The Frailty Model

Visit the link below to download the full version of this book:

https://medipdf.com/product/the-frailty-model/

Click Download Now

To Pascale, Veerle, and Katelijne (L.D.)

To Mia, Hans, and Marleen (P.J.)

Preface

Survival analysis techniques are used in a variety of disciplines including hu-

man and veterinary medicine, epidemiology, engineering, biology and econ-
omy. Proportional hazards models and accelerated failure time models are
classical models that are frequently used for the analysis of univariate (cen-
sored) survival data.
In recent years a number of papers appeared, extending these models to
models that are suitable to handle more complex survival data. In this context
a lot of attention has been paid to frailty models and copula models. These
models provide a powerful tool to analyse clustered survival data.
In this book the focus is on frailty models, but similarities as well as
differences between frailty models and copula models are highlighted. Frailty
models provide a nice way to capture and to describe the dependence of
observations within a cluster and/or the heterogeneity between clusters. The
specification of a frailty model is rather easy. Frailty models are hazard models
having a multiplicative frailty factor. This factor specifies how frail subjects in
a specific cluster are. The hazard factor in the model can depend on covariates
and can be modelled in a parametric or a semiparametric way. Frailty models
are conditional models. The frailty factor is random and therefore a frailty
distribution needs to be specified in the frailty model. Different choices of
frailty distributions are studied in detail in this book; attention is paid to the
relation between the choice of the distribution and the type of dependence it
generates between observations within a cluster.
Once the frailty model is specified we need to fit the model. This is often
not an easy task and in many cases standard software to fit such models is
missing. For particular choices of the frailty distribution, the frailties can be
integrated out with respect to the frailty distribution from the conditional
likelihood resulting in the marginal likelihood. For parametric frailty mod-
els, maximisation of the marginal likelihood then leads to estimates of the
parameters in the model. However, for semiparametric frailty models, more
complex estimation techniques are needed. These techniques include the EM
algorithm, penalised likelihood techniques, and Gibbs sampling in a Bayesian

vii
viii Preface

context. This book does not give an exhaustive overview of all models and
techniques that are available for the analysis of clustered survival data. We
rather aim at an in-depth discussion of the most current techniques and we
want to demonstrate, based on real data sets, how results obtained from the
statistical analysis should be interpreted. By doing so, we hope to bring these
important tools closer to the applied statisticians.
In the two final chapters of the book we consider more complex frailty
models including hierarchical models with more than one level of clustering.
The data sets used in the book are taken from the different disciplines in
which the authors worked during the last ten years. Three data sets are pro-
vided by the European Organisation for Research and Treatment of Cancer
(EORTC, Brussels, Belgium): the periop data set (Example 1.5), the DCIS
data set (Example 1.6), and the bladder cancer data set (Example 1.10). An-
other study in human medicine is a large asthma prevention trial that was run
by Union Chimique Belge (UCB) to assess the effect of a new drug (Example
1.9). An infant mortality data set from Jimma (Ethiopia) resulted from a large
longitudinal study run by the Public Health investigators at Jimma University
(Example 1.11). The other data sets are from veterinary sciences. The first
study (Example 1.1) deals with the epidemiology of East Coast Fever, an im-
portant disease of cattle in eastern and southern Africa. The data are provided
by the Institute of Tropical Medicine (ITM, Antwerp, Belgium). The Depart-
ment of Medical Imaging of Domestic Animals (Ghent University, Belgium)
has provided data on time to diagnosis of fracture healing in dogs (Example
1.2). We further include two data sets from the Department of Physiology
and Biometrics (Ghent University, Belgium) on mastitis and udder infection
(Examples 1.3 and 1.4). Finally the Department of Reproduction, Obstetrics
and Herd Health (Ghent University, Belgium) has kindly provided a culling
data set (Example 1.7) and an insemination data set (Example 1.8).
This book is intended for students and for applied statisticians. It can be
used as a graduate course for students who have had a course on survival
analysis and a course on statistical inference. Applied statisticians will find
a lot of examples in the book, which will help to get a better understanding
of the theoretical ideas. The code of the programs used in the examples of
the book is freely available from the Springer Verlag website. This gives the
readers the opportunity to try the examples themselves and to adapt the code
or write new code for their particular problems.
Selected parts of the book have been used as teaching material for a series
of lectures at the Biometrics and Clinical Informatics Department of Johnson
& Johnson Pharmaceutical Research and Development in Beerse, Belgium
(October-November 2005) and for workshops in Nairobi (Kenya, 2004) and
Addis Ababa (Ethiopia, 2005), sponsored by the Flemish Interuniversitary
Council - University Development Cooperation. We also gratefully acknowl-
edge the financial support from the IAP research network P6/03 of the Belgian
Government (Belgian Science Policy).
 Preface ix

