John O'Quigley - Survival Analysis - Proportional and Non-Proportional Hazards Regression (Springer The Data Sciences) - Springer (2021)

John O’Quigley
Survival
Analysis
Proportional and Non-Proportional
Hazards Regression
Survival Analysis
John O’Quigley
Survival Analysis
Proportional and Non-Proportional Hazards
Regression
123
John O’Quigley
Department of Statistical Science
University College London
London WC1E 6BT, UK
ISBN 978-3-030-33438-3 ISBN 978-3-030-33439-0 (eBook)

https://doi.org/10.1007/978-3-030-33439-0
Mathematics Subject Classification: 62-07, 62B10, 62F03, 62F05, 62F07, 62F40, 62G30, 62G32,
62G10, 62H17, 62J02, 62J05, 62L05, 62L10, 62L12, 62N01, 62N02, 62N05, 62N86, 62P05, 62P10,
91B02, 92B15, 92C60, 92C50, 92D30, 97K80
© Springer Nature Switzerland AG 2021

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
In loving memory of my father, Arthur
O’Quigley, whose guiding strategy when
confronting the most challenging of problems
– use a sense of proportion – would have
baffled many a latter-day epidemiologist.
Preface
In common with most scientific texts this one is the result of several iterations. The
seed for the iterations was the 2008 text Proportional Hazards Regression and the
first step was in the direction of a second edition to that work. The next several
iterations from there took us in a number of different directions before slowly
converging to a text that appears to be more in harmony with modern approaches to
data analysis, graphical techniques in particular. These are given greater emphasis
in this book and, following recent work on the regression effect process, it seems
clear that this process, first presented in O’Quigley (2003) ought be given a more
prominent place. Just about everything we need, from complex estimation to simple
tests, can be readily obtained from the regression effect process. Formal analysis
can be backed up by intuitive graphical evaluation based directly on this process.
Proportional hazards models and their extensions (models with time-dependent
covariates, models with time-dependent regression coefficients, stratified models,
models with random coefficients, and any mixture of these) can be used to char-
acterize just about any applied problem to which the techniques of survival analysis
are appropriate. This simple observation enables us to find an elegant statistical
expression for all plausible practical situations arising in the analysis of survival
data. We have a single unifying framework. In consequence, a solid understanding
of the framework itself offers the investigator the ability to tackle the thorniest of
questions which may arise when dealing with survival data.
Our goal is not to present or review the very substantial amount of research that
has been carried out on proportional hazards and related models. Rather, the goal is
to consider the many questions which are of interest in a regression analysis of
survival data, questions relating to prediction, goodness of fit, model construction,
inference and interpretation in the presence of misspecified models. To address
these questions the standpoint taken is that of the proportional hazards and the
non-proportional hazards models.
This standpoint is essentially theoretical in that the aim is to put all of the
inferential questions on a firm conceptual footing. However, unlike the commonly
preferred approach among academic statisticians, based almost entirely on the
construction of stochastic integrals and the associated conditions for the valid
vii
viii Preface
application of the martingale central limit theorem for multivariate counting pro-
cesses, we work with more classical and better known central limit theorems. In
particular we appeal to theorems dealing with sums of independent but not nec-
essarily identically distributed univariate random variables and of special interest is
the classical functional central limit theorem establishing the Brownian motion limit
for standardized univariate sums.
So, we avoid making use of Rebolledo’s central limit theorem for multivariate
counting processes, including the several related works required in order to validate
this theorem, the Lenglart-Rebolledo inequality for example. Delicate
measure-theoretic arguments, such as uniform integrability, borrowed from math-
ematical analysis can then be wholly avoided. In our view, these concepts have
strayed from their natural habitat—modern probability theory—to find themselves
in a less familiar environment—that of applied probability and applied statistics—
where they are not well understood nor fully appreciated. While it could be
advanced that this abstract theory affords greater generality, the author has yet to
see an applied problem in which the arguably less general but more standard and
well known central limit theorems for univariate sums fail to provide an equally
adequate solution. None of the results presented here lean on Rebolledo’s central
limit theorem.
A central goal of this text is to shift the emphasis away from such abstract theory
and to bring back the focus on those areas where, traditionally, we have done well
—careful parsimonious modelling of medical, biological, physical and social
phenomena. Our aim is to put the main spotlight on robust model building using
analytical and graphical techniques. Powerful tests, including uniformly most
powerful tests, can be obtained, or at least approximated, under given conditions.
We discuss the theory at length but always looked at through the magnifying glass
of real applied problems.
I would like to express my gratitude to several colleagues for their input over
many years. The collaborative work and countless discussions I have had with
Philippe Flandre, Alexia Iasonos, Janez Stare and Ronghui Xu in particular have had
a major impact on this text. I have worked at several institutions, all of them having
provided a steady support to my work. Specifically, I would like to acknowledge: the
Department of Mathematics of the University of California at San Diego, U.S.A., the
Fred Hutchinson Cancer Research Center, Seattle, U.S.A., the Department of
Mathematics at the University of Leeds, U.K., the Department of Mathematics at
Lancaster University, U.K., the Department of Biostatistics at the University of
Washington, U.S.A., the Department of Biostatistics at Harvard University, Boston,
U.S.A., the Division of Biostatistics at the University of Virginia School of
Medicine, the Laboratory for Probability, Statistics and Modelling, University of
Paris, Sorbonne and last (but not least), the Department of Statistical Science,
University College London. I am very grateful to them all for this support.
A final word. At the heart of our endeavor is the goal of prediction, the idea of
looking forward and making statements about the future. Perhaps somewhat
paradoxically, this only makes real sense when looking backwards, that is when all
the observations are in, allowing us to make some general summary statements.
Preface ix
Public health policy might use such statements to make planning decisions, clinical
trials specialists may use them to increase the power of a randomized study. But, to
use them to make individual predictions is unlikely to be of any help and could
potentially be of harm. Even for the terminally ill patient it is not rare for those with
the very worst prognosis to confound the predictions while others, looking much
better in theory, can fare less well than anticipated. Such imprecision can only be
greatly magnified when dealing with otherwise healthy subjects, the example of the
BRCA1 and BRCA2, so-called, susceptibility genes being a striking one. We
consider this question in the chapter dealing with epidemiology. While past history
is indispensable to improving our understanding of given phenomena and how they
can describe group behaviour, on an individual level, it can never be taken to be a
reliable predictor of what lies ahead. Assigning probabilities to individuals, or to
single events—however accurate the model—is not helpful.
When anyone asks me how I can best describe my experience in nearly 40 years at
sea, I merely say ... uneventful. Captain Edward John Smith, RMS Titanic April,
1912.
London, UK John O’Quigley

April 2021
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Context and motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Some examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Main objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Neglected and underdeveloped topics . . . . . . . . . . . . . . . . . . . 7
1.6 Model-based prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.7 Data sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.8 Use as a graduate text . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.9 Classwork and homework . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Survival analysis methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Basic tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Some potential models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.5 Censoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.6 Competing risks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3 Survival without covariates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.3 Parametric models for survival functions . . . . . . . . . . . . . . . . 50
3.4 Empirical estimate (no censoring) . . . . . . . . . . . . . . . . . . . . . 56
3.5 Kaplan-Meier (empirical estimate with censoring) . . . . . . . . . . 58
3.6 Nelson-Aalen estimate of survival . . . . . . . . . . . . . . . . . . . . . 68
3.7 Model verification using empirical estimate . . . . . . . . . . . . . . 69
3.9 Outline of proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
xi
xii Contents
4 Proportional hazards models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

4.1 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.3 General or non-proportional hazards model . . . . . . . . . . . . . . . 77
4.4 Proportional hazards model . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.5 Cox regression model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.6 Modeling multivariate problems . . . . . . . . . . . . . . . . . . . . . . . 87
5 Proportional hazards models in epidemiology . . . . . . . . . . . . . . . . . 97
5.1 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.3 Odds ratio, relative risk, and 2 2 tables . . . . . . . . . . . . . . . . 98
5.4 Logistic regression and proportional hazards . . . . . . . . . . . . . . 102
5.5 Survival in specific groups . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.6 Genetic epidemiology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
6 Non-proportional hazards models . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.1 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.3 Partially proportional hazards models . . . . . . . . . . . . . . . . . . . 120
6.4 Partitioning of the time axis . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.5 Time-dependent covariates . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.6 Linear and alternative model formulations . . . . . . . . . . . . . . . 136
7 Model-based estimating equations . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.1 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.3 Likelihood solution for parametric models . . . . . . . . . . . . . . . 143
7.4 Semi-parametric estimating equations . . . . . . . . . . . . . . . . . . . 151
7.5 Estimating equations using moments . . . . . . . . . . . . . . . . . . . 153
7.6 Incorrectly specified models . . . . . . . . . . . . . . . . . . . . . . . . . . 164
7.7 Estimating equations in small samples . . . . . . . . . . . . . . . . . . 176
8 Survival given covariate information . . . . . . . . . . . . . . . . . . . . . . . 191
8.1 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
8.3 Probability that Ti is greater than Tj . . . . . . . . . . . . . . . . . . . . 193
8.4 Conditional survival given Z 2 H . . . . . . . . . . . . . . . . . . . . . 195
Contents xiii
8.5 Other relative risk forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201

8.6 Informative censoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
9 Regression effect process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
9.1 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
9.3 Elements of the regression effect process . . . . . . . . . . . . . . . . 217
9.4 Univariate regression effect process . . . . . . . . . . . . . . . . . . . . 224
9.5 Regression effect processes for several covariates . . . . . . . . . . 231
9.6 Iterated logarithm for effective sample size . . . . . . . . . . . . . . . 235
10 Model construction guided by regression effect process . . . . . . . . . 261
10.1 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
10.3 Classical graphical methods . . . . . . . . . . . . . . . . . . . . . . . . . . 263
10.4 Confidence bands for regression effect process . . . . . . . . . . . . 267
10.5 Structured tests for time dependency . . . . . . . . . . . . . . . . . . . 269
10.6 Predictive ability of a regression model . . . . . . . . . . . . . . . . . 270
10.7 The R2 estimate of X2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
10.8 Using R2 and fit to build models . . . . . . . . . . . . . . . . . . . . . . 278
10.9 Some simulated situations . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
10.10 Illustrations from clinical studies . . . . . . . . . . . . . . . . . . . . . . 291
11 Hypothesis tests based on regression effect process . . . . . . . . . . . . . 301
11.1 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
11.3 Some commonly employed tests . . . . . . . . . . . . . . . . . . . . . . 302
11.4 Tests based on the regression effect process . . . . . . . . . . . . . . 308
11.5 Choosing the best test statistic . . . . . . . . . . . . . . . . . . . . . . . . 327
11.6 Relative efficiency of competing tests . . . . . . . . . . . . . . . . . . . 334
11.7 Supremum tests over cutpoints . . . . . . . . . . . . . . . . . . . . . . . . 335
11.8 Some simulated comparisons . . . . . . . . . . . . . . . . . . . . . . . . . 337
11.9 Illustrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
11.10 Some further thoughts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
xiv Contents
Appenidx A. Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351

A.1 Essential tools for survival problems . . . . . . . . . . . . . . . . . . . . . . 351
A.2 Integration and measure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
A.3 Random variables and probability measure . . . . . . . . . . . . . . . . . 354
A.4 Convergence for random variables . . . . . . . . . . . . . . . . . . . . . . . 355
A.5 Topology and distance measures . . . . . . . . . . . . . . . . . . . . . . . . . 356
A.6 Distributions and densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
A.7 Multivariate and copula models . . . . . . . . . . . . . . . . . . . . . . . . . . 362
A.8 Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
A.9 Order statistics and their expectations . . . . . . . . . . . . . . . . . . . . . 366
A.10 Approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
Appendix B. Stochastic processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
B.1 Broad overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
B.2 Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
B.3 Counting processes and martingales . . . . . . . . . . . . . . . . . . . . . . 385
B.4 Inference for martingales and stochastic integrals . . . . . . . . . . . . 392
Appendix C. Limit theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
C.1 Empirical processes and central limit theorems . . . . . . . . . . . . . . 401
C.2 Limit theorems for sums of random variables . . . . . . . . . . . . . . . 402
C.3 Functional central limit theorem . . . . . . . . . . . . . . . . . . . . . . . . . 405
C.4 Brownian motion as limit process . . . . . . . . . . . . . . . . . . . . . . . . 408
C.5 Empirical distribution function. . . . . . . . . . . . . . . . . . . . . . . . . . . 410
Appendix D. Inferential tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
D.1 Theory of estimating equations . . . . . . . . . . . . . . . . . . . . . . . . . . 413
D.2 Efficiency in estimation and in tests . . . . . . . . . . . . . . . . . . . . . . 420
D.3 Inference using resampling techniques . . . . . . . . . . . . . . . . . . . . . 422
D.4 Conditional, marginal, and partial likelihood . . . . . . . . . . . . . . . . 425
Appendix E. Simulating data under the non-proportional hazards
model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433
E.1 Method 1—Change-point models . . . . . . . . . . . . . . . . . . . . . . . . 433
E.2 Method 2—Non-proportional hazards models . . . . . . . . . . . . . . . 437
Further exercises and proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
Summary of main notation
W, WðtÞ Brownian motion on the interval ½0; 1

B, BðtÞ, W 0 Brownian bridge on the interval ½0; 1
IðAÞ, 1A Indicator function for the set A
PðAÞ, PrðAÞ, PðAÞ Probability measure for the set A
L Convergence in distribution (law)
!
P Convergence in probability
!
L1 Convergence in mean
!
a:s: Almost surely, w:p:1
Mpp ðRÞ Matrices of dimension p p with real coefficients
kak Supremum norm of the vector a 2 Rp
k Ak Supremum norm of the matrix A 2 Mpp ðRÞ
b
F Kaplan-Meier estimator of F
n Sample size
p Dimension of covariable space
Ti Time to event for individual i
Xi Observed time under study for individual i
di Censoring indicator variable for individual i
Zi Vector of p covariables for individual i
Zi
ðlÞ l lème covariable of individual i, l ¼ 1; . . .; p
kn Number of effective (usable) failures
b Regresson coefficient, log relative-risk
bðtÞ Time-dependent regression coefficient
ZðtÞ Value of covariate for individual failing at time t
pi ðbðtÞ; tÞ Given t, probability that individual i is selected for failure
under NPH model with parameter bðtÞ
xv
xvi Summary of main notation
E bðtÞ ðZjtÞ Expectation with respect to pi ðbðtÞ; tÞ

NPH model with coefficient bðtÞ
V bðtÞ ðZjtÞ Variance with respect to pi ðbðtÞ; tÞ
NPH model with coefficient bðtÞ
U n ðbðtÞ; tÞ Regression effect process evaluated at time t
Chapter 1
Introduction
1.1 Chapter summary
The chapters of this book correspond to a collection of broad themes. Taken

together, these themes make up a sizeable component of what can be consid-
ered to be the topic of modern survival analysis. A feature that we maintain
throughout the text is to summarize the salient features of each chapter in a
brief section headed “Chapter summary”. Section 1.1 is the first such summary.
Following these summaries comes a section headed “Context and motivation”,
the purpose of which is to motivate our interest as well as to recall any key
notions we may need as an aid in pursuing the topic of the chapter. In this
introduction we consider a number of real-life examples in which the methods
of survival analysis, proportional and non-proportional hazards models, in par-
ticular, have been used to advantage. Following this we highlight in a section
entitled “Objectives” the way in which the described methodology can address
many of the important questions that arise in the very diverse set of motivating
examples. A large number of examples are worked throughout this book and we
indicate where to find the relevant data for those readers interested in confirming
the results for themselves before working on their own particular projects. For
teaching purposes a section suggesting problems for classwork and homework is
given. Formal proofs and proof outlines are removed from the main text into
their own section at the end of each chapter.
1.2 Context and motivation
Applications of survival analysis methodology appear to be ever-widening.

Recent areas of interest include unemployment data, lengths of time on and
off welfare, for example, and some unusual applications such as interviewer bias
c Springer Nature Switzerland AG 2021

1
J. O’Quigley, Survival Analysis,
https://doi.org/10.1007/978-3-030-33439-0_1
2 Chapter 1. Introduction
in sociological surveys. The more classic motivating examples behind the bulk
of theoretical advances made in the area have come from reliability problems in
engineering and, especially, clinical studies in chronic diseases such as cancer and
AIDS. Many of the examples given in this text come from clinical research. Par-
allels for these examples from other disciplines are usually readily transposable.
Proportional hazards regression is now widely appreciated as a very powerful
technique for the statistical analysis of broad classes of problems. The ability to
address issues relating to complex time-dependent effects and non-proportional
hazards structures broadens this class yet further.
1.3 Some examples
Novel uses as well as innovatory applications of the basic approach continue

to grow. The following examples represent a small selection of problems for
which the proportional hazards model has been shown to be useful as an analytic
tool. Adding non-proportional hazards models to our toolkit provides us with an
array of statistical techniques capable of tackling almost any problem in survival
analysis.
Clinical trials in chronic disease

Patients, identified as having some chronic illness under study, are recruited to
a clinical trial. A new, proposed, treatment, if effective, should prolong survival.
The idea is that instead of aiming for a “cure” we aim to improve survival.
In cancer research it has been common practice to equate cure with five years
survival without relapse. Such a quantity may be difficult to estimate if some
patients leave the study or die from unrelated causes during the five-year period.
Survival methods address such difficulties. Proportional and non-proportional
hazards models enable us to investigate the dependence of survival on controlled
and uncontrolled sources of variation. Now, suppose that therapy A and ther-
apy B have the same five-year survival rates but different survival curves. As
an illustration consider curative treatments based on surgery having very high
initial risks but, conditional upon having survived some length of time, a reduced
long-term risk. Such differences in overall survival are important and would be
considered relevant when weighing the relative advantages and disadvantages of
any proposed new treatment. In actual studies patients may be lost to follow-up
at any time. Proportional hazards and non-proportional hazards modeling can
allow us to correctly use the partial information provided by such patients. These
models provide the means to isolate as much of the treatment effect as we can
from the several interrelated confounding effects.
1.3. Some examples 3
Applications in epidemiology studies

The time-to-event variable in these studies is almost always age. A large number
of individuals are studied across a relatively wide distribution of ages. Incidence
of some target disease is measured. This incidence will typically vary with age
and, in the majority of cases, the single most important source of variability
in the observed incidence rates comes from age itself. Cancer is a particularly
striking example. Here the variable age plays the role of the time variable in
a survival model. Prolonging survival amounts to living to a longer age. How
does incidence relate to risk variables, obvious examples being smoking, asbestos
exposure or other industrial exposures. In practice the amount of time needed for
such prospective studies may be too great and we may prefer to use a case-control
study where, rather than compare the age distributions for different exposure
categories, we compare the exposure distributions at different ages.
In doing this our purpose is to address some key feature of the conditional
distribution of age given covariate information via a study of the conditional
distribution of the covariates given age. The question arises as to the validity
of this and to the efficiency of such a study design. Additional complications
arise when the exposure categories change for some individuals. At any given
age a particular individual can have a unique and complex exposure history. It
would not be possible to study the relationship between any unique exposure
history and disease outcome without assuming some structure, i.e., the potential
ways in which the different components comprising the exposure history relate
to one another and, more importantly, to outcome. The way to achieve this is
via a statistical model, specifically, in the context of this work, the proportional
hazards and non-proportional hazards regression model.
Genetic susceptibility to disease incidence

Recent years have seen a growing interest in identifying genetic mutations that
may be linked to the occurrence of some particular disease. The use of registry
data, such as the SEER data, in conjunction with a survival model allows us to
quantify the risk associated with different classifications, in particular a classi-
fication based on genetic profiles and so-called susceptibility mutations. Family
studies can be of great value in bringing to light potential risk factors although,
unless the risk factors are particularly strong, it will be difficult to detect them
without long follow-up times and careful modeling strategies. These can enable
some degree of control over factors that may mask effects. Risk over several
years including lifetime risk can be calculated using survival methods. That said,
great care is needed in interpretation and we consider this in the chapter on
epidemiology.
Complex time-dependent effects in clinical studies

In clinical research, not only can there be problems with patients going off study
or being removed from the study due to causes believed to be unrelated to the
main outcome under study, it is also not unusual for subjects to enter the study
at different times. In chronic diseases such as AIDS, patients may leave the study
because they are too sick to continue participation. Such loss to follow-up maybe
of a different nature to that described above and may, of itself, provide informa-
tion on the risk of being incident for the outcome. In technical terms we would
describe the situation as being one where the censoring mechanism is not inde-
pendent of the failure mechanism. In order to avoid potentially misleading biases
it would be necessary to appeal to workable models to describe this dependence.
The complexity of potential applications can easily grow beyond the reach
of available methods. These continue to promote and drive the interest for fur-
ther methodological developments. One example, again from AIDS, where the
methods presented here can provide insight concerns the situation where the
time frame is lengthy. Patients may have very involved time-dependent exposure
histories to the different treatments, the class of available treatments themselves
having evolved through time as well as within any given patient. It is quite possi-
ble that no two patients would have exactly the same treatment history. Models
enable us to put some kind of structure on this so that inferences can be made
regarding treatment approaches.
In prognostic studies, many variables are measured on individuals and we
would like to know how individually and collectively these variables influence sur-
vival (prognosis). It may be possible to use such information to stratify future
clinical trials, the goal being to increase precision, or, possibly, to use the informa-
tion in a context of clinical patient management. Further complexity arises with
correlated data. For recurrent conditions such as asthma the same patient may be
studied for more than one principal event. Cooperative clinical trial groups may
use many centers to recruit patients, the center affects themselves not being
of any intrinsic interest, but nonetheless contributing a correlation factor for
within-center observations. Similar considerations arise in familial data in which
the presence of genetic characteristics may influence the survival probabilities col-
lectively. Proportional hazards models including random component coefficients,
equivalently non-proportional hazards models, can be used to efficiently analyze
such data structures.
1.3. Some examples 5
Insurance, economics, and sociology

An area of application where the law of large numbers comes directly into play is
that of insurance. The calculus of probability together with very large object pop-
ulations enable effective prediction. We may wish to know how demographic and
other variables influence insurance risk. Further examples arise in car insurance;
how long on average does it take to be involved in a traffic accident and, more
importantly, which variables influence this time. Such information will enable the
insurer to correctly price policies for different driver profiles.
In economics there are many possibilities. How long does it take for a price
index to increase to a certain value, to regain a previous value or to change
by some given amount? In this and other examples the technical term “failure”
may in a more common everyday sense correspond to success. Depending on the
context we may wish to prolong or reduce the average time it takes to observe
some kind of event. Failure models have been used to study, under differing
circumstances, the average time taken for a welfare recipient to come off welfare,
an event in social terms that we would regard as a success.
Financial analysis
Mathematical techniques of stochastic integration have seen wide application in
recent years in the study of financial products such as derivatives, futures, and
other pricing schemes for certain kinds of options. An alternative and potentially
more flexible approach to many of these questions is via regression modeling.
The increasing volatility of many of the markets has also given rise to the use
of survival modeling as a tool to identify, among the large-scale borrowers, large
industries, countries, and even financial institutions themselves the ones which
are most likely to fail on their repayment schedule. Modeling techniques can
make such analyses more precise, identifying just which factors are most strongly
predictive.
Manufacturing and reliability

One of the early applications of survival methods concerns the reliability of com-
ponents as well as mechanical, electrical, or electronic systems. Careful modeling
can enable us to use various sources of information to predict the probability of
failure before some given time. Many consumers will have first-hand experience of
such modeling whereby it is possible to keep small the number of manufactured
products failing before the expiration of the guarantee while, at the same time,
ensuring that the probability of survival well beyond the guarantee is not so large
as to ultimately impact consumer demand. It can be of interest to model more
than one endpoint, tracing out the lifetime of a product. The first event may be
something minor and subsequent events of increasing degrees of seriousness, or
involving different components, until the product is finally deemed of no further
use. The many different paths through these states which any product may follow
in the course of a lifetime can be very complex. Models can usefully shed light
on this.
1.4 Main objectives
The purpose of this book is to provide the structure necessary to the building
of a coherent framework in which to view proportional and non-proportional
hazards regression. The essential idea is that of prediction, a feature common to
all regression models, but one that sometimes slips from sight amid the wealth
of methodological innovations that has characterized research in this area. Our
motivation should mostly derive from the need to obtain insights into the complex
data sets that can arise in the survival setting, keeping in mind the key notion
that the outcome time, however measured, and its dependence on other factors
is at the center of our concerns.
The predictive power of a model, proportional and non-proportional hazards
models, in particular, is an area that we pay a lot of attention to. It is neces-
sary to investigate how we can obtain predictions in the absence of information
(often referred to as explanatory variables or covariate information) which relate
to survival and, subsequently, obtaining predictions in the presence of such infor-
mation. How best such information can be summarized brings us into the whole
area of model adequacy and model performance. Measures of explained varia-
tion and explained randomness can indicate to what extent our prediction accu-
racy improves when we include additional covariate information into any model.
They also play a central role in helping identify the form of the unknown time-
dependent regression effects model that is taken to generate the observations.
In order to give the reader, new to the area, a feeling as to how the Cox
model fits into the general statistical literature we provide some discussion of
the original paper of Professor Cox in 1972, some of the background leading up
to that seminal paper, and some of the scientific discussions that ensued. The
early successes of the model in characterizing and generalizing several classes of
statistics are described.
As is true of much of science—statistics is no exception—important ideas
can be understood on many levels. Some researchers, with a limited training in
even the most basic statistical methods, can still appreciate the guiding principles
behind proportional hazards regression. Others are mainly interested in some of
the deeper inferential mathematical questions raised by the estimation techniques
1.5. Neglected and underdeveloped topics 7
employed. Hopefully both kinds of readers will find something in this book to their
taste. The aim of the book is to achieve a good balance, a necessary compromise,
between the theoretical and the applied. This will necessarily be too theoretical for
some potential readers, not theoretical enough for others. Hopefully the average
gap is not too great.
The essential techniques from probability and statistics that are needed
throughout the text are gathered together in several appendices. These can be
used for reference. Any of these key ideas called upon throughout the book can
be found more fully developed in these appendices. The text can therefore serve
as a graduate text for students with a relatively limited background in either
probability or statistics. Advanced measure theory is not really necessary either
in terms of understanding the proportional hazards model or for gaining insight
into applied problems. It is not emphasized in this book and this is something
of a departure from a number of other available texts which deal with these and
related topics. Proofs of results are clearly of importance, partly to be reassured
as to the validity of the techniques we apply, but also in their own right and of
interest to those focusing mostly on the methods. In order not to interrupt the
text with proofs, we give theorems, corollaries, and lemmas, but leave the proofs
to be gathered together at the end of each chapter. The reader, less interested in
the more formal presentation in terms of theorems and proofs, will nonetheless,
it is hoped, find the style helpful in that, by omitting the proofs at the time of
development, the necessary results are organized and brought out in a sharper
focus. While our motivation always comes from real practical problems, the pre-
sentation aims to dig deep enough into the mathematics so that full confidence
can be obtained in using the results as well as using the machinery behind the
results to obtain new ones if needed.
1.5 Neglected and underdeveloped topics
The non-proportional hazards model, including as a special case the proportional

hazards model, is so general in nature that essentially all relative risk models
come under its umbrella. There are one or two exceptions, such as continuous
covariates, but they can be approximated to as high a degree as we wish. This
means that we have a full spectrum of all relative risk models where one extremity
is an absence of any restrictive assumption while the other is that of proportional
hazards. The true, generally unknown, model lies somewhere between the two.
If our interest goes beyond that of just relative risk to, say, the underlying haz-
ard functions, or intensity functions, then further restrictions can be made, an
example being the use of parametric models such as the Weibull model. These
are described in this text but little emphasis is made and the greatest part of our
focus is on relative risk functions and their evolution through time.
It may be that certain parametrizations lead to simplification, one example

arising in cardiovascular studies where an additive survival model allows a more
parsimonious overall description than would a mathematically equivalent, non-
proportional hazards model. Nevertheless, if only conceptually, it helps to keep
in mind that models such as the additive and related models, at first glance
appearing to lie outside of our general formulation, can indeed be expressed as
non-proportional hazards one with a regression function that is time dependent.
This thinking guides our presentation. As a result certain topics that are often
given an important place in survival analysis texts are neglected here. Some of
these are briefly described below.
Left censoring and left truncation

Just as right censoring at time C means that the true, unobserved, time of interest
is greater than C, left censoring at time C means that the true, unobserved, time
of interest is lesser than C. Examples of left censoring in clinical trials are hard to
find but, in epidemiology, it is conceivable that a study would include individuals
for whom the event of interest has already occurred without knowing precisely
when the event occurred; simply that it would be lesser than some known time
point, often their time of entry into the study. In this text we do not study left
censoring and the reader interested in analyzing problems in which left censoring
arises might begin by looking at some specialist epidemiological texts that deal
with longitudinal observations.
Note that observations that are only recorded to occur within some time
interval may, in a formal way, be analyzed by combining methods that can simul-
taneously accommodate both left and right censoring. Again, we do not treat
such problems in this book. Another type of observation that can arise may result
from a phenomenon known as left truncation. If, having encountered the event
of interest, before the study begins, the probability of being included in the study
is small, possibly zero, then this is known as left truncation. It is easy to see
how this phenomenon could lead to substantial bias if ignored. An illustration
of left truncation can be seen in studies that focus on factors that might influ-
ence the rate of spontaneous abortion following conception. This rate is likely to
be underestimated since the smallest times may often correspond to situations
where the event is not recorded, the woman may not even be aware that she was
pregnant.
Studying the impact of genetic and biomarkers on survival can run into the
problem of left truncation. The decision to focus interest on some biomarker may
not have been made at the beginning of the study so that, for those subjects
who have left the study due to failure or other forms of censoring, it is no longer
possible to ascertain to which of the different groups they may belong. As a
result they are often ignored in analyses resulting in potentially strong biases.
In this text we provide survival estimates that would overcome such biases but,
otherwise, we do not study left truncation in its own right. Readers looking for
more on this topic are referred to Hyde (1977), Tsai et al. (1987), Keiding et al.
(1987), Keiding and Gill (1990), Jiang et al. (2005), Shen (2006), Cain et al.
(2011), and Geskus (2011).
Learning and classification

The ability, and the need, to deal with very large sets of data has been the
motor for many developments of the past two decades in a discipline that is
neither statistics nor computer science—we could describe it as a hybrid or an
intersection of the two. Under this heading we might include machine learning,
supervised and unsupervised, deep learning, classification and regression trees,
random forests, and artificial intelligence. In the field of survival analysis the
work of Breiman (2017) anticipated many of these developments. The bulk of
the problems of machine learning, in the context of survival analysis, are directly
addressed in this text albeit using terms that are more familiar to those from
the less recent schools of statistics. While prediction, and the assessment of
predictive strength, is a running theme in this work, we do not make an appeal
to the techniques of classification. The literature has many examples of model
construction based on classification techniques, for example, Ciampi et al. (1988),
Ciampi et al. (1991), LeBlanc and Crowley (1992), Schumacher et al. (2003),
Koziol et al. (2003), Loh (2011), and Loh (2014) and there are very many more.
Linden and Yarnold (2017) provides a specific focus on clinical practice. This is
a large topic and will be the focus of a future text but is not considered here in
any great depth.
Kernel methods and local smoothing

Local smoothing of the baseline hazard, or of the residuals, although an interest-
ing topic, is not given any real space here. The same goes for kernel estimates of
the density function or the use of polynomial splines. The view taken here is that,
for the great majority of applications, cumulative quantities such as the empiri-
cal distribution function do “enough” smoothing. However, such work is valuable
and certainly of interest. An absence of coverage in this text should not be taken
as implying that such techniques have no importance. The choice of how much
weight to afford different topics, including the assignment of a zero weight, is
often a subjective one—a matter of taste. We have little familiarity with applied
problems that would benefit from the use of these smoothing techniques. One
way to view smoothing is to see it as a way to inject some structure into a model
with too many moving parts, as can occur when there are a large number of
potential parameters to estimate. The goal then would then be to bring some
additional constraints to the model, in some sense tightening the overall struc-
ture. In this work, our approach to model building takes something of a contrary
standpoint. We begin mostly with a quite rigid framework for the model and,
using the techniques of fit and predictive performance, gradually weaken certain

restrictions.
There is more than one way to build more complex models from initially sim-
pler ones. In the main we tend to do this by weakening the constraints around
fixed parameters, for example, stepping up from a proportional hazards model
to a piecewise proportional hazards model. A different approach focuses on the
undetermined randomness and by broadening the probabilistic base from which
observations can be made. Frailty models come under this heading so that unob-
served quantities enter our inferential procedures indirectly. These can allow some
simple parameters, often no more than the variance of the frailty distribution, to
be replaced by data-driven estimates, thereby taking up some of the slack in order
to provide a potentially better fit for the model. We give relatively little weight to
the study of frailty models, outside the context of goodness of fit. Random effects
models, although of the same flavor, do not come under this heading and, while
these are described in the text, for frailty models, if we are to understand the
term “frailty” as meaning a random effect for an individual (without repeated
measurements), as opposed to a group (random effects models), then this is
just another way of expressing lack of fit. In other words a proportional hazards
model with a frailty term is equivalent to a non-proportional hazards model.
Tests, described in the literature, for heterogeneity of individuals in this context
can be misleading. This has been described in O’Quigley and Stare (2002).
Partial likelihood
The term “partial likelihood” is common in survival texts as a description of a
particular approach to inference. It is not a very useful concept in our view, and
perhaps it is time to put the term “partial likelihood” to a long deserved rest.
In his introductory paper on proportional hazards models, Professor Sir David
Cox (Cox, 1972) made no reference at all to partial likelihood. He presented a
suitable likelihood with which to make inference, given the observations, and, no
doubt unintentionally, generated some confusion by referring to that likelihood
as a conditional likelihood. Some of the discussants of Cox’s paper picked up
on that and were puzzled as to the sense of a conditional likelihood since the
statistic upon which we would be conditioning is not very transparent. These
discussants pointed out that there is no obvious statistic whose observed value
is taken as fixed before we set about making inferences. Nonetheless, there is
a lot of sequential conditioning that leads to the expression, first derived by
Professor Cox, and, for want of a better term, “conditional likelihood” was not
so bad. Cox’s choice for likelihood was a fully legitimate one. It was also a very
good one, having among several properties that of leaving inference unaffected
by increasing transformations on the time variable. More explanation may have
been needed but, instead, regrettably, we all moved off in something of a wrong
direction: not dramatically wrong but wrong enough to cloud and confuse issues
that ought, otherwise, have been quite plain.
These were very early days for proportional hazards regression and, for those
concerned, it was not always easy to see clearly ahead. Efforts were made either
to justify the likelihood given by Cox, as arising within a particular conditional
sampling framework or to appeal to different techniques from probability that,
by good fortune, led us to the same estimating equations like the ones derived
by Cox. All of this had two unfortunate consequences that turned out to be very
significant for the later development of the field. Those initial worries about the
nature, and indeed the legitimacy, of the likelihood, took us on something of
an arduous path over the next few years. In the late seventies, early eighties,
a solution to the problem was believed to have been found. And in the place
where we might least expect to find it: within the very abstract French school
of probability. It was argued that the key to the solution lay in the central limit
theorem of Rebolledo (1978) for multivariate counting processes.
Now, while stochastic processes, and counting processes, in particular, are at
the very heart of survival analysis, this particular branch of the French school
of probability, built around Lenglart’s theorem, Rebolledo’s multivariate central
limit theorem for stochastic integrals, Jacod’s formula, Brémaud’s conditions,
bracket processes, and locally square integrable martingales, is, and has remained,
almost inaccessible to those lacking a strong background in real analysis. This
development had the unfortunate effect of pushing the topic of survival analysis
beyond the reach of most biostatisticians. Not only that but the field became for
very many years too focused on the validity of Rebolledo’s findings under different
conditions. The beauty of the Cox model and the enormous range of possibilities
it opened up took second place so that questions at the heart of model building:
predictive strength, model validation, goodness of fit, and related issues did not
attract the attention that they deserved.
Our goal is not to use the benefit of hindsight to try to steer the course of
past history—we couldn’t anyway—but we do give much less emphasis to the
multivariate central limit theorem of Rebolledo in favor of more classical central
limit theorems, in particular the functional central limit theorem. This appears
to allow a very simple approach to inference. No more than elementary calculus
is required to understand the underpinnings of this approach.
The concept of partial likelihood as a general technique of inference in its own
right, and for problems other than inference for the proportional hazards model
has never really been thoroughly developed. One difficulty with the concept of
partial likelihood, as currently defined, is that, for given problems, it would not
be unique. For general situations then it may not be clear as to the best way
to proceed, for instance, which of many potential partial likelihoods ought we
choose to work with. For these reasons we do not study partial likelihood as a tool
for inference and the concept is not given particular weight. This is a departure
from several other available texts on survival analysis. In this work, we do not
view partial likelihood as anything other than a regular likelihood consequent

upon a particular sampling structure. As such, it inherits all of those desirable
properties that we know to hold.
Joint survival-covariate models

An earlier version of this book in development included a chapter on joint survival-
covariate models. It was decided to remove this chapter for the following reasons.
Firstly, the contributions of Henderson et al. (2000), Tsiatis and Davidian (2004),
as well as the text by Rizopoulos (2012), cover the field very thoroughly and,
having not spent enough time in this area, I have nothing really to add in the
way of commentary or insight. Secondly, in order to be coherent with the presen-
tation here that hinges upon the regression effect process in transformed time,
we would need to build a parallel setup for joint models. While this would not be
conceptually difficult it remains to be carried out. Finally, the motivation to do
this work is not strong since real applications of joint models are quite limited.
There is an extensive literature on this topic and the interested reader is referred
to; Hsieh et al. (2006), R. Brown and G. Ibrahim (2003), Huang and Liu (2007),
Song and Wang (2008), Yu and Sandler (2008), Crowther et al. (2013), Chen
et al. (2014) and Rizopoulos and Lesaffre (2017).
An important misconception that requires clarification is that joint modeling
is necessary whenever the covariate—a common example being immunological
status in HIV studies—changes through time. Our main goal is to study the
dependence of survival on such measures and, for this, we need to study con-
ditional distributions; survival time given the covariate or the covariate given
survival time. Proportional hazards and non-proportional hazards models can
handle time-dependent covariates very well. Modeling, and learning about, the
covariate process itself through time, provides us with very little additional infor-
mation in the best of cases and, in most cases, it does not provide us with
predictive information at all. None of this is to say that the study of joint models
has little value. Not at all. It is a very interesting activity but is more focused on
determining plausible models for a whole collection of relevant processes taken
together. In this book our focus is overwhelmingly on prediction, improvement
of model fit, and the development of the most powerful tests in given situations.
For these problems conditional distributions tell us all we need to know. Model-
ing the marginal covariate process itself does not help much, if at all, with this
endeavor.
1.6 Model-based prediction
Predictive indices
The main purpose of model construction is ultimately that of prediction. The
amount of variation explained is viewed naturally as a quantity that reflects the
1.6. Model-based prediction 13
predictive power of any model. This is a direct consequence of the Chebyshev

inequality. And yet, considerable confusion on the topic of explained variation is
prevalent among statisticians. This is to some extent surprising since the main
ideas are very simple. The origin of the confusion for many was in a paper of
Draper (1984), echoed subsequently by Healy (1984), which was strongly critical
of the R2 measure of explained variation for multi-linear models. Several authors,
Kvaalseth (1985), Scott and Wild (1991), and Willett and Singer (1988), followed
in the wake of these two papers, adding further nails, it was argued, to the coffin
of explained variation.
However, Draper’s paper was in error. He was certainly the first to notice this
and he wrote a retraction (Draper, 1985)—changing entirely the conclusion of
his earlier paper from “R2 is misleading” to “R2 is a useful indicator” (his italics).
By one of those curious twists governing the survival mechanisms of scientific
ideas, few workers in the area, and apparently none of those building on Draper’s
1984 paper, seem to be aware of the retraction. At the time of writing this text,
the retraction has been cited three times, two of which were by this author and
a colleague. The original contribution is cited 39 times while, of greater impact,
the paper of Kvalseth has been cited over one thousand times. It continues to
be frequently quoted and its erroneous conclusions are widely taken to be valid.
Indeed, there is little mystery to the idea of explained variation. Note that, by
virtue of the Chebyshev-Bienaymé inequality, explained variation directly quanti-
fies predictive ability. In O’Quigley (2008) a basic theory of explained variation
was outlined. It helps to refer to that theory in order to have a good understand-
ing of what needs to be thought through when we wish to consider the particular
case of explained variation for survival models. In the light of this theory and the
main results concerning inference it is relatively straightforward to develop suit-
able measures of explained variation (O’Quigley and Flandre (1994), O’Quigley
and Xu (2001)). The related topic, explained randomness, in view of its direct
connection to likelihood, is also discussed at length in O’Quigley (2008) In a way
similar to explained variation, although leaning on the concepts of entropy rather
than the Chebyshev-Bienaymé inequality, explained randomness will also trans-
late predictive ability. For normal models these different population quantities
happen to coincide. For non-normal models on the other hand, when it might
be argued that measures of variance are less immediate as a means to quantify
predictive effects, a case can be made for preferring explained randomness over
explained variation. In this text we work with explained variation but this is not
because of any perceived advantage over explained randomness. In practice we
would anticipate a very high level of agreement in analyzing data whichever of
the two measures are used.
Among the users of statistical models it can be argued that there are two
quite different schools of thought; the first sees a model as an approximation
to some, infinitely more complex, reality: that is the model is no more than
a tool, taken as a more-or-less refined means to achieving some specific end,
usually that of prediction of some quantity of interest. Any model is then simply
judged by its predictive performance. The second school sees the statistician’s
job as tracking down the “true” model that can be considered to have generated
the data. The position taken in this work is very much closer to the first than
the second. This means that certain well-studied concepts, such as efficiency,
a concept which assumes our models are correct, are given less attention than
is often the case. The regression parameter β in our model is typically taken
to be some sort of average where the proportional hazards model corresponds
to a summary representation of the broader non-proportional hazards model.
We view the restricted model as a working model and not some attempt to
represent the true situation. Let’s remind ourselves that the proportional hazards
model stipulates that effects, as quantified by β, do not change through time.
In reality the effects must surely change, hopefully not too much, but absolute
constancy of effects is perhaps too strong an assumption to hold up precisely. The
working model enables us to estimate useful quantities, one of them being average
regression effect, the average of β taken through time. Interestingly, the usual
partial likelihood estimator in the situation of changing regression effects does
not estimate an average effect, as is often believed. Even so, we can estimate an
average effect but we do require an estimator different from that commonly used.
Details are given in Xu and O’Quigley (2000), O’Quigley (2008) and recalled in
Section 7.6.
Interpreting predictive indices

Is 0.48 greater than 0.16? If so, then by how much. We might think that the
answer to that is easy, Indeed, the answer is easy when the discussion concerns
real numbers. If, on the other hand, the discussion concerns something more con-
crete, something that is hopefully quantified by these numbers then the question
can only be answered by first identifying just what it is we are trying to quantify
and, secondly, how that quantity relates to the real numbers used in its place.
Several popular measures of predictive strength of a survival model fail this sim-
ple test. For example, for the values 0.48 and 0.16, obtained from the popular
C-statistic, it is not even possible to state that the first of these two numbers
translates greater predictive strength than the second, let alone consider that,
in some well-defined sense, the predictive power of 0.48 is 3 times that of 0.16.
As a result this statistic lacks any kind of interpretation. Studies that make use
of estimated values of the C-statistic in their analyses, as a means to establish
orders of importance on the variables in their investigation, need to carefully take
account of the censoring as described by Gönen and Heller (2005). Otherwise,
ordering the importance of these variables alphabetically, or even randomly, may
work no less well. However, even if we follow Gonen and Heller’s prescription
to fix the censoring problem, the interpretation difficulty remains. Certainly, it is
now possible to be sure that an ordering of the indices does translate an order-
1.6. Model-based prediction 15
ing in terms of predictive strength. But the scale is still not something that we
can grasp on an intuitive level. The great majority of the so-called predictive
indices reviewed by Choodari-Oskooei et al. (2012) offer no advantage over the
C-statistic and fail to provide any useful statistical insight into the problem of
prediction for a survival model.
Pepe et al. (2015) shows that some alternative measures, such as the net
reclassification index (Pencina et al., 2008) do no better than the measures they
set out to improve upon. The main problem is almost always that of interpre-
tation and, in this context, improvised solutions that take care of the censoring
have rarely been met with success. The way to proceed, described in Section
3.9 of O’Quigley (2008), is to start from first principles, considering the elemen-
tary definitions of explained variation, what they mean, and how they might be
estimated in the presence of independent censoring. Some suggestions in the
literature have no clear meaning and it is something of a puzzle that they were
ever used at all, let alone that they continue to attract interest. The so-called
proportion of treatment effect explained (Freedman et al., 1992), believed to
quantify how much of the treatment effect can be explained by another variable,
typically some kind of surrogate endpoint for treatment, enjoys no interpretation
as a proportion. Flandre and Saidi (1999) finds values for this quantity, using
data from published clinical trials, ranging from −13% to 249%. These are the
point estimates and not the endpoints of the confidence intervals which reach
as far as 416%. There was nothing unusual in the data from those examples
and we can conclude that Freedman’s measure is not any kind of a percentage
which, necessarily, would lie between zero and one hundred percent. What it is
we cannot say although there is no reason to believe that 249% corresponds to
stronger predictive effects than −13%.
These measures are quite useless. Some measures, while not necessarily being
useless, appear to be not very useful. Take, for example, the lifetime risk of
getting breast cancer for carriers of the BRCA1 and BRCA2 genetic mutation. A
few years back we learned that a celebrity was deemed to have an 87% lifetime
risk of getting breast cancer on the basis of her BRCA status. Despite enjoying
full health, as a result of that risk she chose to have a double mastectomy. This
so-called risk prompting such drastic action became not only a topic of discussion
in scholarly medical journals but also made it into the mainstream press where,
almost universally, the celebrity was praised for her good judgment and her pro-
active stance, a stance taken to avoid the perils that appeared to be confronting
her. As far as we are aware, there was little or no discussion on the obvious
question ... what is meant by the number 87%. Lifetime risk is near impossible
to interpret (O’Quigley, 2017) and, by any standards, cannot be considered a
useful measure of anything. More interpretable measures are available and we
take the view that lifetime risk is not only not useful but is in fact misleading.
It does not measure any clinical or physiological characteristic and basing major
treatment and surgical decisions on its calculated value makes no sense. We
return to this in the chapter covering applications in epidemiology, although not

in great depth, noting, for now, that, without doubt, the lack of understanding of
the advisors concerning the complex meaning of lifetime risk will have contributed
in part to a lack of clarity that is crucial to the making of an informed decision
before undertaking mutilating life-changing surgery.
1.7 Data sets
The importance of working through the calculations on actual data cannot be

overemphasized. The reader is encouraged to use their own or any of the many
available published data sets, as well as simulated data sets corresponding to
specific conditions. In the text we often refer to the Freireich data, described by
several authors including Kalbfleisch and Prentice (2002) and Professor Cox in his
1972 founding paper. These data arose in the context of a balanced comparison of
two treatment groups. We also refer to breast cancer data which were gathered at
the Institut Curie in Paris, France over a near thirty-year period. Finally, another
important data set used here to illustrate many concepts was obtained in the
context of a survival study in gastric cancer carried out at the St James Hospital,
Leeds, U.K. (see, for example, Rashid et al. (1982). An interesting case of non-
proportional hazards arose in a clinical trial studied by Stablein et al. (1981).
The regression effect appeared to change direction during the study. We refer to
this in the text as the Stablein data. All of the examples and illustrations in this
text can be confirmed using available software and the original data the source
of which is referenced.
1.8 Use as a graduate text
The text can be used as support for an introductory graduate course in survival
analysis with particular emphasis on proportional and non-proportional hazards
models. The approach to inference is more classical than often given in such
courses, steering mostly away from the measure-theoretic difficulties associated
with multivariate counting processes and stochastic integrals and focusing instead
on the more classical, and well known, results of empirical processes. Brownian
motion and functions of Brownian motion play a central role. Exercises are pro-
vided in order to reinforce the coursework. Their aim is not so much to help
develop a facility with analytic calculation but more to build insight into the
important features of models in this setting. Some emphasis then is given to
practical work carried out using a computer. No particular knowledge of any
specific software package is assumed.
1.9. Classwork and homework 17
1.9 Classwork and homework
1. For the several examples described in Section 1.2 write down those features
of the data which appear common to all examples. Which features are
distinctive?
2. In Section 1.5 it is claimed that cumulative quantities may do enough

smoothing in practice. How do you understand this idea of “enough
smoothing?” What does a statistician aim to achieve by smoothing?
3. Suppose that the methods of survival analysis were not available to us.
Suggest how we might analyze a randomized clinical trial using (i) multiple
linear regression, (ii) multiple logistic regression.
4. For Exercise 3, describe the shortcomings that we expect to be associated

with either type of analysis. How are these shortcomings amplified by the
presence of increasing censoring, and in what way do we anticipate the
techniques of survival analysis to address these shortcomings?
5. Consider a long-term survival study in which the outcome of interest is sur-

vival itself and we wish to study any possible dependence of this upon two
other variables, one binary and one continuous. The marginal distribution
of survival is unknown but suspected to be very skew. For the data at hand
there is no censoring, i.e., all the actual survival times have been observed.
Describe at least one possible approach, aside from those of survival anal-
ysis, to analyzing such data. Do you think more could been obtained from
the data using survival techniques? Explain.
6. How would you describe to a non-scientist the purpose of a statistical

model, explaining the issues involved with misspecified models. How would
you describe to a physicist the difference between statistical models which
may be employed in epidemiology and statistical models that he or she may
be used to elaborate theories in quantum mechanics.
7. Consider a joint model for survival and an individual’s covariate path

through time. Suppose that the parametrization is such that the model
for the covariate path and the model for survival do not share any param-
eters. Will knowledge, provided by data, on the parameters that model the
covariate path, help improve inference for the regression coefficients of the
survival model? If not, why not?
8. In the context of joint modeling, find a way to make use of information on

the individual’s covariate process to improve inference on the survival pro-
cess without appealing to any structure that postulates shared parameters
between models. Otherwise, how might we decide the plausibility of such
a structure?
9. Check out the definitions for explained variation and explained randomness
for bivariate continuous distributions. Show that, in the case of a multi-
normal model, explained randomness and explained variation coincide.
10. Suppose that T1 and T2 are two survival variates considered in a competing
risks setting. We suppose independence. Imagine though that the observa-
tions of T2 can be modeled by treating T1 as a time-dependent covariate in
a Cox model. When the regression coefficient is non-zero describe how this
would impact the Kaplan-Meier curves obtained for a set of observations;
T21 ,..., T2n . What can be said when the model is an NPH one and the
regression coefficient a function of time?
Chapter 2
Survival analysis methodology
2.1 Chapter summary
We recall some elementary definitions concerning the probability distributions,

putting an emphasis toward one minus the usual cumulative distribution function,
i.e., the survival function. This is also sometimes called the survivorship function.
The closely related hazard function has, traditionally, been the most popular
function around which to construct models. For multistate models it can be
helpful to work with intensity functions, rather than hazard functions since these
allow the possibility of moving in and out of states. This is facilitated by the very
important function, Y (t), the “at-risk” indicator. A number of special parametric
cases of proportional hazards models are presented. The issue of censoring and the
different kinds of censoring is discussed. The “at-risk” indicator Yi (w, t), taking
the value one when the subject i is at risk of making a transition of a certain kind,
indicated by w, makes it particularly simple to address more complex issues in
survival such as repeated events, competing risks, and multistate modeling. We
consider some tractable parametric models, the exponential model in particular.
Survival time T will be a positive random variable, typically right skewed and
with a non-negligible probability of sampling large values, far above the mean.
The fact that an ordering, T1 > T2 , corresponds to a solid physical interpretation
has led some authors to argue that time is somehow different from other contin-
uous random variables, reminiscent of discussion among early twentieth-century
physicists about the nature of time “flowing inexorably in and of itself”. These
characteristics are sometimes put forward as a reason for considering techniques
other than the classic techniques of linear regression. From a purely statistical

19
https://doi.org/10.1007/978-3-030-33439-0_2
20 Chapter 2. Survival analysis methodology
viewpoint, this reasoning is incorrect. Elementary transformations fix the skew-

ness problems which, in consequence, reveal themselves as quite superficial. Nor
is there any worthwhile, statistical, distinction between time and, say, height or
weight. The reason for considering particular techniques, outside of the classical
ones of linear regression, is the presence of censoring. In early work, censoring
came to be viewed as a nuisance feature of the data collection, hampering our
efforts to study the main relationships of interest. A great breakthrough occurred
when this feature of the data, the censoring, was modeled by the “at-risk” func-
tion. Almost immediately it became clear that all sorts of much more involved
problems; competing risks, repeated events, correlated outcomes, could all be
handled with almost no extra work. Careful use of the “at-risk” indicator was all
that would be required. At the heart then of survival analysis is the idea of being
at risk for some event of interest taking place in a short time frame (for theoreti-
cal study this short time will be made arbitrarily small). Transition rates are then
very natural quantities to consider. In epidemiology these ideas have been well
rooted for a half-century where age-dependent rates of disease incidence have
been the main objects under investigation.
2.3 Basic tools
Time and risk

The insurance example in the introduction highlights an obvious, but an impor-
tant, issue. If driver A, on average, has a higher daily risk than driver B, then
his or her mean time to be involved in an accident will be shorter. Conversely, if
driver B has a longer mean time to accident, then he has, on average, a lower
daily risk. For many examples we may tend to have in mind the variable time
and how it is affected by other variables. But we can think equally well in terms
of risk over short time periods, a viewpoint that we will see generalizes more
readily to be able to deal with complicated situations. The connection between
time and risk is outlined more formally below.
Conditional distributions
Our goal is to investigate dependence, its presence, and, if present, the degree of
dependence. The information on this is conveyed to us via the conditional distri-
bution of time given the covariate, keeping in mind that this covariate may be a
combination of several covariates. The conditional distribution of the covariate
given time may not immediately strike us as particularly relevant. However, it is
very relevant because: (1) the marginal distribution of time and the marginal dis-
tribution of the covariate contain little or no information on dependency so that
the two conditional distributions can be viewed, in some sense, as being equiv-
alent and, (2) problems relating to censoring are near intractable when dealing
with the conditional distribution of time given the covariate whereas they become
2.3. Basic tools 21
somewhat straightforward when we consider the conditional distribution of the

covariate given time. This is a very central idea and runs throughout this text.
Hazard and related functions

The purpose here is to continue the introduction of preliminary notions and some
basic concepts. Before discussing data and estimation we consider the problem
in its most simplified form as that of the study of the pair of random variables
(T, Z), T being the response variable “survival” of principal interest and Z an
associated “explanatory” variable. There would be little difficulty in applying the
host of techniques from linear regression to attacking this problem were it not
for the presence of a “censoring” variable C. The particularity of C is that, when
observed, i.e., C = c, we are no longer able to observe values of T for which
T > c. Also, in most cases, when T is observed, we are no longer able to observe
C. Nonetheless an observation on one tells us something about the other, in
particular, that it must assume some greater value.
Although the joint distribution of (T, Z) can be of interest, we are particularly
interested in the conditional distribution of T given Z. First let us consider T
alone. The probability density function of T is defined as
1
f (t) = lim Pr(t < T < t + Δt), (2.1)
Δt→0+ Δt
where limΔt→0+ means that Δt goes to 0 only through positive values. We
t
define as usual F (t) = 0 f (u)du. The survivorship function is written as S(t) =
1 − F (t). If we view the density as the unconditional failure rate, we can define
a conditional failure rate as being the same quantity after having accounted for
the fact that the individual has already survived until the time point t. We call
this λ(t) and we define
1
λ(t) = lim Pr(t < T < t + Δt|T > t). (2.2)
Δt→0+ Δt
It helps understanding to contrast Equation (2.2) and (2.1) where we see that
λ(t) and f (t) are closely related quantities. In a sense, the function f (t) for all
values of t is seen from the standpoint of an observer sitting at T = 0, whereas,
for the function λ(t), the observer moves along with time looking at the same
quantity but viewed from the position T = t. Analogous to a density, conditioned
by some event, we can define
1
λ(t|C > t) = lim Pr(t < T < t + Δt|T > t, C > t). (2.3)
Δt→0+ Δt
The conditioning event C > t is of great interest since, in practical investigations,
all our observations at time t have necessarily been conditioned by the event. All
associated probabilities are also necessarily conditional. But note that, under an
independent censoring mechanism, we have that λ(t|C > t) = λ(t). This result
underlies the great importance of certain assumptions, in this case that of inde-
pendence between C and T . The conditional failure rate, λ(t), is also sometimes
referred to as the hazard function, the force of mortality, the instantaneous failure
rate or the age-specific failure rate. If we consider a small interval then λ(t) × Δt
closely approximates the probability of failing in a small interval for those aged
t, the approximation improving as Δt goes to zero. If units are one year then
these are yearly death rates. The
t cumulative hazard function is also of interest
and this is defined as Λ(t) = 0 λ(u)du. For continuous λ(t), using elementary
calculus we can see that:
λ(t) = f (t)/S(t) , S(t) = exp{−Λ(t)} , f (t) = λ(t) exp{−Λ(t)}.
Although mathematically equivalent, we may prefer to focus attention on one

function rather than another. The survival function, S(t), is the function dis-
playing most clearly the information the majority of applied workers are seeking.
The hazard function, λ(t), of central concern in much theoretical work, provides
the most telling visual representation of time effects. An important function, of
theoretical and practical interest, is the conditional survivorship function,
S(t, u) = Pr(T > t|T > u) = exp{Λ(u) − Λ(t)} , (u < t).
From this it is clear that S(t, u) = S(t)/S(u) and that S(u, u) = 1 so that it is as
though the process had been restarted at time t = u. Other quantities that may
be of interest in some particular contexts are the mean residual lifetime, m(t),
and the mean time lived in the interval [0, t], μ(t), defined as
t
m(t) = E(T − t|T ≥ t), μ(t) = S(u)du. (2.4)
0
Like the hazard itself, these functions provide a more direct reflection on the
impact of having survived until time t. The mean residual lifetime provides a very
interpretable measure of how much more time we can expect to survive, given
that we have already reached the timepoint t. This can be useful in actuarial appli-
cations. The mean time lived in the interval [0, t] is not so readily interpretable,
requiring a little more thought (it is not the same as the expected lifetime given
that T < t). It has one strong advantage in that it can be readily estimated from
right-censored data in which, without additional assumptions, we may not even be
able to estimate the mean itself. The functions m(t) and μ(t) are mathematically
equivalent to one another as well as the three described above and, ∞for example,
a straightforward integration by parts shows that m(t) = S −1 (t) t S(u)du and
that μ(∞) = E(T ). If needed, it follows that the survivorship function can be
expressed in terms of the mean residual lifetime by
2.3. Basic tools 23
t
−1 −1
S(t) = m (t)m(0) exp − m (u)du .
0
We may wish to model directly in terms of m(t), allowing this function to depend
on some vector of parameters θ. If the expression for m(t) is not too intractable
then, using f (t) = −S (t) and the above relationship between m(t) and S(t),
we can write down a likelihood for estimation purposes in the situation of inde-
pendent censoring. An interesting and insightful relationship (see, for instance,
the Kaplan-Meier estimator) between S(t) and S(t, u) follows from considering
some discrete number of time points of interest. Thus, for any partition of the
time axis, 0 = a0 < a1 <, . . . , an = ∞, we see that

S(aj ) = S(aj−1 )S(aj , aj−1 ) = S(a , a−1 ). (2.5)
≤j
The implication of this is that the survival function S(t) can always be viewed
as the product of a sequence of conditional survival functions, S(t, u). This
simple observation often provides a foundation that helps develop our intuition,
one example being the case of competing risks looked at later in this chapter.
Although more cumbersome, a theory could equally wellbe constructed for the
discrete case whereby f (ti ) = Pr(T = ti ) and S(ti ) = ≥i f (t ). We do not
explore this here.
Intensity functions and compartment models

Modern treatment of survival analysis tends to focus more on intensity than
hazard functions. This leads to great flexibility, enabling, for example, the con-
struction of simple models to address questions in complex situations such as
repeated events (Andersen and Gill, 1982). We believe that both concepts can
be useful and we will move back and forth between them according to the appli-
cation. Intensity functions find their setting in the framework of stochastic pro-
cesses where the random nature of T is suppressed, t being taken simply as an
index to some stochastic process. The counting process N (t), takes the value 0 at
Alive Dead
Figure 2.1: An elementary alive/dead compartment model with two states
t = 0, remaining at this same value until some time point, say T = u, at which
the event under study occurs and then N (t) = 1t≥u . We can then define, in an
infinitesimal sense, i.e., the equality only holds precisely in the limit as dt goes
to zero through positive values
Pr{(N (t) − N (t − dt) = 1|Ft−dt )} = α(t)dt (2.6)
where Ft−dt , written as Ft− when we allow dt > 0 to be arbitrarily close to zero,
is the accumulated information, on all processes under consideration, observed
up until time t − dt (Figure 2.1).
The observed set Ft− is referred to as the history at time t. The set is nec-
essarily non-decreasing in size as t increases, translating the fact that more is
being observed or becoming known about the process. The Kolmogorov axioms of
probability, in particular sigma additivity, may not hold for certain noncountable
infinite sets. For this reason probabilists take great care, and use considerable
mathematical sophistication, to ensure, in broad terms, that the size of the set
Ft− does not increase too quickly with t. The idea is to ensure that we remain
within the Kolmogorov axiomatic framework, in particular that we do not vio-
late sigma additivity. Much of these concerns have spilled over into the applied
statistical literature where they do not have their place. No difficulties will arise
in applications, with the possible exception of theoretical physics, and the prac-
titioner, unfamiliar with measure theory, ought not to be deterred from applying
the techniques of stochastic processes simply because he or she lacks a firm grasp
of concepts such as filtrations. It is hard to imagine an application in which a lack
of understanding of the term “filtration” could have led to the error. On the other
hand, the more accessible notions of history, stochastic process, and conditioning
sets are central and of great importance both to understanding and to deriving
creative structures around which applied problems can be solved. Viewing t as
an index to a stochastic process rather than simply the realization of a random
variable T , and defining the intensity process α(t) as above, will enable great
flexibility and the possibility to model events dynamically as they unfold.
Alive in State A Alive in State B Alive in State C
Dead
Figure 2.2: A compartment model with 3 covariate levels and an absorbing state.
Model is fully specified by the 3 transition rates from states A, B, or C to the
state Dead
2.3. Basic tools 25
At-risk functions Y (t), Y (w, t) and multistate models

The simplest case we can consider occurs when following a randomly chosen
subject through time. The information in Ft− tells us whether or not the event
has yet occurred and if the subject is still at risk, i.e., the set Ft− is providing
the same information as an observation on the function Y (t) where we take Y (t)
to be left continuous, assuming the value one until the occurrence of an event,
or removal from observation, at which time it assumes the value zero. If the
simple fact of not having been removed from the study, the event (C > t) is
independent of the event (t < T < t + dt), then conditioning on Y (t) = 1 is the
same as conditioning on T > t. Referring then to Equation 2.2, it is clear that
if Y (t) = 0 then α(t) = 0 and, if Y (t) = 1 then α(t) = λ(t). Putting these two
results together we have
α(t) = Y (t)λ(t). (2.7)
This relation is important in that, under the above condition, referred to as the
independent censoring condition, the link between the intensity function and the
hazard function is clear. Note that the intensity function is random since Y is
random when looking forward in time. Having reached some time point, t say,
then α(t) is fixed and known since the function Y (u), 0 < u < t is known and
Y (t) is left continuous.
We call Y (·) the “at risk” function (left continuous specifically so that at
time t the intensity function α(t) is not random). The idea generalizes readily
and in order to cover a wide range of situations we also allow Y to have an
argument w where w takes integer values counting the possible changes of state.
For the ith subject in any study we will typically define Yi (w, t) to take the value
1 if this subject, at time t, is at risk of making a transition of type w, and 0
otherwise. Figure 2.2 summarizes a situation in which there are four states of
interest, an absorbing state, death, and three states from which an individual
is able to make a transition into the death state. Transitions among the three
non-death states themselves cannot occur. Later we will consider different ways
of modeling such a situation, depending upon further assumptions we may wish
or not wish to make.
In Figure 2.3 there is one absorbing state, the death state, and two non-
absorbing states between which an individual can make transitions. We can define
w = 1 to indicate transitions from state 1 to state 2, w = 2 to indicate transitions
from state 2 to state 1, w = 3 to indicate transitions from state 1 to state 3 and,
finally, w = 4 to indicate transitions from state 2 to state 3. Note that such an
enumeration only deals with whether or not a subject is at risk for making the
transition, the transition probabilities (intensities) themselves could depend on
the path taken to get to the current state.
State 1 State 2
No symptoms Progression
State 3
Dead
Figure 2.3: Binary time-dependent covariate and single absorbing state. Prob-
abilistic structure is fully specified by the transition rates at time t, αij (t) for
leaving state i for state j. Fix and Neyman’s illness-death model, also known as
the semi-competing risks model, is a special case when α21 (t) = 0, ∀t.
We can then appreciate why it can be helpful to frame certain questions in

terms of compartment models, intensity functions and the risk function. Rather
complex situations can be dealt with quite straightforwardly, the figures illustrat-
ing simple cases where we can use the argument w in Yi (w, t) to indicate, at any
t, which kinds of transition any given subject i is available to make. In Figure 2.4
there are two absorbing states, one of which can only be reached from state 2.
The transition rate between state 2 and state 4 may or may not depend on the
number of times a subject moves between states 1 and 2. Allowing for transi-
tions between states greatly adds to the flexibility of any model so that, in Figure
2.2, although the explanatory variable (state) has three levels, the model is, in
principle, much simpler than that described in Figure 2.3 where the explanatory
variable maybe characterized by the history as well as the current state.
At-risk indicator Y (w, t) and repeated events

Some studies have the particularity that an occurrence of the event of interest
does not remove the subject from further observation. Additional events, of the
same or of different types, may happen. An example is benign breast disease,
potentially followed by malignant disease. A patient may have several incidences
of benign breast disease at different intervals of time. Following any one of
these incidences, or even before such an incidence takes place the subject may
become an incident for malignant disease. If our interest is essentially focussed
on the incidence of malignant disease then we would treat the time-dependent
history of benign breast disease as a potential explanatory variable for incidence
of malignant disease. However, we may also be interested in modeling directly
2.3. Basic tools 27
State 1 State 2
State 3 State 4
Figure 2.4: Binary time-dependent covariate and two absorbing states. Transi-
tions to state 4 pass through state 2, i.e., αj4 (t) = 0, j = 2, ∀t. Note also that
α4j (t)=0.
the repeated incidence of benign breast disease in its own right. Clearly a patient
can only be at risk of having a third incident of benign breast disease if she has
already suffered two earlier incidents. We can model the rate of incidence for the
j th occurrence of benign disease as,
αj (t) = Y (j, t)λj (t − tj−1 ), (2.8)
where t0 = 0 and tj is the observed occurrence of the j th event. Different options

may be considered for modeling λj (t). Usually there will be at least one covariate,
Z, indicating two distinct prognostic groups, possibly established on the basis
of different treatments. The model will involve coefficients multiplying Z and
thereby quantifying treatment effect. Allowing these coefficients to also depend
upon j provides the broadest generality and is equivalent to analyzing separate
studies for each of the occurrences. Stronger modeling, imposing greater struc-
ture, might assume that the coefficients do not depend upon j, in which case
the information provided by a subject having three incident cases is compara-
ble to that of three independent subjects each providing information on a single
incident. So-called marginal models have been proposed in this context. Here,
it would be as though the subject, after an event, starts the clock from zero
and, aside from covariate information, is deemed to be in the same position as
another subject who has just entered the study without having yet suffered a sin-
gle event. A lot of information would appear to be thereby gained but the setup
seems rather artificial and implausible. Starting the clock from zero, after each
event, is sensible but it is more realistic to assume that the underlying hazard
rates, i.e., those not adjusted by covariate information, would change with the
number of prior incidents. In other words a more plausible model would condition
on this information allowing the baseline hazard rate to change according to the
number of events counted so far. Stratified proportional hazards models have

been successfully used in order to analyze such problems.
2.4 Some potential models
Simple exponential
The simple exponential model is fully specified by a single parameter λ. The
hazard function, viewed as a function of time, does not in fact depend upon
time so that λ(t) = λ. By simple calculation we find that Pr(T > t) = exp(−λt).
Note that E(T ) = 1/λ and, indeed, the exponential model is often parameter-
ized directly in terms of the mean θ = E(T ) = 1/λ. Also Var(t) = 1/λ2 . This
model expresses the physical phenomenon of no aging or wearing out since, by
elementary calculations, we obtain S(t + u, u) = S(t); the probability of surviving
a further t units of time, having already survived until time u, is the same as that
associated with surviving the initial t units of time. The property is sometimes
referred to as the lack of memory property of the exponential model.
For practical application the exponential model may suggest itself in view
of its simplicity or sometimes when the constant hazard assumption appears
realistic. A good example is that of a light bulb which may only fail following
a sudden surge in voltage. The fact that no such surge has yet occurred may
provide no information about the chances for such a surge to take place in the
next given time period. If T has an exponential distribution with parameter λ then
λT has the so-called standard exponential distribution, i.e., mean and variance
are equal to one.
Recall that for a random variable Y having normal distribution N (μ, σ 2 )
it is useful to think in terms of a simple linear model Y = μ + σ, where
has the standard distribution N (0, 1). As implied above, scale changes for the
exponential model lead to a model still within the exponential class. However,
this is no longer so for location changes so that, unlike the normal model in which
linear transformations lead to other normal models, a linear formulation for the
exponential model is necessarily less straightforward. It is nonetheless of interest
to consider the closest analogous structure and we can write
Y = log T = α + bW, (2.9)
where W has the standard extreme value density f (w) = exp{w − exp(w)}.
When α = 0 we recover an exponential model for T with parameter b, values
other than zero for α pushing the variable T out of the restricted exponential
class into the broader Weibull class discussed below.
2.4. Some potential models 29
Proportional hazards exponential

In anticipation of the central topic of this book (that of heterogeneity among the
subjects under study) imagine that we have two groups, indicated by a binary
variable Z = 0 or Z = 1. For Z = 0 the subjects follow an exponential law with
parameter λ0 . For Z = 1 the subjects follow an exponential law with parameter
λ1 . It is clear that for the hazard functions there exists real β (= log λ1 − log λ0 )
such that
λ(t|Z) = λ(t|Z = 0) exp(βZ) = λ0 exp(βZ). (2.10)
The important point to note here is that the ratio of the hazards, λ(t|Z =
1)/λ(t|Z = 0) does not involve t. It also follows that S(t|Z = 1) = S(t|Z = 0)α
where α = exp(β). The survival curves are power transformations of one another.
This is an appealing parameterization since, unlike a linear parameterization,
whatever the true value of β, the constraints that we impose upon S(t|Z = 1)
and S(t|Z = 0) in order to be well-defined probabilities, i.e., remaining between
0 and 1, are always respected. Such a model is called a proportional hazards
model. For three groups we can employ two indicator variables, Z1 and Z2 , such
that, for group 1 in which the hazard rate is equal to λ0 , Z1 = 0 and Z2 = 0, for
group 2, Z1 = 1 and Z2 = 0 whereas for group 3, Z1 = 0 and Z2 = 1. We can
then write;
λ(t|Z) = λ0 exp(β1 Z1 + β2 Z2 ), (2.11)
where λ0 = λ(t|Z1 = Z2 = 0). It is worthwhile bringing the reader’s attention to

just where the constraints of the model express themselves here. They concern
the hazard rates for all groups, which are assumed to be constant. Given this
constraint there are no further constraints concerning the relationship between
the groups. Suppose, though, that we were to consider a further group, group 4,
defined by Z1 = 1 and Z2 = 1. In order to add a fourth group without introducing
a further binary coding variable Z3 , we introduce the constraint that the hazard
for group 4 is simply expressed in terms of the hazards for groups 2 and 3. Such
assumptions are commonly made in routine data analysis but, nonetheless, ought
to come under critical scrutiny. We return to this issue in later chapters. The
extension to many groups follows in the same way. For this we take Z to be a p-
dimensional vector of indicator variables and β a vector of parameters having the
same dimension as Z, the product βZ in Equation 2.10 now implying an inner
product, i.e., βZ = pi=1 βi Zi . In this case, the proportional hazards exponential
model (2.10) implies that every group follows some simple exponential law, a
consequence being that the survivorship function for any group can be expressed
as a power transformation of any other group. Once again, it is important to keep
in mind just which assumptions are being made, the potential impact of such
assumptions on conclusions, as well as techniques for bringing under scrutiny
these assumptions. The proportional hazards constraint then appears as a very

natural one in which we ensure that the probabilities S(t|z) and subsequent
estimates always remain between 0 and 1. A linear shift added to S(t|0) would
not allow for this. We do nonetheless have a linear shift although on a different,
and thereby more appropriate, scale and we can write

p
log − log S(t|Z) = log − log S(t|0) + βi Zi .
i=1
This formulation is the same as the proportional hazards formulation. Noting that
− log S(T |Z = z) is an exponential variate some authors prefer to write a model
down as a linear expression in the transformed random variable itself with an
exponential error term. This then provides a different link to the more standard
linear models we are familiar with.
Piecewise exponential
The lack of flexibility of the exponential model will often rule it out as a potential
candidate for application. Many other models, only one or two of which are
mentioned here, are more tractable, a property stemming from the inclusion
of at least one additional parameter. Even so, it is possible to maintain the
advantages of the exponential model’s simplicity while simultaneously gaining
flexibility. One way to achieve this is to construct a partition of the time axis
0 = a0 < a1 < . . . < ak = ∞. Within the jth interval (aj−1 , aj ) , (j = 1, . . . , k) the
hazard function is given by λ(t) = λj . We can imagine that this may provide quite
a satisfactory approximation to a more involved smoothly changing hazard model
in which the hazard function changes through time. We use S(t) = exp{−Λ(t)}
to obtain the survival function where

k
k
Λ(t) = I(t ≥ aj )λj (aj − aj−1 ) + I(aj−1 ≤ t < aj )λj (t − aj−1 ). (2.12)
j=1 j=1
Properties such as the lack of memory property of the simple exponential have
analogs here by restricting ourselves to remaining within an interval. Another
attractive property of the simple exponential is that the calculations are straight-
forward and can be done by hand and, again, there are ready analogues for the
piecewise case. Although the ready availability of sophisticated computer pack-
ages tends to eliminate the need for hand calculation, it is still useful to be able
to work by hand if for no other purposes than those of teaching. Students gain
invaluable insight by doing these kinds of calculations the long way.
Proportional hazards piecewise exponential

In the same way as for the simple exponential model, for two groups, indicated
by a binary variable Z = 0 or Z = 1, each having constant piecewise rates on the
same intervals, it is clear that there exists βj such that, for t ∈ [aj−1 , aj ),
λ(t|Z) = λ(t|Z = 0) exp(βj Z) = λ0 (t) exp{β(t)Z}, (2.13)

where we now have a function β(t) = kj=1 βj I(aj−1 ≤ t < aj ). This can be
described as a nonproportional hazards model and, if, under a further restric-
tion that β(t) is a constant function of time, i.e., β1 = β2 = · · · = βk = β,
then, as for the simple exponential model, we have S(t|Z = 1) = S(t|Z = 0)α
where α = exp(β) and, once again, such a model is called a proportional hazards
model. The model can once more be described in terms of a linear translation
on log − log S(t|z).
Weibull model
Another way to generalize the exponential model to a wider class is to consider
a power transformation of the random variable T . For any positive γ, if the
distribution of T γ is exponential with parameter λ, then the distribution of T
itself is said to follow a Weibull model whereby
f (t) = λγ(λt)γ−1 exp{−(λt)γ }
and S(t) = exp −(λt)γ . The hazard function follows immediately from this and
we see, as expected, that when γ = 1 an exponential model with parameter λ is
recovered. It is of interest to trace out the possible forms of the hazard function
for any given λ. It is monotonic, increasing for values of γ greater than 1 and
decreasing for values less than 1. This property, if believed to be reflected in some
given physical situation, may suggest the appropriateness of the model for that
same situation. An example might be the time taken to fall over for a novice one-
wheel skateboard enthusiast—the initial hazard may be high, initially decreasing
somewhat rapidly as learning sets in and thereafter continuing to decrease to
zero, albeit more slowly.
The Weibull model, containing the exponential model as a special case, is
an obvious candidate structure for framing questions of the sort—is the hazard
decreasing to zero or is it remaining at some constant level? A null hypothesis
would express this as H0 : γ = 1. Straightforward integration shows that E(T r ) =
λ−r Γ(1 + r/γ) where Γ(·) is the gamma function,
∞
Γ(p) = up−1 e−u du p > 0.
0
For p integer Γ(p) = (p − 1)! The mean and the variance are λ−1 Γ(1 + 1/γ)
and λ−2 Γ(1 + 2/γ) − E 2 , respectively. The Weibull model can be motivated by
the theory of statistics of extremes. The distribution coincides with the limiting
distribution of the smallest of a collection of random variables, under broad
conditions on the random variables in question (Kalbfleisch and Prentice, 2002,
page 48).
Proportional hazards Weibull

Once again, for two groups indicated by a binary variable Z = 0 or Z = 1, shar-
ing a common γ but different values of λ, then there exists a β such that
λ(t|Z)/λ(t|Z = 0) = exp(βZ). Since, as above, the right-hand side of the equa-
tion does not depend on t, then we have a proportional hazards model. This situ-
ation and the other two described above are the only common parametric models
that come under the heading proportional hazards models by simply expressing
the logarithm of the location parameter linearly in terms of the covariates. The
situation for more than two groups follows as before. Consider however a model
such as
λ(t|Z) = λγ(λt)γ−1 exp(βZ), (2.14)
in which Z indicates three groups by assuming the values Z = 1, 2, 3.

Unlike the model just above in which three groups were represented by two
distinct binary covariates, Z1 and Z2 , we have only one covariate. In the context
of estimation and a given set of data we will almost invariably achieve greater pre-
cision in our estimates when there are less parameters to estimate. We would then
appear to gain by using such a model. As always though, any such gain comes at
a price and the price here is that we have made much stronger assumptions. We
are assuming that the signed “distance” between groups 1 and 2, as measured by
the logarithm of the hazard, is the same as the signed distance between groups 2
and 3. If this is not the case in reality then we are estimating some sort of com-
promise, the exact nature of which is determined by our estimating equations. In
an extreme case in which the distances are the same but the signs are opposite
we might erroneously conclude that there is no effect at all. At the risk of being
repetitive, it cannot be stressed too much just how important it is to identify the
assumptions we are making and how they may influence our conclusions. Here
the assumptions concern both the parametric form of the underlying risk as well
as the nature of how the different groups are related. Allowing a shape parame-
ter γ to be other than one provides a more flexible model for the underlying risk
than that furnished by the simple exponential model. The choice of covariate
coding, on the other hand, is more restrictive than the earlier choice. All of this
needs to be studied in applications. An interesting point is that, for the three
group cases defined as above, the “underlying” hazard, λ(t|Z = 0) = λγ(λt)γ−1
does not correspond to the hazard for any of the three groups under study. It is
common in practice to consider a recoding of Z, a simple one being Z − Z̄, so

that the underlying hazard will correspond to some kind of average across the
groups. For the case just outlined, another simple recoding is to rewrite Z as
Z − 2, in which case the underlying hazard corresponds to the middle group, the
other two groups having hazard rates lower and greater than this, respectively.
Log-minus-log transformation
As a first step to constructing a model for S(t|Z) we may think of a linear
shift, based upon the value of Z, the amount of the shift to be estimated from
data. However, the function S(t|Z) is constrained, becoming severely restricted
for both t = 0 and for large t where it approaches one and zero respectively.
Any model would need to accommodate these natural constraints. It is usu-
ally easiest to do this by eliminating the constraints themselves during the ini-
tial steps of model construction. Thus, log S(t|Z) = −Λ(t) is a better starting
point for modeling, weakening the hold the constraints have on us. However,
log − log S(t|Z) = log Λ(t) is better still. This is because log Λ(t) can take any
value between −∞ and +∞, whereas Λ(t) itself is constrained to be positive.
The transformation log − log S(t|Z) is widely used and is called the log-minus-
log transformation. The above cases of the exponential and Weibull proportional
hazards models, as already seen, fall readily under this heading.
Other models
The exponential, piecewise exponential, and Weibull models are of particular
interest to us because they are especially simple and of the proportional hazards
form. Nonetheless there are many other models which have found use in practical
applications. Some are directly related to the above, such as the extreme value
model in which
t−μ
S(t) = exp − exp ,
σ
since, if T is Weibull, then log T is extreme value with σ = 1/γ and μ = log λ.
These models may also be simple when viewed from some particular angle. For
instance, if M (s) is the moment-generating function for the extreme value density
then we can readily see that M (s) = Γ(1 + s). A distribution, closely related to
the extreme value distribution (Balakrishnan and Johnson, 1994), and which has
found wide application in actuarial work is the Gompertz where
S(t) = exp βα−1 (1 − eαt ) .
The hazard rates for these distributions increase with time, and, for actuar-
ial work, in which time corresponds to age, such a constraint makes sense for
studying disease occurrence or death. The normal distribution is not a natural
candidate in view of the tendency for survival data to exhibit large skewness, not
forgetting that times themselves are constrained to be positive. The log-normal

distribution has seen some use but is most often replaced by the log-logistic,
similar in shape apart from the extreme tails, and much easier to work with. The
form is particularly simple for this model and we have
S(t) = (1 + (αt)γ )−1 .
For two groups, sharing a common γ but different values of α it is interesting to

note that the hazard ratio declines monotonically with time t to its asymptotic
value of one. Such a model may be appropriate when considering group effects
which gradually wane as we move away from some initial time point.
Parametric proportional hazards models

In principle, for any parametric form, the above providing just a very few exam-
ples, we can make a straightforward extension to two or more groups via a
proportional hazards representation. For example, if the survivorship functions of
two groups are S(t|Z = 1) and S(t|Z = 0) then we can introduce the parameter
α to model one group as a power transform of the other. Rewriting α to include
Z via α = exp (βZ) then we have an expression involving the regressors,
log − log S(t|Z) = log − log S(t|Z = 0) + βZ. (2.15)
All parameters, including β, can be estimated using standard techniques, max-

imum likelihood, in particular, the only restriction being that we require some
conditions on the censoring variable C. In practice, standard techniques are rarely
used, most likely as a consequence of the attractive proposal of Cox (1972)
whereby we can estimate β without having to consider the form of S(t|Z = 1) or
S(t|Z = 0). As attractive as the Cox approach is though, we should not overlook
the fact that, in exchange for generality concerning the possible parametric forms
of functions of interest, such as S(t|Z), making inferences on these population
quantities become that much more involved. Parametric proportional hazards
models may be an area that merits renewed interest in applications.
2.5 Censoring
The most important particularity of survival data is the presence of censoring.

Other aspects such as the positivity and skewness of the main random variable
under study, time T , and other complex situations such as repeated measures or
random effects, are not of themselves reasons for seeking methods other than
linear regression. Using transformations and paying careful attention to the struc-
ture of the error, linear models are perfectly adequate for dealing with almost
any situation in which censoring does not arise. It is the censoring that forces us
to consider other techniques. Censoring can arise in different ways.
2.5. Censoring 35
We typically view the censoring as a nuisance feature of the data, and not
of direct interest in its own right, essentially something that hinders us from
estimating what it is we would like to estimate. In order for our endeavors to
succeed we have to make some assumptions about the nature of the censoring
mechanism. The assumptions may often be motivated by convenience, in which
case it is necessary to give consideration as to how well grounded the assumptions
appear to be as well as to how robust are the procedures to depart from any
such assumptions. In other cases the assumptions may appear natural given the
physical context of interest, a common case being the uniform recruitment into
a clinical trial over some predetermined time interval. When the study closes
patients for whom the outcome of interest has not been observed are censored
at study close and until that point occurs it may be reasonable to assume that
patients are included in the study at a steady rate.
It is helpful to think of a randomly chosen subject being associated with a
pair of random variables (T, C), an observation on one of the pair impeding
observation on the other, while at the same time indicating that the unobserved
member of the pair must be greater than the observed member. This idea is
made more succinct by saying that only the random variable X = min(T, C)
can be fully observed. Clearly Pr (X > x) = Pr (T > x, C > x) and we describe
censoring as being independent whenever
Pr (X > x) = Pr (T > x, C > x) = Pr (T > x) Pr (C > x). (2.16)
Type I censoring
Such censoring most often occurs in industrial or animal experimentation. Items
or animals are put on test and observed until failure. The study is stopped at
some time T ∗ . If any subject does not fail it will have observed survival time at
least equal to T ∗ . The censoring times for all those individuals being censored
are then equal to T ∗ . Equation (2.16) is satisfied and so this is a special case of
independent censoring, although not very interesting since all subjects, from any
random sample, have the same censoring time.
Type II censoring
The proportion of censoring is determined in advance. So if we wish to study
100 individuals and observed half of them as failures we determine the number
of failures to be 50. Again all censored observations have the same value T ∗
although, in this case, this value is not known in advance. This is another special
case of independent censoring.
Type III censoring

In a clinical trial patients enter randomly. A model for entry is often assumed to
be uniform over a fixed study period, anywhere from a few months to several
years but determined in advance. Survival time is the time from entry until the
event of interest. Subjects can be censored because (1) the end of the study
period is reached, (2) they are lost to follow-up (3) the subject fails due to
something unrelated to the event of interest. This is called random censoring.
So, unlike for Type I or Type II censoring, for a random sample C1 , . . . , Cn , the
Ci could all be distinct.
For a random sample of pairs (Ti , Ci ), i = 1, . . . , n, we are only able to
observe Xi = min(Ti , Ci ). A fundamental result in this context was discovered
by Tsiatis (1975). The result says that, for such data, we are unable to estimate
the joint distribution of the pair (T, C). Only the marginal distributions can be
estimated under the independent censoring assumption, the assumption itself not
being testable from such data. It is common then to make the assumption of
independent censoring, sometimes referred to as non-informative censoring, by
stipulating that
Pr (Xi > x) = Pr (Ti > x, Ci > x) = Pr (Ti > x) Pr (Ci > x). (2.17)
The assumption is strong but not entirely arbitrary. For the example of the clinical
trial with a fixed closing date for recruitment it seems reasonable to take the
length of time from entry up until this date as not being associated with the
mechanism generating the failures. For loss to follow-up due to an automobile
accident or due to leaving the area, again the assumption may be reasonable,
or, at least, a good first approximation to a much more complex, unknown, and
almost certainly unknowable, reality.
Informative censoring, Koziol-Green model

When censoring is informative, which we can take to be the negation of non-
informative, then it is no longer possible to estimate the main quantities of inter-
est without explicitly introducing some model for the censoring. The number of
potential models relating C and T is infinite and, in the absence of special knowl-
edge, it can be helpful to postulate some simple relationship between the two,
the proportional hazards model itself having been used in this context (CsÖrgÓ
and Horvath, 1981; Slud and Rubinstein, 1983). This model was first introduced
by Koziol and Green (1976) and has a wide range of applicability (Gaddah and
Braekers, 2009; Gather and Pawlitschko, 1998; Stute, 1992). Obvious examples
might be surrogate endpoints in the study of the evolution of AIDS following
treatment, where, for falling CD4 cell counts, below a certain point patients can
be withdrawn from the study. Censoring here is clearly informative. This will be
the case whenever the fact of removing a subject, yet to experience the event of
2.5. Censoring 37
interest, from the study implies a change in risk. Informative censoring is nec-
essarily more involved than non-informative censoring and we have to resort to
more elaborate models for the censoring itself in order to make progress. If, as
might be the case for a clinical trial where the only form of censoring would
be the termination of the study, we know for each subject, in advance, their
censoring time C, we might then postulate that
log − log S(t) = log − log(S(t|C < t) + βI(C > t).
This would be a proportional hazards model for a dependent censoring mecha-

nism. More generally we would not know C in advance of making observations
on T, but we could write down a similar model in terms of intensity functions,
viewing the censoring indicator as a predictable stochastic process. For the pur-
poses of estimation we may require empirical quantities indicating how the risk
changes once censoring is observed, and for this we need to be able to compare
rates between those censored at some point and those who are not. Mostly, once
censoring has occurred, it is no longer possible to observe the main event under
study so that, for data of this nature, we are not able to estimate parameters
of interest without further assumptions. These assumptions are usually that the
censoring is independent of the failure process or that it is conditionally inde-
pendent given covariate values. The paper of Tsiatis (1975) demonstrates this
intuitive observation formally.
Marginal and conditionally independent censoring

When considering many groups, defined by some covariate value Z, there are
essentially two types of independence commonly needed. The stronger assump-
tion is that of marginal independence in which the variables T and C as well
as C and Z are pairwise independent. The censoring distribution for C is the
same for different values of Z. A weaker assumption that is often made is that
of conditional independence. Here, the pair (T, C) are independent given Z. In
other words, for each possible value of Z, the pair (T, C) is independent, but
the censoring distribution C can be different for different values of Z. A nice
illustration of conditional independence was shown by Cox (1972) in his analysis
of the Freireich data. The assumption appeared to be a reasonable one. The
stronger assumption of marginal independence is often needed in more complex
situations, an example being the consistent estimation of R2 , and needs to be
borne in mind and examined critically. We return to this later in the text.
Finite censoring support

Many mathematical issues simplify immediately when the failure variable T is
continuous, as we generally suppose, but that the censoring variable is restricted
to having support on some finite subset. We can imagine that censoring times
are only allowed to take place on the set {a0 , a1 , . . . , ak }. This is not a practi-
cal restriction since we can make the division (aj , aj−1 ) as fine as we wish. We
will frequently need to consider the empirical distribution function and analogues
(Kaplan-Meier estimate, Nelson-Aalen estimate) in the presence of censoring. If
we adopt this particular censoring setup of finite censoring support, then gener-
alization from the empirical distribution function to an analogue incorporating
censoring is very straightforward. We consider this in greater detail when we
discuss the estimation of marginal survival.
Competing risks and censoring

Recalling the “at-risk” indicator function, Yi (w, t), which takes the value one
if, at time t, the i th subject is at risk of making a transition of type w, and
is zero otherwise, we can imagine a simple situation in which w takes only one
of two values. Calling these w = 1 and w = 2, consider a constraint whereby
Yi (1, t) = Yi (2, t). In other words, if the ith subject is at risk of one kind of
transition, then he or she is also at risk of the other kind. If the subject is
no longer at risk then this means that they are not at risk for either kind of
transition. Thus, if a subject suffers an event of type w = 1 then he or she is no
longer considered at risk of suffering an event of type w = 2, and conversely.
This is the situation of competing risks. As long as the subject is at risk,
then either of the event types can occur. Once one type of event has occurred,
then it is no longer possible to observe an occurrence of an event of the other
type. Such a construction fits in immediately with the above models for survival
involving censoring. If at time t = t1 an event of type w = 1 takes place, then,
as far as events of type w = 2 are concerned, the subject is simply censored at
t = t1 . A subject may be at risk of death from stroke or at risk from either stroke
or cirrhosis of the liver. Once one of the types of death has occurred, then the
other type of event can no longer be observed. We will assume that the subject
is censored at this point, in as much as our attention focuses on the second
type of event, and the above discussion on the different censoring models applies
in the same way. We will need to make some assumptions, most often that of
independent censoring or that of independent censoring conditional on covariate
information in order to make progress.
2.6 Competing risks
A number of models for competing risks are in use and we consider these briefly
without going into technical detail. We might view the simplest approach to this
question as one, seen from a somewhat abstract theoretical angle, where everyone
will ultimately suffer any and all events given enough time. We only observe the
first such event and all others are censored at that time. This corresponds to
the competing risks model based on latent events and with an independence
2.6. Competing risks 39
assumption (CRM-I). An alternative theoretical construct, CRM-J, requires us

to break down the event space into a partition, i.e., only one of the event types
can be observed and, for a failure, one of the event types must be observed. If
we use the random variable D to denote the type of failure then we can write
1
λk (t) = lim Pr(t < T < t + Δt, D = k|T > t), (2.18)
Δt→0+ Δt
described as the cause-specific hazard rate. This is straightforward enough,
notwithstanding the strong assumptions relating to the partition, so that, given
data in the form of risk sets and events of ttype k taking place within relatively
short intervals, we can estimate Λk (t) = 0 λ(u)du. From this we have the sur-
vivorship function, Sk (t) = exp{−Λk (t)}. The interpretation of this is not so
easy. It is certainly not accessible to non-statisticians if only for the fact that it
does not stand alone and depends crucially on what other causes of failure are
under consideration. This is immediately apparent from the expression for the sur-
vivorship function of T , the first observed failure, as, S(t) = exp {− k Λk (t)} .
The authors on this topic will often point out that the interpretation of Sk (t)
is difficult, in particular since as t increases without bound Sk (t) will not tend to
zero. While this is clear from its definition it is not our biggest problem here. Our
main problem is one of interpretation. As far as epidemiologists are concerned, the
CRM-J model will be challenging to fit since we are generally unable to estimate
the rates specified by this model using available registry data. These data take
the form of numbers of individuals at the beginning of some age interval “at risk”
of the event in question. The number of such events occurring in the interval
can be counted. While the ratio of these provides valid estimates of short-term
rates, this validity does lean on a strong untestable assumption which is that, for
those subjects lost to follow-up during any interval, had we somehow managed
to not lose them to follow-up then the subsequent rates we are trying to estimate
remain unaffected.
This model for a joint outcome (CRM-J) contrasts with CRM-I for competing
risks where everyone suffers everything ultimately. While any event will, in a
similar way, censor all others, our outcome variable is now a joint one, the time
to the event together with the type of event observed. A third model (CRM-ID)
is a competing risks model built around the illness-death (hence the ID) model
of Fix and Neyman. This model is also referred to as the semi-competing risks
model and has been used to advantage in Fine and Gray (1999), Fine et al.
(2001), Ghosh (2006) and Haneuse and Lee (2016). Lee et al. (2015) considered
inference for such models from a Bayesian viewpoint while some of the difficulties
in likelihood construction were studied in Hsieh and Huang (2012). The impact
that competing risks can have on the confounding of the main treatment effect
has recently been discussed in Huang R. and Dulai (2020).
Models CRM-I and CRM-J have their advantages and drawbacks. The model
based on a joint outcome may seem the most appealing, in some ways the more
realistic, but there are many technical issues associated with this model and
there is considerable practical difficulty in outlining an appropriate partition of
the event space. The survival probabilities of main interest are confounded with
all of the other survival probabilities and, necessarily, dependent on the chosen
partition. This can be problematic from the point of view of interpretation. It is
also very difficult in terms of finding suitable data that allow estimation of the
main quantities of interest. Competing risks models have seen greater application
in epidemiology than in clinical research and we consider this again in the context
of genetic epidemiology (Chapter 5).
For the simplest case, that of just two competing risks, say failure and censor-
ing, a quite fundamental result was derived by Tsiatis (1975). The result indicates
that the marginal distributions of failure or censoring—viewed as competing risks,
i.e., the observation on one impedes the observation on the other—cannot, under
general conditions, be estimated. These distributions are not identifiable. In order
to find consistent estimates of these distributions from observations arising under
such competing risks, it is necessary to add further restrictive conditions. The
most common condition used in this context is that of independence. Tsiatis
(1975), in a paper that is something of a cornerstone to subsequent theory,
showed that we can obtain consistent estimates using, for example, the Kaplan-
Meier estimator, under the assumption of independence.
Consider a clinical trial where patients enter the study according to a uniform
law over some time interval. This appears to be a somewhat reasonable model.
Their time to incidence of some event of interest is then recorded. At some point
the study is closed and the available data analyzed. Not only will we consider
those patients yet to experience the event to be censored but such censoring will
be taken to be independent. The assumption seems reasonable. It would break
down of course if the recruitment over the interval was subject to change, in
particular, the more robust patients being recruited early on. So, even in such a
simple case, the needed independence assumption ought to come under scrutiny.
We can encounter other situations where the independence assumption requires
something of a stretch of the imagination. For example, if patients who are
responding poorly to treatment were to be removed from the study.
Next, let’s consider a more realistic but greatly more complex situation in
which there are several competing risks. We cannot reasonably assume indepen-
dence. The occurrence of any particular outcome will prevent us from observing
all of the other outcomes. In real-life situations this is what happens of course,
when dealing with an absorbing state such as death. If a patient were to die of
heart failure then he or she will no longer be in a position to die of anything else.
In order to make progress in analyzing such data it is necessary to make a large
number of, essentially, unverifiable assumptions. In making these assumptions
we need to be guided by a basic principle in all statistical modeling, the prin-
ciple of parsimony. The situation we wish to model is, almost always, infinitely
more complex than that we can realistically hope to model accurately. Even if we
could find a plausible model deemed to describe the mechanism that generates
the observations, the lack of precision of our parameter estimates will be such
as to drown in noise any worthwhile signal we are aiming to identify. Modeling
requires something of the artist’s ability to sense when, in trying too hard, their
work can fall short of the goal in mind. The dilemma confronting the statisti-
cal modeler is not altogether different. Competing risks models are particularly
complex and great care is called for in order to remain fully confident in our
conclusions. The field is very large and several published texts are devoted solely
to the topic. We do not have the space here to do justice to this extensive body
of work apart from to make some observations on the 3 main approaches that
have been widely adopted. For those new to the field two clear overviews are
given by Putter et al. (2007) and Xu and Tai (2010). There is a large body of
literature on these topics and suggested reading would include Crowder (1994),
Tsiatis (2005), Moeschberger and Kochar (2007), Andersen and Rosthoj (2002),
Bakoyannis and Touloumi (2012) and Austin and Fine (2017). The impact of
competing risks on clinical trials is looked at in Satagopan and Auerbach (2004)
and Freidlin and Korn (2005).
In order to assess the impact of model choice on estimation we can consider
the development more formally. For CRM-I, the independence assumption for
all causes is made together with the idea of latent event times. In this setup
every individual is taken, initially, to be at risk of an event of all causes. If we
have m competing risks and times to event, T1 , T2 , ..., Tm , then we only get to
observe X = min(T1 , T2 , ..., Tm ) and all event times greater than the minimum
remain unobserved and recorded as censored at X. This structure is simple and
appealing. Cancer registries such as the SEER registries compiling data from
several millions of individuals provide information of a form that can be used by
CRM-I. Very large numbers of individuals, at given ages, are taken to be at risk
from one of the possible outcome events under study. The number of occurrences
of any particular cause over some small age interval divided by the number at
risk, either at the beginning of the interval or some average over the interval,
provide an estimate of the age-specific incidence rate for that particular cause.
This is then repeated for the next adjacent interval. Cumulative incidence follows
from just adding together the age-specific incidence rates. Using the elementary
formulas introduced in Section 2.3 we can obtain a survival probability. This
probability is not easily interpreted, leaning as it does on many near impossibly
restrictive assumptions and, more bothersome still, an idealized characterization
of reality that is so far removed from human existence as to make its direct
applicability problematic.
The implications of choosing CRM-J rather than CRM-I are highlighted by
consideration of Equation 2.5. At some time point, let’s say aj , we have that
S(aj ) = SI (aj−1 )S(aj , aj−1 ) (2.19)

where SI (t) is the survival function for the event of interest under an assumption
of independence of the competing events, i.e., SI (t) = S(t) for all t. In Equation
2.5 there was only a single competing event, the censoring itself, taken to be
independent, but let us extend that to all competing events. The left-hand side
of the equation gives the probability of surviving beyond time point aj and this
is the simple product of the probabilities of reaching time point aj−1 together
with the probability of surviving the interval (aj−1 , aj ). These probabilities are
not affected by an independent censoring mechanism, the working assumption
that we are making. So, when it comes to estimation we can use convergent
estimators of survival, such as the Kaplan-Meier or the Nelson-Aalen formulas
together with the observed rate estimated from the interval (aj−1 , aj ). From
a theoretical viewpoint any subject censored before time aj−1 remains in an
abstract sense at risk for the event under study albeit, given actual data, for this
subject, the event can not be observed. The event for this subject is considered
a latent variable, it exists but is not observable.
For the competing risk model CRM-J, this is no longer true. Once one type of
event has been observed then all other types of event are considered to not exist,
we do not have a latent variable structure. In consequence, in the above formula,
S(aj ) describes the probability of the event of interest occurring after time point
aj , and this is the product of the incidence rate for this event between aj−1 and
aj and the probability, SJ aj−1 that all events—the one of interest together with
all competing events—occur some time later than aj−1 . In this case we have:
S(aj ) = SJ (aj−1 )S(aj , aj−1 ). (2.20)
Clearly, SI (t) and SJ (t) in Equations 2.19 and 2.20 may differ so that in turn,
the cumulative incidence rates, 1 − S(aj ), depending on whether model CRM-I
or CRM-J is assumed, can also differ. For a breast cancer study, Satagopan and
Auerbach (2004) found the cumulative incidence rates based on CRM-I or CRM-
J to all but coincide while, for an example in haematologic malignancy, the two
working models led to quite different estimates.
The working assumption behind the characterization of CRM-I in terms of
latent variables is that, ultimately, if given enough time, then everyone dies from
everything. Of course this is a hypothetical construct and while of value in the
appropriate context, it has little role to play in individual prediction. Such pre-
diction can give an entirely false impression on the non-specialist. For any cause,
say, for example, breast cancer, the only reason registry data will furnish a prob-
ability that grows to anything less than one (100%) is a lack of observations,
a limitation on the overall time span. This is a somewhat abstract, theoretical
observation—its mathematical expression comes from the divergence of the har-
monic series—that does require we think carefully about how we interpret any
cumulative risks. The age-specific risks are calculated by constantly updating
the risk sets, having removed deaths from all other causes in the denominator.
Essentially, in this calculation of age-specific risk, via this constant updating, we

are not allowing a subject to fail through any cause other than that under study.
Only a severe lack of observations on highest age groups prevents the lifetime
risk—the probability of being incident or of dying from the specified cause—
from gradually approaching 100%. Lifetime risk will typically be provided over
intervals not going much beyond 75 or 80 years. This does not make it any more
meaningful and clinical decisions, such as prophylactic mastectomies, need to be
based on considerations other than such probabilities. These probabilities are near
impossible to give a clear interpretation to, and this is by experts in probability
theory. They should be considered impossible to interpret by non-specialists and
ought not to be used to advise individuals on their chances of being incident for
some disease. We return to the problems of lifetime risk in the chapter dealing
with epidemiology and, in particular, we review the assessments of breast cancer
risk and its relation to the genetic markers, BRCA1 and BRCA2.
The several difficulties associated with CRM-I, in particular the interpretation
problems have led to more emphasis being given to CRM-J. The structure of
CRM-J has attracted a lot of theoretical interest and there now exists a large
body of work that we will not be able to review here. The main advantage of
CRM-J is it is more straightforward interpretability although, even here, this
is not without problems. The main disadvantage is that there exists very little
data suitable for analysis. The many data banks held in the registries cannot be
readily analyzed via CRM-J. For whatever disease is in question, the registries are
based on large risk sets and the process of counting the incident cases over some
relatively short age interval. All other events during the interval are removed
from further assessment without having identified the nature or reason for their
removal. This fits in well with model CRM-I. The information needed to estimate
risks within the structure of CRM-J is not generally available. That said, there are
several examples in the literature where a solution is cobbled together by taking
death as a competing risk for incidence of some outcome. All other risks are
assumed independent and rates of death, obtained from life tables, can be used
as a rough approximation to the unknown survivorship function of all events.
There are many more things to worry about with competing risks models. We
open up a huge topic simply by aiming to relax independence assumptions. We
also need to give thought to the competing causes we wish to consider. There
is an almost unlimited spectrum of different partitions of the space of causes we
might wish to analyze and, for any individual cause, associated probabilities of
future transitions will depend not only on its own status, and related risk factors,
but also on the status of those other causes we wish to include in the analysis.
It all sounds very complex. It is very complex and yet the tools can be invaluable
to the investigator who does not have to rely on only one of CRM-I or CRM-J.
They are at liberty to employ them both depending on the context. They can
both throw some light on the infinitely complex array of factors that relate to
prognosis. Interpreting survival probabilities is almost impossibly difficult and,
fortunately, this is not the goal. The goal is to provide no more than some rough
and ready tools to help the investigators gain better insight into the myriad of
factors, and their approximate relative importance, that may be implicated in the
outcomes on which we are focused.
A slightly modified structure, that is well described within the framework
of the compartment models described earlier, gives rise to the so-called semi-
competing risks models. For these models, one of the competing causes is not
an absorbing state so that, having arrived in the non-absorbing state, subjects
are still free to leave this state in order to transition to the absorbing state.
Seen from the viewpoint of a compartment model it is clear that semi-competing
risks models coincide with the classical illness-death model (Fix and Neyman,
1951). For this reason we refer to the model as CRM-ID. Some of the underlying
assumptions can differ however, in particular the existence and behavior of latent
times, and these assumptions will impact the task of estimation. There is an
extensive literature on such models and the methods of this book can be adapted
with more or less effort in order to deal with inferential questions that arise in
this context. In this text our central focus is on regression and while we do not
study in depth the question of regression with competing risks models there are
several papers that do, among which Larson and Dinse (1985), Lunn and McNeil
(1995), Kim (2007), Scheike and Zhang (2008), Lau and Gange (2009), Dignam
and Kocherginsky (2012), and Haller et al. (2013).
Which of the 3 main models, CRM-I, CRM-J, and CRM-ID, we may use
for analysis will mostly be guided by the application. The different underlying
assumptions can have a more or less strong impact on inferences. The most
useful focus is regression modeling, which can involve assessing the impact of
certain risk factors after having adjusted for others, the joint impact of these risk
factors and the relative importance of different risk factors. In this, arguably the
most important, setting the model choice tends to be relatively robust. If one
model provides a hierarchy of importance of risk factors then, in the main, the
same or a very similar hierarchy obtains with other models. The main differences
due to model choice will occur when we consider absolute rather than relative
quantities such as the probability of surviving so many years without being inci-
dent of some particular disease. A great deal of circumspection is recommended
when describing absolute rather than relative assessments, especially when any
such assessments can be heavily model dependent. The author’s own limited
experiment in asking fellow statisticians—mostly not in the particular field—to
interpret the 87% chance of breast cancer risk given to some carriers of a mutated
gene was revealing. Not one person got it right—the author is no exception, at
least initially—leading to the inescapable conclusion that this number, 87%, is
almost impossibly difficult to interpret. For non-specialists, and most particularly
the carriers themselves, it is hard to understand the wisdom behind asking them
to make life-changing decisions on the basis of any such number. We return to
this in the chapter on epidemiology and, here, we point this out as a way to
underline an important fact: no less effort ought to go into the purpose of our
competing risks models than in their construction.
The main focus of this text is on regression and we do not investigate the
specifics of this in the competing risks setting. Many ideas carry though. Many
do not. An introduction to regression modeling of the sub-distribution function
for models CRM-J and CRM-ID can be found in Fine and Gray (1999) and Zhang
and Fine (2011).
1. Using the definition for λ(t) = f (t)/S(t), show that S(t) = exp{−Λ(t) and
that f (t) = λ(t) exp{−Λ(t).
2. For a Weibull variate with parameters λ and k, derive an expression for the
conditional survivorship function S(t + u, u). How does this function vary
with t for fixed u? With u for fixed t?
3. Use numerical integration to calculate the mean residual lifetime m(t)

and the mean time lived in the interval [0, t], μ(t) for the Weibull with
parameters 2 and 1.5. With parameters 2 and 0.7. Plot these as functions
of time t.
4. Consider two groups in which f (t) = λγ(λt)γ−1 exp{−(λt)γ }, i.e., each

group follows a Weibull distribution where, for the first group, λ = λ1 ,
γ = γ1 and, for the second, λ = λ2 , γ = γ2 . Under which conditions will
this situations be described by proportional hazards?
5. Undertake a numerical and graphical study of the conditional survivorship

function, S(t + u, u), for the Weibull model, the extreme value model,
the Gompertz model and the log-logistic model. What conclusions can be
drawn from this?
6. Repeat the previous class project, focusing this time on the mean residual
lifetime. Again what conclusions can be drawn from the graphs.
7. Consider a disease with three states of gravity (state 1, state 2, and state
3), the severity corresponding to the size of the number. State 4 corre-
sponds to death and is assumed to follow state 3. New treatments offer
the hope of prolonged survival. The first treatment, if it is effective, is antic-
ipated to slow down the rate of transition from state 2 to state 3. Write
down a compartmental model and a survival model, involving a treatment
indicator, for this situation. A second treatment, if effective, is anticipated
to slow down all transition rates. Write down the model for this. Write
down the relevant null and alternative hypotheses for the two situations.
8. Consider a non-degenerative disease with several states; 1, 2, . . . , counting

the occurrence of these together with a disease state indicating a pro-
gression to something more serious, e.g., benign and malignant tumors or
episodes of mild asthma with the possibility of progression to a more seri-
ous respiratory ailment. Write down possible models for this and how you
might formulate tests of hypotheses of interest under varying assumptions
on the role of the less serious states.
9. Suppose we have data; T1 , . . . , Tn , from a Weibull distribution in which the

shape parameter γ is known to be equal to 1.3. Use the delta-method to
find an estimate for the variance of the estimated median (transform to a
standard form).
10. For a proportional hazards Weibull model describe the relationship between
the respective medians.
11. Investigate the function S(t, u) for different parametric models described
in this chapter. Draw conclusions from the form of this two-dimensional
function and suggest how we might make use of these properties in order
to choose suitable parametric models when faced with actual data.
12. Consider two possible structures for a parametric proportional hazards

model;
log S(t|Z) = log{S[t|E(Z)]} exp(βZ)

log S(t|Z) = log{ES[t|Z]} exp(βZ).
How do the interpretations differ and what difficulties are likely to be

encountered in fitting either of the models?
13. Consider a clinical trial comparing two treatments in which patients enter
sequentially. Identify situations in which an assumption of an independent
censoring mechanism may seem a little shaky.
14. On the basis of a single data set, fit the exponential, the Weibull, the
Gompertz, and the log-normal models. On the basis of each model estimate
the mean survival. On the basis of each model estimate the 90th percentile.
What conclusions would you draw from this?
15. Suppose our focus of interest is on the median. Can you write down a model
directly in terms of the median. Would there be any advantage/drawback
to modeling in this way rather than modeling the hazard and then obtaining
the median via transformations of the hazard function?
16. A cure model will typically suppose two populations; one for which the
probability of an event follows some distribution and another for which
the hazard rate can be taken to be zero. The ratio of the sizes of the
populations is θ/(1 − θ). Write down such a model and indicate how we
could go about estimating the mixing parameter, θ. How might a Bayesian
tackle such a problem.
17. An alternative to a cure model is based on an approach akin to that used

in relative survival. For some reference population the hazard rate is given
by λ0 (t) whereas, for the population of interest, it is given by λ1 (t). For
t < τ, we will assume that the rate λ1 (t) is strictly greater than λ0 (t) as
a result of some risk factor, most often treatment, but that for t > τ, the
two rates are so close as to be essentially indistinguishable. Discuss the
advantages and drawbacks, in terms of estimation and interpretability, of
this approach when compared to that described in the previous question.
Chapter 3
Survival without covariates
3.1 Chapter summary
The marginal survival function is of central interest even when dealing with covari-
ates. We need to find good estimates of this function and we use those estimates
in several different contexts. Good estimates only become possible under certain
assumptions on the censoring mechanism. Attention is paid to the exponential
and piecewise exponential models, both of which are particularly transparent.
The exponential model, fully characterized by its mean, can appear over restric-
tive. However, via the probability integral transform and empirical estimates of
marginal survival, it can be used in more general situations. The piecewise expo-
nential is seen, in some sense, to lie somewhere between the simple exponential
and the empirical estimate. Particular attention is paid to empirical processes
and how the Kaplan-Meier estimator, very commonly employed in survival-type
problems, can be seen to be a natural generalization of the empirical distribution
function. In the presence of parametric assumptions, it is also straightforward to
derive suitable estimating equations. The equations for the exponential model
are very simple.
Our interest is mostly in the survival function S(t). Later we will focus on how
S(t), written as S(t|Z), depends on covariates Z. Even though such studies of
dependence are more readily structured around the hazard function λ(t|Z), the
most interpretable quantity we often would like to be able to say something about
is the survival function itself. In order to distinguish the study of the influence
of Z on S(t|Z) from the less ambitious goal of studying S(t), we refer to the
former as conditional survival and the latter as marginal survival.

49
https://doi.org/10.1007/978-3-030-33439-0_3
50 Chapter 3. Survival without covariates
Since we will almost always have in mind some subset Z from the set of
all possible covariates, and some distribution for this subset, we should remind
ourselves that, although Z has been “integrated out” of the quantity S(t), the
distribution of Z does impact S(t). Different experimental designs will gener-
ally correspond to different S(t). Marginal survival, S(t), corresponds to two
situations: (i) the subjects that are considered as i.i.d. replicates from a single
population or (ii) the subjects can be distinct, from many, and potentially an infi-
nite number of populations, each population being indexed by a value of some
covariate Z. It may also be that we have no information on the covariates Z
that might distinguish these populations. In case (ii), S(t) is an average over
these several populations, not necessarily representing any particular population
of interest in itself. It is important to appreciate that, in the absence of distribu-
tional assumptions, and the absence of observable Z, it is not possible, on the
basis of data, to distinguish case (i) from case (ii). The homogeneous case then
corresponds to either case (i) or case (ii) and it is not generally useful to speculate
on which of the cases we might be dealing with. They are not, in the absence
of observable Z, identifiable from data. We refer to S(t|Z) as the conditional
survival function given the covariate Z. This whole area, the central focus of this
work, is studied in the following chapters. First, we need to consider the simpler
case of a single homogeneous group.
3.3 Parametric models for survival functions
Let us suppose that the survival distribution can be completely specified via
some parametric model, the parameter vector being, say, θ. We take θ to be a
scalar in most cases in order to facilitate the presentation. The higher-dimensional
generalization is, in most cases, very straightforward.
Maximum likelihood estimation

The data will consist of the n i.i.d. pairs (xi , δi ) ; i = 1, . . . , n. We assume an
independent censoring mechanism. This leads to the important lemma:
Lemma 3.1. Under an independent censoring mechanism the log-likelihood

can be written as log L(θ) = ni=1 log Li (θ), where
log Li (θ) = δi log f (xi ; θ) + (1 − δi ) log S(xi ; θ) (3.1)
This covers the majority of cases in which parametric models are used.
Later, when we focus on conditional survival involving covariates Z, rather than
marginal survival, the same arguments follow through. In this latter case the
common assumption, leading to an analogous expression for the log-likelihood,
is that of conditional independence of the pair (T, C) given Z.
3.3. Parametric models for survival functions 51
Estimating equation
The maximum likelihood estimate is obtained as the value of θ, denoted θ̂, which
maximizes L(θ) over the parameter space. Such a value also maximizes log L(θ)
(by monotonicity) and, in the usual case where log L(θ) is a continuous function
of θ this value is then the solution to the estimating equation (see Appendix
D.1),
U (θ) = ∂ log L(θ)/∂θ = ∂ log Li (θ)/∂θ = 0.
i
Next, notice that at the true value of θ, denoted θ0 , we have Var{U (θ0 )} =
EU 2 (θ0 ) = EI(θ0 ) where
n
n

I(θ) = Ii (θ) = −∂ 2 log L(θ)/∂θ2 = − ∂ 2 log Li (θ)/∂θ2 .
i=1 i=1
As for likelihood in general, some care is needed in thinking about the meaning
of these expressions and the fact that the operators E(·) and Var(·) are taken
with respect to the distribution of the pairs (xi , δi ) but with θ0 fixed. The score
equation is U (θ̂) = 0 and the large sample variance is approximated by Var(θ̂) ≈
1/I(θ̂). It is usually preferable to base calculations on I(θ̂) rather than EI(θ̂),
the former being, in any event, a consistent estimate of the latter (after dividing
both sides of the equation by n). The expectation itself would be complicated to
evaluate, involving the distribution of the censoring, and unlikely, in view of the
study by Efron and Hinkley (1978) to be rewarded by more accurate inference.
Newton-Raphson iteration is set up from
θ̂j+1 = θ̂j + I(θ̂j )−1 U (θ̂j ) , j ≥ 1, (3.2)
where θ̂1 is some starting value, often zero, to the iterative cycle. The Newton-
Raphson formula arises as an immediate application of the mean value theorem
(Appendix A). The iteration is brought to a halt once we achieve some desired
level of precision.
Large sample inference can be based on any one of the three tests based
on the likelihood function; the score test, the likelihood ratio test, or the Wald
test. For the score test there is no need to estimate the unknown parameters.
Many well-established tests can be derived in this way. In exponential families,
also the so-called curved exponential families (Efron et al., 1978), such tests
reduce to contrasting some observed value to its expected value under the model.
Confidence intervals with optimal properties (Cox and Hinkley, 1979) can be
constructed from uniformly most powerful tests. For the exponential family class
of distributions the likelihood ratio forms a uniformly most powerful test and,
as such, allows us to obtain confidence intervals with optimal properties. The
other tests are asymptotically equivalent so that confidence intervals based on
the above test procedures will agree as sample size increases. Also we can use
such intervals for other quantities of interest such as the survivorship function
which depends on these unknown parameters.
Estimating the survival function

We can estimate the survival function as S(t; θ̂). If Θα provides a 100(1 − α)%
confidence region for the vector θ then we can obtain a 100(1 − α)% confidence
region for S(t; θ) in the following way. For each t let
Sα+ (t; θ̂) = sup S(t; θ) , Sα− (t; θ̂) = inf S(t; θ) , (3.3)
θ∈Θα θ∈Θα
then Sα+ (t; θ̂) and Sα− (t; θ̂) form the endpoints of the 100(1 − α)% confidence
interval for S(t; θ). Such a quantity may not be so easy to calculate in general,
simulating from Θα or subdividing the space being an effective way to approx-
imate the interval. Some situations nonetheless simplify such as the following
example, for scalar θ, based on the exponential model in which S(t; θ) is mono-
tonic in θ. For such cases it is only necessary to invert any interval for θ to obtain
an interval with the same coverage properties for S(t; θ).
Exponential survival
For this model we only need estimate a single parameter, λ which will then
determine the whole survival curve. Referring to Equation 3.1 in which, for δi = 1,
the contribution to the likelihood is, f (xi ; λ) = λ exp(−λxi ) and, for δi = 0, the
contribution is, S(xi ; λ) = exp(−λxi ). Equation 3.1 then becomes:
n

log L(λ) = k log λ − λ xj , (3.4)
j=1

where k = n N (∞). Differentiating this and equating with zero we find that
n i=1 i
λ̂ = k/ j=1 xj . Differentiating a second time we obtain I(λ) = k/λ2 . Note, by
conditioning upon the observed number of failures k, that EI(λ) = I(λ), the
observed information coinciding with the expected Fisher information, a property
of exponential families, but which we are not generally able to recover in the
presence of censoring.
An ancillary argument would nonetheless treat k as being fixed and this is
what we will do as a general principle in the presence of censoring, the observed
information providing the quantity of interest. Some discussion of this is given
by Efron and Hinkley (1978) and Barndorff-Nielsen and Cox (1994). We can now
write down an estimate of the large sample variance which, interestingly, only
depends on the number of observed failures. Thus, in order to correctly estimate
the average, it is necessary to take into account the total time on study for both
the failures and those observations that result in censoring. On the other hand,
given this estimate of the average, the precision we will associate with this only
depends on the observed number of failures. This is an important observation
and will be made again in the more general stochastic process framework.
Multivariate setting
In the majority of applications, the parameter θ will be a vector of dimension p.
The notation becomes heavier but otherwise everything is pretty much the same.
The estimating equation, U (θ) = 0 then corresponds to a system of p estimating
equations and I(θ) is a p × p symmetric matrix in which the (q, r) th element
is given by −∂ 2 log L(θ)/∂θq ∂θr where θq and θr are elements of the vector θ.
Also, the system of Newton-Raphson iteration can be applied to each one of the
components of θ so that we base our calculations on solving the set of equations:
θ̂j+1,q = θ̂j,q + I(θ̂j )−1 U (θ̂j ) , q = 1, . . . , p ; j ≥ 1, (3.5)
where, in this case, θ̂1 is a vector of starting values to the iterative cycle, again
most often zero.
Example 3.1. For the Freireich data (Cox, 1972), we calculate for the 6-MP
group λ̂ = 9/359 = 0.025. For the placebo group we obtain λ̂ = 21/182 = 0.115.
Furthermore in the 6-MP group we have Var (λ̂) = 9/(359)2 = 0.000070 whereas
for the placebo group we have Var (λ̂) = 21/(182)2 = 0.0006.
The non-parametric empirical estimate (described below) agrees well with
curves based on the exponential model and this is illustrated in Figure 3.1. Infer-
1.0
0.8
Survival
0.6
0.4
0.2
0.0
0 5 10 15 20 25 30 35
Time
Figure 3.1: Exponential fitted curves to two-group Freireich data.

ence can be based on an appeal to the usual large sample theory. In this particular
case, however, we can proceed in a direct way by recognizing that, for the case of

no censoring k = n, the sample size and n j=1 Tj is a sum of n independent ran-
dom variables each exponential with parameter λ. We can therefore treat n/λ̂
as a gamma variate with parameters (λ, n). When there is censoring, in view
of the consistency of λ̂, we can take k/λ̂ as a gamma variate with parameters
(λ, k), when k < n. This is not an exact result, since it hinges on a large sample
approximation, but it may provide greater accuracy than the large sample normal
approximation.
In order to be able to use standard tables we can multiply each term of the
sum by 2λ since this then produces a sum of n exponential variates, each with
variance 2. Such a distribution is a gamma (2, n), equivalent to a chi-square
distribution with 2n degrees of freedom Evans et al. (2001). Taking the range of
values of 2kλ/λ̂ to be between χα/2 and χ1−α/2 gives a 100(1 − α)% confidence
interval for λ. For the Freireich data we find 95% CI = (0.0115, 0.0439). Once
we have intervals for λ, we immediately have intervals with the same coverage
properties for the survivorship function, this being a monotonic function of λ.
Denoting the upper and lower limits of the 100(1 − α)% confidence interval
Sα+ (t; λ̂) and Sα− (t; λ̂) respectively, we have:

+
λ̂χα/2 −
λ̂χ1−α/2
Sα (t; λ̂) = exp − t , Sα (t; λ̂) = exp − t .
2k 2k
An alternative approximation, based on large sample normality, and also very

simple in form, can be written as:

+ λ̂ √
Sα (t; λ̂) ≈ exp − √ ( k − z1−α/2 )t , (3.6)
k
where, this time, the corresponding expression for Sα− (t; λ̂) obtains by replacing
z1−α/2 by zα/2 .
Piecewise exponential model

We can achieve considerably greater flexibility than the standard exponential
model by constructing a partition of the time axis 0 = a0 < a1 < . . . < ak =
∞. Within the jth interval (aj−1 , aj ) , (j = 1, . . . , k) the hazard function
is given by λ(t) = λj . Using Equations 2.12 and (3.1) and equating first
derivatives to zero we find

λ̂j = kj / {(x − aj−1 )I(x < aj ) + (aj − aj−1 )I(x ≥ aj )}.
:x >aj−1
Differentiating a second time we find I(λ1 , ..., λk ) to be block diagonal where

the (j, j) th element is given by Ijj = kj /λ2j . In view of the orthogonality of the
parameter estimates, we can construct a 100(1 − α)% simultaneous confidence
interval for λ1 , . . . , λk by choosing a sequence of αj such that 1− j (1−αj ) = α.
An estimate of the survivorship function derives from Λ̂(t) in which we

define Ŝ(t; Λ̂) = exp{−Λ̂(t)} and where

Λ̂(t) = λ̂j (aj − aj−1 ) + λ̂ (t − a−1 )I(a−1 ≤ t < a ). (3.7)
j:aj ≤t
Confidence intervals for this function can be based on Equation 3.3. We can view
the simple exponential survival model as being at one extreme of the parametric
spectrum, leaning as it does on a single parameter, the mean. It turns out that
we can view the piecewise exponential model, with a division so fine that only
single failures occur in any interval, as being at the other end of the parametric
spectrum, i.e., a non-parametric estimate. Such an estimate corresponds to that
obtained from the empirical distribution function. This is discussed below.
Other parametric models

The exponential and piecewise exponential models hold a special place in the
survival literature for a number of reasons. The models are simple in form, simple
to understand, have a clear physical property that we can interpret (lack of
memory property) and the parameters can be estimated so easily that analysis
based on such models clearly falls under the heading of desirable procedures as
defined by Student; ones that can be calculated on the back of a railway ticket
while awaiting the train. The piecewise model also allows quite considerable
flexibility.
Given these facts, it is hard to justify the use of other parametric models unless
motivated by some compelling physical argument. Cox (1958) used the Weibull
model in analyzing the strengths of wool, but the model was not just pulled
out of the air. Physical considerations and the fact that the Weibull distribution
obtains as the limiting case of the minimum of a collection of random variables
provide a powerful case in its favor. Another physical case might be the sum of
exponential variates, each having the same parameter, since this can be seen to
be a gamma variate. Generally, we may have no good reason to believe some
model over most others is likely to provide a good fit for data. In these cases,
given the generally good performance of empirical estimates, it may be preferable
to base our analyses on these.
3.4 Empirical estimate (no censoring)
Empirical or non-parametric estimates, i.e., those making no model assumptions,

can be most readily understood in the context of some finite division of the time
axis. Theoretical results stem from continuous-time results, the transition from
the discrete to the continuous, as a consequence of sample size n increasing
without bound, presenting no conceptual difficulty. Unlike the discussion on the
piecewise exponential model in which we most often anticipate having very few
intervals, rarely more than three or four—the idea being to pool resources (esti-
mating power) per interval while keeping within the bounds of serious model
violation—for empirical estimates we do the opposite.
We imagine a fine division of the real line in which, given real data, the vast
majority of the intervals are likely to contain no observations. Consider the time
interval to be fixed, divided equally into k non- overlapping intervals (aj−1 , aj ],
j = 1, . . . , k, the notation “(” indicating that the interval is open on the left, and
“]” closed on the right, i.e., aj−1 does not belong to the interval but aj does. We
have that Pr(T ≥ aj ) = Pr(T > aj ) = S(aj ) and that Pr(T > aj |T > aj−1 ) =
S(aj , aj−1 ). Recall from Section 2.3 that
S(aj ) = S(aj−1 )S(aj , aj−1 ) = S(a , a−1 ).

