample is very large. In small studies,
there may be more parameters than observations. Even in moderate studies,
there will be insucient information to obtain precise estimates of the pa-
rameters. The second condition is that there is random variation in the slopes
among subjects (Var[d
i2
] 0). Of practical note, this random eect may exist
in Model 1 but disappear as more variation of the slopes is explained. In the
NSCLC example, there was signicant variation among the slopes in models
1 and 2, but the addition of baseline as an interaction explained approxi-
mately 70% of the variability of the slopes. Although all variance parameters
were estimable in this example, it is likely that as the variance approaches
zero problems with convergence of the algorithm will develop and the second
random eect will need to be dropped from the model.
Random-coecient mixture model
The conditional linear model has been described as a random-coecient pat-
tern mixture model [92] and a random-eects mixture model [67]. Each subject
has a unique pattern dened by the dropout time, covariates, and the baseline
measure. The restrictions placed on the model are dened by the conditional
linear model (Equation 9.12).
c2638 C2638TexturesC09 February 12, 2002 11:42
JOINT MIXED-EFFECTS AND TIME TO DROPOUT 193
9.3 Joint mixed-eects and time to dropout
To begin, consider the same basic mixed-eects model described by Equations
9.6 through 9.8, where changes in HRQoL are described by a simple two-
stage mixed-eects model. Each individuals HRQoL is described by a linear
function of time, with random variation among subjects of the intercept,
i1
,
and the linear rate of change (slope),
i2
. The time of an event associated
with the discontinuation of HRQoL measurement (T
D
i
) is incorporated into
the model by allowing some function of the time to discontinuation (f(T
D
i
)) to
be correlated with the random eects of the longitudinal model for HRQoL [30,
122, 133, 134].
_

i
f
_
T
D
i
_
_
N
_
_

T
_
,
_
D
bt
(
bt
)

2
__
. (9.20)
It is important to note that T
D
i
does not have to be the time after which
no HRQoL assessments were made but can, instead, be the time to any event
associated with dropout. This is helpful when some of the dropout is unrelated
to the outcome. For example, if the event were termination of therapy due to
toxicity or lack of ecacy, then individuals who drop out for unrelated reasons
would not be treated as if they experienced a negative outcome.
The rst implication of this model is that the expected changes in HRQoL
are a function of the dropout time. This is illustrated in Figure 9.1. Note that
the patients with earlier dropout start with lower FACT-Lung TOI scores
and decline more rapidly than subjects who remain longer. In the model
(Equation 9.20), both the intercepts (
i1
) and the slopes (
i2
) are positively
Figure 9.1. Expected change in FACT-Lung TOI among NSCLC patients given
death occurred at 1, 3, 6, and 12 months.
c2638 C2638TexturesC09 February 12, 2002 11:42
194 RANDOM-EFFECTS MIXTURE
correlated with longer survival (
bt
> 0). More formally, the conditional
expectation of the random eects is a function of the dropout time.
E
_

T
D
i

= +
bt

2
_
f
_
T
D
i
_

t
_
. (9.21)
In the example presented in Figure 9.1, the expected intercept for a subject
surviving to T
D
i
is
E
_

i1

T
D
i

= 65.3 + 8.381 1.721

1
_
ln
_
T
D
i
_
+ 0.088
_
= 65.7 + 4.870 ln
_
T
D
i
_
.
Thus, the average baseline score is 59.0 for a patient who survives 3 months
(ln(0.25 years) = 1.386) and 65.7 for a patient surviving 12 months (ln(1.00
years) = 0.00). The expected change over time is
E
_

i2

T
D
i

= 1.78 + 1.974 1.721

1
_
ln
_
T
D
i
_
+ 0.088
_
= 1.77 + 1.147 ln
_
T
D
i
_
.
The predicted decline is 3.4 points per month for a patient who survives 3
months but only 1.8 points per month for a patient surviving 12 months.
This model diers from the conditional linear model in several respects.
1. The joint model allows the intercept and slope to be related to the time of
dropout.
2. There is an added restriction that the relationships are linear functions of
f(T
D
i
).
E
_

T
D
i

= +
bt

2
_
f
_
T
D
i
_

t
_
=
bt

t
. .

h0
+
bt

2
. .

h1
f
_
T
D
i
_
. (9.22)
3. The algorithms developed by Schluchter [133] and DeGruttola and Tu [30]
allow censoring of T
D
i
. This allows us to dierentiate a subject who com-
pletes the 6-month evaluation but dies shortly thereafter (e.g., T
D
i
= 8
months) from a subject who remains alive (e.g., T
D
i
> 8 months). This
frees the analyst from assigning an arbitrary dropout time to subjects who
complete the study.
The second implication of the joint model is that the expected time of
dropout is a function of the initial HRQoL scores as well as the rate of change
over time. More formally, the conditional distribution of the dropout times is
a function of the random eects.
E
_
f
_
T
D
i
_

=
t
+

bt
D
1

i
. (9.23)
Testing MAR vs. MNAR under the assumptions of the joint model
Under this model, the missing data are nonignorable if the random eects
(particularly the slope) are correlated with the time of dropout (
bt
= 0). For
example, if patients with more rapid rates of decline in their HRQoL scores
c2638 C2638TexturesC09 February 12, 2002 11:42
JOINT MIXED-EFFECTS AND TIME TO DROPOUT 195
failed earlier (and thus were more likely to miss subsequent assessments), the
random eect corresponding to the rate of change is positively correlated with
the failure time.
Again, it should be emphasized that the lack of signicant correlation im-
plies the the missing data are ignorable only if the model for dropout is correct.
If dropout was the result of a sudden change in the outcome rather than a
gradual change (as measured by slopes), then the dropout model is misspeci-
ed and would not identify the nonignorable process.
Selection or mixture model?
Little [92] classies this model as a random-coecient selection model with
nonignorable random-coecient-based dropout ([T
D
i
|X
i
,
i
]). In contrast,
Hogan and Laird [67] describe this model as a random-eects mixture model.
This model has also been referred to as a shared-parameter model [51, 67,
155, p. 326]. We considered the observed data to include the observed HRQoL
responses (Y
obs
i
) as well as the time to dropout (T
D
i
). The assumption is
that the data are missing at random (MAR) conditional on the augmented
observed data. In this framework, the choice of the event that denes T
D
i
might not be strictly the time of dropout but, rather, a more specic event
related to both discontinuation of HRQoL assessment and HRQoL itself. For
example, the event might be a specic cause of dropout such as the lack of
ecacy. Other causes of dropout that are not associated with the treatment
would be treated as if T
D
i
were censored for administrative reasons.
NSCLC example
A practical requirement of the joint model is nonzero variation in the random
eects. If the variance of the random eects is close to zero, it is dicult, if
not impossible, to estimate the covariance between the random eects and
the time to dropout. Prior to embarking on analyses using the joint model, it
is wise to check the estimates of variance in the simpler mixed-eects model.
The COVTEST option in the MIXED procedure statement generates these esti-
mates and their estimated standard errors.
PROC MIXED DATA=BOOK.EXAMPLE2 COVTEST;
CLASS TAXOL;
MODEL FACT_T2=TAXOL TAXOL*MONTHS\NOINT;
RANDOM INTERCEPT MONTHS\SUBJECT=CASEID TYPE=UN;
RUN;
The estimated variances of the random eects for all subscales of the FACT-
Lung are summarized in Table 9.6. The results suggest that there are no
problems estimating the covariance of the random intercepts with time to
dropout, but problems will denitely occur with the social and physical well-
being subscale and possibly for the emotional well-being subscale.
In the NSCLC study, the protocol specied that HRQoL was to be collected
until the nal follow-up, regardless of disease progression or discontinuation
c2638 C2638TexturesC09 February 12, 2002 11:42
196 RANDOM-EFFECTS MIXTURE
Table 9.6. Estimates of variance of random eects for subscales of the FACT-Lung
in the NSCLC study.
Intercept Slope
Subscale Var (S.E.) p Var (S.E.) p
Functional well-being 264.8 (28.0) <0.001 4.46 (1.81) 0.006
Lung CA specic issues 154.1 (16.0) <0.001 2.21 (0.97) 0.01
Emotional well-being 203.8 (19.3) <0.001 1.26 (0.91) 0.08
Physical well-being 177.0 (24.6) <0.001 1.13 (1.43) 0.22
Social well-being 110.8 (12.5) <0.001 No estimate
FACT-Lung TOI 148.6 (15.3) <0.001 1.81 (0.89) 0.02
No estimate: Note in log G not positive definite.
Table 9.7. Joint model for FACT-Lung TOI and various measures of the time to
dropout (T
D
).
Correlation (
i
, T
D
) Change in TOI
Dropout event (T
D
) Intercept Slope Estimate S.E.
None (MLE) 0.86 0.17
End of therapy 0.41 0.22 1.24 0.13
Last HRQoL assessment 0.40 0.35 1.76 0.13
ln(event free survival) 0.43 0.58 1.44 0.13
ln(survival) long follow-up 0.46 0.73 1.78 0.13
ln(survival) short follow-up 0.52 0.82 1.78 0.16
Estimates of correlation between time (T
D
) and random eects (
i
) and linear
change in FACT-Lung TOI (

2
).
of treatment. Thus, theoretically, death is the event that censored the mea-
surement of HRQoL. In practice, it became dicult to follow patients after
disease progression for various reasons. In addition to time to death, one might
consider time to the last HRQoL measurement as a candidate for the joint
model. Finally, there is the possibility that the rate of change in HRQoL
depends on other clinical events, such as time to disease progression.
Table 9.7 summarizes the results from a number of joint models for the
NSCLC study and compares the estimates to the corresponding mixed-eects
model. First, visual examination of the distribution of the dropout-event times
(T
D
i
) suggested that the times of the last HRQoL measurement and the end of
therapy had roughly symmetric (although not strictly normal) distributions,
while the times to disease progression and death were skewed to the right.
As a result, a log transformation was used for time to death (survival) and
time to disease progression (event-free survival). We also considered a model
c2638 C2638TexturesC09 February 12, 2002 11:42
JOINT MIXED-EFFECTS AND TIME TO DROPOUT 197
with survival measured with a short follow-up of only 6 months and a model
with extended follow-up (>5 years), where 90% of the deaths were observed.
All the events were moderately correlated with the intercept ( in the range
of 0.4 to 0.5). The correlation with the slopes was weakest for the time to
treatment discontinuation ( 0.2) and last HRQoL assessment ( 0.35).
The weaker correlations are expected, as there are a wide range of reasons for
these events. There was a stronger correlation with death than disease pro-
gression ( 0.6). One explanation for this is that disease progression is often
determined with radiological imaging, before disease progression produces
symptoms that would impact the subjects HRQoL. The strong correlations
( > 0.7) of change with time to death t with the observation that deteri-
oration in physical and functional well-being accelerate in the months prior
to death. It is also likely that events distant from the period of the study are
not correlated with the changes in the FACT-Lung TOI during the 6 months
of the study.
Initial estimates
The algorithms for estimation of the joint model require initial estimates of the
parameters. Good starting estimates speed up the convergence of the program,
decrease the chance of nding a local minimum, and, for some algorithms,
avoid nonpositive denite covariance matrices. The following procedure is
suggested.
1. Fit a mixed-eects model for the HRQoL data, assuming the data are MAR.
Use these estimates for , D, and
2
.
2. Estimate the distribution of f(T
D
i
). To estimate
T
and
2
when a natural
log transformation is used and T
D
i
is censored for some subjects:
PROC LIFEREG;
MODEL TSURV*FAIL(0)= / DISTRIBUTION=LNORMAL;
RUN;
When the times (T
D
) are assumed to be normally distributed, the estimates
can be obtained using
PROC LIFEREG;
MODEL TSURV*FAIL(0)= /DISTRIBUTION=NORMAL;
RUN;
3. Assume that the random eects are weakly correlated ( 0.2) with
f(T
D
i
). Thus,

1bt
= 0.2
_

11

2
,

2bt
= 0.2
_

22

2
,
where
11
and
22
are the diagonal elements of D.
c2638 C2638TexturesC09 February 12, 2002 11:42
198 RANDOM-EFFECTS MIXTURE
Extension to more complex mixed-eects models
The joint model can be extended to more complex models [41, 42, 122]. Con-
sider the more general mixed-eects model for the HRQoL outcome:
Y
i
= X
i
+Z
i
d
i
+e
i
. (9.24)
The marginal model (X
i
) does not need to be constrained to be linear over
time but could contain polynomial or piecewise regression terms. Further, the
random eects do not have to have a one-to-one correspondence with the xed
eects (X
i
= Z
i
). In this more general model, the distribution displayed in
Equation 9.20 is modied by replacing
i
with d
i
and with 0.
_
d
i
f(T
i
)
_
N
_
_
0

T
_
,
_
D
bt
(
bt
)

2
__
. (9.25)
Renal cell carcinoma example
This extension to more complex models is illustrated in the renal cell carci-
noma example, where the change in the FACT-TOI was not linear over time.
Recall that a piecewise linear model with changes in the slope at 2, 8, and
17 weeks was used to describe change in HRQoL over time. One might con-
sider adding additional random eects corresponding to these terms in the
piecewise linear model. Algorithms for estimating models with more random
eects often fail to converge unless there is sucient information in the data
to estimate the variance of these random eects. This will not be the case
unless there is extended follow-up in a large number of subjects.
In the renal cell carcinoma study, the follow-up on most subjects is relatively
short and we are limited to a model with only two random eects. In this study,
both the intercepts and slopes are correlated with the log transformed time of
death. The expected changes in the FACT-BRM TOI for patients surviving 8,
17, and 52 weeks are displayed in Figure 9.2. As seen in the previous example,
subjects with longer survival tend to start with higher FACT-BRM TOI
scores. A little less obvious in this gure, because of the shapes of the curves,
are the increased initial rates of decline of the subjects with shorter survival.
The overall estimates in the two treatment groups are displayed in Figure 9.3.
Although the shapes of the curves are very similar, the estimates of the rate of
change in the FACT-BRM TOI scores change in the later phase of the study.
Under the MAR assumption, HRQoL appears to be improving after week 17,
whereas under the joint model assumptions the proles have attened.
9.4 Selection model for monotone dropout
Diggle and Kenward [36] proposed a selection model for monotone miss-
ing data. They dened the dropout process in terms that correspond to
MCAR, MAR, and MNAR. Completely random dropout (CRD) corresponds
to MCAR, where dropout is completely independent of measurement. Random
dropout (RD) corresponds to MAR, where dropout depends on only observed
c2638 C2638TexturesC09 February 12, 2002 11:42
SELECTION MODEL FOR MONOTONE DROPOUT 199
Figure 9.2. Expected change in FACT-BRM TOI among renal cell carcinoma pa-
tients given that death occurred at 8, 17, or 52 weeks. Upper gure corresponds to
interferon arm (IFN) and lower gure corresponds to interferon plus cis-retinoic
acid (IFN+CRA) arm of trial.
measurements (prior to dropout). Informative dropout (ID) corresponds to
MNAR, where dropout depends on unobserved measures. The response model
is the same multivariate linear regression model that we have used throughout,
f(Y
i
|). For the dropout model they assume a logistic regression, f(R
i
|Y
i
, ),
which may depend on covariates X
i
(CRD), previous observations (RD), or
the unobserved measurement at the time of dropout (ID). The conditional
c2638 C2638TexturesC09 February 12, 2002 11:42
200 RANDOM-EFFECTS MIXTURE
Figure 9.3. Expected change in FACT-TOI among patients with renal cell carci-
noma. Solid line corresponds to interferon arm; dashed line corresponds to interferon
plus cis-retinoic acid arm of trial.
probability of dropout at time j, given the history through time j 1, is
denoted by
P
j
= Pr
_
T
D
i
= t
j

Y
i1
, . . . , Y
ij1
_
. (9.26)
The logistic linear model for the dropout process takes the form
logit[P
j
] =
0
X
i
+
1
y
ij
+
2
y
ij1
. (9.27)
c2638 C2638TexturesC09 February 12, 2002 11:42
SELECTION MODEL FOR MONOTONE DROPOUT 201
If
2
= 0, then dropout depends on the previously observed response (y
ij1
),
and if
1
= 0, then dropout depends on the current unobserved response (y
ij
).
Therefore, if
1
=
2
= 0, the dropout process is completely random (CRD). If

1
= 0 and
2
= 0, then dropout depends only on the previously observed re-
sponses and the process is RD. If
1
= 0, then dropout depends on the current
data (y
ij
) and the process is informative (ID). If the dropout process is CRD
or RD, the response model and the dropout model can be estimated separately.
However, if the dropout process is ID, and must be estimated jointly.
Outcome-dependent selection model
Little [92] classies this model as a random coecient selection model with
nonignorable outcome-based dropout ([R
i
|X
i
, Y
obs
i
, Y
mis
i
]).
NSCLC example
The NSCLC study is used to illustrate this approach. Note that this is for illus-
trative purposes only, as there are some constraints imposed by the model and
the estimation procedure. First, we will have to exclude 15% of the patients
who have intermittent missing data as well as 6% who have no data. This
leaves 471 subjects. Second, we will have to assume that the interval between
the third and fourth measurements is equivalent to the intervals between the
rst, second, and third with respect to the dropout model. In practice, we
would not use a model that excludes such a large portion of the data, nor
would we assume equal spacing of the assessments.
In the following example, we use the Oswald library of functions written for
S-Plus. A full discussion of the software is presented in the next section. The
rst step is to explore models describing the dropout process under the CRD
assumption. In model 1 (CRD1), we assume that the probability of dropout
is the same for the second, third, and fourth assessment (X
i1
= 1).
logit[P
j
|X
i
] =
01
(9.28)
Pr
_
T
D
i
= t
j
_
=
e

01
1 +e

01
, j = 2, 3, 4. (9.29)
The estimated parameters for the model are displayed in Table 9.8. The esti-
mated probability of dropout (Table 9.9) at each follow-up visit is

P
j
= e
0.806
/(1 +e
0.806
) = 0.31.
This implies that 31% of the remaining subjects drop out at each of the follow-
up visits. The probability of completing the HRQoL assessment at the nal
visit is the product of the probability of remaining in the study at each follow-
up. Thus, the model predicts that 69% of the subjects will have an assessment
at 6 weeks [(1

P
2
) = 0.69], 48% at 12 weeks [(1

P
2
)(1

P
3
) = 0.48],
and 33% at 26 weeks [(1

P
2
)(1

P
3
) (1

P
4
) = 0.33]. The estimated
probabilities of dropout and of remaining in the study are compared to the
observed probabilities in Tables 9.9 and 9.10.
c2638 C2638TexturesC09 February 12, 2002 11:42
202 RANDOM-EFFECTS MIXTURE
Table 9.8. Diggle and Kenward models assuming completely random dropout.
Parameter CRD1 CRD2 CRD3 CRD4

1
Intercept 63.9 63.9 63.9 63.9

2
Group 1.69 1.69 1.69 1.69

3
Time 0.266 0.266 0.266 0.266

4
Group:time 0.073 0.073 0.073 0.073

0
Intercept 0.806 0.597 0.933 0.549
0.954 1.005
0.399 0.198
Group 0.307 0.592
0.071
0.285
2 log L 16179.39 16174.77 16166.90 16158.19
Comparison CRD1 CRD1 CRD3
LR statistic
2
1
= 4.62
2
2
= 13.53
2
3
= 8.69
p-value 0.032 0.0012 0.034
Table 9.9. Probability of dropout at the jth assessment if observed at j1 (Pr[R
ij
=
0|R
ij1
= 1]).
Group Weeks Obs CRD1 CRD2 CRD3 CRD4
No Taxol 6 0.37 0.31 0.35 0.28 0.37
12 0.27 0.31 0.35 0.28 0.27
26 0.45 0.31 0.35 0.40 0.45
Taxol 6 0.24 0.31 0.29 0.28 0.24
12 0.28 0.31 0.29 0.28 0.28
26 0.38 0.31 0.29 0.40 0.38
In model 2 (CRD2), we allow the probability to dier across the two treat-
ments by adding an indicator for the Taxol group (X
i2
= 0 if No Taxol, X
i2
= 1
if Taxol).
logit[P
j
, X
i
] =
01
+
02
X
i2
, (9.30)
Pr
_
T
D
i
= t
j

X
i2
= 0
_
=
e

01
1 +e

0
, (9.31)
Pr
_
T
D
i
= t
j

X
i2
= 1
_
=
e

01
+
02
1 +e

01
+
02
. (9.32)
The estimated probability of dropout is higher among the subjects not as-
signed to the Taxol therapy (

P
k
= 0.35) than among those assigned to the
Taxol therapy (

P
k
= 0.29).
In model 3 (CRD3), we assume that the probability is the same across the
two treatments but diers by time (X
i1
= 1 if t
2
, X
i2
= 1 if t
3
, X
i3
= 1
c2638 C2638TexturesC09 February 12, 2002 11:42
SELECTION MODEL FOR MONOTONE DROPOUT 203
Table 9.10. Probability of a subject remaining on study at the jth assessment
(Pr[R
ij
= 1]).
Group Weeks j Obs CRD1 CRD2 CRD3 CRD4
No Taxol 6 2 0.63 0.69 0.65 0.72 0.63
12 3 0.41 0.48 0.42 0.52 0.46
26 4 0.25 0.33 0.27 0.31 0.25
Taxol 6 2 0.76 0.69 0.71 0.72 0.76
12 3 0.54 0.48 0.51 0.52 0.54
26 4 0.34 0.33 0.36 0.31 0.34
if t
4
, 0 otherwise).
logit[P
j
, X
i
] =
01
X
i1
+
02
X
i2
+
03
X
i3
(9.33)
Pr
_
T
D
i
= t
2
_
=
e

01
1 +e

01
(9.34)
Pr
_
T
D
i
= t
3
_
=
e

02
1 +e

02
(9.35)
Pr
_
T
D
i
= t
4
_
=
e

03
1 +e

03
(9.36)
The predicted dropout rate at 6 weeks (

P
k
=0.28) is very similar to that at
12 weeks (

P
k
=0.28) and increases at 26 weeks (

P
k
= 0.40). This model ts
the observed data better than the model where we assume a constant dropout
rate (see Table 9.8).
Finally, in model 4 (CRD4), we allow the probability to dier across the
two treatments by adding indicators for the Taxol group at each time (X
i4
=
X
i5
= X
i6
= 0 if No Taxol, X
i2
= X
i5
= X
i6
= 1 if Taxol).
logit[P
j
, X
i
] =
01
X
i1
+
02
X
i2
+
03
X
i3
+
04
X
i4
+
05
X
i5
+
06
X
i6
, (9.37)
Pr
_
T
D
i
= t
2

X
i4
= 0
_
=
e

01
1 +e

01
, (9.38)
Pr
_
T
D
i
= t
2

X
i4
= 1
_
=
e

01
+
04
1 +e

01
+
04
, (9.39)
Pr
_
T
D
i
= t
3

X
i5
= 0
_
=
e

02
1 +e

02
, (9.40)
Pr
_
T
D
i
= t
3

X
i5
= 1
_
=
e

02
+
05
1 +e

02
+
05
, (9.41)
Pr
_
T
D
i
= t
4

X
i6
= 0
_
=
e

03
1 +e

03
, (9.42)
Pr
_
T
D
i
= t
4

X
i6
= 1
_
=
e

03
+
06
1 +e

03
+
06
. (9.43)
c2638 C2638TexturesC09 February 12, 2002 11:42
204 RANDOM-EFFECTS MIXTURE
This model ts the observed data better than either of the previous two models
(see Table 9.8). Note that none of the parameter estimates for the HRQoL data
changes. This is to be expected, as the longitudinal model and the dropout
model are estimated separately.
Building on the best CRD model, we t the corresponding RD model by
allowing dropout to depend on the value of the previous observation (Y
ij1
).
Note that this model assumes that this relationship is the same for each follow-
up assessment and the previous measurement. Thus, the relationship between
dropout at 6 weeks and the baseline value of the FACT-Lung TOI is assumed
to be the same as between dropout at 26 weeks and the 12-week value of the
FACT-Lung TOI.
logit[P
j
, X
i
] =
01
X
i1
+
02
X
i2
+
03
X
i3
+
04
X
i4
+
05
X
i5
+
06
X
i6
+
2
Y
ij1
, (9.44)
Pr
_
T
D
i
= t
2

X
i4
= 0
_
=
e

01
+
2
Y
ij1
1 +e

01
+
2
Y
ij1
, (9.45)
Pr
_
T
D
i
= t
2

X
i4
= 1
_
=
e

01
+
04
+
2
Y
ij1
1 +e

01
+
04
+
2
Y
ij1
, (9.46)
and so forth. This RD model ts the data much better than the CRD model,
with strong evidence to reject the CRD assumption (Table 9.11). As ex-
pected, there is still no change in the mean and variance parameters for the
Table 9.11. Diggle and Kenward models contrasting completely random dropout
(CRD), random dropout (RD), and informative dropout (ID).
Parameter CRD4 RD4 ID4

1
Intercept 63.9 63.9 63.9

2
Group 1.69 1.69 1.70

3
Time 0.266 0.266 0.254

4
Group:Time 0.073 0.073 0.068

0
Time2 0.549 1.872 1.810
Time3 1.005 1.377 1.317
Time4 0.198 2.286 2.228
Time2 group 0.592 0.574 0.577
Time3 group 0.071 0.134 0.002
Time4 group 0.285 0.274 0.281

1
0.0022

2
0.0386 0.0397
2 log L 16158.19 16078.13 16078.11
Comparison CRD4 RD3
LR statistic
2
1
= 80.1
2
1
= 0.018
p-value <0.001 0.89
c2638 C2638TexturesC09 February 12, 2002 11:42
SELECTION MODEL FOR MONOTONE DROPOUT 205
longitudinal HRQoL data. The RD model predicts that the probability of
dropout will decrease with increasing FACT-Lung TOI scores during the pre-
vious assessment. Thus, in the patient assigned to the treatment that did not
include Taxol, the predicted probabilities of dropout at 6 weeks are 44, 35,
and 27% for subjects with baseline scores of 55, 65, and 75, respectively.