We thank all those who provided in one way or another support to

this book project: Bart Ampe, Sarne De Vliegher, Els Goetghebeur, Klara
Goethals, Hans Laevens, Geert Opsomer, and Marije Risselada (Ghent Univer-
sity, Belgium), Tomasz Burzykowski, Goele Massonnet, and Noël Veraverbeke
(Hasselt University, Belgium), Arnost Komárek (Catholic University Leuven,
Belgium), Catherine Legrand and Ingrid Van Keilegom (Université Catholique
de Louvain, Belgium), Dirk Berkvens (ITM, Belgium), Richard Sylvester
(EORTC, Belgium), Catherine Fortpied (UCB, Belgium), Vincent Ducrocq
(INRA, Jouy-en-Josas, France), Andreas Wienke (Martin-Luther-Universität
Halle-Wittenberg, Germany), Samuli Ripatti (Karolinska Institute, Sweden),
Virginie Rondeau (INSERM, Bordeaux, France), Elisa Molanes (University of
La Coruña, Spain), Mekonnen Assefa and Fasil Tessema (Jimma University,
Ethiopia), and Daan de Waal (University of the Free State, South Africa).
A word of special thanks from Paul Janssen goes to his next-door colleague
for many years, Noël Veraverbeke, for many discussions on different issues in
survival analysis and in statistics in general.
Luc Duchateau would like to thank Herman Ramon, Paul Darius, Bruno
Goddeeris, Jef Vercruysse, and Christian Burvenich for their stimulating sup-
port in his different research undertakings.

Luc Duchateau
Paul Janssen
September 2007
Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Glossary of Definitions and Notation . . . . . . . . . . . . . . . . . . . . . . . . . . xv

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Survival analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.4.1 Survival likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.2 Proportional hazards models . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4.3 Accelerated failure time models . . . . . . . . . . . . . . . . . . . . . 26
1.4.4 The loglinear model representation . . . . . . . . . . . . . . . . . . 30
1.5 Semantics and history of the term frailty . . . . . . . . . . . . . . . . . . . 32

2 Parametric proportional hazards models with gamma

frailty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.1 The parametric proportional hazards model with frailty term . 44
2.2 Maximising the marginal likelihood: the frequentist approach . 45
2.3 Extension of the marginal likelihood approach to
interval-censored data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.4 Posterior densities: the Bayesian approach . . . . . . . . . . . . . . . . . . 65
2.4.1 The Metropolis algorithm in practice for the
parametric gamma frailty model . . . . . . . . . . . . . . . . . . . . . 65
2.4.2 ∗ Theoretical foundations of the Metropolis algorithm . . . 74
2.5 Further extensions and references . . . . . . . . . . . . . . . . . . . . . . . . . . 75