≤j
For each t = a , ( > 0), the empirical estimate of S(t) based on a sample of size
n, and denoted Sn (t), is simply the observed number of observations that are
greater than t. For a random sample of observations Ti , (i = 1, . . . , n) we use the
indicator variable I(·) to describe whether or not the subject i survives beyond
point t, i.e., for t = aj ,
n j
1
Sn (aj ) = I(Ti > aj ) = Sn (a , a−1 ) (3.8)
n
i=1 =1
in which the empirical Sn (a , a−1 ) is defined in an entirely analogous way to

Sn (a ), i.e.,
n
n

1
Sn (a , a−1 ) = I(Ti > a ) , n = I(Ti ≥ a ). (3.9)
n−1
i=1 i=1
It is readily seen, and instructive for understanding the Kaplan-Meier estimate of

the next section, that, if no failure is observed in (a−1 , a ] then Sn (a , a−1 ) = 1,
whereas for an observed failure, Sn (a , a−1 ) = (n−1 − 1)/n−1 . The empirical
distribution has been well studied and, in particular, we have:
3.4. Empirical estimate (no censoring) 57
Lemma 3.2. For any fixed value of t, Sn (t) is asymptotically normal with
mean and variance given by:
E{Sn (t)} = S(t) ; Var {Sn (t)} = F (t)S(t)/n. (3.10)
We should understand the operators E and Var to refer to expectations over a

set of repetitions, the number of repetitions increasing without bound, and with
n and t fixed. As an estimator of S(t), the function Sn (t) is then unbiased. The
variance, as we might anticipate, is the variance of a binomial variate with param-
eters n and S(t). These two moments, together with a normal approximation of
DeMoivre-Laplace (see Appendix C.2), enable the calculation of approximate
confidence intervals. The result, which is well known (Glivenko-Cantelli), enables
us to carry out inference for S(t) on the basis of Sn (t) for any given point t.
Very often we will be interested in the whole function Sn (t), over all values of
t and, in this case, it is more helpful to adopt the view of Sn (t) as a stochastic
process. In this regard we have
√
Theorem 3.1. n{Sn (t) − S(t)} is a Gaussian process with mean zero and
covariance given by
√ √
Cov [ n{Sn (s)}, n{Sn (t)}] = F (s){1 − F (t)}. (3.11)
An important consequence of the theorem arises when the distribution of T is

uniform, for, in this case, F (s) = s and all the conditions are met for the process
√
n{Fn (t) − t} (0 ≤ t ≤ 1) to converge in distribution to the Brownian bridge.
Finally these results apply more generally than just to the uniform case for,
as long as T has a continuous distribution, there exists a unique monotonic
transformation from T to the uniform, such a transformation not impacting
√
n{Fn (t) − F (t)} itself. In particular this enables us to use the result of
Appendix B.2 to make inference for an arbitrary continuous cumulative
√
distribution function, F (t), whereby, for Wn (t) = n{Fn (t) − F (t)},
∞

Pr sup |Wn (t)| ≤ D → 1 − 2 (−1)k+1 exp(−2k 2 D2 ) , D ≥ 0.
t
k=1
(3.12)
Most often we are interested in events occurring with small probability in which
case a good approximation obtains by only taking the first term of the sum, i.e.,
√
k = 1. Under this approximation | n{Fn (t) − F (t)}| will be greater than about
1.4 less than 5% of the time. This is a simple and effective working rule.
3.5 Kaplan-Meier (empirical estimate with censoring)
The impact of censoring is to reduce the number of observations T1 , . . . , Tn so

that some, a number strictly less than n, are actually observed. The empirical
estimate of the standard distribution function is no longer available to us so that
n
1
F̂emp (t) = I(Ti ≤ t), 0≤t≤T,
n
i=1
cannot be calculated. Simply taking the observed censored times as though they
were failure times will clearly not work so that the estimator;
n n
1 1
1− Yi (t) = 1Xi ≤ t , 0≤t≤T
n n
i=1 i=1
will exhibit increasing bias as the amount of censoring increases. This follows
since we will be estimating P (T ≤ t, C ≤ t) and this underestimates P (T ≤ t).
A little more work is needed although it is only really a question of treating the
various needed ingredients in a sensible way for things to work. The most famous
empirical estimate is that of Kaplan and Meier (1958) which we can show to be
consistent under an independent censoring mechanism.
Definition 3.1. The Kaplan-Meier estimator of F , denoted F̂ , is written:

δi
F̂ (t) = 1 − 1 − n , 0≤t≤T.
j=1 Yj (Xi )
i = 1, . . . , n
Xi ≤t
This function is piecewise continuous. At the jump points of discontinuity, for

elementary theoretical reasons, it is helpful to define it as being right continuous
with a left-hand limit (CADLAG, which is a French translation of that definition).
When there is no censoring, for a sample of size n, these jumps all have the same
size 1/n. In the presence of censoring these jumps will become larger as we move
through time. In some sense, the observations lost to censoring, redistribute a part
of their probability mass, over the observations strictly greater than themselves.
This idea can be made precise and we do so below. The Kaplan-Meier estimate
is more general than the usual empirical estimate Fn in that, in the absence
of censoring, the former will reduce to the latter. Like Fn (t), the Kaplan-Meier
estimator is consistent for F (t) although this requires the additional condition
of independence between the censoring and failure mechanisms. Also, we write
Ŝ(t) = 1 − F̂ (t). Calculation is simple, and like the usual empirical estimate, can
be done manually. To see this, consider the 7 observations in Table 3.1. The 3
observations, 7, 13, and 22 are censored times. The other observations are failure
times. When t ∈ [Xi , Xi+1 [ then Ŝ(t) = Ŝ(Xi ). The censored observations play a
3.5. Kaplan-Meier (empirical estimate with censoring) 59
X 2 4 7 9 13 16 22
δ 1 1 0 1 0 1 0
Table 3.1: Taken from Chauvel (2014) as an illustration in Kaplan-Meier calcu-

lation
0.857
0.714
Sˆ ( T|Z )
0.536
0.268
0 2 4 7 9 13 16 22
T
Figure 3.2: Estimated Kaplan–Meier curve
n n
Xi j=1
Yj (Xi ) 1 − δi / j=1
Yj (Xi ) Ŝ(Xi )
2 7 1-1/7 1-1/7 0.857
4 6 1-1/6 (1-1/7)(1-1/6) 0.714
7 5 1 0.714
9 4 1-1/4 (1-1/7)(1-1/6)(1-1/4) 0.536
13 3 1 0.536
16 2 1-1/2 (1-1/7)(1-1/6)(1-1/4)(1-1/2)0.268
22 1 1 0.268
Table 3.2: Chauvel (2014) illustrates the step-by-step Kaplan-Meier calculations
role for values lower than themselves. The Kaplan-Meier curve does not change
at the censored observation time and, beyond this, the role of the censored
observation is indirect. It does not influence the actual calculation. Figure 3.2
shows the estimated Kaplan–Meier curve, Ŝ(t). Unlike the empirical function,
1 − Fn (t), the curve does not reach zero in this example and this will be the
case whenever the last observation is a censored observation. The jumps in the
Kaplan-Meier curve play an important role in proportional and non-proportional
hazards modeling. The size of these jumps at time t, written dŜ(t), are readily
described and we have (Table 3.2):
Proposition 3.1. For t an observed event time and where Ŝ(t− ) = lim Ŝ(s),
s→t−
we can write,
Ŝ(t− )
dŜ(t) = Ŝ(t) − Ŝ(t− ) = n , (3.13)
j=1 Yj (t)
In later chapters we will see how to make full use of these increments. They can
be viewed as weights or as approximations to infinitesimal contributions to an
integral with respect to the survival function. These are sometimes referred to as
Kaplan-Meier integrals and, in the light of Helly-Bray’s theorem (Appendix A),
allow us to consistently estimate functions of T of interest in the presence of cen-
soring. Details can be found in Stute (1995) where, in particular, under general
conditions, we can obtain the asymptotic normality of these integrals. Another
important example in which we use standard software to fit a proportional haz-
ards model to data generated under a non-proportional hazards mechanism has
been studied by Xu and O’Quigley (2000) and O’Quigley (2008). The use of the
Kaplan-Meier increments enables us to consistently estimate an average regres-
sion effect Eβ(T ) when β(t) is not a constant. Failing to correctly incorporate
the Kaplan-Meier increments into the estimating equation—a common oversight
encouraged by the availability of a lot of standard software—will lead to serious
bias in the estimation of Eβ(T ).
Approach using finite censoring support

First, recall the model for finite censoring support, described in Section 2.5.
Such a model adds no additional practical restriction but helps us to avoid much
mathematical complexity, leading to a simple and useful theorem (Theorem 3.2).
The theorem enables us to circumvent the difficulty arising from the fact that

we are unable to obtain n i=1 I(Ti ≥ aj ) as in Equation 3.8. This is because
at aj we cannot ascertain whether or not earlier observed values of Xi (δi = 0),
arising from the pair (Ti , Ci ) in which Ci = Xi , are such that Ti > aj . Since
Xi = min(Ti , Ci ) we do not observe the actual value of Ti when Xi = Ci . The
trick is to notice in the right- hand side of (3.8) that we are able to obtain the
empirical estimates, Gn (a , a−1 ) for G(a , a−1 ) = Pr(X ≥ a |X > a−1 ). But,
unlike Equation 3.8
n
1
Gn (aj ) = I(Xi ≥ aj ) = Gn (a , a−1 ). (3.14)
n
i=1 ≤j
This is because of the non-zero masses being associated to the times at which the
censorings occur. Nonetheless,
n the rest follows through readily, although, unlike
(3.9), we now define n = i=1 I(Xi ≥ a ), noting that this definition contains
(3.9) as a special case when there is no censoring.
Kaplan-Meier estimate based on finite censoring support

The distribution of X = min(T, C) is mostly of indirect interest but turns out to
be important in view of the following theorem and corollary.
Theorem 3.2. Under an independent censoring mechanism,
G(a , a−1 ) = S(a , a−1 ). (3.15)
Corollary 3.1. Under an independent censoring mechanism, a consistent estima-

tor of S(t) at t = aj is given by Ŝ(t) = ≤j Gn (a , a−1 ).
Corollary 3.2. In the absence of censoring Ŝ(t) = Sn (t).
The estimator of the corollary is known as the Kaplan-Meier estimator. In

the light of our particular setup, notably the use of the independent censoring
mechanism having finite support (Section 2.5), it can be argued that our estimate
is only available at distinct values of t = aj . But this is not a practical restriction
since our finite division can be as fine as we wish, the restriction amounting to
limiting accuracy to some given number of decimal places. One way or another it
is not possible to completely avoid some mathematical fine points, our preference
being to use the mechanism of Section 2.5.
In practice we talk about the Kaplan-Meier estimate at time point t, the
underlying support of C and T rarely being a central concern. The one exception
to this arises when the range of C is less than that for T . A single sampled Ti
which is greater than the greatest value that can be taken by Ci will necessarily
appear as a censored observation. In this case the observed empirical Kaplan-
Meier estimate is never able to reach zero. The true distribution F (t) cannot be
estimated at values greater than the upper limit for the support of C. This is
intuitively clear. Again, in practice, this ought not to be a real concern. If we have
no information on something (in this case a certain upper region of the survival
curve), it is most likely too optimistic to hope to be able to carry out useful
estimation there, unless, of course, we make additional parametric assumptions,
which amount to saying that information obtained over some range can tell us
a lot about what is taking place elsewhere, and certainly more than just the
provision of bounds on F (t) for t greater than the upper limit of C.
Remarks on the Kaplan-Meier curve

The Kaplan-Meier (1958) product-limit estimator provides a non-parametric esti-
mate for the survival function S(t) under the independent censorship assumption.
It is a generalization of the commonly known empirical distribution function to the
case of censoring since, in the absence of censoring, the two coincide. Expressing
the estimate Ŝ(t) in terms of the fine division (a , a−1 ), = 1, . . . , N, is concep-
tually useful. However, since intervals not containing events produce no change
in Ŝ(t) it is, for the purposes of practical calculation, only necessary to consider
evaluation at the distinct observed failure times t1 < t2 < · · · < tk . All divisions
of the time interval (a , a−1 ), = 1, . . . , N, for different N , lead to the same
estimate Ŝ(t), provided that the set of observed failure points is contained within
the set {a ; = 1, . . . , N }. A minimal division of the time axis arises by taking
the set {a } to be the same as the set of the distinct observed failure times. So,
for practical purposes aj = tj , j = 1, . . . , k. and the Kaplan-Meier estimate can
be defined as
nj − dj dj
Ŝ(t) = = 1− . (3.16)
nj nj
j:tj <t j:tj <t

where dj = δj , the sum being over the ties at time tj . If there are no ties at
tj then dj = δj . Note that Ŝ(t) is a left-continuous step function that equals
1 at t = 0 and drops immediately after each failure time tj . The estimate does
not change at censoring times. When a censoring time and a failure time tj
are recorded as equal, the convention is that censoring times are adjusted an
infinitesimal amount to the right so that the censoring time is considered to
be infinitesimally larger that tj . Any subjects censored at time tj are therefore
included in the risk set of size nj , as are those that fail at tj . This convention is
sensible because a subject censored at time tj almost certainly survives beyond
tj . Note also that when the last observation is a censoring time rather than
a failure time, the KM estimate is taken as being defined only up to this last
observation.
Continuous version of Kaplan-Meier curve

The Kaplan-Meier curve is so useful and so commonly employed that we reserve
the most standard notation for this estimate, i.e., Ŝ(t). For the majority of
applications this is all we need. However, there are some situations in which it
is useful to be able to invert the function in order to estimate S −1 (t). Now, the
usual Kaplan-Meier estimate takes a constant value between adjacent failures,
tj−1 and tj and so, in general, we are unable to invert this function in a unique
way. This is not a serious concern and any intuitive solution such as taking Ŝ(t)
to be S(tj−1 ) or S(tj ) for all t ∈ (tj−1 , tj ) would be fine. Rather than this
we suggest a simple linear interpolation. We then define the continuous (and
invertible since strictly decreasing) function S̄(t) to be given by
t − tj−1
S̄(t) = Ŝ(tj−1 ) + Ŝ(tj ) − Ŝ(tj−1 ) ; t ∈ (tj−1 , tj ). (3.17)
tj − tj−1
Note that at the distinct failure times tj the two estimates, Ŝ(t) and S̄(t) coin-
cide. An example of where we make an appeal to S̄(t) is illustrated in the two-
group exponential model where a transformation to exponentiality for one group

is then applied to the other group.
Precision of Kaplan-Meier estimate

Keeping in mind the results of Appendices A.2 and A.10 we can derive a Taylor
series approximation to a function of random variables wherever the function
of expectation converges in probability to the expectation of the function. An
immediate application leads to the following theorem.
Theorem 3.3. For each t = a , the estimate Ŝ(t) is asymptotically normal

with mean S(t) and variance:
dm
Var S(a ) ≈ S(a )2 . (3.18)
nm (nm − dm )
m≤ m≤
The above expression for the variance of Ŝ(t) is known as Greenwood’s formula.
Breslow and Crowley (1974) in a detailed large sample study of the Kaplan-
Meier estimator obtained a result asymptotically equivalent to the Greenwood
formula, making a slight correction to overestimation of the variation in the
estimated survival probability. The formula’s simplicity, however, made it the
most commonly used when computing the variance of the Kaplan-Meier estimate
and related quantities. We also have:
Corollary 3.3. When there is no censoring, the approximation Var Ŝ(t)

from Theorem 3.3 reduces to the usual binomial variance estimate
Ŝ(t){1 − Ŝ(t)}/n.
The usual use to which we put such variance estimates is in obtaining approxi-
mate confidence intervals. Thus, using the large sample normality of Ŝ(t), adding
and subtracting to this z1−α/2 (the 1 − α/2 quantile from the standard normal
distribution) multiplied by the square root of the variance estimate, provides
approximate 100(1 − α)% confidence intervals for Ŝ(t). As mentioned before the
constraints on Ŝ(t), lying between 0 and 1, will impact the operating characteris-
tics of such intervals, in particular, it may not be realistic, unless sample sizes are
large, to limit attention to symmetric intervals around Ŝ(t). Borgan and Liestol
(1990) investigate some potential transformations, especially the log-minus-log
transformation discussed in Section 2.4, leading to
Corollary 3.4. Let w(α) = Var1/2 Ŝ(t)z1−α/2 /Ŝ(t) log Ŝ(t). For each t = a , a
100(1 − α)% confidence intervals for Ŝ(t) can be approximated by
{Ŝ(t)exp −w(α) , Ŝ(t)exp w(α) } (3.19)

The same arguments which led to Greenwood’s formula also lead to approxi-
mate variance expressions for alternative transformations of the survivorship func-
tion. In particular we have
Corollary 3.5. For each t = a the estimate log Ŝ(t) is asymptotically normal
with asymptotic mean log S(t) and variance
dm
Var log S(a ) ≈ (3.20)
nm (nm − dm )
m≤ m≤
Corollary 3.6. For each t = a the estimate log Ŝ(t)/{1 − Ŝ(t)} is asymp-
totically normal with asymptotic mean log S(t)/{1 − S(t)} and variance
S(a ) dm
Var log ≈ {1 − S(a )}−2 . (3.21)
1 − S(a ) nm (nm − dm )
m≤ m≤
Confidence intervals calculated using any of the above results will be of help
in practice. Following some point estimate, obtained from Ŝ(t) at some given
t, these intervals are useful enough to quantify the statistical precision that we
wish to associate with the estimate. All of the variance estimates involve a com-
parable degree of complexity of calculation so that choice is to some extent a
question of taste. Nonetheless, intervals based on the log-minus-log or the logit
transformation will behave better for smaller samples, and guarantee that the
endpoints of the intervals themselves stay within the interval (0,1). This is not
so for the Greenwood formula, the main argument in its favor being that it has
been around the longest and is the most well known. For moderate to large sam-
ple sizes, and for Ŝ(t) not too close to 0 or 1, all the intervals will, for practical
purposes, coincide.
Observed differences between Kaplan-Meier curves

A common and fundamental goal in survival studies is to make a statistical
decision as to whether the observed differences between two Kaplan-Meier curves
can be attributed to sampling differences from the same common population
or whether the observed differences are indicative of two distinct populations.
We consider formal testing in later chapters and here we make a few general
observations. The most common test used here is the so-called log-rank test.
It can be shown to be the most powerful test against local departures that can
be described as having a proportional hazards nature (Peto and Peto, 1972), in
other words the log-log transformations of the two survival curves is a constant
that does not change with time. If the hazard ratio is not constant over time,
the log-rank test can suffer significant power losses (Leurgans, 1983, 1984), so
much so that the test can fail to detect differences that are clearly indicative of
real effects in many cases (Figure 3.3).
Figure 3.3: Kaplan-Meier curves obtained from the Curie breast cancer study.
A proportional hazards assumption suggests itself as a possibility to model the
observed differences. A good candidate test would be the log-rank test.
The right-hand figure in Figure 3.4 represents Kaplan-Meier estimators of the

survival of 2 groups of patients with breast cancer followed at the Institut Curie
for a period of 12 years. For the first group, the diameter of the tumor of the
patients is less than 60mm and for the other group the diameter is greater than
60mm. There were a total of 339 patients in the study. The differential survival
experience of the two groups is quite clear from the figure. At the same time, a
rough and ready glance suggests that these differences are not likely to be well
described by a proportional hazards model. During the first 60 months, the
advantage seen by one group is clear and the curves move further and further
apart. The effect appears to be a strong one. However, beyond this period,
the slopes in the curves are much more similar and suggest a very considerable
attenuation of the initial effect. Some 80% of patients with tumor size less than
60mm survive at least 75 months. This duration drops all the way down to 28
months for patients with a tumor size greater than 60mm. There would seem
little doubt as to the prognostic impact of tumor size and yet the classical log-
rank test fails to achieve significance in this example. Of course this could always
be due to the fact that we cannot rule out with great enough conviction the
possibility of having observed a large but non-significant difference. Although, it
seems more likely that there is a real, and clinically important difference here,
but that the lack of proportionality of risks clouds the issue and puts a spoke in
the wheel of the usual log-rank test. A simple clinical explanation might be that
tumor size is of considerable importance in the short term but that, after having
managed to survive beyond some point, this importance is much diminished. A
more likely clinical explanation is that tumor size maintains its importance but
that it is measured with some degree of error. A number of patients would have
Figure 3.4: Kaplan-Meier curves obtained from a randomized clinical trial and
a study in breast cancer. In both cases, a proportional hazards assumption is
doubtful. The log-rank test would show poor performance in such situations.
been misclassified and, as we will see in later theoretical work, the consequences
of this would be to produce an impression of a diminishing regression effect.
Figure 3.4 also illustrates two Kaplan-Meier curves taken from a randomized
clinical trial in lung cancer. There are three groups and the short-term effects and
long term effects appear to indicate no real treatment effect. Initially, the curves
more or less coincide, as they do in the long term. However, for a significant part
of the study, say between the median and the 10th percentile, there appears to
be a real advantage in favor of the two active treatment groups. The log-rank
test fails to detect this and more suitable tests, described in later chapters, are
able to confirm the treatment differences. The clear lack of proportionality of
hazards is at the root of the problem here.
Striving to obtain tests that to a greater or lesser extent can reverse the power
deficit arising as a result of non-proportional hazards has stimulated the work of
many authors in this field. Weighting the linear contributions to the log-rank
test, leading to the so-called weighted log-rank tests, has a long history (Fleming
and Harrington, 1991; Gehan, 1965; Peto and Peto, 1972). Prentice (1978) and
others showed that a weighting that mirrors the actual form of the departure from
proportional hazards will lead to powerful alternative tests. What is trickier, but
arguably much more relevant, is to obtain tests that will have good performance
in both situations, that will be close to optimal under proportional hazards, if not
quite optimal, but that will retain good power in situations where the alternative
is significantly remote from that of proportional hazards. We consider this in later
chapters.
Redistribution to the right algorithm

Another way to look at the KM estimate, which turns out to be conceptually
useful and of value to later developments, is to focus on the increments, or step
size, of the function at points where it changes. If we denote tj + the time instant
immediately after tj , then Ŝ(tj ) − Ŝ(tj +) is the stepsize, or jump, of the KM
curve at time tj . From Equation 3.16 we see that
nj − dj
Ŝ(tj +) = Ŝ(tj ) · ,
nj
so the stepsize is Ŝ(tj )·dj /nj . That is to say, when the total “leftover" probability

mass is Ŝ(tj ), each observed failure gets one-nj th of it, where nj = n i=1 Yi (tj )
is the number of subjects at risk at time tj . In the absence of censoring this
corresponds exactly to the way in which the empirical estimate behaves. When
there is censoring, then one way of looking at a censored observation is to consider
that the mass that would have been associated with it is simply reallocated to all
of those observations still remaining in the risk set (hence the term “redistribution
to the right.”)
Kaplan-Meier estimates of median and mean survival

Since the support of T is only on the positive real line the distributions we
deal with are asymmetrical. In consequence it is more common to take as sum-
mary measures simple functions of different quantiles, most often the median or
interquartile range, rather than the mean and variance. Nonetheless, the mean
is of interest and in special cases, such as the exponential distribution, relates
directly to the median via a constant scaling factor. Also, of course, there is no
compelling reason not to work with symmetric distributions defined for say log T
and then transform back to T , although this is not very commonly done. From
Section 2.3 we can estimate the expected lifetime over some given interval [0, t]
as
t
μ̂(t) = Ŝ(u)du, (3.22)
0
the mean itself being then estimated by μ̂(∞). However, the theory comes a
little unstuck here since, not only must we restrict the time scale to be within
the range determined by the largest observation, the empirical distribution itself
will not correspond to a probability distribution whenever F̂ (t) = 1 − Ŝ(t) fails to
reach one. In practice then it makes more sense to consider mean life time μ̂(t)
over intervals [0, t], acknowledging that t needs to be kept within the range of
our observations. The following result provides the required inference for μ̂(t);
Lemma 3.3. For large samples, μ̂(t) can be approximated by a normal distribu-
tion with E μ̂(t) = μ(t) and
{μ̂(t) − μ̂(am )}2 dm
Var μ̂(t) ≈ . (3.23)
nm (nm − dm )
m≤ m≤
In view of the consistency of the Kaplan-Meier estimate it is, in principle,

straightforward to obtain estimates for any desired quantile, or function of the
quantiles, such as the median or interquartile range. However, again, we can be
limited by the observations and, if Ŝ(t), for the largest observed survival time,
is not less than p then we are not able to obtain point estimates of quantiles
corresponding to such values, although we could obtain interval estimates. We
define the pth quantile ξp to satisfy F (ξp ) = p. The Kaplan-Meier function, or
indeed the uncensored empirical distribution function itself, being a step function,
is not invertible. To overcome this, we define the estimate ξˆp to be the smallest
value of t such that F̂ (ξp ) ≥ p. In order to make inferences for ξˆp we can appeal
to some basic results from Appendix A.9 to obtain:
Lemma 3.4. For large samples, ξˆp approaches a normal distribution with E ξˆp =
ξp and
dm (1 − p2 )f −2 (ξˆp )
Var ξˆp ≈ . (3.24)
nm (nm − dm )
m≤ m≤
It is difficult to use the above result in practice in view of the presence of the
density f (·) in the expression. Smoothing techniques and the methods of density
estimation can be used to make progress here but our recommendation would be
to use a more direct, albeit more heavy, approach constructing intervals based
on sequences of hypothesis tests. In principle at least the programming of these
is straightforward.
3.6 Nelson-Aalen estimate of survival
An alternative approach to estimating the empirical survival distribution, adapted

to accommodate censoring, and, essentially, equivalent to the Kaplan-Meier esti-
mate arises from considerations of the basic t formulae in Section 2.3. Recalling
that S(t) = exp{−Λ(t)} and that Λ(t) = 0 λ(u)du we can consider empirical
estimates of Λ(t). The integral can be approximated by a Riemann sum so that
for t = aj
aj
λ(a )(a − a−1 ) → λ(u)du (3.25)
≤j 0
as a − a−1 goes to zero. Now, applying a local linearization, we obtain
λ(a )(a − a−1 ) ≈ P (a < T < a−1 |T > a−1 ) = 1 − S(a , a−1 ).
Applying Theorem 3.2 and then, first replacing S(a , a−1 ) by G(a , a−1 ), sec-
ond replacing G(a , a−1 ) by Gn (a , a−1 ) i.e., d /n , we obtain, at t = aj ,

Λ̃(aj ) = j=1 d /n as a consistent estimator for Λ(t). The resulting estimator
3.7. Model verification using empirical estimate 69
S̃(t) = exp{−Λ̃(t)} (3.26)
is called the Nelson-Aalen estimate of survival. Recalling the Taylor series expan-
sion, exp(x) = 1 + x + x2 /2! + · · · , for small values of x we have exp(−x) =
1 − x + O(x2 ), the error of the approximation being strictly less than x2 /2 since
the series is convergent with alternating sign. Applying this approximation to
S̃(t) we recover the Kaplan-Meier estimate described above. In fact we can use
this idea to obtain:
Lemma 3.5. Under the Breslow-Crowley conditions, |S̃(t) − Ŝ(t)| converges
almost surely to zero.
In view of the lemma, large sample results for the Nelson-Aalen estimate
can be deduced from those already obtained for the Kaplan-Meier estimate.
This is the main reason that there is relatively little study of the Nelson-Aalen
estimate in its own right. We can exploit the wealth of results for the Kaplan-
Meier estimate that are already available to us. Indeed, in most practical finite
sample applications, the level of agreement is also very high and the use of
one estimator rather than the other is really more a question of taste than any
theoretical advantage. In some ways the Nelson-Aalen estimate appears very
natural in the survival setting, and it would be nice to see it used more in
practice.
3.7 Model verification using empirical estimate
A natural approach to model assessment, i.e., whether or not some parametric

model appears as a reasonable choice for the observed data, is to contrast the
empirical estimates to those leaning on the model assumptions, the role of the
data being reduced to that of providing estimates for any unknown parameters.
We do not propose tackling the broad issues of goodness of fit until later but
notice that, if the assumed parametric form is reasonable, then the model-based
estimates and the empirical estimates ought to broadly agree. For the Weibull
model, for example, we know that S(t) = exp{−(λt)p }. Therefore, a plot, from
the Kaplan-Meier, or Nelson-Aalen estimate, of log{− log Ŝ(t)} against log t
should be linear with an intercept equal to p log λ and slope equal to p. For
all the other parametric models it is usually possible to derive similar construc-
tions. A visual impression of the adequacy of any postulated model is obtained,
alongside the possibility of obtaining simple parameter estimates for the unknown
parameters. Such estimates are typically less efficient than maximum likelihood
estimates so, in a more thorough analysis, we may wish to use them either as
a rough guide or as a first approximation in some iterative scheme. In practice
it is often the case that quite different parametric assumptions, unless partic-
ularly restrictive like that for the exponential model, will produce very similar
survival curves. Important differences between competing parametrizations tend
to manifest themselves mostly in the tails of the distribution where there may be
few observations. As a goodness of fit tool these procedures are not usually very
powerful.
Empirical estimates and exponential analysis

Referring to Appendix A.6 and Theorem A.8, we have the important result that,
for any continuous positive random variable T , with distribution function F (t),
T
the variate Λ(T ) = 0 f (u)/[1−F (u)]du has a standard exponential distribution.
As a consequence, if we consider the empirical survivorship function, Ŝ(t), then
we can take the observations − log Ŝ(Xi ) as arising from a standard exponential
distribution. All of the simple results that are available to us when data are
generated by an exponential distribution can be used. In particular, if we wish
to compare the means of two distributions, both subject to censoring, then we
can transform one of them to standard exponential via its empirical survival
function, then use this same transformation on the other group. The simple
results for contrasting two censored exponential samples can then be applied
even though, at least initially, the data arose from samples generated by some
other mechanism.
1. Write down the estimating equations for a Weibull model based on maxi-
mum likelihood. Write down the estimating equations for a Weibull model
based on the mean and variance.
2. Consider the following observations; 1, 3, 4, 4, 5, 6∗ , 6, 7, 9∗ , 16 where a ∗

indicates a censored observation. Fit a Weibull model to these observations
based on (i) maximum likelihood, (ii) method of moments.
3. For the data of the previous question, calculate and plot the survivorship
function. Calculate an approximate 90% confidence interval for S(4). Do
the same for S(7).
4. Describe how you might calculate a simultaneous 90% confidence interval

for S(4) and S(7) together.
5. Compare the variance expression for S(7) with that approximated by the
binomial formula based on Ŝ(7) and 8 failure times.
6. Take 100 bootstrap samples, fit the Weibull model to each one separately
and estimate S(4) and S(7). Calculate empirical variances based on the
100 sample estimates. How do these compare with those calculated on the
basis of large sample theory.
7. In the previous question, rather than fit the model, we could calculate

empirical estimates of S(4) and S(7). Describe possible difficulties with
this approach.
8. For small samples generated via an exponential distribution with unknown

mean, discuss the relative merits of the different possible confidence interval
approximations for the survival function. Describe a study you might set
up in order to make a recommendation regarding the “best” confidence
interval to work with. How do you understand “best” in this context.
9. Simulate 100 uncensored observations from a standard exponential distri-

bution. Calculate the empirical survival function Sn (t). Choose two points,
s and t and repeat the whole process 100 times obtaining 100 pairs of
S100 (s) and S100 (t). Calculate the empirical covariance between S100 (s)
and S100 (t). Use different values of s and t to suggest the validity of
Theorem 3.1.
√
10. Explain why the result of Equation 3.12 for Wn (t) = n{Fn (t) − F (t)},
when F (t) is uniform continues to hold for any other continuous distribu-
tion.
11. Explain the importance of Theorem 3.2 and how it is used in order to
obtain consistent estimates of survival in the presence of an independent
censoring mechanism.
12. Simulate 200 uncensored observations from the log-normal distribution.

Suppose that we had been led to believe that the observations had been
generated from a Weibull law. Carry out graphical procedures to challenge
the validity of the assumption. Repeat the exercise under the supposition
that the observations had been generated via a log-logistic model.
13. For the 200 observations of the previous question, introduce an independent
censoring mechanism so that approximately half of the observations are
censored. Calculate the logarithm of the Kaplan-Meier and Nelson-Aalen
estimates and plot one against the other. Fit a least squares line to the
plot and comment on the values of the slope.
14. Use the results of Lemma 3.3 to show that μ̂(t) is consistent for μ(t).
15. Show that when there is no censoring, the Greenwood estimate of the vari-
ance of the Kaplan-Meier estimate reduces to the usual variance estimate
for the empirical distribution function. Conclude from this that confidence
intervals based on the Greenwood estimate of variance are only valid at
a single given time point, t, and would not provide bounds for the whole
Kaplan-Meier curve.
16. Carry out a study on the coverage properties based on Ŝ(t), log Ŝ(t) and
log Ŝ(t)/{1 − Ŝ(t)}. Describe what you anticipate to be the relative merits
of the different functions.
17. Malani (1995) outlined an operationally simple approach for estimating

survival in the presence of a dependent censoring mechanism. The method
requires that the dependency be captured via some explanatory variable.
Appealing to the redistribution to the right algorithm, each censored obser-
vation has its remaining mass redistributed. However, unlike the simple ver-
sion of the algorithm, Malani proposes to only redistribute among subjects
in the risk set sharing the same covariate value as the subject censored at
that point. Give an intuitive explanation as to why this would work.
18. Following the idea of Malani, suppose, in the presence of dependent cen-
soring, we obtained a Nelson-Aalen estimate of survival for each level of the
covariate. Subsequently, appealing to the law of total probability, we esti-
mate marginal survival by a linear combination of these several estimates.
Comment on such an estimate and contrast it with that of Malani.
19. Recall that the uncensored Kaplan-Meier estimator, i.e., the usual empirical
estimate, is unbiased. This is no longer generally so for the Kaplan-Meier
estimate. Can you construct a situation in which the estimate of Equation
3.17 would exhibit less bias than the Kaplan-Meier estimate?
20. Using data from a cancer registry, show how you could make use of the
piecewise exponential model to obtain conditional survival estimates for
S(T > t + s|T > s).
3.9 Outline of proofs
Lemma 3.1 and Theorem 3.2: There are two possibilities; the observation xi
corresponds to a failure, the observation corresponds to a censoring time. For
the first possibility let dxi be an infinitesimally small interval around this point.
The probability that we can associate with this event is Pr(T ∈ dxi , C > xi ) =
Pr(T ∈ dxi ) × pr(C > xi ) i.e. f (xi ; θ)dxi × G(xi ; θ). For a censored observation
at time xi (δi = 0) we have Pr(C ∈ dxi , T > xi ) = Pr(C ∈ dxi ) × pr(T > xi ) i.e.,
g(xi ; θ)dxi × S(xi ; θ). We can then write the likelihood as
L(θ) = f (xi ; θ)dxi G(xi ; θ) × g(xi ; θ)dxi S(xi ; θ)

δi =1 δi =0
In most cases the further assumption that the censoring itself does not depend
on θ may be reasonable. Taking logs and ignoring constants we obtain the result.
For Theorem 3.2 note that
3.9. Outline of proofs 73
Pr{Xi ≥ a |Xi > a−1 } = Pr{Ti ≥ a |Xi > a−1 } × Pr{Ci ≥ a |Xi > a−1 }
= Pr{Ti ≥ a |Ci > a−1 , Ti > a−1 } = Pr{Ti ≥ a |Ti > a−1 }
by independence.
Theorem 3.3: We have Var (log S(a ) ≈ m≤ Var (log(1 − πm )). This is
−2 Var (π ) which we write as
−2 π (1 − π )/n
m≤ (1−πm ) m m≤ (1−πm ) m m m
where
nm corresponds to the number at risk. i.e. the denominator. We write
m≤ dm /rm (rm − dm ) Use log function and a further application of delta
method to obtain:

Var S(a ) ≈ S(a )2 dm /rm (rm − dm ).
m≤ m≤
An approach not using the delta method follows Greenwood (1926). Let t0 <
t1 < · · · < tk . An estimate of the variance of the survival probability P of the form
P = p1 × p2 × .... × pk , where each pi is the estimated probability of survival from
time ti−1 to time ti , qi = 1 − pi and P is therefore the estimated probability of
survival from t0 to tk . Assuming that the pi ’s are independent of one another, we
have E(P ) = E(p1 ) × E(p2 ) × .... × E(pk ), as well as, E(P 2 ) = E(p21 ) × E(p22 ) ×
.... × E(p2k ), and E(p2i ) = (Epi )2 + σi2 , where σi2 = Var(pi ). Then

Var(P ) = E(P 2 ) − {E(P )}2 = {E(P )}2 ( 1 + σ12 /(Ep1 )2 1 + σ22 /(Ep2 )2 · · ·

1 + σk2 /(Epk )2 − 1) ≈ {E(P )}2 (σ12 /(Ep1 )2 + σ22 /(Ep2 )2 + .... + σk2 /(Epk )2 ).
Now condition on E(pi ) and the number of observations ni to which pi is applied,

then σi2 = Epi (1 − Epi )/ni . Substituting pi for E(pi ), we obtain estimates of
the variance
k k

ˆ
Var(P ) = P 2{ (1 + qi /ni pi ) − 1} ≈ P 2 qi /ni pi .
i=1 i=1
Proposition 3.1: Take l ∈ {1, . . . , n}. If subject l fails at time t, then we have
that Xl = t and δl = 1, so that denoting ΔŜ(t) = Ŝ(t) − Ŝ(t− ). then
δi δi
ΔŜ(t) = 1 − n − 1 − n =
j=1
Yj (Xi ) j=1
Yj (Xi )
i = 1, . . . , n i = 1, . . . , n
Xi ≤t Xi ≤t−
δi δl Ŝ(t− )δl Ŝ(t− )

1 − n 1 − n −1 = n = n .
j=1
Yj (Xi ) j=1
Yj (Xl ) j=1
Yj (t) j=1
Yj (t)
i = 1, . . . , n
Xi ≤t−
Chapter 4
Proportional hazards models
4.1 Chapter summary
We consider several models that describe survival in the presence of observ-

able covariates, these covariates measuring subject heterogeneity. The most gen-
eral situation can be described by a model with a parameter of high, possibly
unbounded, dimension. We refer to this as the general or non-proportional haz-
ards model since dependence is expressed via a parameter, β(t), that is not
constrained or restricted. Proportional hazards models have the same form but
constrain β(t) to be a constant. We write the constant as β, sometimes β0 , since
it does not change with time. When the covariate itself is constant, the depen-
dence structure corresponds to the Cox regression model. We describe this model,
its connection to the well-known log-rank test, and its use in many applications.
We recall the founding paper of Cox (Cox, 1972) and the many discussions that
surrounded that paper. Some of the historical backgrounds that lay behind Cox’s
proposal is also recalled in order to for the new reader to quickly appreciate that,
brilliant though Professor Cox’s insights were, they leant on more than just his
imagination. They did not emerge from a vacuum. Some discussion on how the
model should be used in practice is given.
The presence of subject heterogeneity, summarized by risk factors Z, known or

suspected of being related to S(t), is our central concern. The previous chapter
dealt with the issue of marginal survival, i.e., survival ignoring any indicator of
heterogeneity and which treats the data in hand as though the observations
came from a single population. In Figure 4.1 there are two groups. This can be
described by two distinct Kaplan-Meier curves or, possibly, two independently
75
https://doi.org/10.1007/978-3-030-33439-0_4
76 Chapter 4. Proportional hazards models
calculated fitted parametric curves. If, however, the curves are related, then each
estimate provides information not only about its own population curve but also
about the other group’s population curve. The curve estimates would not be inde-
pendent. Exploiting such dependence can lead to considerable gains in our esti-
mating power. The agreement between an approach modeling dependence and
1.0
0.8
survival function
0.6
0.4
0.2
0 50 100 150 200

survival time(months)
Figure 4.1: Kaplan-Meier survival curves and PH model curves for two groups
defined by a binary covariate. Dashed lines represent PH estimates. The fit is
adequate up to 150 months, after which the fit becomes progressively poorer.
one ignoring it can be more or less strong and, in Figure 4.1, agreement is
good apart from observations beyond 150 months where a proportional hazards
assumption may not hold very well. Returning to the simplest case, we can imag-
ine a compartmental model describing the occurrence of deaths independently
of group status in which all individuals are assumed to have the same hazard
rates. As pointed out in the previous chapter, the main interest then is in the
survival function S(t) when the Z are either unobservable or being ignored. Here
we study the conditional survival function given the covariates Z and we write
this as S(t|Z). In the more complex situations (multicompartment models, time-
dependent Z) it may be difficult, or even impossible, to given an interpretation
to S(t) as an average over conditional distributions, but the idea of condition-
ing is still central although we may not take it beyond that of the probability
of a change of state conditional upon the current state as well as the relevant
covariate history which led to being in that state.
The goal here is to consider models with varying degrees of flexibility applied
to the summary of n subjects each with an associated covariate vector Z of
dimension p. The most flexible models will be able to fully describe any data at
hand but, as a price for their flexibility, little reduction in dimension from the
n × p data matrix we begin with. Such models will have small bias in prediction
compared with large sampling errors. The most rigid models can allow for striking
reductions in dimension. Their consequent impact on prediction will be associated
with much smaller sampling errors. However, as a price for such gains, the biases
4.3. General or non-proportional hazards model 77
in prediction can be large. The models we finally work with will lie between these
two extremes. Their choice then depends on an artful balance between the two
conflicting characteristics. A central task, guided by the principle of parsimony, is
to use as few parameters as possible to achieve whatever purpose we have in mind.
4.3 General or non-proportional hazards model
In the most straightforward cases we can express the conditional dependence

of survival upon fixed covariates in terms of the hazard function. A general
expression for the hazard function, given the value of the covariate Z is:
λ(t|Z) = λ0 (t) exp{β(t)Z}, (4.1)
where λ(t|·) is the conditional hazard function, λ0 (t) the baseline hazard
corresponding to Z = 0, and β(t) a time-varying regression effect. Whenever
Z has dimension greater than one we view β(t)Z as an inner product in which
β(t) has the same dimension as Z so that β(t)Z = β1 (t)Z1 +, · · · , +βp (t)Zp .
Recalling the discussion of Chapter 3, we are mostly interested in situations

where observations on Z can be made in the course of any study. In Equation
4.1 Z is not allowed to depend upon time. If we also disallow the possibility of
continuous covariates, which, in practice, we can approximate as accurately as
we wish via high dimensional Z together with β(t) of the same dimension, we
see that model (4.1) is completely general and, as such, not really a model. It is
instead a representation, or re-expression, of a very general reality, an expression
that is convenient and which provides a framework to understanding many of
the models described in this chapter. At the cost of losing the interpretation of
a hazard function, we can immediately generalize (4.1) to
λ(t|Z) = λ0 (t) exp{β(t)Z(t)}. (4.2)
As long as we do not view Z(t) as random, i.e., the whole time path of Z(t) is
known at t = 0, then a hazard function interpretation for λ(t|Z) is maintained.
Otherwise we lose the hazard function interpretation, since this requires knowl-
edge of the whole function at the origin t = 0, i.e., the function is a deterministic
and not a random one. In some ways this loss is of importance in that the equiv-
alence of the hazard function, the survival function, and the density function
means that we can easily move from one to another. However, when Z(t) is ran-
dom, we can reason in terms of intensity functions and compartmental models,
a structure that enables us to deal with a wide variety of applied problems such
as clinical trials using cross-over designs, studies in HIV that account for var-
ied accumulated treatment histories and involved epidemiological investigations
in which exposure history over time can be complex. The parameter β(t) is of
infinite dimension and therefore the model would not be useful without some
restrictions upon β(t).
4.4 Proportional hazards model
Corresponding to the truth or reality under scrutiny, we can view Equation 4.2
as being an extreme point on a large scale which calibrates model complexity.
The opposite extreme point on this scale might have been the simple exponential
model, although we will start with a restriction that is less extreme, specifically
the proportional hazards model in which β(t) = β so that;
λ(t|Z) = λ0 (t) exp{βZ(t)}. (4.3)
Putting restrictions on β(t) can be done in many ways, and the whole art of sta-
tistical modeling, not only for survival data, is in the search for useful restrictions
upon the parameterization of the problem in hand. Our interpretation of the word
“useful” depends very much on the given particular context. Just where different
models find themselves on the infinite scale between Equation 4.3 and Equation
4.2 and how they can be ordered is a very important concept we need to master
if we are to be successful at the modeling process, a process which amounts to
feeling our way up this scale (relaxing constraints) or down this scale (adding
constraints), guided by the various techniques at our disposal. From the outset it
is important to understand that the goal is not one of establishing some unknown
hidden truth. We already have this, expressed via the model described in Equation
4.1. The goal is to find a much smaller, more restrictive model, which, for practical
purposes is close enough or which is good enough to address those questions that
we have in mind; for example, deciding whether or not there is an effect of treat-
ment on survival once we have accounted for known prognostic factors which may
not be equally distributed across the groups we are comparing. For such purposes,
no model to date has seen more use than the Cox regression model (Figure 4.2).
4.5 Cox regression model
In tackling the problem of subject heterogeneity, the Cox model has enjoyed
outstanding success, a success, it could be claimed, matching that of classic
multilinear regression itself. The model has given rise to considerable theoretical
work and continues to provoke methodological advances. Research and develop-
ment into the model and the model’s offspring have become so extensive that
we cannot here hope to cover the whole field, even at the time of writing. We
aim nonetheless to highlight what seems to be the essential ideas and we begin
with a recollection of the seminal paper of D.R. Cox, presented at a meeting of
the Royal Statistical Society in London, England, March 8, 1972.
4.5. Cox regression model 79
Figure 4.2: An illustration of survival curves and associated hazard functions for
a proportional hazards model.
Regression models and life tables (D.R. Cox 1972)

After summarizing earlier work on the life table (Chang, 1968; Kaplan and Meier,
1958), Professor Cox introduced his, now famous, model postulating a simplified
form for the relationship between the hazard function λ(t), at time t and the value
of an associated fixed covariate Z. As its name suggests, the proportional hazards
model assumes that the hazard functions among subjects with different covariates
are proportional to one another. The hazard function can then be written:
λ(t|Z) = λ0 (t) exp{βZ}, (4.4)
where λ0 (t) is a fixed “baseline” hazard function, and β is a relative risk param-
eter to be estimated. Whenever Z = 0 has a concrete interpretation (which we
can always obtain by re-coding) then so does the baseline hazard λ0 (t) since, in
this case, λ(t|Z = 0) = λ0 (t). As mentioned just above, when Z is a vector of
covariates, then the model is the same, although with the product of vectors βZ
interpreted as an inner product. It is common to replace the expression βZ by
β Z or β T Z where β and Z are p × 1 vectors, and a b, or aT b denote the inner
product of vectors a and b. Usually, though, we will not distinguish notationally
between the two situations since the former is just a special case of the latter.
We write them both as βZ. Again we can interpret λ0 (t) as being the hazard
corresponding to the group for which the vector Z is identically zero.
The model is described as a multiplicative model, i.e., a model in which factors
related to the survival time have a multiplicative effect on the hazard function.
An illustration in which two binary variables are used to summarize the effects
of four groups is shown in Figure 4.3. As pointed out by Cox, the function (βZ)
can be replaced by any function of β and Z, the positivity of exp(·) guaranteeing
that, for any hazard function λ0 (t), and any Z, we can always maintain a hazard
function interpretation for λ(t|Z). Indeed it is not necessary to restrict ourselves
h(t)
Figure 4.3: Proportional hazards with two binary covariates indicating 4 groups.
Log-hazard rate written as h(t) = log λ(t).
to exp(·), and we may wish to work with other functions R(·), although care is
required to ensure that R(·) remains positive over the range of values of β and Z
of interest. Figure 4.3 represents the case of two binary covariables indicating four
distinct groups (in the figure we take the logarithm of λ(t)) and the important
thing to observe is that the distance between any two groups on this particular
scale, i.e., in terms of the log-hazards, does not change through time. In view of
the relation between the hazard function and the survival function, there is an
equivalent form of Equation 4.4 in terms of the survival function. Defining S0 (t)
to be the baseline survival function; that is, the survival function corresponding
to S(t|Z = 0), then, for scalar or vector Z, we have that
S(t|Z) = {S0 (t)}exp(βZ) . (4.5)
When the covariate is a single binary variable indicating, for example, treatment
groups, the model simply says that the survival function of one group is a power
transformation of the other, thereby making an important connection to the class
of Lehmann alternatives (Lehmann et al., 1953).
Cox took the view that “parametrization of the dependence on Z is required so
that our conclusions about that dependence are expressed concisely”, adding that
any choice “needs examination in the light of the data”. “So far as secondary fea-
tures of the system are concerned ... it is sensible to make a minimum of assump-
tions.” This view led to focusing on inference that allowed λ0 (t) to remain arbi-
trary. The resulting procedures are nonparametric with respect to t in that infer-
ence is invariant to any increasing monotonic transformation of t, but parametric
in as much as concerns Z. For this reason the model is often referred to as Cox’s
semi-parametric model. Let’s keep in mind, however, that it is the adopted infer-
ential procedures that are semi-parametric rather than the model itself. Although,
of course, use of the term λ0 (t) in the model, in which λ0 (t) is not specified,
implies the use of procedures that will work for all allowable functions λ0 (t).
Having recalled to the reader how inference could be carried out following
some added assumptions on λ0 (t), the most common assumptions being that
λ0 (t) is constant, that λ0 (t) is a piecewise constant function, or that λ0 (t)
is equal to tγ for some γ, Cox presented his innovatory likelihood expression
for inference, an expression that subsequently became known as a partial likeli-
hood (Cox, 1975). We look more closely at these inferential questions in later
chapters. First note that the quantity λ0 (t) does not appear in the likelihood
expression given by
n
δi
exp(βZi )
L(β) = n , (4.6)
i=1 j=1 j (Xi ) exp(βZj )
Y
and, in consequence, λ0 (t) can remain arbitrary. Secondly, note that each term in
the product is the conditional probability that at time Xi of an observed failure,
it is precisely individual i who is selected to fail, given all the individuals at risk
and given that one failure would occur. Taking the logarithm in Equation 4.6
and its derivative with respect to β, we obtain the estimating equation which,
upon setting equal to zero, can generally be solved without difficulty using the
Newton-Raphson method, to obtain the maximum partial likelihood estimate β̂
of β. We will discuss more deeply the function U (β) under the various approaches
to inference. We can see already that it has the same form as that encountered
in the standard linear regression situation where the observations are contrasted
to some kind of weighted mean. The exact nature of this mean is described later.
Also, even though the expression
n
n
j=1 Yj (Xi )Zj exp(βZj )
U (β) = δi Zi − n (4.7)
i=1 j=1 Yj (Xi ) exp(βZj )
looks slightly involved, we might hope that the discrepancies between the Zi and
the weighted mean, clearly some kind of residual, would be uncorrelated, at least
for large samples, since the Zi themselves are uncorrelated.
All of this turns out to be so and makes it relatively easy to carry out appropri-
ate inference. The simplest and most common approach to inference is to treat
β̂ as asymptotically normally distributed with mean β and large sample vari-
ance I(β̂)−1 , where I(β), called the information in view of its connection to the
likelihood, is the second derivative of − log L(β) with respect to β, i.e., letting
n n 2
2
j=1 Yj (Xi )Zj exp(βZj ) j=1 Yj (Xi )Zj exp(βZj )
Ii (β) = n − n , (4.8)
j=1 Yj (Xi ) exp(βZj ) j=1 Yj (Xi ) exp(βZj )

then I(β) = n i=1 δi Ii (β). Inferences can also be based on likelihood ratio meth-
ods. A third possibility, which is sometimes convenient, is to base tests on the
score U (β), which in large samples can be considered to be normally distributed
with mean zero and variance I(β). Multivariate extensions are completely natu-
ral, with the score being a vector and I an information matrix.
Early applications of the model

The first success of the model was in its use for the two-sample problem, i.e.,
testing the null hypothesis of no difference in the underlying true survival
curves
for two groups. In this case Cox showed that the test statistic U (0)/ I(0) is
formally identical to a test, later known under the heading of the log-rank test,
obtained by setting up at each failure point a 2 × 2 contingency table, group
against failed/survived, and combining the many 2 × 2 tables. As in a standard
analysis of a single such contingency table we use the marginal frequencies to
obtain estimates of expected rates under the null hypothesis of no effect. Assum-
ing, as we usually do here, no ties we can obtain a table such as described in
Table 4.1 in which, at time t = Xi the observed failure occurs in group A and
there are nA (t) and nB (t) individuals at risk in the respective groups.
Time point t = Xi Group A Group B Totals

Number of failures 1 0 1
Number not failing nA (t) − 1 nB (t) nA (t) + nB (t) − 1
Total at risk nA (t) nB (t) nA (t) + nB (t)
Table 4.1: 2 × 2 table at failure point t = Xi for group A and group B.
The observed rates and the expected rates are simply summed across the
distinct failure points, each of which gives rise to its own contingency table
where the margins are obtained from the available risk sets at that time. From
the above, if Zi = 1 when subject i is in group A and zero otherwise, then
elementary calculation gives that
n
n

U (0) = δi {Zi − π(Xi )} , I(0) = δi π(Xi ){1 − π(Xi )}
i=1 i=1
where π(t) = nA (t)/{nA (t) + nB (t)}. The statistic U then contrasts the
observations with their expectations under the null hypothesis of no effect. This
expectation is simply the probability of choosing, from the subjects at risk, a
subject from group A. The variance expression is the well-known expression for
a Bernoulli variable. Readers interested in a deeper insight into this test should
also consult (Cochran, 1954; Mantel, 1963; Mantel and Haenszel, 1959; Peto
and Peto, 1972). As pointed out by Cox, “whereas the test in the contingency
table situation is, at least in principle, exact, the test here is only asymptotic ...”
This statement is not fully precise since there is still an appeal to the DeMoivre-
Laplace approximation. Nonetheless, we can understand his point.
However, the real advantage of Cox’s approach was that while contributing
significantly toward a deeper understanding of the log-rank and related tests, it
opened up the way for more involved situations; additional covariates, continuous
covariates, random effects, and, perhaps surprisingly, in view of the attribute
“proportional hazards”, a way to tackle problems involving time-varying effects
or time-dependent covariates. Cox illustrated his model via an application to the
now famous Freireich data (Acute Leukemia Group B et al., 1963) describing a
clinical trial in leukemia in which a new treatment was compared to a placebo.
Treating the two groups independently and estimating either survivorship function
using a Kaplan-Meier curve gave good agreement with the survivorship estimates
derived from the Cox model. Such a result can also, of course, be anticipated
by taking a log(− log) transform of the Kaplan-Meier estimates and noting that
they relate to one another via a simple shift. This shift exhibits only the weakest,
if any, dependence on time itself.
Figure 4.4: Kaplan-Meier curves and model-based curves for Freireich data.
Dashed lines represent Model-based estimates; exponential model (left), Cox
model (right).
Recovering the usual two-group log-rank statistic as a special case of a test
based on model (4.4) is reassuring. In fact, exactly the same approach extends
to the several group comparison (Breslow, 1972). More importantly, Equation
4.4 provides the framework for considering the multivariate problem from its
many angles; global comparisons of course but also more involved conditional
comparisons in which certain effects are controlled for while others are tested.
We look at this in more detail below under the heading “Modeling multivariate
problems”. The partially proportional hazards model (in particular the stratified
model) was to appear later to Cox’s original work of 1972 and provide great
flexibility in addressing regression problems in a multivariate context. The early

applications of the model provided some added theoretical structure to procedures
already in use. It was not long before statisticians came to realize that the model
would allow us to go well beyond those procedures and the discussion to Cox’s
paper already anticipated some of these later developments.
Discussion of Professor Cox’s paper

Kalbfleisch and Prentice took issue with Cox’s naming of the likelihood used
for inference as a “conditional” likelihood. They pointed out that the likelihood
expression is not obtainable as a quantity proportional to a probability after
having conditioned on some event. Conditioning was indeed taking place in the
construction of the likelihood expression but in a sequential manner, a dynamic
updating whose inferential home would later be seen to lie more naturally within
the context of stochastic processes, indexed by time, rather than regular likeli-
hoods, whether marginal or conditional.
The years following this discussion gave rise to a number of papers inves-
tigating the nature of the “conditional” likelihood proposed in Cox’s original
paper. Given the striking success of the model, together with the suggested
likelihood expression, in reproducing and taking further a wide range of statis-
tics then in use, most researchers agreed that Cox’s proposal was correct. They
remained uncertain, though, as to how to justify the likelihood itself. This think-
ing culminated in several major contributions; those of Cox (1975), Prentice and
Kalbfleisch (1979), Aalen (1978) and Andersen and Gill (1982), firmly establish-
ing the likelihood expression of Cox.
It turned out that Cox was correct, not just on the appropriateness of his pro-
posed likelihood expression but also in describing it as a “conditional” likelihood,
this description being the source of all the debate. Although Cox’s likelihood
derivation may not have been conditional, in the sense of taking as observed
some single statistic upon which we condition before proceeding, his likelihood is
not only very much a conditional one but also it conditions in just the right way.
Not in the most straightforward sense whereby all the conditioning is done in one
go, but in the sense of sequentially conditioning through time. Cox’s “conditional”
likelihood is now called a “partial” likelihood although, as an inferential tool in its
own right, i.e., as a tool for inference independent of the choice of any particular
model the partial likelihood turned out not to be a particularly useful concept.
Professor Downton of the University of Birmingham and Professor Peto of
the University of Oxford pointed out the connection to rank test procedures.
Although the formulation of Cox allowed the user to investigate more complex
structures, many existing setups, framed in terms of tests based on the ranks,
could be obtained directly from the use of the Cox likelihood. The simplest exam-
ple was the sign test for the median. Using permutation arguments, other tests
of interest in the multivariate setting could be obtained, in particular tests anal-
ogous to the Friedman test and the Kruskal-Wallis test. Richard Peto referred
to some of his own work with Julian Peto. Their work demonstrated the asymp-
totic efficiency of the log-rank test and that, for the two-group problem and for
Lehmann alternatives, this test was locally most powerful. Since the log-rank
test coincides with a score test based on Cox’s likelihood, Peto argued that Cox’s
method necessarily inherits the same properties.
Professor Bartholomew of the University of Kent considered a lognormal
model in current use and postulated its extension to the regression situation by
writing down the likelihood. Such an analysis, being fully parametric, represents
an alternative approach since the structure is not nested in a proportional hazards
one. Bartholomew made an insightful observation that allowing for some depen-
dence of the explanatory variable Z on t can enable the lognormal model and
a proportional hazards model to better approximate each other. This is indeed
true and allows for a whole development of a class of non-proportional hazards
models where Z is a function of time and within which the proportional hazards
model arises as a special case.
Professors Oakes and Breslow discussed the equivalence between a saturated
piecewise exponential model and the proportional hazards model. By a satu-
rated piecewise exponential model we mean one allowing for constant hazard
rates between adjacent failures. The model is data dependent in that it does not
specify in advance time regions of constant hazard but will allow these to be
determined by the observed failures. From an inferential standpoint, in particular
making use of likelihood theory, we may expect to run into some difficulties.
This is because the number of parameters of the model (number of constant
hazard rates) increases at the same rate as the effective sample size (number of
observed failure times). However, the approach does nonetheless work, although
justification requires the use of techniques other than standard likelihood. A sim-
ple estimate of the hazard rate, the cumulative hazard rate, and the survivorship
function are then available. When β = 0 the estimate of the cumulative hazard
rate coincides with that of Nelson (1969).
Professor Lindley of University College London writes down the full likelihood
which involves λ0 (t) and points out that, since terms involving λ0 (t) do not factor
out we cannot justify Cox’s conditional likelihood. If we take λ0 (t) as an unknown
nuisance parameter having some prior distribution, then we can integrate the full
likelihood with respect to this in order to obtain a marginal likelihood (this would
be different to the marginal likelihood of ranks studied later by Kalbfleisch and
Prentice (1973). Lindley argues that the impact of censoring is greater for the
Cox likelihood than for this likelihood which is then to be preferred. The author
of this text confesses to not fully understanding Lindley’s argument and there
is some slight confusion there since, either due to a typo or to a subtlety that
escapes me, Lindley calls the Cox likelihood a “marginal likelihood” and what I
am referring to as a marginal likelihood, an “integrated likelihood”. We do, of
course, integrate a full likelihood to obtain a marginal likelihood, but it seems
as though Professor Lindley was making other, finer, distinctions which are best
understood by those in the Bayesian school. His concern on the impact of cen-
soring is echoed by Mr. P. Glassborow of British Rail underlining the strength

behind the independent censoring assumption, an assumption which would not
be reasonable in many practical cases.
Professor Zelen, a pioneer in the area of regression analysis of survival data,
pointed out important relationships in tests of regression effect in the propor-
tional hazards model and tests of homogeneity of the odds ratio in the study
of several contingency tables. Dr. John Gart of the National Cancer Institute
also underlined parallels between contingency table analysis and Cox regression.
These ideas were to be developed extensively in later papers by Ross Prentice
and Norman Breslow in which the focus switched from classical survival anal-
ysis to studies in epidemiology. The connection to epidemiological applications
was already alluded to in the discussion of the Cox paper by Drs. Meshalkin
and Kagan of the World Health Organization. Finally, algorithms for carrying
out an analysis based on the Cox model became quickly available thanks to two
further important contributions to the discussion of Cox’s paper. Richard Peto
obtained accurate approximations to the likelihood in the presence of ties, obviat-
ing the need for computationally intensive permutation algorithms, and Susannah
Howard showed how to program efficiently by exploiting the nested property of
the risk sets in reversed time.
Historical background to Cox’s paper

Alternative hypotheses to a null which assumes that two probabilities are equal,
such as in Equation 4.5, taking the form of a simple power transformation, have
a long history in statistical modeling. Such alternatives which, in the special
case where the probabilities in question are survival functions, are known as
Lehmann alternatives (Lehmann et al., 1953). Lehmann alternatives are natural
in that, under the restriction that the power term is positive, always achievable by
reparameterizing the power term to be of an exponential form; then, whatever
the actual parameter estimates, the resulting probability estimates satisfy the
laws of probability. In particular, they remain in the interval (0,1).
Linear expressions for probabilities are less natural although, at least prior to
the discovery of the logistic and Cox models, possibly more familiar. Feigl and
Zelen (1965) postulated a linear regression for the location parameter, λ0 , of an
exponential law. In this case the location parameter and the (constant) hazard
coincide so that the model could be written:
λ(t|Z) = λ0 exp{βZ}. (4.9)
In Feigl and Zelen their model was not written exactly this way, expressed as
λ = α + βZ. However, since λ is constant, the two expressions are equivalent
and highlight the link to Cox’s more general formulation. Feigl and Zelen only
considered the case of uncensored data. Zippin and Armitage (1966) used a
modeling approach, essentially the same as that of Feigl and Zelen, although
4.6. Modeling multivariate problems 87
allowing for the possibility of censoring. This was achieved by an assumption

of independence between the censoring mechanism and the failure mechanism
enabling an expression for the full likelihood to be obtained. Further discussion on
these ideas can be found in Myers et al. (1973) and Brown (1975). The estimates
of the survival function for the different groups in the Freireich study, based on
a simple exponential model or a Cox model, are shown in Figure 4.4. For these
data the level of agreement between the two approaches appears to be high.
This early work on the exponential model certainly helped anticipate the more
general development of Cox and, for many more straightforward comparisons,
such as the one illustrated by the Freireich data, it is perhaps unfortunate that
the exponential model has been relegated to a historical role alone and is rarely,
if ever, used in current practical analysis of similar data. Estimation for the
exponential model is particularly simple and can be carried out without the use
of a computer. This would often make the model a good starting point before
building more complex models. However, modern statistical analysis will tend to
jump straight into the problem backed up by powerful programs and as a result,
there is much less call for making use of methods that can be calculated using
no more than a pencil and paper.
4.6 Modeling multivariate problems
The strength of the Cox model lies in its ability to describe and characterize
involved multivariate situations. Crucial issues concern the adequacy of fit of
the model, how to make predictions based on the model, and how strong is the
model’s predictive capability. These are considered in detail later. Here, in the
following sections and in the chapter on inference we consider how the model
can be used as a tool to formulate questions of interest to us in the multivari-
ate setting. The simplest case is that of a single binary covariate Z taking the
values zero and one. The zero might indicate a group of patients undergoing a
standard therapy, whereas the group for which Z = 1 could be undergoing some
experimental therapy. Model 4.4 then indicates the hazard rate for the standard
group to be λ0 (t) and for the experimental group to be λ0 (t) exp(β). Testing
whether or not the new therapy has any effect on survival translates as testing
the hypothesis H0 : β = 0. If β is less than zero then the hazard rate for the
experimental therapy is less than that for the standard therapy at all times and
is such that the arithmetic difference between the respective logarithms of the
hazards is of magnitude β. Suppose the problem is slightly more complex and we
have two new experimental therapies. We can write:
λ(t|Z) = λ0 (t) exp{β1 Z1 + β2 Z2 }
and obtain Table 4.2. As we shall see the two covariate problem is very much more
complex than the case of a single covariate. Not only do we need to consider the
Treatment group Z1 Z2 Log of group effect

Standard therapy 0 0 0
Experimental therapy 1 1 0 β1
Experimental therapy 2 0 1 β2
Table 4.2: Effects for two treatment groups.
effect of each individual treatment on the hazard rate for the standard therapy
but we also need to consider the effect of each treatment in the presence or
absence of the other as well as the combined effect of both treatments together.
The particular model form in which we express any relationships will typically
imply assumptions on those relationships and an important task is to bring under
scrutiny (goodness of fit) the soundness of any assumptions.
It is also worth noting that if we are to assume that a two-dimensional covari-
ate proportional hazards model holds exactly, then, integrating over one of the
covariates to obtain a one-dimensional model will not result (apart from in very
particular circumstances) in a lower-dimensional proportional hazards model. The
lower-dimensional model would be in a much more involved non-proportional
hazards form. This observation also holds when adding a covariate to a one-
dimensional proportional hazards model, a finding that compels us, in realistic
modeling situations, to only ever consider the model as an approximation.
By extension the case of several covariates becomes rapidly very complicated.
If, informally, we were to define complexity as the number of things you have to
worry about, then we could, even more informally, state an important theorem.
Theorem 4.1. (Informal non-mathematical theorem). The complexity of

any regression problem grows exponentially with the number of covariates
in the equation.
Obviously such a theorem cannot hold in any precise mathematical sense without
the need to add conditions and restrictions such that its simple take-home mes-
sage would be lost. For instance, if each added covariate was a simple constant
multiple of the previous one, then there would really be no added complexity.
But, in some broad sense, the theorem does hold and to convince ourselves of
this we can return to the case of two covariates. Simple combinatorial arguments
show that the number of possible hypotheses of potential interest is increasing
exponentially. But it is more complex than that. Suppose we test the hypothesis
H0 : β1 = β2 = 0. This translates the clinical null hypothesis: neither of the exper-
imental therapies impacts survival against the alternative, H1 : ∃ βi = 0, i = 1, 2.
This is almost, yet not exactly, the same as simply regrouping the two experi-
mental treatments together and reformulating the problem in terms of a single
binary variable.
Next, we might consider testing the null hypothesis H0 : β1 = 0 against

the alternative hypothesis H1 : β1 = 0. Such a test focuses only on the first
experimental treatment, but does not, as we might at first imagine, lump together
both the second experimental treatment and the standard treatment. This test
makes no statement about β2 and so this could indeed take the value zero (in
which case the standard and the second experimental therapy are taken to be the
same) or any other value in which case, detecting a nonzero value for β1 translates
as saying that this therapy has an effect different to the standard regardless of the
effect of the second experimental therapy. Clearly this is different from lumping
together the second experimental therapy with the standard and testing the two
together against the first experimental therapy. In such a case, should the effect
of the first experimental therapy lie somewhere between that of the standard and
the second, then, plausibly, we might fail to detect a nonzero β1 even though
there exist real differences between the standard and the first therapy.
All of this discussion can be repeated, writing β1 in the place of β2 . Already,
we can see that there are many angles from which to consider an equation such
as the above. These angles, or ways of expressing the scientific question, will
impact the way of setting up the statistical hypotheses. In turn, these impact our
inferences.
Another example would be testing the above null hypothesis H0 : β1 = β2 = 0
against an alternative H1 : 0 < β1 < β2 instead of that initially considered (i.e.,
H1 : ∃ βi = 0, i = 1, 2). The tests, and their power properties, would not typically
be the same. We might consider re-coding the problem, as in Equation 4.10, so
that testing H0 : β1 = 0 against H1 : β1 = 0 corresponds to testing for an effect
in either group. Given this effect we can test H0 : β2 = 0 against H1 : β2 = 0
which will answer the question as to whether, given that their exists a treatment
effect, it is the same for both of the experimental treatments:
λ(t|Z) = λ0 (t) exp{β1 Z1 + (β1 + β2 )Z2 }

= λ0 (t) exp{β1 (Z1 + Z2 ) + β2 Z2 }. (4.10)
Note that fitting the above models needs no new procedures or software for
example, since both cases come under the standard heading. In the first equation
all we do is write α1 = β1 and α2 = β1 + β2 . In the second we simply redefine
the covariates themselves. The equivalence expressed in the above equation is
important. It implies two things. Firstly, that this previous question concerning
differential treatment effects can be re-expressed in a standard way enabling us
to use existing structures, and computer programs. Secondly, since the effects in
our models express themselves via products of the form βZ, any re-coding of β
can be artificially carried out by re-coding Z and vice versa. This turns out to be
an important property and anticipates the fact that a non-proportional hazards
model β(t)Z can be re-expressed as a time-dependent proportional hazards model
βZ(t). Hence the very broad sweep of proportional hazards models.
It is easy to see how the above considerations, applied to a situation in which

we have p > 2 covariates, become very involved. Suppose we have four ordered
levels of some risk factor. We can re-code these levels using three binary covariates
as in Table 4.3. For this model we can, again, write the hazard function in terms
Risk factor Z1 Z2 Z3 Log of risk factor effect

Level 1 0 0 0 0
Level 2 1 0 0 β1
Level 3 0 1 0 β2
Level 4 0 0 1 β3
Table 4.3: Coding for four ordered levels of a risk factor.
of these binary coding variables, noting that, as before, there are different ways
of expressing this. In standard form we write
λ(t|Z) = λ0 (t) exp{β1 Z1 + β2 Z2 + β3 Z3 }
so that the hazard rate for those exposed to the risk factor at level i, i = 1, . . . , 4,
is given by λ0 (t) exp(βi ) where we take β0 = 0. Our interest may be more on
the incremental nature of the risk as we increase through the levels of exposure
to the risk factor. The above model can be written equivalently as
λ(t|Z) = λ0 (t) exp{β1 Z1 + (β1 + β2 )Z2 + (β1 + β2 + β3 )Z3 }

= λ0 (t) exp{β1 (Z1 + Z2 + Z3 ) + β2 (Z2 + Z3 ) + β3 Z3 } (4.11)
so that our interpretation of the βi is in terms of increase in risk. The coefficient

β1 in this formulation corresponds to an overall effect, common to all levels
above the lowest. The coefficient β2 corresponds to the amount by which the
log-hazard rate for the second level differs from that at the first. Here then, a
value of β2 equal to zero does not mean that there is no effect at level 2, simply
that the effect is no greater than that already quantified at level 1. The same
arguments follow for levels 3 and 4.
Writing the model in these different ways is not changing the basic model.
It changes the interpretation that we can give to the different coefficients. The
equivalent expression shown in Equation 4.11, for example, means that we can
carefully employ combinations of the covariates in order to use existing software.
But we can also consider the original coding of the covariates Z. Suppose that,
instead of the coding given in Table 4.3, we use the coding given in Table 4.4.
This provides an equivalent description of the four levels. As we move up the
levels, changing from level i to level i + 1, the log hazard is increased by βi .
Let’s imagine a situation, taken from Table 4.4, in which β1 = β2 = β3 .
Real situations may not give rise to strict equalities but may well provide good
first approximations. The hazards at each level can now be written very simply as
Risk factor Z1 Z2 Z3 Log of risk factor effect

Level 1 0 0 0 0
Level 2 1 0 0 β1
Level 3 1 1 0 β1 + β2
Level 4 1 1 1 β1 + β2 + β3
λ0 (t) exp(jβ1 ) for j = 0, 1, 2, 3, and this is described in Table 4.5. Taking β1 = β,

we are then able to write a model for this situation as λ(t|Z) = λ0 (t) exp(βZ), in
which the covariate Z, describing group level, takes the values 0 to 3. This model
has a considerable advantage over the previous one, describing the same situation
of four levels, in that only a single coefficient appears in the model as opposed to
three. We will use our data to estimate just a single parameter. The gain is clear.
Risk factor Z Log of risk factor effect

Level 1 0 0
Level 2 1 β
Level 3 2 2β
Level 4 3 3β
The cost, however, is much less so, and is investigated more thoroughly in the
chapters on prediction (explained variation, explained randomness) and goodness
of fit. If the fit is good, i.e., the assumed linearity is reasonable, then we would
certainly prefer the latter model to the former. If we are unsure we may prefer
to make less assumptions and use the extra flexibility afforded by a model which
includes three binary covariates rather than a single linear covariate. In real data
analytic situations we are likely to find ourselves somewhere between the two,
using the tools of fit and predictability to guide us.
Returning once more to Table 4.4 we can see that the same idea prevails for
the βi not all assuming the same values. A situation in which four ordered levels
is described by three binary covariates could be recoded so that we only have a
single covariate Z, together with a single coefficient β. Next, suppose that in the
model, λ(t|Z) = λ0 (t) exp(βZ), Z not only takes the ordered values, 0, 1, 2 and
3 but also all of those in between. In a clinical study this might correspond to
some prognostic indicator, such as blood pressure or blood cholesterol, recorded
continuously and re-scaled to lie between 0 and 3.
Including the value of Z, as a continuous covariate, in the model amounts to
making very strong assumptions. It supposes that the log hazard increases by the
same amount for every given increase in Z, so that the relative risk associated
with Δ = z2 − z1 is the same for all values of z1 between 0 and 3 − Δ. Let’s make
things a little more involved. Suppose we have the same continuous covariate,
this time let’s call it Z1 , together with a single binary covariate Z2 indicating
one of two groups. We can write
λ(t|Z1 , Z2 ) = λ0 (t) exp(β1 Z1 + β2 Z2 ).
Such a model supposes that a given change in exposure Z1 results in a given

change in risk, as just described, but that, furthermore, this resulting change is the
same at both levels of the discrete binary covariate Z2 . This may be so but such
strong assumptions must be brought under scrutiny. Given the ready availability
of software, it is not at all uncommon for data analysts to simply “throw in” all
of the variables of interest, both discrete and continuous, without considering
potential transformations or re-coding, turn the handle, and then try to make
sense of the resulting coefficient estimates together with their standard errors.
Such an exercise will rarely be fruitful. In this respect it is preferable to write
one’s own computer programs when possible or to use available software such as
the R package, which tends to accompany the user through model development.
Packages that present a “complete” one-off black-box analysis based on a single
model are unlikely to provide much insight into the nature of the mechanisms
generating the data at hand.
The user is advised to exercise great care when including continuous covari-
ates in a model. We can view a continuous covariate as equivalent to an infinite-
dimensional vector of indicator variables so that, in accordance with our informal
theorem of complexity, the number of things we need worry about is effectively
infinite. Let us not however overstate things, and it is of course useful to model
continuous covariates. But be wary. Also consider the model
λ(t|Z) = λ0 (t) exp(β1 Z + β2 Z 2 ).
If Z is binary then Z 2 = Z and there is no purpose to the second term in the

equation. If Z is ordinal or continuous then the effect of Z is quadratic rather
than linear. And, adding yet higher-order terms enables us, at least in principle,
to model other nonlinear functions. In practice, in order to carry out the analysis,
we would use existing tools by simply introducing a second variable Z2 defined
by Z2 = Z 2 ; an important observation in that the linear representation of the
covariate can be relaxed with relatively little effort. For example, suppose that
the log-relative risk is expressed via some smooth function ψ(z) of a continuous
covariate z. Writing the model
λ(t|Z) = λ0 (t) exp{βψ(Z)}
supposes that we know the functional form of the relative risk, at least up to the
constant multiple β. Then, a power series approximation to this would allow us to

write ψ(Z) = βj Z j in which any constant term β0 is absorbed into λ0 (t). We
then introduce the covariates Zj = Z j to bring the model into its standard form.
1. One of the early points of discussion on Cox’s 1972 paper was how to deal
with tied data. Look up the Cox paper and write down the various different
ways that Cox and the contributors to the discussion suggested that tied data
be handled. Explain the advantages and disadvantages of each approach.
2. One suggestion for dealing with tied data, not in that discussion, is to simply
break the ties via some random split mechanism. What are the advantages
and drawbacks of such an approach?
3. As an alternative to the proportional hazards model consider the two models

(i) S(t|Z) = S0 (t) + βZ, and (ii) logitS(t|Z) = logitS0 (t) + βZ. Discuss the
relative advantages and drawbacks of all three models.
4. Show that the relation; S(t|Z) = {S0 (t)}exp(βZ) implies the Cox model and
vice versa.
5. Suppose that we have two groups and that a proportional hazards model
is believed to apply. Suppose also that we know for one of the groups that
the hazard rate is a linear function of time, and equal to zero at the origin.
Given data from such a situation, suggest different ways in which it can
be analyzed and the possible advantages and disadvantages of the various
approaches.
6. Explain in what sense the components of Equation 4.7 and equation (4.8)
can be viewed as an equation for the mean and an equation for the variance.
7. Using equations (4.7) and (4.8) work out the calculations explicitly for the
two-group case, i.e., the case in which there are n1 (t) subjects at risk from
group 1 at time t and n2 (t) from group 2.
8. Suppose that we have available software able to analyze a proportional haz-

ards model with a time-dependent covariate Z(t). Suppose that, for the
problem in hand the covariate, Z, does not depend on time. However, the
regression effect β(t) is known to decline as an exponential function of time.
How would you proceed?
9. Suppose we fit a proportional hazards model, using some standard software,

to a continuous covariate Z defined on the interval (1,4). Unknown to us
our model assumption is incorrect and the model applies exactly to log Z
instead. What effect does this have on our parameter estimate?
10. Consider an experiment in which there are eight levels of treatment. The
levels are ordered. The null hypothesis is that there is no treatment effect.
The alternative is that there exists a non-null effect increasing with level
until it reaches one of the levels, say level j, after which the remaining levels
all have the same effect as level j. How would you test for this?
11. Write down the joint likelihood for the underlying hazard rate and the regres-
sion parameter β for the two-group case in which we assume the saturated
piecewise exponential model. Use this likelihood to recover the partial like-
lihood estimate for β. Obtain an estimate of the survivorship function for
both groups.
12. For the previous question derive an approximate large sample confidence
interval for the estimate of the survivorship function for both groups in
cases: (i) where the parameter β is exactly known, (ii) where the parameter
is replaced by an estimate with approximate large sample variance σ 2 .
13. Carry out a large sample simulation for a model with two binary variables.
Each study is balanced with a total of 100 subjects. Choose β1 = β2 = 1.5
and simulate binary Z1 and Z2 to be uncorrelated. Show the distribution of
β̂1 in two cases: (i) where the model used includes Z2 , (ii) where the model
used includes only Z1 . Comment on the distributions, in particular the mean
value of β̂1 in either case.
14. In the previous exercise, rather than include in the model Z2 , use Z2 as a
variable of stratification. Repeat the simulation in this case for the stratified
model. Comment on your findings.
15. Consider the following regression situation. We have one-dimensional covari-
ates Z, sampled from a density g(z). Given z we have a proportional hazards
model for the hazard rates. Suppose that, in addition, we are
in a position
to know exactly the marginal survivorship function S(t) = S(t|z)g(z)dz.
How can we use this information to obtain a more precise analysis of data
generated under the PH model with Z randomly sampled from g(z)?
16. Suppose we have two groups defined by the indicator variable Z = {0, 1}.
In this example, unlike the previous in which we know the marginal survival,
we know the survivorship function S0 (t) for one of the groups. How can this
information be incorporated into a two-group comparison in which survival
for both groups is described by a proportional hazards model? Use a likelihood
approach.
17. Use known results for the exponential regression model in order to construct
an alternative analysis to that of the previous question based upon likelihood.
18. A simple test in the two-group case for absence of effects is to calculate the
area between the two empirical survival curves. We can evaluate the null
distribution by permuting the labels corresponding to the group assignment

indicator Z. Carry out this analysis for the Freireich data and obtain a p-
value. How does this compare with that obtained from an analysis based on
the assumption of an exponential model and that based on partial likelihood?
19. Carry out a study, i.e., the advantages, drawbacks, and potentially restrictive
assumptions of the test of the previous example. How does this test compare
with the score test based on the proportional hazards model?
20. Obtain a plot of the likelihood function for the Freireich data. Using simple
numerical integration routines, standardize the area under the curve to be
equal to one.
21. For the previous question, treat the curve as a density. Use the mean as
an estimate of the unknown β. Use the upper and lower 2.5% percentiles
as limits to a 95% confidence interval. Compare these results with those
obtained using large sample theory.
22. Suppose we have six ordered treatment groups indicated by Z = 1, . . . , 6. For

all values of Z ≤ the hazards are the same. For Z > the hazards are again
the same and either the same as those for Z ≤ or all strictly greater than
for Z ≤ . The value of is not known. How would you model and set up
tests in this situation?
Chapter 5
Proportional hazards models in epidemiology
5.1 Chapter summary
The basic questions of epidemiology are reconsidered in this chapter from the
standpoint of a survival model. We rework the calculations of relative risk, where
the time factor is now age, and we see how our survival models can be used to con-
trol for the effects of age. Series of 2×2 tables, familiar to epidemiologists, can be
structured within the regression model setting. The well-known Mantel-Haenszel
test arises as a model-based score test. Logistic regression, conditional logistic
regression as well as stratified regression are all considered. These various mod-
els, simple proportional hazards model, stratified models, and time-dependent
models can all be exploited in order to better evaluate risk factors, how they
interrelate, and how they relate to disease incidence in various situations. The
use of registry data is looked at in relation to the estimation of survival in specific
risk sub-groups. This motivates the topic of relative survival.
By considering the time variable to correspond to age we are able to set up

powerful regression models that can be used in epidemiological applications,
including case-control studies, retrospective, and prospective studies. For many
chronic diseases, such as cancer, age is the most important risk factor and we
can control for it in several ways. Just as we do for survival time, where we
sequentially condition on time to remove its influence, the same thing can be
done for age. Relative risk models used in epidemiology come under these same
headings. For relative risk models the time component is usually taken to be age
and great generalization, e.g., period or cohort analysis is readily accomplished.
Time-dependent covariates, Z(t), in combination with the at-risk indicator, Y (t),

97
https://doi.org/10.1007/978-3-030-33439-0_5
98 Chapter 5. Proportional hazards models in epidemiology
can be used to describe states. Multistate models in which subjects can move in
and out of different states, or into an absorbing state such as death, can then
be analyzed using the same methodology.
5.3 Odds ratio, relative risk, and 2 × 2 tables
For arbitrary random variables X and Y with joint density f (x, y), conditional
densities g(x|y) and h(y|x), marginal densities v(x) and w(y), we know that
f (x, y) = g(x|y)w(y) = h(y|x)v(x),
so that, in the context of postulating a model for the pair (X, Y ), we see that
there are two natural potential characterizations. Recalling the discussion from
Section 2.3 note that, for survival studies, our interest in the binary pair (T, Z),
time and covariate, can be seen equivalently from the viewpoint of the conditional
distribution of time given the covariate, along with the marginal distribution of
the covariate, or from the viewpoint of the conditional distribution of the covari-
ate given time, along with the marginal distribution of time. This equivalence we
exploit in setting up inference where, even though the physical problem concerns
time given the covariate, our analysis describes the distribution of the covariate
given time.
In epidemiological studies the variable time T is typically taken to be age.
Calendar time and time elapsed from some origin may also be used but, mostly,
the purpose is to control for age in any comparisons we wish to make. Usually we
will consider rates of incidence of some disease within small age groups or possibly,
via the use of models, for a large range of values of age. Unlike the relatively
artificial construction of survival analysis which exploits the equivalent ways of
expressing joint distributions, in epidemiological studies our interest naturally
falls on the rates of incidence for different values of Z given fixed values of age
T . It is not then surprising that the estimating equations we work with turn out
to be essentially the same for the two situations.
The main results of proportional hazards regression, focused on the condi-
tional distribution of the covariable given time, rather than the other way around,
apply more immediately and in a more natural way in epidemiology than in sur-
vival type studies. We return to this in the later chapters that consider inference
more closely. One important distinction, although already well catered for by use
of our “at risk” indicator variables, is that for epidemiological studies the subjects
in different risk sets are often distinct subjects. This is unlike the situation for
survival studies where the risk sets are typically nested. Even so, as we will see,
the form of the equations is the same, and software which allows an analysis of
survival data will also allow an analysis of certain problems in epidemiology.
5.3. Odds ratio, relative risk, and 2 × 2 tables 99
For a binary outcome, indicated by Y = 1 or Y = 0, and a binary risk or

exposure factor, Z = 1 or Z = 0, the relative risk is defined as the ratio of
the probabilities P (Y = 1|Z = 1)/P (Y = 1|Z = 0) and the, related, odds
ratio ψ as
P (Y = 1|Z = 1)P (Y = 0|Z = 0)

ψ= .
P (Y = 1|Z = 0)P (Y = 0|Z = 1)
In the above and in what follows, in order for the notation not to become too
cluttered, we write Pr (A) = P (A). Under a “rare disease assumption”, i.e., when
P (Y = 0|Z = 0) and P (Y = 0|Z = 1) are close to 1, then the odds ratio and
relative risk approximate one another.
One reason for being interested in the odds ratio, as a measure of the impact
of different levels of the covariate (risk factor) Z follows from the identity
P (Y = 1|Z = 1)P (Y = 0|Z = 0) P (Z = 1|Y = 1)P (Z = 0|Y = 0)

= .
P (Y = 1|Z = 0)P (Y = 0|Z = 1) P (Z = 1|Y = 0)P (Z = 0|Y = 1)
(5.1)
Thus, the impact of different levels of the risk factor Z can equally well be
estimated by studying groups defined on the basis of this same risk factor and
their corresponding incidence rates of Y = 1. This provides the rationale for the
case-control study in which, in order to estimate ψ, we make our observations on
Z over fixed groups of cases and controls (distribution of Y fixed), rather than
the more natural, but practically difficult if not impossible, approach of making
our observations on Y for a fixed distribution of Z. Assumptions and various
subtleties are involved. The subject is vast and we will not dig too deeply into
this. The points we wish to underline in this section are those that establish the
link between epidemiological modeling and proportional hazards regression.
Series of 2 × 2 tables
The most elementary presentation of data arising from either a prospective study
(distribution of Z fixed) or a case-control study (distribution of Y fixed) is in
the form of a 2 × 2 contingency table in which the counts of the number of
observations are expressed. Estimated probabilities or proportions of interest are
readily calculated.
In Table 5.1, a1∗ = a11 + a12 , a2∗ = a21 + a22 , a∗1 = a11 + a21 , a∗2 =
a12 + a22 and a∗∗ = a1∗ + a2∗ = a∗1 + a∗2 . For prospective studies the pro-
portions a11 /a∗1 and a12 /a∗2 estimate the probabilities of being a case (Y = 1)
for both exposure groups while, for case-control studies, the proportions a11 /a1∗
and a21 /a2∗ estimate the probabilities of exhibiting the risk or exposure factor
Z =1 Z =0 totals
Y =1 a11 a12 a1∗
Y =0 a21 a22 a2∗
Totals a∗1 a∗2 a∗∗
Table 5.1: Basic 2 × 2 table for cases (Y = 1) and controls (Y = 0).
(Z = 1) for both cases and controls. For both types of studies we can esti-
mate ψ by the ratio (a11 a22 )/(a21 a12 ), which is also the numerator of the usual
chi-squared test for equality of the two probabilities. If we reject the null hypoth-
esis of the equality of the two probabilities we may wish to say something about
how different they are based on the data from the table.
As explained below, in Section 5.4, quantifying the difference between two
proportions is not best done via the most obvious, and simple, arithmetic differ-
ence. There is room for more than one approach, the simple arithmetic difference
being perfectly acceptable when sample sizes are large enough to be able to use
the De Moivre-Laplace approximation (Appendix C.2) but, more generally, the
most logical in our context is to express everything in terms of the odds ratio.
We can then exploit the following theorem:
Theorem 5.1. Taking all the marginal totals as fixed, the conditional distri-
bution of a11 is written

a1∗ a2∗ a1∗ a2∗
P (a|a1∗ , a2∗ , a∗1 , a∗2 ) = ψa ψu ,
a a∗1 − a u
u a∗1 − u
the sum over u being over all integers compatible with the marginal totals.
The conditionality principle appears once more, in this instance in the form of
fixed margins. The appropriateness of such conditioning, as in other cases, can
be open to discussion. And again, insightful conditioning has greatly simplified
the inferential structure. Following conditioning of the margins, it is only nec-
essary to study the distribution of any single entry in the 2 × 2 table, the other
entries being then determined. This kind of approach forms the basis of the well-
known Fisher’s exact test. It is usual to study the distribution of a11 . A non-linear
estimating equation can be based on a11 − E(a11 ), expectation obtained from
Theorem 5.1, and from which we can estimate ψ and associate a variance term
with the estimator. The non-linearity of the estimating equation, the only approx-
imate normality of the estimator, and the involved form of variance expressions
has led to much work in the methodological epidemiology literature; improving
the approximations, obtaining greater robustness and so on. However, all of this
can be dealt with in the context of a proportional hazards (conditional logistic)
5.3. Odds ratio, relative risk, and 2 × 2 tables 101
Table i Z =1 Z =0 Totals
Y =1 a11 (i) a12 (i) a1∗ (i)
Y =0 a21 (i) a22 (i) a2∗ (i)
Totals a∗1 (i) a∗2 (i) a∗∗ (i)
Table 5.2: 2 × 2 table for ith age group of cases and controls.
regression model. Since it would seem more satisfactory to work with a single
structure rather than deal with problems on a case-by-case basis, our recom-
mendation is to work with proportional and non-proportional hazards models.
Not only does a model enable us to more succinctly express the several assump-
tions which we may be making, it offers, more readily, well-established ways of
investigating the validity of any such assumptions. In addition the framework for
studying questions such as explained variation, explained randomness and partial
measures of these is clear and requires no new work.
The “rare disease” assumption, allowing the odds ratio and relative risk to
approximate one another, is not necessary in general. However, the assumption
can be made to hold quite easily and is therefore not restrictive. To do this
we construct fine strata, within which the probabilities P (Y = 0|Z = 0) and
P (Y = 0|Z = 1) can be taken to be close to 1. For each stratum, or table,
we have a 2 × 2 table as in Table 5.2, indexed by i. Each table provides an
estimate of relative risk at that stratum level and, assuming that the relative
risk itself does not depend upon this stratum, although the actual probabilities
themselves composing the relative risk definition may themselves depend upon
strata, then the problem is putting all these estimates of the same thing into
a single expression. The most common such expression for this purpose is the
Mantel-Haenszel estimate of relative risk.
Mantel-Haenszel estimate of relative risk

The, now famous, Mantel-Haenszel estimate of relative risk was described by
Mantel and Haenszel (1959) and is particularly simple to calculate. Let us first
refer to the entries of observed counts in Table 5.3.
Table i Z=1 Z=0 Totals Table i Z =1 Z=0 Totals

Y =1 a11 (i) a12 (i) a1∗ (i) Y =1 e11 (i) e12 (i) e1∗ (i)
Y =0 a21 (i) a22 (i) a2∗ (i) Y =0 e21 (i) e22 (i) e2∗ (i)
Totals a∗1 (i) a∗2 (i) a∗∗ (i) Totals e∗1 (i) e∗2 (i) e∗∗ (i)
Table 5.3: 2 × 2 table for ith age group of cases and controls. Left-hand table:
observed counts. Right-hand table: expected counts.
If we first define for the i th sub-table Ri = a11 (i)a22 (i)/a∗∗ (i) and
Si = a12 (i)a21 (i)/ a∗∗ (i), then the Mantel-Haenszel summary relative risk

estimate across the tables is given by ψ̂M H = i Ri / i Si . Breslow (1996)
makes the following useful observations concerning ψ̂M H and β̂M H = ψ̂M H .
First, E(Ri ) = ψi E(Si ) where the true odds ratio in the ith table is given
by ψi . When all of these odds ratios coincide then ψ̂M H is the solution

to the
unbiased estimating equation; R − ψS = 0, where R = i Ri and
S = i Si .
Under an assumption of binomial sampling, Breslow shows that the variances of

the individual contributions to the estimating equation are such that the quantity
2a2∗∗ (i)Var (Ri − ψSi ) can be equated to
E {[a11 (i)a22 (i) + ψa12 (i)a21 (i)] [a11 (i) + a22 (i) + ψ (a12 (i) + a21 (i))]} ,
from which, by a simple application of the delta method (Appendix A.10), we

can obtain estimates of the variance of ψ̂M H .
5.4 Logistic regression and proportional hazards
Without any loss in generality we can express the two probabilities of interest,
P (Y = 1|Z = 1) and P (Y = 1|Z = 0) as simple power transforms of one another.
This follows, since, whatever the true values of these probabilities, there exists
some positive number α such that P (Y = 1|Z = 1) = P (Y = 1|Z = 0)α . The
parameter α is constrained to be positive in order that the probabilities themselves
remain between 0 and 1. To eliminate any potential dangers that may arise,
particularly in the estimation context where, even though the true value of α
is positive, the estimate itself may not be, a good strategy is to re-express this
parameter as α = exp(β). We then have
log −{log P (Y = 1|Z = 1)} = log −{log P (Y = 1|Z = 0)} + β. (5.2)
The parameter β can then be interpreted as a linear shift in the log-log trans-
formation of the probabilities, and can take any value between −∞ and ∞, the
inverse transformations being one-to-one and guaranteed to lie in the interval
(0,1). An alternative model to the above is
logit P (Y = 1|Z = 1) = logit P (Y = 1|Z = 0) + β. (5.3)
where the logit transformation, again one-to-one, is defined by logit θ = log{θ/(1−

θ)}. Although a natural model, the model of Equation 5.2 is not usually pre-
ferred to that of Equation 5.3, motivated in an analogous way (i.e., avoiding
5.4. Logistic regression and proportional hazards 103
constraints) but having a slight advantage from the viewpoint of interpretation.

This is because the parameter β is the logarithm of the odds ratio, i.e., β = log ψ.
In the light of the equivalence of the odds for disease given the risk factor and
the odds for the risk factor given the disease, as expressed in Equation 5.1, we
conclude immediately that, equivalent to the above model involving β, expressed
in Equation 5.3, we have a model expressing the conditional probability of Z given
Y and using the same β. This highlights an important feature of proportional
hazards modeling whereby we focus attention on the conditional distribution of
the covariates given an event yet, when thinking of the applied physical problem
behind the analysis, we would think more naturally in terms of the conditional
distribution of the event given the covariates. The essential point is that the
unknown regression parameter, β, of interest to us is the same for either situation
so that in place of Equation 5.3, we can write
logit P (Z = 1|Y = 1) = logit P (Z = 1|Y = 0) + β. (5.4)
Since the groups are indicated by a binary Z, we can exploit this in order to
obtain the more concise notation, now common for such models, whereby
logit P (Y = 1|Z) = logit P (Y = 1|Z = 0) + βZ. (5.5)
As we have tried, in as much as is possible throughout this text, to restrict atten-

tion to a single explanatory variable, this is once more the case here. Extension
to multiple explanatory variables, or risk factors, is immediate and, apart from
the notation becoming more cumbersome, there are no other concepts to which
to give thought. We write the model down, as above in Equation 5.5, and use
several binary factors Z (Z now a vector) to describe the different group levels.
The coefficients β (β now a vector) then allow the overall odds ratio to be mod-
eled or, allows the modeling of partial odds ratios whereby certain risk factors are
included in the model, and our interest focuses on those remaining after having
taken account of those already included. The above model can also be written
in the form
P (Y = 1|Z)
= exp(β0 + βZ) , (5.6)
1 − P (Y = 1|Z)
where β0 = logit P (Y = 1|Z = 0). Maintaining an analogy with the usual linear
model we can interpret β0 as an intercept, simply a function of the risk for a
“baseline” group defined by Z = 0.
Assigning the value Z = 0 to some group and thereby giving that group
baseline status is, naturally, quite arbitrary and there is nothing special about
the baseline group apart from the fact that we define it as such. We are at lib-
erty to make other choices and, in all events, the only quantities of real interest to
us are relative ones. In giving thought to the different modeling possibilities that
arise when dealing with a multivariate Z, the exact same kind of considerations,
already described via several tables in the section on modeling multivariate prob-
lems will guide us (see Section 4.6 and those immediately following it). Rather
than repeat or reformulate those ideas again here, the reader, interested in these
aspects of epidemiological modeling, is advised to go over those earlier sections.
Indeed, without a solid understanding as to why we choose to work with a par-
ticular model rather than another, and as to what the different models imply
concerning the complex inter-relationships between the underlying probabilities,
it is not really possible to carry out successful modeling in epidemiology.
Stratified and conditional logistic regression

In the above model, and Z being multivariate, we may wish to include alongside
the main factors under study, known risk factors, and particularly risk factors
such as age, or period effects, for which we would like to control. Often age
alone is the strongest factor and its effect can be such that the associated errors
of estimation in quantifying its impact can drown the effect of weaker risk factors.
A powerful approach in controlling for such factors, S, is to appeal to the idea
of stratification. This means that analysis is carried out at each level of S and,
within a level, we make the same set of assumptions concerning the principle
factors under study. We write
P (Y = 1|Z, S)
= exp(β0 + βZ) , (5.7)
1 − P (Y = 1|Z, S)
where, in the same way as before, β0 = logit P (Y = 1|Z = 0, S). The important
aspect of a stratified model is that the levels of S only appear on the left-hand
side of the equation.
We might conclude that this is the same model as the previous one but it is
not quite and, in later discussions on inference, we see that it does impact the
way in which inferences are made. It also impacts interpretation. In the simpler
cases, in as far as β is concerned, the stratified model is exactly equivalent to a
regular logistic model if we include in the regression function indicator variables,
of dimension one less than the number of strata. However, when the number of
strata is large, the use of the stratified model enables us to bypass estimation of
the stratum-level effects. If these are not of real interest then this may be useful in
that it can result in gains in estimating efficiency even though the underlying mod-
els may be equivalent. In a rough intuitive sense we are spending the available esti-
mating power on the estimation of many less parameters, thereby increasing the
precision of each one. This underlines an important point in that the question of
stratification is more to do with inference than the setting up of the model itself.
This last remark is even more true when we speak of conditional logistic
regression. The model will look almost the same as the unconditional one but
the process of inference will be quite different. Suppose we have a large number
of strata, very often in this context defined by age. A full model would be as in
5.4. Logistic regression and proportional hazards 105
Equation 5.6, including in addition to the risk factor vector Z, a vector parameter
of indicator variables of dimension one less than the number of strata. Within
each age group, for the sake of argument let’s say age group i, we have the simple
logistic model. However, rather than write down the likelihood in terms of the
products P (Y = 1|Z) and P (Y = 0|Z) we consider a different probability upon
which to construct the likelihood, namely the probability that the event of inter-
est, the outcome or case in other words, occurred on an individual (in particular
the very individual for whom the event did occur, given that one event occurred
among the set S{i} of the a∗∗ (i) cases and controls. Denoting Zi to be the risk
factor for the case,
corresponding to the age group i, then this probability is sim-
ply exp(βZi )/ I[j ∈ S{i}] exp(βZj ). The likelihood is then the product of such
terms across the number of different age groups for which a case was selected.
If we carefully define the “at-risk” indicator Y (t) where t now represents age,
we can write the conditional likelihood as
n
δi
exp(βZi )
L(β) = n . (5.8)
Here we take the at-risk indicator function to be zero unless, for the subject j,
Xj has the same age, or is among the same age group as that given by Xi .
In this case the at-risk indicator Yj (Xi ) takes the value one. To begin with, we
assume that there is only a single case per age group, that the ages are distinct
between age groups, and that, for individual i, the indicator δi takes the value
one if this individual is a case. Use of the δi would enable us to include in an
analysis sets of controls for which there was no case. This would be of no value
in the simplest case but, generalizing the ideas along exactly the same lines as
for standard proportional hazards models, we could easily work with indicators
Y (t) taking the value one for all values less than t and becoming zero if the
subject becomes incident or is removed from the study. A subject is then able
to make contributions to the likelihood at different values of t, i.e., at different
ages, and appears therefore in different sets of controls. Indeed, the use of the
risk indicator Y (t) can be generalized readily to other complex situations.
One example is to allow it to depend on two time variables, for example, an
age and a cohort effect, denoting this as Y (t, u). Comparisons are then made
between individuals having the same age and cohort status. Another useful gen-
eralization of Y (t) is where individuals go on and off risk, either because they
leave the risk set for a given period or, possibly, because their status cannot be
ascertained. Judicious use of the at-risk indicator Y makes it possible then to
analyze many types of data that, at first glance, would seem quite intractable.
This can be of particular value in longitudinal studies involving time-dependent
measurements where, in order to carry out unmodified analysis we would need, at
each observed failure time, the time-dependent covariate values for all subjects at
risk. These would not typically all be available. A solution based on interpolation,
assuming that measurements do not behave too erratically, is often employed.
Alternatively we can allow for subjects for whom, at an event time, no reliable
measurement is available, to simply temporarily leave the risk set, returning later
when measurements have been made.
The striking thing to note about the above conditional likelihood is that it
coincides with the expression for the partial likelihood. This is no real coincidence
of course and the main theorems of proportional hazards models apply equally
well here. This result anticipates a very important concept, and that is the idea
of sampling from the risk set. The difference between the Y (t) in a classical
survival study, where it is equal to one as long as the subject is under study
and then drops to zero, as opposed to the Y (t) in the simple epidemiological
application in which it is zero most of time, taking the value one when indicating
the appropriate age group, is a small one. It can be equated with having taken
a small random sample from a conceptually much larger group followed since
time (age) is zero. On the basis of the above conditional likelihood we obtain
the estimating equation
n
n
j=1 Yj (Xi )Zj exp(βZj )
U (β) = δi Zi − n , (5.9)
which we equate to zero in order to estimate β. The equation contrasts the same
quantities written down in Table 5.3 in which the expectations are taken with
respect to the model. The estimating equations are then essentially the same
as those given in Table 5.3 for the Mantel-Haenszel estimator. Furthermore,
taking the second
derivative of the expression for the log-likelihood, we have
that I(β) = n i=1 i Ii (β) where
δ
n n 2
2
j=1 Yj (Xi )Zj exp(βZj ) j=1 Yj (Xi )Zj exp(βZj )
Ii (β) = n − n ,
j=1 Yj (Xi ) exp(βZj ) j=1 Yj (Xi ) exp(βZj )
(5.10)
n
then I(β) = i=1 δi Ii (β). Inferences can then be carried out on the basis of
these expressions. In fact, once we have established the link between the applied
problem in epidemiology and its description via a proportional hazards model,
we can then appeal to those model-building techniques (explained variation,
explained randomness, goodness of fit, conditional survivorship function etc.)
which we use for applications in time to event analysis. In this context the building
of models in epidemiology is no less important, and no less delicate, than the
building of models in clinical research. Although many of the regression modeling
ideas came to the field of epidemiology later than they did for clinical research,
several of the deeper concepts such as stratification, risk-set sampling and non-
nested studies were already well known to epidemiologists. As a result some
5.5. Survival in specific groups 107
of the most important contributions to the survival literature have come from
epidemiologists. For readers looking for more of an epidemiological flavor to
survival problems we would recommend taking a look at Breslow (1978), Annesi
and Lellouch (1989), Rosenberg and Anderson (2010), Cologne et al. (2012),
Xue et al. (2013), Moolgavkar and Lau (2018) and Fang and Wang (2020).
5.5 Survival in specific groups
The issue of competing risks is never far from our attention, no less so when we
discuss problems in epidemiology. Almost without exception, outcomes other than
that of the investigators’ main focus, will hinder its observation and will bring into
play one of the approaches, CRM-I, CRM-J or CRM-ID (Section 2.6), in order to
enable our work to progress. Some obvious questions such as how would a given
population fare if we were able to remove some particular cause of death turns
out to be almost impossibly difficult to even correctly formulate. How probable
is it that a member of some high-risk group will succumb to the disease is again
a question that is very difficult to address. This may at first seem surprising.
Survival methods can help and, in this and the following section dealing with
genetic epidemiology, we will see how very great care is needed to avoid making
quite misleading inferences. One case in point is that of the so-called breast
cancer susceptibility genes, BRCA1 and BRCA2. Any additional risk to carriers,
if indeed there is any additional risk, has been very greatly exaggerated.
Relative survival and registry data

Under CRM-I (see Section 1.5) we can follow a large number of subjects for
an interval of short duration, noting the number available to suffer the event of
interest during the interval, together with the number of cases occurring. CRM-I
enables us to obtain an estimate (slightly biased downwards due to censorings
during the interval) of the disease rate as the ratio of the number of events to
the number available at the beginning of the interval. The smaller the interval
size the better but, of course, in practice we are limited by the amount of data
available as well as the mechanisms by which we can make accurate recordings.
Typically these intervals are of 5-years durations.
At some time and place we may have large enough observations to enable
accurate estimates of disease rates. Under some assumptions differential rates
can be smoothed, often using Bayesian techniques, and this is a whole field that
has many overlaps with the main strands of this text. Suppose, for some popula-
tion, we have large enough estimates from registry data, that we can consider the
survival probabilities known, denoted as S0 (t). Within this population we have
all disease factors acting, and represented in accordance to their observed fre-
quencies. We may be interested to know how some specific sub-group, say those
who have been incident for disease Z fare with respect to the whole population
at large. We might postulate an elementary model for this as;
S(t|Z) = S0 (t) × α(t) (5.11)
where S(t|Z) is the survival of the sub-group of interest and S0 (t) the population
survival. For this model, we refer to α(t) as relative survival. It is described as
a measure of net survival for this group in the absence of all those other factors
that influence the reference group. We often assume that the size of the group
Z is not large enough to have a significant impact on S0 (t). In consequence, the
sub-group, S(t|Z) forms a negligible component of S0 (t) so that α(t) will be
described as the ratio of survival in the sub-group with respect to the expected
population survival unaffected by whatever handicap the sub-group suffers from.
Equation 5.11 is very popular in this context due to its simple interpretation. As
pointed out in Section 2.4 it is generally preferable to appeal to the log-minus-log
transformation as the basis for a model. We would then have a non-proportional
hazards model alternative to Equation 5.11 as
log{− log S(t|Z)} = β(t) + log{− log S0 (t)}. (5.12)
Aside from taking logs it is no more difficult to estimate β(t) than α(t) and,
should β(t) appear to be reasonably constant over time, then we could appeal to
t=age n(t) nC (t) I(t) IH (t)

00-05 10,000,000 0 0 0
05-10 9,895,159 2 1 in 4,947,580 16 in 4,947,580
10-15 9,887,872 26 1 in 380,303 16 in 4,947,580
15-20 9,879,578 173 1 in 57,107 16 in 57,107
20-25 9,862,951 1087 1 in 9,073 16 in 9,073
25-30 9,840,178 4795 1 in 2,052 16 in 2,052
30-35 9,810,380 14,529 1 in 675 16 in 675
35-40 9,763,506 34,102 1 in 286 16 in 286
40-45 9,681,029 69,466 1 in 139 16 in 139
45-50 9,538,425 106,633 1 in 89 16 in 89
50-55 9,324,938 136,860 1 in 68 16 in 68
55-60 9,034,317 168,931 1 in 53 16 in 53
60-65 8,634,150 189,344 1 in 46 16 in 46
65-70 8,088,535 198,438 1 in 41 16 in 41
70-75 7,365,731 189,195 1 in 39 16 in 39
00-75 1 in 8 = 12.1% 7 in 8 = 87%
Table 5.4: Based on NIH-SEER data. n(t) =# at-risk at age t, nC (t) =# cases
between t and t + 5, I(t) = overall incidence rate, IH (t) = incidence rate (high
risk, BRCA1/2) on the basis of a relative risk of 16.
5.5. Survival in specific groups 109
a simple proportional hazards model to describe the observations. An important

contribution in this field is that of Perme et al. (2012).
Using models crm-i and crm-j to expoit registry data

In epidemiology, the age-specific failure rate, λ(t), characterizes the distribution
of the random variable T, the time until the relevant event. Our illustration here
focuses on breast cancer. In the simplest construct the only cause of death is
breast cancer, so that the probability that the incidence of breast cancer occurs
before time t is given by F (t) = 1 − exp{−Λ(t)} where Λ(t) is the cumulative
t
risk over time. Note that Λ(t) = 0 λ(u)du. Under CRM-I we consider censoring
random variables, C1 , ... Ck , describing the occurrence of competing events. If
any one of them occurs before T = t, then we are unable to observe our outcome
of interest should it take place itself any time later than t. The risk function
of interest—in as much as it’s the only one that can be estimated by real data
under CRM-I—is then rather different to λ(t) and can be expressed as
λ{t|Cmin > t} = lim (Δt)−1 Pr t < T < t + Δt|T > t, Cmin > t .
Δt→0+
where Cmin = min(C1 , ..., Ck ). The conditioning event Cmin > t is of central
importance in epidemiological studies since, in practical investigations—in par-
ticular the compilation of registry data—all our observations at time t have
necessarily been conditioned by the events, T > t and Cmin > t. All associated
probabilities are also necessarily conditional. But note that, under an indepen-
dent censoring mechanism, λ(t|Cmin > t) = λ(t). This result is crucial since,
given independence of the competing causes of failure, and, by partitioning the
interval (0,t) into t0 < t1 <, ..., < tm where t0 = 0 and tm = t then, as a conse-
quence of the above expression we can write:
t m

Λ(t) = λ{u|Cmin > u}du ≈ Pr{tj−1 < T < tj |T > tj−1 , Cmin > tj−1 }
0 j=1
(5.13)
so that we can approximate F (t) via F (t) = 1 − exp{−Λ(t)}. When using this
approximation with real data, the quality of the approximation can be investigated
on a case-by-case basis. It is common, in epidemiological applications, to use
gaps, tj − tj−1 , as large as 5 years. Equation 5.13 can be directly used and
interval probability estimates taken directly from registry data. Data such as the
SEER data, for very large samples, indicate those subjects entering each 5-year
age interval, i.e., those satisfying the conditioning restriction in Equation 5.13,
and just how many observed cases could be counted during that 5-year interval.
For each component to the sum we have an empirical estimate of the required
probability, the numerator being simply the number of cases seen during the
interval and the denominator typically the number of individuals satisfying the
criteria to be “at-risk” at the beginning of the interval. Attempts to improve the
approximation may involve looking at the average number at-risk throughout the
interval rather than just at the beginning or some other model to improve the
numerical approximation to the integral. Working with model CRM-J, rather than
CRM-I will, in many cases, and these can include breast cancer, have little impact
on the estimation of the cumulative incidence rate (Satagopan and Auerbach,
2004).
Equation 5.13 is the basis for estimating probabilities of cancer incidence over
given periods. Cancer registry data, such as the SEER data set, provide for the
numbers exposed to risk of breast and other cancers and the number of cases
observed during those 5-year intervals. These two numbers provide the numerator
and the denominator to our conditional probability estimates. We can first use
Table 5.4 to confirm the widely quoted figure of one woman in eight will have
breast cancer in her life. If we let H = 1/39 + 1/41 + 1/46 + ... + 1/4, 947, 580,
in Equation 5.13, then we find that 1 − exp(−H) = 0.121 ≈ 1/8, a much quoted
result in not just scientific journals but in everyday popular sources. Note that
this estimated probability is unaltered in the first 3 figures if we start the clock
from age 20 rather than age zero. The reason for this can be seen in the table
since those early rates are so small as to be negligible. On the other hand, at the
top end, the summation stops for women over 75 years of age. Some authors go
beyond 75 years when evaluating lifetime risk (Brose et al. (2002), for example,
calculate out to age 110 years).
Our estimates are limited by the rapidly declining number of observations
for the higher age groups but, theoretically, if we were able to estimate the
death rate due to cancer without fixing some upper limit then it would go to
one hundred percent. This is easily seen intuitively since, everyone has to die
eventually, and the calculations do not allow for anyone to die of causes other
than breast cancer (in technical language, other causes of death are “censored
out” under CRM-I when we calculate the risk sets). When appealing to model
CRM-J, instead of CRM-I, we will need to introduce into the calculation the
marginal survival time to either death or breast cancer incidence (see Section
2.6) which is often approximated from life tables by registry death rates. In this
case the choice of model did not have any significant impact on estimation.
5.6 Genetic epidemiology
It is not easy to find a subject more fascinating than genetics. Seyerle and Avery
(2013) describe genetic epidemiology as being at the crossroads of epidemiology
and genetics, the discipline’s aim being to identify the myriad of relationships
between inherited risk factors and disease etiology. The field is truly a vast one and
our purpose here is very limited—to consider the contribution to such an aim that
can be made by modern techniques of survival analysis, proportional and non-
5.6. Genetic epidemiology 111
proportional hazards models in particular. Many of the more notable advances in

genetics have not come from statistical modeling but from more direct combina-
torial probabilistic calculation. Mendelian inheritance, Hardy-Weinberg equilib-
rium equations, and pedigree analysis all come under this heading. The genetic
nature of certain diseases, their observed frequencies being well explained by
given mechanisms, for example, a recessive normal allele together with a domi-
nant abnormal allele, have added greatly to our understanding of diseases such
as Huntington’s. This is all the more significant as diseases like Huntington’s may
not become manifest until well beyond the sufferer’s age of producing offspring.
Now, while we can well describe the example of Huntington’s using probabil-
ity models, the probabilistic mechanism of these models is fully understood. It is
a combinatorial one. The situation is very different when we consider so-called
“cancer susceptibility” genes. Unlike Huntington’s where the dominant inher-
ited gene has been traced to a mutation of a specific nature—an approximate
tripling of the frequency of a particular nucleotide sequence—when we consider
susceptibility genes we have nothing particular in mind and while several types
of polymorphisms may be considered unharmful others will be considered “dele-
terious”. This number has grown over the years and the number of deleterious
mutations believed to be associated with breast cancer, some on chromosome
17 and some on chromosome 13, are now estimated to be in the hundreds.
Establishing family association

Inheriting a defective allele, considered either dominant or recessive, is well
described by simple probabilistic models. Observations on families are necessarily
very sparse and so studies will rely on collecting families and appealing to likeli-
hood models to provide complete descriptions of the data. Even in the absence
of any clear indication as to the location of an implicated gene, or genes, the
distribution of diseases such as Huntington’s can be accurately described. This
is a very different situation to that we encounter when we set out to investigate
so-called “cancer susceptibility” genes. In this latter situation the room for errors
is very large; so much so that some of the more alarming claims concerning the
impending risks faced by carriers of, for example, the BRCA1 and BRCA2 muta-
tions, need to be taken with a great deal of circumspection. What is more, these
errors, whether statistical (lack of precision), logical (the presence of flaws in the
reasoning), or interpretative (the difficulty in translating results into something
understandable) will amplify one another. Without some understanding, however
limited, of these errors and their sources, the carrier is in no position to make
any kind of informed decision. A decision based on little more than some simple
take-home number, such as lifetime risk, would seem to be most unwise. The
techniques of survival analysis can provide deeper insight. We describe this below
but, first, let us revisit those approaches that will greatly overestimate the degree
of family association.
Early studies on family association in breast cancer used case-control method-

ology. Making the usual efforts to control for age and possibly other risk factors,
the following question was put to both the case and the control subject: do
you have one or more family members diagnosed with breast cancer? The esti-
mated probability of a positive response for the cases turned out to be more
than twice that of the controls. When the question was changed to two or more
family members, this ratio increased further to more than three. At first glance
this may appear to provide solid evidence in favor of a family association with a
strong relative risk. The logic is however flawed. To see this, we first define Y
to be the total number of cases in the family, i.e., at least one for the case and
no more than n − 1, where n is the number of relevant family members, for the
control. We express our null hypothesis as
H0 : Pr (Y ≥ 2 | Y ≥ 1) = Pr (Y ≥ 1 | Y ≥ 0)
and we estimate the relative risk, ψ, by
Pr (Y ≥ 2 | Y ≥ 1)
ψ= .
Pr (Y ≥ 1 | Y ≥ 0)
We can re-express the null hypothesis—are cases no more likely than controls
to have other family members with breast cancer—as; H0 : ψ = 1. Now, while
the hypothesis, H0 : Pr (Y ≥ 2 | Y ≥ 1, n) = Pr (Y ≥ 1 | Y ≥ 0, n) will correctly
control the size of the test, the more obvious formulation, H0 : Pr (Y ≥ 2 | Y ≥
1) = Pr (Y ≥ 1 | Y ≥ 0) will not. Indeed, under H0 it is not generally true that
ψ = 1. We have that;

Pr (Y ≥ 2 | Y ≥ 1) = Pr (Y ≥ 2 | Y ≥ 1, n) × g(n | Y ≥ 1) (5.14)
n

Pr (Y ≥ 1 | Y ≥ 0) = Pr (Y ≥ 1 | Y ≥ 0, n) × g(n | Y ≥ 0) (5.15)
n
Under the null hypothesis we would like for ψ, the ratio of Equation 5.14 to
Equation 5.15, to be equal to one. This however is not the case. Choosing a
particular distribution for g(n), specifically one with support restricted to n =
{3, 25}, and with g(3)/g(25) for the under 50 age group twice that of the other
group, we obtain the results shown in Table 5.5. In brackets are the values taken
from early works that supposedly justify the conclusion of family association.
Table 5.5 is based on an independence assumption and we obtain very similar
results to those taken from the literature. Such results do not therefore give
support to a conclusion of family association.
Of course, the particular distribution chosen for g(n) is not at all plausible.
This is however beside the point since it underlines what we need to know and
# 1st degree relatives ψ (age<50 ) ψ (age>50)

1 or more 2.10 (2.14) 1.76 (1.65)
2 or more 3.71 (3.84) 2.42 (2.61)
Table 5.5: Ratio of the left hand sides of Equation 5.14 to Equation 5.15 under
Binomial sampling with unknown n and independence. In brackets, ψ from case
control studies, the strong similarity of the numbers tending to contradict a
conclusion of dependence.
that is that, even under independence, the null hypothesis will not hold. The
reason for this is that knowing that Y ≥ 1 tells us something about n, i.e., a case
is more likely to be associated with another case in the family than is a control.
This dependence will disappear if we are able to condition (take as fixed) n. This
is not usually feasible and, in any event, the bias would not disappear due to
other less obvious biases, recall bias in particular. In statistical terms, the test
cannot be shown to be unbiased (note that the meaning of the term unbiased
differs when we refer to a test rather than an estimator) and will fail to control
for the false-positive rate.
Identifying implicated gene or genes

The null hypothesis of no family association is unlikely to be fully satisfied since
many factors, environmental, socioeconomic, and other less obvious ones are
potentially involved and shared to some degree within families. Had early studies
failed to reject a null hypothesis of no family association that would have brought
the discussion, at least for the time being, to a close. Instead, the apparent
discovery of strong family association was enough to conclude that genetics was
playing a part, a conclusion followed by the natural question ... which gene or
genes are implicated.
Moving beyond the relatively simple, albeit difficult, question of the detection
of family association onto that of the identification of responsible mutations
and their location is no small task. The difficulty here can only be described
as herculean. The studies leading to the identification of BRCA1 made use of
quite small data sets consisting of families containing varying numbers of cases.
Likelihood methods (LOD scores) were calculated to see which mutations were
most strongly associated with the observed outcomes, these outcomes being the
numbers of cases. The statistical problem here is that of looking for a needle
in a haystack, sometimes referred to as the “fishing expedition” problem. The
likelihood will not identify with any reasonable degree of accuracy a specific
location on the genome. For sure, some location will correspond to the largest
likelihood, while others, possibly very far removed in genetic distance, will also
produce likelihoods that can lie a negligibly small distance from the maximum.
The point estimate is very imprecise.
Essentially, the likelihood (LOD scores) method is of little help in cases like
this. For more regular problems the log-likelihood that is used to estimate an
unknown parameter will furnish us with an estimating equation, the zero of which
corresponds to our parameter estimate. Small perturbations in the likelihood
around its maximum correspond to small perturbations in our parameter estimate.
We have continuity but, usually more, we have a log-likelihood that is twice
differentiable allowing us, after verifying that the order of limiting processes can
be switched, to provide an assessment of the precision of our estimate. For our
genome-wide search we have no such conditions and no such reassuring estimates
of precision. So much so that, for the kind of limited data available, the gene
location identified by a likelihood maximum has much less chances of being the
quantity we are seeking than has the sum total of all other locations. If, as a
statistical technique, likelihood was up to the task, it would tell us that, even if
our best estimate of the location on the genome is not estimated with perfect
precision, then the location must be nearby. However, no such statement can be
made, even in any approximate way, telling us that a heavy reliance on likelihood
is not warranted.
In order to get a better insight into the size of the challenge here, consider
an experiment quite unrelated to genetics. Suppose we have a fair deck of cards
and one card is randomly chosen. The outcome of interest is obtaining a black
queen. The probability of this outcome is 1/26. Suppose now that we have a
second deck, this time a very flawed deck in which half of the cards are black
queens. On the basis of observations our goal is to identify the flawed deck. To
this end we will use the likelihood function. The likelihood that a random draw
obtains a black queen is 13 times higher in the second deck than the first. If,
for each deck, 10 cards were drawn with replacement and we count the number
of times X , X = 0, ...., 10, we observe a black queen, then, identifying (using
the likelihood) the fair (1/26) from the flawed (13/26) deck, we would choose
incorrectly the flawed deck less than one time in ten thousand. We are just about
certain to correctly identify the faulty deck.
How would this work though when our faulty deck is surrounded by, not
one, but one hundred thousand fair decks. Choosing the deck that maximizes
the likelihood will very rarely lead to the correct choice. We are far more likely
to incorrectly identify a fair deck than correctly identify the faulty deck. And
this despite the huge difference in risk, 13:1. Note also, returning to the case of
interest, that of the BRCA gene, the relative risk is not generally believed to be
this large. And of course, while the card example relates to a search across one
hundred thousand candidate choices, in the case of alleles defined by nucleotide
variations, we are talking about millions of choices.
An approach based on the likelihood can be enhanced using techniques
such as linkage analysis that will weigh more strongly the relationships mother-
daughter to mother-niece. It will nonetheless do little to refine what remains a
rough and ready approach involving a lot of statistical uncertainty. Although it
can be argued that the best estimate of the implicated gene is the likelihood-
based estimate, we ought not to overlook the great deal of imprecision associated
with this estimate. More problematic is the highly erratic nature of an estimate,
such as this one, where the parameter of interest (the gene) is not a contin-
uous function of the log-likelihood. Very small non-significant differences from
the maximum of the likelihood (LOD score) can be associated with genes that
are not only very far removed from the so-called BRCA genes but they may well
be found on entirely different chromosomes. The situation here is quite different
from that we are familiar with when the log-likelihood is a continuous differen-
tiable function of the unknown parameter(s), enabling us to structure reliable
inference based on estimating equations. All we have here is the best point esti-
mate of the implicated gene and no measure at all regarding the amount of
sampling uncertainty that ought to be associated with it.
Establishing family association using survival analysis

The case-control methodology used to detect and quantify the degree of fam-
ily association with breast cancer is too inaccurate to allow reliable estimation.
Biases such as recall bias—cases are much more likely than controls to inves-
tigate the cause of death for some remote aunt, otherwise barely heard of or
mentioned—and the impracticability of controlling for family size make infer-
ences unconvincing.
Survival analysis provides a way forward. First, we need to directly acknowl-
edge the time (age) component of the study. A case-control study mostly ignores
this although some efforts are made by age matching. The key concept here is
that of a time-dependent covariate indicating, for any individual, how many other
family members have been incident for breast cancer. At some age, an individual
may have no family members that have been incident for breast cancer, while,
six months later, the same individual could signal two family members incident
for breast cancer. Not only that but, if we have the ages of the individual at
which the other members became incident for breast cancer, then we are in a
position to structure our time-dependent covariate, Z(t) so that it can assume
values, Z(t) = 0, 1, 2, ..., as age t ranges from 0 to some upper limit. In this way
we can see that not only can a control later become a case but, the risk factor
itself is time dependent.
We consider models that will accommodate such situations in the following
chapter. More than one model could be of interest and, in particular, we may
wish to consider making use of stratified models where the stratification variable
would be indexed to the size of the family. The logarithm of the relative risk—
point and interval estimates—can be obtained from inference for the regression
coefficient. A test of the null hypothesis: there is no family association, is readily
expressed in terms of a test on the regression coefficient, β. The maximum
likelihood estimate, β̂, quantifies the observed degree of family association.
Quantifying individual risk

Should a healthy individual be told that she or he carries one of the genetic
mutations believed to be associated with increased risk of breast cancer the
question is how to best advise them. Guidance on providing any such advice is
well beyond the capacity of this work. That said, if any part of this guidance is
built around a statistical assessment of the risks involved, we could say to the
carrier the following: focus only on short-term risks; long-term risks are hopelessly
misleading. While short-term risks can be readily understood as being the ratio of
those who fall victim to the disease divided by all of those who might have become
such a victim, long-term risks have no such interpretation and lean wholly on
complex models built on an array of untestable assumptions that are, to say the
very least, somewhat unlikely. Not only that but, whereas an estimation of short-
term risks will most likely produce percentages offering some reassurance, an
estimation of long-term risks, unless dealing with a statistical expert, are likely
to produce percentages that provoke considerable alarm. Claims such as 45%
to 65% of BRCA1/2 carriers will get breast cancer (ww5.komen.org) are very
misleading and have no useful scientific foundation. No less alarmists are claims
that an intervention such as a double mastectomy will reduce the probability
of breast cancer by a large amount. Figures of 90% or more are often quoted.
Again, such claims have no statistical or scientific foundation.
O’Quigley (2017) explains why long-term risks have no fruitful meaning that
could be of any help to a carrier. In that paper the example illustrates the case of
25-year-old Sophie who has been given an 87% lifetime risk of developing breast
cancer. The same data and assumptions simultaneously show that Sophie has a
99.9% chance of attaining age 30 cancer free and, if she succeeds—the chances
are overwhelmingly in her favor—she then has a 98% change of reaching age
35 cancer free. Indeed, since the biggest constituent to lifetime risk comes after
age 70, Sophie’s chances of reaching age 65 cancer free do not differ dramati-
cally from that of non-carriers. And these calculations make no provision for the
potentially very large statistical inaccuracies in quantifying family association as
well as imprecision in the location of the implicated genes. The calculations take
their starting point to be the most pessimistic one—from the wide range of esti-
mated relative risks found in the literature, close to the highest is assumed—so
that, in truth, there is very much less cause for alarm than that initially provoked
in Sophie upon being first informed of her lifetime risk.
Even if we were able to provide some meaningful interpretation of lifetime risk
we would have to say that, as a summary of the risks facing a carrier, it is at its
best, woefully inadequate. Aside from short-term risks, there are other summary
statistics, over a lifetime, that the carrier can make sense of. For example, if she
imagines herself together with 4 friends of similar age and health characteris-
tics, she would readily understand that, from the 5 of them, all else being equal,
her probability of not being the first one to die would be 0.8. Death from all
causes is what matters and the question now becomes, given her BRCA status,
how does that probability change. The answer to that is ... not much. It can
be estimated using registry data and some calculations involving order statistics
(Appendix A.9). It would greatly strengthen intuitive understanding of the impli-
cations of carrier status. The idea is to come up with quantifiers of risk that can
be understood on some level as opposed to lifetime risk which a carrier will not
understand. More work could be done. The example of the first to die out of a
small group of friends is immediately extended to not being one of the first two
and how this might change given BRCA status. As well as short-term probabil-
ities, another useful quantity from survival analysis, is the expected remaining
lifetime at some point. For a 25-year-old carrier, assuming that surgery would
put them in the same position as a non-carrier, we can calculate a difference in
mean remaining lifetimes of 3.7 years. We might tell a carrier that they have an
87% probability of getting breast cancer without surgery. Or that, with surgery,
they will be expected to gain, on average, some 3.7 years a half-century down
the line. The carrier would interpret those two statements very differently. And
yet they use the very same data, the same assumptions, and the same belief as
to the existence of the BRCA gene as well as the unfavorable outcome associated
with many of the variant polymorphisms.
1. Consider an epidemiological application in which workers may be exposed

to some carcinogen during periods in which they work in some particular
environment. When not working in that particular environment their risk
falls back to the same as that for the reference population. Describe this
situation via a proportional hazards model with time-dependent effects. How
do you suggest modifying such a model if the risk from exposure rather than
falling back to the reference group once exposure is removed is believed to
be cumulative?
2. Write down a conditional logistic model in which we adjust for both age and
cohort effects where cohorts are grouped by intervals of births from 1930-35,
1936-40, 1940-45, etc. For such a model is it possible to answer the question:
was there a peak in relative risk during the nineteen sixties?
3. Consider a study of an industrial risk factor such as asbestos exposure. How

might the presence of non-proportional hazards manifest itself in such a
setting. Argue for some plausible situations. How might this be checked?
4. An investigation into the potential carcinogenic properties of an industrial

compound has evidence to suggest that it may influence the rate of some
recurrent secondary risk factor but little evidence that it would directly
impact the main outcome. An example might be recurrent liver disease,
a risk factor for liver cancer, but no direct influence on liver cancer itself.
Describe this situation via the use of compartment models. Complete the
description via the use of regression models. What assumptions do you need
to make? How can you go about testing the validity of those assumptions?
5. A biological marker is believed to be raised in the presence of tumor activity.

