Free ebooks ==> www.Ebook777.

com

www.Ebook777.com
 Free ebooks ==> www.Ebook777.com

Frailty Models in
Survival Analysis

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC

C7388_FM.indd 1 6/30/10 1:10:24 PM

 Free ebooks ==> www.Ebook777.com

Editor-in-Chief

Shein-Chung Chow, Ph.D.

Professor
Department of Biostatistics and Bioinformatics
Duke University School of Medicine
Durham, North Carolina, U.S.A.

Series Editors

Byron Jones Jen-pei Liu

Senior Director Professor
Statistical Research and Consulting Centre Division of Biometry
(IPC 193) Department of Agronomy
Pfizer Global Research and Development National Taiwan University
Sandwich, Kent, U. K. Taipei, Taiwan

Karl E. Peace Bruce W. Turnbull

Georgia Cancer Coalition Professor
Distinguished Cancer Scholar School of Operations Research
Senior Research Scientist and and Industrial Engineering
Professor of Biostatistics Cornell University
Jiann-Ping Hsu College of Public Health Ithaca, New York
Georgia Southern University
Statesboro, Georgia

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC

C7388_FM.indd 2 6/30/10 1:10:25 PM

 Free ebooks ==> www.Ebook777.com

Published Titles

1. Design and Analysis of Animal Studies in 19. Statistical Design and Analysis of Stability
Pharmaceutical Development, Studies, Shein-Chung Chow
Shein-Chung Chow and Jen-pei Liu 20. Sample Size Calculations in Clinical Research,
2. Basic Statistics and Pharmaceutical Statistical Second Edition, Shein-Chung Chow,
Applications, James E. De Muth Jun Shao, and Hansheng Wang
3. Design and Analysis of Bioavailability and 21. Elementary Bayesian Biostatistics,
Bioequivalence Studies, Second Edition, Revised Lemuel A. Moyé
and Expanded, Shein-Chung Chow and 22. Adaptive Design Theory and Implementation
Jen-pei Liu Using SAS and R, Mark Chang
4. Meta-Analysis in Medicine and Health Policy, 23. Computational Pharmacokinetics, Anders Källén
Dalene K. Stangl and Donald A. Berry 24. Computational Methods in Biomedical Research,
5. Generalized Linear Models: A Bayesian Ravindra Khattree and Dayanand N. Naik
Perspective, Dipak K. Dey, Sujit K. Ghosh, 25. Medical Biostatistics, Second Edition,
and Bani K. Mallick A. Indrayan
6. Difference Equations with Public Health 26. DNA Methylation Microarrays: Experimental
Applications, Lemuel A. Moyé and Design and Statistical Analysis,
Asha Seth Kapadia Sun-Chong Wang and Arturas Petronis
7. Medical Biostatistics, Abhaya Indrayan and 27. Design and Analysis of Bioavailability and
Sanjeev B. Sarmukaddam Bioequivalence Studies, Third Edition,
8. Statistical Methods for Clinical Trials, Shein-Chung Chow and Jen-pei Liu
Mark X. Norleans 28. Translational Medicine: Strategies and
9. Causal Analysis in Biomedicine and Statistical Methods, Dennis Cosmatos and
Epidemiology: Based on Minimal Sufficient Shein-Chung Chow
Causation, Mikel Aickin 29. Bayesian Methods for Measures of Agreement,
10. Statistics in Drug Research: Methodologies and Lyle D. Broemeling
Recent Developments, Shein-Chung Chow 30. Data and Safety Monitoring Committees in
and Jun Shao Clinical Trials, Jay Herson
11. Sample Size Calculations in Clinical Research, 31. Design and Analysis of Clinical Trials with Time-
Shein-Chung Chow, Jun Shao, and to-Event Endpoints, Karl E. Peace
Hansheng Wang 32. Bayesian Missing Data Problems: EM, Data
12. Applied Statistical Design for the Researcher, Augmentation and Noniterative Computation,
Daryl S. Paulson Ming T. Tan, Guo-Liang Tian, and Kai Wang Ng
13. Advances in Clinical Trial Biostatistics, 33. Multiple Testing Problems in Pharmaceutical
Nancy L. Geller Statistics, Alex Dmitrienko, Ajit C. Tamhane,
14. Statistics in the Pharmaceutical Industry, and Frank Bretz
Third Edition, Ralph Buncher and Jia-Yeong Tsay 34. Bayesian Modeling in Bioinformatics,
15. DNA Microarrays and Related Genomics Dipak K. Dey, Samiran Ghosh, and
Techniques: Design, Analysis, and Interpretation Bani K. Mallick
of Experiments, David B. Allsion, Grier P. Page, 35. Clinical Trial Methodology, Karl E. Peace
T. Mark Beasley, and Jode W. Edwards and Ding-Geng (Din) Chen
16. Basic Statistics and Pharmaceutical Statistical 36. Monte Carlo Simulation for the Pharmaceutical
Applications, Second Edition, James E. De Muth Industry: Concepts, Algorithms, and Case
17. Adaptive Design Methods in Clinical Trials, Studies, Mark Chang
Shein-Chung Chow and Mark Chang 37. Frailty Models in Survival Analysis,
18. Handbook of Regression and Modeling: Andreas Wienke
Applications for the Clinical and Pharmaceutical
Industries, Daryl S. Paulson