3 Alternatives for the frailty model . . . . . . . . . . . . . . . . . . . . . . . . . . 77

3.1 The fixed effects model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.1.1 The model specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

xi
xii Contents

3.1.2 ∗ Asymptotic efficiency of fixed effects model parameter

estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.2 The stratified model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.3 The copula model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.3.1 Notation and definitions for the conditional, joint, and
population survival functions . . . . . . . . . . . . . . . . . . . . . . . 93
3.3.2 Definition of the copula model . . . . . . . . . . . . . . . . . . . . . . 95
3.3.3 The Clayton copula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.3.4 The Clayton copula versus the gamma frailty model . . . 99
3.4 The marginal model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
3.4.1 Defining the marginal model . . . . . . . . . . . . . . . . . . . . . . . . 104
3.4.2 ∗ Consistency of parameter estimates from marginal
model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
3.4.3 Variance of parameter estimates adjusted for
correlation structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.5 Population hazards from conditional models . . . . . . . . . . . . . . . . 111
3.5.1 Population versus conditional hazard from frailty
models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.5.2 Population versus conditional hazard ratio from frailty
models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.6 Further extensions and references . . . . . . . . . . . . . . . . . . . . . . . . . . 116

4 Frailty distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

4.1 General characteristics of frailty distributions . . . . . . . . . . . . . . . 118
4.1.1 Joint survival function and the Laplace transform . . . . . 119
4.1.2 Population survival function and the copula . . . . . . . . . . 120
4.1.3 Conditional frailty density changes over time . . . . . . . . . . 122
4.1.4 Measures of dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.2 The gamma distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.2.1 Definitions and basic properties . . . . . . . . . . . . . . . . . . . . . 130
4.2.2 Joint and population survival function . . . . . . . . . . . . . . . 131
4.2.3 Updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.2.4 Copula form representation . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.2.5 Dependence measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
4.2.6 Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.2.7 ∗ Estimation of the cross ratio function: some theoretical
considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
4.3 The inverse Gaussian distribution . . . . . . . . . . . . . . . . . . . . . . . . . 150
4.3.1 Definitions and basic properties . . . . . . . . . . . . . . . . . . . . . 150
4.3.2 Joint and population survival function . . . . . . . . . . . . . . . 152
4.3.3 Updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
4.3.4 Copula form representation . . . . . . . . . . . . . . . . . . . . . . . . . 158
4.3.5 Dependence measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
4.3.6 Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
 Contents xiii

4.4 The positive stable distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

4.4.1 Definitions and basic properties . . . . . . . . . . . . . . . . . . . . . 164
4.4.2 Joint and population survival function . . . . . . . . . . . . . . . 167
4.4.3 Updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
4.4.4 Copula form representation . . . . . . . . . . . . . . . . . . . . . . . . . 171
4.4.5 Dependence measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
4.4.6 Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
4.5 The power variance function distribution . . . . . . . . . . . . . . . . . . . 177
4.5.1 Definitions and basic properties . . . . . . . . . . . . . . . . . . . . . 177
4.5.2 Joint and population survival function . . . . . . . . . . . . . . . 181
4.5.3 Updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
4.5.4 Copula form representation . . . . . . . . . . . . . . . . . . . . . . . . . 185
4.5.5 Dependence measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
4.5.6 Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
4.6 The compound Poisson distribution . . . . . . . . . . . . . . . . . . . . . . . . 190
4.6.1 Definitions and basic properties . . . . . . . . . . . . . . . . . . . . . 190
4.6.2 Joint and population survival functions . . . . . . . . . . . . . . 192
4.6.3 Updating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
4.7 The lognormal distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
4.8 Further extensions and references . . . . . . . . . . . . . . . . . . . . . . . . . . 196