Studies on the marginal distribution of this marker show it to have a highly
skewed distribution. Also, there is a lower threshold below which values can-
not be detected. What kind of model might be of value in trying to detect
and quantify the role of this marker in cancer incidence.
6. Simulate the number of cases per family in the following way. First choose
family size, n, based on a mixture of two Poisson laws: one with a mean equal
to 6 and one with a mean equal to 20 and where P (20) = 1 − P (8) = 0.1.
Next, use a Binomial law B(n, p), where p = 0.15, to simulate the number of
observed cases. Note down this number and repeat for 200 families. Finally,
produce a histogram of the observed number of cases per family; 0, 1, 2 ...
7. For the data obtained from the previous exercise, ask a classmate or colleague
to look at the data and, in the absence of any further analysis, to judge
whether or not there appears to be evidence of a family effect. Do they
consider this effect to be absent or weak, moderate to strong, or very strong?
Do not provide any explanation as to how the data were obtained.
8. If given data like that of the previous question and no indication as to how
the data were obtained how might you go about testing a null hypothesis—
there is no family association. What would be the nature of the statistic
that would lead to a consistent test while maintaining power against the
alternative hypothesis.
9. Imagine that a friend, with no knowledge of epidemiology or statistics, is

told that they have a 90% lifetime probability of suffering some event. Is it
possible to explain to them just what this means? If not, why not? If so,
then what meaning—understandable to such a friend—could be furnished.
Chapter 6
Non-proportional hazards models
6.1 Chapter summary
The most general model, described in Chapter 4 covers a very broad spread of
possibilities and, in this chapter, we consider some special cases. Proportional
hazards models, partially proportional hazards models (O’Quigley and Stare,
2002), stratified models, or models with frailties or random coefficients all arise
as special cases of this model (Xu and O’Quigley, 2000). One useful parame-
terization (O’Quigley and Pessione, 1991; O’Quigley and Prentice, 1991) can
be described as a non-proportional hazards model with intercept. Changepoint
models are a particular form of a non-proportional hazards model with intercept
(O’Quigley and Natarajan, 2004). Any model can be viewed as a special case of
the general model, lying somewhere on a conceptual scale between this general
model and the most parametric extreme, which would be the simple exponential
model. Models can be placed on this scale according to the extent of model con-
straints and, for example, a random effects model would lie somewhere between
a stratified model and a proportional hazards model.
It is very easy to extend the simple proportional hazards model to deal with
much more general problems. Allowing regression effects, β(t), to change with
time introduces great generality to our model structure. We can then view the
proportional hazards structure as a special case of a non-proportional hazards
structure. Intermediary cases, where some parameters do not depend on time,
and some do, is also very easily accommodated. The fact that so many differ-
ent model structures can be all put under the same heading brings two major
benefits. The first is that a general understanding of the overall model struc-

119
https://doi.org/10.1007/978-3-030-33439-0_6
120 Chapter 6. Non-proportional hazards models
ture allows a better understanding of the specific structure of the special cases.
The second benefit of this generality is that we only need to tackle inferential
questions in the broadest setting. Applications to special cases are then almost
immediate. An important result is that any non-proportional hazards model can,
under a suitable transformation, be made into a proportional hazards one. The
transformation depends on the unknown regression function, β(t), and so is of
no obvious practical value. However, since we have optimality results for the
proportional hazards setting, we can use this result, in a theoretical framework,
to investigate for instance how well any test performs when contrasted to the
(generally unavailable) optimal test.
6.3 Partially proportional hazards models
In the case of a single binary variable, model (4.2) and model (4.4) represent
the two extremes of the modeling options open to us. Under model (4.2) there
would be no model constraint and any consequent estimation techniques would
amount to dealing with each level of the variable independently. Under model
(4.4) we make a strong assumption about the nature of the relative hazards,
an assumption that allows us to completely share information between the two
levels. There exists an important class of models lying between these extremes
and, in order to describe this class, let us now imagine a more complex situation;
that of three groups, A, B, and C, identified by a vector Z of binary covariates;
Z = (Z2 , Z3 ). This is summarized in Table 6.1.
Z2 Z3 Log of group effect Relative risk wrt A

Group A 0 0 0 1
Group B 1 0 β2 exp(β2 )
Group C 1 1 β2 + β3 exp(β2 + β3 )
Table 6.1: Coding for three groups. Impact on relative risks.
We are mainly interested in a treatment indicator Z1 , mindful of the fact

that the groups themselves may have very different survival probabilities. Under
model (4.4) we have
λ(t|Z) = λ0 (t) exp{β1 Z1 + β2 Z2 + β3 Z3 }. (6.1)
Our assumptions are becoming stronger in that not only are we modeling the
treatment affect via β1 but also the group effects via β2 and β3 . Expressing this
problem in complete generality, i.e., in terms of model (4.2), we write
λ(t|Z) = λ0 (t) exp{β1 (t)Z1 + β2 (t)Z2 + β3 (t)Z3 }. (6.2)

6.3. Partially proportional hazards models 121
Unlike the simple case of a single binary variable where our model choices were
between the two extremes of model (4.2) and model (4.4), as the situation
becomes more complex, we have open to us the possibility of a large number
of intermediary models. These are models that make assumptions lying between
model (4.2) and model (4.4) and, following O’Quigley and Stare (2002) we call
them partially proportional hazards models. A model in between (6.1) and (6.2) is
λ(t|Z) = λ0 (t) exp{β1 Z1 + β2 (t)Z2 + β3 (t)Z3 }. (6.3)
Alive in Stratum 1 Alive in Stratum 2 Alive in Stratum 3
Dead
Figure 6.1: A stratified model with transitions only to death state. For all 3
transitions the log-relative risk is given by β. The base rates, however, λ0w , w =
1, 2, 3 depend on the stratum w. Risk sets are stratum specific.
This model is of quite some interest in that the strongly modeled part of
the equation concerns Z1 , possibly the major focus of our study. Figure 6.1
illustrates a simple situation. The only way to leave any state is to die, the
probabilities of making this transition varying from state to state and the rates
of transition themselves depending on time. Below, under the heading time-
dependent covariates, we consider the case where it is possible to move within
states. Here it will be possible to move from a low-risk state to a high-risk state,
to move from either to the death state, but to also, without having made the
transition to the absorbing state, death, to move back from high-risk to low-risk.
Stratified models
Coming under the heading of a partially proportional hazards model is the class
of models known as stratified models. In the same way these models can be
considered as being situated between the two extremes of Equation 4.2 and
Equation 4.3 and have been discussed by Kalbfleisch and Prentice (2002) among
others. Before outlining why stratified models are simply partially proportional
hazards models we recall the usual expression for the stratified model as
λ(t|Z(t), w) = λ0w (t) exp{βZ(t)}, (6.4)

where w takes integer values 1, . . . , m. If the coefficient β were allowed to depend

on each stratum, indicated by w, say β(w), then this would exactly correspond
to a situation in which we consider each stratum independently, i.e., we have
independent models for each stratum. This would be nothing more than m sepa-
rate, independent, proportional hazards models. The estimation of β(w) for one
model has no impact on the estimation of β(w) for another. If we take β to
be common to the different strata, which is of course the whole purpose of the
stratified model, then, using data, whatever we learn about one stratum tell us
something about the others. They are no longer independent of one another.
Stratified models are necessarily broader than (4.3), lying, in the precise sense
described below, between this model and the non-proportional hazards model
(4.2). To see this, consider a restricted case of model (4.2) in which we have two
binary covariates Z1 (t) and Z2 (t). We put the restriction on the coefficient β2 ,
constrained to be constant in time. The model is then
λ{t|Z1 (t), Z2 (t)} = λ0 (t) exp{β1 (t)Z1 (t) + β2 Z2 (t)}, (6.5)
a model clearly lying, in a well-defined way, between models (4.3) and (4.2). It
follows that
λ{t|Z1 (t) = 0, Z2 (t)} = λ0 (t) exp{β2 Z2 (t)}
and
λ{t|Z1 (t) = 1, Z2 (t)} = λ∗0 (t) exp{β2 Z2 (t)},
where λ∗0 (t) = λ0 (t)eβ1 (t) . Recoding the binary Z1 (t) to take the values 1 and
2, and rewriting λ∗0 (t) = λ02 (t), λ0 (t) = λ01 (t) we recover the stratified PH
model (6.4) for Z2 (t). The argument is easily seen to be reversible and readily
extended to higher dimensions so we can conclude an equivalence between the
stratified model and the partially proportional hazards model in which some
of the β(t) are constrained to be constant. We can exploit this idea in the
goodness of fit or the model construction context. If a PH model holds as a
good approximation, then the main effect of Z2 say, quantified by β2 , would be
similar over different stratifications of Z1 and remain so when these stratifications
are re-expressed as a PH component to a two-covariate model. Otherwise the
indication is that β1 (t) should be allowed to depend on t. The predictability of
any model is studied later under the headings of explained variation and explained
randomness and it is of interest to compare the predictability of a stratified model
and an un-stratified one. For instance, we might ask ourselves just how strong is
the predictive strength of Z2 after having accounted for Z1 . Since we can account
for the effects of Z1 either by stratification or by its inclusion in a single PH model
we may obtain different results. Possible discrepancies tell us something about
our model choice.
The relation between the hazard function and the survival function follows
as a straightforward extension of (4.5). Specifically, we have

S(t|Z) = φ(w){S0w (t)}exp(βZ) , (6.6)
w
where S0w (t) is the corresponding baseline survival function in stratum w and
φ(w) is the probability of coming from that particular stratum. This is then
slightly more involved than the nonstratified case in which, for two groups, the
model expresses the survival function of one group as a power transformation
of the other. The connection to the class of Lehmann alternatives is still there
although somewhat weaker. For the stratified model, once again the quantity
λ0w (t) does not appear in the expression for the partial likelihood given now by
n
δi
exp(βZi )
L(β) = n (6.7)
j=1 Yj {wi (Xi ), Xi } exp(βZj )
i=1
and, in consequence, once again, λ0w (t) can remain arbitrary. Note also that
each term in the product is the conditional probability that at time Xi of an
observed failure, it is precisely individual i who is selected to fail, given all the
individuals at risk from stratum w and that one failure from this stratum occurs.
The notation wi (t) indicates the stratum in which the subject i is found at
time t. Although we mostly consider wi (t) which do not depend on time, i.e.,
the stratum is fixed at the outset and thereafter remains the same, it is almost
immediate to generalize this idea to time dependency and we can anticipate the
later section on time-dependent covariates where the risk indicator Yj {wi (t), t}
is not just a function taking the value one until it drops at some point to zero,
but can change between zero and one with time, as the subject moves from one
stratum to another. For now the function Yj {wi (t), t} will be zero unless the
subject is at risk of failure from stratum wi , i.e., the same stratum in which the
subject i is to be found. Taking the logarithm in (6.7) and derivative with respect
to β, we obtain the score function
n
n
j=1 Yj {wi (Xi ), Xi }Zj exp(βZj )
U (β) = δi Zi − n , (6.8)
i=1 j=1 Yj {wi (Xi ), Xi } exp(βZj )
which, upon setting equal to zero, can generally be solved without difficulty using
standard numerical routines, to obtain the maximum partial likelihood estimate
β̂ of β. The parameter β then is assumed to be common across the different
strata.
Inferences about β are made by treating β̂ as asymptotically normally dis-
tributed with mean β and variance I(β̂)−1 , where, now, I(β) is given by

I(β) = n i=1 δi Ii (β). In this case each Ii is, as before, obtained as the derivative
of each component to the score statistic U (β). For the stratified score this is
n 2
n 2
j=1 Yj {wi (Xi ), Xi }Zj exp(βZj ) j=1 Yj {wi (Xi ), Xi }Zj exp(βZj )
Ii = n − n .
j=1 Yj {wi (Xi ), Xi } exp(βZj ) j=1 Yj {wi (Xi ), Xi } exp(βZj )
The central notion of the risk set is once more clear from the above expressions
and we most usefully view the score function as contrasting the observed covari-
ates at each distinct failure time with the means of those at risk from the same
stratum. A further way of looking at the score function is to see it as having put
the individual contributions on a linear scale. We simply add them up within a
stratum and then, across the strata, it only remains to add up the different sums.
Once again, inferences can also be based on likelihood ratio methods or on the
score U (β), which in large samples can be considered to be normally distributed
with mean zero and variance I(β).
Multivariate extensions follow as before. For the stratified model the only
important distinction impacting the calculation of U (β) and Ii (β) is that the
sums are carried out over each stratum separately and then combined at the
end. The indicator Yj {wi (Xi )} enables this to be carried out in a simpler way as
indicated by the equation. The random effects model has proved to be a valuable
tool in applications and this can be seen in several detailed applications (Binder,
1992; Carlin and Hodges, 1999; Collaboration, 2009; Natarajan and O’Quigley,
2002; O’Quigley and Stare, 2002; Zhou et al., 2011; Hanson, 2012).
Random effects and frailty models

Also coming under the heading of partially proportional hazards model are the
classes of models, which include random effects. When the effects concern a single
individual such models have been given the heading frailty models (Vaupel et al.,
1979) since, for an individual identified by w, we can write λ0w (t) = αw λ0 (t)
implying a common underlying hazard λ0 (t) adjusted to each individual by a
factor, the individual’s frailty, unrelated to the effects of any other covariates
that are quantified by the regression coefficients. The individual effects are then
quantified by the αw .
Although of some conceptual interest, such models are indistinguishable from
models with time-dependent regression effects and therefore, unless there is some
compelling reason to believe (in the absence of frailties) that a proportional
hazards model would hold, it seems more useful to consider departures from
proportional hazards in terms of model (4.2). On the other hand, random effects
models, as commonly described by Equation (6.9) in which the αw identifies a
potentially large number of different groups, are interesting and potentially of
use. We express these as
λ(t|Z(t), w) = αw λ0 (t) exp{βZ(t)}. (6.9)
These models are also partially proportional in that some effects are allowed not
to follow a proportional hazards constraint. However, unlike the stratified models
described above, restrictions are imposed. The most useful view of a random
effects model is to see it as a stratified model with some structure imposed upon
the strata. A random effects model is usually written
λ(t|Z(t), w) = λ0 (t) exp{βZ(t) + w}, (6.10)
in which we take w as having been sampled from some distribution G(w; θ).
Practically there will only be a finite number of distinct values of w, however
large. For any value w we can rewrite λ0 (t)ew = λ0w (t) and recover model (6.4).
For the right-hand side of this equation, and as we might understand from (6.4),
we suppose w to take the values 1,2, ... The values on the left-hand side, being
generated from G(·) would generally not be integers but this is an insignificant
notational issue and not one involving concepts. Consider the equation to hold. It
implies that the random effects model is a stratified model in which added struc-
ture is placed on the strata. In view of Equation 6.5 and the arguments following
this equation we can view a random effects model equivalently as in Equation
4.2 where, not only are PH restrictions imposed on some of the components of
β(t), but the time dependency of the other components is subject to constraints.
These latter constraints, although weaker than imposing constancy of effect, are
all the stronger as the distribution of G(w; θ) is concentrated. In applications it
is common to choose forms for G(w; θ) that are amenable to ready calculation
(Duchateau and Janssen, 2007; Gutierrez, 2002; Hanagal, 2011; Li and Ryan,
2002; Liu et al., 2004; Wienke, 2010).
Structure of random effects models

Consider firstly the model of Equation 4.3. Suppose we have one main variable,
possibly a treatment variable of interest, coded by Z1 = 0 for group A and Z1 = 1
for group B. The second variable, say a center variable, which may or may not
have prognostic importance and for which we may wish to control for possible
imbalance is denoted by Z2 . A strong modeling approach would include both
binary terms in the model so that the relationship between the hazard functions
is as described in the left-hand side of Figure 6.2. If our main focus is on the effect
of treatment, believed to be comparable from one center to another, even though
the effects of the centers themselves are not absent, it makes sense to stratify.
This means that we do not attempt to model the effects of the centers but,
instead, remove any such potential effects from our analysis. This is nice in that
it allows for rather greater generality than that illustrated in the left-hand side
of Figure 6.2. We maintain an assumption of constant treatment effect but the
center effects can be arbitrary. This is illustrated on the right-hand side of Figure
Figure 6.2: An illustration of the impact of a proportional hazards assumption

and a partially proportional hazards assumption (stratified model (b)). Base rate
depends on stratum while regression coefficient β is the same.
Figure 6.3: An illustration on the impact of a proportional hazards assumption

and a partially proportional hazards assumption (random effects model). Stratum
effect is sampled from some distribution, G(w). Regression effect (β) common
across strata
6.2. The illustration makes it clear that, under the assumption, a weaker one
than that implied by Equation 4.3, we can estimate the treatment effect whilst
ignoring center effects. A study of these figures is important to understanding
what takes place when we impose a random effects model as in Equation 6.10.
For many centers, Figure 6.3, rather than having two curves per center, parallel
but otherwise arbitrary, we have a family of parallel curves. We no longer are able
to say anything about the distance between any given centers, as we could for the
model of Equation 4.3, a so-called fixed effects model, but the distribution of the
distances between centers is something we aim to quantify. This is summarized
by the distribution G(w; θ) and our inferences are then partly directed at θ.
Random effects models versus stratified models

The stratified model is making weaker assumptions than the random effects
model. This follows since the random effects model is just a special case of a
stratified model in which some structure is imposed upon the differences between
strata. The stratified model not only leaves any distribution of differences between
strata unspecified, but it also makes no assumption about the form of any given
stratum. Whenever the random effects model is valid, then so also is the stratified
model, the converse not being the case.
It may then be argued that we are making quite a strong assumption when we
impose this added structure upon the stratified model. In exchange we would hope
to make non-negligible inferential gains, i.e., greater precision of our estimates
of errors for the parameters of main interest, the treatment parameters.
In practice gains tend to be small for most situations and give relatively little
reward for the extra effort made. Since any such gains are only obtainable under
the assumption that the chosen random effects model generates the data, actual
gains in practice are likely to be yet smaller and, of course, possibly negative
when our additional model assumptions are incorrect. A situation where gains
for the random effects model may be of importance is one where a non-negligible
subset of the data include strata containing only a single subject. In such a case
simple stratification would lose information on those subjects. A random effects
model, assuming the approximation to be sufficiently accurate, enables us to
recover such information.
Efficiency of random effects models

Most of our discussion here focuses on different possible representations of the
infinitely complex reality we are hoping to model. Our purpose in modeling is,
ultimately, to draw simple, at least clear-cut, inferences. The question of inference
no longer concerns the general but rather the specific data set we have at hand.
If our main concern is on estimating risk functions then the question becomes,
to what extent do we gain by including in our inferential setup the presence of
random effect terms. Since our main objective is estimation and quantification of
regression parameters enabling us to say something about the risk factors under
study, the idea behind the inclusion of additional random effect terms is to make
more precise this estimation and quantification.
As already argued above the inclusion of individual random effects (frailties)
is of no practical interest and simply amounts to expressing the idea, albeit in
an indirect way, of model inadequacy (O’Quigley and Stare, 2002). We therefore
assume that we are dealing with groups, some of which, but not all, may only
include an isolated individual. We know that a partial likelihood analysis, stratified
by group, is estimating the same regression parameter. Inference is based on
the stratified score statistic. We contrast the observed covariate value with its
estimated expectation under the model. Different model assumptions will impact
this estimated expectation and it is here that any efficiency gains can be made.
For a stratified model, these estimated expectations may be with respect to
relatively small risk sets. A random effects model on the other hand, via the
100 × 5 250 × 2 25 × 20
Ignoring effect 0.52 (0.16) 0.51 (0.16) 0.54 (0.16)
Random effect model 1.03 (0.19) 0.99 (0.22) 1.01 (0.17)
Stratified model 1.03 (0.22) 1.02 (0.33) 1.01 (0.18)
Table 6.2: Simulations for three models under different groupings.
inclusion of a different w per group, will estimate the relevant expectations over
the whole risk set and not just that relative to the group defined by the covariate
value.
Comparisons for the stratified model are made with respect to the relatively
few subjects of the group risk sets. This may lead us to believe that much infor-
mation could be recovered were we able to make the comparison, as does the
alternative random effects analysis, with respect to the whole risk set. Unfortu-
nately this is not quite so because each contribution to the score statistic involves
a difference between an observation on a covariate and its expectation under the
model and the “noise” in the expectation estimate is of lower order that the
covariate observations themselves. There is not all that much to be gained by
improving the precision of the expectation estimate.
In other words, using the whole of the risk set or just a small sample from
it will provide similar results. This idea of risk set sampling has been studied in
epidemiology and it can be readily seen that the efficiency of estimates based on
risk set samples of size k, rather than the whole risk set, is of the order
⎧ ⎫
k ⎨ ⎬
n
1
1+ . (6.11)
k+1 ⎩ n(n − j + 1) ⎭
j=1
This function increases very slowly to one but, with as few as four subjects on
average in each risk set comparison, we have already achieved 80% efficiency.
With nine subjects this figure is close to 90%. Real efficiency will be higher for
two reasons: (1) the above assumes that the estimate based on the full risk set
is without error, (2) in our context we are assuming that each random effect w
is observed precisely.
Added to this is the fact that, since the stronger assumptions of the random
effects model must necessarily depart to some degree from the truth, it is by no
means clear that there is much room to make any kind of significant gains. As
an aside, it is of interest to note that, since we do not gain much by considering
the whole of the risk set as opposed to a small sample from it, the converse
must also hold, i.e., we do not lose very much by working with small samples
rather than the whole of the risk set. In certain studies, there may be great
economical savings made by only using covariate information, in particular when
time dependent, from a subset of the full risk set.
6.4. Partitioning of the time axis 129
Table 6.2 was taken O’Quigley and Stare (2002). The table was constructed
from simulated failure times where the random effects model was taken to be
exactly correct. Data were generated from this model in which the gamma frailty
had a mean and variance equal to one. The regression coefficient of interest
was exactly equal to 1.0. Three situations were considered; 100 strata each of
size 5, 250 strata each of size 2, and 25 strata each of size 20. The take-home
message from the table is that, in these cases for random effects models, not
much is to be gained in terms of efficiency. Any biases appear negligible and the
mean of the point estimates for both random effects and stratified models, while
differing notably from a crude model ignoring model inadequacy, are effectively
indistinguishable. As we would expect there is a gain for the variance of estimates
based on the random effects model but, even for highly stratified data (100 × 5),
any gain is very small. Indeed for the extreme case of 250 strata, each of size
2, surely the worst situation for the stratified model, it is difficult to become
enthusiastic over the comparative performance of the random effects model.
We might conclude that we only require around 80% of the comparative
sample size needed for estimating relative risk based on the stratified model.
But, such a conclusion, leaning entirely on the assumption that we know not
only the class of distributions from which the random effects come but also the
exact value of the population parameters, suggests, in practice, that the hoped
for gain, in this most hopeful of cases, is more likely to be greater than the
80% indicated by our calculations. The only real situation that can be clearly
disadvantageous to the stratified model is one where a non-negligible subset of
the strata are seen to only contain a single observation. For such cases, and
assuming a random effects model to provide an adequate fit, information from
states with a single observation (which would be lost by a stratified analysis) can
be recovered by a random effects analysis.
6.4 Partitioning of the time axis
Recalling the general model, i.e., the non-proportional hazards model for which
there is no restriction on β(t), note that we can re-express this so that the
function β(t) is written as a constant term, the intercept, plus some function of
time multiplied by a constant coefficient. Writing this as
λ(t|Z) = λ0 (t) exp{[β0 + θQ(t)]Z}, (6.12)
we can describe the term β0 as the intercept and Q(t) as reflecting the nature
of the time dependency. The coefficient θ will simply scale this dependency and
we may often be interested in testing the particular value, θ = 0, since this value
corresponds to a hypothesis of proportional hazards. Fixing the function Q(t)
to be of some special functional form allows us to obtain tests of proportional-
ity against alternatives of a particular nature. Linear or quadratic decline in the
log-relative risk, changepoint, and crossing hazard situations are all then easily
accommodated by this simple formulation. Tests of goodness of fit of the pro-
portional hazards assumption can be then be constructed which may be optimal
for certain kinds of departures.
Although not always needed it can sometimes be helpful to divide the time
axis into r non-overlapping intervals, B1 , . . . , Br in an ordered sequence beginning
at the origin. In a data-driven situation these intervals may be chosen so as to
have a comparable number of events in each interval or so as not to have too
few events in any given interval. Defined on these intervals is a vector, also of
dimension r, of some known or estimable functions of time, not involving the
parameters of interest, β. This is denoted Q(t) = {Q1 (t), . . . , Qr (t)} This model
is then written in the form
λ(t|Z) = λ0 (t) exp{[β + θQ(t)]Z}, (6.13)
where θ is a vector of dimension r. Thus, θQ(t) (here the usual inner product) has
the same dimension as β, i.e., one. In order to investigate the time dependency of
particular covariates in the case of multivariate Z we would have β of dimension
greater than one, in which case Q(t) and θ are best expressed in matrix notation
(O’Quigley and Pessione, 1989).
Here, as through most of this text, we concentrate on the univariate case since
the added complexity of the multivariate notation does not bring any added light
to the concepts being discussed. Also, for the majority of the cases of interest,
r = 1 and θ becomes a simple scalar. We will often have in mind some partic-
ular form for the time-dependent regression coefficient Q(t), common examples
being a linear slope (Cox, 1972), an exponential slope corresponding to rapidly
declining effects (Gore et al., 1984) or some function related to the marginal
distribution, F (t) (Breslow et al., 1984). In practice we may be able to estimate
this function of F (t) with the help of consistent estimates of F (t) itself, in par-
ticular the Kaplan-Meier estimate. The non- proportional hazards model with
intercept is of particular use in questions of goodness of fit of the proportional
hazards model pitted against specific alternatives. These specific alternatives can
be quantified by appropriate forms of the function Q(t). We could also test a
joint null hypothesis H0 : β = θ = 0 corresponding to no effect, against an alter-
native H1 , either θ or β nonzero. This leads to a test with the ability to detect
non-proportional hazards, as well as proportional hazards departures to the null
hypothesis of no effect. We could also test a null hypothesis H0 : θ = 0 against
H1 : θ = 0, leaving β itself unspecified. This would then provide a goodness of
fit test of the proportional hazards assumption. We return to these issues later
on when we investigate in greater detail how these models give rise to simple
goodness of fit tests.
6.5. Time-dependent covariates 131
Changepoint models
A simple special case of a non-proportional hazards model with an intercept is
that of a changepoint model. O’Quigley and Pessione (1991), O’Quigley (1994)
and O’Quigley and Natarajan (2004) develop such models whereby we take the
function Q(t) to be defined by, Q(t) = I(t ≤ γ) − I(t > γ) with γ an unknown
changepoint. This function Q(t) depends upon γ but otherwise does not depend
upon the unknown regression coefficients and comes under the above heading
of a non-proportional hazards model with an intercept. For the purposes of a
particular structure for a goodness of fit test we can choose the intercept to be
equal to some fixed value, often zero (O’Quigley and Pessione, 1991). The model
is then
λ(t|Z) = λ0 (t) exp{[β + αQ(t)]Z(t)}. (6.14)
The parameter α is simply providing a scaling (possibly of value zero) to the
time dependency as quantified by the function Q(t). The chosen form of Q(t),
itself fixed and not a parameter, determines the way in which effects change
through time; for instance, whether they decline exponentially to zero, whether
they decline less rapidly or any other way in which effects might potentially
change through time.
Inference for the changepoint model is not straightforward and in the later
sections dealing with inference we pay particular attention to some of the diffi-
culties raised. Note that were γ to be known, then inference would come under
the usual headings with no additional work required. The changepoint model
expressed by Equation 6.14 deals with the regression effect changing through
time and putting the model under the heading of a non-proportional hazards
model. A related, although entirely different model, is one which arises as a sim-
plification of a proportional model with a continuous covariate and the idea is to
replace the continuous covariate with a discrete classification.
The classification problem itself falls into two categories. If we are convinced
of the presence of effects and simply wish to derive the most predictive clas-
sification into, say, two groups, then the methods using explained randomness
or explained variation will achieve this goal. If, on the other hand, we wish to
test a null hypothesis of absence of effect, and, in so doing, wish to consider all
possible classifications based on a family of potential cutpoints of the continuous
covariate, then, as mentioned above, special techniques of inference are required.
We return to these questions in later chapters where they are readily addressed
via the regression effect process.
6.5 Time-dependent covariates
In all of the above models we can make a simple change by writing the covariate
Z as Z(t), allowing the covariate to assume different values at different time
points. Our model then becomes
State 1 State 2
No symptoms Progression
State 3
Dead
Figure 6.4: Compartment model where ability to move between states other than
death state can be characterized by time-dependent indicator covariates Z(t).
Any paths not contradicting arrows are allowed.
λ(t|Z(t)) = λ0 (t) exp{β(t)Z(t)} (6.15)
and allows situations such as those described in Figure 6.4 to be addressed. As

we change states the intensity function changes. This enables us to immediately
introduce further refinement into a simple alive/dead model whereby we can sup-
pose one or more intermediary states. A subject can move across states thereby
allowing prognosis to improve or to worsen, the rates of these changes themselves
depending upon other factors. The state death is described as an absorbing state
and so we can move into this state but, once there, we cannot move out of it
again.
Mostly we will work with the proportional hazard restriction on the above
model so that
λ(t|Z(t)) = λ0 (t) exp{βZ(t)}, (6.16)
Such a simple, albeit very much more sophisticated, model than our earlier one
describes a broad range of realistic situations. We will see that models with time-
dependent covariates do not raise particular difficulties, either computationally
or from the viewpoint of interpretation, when we deal with inference. The model
simply says that the effect of the covariate remains constant, i.e., the regression
coefficient remains constant, but that the covariate, or state, can itself change
with time. Models with time-dependent covariates can also be used as a purely
artificial construction in order to be able to express non-proportional hazards
models in a proportional hazards form.
We can also imagine a slightly more involved situation than the above. Sup-
pose that the covariate Z remains fixed, but that a second covariate, known
to influence survival, also needs to be accounted for. Furthermore this second
covariate is time dependent. We could, of course, simply use the above model
extended to the case of two covariates. This is straightforward, apart from the
fact that, as previously underlined by the complexity theorem, care is needed.
If, however, we do not wish to model the effects of this second covariate, either
because it is only of indirect concern or because its effects might be hard to
model, then we could appeal to a stratified model. We write:
λ(t|Z(t), w(t)) = λ0w(t) (t) exp{βZ(t)}, (6.17)
where, as for the non-time-dependent case, w(t) takes integer values 1, ..., m
indicating status. The subject can move in and out of the m strata as time
proceeds. Two examples illustrate this. Consider a new treatment to reduce the
incidence of breast cancer. An important time-dependent covariate would be
the number of previous incidents of benign disease. In the context of inference,
the above model simply means that, as far as treatment is concerned, the new
treatment and the standard are only ever contrasted within patients having the
same previous history. These contrasts are then summarized in final estimates
and possibly tests. Any patient works her way through the various states, being
unable to return to a previous state. The states themselves are not modeled. A
second example might be a sociological study on the incidence of job loss and
how it relates to covariates of main interest such as training, computer skills, etc.
Here, a stratification variable would be the type of work or industry in which the
individual finds him or herself. Unlike the previous example a subject can move
between states and return to previously occupied states.
Time-dependent covariates describing states can be used in the same way
for transition models in which there is more than one absorbing “death” state.
Many different kinds of situations can be constructed, these situations being
well described by compartment models with arrows indicating the nature of the
transitions that are possible (Figure 6.5). For compartment models with time-
dependent covariates there is a need for some thought when our interest focuses
on the survival function. The term external covariate is used to describe any
covariate Z(t) such that, at t = 0, for all other t > 0, we know the value of
State 1 State 2
State 3 State 4
Figure 6.5: Compartment model with 2 absorbing “death” states. State 4 cannot
be reached from State 3 and can only be reached from State 1 indirectly via
State 2.
Z(t). The paths can be described as deterministic. In the great majority of the
problems that we face this is not the case and a more realistic way of describing
the situation is to consider the covariate path Z(t) to be random. Also open to
us as a modeling possibility, when some covariate Z1 (t) is of secondary interest
assuming a finite number of possible states, is to use the at-risk function Y (s, t).
This restricts our summations to those subjects in state s as described above for
stratified models.
Time-dependent covariates and non-PH models

A non-proportional hazard model with a single constant covariate Z is written
λ(t|Z) = λ0 (t) exp{β(t)Z}. (6.18)
The multivariate extension is immediate and, in keeping with our convention

of only dealing with the univariate problem whenever possible, we focus our
attention on the simple product β(t)Z at some given point in time t. If we define
β0 = β(0) = 0 , we can rewrite this product as β0 Z(t) where Z(t) = Zβ(t)/β0 . We
could take any other time point t and, once again, we observe that we can rewrite
the product β(t )Z as β0 Z(t ) where Z(t ) = Zβ(t )/β0 . This equivalence is then
true for any, and all, values of t. Thus, a non-proportional hazards model with
a constant covariate can be re-expressed, equivalently, as a simple proportional
hazards model with a time-dependent covariate.
It is almost immediate, and perhaps worth carrying out as an exercise, to
show that we can reverse these steps to conclude also that any model with
time-dependent covariates can be expressed in an equivalent form as a non-
proportional hazards model. In conclusion, for every non-proportional hazards
model there exists an equivalent proportional hazards model with time-dependent
covariates. Indeed, it also clear that this argument can be extended. For, if we
have the model: λ(t|Z) = λ0 (t) exp{β(t)Z(t)}, then, via a re-expression of the
model using β0 Z ∗ (t) where Z ∗ (t) = Z(t)β(t)/β0 , we can construct a propor-
tional hazards model with a time-dependent regression effect from a model which
began with both time-dependent regression effects as well as time changing
regression coefficient. For reference purposes, it is worth putting this summary
in the form of a lemma.
Lemma 6.1. For given β(t) and covariate Z(t) there exists a constant β0
and time-dependent covariate Z ∗ (t) so that λ(t|Z(t)) = λ0 (t) exp{β(t)Z(t)} =
λ0 (t) exp{β0 Z ∗ (t)} .
The important thing to note is that we have the same λ0 (t) either side of
the equation and that, whatever the value of λ(t|Z(t)), for all values of t, these
values are exactly reproduced by either expression, i.e., we have equivalence.
This equivalence is a formal one and does not of itself provide any new angle
on model development. The function β(t) will not generally be known to us. This
equivalence may be exploited nonetheless in theoretical investigation. We can use

it to obtain the most powerful unbiased test in chosen situations. We can also
make use of the idea to adopt certain software, as well as user-written code, that
already caters for time-dependent covariables. Almost no extra work is needed to
use these programs should we wish to study particular types of non-proportional
hazards models characterized by different β(t).
We can end this section and round off connections to earlier sections by con-
sidering the likelihood. The sequential nature of these studies and the conditional
arguments that run throughout this work means—leaving aside at least for now
questions of computing—that no extra work is needed. We sequentially condition
on time (this enables us in later chapters to view everything from the angle of
stochastic processes) so that, at a given time t = Xi , all that is needed is infor-
mation available at that time point Xi . The likelihood and resulting estimating
equations look almost identical to what we had in the case of constant covari-
ates. In particular, the function λ0 (t) is not involved. The likelihood expression
is given by
n
δi
exp(βZi (Xi ))
L(β) = n , (6.19)
i=1 j=1 j (Xi ) exp(βZj (Xi ))
Y
As before, taking the logarithm in Equation 6.19 and its derivative with respect
to β, we obtain the estimating equation which, upon setting equal to zero,
can generally be solved without difficulty using the Newton-Raphson method.
Chapter 7 on estimating equations looks at this more closely. The form of U (β)
is as before with only the time dependency marking any distinction.
n
n
j=1 Yj (Xi )Zj (Xi ) exp(βZj (Xi ))
U (β) = δi Zi (Xi ) − n (6.20)
i=1 j=1 Yj (Xi ) exp(βZj (Xi ))
The residual interpretation of the contributions to the estimating equation

remains. In order for β̂ to behave as in the non-time-dependent case some
weak restrictions on the allowable covariate functions would be needed. We
leave this to later chapters, and treating β̂ to be asymptotically normally dis-
tributed with mean β and large sample variance I(β̂)−1 , where I(β) is, once
again, minus the second derivative of log L(β) with respect to β. The expression

for I(β) = n i=1 δi Ii (β) would be analogous to that given by Equation 4.8. We
might also keep in mind that Lemma 11.1 allows us to formally work with this
structure for time-dependent covariates in order to cater for non-proportional
hazards situations.
6.6 Linear and alternative model formulations
Throughout the history of survival analysis some researchers have preferred to

frame the many questions raised within the setting of linear models. Certainly
there is an advantage in having a very substantial amount of literature dealing
with the linear model in other circumstances. However, this advantage is more
of a theoretical than a practical one and is lost under the simple device of taking
logarithms. Linear models are simply less natural in the context of rates and
probabilities and, while we recall some of the main ideas here, our purpose is
more to do with completeness than a belief that linear models offer something not
adequately catered for by PH and NPH models. That said, we should always keep
our eye on parsimony principles. In some situations an additive model may provide
a more parsimonious description of the data. One example that comes to mind
arises in cardiovascular studies where, for certain risk factors, a constant additive
handicap for those at risk provides a simpler model than the more involved relative
risk one where there are complex time dependencies that have to be catered for.
Additive models
Instead of considering a model with multiplicative risks, some authors appeal to
a structure more familiar to the one we know from linear regression, e.g., the
additive intensity model proposed by Aalen (1989, 1980). This is written:
λ(t | Z) = β0 (t) + β T (t)Z(t), (6.21)
where the vector of regression coefficients, β(t) = (β0 (t), β1 (t), . . . , βp (t)) depends
t
on time. The estimation of the cumulative effects 0 βi (s)ds for i = 1, . . . , p is
done by weighted least squares. McKeague and Sasieni (1994) studied the case
of fixed effects for some covariates and time-dependent effects for the others
creating a model analogous to the partially proportional hazards model. Lin and
Ying (1995) proposed a more complex combined additive and multiplicative risk
model written:
λ(t | X, Z) = Y (t)g{β T X(t)} + λ0 (t)h{γ T Z(t)},
where (X(t), Z(t)) is a vector of p + q time-dependent covariables β ∈ Rp and

γ ∈ Rq and where, again, Y (t) is the indicator of being at risk at time t. The
functions h and g are link functions. For example, we can choose h(x) = exp(x)
or 1 + x and g(x) = x or exp(x). This model class aims to include both the
proportional hazard model as well as the additive model for particular values
of h and g. Scheike and Zhang (2002) investigated a further combination of
multiplicative and additive intensities with the model given by
λ(t | X, Z) = Y (t)β(t)T X(t) exp(γ T Z(t)),

6.6. Linear and alternative model formulations 137
where β(t) is time dependent. Aalen et al. (2008) presents an analysis of the
properties of these kinds of model. Beran (1981) and McKeague and Utikal
(1990) studied the model,
λ(t | Z) = Y (t)α(t, Z(t)), (6.22)
without making explicit the function α, which can be estimated by kernel methods
and local smoothing.
Linear transformation models

Elementary manipulation allows us to see that the proportional hazards model
with fixed covariables, Z, can be written
t
log {− log (S(t | Z)} = log λ0 (s)ds + β T Z,
0
where, as usual, λ0 is an arbitrary underlying risk function. Using this as a starting

point we can generalize in the following way:
h (S(t | Z)) = g(t) + β T Z, (6.23)
where h is a link function, g an unspecified function, both of which are strictly

increasing. Model (6.23) can be re-expressed as
g(T ) = −β T Z + ε, (6.24)
where ε is a random variable with distribution function 1 − h−1 . Equation (6.24)

describes a class of transformation models for T that results in a linear relation-
ship with Z (Cheng et al., 1995; Dabrowska and Doksum, 1988; Lin, 2013; Wei,
1992). Linear models have been very well studied and their properties are readily
anticipated. For this reason it can seem attractive to bring these non-linear mod-
els under a linear umbrella. One particular case is the proportional odds model
(Bennett, 1983a) which follows from fixing

x
h(x) = logit(x) = log .
1−x
Note that, for

1 − S(x | Z)
h(x) = log and g(t) = φ log(t),
S(x | Z)
we recover the log-logistic model (Bennett, 1983b). The transformation g = log,

produces a class of models referred to as accelerated failure time models. The
survival function, given Z, is given by

S(t | Z) = S0 t exp(β T Z) , 0≤t≤T, (6.25)
where S0 is an underlying survival function corresponding to the hazard λ0 . The

formula (6.25) shows that the covariates have a multiplicative effect on time,
which implies that they accelerate or slow down the death rate, according to the
sign of β T Z (Martinussen and Peng, 2013). The class of linearly transformed
models is more general than the proportional hazard model but is less flexible,
since it does not easily allow for the use of time-dependent covariates. Exten-
sions to accommodate other common situations such as multiple events is also
awkward. Finally, great care is needed when using mixture type models of this
nature in practice. It is frequently the case (Venzon and Moolgavkar, 1988) that
they fail to be invariant to simple changes in the coding. Coding the treatment
0, rather than 1, can have a significant impact on the p-value. There is no good
reason to choose a particular coding, say 1 for males and 0 for females rather
than the other way around, and it is more than problematic that the final result
would depend on that choice. The Yang and Prentice model, described in the
following section, also suffers from a lack of invariance to coding.
Yang and Prentice model

In the presence of a binary covariate Z describing the two-group problem, Yang
and Prentice (2005) proposed modeling the survival time in both groups by
θ1 θ0
λT (t) = λP (t), 0≤t≤T, (6.26)
θ1 + (θ0 − θ1 )SP (t)
where θ1 and θ0 are two positive parameters, SP is the survival function in

the first group, λP and λT are the hazard rate in the first and second groups,
respectively. The proportional hazards model (θ1 = θ0 ) and the proportional odds
model (θ0 = 1) are special cases. The model appears to be flexible allowing for
the possibility of modeling risks that cross over in time. This property can be
exploited as a means to obtain tests of greater generality than, say, the log-
rank test. The parameters of the model are such that θ1 = limt↓0 λT (t)/λP (t)
and θ0 = limt↑T λT (t)/λP (t). Thus θ1 represents the relative risk in the short
term and θ0 the long-term relative risk. The estimation in the model is done by
maximizing an equivalent of the model-specific partial likelihood. The authors
suggest the introduction of Z covariates by replacing the survival functions by
T T
ST (t)exp(β Z) and SP (t)exp(β Z) .
While these ideas are very interesting and provide a lot of insight into the
possible mechanisms that generate the observations, at the time of writing, the
test based on this model cannot be recommended for general use. There are two
reasons for this. The first, identified by Chauvel and O’Quigley (2014) is the poor
control on Type 1 error offered by the test. The good power properties are in
part a reflection of this. Indeed in several situations of proportional hazards the
test outperformed the log-rank test, something that is not theoretically possible.
So, some kind of adjustment is needed before too much reliance be put on the
calculated p-value. The second reason is rather more serious, not easily overcome,
and stems from the somewhat non-linear specification of the model. Different
codings will give different results. Coding the treatment variable (1,0) rather than
(0,1) will lead to different answers. Since there is often no natural way to code
this is very problematic. It is quite plausible, for example, that the same set of
observations allows us to conclude that the survival experience of women differs
significantly from that of men and that, simultaneously, the survival experience
of men does not significantly differ from that of women. This is clearly not a
coherent summary of the data and has to be fixed before we could have confidence
in any results based on this test. A simple potential solution would be to calculate
both p-values say and then take the minimum. This would solve the coherence
problem but would certainly exacerbate the difficulty in exercising good control
over Type 1 error. And in the situation of several levels, or several covariates, it
would seem to be very difficult to anticipate the operating characteristics in any
broad sense.
Nonetheless, there is something rather attractive in this formulation and it
would be worth the extra effort to fix this coherence problem. The root of the
difficulty would stem from the lack of symmetry in Equation 6.26 and this obser-
vation may open up a path toward a solution. Further discussion is given in
Flandre and O’Quigley (2020).
1. Consider the approach of partitioning the time axis as a way to tackle non-
proportional hazards. For example, we might choose a partition consisting
of 3 intervals, any one of which respects the proportional hazards constraint
with an interval-specific regression coefficient. Find the simplest way of
writing this model down. How might you express a null hypothesis of no
effect against an alternative of a non-null effect. How would this expression
change if, aside from the null, the only possibility is of an effect that
diminishes through time.
2. When biological or other measures are expensive to make, sampling from
the risk set can enable great savings to be made. Explain why this would
be the case and consider the criteria to be used to decide the size of the
risk set sampled. Look up and describe the case-cohort design which has
been successfully used in large epidemiological studies.
3. Show formally that a non-proportional hazards model with a constant
covariate is equivalent to a proportional hazards one with a time-dependent
covariate. Does any such result hold for a non-proportional hazards model
with a non-constant covariate.
4. Suppose, for a binary group indicator, that the observations are generated
by a linear model with a constant regression coefficient. Show that such a
model is equivalent to a proportional hazards model with a time-dependent
covariate. Conclude that a linear model is equivalent to a non-proportional
hazards one.
5. Consider a randomized clinical trial comparing 2 treatments. In order to

obtain sufficient recruitment a total of 25 medical centers participate in the
trial. The prognosis varies widely among centers and so center effects need
to be taken into account to avoid any real effects of the new treatment
being drowned by noise. As a statistician your opinion is asked. There are
3 possibly approaches: (1) make use of a proportional hazards model that
includes a term of dimension 24 to account for center effects, (2) make
use of a stratified proportional hazards model to account for center effects,
(3) make use of a random effects model to account for center effects, or
(4) ignore center effects and carry out a simple log-rank test for possible
differences in treatment outcomes. What advice would you give? And why?
6. Again consider the setting of a randomized clinical trial comparing 2 treat-

ments. Suppose that we base our analysis on a proportional hazards model
but that the mechanism generating the observations is in fact a non-
proportional hazards one where the effects decline through time. Describe
how inference will be impacted by this. Next, suppose that we assume a
broader non-proportional hazards model in which regression effects may
decline but may also remain constant. It turns out that the mechanism
generating the observations is a proportional hazards one. Describe how,
in this case, inference will be affected by our assumptions.
7. Show formally that, if our covariate space does not include continu-
ous covariates, then any true situation can be modeled precisely by a
non-proportional hazards model. Conclude that, for any arbitrary situa-
tion, including that of continuous covariates, we can postulate a non-
proportional hazards model that is arbitrarily close to that generating the
observations.
Chapter 7
Model-based estimating equations
7.1 Chapter summary
The regression effect process, described in Chapter 9, shapes our main approach
to inference. At its heart are differences between observations and their model-
based expectations. The flavor is very much that of linear estimating equations
(Appendix D.1). Before we study this process, we consider here an approach
to inference that makes a more direct appeal to estimating equations. The two
chapters are closely related and complement one another. This chapter leans
less heavily on stochastic processes and links in a natural and direct way to the
large body of theory available for estimating equations. Focusing attention on
the expectation operator, leaning upon different population models and different
working assumptions, makes several important results transparent. For example,
it is readily seen that the so-called partial likelihood estimator is not consistent for
average effect, E{β(T )}, under independent censoring and non-constant β(t).
One example we show, under heavy censoring, indicates the commonly used
partial likelihood estimate to converge to a value greater than 4 times its true
value. Linear estimating equations provide a way to investigate statistical behavior
of estimates for small samples. Several examples are considered.
This chapter and Chapter 9 provide the inferential tools needed to analyze survival
data. Either approach provides several techniques and results that we can exploit.
Taken together we have a broad array of methods, and their properties, that,
when used carefully, can help gain deep understanding of real datasets that arise
in the setting of a survival study.

141
https://doi.org/10.1007/978-3-030-33439-0_7
142 Chapter 7. Model-based estimating equations
The earlier chapter on marginal survival is important in its own right and we
lean on the results of that chapter throughout this work. We need to keep in mind
the idea of marginal survival for two reasons: (1) it provides a natural backdrop to
the ideas of conditional survival and (2), together with the conditional distribution
of the covariate given T = t, we are able to consider the joint distribution of
covariate and survival time T. A central concern is conditional survival, where
we investigate the conditional distribution of survival given different potential
covariate configurations, as well as variables such as time elapsed. More generally
we are interested in survival distributions corresponding to transitions from one
state to another, conditional on being in some particular state or of having
mapped out some particular covariate path. The machinery that will enable us
to obtain insight into these conditional distributions is that of proportional and
non-proportional hazards regression.
When we consider any data at hand as having arisen from some experi-
ment, the most common framework for characterizing the joint distribution of
the covariate Z and survival T is one where the distribution of Z is fixed and
known, and the conditional survivorship distribution is the subject of our infer-
ential endeavors. Certainly, this characterization would bring the model to most
closely resemble the experiment as set-up. However, in order to accommodate
censoring, it is more useful to characterize the joint distribution of Z and T via
the conditional distribution of Z given T = t and the marginal distribution of
T . This is one of the reasons why we dealt first with the marginal distribution
of T. Having dealt with that we can now focus our attention on the conditional
distribution of Z given T . We can construct estimating equations based on these
ideas and from these build simple tests or make more general inferences.
The main theorem of proportional hazards regression, introduced by O’Quigley
(2008), generalizes earlier results of Schoenfeld (1980), O’Quigley and Flandre
(1994), and Xu and O’Quigley (2000). The theorem has several immediate corol-
laries and we can use these to write down estimating equations upon which we
can then construct suitable inferential procedures for our models. The regression
effect process of subsequent chapters also has a clear connection to this theorem.
While a particular choice of estimating equation can result in high efficiency when
model assumptions are correct or close to being correct, other equations may be
less efficient but still provide estimates which can be interpreted when model
assumptions are incorrect. For example, when the regression function β(t) might
vary with time, we are able to construct an estimating equation, the solution of
which provides a consistent estimate of E{β(T )}, the average effect. The usual
partial likelihood estimate fails to achieve this, and the resulting errors, depend-
ing on the level of censoring, even when independent, can be considerable. This
can be seen in Table 7.1 where the resulting bias, for high censoring rates, can
be more than 400%. Most currently available software makes no account of this
and, given that a non-constant β(t) will be more the rule than the exception,
7.3. Likelihood solution for parametric models 143
we ought to be very cautious in interpreting the majority of estimates of average

effect, E{β(T )}, unless: (i) the function β(t) is close to being a constant or (ii)
the dataset contains little censoring.
7.3 Likelihood solution for parametric models
For almost any statistical model, use of the likelihood is usually the chosen
method for dealing with inference on unknown parameters. Bayesian inference,
in which prior information is available, can be viewed as a broadening of the
approach and, aside from the prior, it is again the likelihood function that will be
used to estimate parameters and carry out tests. Maximum likelihood estimates,
broadly speaking, have good properties and, for exponential families, a class to
which our models either belong or are close, we can even claim some optimality.
A useful property of maximum likelihood estimators of some parameter is that
the maximum likelihood estimator of some monotonic function of the parameter
is the same monotonic function of the maximum likelihood estimator of the
parameter itself. Survival functions themselves will often come under this heading
and so, once we have estimated parameters that provide the hazard rate, then we
immediately have estimates of survival. Variance expressions are also obtained
quite easily, either directly or by the approximation techniques of Appendix A.10.
Keeping in mind that our purpose is to make inference on the unknown regression
coefficients, invariant to monotonic increasing transformations on T, we might
also consider lesser used likelihood approaches such as marginal likelihood and
conditional likelihood. It can be seen that these kinds of approaches lead to
the so-called partial likelihood. In practice we will treat the partial likelihood as
though it were any regular likelihood, the justification for this being possible
through several different arguments.
For fixed covariates, in the presence of parametric assumptions concerning
λ0 (t), inference can be carried out on the basis of the following theorem that
simply extends that of Theorem 3.1. We suppose that the survival distribution
is completely specified via some parametric model, the parameter vector being
say θ. A subset of θ is a vector of regression coefficients, β, to the covariates in
the model. The usual working assumption is that of a conditionally independent
censoring mechanism, i.e., the pair (T, C) is independent given Z. This would
mean, for instance, that within any covariate grouping, T and C are independent
but that C itself can depend on the covariate. Such dependence would generally
induce a marginal dependency between C and T .
Theorem 7.1. Under a conditionally independent censoring mechanism the log-

likelihood log L(θ) can be written as log L(θ) = n
i=1 log Li (θ) where
log Li (θ) = I(δi = 1) log f (xi |zi ; θ) + I(δi = 0) log S(xi |zi ; θ). (7.1)
The maximum likelihood estimates obtain the values of θ, denoted θ̂, that
maximize log L(θ) over the parameter space. For log L(θ) a differentiable function
value is then the solution to the estimating equation U (θ) = 0 where
of θ, this
U (θ) = i ∂ log Li (θ)/∂θ. Next notice that, at the true value of θ, i.e., the value
which supposedly generates the observations, denoted θ0 , we have Var(U (θ0 )) =
E U 2 (θ0 ) = E I(θ0 ) where
n
n

2 2
I(θ) = Ii (θ) = −∂ log L(θ)/∂θ = − ∂ 2 log Li (θ)/∂θ2 .
i=1 i=1
As for likelihood in general, some care is needed in thinking about the meaning
of these expressions and the fact that the operators E(·) and Var(·) are taken
with respect to the distribution of the pairs (xi , δi ) but with θ0 fixed. The score
equation is U (θ̂) = 0 and the large sample variance is approximated by Var(θ̂) ≈
1/I(θ̂). Newton-Raphson iteration is set up by a simple application of the mean
value theorem so that
θj+1 = θj + I(θj )−1 U (θj ) , j ≥ 1, (7.2)
where θ1 is some starting value, often zero, to the iterative cycle. The iteration is
brought to a halt once we achieve some desired level of precision. An interesting
result that is not well known and is also, at first glance, surprising is that θ2
is a fully efficient estimator. It does no less well than θ̂ and this is because,
while subsequent iterations may bring up closer and closer to θ̂, they do not, on
average, bring us any closer to θ0 . A one step estimator such as θ2 can save time
when dealing with onerous simulations.
Note that likelihood theory would imply that we work with the expected
information (called Fisher information) E{I(θ)} but in view of Efron and Hink-
ley (1978) and the practical difficulty of specifying the censoring we usually prefer
to work with a quantity allowing us to consistently estimate the expected infor-
mation, in particular the observed information.
Large sample inference can be based on any one of the three tests derived
from the likelihood. For the score test there is no need to carry out parameter
estimation or to maximize some function. Many well-established tests can be
derived in this way. In exponential families, also the so-called curved exponential
families (Efron et al., 1978), such tests reduce to contrasting some observed
value to its expected value under the model. Good confidence intervals (Cox and
Hinkley, 1979) can be constructed from “good” tests. For the exponential family
class of distributions the likelihood ratio forms a uniformly most powerful test
and, as such, qualifies as a “good” test in the sense of Cox and Hinkley. The
other tests are asymptotically equivalent so that confidence intervals based on
the above test procedures will agree as sample size increases. Also, we can use
such intervals for other quantities of interest such as the survivorship function
since this function depends on these unknown parameters.
Recall from Chapter 3 that we can estimate the survival function as S(t; θ̂).
If Θα provides a 100(1 − α)% confidence region for the vector θ, then we can
obtain a 100(1 − α)% confidence region for S(t; θ) in the following way. For each
t let
Sα+ (t; θ̂) = sup S(t; θ) , Sα− (t; θ̂) = inf S(t; θ) , (7.3)
θ∈Θα θ∈Θα
then Sα+ (t; θ̂) and Sα− (t; θ̂) form the endpoints of the 100(1 − α)% confidence
interval for S(t; θ). Such a quantity may not be so easy to calculate in general,
simulating from Θα or subdividing the space being an effective way to approx-
imate the interval. Some situations nonetheless simplify. The most straight-
forward is where the survival function is a monotonic function of the one-
dimensional parameter θ. As an illustration, the scalar location parameter, θ,
for the exponential model corresponds to the mean. We have that S(t; θ) is
monotonic in θ. For such cases it is only necessary to invert any interval for θ
to obtain an interval with the same coverage properties for S(t; θ). Denoting the
upper limit of the 100(1 − α)% confidence interval for θ as θα + and the lower
−
limit of the 100(1 − α)% confidence interval for θ as θα , we can then write:
Sα+ (t; θ̂) = S(t; θα
− ) and S − (t; θ̂) = S(t; θ + ). Note that these intervals are calcu-
α α
lated under the assumption that t is fixed. For the exponential model, since the
whole distribution is defined by θ, the confidence intervals calculated pointwise
at each t also provide confidence bands for the whole distribution.
Likelihood solution for exponential models

As for the case of a single group, an analysis based on the exponential model
is particularly simple. For this reason alone it is of interest but also (see sec-
tion below on the nonparametric exponential model) the results are much more
general than are often supposed. We restrict attention to the two-group case in
order to enhance readability. The two groups are defined by the binary covariate
Z taking the value either zero or one. The extension to higher dimensions is all
but immediate. For the two-group case we will only need to concern ourselves
with two parameters, λ1 and λ2 , for which, once we have them or consistent
estimates of them, we have the whole survival experience (or estimates of this)
for both groups.
Expressing the model in proportional hazards form we can write: λ1 = λ
and λ2 = λ exp(β). Referring to Equation 7.1 then, if individual i corresponds
to group 1, his or her contribution to the likelihood is f (xi ; λ) = λ exp(−λxi )
when δi = 1, whereas for δi = 0, the contribution is S(xi ; λ) = exp(−λxi ). If
the individual belongs to group 2 the likelihood contribution would be either
λ exp(β) exp(−eβ λxi ) or exp(−eβ λxi ) according to whether δi is equal to one
or zero. We use the variable wi = I(zi = 1) to indicate which group the subject
is from. From this and Theorem 7.1 we have:
Corollary 7.1. For the 2-sample exponential model, the likelihood satisfies
⎧ ⎫
⎨n n
⎬
log L(λ, β) = k log λ + βk2 − λ xj (1 − wj ) + eβ xj wj ,
⎩ ⎭
j=1 j=1
where wi = 1zi =1 and where there are k1 distinct failures in group 1, k2 in group
2, and k = k1 + k2 .
Differentiating the log-likelihood with respect to both λ and β and equating
both partial derivatives to zero we readily obtain an analytic solution to the pair
of equations given by:
Corollary 7.2. The maximum likelihood estimates β̂ and λ̂ for the two-group
exponential model are written as
k2 k1 k1
β̂ = log n − log n ; λ̂ = n .
x w
j=1 j j j=1 j (1 − wj )
x j=1 j (1 − wj )
x

It follows immediately that λ̂1 = λ̂ and that λ̂2 = λ̂ exp(β̂) = k2 / nj=1 xj wj .
In order to carry out tests and construct confidence intervals we construct the
matrix of second derivatives of the log-likelihood, I(λ, β), obtaining

−∂ 2 log L(λ, β)/∂λ2 −∂ 2 log L(λ, β)/∂λ∂β k/λ2 eβ j xj wj
= β β .
−∂ 2 log L(λ, β)/∂λ∂β −∂ 2 log L(λ, β)/∂β 2 e j xj wj λe j xj wj
The advantage of the two parameter case is that the matrix can be explicitly
inverted. We then have:

Corollary 7.3. Let D = λ−1 eβ j xj wj {k − λeβ j xj wj }. Then, for the two-
group exponential model the inverse of the information matrix is given by

λeβ j xj wj −eβ j xj wj
I −1 (λ, β) = D−1 .
−eβ j xj wj k/λ2
The score test is given by XS2 = U (λ̂, 0)I −1 (λ̂, 0)U (λ̂, 0). Following some

simple calculations and recalling that exp(−β̂) = n j=1 xj wj /k2 , we have:
Corollary 7.4. For the two-group exponential model the score test is given by
XS2 = k −1 k1 k2 exp(−β̂){1 − exp(β̂)}2 .
At first glance the above expression, involving as it does β̂, might appear
to contradict our contention that the score statistic does not require estimation
of the parameter. There is no contradiction although we consider in the above

expression λ to be a nuisance parameter not specified under the null hypothesis.
This parameter value does require estimation, although still under the null. The
regression parameter itself turns out to have a simple explicit form, and it is this
same term that appears in the score statistic. Typically, for other models, the
maximum likelihood estimate would not have an explicit analytic form. We do
not need to estimate it in order to evaluate the score statistic. On the other
hand, both the Wald test and the likelihood ratio test do require estimation
under the alternative. The calculations in this specific case can be carried out
straightforwardly and we also have a relatively simple, and again explicit solution
(i.e., not requiring the finding of an iterative solution to the likelihood equation)
for the likelihood ratio test. We have then the following two corollaries:
2
Corollary 7.5. For the two-group exponential model the likelihood ratio test XL
is given by
2 k2 k1 k
XL = 2 k2 log + k1 log − k log .
j xj wj j xj (1 − wj ) j xj
The third of the tests based on the likelihood, the Wald test, is also straight-
forward to calculate and we have the corresponding lemma:
Corollary 7.6. For the two-group exponential model, the Wald test is given by
2
XW = k −1 k1 k2 β̂ 2 .
For large samples we anticipate the three different tests to give very similar
results. For smaller samples the Wald test, although the most commonly used,
is generally considered to be the least robust. In particular, a monotonic trans-
formation of the parameter will, typically, lead to a different value of the test
statistic.
For the Freireich data, the maximum likelihood estimates of the hazard rates
in each group are
λ̂1 = 9/359 = 0.025 , Var(λ̂1 ) = 9/(359)2 = 0.000070,

λ̂2 = 21/182 = 0.115 , Var(λ̂2 ) = 21/(182)2 = 0.00063.
We might note that the above results are those that we would have obtained
had we used the exponential model separately in each of the groups. In this par-
ticular case then the model structure has not added anything or allowed us to
achieve any greater precision in our analysis. The reason is simple. The expo-
nential model only requires a single parameter. In the above model we have two
groups and, allowing these to be parameterized by two parameters, the rate λ
and the multiplicative factor exp(β), we have a saturated model. The saturated
model is entirely equivalent to using two parameters, λ1 and λ2 , in each of the
two groups separately. More generally, for exponential models with many groups
or with continuous covariates, or for other models, we will not usually obtain the
same results from separate analyzes as those we obtain via the model structure.
The model structure will, as long as it is not seriously misspecified, usually lead
to inferential gains in terms of precision of parameter estimates.
Since exp(β̂) = 21/182 × 359/9 = 4.60 we have that the estimate of the
log-relative risk parameter β̂ is 1.53. We also have that the score test XS2 =
17.8, the Wald test XW 2 = 14.7, and the likelihood ratio test X 2 = 16.5. The
L
agreement between the test statistics is good and, in all cases, the significance
level is sufficiently strong to enable us to conclude in favor of clear evidence of a
difference between the groups.

Had there been no censoring then k = n, the sample size, and n j=1 tj cor-
responds to a sum of n independent random variables each exponential with
parameter λ. We could therefore treat n/λ̂ as a gamma variate with parameters
(λ, n). In view of the consistency of λ̂, when there is censoring, we can take k/λ̂
as a gamma variate with parameters (λ, k), when k < n. This is not an exact
result, since it hinges on a large sample approximation, but it may provide greater
accuracy than the large sample normal approximation.
Recall from Chapter 3 that we can make use of standard tables by multiplying
each term of the sum by 2λ. The result of this product is a sum of n exponential
variates in which each component of the sum has variance equal to 2. This
corresponds to a gamma (2, n) distribution which is also equivalent to a chi-
square distribution with 2n degrees of freedom. Taking the range of values of
2kλ/λ̂ to be between χα/2 and χ1−α/2 gives a 100(1 − α)% confidence interval
for λ. For the Freireich data we obtained a 95% CI = (0.0115, 0.0439). On the
basis of intervals for λ, we can obtain intervals for the survivorship function which
is, in this particular case, a monotonic function of λ. The upper and lower limits
of the 100(1 − α)% confidence interval are denoted by Sα+ (t; λ̂) and Sα− (t; λ̂),
respectively. We write:

+ − λ̂χα/2 λ̂χ1−α/2
[Sα (t; λ̂), Sα (t; λ̂)] = exp − t , exp − t . (7.4)
2k 2k
A different approximation is described in Section 7.3 in which

+ λ̂ √
Sα (t; λ̂) ≈ exp − √ k − z1−α/2 t , (7.5)
k
where the corresponding expression for Sα− (t; λ̂) is obtained using the percentiles,
z1−α/2 by zα/2 of the standard normal distribution. Agreement between these
two approximations appears to be very close. It would be of interest to have a
more detailed comparison between the approaches.
Nonparametric exponential analysis

For the one-sample case we have already seen how, referring to Section A.6
and Theorem A.8, we are able to make use of the result that, for any contin-
uous positive random variable T , with distribution function F (t), the variate
T
Λ(T ) = 0 f (u)/[1 − F (u)]du has a standard exponential distribution. For the
one-sample case we can work with the empirical survival function Ŝ(t) appeal-
ing to the result that the observations − log Ŝ(Xi ) can be taken to have been
sampled from a standard exponential distribution.
In the two-group and, by extension, many-group case it will be necessary to
transform observations from a group other than that giving rise to the group
estimate, say ŜG (t). To facilitate this, as described in Chapter 3, we work with
the continuous version of the Kaplan-Meier estimate, S̄(t). Note that a two-group
proportional hazards model can be expressed as S2 (t) = S1α (t) where α = exp(β).
Taking logarithms then enables an analytic expression for β as
β = log{− log S2 (t)} − log{− log S1 (t)}. (7.6)
For the case of three groups, defined by the pair of binary indicator variables, Z1
and Z2 , the model states that S(t|Z1 , Z2 ) = S1α (t) where, in this more complex
set-up, log α = β1 Z1 + β2 Z2 . Here, in exactly the same way, we obtain ana-
lytic expressions for β1 and β2 as the arithmetic difference between the log-log
transformations of the respective marginal survival curves.
For two independent groups G1 and G2 we can consider two separate estima-
tors, Ŝ1 (t) and Ŝ2 (t). Since we are assuming a proportional hazards model, we will
carry over this restriction to the sample-based estimates whereby Ŝ2 (t) = Ŝ1α (t)
and where, as before, α = exp(β). In view of the above result we have:
Lemma 7.1. A consistent estimator of β obtains from

nm
1
β̂ = log{− log S̄2 (Xi )} − log{− log S̄1 (Xi )} , (7.7)
nm
i=1

where nm = i 1Xi ≤m , m = max{Xi |Xi ∈ G }, and m = min(m1 , m2 ).
All of the simple results that are available to us when data are generated by
an exponential distribution can be used. In particular, if we wish to compare the
means of two distributions, both subject to censoring, then we can transform one
of them to standard exponential via its empirical survival function, then use this
same transformation on the other group. The simple results for contrasting two
censored exponential samples can then be applied even though, at least initially,
the data arose from samples generated by some other mechanism.
Estimating equations and model adequacy

A goodness-of-fit test is simply a test of a particular kind of hypothesis, notably a
data-driven hypothesis where certain parameters, under the original specification
of the model, have been replaced by estimates. We are then able to place these
kinds of test under the same heading as the other tests, whether or not they
appeal to any approach using likelihood. More formally, instead of testing a
hypothesis of the type H0 : β = β0 , for some β0 or set of β0 given in advance,
we test a hypothesis of the type H0 : β = β̂. Evaluation of adequacy that goes
beyond simple tests is helpful and knowing the form that certain processes under
the model will take can furnish us with powerful techniques. One good choice
can be based on processes that, under the model, will approximate that of a
Brownian bridge. The construction, evaluated in greater detail in Chapter 9, is
based upon being able to estimate the first two moments of the covariate at
each failure point Xi . For parametric models using likelihood the dependence is
typically expressed via the conditional distributions of time given the covariate,
i.e., f (t|z), that is unlike the case based on the main theorem and the processes
described earlier. However, by applying Bayes’ theorem successively we can obtain
necessary expectations in terms of the conditional probabilities of T given Z, i.e.,
via the use of f (t|z). Specifically we have:

zf (t|z)dG(z) f (t|z)
E Z − Eβ (Z|t) = z− Z dG(z), (7.8)
Z Z
f (t|z)dG(z) Z
f (t|z)dG(z)
where we use the notation Eβ to denote an inner expectation conditional not

only on the assumed model but also on the observations k and n, in addition to
the configuration of censorings. The notation Eβ is not of great interest here, in
the context of a fully parameterized model, but is of great importance in later
sections and chapters. Figure 9.3(b) shows a bridged process for the Freireich
data where the fitted model is a parametric Weibull model. The very good fit is
confirmed by inspection of the curve. The essential difference between parametric
and nonparametric, or semi-parametric, approaches is in the calculation of the
means and variances. The observations are structured in the same way and it is
the model, of whatever nature, that provides the necessary first two moments for
us. In the parametric set-up, since the model specifies the distribution of time
given the covariate, and we need to consider the distribution of the covariate
given time, we make a simple appeal to Bayes’ theorem. In the two-group case
with constant hazard rates and where there is no censoring the above expression
can be simplified. Letting the proportion of the first group be given by π1 , the
second by π2 such that (π1 + π2 = 1) and writing ψ(v) = a(v)/{π1 e−v + a(v)}
where a(v) = π2 eβ exp(−veβ ) then
t t
p
E Z − Eβ (Z|t) = π1 {−ψ(v)}p e−v dv + π2 {1 − ψ(v)} eβ exp(−veβ )dv.
0 0
7.4. Semi-parametric estimating equations 151
The most immediate departure from proportional hazards would be one where
effects decline through time and, as mentioned before, perhaps even changing
direction (Stablein et al., 1981). The standardized cumulative score, however the
moments are calculated, would, under such departures, increase (or decrease)
steadily until the increase (or decrease) dies away and the process would pro-
ceed on average horizontally or, if effects change direction, make its way back
toward the origin. A test based on the maximum would have the ability to pick
up this kind of behavior. Again, the visual impression given by the cumulative
standardized score process can, of itself, suggest the nature of the departure
from proportional hazards. This kind of test is not focused on parameters in the
model other than the regression effect given by β or possibly β(t). The goal is to
consider the proportionality or lack of such proportionality and, otherwise, how
well the overall model may fit is secondary (Figure 7.1).
7.4 Semi-parametric estimating equations
Remind ourselves that the data consist of the observations (Zi (t), Yi (t), (t ≤
Xi ), Xi ; i = 1 . . . n). The Zi are the covariates (possibly time-dependent), the
Xi = min(Ti , Ci ), the observed survival which is the smallest of the censoring
time and the actual survival time, and the Yi (t) are time-dependent indicators
taking the value one as long as the i th subject is at risk at time t and zero
otherwise. For the sake of large sample constructions we make Yi (t) to be left
continuous. At some level we will be making an assumption of independence, an
Figure 7.1: Graphical fit for two parametric models. A slight improvement in fit is

seen for the Weibull model over the exponential model. Both are well within limits
of significance and either assumption appears reasonable. There is a hint of weak
non-significant time dependence that the Weibull model better accommodates.
assumption that can be challenged via the data themselves, but that is often left
unchallenged, the physical context providing the main guide. Mostly, we think
of independence as existing across the indices i (i = 1, . . . , n), i.e., the triplets
{Zi (t), Yi (t), Xi ; i = 1, . . . , n}. It is helpful to our notational construction to have:
Definition 7.1. Let Z(t) be a data-based step function of t, everywhere

equal to zero except at the points Xi , i = 1, ..., n, at which the function

takes the value Zi (Xi ) so that Z(t) = n i=1 Zi (t) 1Xi =t . We assume that
|Zi | is bounded, if not the definition is readily broadened.
The reason for this definition is to unify notation. Our practical interest will be on
sums of quantities such as Zi (Xi ) with i ranging from 1 to n. Using the Stieltjes
integral (Appendix A), we will be able to write such sums as integrals with respect
to an empirical process. In view of the Helly-Bray theorem (Appendix A.2) this
makes it easier to gain an intuitive grasp on the population structure behind
the various statistics of interest. Both T and C are assumed to have supports
on some finite interval, the first of which is denoted T . The time-dependent
covariate Z(·) is assumed to be a left-continuous stochastic process and, for
notational simplicity, is taken to be of dimension one whenever possible.
We use the function Pr(A) to return the probability measure associated with
the event A, the reference sets for this being the largest probability space in the
context. In other words, we have not restricted our outcome space by conditioning
on any particular events. It is usually clear which probability space is assumed,
if not we include F to denote this space preceded by a colon, i.e., we write
Pr(A : F). The function P(A) also returns a probability measure associated with
the event A and we reserve this usage for those cases where significant reduction
of the original probability space has taken place. In other words, we view P(A) as
a probability arising after conditioning, and, in general, after conditioning on a
substantial part of the data structure. Again, the context is usually sufficient to
know what is being conditioned on. If not this is made explicit. It is of interest,
although not generally exploited, to observe that, under repeated sampling, A
under P(A) will generally converge in distribution to that of A under Pr(A). The
discrete probabilities πi (β(t), t), defined in Equation 7.10, are so central to the
development that they are given a notation all of their own. The expectation
operator E(·) is typically associated with Pr(·) whereas the expectation opera-
tor E(·|t) is associated with the model-based πi (β(t),
t t). Let F (t) = Pr(T < t),
D(t) = Pr (C < t), and H(t) = F (t){1 − D(t)} − 0 F (u)dD(u).
For each subject i we observe Xi = min(Ti , Ci ), and δi = I(Ti ≤ Ci ) so that
δi takes the value one if the ith subject corresponds to a failure and is zero if the
subject corresponds to a censored observation. A more general situation allows a
subject to be dynamically censored in that he or she can move in and out of the
risk set. To do this we define the “at-risk” indicator Yi (t) where Yi (t) = I(Xi ≥ t).
The events on the i th individual are counted by Ni (t) = I{Ti ≤ t, Ti ≤ Ci },

and N̄ (t) = n 1 Ni (t) counts the number of events before t. Some other sums
7.5. Estimating equations using moments 153
of observations will frequently occur. In order to obtain an angle on empirical

moments under the model, Andersen and Gill (1982) define
n

S (r) (β, t) = n−1 Yi (t)eβZi (t) Zi (t)r , s(r) (β, t) = ES (r) (β, t),
i=1
for r = 0, 1, 2, where the expectations are taken with respect to the true distri-
bution of (T, C, Z(·)). Define also
S (2) (β, t) S (1) (β, t)2 s(2) (β, t) s(1) (β, t)2
V (β, t) = (0)
− (0) , v(β, t) = (0) − . (7.9)
S (β, t) S (β, t) 2 s (β, t) s(0) (β, t)2
The Andersen and Gill notation is now classic in this context. Their notation lends
itself more readily to large sample theory based upon Rebolledo’s multivariate
central limit theorem for martingales and stochastic integrals. We will keep this
notation in mind for this chapter although, for subsequent chapters, we use
a lighter notation since our approach to inference does not appeal to special
central limit theorems (Rebolledo’s theorem in particular). One reason for using
the Andersen and Gill notation in this chapter is to help the reader familiar with
that theory to join up the dots and readily see the connections with chapters 9,
10, and 11. The required conditions for the Andersen and Gill theory to apply are
slightly broader than those of our development although this advantage is more of
a theoretical than a practical one. For their results, as well as ours, the censorship
is restricted in such a way that, for large samples, there remains information on
F in the tails. The conditional means and the conditional variances, Eβ(t) (Z|t)
Vβ(t) (Z|t), introduced immediately below, are related to the above via V (β, t) ≡
Vβ (Z|t) and S (1) (β, t)/S (0) (β, t) ≡ Eβ (Z|t). In the counting process framework
of Andersen and Gill (1982), we imagine n as remaining fixed and the asymptotic
results obtaining as a result of asymptotic theory for n-dimensional counting
processes, in which we understand the expectation operator E to be with respect
to infinitely many repetitions of the process. Subsequently we allow n to increase
without bound. For the quantities Eβ(t) (Z k |t) we take the E operator to be these
same quantities when n grows without bound.
7.5 Estimating equations using moments
We most often view time as providing the set of indices to certain stochastic
processes, so that, for example, we consider Z(t) to be a random variable having
different distributions for different t. Also, the failure time variable T can be
viewed as a non-negative random variable with distribution F (t) and, whenever
the set of indices t to the stochastic process coincides with the support for T ,
then not only can we talk about the random variables Z(t) for which the dis-
tribution corresponds to Pr(Z ≤ z|T = t) but also marginal quantities such as
the random variable Z(T ) having distribution G(z) = Pr(Z ≤ z). An important
result concerning the conditional distribution of Z(t) given T = t follows. How-
ever, the true population joint distribution of (T, Z) turns out to be of little
interest. Concerning T , we will view its support mostly in terms of providing
indices to a stochastic process. The variable Z depends of course on rather arbi-
trary design features. In order to study dependency, quantified by β(t), we focus
on the conditional distribution of Z given T = t.
Definition 7.2. The discrete probabilities πi (β(t), t) are given by
Yi (t) exp{β(t)Zi (t)}

πi (β(t), t) = n . (7.10)
j=1 Yj (t) exp{β(t)Zj (t)}
The πi (β(t), t) are easilyseen to be bonafide probabilities (for all real values of
β(t)) since πi ≥ 0 and i πi = 1. Note that this continues to hold for values
of β(t) different from those generating the data, and even when the model is
incorrectly specified. As a consequence, replacing β by β̂ results in a probability
distribution that is still valid but different from the true one. Means and vari-
ances with respect to this distribution maintain their interpretation as means and
variances.
Under the proportional hazards assumption, i.e., the constraint β(t) = β, the
product of the π’s over the observed failure times gives the partial likelihood (Cox
1972, 1975). When β = 0, πi (0, t) is the empirical distribution that assigns equal
weight to each sample subject in the risk set. Based on the πi (β(t), t) we have:
Definition 7.3. Conditional moments of Z with respect to πi (β(t), t) are given
by
n

k
Eβ(t) (Z |t) = Zik (t)πi (β(t), t) , k = 1, 2, . . . , . (7.11)
i=1
These two definitions are all that we need in order to set about building the
structures upon which inference is based. This is particularly so when we are able
to assume an independent censoring mechanism, although the weaker assumption
of a conditionally independent censoring mechanism (see Chapter 2) will mostly
cause no conceptual difficulties; simply a slightly more burdensome notation.
Another, somewhat natural, definition will also be appealed to on occasion and
this concerns unconditional expectations.
Definition 7.4. Marginal moments of Z with respect to the bivariate distribution
characterized by πi (β(t), t) and F (t) are given by

Eβ(t) (Z ) = Eβ(t) (Z k |t)dF (t) , k = 1, 2, . . . , .
k
(7.12)
Recall that for arbitrary random variables A and B, assuming expectation to

be defined, we have the result of double expectation whereby E(A) = EE(A|B).
This is the motivation behind the above definition. Once again, these expecta-
tions are to be interpreted as population quantities in as much as β(t) and F (t)
are taken to be known. They can also, of course, be viewed as sample-based
quantities since n is finite and the Yi (t) are random until time point t. At the
end of the study the paths of all the Yi (t) are known and we are, to use a common
expression, “conditioning on the data.” The art of inference, and its understand-
ing, stem, to a great extent, from knowing which aspects of an experiment to
view as random (given that once the experiment is over there is not really any-
thing truly random). Also which distributions are relevant and these can change
so that, here for example, we should think carefully about the meaning of the
expectation operators E and E in its particular context. These expectations are
still well defined, but with respect to different distributions; when replacing β by
β̂, when replacing F by Fn and F̂ , and when allowing n to go to infinity. The
quantity φ of the following definition is not of any essential interest, featuring in
the main theorem but disappearing afterwards.
Definition 7.5. In order to distinguish conditionally independent censoring

from independent censoring we introduce BC (t, z) = P (C ≥ t|z) and define
φ(z, t) where

φ(z ∗ , t) = BC (t, z ∗ )−1 × P (C ≥ t|z)dG(z).
Note that when censoring does not depend upon z then φ(z, t) will depend upon
neither z nor t and is, in fact, equal to one. Otherwise, under a conditionally
independent censoring assumption, we can consistently estimate φ(z, t) and we
call this φ̂(z, t). This is not explored in this text.
Theorem 7.2. (O’Quigley 2003). Under model (4.2) and assuming β(t)
known, the conditional distribution function of Z(t) given T = t is given by

z ≤z Yi (t) exp{β(t)zi (t)}φ̂(zi , t)
P{Z(t) ≤ z|T = t} = ni . (7.13)
j=1 Yj (t) exp{β(t)zj (t)}φ̂(zj , t)
The theorem, which we refer to as the main theorem of proportional hazards

regression, has many important consequences including
Corollary 7.7. Under model (4.2) and an independent censorship, assuming β(t)
known, the conditional distribution function of Z(t) given T = t is given by
n

P(Z(t) ≤ z|T = t) = πj (β(t), t)I(Zj (t) ≤ z). (7.14)
j=1
The observation we would like to make here is that we can fully describe a
random variable indexed by t, i.e., a stochastic process. This idea underlies the
development of the regression effect process described in the following chapters.
All of our inferences can be based on this. In essence, we first fix t and then
we fix our attention on the conditional distribution of Z given that T = t. This
distribution brings into play the models of interest. These models are character-
ized by this distribution making it straightforward to construct tests, to estimate
parameters, and to build confidence regions. Indeed, under the broader censoring
definition of conditional independence, common in the survival context, we can
still make the same basic observation. In this case we condition upon something
more complex than just T = t. The actual random outcome that we condition
upon is of less importance than the simple fact that we are able to describe sets
of conditional distributions all indexed by t, i.e., a stochastic process indexed by
t. Specifically
Corollary 7.8. For a conditionally independent censoring mechanism we

have
n

P(Z(t) ≤ z|T = t, C > t) = πj (β(t), t)I(Zj (t) ≤ z). (7.15)
j=1
Whether we condition on the event T = t or the event (T = t, C > t), we iden-

tify a random variable indexed by t. This is all we need to construct appropriate
stochastic processes (functions of Z(t)) enabling inference. Again simple appli-
cations of Slutsky’s theorem show that the result still holds, as a large sample
approximation, for β(t) replaced by any consistent estimate. In particular, when
the hypothesis of proportionality of risks is correct, the result holds for the esti-
mate β̂. The following two corollaries follow immediately from those just above
and form the basis for simple inference. For integer k we have:
Corollary 7.9. Eβ̂(t) (Z k |t) provides a consistent estimate of Eβ(t) (Z k (t)|t),
under model (4.2). In particular Eβ̂ (Z k |t) provides a consistent estimate of
Eβ (Z k (t)|t), under the model expressed by Equation 4.3.
Furthermore, once again working under the model, we consider
Definition 7.6. Vβ(t) (Z|t) = Eβ(t) (Z 2 |t) − Eβ(t)

2 (Z|t) .
In practical data analysis the quantity β(t) may be replaced by a value con-
strained by some hypothesis or an estimate. The quantity Vβ(t) (Z|t) can be
viewed as a conditional variance which may vary little with t, in a way analo-
gously to the residual variance in linear regression which, under classic assump-
tions, remains constant with different levels of the independent variable. Since
Vβ(t) (Z|t) may change with t, even if not a lot, it is of interest to consider some
average quantity and so we also introduce

Definition 7.7. σ 2 = E Vβ(t) (Z) = Vβ(t) (Z|t)dF (t) .
Interpretation requires some care. For example, although E Vβ̂ (Z|t) is, in
some sense, a marginal quantity, it is not the marginal variance of Z since we
have neglected the variance of Eβ(t) (Z(t)|t) with respect to the distribution of T.
The easiest case to interpret is the one where we have an independent censoring
mechanism (Equation 7.14). However, we do not need to be very concerned
about any interpretation difficulty, arising for instance in Equation 7.15 where
the censoring time appears in the expression, since, in this or the simpler case,
all that matters to us is that our observations can be considered as arising from
some process, indexed by t and, for this process, we are able, under the model,
to consistently estimate the mean and the variance of the quantities that we
observe. It is also useful to note another natural relation between Vβ (Z|t) and
Eβ (Z|t) since
Vβ (Z|t) = {∂ Eθ (Z|t)/∂θ.}θ=β (7.16)
This relation is readily verified for fixed β. In the case of time-dependent β(t)
then, at each given value of t, it is again clear that the same relation holds. The
result constitutes one of the building blocks in the overall inferential construction
and, under weak conditions, essentially no more than Z being bounded, then it
also follows that

Vβ (Z|t) = ∂ Eβ (Z|t)/∂β = ∂ Eβ (Z|t) /∂β.
Essentially all the information we need, for almost any conceivable statistical
goal, arising from considerations of any of the models considered, is contained
in the joint probabilities πi (β(t), t) of the fundamental definition 7.2. We are
often interested, in the multivariate setting for example, in the evaluation of the
effects of some factor while having controlled for others. This can be immediately
accommodated. Specifically, taking Z to be of some dimension greater than one
(β being of the same dimension), writing Z T = (Z1T , Z2T ) and ZiT = (Z1i T , Z T ),
2i
and then summing over the multivariate probabilities, we have two obvious exten-
sions to Corollaries 7.7 and 7.8.
Corollary 7.10. Under model (4.2) and an independent censorship, assuming

β(t) known, the conditional distribution function of Z2 (t) given T = t is given
by
n

P(Z2 (t) ≤ z|T = t) = πj (β(t), t)I(Z2j (t) ≤ z). (7.17)
j=1
The corollary enables component wise inference. We can consider the com-
ponents of the vector Zi individually. Also we could study some functions of the
components, usually say a simple linear combination of the components such as
the prognostic index. Note also that
Corollary 7.11. For a conditionally independent censoring mechanism we have

n

P(Z2 (t) ≤ z|T = t, C > t) = πj (β(t), t)I(Z2j (t) ≤ z), (7.18)
j=1
where in Definition 7.2 for πj (β(t), t) we take β(t)Zj (t) to be an inner product,
which we may prefer to write using boldface or as β(t)T Zj (t) and where Zj (t)
are the observed values of the vector Z(t) for the jth subject. Also, by Z2 (t) ≤ z
we mean that all of the scalar components of Z2 (t) are less than or equal to
the corresponding scalar components of z. As for the corollaries and definitions
following Corollaries 7.7 and 7.8 they have obvious equivalents in the multivariate
setting and so we can readily write down expressions for expectations, variances,
and covariances as well as their corresponding estimates.
Moments for stratified models

Firstly we recall from the previous chapter that the stratified model is simply a
partially proportional hazards model in which some of the components of β(t)
remain unspecified while the other components are constant terms. The definition
for the stratified model was
λ(t|Z(t), s) = λ0s (t) exp{β(t)Z(t)},
where s takes integer values 1, ..., m. In view of the equivalence between strat-
ified models and partially proportional hazards models described in the previ-
ous chapter, the main theorem and its corollaries apply immediately. However,
in light of the special importance of stratified models, as proportional hazards
models with relaxed assumptions, it will be helpful to our development to devote
a few words to this case. Analogous to the above definition for πi (β(t), t), and
using the, possibly time-dependent, stratum indicator s(t) we now define these
probabilities via
Definition 7.8. For the stratified model, having strata s = 1, . . . , m, the dis-
crete probabilities πi (β(t), t) are now given by
Yi {s(t), t} exp{β(t)Zi (t)}

πi (β(t), t) = n . (7.19)
j=1 Yj {s(t), t} exp{β(t)Zj (t)}
When there is a single stratum then this definition coincides with the earlier
one and, indeed, we use the same πi (β(t), t) for both situations, since it is only
used indirectly and there is no risk of confusion. Under equation (4.3), i.e., the
constraint β(t) = β, the product of the π’s over the observed failure times gives
the so-called stratified partial likelihood (Kalbfleisch and Prentice, 2002). The
series of above definitions for the non-stratified model, in particular Definition
7.2, theorems, and corollaries, all carry over in an obvious way to the stratified
model and we do not propose any additional notation. It is usually clear from
the context although it is worth making some remarks. Firstly, we have no direct
interest in the distribution of Z given t (note that this distribution depends on
the distribution of Z given T > 0, a distribution which corresponds to our design
and is quite arbitrary).
We will exploit the main theorem in order to make inferences on β and, in the
stratified case, we would also condition upon the strata from which transitions
can be made. In practice, we contrast the observations Zi (Xi ), made at time
point Xi at which an event occurs (δi = 1) with those subjects at risk of the same
event. The “at-risk” indicator, Y (s(t), t), makes this very simple to express. We
can use Y (s(t), t) to single out appropriate groups for comparison. This formalizes
a standard technique in epidemiology whereby the groups for comparison may be
matched by not just age but by other variables. Such variables have then been
controlled for and eliminated from the analysis. Their own specific effects can
be quite general and we are not in a position to estimate them. Very complex
situations, such as subjects moving in and out of risk categories, can be easily
modeled by the use of these indicator variables.
Moments for other relative risk models

Instead of Equation 4.2 some authors have suggested a more general form for
the hazard function whereby
λ(t|Z) = λ0 (t)R{β(t)Z}, (7.20)
and where, mostly, β(t) is not time-varying, being equal to some unknown con-
stant. The most common choices for the function R(r) are exp(r), in which
case we recover the usual model, and 1 + r which leads to the so-called addi-
tive model. Since both λ(t|Z) and λ0 are necessarily positive we would generally
need constraints on the function R(r). In practice this can be a little bothersome
and is, among several other good reasons, a cause for favoring the multiplicative
risk model exp(r) over the additive risk model 1 + r. If we replace our earlier
definition for πi (β(t), t) by
Definition 7.9. The discrete probabilities πi (β(t), t) are given by
Yi (t)R{β(t)Zi (t)}
πi (β(t), t) = n , (7.21)
j=1 Yj (t)R{β(t)Zj (t)}
following which all of the above definitions, theorems, and corollaries have imme-
diate analogues and we do not write them out explicitly. Apart from one inter-
esting exception, which we look at more closely in the chapters dealing with
inference, there are no particular considerations we need to concern ourselves
over if we choose R(r) = 1 + r rather than R(r) = exp(r).
What is more, if we allow the regression functions, β(t), to depend arbitrarily
upon time then, given either model, the other model exists with a different
function of β(t). The only real reason for preferring one model over another
would be due to parsimony; for example, we might find in some given situation
that in the case of the additive model the regression function β(t) is in fact
constant unlike the multiplicative model where it may depend on time. But
otherwise both functions may depend, at least to some extent, on time and then
the multiplicative model ought to be preferred since it is the more natural. We
say the more natural because the positivity constraint is automatically satisfied.
Transformed covariate models

For some transformation ψ of the covariate we can postulate a model of the
form
λ(t|Z(t)) = λ0 (t) exp{β(t)ψ[Z(t)]}. (7.22)
All of the calculations proceed as above and no real new concept is involved.
Such models can be considered in the case of continuous covariates, Z, which
may be sufficiently asymmetric, implying very great changes of risk at the high
or low values, to be unlikely to provide a satisfactory fit. Taking logarithms,
or curbing the more extreme values via a defined plateau, or some other such
transformation will produce models of potentially wider applicability. Note that
this is a different approach to work with, say,
λ(t|Z(t)) = λ0 (t) exp{β(t)Z(t)},
and using the main theorem, in conjunction with estimating equations described
here below and basing inference upon the observations ψZ(Xi ) and their expec-
tations under this model. In this latter case we employ ψ in the estimating
equation as a means to obtain greater robustness or to reduce sensitivity to large
observations. In the former case the model itself is different and would lead to
different estimates of survival probabilities.
Our discussion so far has turned around the hazard function. However, it
is equally straightforward to work with intensity functions and these allow for
increased generality, especially when tackling complex time-dependent effects.
O’Brien (1978) introduced the logit-rank test for survival data when investi-
gating the effect of a continuous covariate on survival time. His purpose was
to construct a test that was rank invariant with respect to both time and the
covariate itself. O’Quigley and Prentice (1991) showed how a broad class of rank
invariant procedures can be developed within the framework of proportional haz-
ards models. The O’Brien logit-rank procedure was a special case of this class.
In these cases we work with intensity rather than hazard functions. Suppose
then that λi (t) indicates an intensity function for the ith subject at time t. A
proportional hazards model for this intensity function can be written as
λi (t) = Yi (t)λ0 (t) exp{βZi (t)},
where Yi (t) indicates whether or not the ith subject is at risk at time t, λ0 (t)
the usual “baseline” hazard function, and Zi (t) is a constructed covariate for
the ith subject at time t. Typically, Zi (t) in the estimating equation is defined
as a function of measurements on the ith subject alone, but it can be defined
more generally as Zi (t) = ψi (t, Ft ) for ψ some function of Ft , the collective
failure, censoring, and covariate information prior to time t on the entire study
group. The examples in O’Quigley and Prentice (1991) included the rank of the
subject’s covariate at Xi and transformations on this such as the normal order
statistics. This represents a departure from most regression situations because
the value used in the estimating equation depends not only on what has been
observed on the particular individual but also upon what has been observed on
other relevant subsets of individuals.
Structured time effects for β(t)

In the review of goodness-of-fit tests for survival models, O’Quigley and Xu
(2014) gave particular consideration to the non-proportional hazards model with
intercept described in Chapter 4. Using this model and the usual likelihood pro-
cedures we can test for several specific departures from proportional hazards.
For the sake of simplicity of exposition we continue to limit attention to the
single variable case. Extension to higher dimensions is immediate. Also we will
sometimes write Z, instead of Z(t), in order for the notation to not become over
cluttered. It can always be replaced by Z(t) without additional concerns. We
write
λ(t|Z) = λ0 (t) exp{[β + αQ(t)]Z(t)}, (7.23)

where Q(t) is a function of time that does not depend on the parameters β and
α. Under a null hypothesis that a proportional hazards model is adequate, i.e.,
H0 : α = 0, we recover a proportional hazards model. In the context of sequential
group comparisons of survival data, the above model has been considered by
Tsiatis (1982) and Fleming and Harrington (1984). In keeping with the usual
notation we denote Eβ,α (Z|t) to be the expectation taken with respect to the
probability distribution πi (β, α, t), where
Yi (t) exp{[β + αQ(t)]Zi (t)}

πi (β, α, t) = n . (7.24)
j=1 Yj (t) exp{[β + αQ(t)]Zj (t)}
Lemma 7.2. The components of the score vector U (β, α) can be expressed as
n

Uβ (β, α) = δi {Zi (Xi ) − Eβ,α (Z|Xi )}, (7.25)
i=1
n
Uα (β, α) = δi Q(Xi ){Zi (Xi ) − Eβ,α (Z|Xi )}. (7.26)
i=1
A test would be carried out using any one of the large sample tests aris-
ing from considerations of the likelihood. If we let β̂ be the maximum par-
tial likelihood estimate of β under the null hypothesis of proportional hazards,
i.e., H0 : α = 0, then Uβ (β̂, 0) = 0. The score test statistic arising under H0 is
−1
B = Uα (β̂, 0)G−1 Uα (β̂, 0), where G = I22 − I21 I11 I12 and G−1 is the lower
−1
right corner element of I . The elements of the information matrix required to
carry out the calculation are given below in Lemma 7.3. Under H0 , the hypothesis
of proportional hazards, B has asymptotically a χ2 distribution with one degree
of freedom.
Lemma 7.3. Taking k = 1, 2 and = 1, 2, the components of I are

Uββ Uβα I11 I12
I(β, α) = − = where
Uαβ Uαα I21 I22
n

Ik (β, α) = δi Q(Xi )k+−2 {Eβ,α (Z 2 |Xi ) − Eβ,α
2
(Z|Xi )} .
i=1
The literature on the goodness-of-fit problem for the Cox model has consid-
ered many formulations that correspond to particular choices for Q(t). The first
of these was given in the founding paper of Cox (1972). Cox’s suggestion was
equivalent to taking Q(t) = t. Defining Q(t) as a two-dimensional vector, Sta-
blein et al. (1981) considered Q(t) = (t, t2 ) , and Brown (1975), Anderson and
Senthilselvan (1982), O’Quigley and Moreau (1984), Moreau et al. (1985), and
O’Quigley and Pessione (1989) assumed Q(t) to be constant on predetermined

intervals of the time axis, i.e., Q(t) is a step function.
Although in the latter cases there is more than one parameter associated with
Q(t), the computation of the test statistic is similar. Murphy (1993) studied the
size and the power of the test of Moreau and colleagues and found that, although
it is consistent against a wide class of alternatives to proportional hazards, the
Moreau test is nonetheless an omnibus test that is used to greatest advantage
when there is no specific alternative in mind.
We can choose Q(t) to be an unknown function and use the available data
to provide an estimate of this function. For instance, Breslow et al. (1984) chose
Q(t) = Λ(t) and replaced this unknown function by the Nelson estimator. Because
the estimates at time t depend only on the history of events up to that time, the
development of Cox (1975) and Andersen and Gill (1982) and thus, the usual
asymptotic theory, still applies. Breslow et al. (1984) showed that their choice
Q(t) = Λ(t) has good power against the alternative of crossing hazards. Tsiatis
(1982) and Harrington and Fleming (1982) used score processes based on the
non-proportional hazards model with intercept to derive sequential tests. They
showed that after Q(t) has been replaced by its estimate, the score process at dif-
ferent time points converges in distribution to a multivariate normal. Harrington
and Fleming focus particular interest on the Gρ family, where Q(t) = S(t)ρ .
Another special case is described in O’Quigley and Pessione (1991), O’Quigley
and Natarajan (2004), and O’Quigley and Flandre (1994), where Q(t) = I(t ≤
γ) − I(t > γ) with γ an unknown changepoint. When γ is known the test statistic
can be evaluated with no particular difficulty. For γ unknown, we would maximize
over all possible values of γ and special care is required for the resulting infer-
ence to remain valid. O’Quigley and Pessione (1991) showed that tests using a
changepoint model can be powerful for testing the equality of two survival distri-
butions against the specific alternative of crossing hazards. Also, such tests suffer
only moderate losses in power, when compared with their optimal counterparts,
if the alternative is one of proportional hazards.
Lin (1991) and Gill and Schumacher (1987) have taken a slightly different
approach to work with the function Q(t) and introduce it directly into a weighted
score. This can be written as
n

UQ (β) = δi Q(Xi ){Zi (Xi ) − E(Z|Xi ; β)}, (7.27)
i=1
where the only key requirement is that Q(·) be a predictable process (see Sec-
tion B.4) that converges in probability to a non-negative bounded function uni-
formly in t. Let β̂Q be the zero of (7.27) and β̂ be the partial likelihood esti-
mate. Under the assumption that the proportional hazards model holds and that
(Xi , δi , Zi ) (i = 1, . . . , n) are i.i.d. replicates of (X, δ, Z), n1/2 (β̂Q − β̂) is asymp-
totically normal with zero mean and covariance matrix that can be consistently
estimated. It then follows that a simple test can be based on the standardized
difference between the two estimates. Lin (1991) showed such a test to be con-
sistent against any model misspecification under which βQ = β, where βQ is the
probability limit of β̂Q . In particular, it can be shown that choosing a mono-
tone weight function for Q(t) such as F̂ (t), where F̂ (·) is the Kaplan-Meier
estimate, is consistent against monotone departures (e.g., decreasing regression
effect) from the proportional hazards assumption.
7.6 Incorrectly specified models
For multinormal linear regression involving p regressors we can eliminate from

consideration some of these and focus our attention on models involving the
remaining regressors strictly less than p. We could eliminate these by simple inte-
gration, thereby obtaining marginal distributions. Under the usual assumptions of
multiple linear regression the resulting lower dimensional model remains a multi-
normal one. As an example, in the simple case of a two-dimensional covariate
normal model, both the marginal models involving only one of the two covari-
ates are normal models. However, for non-linear models this result would only
be expected to hold under quite unusual circumstances. Generally, for non-linear
models, and specifically proportional hazards models, the result will not hold so
that if the model is assumed true for a covariate vector of dimension p, then,
for any sub-model, of dimension less than p, the model will not hold exactly. A
corollary to this is that no model of dimension greater than p could exactly follow
a proportional hazards prescription if we claim that the model holds precisely for
some given p covariates.
These observations led some authors to claim that “forgotten” or “over-
looked” variables would inevitably lead to misleading results. Such a claim implies
that all analyses based on proportional hazards models are misleading and since,
to say the least, such a conclusion is unhelpful we offer a different perspec-
tive. This says that all practical models are only ever approximately correct. In
other words, the model is always making a simplifying assumption, necessar-
ily overlooking potential effects as well as including others which may impact
the proportionality of those key variables of interest. Our task then focuses on
interpreting our estimates when our model cannot be exactly true. In terms of
analyzing real data, it makes much more sense to take as our underlying working
assumption that the model is, to a greater or lesser degree, misspecified.
A model can be misspecified in one of two clear ways; the first is that the
covariate form is not correctly expressed and the second is that the regression
coefficient is not constant through time. An example of the first would be that
the true model holds for log Z but that, not knowing this, we include Z in the
model. An example of the second might have β(t) declining through time rather
than remaining constant.
7.6. Incorrectly specified models 165
It has been argued that the careful use of residual techniques can indicate
which kind of model failure may be present. This is not so. Whenever a poor
fit could be due to either cause it is readily seen that a misspecified covariate
form can be represented correctly via a time-dependent effect. In some sense the
two kinds of misspecification are unidentifiable. We can fix the model by working
either with the covariate form or the regression coefficient β(t). Of course, in
certain cases, a discrete binary covariate describing two groups, for example, there
can only be one cause of model failure—the time dependency of the regression
coefficient. This is because the binary coding imposes no restriction of itself since
all possible codings are equivalent.
Average regression effect

The important issue is then the interpretation of an estimate say β̂ under a pro-
portional hazards assumption when, in reality, the data are generated under the
broader non-proportional hazards model with regression coefficient function β(t).
This is not a straightforward endeavor and the great majority of the currently used
procedures, including those proposed in the widely distributed R, SAS, STATA,
and S-Plus packages, produce estimates which cannot be interpreted
unless there
is no censoring. To study this question we first define μ = β(t)dF (t), which is
an average of β(T ) with respect to the distribution F (t).
It is of interest to consider the approximation,
n

P(Z(t) ≤ z|T = t, C > t) ≈ πj (μ, t)I(Zj (t) ≤ z) (7.28)
j=1
and, for the case of a model making the stronger assumption of an indepen-
dent censoring mechanism as opposed to a conditionally independent censoring
mechanism given the covariate, we have
n

P(Z(t) ≤ z|T = t) ≈ πj (μ, t)I(Zj (t) ≤ z). (7.29)
j=1
For small samples it will be unrealistic to hope to obtain reliable estimates of

β(t) for all of t so that, often, we take an estimate of some summary measure, in
particular μ. It is in fact possible to construct an estimating equation which pro-
vides an estimate of μ without estimating β(t) (Xu and O’Quigley, 2000) and it
is very important to stress that, unless there is no censoring, the usual estimating
equation which leads to the partial likelihood estimate does not accomplish this.
Some thought needs to be given to the issues arising when our estimating
equation is based on certain assumptions (in particular, a proportional hazards
assumption), whereas the data themselves can be considered to have been gen-
erated by something broader (in particular, a non-proportional hazards model).
To this purpose we firstly consider a definition that will allow us to anticipate

just what is being estimated when the data are generated by model (4.2) and
we are working with model (4.3). This is contained in the definition for β ∗ just
below.
Let’s keep in mind the widely held belief that the partial likelihood estimate
obtained when using a proportional hazards model in a situation where the data
are generated by a broader model must correspond to some kind of average
effect. It does correspond to something (as always) but nothing very useful and
not something we can hopefully interpret as an average effect. Firstly we need
Definition 7.10. Let β ∗ be the constant value satisfying

Eβ ∗ (Z|t)dF (t) = Eβ(t) (Z|t)dF (t). (7.30)
T T
The definition enables us to make sense out of using estimates based on (4.3)
when the data are in fact generated by (4.2). Since we can view T as being
random, whenever β(t) is not constant, we can think of having sampled from
β(T ). The right-hand side of the above equation is then a double expectation
and β ∗ , occurring in the left-hand side of the equation, is the best fitting value
under the constraint that β(t) = β. We can show the existence and uniqueness of
solutions to Equation 7.30 (Xu and O’Quigley, 2000). More importantly, β ∗ can
be shown to have the following three properties: (i) under model (4.3) β ∗ = β;
(ii) under a subclass of the broad class of models known as the Harrington-
Fleming models, we have an exact result in thatβ ∗ = T β(t)dF (t); and (iii) for
very general situations we can write that β ∗ ≈ T β(t)dF (t), an approximation
which is in fact very accurate. Estimates of β ∗ are discussed in Xu and O’Quigley
(2000) and, in the light of the foregoing, we can take these as estimates of μ.
Families of estimating equations

The above setting helps us anticipate the properties of the estimators we will be
using. First, recall our definition of Z(t) as a step function of t with discontinuities
at the points Xi , i = 1, ..., n, at which the function takes the value Zi (Xi ). Next,
consider Fn (t), the empirical marginal distribution function of T . Note that Fn (t)
coincides with the Kaplan-Meier estimate of F (t) in the absence of censoring.
When there is no censoring, a sensible estimating equation (which we will see also
arises as the derivative of a log likelihood, as well as the log partial likelihood) is

U1 (β) = {Z(t) − Eβ (Z|t)}dFn (t) = 0. (7.31)
The above integral is simply the difference of two sums, the first the empirical
mean without reference to any model and the second the average of model-based
means. It makes intuitive sense as an estimating equation and the only reason
for writing the sum in the less immediate form as an integral is that it helps
p
understand the large sample theory when Fn (t) → F (t). Each component in the
above sum includes the size of the increment, 1/n, a quantity that can then
be taken outside of the summation (or integral) as a constant factor. Since the
right-hand side of the equation is identically equal to zero, the incremental size
1/n can be canceled, enabling us to rewrite the equation as

U2 (β) = {Z(t) − Eβ (Z|t)}dN̄ (t) = 0. (7.32)
It is this expression the integral is taken with respect to increments dN̄ (t), rather
than with respect to dFn (t) that is the more classic representation in this context.
The expression equates U2 (β) in terms of the counting processes Ni (t). These
processes, unlike the empirical distribution function, are available in the presence
of censoring. It is the above equation that is used to define the partial likelihood
estimator, since, unless the censoring is completely absent, the quantity U1 (β)
is not defined.
Now, suppose that two observers were to undertake an experiment to estimate
β. A certain percentage of observations remain unobservable to the first observer
as a result of an independent censoring mechanism but are available to the second
observer. The first observer uses Equation 7.32 to estimate β, whereas the second
observer uses Equation 7.31. Will the two estimates agree? By “agree” we mean,
under large sample theory, will they converge to the same quantity. We might
hope that they would; at least if we are to be able to usefully interpret estimates
obtained from Equation 7.32. Unfortunately though (especially since Equation
7.32 is so widely used), the estimates do not typically agree, the greater the degree
of censoring, even when independent, the greater the disagreement. Table 7.1
below indicates just how severe the disagreement might be. However, the form of
U1 (β) remains very much of interest and, before discussing the properties of the
above equations let us consider a third estimating equation which we write as

U3 (β) = {Z(t) − Eβ (Z|t)}dF̂ (t) = W (t){Z(t) − Eβ (Z|t)}dN̄ (t) = 0, (7.33)

upon defining the stochastic process W (t) = Ŝ(t){ n i=1 Yi (t)}
−1 . For practical
calculation note that W (Xi ) = F̂ (Xi +) − F̂ (Xi ) at each observed failure time
Xi , i.e., the jump in the KM curve. When there is no censoring, then clearly
U1 (β) = U2 (β) = U3 (β).
More generally U1 (β) may not be available and solutions to U2 (β) = 0 and
U3 (β) = 0 do not coincide or converge to the same population counterparts
even under independent censoring. They would only ever converge to the same
quantities under the unrealistic assumption that the data are exactly generated
by a proportional hazards model. As argued in the previous section we can assume

that this never really holds in practical situations.
Many other possibilities could be used instead of U3 (β), ones in which other
consistent estimates of F (t) are used in place of F̂ (t), for example, the Nelson-
Aalen estimator or, indeed, any parametric estimate for marginal survival. If we
were to take the route of parametric estimates of marginal survival, we would need
to be a little cautious since these estimates could also contain information on the
parameter β which is our central focus. However, we could invoke a conditional
argument, i.e., take the marginal survival estimate as fixed and known at its
observed value or argue that the information contained is so weak that it can
be ignored. Although we have not studied any of these we would anticipate the
desirable properties described below to still hold. Stronger modeling assumptions
are also possible (Klein et al., 1990; Moeschberger and Klein, 1985).
Note also that the left-hand side of the equation is a special case of the
weighted scores under the proportional hazards model (Harrington and Fleming,
1982; Lin, 1991; Newton and Raftery, 1994). However those weighted scores
were not proposed with the non-proportional hazards model in mind, and the
particular choice of W (·) used here was not considered in those papers. Indeed
other choices for the weights will lead to estimators closer to the partial likelihood
itself, in the sense that under a non-proportional hazards model and in the pres-
ence of censoring, the broader class of weighted estimates will not converge to
quantities that remain unaffected by an independent censoring mechanism. On
the other hand, the estimating equation based on U3 is in the same spirit as the
approximate likelihood of Oakes (1986) for censored data and the M-estimate of
Zhou and Li (2008) for censored linear models. Hjort (1992) also mentioned the
use of the reciprocal of the Kaplan-Meier estimate of the censoring distribution
as weights in parametric survival models, and these weights are the same as W (·)
defined here. For the random effects model—a special case of this is the stratified
model which, in turn, can be expressed in the form (4.2)—we can see, even when
we know that (4.3) is severely misspecified, that we can still obtain estimates of
meaningful quantities. The average effect resulting from the estimating equation
U3 is clearly of interest.
For the stratified model, Z(Xi ) is contrasted with its expectation Eβ (Z|Xi , s).
Here, the inclusion of s is used to indicate that if Zi belongs to stratum s then
the reference risk set for Eβ (Z|Xi , s) is restricted to members of this same stra-
tum. Note that for time-dependent s(t) the risk set is dynamic, subjects entering
and leaving the set as they become at risk. The usual estimating equation for
stratified models is again of the form U (β) and, for the same reasons as recalled
above and described more fully in Xu and O’Quigley (2000) and we might prefer
to use

Us (β) = {Z(t) − Eβ (Z|t, s)}dF̂ (t) = 0 . (7.34)
Even weaker assumptions (not taking the marginal F (t) to be common across
strata) can be made and, at present, this is a topic that remains to be studied.
Zeros of estimating equations

Referring back to Section 7.5 we can immediately deduce that the zeros of the
estimating equations provide consistent estimates of β under the model. Below
we consider zeros of the estimating equations when the model is incorrectly
specified. This is important since, in practice, we can assume this to be the case.
Most theoretical developments proceed under the assumptions that the model is
correct. We would have that β̂ where U2 (β̂) = 0 is consistent for β. Also β̃ where
U3 (β̃) = 0 is consistent for a parameter of interest, namely the average effect.
From the mean value theorem we write

∂U2 (β)
U2 (β̂) = U2 (β0 ) + (β̂ − β0 ) ,
∂β β=ξ
where ξ lies strictly on the interior of the interval with endpoints β0 and β̂. Now
n
U2 (β̂) = 0 and U2 (ξ) = n i=1 δi Vξ (Z|Xi ) so that Var(β̂) ≈ 1/ i=1 δi Var(Z|Xi ).
This is the Cramer-Rao bound and so the estimate is a good one. Although the
sums are of variables that we can take to be independent they are not iden-
tically distributed. Showing large sample normality requires verification of the
Lindeburgh condition which, although awkward, is not difficult. All the neces-
sary ingredients are then available for inference. However, as our recommended
approach, we adopt a different viewpoint based on the functional central limit
theorem rather than a central limit theorem for independent variables. This is
outlined in some detail in Chapter 9.
Properties of solutions to estimating equations

The reason for considering estimating equations other than (7.32) is because of
large sample properties. Without loss of generality, for any multivariate categor-
ical situation, a non-proportional hazards model (Equation 4.2) can be taken to
generate the observations. Suppose that for this more general situation we fit
the best available model, in particular the proportional hazards model (Equation
4.3). In fact, this is what always takes place when fitting the Cox model to data.
It will be helpful to have the following definition:
Definition 7.11. The average conditional variance A(β) is defined as
∞

A(β) = Eβ (Z 2 |t) − Eβ2 (Z|t) dF (t).
0
Note that the averaging does not produce the marginal variance. For that
we would need to include a further term which measures the variance of the
conditional expectations. Under the conditions on the censoring of Breslow and
Crowley (1974), essentially requiring that, for each t, as n increases, the infor-
mation increases at the same rate, then nW (t) converges in probability to w(t).
Under these same conditions, recall that the probability limit as n → ∞ of Eβ (Z|t)
under model (4.2) is Eβ (Z|t), that of Eβ (Z 2 |t) is Eβ (Z 2 |t), and that of Vβ (Z|t)
is Vβ (Z|t). The population conditional expectation and variance, whether the
model is correct or not, are denoted by E(Z|t) and V (Z|t), respectively. We
have an important result due to Struthers and Kalbfleisch (1986).
Theorem 7.3. Under model 4.2 the estimator β̂, such that U2 (β̂) = 0, con-
verges in probability to the constant βP L , where βP L is the unique solution
to the equation
∞

w−1 (t) E(Z|t) − Eβ (Z|t) dF (t) = 0, (7.35)
0
provided that A(βP L ) is strictly greater than zero.
Should the data be generated by model (4.3) then βP L = β, but otherwise

the value of βP L would depend upon the censoring mechanism in view of its
dependence on w(t). Simulation results below on the estimation of average effect
show a very strong dependence of βP L on an independent censoring mechanism.
Of course, under the unrealistic assumption that the data are exactly generated
by the model, then, for every value of t, the above integrand is identically zero,
thereby eliminating any effect of w(t). In such situations the partial likelihood
estimator is more efficient and we must anticipate losing efficiency should we use
the estimating equation U3 (β) rather than the estimating equation U2 (β).
Viewing the censoring mechanism as a nuisance feature of the data we might
ask the following question: was it possible to remove the censoring then to which
population value do we converge? We would like an estimating equation that,
in the presence of an independent censoring mechanism, produces an estimate
that converges to the same quantity we would have converged to had there been
no censoring. The above estimating equation (7.33) has this property. This is
summarized in the following theorem of Xu and O’Quigley (2000), which is an
application of Theorem 3.2 in Lin (1991).
Theorem 7.4. Under model 4.2 the estimator β̃, such that U3 (β̃) = 0, con-
verges in probability to the constant β ∗ , where β ∗ is the unique solution to
the equation
∞

E(Z|t) − Eβ (Z|t) dF (t) = 0, (7.36)
0
provided that A(β ∗ ) is strictly greater than zero.

None of the ingredients in the above equation depends on an independent cen-

soring mechanism. In consequence the solution itself, β = β ∗ , is not influenced
by the censoring. Thus the value we estimate in the absence of censoring, β ∗ ,
is the same as the value we estimate when there is censoring. A visual inspec-
tion of equations (7.35) and (7.36) suffices to reveal why we argue in favor of
(7.33) as a more suitable estimating equation than (7.32) in the presence of non-
proportional hazard effects. Furthermore, the solution to (7.33) can be given a
strong interpretation in terms of average effects. We return to this in more detail,
but we can already state a compelling argument for the broader interpretability
of β ∗ .
Large sample properties of β̃
We have that Eβ (Z|t)=S (1) (β, t)/S (0) (β, t), and that W (t) = Ŝ(t)/{nS (0) (0, t)}.
Under an independent censoring mechanism, s(1) (β(t), t)/s(0) (β(t), t) =
E{Z(t)|T = t}, and s(1) (β, t)/s(0) (β, t) is what we get when we impose a con-
stant β through time in place of β(t), both of which do not involve the cen-
soring distribution. In addition v(t) = v(β(t), t) = Var{Z(t)|T = t}. We take it
that nW (t) converges in probability to a non-negative bounded function w(t)
uniformly in t. Then we have w(t) = S(t)/s(0) (0, t). Using the same essential
approach as that of Andersen and Gill (1982) it is seen, under the model and an
independent censoring mechanism, that the marginal distribution function of T
can be written as
t
F (t) = w(t)s(0) (β(t), t)λ0 (t)dt. (7.37)
0
Theorem 7.5. (Xu 1996). Under the non-proportional hazards model and
an independent censorship the estimator β̃ converges in probability to the
constant β ∗ , where β ∗ is the unique solution to the equation
∞
s(1) (β(t), t) s(1) (β, t)
− dF (t) = 0, (7.38)
0 s(0) (β(t), t) s(0) (β, t)
∞
provided that 0 v(β ∗ , t)dF (t) > 0.
It is clear that equation (7.38) does not involve censoring. Neither then does
the solution to the equation, β ∗ . As a contrast the maximum partial likelihood
estimator β̂P L from the estimating equation U2 = 0 converges to the solution of
the equation
∞
s(1) (β(t), t) s(1) (β, t)
− s(0) (β(t), t)λ0 (t)dt = 0. (7.39)
0 s(0) (β(t), t) s(0) (β, t)
This result was obtained by Struthers and Kalbfleisch (1986). If the data be
generated by the proportional hazards model, then the solutions of (7.38) and
(7.39) are both equal to the true regression parameter β. In general, however,
these solutions will be different, the solution to (7.39) depending on the unknown
censoring mechanism through the factor s(0) (β(t), t). The simulation results of
Table 7.1 serve to underline this fact in a striking way. The estimate β̃ can be
shown to be asymptotically normal with mean zero and variance that can be
written down. The expression for the variance is nonetheless complicated and is
not reproduced here since it is not used. Instead we base inference on functions
of Brownian motion which can be seen to describe the limiting behavior of the
regression effect process.
Interpretation as average effect

In Section 7.5 the average effect, μ = β(t)dF (t), is seen to coincide, in sev-
eral special cases, with β ∗ , the solution to the large sample equivalent of the
estimating equation U3 (β). More generally, and this result is supported by the
entries in Table 7.1, β ∗ and μ = β(t)dF (t) will be so close that, for practical
purposes, that could be identified as being the same quantity. In other words,
Equation 7.38 can be viewed as an average regression effect. In the equation
s(1) (β(t), t)/s(0) (β(t), t) = E{Z(t)|T = t}, and s(1) (β ∗ , t)/s(0) (β ∗ , t) results
when β(t) is restricted to be a constant; the difference between these two is zero
when integrated out with respect to the marginal distribution of failure time.
Suppose, for instance, that β(t) decreases over time, then earlier on β(t) > β ∗
and s(1) (β(t), t)/s(0) (β(t), t) > s(1) (β ∗ , t)/s(0) (β ∗ , t); whereas later we would
have the opposite effect whereby β(t) < β ∗ and s(1) (β(t), t)/s(0) (β(t), t) <
s(1) (β ∗ , t)/s(0) (β ∗ , t). We can write: v(β, t) = ∂/∂β{s(1) (β, t)/s(0) (β, t)} and,
applying a first-order Taylor series approximation to the integrand of (7.38), we
have
∞
v(t){β(t) − β ∗ }dF (t) ≈ 0, (7.40)
0
where v(t) = v(β(t), t) = Var{Z(t)|T = t}. Therefore

∞
∗ v(t)β(t)dF (t)
β ≈ 0 ∞ (7.41)
0 v(t)dF (t)
is a weighted average of β(t) over time. According to Equation 7.41 more weights
are given to those β(t)’s where the marginal distribution of T is concentrated,
which simply means that, on average, we anticipate there being more individuals
subjected to those particular levels of β(t). The approximation of Equation 7.41
also has an interesting connection with Murphy and Sen (1991), where they show
that if we divide the time domain into disjoint intervals and estimate a constant
β on each interval, in the limit as n → ∞ and the intervals become finer at a
certain rate, the resulting β̂(t) estimates β(t) consistently. In their large sample
studies, they used a (deterministic) piecewise-constant parameter β̄(t), which is

equivalent to Equation 7.41 restricted to individual intervals. They showed that
β̄(t) is thebest approximation to β̂(t), in the sense that the integrated squared
difference {β̂(t) − β̄(t)}2 dt → 0 in probability as n → ∞, at a faster rate than
any other choice of such piecewise-constant parameters. In Equation (7.41) if
v(t), the conditional variance of Z(t), changes relatively little with time apart
from for large t, when the size of the risk sets becomes very small, we can make
the approximation v(t) ≡ c and it follows that
∞
∗
β ≈ β(t)dF (t) = E{β(T )}. (7.42)
0
In practice, v(t) will often be approximately constant, an observation supported

by our own practical experience as well as with simulated datasets. For a com-
parison of two groups coded as 0 and 1, the conditional variance is of the form
p(1 − p) for some 0 < p < 1, and this changes relatively little provided that,
throughout the study, p and 1 − p are not too close to zero. The approximate
constancy of this conditional variance is used in the sample size calculation for
two-group comparisons (Kim and Tsiatis, 1990). In fact, we only require the
weaker condition that Cov(v(T ), β(T )) = 0 to obtain Equation 7.42, a constant
v(t) being
a special case of this. Even when this weaker condition does not hold
exactly, β(t)dF (t) will still be close to β ∗ .

β1 β2 t0 % censored β∗ β(t)dF (t) β̃ β̂PL
1 0 0.1 0% 0.156 0.157 0.155 (0.089) 0.155 (0.089)
17% 0.156 0.157 0.158 (0.099) 0.189 (0.099)
34% 0.156 0.157 0.160 (0.111) 0.239 (0.111)
50% 0.156 0.157 0.148 (0.140) 0.309 (0.130)
67% 0.156 0.157 0.148 (0.186) 0.475 (0.161)
76% 0.156 0.157 0.161 (0.265) 0.654 (0.188)
3 0 0.05 0% 0.721 0.750 0.716 (0.097) 0.716 (0.097)
15% 0.721 0.750 0.720 (0.106) 0.844 (0.107)
30% 0.721 0.750 0.725 (0.117) 1.025 (0.119)
45% 0.721 0.750 0.716 (0.139) 1.294 (0.133)
60% 0.721 0.750 0.716 (0.181) 1.789 (0.168)
67% 0.721 0.750 0.739 (0.255) 2.247 (0.195)

Table 7.1: Comparison of β ∗ , β(t)dF (t), and the estimates β̃ and β̂P L .