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC

C7388_FM.indd 3 6/30/10 1:10:25 PM

Free ebooks ==> www.Ebook777.com

www.Ebook777.com
 Free ebooks ==> www.Ebook777.com

Frailty Models in
Survival Analysis

Andreas Wienke
Institute of Medical Epidemiology, Biostatistics, and Informatics
Martin-Luther-University Halle-Wittenberg, Germany

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC

C7388_FM.indd 5 6/30/10 1:10:25 PM

 Free ebooks ==> www.Ebook777.com

Chapman & Hall/CRC

Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2011 by Taylor and Francis Group, LLC

Chapman & Hall/CRC is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed in the United States of America on acid-free paper

10 9 8 7 6 5 4 3 2 1

International Standard Book Number: 978-1-4200-7388-1 (Hardback)

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

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

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

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

Library of Congress Cataloging‑in‑Publication Data

Wienke, Andreas.
Frailty models in survival analysis / author, Andreas Wienke.
p. cm. --  (Chapman & Hall/CRC biostatistics series)
“A CRC title.”
Includes bibliographical references and index.
ISBN 978-1-4200-7388-1 (hardcover : alk. paper)
1.  Failure time data analysis--Mathematics. 2.  Survival analysis
(Biometry)--Mathematics. 3.  Mortality--Mathematical models. 4. 
Demography--Mathematics.  I. Title. II. Series.

QA280.W54 2011
519.5’46--dc22 2010021869

Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com

and the CRC Press Web site at

http://www.crcpress.com

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC

C7388_FM.indd 6 6/30/10 1:10:26 PM

 Free ebooks ==> www.Ebook777.com

All models are wrong, some are useful. (Box 1976)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
Free ebooks ==> www.Ebook777.com