P
2
=
e

01
+
2
Y
ij1
1 +e

01
+
2
Y
ij1
=
e
1.8720.038555
1 +e
1.8720.038555
= 0.44, X
i4
= 0, Y
i1
= 55,

P
2
=
e

01
+
2
Y
ij1
1 +e

01
+
2
Y
ij1
=
e
1.8720.038565
1 +e
1.8720.038565
= 0.35, X
i4
= 0, Y
i1
= 65,

P
2
=
e

01
+
2
Y
ij1
1 +e

01
+
2
Y
ij1
=
e
1.8720.038575
1 +e
1.8720.038575
= 0.27, X
i4
= 0, Y
i1
= 75.
It is also possible to let dropout depend on the observed scores lagging back
two observations. However, to do this in the current study, we would have had
to restrict the data to subjects with the rst two observations; less than half
of the subjects would remain. Thus, in this study, we are limited to a single
lag.*
Finally, we examine the ID model by adding the term that allows dropout
to depend on the current unobserved FACT-Lung TOI measure (Y
ij
).
logit[P
j
, X
i
)] =
01
X
i1
+
02
X
i2
+
03
X
i3
+
04
X
i4
+
05
X
i5
+
06
X
i6
+
1
Y
ij
+
2
Y
ij1
, (9.47)
Pr
_
T
D
i
= t
2
|X
i4
= 0
_
=
e

01
+
1
Y
ij
+
2
Y
ij1
1 +e

01
+
2
Y
ij1
, (9.48)
Pr
_
T
D
i
= t
2
|X
i4
= 1
_
=
e

01
+
1
Y
ij
+
04
+
2
Y
ij1
1 +e

01
+
04
+
2
Y
ij1
, (9.49)
and so forth. The ID model was not signicantly dierent from the RD model.
This is rather surprising, as there was evidence of nonignorable missing data
for all other models considered in this and the previous chapter. This illus-
trates a very important point. The failure to reject a hypothesis of informa-
tive dropout (ID vs. RD) is not conclusive proof that the dropout is ignorable
(RD). The assumption of random dropout (RD) is acceptable only if the ID
model was the correct alternative and the parametric form of the ID process
is correct. If, for example, either the logistic model was inappropriate or the
assumption that the relationship between dropout and Y
ij
is the same at 6,
12, and 26 weeks is incorrect, then the failure to reject the RD model may be
incorrect. It may also be the case, in this study, that the omission of the second
random eect** corresponding to variation in the rates of change among in-
dividuals may be the critical dierence.
* The Oswald routine will run without error messages other than the note that the algo-
rithm failed to converge. The resulting parameters describing dropout at 6 weeks make
no sense.
** Oswald (Version 3.4) does not allow this option.
c2638 C2638TexturesC09 February 12, 2002 11:42
206 RANDOM-EFFECTS MIXTURE
Oswald program
For this example, I have used S-Plus (Version 4.5) with the Oswald (Version
3.4) library of functions [109] and have reset the S default parameterization,
which uses Helmert contrasts, to the more familiar reference cell design
matrices.
> library(oswald)
> options(contrasts=c("contr.treatment","contr.poly"))
First, two data sets are imported into S-Plus. The rst data set, patient,
has one record per subject and contains the CASEID in column 1 and a treat-
ment indicator for Taxol (TAXOL) in column 2. The second data set, assmnt,
consists of four rows for each subject corresponding to the four planned as-
sessments. It contains the CASEID in column 1, the FUNO in column 2, the
FACT-Lung TOI score (FACT.T2)* in column 4 and the planned time of the
assessment (0, 6, 12, or 26 weeks) in column 6. An ldamat object FACT is
created from three data objects, fact.t2, time, and caseid.
> fact.t2 <- assmnt[,4] # Reads FACT_T2 into FACT.T2
> time <- assmnt[,6]
> caseid <- assmnt[,1]
> fact <- conv.ldamat(fact.t2,time,caseid)
The next S-Plus statement creates a full-rank reference cell design matrix for
GROUPS such that when GROUPS is included in the model the intercept cor-
responds to the patients in the No Taxol group and the GROUP parameter
to the dierence between the two groups.
> groups(fact) <- patient[,2] # Reads TAXOL and creates GROUP
Finally, a design matrix (DT) is created that has indicators for the dierent
times of dropout.
> temp <- assmnt[,2] # Reads FUNO into TEMP
> temp[temp==1] <- 0 # Replaces values of 1 with 0
> dt <- as.factor(temp) # Creates design matrix DT
Note that if there were no dropout at either baseline or 6 weeks, we would set
both the rows corresponding to the rst and second visit to 0.
> temp[temp==1|temp==2] <- 0 # Replaces values of 1 or 2 by 0
Longitudinal model
The pcmid function in the Oswald library ts a parametric correlation model
with informative dropout. It consists of several parts. The longitudinal model
(E[Y
i
] = X
i
, Var[Y
i
] =
2
J +
2
H +
2
I) is dened by formula, vparms,
* S-Plus converts FACT T2 to FACT.T2 when imported from the SAS data set.
c2638 C2638TexturesC09 February 12, 2002 11:42
SELECTION MODEL FOR MONOTONE DROPOUT 207
Table 9.12. Output from pcmid function for a longitudinal model with compound
symmetry.
Longitudinal Data Analysis Model assuming completely random dropout
Call: pcmid(formula=fact~group*time,vparms=c(0,120,0),correxp = 1)
Analysis Method: Maximum Likelihood(ML)
Correlation structure: exp(- phi*|u| ^ 1 )
Maximized log-likelihood: [1] -7438.93
Mean Parameters:
(Intercept) group time group:time
PARAMETER 63.875429 1.694745 -0.2662796 0.07277491
STD.ERROR 1.114316 1.344469 0.1108001 0.13124155
Variance Parameters:
nu.sq sigma.sq tau.sq phi
0 148.1602 120.2703 0
Maximization converged after 48 iterations.
and correxp. I is a diagonal matrix of 1s, J is a matrix of 1s, and H is a
correlation matrix where the elements are
rc
= e
|rc|
.
formula Denes structure of the longitudinal model
vparms=c(#,#,#) Initial estimates of the variance parameters
2
,
2
,
and
correxp=# Exponent in the correlation function
For example, if # indicates a nonzero value, the command
> pcmid(formula=fact~group*time, vparms=c(0,#,0), correxp=1)
denes a model with xed-eect parameters labeled as (Intercept), group,
time, and group:time (Table 9.12). These correspond to the intercept in the
No Taxol group, the dierence between the intercepts for the two groups,
the slope in the No Taxol group, and the eect of Taxol on the slope. Setting
the rst and third numbers in the vparms arguments to zero constrains
2
and
to be zero. Thus, = e
|rc|
= e
0
= 1, H = J, and Var[Y
i
] =
2
J +
2
I.
This results in a variance structure with compound symmetry (Table 9.12).
Although any positive number could be used as an initial estimate of
2
, a close
guess will reduce the computing time and decrease the possibility of conver-
gence to a local maximum. One might also assume a random subject eect
where the residual errors have a serial correlation (AR(1)).
> pcmid(formula=fact~group*time, vparm=c(#,0,#), correxp=1)
In this model,
2
is constrained to zero; thus, Var[Y
i
] =
2
J +
2
H.
c2638 C2638TexturesC09 February 12, 2002 11:42
208 RANDOM-EFFECTS MIXTURE
Dropout model
The dropout model is dened by the dropmodel, drop.cov.parms, and drop.
parms arguments of the pcmid function.
dropmodel A model dening the covariate part of the dropout model
(the parameters of
0
)
drop.cov.parms Initial estimates of the parameters of
0
; initial estimate
of 0 will constrain parameter to be 0
drop.parms Initial estimates of
1
,
2
, etc.; initial estimate of 0 will
constrain the parameter to be 0
When drop.parms is omitted or dened as drop.parms=c(0,0), both
1
and
2
are constrained to zero and a completely random dropout (CRD) model
is assumed. Table 9.13 illustrates this by showing the coecients under y.d
and y.d-1 to be zero. Providing an initial estimate of
2
assumes a random
dropout (RD) model drop.parms=c(0,#) dependent on the previous obser-
vation. The output (Table 9.14) now displays the estimate of
2
. Similarly,
Table 9.13. Output from pcmid function assuming completely random dropout
(CRD).
Longitudinal Data Analysis Model
assuming random dropout based on 1 previous observations
Call: pcmid(formula=fact~group*time, vparms=c(0,120,0), correxp=1,
dropmodel=~1+dt*group,drop.cov.parms=c(0,-.5,-1,-.2,0,-.5,-.1,-.3),
drop.parms=c(0,0), reqmin = 1e-012)
Analysis Method: Maximum Likelihood(ML)
Correlation structure: exp(- phi*|u| ^ 1 )
Maximized log-likelihood: [1] -8079.097
Mean Parameters:
(Intercept) group time group:time
PARAMETER 63.875429 1.694745 -0.2662796 0.07277491
STD.ERROR 1.114316 1.344469 0.1108001 0.13124155
Variance Parameters:
nu.sq sigma.sq tau.sq phi
0 148.1602 120.2703 0
Dropout parameters:
(I) dt2 dt3 dt4 group dt2grp dt3grp dt4grp y.d y.d-1
0 -0.5493 -1.0046 -0.1978 0 -0.5916 0.0708 -0.2854 0 0
Maximization converged after 2688 iterations.
c2638 C2638TexturesC09 February 12, 2002 11:42
SELECTION MODEL FOR MONOTONE DROPOUT 209
Table 9.14. Output from pcmid function assuming random dropout (RD).
Longitudinal Data Analysis Model
assuming random dropout based on 1 previous observations
Call: pcmid(formula=fact~group*time,vparms=c(0,120,0),correxp=1,
dropmodel=~1+dt*group,drop.cov.parms=c(0,1.9,1.4,2.3,0,-.6,.1,-.3),
drop.parms = c(0, -0.04), reqmin = 1e-012)
Analysis Method: Maximum Likelihood (ML)
Correlation structure: exp(- phi *|u| ^ 1 )
Maximized log-likelihood: [1] -8039.064
Mean Parameters:
(Intercept) group time group:time
PARAMETER 63.875423 1.694822 -0.2662785 0.0727742
STD.ERROR 1.114316 1.344469 0.1108001 0.1312415
Variance Parameters:
nu.sq sigma.sq tau.sq phi
0 148.1605 120.2702 0
Dropout parameters:
(I) dt2 dt3 dt4 group dt2grp dt3grp dt4grp y.d y.d-1
0 1.872 1.377 2.286 0 -0.5736 0.1345 -0.2735 0 -0.0385
Maximization converged after 2779 iterations.
providing initial estimates of both
2
and
3
assumes a random dropout (RD)
model drop.parms=c(0,#) dependent on the two previous observations. Fi-
nally, including an initial estimate of
1
allows an informative dropout (ID)
model drop.parms=c(#,#).
The arguments dropmodel and drop.cov.parms dene the part of the
dropout model described by
0
. For example, the model labeled CRD1 in
the preceding example is specied as
> pcmid(..., dropmodel~1, drop.cov.parms=c(\#))
Model CRD2 includes the covariate group:
> pcmid(..., dropmodel~1+group, drop.cov.parms=c(\#,\#))
Model CRD3 allows a dierent dropout rate at each of the follow-up visits
and constrains the intercept to be zero. This is equivalent to using a CLASS
DT; statement and a NOINT option in the model statement when using SAS.
> pcmid(..., dropmodel~1+dt, drop.cov.parms=c(0,\#,\#,\#))
c2638 C2638TexturesC09 February 12, 2002 11:42
210 RANDOM-EFFECTS MIXTURE
Finally, the model CRD4 is specied as follows, with two parameters con-
strained to be zero to obtain a full rank model.
..., dropmodel~1+dt*group, drop.cov.parms=c(0,\#,\#,\#,0,\#,\#,\#))
Oswald warnings
The Oswald documentation states the need to double-check the convergence
of each model by changing the parameter estimates and modifying the con-
vergence criteria. This is absolutely critical, as the routine is very prone to
nd local maxima and print a message that the Maximization converged
.... For the CRD and RD models, one way to check is to ensure that the
HRQoL mean and variance parameters in each model are converging to the
same values. This does not guarantee that the maximum is found, but if they
dier, one or more of the models have not been maximized.
9.5 Advanced readings
Intermittent missing data
In addition to these monotone patterns of missing data due to dropout, non-
monotone patterns may exist. Troxel [151] describes a likelihood method with
Markovian correlation structure in which the missingness follows a logistic
model. It is an extension of the model proposed by Diggle and Kenward [36]
and utilizes numerical integration to solve the problem.
More selection models
Heckman probit stochastic dropout model
This model assumes that nonresponse occurs when response crosses a
threshold.
Pr(r
i2
= 0|y
i1
, y
i2
, ) = (
0
+
1
(y
i1
+
2
y
i2
)), (9.50)
where () is the normal cumulative distribution function and
0
,
1
,
2
are
unknown parameters. If
1
= 0, then the data are MCAR and non-ignorable
if both
1
= 0 and
2
= 0. Estimation relies heavily on normal assumptions.
Wu and Carroll
Wu and Carroll [168] propose a selection model that is applicable to growth
curve models with the individual slope parameters related to nonresponse for
later observations (right censoring) through a probit model.
Mori
Mori et al. [103] propose empirical Bayes estimates of the individual sub-
ject slopes that adjust for informative dropout. When a patient is censored
early, the time of censoring dominates the estimate of the change in HRQoL,
c2638 C2638TexturesC09 February 12, 2002 11:42
SUMMARY 211
whereas when dropout occurs later, the ordinary least-squares estimate of the
individuals slope dominates the estimate.
Nonparametric analyses
Yao et al. [170] describe an extension of a nonparametric analysis of lon-
gitudinal data with nonignorable dropout. The advantage of the proposed
method is that it can be applied to settings where the HRQoL assessments do
not follow the planned timing of the assessments and the distribution of the
QoL scores is non-normal. There is, however, a strong assumption made in
this extension. Specically, this approach assumes that the dierence in QoL
between the two treatment arms does not vary over time. This assumption
severely limits the settings where this method is useful.
9.6 Summary
All these models require the analyst to make strong assumptions.
These assumptions cannot be formally tested. Defense of the assumptions
must be made on a clinical basis rather than statistically.
Lack of evidence of nonignorable missing data for any particular model
does not prove that the missing data are ignorable.
Estimates are not robust to model misspecication; thus, more than one
model should be considered, which is a sensitivity analysis.
c2638 C2638TexturesC09 February 12, 2002 11:42
c2638 C2638TextureC10 February 12, 2002 11:45
CHAPTER 10
Summary Measures
10.1 Introduction
In most clinical trials, investigators assess HRQoL longitudinally over the
period of treatment and, in some trials, subsequent to treatment. Each as-
sessment involves multiple scales that measure the general and disease-specic
domains of HRQoL. For example, in the NSCLC study, there are three treat-
ment arms, four assessments, and ve subscales of the FACT-L. Thus, most
studies entail multiple HRQoL measures that are assessed periodically. As a
result, addressing the problem of multiple comparisons is one of the analytic
challenges in these trials [79]. Not only are there concerns about Type I errors,
but large numbers of statistical tests generally result in a confusing picture of
HRQoL that is dicult to interpret [31].
Addressing multiplicity of endpoints
There are three general approaches to reduce the impact of multiplicity:
1. A priori specication of a limited number of conrmatory tests
2. Multiple comparison procedures including alpha adjustments (e.g.,
Bonferroni) and closed multiple testing
3. The use of summary measures or statistics
In practice, a combination of focused hypotheses, summary measures, and
multiple comparison procedures is necessary in most clinical trails. In this
chapter, the computations of summary measures and statistics are presented
with details concerning how their derivation is aected by missing data.
Selected multiple comparison procedures are presented in the following
chapter.
Summary measures vs. summary statistics
The terms summary measures and summary statistics are often interchanged.
The following denitions are used in this chapter. A summary measure re-
duces the repeated observations on each individual to a single number (see
Table 10.1). The procedure is rst to summarize the data within an individual
subject by calculating a single value (summary measure) for each individual
and then to perform an analysis of the summary measures. For example, we
can estimate the rate of change experienced by each individual using ordinary
c2638 C2638TextureC10 February 12, 2002 11:45
214 SUMMARY MEASURES
Table 10.1. Comparison of summary measures and summary statistics for combining
information across time.
Method Summary measures (S
i
=

J
j=1
w
j
f(Y
ij
))
Strategy 1. Summarize data within ith patient (over J repeated
evaluations)
2. Univariate analysis of S
i
Advantage Easy to describe and interpret
Disadvantages Often dicult to develop strategy for handling every
missing value on an individual basis; some measures
are biased, depending on pattern and reasons
for missing data
Method Summary statistics (S
h
=

J
j=1
w
j
g(
hj
))
Strategy 1. Fit multivariate model, estimating means (or para-
meters) for repeated measures or mixed-eects
model
2. Compute summary statistic (generally linear comb-
ination of parameters)
3. Test hypothesis (H
0
: S
j
= S
j
or S
j
S
j
= 0)
Advantage Strategies for handling missing data are model based
Disadvantage Harder to describe procedure
least squares (OLS) and then apply a two-sample t-test to compare the esti-
mates in two treatment groups. Other examples of summary measures include
the average of the within-subject post-treatment values [31, 54, 97], last value
minus baseline [54], average rate of change over time (or slope) [97], maximum
value [31, 97, 117], area under the curve [20, 97], and time to reach a peak or
prespecied value [97, 117].
In contrast, a summary statistic reduces the measurements on a group of
individuals to a single value. Data are initially analyzed using multivariate
techniques for longitudinal models and summary statistics are constructed
from the estimates of the parameters. For example, the mean rate of change
(or slope) for each treatment group is estimated using a mixed-eects model
and the dierences in the slopes between the two treatment groups are tested.
Other examples include the average of the group means during the post-
treatment period and area under the mean HRQoL vs. time curve.
Missing data are dealt with dierently in the construction of summary mea-
sures and statistics [39]. For summary measures (see Section 10.3), we must
develop a procedure to handle missing data at the subject level, possibly by
using interpolation and extrapolation or by imputing missing observations.
For summary statistics (see Section 10.4), missing data handled by the selec-
tion of the analytic model is described in Chapters 5, 8, and 9.
c2638 C2638TextureC10 February 12, 2002 11:45
CHOOSING A SUMMARY MEASURE 215
Strengths and weaknesses
Easier interpretation
There are several motivations for the use of summary measures in the analysis
of a longitudinal study of HRQoL [39]. The primary advantage of summary
measures is that they facilitate interpretation. Not only is the number of
comparisons reduced, but measures such as the rate of change and the area
under the curve are familiar concepts in clinical medicine.
Increased power
Summary measures and statistics also have greater power to detect small
but consistent dierences that may occur over extended periods of time or
multiple domains of HRQoL. In contrast to the Bonferroni correction or a
general multivariate test (Hotellings T), these summary statistics are more
sensitive to the settings where there are consistent dierences over time. To
illustrate, consider the two hypothetical examples displayed in Figure 10.1. In
the rst example (Figure 10.1A), the measure of HRQoL is consistently bet-
ter in one treatment during all four post-baseline assessments. In the second
example (Figure 10.1B), the second treatment has a negative impact (tox-
icity) 1 month postdiagnosis, but this dierence almost disappears by the
third month and begins to reverse by the ninth month. Although the dier-
ences between the groups in both examples are of clinical interest, in most
clinical trials one would wish to have test procedures that are more sensitive
to (or have greater power to detect) the consistent dierences displayed in
Figure 10.1A.
Weakness
The primary weakness of summary measures or statistics is that dierences in
specic domains of HRQoL or transient dierences at specic points in time
may be obscured. Investigations that are conrmatory, designed to test the
question of dierences in overall HRQoL, may benet from the use of sum-
mary measures. This is especially useful in studies where early toxicity in one
treatment group may be balanced by later improvements in disease control. In
contrast, studies that have more exploratory objectives of identifying which
aspects of HRQoL are impacted by the disease or a particular therapy may
not benet from the use of summary measures or statistics.
10.2 Choosing a summary measure
The selection of a summary measure as the endpoint in a clinical trial depends
on the objective of the investigation. The summary measure should have a
clear clinical relevance and ideally should be determined prior to data collec-
tion [22, 97]. Posing the research question simply (e.g.,How do the groups dif-
fer?) is too vague to indicate the correct measure. The process of identifying
the appropriate summary measure often requires the investigators to decide
c2638 C2638TextureC10 February 12, 2002 11:45
216 SUMMARY MEASURES
Figure 10.1. Hypothetical studies with (A) consistent dierences in HRQoL over
time or (B) a large dierence at a single point in time. Summary measure is the mean
of the post-baseline assessments. Means are joined by horizontal line. Vertical lines
indicate 1 standard error. F-test indicates multivariate test of group dierences at
assessments 2, 3, 4, and 5. t-tests indicate test of group dierence at that assessment
or for the summary measure.
c2638 C2638TextureC10 February 12, 2002 11:45
CONSTRUCTING SUMMARY MEASURES 217
what specic questions the HRQoL data are expected to answer. For example,
Curren et al. [22] describe a hypothetical trial where the objective is to reduce
toxicity and maintain acceptable HRQoL. In this setting one might want to
study the worst (minimum) HRQoL score that occurred for each individual.
The selection also depends on the expected pattern of change across time
and patterns of missing data. Consider several possible patterns of change in
HRQoL across time (Figure 10.2). One prole is a steady rate of change over
time, reecting either a constant decline in HRQoL (Figure 10.2A) or a con-
stant improvement (Figure 10.2B). The rst pattern is typical of patients with
progressive disease where standard therapy is palliative rather than curative.
This is the pattern observed in the NSCLC study (Example 2). This pattern
of change suggests that the rate of change or slope is a possible choice of a
summary measure. In contrast, measures such as the change from baseline to
the last measure might not be desirable if patients who fail earlier and thus
drop out of the study earlier have smaller changes than those patients with
longer follow-up.
An alternative prole is an initial rapid change with a subsequent plateau
after the maximum therapeutic benet is realized (Figure 10.2D). This might
occur for therapies where the dose needs to be increased slowly over time or
where there is a lag between the time when therapy is initiated and the time
when maximal benet is achieved. This prole illustrates the importance
of identifying the clinically relevant question a priori. If the objective is to
identify the therapy that produces the most rapid improvement in HRQoL,
the time to reach a peak or prespecied value is good choice. If, in contrast,
the ultimate level of benet is more important than the time to achieve the
benet, then a measure such as the post-treatment mean or mean change
relative to baseline is desirable. More than one summary measure can be con-
structed to investigate dierent questions. It is rare that more than two are
clinically relevant.
A third pattern of change could occur with a therapy that has transient ben-
ets or toxicity (Figures 10.2E and F). For example, individuals may expe-
rience transient benets and then return to their baseline levels after the
eect of the therapy has ceased. Alternatively, a toxic therapy for cancer may
signicantly reduce HRQoL during therapy but ultimately result in a better
HRQoL following therapy than the patient was experiencing at the time of dia-
gnosis [48, 86]. For these more complex patterns of change over time, a measure
such as the area under the curve might be considered as a summary of both
early and continued eects of the therapy.
10.3 Constructing summary measures
The approach for constructing summary measures is to reduce the repeated
measures from each individual to a single value. This approach is popular be-
cause simple univariate tests (e.g., t-tests) can be used for the analysis. For ex-
ample, one might compute the slope of the scores observed for each subject and
test the hypothesis that the slopes diered among the experimental groups.
c2638 C2638TextureC10 February 12, 2002 11:45
218 SUMMARY MEASURES
Figure 10.2. Patterns of change in HRQoL over time. (From Fairclough, D.L., in
Quality of Life Assessment in Clinical Trials: Methods and Practice, Staquet, M.J.
et al., Eds., Oxford University Press, New York, 1998, Chap. 13. With permission.)
c2638 C2638TextureC10 February 12, 2002 11:45
CONSTRUCTING SUMMARY MEASURES 219
Figure 10.2. Continued
c2638 C2638TextureC10 February 12, 2002 11:45
220 SUMMARY MEASURES
When the period of observation varies widely among subjects or data col-
lection stops for some reason related to the outcome, this construction of
summary measures is challenging. Easy xes without careful thought only
hide the problem. Many of the procedures for constructing summary mea-
sures assume data are missing completely at random (MCAR) [110] and are
inappropriate in studies of HRQoL.
Notation
The construction of a summary measure that reduces the set of J measure-
ments (Y
ij
) on the ith individual to a single value (S
i
) can be described as a
weighted sum of the measurements (Y
ij
) or a function of the measurements
(f(Y
ij
)). The general form is
S
i
=
J

j=1
w
j
f(Y
ij
). (10.1)
Table 10.2 illustrates how this applies to a number of examples. Although
in some cases, such as the average rate of change (slope), there are easier
ways of constructing the summary measures, the table illustrates the common
structure of all summary measures [39].
Missing data
The calculation of summary measures is straightforward in the absence of miss-
ing and mistimed observations. It becomes much more challenging when there
are missing data or the length of follow-up diers by individual. Strategies for
handling missing data often need to be developed on a case-by-case basis.
All of the following approaches make assumptions that may or may not be
Table 10.2. Examples of summary measures constructed within individuals across
time: S
i
=

n
j=1
w
j
f(Y
ij
).
Description f(Y
ij
) Weights (w
j
)
Average rate of change Y
ij


Y
i
(t
i

t)/(

t
i

t )
2
Mean of r follow-up Y
ij
0, . . . ,
1
r
, . . . ,
1
r
observations
Mean of r follow-up, Y
ij
1, . . . ,
1
r
, . . . ,
1
r
baseline
Area under the curve Y
ij
t
2
t
1
2
, . . . ,
t
j+1
t
j1
2
, . . . ,
t
n
t
n1
2
Average of ranks Rank(Y
ij
) 1/n, . . . , 1/n
Minimum value Min(Y
i
) 1
c2638 C2638TextureC10 February 12, 2002 11:45
CONSTRUCTING SUMMARY MEASURES 221
Table 10.3. Possible strategies for handling missing data in the calculation of sum-
mary measures.
Problem Solution Assumptions/limitations
Intermittent Interpolation Change approximately linear over
missing period; t
ij1
and t
ij+1
measured
values under similar conditions
Mistimed last Estimate Y
iT
Y
iT
=
(T t
iJ1
)Y
iJ
+ (T t
iJ
)Y
iJ1
t
iJ
t
iJ1
observation
(t
iJ
= T)
Incomplete Extrapolation Potential bias favoring early
follow-up or LVCF dropout
Death Minimum Potential bias favoring early death
value
Zero Mimics survival analysis
reasonable in specic settings, and it is advisable to examine the sensitivity
of the conclusions to the various assumptions (Table 10.3).
For example, if intermediate observations are missing, one could interpolate
between observations. This approach assumes that these intermediate missing
values are random and that the previous and following measures are taken un-
der similar conditions. Interpolation is reasonable when the expected pattern
of change over time is roughly linear (see Figures 10.2A and B) or constant.
Interpolation would not make sense in the design used for the adjuvant breast
cancer study (Example 1) for missing assessments during therapy. Inter-
polation between the pre- and post-therapy assessments would obviously over-
estimate the HRQoL of subjects during therapy. Also, interpolation during
the early phase of the renal cell carcinoma study (Example 3) would not
make sense given the rapid drop in HRQoL followed by recovery. Even in the
NSCLC study, interpolation assumes that the missingness is not the result of
some event, such as an episode of acute toxicity, that would cause the patient
to miss the assessment and have poorer HRQoL.
Dropout and other causes of dierences in the length of follow-up among
individuals are another challenge. For a patient who dropped out, the last
measurement could be imputed by (1) carrying the last value forward, (2) ex-
trapolating from the last two observations, or (3) implementing one of the
imputation techniques described in Chapters 6 and 7. The most popular ap-
proach is to carry the value of the last observation forward. Unfortunately, this
strategy is often implemented without careful thought. This is a reasonable
strategy if the expected pattern of change within individuals is either a steady
improvement (Figure 10.2B) or improvement with a plateau (Figure 10.2D).
However, with other patterns of change (Figures 10.2A, C, and E), this strat-
egy would bias the results in favor of interventions with early dropout.
c2638 C2638TextureC10 February 12, 2002 11:45
222 SUMMARY MEASURES
Assigning zero is a valid approach for HRQoL scores that are explicitly
anchored at zero for the health state of death. These are generally scores
measured using time trade-o (TTO) or standard gamble (SG) techniques to
produce utility measures. However, the majority of HRQoL instruments are
developed to maximize discrimination among patients. In these instruments a
value of zero would correspond to the worst possible outcome on every ques-
tion. Assigning zero also has some statistical implications. If the proportion
of deaths is substantial, the observations may mimic a binomial distribution
and the results approximate a KaplanMeier analysis of survival. When the
number of deaths is small, the zero values will violate the multivariate normal
distribution assumptions of the analysis methods.
Average rate of change (slopes)
When the changes over time are approximately linear, the average rate of
change (or slope) may provide an excellent summary of the eect of an inter-
vention. When there is virtually no dropout during the study or the dropout
occurs only during the later part of the study, it is feasible to t a simple re-
gression model to the available data on each individual. This is often referred
to as the ordinary least-squares slope (OLS slope).

i2
is the estimated slope
from a simple linear regression:

OLS
i
=
J

(X
ij


X
i
)(Y
ij


Y
i
)
(X
ij


X
i
)
2
= (X

i
X
i
)
1
X

i
Y
i
. (10.2)
NSCLC example
At rst glance, estimating individual slopes might seem to be an approach
for the NSCLC study. But recall that a fourth of the subjects had only a
single HRQoL assessment and would contribute nothing to the analysis. Fur-
ther, less than half of the subjects had three or more observations, so that
a substantial proportion of the subjects is likely to have unstable estimates
of their slopes. Finally, there is a suspicion that the data are MNAR. As a
practical design issue, it is advisable to plan frequent assessments during the
early phase of a trial if one is intending to use this approach and dropout is
likely.
Missing data
Obviously, slopes for each individual can be estimated if all subjects have two
or more observations. However, the estimates of the slopes will have a large
associated error when the available observations span a short period of time
relative to the entire length of the study. The wide variation in slopes esti-
mated with OLS is displayed in Figure 10.3; note the wide range of estimates
for subjects with only two measures. If there is a substantial proportion of
subjects with only one or two observations, it may be necessary to use either
c2638 C2638TextureC10 February 12, 2002 11:45
CONSTRUCTING SUMMARY MEASURES 223
(a)
(b)
Figure 10.3. Distribution of summary measures using dierent methods of estimat-
ing the slope. (a) Ordinary least squares (OLS); (b) empirical Bayes (EB); (c) last
value carried forward (LVCF), (d) multiple imputation (MI).
imputed values for later missing values or the empirical Bayes (EB) estimates
from a mixed-eects model. Because this second estimate also uses information
from other individuals, it is more stable, especially with highly unbalanced
longitudinal studies where some individuals have only a few observations.
The EB slopes display shrinkage toward the average slope, as the estimator
is a weighted average of the observed data and the overall average slope [82]
(see Section 5.6). For individuals with only a few observations, the EB slope is
c2638 C2638TextureC10 February 12, 2002 11:45
224 SUMMARY MEASURES
(c)
(d)
Figure 10.3. Continued
very close to the overall average. This is illustrated by the narrow distribution
of estimates displayed in Figure 10.3. Note that both approaches assume the
data are missing at random (MAR). The distribution of the slopes derived us-
ing LVCF illustrates the tendency of these values to center around zero. This
is by denition for those who drop out after the rst observation. Finally, the
distribution of slopes is displayed from the multiply imputed data. Values are
more widely distributed than for either the EB or LVCF strategies. Clearly,
the distribution of the summary measures can be sensitive to the method used
to handle the missing data.
c2638 C2638TextureC10 February 12, 2002 11:45
CONSTRUCTING SUMMARY MEASURES 225
Figure 10.4. Calculation of the AUC using a trapezoidal approximation.
Area under the curve
Another measure summarizing HRQoL over time is the area under the curve
(AUC) of the HRQoL scores (Y
ij
) plotted against time (t) [19, 94]. The AUC
for the ith individual can be estimated using a trapezoidal approximation
from the observed scores (Figure 10.4). The area of each trapezoid is equal to
the product of the height at the midpoint ((Y
ij
+ Y
i(j1)
)/2) and the width
of the base (t
j
t
j1
). The total area is calculated by adding areas of a series
of trapezoids
S
i
= AUC
i
=
J

j=2
(t
j
t
j1
)
Y
ij
+Y
i(j1)
2
. (10.3)
With a little algebraic manipulation, this calculation can also be expressed as
a weighted function of the HRQoL scores, where the weights are determined
by the spacing of the assessments over time:
AUC
i
=
t
2
t
1
2
Y
i1
+
J1

j=2
t
j+1
t
j1
2
Y
ij
+
t
J
t
J1
2
Y
iJ
. (10.4)
Missing data
There are a number of options for constructing a measure of the AUC in the
presence of missing data (see Table 10.3). All have some limitations. The best
choice will depend on the objectives and the reasons for missing assessments
or dropout.
c2638 C2638TextureC10 February 12, 2002 11:45
226 SUMMARY MEASURES
Figure 10.5. Contrast of two subjects with dierent lengths of follow-up. Both
subjects have the same AUC when adjusted by each individuals length of follow-up.
Fayers and Machin [45] describe one alternative, the average AUC, which
may be useful in some settings where the length of follow-up varies among
subjects. This summary measure is calculated by dividing the total AUC
by the total time of observation. S
i
= AUC
i
/T
i
, where T
i
is the total time of
observation for the ith individual. If the research question concerns the average
HRQoL of patients during a limited period such as the duration of a specic
therapy, this may be a reasonable approach. However, if early termination due
to toxicity or lack of ecacy results in similar or higher summary scores, then
the results are biased. This is illustrated in Figure 10.5, where the average
c2638 C2638TextureC10 February 12, 2002 11:45
CONSTRUCTING SUMMARY MEASURES 227
areas are the same for the two curves but the scores decline more rapidly in
the subject who terminated the study early. As always, the choice between
these two approaches depends primarily on the research question.
The issue of selecting a strategy for computing the AUC also occurs when
patients die during the study. One strategy is to extrapolate the curve to
zero at the time of death. Other proposed strategies include assigning values
of (1) the minimum HRQoL score for that individual or (2) the minimum
HRQoL score for all individuals [68]. One strategy will not work for all studies.
Whichever strategy is chosen, it needs to be justied and the sensitivity of
the results to the assumptions examined.
Dierences at baseline
A practical problem with the use of the AUC or the mean of post-baseline
measures as a summary measure occurs when the baseline HRQoL scores
dier among groups. If one group contains more individuals who score their
HRQoL consistently higher than other individuals, small (possibly statistically
nonsignicant) dierences are magnied over time in the summary measure.
One possible solution is to calculate the AUC relative to the baseline score.
AUC

i
= AUC
i
Y
i1
(t
J
t
1
)
=
_
t
2
t
1
2
(t
J
t
1
)
_
Y
i1
+
J1

j=2
t
j+1
t
j1
2
Y
ij
+
t
J
t
J1
2
Y
iJ
.
(10.5)
Average of ranks
Another method proposed for combining information across repeated mea-
sures uses ranks [108]. The measurements on all subjects are ranked at each
time point and then the average of the ranks across the n time points is com-
puted for each individual. In the notation of Equation 10.1 and Table 10.2,
f(Y
ij
) is the rank of Y
ij
among all values of the jth measure and S
i
is the
average of the ranks for the ith individual (w
j
= 1/n). When the data are
complete, this procedure is straightforward.
Missing data
If the reasons for missing data were known and one could make reasonable
assumptions about the ranking of HRQoL in patients at each time point, one
could possibly adapt this approach to a study with missing data. For example,
it would not be unreasonable to assume that the HRQoL of patients who died
or left the study due to excessive toxicity was worse than the HRQoL of those
patients who remained on therapy. They would then be assigned the lowest
possible rank for measurements scheduled after death or during the time of
excessive toxicity. Other strategies that might be considered are discussed in
Chapter 6 (see Table 6.7).
c2638 C2638TextureC10 February 12, 2002 11:45
228 SUMMARY MEASURES
Univariate analysis of summary measures
After constructing the summary measures, statistical tests (e.g., two-sample
t-test, Wilcoxon rank sum test) are used to compare treatment groups. In most
cases these tests assume that the observations are independent and identically
distributed (i.i.d). The assumption of homogeneity of variance of the summary
measures is often violated when the duration of follow-up is vastly dierent
among subjects. For example, slopes estimated for individuals who drop out
of the study early will have greater variation than slopes for those who com-
plete the study. One solution is to weight the slopes using their standard
errors, giving greater weight to those individuals with the longest follow-up.
Unfortunately, unless dropout is MCAR, this will bias the analysis. The alter-
native procedure is to ignore the possible heterogeneity, giving equal weight
to all subjects. This later approach is generally more appropriate in studies
of HRQoL where the MCAR assumption is rarely appropriate.
Stratied analysis of summary measures
Dawson and Lagokos [2729] propose a strategy for handling monotone
dropout. First, we identify strata (g) dened by the dropout pattern. Sum-
mary measures are then calculated for each individual within the strata using
just the observed data. This avoids the need to extrapolate past the point of
dropout. Then a standardized test statistic, Z
g
, for the two-arm comparison
is calculated for each stratum. Dawson and Lagokos [2729] use the nonpara-
metric van der Waerden scores in their applications. These stratum-specic
test statistics are then combined into an overall test statistic Z =

w
g
Z
g
/
_

w
2
g
. Dawson uses w
g
=
_
j
g
n
g1
n
g2
/(n
g1
+n
g2
) as a general set of weights
[29] and w
g
=
_

(x
gj
x
g
)
2
n
g1
n
g2
/(n
g1
+n
g2
) when the summary mea-
sures are slopes [28].* Note that these weights will tend to increase the con-
tribution of strata with longer follow-up.
When the data are missing completely at random, this approach will in-
crease the power, since the across-strata variation is removed [28, 29]. How-
ever, it is unclear how it will perform when dropout is nonignorable and the
pattern of dropout diers by treatment group [22]. For example, there may not
be any treatment dierence in the summary statistic conditional on dropout:
All Z
g
0 Z 0. However, if the dropout patterns dier and the summary
measures dier across the strata, then there are real dierences that are not
detected. To illustrate, consider a hypothetical example (Table 10.4) where
there is extensive early dropout in one arm of the study. Higher scores indicate
an improvement in HRQoL. Among those who drop out early (g = 1), the
average summary measure score is 5, indicating a decline in HRQoL regard-
less of treatment arm in these subjects. Similarly, among those who remain
on the study (g = 2), the summary measures average 20 in both groups. If we
* n
gh
is the number of subjects in the gth stratum and the hth treatment arm. j
g
is
the number of observations used to compute the summary measure in the gth group. x
gj
denotes the times when subjects in the gth group contribute observations.
c2638 C2638TextureC10 February 12, 2002 11:45
CONSTRUCTING SUMMARY MEASURES 229
Table 10.4. Hypothetical example illustrating problem with stratied analysis of
summary measures when dropout rates dier across groups.
Stratum (g) Treatment Subjects (n
gh
)

S
h(g)
Z
g
1 A 10 5 0
B 40 5
2 A 40 20 0
B 10 20
Table 10.5. Actual weights (w
j
) and relative weights for the NSCLC example. |w
1
|
is constrained to be 1 for relative weights.
Measure T1 T2 T3 T4
Actual weights
Slope 0.046 0.020 0.007 0.059
Mean dierence 1.000 0.333 0.333 0.333
AUC 0.125 0.250 0.375 0.250
Relative weights
Slope 1.000 0.429 0.143 1.286
Mean dierence 1.000 0.333 0.333 0.333
AUC 1.000 2.000 3.000 2.000
test for dierences within each stratum, we observe none, so the standardized
test statistics are both approximately 0 (Z
g
0). However, it is clear that in
arm A there are more individuals who remained on the study and experi-
enced a greater benet than on arm B. This example is highly exaggerated,
but it illustrates the problem. One could imagine a similar, but less dramatic,
scenario on a placebo-controlled trial where there was early dropout among
individuals who did not perceive any benet from the intervention.
NSCLC example
In this example, all ve subscales of the FACT-Lung instrument will be
analyzed. Three summary measures are examined: the average change (slope),
the mean of the post-randomization assessments minus baseline, and the AUC.
The weights for the three summary measures are displayed in Table 10.5. Note
that the weights for the slope and the mean dierence are similar, weight-
ing early scores negatively and later scores positively. The biggest dierence
among the three summary measures is the much greater weights for the last
assessment when the summary measure is the slope. As approximately
2
3
of
the subjects are missing this assessment, the slope is most sensitive to the
manner in which missing values are handled.
To illustrate the sensitivity of the summary measures to the method of
handling missing values, two approaches were used. In the rst, missing values
c2638 C2638TextureC10 February 12, 2002 11:45
230 SUMMARY MEASURES
(a)
(b)
Figure 10.6. Mean dierence between postrandomization and prerandomization
assessments with missing data handled by multiple imputation (MI MCMC) and
LVCF. Pairs display mean 1 standard error for no Taxol (a) and Taxol (b) groups.
were handled by multiple imputation using the MCMC approach for multi-
variate normal data described in Chapter 7. All ve subscales were jointly
modeled with age, log survival, best response, and early progressive disease. In
the second, the last value was carried forward (LVCF). The patterns across the
ve subscales were very similar for all three summary measures. The estimates
and their standard errors for the mean dierence between postrandomization
and baseline are displayed in Figure 10.6.
c2638 C2638TextureC10 February 12, 2002 11:45
SUMMARY STATISTICS ACROSS TIME 231
10.4 Summary statistics across time
In contrast to summary measures, summary statistics are constructed by com-
bining group measures over time. Examples of these group measures are the
parameter estimates from growth curve models or means from repeated mea-
sures models. In practice, the use of summary statistics rather than summary
measures is often preferable because it may be much easier to develop a
model-based method for handling missing data than to develop strategies
for imputing each individual missing value required to compute the summary
measures.
Notation
The general procedure for the construction of summary statistics is to obtain
parameter estimates for the hth group (

hj
) and then reduce the set of J
estimates to a single summary statistic:

S
h
=
J

j=1
w
j
g(

hj
). (10.6)
In a repeated measures design, g(

hj
) is generally a vector of the estimates of
the means over time (
hj
) adjusted for important covariates [39]. In a growth
curve model, g(

hj
) contains the parameter estimates of the polynomial or
piecewise regression model. g(

hj
) may also be a direct estimate of a summary
statistic such as the slope or a test statistic such as the ratio of the estimate
to its standard error (t
hj
).
If the data are complete, one can use any multivariate analysis method to
estimate
hj
. When data are missing, one must use an analytic method that is
appropriate for the missing data mechanism as described in Chapter 5, 8, or 9.
Area under the curve
Repeated measures
When the model has been structured as a repeated measures analysis, the
AUC can be estimated for each of the H groups by using a trapezoidal
approximation.
AUC
h
(t
J
) =

S
h
=
J

j=2
(t
j
t
j1
)

hj
+
h(j1)
2
. (10.7)
The equation can be rewritten as a weighted function of the means:
AUC
h
(t
J
) =

S
h
=
t
2
t
1
2

h1
+
J1

j=2
t
j+1
t
j1
2

hj
+
t
J
t
J1
2

hJ
. (10.8)
For example, in the NSCLC study, the planned times of the assessments
were 0, 6, 12, and 26 weeks. The AUC over 26 weeks for the hth group is
c2638 C2638TextureC10 February 12, 2002 11:45
232 SUMMARY MEASURES
Table 10.6. Examples of summary statistics:

S
h
=

p
j=1
w
j
g(

hj
).
Description g(

hj
) Weights (w
j
)
Mean of r follow-up
hj
0, . . . ,
1
r
, . . . ,
1
r
observations
Mean of follow-up
hj
1, . . . ,
1
r
, . . . ,
1
r
rst observation
Average rate of change

h2
(slope)
Area under the curve
hj
t
2
t
1
2
, . . . ,
t
j+1
t
j1
2
, . . . ,
t
J
t
J1
2
Inverse correlation
hj
/
(
hj
)
J

R
1
J = (1, . . . , 1)

, R = correlation of Y
i
= (Y
i1
, . . . , Y
in
)

.
AUC
h
(26) =
t
2
t
1
2

h1
+
t
3
t
1
2

h2
+
t
4
t
2
2

h3
+
t
4
t
3
2

h4
=
6 0
2

h1
+
12 0
2

h2
+
26 6
2

h3
+
26 12
2

h4
.
As discussed before, dividing this quantity by t
J
or scaling time so that t
J
= 1
has two advantages. First, the statistic is independent of the units of time. Sec-
ond, the statistic is in the range of values for the scale and has the alternative
interpretation of the average score over the period of interest.
Growth curve models
When the model has been structured as a growth curve model, the AUC is
computed by integration from 0 to T. If a polynomial model has been used
to describe the change in HRQoL over time:
AUC
h
(T) =

S
h
=
_
T
t=0
J

j=0

hj
t
j
t =
J

j=0
T
j+1
j + 1

hj
, (10.9)
where

h0
is the estimate of the intercept for the hth group,

h1
is the linear
coecient for the hth group, etc.
For example, in the renal cell carcinoma study, the AUC is
AUC
h
(T) =
J

j=0
T
j+1
j + 1

hj
=
T
1
1

h0
+
T
2
2

h1
+
T
3
3

h2
+
T
4
4

h3
+
T
5
5

h4
+
T
6
6

h5
.
c2638 C2638TextureC10 February 12, 2002 11:45
SUMMARY STATISTICS ACROSS TIME 233
If we are interested in the AUC during the initial period of therapy, dened
as the rst 8 weeks or 1.84 months, the AUC is
AUC
h
=
1.84
1
1
. .
1.84

h0
+
1.84
2
2
. .
1.6925

h1
+
1.84
3
3
. .
2.076

h2
+
1.84
4
4
. .
2.86

h3
+
1.84
5
5
. .
4.22

h4
+
1.84
6
6
. .
6.46

h5
.
The estimates and their dierences for 8 weeks are generated by the
ESTIMATE statement as illustrated below:
ESTIMATE IFN only - AUC 8wks
TREAT 1.840 0
TREAT*MONTHS 1.6925 0
TREAT*MONTHS*MONTHS 2.076 0
TREAT*MONTHS*MONTHS*MONTHS 2.86 0
TREAT*MONTHS*MONTHS*MONTHS*MONTHS 4.22 0
TREAT*MONTHS*MONTHS*MONTHS*MONTHS*MONTHS 6.46
/DIVISOR=1.84;
ESTIMATE IFN +CRA - AUC 8wks
TREAT 0 1.840
TREAT*MONTHS 0 1.6925
TREAT*MONTHS*MONTHS 0 2.076
TREAT*MONTHS*MONTHS*MONTHS 0 2.86
TREAT*MONTHS*MONTHS*MONTHS*MONTHS 0 4.22
TREAT*MONTHS*MONTHS*MONTHS*MONTHS*MONTHS 0 6.46
/DIVISOR=1.84;
ESTIMATE Difference - AUC 8wks
TREAT -1.840 1.840
TREAT*MONTHS -1.6925 1.6925
TREAT*MONTHS*MONTHS -2.076 2.076
TREAT*MONTHS*MONTHS*MONTHS -2.86 2.86
TREAT*MONTHS*MONTHS*MONTHS*MONTHS -4.22 4.22
TREAT*MONTHS*MONTHS*MONTHS*MONTHS*MONTHS -6.46 6.46
/DIVISOR=1.84;
Estimates for longer periods of time, such as 26 and 52 weeks, are generated
using the same procedure. In this study, the results would appear as follows
where T is rescaled to be equal to 1. The estimates are now on the same scale
as the original measure and have the alternative interpretation of the average
score over the period from 0 to T.
Standard
Label Estimate Error DF t Value Pr > |t|
IFN only - AUC 8wks 61.97 1.63 468 38.04 <.0001
IFN +CRA - AUC 8wks 61.30 1.65 468 37.16 <.0001
Difference - AUC 8wks -0.67 2.32 468 -0.29 0.7739
c2638 C2638TextureC10 February 12, 2002 11:45
234 SUMMARY MEASURES
IFN only - AUC 26wks 64.35 1.66 468 38.65 <.0001
IFN +CRA - AUC 26wks 59.30 1.71 468 34.64 <.0001
Difference - AUC 26wks -5.05 2.39 468 -2.11 0.0350
IFN only - AUC 52wks 64.73 1.75 468 36.94 <.0001
IFN +CRA - AUC 52wks 61.27 1.82 468 33.61 <.0001
Difference - AUC 52wks -3.45 2.53 468 -1.37 0.1726
For a piecewise regression model, one could estimate the means at each knot
and then use a trapezoidal approximation. Integration provides a more direct
method:
AUC
h
(T) =

S
h
=
_
T
t=0
J

j=0

hj
t
[j]
t
=

h0
T +
J

j=1

hj
max(T T
[j]
, 0)
2
2
, (10.10)
where

h0
is the estimate of the intercept for the hth group,

h1
is the initial
slope, and

hj
, j > 2 is the change in slope at each of the knots, T
[l]
. t
[0]
= 1,
t
[1]
= t, and t
[j]
= max(t T
[j]
).
In the renal cell carcinoma study, the piecewise regression model has two
knots, one at 2 weeks and one at 8 weeks. Thus, the estimated AUC (rescaled
to T = 1) is
AUC
h
(T) =
_

h0
T +

h1
T
2
2
+

h2
max(T 2, 0)
2
2
+

h3
max(T 8, 0)
2
2
__
T.
If we are interested in the early period dened as the rst 8 weeks, then
AUC
h
(8) =
_

h0
8 +

h1
8
2
2
..
=32
+

h2
max(8 2, 0)
2
2
. .
=18
+

h3
max(8 8, 0)
2
2
. .
=0
__
8.
The ESTIMATE statements follow directly from these equations, as illustrated
below for the calculation of the AUC at 8 weeks for each group and their
dierence.
ESTIMATE IFN only - AUC 8wks TREAT 8 0
TREAT*WEEKS 32 0
TREAT*WEEK2 18 0/DIVISOR=8;
ESTIMATE IFN +CRA - AUC 8wks TREAT 0 8
TREAT*WEEKS 0 32
TREAT*WEEK2 0 18/DIVISOR=8;
ESTIMATE Difference - AUC 8wks TREAT -8 8
TREAT*WEEKS -32 32
TREAT*WEEK2 -18 18/DIVISOR=8;
Estimates for longer periods of time, such as 26 and 52 weeks, are generated
using the same procedure. The results would appear as follows:
c2638 C2638TextureC10 February 12, 2002 11:45
SUMMARIZING ACROSS HRQoL DOMAINS OR SUBSCALES 235
Standard
Label Estimate Error DF t Value Pr>|t|
IFN only - AUC 8wks 63.55 1.54 472 41.30 <.0001
IFN +CRA - AUC 8wks 61.18 1.58 472 38.77 <.0001
Difference - AUC 8wks -2.37 2.20 472 -1.08 0.2822
IFN only - AUC 26wks 63.16 1.57 472 40.12 <.0001
IFN +CRA - AUC 26wks 59.51 1.61 472 36.91 <.0001
Difference - AUC 26wks -3.66 2.25 472 -1.62 0.1051
IFN only - AUC 52wks 64.44 1.72 472 37.36 <.0001
IFN +CRA - AUC 52wks 61.88 1.71 472 36.16 <.0001
Difference - AUC 52wks -2.70 2.43 472 -1.11 0.2668
The interpretation is that there is, on average, a 2- to 4-point dierence be-
tween the two groups. Although there are consistent dierences between the
two groups, the magnitude is small. The standard deviation of the FACT-
BRM TOI at baseline is approximately 17 points, and dierences between 2
and 4 points are not clinically signicant.
For the purpose of illustration, we have assumed the data are MAR and
estimated the parameters using REML. However, because we suspect that the
data are MNAR, these summary measures should be used in the sensitivity
analysis as described in Chapters 8 and 9.
10.5 Summarizing across HRQoL domains or subscales
Summarizing across HRQoL domains or subscales is useful when the objec-
tive is to obtain increased power to test whether a hypothesis of interest is
consistently dierent across scales. However, a change in one domain may
be obscured by changes in the opposite direction in another. Alternative
methods such as multiple testing procedures (Chapter 11) should be con-
sidered when dierences among the individual HRQoL domains are of inter-
est. The same two approaches can be used to compute summary measures
or statistics across domains or subscales as were used to summarize across
time.
Summary measures
Most HRQoL instruments create summary measures. Actually, each subscale
is inherently a summary measure constructed by combining the responses to
individual items. However, for most instruments, summary measures are also
constructed from the subscales. Some options for weighting include
1. Weights proportional to the number of questions for each domain
2. Equal weights for each domain
3. Factor analytic weights
c2638 C2638TextureC10 February 12, 2002 11:45
236 SUMMARY MEASURES
4. Weights derived from patient preferences
5. Statistically derived weights (e.g., correlation among domains)
Weighting proportional to the number of questions
Many HRQoL instruments compute an overall score where domains are
weighted in proportion to the number of questions asked about each domain.
In the FACT-Lung and FACT-BRM instruments, there are a number of sum-
mary measures. For both, a total score is computed by summing all subscales
(Table 10.7). Another summary measure, the Trial Outcome Index, is also
constructed by summing the physical well-being, functional well-being, and
lung cancerspecic items to create the FACT-Lung TOI and by summing the
physical well-being, functional well-being, and BRM-specic items to create
the FACT-BRM TOI. There is an implicit weighting of the subscales that is
dependent on the number of items. For example, because there are only ve
questions assessing emotional well-being, this domain receives less weight than
the other domains of physical, functional, and social well-being (Table 10.7).
Similarly, because there are 13 BRM-specic questions and 7 lung cancer
specic questions, the physical and functional well-being subscales have less
weight in the FACT-BRM TOI than they do in the FACT-Lung TOI.
Factor analytic weights
The physical component scale (PCS) and the mental component scale (MCS)
of the Medical Outcome Study SF-36 are examples of summary scores where
weights are derived from factor analysis. These widely used scores present an
excellent example of summary scores that need to be carefully examined when
interpreting the results of a trial. The weights used to construct the summary
Table 10.7. Summary measures from the FACT-Lung and FACT-BRM Version 2
instruments.
Total score TOI
Subscale Items Lung BRM Lung BRM
Physical well-being 7 0.200 0.171 0.333 0.259
Social/family well-being 7 0.200 0.171
Relationship with doctor 2
a
0.057 0.049
Emotional well-being 5
b
0.143 0.122
Functional well-being 7 0.200 0.171 0.333 0.259
Lung cancer subscale 7 0.200 0.333
BRM physical subscale 7 0.171 0.259
BRM mental subscale 6 0.146 0.222
Weights based on number of items used in constructing the total scores and the
Trial Outcome Index (TOI).
a
Dropped from Version 4.
b
Six items in Versions 3 and 4.
c2638 C2638TextureC10 February 12, 2002 11:45
SUMMARIZING ACROSS HRQoL DOMAINS OR SUBSCALES 237
Table 10.8. Factor analytic weights for the PCS and the MCS of the Medical Out-
come Study SF-36 from a U.S. general population sample [158].
Component PF RP BP GH V SF RE MH
Physical 0.424 0.351 0.318 0.250 0.029 0.08 0.192 0.221
Mental 0.230 0.123 0.123 0.97 0.235 0.269 0.434 0.486
PF=physical function, RP=role-physical, BP=bodily pain, GH=general health, V=
vitality, SF=social function, RE=role-emotional, MH=mental health.
scores were derived from an orthogonal factor rotation. In the resulting scor-
ing algorithms, physical function, role physical, and bodily pain subscales
make a modest negative contribution to the MCS, and the mental health and
role-emotional subscales make a modest negative contribution to the PCS
score (Table 10.8). These negative contributions can produce surprising re-
sults. Simon et al. [139] describe this in a study of antidepressant treatment
where there were modest positive eects over time on the physical function,
role-physical, bodily pain, and general health, but a negative score (nonsignif-
icant) was observed for the PCS because of the very strong positive eects in
the remaining scales. The negative contribution of these remaining subscales
overwhelmed the smaller contributions of the rst four subscales. As a result,
the authors suggest very cautious interpretation of these summary scales when
the treatment has a strong eect on subscales with negative coecients.
Patient weights
Weights derived based on patient preferences are currently uncommon, but
they have a conceptual lure. For example, it is possible that most patients
on adjuvant breast cancer therapy do not consider that physical well-being
during a limited treatment period and disease-free survival have equal weight.
They are willing to undergo adjuvant therapy with its associated morbidity
including severe fatigue, hair loss, and weight gain even after there is no
detectable disease to maximize the chance that the disease will never recur. If
weights that reect patient preferences or values are available, they are likely
to be more relevant than statistically derived weights.
A series of experimental questions are included in the FACT instruments
(Versions 1 to 3) asking the patient to indicate how much the items in each
subscale aect the patients quality of life. For example, following the seven
questions asking the patient to comment on energy, pain, nausea, side eects,
ability to meet needs of family, feeling ill, and spending time in bed, the fol-
lowing question appears:
Looking at the above 7 questions, how much would you say your PHYSICAL
WELL-BEING aects your quality of life?
0 1 2 3 4 5 6 7 8 9 10
Not at all Very much so
c2638 C2638TextureC10 February 12, 2002 11:45
238 SUMMARY MEASURES
Similar questions appear for the other subscales. Responses to these questions
could be used to weight the responses to each of the subscales.
Interestingly, these experimental questions have become optional in Version
4 and are not recommended for use in clinical trials. The developer cites
two basic reasons [17]. First, using weighted and unweighted scores produces
essentially the same results in all analyses examinined. Second was the concern
that the respondents were answering the questions as intended. Although some
appeared to answer as a true weight, others seemed to answer the questions as
a summary of their response to the items, and as many as 15 to 25% did not
seem to understand at all and left the questions blank or responded in rather
unusual ways. Other considerations may be (1) a more-complicated scoring
system that would preclude hand-scoring of the scale, (2) requirements for
additional validation studies, and (3) nonequivalence of scales from study to
study.
Statistically derived weights: Inverse correlation
In a procedure proposed by OBrien [108], the weights are proportional to
the inverse of the correlation of the K domains. The result is that the more
strongly correlated domains are down-weighted. Two subscales that measure
very similar aspects of HRQoL will contribute less to the overall score than
two subscales that measure very dierent aspects of HRQoL. For example,
consider the NSCLC results presented previously in this chapter. The sub-
scales measuring the physical/functional aspects of the patients well-being
are more strongly correlated with each other than the emotional and social as-
pects (Table 10.9). Note that the physical well-being scores have the strongest
correlation with the other subscales ( = 0.450.77) and the social well-being
scale has the weakest correlation with the other subscales ( = 0.36 0.45).
The procedure is rst to compute standardized scores (z
hik
) for each of the K
domains or subscales. Then, compute a weighted sum of these standardized
scores using the column sum of the inverse of the correlation (Table 10.10).
Thus,

S
hi
= J

R
1
z
hi
, where z
hi
= (z
hi1
, . . . , z
hiK
)

, J = (1, . . . , 1)

, and R
is the estimated common correlation matrix of the observations (

).
Summary measures with weights based on the inverse correlation are the
most powerful [108, 117] and do not require specication prior to analysis.
Table 10.9. Correlations among the AUC summary measures in the NSCLC study.
Physical Functional Emotional Social Lung CA
Physical 1.00 0.77 0.64 0.45 0.65
Functional 0.77 1.00 0.56 0.36 0.68
Emotional 0.64 0.56 1.00 0.41 0.49
Social 0.45 0.36 0.41 1.00 0.36
Lung CA 0.65 0.68 0.49 0.36 1.00
c2638 C2638TextureC10 February 12, 2002 11:45
SUMMARIZING ACROSS HRQoL DOMAINS OR SUBSCALES 239
Table 10.10. Inverse correlation (R
1
) and weights (J

