The Statistical Analysis Of Multivariate Time

Data A Marginal Modeling Approach Prentice

download

https://ebookbell.com/product/the-statistical-analysis-of-
multivariate-time-data-a-marginal-modeling-approach-
prentice-10510216

Explore and download more ebooks at ebookbell.com

Here are some recommended products that we believe you will be
interested in. You can click the link to download.

The Statistical Analysis Of Multivariate Failure Time Data A Marginal

Modeling Approach Ross L Prentice

https://ebookbell.com/product/the-statistical-analysis-of-
multivariate-failure-time-data-a-marginal-modeling-approach-ross-l-
prentice-10441688

The Statistical Analysis Of Multivariate Failure Time Data Prentice Rl

https://ebookbell.com/product/the-statistical-analysis-of-
multivariate-failure-time-data-prentice-rl-12083448

Advances In Multivariate Data Analysis Proceedings Of The Meeting Of

The Classification And Data Analysis Group Cladag Of The Italian
Statistical Society University Of Palermo July 56 2001 1st Edition
Carmela Cappelli
https://ebookbell.com/product/advances-in-multivariate-data-analysis-
proceedings-of-the-meeting-of-the-classification-and-data-analysis-
group-cladag-of-the-italian-statistical-society-university-of-palermo-
july-56-2001-1st-edition-carmela-cappelli-4238520

The Statistical Analysis Of Quasiexperiments Reprint 2019 Christopher

H Achen

https://ebookbell.com/product/the-statistical-analysis-of-
quasiexperiments-reprint-2019-christopher-h-achen-51822324
The Statistical Analysis Of Doubly Truncated Data With Applications In
R 1st Edition Jacobo De Ualvarez

https://ebookbell.com/product/the-statistical-analysis-of-doubly-
truncated-data-with-applications-in-r-1st-edition-jacobo-de-
ualvarez-36306410

The Statistical Analysis Of Intervalcensored Failure Time Data 1st

Edition Jianguo Sun Auth

https://ebookbell.com/product/the-statistical-analysis-of-
intervalcensored-failure-time-data-1st-edition-jianguo-sun-
auth-4271968

The Statistical Analysis Of Functional Mri Data 1st Edition Nicole

Lazar Auth

https://ebookbell.com/product/the-statistical-analysis-of-functional-
mri-data-1st-edition-nicole-lazar-auth-4272032

The Statistical Analysis Of Failure Time Data Second Edition 2nd

Edition John D Kalbfleisch

https://ebookbell.com/product/the-statistical-analysis-of-failure-
time-data-second-edition-2nd-edition-john-d-kalbfleisch-4312398

The Statistical Analysis Of Time Series Theodore W Anderson

https://ebookbell.com/product/the-statistical-analysis-of-time-series-
theodore-w-anderson-6847034
The Statistical Analysis of
Multivariate Failure Time
Data
A Marginal Modeling Approach
MONOGRAPHS ON STATISTICS AND APPLIED PROBABILITY

Editors: F. Bunea, P. Fryzlewicz, R. Henderson, N. Keiding, T. Louis, R. Smith,

and W. Wong

Recently Published Titles

Asymptotic Analysis of Mixed Effects Models

Theory, Applications, and Open Problems
Jiming Jiang 155

Missing and Modified Data in Nonparametric Estimation with R Examples

Sam Efromovich 156

Probabilistic Foundations of Statistical Network Analysis

Harry Crane 157

Multistate Models for the Analysis of Life History Data

Richard J. Cook and Jerald F. Lawless 158

Nonparametric Models for Longitudinal Data

with Implementation in R
Colin O. Wu and Xin Tian 159

Multivariate Kernel Smoothing and Its Applications

José E. Chacón and Tarn Duong 160

Sufficient Dimension Reduction

Methods and Applications with R
Bing Li 161

Large Covariance and Autocovariance Matrices

Arup Bose and Monika Bhattacharjee 162

The Statistical Analysis of Multivariate Failure Time Data: A Marginal

Modeling Approach
Ross L. Prentice and Shanshan Zhao 163

For more information about this series please visit: http://crcpress.com/go/monographs

The Statistical Analysis of
Multivariate Failure Time
Data
A Marginal Modeling Approach

Ross L. Prentice
Shanshan Zhao
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2019 by Taylor & Francis Group, LLC

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

No claim to original U.S. Government works

Printed on acid-free paper

Version Date: 20190416

International Standard Book Number-13: 978-1-482-256557-4 (Hardback)

This book contains information obtained from authentic and highly regarded sources. Reason-
able efforts have been made to publish reliable data and information, but the author and publisher
cannot assume responsibility for the validity of all materials or the consequences of their use. The
authors and publishers have attempted to trace the copyright holders of all material reproduced in
this publication and apologize to copyright holders if permission to publish in this form has not
been obtained. If any copyright material has not been acknowledged please write and let us know
so we may rectify in any future reprint.

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

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

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

Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
To our spouses, children, and grandchildren.
 Contents

Preface xi

1 Introduction and Characterization of Multivariate Failure Time Dis-

tributions 1
1.1 Failure Time Data and Distributions 1
1.2 Bivariate Failure Time Data and Distributions 4
1.3 Bivariate Failure Time Regression Modeling 8
1.4 Higher Dimensional Failure Time Data and Distributions 9
1.5 Multivariate Response Data: Modeling and Analysis 11
1.6 Recurrent Event Characterization and Modeling 12
1.7 Some Application Settings 13
1.7.1 Aplastic anemia clinical trial 13
1.7.2 Australian twin data 14
1.7.3 Women’s Health Initiative hormone therapy trial 15
1.7.4 Bladder tumor recurrence data 17
1.7.5 Women’s Health Initiative dietary modification trial 19

2 Univariate Failure Time Data Analysis Methods 25

2.1 Overview 25
2.2 Nonparametric Survivor Function Estimation 25
2.3 Hazard Ratio Regression Estimation Using the Cox Model 28
2.4 Cox Model Properties and Generalizations 31
2.5 Censored Data Rank Tests 32
2.6 Cohort Sampling and Dependent Censoring 33
2.7 Aplastic Anemia Clinical Trial Application 35
2.8 WHI Postmenopausal Hormone Therapy Application 36
2.9 Asymptotic Distribution Theory 40
2.10 Additional Univariate Failure Time Models and Methods 44
2.11 A Cox-Logistic Model for Continuous, Discrete or Mixed Failure
Time Data 45

3 Nonparametric Estimation of the Bivariate Survivor Function 51

3.1 Introduction 51
3.2 Plug-In Nonparametric Estimators of F 52
3.2.1 The Volterra estimator 52
3.2.2 The Dabrowska and Prentice–Cai estimators 55

vii
viii
3.2.3 Simulation evaluation 57
3.2.4 Asymptotic distributional results 59
3.3 Maximum Likelihood and Estimating Equation Approaches 60
3.4 Nonparametric Assessment of Dependency 62
3.4.1 Cross ratio and concordance function estimators 62
3.4.2 Australian twin study illustration 63
3.4.3 Simulation evaluation 65
3.5 Additional Estimators and Estimation Perspectives 65
3.5.1 Additional bivariate survivor function estimators 65
3.5.2 Estimation perspectives 67

4 Regression Analysis of Bivariate Failure Time Data 71

4.1 Introduction 71
4.2 Independent Censoring and Likelihood-Based Inference 72
4.3 Copula Models and Estimation Methods 74
4.3.1 Formulation 74
4.3.2 Likelihood-based estimation 75
4.3.3 Unbiased estimating equations 76
4.4 Frailty Models and Estimation Methods 78
4.5 Australian Twin Study Illustration 79
4.6 Regression on Single and Dual Outcome Hazard Rates 79
4.6.1 Semiparametric regression model possibilities 79
4.6.2 Cox models for marginal single and dual outcome hazard
rates 80
4.6.3 Dependency measures given covariates 82
4.6.4 Asymptotic distribution theory 82
4.6.5 Simulation evaluation of marginal hazard rate estimators 85
4.7 Breast Cancer Followed by Death in the WHI Low-Fat Diet Inter-
vention Trial 89
4.8 Counting Process Intensity Modeling 91
4.9 Marginal Hazard Rate Regression in Context 92
4.9.1 Likelihood maximization and empirical plug-in estimators 92
4.9.2 Independent censoring and death outcomes 92
4.9.3 Marginal hazard rates for competing risk data 93
4.10 Summary 94

5 Trivariate Failure Time Data Modeling and Analysis 99

5.1 Introduction 99
5.2 Nonparametric Estimation of the Trivariate Survivor Function 100
5.2.1 Dabrowska-type estimator development 100
5.2.2 Volterra estimator 104
5.2.3 Trivariate dependency assessment 105
5.2.4 Simulation evaluation and comparison 106
5.3 Trivariate Regression Analysis via Copulas 109
 ix
5.4 Regression on Marginal Single, Double and Triple Failure Hazard
Rates 110
5.5 Simulation Evaluation of Hazard Ratio Estimators 113
5.6 Postmenopausal Hormone Therapy in Relation to CVD and Mortal-
ity 115

6 Higher Dimensional Failure Time Data Modeling and Estimation 119

6.1 Introduction 119
6.2 Nonparametric Estimation of the m-Dimensional Survivor Function 120
6.2.1 Dabrowska-type estimator development 120
6.2.2 Volterra nonparametric survivor function estimator 123
6.2.3 Multivariate dependency assessment 124
6.3 Regression Analysis on Marginal Single Failure Hazard Rates 125
6.4 Regression on Marginal Hazard Rates and Dependencies 129
6.4.1 Likelihood specification 129
6.4.2 Estimation using copula models 130
6.5 Marginal Single and Double Failure Hazard Rate Modeling 133
6.6 Counting Process Intensity Modeling and Estimation 136
6.7 Women’s Health Initiative Hormone Therapy Illustration 137
6.8 More on Estimating Equations and Likelihood 140

7 Recurrent Event Data Analysis Methods 143

7.1 Introduction 143
7.2 Intensity Process Modeling on a Single Failure Time Axis 144
7.2.1 Counting process intensity modeling and estimation 144
7.2.2 Bladder tumor recurrence illustration 146
7.2.3 Intensity modeling with multiple failure types 148
7.3 Marginal Failure Rate Estimation with Recurrent Events 149
7.4 Single and Double Failure Rate Models for Recurrent Events 151
7.5 WHI Dietary Modification Trial Illustration 151
7.6 Absolute Failure Rates and Mean Models for Recurrent Events 152
7.7 Perspective on Regression Modeling via Intensities and Marginal
Models 153

8 Additional Important Multivariate Failure Time Topics 157

8.1 Introduction 157
8.2 Dependent Censorship, Confounding and Mediation 158
8.2.1 Dependent censorship 158
8.2.2 Confounding control and mediation analysis 164
8.3 Cohort Sampling and Missing Covariates 166
8.3.1 Introduction 166
8.3.2 Case-cohort and two-phase sampling 166
8.3.3 Nested case–control sampling 169
8.3.4 Missing covariate data methods 170
8.4 Mismeasured Covariate Data 171
x
8.4.1 Background 171
8.4.2 Hazard rate estimation with a validation subsample 171
8.4.3 Hazard rate estimation without a validation subsample 172
8.4.4 Energy intake and physical activity in relation to chronic
disease risk 174
8.5 Joint Modeling of Longitudinal Covariates and Failure Rates 177
8.6 Model Checking 180
8.7 Marked Point Processes and Multistate Models 181
8.8 Imprecisely Measured Failure Times 182

Glossary of Notation 187

Appendix A: Technical Materials 191

A.1 Product Integrals and Stieltjes Integration 191
A.2 Generalized Estimating Equations for Mean Parameters 193
A.3 Some Basic Empirical Process Results 194

Appendix B: Software and Data 197

B.1 Software for Multivariate Failure Time Analysis 197
B.2 Data Access 199

Bibliography 201

Author Index 213

Subject Index 219

 Preface

Over the past several decades statistical methods for the analysis of univariate failure
time data have become well established, with Kaplan–Meier (KM) curves, censored
data rank tests, and especially Cox regression, as core analytic tools. These tools arise
substantially from nonparametric and semiparametric likelihood considerations. The
likelihood developments have mostly started with a focus on the survivor function
F, defined by F(t) = P(T > t), for a random failure time variate T > 0; or, in con-
junction with a corresponding covariate p-vector z = (z1 , z2 , . . . , z p ), on the survivor
function given z, defined by F(t; z) = P(T > t; z).
Because of the usual presence of right censoring, the hazard function λ , de-
fined with absolutely continuous failure times by λ (t : z) = f (t; z)/F(t; z), where
f (t; z) = −dF(t; z)/dt is the failure time density (given z) at follow-up time t, plays
a key role in the likelihood-based developments. In fact, a focus on hazard rates
over time allows considerable modeling flexibility compared to that for F given z,
since hazard rates can be allowed to depend, not just on a set of “baseline” covari-
ates, but also on covariate changes and on the occurrence of events of various types
arising during the follow-up of study subjects. As a result, hazard, and more specif-
ically intensity, rate regression analyses are sometimes referred to as event history
analyses. The Cox (1972) model stands as the major tool for univariate failure time
regression analysis. Under this model λ (t; z) is written as a product of a baseline
rate λ {t; z = (0, . . . , 0)}, hereafter λ (t; 0), and a hazard ratio factor that is usually
expressed in exponential form as exp(zβ ), with column vector β = (β1 , . . . , β p )0 .
The hazard ratio parameter β is often the principal target of estimation in univari-
ate failure time analysis, and efficient and reliable procedures are available for its
estimation, via Cox’s (1975) partial likelihood method. There are other classes of
hazard rate regression models, including accelerated failure time models that allow
covariates to affect the “speed” at which an individual traverses the failure time axis,
and transformation models that combine these and other classes of regression mod-
els, but the vast majority of applications in various disciplines and settings focus on
the Cox model. This model is quite comprehensive when the time-dependent covari-
ate feature is fully exercised. While it is clearly useful to have more than a single
framework for failure time regression analysis, other classes of semiparametric re-
gression models typically do not enjoy the same utility of parameter interpretation,
ease of modeling, and computational efficiency as does the Cox model. There are
a number of books that provide excellent accounts of many of these univariate fail-
ure time analysis developments, including early books by Kalbfleisch and Prentice
(1980, 2002), Breslow and Day (1980, 1987), Lawless (1983, 2002), Cox and Oakes