www.Ebook777.com
 Free ebooks ==> www.Ebook777.com

Contents

List of Tables xii

List of Figures xv

Preface xix

1 Introduction 1
1.1 Goals and Outline . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Survival Analysis 15
2.1 Basic Concepts in Survival Analysis . . . . . . . . . . . . . . 15
2.2 Censoring and Truncation . . . . . . . . . . . . . . . . . . . 19
2.3 Parametric Models . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.1 Exponential distribution . . . . . . . . . . . . . . . . . 28
2.3.2 Weibull distribution . . . . . . . . . . . . . . . . . . . 30
2.3.3 Log-logistic distribution . . . . . . . . . . . . . . . . . 32
2.3.4 Gompertz distribution . . . . . . . . . . . . . . . . . . 33
2.3.5 Log-normal distribution . . . . . . . . . . . . . . . . . 34
2.3.6 Gamma distribution . . . . . . . . . . . . . . . . . . . 36
2.3.7 Pareto distribution . . . . . . . . . . . . . . . . . . . . 37
2.4 Estimation of Survival and Hazard Functions . . . . . . . . . 38
2.4.1 Kaplan–Meier estimator . . . . . . . . . . . . . . . . . 38
2.4.2 Nelson–Aalen estimator . . . . . . . . . . . . . . . . . 41
2.5 Regression Models . . . . . . . . . . . . . . . . . . . . . . . . 43
2.5.1 Proportional hazards model . . . . . . . . . . . . . . 43
2.5.2 Accelerated failure time model . . . . . . . . . . . . . 50
2.6 Identifiability Problems . . . . . . . . . . . . . . . . . . . . . 52

3 Univariate Frailty Models 55

3.1 The Concept of Univariate Frailty . . . . . . . . . . . . . . . 57
3.2 Discrete Frailty Model . . . . . . . . . . . . . . . . . . . . . 65
3.3 Gamma Frailty Model . . . . . . . . . . . . . . . . . . . . . 72
3.3.1 Parametric gamma frailty model . . . . . . . . . . . . 79
3.3.2 Semiparametric gamma frailty model . . . . . . . . . 84
3.3.3 Gamma frailty model for current status data . . . . . 90
3.3.4 Extensions of the gamma frailty model . . . . . . . . . 92

ix
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
x

3.4 Log-normal Frailty Model . . . . . . . . . . . . . . . . . . . 96

3.4.1 Parametric log-normal frailty model . . . . . . . . . . 98
3.4.2 Semiparametric log-normal frailty model . . . . . . . . 99
3.5 Inverse Gaussian Frailty Model . . . . . . . . . . . . . . . . 101
3.6 Positive Stable Frailty Model . . . . . . . . . . . . . . . . . 105
3.7 PVF Frailty Model . . . . . . . . . . . . . . . . . . . . . . . 108
3.8 Compound Poisson Frailty Model . . . . . . . . . . . . . . . 111
3.9 Quadratic Hazard Frailty Model . . . . . . . . . . . . . . . . 116
3.10 Lévy Frailty Models . . . . . . . . . . . . . . . . . . . . . . 120
3.11 Log-t Frailty Model . . . . . . . . . . . . . . . . . . . . . . . 121
3.12 Univariate Frailty Cure Models . . . . . . . . . . . . . . . . 122
3.13 Missing Covariates in Proportional Hazards Models . . . . . 127

4 Shared Frailty Models 131

4.1 Marginal versus Frailty Model . . . . . . . . . . . . . . . . . 132
4.2 The Concept of Shared Frailty . . . . . . . . . . . . . . . . . 135
4.3 Shared Gamma Frailty Model . . . . . . . . . . . . . . . . . 139
4.3.1 Parametric shared gamma frailty model . . . . . . . . 140
4.3.2 Semiparametric shared gamma frailty model . . . . . 142
4.3.3 Shared gamma frailty model for current status data . 145
4.4 Shared Log-normal Frailty Model . . . . . . . . . . . . . . . 148
4.5 Shared Positive Stable Frailty Model . . . . . . . . . . . . . 150
4.6 Shared Compound Poisson/PVF Frailty Model . . . . . . . . 151
4.7 Shared Frailty Models More General . . . . . . . . . . . . . . 152
4.8 Dependence measures . . . . . . . . . . . . . . . . . . . . . . 153
4.9 Limitations of the Shared Frailty Model . . . . . . . . . . . 158