R
1
) of the AUC summary
measures in the NSCLC study.
Physical Functional Emotional Social Lung CA
Physical 3.21 1.57 0.76 0.40 0.51
Functional 1.57 2.91 2.25 0.08 0.86
Emotional 0.76 0.25 1.79 0.25 0.12
Social 0.40 0.08 0.25 1.30 0.13
Lung CA subscale 0.51 0.86 0.12 0.13 2.02
Sum (J

R
1
) 0.02 0.31 0.40 0.59 0.40
The disadvantages are that they vary from study to study [20], and they may
not reect the importance that patients place on dierent domains.
Summary statistics
The weighted average of the individual values proposed by OBrien [108] also
extends to the construction of summary statistics with use of a weighted
average of asymptotically normal test statistics such as the two-sample
t-statistic.
t
hk
=

hk

(

)hk
, t
h
= (t
h1
, . . . , t
hK
)

, (10.11)

S
h
= J

R
1
t
h
, J = (1, . . . , 1)

. (10.12)
R is the estimated common correlation matrix of either the raw data in each
treatment group (

) or the pooled correlation matrix of the estimated means

(
jk
). An alternative way to express the summary statistic is

S
h
=
K

k=1
w
k
g(

kj
), (10.13)
where g(

kj
) =
kj
/
(
kj
)
and (w
1
, . . . , w
K
) = J

R
1
. Because

S
h
is a linear
combination of asymptotically normal parameter estimates, the asymptotic
variance of the summary measure is
Var(

S
h
) = W

Cov[g(

hk
)]W. (10.14)
For two treatment groups, we can test the hypothesis S
1
= S
2
or = S
1
S
2
=
0 using a t-test with N 4 degrees of freedom:
t
N4
= (

S
1


S
2
)[Var(

S
1


S
2
)]
1/2
=

[Var(

)]
1/2
(10.15)
c2638 C2638TextureC10 February 12, 2002 11:45
240 SUMMARY MEASURES
for small samples [117]. More generally, for large samples we can test the
hypothesis S
1
= S
2
= S
k
using a Wald
2
statistic:

2
k1
=

[Var(

)]
1

,

= (

S
2


S
1
, . . . ,

S
n


S
1
)

. (10.16)
10.6 Advanced notes
Nonparametric procedures
Based on studies of the power and size of the test, OBrien [108] recommended
the nonparametric procedure (Section 10.3) for general use and the inverse
correlation procedure for normally distributed data with moderate or large
sample sizes.
Combining HRQoL and time to event
Pocock et al. [118] suggest an extension of OBriens weighted average z-scores
to the combination of any asymptotically normal test statistics, illustrat-
ing the concept with the long-rank statistics and a binary endpoint. Fair-
clough [39] describes an application to repeated assessment of HRQoL and
survival where the logarithmic transformation of the survival times is assumed
to be normally distributed. Because

S
k
is a linear combination of asymptot-
ically normal parameter estimates, the asymptotic variance of the summary
measure is Var(

S
k
) = W

Cov[g(

jk
)]W, where W = (w
1
, . . . , w
n
). For two
treatment groups, we can test the hypothesis S
1
= S
2
using a t-statistic, t =
(

S
1

S
2
)/Var(

S
1

S
2
)
1/2
, with df =N 4 for small samples [118]. More gen-
erally, for large samples, we can test the hypothesis S
1
= S
2
= = S
K
using
a Wald
2
statistic:
2
K1
=

[Cov(

)]
1

, where

= (

S
2

S
1
, . . . ,

S
n

S
1
).
Area under the curve
One might be inclined to present the AUC values calculated to the time of
censoring, as one would present survival data. This approach would appear to
have the advantages of displaying more information about the distribution of
the AUC values and accommodating administrative censoring. Unfortunately,
administrative censoring is informative on the AUC scale [55, 56, 80] and the
usual KaplanMeier estimates are biased. Specically, if the missing data are
due to staggered entry and incomplete follow-up is identical for two groups, the
group with poorer HRQoL will have lower values of the AUC and is censored
earlier on the AUC scale. Knowing when a subject is censored on the AUC
scale gives us some information about the AUC score and, thus, the censoring
is informative. Korn [80] suggests an improved procedure to reduce the bias of
the estimator by assuming that the probability of censoring in short intervals
is independent of the HRQoL measures prior to that time. Although this
assumption is probably not true, if the HRQoL is measured frequently and
the relationship between HRQoL and censoring is weak, the violation may be
small enough that the bias in the estimator will also be small.
c2638 C2638TextureC10 February 12, 2002 11:45
SUMMARY 241
Latent variable models
The multiple measures of the dierent aspects of HRQoL can be considered
to be measurement of an unobservable latent variable. Latent variable models
for longitudinal studies have been used by Zwinderman [174] for dichotomous
outcomes and by Busch et al. [12] for continuous normally distributed out-
comes. In both applications, the issue of missing assessments was addressed,
with the subsequent analyses based on the assumption of MAR.
10.7 Summary
Summary measures and statistics
1. Facilitate interpretation by reducing the number of comparisons
2. Increase the power to detect consistent dierences over time or over
HRQoL domains
3. May obscure dierences in dierent directions.
Selection of a summary measure or statistic should depend on the research
question, expected pattern of change over time, and missing-data patterns.
Statistically derived weights are ecient but have the limitations of varying
from study to study and not reecting patient values.
c2638 C2638TextureC10 February 12, 2002 11:45
c2638 C2638TextureC11 February 12, 2002 11:59
CHAPTER 11
Multiple Endpoints
11.1 Introduction
The issue of multiple comparisons in clinical trials assessing HRQoL arises
from three sources: (1) multiple HRQoL measures (scales or subscales), (2) re-
peated postrandomization assessments, and (3) multiple (three or more)
treatment arms. As a result, the problem of multiple comparisons is one of
the major analytic challenges in these trials [79]. For example, in the NSCLC
study, there are ve major subscales in the FACT-Lung instrument (physical,
functional, emotional, and social/family well-being, plus the disease-specic
concerns). There are three follow-up assessments, at 6, 12, and 26 weeks, and
three treatment arms. If we consider the three possible pairwise comparisons
of the treatment arms at each of the three follow-ups for the ve primary
subscales, we have 45 tests. Not only does this create concerns about Type I
error, but reports containing large numbers of statistical tests generally result
in a confusing picture of HRQoL.
Although widely used in the analysis of quality of life in clinical trials [132],
univariate tests of each HRQoL domain or scale and time point can seriously
inate the Type I (false-positive) error rate for the overall trial such that the
analyst is unable to distinguish between the true and false-positive dierences.
Post hoc adjustment is often infeasible because, at the end of analysis, it is
impossible to determine the number of tests performed. There are three typ-
ical strategies to reduce the problem: (1) a priori specication of a limited
number of conrmatory tests, (2) multiple comparison procedures including
alpha adjustments (e.g., Bonferroni) and closed multiple testing, and (3) the
use of summary measures or statistics (see Chapter 10). In practice, a com-
bination of focused hypotheses, summary measures, and multiple comparison
procedures is necessary.
Limiting the number of conrmatory tests
One recommended solution to the multiple comparison problem is to specify
a limited number (3) of a priori endpoints in the design of the trial [59]
(Table 11.1). While theoretically improving the overall Type I error rate
for the study, in practice investigators are reluctant to ignore the remain-
ing data. Descriptive or graphical comparisons of the remaining scales and/or
time points [135] often accompany the formal hypothesis tests. However, these
presentations generally include condence intervals that can be disguised sta-
tistical tests. A more important critique of this approach is an ethical question
about collection of data that will not be used in the primary analysis.
c2638 C2638TextureC11 February 12, 2002 11:59
244 MULTIPLE ENDPOINTS
Table 11.1. General strategies for addressing concerns about multiple comparison
issues in HRQoL studies.
Strategy Limit conrmatory tests
Advantage Does not require any adjustments or special
statistical procedures
Disadvantages Does not strictly control Type I error
What is done about exploratory analyses?
Ethical issue: What is justication for collection
of HRQoL data not required for conrmatory
tests?
Strategy Summary measures and statistics
Advantages Reduce number of comparisons
Increase power to detect consistent small
dierences
Disadvantage Obscure dierences in opposite directions
Strategy Multiple comparison procedures
Advantage Control of Type I errors
Disadvantage Decreased power
Summary measures and statistics
Well-chosen summary measures or statistics often have greater power to de-
tect patterns of consistent HRQoL dierences across time or measures. They
also facilitate interpretation, especially when used to summarize information
over time. However, they also have the potential to blur dierences when
components of the summary measure are in opposite directions. Additional
details about their advantages and disadvantages are discussed in detail in the
previous chapter. In this chapter, we will assume that a summary measure or
statistic is used to reduce the the multiplicity of the repeated assessments
over time. This is partially to simplify the presentation but also because it is
a good strategy.
Multiple comparison procedures
This chapter describes a number of procedures that can be used to control the
Type I error rate for K multiple comparisons. Although the procedures can
generally be applied to any set of hypotheses, the discussion here is presented
for a typical situation of two treatment groups with K distinct measures of
HRQoL. Some special cases that are not unique to HRQoL studies, such as
designs with three or more ordered treatment arms, will not be addressed.
The reader is advised to consult texts on clinical trial design and multiple
comparisons.
c2638 C2638TextureC11 February 12, 2002 11:59
BACKGROUND CONCEPTS AND DEFINITIONS 245
11.2 Background concepts and denitions
Let H
0(1)
, . . . , H
0(K)
denote the K null hypotheses that are to be tested and
let H
0
denote the set of K hypotheses. T
(1)
, . . . , T
(K)
denotes the correspond-
ing K test statistics. The obtained p-value, p
(k)
, denotes the (unadjusted)
probability of observing the test statistic, T
(k)
, or a more extreme value if the
null hypothesis, H
0(k)
, is true. Finally,
(k)
is the signicance level for testing
the kth hypothesis.
When describing the multiple comparison procedures, it is often useful to
order the obtained p-values from smallest to largest: p
[1]
p
[K]
and
then to reorder the corresponding hypotheses H
0[1]
, . . . , H
0[K]
, test statistics
T
[1]
, . . . , T
[K]
, and signicance levels
[1]

[k]
for the multiple compar-
ison procedures. Note that square brackets are used to indicate the ordered
values.
Univariate vs. multivariate test statistics
In this chapter we use the term univariate test statistic to refer to the kth test
statistic, T
[k]
. In some cases this is a test statistic derived from a longitudinal
(multivariate) analysis. For example, a set of ve univariate test statistics (t-
statistics) could be generated from tests of dierences in the slopes from the
longitudinal analyses of the ve distinct subscales in the lung cancer trial.
There is no restriction on whether the K univariate statistics are generated
from K separate models or from a joint multivariate model.
In contrast, a multivariate test statistic corresponds to the joint test of the
K hypotheses, such as an F-test of treatment dierences from a MANOVA
of summary measures (e.g., AUC) for K subscales. The critical dierence is
that the multivariate test statistic is a function of the correlation among the
subscales and the univariate test statistics are not.
Familywise and experimentwise error rates
There are two ways of controlling the probability of erroneously rejecting
at least one null hypothesis. A test or multiple testing procedure controls the
familywise error rate or global -level if the probability of incorrectly rejecting
at least one true null hypothesis does not exceed when all null hypotheses
are simultaneously true. In contrast, a test or multiple testing procedure con-
trols the experimentwise error rate or the multiple -level if the probability
of incorrectly rejecting at least one true null hypothesis does not exceed re-
gardless of which (if any) null hypotheses are true. A procedure that controls
the multiple -level also controls the global -level. The converse is not true.
Control of the experimentwise error rate is more important in conrmatory
trials or those from which decisions that will aect clinical practice will be
made.
c2638 C2638TextureC11 February 12, 2002 11:59
246 MULTIPLE ENDPOINTS
Global vs. individual tests
A global test generates a single statistic for testing the overall treatment eect
and results in the acceptance or rejection of a set of hypotheses.
H
0
: H
0(1)
, . . . , H
0(K)
. (11.1)
For example, one might make the decision to approve a drug if the null hy-
pothesis was rejected on the basis of a global test. When the overall test of H
0
has been rejected, the question remains, Which of the individual hypotheses
can be rejected? A global test does not allow inferences to be made about
individual endpoints.
There are a number of strategies for handling the individual tests. The
loosest strategies require only that the global test be rejected at before ex-
amining the univariate tests each at . These strategies control only the global
-level. At the other extreme are formal closed testing procedures designed to
control the multiple -level. A detailed example is shown later. As the testing
procedures increase control on the -level for the worst-case scenario (inde-
pendent tests), they decrease the power to detect dierences for more typical
settings (correlated tests).
11.3 Multivariate statistics
Global tests
Examples of multivariate statistics for a global test are the HotellingLawley
trace for MANOVA and likelihood ratio tests. Constructing multivariate
statistics may be feasible with simple designs (e.g., cross-sectional or pre/
postdesigns) with minimal missing data. For purposes of illustration, consider
the adjuvant breast cancer study where the Breast Chemotherapy Question-
naire has seven subscales. If our intent is to control the Type I error for testing
treatment eects separately for on-therapy and post-therapy, we could per-
form two multivariate analyses of variance (Table 11.2). By using the change
from baseline as the outcome, the global test of treatment eects on the seven
BCQ measures is highly signicant during therapy(F
7,135
= 5.73, p = 0.0001)
but not signicant post-therapy (F
7,125
= 0.34, p = 0.93).
Closed multivariate testing procedures
Although the global test allows us to make a decision about the multiple end-
points, we are rarely satised with only being able to say that there is a dier-
ence for at least one endpoint. The question always follows: Which endpoints
are aected by the intervention? If we wish to control only the global -level,
we can specify the following procedure. First, perform the multivariate test for
the K endpoints, and then if (and only if) that test is signicant at , perform
the K univariate tests at . In the breast cancer trial, no further tests are
performed for the post-therapy comparisons. Examination of the univariate
c2638 C2638TextureC11 February 12, 2002 11:59
MULTIVARIATE STATISTICS 247
Table 11.2. Results of the multivariate analyses of change from baseline in the ad-
juvant breast cancer trial.
During/ Post/
Pretherapy Pretherapy
F
7,135
= 5.73, F
7,125
= 0.34,
Global Test p = 0.0001 p = 0.93
Univariate t p t p
Hair loss 0.13 0.89 NR NS
Well-being 1.39 0.17 NR NS
Symptoms 4.33 0.0001 NR NS
Trouble 1.30 0.20 NR NS
Fatigue 4.52 0.0001 NR NS
Emotional 0.71 0.48 NR NS
Nausea 1.57 0.12 NR NS
Nominal p-values are reported for univariate tests.
NS = not signicant, NR = not reported.
Table 11.3. Breast cancer trial: Correlations between the change scores.
1 2 3 4 5 6 7
Hair loss 1 0.36 0.38 0.07 0.27 0.35 0.16
Well-being 2 0.31 0.26 0.01 0.23 0.48 0.12
Symptoms 3 0.36 0.24 0.10 0.44 0.28 0.41
Trouble 4 0.10 0.08 0.15 0.15 0.14 0.04
Fatigue 5 0.27 0.27 0.33 0.10 0.46 0.24
Emotional 6 0.27 0.34 0.35 0.35 0.43 0.15
Nausea 7 0.26 0.12 0.42 0.01 0.08 0.16
Upper triangle contains the Pearson correlations for during therapy and
lower triangle contains the Pearson correlations for post-therapy.
tests reveals signicant dierences for the symptoms and fatigue subscales
(unadjusted p < 0.05).
The closed-testing procedure proposed by Marcus et al. [96] for controlling
the multiple -level is more complicated. After rejecting the multivariate test
for the K endpoints H
0(1)
, . . . , H
0(K)
, a stepdown procedure is used. In this
closed-testing procedure, sequentially smaller subsets of the K endpoints are
tested only if all the tests containing a subset of the K hypotheses have been
previously rejected (p < ). Basically, to reject the kth hypothesis H
0(k)
, all
possible subsets of hypotheses containing the kth hypothesis must be rejected.
In some cases, this could involve the testing of 2
K
1 sets of hypotheses.
In the breast cancer example, the multivariate test of dierences during
therapy for all K scales was rst rejected as previously described. The next
c2638 C2638TextureC11 February 12, 2002 11:59
248 MULTIPLE ENDPOINTS
step was to examine the K multivariate tests for K 1 endpoints. All seven
multivariate tests of the possible subsets of six subscales were rejected, each
at = 0.05. Because all were rejected, all possible K 2 combinations were
then tested again at = 0.05. Only 1 of the 21 multivariate tests was not
rejected, the one not containing the fatigue and symptoms scales. Following
the procedure, only hypotheses containing the fatigue and symptoms scales
were subsequently tested. The joint test of these two endpoints was rejected, as
were the individual univariate tests. The conclusion is that there are signicant
dierences in the endpoints identied as fatigue and symptoms.
Limitations
There are several limitations of these multivariate procedures for HRQoL
studies. The rst is that describing the sequential method in clinical journals
is challenging. Only the p-value for the rst global test can be reported; the
subsequent test statistics are unadjusted for the conditional tests. This is likely
to create confusion for reviewers and readers.
The second is the lack of sensitivity to dierences in the same direction
across the multiple endpoints. These multivariate statistics test a hypothesis
of no treatment dierences (H
0
) for the K endpoints against a general alter-
native (H
1
) that there is at least one endpoint where there are treatment
dierences. These tests may sometimes be hard to interpret when the results
are counterintuitive [31]. For example, the test may be statistically signicant
when HRQoL is better in one treatment arm for one endpoint and worse
for another endpoint and not signicant when one treatment appears to be
consistently better over all endpoints [39].
The nal limitation is practical. It is very cumbersome (although possible)
to set up a full multivariate longitudinal model. In the breast cancer study
this would entail 21 repeated measures (7 measures 3 times). Because of the
Table 11.4. Two testing procedures for identifying individual dierences from mul-
tivariate tests.
Strategy Control of Global -level
1. Test multivariate H
0
: H
0(1)
, . . . , H
0(K)
2. If p < perform K univariate tests ()
Advantage Minimal loss of power
Disadvantage Controls only global Type I error
Strategy Control of Multiple -level
1. Multivariate test: H
0(1)
, . . . , H
0(K)
2. If multivariate test rejected p < Sequentially
test subsets the K hypotheses
Advantage Controls multiple and global Type I error
Disadvantages 1. More contrasts to program
2. Reporting of p-values is complicated
c2638 C2638TextureC11 February 12, 2002 11:59
UNIVARIATE STATISTICS 249
large number of covariance parameters, algorithms are very slow. Implement-
ing procedures and sensitivity analyses for missing data will be even more
dicult. Solely for the purposes of illustration, we simplied the example by
deciding to control the error rate separately for the comparisons of treatment
eects during and after therapy and to ignore the loss of subjects with missing
follow-up assessments. In practice we would have to deal with these issues.
11.4 Univariate statistics
In general, multiple comparison procedures using a set of univariate test statis-
tics are much easier to implement and report than those using multivariate
test statistics. There are a number of other practical advantages.
1. In the context of HRQoL studies, the major advantage is that any of the
analytic methods described in the previous chapters that generate a p-value
associated with a test of treatment dierences can be used.
2. Another advantage is that the procedures can be adapted to one-sided tests.
3. The univariate analyses do not require that all K endpoints are measured
on all subjects (although all issues of nonrandom missing data are still
applicable). This is an advantage if a subset of the K endpoints is not
measured because a translation is not available for some of the HRQoL
measures.
4. Another advantage is the ease of application of nonparametric methods.
NSCLC example
The advantage of the univariate approach is illustrated with the NSCLC ex-
ample for the summary measures described in Chapter 10. As the missing
values were handled using multiple imputation, a multivariate analysis of the
M imputed data sets would be complex. In contrast, the univariate analy-
ses are straigtforward, and the results for the AUC summary measure are
displayed in Figure 11.1 and Table 11.5.
Alpha adjustments for K univariate tests
Bonferroni adjustment
A traditional method to address the multiple comparison problem is to utilize
a procedure such as the Bonferroni correction, which adjusts the test statistics
on K endpoints. The global test is based on the smallest p-value, p
[1]
. The
global null hypothesis is rejected when
p
[1]
/K. (11.2)
For individual endpoints the procedure is to accept as statistically signicant
only those tests with p-values that are less than /K.
The Bonferroni procedure controls the experimentwise error rate but is well
known to be quite conservative. If the K outcomes are uncorrelated (the tests
c2638 C2638TextureC11 February 12, 2002 11:59
250 MULTIPLE ENDPOINTS
Figure 11.1. Area under the curve (AUC) with missing data handled by multiple
imputation (MI MCMC). Pairs display mean 1 standard error for no Taxol (left)
and Taxol (right) groups.
are independent) and the null hypotheses are all true, then the probability of
rejecting at least one of the K hypotheses is approximately* K when is
small. The Bonferroni focuses on the detection of large dierences in one or
more endpoints and is insensitive to a pattern of smaller dierences that are
all in the same direction.
Strategy Bonferroni Corrections
p < /K reject H
0
Advantage Well known, simple
Disadvantage Conservative if K is large and endpoints are
correlated
Strategy Sequential Rejective Bonferroni Procedure
Holm [69]
Order univariate p-values: p
[1]
p
[K]
p
[1]
/K reject H
0[1]
, else stop
p
[2]
/(K 1) reject H
0[2]
, else stop
etc.
Advantage More powerful than Bonferroni when some H
0
are
rejected with high probability (large dierence
for some endpoints)
Disadvantage Not much gain when most H
0
are approximately
true (small dierences for all endpoints)
* Pr[min(p-value) ] = 1 (1 )
K
K.
c2638 C2638TextureC11 February 12, 2002 11:59
UNIVARIATE STATISTICS 251
T
a
b
l
e
1
1
.
5
.
S
u
m
m
a
r
y
o
f
u
n
i
v
a
r
i
a
t
e
t
e
s
t
s
t
a
t
i
s
t
i
c
s
b
a
s
e
d
o
n
a
n
a
l
y
s
i
s
o
f
A
U
C
c
a
l
c
u
l
a
t
e
d
u
s
i
n
g
m
u
l
t
i
p
l
y
i
m
p
u
t
e
d
d
a
t
a
f
o
r
l
u
n
g
c
a
n
c
e
r
s
t
u
d
y
a
n
d
a
d
j
u
s
t
e
d

l
e
v
e
l
s
f
o
r
g
l
o
b
a
l
t
e
s
t
o
f
K
=
5
e
n
d
p
o
i
n
t
s
.
U
n
i
v
a
r
i
a
t
e
L
u
n
g
E
m
o
t
i
o
n
a
l
S
o
c
i
a
l
P
h
y
s
i
c
a
l
F
u
n
c
t
i
o
n
a
l
a
n
a
l
y
s
i
s
k
=
1
k
=
2
k
=
3
k
=
4
k
=
5
D
i

e
r
e
n
c
e
5
.
4
5
3
.
3
1
2
.
5
0
0
.
7
5

0
.
0
8
S
t
a
n
d
a
r
d
e
r
r
o
r
1
.
6
3
1
.
8
9
1
.
4
5
2
.
0
2
2
.
0
1
T
-
s
t
a
t
i
s
t
i
c
3
.
3
5
1
.
7
5
1
.
7
2
0
.
3
7

0
.
0
4
p
[
k
]
0
.
0
0
0
9
0
.
0
8
1
0
.
0
8
7
0
.
7
1
0
.
9
7
M
e
t
h
o
d

[
k
]

L
e
v
e
l
s
f
o
r
G
l
o
b
a
l
T
e
s
t
B
o
n
f
e
r
r
o
n
i

1
/
K
0
.
0
1
0
0
R
u
g
e
r

/
K
,
K

=
3
0
.
0
3
0
0
S
i
m
e
s

k
/
K
0
.
0
1
0
0
0
.
0
2
0
0
0
.
0
3
0
0
0
.
0
4
0
0
0
.
0
5
0
0
H
o
m
m
e
l

k
/
K
C
0
.
0
0
4
4
0
.
0
0
8
8
0
.
0
1
3
1
0
.
0
1
7
5
0
.
0
2
1
9
H
o
l
m
s

/
(
K

k
+
1
)
0
.
0
1
0
0
0
.
0
1
2
5
0
.
0
1
6
7
0
.
0
2
5
0
0
.
0
5
0
0
N
o
t
e
:
C
=

K k
=
1
(
1
/
k
)
.
c2638 C2638TextureC11 February 12, 2002 11:59
252 MULTIPLE ENDPOINTS
R ugers inequality
R uger proposed an alternative procedure for situtations where, rather than be-
ing interested in only the smallest p-value, one is interested in the K

smallest
p-values (K

K). For example, one might propose that a general HRQoL

benet is demonstrated by signicant improvement in at least half of the mea-
sured domains without specifying which domains. The procedure is to order
the p-values from smallest to largest:
p
[1]
p
[2]
p
[K]
(11.3)
and reject the global test if
p
[K

]
<
K

K
. (11.4)
If this represents three of the ve subscales in the lung cancer study, the
critical value for the global test is K

/K = 0.05(3/5) = 0.03 for the third-

smallest p-value. If and only if the global test is rejected are the K individual
comparisons made. In the lung cancer study, the global test will not be rejected
and no further comparisons are made.
Simes global test
Hommel [70] and Simes [138] both proposed a global test that avoids specifying
K

. In the Simes test, the global hypothesis of no treatment dierences is

rejected if
p
[k]

k
K
for any k = 1, . . . , K. (11.5)
This procedure is less conservative than the Bonferroni procedure. However,
the procedure does not strictly control the Type I error rate to be less than
except when the tests are independent. For the Hommel procedure, the global
hypothesis is rejected if
p
[k]

k
K

1 +
1
2
+ +
1
K

for any k = 1, . . . , K. (11.6)

Although Hommels procedure strictly controls the Type I error rate to be
less than or equal to , it does so at the cost of power for the procedure,
especially when the tests are expected to be almost independent. However, it
is rarely the case (if ever) that the dierent HRQoL domains are independent
(see Table 11.3). Even though the Simes procedure may inate the Type I
error rate two- to three-fold (depending on K) for independent endpoints, it
is likely to be very practical (and denitely more powerful) for moderately
correlated endpoints typical of HRQoL studies.
c2638 C2638TextureC11 February 12, 2002 11:59
UNIVARIATE STATISTICS 253
Note that both procedures provide only a global test and do not test for
the individual endpoints.
Sequential rejective Bonferroni procedure
Holm proposed less conservative alternative to the classical Bonferroni ad-
justments which allows tests of the individual hypotheses [69]. Again, the
univariate p-values are ordered from smallest to largest. Then, in a stepwise
procedure, the smallest p-value, p
[1]
, is compared to /K. If p
[1]
/K, we
reject the corresponding hypothesis and continue with the stepwise procedure,
comparing p
[k]
to /(K k + 1). If H
0[k1]
has been rejected, evaluate
p
[k]
/(K k + 1). (11.7)
At the point where any null hypothesis cannot be rejected, the procedure
stops and all remaining null hypotheses are accepted.
When all null hypotheses are approximately true, this procedure does not
provide much of an advantage. But when some of the null hypotheses are not
true, this procedure substantially increases the power to detect the moderate
eects of the treatment.
p-value adjustments
Reporting the results of the individual tests is facilitated by the use of adjusted
p-values, p
adj
(k)
. These p-values do not reect the exact probability of observing
the test statistic conditional on the testing procedure but do allow the reader
an easy way to evaluate the results relative to the overall signicance level, .
If
[k]
= f
[k]
, then the adjusted p-values can be computed as
p
adj
[k]
= min(p
[k]
/f
[k]
, 1). (11.8)
For example, in the lung cancer study, the ordered p-values are
p
[1]
=0.0009, p
[2]
=0.081, p
[3]
=0.087, p
[4]
=0.71, p
[5]
=0.97.
For the Simes procedure, f
[k]
= k/K and the adjusted p-values are
p
adj
[1]
=0.0045, p
adj
[2]
=0.20, p
adj
[3]
=0.14, p
adj
[4]
=0.89, p
adj
[5]
=0.97.
This is a bit odd because the adjusted p-values do not reect the original
ordering. An alternative computation corrects this problem. Starting with
the largest unadjusted p-value,
p
adj
[K]
= p
[K]
/f
[K]
, (11.9)
p
adj
[k]
= min

p
[k]
/f
[k]
, p
adj
[k+1]