5 The semiparametric frailty model . . . . . . . . . . . . . . . . . . . . . . . . . . 199

5.1 The EM algorithm approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
5.1.1 Description of the EM algorithm . . . . . . . . . . . . . . . . . . . . 199
5.1.2 Expectation and maximisation for the gamma frailty
model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
5.1.3 Why the EM algorithm works for the gamma frailty
model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
5.2 The penalised partial likelihood approach . . . . . . . . . . . . . . . . . . . 210
5.2.1 The penalised partial likelihood for the normal
random effects density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
5.2.2 The penalised partial likelihood for the gamma frailty
distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
5.2.3 Performance of the penalised partial likelihood
estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
5.2.4 Robustness of the frailty distribution assumption . . . . . . 228
5.3 Bayesian analysis for the semiparametric gamma frailty
model through Gibbs sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
5.3.1 The frailty model with a gamma process prior for the
cumulative baseline hazard for grouped data . . . . . . . . . . 234
5.3.2 The frailty model with a gamma process prior for the
cumulative baseline hazard for observed event times . . . 239
5.3.3 The normal frailty model based on Poisson likelihood . . 244
5.3.4 Sampling techniques used for semiparametric frailty
models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
xiv Contents

5.3.5 Gibbs sampling, a special case of the Metropolis–

Hastings algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
5.4 Further extensions and references . . . . . . . . . . . . . . . . . . . . . . . . . . 258

6 Multifrailty and multilevel models . . . . . . . . . . . . . . . . . . . . . . . . . 259

6.1 Multifrailty models with one clustering level . . . . . . . . . . . . . . . . 260
6.1.1 Bayesian analysis based on Laplacian integration . . . . . . 260
6.1.2 Frequentist approach using Laplacian integration . . . . . . 268
6.2 Multilevel frailty models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
6.2.1 Maximising the marginal likelihood with penalised
splines for the baseline hazard . . . . . . . . . . . . . . . . . . . . . . 277
6.2.2 The Bayesian approach for multilevel frailty models
using Gibbs sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
6.3 Further extensions and references . . . . . . . . . . . . . . . . . . . . . . . . . . 286

7 Extensions of the frailty model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287

7.1 Censoring and truncation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
7.2 Correlated frailty models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
7.3 Joint modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
7.4 The accelerated failure time model . . . . . . . . . . . . . . . . . . . . . . . . . 292

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295

Applications and Examples Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309

Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314

Glossary of Definitions and Notation

Definitions are given for the case of the shared frailty model with one cluster-
ing level, with the first index i referring to the cluster and the second index j
referring to the j th observation within that cluster. For univariate frailty mod-
els or models without clustering, the first index i refers to the ith observation.
For hierarchical models with two clustering levels, the first index i refers to
the highest clustering level, the second index j to the second clustering level
and the third index k to the k th observation within cluster j which is nested
within cluster i.

The indexing for the conditional, joint and population density and survival
functions can be deduced from the corresponding hazard functions.

β the fixed effects vector (of dimension p)

Γ (.) the gamma function
δij censoring indicator with δij = 1 if the event time
is observed, otherwise zero
ζ (ξ, θ, β)
ζ (t1 , t2 ) the cross ratio function
θ the variance of a frailty density with mean one
(except in the case of the positive stable frailty distribution)
ξ vector containing the parameters of the baseline hazard h0 (t)
τ Kendall’s tau
Φ the acceleration factor
in the accelerated failure time model
φ (t1 , t2 ) the ratio of the joint density function and the product
of the two population density functions
ψ (t1 , t2 ) the ratio of the joint survival function
and the product of the two population survival functions