The accuracy of β(t)dF (t) in approximating β ∗ was further studied in
simulations by Xu and O’Quigley (2000). Some of those findings are shown in
Table 7.1 and these are typical of the findings from a wide variety of other
situations. The results are indeed striking. It is also most likely true that it is not
well known just how strong is the dependence of the partial likelihood estimator
on an independent censoring mechanism when the data are generated by a non-

proportional hazards model. Since, in practical data analysis, such a situation
will almost always hold, we ought to be rather more circumspect about the usual
estimators furnished by standard software.
In the table the data are simulated from a simple two-step time-varying regres-
sion coefficients model, with baseline hazard λ0 (t) = 1, β(t) = β1 when t < t0
and β2 otherwise. The covariate Z is distributed as Uniform(0,1). At time t0 a
certain percentage of subjects at risk are censored. The value β̂P L is the partial
likelihood estimate when we fit a proportional hazards model to the data. Table
7.1 summarizes
the results of 200 simulations with sample size of 1600. We see
that β(t)dF (t) is always close to β ∗ , for the values of β that we might see
in practice. The most important observation to be made from the table is the
strong dependence of β̂P L on an independent censoring mechanism, the value
to which it converges changing substantially as censoring increases. The censor-
ing mechanism here was chosen to emphasize the difference between β̂P L and
β̃, since β̃ puts (asymptotically) the correct weights on the observations before
and after t0 . In other cases the effect of censoring may be weaker. Nonetheless,
it is important to be aware of the behavior of the partial likelihood estimator
under independent censoring and non-proportional hazards and the subsequent
difficulties in interpreting the partial likelihood estimate in general situations.
The bracketed figures in Table 7.1 give the standard errors of the estimates
from the simulations. From these we can conclude that any gains in efficiency of
the partial likelihood estimate can be very quickly lost to biases due to censoring.
When there is no censoring the estimators are the same. As censoring increases
we see differences in the standard errors of the estimates, the partial likelihood
estimate being more efficient; but we also see differences in the biases. Typically,
these latter differences are at least an order of magnitude greater.
Residuals from estimating equations

In linear regression we postulate a linear dependence of the response variable on
the explanatory variable or variables. Once unknown parameters have been fitted
we can use the model to obtain the conditional mean of the response variable
given the explanatory variables. For a “good” or well performing model we would
like for the conditional mean of the response variable to plausibly arise from the
model. Usually we will only have one observation per subject and we can study
the discrepancy between this observation and that predicted by the model. These
discrepancies are typically referred to as the residuals.
Trends in the residuals or any kind of systematic behavior not described
by the model are indications of a lack of fit or that a more complex model
is likely to provide a better description of the observations. Checking model
adequacy is a fundamental part of model construction and we return to this in
later chapters. The key difference to residuals in the proportional hazards setting
from residuals in the classical linear setting was pointed out by Schoenfeld (1982).
In the linear setting we study the discrepancies between the observed responses
and their corresponding model-based predictions. In the proportional hazards
setting, rather than being the response variable, survival, it is the covariate given
survival that provides the basis for suitable residuals. On a deeper level, there is
no real difference between these two settings once we accept that our outcome
variable is not survival (since it only determined up to its rank) and is indeed
the covariate, or covariate vector, that is observed, given T = t. This is of course
anticipated in the main theorem of this chapter.
Schoenfeld considered a vector of covariates Z that were fixed over time.
Once we fix time however, the extension of the Schoenfeld residuals to time-
dependent covariates is immediate and, aside from the need for great care in the
calculations, presents no added difficulty. The Schoenfeld residuals are defined
at each death time t by
rj (t : β) = Zj (t) − Eβ (Z | t) , (7.43)
where j indexes the individual failing at time t. These residuals have formed
the basis for many goodness-of-fit tests as well as the basis for graphical pro-
cedures, (Grambsch and Therneau, 1994; Lin, 1991; Lin et al., 1993; O’Quigley
and Pessione, 1991; Wei, 1984).
Given the basic nature of martingales as discrepancies between random vari-
ables conditional upon a history and their corresponding conditional expectation
it is natural to consider such objects as a way to create a large class of residuals.
All that is needed is the history, some function of this—technically referred to
as a filtration, and the conditional expectations. These discrepancies have the
martingale property, essentially no more than the expectations exist, and a large
class of residuals becomes available to us. These are known as martingale residu-
als and, of course, the Schoenfeld residuals are a particular case. In our view, the
Schoenfeld residuals are not just a special case of the martingale residuals but,
in practice, are the only martingale residuals of real interest. We might qualify
that statement by adding after some appropriate standardization. For this rea-
son we do not dwell on any particular features of martingale residuals and we
present little of the work that has been done on these. We recall the bare bones
nonetheless for completeness. For subject j at time t, the martingale residuals
are (Appendix B.3)
t
Mj (t) = Nj (t) − Λ̂j (t), avec Λ̂j (t) = πj (β̂, s)dN̄ (s).
0
The definition follows the usual Doob-Meyer decomposition of a counting process

into a martingale and a compensator (Andersen and Gill, 1982). For the usual
proportional hazards model with parameter β, the compensator of the process Nj
t
at time t is 0 λ(s | Zj )ds and is estimated by Λ̂j (t). If we can consider the model
to have generated the observations, then M1 (t), . . . , Mn (t) are observations of
independent martingales (Gill, 1980), which means that they should be centered
about zero and uncorrelated. These properties can be observed on a graph of
the residual, Mj (T ), j = 1, . . . , n. Barlow
and Prentice (1988) studied a class
of martingale residuals by focusing on φ(t)Mj (t)dt where φ is a predictable
function. For a predictable function, φ(t), conditioning
on the history at time
t allows us to treat φ(t) as a constant so that φ(t)Mj (t)dt will also be a
martingale. These residuals have formed the basis of tests of fit by Kay (1977),
Lin (1991), Lin et al. (1993) as well as Therneau et al. (1990). As already
mentioned, we mostly focus attention on the Schoenfeld residuals which arise
under the definition, φ = Zj (Schoenfeld, 1982).
7.7 Estimating equations in small samples
The finite sample distribution of the score statistic is considered more closely.
Since the other test statistics are derived from this, and the regression coeffi-
cient itself a monotonic function of the score, it is enough to restrict attention
to the score statistic alone. One direct approach leads to a simple convolution
expression which can be evaluated by integrated integrals. It is also possible to
make improvements to the large sample normal approximation via the use of
saddlepoint approximations or Cornish-Fisher expansions. For these we can use
the results of Corollary 7.9. Corrections to the distribution of the score statistic
can be particularly useful when the distribution of the explanatory variable is
asymmetric. Corrections to the distribution of the score equation have a rather
small impact in the case of the fourth moment but can be of significance in the
case of the third moment. The calculations themselves are uncomplicated and
simplify further in the case of an exponential distribution. Since we can transform
an arbitrary marginal distribution to one of the exponential form, while preserv-
ing the ranks, we can then consider the results for the exponential case to be of
broader generality. The focus of our inferential efforts, regardless of the particular
technique we choose, is mostly the score statistic. For this statistic, based on the
properties of the estimating equation, we can claim large sample normality.
Recall that our underlying probability model is focused on the distribution
of the covariate, or covariate vector, at each time t given that the subject is
still at risk at this time. From this the mean and the variance of the conditional
distribution can be consistently estimated and this is typically the cornerstone
for the basis of any tests or confidence interval construction. Implicitly we are
summarizing these key distributions by their means and variances or, at least, our
best estimates of these means and variances. The fact that it is the distributions
themselves that are of key interest, and not just their first two moments, suggests
that we may be able to improve the accuracy of any inference if we were to take
into account higher order moments. As a consequence of the main theorem it
turns out that this is particularly straightforward, at least for the third moment
and relatively uncomplicated for the fourth moment. In fact, corrections based
7.7. Estimating equations in small samples 177
on the fourth moment seem to have little impact and so only the third moment
might be considered in practice.
We can incorporate information on these higher moments via a Cornish-Fisher
expansion or via the use of a saddlepoint approximation. Potential improvements
over large sample results would need to be assessed on a case-by-case basis,
often via the use of simulation. Some limited simulations are given in O’Quigley
(2008) and suggest that these small sample corrections can lead to more accurate
inference, in particular for situations where there is strong group imbalance.
Edgeworth and saddlepoint approximations

For the proportional hazards model it is relatively easy to obtain Edgeworth
corrections (Appendix A) to the score statistic. For the particular case of the
multiplicative risk function the saddlepoint approximation is also straightforward.
In either case we make progress by evaluating the cumulant generating function
K(θ) given by,

K(θ) = log E(e ) = log EE(e |t) = log E(eθZ |t)dF (t),
θZ θZ
the trick of double expectation leading to quantities we can readily estimate in

view of the results of Section 7.5. Furthermore, applying Corollary 7.9, we have
that the difference between E(eθZ |t) and Eβ̂ (eθZ |t) or Eβ (eθZ |t) converges in
probability to zero. For the multiplicative model we can then use the estimate
K̂(θ) where

i j Yj (Xi ) exp[(θ + β)Zj ]
exp{K̂(θ)} = dF̂ (t),
i j Yj (Xi ) exp{βZj }
leading to the results, after some elementary manipulation, that

K̂ (0) = Eβ (Z|t)dF̂ (t) ; K̂ (0) = Vβ (Z|t)dF̂ (t).
The following results enable us to obtain the needed terms.
Lemma 7.4. Letting A(θ) = exp{K(θ)} then:

∂ p A(θ)
= Eβ (Z p (t)|t)dF̂ (t) (7.44)
∂θp θ=0
The first two derivatives of A(θ) are well known and widely available from
any software which fits the proportional hazards model. The third and fourth
derivatives are a little fastidious although, nonetheless straightforward to obtain.
Pulling all of these together we have:
Lemma 7.5. The first four derivatives of K(θ) are obtained from:
A (θ)/A(θ) = K (θ),
A (θ)/A(θ) = [K (θ)]2 + K (θ),
A (θ)/A(θ) = [K (θ)]3 + 3K (θ)K (θ) + K (θ),
A(4) /A(θ) = [K (θ)]4 + 6[K (θ)]2 K (θ) + 4K (θ)K (θ) + 3[K (θ)]2 + K (4) (θ).
We can use these results in either a saddlepoint approximation or an Edgeworth

approximation allowing us to gain greater accuracy than that provided by the
usual assumption of large sample normality. In the light of several comparative
studies of the relative merits of the two kinds of approximation we cannot really
anticipate obtaining any clear answer as to which is the best to use. It will depend
on the particular case under study and, typically, there is very little to choose
between the two. The above results can be used in both cases. Once the particular
parameters of a study (total sample size, number of groups, group imbalance,
approximate distribution of the covariates) are known, then simulations can help
answer this question. Simulations show that real advantages, in particular for the
additive model, can be obtained by making these adjustments. Some preliminary
work can lead to further simplification, and subtracting off the mean allows us
to ignore K (θ), which is equal then to zero at θ = 0. Instead of referring the
test statistic to the percentage points Zα of the normal distribution we use,
Lα = Zα + (Zα2 − 1)E{U 3 (β̂, ∞)}/6

+ (Zα3 − 3Zα ) E{U 4 (β̂, ∞)} − 3E 2 {U 2 (β̂, ∞)} /24. (7.45)
The adjustment can be used either when carrying out a test of a point hypothesis
or when constructing test-based confidence intervals. In the simulations we can
see that the correction, relatively straightforward to implement, leads to improved
control on type I error in a number of situations.
Distribution of estimating equation

The parameter β is estimated by equating to zero the score function, U (β).
Thus, U (β̂) = 0 and the distribution of β̂ can be investigated via the estimating
equation, since
Pr (β̂ < b) = Pr {U (b) < 0} (7.46)
We can use what we know about U to make statements about β̂. This only
works because U (β) is monotonic in β and the same will continue to hold in the
multivariate setting when considering components of the vector U . Formulating

the probability statement about β̂ in terms of U is particularly convenient. Two
simple illustrations of this are: (1) Bayesian inference and (2) the exact distri-
bution of a sum of independent, not necessarily identically distributed, random
variables. If we have prior information on β, in the form of the density q(β), then
we can write
Pr (β̂ < b) = Pr {U (b) < 0}q(b)db,
being an expression in terms of total probability rather than the usual Bayes’
formula since, instead of a data statistic depending on the model parameter, we
have a direct expression for the parameter estimate. For the small sample exact
distribution of the sum we use the following lemma:
Lemma 7.6. Let U1 , . . . , Un be independent, not necessarily identically dis-

tributed, continuous random variables with densities p1 (x) to pn (s), respec-

tively. Let Sn = n j=1 Uj . Then the density, qn (s) = dQn (s), of Sn is given
by
∞
dQn (s) = dQn−1 (s − u)dPn (u)du .
−∞
We use the above form dQn (s) in order to accommodate the discrete and the
continuous cases in a single expression. The lemma is proved by recurrence of an
elementary convolution result (see for example Kendall et al. (1987)). Following
Cox (1975) and Andersen et al. (1993), Andersen (1982), and Andersen and
Gill (1982) we will take the contributions to the score statistic U (Xi ) to be
independent with different distributions given by Theorem 7.2. We can then
apply the result by letting Ui = Hi (Xi ) where the i indices now run over the
k, rather than n failure times. The distribution of U1 is given by G0 (X1 ), of
U2 by G0 (X2 ) and we can then construct a sequence of equations based on the
above expression to finally obtain the distribution Qk (s) of the sum. Any prior
information can be incorporated in this expression in the same way as before.
Higher order moments

It is possible, in a way similar to that leading to the variance of the score statistic,
to obtain third and fourth moments. We can then use the estimated cumulative
hazard equation to derive data-based estimates. Since, under the model, we have
that E{U (β, t)} = 0 then
n
∞
3
E{U (β, ∞)} = E{Yi (s)Hi3 (s)}λi (s)ds. (7.47)
0 i=1
We can estimate E{U 3 (β, ∞)} consistently by replacing λi (s)ds and λi (s1 )
λj (s2 ) ds1 ds2 by R{β̂Zi (s)}dΛ̂0 (s) and R{β̂Zi (s1 )}R{β̂Zj (s2 )}dΛ̂0 (s1 )dΛ̂0 (s2 ),
respectively.
n
∞
E{U 4 (β, ∞)} = E{Yi (s)Hi4 (s)}λi (s)ds
0 i=1
∞ ∞
n
+6 E{Yi (s1 )Hi2 (s1 )}E{Yj (s2 )Hj2 (s2 )}λi (s1 )λj (s2 )ds1 ds2 .
0 0 i=1 j>i
(7.48)
Again, we can estimate E{U 4 (β, ∞)} consistently by replacing λi (s)ds

and λi (s1 ) λj (s2 )ds1 ds2 by R{β̂Zi (s)}dΛ̂0 (s) and R{β̂Zi (s1 )}R{β̂Zj (s2 )}
dΛ̂0 (s1 )dΛ̂0 (s2 ), respectively. Note that we can replace E{Yi (s)Hi2 (s)},
E{Yi (s)Hi3 (s)}, E{Yi (s)Hi4 (s)} by observed values at time of failure or by
an average taken over the risk set.
Integral transform of the baseline hazard

Note that for an arbitrary continuous distribution function, F (t), the probability
integral transform tells us that the variable Y = − log{1 − F (T )} has a standard
exponential distribution. In the case of two groups, and using F̂ in place of F, we
can transform a Kaplan-Meier curve into one approaching a standard exponential
for one group and, using the same transformation, into one approaching an expo-
nential distribution with parameter exp(β) in the other. For this reason a study
of the moment adjustments for the special case of an exponential distribution
can be of value since our interest in the baseline hazard itself is only accessory.
The necessary calculations simplify.
Being a special case of the proportional hazards model, an analysis based on
the case of a constant hazard enables us to compare our estimates with fixed pop-
ulation values. Furthermore, the estimates can still be useful when the true model
is different from the exponential one. Suppose that we have a binary covariate
(0,1) denoting group membership of which there are π1 in the first group and
π2 in the second (π1 + π2 = 1), survival time is distributed exponentially with
underlying hazard equal to 1 in the first group and eβ in the second and there is
no censoring. Then:
p t t
E Z − Eβp (Z|t) = π1 {−ψ(v)}p e−v dv + π2 {1 − ψ(v)} eβ exp(−veβ )dv,
0 0
where
π2 eβ exp(−veβ )
ψ(v) = .
π1 e−v + π2 eβ exp(−veβ )
Evaluating the above formula under β = 0, we have:
Corollary 7.12. The second, third, and fourth moments of U are given by
the following where Us is U standardized to have unit variance:
E{U 2 (0, ∞)} = nπ1 π2 ; E{Us3 (0, ∞)} = n−1/2 (π1 − π2 )(π1 π2 )−1/2
E{Us4 (0, ∞)} =n −1
(π1 + π23 )(π1 π2 )−1 + 3n−1 (n − 1).
3
More generally, consider the case of a continuous variable Z with support I and
density f . Furthermore, suppose that survival time is distributed exponentially
with underlying hazard equal to λ0 R(βz) with R(βz) = 1 for β = 0, and that
there is no censoring. Then, for β = 0 and p ≥ 2, we have:
n
∞ ∞ !
E{Yi (s)Hip (s)}λi (s)ds = {z − E(Z)}p λ0 e−λ0 t dt f (z)dz
0 i=1 I 0
and this integral can be readily evaluated so that the right-hand term becomes:
∞
{z −E(Z)}p λ0 e−λ0 t dt f (z)dz = {z − E(Z)}p f (z)dz = E{Z −E(Z)}p .
I 0 I
Therefore, the required terms can easily be evaluated from the central moments
of Z. Specifically, taking the subscript s to refer to the standardized variable, we
obtain:
Corollary 7.13. The second, third, and fourth moments of U are given, respec-
tively, by:
E(U 2 (0, ∞)) = nE{Z − E(Z)}2 ,

−3/2
E(Us3 (0, ∞)) = n−1/2 E{Z − E(Z)}3 E{Z − E(Z)}2 ,
−2
E(Us4 (0, ∞)) = n−1 E{Z − E(Z)}4 E{Z − E(Z)}2 + 3n−1 (n − 1),
where the subscript s refers to the standardized variable. Note that these results
also hold for a discrete variable Z.
Integral transform of the conditional covariate

distribution
Consider the case of a continuous covariate Z. A more sure way, albeit a more
onerous one, to correct for asymmetries in the conditional covariate distribution
is to, once again, lean on the probability integral transform. The need to evalu-
ate higher order moments then disappears since, by construction, the odd order
moments will be zero and the fourth very close to its normal counterpart. Denote
by Ĝ(z|t) the estimated conditional distribution of Z given that T = t, C > t, i.e.,
n

Ĝ(z|t) = P̂ (Z(t) ≤ z|T = t, C > t) = πj (β, t)I(Zj (t) ≤ z). (7.49)
j=1
Note that the definition of πj (β, t) restricts the subjects under consideration to
those in the risk set at time t. The cumulative distribution Ĝ(z|t) is restricted by
both z and t. We will need to invert this function, at each point Xi corresponding
to a failure. Assuming no ties in the observations (we will randomly break them
if there are any) then, at each time point Xi , we order the observations Z in the
risk set. We express the order statistics as Z(1) < Z(2) < . . . < Z(ni ) where there
are ni subjects in the risk set at time Xi . We define the estimator G̃(z|Xi ) at
time t = Xi and for z ∈ (Z(m) , Z(m+1) ) by
z − Z(m)
G̃(z|Xi ) = Ĝ(Z(m) |Xi ) + Ĝ(Z(m+1) |Xi ) − Ĝ(Z(m) |Xi ) ,
Z(m+1) − Z(m)
noting that, at the observed values Z(m) , m = 1, . . . , ni , the two estimators coin-
cide so that G̃(z|Xi ) = Ĝ(z|Xi ) for all values of z taken in the risk set at
time Xi . Otherwise, G̃(z|Xi ) linearly interpolates between adjacent values of
the observed order statistics Z(m) , m = 1, . . . , ni . Also, we are assuming no ties,
in which case, the function G̃(z|Xi ), between the values Z(1) and Z(ni ) , is a
strictly increasing function and can thereby be inverted. We denote the inverse
function by G̃−1 (α) , 0 < α < 1.
Our purpose is achieved by using, instead of G̃−1 (α) which would take us back
to where we began, the inverse of the cumulative normal distribution Φ−1 (α).
We define the transform
∗
Z(m) = Φ−1 G̃(Z(m) |Xi ), (7.50)
noting that the transform is strictly increasing so that the order of the covari-
ate observations in the risk set is respected. We are essentially transforming to
normality via the observed empirical distribution of the covariate in the risk set.
Under the null hypothesis that β = 0 the cumulative distribution G̃(Z(m) |Xi )
is discrete uniform where each atom of probability has mass 1/ni . Thus, the
∗ , m = 1, . . . , n will be close (the degree of closeness increasing with n )
Z(m) i i
to the expectation of the mth smallest order statistic from a normal sample of
size ni (Appendix A). The statistic U (β) is then a linear sum of zero mean
and symmetric variables that will be closer to normal than that for the untrans-
formed sequence. At the same time any information in the covariate is captured
via the ranks of the covariate values among those subjects at risk and so local
power to departures from the null would be model dependent. Under the null
the suggested transformation achieves our purpose, the mean of U (0) is zero
and the distribution of U (0) is symmetric. Under the alternative, however, we
would effectively have changed our model by the transformation and a choice
of model which coincides with the mechanism generating the selection from the
risk set would maximize power. The above choice would not necessarily be the
most efficient. An expression for the statistical efficiency of using some particular
covariate transformation model when another one generates the observations is
given in O’Quigley and Prentice (1991).
One way to maintain exact control over type I error using Z(m) ∗ =
Φ−1 G̃(Z(m) |Xi ) is to consider, at each observed failure time, alongside Z(m)
∗ ,
∗
its reflection in the origin −Z(m) , such values, and any more extreme in absolute
value, arising with the same probability under the null hypothesis of no effect.
A nonparametric test considers the distribution of the test statistic under all
possible configurations of the vector of dimension equal to the number of the
∗
observed failures having entries Z(m) ∗ .
or −Z(m)
The number of possibilities grows exponentially so that it is possible, with
even quite small samples, to achieve almost exact control over type I error.
The significance level is simply the number of tests with more extreme values
than those obtained by the configuration that corresponds to the observed data
themselves. This approach would be very attractive apart from the drawback
of the intensity of calculation. With as few as 10 observations per group, in a
two-group case, the number of cases to evaluate is over one million. Finding, say,
the most extreme five percent of these requires comparisons taking us into the
thousands of billions. Approximations are therefore unavoidable.
Robinson (1982) developed a simple saddlepoint approximation to the densi-

ties corresponding to paired data. Relabelling the elements of this sum as yi ,

i = 1, . . . , 2k , and the null density as ψ(u) then, letting yj tanh(λu yj ) =
uk, following Robinson’s development and regrouping terms, we obtain
12
k2
ψ(u) = 2 exp log cosh(λu yj ) − λu u ,
2π yj sech2 (λu yj )
(7.51)
For any values of u, λu is obtained as a solution to this second equation. In the

above expressions all sums range from 1 to 2k . We can obtain the significance
level by numerical integration, in particular via the use of orthogonal polynomials
(Abramowitz and Stegun, 1965). Alternatively we can work directly with an
approximation to the cumulative distribution itself (Daniels, 1983, 1954, 1980,
1987). Both require numerical approximation and the results, and effort involved,
are, for practical purposes, the same.
Illustration of small sample corrections

Rashid et al. (1982) investigated the prognostic influence of the levels of five dif-
ferent acute phase reactant proteins on survival in gastric cancer. The currently
recognized most important prognostic factor in the disease is stage, determined
at the time of operation. A staging classification simpler than the usual TNM
one was developed in the paper of Rashid and proved itself to be very predictive
of survival. On the basis of this system, it was possible to strongly discrimi-
nate among different prognostic groups. However, since the staging information
was only available at the time of surgical intervention, it could not be used to
identify the patients with a poorer prognosis for whom intervention should pos-
sibly be avoided. The rationale for the use of the acute phase reactant protein
is that any such information can be used pre-operatively. A statistical difficulty,
however, arises due to the highly non-normal, in particular highly skewed, distri-
butional behavior exhibited by these protein measurements. It has been noted by
Kalbfleisch and Prentice (1973) that outliers can have an unbounded effect on
estimates and test statistics.
A first approach might be to dichotomize, whereby all the high values are
grouped together, and indeed this is the usual way in which such data are han-
dled, the notion of “normal” and “raised” levels being common in the medical lit-
erature. Such an approach, however, sacrifices information and, given that power
may be lacking due to modest sample sizes, this may not be the best approach.
A second approach (O’Brien, 1978; O’Brien and Fleming, 1987; O’Quigley and
Prentice, 1991) transforms the explanatory variables to some familiar scale (uni-
form or normal order statistics for example). A third approach leaves the covariate
scale as observed and makes higher order corrections to the score statistic to com-
pensate for the induced lack of normality. This third approach has an advantage
over the second in that the (arbitrary) choice of scaling, necessarily impacting the
result, is avoided. The second approach has an advantage over the third in that
inference is rank invariant, not only with respect to the time variable, but also
with respect to the explanatory variable. The data are taken from Rashid et al.
(1982) where the focus was on the continuously measured variable C-reactive
protein and its impact on prognosis. In the original study, in addition to the
well-known prognostic indicators such as stage and tumor histology, there was
interest in the degree to which the pre-operative biological measurements might
on their own indicate prognostic effects. In such studies it is common to define
some kind of cut-off for such measures below which the patients are considered
to be within the normal range, and beyond which the tumor is suspected of being
particularly aggressive. The reason to consider the original measurements rather
than a new variable defined on the basis of a cut-off is that there may be a
gradual worsening of prognosis rather than any sudden phenomenon in effect.
A two-sided test based on the score statistic produces the value p = 0.034, in
reasonably close agreement with the Wald test (p = 0.042), although sufficiently
removed from the likelihood ratio test (p = 0.066) to suggest that the large
sample approximations may be slightly suspect. Carrying out a third moment

correction to the score statistic, the value p = 0.060 was obtained and increases to
p = 0.063 when we apply a fourth moment correction. This is in closer agreement
with the likelihood ratio test and indicates that the uncorrected score statistic
may be slightly underconservative. The values p = 0.06 or 0.07 appear then to
more accurately reflect the percentile under the null. Whether making a test-
based decision or using a test as a means to construct confidence intervals,
inference will be more accurate when the p-value is more accurately obtained. It
could be argued of course that, in this particular case, it would have been more
straightforward to just calculate the likelihood ratio test which seems to be more
accurate. However, in other cases, in which there is lack of agreement between
the likelihood ratio test and the score test, we have no way of knowing which
is the more reliable. Indeed all of the corrections outlined here could be equally
well applied to the likelihood ratio test (via a Taylor expansion) instead of the
score statistic.
Assessing accuracy
The most useful tool in assessing which of the several approaches is likely to
deliver the best rewards is that of simulation. It is difficult otherwise because,
even when we can show that taking into account higher moments will reduce the
order of error in an estimate, the exact value of these moments is not typically
known. The further error involved in replacing them by estimates involving error
can often lead us back to an overall order of error no less than we had in the first
place. In some cases we can carry out exact calculation. Even here though caution
is needed since if we need to evaluate integrals numerically, although there is no
statistical error involved, there is a risk of approximation error. Among the three
available tests based on the likelihood; the score test, likelihood ratio, and the
Wald test, the score test is arguably the most satisfactory. Although all three
are asymptotically equivalent, the Wald test’s sensitivity to parameterization has
raised questions as to its value in general situations. For the remaining two, the
score test (log-rank test) has the advantage of not requiring estimation under the
alternative hypothesis and has nice interpretability in terms of simple comparisons
between observed and expected quantities. Indeed it is this test, the log-rank
test in the case of a discrete covariate, that is by far the most used. The higher
moments are also evaluated very easily, again not requiring estimation under
the alternative hypothesis, and therefore it is possible to improve the accuracy
of inference based on the score test at little cost. Only tests of the hypothesis
H0 : β = 0 have been discussed. More generally, we may wish to consider testing
H0 : β = β0 , β0 = 0, such a formulation enabling us to construct confidence
intervals about non-null values of β. The same arguments apply to this case also
and, by extension, will lead to intervals with more accurate coverage properties.
Estimation and tests

1. Show that, under an independent censoring mechanism, Ĥ(t), as defined in
Section 7.4, provides a consistent estimate of H(t).
2. Show that the variance expression V (β, t) using the Andersen and Gill nota-
tion (see Section 7.4) is the same as Vβ (Z|t) using the notation of Section
7.5. Explain why Var(Z|t) is consistently estimated by Vβ̂ (Z|t) but that
Var(Z|t) is not generally equal to v(β, t).
3. For the general model, suppose that β(t) is linear so that β(t) = α0 + βt.
Show that Eβ(t) (Z k |t) does not depend upon α0 .
4. Sketch an outline of a proof that Var(Z|t) is consistently estimated by

Vβ̂ (Z|t) and that E Var(Z|t) is consistently estimated by E Vβ̂ (Z|t).

5. As for the previous question, indicate why Vβ̂ (Z|t)dF̂ (t) would be consis-
tent for E Var(Z|t).
Vβ (Z|t) =∂ Eβ (Z|t)/∂β and identify

6. Show that the conditions
for the rela-
tionship Vβ (Z|t) = ∂ Eβ (Z|t)/∂β = ∂ Eβ (Z|t) /∂β to hold.
7. Consider some parametric non-proportional hazards model (see Chapter 2),

in which the conditional density of T given Z = z is expressed as f (t|z). Sup-
pose the marginal distribution of Z is G(z). Write down estimating equations
for the unknown parameters based on the observations Zi at the failure times
Xi .
8. Use some dataset to fit the proportional hazards model. Estimate the param-
eter β on the basis of estimating equations for the observations Zi2 rather
√
than Zi . Derive another estimate based on estimating equations for Z i .
Compare the estimates.
9. Write down a set of estimating equations based on the observations Zip ,

p > 0 , i = 1, . . . , n. Index the estimate β̂ by p, i.e., β̂(p). For a given dataset,
plot β̂(p) as a function of p.
10. Use analytical or heuristic arguments to describe the expected behavior of

β̂(p) as a function of p under (1) data generated under a proportional hazards
model, (2) data generated under a non-proportional hazards model where
the effect declines monotonically with time.
11. Consider a proportional hazards model in which we also know that the
marginal survival is governed by a distribution F (t; θ) where θ is not known.
Suppose that it is relatively straightforward to estimate θ, by maximum
likelihood or by some graphical technique. Following this we base an esti-

mating
equation for the unknown regression coefficient, β, on U (β|θ̂) =
{Z(t) − Eβ (Z|t)}dF (t; θ̂). Comment on this approach and on the proper-
ties you anticipate it conferring on the estimate β̂.
12. Use the approach of the preceding question on some dataset by (1) approxi-
mating the marginal distribution by an exponential distribution, (2) approx-
imating the marginal distribution by a log-normal distribution.
13. Using again the approach of the previous two questions show that if the
proportional hazards models are correctly specified then the estimate β̂ based
on F (t; θ) is consistent whether or not the marginal model F (t; θ) is correctly
specified.
14. Supposing that the function β(t) is linear so that β(t) = α0 + βt. Show how
to estimate the function β(t) in this simple case. Note that we can use this
model to base a test of the proportional hazards assumption via a hypothesis
test that H0 : β = 0, α0 = 0 (Cox 1972).
15. Investigate the assertion that it is not anticipated for v(t), the conditional
variance of Z(t), to change much with time. Use the model-based estimates
of v(t) and different datasets to study this question informally.
16. In epidemiological studies of breast cancer it has been observed that the
tumor grade is not well modeled on the basis of a proportional hazards
assumption. A model allowing a monotonic decline in the regression coeffi-
cient β(t) provides a better fit to observed data. On the basis of observa-
tions some epidemiologists have argued that the disease is more aggressive
(higher grade) in younger women. Can you think of other explanations for
this observed phenomenon?
17. Try different weights in the weighted log-rank test and apply these to a
dataset. Suppose we decide to use the weight that leads to the most sig-
nificant result. Would such an approach maintain control over Type I error
under the null hypothesis of no association? Suggest at least two ways in
which we might improve control over the Type I error rate.
18. Use a two-sample dataset such as the Freireich data and carry out a one-
sided test at the 5% level. How does the p-value change if we make the
Edgeworth correction given in Equation 7.45.
19. Repeat the above question but this time using a saddlepoint approximation.
20. Using bootstrap resampling, calculate a 95% confidence interval for the esti-
mated regression coefficient for the above data by the percentile method
on β̂. Compare this to a 95% confidence interval obtained by inverting the
monotone function U (b) and by determining values of b for which the esti-
mated Pr {U (b) < 0} ≤ 0.025 and Pr {U (b) > 0} ≤ 0.025.
21. Consider the following two priors on β : (1) Pr (β < −1) = 0.1; Pr (β > 2) =
0.1; Pr (−1 < β < 2) = 0.8, (2) Pr (β < 0) = 0; Pr (β > 1) = 0.2; Pr (0 < β <
1) = 0.8. Using these priors repeat the above confidence interval calculations
and comment on the impact of the priors.
22. In clinical trials and many epidemiological investigations we are often in a

position to know ranges of implausible values of the regression coefficient.
Should we incorporate this knowledge into our inferential calculations, and
if so, how?
23. Either use an existing dataset or generate censored data with a single contin-
uous covariate. Evaluate the empirical distribution of the covariate at each
failure time in the risk set. Use several transformations of this distribution,
e.g., to approximately normal, exponential , or uniform, and take as a test
statistic the maximum across all considered transformations. How would you
ensure correct control of type I error for this test? What are the advantages
and drawbacks of this test?
Models and likelihood

24. Describe why the assumption of marginal independence is a stronger assump-
tion than that of conditional independence. Describe situations in which each
of these assumptions appears to be reasonable. How is the likelihood function
impacted by the assumptions?
25. Suppose that the censoring mechanism is not independent of the survival
mechanism, in particular suppose that
log Pr (T > x + u|C = u, T > u) = 2 log Pr (T > x + u|T > u).
Write down the likelihood for a parametric model for which the censoring
mechanism is governed by this equation. Next, suppose that we can take the
above equation to represent the general form for the censoring model but
that, instead of the constant value 2, it depends on an unknown parameter,
i.e., the number 2 is replaced by α. What kind of data would enable us to
estimate the parameter α?
26. As a class project, simulate data with a dependent censoring mechanism as

above with an unknown parameter α. Investigate the distribution of α̂ via
1000 simulations.
27. Fit a Weibull proportional hazards model to data including at least two
binary regressors; Z1 and Z2 . Calculate a 90% confidence intervals for the
probability that the most unfavorable prognosis among the 4 groups has
a survival greater than the marginal median. Calculate a 90% confidence
interval that a subject chosen randomly from either the most unfavorable, or
the second most unfavorable, group has a survival greater than the marginal
median.
28. As a generalization of the previous exercise, consider a parametric model with

covariate vector Z of dimension p. The p-dimensional model is considered
to provide a good approximation to the underlying mechanism generating
the observations. On the basis of the fitted p-dimensional model, write down
an expression for a confidence interval for the probability of surviving longer
than t for a subject in which Z1 , the first component of Z, is equal to z1 .
29. For the two-sample exponential model, write down the likelihood and confirm
the maximum likelihood estimates given in Section 7.3. Calculate the score
test, the likelihood ratio test, and Wald’s test, and compare these with the
expressions given in Section 7.3. For a dataset with a single binary covariate
calculate and compare the three test statistics.
30. On the basis of data, estimate the unknown regression coefficient, β, as the
expected value of the conditional likelihood (see Appendix D.4). Do this
for both an exponential based likelihood and the partial likelihood. Next,
consider the distribution of log β in this context and take an estimate as
exp E(log β). Do you anticipate these two estimators to agree? Note that the
corresponding maximum likelihood estimators do agree exactly. Comment.
Theorem 7.2 Again, since λ(t|Z(t) = z) = λ0 (t) exp{β(t)z}, then f (t|Z(t) =

z) = λ0 (t) exp{β(t)z} S(t|Z(t) = z), where S(t|Z(t) = z) is the conditional
survival function. By Bayes’ rule, we can express the conditional density of Z(t)
given T = t as
f (t|z)g(z) λ0 (t)eβ(t)z S(t|z)g(z) eβ(t)z h(z|T ≥ t)

ft (z|T = t) = = =
f (t|z)g(z)dz λ0 (t)eβ(t)z S(t|z)g(z)dz eβ(t)z h(z|T ≥ t)
where h(z|T ≥ t) is the conditional density of Z(t) given T ≥ t. From elementary

probability calculus, for sets A and B we have:
P (A) = P (A|B)P (B)/P (B|A)
so that we can write:
h(z|T ≥ t) = f (z|T ≥ t, C ≥ t)P (C ≥ t)/P (C ≥ t|z, T ≥ t)

and, by our conditional independence assumption, the denominator simplifies so

that
P (C ≥ t|z, T ≥ t) = P (C ≥ t|z).

Now, by the law of total probability, we have P (C ≥ t) = P (C ≥ t|z)g(z)dz,
where the integral is over the domain of definition of Z. Next we replace P (Z ≤

z|T ≥ t, C ≥ t) by the consistent estimate {Zj (t)≤z} Yj (t)/ n 1 Yj (t), which is
simply the empirical distribution in the risk set, which leads to

z ≤z Yi (t) exp{β(t)zi (t)}φ̂(zi , t)
P̂ {Z(t) ≤ z|T = t} = ni .
j=1 Yj (t) exp{β(t)zj (t)}φ̂(zj , t)
An application of Slutsky’s theorem enables us to claim the result continues to

hold whenever β(t) is replaced by any consistent estimator β̂(t), in particular the
partial likelihood estimator when we assume the more restricted model to hold.
Chapter 8
Survival given covariate information
8.1 Chapter summary
We begin by considering the probability that one subject with particular covariates
will have a greater survival time than another subject with different covariates,
i.e., Pr (Ti > Tj |Zi , Zj ). Note that this also provides a Kendall τ -type measure of
predictive strength (Gönen and Heller, 2005) and, although not explored in this
work, provides a potential alternative to the R2 that we recommend. Confidence
intervals are simple to construct and maintain the same coverage properties as
those for β. Using the main results of Chapter 7 we obtain a simple expression
for survival probability given a particular covariate configuration, i.e., S(t|Z ∈ H)
where H is some given covariate subspace. When the subspace is the full covariate
space then this function coincides with S(t) and the estimate coincides with the
Kaplan-Meier estimate. Simple adjustments to cater for the classical difficulty
of Kaplan-Meier estimates not necessarily reaching zero are provided. Several
different situations are highlighted including survival under informative censoring.
We would like to know how marginal survival is impacted by a knowledge of

covariate information. In the simplest case, given such information, we might
ask how much more likely is it that one individual outlasts another. One of the
main purposes of survival analysis is to obtain an estimate of the survivorship
function given certain covariate patterns. Although inference for the proportional
hazards model ignores specification of the baseline hazard rate, thereby leaving
the baseline survivorship function as well as conditional survivorship functions
undetermined, it is common to carry out further estimation on these quantities.

191
https://doi.org/10.1007/978-3-030-33439-0_8
192 Chapter 8. Survival given covariate information
The provision of such information may help guide decision making in an applied
context.
While it is usually technically difficult to estimate densities and hazards (some
kind of smoothing typically being required), it is easier to estimate cumula-
tive hazards and distribution (survivorship) functions. These have already been
smoothed, in some sense, via the summing inherently taking place.
Breslow (1972), Breslow and Crowley (1974), using an equivalence between
the proportional hazards model and a piecewise exponential regression model,
with as many parameters as there are failure times, derived a simple expression
for conditional survival given covariate information. An expression for the variance
of the Breslow estimate was derived by O’Quigley (1986). Appealing to Bayes’
rule, Xu and O’Quigley (2000) obtained the simple expression
∞ ∞
S(t|Z ∈ H) = P (Z ∈ H|u)dF (u) P (Z ∈ H|u)dF (u),
t 0
from which an estimate of conditional survival, making a direct appeal to the

main theorem, follows (O’Quigley et al., 2005; Xu and Adak, 2002; Xu and
O’Quigley, 2000). In most practical applications the two estimators behave sim-
ilarly. We prefer the second in view of its closer association with basic inference.
It is more readily generalized to deal with non-proportional hazards models, in
particular the stratified model and models that include random effects. Another
predictive quantity of interest is Pr (Ti > Tj |Zi , Zj ) which gives the probability
for an individual to outlast another given their respective covariate information.
Finally, survival, given a non-independent competing risk or non-independent
censoring, is considered (Flandre and O’Quigley, 1995; O’Quigley and Flandre,
2006).
The question we would like to answer is expressed straightforwardly as the
probability of survival time being greater than t given that Z belongs to some
subset H, i.e., Pr(T > t|Z ∈ H). Via Bayes’ rule, this probability can be imme-
diately expressed in terms of the conditional distribution of Z at T = t, together
with the marginal distribution of T . An estimate of this conditional distribution
is available as a consequence of the main theorem of Chapter 7. The unknown
marginal distribution of T can be replaced by the Kaplan-Meier estimate, or
some other distribution if we wish to investigate effects in different contexts.
We mostly limit attention to the case where the covariate Z is assumed to be
time-invariant. The situation becomes more complicated in the presence of time-
dependent covariates because of certain restrictions, for example, Z(·) should
be an external covariate in order for S(t|z) to be interpretable (Kalbfleisch and
Prentice, 2002; Keiding, 1999; Keiding and Gill, 1990; Lin, 1994; Lin and Ying,
1994).
8.3. Probability that Ti is greater than Tj 193
8.3 Probability that Ti is greater than Tj
If two individuals are independently sampled from the same distribution then, by
simple symmetry arguments, it is clear that the probability of the first having a
longer survival time than the second is just 0.5. If, instead of sampling from the
same distribution, each individual is sampled from a distribution determined by
the value of their covariate information, then, the stronger the impact of this
covariate information, the further away from 0.5 will this probability be. When
the covariates do not depend on time then this probability is very easily evalu-
ated using:
Theorem 8.1. For subjects i and j, having covariate values Zi and Zj then,
under the proportional hazards model, we can write
exp(βZj )
Pr (Ti > Tj |Zi , Zj ) = .
exp(βZi ) + exp(βZj )
An important observation to make is that the expression does not involve Λ0 (t).
If we define ψ(a, b : β) to be exp(βb)/{exp(βb) + exp(βa)} we then have:
Corollary 8.1. A consistent estimate of Pr (Ti > Tj |Zi , Zj ), under the pro-
portional hazards model is given by ψ(Zi , Zj : β̂) and Var log{ψ/(1 − ψ)} ≈
(Zj − Zi )2 Var(β̂).
The approximation in the corollary arises from an immediate application of
the mean value theorem (Appendix A). In the theorem and corollary it is assumed
that Zi and Zj are scalars and that the model involves only a one-dimensional
covariate. Extension to the multivariate case is again immediate, and instead of
β̂Zi in ψ(Zi , Zj : β̂) being a scalar it can be replaced by the usual inner product
(prognostic index). Suppose that the dimension of β and Z is p and that we use
the notation Zjr to indicate, for subject j, the rth component of Zj . Applying
the delta method (Appendix A.10),

p
p
Var log{ψ/(1 − ψ)} ≈ (Zjr − Zir )(Zjs − Zis )Cov (β̂r , β̂s ).
r=1 s=1
Use of ψ would seem to be a particularly simple and transparent way in which to

summarize the impact, or predictive strength, of regression effects. We know that
significance levels alone, directly dependent as they are on sample sizes (more
precisely the amount of uncensored observations), are not indicative of predic-
tive strength. For this we need to make use of explained variation or explained
randomness measures. Whereas these measures are being averaged over some
covariate distribution it is also helpful to have specific measures given two par-
ticular covariate configurations. If a patient is told that they have an unfavorable
prognosis in the light of studies on their prognostic variables it can be helpful to
add to that some idea on just how unfavorable it is. If ψ(Zi , Zj : β̂) remains close
1.0
0.8
Xu-O’Quigley
Kaplan-Meier
0.6
survival
0.4
0.2
0.0
0 20 40 60 80
time
Figure 8.1: Kaplan-Meier survival plot for gastric cancer data and model-based
plot based on selection of the covariate CEA to lie in interval (0,100).
to 0.5, then, regardless of significance which may be of importance in considering

the public health effects for large groups, an individual might reasonably feel that
he or she is not particularly disadvantaged by such an unfavorable prognosis. All
of the above is very straightforward when dealing with the evaluation of covariate
“points” as given by Zi and Zj .
More generally, as discussed in Section 8.4, we may wish to consider some
range within the covariate space in which case we specify Pr (Ti > Tj |Zi ∈
Hi , Zj ∈ Hj ). To do this in practice we would need to integrate or sum over
the range of points of Zi and Zj , contained in Hi and Hj , respectively, with
respect to the densities of Zi and Zj within these sets.
The connection between the population explained variation Ω2 and ψ(Zi , Zj :
β) would be worthy of further investigation. Xu (1996) showed that, for fixed
values of the covariates, and, regarding both Ω2 and ψ(Zi , Zj : β) as functions
of strength of effect as measured by β, then these two quantities are mono-
tonic increasing functions of one another. There is a one-to-one correspondence
between them so that, clearly, in some sense, they are measuring the same phe-
nomenon but on different scales.
Another potential application of ψ(Zi , Zj : β) is in relative survival where
we take Zj to be a value across some reference group or population and Zi to
be a value for some group under study, for instance, a group having recently
been treated for some particular chronic disease. The negative effects of belong-
ing to that group, as opposed to the reference group, are then directly quanti-
fied by ψ. Further developments to these ideas immediately suggest themselves
and, foremost, making use of time itself. Rather than limiting our attention to
Pr (Ti > Tj |Zi , Zj ) we can explicitly introduce into this quantity the amount of
time elapsed. We would not only condition on the values of Zi and Zj but also
that both Ti and Tj are greater than t, bring our attention then to Pr (Ti >
Tj |Zi , Zj , Ti > t, Tj > t). The resulting expressions would be more involved than
8.4. Conditional survival given Z ∈ H 195
in Theorem 8.1 but could be worked out. Applications to cure studies follow since
we could conceive of situations in which this expression diminishes with t, becom-
ing, at some point, sufficiently close to the value 0.5 to claim that the exposed
group no longer carries a disadvantage when compared to the reference group.
8.4 Conditional survival given Z ∈ H
Under the proportional hazards model, the conditional survival probability

S(t|z) = Pr (T > t|Z = z) can be estimated using the development of
Breslow (1972), Breslow and Crowley (1974) whereby
δi
Ŝ(t|z) = exp −Λ̂0 (t)eβ̂z ; Λ̂0 (t) = n . (8.1)
j=1 Yj (Xi )e
β̂Zj
Xi ≤t
An expression for the large sample variance of Y = log − log S(t|z) was obtained
by O’Quigley (1986). Symmetric intervals for Y can then be transformed into
more plausible ones, at least having better coverage properties according to the
arguments of O’Quigley (1986), by simply applying the exponential function
twice. The Breslow estimate concerns a single point z. It is a natural question
to ask what is the survival probability given that the covariates belong to some
subset H. The set H may denote for example an age group, or a certain range
of continuous measurement, or a combination of those.
In general we assume H to be a subset of the p-dimensional Euclidean space.
A natural approach may be to take the above formula, which is applied to a
point, and average a set of curves over all points belonging to the set H of
interest. For this we would need some distribution for the z across the set H.
Keiding (1995) has a discussion on expected survival curves over a historical, or
background, population, where the main approaches are to take an average of
the individual survival curves obtained from the above equation. See also Sasieni
(2003). Following that, one might use the equation to estimate S(t|z) for all
z in H, then average over an estimated distribution of Z. Xu and O’Quigley
(2000) adopted a different starting point in trying to estimate directly the survival
probabilities given that Z ∈ H. Apart from being direct, this approach is the more
natural in view of the main theorem of Section 7.5. What is more, the method
can also have application to situations in which the regression effect varies with
time. In the following, for notational simplicity, we will assume p = 1. Extensions
to p > 1 are immediate. As for almost all of the quantities we have considered it
turns out to be most useful to work with the conditional distribution of Z given
T = t rather than the other way around. Everything is fully specified by the joint
distribution of (T, Z) and we keep in mind that this can be expressed either as
the conditional distribution of T given Z, together with the marginal distribution
of Z or as the conditional distribution of Z given T, together with the marginal
distribution of T.
Using Bayes’ formula to rewrite the conditional distribution of T given infor-

mation on Z, we have:
∞
P (Z ∈ H|u)dF (u)
S(t|Z ∈ H) = t∞ . (8.2)
0 P (Z ∈ H|u)dF (u)
This is a very simple and elegant expression and we can see from it how con-
ditioning on the covariates modifies the underlying survival distribution. If H
were to be the whole domain of definition of Z, in which case Z is contained in
H with probability one, then the left-hand side of the equation simply reduces
to the marginal distribution of T . This is nice and, below, we will see that we
have something entirely analogous when dealing with sample-based estimates
whereby, if we are to consider the whole of the covariate space, then we simply
recover the usual empirical estimate. In particular this is just the Kaplan-Meier
estimate when the increments of the right-hand side of the equation are those
of the Kaplan-Meier function. The main theorem of Section 7.5 implies that
P(Z ∈ H|t) can be consistently estimated from
Lemma 8.1. P̂ (Z ∈ H|t) is consistent for the probability P (Z ∈ H|t) where

Yj (t) exp{β̂Zj }
P̂ (Z ∈ H|t) = πj (β̂, t) = H . (8.3)
{j:Zj ∈H}
Yj (t) exp{β̂Zj }
This striking, and simple, result is the main ingredient needed to obtain survival
function estimates conditional on particular covariate configurations. The rest
essentially the step increments in the Kaplan-Meier curve are readily available.
Unfortunately, a problem that is always present when dealing with censored data
remains and that is the possibility that the estimated survival function does not
decrease all the way to zero. This will happen when the largest observation is not
a failure. To look at this more closely, let F̂ (·) = 1 − Ŝ(·) be the left-continuous
Kaplan-Meier (KM) estimator of F (·). Let 0 = t0 < t1 < ... < tk be the distinct
failure times, and let W (ti ) = dF̂ (ti ) be the stepsize of F̂ at ti . If the last
observation is a failure, then,
∞
P̂ (Z ∈ H|u)dF̂ (u) t >t P̂ (Z ∈ H|ti )W (ti )
Ŝ(t|Z ∈ H) = t∞ = ki .
0 P̂ (Z ∈ H|u)dF̂ (u) i=1 P̂ (Z ∈ H|ti )W (ti )
(8.4)

When the last observation is not a failure and k1 W (ti ) < 1, an application of
the law of total probability indicates that the quantity B1 where B1 = P̂ (Z ∈
H|T > tk )Ŝ(tk ) should be added to both the numerator and the denominator
in (8.4) This is due to the fact that the estimated survival distribution is not
summing to one. Alternatively, we could simply reduce the support of the time
frame to be less than or equal to the greatest observed failure. In addition,
using the empirical estimate over all the subjects that are censored after the last
observed failure, we have:

Yj (tk +)
P̂ (Z ∈ H|T > tk ) = H , (8.5)
Yj (tk +)
where tk + denotes the moment right after time tk . Therefore we can write:
k
ti >t P̂ (Z ∈ H|ti )W (ti ) + P̂ (Z ∈ H|T > tk ){1 − 1 W (ti )}
Ŝ(t|Z ∈ H) = k k .
1 P̂ (Z ∈ H|ti )W (ti ) + P̂ (Z ∈ H|T > tk ){1 − 1 W (ti )}
The above estimate of the conditional survival function is readily calculated,

since each term derives from standard procedures of survival analysis to fit the
Cox model. An attractive aspect of the approach is that when H includes all the
possible values of z, the estimator simply becomes the Kaplan-Meier estimator of
the marginal survival function. The estimate of the conditional survival probability
P (T > t + u|T > u, Z ∈ H) can also be nicely written in the form of a simple

ratio where the numerator is given by ti >t+u Ci + B1 and the denominator by

ti >t Ci + B1 and where Ci = P̂ (Z ∈ H|ti )W (ti ). For the gastric cancer data
it is interesting to contrast the survival estimate based on these calculations
for the quantity Pr {T > t|Z1 ∈ (0, 100)}, where Z1 is the tumor marker CEA
and the simple Kaplan-Meier estimator based on the subset of the data defined
by tumor marker CEA less than 100. This is shown in Figure 8.1 and there is
good agreement between the model-based estimator of Xu-O’Quigley and the
Kaplan-Meier estimate. This, although not used here as a goodness-of-fit test in
its own right, can be taken to indicate that the model appears reasonable over
the specified range of the covariate.
Variance approximations for Ŝ(t|Z ∈ H)
In this section we assume that the proportional hazards model holds with true
β = β0 . To obtain the asymptotic variance of (8.3) at each t, we use the approach
of Link (1984) and O’Quigley (1986) which is tractable and can be computed
using standard packages for fitting the Cox regression model. Our experience
is that this provides good estimates when compared with methods such as the
bootstrap. There are two sources of variation in Ŝ(t|Z ∈ H); one caused by the
estimate of conditional probability of survival with given β, the other by the
uncertainty in β̂. Using a first-order Taylor series expansion we have:
∂ Ŝ(t|Z ∈ H)
Ŝ(t|Z ∈ H) = Ŝ(t|Z ∈ H)|β0 + (β̂ − β0 ) β=β̇ , (8.6)
∂β
where β̇ lies on the line segment between β0 and β̂. We then need to bring
together some results.
Lemma 8.2. The quantity {∂ Ŝ(t|Z ∈ H)/∂β} |β=β̇ is asymptotically constant.
Lemma 8.3. The quantity β̂ − β0 is asymptotically uncorrelated with Ŝ(t|Z ∈

H)|β0 .
Corollary 8.2. The variance of Ŝ(t|Z ∈ H) is approximated by

2
∂ Ŝ(t|Z ∈ H)
Var{Ŝ(t|Z ∈ H)} ≈ Var{Ŝ(t|Z ∈ H)|β0 } + β=β̇ Var(β̂).
∂β
(8.7)
The first term in (8.7) gives the variation due to the estimation of the conditional
survival, the second term the variation caused by β̂. Details are given at the end
of the chapter. In addition, we have:
Theorem 8.2. Under the proportional hazards model Ŝ(t|Z ∈ H) is asymptoti-
cally normal.
As a consequence one can use the above estimated variance to construct
confidence intervals for S(t|Z ∈ H) at each t.
√
Theorem 8.3. nU (β0 ) is asymptotically equivalent to n−1/2 n
1 ωi (β0 ), where
1 1
s(1) (β, t) s(1) (β, t)
ωi (β) = Zi − dNi (t) − Yi (t)eβZi Zi − λ0 (t)dt
0 s(0) (β, t) 0 s(0) (β, t)
and Ni (t) = I{Ti ≤ t, Ti ≤ Ci }.

1.0
0.8
0.6
survival
0.4
0.2
Xu-O’Quigley
Kaplan-Meier
Breslow
0.0
0 20 40 60 80
time
Figure 8.2: Survival probabilities for myeloma data based on prognostic index of
lower 0.33 percentile using Kaplan-Meier, Breslow, and Xu-O’Quigley estimators.
So all that remains to show the asymptotic normality of Ŝ(t∗ |Z ∈ H) is to show

that the numerator of Ŝ(t∗ |Z ∈ H) |β0 is also asymptotically equivalent to 1/n
times a sum of n i.i.d. random variables like the above, since the denominator
of it we know is consistent for P(Z ∈ H). To avoid becoming too cluttered we
drop the subscript of β0 in Ŝ(t∗ |Z ∈ H)|β0 . The numerator of Ŝ(t∗ |Z ∈ H)
∞ √ ∞
is t∗ P̂ (Z ∈ H|t)dF̂ (t). Note that n{ t∗ P̂ (Z ∈ H|t)dF̂ (t) − P (Z ∈ H, T >
√ ∞ √ ∞
t∗ )} = n t∗ P (Z ∈ H|t)d{F̂ (t) − F (t)} + n t∗ {P̂ (Z ∈ H|t) −
√ ∞
P (Z ∈ H|t)}d{F̂ (t) − F (t)} + n t∗ {P̂ (Z ∈ H|t) − P (Z ∈ H|t)}dF (t). Now
√
n{F̂ (t) − F (t)} converges in distribution to a zero-mean Gaussian process.
Therefore, the second term in the above expression is op (1). The last term is
A1 + op (1) where
1
√ S (H) (β0 , t) s(H) (β0 , t)S (0) (β0 , t)
A1 = n − dF (t)
t∗ s(0) (β0 , t) s(0) (β0 , t)2
n 1
Yi (t)eβ0 Zi s(H) (β0 , t)
= n−1/2 (0)
Qi − (0) dF (t).
i=1 t
∗ s (β0 , t) s (β0 , t)
As for the first term, we can use Theorem II.5 of Xu (1996), which derives from
a result of Stute (1995). With φ(t) = 1[t∗ ,1] (t)P (Z ∈ H|t) in the theorem, the
√
first term is equal to n−1/2 n i=1 νi + nRn , where |Rn | = op (n
−1/2 ) and ν’s
are i.i.d. with mean zero, each being a function of Xi and δi .
Relative merits of competing estimators

A thorough study of the relative merits of the different estimators has yet to
be carried out. For any such study the first thing to consider would be the
chosen yardstick with which to evaluate any estimate. For example, should the
whole curve be considered or only some part of it? Should a “distance” measure
be an average discrepancy, the greatest discrepancy over some range or some
weighted discrepancy? It is even very possible that some estimator would outper-
form another with respect to one distance measure and perform less well than
the competitor with respect to another distance measure. In addition to this it
is also quite possible for some estimator to maintain an advantage for certain
population situations but to lose this advantage in other situations. In the light
of these remarks it may then appear difficult to obtain a simple unequivocal find-
ing in favor of one or another estimator. Nonetheless it would be nice to know
more and further work here would be of help. In the meantime it helps to provide
us with some insight by considering various situations which have arisen when
looking at real datasets.
Prognostic biomarkers in gastric cancer

Rashid et al. (1982) studied a group of gastric cancer patients. The goal of
the study was to determine the prognostic impact of certain acute phase reac-
1.0
0.8
0.6
survival
0.4
0.2
Xu-O’Quigley
Kaplan-Meier
Breslow
0.0
0 20 40 60 80
time
Figure 8.3: Survival probabilities for myeloma data based on prognostic index
between 0.33 and 0.66 percentiles using Kaplan-Meier, Breslow, and Xu-
O’Quigley estimators.
tant proteins measured pre-operatively. This biological information could then

be used in conjunction with clinical information obtained at the time of surgical
intervention to investigate the relative prognostic impact of the different factors.
There were 104 patients and five covariates: stage (degree of invasion), ACT pro-
tein (α1 -anti chimotrypsin), CEA (carcino embryonic antigen), CRP (C-reactive
protein), and AGP (alpha glyco protein).
Although it is known that stage has strong predictive capability for survival, it
can only be determined after surgery. Of interest was the prediction of a patient’s
survival based on pre-operative measurements alone, an assessment of which
might be used to help guide clinical decision making. Values of certain covariates
such as CEA are very skew and have a wide range from below 1 to over 900.
After a log transformation of CEA, a proportional hazards model with the four
pre-operative covariates included was not rejected by the data. In fitting such a
model, CRP and AGP were found to be insignificant at 0.05 level in the presence
of ACT and logCEA. Therefore only the latter two were retained for subsequent
analysis. The regression coefficients for ACT and logCEA were calculated to be
1.817 and 0.212, with standard errors 0.41 and 0.07, respectively. This gives a
range of 0.92–4.48 for the estimated prognostic index β z (Andersen et al., 1983;
Altman and Andersen, 1986).
If we divide the patients into three groups, with low, median, and high risks,
according to the prognostic index <2, 2-3, and >3, we can predict the survival
probabilities in each risk group. It was then possible to estimate the survival
curves for these three groups, and these were calculated by Xu and O’Quigley
(2000). The curves can be compared with the empirical Kaplan-Meier curves
which make no appeal to the model. Agreement is strong. We also chose to
define the set H by all those patients having values of CEA less than 50. This
8.5. Other relative risk forms 201
1.0
0.8
0.6
survival
0.4
0.2
Xu-O’Quigley
Kaplan-Meier
Breslow
0.0
0 20 40 60 80
time
Figure 8.4: Survival probabilities for myeloma data based on prognostic index for
values greater than the 0.33 percentile using 3 different estimators.
was not very far from the median value and provided enough observations for a
good empirical Kaplan-Meier estimate. The empirical estimate and the model-
based estimates show good agreement. For the three prognostic groups, defined
on the basis of a division of the prognostic index from the multivariate model,
the empirical Kaplan-Meier estimates, the Breslow estimates, and the Xu and
O’Quigley estimates are shown in Figures 8.2, 8.3, and 8.4. Again agreement is
strong among the three estimators.
8.5 Other relative risk forms
The estimator described above is not limited to proportional hazards models. The
formula itself came from a simple application of Bayes’ rule, and the marginal
distribution of T can always be estimated by the Kaplan-Meier estimator or some
other estimator if we wish to use other assumptions. Taking a general form of
relative risk r(t; z), so that λ(t|z) = λ0 (t)r(t; z). Assume also that r(t; z) can
be estimated by r̂(t; z), for example, that it has a known functional form and a
finite or infinite dimensional parameter that can be consistently estimated. Special
cases of r(t; z) are exp(βz), 1 + βz, and exp{β(t)z}. Since the main theorem
of Section 7.5 extends readily to other relative risk models it is straightforward
to derive analogous results to those above. We are still able to estimate S(t|Z ∈
H), with P̂ (Z ∈ H|t) = {j:Zj ∈H} πj (t). This is an important extension of
the estimator since we may wish to directly work with some form of a non-
proportional hazards model.
Stratified and random effects models

In studies where there is stratification, in particular, when there are many strata,
the approach can be a relatively simple one to estimate conditional survivorship
as it does not require the estimation of the baseline hazards. Suppose that V
is the stratification variable, and we are interested in the survival given Z ∈ H.

Note P (Z ∈ H|t) = v P (Z ∈ H|t, V = v)P (V = v). We can estimate P (Z ∈

H|t, V = v) by {j:Zj ∈H} πjv (β̂, t), where πjv (β, t) is the conditional probability
defined within strata v, and estimate P (V = v) by the empirical distribution of
V . Similarly to the stratified case, (8.3) can also be used to estimate survival
under random effects models arising from clustered, such as genetic or familial
data. The frailty models under such settings can be written as
λij (t) = λ0 (t)ωi exp{βzij }, (8.8)
where λij is the hazard function of the jth individual in the ith cluster. This
is the same as a stratified model, except that we do not observe the values of
the “stratification variable” ω; but such values are not needed in the calculation
described above for stratified models. So the procedure described above can be
used to estimate S(t|Z ∈ H). In both cases considered here, we need reasonable
stratum or cluster sizes in order to get a good estimate of P (Z ∈ H|t, v) or
P (Z ∈ H|t, ω).
Conditional independent censorship

It is also possible to generalize (8.3) to the cases where C and T are independent
given Z. First let us assume that the covariate Z is discrete with finitely many
categories. Then, the KM estimate can be replaced by a weighted Kaplan-Meier
(WKM) estimate (Murray and Tsiatis, 1999, 2001, 1996), which still consistently
estimates F (·) under the conditional independent censorship. The WKM estimate
calculates the subgroup KM estimates within each category of the covariate
values, and then weights these subgroup estimates by the empirical distribution
of Z. For the conditional distribution of Z given T = t, from the proof of the
main theorem (Section 7.5),
eβz S(t|z)g(z)
f (z|T = t) = . (8.9)
eβz S(t|z)g(z)dz
If we estimate S(t|z) by the subgroup KM estimate within the category of value

z, and the marginal g(z) by the empirical probabilities, we are still able to consis-
tently estimate the conditional distribution of Z given T = t and thus S(t|Z ∈ H).
For other types of covariate distribution such as continuous covariates, we need
to incorporate the covariates into categories, and Murray and Tsiatis (1996)
suggested guidelines which could be useful in practice.
8.5. Other relative risk forms 203
Sparse data in high dimensions

When data are sparse in high dimensions, i.e., multiple covariates, the πj (β̂, t)’s
used to estimate the conditional distribution of Z given T = t may encounter
some difficulties, because they are like empirical distributions. In fact, they are
obtained through the empirical distribution of Z given T ≥ t. In this case, as
seen in the gastric cancer example, we consider H as the Cartesian product of
the range for that component with (−∞, ∞) for the remaining components. Of
course the actual range of the covariate vector is reflected in the data itself.
However, sometimes we might specify H as a relatively small “block” in a p-
dimensional space. For example, for the gastric cancer data, we might ask what
is the survival probabilities given that CEA is greater than 10 but less than 15. In
this type of situation there might be so few observations in H that quite early on
the risk sets may no longer contain any subjects with covariates in H, allowing
for {πj (β, t)}j to be a sufficiently reliable estimate of the conditional distribution
of Z given T = t.
In many practical cases one is likely to ask for the conditional survival prob-
abilities given a range for a single component of the covariate vector, while
allowing the other components to vary freely. We can proceed as follows. Denote
the prognostic index η = β z. Under the model two individuals should have the
same survival probabilities as long as η1 = η2 . That is, conditioning on Z ∈ H is
equivalent to conditioning on η ∈ β H. Therefore, S(t|Z ∈ H) = S(t|η ∈ β H),
where β H = {β z|z ∈ H} and should contain potentially more observations than
H does. This way the p-dimensional vector of covariates is reduced to the one-
dimensional prognostic index, and the same number of observations in the risk
set is now used to estimate the one-dimensional conditional distribution of η.
Thus Ŝ(t|Z ∈ H) becomes:

t >t P̂ (η ∈ β H|ti )W (ti ) + P̂ (η ∈ β H|T > tk )Ŝ(tk )
Ŝ(t|η ∈ β H) = ki ,
i=1 P̂ (η ∈ β H|ti )W (ti ) + P̂ (η ∈ β H|T > tk )Ŝ(tk )

where P̂ (η ∈ β H|t) = {j:ηj ∈β̂ H} πj (β̂, t), and

P̂ (η ∈ β H|T > tk ) = Yj (tk +)/ Yj (tk +).
{j:ηj ∈β̂ H}
As before, since one can consistently estimate β, the above expression still pro-
vides a consistent estimate of S(t|Z ∈ H). Note that when z is a single covariate,
z ∈ H is exactly the same as η ∈ βH (unless β = 0 in which case the covariates
have no predictive capability), so the above is consistent with the one-dimensional
case developed earlier. While we regard the above as one possible approach under
high dimensions when there are not “enough" observations falling into the ranges
of covariates of interest, the variance estimation and the asymptotic properties

seem to be more complicated as the estimate of β enters both η and the set β H.
When there is a need to use the estimate based on β H, resampling methods such
us bootstrap could be employed for inferential purposes. There are many potential
areas for research here in order to further develop these techniques. Comparative
studies would be of value. These would not be straightforward since compar-
ing competing estimated curves requires considering the whole of the curve. For
instance, at some points in time one estimator may outperform another whereas,
at a later point, it may be the converse that holds. Some measure of overall
distance such as the maximum or average discrepancy could be considered.
8.6 Informative censoring
Events that occur through time, alongside the main outcome of interest, may
often provide prognostic information on the outcome itself. These can be viewed
as time-dependent covariates and, as before, in light of the main theorem (Section
7.5), it is still straightforward to use such information in the expression of the
survivorship function. Since, in essence, we sum, or integrate, future information,
it can be necessary to postulate paths that the covariate process might take.
Emphasis remains on the conditional distribution of survival given current and
evolving covariate information. This differs slightly from an approach, more even
handed with respect to the covariate process alongside the survival endpoints,
that makes an appeal to joint modeling. Even so, the end goal is mostly the
same, that of characterizing the survival experience given covariate information.
Paths that remain constant are the easiest to interpret and, in certain cases,
the simple fact of having a value tells us that the subject is still at risk for the
event of interest Kalbfleisch and Prentice (2002). The use of the main theorem
(Section 7.5) in this context makes things particularly simple since the relevant
probabilities express themselves in terms of the conditional distribution of the
covariate at given time points. We can make an immediate generalization of
Equation 8.2 if we also wish to condition on the fact that T > s. We have:
∞
P (Z ∈ H|u)dF (u|u > s)
S(t + s|Z ∈ H, T > s) = t+s
∞ ,
s P (Z ∈ H|u)dF (u|u > s)
and, in exactly the same way as before, and assuming that the last observation
is a failure, we replace this expression in practice by its empirical equivalent
∞
P̂ (Z ∈ H|u)dF̂ (u|u > s) t >t+s P̂ (Z ∈ H|ti )W (ti )
Ŝ(t + s|Z ∈ H, T > s) = t+s
∞ = i .
s
P̂ (Z ∈ H|u)dF̂ (u|u > s) ti >s P̂ (Z ∈ H|ti )W (ti )
8.6. Informative censoring 205

When the last observation is not a failure and k1 W (ti ) < 1 we can make a
further adjustment for this in the same way as before.
One reason for favoring the Xu-O’Quigley estimate of survival over the Bres-
low estimate is the immediate extension to time-dependent covariates and to
time-dependent covariate effects. Keeping in mind the discussion of Kalbfleisch
and Prentice (2002), concerning internal and external time-dependent covariates,
whether or not these are determined in advance or can be considered as an indi-
vidual process generated through time, we can, at least formally, leaving aside
interpretation questions, apply the above formulae. The interpretation questions
are solved by sequentially conditioning on time as we progress along the time axis
and, thereby, the further straightforward extension which seeks to quantity the
probability of the event T > t, conditioned by T > s where s < t, is particularly
useful.
Survival estimation using surrogate endpoints

For certain chronic diseases, a notable example being HIV, researchers have con-
sidered the need for clinical evaluation that might yield quicker results. Surrogate
endpoints, viewed as time-dependent covariates, can help in addressing this issue
and have received attention in the medical statistical literature (Ellenberg and
Hamilton, 1989; Hillis and Seigel, 1989; Wittes et al., 1989). Herson (1989)
wrote that “a surrogate endpoint is one that an investigator deems as correlated
with an endpoint of interest but that can perhaps be measured at lower expense
or at an earlier time than the endpoint of interest.” Prentice (1989) defined a
surrogate endpoint as “a response variable for which a test of the null hypothesis
of no relationship to the treatment groups under comparison is also a valid test
of the corresponding null hypothesis based on the true endpoint.”
We can view a surrogate endpoint to be a time-dependent response variable
of prognostic value obtained during follow-up, which indicates an objective pro-
gression of disease. In the survival context a surrogate variable is most often
a discrete endpoint indicating, in some way, disease progression. Flandre and
O’Quigley (1995) proposed a two-stage procedure for survival studies when a
surrogate or intermediate time-dependent response variable is available for some
patients. In the presence of a time-dependent intermediary event we can write:
∞ t ∞
S(t) = S(t|c)g(c)dc = S(t|c < t)g(c)dc + S(t|c ≥ t)g(c)dc.
0 0 t
If the effect of the occurrence of the intermediary event is to change the hazard
function of death λ(t) from λ1 (t) to λ2 (t), that is: λ(t) = λ1 (t) if t ≤ C and
is equal to λ2 (t) otherwise then λ1 (t) = λ2 (t) when the intermediary response
variable has no influence on survival. When λ2 (t) < λ1 (t) or λ2 (t) > λ1 (t) then
the intermediary, or surrogate, response variable carries relatively favorable or
unfavorable predictions of survival. Thus the quantity π(t) = λ2 (t)/λ1 (t) is a
measure of the effect of the surrogate response on survival. When f1 (t) and
f2 (t) are the density functions, S1 (t) and S2 (t) the survivorship functions corre-
sponding to the hazard functions λ1 (t) and λ2 (t), respectively, then the marginal
survival function is
t c t t
S(t) = exp − λ1 (u)du + λ2 (u)du dG(c) + exp − λ1 (u)du G(t).
0 0 c 0
In the first stage of a two-stage design, all patients are followed to the end-
point of primary concern and for some subset of the patients there will be the
surrogate information collected at an intermediary point during the follow-up.
The purpose of the first stage is to estimate the relationship between the occur-
rence of the surrogate response variable and the remaining survival time. This
information can then be used in the second stage, at which time, for patients
who reach the surrogate endpoint, follow-up is terminated. Such patients could
be considered as censored under a particular dependent censorship model, the
censoring being, in general “informative.” The Kaplan-Meier estimator will not
generally be consistent if the survival time and an informative censoring time are
dependent but treated as though they were independent. Flandre and O’Quigley
(1995) proposed a nonparametric estimator of the survival function for data col-
lected in a two-stage procedure. A nonparametric permutation test for comparing
the survival distributions of two treatments using the two-stage procedure is also
readily derived.
The idea behind a two-stage design in the context of a time-dependent sur-
rogate endpoint is to reduce the overall duration of the study. This potential
reduction occurs at the second stage, where follow-up is terminated on patients
for whom the surrogate variable has been observed. The first stage is used to
quantify the strength of the relationship between occurrence of the surrogate
variable and subsequent survival. It is this information, obtained from the first
stage analysis, that will enable us to make inferences on survival on the basis of
not only observed failures, but also observed occurrences of the surrogate. In the
context of clinical trials, as pointed out by Prentice (1989), the surrogate variable
must attempt to “capture” any relationship between the treatment and the true
endpoint. We may wish to formally test the validity of a surrogate variable before
proceeding to the second stage using a standard likelihood ratio test.
In the first stage N1 patients are enrolled and followed to the endpoint of
primary concern (e.g., death) or to censorship, as in a classical study, and infor-
mation concerning the surrogate variable is recorded. Survival time is then either
the true survival time or the informative censoring time. The information avail-
able for some patients will consist of both time until the surrogate variable and
survival time, while for others (i.e., those who die or are censored without the
occurrence of the surrogate variable) it consists only of survival time. In the sec-
ond stage a new set of patients (N2 ) is enrolled in the study. For those patients,
8.6. Informative censoring 207
the follow-up is completed when the surrogate variable has been observed. Thus,
the information collected consists only of one time, either the time until the sur-
rogate variable is reached or the survival time. In some cases the two stages may
correspond to separate and distinct studies; the second stage being the clinical
trial of immediate interest while the first stage would be an earlier trial carried
out under similar conditions.
When parametric models are assumed, then the likelihood function can be
obtained directly and provides the basis for inference. The main idea follows
that of Lagakos et al. (1978), Lagakos (1976, 1977) who introduced a stochastic
model that utilizes the information on a time-dependent event (auxiliary variable)
that may be related to survival time. By taking λ0 (t) = λ2 , where λ0 (.) is the
hazard function for the occurrence of the surrogate response, λ1 (t) = λ1 , and
λ2 (t) = λ3 , the survival function has the marginal distribution function given
by Lagakos. The Lagakos model itself is a special case of the bivariate model
of Freund (1961), applicable to the lifetimes of certain two-component systems
where a failure of the first or second component alters the failure rate of the
second or first component from β to β or (α to α ). By taking α = λ1 (t),
α = λ2 (t), and β = β = λ0 (t), the Freund model can be viewed as a special
case of the model described above.
Slud and Rubinstein (1983) make simple nonparametric assumptions on the
joint density of (T, C) and consider the function ρ(t) defined by
Pr (t < T < t + δ|T > t, C < t)

ρ(t) = lim .
δ→ 0 Pr (t < T < t + δ|T > t, C ≥ t)
It is possible to make use of this function in order to derive a nonparametric

estimation of S(t) for use with the two-stage procedure with surrogate endpoints
and which accounts for the dependent censoring. The authors present a class of
nonparametric assumptions on the conditional distribution of T given C which
leads to a consistent generalization of the Kaplan-Meier survival curve estimator.
Individual values of ρ(t) > 1 mean that, after censoring, the hazard risk of death
for the patient is increased, and thus if we make an assumption of independence
of T and C, the Kaplan-Meier estimator will tend to overestimate survival. On the
other hand, if ρ(t) < 1, the independence assumption will lead to underestimation
of the survival curve. Recalling the usual notation whereby δi = 1 if Ti ≤ Ci and
Xi = min(Ti , Ci ), then Xi defines the observed portion of survival time. The
Slud and Rubinstein estimator is given by
⎧ ⎫
⎨
d(t)−1

d(t)
n − 1 ⎬
Ŝρ (t) = N −1 n(t) +
i
Wk , (8.10)
⎩ ni + ρi − 1 ⎭
k=0 i=k+1

where n(t) = I(Xi > t), nj = i I(Xi ≥ Xj ), d(t) = I(δi = 1, Xi ≤ t),
ρi = ρ(ti ) and Wj is the number of patients censored between two consecutively
ordered failure times Xj and Xj+1 . When ρi = 1 it follows that Ŝρ (t) reduces to
the usual Kaplan-Meier estimator. This model is a special case of the nonpara-
metric assumption presented by Slud and Rubinstein. The focus here is not on
the dependence of T and C but on the dependence of T and Cs where Cs is a
dependent censoring indicator, in particular a surrogate endpoint. The function
of interest is
Pr (t < T < t + δ|T > t, Cs < t)
ρs (t) = lim .
δ→ 0 Pr (t < T < t + δ|T > t, Cs ≥ t)
This function is equivalent to the function π(t) and can be estimated from data
from the first stage. Suppose that the conditional hazard, λ(t|z), of death at
t given Z = z has the form h0 (t) exp(βz(t)) where zi (t) takes the value 0 if
ti ≤ ci and value 1 if ti > ci then ρs (t) = ρs = exp(β). Thus, an estimate of ρs
is given by exp(β̂). The estimate of β using data from the first stage quantifies
the increase in the risk of death occurring after the surrogate variable has been
observed. The first stage is viewed as a training set of data to learn about the
relationship between the potential surrogate endpoint and the survival time.
The estimator is constructed from the entire sample N (N = N1 +N2 ). In the
sample of size N , the ordered failure times Xi for which δi = 1 are X1 ≤ . . . ≤ Xd ,
where d is the number of deaths of patients enrolled either in the first stage or in
the second stage. Using the notation that i = 1 if Xi > ci and 0 otherwise, the
random variable V defines either the observed survival time or the time to the
surrogate variable. For patients in the second stage, let us denote the number
of Xi with i = 1 and δi = 0 between Xj and Xj+1 by Wj . In this way, Wj
denotes the number of individuals from the second stage having a surrogate

response between two consecutive failure times Xj and Xj+1 . Let Wj denote
either the number of individuals censored in the first stage or the number of
individuals censored without a surrogate response in the second stage between
Xj and Xj+1 . Clearly, Wj is the number of patients censored with “informative
censoring” between two consecutively ordered failure times Xj and Xj+1 while

Wj is the number of patients censored with “non-informative censoring” between
two consecutively ordered failure times Xj and Xj+1 . Finally nj is the number
of i with Xi ≥ Xj . The product-limit estimator then becomes
⎧ ⎫
⎨
d(t)−1

d(t)

d(t)−1
ni − 1 ⎬
d(t)
ni − 1
Ŝ(t; ρs ) = N −1 n(t) + Wk + Wk .
⎩ ni + ρs − 1 ni ⎭
k=0 i=k+1 k=0 i=k+1
Notice that when ρs = 1 (i.e., the occurrence of the surrogate variables has
no influence of survival) then Ŝ(t; ρs ) is simply the Kaplan-Meier estimator.
Considering the two-sample case, for example, a controlled clinical trial where
group membership is indicated by a single binary variable Z, then the survival
function, corresponding to z = 0, is denoted by S0 (t) and the survival function

corresponding to z = 1 is denoted by S1 (t). A nonparametric permutation test
for testing the null hypothesis H0 : S0 = S1 versus H1 : S0 = S1 can be easily
constructed (Flandre and O’Quigley, 1995). The test accounts for the presence
of dependent censoring as described and, since we would not want to assume
that the effect of the surrogate is the same in both groups, then ρs0 need not
be equal to ρs1 . A test can be based on the statistic Y defined by
∞
Y (w; ψ) = w(t)ψ(Ŝ0 (t; ρ̂s0 ), Ŝ1 (t; ρ̂s1 ))dŜ0 (t; ρ̂s0 ), (8.11)
0
where ψ(a, b) is some distance function (metric) between a and b at the point
s and w(s) is some positive weighting function, often taken to be 1 although a
variance-stabilizing definition such as w(t)2 = Ŝ0 (t; ρ̂s0 )(1 − Ŝ0 (t; ρ̂s0 )) can be
useful in certain applications. The choice ψ(a, b) = a − b leads to a test with good
power against alternatives of stochastic ordering, whereas the choice ψ(a, b) =
|a − b| or ψ(a, b) = (a − b)2 , for instance, would provide better power against
crossing hazard alternatives. Large sample theory is simplified by assuming that
the non-informative censoring distributions are identical in both groups and this
may require more critical examination in practical examples. Given data we can
observe some value Y = y0 from which the significance level can be calculated by
randomly permuting the (0, 1) labels corresponding to treatment assignment. For
the ith permutation (i = 1, . . . , np ) the test statistic can be calculated, resulting
in the value say yi . Out of the np permutations suppose that there are n+ values
of yi greater than or equal to y0 and, therefore, np − n+ values of yi less than y0 .
The significance level for a two-sided test is then given by 2 min(n+ , np −n+ )/np .
In practice we sample from the set of all permutations so that np does not
correspond to the total number of possible permutations but, rather, the number
actually used, of which some may even be repeated. This is the same idea that
is used in bootstrap resampling.
1. For subjects i and j with covariate values Zi and Zj , write down the proba-
bility that the ratio of the survival time for subject i to the survival time for
subject j is greater than 2/3.
2. Calculate an estimate of the above probability for the Freireich data in which
Zi = 1 and Zj = 0. Derive an expression for a 95% confidence interval for this
quantity and use it to derive a confidence interval for use with the Freireich
data.
3. Referring to the paper of Kent and O’Quigley (1988), consider the approx-
imation suggested in that paper for the coefficient of randomness. Use this
to motivate the quantity Pr (Ti > Tj |Zi , Zj ) as a measure of dependence

analogous to explained variation.
4. Consider some large dataset in which there exist two prognostic groups.
Divide the time scale into m non-overlapping intervals, a0 = 0 < a1 < ... <
am = ∞. Calculate Pr (Ti > Tj |Zi , Zj , Tj > ak ) for all values of k less than
m and use this information to make inferences about the impact of group
effects through time.
5. Write down the likelihood for the piecewise exponential model in which,
between adjacent failure times, the hazard can be any positive value (Bres-
low, 1972). Find an expression for the cumulative hazard function and use
this to obtain an expression for the survivorship function. Although such
an estimator can be shown to be consistent (Breslow and Crowley, 1974),
explain why the usual large sample likelihood theory would fail to apply.
6. Consider again the two-group case. Suppose we are told that the survival
time of a given subject is less than t0 . We also know that the groups are
initially balanced. Derive an expression for the probability that the subject in
question belongs to group 1. For the Freireich data, given that a subject has
a survival time less than 15 weeks, estimate the probability that the subject
belongs to the group receiving placebo.
7. Derive an estimator analogous to Equation 8.4 but based on the Nelson-

Aalen estimator for marginal survival. By adding to the marginal cumulative
hazard some arbitrary increasing unbounded function of time, obtain a sim-
pler expression for the conditional survival estimate.
8. Use the delta method (Section A.10) to obtain Equation 8.7.
9. Use the results of Andersen and Gill (1982) to conclude that β̂ −β0 is asymp-
totically uncorrelated with Ŝ(t|Z ∈ H)|β0 .
10. On the basis of a large dataset construct some simple prognostic indices
using the most important risk factors. Divide into 3 groups the data based
on the prognostic index and calculate the different survival estimates for each
subgroup. Comment on the different features of these estimators as observed
in this example. How would you investigate more closely the relative benefits
of the different estimators?
11. Show that when Ti and Tj have the same covariate values, then Pr (Ti >
Tj ) = 0.5.
12. For the Freireich data calculate the probability that a randomly chosen sub-
ject from the treated group lives longer than a randomly chosen subject from
the control group.
13. Generate a two-variate proportional hazards model in which Z1 is binary and

Z2 is uniform on the interval (0, 1). The regression coefficients β1 and β2
take values (i) 0.5, 0.5; (ii) 0.5, 1.0; and (iii) 1, 2, respectively. For all three
situations, and taking λ0 = 1, calculate and compare the six survival curves
for Z1 = 0, 1 and Z2 = 0.25, 0.50, and 0.75.
Theorem 8.2 We know how to estimate the asymptotic variance of β̂ under the
model. So all that remains for the second term on the right-hand side of (8.7) is
to calculate the partial derivative of Ŝ(t|Z ∈ H) with respect to β. For this we
have:

∂ ( ti >t Di )( ti ≤t Ci ) − ( ti ≤t Di )( ti >t Ci + B1 )
Ŝ(t|Z ∈ H) = ,
∂β ( ki=1 Ci + B1 )2
(8.12)
where Ci and B1 are the same as above and

∂
Di = P̂ (Z ∈ H|ti )W (ti ),
∂β
with
∂
P̂ (Z ∈ H|t) = P̂ (Z ∈ H|t){Eβ (Z|t; π H ) − Eβ (Z|t; π)},
∂β
where
Eβ (Z|t; π) = Yj (t)Zj exp{β̂Zj }/ Yj (t) exp{β̂Zj },
and
Eβ (Z|t; π H ) = Yj (t)Zj exp{β̂Zj }/ Yj (t) exp{β̂Zj }.
H H
The first term on the right-hand side of (8.7) can be estimated using Greenwood’s
formula
⎧ ⎫
⎨ q̂ ⎬
ˆ Ŝ(t|Z ∈ H)|β } ≈ Ŝ(t|Z ∈ H)|2
Var{ (1 +
i
) − 1 , (8.13)
0 β0
⎩ ni p̂i ⎭
ti ≤t
where k
Cj + B1
p̂i = P̂ (T > ti |T > ti−1 , Z ∈ H) = i+1 k
,
i Cj + B1

q̂i = 1 − p̂i and ni = Yj (ti ). Then each p̂i is a binomial probability based
on a sample of size ni and Ŝ(t|Z ∈ H) = ti ≤t p̂i . The p̂i ’s may be treated
as conditionally independent given the ni ’s, with β0 fixed. Thus, Greenwood’s
formula applies. All the quantities involved in (8.12) and (8.13) are those routinely
calculated in a Cox model analysis.
Theorem 8.3 Xu (1996) derived the asymptotic normality of Ŝ(t|Z ∈ H) under

the proportional hazards model at fixed points t = t∗ . Let’s take Qi = 1 if Zi ∈ H,
and 0 otherwise and follow, fairly closely, the notational set-up of Andersen and
Gill (1982) and Gill from which
β(t) Zi (t) Z (t)r ,
S (r) (t) = n−1 n i=1 Yi (t)e i s(r) (t) = ES (r) (t),
(r) −1

(t) (r)
S (β, t) =n n
Y (t)e i Zi (t) ,
β Z r s (β, t) = ES (r) (β, t),
i=1 i
S (H) (β, t) = n−1 n Q Y (t)eβZi , s(H) (β, t) = ES (H) (β, t),
1n i i
S (H1) (β, t) =n −1
1 Qi Yi (t)Zi e
βZ i , s(H1) (β, t) = ES (H1) (β, t),
for r = 0, 1, 2,. Next rewrite
P̂ (Z ∈ H|t) = S (H) (β̂, t)/S (0) (β̂, t) , Eβ (Z|t; π H ) = S (H1) (β̂, t)/S (H) (β̂, t).
Using the main theorem of Section 7.5 we have s(H) (β0 , t)/s(0) (β0 , t) = P (Z ∈
H|t). Under the usual regularity and continuity conditions (Xu, 1996) it can
be shown that {∂ Ŝ(t∗ |Z ∈ H)/∂β} |β=β̇ is asymptotically constant. Now β̂ −
β0 = I −1 (β̌)U (β0 ) where β̌ is on the line segment between β̂ and β0 , U (β) =
∂ log L(β)/∂β and I(β) = −∂U (β)/∂β. Combining these we have:
√ √ √ ∂ Ŝ(t∗ |Z ∈ H)
nŜ(t∗ |Z ∈ H) = nŜ(t∗ |Z ∈ H)|β0 + I −1 (β̌) nU (β0 ) β=β̇ .
∂β
Andersen and Gill (1982) show that I(β̌) converges in probability to a well-defined
population parameter. In the following theorem, Lin and Wei (1989) showed that
U (β0 ) is asymptotically equivalent to 1/n times a sum of i.i.d. random variables:
√
Theorem 8.4. (Lin and Wei, 1989) nU (β0 ) is asymptotically equivalent to

n−1/2 × n1 ωi (β0 ), where Ni (t) = I{Ti ≤ t, Ti ≤ Ci } and
1 1
s(1) (β, t) s(1) (β, t)
ωi (β) = Zi − (0) dNi (t)− Yi (t)eβZi Zi − λ0 (t)dt.
0 s (β, t) 0 s(0) (β, t)
It only then remains to show the asymptotic normality of Ŝ(t∗ |Z ∈ H). We see
that the numerator of Ŝ(t∗ |Z ∈ H) |β0 is asymptotically equivalent to 1/n times
a sum of n i.i.d. random variables like the above, since the denominator of it we
know is consistent for P(Z ∈ H). We drop the subscript of β0 in Ŝ(t∗ |Z ∈ H)|β0 .
∞
The numerator of Ŝ(t∗ |Z ∈ H) is t∗ P̂ (Z ∈ H|t)dF̂ (t). Note that
∞
√
n{ P̂ (Z ∈ H|t)dF̂ (t) − P (Z ∈ H, T > t∗ )}
t∗
∞
√
= n P (Z ∈ H|t)d{F̂ (t) − F (t)}
t∗
∞
√
+ n {P̂ (Z ∈ H|t) − P (Z ∈ H|t)}d{F̂ (t) − F (t)}
∗
t ∞
√
+ n {P̂ (Z ∈ H|t) − P (Z ∈ H|t)}dF (t).
t∗
√
Now n{F̂ (t)−F (t)} converges in distribution to a zero-mean Gaussian process.
Therefore the second term on the right-hand side of the preceding equation is
op (1). The last term is A1 + op (1) (see also Lemma II.4 of Xu (1996), where
1
√ S (H) (β0 , t) s(H) (β0 , t)S (0) (β0 , t)
A1 = n − dF (t)
t∗ s(0) (β0 , t) s(0) (β0 , t)2
n 1
−1/2 Yi (t)eβ0 Zi s(H) (β0 , t)
=n Qi − dF (t).
∗ s(0) (β0 , t) s(0) (β0 , t)
i=1 t
As for the first term on the right-hand side of the equation preceding the
one immediately above, we use Theorem II.5 of Xu (1996). With φ(t) =
1[t∗ ,1] (t)P (Z ∈ H|t) in the theorem, the first term in this equation is equal
to

n
√
n−1/2 νi + nRn ,
i=1
where |Rn | = op (n−1/2 )and ν’s are i.i.d. with mean zero, each being a function
of Xi and δi . Thus the proof is complete.
Chapter 9
Regression effect process
9.1 Chapter summary
In this chapter we describe the regression effect process. This can be established in
different ways and provides all of the essential information that we need in order to
gain an impression of departures from some null structure, the most common null
structure corresponding to an absence of regression effect. Departures in specific
directions enable us to make inferences on model assumptions and can suggest,
of themselves, richer more plausible models. The regression effect process, in its
basic form, is much like a scatterplot for linear regression telling us, before any
formal statistical analysis, whether the dependent variable really does seem to
depend on the explanatory variable as well as the nature, linear or more complex,
of that relationship. Our setting is semi-parametric and the information on the
time variable is summarized by its rank within the time observations. We make
use of a particular time transformation and see that a great body of known theory
becomes available to us immediately. An important objective is to glean what
we can from a graphical presentation of the regression effect process. The two
chapters following this one—building test statistics for particular and general
situations, and robust, effective model-building—lean heavily on the results of
this chapter.
O’Quigley (2003) developed a process based on the regression effect of a pro-

portional or a non-proportional hazards model. In that paper it was shown how
an appropriate transformation of the time scale allowed us to visualize this pro-
cess in relation to standard processes such as Brownian motion and the Brownian
bridge, two processes that arise as limiting results for the regression effect process

215
https://doi.org/10.1007/978-3-030-33439-0_9
216 Chapter 9. Regression effect process
Figure 9.1: The regression effect process as a function of time for simulated data
under the proportional hazards model for different values of β.
under certain conditions. O’Quigley (2003) and O’Quigley (2008) indicate how
we can move back to a regression effect process on the original scale if needed.
While conceptually the multivariate situation is not really a greater challenge than
the univariate case, the number of possibilities increases greatly. For instance,
two binary covariates give rise to several processes; two marginal processes, two
conditional processes, and a process based on the combined effect as expressed
through the prognostic index. In addition we can construct processes for stratified
models. All of these processes have known limiting distributions when all model
assumptions are met. However, under departures from working assumptions, we
are faced with the problem of deciding which processes provide the most insight
into the collection of regression effects, taken as a whole or looked at individually
(Figure 9.1).
In the presentation of this chapter we follow the outline given by O’Quigley
(2003) and O’Quigley (2008). The key results that enable us to appeal to the form
of Brownian motion and the Brownian bridge are proven in O’Quigley (2003),
O’Quigley (2008), Chauvel and O’Quigley (2014, 2017), and Chauvel (2014).
The processes of interest to us are based on cumulative quantities having the
flavor of weighted residuals, i.e., the discrepancy between observations and some
mean effect determined by the model. The model in turn can be seen as a
mechanism that provides a prediction and this prediction can then be indexed
by one or more unknown parameters. Different large sample results depend on
what values we choose to use for these parameters. They may be values fixed
under an assumption given by a null hypothesis or they may be some or all of the
values provided by estimates from our data. In this latter case the process will
tell us something about goodness of fit (O’Quigley, 2003). In the former case the
process will tell us about different regression functions that can be considered to
be compatible with the observations.
9.3. Elements of the regression effect process 217
9.3 Elements of the regression effect process
In the most basic case of a single binary covariate, the process developed below
in Section 9.4 differs from that of Wei (1984) in three very simple ways: the
sequential standardization of the cumulative process, the use of a transformed
time scale, and the direct interpolation that leads to the needed continuity. These
three features allow for a great deal of development, both on a theoretical and
practical level.
Taking the regression parameter to be β, then at time t, the process of Wei

is given by
n t

U (β, t) = Zi (s) − Eβ (Z | s) dNi (s), 0≤t≤T. (9.1)
i=1 0
The process corresponds to a sum which can be broken down into a series of
sequential increments, each increment being the discrepancy between the obser-
vation and the expected value of this observation, conditional upon time s and the
history up to that point. It is assumed that the model generates the observations.
Following the last usable observation (by usable we mean that the conditional
variance of the covariable in the risk set is strictly greater than zero), the result-
ing quantity is the same as the score obtained by the first derivative of the log
likelihood before time transformation. For this reason the process is also referred
to as the score process. In view of the immediate generalization of the score
process and the use we will make of this process we prefer to refer to it as the
regression effect process.
The point of view via empirical processes has been adopted by other authors.
Arjas (1988) developed a process that is similar to that considered by Wei.
Barlow and Prentice (1988) developed a class of residuals by conditioning in
such a way—not unlike our own approach—that the residuals can be treated as
martingales. The term martingale residuals is used. In essence any weight that
we can calculate at time t, using only information available immediately prior
to that time, can be used without compromising the martingale property. Such
quantities can be calculated for each individual and subsequently summed. The
result is a weighted process that contains the usual score process as a particular
case (Lin et al., 1993; Therneau and Grambsch, 2000). The constructions of
all of these processes take place within the context of the proportional hazards
model. Extending the ideas to non-proportional hazards does not appear to be
very easy and we do not investigate it outside of our own construction based on
sequential standardization and a transformation of the time scale. Wei (1984)
was interested in model fit and showed how a global, rather than a sequential
standardization, could still result in a large sample result based on a Brownian
bridge.
Sequential standardization and time transformation lead to great simplifica-

tion, to immediate extensions to the multivariate case, and to covariate depen-
dent censoring. The simplicity of the resulting graphics is also compelling and,
finally, the simple theorems enable us to use these results in practice with con-
fidence. Although more challenging in our view, other authors have had success
in extending the results of Wei (1984). Haara (1987) was able to find analo-
gous results in the case of non-binary covariables while Therneau et al. (1990)
considered a global standardization leading to a process that would look like a
Brownian bridge under the condition that the covariables were uncorrelated. This
would not typically be a realistic assumption. Given the very limited number of
situations, under global standardization, in which we can appeal to large sample
results looking like Brownian motion or the Brownian bridge, Lin et al. (1993)
adopted a different approach based on martingale simulation. This is more gen-
eral and while the resulting process has no form that can be anticipated we can
appeal to computer power to generate a large number of simulated processes
under the appropriate hypothesis. Tracing out all of these processes, over the
points of evaluation, provides us with envelopes that are indexed by their per-
cent coverage properties. If the observed process lies well within the envelope,
we may take that as evidence that any violation of the model assumptions is not
too great. “Well within” is not easily defined so that, again, it seems preferable
to have procedures based on more analytic results. Our own preference is for
sequential rather than global standardization and time transformation since, as
we see more fully below, this avoids having to deal with a lot of messy looking
figures. More importantly, it provides an immediate impression as to directions in
which to look if we have doubts about the underlying assumption of proportion-
ality of hazards. Lin et al. (1996) extended this envelope approach for situations
of covariate dependent censoring.
In order to give ourselves a foretaste as to where this is all pointing, consider a
simple example taken from the breast cancer study at the Curie Institute in Paris.
One covariable of interest was age at diagnosis categorized into two groups, the
hypothesis being that the younger group may harbor more aggressive tumors. The
results are illustrated in Figure 9.2. The leftmost figure shows the Kaplan-Meier
curves for the two groups with some suggestion that there may be a difference.
The middle figure gives the regression effect process in which the arrival point
of the process is close to the value −2.0, indicating a result that would be
considered to be borderline significant at 5%. Of equal importance is the form
of the regression effect process which, as we will show in this chapter, will look
like Brownian motion with linear drift under a non-null effect of a proportional
hazards nature. In Appendix B we see how this transforms to a Brownian bridge,
the 95% limits of the maximum which are shown for reference in the rightmost
figure. The bridge process lies well within those limits and shows no obvious sign
of departure from a Brownian bridge assumption allowing us to conclude that
there is no compelling reason to not accept a working assumption of proportional
Figure 9.2: Kaplan-Meier curves for the two age groups from the Curie Institute
breast cancer study. The regression effect process U ∗ (t) approximates that of
a linear drift while the transformation U ∗ (t) − tU ∗ (1) of the rightmost figure
appears well approximated by a Brownian bridge. This indicates a satisfactory fit.
hazards. All of this can be quantified and framed within more formal inferential
structures. The key point here is that before any formal inference takes place, the
eyeball factor—a quick glance at the Kaplan-Meier curves, the regression effect
process, and its transform—tells us a lot.
The empirical process we construct in the following sections allows us to
perform hypothesis tests on the value of the parameter and study the adequacy
of the non-proportional hazards model without having to estimate the regres-
sion coefficient. In addition, increments of the process are standardized at each
time point, rather than adopting an overall standardization for all increments as
in Wei (1984). This makes it possible to take into account multiple and cor-
related covariates because, unlike the non-standardized score process (9.1), the
correlation between covariates is not assumed to be constant over time. We plot
the process according to ranked times of failure, in the manner of Arjas (1988),
to obtain a simple and explicit asymptotic distribution. Before constructing the
process, we need to change the time scale. This transformation is crucial for
obtaining asymptotic results for the process without having to use the theorem
of Rebolledo (1980), nor the inequality of Lenglart (1977), which are classical
tools in survival analysis. In the following section, we present the time transfor-
mation used. In Section 9.4, we define the univariate regression effect process
and study its asymptotic behavior. In Section 9.5, we look at the multivariate
case. Section 9.8 gathers together the proofs of the theorems and lemmas that
are presented.
Transforming the time scale

In the proportional hazards model, a strictly monotonically increasing transforma-
tion in time does not affect inference on the regression coefficient β. Indeed, the
actual times of death and censored values are not used for inference; only their
order of occurrence counts. Among all possible monotonic increasing transforma-
tions we choose one of particular value. The order of observed times is preserved
and, by applying the inverse transformation, we can go back to the initial time
scale. Depending on how we carry out inference, there can be a minor, although
somewhat sticky, technical difficulty, in that the number of distinct, and usable,
failure times, depends on the data and is not known in advance. Typically we
treat this number as though it were known and, indeed, our advice is to do just
that in practice. This amounts to conditioning on the number of distinct and
usable failure times. We use the word “usable” to indicate that information on
regression effects is contained in the observations at these time points. In the
case of 2 groups for example, once one group has been extinguished, then, all
failure times greater than the greatest for this group carry no information on
regression effects.
Definition 9.1. With this in mind we define the effective sample size as
n

kn = δi × Δ(Xi ) where Δ(t) = 1Vβ(t) (Z|t)>0
i=1
If the variance is zero at a time of death, this means that the set of individuals
at risk is composed of a homogeneous population sharing the same value of the
covariate. Cleary the outcome provides no information on differential rates as
quantified by the regression parameter β. This translates mathematically as zero
contribution of these times of death to inference, since the value of the covariate
of the individual who dies is equal to its expectation.
We want to avoid such variances from a technical point of view, because
later we will want to normalize by the square root of the variance. For example,
in the case of a comparison of two treatment groups, a null conditional variance
corresponds to the situation in which one group is now empty but individuals from
the other group are still present in the study. Intuitively, we understand that these
times of death do not contribute to estimation of the regression coefficient β since
no information is available to compare the two groups. Note that the nullity of
a conditional variance at a given time implies nullity at later time points. Thus,
according to the definition of kn , the conditional variances Vβ(t) (Z|t) calculated
at the first kn times of death t are strictly greater than zero. For a continuous
covariate when the last time point observed is a death, we have the equality

(algebraic if conditioning, almost sure otherwise) that kn = n i=1 δi −1. In effect,
the conditional variance Vβ(t) (Z|t) is zero at the last observed time point t, and
almost surely is the only time at which the set of at-risk individuals shares the
same value of the covariate. If, for a continuous covariate, the last observed time

point is censored, then we have that kn = n i=1 δi .
Definition 9.2. Set φn (0) = 0 and consider the following transformation φn

of the observed times Xi , i = 1, ..., n:

1 # j : j = 1, . . . , n, Xj < Xi , N̄ (Xj ) = N̄ (Xi )
φn (Xi ) = N̄ (Xi ) + (1 − δi )
kn # j : j = 1, . . . , n, N̄ (Xj ) = N̄ (Xi )
We define the inverse φ−1

n of φn on [0, 1] by
φ−1
n (t) = inf {Xi , φn (Xi ) ≥ t, i = 1, . . . , n} , 0 ≤ t ≤ 1.
Recall that the counting process {N̄ (t)}t∈[0,T ] has a jump of size 1 at each
time of death. Thus, on the new time scale representing the image of the
observed times {X1 , . . . , Xn } under φn , the values in the set {1/kn , 2/kn , . . . , 1}
correspond to times of death, and the ith time of death ti is such that
ti = i/kn , i = 1, . . . , kn where t0 = 0. The set {1/kn , 2/kn , . . . , 1} is included
in but is not necessarily equal to the transformed set of times of death.
Inference based on transformed and original time scale

Inference for the regression coefficient is unaffected by monotonic increasing
transformations on the time scale and, in particular, the transformation of Def-
inition 9.2. The nature of effects will be very visible on the transformed time
scale and less so on the original scale. Nonetheless we can use the transform,
and its inverse, to move back and forward between scales. The transformation,
φn (Xi ), is not independent of the censoring and so some caution is needed in
making statements that we wish to hold regardless of the censoring. A regression
effect process with linear drift implies and is implied by proportional hazards.
Constancy of effect on one scale will then imply constancy of effect on the other.
We can say more. Suppose we have a changepoint model. The ratio of the early
regression coefficient to the later regression coefficient will be the same on either
scale.
Working with the first kn of the transformed failures, we will therefore restrict
ourselves to values less than or equal to 1. As the importance of the transforma-
tion applied to the observed times is to preserve order, various transformations—
treating censoring times differently—could be used. We choose to uniformly dis-
tribute censoring times between adjacent times of death. The time τ0 on this
new scale corresponds to the (100 × τ0 )th percentile for deaths in the sample.
For example, at time τ0 = 0.5, half of the deaths have been observed. At time
τ0 = 0.2, 20% of the deaths have occurred. The inverse of φn can be obtained
easily and with it we can interpret the results on the initial time scale.
Conditional means and variances

Our next task is that of sequential standardization. We now work with the stan-
dardized time scale in [0, 1]. With this time scale, for t ∈ [0, 1], we define the haz-
ard index Yi∗ (t) and individual counting process Ni∗ (t) for individuals i = 1, . . . , n
by
Yi∗ (t) = 1φ (X ) ≥ t , Ni∗ (t) = 1φ (X ) ≤ t, δ = 1 .

n i n i i
Each time of death ti is a (Ft∗ )t∈[0,1] -stopping time where, for t ∈ [0, 1], the
σ-algebra Ft∗ is defined by

Ft∗ = σ Ni∗ (u), Yi∗ (u+ ), Zi ; i = 1, . . . , n ; u = 0, 1, . . . , tkn ,
where · is the floor function and Yi∗ (t+ ) = lim Yi∗ (s). Notice that if 0 ≤ s <
s→t+
t ≤ 1, Fs∗ ⊂ Ft∗ .
Remark. Denote a ∧ b = min(a, b), for a, b ∈ R. If the covariates are time-
dependent, for t ∈ [0, 1], the σ-algebra Ft∗ is defined by

u +
Ft∗ =σ Ni∗ (u), Yi∗ (u+ ), Zi φ−1
n ∧ Xi ; i = 1, . . . , n ; u = 0, 1, . . . , tkn
kn
since, as we recall, the covariate Zi (·) is not observed after time −1Xi , i =
1, . . . , n. To simplify notation in the following, we will write Zi φn (t) in
+
the place of Zi φ−1
n (t) ∧ Xi , , t ∈ [0, 1]. We define the counting process
associated with transformed times, with jumps of size 1 at times of death in
the new scale, by
n

N̄ ∗ (t) = 1φ (X ) ≤ t, δ = 1 , 0 ≤ t ≤ 1.
n i i
i=1
This counting process satisfies the following result:

Proposition 9.1. (Chauvel 2014). ∀ t ∈ [0, 1], N̄ ∗ (t) = kn t.
In other words, t0 = 0 and N̄ ∗ (t) has jumps at times tj = j/kn for j =
1, . . . , kn . This leads to the following lemma, which will be of use in proving
Theorem 9.2:
Lemma 9.1. (Chauvel and O’Quigley 2014). Let t0 = 0, n ∈ N and let
{An (t), t ∈ {t1 , . . . , tkn }} be a process with values in R. Suppose that
P
sup |An (ti ) − a(ti )| −→ 0,
i=1,...,kn n→∞
where a is a function defined and bounded on [0, 1]. Then,

s s
1 P
sup ∗
An (t)dN̄ (t) − a(t)dt −→ 0.
s∈[0,1] kn 0 0 n→∞
The proofs are given at the end of the chapter. To ease notation, let us define
a process Z which is 0 everywhere except at times of death. At each such time,
the process takes the value of the covariate of the individual who fails at that
time.
Definition 9.3. We denote Z = {Z(t), t ∈ [0, 1]} a process such that
n

Z(t) = Zi (Xi ) 1φ (X ) = t, δ = 1 , t ∈ [0, 1]. (9.2)
n i i
i=1
The family of probabilities {πi (β(t), t), i = 1, . . . , n}, with t ∈ [0, T ] can be
extended to a non-proportional hazards model on the transformed scale as fol-
lows.
Definition 9.4. For i = 1, . . . , n and t ∈ [0, 1], the probability that individual
i dies at time t under the non-proportional hazards model with parameter
β(t), conditional on the set of individuals at risk at time of death t, is defined
by

Yi∗ (t) exp β(t)T Zi φ−1 n (t) 1kn t∈N
πi (β(t), t) = n ∗
−1
.
T
j=1 Yj (t) exp β(t) Zj φn (t)
Remark. The only values at which these quantities will be calculated are the
times of death t1 , . . . , tkn . We nevertheless define Z(t) and {πi (β(t), t), i =
1, . . . , n} for all t ∈ [0, 1] in order to be able to write their respective sums
as integrals with respect to the counting process N̄ ∗ . The values 0 taken
by them at times other than that of death are chosen arbitrarily as long
as they remain bounded by adjacent failures. Note also that at t0 = 0, we
have Z(0) = 0. Expectations and variances with respect to the family of
probabilities {πi (β(t), t), i = 1, . . . , n} can now be defined.
Definition 9.5. Let t ∈ [0, 1]. The expectation and variance of Z with respect
to the family of probabilities {πi (β(t), t), i = 1, . . . , n} are respectively a vector
in Rp and a matrix in Mp×p (R) such that
n

Eβ(t) (Z | t) = Zi φ−1
n (t) πi (β(t), t), (9.3)
i=1
and
n
⊗2
Vβ(t) (Z | t) = Zi φ−1
n (t) πi (β(t), t) − Eβ(t) (Z | t)⊗2 . (9.4)
i=1
∂
It can be shown that the Jacobian matrix of Eβ(t) (Z | t) is E (Z | t) =
∂β β(t)
Vβ(t) (Z | t), for t ∈ [0, 1].
Proposition 9.2. Let t ∈ [0, 1]. Under the non-proportional hazards model
with parameter β(t), the expectation Eβ(t) (Z(t) | Ft∗− ) and variance
Vβ(t) (Z(t) | Ft∗− ) of the random variable Z(t) given the σ-algebra Ft∗− are
Eβ(t) (Z(t) | Ft∗− ) = Eβ(t) (Z | t) , Vβ(t) (Z(t) | Ft∗− ) = Vβ(t) (Z | t).
In conclusion, for j = 1, . . . , kn and time of death tj , by conditioning on the set

of at-risk individuals, the expectation of Z(tj ) is Eβ(tj ) (Z | tj ) and its variance-
covariance matrix is Vβ(tj ) (Z | tj ). Knowledge of these moments allows us to
define the regression effect process, which we will study separately for the univari-
ate (p = 1) and multivariate (p > 1) cases. It is worth adding that the univariate
process has an interest well beyond the case of simple models involving only a
single covariate. For multivariate models there are several univariate processes
that will reveal much of what is taking place. These include the process for the
prognostic index, processes for individual covariates after having adjusted via the
model for the effects of some or all of the other covariates, processes for stratified
models, and indeed any process that can be derived from the model.
9.4 Univariate regression effect process
In this section, we will focus more closely on the univariate framework (p =

1), i.e., when there is only one covariate Z(t) in the model. In some sense
higher dimensions represent an immediate generalization of this but, for higher
dimensions, in view of the complexity theorem (Theorem 4.1) there are many
more things to consider, both conceptually and operationally. In order to gain the
firmest grip possible on this we look closely at the univariate case which contains
all of the key ideas. It is always a good principle in statistics to condition on as
many things as we can. We can go too far and condition on aspects of the data
that are not uninformative with regard to our unknown regression parameter.
This would certainly lead to error so care is always called for. In some ways
conditioning is an art, knowing just how far to go. One thing that can guide us
is to what extent, after such conditioning, we can recover procedures that have
been established under quite different principles. Here, we will see for instance
that a lot of conditioning will result in the well-known log-rank test which, at
the very least, gives us some confidence.
Sequential conditioning through time, i.e., conditioning on the times of
death and the sets of individuals at risk, as first proposed by Cox (1972),
leads to a very natural framework for inference. An individual i who dies at
tj is viewed as having been selected randomly (under the model) from the set
9.4. Univariate regression effect process 225
{i : Yi∗ (tj ) = 1; i = 1, . . . , n} of individuals at risk of death at time tj with

probability πi (β(tj ), tj ). Under this conditioning, at the time of death tj , the
information necessary to calculate the expectation and variance of Z(tj ) with
respect to {πi (β(tj ), tj ); i = 1, . . . , n} is available for j = 1, . . . , kn . The condi-
tional variances Vβ(t) (Z | t) calculated for the first kn times of death t are strictly
greater than zero. Thus,
∀ t ∈ {t1 , . . . , tkn }, Vβ(t) (Z | t) > 0 a.s.
The variables Z(tj ) may be standardized. By sequentially summing all of these

standardized variables, we obtain a standardized score which is the basis of our
regression effect process. A further conditioning argument whereby the number
of discrete failure times at which the variance Vβ(t) (Z | t) is strictly greater than
zero leads, once more, to great simplification. For completeness the unconditioned
case is worked through at the end of the chapter. As argued above, the most
useful approach is a conditional one in which we take kn to be equal to its
observed value.
Definition 9.6. Let j = 0, 1, . . . , kn . The regression effect process evaluated

for the parameter β(tj ) at time tj is defined by Un∗ (·, ·, 0) = 0 and
tj
1
Un∗ α(tj ), β(tj ), tj =√ Vα(s) (Z | s)−1/2 Z(s) − Eβ(s) (Z | s) dN̄ ∗ (s).
kn 0
where, by using two parameter functions, α(t) and β(t), we allow ourselves
increased flexibility. The α(t) concerns the variance and allowing this to not be
tied to β(t) may be of help in some cases where we would like estimates of
the variance to be valid both under the null and the alternative. This would be
similar to what takes place in an analysis of variance where the residual variance
is estimated in such a way as to remain consistent both under the null and the
alternative. Our study on this, in this particular context, is very limited, and this
could be something of an open problem. For the remainder of this text we will
suppose that α(t) = β(t) and we will consequently write Un∗ (β(tj ), tj ) as only
having two arguments.
Remark. For computational purposes, and for j = 1, . . . , kn , the variable

Un∗ (β(tj ), tj ) can also be written as the sum:
j
1
Un∗ (β(tj ), tj ) = √ Vβ(ti ) (Z | ti )−1/2 Z(ti ) − Eβ(ti ) (Z | ti ) , (9.5)
kn i=1
noting that all of the ingredients in the above expression are obtained rou-
tinely from all of the currently available software packages.
Remark. Let w be a deterministic or random (Ft∗ )t∈[0,1] -predictable function

from [0, 1] to R. According to Proposition 9.2, under the non-proportional
hazards model with parameter β(t), conditional on the set of individuals at
risk at time of death tj , the expectation of w(tj )Z(tj ) is w(tj )Eβ(tj ) (Z | tj )
and its variance w(tj )2 Vβ(tj ) (Z | tj ) with respect to the family of probabilities
{πi (β(tj ), tj ); i = 1, . . . , n}. By summing the standardized weighted variables,
we obtain the same definition of the process as in Definition 9.6. We will come
back to this remark in Section 11.4 for the weighted log-rank test.
We can now define the process Un∗ on [0, 1] by linearly interpolating the
kn + 1 random variables {Un∗ (β(tj ), tj ), j = 0, 1, . . . , kn }.
Definition 9.7. The standardized score process evaluated at β(t) is defined
by {Un∗ (β(t), t), t ∈ [0, 1]}, where for j = 0, . . . , kn and t ∈ [tj , tj+1 [,
Un∗ (β(t), t) = Un∗ (β(tj ), tj ) + (tkn − j) {Un∗ (β(tj+1 ), tj+1 ) − Un∗ (β(tj ), tj )} .
By definition, the process Un∗ (β(·), ·) is continuous on the interval [0, 1]. The pro-
cess depends on two parameters: the time t and regression coefficient β(t). For
the sake of clarity, we recall that the temporal function β : t → β(t) is denoted
β(t), which will make it easier to distinguish the proportional hazards model
with parameter β from the non-proportional hazards model with parameter β(t).
Under the non-proportional hazards model with β(t), the increments of the pro-
cess Un∗ are centered with variance 1. Moreover, the increments are uncorrelated,
as shown in the following proposition.
Proposition 9.3. (Cox 1975). Let Dn∗ (β(tj ), tj ) = Un∗ (β(tj ), tj ) − Un∗ (β(tj−1 ),
tj−1 ). Under the non-proportional hazards model with parameter β(t), the ran-
dom variables Dn∗ (β(tj ), tj ) for j = 1, 2, . . . , kn are uncorrelated.
The uncorrelated property of the increments is also used in the calculation of
log-rank statistic. All of these properties together make it possible to obtain our
essential convergence results for the process Un∗ (β(t), t).
Working assumptions for large sample theory

The space (D[0, 1], R) comes equipped with uniform convergence. For 0 ≤ t ≤ 1
and for r = 0, 1, 2, we can define:
n
1 ∗ r −1
S (r) {β(t), t} = Yi (t)Zi φ−1
n (t) exp β(t)Zi φn (t) .
n
i=1
Note that for t ∈ {t1 , t2 , . . . , tkn },
∂ S (1) (β(t), t)

Eβ(t) (Z | t) = log S (b, t)
(0)
= (0) ,
∂b b=β(t) S (β(t), t)
and
2
∂2 S (2) (β(t), t) S (1) (β(t), t)
Vβ(t) (Z | t) = 2 log S (0) (b, t) = (0) − .
∂b b=β(t) S (β(t), t) S (0) (β(t), t)
Suppose now the following conditions:
A1
(Asymptotic stability). There exists
some δ1 > 0, a neighborhood B =
γ, supt∈[0,1] |γ(t) − β(t)| < δ1 of β of radius δ1 containing the zero func-
tion, and functions s(r) defined on B × [0, 1] for r = 0, 1, 2, such that
√
P
n sup S (r) (γ(t), t) − s(r) (γ(t), t) −→ 0. (9.6)
t∈[0,1],γ∈B n→∞
A2 (Asymptotic regularity). The deterministic functions s(r) , r = 0, 1, 2,

defined in A1 are uniformly continuous for t ∈ [0, 1] and bounded in
B × [0, 1]. Furthermore, for r = 0, 1, 2, and t ∈ [0, 1], s(r) (·, t) is a con-
tinuous function in B. The function s(0) is bounded below by a strictly
positive constant.
By analogy with the empirical quantities, for t ∈ [0, 1] and γ ∈ B, denote
s(1) (γ(t), t) s(2) (γ(t), t) 2

e(γ(t), t) = , v(γ(t), t) = − e(γ(t), t) . (9.7)
s(0) (γ(t), t) (0)
s (γ(t), t)
∂
A3 (Homoscedasticity). For any t ∈ [0, 1] and γ ∈ B, we have v(γ(t), t) = 0.
∂t
A4 (Uniformly bounded covariates). There exists L ∈ R∗+ such that
sup sup |Zi (u)| ≤ L.

i=1,...,n u∈[0,T ]
The first two conditions are standard in survival analysis. They were introduced by
Andersen and Gill (1982) in order to use theory from counting processes and mar-
tingales such as the inequality of Lenglart (1977) and the theorem of Rebolledo
(1980). The variance Vβ(t) (Z | t) is, by definition, an estimator of the variance
of Z given T = t under the non-proportional hazards model with parameter β(t).
Thus, the homoscedasticity condition A3 means that the asymptotic variance is
independent of time for parameters close to the true regression coefficient β(t).
This condition is often implicitly encountered in the use of the proportional
hazards model. This is the case, for example, in the estimation of the variance
of the parameter β, or in the expression for the log-rank statistic, where the
contribution to the overall variance is the same at each time of death. Indeed,
the overall variance is an unweighted sum of the conditional variances. The
temporal stability of the variance has been pointed out by several authors, notably
Grambsch and Therneau (1994), Xu (1996), and Xu and O’Quigley (2000). Next,
we need two lemmas. Proofs are given at the end of the chapter.
Lemma 9.2. Under conditions A1 and A3, for all t ∈ [0, 1] and γ ∈ B,
v(γ(t), t) > 0.
Lemma 9.3. Under hypotheses, A1, A2, and A3, and if kn = kn (β0 ), for all
γ1 ∈ B, there exist constants C(γ1 ) et C(β0 ) ∈ R∗+ such that
√ P
n sup Vγ (t) (Z | t) − C(γ1 ) −→ 0, (9.8)
1 n→∞
t∈{t1 ,t2 ,...,tkn }

√ Vγ1 (t) (Z | t) C(γ1 ) P
n sup − −→ 0, (9.9)
V (Z | t) C(β0 ) n→∞
t∈{t1 ,t2 ,...,tkn } β 0

V C(γ1 )
√ γ1 (t) (Z | t) P
n sup − −→ 0. (9.10)
t∈{t1 ,t2 ,...,tkn } Vβ0 (Z | t) C(β0 )
1/2 1/2 n→∞
Theorem 9.1. Under the non-proportional hazards model with parameter

β(t),
L
Un∗ (β(·), ·) −→ W(t),
n→+∞
where W(t) is a standard Brownian motion.
The proof, given at the end of the chapter, is built on the use of the functional
central limit theorem for martingale differences introduced by Helland (1982).
This theorem can be seen as an extension of the functional central limit theo-
rem of Donsker (1951) to standardized though not necessarily independent nor
identically distributed variables. Figure 9.3(a) shows the result of simulating a
standardized score process Un∗ (0, t) as a function of time t, when the true param-
eter is constant. The immediate visual impression would lead us to see Brownian
motion—standard or with linear drift—as a good approximation. We will study
how to calculate this approximation on a given sample in Chapters 10 and 11.
Theorem 9.2. (Chauvel and O’Quigley 2014). Let β0 be a constant func-

tion on B. Under the non-proportional hazards model with parameter β(t),
consider
Vβ̃(s) (Z | s)
1 t
An (t) = {β(s) − β0 } dN̄ ∗ (s), 0 ≤ t ≤ 1, (9.11)
kn 0 Vβ0 (Z | s)1/2
where β̃ is in the ball with center β and radius supt∈[0,1] |β(t) − β0 |. Then,
there exist two constants C1 (β, β0 ) and C2 ∈ R∗+ such that C1 (β, β)=1 and
L
Un∗ (β0 , ·) − kn An −→ C1 (β, β0 )W(t). (9.12)
n→∞
Figure 9.3: The process Un∗ (0, ·) as a function of time for simulated data under
the proportional hazards model. For left-hand figure, β = 1.0. For right-hand
figure, the absence of effects manifests itself as approximate standard Brownian
motion.
Furthermore,
t
P
sup An (t) − C2 {β(s) − β0 } ds −→ 0. (9.13)
0≤t≤1 0 n→∞
Remark. Theorem 9.1 is a special case of Theorem 9.2 when β(t) = β0 . In

effect, in Theorem 9.2, when β(t) = β0 , the constant C1 (β, β) is equal to 1,
the drift is zero, and the limit process of Un∗ (β0 , ·) is a standard Brownian
motion. The proof of the theorem can be found in Section 9.8. The drift
of the standardized score process at time t tends to infinity when kn →
∞, according to equation (9.12). In practice, we work with a large enough
fixed kn and β0 = 0. The theorem indicates that the drift of the process
Un∗ (0, ·) is proportional to β(s). We will distinguish between two cases in
which the regression coefficient β(t) is non-zero. In the first, consider the
proportional hazards model, that is, β(t) is constant over time. From Theorem
9.2, the process Un∗ (0, ·) can be approximated by a Brownian motion with
a linear drift. This is illustrated in Figure 9.3(a) with a process that has
been simulated under the proportional hazards model with non-zero β. The
drift can be approximated by a straight line, indicating that the regression
coefficient is constant in time. The coefficient β, which has the same sign as
the slope of the line according to Theorem 9.2, is positive.
The second case corresponds to the situation where there is a non-constant
regression coefficient over time. Several scenarios are possible. For example, the
effect could be constant for an initial period and then quickly decline beyond
Figure 9.4: The process Un∗ (0, ·) as a function of time for simulated data under the
non-proportional hazards model with different values of β(t). Temporal effects
are clearly reflected in the process on the transformed time scale.
some point τ , piecewise constant over time, increase or decrease continuously,
and so on. In all these situations, the cumulative effect β(s)ds will be seen in
the drift of the process (Theorem 9.2) and will, of itself, describe to us the nature
of the true effects. For example, Figure 9.4(a) shows a simulated process with a
piecewise-constant effect β(t) = 1t≤0.5 . Before t = 0.5, we see a positive linear
trend corresponding to β = 1, and for t > 0.5, the coefficient becomes zero and
the drift of the process Un∗ (0, ·) is parallel to the time axis. Figure 9.4(b) shows
the standardized score process for simulated data under failure time model with
regression function, β(t) = 1t≤1/3 + 0.5 1t≥2/3 . The drift of the process can be
separated into three linear sections reflecting the effect’s strength, each of which
can be approximated by a straight line. The slope of the first line appears to be
twice that of the last, while the slope of the second line is zero.
In summary, whether the effect corresponds to proportional or non-proportional
hazards, all information about the regression coefficient β(t) can be found in the
process Un∗ (0, ·). We can then use this process to test the value of the regres-
sion coefficient and evaluate the adequacy of the proportional hazards model. A
simple glance at the regression effect process is often all that is needed to let
us know that there exists an effect and the nature of that effect. More precise
inference can then follow. Before delving into these aspects in Chapters 10 and
11, we consider more closely the standardized score process for the multivariate
case. As mentioned earlier the univariate process, even in the multivariate setting,
will enable us to answer most of the relevant questions, such as the behavior of
many variables combined in a prognostic index, or, for example, the impact of
some variable after having controlled for the effects of other variables, whether
via the model or via stratification.
9.5. Regression effect processes for several covariates 231
9.5 Regression effect processes for several covariates
Here we extend the regression effect process of the previous section to the case
of several covariates grouped in a vector Z(t) of dimension p > 1. Suppose that
we have the non-proportional hazards model with regression function β(t) and a
vector of covariates Z(t) which are functions from [0, 1] to Rp . Following Chauvel
(2014), let t ∈ [0, 1] and
n

S (0) (β(t), t) = n−1 Yi∗ (t) exp β(t)T Zi (φ−1
n (t)) ∈ R,
i=1
n

S (1) (β(t), t) = n−1 Yi∗ (t)Zi (φ−1 T −1 p
n (t)) exp β(t) Zi (φn (t)) ∈ R ,
i=1
n

S (2) (β(t), t) = n−1 Yi∗ (t)Zi (φ−1
n (t))
⊗2
exp β(t)T Zi (φ−1
n (t)) ∈ Mp×p (R).
i=1
For t ∈ {t1 , . . . , tkn }, the conditional expectation Eβ(t) (Z | t) and the conditional
variance-covariance matrix Vβ(t) (Z | t) defined in (9.3) and (9.4) can be written
as functions of S (r) (β(t), t) (r = 0, 1, 2):
S (1) (β(t), t) S (2) (β(t), t)

Eβ(t) (Z | t) = , Vβ(t) (Z | t) = − Eβ(t) (Z | t)⊗2 .
S (0) (β(t), t) S (0) (β(t), t)
which is the same as ∂Eβ(t) (Z | t)/∂β. As the matrix Vβ(t) (Z | t) is symmetric

and positive definite, there exists an orthogonal matrix Pβ(t) (t) and a diagonal
matrix Dβ(t) (t) such that
Vβ(t) (Z | t) = Pβ(t) (t)Dβ(t) (t)Pβ(t) (t)T .
This leads us to define the symmetric matrix Vβ(t) (Z | t)−1/2 as
Vβ(t) (Z | t)−1/2 = Pβ(t) (t)Dβ(t) (t)−1/2 Pβ(t) (t)T .
By the definition of k
n = kn (β) in the multivariate case, and since kn ≤ kn

almost surely (Equation (9.23)), the conditional variance-covariance matrices
Vβ(t) (Z | t) calculated for the first kn times of death t are positive definite, and

∀ t ∈ {t1 , . . . , tkn }, Vβ(t) (Z | t)−1 ≤ CV−1 a.s.
Definition 9.8. Let j = 0, 1, . . . , kn . The multivariate standardized score calcu-

lated at time tj for parameter β(tj ) is defined by
tj
1
Un∗ (β(tj ), tj ) = √ Vβ(s) (Z | s)−1/2 Z(s) − Eβ(s) (Z | s) dN̄ ∗ (s),
kn 0
where
tj tj tj
a(s)dN̄ ∗ (s) = a1 (s)dN̄ ∗ (s), . . . , ap (s)dN̄ ∗ (s) ,
0 0 0
for a = (a1 , . . . , ap ), where the ai are functions from [0, 1] to R.
Remark. As in the univariate case, the random variable Un∗ (β(tj ), tj ) can be
written as a sum:
j
1
Un∗ (β(tj ), tj ) = √ Vβ(ti ) (Z | ti )−1/2 Z(ti ) − Eβ(ti ) (Z | ti ) .
kn i=1
The standardized score is a random function defined at times t1 , . . . , tkn tak-

ing values in Rp , whose definition can be extended to [0, 1] by linear interpolation.
Definition 9.9. The multivariate standardized score process {Un∗ (β(t), t),
t ∈ [0, 1]} evaluated at β(t) is such that, for j = 0, 1, . . . , kn and t ∈ [tj , tj+1 [,
Un∗ (β(t), t) = Un∗ (β(tj ), tj ) + (tkn − j) {Un∗ (β(tj+1 ), tj+1 ) − Un∗ (β(tj ), tj )} .
Remark. We can readily construct several univariate processes on the basis

of the multivariate model. Suppose for example that we have 2 binary covari-
ates Z1 and Z2 . If we stratify on the basis of Z1 then we would have a single
regression effect processes for Z2 that can be studied, as in the simple case.
The same is true if we stratify on the basis of Z2 and study Z1 . If, instead
of stratifying, we fit both covariables into the model, then, we can consider
processes for each variable separately. We could also consider conditional pro-
cesses, for example where we look at Z2 and fix Z1 , leading to two conditional
processes. And, of course, the prognostic index would define a joint process
where both variables are included in a linear combination. Definition 9.9 is
interesting because it allows us to take into account the degree of association
between the covariates themselves as well as how they relate, individually
and collectively, to survival outcome.
Working assumptions in the multivariate case

Recall that the norms of vectors in Rp and matrices in Mp×p (R) are their sup
norms, as defined in Appendix A. The function space (D[0, 1], Rp ) is equipped
with the Skorokhod product topology. To establish the limit behavior of the pro-
cess, we refer to Chauvel and O’Quigley (2014, 2017), and suppose the following
conditions to hold:
9.5. Regression effect processes for several covariates 233
B1 (Asymptotic stability). There exists δ2 >0, abounded neighborhood B of β of

radius δ2 containing the zero function: B = γ, supt∈[0,1] γ(t) − β(t) < δ2 ,
and functions s(r) defined on B ×[0, 1] for r = 0, 1, and 2, respectively, taking
values in R, Rp , and the matrix space Mp×p (R), such that
√
(r) P
n sup S (γ(t), t) − s(r) (γ(t), t) −→ 0.
t∈[0,1]},γ∈B n→∞
B2 (Asymptotic regularity). The deterministic functions s(r) , r = 0, 1, 2, defined

in B1 are uniformly continuous for t ∈ [0, 1] and bounded in B × [0, 1]. Fur-
thermore, for r = 0, 1, 2 and t ∈ [0, 1], s(r) (·, t) is a continuous function in
B . The function s(0) is bounded below by a strictly positive constant.
B3 (Homoscedasticity). There exists a symmetric and positive definite matrix

Σ ∈ Mp×p (R) and a sequence of positive constants (Mn )n such that
limn→∞ Mn = 0 which satisfy for all γ ∈ B , t ∈ [0, 1] and i, j = 1, . . . , p,

∂2

Eβ (Z | t) ≤ Mn a.s., (9.14)
∂βi ∂βj β=γ(t)

√
P
n sup Vγ(t) (Z | t)1/2 − Σ1/2 −→ 0. (9.15)
t∈{t1 ,t2 ,...,tkn },γ∈B n→∞
B4 (Uniformly bounded covariates). There exists some L ∈ R∗+ such that

(j)
sup sup sup Zi (u) ≤ L .
i=1,...,n u∈[0,T ] j=1,...,p
These conditions are extensions of hypotheses A1, A2, A3, and A4 from Section
9.4 to the multiple covariates case. We make the additional assumption that the
conditional variance-covariance matrices do not depend on the parameter when it
is close to the true regression parameter. This is a technical assumption required
due to the use of the multidimensional Taylor-Lagrange inequality in the proofs.
In the univariate case, we do not need this hypothesis thanks to the availability
of a Taylor-Lagrange equality.
Proposition 9.4. (Chauvel 2014). Under hypotheses B1, B3, and B4, we have
the following:
√ P
n sup Vγ(t) (Z | t) − Σ −→ 0. (9.16)
t∈{t1 ,t2 ,...,tkn },γ∈B n→∞
The proposition is proved at the end of the chapter. The following theorem gives
the limit behavior of the univariate regression effect process. Let a = (a1 , . . . , ap )
be a function from Rp to [0, 1]. Denote
t t t
a(s)ds = a1 (s)ds, . . . , ap (s)ds , 0 ≤ t ≤ 1.
0 0 0
Suppose γ ∈ B . Let Un∗ (γ, ·) be the multivariate regression effect process cal-
culated at γ. To construct this process, we consider kn = kn (γ), which can be
estimated by k
n (γ). Recall that kn (γ) ≤ kn (γ) almost surely (Equation 9.23)

and by the definition of kn (γ) in the multivariate case, we have:

∀ t ∈ {t1 , . . . , tkn }, Vγ(t) (Z | t)−1 ≤ CV−1 a.s. (9.17)
Theorem 9.3. (Chauvel and O’Quigley 2014). Let β0 be a constant func-

tion on B . Under the non-proportional hazards model with regression coeffi-
cient β(t), if the multivariate regression effect process is calculated at param-
eter β0 , we have the convergence in distribution result:
L
Un∗ (β0 , ·) − kn Bn −→ Wp (t), (9.18)
n→∞
where Wp (t) is a standard p-dimensional Brownian motion, and for each

t ∈ [0, 1],
tkn
1
Bn (t) = Vβ0 (Z | ti )−1/2 Eβ(ti ) (Z | ti ) − Eβ0 (Z | ti ) .
kn
i=1
Furthermore,
t
P

sup Bn (t) − Σ1/2
{β(s) − β0 } ds −→ 0. (9.19)
n→∞
t∈[0,1] 0
We recall that a p-dimensional Brownian motion is a vector of p independent

Brownian motions, each in R.
Remark. The constant C1 (β, β0 ) in Theorem 9.2 in which the convergence

of the univariate process Un∗ is established has no multivariate equivalent.
This is a consequence of Equation 9.14 in hypothesis B3, which is used in the
proof in a multidimensional Taylor-Lagrange inequality, but is not needed
in the univariate case. The constant C2 in Theorem 9.2 is equivalent to the
matrix Σ1/2 . In practice, for a large enough fixed value of n, the multivariate
standardized score process Un∗ (0, ·) can thus be approximated by a Brownian
motion whose drift depends on the true parameter β(t) of the model. Each
9.6. Iterated logarithm for effective sample size 235
√ t
component of the drift kn Σ1/2 0 β(s)ds represents a linear combination of
t t
the cumulative effects 0 β1 (s)ds, . . . , 0 βp (s)ds, except when the variance-
covariance matrix Σ is diagonal. Under the theorem’s hypotheses, a direct
consequence of Equation 9.16 is
k
1
n
P
Σ̂ := Vβ0 (Z | ti ) −→ Σ, (9.20)
kn n→∞
i=1
which means that Σ̂ is an estimator that converges to Σ. The following

corollary allows us to decorrelate the components of the process Un∗ so
that the drift of each component represents the corresponding effect (in
β1 (t), . . . , βp (t)) separately.
Corollary 9.1. (Chauvel 2014). Under the conditions of Theorem 9.3,
L
−1/2 U ∗ (β0 , ·) −
Σ −1/2 Bn
kn Σ −→ Σ−1/2 Wp ,
n n→∞
and t
−1/2 P

sup Σ Bn (t) − {β(s) − β0 } ds −→ 0.
n→∞
t∈[0,1] 0
Thus, when there are several covariates, the drifts of the process
−1/2 U ∗ (β0 , ·) help us to make inference on the regression coefficients. Each
Σ n
drift represents a cumulative coefficient and inference is done separately for each
drift, as in the univariate case.
9.6 Iterated logarithm for effective sample size
This section is focused on a technical detail first studied by Chauvel (2014).