xi
xii PREFACE
(1984), Fleming and Harrington (1991), and Andersen, Borgan, Gill, and Keiding
(1993), among several others. These books vary in the extent of their coverage and
in their degree of technical formality. Some of them provide a rather thorough ac-
count of asymptotic distribution theory via counting process and martingale meth-
ods. For example, Andersen et al. (1993) present comprehensive distribution theory
using martingale and counting process convergence theory for a fairly broad range
of topics, from an event history analysis perspective.
The methods just alluded to can be obtained alternatively by plugging empiri-
cal hazard rate estimators into a representation for F given Z, where Z is a possibly
evolving covariate history, or by using a mean parameter estimating equation ap-
proach (e.g., Liang & Zeger, 1986) for parameter estimation. The utility of each of
these approaches can be considered for the analysis of multivariate failure time data.
A mature theory for the analysis of multivariate failure time data has been slow
to develop. Most of the books just mentioned include some account of analysis
methods for select types of multivariate failure time data; for example, Kalbfleisch
and Prentice (2002) include chapters on correlated failure time methods and on
recurrent event methods. However, consensus is still lacking on such basic top-
ics as the preferred means of estimating the joint survivor function F, defined by
F(t1 ,t2 ) = P(T1 > t1 , T2 > t2 ), for a pair of failure time variates (T1 , T2 ), even in
the homogeneous (non-regression) situation. Also, a compelling semiparametric ap-
proach to the conceptualization and modeling of dependency among failure time
variates, that is suited to right censored data, has yet to be established.
In more recent times, specialized books have been devoted to multivariate fail-
ure time data analysis. Hougaard (2000) provides a thorough account of multivariate
failure time methods proposed up to that point in time, with considerable emphasis
on the use of frailty models to induce dependencies, and on the use of transformation
models for marginal distributions in conjunction with copula models for dependency.
Crowder (2012) provides an update on these same approaches along with a detailed
account of parametric methods for multivariate failure time data analysis. Cook and
Lawless (2007) provide an excellent account of the more specialized topic of recur-
rent event data analysis methods.
The frailty approach has also been the subject of specialized books (Duchateau &
Janssen, 2010; Wienke, 2011) on data modeling and analysis. Frailty models are usu-
ally formulated by multiplying univariate hazard rate models by a positive random ef-
fect, or frailty, variate that is shared by failure time outcomes that may be dependent.
These methods are well suited to assessing comparative hazard rates for individuals
in predefined clusters, which could be an important application goal. Frailty models,
on the other hand are not so well suited to estimating population-averaged regres-
sion effects. Regression coefficients in frailty models have an interpretation that is
conditional on the random effect for the correlated outcomes, while corresponding
marginal regression associations typically have a complex form. Additionally, frailty
factors may need to be allowed to vary over the follow-up period of the study to
span a broad range of dependencies among failure times, adding further complexity.
Nevertheless, much useful information can often be extracted by the careful applica-
tion of frailty models. Therneau and Grambsch (2000), provide a valuable account of
 xiii
Cox model extensions using frailty models with emphasis on model testing and the
adaptation of existing software.
The copula approach to multivariate failure time data analysis has also received
considerable attention. This approach embraces univariate survivor function mod-
els, including those incorporating time-dependent covariates, for marginal survivor
functions, and brings these together using a parametric copula model to yield a joint
survivor function for a multivariate failure time variate given covariates. As such
marginal regression parameters retain the usual population averaged interpretation
that attends corresponding univariate data analyses. The models are typically applied
in a two-stage fashion (e.g., Shih & Louis, 1995) so that dependency assumptions im-
posed through the choice of copula neither bias or enhance the efficiency of marginal
hazard rate regression analyses compared to univariate failure time data analyses.
With a choice of copula function that is in good agreement with the data, this ana-
lytic approach can lead to useful summary measures of dependence between depen-
dent failure times, as is a major goal in some application settings, such as family
studies in genetic epidemiology. As typically applied, however, copula models tend
to impose strong assumptions on the nature of dependencies among failure times,
and are not convenient for allowing such dependencies to depend on covariates that
may be evolving over the study follow-up period(s).
Counting process intensity modeling provides an important approach to the re-
gression analysis of multivariate failure time data, with Andersen and Gill’s (1982)
development of distribution theory for semiparametric intensity models of multi-
plicative (Cox model) form standing out as a major development. When individual
study subjects have multiple failure times, of the same or different types, on the
same failure time axis, intensity modeling can lead to valuable insights into how fail-
ure rates at a given follow-up time depend on the preceding covariates and failure
histories for the individual. Often, however, primary interest resides in the depen-
dence of hazard rates on preceding covariate, but not preceding “counting process”
history, requiring a different regression methodology. Also the counting process in-
tensity modeling is not suited to multiple failure types for an individual with failure
types each having different potential censoring times, as may occur if failure types
have their own outcome ascertainment processes. Furthermore, the nice martingale
convergence results that attend multivariate failures on a single failure time axis have
not been extended to multiple time axis (see Andersen et al., 1993 for discussion).
Also see Aalen, Borgan, and Gjessing (2010) for an account of the various modeling
approaches mentioned above, with a major emphasis on counting process intensity
modeling, as well as the comprehensive recent book by Cook and Lawless (2018) on
multistate models for life history data analysis.
A marginal modeling approach can substantially fill the gap left by the previ-
ously mentioned modeling approaches. This approach, championed by Danyu Lin,
L. J. Wei and colleagues considers semiparametric models for the single failure haz-
ard rates for a multivariate failure time response (e.g., Spiekerman & Lin, 1998;
Lin, Wei, Yang, & Ying, 2000). Estimating equations have been developed for haz-
ard rate parameters, and empirical process convergence results lead to corresponding
distribution theory. The present authors have recently proposed an extension of these
xiv PREFACE
methods using semiparametric models for both single and dual outcome hazard rates
for correlated failure time outcomes. This extension leads to novel survivor function
estimators for pairs of failure time outcomes that may be accompanied by evolving
covariates, and to semiparametric estimates of dependency between pairs of failure
time variates. Importantly, these dual outcome hazard rate methods can provide ad-
ditional insight into the effects of the treatments or exposures under study, beyond
single outcome analyses. A primary aim of the presentation here is to provide a sum-
mary of these marginal methods. The frailty, copula and counting process intensity
approaches have been well covered in other venues, but we provide sufficient cov-
erage here to offer the reader an introduction to these methods, and to provide some
comparisons and contrasts with the marginal modeling approach.
While it is not feasible to provide a thorough account of asymptotic distribu-
tion theory for the estimators presented, we sketch the main elements of such theory
development, based mainly on empirical process theory, and make some use also
of martingale convergence results. While our account does not provide a complete
and rigorous development, we hope to provide enough detail to provide insight into
associated distributional results, and to provide a useful linkage to sources that pro-
vide such rigor and completeness. Also, our presentation emphasizes the modeling
of observable quantities, with little attention given, for example, to the counterfactual
approach to causal inference with observational data, the latent failure time approach
to competing risks, or even to random effects models generally. These approaches
each involve assumptions about unobservable quantities. As such they may be best
thought of as adjuncts to observable data modeling methods that may be consid-
ered to address further data analytic questions, but typically do so under additional
untestable model assumptions.
The present effort is essentially a research monograph. We do not attempt to
present a compendium of the rather voluminous set of methods that have been pro-
posed for some aspect of multivariate failure time data analysis, and we apologize to
other authors if their methodologic contributions are not discussed comprehensively,
or even at all. Furthermore, the data analysis methods emphasized require moderately
large numbers of pairs of outcome events during cohort follow-up, precluding useful
illustrations using classroom-type data sets. Partly for this reason we have chosen
to illustrate these multivariate methods using data from large cohorts in the national
Women’s Health Initiative in which we are engaged. The methods emphasized may
require novel software. We provide a description of, and link to, the software used in
most of our illustrations in appendix materials.
This book is intended primarily for statistical and biostatistical researchers as a
source of useful and interpretable data analysis methods, and as a basis for ideas
for further methodology development. The book could also serve as a text for a
graduate-level course in statistics or biostatistics among students having a reason-
able command of calculus and probability theory. Prior exposure to univariate failure
time methods for such students would also be helpful, though a presentation of the
“core” univariate methods mentioned above is included here. A set of exercises is in-
cluded with each chapter to enhance the usefulness of the presentation as a graduate
text. Many of the examples used to illustrate the methods described are drawn from
 xv
epidemiologic cohort studies or from randomized controlled clinical trials, and this
book may also serve as a useful reference source for quantitative epidemiologists,
and for biomedical scientists more generally, who are working with data of the type
discussed here. Statistical and biostatistical practitioners can derive utility from our
presentation, since we attempt to describe the analytic procedures and underlying
concepts in terms that are mostly not highly technical.
The authors would like to thank Aaron Aragaki for help with Women’s Health
Initiative illustrations, including creation of the platter plots shown on the front cover;
and to thank Noelle Noble for tremendous help with manuscript preparation.

Ross L. Prentice and Shanshan Zhao

 Chapter 1

Introduction and Characterization of

Multivariate Failure Time Distributions

1.1 Failure Time Data and Distributions 1

1.2 Bivariate Failure Time Data and Distributions 4
1.3 Bivariate Failure Time Regression Modeling 8
1.4 Higher Dimensional Failure Time Data and Distributions 9
1.5 Multivariate Response Data: Modeling and Analysis 11
1.6 Recurrent Event Characterization and Modeling 12
1.7 Some Application Settings 13
1.7.1 Aplastic anemia clinical trial 13
1.7.2 Australian twin data 13
1.7.3 Women’s Health Initiative hormone therapy trial 15
1.7.4 Bladder tumor recurrence data 17
1.7.5 Women’s Health Initiative dietary modification trial 19

1.1 Failure Time Data and Distributions

This book is concerned with methods for the analysis of time-to-event data, with
events generically referred to as failures. Typically there is an underlying study popu-
lation from which a cohort of individuals is selected and followed forward to observe
the times to occurrences of events of interest. Data analysis goals may include esti-
mation of the failure time distribution, and study of dependencies of failure times on
study subject characteristics, exposures, or treatments, generically referred to as co-
variates. The nature and strength of dependencies among the failure times themselves
may be also of interest in some settings. The types of mean and covariance models
used for the analysis of multivariate quantitative response data more generally can be
considered for multivariate failure time data analyses, but there are some important
features of multivariate time to response data that need to be acknowledged, as is
elaborated below.

1
2 INTRODUCTION AND CHARACTERIZATION
Failure time methods have application in many subject matter and research ar-
eas, including biomedical, behavioral, physical, and engineering sciences, and vari-
ous industrial settings. Most of the illustrations in this book will be drawn from the
biomedical research area in which the authors are engaged.
A major reason for specialized statistical methods for failure time data analysis is
the usual presence of right censoring since some, or perhaps most, study subjects will
not have experienced the event or events of interest at the cutoff date for data analysis,
and some may have discontinued participation in study follow-up procedures used to
ascertain failures prior to such cutoff date. The usual assumption about censoring
is that of independence, which requires the set of subjects who are uncensored and
continuing to be monitored for failures at any follow-up time to be representative of
individuals at risk for failure in the study population in terms of their failure rates,
conditional on a specified set of covariates.
Most failure time methods have focused on a single failure time variate T > 0.
The distribution of T can be characterized by its survivor function F, where F(t) =
P(T > t) and P denotes probability, which can be thought of in terms of frequencies
in the underlying, typically conceptual, study population. To accommodate contin-
uous, discrete and mixed failure times, F is usually assumed to be continuous from
the right with left-hand limits, whence the probability function corresponding to F is
defined at time t by −F(dt) where

 lim [{F(t) − F(t + ∆t)}/∆t]dt if t is a continuity point of F
−F(dt) = ∆t↓0
F(∆t) if t is a mass point of F,

and where F(∆t) = F(t − ) − F(t), with F(t − ) = limF(s).

s↑t
The need to accommodate right censoring in data analysis leads naturally to a
focus on the corresponding (cumulative) hazard function Λ, specified by

Λ(dt) = −F(dt)/F(t − ).

Λ(dt) is referred to as the hazard rate at follow-up time t. It is the failure probability,
or failure probability element if t is a continuity point of F, at time t conditional
on lack of failure prior to t. The absolutely continuous, discrete and mixed special
cases can be combined using Stieltjes integrals so that, for example, the probabil-
ity distribution function F and the (cumulative) hazard function at time t are given,
respectively, by
Z t Z t
F(t) = 1 − F(t) = 1 + F(ds) and Λ(t) = Λ(ds).
0 0

Also F can be written in product integral notation as

t
F(t) = ∏ {1 − Λ(ds)} (1.1)
0
FAILURE TIME DATA AND DISTRIBUTIONS 3
b
where ∏ denotes the product integral over (a, b]. Expression (1.1) reduces to
a
 Zt 
F(t) = exp − Λ(ds) (1.2)
0

if T is absolutely continuous. Product integrals are quite useful for likelihood con-
struction, and for data analysis more generally, with failure time data. In part this
utility arises because product integration allows discrete and continuous components
of the distribution of F to be included in a single notation. Appendix A provides
some background on the definition and properties of Stieltjes integrals and product
integration.
One could speculate that failure time data would be analyzed mostly using linear
models, or generalized linear models, or other standard statistical procedures, were
it not for the presence of right censoring. This would be an oversimplification, how-
ever, since a focus on hazard functions is helpful for statistical modeling and leads
to useful parameter interpretation regardless of the presence of censoring. Impor-
tantly, hazard rate modeling allows inference to be made on failure rates that evolve
over the cohort follow-up period in a manner that may depend on covariates that are
also changing over time, or on other types of events that are experienced by indi-
viduals during the study follow-up period. Also, in some contexts the inclusion of
time-varying covariates may be necessary for an independent censoring assumption
to be plausible. The simplest type of covariate is a fixed vector z = (z1 , . . . , zq ) as-
certained at time zero (e.g., at or before the date of enrollment in a cohort study) for
each study subject. Such a “baseline” covariate q-vector can include various aspects
of the individual’s preceding history that may be of interest in relation to subsequent
survival probabilities or hazard rates. In an observational epidemiology context, for
example, z may include exposures or characteristics that may be associated with the
risk of occurrence of a study disease. In an industrial accelerated lifetime product
testing situation z may include the temperature, or stress level, applied to a prod-
uct to produce early breakdowns of various types. In a randomized controlled trial
(RCT) z would typically include indicator variables for treatment assignments, pos-
sibly along with product terms between these indicator variables and other factors
that may influence the magnitude of any treatment effect on a failure time outcome.
In general there may be a covariate process, Z, that evolves over the study follow-
up period. For notation, one can denote by z(s) = {z1 (s), z2 (s), . . .} the covariate
value at follow-up time s ≥ 0 and write Z(t) = z(0) ∨ {z(s); 0 < s < t} to denote
the collection of an individual’s covariate data up to time t, including baseline co-
variates that can include exposures and characteristics that pertain to the time period
prior to study enrollment. Here Z(t) is usually defined to involve sample paths that
are continuous from the left; that is Z(t) = lim Z(s) when t > 0, so that the history
s↑t
Z(t) does not include covariate “process” jumps occurring at time t. It is often useful
to specify a statistical model for Λ given Z. In fact many, perhaps most, univariate
failure time data analysis applications involve the modeling and estimation of hazard
rates of this type. This statistical approach, and the use of the Cox (1972) regres-
sion model specifically with its nonparametric “baseline” hazard function, will be
4 INTRODUCTION AND CHARACTERIZATION
described in Chapter 2. There are many variants of the simple cohort study design
in which data are ascertained on a random sample of a study population conditional
on Z(0); for example, study subjects may enter the cohort late (t > 0), or be ascer-
tained with varying selection probabilities according to their covariate histories, or
according to their failure experience during study follow-up; there may be variations
in how and when covariates are measured, and in the reliability of those measure-
ments. Also subjects may cease continued study participation, censoring their failure
times, for complex reasons, possibly including their experiences during the study
follow-up period. These and other study features necessitate a corresponding rich
class of statistical models and methods for data analysis. Hazard rate models provide
a major unifying concept, and related methods have come to be known as event his-
tory analysis methods. The usual presence of right censoring has implications for the
types of statistical models that can be reliably applied, even if the covariates are time-
independent. For example linear regression models and maximum likelihood-based
regression parameter estimators have valuable robustness properties in that consis-
tent estimators of regression parameters typically arise even if the parametric form
for the linear model error variable is misspecified, in the absence of censoring or
other forms of missing or incomplete data. Unfortunately, this robustness is not re-
tained in the presence of right censoring, motivating the application of nonparametric
or semiparametric models, and estimation procedures that will have good behavior
regardless of the value of a nonparametric model component. This topic, too, will be
elaborated in Chapter 2 for univariate failure time data.

1.2 Bivariate Failure Time Data and Distributions

Multivariate failure time data arise when univariate failure times for individuals in
the study cohort need not be statistically independent, or when multiple events of
the same or of different types are recorded for the same individual in a study cohort.
Noteworthy special cases include RCTs where the treatment(s) may affect two or
more failure time outcomes; epidemiologic follow-up studies of family members in
an attempt to learn about inherited and shared environment in relation to the risk of a
disease; studies of the timing of multiple disease recurrences among treated patients;
or studies of the sequence of breakdown times for a manufactured product. While
univariate failure time data analysis methods can be regarded as well-established
and mature, the same cannot be said for multivariate failure time analysis methods.
Even such basic topics as nonparametric estimation of the bivariate survivor function
remain the topic of continuing statistical research.
Let T1 > 0 and T2 > 0 denote a pair of failure time variates that may be dependent.
Their joint survivor function F is given by F(t1 ,t2 ) = P(T1 > t1 , T2 > t2 ). The two
variates may be subject to the same, or different, censoring patterns complicating the
estimation of F. F can be characterized by its marginal survivor functions, given by
F(t1 , 0) and F(0,t2 ), and its so-called double failure hazard function Λ11 , where

Λ11 (dt1 , dt2 ) = F(dt1 , dt2 )/F(t1− ,t2− ). (1.3)

BIVARIATE FAILURE TIME DATA AND DISTRIBUTIONS 5
Here the bivariate probability element F(dt1 , dt2 ) can be written



{∂ 2 F(t1 ,t2 )/∂t1 , ∂t2 }dt1 dt2 if (t1 ,t2 ) is a continuity point of F
{−∂ F(t1 , ∆t2 )/∂t1 }dt1 if F is continuous in t1 but not in t2 at (t1 ,t2 )



F(dt1 , dt2 ) = {−∂ F(∆t1 ,t2 )/∂t2 }dt2 if F is continuous in t2 but not in t1 at (t1 ,t2 )
F(∆t1 , ∆t2 ) = F(t − ,t − ) − F(t1 ,t − ) − F(t − ,t2 ) + F(t1 ,t2 )




 1 2 2 1
 if F is discontinuous in both components at (t1 ,t2 ),

where ∂ denotes partial derivative.