, k = K 1, . . . , 1. (11.10)
c2638 C2638TextureC11 February 12, 2002 11:59
254 MULTIPLE ENDPOINTS
In the example, this aects only the second p-value, p
[2]
. The adjusted p-values
are
p
adj
[1]
=0.0045, p
adj
[2]
=0.14, p
adj
[3]
=0.14, p
adj
[4]
=0.89, p
adj
[5]
=0.97.
11.5 Resampling techniques
The major limitation of all of the previously described global tests is that
they were developed to control the Type I error rate under the most conser-
vative condition, K independent tests. However, in most studies of HRQoL,
the K endpoints are moderately correlated. As a result, these procedures are
very conservative and the power to detect meaningful dierences is severely re-
duced. Application of the resampling technique has been proposed as a method
to address this problem. The general idea is to obtain an estimate of the dis-
tribution of the cut-o test statistics (T
COT
) for the multiple comparison
procedure for endpoints with an unknown correlation structure. A bootstrap
procedure was rst proposed by Westfall and Young [161] for test statistics
from binomial distributions and adapted by Reitmeir and Wasser [123] for
multiple comparisons of K endpoints between two treatment groups. The
procedure is as follows:
1. Identify the statistic for the global test (T
COT
or p
COT
) and calculate it
from the observed data (Table 11.6).
2. Draw a random sample of subjects with replacement (bootstrap sample)
from the pooled sample of the same size as the original sample.
3. Recompute the statistic from the data associated with the subjects drawn
in the bootstrap sample, p
b
COT
.
4. Repeat the previous two steps B times. B = 10,000 is recommended [123].
Since the bootstrap statistics were calculated under the null hypothesis by
generating the bootstrap samples from the pooled sample, the distribution
of the test statistic is approximated by the distribution of the p
b
COT
. The p-
value for the global test is the proportion of the B bootstrap statistics that
are equal to or more extreme than the observed data statistic (p
b
COT
p
COT
or T
b
COT
T
COT
). If this proportion is less than , the global test is rejected.
Table 11.6. Examples of test statistics for resampling procedure.
Test statistic
Technique T
COT
p
COT
Bonferroni T
[1]
p
[1]
R uger T
[K

]
K

p
[K

]
K

Simes/Hommel max(T
[k]
/k) min(p
[k]
/k)
c2638 C2638TextureC11 February 12, 2002 11:59
SUMMARY 255
11.6 Summary
Multiple comparisons create a major analytic problem in clinical trials with
HRQoL assessments.
Summary measures and statistics combining information across time are ex-
tremely useful for simplifying interpretation, increasing the power to detect
dierences in the same direction, and reducing the number of comparisons.
Multiple comparison procedures are most useful for measures of the multi-
ple domains of HRQoL, especially when there is a concern about obscuring
eects of components that move in opposite directions with the use of sum-
mary measures.
Combining summary measures or statistics (which tends to increase power)
with multiple comparison procedures (which reduce power) is a practical
strategy.
A multiple comparison procedure based on K univariate tests is likely to
be easier to report than one based on multivariate tests.
Selection among the various univariate procedures depends on the research
objectives. The Bonferroni procedure will identify a single strong dierence,
the R uger procedure will identify a prespecied number of K

moderate
eects, and the Simes and Hommel procedures are a compromise.
Strict control of the -level for independent endpoints is overly conservative
for correlated HRQoL measures and more relaxed procedures are preferable
(e.g., Simes is better than Hommel procedure).
Resampling techniques are especially useful for studies where power is likely
to be an issue.
c2638 C2638TextureC11 February 12, 2002 11:59
c2638 2638TexturesC12 February 7, 2002 12:27
CHAPTER 12
Design: Analysis Plans
12.1 Introduction
The analysis plan is an essential part of any randomized clinical trial. Most
statisticians have considerable experience with writing adequate analysis plans
for studies that have one or two univariate endpoints. For example, clini-
cal trials of treatments for cancer have a primary and one to two secondary
endpoints, which are either the time-to-an-event (e.g., disease progression or
death) or a binary outcome such as a complete or partial response (measured
as the change in tumor size). The analysis of these univariate outcomes is
straightforward.
When multiple measures are assessed repeatedly, the choices for analysis are
much more complex. This is true whether the outcome is HRQoL or another
set of longitudinal measures. For example, an analgesic trial might include
longitudinal patient ratings of the average and worst pain as well as use of ad-
ditional medication for uncontrolled episodes of pain. Unfortunately, HRQoL
is often designated as a secondary endpoint and the development of an anal-
ysis is postponed until after the trial begins. Because the analyses are more
complex and may require auxiliary information to be gathered, more atten-
tion is required during protocol development to ensure an analysis that will
answer the research objectives.
At the point that a detailed analysis plan is written, it will become very
clear whether the investigators have a clear view of their research objectives.
Choices for strategies of handling multiple comparisons will require clarity
about the exact role of the HRQoL data. Will it be used in any decision
process about the intervention or will it play only an explanatory or supportive
role? The choice to use a specic summary measure or to perform tests at
each follow-up will require clarity about whether the intent is to explore the
patterns of dierences over time or to identify a general benet. The choice of
analysis methods and strategies for handling missing data will require clarity
about the population of inference. For example, are the conclusions intended
for all those started on the intervention (intent-to-treat) or conditional on
some criteria (e.g., survival, only while on therapy, etc.).
This chapter focuses on selected issues that are critical to the develop-
ment of a detailed analysis plan for the HRQoL component of a clinical trial
(Table 12.1). Even when it is a policy not to include details of the analysis of
secondary endpoints in the protocol, these details should be established before
the protocol is nalized. Failure to do so may result in the omission of critical
information from the data collection or the ability to make a denitive infer-
ence from the data.
c2638 2638TexturesC12 February 7, 2002 12:27
258 DESIGN: ANALYSIS PLANS
Table 12.1. Checklist for statistical analysis plan.
Endpoints
Denitive research objectives and specic a priori hypotheses.
Superiority vs. equivalence.
Dene primary and secondary endpoints; summary measures.
Scoring of HRQoL instruments
Specify method (or reference if it contains specic instructions).
State how missing responses to individual items are handled.
Primary analysis
Power or sample size requirements.
Procedures for handling multiplicity of HRQoL measurements:
What summary measures are proposed?
Adjustment procedures for multiple comparisons.
What statistical procedure is used to model the repeated measures/
longitudinal data?
What assumptions are made about the missing data? What are the ex-
pected rates of missing data and how do they aect the analysis?
Sensitivity analysis
What is the plan for sensitivity analyses?
What models were considered and why were the particular models
selected?
Is the description specic enough for another analyst to implement the
analysis?
Secondary analysis
If some scales are excluded from the primary analysis, what exactly will
be reported for these scales?
What is the justication for including these assessments if they are not
part of the primary analysis?
What exploratory analyses are planned (e.g., psychometric characteris-
tics of HRQoL instrument, relationship between clinical outcomes and
HRQoL measures, treatment eects in specic subgroups)?
12.2 General analysis planWho is included?
If the data analysis is limited to any subset of the randomized patients, this
must be described explicitly and justied by the objectives. One example is
the omission of a small proportion of subjects because appropriate transla-
tions are not available. At the other extreme is limiting the analysis to those
who complete treatment or to survivors. A cautionary note is important at
this point. The more the subset of patients included in the analysis deviates
from the randomized population, the greater the potential for selection bias.
Dierences between the two populations may be attributable to selection
rather than the eects of treatment. Some exclusions, such as failure to start
c2638 2638TexturesC12 February 7, 2002 12:27
MODELS FOR LONGITUDINAL DATA 259
therapy, are easily justied in a blinded study if the exclusion is equally likely
across treatment arms. Other exclusions, such as no follow-up HRQoL, may
depend on treatment and should be very carefully justied with plans for
additional analysis and documentation of the impact of these exclusions.
12.3 Models for longitudinal data
Missing HRQoL assessments are generally not random. This may introduce
bias that cannot be controlled with statistical methods [45] or through impu-
tation [121]. As indicated by the number of chapters in this book, handling
missing data is a challenge and requires careful planning. Appropriate strate-
gies will depend on the objectives and the nature and extent of the missing
data. No single approach is useful in all settings [6].
Ignorable missing data
With minimal missing data or the expectation that the data will be pre-
dominately MAR or MCAR, the recommendation is to use a likelihood-based
method that uses all available data (see Chapters 3 and 5). Depending on the
design of the study, the proposal can be to use either a repeated measures
model (Section 3.3) or a growth curve model (Section 3.4). Even when there
is a concern that the data may be MNAR, these models may be proposed
for (1) initial analysis to be followed by sensitivity analyses, (2) a means for
identifying feasible models for the covariance structure, and (3) generating
initial estimates required for other models.
Nonignorable missing data
When there is any suspicion that the missing data will be related to events
or outcomes that will aect HRQoL, a strategy for checking the sensitivity
of the results should be included in the analysis plan. Because, as we noted
in Chapter 9, it is possible a particular model may suggest ignorable missing
data when others contradict that nding, the analyst should consider at least
two candidate models for the proposed sensitivity analysis.
Event-driven designs and repeated measures models
When the design is event driven or a repeated measures models is appropriate
because there will be a limited number of assessments, the following possibil-
ities exist among the methods described in this book. For the simplest design
with only two assessments (pre/post) and with missing data limited to follow-
up, the sensitivity analysis that includes the CCMV restriction and Browns
protective estimate is a feasible strategy (see Section 8.2). When the analyst
has good auxiliary information about the subjects cause of dropout and sta-
tus after dropout, multiple imputation using MCMC or sequential univariate
regression is a possible approach. In the absence of any auxiliary information
c2638 2638TexturesC12 February 7, 2002 12:27
260 DESIGN: ANALYSIS PLANS
or as a second strategy in a sensitivity analysis, a pattern mixture model with
the NCMV restrictions is a possibility (see Section 8.3). Finally, if the pattern
of dropout is monotone, the observations are equally spaced*, and dropout
occurs after two or more assessments, the analyst could consider the selection
model proposed by Diggle and Kenward (see Section 9.4).
Time-driven designs and growth curve models
A similar logic applies to time-driven designs with growth curve models. Two
of the methods proposed in the book rely on the presence of variation in the
slopes among subjects. When the change over time is expected to be linear, the
conditional linear model is feasible (see Section 9.2). The joint mixed eects
model can also be considered when the change over time is linear or has a
more complex form (see Section 9.3). The joint model also allows the time
to dropout (or other related event) to be censored. If there is no variation
of the slopes among the subjects, a pattern mixture model with parametric
restrictions is possible (see Section 8.4). The term pattern mixture model refers
to a large class of models, and it is necessary to specify in the analysis plan
what restrictions will be used to overcome the problem of underidentication.
Modication of analysis plan
Most of the analytic methods for nonignorable missing data are dependent
on the characteristics of the data. For example, neither the conditional linear
model nor the joint mixed-eects/dropout model (Chapter 9) is feasible if
there is no random variation in the slopes among subjects. Thus, nal decisions
about the analytic methods may require preliminary looks at blinded data.
Initial analysis plans should consider possible options, followed by an updated
plan based on preliminary analyses of blinded data.
12.4 Multiplicity of endpoints
The choice of strategies for handling the multiplicity of endpoints will depend
on a number of issues; the most important issues are the intent and objectives
of the study.
Primary vs. secondary endpoints
At this point the role of the HRQoL assessments in the clinical trial must
be claried. The intent of the trial could be as strong as demonstrating the
ecacy of a drug solely on the basis of HRQoL for the purposes of product
registration, advertising claims, or changing practice. At the other extreme,
the intent may be solely exploratory with no decisions or claims to be made
on the basis of the HRQoL assessments. Most likely, the role of HRQoL will be
* May be unequally spaced if the dependence of dropout on previous observations is not a
function of the time between measurements.
c2638 2638TexturesC12 February 7, 2002 12:27
MULTIPLICITY OF ENDPOINTS 261
somewhere between the two extremes. For example, when there are no eective
therapies for a life-threatening condition, a decision to declare a drug eective
might be made solely on the basis of clinical measures. Tumor response (mea-
surable shrinkage of a tumor) and the time to disease progression are typical
measures of response for cancer treatments. On the other hand, a decision
to declare a drug benecial might be made on the basis of either survival or
improvements in a global measure of HRQoL.
If the intent is to make a decision about the ecacy or benets of the in-
tervention based on the dimensions of HRQoL and other primary endpoints,
then HRQoL is taking on the role of a primary endpoint (regardless of the
language used in the protocol). This implies a need for an explicit strategy
for handling or reducing multiple comparisons. Note that an explicit strategy
does not always mean the use of a formal procedure to control the experiment-
wise error rate simultaneously for all endpoints but may include other options,
such as the use of summary measures or controlling error rates within clusters
of related endpoints [119]. When the HRQoL assessments are considered sup-
portive of primary endpoints, then HRQoL is taking the role of a secondary
endpoint. As secondary endpoints, less stringent criteria are required for con-
trolling error rates. It is still wise to consider options to reduce the multiplicity
of endpoints as a means to improve interpretability of the results.
The intent with respect to the multiple dimensions of HRQoL needs to be
claried. Is the intent to claim a generic HRQoL benet or only benets in
certain dimensions? If generic, what are the criteria that will establish that
inference?
Summary measures
As a rst step, identify summary measures consistent with the objectives that
will reduce multiplicity of data over time. As discussed in detail in Chapter 10,
the choice will be determined by the expected pattern of change over time and
the period of the study where dierences are clinically relevant. Well-chosen
summary measures or statistics often have greater power to detect patterns
of consistent HRQoL dierences across time or measures. This is a particular
advantage in smaller studies, where the power to detect meaningful dierences
is limited.
Multiple comparison procedures
The next step is to consider strategies for controlling the Type I error rate.
Consider several scenarios for drugs to treat a fatal disease. In the rst sce-
nario, there are no eective drugs. Thus, either extending the duration of
survival or the quality of life during survival could be considered a benet.
If this is the rst in a series of trials and the decision is only whether to
continue development of the drug, then no multiple comparison procedures
are mandatory. However, if the intent is to ask for approval of the drug if ei-
ther survival or HRQoL is improved, then an explicit procedure is necessary.
c2638 2638TexturesC12 February 7, 2002 12:27
262 DESIGN: ANALYSIS PLANS
There are still numerous options. One is based on the argument that because
there are no eective drugs, the analyses of survival and HRQoL should each
be performed at = 0.05, with multiple comparison procedures for the K
HRQoL measures. At worst, the experimentwise error rate is no more than
10%. In a second scenario, there exist current therapies that delay the pro-
gression of the disease but do not result in a cure. If the intent was to request
approval if either survival or HRQoL was improved, an option is to split the
alpha between survival (
1
) and HRQoL (
2
). However, if the intent is to
ask for approval only if the drug shows a survival benet and then examine
HRQoL for other reasons, then the analyses of survival and HRQoL would
each be performed at = 0.05. If possible, adopt the same procedure used
for all other clinical endpoints.
12.5 Sample size and power
Although repeated measures and growth curve models are often used for the
analysis of longitudinal studies, it is rare to see sample size calculations for
clinical trials that correspond to these analyses. One of the most common
approaches is to pick a single point in time (often the last), calculate the
sample size for the expected dierence in the group means, and inate that
sample size for dropout. Basically, this calculation provides the sample size for
a repeated univariate analysis. In most cases, this will provide a conservative
(larger than necessary) estimate of the sample size. In this section, a procedure
is described for the estimation of the sample size when the analysis is based
on MLE of a repeated measures or growth curve model with missing data.
Simple linear combinations of
First, consider the case where the hypothesis of interest is a univariate test of
a linear combination of the parameters:
H
0
: = C = 0 vs. H
A
: =

,
where C is a known 1 p vector and

is the value of under the alternative

hypothesis. The hypothesis can be as simple as: the means at a specic time
point are equal across treatment groups. For example, using the cell means
model notation, the comparison of the means at the time of the fourth assess-
ment is H
0
: =
B4

A4
= 0. When the endpoint is a summary measure,
the hypothesis is a linear function of all the parameters.
To illustrate, assume the design for the study includes two groups (A and
B) with four repeated measurements of HRQoL: baseline and three follow-ups.
We wish to have adequate power to detect dierences between the two groups
when the true dierences are 0.4, 0.5, and 0.6 times the standard deviation
at 6, 12, and 26 weeks, respectively. Thus,
B1

A1
= 0.0,
B2

A2
=
0.4,
B3

A3
= 0.5, and
B4

A4
= 0.6. If the summary measure is
the dierence between the baseline assessment and the average of the three
c2638 2638TexturesC12 February 7, 2002 12:27
SAMPLE SIZE AND POWER 263
follow-up assessments, the null hypothesis is
_

B2
+
B3
+
B4
3

B1
_
=
_

A2
+
A3
+
A4
3

A1
_
. (12.1)
If the endpoint is the area under the HRQoL vs. time curve (AUC), using a
trapezoidal approximation, the null hypothesis is

B1
+
B2
2
(6 0) +

B2
+
B3
2
(12 6) +

B3
+
B4
2
(26 12)
=

A1
+
A2
2
(6 0) +

A2
+
A3
2
(12 6) +

A3
+
A4
2
(26 12)
or
3
B1
+ 6
B2
+ 10
B3
+ 7
B4
= 3
A1
+ 6
A2
+ 10
A3
+ 7
A4
. (12.2)
Secondary endpoints might include the change from baseline to each follow-up
assessment:
(
B2

B1
) = (
A2

A1
), (12.3)
(
B3

B1
) = (
A3

A1
), (12.4)
(
B4

B1
) = (
A4

A1
). (12.5)
Rearranging Equations 12.1 through 12.5 in the form = C yields

1
=
_

B2
+
B3
+
B4
3

B1
_

A2
+
A3
+
A4
3

A1
_
, (12.6)

2
= (3
B1
+ 6
B2
+ 10
B3
+ 7
B4
) (3
A1
+ 6
A2
+ 10
A3
+ 7
A4
) ,
(12.7)

3
= (
B2

B1
) (
A2

A1
), (12.8)

4
= (
B3

B1
) (
A3

A1
), (12.9)

5
= (
B4

B1
) (
A4

A1
). (12.10)
Under the alternative hypothesis,
B1

A1
= 0.0,
B2

A2
= 0.4,

B3

A3
= 0.5, and
B4

A4
= 0.6. Thus,

1
= (
B1

A1
) +

B2

A2
3
+

B3

A3
3
+

B3

A3
3
= 0 + 0.4/3 + 0.5/3 + 0.6/3 = 0.5, (12.11)

2
= 3(
B1

A1
) + 6(
B2

A2
) + 10(
B3

A3
) + 7(
B4

A4
)
= 3(0) + 6(0.4) + 10(0.5) + 7(0.6) = 10.5, (12.12)

3
= (
B1

A1
) + (
B2

A2
) = 0 + 0.4 = 0.4, (12.13)

4
= (
B1

A1
) + (
B3

A3
) = 0 + 0.5 = 0.5, (12.14)

5
= (
B1

A1
) + (
B4

A4
) = 0 + 0.6 = 0.6. (12.15)
c2638 2638TexturesC12 February 7, 2002 12:27
264 DESIGN: ANALYSIS PLANS
Basic assumptions
The following procedure is based on asymptotic approximations of the co-
variance of the estimated parameters. It is applicable to any linear model,
E[Y
i
] = X
i
+
i
, where the data are approximately multivariate normal
(Y
i
N(X
i
,
i
)) and the proposed analysis is maximum likelihood estima-
tion. There is an implied assumption that missing data will be MAR and
ignorable for the specied model.
If the sample size is suciently large, then the distribution of the test
statistic z

= /
_

/N is well approximated by a univariate standard normal

distribution, where
_

/N is the standard error of and

2

is N times the
variance of the estimate of . That is,

= N Var(

).
The general form of the sample size approximation for a two-sided test is
N = (z
/2
+z

)
2

2

. (12.16)
For any hypothesis of the form = C, we can estimate the required sample
size if we can determine
2

and
2

.
Equation 12.16 should appear familiar, as there are two well-known univari-
ate sample size formulas that are special cases of this general formula. In the
rst case, a test of the equality of means from two independent samples of the
same size with equal variance corresponds to N = n
A
+ n
B
,

=
A

B
,

= N Var(
A

B
) = N 2
2
Y
/n = 4
2
Y
. Thus, the total sample size is
N = (z
/2
+z

)
2

2

= (z
/2
+z

)
2
4
2
Y
(
A

B
)
2
(12.17)
or the more familiar formula for the number required in each group
n = (z
/2
+z

)
2
2
2
Y
/(
A

B
)
2
.
The second special case is the test of equality of means in a paired sample:
N = n,

=
A

B
,
2

= N Var(
A

B
) = N 2
2
Y
(1 )/N =
2
2
Y
(1 ). Thus, the number of pairs is
N = (z
/2
+z

)
2

2

= (z
/2
+z

)
2
2
2
Y
(1 )
(
A

B
)
2
. (12.18)
Incomplete designs
We assume that we can reasonably estimate the timing of the assessments
(and thus the design matrix for each individual, X
i
) and the proportion of
individuals who drop out of the study at each assessment, such that there are
K patterns of observations,
c2638 2638TexturesC12 February 7, 2002 12:27
SAMPLE SIZE AND POWER 265
Table 12.2. Proportions in each pattern with equal allocation.
Assessments
Pattern Group T1 T2 T3 T4 p
k
1 A X X X X 0.32
2 A X X X 0.10
3 A X X 0.05
4 A X 0.025
5 A 0.005
6 B X X X X 0.32
7 B X X X 0.10
8 B X X 0.05
9 B X 0.025
10 B 0.005
X indicates an observed assessment of HRQoL. p
k
is the
proportion of the total sample with the indicated pattern.
where
n
k
= the number of subjects with the design matrix, X
k
, for the kth
pattern
p
k
= the proportion of subjects, p
k
= n
k
/N, with the kth pattern of
assessments (

K
p
k
= 1)
Let us assume that dropout is equal in the two groups and that all missing
data are due to dropout. The rate of dropout is 1, 5, 10, and 20% at each of
the four assessments (T1 to T4). Thus, 64% will complete all assessments. If
subjects are equally allocated to both groups, the pattern of observations will
appear as displayed in Table 12.2.
In our example, the design matrix (cell means model) for the subjects in
group A with all four observations is
X
k
=
_

_
1 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0
0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0
_

_
, p
k
= 0.32.
The design matrix for the subjects in group A with only the rst two obser-
vations is
X
k
=
_
1 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0
_
, p
k
= 0.05.
Finally, it is assumed that the total number of subjects is large and
[

N
i
X

1
i
X
i
]
1
is a valid approximation of the covariance of the param-
eter estimates (

). This variance can be expressed as a function of the total

c2638 2638TexturesC12 February 7, 2002 12:27
266 DESIGN: ANALYSIS PLANS
number of subjects, N:
Var[

] =
_
N

i=1
X

1
i
X
i
_
1
=
_
K

k=1
n
k
X

1
k
X
k
_
1
=
1
N
_
K

k=1
p
k
X

1
k
X
k
_
1
. (12.19)
The variance of is C Var(

)C

and

= N Var(

) = NC Var(

)C

= C
_
K

k=1
p
k
X

1
k
X
k
_
1
C

. (12.20)
For incomplete designs, the sample size approximation can be written as
N = (z
/2
+z

)
2
C
_
K

k=1
p
k
X

1
k
X
k
_
1
C

/
2

. (12.21)
Example 1: Repeated measures
To use Equation 12.21, we need to dene the following: z
/2
, z

, p
k
, X
k
,
k
,
C, and

. The rst two are familiar critical values associated with the Type I
and Type II error rates. For a two-sided test with = 0.05 and 90% power,
z
/2
= 1.96 and z

= 1.282.
To determine the remaining parameters, we need to dene the dropout
pattern, design, covariance of the observations, null hypothesis, and minimum
dierence of interest under the alternative hypothesis. The dropout pattern
was dened above. Let the design be dened as a cell means model: =
[
A1
,
A2
, . . . ,
B4
]

. For a subject in group A with all four assessments,

X
k
=
_

_
1 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0
0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0
_

_
.
For a subject in group B with only the rst two assessments,
X
k
=
_
0 0 0 0 1 0 0 0
0 0 0 0 0 1 0 0
_
.
Based on the previously stated assumptions, we would expect to see 32 and
5% of the sample with these designs, respectively.
If we have good estimates of the variance of the HRQoL measure over time,
we can use them. However, in their absence we will have to assume some
covariance structure for the repeated measures. Suppose the variance is con-
stant over time and = 0.5. Then,
i
=
2
Y
[(1 )I
n
i
+J
n
i
], where I
n
i
is an
n
i
n
i
identity matrix and J
n
i
is an n
i
n
i
matrix of 1s. If we wish to examine
the dierence between the baseline assessment and the average of the three
c2638 2638TexturesC12 February 7, 2002 12:27
SAMPLE SIZE AND POWER 267
follow-up assessments, Equation 12.6 gives C
1
= [3 1 1 1 3 1 1 1]/3
and Equation 12.11 gives
1
= 0.05.
The SAS program for the estimation of
2

for repeated measures design

(Example 1) appears as follows:
DATA WORK.WORK1;
INPUT M GROUP TIME p_k Y @@;
CARDS;
1 1 1 .32 .0 1 1 2 .32 .0 1 1 3 .32 .0 1 1 4 .32 .0
2 1 1 .1 .0 2 1 2 .1 .0 2 1 3 .1 .0
3 1 1 .05 .0 3 1 2 .05 .0
4 1 1 .025 .0
6 2 1 .32 .0 6 2 2 .32 .4 6 2 3 .32 .5 6 2 4 .32 .6
7 2 1 .1 .0 7 2 2 .1 .4 7 2 3 .1 .5
8 2 1 .05 .0 8 2 2 .05 .4
9 2 1 .025 .0
RUN;
PROC MIXED DATA=WORK.WORK1 MAXITER=1 METHOD=ML;
CLASS M GROUP TIME;
WEIGHT P_K;
MODEL Y=GROUP*TIME/NOINT COVB;
REPEATED TIME/SUBJECT=M TYPE=UN;
PARMS (1)
(.5) (1)
(.5) (.5) (1)
(.5) (.5) (.5) (1)/NOITER;
ESTIMATE THETA GROUP*TIME -3 1 1 1 3 -1 -1 -1/DIVISOR=3 E;
RUN;
Note that in the generated data set, all four assessments appear only for
patterns 1 and 6. The fourth assessment is omitted in patterns 2 and 7, the
third and fourth assessments are missing in patterns 3 and 8, all but the initial
assessment are missing in patterns 4 and 9, and patterns 5 and 10 are omitted
completely.
In theory, the calculations in Equation 12.21 could be performed by hand.
However, the inversion of the matrix required for the calculation of
2

is very
time consuming. The calculations can also be performed using programs that
facilitate matrix algebra (e.g., SAS Proc IML). However, the calculation of
2

can be performed using analysis software (e.g., SAS Proc Mixed), by limiting
the software to one iteration and providing starting values for the covariance
parameters. First, a data set is created with one subject for each of the M
patterns of observations. The design matrix (X
k
) is dened as intended in the
analysis, in this case by GROUP and TIME. Weights are equal to the proportion
of subjects in the group (p
k
). Note that there are two groups of subjects
with no assessments that are not included and the total proportion is only
0.99. The outcome (Y
k
) can be any numeric value; however, if the values
of the alternative hypothesis are used, this procedure will also calculate

.
The MIXED procedure is run with only a single iteration (MAXITER=1 in MODEL
c2638 2638TexturesC12 February 7, 2002 12:27
268 DESIGN: ANALYSIS PLANS
statement) and (NOITER in PARMS statement) with the covariance specied
(PARMS statement) and xed (NOITER option).
Note that in the above example, Y and are scaled so that the variance
of Y
ijk
is standard normal (N(0, 1)) and PARMS species the correlation of Y .
It is not necessary to convert everything to a standard normal distribution; it
is critical that one consistently use either the unstandardized or standardized
values of Y (delta) and .
Selected SAS output from PROC MIXED for Example 1 appears as follows:
Covariance Matrix for Fixed Effects
Effect GROUP TIME COL1 COL2 COL3 COL4 COL5 COL6 COL7 COL8
GROUP*TIME 1 1 2.02 1.01 1.01 1.01 0.00 0.00 0.00 0.00
GROUP*TIME 1 2 1.01 2.10 1.03 1.03 0.00 0.00 0.00 0.00
GROUP*TIME 1 3 1.01 1.03 2.26 1.07 0.00 0.00 0.00 0.00
GROUP*TIME 1 4 1.01 1.03 1.07 2.73 0.00 0.00 0.00 0.00
GROUP*TIME 2 1 0.00 0.00 0.00 0.00 2.02 1.01 1.01 1.01
GROUP*TIME 2 2 0.00 0.00 0.00 0.00 1.01 2.10 1.03 1.03
GROUP*TIME 2 3 0.00 0.00 0.00 0.00 1.01 1.03 2.26 1.07
GROUP*TIME 2 4 0.00 0.00 0.00 0.00 1.01 1.03 1.07 2.73
ESTIMATE Statement Results
Parameter Estimate Std Error
Theta -0.50000 1.72635534
The matrix displayed in the SAS output as the Covariance Matrix for
Fixed Effects is N Var(

) = [

K
k=1
p
k
X

1
k
X
k
]
1
. When the outcome
(Y ) is specied as the expected values under the alternative hypothesis, the
ESTIMATE Statement Results contains

= 0.500 as the Estimate.