xv
xvi Glossary of Definitions and Notation

cij the censoring time

Cθ (.) the copula function with parameter vector θ
d the total number of events
di the number of events in cluster i
h0 (t) the baseline hazard function
h0i (t) the baseline hazard function for cluster i
hij (t) the conditional hazard function
hij,c (t) the conditional hazard function for observation j
from cluster i with ui = 1
hi (tni ) the joint conditional hazard function for cluster i
hÜ,f (tn ) the joint hazard function for clusters of size n
with covariate information Ü based on the frailty model
hÜ,p (tn ) the joint hazard function for clusters of size n
with covariate information Ü based on the marginal model
hx,f (t) the population hazard function for subjects
with covariate information x based on the frailty model
hx,p (t) the population hazard function for subjects
with covariate information x based on the marginal model
H0 (t) the cumulative baseline hazard
H the Hessian matrix
I the Fisher information matrix
I the observed information matrix
I(condition) the indicator function equal to one if the condition is true,
zero otherwise
lij the lowerbound of the interval that contains the event time
l(.) the loglikelihood function
L(.) the likelihood function
Lmarg (.) the marginal likelihood function
Lpart (.) the partial likelihood function
L(.) the Laplace transform
ni the number of observations in cluster i
p the dimension of β
r the total number of distinct event times
rij the upperbound of the interval that contains the event time
R (yij ) the risk set for observation yij
including all subjects at risk at time yij
Ri (yij ) the risk set for observation yij
including only subjects from cluster i at risk at time yij
s the number of clusters
S the vector of scores
tij the event time
ui the frailty term for cluster i
wi the random effect for cluster i
W (λ, ρ) the Weibull density with scale parameter λ
and shape parameter ρ
 Glossary of Definitions and Notation xvii

xij
 vector containing covariate information

xij (tij1 ) , . . . , xij tijkij time-varying covariate vectors with kij timepoints
Ü (xt1 , . . . , xtn )

t
t
t
Üi t
xi1 , . . . , xini
yij the minimum of the event time and censoring time
y(1) , . . . , y(r) the ordered event times
zij (yij , δij )
1
Introduction

1.1 Goals
Survival analysis is an important statistical field that is required for data
analysis in different disciplines. There are a large number of textbooks on
the basic concepts of survival analysis, e.g., Collett (2003) and Machin et al.
(2006). In recent years a number of papers have been published extending the
classical survival models to models that are suitable for the analysis of clus-
tered survival data: frailty models. A wide variety of frailty models and several
numerical techniques to fit these models have been studied in these papers.
Only a limited number of textbooks have chapters devoted to frailty mod-
els, all of them containing only some specific aspects on frailty models. Klein
and Moeschberger (1997) have a chapter on the use of the EM algorithm in
semiparametric frailty models. Ibrahim et al. (2001) discuss both parametric
and semiparametric frailty models but only in a Bayesian context. Hougaard
(2000) discusses, among other techniques to model multivariate survival data,
the frailty model mainly assuming a parametric model. Finally, Therneau and
Grambsch (2000) discuss inference for the semiparametric frailty model using
penalised partial likelihood.
The main objective of this book is to study frailty models and the nu-
merical techniques, needed to fit such models, in a unified and detailed way.
Both proportional hazards models and accelerated failure time models are
discussed; we consider parametric as well as semiparametric modelling. We fit
models using both a frequentist and a Bayesian approach. Furthermore, dif-
ferent density functions for the frailty term are studied as each of them puts a
different dependence structure on the data. We also discuss hierarchical mod-
els, i.e., models with more than one level of clustering. We do not attempt to
give an exhaustive overview of the literature on frailty models as it is far too
ambitious to do so; we rather prefer to explain the basic ideas and statistical
techniques in detail. Some sections of the book are marked with an asterisk
to indicate that they deal with more theoretical aspects. These sections can
be skipped without losing the flow of the book.

1
2 1 Introduction

A very important further objective is to make the frailty model techniques

available to a wider audience. In spite of the amount of published material,
the bottleneck is the lack of available public software. We will mainly use R, a
freeware package, to fit the different models, and all the programs used are also
freely available from the Springer Verlag website. Also software available in
STATA and SAS is used to fit frailty models. WinBUGS, a freeware package,
is used to perform most of the Bayesian analyses (Spiegelhalter et al., 2003).