Specifically, since F(t1 ,t2 ) = 1 − P(T1 ≤ t1 ) − P(T2 ≤ t2 ) + P(T1 ≤ t1 , T2 ≤ t2 ),
one can write
Z t1 Z t2
F(t1 ,t2 ) = F(t1 , 0) + F(0,t2 ) − 1 + F(s− −
1 , s2 )Λ11 (ds1 , ds2 ), (1.4)
0 0

defining a Volterra integral equation for F, in terms of F(·, 0), F(0, ·) and Λ11 , that
has a unique solution. That solution is in a rather inconvenient Péano series form.
Nevertheless, the fact that F is determined by its marginal hazard rates, and its double
failure hazard rate provides useful background for bivariate failure time data mod-
eling. See Appendix A for the definition of two-dimensional Stieltjes integrals. The
Péano series solution to (1.4) is given in §3.2.1.
It is natural to consider the joint probability distribution for (T1 , T2 ) as comprised
of its marginal survivor functions, or equivalently its marginal hazard functions, and
a component that measures dependency between T1 and T2 given the marginal dis-
tributions. In fact, the copula approach to bivariate distribution modeling uses this
conceptualization through specification of a parametric model that brings together
the marginal survivor functions for T1 and T2 to form their joint survivor function.
Dependency between T1 and T2 at follow-up time (t1 ,t2 ) can also be character-
ized by comparing the double failure hazard rate Λ11 (dt1 , dt2 ) to the product of cor-
responding single failure hazard rates at (t1 ,t2 ). The single failure hazard rate func-
tions Λ10 and Λ01 are defined respectively by Λ10 (dt1 ,t2− ) = −F(dt1 ,t2− )/F(t1− ,t2− )
and Λ01 (t1− , dt2 ) = −F(t1− , dt2 )/F(t1− ,t2− ) for t1 ≥ 0 and t2 ≥ 0; and by Λ10 (t1 ,t2 ) =
R t1 R t2
0 Λ10 (ds1 ,t2 ) and Λ01 (t1 ,t2 ) = 0 Λ01 (t1 , ds2 ). In particular one can compare
− −
Λ11 (dt1 , dt2 ) to Λ10 (dt1 ,t2 )Λ01 (t1 , dt2 ) at any t1 > 0, t2 > 0 through the ratio

α(t1 ,t2 ) = Λ11 (dt1 , dt2 )/{Λ10 (dt1 ,t2− )Λ01 (t1− , dt2 )}, (1.5)

which expresses double failure rate departure from local independence on a rela-
tive scale. The function α is referred to as the cross ratio function, terminology that
F(dt1 ,dt2 )F(t1− ,t2− )
reflects the expression α(t1 ,t2 ) = F(dt1 ,t2− )F(t1− ,dt2 )
. It provides a possible means of
characterizing dependency between T1 and T2 in that its set of possible values is
essentially unrestricted by the corresponding marginal hazard rates Λ10 (dt1 , 0) and
Λ01 (0, dt2 ). In fact α(t1 ,t2 ) may take any value in [0, ∞) for absolutely continuous
failure times.
It turns out that the bivariate survivor function F is completely determined also
6 INTRODUCTION AND CHARACTERIZATION
by its marginal hazard rate Λ10 (·, 0) and Λ01 (0, ·) and its cross ratio functions. The
joint modeling and estimation of marginal hazard rates and cross ratios, however, is
complicated by the lack of a closed form expression for F in terms of these distribu-
tional components.
There is, however, a closed form expression for the survivor function in terms of
its single failure hazard rates more generally and its cross ratio function. Specifically
F is given, in product integral notation (see Appendix A), by
t1 t2
F(t1 ,t2 ) = ∏{1 − Λ10 (ds1 , 0)} ∏{1 − Λ01 (0, ds2 )}
0 0
t1
Λ11 (ds1 , ds2 ) − Λ10 (ds1 , s−
t2 −
 
2 )Λ01 (s1 , ds2 )
∏ ∏ 1 + {1 − Λ10 (∆s1 , s− )}{1 − Λ01 (s− , ∆s2 )}
0 0 2 1
t1 t2
= ∏{1 − Λ10 (ds1 , 0)} ∏{1 − Λ01 (0, ds2 )}
0 0
t1 t2
{α(s1 , s2 ) − 1}Λ10 (ds1 , s− −
 
2 )Λ01 (s1 , ds2 )
∏∏ 1 + . (1.6)
0 0 {1 − Λ10 (∆s1 , s− −
2 )}{1 − Λ01 (s1 , ∆s2 )}

The first expression is Dabrowska’s (1988) representation of the bivariate sur-

vivor function in terms of single and double failure hazard rates, while the second
simply substitutes (1.5). The first two product integrals in (1.6) are the marginal sur-
vivor functions, while the third involves the cross ratios α, but also the single failure
hazard rates away from the coordinate axes. However, straightforward calculations
give
t2 
{α(t1 , s2 ) − 1}Λ10 (dt1 , s− − 
2 )Λ01 (t1 , ds2 )
Λ10 (dt1 ,t2 ) = 1−{1−Λ10 (dt1 , 0)} ∏ 1+
0 {1 − Λ10 (∆t1 , s− −
2 )}{1 − Λ01 (t1 , ∆s2 )}

from which one sees that Λ10 (dt1 ,t2 ) is determined by the marginal hazard rate
Λ10 (dt1 , 0), the cross ratios α(t1 , s2 ) for 0 < s2 ≤ t2 , the single failure hazard
rates Λ10 (dt1 , s2 ) for 0 < s2 < t2 , and single failure hazard rates Λ01 (t1− , ds2 ) for
0 < s2 ≤ t2 . Note that the right side of this expression does not involve single
failure hazard rates Λ10 (ds1 , s2 ) or Λ01 (s1 , ds2 ) at (s1 , s2 ) = (t1 ,t2 ). This, and a
corresponding expression for Λ01 (t1 , dt2 ), support an inductive proof that F is de-
termined by its marginal hazard rate and cross ratio functions for discrete fail-
ure times (T1 , T2 ). The induction hypothesis specifies this to be true for all points
{(s1 , s2 ); 0 ≤ s1 ≤ t1 , 0 ≤ s2 ≤ t2 , (s1 , s2 ) 6= (t1 ,t2 )}. The single failure hazard rate
expressions at (t1 ,t2 ) then show the hypothesis to hold also at (t1 ,t2 ). Since the in-
duction hypothesis holds trivially along the coordinate axes, it follows that it holds
also throughout the set of grid points (t1 ,t2 ), where P(T1 = t1 ) > 0 and P(T2 = t2 ) > 0,
and hence for the discrete failure time survivor function as a whole.
If the distribution of (T1 , T2 ) includes continuity points then the T1 and T2 axes
can each be partitioned with F, by definition of the product integrals in (1.6), given
by the limit of discrete distributions formed by these partitions as the mesh of the grid
decreases to zero. Each such approximating discrete distribution can be characterized
BIVARIATE FAILURE TIME DATA AND DISTRIBUTIONS 7
by its marginal hazard rate and cross ratio functions, and these functions converge to
those from F as the partition mesh becomes small. It follows that F is uniquely de-
termined generally by its marginal hazard rate and cross ratio functions. In fact, one
could regard marginal hazard rates and cross ratios as key building blocks for bivari-
ate survival data modeling. Note that if F is absolutely continuous (1.6) simplifies
to
 Zt Z t2
1
F(t1 ,t2 ) = exp − Λ10 (ds1 , 0) − Λ01 (0, ds2 )+
0 0
Z t Z t2 
{Λ11 (ds1 , ds2 ) − Λ10 (ds1 , s2 )Λ01 (s1 , ds2 )}
0 0
 Zt Z t2
1
= exp − Λ10 (ds1 , 0) − Λ01 (0, ds2 )+
0 0
Z t1 Z t2 
{α(s1 , s2 ) − 1}Λ10 (ds1 , s2 )Λ01 (s1 , ds2 ) .
0 0

The survivor function F can be characterized also in terms of its marginal hazard
rate functions Λ10 (·, 0) and Λ01 (0, ·) and its “covariance rate” function Ω11 , defined
by Ω11 (t1 ,t2 ) = 0t1 0t2 Ω11 (ds1 , ds2 ) where
R R

Λ11 (ds1 , ds2 ) − Λ10 (ds1 , s− −

2 )Λ01 (s1 , ds2 )
Ω11 (ds1 , ds2 ) = − − .
{1 − Λ10 (∆s1 , s2 )}{1 − Λ01 (s1 , ∆s2 )}

This characterization, from (1.6), expresses F(t1 ,t2 ) as a product of its marginal sur-
vival probabilities F(t1 , 0) and F(0,t2 ) and a factor reflecting dependency between
T1 and T2 over (0,t1 ] × (0,t2 ]. With absolutely continuous failure times the denomi-
nator terms in Ω11 equal one, and Ω11 (dt1 , dt2 ) is simply the difference between the
double failure hazard element at (t1 ,t2 ) and the “local independence” product of the
corresponding single failure hazard elements.
Often with bivariate failure time data, primary interest will focus on marginal
hazard rates and their dependence on covariates. The reader might logically ask, why
not simply apply the well-established univariate failure time methods that were pre-
viously mentioned for inference on marginal hazard rates, while bringing in a com-
plementary dependency function only if there is additional interest in the nature of
any dependency between T1 and T2 . In fact, much of the available literature on copula
models uses this type of two-stage modeling with a parametric “copula” model for F
given its marginal hazard rates. This same approach will be considered in Chapter 4,
but with semiparametric and parametric regression models for marginal hazard rates
and for cross ratios, respectively.
A simple, but important, special case of a bivariate survivor function is provided
by the Clayton–Oakes model

F(t1 ,t2 ) = {F(t1 , 0)−θ + F(0,t2 )−θ − 1}−1/θ ∨ 0, for θ ≥ −1. (1.7)

This joint survivor function (Exercise 1.1) has marginal survivor functions
given by F(t1 , 0) and F(0,t2 ) and is an example of a copula model, wherein the
8 INTRODUCTION AND CHARACTERIZATION
marginal survival probabilities are brought together through a copula function C,
here C(u1 , u2 ) = {u−θ + u−θ 2 − 1}
−1/θ ∨ 0, to give the joint survivor function. The

parameter θ measures dependence between T1 and T2 with (1.7) approaching the

independence special case, F(t1 ,t2 ) = F(t1 , 0)F(0,t2 ), as θ → 0, and the so-called
upper Fréchet bound of F(t1 , 0) ∧ F(0,t2 ) as θ → ∞. Also the distribution for (T1 , T2 )
is everywhere absolutely continuous for θ > −0.5. For θ ≤ 0.5 probability begins to
accumulate along the lower Fréchet bound F(t1 , 0) + F(0,t2 ) − 1 ∨ 0, but with each
(t1 ,t2 ) a continuity point away from this lower bound. As θ → −1 all probability
is eventually assigned to the Fréchet lower bound, and the probability density con-
verges to zero away from this bound. In these expressions, ∧ and ∨ refer to minimum
and maximum, respectively.
At any continuity point (t1 ,t2 ) the cross ratio from (1.7) is simply α(t1 ,t2 ) =
1 + θ . Hence the cross ratio function from (1.7) is assumed to take a constant value,
independently of t1 or t2 . This illustrates a potential limitation in using a copula model
with dependencies between T1 and T2 characterized by one or a few parameters:
Such models may make efficient use of data, for example for marginal hazard rate
estimation, but in doing so they may introduce biases if the dependency modeling
assumptions are not in agreement with available data. For example, if cross ratios
tend to be relatively larger at small compared to large follow-up times, then use of
(1.7) may lead to biased estimates of marginal hazard rates. A modeling challenge
then is to avoid unduly strong assumptions on cross ratios, and on other dependency
functions, in the estimation of F. Nonparametric estimation of F will be discussed
in Chapter 3.

1.3 Bivariate Failure Time Regression Modeling

Suppose now that the failure times (T1 , T2 ) are accompanied by a covariate z =
(z1 , z2 , . . .) available at t1 = 0,t2 = 0. The arguments of the preceding section gen-
eralize directly to show that the survivor function F given z defined by F(t1 ,t2 ; z) =
P(T1 > t1 , T2 > t2 ; z), is uniquely determined by its marginal hazard rate functions
given z and its cross ratio function given z. For example, the marginal hazard rates
are defined by

Λ10 (dt1 , 0; z)= −F(dt1 , 0; z)/F(t1− , 0; z) and Λ01 (0, dt2 ; z)= −F(0, dt2 ; z)/F(0,t2− ; z).

More generally (T1 , T2 ) may be accompanied by an evolving covariate process Z.

Let z(t1 ,t2 ) = {z1 (t1 ,t2 ), z2 (t1 ,t2 ), . . .} denote the covariate value at time (t1 ,t2 ) for
t1 ≥ 0 and t2 ≥ 0 and denote by

Z(t1 ,t2 ) =
{z(s1 , s2 ); s1 = 0 if t1 = 0, s1 < t1 if t1 > 0; and s2 = 0 if t2 = 0, s2 < t2 if t2 > 0}

the covariate history up to (t1 ,t2 ). One can define single and double failure hazard
HIGHER DIMENSIONAL FAILURE TIME DATA AND DISTRIBUTIONS 9
rates at (t1 ,t2 ) given Z(t1 ,t2 ), respectively, by

Λ10 {dt1 ,t2− ; Z(t1 ,t2 )} = P{T1 ∈ [t1 ,t1 +dt1 ); T1 ≥ t1 , T2 ≥ t2 , Z(t1 ,t2 )},
Λ01 {t1− , dt2 ; Z(t1 ,t2 )} = P{T2 ∈ [t2 ,t2 +dt2 ); T1 ≥ t1 , T2 ≥ t2 , Z(t1 ,t2 )}, and
Λ11 {dt1 , dt2 ; Z(t1 ,t2 )} = P{T1 ∈ [t1 ,t1 +dt1 ), T2 ∈ [t2 ,t2 +dt2 );
T1 ≥t1 , T2 ≥t2 , Z(t1 ,t2 )},

and corresponding single and double failure hazard processes Λ10 , Λ01 and Λ11 by
Z t1
Λ10 {t1 ,t2− ; Z(t1 ,t2 )} = Λ10 {ds1 ,t2− ; Z(s1 ,t2 )},
0
Z t2
Λ01 {t1− ,t2 ; Z(t1 ,t2 )} = Λ01 {t1− , ds2 ; Z(t1 , s2 )}, and
0
Z t1 Z t2
Λ11 {t1 ,t2 ; Z(t1 ,t2 )} = Λ11 {ds1 , ds2 ; Z(s1 , s2 )}.
0 0

Regression modeling can focus, for example, on marginal hazard rates

Λ10 {dt1 , 0; Z(t1 , 0)} for t1 ≥ 0 and Λ01 {0, dt2 ; Z(0,t2 )} for t2 ≥ 0, and cross ra-
tios α{t1 ,t2 ; Z(t1 ,t2 )} = Λ11 {dt1 , dt2 ; Z(t1 ,t2 )}/ [Λ10 {dt1 ,t2− ; Z(t1 , t2 )}Λ01 {t1− , dt2 ;
Z(t1 ,t2 )}] for t1 > 0 and t2 > 0. Given the common focus on mean and covariance
parameter modeling with uncensored data a focus on marginal hazard rates and a
complementary dependency function seems natural for bivariate failure time mod-
eling. A conceptualization based on a marginal hazard rates and the double failure
hazard rate pattern of co-occurrence for the two failure time events is also very nat-
ural. For example, in a clinical trial context there may be interest in whether a treat-
ment influences the rate of development of two important clinical outcomes jointly.
The joint modeling of marginal single and double failure hazard rates is useful for
these types of applications, and will be a major focus of the presentation in this book
beginning in Chapter 3, with regression extensions in Chapter 4 and in later chapters.