The issue is thoroughly resolved if we are prepared to condition on the observed
effective sample size. That would be our recommendation. If we do not wish to
make use of such a simplifying argument, and we wish to appeal to results that
are available in the framework of the martingale central limit theorem, then we
need to pay some small amount of attention to a technical detail concerning
effective sample size. The martingale central limit theorem requires us to make
use of the concept of predictability. Our useable number of failure points, kn ,
is, technically speaking, not predictable since, until the study is finished or until
such a time as the variance becomes zero, we do not know what its value is. Our
solution to this is simple, just take the value as fixed and known at its observed
value. This amounts to conditioning on a future value, just like obtaining the
Brownian bridge (not a martingale) from Brownian motion (a martingale) by
conditioning on its final value. But, this very simple solution, while fine in our
view, does take the inference away from under the umbrella of the martingale
central limit theorem. A general solution is not hard to find but requires a little
attention to some finicky details. We consider this more closely here. Let us
suppose that there exists α0 ∈]0, 1] such that
k
n (β) a.s.
−→ α0 .
n n→∞
Furthermore, suppose that there exists a sequence (an )n converging to 1 and

∈ N∗ such that for each n ≥ N
N ,
k
n (β)
an ≥ 1, an ≥ α0 a.s. (9.21)
n
This would imply in particular that k n (β) tends to infinity when n tends to
infinity, i.e., the number of deaths increases without bound as the sample size
increases. If Z is a discrete variable, this means that for a sample of infinite size,
n
no group will run out of individuals. If Z is continuous, k n (β) = i=1 δi − 1 or
n −1
kn (β) = i=1 δi almost surely. If so, n kn (β) is an estimator that converges
to P (T ≤ C) and α0 = P (T ≤ C). The law of the iterated logarithm then allows
us to construct a sequence (an ) satisfying Equation 9.21.
Remark. Suppose we have one continuous covariate n withcensoring at the

last observed time point, implying that k n (β) = δ
i=1 i = n
1
i=1 Xi ≤Ci . The
variables δ1 , . . . , δn are independent and identically distributed with finite
second order moments. Their expectation is α0 = P (T ≤ C) and variance
α0 (1 − α0 ). Using the law of the iterated logarithm, we have:
√
n n−1 k n (β) − α0
lim inf = −1 a.s.
n→∞ 2α0 (1 − α0 ) log log(n)
Thus,
√
n n−1 k n (β) − α0
∀ ε > 0, ∃ Nε ∈ N∗ , ∀ n ≥ Nε , ≥ −1 − ε a.s.
2α0 (1 − α0 ) log log(n)
Setting ε = 1/2, ∃ N1/2 ∈ N∗ , ∀ n ≥ N1/2 ,

√
n n−1 k n (β) − α0 3 log log(n) 1 2α0
≥− a.s. , ≤ . (9.22)
2α0 (1 − α0 ) log log(n) 2 n 3 1 − α0
Suppose that n ≥ N1/2 . Then, with the help of Equation 9.22,
k
n (β) 3
≥ α0 − √ α0 (1 − α0 ) log log(n),
n 2n
and by setting
−1
3
an = 1 − √ (1 − α0 ) log log(n) ,
2α0 n
we find n−1 k
n (β)an ≥ α0 . From Equation 9.22, an ≥ 1 and we therefore have
limn→∞ an = 1. We have thus constructed a sequence (an ) which satisfies
Equation 9.21.
The random variable k n (β) is known only at the end of data collection. It
cannot be used during the experiment to construct a predictable process (in the
sense of an increasing family of σ-algebras as described below). To overcome
this we define the deterministic quantity kn (β) = na−1n α0 , in order to obtain
kn (β)
theoretical results. We have the almost sure convergence −→ 1, and
kn (β) n→∞
Equation 9.21 implies that
,
∀n ≥ N k
n (β) ≥ kn (β) a.s. (9.23)

1. Simulate 30 i.i.d. uniform variates, Xi , i = 1, ..., 30. Calculate Si = ij=1 Xj .
Join, using straight lines, the points Si and Si+1 . We call this a process.
Make a graphical plot of the process based on Si versus i, for i = 1 through
i = 30. Evaluate theoretically, E (S10 ), E (S20 ), and E (S30 ). Next evaluate
Var (S10 ), Var (S20 ), and Var (S30 ). Graph the theoretical quantities against
their observed counterparts.
2. Repeat the experiment in the above question for XiC = Xi − 0.5.
√
3. Repeat the experiment of Question 1, for XiC = (Xi − 0.5)/ 2.5. What can
we say here? Where does the number 2.5 come from? If you had to make a
guess at Var (S19.5 ) what value would you choose?
4. Let W(t) be Brownian motion on the interval (0, 1). Consider the time points
t where t = 1/30, 2/30, ..., 29/30, 1. Write down E W(t) and Var W(t) for
these points. Compare these values with those calculated above.
5. Calculate Si0 = Si − i × S30 /30, based upon XiC . Make some general obser-
vations on this process.
√
6. Repeat each set of simulations; XiC = (Xi − 0.5)/ 2.5, 1000 times. For
each of these repetitions, note down the empirically observed averages of
the means and variances at each value of i for i = 1, ..., 30. Rescale the
interval for i to be the interval (0,1). Plot the average mean and the average
variance against i/30. What conclusions can be made?
√ √
7. Let XiC take the value Xi / 2.5 for i=1,...,10; √
the value (Xi − 0.25)/ 2.5,
for i=11,..., 20; and the value (Xi − 0.50)/ 2.5 otherwise. Plot these
series of values against i. What conclusions can be drawn? Calculate Si0 =
Si − i × S30 /30. What can be said about this process?
8. The condition of homoscedasticity is helpful in obtaining several of the large

sample approximations. In practice the condition is typically respected to a
strong degree. Can you construct a situation in which the assumption may
not be a plausible one? What would be the implications of this for inference?
We provide here some detail to the proofs needed to support the main findings
of this chapter. Broad outlines for these proofs have been given in Chauvel and
O’Quigley (2014, 2017) and further details can be found in Chauvel (2014).
Proposition 9.1 The counting process N̄ on the initial time scale takes values
in N. Thus, for all u ∈ [0, T ],
{x ∈ R : N̄ (u) ≤ x} = {x ∈ R : N̄ (u) ≤ x}.
The inverse process is such that N̄ −1 (j) is the jth ordered time of death, j =
1, . . . , kn . Note that N̄ (N̄ −1 (j)) = j. Thus, by the definition of N̄ ∗ , we have for
0 ≤ t ≤ 1,
n
n

N̄ ∗ (t) = 1 = 1
{N̄ (Xi ) ≤ kn t, δi = 1} {Xi ≤ N̄ −1 (kn t), δi = 1}
i=1 i=1
= N̄ (N̄ −1 (kn t)) = kn t
Lemma 9.1 Suppose that ε > 0. Note that

s s
1
P sup ∗
An (t)dN̄ (t) −
a(t)dt > ε
0≤s≤1 kn 0 0

1 s ε
≤P sup {An (t) − a(t)} dN̄ (t) >
∗
0≤s≤1 kn 0 2
s s
1 ε
+P sup ∗
a(t)dN̄ (t) −
a(t)dt > (9.24)
0≤s≤1 kn 0 0 2
Suppose also that s ∈ [0, 1]. For the first term we have:

kn s
1 s ∗

1

{An (t) − a(t)} d N̄ (t) = (A (t
n i ) − a(t ))
i
kn 0 kn
i=1
kn s
≤ sup |An (ti ) − a(ti )| ≤ sup |An (ti ) − a(ti )| .
kn i=1,..., kn s i=1,...,kn
Therefore,

1 s
sup {An (t) − a(t)} dN̄ (t) ≤ sup |An (ti ) − a(ti )| ,
∗
k
0≤s≤1 n 0 i=1,...,kn
P
and supi=1,...,kn |An (ti ) − a(ti )| −→ 0. As a result,
n→∞

1 s ∗
ε

lim P sup {An (t) − a(t)} dN̄ (t) > = 0.
n→∞ 0≤s≤1 kn 0 2
For the second term on the right-hand side of Equation 9.24 we have:

s s kn s
s
1 1
sup ∗
a(t)dN̄ (t) −
a(t)dt = sup a (ti ) − a(t)dt .
0≤s≤1 kn 0 0 0≤s≤1 kn i=1 0
Note that an is piecewise constant on the interval [0, 1] such that an (t) = a (ti )
for t ∈ [ti , ti+1 [, i = 0, . . . , kn . We then have:
kn s s
1 kn s kn s
a (ti ) = an (t)dt − s − a .
kn 0 kn kn
i=1
As a result we have:
s s
1
sup ∗
a(t)dN̄ (t) − a(t)dt
0≤s≤1 kn 0 0
s s
kn s kn s
≤ sup an (t)dt − a(t)dt + sup s − a
0≤s≤1 0 0 0≤s≤1 kn kn
s s
1 kn s
≤ sup an (t)dt − a(t)dt + sup a ,
0≤s≤1 0 0 kn 0≤s≤1 kn
Note that limn→∞ sups∈[0,1] |an (s) − a(s)| = 0 by the uniform continuity of the
function a. This implies uniform convergence
s s

lim sup an (t)dt − a(t)dt = 0.
n→∞ s∈[0,1] 0 0

kn s
Furthermore, a is bounded so that sup0≤s≤1 a < ∞, which shows
kn
the result since
s s
1
lim sup a(t)dN̄ ∗ (t) − a(t)dt = 0, (9.25)
n→∞ 0≤s≤1 kn 0 0
Proposition 9.2 Let t ∈ [0, 1]. The expectation of Z(t) (defined in (9.2)) condi-
tional on the σ-algebra Ft∗− is
n

Eβ(t) (Z(t) | Ft∗− ) = Eβ(t) Zi (Xi ) 1φ (X ) = t, δ = 1 | Ft∗−
n i i
i=1
n

= Eβ(t) Zi φ−1
n (t) 1φ (X ) = t, δ = 1 | Ft∗−
n i i
i=1
n

= Zi φ−1 ∗
n (t) Eβ(t) 1φn (Xi ) = t, δi = 1 | Ft−
i=1
n

= Zi φ−1 ∗
n (t) Pβ(t) φn (Xi ) = t, δi = 1 | Ft− .
i=1
−1 −1 −
Indeed, Z and φ−1 n are left continuous, so Zi φn (t) = Zi φn (t ) is Ft−
∗
measurable. By definition, the probability that individual i has time of death t

under the model with
parameter β(t), conditional
on the set of at-risk individuals
at time t, is Pβ(t) φn (Xi ) = t, δi = 1 | Ft∗− = πi (β(t), t). Thus,
Eβ(t) (Z(t) | Ft∗− ) = Eβ(t) (Z | t).
The same reasoning for the second-order moment gives the variance result.
Proposition 9.3 Let i, j = 1, 2, . . . , kn , i < j. By definition, the random variable

Vβ(ti ) (Z | ti )−1/2 Z(ti ) − Eβ(ti ) (Z | ti )
is Ft∗i measurable. As the σ-algebras satisfy Ft∗i ⊂ F ∗− , this random variable is

tj
also F ∗− measurable. Furthermore, from Proposition 9.2,
tj

E Vβ(tj ) (Z | tj )−1/2 Z(tj ) − Eβ(tj ) (Z | tj ) Ft∗−
j

= Vβ(tj ) (Z | tj )−1/2 E Z(tj ) − Eβ(tj ) (Z | tj ) Ft∗− = 0.
j
Thus,

E Vβ(ti ) (Z | ti )−1/2 Z(ti ) − Eβ(ti ) (Z | ti ) VjZ Z(tj ) − Eβ(tj ) (Z | tj )

=E E Vβ(ti ) (Z | ti )−1/2 Z(ti )−Eβ(ti ) (Z | ti ) VjZ Z(tj )−Eβ(tj ) (Z | tj ) F ∗−
tj

∗
= E Vβ(ti ) (Z | ti )−1/2 Z(ti )−Eβ(ti ) (Z | ti ) E VjZ Z(tj )−Eβ(tj ) (Z | tj ) F − = 0.
tj
where we have written VjZ in place of Vβ(tj ) (Z | tj )−1/2
Lemma 9.2 Appealing to the law of large numbers, condition A1 implies that
for r = 0, 1, 2, we have:
s(r) (γ(0), 0) = E S (r) (γ(0), 0) = E {Y (0)Z(0)r exp(γ(0)Z(0))} ,
which is E {Z(0)r exp(γ(0)Z(0))} since Y (0) = 1X≥0 = 1. As a result,

2
E Z(0)2 exp(γ(0)Z(0)) E {Z(0) exp(γ(0)Z(0))}
v(γ(0), 0) = −
E {exp (γ(0)Z(0))} E {exp (γ(0)Z(0))}
2
= E Z(0)2 Φγ(0) (Z(0)) − E Z(0)Φγ(0) (Z(0)) ,
where
exp(γ(0)z)
Φγ(0) (z) = , z ∈ R.
E (exp (γ(0)Z(0)))
Furthermore, denoting PZ(0) to be the distribution function of the random
variable Z(0) and X having distribution function PX such that dPX (x) =
Φγ(0) (x)dPZ(0) (x), x ∈ R, we have:
v(γ(0), 0) = E(X 2 ) − E(X)2 = V (X).
The function Φγ(0) being invertible on R, X is distributed according to a Dirac

law if and only if Z(0) follows a Dirac law. We know that V (Z(0)) > 0 which is
equivalent to saying that Z(0) does not follow a Dirac law. Thus, X does not
follow a Dirac law and v(γ(0), 0) = V (X) > 0. The hypothesis A3 concerning
homoscedasticity indicates that, for all t ∈ [0, 1], v(γ(t), t) = v(γ(0), 0), which
leads to the conclusion.
Lemma 9.3 Assume that conditions A1 and A2 hold and that γ1 ∈ B. Recall the
√
definition of v(γ1 (·), ·). We write supn
k to denote n supt∈{t1 ,t2 ,...,tk } so that
n
we have:

supn
k Vγ1 (t) (Z | t) − v(γ1 (t), t)

(2) 2 2
nS (γ1 (t), t) S (1) (γ1 (t), t) s(2) (γ1 (t), t) s(1) (γ1 (t), t)
= supk (0) − − −
S (γ1 (t), t) S (0) (γ1 (t), t) s(0) (γ1 (t), t) s(0) (γ1 (t), t)

(2) 2 2
√ S (γ1 (t), t) S (1) (γ1 (t), t) s(2) (γ1 (t), t) s(1) (γ1 (t), t)
≤ n sup (0) − − −
t∈[0,1] S
(γ1 (t), t) (0)
S (γ1 (t), t) (0)
s (γ1 (t), t) s(0) (γ1 (t), t)
≤ Vn + Wn ,
where
√ S (2) (γ (t), t) s(2) (γ (t), t)
1 1
Vn = n sup (0) − (0) ,
t∈[0,1] S (γ1 (t), t) s (γ1 (t), t)
and 2 2

√ S (1) (γ1 (t), t) s (γ1 (t), t)
(1)
Wn = n sup − .
t∈[0,1] S
(0) (γ1 (t), t) s(0) (γ1 (t), t)
We can study the large sample behavior of these two terms separately. Denote
(0)
m0 , M1 , and M2 to be strictly positive constants such that s (γ(t), t) ≥ m0 et
(i)
s (γ(t), t) ≤ Mi , (i = 1, 2) for all t ∈ [0, 1] and γ ∈ B. Their existence follows
from condition A2. We have:

√ 1
Vn ≤ n sup (0) S (2) (γ1 (t), t) − s(2) (γ1 (t), t)
t∈[0,1] S (γ1 (t), t)

√ (2) 1 1

+ n sup s (γ1 (t), t) − (0)
t∈[0,1] S (γ1 (t), t) s (γ1 (t), t)
(0)
√
1
≤ n sup (0) sup S (2) (γ1 (t), t) − s(2) (γ1 (t), t)
t∈[0,1] S (γ 1 (t), t) t∈[0,1]

√ 1 1
+ M2 n sup (0) − (0) . (9.26)
S (γ (t), t) s (γ (t), t)
t∈[0,1] 1 1
Furthermore,

√ 1 1
n sup (0) − (0)
t∈[0,1],γ∈B S (γ(t), t) s (γ(t), t)

(0) (0) (γ(t), t)
√ s (γ(t), t) − S
= n sup (0) (γ(t), t) s(0) (γ(t), t)
t∈[0,1],γ∈B S
√
1
≤ m−10 n sup s(0) (γ(t), t) − S (0) (γ(t), t) sup .
t∈[0,1],γ∈B t∈[0,1],γ∈B S (0) (γ(t), t)
√

From Equation 9.6, n supt∈[0,1],γ∈B s(0) (γ(t), t) − S (0) (γ(t), t) = oP (1),
and since s(0) is bounded below uniformly by m0 , we have
supt∈[0,1],γ∈B S (0) (γ(t), t)−1 = OP (1), which indicates that this term is bounded
after a certain rank with high probability. As a result,

√ 1 1 P
n
sup (0) − (0) −→ 0. (9.27)
t∈[0,1],γ∈B S (γ(t), t) s (γ(t), t) n→∞
In addition, under condition A1,

√
P
n sup S (2) (γ(t), t) − s(2) (γ(t), t) −→ 0,
t∈[0,1],γ∈B n→∞
−1
and since supt∈[0,1],γ∈B S (0) (γ(t), t) = OP (1), we have convergence for
√
1 (2) (2) P
n sup (0) (γ (t), t)
sup S (γ1 (t), t) − s (γ1 (t), t) −→ 0.
t∈[0,1] S 1 t∈[0,1] n→∞
Equation 9.26 allows us to conclude that Vn converges in probability to 0 when

n → ∞. Similar arguments show that we can bound the term Wn by

√ 1
Wn ≤ n sup (0) S (1)
(γ (t), t) 2
− s (1)
(γ (t), t) 2
(γ1 (t), t)2
1 1
t∈[0,1] S

√ 1 1
+ n sup s(1) (γ1 (t), t)2 −
t∈[0,1]
(0)
S (γ1 (t), t) 2 s (γ1 (t), t)
(0) 2
√
1 (1) 2 (1) 2
≤ n sup (0) sup S (γ 1 (t), t) − s (γ 1 (t), t)
t∈[0,1] S (γ1 (t), t)2 t∈[0,1]

√ 1 1
+ M12 n sup (0) − (0) . (9.28)
S (γ (t), t)
t∈[0,1] 1
2 s (γ (t), t)
1
2
We have:

√ 1 1
n sup −
t∈[0,1],γ∈B S
(0) (γ(t), t)2 (0)
s (γ(t), t) 2

√ 1 1 1 1
= n sup − +
t∈[0,1],γ∈B
(0) (0)
S (γ(t), t) s (γ(t), t) S (γ(t), t) s (γ(t), t)
(0) (0)

√ 1 1 1
≤ n −1
sup
t∈[0,1],γ∈B S
(0) (γ(t), t) − s(0) (γ(t), t) t∈[0,1],γ∈B
sup
S (0) (γ(t), t)
+ m0 .
−1
Equation 9.27 and the fact that supt∈[0,1],γ∈B S (0) (γ(t), t) = OP (1) imply
that

√
1 1 P
n sup 2 − 2 −→ 0.

t∈[0,1],γ∈B S (0) (γ(t), t) s (γ(t), t)
(0) n→∞
Similar arguments can be used to show that

√
P
n sup S (1) (γ(t), t)2 − s(1) (γ(t), t)2 −→ 0.
t∈[0,1],γ∈B n→∞
−2
Since supt∈[0,1],γ∈B S (0) (γ(t), t) = OP (1), we have the convergence of
√
1 (1) 2 (1) 2 P
n sup (0) (γ (t), t)2
sup S (γ1 (t), t) − s (γ1 (t), t) −→ 0.
t∈[0,1] S 1 t∈[0,1] n→∞
As a result, following Equation 9.28, Wn converges in probability to 0 when

n → ∞. In conclusion,
√ P
n sup Vγ (Z | t) − v(γ1 (t), t) −→ 0.
1 (t) n→∞
t∈{t1 ,...,tkn }
Under condition A3, there exists a positive constant C(γ1 ) such that v(γ1 (t), t) =
C(γ1 ) for all t ∈ [0, 1]. Lemma 9.2 indicates the strict positivity of C(γ1 ), and
Equation 9.8 is thus demonstrated.
Equations 9.9 and 9.10. Let γ1 ∈ B. Following Equation 9.8, under conditions
A1, A2, and A3, there exist constants C(γ1 ), C(β0 ) > 0 such that
√ P
n sup Vγ (Z | t) − C(γ1 ) −→ 0, (9.29)
1 (t) n→∞
t∈{t1 ,t2 ,...,tkn }
√ P
n sup Vβ (Z | t) − C(β0 ) −→ 0. (9.30)
0 n→∞
t∈{t1 ,t2 ,...,tkn }
Since kn = kn (β0 ), Equation 9.9 is confirmed:
∀ t ∈ {t1 , . . . , tkn }, Vβ0 (Z | t) > CV w.p. 1
We then have:

V
n γ1 (t)
(Z | t) C(γ1 ) 1
supk − n
≤ supk V (Z | t) − C(γ1 )
Vβ0 (Z | t) C(β0 ) Vβ0 (Z | t) γ1 (t)

n
1 1
+ C(γ1 ) supk −
Vβ0 (Z | t) C(β0 )

n
1 1
−1
≤ CV supk Vγ1 (t) (Z | t)−C(γ1 ) + C(γ1 ) supk n
− .
V (Z | t) C(β0 ) β0
Moreover,

1
1 n Vβ0 (Z | t) − C(β0 )
sup V − = supk
t∈{t1 ,t2 ,...,tkn } β0 (Z | t) C(β0 ) C(β0 )Vβ0 (Z | t)

≤ C(β0 )−1 CV−1 supn
k Vβ0 (Z | t) − C(β0 ) .
The result 9.30 then implies:

√ 1 1 P
n sup − −→ 0.
V C(β0 )
t∈{t1 ,t2 ,...,tkn } β0 (Z | t) n→∞
This convergence together with that of Equation 9.29 lead to the conclusion that

√ Vγ1 (t) (Z | t) C(γ1 ) P
n sup − −→ 0.
C(β0 ) n→∞
t∈{t1 ,t2 ,...,tkn } Vβ0 (Z | t)
The same arguments can be used to prove Equation 9.10.
Theorem 9.1 This is a special case of the more general proof given for Theorem
9.2.
Theorem 9.2 Let t ∈ [0, 1]. Under hypotheses A1, A2, A3, and A4, we suppose
that β(t) is the true value for the model and that the standarized score, Un∗ , is
evaluated at the value β0 ∈ B. Recall the equalities of Proposition 9.2:

Eβ(t) (Z | t) = Eβ(t) Z(t) Ft∗− , Vβ0 (Z | t) = Vβ0 Z(t) Ft∗− .
Define the process {Un∗∗ (β0 , t), t ∈ [0, 1]} by

t
∗∗ Z(s) − Eβ0 Z(s) Fs∗− ∗
Un (β0 , t) = 1/2 dN̄ (s), 0 ≤ t ≤ 1. (9.31)
0 kn Vβ Z(s) F ∗−
0 s
This process is piecewise constant with a jump at each time of event. The out-
line of the proof is the following. Firstly, we will decompose Un∗∗ (β0 , ·) into 2
processes belonging to (D[0, 1], R). We will show that the first process converges
in distribution to a Brownian motion by appealing to a theorem for the con-
vergence of differences of martingales. The convergence of the second process
can be obtained by making use of Lemma 9.1. The large sample behavior of
the process Un∗ (β0 , ·) belonging to (C[0, 1], R), which is the linear interpolation
of the variables {Un∗∗ (β0 , ti ), i = 1, . . . , kn }, will be obtained by showing that
the difference between Un∗ and Un∗∗ tends in probability to 0 when n → ∞. A
Taylor-Lagrange expansion of the conditional expectation Eβ(t) {Z(t) | Ft∗− } at
the point β(t) = β0 gives:

∗ ∗ ∂ ∗
Eβ0
Z(t) Ft− = Eβ(t) Z(t) Ft− + (β0 − β(t)) Eβ Z(t) Ft−
,
∂β β=β̃(t)
(9.32)
where β̃ belongs to the ball centered at β having radius supt∈[0,1] |β(t) − β0 |.

Recall that
∂
Eβ Z(t) Ft∗− = Vβ Z(t) Ft∗− .
∂β
∗∗
√
Therefore Un (β0 , t) = Xn (t) + kn An (t), where

1 t Vβ̃(s) Z(s) Fs∗− ∗
An (t) = {β(s) − β0 } 1/2 dN̄ (s),
kn 0 Vβ0 Z(s) Fs− ∗
t
1 Z(s) − Eβ(s) Z(s) Fs∗− ∗
Xn (t) = 1/2 1/2 dN̄ (s).
kn 0 Vβ0 Z(s) Fs∗−
We need to consider the limiting behavior of An . We have:

t

sup An (t) − C2 {β(s) − β0 } ds
t∈[0,1] 0

t
1 Vβ̃(s) Z(s) Fs∗−
≤ sup |β(t) − β0 | sup dN̄ (s) − C2 t
∗
1/2
t∈[0,1] t∈[0,1] kn 0 Vβ Z(s) F ∗−
0 s

t
1 Vβ̃(s) Z(s) Fs∗−
≤ δ1 sup dN̄ (s) − C2 t ,
∗
(9.33)
1/2
t∈[0,1] kn 0 Vβ Z(s) F ∗−
0 s
since β0 ∈ B with radius δ1 (condition A1). Under conditions A1, A2, and A3,
Equation 9.10 of Lemma 9.3 indicates the existence of a constant C2 > 0 such
that

Vβ̃(s) Z(s) Fs∗− P
sup − C2 −→ 0.
1/2
s∈{t1 ,t2 ,...,tkn } Vβ Z(s) F ∗− n→∞
0 s
Lemma 9.1 then implies:

t
1 Vβ̃(s) Z(s) Fs∗− P
sup d N̄ ∗
(s) − C t −→ 0.
1/2 2
t∈[0,1] kn 0 Vβ Z(s) F ∗− n→∞
0 s
Equation (9.33) now indicates that the convergence of Equation 9.13 is shown.
We see that Xn converges in distribution to Brownian motion by making use of
the functional central limit theorem for the differences of martingales developed
by Helland (1982) and described in Theorem C.4. Let j = 1, . . . , kn . We write
ξj,kn as the jth increment of Xn :

Z(tj ) − Eβ(tj ) Z(tj ) F ∗−
tj
ξj,kn = 1/2
.

∗
kn Vβ0 Z(tj ) F −
tj

The increment ξj,kn is Ft∗j -measurable and Eβ(tj ) ξj,kn F ∗− = 0. The pro-
tj
cess N̄ ∗ contains a jump at each time of an event tj = j/kn (Proposition 9.1),
we then have:
tkn
tkn /kn Z(s) − E
β(s) Z(s) | Fs∗−
∗
Xn (t) = 1/2 d N̄ (s) = ξj,kn ,
0 kn Vβ0 Z(s) Fs∗− j=1
where the function t −→ tkn is positive, increasing with integer values and
right continuous. To check the assumption of Theorem C.4 we need to consider
the expectation, the variance, and the Lyapunov condition: We begin with the
expectation.
a. (Expectation). The inclusion of the σ-algebras Ft∗j−1 ⊂ F ∗− leads to

tj

∗
Eβ(tj−1 ) ξj,kn Ft∗j−1 = Eβ(tj−1 ) Eβ(tj ) ξj,kn Ft∗− Ft
j−1 = 0.
j
(9.34)
b. (Variance). Under the hypotheses A1, A2, and A3, following Lemma 9.3,
there exists a constant C1 (β, β0 ) > 0 such that

Vβ(tj ) Z(tj ) F ∗−
tj P
sup − C1 (β, β0 )
2
−→ 0.
j=1,...,kn
n→∞
β0
V Z(t ) F ∗

j − tj
In particular, C1 (β, β0 ) = 1 whenever β(t) = β0 . Let s ∈ {t1 , t2 , . . . , tkn }.

From Equation 9.9,

Vβ0 Z(s) Fs∗− > CV a.s.
We also have, according to the hypothesis A4, the upper bound,

n

|Z(s)| = Zi (Xi ) 1φ (X ) = s, δ = 1
n i i
i=1
n

≤ sup |Zi (Xi )| 1φ (X ) = s, δ = 1 ≤ L, (9.35)
i=1,...,n n i i
i=1
n
since s ∈ {t1 , t2 , . . . , tkn } donc i=1 1φn (Xi ) = s, δi
= 1 = 1. As a result,

Vβ(s) Z(s) Fs∗− ≤ Eβ(s) Z(s)2 Fs∗− ≤ L2 .
In conclusion, we have, almost surely, the upper bound,

tj L2
sup ≤ ,
CV
j=1,...,kn
Vβ0 Z(tj ) F ∗−
tj
and an application of the dominated convergence theorem implies that

tj L1
sup
− C1 (β, β )2
0 −→ 0. (9.36)
j=1,...,kn n→∞

Vβ0 Z(tj ) Ft−
∗

j
Next, we have that

tkn

2 ∗ L1
Eβ(tj−1 ) ξj,kn
Ftj−1 −→ C1 (β, β0 )2 t.
n→∞
j=1
Vβ(t ) (Z(tj ) | F ∗− )
j tj
and, to keep things uncluttered, we first write W0F (tj ) = ,
Vβ0 (Z(tj ) | F ∗− )
tj
then,

tkn 2 ∗
E ξ F − C (β, β )2
t
β(tj−1 ) j,kn tj−1 1 0
j=1

tkn
= Eβ(tj−1 ) Eβ(tj ) ξj,k 2 F ∗− Ft∗j−1 − C1 (β, β0 )2 t
n tj
j=1

tkn
1
= Eβ(tj−1 ) W0F (tj ) Ft∗j−1 − C1 (β, β0 )t
kn j=1
tkn
1 F 2

2 tkn

≤ ∗
Eβ(tj−1 ) W0 (tj ) − C1 (β, β0 ) Ftj−1 + C1 (β, β0 ) − t .
kn kn
j=1
As a result of the above we now have:

⎛ ⎞
tkn

E ⎝ 2
Eβ(tj−1 ) ξj,kn
Ft∗
j−1
− C1 (β, β0 )2 ⎠
t
j=1
tkn
1 tkn
≤ E W0F (tj ) − C1 (β, β0 )2 + C1 (β, β0 )2 − t
kn kn
j=1

tkn tkn
≤ sup E W0F (tj ) − C1 (β, β0 )2 + C1 (β, β0 )2 − t
kn j=1,..., tkn kn

tkn
F 2 2 tkn
≤ E sup W0 (tj ) − C1 (β, β0 ) + C1 (β, β0 ) − t
kn j=1,..., tkn kn

tkn tkn

≤ E sup W0F (tj ) − C1 (β, β0 )2 + C1 (β, β0 )2 − t .
kn j=1,...,kn kn

tkn 2 ∗
Equation 9.36 then leads to the conclusion that j=1 Eβ ξj,kn
Ftj−1
converges in mean and, as a result, in probability to C1 (β, β0 )2 t when
n → ∞.
tk 3
c. We have the convergence in mean of j=1n Eβ(tj−1 ) ξj,k n
| Ft∗j−1 to 0
when n → ∞. Furthermore,
⎛ ⎞
tkn tkn
3 3 ∗
E ⎝ Eβ(tj−1 ) ξj,kn | Ft∗j−1 ⎠ ≤ E Eβ(tj−1 ) ξj,k n
| Ftj−1
j=1 j=1
tkn

= E ξj,k
3
n
, (9.37)
j=1
and keeping in mind that for j = 1, . . . , kn , then
Z(tj ) − Eβ(tj ) Z(tj ) | F ∗−

tj
ξj,kn = 1/2 .

kn Vβ0 Z(tj ) F −
∗
tj

3
E Z(t) − Eβ(t) Z(t)|Ft∗− ≤ 8L3 .

Also, on the basis of Equation 9.9, t ∈ {t1 , t2 , . . . , tkn } ,Vβ0 Z(t)|Ft∗− is
strictly greater than CV . As a result, Equation 9.37 becomes:
⎛ ⎞
tkn

⎝
E Eβ(tj−1 ) ξj,kn | Ftj−1 ⎠
3 ∗
j=1
kn
3
1 8L3
≤ E Z(tj ) − Eβ(t ) Z(tj )|F − ≤∗
,
(kn CV )3/2 j tj 1/2 3/2
j=1 kn CV
(9.38)
tk 3
and j=1n Eβ(tj−1 ) ξj,k n
| Ft∗j−1 converges in mean to 0 when n → ∞
and, in consequence, in probability. The Lyapunov condition is verified.
We can now apply Theorem C.4 and conclude that

L
Xn −→ C1 (β, β0 )W,
n→∞
where the constant C1 (β, β0 )2 is the ratio of two asymptotic conditional variances
v evaluated at β and β0 . When β = β0 , we have C1 (β, β0 ) = 1. We have the
relation

Un∗ (β0 , ·) − kn An = Un∗ (β0 , ·) − Un∗∗ (β0 , ·) + Un∗∗ (β0 , ·) − kn An
= Un∗ (β0 , ·) − Un∗∗ (β0 , ·) + Xn . (9.39)
We have then established the limiting behavior of Xn when n → ∞. Next we look

at the convergence in probability of Un∗ (β0 , ·) − Un∗∗ (β0 , ·) to zero when n → ∞
with respect to the uniform norm. Let ε > 0,

P ( Un∗ (β0 , ·) − Un∗∗ (β0 , ·) ≥ ε) = P sup |Un∗ (β0 , t) − Un∗∗ (β0 , t)| ≥ ε
t∈[0,1]

=P sup |Un∗ (β0 , ti ) − Un∗∗ (β0 , ti−1 )| ≥ε
i=1,...,kn
⎛ ⎞

Z(t ) − E ∗
⎜ 1 j β0 Z(tj ) | Ft− ⎟
⎜ j ⎟ 2L
= P ⎜√ sup ≥ ε⎟ ≤ P √ ≥ε ,
⎝ kn i=1,...,kn 1/2 ⎠ kn CV
V ∗
β0 Z(tj ) F − tj
P
Therefore, Un∗ (β0 , ·)−Un∗∗ (β0 , ·) −→ 0. Slutsky’s theorem applied to the decom-
n→∞
position of Equation 9.39 leads us to conclude that
L
Un∗ (β0 , ·) − kn 1/2 An −→ C1 (β, β0 )W,
n→∞
in (C[0, 1], R) with respect to the topology of uniform convergence.
Proposition 9.4 Recall that Ax ≤ p A x et AB ≤ p A B for A, B ∈

Mp×p (R) et x ∈ Rp . Note that, if A is diagonal and its elements are positive,
we have:
1/2
1/2 1/2
A = max A1/2 ≤ max (A)ll = A .
l=1,...,p ll l=1,...,p
Following the condition

B3,Σ is a symmetric matrix having real coefficients so

that Σ1/2 exists and Σ1/2 < ∞. We have:
√
n sup Vγ(t) (Z | t) − Σ
t∈{t1 ,t2 ,...,tkn },γ∈B
√

1/2 1/2 1/2 1/2
= n sup Vγ(t) (Z | t) − Vγ(t) (Z | t) Σ + Vγ(t) (Z | t) Σ − Σ
t∈{t1 ,t2 ,...,tkn },γ∈B
√

1/2 1/2
≤p n sup Vγ(t) (Z | t) − Σ
t∈{t1 ,t2 ,...,tkn },γ∈B

1/2 1/2
× sup Vγ(t) (Z | t) + Σ .
t∈{t1 ,t2 ,...,tkn },γ∈B
Now, for all t ∈ {t1 , . . . , tkn } and γ ∈ B , according to the condition B4,

Vγ(t) (Z | t) ≤ Eγ Z ⊗2 | t + ⊗2
Eγ (Z | t) ≤ 2L 2 . (9.40)
Therefore, in choosing an orthogonal transformation matrix Pγ(t) (t) of norm 1

in the diagonalisation of the symmetric positive definite matrix Vγ(t) (Z | t) =
Pγ(t) (t) Dγ(t) (t) Pγ(t) (t)T with Dγ(t) (t) diagonal, we have:
2 1/2

Vγ(t) (Z | t)1/2 ≤ p2 Pγ(t) (t) Dγ(t) (t)1/2 ≤ p2 Dγ(t) (t)
1/2

= p2 Pγ(t) (t)T Pγ(t) (t)Dγ(t) (t)Pγ(t) (t)T Pγ(t) (t)
2 1/2 1/2

≤ p2 p2 Pγ(t) (t)Dγ(t) (t)Pγ(t) (t)T Pγ(t) (t) = p3 Vγ(t) (Z | t)
√
≤ 2p3 L .
√

from which, supt∈{t1 ,t2 ,...,tk },γ∈B Vγ(t) (Z | t)1/2 ≤ 2p3 L , and
n
√
n sup Vγ(t) (Z | t) − Σ
t∈{t1 ,t2 ,...,tkn },γ∈B
√ √
1/2
≤p 2p3 L + Σ1/2 n sup 1/2
Vγ(t) (Z | t) − Σ . (9.41)
t∈{t1 ,t2 ,...,tkn },γ∈B
The result then follows from Equation 9.15.
Theorem 9.3 The proof follows the same outline as in the univariate setting.
Recall that the space of cadlag functions of (D[0, 1], Rp ) is equipped with the
product topology of Skorokhod, described in Definition A.5. Furthermore, the
norm of the vector in Rp or a matrix of Mp×p (R) is the sup norm. The pro-
cess being evaluated at β0 ∈ B , we have kn = kn (β0 ) and the upper bound of
Equation 9.17 becomes:

∀ t ∈ {t1 , . . . , tkn }, Vβ0 (t) (Z | t)−1 ≤ CV−1 a.s. (9.42)
We take Un∗∗ to be a cadlag process having a jump at each event time ti =

i/kn , i = 1, . . . , kn , such that
tkn
1 −1/2
Un∗∗ (β0 , t) = √ Vβ0 Z(ti ) Ft∗i − Z(ti ) − Eβ0 Z(ti ) Ft∗i − , 0 ≤ t ≤ 1.
kn
i=1
Let γ ∈ B . Recall the equalities of Proposition 9.2:

Eγ(ti ) (Z | ti ) = Eγ(ti ) Z(ti ) Ft∗i − , Vγ(ti ) (Z | ti ) = Vγ(ti ) Z(ti ) Ft∗i − .
√
Note that Un∗∗ (β0 , t) = Xn (t) + kn Bn (t) for t ∈ [0, 1],
tkn
1
Xn (t) = √ Vβ0 (Z | ti )−1/2 Z(ti ) − Eβ(ti ) (Z | ti )
kn i=1
tkn
1
Bn (t) = Vβ0 (Z | ti )−1/2 Eβ(ti ) (Z | ti ) − Eβ0 (Z | ti ) .
kn
i=1
We have:

∗
Un (β0 , ·) − kn Bn − Wp

≤ Un∗ (β0 , ·) − Un∗∗ (β0 , ·) + Un∗∗ (β0 , ·) − kn Bn − Wp
= Un∗ (β0 , ·) − Un∗∗ (β0 , ·) + Xn − Wp . (9.43)
Let us consider the limiting behavior of the two terms on the right of the inequal-
ity. Considering just the first term and letting ε > 0, we have:

P sup Un∗ (β0 , t) − Un∗∗ (β0 , t) ≥ ε
t∈[0,1]

∗∗ i i − 1
=P sup ∗∗
Un β0 , kn − Un β0 , kn ≥ ε

i=1,...,kn

1
=P √ sup Vβ0 (Z | ti )−1/2 Z(ti ) − Eβ0 (Z | ti ) ≥ ε
kn i=1,...,kn

p −1/2
≤P √ sup Vβ0 (Z | ti ) Z(ti ) − Eβ0 (Z | ti ) ≥ ε . (9.44)
kn i=1,...,kn
We also have, for i = 1, . . . , kn ,

−1/2 −1 ∗ 1/2
Vβ Z(ti ) F ∗ − = Vβ Z(ti ) F ∗ − V Z(t i Fti −
)
0 ti 0 ti β0

∗ −1 ∗ 1/2
≤ p V
β0 Z(t i Fti −
) V
β0 Z(t i Fti −
)

√
≤ CV−1 2p4 L , (9.45)
where the last inequality is obtained from Equations 9.41 and 9.42. Also we have:

Z(ti ) − Eβ0 Z(ti ) Ft∗i − ≤ Z(ti ) + Eβ0 Z(ti ) Ft∗i − ≤ 2L .
The inequality 9.44 implies that

√
2 2
P sup Un∗ (β0 , t) − Un∗∗ (β0 , t) ≥ε ≤P √ p5 CV−1 (L )2 ≥ ε ,
t∈[0,1] kn
(9.46)
and lim P supt∈[0,1] Un∗ (β0 , t) − Un∗∗ (β0 , t) ≥ ε = 0. This implies the con-
n→∞
vergence in probability of Un∗ (β0 , ·) − Un∗∗ (β0 , ·) to 0 when n → ∞ in the space
(D[0, 1], Rp ) equipped with the Skorohod product topology as a consequence of
the results of Section A.5. We now consider the second term. The convergence in
distribution of Xn to Wp is obtained by an application of the multivariate func-
tional central limit theorem C.5 described in Helland (1982). Before applying this
theorem we need to check some working hypotheses. Let i ∈ {1, 2, . . . , kn } and
t ∈ [0, 1]. Note that
1 −1/2
1
ξi,kn = (ξi,k n
p
, . . . , ξi,k n
) = √ Vβ0 Z(ti ) Ft∗i − Z(ti ) − Eβ(ti ) Z(ti ) Ft∗i −
kn
tk
The ith increment in Rp of the process Xn . We have Xn (t) = i=1n ξi,kn .

Note that ξi,kn is Ft∗i -measurable and that Eβ(ti ) ξi,kn F ∗− = 0. Let l, m ∈
ti
{1, . . . , p}, l = m. We have el the lth vector of the canonical base of Rp : all of
the elements are zero with the exception of the lth which is 1. Also, eTl em = 0,
eTl el = 1 and el = maxi=1,...,p (el )i = 1. For i = 1, . . . , kn , we have:
l
ξi,k n
= eTl ξi,kn = ξi,k
T
e ∈ R.
n l
(9.47)
a.’ (Table of martingale differences.) The inclusion of the σ-algebras Ft∗i−1 ⊂

F ∗− and the centering of the increments implies that
ti

∗ ∗ ∗
Eβ(ti−1 ) l
ξi,k Fti−1 = Eβ(ti−1 ) Eβ(ti ) l
ξi,k F − Fti−1 = 0.
n n ti
b.’ (Non-correlation.) Using Equation 9.47, we have:

l
Eβ(ti ) ξi,k ξm
n i,kn
Ft∗− = eT T ∗
l Eβ(ti ) ξi,kn ξi,kn Ft− em
i i

−1/2 −1/2
1 T
= el Vβ0 Z(ti ) Ft∗− Vβ(ti ) Z(ti ) Ft∗− Vβ0 Z(ti ) Ft∗− em
kn i i i
1 T −
= e V ti , β0 , β(ti ) em ,
kn l
where

V t−

Z(ti ) F ∗−
−1/2

Vβ(ti ) Z(ti ) F ∗−

Z(ti ) F ∗−
−1/2
i , β0 , β(ti ) = Vβ0 ti ti
Vβ0
ti
.
Note that Ip is the identity matrix of Mp×p (R). Also, eTl Ip em = 0. We

have:
⎛ ⎞

tkn
⎝
E Eβ(ti−1 ) ξi,kn ξi,kn Fti−1 ⎠
l m ∗
i=1
⎛ ⎞
tkn
1 − ∗
=E ⎝ Eβ(ti−1 ) el V ti , β0 , β(ti ) em Fti−1 ⎠
T
kn i=1
1
tkn

≤ E eTl V t−
i , β0 , β(ti ) em
kn
i=1
tkn
1

= E eTl V t−
i , β0 , β(ti ) − Ip em
kn
i=1
tkn
T
≤ sup E el V t−
i , β0 , β(ti ) − Ip em
kn i=1,..., tkn

tkn
T
≤ E sup el V t−
i , β0 , β(t i ) − Ip em
kn i=1,..., tkn

2 tkn
−
≤p eTl em E sup V t , β0 , β(ti ) − Ip
i
kn i=1,..., tkn

2 tkn
−
=p E sup V t , β0 , β(ti ) − Ip
i
kn i=1,..., tkn

2 tkn
−
≤p E sup V t , β0 , β(ti ) − Ip . (9.48)
i
kn i=1,...,kn
We need to show that

− L1
sup V t , β0 , β(ti ) − Ip −→ 0.
i n→∞
i=1,...,kn
We have:
−
sup V ti , β0 , β(ti ) − Ip
i=1,...,kn

−1/2

= sup Vβ0 Z(ti ) F
∗
Vβ(ti ) Z(ti ) F
∗
i=1,...,kn
t
−
i
t
−
i

∗ ∗ −1/2
−Vβ0 Z(ti ) F − Vβ0 Z(ti ) F −
t
i
t
i

−1/2 2
V ∗
≤p
2
sup Vβ0 Z(ti ) F ∗ sup β(ti ) Z(ti ) F −
i=1,...,kn
t
−
i
i=1,...,kn
t
i

−Vβ0
∗
Z(ti ) F − .
t
i

sup V t− i , β0 , β(ti ) − Ip

i=1,...,kn
2
−2
≤ 2p10 CV L sup Vβ(ti ) Z(ti ) Ft∗− − Vβ0 Z(ti ) Ft∗−
i=1,...,kn i i
2
−2
≤ 2p10 CV L sup Vβ(ti ) Z(ti ) Ft∗− − Σ +
i=1,...,kn i

Σ − V ∗
sup β0 Z(ti ) F − .
t
i=1,...,kn i
Finally, the convergence of Equation 9.16 implies the convergence in proba-

bility,
− P
sup V t , β0 , β(ti ) − Ip −→ 0.
i n→∞
i=1,...,kn

Equations 9.40 and 9.45 indicate that supi=1,...,kn V t−i , β0 , β(ti ) − Ip

is bounded by a finite constant. The convergence is then again a convergence
in mean.
L1
sup V t− i , β0 , β(ti ) − Ip
−→ 0. (9.49)
i=1,...,kn n→∞
Equation 9.48 leads us to conclude that

kn t
L1
l
Eβ(ti−1 ) (ξi,k ξ m |Ft∗− )
n i,kn
−→ 0,
i−1 n→∞
i=1
and the convergence is one in probability.
c.’ (Variance.) Similar arguments to a.’ and b.’ can be used, implying the equality,
2 ∗
tkn
l
Eβ(ti−1 ) ξi,kn Fti−1 − t
i=1
tkn
1 − kn t
= Eβ(ti−1 ) eT
l V ti , β0 , β(ti ) − Ip el Ft∗i−1 + −t .
kn kn
i=1
This leads to
⎛ ⎞
tkn
2
E ⎝ Eβ(ti−1 ) l
ξi,k n
Ft∗
i−1 − t ⎠

i=1

tkn
1 kn t

≤ E eTl V t−
i , β0 , β(t i ) − Ip el + − t
kn kn
i=1
tkn
p2 kn t
≤ E V t− i , β0 , β(t i ) − I p
+
− t

kn kn
i=1

kn t kn t
≤ p2 sup E V t− i , β 0 , β(t i ) − Ip
+
− t

kn i=1,...,kn kn

kn t kn t
≤ p2 E sup V t− i , β 0 , β(t i ) − Ip
+
− t .

kn i=1,...,kn kn
Equation 9.49 then leads to the conclusion that

tkn 2 L1
l F∗ −→
i=1 Eβ(ti−1 ) ξi,k n ti−1 n→∞
t, indicating convergence in
probability

l l
d.’ (Lyapunov condition.) We have ξi,k n
≤ max l=1,...,p ξi,kn = ξi,kn .
Therefore,
kn t
l 3 ∗
kn t
l 3
kn
3
E Eβ(ti−1 ) ξi,k Ft ≤ E ξi,k ≤ E ξi,kn .
n i−1 n
i=1 i=1 i=1
Using Equation 9.45 and condition B4, we have the bound,

⎛ ⎞
kn t

l 3 ∗
E⎝ Eβ(ti−1 ) ξi,kn Fti−1 ⎠
i=1
kn −1/2
p3 3 ∗ 3
≤ 3/2
E Vβ0 Z(ti ) Ft∗i − Z(ti ) − Eβ(ti ) Z(ti ) Fti −
kn i=1
p 3 √ 3
≤ 1/2
CV−1 2p4 L (2L )3 .
kn

kn t l 3 ∗
In consequence, i=1 Eβ(ti−1 ) ξi,kn Fti−1 converge to 0 in mean
and therefore in probability when n → ∞.
In conclusion, all of the hypotheses of Theorem C.5 are met, which allows us
to conclude that Xn converges to Wp when n → ∞ in the space (D[0, 1], Rp )
equipped with the product topology of Skorohod. It then follows, as a result of
Equation 9.43 and the convergence result of the first term on the right of Equa-
tion 9.46, Equation 9.18 is proved. To close the demonstration of the theorem it
remains to prove Equation 9.19. Let t ∈ [0, 1]. The multivariate Taylor-Lagrange
inequality implies that for s ∈ {t1 , . . . , tkn },

Eβ(s) (Z | s) − Eβ (Z | s) − Vβ (Z | s) (β(s) − β0 ) ≤ 1 Mn β(s) − β0 2
.
0 0
2
(9.50)
As a result,
t

Bn (t) − Σ1/2 {β(s) − β0 } ds

0
tk
n
1/2 t
=
1
Vβ0 (Z | ti ) −1/2
Eβ(ti ) (Z | ti ) − Eβ0 (Z | ti ) − Σ

{β(s) − β0 } ds
kn i=1 0
1
tkn

≤ Vβ0 (Z | ti )−1/2 Eβ(ti ) (Z | ti ) − Eβ0 (Z | ti ) − Vβ0 (Z | ti ) {β(ti ) − β0 }
kn
i=1
1
tkn

+ Vβ0 (Z | ti )1/2 − Σ1/2 {β(ti ) − β0 }
kn
i=1
tkn t

1 tkn
+ Σ1/2 β(ti ) − β(s)ds + β0 −t
kn
i=1 0
kn

tkn

tkn

≤
pMn Vβ0 (Z | ti )−1/2 β(ti ) − β0 2 + p Vβ0 (Z | ti )1/2 − Σ1/2 β(ti ) − β0
2kn kn
i=1 i=1
tk t
1/2
1
n
tkn − t

+p Σ k β(ti ) − β(s)ds + Σ1/2 β0
n i=1 0 k n
tkn pMn
≤ sup β(ti ) − β0 2 sup Vβ0 (Z | ti )−1/2
kn 2 i=1,...,tkn i=1,...,tkn
tkn
+p sup β(ti ) − β0 sup Vβ0 (Z | ti )1/2 − Σ1/2
kn i=1,...,tkn i=1,...,tkn
tk t

tkn − t
n
+ p Σ1/2 sup
1
βl (s)ds + p Σ1/2 β0 .
βl (ti ) −
l=1,...,p kn 0 kn
i=1
Since β0 and 0 ∈ B donc β0 ≤ 2δ2 et supi=1,..., tkn β(ti ) − β0 ≤

supt∈[0,1] β(t) − β0 ≤ δ2 . Now, Equation 9.45 implies that

−1/2 √
sup Vβ Z(ti ) F ∗ − ≤ 2C −1 p4 L .
0 ti V
i=1,...,kn
Therefore,
t

Bn (t) − Σ1/2 {β(s) − β0 } ds

0
tkn p5 2 −1 tk

≤ Mn √ δ2 CV L + p n δ2 sup Vβ0 (Z | ti )1/2 − Σ1/2
kn 2 kn i=1,...,kn

tkn t
1/2 1 tkn
+ p Σ sup βl (ti ) −
βl (s)ds + 2δ2 p − t Σ1/2 .
k
l=1,...,p n i=1 0 kn
Bringing in the sup, we have:

t

sup B
n (t) − Σ1/2
{β(s) − β0 } ds

t∈[0,1] 0
p5

≤ Mn √ δ22 CV−1 L + pδ2 sup Vβ0 (Z | ti )1/2 − Σ1/2
2 i=1,...,kn

tkn t
1/2 1 2
+ p Σ sup sup βl (ti ) − βl (s)ds + δ2 p Σ1/2 .
t∈[0,1] l=1,...,p kn i=1 0 kn
(9.51)
According to B3, limn→∞ Mn = 0, and

P
sup Vγ(t) (Z | t)1/2 − Σ1/2 −→ 0.
t∈{t1 ,t2 ,...,tkn },γ∈B n→∞
Let l = 1, . . . , p. The function βl is bounded, allowing us to apply Equation 9.25

of the proof of Lemma 9.1 so that

tkn t
1
lim sup βl (ti ) − βl (s)ds = 0.
n→∞ t∈[0,1] kn 0
i=1
It follows that
tk t tk t

1 1
n p n

sup sup βl (ti ) − βl (s)ds ≤ sup βl (ti ) − βl (s)ds ,
t∈[0,1] l=1,...,p kn 0 t∈[0,1] kn 0
i=1 l=1 i=1
and since p is finite,

tkn t
1
lim sup sup β(ti )l − β(s)l ds = 0.
n→∞ t∈[0,1] l=1,...,p kn 0
i=1
According to Equation 9.51, the convergence of Equation 9.19 is then established.

Chapter 10
Model construction guided by regression

effect process
10.1 Chapter summary
The basic ideas of linear regression run through this chapter. It is structured
as follows: In Section 10.3, we present a literature review of graphical methods
and goodness of fit tests for proportional hazards models. In Section 10.6, we
consider the R2 coefficient for non-proportional hazards models, which is the
measure of predictive ability used from here on. Although other suggestions for
R2 have been discussed extensively in the literature O’Quigley (2008), we limit
our attention to this particular one in view of its many desirable properties, most
importantly a key property, not known to exist for rival measures, that the popula-
tion equivalent for R2 , Ω2 , is maximized for the model closest to that generating
the observations. Section 10.4 presents a graphical method and a goodness of fit
test for proportional hazards models, based on the regression effect process. We
also present a method for constructing non-proportional hazards models from
this process and the R2 coefficient. Next, simulations are presented in Section
10.9 comparing the performance of the goodness of fit test we have developed
with some standard tests. The method for constructing non-proportional hazards
models will then be illustrated on simulated data. Applications on real data are
presented in Section 10.10. We terminate the chapter with a discussion.
In linear regression it is clear that bringing a non-linear model, via transformation,

under a linear framework will improve predictability. Predictability is also going
to improve if, via regression on covariates, we reduce the residual variance. We
see in this chapter that these two distinct, although not orthogonal, perspectives

261
https://doi.org/10.1007/978-3-030-33439-0_10
262 Chapter 10. Model construction guided . . .
can be investigated to improve predictive strength. These two perspectives, fit

and the quantification of the strength of regression effect, are not only related
to one another but each relates directly to the regression effect process of the
previous chapter. These loose intuitive ideas are made firm via some important
theorems and corollaries. These results can then be brought together to guide
model building. Several detailed examples highlight their practical importance.
The statistical properties of the regression effect process established in Chap-
ter 9 can be used to guide model building and, in particular, for non-proportional
hazards models. It turns out that the regression effect process essentially contains
all the information we need and can steer us in the direction of better models.
These resulting models will necessarily have good fits, and optimal predictive
power as measured by the R2 coefficient given the available information. We will
see how goodness of fit tests can be structured under the heading of the regres-
sion effect process. This will apply at all levels of model complexity, from the
simple proportional hazards model for a binary covariate to higher dimensional
models in which effects may be a mixture of proportional and non-proportional
ones. Confidence bands for the process corresponding to a constant effect can
provide some initial building blocks. As described in the introductory chapter,
goodness of fit and predictive ability are two very different aspects of statistical
models, and the link between the two is not always clear. In this chapter, we con-
sider more formally how improving the fit of a non-proportional hazards model
will increase its predictive power.
The purpose of a statistical model when representing data is twofold: first
to simplify as much as possible the several relationships and sub-relationships
that may be present in the data, and secondly, to use such a structure as an
inferential tool. An important question raised by the use of a particular model
on observed data is its goodness of fit; i.e., is it plausible to consider that the
data can be viewed as having been generated from that model? Procedures for
evaluating the goodness of fit can be used to study the quality of a model as
a whole, or certain parts only. For example, in a clinical trial context, we may
question whether the working hypotheses concerning differences in treatment are
satisfied in the presence of other covariates, without considering how well these
covariates are modeled. The goodness of fit of a model can be assessed with
hypothesis testing or using graphical methods. The interpretation of goodness of
fit tests must be performed with care. Most of them are powerful for detecting
specific alternative hypotheses, e.g., an effect that increases or decreases over
time. We need to have an idea of the type of alternative hypothesis to consider
when looking at a dataset, but this is not always realistic. Moreover, the result of
a test is often limited to reading off the p-value. This value can decrease simply
by increasing the number of observations, and not with a necessarily better fit.
Rejecting the null hypothesis of a good fit of the model to the data does
not generally get us very far. We need some kind of pointers on where to go
to improve fit. Graphical methods, for their part, often have the advantage of
10.3. Classical graphical methods 263
suggesting visually the type of alternative hypothesis, i.e., more realistic models
to consider. Even here, results can be confusing or difficult to summarize if we
are not used to interpreting such plots. In our view the regression effect process,
processes in the multivariate case, can ease interpretation and play a fundamental
role in model construction. Before looking more closely at the regression effect
process and especially the time-transformed regression effect process, we present
a brief overview of some of the less recent proposals in the literature.
10.3 Classical graphical methods
The first graphical method proposed for testing the proportional hazards hypoth-
esis for categorical variables was proposed by Kay (1977). The method is based
on the following formulation of the proportional hazards model:
t
T
log {− log(S(t|Z))} = β Z + log λ0 (s)ds .
0
Thus, if the covariate Z has categories z1 , . . . , zM , the lines log(− log(S(t|Z =

zm )) plotted as a function of time for each category m = 1, . . . , M, are parallel.
In practice, the conditional survival function S(·|Z) is replaced by its Kaplan-
Meier estimator Ŝ(·|Z) (Equation 3.1). There are certain disadvantages to this
method. For instance, continuous covariates have to be discretized, and the
choice of points at which to cut is not always obvious (Andersen, 1982). Second,
the parallel feature we are looking for between the estimated, noisy functions is
not always readily detectable. Another way to proceed is to plot the difference
between the functions and look at whether this remains constant over time. Thus,
with two categories, we can plot one as a function of the other at different times,
which should be reasonably well approximated by a straight line (Andersen, 1982).
The larger variance in the Kaplan-Meier estimators later in the study where often
there will be few remaining subjects can sometimes obscure a picture of simple
linearity. We may suspect a deviation from linearity at these later time points
that is no more than a manifestation of increased noise.
Other procedures are essentially based on studying residuals and these, quite
naturally, bring us closer to our own approach based on the regression effect
process. The first such residuals used in graphical goodness of fit methods
were those presented in Cox and Snell (1968). Their distribution under the
Cox model was studied by Kay (1977). Lagakos (1981), O’Quigley (1982) and
Crowley and Storer (1983) have underlined their weak power for detecting
poor fits. Furthermore, these residuals also reflect variability in the covariate
(Baltazar-Aban and Peña, 1995). Graphical methods for examining the goodness
of fit of proportional hazards models based on residuals can be divided into
two categories according to whether the residuals are cumulative or not. Within
the non-cumulative methods, a large class of martingale residuals from Barlow
and Prentice (1988) can be examined by plotting its elements as a function of

observed times of deaths or their ranks. The martingale residual of individual
j = 1, . . . , n, at time t is defined by
t
Mj (t) = Nj (t) − Λ̂j (t), with Λ̂j (t) = πj (β̂, s)dN̄ (s),
0
where β̂ is the partial maximum likelihood estimator of β. If the model is valid,

M1 (t), . . . , Mn (t) are the outputs of independent martingales (Gill, 1980). The
class of martingale residuals in Barlow and Prentice (1988) is then defined by
t
φ(s)Mj (s)ds, j = 1, . . . , n, 0≤t≤T,
0
where φ is a deterministic or random and predictable function of time. The

integral of a predictable process with respect to the martingale residuals is a mar-
tingale with known properties, including being centered and uncorrelated. This
can be seen on the residuals plot if the model is valid. The residuals in Schoenfeld
(1982), the weighted ones of Schoenfeld (Lin, 1991), and those in Kay (1977),
belong to this class of residuals. The Schoenfeld ones have an important role in
goodness of fit methods. Recall that these are defined at each time of death t by
rβ (t) = Zj (t) − Eβ (Z|t) , (10.1)
where j is the individual who dies at time t. Grambsch and Therneau (1994)
have suggested plotting the Schoenfeld residuals standardized over time in order
to examine the validity of the proportional hazards model, and if it is rejected,
obtain some idea of the time-dependent effect. Essentially, they have shown that
for each time of death t:

−1
E Vβ̂ (Z|t) rβ̂ (t) + βˆj ≈ βj (t), j = 1, . . . , p,
j

where β̂ = βˆ1 , . . . , βˆp is the partial maximum likelihood estimator of the pro-
portional hazards model and β(t) = (β1 (t), . . . , βp (t)) the true time-dependent
model of the non-proportional hazards model. This approach is used often, and
is implemented in the survival package in R. Furthermore, the authors have
proposed a confidence band for the regression effect, but this is not reliable if the
proportional hazards model is not satisfied. More recently, Sasieni and Winnett
(2003) have suggested using the smoothed residuals of martingale differences.
These can be plotted as a function of time for a fixed value of the covariate, allow-
ing a glimpse at temporal dependency. These can also be plotted as a function
of a covariate at a fixed time, to study how to introduce it into the model. In the
former, several plots need to be output for the various values of the covariates,
10.3. Classical graphical methods 265
and in the latter, for various times. In practice, it may be difficult to interpret
all of these. Also, these methods—based on non-cumulative residuals—require
smoothing functions to join up the residuals. As underlined by Lin et al. (1993),
results are often sensitive to the choice of smoothing technique employed. Dif-
ferent smoothings can lead to very different conclusions. Examining the quality
of proportional hazards models with such methods can be tricky.
To get around these kinds of problems, several authors have proposed using
cumulative martingale residuals. Arjas (1988) suggests plotting the expected
number of deaths under the proportional hazards model as a function of the ranks
of the times of death. The model fits the data well if the resulting curve is close to
the diagonal. A score process, inspired by the goodness of fit test of Wei (1984),
was introduced by Therneau et al. (1990) and is denoted {U (β, t), 0 ≤ t ≤ T },
where
n t
U (β, t) = Zi (s) − Eβ (Z|s) dNi (s), 0≤t≤T. (10.2)
i=1 0
Under the proportional hazards model, a standardized version of {U (β̂, t),

0 ≤ t ≤ T } converges to a Brownian bridge. A test statistic corresponding to
the supremum of the absolute value of the standardized process can then be
applied. The limit distribution of the statistic is the Kolmogorov distribution.
Wei (1984) proposed this test for the binary covariate case and Therneau et al.
(1990) extended it to a time-dependent covariate. In the presence of correlated
covariates, the limit distribution of the statistic is no longer a Brownian bridge.
Lin et al. (1993) showed that the non-standardized score process (10.2) converges
to a centered Gaussian variable whose variance-covariance matrix depends on the
data. Hence, a comparison of the observed process over time with a large number
of processes simulated from the Gaussian limit distribution can give an idea of
the asymptotic distribution of the process’s supremum. In practice, interpretation
of such plots is not always easy, notably because the envelope of the process is
not deterministic and contains noise.
The classic case of linear regression with Gaussian noise and covariates cor-
responds to a very special kind of regression model. Any sub-model, i.e., all
conditional and marginal models, are also linear and the residual noise remains
Gaussian albeit with a potentially different variance. This type of model structure
is, however, just about unique, and with very few conditions, can even be taken as
a characterization of the multinormal model. For non-linear models, and specifi-
cally the proportional hazards models with several covariates, the sub-models will
not belong to the same class. Thus, evaluating the goodness of fit of a multivari-
ate proportional hazards models by looking individually at the univariate models
is not really addressing the relevant question. In fact, in the absence of available
tools for checking the overall validity of the model, most methods for testing the
proportional hazards hypothesis with one covariate suppose that it is true for the
others, an assumption that is generally false (Scheike and Martinussen, 2004).
Many methods depend on the covariance between the covariates, such as the
ones proposed by Grambsch and Therneau (1994) and Lin et al. (1993). To deal
with this problem, Scheike and Martinussen (2004), looking at a non-proportional
hazards model, developed estimation techniques and goodness of fit tests for the
proportional hazards model for each covariate, while allowing for the possibility
that other covariates have time-dependent effects. Their simulations show that the
method works well compared to the often-used ones of Grambsch and Therneau
(1994) and Lin et al. (1993) when the proportional hazards hypothesis is invalid
and/or there are correlated covariates. Their test statistics depend on the estima-
tion of the regression parameter, which requires a hard-to-understand algorithm
based on kernel smoothing. Also, the form of the regression parameter estimator
is not a smooth and explicit function of time. The goodness of fit tests are of the
Kolmogorov-Smirnov and Cramér-von Mises type and the p-values are calculated
by simulation since the limiting distributions of the statistics are not known.
For more details on these less recent goodness of fit tests for proportional
hazards models, the reader may wish to consult the works by Therneau and
Grambsch (2000), Klein and Moeschberger (2003) and Martinussen and Scheike
(2006). Our preference as outlined below is to make full use of known and estab-
lished properties of the regression effect process which, depending on the cir-
cumstances, we will approximate by a Brownian motion or a Brownian bridge.
We make use of Theorem 9.2 indicating how a correctly specified model corre-
sponds to a Brownian motion with linear drift. This theorem, together with an
elementary conditioning argument, leads to
Corollary 10.1. Under the non-proportional hazards model with parameter

β(t),
L
Un∗ (β(t), t) − t × Un∗ (β(t), 1) −→ W 0 (t),
n→+∞
where W 0 (t) , t ∈ (0, 1), is the Brownian bridge.
As a corollary, the result is slightly misplaced in the text, and if we were guided
only by logical flow, then it would follow immediately from Theorem 9.2. We
present it here because this is where we need it, as a key support to assessing
goodness of fit. It tells us that the above elementary transformation will look
like a Brownian bridge in practice, when our chosen model can be considered
well specified and one that could plausibly have generated the observations.
We will see this in later real examples, but as an appetizer, the reader might
consider Figure 10.1 where—leaving aside any formal testing procedure—the
visual impression is enough to tell us that, in this case, a non-proportional hazards
model with time-dependent regression effects provides a much better fit to the
observations than does the simple proportional hazards model. In a multivariate
setting we can obtain several such graphs; examples being the process for the
prognostic index, the process for particular covariates where others are kept fixed
and processes for stratified models.
10.4. Confidence bands for regression effect process 267
Figure 10.1: Dotted lines show upper 5% limit of the supremum of a Brownian
bridge process. Left-hand process for PH model encroaches on this limit. Right-
hand process for the NPH model with time-dependent effects shows a much
improved fit. The process appears to be not unlike a Brownian bridge
10.4 Confidence bands for regression effect process
The regression effect process provides us with insight into both the strength
of effect as well as the fit of the assumed model. No less importantly, when
the mechanism generating the observations differs from the assumed model,
the regression effect process will indicate to us in which direction to look for a
better fitting model. Here, we consider confidence bands for the process. These
offer some help in seeing whether the model’s assumed functional form for the
parameter can be deemed sufficiently plausible.
Let us consider the non-proportional hazards model with regression coefficient
β(t) = (β1 (t), . . . , βp (t)), and suppose that conditions B1, B2, B3 and B4 of
Chapter 9 hold. We suppose furthermore than the sequence (Mn )n in hypothesis
√
B3 is such that nMn → 0 as n → ∞. For i = 1, . . . , p, consider the following
null and alternative hypotheses:
H0,i : ∃ bi , ∀ t, βi (t) = bi , H1,i : bi , ∀ t, βi (t) = bi .
Under the hypothesis H0,i , the covariate Z (i) is consistent with the proportional
hazards model, while under H1,i , its coefficient changes over time. The propo-
sition below allows us to define the limit behavior of the ith component of the
regression effect process under H0,i . This then allows us to construct confidence
bands for the constant nature of a parameter around the corresponding compo-
nent of the process, as well as a goodness of fit test for the proportional hazards
model. Recall that the limit distribution of the supremum of the absolute value
of a Brownian bridge is the Kolmogorov distribution. For α = 5%, the upper
quantile of order α of this distribution is a(α) = 1.358.
Corollary 10.2. (Confidence bands). Let i ∈ {1, . . . , p} and β0 ∈ B . Under

H0,i , i.e., under the proportional hazards model with a constant parameter
b not necessarily equal to β0 , and denoting a(α) the upper quantile of order
α of the Kolmogorov distribution, we have:

lim P ∀ t ∈ [0, 1], Σ̂−1/2 Un∗ (β0 , t) ∈ BCi (α) = 1 − α,
n→+∞ i
where BCi (α) is a level 1 − α confidence band given by:

−1/2 −1/2
t Σ̂−1/2 Un∗ (β0 , 1) − Σ̂·,i a(α); t Σ̂−1/2 Un∗ (β0 , 1) + Σ̂·,i a(α) .
i 2 i 2
Corollary 10.3. (Goodness of fit test). Let i ∈ {1, . . . , p} and β0 ∈ B . Under

the non-proportional hazards model with parameter β(t) not necessarily
equal to β0 , the hypothesis H0,i is rejected at asymptotic level α if:

−1/2 −1
Σ̂·,i sup Σ̂−1/2 {Un∗ (β0 , t) − tUn∗ (β0 , 1)} ≥ a(α),
2 t∈[0,1] i
where a(α) is the order α upper quantile of the Kolmogorov distribution.

Denoting K a random variable from this distribution, the test’s asymptotic
p-value is:

−1/2 −1 −1/2 ∗ ∗
P K ≥ Σ̂·,i sup Σ̂ {Un (β0 , t) − tUn (β0 , 1)} .
2 t∈[0,1] i
If the ith component of the process Σ̂−1/2 Un∗ (β0 , ·) exits the confidence band
IC(α)i or the p-value is less than α, the hypothesis H0,i that the effect βi (t)
is constant is rejected at level α. However, by simultaneously testing several
hypotheses H0,i for different covariates, we see inflation in the overall level of
the test. This means that a process can exit the confidence band even if the
corresponding effect is constant over time, with an overall level greater than α
(Figure 10.2).
This could be a problem if our definitive conclusion on the goodness of fit
of the proportional hazards model is based on these confidence bands or the
goodness of fit test. This is not the case however since such formal testing does
not have any central role in model building. Confidence bands are just one tool
in model construction. The non-detection of a constant effect may be corrected
10.5. Structured tests for time dependency 269
Figure 10.2: Process and confidence intervals under PH assumption. Left-hand

figure shows clear evidence of non proportionality whereas right-hand figure, for
a fitted piecewise model, appears to provide an adequate fit.
in later model selection steps respecting the proportional hazards hypothesis, as

described in the next section. Note that the proportional hazards hypothesis is
evaluated for each of the covariates, taking into account the correlation between
them and allowing for the possibility of other time-dependent effects. Simulations
comparing the performance of this goodness of fit test for proportional hazards
models with other common tests are presented in Section 10.9.
10.5 Structured tests for time dependency
In his article introducing the proportional hazards model, Cox (1972) suggested
replacing the constant regression coefficient by a non-constant one, then testing
this postulated time dependency. Formally, this amounts to considering the class
of tests based on the non-proportional hazards model with parameter
β(t) = β0 + β1 g(t),
where β0 and β1 are constant parameters and g is a function of time. This

means testing the null hypothesis concerning the validity of the proportional
hazards model H0 : β1 = 0 against the alternative H1 : β1 = 0 (Grambsch and
Therneau, 1994; Therneau and Grambsch, 2000). A wide range of g are possible
(O’Quigley, 2008). If g is a known function of time, this reverts to running a score
test as proposed in Cox (1972). The function g can also be a predictable process.
For example, g(t) = N̄ (t− ) corresponds to the test in Breslow et al. (1984) based
on the ranked times of events. One last example is a piecewise constant function
g with pre-defined change-points (Moreau et al., 1985; O’Quigley and Pessione,

1989; Schoenfeld, 1980).
Wei (1984) has developed a goodness of fit test for one binomial variable, with
test statistic supt |U (β̂, t)|, where β̂ is the maximum partial likelihood estimator.
This corresponds to applying the test for the supremum of a Brownian bridge,
and the asymptotic distribution of the statistic is the Kolmogorov one. This test
has inspired several graphical methods, which were mentioned in the previous
section (Lin et al., 1993, 1996; Therneau et al., 1990). Schoenfeld (1980) has
developed a χ2 test which partitions the covariate and time space and compares
the observed values in the various partitions to the expected ones under the
model. Lin (1991) has proposed a score test based on a weighted maximum log-
likelihood estimator with regression coefficient β, noting that under this model,
the limiting distribution of the difference between this estimator and the usual
partial maximum likelihood one is a centered normal distribution.
In addition, Murphy and Sen (1991) and Gray (1992) have proposed goodness
of fit tests for the proportional hazards model, modeling the regression coefficient
β(t) in a non-proportional hazards model by projection onto basis functions, then
testing whether the projections are time-dependent.
The collection of goodness of fit procedures that we recommend is based
on the results in Theorems 9.2 and 9.3 which indicate that the form of time
dependency for the regression coefficient expresses itself via the regression effect
process. Thus, testing the goodness of fit of proportional hazards models corre-
sponds to examining the linearity of the process’s drift. Before looking at this
aspect more closely, let us first consider the notion of explained variance in the
non-proportional hazards model, as well as its estimator, the R2 coefficient which
allows us to measure the predictive strength of any model. Again, all of the predic-
tive information is contained in the regression effect process and so, structuring
model building around a careful study of this process allows us to address all of
the key questions that arise.
10.6 Predictive ability of a regression model
Explained variance
Let X be a random variable whose second-order moment exists. The Bienaymé-
Chebyshev inequality:
∀ ε > 0, P (|X − E (X)| ≥ ε) ≤ ε−2 Var (X)
says that the spread of X around its expectation can be bounded in terms of its
variance. The smaller the variance, the greater the probability that X is closer, on
average, to its expectation. Using a statistical model, the idea of evaluating the
proximity of a response variable to its predicted value, in terms of the variance,
would appear natural. The smaller the variance, the better the quality of the
10.6. Predictive ability of a regression model 271
prediction. For random variables X and Y whose second-order moments exist,

the variance can be always expanded as:
Var (Y ) = Var (E (Y |X)) + E (Var (Y |X)), (10.3)
which then allows us to define the explained variance parameter. Suppose that
we want to model a real-valued response variable Y with the help of a vector of
explanatory variables X. If the chosen model and the covariate information, X,
allow a good characterization of Y , then the expected values of Y given X will
show relatively large dispersion when compared to the simple imprecision (noise)
associated with any given fixed value of X. Whenever Var(E(Y |X)) is large
compared to the average of Var(Y |X), then prediction becomes easy. Conversely,
whenever the opposite holds, then the residual noise will mask to a greater or
lesser extent any underlying signal. It makes sense to describe Var(E(Y |X))
as the signal and E(Var(Y |X)) as the noise, so that we can write: Var(Y ) =
signal + noise. We can formally introduce a parameter that can be viewed as the
explained variation in Y given X via,
Definition 10.1. The proportion of the variance of Y explained by X and

the model, denoted Ω2Y |X , is:
Var (E (Y |X)) Var (Y ) − E Var (Y |X) signal

Ω2Y |X = = = . (10.4)
Var (Y ) Var (Y ) signal + noise
This Ω2Y |X is commonly known as the explained variance, and corresponds to

the ratio of the variance of the expected values of the response variable given the
variables X and the model, over the marginal variance of the response variable.
In a non-proportional hazards model, two coefficients can be defined, depending
on whether T is the response variable and Z the explanatory variable, or vice
versa.
The explained variance of T given Z.

Replacing Y by T and X by Z in formula (10.4), we could define the explained
variance of T given Z by Ω2T |Z = Var E(T |Z)/Var (T ). Intuitively, this seems
to be what we should use to represent the predictive ability of the proportional
hazards model. Indeed, when we construct a survival model using data, one of
our goals is to predict an individual’s survival, taking into account the values of
some of their characteristics given in the vector Z. We would therefore like to
calculate the predictive power of Z on T . The following result throws some light
on why this coefficient may not be the most useful measure of predictive ability
in the proportional hazards model context.
Proposition 10.1. Under the proportional hazards model with one covariate
Z ∈ R and a constant risk λ0 (t) = a, a > 0, we have Ω2T |Z ≤ 1/2.
If the regression coefficient β in the proportional hazards model is large, this

means that the association between T and Z is strong and the probability that
an individual dies earlier or later will be strongly influenced by the value of Z. As
this value influences survival to a greater extent as the value of the regression
coefficient increases (in absolute value), the predictive power of Z on T should
increase when |β| does, and not be bounded by a value strictly less than 1.
Proposition 10.1 prevents this happening.
Furthermore, using formula (10.14), we can show that if Z is Bernoulli with
parameter 1/2, then Ω2T |Z is bounded above by 1/3, and that this upper bound
changes as the Bernoulli parameter does. Thus, the comparison of two values of
Ω2T |Z for different marginal distributions of Z is difficult since the coefficients are
not on the same scale. This is illustrated in Figure 10.3 with conditional survival
curves under the proportional hazards model with β = 1.5 or β = 4, λ0 (t) = 1, and
a binary covariate Z. Notice that the conditional survival curves are further away
from each other when β is larger. This should improve the prediction quality since
the covariate is more strongly linked to survival. However, using equation (10.14),
we can show that for β = 1.5 and Z ∼ B(0.5), and for β = 4 and Z ∼ B(0.3),
the value of the explained variance is the same: Ω2T |Z = 0.22.
Figure 10.3: Conditional survival curves S(·|Z) as a function of time under the
proportional hazards model with λ0 (t) = 1, β = 1.5 or 4, and binary covariates.
Also, O’Quigley and Flandre (1994) have shown that a non-linear trans-
form which changes the time axis but keeps the order of deaths the same alters
the explained variance of T given Z. For proportional hazards models however,
changing the time scale in this way does not change the estimation of the regres-
sion coefficient β. The predictive ability of the model should not change. The
explained variance of T given Z would therefore seem to be a poor indicator of
10.6. Predictive ability of a regression model 273
the predictive power of variables in proportional hazards models, and even less so
in non-proportional ones. Let us now look at the explained variance of Z given T .
The explained variance of Z given T

Recall that Z = Z(t) is a vector containing p possibly time-dependent covariates.
The explained variance of Z given T under the non-proportional hazards model is
not defined in the same way for the univariate case (p = 1) as for the multivariate
(p > 1) one. In what follows, we will only look at the explained variance based
on the distribution of Z conditional on T , and to simplify notation, we will not
always specify the conditional nature of this distribution.
Definition 10.2. In the univariate non-proportional hazards model, the

explained variance, given as a function of the time-dependent regression
coefficient β(t), is defined as:
Var (Eβ(t) (Z|T = t)) E (Varβ(t) (Z|T = t))

Ω2 (β(t)) = = 1− . (10.5)
Var (Z) Var (Z)
When Z(t) is a vector containing p > 1 covariates, at time t an individual i is

characterized by its real-valued prognostic index ηi (t) = β(t)T Zi (t), which is a
realization of the random variable η(t) = β(t)T Z(t). Hence, the model’s predic-
tive ability can equally be calculated using either Z or η (Altman and Ander-
sen, 1986; Andersen et al., 1983). Note also that any time dependency in β(t)
becomes very quickly intractable when working with the conditional distribution
of T given Z whereas, for the conditional distribution of Z(t) given T = t, no
additional concerns are raised.
Definition 10.3. The explained variance of the non-proportional hazards model

with more than one covariate can be defined as a function of β:
Var(E(η|T )) E(Var(η|T ))
Ω2 (β(t)) = = 1− , η(t) = β(t)T Z(t). (10.6)
Var(η) Var(η)
We can define the expectation with respect to the distribution of T of the

variance of the covariate or the prognostic index evaluated with the coefficient
α2 (t) under the non-proportional hazards model with p-dimensional parameter
α1 (t). In most cases we will have, α2 (t) = α1 (t) = β̂(t) and the extra flexibility is
not needed. However, for testing some null hypotheses and for building particular
tests having certain optimality properties it is useful to have this added generality.
With this in mind, we write:
1
2
Q F, α1 (t), α2 (t) = Eα1 (t) Z(t) − Eα2 (t) (Z|T = t) T = t dF (t)
0
(10.7)
where F is the cumulative distribution function of T . It is always easier to think

in terms of a univariate “response” variable, Z(t), as in the above expression. In
practice we will often be focused on multivariate outcomes and this can come in
one of two forms. The first is again just a univariate “response” where the other
variables have been accounted for via the model. The second form is where we
summarize the vector of outcome variables as a single variable, in particular a
linear sum, α2 (t)T Z(t) and, usually, the prognostic index itself so that α2 (t) =
β̂(t). We then have:
Ω2 (β(t)) = 1 − Q−1 (F, 0, β(t)) Q(F, β(t), β(t)).
This way of writing the explained variance will be used later to construct an
estimator. The difficulties we had with the explained variance based on the dis-
tribution of T given Z in the previous section disappear when considering the
explained variance of Z conditional on T , thanks to the following result.
Proposition 10.2. Under the non-proportional hazards model with param-

eter β(t), and for a p-dimensional covariate vector (p ∈ N∗ ), the following
three statements hold:
1. Ω2 (0) = 0 and 0 ≤ Ω2Z|T (β(t)) ≤ 1,
2. Ω2 is invariant for transformations that are strictly increasing in T as

well as for transformations that are linear in Z,
3. If β(t) = β0 is constant and p = 1, Ω2 is an increasing function of |β0 |

and lim Ω2 (β) = 1.
|β|→∞
Proof of Proposition 10.2. In Chapter 13 of O’Quigley (2008) and in O’Quigley

and Xu (2001), Ω2 is a function of a constant regression parameter, correspond-
ing to the proportional hazards model. Extending this to the time-dependent
parameter case is trivial. For the multivariate case, see Xu (1996).
Generalizing the third statement above requires a little thought. Since β(t) is no
longer constant it may not be immediately clear what we mean when we talk of
increasing β. The following is however true. For any model based on β(t)Z, there
exists an equivalent model based on β ∗ Z ∗ (t) in which the time dependency has
now been absorbed by Z ∗ (t). This is a purely formal manipulation but allows us to
place any non-proportional hazards model under a proportional hazards heading.
As a result we see that, given Z ∗ (t), it is true that Ω2 is an increasing function
of |β ∗ |. What we have is that part 3 of Proposition 10.2 holds in a particular
direction of the functional coefficient space, the direction that remains within
the functional form of the model. The explained variance coefficient remains
unchanged when applying strictly increasing transformations in time. Only the
ranks matter. In consequence, we can continue to work with the standardized
10.7. The R2 estimate of Ω2 275
time scale described in Chapter 9. We view the explained variance as a population

parameter that, typically, needs to be estimated.
10.7 The R2 estimate of Ω2
Numerous measures of the predictive ability of proportional hazards models have

been proposed in the literature (Choodari-Oskooei et al., 2012). Some of these are
called “R2 coefficients”, which classically refer to estimators of Ω2 or “explained
variance”, without actually being estimators of the explained variance as defined
in Section 10.6. Most of these measures are entirely ad hoc and constructed with-
out any theoretical justifications. Furthermore, they are almost always defined for
proportional hazards models, but for the most part their behavior in the presence
of poorly-specified models has not been studied. Corresponding measures for the
situation of non-proportional hazards are not usually developed and it is not clear
how any of these measures work in the situation of non-proportional hazards. For
example, when considering rival candidate non-proportional hazards models, we
would like the model that is closer to the mechanism generating the observations
to result in a greater value of R2 . It is not our goal here to thoroughly review
all such estimators. Instead, we will concentrate on a specific estimator of Ω2 :
an R2 coefficient constructed with the same residuals as the regression effect
process, introduced by O’Quigley and Flandre (1994).
After defining Ω2 , we will extend the definition to non-proportional hazards
models. We also show Ω2 to have all of the desirable properties that we would
hope for such a measure. Some simple but strong results indicate why the R2
coefficient, a consistent estimator of Ω2 , based on the regression effect process
turns out to be a fundamental tool in model construction. It guides us away from
poorly fitting models toward better fitting models, and ultimately, will assume
its greatest value when the chosen model coincides with the mechanism that can
be viewed as having truly generated the observations.
In order to obtain our needed theorems we extend the Schoenfeld residuals,
defined in Equation (7.43), to the non-proportional hazards model with parameter
β(t). Furthermore, we keep in mind that this is carried out on the transformed
scale of the previous chapter so that:
rβ(t) (t) = Z(t) − Eβ(t) (Z|t), 0 ≤ t ≤ 1,
where Z(t) is the value of the covariate of the individual who has died at t and
Eβ(t) (Z|t) its expectation under the model. Let us denote F̃ the estimator of the
empirical cumulative distribution function of T on the transformed scale:
n
1 ∗ 1
F̃ (t) = N̄ (t) = 1φ (X ) ≤ t, δ = 1 , 0 ≤ t ≤ 1.
kn kn n i i
i=1
In the presence of uncensored data, F̃ corresponds to the usual estimator of the

empirical cumulative distribution function of T on the transformed scale. The
value Q(F, α1 (t), α2 (t)) in equation (10.7) can be estimated in the univariate
case by:
1 kn
2 1 2
Q̃(F̃ , α1 (t), α2 (t)) = rα1 (s) (s) dF̃ (s) = rα1 (ti ) (ti ) . (10.8)
0 kn
i=1
where, although no role is being played by α2 (t), we keep it as an argument to

Q̃(F̃ , α1 (t), α2 (t)) in order to keep the notation simple. The reason for this is
that, in the multivariate case, where we work with the same equation but replace
rα1 (s) (s) by α2 (s)T rα1 (s) (s), we can put both cases under the same heading.
We now have the ingredients we need for the main definition of this section.
Definition 10.4. The R2 is defined as R2 = R2 (β̂(t)), where β̂(t) is a con-

vergent estimator of β(t), and for any function α from [0, 1] to Rp ,
R2 (α(t)) = 1 − Q̃−1 (F̃ , 0, α(t)) Q̃(F̃ , α(t), α(t)) (10.9)
With one covariate in the model, the R2 coefficient is the ratio of the sum of the
squared Schoenfeld residuals for the model with parameter β̂(t), over the sum of
the squared residuals for the model whose parameter is zero. When instead the
model has several covariates, the coefficient is still the ratio of sums of squared
residuals, but the residuals are now evaluated using the prognostic indicator
β̂(t)T Z(t) rather than the covariate Z(t) directly. A quantity very closely related
to R2 (α(t)) is RS2 (α(t)) defined as:
2
RS (α(t)) = 1 − Q̂−1 (F̂ , 0, α(t)) Q̃(F̂ , α(t), α(t)) , (10.10)
the only change from Equation 10.9 being that F̃ is replaced by F̂ . In practical
work this is only very rarely of interest and we can safely ignore RS 2 (α(t)) and
2 2
only keep in mind R (α(t)). The purpose of RS (α(t)) is for theoretical study. In
the absence of censoring RS 2 (α(t)) and R2 (α(t)) coincide, but when censoring is
present, we will need work with RS 2 (α(t)) in order to demonstrate large sample
consistency. The explanation for this is that the estimator F̃ does not converge
to the cumulative distribution function of T. As a consequence the R2 coefficient
will not generally converge to the explained variance Ω2 . Noting that dF̂ = −dŜ
then −1
dŜ(ti ) = Ŝ φ−1
n (ti ) − Ŝ φn (ti )
−
is the jump in the estimator at time of death ti for i = 1, . . . , kn (Equation 3.13).

This weighted modification to the R2 coefficient with jumps from the Kaplan-
Meier estimator allows us to obtain convergence of RŜ 2 (β̂(t)) to the proportion
of explained variance Ω2 when β̂(t) is a convergent estimator of β(t) (O’Quigley

10.7. The R2 estimate of Ω2 277
and Xu, 2001; Xu, 1996). However, in real-world settings, it has been observed
that R2 depends very weakly on the censoring, even for rates as high as 90% and
that, in the great majority of applications, the standard deviation of R2 will be an
order of magnitude greater than any bias. Simulation work by Choodari-Oskooei
et al. (2012) has confirmed the independence of the unweighted R2 coefficient
with respect to censoring, and a higher variance for RŜ 2 in return for its lack of
bias. For these reasons, in our work we mostly work with the unweighted R2
coefficient in real applications.
The coefficient of explained variance Ω2 (β(t)) can be estimated with R2 (β̂(t)),
where β̂(t) is a consistent estimator of the true regression parameter β(t). The
following theorem justifies the use of the R2 coefficient when evaluating the
goodness of fit of the non-proportional hazards model, and gives its asymptotic
behavior under poorly-specified models.
Theorem 10.1. Under the non-proportional hazards model with parameter

β(t), if p = 1 and conditions A1, A2, A3 and A4 of Section 9.4 hold, then
there exists a constant C(β) > 0 such that for all α ∈ B:
1
2 C(β) + 0 (e(α(t), t) − e(β(t), t))2 dt
lim R (α(t)) = 1 − 1 , (10.11)
n→∞ C(β) + 0 (e(0, t) − e(β(t), t))2 dt
and
arg max lim R2 (b(t)) = β a.s.
b∈B n→+∞
If p > 1 and conditions B1, B2, B3 and B4 of Section 9.5 hold, for all α ∈ B
with α = 0, we have:
1 1 2
2 0 α(t)T Σα(t)dt + 0 α(t)T {e(β(t), t) − e(α(t), t)} dt
lim R (α(t)) = 1 − 1 1 2
.
n→∞ α(t)T Σα(t)dt + (α(t)T {e(β(t), t) − e(0, t)}) dt
0 0
For t ∈ [0, 1] and γ a function from [0, 1] to Rp , recall the definition:
e(γ(t), t) = s(1) (γ(t), t)/s(0) (0, t).
Theorem 10.1 says that in the univariate case, if β(t) is the true regression
coefficient and if the sample size is large enough, the maximum of the R2 function
is reached for the true parameter β(t). Generalizing this to the multivariate case
is not a simple task. Consider the two covariate case Z (1) and Z (2) with effects
β1 (t) and β2 (t). If we suppose that β2 (t) is known, this becomes equivalent to
estimating β1 (t) in the univariate model. Theorem 10.1 then applies and β1 (t)
reaches the maximum for R2 in the limit. In practice, β2 (t) is not known but
we can get close to it using a convergent estimator. By conditioning on this
estimator, i.e., treating β2 (t) as a known constant, the result again holds for
β1 (t). We can therefore consider than the result of Theorem 10.1 applies for
each variable taken separately, conditional on convergent estimators for all of the
other regression coefficients.
The R2 coefficient and the regression effect process are constructed using the
same residuals. The regression effect process allows us to check the goodness of
fit of the non-proportional hazards model, while the R2 coefficient is a measure of
the model’s predictive ability. Though these aspects are different, their construc-
tion using the same quantities is natural. This means that in essence all of the
relevant information regarding the predictive power of a model as well as model
adequacy is contained in the regression effect process. It is natural to expect that
introducing additional covariates into any model will result in some increase, if
only an apparent one, in predictive power. Here, we see that for any given set of
covariables, improvements in fit will result in improvements in predictive power.
Formal theorems enable us to safely rely on this result as a means to guide model
construction.
10.8 Using R2 and fit to build models
Using the results of Theorem 9.3 and its corollary, the regression effect process Un∗
can be used to determine the form of the multivariate regression coefficient β(t).
The process’s drift will mirror the form of β(t). No other techniques like local
smoothing, projection onto basis functions, or kernel estimation are necessary
(Cai and Sun, 2003; Hastie and Tibshirani, 1990; Scheike and Martinussen, 2004).
For example, as illustrated in Figure 9.4a, an effect which is constant until time
τ , followed by a zero-valued effect is easy to spot, even for small sample sizes.
This simple notion generalizes immediately. Suppose that the p components of
a time-dependent regression parameter
β(t) = (β1 (t), . . . , βp (t))
can be written βj (t) = β0,j hj (t) for j = 1, . . . , p, with
• β0 = (β0,1 , . . . , β0,p ) ∈ Rp an unknown constant (over time) regression

parameter,
• h = (h1 , . . . , hp ) a known function from [0, 1] to Rp .
Chauvel (2014) obtains a convergent estimator of β(t) as β(t) = (β p (t))

1 (t), . . . , β
j (t) = β
with β
0,j hj (t) (j = 1, . . . , p), where β0 = (β0,1 , . . . , β0,p ) is the maximum
partial likelihood estimator of β0 . This estimate can be obtained by transferring
the temporal part h of the regression coefficient into the covariate (Cox, 1972).
This function h can be determined graphically using the regression effect process
(see the examples in Section 10.9). For example, if the jth component of the
10.8. Using R2 and fit to build models 279
drift is a linear function, then the function hj is constant. If the drift is concave,
hj is decreasing, and if convex, increasing. Also, the confidence bands defined in
Section 10.4 can help to evaluate the plausibility of a constant effect over time
for βj (t), resulting in a constant function hj for j = 1, . . . , p.
In the presence of non-proportional hazards, we need a more complex strategy,
one that takes on board both the predictive strength of the covariates as well as
the accuracy of the modeling of the time dependency. We propose using the R2
coefficient which—on top of indicating the predictive abilities of the model—also
converges to a function whose maximum is attained when the true function h is
chosen (Theorem 10.1). When several models provide plausible suggestions for h,
the estimators β̂(t) and corresponding R2 coefficients should be evaluated. The
time-dependent function h retained is the one that maximizes the R2 . In this
way, we obtain a non-proportional hazards model with a good fit and maximal
predictive ability. The measure of predictive ability is maximized over the set
B of m preselected time-dependent regression effects. More formally, denote
B = {β1 (t), . . . , βm (t)} the set of m functions from [0, 1] to Rp . The regression
coefficient β ∗ (t) selected is the one for which
β ∗ (t) = arg max R2 (α(t)) .

α(t)∈B
Note that the set B may be arbitrarily large and not even limited to countable
sets, albeit such a level of generalization would not bring any practical benefit.
The following theorem shows the equivalence between this maximization problem
when n → ∞ and one in which we minimize an L2 norm.
Theorem 10.2. (Chauvel and O’Quigley 2017). Let p = 1. Suppose that

conditions A1, A2, A3 and A4 hold, β(t) ∈ B and B ⊂ B. Under the non-
proportional hazards model with regression parameter β(t), where β(t) is
not necessarily in B, the coefficient
β ∗ (t) = arg max lim R2 (α(t))

α(t)∈B n→∞
is a solution to
β ∗ (t) = arg min β(t) − α(t)2 ,
α(t)∈B
1
where for α(t) ∈ B, β(t) − α(t)22 = 0 (β(t) − α(t))2 dt.
In other words, for a large enough sample size, selecting the regression coef-
ficient which maximizes the R2 is the same as choosing the coefficient closest to
the true regression function in terms of L2 distance. It is possible that the true
regression coefficient, if it exists, is not in the set B. We can, however, make it
arbitrarily close. In any event, models are chosen to represent data either because
they fit well, or for their predictive performance. Many models may correspond to
one or both of these aspects, not just the “true model”. Here, priority is given to
goodness of fit, with the selection of candidate regression coefficients, and only
then is a model’s predictive performance considered. We have chosen to work
with the R2 coefficient, but any measure of predictive performance that satisfies
Theorem 10.1 could be considered in principle.
When the trend in the process’s drift is a concave function, the effect
decreases with time, and if it is convex, the effect increases. In order the get the
largest possible R2 , it would be possible to construct a time-dependent effect
which is closer and closer to the observed drift of the process, like for example
a piecewise constant effect with many change-points. In general, this will lead
to overfitting, and interpretation of the coefficient will not be straightforward.
A compromise needs to be made between high predictive ability and simplic-
ity of the coefficient, notably for interpretation purposes. This is comparable to
the linear model situation in which the estimated explained variance—the R2
coefficient—is positively biased and increasingly so as the model’s dimension
increases. A balance needs to be found between the goal of improving predictive
strength and the danger of poor prediction due to overfitting. This boils down
to respecting the elementary principles of parsimonious model building.
10.9 Some simulated situations
Chauvel and O’Quigley (2017) present two distinct simulated situations. The
first looks at the goodness of fit of proportional hazards models, and the second
illustrates methods for constructing non-proportional hazards models with the
help of the regression effect process and the R2 coefficient. We generate samples
under the non-proportional hazards model as described in the appendix. Let us
consider the behavior of the goodness of fit test for the proportional hazards
model presented in Section 10.4. This will be compared to standard tests which
are also based on graphical methods, allowing visualization of the form of the
time-dependent effect in the case of poor model fitting.
The first comparison test—proposed by Lin et al. (1993)—is based on the
standadized score process defined in equation (10.2). This is different with the
regression effect process defined and studied here because the increments of the
process of Lin et al. (1993) are non-standardized Schoenfeld residuals summed
with respect to times of death in the initial time scale, and an overall standard-
ization is applied to all increments. The test statistic is the supremum over time
of the absolute value of the globally standardized process. If the covariates are
non-correlated, the statistic’s limit distribution is the Kolmogorov one (Therneau
et al., 1990). If they are correlated, the limit distribution is not known but can be
evaluated numerically using Monte Carlo simulation. Thus, the confidence enve-
lope of the process and the goodness of fit test are obtained by simulating N
processes according to the process’s Gaussian limit distribution. We have chosen
N = 103 here. The R code for evaluating the process, the simulated envelope,
10.9. Some simulated situations 281
and the goodness of fit test can be found in the timereg package (Martinussen
and Scheike, 2006). Below we refer to this test with the abbreviation LWY.
A second goodness of fit test for the proportional hazards model is due to
Grambsch and Therneau (1994). They consider an effect of the form β(t) =
β0 + β1 log(t) and propose a score test for the null hypothesis H0 : β1 = 0 versus
H1 : β1 = 0. This test can be found in the survival package in R. We label this
test GT in the following.
The final goodness of fit test compared here was proposed by Scheike and
Martinussen (2004). Under the non-proportional hazards model, they propose a
Cramér-von Mises-type test based on the cumulative value of the estimator of
β(t). This estimate is made using an algorithm involving kernel smoothing. Code
for this test can be found in the timereg package in R. This test is labeled SM
below. These authors also proposed a Kolmogorov-Smirnov test, but based on
their simulations it appears less powerful, so we have not retained it here. The
test based on the regression effect process introduced in Section 10.4 is denoted
Un∗ in the tables taken from Chauvel (2014), Chauvel and O’Quigley (2017).
Simple two covariate example

Data were simulated under the proportional hazards model with two centered
normal covariates Z (1) and Z (2) with variance 2 and correlation ρ(Z (1) , Z (2) )
of 0, 0.3, 0.5 or 0.7. The sample size was either n = 50, n = 100 or n = 200.
The mean rate of censoring wa 25%: data were censored at the time 1.5 on the
initial scale. The non-proportional hazards model is

λ(t|Z (1) , Z (2) ) = exp 0.5Z (1) + β2 (t)Z (2) ,
where β2 (t) takes the values 0, 0.5, 1t≤0.5 , 1.5 1t≤0.5 or 0.51t≤0.5 − 0.51t≥0.5
on the transformed time scale. The covariate Z (1) respects the proportional
hazards hypothesis. In the first two cases, Z (2) does too. In the next two, the
hypothesis is not satisfied because the effect is constant at the start and becomes
zero when half of the deaths have been observed. In the final case, the hypothesis
is not satisfied and the type of effect changes half-way through the study (the
variable’s effect on survival is positive at the start and negative at the end). This
situation corresponds to risks which cross over each other. For each trial setting,
3000 samples were simulated in order to evaluate the level and empirical power
of the tests.
Some summary results

Tables 10.1 and 10.2 show the results for the two proportional hazards models.
We see that the level is well-estimated at 5% for each situation and each test,
except for the Scheike and Martinussen test, which has a level above 5% for
small sample sizes (n = 50 and n = 100).
LWY GT SM Un∗
n ρ(Z (1) , Z (2) ) β1 (t) β2 (t) β1 (t) β2 (t) β1 (t) β2 (t) β1 (t) β2 (t)
50 0.0 6.3 5.3 5.5 4.4 7.4 7.0 3.3 2.9
50 0.3 5.4 4.8 5.2 4.6 8.1 7.0 2.8 3.1
50 0.5 5.0 4.7 5.2 4.5 6.8 6.7 3.0 2.6
50 0.7 5.7 6.5 5.0 5.1 7.2 8.8 3.1 3.2
100 0.0 5.4 5.7 4.8 5.0 5.7 6.5 3.4 3.7
100 0.3 5.5 5.2 4.9 4.7 5.9 5.9 3.4 3.6
100 0.5 5.7 5.9 4.5 4.8 5.4 6.6 3.4 3.5
100 0.7 5.4 5.4 5.7 6.1 6.5 6.7 3.9 3.7
200 0.0 4.6 5.6 4.3 5.1 5.0 5.7 3.9 3.9
200 0.3 5.5 5.7 4.9 4.8 5.2 5.4 4.5 4.2
200 0.5 6.2 5.7 4.4 4.9 4.9 5.1 3.8 3.9
200 0.7 5.5 5.2 5.7 5.0 5.3 5.0 4.6 3.8
400 0.0 5.4 5.7 4.7 4.7 4.6 4.7 4.5 4.6
400 0.3 5.6 4.5 5.1 4.4 5.0 5.2 4.4 3.6
400 0.5 4.8 5.6 4.8 5.2 5.3 5.2 4.4 5.2
400 0.7 6.0 5.2 5.1 5.3 5.7 5.9 5.1 5.0
Table 10.1: Empirical level of tests (in %) for β2 (t) = 0. Taken from Chauvel
(2014).
LWY GT SM Un∗
50 0.0 5.3 5.9 4.8 4.9 7.3 6.7 3.0 2.8
50 0.3 6.1 5.4 5.0 4.5 7.8 7.0 3.2 3.0
50 0.5 5.1 5.1 5.0 4.5 7.0 7.2 2.9 2.6
50 0.7 5.4 5.5 5.1 5.0 8.1 7.7 3.1 3.0
100 0.0 5.9 5.3 4.3 4.3 5.6 5.3 3.6 3.3
100 0.3 5.0 5.7 4.9 4.3 5.3 5.0 3.5 3.8
100 0.5 5.4 5.5 4.8 4.9 6.1 6.7 3.5 3.4
100 0.7 5.4 5.8 5.6 5.4 5.4 6.0 3.6 3.6
200 0.0 5.2 5.5 3.9 4.6 5.3 5.7 3.5 3.4
200 0.3 6.0 5.6 4.8 4.1 5.7 4.9 4.1 3.6
200 0.5 5.7 5.8 4.9 4.5 5.3 5.0 4.0 4.1
200 0.7 4.6 4.9 5.1 4.4 5.2 4.7 3.4 4.2
400 0.0 4.7 5.6 4.7 5.6 4.6 5.4 3.9 4.6
400 0.3 4.2 5.0 4.5 4.4 5.2 4.6 4.2 4.5
400 0.5 6.6 5.3 5.8 4.9 5.1 5.4 4.6 5.0
400 0.7 5.7 6.3 5.2 4.9 5.1 5.3 4.9 5.1
Table 10.2: Empirical level of tests (in %) for β2 (t) = 0.5. Taken from Chauvel
(2014)
LWY GT SM Un∗
50 0.0 5.3 14.9 4.4 13.6 7.2 7.8 2.6 11.3
50 0.3 6.3 14.7 5.0 12.8 7.4 7.6 3.3 9.6
50 0.5 6.1 15.1 4.6 11.9 6.8 7.6 2.7 8.3
50 0.7 9.3 16.8 5.4 11.0 6.3 7.4 2.8 7.9
100 0.0 4.6 28.3 4.8 21.2 5.2 9.0 2.4 22.4
100 0.3 6.4 27.9 5.2 20.7 5.8 8.3 4.0 21.7
100 0.5 9.4 30.6 5.0 20.3 5.5 9.1 4.1 19.8
100 0.7 14.6 30.0 5.5 15.1 5.8 7.2 3.4 13.5
200 0.0 5.5 55.5 4.6 37.9 5.4 14.7 3.7 50.4
200 0.3 8.2 57.5 4.7 38.4 4.9 13.2 4.1 50.7
200 0.5 13.9 59.2 5.0 34.0 4.4 12.8 4.6 42.7
200 0.7 29.3 59.9 4.9 23.8 5.3 9.2 3.8 28.3
400 0.0 5.3 89.6 4.7 60.8 4.4 23.4 4.3 87.6
400 0.3 11.5 90.4 5.1 62.5 5.4 25.9 4.2 86.0
400 0.5 27.7 91.2 5.6 56.0 5.1 21.5 4.6 79.6
400 0.7 56.3 92.5 5.5 46.2 4.9 16.8 4.7 63.3
Table 10.3: Empirical level (β1 (t) column) and power (β2 (t) column) of several
tests (in %) for β2 (t) = 1t≤0.5 . Taken from Chauvel (2014).
In the non-proportional hazards setting (Tables 10.3, 10.4 and 10.5), an

increase in correlation between covariates corresponds to an increase in the esti-
mated level of the LWY test. This has already been pointed out by Scheike and
Martinussen (2004), meaning that this test is not necessarily a good choice when
covariates are correlated. Such correlation will occur often in practice. The other
tests do not have this kind of problem. However, in their simulations, Scheike and
Martinussen (2004) do find this type of inflation in the GT test level in the pres-
ence of correlation between covariates. This is likely because the non-constant
effects β2 (t) considered are relatively simple. The GT test supposes that the
coefficient β2 (t) is constant while the test is run for β1 (t), so it must be used
with caution. Note that the power of the GT test is equivalent or inferior to that
of the goodness of fit test based on the regression effect process.
Lastly, the SM test is less powerful than the goodness of fit test based on
the regression effect process in the situations considered by Chauvel (2014).
Further, for smaller sample sizes (n = 50), the SM test provides inadequate
control on the 5% level. In conclusion, the simulations show that the goodness of
fit test based on the regression effect process is the most powerful among those
considered in the situations looked at. This test is not affected by correlation
between covariates, and its level is well-calibrated, even for small sample sizes.
LWY GT SM Un∗
50 0.0 7.0 51.6 5.6 57.4 8.9 53.4 2.9 56.3
50 0.3 5.8 44.9 5.4 53.1 7.4 47.5 3.1 49.5
50 0.5 10.0 41.9 5.9 49.2 8.0 43.1 3.4 41.9
50 0.7 14.6 37.3 5.5 37.2 7.9 31.4 2.3 27.4
100 0.0 7.5 82.2 4.9 85.4 6.0 81.9 3.4 88.0
100 0.3 7.0 80.9 5.5 83.9 6.2 80.0 4.3 86.6
100 0.5 13.9 76.5 5.6 78.6 6.4 72.4 3.8 77.9
100 0.7 26.1 70.4 5.7 63.6 5.4 56.5 2.8 58.8
200 0.0 8.7 99.0 5.3 98.9 5.1 98.6 4.6 99.7
200 0.3 8.4 98.5 6.0 98.9 5.4 97.9 4.4 99.5
200 0.5 22.3 97.5 5.9 97.4 5.1 95.8 4.5 98.6
200 0.7 53.4 96.6 5.7 92.4 5.1 88.3 3.6 92.9
400 0.0 13.7 100.0 5.0 100.0 4.9 100.0 4.6 100.0
400 0.3 11.4 100.0 5.4 100.0 4.7 100.0 4.3 100.0
400 0.5 43.5 100.0 6.1 100.0 4.6 99.9 5.0 100.0
400 0.7 84.5 100.0 5.7 99.6 4.9 99.4 5.0 100.0
tests (in %) for β2 (t) = 1.51t≤0.5 . Taken from Chauvel (2014).
LWY GT SM Un∗
50 0.0 5.4 14.4 4.6 12.6 6.8 7.8 2.8 9.8
50 0.3 5.8 14.9 4.9 13.0 7.0 8.1 2.7 9.8
50 0.5 8.0 15.2 5.4 12.6 7.1 8.4 3.2 9.4
50 0.7 8.4 15.2 5.3 9.8 6.6 6.7 3.0 6.7
100 0.0 6.2 26.8 5.4 21.6 5.9 9.1 3.4 21.3
100 0.3 6.8 29.0 5.0 21.3 5.5 8.7 3.6 21.6
100 0.5 9.8 29.6 5.2 20.1 5.8 8.2 4.3 19.9
100 0.7 15.2 31.1 5.3 16.0 5.4 7.5 3.1 14.6
200 0.0 6.0 52.8 5.2 35.8 5.5 12.4 4.2 47.7
200 0.3 8.3 56.7 5.3 37.1 5.1 13.0 4.3 48.6
200 0.5 14.9 59.7 4.7 34.0 5.1 11.6 3.6 43.0
200 0.7 29.9 61.8 6.2 26.0 5.2 10.2 4.3 29.7
400 0.0 5.4 87.3 4.5 62.4 5.6 25.0 4.3 86.2
400 0.3 11.1 90.0 5.3 61.6 5.3 22.4 4.3 85.2
400 0.5 27.0 91.7 4.5 55.4 4.1 18.7 4.1 79.6
400 0.7 56.6 93.2 4.9 46.6 4.6 15.4 4.0 64.5
tests (in %) for β2 (t) = 0.51t≤0.5 − 0.51t≥0.5 . Taken from Chauvel (2014).
Simple departures from proportional hazards models

Via a number of examples, Chauvel and O’Quigley (2017) and Chauvel (2014)
illustrate how to construct non-proportional hazards models starting with the
regression effect process and the R2 coefficient. To begin with, we will illustrate
the method in the univariate setting before moving on to the multivariate one.
We will draw confidence bands for each process as described in Section 10.4 in
order to get an idea of the validity of the proportional hazards hypothesis. The R
package PHeval1 developed by Chauvel (2014) allows us to evaluate and plot the
regression effect process, as well as calculate the R2 coefficient. We will proceed
with a simulated data set of size n = 200 and one covariate Z following a Bernoulli
distribution with parameter 0.5. Censoring times are simulated independently of
event times from a uniform distribution on [0, 3]. We first simulate a sample
under the proportional hazards model with parameter 1:
λ(t|Z) = exp(Z). (10.12)
Figure 10.4: The regression effect process Un∗ (0, ·) (solid line) and its confidence
band (dotted lines) for the simulated dataset with β(t) = 1 (left-hand) and trans-
formed process (right-hand).
The rate of censoring is 18%. The regression effect process Un∗ (0, ·) and its
confidence band under the proportional hazards model are plotted as a function
of time in Figure 10.4. We observe drift, which implies a non-zero effect. The
drift would appear to be linear and the process does not exit the 95% confidence
bands; thus, the proportionality hypothesis appears reasonable. As the drift is
increasing, the coefficient will be positive. The estimator based on the partial
maximum likelihood of the proportional hazards models gives 1.08 and the R2
1 Available at http://cran.r-project.org/web/packages/PHeval/index.html
coefficient is 0.21. Note that the transformation to the Brownian bridge is valid
when applied to Brownian motion with drift, as long as the drift is linear. The
slope term disappears in the transformation.
A simple change-point model

Suppose we simulate a dataset under the change-point model
λ(t|Z) = exp(β(t)Z) ; β(t) = 1t≤0.4
with t on the initial time scale. The coefficient is equal to 1 for the first part
of the study, then zero afterwards. The simulated sample has 26% censoring.
Figure 10.5(a) shows the regression effect process Un∗ (0, ·) as a function of the
transformed time between 0 and 1. Drift can be seen, indicating the presence of
a non-zero effect. The drift does not seem to be linear and the process exits the
95% confidence band. There is evidence of an effect that changes over time. The
drift appears linear until time 0.6 on the transformed time scale, and the effect
appears to be weaker afterwards. This suggests that a single regression coefficient
alone may not be adequate to provide a good summary of the observations. A
more involved model would postulate a constant effect until time 0.6 on the
transformed time scale, then zero after, denoted β0.6a (t) = β0 1t≤0.6 . Here, the
observed drift after time 0.6 corresponds to noise. The effect after 0.6 could also
be non-zero, so we consider the effect
β0.6b (t) = β0 1{1 ,

t≤0.6 + C0.6 1t>0.6 }
Figure 10.5: The regression effect process Un∗ (0, ·) for the dataset simulated
according to a change-point model.
with unknown β0 and C0.6 . The value C0.6 is the one multiplied by the regression
coefficient in the second part of the study. In Figure 10.5(b), two straight lines
have been fitted to the process using linear regression, before and after t = 0.6.
The ratio of the second one’s slope to the first’s gives C0.6 = −0.17.
We also consider other change-point times at t = 0.4, t = 0.5, and t = 0.7,
and in particular the coefficients β0.4a (t) = β0 1t≤0.4 , β0.5a (t) = β0 1t≤0.5 and
β0.7a (t) = β0 1t≤0.7 , which are zero in the second part of the study. The piecewise
constant coefficients are not zero in the second part of the study; β0.4b , β0.5b
and β0.7b were considered and plotted over the process in Figure 10.6. Fitting
was performed using linear regression.
The shape of the regression effect process indicates that the effect β(t)
decreases with time. Several models with continuously decreasing effects β(t) =
β0 h(t) were selected together with the already mentioned coefficients. For each
model, the maximum likelihood estimator β̂0 and the R2 coefficient was calcu-
lated. The coefficients β0.6a (t) and β0.6b (t) give the largest value of R2 here,
0.25, which corresponds to an increase of 80% in predictive ability with respect
to the proportional hazards model, which had R2 = 0.14. Of the two models
with excellent predictive power, we retain the one with regression parameter
β0.6a (t) = β0 1t≤0.6 . Indeed, as the two models fit the data well and have good
predictive ability, we would choose to retain the one with the simplest coefficient.
The time t = 0.6 on the transformed scale corresponds to the time 0.39 on the
initial scale. Recall that on the initial scale, the change-point time used to sim-
ulate the data was 0.4. The regression coefficient selected is therefore close to
the one used to simulate the data (Table 10.6).
Figure 10.6: Regression effect process Un∗ (0, ·) (solid line) and piecewise con-
stant coefficient fits (dotted lines) with a change-point at time t0 , for a dataset
simulated according to a change-point model.
R2 0.25 0.25 0.24 0.23 0.23 0.23

β(t) β0.6a (t) β0.6b (t) β0 (1 − t)2 β0.5a (t) β0.5b (t) β0 (1 − t)
β̂0 1.46 1.42 2.78 1.58 1.57 1.90
R2 0.22 0.22 0.21 0.19 0.17 0.14
β(t) β0.7b (t) β0.7a (t) β0 (1 − t2 ) β0.4b (t) β0.4a (t) β0
β̂0 1.20 1.21 1.38 1.57 1.51 0.74
Table 10.6: R2 coefficients and partial maximum likelihood estimators β̂0 for
the dataset simulated under a change-point model.
Steadily decreasing regression effects

Consider now the case of a continuously decreasing effect. We simulate data
using β(t) = 3(1 − t)2 on the transformed scale [0, 1] with n = 200 individuals
and uniform censoring on [0, 3].
The regression effect process Un∗ (0, ·) is plotted as a function of time in Figure
10.7(a) with a confidence band within which the process would lie with high
probability under a proportional hazards hypothesis. However, the process exits
the 95% confidence band, indicating that the proportional hazards assumption
is not wholly satisfactory. The trend in the drift is concave, with the effect
appearing to decrease over time. Among other possibilities, the effect could be
linear, quadratic, or even piecewise constant. For the latter, the trend would
seem to be linear up to t = 0.6, corresponding to a constant coefficient. Then,
the drift moves to smaller values, leading us to consider the coefficient β06a (t) =
β0 (1t≤0.6 + C06 1t>0.6 ) , where β0 and C06 are unknown. C06 is the value by
which the coefficient is multiplied in the second part of the study. Using linear
Figure 10.7: The regression effect process Un∗ (0, ·) for a dataset simulated with
a continuously decreasing effect.
regression, two lines have been fitted to the process, before and after the change-
point at t = 0.6 (Figure 10.7(b)). The ratio of the second slope to the first gives
the value C06 = −0.2. We also considered the coefficient β06b (t) = β0 1t≤0.6 ,
which equals zero after t = 0.6. Using the same methodology, we looked at
piecewise constant coefficients with change-points at t = 0.4, 0.5 or 0.7 (see
the definitions of the coefficients in the previous example). Several models with
continuously decreasing effects over time were investigated.
R2 0.36 0.34 0.34 0.34 0.31 0.31 0.31

β(t) β0 (1 − t)2 β0.6b (t) β0.6a (t) β0 (1 − t) β0.5a (t) β0 (1 − t2 ) β0.5b (t)
β̂0 3.54 1.64 1.65 2.34 1.78 1.69 1.77
R2 0.31 0.31 0.30 0.29 0.28 0.21
β(t) β0.7b (t) β0.7a (t) β0.4b (t) β0 exp(−t) β0.4a (t) β0
β̂0 1.44 1.43 2.01 1.79 1.98 0.95
Table 10.7: R2 coefficients and maximum partial likelihood estimators β̂0 for
the dataset simulated under some different NPH models. Taken from Chauvel
(2014).
The R2 coefficients and partial maximum likelihood estimates β̂0 of β0 are

shown in Table 10.7. The weakest R2 corresponds to the proportional hazards
model. The model selected is the one with the largest R2 , corresponding here to
the model associated with the coefficient β(t) = β0 (1−t)2 , where β0 is estimated
as 3.54. The R2 of this model is 0.36, corresponding to an increase of 70% in
predictive ability. Again, we see that the data were simulated under the model
that was correctly selected.
Graphs in the multivariate case

Here, we simulated two centered normal covariates Z (1) and Z (2) with variance
1 and correlation 0.5 according to:

λ(t|Z) = exp β1 (t)Z (1) + β2 Z (2) ,
with β1 (t) = 1t≤0.7 and β2 = −1. For this example, we chose to simulate, with-
out censoring, n = 200 individuals. Each component of the bivariate process
Σ̂−1/2 Un∗ (0, ·) and its corresponding 95% confidence bands are plotted as a func-
tion of the transformed time between 0 and 1 in Figures 10.8(a) and 10.8(b). The
proportional hazards hypothesis for covariate Z (1) is rejected at the 5% level since
the process exits the confidence bands. The shape of the process’s drift suggests a
piecewise constant effect with a change-point at t0 = 0.6 on the transformed time
scale. As in the univariate case, two fits were implemented using linear regression,
one before t0 = 0.6 and one after. The ratio of the slopes is −0.12, so we con-
sider the regression effect β1 (t) = β1 B0.6 (t), with B0.6 (t) = 1t≤0.6 −0.12 1t≥0.6 .
Other piecewise constant regression coefficients β(t) = β1 Bt0 (t) were also inves-
Figure 10.8: The regression effect process Σ̂−1/2 Un∗ (0, ·) (solid lines), confidence
bands (dashed lines) and piecewise constant coefficient fits (dotted lines) for
simulated multivariate data.
tigated, with change-points at times t0 ∈ {0.45, 0.5, . . . , 0.7} on the transformed

time scale. For each time t0 , the ratio of the slopes was calculated in order to
determine the value to be multiplied by the coefficient in the second part of
the study. The second covariate, Z (2) , appears to have a constant regression
effect since the process stays well inside the 95% confidence bands, and the drift
appears close to being linear (Figure 10.8b). Therefore, we would choose to work
with the regression coefficient β2 (t) = β2 (Table 10.8).
β1 (t) β1 β1 B0.45 (t) β1 B0.5 (t) β1 B0.55 (t) β1 B0.6 (t) β1 B0.7 (t)
β̂1 0.45 0.93 0.96 0.89 0.95 0.72

β̂2 -0.73 -0.72 -0.73 -0.74 -0.79 -0.77
R2 0.24 0.35 0.37 0.35 0.39 0.32
Table 10.8: Partial maximum likelihood estimates β̂(t) and R2 coefficients for
simulated multivariate data. Best fitting model has R2 = 0.39.
Estimation results are shown in Table 10.8. The proportional hazards model
gives an R2 of 0.24. The largest R2 is obtained with β1 (t) = β1 B0.6 (t), increasing
the predictive ability by 60% with respect to the proportional hazards model. In
conclusion, we retain the model with β1 (t) = β1 B0.6 (t) and β2 (t) = β2 . Note
that the time t0 = 0.6 on the transformed scale corresponds to the time 0.68 on
the initial one, which is close to the 0.7 used to simulate the data.
10.10. Illustrations from clinical studies 291
10.10 Illustrations from clinical studies
Freireich leukemia study

The Freireich dataset is often referred to in the survival analysis setting (Acute
Leukemia Group B et al., 1963; Cox, 1972). These observations are generally
taken to respect the proportional hazards hypothesis. Half of the 42 leukemia
patients in the study received a new treatment: 6-Mercaptopurine, while the
other half received the control. The regression effect process Un∗ (0, ·) and its
confidence band are plotted as a function of the transformed time scale [0, 1] in
Figure 10.9. As expected, the process does not exit the 95% confidence band;
the proportional hazards model fits the data well.
Clinical trial in breast cancer

In a clinical trial context, data were collected from 1504 breast cancer patients
at the Curie Institute, Paris. After initially successful treatment, patients were
followed over a 15-year period. The rate of censoring in the study was 24%. One
of the study goals was to build a descriptive survival model leading to a better
understanding of the influence of various prognostic factors on post-treatment
survival. As the effect of some of the factors might change over time, the resulting
prognostic rules, i.e., linear combinations of effects and the prognostic factors,
should reflect this.
The prognostic factors in the database were the presence of the progesterone
receptor, tumor size above 60 mm, and cancer stage above 2. The multivariate
regression effect process Σ̂−1/2 Un∗ (0, ·) and its confidence bands are plotted as
Figure 10.9: Kaplan-Meier curves and model based curves for Freireich data.
Regression effect process indicates strong effects and good model fit.
Figure 10.10: Kaplan-Meier curves for variable tumor size for Curie Institute
breast cancer study. Regression effect process indicates clear effects that diminish
gradually with time.
a function of the transformed time [0, 1] in Figure 10.10. The process exits the
95% confidence band and the effect appears not to be constant, with a gradually
decreasing slope over time. Hence, one model that fits the data better than
the proportional hazards one has a piecewise constant effect for the tumor size
variable, with a change-point at t = 0.2 on the transformed time scale, and
constant effects for the other two prognostic factors. As in the simulations, two
straight lines have been fitted to the process, before and after t = 0.2, which
leads us to consider a regression effect for the tumor size variable of βsize (t) =
β0 (1t≤0.2 + 0.24 1t≥0.2 ). The predictive ability of this model with respect to the
proportional hazards one corresponds to an increase of more than 30% in the
R2 , moving from 0.29 to 0.39 (Table 10.9).
Figure 10.11 shows the process for the presence of progesterone receptor effect
as a function of the transformed time scale. The drift is not entirely linear but
the process stays within the confidence bands corresponding to a constant effect
over time. We considered several potential regression effects: a change-point
model with a jump at t = 0.5 on the transformed time scale, i.e., βrec0 (t) =
β0 (1t≤0.5 + 0.39 1t>0.5 ), and several continuous effects: βrec1 (t) = β0 (1 − t),
βrec2 (t) = β0 (1−t)2 , βrec3 (t) = β0 (1−t2 ) and βrec4 (t) = β0 log(t). Figure 10.12
shows the process for the cancer grade effect. It touches the lower bound in the
constant effect confidence band, but does not breach it. The drift does not seem
to be linear, and gives the impression of a negative effect that decreases over
time. The simplest effect possible, i.e., constant over time, would nevertheless
appear to be a good candidate; however, in the model construction context, we
will also consider time-dependent effects. In particular, we examine the piecewise

constant effect βgrade (t) = β0 (t)(1t≤0.4 + 0.691t>0.4 ).
We looked at all combinations of these effects, as well as the proportional
hazards model. For each combination, the R2 was calculated and the regression
effects were estimated by maximizing the partial likelihood.
Progesterone
Tumor size Stage R2
receptor
0.84 1.03 -0.68 0.29
1.77(1t≤0.2 + 0.241t>0.2 ) 1.03 -0.66 0.39
0.85 −1.02 log(t) -0.67 0.39
1.74(1t≤0.2 + 0.241t>0.2 ) −1.02 log(t) -0.66 0.51
1.72(1t≤0.2 + 0.241t>0.2 ) −1.02 log(t) −0.82(1t≤0.4 + 0.691t>0.4 ) 0.52
Table 10.9: Partial likelihood estimates and R2 coefficients for the clinical trial
data.
Some results are shown in Table 10.9. The proportional hazards model gives
an R2 of 0.29. As we have mentioned previously, a model allowing the tumor
size effect to change in a simple way over time gives a great improvement in
predictive ability (in the order of 30%). The largest R2 is obtained with piecewise
constant effects for tumor size and grade, and an effect proportional to log(t)
for the progesterone receptor. The predictive ability of this model is 80% better
than the proportional hazards one for the three prognostic factors. If we instead
introduce a time-dependent effect for the cancer grade while also allowing the
tumor size and progesterone effects to change over time, we get an increase in
Figure 10.11: Kaplan-Meier curves for variable progesterone status for breast can-
cer study. Regression effect process is suggestive of some weak time dependency.
Figure 10.12: Kaplan-Meier curves for variable tumor grade for breast cancer
study. Process indicates the presence of effects that appear to diminish with
time.
the R2 from 0.51 to 0.52. Such a small improvement cannot justify the added
complexity of this model; therefore, we retain the one with constant cancer stage
effect and non-constant effects for tumor size and progesterone.
Building more complex models one step at a time

The regression effect process is a tool that allows us to construct non-proportional
hazards models. The idea of the process is to use the raw data without estimat-
ing any parameters, and graphically evaluate the temporal form of the regression
effect, as is the case for scatterplots in two dimensions for uncensored data. Note
that processes developed until now for investigating the goodness of fit of pro-
portional hazards models use an estimate of the regression parameter. Another
difference between the regression effect process and others found in the literature
is in terms of standardization. Two types are applied: one to the time, the other
to increments of the process. We have chosen to standardize each increment
rather than apply an overall standardization. This allows us to analytically obtain
the process’s limit distribution, as well as decorrelate covariates. Knowledge of
the process’s asymptotic distribution means that it is possible to derive a good-
ness of fit test and confidence bands for a constant effect in the proportional
hazards model for each component of the process. These tests are not affected
by correlations between covariates since this has been taken into account in the
construction of the process (Figure 10.13). Related goodness of fit and predictive
ability procedures that complement the above techniques for statistical modeling
provide us with a coherent model construction methodology. Intuitively, models
Figure 10.13: Fitting a simple change-point model to the regression effect process
for the variable receptor in the breast cancer study in order to obtain a better fit.
constructed in this way should have better predictive abilities, as is confirmed

by the theoretical and practical results of the chapter. Change-point models,
that is, non-proportional hazards (NPH) models with piecewise constant effects,
are more sophisticated than proportional hazards ones. The model construction
method presented here could potentially be used to estimate the change-point
locations, as well as the various values of the regression parameter. In some ways,
analogous to the linear model in which the addition of a quadratic term is the
first step in a direction that moves away from a linearity assumption, we can see
a model with a single change-point as the first step away from a PH model in
the direction of NPH models. In the examples we have considered, we looked at
the one change-point case, and the ratio of the two regression coefficients was
estimated by linear regression using the slope of the regression effect process. It
may instead be possible to estimate the two parts of the coefficient by maximiz-
ing the partial likelihood (Figure 10.14). In the method we have proposed, users
play an important role in determining the regression coefficients that lead to the
data being modeled satisfactorily. We chose not to estimate the parameter using
basis functions like splines or histograms, so as to be able to interpret the regres-
sion coefficients more easily. The retained model will fit the data well and have
good predictive abilities. There may be other candidate models, and for these,
we could go through the same steps. Our viewpoint is that the model is no more
than an efficient way of summarizing the infinitely complex mechanisms that can
be considered to have generated the data. Necessarily, there can be more than a
single candidate model, not only in terms of the inclusion of certain regressors,
but also in terms of the particular model structure. The tools used here allow us
Figure 10.14: Right-hand figures indicates an improvement of fit for a simple

change-point model when compared to a basic PH model.
to move away from models showing poor adequacy in the direction of models
with improved adequacy. The sense of adequacy is both in terms of providing
a model based description that can plausibly be considered to have generated
the observations, and in terms of the strength of prediction. These two senses
feed off one another so that model construction is to some degree of an iterative
nature. In Figure 10.15 we can see a further small improvement in the model fit
Figure 10.15: Right-hand figures indicates improvement of fit for a double change-

point model when compared to a simple change-point model.
by increasing the complexity of the model from a single change-point to a model

with two change-points. The improvement in fit is very modest and probably not
worth the extra complexity that is called for. The predictive capability of the two
models barely differs. We can use the difference between the regression effect
process and twice the area under the curve to evaluate the degree of deviation
from the proportional hazards hypothesis (Chapter 11). If the proportional haz-
ards model is valid, this difference ought to be small and a statistical test can
readily be built to test it. When there is a single covariate, knowledge of the limit
distribution of this statistic leads to the construction of a goodness of fit test.
Extending this to the multivariate case is not obvious and needs further study.
Under the non-proportional hazards model with parameter β(t), the regres-
sion effect process {Un∗ (β(t), t), 0 ≤ t ≤ 1} converges to standard Brownian
motion.
It would be interesting
to study the limit behavior of the process
t), 0 ≤ t ≤ 1 , where β(t)
Un∗ (β(t), is a convergent estimator of β(t). Knowl-
edge of the asymptotic distribution of this process would allow us to examine
whether the estimator of β(t) selected by the model construction method is close
where β(t)
to the true parameter of the model. If the process is evaluated at β(t),
is the partial maximum likelihood estimator of β(t), then the process is no longer
(Ft∗ )-predictable. Though our asymptotic results are based on this property, it
is not simple to derive the asymptotic distribution. Nevertheless, note that the
goodness of fit of the non-proportional hazards model with constant parameter
β0 (t) can be evaluated using the test with null hypothesis H0 : β(t) = β0 (t)
studied in Chapter 11.
1. It is not uncommon in the applied scientific literature to encounter state-

ments such as “R2 indicated an excellent fit at over 70%. Explain why
this statement is incorrect. Find examples from linear regression and logistic
regression where poor fit is associated with a large value of R2 and where a
good fit is associated with small values of R2 .
2. Describe the advantages of graphical goodness of fit procedures over formal
goodness of fit tests.
3. The fit for a proportional hazards model with a continuous covariate may be
poor in two distinct ways: a regression coefficient that turns out to be not
independent of time or a poorly-specified functional form for the continuous
covariate. Show that no formal test can distinguish either situation from the
other.
4. To fit the drift parameter for a Brownian motion with linear drift it suffices
to draw a straight line between the origin and the arrival point. Show how
this idea extends to non-linear drift.
5. Show how we can approximate the drift parameter function for an observed
Brownian motion with non-linear drift by using multiple change-points. How
would you make use of R2 to obtain the best possible fit while minimizing
the number of change-points?
6. Given two models: one has a higher R2 than the other but shows a signif-
icantly poorer fit. Which of the two models would you recommend? Give
reasons.
7. On the basis of a real data set make use of stepwise model building techniques
to obtain the model that appears to be the most satisfactory. At each step
argue the basis underlying any specific decision. On what basis would we
conclude that, given the information we have, we are unlikely to find a better
performing model.
8. Consider a model with a single continuous covariate. For the purposes of

interpretability it may be desired to use the continuous covariate to establish
two distinct groups. How would the techniques of this chapter help do that?
How would you decide if the two group model was preferable to the model
with a continuous covariate, or vice versa?
9. Consider a model with a binary covariate. Describe the impact of increasing

censoring on the regression effect process under; (1) independent censoring
and (2) covariate dependent censoring, otherwise known as conditionally
independent censoring.
Proposition 10.1 Under the hypotheses of Proposition 10.1, the distribution of T

conditional on Z is an exponential E (α exp(βZ)). Thus, E(T |Z) = exp(−βZ)/α
and E(T 2 |Z) = 2 exp(−2βZ)/α2 . We have therefore that:
Var(T ) = E(T 2 ) − E(T )2 = E(E(T 2 |Z)) − E(E(T |Z))2

1 1
= 2 2E(exp(−2βZ)) − E(exp(−βZ))2 = 2 2 φZ (−2β) − φZ (−β)2 ,
α α
where φZ is the moment generating function of Z. Also:
1
Var(T |Z) = exp(−2βZ),
α2
and
1 1
E(Var(T |Z)) = 2
E(exp(−2βZ)) = 2 φZ (−2β). (10.13)
α α
Using these equations and the definition of Ω2T |Z , we get:
φZ (−2β) 1
Ω2T |Z = 1 − = 1− . (10.14)
2φZ (−2β) − φZ (−β)2 2 − φZ (−β)2 φ−1
Z (−2β)
As the moment generating function is positive, the result follows.