1.2 Outline

In this first chapter, we introduce different data sets with time to event out-
comes that will be used throughout the book to demonstrate the frailty model
fitting techniques and the practical interpretation of the results. The introduc-
tory chapter further describes and explains some basic and general concepts in
survival analysis to lay the foundation for the more advanced frailty models.
In Chapter 2, the simplest frailty model is introduced: the parametric
model with clusters sharing the same gamma distributed frailty term. The
simplicity is due to the fact that the frailties appearing in the conditional
likelihood can be integrated out in a simple way to obtain the marginal like-
lihood that can then be maximised to obtain parameter estimates. Both the
frequentist and the Bayesian approach are considered.
In Chapter 3, we consider models that, although different from frailty
models, take the clustering of the data into account: the fixed effects model,
the stratified model, the copula model, and the marginal model with robust
variance estimators. We also demonstrate the relation between conditional
and marginal models: given the hazard function and the hazard ratio in the
conditional model and given the frailty distribution, the hazard function and
the hazard ratio in the population can easily be obtained. These functions are
important to understand how the population evolves over time.
In Chapter 4, different distributions for the frailty term are discussed and
the type of dependence that they induce on the event times in the cluster
is studied. The gamma distribution is most often used in practice. It be-
longs to the power variance function family proposed by Hougaard (1986a);
other members of this family are also discussed: the inverse Gaussian and the
positive stable distribution. The compound Poisson distribution is a further
extension of the power variance function family. Characteristics of these distri-
butions follow easily from their corresponding Laplace transforms. In practice
also the lognormal distribution is often used for the frailty term. The use of
the lognormal distribution mainly originates from work by McGilchrist and
Aisbett (1991). They look at survival data from a mixed models viewpoint
and therefore assume a normal distribution for the random effects which ap-
pear in the model formulation of the loghazard function. A drawback of the
lognormal frailty distribution is that the Laplace transform does not take a
 1.3 Examples 3

simple form and hence the dependence imposed by the lognormal distribution
is difficult to evaluate.
In Chapter 5, the semiparametric model is studied. Different numerical
techniques have been proposed to fit semiparametric models, such as the
EM algorithm (Klein, 1992), the penalised partial likelihood maximisation
(Therneau et al., 2003), in a Bayesian context Laplacian integration (Ducrocq
and Sölkner, 1994) or the MCMC algorithm (Clayton, 1991). The different
techniques are studied in detail and the specific features of the different ap-
proaches will be compared.
Chapter 6 deals with frailty models with more than one frailty term. The
chapter has two parts. First we discuss the case of two frailty terms within
the same cluster, which occurs for instance if the objective in a multicentre
trial consists of both looking at the heterogeneity between centres but also
at the heterogeneity of the treatment effect between centres (Legrand et al.,
2005). In the second part, nested or hierarchical models are discussed, with
one clustering level nested within another clustering level.
In Chapter 7, further extensions of frailty models are discussed: frailty
models for data with different censoring and truncation characteristics, cor-
related frailty models, joint modelling, and the accelerated failure time frailty
model.

1.3 Examples

Different techniques to fit frailty models for clustered survival data are dis-
cussed in this book. All these techniques are applied to real data sets. We made
a selection of examples so that we can demonstrate in later chapters how spe-
cific aspects of the data determine the appropriate inferential methodology
used for the statistical analysis. A few important aspects that we will men-
tion at this stage are (i) the cluster size, e.g., bivariate clusters versus clusters
with large cluster sizes; (ii) the nesting or the hierarchy present in the design
of the experiment; (iii) presence or absence of event ordering within a cluster,
e.g., ordering is present in recurrent event data sets. Most of the examples
do not have such ordering. If there exists an ordering, either in space or in
time, we can still use techniques that do not take the ordering into account,
although more specific models might be more relevant. We will therefore men-
tion, whenever needed, that events in a cluster are ordered.
It is not our objective to give the complete analysis of the data sets; the
data sets are rather used to illustrate how the different inferential techniques
can be applied in practice.
We first consider data sets with one level of clustering. The simplest such
data set consists of clusters of size one. Such a data set is given in Example
1.1. The models fitted to these data are called univariate frailty models. The
simplest extension of this type of data set is to enlarge the cluster from one unit
to two units. This type of data, called bivariate survival data, has received