1.4 Higher Dimensional Failure Time Data and Distributions

The concepts of the preceding sections can be generalized to more than two failure
time variates, as will be elaborated in Chapters 5 and 6. For the present, consider
three absolutely continuous failure time variates T1 , T2 and T3 . The survivor function
F, defined by F(t1 ,t2 ,t3 ) = P(T1 > t1 , T2 > t2 , T3 > t3 ) for ti ≥ 0, i = 1, 2, 3 can be
written using concepts given in Dabrowska (1988) as
Z t1 Z t2 Z t3
F(t1 ,t2 ,t3 ) = exp{− Λ100 (ds1 , 0, 0) − Λ010 (0, ds2 , 0) − Λ001 (0, 0, ds3 )
0 0 0
Z t1 Z t2 Z t1 Z t3 Z t2 Z t3
+ Ω110 (ds1 , ds2 , 0) + Ω101 (ds1 , 0,ts3 ) + Ω011 (0, ds2 , ds3 )
0 0 0 0 0 0
Z t1 Z t2 Z t3
− Ω111 (ds1 , ds2 , ds3 )}, (1.8)
0 0 0
10 INTRODUCTION AND CHARACTERIZATION
where, for example, Ω110 (ds1 , ds2 , s3 ) = {∂ 2 log F(s1 , s2 , s3 )/∂ s1 ∂ s2 }ds1 , ds2 is the
covariance rate for (T1 , T2 ) at (s1 , s2 , s3 ) and

Ω111 (ds1 , ds2 , ds3 ) ={−∂ 3 log F(s1 , s2 , s3 )/∂ s1 ∂ s2 ∂ s3 }ds1 ds2 ds3
=Λ111 (ds1 , ds2 , ds3 ) − Ω110 (ds1 , ds2 , s3 )Λ001 (s1 , s2 , ds3 )
− Ω101 (ds1 , s2 , ds3 )Λ010 (s1 , ds2 , s3 )
− Ω011 (s1 , ds2 , ds3 )Λ100 (ds1 , s2 , s3 )
− Λ100 (ds1 , s2 , s3 )Λ010 (s1 , ds2 , s3 )Λ001 (s1 , s2 , ds3 ).

In this expression Λ111 is the triple failure hazard rate function given by

Λ111 (ds1 , ds2 , ds3 ) = [{−∂ 3 F(s1 , s2 , s3 )/∂ s1 ∂ s2 ∂ s3 }/F(s1 , s2 , s3 )]ds1 ds2 ds3 .

Note that Ω111 (ds1 , ds2 , ds3 ) contrasts the triple failure hazard rate at (s1 , s2 , s3 )
with that under local independence after allowing for dependencies at (s1 , s2 , s3 ) that
emanate from the pairwise marginal covariance rates.
In Chapter 5 we will consider modeling the trivariate survivor function, and re-
gression extensions thereof. As a specific survivor function consider

F(t1 ,t2 ,t3 ) ={F(t1 ,t2 , 0)−θ + F(t1 , 0,t3 )−θ + F(0,t2 ,t3 )−θ
− F(t1 , 0, 0)−θ − F(0,t2 , 0)−θ − F(0, 0,t3 )−θ + 1}−1/θ ∨ 0 (1.9)

for −1 ≤ θ < ∞, which is a trivariate generalization of (1.7). Expression (1.9)

has marginal survivor functions given by F(t1 , 0, 0), F(0,t2 , 0), F(0, 0,t3 ) and cor-
responding pairwise marginal survivor functions given by F(t1 ,t2 , 0), F(t1 , 0,t3 ) and
F(0,t2 ,t3 ). For example, Clayton–Oakes models (1.7) could be specified for these
pairwise marginal survivor functions, each with its own cross ratio parameter. The
parameter θ in (1.9) governs dependencies among the three failure times beyond
those attributable to the pairwise dependencies. As θ → 0, (1.9) approaches

F(t1 ,t2 , 0)F(t1 , 0,t3 )F(0,t2 ,t3 )/{F(t1 , 0, 0)F(0,t2 , 0)F(0, 0,t3 )}.

Also (1.9) approaches the upper Fréchet bound F(t1 ,t2 , 0) ∧ F(t1 , 0,t3 ) ∧ F(0,t2 ,t3 )
as θ → ∞, and the lower Fréchet bound

{F(t1 ,t2 , 0) + F(t1 , 0,t3 ) + F(0,t2 ,t3 ) − F(t1 , 0, 0) − F(0,t2 , 0) − F(0, 0,t3 ) + 1} ∨ 0

as θ → −1. All (t1 ,t2 ,t3 ) values away from this lower bound are continuity points
for (T1 , T2 , T3 ) for any θ ∈ [−1, ∞).
Straightforward calculations from (1.9), at continuity points, show

Ω111 (dt1 , dt2 , dt3 ) = θ {Ω110 (dt1 , dt2 ,t3 )Λ001 (t1 ,t2 , dt3 )
+ Ω101 (dt1 ,t2 , dt3 )Λ010 (t1 , dt2 ,t3 )
+ Ω011 (t1 , dt2 , dt3 )Λ100 (dt1 ,t2 ,t3 )}
− θ 2 Λ100 (dt1 ,t2 ,t3 )Λ010 (t1 , dt2 ,t3 )Λ001 (t1 ,t2 , dt3 ) (1.10)
MULTIVARIATE RESPONSE DATA: MODELING AND ANALYSIS 11
under (1.9), so that the parameter θ governs the magnitude of any trivariate depen-
dency among the three variates.
In the very special case in which each of the three pairwise marginal survivor
functions adheres to (1.7) with the same θ value as in (1.9), this trivariate survivor
function reduces to

F(t1 ,t2 ,t3 ) = {F(t1 , 0, 0)−θ + F(0,t2 , 0)−θ + F(0, 0,t3 )−θ − 2}−1/θ ∨ 0,

which some authors have considered as the trivariate generalization of (1.7). How-
ever, this survivor function may be too specialized for many applications: not only
are the pairwise cross ratios independent of their respective time arguments, but these
marginal cross ratios take the identical value (1 + θ ) for each pair of failure times.
Generalization of (1.9) to an arbitrary number of failure time variates will be given
in Chapter 6, along with estimation procedures for trivariate and higher dimensional
failure time data analysis more generally.

1.5 Multivariate Response Data: Modeling and Analysis

Historically multivariate response data analysis methods have relied on a multivari-
ate normal modeling assumption. Estimation procedures focusing on parameters in
the response mean vector and covariance matrix have long ago been derived using
multivariate t-distributions and Wishart distributions. Many problems of scientific
interest can be formulated in terms of multivariate mean regression parameters, or
covariance matrix patterns. Many of these developments are summarized in the clas-
sic book by Anderson (1984). Because many response variables are decidedly non-
normal in distribution, estimating equations for mean parameters, and for mean and
covariance parameters, subsequently came to provide a central approach to the mod-
eling and analysis of multivariate response data, with major stimulus from the work
of Liang and Zeger (1986) and Zeger and Liang (1986). By construing censored fail-
ure time data as a set of binary variates, “generalized estimating equations” may be
considered also for the analysis of univariate and multivariate failure time data, as
will be illustrated in subsequent chapters. Appendix A provides a brief account of
generalized estimating equations for mean parameter estimation.
Analysis methods based on multivariate normal theory, on more general paramet-
ric maximum likelihood theory or M-estimation, or based on generalized estimat-
ing equations enjoy some valuable robustness to departure from distributional form
working model assumptions. With right censoring however, or with other forms of
censoring or missing data, this robustness is typically lost, providing a reason to em-
phasize nonparametric and semiparametric models and estimation procedures. Also,
as with generalized estimating equations for mean and covariance parameters, with
censored data it is valuable to consider models and estimation procedures for de-
pendency parameters that do not compromise the properties of marginal distribution
parameter estimates. Hence the approach emphasized in subsequent chapters will
estimate marginal distribution parameters using the associated univariate data, with
additional estimating procedures for multivariate parameter estimation. Related pro-
12 INTRODUCTION AND CHARACTERIZATION
cedures using dependency parameter estimates to weight unbiased marginal distribu-
tion parameter estimating functions will also be mentioned.
The discussion of this section focuses on marginal, or population-averaged, quan-
tities. There is also a substantial multivariate failure time literature on the modeling
and estimation of evolving failure rates that include each study subject’s prior failure
time history in the failure rate specification. Such rates are referred to as intensities,
and the associated modeling and estimation procedures play a major role in recurrent
event methods and in more general multistate modeling methods.

1.6 Recurrent Event Characterization and Modeling

In an important subclass of applications there is a single failure time axis with indi-
viduals experiencing a failure continuing to be followed for second and subsequent
failures. Such recurrent failure time process data T1 , T2 , . . . may be subject to a single
censoring process that discontinues the follow-up for the study subject. In many re-
spects the modeling and analysis of this type of recurrent event data is more like the
univariate failure time data modeling mentioned in §1.1 than the correlated failure
time modeling discussed in §§1.2–1.4. For example, recurrent event modeling can
focus on failure rates

Λ{dt; H (t)} = P{failure in [t,t + dt); H (t)} (1.11)

where the conditioning event H (t) includes not only the covariate history Z(t) prior
to time t, but also the failure history for the individual prior to time t. This latter
history is conveniently described by the counting process N, where N(dt) equals the
number of failures experienced by the individual at time t and N(t) = 0t N(ds). Re-
R

current event data frequently involve a large number of failures on individual study
subjects, and the modeling of (1.11) sometimes involves simplifying assumptions as
to how the intensity at time t depends on the preceding failure history {N(s), s < t}
for the individual, with Markov and semi-Markov assumptions commonly imposed.
Of course, there may be recurrent events for several types of failure time variates in
which case the concepts of this and the preceding sections can be combined leading
to the modeling of marginal failure rate processes (1.11), along with corresponding
cross ratio processes for example, in each case with the conditioning event including
not only the covariate, but also the preceding failure history for each event type. The
modeling and analysis of recurrent event data will be discussed in Chapter 7. Impor-
tantly, marginal modeling approaches in which one models failure rates at follow-up
time t as a function of the preceding failure, but not the preceding counting process
history, will also be considered in Chapter 7, along with generalizations to include
failures of various types on the same or different failure time axes.
Chapter 8 considers a variety of additional important topics in the modeling and
analysis of multivariate failure time data, including censoring schemes that are “de-
pendent;” cohort sampling procedures where some components of covariate histories
are assembled only for failing individuals and a subcohort of individuals who are
without failure during certain follow-up periods; data analysis procedures when co-
SOME APPLICATION SETTINGS 13
variate values are subject to measurement error or missing values, and joint models
for covariate histories and failure times, among other topics.
We end this chapter by describing a few application areas encountered in our
applied work. These settings will be used to illustrate modeling and analysis methods
in subsequent chapters.

1.7 Some Application Settings

1.7.1 Aplastic anemia clinical trial
Table 1.1 shows failure time data for 64 patients having severe aplastic anemia. These
data are from Storb et al. (1986) as presented in Kalbfleisch and Prentice (2002, Ta-
ble 1.2). Patients were conditioned with high-dose cyclophosphamide followed by
bone marrow cell infusion from a human lymphocyte antigen (HLA)–matched fam-
ily member. Patients were then randomly assigned to receive either cyclosporine plus
methotrexate (CSP + MTX) or methotrexate alone (MTX) with time from assign-
ment to the diagnosis of stage 2 or greater acute graft versus host disease (A-GVHD)
as a key outcome. Table 1.1 also shows whether (LAF = 1) or not (LAF = 0) each
patient was assigned to a laminar air flow isolation room, and the patient’s age in
years, both at baseline (t = 0). Note that 44 of the 64 patients had right censored
times to A-GVHD, with censored times indicated by an asterisk. A principal study
goal was to assess whether the addition of CSP to the MTX regimen reduced the
A-GVHD risk. An independent censoring assumption requires patients who are still
alive without A-GVHD at the cutoff date for data analysis to be representative of
patients who are without an A-GVHD diagnosis at this censoring time, conditional
on treatment, LAF assignments and age (assuming that these variables are included
and well modeled in the data analysis). This independent censoring assumption is
satisfied for patients alive and without A-GVHD at the time of data analysis. Many
of the asterisks in Table 1.1, however, attend the days from randomization to death
for patients who died without having an A-GVHD diagnosis. It may seem like Table
1.1 provides censored bivariate failure time data on time (T1 ) from randomization to
A-GVHD and time from randomization to death (T2 ), but the table only gives times
to the smaller of T1 and T2 . Competing risk methods (e.g., Kalbfleisch & Prentice,
2002, Chapter 8) indicate that A-GVHD hazard rates (among surviving patients) can
be estimated simply by regarding death times as additional censored observations.
Hence these data can be analyzed as censored univariate failure time data to relate
treatment assignment and other study subject characteristics to A-GVHD incidence,
as will be illustrated in Chapter 2. Genuinely bivariate data could be obtained in this
application by recording also the time to death for patients experiencing A-GVHD.
These more comprehensive data would support estimation of double failure hazard
rates (for A-GVHD followed by death) and their comparison among randomization
groups, as well as analyses of dependency between time to A-GVHD and time to
death at specified values of randomization assignment and other covariates.
14 INTRODUCTION AND CHARACTERIZATION

Table 1.1 Time in days to severe (stage ≥ 2) acute graft versus host disease (A-GVHD), death,
or last contact for bone marrow transplant patients treated with cyclosporine and methotrexate
(CSP + MTX) or with MTX onlya
CSP+MTX MTX
Time LAF Age Time LAF Age Time LAF Age Time LAF Age
3∗ 0 40 324∗ 0 23 9 1 35 104∗ 1 27
8 1 21 356∗ 1 13 11 1 27 106∗ 1 19
10 1 18 378∗ 1 34 12 0 22 156∗ 1 15
12∗ 0 42 408∗ 1 27 20 1 21 218∗ 1 26
16 0 23 411∗ 1 5 20 1 30 230∗ 0 11
17 0 21 420∗ 1 23 22 0 7 231∗ 1 14
22 1 13 449∗ 1 37 25 1 36 316∗ 1 15
64∗ 0 20 490∗ 1 37 25 1 38 393∗ 7 27
65∗ 1 15 528∗ 1 32 25∗ 0 20 395∗ 0 2
77∗ 1 34 547∗ 1 32 28 0 25 428∗ 0 3
82∗ 1 14 691∗ 1 38 28 0 28 469∗ 1 14
98∗ 1 10 769∗ 0 18 31 1 17 602∗ 1 18
155∗ 0 27 1111∗ 0 20 35 1 21 681∗ 0 23
189∗ 1 9 1173∗ 0 12 35 1 25 690∗ 1 9
199∗ 1 19 1213∗ 0 12 46 1 35 1112∗ 1 11
247∗ 1 14 1357∗ 0 29 49 0 19 1180∗ 0 11
Source: Kalbfleisch and Prentice (2002, Table 1.2)
a
Asterisks indicate that time to severe A-GVHD is right censored; that is, the patient died without severe A-GVHD or
was without severe A-GVHD at last contact.

1.7.2 Australian twin data

Duffy, Martin, and Mathews (1990) present analyses of ages at appendectomy for
twin pairs in Australia. This study was conducted, in part, to compare monozygotic
(MZ) and dizygotic (DZ) twins with respect to the strength of dependency in ages
of occurrence between pair members for various outcomes, including vermiform ap-
pendectomy. Twin pairs over the age of 17 years were asked to provide information
on the occurrence, and age at occurrence, of appendectomy and other outcomes. Re-
spondents not undergoing appendectomy prior to survey, or suspected of undergoing
prophylactic appendectomy, give rise to right-censored ages at appendectomy. There
were 1953 twin pairs, comprised of 1218 MZ and 735 DZ pairs. Among MZ pairs
there were 144 pairs in which both members, 304 pairs in which one member, and
770 in which neither member underwent appendectomy. The corresponding numbers
for DZ twins were 63, 208 and 464, respectively.
Comparison of strength of dependency between MZ and DZ twins may provide
insight into the importance of genetic factors for the outcome under consideration,
though the possibility that environmental factors (e.g., diet, physical activity patterns)
are more closely shared by MZ than by DZ twins also needs to be entertained. Also,
the twin pair ascertainment process may need to be modeled for some analytic pur-
poses. Here, for purposes of illustration (Chapter 3), as in Prentice and Hsu (1997),
SOME APPLICATION SETTINGS 15
we will regard these data as a simple cohort study of twin pairs without allowing
for the possible analytic impact of ascertainment criteria, continued survival, and
willingness (of both pair members) to respond to the study survey.