Theorem 10.2 Let t ∈ [0, 1]. Under the conditions of the theorem, a Taylor
expansion of e (α(t), t) gives:
e (α(t), t) = e (β(t), t) − (α(t) − β(t)) v (γ(t), t) ,
where γ is in the ball with center α and radius supt∈[0,1] |α(t) − β(t)|. Under A3,
there exists a constant C(γ) such that v (γ(t), t) = C(γ). Using this expansion
in equation (10.11), we get:
1
2 C(β) + C(γ)2 0 (α(t) − β(t))2 dt
lim R (α(t)) = 1 − 1 .
n→∞ C(β) + 0 (e(0, t) − e(β(t), t))2 dt
As a consequence, maximizing the function α −→ limn→∞ R2 (α) is equivalent

1
to minimizing the function α −→ 0 (α(t) − β(t))2 dt.
Chapter 11
Hypothesis tests based on regression effect

process
11.1 Chapter summary
We revisit the standard log-rank test and several modifications of it that come
under the heading of weighted log-rank tests. Taken together these provide us
with an extensive array of tools for the hypothesis testing problem. Importantly,
all of these tests can be readily derived from within the proportional and non-
proportional hazards framework. Given our focus on the regression effect pro-
cess, it is equally important to note that these tests can be based on established
properties of this process under various assumptions. These properties allow us
to cover a very broad range of situations. With many different tests, including
goodness-of-fit tests, coming under the same heading, it makes it particularly
straightforward to carry out comparative studies on the relative merits of dif-
ferent choices. Furthermore, we underline the intuitive value of the regression
effect process since it provides us with a clear visual impression of the possible
presence of effects as well as the nature of any such effects. In conjunction with
formal testing, the investigator has a powerful tool to study dependencies and
co-dependencies in survival data.
Significance tests have an important role to play in formal decision making but
also as a way to provide the shortest possible confidence intervals in a model
fitting and estimation setting. There are several possible statistics for testing a
null hypothesis of no difference in survival experience between groups against
some alternative hypothesis. The alternative may be of a general nature or may
be more specific. Specific directions moving away from the null will be mirrored,

301
https://doi.org/10.1007/978-3-030-33439-0_11
302 Chapter 11. Hypothesis tests based on regression effect process
in the regression effect process. If the form of this process can be anticipated
ahead of time this will allow us to derive tailor made tests with high power
for certain specific alternatives. Mostly, the null hypothesis corresponds to an
absence of regression effects for two or more distinct groups, i.e., β(t) = 0 for
all t, and, in this case, the test statistics often assume a particularly simple
form. A classical example arises in the context of a clinical trial when we wish
to compare the survival experience of patients from different treatment groups,
the goal being to assess whether or not a new treatment improves survival. For
the sake of simplicity, and clarity of presentation, we mostly consider the case
with two groups defined by a single binary variable. Extensions to several groups
raise both operational and conceptual issues. We may wish to test the impact
of all variables taken together—an example is provided of a clinical trial with
3 treatment arms defined via 2 binary covariates—or we may wish to test the
impact of one variable having controlled, in some way, for the effects of one or
more of the other covariables. Often we will restrict the alternative hypothesis
to one that belongs to the proportional hazards family.
11.3 Some commonly employed tests
Consider a straightforward clinical trial. Patients in the first group receive the
new treatment, indexed by T , and patients in the second group receive a placebo,
indexed by P . A comparison of the two survival functions is carried out by testing
a null hypothesis H0 against an alternative H1 , expressed more formally as:
H0 : ∀t, SP (t) = ST (t), versus H1 : ∃ T such that ∀t ∈ T , SP (t) = ST (t),

where T dt > 0. This is a little cumbersome but is needed to rule out (non-
detectable and thereby not consistent) alternatives defined on a null set, i.e., a
time interval with zero measure. We can write this more succinctly in terms of
the parameter function β(t) so that we consider H0 and H1 where:

H0 : ∀t, β(t) = 0, versus H1 : ∃ T such that T dt > 0 and T β 2 (t)dt > 0.
The above alternative is very general and while it will have power against any
kind of departure from the null, we can increase this power, in some cases greatly
increase this power, by considering more specific alternative hypotheses. This is a
constant theme throughout this chapter. We focus more on two-sided tests, but
one-sided tests can be structured once we figure out how to order different pos-
sible functions for β(t). This is simple when β is constant and does not depend
on t, otherwise the question itself is not always easily framed. A comparison of
survival functions can also be useful in epidemiological studies, to compare for
example the survival of groups of subjects exposed to different environmental
11.3. Some commonly employed tests 303
factors. In these kinds of studies we would usually expect that, under the alter-
native hypothesis, at any time (age), t, the difference between groups would be
in the same direction.
Log-rank and weighted log-rank tests

When trying to detect a difference between survival functions, the log-rank test
(Mantel, 1966) is the most commonly used. The test can be expressed as a
likelihood based score test under a proportional hazards model in which the
covariate is discrete (Cox, 1972). Denote U (0) (resp. −I(0)) the first (resp. sec-
ond) derivative of Cox’s partial log-likelihood with respect to the parameter β,
calculated at β = 0. On the standardized time scale from 0 to 1, these quantities
are respectively
1 kn

U (0) = {Z(s) − E0 (Z | s)} dN̄ ∗ (s) = {Z(tj ) − E0 (Z | tj )} ,
0 j=1
and
1 kn

I(0) = V0 (Z | s)dN̄ ∗ (s) = V0 (Z | ti ).
0 i=1
By definition, the test with statistic U (0)2 /I(0) is a score test for testing
H0 : β = 0 in the proportional hazards model with parameter β.
Definition 11.1. The log-rank test rejects H0 with Type 1 risk α if |Ln | ≥
z α/2 , where z α/2 is the upper quantile of order α/2 of the standardized
normal distribution, and

n k
U (0) Z(tj ) − E0 (Z | tj )
Ln = = (11.1)
I(0) j=1 kn V (Z | t )
i=1 0 i
is the test statistic.
When β ∈ Rp , the statistic L2n is defined by L2n = U (0)T I(0)−1 U (0) and con-
verges to a χ2 distribution with p − 1 degrees of freedom under the null hypothe-
sis. Written in this way, the log-rank test can be extended to continuous variables
(e.g., age) to test for their effect on survival. The test statistic is easy to imple-
ment, and the p-value of the test can be accurately approximated since the
asymptotic distribution is known. The great popularity of this test is in particular
due to it being the most powerful (Peto and Peto, 1972) under local alternative
hypotheses of the proportional hazards type, which are written
H1 : ∃ β0 = 0, ∀ t, β(t) = β0 = 0.
This test is powerful for detecting the presence of an effect β that is constant
in time between the groups, under the proportional hazards model. On the other
hand, in the presence of non-proportional hazards, i.e., an effect β(t) which varies
over time, the log-rank test is no longer optimal in the sense that there exist
alternative consistent tests with greater power. This is the case, for instance,
in the presence of hazards that intersect, which happens when the coefficient
changes sign during the study period (Leurgans, 1983, 1984).
Definition 11.2. The weighted log-rank test has the statistic LWn given by
kn
Z(tj ) − E0 (Z | tj )
LWn = Wn (tj ) , (11.2)
kn 2
j=1 i=1 Wn (ti ) V0 (Z | ti )
where Wn is a function with positive weights, defined on [0, 1].
If we wish to use the techniques of martingale theory to study these kinds of

weights then we will usually need some added restrictions, the most notable
being that Wn be Ft∗ -predictable. In words, at time t we only use information
available up to that time point and we do not let Wn (t) depend on outcomes that
can take place at values greater than or equal to t. Specific choices of weights
lead to the ability to detect different types of temporal effect β(t). Recall that
the transformation φ−1 n , defined on [0, 1], lets us map times t ∈ [0, 1] back to the
initial time scale (see Section 9.3). Several suggestions for weights for particular
situations have been presented, among which:
−1
• Wn (t) = n i=1 Yi φn (t) , the number of individuals at risk at time t (Bres-
low, 1970; Gehan, 1965),

• Wn (t) = Ŝ φ−1 n (t) , the Kaplan-Meier estimator of survival at time t (Pren-
tice, 1978),

• Wn (t) = S̃ φ−1 n (t) , where S̃ is a modified version of the Kaplan-Meier esti-
mator (Prentice, 1978),
n
• Wn (t) = g Yi φ−1
n (t) , for a given function g (for example, g(x) =
√ i=1
x, ∀ x) (Tarone and Ware, 1977),
p q
• Wn (t) = Ŝ φ−1
n (t) 1 − Ŝ φ−1
n (t) , p, q ≥ 0, which are weights of the
family described in Fleming and Harrington (1991, 2005).
Many other weights have been proposed (Lagakos et al., 1990; Mantel and Sta-
blein, 1988; Wu and Gilbert, 2002). Note that the test with the weights proposed
in Gehan (1965) corresponds to a modification of the Wilcoxon (1945) test that
takes censored data into account. This test is equivalent to the two sample
Mann-Whitney test and to the Kruskal-Wallis test when there are more than two
samples (Breslow, 1970). Aalen (1978) and Gill (1980) have shown that weighted
log-rank statistics converge to centered normal distributions. Jones and Crowley
(1989, 1990) have defined a broader class of tests that includes weighted and
unweighted log-rank tests. This class also includes tests for comparing more than
two survival functions, as well as specific constructions such as a trend across
groups.
All of these weighted log-rank tests were developed on the initial time scale.
However, since they are rank tests, they remain unchanged on the transformed
time scale [0, 1] using the mapping φn from Equation (9.2). When the weights
are larger for earlier times, we will put more emphasis there so that such a test
ought to be more sensitive to the detection of early effects. This can be expressed
as the coefficients β(t) decreasing with time (Fig. 11.1).
Figure 11.1: Examples of non-proportionality detectable from the Kaplan-Meier

curves. Correct weighting will significantly improve the power of the log-rank
test.
Many other choices are possible. The weights proposed by Tarone and Ware
(1977) and Fleming and Harrington (1991) involve classes of weights for which
a good choice of function g or (p, q) can help detect late effects, i.e., coefficients
β(t) which increase with time. In immunotherapy trials there is often an early
period in which no effects can be seen, followed by a period in which effects
become manifest. A judicious choice of the weights described by Tarone and Ware
(1977) and Fleming and Harrington (1991) enables us to increase the power of
the basic log-rank test. The optimal choice of g or (p, q) depends on the data,
and may not be easy to choose. Peckova and Fleming (2003) have proposed a
weighted log-rank test whose weights are chosen adaptively as a function of the
data. Such tests will generally display good behavior, although their properties,
both under the null and the alternative, may be difficult to anticipate.
While these tests can exhibit good power under non-proportional hazards-
type alternative hypotheses, they would be expected to lose power with respect to
unweighted log-rank tests under proportional hazards-type alternative hypothe-
ses. Moreover, in order to set suitable weights, if the alternative hypothesis

involves non-proportional hazards, it is necessary to know the type of variation
that is likely to arise (early effects, late effects, larger effects in the middle of
the study, etc.). Yang and Prentice (2010) have proposed an adaptive test which
the authors claim will simplify the log-rank test when dealing with proportional
hazards. Although the model upon which the test is based (see Equation 6.26)
may simplify under some parametric constraint, it does not follow that the same
holds for the test based upon this model. Indeed, unlike the log-rank test, the
test based upon this model is not invariant to the coding of the covariable. The
class is broad and does include the proportional hazards and proportional odds
models. The test statistic is the maximum difference in the absolute value of
the two weighted log-rank statistics LW1,n and LW2,n , with respective weights
Wn1 (t) = λ̂T (t)/λ̂P (t) and Wn2 (t) = λ̂P (t)/λ̂T (t), where the instantaneous haz-
ard estimators in the groups receiving the placebo, λ̂P , and the treatment, λ̂T ,
are based on model (6.26). This test has unacceptable behavior and we would
not recommend its use.
Combining log-rank and weighted log-rank tests

Regulations for clinical trials require that the choice of statistical analysis be
defined in advance, and consequently, a choice of the test to be used needs to
be made before data are collected. It is often difficult to predict whether the
effect of the treatment group on survival time is constant or whether it changes
over time, and if so, what type of change is involved. The log-rank test is the
most commonly used test, due to its optimal local behavior under proportional
hazards-type alternative hypotheses. However, when faced with non-proportional
hazards situations, the low power of the test may prevent detection of the effect
of groups on survival, even when differences are strong and manifest from the
Kaplan-Meier curves themselves. For such reasons, developing a test that behaves
like the log-rank under proportional hazards but that continues to maintain good
power, and indeed overtaking the log-rank, in non-proportional hazards situations
appears as a worthwhile goal. Several attempts in this direction have been made.
The first approach is to combine several weighted log-rank statistics, usually as a
linear combination. For example, to detect the presence of an effect that appears
after a certain time lag, Zucker and Lakatos (1990) have studied a test whose
statistic is the sum of a log-rank statistic and a weighted log-rank one, chosen to
be good at detecting late effects. Lee (1996) has looked at a linear combination
of four weighted log-rank statistics with weights from the family given in Fleming
and Harrington (1991). Denoting by Gp,q the weighted log-rank statistic with
weights Ŝ(t)p (1 − Ŝ(t))q , p, q ≥ 0, the four statistics considered by Lee are:
• G0,0 = Ln , for detecting constant effects (proportional hazards),
• G2,0 , for detecting early effects,

• G0,2 , for detecting late effects,
• G2,2 , for detecting effects in the middle of the study.

Similarly, Wu and Gilbert (2002) have proposed a linear combination of several
weighted log-rank statistics with weights of the form
2
Wna (t) = Ŝ(t) − (aŜ(τ0 ) + 1 − a) , (11.3)
where Ŝ is the Kaplan-Meier estimator, τ0 the maximum follow-up duration,

and a a parameter between 0 and 1 that allows for detection of different types
of time-dependent effect. The disadvantage of all these linear procedures is to
choose the optimal weights to give to each statistic. A poor choice of weights
could result in a significant loss of power, even in situations normally covered
well by one of the included statistics (Lee, 1996). On the other hand, such tests
are of interest if investigators already have an idea of the type of effect expected.
Another possibility for combining statistics is to keep the one with the
maximum absolute value. Tarone (1981) has studied the test whose statistic
is the maximum of the log-rank statistic and the weighted log-rank statistic
with the weights given in Gehan (1965). Self (1991) has investigated a supre-
mum test over the class of weighted log-rank statistics with weights of the form
tθ1 (1 − t)θ2 , θ1 , θ2 ≥ 0. Fleming and Harrington (1991) have studied the prop-
erties of the test whose statistic is the maximum of several statistics Gp,q . Lee
(1996) studied this test using simulation, looking at the class of four statistics
G0,0 , G2,0 , G0,2 and G2,2 . Wu and Gilbert (2002) have also studied the theo-
retical properties of the test whose statistic is the maximum of several weighted
log-rank tests with weights Wna (t) (Equation 11.3). Kosorok and Lin (1999) have
looked at the supremum of a class of statistics indexed by functions, extending
the results of Jones and Crowley (1990) to the multiple covariate case. Finally,
Breslow et al. (1984) have studied the statistic given by the maximum of the
log-rank statistic and a statistic representing the score associated with a covari-
ate, to test the null hypothesis of absence of group and covariate effects. Rather
than considering the previous statistics at the end of the study, one idea is to
work with their supremum over time. The log-rank statistic can then be seen
as a process tracing the history of the effect’s strength. If the effect depends
on time, deviation from the null hypothesis can be more easily detected. The
introduction of weights—the choice of which is always delicate—is possible but
not necessary for detecting time-dependent effects. These tests are called Rényi-
type tests, i.e., modified Kolmogorov-Smirnov tests for comparing distributions
of censored data. Their asymptotic and small-sample behavior have been stud-
ied by several authors, both theoretically and via simulation (Fleming and Har-
rington, 1984; Fleming et al., 1980, 1987; Fleming and Harrington, 1991; Gill,
1980; Schumacher, 1984). Eng and Kosorok (2005) provide the formula for the
number of subjects necessary for the supremum version of the test. Note that
Schumacher (1984) also looked at Cramér-von Mises-type tests (Appendix C).

Kosorok and Lin (1999) have also contributed to the study of supremum-type
tests over time. They looked at both the supremum and infimum over time,
leading to a powerful test for order-based alternative hypotheses, like for exam-
ple H1 : SP (t) < ST (t), ∀t.
Weighted Kaplan-Meier statistics.

Tests other than log-rank-type ones have also been introduced for comparing
survival functions. Pepe and Fleming (1989) have proposed a class of tests based
on the differences between Kaplan-Meier estimators of survival functions. The
differences can either be weighted or unweighted. Shen and Cai (2001) have pro-
posed an adaptive version of these tests, where the statistic is the maximum over
several statistics from the class, with various weights. Lee (2011) has studied the
asymptotic behavior of these tests using simulation. Note however that such tests
are not rank tests, and can be affected by increasing monotone transformations
with respect to time. Moreover, even if such tests are powerful under alternative
hypotheses with ordered survival functions, e.g., H1 : ST (t) > SP (t), ∀t, their
power is low if the functions intersect. For these two reasons, we do not look
further at this type of testing here.
In the following section, we consider a class of tests that we believe essentially
solves all of the problems described above. These tests are constructed via the
regression effect process Un∗ . Several tests are looked at in greater detail, each
based on particular characteristics of the process. In principle there is almost no
limit to the variety of tests that we can create. This great level of generality
can be used to our advantage when we wish, and are able, to introduce into the
problem any knowledge we may have concerning the nature of the departure from
a null hypothesis of absence of effects. Combined tests are easily constructed
as well, and basing them on different aspects of the regression effect process
and its known limiting distribution, makes it very convenient when seeking to
establish joint distributions. It is also reassuring to know that well-known popular
tests, such as the log-rank and weighted log-rank tests, are readily obtained as
special cases of tests based on the regression effect process. Some finite sample
behavior by simulation Chauvel (2014), Chauvel and O’Quigley (2014) is recalled
in Section 11.8 and some real data are analyzed in Section 11.9.
11.4 Tests based on the regression effect process
The asymptotic behavior of the regression effect process Un∗ , given by Theorems
9.1 and 9.2, allows us to construct hypothesis tests on the value of the regression
parameter β(t). The null hypothesis H0 and alternative hypothesis H1 that we
wish to test are
11.4. Tests based on the regression effect process 309

H0 : ∀ t, β(t) = β0 , versus H1 : {β(t) − β0 }2 dt > 0,
T

where β0 is a fixed constant and T dt > 0. The reason for the slightly involved
expression for H1 is to overcome situations in which the test could be inconsis-
tent. It enables us to cater for cases where, for example, β(t) differs from β0 only
on a finite set of points or, to be more precise, on a set of time points of mea-
sure zero. Under H0 , the drift of the regression effect process Un∗ (β0 , ·) at β0 is
zero, and the process converges to standard Brownian motion. We can therefore
base tests on the properties of Brownian motion. When β0 = 0, this amounts to
testing for the absence of an effect between the various values of the covariate
Z. In particular, if the covariate is categorical and represents different treatment
groups, this means testing the presence of a treatment effect on survival. The
β0 = 0 case thus corresponds to the null hypothesis of the log-rank test.
Several possibilities are mentioned in the book of O’Quigley (2008), including
tests of the distance from the origin at time t, the maximal distance covered in
a given time interval, the integral of the Brownian motion, the supremum of a
Brownian bridge, reflected Brownian motion, and the arcsine law. We will take a
closer look at the first two of these, followed by a new proposition. In this section
we look at the univariate case with one covariate and one effect to test (p = 1),
before extending the results to the case of several coefficients.
The following, almost trivial, result turns out to be of great help in our theo-
retical investigations. It is that every non-proportional hazard model, λ(t|Z(t)) =
λ0 (t) exp{β(t)Z(t)}, is equivalent to a proportional hazards model.
Lemma 11.1. For given β(t) and covariate Z(t) there exists a constant β0
and time-dependent covariate Z ∗ (t) so that λ(t|Z(t)) = λ0 (t) exp{β(t)Z(t)} =
λ0 (t) exp{β0 Z ∗ (t)} .
There is no loss of generality if we take β0 = 1. The result is immedi-
ate upon introducing β0 = 0 and defining a time-dependent covariate Z ∗ (t) ≡
Z(t)β(t)β0−1 . The important thing to note is that we have the same λ0 (t) either
side of the equation and that, whatever the value of λ(t|Z(t)), for all values of
t, these values are exactly reproduced by either expression, i.e., we have equiv-
alence. Simple though the lemma is, it has strong and important ramifications.
It allows us to identify tests that are unbiased, consistent, and indeed, most
powerful in given situations. A non-proportional hazards effect can thus be made
equivalent to a proportional hazards one simply by multiplying the covariate Z
by β(t). The value of this may appear somewhat limited in that we do not know
the form or magnitude of β(t). However in terms of theoretical study, this simple
observation is very useful. It will allow us to identify the uniformly most powerful
test, generally unavailable to us, and to gauge how close in performance comes
any of those tests that are available in practice.
Our structure hinges on a useful result of Cox (1975). Under the non-
proportional hazards model with β(t), the increments of the process Un∗ are
centered with variance 1. From Proposition 9.3 we also have that these incre-
ments are uncorrelated, a key property in the derivation of the log-rank statistic.
Based on the regression effect process, two clear candidate statistics stand out
for testing the null hypothesis of no effect. The first is the “distance-traveled” at
time t test and the second is the area under the curve test. Under proportional
hazards the correlation between these two test statistics approaches one as we
move away from the null. Even under the null this correlation is high and we
consider this below. Under non-proportional hazards the two tests behave very
differently. A combination of the two turns out to be particularly valuable for
testing the null against various non-proportional hazards alternatives, including
declining effects, delayed effects, and effects that change direction during the
study.
Distance traveled from the origin at time t

The distance traveled from the origin at time t, and in particular when t = 1,
appears as a logical, and intuitive, candidate for a test. Much more can be said
as we see in this section. Under the null, this distance, Un∗ (0, 1), has expectation
equal to zero. The area under the curve test is the area, positive and negative,
between the process and the x-axis. Again, under the null hypothesis this has
expectation equal to zero. Under proportional hazards, the test based on Un∗ (0, 1)
turns out to be the best available test for the following reasons:
Lemma 11.2. For 0 < s < 1 and under the null hypothesis, H0 : β(t) = 0,
√
Un∗ (0, s)/ s converges in distribution to Φ(z) where Φ(·) is the standard normal
distribution.
Corollary 11.1. Set z α such that α = 2(1 − Φ(z α )). The distance √ traveled at
time t test rejects H0 with a type I error α if |U ∗ (β , t) / t| ≥ z α/2 . The
√ 0
p-value for this test is given by 2 1 − Φ |U ∗ (β0 , t) |/ t .
Corollary 11.2. The distance traveled test is unbiased against proportional
hazards alternatives on the interval (0, t).
Corollary 11.3. Under the null hypothesis, H0 : β(t) = 0, the distance traveled
test is consistent against proportional hazards alternatives.
The lemma is an immediate consequence of the Donsker-Prokhorov invariance
principle. The corollaries follow from the Neyman-Pearson lemma as applied to
the standard normal distribution and one with the same variance but a translated
mean.
Corollary 11.4. Under H1 : β > 0 , assuming that kn /n → C as n → ∞, then,
for t > s
lim E Un∗ (0, t) − E Un∗ (0, s) = R(kn , β)(t − s), lim Var {Un∗ (0, t) − Un∗ (0, s)} = t − s
n→∞ n→∞
where R is monotonic increasing in kn for fixed β and in β for fixed kn .

Corollary 11.5. Under H1 : β > 0 , suppose that P(s) is the p-value for
√
Un∗ (0, s)/ s, then, assuming that kn /n → C as n → ∞, then, for t > s , E P(t) <
E P(s).
By applying the Neyman-Pearson lemma to the normal distribution, the above
two corollaries tell us that the distance traveled test is the uniformly most pow-
erful test of the null hypothesis, H0 : β = 0 against the alternative, H1 : β > 0.
The log-rank test, Ln shares this property and this is formalized in Proposition
11.2. Lemma 11.1 allows us to make a stronger conclusion. Assume the non-
proportional hazards model and consider the null hypothesis, H0 : β(t) = 0 for
all t. Then:
Lemma 11.3. The uniformly most powerful test of H0 : β(t) = 0 for all t, is
the distance traveled test applied to the proportional hazards model λ(t|Z(t)) =
λ0 (t) exp{β0 Z ∗ (t)} in which Z ∗ (t) = β0−1 β(t)Z(t) and where we take 0/0 to
be equal to 1.
Since, under the alternative, we may have no information on the size or shape
of β(t), this result is not of any immediate practical value. It can though act as
a bound, not unlike the Cramer-Rao bound for unbiased estimators and let us
judge how well any test is performing when compared with the optimal test in
that situation.
The following proposition states that the distance test statistic is consistent;
that is, the larger the sample, the better we will be able to detect the alternative
hypothesis of a proportional hazards nature. The result continues to hold for
tests of the null versus non-proportional hazards alternatives but restrictions
are needed, the most common one being that any non-null effect is monotonic
through time.
Proposition 11.1. Under the alternative hypothesis H1 : β(t) = β0 , that is,

under the non-proportional hazards model with parameter β(t) = β0 , we
have:
1
lim P √ |Un∗ (β0 , t) | ≥ z α/2 = 1.
n→∞ t
For a proof see the proof of Proposition 11.5.

Remark. The test can be applied at any time point t, making it flexible.
For example, if we have prior knowledge that the effect disappears beyond
some time point τ , we can apply the test over the time interval (0, τ ). When
such information is not available, we take t = 1, which, under proportional
hazards, will maximize the power.
Log-rank test as a distance traveled test

Instead of standardizing by V0 (Z|ti ) at each ti we might, instead,
use a simple
average of such quantities over the observed failures. Note that V0 (Z|t)dF̂ (t)
is consistent for E(var(Z|t)) under the proportional hazards model with β(t) = 0.
Rather than integrate with respect to dF̂ it is more common, in the counting
process context, to integrate with respect to dN̄ , the two coinciding in the
absence of censoring. If, we replace V0 (Z|s) by V̄0 (Z) then the distance traveled
test coincides exactly with the log-rank test. The following proposition gives an
asymptotic equivalence under the null between the log-rank and the distance
traveled tests:
Proposition 11.2. Let Ln denote the log-rank statistic and let Un∗ (0, 1) be
the distance statistic Un∗ (β0 , 1) evaluated at β0 = 0. We have that
P
|Un∗ (0, 1) − Ln | −→ 0.
n→∞
In order to obtain the result, we consider an asymptotic decomposition of the

statistic Ln in a Brownian motion with an added drift. The decomposition par-
allels that of the regression effect process at time t = 1. The conclusion is then
immediate. A convergence in probability result implies a convergence in distri-
bution result so that, under the conditions, the log-rank test and the distance
from the origin test can be considered to be equivalent, this equivalence holding
under the null and under local alternatives. Results for distant alternatives are
likely to be rather less straightforward, in particular since the censoring, whether
independent or conditionally independent, is likely to be involved.
Under the conditions we have stated, the distance test statistic converges to
that of the log-rank, guaranteeing maximum power under alternative hypotheses
of the proportional hazards type. This result is not surprising since the expres-
sions for the log-rank statistic Ln (Equation 11.1) and the distance test (Equation
9.5) differ only in the standardization used. For the former, standardization takes
place globally for all times at the end of the experiment, while for the latter,
standardization happens at each time of death. However, under the homoscedas-
ticity condition A3, the standardization does not change in expectation over time.
Recall that this condition, the use of which we justified in Section 9.4, has been
used by other authors (Grambsch and Therneau, 1994; Xu, 1996). The result
given in Proposition 11.2 also implies that the distance test is not optimal under
alternative hypotheses of a non-proportional hazards since the power properties
of the log-rank test also hold here. Many other tests can be based on the regres-
sion effect process Un∗ and properties of Brownian motion. We look at several of
these in the rest of this chapter.
Kolmogorov type tests

Since we are viewing the process U ∗ (β̂, β0 , t) as though it were a realization of a
Brownian motion, we can consider some other well-known functions of Brownian
motion. Consider then the bridged process U0∗ (β̂, β0 , t):
Definition 11.3. The bridged process is defined by the transformation
U0∗ (β̂, β0 , t) = U ∗ (β̂, β0 , t) − t × U ∗ (β̂, β0 , 1)
Lemma 11.4. The process U0∗ (β̂, β0 , t) converges in distribution to the Brownian
bridge, in particular E U0∗ (β̂, β0 , t) = 0 and Cov {U0∗ (β̂, β0 , s), U0∗ (β̂, β0 , t)} =
s(1 − t).
The Brownian bridge is also called tied-down Brownian motion for the obvious
reason that at t = 0 and t = 1 the process takes the value 0. Carrying out a test
at t = 1 will not then be particularly useful and it is more useful to consider, as
a test statistic, the greatest distance of the bridged process from the time axis.
We can then appeal to:
Lemma 11.5.

Pr sup |U0∗ (β̂, β0 , u)| ≥ a ≈ 2 exp(−2a2 ), (11.4)
u
which follows as a large sample result since

∞

Pr sup |U0∗ (β̂, β0 , u)| ≤ a → 1 − 2 (−1)k+1 exp(−2k 2 a2 ) , a ≥ 0.
u
k=1
This is an alternating sign series and therefore, if we stop the series at k = 2

then the error is bounded by 2 exp(−8a2 ) which for most values of a that we
will be interested in will be small enough to ignore. For alternatives to the null
hypothesis (β = 0) belonging to the proportional hazards class, the Brownian
bridge test is not generally a consistent test. The reason for this is that the
transformation to the Brownian bridge will have the effect of canceling the linear
drift term. The large sample distribution under both the null and the alternative
will be the same. Nonetheless, the test has value under alternatives of a non-
proportional hazards nature, in particular an alternative in which U ∗ (β̂, β0 , 1) is
close to zero, a situation we might anticipate when the hazard functions cross
over. If the transformation to a Brownian bridge results in a process that would
arise with small probability under the assumption of a Brownian bridge then this
is indicative of the effects of a non-proportional hazards nature.
Reflected Brownian motion

An interesting property of Brownian motion described in Appendix B is the fol-
lowing. Let W(t) be Brownian motion, choose some positive value r and define
the process Wr (t) in the following way: If W(t) < r then Wr (t) = W(t). If
W(t) ≥ r then Wr (t) = 2r − W(t). It turns out that the reflected process Wr (t)
is also Brownian motion (Appendix B). Choosing r to be negative and defin-
ing Wr (t) accordingly we have the same result. The process Wr (t) coincides
exactly with W(t) until such a time as a barrier is reached. We can imagine this
barrier as a mirror and beyond the barrier the process Wr (t) is a simple reflec-
tion of W(t). So, consider the process U r (β̂, β0 , t) defined to be U ∗ (β̂, β0 , t) if
|U ∗ (β̂, β0 , t)| < r and to be equal to 2r − U ∗ (β̂, β0 , t) if |U ∗ (β̂, β0 , t)| ≥ r.
Lemma 11.6. The process U r (β̂, β0 , t) converges in distribution to Brownian
motion, in particular, for large samples, E U r (β̂, β0 , t) = 0 and Cov {U r (β̂, β0 , s),
U r (β̂, β0 , t)} = s.
Under proportional hazards there is no obvious role to be played by U r .
However, imagine a non-proportional hazards alternative where the direction of
the effect reverses at some point, the so-called crossing hazards problem. The
statistic U ∗ (β̂, 0, t) would increase up to some point and then decrease back to a
value close to zero. If we knew this point, or had some reasons for guessing it in
advance, then we could work with U r (β̂, β0 , t) instead of U ∗ (β̂, β0 , t). A judicious
choice of the point of reflection would result in a test statistic that continues to
increase under such an alternative so that a distance from the origin test might
have reasonable power. In practice we may not have any ideas on a potential
point of reflection. We could then consider trying a whole class of points of
reflection and choosing that point which results in the greatest test statistic. We
require different inferential procedures for this.
A bound for a supremum-type test can be derived by applying the results of
Davies (1977, 1987). Under the alternative hypothesis we could imagine incre-
ments of the same sign being added together until the value r is reached, at
which point the sign of the increments changes. Under the alternative hypoth-
esis the absolute value of the increments is strictly greater than zero. Under
the null, r is not defined and, following the usual standardization, this set-
up fits in with that of Davies. We can define γr to be the time point satis-
fying U ∗ (β̂, β0 , γr ) = r. A two-sided test can then be based on the statistic
M = supr {|U r (β̂, β0 , 1)| : 0 ≤ γr ≤ 1}, so that:
1
exp(−c2 /2) 1
Pr {sup |U r (β̂, β0 , 1)| > c : 0 ≤ γr ≤ 1} ≤ Φ(−c) + {−ρ11 (γ)} 2 dγ
2π 0
where Φ denotes the cumulative normal distribution function,
ρ11 (γ) = {∂ 2 ρ(φ, γ)/∂φ2 }φ=γ
and where ρ(γr , γs ) is the autocorrelation function between the processes

U r (β̂, β0 , 1) and U s (β̂, β0 , 1). In general, the autocorrelation function ρ(φ, γ),
needed to evaluate the test statistic is unknown. However, it can be consistently
estimated using bootstrap resampling methods (O’Quigley and Pessione, 1989,
1991). For γr and γs taken as fixed, we can take bootstrap samples from which
several pairs of U r (β̂, β0 , 1) and U s (β̂, β0 , 1) can be obtained. Using these pairs,
an empirical, i.e., product moment, correlation coefficient can be calculated.
Under the usual conditions (Efron, 1981a,b,c, 1987; Efron and Stein, 1981), the
empirical estimate provides a consistent estimate of the true value. This sam-
pling strategy is further investigated in related work by O’Quigley and Natarajan
(2004). A simpler approximation is available (O’Quigley, 1994) and this has the
advantage that the autocorrelation is not needed. This may be written down as
√
Pr {sup |U r (β̂, β0 , 1)| > M } ≈ Φ(−M ) + 2−3/2 Vρ exp(−M 2 /2)/ π,(11.5)

where Vρ = i |U r (β̂, β0 , 1) − U s (β̂, β0 , 1)| , the γi , ranging over (L, U), are the
·) and M is the observed maximum of T (0, β;
turning points of T (0, β; ·). Turning
points only occur at the kn distinct failure times and, to keep the notation
consistent with that of the next section, it suffices to take γi , i = 2, . . . , kn , as
being located half way between adjacent failures. To this set we can add γ1 = 0
and γkn +1 to be any value greater than the largest failure time, both resulting
in the usual constant estimator.
Area under the curve

We now define the area under the regression effect process at time t.
Definition 11.4. For any γ ∈ B, the process representing the area under the
standardized score function at time t, denoted {Jn (γ, t), 0 ≤ t ≤ 1}, is defined
by t
Jn (γ, t) = Un∗ (γ, u)du, 0 ≤ t ≤ 1.
0
Theorem 9.1 and properties of Brownian motion (Bhattacharya and Waymire,
1990) allow us to establish the following result.
Proposition 11.3. Under the proportional hazards model with parameter

β0 , the process representing the area under the standardized score function
converges in distribution to an integrated Brownian motion:
t
L
Jn (β0 , t) −→ W(s)ds, t > 0.
n→∞ 0
Furthermore, the covariance of the process is given by

t s
lim Cov {Jn (β0 , s) , Jn (β0 , t)} = s2 − , 0 ≤ s ≤ t ≤ 1. (11.6)
n→∞ 2 6
t
Recall that the random variable 0 W(s)ds has a centered normal distribution
with variance t3 /3. The proposition above allows us to define a test using the
area under the score process at time t.
Proposition 11.4. Let t ∈ [0, 1]. The statistic for the area under the standardized
score process at time t is Jn (β0 , t). The test for the area under the curve at
time t rejects H0 at asymptotic level α if
1/2
3t−3|Jn (β0 , t) | ≥ z α/2 .

1/2
The asymptotic p-value of the test is 2 1 − Φ 3t−3 |Jn (β0 , t)| .
Remark. As for the distance from origin traveled, we are at liberty to choose
values for t other than t = 1. This opens up the possibility of a very wide
range of potential tests. For now, we only consider t = 1, and to keep the
text simple, we will simply call the test the “area under the curve test”. The
following proposition indicates that the test statistic for the area under the
curve is consistent.
Proposition 11.5. Under the alternative hypothesis H1 : β(t) = β0 , that is,

under the non-proportional hazards model with parameter β(t) = β0 , we
have: 1/2

lim P 3t−3 |Jn (β0 , t) | ≥ z α/2 = 1.
n→∞
Under proportional hazards alternative hypotheses, the distance test has a

power slightly greater than the area under the curve test since it is asymptotically
equivalent to the log-rank test. On the other hand, when the effect β(t) decreases
with time, the area under the curve test will become the more powerful of the
two. This can be seen in an example involving a change-point model in which
the effect is constant up to a certain time τ and vanishes afterwards (see Figure
9.4(a)). Until time τ , the drift of the process Un∗ (β0 , ·) is linear with a strictly
positive slope, but after τ , the drift is approximately parallel to the time axis,
i.e., close to zero. A test based on the distance from the origin will only take into
account the last value of the process, and not the strength of the effect before
τ . This will lessen the power of the distance test compared to that of the area
under the curve test, in which all the information collected over the evolution of
the process is retained.
Remark. According to Propositions 11.1 and 11.5, the distance and area
under the curve statistics are both consistent. Using the covariance function
given below we can work out the√Pitman asymptotic efficiency (see Appendix
D.2). This is, for all t, given by 3/2 = 0.87. A comparison of the test’s pow-
ers for finite-sized samples confirms this finding in Section 11.8. This result
could be used to guide sample size calculations in a clinical trial where there
is concern that, if the null does not hold, there is a high chance that the
alternative will display non-proportional hazards behavior. We could protect

ourselves against losing power by increasing the sample size by 1/0.87 = 1.15,
i.e., by 15% and using the area under the curve test. If, indeed the alterna-
tive manifests proportional hazards behavior then our increased sample size
assures that we keep the same power we would have had with the original
sample and the original test. If on the other hand the alternative is of a non-
proportional hazards nature then we could anticipate large gains in power
when we witness effects waning with time. The following lemma provides the
result needed for the above to hold. (Ross, 1996):
Figure 11.2: Under proportional hazards the upper two figures indicate very close
agreement between log-rank and AUC. Lower curves indicate substantial gains
of AUC over log-rank when effects diminish with time.
Lemma 11.7. Let t ∈ [0, 1]. Under the proportional hazards model with
parameter β0 , the covariance function of Jn (β0 , t) and Un∗ (β0 , t) is such
that
P
Cov {Jn (β0 , t), Un∗ (β0 , t)} −→ t2 /2.
n→∞
The result opens up the possibility for whole classes of tests based on different
ways of combining the component contributions. We can build tests based on
different combinations of the two that can have good power under both types
of alternative hypothesis (proportional and non-proportional hazards). Below, we
define a class of adaptive tests that we describe as restrictive adaptive tests. They
combine the distance statistic Un∗ (β0 , t) with the area under the curve statistic
Jn (β0 , t). The user is allowed to bias the choice, hence the term restrictive.
The resulting statistic takes the value of one or the other of the tests. The
class is adaptive in the sense that users do not directly choose which statistic
to apply; instead, it is chosen according to the data, with a parameter set by
the user. We also present an unrestricted adaptive test that does not depend on
any constraining parameters. We can control the degree to which we want the
data to, in some sense, have their say. At one extreme, no amount of data will
convince us to move away from the distance traveled (log-rank) test, whereas,
on the other hand, a more even handed approach might specify no preference for
either test.
Figure 11.3: Two examples of the regression effect process Un∗ (0, ·) under the
null, β = 0. When reading such a figure, note the different scales.
Integrated log-rank test

Proposition 11.2 shows the large sample equivalence between the log-rank test
and the distance from origin test. An analogous result can be shown in the same
way to establish the equivalence between the area under the curve test and one
based on integrating the log-rank process over the interval (0,1). The distinction
between these two tests is a minor one but worth having in order to complete the
picture of available tests. A test based on integrating the log-rank process over
the interval (0,1) is referred to as the integrated log-rank test. It will offer no
advantage, nor disadvantage, over the area under the curve test under the null
and in the vicinity of the null and so is not studied deeply in its own right. Since
the absolute difference between the area under the curve test and the integrated
log-rank test converges in probability to zero, under the usual conditions on n,
kn , and the censoring, we may sometimes refer to them as essentially the same
thing. The term “integrated log-rank test” may be used informally in either case.
We can glean a lot from the visual impression given by Figure 11.2. The
upper two processes correspond to a case arising under proportional hazards.
The most efficient estimate of the slope of a Brownian motion with linear drift
is given by connecting the origin (0,0) to the point of arrival of the process at
t = 1. The triangular area under this slope corresponds to precisely 0.5 multiplied
by the distance from origin (log-rank test). The variance of this would be 0.25
times the variance of the log-rank statistic and so we see, immediately, that
the area of the triangle can itself be taken to be the log-rank test. Now, under
proportional hazards, we have linear drift so that the integrated log-rank test—
the area under the process—will be almost identical in value to the log-rank test.
The triangular area and the area under the process almost coincide. This is well
confirmed in simulations and, even when the drift is equal to zero, i.e., under
the null hypothesis, the correlation between the log-rank test and the integrated
log-rank test is very high. Let us now consider the lower two graphs of Figure
11.2. We have no difficulty in seeing that the triangular area in this case will be
quite different from the area under the process. Not only that, but in cases like
these, where this area corresponds to the test statistic, it is clear that we will lose
a lot of power by working with the triangular area (log-rank test) rather than
the area under the process (integrated log-rank test). The differences in these
two areas could be large. The lower graphs indicate an effect that diminishes
with time. A yet stronger illustration is presented in Figure 11.4(b) where we can
see that a very great loss of power would be consequent upon the choice of a
statistic based on the triangular area (log-rank) rather than all of the area under
the curve (Fig. 11.3).
Note also that it is easy to encounter such a diminishing effect, even under
proportional hazards. Suppose, for example, that some percentage of the treated
group fails to stick with the treatment. This may be due to undesirable secondary
effects or a lack of conviction concerning the potential of the treatment. Either
way, we will see a dilution of effect through time resulting in a heterogeneous
rather than a homogeneous sample. The impact of this can be a regression process
where the effect appears to diminish with time. When so, it is very likely that the
integrated log-rank would be a more powerful choice of test than the log-rank
test. As a result, in very many practical situations, it can be more logical to
take the integrated log-rank test, rather than the log-rank test, as our working
standard. There are several other possibilities, many of which remain open to
exploration. Combining the log-rank test with the integrated log-rank test can
also provide a valuable tool, in particular in situations where we anticipate a
proportion of non-responders. We look at this in the following section.
Area over the curve

Whereas the area under the curve is easy to visualize, the area over the curve
needs a bit more thought. And, just as the area under the curve will provide
the basis for a test close to optimal when regression effects dissipate with time,
the area over the curve shows itself to be particularly effective when regression
effects are delayed. As well as Un∗ (β(t), t) , and the drift of this process under
the
t null and alternative hypotheses, we are also interested in the behavior of
U ∗ (β(u), u) du.
0 n
Definition 11.5. The area over the curve (AOC) test statistic Tn (β0 , 1) is
given by;
t
Tn (β(t), t) = Un∗ (β(t), t) − Un∗ (β(u), u) du (11.7)
0
Note that the time scale is on the interval (0,1) so that Un (0, 1) corresponds to
the rectangular area formed by the coordinates (0,1) and (0, Un (0, 1)). The area
1
below the regression effect process is given by 0 Un∗ (0, u) du so that subtracting
this from Un (0, 1) results in the shaded area. This area is our test statistic. Before
considering more precisely the statistical properties of the test statistic we can
build our intuition by making some observations. The expected value of the test
statistic under H0 : β(t) = 0 , t ∈ (0, 1), will be zero. This is because the expected
value of the regression effect process and the area under its path are both zero
under H0 . These two quantities are correlated but that does not impact the
Figure 11.4: The regression effect process Un∗ (0, ·) under PH and strong, non-
proportional hazards (NPH) effect. It is clear that, under PH, log-rank test (tri-
angular area) will be close to AUC. Under NPH, the area under the curve will
provide a more powerful test statistic.
expectation. It will impact the variance and we consider this in the next section.
Under a proportional hazards departure from H0 we anticipate a linear drift in
the regression effect process. The test statistic will then be well approximated by
one half of the area given by Un (0, 1) which is, in turn, equivalent to the log-rank
statistic. The test will then be very close to the log-rank test and a small amount
of power is lost. For a delayed effect, as long as it is not too weak and allows
Un (0, 1) to drift ultimately away from the origin, we will obtain a large value of
the test statistic, all the more so as Un (0, 1) finally climbs away from the region
close to the axis.
Lemma 11.8. Under the model, with parameter β0 (t), Tn (β0 (t), t) is a
Gaussian process for which we have; E Tn (β0 (t), t) = 0 and Var Tn (β0 (t), t) =
t(1 − t) + t3 /3. In particular, we have that Var Tn (β0 (t), 1) = 1/3.
Tests based upon Tn (β0 (t), t) would control for Type I error and would be antic-
ipated to have good power, albeit suboptimal, in proportional hazards situations
and to have high power for delayed effect alternatives.
Figure 11.5: Kaplan-Meier curves and corresponding regression effect process

Un∗ (0, ·). The area over the curve is used as the test statistic.
Figure 11.5 is taken from Flandre and O’Quigley (2019) and shows how the
test statistic might look under an alternative of a delayed effect. This study is
fully described in Hodi et al. (2018). For a period not far short of the median
survival times there appears to be no effect. Beyond this the effect appears to be
a significant one that would tend to rule out the plausibility of the null hypothesis
of absence of effect.
Delayed transient effects

The area under the curve and the area over the curve tests, described above,
will exhibit good power under situations of waning effects and delayed effects
respectively. These tests would be competitors for the Harrington-Fleming class

of tests investigated by Lee (1996). These were described under the above head-
ing “Combining log-rank and weighted log-rank tests”. Lee considered a linear
combination of four weighted log-rank statistics with weights from the family
given in Fleming and Harrington (1991). Recall that Gp,q was the weighted
log-rank statistic with weights Ŝ(t)p (1 − Ŝ(t))q , p, q ≥ 0 and that G0,0 = Ln ,
whereas G2,0 was geared to picking up early effects, G0,2 was geared to picking
up late effects and G2,2 for effects in the “middle of the study.”
This is somewhat vague but let’s suppose that it corresponds to transient
effects that do not kick in immediately. The circumstance may not be such
a common one and it is not one that we have studied. We would anticipate
the regression effect process to start out looking like standard Brownian motion
with no drift, followed by a period during which a regression effect may become
manifest and then a later period looking again like Brownian motion, although,
now, shifted with respect to the time axis. Given the anticipated form of the
regression effect process we speculate that a test based on the maximum distance
traveled of the Brownian bridge process, i.e., the process obtained following the
transformation to a tied-down process, would exhibit good power properties.
This has not been studied, and in the absence of a compelling application, may
continue to remain neglected. Certainly the situation of waning effects or of
delayed effects is much more likely to be encountered in practice.
Distribution of signed process

Although mostly of theoretical interest the time that the regression effect process
remains on one side of the time axis is very readily observed. Under the null
hypothesis of no effect, the fact that the regression effect process converges in law
to Brownian motion allows us to immediately deduce the large sample behavior
of this quantity. Let us call Δ(u) the amount of time that the process is above
(below) the time axis by time t = u. We have the following simple proposition:
Proposition 11.6. Suppose that Δ(u) corresponds to the percentage of

time that the process, Un∗ (β(t), t) is greater than zero. Then:
√
lim P [{Δ(u) : Un∗ (β(t), t) > 0} > w] = 2 × π −1 sin−1 w.
n→∞
This follows as an immediate application of the arcsine result for Brownian

motion. Despite its simplicity the test does not appear to be that useful. It
is readily seen to be unbiased and consistent but, in practice, distinguishing the
null from the alternative is not easy. There are a couple of reasons for this. The
power curve is very steep very close to zero and one but, otherwise, power is low.
Under the null it is quite possible and not rare to spend a very large percentage
of time on one side of the time axis.
By plugging in a couple of numbers, we see for example, that, if only one

percent of the time the process finds itself the other side of the time axis then
the probability for this, under the null, is around 0.06, not even achieving the 5%
limit of significance. The second difficulty is that this is, again, a large sample
result. For around one hundred failures, only one of which corresponds to a part
of the curve being the other side of the time axis, we could not claim a significant
result. The approximation is likely to be poor then for sample sizes (failures) not
well into the hundreds. We know, of course, by standard theory, that, under the
null, the process will return to cross the time axis with probability one. We may
have to wait a long time though. What could be said, in practical cases, is that,
if the whole of the process is located on one side of the time axis, then the test
would have a p-value of zero, and without wishing to put too much weight on
that, there does appear to be evidence of some kind of effect.
Restricted adaptive tests

Chauvel and O’Quigley (2014) consider restricted adaptive tests in which we
limit the class of possibilities under consideration, more or less favoring partic-
ular members of the class. For instance, we may prefer to use the log-rank test
unless there are strong indications of non-proportionality of the hazards. We can
introduce a parameter γ to favor one test over the others. In our case, a large
value of γ favors the log-rank test, resulting in a negligible loss of power under
the proportional hazards hypothesis compared to direct application of the log-
rank test. The downside is low power for detecting effects that wane or disappear
through time. Smaller values of γ make it easier to detect effects when hazards
are not proportional.
Under proportional hazards-type alternative hypotheses H1,P H , the process
Un∗ (β0 , ·) can be approximated by a Brownian motion with linear drift, according
to Theorem 9.2. Thus, at time t, twice the integrated process Jn (β0 , t) will be
close in value to the distance from the origin Un∗ (β0 , t). Next, denote by Δn (β0 , t)
the difference between these two processes:
Δn (β0 , t) = 2Jn (β0 , t) − Un∗ (β0 , t).
If the absolute value of Δn (β0 , t) is fairly small we can expect the hazards to be
proportional, and the distance test or, equally so, the log-rank test (Proposition
11.2) to be powerful. If not, the area under the curve test is likely to be more
powerful. In this way, we define a class of restricted adaptive tests that depend
on the threshold γ ≥ 0.
Definition 11.6. The class of statistics for restricted adaptive tests at time
t is denoted Mnγ (β0 , t), γ ≥ 0 , where for γ ≥ 0,
√
Mnγ (β0 , t) = 3|Jn (β0 , t)| − |Un∗ (β0 , t)| 1|Δ (β , t)| ≥ γ + |Un∗ (β0 , t)|.
n 0
We have choice in which value of t to work with, and if effects are known to
drop to zero beyond some point τ then it makes sense to choose t = τ. In most
situations, and wishing to make full use of all of the observations, we will typically
choose t = 1.
Remark. The p-value of the test can be calculated numerically by simulating

N independent pairs (X1 , Y1 ), . . . , (XN , YN ) of centered normal distributions
in R2 with variance-covariance matrix

t t2 /2
.
t2 /2 t3 /3
The p-value can then be taken to be:

N
1
1f (X , Y ) ≥ M γ (β , t) ,
N i i n 0
i=1
where for i = 1, . . . , N ,
√
f (Xi , Yi ) = 3|Xi | − |Yi | 1|2X − Y | ≥ γ + |Yi |.
i i
When γ = 0, Mnγ (β0 , t) is the absolute value of the area under the curve.
When γ → ∞, Mnγ (β0 , t) tends to the absolute value of the distance from the
origin. Between the two, Mnγ (β0 , t) is equal to the absolute value of one or the
other. The optimal choice of γ depends on the unknown form of the alternative
hypothesis H1,P H or H1,N P H . One strategy is to work with a large value of
γ, which essentially amounts to applying the log-rank test and keeping some
power in reserve for unanticipated alternative hypotheses in which the effect
changes strongly over time (for example, a change of sign). However, this test
will be less powerful than the log-rank in detecting the presence of a non-zero
constant effect. If we have no wish to give priority to either of the alternative
hypotheses, rather than estimating the optimal parameter γ and losing power to
detect the alternative hypothesis, we could simply just take the test that provides
the smallest p-value. Of course this then requires adjustment in order to maintain
control on Type 1 error.
Definition 11.7. The statistic for a simple unrestricted adaptive test, in

which no parameter γ is specified, can be written as:

∗
√
Mn (β0 , t) = max |Un (β0 , t)| , 3|Jn (β0 , t)| .
Proposition 11.7. (Chauvel and O’Quigley 2014). Denote φ{·; 0, Σ(t)} the
density of the centered normal distribution in R2 with variance-covariance matrix
√
t 3t2 /2
Σ(t) = √ 2 .
3t /2 t3
The unrestricted adaptive test rejects H0 at asymptotic level α if Mn (β0 , t) ≥ q α ,

where q α is the upper quantile of N (0, Σ(t)) such that:
qα qα
1−2 φ{u, v; 0, Σ(t)}dudv = α.
0 0
The p-value of the test is

Mn (β0 ,t) Mn (β0 ,t)
1−2 φ{u, v; 0, Σ(t)}dudv.
0 0
The p-value can be calculated numerically using simulation by adapting the

method described earlier in the section. The test is adaptive in the sense that
the data will determine which of the two statistics will be used. The possibili-
ties are limited—only two choices are available—but could be readily extended
to a broader range of situations if required. The performance of the test will
lie between that of the distance test and the area under the curve test. Thus,
the alternative hypothesis H1 : β(t) = β0 is detected well in general, whether
the effect depends on time or not. Naturally, if we truly are under proportional
hazards then we anticipate a small loss in power when compared with the dis-
tance traveled or log-rank test. As hazards become more non-proportional, and
specifically if the hazard ratio declines with time, then, again, the adaptive test
will lose a small amount of power compared to the area under the curve test. It
is not possible to find a test that is the best under all circumstances and the way
to view the adaptive test here is as a good compromise test that covers several
possible situations. Unless we have very strong prior knowledge about which of
these situations most likely prevails, then, on average, the adaptive test will fare
very well. We next consider some details concerning the multivariate setting.
Tests based on multivariate regression effect processes

Consider a p-dimensional vector of covariates:

Z(t) = Z (1) (t), Z (2) (t), . . . , Z (p) (t)
and recall that we have already studied the multivariate standardized score pro-
cess in Section 9.5. Here, we suppose that B1, B1, B3 and B4 hold, and that
β, β0 ∈ B . The standardized score process is a random function from [0, 1] to
Rp , written

Un∗ (β0 , t) = Un∗ (β0 , t)1 , Un∗ (β0 , t)2 , . . . , Un∗ (β0 , t)p , 0 ≤ t ≤ 1.
By integrating each component of this process with respect to time, we can also
define the area under the curve process of the multivariate standardized score
process.
Definition 11.8. The area under the curve process of the multivariate standard-
ized score process is a random function from [0, 1] to Rp such that
t t t
Jn (β0 , t) = Un∗ (β0 , s)1 ds, Un∗ (β0 , s)2 ds, . . . , Un∗ (β0 , s)p ds , 0 ≤ t ≤ 1.
0 0 0
Under H0 : β(t) = β0 , Theorem 9.3 implies the following convergence:

L L
Un∗ (β0 , ·) −→ Wp , Jn (β0 , ·) −→ Wp (s)ds.
n→∞ n→∞
To simplify notation, without loss of generality we consider p = 2. Equations 11.6

and Lemma 11.7, which describe the asymptotic behavior of covariance functions
between various random quantities, imply the following result.
Proposition 11.8. (Chauvel and O’Quigley 2014). Let t ∈ [0, 1]. Under the
null hypothesis H0 : β = β0 , that is, under the PH model with parameter β0 , the
vector T
√ √
Un∗ (β0 , t)1 , Un∗ (β0 , t)2 , 3Jn (β0 , t)1 , 3Jn (β0 , t)2
converges in distribution to a centered Gaussian vector with variance-covariance

matrix Σ̃t given by:
⎛ √ 2 ⎞
t 0 3t /2 √ 0
⎜ 0 t 0 3t2 /2⎟
Σ̃t = ⎜ √
⎝ 3t2 /2
⎟.
√ 02 t 3 0 ⎠
0 3t /2 0 t3
We can extend this multivariate result to the distance from the origin, area
under the curve, and restricted, or unrestricted, adaptive tests, in order to test
the null hypothesis H0 : β(t) = β0 against H1 : β(t) = β0 . The first corollary
extends the test to the distance from the origin.
Corollary 11.6. The distance from the origin statistic is given by:
Un (β0 , t) = t−1 Un∗ (β0 , t)21 + Un∗ (β0 , t)22 .
Under H0 : β(t) = β0 , Un (β0 , t) converges in distribution to a chi-squared dis-

tribution with two degrees of freedom χ22 when n → ∞. The distance from
the origin test rejects H0 at asymptotic level α if Un (β0 , t) ≥ Q2 (α), where
P χ22 ≥ Q2 (α) = α.
The next corollary deals with the multivariate area under the curve.
11.5. Choosing the best test statistic 327
Corollary 11.7. The area under the curve statistic is given by:
Jn (β0 , t) = 3t−3 Jn (β0 , t)21 + Jn (β0 , t)22 .
Under H0 : β(t) = β0 , Jn (β0 , t) converges in distribution to a χ22 when n → ∞.

curve test rejects H0 at asymptotic level α if Jn (β0 , t) ≥
The area under the
Q2 (α), where P χ22 ≥ Q2 (α) = α.
Lastly, we also extend the unrestricted adaptive test to the multivariate case.
Corollary 11.8. Denote Mn (β0 , t) the statistic for the multivariate unrestricted
adaptive test, where
Mn (β0 , t) = max {Un (β0 , t), Jn (β0 , t)} .
This test rejects H0 at asymptotic level α if Mn (β0 , t) ≥ Q(α), where

1− 1p2 +q2 ≤Q(α),r2 +s2 ≤Q(α) φ(p, q, r, s; 0, Σ̃1 )dpdqdrds = α,
R4
and φ(· ; 0, Σ̃1 ) is the density of the normal distribution N4 (0, Σ̃1 ). The p-value
of the test is:

1− 1p2 +q2 ≤Mn (β0 ,t),r2 +s2 ≤Mn (β0 ,t) φ(p, q, r, s; 0, Σ̃1 )dpdqdrds.
R4
Conceptually the multivariate situation is no more complex than the univariate

one. The difficulty lies in deciding how to treat the information from the several
univariate processes. This will often boil down to a consideration of the kinds of
assumptions we wish to make. For instance, testing the impact of one variable
while controlling for a second variable through stratification makes a weaker
assumption on the whole model structure than assuming that both regression
effect processes would be linear, whether under the null or the alternative. When
the assumption is a fairly accurate one then we would anticipate gaining a little
in power by working under the assumption than not. How much the gain may
need to be studied on a case by case basis.
11.5 Choosing the best test statistic
In a planned experiment we will work out our sample size and other design
features in line with statistical requirements such as control of Type 1 error and
power. These, in turn, lean upon some postulated possibilities, both under the null
and under the alternative. As we have seen, if the alternative differs significantly
from that postulated we can lose a lot of power. We may be working with a
promising treatment but our inability to see just how the patient population
will respond to the treatment can easily result in a failed study. We would like
to minimize the risk of such failures. This is a goal that the results of this
chapter show to be achievable if we have information on time dependencies under
the alternative hypothesis. Unfortunately our knowledge here may be imprecise.
Nonetheless, even if we fall a little short of some optimal goal, we can still reduce
the risk of trial failure by choosing our test to accommodate effectively plausible
outcomes. We look at some broad considerations in this section.
Proportional hazards alternatives

When the relative risk does not depend on time we know from the theory of
earlier chapters that the log-rank test—equivalently the distance from the origin
test—will be the uniformly most powerful unbiased test; the test of choice in situ-
ations of proportional hazards. The question then is, how plausible is it to assume
proportional hazards. In many cases this would be a reasonable assumption.
The physiological impact of a treatment may confer some advantage to all of
those being treated, and in such a way, that whenever we consider a comparative
study of two groups, one of which has received the treatment while the other
group, otherwise homogeneous, received some other treatment, the advantage
remains a constant one. This ought be a not uncommon occurrence, the advan-
tage, in some sense, being long lasting. If such a set-up appears to describe the
study in hand, at least approximately, then the log-rank test will, most likely, be
the best choice we could make. This partly explains its popularity. In epidemio-
logical applications it is very common for the strongest risk factor by far to be
age. Other risk factors may be an order of magnitude weaker and interact, if
at all, very weakly with age. Once more, a sensible working assumption would
be one of proportional hazards and the best test available to us the log-rank or
distance from the origin test (Fig. 11.6).
It is generally admitted that the data used to illustrate Cox’s method, the
Freireich data, come from a clinical trial where a proportional hazards constraint
could reasonably be assumed to hold. This is confirmed by looking at the Brow-
nian bridge transform of the regression effect process shown in Figure 11.7. For
these data the several ties were randomly split, two such splits being illustrated
in Figure 11.7.
Potential departures from this basic set-up need careful consideration. One
example would be where we have a treatment group that contains a mixture
of responders and non-responders. The responders would behave like subjects in
the reference group. The ratio of the responders to all members of the treat-
ment group will change through time. This will tend to express itself as non-
proportional hazards of the convex kind outlined just below.
Concave non-proportional hazards alternatives

In our formulation, a proportional hazards model is simply a special case of
a non-proportional hazards one. It might be considered to be a proportional
Figure 11.6: Test based on PH assumption near optimal for left hand figure, and
near optimal after recoding via a cutpoint model for right hand figure.
and a non-proportional hazards model simultaneously. Informally though we will

suppose a non-proportional hazards model to be one where the log risk function,
β(t), is not constant in t. This non-proportional class is indeed very large. We
can nonetheless identify at least two very important, and substantial, subclasses
within this general class. We refer to these as the concave and the convex classes.
Let’s deal first with the concave class. The concave class will correspond to
situations where the regression effect diminishes with time. Figure 11.8 shows
an example taken from the Curie Institute breast cancer study. The non-linearity
of the regression effect process is clear and the process lies entirely above the
line connecting the arrival point to the origin. The area under the curve test—
equivalently the integrated log-rank test—is likely to outperform the log-rank
test. This is apparent by considering the area between the process and the straight
line. If negligible then we are close to proportional hazards anyway and little, if
anything, will be lost. If non-negligible then making use of the log-rank test
means that we may well miss effects that are of importance although of a non-
proportional hazards nature.
Convex non-proportional hazards alternatives

Another large class of non-proportional hazard effects can be described as convex
effects. In a way very much analogous to concave effects, if we linearly connect
the arrival point of the process to the origin and find that all of the points on the
line lie above the regression effect process then this is a case of convex effects.
The area under the curve will be a test with behavior yet worse than the log-rank
test since the area between the line and the process is removed, rather than
added, to our test. On the other hand, the area over the curve will be, in general,
Figure 11.7: Regression effect process transformed to a bridge process for the
Freireich data via Wn∗ (t) = Un∗ (β̂, t) − tUn∗ (β̂, 1). The process lies well within the
limits of the supremum of the limiting process, the Brownian bridge.
a more powerful test than either of these two. We consider this below. Before
looking at some particular cases note that the concepts convex and concave,
in this context, while inspired by geometrical concepts can differ slightly. Small
departures from the assumption will imply small departures from test behavior
so, if a curve is mostly concave, then, by and large, it will still be advantageous
to make use of the area under the curve test.
Delayed treatment effects

Suppose, as is frequently observed in immunotherapy trials, that each subject
has a period of varying length during which time no effect is observed. The
effect would kick in at some point once triggered for each individual. We could
postulate, for patient i, an effect βi (t) such that:
λi (t|Z) = λ0 (t) exp{βi (t)Z} = λ0 (t) exp{Wi (t)β0 (t)Z}, (11.8)
where Wi (t) = I(t > τi ). Each patient has their own specific time point τi at
which the effect begins to take a hold. At time t, the percentage of “trigger
points” τi , less than t, will be an increasing function of t, say G(t) and, for the
effect β(t), governing the whole study group:
β(t) = Eβi (t) = E[Wi (t)β0 ] = β0 × Pr (τi ≤ t) = β0 G(t).
The overall impression will be that of a non-proportional hazards effect whereby

the between time points, t = 0 and t = min τi , i = 1, ..., n, the treated group does
not differ to the control group but, beyond which, we would observe a gradual
increase in overall effect. If we knew the form of G(t), we could derive an optimal
test. While this is not available to us, we can obtain a test that will generally
be powerful using the area over the curve statistic. Figure 11.9 is somewhat
Figure 11.8: When the points linearly connecting the arrival point of the process
to the origin lie wholly below the regression effect process we call this concave
effects.
idealized in supposing an initial time period of no effect, common to the whole

group, and beyond which all of the treated group are triggered. The model just
described is no doubt more realistic in postulating a distribution for the triggered
times. The anticipated form of the regression effect process comes under the
heading of convex non-proportional hazards. Again, the area over the curve will
provide a statistic generally more powerful than that of the log-rank test since
the triangular area corresponding to the log-rank test is most likely smaller.
Erosion of regression effect

Suppose that β(t) > 0 for all t. Furthermore, suppose that β(t) is a non-increasing
regression function whereby late in the study it takes on values that are non-
negligibly less than the early values. This will lead to a regression effect process
that is concave in appearance. The kind of situation we will encounter will often
correspond to that illustrated in Figure 11.4(b), possibly with a less pronounced
effect. Nonetheless, as long as the triangular area is notably smaller than the area
under the curve, then the AUC test is to be preferred. We may often anticipate
that effects will diminish through time and, if so, a test based on AUC is likely to
be a better choice than the log-rank test. All of this remains true when β(t) < t
for all t, only now we take the area which is negative to be that between the
process and the time axis, i.e., the area under the curve although, in practice,
above the regression process. Concave processes are more likely to be encountered
in retrospective studies focused on risk factors than in randomized clinical trials.
Consider the Curie breast cancer study (Figure 11.10) where the risk factors grade
and tumor size both show non-proportional hazards effects. One explanation
would be that there is non-negligible error in these kinds of classifications so that
Figure 11.9: An idealized model for immunotherapy effects, absent initially and
then manifesting themselves at and beyond some approximate time point.
the relative balance between groups stabilizes through time. In the long run the
groups will look more similar than they did initially.
Presence of non-responders
Consider a clinical trial for which under the usual arguments, such as those laid
out above, we could take a proportional hazards assumption to be a reasonable
one. In this case the distance from the origin (log-rank) test would be close to
optimal. This assumes however that we are dealing with homogeneous groups; a
treated group and a, for the sake of argument, placebo group. Suppose though
that among those treated there exists a subset of patients who will not be sus-
ceptible to the treatment’s action. For these patients they behave similarly to
patients from the placebo group. The treatment group is not then homogeneous
and is a mixture of both types of patients. Under the alternative hypothesis, the
prognostic impact of belonging to the treatment group will grow in strength.
The reason for this is that the composition of the treatment group will change
through time, losing more of the non responding patients within this group with
respect to the responders.
The relative balance within the treatment group of responders to non-
responders increases with time. The regression effect process will also appear
to increase with time. The regression effect process will look similar to that we
obtain with delayed treatment effect, an example being immunotherapy trials.
Knowing, even if only roughly, how many non-responders we may have at the
outset, would allow us to construct a close to optimal test. Without such knowl-
edge we can still improve on the log-rank test by noting that we will be observing
effects of a convex nature. The area over the curve test would be a good choice
in such cases, in particular when the percentage of non-responders is likely to be
quite large.
Figure 11.10: Curie Institute breast cancer study. Prognostic factors grade and
tumor size both show evidence of waning effects. Grade lies just within the 95%
PH limits. Size clearly violates the PH assumption.
Categorization errors and misclassification

While it is not exceptionally rare for a patient in a clinical trial to end up in
a group different from what he/she was initially assigned to, such cases would
have a negligible effect on subsequent tests. For any given trial there are likely
to be no more than one or very few such occurrences. For descriptive studies
this can be quite different. In prognostic studies we will often find ourselves
dealing with categorizations that are not particularly sharp. Tumor grade, tumor
size, stage, and several other indicators may well be clear cut on either end of the
spectrum, but for those middle categories, it can be very difficult to classify accu-
rately. As a result, prognostic studies will often include variables that will have a
non-negligible subset of subjects who may not be assigned the most appropriate
group. This situation is not altogether unlike that for non-responders although,
here, we would anticipate seeing concave as opposed to convex effects. It could
indeed go either way and each case ought to be considered on its own merits.
An example is seen in the Curie Institute breast cancer study. The variable
grade, when classified into two groups shows clear effects that appear to be of
a concave nature. The reason is that the poorer prognoses will be among the
shorter survival times and, while these may include some misclassified patients,
the percentage of truly labeled poor grades will be higher earlier on than later. As
time unfolds, the initial classification has less relevance. The physical manifesta-
tion of this in the regression effect process is one where we see concave effects.
The area under the curves will generally provide a more powerful test than the
log-rank test.
11.6 Relative efficiency of competing tests
We now know that, under proportional hazards alternatives, the distance from
the origin test—equivalently the log-rank test—is not only unbiased and con-
sistent but is the uniformly most powerful unbiased test (UMPU). What hap-
pens though when these optimality conditions are not met, specifically when
the non-proportionality shown by the regression effect takes a particular form.
We might first consider different codings for Z(t) in the model: λ(t|Z(t)) =
λ0 (t) exp{β(t)ψ[Z(t)]}, and in view of Lemma 11.3, if ψ[Z(t)] = C0 β(t)Z(t)
for any strictly positive C0 , then this results in the most powerful test of
H0 : β(t) ≡ 0. We can make use of results from Lagakos et al. (1984) in order
to obtain a direct assessment of efficiency that translates as the number of extra
failures that we would need to observe in order to nullify the handicap resulting
from not using the most powerful local test. Proposition 11.9 provides the results
we need and its validity only requires the application of Lemma 11.3.
Exploiting the formal equivalence between time-dependent effects and time
dependency of the covariable, we can make use of a result of O’Quigley and
Prentice (1991) which extends a basic result of Lagakos et al. (1984). This
allows us to assess the impact of using the coding ψ[Z(t)] in the model in place
of the optimal coding ψ ∗ [Z(t)] = C0 β(t)Z(t). See also Schoenfeld (1981) and
Lagakos (1988).
Proposition 11.9. An expression for asymptotic relative efficiency is given

by:

[( V[ψ{Z(t)}, ψ ∗ {Z(t)}]λ0 (t)dt)]2
e2 (ψ, ψ ∗ ) = ,
V[ψ{Z(t)}, ψ{Z(t)}]λ0 (t)dt V[ψ ∗ {Z(t)}, ψ ∗ {Z(t)}]λ0 (t)dt
with V(a(t), b(t)) = E{Y (t)a(t)b(t)} − E{Y (t)a(t)}E{Y (t)b(t)}/E{Y (t)}.
The integrals are over the interval (0,1) and the expectations are taken with
respect to the distribution of Ft again on the interval (0,1). We can use this
expression to evaluate how much we lose by using the log-rank statistic when,
say, regression effects do not kick in for the first 25% of the study (on the φn (t)
timescale). It can also inform us on how well we do when using, say, the area
over the curve test as opposed to the optimal (unknown in practice) test. In the
absence of censoring this expression simplifies to that of the correlation between
the different processes. In some cases, an example being the log-rank process,
and the AUC process, the known properties of the limiting processes are readily
available and provide an accurate guide.
11.7. Supremum tests over cutpoints 335
11.7 Supremum tests over cutpoints
The above tests will be optimal, i.e., consistent and uniformly most powerful unbi-
ased tests, in given circumstances. Moving away from particular circumstances,
we will continue to obtain tests with near optimal properties. This requires us
to carefully consider the type of behavior that is likely to be exhibited. When
the alternative mirrors the form of β(t), we have an optimal test. Typically we
look for something broader since β(t) is not known. The three classes consid-
ered above—linear, concave, and convex—are precise enough to obtain powerful
tests, albeit falling short of the optimal test for specific situations. We might
then conclude that we have an array of tools from which we can find tests with
very good performance in a wide array of cases. This is true. Yet, there are still
other ways of tackling the problem.
For example, we may not wish to favor any class of alternatives; linear, con-
cave or convex. The alternative is then very broad. An appropriate class of tests
for this situation, a class that we would anticipate to have good power proper-
ties, might be based on cutpoints. The several examples of Chapter 10 suggest
that change-point models can closely approximate the regression effect process.
In most cases a single cutpoint will be good enough and provide a significantly
improved fit to that attained by a model with a constant regression effect. Extend-
ing to more than a single cutpoint is, in principle, quite straightforward, raising no
additional methodological considerations, although more cumbersome computa-
tionally. Under the null, the cutpoint would not be defined. Under the alternative,
we would have as many cutpoints as we choose to model and for s − 1 cutpoints
we would obtain s regression effect processes, each one, under the null and con-
ditional upon its starting point, being uncorrelated with the others. A single
cutpoint for example leads to two regression effect processes; one from zero up
to the cutpoint, the other from the cutpoint to the point 1 on the transformed
time scale (Fig. 11.11).
Tests within such a structure fall under the heading studied by Davies (1977,
1987) whereby the nuisance parameter (cutpoint) is defined only under the alter-
native and not under the null. In the case of a single cutpoint, we write:
Tn2 (β(t), θ) = θ−1 {Un∗ (β(t), θ)}2 + (1 − θ)−1 {Un+ (β(t), θ)}2 (11.9)
where Un+ (β(t), t) = Un∗ (β(t), 1) −Un∗ (β(t), θ) . The ingredients we need are pro-
vided by the variance-covariance matrix of the unsquared quantities. Writing
these as:
⎛ √ ⎞
Un∗ (β(t), θ)√
/ θ
⎜ Un+ (β(t), 1) / 1 − θ ⎟ I A(θ)
Var ⎜ √
⎝ DU ∗ (β(t), θ) / θ ⎠
⎟=
AT (θ) B(θ)
+
n √
DUn (β(t), 1) / 1 − θ
where DU (θ) = ∂U/∂θ, we have most of what we need. Finally, suppose that
η(θ) are zero-mean normal random variables with variances given by λ1 and λ2
where λi , i = 1, 2, are the eigenvalues of B(θ) − AT (θ)A(θ). Then:
1√
1
P{sup Tn2 (β(t), θ) > u ; 0 < θ < 1 } ≤ P(χ22 > u) + ue−u/2 E||η(θ)||dθ
π 0
Clearly, the calculations are not that straightforward although, with modern soft-
ware such as R, they are readily programmed. In addition, Davies (1987) makes a
number of suggestions for simpler approximations. Some study of these sugges-
tions, in the context of survival data with changepoints, has been carried out by
O’Quigley and Pessione (1991), O’Quigley (1994), and O’Quigley and Natarajan
(2004) and indicate that the accuracy of the approximations is such that they
can be reliably used.
A full study of tests based on taking the supremum over an interval of cut-
points has yet to be carried out. In terms of power and test optimality any
advantages will be necessarily small since we already know how to obtain near
optimal tests. These optimal, and near optimal, tests apply in a range of situa-
tions and include cutpoints. So, despite the extra work, there may not be much
to be gained. One useful advantage is that any such supremum test immediately
furnishes both interval and point estimates of the changepoint itself. This can
be valuable in moving forward from the rejection of a null hypothesis to the
construction of a more plausible hypothesis.
Figure 11.11: Two possible cutpoints for a single cutpoint model. Note the impact
of small changes in cutpoint on the slopes of the regression effect process.
11.8. Some simulated comparisons 337
11.8 Some simulated comparisons
A large simulation study was carried out by Chauvel (2014), some of which was
reproduced in Chauvel and O’Quigley (2014) and Chauvel and O’Quigley (2017).
The behavior of the tests described above for finite sample sizes was compared
with a number of other established tests. The advantages are readily highlighted.
Framework for the simulations

The times of death T are simulated according to an exponential distribution with
parameter 1. The censoring times C, independent of T , also follow an exponential
distribution, whose fixed parameter takes a range of values, chosen to calibrate
the rate of censoring, as shown in Table 11.1. The rate of censoring in the sample
is thus 30%, 50% or 70%. Data is simulated with one covariate (p = 1). This
covariate Z follows either a Bernoulli distribution with parameter 1/2, a normal
distribution
with
√ mean√1/2 and variance 1/4, a continuous uniform distribution
on 0.5 × 1 − 3, 1 + 3 , or an exponential distribution with parameter 1/2.
These parameters have been chosen in order to make the first two moments
identical. The sample size is either 60, 100 or 200. The tests have been cal-
ibrated so that α = 0.05. Several regression coefficients were considered. The
β(t) = 0 for all t ∈ [0, 1] case helps us check the level of the tests. The β(t) = 0.5
(or β(t) = 0.8) for all t ∈ [0, 1] cases are the two proportional hazards situations
we choose to look at. Other cases we consider involve non-proportional hazards,
with either a continuous decreasing effect (β(t) = 1.5(1 − t) or β(t) = 2(1 − t)2 ,
for t ∈ [0, 1]) or hazards that cross over: (β(t) = −1.5 1t≤0.5 + 1.5 1t≥0.5 for
t ∈ [0, 1], or with swapped signs). We also look at three change-point models
(β(t) = 1t≤0.3 , β(t) = 1t≤0.5 and β(t) = 1t≤0.7 , t ∈ [0, 1]). For each regression
coefficient and each rate of censoring considered, 3000 samples are generated as
described above and used to evaluate the empirical level and power of each test.
λ 0.5 1 2
P (C ≤ T ) 0.3 0.5 0.7
Table 11.1: Link between the rate of censoring P (C ≤ T ) and parameter λ of

the censoring distribution E(λ).
Remark. The theorems of Section 9 were proven for uniformly bounded

covariates. These theorems form the basis of the tests presented in this chap-
ter. However, here we run simulations with normal and exponential covariates
since these cases are often found in applications, even though they are indeed
not uniformly bounded. More examples can be found in Chauvel (2014).
Mnγ (0, 1)
β(t) Log-rank Jn (0, 1) γ = 0.5 γ = 1 γ = 1.5 γ=2
0 4.9 5.2 5.0 5.1 4.8 4.6
0.5 50.9 40.9 42.7 46.3 49.5 50.3
0.8 88.1 76.8 79.2 83.5 86.6 87.5
1t≤0.3 21.8 38.5 35.4 32.0 25.3 21.2
1t≤0.5 54.1 72.6 70.1 66.7 58.3 52.8
1t≤0.7 83.4 89.0 88.3 85.5 82.3 80.9
−1.5 1t≤0.5
16.8 85.3 82.7 83.4 84.4 80.0
+1.5 1t≥0.5
1.5 1t≤0.5
15.5 85.1 82.3 83.2 84.4 79.5
−1.5 1t≥0.5
1.5 (1 − t) 84.7 91.6 91.0 88.8 85.2 83.0
2 (1 − t)2 73.8 90.6 89.2 86.0 78.7 72.8
Table 11.2: Empirical level of significance and power of each test (in %) based
on 3000 simulated datasets for each β(t). Each dataset has 100 elements (50
subjects per group). The rate of censoring is fixed at 30%.
Setting the calibration parameter γ

Simulations were first run to calibrate γ for the restricted adaptive test with
statistic Mnγ (0, 1). Note that the advantage of this test is that it loses very little
power compared with the log-rank test under the proportional hazards hypothesis,
while gaining power under non-proportional hazards-type alternative hypotheses.
The sample size was set to n = 100, with the covariate Z simulated following
a Bernoulli with parameter 1/2 and a 30% rate of censoring. For each regres-
sion coefficient β(t), the log-rank test, area under the curve test with statistic
Jn (0, 1), and restricted adaptive test Mnγ (0, 1) were run, with γ = 0.5, 1, 1.5,
and 2. The results are given in Table 11.2. When γ = 2, the power of the test
with statistic Mnγ (0, 1) is very close to that of the log-rank test, except when
the risks intersect, i.e., when the regression coefficient changes sign over time.
In such situations, the test with statistic Mn2 (0, 1) is much more powerful than
the log-rank. In other situations, the test is weak in detecting the presence of
an effect that changes with time, inheriting the behavior of the log-rank. When
γ = 0.5, the power of the test with statistic Mnγ (0, 1) is now instead very close
to that of the area under the curve test. Hence, we see a large loss in power
(around 10%) with respect to the log-rank test under proportional hazards-type
alternative hypotheses. Lastly, the γ = 1 and γ = 1.5 cases are between these two
extremes. The power of the test with statistic Mn1 (0, 1) is closer to that of the
area under the curve test, while the behavior under the alternative hypothesis of
the test with statistic Mn1.5 (0, 1) is closer to that of the log-rank. Unlike the pre-
11.8. Some simulated comparisons 339
vious cases, these tests have acceptable power under all alternative hypotheses
considered. On this basis we find γ = 1.5 provides a good all-round compromise
and tends to keep to a minimum any loss in power with respect to the log-rank
test under proportional hazards-type alternative hypotheses.
Performance of the tests

Chauvel and O’Quigley (2014) examined the performance at t = 1 of the dis-
tance from the origin test, area under the regression effect process test, and
restricted adaptive tests, with respective statistics Un∗ (0, 1), Jn (0, 1), Mn (0, 1),
and Mn1.5 (0, 1). We also included the log-rank test (Mantel, 1966) and two other
adaptive tests which were briefly described in Section 11.3 and are summarized
below. More results are given in Chauvel (2014). The first test is that of Lee
(1996), whose statistic is the maximum of four weighted log-rank statistics with
weights from the family Gρ,γ given in Fleming and Harrington (1991). These
weights are useful to detect respectively early effects, effects in the middle of
a study, delayed effects between the survival curves of the two groups, and a
proportional hazards situation. The procedure is similar to the test proposed by
Self (1991). The class of tests in Fleming and Harrington (1991) is known to
have good power for detecting time-dependent effects.
The other test statistic, proposed by Yang and Prentice (2010), is the maxi-
mum between two weighted log-rank statistics, with weights based on the infer-
ence of the quantities defined by the model (6.26) of Yang and Prentice (2005).
The authors argue that, asymptotically, this statistic would be equivalent to the
log-rank when hazards are proportional.
Consider the one-covariate case with a Bernoulli distribution with parameter
1/2. Results are presented in Table 11.3 for the n = 100 case, Table 11.4 for
n = 60, and Table 11.5 for n = 200. The conclusion is the same for each sample
size considered.
With the exception of the tests of Lee and Yang and Prentice, test levels are
controlled well at 5%. Inflation of the size of the Yang and Prentice test confirms
their own simulation results for comparable sample sizes and rates of censoring.
As expected (see Proposition 11.2), the power of the distance test is very close
to that of the log-rank’s under proportional and non-proportional hazards-type
alternative hypotheses.
When the hazards are proportional, the area under the curve test is less
powerful than the log-rank, while the power of the restricted adaptive test Mn
lies between the two, with an average loss compared to the log-rank of only 3%.
The powers of the restricted adaptive test Mnγ with γ = 1.5 and the Lee test
are comparable to those of the log-rank. The most powerful test would seem
to be the Yang and Prentice one. This result is surprising since the log-rank
test is the most powerful in this situation (Peto and Peto, 1972). This apparent
contradiction can be explained by poor control of the type I error of the test.
The first non-proportional hazards case involves a piecewise-constant effect,

corresponding to a change-point model. We vary the times at which the changes
occur. The area under the curve test is more powerful than the log-rank and
distance tests. The power of the restricted adaptive test Mn is closest to that of
the area under the curve test. As for the power of the test with statistic Mn1.5 , it
is low and close to the log-rank’s. The tests described in this chapter have better
power than the Lee test. The Yang and Prentice test appears to be a powerful
test but the drawbacks of this test, one being inadequate control over Type 1
error, rule out its use in practice.
In cases where the hazards intersect, the powers of the log-rank, distance,
and Lee tests are low. The Yang and Prentice test undergoes a severe reduction
in power, up to 30%, between an effect having a negative and then a positive
value and a positive followed by a negative one. As for the restricted adaptive
test and the area under the curve test, they have reasonable power in situations
where hazards intersect. In the last set of cases studied, the effect β(t) decreases
continuously with time. All of the test’s powers were similar, with the area under
the curve test being the most powerful.
To summarize the simulations, in general situations in which no hypothesis
is made about the type of effect, use of the restricted adaptive test with statistic
Mn is recommended. The simulation results show good power under alternative
hypotheses of both the proportional and non-proportional hazard types. The Yang
and Prentice test should be used with caution because of its lack of control of
type I errors and dependency on the chosen coding. If the goal is to lose as little
power as possible compared to the log-rank test, we recommend using one of the
tests from the restricted adaptive class with statistics Mnγ . However, these tests
are less powerful than the Mn one for non-proportional hazards. The Lee and
the Yang and Prentice tests were developed to compare survival in two groups
of patients. Thus, they can only be applied when there are binary covariates. We
also studied the behavior of the tests when there are continuous covariates. The
parameters of the distributions were chosen so that the first two moments were
the same. The sample size was fixed at n = 100 and the rate of censoring at
30%. Some of the findings are reproduced in Table 11.6. For a given test and
regression coefficients, the respective powers of each test are similar for the three
covariate distributions.
11.9 Illustrations
We consider several publicly available data sets. Also included are data collected
by the Curie Institute for studying breast cancer and a number of data sets from
the University of Leeds, U.K. focused on the impact of markers on survival in
cancer. Note that the analysis of a multivariate problem often comes down to
considering several univariate regression effect processes. This is because the
relevant questions typically concern the impact of some variable, often treat-
11.9. Illustrations 341
ment, having taken into account several potentially confounding variables. Once
these confounding variables have been included in the model then the process
of interest to us is a univariate one where the expectation and variance have
been suitably adjusted via model estimation. This is also the case when using
stratification or, indeed, when considering the prognostic index which would be
a linear combination of several variables.
Breast cancer study at the Institut Curie, Paris

The dataset is made up of 1504 patients with breast cancer. The data is complex
with several subgroups; for our illustration, we consider a specific subgroup of 332
patients. Among the prognostic factors, we focus our attention on the influence
of tumor size in millimeters (mm) on patient survival. We separate the patients
into two groups depending on whether their tumor size is above or below 60 mm.
Patients were monitored for 12 years. The rate of censoring in the dataset was
22%. Kaplan-Meier estimators of survival are shown in Figure 11.12(a). During
the first 60 months, the estimated survival rates are quite different, correspond-
ing to a strong effect between the patient groups. After 60 months, the curves
get closer together, reflecting a decrease in effect. We see that 80% of patients
whose tumor size is less than 60 mm survive at least 75 months. This drops to
28 months for patients with tumor size above 60 mm.
The corresponding standardized score process is shown in Figure 11.12(b). As
this process is not centered around 0, the presence of drift and thus an effect is
clear. As the drift is not linear, the effect decreases over time. This decrease has an
impact on the log-rank (p = 0.06) and distance from the origin (p = 0.08) tests,
which are not significant. Nevertheless, the area under the curve and restricted
adaptive tests are highly significant, with respective p-values of 0.001 and 0.002.
The conclusion from these two tests is consistent with our analysis of the Kaplan-
Meier estimators.
A large randomized clinical trial

In a recent clinical trial, 1636 patients were randomized to 3 treatment groups
and followed for 65 months. The goal of the trial was to compare the influence
of different treatments on patient survival for individuals with the same illness.
The Kaplan-Meier estimators of the survival curves are shown in Figure 11.13(a).
Patient survival in randomized groups 2 and 3 appears similar, and different to
that of group 1. This is immediately interesting since groups 2 and 3 correspond
to slightly different administrations of the treatment. Independently then, given
the observed survival of group 1, each group appears to differ from this group in
the same way. The survival experiences of the two treated groups all but coin-
cide. Using the treatment covariate, two binary covariates were constructed with
“treatment group 1” as a reference. The corresponding multivariate standardized
score process Un∗ (0, ·) is shown as a function of time in Figure 11.13(b).
Despite the noise, the drifts of the processes are clearly visible, corresponding
to the alternative hypothesis of the presence of an effect—which appears to be
nonlinear. Its value decreases with time in such a way that the multivariate log-
rank test is not significant, with a p-value of 0.10. The multivariate distance
test has a p-value of 0.09. In contrast, the multivariate area under the curve
and multivariate restricted adaptive tests are highly significant, with respective
p-values of 0.005 and 0.008. These tests are less affected by the decrease in
effect, and for this reason they detect it. They thus correspond more closely to
our intuition when looking at the Kaplan-Meier curves which do seem to suggest
the presence of true differential effects.
11.10 Some further thoughts
In order to compare the survival of several groups, the log-rank test is commonly
used. This test is the most powerful for detecting a fixed effect over time. On
the other hand, several authors have shown its weakness in the presence of
an effect which varies. The purpose of Chauvel and O’Quigley (2014) was to
develop a test with good power under both types of alternative hypothesis. The
asymptotic properties of the regression effect process studied in Chapter 9 lead to
the construction of several new tests. The first of these is the distance from the
origin test, which is asymptotically equivalent to the log-rank. There are many
other possibilities including the area under the curve test. This test, slightly less
powerful than the log-rank test in the presence of a constant effect, allows for
significant power gains when there is an effect that changes with time.
Figure 11.12: Kaplan-Meier estimators and standardized score process for the
Curie Institute data.
11.10. Some further thoughts 343
The class of adaptive tests considered will assume either the value of the
distance traveled or the value of the area under the curve. The parameter allows
us to calibrate the proximity of the tests to the log-rank. However, fixing this
parameter may be tricky, so if there is no compelling reason to privilege the
log-rank test, Chauvel and O’Quigley (2014) proposed the use of an unrestricted
adaptive test. This corresponds to a statistic that is the maximum absolute value
of the distance from the origin and the area under the curve.
Simulations show that these adaptive tests perform well compared to other
tests in the literature. In addition to the tests’ control of the level and their good
power properties, they also have the advantage of being easily interpretable and
simple to plot. In applications, at the same time as we look at the p-values,
parameter estimates, and confidence intervals, it would certainly be helpful to
consider plots of the processes in order to get a visual impression of how close or
far we appear to be from the conditions under which we know we would achieve
something close to an optimal test. This tells us something not only about the
potential lack of plausibility of the null hypothesis but also something about the
nature of a more plausible alternative hypothesis.
By calculating the standardized score process Un∗ at β0 , the results can imme-
diately be extended to the null hypothesis H0 : β(t) = β0 , ∀t, against the alter-
native H1 : ∃ t0 , β(t0 ) = β0 . If β0 is non-zero, we are no longer testing for the
presence of a difference in survival between groups, but rather whether the relative
risk between groups is equal to β0 or not. Although we have mostly considered
testing a null against an alternative of a constant value for the regression parame-
ter, it is however possible to consider testing the null hypothesis H0 : β(t) = β0 (t)
against H1 : β(t) = β0 (t), where β0 (t) changes with time.
Figure 11.13: Kaplan-Meier estimators and standardized score processes for clin-
ical trial data with 3 treatment groups.
Other combinations of the distance from the origin and area under the curve
tests could also be studied, rather than the linear combination looked at here:
θUn∗ (β0 , t) + (1 − θ)Jn (β0 , t), 0 ≤ θ ≤ 1.
However, depending on whether the hazards are proportional or not, simulations

tend to show that the greatest power is reached for θ = 1 or θ = 0, respectively.
We then find ourselves back with the restricted adaptive test. Aside from this,
it would also be interesting to study other tests based on the properties of dif-
ferent functions of Brownian motion, which will approximate the limit process of
Un∗ (β0 , ·) under varying assumptions.
We have applied the tests described here with t = 1. It would be possible to
apply them at any time τ if we knew, for example, that the effect became zero
after some time point, τ. This would also allow for increases in the power of
the tests. It may also be of interest to study tests whose statistic is the supre-
mum over time of the distance or area under the curve tests. In the multivariate
case, the tests we have introduced allow us to simultaneously test whether all
parameters are equal to zero. An extension to this could be to find a way to use
the process to test whether specific parameters only were equal to zero. Sup-
pose we had two variables Z1 and Z2 in a non-proportional hazards model with
respective parameters β1 (t) and β2 (t). To test H0 : β1 = 0 versus H1 : β1 = 0,
one idea would be to stratify with respect to the covariate Z2 by estimating
β2 (t) and examining
the process whose construction
is based on the probabilities
πi B̂(t), t , i = 1, . . . , n , with B̂(t) = 0, β̂2 (t) .
1. Suppose we have a study in which, before some time point τ , treatment A

has an advantage over treatment B but, beyond that point, this advantage
becomes a disadvantage. Without any formal analysis how might you obtain
a quick rough-and-ready estimate of τ.
2. For the previous question consider a null hypothesis that deems effects after
τ to be equal to those before τ and both equal to zero. How might you
make use of the known properties of reflected Brownian motion to construct
a consistent test of this hypothesis?
3. Continuing the previous question, how might we make use of Kolmogorov’s
theorem and the Brownian bridge to construct an alternative test to one
based on reflected Brownian motion.
4. Discuss the relative advantages and drawbacks of the two tests described in
the previous two questions. Are you able to make any clear recommendations
for given situations?
5. On the basis of different data sets, calculate the distance from the origin
test, the log-rank test, and the arcsine test. On the basis of these results
what does your intuition tell you?
6. Under proportional hazards alternatives to the null hypothesis of no differ-

ence between two groups, show that the log-rank test, the distance from
the origin test, the area under the curve test are all unbiased and consistent
but that Kolmogorov’s test will not generally be unbiased nor consistent.
7. For some data, construct 100(1−α)% confidence intervals for β in a propor-

tional model based on the distance from the origin test and the area under
the curve test. For a non-proportional hazards model with coefficient β(t),
how might you best go about constructing test based confidence intervals
for β(t).
8. One of the conditions for consistency relates to the rate of censoring; specif-
ically we need to have kn /n → C as n → ∞ where 0 < C < 1. Suppose,
for some true situation, that kn /n → 0 as n → ∞. Indicate why all of the
tests, including the log-rank test, would be inconsistent in such a situation.
9. Under proportional hazards, with coefficient β0 , show that as |β0 | moves

away from zero, the area under the curve test tends to a test that is fully
efficient.
10. For the situation of either proportional hazards or concave non-proportional

hazards, we can calculate two areas: the area under the process and the
area of the triangle defined by the arrival point of the process. The ratio
of these two areas provides a goodness-of-fit statistic. Under proportional
hazards, these two areas will be very close in size. As we move away from
proportional hazards the difference between them will grow. Construct a
formal test based on these two areas. Use the relative difference in these
two areas to construct an index of fit. Discuss the advantages and drawbacks
of such an index in practice.
11. Repeat the previous question, changing the word concave to convex. What
would the impact of this be? Can you generalize the arguments to embrace
the more general setting in which one of three situations may prevail:
proportional hazards, concave non-proportional hazards, and convex non-
proportional hazards.
12. For a clinical trial the presence of a percentage of non-responders suggests

a particular choice of test in order to maximize power. Suppose we have a
different situation that involves heterogeneity; that is one where the effect
of treatment differs from patient to patient. This is the case described by
frailty models. Which test would be the most suitable for this situation.
13. Consider a two-group clinical trial in which, beyond the median of the com-
bined groups there is no further effect. Use the expression for efficiency to
calculate how many more patients would be needed to maintain the power
of the most powerful test when the design is based on the log-rank test.
Hint—use the convergence in distribution result between the log-rank and
the distance traveled tests.
14. Guidance on the choice of tests has so far made little reference to censor-
ing beyond the fact that it is assumed either independent or conditionally
independent of the failure mechanism. The impact of a conditionally inde-
pendent censoring mechanism is potentially greater on the distance from
origin test than it is on the log-rank test. Explain why this would be so.
Propositions 11.1 and 11.5 First, recall the result of Theorem 9.2:
L
Un∗ (β0 , t) − kn An (t) −→ C1 (β, β0 )W(t), 0 ≤ t ≤ 1,
n→∞
where t
P
An (t) −→ C2 {β(s) − β0 } ds, 0 ≤ t ≤ 1.
n→∞ 0
Let t ∈ [0, 1]. The functions ∗
√ Un(β 0 , ·) and An are almost surely bounded so that
t
we can write Jn (β0 , t) − kn 0 An (s)ds as:
t t t
L
Un∗ (β0 , s)ds − kn An (s)ds −→ C1 (β, β0 ) W(s)ds.
0 0 n→∞ 0
In consequence, we have convergence in probability:

t
1 P
√ Un∗ (β0 , t) −→ C2 {β(s) − β0 } ds,
kn n→∞ 0
and t s
1 P
√ Jn (β0 , t) −→ C2 (β(u) − β0 )duds.
kn n→∞ 0 0
√ s
Recall that limn→∞ kn = ∞, and for any s ∈ [0, 1], 0 {β(u) − β0 }du = 0,
which together imply:

1 ∗ α/2 −3 1/2 α/2
lim P √ |Un (β0 , t) | ≥ z = 1, lim P 3t |Jn (β0 , t) | ≥ z = 1.
n→∞ t n→∞
cens. Log- ∗ 1.5 Yang

β(t) Un (0, 1) Jn (0, 1) Mn (0, 1) Mn (0, 1) Lee
(%) rank Prentice
30 4.8 4.5 4.9 5.0 4.4 6.1 6.4
0 50 5.2 5.0 5.4 5.0 5.0 6.3 6.8
70 5.1 4.7 4.8 4.8 4.4 5.8 6.2
30 50.9 49.9 39.6 47.3 48.8 50.4 54.2
0.5 50 51.0 50.2 39.0 47.4 48.9 51.0 53.8
70 51.7 50.8 40.9 48.7 49.7 50.8 55.5
30 87.1 86.6 74.6 84.1 85.3 86.3 89.0
0.8 50 87.2 86.7 75.8 85.3 85.7 86.8 88.4
70 88.0 87.1 76.8 85.8 85.9 87.3 89.0
30 23.7 22.2 40.1 34.6 26.2 19.8 40.4
1t≤0.3 50 18.4 17.6 30.8 26.9 21.0 16.1 34.4
70 13.5 12.7 21.7 18.5 14.4 11.1 26.6
30 55.5 52.4 73.2 69.4 59.0 49.5 69.1
1t≤0.5 50 44.3 41.9 61.0 55.9 46.7 37.3 57.7
70 32.8 30.3 44.4 40.9 32.6 27.2 43.6
30 83.1 80.6 88.2 86.9 82.1 79.0 87.3
1t≤0.7 50 72.4 69.4 77.3 76.4 71.1 66.6 78.1
70 56.0 51.9 60.2 59.0 53.3 49.2 61.2
30 17.1 15.1 85.7 81.3 84.8 27.5 61.5
−1.51t≤0.5
50 13.5 11.7 72.7 67.7 69.6 39.5 42.3
+1.51t≥0.5
70 12.0 11.4 56.0 49.5 49.0 37.0 25.8
30 15.9 13.8 85.3 81.4 84.3 26.0 87.8
1.51t≤0.5
50 14.0 12.4 74.0 68.0 71.0 40.6 77.1
−1.51t≥0.5
70 11.0 10.2 55.2 49.1 48.7 36.7 61.6
30 85.0 82.9 90.7 90.0 84.7 81.3 90.1
1.5 (1 − t) 50 74.5 71.0 81.5 79.9 73.2 68.6 81.2
70 58.0 54.9 64.7 62.8 56.3 52.0 65.8
30 73.5 70.8 89.7 87.0 78.2 67.8 88.1
2 (1 − t)2 50 60.7 57.4 78.7 74.7 64.6 54.0 77.9
70 46.7 43.0 62.1 58.1 47.2 39.4 62.8
Table 11.3: Empirical level of significance and power of each test (in %) based on
3000 simulated datasets for each β(t) and each rate of censoring. Each dataset
has 100 elements (50 subjects per group). The tests have nominal levels of 5%.