The square root of
2

is displayed as the Std Error:

_

= 1.726. We can
use these values in the calculation of the sample size N. For a two-sided test
with a Type I error of 5% and 90% power, the total required sample is 126
(or 63 subjects per arm):
N = (z
1/2
+z

)
2

2

= (1.96 + 1.282)
2
(1.726)
2
/(0.5)
2
= 125.3.
Example 2: Growth curve model
Consider the same two-group design where the four assessments are assumed
to occur at 0, 6, 12, and 26 weeks. The dropout rates are the same. Instead
of using a repeated measures design, we plan to use a piecewise linear growth
curve model. The model is dened with four parameters per group:
1
is the
intercept,
2
is the slope for the rst 6 weeks,
3
is the change in the slope
after 6 weeks, and
4
is the change in the slope after 12 weeks. For a subject
in group A with all four assessments,
X
i
=
_

_
1 0 0 0 0 0 0 0
1 6 0 0 0 0 0 0
1 12 6 0 0 0 0 0
1 26 20 14 0 0 0 0
_

_
.
c2638 2638TexturesC12 February 7, 2002 12:27
SAMPLE SIZE AND POWER 269
Let us consider the sample size required for several endpoints. The rst three
endpoints are the change from baseline to 6, 12, and 26 weeks (Equations 12.3
through 12.5). The fourth endpoint is the trapezoidal approximation of the
area under the HRQoL vs. time curve (Equation 12.2). The nal endpoint
is the same as in the previous example: the averages of the estimates at 6,
12, and 26 weeks minus the baseline assessments. The null hypothesis is that
there is no dierence between the two groups.
The following SAS program is very similar to that used in Example 1; the
major dierence is the setup of the variables used to model change over time.
We use the same method of dening the covariance structure, noting that the
compound symmetry is identical to assuming a random intercept in a mixed-
eects model. If there is sucient information to estimate a more complex
variance structure, such as two random eects, then
i
should be dened as
if the subject would have assessments at 0, 6, 12, and 26 weeks.
DATA WORK.WORK1;
INPUT M GROUP TIME WEIGHT Y @@;
TIME00=TIME;
TIME06=MAX(TIME-6,0);
TIME12=MAX(TIME-12,0);
CARDS;
1 1 0 .32 .0 1 1 6 .32 .0 1 1 12 .32 .0 1 1 26 .32 .0
2 1 0 .10 .0 2 1 6 .1 .0 2 1 12 .1 .0
3 1 0 .05 .0 3 1 6 .05 .0
4 1 0 .025 .0
6 2 0 .32 .0 6 2 6 .32 .4 6 2 12 .32 .5 6 2 26 .32 .6
7 2 0 .10 .0 7 2 6 .10 .4 7 2 12 .10 .5
8 2 0 .05 .0 8 2 6 .05 .4
9 2 0 .025 .0
RUN;
PROC MIXED DATA=WORK.WORK1 MAXITER=1 METHOD=ML;
CLASS M GROUP TIME;
WEIGHT WEIGHT;
MODEL Y=GROUP GROUP*TIME00 GROUP*TIME06 GROUP*TIME12/NOINT COVB;
REPEATED TIME/SUBJECT=M TYPE=UN;
PARMS (1)(.5)(1)(.5)(.5)(1)(.5)(.5)(.5)(1)/NOITER;
ESTIMATE 6 WK GROUP*TIME00 6 -6/E;
ESTIMATE 12 WK GROUP*TIME00 12 -12 GROUP*TIME06 6 -6/E;
ESTIMATE 26 WK GROUP*TIME00 26 -26 GROUP*TIME06 20 -20
GROUP*TIME12 14 -14/E;
ESTIMATE AUC/26 GROUP 26 -26 GROUP*TIME00 338 -338
GROUP*TIME06 200 -200 GROUP*TIME12 98 -98
/DIVISOR=26 E;
ESTIMATE MEAN GROUP*TIME00 44 -44 GROUP*TIME06 26 -26
GROUP*TIME12 14 -14/DIVISOR=3 E;
MAKE ESTIMATE OUT=WORK.ESTIMATE;
RUN;
DATA WORK.SSIZE;
SET WORK.ESTIMATE;
c2638 2638TexturesC12 February 7, 2002 12:27
270 DESIGN: ANALYSIS PLANS
N=((1.96+1.282)*SE/EST)**2;
RUN;
PROC PRINT DATA=WORK.SSIZE;
ID PARM;
VAR N EST SE;
RUN;
The results of the estimate statements are output to a data set and the
total sample size (N) is calculated in a subsequent DATA Step. The output
is displayed below. Note that the estimated sample size varies among the
hypotheses.
PARM N EST SE
6 wk 276.0 -0.40000000 2.04977813
12 wk 190.8 -0.50000000 2.13056662
26 wk 159.7 -0.60000000 2.33866920
AUC/26 149.5 -0.44615385 1.68248549
Mean 125.3 -0.50000000 1.72635534
Other considerations
Intermittent missing data patterns and time-varying covariates
In the previous two examples, we assumed that missing data occurred only
as a result of dropout. We ignored the possibility of intermittent patterns of
missing data. Another problem is that there may be considerable variation
in the exact timing of the assessments. Both will have some impact on the
power (and sample size estimates). If there is enough information to predict
the expected patterns, it is possible to address these problems. One strategy
is to consider all patterns of observed data. However, in most cases this is
cumbersome. An alternative approach is to obtain an estimate of
2

by ran-
domly generating a large (e.g., N
S
= 10000) number of subjects with their
expected pattern of observation (X
i
) and corresponding variance (
i
). Then
estimate
2

= N
S
C[

N
S
X

1
i
X
i
]
1
C

. One can do this again using PROC

MIXED (omitting the WEIGHT statement and modifying the SUBJECT= option).
The value of SE in the output labeled ESTIMATE Statement Results of PROC
MIXED is
2

N
S
, so it is necessary to multiply this value by

N
S
, where N
S
is the number of observations used to simulate the pattern of observations.
If we extend Example 2, adding an additional 10% of missing data at each
point in time due to reasons other than dropout, we obtain estimates of 2.18,
2.26, 2.47, 1.70, and 1.83 for
2

for the ve hypotheses and sample size esti-

mates of 312, 214, 178, 153, and 141, respectively. This represents an ination
of 12 to 13% in the sample size requirements for all but the AUC.
Unequal allocations of subjects to treatment groups
Some clinical trial designs specify an unequal allocation of subjects to treat-
ment arms. Because the sample size calculation is specied in terms of the
c2638 2638TexturesC12 February 7, 2002 12:27
SAMPLE SIZE AND POWER 271
Table 12.3. Proportions in each pattern with unequal (2:1) allocation.
Assessments
Pattern Group T1 T2 T3 T4 p
k
1 A X X X X 0.4267
2 A X X X 0.1333
3 A X X 0.0667
4 A X 0.0333
5 A 0.0067
6 B X X X X 0.2133
7 B X X X 0.0667
8 B X X 0.0333
9 B X 0.0167
10 B 0.0033
X indicates an observed assessment of HRQoL. p
k
is the
proportion of the total sample with the indicated pattern.
total sample size (N), this requires only an adjustment of the proportions
within each pattern of observations (p
k
). For example, if subjects were allo-
cated in a 2:1 ratio to treatments A and B, then the pattern would appear as
displayed in Table 12.3.
Multivariate tests
Sample size approximations can also be created for multivariate tests. For
example, we might wish to test simultaneously for dierences in the change
from baseline in our example.
H
0
: (
B2

B1
) (
A2

A1
) = 0,
(
B3

B1
) (
A3

A1
) = 0,
(
B4

B1
) (
A4

A1
) = 0.
For large samples, one can generalize the univariate z-statistic to a multi-
variate
2
-statistic. The null hypothesis is H
0
: = C = G, where C
(p)
and G
(1)
are known. The
2
test statistic is

(Var())
1

and the power

of the test is
Pr
_

() >
2
,

= 1 , (12.22)
where
2
,
is the critical value for a
2
distribution with degrees of freedom
(). is the noncentrality parameter, such that
=

(Var())
1

= N

_
_
C
_
K

k=1
p
k
X

1
k
X
k
_
1
C

_
_
1
, (12.23)
when = C G is calculated under the alternative hypothesis.
c2638 2638TexturesC12 February 7, 2002 12:27
272 DESIGN: ANALYSIS PLANS
We can again use the output from our SAS program to simplify the calcula-
tions. First, we calculate the critical value. In our example, this is
2
3,=0.05
=
7.815.* Second, we calculate the noncentrality parameter () from Equa-
tion 12.31. In our example, = 14.17.** Finally, we obtain the quantity

(C[

K
k=1
p
k
X

1
k
X
k
]
1
C

)
1
by adding to the Proc MIXED statements.
CONTRAST Change from Baseline
GROUP*TIME00 6 -6,
GROUP*TIME00 12 -12 GROUP*TIME06 6 -6,
GROUP*TIME00 26 -26 GROUP*TIME06 20 -20 GROUP*TIME12 14 -14
/CHISQ;
The estimate of the total sample size is /CHISQ = 14.17/0.086=163.1, where
CHISQ is obtained from the Proc MIXED output.
Small sample size approximations
For previous sample size calculations, the approximations assume the sample
size is suciently large that the asymptotic approximation of the covariance
of the parameters is appropriate and the loss of degrees of freedom when using
t- and F-statistics will not aect the results. When the sample sizes are small,
the above procedures can be used to obtain the rst estimate of the sample
size. The estimation procedure is repeated, using the t- or F-distributions
with updated estimates of the degrees of freedom (), until the procedure
converges.
N = (t
,/2
+t
,
)
2

2

= (t
/2
+t

)
2
C
_
K

k=1
p
k
X

1
k
X
k
_
1
C

/
2

. (12.24)
Restricted maximum likelihood estimation
Kenward and Rogers [78] provide the details for modifying the estimated
covariance of

for restricted maximum likelihood estimation (REML) and
propose an F-distribution approximation for small sample inference.
12.6 Reporting results
All the standard guidelines for reporting clinical trials apply to the reporting of
HRQoL studies. Because all readers are not as familiar with some of the issues
associated with measurement of HRQoL as they are with other clinical end-
points, some additional guidelines are worth mentioning.Based on a systematic
* SAS: c alpha=cinv(1-0.05,3).
** SAS: lambda=cnonct(c alpha,3,1-0.9).
c2638 2638TexturesC12 February 7, 2002 12:27
REPORTING RESULTS 273
Table 12.4. Checklist for reporting requirements for HRQoL trials.
Introduction
Rational for assessing HRQoL in the particular disease and
treatment.
State specic a priori (pretrial) hypotheses.
Methods
Justication for selection of instruments assessing HRQoL (see
Chapter 2). References to instrument and validation studies. De-
tails of psychometric properties if a new instrument. Include copy
of instrument in appendix if previously unpublished. Details of
any modications of the questions or formats.
Details of cross-cultural validation if relevant and previously
unpublished.
Method of administration (self-report, face-to-face, etc.).
Planned timing of the study assessments.
Method of scoring, preferably by reference to a published article
or scoring manual, with details of any deviations.
Interpretation of scores. Do higher values indicate better
outcomes?
Methods of analysis. What analyses were specied a priori and
which were exploratory?
Which dimensions or items of the HRQoL instruments were se-
lected as endpoints prior to subject accrual?
What summary measures and multiple-comparison procedures
were used? Were they specied a priori?
Results
Timing of assessments and length of follow-up by treatment group.
Missing data:
Proportions with missing data and relevant patterns.
How were patients who dropped out of the study handled?
How were patients who died handled?
Results for all scales specied in protocol/analysis plan (negative
as well as positive results).
If general HRQoL benet reported, summary of all dimensions.
If no changes were observed, describe evidence of responsiveness
to measures in related settings and the lack of oor and ceiling
eects in current study.
review of 20 articles reporting HRQoL in cancer trials in 1997, Lee and Chi
[85] identify the major deciencies that should be addressed as (1) the failure
to provide a rationale of HRQoL assessment and (2) inadequate description
of methodology. In addition to the checklist included here (Table 12.4), the
reader is encouraged to read other published guidelines [145].
c2638 2638TexturesC12 February 7, 2002 12:27
274 DESIGN: ANALYSIS PLANS
12.7 Summary
The analysis plan is driven by explicit research objectives.
Analyses of HRQoL data are often more complex, because of the longitudi-
nal and multidimensional nature, than the analyses of traditional univariate
outcomes.
The analysis plan should be developed prior to initiation of the study re-
gardless of whether HRQoL is a primary or secondary endpoint.
c2638 C2638Appendix-I February 7, 2002 12:28
APPENDIX I
Abbreviations
ABB approximate Bayesian bootstrap
ACMV available case missing value
AIC Akaikes information criterion
ANOVA analysis of variance
ARMA autoregressive moving average
AUC area under the curve
BCQ Breast Chemotherapy Questionnaire
BIC Bayesian information criterion
BLUE best linear unbiased estimate
BLUP best linear unbiased predictor
BRM biologic response modier
CAF cyclophosphamide, doxirubicin, and 5-urouracil
CCMV complete case missing value
13-CRA 13-cis-retinoic acid
CRD completely random dropout
EB empirical Bayes
EBLUP empirical best linear unbiased predictor
ECOG PS Eastern Cooperative Oncology Group scale for performance
status
EM estimationmaximization
FACT-BRM Functional assessment of cancer-biological response modi-
ers
FACT-G Functional assessment of cancergeneral questionnaire
FACT-Lung Functional assessment of cancerlung questionnaire
FU follow-up
HRQoL health-related quality of life
ID informative dropout
IFN interferon alpha-2a
LVCF last value carried forward
LOCF last observation carried forward
MANOVA multivariate analysis of variance
MAR missing at random
MCAR missing completely at random
MCMC multiple chain Monte Carlo
MCS mental component scale (SF-36)
MI multiple imputation
ML maximum likelihood
MLE maximum likelihood estimation (estimate)
c2638 C2638Appendix-I February 7, 2002 12:28
276 ABBREVIATIONS
MNAR missing not at random
NCMV nearest case missing value
NSCLC non-small-cell lung cancer
OLS ordinary least squares
PCS physical component scale (SF-36)
RD random dropout
REML restricted maximum likelihood estimation
S.D. standard deviation
S.E. standard error of the mean
SF-36 Medical Outcome Study Short Form36
TOI treatment outcome index
c2638 C2638Appendix-II February 12, 2002 12:1
APPENDIX II
Notation
the vector of xed-eects parameters

h
the vector of xed-eects parameters for the hth group

i
+ d
i

{p}
the vector of xed-eects parameters for subjects in the pth missing
data pattern
B

the vector of xed-eects parameters from the imputation model

(Y
obs
i
= X
obs
i
B

i
)

the vector of estimates of the xed-eects parameters from the impu-

tation model

(m)
the vector of xed-eects parameters used to impute the mth set of
missing values

(m)
the estimates of the xed-eects parameters from the mth set of im-
puted data
d
i
the vector of random eects for the ith individual
D the covariance of the random eects, d
i
; the elements of D are noted
by
rc

i
the vector of residual errors for a xed-eects model (
i
= Y
i
X
i
)
e
i
the vector of residual errors for a mixed-eects model (e
i
= Y
i
X
i

Z
i
d
i
)
h the indicator of the hth group (usually treatment)
i the indicator of the ith independent unit of observations (usually pa-
tient or subject)
j the indicator of the jth measurement on the ith subject
k the indicator of the kth endpoint or HRQoL scale
M a generic term that is associated with the dropout process; may refer
to the time of the event associated with dropout, T
D
i
, the pattern
of missing observations, R
i
, or random eects,
i
or d
i

hj
the mean of the hth group at the time of the jth observation.

{p}
the proportion of subjects in the pth missing data pattern
R
i
a vector of indicators of the missing data pattern for the ith individual,
where r
ij
= 1 if y
ij
is observed and r
ij
= 0 if y
ij
is missing
R
i
the covariance of e
i
= Y
i
X
i
Z
i
d
i
for mixed-eects models and

i
= Y
i
X
i
for repeated measures models; in most mixed-eects
models, the residual errors are uncorrelated, R
i
=
2
I
c2638 C2638Appendix-II February 12, 2002 12:1
278 NOTATION

i
the covariance of the complete data (Y
i
), which is a known function
of the vector of unknown variance parameters, :
i
= f()

rc
the element in the rth row and cth column of R
i

RC
the element in the rth row and cth column of D
i
t
hij
the time elapsed from the beginning of the trial until the obser-
vation of the Y
hij
measure
t
c
hij
the value t
hij
raised to the cth power
t
[c]
hij
the time elapsed from T
[c]
until the observation of the Y
hij
mea-
sure; the brackets distinguish t
[c]
hij
from t
c
hij
T
D
i
the time to an event associated with dropout, such as the time
to the last HRQoL assessment; the time to discontinuation of
treatment; or the time to death

[A.B]
a vector of mean and variance parameters estimated from the
regression of A on B
X
hi
the design matrix of xed covariates corresponding to the complete
data (Y
hi
); the design matrix will include the time of measure-
ment relative to the beginning of the study as well as indicators
of treatment and other explanatory variables
X
i
the design matrix of xed covariates corresponding to the complete
data (Y
i
), where the group indicator (h) has been dropped for
simplication of notation
X
obs
i
the design matrix corresponding to Y
obs
i
included in the imputa-
tion model (Y
obs
i
= X
obs
i
B

i
)
Y
hi
the complete data vector of planned observations of the outcome
for the ith individual in the hth group, which includes both the
observed data Y
obs
hi
and missing Y
mis
hi
observations of HRQoL
Y
i
the complete data vector of planned observations of the out-
come for the ith individual (the indicator of group (h) has
been dropped to simplify notation when it is not necessary for
clarication)
Y
C
i
designates the complete cases as the set of responses from subjects
who completed all possible assessments (e.g., r
ij
= 1 for all
possible y
ij
from the ith subject)
Y
I
i
designates the incomplete cases as the set of responses from sub-
jects who did not complete all possible assessments (e.g., r
ij
= 1
for all possible observations on the ith subject)
Y
hij
the jth observation on the ith individual in the hth group
Y
obs
i
the vector of observed data included in the imputation model
(Y
obs
i
= X
obs
i
B

i
)
(Y, R) the complete data of the joint distribution of Y and R
(Y
obs
, R) the observed data of the joint distribution of Y and R
c2638 C2638Appendix-II February 12, 2002 12:1
NOTATION 279
f(R| Y, X, ) the distribution of R for a given set of outcomes (Y) and
covariates (X)
f(Y, R| , ) the joint density function for the complete data
f(Y
obs
, R| , ) the joint distribution of the observed data
Z
i
the design matrix corresponding to the random eects
(d
i
) on the ith individual
Z
(m)

a vector of random numbers drawn from a standard nor-

mal distribution of the same dimension as for the mth
imputation
c2638 C2638Appendix-II February 12, 2002 12:1
c2638 C2638Appendix-III February 12, 2002 12:2
APPENDIX III
Formal Denitions for Missing Data
Let Y
i
denote a vector of scores from a series of planned HRQoL assessments,
some of which are observed, Y
obs
i
, and some of which are missing Y
mis
i
. X
i
denotes explanatory variables (or covariates) that include indicators of treat-
ment, time since the beginning of treatment or diagnosis, and other subject
characteristics. Let denote the parameters that describe the HRQoL pro-
cess. M
i
denotes a vector of variables representing the missing data process.
Often M
i
= R
i
, where R
i
denotes a vector of binary indicators of whether a
particular HRQoL score was observed (r
ij
= 1) or missing (r
ij
= 0), but it
can also denote the time to dropout, an associated event (T
D
i
), or a random
coecient (
i
). Let denote the parameters that describe the missing data
process.
III.1 Joint distribution of Y and M
Complete data
The complete data for both the measurement and the missing data process is
the joint distribution of Y = (Y
obs
, Y
mis
) and M:
f[Y, M| X, , ]. (III.1)
There are two ways to factor the density function for the complete data. In
the rst factorization, f[Y| M, X, ] is the density function for the measure-
ment process which depends on the missing data pattern (M).
f[Y, M| X, , ] = f[Y| M, X, ]f[M| X, ]. (III.2)
This factorization forms the basis of a family of models referred to as mix-
ture models. In this factorization, the model for the missing data mechanism,
f[M| X, ], does not depend on the HRQoL measurements. In the second fac-
torization, f[Y
i
| X
i
, ] is the density function for the measurement process,
which depends only on the covariates and the parameters .
f[Y, M| X, , ] = f[Y| X, ]f[M| Y, X, ]. (III.3)
The model of the missing-data mechanism, f[M
i
| Y
i
, X
i
, ], may depend on
the covariates, the observed data Y
obs
i
, and the missing data Y
mis
i
. This
factorization forms the basis of a family of models referred to as selection mod-
els. In both classes of models, strong assumptions must be made about the
c2638 C2638Appendix-III February 12, 2002 12:2
282 FORMAL DEFINITIONS FOR MISSING DATA
missing values (Y
mis
). The important distinction between the two is whether
the assumption is part of the model for missingness (M), as in the selection
models, or the model for the outcome (Y), as in the mixture models.
Observed data
The observed data for the measurement and missing data process has the joint
distribution of Y
obs
and M:
f(Y
obs
, M| X, , ) =

f(Y, M| X, , )dY
mis
=

f(Y
obs
, Y
mis
| X, )
f(M| Y
obs
, Y
mis
, X, )dY
mis
. (III.4)
III.2 Missing data mechanism
The following taxonomy is based on a model for the missing data mechanism,
denoted as in the selection models:
f

M
i

Y
obs
i
, Y
mis
i
, X
i
,

. (III.5)
Thus, the missing data mechanism is dened by the dependence of missing-
ness on the covariates (X
i
), the observed responses (Y
obs
i
), and the missing
responses (Y
mis
i
).
MCAR: Missing completely at random
The strongest assumption about the missing data mechanism is that the data
are missing completely at random (MCAR).
f

M
i

Y
obs
i
, Y
mis
i
, X
i

= f[M
i
]. (III.6)
This mechanism assumes that the probability that an observation is missing
is independent of observed HRQoL measures (Y
obs
i
) at other times and the
missing HRQoL measure (Y
mis
i
). Little and Rubin (see Little [92]) distinguish
MCAR from covariate-dependent missingness.
f

M
i

Y
obs
i
, Y
mis
i
, X
i

= f[M
i
| X
i
]. (III.7)
This mechanism assumes that the probability that an observation is observed
is dependent only on covariates (X
i
). Thus, conditional on the covariates, miss-
ingness is independent of observed HRQoL measures (Y
obs
i
) and the missing
HRQoL measure (Y
mis
i
). When the missing data are the result of dropout
and have a monotone pattern, this mechanism is sometimes referred to as
covariate-dependent dropout [92] or completely random dropout (CRD) [36].
c2638 C2638Appendix-III February 12, 2002 12:2
MISSING DATA MECHANISM 283
MAR: Missing at random
The assumptions about the missing data mechanism are relaxed slightly when
we assume the data are missing at random (MAR).
f

M
i

Y
obs
i
, Y
mis
i
, X
i

= f

M
i

Y
obs
i
, X
i

. (III.8)
The probability that an observation is observed may depend on the xed
covariates, X
i
, and the observed data, Y
obs
, but may not depend on the
value of the missing data Y
mis
. Note that when this is true, Y
mis
drops out
of Equation III.4 and the joint distribution of the observed data, (Y
obs
, M),
reduces to
f(Y
obs
, M| X, , ) =

f(Y
obs
, Y
mis
| X, )
f(M| Y
obs
, Y
mis
, X, )dY
mis
=

f(Y
obs
| X, )f(M| Y
obs
, X, )dY
mis
= f(Y
obs
| X, )f(M| Y
obs
, X, ). (III.9)
MAR vs. ignorable
Often we interchange the terms MAR and ignorable. They are not strictly
equivalent. If a dropout process is random (MAR or MCAR), then unbiased
estimates can be obtained from likelihood-based estimation that ignores the
dropout mechanism, provided the parameters describing the measurement
process of HRQoL () are functionally independent of the parameters de-
scribing the dropout process (). This is referred to as the separability or
parameter distinctness condition and is called ignorable by Little and Rubin
[88]. When the separability condition is satised, inference can be based solely
on the marginal density of the observed data [32, 75, 154]:
L(| X, Y
obs
) f(Y
obs
| X, ). (III.10)
This allows straightforward estimation of using procedures such as the EM
algorithm [32, 75, 115].
MNAR: Missing not at random
The nal mechanism relaxes the remaining assumption.
f