1.7.3 Women’s Health Initiative hormone therapy trial

A setting that we will use for illustration in various chapters is the Women’s Health
Initiative (WHI) hormone therapy trials. The WHI, in which the authors are engaged,
is a National Institutes of Health–sponsored disease prevention and population sci-
ence research program among post-menopausal women in the United States. It in-
cludes a multifaceted randomized controlled trial (RCT) of four different preventive
interventions (treatments) in a partial factorial design, and a prospective observa-
tional cohort study. The RCT included 68,132 women, each of whom enrolled ei-
ther in a postmenopausal hormone therapy (HT) trial (27,347 women) or a low-fat
dietary pattern intervention trial (48,835 women), or both, while the observational
study enrolled 93,676 women from essentially the same catchment populations in
proximity to 40 clinical centers across the United States. The HT trials included two
separate randomized, placebo controlled comparisons: conjugated equine estrogen
(CEE-alone; 0.625 mg/d continuous) among women who were post-hysterectomy
(10,739 women), or the same estrogen preparation plus medroxyprogesterone acetate
(CEE + MPA; 2.5 mg/d MPA continuous) among women having a uterus (16,608
women). Postmenopausal estrogens became widely used in the 1960s as an effective
means of controlling vasomotor symptoms associated with the menopause. Proges-
tational agents were added starting in the late 1970s to protect the uterus, when a 5-
to 10-fold increase in uterine cancer risk was observed among CEE-alone users. By
the time the WHI trials began in 1993, about 8 million women were using the pre-
cise CEE-alone regimen studied and about 6 million women were using the precise
CEE + MPA regimen studied, in the United States alone. These potent hormonal
preparations lead to approximate doubling of blood estrogens (estradiol, estrone, es-
trone sulfate) and an approximate doubling of their offsetting sex-hormone binding
globulin, whether or not progestin is included. In a mass spectrometry-based serum
proteomic profiling study, evidence emerged that CEE-alone and/or CEE + MPA
changed circulating protein concentrations for nearly half of the 350 proteins quanti-
fied, including proteins involved in growth factors, inflammation, immune response,
metabolism and osteogenesis, among other biological pathways. Hence it may not be
surprising that the use of these preparations has implications of the risk of important
chronic diseases. The CEE + MPA trial intervention (active treatment or placebo)
was stopped early, in 2002, on the recommendation of the external Data and Safety
Monitoring Committee for the WHI, when an increase in breast cancer risk was ob-
served, and a designated global index, defined as the time to the earliest of coronary
heart disease (CHD), stroke, pulmonary embolism, (invasive) breast cancer, col-
orectal cancer, endometrial cancer, hip fracture or death from any other cause, was
also in the unfavorable direction. These results were quite a shock to researchers
and practitioners, especially concerning CHD, for which an early hazard ratio (HR)
elevation was observed. In fact CHD was the designated primary outcome in the
16 INTRODUCTION AND CHARACTERIZATION
HT trials, and a major risk reduction was hypothesized for both regimens, primarily
based on an extensive observational epidemiology literature. A more sustained HR
elevation in stroke risk was also observed, as was a substantial early HR elevation in
pulmonary embolism, and in venous thromboembolism more generally. On the other
side of the ledger, the rate of colorectal cancer diagnosis was lower in the active treat-
ment compared to the placebo group, and the incidence of hip fracture, and of bone
fractures more generally, were lower with active treatment. Nevertheless, the global
index mentioned above was decidedly in the unfavorable direction, contributing to
the early trial stoppage decision.
The CEE-alone trial intervention was also stopped early, in 2004, about a year be-
fore the planned completion of its intervention phase, substantially because a stroke
risk elevation of similar magnitude to that for CEE + MPA was observed. However,
the global index was essentially null for this trial, due to a balancing of health bene-
fits and risks. In particular breast cancer risk was somewhat lower in the CEE-alone
treatment versus the placebo group, and fracture risk reduction of similar magnitude
to that with CEE + MPA was observed. Follow-up continues, to explore the long-
term effects of an average 5.6 years of CEE + MPA treatment, or an average of 7.1
years of CEE-alone treatment, more than 14 years after the conclusion of the HT
trial’s intervention phase.
In these controlled trials it is natural to define failure times for each woman as the
time from trial enrollment to disease event occurrence. Individual women can be fol-
lowed to observe the time to occurrence of CHD (T1 ), breast cancer (T2 ), stroke (T3 ),
hip fracture (T4 ), among other clinical outcomes. Even though these are among the
most common chronic diseases among postmenopausal women, only a few percent
of trial enrollees experienced any one of these types of failures during trial follow-up,
with other women having censored times, as determined by death, loss to follow-up,
or alive and under active follow-up without the study disease at the cutoff time for
data analysis. Analyses to date have mostly looked at hazard rates for clinical out-
comes one at a time, with follow-up continuing even after the occurrence of nonfatal
events of other disease types. Some univariate failure time data analyses from the HT
trials will be described in Chapter 2 and multivariate analyses will be considered in
Chapters 5 and 6.
The availability of multivariate failure time data (T1 , T2 , . . .) opens up some ad-
ditional data analytic opportunities: Women at elevated risk for stroke also tend to
be at elevated risk for other cardiovascular diseases, including CHD. One can ask
whether the incidence data on CHD can be used to increase the precision of hazard-
based treatment evaluation for stroke, or whether some composite CHD and stroke
outcomes may be informative. This type of “strength borrowing” is commonplace
in other data analytic contexts, for example, using generalized estimating equations,
but is it of practical importance for failure time data and, if so, how should analyses
be carried out? Any such strength borrowing would require, for example, continued
follow-up for T1 following a T2 event, and events of a particular disease type may
then need to distinguish fatal from non-fatal disease events. The nature and strength
of dependencies between non-fatal disease occurrences could be of some interest it-
SOME APPLICATION SETTINGS 17
self in this type of setting, particularly if the study treatment had an influence on such
dependencies.
Some of the diseases relevant to hormone therapy may occur on more than one
occasion during the follow-up of a particular woman as for venous thromboembolic
events, fractures, or even cancers of certain types either through recurrence or a new
primary diagnosis. One can ask whether treatment assessments can be more infor-
mative if the total event history during follow-up, rather than time to the first event
of each specific type, is included for each woman; and whether data summaries, such
as comparisons of estimated mean numbers of events over a follow-up period of a
certain duration may be helpful.
An additional goal of multivariate failure time data analysis in some contexts
may be a summary of the effects of a treatment or an exposure across a range of
health-related outcomes, or across a set of failure-types more generally. While the
construction of summaries of this type may require data beyond that available in a
given study, or judgments by subject matter specialists concerning the severity or
impact of each failure type, the value of suitable summary measures, or indices, is
quite evident for study monitoring and reporting purposes, and it is useful to examine
methodologic approaches to addressing statistical aspects of such data summaries.
See Anderson et al. (2007) for further detail on statistical aspects of the monitoring
and reporting on the WHI hormone therapy trials.
The multivariate failure time variables T1 , T2 , . . . defined above can be considered
in this context as correlated failure times on the same time from randomization axis.
The potential censoring times for failures of different types may vary however, since
outcome ascertainment procedures and outcome data sources may vary by failure
type.

1.7.4 Bladder tumor recurrence data

Table 1.2, obtained from Table 9.2 of Kalbfleisch and Prentice (2002), shows data
from a randomized trial of patient recurrences of superficial bladder tumor recur-
rence as conducted by the Veterans Administration Cooperative Urological Group.
As discussed in Byar (1980) these data were used to compare the frequency of re-
currences among 48 patients assigned to placebo, among whom there were a total
of 87 post-randomization recurrences, and 38 patients assigned to treatment with a
drug called thiotepa, among whom there were 45 recurrences during the trial follow-
up period, which averaged about 31 months. Tumors present at baseline were re-
moved transurethrally prior to randomization. In addition to studying the influence
of thiotepa on recurrence frequency, there was interest in the dependence of such
recurrence rates on the number and size of prerandomization tumors. Note that some
patients experience multiple recurrences during the study follow-up period, and use-
ful data analyses may focus on the mean number of recurrences during a specified
follow-up period. However, care may be needed in the modeling of data of this type
since patients experiencing recurrences could be more likely to cease trial participa-
tion prematurely, giving rise to censoring rates that could depend on the preceding
failure time history for the patient. Also note in Table 1.2 that recurrence times have
18 INTRODUCTION AND CHARACTERIZATION

Table 1.2 Bladder tumor recurrence data

a
Initial Tumors Initial Tumors
Censoringb Recurrence Timesb Censoring Recurrence Times
Number Size Time T1 , T2 . . . Number Size Time T1 , T2 . . .
Placebo Group
1 1 0 1 5 30 2,17,22
1 3 1 2 1 30 3, 6, 8, 12, 26
2 1 4 1 3 31 12, 15, 24
1 1 7 1 2 32
5 1 10 2 1 34
4 1 10 6 2 1 36
1 1 14 3 1 36 29
1 1 18 1 2 37
1 3 18 5 4 1 40 9, 17, 22, 24
1 1 18 12, 16 5 1 40 16, 19, 23, 29, 34, 40
3 3 23 1 2 41
1 3 23 10, 15 1 1 43 3
1 1 23 3, 16, 23 2 6 43 6
3 1 23 3, 9, 21 2 1 44 3, 6, 9
2 3 24 7, 10, 16, 24 1 1 45 9, 11, 20, 26, 30
1 1 25 3, 15, 25 1 1 48 18
1 2 26 1 3 49
8 1 26 1 3 1 51 35
1 4 26 2, 26 1 7 53 17
1 2 28 25 3 1 53 3, 15, 46, 51, 53
1 4 29 1 1 59
1 2 29 3 2 61 2, 15, 24, 30, 34, 39, 43, 49, 52
4 1 29 1 3 64 5, 14, 19, 27, 41
1 6 30 28, 30 2 3 64 2, 8, 12, 13, 17, 21, 33, 49

Thiotepa Group
1 3 1 8 3 36 26, 35
1 1 1 1 1 38
8 1 5 5 1 1 39 22, 23, 27, 32
1 2 9 6 1 39 4, 16, 23, 27, 33, 36, 37
1 1 10 3 1 40 24, 26, 29, 40
1 1 13 3 2 41
2 6 14 3 1 1 41
5 3 17 1, 3, 5, 7, 10 1 1 43 1, 27
5 1 18 1 1 44
1 3 18 17 6 1 44 22, 20, 23, 27, 38
5 1 19 2 1 2 45
1 1 21 17, 19 1 4 46 2
1 1 22 1 4 46
1 3 25 3 3 49
1 5 25 1 1 50
1 1 25 4 1 50 4, 24, 47
1 1 26 6, 12, 13 3 4 54
1 1 27 6 2 1 54 38
2 1 29 2 1 3 59
Source: Kalbfleisch and Prentice (2002, p. 292)
a
Initial number of tumors of 8 denotes 8 or more; Size denotes size of largest such tumor in centimeters.
b
Censoring and recurrence times are measured in months.

been grouped into months, giving a moderate number of tied recurrence times. Data
analysis methods that can accommodate these types of tied times without incurring
appreciable bias are needed for this and other applications.
SOME APPLICATION SETTINGS 19
1.7.5 Women’s Health Initiative dietary modification trial
As introduced in §1.7.3 a total of 48,835 post-menopausal US women were assigned
to either a low-fat dietary pattern intervention (40%) or to a usual diet comparison
group (60%) as a part of the multifaceted Women’s Health Initiative clinical trial. In-
tervention group women were taught nutritional and behavioral approaches to mak-
ing a major dietary change, in groups of size 10–15 led by nutritionists. The dietary
goals included fat reduction to 20% of energy (calories), fruit and vegetable increase
to 5 servings/day, and grains increase to 6 servings/day. The comparison group re-
ceived printed health-related materials only. Breast and colorectal cancer incidence
were designated primary outcomes for disease risk reduction, and coronary heart
disease (CHD) and total cardiovascular disease (CVD) incidence were designated
secondary trial outcomes.
The trial proceeded to its planned termination (March 31, 2005) at which time
breast cancer incidence results were in the favorable direction for intervention versus
comparison-group women, but not statistically significant (p = 0.09) at conventional
levels. There was no evidence of an intervention influence on colorectal cancer inci-
dence; and CHD and overall CVD results were also neutral in spite of evidence of
favorable change in low-density lipoprotein cholesterol among intervention, but not
comparison group, women.
These findings were somewhat disappointing, given the magnitude of effort re-
quired to mount such a large, complex trial. However, the adherence of intervention
women to dietary fat goals was only about 70% of that anticipated in the trial de-
sign resulting in loss of power for trial outcomes, and the differential breast cancer
incidence in the intervention versus the comparison group was also about 70% of
that projected in the trial design. Also, this nutritional and behavioral intervention
can be projected to favorably influence a range of other important outcomes, includ-
ing disease-specific and total mortality. This opens the possibility of more definitive
results for composite outcomes, such as breast cancer followed by death from any
cause. In spite of much reduced incidence rates for composite outcomes (double fail-
ures) of this type, randomization comparisons may have greater power than either
of the marginal hazard rate comparisons, depending on the strength or relationship
between the double failure hazard rates and the randomization indicator variable. HI
investigators have recently conducted further trial analyses of this type, finding nom-
inally significant intention-to-treat effects on breast cancer followed by death, and on
diabetes requiring insulin injections. The related bivariate failure time analyses will
provide illustration in Chapters 4 and 7. Another valuable development in a recent
round of data analysis was the identification of post-randomization confounding by
differential use of statins between randomization groups among women who were
hypertensive at baseline or who had prior cardiovascular disease. Statin use in the
trial cohort increased markedly during the trial follow-up period, and these potent
preparations are known to have a strong influence on low-density lipoprotein choles-
terol concentrations in the blood, and on coronary heart disease incidence. In contrast
there was no evidence of such confounding among baseline healthy (normotensive,
without prior CVD) women, and in this stratum intervention group, women experi-
20 INTRODUCTION AND CHARACTERIZATION
enced a Bonferroni–adjusted significantly lower CHD incidence compared to com-
parison group women. Trial women have continued to be followed for clinical out-
comes during the post-intervention period, and new information is still emerging on
the effects of this low-fat intervention program on clinical outcomes during the com-
bined intervention and post-intervention follow-up periods. Statistical methods are
needed to take full advantage of the wealth of data obtained in this type of enterprise,
including analyses of single and double hazard failure rates in relation to random-
ization indicator variables and in relation to other study subject characteristics and
exposures over the trial follow-up period.

BIBLIOGRAPHIC NOTES
There is a long history of modeling failure time data using survivor and hazard
functions. The preface lists a number of books that describe these functions, and
their estimation under independent censorship, in some detail, including early books
by Kalbfleisch and Prentice (1980, 2002), Breslow and Day (1980, 1987), Lawless
(1983, 2002), Cox and Oakes (1984), Fleming and Harrington (1991), and Andersen
et al. (1993). A thorough account of product integration, as in (1.1), is given by
Gill and Johansen (1990) with applications to failure time data (see Appendix A).
Dabrowska (1988) provided the nice representation (1.6), which expresses the sur-
vivor function in terms of its marginal hazard rates and dependency rates that contrast
the double failure hazard rate to the product of corresponding single failure hazard
rates locally. Dabrowska (1988) also alludes to higher dimensional representation
from which (1.8) derives. See also Gill and Johansen (1990) and Prentice and Zhao
(2018) for such higher dimensional representation. Clayton (1978) introduced the
bivariate survivor function model (1.7) with θ > 0, which was further developed
by Oakes (1982, 1986, 1989). This model was generalized to higher dimensions in
Prentice (2016). The focus here on modeling marginal hazard rates, and on marginal
single and double failure hazard rates, will be used in an attempt to provide a unified
presentation throughout this book. Key references for marginal hazard rate analyses
for single failure hazard rates include Wei, Lin, and Weissfeld (1989), Spiekerman
and Lin (1998), and Lin et al. (2000). The literature on the modeling and analysis of
recurrent events is described in some detail in Cook and Lawless (2007) while the
same authors have recently provided (Cook & Lawless, 2018) a detailed account of
multistate models for event history analyses more generally. Anderson (1984) pro-
vides a unified record of multivariate normal-based modeling and estimation proce-
dures. Key references for mean and covariance estimation with (uncensored) discrete
and continuous data include Liang and Zeger (1986), Zeger and Liang (1986), and
Prentice and Zhao (1991).
SOME APPLICATION SETTINGS 21
EXERCISES AND COMPLEMENTS

Exercise 1.1
Show that the Clayton–Oakes bivariate survivor function (1.7) for failure time vari-
ates T1 , and T2 given by

F(t1 ,t2 ) = {F(t1 , 0)−θ + F(0,t2 )−θ − 1}−1/θ ,

for θ > 0 has marginal survivor functions given by F(t1 , 0) and F(0,t2 ) and a time-
independent cross ratio function equal to 1 + θ . Also, by considering log F(t1 ,t2 ),
show this survivor function converges to the independence special case F(t1 ,t2 ) =
F(t1 , 0)F(0,t2 ) as θ ↓ 0. Further show that this survivor function can be extended to

F(t1 ,t2 ) = {F(t1 , 0)−θ + F(0,t2 )θ − 1}−1/θ ∨ 0

to allow negative dependencies (θ < 0), that this distribution approaches the upper
Fréchet bound of F(t1 , 0) ∧ F(0,t2 ) for maximal positive dependency as θ → ∞, and
approaches the lower Fréchet bound of {F(t1 , 0) + F(0,t2 ) − 1} ∨ 0 for maximal neg-
ative dependency as θ → −1. Comment on the extent to which absolute continuity
is retained as θ becomes increasingly negative.