β(t) Un (0, 1) Jn (0, 1) Mn (0, 1) Mn (0, 1) Lee
(%) rank Prentice
30 4.9 4.7 4.7 4.3 4.7 5.7 6.6
0 50 5.4 5.5 4.9 5.1 5.3 6.5 7.1
70 4.9 4.7 4.3 4.4 4.7 5 4.9
30 35.6 34.6 26.9 32.5 33.5 36.4 38.1
0.5 50 27.6 26.7 19.5 25.4 26.0 27.7 29.7
70 19.9 19.9 14.8 18.6 19.3 21.1 21.1
30 68.2 67.6 53.7 64.8 66.0 68.9 70.4
0.8 50 53.7 52.8 40.8 50.6 51.3 54.6 56.2
70 38.9 38.2 28.7 36.6 37.1 40.0 39.4
30 16.5 15.2 25.4 22.4 17.6 14.9 29.3
1t≤0.3 50 13.5 12.7 19.9 17.6 13.7 12.5 23.6
70 10.5 10.2 15.0 13.2 10.5 9.2 16.6
30 36.8 34.1 50.5 46.3 38.2 33.0 49.2
1t≤0.5 50 29.3 26.7 40.2 36.6 28.9 25.0 40.1
70 23.1 21.0 28.7 26.5 21.6 18.8 29.0
30 61.9 58.3 66.4 64.7 59.1 57.0 67.4
1t≤0.7 50 51.6 47.9 53.8 54.4 48.8 47.2 57.5
70 38.4 34.9 37.9 38.4 34.8 33.4 40.2
30 11.9 10.2 65.8 59.4 61.0 15.6 42.7
−1.51t≤0.5
50 10.7 9.3 51.8 45.8 43.3 22.4 29.6
+1.51t≥0.5
70 9.4 8.4 36.6 30.8 25.5 20.2 18.9
30 13 11.2 66.6 59.9 61.7 16.8 67.6
1.51t≤0.5
50 11.2 9.7 51.3 44.8 43.4 22.6 51.3
−1.51t≥0.5
70 8.7 8.1 36.7 31.2 26.0 19.1 38.6
30 65.8 62.4 71.6 70.7 63.8 61.4 73.8
1.5 (1 − t) 50 53.1 48.6 57.7 57.1 50.3 46.9 60.9
70 40.8 37.6 42.4 42.1 37.8 36.3 44.7
30 53.3 49.4 69.8 65.6 55.5 48.4 71.0
2 (1 − t)2 50 44.9 41.3 57.4 53.4 44.2 39.3 58.3
70 29.7 27.6 37.5 34.7 29.2 25.2 37.5

β(t) Un (0, 1) Jn (0, 1) Mn (0, 1) Mn (0, 1) Lee
(%) rank Prentice
30 5.5 5.4 4.6 5.2 5.1 6.7 6.3
0 50 5.2 5.3 4.9 5 5.2 6.2 6.6
70 5.2 5.2 5.5 5.3 5.1 5.8 6.4
30 79.8 79.3 68.4 76.7 78.8 78.4 81.7
0.5 50 67 66.3 54.8 64.2 65.4 64.2 69
70 50.3 48.9 41.1 47.9 48.2 47.2 52.2
30 99.2 99.1 97.1 98.9 98.8 99.1 99.3
0.8 50 96.9 96.5 91.7 96.3 95.7 95.9 97.1
70 86.7 86.3 76.8 84.8 85 84.6 87.8
30 40.3 38.9 69.6 63.2 52.4 33.7 62.7
1t≤0.3 50 29.9 28.8 55.6 49 38.4 24.3 49.5
70 22.3 21.5 39.8 35.2 26.4 17.6 39.9
30 82.9 81.6 96.4 95 89.5 78 93.3
1t≤0.5 50 69.8 68.2 88.7 85.4 76.2 63.4 83.6
70 53.7 51 72.2 67.9 57.4 45.1 67.5
30 98.2 97.9 99.3 99.1 98.3 97.2 99
1t≤0.7 50 94.2 93 96.7 96.1 94 91.3 96.2
70 83.8 80.9 88.1 87.3 82.7 77.8 87.8
30 23.9 19.9 98.9 98.4 98.8 56.4 94
−1.51t≤0.5
50 20.7 17.2 95.3 93.5 95.1 73.3 70
+1.51t≥0.5
70 15.6 13 84.2 80.4 83.1 63.3 41.8
30 23.7 19.7 98.7 98.3 98.7 53.9 98.8
1.51t≤0.5
50 20.6 17 94.8 92.9 94.3 74 96.9
−1.51t≥0.5
70 15.7 13.7 83.1 79.3 81.6 61 88.5
30 98.7 98.4 99.8 99.7 98.8 98.3 99.6
1.5 (1 − t) 50 95.5 94.2 98.4 97.9 95.8 93 97.8
70 86.5 84.3 92.1 91.1 86.6 81.7 91.4
30 95.5 94.8 99.7 99.5 98 93.3 99.5
2 (1 − t)2 50 89 87.1 98.1 97.3 93.7 84.1 97.3
70 73.1 70.1 89.7 86.9 78.3 65.7 87.8
β(t) Covariate Un∗ (0, 1) Jn (0, 1) Mn (0, 1) Mn1.5 (0, 1)

Normal 4.9 4.5 4.9 4.9
0 Uniform 5.4 4.8 5.2 5.5
Exponential 5.0 5.0 5.0 5.3
Normal 46.9 40.6 45.9 46.4
0.5 Uniform 48.7 41.0 46.5 48.2
Exponential 48.3 38.5 45.4 47.9
Normal 84.6 76.4 83.5 84.1
0.8 Uniform 84.5 76.3 82.1 84.0
Exponential 84.7 76.1 83.0 83.9
Normal 22.4 41.7 37.2 28.2
1t≤0.3 Uniform 21.2 41.1 35.7 27.4
Exponential 21.6 42.1 36.1 27.7
Normal 49.1 71.2 66.7 56.4
1t≤0.5 Uniform 50.6 71.5 66.7 56.9
Exponential 49.0 71.5 67.5 56.9
Normal 77.0 87.3 84.9 79.4
1t≤0.7 Uniform 75.9 85.6 83.6 77.5
Exponential 75.7 85.9 83.7 78.1
Normal 9.8 71.9 66.2 70.1
−1.51t≤0.5
Uniform 9.1 73.3 67.9 71.4
+1.51t≥0.5
Exponential 9.6 73.1 67.9 71.2
Normal 9.4 74.8 69.6 73.0
1.51t≤0.5
Uniform 9.5 73.8 68.2 71.9
−1.51t≥0.5
Exponential 9.6 72.5 67.6 70.5
Normal 81.5 91.4 89.4 84.0
1.5 (1 − t) Uniform 79.9 91.3 89.9 83.2
Exponential 79.7 90.5 88.7 83.
Normal 70.9 90.5 87.9 79.7
2 (1 − t)2 Uniform 68.7 89.6 86.4 78.6
Exponential 70.2 90.2 87.1 79.1
Table 11.6: Empirical levels of significance and test powers (in %) based on 3000
datasets for each β(t) and covariate distribution. The rate of censoring is set at
30%. Each dataset is of size 100.
Appendix A
Probability
A.1 Essential tools for survival problems
We recall some of the fundamental tools used to establish the inferential basis
for our models. The main ideas of stochastic processes, in particular Brownian
motion and functions of Brownian motion, are explained in terms that are not
overly technical. The background to this, i.e., distribution theory and large sam-
ple results, is recalled. Rank invariance is an important concept, i.e., the ability
to transform some variable, usually time, via monotonic increasing transforma-
tions without having an impact on inference. These ideas hinge on the theory
of order statistics and the basic notions of this theory are recalled. An outline
of the theory of counting processes and martingales is presented, once again
without leaning too heavily upon technical measure-theoretic constructions. The
important concepts of explained variation and explained randomness are outlined
in elementary terms, i.e., only with reference to random variables, and at least
initially, making no explicit appeal to any particular model. This is important
since the concepts are hardly any less fundamental than a concept such as vari-
ance itself. They ought therefore stand alone, and not require derivation as a
particular feature of some model. In practice, of course, we may need to estimate
conditional distributions and making an appeal to a model at this point is quite
natural.
A.2 Integration and measure
The reader is assumed to have some elementary knowledge of set theory and
calculus. We do not recall here any of the basic notions concerning limits, con-
tinuity, differentiability, convergence of infinite series, Taylor series, and so on
and the rusty reader may want to refer to any of the many standard calculus
© Springer Nature Switzerland AG 2021 351

https://doi.org/10.1007/978-3-030-33439-0
352 Appendix A. Probability
texts when necessary. One central result which is frequently called upon is the
mean value theorem. This can be deduced as an immediate consequence of the
following result known as Rolle’s theorem:
Theorem A.1. If f (x) is continuously differentiable at all interior points of the
interval [a, b] and f (a) = f (b), then there exists a real number ξ ∈ (a, b) such
that f (ξ) = 0.
A simple sketch would back up our intuition that the theorem would be
correct. Simple though the result appears to be, it has many powerful implications
including:
Theorem A.2. If f (x) is continuously differentiable on the interval [a, b], then
there exists a real number ξ ∈ (a, b) such that
f (b) = f (a) + (b − a)f (ξ).
When f (x) is monotone then ξ is unique. This elementary theorem can form
the basis for approximation theory and series expansions such as the Edgeworth
and Cornish-Fisher (see Section A.10). For example, a further immediate corollary
to the above theorem obtains by expanding in turn f (ξ) about f (a) whereby:
Corollary A.1. If f (x) is at least twice differentiable on the interval [a, b] then
there exists a real number ξ ∈ (a, b) such that
(b − a)2
f (b) = f (a) + (b − a)f (a) + f (ξ).
2
The ξ of the theorems and corollary would not typically be the same and we
can clearly continue the process, resulting in an expansion of m + 1 terms, the
last term being the m th derivative of f (x), evaluated at some point ξ ∈ (a, b) and
multiplied by (b − a)m /m!. An understanding of Riemann integrals as limits of
sums, definite and indefinite integrals, is mostly all that is required to follow the
text. It is enough to know that we can often interchange the limiting processes
of integration and differentiation. The precise conditions for this to be valid are
not emphasized. Indeed, we almost entirely avoid the tools of real analysis. The
Lebesgue theory of measure and integration is on occasion referred to, but a lack
of knowledge of this will not hinder the reader. Likewise we will not dig deeply
into the measure-theoretic aspects of the Riemann-Stieltjes integral apart from
the following extremely useful construction:
Definition A.1. The Riemann integral of the function f (x) with respect to x, on
the interval [a, b], is the limit of a sum Δi f (xi−1 ), where Δi = xi − xi−1 > 0,
for an increasing partition of [a, b] in which max Δi goes to zero.
b
The limit is written a f (x)dx and can be seen to be the area under the
curve f (x) between a and b. If b = ∞ then we understand the integral to exist
if the limit exists for any b > 0, the result itself converging to a limit as b → ∞.
A.2. Integration and measure 353
Similarly for a = −∞. Now, instead of only considering small increments in x,

i.e., integrating with respect to x, we can make use of a more general definition.
We have:
Definition A.2. The Riemann-Stieltjes

integral of the function f (x) with respect
to g(x) is the limit of a sum {g(xi ) − g(xi−1 )}f (xi−1 ), for an increasing
partition of [a, b] in which, once again, max Δi goes to zero.
b
The limit is written a f (x)dg(x) and, in the special case where g(x) = x,
reduces to the usual Riemann integral. For functions, necessarily continuous,
whereby
x g(x) is an antiderivative of, say, h(x) and can be written g(x) =
h(u)du then the Stieltjes integral coincides with the Riemann integral
−∞
f (x)h(x)dx. On the other hand whenever g(x) is a step function with a finite
or a countable number of discontinuities then f (x)dg(x) reduces to a sum, the
only contributions arising at the discontinuities themselves. This is of great impor-
tance in statistical applications where step functions naturally arise as estimators
of key functions. A clear example of a step function of central importance is the
empirical distribution function, Fn (x) (thisis discussed in detail in Appendix D).
We can then write the sample mean x̄ = udFn (u) and the population mean
μ = udF (u), highlighting an important concept, that fluctuations in the sample
mean can be considered a consequence of fluctuations in Fn (x) as an estimate
of F (x). Consider the following theorem, somewhat out of sequence in the text
but worth seeing here for its motivational value. The reader may wish to take a
glance ahead at Appendix A.4 and Appendix C.5.
Theorem A.3. For every bounded continuous

function h(x), if Fn (x) con-
verges
in distribution to F (x), then h(x)dF n (x) converges in distribution to
h(x)dF (x).
This is the Helly-Bray theorem. The theorem will also hold (see the Exercises)
when h(x) is unbounded provided that some broad conditions are met. A deep
study of Fn (x) as an estimator of F (x) is then all that is needed to obtain insight
into the sample behavior of the empirical mean, the empirical variance and many
other quantities. Of particular importance for the applications of interest to us
here, and developed, albeit very briefly, in Appendix B.3, is the fact that, letting
M (x) = Fn (x) − F (x), then

E h(x)dM (x) = h(x)dF (x) − h(x)dF (x) = 0, (A.1)
a seemingly somewhat innocuous result until we interchange the order of inte-

gration (expectation, denoted by E being an integral operator), and under some
very mild conditions on h(x) described in Appendix B.3, we obtain a formulation
of great generality and into which can be fit many statistical problems arising in
the context of stochastic processes (see Appendix B.3).
A.3 Random variables and probability measure
The possible outcomes of any experiment are called events where any event
represents some subset of the sample space. The sample space is the collection
of all events, in particular the set of elementary events. A random variable X is
a function from the set of outcomes to the real line. A probability measure is a
function of some subset of the real line to the interval [0,1]. Kolmogorov (2018)
provides axioms which enable us to identify any measure as being a probability
measure. These axioms appear very reasonable and almost self-evident, apart
from the last, which concerns assigning probability measure to infinite collections
of events. There are, in a well defined sense, many more members in the set of
all subsets of any infinite set than in the original set itself, an example being
the set of all subsets of the positive integers which has as many members as the
real line. This fact would have hampered the development of probability without
the inclusion of Kolmogorov’s third axiom which, broadly says that the random
variable is measurable, or, in other words, that the sample space upon which the
probability function is defined is restricted in such a way that the probability we
associate with the sum of an infinite collection of mutually exclusive events is
the same as the sum of the probabilities associated with each composing event.
We call such a space a measurable space or a Borel space, the core idea being
that the property of additivity for infinite sums of probabilities, as axiomatized by
Kolmogorov, holds. The allowable operations on this space are referred to as a
sigma-algebra. Subsets of a sigma-algebra—the most common case being under
some kind of conditioning—are referred to as sub sigma-algebras and inherit the
axiomatic properties defined by Kolmogorov.
A great deal of modern probability theory is based on measure-theoretic ques-
tions, questions that essentially arise from the applicability or otherwise of Kol-
mogorov’s third axiom in any given context. This is an area that is highly techni-
cal and relatively inaccessible to non-mathematicians, or even to mathematicians
lacking a firm grounding in real analysis. The influence of measure theory has
been strongly felt in the area of survival analysis over the last 20 or so years
and much modern work is now of a very technical nature. Even so, none of the
main statistical ideas, or any of the needed demonstrations in this text, require
such knowledge. We can therefore largely avoid measure-theoretic arguments,
although some of the key ideas that underpin important concepts in stochas-
tic processes are touched upon whenever necessary. The reader is expected to
understand the meaning of the term random variable on some level.
Observations or outcomes as random variables and, via models, the proba-
bilities we will associate with them are all part of a theoretical, and therefore
artificial, construction. The hope is that these probabilities will throw light on
real applied problems and it is useful to keep in mind that, in given contexts,
there may be more than one way to set things up. Conditional expectation is a
recurring central topic but can arise in ways that we did not originally anticipate.
A.4. Convergence for random variables 355
We may naturally think of the conditional expected survival time given that a
subject begins the study under, say, some treatment. It may be less natural to
think of the conditional expectation of the random variable we use as a treatment
indicator given some value of time after the beginning of treatment. Yet, this
latter conditional expectation, as we shall see, turns out to be the more relevant
for many situations.
A.4 Convergence for random variables
Simple geometrical constructions (intervals, balls) are all that are necessary to
formalize the concept of convergence of a sequence in real and complex analy-
sis. For random variables there are a number of different kinds of convergence,
depending upon which aspect of the random variable we are looking at. Consider
any real value Z and the sequence Un = Z/n. We can easily show that Un → 0
as n → ∞. Now let Un be defined as before except for values of n that are
prime. Whenever n is a prime number then Un = 1. Even though, as n becomes
large, Un is almost always arbitrarily close to zero, a simple definition of conver-
gence would not be adequate and we need to consider more carefully the sizes
of the relevant sets in order to accurately describe this. Now, suppose that Z
is a uniform random variable on the interval (0,1). We can readily calculate the
probability that the distance between Un and 0 is greater than any arbitrarily
small positive number and this number goes to zero with n. We have conver-
gence in probability. Nonetheless there is something slightly erratic about such
convergence, large deviations occurring each time that n is prime. When possi-
ble, we usually prefer a stronger type of convergence. If, for all integer values m
greater than n and as n becomes large, we can assert that the probability of the
distance between Um and 0 being greater than some arbitrarily small positive
number goes to zero, then such a mode of convergence is called strong conver-
gence. This stronger convergence is also called convergence with probability one
or almost sure convergence. Consider also (n + 3)Un . This random variable will
converge almost surely to the random variable Z. But, also, we can say that the
distribution of loge (n + 3)Un , at all points of continuity z, becomes arbitrarily
close to that of a standard exponential distribution. This is called convergence
in distribution. The three modes of convergence are related by:
Theorem A.4. Convergence with probability one implies convergence in proba-
bility. Convergence in probability implies convergence in distribution.
Note also that, for a sequence that converges in probability, there exists a
subsequence that converges with probability one. This latter result requires the
tools of measure theory and is not of wide practical applicability since we may
not have any obvious way of identifying such a subsequence. Added conditions
can enable the direction of the “implies” arrow to be inverted. For example
convergence in distribution implies convergence in probability when the limiting
random variable is constant. In theoretical work it can sometimes be easier to

obtain results for weak rather than strong convergence. However, in practical
applications, we usually need strong (almost sure, “with probability one”) con-
vergence since this corresponds in a more abstract language to the important
idea that, as our information increases, our inferences becomes more precise.
Convergence of functions of random variables

In constructing models and establishing inference for them we will frequently
appeal to two other sets of results relating to convergence. The first of these is
that, for a continuous function g(z), if Zn converges in probability to c, then
g(Zn ) converges in probability to g(c) and, if Zn converges in distribution to Z,
then g(Zn ) converges in distribution to g(Z). The second set, Slutsky’s theorem
(a proof is given in Randles and Wolfe (1979)), enables us to combine modes of
convergence. In particular, for modeling purposes, if a convergence in distribution
result holds when the parameters are known, then it will continue to hold when
those same parameters are replaced by consistent estimators. This has great
practical value.
A.5 Topology and distance measures
Here, we describe some further tools that are helpful in determining large sample
behavior. Such behavior, in particular almost sure convergence and the law of
the iterated logarithm allow us to anticipate what we can expect in moderate to
large sample sizes. Small sample behavior is considered separately.
Compact topological spaces

For readers interested in the proofs given at the end of the chapters, we will need
to appeal on occasion to more formal concepts of convergence. These require
that the way in which we measure distance be made formal and the spaces in
which sit the functions of interest need to be defined. The main results we need
are obtained for spaces that are bounded and closed and these are called compact
spaces.
Topology of the spaces D[0, 1] and C[0, 1]

We denote by (C[0, 1], E) the space of continuous functions defined on the closed
interval [0, 1] assuming values in the space E and (D[0, 1], E) the space of right
continuous functions with left limits (called cadlag, a French acronym of this
definition). These are defined on [0, 1] with values in E. The context in which we
use these concepts relates to random elements of these two spaces with E = Rp ,
p ∈ N∗ . In our context, the interval [0, 1] corresponds to a transformation of time
to this interval, and the context usually makes it clear whether we are referring
A.5. Topology and distance measures 357
to transformed time or time on the original scale. Moving between the scales is
described precisely.
The interval [0,1] arises in a very natural way when dealing with distribu-
tion functions and it also appears natural to work with uniform distance and to
consider uniform convergence as the basic concept behind our metric definition.
With this in mind we have:
Definition A.3. The uniform distance d between two elements of (C[0, 1], R) is
defined by:
d(f, g) = sup |f (t) − g(t)| , f, g ∈ (C[0, 1], R). (A.2)
0≤t≤1
In order to get around the problem of functions with jumps at the disconti-
nuities, Skorokhod developed a more suitable metric providing the basis for the
Skorokhod topology. Specifically, we have the definition:
Definition A.4. Skorokhod distance δ is defined by:
δ(f, g) = inf {d(f, g(λ)) ∨ d(λ, I)} , f, g ∈ (D[0, 1], R), (A.3)
λ∈Λ
where I is the identity transform on [0, 1], Λ indicates the class of structure
preserving homomorphisms of [0, 1] into itself such that λ(0) = 0 and λ(1) = 1
for all f λ ∈ Λ and a ∨ b = max(a, b).
The idea is to not allow any jumps to dominate and is more fully explored
and explained in Billingsley (1999). Note also that
δ(f, g) ≤ d(f, g), f, g ∈ (D[0, 1], R). (A.4)
which implies the following result:

Proposition A.1. (Billingsley, 1999). Consider the sequence of functions, (fn )n∈N
of (D[0, 1], R) and f ∈ (D[0, 1], R). If
lim d(fn , f ) = 0, then lim δ(f, gn ) = 0.

n→+∞ n→+∞
In other words, in the space (D[0, 1], R), convergence with respect to a topol-
ogy of uniform convergence implies convergence with respect to the topology of
Skorokhod. At the same time, if the limit is a continuous function then the two
forms of convergence are equivalent.
Proposition A.2. (Billingsley, 1999). For all f ∈ (C[0, 1], R) ; (gn )n∈N ∈
(D[0, 1], R),
lim d(f, gn ) = 0 ⇐⇒ lim δ(f, gn ) = 0.

n→+∞ n→+∞
Some simple properties and the fact that cadlag functions arise naturally
when considering empirical distribution functions, enable us to find results easily.
The uniform limit of a sequence of cadlag functions is cadlag and the proposition
shows that it is equivalent to show convergence of a sequence of cadlag functions
to a continuous function with a uniform convergence topology or the Skorokhod
topology.
Topologies of (D[0, 1], Rp )

When we are specifically interested in higher dimensions then we can consider
the space of cadlag functions of [0, 1] in Rp where p is an integer strictly greater
than 1. In this case, several distances can be considered and, for the proofs
in the multivariate setting given here, we make use of the product distance of
Skorokhod (Chauvel, 2014).
Definition A.5. Let f = (f1 , . . . , fp ) and g = (g1 , . . . , gp ) be two functions on
(D[0, 1], Rp ), with fi and gi functions on (D[0, 1], R), for i = 1, . . . , p. The Sko-
rokhod product distance between f and g is defined by:
p

δp (f, g) = δ(fi , gi ).
i=1
In some cases it is worth considering the Skorokhod distance (Equation A.3) in

which the absolute value in Definition A.2 of uniform distance is replaced by the
sup norm such that a = maxi=1,...,p |ai |, for a ∈ Rp . The induced topology
corresponding to this distance is referred to as having the strong Skorokhod
property. The following proposition also turns out to be a useful tool.
Proposition A.3. For all functions in f ∈ (C[0, 1], Rp ) and for all function
sequences, (gn )n∈N∗ ∈ (D[0, 1], Rp ), if (gn )n∈N∗ converges to f with respect
to uniform convergence topology in Rp , i.e.
lim sup f (t) − gn (t) = 0, then lim δp (f, gn ) = 0.

n→∞ 0≤t≤1 n→∞
To summarize, if a sequence of functions of (D[0, 1], Rp ) converges to a

function of (C[0, 1], Rp ) with respect to the topology of uniform convergence,
then we can say that convergence is also obtained with respect to the product
topology of Skorokhod. We will use these results in the chapters concerning the
regression effect process in order to anticipate the properties for large samples of
certain empirical processes.
A.6 Distributions and densities
We anticipate that most readers will have some familiarity with the basic ideas
of a distribution function F (t) = Pr (T < t), a density function f (t) = dF (t)/dt,
expectation and conditional expectation, the moments of a random variable, and
A.6. Distributions and densities 359
other basic tools. Nonetheless we will go over these elementary notions in the
context of survival in the next chapter. We write

E ψ(T ) = ψ(t)f (t)dt = ψ(t)dF (t)
for the expected value of the function ψ(T ). Such an expression leaves much
unsaid, that ψ(t) is a function of t and therefore ψ(T ) itself random, that the
integrals exist, the domain of definition of the function being left implicit, and
that the density f (t) is an antiderivative of the cumulative distribution F (t) (in
fact, a slightly weaker mathematical construct, absolute continuity, is enough but
we do not feel the stronger assumption has any significant cost attached to it).
There is a wealth of solid references for the rusty reader on these topics, among
which Billingsley (1999), Rao et al. (1973) and Serfling (2009) are particularly
outstanding. It is very common to wish to consider some transformation of a
random variable, the simplest situation being that of a change in origin or scale.
The distribution of sums of random variables arises by extension to the bivariate
and multivariate cases.
Theorem A.5. Suppose that the distribution of X is F (x) and that F (x) =
f (x). Suppose that y = φ(x) is a monotonic function of x and that φ−1 (y) = x.
Then, if the distribution of Y is G(y) and G (y) = g(y),

dφ(x) −1
G(y) = F {φ−1 (y)} ; g(y) = f {φ−1 (y)} (A.5)
dx x=φ−1 (y) .
Theorem A.6. Let X and Y have joint density f (x, y). Then the density g(w)
of W = X + Y is given by
∞ ∞
g(w) = f (x, w − x)dx = f (w − y, y)dy. (A.6)
−∞ −∞
A result for W = X − Y follows immediately and, in the case of X and Y

being independent, the corresponding expression can also be written down readily
as a product of the two respective densities. Similar results hold for the product
or ratio of random variables (see Rohatgi and Saleh (2015), Chapter 8) but,
since we have no call for them in this work, we do not write them down here. An
immediate corollary that can give an angle on small sample behavior of statistics
that are written as sums is:
Corollary A.2. Let X1 , . . . , Xn be independent, not always identically distributed,
continuous
random variables with densities f1 (x) to fn (s) respectively. Let
Sn = n j=1 Xj . Then the density, gn (s), of Sn is given by
∞
gn (s) = gn−1 (s − x)fn (x)dx.
−∞
This result can be used iteratively for building up successive solutions by

carrying out the integration. The integration itself will mostly be not particularly
tractable and can be evaluated using numerical routines. Note the difference
between making a large sample statistical approximation to the sum and that
of a numerical approximation to the integral. The integral expression itself is an
exact result.
Normal distribution
A random variable X is taken to be a a normal variate with parameters μ and
σ when we write X ∼ Φ(μ, σ 2 ). The parameters μ and σ 2 are the mean and
variance, respectively, so that σ −1 (X − μ) ∼ Φ(0, 1). The distribution Φ(0, 1) is
called the standard normal. The density of the standard normal variate, that is,
having mean zero and variance one, is typically denoted φ(x) and the cumulative
distribution Φ(x). The density f (x), for x ∈ (−∞, ∞) is given by
2
1 1 x−μ
f (x) = φ(x) = √ exp − .
2πσ 2 σ
For stochastic processes described below, Brownian motion relates to a Gaussian

process, that is, it has been standardized, in an analogous way that the standard
normal relates to any other normal distribution. For the normal distribution, all
cumulants greater than 2 are equal to zero. Simple calculations show that, for
X ∼ N (0, 1), then E(X r ) = (r − 1)(r − 3) . . . 3.1. Thus, all odd moments are
equal to zero and all even moments are expressible in terms of the variance.
The normal distribution is of very great interest in view of it frequently being
the large sample limiting distribution for sums of random variables. These arise
naturally via simple estimating equations. These topics are looked at in greater
detail below.
The multivariate normal can be characterized in various ways. If and only if all
marginal distributions and all conditional distributions are normal then we have
multivariate normality. If and only if all linear combinations are univariate normal
then we have multivariate normality. It is only ∞ necessary to be able to evaluate
the standard normal integral, Φ(x) = 1 − x φ(x)dx, since any other normal
distribution, f (x), can be put in this form via the linear transformation (X −μ)/σ.
Tables, calculator, and computer routines can approximate the numerical integral.
Otherwise, it is worth bearing in mind the following;
Lemma A.1. Upper and lower bounds for the normal integral can be obtained
from
∞
x −x2 /2 2 1 2
2
e < e−u /2 du < e−x /2 .
1+x x x
A.6. Distributions and densities 361
The lemma tells us that we expect 1 − Φ(x) to behave like φ(x)/x as x

increases. The ratio φ(x)/x is known as Mill’s ratio. Approximate calculations
are then possible without the need to resort to sophisticated algorithms, although,
in modern statistical analysis, it is now so commonplace to routinely use com-
puters that the value of the lemma is rather limited. The normal distribution
plays an important role in view of the central limit theorem described below
but also note the interesting theorem of Cramér (2004) whereby, if a finite sum
of independent random variables is normal, then each variable itself is normal.
Cramer’s theorem might be contrasted with central limit theorems whereby sums
of random variables, under broad conditions, approach the normal as the sum
becomes infinitely large. The normal distribution is important since it provides the
basis for Brownian motion and this is the key tool that we will use for inference
throughout this text.
Uniform distribution and the probability integral transform

For the standard uniform distribution in which u ∈ [0, 1], f (u) = 1 and F (u) = u.
Uniform distributions on the interval [a, b] correspond to the density f (u) =
1/(b − a) but much more important is the fact that for any continuous distribu-
tion, G(t), we can say:
Theorem A.7. For the random variable T, having distribution G(t), letting
U1 = G(T ) and U2 = 1 − G(T ), then both U1 and U2 have a standard uniform
distribution.
This central result, underpinning a substantial body of work on simulation
and resampling, is known as the probability integral transform. Whenever we can
invert the function G, denoted G−1 , then, from a single uniform variate U we
obtain the two variates G−1 (U ) and G−1 (1 − U ) which have the distribution
G. The two variates are of course not independent but, in view of the strong
linearity property of expectation (the expectation of a linear function of random
variables is the same linear function of the expectations), we can often use this to
our advantage to improve precision when simulating. Another interesting conse-
quence of the probability integral transform is that there exists a transformation
of a variate T , with any given distribution, into a variate having any other chosen
distribution. Specifically, we have:
Corollary A.3. For any given continuously invertible distribution function H, and
continuous distribution G(t), the variate H −1 {G(T )} has distribution H.
In particular, it is interesting to consider the transformation Φ−1 {Gn (T )}
where Gn is the empirical estimate (discussed below) of G. This transformation,
which preserves the ordering, makes the observed distribution of observations
as close to normal as possible. Note that since the ordering is preserved, use
of the transformation makes subsequent procedures nonparametric in as much
as the original distribution of T has no impact. For the problems of interest to
us in survival analysis we can use this in one of two ways: firstly, to transform
the response variable time in order to eliminate the impact of its distribution,
and secondly, in the context of regression problems, to transform the distribution
of regressors as a way to obtain greater robustness by reducing the impact of
outliers.
Exponential distribution and cumulative hazard transformation

The standard exponential distribution is defined on the positive real line (0, ∞).
We have, for u ∈ (0, ∞), f (u) = exp(−u) and F (u) = 1 − exp(−u). An expo-
nential distribution with mean 1/α and variance 1/α2 has density f (u) =
α exp(−αu) and cumulative distribution F (u) = 1 − exp(−αu). The density
of a sum of m independent exponential variates having mean 1/α, is an
Erlang density whereby f (u) = α(αu)m−1 exp(−αu)/Γ(m) and where Γ(m) =
∞ m−1 du. The gamma distribution has the same form as the Erlang
0 exp(−u)u
although, for the gamma, the parameter m can be any real positive number and is
not restricted to being an integer. An exponential variate U can be characterized
as a power transformation on a Weibull variate in which F (t) = 1 − exp[(−αt)k ].
Finally, we have the important result:
Theorem A.8. For any continuous T positive random variable T , with distribution
function F (t), the variate U = 0 f (u)/[1 − F (u)]du has a standard exponential
distribution.
This result is important in survival modeling and we appeal to it frequently.
t
The function f (t)/[1 − F (t)] is known as the hazard function and 0 f (u)/[1 −
F (u)]du as the cumulative hazard function. The transformation is called the
cumulative hazard transformation.
A.7 Multivariate and copula models
The multivariate normal distribution, like its univariate counterpart, is completely

specified by two compound quantities, the vector of means and the variance-
covariance matrix. However, the assumptions of the multivariate situation are
much more restrictive than anything needed for the simple univariate case.
Indeed, if we use these restrictions to characterize the multivariate distribution,
then we see straight away how strong are the assumptions we make. For example,
if all marginal distributions are assumed normal, and all conditional distributions
are assumed normal, this is a minimal characterization of the multivariate normal
distribution. If all linear combinations of the components or sub-components of a
multivariate distribution are univariate normal then this again characterizes mul-
tivariate normality. In consequence an assumption of multivariate normality is a
very strong one. If we are comfortable in making this assumption the benefits are
great. All of the regressions are linear with constant expressions for the variances
A.7. Multivariate and copula models 363
and covariances. Measures of explained variation are obtained immediately in

terms of the covariances and variances, i.e., the correlation coefficients.
Finding useful models with weaker assumptions is not easy, in particular if
we wish to have some easily interpretable parameter that quantifies dependency
such as the correlation coefficient. Multivariate normal distributions have normal
marginals but, as mentioned above, having normal marginals is not enough to
result in a multivariate normal distribution. An infinity of distinct multivariate
distributions can have the same univariate marginals. Indeed, for any pair of
univariate marginals there exists an infinite number of multivariate distributions.
Even something as simple as having uniform marginals, unless independent, is
associated with an unlimited choice of candidate multivariate distributions having
those same marginals.
One approach to constructing multivariate distributions from univariate
marginals is described as a copula model. Again, the possibilities are infinite. For
example, suppose we have distribution functions, F (x) and G(y), with respective
densities, f (x) and g(y) for the variables X and Y. Consider the bivariate density
function, h(x, y ; θ) where;
(2)
h(x, y ; θ) = Φθ Φ−1 (ux ), Φ−1 (uy ) (A.7)
(2)
where Φθ indicates the standardized bivariate normal distribution with correla-
tion coefficient θ, Φ(·) the univariate normal distributions and ux , uy , indepen-
dent uniform variates. The probability integral transform, described just above,
tells us that, for continuous F (x) and G(y), we have a one-to-one correspon-
dence with the uniform distribution. We can break down the steps by starting
with X and Y , transforming these to the uniform interval via ux = F (x) and
uy = G(y), subsequently transforming once more to normal marginals via Φ−1 (·),
and finally, via the bivariate normal distribution with parameter θ, creating the
bivariate model with normal (0,1) marginals and association parameter θ. We
can then use F −1 and G−1 to return to our original scale. This is referred to as
the normal copula and the creation of a bivariate model with a particular asso-
ciation parameter, θ, is, in this case, quite transparent. Also, the dependency
parameter, θ, has a concrete interpretation as a measure of explained variation
on the transformed scales. Clearly, this set-up is readily generalized by replacing
the first and second occurrences of Φ in Equation A.7 by any other distribution
(2)
functions, not necessarily the same, and by replacing Φθ by a different joint
distribution with a different dependency structure.
There are very many more, in fact an infinite number of alternative ways to
create bivariate models with some association structure from given marginals.
Many of these have been used in the survival context, in particular when dealing
with the problem of competing risks, or when considering surrogate endpoints.
These are mentioned in the main text. Building and exploiting the structure of
copula models is a field in its own right discussed in many texts and articles. A
thorough and clear discussion is given in Trivedi and Zimmer (2007).
A.8 Expectation
It is worth saying a word or two more about expectation as a fundamental

aspect of studies in probability. Indeed it is possible for the whole theory to
be constructed with expectation as a starting point rather than the now classical
axiomatic structure to probability. For a function of a random variable T , ψ(T )
say, as stated at the beginning of the previous section, we write, E( ψ(T ) of this
function via

E ψ(T ) = ψ(t)f (t)dt = ψ(t)dF (t),
where the integrals, viewed as limiting processes, are all assumed to converge.
The normal distribution function for a random variable X is completely specified
by E(X) and E(X 2 ). In more general situations we can assume a unique corre-
spondence between the moments of X, E(X r ) , r = 1, 2, . . . , and the distribution
functions as long as these moments all exist. While it is true that the distribution
function determines the moments the converse is not always true. However, it is
almost always true (Kendall et al., 1987) and, for all the distributions of interest
to us here, the assumption can be made without risk. It can then be helpful to
view each moment, beginning with E(X), as providing information about F (x).
This information typically diminishes quickly with increasing r. We can use this
idea to improve inference for small samples when large sample approximations
may not be sufficiently accurate. Moments can be obtained from the moment
generating function, M (t) = E{exp(tX)} since we have:

Lemma A.2. If exp(tx)f (x)dx < ∞ then
r
r ∂ M (t)
E(X ) = , for all r.
∂tr t=0
We often focus on the variance function which is also an expectation and is

of particular interest to one of our central goals here, that of constructing useful
measures of the predictive strength of any model. At the root of the construction
lie two important inequalities, the Chebyshev-Bienaymé inequality and Jensen’s
inequality described below. For this we first need:
Definition A.6. The real-valued function w(x) is called “convex” on some inter-
val I (an infinite set and not just a point) whenever, for x1 , x2 ∈ I and for
0 ≤ λ ≤ 1, we have
w[λx1 + (1 − λ)x2 ] ≤ λw(x1 ) + (1 − λ)w(x2 ).

A.8. Expectation 365
It is usually sufficient to take convexity to mean that w (x) and w (x) are
greater than or equal to zero at all interior points of I since this is a consequence
of the definition. We have (Jensen’s inequality):
Lemma A.3. If w is convex on I then, assuming expectations exist on this

interval, w[E(X)] ≤ E[w(X)]. If w is linear in X throughout I, that is, w (x) = 0
when twice differentiable, then equality holds.
For the variance function we see that w(x) = x2 is a convex function and so
the variance is always positive. The further away from the mean, on average, the
observations are to be found, then the greater the variance. We return to this
in Chapter 10. Although very useful, the moment-generating function, M (t) =
E{exp(tX)} has a theoretical weakness in that the integrals may not always
converge. It is for this, mainly theoretical, reason that it is common to study
instead the characteristic function, which has an almost identical definition, the
only difference being the introduction of complex numbers into the setting. The
characteristic function, denoted by φ(t), always exists and is defined as:
∞
φ(t) = M (it) = exp(itx)dF (x) , i2 = −1.
−∞
Note that the contour integral in the complex plane is restricted to the whole real
axis. Analogous to the above lemma concerning the moment-generating function
we have: r
r r ∂ φ(t)
E(X ) = (−i) , for all r.
∂tr t=0
This is important in that it allows us to anticipate the cumulative generating
function which turns out to be of particular importance in obtaining improved
approximations to those provided by assuming normality. We return to this below
in Section A.10. If we expand the exponential function then we can write:
∞ ∞

r
φ(t) = exp(itx)dF (x) = exp κr (it) /r!
−∞ r=1
and, identifying κr as the coefficient of (it)r /r! in the expansion of log φ(t). The
function ψ(t) = log φ(t) is called the cumulative generating function. When this
function can be found then the density f (x) can be defined in terms of it. We
have the important relation
∞ ∞
1
f (x) = e−itx φ(t)dt , φ(t) = eitx f (x)dx .
2π −∞ −∞
It is possible to approximate the density f (x) by working with i.i.d.

nobservations
X1 , · · · , Xn and the empirical characteristic function φ(t) = n−1
i=1 exp(itxi )
which can then be inverted. It is also possible to approximate the integral using a
method of numerical analysis, the so-called method of steepest descent, to obtain

a saddlepoint approximation (Daniels, 1983, 1954, 1980, 1987). We return to this
approximation below in Section A.10.
A.9 Order statistics and their expectations
The normal distribution and other parametric distributions described in the

next chapter play a major role in survival modeling. However, the robustness
of any inferential technique to particular parametric assumptions is always a con-
cern. Hopefully, inference is relatively insensitive to departures from parametric
assumptions or is applicable to whole families of parametric assumptions. The
most common way to ensure this latter property is via the theory of order statistics
which we recall here. Consider the n independent identically distributed (i.i.d.)
random variables: X1 , X2 , ... , Xn and a single realization of these that we can
order from the smallest to the largest: X(1) ≤ X(2) ≤ · · · ≤ X(n) . Since the Xi
are random, so also are the X(i) , and the interesting question concerns what we
can say about the probability structure of the X(i) on the basis of knowledge of
the parent distribution of Xi . In fact, we can readily obtain many useful results
which, although often cumbersome to write down, are in fact straightforward.
Firstly we have:
Theorem A.9. Taking P (x) = Pr (X ≤ x) and Fr (x) = Pr (X(r) ≤ x) then:
n
n
Fr (x) = P i (x)[1 − P (x)]n−i . (A.8)
i
i=r
This important result has two immediate and well-known corollaries dealing
with the maximum and minimum of a sample of size n.
Corollary A.4.
Fn (x) = P n (x) , F1 (x) = 1 − [1 − P (x)]n (A.9)
In practice, in order to evaluate Fr (x) for other than very small n, we

exploit the equivalence between partial binomial sums and the incomplete
1 a−1
beta function. Thus,
π a−1 if, for a > 0, b > 0, B(a, b) = 0 t (1 − t)b−1 dt and
Iπ (a, b) = 0 t (1 − t) b−1 dt/B(a, b), then putting P (x) = π, we have that
Fr (x) = Iπ (r, n − r + 1). These functions are widely tabulated and also avail-
able via numerical algorithms to a high level of approximation. An alternative,
although less satisfying, approximation would be to use the DeMoivre-Laplace
normal approximation to the binomial sums. Differentiation of (A.8) provides the
density which can be written as
1
fr (x) = P r−1 (x)[1 − P (x)]n−r p(x). (A.10)
B(r, n − r + 1)
A.9. Order statistics and their expectations 367
Since we have a relatively straightforward expression for the distribution function

itself, then this expression for the density is not often needed. It can come in
handy in cases where we need to condition and apply the law of total probability.
Expressions for f1 (x) and fn (x) are particularly simple and we have:
Corollary A.5.
f1 (x) = n[1 − P (x)]n−1 p(x) , fn (x) = nP n−1 (x)p(x). (A.11)
More generally it is also straightforward to obtain

Theorem A.10. For any subset of the n order statistics: Xn1 , Xn2 , ... , Xnk ,
1 ≤ n1 ≤ . . . ≤ nk , the joint distribution f (x1 , . . . , x2 ) is expressed as
⎡ ⎤
k
k

[P (xj+1 ) − P (xj )]nj+1 −nj −1
f (x1 , . . . , xk ) = n! ⎣ p(xj )⎦ (A.12)
(nj+1 − nj − 1)!
j=1 j=0
in which p(x) = P (x). This rather involved expression leads to many useful
results including the following corollaries:
Corollary A.6. The joint distribution of X(r) and X(s) is
j
n
n!
Frs (x, y) = P i (x)[P (y) − P (x)]j−i [1 − P (y)]n−j .
i!(j − i)!(n − j)!
j=s i=r
The joint distribution of X(r) and X(s) is useful in establishing a number

of practical results such as the distribution of the range, the distribution of the
interquartile range and an estimate for the median among others. Using the result
(Appendix A.6) for the distribution of a difference, a simple integration then leads
to the following:
Corollary A.7. Letting Wrs = X(s) − X(r) then: in the special case of a parent
uniform distribution we have
1
f (wrs ) = ws−r−1 (1 − wrs )n−s+r . (A.13)
B(s − r, n − s + r + 1) rs
Taking s = n and r = 1, recalling that B(α, β) = Γ(α)Γ(β)/Γ(α + β) and

that Γ(n) = n!, then we have the distribution of the range for the uniform.
Corollary A.8. Letting w = U(n) − U(1) be the range for a random sample of
size n from the standard uniform distribution, then the cumulative distribution
is given by
FU (w) = nwn−1 − (n − 1)wn . (A.14)
Straightforward differentiation gives fU (w) = n(n − 1)wn−2 (1 − w), a simple

and useful result. For an arbitrary distribution, F (·) we can either carry out the
same kind of calculations from scratch or, making use once more of the probability
integral transform (see Section A.3), use the above result for the uniform and
transform into arbitrary F . Even this is not that straightforward since, for some
fixed interval (w1 , w2 ), corresponding to w = w2 − w1 from the uniform, the
corresponding F −1 (w2 ) − F −1 (w1 ) depends not only on w2 − w1 but on w1
itself. Again we can appeal to the law of total probability, integrating over all
values of w1 from 0 to 1 − w. In practice, it may be good enough to divide the
interval (0, 1 − w) into a number of equally spaced points, ten would suffice, and
simply take the average. Interval estimates for any given quantile, defined by
P (ξα ) = α, follow from the basic result and we have:
Corollary A.9. In the continuous case, for r < s, the pair (X(r) , X(s) ) covers ξα
with probability given by Iπ (r, n − r + 1) − Iπ (r, n − s + 1).
Theorem A.11. For the special case in which n1 = 1, n2 = 2, ... nn = n, then

n

f (x1 , . . . , xn ) = n! p(xj ). (A.15)
j=1
A characterization of order statistics: Markov property

The particularly simple results for the exponential distribution lead to a very useful
and powerful characterization of order statistics. If Z1 , . . . , Zn are i.i.d. exponen-
tial variates with parameter λ, then an application of Corollary A.4 shows that
the minimum of Z1 to Zn has itself an exponential distribution with parameter
nλ. We can define the random variable Y1 to be the gap time between 0 and
the first observation, Z(1) . The distribution of Y1 (equivalently Z(1) ) is expo-
nential with parameter nλ. Next, we can define Y2 to be the gap Z(2) − Z(1) .
In view of the lack of memory property of the exponential distribution, once
Z(1) is observed, the conditional distribution of each of the remaining (n − 1)
variables, given that they are all greater than the observed time Z(1) , remains
exponential with parameter λ. The variable Y2 is then the minimum of (n − 1)
i.i.d. exponential variates with parameter λ. The distribution of Y2 is therefore,
once again, exponential, this time with parameter (n − 1)λ. More generally we
have the following lemma:
Lemma A.4. If Z(1) , . . . , Z(n) are the order statistics from a sample of size n of
standard exponential variates, then, defining Z(0) = 0,
Yi = Z(i) − Z(i−1) , i = 1, . . . , n
are n independent exponential variates in which E(Yi ) = 1/(n − i + 1).

This elementary result is very important in that it relates the order statistics
directly to sums of simple independent random variables which are not themselves
order statistics. Specifically we can write:
A.9. Order statistics and their expectations 369
r
r

Z(r) = {Z(i) − Z(i−1) } = Yi ,
i=1 i=1
leading to the immediate further lemma:
Lemma A.5. For a sample of size n from the standard exponential distribution
and letting αi = 1/(n − i + 1), we have:
r
r
r
r

E[Z(r) ] = E(Yi ) = αi , Var [Z(r) ] = Var (Yi ) = αi2 .
i=1 i=1 i=1 i=1
The general flavor of the above result applies more generally than just to the
exponential and, applying the probability integral transform (Section A.6), we
have:
Lemma A.6. For an i.i.d. sample of size n from an arbitrary distribution, G(x),
the rth largest order statistic, X(r) can be written:
X(r) = G−1 {1 − exp(−Y1 − Y2 − · · · − Yr )},
where the Yi are independent exponential variates in which E(Yi ) = 1/(n−i+1).
One immediate conclusion that we can make from the above expression is
that the order statistics from an arbitrary distribution form a Markov chain.
The conditional distribution of X(r+1) given X(1) , X(2) , . . . , X(r) depends only
on the observed value of X(r) and the distribution of Yr+1 . This conditional
distribution is clearly the same as that for X(r+1) given X(r) alone, hence the
Markov property. If needed we can obtain the joint density, frs , of X(r) and
X(s) , (1 ≤ r < s ≤ n) by a simple application of Theorem A.10. We then write:
n! P r−1 (x)p(x)p(y)[P (y) − P (x)]s−r−1 [1 − P (y)]n−s

frs (x, y) = .
(r − 1)!(s − r − 1)!(n − s)!
From this we can immediately deduce the conditional distribution of X(s) given
that X(r) = x as:
(n − r)! p(y)[P (y) − P (x)]s−r−1 [1 − P (y)]n−s

fs|r (y|x) = .
(s − r − 1)!(n − s)! [1 − P (x)]n−r
A simple visual inspection of this formula confirms again the Markov property.
Given that X(r) = x we can view the distribution of the remaining (n − r) order
statistics as an ordered sample of size (n − r) from the conditional distribution
P (u|u > x).
Expected values of order statistics

Given the distribution of any given order statistic we can, at least in principle,
calculate any moments, in particular the mean, by applying the basic definition.
In practice, this may be involved and there may be no explicit analytic solution.
Integrals can be evaluated numerically but, in the majority of applications, it can
be good enough to work with accurate approximations. The results of the above
subsection, together with some elementary approximation techniques are all that
we need. Denoting the distribution of X as P (x), then the probability integral
transform (Appendix A.6) provides that U = P (X) has a uniform distribution.
The moments of the order statistics from a uniform distribution are particularly
simple so that: E{U(r) } = pr = r/(n + 1)). Denoting the inverse transformation
by Q = P −1 , then
X(r) = P −1 {U(r) } = Q{U(r) }.
Next, we can use a Taylor series development of the function X(r) about the pr
so that:
X(r) = Q(pr ) + {U(r) − pr }Q (pr ) + {U(r) − pr }2 Q (pr )/2 + · · ·
and, taking expectations, term by term, we have:

pr qr pr qr 1 1
E{X(r) } ≈ Q(pr )+ Q (pr )+ (qr − pr )Q (pr ) + pr qr Q (pr )
2(n + 2) (n + 2)2 3 8
and
pr q r
Var {X(r) } = [Q (pr )]2 +
2(n + 2)
pr q r
2(qr − pr )Q (pr )Q (pr ) + pr qr Q (pr )Q (pr ) + [Q (pr )]2 .
(n + 2)2
It is straightforward to establish some relationships between the moments of the

order statistics and the moments from the parent distribution. Firstly note that:
n m n m

k k
E X(r) =E Xr ,
r=1 r=1
so that, if μ and σ 2
are the mean and variance in the parent population, then
n n 2 2 2 2
μ
r=1 r = nμ and r=1 E{X(r) } = nE(X ) = n(μ + σ ).
Normal parent distribution

For the case of a normal parent the expected values can be evaluated precisely
for small samples and the approximations themselves are relatively tractable for
larger sample sizes. One approach to data analysis in which it may be desirable to
A.10. Approximations 371
have a marginal normal distribution in at least one of the variables under study is
to replace the observations by the expectations of the order statistics. These are
sometimes called normal scores, typically denoted by ξrn = E(X(r) ) for a random
sample of size n from a standard normal parent with distribution function Φ(x)
and density φ(x). For a random sample of size n from a normal distribution with
mean μ and variance σ 2 we can reduce everything to the standard case since
E(X(r) ) = μ + ξrn σ. Note that, if n is odd, then, by symmetry, it is immedi-
ately clear that E(X(r) ) = 0 for all r that are odd. We can see that E(X(r) ) =
−E(X(n−r+1) ). For n as small as, say, 5 we can use integration by parts to
evaluate ξr5 for different values of r. For example, ξ55 = 5 4Φ (x)φ2 (x)dx
3
which then simplifies to: ξ55 = 5π −1/2 /4 + 15π −3/2 sin−1 (1/3)/2 = 1.16296.
Also, ξ45 = 5π −1/2 /2 − 15π −3/2 sin−1 (1/3) = 0.49502 and ξ35 = 0. Finally,
ξ15 = −1.16296 and ξ25 = −0.49502. For larger sample sizes in which the inte-
gration becomes too fastidious we can appeal to the above approximations using
the fact that
1 Q 1 + 2Q2 Q(7 + 6Q2 )
Q (pr ) = , Q (pr )= 2 , Q (pr )= 3 , Q (pr ) = .
φ(Q) φ (Q) φ (Q) φ4 (Q)
The above results arise from straightforward differentiation. Analogous calcula-

tions can be used to obtain exact or approximate expressions for Cov {X(r) , X(s) }.
A.10 Approximations
Approximations to means and variances for functions of T (δ-method)

Consider some differentiable monotonic function of X, say ψ(X). Our particular
concern often relates to parameter estimates in which case the random variable
X would be some function of the n i.i.d. data values, say θn as an estimator of
the parameter θ. In the cases of interest, θn converges with probability one to
θ and so also does ψ(θn ) to ψ(θ). Although θn may not be unbiased for θ, for
large samples, the sequence E(θn ) converges to E(θ) = θ. Similarly E[ψ(θn )]
converges to ψ(θ). The mean value theorem (Section A.2) enables us to write
(θn − θ)2
φ(θn ) = ψ(θ) + (θn − θ)φ (θ) + ψ (ξ) (A.16)
2
for ξ ∈ (θ ± θn ) Rearranging this expression, ignoring the third term on the right-
hand side, and taking expectations we obtain
Var {ψ(θn )} ≈ E{ψ(θn ) − ψ(θ)}2 ≈ {ψ (θ)}2 Var (θn ) ≈ {ψ (θn )}2 Var (θn )
as an approximation to the variance. The approximation, once obtained in any

given setting, is best studied on a case-by-case basis. It is an exact result for
linear functions. For these, the second derivative is equal to zero and, more
generally, the smaller the absolute value of this second derivative, the better
we might anticipate the approximation to be. For θn close to θ the squared
term will be small in absolute value when compared with the linear term, an
additional motivation to neglecting the third term. For the mean, the second
term of Equation (A.16) is zero when θn is unbiased, otherwise close to zero
and, this time, ignoring this second term, we obtain
1
E{ψ(θn )} ≈ ψ(θn ) + Var (θn )ψ (θn ) (A.17)
2
as an improvement over the rougher approximation based on the first term alone
of the above expression. Extensions of these expressions to the case of a consistent
estimator ψ(θn ) = ψ(θ1n , . . . , θpn ) of ψ(θ) proceeds in the very same way, only
this time based on a multivariate version of Taylor’s theorem. These are:
p
p
∂ψ(θ) ∂ψ(θ)
Var {ψ(θn )} ≈ Cov (θjn , θmn ) ,
∂θj ∂θm
j=1 m≥j
1 ∂ 2 ψ(θn )
E{ψ(θn )} ≈ ψ(θ1n , . . . , θpn ) + Cov (θjn , θmn ).
2 m
∂θj ∂θm
j
When p = 1 then the previous expressions are recovered as special cases. Again,
the result is an exact one in the case where ψ(·) is a linear combination of the
components θj and this helps guide us in situations where the purpose is that of
confidence interval construction. If, for example, our interest is on ψ and some
strictly monotonic transformation of this, say ψ ∗ , is either linear or close to linear
in the θj , then it may well pay, in terms of accuracy of interval coverage, to use
the delta-method on ψ ∗ , obtaining the end points of the confidence interval for
ψ ∗ and subsequently inverting these, knowing the relationship between ψ and
ψ ∗ , in order to obtain the interval of interest for ψ. Since ψ and ψ ∗ are related by
one-to-one transformations then the coverage properties of an interval for ψ ∗ will
be identical to those of its image for ψ. Examples in this book include confidence
intervals for the conditional survivorship function, given covariate information,
based on a proportional hazards model as well as confidence intervals for indices
of predictability and multiple coefficients of explained variation.
Cornish-Fisher approximations
In the construction of confidence intervals, the δ-method makes a normal-
ity approximation to the unknown distribution and then replaces the first two
moments by local linearization. A different approach, while still working with a
normal density φ(x) = (2π)−1/2 exp(−x2 /2), in a way somewhat analogous to
the construction of a Taylor series, is to express the density of interest, f (x), in
terms of a linear combination of φ(x) and derivatives of φ(x). Normal distribu-
tions with non-zero means and variances not equal to one are obtained by the
usual simple linear transformation and, in practical work, the simplest approach
is to standardize the random variable X so that the mean and variance corre-
sponding to the density f (x) are zero and one, respectively.
The derivatives of φ(x) are well-known, arising in many fields of mathe-
matical physics and numerical approximations. Since φ(x) is simply a constant
multiplying an exponential term it follows immediately that all derivatives of φ(x)
are of the form of a polynomial that multiplies φ(x) itself. These polynomials
(apart from an alternating sign coefficient (−1)i ) are the Hermite polynomials,
Hi (x) , i = 0, 1, . . . , and we have:
H0 = 1 , H1 = x , H2 = x2 − 1 , H3 = x3 − 3x , H4 = x4 − 6x2 + 3 ,
with H5 and higher terms being calculated by simple differentiation. The polyno-
mials are of importance in their own right, belonging to the class of orthogonal
polynomials and useful in numerical integration. Indeed, we have that:
∞ ∞
2
Hi (x)φ(x)dx = i! , i = 0, . . . : Hi (x)Hj (x)φ(x)dx = 0 , i = j.
−∞ −∞
This orthogonality property is exploited in order for us to obtain explicit expres-

sions for the coefficients in our expansion. Returning to our original problem we
wish to determine the coefficients ci in the expansion
∞

f (x) = ci Hi (x)φ(x) (A.18)
i=0
and, in order to achieve this we multiply both sides of equation (A.18) by Hj (x),
subsequently integrating to obtain the coefficients

1 ∞
cj = f (x)Hj (x)dx. (A.19)
j! −∞
Note that the polynomial Hj (x) is of order j so that the right-hand side of
equation (A.19) is a linear combination of the moments, (up to the jth), of
the random variable X having associated density f (x). These can be calculated
step-by-step. For many standard densities several of the lower-order moments
have been worked out and are available. Thus, it is relatively straightforward to
approximate some given density f (x) in terms of a linear combination of φ(x).
The expansion of Equation (A.18) can be used in theoretical investigations
as a means to study the impact of ignoring higher-order terms when we make
a normal approximation to the density of X. We will use the expansion in an
attempt to obtain more accurate inference for proportional hazards models fitted
using small samples. Here the large sample normal assumption may not be suf-
ficiently accurate and the approximating equation is used to motivate potential
improvements obtained by taking into account moments of higher order than
just the first and second. When dealing with actual data, the performance of any
such adjustments need to be evaluated on a case-by-case basis. This is because
theoretical moments will have to be replaced by observed moments and the sta-
tistical error involved in that can be of the same order, or greater, than the error
involved in the initial normal approximation. If we know or are able to calculate
the moments of the distribution, then the ci are immediately obtained. When
the mean is zero we can write down the first four terms as:
c0 = 1 , c1 = 0 , c2 = (μ2 − 1)/2 , c3 = μ3 /6 , c4 = (μ4 − 6μ2 + 3)/24 ,
from which we can write down an expansion in terms of φ(x) as
f (x)=φ(x){1 + (μ2 − 1)H2 (x)/2+μ3 H3 (x)/6 + (μ4 − 6μ2 + 3)H4 (x)/24 + · · · }.
This series is known as the Gram-Charlier series, and stopping the development
at the fourth term corresponds to making corrections for skewness and kurtosis.
In the development of the properties of estimators in the proportional hazards
model we see that making corrections for skewness can help make inference more
accurate, whereas, at least in that particular application, corrections for kurtosis
appear to have little impact (Chapter 7).
Saddlepoint approximations
A different, although quite closely related, approach to the above uses saddlepoint
approximations. Theoretical and practical work on these approximations indicate
them to be surprisingly accurate for the tails of a distribution. We work with the
inversion formula for the cumulant generating function, a function that is defined
in the complex plane, and in this two-dimensional plane, around the point of
interest (which is typically a mean or a parameter estimate) the function looks
like a minimum in one direction and a maximum in an orthogonal direction: hence
the name “saddlepoint.” Referring back to Section A.8 recall that we identified
κr as the coefficient of (it)r /r! in the expansion of the cumulant generating
function K(t) = log φ(t) where φ(t) is the characteristic function. We can exploit
the relationship between φ(t) and f (x); that is:
∞ ∞
1
f (x) = e−itx φ(t)dt , φ(t) = eitx f (x)dx .
2π −∞ −∞
to approximate f (x) by approximating the integral. The numerical technique

that enables this approximation to be carried out is called the method of steepest
descent and is described in Daniels (1954). The approximation to f (x) is simply
denoted as fs (x), and carrying through the calculations, we find that
1/2
n
fs (x) = exp[n{K(λx ) − xλx }] (A.20)
2πK (λx )
in which the solution to the differential equation in λ, K (λ) = x is given by

λx . Our notation here of x as a realization of some random variable X is not
specifically referring to our usual use of X as the minimum of survival time T
and the censoring time C. It is simply the variable of interest and that variable,
in our context, will be the score statistic arising from the estimating equation
(Chapter 7). For now, we assume the score to be composed of n contributions
so that we view x as a mean based on n observations. Since, mostly, we are
interested in the tails of the distribution, it can often help to approximate the
cumulative distribution directly rather than make a subsequent appeal to numer-
ical integration. Denoting the saddlepoint approximation to the cumulative dis-
tribution by Fs (x), we write
Fs (x) = Φ(ux ) + φ(ux )(u−1 −1

x + vx ) (A.21)
x
where φ(x) indicates the standard normal density, Φ(x) = −∞ φ(u)du, the cumul-
ative normal, ux = [2n{xλx − K(λx )}]1/2 sgn(λx ), and vx = λx {nK (λx )}1/2 .
Since we are only concerned with tail probabilities we need not pay attention to
what occurs around the mean. If we do wish to consider Fs (x), evaluated at the
mean, the approximation is slightly modified and the reader is referred to Daniels
(1987).
Appendix B
Stochastic processes
B.1 Broad overview
We define a stochastic process to be a collection of random variables indexed

by t ∈ T . We write these as X(t), or Xt , and take t to be fixed. If the set T
has only a finite or a countably infinite number of elements then X(t) is referred
to as a discrete-time process. We will be most interested in continuous-time
processes. In applications we can standardize by the greatest value of t in the set
T that can be observed, and so we usually take sup{t : t ∈ T } = 1. We also take
inf{t : t ∈ T } = 0. We will be especially interested in observations on any given
process between 0 and t. We call this the sample path.
Independent increments and stationarity

Consider some partition of (0,1) in which 0 = t0 < t1 < t2 < · · · < tn = 1. If the
set of random variables X(ti ) − X(ti−1 ) i = 1, . . . , n are independent then the
stochastic process X(t) is said to have independent increments. Another impor-
tant property is that of stationarity. We say that a stochastic process X(t) has
stationary increments if X(s + t) − X(s) has the same distribution for all values
of s. Stationarity indicates, in as much as probabilistic properties are concerned,
that when we look forward, from the point s, a distance t, the only relevant
quantity is how far forward t we look. Our starting point itself is irrelevant. As
we progress through time, everything that we have learned is summarized by the
current position. It can also be of value to consider a process with a slighter
weaker property, the so-called second-order stationarity. Rather than insist on
a requirement for the whole distribution we limit our attention to the first two
moments and the covariance between X(s + t) and X(s) which depends only
upon |t|. Our main focus is on Gaussian processes which, when they have the
property of second-order stationarity, will in consequence be stationary processes.

https://doi.org/10.1007/978-3-030-33439-0
378 Appendix B. Stochastic processes
Also, simple transformations can produce stationary processes from nonstation-

ary ones, an example being the transformation of the Brownian bridge into an
Ornstein-Uhlenbeck process.
Gaussian processes
If for every partition of (0,1), 0 = t0 < t1 < t2 < · · · < tn = 1, the set of random
variables X(t1 ), . . . , X(tn ) has a multivariate normal distribution, then the pro-
cess X(t) is called a Gaussian process. Brownian motion, described below, can be
thought of as simply a standardized Gaussian process. A Gaussian process being
uniquely determined by the multivariate means and covariances it follows that
such a process will have the property of stationarity if for any pair (s, t : t > s),
Cov {X(s), X(t)} depends only on (t − s). In practical studies we will often deal
with sums indexed by t and the usual central limit theorem will often underlie
the construction of Gaussian processes.
B.2 Brownian motion
Consider a stochastic process X(t) on (0, 1) with the following three properties:
1. X(0) = 0, i.e., at time t = 0 the starting value of X is fixed at 0.
2. X(t) , t ∈ (0, 1) has independent stationary increments.
3. At each t ∈ (0, 1) the distribution of X(t) is N (0, t).
This simple set of conditions completely describes a uniquely determined stochas-
tic process called Brownian motion. It is also called the Wiener process or Wiener
measure. It has many important properties and is of fundamental interest as a
limiting process for a large class of sums of random variables on the interval (0,1).
An important property is described in Theorem B.1 below. Firstly we make an
attempt to describe just what a single realization of such a process might look
like. Later we will recognize the same process as being the limit of a sum of
independent random contributions. The process is continuous and so, approxi-
mating it by any drawing, there cannot be any gaps. At the same time, in a sense
that can be made more mathematically precise, the process is infinitely jumpy.
Nowhere does a derivative exist. Figure B.1 illustrates this via simulated approx-
imations. The right-hand figure could plausibly be obtained from the left-hand
one by homing in on any small interval, e.g., (0.20, 0.21), subtracting off the
value observed at t = 0.20, and rescaling by a multiple of ten to restore the inter-
val of length 0.01 to the interval (0,1). The point we are trying to make is that
the resulting process itself looks like (and indeed is) a realization of Brownian
motion. Theoretically, this could be repeated without limit which allows us to
understand in some way how infinitely jumpy is the process. In practical exam-
ples we can only ever approximate the process by linearly connecting up adjacent
simulated points.
B.2. Brownian motion 379
Figure B.1: Two independently simulated Brownian motions on the interval (0,1).
Theorem B.1. Conditioning on a given path we have:
Pr {X(t + s) > x|X(s) = xs , X(u), 0 ≤ u < s}

= Pr {X(t + s) > x|X(s) = xs .}
So, when looking ahead from time point s to time point t + s, the previous
history indicating how we arrived at s is not relevant. The only thing that matters
is the point at which we find ourselves at time point s. This is referred to as the
Markov property. The joint density of X(t1 ), . . . , X(tn ) can be written as:
f (x1 , x2 , . . . , xn ) = ft1 (x1 )ft1 −t2 (x2 − x1 ) · · · ftn −tn−1 (xn − xn−1 )
This follows from the independent stationary increment condition. A consequence

of the above result is that we can readily evaluate the conditional distribution of
X(s) given some future value X(t) (t > s). Applying the definition for conditional
probability we have the following:
Corollary B.1. The conditional distribution of X(s) given X(t) (t > s) is normal
with a mean and a variance given by:
E{X(s)|X(t) = w} = ws/t , Var {X(s)|X(t) = w} = s(t − s)/t .
This result helps provide insight into another useful process, the Brownian
bridge described below. Other important processes arise as simple transformations
of Brownian motion. The most obvious to consider is where we have a Gaussian
process satisfying conditions (1) and (2) for Brownian motion but where, instead
of the variance increasing linearly, i.e., Var X(t) = t, the variance increases either
too quickly or too slowly so that Var X(t) = φ(t) where φ(·) is some monotonic
increasing function of t. Then we can transform the time axis using φ(·) to
produce a process satisfying all three conditions for Brownian motion. Consider
also the transformation
V (t) = exp(−αt/2)X{exp(αt)}
where X(t) is Brownian motion. This is the Ornstein-Uhlenbeck process. It is

readily seen that:
Corollary B.2. The process V (t) is a Gaussian process in which E{V (t)} = 0
and Cov {V (t), V (s)} = exp{−α(t − s)/2}.
Note that VarV (t) is a constant not depending on t, so that V (t) is often
referred to as variance stabilizing.
Time-transformed Brownian motion

Consider a process, X ψ (t), defined via the following three conditions, for some
continuous ψ such that, ψ(t ) > ψ(t) (t > t); (1) X ψ (0) = 0 (2) X ψ (t) , t ∈ (0, 1)
has independent stationary increments; (3) at each t ∈ (0, 1) the distribution of
X ψ (t) is N {0, ψ(t)}. The usual Brownian motion described above is exactly this
process when ψ(t) = t. However, in view of the continuity and monotonicity of
ψ, there exists an inverse function ψ −1 such that ψ −1 {ψ(t)} = t. Clearly, we can
transform the process X ψ (t) by multiplying, at each t, by t/ψ(t), and, defining

0/ψ(0) = 0. The resulting process we can call X(t) and it is readily seen that
this process is standard Brownian motion. Thus, the only crucial assumption in
Brownian motion is that of independent increments. Once we can assert this
to be the case, it is only a question of scale and location to obtain standard
Brownian motion.
Brownian bridge
Let W(t) be Brownian motion. We know that W(0) = 0. We also know that with
probability one the process W(t) will return at some point to the origin. Let’s
choose a point, and in particular the point t = 1 and consider the conditional
process W 0 (t), defined to be Brownian motion conditioned by the fact that
W(1) = 0. For small t this process will look very much like the Brownian motion
from which it is derived. As t goes to one the process is pulled back to the
origin since at t = 1 we have that W 0 (1) = 0 and W(t) is continuous. Also
W 0 (0) = W(0) = 0. Such a process is called tied down Brownian motion or the
Brownian bridge. We will see below that realizations of a Brownian bridge can
be viewed as linearly transformed realizations of Brownian motion itself, and vice
versa.
From the results of above the section we can investigate the properties of
W 0 (t). The process is a Gaussian process so we only need to consider the mean
and covariance function for the process to be completely determined. We have:
E{W(s)|W(1) = 0} = 0 for s < t .
This comes immediately from the above result. Next we have:
Theorem B.2.
Cov (W(s), W(t)|W(1) = 0) = s(1 − t). (B.1)
This provides a simple definition of the Brownian bridge as being a Gaussian

process having mean zero and covariance function s(1 − t) , s < t. An alternative
way of constructing the Brownian bridge is to consider the process defined as:
W 0 (t) = W(t) − tW(1) , 0 ≤ t ≤ 1.
Clearly W 0 (t) is a Gaussian process. We see that
E{W(0)} = W(0) = E{W(1)} = W(1) = E{W(t)} = 0
so that the only remaining question is the covariance function for the process to
be completely and uniquely determined. The following corollary is all we need.
Corollary B.3. The covariance function for the process defined as W 0 (t) is,
Cov {W 0 (s), W 0 (t)} = s(1 − t) s < t.
This is the covariance function for the Brownian bridge developed above
and, by uniqueness, the process is therefore itself the Brownian bridge. Such a
covariance function is characteristic of many observed phenomena. The covari-
ance decreases linearly with distance from s. As for Brownian motion, should the
covariance function decrease monotonically rather than linearly, then a suitable
transformation of the time scale enables us to write the covariance in this form.
At t = s we recover the usual binomial expression s(1 − s).
Notice that not only can we go from Brownian motion to a Brownian bridge
via the simple transformation
W 0 (t) = W(t) − tW(1) , 0 ≤ t ≤ 1 ,
but the converse is also true, i.e., we can recover Brownian motion, X(t), from
the Brownian bridge, Z(t), via the transformation
t
X(t) = (t + 1)Z . (B.2)
t+1
To see this, first note that, assuming Z(t) to be a Brownian bridge, then X(t)
is a Gaussian process. It will be completely determined by its covariance process
Cov {X(s), X(t)}. All we then require is the following lemma:
Lemma B.1. For the process defined in (B.2), Cov {X(s),X(t)} = s.
The three processes: Brownian motion, the Brownian bridge, and the Ornstein-
Uhlenbeck are then closely related and are those used in the majority of applica-
tions. Two further related processes are also of use in our particular applications:
integrated Brownian motion and reflected Brownian motion.
Integrated Brownian motion

t
The process Z(t) defined by: Z(t) = 0 W(u)du, where W(t) is Brownian motion
is called integrated Brownian motion. Note that: dZ(t)/dt = W(t) so that, for
example, in the context of a model of interest, should we be able to construct a
process converging in distribution to a process equivalent to Brownian motion,
then the integrated process will converge in distribution to a process equivalent to
integrated Brownian motion. We can see (by interchanging limits) that Z(t) can
be viewed as the limit of a sum of Gaussian processes and is therefore Gaussian.
Its nature is completely determined by its mean and covariance function. We
have that:
t t
E{Z(t)} = E W(u)du = E{W(u)}du = 0. (B.3)
0 0
For s < t we have:
Lemma B.2. The covariance function for Z(s) and Z(t) is
Cov {Z(s), Z(t)} = s2 (t/2 − s/6) . (B.4)
Lemma B.3. The covariance function for Z(t) and W(t) is:
Cov {Z(t), W(t)} = t2 /2. (B.5)
For a model in which inference derives from cumulative sums, this would
provide a way of examining how reasonable are the underlying assumptions if
repetitions are available. Repetitions can be obtained by bootstrap resampling if
only a single observed process is available. Having standardized, a plot of the log-
covariance function between the process and the integrated process against log-
time ought to be linear with slope of two and intercept of minus log 2 assuming
that model assumptions hold.
Reflected Brownian motion

Suppose we choose some positive value r and then define the process Wr (t) as
a function of Brownian motion, W(t), in the following way: If W(t) < r then
Wr (t) = W(t). If W(t) ≥ r then Wr (t) = 2r − W(t). We have:
Lemma B.4. Wr (t) is a Gaussian process, EWr (t) = 0, Cov{Wr (s), Wr (t)} = s
when s < t.
Thus, Wr (t) is also Brownian motion. Choosing r to be negative and defining
Wr (t) so that, when W(t) > r then Wr (t) = W(t). If W(t) ≤ r then Wr (t) =
2r − W(t). accordingly we have the same result. The process Wr (t) coincides
exactly with W(t) until such a time as a barrier is reached. We can imagine this
barrier as a mirror, and beyond the barrier the process Wr (t) is a simple reflection
of W(t). The interesting thing is that the resulting process is itself Brownian
motion. One way of conceptualizing the idea is to imagine a large number of
realizations of a completed Brownian motion process sampled independently.
Imagine then these same realizations with a reflection applied. Then, whatever
the point of reflection, if we consider the two collected sets of realizations, our
overall impression of the behavior of the two processes will be the same. The
value of this construction is to be seen in situations where, at some point in
time, corresponding to some expected point of reflection under a hypothesis of
drift, the drift changes direction. Under the hypothesis of Brownian motion, both
Brownian motion, and Brownian motion reflected at some point, will look alike
and will obey the same probability laws. Under an alternative hypothesis of drift,
however (see below), the behaviors will look quite different. This observation
enables a simple construction with which to address the problem of crossing
hazards.
Maximum of a Brownian motion on (0,1)

A useful further result can be immediately obtained from the preceding one deal-
ing with reflected Brownian motion. Suppose that W(t) is a Brownian motion.
We might wish to consider the process M (t) = supu∈(0,t) W(u), which is the
greatest value obtained by the process W(u) in the interval (0, t). The greatest
absolute distance is also of interest but, by symmetry arguments, this can be
obtained immediately from the distribution of M (t). Another related question,
useful in interim analyses, is the distribution of W(t) given the maximum M (t)
obtained up until that time point. We have the following:
Lemma B.5. If W(t) is standard Brownian motion on (0,1) and M (t) the max-
imum value attained on the interval (0, t), i.e., M (t) = supu∈(0,t) W(u), then
Pr {M (t) > a} = 2 Pr {W(t) > a}.
This is a simple and elegant result and enables us to make simultaneous

inference very readily. Sometimes, when using a Brownian motion approximation
for a process, we may want to, for example, describe an approximate confidence
interval for the whole process rather than just a confidence interval at a single
point t. In such a case the above result comes into play. The joint distribution is
equally simple and we make use of the following:
Lemma B.6. If W(t) is standard Brownian motion and M (t) the maximum value
attained on the interval (0, t), i.e., M (t) = supu∈(0,t) W(u), then
Pr {W(t) < a − b , M (t) > a} = Pr {W(t) > a + b}.
The conditional distribution Pr {W(t) < a − b |M (t) > a} can then be derived
immediately by using the results of the two lemmas.
Brownian motion with drift

We will see that simple Brownian motion provides a good model for describing
score statistics, or estimating equations, once standardized. This is because we
can visualize these sums as approximating a limiting process arising from summing
increments, for which the expected value is equal to zero. The setting in which
we study such sums is typically that of evaluating some null hypothesis, often
one of some given effect, H0 : β = β0 , but sometimes a less obvious one, in
the goodness-of-fit context, for example, whereby we can have, H0 : β(t) = β̂.
Almost invariably, when we consider a null hypothesis, we have an alternative
in mind, frequently a local or first alternative to the null. For a null hypothesis
of Brownian motion, a natural and immediate alternative is that of Brownian
motion with drift. Consider then the stochastic process X(t) defined by:
X(t) = W(t) + μt
where W (t) is Brownian motion. We can immediately see that E{X(t)} = μt and
Var {X(t)} = t As for Brownian motion Cov {X(s), X(t)} = s , s < t. Alternatively
we can define the process in a way analogous to our definition for Brownian
motion as a process having the following three properties:
1. X(0) = 0.
2. X(t) , t ∈ (0, 1) has independent stationary increments.
3. At each t ∈ (0, 1), X(t) is N (μt, t).
Clearly, if X(t) is Brownian motion with drift parameter μ, then the process
X(t) − μt is standard Brownian motion. Also, for the more common situation in
which the mean may change non-linearly with time, provided the increments are
independent, we can always construct a standard Brownian motion by first sub-
tracting the mean at time t, then transforming the timescale in order to achieve
a linearly increasing variance. Note that for non-linear differentiable functions of
drift, these can always be approximated locally by a linear function so that the
B.3. Counting processes and martingales 385
essential nature of the process, whether the drift is linear or smoothly non-linear,
is the same. We will make use of this idea in those chapters devoted to fit and
model building.
Probability results for Brownian motion

There are a number of well-established and useful results for Brownian motion and
related processes. The arcsine law can be helpful in comparing processes. Defining
X + (t) to be the time elapsed from the origin that the Brownian process remains
√
positive, i.e., sup{t : X(s) > 0 : 0 < s < t} then Pr (X + < x) = (2/π) sin−1 x
This law can be helpful in comparing processes and also in examining underly-
ing hypotheses. For the Brownian bridge the largest distance from the origin in
absolute value has a known distribution given in a theorem of Kolmogorov:
∞

Pr sup |W0 (t)| ≤ α → 1 − 2 (−1)k+1 exp(−2k 2 α2 ) , α ≥ 0. (B.6)
t
k=1
The sum can be seen to be convergent since this is an alternating sign series
in which the kth term goes to zero. Furthermore, the error in ignoring all terms
higher than the nth is less, in absolute value, than the size of the (n + 1)th
term. Given that the variance of W0 (t) depends ont it is also of interest to
study the standardized distribution B0 (t) = W0 (t)/ t(1 − t). This is, in fact,
the Ornstein-Uhlenbeck process. Simple results for the supremum of this are not
possible since the process becomes unbounded at t = 0 and t = 1. Nonetheless,
if we are prepared to reduce the interval from (0, 1) to (ε1 , ε2 ) where ε1 > 0 and
ε2 < 1 then we have an approximation due to Miller and Siegmund (1982):

4φ(α) 1 ε2 (1 − ε1 )
Pr sup |B0 (t)| ≥ α ≈ + φ(α) α − log , (B.7)
t α α ε1 (1 − ε2 )
where φ(x) denotes the standard normal density. This enables us to construct
confidence intervals for a bridged process with limits themselves going to zero
endpoints. To obtain these we use the fact that Pr {W0 (t) > α} =
at the
Pr { t(1 − t)B0 (t) > α}. For most practical purposes though it is good enough
to work with Equation B.6 and approximate the infinite sum by curtailing sum-
mation for values of k greater than 2.
B.3 Counting processes and martingales
While the basic ideas behind counting processes and martingales are both simple
and natural, if we wish to maintain the fullest generality—in practice allowing the
time interval to stretch without limit, as well as allowing the unboundedness for
covariates—then things become difficult and very technical. Rebolledo’s multi-
variate central limit theorem for martingales requires great care in its application.
Our preference is to assume for both time and the covariates, at least for the
purposes of obtaining large sample results, support on finite bounded intervals.
This enables us to work with standard well-known central limit theorems. Specif-
ically we put time on the interval (0,1). In this appendix we motivate the use
of counting processes and martingales in the context of survival analysis prob-
lems and we describe their application in real situations. This also helps establish
the links with other works in the field that do make an appeal to Rebolledo’s
martingale central limit theorem. The goal of this appendix is to provide some
understanding of the probability structure upon which the theory is based.
Martingales and stochastic integrals

Recalling the discussion of Section A.2 and that, for a bounded function H(x)
and the empirical distribution
function Fn (x), we have, by virtue
of the Helly-
Bray theorem, that H(x)dFn (x) converges in distribution to H(x)dF (x). If
we define M (x) = Fn (x) − F (x) and change the order of integration, i.e., move
the expectation operator, E, outside the integral, then

E H(x)dM (x) = 0.
This expression is worth dwelling upon. We think of E as being an integral

operator or as defining some property of a random variable, specifically a measure
of location. The random variable of relevance is not immediately apparent but can
be seen to be Fn (x), an n − dimensional function from the observations to the
interval [0, 1]. We can suppose, at least initially, the functions F (x) and H(x)
to be fixed and known. Our conceptual model allows the possibility of being
able to obtain repetitions of the experiment, each time taking n independent
observations. Thus, for some fixed given x, the value of Fn (x) will generally
vary from one experiment to the next. We view x as an argument to a function,
and Fn (x) as being random having a distribution studied below in Section C.2.
Recalling Section A.2 on integration, note that we can rewrite the above equation
as:

E lim {M (xi ) − M (xi−1 )}H(xi−1 ) = 0, (B.8)
max Δi →0
where Δi = xi − xi−1 > 0 and where, as described in Section A.2 the summation
is understood to be over an increasing partition in which Δi > 0 and max Δi
goes to zero. Now, changing the order of taking limits, the above expression
becomes

lim E{[M (xi ) − M (xi−1 )]H(xi−1 )} = 0, (B.9)
max Δi →0
a result which looks simple enough but that has a lot of force when each of the
infinite number of expectations can be readily evaluated. Let’s view Equation B.9
in a different light, one that highlights the sequential and ordered nature of the
partition. Rather than focus on the collection of M (xi ) and H(xi ), we can focus
our attention on the increments M (xi ) − M (xi−1 ) themselves, the increments
being multiplied by H(xi−1 ), and, rather than work with the overall expectation
implied by the operator E, we will set up a sequence of conditional expectations.
Also, for greater clarity, we will omit the term limmax Δi →0 altogether. We will
put it back when it suits us. This lightens the notation and helps to make certain
ideas more transparent. Later, we will equate the effect of adding back in the term
limmax Δi →0 to that of replacing finite differences by infinitesimal differences.
Consider then

U= {M (xi ) − M (xi−1 )}H(xi−1 ), (B.10)
and unlike the preceding two equations, we are able to greatly relax the require-
ment that H(x) be a known function or that M (x) be restricted to being the
difference between the empirical distribution function and the distribution func-
tion. By sequential conditioning upon F(xi ) where F(xi ) are increasing sequence
of sets denoting observations on M (x) and H(x), for all values of x less than
or equal to xi , we can derive results of wide applicability. In particular, we can
now take M (x) and H(x) to be stochastic processes. Some restrictions are still
needed for M (x), in particular that the incremental means and variances exist.
We will suppose that
E{M (xi ) − M (xi−1 )|F(xi−1 )} = 0 , (B.11)
in words, when given F(xi−1 ), the quantity M (xi−1 ) is fixed and known and the
expected size of the increment is zero. This is not a strong requirement and only
supposes the existence of the mean. If the expected size of the increment is other
than zero, then we can subtract this difference to recover the desired property.
Furthermore, given F(x), the quantity H(x) is fixed. The trick is then to exploit
the device of double expectation whereby for events, A and B, it is always true
that E(A) = EE(A|B). In the context of this expression, B = F(xi−1 ), leading
to

E(U ) = H(xi−1 )E{M (xi ) − M (xi−1 )|F(xi−1 )} = 0, (B.12)
and under the assumption that the increments are uncorrelated we have the
variance is the sum of the variance of each component to the sum. Thus

Var(U ) = E{H 2 (xi−1 )[M (xi ) − M (xi−1 )]2 |F(xi−1 )}. (B.13)
In order to keep the presentation uncluttered we use a single operator E in the

above expressions, but there are some subtleties that ought not to go unremarked.
For instance, in Equation B.13, the inner expectation is taken with respect to rep-
etitions over all possible outcomes in which the set F(xi−1 ) remains unchanged,
whereas the outer expectation is taken with respect to all possible repetitions. In
Equation B.12 the outer expectation, taken with respect to the distribution of all
potential realizations of all the sets F(xi−1 ), is not written and is necessarily zero
since all of the inner expectations are zero. The analogous device to double expec-
tation for the variance is not so simple since Var(Y ) = E Var(Y |Z)+Var E(Y |Z).
Applying this we have
Var {M (xi ) − M (xi−1 )} = E Var{M (xi ) − M (xi−1 )|F(xi−1 )} (B.14)
since Var E{M (xi ) − M (xi−1 )|F(xi−1 )} is equal to zero, this being the case
because each term is itself equal to the constant zero. The first term also requires
a little thought, the outer expectation indicated by E being taken with respect
to the distribution of F(xi−1 ), i.e., all the conditional distributions M (x) and
H(x) where x ≤ xi−1 . The next key point arises through the sequential nesting.
These outer expectations, taken with respect to the distribution of F(xi−1 )
are the same as those taken with respect to the distribution of any F(x) for
which x ≥ xi−1 . This is an immediate consequence of the fact that the lower-
dimensional distribution results from integrating out all the additional terms in
the higher-dimensional distribution. Thus, if xmax is the greatest value of x
for which observations are made then we can consider that all of these outer
expectations are taken with respect to F(xmax ). Each time that we condition
upon F(xi−1 ) we will treat H(xi−1 ) as a fixed constant and so it can be simply
squared and moved outside the inner expectation. It is still governed by the outer
expectation which we will take to be with respect to the distribution of F(xmax ).
Equation B.13 then follows.
Making a normal approximation for U , and from the theory of estimating
equations, given any set of observations, that U depends monotonically on some
parameter β, then it is very straightforward to set up hypothesis tests for β = β0 .
Many situations, including that of proportional hazards regression, lead to esti-
mating equations of the form of U. The above set-up, which is further developed
below in a continuous form, i.e., after having “added in” the term limmax Δi →0 ,
applies very broadly. We need the concept of a process, usually indexed by time
t, the conditional means and variances of the increments, given the accumulated
information up until time t.
We have restricted our attention here to the Riemann-Stieltjes definition of
the integral. The broader Lebesgue definition allows the inclusion of subsets
of t tolerating serious violations of our conditions such as conditional means
and variances not existing. The conditioning sets can be also very much more
involved. Only in a very small number of applications has this extra generality
been exploited. Given that it makes the main ideas much more difficult to all but
those familiar with measure theory, it seems preferable to avoid it altogether. As
already mentioned, in exchange for a reduction in generality by the imposition

of some constraints, e.g., putting the time frame on the interval (0,1), we are
able to appeal to standard central limit theorems in order to obtain large sample
behavior.
Counting processes
The above discussion started off with some consideration of the empirical cumula-
tive distribution function Fn (t) which is discussed in much more detail in Section
C.5. Let’s consider the function N (t) = {nFn (t) : 0 ≤ t ≤ 1}. We can view this as
a stochastic process, indexed by time t so that, given any t we can consider N (t)
to be a random variable taking values from 0 to n. We include here a restriction
that we generally make which is that time has some upper limit, without loss
of generality, we call this 1. This restriction can easily be avoided but it implies
no practical constraint and is often convenient in practical applications. We can
broaden the definition of N (t) beyond that of nFn (t) and we have:
Definition B.1. A counting process N = {N (t) : 0 ≤ t ≤ 1} is a stochastic pro-
cess that can be thought of as counting the occurrences (as time t proceeds) of
certain type of events. We suppose these events occur singly.
Very often N (t) can be expressed as the sum of n individual counting pro-
cesses, Ni (t), each one counting no more than a single event. In this case Ni (t)
is a simple step function, taking the value zero at t = 0 and jumping to the
value one at the time of an event. The realizations of N (t) are integer-valued
step functions with jumps of size +1 only. These functions are right continuous
and N (t) is the (random) number of events in the time interval [0, t]. We asso-
ciate with the stochastic process N (t) an intensity function α(t). The intensity
function serves the purpose of standardizing the increments to have zero mean.
In order to better grasp what is happening here, the reader might look back to
Equation B.11 and the two sentences following that equation. The mean is not
determined in advance but depends upon Ft− where, in a continuous framework,
Ft− is to Ft what F(xi−1 ) is to F(xi ). In technical terms:
Definition B.2. A filtration, Ft , is an increasing right continuous family of sub
sigma-algebras (see A3 for the meaning of sigma-algebra).
This definition may not be very transparent to those unfamiliar with the
requirement of sigma additivity for probability spaces and there is no real need
to expand on it here. The requirement is a theoretical one which imposes a
mathematical restriction on the size, in an infinite sense, of the set of subsets
of Ft . The restriction guarantees that the probability we can associate with any
infinite sum of disjoint sets is simply the sum of the probabilities associated
with those sets composing the sum. For our purposes, the only key idea of
importance is that Ft− is a set containing all the accumulated information (hence
“increasing”) on all processes contained in the past up until but not including
the time point t (hence “right continuous”). We write, α = {α(t) : 0 ≤ t ≤ 1}

where
α(t)dt = Pr {N (t) jumps in [ t, t + dt)|Ft− } = E{dN (t)|Ft− },
the equality being understood in an infinitesimal sense, i.e., the functional part
of the left-hand side, α(t), is the limit of the right-hand side divided by dt > 0
as dt goes to zero. In the chapter on survival analysis we will see that the hazard
function, λ(t), expressible as the ratio of the density, f (t), to the survivorship
function, S(t), i.e., f (t)/S(t), can be expressed in fundamental terms by first
letting Y (t) = I(T ≥ t). Under this interpretation, we can also write
λ(t)dt = Pr {N (t) jumps in [ t, t + dt)|Y (t) = 1} = E{dN (t)|Y (t) = 1}.
It is instructive to compare the above definitions of α(t) and λ(t). The first def-
inition is the more general since, choosing the sets Ft to be defined from the
at-risk function Y (t) when it takes the value one, enables the first definition to
reduce to a definition equivalent to the second. The difference is an important
one in that if we do not provide a value for I(T ≥ t) then this is a (0, 1) ran-
dom variable, and in consequence, α(t) is a (0, λ(t)) random variable. For this
particular case we can express this idea succinctly via the formula
α(t)dt = Y (t)λ(t)dt. (B.15)
Replacing Y (t) by a more general “at risk” indicator variable will allow for great
flexibility, including the ability to obtain a simple expression for the intensity
in the presence of censoring as well as the ability to take on-board multistate
problems where the transitions are not simply from alive to dead but from, say,
state j to state k summarized via αjk (t)dt = Yjk (t)λjk (t)dt in which Yjk (t) is
left continuous and therefore equal to the limit Yjk (t − ) as > 0 goes to zero
through positive values, an indicator variable taking the value one if the subject is
in state j and available to make a transition to state k at time t − as → 0. The
hazards λjk (t) are known in advance, i.e., at t = 0 for all t, whereas the αjk (t)
are randomly viewed from time point s where s < t, with the subtle condition of
left continuity which leads to the notion of “predictability” described below. The
idea of sequential standardization, the repeated subtraction of the mean, that
leans on the evaluation of intensities, can only work when the mean exists. This
requires a further technical property, that of being “adapted.” We say
Definition B.3. A stochastic process X(t) is said to be adapted to the filtration
Ft if X(t) is a random variable with respect to Ft .
Once again the definition is not particularly transparent to nonprobabilists
and the reader need not be over-concerned since it will not be referred to here
apart from in connection with the important concept of a predictable process.
The basic idea is that the relevant quantities upon which we aim to use the
tools of probability modeling should all be contained in Ft . If any probability

statement we wish to construct concerning X(t) cannot be made using the set
Ft but requires the set Ft+u , where u > 0, then X(t) is not adapted to Ft .
In our context just about all of the stochastic processes that are of interest to
us are adapted and so this need not be a concern. A related property, of great
importance, and which also will hold for all of those processes we focus attention
on, is that of predictability. We have
Definition B.4. A real-valued stochastic process, H(t), that is left continuous

and adapted to the filtration Ft is called a predictable process.
Since H(t) is adapted to Ft it is a random variable with respect to Ft . Since

the process is left continuous it is also adapted to Ft− . Therefore, whenever
we condition upon Ft− , H(t) is simply a fixed and known constant. This is
the real sense of the term “predictable” and, in practice, the property is a very
useful one. It is frequently encountered in the probabilistic context upon which a
great number of tests are constructed. Counting processes can be defined in many
different ways and such a formulation allows for a great deal of flexibility. Suppose
for instance that we have events of type 1 and events of type 2, indicated by
N1 (t) and N2 (t) respectively. Then N (t) = N1 (t)+N2 (t) counts the occurrences
of events of either type. For this counting process we have
α(t)dt = P (N (t) jumps in [ t, t + dt)|Ft− ),
i.e., the same as P (N1 (t) or N2 (t) jump in [ t, t + dt)|Ft− ) and, if as is reason-
able in the great majority of applications, where, we assume to be negligible the
probability of seeing events occurring simultaneously compared to seeing them
occur singly, then
α(t)dt = E{dN1 (t) + dN2 (t)|Ft− } = α1 (t) + α2 (t).
This highlights a nice linearity property of intensities, not shared by probabilities

themselves. For example, if we consider a group of n subjects and n individual
counting processes Ni (t), then the intensity function, α(t), for the occurrence

of an event, regardless of individual, is simply αi (t). This result does not
require independence of the processes, only that we can consider as negligible
the intensities we might associate with simultaneous events.
Another counting process of great interest in survival applications concerns
competing risks. Suppose there are two types of event but that they cannot both
be observed. The most common example of this is right censoring where, once
the censoring event has occurred, it is no longer possible to make observations
on Ni (t). This is discussed more fully in the following chapters and we limit
ourselves here to the observation that Ni (t) depends on more than one variable.
In the absence of further assumptions, we are not able to determine the intensity
function, but if we are prepared to assume that the censoring mechanism is
independent of the failure mechanism, i.e., that Pr (Ti > t|Ci > c) = Pr (Ti > t),
then a simple result is available.
Theorem B.3. Let the counting process, Ni (t), depend on two independent and
positive random variables, Ti and Ci such that Ni (t) = I{Ti ≤ t, Ti ≤ Ci }. Let
Xi = min(Ti , Ci ), Yi (t) = I(Xi ≥ t); then Ni (t) has intensity process
αi (t)dt = Yi (t)λi (t)dt. (B.16)
The counting process, Ni (t), is one of great interest to us since the response
variable in most studies will be of such a form, i.e., an observation when the event
of interest occurs but an observation that is only possible when the censoring
variable is greater than the failure variable. Also, when we study a heterogeneous
group, our principal focus in this book, the theorem still holds in a modified form.
Thus, if we can assume that Pr (Ti > t|Ci > c, Z = z) = Pr (Ti > t|Z = z), we
then have:
Theorem B.4. Let the counting processes, Ni (t), depend on two independent
and positive random variables, Ti and Ci , as well as Z such that
Ni (t) = I{Ti ≤ t, Ti ≤ Ci , Z = z}. (B.17)
Then the intensity process for Ni (t) can be written as αi (t, z)dt = Yi (t)λi (t, z)dt.
The assumption needed for Theorem B.4, known as the conditional indepen-
dence assumption, is weaker than that needed for B.3 in that the latter theorem
contains the former as a special case. Note that the stochastic processes Yi (t)
and αi (t) are left continuous and adapted to Ft . They are therefore predictable
stochastic processes, which means that, given Ft− , we treat Yi (t), αi (t) and,
assuming that Z(t) is predictable, αi (t, z) as fixed constants.
B.4 Inference for martingales and stochastic integrals
The reader might look over Section B.3 for the probability background behind
martingales. A martingale M = {M (t) : t ≥ 0} is a stochastic process whose
increment over an interval (u, v], given the past up to and including time u, has
expectation zero, i.e., E{M (v) − M (u)|Fu } = 0 for all 0 ≤ u < v < 1. Equation
B.11 provides the essential idea for the discrete time case. We can rewrite the
above defining property of martingales by taking the time instants u and v to be
B.4. Inference for martingales and stochastic integrals 393
just before and just after the time instant t. Letting both v and u tend to t and
u play the role of t−, we can write;
E{dM (t)|Ft− } = 0. (B.18)
Note that this is no more than a formal way of stating that, whatever the history
Ft may be, given this history, expectations exist. If these expectations are not
themselves equal to zero then we only need to subtract the nonzero means to
achieve this end. A counting process Ni (t) is not of itself a martingale but note,
for 0 ≤ u < v ≤ 1, that E{Ni (v)|Fu } > E{Ni (u)|Fu } and, as above, by taking
the time instants u and v to be just before and just after the time instant t,
letting v and u tend to t and u play the role of t−, we have
E{dNi (t)|Ft− } > 0. (B.19)
A stochastic process Ni (t) with the above property is known as a submartin-

gale. Again, providing expectations are finite, it is only a matter of subtracting
the sequentially calculated means in order to bring a submartingale under the
martingale heading. This idea is made precise by the theorem of Doob-Meyer.
Doob-Meyer decomposition
For the submartingale Ni (t), having associated intensity process α(t), we have
from Equation B.15 that E{dN (t)|Ft− } = α(t)dt. If we write dM (t) = dN (t) −
α(t)dt then E{dM (t)|Ft− } = 0. Thus M (t) is a martingale. For the counting
processes
t of interest to us we will always be able to integrate α(t) and we define
A(t) = 0 α(t). We can write
Ni (t) = Mi (t) + Ai (t). (B.20)
Such a decomposition of a submartingale into the sum of a martingale and a

predictable stochastic process, Ai (t), is an example of a more general theorem
for such decompositions known as the Doob-Meyer theorem. It can be applied to
quite general submartingales, the precise conditions under which require measure-
theoretic arguments. For the counting processes of interest to us in survival
analysis the theorem always applies. The predictable process Ai (t) is called the
compensator of Ni (t). In simple terms the compensator is used to make the
means zero thereby producing the martingales that our theory needs.
The compensator Ai (t)

A counting process, Ni (t), is simply a random variable indexed by t. This is
the definition of a stochastic process so that Ni (t) is, in particular, a stochastic
process. For the majority of applications in survival analysis, Ni (t) will count no
further than one; at the outset, Ni (0) takes the value zero and, subsequently, the
value one for all times greater than or equal to that at which the event of interest
occurs. But, generally, Ni (t) may assume many, or all, integer values. Note that
any sum of counting processes can be immediately seen to be a counting process
in its own right. An illustrative example could be the number of goals scored
during a soccer season by some team. Here, the indexing variable t counts the
minutes from the beginning of the season. The expectation of Ni (t) (which must
exist given the physical constraints of the example) may vary in a complex way
with t, certainly non-decreasing and with long plateau when it is not possible for
a goal to be scored, for instance when no game is being played. At time t = 0,
it might make sense to look forward to any future time t and to consider the
expectation of Ni (t).
As the season unfolds, at each t, depending on how the team performs, we
may exceed, possibly greatly, or fall short of, the initial expectation of Ni (t).
As the team’s performance is progressively revealed to us, the original expecta-
tions are of diminishing interest and it is clearly more useful to consider those
conditional expectations in which we take account of the accumulated history
at time point t. Working this out as we go along, we determine Ai (t) so that
Ni (t) − Ai (t), given all that has happened up to time t, has zero expectation.
When αi (s) is the intensity function for Ni (s), then
t
Ai (t) = αi (s)ds
0
and this important result is presented in Theorem B.5 given immediately below.
Predictable variation process

Linear statistics of the form U described in Section D.1, following standardiza-
tion, are, not surprisingly, well approximated by standard normal variates. We
will see this below using results for estimating equations and results for sums of
independent, although not necessarily identical, random variables. The martin-
gale central limit theorem can also be used in this context, and for all of our
applications, it is possible to apply it in a simple form avoiding measure-theoretic
arguments. Such an approach would then coincide with standard results for sums
of independent random variables. In order to standardize U we will require an
estimate of the variance as well as the mean. Unlike, say, Brownian motion or the
Ornstein-Uhlenbeck processes mentioned in the previous chapter, where at time
t we have a very simple expression for the variance, the variance of U can be
complex and will clearly depend on H(x). One way of addressing this question
is through the use of the predictable variation process. We know from the above
that:
E{dN (t)|Ft− } = α(t)dt , E{dM (t)|Ft− } = 0.
Conditional upon Ft− , we can view the random variable dN (t) as a Bernoulli
(0,1) having mean α(t)dt and variance given by α(t)dt{1 − α(t)dt}. In contrast,
the random variable dM (t), conditional on Ft− , has mean zero and the same
variance. This follows since, given Ft− , α(t) is fixed and known. As usual, all
the equations are in an infinitesimal sense, the equal sign indicating a limiting
value as dt → 0. In this sense α2 (t)(dt)2 is negligible when compared to α(t)dt
since the ratio of the first to the second goes to zero as t goes to zero. Thus,
the incremental variances are simply the same as the means, i.e., α(t)dt. This,
of course, ties in exactly with the theory for Poisson counting processes.
Definition B.5. The predictable variation process of a martingale M (t), denoted
by M = {M (t) : t ≥ 0} is such that
dM (t) = E{[dM (t)]2 |Ft− } = Var{dM (t)|Ft− }. (B.21)
The use of pointed brackets has become standard notation here and, indeed,
the process is often referred to as the pointed brackets process, Note that M
is clearly a stochastic process and that the process is predictable and non-
decreasing. It can be thought of as the sum of conditional variances of the
increments of M over small time intervals partitioning [0, t], each conditional
variance being taken given what has happened up to the beginning of the corre-
sponding interval. We then have the following important result:
t
Theorem B.5. Let Mi (t) = Ni (t) − Ai (t) where Ai (t) = 0 αi (s)ds. Then
Mi (t) = Ai (t). (B.22)
Corollary B.4. Define, for all t and i = j, the predictable covariation process,
Mi , Mj , of two martingales, Mi and Mj , analogously to the above. Then
Mi , Mj (t) = 0. (B.23)
The corollary follows readily if, for i = j, the counting processes Ni (t) and
Nj (t) can never jump simultaneously. In this case the product dNi (t)dNj (t)
is always equal to zero. Thus, the conditional covariance between dNi (t) and
dNj (t) is −αi (t)dt · αj (t)dt.
Stochastic integrals
The concept of a stochastic integral is very simple; essentially we take a
Riemmann-Stieltjes integral, from zero to time point t, of a function which,
at the outset when t = 0 and looking forward, t would be random. t Examples
of most immediate interest to us are: N (t) = 0 dN (s), A(t) = 0 dA(s) and
t t
M (t) = 0 dM (s). Of particular value are integrals of the form 0 H(s)dM (s)
where M (s) is a martingale and H(s) a predictable function. By predictable we
mean that if we know all the values of H(s) for s less than t then we also know
H(t), and this value is the same as the limit of H(s) as s → t for values of s
less than t.
The martingale transform theorem provides a tool for carrying out inference
in the survival context. Many statistics arising in practice will be of a form U
described in Appendix D on estimating equations. For these the following result
will find immediate application:
Theorem B.6. Let M be a martingale and H a predictable stochastic process.
Then M ∗ is also a martingale where it is defined by:
t
∗
M (t) = H(s)dM (s). (B.24)
0
Corollary B.5. The predictable variation process of the stochastic process M ∗ (t)
can be written
t t
M ∗ (t) = H(s)2 dM (s) = H 2 (s)dA(s). (B.25)
0 0
There is a considerable theory for stochastic integrals, much of it developed

in the econometric and financial statistical literature. For our purposes and for
all the tests that have been developed in the framework of counting processes
for proportional hazards models, the above theorem and corollary are all that
are needed. A very considerable array of test procedures come directly under
this heading. Many modern approaches to survival analysis lean on the theory of
stochastic integrals in order to carry out inference. The main approaches here are
sometimes different, based instead on the main theorem of proportional hazards
regression, Donsker’s theorem, and known results concerning Brownian motion
and functions of Brownian motion. Behind both approaches are the ideas relating
to the limits of sums of conditionally independent increments. In some situations
the resulting statistics that we use are identical.
Central limit theorem for martingale processes

The hard part of the work in obtaining a large sample result for stochastic inte-
grals is to find a convergence in probability result for the variation process M (t).
If this limit exists then we write it as A(t). The result can be summarized in a
theorem which depends on two conditions.
Theorem B.7. Suppose we have the following two conditions:
1. As n increases without bound, M (t) converges in probability to A(t),
2. As n increases, the jumps in M (t) tend to zero.
Then, the martingale M (t) converges to a Gaussian process with mean zero and
variance A(t).
Added conditions can make it easier to obtain the first one of these conditions,
and as a result, there are a number of slightly different versions of these two
criteria. The multivariate form has the same structure. In practical situations, we
take M (∞) to be N (0, σ 2 ) where we estimate σ 2 by M (∞).
Censoring and at-risk functions

The counting process structure does well on an intuitive level in dealing with
the concept of censoring and the concept of being at risk. We will only observe
the actual counts but we can imagine that the probability, more precisely the
intensity when referring to infinitely small time periods, can change in complex
ways through time. In particular, there may be time periods when, although a key
event of interest may occur, we are unable to observe it because of some censoring
phenomenon. As an example, allow N (t) to count some event of interest and
define Y (t) to take the value zero when t lies in the semi-closed interval (a, b],
(when we are unable to observe the event of interest) and the value one otherwise.
The counting process N ∗ (t) where,
t
N ∗ (t) = Y (s)dN (s), (B.26)
0
counts observable events. If the censoring does not modify the compensator,
A(t), of N (t), then N (t) and N ∗ (t) have the same compensator. The differ-
ence, M ∗ (t) = N ∗ (t) − A(t) would typically differ from the martingale M (t) but
would nonetheless still be a martingale in its own right. In addition, it is easily
anticipated how we might go about tackling the much more complex situation in
which the censoring would not be independent of the failure mechanism. Here,
the compensators for N ∗ (t) and N (t) do not coincide. For this more complex
case, we would need some model, A∗ (t), for the compensator of N ∗ (t) in order
that M ∗ (t) = N ∗ (t) − A∗ (t) would be a martingale.
The most common and the simplest form of the at-risk indicator Y (t) is one
where it assumes the value one at t = 0, retaining this value until censored or
failed, beyond which time point it assumes the value zero. When dealing with
n individuals, and n counting processes, we can write N̄ (t) = n i=1 Ni (t) and
use the at-risk indicator to denote the risk set. If Yi (t) refers to individual i,

then Ȳ (t) = n i=1 Yi (t) is the risk set at time t. The compensator for Ni (t) is
αi (t) = Yi (t)λi (t), where λi (t) is the hazard for subject i, written simply as λ(t)
in the case of i.i.d. replications. Then, the compensator, Ā(t), for N̄ (t) is:
t t

Ā(t) = { ni=1 Yi (s)}λ(s)ds = Ȳ (s)λ(s)ds.
0 0
The intensity process for N̄ (t) is then given by Ȳ (t)λ(t). The multiplicative
intensity model Aalen (1978) has as its cornerstone the product of the fully
observable quantity Ȳ (t) and the hazard rate, λ(t) which, typically, will involve
unknown model parameters. In testing specific hypotheses we might fix some of
these parameters at particular population values, most often the value zero.
Nonparametric statistics
The multiplicative intensity model just described and first recognized by Aalen
(1978) allows a simple expression, and simple inference, for a large number of
nonparametric statistics that have been used in survival analysis over the past half
century. Martingales are immediate candidates for forming an estimating equation
with which inference can be made on unknown parameters in the model. For
our specific applications, these estimating equations will almost always present
themselves in the form of a martingale. For example, the martingale structure
can be used to underpin several nonparametric tests. Using a martingale as an
estimating equation we can take N̄ (t) as an estimator of its compensator Ā(t).
Dividing by the risk set (assumed to be always greater than zero), we have:
t t
dN̄ (t)
− λ(s)ds,
0 Ȳ (t) 0
a quantity which is a martingale and has, in consequence, zero t expectation.

An estimate of the cumulative risk Λ̂(t) is given by Λ̂(t) = 0 Ȳ (t)−1 dN̄ (t)
which is the estimator of Nelson-Aalen. Simple nonparametric statistics for the
comparison of two groups can be obtained immediately from this. If we consider
some predictable weighting process W (s) (for an explanation of what we mean
by “predictable” see the beginning of this appendix), then define
t
dN̄1 (s) dN̄2 (s)
K(t) = W (s) − (B.27)
0 Y¯1 (s) Y¯2 (s)
where a subscript 1 denotes subjects from group 1 and a 2 from group 2. The
choice of the weighting function W (s) can be made by the user and might be
chosen when some particular alternative is in mind. The properties of differ-
ent weights were investigated by Prentice (1978) and Harrington and Fleming
(1982). We can readily claim that K(∞) converges in probability to N (0, σ 2 ).
We estimate σ 2 by K(∞) where
t 2 t 2
W (s) W (s)
K(t) = dN̄1 (s) + dN̄2 (s). (B.28)
0 Y¯1 (s) 0 Y¯2 (s)
The choice W (s) = Y¯1 (s)Y¯2 (s)/[Y¯1 (s) + Y¯1 (s)] leads to the log-rank test statis-
tic and would maintain good power under a proportional hazards alternative of
a constant group difference as opposed to the null hypothesis of no group differ-
ences. The choice W (s) = Y¯1 (s)Y¯2 (s) corresponds to the weighting suggested
by Gehan (1965) in his generalization of the Wilcoxon statistic. This test may
offer improved power over the log-rank test in situations where the group differ-
ence declines with time. These weights and therefore the properties of the test
are impacted by the censoring, and in order to obtain a test free of the impact
of censoring, Prentice (1978) suggested the weights, W (s) = Ŝ1 (s)Ŝ2 (s). These
weights would also offer the potential for improved power when the regression
effect declines with time. These weights are also considered in the chapter on
test statistics.
Appendix C
Limit theorems
C.1 Empirical processes and central limit theorems
We outline the main theorems providing inference for sums of random variables.
The theorem of de Moivre-Laplace is a well-known special case of the central
limit theorem and helps provide the setting. Our main interest is in sums which
can be considered to be composed of independent increments. The empirical
distribution function Fn (t) is readily seen to be a consistent estimator for F (t)
at all continuity points of F (t). However, we can also view Fn (t) as a constant
number multiplying a sum of independent Bernoulli variates and this enables us
to construct inference for F (t) on the basis of Fn (t). Such inference can then
be extended to the more general context of estimating equations. Inference for
counting processes and stochastic integrals is described since this is commonly
used in this area and, additionally, shares a number of features with an approach
based on empirical processes.
The importance of estimating equations is stressed, in particular equations
based on the method of moments and equations derived from the likelihood.
Resampling techniques can also be of great value for problems in inference. All
of the statistics that arise in practical situations of interest can be seen quite
easily to fall under the headings described here. These statistics will have known
large sample distributions. We can then appeal immediately to known results from
Brownian motion and other functions of Brownian motion. Using this approach
to inference is reassuring since (1) the building blocks are elementary ones, well-
known to those who have followed introductory courses on inference (this is
not the case, for instance, for the martingale central limit theorem) and (2) we
obtain, as special cases, statistics that are currently widely used, the most notable
examples being the partial likelihood score test and weighted log-rank statistics.
However, we will obtain many more statistics, all of which can be seen to sit in a

https://doi.org/10.1007/978-3-030-33439-0
402 Appendix C. Limit theorems
single solid framework and some of which, given a particular situation of interest,
will suggest themselves as being potentially more suitable than others.
C.2 Limit theorems for sums of random variables
The majority of statistics of interest that arise in practical applications are directly
or indirectly (e.g., after taking the logarithm to some base) expressible as sums of
random variables. It is therefore of immense practical value that the distribution
theory for such sums can, in a wide variety of cases, be approximated by normal
distributions. Moreover, we can obtain some idea as to how well the approxima-
tion may be expected to behave. It is also possible to refine the approximation.
In this section we review the main limit theorems applicable to sums of random
variables.
Weak law of large numbers for non-correlated variables

A central theorem provides the conditions under which we obtain the convergence
in probability of the empirical mean for centered variables.
Theorem C.1. Let X1 , X2 , . . . , Xn be random variables having mean zero and
uncorrelated. We assume that there exists a constant C ∈ R+ such that V (X) =
Var(X), V (Xi ) ≤ C, for i = 1, . . . , n. Then we have:
n
1 P
Xi −→ 0.
n n→∞
i=1
Theorem of De Moivre-Laplace

Let Nn = n i=1 Xi be the number of successes in n independent Bernoulli trials
Xi , each trial having probability of success equal to p. Then

{Nn − np}/ np(1 − p) → N (0, 1)
where → means convergence in distribution. This is the oldest result of a central

limit type and is the most well-known special case of the more general result,
just below, for sums of independent and identically distributed random variables.
Central limit theorem for i.i.d. variables

Let Xi , i = 1, 2, . . . be independent
2 random variables having
the same distribution
F (.). We assume that u dF (u) < ∞. Let σ 2 = u2 dF (u) − μ2 where μ =

udF (u). Let x̄ = udFn (u) where Fn (t) = n−1 n i=1 I(Ti ≤ t). Then the
central limit theorem states that
√
σ n(x̄ − μ) → N (0, 1).
C.2. Limit theorems for sums of random variables 403
Less formally we state that x̄ converges to a normal distribution with mean μ and
variance σ 2 /n. This result is extremely useful and also quite general. For example,
applying the mean value theorem, then for g(x̄), where g(x) is a differentiable
function of x, we can see, using the same kind of informal statement, that g(x̄)
converges to a normal distribution with mean g(μ) and variance {g (μ)}2 σ 2 /n.
Central limit theorem for independent variables

For nonidentically distributed random variables the problem is very much more
involved. This is because of the large number of potential situations that need
be considered. The most succinct solution appeared as the condition described
below. Let Xi , i= 1, 2, . . . be independent random variables having distributions
2 2 2
2
i (.). Let σi = u dFi (u) < ∞ and μi = udFi (u). Let Bn = σi and define
F
to be an integral over the real line such that |t − μi | > Bn . Introduce the
following:
−2
Condition C.1. For each > 0, Bn (t − μi )2 → 0 , as n → ∞ If this con-
dition is satisfied then

nBn−1 (x̄ − n−1 μi ) → N (0, 1).
This condition is known as the Lindeberg condition. The statement is an “only

if” statement. Less formally we say that x̄ converges to a normal distribution with

mean μi /n and variance Bn2 /n2 . The condition is simply a way of formulating
or expressing mathematically the need that the sum be composed of independent
“relevant” contributions. If a single term or group of terms dominate the sum
such that the remaining contributions are in some sense negligible, then our
intuition may tell us it would not be reasonable to anticipate the central limit
theorem to generally apply. This could happen in various ways, in particular if
there is “too much” information in the tails of the distributions, i.e., the tails are
too heavy or σi2 diminishes with increasing i at too fast a rate. It follows from
the Lindeberg condition that
Bn−2 σn2 → 0 , Bn → ∞ , as n → ∞.
It can be fairly easily shown that the condition below implies the Lindeberg
condition and provides a more ready way of evaluating whether or not asymptotic
normality obtains.
Condition C.2. The Lindeberg condition holds if, for k > 2,

Bn−k κk → 0 as n → ∞.
Central limit theorem for dependent variables

Let Xi , i =
1, 2, . . . be a sequence of random variables having distributions Fi (.).

Let σi2 = u2 dFi (u) < ∞ and μi = udFi (u). As before, let Bn2 = σi2 . Then,
under certain conditions:

nBn−1 x̄ − n−1 μi → N (0, 1).
As we might guess, the conditions in this case are much more involved and we
need to use array notation in order to express the cross dependencies that are
generated. If we take an extreme case we see immediately why the dependen-
cies have to be carefully considered for, suppose Xi = αi−1 Xi−1 where the αi

are nonzero deterministic coefficients such that n 1 αi → 1, then clearly Xn con-
verges in distribution to X1 which can be any chosen distribution. In rough terms,
there needs to be enough independence between the variables for the result to
hold. Describing what is meant by “enough” is important in certain contexts,
time series analysis being an example, but, since it is not needed in this work, we
do not spend any time on it here. A special case of nonidentical distributions, of
value in survival analysis, is the following.
Central limit theorem for weighted sums of i.i.d. variables

Let Xi , i = 1, 2, . . . be independent random variables
having the same distribu-
tion F (.). Let σ 2 = u2 dF (u) < ∞ and μ = udF (u). Letai , i = 1, . . . , n,
be constants, Sn = n−1/2 n i=1 ai (X i − μ) and σ 2 (n) = σ 2
S
n 2
i=1 ai /n, then
2 1/2
Sn /σS (n) → N (0, 1) where σS (n) = {σS (n)} , whenever the following con-
dition holds;
Condition C.3. The coeficients ai are constants and are such that
max |a |
i → 0 .
n 2
j=1 aj
Many statistics arising in nonparametric theory come under this heading, e.g.,
linear sums of ranks. The condition is a particularly straightforward one to verify
and leads us to conclude large sample normality for the great majority of the com-
monly used rank statistics in nonparametric theory. A related condition, which is
sometimes of more immediate applicability, can be derived as a consequence of
the above large sample result together with an application of Slutsky’s theorem.
Suppose, as before, that Xi , i = 1, 2, . . .are independent random
variables having
the same distribution F (.), that σ 2 = u2 dF (u) < ∞, μ = udF (u) and that
, n, are constants. Again, letting Sn = n−1/2 n
ai , i = 1, . . . i=1 ai (Xi − μ) and
2 2 n 2 /n, then S → N (0, σ 2 α2 ) where:
σS (n) = σ a
i=1 i n
C.3. Functional central limit theorem 405
Condition C.4. The mean of the constant coeficients ai converges and

n
1 2
aj → α2 , 0 < α2 < ∞.
n
j=1
The condition is useful in that it will allow us to both conclude normality for
the linear combination Sn , and at the same time, provide us with a variance for
the linear combination. Weighted log-rank statistics and score statistics under
non-proportional hazards models can be put under this heading. The weights in
that case are not fixed in advance but, since the weights are typically bounded
and converge to given quantities, it is relatively straightforward to put in the
extra steps to obtain large sample normality in those cases too.
C.3 Functional central limit theorem
If we limit our attention to sums of random variables, each of which is indexed

by a value t lying between 0 and 1 (we lose no generality in practice by fixing
an upper limit 1 rather than infinity), we can obtain many useful results. The
important idea here is that of the order among the random variables indexed
by t, since t will be a real number between 0 and 1. As always the sums of
interest will be finite, sums of quantities evaluated at some finite number of time
points on the interval (0,1), and, we will appeal to known results concerning the
limiting continuous distributions, as the interval is “filled out,” as a means to
approximate the exact, but necessarily complicated, finite sample distributions.
For this reason it is helpful to begin reasoning in terms of sums, indexed by a
finite number of points, and consider what such sums look like as the number of
points increases without limit.
Sums of i.i.d. variables on interval (0,1)

Imagine a process starting at the origin and making successive displacements
Xi , i = 1, . . . n, where the Xi are all independent. For every k, where 1 ≤ k ≤ n,

the total distance traveled from the origin can be represented by Uk = i≤k Xi
(random walk). The simplest way of looking at such a process is to consider the
interval (0,1) divided into n equal nonoverlapping intervals each of size 1/n. This
can only be achieved in one way. We make observations Xi , and therefore Ui , at
the points t = i/n , i = 1, . . . , n. The increments Xi are independent. We have
E(Xi ) = 0 and Var(Xi ) = σ 2 < ∞. In consequence we see that E(Uk ) = 0, that
E(Uk2 ) = kσ 2 and, in view of the central limit theorem, that E(Uk ) → 0 , ∀ odd.
We make the process continuous by linearly interpolating between the points
at which Ui , i = 1, . . . , n is defined. Note that there are much more general
developments of the limiting process than we obtain here (Brownian motion)
and that continuity can be demonstrated as a property of the limiting process.
However, it seems easier to construct the process already having continuity as a

property for finite situations. This avoids technical difficulties and, perhaps more
importantly, helps illustrate why and how, in practice, we can construct processes
that will look like Brownian motion. Indeed, not only will these processes look like
Brownian motion, but their probabilistic behavior, of practical interest to us, can
be accurately approximated by the known properties of Brownian motion. Finally,
just as in the standardizations of the preceding sections, we need to standardize
the variance of our process. This we do by considering the sum
√ √
Uk∗ = (σ n)−1 Uk = (σ n)−1 Xi ,
i≤k
from which, letting t = k/n, we readily obtain the mean and the variance of Uk∗
as
√ √
E(Uk∗ ) = (σ n)−1 E(Xi ) = 0 ; Var (Uk∗ ) = (σ n)−2 kσ 2 = t. (C.1)
i≤k
Although this and the following section are particularly simple, the reader should
make sure that he or she has a very solid understanding as to what is taking
place. It underscores all the main ideas behind the methods of inference that are
used. An example of such a process in which σ 2 = 1 and n = 30 is shown in
Figure C.1 As for Var (Uk∗ ) we see in the same way that;
Theorem C.2. For k < m, Cov (Uk∗ , Um

∗ ) = t where t = k/n.
The important thing to note is that the increments are independent, implying
convergence to a Gaussian process. All we then need is the covariance process.
Figure C.1 and Figure C.2 represent approximations to Brownian motion in view
of discreteness and the linear interpolation. The figures indicate two realizations
from the above summed processes, and the reader is encouraged to carry out
his or her own such simulations, an easy exercise, and yet invaluable in terms
of building good intuition. An inspection of any small part of the curve (take,
for example, the curve between 0.30 and 0.31 where the curve is based on less
than 100 points), might easily be a continuous straight line, nothing at all like
the limiting process, Brownian motion. But imagine, as very often is the case
in applications, that we are only concerned about some simple aspect of the
process, for instance, the greatest absolute distance traveled from the origin for
the transformed process, tied down at t = 1. With as few as 30 observations our
intuition would correctly lead us to believe that the distribution of this maximum
will be accurately approximated by the same distribution, evaluated under the
assumption of a Brownian bridge. Of course, such a statement can be made more
technically precise via use, for example, of the law of the iterated logarithm or
the use of Berry-Esseen bounds.
C.3. Functional central limit theorem 407
Figure C.1: Two independent simulations of sums of 30 points on interval (0,1).
Sums of independent variables on (0,1)

The simple set-up of the previous section, for which the large sample theory,
described above, is well established, can be readily extended to the non i.i.d.
case. Let’s begin by relaxing the assumption that the variances of the Xi do
not depend on i. Suppose that as before E(Xi ) = 0 and let Var(Xi ) = σi2 < ∞.
Then clearly the process
√
Uk∗ = ( n)−1 σi−1 Xi
i≤k
will look like the process defined above. In particular, straightforward manipula-
tion as above shows that E(Uk∗ ) = 0 and Cov (Uk∗ , Um ∗ ) = t where k < m and
t = k/n. We allow k to increase at the same rate, i.e., k = nt where 0 < t < 1.
As n → ∞ the number of possible values of t, t ∈ (0, 1) also increases without
limit to the set of all rationals on this interval. We can also suppose that as
k, n → ∞ ; k < n, such that k/n = t then σt2 converges almost everywhere to
some function σ 2 (t). We then allow n to increase without bound.
The functional central limit theorem states that the above process goes to
a limit. The limiting process is defined on the real interval. Choosing any set of
points {t1 , . . . , tk } , (0 < ti < 1 , i = 1, . . . , k) then the process Ut∗1 , Ut∗2 , . . . , Ut∗k
converges in distribution to the multivariate normal. As indicated above the
covariance only depends on the distance between points so that Cov {Us∗ , Ut∗ } =
s ; s < t. The basic idea is that the increments, making up the sum U ∗ (t), get
smaller and smaller as n increases. The increments have expectation equal to
zero unless there is drift. Also, the way in which the increments become smaller
√
with n is precisely of order n. The variance therefore increases linearly with
time out in the process. In practical applications, it is only necessary that the
increments be independent and that these increments have a finite variance. It
Figure C.2: Two independent simulations of sums of 500 points on interval (0,1).
is then straightforward to carry out a time transformation to obtain the limiting

process as an immediate consequence of the functional central limit theorem.
The functional central limit theorem differs very little in essence from the usual
central limit theorem, from which it derives. The key additional idea is that of
sequential standardization. It is all very simple but, as we shall see, very powerful.
C.4 Brownian motion as limit process
Perhaps the most well-known application of the above is to the empirical distribu-
tion function. A great deal can be said about the empirical distribution function
by appealing to the large sample results that can be anticipated from Donsker’s
theorem. Brownian motion can be seen to be the limit process obtained by apply-
ing the functional central limit theorem. We make wide use of these results in
this book. Donsker’s theorem focuses on the case of linear interpolation of a
stochastic process having independent increments with the same distribution
and, specifically, having the same variance.
Theorem C.3. (Donsker, 1951). Let (ξn )n∈N be a sequence of i.i.d. random
variables on the probability space (Ω, F, P ) such that E(ξn ) = 0 et V (ξn ) = σ 2
for all n ∈ N, then
L
Xn −→ W(t),
n→∞
where, for all t ∈ [0, 1],
nt
1 1
Xn (t) = √ ξi + (nt − nt) ξ nt +1 .
σ n σ (n)
i=1
We can relax the assumption of the identical distribution and limit our atten-
tion to zero mean random variables not having the same distribution, and specif-
C.4. Brownian motion as limit process 409
ically having different variances, we can still apply a functional central limit the-
orem described by Helland (1982). The main idea here is to view the sequences
of random variables as martingale differences with respect to a particular family
of σ-algebras. Unlike Donsker we do not use linear interpolation. In practice of
course linear interpolation does not add on any essential restriction and, impor-
tant properties such as continuity, follow very easily.
Theorem C.4. (Helland, 1982). Let ξj,n be a random variable defined on the
probability space (Ω, F, P ), for j = 1, . . . , n. Let Fj,n be a σ-algebra such that
ξj,n is Fj,n -measurable and Fj−1,n ⊂ Fj,n ⊂ F for all j = 2, . . . , n. Let rn be
a function defined on R+ such that rn (t) is a stopping time with respect to
Fj,n , j = 1, . . . , n. We suppose that the paths rn are right continuous and increas-
ing with rn (0) = 0. Note that,
rn (t)

Xn (t) = ξj,n . (C.2)
j=1
t
Let f be a positive measurable function such that 0 f 2 (s)ds < ∞, ∀t > 0. When
the following conditions are verified:
a. ξj,n is a difference of martingales, i.e.,
E(ξj,n | Fj−1,n ) = 0, j = 1, . . . , n,
rn (t) 2 |F P t 2
b. j=1 E(ξj,n j−1,n ) −→ f (s)ds,
n→∞ 0
rn (t) 2 I(|ξ P

c. j=1 E(ξj,n j,n | > ε)|Fj−1,n ) −→ 0, ∀ε > 0,
n→∞
Then,
L
Xn −→ f W,
n→∞
t
where f W(t) = 0 f (s)dW(s) for t ∈ R+ .
When f is constant and equal to one, the process Xn converges in distribution
toward standard Brownian motion. Condition (c) follows from the condition of
Lyapunov, denoted (c’) such that
rn (t)
P
2+δ
∃ δ > 0, E(ξj,n |Fj−1,n ) −→ 0. (C.3)
n→∞
j=1
These somewhat elementary results provide the framework in which we can read-
ily anticipate the large sample behavior of the processes of interest to us. Such
behavior we outline specifically at the relevant place via key theorems.
(l)
Theorem C.5. (Helland, 1982). Let n, m ∈ N∗ . Let ξj,n be a random variable
space (Ω, F, P ), for j = 1, .. . , n and l = 1, . . . , m. We
defined on the probability
(l)
assume that the sets ξj,n , j = 1, . . . , n, n = 1, 2, . . . are tables of martingale
differences with respect to the increasing sequence of σ-algebras (Fj,n )j=1,2,...,n ,
for l = 1, . . . , m. Let rn be a function defined on R+ such that rn (t) is a stopping
time with respect to Fj,n , j = 1, . . . , n. We suppose also that the paths rn are
right continuous and increasing with rn (0) = 0. We have:
rn (t)

(1) (m) (l) (l)
Xn (t) = Xn (t), . . . , Xn (t) , Xn (t) = ξj,n , l = 1, . . . , m. (C.4)
j=1
t
Let f1 , . . . , fm , m be positive measurable functions such that 0 fl2 (s)ds <
∞, ∀t > 0, for l = 1, . . . , m. When the following conditions hold the, for all
i, l = 1, . . . , m :
rn (t)
(l) (i) P
a. j=1 E ξj,n ξj,n Fj−1,n −→ 0, for t > 0 and l = i.
n→∞

rn (t) (l) 2 P t 2
b. j=1 E ξj,n Fj−1,n −→ f (s)ds,
n→∞ 0 l

rn (t) (l) 2 (l) P
c. j=1 E ξj,n I(|ξj,n | > ε) Fj−1,n −→ 0, ∀ε > 0,
n→∞
L
Then, Xn −→ f1 dW1 , . . . , fm dWm with respect to the product
n→∞
t t
topology of Skorokhod where 0 fl dWl = 0 fl (s)dWl (s) for t ∈ R+ , l = 1, . . . , m
and W1 , . . . , Wm are m independent Brownian motions.
C.5 Empirical distribution function
The above results can be directly applied to the sample empirical distribution
function Fn (t), defined for a sample of size n (uncensored) to be the number of

observations less than or equal to t divided by n, i.e., Fn (t) = n−1 n
i=1 I(Ti ≤ t).
For each t, and we may assume F (t) to be a continuous function of t, we would
hope that Fn (t) converges to F (t) in probability. This is easy to see but, in fact,
we have stronger results, starting with the Glivenko-Cantelli theorem whereby:
Dn = sup |Fn (t) − F (t)| = sup |Sn (t) − S(t)|

0≤t≤∞ 0≤t≤∞
converges to zero with probability one and where Sn (t) = 1 − Fn (t). This is
analogous to the law of large numbers, and although important, is not all that
informative. A central limit theorem can tell us much more and this obtains
C.5. Empirical distribution function 411
by noticing how Fn (t) will simulate a process relating to the Brownian bridge.
To see this it suffices to note that nFn (t), for each value of t, is a sum of
independent Bernoulli variables. Therefore, for each t as n → ∞, we have that
√
n{Fn (t) − F (t)} converges to normal with mean zero and variance F (t){1 −
F (t)}. We have marginal normality. However, we can claim conditional normality
in the same way, since, for each t and s (s < t), nFn (t) given nFn (s), is also a
sum of independent Bernoulli variables. Take k1 and k2 to be integers (1 < k1 <
k2 < n) such that k1 /n is the nearest rational smaller than or equal to s (i.e.,
k1 = max j ; j ∈ {1, . . . , n, }, j ≤ ns) and k2 /n is the nearest rational smaller than
or equal to t. Thus, k1 /n converges with probability one to s, k2 /n to t, and
k2 − k1 increases without bound at the same rate as n. We then have:
√
Theorem C.6. n{Fn (t) − F (t)} is a Gaussian process with mean zero and
covariance given by:
√ √
Cov [ n{Fn (s)}, n{Fn (t)}] = F (s){1 − F (t)}. (C.5)
√
It follows immediately that, for T uniform, the process n{Fn (t) − t} (0 ≤
t ≤ 1), converges in distribution to the Brownian bridge. But note that, whatever
the distribution of T , as long as it has a continuous distribution function, mono-
√
tonic increasing transformations on T leave the distribution of n{Fn (t)−F (t)}
unaltered. This means that we can use the Brownian bridge for inference quite
generally. In particular, consider results of the Brownian bridge, such as the distri-
bution of the supremum over the interval (0,1), that do not involve any particular
value of t (and thereby F (t)). These results can be applied without modification
√
to the process n{Fn (t) − F (t)} whether or not F (t) is uniform. Among other
useful results concerning the empirical distribution function we have:
Law of iterated logarithm

This law tells us something about the extreme deviations of the process. The
following theorem (Serfling, 2009) provides the rate with n at which the largest
absolute discrepancy between Fn (t) and F (t) is tending to zero.
Theorem C.7. With probability one,
√
nDn 1
lim = sup{F (t)(1 − F (t))} 2 (C.6)
n→∞ (2 log log n) 12
For the most common case in which F (t) is continuous, we know that
sup F (t)(1 − F (t)) is equal to 0.5. As an illustration, for 50 subjects, we find that
Dn is around 0.12. For 50 i.i.d. observations coming from some known or some
hypothesized distribution, if the hypothesis is correct then we expect to see the
greatest discrepancy between the empirical and the hypothesized distribution to
be close to 0.12. Values far removed from that might then be indicative of either
a rare event or that the assumed distribution is not correct. Other quantities,
indicating how close to F (t) we can anticipate Fn (t) to be, are of interest, one
in particular being;
∞
Cn = n {Fn (t) − F (t)}2 f (t)dt.
0
As for Dn the asymptotic distribution of Cn does not depend upon F (t). For
this case the law of the iterated logarithm is expressed as follows:
Theorem C.8. With probability one,
Cn 1
lim = . (C.7)
n→∞ (2 log log n)
1
2 π2
For Dn , inference can be based on the maximum of a Brownian bridge. In

the case of Cn inference is less straightforward and is based on the following
lemma;

Lemma C.1. Letting η = ∞ 2
j=1 χj (πj)
−2 where the χ2 are independent chi-
j
square variates then
lim P (Cn ≤ c) = P (η ≤ c). (C.8)

n→∞
The results for both Dn and Cn are large sample ones but can nonetheless
provide guidance when dealing with actual finite samples. Under assumed models
it is usually possible to calculate the theoretical distribution of some quantity
which can also be observed. We are then able to contrast the two and test the
plausibility of given hypotheses.
Appendix D
Inferential tools
D.1 Theory of estimating equations
Most researchers, together with a large section of the general public, even
if uncertain as to what the study of statistics entails, will be familiar with

the concept, if not the expression itself, of the type T̄ = n−1 n i=1 Ti . The
statistician
may formulate this in somewhat more abstract terms, stating that;
T̄ = n−1 n i=1 T i is a solution to the linear estimating equation for the param-
eter μ, the population mean of the random variable T , in terms of the n i.i.d.
replicates of T. The estimating equation is simply μ − n−1 n i=1 Ti = 0. This
basic idea is very useful in view of the potential for immediate generalization.
The most useful approach to analyzing data is to postulate plausible mod-
els that may approximate some unknown, most likely very complex, mechanism
generating the observations. These models involve unknown parameters and we
use the observations, in conjunction with an estimating equation, to replace the
unknown parameters by estimates. Deriving “good” estimating equations is a
sizeable topic whose surface we only need to scratch here. We appeal to some
general principles, the most common of which are very briefly recalled below,
and note that, unfortunately, the nice simple form for the estimating equation
for μ just above is more an exception than the rule. Estimating equations are
mostly nonlinear and need to be solved by numerical algorithms. Nonetheless, an
understanding of the linear case is more helpful than it may at first appear since
solutions to the nonlinear case are achieved by local linearization (called also
Newton-Raphson approximation) in the neighborhood of the solution. A funda-
mental result in the theory of estimation is described in the following theorem.
Firstly, we define two important functions, L(θ) and I(θ) of the parameter θ by
L(θ) = f (t1 , t2 , . . . , tn ; θ) ; I(θ) = −∂ 2 log L(θ)/∂θ2 . (D.1)

https://doi.org/10.1007/978-3-030-33439-0
414 Appendix D. Inferential tools
We refer to L(θ) as the observed likelihood, or simply just the likelihood (note
that, for n = 1, the expected log-likelihood is the negative of the entropy, also
called the information). When the observations Ti , i = 1, . . . , n, are indepen-
n
dent and identically
n distributed then we can write L(θ) = i=1 f (ti ; θ) and
log L(θ) = i=1 log f (ti ; θ). We refer to I(θ) as the information in the sam-
ple. Unfortunately the negative of the entropy is also called the information (the
two are of course related, both quantifying precision is some sense). The risks
of confusion are small given that the contexts are usually distinct. The function
I(θ) is random because it depends on the data and reaches a maximum in the
neighborhood of θ0 since this is where the slope of the log-likelihood is changing
the most quickly.
Theorem D.1. For a statistic T we can write the following;
Var(T ) ≥ {∂ E(T )/∂θ}2 /E{I(θ)}.
This inequality is called the Cramer-Rao inequality (Cox and Hinkley, 1979).
When T is an unbiased estimate of θ then ∂ E(T )/∂θ = 1 and Var(T ) ≥
1/E{I(θ)}. The quantity 1/E{I(θ)} is called the Cramer-Rao bound. Taking
the variance as a measure of preciseness then, given unbiasedness, we prefer the
estimator T that has the smallest variance. The Cramer-Rao bound provides the
best that we can do in this sense, and below, we see that the maximum likeli-
hood estimator achieves this for large samples, i.e., the variance of the maximum
likelihood estimator becomes progressively closer to the bound as sample size
increases without limit.
Basic equations
For a scalar parameter θ0 we will take some function U (θ) that depends on the
observations as well as θ. We then use U (θ) to obtain an estimate θ̂ of θ0 via
an estimating equation of the form
U (θ̂) = 0. (D.2)
This is too general to be of use and so we limit the class of possible choices of
U (·). We require the first two moments of U to exist in which case, without loss
of generality we can say
E U (θ0 ) = 0, Var U (θ0 ) = σ 2 . (D.3)
Two widely used methods for constructing U (θ) are described below. It is quite
common that U be expressible as a sum of independent and identically distributed
contributions, Ui , each having a finite second moment. An immediate applica-
tion of the central limit theorem then provides the large sample normality for
U (θ0 ). For independent but nonidentically distributed Ui , it is still usually not
D.1. Theory of estimating equations 415
difficult to verify the Lindeberg condition and apply the central limit theorem
for independent sums. Finally, in order for inference for U to carry over to θ̂,
some further weak restrictions on U will be all we need. These require that U
be monotone and continuous in θ and differentiable in some neighborhood of θ0 .
This is less restrictive than it sounds. In practice it means that we can simply
apply the mean value theorem (A.2) whereby:
Corollary D.1. For any > 0, when θ̂ lies in an interval (θ0 − , θ0 + ) within
which U (θ) is continuously differentiable, then there exists a real number ξ ∈
(θ0 − , θ0 + ) such that
U (θ̂) = U (θ0 ) − (θ̂ − θ0 )I(ξ).
This expression is useful for the following reasons. A likelihood for θ will, with
increasing sample size, look more and more normal. As a consequence, I(ξ) will
look more and more like a constant, depending only on sample size, and not ξ
itself. This is useful since ξ is unknown. We approximate I(ξ) by I(θ̂). We can
then express θ̂ in terms of approximate constants and U (θ̂) whose distribution
we can approximate by a normal distribution.
Finding equations
The guiding principle is always the same, that of replacing unknown parameters
by values that minimize the distance between empirical (observed) quantities
and their theoretical (model-based) equivalents. The large range of potential
choices stem from two central observations: (1) there can be many different
definitions of distance (indeed, the concept of distance is typically made wider
than the usual mathematical one which stipulates that the distance between a
and b must be the same as that between b and a) and (2) there may be a number
of competing empirical and theoretical quantities to consider. To make this more
concrete, consider a particular situation in which the mean is modeled by some
parameter θ such that Eθ (T ) is monotone in θ. Let’s say that the true mean
E(T ) corresponds to the value θ = θ0 . Then the mean squared error, variance
about a hypothesized Eθ (T ), let’s say σ 2 (θ), can be written as
σ 2 (θ) = E{T − Eθ (T )}2 = E{T − Eθ (T )}2 + {Eθ (T ) − Eθ0 (T )}2 .
The value of θ that minimizes this expression is clearly θ0 . An estimating equation

derives from minimizing the empirical equivalent of σ 2 (θ). Minimum chi-squared
estimates have a similar motivation. The idea is to bring, in some sense, via our
choice of parameter value, the hypothesized model as close as possible to the
data. Were we to index the distribution F by this parameter, calling this say
Fθ (t), we could re-express the definition for Dn (θ) given earlier as
Dn (θ) = sup |Fn (t) − Fθ (t)|.