M
i

Y
obs
i
, Y
mis
i
, X
i

= f

M
i

Y
obs
i
, Y
mis
i
, X
i

. (III.11)
The probability that an observation is missing may depend on the value of
the missing data Y
mis
.
The terms outcome-based and random-eect dependent have been used to
distinguish between missing data mechanisms dependent on Y
mis
i
and those
dependent on
i
= +d
i
in mixed-eects models.
c2638 C2638Appendix-III February 12, 2002 12:2
c2638 C2638texRef February 7, 2002 12:29
References
[1] Aaronson, N.K. et al., The European Organization for Research and
Treatment of Cancer QLQ-30: a quality-of-life instrument for use in
international clinical trials in oncology, J. Natl. Cancer Inst., 85: 365
376, 1993.
[2] Bacik, J. et al., The Functional Assessment of Cancer Therapy-BRM
(FACT-BRM): a new tool for the assessment of quality of life in patients
treated with biologic response modiers, submitted.
[3] Bacik, J. et al., Quality of life analysis for a randomized Phase III
trial of interferon alfa-2a versus IFN plus 13-cis-retinoic acid in patients
with advanced renal cell carcinoma, in ASCO Proceedings, New Orleans,
2000.
[4] Barnard, J. and Meng, X.L., Applications of multiple imputation in
medical statistics: from AIDS to NHANES. Stat. Meth. Med. Res., 8:
1736, 1999.
[5] Barnard, J. and Rubin, D.B., Small-sample degrees of freedom with
multiple imputation, Biometrika, 86: 948955, 1999.
[6] Bernhard, J. et al., Missing quality of life data in clinical trials: serious
problems and challenges, Stat. Med., 17: 517532, 1998.
[7] Bonomi, P. et al., Comparison of survival and quality of life in ad-
vanced non-small-cell lung cancer patients treated with two dose lev-
els of Paclitaxel combined with cisplatin versus etoposide-cisplatin:
results of an Eastern Cooperative Group trial, J. Clin. Oncol., 18:
62331, 2000.
[8] Bouchet, C. et al., Selection of quality-of-life measures for a prevention
trial: a psychometric analysis, Controlled Clin. Trials, 21: 3043, 2000.
[9] Brooks, M.M. et al., Quality of life at baseline: is assessment after ran-
domization valid? Med. Care, 36: 15151519, 1998.
[10] Brown, C.H., Protecting against nonrandomly missing data in longitu-
dinal studies, Biometrics, 46: 143155, 1990.
[11] Buck, S.F., A method of estimation of missing values in multivariate
data suitable for use with an electronic computer, J. R. Stat. Soc. Ser.
B, 22: 302306, 1960.
[12] Busch, P. et al., Life quality assessment of breast cancer patients receiv-
ing adjuvant therapy using incomplete data, Health Econ., 3: 203220,
1994.
[13] Cella, D.F., Skeel, R.T., and Bonomi, A.E., Policies and Procedures
Manual, Eastern Cooperative Oncology Group Quality of Life Subcom-
mittee, unpublished, 1993.
c2638 C2638texRef February 7, 2002 12:29
286 REFERENCES
[14] Cella, D.F. et al., The Functional Assessment of Cancer Therapy
(FACT) scales: development and validation of the general measure, J.
Clin. Oncol., 11: 570579, 1993.
[15] Cella, D.F. et al., Reliability and validity of the Functional Assessment
of Cancer TherapyLung (FACT-L) quality of life instrument, Lung
Cancer, 12: 199220, 1995.
[16] Cella, D.F., Bonomi, A.E., Measuring quality of life: 1995 update. On-
cology, 9 11(supplement) 4760, 1995.
[17] Cella, D.F., Personal communication.
[18] Cohen, J., Statistical Power Analysis for the Behavioral Sciences, 2nd
ed., Lawrence Erlbaum Associates, Hillsdale, NJ, 1988.
[19] Cook, N.R., An imputation method for nonignorable missing data in
studies of blood pressure, Stat. Med., 16: 27132728, 1997.
[20] Cox, D.R. et al., Quality-of-life assessment: can we keep it simple? (with
discussion), J. R. Stat. Soc. A, 155: 353393, 1992.
[21] Crawford, S.L., Tennstedt, S.L., and McKinlay, J.B., A comparison of
analytic methods for non-random missingness of outcome data, J. Clin.
Epidemiol., 48: 209219, 1995.
[22] Curren, D. et al., Summary measures and statistics in the analy-
sis of quality of life data: an example from an EORTC-NCIC-SAKK
locally advanced breast cancer study, Eur. J. Cancer, 36: 834844,
2000.
[23] Curren, D., Analysis of Incomplete Longitudinal Quality of Life Data,
doctoral disertation, Linburgs Universitair Centrum, June 2000.
[24] Center for Drug Evaluation and Research, Food and Drug Administra-
tion, E9 Statistical Principles for Clinical Trials, Fed. Regis., 63: 49584
49598, 1998.
[25] Converse, J.M. and Presser, S., Survey Questions: Handcrafting the
Standardized Questionnaire. Survey Research Methods., Sage Publica-
tions, Newbury Park, CA, 1986.
[26] Daniels, M.J. and Hogan, J.W., Reparameterizing the pattern mixture
model for sensitivity analyses under informative dropout, Biometrics,
56: 12411248, 2000.
[27] Dawson, J.D. and Lagakos, S.W., Analyzing laboratory marker changes
in AIDS clinical trials, J. Acq. Immune Deciency Syndr., 4: 667676,
1991.
[28] Dawson, J.D. and Lagakos, S.W., Size and power of two-sample tests of
repeated measures data, Biometrics, 49: 10221032, 1993.
[29] Dawson, J.D., Stratication of summary statistic tests according to miss-
ing data patterns, Stat. Med., 13: 18531863, 1994.
[30] DeGruttola, V. and Tu, X.M., Modeling progression of CD-4 lymphocyte
count and its relationship to survival time, Biometrics, 50: 10031014,
1994.
[31] DeKlerk, N.H., Repeated warnings re repeated measures, Aust. N.Z. J.
Med., 16: 637638, 1986.
c2638 C2638texRef February 7, 2002 12:29
REFERENCES 287
[32] Dempster, A.P., Laird, N.M., and Rubin, D.B., Maximum likelihood
estimation from incomplete data via the EM algorithm (with discussion),
J. R. Stat. Soc. Ser. B, 39: 138, 1977.
[33] Devellis, R.F., Scale Development: Theory and Application, Sage Publi-
cations, Newbury Park, CA, 1991.
[34] Diehr, P. et al., Including deaths when measuring health status over
time, Med. Care, 33(suppl.): AS164172, 1995.
[35] Diggle, P.J., Liang, K.Y., and Zeger, S.L., Analysis of Longitudinal Data,
Oxford Science Publications, Clarendon Press, Oxford, U.K., 1994.
[36] Diggle, P.J. and Kenward, M.G., Informative dropout in longitudinal
data analysis (with discussion), Appl. Stat., 43: 4993, 1994.
[37] Fairclough, D.L. and Cella, D.F., A cooperative group report on quality
of life research: lessons learned. Eastern Cooperative Oncology Group
(ECOG), J. Natl. Cancer Inst., 40: 7375, 1996.
[38] Fairclough, D.L. and Cella, D.F., Functional assessment of cancer ther-
apy (FACT-G): non-response to individual questions, Qual. Life Res.,
5: 321329, 1996.
[39] Fairclough, D.L., Summary measures and statistics for comparison of
quality of life in a clinical trial of cancer therapy, Stat. Med., 16: 1197
1209, 1997.
[40] Fairclough, D.L., Method of analysis for longitudinal studies of health-
related quality of life, in Quality of Life Assessment in Clinical Trials:
Methods and Practice, Staquet, M.J., Hayes, R.D., and Fayers, P.M.,
Eds., Oxford University Press, New York, 1998, chap. 13.
[41] Fairclough, D.L., Peterson, H., and Chang, V., Why is missing quality
of life data a problem in clinical trials of cancer therapy? Stat. Med., 17:
667678, 1998.
[42] Fairclough, D.L. et al., Comparison of model based methods dependent
of the missing data mechanism in two clinical trials of cancer therapy,
Stat. Med., 17: 781796, 1998.
[43] Fairclough, D.L. et al., For the Eastern Cooperative Oncology Group,
Quality of life and quality adjusted survival for breast cancer patients
recieving adjuvant therapy, Qual. Life Res., 8: 723731, 1999.
[44] Fayers, P.M. and Hand, D.J., Factor analysis, causal indicators and qual-
ity of life, Qual. Life Res., 6: 139150, 1997.
[45] Fayers, P.M. and Machin, D., Quality of Life: Assessment, Analysis and
Interpretation, John Wiley & Sons, New York, 2000.
[46] Feeny, D. et al., A comprehensive multi-attribute system for classifying
the health status of survivors of childhood cancer, J. Clin. Oncol., 10:
923928, 1992.
[47] Fetting, J. et al., CAF vs. a 16-week multidrug regimen as adjuvant
therapy for receptor-negative, node positive breast cancer: an intergroup
study, Proc. Am. Soc. Clin. Oncol., 114(83), 1995.
[48] Fetting, J.J. et al., A 16-week multidrug regimen versus cyclophos-
phamide, doxorubicin and 5-urouracil as adjuvant therapy for
c2638 C2638texRef February 7, 2002 12:29
288 REFERENCES
node-positive, receptor negative breast cancer: an intergroup study, J.
Clin. Oncol., 16: 23822391, 1998.
[49] Fleiss, J.L., The Design and Analysis of Clinical Experiments, John
Wiley & Sons, New York, 1986.
[50] Floyd, J. and Fowler, J.R., Survey Research Methods, Sage Publications,
Newbury Park, CA, 1988.
[51] Follman, D. and Wu, M., An approximate generalized linear model with
random eects for informative missing data, Biometrics, 51: 151168,
1995.
[52] Frank-Stromborg, M. and Olsen, S., Instruments for Clinical Health
Care Research, 2nd ed., Jones & Bartlett, Boston, 1997.
[53] Friedman, L.M., Furberg, C.D., and Demets, D., Fundamentals of Clin-
ical Trials, 2nd ed., John Wright, Boston, 1985.
[54] Frison, L. and Popcock, S.J., Repeated measures in clinical trials: anal-
ysis using mean summary statistics and its implications for design, Stat.
Med., 11: 16851704, 1992.
[55] Gelber, R.D., Gelman, R.S., and Goldhirsh, A., A quality of life oriented
endpoint for comparing therapies, Biometrics, 45: 781795, 1998.
[56] Glasziou, P.P., Simes, R.J., and Gelber, R.D., Quality adjusted survival
analysis, Stat. Med., 9: 12591276, 1990.
[57] Glynn, R.J., Laird, N.M., and Rubin, D.B., Multiple imputation in mix-
ture models for non-ignorable nonresponse with follow-ups, J. Am. Stat.
Assoc., 88: 984993, 1993.
[58] Goldhirsch, A. et al., Costs and benets of adjuvant therapy in breast
cancer: a quality adjusted survival analysis, J. Clin. Oncol., 7: 3644,
1989.
[59] Gotay, C.C. et al., Building quality of life assessment into cancer treat-
ment studies, Oncology, 6: 2528, 1992.
[60] Gould, A.L., A new approach to the analysis of clinical drug trials with
withdrawals, Biometrics, 36: 721727, 1980.
[61] Guyatt, G.H. et al., A comparison of Likert and visual analogue scales
for measuring change in function, J. Chronic Dis., 40: 11291133,
1987.
[62] Guyatt, G. et al., Glossary, Controlled Clin. Trials, 12: 274S280S, 1991.
[63] Heitjan, D.A. and Landis, J.R., Assessing secular trends in blood pres-
sure: a multiple imputation approach, J. Am. Stat. Assoc., 89: 750759,
1994.
[64] Heyting, A., Tolbomm, T.B.M., and Essers, J.G.A., Statistical handling
of dropouts in longitudinal clinical trials, Stat. Med., 11: 20432061,
1992.
[65] Hicks, J.E. et al., Functional outcome update in patients with soft tissue
sarcoma undergoing wide local excision and radiation (Abstr.), Arch.
Phys. Med. Rehabil., 66: 542543, 1985.
[66] Hogan, J.W. and Laird, N.M., Mixture models for the joint distribu-
tion of repeated measuress and event times, Stat. Med., 16: 239257,
1997.
c2638 C2638texRef February 7, 2002 12:29
REFERENCES 289
[67] Hogan, J.W. and Laird, N.M., Model-based approaches to analyzing
incomplete longitudinal and failure time data, Stat. Med., 16: 259272,
1997.
[68] Hollen, P.J. et al., A dilemma in analysis: issues in serial measurement
of quality of life in patients with advanced lung cancer, Lung Cancer,
18: 119136, 1997.
[69] Holm, S., A simple sequentially rejective multivariate test procedure,
Scan. J. Statis., 6: 6570, 1979.
[70] Hommel, G., Tests of overall hypothesis for arbitrary dependence struc-
tures, Biom. J., 25: 423430, 1983.
[71] Hopwood, P. et al., Survey of the administration of quality of life ques-
tionnaires in three multicentre randomised trials in cancer, Eur. J. Can-
cer, 90: 4957, 1997.
[72] H urny, C. et al., Quality of life measures for patients receiving adjuvant
therapy for breast cancer: an international trial, Eur. J. Cancer, 28:
118124, 1992.
[73] H urny, C. et al., The perceived adjustment to chronic illness scale
(PACIS): a global indicator of coping for operable breast cancer patients
in clinical trials, Supportive Care Cancer, 1: 200208, 1993.
[74] H urny, C. et al., Impact of adjuvant therapy on quality of life in node-
positive patients with operable breast cancer, Lancet, 347: 12791284,
1996.
[75] Jennrich, R. and Schluchter, M., Unbalanced repeated-measures models
with structured covariance matrices, Biometrics, 42: 805820, 1986.
[76] Jones, R.H. Longitudinal Data with Serial Correlation: A State-Space
Approach, Chapman & Hall, London, 1993.
[77] Kaplan, R.M. and Bush, J.W., Health-related quality of life measure-
ment for evaluation research and policy analysis, Health Psychology, 1:
6180, 1982.
[78] Kenward, M.G. and Roger, J.H., Small sample inference for xed ef-
fects from restricted maximum likelihood, Biometrics, 53: 983997,
1997.
[79] Korn, E.L. and OFallon, J., Statistical Considerations, Statistics Work-
ing Group. Quality of Life Assessment in Cancer Clinical Trials, Report
on Workshop on Quality of Life Research in Cancer Clinical Trials, Di-
vision of Cancer Prevention and Control, National Cancer Institute,
1990.
[80] Korn, D.L., On estimating the distribution function for quality of life in
cancer clinical trials, Biometrika, 80: 535542, 1993.
[81] Kosinski, M. et al., Improving estimates of SF-36 health survey scores
for respondents with missing data, Med. Outcomes Trust Monitor, 5(1):
810, 2000.
[82] Laird, N.M. and Ware, J.H., Random-eects models for longitudinal
data, Biometrics, 38: 963974, 1982.
[83] Lauter, J., Exact t and F tests for analyzing studies with multiple end-
points, Biometrics, 52: 964970, 1996.
c2638 C2638texRef February 7, 2002 12:29
290 REFERENCES
[84] Lavori, P.W., Dawson, R., and Shera, D., A multiple imputation strategy
for clinical trials with truncation of patient data, Stat. Med., 14: 1912
1925, 1995.
[85] Lee, C.W. and Chi, K.N., The standard of report of health-related qual-
ity of life in clinical cancer trials, J. Clin. Epidemiol., 53: 451458, 2000.
[86] Levine, M. et al., Quality of life in stage II breast cancer: an instrument
for clinical trials, J. Clin. Oncol., 6: 17981810, 1988.
[87] Littel, R.C. et al., SAS System for Mixed Models, SAS Institute, Inc.,
Cary, NC, 1996.
[88] Little, R.J. and Rubin, D.B., Statistical Analysis with Missing Data,
John Wiley & Sons, New York, 1987.
[89] Little, R.J.A., A test of missing completely at random for multivariate
data with missing values, J. Am. Stat. Assoc., 83: 11981202, 1988.
[90] Little, R.J.A., Pattern-mixture models for multivariate incomplete data,
J. Am. Stat. Assoc., 88: 125134, 1993.
[91] Little, R.J.A., A class of pattern-mixture models for multivariate incom-
plete data, Biometrika, 81: 471483, 1994.
[92] Little, R.J.A., Modeling the dropout mechanism in repeated-measures
studies, J. Am. Stat. Assoc., 90: 11121121, 1995.
[93] Little, R.J.A. and Wang, Y., Pattern-mixture models for multivariate
incomplete data with covariates, Biometrics, 52: 98111, 1996.
[94] Little, R. and Yau, L., Intent-to-treat analysis for longitudinal studies
with dropouts, Biometrics, 52: 13241333, 1996.
[95] McDowell, I. and Newell, C., Measuring Health: A Guide to Rating Scales
and Questionnaires, 2nd ed., Oxford University Press, New York, 1996.
[96] Marcus, R., Peritz, E., and Gabriel, K.R., On closed testing procedures
with special reference to ordered analysis of variance, Biometrika, 63:
655660, 1976.
[97] Matthews, J.N.S. et al., Analysis of serial measurements in medical re-
search, Br. Med. J., 300: 230235, 1990.
[98] McNeil, B.J., Weichselbaum, R., and Pauker, S.G., Tradeos between
quality and quantity of life in laryngeal cancer, N. Engl. J. Med., 305:
982987, 1981.
[99] Meinert, C.L., Clinical Trials: Design, Conduct and Analysis, Oxford
University Press, New York, 1986.
[100] Michiels, B., Molenberghs, G., and Lipsitz, S.R., Selection models and
pattern-mixture models for incomplete data with covariates, Biometrics,
55: 978983, 1999.
[101] Miller, G.A., The magic number seven plus or minus two: some limits on
our capacity for information processing, Psychol. Bull., 63: 8197, 1956.
[102] Moinpour, C. et al., Quality of life endpoints in cancer clinical trials:
review and recommendations, J. Natl. Cancer Inst., 81: 485495, 1989.
[103] Mori, M., Woodworth, G.G., and Woolson, R.F., Application of empir-
ical Bayes inference to estimation of rate of change in the presence of
informative right censoring, Stat. Med., 11: 621631, 1992.
c2638 C2638texRef February 7, 2002 12:29
REFERENCES 291
[104] Motzer, R.J. et al., Phase III trial of interferon alfa-2a with or without
13-cis-retinoic acid for patients with advanced renal cell carcinoma, J.
Clin. Oncol., 18: 29722980, 2000.
[105] Murray, G.D. and Findlay, J.G., Correcting for the bias caused by
dropouts in hypertension trials, Stat. Med., 7: 941946, 1988.
[106] Neter, J. and Wasserman, W., Applied Linear Statistical Models, Richard
Irwin, Homewood, IL, 1974, 313317.
[107] Norton, N.J. and Lipsitz, S.R., Multiple imputation in practice: compar-
ison of software packages for regression models with missing variables,
Am. Stat., 55: 244254, 2001.
[108] OBrien, P.C., Procedures for comparing samples with multiple end-
points, Biometrics, 40: 10791087, 1984.
[109] Smith, D.M., Oswald: Object-Oriented Software for the Analysis of Lon-
gitudinal Data in S., 1996, available at http://www.maths.lancs.ac.uk/
Software/Oswald.
[110] Omar, P.Z. et al., Analysing repeated measurements data: a practical
comparison of methods, Stat. Med., 18: 15871608, 1999.
[111] Pater, J. et al., Eects of altering the time of administration and the
time frame of quality of life assessments in clinical trials: an example
using the EORTC QLQ-C30 in a large anti-emetic trial, Qual. Life Res.,
7: 273778, 1998.
[112] Patrick, D. and Erickson, P., Health Status and Health Policy: Allo-
cating Resources to Health Care, Oxford University Press, New York,
1993.
[113] Piantadosi, S., Clinical Trials: A Methodologic Perspective, John Wiley
& Sons, New York, 1997.
[114] Pledger, G. and Hall, D., Withdrawals from drug trials (letter to editor),
Biometrics, 38: 276278, 1982.
[115] SAS Proc Mixed, in The MIXED Procedure. SAS/STAT Software.
Changes and Enhancements (through Release 6.11), SAS Institute, Cary,
NC, 1996, chap. 18.
[116] Bernard, G., et al., Factors associated with noncompliance with self-
administered quality of life assessments of node-positive patients receiv-
ing adjuvant therapy for operable breast cancer, Stat. Med., 17: 587603,
1997.
[117] Pocock, S.J., Geller, N.L., and Tsiatis, A.A., The analysis of multiple
endpoints in clinical trials, Biometrics, 43: 487498, 1987.
[118] Pocock, S.J., Hughes, M.D., and Lee, R.J., Statistical problems in the
reporting of clinical trials, N. Engl. J. Med., 317: 426432, 1987.
[119] Proschan, M.A. and Waclawiw, M.A., Practical guidelines for multipli-
city adjustment in clinical trials, Controlled Clin. Trials, 21: 527539,
2000.
[120] Raboud, J.M. et al., Estimating the eect of treatment on quality of life
in the presence of missing data due to dropout and death, Qual. Life
Res., 7: 487494, 1998.
c2638 C2638texRef February 7, 2002 12:29
292 REFERENCES
[121] Revicki, D.A. et al., Imputing physical function scores missing owing to
mortality: results of a simulation comparing multiple techniquest, Med.
Care, 29: 6171, 2000.
[122] Ribaudo, H.J., Thompson, S.G., and Allen-Mersh, T.G., A joint analysis
of quality of life and survival using a random-eect selection model, Stat.
Med., 19: 32373250, 2000.
[123] Reitmeir, J. and Wassmer, G., Resampling-based methods for the anal-
ysis of multiple endpoints in clinical trials, Stat. Med., 18: 34553462,
1999.
[124] Rosenbaum, P.R. and Rubin, D.B., Reducing bias in observational stud-
ies using subclassication in the propensity score, J. Am. Stat. Assoc.,
79: 516524, 1984.
[125] Rubin, D.B. and Schenker, N., Multiple imputation for interval estima-
tion from simple random samples with ignorable nonresponse, J. Am.
Stat. Assoc., 81: 366374, 1986.
[126] Rubin, D.B., Multiple Imputation for Nonresponse in Surveys, John
Wiley and Sons, New York, 1987.
[127] Rubin, D.B. and Schenker, N., Multiple imputation in health-care data
bases: an overview and some applications, Stat. Med., 10: 585598,
1991.
[128] Rubin, D.B., Multiple imputation after 18+ years, J. Am. Stat. Assoc.,
91: 473489, 1996.
[129] Schafer, J., Analysis of Incomplete Multivariate Data, Monograph on
Statistics and Applied Probability 72, Chapman & Hall, London,
1997.
[130] Schipper, H. et al., Measuring the quality of life of cancer patients: The
Functional Living IndexCancer: development and validation, J. Clin.
Oncol., 2: 472482, 1984.
[131] Schipper, H., Guidelines and caveats for quality of life measurement in
clinical practice and research, Oncology, 4: 5157, 1990.
[132] Schluchter, M.D., 5V: unbalanced repeated measures models with struc-
tured covariance matrices, in BMDP Statistical Software Manual, 2,
Dixon, W.J., Ed., University of California Press, Berkeley, CA, 1990,
12071244, 13221327.
[133] Schluchter, M.D., Methods for the analysis of informatively censored
longitudinal data, Stat. Med., 11: 18611870, 1992.
[134] Schluchter, D.M., Green, T., and Beck, G.J., Analysis of change in the
presence of informative censoring: application to a longitudinal clinical
trial of progressive renal disease, Stat. Med., 20: 9891007, 2001.
[135] Schumacher, M., Olschewski, M., and Schulgen, G., Assessment of qual-
ity of life in clinical trials, Stat. Med., 10: 19151930, 1991.
[136] Schwartz, D., Flamant, R., and Lellouch, J., Clinical Trials, Academic
Press, London, 1980.
[137] Searle, S.R., Linear Models, John Wiley & Sons, New York, 1971, 4647.
[138] Simes, R.J., An improved Bonferroni procedure for multiple tests of
signicance, Biometrika, 73: 751754, 1986.
c2638 C2638texRef February 7, 2002 12:29
REFERENCES 293
[139] Simon, G.E. et al., SF-Summary Scores: are physical and mental health
truly distinct, Med. Care, 36: 567572, 1998.
[140] Solas for Missing Data Analysis 1.0 User Reference, Statistical Solutions
Ltd, Cork, Ireland, 1997.
[141] Spilker, B., Guide to Clinical Trials, Lippincott-Williams & Wilkins,
Baltimore, MD, 1991.
[142] Spilker, B., Ed., Quality of Life and Pharmacoeconomics in Clinical Tri-
als, Lippincott-Raven, Philadelphia, 1996.
[143] Schuman, H. and Presser, S., Questions and Answers in Attitude Sur-
veys: Experiments on Question Form, Wording and Context, Academic
Press, New York, 1981.
[144] Staquet, M. et al., Editorial: Health-related quality of life research, Qual.
Life Res., 1: 3, 1992.
[145] Staquet, M. et al., Guidelines for reporting results of quality of life as-
sessments in clinical trials, Qual. Life Res., 5: 496502, 1996.
[146] Staquet, M.J., Hayes, R.D., and Fayers, P.M., Quality of Life Assess-
ment in Clinical Trials: Methods and Practice, Oxford University Press,
New York, 1998.
[147] Streiner, D.L. and Norman, G.R., Health Measurement ScalesA Prac-
tical Guide to Their Development and Use, 2nd ed., Oxford Medical
Publications, Oxford University Press, New York, 1995.
[148] Sugarbaker, P.H. et al., Quality of life assessment of patients in extremity
sarcoma clinical trials, Surgery, 91: 1723, 1986.
[149] Tandon, P.K., Applications of global statistics in analyzing quality of
life data, Stat. Med., 9: 819827, 1990.
[150] Torrance, G.W., Thomas, W.H., and Sackett, D.L., A utility maximizing
model for evaluation of health care programs, Health Serv. Res., 7: 118
133, 1971.
[151] Troxel, A. et al., The generalized estimating approach when data are
not missing completely at random, J. Am. Stat. Assoc., 92: 13201329,
1997.
[152] Troxel, A. et al., Statistical analysis of quality of life data in cancer
clinical trials, Stat. Med., 17: 653666, 1997.
[153] VanBuuren, S., Boshuizen, H.C., and Knook, D.L., Multiple imputation
of missing blood pressure covariates in survival analysis, Stat. Med., 18:
681694, 1999.
[154] Verbeke, G. and Molenberghs, G., Eds., Linear Mixed Models in Prac-
tice: A SAS-Orientated Approach, Lecture Notes in Statistics, Springer,
New York, 1997.
[155] Verbeke, G. and Molenberghs, G., Linear Mixed Models for Longitudinal
Data, Springer, New York, 2000.
[156] Ware, J.E. et al., Choosing measures of health status for individuals in
general populations, Am. J. Public Health, 71: 620625, 1981.
[157] Ware, J.E. et al., SF-36 Health Survey: Manual and Interpretation
Guide, The Health Institute, New England Medical Center, Boston,
1993.
c2638 C2638texRef February 7, 2002 12:29
294 REFERENCES
[158] Ware, J., Kosinski, M., and Keller, S.D., SF-36 Physical and Mental
Component Summary Scales: A Users Manual, The Health Institute,
New England Medical Center, Boston, 1993.
[159] Weeks, J., Quality-of-life assessment: performance status upstaged? J.
Clin. Oncol., 10: 18271829, 1992.
[160] Wei, L.J. and Johnson, W.E., Combining dependent tests with incom-
plete repeated measurements, Biometrika, 72: 359364, 1985.
[161] Westfall, P.H. and Young, S.S., p-value adjustment for multiple testing
in multivariate bionomial models, J. Am. Stat. Assoc., 84: 780786, 1989.
[162] Wiklund, I., Dimen as, E., and Wahl, M., Factor of importance when
evaluating quality of life in clinical trials, Controlled Clin. Trials, 11:
169179, 1990.
[163] Wilson, I.B. and Cleary, P.D., Linking clinical variables with health-
related quality of lifea conceptual model of patient outcomes, J. Am.
Med. Assoc., 273: 5965, 1995.
[164] World Health Organization, Constitution of the World Health Organi-
zation, Basic Documents, WHO, Geneva, 1948.
[165] World Health Organization, The First Ten Years of the World Health
Organization, WHO, Geneva, 1958.
[166] Wu, M.C. and Bailey, K.R., Analyzing changes in the presence of infor-
mative right censoring caused by death and withdrawal, Stat. Med., 7:
337346, 1988.
[167] Wu, M.C. and Bailey, K.R., Estimation and comparison of changes in
the presence of informative right censoring: conditional linear model,
Biometrics, 45: 939955, 1989.
[168] Wu, M.C. and Carroll, R.J., Estimation and comparison of changes in
the presence of informative right censoring by modeling the censoring
process, Biometrics, 44: 175188, 1988.
[169] Yabro, K.R., Linas, B.P., and Schulman, K., Evaluation of quality of
life for diverse patient populations, Breast Cancer Res. Treatment, 40:
87104, 1996.
[170] Yao, Q., Wei, L.J., and Hogan, J.W., Analysis of incomplete repeated
measurements with dependent censoring times, Biometrika, 85: 139149,
1998.
[171] Young, T. and Maher, J., Collecting quality of life data in EORTC
clinical trialswhat happens in practice? Psycho-oncology, 8: 260263,
1999.
[172] Zeger, S.L. and Liang, K.Y., An overview of methods for the analysis of
longitudinal data, Stat. Med., 11: 18251839, 1992.
[173] Zhang, J. et al., Some statistical methods for multiple endpoints in clin-
ical trials, Controlled Clin. Trials, 18: 204221, 1997.
[174] Zwinderman, A.H., The measurement of change of quality of life in clin-
ical trials, Stat. Med., 9: 931942, 1990.