Exercise 1.2
The trivariate survivor function, for failure time variates T1 , T2 and T3 given by

F(t1 ,t2 ,t3 ) = {F(t1 , 0, 0)−θ + F(0,t2 , 0)−θ + F(0, 0,t3 )−θ − 2}−1/θ ∨ 0

θ ≥ −1 is sometimes considered as the trivariate generalization of the Clayton and

Oakes model (1.7). Show that the pairwise marginal cross ratio functions are each
constant and equal to 1+θ at all continuity points. Discuss whether this model would
be suited to the analysis of a family breast cancer data set in which T1 , T2 , and T3
denote ages of breast cancer occurrence for an index case (T1 ), her sister (T2 ) and
her daughter (T3 ).
Consider the more general trivariate survivor function (1.9). Show that the pair-
wise marginal survivor functions are given by F(t1 ,t2 , 0), F(t1 , 0,t3 ) and F(0,t2 ,t3 ).
Suppose that these marginal distributions have time-independent cross ratios of
1 + θ110 , 1 + θ101 and 1 + θ011 respectively at continuity points for the three vari-
ates. Derive the trivariate hazard rate function Λ111 and show that Λ111 approaches
zero at all continuity points as θ → −1. From this calculate the trivariate dependency
function Ω111 given in (1.10). Discuss any implications for modeling the distribution
of (T1 , T2 , T3 ).
22 INTRODUCTION AND CHARACTERIZATION
Exercise 1.3
For absolutely continuous failure time variates T1 , T2 and T3 show that
Z t1 Z t2 Z t3
{d 3 log F(s1 , s2 , s3 )/ds1 , ds2 , ds3 }ds1 , ds2 , ds3
0 0 0
= log F(t1 ,t2 ,t3 ) − log F(t1 ,t2 , 0) − log F(t1 , 0,t3 ) − log F(0,t2 ,t3 )
+ log F(t1 , 0, 0) + log F(0,t2 , 0) + log F(0, 0,t3 ).

Also show that

d 3 F(s1 , s2 , s3 ) . d 3 log F(s1 , s2 , s3 )
F(s1 , s2 , s3 ) =
ds1 ds2 ds3 ds1 ds2 ds3
2
d log F(s1 , s2 , s3 ) d log F(s1 , s2 , s3 )
+
ds1 ds2 ds3
2
d log F(s1 , s2 , s3 ) d log F(s1 , s2 , s3 )
+
ds1 ds3 ds2
d2 log F(s1 , s2 , s3 ) d log F(s1 , s2 , s3 )
+
ds2 ds3 ds1
d log F(s1 , s2 , s3 ) d log F(s1 , s2 , s3 ) d log F(s1 , s2 , s3 )
+ .
ds1 ds2 ds3

Integrate this latter expression over (0,t1 ] × (0,t2 ] × (0,t3 ] and combine with the
former expression to yield the absolutely continuous survivor function representation
(1.8).

Exercise 1.4
Derive a test for whether the censoring times in Table 1.2, in a given treatment group,
depend on the preceding bladder tumor recurrence pattern for each patient as ob-
served during trial follow-up.

Exercise 1.5
Consider discrete failure time variates (T1 , T2 ). Show that the quantity in square
brackets in the double product integral on the right side of (1.6) can be expressed
as F(s− − − −
1 , s2 )F(s1 , s2 )/{F(s1 , s2 )F(s1 , s2 )} and thereby show through massive can-
cellation that this double product integral reduces to F(t1 ,t2 )/{F(t1 , 0)F(0,t2 )}, so
that the right side of (1.6) equals F(t1 ,t2 ).

Exercise 1.6
Consider m > 2 failure time variates T1 , . . . , Tm having joint survivor function (θ ≥
−1) (Prentice, 2016) given by
SOME APPLICATION SETTINGS 23

F(t1 , . . . ,tm ) = {F(t1 , . . . ,tm−1 , 0)−θ

+ F(t1 , . . . ,tm−2 , 0,tm )−θ + · · · + F(0,t2 , . . . ,tm )−θ
− F(t1 , . . . ,tm−2 , 0, 0)−θ − F(t1 , . . . ,tm−3 , 0,tm−1 , 0)−θ − · · · −
F(0,t2 , . . . ,tm−1 , 0)−θ − · · · − F(0, 0,t3 , . . . ,tm )−θ
+F(t1 , . . . ,tm−3 , 0, 0, 0)−θ + · · · + F(0, 0, 0,t4 , . . . ,tm )−θ − · · · −
+(−1)m−2 F(t1 , 0 . . . , 0)−θ +(−1)m−2 F(0,t2 , 0 . . . , 0)−θ + · · · +
(−1)m−2 F(0, . . . , 0,tm )−θ +(−1)m−1 }−1/θ ∨ 0
where, for example, the second component of this expression consists of all
m(m − 1)/2 pairs with two arguments equal to zero and the third component is
composed of all m(m − 1)(m − 2)/6 triplets with three arguments equal to zero.
Show that the upper bound for this survival probability given by F(t1 , . . . ,tm−1 , 0) ∧
F(t1 , . . . ,tm−2 , 0,tm ) ∧ . . . F(0,t2 , . . . ,tm ) is approached as θ → ∞, and that the lower
bound is given by the expression above evaluated at θ = −1. Also show that as
θ → 0 this survival probability approaches the product of marginal survival proba-
bilities having a positive coefficient divided by the product of all survival probabili-
ties having a negative coefficient on the right side of the above expression. Can you
develop an expression for F(t1 , . . . ,tm ) in terms of marginal survival probabilities of
dimension q or less, where 1 ≤ q < m upon assuming that all marginal probabilities
of dimension more than q have a survivor function of this same form, with the same
θ value?

Exercise 1.7
Suppose that failure time variates T1 and T2 are statistically independent given the
value of a shared random effect W , where W is a gamma variate rescaled to have
mean one and variance θ > 0 that acts multiplicatively on the hazard rate. Show that
the Clayton (1978) model
F(t1 ,t2 ) = {F(t1 , 0)−θ + F(0,t2 )−θ − 1}1/θ
then arises by integrating over the joint distribution of the random effect, which is
often referred to as a “frailty” variate in this context. Generalize this result to m > 2
failure time variates and derive pairwise marginal cross ratio functions for each pair
of the m variates.

Exercise 1.8
Consider the WHI hormone therapy trial context of §1.7.3 with T1 defined as time
from randomization to CHD and T2 time from randomization to stroke. Describe
the difference in interpretation between the marginal T1 hazard process Λ10 given
by Λ10 {dt1 , 0; Z(t1 , 0)} and the recurrent event intensity process for T1 , given by
Λ1 {dt1 ; Ht } = P{CHD event in [t,t + dt); Ht } as in (1.11).
24 INTRODUCTION AND CHARACTERIZATION
Exercise 1.9
In the same WHI hormone therapy trial context (§1.7.3) define T1 as time from ran-
domization to breast cancer diagnosis, and T2 as time from breast cancer diagnosis
to death following breast cancer. Write down expressions for hazard rate processes
for T1 and for T2 given T1 ≤ t1 . Can you develop an expression for the hazard rate
for the composite time from randomization to death following breast cancer outcome
T3 = T1 + T2 in terms of these component hazard functions. Discuss the advantages
and disadvantages of comparing randomization groups in terms of T3 hazard rates
versus separate analyses for T1 and T2 hazard rates.
 Chapter 2

Univariate Failure Time Data Analysis

Methods

2.1 Overview 25
2.2 Nonparametric Survivor Function Estimation 25
2.3 Hazard Ratio Regression Estimation Using the Cox Model 27
2.4 Cox Model Properties and Generalizations 31
2.5 Censored Data Rank Tests 32
2.6 Cohort Sampling and Dependent Censoring 33
2.7 Aplastic Anemia Clinical Trial Application 35
2.8 WHI Postmenopausal Hormone Therapy Application 35
2.9 Asymptotic Distribution Theory 39
2.10 Additional Univariate Failure Time Models and Methods 44
2.11 A Cox-Logistic Model for Continuous, Discrete or Mixed Failure Time Data 45

2.1 Overview
There is an extensive literature on statistical modeling and estimation for a univari-
ate failure time variate T > 0. In this chapter some core methods, which we will
build upon in subsequent multivariate failure time methods presentations, will be de-
scribed. The core methods include Kaplan–Meier survivor function estimation, Cox
model hazard ratio parameter estimation with its associated logrank test, as well as
other censored data rank tests. The presentation will focus on nonparametric and
semiparametric likelihood formulations for estimator development for reasons men-
tioned in Chapter 1. A brief account of asymptotic distribution theory for these testing
and estimation procedures will also be given.

2.2 Nonparametric Survivor Function Estimation

Consider a failure time variate with T > 0 having survivor function F, so that F(t) =
P(T > t) for any t ≥ 0. Suppose that n individuals are chosen at random from a study

25
26 UNIVARIATE FAILURE TIME
population, and are followed forward from t = 0 to observe individual failure times,
subject to independent right censoring. Here, independent right censoring means that
the set of individuals without prior failure or censoring has a hazard rate equal to that
for the study population, at any follow-up time t > 0. Denote by t1 < t2 < · · · < tI
the ordered distinct failure times in the sample, and suppose that di individuals fail
at ti , out of the ri individuals who are without failure or censoring prior to time
ti , i = 1, . . . , I. A nonparametric likelihood function for F can be written
n h i
L = ∏ {−F(dsk )}δk F(sk )1−δk , (2.1)
k=1

where sk is observed and is the smaller of the failure or censoring time for the kth
individual in the sample and δk takes a value of 1 if sk is uncensored and a value
0 if sk is censored. Expression (2.1) can be maximized within the class of discrete,
continuous and mixed survivor functions by placing mass (probability) only at the
observed uncensored failure times, or on the half line beyond the largest sk value if
uncensored. Doing so yields a discrete, step function estimator of F, starting with
F(0) = 1. Substituting

F(ti ) = ∏{1 − Λ(dt` )} and − F(dti ) = ∏{1 − Λ(dt` )}Λ(dti )

`≤i `<i

using the discrete special case of (1.1) and collecting terms, then gives a partially
maximized likelihood of
I
L = ∏[Λ(dti )di {1 − Λ(dti )}ri −di ], (2.2)
i=1

which is maximized at Λ(dti ) = di /ri , i = 1, . . . , I, yielding the well-known Kaplan–

Meier (1958) product limit survivor function estimator

F̂(t) = ∏{1 − di /ri }. (2.3)

ti ≤t

Of course F is not identifiable at times where no individuals are “at risk” for
failure, so F̂ is undefined beyond the largest follow-up time (i.e., the largest sk value)
observed in the sample. The corresponding hazard function estimator Λ̂ is given by

Λ̂(t) = ∑ di /ri ,
ti ≤t

which is often referred to as the Nelson–Aalen estimator. Note that, in keeping with
(1.1)
t
F̂(t) = ∏{1 − Λ̂(ds)}.
0
Informally, the ith factor in (2.2) can be recognized as a binomial likelihood for
Λ(dti ). Conditioning on all failure and censoring information prior to ti , thereby fix-
ing ri , shows Λ̂(dti ) = di ri−1 to have a conditional, and hence an unconditional, mean
NONPARAMETRIC SURVIVOR FUNCTION ESTIMATION 27
of Λ(dti ). Similarly, for i < j, conditioning on all failure and censoring information
prior to t j fixes Λ̂(dti ) and shows Λ̂(dti ) and Λ̂(dt j ) to be conditionally, and hence
unconditionally, uncorrelated. Application of the delta method then leads to
ˆ F̂(t) = F̂(t)2 ∑ {di /(ri − di )}
var
ti ≤t

as a variance estimator for F̂(t), referred to as the Greenwood formula. To avoid

influences from the range restrictions, 0 ≤ F(t) ≤ 1, one can apply a normal distribu-
tion approximation to the unconstrained function given by log{−logF(t)} leading,
for example, to an approximate 95% confidence interval for F(t) of
F̂(t) exp{±1.96v̂(t)} (2.4)
where v̂(t)2 = var
ˆ F̂(t)/{logF̂(t)}2 at times where F̂(t) > 0.
One can also obtain the Kaplan–Meier estimator using mean parameter estimat-
ing equations as follows: Reconstrue the censored failure time data (sk , δk ) as a se-
quence of uncorrelated binary variates
(
1 if sk = ti and δk = 1
Wki =
0 otherwise

and also define “at-risk” indicator variables

(
1 if sk ≥ ti
Yki =
0 otherwise

for i = 1, . . . , I, for each k = 1, . . . , n. Under a discrete failure time model with hazard
rate Λ(dti ) at T = ti , i = 1, . . . , I one has
µki =E(Wki ) = Yki Λ(dti ), partial derivatives ∂ µki /∂ Λ(dti ) = Yki ,
(
µki (1 − µki ) if i = j
and cov{(Wki − µki )(Wk j − µk j )} =
0 otherwise.

This leads to mean parameter estimating equations (see Appendix A, expression

A.5)
n
∑ Yki (Wki − µki ) = 0, for i = 1, . . . , I
k=1
and to the Kaplan–Meier (KM) estimator. Hence the nonparametric maximum like-
lihood, the survivor function or hazard function representation plug-in, and the mean
parameter estimating equation approaches each lead to the same nonparametric sur-
vivor function estimator (2.3). F̂ has been shown to be strongly consistent for F,
and n1/2 (F̂ − F) has been shown to be weakly convergent to a mean zero Gaussian
process over a time period [0, τ], where τ is in the support of the observed follow-up
times. Moreover, in keeping with its nonparametric maximum likelihood develop-
ment, F̂ has also been shown to be nonparametric efficient as an estimator of F over
[0, τ]. Some detail on these asymptotic results will be given in §2.9.
28 UNIVARIATE FAILURE TIME
2.3 Hazard Ratio Regression Estimation Using the Cox Model
Suppose now that a covariate history {Z(t),t ≥ 0} is recorded for each individual in
the study sample. The hazard rate regression model introduced by Sir David Cox rev-
olutionized the analysis of censored failure time data, and Cox (1972) is one of the
most highly cited statistical papers of all time. The Cox model, for an absolutely con-
tinuous failure time variable T , specifies a hazard rate at time t given the preceding
covariate history Z(t), having the multiplicative form

λ {t; Z(t)} = λ0 (t) exp{x(t)β }, (2.5)

where x(t) = {x1 (t), . . . , x p (t)} is a modeled covariate p-vector comprised of data-
analyst-defined functions of Z(t) and possibly product terms between such functions
and t. This modeled regression variable, with sample paths that are continuous from
the left with limits from the right, is intended to “capture” the dependence of the
hazard rate at time t on the preceding covariate history, through the value of the
hazard ratio parameter β 0 = (β1 , . . . , β p ), where a prime (0) denotes vector transpose.
The function λ0 in (2.5) is referred to as the baseline hazard function, and λ0 (t) is the
hazard rate at a reference covariate history Z0 (t) for which the modeled covariate is
x(t) ≡ 0, a zero vector for all t. The hazard process model given Z is semiparametric
with the p-vector β and the nonparametric function Λ0 , where Λ0 (t) = 0t λ0 (s)ds, as
R

parameters to be estimated.
Denote by T1 , . . . , Tn the underlying failure times for a random sample of size
n from a study population followed forward in time from t = 0. Suppose that Tk is
subject to right censoring by a variate Ck , so that one observes Sk = Tk ∧Ck and non-
censoring indicator variable δk = I[Sk = Tk ]. Suppose also that covariate histories
Zk (Sk ) are recorded, k = 1, . . . , n. An independent censoring assumption requires the
hazard rate λ {t; Z(t)} in (2.5) to equal the same hazard rate, but with Ck ≥ t added to
the conditioning event. That is, independent censorship implies that the subset of in-
dividuals who are without prior failure or censoring at any follow-up time t, referred
to as the “risk set” at time t and denoted R(t), is representative of the study popu-
lation in terms of hazard rate at t given Z(t). This assumption needs to be carefully
considered in the context of specific applications, and may be able to be relaxed as
necessary.
A semiparametric likelihood function, analogous to (2.1), can be written
n n oδk  Zs
k