0≤t≤∞
Minimizing Dn (θ) with respect to θ will often provide a good, although not
necessarily very tractable, estimating equation. The same will apply to Cn and
related expressions such as the Anderson-Darling statistic. We will see later that
the so-called partial likelihood estimate for the proportional hazards model can
be viewed as an estimate arising from an empirical process. It can also be seen
as a method of moments estimate and closely relates to the maximum likelihood
estimate. Indeed, these latter two methods of obtaining estimating equations are
those most commonly used, and in particular, the ones given the closest attention
in this work. It is quite common for different techniques, and even contending
approaches from within the same technique, to lead to different estimators. It is
not always easy to argue in favor of one over the others.
Other principles can sometimes provide guidance in practice, the principle
of efficiency holding a strong place in this regard. The idea of efficiency is to
minimize the sample size required to achieve any given precision, or equivalently,
to find estimators having the smallest variance. However, since we are almost
always in situations where our models are only approximately correct, and on
occasion, even quite far off, it is more useful to focus attention on other qualities
of an estimator. How can it be interpreted when the data are generated by a
mechanism much wider than that assumed by the model? How useful is it to
us in our endeavor to build predictive models, even when the model is, at least
to some extent, incorrectly specified. This is the reality of modeling data and
efficiency, as an issue for us to be concerned with, does not take us very far. On
the other hand, estimators that have demonstrably poor efficiency, when model
assumptions are correct, are unlikely to redeem themselves in a broader context
and so it would be a mistake to dismiss efficiency considerations altogether even
though they are rather limited.
Method of moments
This very simple method derives immediately as an application of the Helly-Bray
theorem (Theorem A.3). The idea is to equate population moments
to empirical
ones obtained from the observed sample. Given that μ = xdF (x), the above
example is a special case since, we can write μ̄ = x̄ = xdFn (x). Properties of the
estimate can be deduced from the well-known properties of Fn (x) as an estimate
of F (x) (see Section C.5). For the broad exponential class of distributions, the
method of moments estimator, based on the first moment, coincides with the
maximum likelihood estimator recalled below. In the survival context we will
see that the so-called partial likelihood estimator can be viewed as a method of
moments estimator. The main difficulty with method of moments estimators is
that they are not uniquely defined for any given problem. For example, suppose
we wish to aim to estimate the rate parameter λ from a series of observations,
assumed to have been generated by a Poisson distribution. We can use either the
empirical mean or the empirical variance as an estimate for λ. Typically they will
not be the same. Indeed we can construct an infinite class of potential estimators
as linear combinations of the two.
Maximum likelihood estimation

A minimum chi-square estimator for θ derives by minimizing an expression of
variance. It can also appear equally natural to minimize an estimate of the entropy
as a function of θ, i.e., maximize the observed information as a function of θ.
We write the information as V (θ) where
V (θ) = E log f (T ; θ).
Given the observations, T1 , . . . , Tn , we replace the unknown function V (θ) by

V̄ (θ) = n−1 n i=1 log f (Ti , θ). The maximization is easily accomplished when the
parameter or parameter vector θ can only assume some finite number of discrete
values. It is then sufficient to examine all the cases and select θ such that V̄ (θ) is
maximized. For all the models under consideration here we can assume that V (θ)
and V̄ (θ) are continuous smooth functions of θ. By smooth we mean that the first
two derivatives, at least, exist for all values of θ. This is not at all a restrictive
assumption and models that do not have such differentiability properties can
nearly always be replaced by models that do via useful reparameterization. For
instance, there are cases where a model, defined for all positive real θ, may break
down at θ = 0, the entropy not being differentiable at that point, whereas under
the reparameterization θ = exp(α) for α defined over the whole real line, the
problem disappears.
Two fundamental theorems and three corollaries enable us to appreciate the
great utility of the maximum likelihood approach. All that we need are “suitable
regularity conditions.” We return to these immediately below. Assuming these
conditions (a valid assumption for all the models in this book), we have a number
of important results concerning V (θ) and consistent estimates of V (θ).
Theorem D.2. Viewed as a function of θ, V (θ) satisfies

∂V (θ) ∂ log f (T ; θ)
=E = 0. (D.4)
∂θ θ=θ0 ∂θ θ=θ0
Note the switching of the operations, integration and differentiation, in the

above equations. In many texts describing the likelihood method it is common
to only focus on the second part of the equation. It helps understanding to also
keep the first part of the equation in mind since this will enable us to establish
the solid link between the information measure and likelihood. Having divided by
sample size we should view the log-likelihood as an empirical estimate of V (θ).
The law of large numbers alone would provide us with a convergence result but
we can do better, in terms of fitting in with elementary results for estimating
equations, by assuming some smoothness in V (θ) and a consistent estimator,

V̄ (θ), as functions of θ. More precisely:
Corollary D.2. Let V̄ (θ) be a consistent estimate of V (θ). Then {∂ V̄ (θ)/∂θ}θ=θ0
converges, with probability one, to zero.
The expression “suitable regularity conditions” is not very transparent, and
for those lacking a good grounding in analysis, or simply a bit rusty, it might be
less than illuminating. We dwell on it for a moment since it appears frequently
in the literature. We require continuity of the function V (θ), at least for general
situations, in order to be able to claim that as V (θ) approaches V (θ0 ) then
θ approaches θ0 . This is nearly enough, although not quite. Differentiability is
a stronger requirement since we can see that a function that is differentiable
at some point must also be continuous at that same point. The converse is
not so and can be seen immediately in a simple example, y = |x|, a function
that is continuous everywhere but not differentiable at the origin. We need the
differentiability condition in order to obtain the estimating equation and just
a tiny bit more, the tiny bit more not being easily described but amounting to
authorizing the switching of the processes of integration and differentiation. Such
switching has to take place in order to be able to demonstrate the validity of
the above theorem, and the one just below. All of this is summarized by the
expression “suitable regularity conditions” and the reader need not worry about
them since they will hold in all the practical cases of interest to us. The main
result follows as a further corollary:
Corollary D.3. Suppose that U (α) = {∂ V̄ (θ)/∂θ}θ=α and that, for sample size
n, θ̂n is the solution to the equation U (θ) = 0. Then, if V̄ (θ) is consistent for
V (θ), θ̂n converges with probability one to θ0 .
This is almost obvious, and certainly very intuitive, but it provides a solid
foundation to likelihood theory. The result is a strong and useful one, requiring
only the so-called regularity conditions referred to above and that V̄ (θ) be con-
sistent for V (θ). Independence of the observations or that the observations arise
from random sampling is not appealed to or needed. However, when we do have
independent and identically distributed observations then, not only do our oper-
ations become very straightforward, we also can appeal readily to central limit
theorems, initially for the left-hand side of the estimating equation, and then, by
extension, to a smooth function of the estimating equation. The smooth function
of interest is of, course, θ̂n itself.
Corollary D.4. Suppose that T1 , . . . , Tn are independent identically distributed
random variables having density f (t; θ0 ). Then θ̂n converges
with probability one
to θ0 where θ̂n is such that U (θ̂n ) = 0, where U (θ) = n i=1 Ui (θ) and where
Ui (α) = {∂log f (Ti ; θ)/∂θ}θ=α .
We can say much more about θ̂n . A central limit theorem result applies imme-
diately to U (θ0 ), and via Slutsky’s theorem, we can then also claim large sample
normality for U (θ̂n ). By expressing θ̂n as a smooth (not necessarily explicit) func-
tion of U we can also then claim large sample normality for θ̂n . The fact that
U (θ̂n ) (having subtracted off the mean and divided by its standard deviation)
will converge in distribution to a standard normal and that a smooth function
of this, notably θn , will do the same, does not mean that their behavior can be
considered to be equivalent. The result is a large sample one, i.e., as n tends to
infinity, a concept that is not so easy to grasp, and for finite samples, behavior
will differ. Since U is a linear sum we may expect the large sample approxima-
tion to be more accurate, more quickly, than for θn itself. Inference is often more
accurate if we work directly with U (θ̂n ), exploiting the monotonicity of U (·), and
inverting intervals for U (θ̂n ) into intervals for θ̂n . In either case we need some
expression for the variance and this can be obtained from the second important
theorem:
Theorem D.3. Viewed as a function of θ, V (θ) satisfies:

2
∂V (θ) −∂ 2 V (θ) −∂ 2 log f (T ; θ)
= =E . (D.5)
∂θ θ=θ0 ∂θ2 θ=θ0 ∂θ2 θ=θ0
As in the previous theorem, note the switching of the operations of integration

(expectation) and differentiation. Since E U (θ0 ) = E{∂log f (T ; θ)/∂θ}θ=θ0 =
0, then, from the above theorem, Var U (θ0 ) = E{∂ 2 log f (T ; θ)/∂θ2 }θ=θ0 . In
practical applications we approximate the variance expression by replacing θ0 by
its maximum likelihood estimate. It is also interesting to note that the above
inequality will usually break down when the model is incorrectly specified and
that, in some sense, the further away is the assumed model from that which
actually generates the data, then the greater the discrepancy between the two
quantities will be. This idea can be exploited to construct goodness-of-fit tests
or to construct more robust estimators.
Accurate inference for linear estimating equations

For the case described above where T1 , . . . , Tn are independent identically dis-
tributed random variables having density f (t; θ0 ), we write the estimating equa-
tion in terms of U (T1 , . . . , Tn : θ0 ) having solution θ̂n where U (T1 , . . . , Tn :
θ̂n ) = 0. We already know that θ̂n converges with probability one to θ0 . In
n
the linear case we write: U (θ) = n i=1 Ui (θ) = i=1 U(i) (θ) where the paren-
thesized subscript denotes order statistics (Appendix A.9) and where Ui (α) =
{∂log f (Ti ; θ)/∂θ}θ=α . The large sample theory applies but consider also the
n
estimate θ̂nw , obtained from the weighted estimating equation, i=1 Wi U(i)
(θ̂nw) = 0, where for all i, Wi > 0. In general, the large sample properties of θ̂nw
and θ̂n are the same but can differ for finite samples. Naturally, some condi-
tions are needed, specifically that Cov (Wi , Ui ) = 0 and with some restrictions
on the relative sizes of the Wi . Standardization will solve most problems, usually
n
achieved via the constraint, i=1 Wi = 1. We might wish to make the con-
dition, Cov (Wi , Ui ) = 0, a large sample rather than a finite sample result and
√
that would also work providing the rate of convergence is higher than n. Sup-
pose that X i , i = 1, ..., n, are standard exponential variates and we define Wi by:
Wi = Xi /{ n j=1 Xj }.
Rubin (1981) and, more recently, Xu and Meeker (2020) studied inference
based on such weighted estimating equations. Inference is fully efficient and
particularly easy to carry out. We can view inference as being conditional upon the
data, i.e., the observations T1 ... Tn . Rather than consider theoretical replications
of U1 ... Un , we treat these as being fixed. In particular, we can write down the
empirical distribution for the ordered U(1) ... U(n) . If we then make an appeal to
the probability integral transform (Appendix A.6), we see that the Wi , the gaps
between the cumulative observations, i.e., the empirical distribution function, are
exactly distributed as described above. This is a consequence of the fact that the
gaps between uniform order statistics are distributed as: Wi = Xi /{ n j=1 Xj }
(Appendix A.9). Large sample inference will agree to a very high degree with
inference based on normal approximations, and for finite samples, we can often
obtain more precise and more robust results.
Three further advantages of such an approach to inference for linear esti-
mating equations are worthy of mention. The first is that we can either do any
calculations analytically, given the particularly simple form of the exponential
distribution, or via simulated replications which, again, are very simple to set up.
For small samples where the bootstrap may fail as a result of unboundedness of
the estimator itself, use of the continuously weighted estimating equation will
avoid such difficulties. The bootstrap for survival data is described below and it
is not difficult to see how the non-zero chance of obtaining samples with infinite
parameter estimates can cause problems. The wild bootstrap is an alternative
way around this problem but we do not investigate that here.
Finally, when a parameter is defined in terms of large sample results for an
estimating equation, we are able to interpret what is being estimated when our
model assumptions are not correct. But, if our estimating equations are linear, we
can do more and we can use these weighted linear estimating equations to make
valid inferences even when the data are generated by a model that is different
from that assumed. This is of great value in cases where we assume some narrow
model, e.g., proportional hazards, but the mechanism generating the observations
is a broader one.
D.2 Efficiency in estimation and in tests
Relative efficiency for consistent estimators

Intuitively we would prefer the estimator θ̂ to θ̃ of the parameter θ if, on average,
it was the closer of the two. The word “closer” gives us the clue that this is not
D.2. Efficiency in estimation and in tests 421
going to be straightforward since there is a very large number of possibilities

for distance that we could choose to work with. Suppose we choose to work
with Euclidean distance. Assuming θ to be fixed, a reasonable guide could be
provided by E(θ̂ − θ)2 , the mean squared error. This would depend on both the
bias and the variance of the distribution of θ̂. In practice we will mostly work
with estimators that are consistent or asymptotically unbiased so that our focus
can then be limited to variances of the distributions for θ̂ and θ̃. Recalling the
expressions for the likelihood, L(θ), and information I(θ) of the parameter θ by
L(θ) = f (t1 , t2 , . . . , tn ; θ) ; I(θ) = −∂ 2 log L(θ)/∂θ2 ,
for a statistic T we can write the following:
Var(T ) ≥ {∂ E(T )/∂θ}2 /E{I(θ)}.
which is the Cramer-Rao inequality (Cox and Hinkley, 1979). When T is an unbi-
ased estimate of θ then ∂ E(T )/∂θ = 1 and Var(T ) ≥ 1/E{I(θ)}. The quantity
1/E{I(θ)} is called the Cramer-Rao bound and it provides the best that we can
do. This bound divided by the variance of any alternative unbiased estimator
measures the relative efficiency of the estimator. This would depend on sample
size n and, when this ratio achieves a limit as n increases without bound then
we call this the asymptotic relative efficiency. Since the ratio of the variance of
the maximum likelihood estimator and this bound tends to one with increasing
n we can see that the maximum likelihood estimator is fully efficient.
Relative efficiency for consistent tests

Against local alternatives, indeed against any alternatives, the power of a con-
sistent test tends to 100% as sample size increases without bound. In order to
weigh up the relative advantages of one consistent test over another, in terms of
power, it is common to investigate relative efficiency at some different, plausi-
ble, sample sizes. Typically simulations are used, and assuming some smoothness
with sample size, we are able to obtain a good sense of the advantages or disad-
vantages of different consistent tests. Asymptotic efficiency is, again, a limiting
behavior although somewhat more complex than for estimation in that not only
the parameter stipulated by the test, say θ = θ0 comes into play but also the
size, α and power, β, need be taken into account. For two consistent tests, T1
and T2 , and for given fixed α and β, we require n1 and n2 subjects. The relative
efficiency, e(T1 , T2 ) is defined to be
e(T1 , T2 ) = n1 /n2 .
In this expression, if we know which is the more powerful test, then this would
correspond to n1 so that this ratio lets us know how much extra work is needed,
in terms of sample size, if we prefer to use test T2 . If, for α and β fixed, for
a sequence of local alternatives to the null hypothesis, H0 : θ = θ0 , where θ

√
approaches θ0 at rate n, the ratio n1 /n2 approaches a limit, then we call this
limit, e(T1 , T2 ) the Pitman relative efficiency of the tests. Two other measures
of asymptotic relative efficiency are commonly used, the Hodges-Lehmann and
the Bahadur measures. For the Hodges-Lehmann measure, we fix β and θ and
consider the limit as α → 0. For the Bahadur measure, we fix α and θ and consider
the limit as β → 1. Note further (Van Eeden, 1963) that, under broad conditions,
for Pitman asymptotic relative efficiency, e(T1 , T2 ) = ρ2 (T1 , T2 ) where ρ2 (T1 , T2 )
is the limiting correlation between T1 and T2 .
D.3 Inference using resampling techniques
Bootstrap resampling
The purpose of bootstrap resampling is twofold: (1) to obtain more accurate
inference, in particular more accurate confidence intervals, than is available via
the usual normal approximation, and (2) to facilitate inference for parameter
estimators in complex situations. A broad discussion including several challenging
applications is provided by Politis (1998). Here we will describe the basic ideas in
so far as they are used for most problems arising in survival analysis. Consider the
empirical distribution function Fn (t) as an estimate for the unknown distribution
function F (t). The observations are T1 , T2 , ... , Tn . A parameter of interest,
such as the mean, the median, some percentile, let’s say θ, depends only on F .
This dependence can be made more explicit by writing θ = θ(F ). The core idea
of the bootstrap can be summarized via the simple expression θ̃ = θ(Fn ) as an
estimator for θ(F ).
Taking infinitely many i.i.d. samples, each of size n, from F would provide
us with the exact sampling properties of any estimator θ̃ = θ(Fn ). If, instead
of taking infinitely many samples, we were to take a very large number, say
B, of samples, each sample again of size n from F , then this would provide
us with accurate approximations to the sampling properties of θ̃, the errors of
the approximations diminishing to zero as B becomes infinitely large. Since F
is not known we are unable to carry out such a prescription. However, we do
have available our best possible estimator of F (t), the empirical distribution
function Fn (t). The bootstrap idea is to sample from Fn (t), which is known and
available, instead of from F (t) which, apart from theoretical investigations, is
typically unknown and unavailable.
Empirical bootstrap distribution

The conceptual viewpoint of the bootstrap is to condition on the observed T1 , ...
Tn and its associated empirical cumulative distribution function Fn (t), thereafter
treating these quantities as though they were a population of interest, rather than
a sample. From this “population” we can draw samples with replacement, each
D.3. Inference using resampling techniques 423
sample having size n. We repeat this whole process B times where B is a large
number, typically in the thousands. Each sample is viewed as an i.i.d. sample
from Fn (t). The i th resample of size n can be written T1i ∗ , T ∗ , ..., T ∗ and has
2i ni
∗i
empirical distribution Fn (t). For any parameter of interest θ, the mean, median
coefficient of variation for example, it is helpful to remind ourselves of the several
quantities of interest, θ(F ), θ(Fn ), θ(Fn∗i ) and FB (θ), the significance of each of
these quantities needing a little explanation. First, θ(F ) is simply the population
quantity of interest. Second, θ(Fn ) is this same quantity defined with respect to
the empirical distribution of the data T1 , ... , Tn . Third, θ(Fn∗i ) is again the same
quantity defined with respect to the i th empirical distribution of the resamples
∗ , T ∗ , ..., T ∗ . Finally, F (θ) is the bootstrap distribution of θ(F ∗i ), i.e., the
T1i 2i ni B n
empirical distribution of θ(Fn∗i ) (i = 1, . . . , B).
To keep track of our asymptotic thinking we might note that, as B → ∞ ,
udFB (u) converges in probability to θ(Fn ) and, as n → ∞ , θ(Fn ) converges in
probability to θ(F ). Thus, there is an important conceptual distinction between
FB and the other distribution functions. These latter concern the distribution of
the original observations or resamples of these observations. FB itself deals with
the distribution of θ(Fn∗i ) (i = 1, . . . , n) and therefore, when our focus of interest
changes from one parameter to another, from say θ1 to θ2 the function FB will
be generally quite different. This is not the case for F , Fn , and Fn∗i which are
not affected by the particular parameter we are considering. Empirical quantities
with respect to the bootstrap distribution, FB are evaluated in a way entirely
analogous to those evaluated with respect to Fn . For example,
2
2
Var {θ(Fn )} = σB = u2 dFB (u) − udFB (u) , (D.6)
where it is understood that the variance operator, Var() is with respect to the
distribution FB (t). Of greater interest in practice is the fact that Var {θ(Fn )},
where Var() is with respect to the distribution FB (t)}, can be used as an esti-
mator of Var {θ(Fn )}, where Var(·) is with respect to the distribution F (t).
Bootstrap confidence intervals

For the normal distribution the standardized percentiles zα are defined from
Φ(zα ) = α. Supposing that Var θ(Fn ) = σ 2 , the variance operator being taken
with respect to F , the scaled percentiles Qα are then given by Qα = σzα , leading
to a normal approximation for a confidence interval for θ as
I1−α (θ) = {θ(Fn ) − Q1−α/2 , θ(Fn ) − Qα/2 } (D.7)
which obtains from a rearrangement of the expression Pr [σzα/2 < θ(Fn ) − θ <
σz1−α/2 ] = 1 − α. Since σ 2 is not generally available we would usually work
2 . Instead of using the normal approximation it is possible to define Q
with σB α
differently, directly from the observed bootstrap distribution FB . We can define

Qα via the equation FB (Qα ) = α. In view of the finiteness of B (and also n) this
equation may not have an exact solution and we will, in practice, take the nearest
point from FB−1 (α) as Qα . The values of Qα/2 and Q1−α/2 are then inserted
into equation (D.7). Such intervals are referred to as bootstrap “root” intervals.
The more common approach to constructing bootstrap confidence intervals is,
however, slightly different and has something of a Bayesian or fiducial inference
flavor to it. We simply tick off the percentiles, Qα/2 and Q1−α/2 and view the
distribution FB as our best estimate of a distribution for θ. The intervals are
then written as
I1−α (θ) = {θ(Fn ) + Qα/2 , θ(Fn ) + Q1−α/2 }. (D.8)
These intervals are called percentile bootstrap confidence intervals. Whenever the
distribution FB is symmetric then the root intervals and the percentile intervals
coincide since Qα/2 + Q1−α/2 = 0. In particular, they coincide if we make a
normal approximation.
Accuracy of bootstrap confidence intervals

Using theoretical tools for investigating statistical distributions, the Edgeworth
expansion in particular, it can be shown that the accuracy of the three types of
interval described above is the same. The argument in favor of the bootstrap is
then not compelling, apart from the fact that they can be constructed in cases
where variance estimates may not be readily available. However, it is possible
to improve on the accuracy of both the root and the percentile intervals. One
simple, albeit slightly laborious, way to accomplish this is to consider studentized
methods. By these we mean, in essence, that the variance in the “population” Fn
is not considered fixed and known for each subsample but is estimated several,
precisely B, times across the B bootstrap samples.
To get an intuitive feel for this it helps to recall the simple standard set-up we
work with in the comparison of two estimated means, X̄1 and X̄2 where, when
the variance σ 2 is known we use as test statistic, (X̄1 − X̄2 )/σ, referring the
result to standard normal tables. When σ 2 is unknown and replaced by its usual
empirical estimate s2 then, for moderate to large samples, we do the same thing.
However, for small samples, in order to improve on the accuracy of inference, we
appeal to the known distribution of (X̄1 − X̄2 )/s when the original observations
follow a normal distribution. This distribution was worked out by Student and is
his so-called Student’s t distribution. For the i th bootstrap sample our estimate
of the variance is
2
Var {θ(Fn∗i )} = σ∗i
2
= u2 dFn∗i (u) − udFn∗i (u) (D.9)
D.4. Conditional, marginal, and partial likelihood 425
and we then consider the standardized distribution of the quantity, θ(Fn∗i )/σ∗i .
The essence of the studentized approach, having thus standardized, is to use the
bootstrap sampling force to focus on the higher-order questions, those concern-
ing bias and skewness in particular. Having, in some sense, spared our bootstrap
resources from being dilapidated to an extent via estimation of the mean and
variance, we can make real gains in accuracy when applying the resulting distri-
butional results to the statistic of interest. For most day-to-day situations this is
probably the best approach to take. It is computationally intensive (no longer a
serious objection) but very simple conceptually. An alternative approach, with the
same ultimate end in mind, is not to standardize the statistic but to make adjust-
ments to the derived bootstrap distribution, the adjustments taking into account
any bias and skewness. These lead to the so-called bias corrected, accelerated
intervals (often written as BCa intervals).
D.4 Conditional, marginal, and partial likelihood
A reasonably accurate description of a mathematical function, including multi-

dimensional functions, would be that of a recipe. Throw in the ingredients (argu-
ments) and out comes the result. The result is unique as long as we use the same
ingredients. The likelihood comes under this heading and we refer to it as the
likelihood function. It is a function of a model’s parameters (arguments). The
data involved in the likelihood expression are taken to be fixed. As is so often
the case, we will have conditioned on (taken as fixed) some aspect of the obser-
vations and then considered the likelihood as a function of the parameters. The
precise nature of the functional output will depend upon the probability measure
we associate with the observations. For any given model, and set of observations,
it is not common to give much thought to the type of likelihood on which to
base inference. We tend to talk about the ”likelihood of the data,” as though it
were clear and unambiguous.
While this is often the case we might remind ourselves that there may be
many, even an infinite number of potential likelihoods we are at liberty to choose.
Take, for example, the i.i.d. observations, X1 ... Xn . We may decide to neglect
all contributions with an odd index reducing our sample size by 50%. Now,
none of the large sample properties that underlie our motivation are in any way
compromised. They work no less well with only half of the sample as with the
full sample, at least in as far as asymptotic properties are concerned. Finite
samples are, of course, another story and even though large sample properties are
unaffected, none of us would have any difficulty in choosing whether to work with
all of the observations or only half of them. The point is an important one even so,
and in many cases, it may not be at all clear, given several competing likelihoods,
which one will provide the more precise results. Deciding that some feature of the
observations ought to be treated as though they were fixed and known will lead
to a number of possibilities coming under the heading of conditional likelihood.
Empirical likelihood in which certain parameters are assumed fixed at some value
will also come under this heading. We next describe an interesting approach to
inference in which those aspects of the observations that impact the precision
and not the location of the parameter of interest could be assumed fixed at the
values observed in the data.
Conditional likelihood
A conditional likelihood, that we may like to view as a conditional density, sim-
ply fixes some parameters at certain values. An obvious example is that of a
parameter that is a function of some variance, a quantity which indicates the
precision in an estimate of other parameters but tells us nothing about where
such parameters may be located. It has been often argued that it is appropriate
to condition on such quantities. The variance parameter is then taken as fixed
and known and the remaining likelihood is a conditional likelihood.
Cox (1958) gives an example of two instruments measuring the same quantity
but having very different precision. His argument, which is entirely persuasive,
says that we should use the conditional likelihood and not the full likelihood
that involves the probabilities of having chosen one or the other instrument.
The fact that we may have chosen the less precise instrument at a given rate
of the time is neither here nor there. All that matters are the observations and
the particular instruments from which they arise. Another compelling example
arises in binomial sampling. It is not at all uncommon that we would not know
in advance the exact number n of repetitions of the experiment. However, we
rarely would think of working with a full likelihood in which the distribution of n
is explicitly expressed. Typically, we would simply use the actual value of n that
we observed. This amounts to working with a conditional likelihood.
Fisher (1934) derived an exact expression for the distribution of the maximum
likelihood estimate for a location or scale parameter conditional on observed
spread in the data. Fisher’s expression is particularly simple and corresponds to
the likelihood function itself standardized so that the integral with respect to
the parameter over the whole of the parameter space is equal to one. It is quite
an extraordinary result in its generality and the fact that it is exact. Mostly we
are happy to use large sample approximations based on central limit theorems
for the score statistic and Taylor series approximations applied to a development
around the true value of an unknown parameter. Here we have an exact result
regardless of how small our sample is as long as we are prepared to accept the
argument that it is appropriate to condition on the observed spread, the so-called
“configuration” of the sample. For a model in which the parameter of interest is a
location parameter, the configuration of the sample is simply the set of distances
between the observations and the empirical mean. For a location-scale family
this set consists of these same quantities standardized by the square root of the
variance.
Fisher’s results were extended to the very broad class of models coming under
the heading of curved exponential family models (Efron et al., 1978). This exten-
sion was carried out for more general situations by conditioning on the observed
information in the sample (a quantity analogous to the information contained in
the set of standardized residuals). Although no longer an exact result the result
turns out to be very accurate and has been extensively studied by Barndorff-
Nielsen and Hall (1988). For the proportional hazards model we can use these
results to obtain g(β) as the conditional distribution of the maximum likelihood
estimator via the expression

g(u) = π(u, Xi ) π(u, Xi ) du.
i u i
In Figure D.1 we illustrate the function g(u), i.e., the estimated conditional dis-
tribution of the maximum likelihood estimator for the proportional hazards model
given the two sample data from the Freireich study. It is interesting to compare
the findings with those available via more standard procedures. Rather than base
inference on the mode of this distribution (i.e., the maximum partial likelihood
estimate) and the second derivative of the log-likelihood we can consider
2
β̃ = ug(u)du , v(β) = u2 g(u)du − ug(u)du ,
u u u
and base tests, point, and interval estimation on these. Estimators of this kind
have been studied extensively by Pitman (1948), and generally, have smaller mean
squared error than estimators based on the mode. Note also, in a way analogous
to the calculation of bootstrap percentile intervals, that we can simply consider
the curve g(u) and tick off areas of size α/2 on the upper and lower halves of
the function to obtain a 100(1 − α)% confidence interval for β. Again, analogous
to bootstrap percentile intervals, these intervals can pick up asymmetries and
provide more accurate confidence intervals for small samples than those based
on the symmetric large sample normal approximation. For the Freireich data,
agreement between the approaches is good and that is to be expected since, in
this case, the normal curve is clearly able to give a good approximation to the
likelihood function.
For higher dimensions the approach can be extended almost immediately, at
least conceptually. A potential difficulty arises in the following way: suppose we
are interested in β1 and we have a second parameter, β2 , in the model. Logically
we would just integrate the two-dimensional density with respect to β2 , leaving
us with a single marginal density for β1 . This is straightforward since we have
all that we need to do this, although of course there will usually be no analytical
solution and it will be necessary to appeal to numerical techniques. The difficulty
occurs since we could often parameterize a model differently, say incorporating
β22 instead of β2 alone. The marginal distribution for β1 , after having integrated
Figure D.1: The standardized (i.e., area integrates to one) partial likelihood func-
tion for the proportional hazards model based on the Freireich data.
out β2 , will not generally be the same as before. Nonetheless, for the proportional
hazards model at least, the parameterization is quite natural and it would suffice
to work with the models as they are usually expressed.
Partial likelihood
Again, starting from a full likelihood, or full density, we can focus interest on
some subset, “integrating out” those parameters of indirect interest. Integrating
the full density (likelihood) with respect to those parameters of secondary interest
produces a marginal likelihood. Cox (1975) develops a particular expression for
the full likelihood in which an important component term is called the partial
likelihood. Careful choices lead us back to conditional likelihood and marginal
likelihood and Cox then describes these as special cases. For a given problem, Cox
provides some guidelines for finding partial likelihoods: (1) the omitted factors
should have distributions depending in an essential way on nuisance parameters
and should contain no information about the parameters of interest and (2)
incidental parameters, in particular the nuisance parameters, should not occur in
the partial likelihood.
Lemma D.1. For the proportional hazards model with constant effects:
n
δi
exp(βZi )
L(β) = n (D.10)
i=1 j=1 j (Xi ) exp(βZj )
Y
satisfies the two guidelines provided by Cox.

For this particular case then the term “partial likelihood” applies. We use
the term more generally for likelihoods of this form even though we may not be
able to verify the extent to which the guidelines apply. Unfortunately, at least
when compared to usual likelihood, marginal and conditional likelihood, or profile
likelihood (maximizing out rather than integrating out nuisance parameters),
partial likelihood is a very difficult concept. For general situations, it is not at
all clear how to proceed. Furthermore, however obtained, such partial likelihoods
are unlikely to be unique. Unlike the more commonly used likelihoods, great
mathematical skill is required and even well steeled statistical modelers, able to
obtain likelihood estimates in complex applied settings, will not generally be able
to do the same should they wish to proceed on the basis of partial likelihood.
For counting processes Andersen et al. (1993), it requires some six pages (pages
103–109), in order to formulate an appropriate partial likelihood.
However, none of this impedes our development since the usual reason for
seeking an expression for the likelihood is to be able to take its logarithm, differ-
entiate it with respect to the unknown parameters, and equating this to zero, to
enable the construction of an estimating equation. Here we already have, via the
main theorem (Section 7.5), or stochastic integral considerations, appropriate
estimating equations. Thus, and in light of the above mentioned difficulties, we
do not put emphasis on partial likelihood as a concept or as a general statistical
technique. Nonetheless, for the main models we consider here, the partial like-
lihood, when calculated, can be seen to coincide with other kinds of likelihood,
derived in different ways, and that the estimating equation, arising from use of
the partial likelihood, can be seen to be a reasonable estimating equation in its
own right, however obtained. The above expression was first presented in Cox’s
original paper where it was described as a conditional likelihood.
Cox’s conditional likelihood

At any time point t there is a maximum of n subjects under study. We can index
the jth subject as having hazard rate λj (Xi ) at time point Xi . In other words
we let Yj (t) take the value 1 if an individual is available to make a transition,
zero otherwise. Then, for the j th subject we have an intensity function αj (t) =
Yj (t)λj (t). The n patients can be viewed as a system. At time point t the system
either remains the same (no failure), changes in a total of n possible ways, in
which a single patient fails or the system may change in a more complicated way
in which there can be more than a single failure.
If we are prepared to assume that at any given point there can be no more
than a single failure, then the system can change in a maximum of n ways. This
defines a simple stochastic process. Notice that, conditional upon there being
a transition, then a straightforward application of Bayes formula enables us to
deduce the probability that the transition is of type j. This is simply
α (X ) Yj (Xi )λj (Xi ) exp(βZj )

n j i = n = n .
=1 α (X i ) =1 Y (X i )λ (X i ) =1 Y (Xi ) exp(βZ )
For subjects having either failed or being lost to follow-up before time t we can
still carry out the sum over all n subjects in our evaluation at time t. This is
because of the indicator variable Yj (t) that takes the value zero for all such
subjects, so that their transition probabilities become zero. The same idea can
be expressed via the concept of risk sets, i.e., those subjects alive and available
to make the transition under study. However, whenever possible, it is preferable
to make use of the indicator variables Yj (t), thereby keeping the sums over n.
Multiplying all the above terms over the observed failure times produces L(β).
In his 1972 paper Cox described this as a conditional likelihood and suggested it
be treated as a regular likelihood for the purposes of inference. In their contribu-
tion to the discussion of Cox’s paper, Kalbfleisch and Prentice point out that the
above likelihood does not have the usual probabilistic interpretation. If we take

times to be fixed then exp(βZj )/ Y (Xi ) exp(βZ ) is the probability of the
subject indexed by j failing at Xi and that all other subjects, regardless of order,
occur after Xi . Cox’s deeper study (Cox, 1975) into L(β) led to the recognition
of L(β) as a partial likelihood and not a conditional likelihood in the usual sense.
The flavor of L(β) is nonetheless that of a conditional quantity, even if
the conditioning is done sequentially and not all at once. Cox’s discovery of
L(β), leading to a host of subsequent applications (time-dependent effects, time-
dependent covariates, random effects), represents one of the most important
statistical advances of the twentieth century. Although it took years of subse-
quent research in order to identify the quantity introduced by Cox as the relevant
quantity with which to carry out inference, and although it was argued that Cox’s
likelihood was not a conditional likelihood in the usual sense (where all of the
conditioning is done at once), his likelihood was all the same the right quantity
to work with.
Marginal likelihood of ranks

Kalbfleisch and Prentice (1973) pointed out, that for fixed covariates Z the
partial likelihood coincides with the likelihood, or probability, of the rank vector
occurring as observed, under the model. We then have an important result;
Theorem D.4. Under an independent censoring mechanism the probability of

observing the particular order of the failures is given by L(β).
Alternatively we can express the likelihood as a function of the regression

vector β and the underlying failure rate λ0 (t). Writing this down we have:
n
!δi !1−δi
L(β, λ0 (t)) = λ0 (t)eβZi S0 (Xi )exp(βZi ) S0 (Xi )exp(βZi ) .
i=1
From this we can break the likelihood into two components. We then have:
Theorem D.5. L(λ0 , β) can be written as the product Lλ (λ0 , β)L(β).

This argument is also sometimes given to motivate the idea of partial likeli-
hood, stating that the full likelihood can be decomposed into a product of two
terms, one of which contains the nuisance parameters λ0 (t) inextricably mixed
in with the parameter of interest β and a term that only depends upon β. This
second term is then called the partial likelihood. Once again, however, any such
decomposition is unlikely to be unique and it is not clear how to express in precise
mathematical terms just what we mean by “inextricably mixed in” since this is
more of an intuitive notion suggesting that we do not know how to separate the
parameters. Not knowing how to separate the parameters do not mean that no
procedure exists that might be able to separate them. And, if we were to sharpen
the definition by stating, for example, that within some large class there exists no
transformation or re-parameterization that would separate out the parameters,
then we would be left with the difficulty of verifying this in practice, a task that
would not be feasible.
Appendix E
Simulating data under the non-proportional

hazards model
We look at two methods described by Chauvel (2014) that will allow us to sim-
ulate samples {Ti , Zi , i = 1, . . . , n} under the non-proportional hazards model.
The first allows us to simulate data when we have a piecewise constant β(t),
corresponding to a change-point model, using the distribution of T given Z. The
second is based on the distribution of Z given T , and allows us to generate data
for any β(t) which is non-constant over time. Censoring is independent of this;
we thus simulate the times {C1 , . . . , Cn } independently of the data.
E.1 Method 1—Change-point models
Consider the non-proportional hazards model with a piecewise constant regression

coefficient given by: for any t ∈ [0, T ]:
k

β(t) = βi 1ti ≤ t < ti+1 , t1 = 0, tk+1 = +∞,
i=1
where the values β1 , . . . , βk are constants. To begin with, we suppose that the
k − 1 break-points t2 , t3 , . . . , tk are known and we show how to simulate data
under the model with parameter β(t). After that, we show how to define the
change-points corresponding to the places where the value of β(t) changes.
Simulating the data

To simplify the presentation, we will work with a constant (over time) one-
dimensional covariate Z ∈ R and suppose that the base risk is constant: λ0 (t) = 1
for all t ∈ [0, T ]. Generalizing this to several (potentially time-dependent) covari-
ates is straightforward. Data is thus simulated under the model:

https://doi.org/10.1007/978-3-030-33439-0
434 Appendix E. Simulating data under the non-proportional . . .
" k
#

λ(t | Z) = λ0 (t) exp {β(t)Z} = exp Z βi 1ti ≤ t < ti+1 , t ∈ [0, T ].
i=1
(E.1)
Calculating S(. | Z)
Let t ∈ [0, T ] and calculate the value of the conditional survival S(t | Z). If t is
between t1 = 0 and t2 , the survival is
t t
S (t | Z) = exp − λ (u | Z) du = exp − exp (β (u) Z) du
0 0
t
= exp − exp (β1 Z) du = exp (−t exp (β1 Z)) .
t1 =0
Then, if t is between tj and tj+1 , j ∈ {2, ..., k}, the survival of T given Z
calculated at t is
t t
S (t | Z) = exp − λ (u | Z) du = exp − exp (β (u) Z) du
0 0
" #
ti+1
j−1 t
= exp − exp (βi Z) du − exp (βj Z) du
i=1 ti tj
" j−1
#

= exp − (ti+1 − ti ) exp (βi Z) − (t − tj ) exp (βj Z) .
i=1
For simulating a random variable whose survival function, given the covariates Z,
is S(. | Z), we use the fact that, if U ∼ U[0, 1], then F −1 (U ) has the cumulative
distribution function F . All that is needed then is to inverse the cumulative
distribution function F (. | Z) = 1 − S(. | Z).
The inverse of F (. | Z)
Let γ ∈ [0, 1] be such that
F (tj | Z) ≤ γ < F (tj+1 | Z), j ∈ {1, ..., k}.
We are looking for the t such that γ = F (t | Z). If 0 ≤ γ < F (t2 | Z):
γ = F (t | Z) = 1 − exp (−t exp (β1 Z)) ,
which is equivalent to t = − log (1 − γ) exp (−β1 Z) . If γ is such that F (tj | Z) ≤

γ < F (tj+1 | Z), then for j ∈ {2, ..., k} we have
E.1. Method 1—Change-point models 435
" j−1
#

γ = F (t | Z) = 1 − exp − (ti+1 − ti ) exp (βi Z) − (t − tj ) exp (βj Z) ,
i=1
which is the same as

j−1

− log (1 − γ) = (ti+1 − ti ) exp (βi Z) + (t − tj ) exp (βj Z)
i=1
j
= ti {exp (βi−1 Z) − exp (βi Z)} + t exp (βj Z) .
i=2
In conclusion:
" j
#

t = exp (−βj Z) − log (1 − γ) − ti {exp (βi−1 Z) − exp (βi Z)} .
i=2
Thus, overall, for any γ ∈ [0, 1], we have:
" " ##

k
j
F −1 (γ | Z) = e−βj Z − log(1 − γ)− ti (eβi−1 Z − eβi Z ) 1F (tj |Z)≤γ<F (tj+1 |Z)

j=2 i=2
− log(1 − γ)e−β1 Z 1γ≤F (t2 |Z) .
Starting with a variable Z, a coefficients vector (β1 , ..., βk ), and a vector with
the break-points (t2 , ..., tk ), we simulate U from the uniform distribution U[0, 1];
then, F −1 (U | Z) is a random variable with cumulative distribution function
F (. | Z). This variable is the survival time T and is coherent with the non-
proportional hazards model.
The break-point times need to be selected in such a way that we do not run
out of individuals. In effect, the times need to be calibrated so that individuals
are present and of sufficient number in each interval [tj , tj+1 [. The choice of
times is made as a function of the coefficients β1 , . . . , βk that have been selected
for the study.
Choosing change-point locations

Break-point times t2 , ..., tk−1 are chosen numerically in such a way that there is
a comparable number of individuals in each time interval. Consider the following
example: we wish to simulate survival data with model (E.1) and three successive
values of β:
3

β(t) = βi 1ti ≤ t < ti+1 , t1 = 0, t4 = +∞,
i=1
436 Appendix E. Simulating data under the non-proportional . . .
and β1 = 2, β2 = 1 and β3 = 0. Before being able to simulate the data using the
method presented in Section E, we need to choose the values t2 and t3 .
Step 1: We simulate survival data for 2000 individuals according to model (E.1)
with β(t) = β1 = 2 for all t. We then calculate the corresponding Kaplan-Meier
estimator. Next, we select a time such that 1/3 of individuals have died before
that time. This value corresponds to t2 . In our example, we get t2 = 0.1.
Step 2: We simulate a new dataset with 2000 individuals according to model

(E.1) with β(t) = β1 1t≤t2 + β2 1t≥t2 = 21t≤t2 + 1t≥t2 . We then calculate the
corresponding Kaplan-Meier estimator. The time t3 corresponds to when 2/3
of the individuals have died. In our example, t3 = 0.4. We simulated a dataset
with the break-points determined as above. Figure E.1 shows the Kaplan-Meier
estimator conditional Ŝ(· | Z) as a function of time, as well as
on the covariates
its transform. log − log Ŝ(· | Z) . After applying this transform, each line on
the plot represents the log of the instantaneous hazard for one group. We see that
there is a different distance between the lines in each of the three time periods
(between 0 and t2 = 0.1, t2 and t3 = 0.4, and after t3 ). We choose these time
intervals so that the same number of deaths occurs in each. It is possible instead
to have unequal numbers, as long as we keep in mind that we need individuals
in each time period; otherwise, the corresponding effect will not be represented
in the dataset.
Figure E.1: The Kaplan-Meier

estimator
conditional on the covariate Ŝ(· | Z) and
its transform log − log Ŝ(· | Z) as a function of time.
E2 Method 2—Non-proportional hazards models 437
E.2 Method 2—Non-proportional hazards models
In order to generate data under the more general non-proportional hazards mod-
els, we use the conditional distribution of Z given T . We work with a single
covariate Z that does not change with time; extensions to multiple covariates
are straightforward with the help of the prognostic index.
We simulate the vectors for times of death (T1 , . . . , Tn ), censoring (C1 , . . . , Cn ),
and covariate vectors (Z1 , . . . , Zn ) independently, via their marginal distributions.
With the help of the probabilities {πi (β(t), t) , i = 1, . . . , n} , we connect up the
covariates and the ranked times of death in the following way:
1. We order the times of death: T(1) ≤ T(2) ≤ · · · ≤ T(n) .
2. We set R = {1, . . . , n}, which corresponds to the current set of at-risk

individuals.
3. For i from 1 to n:

– We calculate πj β(T(i) ), T(i) for each individual j ∈ R, setting
Yl (t) = 1 if l ∈ R and Yl (t) = 0 otherwise.
– We randomly select an individual
j ∗ in R, where individual j in R is
chosen with probability πj β(T(i) ), T(i) .
– We link Zj ∗ with T(i) .
– We remove j ∗ from R.
4. We set Xi = min (Ti , Ci ) and δi = 1Ti ≤Ci for i = 1, . . . , n.
Hence, in our dataset {Xi , δi , Zi ; i = 1, . . . , n}, times of death are simulated

under the non-proportional hazards model with parameter β(t), and censoring
is independent of times of death. Abrahamowicz et al. (1996) have proposed a
similar method in which the sampling probabilities for Z are not the same when
there is censoring as when there is not.
Further exercises and proofs
The appendices are referred to throughout the text and provide the background
to many important results. They are of interest in their own right and, for this
reason, we include here proofs of some key results as well as classwork and
exercises. These are for the benefit of instructors and students who wish to dig a
little more deeply into this background. A course in survival analysis might well
include certain aspects covered in these appendices and this would depend on
the type of course being given as well as the flavor that the instructor wishes for
it to assume.
Random variables, distributions, order statistics

1. Use a simple sketch to informally demonstrate the mean value theorem.
2. Newton-Raphson iteration provides sequentially updated estimates to the

solution to the equation f (x0 ) = 0. At the nth step, we write xn+1 = xn −
f (xn )/f (xn ) and claim that xn converges (in the analytical sense) to x0 . Use
the mean value theorem, and again, a simple sketch to show this. Intuitively,
which conditions will lead to convergence and which ones can lead to failure of
the algorithm.
3. Let g(x) take the value 0 for −∞ < x ≤ 0 : 1/2 for 0 < x ≤ 1 ; 1 for 1 < x ≤ 2 :
and 0 otherwise. Let f (x) = x2 + 2. Evaluate the Riemann-Stieltjes integral of
f (x) with respect to g(x) over the real line.

4. Note that n i=1 i = n(n + 1)/2. Describe a function such that a Riemann-
Stieltjes integral of it is equal to n(n + 1)/2. Viewing integration as an area
under a curve, conclude that this integral converges to n2 as n becomes large.

5. Suppose that in the Helly-Bray theorem for h(x)dFn (x), the function h(x)
is unbounded. Break the integral into components over the real line. For regions
where h(x) is bounded the theorem holds. For the other regions obtain conditions
that would lead to the result holding generally.

https://doi.org/10.1007/978-3-030-33439-0
440 Further exercises and proofs
6. Prove the probability integral transformation by finding the moment-generating

function of the random variable Y = F (X) where X has the continuous cumu-
lative distribution function F (x) and a moment-generating function that exists.
7. If X is a continuous random variable with probability density function f (x) =

2(1 − x) , 0 < x < 1, find that transformation Y = ψ(X) such that the random
variable Y has the uniform distribution over (0,2).
8. The order statistics for a random sample of size n from a discrete distribution
are defined as in the continuous case except that now we have X(1) ≤ X(2) ≤
· · · ≤ X(n) . Suppose a random sample of size 5 is taken with replacement from
the discrete distribution f (x) = 1/6 for x = 1, 2, . . . , 6. Find the probability mass
function of X(1) , the smallest order statistic.
9. Ten points are chosen randomly and independently on the interval (0,1). Find
(a) the probability that the point nearest 1 exceeds 0.8, (b) the number c such
that the probability is 0.4 that the point nearest zero will exceed c.
10. Find the expected value of the largest order statistic in a random sample of
size 3 from (a) the exponential distribution f (x) = exp(−x) for x > 0, (b) the
standard normal distribution.
11. Find the probability that the range of a random sample of size n from the
population f (x) = 2e−2x for x ≥ 0 does not exceed the value 4.
12. Approximate the mean and variance of (a) the median of a sample of size
13 from a normal distribution with mean 2 and variance 9, (b) the fifth-order
statistic of a random sample of size 15 from the standard exponential distribution.
13. Simulate 100 observations from a uniform distribution. Do the same for
an exponential, Weibull, and log-logistic distribution with different parameters.
Next, generate normal and log-normal variates by summing a small number of
uniform variates. Obtain histograms. Do the same for 5000 observations.
14. Obtain the histogram of 100 Weibull observations. Obtain the histogram of
the logarithms of these observations. Compare this with the histogram obtained
by the empirical transformation to normality.
15. Suppose that T1 , . . . , Tn are n exponential variates with parameter λ. Show

that, under repeated sampling, the smallest of these also has an exponential
distribution. Is the same true for the largest observation? Suppose we are only
give the value of the smallest of n observations from an exponential distribution
with parameter λ. How can this observation be used to estimate λ.
Further exercises and proofs 441
16. Suppose that Xi i = 1, . . . , n are independent exponential variates with param-

eter λ. Determine, via simple calculation, the variance of min(X1 , . . . , Xn ).
17. Having some knowledge of the survival distribution governing observations

we are planning to study, how might we determine an interval of time to obtain
with high probability a given number of failures? How should we proceed in the
presence of censoring?
18. Derive the Bienaymé-Chebyshev inequality. Describe the advantages and

drawbacks of using this inequality to construct confidence intervals in a gen-
eral setting.
19. Recall that the information contained in the density g(t) with respect to
h(t) is defined by V (g, h) = E log g(T ) = log g(t)h(t)dt and the entropy is just
−V (g, h). Suppose that the entropy depends on a parameter θ and is written
−Vθ (f, f ). Consider −Vα (f, f ) as a function of α. Show that this function is
maximized when α = θ.
20. For random variables, T and Z with finite second moments, use the device
of double expectation to show that: Var (T ) = Var E(T |Z) + E Var (T |Z). Thus,
the total variance, Var (T ), breaks down into two parts. Why is this breakdown
interpreted as one component corresponding to “signal” and one component
corresponding to “noise?”
21. Suppose that θn converges in probability to θ and that the variance of θn is

given by ψ(θ)/n. Using Equation A.17, find a transformation of θn for which, at
least approximately, the variance does not depend on θ.
22. Consider a stochastic process X(t) on the interval (2, 7) with the following
properties: (a) X(0) = 2, (b) X(t) , t ∈ (2, 7) has increments such that (c), for
each t ∈ (2, 7) the distribution of X(t) is Weibull with mean 2+λtγ . Can these
increments be independent and stationary? Can the process be described using
the known results of Brownian motion?
23. For Brownian motion, explain why the conditional distribution of X(s) given
X(t) (t > s) is normal with E{X(s)|X(t) = w} = ws/t and Var {X(s)|X(t) =
w} = s(t − s)/t. Deduce the mean and the covariance process for the Brownian
bridge.
24. The Ornstein-Uhlenbeck process can be thought of as transformed Brownian

motion in which the variance has been standardized. Explain why this is the case.
25. Reread the subsection headed “Time-transformed Brownian motion” (Sec-

tion B.2) and conclude that the only essential characteristic underwriting the
construction of Brownian motion is that of independent increments.
26. Find the value of t ∈ (0, 1) for which the variance of a Brownian bridge is
maximized.
27. Suppose that under H0 , X(t) is Brownian motion. Under H1 , X(t) is Brow-
nian motion with drift, having drift parameter 2 as long as X(t) < 1 and drift
parameter minus 2 otherwise. Describe likely paths for reflected Brownian motion
under both H0 and H1 . As a class exercise simulate ten paths under both
hypotheses. Comment on the resulting figures.
Approximations and large sample theory

1. Suppose that ∞ i=1 ai = 1. Also
suppose that there exist random variable Xi
such that Xi = ai X1 . Show that Xi , standardized by mean and variance,
would not generally tend to a normal distribution. Comment.
2. Describe in simple terms what is quantified by the Lindeburg condition.
3. For a weighted sum of independent identically distributed random variables,

suppose that each variable is standard uniform and the weights ai are defined as
ai = 10i . Will a central limit result hold, and if not, why not.
4. As a class project, construct graphs based on summing (1) 20 uniform random

variates, (2) 100 uniform random variates, and (3) 10000 uniform random vari-
ates. Standardize the x axis to lie between 0 and 1 and the y axis such that the
mean is equal to zero and the variance increases linearly with x. Replicate each
graph ten times. What conclusions can be drawn from the figures, in particular
the influence of the number of variates summed?
5. Repeat the above class exercise, replacing the uniform distribution by (1) the
log-logistic distribution with different means and variances, (2) the exponential
distribution, and (3) the normal distribution. Again replicate each graph ten
times. Comment on your findings.
6. Consider two hypotheses for the sequence of observations: X1 , . . . , Xn in which

μi = E(Xi ); H1 : μi = 0 , ∀i, against H1 : μi = b2 i for some b = 0. Construct
different tests based on Brownian motion that would enable us to test H0 versus
H1 . Discuss the relative merits of the different tests.
7. Simulate 20 values from a standard exponential distribution and evaluate the

greatest absolute distance between Fn (t) and F (t). Repeat this 1000 times, store
the 1000 values of D20 , and then plot a histogram of these values. Add to the
histogram the distribution of D20 using the Brownian bridge approximation.
8. For a sample size of 220, provide an approximate calculation of how large you
would anticipate the greatest discrepancy between Fn (t) and F (t) to be.
9. Let T1 , . . . , Tn be i.i.d. observations from an exponential distribution with mean

θ. Obtain an estimating equation for θ in terms of E(Ti ). Obtain an estimating
equation for θ in terms of Var (Ti ). Which of the two equations, if any, would
be the most preferable. Give reasons.
10. Consider the following bivariate situation. We have a random variable, T ,

which is continuous and a binary variable Z, which take the values 0 or 1. Given
that T = 0, Pr (Z = 0) = 0.5. Given that Z = 0, the distribution of T is exponen-
tial with mean equal to 1. Given that Z = 1 the distribution of T is exponential
with mean equal to exp(β). Given n pairs of observations (Ti , Zi ) our purpose is
to estimate the parameter β. Considering the values of Z as being fixed, obtain
an estimating equation based on the observations Ti (i = 1, . . . , n). Secondly, we
could view the random aspect of the experiment differently and now take the
observations T1 to Tn as being fixed. Derive the conditional distribution of Z
given T = Ti . Use this to obtain a different estimating equation. Discuss the
relative merits and disadvantages of the two sets of estimating equations.
11. In the next chapter we discuss the idea of right censoring where, for certain
of the observations Ti , the exact value is not known. All that we can say for sure
is that it is greater than some censoring time. How might the discussion of the
previous exercise on the two types of estimating equations arising from reversing
the conditioning variable have a bearing on this.
12. Consider the pair of independent random variables (Yi , Xi ), i = 1, . . . , n. The

null hypothesis is that Yi = φ(Xi ) + i where i is an error term independent of
the pair (Yi , Xi ) and where φ(u) is a nondecreasing function of u. Describe how
you would carry out tests based on Dn and Cn of Section C.5.
13. By investigating different classes of functions φ, describe the relative advan-

tages and disadvantages of tests based upon Cn rather than Dn .
Outline of proofs
Proof of Theorem A.2, Theorem A.7 and Corollary A.9. The importance of the
first of these two theorems is difficult to overstate. All the useful large sample
results, for instance, hinge ultimately on the theorem. An elegant proof of the
theorem, together with well thought out illustrations and some examples, is given
in Shenk (1979). For Theorem A.7 let T have a continuous and invertible distri-
bution function F (t) and let U = F (T ). The inverse function is denoted F −1 so
that F −1 {F (t)} = t. Then
P (U < u) = P {F (T ) < u} = P {T < F −1 (u)}

= P {T < F −1 F (t)} = P {T < t} = F (t) = u.
Thus U has the distribution of a standard uniform. The proof of the corollary
leans upon some straightforward manipulation of elementary events. An outline
is provided in David (1994).
Proof of Theorem A.9 and Corollary A.4. The theorem states that, letting F (x) =
P (X ≤ x) and Fr (x) = P (X(r) ≤ x) then;
n
n
Fr (x) = F i (x)[1 − F (x)]n−i .
i
i=r
Recall that the Xi from the parent population with distribution F (x) are i.i.d.
The event that X(r) ≤ x is the event that at least r of the Xi are less than or
equal to x. This is then the sum of the binomial probabilities summed over all
values of i greater than or equal to r. The first part of the corollary is clear upon
inspection. For the second part note that
n
n n
F i (x)[1 − F (x)]n−i = 1 , F 0 (x)[1 − F (x)]n = [1 − F (x)]n .
i 0
i=0
Proof of Proposition A.3. Let n ∈ N∗ and the functions f = (f 1 , . . . , f p ) ∈

(C[0, 1], Rp ) and gn = (gn1 , . . . , gnp ) ∈ (D[0, 1], Rp ). We have the relation,
p
1
sup f (t) − gn (t) = max d(f i , gni ) ≥ d(f i , gni )
0≤t≤1 i=1,...,p p
i=1
p
1 1
≥ δ(f i , gni ) = δp (f, gn ),
p p
i=1
where the final inequality is obtained by applying Equation (A.4). The conclusion
follows.
Proof of Theorem B.1 and Corollaries B.1 and B.2. We have that:
Pr {X(t + s) > x|X(s) = xs , X(u), 0 ≤ u < s}

= Pr {X(t + s) − X(s) > x − xs |X(s) = xs , X(u), 0 ≤ u < s}
= Pr {X(t + s) − X(s) > x − xs } = Pr {X(t + s) > x|X(s) = xs }.
Corollary B.1 follows since:
fs|t (x|w) = fs (x)ft−s (w − x)/ft (w) = const exp {−x2 /2s − (w − x)2 /2(t − s)}
= const exp {−t(x − ws/t)2 /2s(t − s)}.
For Corollary B.2 note that the process V (t) is a Gaussian process in which:
E{V (t)} = 0 , Cov {V (t), V (s)} = E{exp (−αt/2) exp (−αs/2)X(eαt )X(eα s)}
This can be written: exp (−α(t + s)/2) × E{X(eαt )X(eα s)} which in turn is
then equal to exp(−α(t + s)/2) exp αt and this is just exp (−α(t − s)/2).
Proof of Theorem B.2 and Corollary B.3. Note that:
Cov (X(s), X(t)|X(1) = 0) = E(X(s), X(t)|X(1) = 0)

= E{E(X(s), X(t)|X(t), X(1) = 0|X(1) = 0)}.
= E{X(t)E(X(s)|X(t))|X(1) = 0}
= E{X(t)(s/t)X(t)|X(1) = 0}
= (s/t)E{X 2 (t)|X(1) = 0} = (s/t)t(1 − t) = s(1 − t).
Corollary B.3 can be deduced from the this theorem via the following simple
steps: For s < t,
Cov {W 0 (s), W 0 (t)} = E{W (s) − sW (1)}{W (t) − tW (1)}
and
E{W (s)W (t) − sW (1)W (t) + stW (1)W (1) − tW (1)W (s)}
= s − st + st − st = s(1 − t).
Proof of Lemma B.1.
Cov {X(s), X(t)} = (s + 1)(t + 1) × Cov (Z(t/(t + 1)), Z(s/(s + 1))).
Next, from the definition of the Brownian bridge, Z(t) = W (t) − tW (1) so that,
expanding the above expression we obtain:
Cov {W (t/(t + 1)), W (s/(s + 1))} − t/(t + 1) × Cov {W (1), W (s/(s + 1))} −
s/(s + 1)Cov {W (1), W (t/(t + 1))} + st/((s + 1)(t + 1)) × Cov {W (1), W (1)}
= s(t + 1) − st − st + st = s.
Proof of Lemma B.2. Cov {Z(s),

tZ(t)} = E{Z(s)Z(t)}.
s
We write this as E 0 X(y)dy 0 X(u)du. Bringing together the two integral
operators we have:
s t s t
E X(y)X(u)dydu = EX(y)X(u)dydu
0 0 0 0
s t s u t
= min(y, u)dydu = ydy + udydu = s2 (t/2 − s/6).
0 0 0 0 u
Proof of Theorem C.1. Assume that ε > 0. The variables X1 , X2 , . . . , Xn are

uncorrelated so thatV ( n X
i=1 i ) = n
i=1 V (Xi ). A simple application of the
Chebyshev-Bienaymé inequality implies that:
" n # " n #
1 1 1
n
1 C
P Xi > ε ≤ 2 V Xi = 2 2 V (Xi ) ≤ 2 .
n ε n n ε nε
i=1 i=1 i=1

1 n

As a result, limn→∞ P Xi > ε = 0 which is what we needed to
n i=1
show.
Proof of Theorem C.2. Suppose that k < m,
Cov (Sk∗ , Sm
∗
) = E(Sk∗ Sm
∗
) − E(Sk∗ )E(Sm
∗
)
√ −2
= (σ n) E{Sk Sm }
√
= (σ n)−2 E{Sk (Sk + Xi )}.
k<i≤m
Developing the bracket gives,

√
(σ n)−2 {E(Sk2 ) + E(Sk )E( Xi )}
k<i≤m
√ √
= (σ n)−2 E(Sk2 ) = (σ n)−2 kσ 2 = k/n = t.
Proof of Theorem C.6 and B.6. For the first of these two theorems, note that the
√
marginal and conditional normality of any sequence of n{Fn (ti ) − F (ti )}, (0 <
√
t1 < . . . < tm ), for some m, indicates the multivariate normality of n{Fn (ti ) −
√
F (ti )}, (0 < t1 < . . . < tm ). Thus n{Fn (t) − F (t)} is a Gaussian process. It
has mean zero and variance F (t){1 − F (t)}. To obtain the covariance, note first
that the indicator variables I(Ti ≤ t) and I(Tj ≤ s) are independent for all t and
s and i = j. It only remains to evaluate, for s < t
Cov {I(Ti ≤ t), I(Ti ≤ s)} = E{I(Ti ≤ t)I(Ti ≤ s)} − E{I(Ti ≤ t)}E{I(Ti ≤ s)}
= F (s) − F (t)F (s) = F (s){1 − F (t)}.
For the second of these two theorems, note that:
E{dM (t)|Ft− } = E{H(t)dM (t)|Ft− } = H(t)E{dM (t)|Ft− } = 0.
Furthermore,
Var{dM (t)|Ft− } = Var{H(t)dM (t)|Ft− } = H(t)2 Var{dM (t)|Ft− } = H(t)2 dM (t).
Similarly,
$ % t
HdM, KdM (t) = H(s)K(s)dM, M (s).
0
Proof of Theorem D.4. Suppose k individuals are observed to fail at times

ti , ..., tk and have corresponding explanatory variables zi to zk . Assume times
are ordered. Then the probability, conditional upon the observed failure times, of
obtaining a particular ordering is:
∞ ∞ ∞
Pr (r = (1), (2), ..., (n)) = Pr (t1 < t2 < t3 ) = ... f (t1 |z1 )
0 t1 tn−1
∞ ∞ ∞ n

...f (tn |zn )dtn ...dt1 = ... {h0 (ti )eβzi exp[−H0 (ti )eβzi ]}
0 t1 tn−1 i=1
n
n

= {exp(βZi )/ Yj (Xi ) exp(βZj }δi
i=1 j=1
Bibliography
O. Aalen, A model for nonparametric regression analysis of counting processes,

in Mathematical Statistics and Probability Theory. Lecture Notes in Statistics
2 (Springer, New York, 1980), pp. 1–25
O. Aalen, O. Borgan, H. Gjessing, Survival and Event History Analysis: A Process

Point of View (Springer, New-York, 2008)
O.O. Aalen, Nonparametric inference for a family of counting processes. Ann.

Stat. 6, 701–726 (1978)
O.O. Aalen, A linear regression model for the analysis of life times. Stat. Med.
8, 907–925 (1989)
M. Abrahamowicz, T. MacKenzie, J. Esdaile, Time-dependent hazard ratio: mod-

eling and hypothesis testing with application in lupus nephritis. J. Am. Stat.
Assoc. 91, 1432–1439 (1996)
M. Abramowitz, I.A. Stegun, Handbook of mathematical functions with formulas,

graphs, and mathematical table, in US Department of Commerce. National
Bureau of Standards Applied Mathematics series 55 (1965)
Acute Leukemia Group B, E.J. Freireich, E. Gehan, E. Frei, L.R. Schroeder, I.J.
Wolman, R. Anbari, E.O. Burgert, S.D. Mills, D. Pinkel, O.S. Selawry, J.H.
Moon, B.R. Gendel, C.L. Spurr, R. Storrs, F. Haurani, B. Hoogstraten, S. Lee,
The effect of 6-mercaptopurine on the duration of steroid-induced remissions
in acute leukemia: a model for evaluation of other potentially useful therapy.
Blood 21(6), 699–716 (1963)
D.G. Altman, P.K. Andersen, A note on the uncertainty of a survival probability

estimated from Cox’s regression model. Biometrika 73(3), 722–724 (1986)
S.Z. Abildstrom, P.K. Andersen, S. Rosthoj, Competing risks as a multi-state

model. Stat. Methods Med. Res. 11(2), 203–215 (2002)
P. Andersen, Testing goodness of fit of Cox’s regression and life model. Biometrics

38, 67–77 (1982)
https://doi.org/10.1007/978-3-030-33439-0
450 Bibliography
P.K. Andersen, R.D. Gill, Cox’s regression model for counting processes: a large
sample study. Ann. Stat. 10, 1100–1120 (1982)
P.K. Andersen, O. Borgan, R. Gill, N. Keiding, Statistical Models Based on

Counting Processes (Springer, New York, 1993)
P.K. Andersen, E. Christensen, L. Fauerholdt, P. Schlichting, Evaluating prog-

noses based on the proportional hazards model. Scand. J. Stat. 10(2), 141–144
(1983)
J.A. Anderson, A. Senthilselvan, A two-step regression model for hazard function.

J. R. Stat. Soc. Ser. C 31, 44–51 (1982)
I. Annesi, T. Moreau, J. Lellouch, Efficiency of the logistic regression and cox

proportional hazards models in longitudinal studies. Stat. Med. 8(12), 1515–
1521 (1989)
E. Arjas, A graphical method for assessing goodness of fit in Cox’s proportional

hazards model. J. Am. Stat. Assoc. 83, 204–212 (1988)
P.C. Austin, J.P. Fine, Accounting for competing risks in randomized controlled
trials: a review and recommendations for improvement. Stat. Med. 36(8),
1203–1209 (2017)
G. Bakoyannis, G. Touloumi, Practical methods for competing risks data: a

review. Stat. Methods Med. Res. 21(3), 257–272 (2012)
N. Balakrishnan, N. Johnson, Samuel kotz. Continuous univariate distributions,

vol. 1 (1994)
I. Baltazar-Aban, E. Peña, Properties of hazard-based residuals and implications

in model diagnostics. J. Am. Stat. Assoc. 90, 185–219 (1995)
W.E. Barlow, R.L. Prentice, Residuals for relative risk regression. Biometrika
75(1), 65–74 (1988)
O.E. Barndorff-Nielsen, P. Hall, On the level-error after Bartlett adjustment of

the likelihood ratio statistic. Biometrika 75(2), 374–378 (1988)
O.E. Barndorff-Nielsen, D.R. Cox, Inference and Asymptotics, tome 13 (Chapman

& Hall London, 1994)
S. Bennett, Analysis of survival data by the proportional odds model. Stat. Med.
2, 273–277 (1983a)
S. Bennett, Log-logistic regression models for survival data. Appl. Stat. 32, 165–
171 (1983b)
Bibliography 451
R. Beran, Nonparametric regression with randomly censored survival data. Tech-

nical Report, University of California, Berkeley (1981)
R.N. Bhattacharya, E.C. Waymire, Stochastic Processes with Applications

(Wiley, New York, 1990)
P. Billingsley, Convergence of Probability Measures, 2nd edn. (Wiley, New York,

1999)
D.A. Binder, Fitting cox’s proportional hazards models from survey data.
Biometrika 79(1), 139–147 (1992)
O. Borgan, K. Liestol, A note on confidence intervals and bands for the survival
function based on transformations. Scand. J. Stat. 35–41 (1990)
L. Breiman, Classification and Regression Trees (Routledge, 2017)
N. Breslow, A generalized Kruskal-Wallis test for comparing K samples subject

to unequal patterns of censorship. Biometrika 57, 579–594 (1970)
N. Breslow, Discussion of the paper by D. R. Cox. J. R. Stat. Soc. B 34, 216–217

(1972)
N. Breslow, The proportional hazards model: applications in epidemiology. Com-

mun. Stat. Theory Methods 7(4), 315–332 (1978)
N. Breslow, J. Crowley, A large sample study of the life table and product limit
estimates under random censorship. Ann. Stat. 437–453 (1974)
N. Breslow, L. Elder, L. Berger, A two sample censored-data rank test for accel-
eration. Biometrics 40, 1042–1069 (1984)
N.E. Breslow, Statistics in epidemiology: the case-control study. J. Am. Stat.

Assoc. 91(433), 14–28 (1996)
M.S. Brose, T.R. Rebbeck, K.A. Calzone, J.E. Stopfer, K.L. Nathanson, B.L.
Weber, Cancer risk estimates for brca1 mutation carriers identified in a risk
evaluation program. J. Natl. Cancer Inst. 94(18), 1365–1372 (2002)
C. Brown, On the use of indicator variables for studying the time-dependence of

parameters in a response-time model. Biometrics 31, 863–872 (1975)
Z. Cai, Y. Sun, Local linear estimation for time-dependent coefficients in Cox’s

regression models. Scand. J. Stat. 30, 93–111 (2003)
H.-S.D. Cain, C. Kevin, R.J. Little, B. Nan, M. Yosef, J.R. Taffe, M.R. Elliott,
Bias due to left truncation and left censoring in longitudinal studies of devel-
opmental and disease processes. Am. J. Epidemiol. 173(9), 1078–1084 (2011)
452 Bibliography
B.P. Carlin, J.S. Hodges, Hierarchical proportional hazards regression models for
highly stratified data. Biometrics 55(4), 1162–1170 (1999)
C.L. Chang, Introduction to Stochastic Processes in Biostatistics (Wiley, New

York, 1968)
C. Chauvel, Empirical Processes for Inference in the Non-Proportional Hazards

Model. Thèse de doctorat, PhD thesis, Université Pierre et Marie Curie-Paris
6, Paris (2014)
C. Chauvel, J. O’Quigley, Tests for comparing estimated survival functions.

Biometrika 101(3), 535–552 (2014)
C. Chauvel, J. O’Quigley, Survival model construction guided by fit and predictive

strength. Biometrics 73(2), 483–494 (2017)
Q. Chen, R.C. May, J.G. Ibrahim, H. Chu, S.R. Cole, Joint modeling of longitu-
dinal and survival data with missing and left-censored time-varying covariates.
Stat. Med. 33(26), 4560–4576 (2014)
S. Cheng, L. Wei, Z. Ying, Analysis of transformation models with censored data.

Biometrika 82, 835–845 (1995)
B. Choodari-Oskooei, P. Royston, M.K.B. Parmar, A simulation study of predic-

tive ability measures in a survival model I: explained variation measures. Stat.
Med. 31(23), 2627–2643 (2012)
A. Ciampi, J. Lawless, S. McKinney, K. Singhal, Regression and recursive par-

tition strategies in the analysis of medical survival data. J. Clin. Epidemiol.
41(8), 737–748 (1988)
A. Ciampi, Z. Lou, Q. Lin, A. Negassa, Recursive partition and amalgamation

with the exponential family: theory and applications. Appl. Stoch. Models Data
Anal. 7(2), 121–137 (1991)
W.G. Cochran, Some methods for strengthening the common χ 2 tests. Biomet-
rics 10(4), 417–451 (1954)
Fibrinogen Studies Collaboration, Measures to assess the prognostic ability of the

stratified cox proportional hazards model. Stat. Med. 28(3), 389–411 (2009)
J. Cologne, W.-L. Hsu, R.D. Abbott, W. Ohishi, E.J. Grant, S. Fujiwara, H.M.
Cullings, Proportional hazards regression in epidemiologic follow-up studies:
an intuitive consideration of primary time scale. Epidemiology 565–573 (2012)
D.R. Cox, Planning of Experiments (1958)
D.R. Cox, Regression models and life–tables (with discussion). J. R. Stat. Soc.
Ser. B 34(2), 187–220 (1972)
Bibliography 453
D.R. Cox, Partial likelihood. Biometrika 63, 269–276 (1975)
D.R. Cox, D.V. Hinkley, Theoretical Statistics (Chapman and Hall/CRC, 1979)
D.R. Cox, E.J. Snell, A general definition of residuals. J. R. Stat. Soc. Ser. B
30(2), 248–275 (1968)
H. Cramér, Random Variables and Probability Distributions, tome 36 (Cambridge

University Press, 2004)
M. Crowder, Identifiability crises in competing risks. Int. Stat. Rev. 379–391

(1994)
J.J. Crowley, B.E. Storer, Comment on ’A reanalysis of the Stanford Heart Trans-
plant Data’, by M. Aitkin, N. Laird and B. Francis. J. Am. Stat. Assoc. 78,
277–281 (1983)
M.J. Crowther, K.R. Abrams, P.C. Lambert, Joint modeling of longitudinal and
survival data. Stata J. 13(1), 165–184 (2013)
S. CsÖrgÓ, L. Horvath, On the koziol-green model for random censorship.

Biometrika 68(2), 391–401 (1981)
D.M. Dabrowska, K.A. Doksum, Partial likelihood in transformation models with

censored data. Scand. J. Stat. 18, 1–23 (1988)
H. Daniels, Saddlepoint approximations for estimating equations. Biometrika

70(1), 89–96 (1983)
H.E. Daniels, Saddlepoint approximations in statistics. Ann. Math. Stat. 631–650

(1954)
H.E. Daniels, Exact saddlepoint approximations. Biometrika 67(1), 59–63 (1980)
H.E. Daniels, Tail probability approximations. Int. Stat. Rev./Revue Interna-

tionale de Statistique 37–48 (1987)
H. David, Concomitants of extreme order statistics, in Extreme Value Theory

and Applications (Springer, 1994), pp. 211–224
R.B. Davies, Hypothesis testing when a nuisance parameter is present only under
the alternative. Biometrika 64(2), 247–254 (1977)
R.B. Davies, Hypothesis testing when a nuisance parameter is present only under
the alternative. Biometrika 74(1), 33–43 (1987)
Z.-Q. Dignam, J. James, M. Kocherginsky, The use and interpretation of com-

peting risks regression models. Clin. Cancer Res. 18(8), 2301–2308 (2012)
454 Bibliography
M. Donsker, An invariance principle for certain probability limit theorems. Mem.

Am. Math. Soc. 6, 1–10 (1951)
N.R. Draper, The Box-Wetz criterion versus R2. J. R. Stat. Soc. Ser. A (General)
147(1), 100–103 (1984)
N.R. Draper, Corrections: the Box-Wetz criterion versus R2. J. R. Stat. Soc. Ser.
A (General) 148(4), 357–357 (1985)
L. Duchateau, P. Janssen, The Frailty Model (Springer Science & Business

Media, 2007)
B. Efron, Censored data and the bootstrap. J. Am. Stat. Assoc. 76(374), 312–
319 (1981a)
B. Efron, Nonparametric estimates of standard error: the jackknife, the bootstrap

and other methods. Biometrika 68(3), 589–599 (1981b)
B. Efron, Nonparametric standard errors and confidence intervals. Can. J. Stat.

9(2), 139–158 (1981c)
B. Efron, Better bootstrap confidence intervals. J. Am. Stat. Assoc. 82(397),

171–185 (1987)
B. Efron, D.V. Hinkley, Assessing the accuracy of the maximum likelihood esti-
mator: observed versus expected Fisher information. Biometrika 65, 457–483
(1978)
B. Efron, C. Stein, The jackknife estimate of variance. Ann. Stat. 586–596 (1981)
B. Efron et al., The geometry of exponential families. Ann. Stat. 6(2), 362–376
(1978)
S.S. Ellenberg, J.M. Hamilton, Surrogate endpoints in clinical trials: cancer. Stat.
Med. 8(4), 405–413 (1989)
K.H. Eng, M.R. Kosorok, A sample size formula for the supremum log-rank
statistic. Biometrics 61, 86–91 (2005)
M. Evans, N. Hastings, B. Peacock, Statistical Distributions (2001)
G. Fang, W. Liu, L. Wang, A machine learning approach to select features impor-

tant to stroke prognosis. Comput. Biol. Chem. 107316 (2020)
P. Feigl, M. Zelen, Estimation of exponential survival probabilities with concomi-

tant information. Biometrics 21(4), 826–838 (1965)
J.P Fine, R.J. Gray, A proportional hazards model for the subdistribution of a
competing risk. J. Am. Stat. Assoc. 94(446), 496–509 (1999)
Bibliography 455
J.P. Fine, H. Jiang, R. Chappell, On semi-competing risks data. Biometrika

88(4), 907–919 (2001)
R.A. Fisher, Two new properties of mathematical likelihood. Proceedings of the

Royal Society of London. Series A, Containing Papers of a Mathematical and
Physical Character, vol. 144(852), pp. 285–307 (1934)
E. Fix, J. Neyman, A simple stochastic model of recovery, relapse, death and loss
of patients. Hum. Biol. 23(3), 205–241 (1951)
P. Flandre, J. O’Quigley, A two-stage procedure for survival studies with surrogate

endpoints. Biometrics 969–976 (1995)
P. Flandre, J. O’Quigley, Comparing Kaplan–Meier curves with delayed treatment

effects: applications in immunotherapy trials. J. R. Stat. Soc. Ser. C (Appl.
Stat.) (2019)
P. Flandre, J. O’Quigley, The short-term and long-term hazard ratio model:

parameterization inconsistency. Am. Stat. 1–7 (2020)
P. Flandre, Y. Saidi, Estimating the proportion of treatment effect explained by

a surrogate marker by DY Lin, TR Fleming and V. De Gruttola. Stat. Med.
16, 1515–1527 (1997). Stat. Med. 18(1), 107–109 (1999)
T.R. Fleming, D.P. Harrington, Counting Processes and Survival Analysis (Wiley,
New York, 1991)
T.R. Fleming, D.P. Harrington, Counting Processes and Survival Analysis, 2nd
edn. (Wiley, New York, 2005)
T.R. Fleming, D.P. Harrington, Evaluation of censored survival data test proce-
dures based on single and multiple statistics, in Topics in Applied Statistics
(Marcel Dekker, New York, 1984), pp. 97–123
T.R. Fleming, D.P. Harrington, M. O’Sullivan, Supremum versions of the log-

rank and generalized Wilcoxon statistics. J. Am. Stat. Assoc. 82, 312–320
(1987)
T.R. Fleming, J.R. O’Fallon, P.C. O’Brien, D.P. Harrington, Modified

Kolmogorov-Smirnov test procedures with application to arbitrarily right-
censored data. Biometrics 36, 607–625 (1980)
L.S. Freedman, B.I. Graubard, A. Schatzkin, Statistical validation of intermediate

endpoints for chronic diseases. Stat. Med. 11(2), 167–178 (1992)
B. Freidlin, E.L. Korn, Testing treatment effects in the presence of competing

risks. Stat. Med. 24(11), 1703–1712 (2005)
456 Bibliography
J.E. Freund, A bivariate extension of the exponential distribution. J. Am. Stat.

Assoc. 56(296), 971–977 (1961)
A. Gaddah, R. Braekers, Weak convergence for the conditional distribution func-
tion in a koziol-green model under dependent censoring. J. Stat. Plann. Infer.
139(3), 930–943 (2009)
U. Gather, J. Pawlitschko, Estimating the survival function under a generalized
koziol-green model with partially informative censoring. Metrika 48(3), 189–
207 (1998)
E. Gehan, A generalized Wilcoxon test for comparing arbitrarily singly-censored
samples. Biometrika 52, 203–223 (1965)
R.B. Geskus, Cause-specific cumulative incidence estimation and the fine and
gray model under both left truncation and right censoring. Biometrics 67(1),
39–49 (2011)
D. Ghosh, Semiparametric inferences for association with semi-competing risks
data. Stat. Med. 25(12), 2059–2070 (2006)
R. Gill, Censoring and stochastic integrals. Mathematical Center Tract 124
(1980)
R. Gill, M. Schumacher, A simple test of the proportional hazards assumption.
Biometrika 74(2), 289–300 (1987)
S.M. Gore, S.J. Pocock, G.R. Kerr, Regression models and non-proportional haz-
ards in the analysis of breast cancer survival. J. R. Stat. Soc. Ser. C (Appl.
Stat.) 33(2), 176–195 (1984)
P.M. Grambsch, T.M. Therneau, Proportional hazards tests and diagnostics
based on weighted residuals. Biometrika 81(3), 515–526 (1994)
R. Gray, Flexible methods for analyzing survival data using splines, with applica-
tions to breast cancer prognosis. J. Am. Stat. Assoc. 87, 942–951 (1992)
M. Greenwood, A Report on the Natural Duration of Cancer. Reports of Public
Health and Related Subjects, vol. 33 (HMSO, London, 1926)
R.G. Gutierrez, Parametric frailty and shared frailty survival models. Stata J.
2(1), 22–44 (2002)
M. Gönen, G. Heller, Concordance probability and discriminatory power in pro-
portional hazards regression. Biometrika 92, 965–970 (2005)
P. Haara, A note on the asymptotic behaviour of the empirical score in Cox’s
regression model for counting processes, in Proceedings of the 1st World
Congress of the Bernoulli Society (VNU Science Press, Tashkent, Soviet Union,
1987), pp. 139–142
Bibliography 457
B. Haller, G. Schmidt, K. Ulm, Applying competing risks regression models: an

overview. Lifetime Data Anal. 19(1), 33–58 (2013)
D.D. Hanagal, Modeling Survival Data Using Frailty Models (Chapman and Hall,
CRC, 2011)
S. Haneuse, K.H. Lee, Semi-competing risks data analysis: accounting for death
as a competing risk when the outcome of interest is nonterminal. Circ. Car-
diovasc. Qual. Outcomes 9(3), 322–331 (2016)
T.E. Hanson , A. Jara, L. Zhao, A Bayesian semi-parametric temporally stratified

proportional hazards model with spatial frailties. Bayesian Anal. 7, 147–188
(2012)
D.P. Harrington, T.R. Fleming, A class of rank test procedures for censored
survival data. Biometrika 69(3), 553–566 (1982)
T. Hastie, R. Tibshirani, Exploring the nature of covariate effects in the propor-

tional hazards model. Biometrics 46, 1005–1016 (1990)
M. Healy, The use of R2 as a measure of goodness of fit. J. R. Stat. Soc. Ser. A

(General) 147(4), 608–609 (1984)
I. Helland, Central limit theorems for martingales with discrete or continuous

time. Scand. J. Stat. 9, 79–94 (1982)
R. Henderson, P. Diggle, A. Dobson, Joint modelling of longitudinal measure-

ments and event time data. Biostatistics 1(4), 465–480 (2000)
J. Herson, The use of surrogate endpoints in clinical trials (an introduction to a

series of four papers). Stat. Med. 8(4), 403–404 (1989)
A. Hillis, D. Seigel, Surrogate endpoints in clinical trials: ophthalmologic disor-

ders. Stat. Med. 8(4), 427–430 (1989)
N.L. Hjort, On inference in parametric survival data models. Int. Stat. Rev./Revue
Internationale de Statistique 355–387 (1992)
F.S. Hodi, V. Chiarion-Sileni, R. Gonzalez, J.-J. Grob, P. Rutkowski, C.L. Cowey,

C.D. Lao, D. Schadendorf, J. Wagstaff, R. Dummer et al., Nivolumab plus
ipilimumab or nivolumab alone versus ipilimumab alone in advanced melanoma
(checkmate 067): 4-year outcomes of a multicentre, randomised, phase 3 trial.
Lancet Oncol. 19(11), 1480–1492 (2018)
F. Hsieh, Y.-K. Tseng, J.-L. Wang, Joint modeling of survival and longitudinal
data: likelihood approach revisited. Biometrics 62(4), 1037–1043 (2006)
458 Bibliography
J.-J. Hsieh, Y.-T. Huang, Regression analysis based on conditional likelihood

approach under semi-competing risks data. Lifetime Data Anal. 18(3), 302–
320 (2012)
X. Huang, L. Liu, A joint frailty model for survival and gap times between recur-
rent events. Biometrics 63(2), 389–397 (2007)
R. Huang, R. Xu, P.S. Dulai, Sensitivity analysis of treatment effect to unmea-

sured confounding in observational studies with survival and competing risks
outcomes. Stat. Med. 1–15 (2020)
J. Hyde, Testing survival under right censoring and left truncation. Biometrika
64(2), 225–230 (1977)
H. Jiang, J.P. Fine, R. Chappell, Semiparametric analysis of survival data with left
truncation and dependent right censoring. Biometrics 61(2), 567–575 (2005)
M.P. Jones, J. Crowley, A general class of nonparametric tests for survival anal-
ysis. Biometrics 45, 157–170 (1989)
M.P. Jones, J. Crowley, Asymptotic properties of a general class of nonparametric

tests for survival analysis. Ann. Stat. 18, 1203–1220 (1990)
J.D. Kalbfleisch, R.L. Prentice, Marginal likelihood based on Cox’s regression

and life model. Biometrika 60, 267–278 (1973)
J.D. Kalbfleisch, R.L. Prentice, The Statistical Analysis of Failure Time Data
(Wiley, 2002)
E.L. Kaplan, P. Meier, Nonparametric estimation from incomplete observations.

J. Am. Stat. Assoc. 53(282), 457–481 (1958)
R. Kay, Proportional hazard regression models and the analysis of censored sur-
vival data. J. R. Stat. Soc. Ser. C (Appl. Stat.) 26(3), 227–237 (1977)
N. Keiding, Historical controls and modern survival analysis. Lifetime Data Anal.
1(1), 19–25 (1995)
N. Keiding, Event history analysis and inference from observational epidemiology.

Stat. Med. 18(17–18), 2353–2363 (1999)
N. Keiding, T. Bayer, S. Watt-Boolsen, Confirmatory analysis of survival data

using left truncation of the life times of primary survivors. Stat. Med. 6(8),
939–944 (1987)
N. Keiding, R.D. Gill, Random truncation models and Markov processes. Ann.
Stat. 582–602 (1990)
Bibliography 459
M.G. Kendall, A. Stuart, J.K. Ord, S.F. Arnold, Kendall’s advanced theory of
statistics (1987)
J.T. Kent, J. O’Quigley, Measures of dependence for censored survival data.
Biometrika 75(3), 525–534 (1988)
H.T. Kim, Cumulative incidence in competing risks data and competing risks
regression analysis. Clin. Cancer Res. 13(2), 559–565 (2007)
K. Kim, A.A. Tsiatis, Study duration for clinical trials with survival response and
early stopping rule. Biometrics 81–92 (1990)
J.P. Klein, M.L. Moeschberger, Survival Analysis Techniques for Censored and
Truncated Data (Springer, 2003)
J.P. Klein, S.-C. Lee, M. Moeschberger, A partially parametric estimator of sur-
vival in the presence of randomly censored data. Biometrics 795–811 (1990)
A.N. Kolmogorov, Foundations of the Theory of Probability: Second English
Edition (2018)
M.R. Kosorok, C.Y. Lin, The versatility of function-indexed weighted log-rank
statistics. J. Am. Stat. Assoc. 94, 320–332 (1999)
J.A. Koziol, S.B. Green, A Cramer-von Mises statistic for randomly censored
data. Biometrika 63(3), 465–474 (1976)
J.A. Koziol, J.-Y. Zhang, C.A. Casiano, X.-X. Peng, F.-D. Shi, A.C. Feng, E.K.
Chan, E.M. Tan, Recursive partitioning as an approach to selection of immune
markers for tumor diagnosis. Clin. Cancer Res. 9(14), 5120–5126 (2003)
T.O. Kvaalseth, Cautionary note about R2. Am. Stat. 39(4), 279–285 (1985)
S. Lagakos, D. Schoenfeld, Properties of proportional-hazards score tests under
misspecified regression models. Biometrics 1037–1048 (1984)
S. Lagakos, The graphical evaluation of explanatory variables in proportional
hazards regression models. Biometrika 68, 93–98 (1981)
S. Lagakos, The loss in efficiency from misspecifying covariates in proportional
hazards regression models. Biometrika 75(1), 156–160 (1988)
S.W. Lagakos, A stochastic model for censored-survival data in the presence of
an auxiliary variable. Biometrics 551–559 (1976)
S.W. Lagakos, Using auxiliary variables for improved estimates of survival time.
Biometrics 399–404 (1977)
S.W. Lagakos, L.L. Kim, J.M. Robins, Adjusting for early treatment termination
in comparative clinical trials. Stat. Med. 9, 1417–1424 (1990)
460 Bibliography
S.W. Lagakos, C.J. Sommer, M. Zelen, Semi-Markov models for partially cen-
sored data. Biometrika 65(2), 311–317 (1978)
M.G. Larson, G.E. Dinse, A mixture model for the regression analysis of compet-
ing risks data. J. R. Stat. Soc. Ser. C (Appl. Stat.) 34(3), 201–211 (1985)
B. Lau, S.R. Cole, S.J. Gange, Competing risk regression models for epidemiologic
data. Am. J. Epidemiol. 170(2), 244–256 (2009)
M. LeBlanc, J. Crowley, Relative risk trees for censored survival data. Biometrics
411–425 (1992)
J. Lee, Some versatile tests based on the simultaneous use of weighted log-rank
statistics. Biometrics 52, 721–725 (1996)
K.H. Lee, S. Haneuse, D. Schrag, F. Dominici, Bayesian semi-parametric anal-

ysis of semi-competing risks data: investigating hospital readmission after a
pancreatic cancer diagnosis. J. R. Stat. Soc. Ser. C, Appl. Stat. 64(2), 253
(2015)
S.-H. Lee, Maximum of the weighted Kaplan-Meier tests for the two-sample
censored data. J. Stat. Comput. Simul. 81, 1017–1026 (2011)
E.L. Lehmann et al., The power of rank tests. Ann. Math. Stat. 24(1), 23–43
(1953)
E. Lenglart, Relation de domination entre deux processus. Annales de l’Institut

Henri Poincaré 13, 171–179 (1977)
S. Leurgans, Three classes of censored data rank tests: strengths and weaknesses
under censoring. Biometrika 70, 651–658 (1983)
S. Leurgans, Asymptotic behavior of two-sample rank tests in the presence of

random censoring. Ann. Stat. 12, 572–589 (1984)
Y. Li, L. Ryan, Modeling spatial survival data using semiparametric frailty models.
Biometrics 58(2), 287–297 (2002)
D. Lin, Goodness-of-fit analysis for the Cox regression model based on a class of
parameter estimators. J. Am. Stat. Assoc. 86, 153–180 (1991)
D. Lin, Cox regression analysis of multivariate failure time data: the marginal
approach. Stat. Med. 13(21), 2233–2247 (1994)
D. Lin, J. Robins, L. Wei, Comparing two failure time distributions in the presence
of dependent censoring. Biometrika 83, 381–393 (1996)
D. Lin, L. Wei, Z. Ying, Checking the Cox model with cumulative sums of mar-
tingale based residuals. Biometrika 80, 557–572 (1993)
Bibliography 461
D.Y. Lin, Z. Ying, Semiparametric analysis of the additive risk model. Biometrika
81(1), 61–71 (1994)
D.Y. Lin, Transformation models, in Handook of Survival Analysis (Chapman &

Hall/CRC, 2013)
D.Y. Lin, L.-J. Wei, The robust inference for the cox proportional hazards model.
J. Am. Stat. Assoc. 84(408), 1074–1078 (1989)
D.Y. Lin, Z. Ying, Semiparametric analysis of general additive-multiplicative

intensity for counting processes. Ann. Stat. 23, 1712–1734 (1995)
A. Linden, P.R. Yarnold, Modeling time-to-event (survival) data using classifica-

tion tree analysis. J. Eval. Clin. Pract. 23(6), 1299–1308 (2017)
C.L. Link, Confidence intervals for the survival function using Cox’s proportional-
hazard model with covariates. Biometrics 601–609 (1984)
L. Liu, R.A. Wolfe, X. Huang, Shared frailty models for recurrent events and a
terminal event. Biometrics 60(3), 747–756 (2004)
W.-Y. Loh, Classification and regression trees. Wiley Interdisc. Rev. Data Min.
Knowl. Disc. 1(1), 14–23 (2011)
W.-Y. Loh, Fifty years of classification and regression trees. Int. Stat. Rev. 82(3),
329–348 (2014)
M. Lunn, D. McNeil, Applying Cox regression to competing risks. Biometrics

524–532 (1995)
H.M. Malani, A modification of the redistribution to the right algorithm using

disease markers. Biometrika 82(3), 515–526 (1995)
N. Mantel, Chi-square tests with one degree of freedom; extensions of the Mantel-
Haenszel procedure. J. Am. Stat. Assoc. 58(303), 690–700 (1963)
N. Mantel, Evaluation of survival data and two new rank order statistics arising
in its consideration. Cancer Chemother. Rep. 50, 163–70 (1966)
N. Mantel, W. Haenszel, Statistical aspects of the analysis of data from retro-

spective studies of disease. J. Natl. Cancer Inst. 22(4), 719–748 (1959)
N. Mantel, D.M. Stablein, The crossing hazard function problem. The Statistician
37, 59–64 (1988)
T. Martinussen, L. Peng, Alternatives to the Cox model, in Handook of Survival

Analysis, ed. by J.P. Klein, H.C. van Houwelingen, J.G. Ibrahim, T.H. Scheike
(Chapman & Hall/CRC, 2013)
462 Bibliography
T. Martinussen, T.H. Scheike, Dynamic Regression Models for Survival Data

(Springer, 2006)
I.W. McKeague, P.D. Sasieni, A partly parametric additive risk model. Biometrika
81, 501–514 (1994)
I.W. McKeague, K.J. Utikal, Inference for a nonlinear counting process regression
model. Ann. Stat. 18, 1172–1187 (1990)
M.L. Moeschberger, J.P. Klein, A comparison of several methods of estimating

the survival function when there is extreme right censoring. Biometrics 41(1),
253–259 (1985)
M.L. Moeschberger, K.P. Tordoff, N. Kochar, A review of statistical analyses for

competing risks. Handb. Stat. 27, 321–341 (2007)
S.H. Moolgavkar, E.T. Chang, H.N. Watson, E.C. Lau, An assessment of the
cox proportional hazards regression model for epidemiologic studies. Risk Anal.
38(4), 777–794 (2018)
T. Moreau, J. O’Quigley, M. Mesbah, A global goodness–of–fit statistic for the

proportional hazards model. J. R. Stat. Soc. Ser. C 34(3), 212–218 (1985)
S. Murphy, Testing for a time-dependent coefficient in Cox’s regression model.

Scand. J. Stat. 20, 35–50 (1993)
S.A. Murphy, P.K. Sen, Time-dependent coefficients in a Cox-type regression

model. Stoch. Process. Appl. 39, 153–180 (1991)
S. Murray, A.A. Tsiatis, Sequential methods for comparing years of life saved in
the two sample censored data problem. Biometrics 55 (1999)
S. Murray, A.A. Tsiatis, Using auxiliary time-dependent covariates to recover

information in nonparametric testing with censored data. Lifetime Data Anal.
7 (2001)
S. Murray, A.A. Tsiatis, Nonparametric survival estimation using prognostic lon-

gitudinal covariates. Biometrics 137–151 (1996)
M.H. Myers, B.F. Hankey, N. Mantel, A logistic-exponential model for use with
response-time data involving regressor variables. Biometrics 257–269 (1973)
L. Natarajan, J. O’Quigley, Predictive capability of stratified proportional hazards

models. J. Appl. Stat. 29(8), 1153–1163 (2002)
W. Nelson, Hazard plotting for incomplete failure data. J. Qual. Technol. 1(1),
27–52 (1969)
Bibliography 463
M.A. Newton, A.E. Raftery, Approximate Bayesian inference with the weighted
likelihood bootstrap. J. R. Stat. Soc. Ser. B (Methodological) 56(1), 3–26
(1994)
D. Oakes, An approximate likelihood procedure for censored data. Biometrics

177–182 (1986)
P.C. O’Brien, A nonparametric test for association with censored data. Biometrics
243–250 (1978)
P.C. O’Brien, T.R. Fleming, A paired Prentice-Wilcoxon test for censored paired
data. Biometrics 169–180 (1987)
J. O’Quigley, Regression models and survival prediction. Statistician 31, 106–116

(1982)
J. O’Quigley, Confidence intervals for the survival function in the presence of

covariates. Biometrics 42(1), 219–220 (1986)
J. O’Quigley, On a two-sided test for crossing hazards. J. R. Stat. Soc. Ser. D

(The Statistician) 43(4), 563–569 (1994)
J. O’Quigley, Khmaladze-type graphical evaluation of the proportional hazards

assumption. Biometrika 90, 577–584 (2003)
J. O’Quigley, Proportional Hazards Regression (Springer, New York, 2008)
J. O’Quigley, Faulty brca1, brca2 genes: how poor is the prognosis? Anna. Epi-
demiol. 27(10), 672–676 (2017)
J. O’Quigley, P. Flandre, Predictive capability of proportional hazards regression.

Proc. Natl. Acad. Sci. 91(6), 2310–2314 (1994)
J. O’Quigley, P. Flandre, Quantification of the prentice criteria for surrogate

endpoints. Biometrics 62(1), 297–300 (2006)
J. O’Quigley, T. Moreau, Testing the proportional hazards regression model

against some general alternatives. Revue d’epidemiologie et de sante publique
32(3–4), 199–205 (1984)
J. O’Quigley, L. Natarajan, Erosion of regression effect in a survival study. Bio-

metrics 60(2), 344–351 (2004)
J. O’Quigley, F. Pessione, Score tests for homogeneity of regression effect in the

proportional hazards model. Biometrics 45, 135–144 (1989)
J. O’Quigley, F. Pessione, The problem of a covariate-time qualitative interaction

in a survival study. Biometrics 47, 101–15 (1991)
464 Bibliography
J. O’Quigley, R.L. Prentice, Nonparametric tests of association between survival

time and continuously measured covariates: the logit-rank and associated pro-
cedures. Biometrics 117–127 (1991)
J. O’Quigley, J. Stare, Proportional hazards models with frailties and random

effects. Stat. Med. 21(21), 3219–3233 (2002)
J. O’Quigley, R. Xu, Explained variation in proportional hazards regression, in

Handbook of Statistics in Clinical Oncology, ed. by J. Crowley, A. Hoering
(Marcel Dekker, New York, 2001), pp. 397–410
J. O’Quigley, R. Xu, Goodness of fit in survival analysis. Wiley StatsRef: Statistics

Reference Online (2014)
J. O’Quigley, R. Xu, J. Stare, Explained randomness in proportional hazards

models. Stat. Med. 24(3), 479–489 (2005)
M. Peckova, T.R. Fleming, Adaptive test for testing the difference in survival
distributions. Lifetime Data Anal. 9, 223–238 (2003)
M.J. Pencina, R.B. D’Agostino, R.S. Vasan, Evaluating the added predictive
ability of a new marker: from area under the ROC curve to reclassification and
beyond. Stat. Med. 27, 157–172 (2008)
M.S. Pepe, T.R. Fleming, Weighted Kaplan-Meier statistics: a class of distance

tests for censored survival data. Biometrics 45, 497–507 (1989)
M.S. Pepe, J. Fan, Z. Feng, T. Gerds, J. Hilden, The net reclassification index
(nri): a misleading measure of prediction improvement even with independent
test data sets. Stat. Biosci. 7(2), 282–295 (2015)
M.P. Perme, J. Stare, J. Estève, On estimation in relative survival. Biometrics

68(1), 113–120 (2012)
R. Peto, J. Peto, Asymptotically efficient rank invariant test procedures. J. R.

Stat. Soc. Ser. A 135, 185–207 (1972)
E. Pitman, Nonparametric statistical inference. Lecture Notes (Institute of Statis-

tics, University of North Carolina, 1948)
D.N. Politis, Computer-intensive methods in statistical analysis. IEEE Signal Pro-

cess. Mag. 15(1), 39–55 (1998)
R.L. Prentice, Linear rank tests with right censored data. Biometrika 65, 167–179
(1978)
R.L. Prentice, Surrogate endpoints in clinical trials: definition and operational

criteria. Stat. Med. 8(4), 431–440 (1989)
Bibliography 465
R.L. Prentice, J.D. Kalbfleisch, Hazard rate models with covariates. Biometrics
25–39 (1979)
H. Putter, M. Fiocco, R.B. Geskus, Tutorial in biostatistics: competing risks and

multi-state models. Stat. Med. 26(11), 2389–2430 (2007)
E.R. Brown, J.G. Ibrahim, A Bayesian semiparametric joint hierarchical model

for longitudinal and survival data. Biometrics 59(2), 221–228 (2003)
R.H. Randles, D.A. Wolfe, Introduction to the theory of nonparametric statistics.

Introduction to the Theory of Nonparametric Statistics, by Randles, Ronald
H.; Wolfe, Douglas A. Wiley, New York, c1979. Wiley series in probability and
mathematical statistics (1979)
C.R. Rao, C.R. Rao, M. Statistiker, C.R. Rao, C.R. Rao, Linear Statistical Infer-
ence and Its Applications, tome 2 (Wiley, New York, 1973)
S. Rashid, J. O’Quigley, A. Axon, E. Cooper, Plasma protein profiles and prog-

nosis in gastric cancer. Br. J. Cancer 45(3), 390 (1982)
R. Rebolledo, Sur les applications de la théorie des martingales à l’étude statis-

tique d’une famille de processus ponctuels. Dans Journées de Statistique des
Processus Stochastiques (Springer, 1978), pp. 27–70
R. Rebolledo, Central limit theorems for local martingales. Z. Wahrsch. Verw.

Gebiete 51, 269–286 (1980)
D. Rizopoulos, Joint Models for Longitudinal and Time-to-Event Data: With

Applications in R. (Chapman and Hall/CRC, 2012)
D. Rizopoulos, G. Molenberghs, E.M.E.H. Lesaffre, Dynamic predictions with

time-dependent covariates in survival analysis using joint modeling and land-
marking. Biometrical J. 59(6), 1261–1276 (2017)
J. Robinson, Saddlepoint approximations for permutation tests and confidence

intervals. J. R. Stat. Soc. Ser. B 44, 91–101 (1982)
V.K. Rohatgi, A.K.M.E. Saleh, An Introduction to Probability and Statistics

(Wiley, 2015)
P.S. Rosenberg, W.F. Anderson, Proportional hazards models and age-period-

cohort analysis of cancer rates. Stat. Med. 29(11), 1228–1238 (2010)
S. Ross, Stochastic Processes, 2nd edn. (Wiley, New York, 1996)
D.B. Rubin, The Bayesian bootstrap. Ann. Stat. 130–134 (1981)
P. Sasieni, On the expected number of cancer deaths during follow-up of an

initially cancer-free cohort. Epidemiology 14(1), 108–110 (2003)
466 Bibliography
P. Sasieni, A. Winnett, Martingale difference residuals as a diagnostic tool for

the Cox model. Biometrika 90, 899–912 (2003)
J.M. Satagopan, L. Ben-Porat, M. Berwick, M. Robson, D. Kutler, A.D. Auer-

bach, A note on competing risks in survival data analysis. Br. J. Cancer 91(7),
1229–1235 (2004)
T.H. Scheike, T. Martinussen, On estimation and tests of time-varying effects in

the proportional hazards model. Scand. J. Stat. 31, 51–62 (2004)
T.H. Scheike, M. Zhang, An additive-multiplicative Cox-Aalen regression model.

Scand. J. Stat. 29, 79–92 (2002)
T.H. Scheike, M.-J. Zhang, Flexible competing risks regression modeling and
goodness-of-fit. Lifetime Data Anal. 14(4), 464 (2008)
D. Schoenfeld, Chi-squared goodness-of-fit tests for the proportional hazards

regression model. Biometrika 67, 145–153 (1980)
D. Schoenfeld, The asymptotic properties of nonparametric tests for comparing

survival distributions. Biometrika 68(1), 316–319 (1981)
D. Schoenfeld, Partial residuals for the proportional hazards regression model.

Biometrika 69, 239–241 (1982)
M. Schumacher, Two-sample tests of Cramer-von Mises and Kolmogorov-

Smirnov-type for randomly censored data. Int. Stat. Rev. 62, 263–281 (1984)
M. Schumacher, E. Graf, T. Gerds, How to assess prognostic models for survival

data: a case study in oncology. Methods Inf. Med. 42(05), 564–571 (2003)
A. Scott, C. Wild, Transformations and R 2. Am. Stat. 45(2), 127–129 (1991)
S. Self, An adaptive weighted log-rank test with application to cancer prevention

and screening trials. Biometrics 47, 975–86 (1991)
R.J. Serfling, Approximation Theorems of Mathematical Statistics, tome 162

(Wiley, 2009)
A.A. Seyerle, C.L. Avery, Genetic epidemiology: the potential benefits and chal-
lenges of using genetic information to improve human health. N. C. Med. J.
74(6), 505–508 (2013)
P.-S. Shen, An inverse-probability-weighted approach to estimation of the bivari-

ate survival function under left-truncation and right-censoring. J. Stat. Plann.
Infer. 136(12), 4365–4384 (2006)
Y. Shen, J. Cai, Maximum of the weighted Kaplan-Meier tests with application

to cancer prevention and screening trials. Biometrics 57, 837–843 (2001)
Bibliography 467
A. Shenk, Calculus and Analytic Geometry (Santa Monica, 1979)
E.V. Slud, L.V. Rubinstein, Dependent competing risks and summary survival
curves. Biometrika 70(3), 643–649 (1983)
X. Song, C.Y. Wang, Semiparametric approaches for joint modeling of longitudi-

nal and survival data with time-varying coefficients. Biometrics 64(2), 557–566
(2008)
D. Stablein, W. Carter, J. Novax, Analysis of survival data with non-proportional

hazard functions. Control. Clin. Trials 2, 149–159 (1981)
C.A. Struthers, J.D. Kalbfleisch, Misspecified proportional hazards model.

Biometrika 73, 363–369 (1986)
W. Stute, Strong consistency under the koziol-green model. Stat. Prob. Lett.
14(4), 313–320 (1992)
W. Stute, The central limit theorem under random censorship. Ann. Stat. 23,
422–439 (1995)
R.E. Tarone, On the distribution of the maximum of the log-rank statistic and
the modified Wilcoxon statistic. Biometrics 37, 79–85 (1981)
R.E. Tarone, J.H. Ware, On distribution-free tests for equality for survival distri-
butions. Biometrika 64, 156–160 (1977)
T.M. Therneau, P.M. Grambsch, Modeling Survival Data: Extending the Cox
Model (Springer, New York, 2000)
T. Therneau, P. Grambsch, T. Fleming, Martingale-based residuals for survival

models. Biometrika 77(1), 147–160 (1990)
P.K. Trivedi, D.M. Zimmer, Copula Modeling: An Introduction for Practitioners

(Now Publishers Inc., 2007)
W.-Y. Tsai, N.P. Jewell, M.-C. Wang, A note on the product-limit estimator
under right censoring and left truncation. Biometrika 74(4), 883–886 (1987)
A. Tsiatis, A nonidentifiability aspect of the problem of competing risks. Proc.

Natl. Acad. Sci. 72(1), 20–22 (1975)
A. Tsiatis, Competing risks. Encycl. Biostat. 2 (2005)
A.A. Tsiatis, Group sequential methods for survival analysis with staggered entry.
Lect. Notes Monogr. Ser. 2, 257–268 (1982)
A.A. Tsiatis, M. Davidian, Joint modeling of longitudinal and time-to-event data:

an overview. Statistica Sinica 809–834 (2004)
468 Bibliography
C. Van Eeden, The relation between pitman’s asymptotic relative efficiency of

two tests and the correlation coefficient between their test statistics. Ann.
Math. Stat. 34(4), 1442–1451 (1963)
J.W. Vaupel, K.G. Manton, E. Stallard, The impact of heterogeneity in individual

frailty on the dynamics of mortality. Demography 16(3), 439–454 (1979)
D.J. Venzon, S.H. Moolgavkar, Origin-invariant relative risk functions for case-
control and survival studies. Biometrika 75(2), 325–333 (1988)
L. Wei, Testing goodness-of-fit for proportional hazards model with censored

observations. J. Am. Stat. Assoc. 79, 649–652 (1984)
L. Wei, The accelerated failure time model: a useful alternative to the Cox regres-
sion model in survival analysis. Stat. Med. 11, 1871–1879 (1992)
A. Wienke, Frailty Models in Survival Analysis (CRC Press, 2010)
F. Wilcoxon, Individual comparisons by ranking methods. Biometrics Bull. 1,

80–83 (1945)
J.B. Willett, J.D. Singer, Another cautionary note about R 2: its use in weighted
least-squares regression analysis. Am. Stat. 42(3), 236–238 (1988)
J. Wittes, E. Lakatos, J. Probstfield, Surrogate endpoints in clinical trials: car-

diovascular diseases. Stat. Med. 8(4), 415–425 (1989)
L. Wu, P.B. Gilbert, Flexible weighted log-rank tests optimal for detecting early
and/or late survival differences. Biometrics 58, 997–1004 (2002)
L. Xu, C. Gotwalt, Y. Hong, C.B. King, W.Q. Meeker, Applications of the

fractional-random-weight bootstrap. Am. Stat. 1–21 (2020)
J. Xu, J.D. Kalbfleisch, B. Tai, Statistical analysis of illness-death processes and

semicompeting risks data. Biometrics 66(3), 716–725 (2010)
R. Xu, Inference for the proportional hazards model. Thèse de doctorat, University
of California, San Diego (1996)
R. Xu, S. Adak, Survival analysis with time-varying regression effects using a

tree-based approach. Biometrics 58(2), 305–315 (2002)
R. Xu, J. O’Quigley, Proportional hazards estimate of the conditional survival

function. J. R. Stat. Soc. Ser. B (Stat. Methodol.) 62(4), 667–680 (2000)
R. Xu, J. O’Quigley, Estimating average regression effect under non-proportional

hazards. Biostatistics 1(4), 423–439 (2000)
Bibliography 469
X. Xue, M.Y. Kim, M.M. Gaudet, Y. Park, M. Heo, A.R. Hollenbeck, H.D.
Strickler, M.J. Gunter, A comparison of the polytomous logistic regression and
joint cox proportional hazards models for evaluating multiple disease subtypes
in prospective cohort studies. Cancer Epidemiol. Prev. Biomark. 22(2), 275–
285 (2013)
S. Yang, R. Prentice, Semiparametric analysis of short term and long term relative
risks with two sample survival data. Biometrika 92, 1–17 (2005)
S. Yang, R. Prentice, Improved logrank-type tests for survival data using adaptive
weights. Biometrics 66, 30–38 (2010)
M. Yu, J.M.G. Taylor, H.M. Sandler, Individual prediction in prostate cancer

studies using a joint longitudinal survival-cure model. J. Am. Stat. Assoc.
103(481), 178–187 (2008)
M.J. Zhang, X. Zhang, J. Fine, A proportional hazards regression model for the
subdistribution with right-censored and left-truncated competing risks data.
Stat. Med. 30(16), 1933–1951 (2011)
B. Zhou, A. Latouche, V. Rocha, J. Fine, Competing risks regression for stratified

data. Biometrics 67(2), 661–670 (2011)
M. Zhou, G. Li, Empirical likelihood analysis of the Buckley-James estimator. J.

Multivar. Anal. 99(4), 649–664 (2008)
C. Zippin, P. Armitage, Use of concomitant variables and incomplete survival

information in the estimation of an exponential survival parameter. Biometrics
665–672 (1966)
D.M. Zucker, E. Lakatos, Weighted log rank type statistics for comparing survival
curves when there is a time lag in the effectiveness of treatment. Biometrika
77, 853–864 (1990)
Index
A Left censoring; truncation, 8

Approximations, 371 Type I censoring, 35
Cornish-Fisher approximation, 177, Type II censoring, 35
372 Type III censoring, 36
Saddlepoint approximation, 177, 374 Central limit theorems
Arcsine test, 322 Dependent variables, 404
At-risk indicator, 19, 25 Functional central limit theorem,
Repeated events, 26 408
Average value of β(T ), 165 I.I.D. variables, 402
Axioms of probability, 24, 354 Independent variables, 403
Linear sums of ranks, 404
B Stochastic integrals, 396
Brownian bridge Sums of random variables, 405
Covariance process, 381 Change-point models, 286
Brownian motion, 408 Characteristic function, 365
Maximum of process, 383 Clinical trials, 2
Probability results, 385 Coding of groups, 87
Process with drift, 384 Coding variables, 90
Properties of drift process, 384 Competing risks, 38, 391
Brownian motion based tests Conditional logistic regression, 104
Brownian bridge, 312 Conditional survivorship function, 22
Convergence, 355
C Convergence in distribution, 355
Censoring, 34 Functions, 356
Competing risks, 38 Counting processes, 385, 389, 391
Conditionally independent censor- Adapted process, 392
ing, 37, 202 Filtrations, 389
Dependent censoring, 207 Predictability, 391
Finite censoring support, 37 Covariate distribution
Independent censoring, 37 Continuous, 91
Informative censoring, 36, 204 Moments, 179
Koziol-Green model, 36 Cox regression model, 78

https://doi.org/10.1007/978-3-030-33439-0
472 Index
Historical background, 86 G
Cramer-Rao inequality, 414 Glivenko-Cantelli theorem, 57
Cumulative generating function, 365 Goodness of fit, 278
Cumulative hazard transform, 362 Goodness-of-fit tests, 69
Graphical methods, 263
D Greenwood’s formula, 211
DeMoivre-Laplace approximation, 366
Densities, 358 H
Density of a sum, 359 Hazard and related functions, 21
Distributions, 358 Hazard function, 390
Difference of random variables, 359 Helly-Bray theorem, 152, 353, 416
Sums of random variables, 359 High dimensional sparse data, 203
Hypothesis tests, 301
E Area above curve, 320
Empirical distribution function, 353 Area under curve, 315
Empirical estimates Combined tests, 306
Model verification, 69 Concave alternatives, 328
Epidemiology, 97 Convex alternatives, 329
Estimating equations, 141, 413 Delayed effect, 330
Censoring, 170 Diminishing effect, 331
Distribution of solution, 178 Distance traveled, 310
Large sample properties, 169 Integrated log-rank, 318
Marginal survival, 51 Kolmogorov type tests, 312
Maximum likelihood, 50, 417 Log-rank test, 303
Method of moments, 416 Maximum statistics, 307
Minimum chi-square, 415 Non-responders, 332
Misspecified models, 164 Reflected Brownian motion, 313
Moments, 156 Restrictive adaptive, 323
Newton-Raphson iteration, 51 Supremum over cutpoints, 335
Non-proportional hazards, 169 Weighted Kaplan-Meier, 308
Proportional hazards, 153 Weighted log-rank, 304
Regularity conditions, 418
Relative risk models, 159 I
Residuals, 174 Individual risks, 116
Semi-parametric, 151 Integrals
Small samples, 176 Lebesgue, 388
Stratified models, 158 Riemann-Stieltjes, 388
Expectation, 364 Intensity functions, 23
Compartment models, 23
F
Finance and insurance, 5 J
Frailty model, 9 Jensen’s inequality, 364
Freireich data, 53, 147 Joint survival-covariate model, 12
Index 473
K N
Kaplan-Meier estimate, 23, 56, 61, 62 Nelson-Aalen estimate, 68
Continuous version, 62 Non-proportional hazards, 77, 119, 161
Greenwood’s formula, 63 Normal distribution, 358
Mean and Median, 67 Mill’s ratio, 361
Precision, 63
Redistribution to the right, 66 O
Transformations, 63 Observed information, 52
Variance, 63, 67 Order statistics, 366
Distribution of difference, 367
Expected values, 370
L Joint distribution, 367
Large sample theory, 226 Markov property, 368
Law of iterated logarithm, 411 Maximum of sample, 366
Learning and classification, 9 Minimum of sample, 366
Likelihood Normal parent, 370
Conditional, 426
Cox’s conditional likelihood, 429 P
Exponential model, 145 Parametric goodness-of-fit tests, 150
Marginal, 430 Permutation test, 209
Nonparametric exponential, 149 Predictive indices, 12
Parametric models, 143 Interpretation, 14
Partial, 10, 428 Probability integral transform, 181, 361
Linear models Probability that Ti > Tj , 193
Additive, 136 Prognostic biomarkers, 199
Transformation models, 137 Proportional hazards, 78
Log-minus-log transformation, 33 Applications in epidemiology, 98
Logistic regression, 102 Average effect, 172
Changepoint models, 131
Cox model, 79
M
Explained variation, 270
Mantel-Haenszel estimate, 101
Models with intercept, 129
Marginal survival, 49
Partial, 120
Kaplan-Meier estimator, 49
Predictive ability, 270
Martingales, 385, 392
Random effects, frailties, 124
Compensator, 393 Stratified models, 121
Doob-Meyer decomposition, 393 Time-dependent effects, 131
Predictable variation process, 394
Stochastic integrals, 386, 392, 395 R
Mean residual lifetime, 22 Random variables, 354
Mean value theorem, 352 Registry data, 107
Model building, 294 Regression effect process, 215
Multistate models, 25 Concave effects, 288
Multivariate normal distribution, 360 Confidence bands, 267
474 Index
Sample size, 220, 235 Surrogate endpoints, 205

Several covariates, 231 Survival function
Time scale, 219 Covariate information, 192
Regression models, 75, 97, 119 Empirical estimate, 56, 58
Relative efficiency, 334 Estimation, 52, 197
Relative survival, 107, 194 estimation, 145
Resampling methods, 422 Exponential model, 52
Accuracy, 424 Given Z ∈ H, 195
Boostrap confidence intervals, 423 Kaplan-Meier estimate, 58
Bootstrap, 422 Piecewise exponential, 54
Empirical bootstrap, 422 Survival models, 28
Studentized bootstrap, 424 Exponential, 28
Riemann integral, 352 Extreme value, 33
Riemann-Stieltjes integral, 353 Freund model, 207
Rolle’s theorem, 352 Gompertz, 33
Lagakos model, 207
S Log-normal, 34
Saddlepoint approximation, 366 Nonparametric exponential, 70
Stochastic processes, 377 Parametric proportional hazards,
Brownian bridge, 379, 380 34
Brownian motion, 378 Piecewise exponential, 30
Gaussian processes, 378 Proportional hazards exponential,
Independent increments, 377 29
Integrated Brownian motion, 382 Proportional hazards piecewise expo-
Ornstein-Uhlenbeck process, 380 nential, 31
Reflected Brownian motion, 383 Proportional hazards Weibull, 32
Stationarity, 377 Random effects model, 201
Time-transformed Brownian motion, Stratified model, 201
380 Weibull, 31
Sums of random variables, 402
T
Weighted sums, 404
Transformed covariate models, 160
Central limit theorem for depen-
Two stage designs, 206
dent variables, 404
Central limit theorem for i.i.d. vari- X
ables, 402 Xu-O’Quigley estimator, 205
Central limit theorem for indepen-
dent variables, 403 Y
De Moivre-Laplace, 402 Yang and Prentice model, 138
Empirical distribution function, 410
Functional central limit theorem,
405
Lindeberg condition, 403
Sums on (0,1), 405

John O'Quigley - Survival Analysis - Proportional and Non-Proportional Hazards Regression (Springer The Data Sciences) - Springer (2021)

Uploaded by

Copyright:

Available Formats

John O'Quigley - Survival Analysis - Proportional and Non-Proportional Hazards Regression (Springer The Data Sciences) - Springer (2021)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

John O'Quigley - Survival Analysis - Proportional and Non-Proportional Hazards Regression (Springer The Data Sciences) - Springer (2021)

Uploaded by

Copyright:

Available Formats

John O’Quigley

ISBN 978-3-030-33438-3 ISBN 978-3-030-33439-0 (eBook)

© Springer Nature Switzerland AG 2021

London, UK John O’Quigley

4 Proportional hazards models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

8.5 Other relative risk forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201

Appenidx A. Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351

W, WðtÞ Brownian motion on the interval ½0; 1

E bðtÞ ðZjtÞ Expectation with respect to pi ðbðtÞ; tÞ

1.1 Chapter summary

The chapters of this book correspond to a collection of broad themes. Taken

1.2 Context and motivation

Applications of survival analysis methodology appear to be ever-widening.

c Springer Nature Switzerland AG 2021

1.3 Some examples

Novel uses as well as innovatory applications of the basic approach continue

Clinical trials in chronic disease

Applications in epidemiology studies

Genetic susceptibility to disease incidence

Complex time-dependent effects in clinical studies

Insurance, economics, and sociology

Manufacturing and reliability

1.4 Main objectives

1.5 Neglected and underdeveloped topics

The non-proportional hazards model, including as a special case the proportional

It may be that certain parametrizations lead to simpliﬁcation, one example

Left censoring and left truncation

Learning and classification

Kernel methods and local smoothing

using the techniques of ﬁt and predictive performance, gradually weaken certain

view partial likelihood as anything other than a regular likelihood consequent

Joint survival-covariate models

1.6 Model-based prediction

predictive power of any model. This is a direct consequence of the Chebyshev

Interpreting predictive indices

return to this in the chapter covering applications in epidemiology, although not

1.7 Data sets

The importance of working through the calculations on actual data cannot be

1.8 Use as a graduate text

1.9 Classwork and homework

2. In Section 1.5 it is claimed that cumulative quantities may do enough

4. For Exercise 3, describe the shortcomings that we expect to be associated

5. Consider a long-term survival study in which the outcome of interest is sur-

6. How would you describe to a non-scientist the purpose of a statistical

7. Consider a joint model for survival and an individual’s covariate path

8. In the context of joint modeling, ﬁnd a way to make use of information on

Survival analysis methodology

2.1 Chapter summary

We recall some elementary deﬁnitions concerning the probability distributions,

2.2 Context and motivation

c Springer Nature Switzerland AG 2021

viewpoint, this reasoning is incorrect. Elementary transformations ﬁx the skew-

2.3 Basic tools

Time and risk

somewhat straightforward when we consider the conditional distribution of the

Hazard and related functions

λ(t) = f (t)/S(t) , S(t) = exp{−Λ(t)} , f (t) = λ(t) exp{−Λ(t)}.

Although mathematically equivalent, we may prefer to focus attention on one

S(t, u) = Pr(T > t|T > u) = exp{Λ(u) − Λ(t)} , (u < t).