5 Correlated Frailty Models 161

5.1 The Concept of Correlated Frailty . . . . . . . . . . . . . . . 163
5.2 Correlated Gamma Frailty Model . . . . . . . . . . . . . . . 165
5.3 Correlated Log-normal Frailty Model . . . . . . . . . . . . . 177
5.4 MCMC Methods for the Correlated Log-normal Frailty Model 181
5.5 Correlated Compound Poisson Frailty Model . . . . . . . . . 185
5.6 Correlated Quadratic Hazard Frailty Model . . . . . . . . . . 189
5.7 Other Correlated Frailty Models . . . . . . . . . . . . . . . . 193
5.8 Bivariate Frailty Cure Models . . . . . . . . . . . . . . . . . 195
5.9 Comparison of Different Estimation Strategies . . . . . . . . 198
5.10 Dependent Competing Risks in Frailty Models . . . . . . . . 204

6 Copula Models 209

6.1 Shared Gamma Frailty Copula . . . . . . . . . . . . . . . . 210
6.2 Correlated Gamma Frailty Copula . . . . . . . . . . . . . . 212
6.3 General Correlated Frailty Copulas . . . . . . . . . . . . . . 215
6.4 Cross-Ratio Function . . . . . . . . . . . . . . . . . . . . . . 220

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
xi

7 Different Aspects of Frailty Modeling 223

7.1 Dependence and Interaction between Frailty and Observed
Covariates . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
7.2 Cox Model with General Gaussian Random Effects . . . . . 226
7.3 Nested Frailty Models . . . . . . . . . . . . . . . . . . . . . . 227
7.4 Recurrent Event Time Data . . . . . . . . . . . . . . . . . . 228
7.5 Tests for Heterogeneity . . . . . . . . . . . . . . . . . . . . . 230
7.6 Log-Rank Test in Frailty Models . . . . . . . . . . . . . . . . 231
7.7 Time-Dependent Frailty Models . . . . . . . . . . . . . . . . 232
7.8 Identifiability of Frailty Models . . . . . . . . . . . . . . . . 235
7.8.1 Univariate frailty models . . . . . . . . . . . . . . . . 235
7.8.2 Multivariate frailty models . . . . . . . . . . . . . . . 237
7.9 Applications of Frailty Models . . . . . . . . . . . . . . . . . 238
7.10 Software for Frailty Models . . . . . . . . . . . . . . . . . . . 240
7.10.1 R packages . . . . . . . . . . . . . . . . . . . . . . . . 241
7.10.2 SAS packages . . . . . . . . . . . . . . . . . . . . . . . 242
7.10.3 STATA package . . . . . . . . . . . . . . . . . . . . . . 242

A Appendix 243
A.1 Bivariate Lifetime Models . . . . . . . . . . . . . . . . . . . . 243
A.2 Correlated Gamma Frailty Model . . . . . . . . . . . . . . . 245
A.3 Correlated Compound Poisson Frailty Model . . . . . . . . 247
A.4 Correlated Quadratic Hazard Frailty Model . . . . . . . . . 249
A.5 Dependent Competing Risks Model . . . . . . . . . . . . . . 253
A.6 Quantitative Genetics . . . . . . . . . . . . . . . . . . . . . 261

References 265

Index 299

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
Free ebooks ==> www.Ebook777.com

www.Ebook777.com
 Free ebooks ==> www.Ebook777.com

List of Tables

1.1 Data of five patients in the EBCT study . . . . . . . . . . . . 4

1.2 Description of the EBCT study population . . . . . . . . . . 4
1.3 Data of five patients of the malignant melanoma study . . . . 5
1.4 Data of five patients in the Halluca study . . . . . . . . . . . 6
1.5 Description of the Halluca study population . . . . . . . . . . 7
1.6 Data of three Danish twin pairs . . . . . . . . . . . . . . . . . 8
1.7 Causes of death in the Danish twin population . . . . . . . . 9
1.8 Cause of death groups by ICD number . . . . . . . . . . . . . 10
1.9 Data of three Danish twin pairs with covariates . . . . . . . . 11
1.10 Study population by sex, BMI, and smoking . . . . . . . . . . 11
1.11 Data of three Swedish twin pairs . . . . . . . . . . . . . . . . 12
1.12 Breast cancer in Swedish twins (old cohort) . . . . . . . . . . 13
1.13 Breast cancer in Swedish twins (old and middle cohort) . . . 13
1.14 Data of five patients of the hepatitis study . . . . . . . . . . . 14