L = ∏ λ0 (sk )exk (sk )β exp − exk (u)β λ0 (u)du ,
k=1 0

where (sk , δk ) and Zk (sk ), k = 1, . . . , n are the observed data in the study sample.
Several approaches have been considered for dealing with the nonparametric as-
pect of this model, including partial likelihood (Cox, 1972, 1975), marginal likeli-
hood (Kalbfleisch & Prentice, 1973), and approximate likelihood (Breslow, 1974)
methods.
The Breslow (1974) approach begins by noting that the above likelihood can be
maximized by placing all failure probability within the risk region of the data, de-
fined by R = {t; sk ≥ t for some k ∈ (1, . . . , n)}, on the observed uncensored failure
HAZARD RATIO REGRESSION ESTIMATION USING THE COX MODEL 29
times t1 < t2 < · · · < tI in the sample. This implies that censored sk values can be
replaced by censored values at the immediately preceding uncensored failure time in
the sample without diminishing the likelihood. Now if one approximates the base-
line rate function by λ0 (t) = λi whenever t ∈ (ti−1 ,ti ], i = 1, . . . , I one can write the
likelihood, following the censored data shifting just mentioned, as
" #
I n o
xi (ti )β xk (ti )β
L = ∏ ∏ λi e ∏ exp −e λi ∆ti (2.6)
i=1 k∈D(ti ) k∈R(ti )

where D(ti ) denotes the set of di individuals having uncensored failures at T = ti , and
∆ti = ti −ti−1 , i = 1, . . . , I, with t0 = 0. In this form L can be recognized as having the
form of a parametric likelihood to which application of standard likelihood methods
can be considered.
Specifically, one can solve the equations ∂ log L/∂ λi = 0, i = 1, . . . I explicitly
giving ( )
λ̂i (β ) = di / ∆ti ∑ exk (ti )β , i = 1, . . . , I,
k∈R(ti )
which can be inserted into (2.6) to give the profile likelihood
 ,( )di 
I
L(β ) = ∏  ∏ exk (ti )β ∑ exk (ti )β  , (2.7)
i=1 k∈D(ti ) k∈R(ti )

which is Breslow’s tied data approximation to the Cox partial likelihood.

The corresponding estimating function for β is
U(β ) = ∂ log L(β )/∂ β
I
= ∑ Ui (β )
i=1
" #
I 
0 0
=∑ ∑ xk (ti ) − di ∑ xk (ti ) exp{xk (ti )β } ∑ exp{xk (ti )β } .
i=1 k∈D(ti ) k∈R(ti ) k∈R(ti )
(2.8)
Conditional on all failure, censoring and covariate information up to ti , one can
see, under (2.5), that the conditional, and hence also the unconditional, expectation
of Ui (β ) is zero for all, i = 1, . . . I, so U(β ) = 0 provides an unbiased estimating
equation for β . Similarly for i < j conditioning on all failure, censoring and covariate
information up to t j fixes Ui (β ) and gives a conditional, and hence unconditional
expectation of zero for Ui (β )U j (β ), all (i, j). Hence I(β ) = −∂ 2 log L/∂ β ∂ β 0 has
expectation equal to the variance matrix for U(β ). Direct calculation gives
I ∑k∈R(ti ) xk (ti )0 xk (ti )exk (ti )β

I(β ) = ∑ di
i=1 ∑k∈R(ti ) exk (ti )β
∑k∈R(ti ) xk (ti )0 exk (ti )β ∑k∈R(ti ) xk (ti )exk (ti )β

− . (2.9)
{∑k∈R(ti ) exk (ti )β }2
30 UNIVARIATE FAILURE TIME
Under mild conditions β̂ that maximizes (2.7), which is sometimes referred to as
the maximum partial likelihood estimate of β , has an asymptotic distribution, with
consistent variance matrix estimator, given by
1
n 2 (β̂ − β ) ∼ N {0, nI(β̂ )−1 }

where N denotes a normal distribution. Also, the corresponding baseline hazard

estimator Λ̂0 , where
"  #
Λ̂0 (t) = Λ̂0 (t, β̂ ) = ∑ di ∑ exp{xk (ti )β̂ } (2.10)
ti ≤t k∈R(ti )

1 1
is such that n 2 {Λ̂0 (·) − Λ0 (·)} converges jointly with n 2 (β̂ − β ) to a zero mean
Gaussian process over a time period [0, τ], where τ is in the support of the follow-up
times (S values). The estimator β̂ has also been shown to be semiparametric efficient
under (2.5) in the special case of time-independent covariates.
The Cox likelihood can be derived similarly by setting the overall sample empir-
ical hazard rates di /ri equal to the model-based average ∑`∈R(ti ) λ0 (ti )ex` (ti )β /ri and
solving for λ0 (ti ) at each i = 1, . . . , I. Plugging these values into the semiparametric
likelihood gives (2.7).
A mean parameter estimating equation development can also be considered: As
above set
( (
1 Sk = ti , and δk = 1 1 Sk ≥ ti
Wki = , and Yki = , for all (k, i).
0 otherwise 0 otherwise

Under a model of the form (2.5) one has µki = E(Wki ) = Yki eαi +xk (ti )β , where αi =
log Λ0 (dti ), the uncorrelatedness of Wki , i = 1, . . . , I for each k, and mean parameter
estimating equations (Appendix A, A.5) for (α1 , . . . , αI ) that solve
n
∑ (Wki − µki ) = 0, for i = 1, . . . , I.
k=1

Solving these equations gives

 n
eαi = di ∑ Yki exk (ti )β , i = 1, . . . , I,
k=1

which can be inserted into the estimating equation for β giving ∑Ii=1 Wki xki (ti )0 −
di ∑Ii=1 {∑nk=1 Yki xk (ti )0 exk (ti )β / ∑nk=1 Yki exk (ti )β } = 0, yielding (2.8) and (2.10). Also
the model-based variance estimator, using the notation of Appendix A, is
!−1
n I
∑∑ Yki D̂0kiV̂ki−1 D̂ki ,
k=1 i=1
COX MODEL PROPERTIES AND GENERALIZATIONS 31
where ˆ denotes evaluation at (α̂1 , . . . , α̂I , β̂ ) solving these equations, which equals
I(β̂ )−1 from (2.9).
It follows that maximum likelihood, empirical hazard plug-in, and mean param-
eter estimating functions again agree for this rather general regression estimation
problem. Generalizations of these approaches will be considered in later chapters
to manage nonparametric aspects of models specified for multivariate failure time
regression estimation.
The asymptotic developments mentioned above assume absolutely continuous
failure times, so that technically we should have di = 1, i = 1, . . . , I. However (β̂ , Λ̂0 )
as described above can tolerate some tied failure times without incurring appreciable
asymptotic bias. A rule of thumb may be that the number of ties di should not be
more than a few percent (e.g., 5%) of the size, ri , of the corresponding risk set at
uncensored failure times. Nearly all available computer software for the Cox model
allows tied failure times, and applies the expressions given above, even though more
sophisticated approximations have been proposed for handling tied failure times.

2.4 Cox Model Properties and Generalizations

A few points can be made about the Cox model estimation procedure described
above. First, it is easy to see that the hazard ratio function comparing covariate his-
tories Z1 and Z2 is given by

λ {t; Z1 (t)}/λ {t; Z2 (t)} = exp{x1 (t) − x2 (t)}β

at time t, and does not depend on the choice of baseline covariate history Z0 .
Also, it is worth commenting that the time-varying feature of (2.5) can be quite
powerful. This feature allows hazard ratios for a specific covariate to vary in a user-
defined fashion as a function of follow-up time, and it allows hazard rates to be
defined that condition on stochastic covariates that are recorded during study follow-
up.
From (2.10) one can specify
(  )
F̂{t; Z(t)} = ∏ 1 − di exp{xi (t)β̂ } ∑ exk (ti )β̂ ,
ti ≤t k∈R(ti )

which has a survivor function interpretation with time-independent covariates Z(t) ≡

z, or with evolving covariates that are external to the failure process in the sense that
the covariate paths are unaffected by the failure time process under study. Note that
F̂ reduces to the Kaplan–Meier estimator at β̂ = 0.
A simple, but very useful, relaxation of the Cox model (2.5) allows the baseline
hazard function to vary among strata, which also may be time-dependent. Specifi-
cally, one can write
λ {t; Z(t)} = λ0u (t) exp{x(t)β }, (2.11)
where stratum u = u{t, Z(t)} ∈ {1, . . . , m}, has sample paths that are continuous from
the left with limits from the right. Under (2.11) the study population is partitioned
32 UNIVARIATE FAILURE TIME
into m mutually exclusive strata at each follow-up time t. For example, this stratifi-
cation feature can allow the baseline hazard rate at time t to vary in an unrestricted
manner for some variables defined by {t, Z(t)}, while imposing a parametric hazard
ratio model on variables of principal interest that are used to define x(t). The hazard
ratio parameter β can be estimated under (2.11) using a product of factors (2.7), one
from each stratum, while U(β ) and I(β ) are given by sums over the strata of (2.8) and
(2.9). For example, in large epidemiologic cohort studies with most observed times
censored it is often possible to incorporate extensive baseline stratification, thereby
enhancing confounding control, with little effect on the efficiency with which hazard
ratio parameters of primary interest are estimated.
In some settings a univariate failure time variate T > 0 may be accompanied by a
failure type J ∈ {1, . . . , q}, sometimes referred to as a “mark.” The hazard ratio meth-
ods of the preceding section extend readily to failure type–specific hazard models
λ j {t; Z(t)} = λ0 j (t) exp{x(t)β j }, j = 1, . . . , q (2.12)
simply by applying a (partial) likelihood function that is a product of terms (2.7)
for each type-specific hazard ratio parameter, following the imposition of additional
censorship wherein follow-up times for estimating β j are additionally censored at the
time of failure of any type other than j. Data of this type are sometimes formulated
in terms of potential failure times, say U1 , . . . ,Uq with T = min(U1 , . . . ,Uq ), and the
estimation of marginal hazard rates, or other distributional characteristics, for these
latent failure times is referred to as the competing risk problem. However, the joint
distribution of (U1 , . . . ,Uq ) given Z, for a time-independent covariate, is not identifi-
able without strong additional assumptions, such as independence among such times,
unless genuinely multivariate failure time data are available. Genuinely multivariate
failure time data arise when individuals continue to be followed beyond the time of
their first failure to observe second and subsequent failure times for the individual.
The analysis of such multivariate failure time data is a major focus of this book.
If the failure time variate, T , includes discrete elements then some care may be
needed in applying (2.5) since the hazard rates at a mass point cannot exceed one.
Depending on the distribution of the modeled covariate x(t), a different regression
model form, such as a logistic hazard rate model, may be preferable if T has some
large point masses. This topic will be elaborated in §2.11.

2.5 Censored Data Rank Tests

In addition to Kaplan–Meier survival curves and Cox regression, censored data rank
tests that are used to test equality of two or more survival curves, can be included
among core univariate failure time methods.
The score test
" #
I
0 di 0
U(0) = ∑ ∑ xk (ti ) − ∑ xk (ti ) (2.13)
i=1 k∈D(t ) ri k∈R(t )
i i

from (2.8) can be used to test β = 0, by comparing

U(0)0V (0)−1U(0)
COHORT SAMPLING AND DEPENDENT CENSORING 33
to an asymptotic chi-square distribution on p degrees of freedom where V (0) is a
variance estimator for U(0) with x(t) is composed of indicator variables for p of
p + 1 samples whose failure rates are being compared. This p + 1 sample comparison
test is referred to as the logrank test. The corresponding hazard rates under (2.5)
are proportional to each other, so it should not be surprising that the logrank test
has attractive power properties for testing the null hypothesis against proportional
hazards alternatives. This class of null hypothesis tests can be broadened to
( )
I
d i
∑ ∑ h(ti )xk (ti )0 − ri ∑ h(ti )xk (ti )0 (2.14)
i=1 k∈D(t ) i k∈R(t ) i

by introducing a weight function h, where h(t) can depend on failure and censoring
information prior to t, in order to provide, for example, greater sensitivity to hazard
rate differences early versus late in the follow-up period, or vice versa. A correspond-
ing variance estimator under β = 0 is given by
I
V (0) = ∑ h(ti )2Vi
i=1

where (Vi ) j j = ri j (ri − ri j )di (ri − di )ri−2 (ri − 1)−1 , j = 1, . . . , p and (Vi ) jk =
−ri j rik di (ri −di )ri−2 (ri −1)−1 , j 6= k , where, for example, ri j is the size of the risk set
in the jth sample at ti . Tests of the form (2.14) are known as weighted logrank tests.
In addition to h(t) ≡ 1, another commonly used test specifies h(t) = F̂(t), with F̂
a survivor function estimator under the null hypothesis. This generalized Wilcoxon
test applies greater weight to early failures, compared to the logrank test. Weight
functions h that may depend on failure, but not censoring, distribution estimates are
preferable for test interpretation.

2.6 Cohort Sampling and Dependent Censoring

There are many variations in the criteria for study subject selection and follow-up
with univariate failure time data. For example, if T is defined to be age at disease
occurrence in an epidemiologic cohort study with a minimum age criterion, there
would typically be late entry into the cohort as study subjects are enrolled at ages
beyond the specified minimum. The hazard rate methods described above readily
adapt to such late entries by adjusting the risk sets R(t) to include only subjects who
are under active follow-up at time t.
Since some cohort studies involve very heavy censorship, with only a few per-
cent of study subjects experiencing the failure time outcome under study, it may be
inefficient or impractical to assemble cohort histories on the entire study cohort. Such
assembly, for example, may involve expensive laboratory analysis of stored biospeci-
mens. Cohort sampling methods, such as matched case–control or case–cohort meth-
ods are often applied. Hazard ratio parameter estimates under sampling in which each
failing individual (a case) is matched to one or more cohort members (controls) who
are without failure at the time of case failure occurrence (t) can be derived by ap-
plying (2.7) with D(t) and R(t) respectively the case and the combined case and its
Other documents randomly have
different content
Teut. huysken, theca.

HOG, s. A young sheep, before it has lost its first fleece.

Gl. Complaynt.
Stat. Acc.
L. B. hoggaeius, a young sheep of the second year.

HOG, s. In the diversion of curling, the name given to a stone which

does not go over the distance score, S.
Graeme.

To HOG, HOGG, v. a. To shog, Ang.

Old Ball.
Isl. hagg-a, commoveo, quasso.

HOGERS, HOGGERS, s. pl. Coarse stockings without feet, S.

Ross.

HOGMANAY, HOGMENAY, s.
1. The last day of the year, S.
2. The entertainment given to a visitor on this day; or a gift
conferred on those who apply for it, S.
J. Nicol.
The origin is quite uncertain.

HOGRY-MOGRY, adj Slovenly, Loth. corr. from hugger-mugger, E.

V. Hudge-mudge.

HOG-SCORE, s. A distance-line, in curling, drawn across the rink or

course, S.
 Burns.

HOG-SHOUTHER, s. A game in which those who amuse themselves

justle each other by the shoulders, S.
Burns.
Isl. hogg-a, to strike.
To Hogshouther, v. a. To justle with the shoulder.
Burns.

To HOY, v. a.
1. To incite, a term used as to dogs, S.
Burns.
2. To chase or drive away.
Lyndsay.
Isl. ho-a, greges convocare vel agere.

HOYES, s.
1. A term used in public proclamations, calling attention, S.
Skene.
O. Fr. oyez, hear ye.
2. Used as equivalent to hue, in the phrase hue and cry.
Stat. Rob. I.

HOIF, HOFF, HOVE, HOUFF, HUFE, s.