Index

13-cis -retinoic acid (13-CRA), 1518

A

Aaronson, N.K., 1
ABB (Approximate Bayesian bootstrap),
142146;

see also

Covariance
structure; Multiple
imputation
ACMV restriction,

see

Available case
missing value (ACMV)
restriction
Adjuvant breast cancer study, 710
appropriateness of, 30
global tests, 246
multivariate analyses, 247248,
248249
repeated measures cell means model,
4647
selection between models, 42
timing of, 25
Administration of assessments, 3132
Age, and missing data, 10, 11, 47, 48, 76,
77, 86
Akaikes Information Criterion (AIC), 46;


see also

Models
Alpha adjustments, 249253
Analysis plans, 257274, 260261;

see
also

Multiple endpoints;
Summary measures;
Summary statistics
general analysis plan, 258259
introduction, 257258
models for longitudinal data, 259260
multiple comparisons, 244, 261262
reporting results, 272273
sample size and power, 262272
basic assumptions, 264
growth curve model, 268272
incomplete designs, 264266
repeated measures example,
266268
simple linear combinations of ,
262263
secondary endpoints, 260261, 263
Analytic models, vs. imputation, 151
Approximate Bayesian bootstrap (ABB),
142146;

see also

Covariance
structure; Multiple
imputation
Area under the curve (AUC), 225227
with missing data handled by multiple
imputation, 250
and summary statistics across time,
231235
ARMA (autoregressive moving average),
50
Assessments,

see

Study design and
protocol development
AT MEANS option, 107108
AUC,

see

Area under the curve
Autoregressive covariance structure, 45,
49, 50
Autoregressive error structure, 58, 60
Autoregressive moving average (ARMA),
50
Available case missing value (ACMV)
restriction, 164, 167, 169;

see
also

Pattern mixture models
Available data,

see

Missing at random
(MAR)
Average of ranks, 227
Average rate of change (slopes), 222

C2638_frame_IDX Page 295 Tuesday, February 12, 2002 2:21 PM

296 INDEX

B

Bailey, K.R., 183, 184, 185, 186, 192
Barnard, J., 131, 148
Baseline
adding other covariates, 110112
assessment as covariate, 105108
change from, 108110
Bayes estimates, empirical, 112113,
210211, 223224
Bayesian bootstrap, approximate (ABB),
142146
Bayesian Information Criterion (BIC), 46
BCQ (Breast Chemotherapy
Questionnaire), 8, 30, 246
Best linear unbiased estimate (BLUE),
113
Best linear unbiased predictor (BLUP),
113, 141
Between imputation variance, 132, 147,
150
BIC (Bayesian Information Criterion ), 46
Biologic response modifiers (BRM), 1516
Bivariate correlation, 76
Bivariate data, 155161
Browns protective restrictions,
158160
complete-case missing variable
(CCMV) restriction, 156158
non-small-cell lung cancer (NSCLC)
study example, 156, 158
sensitivity analyses with intermediate
restrictions, 160161
BLUE (best linear unbiased estimate),
113
BLUP (best linear unbiased predictor),
113
Bonferroni adjustment, 215, 249250,
253254
Bonomi, A.E., 2
Bootstrap procedure, 192, 254
Bouchet, C., 27
Breast cancer study, adjuvant,

see

Adjuvant breast cancer study
Breast Chemotherapy Questionnaire
(BCQ), 8, 30, 246
BRM (biologic response modifiers), 1516
Brown, C.H., 158160
Browns protective estimate, 259
Browns protective restrictions, 158160;


see also

Pattern mixture
models
Buck's conditional means, 123
Building models,

see

Models
Bush, J.W., 2

C

CAF (Cyclophosphamide,doxirubicin,and
5-flurouracil) treatment, 710
Cancer,

see

Adjuvant breast cancer study;
Non-small-cell lung cancer
(NSCLC)
Carcinoma study,

see

Renal cell
carcinoma study
Carroll, R.J., 210
Casewise deletion, 97
CCMV restriction,

see

Complete-case
missing value (CCMV)
restriction
Ceiling effect,

see

Validity of instruments
Cella, D.F., 2
Cell means model, 46, 5253;

see also

Repeated measures
Change points,

see

Growth curve models
Characteristics of patients,

see

Covariates
Cholesky decomposition, 134
Cisplatin, treatment with, 1012
Cleary, P.D., 12
Clinical trial endpoints,

see

Endpoints
Closed tests, 246248
Closest neighbor and predictive mean
matching, 140142
Complete-case missing value (CCMV)
restriction, 179180;

see also

Pattern mixture models
bivariate data, 156158
comparison of ACMV, NCMV and,
167, 169
monotone dropout, 162167
Complete cases, 73, 97101
Complete data, use of term, 73
Completely random dropout (CRD), 198,
201, 204
Compound symmetry, 44, 45, 49, 50, 59;


see also

Covariance structure

C2638_frame_IDX Page 296 Tuesday, February 12, 2002 2:21 PM

INDEX 297

Conditional linear model, 184, 185192
assumptions, 192
and designing analysis plan, 260
estimation of the standard errors, 192
non-small-cell lung cancer (NSCLC)
study example, 187192
random-coefficient mixture model,
192
testing MAR vs.MNAR under
assumptions of, 187
Conditional predicted values, 123125
Condition-driven designs, 23
Confirmatory tests, 243244
Construct validity, 29
CONTRAST statement., 53
Convergent validity, 29
Covariance structure, 5759;

see also

Repeated measures models
piecewise linear regression model,
6667
polynomial growth curve models,
5963
Covariates, 4650
baseline assessment as, 105108
dropout dependent on, 7576
time-varying, 270
COVB option, 174
COVTEST option, 195
CRD (completely random dropout), 198,
201, 204
Criterion validity, 29
Cubic terms, 54
Cultures, and instrument
appropriateness, 30
Curren, D., 161, 217
Cut-off test statistics, 254
Cyclic therapy, 25
Cyclophosphamide,doxirubicin,and 5-
flurouracil (CAF) treatment,
710
Cytokine treatment, 15

D

Daniels, M.J., 154, 178
Data collection and management, 32, 34,
35
Dawson, J.D., 228
DeGruttola, V., 184, 194
Delta method, 173174
Designing analysis plan,

see

Analysis
plans
Diggle, P.J., 184, 198, 202, 204, 210, 260
Discontinuation of therapy, assessment
after, 27
Disease-specific instruments, 4;

see also

Measures
Divergent validity, 29
Domains, HRQoL, 235240
Dropout;

see also

Missing data; Monotone
dropout; Nonmonotone
dropout
informative, 199, 201, 205, 206207
intermittent, 7172, 146147
joint mixed-effects and time to,

see

Joint mixed-effects model
model for, 208210
random, 201, 204, 205
time of, 185187
use of term, 69
Duration of assessments, 2627

E

EBLUP (empirical best linear unbiased
predictors), 123
Economic evaluations, 28
EM algorithm, 140, 178, 283
Empirical Bayes estimates
of individual subject slopes, for
informative dropout, 210211
use with ignorable missing data, 112114
use with missing data, 223224
Empirical best linear unbiased predictors
(EBLUP), 123
Endpoints;

see also

Multiple endpoints
checklist for statistical analysis plan,
258
primary, 260261
secondary, 260261, 263
EORTC QLQ-C30 (European
Organization for Research
and Treatment of Cancer), 3,
5, 6;

see also

Measures
Erickson, P., 2
Error rates, 245,

see

Type I error rate;
Type II error rate

C2638_frame_IDX Page 297 Tuesday, February 12, 2002 2:21 PM

298 INDEX

Estimable functions, 113
Etoposide-cisplatin treatment, 1012
European Organization for Research and
Treatment of Cancer (EORTC
QLQ-C30), 3, 5, 6;

see also

Measures
Event-driven designs, 23, 259260
Exclusion, of observations/subjects,
104105
Experimentwise error rates, 245
Explanatory investigational aims, 2122
Explicit regression models, 118125
Explicit univariate regression
assumptions, 135136
computation of imputed values, 134
extensions to longitudinal studies, 135
non-small-cell lung cancer (NSCLC)
study, 136140
Extensions to longitudinal studies, 144

F

Face validity, 29
FACT,

see

Functional Assessment of
Cancer Therapy
FACT BRM,

see

Functional Assessment
of Cancer Therapy -Biologic
Response Modifiers (FACT
BRM)
FACT-G (Functional Assessment of
Cancer Therapy Version 3),
1518
FACT-L (Functional Assessment of
Cancer Therapy,Lung Cancer
Version 2 ), 12
FACT-Lung Trial Outcome Index (FACT-
Lung TOI), 12, 7880, 8586
Factor-analytic weights, 3637, 236237
FACT-TOI (FACT Trial Outcome Index),
55, 198, 200
Fairclough, D.L., 130
Familywise error rates, 245
Fayers, P.M., 36, 225
Findlay, J.G., 78
Fixed effects, 57;

see also

Mixed-effects
models
FLIC (Functional Living Index -Cancer),
3
Floor effect, 30;

see also

Validity of
instruments
Follow-up HRQoL assessments, 24
Follow-up procedures, explicit, 34
Frequency of assessments, 2526
Functional Assessment of Cancer
Therapy (FACT), 3
global index vs. profile of domain-
specific measures, 5
period of recall, 6
sample questions, 12, 36
Functional Assessment of Cancer
Therapy,Lung Cancer Version
2 (FACT-L), 12
Functional Assessment of Cancer
Therapy -Biologic Response
Modifiers (FACT BRM)
renal cell carcinoma study, 87, 89, 90,
198, 199
sample questions from, 16
standard deviation of baseline, 235
summary measures, 236
Functional Assessment of Cancer
Therapy Version 3 (FACT-G),
1518
Functional Living Index -Cancer (FLIC),
3

G

General analysis plan, 258259
Generic instruments, 4, 28
Global index of HRQoL, 45
Global tests, 246
Gould, A.L., 128
GROUP parameter, 206
Growth curve models, 5467;

see also

Mixed-effects models;
Piecewise linear regression
and area under the curve (AUC),
232233
covariance structure, 5759
and designing analysis plan, 260
model for means, 5457
for non-small-cell lung cancer
(NSCLC) study, 43
for renal cell carcinoma study, 43
sample size and power, 268272

C2638_frame_IDX Page 298 Tuesday, February 12, 2002 2:21 PM

INDEX 299

SAS example of piecewise linear
regression model, 6567
SAS example of polynomial growth
curve model, 5965
Guttman scale, 38
Guyatt, G., 27

H

Half rule, 3738
Hall, D., 128
Hand, D.J., 36
Health, defined, 1
Health-related quality of life (HRQoL),
use of term, 2
Health status assessment, 24
Heckman probit stochastic dropout
model, 210
Heterogeneous autoregressive covariance
structure, 45, 49, 50
Heterogeneous compound covariance
structure, 45, 49, 50
Heterogeneous Toplitz covariance
structure, 45, 49, 50
Heyting, A., 128
Higher-order terms, 54
Hogan, J.W., 84, 154, 178, 195
Hommel,G., 252
Hopwood, P., 31, 3435
Hotelling-Lawley trace, 246
Hypothesis testing
building growth curve models, 6465
building repeated measures models,
5253
pattern-mixture model, 177
polynomial growth curve models,
6465

I

ID (informative dropout), 199, 201, 205
Ignorable missing data, 93114;

see also

Imputation; Missing data
adding other baseline covariates,
110112
baseline assessment as a covariate,
105108
change from baseline, 108110
and designing analysis plan, 259
empirical Bayes estimates, 112113
introduction, 9394
multivariate methods, 96105
complete case analysis, 97101
exclusion of observations, 105
exclusion of subjects, 104105
maximum likelihood estimation
with all available data,
101104
repeated univariate analyses, 9496
Imputation,

see

Multiple imputation;
Simple imputation
Incomplete cases,

see

Missing data
Individual tests, vs. global, 246
Information criterion, 46, 49, 52
Informative dropout (ID), 199, 201, 205
informative missing data,

see

Missing
data
Initial assessments,

see

Analysis plans
Instruments
defined, 27
disease-specific instruments, 4
generic, 4
scoring, 3539
selecting, 2730
validity and reliability, 2930
Intent-to-treat analysis, 22, 257;

see also

Research objectives
Interferon alpha-2, 1518
Intermittent dropout, 7172, 146147
Intermittent missing data, 210, 270
International assessments, 30
Interpolation, 221
Interviews, for assessments, 31
Inverse correlation, 238239
Investigation aims,

see

Research
objectives
Item-nonresponse,

see

Missing data

J

Joint mixed-effects model, 184, 185,
193198
and designing analysis plan, 260

C2638_frame_IDX Page 299 Tuesday, February 12, 2002 2:21 PM

300 INDEX

extension to more complex mixed-
effects models, 198
initial estimates, 197
non-small-cell lung cancer (NSCLC)
study example, 195197
renal cell carcinoma example, 198
testing MAR vs.MNAR under the
assumptions of, 194195

K

Kaplan, R.M., 2
Kenward, M.G., 184, 198, 202, 204, 210,
260, 272

L

Lagakos, S.W., 228
Laird, N.M., 84, 195
Languages, and instrument
appropriateness, 30
Large-sample inferences for

2

, 160161
Last observation carried forward (LOCF),
125128
Last value carried forward (LVCF),
125128, 223, 224
Lavori, P.W., 144, 145
Likelihood ratio tests, 4446
Likert scale, 6, 36
Linear minimum variance unbiased
estimator (LMVUB), 186
Lipsitz, S.R., 131
Listwise deletion, 97
Little, R., 135
Little, R.J.A., 79, 81, 84, 87, 160, 161, 178,
195
LMVUB (linear minimum variance
unbiased estimator), 186
LOCF (last observation carried forward),
125128
Logistic selection model, 185
Longitudinal studies, models for,

see

Models
LSMEANS option, 111
LSMEANS statement, 107108
Lung cancer study,

see

Non-small-cell
lung cancer (NSCLC)
LVCF (last value carried forward),
125128, 223, 224

M

Machin, D., 225
Macros, for scoring, 3839
M analyses, combining, 147150
MANOVA statement, 98
Marcus, R., 247
Marginal model, 198
Markov Chain Monte Carlo (MCMC)
method, 147
Maximization converged message, 210
Maximum likelihood (ML) ratio test, 44
Maximum likelihood estimation (MLE),
97
MCMC (Markov Chain Monte Carlo)
method, 147
MCS (Mental component scale), 236237
Mean value substitution, 117118
Measures;

see also

Models; Study design
and protocol development
Breast Chemotherapy Questionnaire
(BCQ), 8, 30, 246
disease-specific instruments, 4
EORTC QLQ-C30 (European
Organization for Research
and Treatment of Cancer), 3,
5, 6
FACT-Lung Trial Outcome Index
(FACT-Lung TOI), 12, 7880,
8586
FACT-TOI (FACT Trial Outcome
Index), 55, 198, 200
generic instruments, 4, 28
global index vs. profile of domain-
specific measures, 46
Guttman scale, 38
half rule, 3738
health status vs. patient preferences,
24
Medical Outcomes Study SF-36, 7,
236237
multi-attribute, 34
objective vs. subjective assessments, 4

C2638_frame_IDX Page 300 Tuesday, February 12, 2002 2:21 PM

INDEX 301

period of recall, 67
response format, 6
reverse coding, 35
scoring, 3539, 258
standard gamble utilities, 3
summated ratings, 3536
time trade-off (TTO), 222
utility, 222
Medical Outcomes Study SF-36, 7,
236237;

see also

Measures
Meng, X.L., 131
Mental component scale (MCS), 236237
Metastatic disease (SX_MET ), 188
MI (multiple imputation), 223
Missing at random (MAR)
analytic methods, 8081
identification of dependence on
observed data, 7880
vs. MCAR, 8183
vs. MNAR, 194195
overview, 74
Missing completely at random (MCAR),
7577;

see also

Missing data
formal definitions for, 282283
vs. MAR, 8183
overview, 74
Missing data, 6991;

see also

Area under
the curve (AUC); Complete-
case missing value (CCMV)
restriction; Ignorable missing
data; Imputation; Missing at
random (MAR); Missing
completely at random
(MCAR); Missing not at
random (MNAR)
acceptable amount, 7071
and calculation of summary measures,
220222
documentation, 3435
formal definitions for, 280283
intermittent, 210, 270
notation, 7277
patterns of, 7172
preventing, 3235, 71
renal cell carcinoma study, 8790
similar patterns of dropout among
intervention arms, 71
use of term, 69
why problematic, 6970
Missing not at random (MNAR), 74,
8487;

see also

Missing data
formal definitions for, 283
vs. MAR under assumptions of joint
model, 187
Mixed-effects models, 57, 187188;

see
also

Covariance structure;
Pattern mixture models; SAS
examples
fixed effects, 57
random effects, 5758, 62, 183, 184
MIXED procedure, 112, 195
for cell means model, 4748
repeated measures model, 51
Mixture models, 84, 183;

see also

Joint
mixed-effects
ML (maximum likelihood) ratio test, 44
MLE (maximum likelihood estimation),
97
Models, 4168;

see also

Covariance
structure
analytic models, 4346
growth curve models, 5467
covariance structure, 5759
model for means, 5457
for non-small-cell lung cancer
(NSCLC) study, 43
for renal cell carcinoma study, 43
SAS example of piecewise linear
regression model, 6567
SAS example of polynomial growth
curve model, 5965
information criterion, 46, 49, 52
introduction, 4143
likelihood ratio tests, 4446
nested, 4446
repeated measures models, 4653
for adjuvant breast cancer study,
42
covariance structures, 4850
hypothesis testing, 5253
mean structure, 4648
SAS example of repeated measures
model, 5052
underidentified, 84
MODEL statement, 48, 89, 188, 189
MODEL statements, 59
Monotone dropout, 7172, 161167
comparison of CCMV,ACMV,and
NCMV estimates, 167

C2638_frame_IDX Page 301 Tuesday, February 12, 2002 2:21 PM

302 INDEX

complete-case missing value
restriction, 162167
non-small-cell lung cancer (NSCLC)
study example, 162
selection model for, 198210
Heckman probit stochastic dropout
model, 210
intermittent missing data, 210
nonparametric analyses, 211
non-small-cell lung cancer
(NSCLC) study example,
201205
Oswald program, 206210
outcome-dependent selection
model, 201
Mori, M., 210
Multi-attribute assessment measures,
34
Multi-item scales, scoring, 3537
Multiple comparisons, 244, 261262;

see
also

Summary measures;
Summary statistics
alpha adjustment, 249253
Bonferroni adjustment, 215, 249250,
253254
closed tests, 246248
global tests, 246
post hoc adjustment, 243
p -value adjustments, 253254
Rgers inequality, 252
Simes' global test, 252
Stepdown procedure, 247
Multiple endpoints, 243255
background concepts and definitions,
245246
and designing analysis plan, 260262
introduction, 243244
limiting the number of confirmatory
tests, 243244
multiple comparison procedures, 244
multivariate statistics, 246249
resampling techniques, 254
summary measures and statistics, 244
univariate statistics, 249254
alpha adjustments for K univariate
tests, 249253
non-small-cell lung cancer
(NSCLC) study example, 249
sequential rejective Bonferroni
procedure, 253254
Multiple imputation (MI), 131152, 223,
230;

see also

Pattern mixture
models
analysis of M data sets, 132
approximate Bayesian bootstrap,
142146
closest neighbor and predictive mean
matching, 140142
combining M analyses, 132, 147150
explicit univariate regression
assumptions, 135136
computation of imputed values,
134
extensions to longitudinal studies,
135
identification of imputation model,
133134
non-small-cell lung cancer
(NSCLC) study, 136140
generation of M imputed data sets,
132
implications for design, 152
imputation vs.analytic models, 151
introduction, 131
multivariate procedures for
nonmonotone missing data,
146
predictive mean matching, 140142
propensity scores, 144
selection of procedures, 131132
sensitivity analyses, 150151
Multivariate analysis of variance, 97101
Multivariate logistic regression, 76
Multivariate methods, 96105
complete case analysis, 97101
exclusion of observations, 105
exclusion of subjects, 104105
maximum likelihood estimation with
all available data, 101104
Multivariate model, 48
Multivariate statistics, 246249
Murray, G.D., 78

N

NCMV (neighboring case missing value)
restriction, 164167
Negative quadratic terms, 191

C2638_frame_IDX Page 302 Tuesday, February 12, 2002 2:21 PM

INDEX 303

Neighboring case missing value (NCMV)
restriction, 164167, 169;

see
also

Pattern mixture models
Nested models, 4446
Non-ignorable missing data, 145146,
183184, 259;

see also

Conditional linear model;
Joint mixed-effects; Mixture
models
Nonmonotone dropout, 7172, 146147
Nonparametric analyses, 178, 211
Nonparametric van der Waerden scores,
228
Non-small-cell lung cancer (NSCLC)
study, 1015
average rate of change (slopes), 222
baseline assessment as covariate,
107108
complete case analysis (MANOVA),
99101
conditional linear model, 187192
constructing summary measures,
229230
explicit univariate regression,
136140
maximum likelihood estimation,
102104
missing data, 72, 73, 76, 8587
monotone dropout, 162
multivariate procedures for
nonmonotone missing data,
146147
pattern mixture models, 154155, 156,
158
repeated univariate analyses, 96
selection between models for
assessment, 43
selection model for monotone dropout,
201205
summary measures, 229230
underestimation of variance, 129130
univariate statistics, 249
use of imputation, 116117
Norton, N.J., 131
NSCLC study,

see

Non-small-cell lung
cancer; Non-small-cell lung
cancer (NSCLC)

O

OBrien, P C., 238, 239
Objective assessments, 4
Objectives of research, 2022, 2627
Observations, exclusion of, 105
ODS statement, 83
Ordinary least squares (OLS), 213214,
222, 223
Oswald program, 206210
Outcome-dependent selection models,
184, 201
Overidentified models, 178

P

Paper-and-pencil self-reports, 31
Parametric models, 167177
with informative dropout, 206207
linear trends, 168172
SAS example of pattern-mixture
model, 174177
variance estimation, 172174
Patient characteristics,

see

Covariates
Patient preference assessment, 24
Patient preferences, 24
Patient weights, 237238
Patrick, D., 2
Pattern mixture models, 84, 153181,
183, 260
additional reading, 178
algebraic details, 178180
bivariate data, 155161
Browns protective restrictions,
158160
complete-case missing variable
(CCMV)restriction, 156158
non-small-cell lung cancer
(NSCLC) study example, 156,
158
sensitivity analyses with
intermediate restrictions,
160161
introduction, 153155
monotone dropout, 161167
comparison of CCMV,ACMV,and
NCMV estimates, 167

C2638_frame_IDX Page 303 Tuesday, February 12, 2002 2:21 PM

304 INDEX

complete-case missing value
restriction, 162167
non-small-cell lung cancer
(NSCLC) study example, 162
parametric models, 167177
linear trends, 168172
SAS example of pattern-mixture
model, 174177
variance estimation, 172174
SAS example of, 174177
Pcmid function, 206207, 208
PCS (physical component scale), 236237
Period of recall, 67
Personnel,

see

Staff
Physical component scale (PCS), 236237
Piecewise linear regression, 5457, 6667
Pledger, G., 128
Polynomial growth curve models, 54, 55,
56
SAS example of, 5965
covariance structure, 5963
estimation, 6364
fully parameterized model for the
means, 59
hypothesis testing, 6465
model reduction, 63
Pooling estimates, 176177
Post hoc adjustment, 243
Pragmatic investigational aims, 2122
Predictable functions, 113
Predictive mean matching, 140142
Primary endpoints, vs. secondary,
260261
Probit stochastic dropout model, 210
Propensity scores, 144
Protocol development,

see

Study design
and protocol development
P-value adjustments, 253254;

see also

Univariate analysis

Q

Quadratic terms, 54
Quality-adjusted life years, 28
Quality of life, use of term, 2
Questionnaires,

see

Instruments;
Measures

R

Raboud, J.M. et al., 130
Random-coefficient mixture model, 192
Random-coefficient selection models, 184,
195;

see also

Conditional
linear model
Random dropout (RD), 198199, 201, 204,
205
Random effects, 5758, 62, 183, 184;

see
also

Mixed-effects models
RANDOM statement, 59, 60, 61, 62
Rating,

see

Scoring
RCORR option, 51
RD (random dropout), 198199, 201, 204,
205
Recall, period of, 67
Regression models
explicit, 118125
assumptions, 135136
computation of imputed values,
134
extensions to longitudinal studies,
135
non-small-cell lung cancer
(NSCLC) study, 136140
univariate, 120122
Reitmeir, J., 254
Reliability of instruments, 2930
REML (restricted maximum likelihood )
estimates, 97101, 272
REML (restricted maximum likelihood )
ratio test, 44
Renal cell carcinoma study, 1518
area under the curve (AUC), 232, 234
covariance structures tested for, 61
extension to more complex mixed-
effects models, 198
missing data, 8790
piecewise linear model for, 5457
selection between models for, 43
summary measures, 221, 232233,
234
Repeated measures models, 4142, 4653
for adjuvant breast cancer study, 42
covariance structures, 4850
and designing analysis plan, 259260
hypothesis testing, 5253
mean structure, 4648

C2638_frame_IDX Page 304 Tuesday, February 12, 2002 2:21 PM

INDEX 305

SAS example of repeated measures
model, 5052
Repeated univariate analyses, 9496
Resampling techniques, 254
Research objectives, 2022, 2627
Residual errors, variance of, 58
Response format of questionnaires, 6
Restricted maximum likelihood (REML)
estimates, 97101, 272
Restricted maximum likelihood (REML)
ratio test, 44
Reverse coding, 35
Roger, J.H.,, 272
R option, 51
Rubin, D.B., 145146, 148
Rgers inequality, 252

S

Sampling-based imputation, 142146
SAS examples
growth curve models, 5965
covariance structure, 5963
estimation, 6364
fully parameterized model for the
means, 59
hypothesis testing, 6465
model reduction, 63
piecewise linear regression, 6567
pattern-mixture model, 174177
polynomial growth curve models,
5965
repeated measures models, 5052
scoring, 38
SAS FREQ procedure, 174
SAS IML procedure, 176
SAS macro variables, 174
Schenker, N., 145146
Schipper, H., 31
Schluchter, M.D., 183, 184, 194
Scoring
checklist for statistical analysis plan,
258
instruments for, 3539
SD (standard deviation), 129130
SE (standard error ), 129130
Secondary endpoints, 260261, 263
Selection models, 84, 184;

see also

Joint
mixed-effects
for monotone dropout, 198210
Heckman probit stochastic dropout
model, 210
intermittent missing data, 210
nonparametric analyses, 211
non-small-cell lung cancer
(NSCLC) study example,
201205
Oswald program, 206210
outcome-dependent selection
model, 201
Selection of subjects, 2223
Sensitivity analyses, 150151
Sensitivity analysis, 130, 160161, 258
Sequential rejective Bonferroni
procedure, 253254
Shared-parameter model, 195
Short-term fluctuations, minimizing, 6
Simes' global test, 252253
Simes, R.J., 252
Simes procedure, 253
Simon, G.E., 237
Simple imputation, 115130
explicit regression models, 118125
last value carried forward (LVCF),
125128
limitations of, 116
mean value substitution, 117118
non-small-cell lung cancer (NSCLC)
study, 116117
sensitivity analysis, 130
underestimation of variance, 128130
Small sample size approximations, 272
Square root method, 134
Staff, training to avoid missing data,
3435
Standard deviation (SD), 129130
Standard error (SE), 129130
Standard gamble utilities, 3
Statistics guiding model reduction, 4446
Stepdown procedure, 247
Stratified analysis, of summary
measures, 228229
Structured covariance, 4950
Study design and protocol development,
1939;

see also

Models
avoiding missing data, 3235
background and rationale, 1920

C2638_frame_IDX Page 305 Tuesday, February 12, 2002 2:21 PM

306 INDEX

data collection and management, 32
introduction, 19
longitudinal designs, 2327
mode of administration, 3132
research objectives, 2022
scoring instruments, 3539
selection of a quality of life measure,
2729
selection of subjects, 2223
Subjective assessments, 4
Subjects of assessments
exclusion of, 104105
selecting, 2223
unequal allocations to treatment
groups, 270271
Subscales, HRQoL, summarizing across,
235240
Summary measures, 213241;

see also

Summary statistics
addressing multiplicity of endpoints,
213
area under the curve, 240
choosing, 215217
combining HRQoL and time to event,
240
constructing, 217230
average of ranks, 227
average rate of change (slopes),
222227
non-small-cell lung cancer
(NSCLC) study example,
229230
notation, 220
when missing data, 220222
and multiple endpoints, 244
nonparametric procedures, 240
options for weighting, 235239
renal cell carcinoma study, 221,
232233, 234
stratified analysis of, 228229
strengths and weaknesses, 215
summarizing across HRQoL domains
or subscales, 235240
vs. summary statistics, 213214
and summary statistics, 213215
univariate analysis of, 228
weighting, 235239
Summary statistics, 213215;

see also

Summary measures
and multiple endpoints, 244
summarizing across HRQoL domains
or subscales, 239240
Summated ratings, 3536
SX_MET (metastatic disease), 188

T

Taylor series approximation, 173174
Terminal dropout,

see

Monotone dropout
Therapy, assessment after
discontinuation of, 27
Time-driven designs, 24, 260
Time-to-event mixture, 183
Time trade-off (TTO), 222
Time trade utilities, 3
Time-varying covariates, 270
Timing, of HRQoL assessments, 810
follow-up, 2425
initial, 24
Toplitz covariance structure, 45, 49, 50,
52
Tracking patients in studies, 32
Training research personnel, to avoid
missing data, 34
TRANSPOSE procedure, 98
Treatments, assessment after
discontinuation of, 27
Trial Outcome Index (FACT-BRM TOI),
16, 236
Trials,

see

Study design and protocol
development
Tu, X.M., 184, 194
Tumor response, 261
Two-sample t -test, 228
Type I error rate
causes of, 95, 110, 213, 243
controlling, 246, 248, 261, 266
and Simes' global test, 252
Type II error rate, 266

U

Underestimation of variance, 128130
Underidentified models, 84
Unequal allocations of subjects to
treatment groups, 270271

C2638_frame_IDX Page 306 Tuesday, February 12, 2002 2:21 PM

INDEX 307

Univariate analysis, 228
Univariate regression, 48, 120122
assumptions, 135136
computation of imputed values, 134
extensions to longitudinal studies, 135
non-small-cell lung cancer (NSCLC)
study, 136140
Univariate statistics, 249254
alpha adjustments for K univariate
tests, 249253
vs. multivariate, 245
non-small-cell lung cancer (NSCLC)
study example, 249
sequential rejective Bonferroni
procedure, 253254
use of term, 245
Unstructured covariance, 4849, 50
Utility measures, 222

V

Validity of instruments, 2930
VanBuuren, S., 146
Variance estimation, 128130, 172174
Variation in responses, 29
Visual analog scale (VAS), 6

W

Wang, Y., 178
Ware, J.E. et al., 28
Wassmer, G., 254
Weighting summary measures, 235239
Westfall, P.H., 254
WHO (World Health Organization), 1
Wilcoxon Rank Sum test, 128, 228
Wilson, I.B., 12
Within imputance variation,

see

Multiple
imputation (MI)
World Health Organization (WHO), 1
Wu, M.C., 183, 184, 185, 186, 192, 210

Y

Yao, Q., 178
Yau, L., 135, 211
Young, S.S., 254

Z

Zero, assigning to HRQoL scores, 222

C2638_frame_IDX Page 307 Tuesday, February 12, 2002 2:21 PM

C2638_frame_IDX Page 308 Tuesday, February 12, 2002 2:21 PM