2.1 Parametric proportional hazards models for Halluca data . . 45

2.2 Proportional hazards models for Halluca data . . . . . . . . . 49
2.3 Parametric AFT models for Halluca data . . . . . . . . . . . 51

3.1 Parametric binary frailty model for Halluca data . . . . . . . 69

3.2 Parametric discrete frailty models for Halluca data . . . . . . 71
3.3 Parametric gamma frailty model for Halluca data . . . . . . . 79
3.4 Parametric gamma frailty models for Halluca data . . . . . . 81
3.5 Parametric gamma frailty model for the EBCT data . . . . . 82
3.6 Parametric gamma frailty models for the EBCT data . . . . . 83
3.7 Semiparametric gamma frailty model for the EBCT data . . 86
3.8 Semiparametric gamma frailty model for Halluca data . . . . 88
3.9 Univariate gamma frailty model for hepatitis A and B current
status data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.10 Parametric log-normal frailty models for Halluca data . . . . 98
3.11 Semiparametric log-normal frailty model for Halluca data . . 100
3.12 Parametric positive stable frailty models for Halluca data . . 107
3.13 PH model, gamma frailty model, and compound Poisson/PVF
frailty model applied to the malignant melanoma data . . . . 115
3.14 Parametric quadratic hazard frailty models for Halluca data . 119
3.15 Frailty and frailty cure models for the breast cancer data of
Swedish twins . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

xiii
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
xiv

4.1 Parametric shared gamma frailty models for Halluca data . . 142
4.2 Shared gamma frailty model for Halluca data . . . . . . . . . 144
4.3 Simulation of shared gamma frailty in current status data I . 146
4.4 Simulation of shared gamma frailty in current status data II . 147
4.5 Analysis of hepatitis A and B current status data with the
shared gamma frailty model . . . . . . . . . . . . . . . . . . . 147
4.6 Parametric shared log-normal frailty models for Halluca data 148
4.7 Shared log-normal frailty model for Halluca data . . . . . . . 149

5.1 Correlated gamma frailty model for Danish twins . . . . . . . 168

5.2 Heritability of cause of death . . . . . . . . . . . . . . . . . . 169
5.3 Correlated gamma frailty model for Danish twin data . . . . 172
5.4 Simulation of the correlated gamma frailty model (ρ = 0.5)
with current status data . . . . . . . . . . . . . . . . . . . . . 174
5.5 Simulation of the correlated gamma frailty model (ρ = 0.25)
with current status data . . . . . . . . . . . . . . . . . . . . . 175
5.6 Simulation of the correlated gamma frailty model (ρ = 0.50)
with current status data . . . . . . . . . . . . . . . . . . . . . 175
5.7 Simulation of the correlated gamma frailty model (ρ = 0.75)
with current status data . . . . . . . . . . . . . . . . . . . . . 175
5.8 Simulation of the correlated gamma frailty model (ρ = 1.00)
with current status data . . . . . . . . . . . . . . . . . . . . . 175
5.9 Analysis of hepatitis A and B current status data with the
correlated gamma frailty model . . . . . . . . . . . . . . . . . 176
5.10 Correlated log-normal frailty model for Danish twins I . . . . 178
5.11 Correlated log-normal frailty model for Danish twins II . . . 179
5.12 Correlated log-normal frailty model . . . . . . . . . . . . . . . 184
5.13 Correlated frailty models for breast cancer of Swedish twins . 187
5.14 Correlated gamma frailty models with/without cure fraction
applied to breast cancer in Swedish twins . . . . . . . . . . . 196
5.15 Parameter estimation in the simulation study . . . . . . . . . 197
5.16 Model 1 with two covariates, simulated data . . . . . . . . . . 200
5.17 Model 2 with two covariates, simulated data . . . . . . . . . . 201
5.18 Model 3 with two covariates, simulated data . . . . . . . . . . 202
5.19 Models 1 – 3 with two covariates, simulated data . . . . . . . 203