1. A hall.
Bellenden.
Su. G. hof, aula.
2. A burial-place. The principal place of interment at Dundee is
called the houff.
Isl. hof, atrium, Germ. hof, area, kirchhof, area ante
templum.
3. A haunt, S.
Burns.
A. S. hofe, Germ. hof, a house.
4. A place where one wishes to be concealed.
Ferguson.
A. S. hofe, spelunca, a den.

HOISPEHOY, s. A game used in Banffshire, similar to Hide and Seek.

O. Fr. oyez, hear, and espier, to spy; q. listen, I espy you.

To HOIST, v. n. To cough.
V. Host.

To HOIT, HOYTE, v. n. To move with expedition, but stiffly and

clumsily, S.
Burns.
Isl. haut-a, cursitare more detentae volucris.
Hoit, s. A hobbling motion, S. B.

HOLYN, HOLENE, s. The holly, S.

A. S. holen, id.
Wallace.

To HOLK, HOUK, HOWK, v. a.

1. To dig, S.
Douglas.
2. Metaph. to search.
R. Bruce.
Su. G. holk-a, cavare, from hol, cavus.

HOLKIS, s. pl. A disease of the eye; heuck, S. B.

Douglas.

To HOLL, v. n. To excavate, S.
A. S. hol-ian, id.
Holl, Howe, adj.
1. Hollow, deep; how, S.
Palice Hon.
2. Concave.
Douglas.
3. Giving a hollow sound, S.
Burns.
Isl. hol-ur, cavus, concavus.
Holl, s. Hold of a ship.
Wallace.

HOLLIGLASS, s. A character in old Romances.

Poems 16th Cent.
Belg. Uyle-spegel, i. e. Owl-glass; the original work being
written in Dutch.

HOLLION, s. Conjoined with hip, Ang.

Morison.

HOLM, HOWM, s. The level low ground on the banks of a river, S.,
hoam, S. B.
Isl. hwam-r, a little valley.
Wyntown.

HOLT, s. A wood; as in E.

HOLT, s.
1. High and barren ground.
Douglas.
Isl. hollt, terra aspera et sterilis.
2. A very small hay cock, or a small quantity of manure before it is
spread, Dumfr.
Statist. Acc.

HOME, adj. Close, urgent, S.

Spalding.

HOMELTY-JOMELTY, adj. Clumsy and confused in manner.

Dunbar.
Perh. from whummil, and jumble.

HOMYLL, adj. Having no horns; S. hummil, hummilt.

V. Hummil.
Bellenden.
Isl. hamla, membri mutilatione impedire.

To HOMOLOGATE, v. a. To give an indirect approbation of any thing,

S.
Burnet.
HONE, s. Delay.
Barbour.
Apparently from Hove, How, q. v.

HONEST, adj. Honourable.

Wyntown.
Honest-like, adj.
1. Goodly; as regarding the person, S.
2. As respecting dress; not shabby, S.
3. Having the appearance of liberality, or of plenty, S.
Honesty, s.
1. Respectability.
Wyntown.
2. Liberality, S.
Rutherford.
3. Decency, as becoming one's station, S.
Lat. honest-us, kind; decent.
Kelly.

HOO, s. Delay.
V. Hove.
Wallace.

HOO, s. Cap.
V. How.

HOODED CROW, The pewit gull, Orkn.

Barry.
To HOOL, v. a. To conceal, S. B.
Shirrefs.
Su. G. hoel-ja, operire; Moes. G. hul-jan.

HOOLIE, adj. Slow.

V. Huly.

To HOP, HAP, v. n. To dance.

Douglas.
Teut. hopp-en, salire, saltare.

HOP, HOPE, s. A sloping hollow between two hills, or the hollow that
forms two ridges between one hill, South of S.
Wallace.
Celt. hope, petite vallée entre des montagnes.

HOPE, s. A small bay, Orkn.

Wyntown.
Isl. hop, recessus maris.

HORIE GOOSE, the Brent goose; also horra, Orkn.

Statist. Acc.

HORN, s. A vessel for holding liquor; figuratively used for its

contents, S.
Isl. horn, poculum.
Ramsay.
Horn, s. An excrescence on the foot, a corn, S. B.
Sw. lik-thorn, id. q. a body-horn.
Horn, s. To put to the horn, to denounce as a rebel; a forensic
phrase; from the formality of blowing a horn, S.
Bellenden.
Horning, s. Or, Letter of Horning, a letter issued from his Majesty's
Signet, and directed to a Messenger, who is required to charge a
debtor to pay the debt for which he is prosecuted, or perform the
obligation within a limited time, under the pain of rebellion, S.
Erskine.
Horne, s. One of the constellations.
Doug.
Horn-daft, adj. Outrageous; perhaps in allusion to an animal that
pushes with the horn, S.

HORRING, s. Abhorrence.
Buchanan.

HORSE, s. A faucet, S. B.

HORSE-COUPER, s. A horse-dealer, S.
Colvil.

HORSE-GANG, s. A certain quantity of land, S.

Pennant.

HORSEGOUK, s. The green sand-piper, Shetl.

Dan. horse gioeg.
HORSE-KNOT, s. Common black knapweed, S.
Lightfoot.

HORSE-MUSCLE, s. The pearl oyster, S.

Statist. Acc.

HOSE-FISH, s. The cuttle-fish, S. O-fish, Loth.

Sibbald.

HOSE-NET, s.
1. A small net, affixed to a pole, resembling a stocking, S.
2. In a hose-net, in an entanglement, S.
R. Bruce.

To HOST, HOIST, v. n.
1. To cough, S.
Henrysone.
2. Metaph. to belch up; applied to the effusions of grief or
displeasure.
Doug.
3. To hem, S.
A. S. hweost-an, Su. G. host-a, id.
Host, Hoast, Hoist, s.
1. A single act of coughing, S.
Dunbar.
2. A settled cough, S.
K. Hart.
3. A hem, S.
4. Denoting what is attended with no difficulty or hesitation. It did
na cost him a host, S.
Ross.
A. S. hweost, Belg. hoest, id.
HOSTA, interj. Expressing surprise, and perhaps hesitation, Ang.
Shirrefs.
Moes. G. haus-jan, audire.

To HOSTAY, v. a. To besiege.
Fr. hostoyer, id.
Wyntown.

HOSTELER, HOSTELLAR, s. An innkeeper.

Fr. hostelier, id.
Wallace.
Hostillar, Hostillarie, s. An inn.
Fr. hostelerie, id.
Acts Ja. I.

To HOTCH, v. n. To move the body by sudden jerks, S.

Teut. huts-en, Belg. hots-en, Fr. hoch-er, to jog.

HOTCH-POTCH, s. A dish of broth, made with mutton or lamb, cut

into small pieces, together with green peas, carrots, turnips, and
sometimes parsley or celery, served up with the meat in it, S.
Teut. huts-pot, Fr. hochepot.

To HOTTER, v. a. To crowd together; expressive of individual motion,

S. O.
Teut. hott-en, coalescere.
J. Nicol.
HOU, s. A roof-tree.
V. How, s. 4.
Ramsay.

To HOUD, v. n.
1. To wriggle, S.
2. To move by succussation, Loth.
Houd, s. The act of wriggling, S. B.

To HOVE, HOW, HUFE, HUFF, v. n.

1. To lodge.
Barbour.
2. To stay, to tarry.
Douglas.
Germ. hof-en, domo et hospitio excipere.

To HOVE, v. n.
1. To swell, S.
Hogg.
2. To rise, to ascend.
Polwart.
Dan. hov-er, to swell.

HOVE. Arthur's Hove, the ancient building called Arthur's Oon, S.

Bellenden.

HOUFF, s. A haunt.
V. Hoif.
To Houff, v. n. To take shelter, S.
HOUFFIT, part. Heaved.
K. Hart.

HOUGH, adj. Having a hollow sound.

Glanville.

HOUGH, adj. (gutt.)

1. Low, mean; pron. hogh.
Ramsay.
2. In a poor state of health, S.

To HOUK, v. a. Expl. To heap.

Gl. Sibb.

HOUK, s. A large ship.

Douglas.
Su. G. holk, navis oneraria.

HOURIS, s. pl.
1. Matins.
Bellenden.
2. Metaph. the chanting of birds.
Dunbar.
Fr. heures, a book of prayers for certain hours.

HOURS. Ten hours, ten o'clock, S.

Acts Ja. I.
Fr. qu'elle heure, S. what hours?
HOUSS, s. A castle.
Wallace.
Su. G. hus, castellum, arx.

HOW, adj. Hollow.

V. Holl.
How, s.
1. Any hollow place, S.
Ross.
2. A plain, S.
Statist. Acc.
3. The hold of a ship.
Douglas.
4. Dung in the howes, overturned.
Baillie.

HOW, s. A tumulus, Orkn.

Stat. Acc.
Isl. haug, Su. G. hoeg, a sepulchral mound.

HOW, s.
1. A coif or hood. S. B. pron. hoo.
Kelly.
Belg. huyve, Dan. hue, id.
2. A chaplet.
Douglas.
A. S. hufe, tiara.
3. Sely how, also happy how, a membrane on the head, with which
some children are born; pron. hoo, S. B.
Ruddiman.
HOW, HOU, HOO, s. A piece of wood, which joins the couple-wings
together at the top, on which rests the roof-tree of a thatched
house, S.
Ramsay.
Su. G. huf, summitas tecti.

HOW, s. A hoe, S.
Fr. houe.
Barbour.

HOW, HOU, s.
1. The sound made by the owl.
Fr. hu-er, to hoot.
Doug.
2. A sea cheer.
Complaynt S.

To HOWD, v. a. To act as a midwife, S.

Isl. iod, childbirth, iod sott, the pangs of childbirth.
Howdy, s. A midwife, S.
Ramsay.
Su. G. iodgumma, id. i. e. as frequently expressed in S. a
houdy-wife.

To HOWDER, v. n. To move by succussation, S.

Ferguson.

To HOWDER, v. a. To hide, Loth.

Ramsay.
Howdrand, part. pa. Hiding.
Dunbar.
S. B. hode, to hide; or Teut. hoeder, receptaculum.

HOWE, interj. A call. S.

Douglas.
Dan. hoo, Fr. ho, id.

HOWIE, CASTLE-HOWIE, s. The name given Orkn. to such of the

Picts' houses as still appear like tumuli.
From How, a tumulus, q. v.

HOWYN, part. pa. Baptised.

Wynt.

HOWLLIS HALD, a ruin; q. an owl's habitation.

Dunbar.

HOWPHYN, s. A term of endearment, equivalent to E. darling.

Evergreen.
C. B. hoffdyn, one who is beloved.

HOW SA, adv. Although.

Barbour.

HOWTOWDY, s. A hen that has never laid, S.

Fr. hustaudeau, hutaudeau, any well-grown pullet.
HUBBILSCHOW, HOBBLESHOW, s. A hubbub, a tumult, S.
Ross.
Teut. hobbel-en, inglomerare; schowe, spectaculum.

To HUCK, v. n. To hesitate as in a bargain, q. to play the huckster,

Z. Boyd.

HUCKIE-BUCKIE, s. A play, in which children slide down a hill on

their hunkers, Loth.
V. Hunker.

HUD, s. The trough employed by masons for carrying mortar, Loth.

To HUD, v. n. To hide.
V. Hod.
Leg. St Androis.

HUDDERIN, HUDERON, part. adj. Flabby in person, and slovenly.

Ang. pron. hutherin.
Kelly.
2. Ugly, hideous, Aberd.
Journ. Lond.
3. Empty, ill-filled, Orkn.
Teut. huyder-en, to have the udder distended.
Huddroun, s. Belly-huddroun, s. A gluttonous sloven.
Dunbar.

HUDDY CRAW, HODDIE, s. The carrion crow, S. B. hoddy craw, S. A.

huddit crau.
Complaynt S.
HUDDS, s. A kind of clay hardened, used for a back to a grate,
Dumfr.
Stat. Acc.

HUDDUM, HUDDONE, s. A kind of whale.

Douglas.

HUDGE-MUDGE, adj. Clandestinely, S. B.

Poems Buch. Dial.
Su. G. miugg, secretly, compounded with hug-a, to
meditate, O. Teut. huggh-en, to observe.

HUD-PYKE, s. A miser.
Dunbar.
Su. G. pick-hogad, qui avide desiderat.

HUFUD, s. A stroke on the head, S. B.

HUGGRIE-MUGGRIE, adv. Hugger-mugger, Fife.

V. Hudge-mudge.

HUICK, s. A small rick of corn, Banffs.

To HUIK, v. a. To consider, to regard.

Chron. S. P.
Teut. huggh-en, observare, considerare.

HUKEBANE, s. Huckle-bone, S. B.
Su. G. Isl. huk-a, inclinare se.
 Dunbar.

To HUKE, Perhaps, to tack;

Teut. huck-en, incurvare.
Maitland P.

HULGIE-BACK, s. Hump back.

Gl. Ross.
Hulgie-backed, adj. Hump-backed, S. B.
Ross.
Su. G. hulkig, convexus, E. hulch.

HULY, HOOLIE, adj. Slow, moderate, S. heelie, Aberd.

Douglas.
Hove, to stay, S., or Su. G. hoflig, moderate.

HULLION, s. A sloven, Fife.

HULLCOCK, s. The Smooth hound, a fish, Orkn.

HULTER CORN, s. The same with shilling, Aberd. q. hulled.

Stat. Acc.

HUM, s. A sham, S.
Su. G. hum, an uncertain rumour.

To HUM, To feed, as birds do their young, by billing, Ang.

HUMANITY, s. The study of the Latin language.
Hence the Humanity Class, that in which this is taught; and the
teacher, the Professor of humanity.
Lat. Literae Humaniores.
Stat. Acc.

HUMDRUM, s. Dejection, S. B.
Ross.
Isl. humm-a, admurmurare, and drom-a, tarde et lente
gradi.

HUMEST, adj. Uppermost.

V. Umast.
Wallace.

HUMLY, adj. Humble.

Bellenden.

HUMLOIK, s. Hemlock.
Lyndsay.

HUMMEL, s. A drone.
Dunbar.
Germ. hummel, fucus.

To HUMMEL, v. a. To hummil bear, to separate the grain of barley

from the beards, S. B.
Hummel-corn, s. Grain which wants a beard, as pease, &c. S. B.
Stat. Acc.
Su. G. haml-a, to mutilate.

HUMMEL, adj. Wanting horns.

V. Homyll.

HUMMIE, s. The game otherwise called shintie, Loth.

HUMSTRUM, s. A pet.
Gl. Shirr.
Hum, as in hum-drum, and strum, q. v.

HUND, s.
1. A dog, S.
Dunbar.
Moes. G. hunds, A. S. hund, canis.
2. An avaricious person, S.
Teut. hond, homo avarus.

HUNE, s. Delay.
V. Hone.
Dunbar.

To HUNE, v. n. To emit a querulous sound, Ang.

Su. G. hwin-a, lugere.

HUNGRY GROUND, ground, by superstition, believed to be so much

under the power of enchantment, that he who passes over it would
infallibly faint, if he did not use something for the support of nature,
West of S.
To HUNKER, v. n. To squat down.
Gl. Shirr.
To Hunker, v. a. The same.
Pop. Ball.
Hunkers, s. pl. To sit on one's hunkers, to sit with the hips hanging
downwards, S.
Isl. huk-a, incurvare se modo cacantis.

HUPES of a mill, s. pl. The circular wooden frame, which surrounds

the millstones, Loth. q. hoops.

To HUR, v. n. To snarl.
Muses Thren.
Lat. hirr-ire, id.

HURBLE, s. A lean or meagre object, S. B.

HURCHAM, adj. Like a hedgehog.

Dunbar.

HURCHEON, s. A hedgehog, S.

HURD, HURDE, s. A hoard, S.

Wyntown.

HURDIES, s. pl. The buttocks, S.

Lyndsay.
Welcome to our website – the perfect destination for book lovers and
knowledge seekers. We believe that every book holds a new world,
offering opportunities for learning, discovery, and personal growth.
That’s why we are dedicated to bringing you a diverse collection of
books, ranging from classic literature and specialized publications to
self-development guides and children's books.

More than just a book-buying platform, we strive to be a bridge

connecting you with timeless cultural and intellectual values. With an
elegant, user-friendly interface and a smart search system, you can
quickly find the books that best suit your interests. Additionally,
our special promotions and home delivery services help you save time
and fully enjoy the joy of reading.

Join us on a journey of knowledge exploration, passion nurturing, and

personal growth every day!

ebookbell.com