6.1 Correlated gamma frailty and copula models for Danish twins 214

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com

List of Figures

1.1 Median survival in days of the largest 25 diagnosing units . . 7

2.1 Event and censoring times of patients in a clinical trial . . . . 20

2.2 Event times of patients in a clinical trial with truncation . . . 25
2.3 Weibull hazard functions with different shape parameters . . 31
2.4 Log-logistic hazard functions with different parameters . . . . 32
2.5 Gompertz hazard functions with different shape parameters . 34
2.6 Log-normal hazard functions with different parameters . . . . 35
2.7 Kaplan–Meier curves for different patient groups in the EBCT
study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.8 Nelson–Aalen curves for different patient groups in the EBCT
study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.1 Conditional and unconditional hazard functions of a hetero-

geneous population . . . . . . . . . . . . . . . . . . . . . . . . 62
3.2 Mortality hazards of two populations with Gompertz hazard
and their mixed hazard . . . . . . . . . . . . . . . . . . . . . 66
3.3 Proportions of two sub-populations with Gompertz hazard . . 68
3.4 Probability density functions of different gamma distributions 72
3.5 Baseline hazard functions of two populations with Gompertz
hazard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.6 Unconditional hazard functions of populations with Gompertz
baseline hazard and gamma-distributed frailty . . . . . . . . . 77
3.7 Age-specific prevalence of hepatitis A . . . . . . . . . . . . . . 90
3.8 Age-specific prevalence of hepatitis B . . . . . . . . . . . . . . 91
3.9 Probability density functions of log-normal distributions . . . 97
3.10 Probability density functions of inverse Gauss distributions . 101
3.11 Kaplan–Meier estimator and compound Poisson frailty model
for Swedish breast cancer data . . . . . . . . . . . . . . . . . 126

5.1 Cause-specific frailties and their correlations in a twin pair . . 206

7.1 Time-varying frailties of two relatives in the Paik model . . . 234

xv
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
Free ebooks ==> www.Ebook777.com

www.Ebook777.com
 Free ebooks ==> www.Ebook777.com

Symbol Description
ACE genetic model M0 generic symbol for a
AFT accelerated failure cumulative baseline
time hazard
AIC Akaike information log N (m, s2 ) log-normal distribution
criterion with parameters m, s2
AR autoregressive process MCEM Markov Chain EM
BMI body mass index MCMC Markov Chain Monte
BLUB best linear unbiased Carlo
predictor ML maximum likelihood
CAD coronary artery MZ monozygotic
disease NSCLC non-small cell lung
CSRF corrected scale reduc- carcinoma
tion factor P(A) probability of event A
CHD coronary heart disease pdf probability density
cdf cumulative density function
function PH proportional hazards
DIC Bayesian information PPL penalized partial likelihood
criterion PVF power variance function
DZ dizygotic S generic symbol for a
E expectation survival function
EBCT electron-beam com- t+ truncation time
puted tomography TNM classification of malig-
ECOG Eastern Cooperative nant tumours
Oncology Group V variance
EM expectation-maximi- U (a, b) uniform distribution in
zation the interval [a, b]
f generic symbol for pdf N (µ, σ 2 ) normal distribution
F generic symbol for cdf with parameters µ, σ 2
Γ gamma function Exp(λ) exponential distribution
H0 null hypothesis with parameter λ
HA alternative hypothesis W (λ, ν) Weibull distribution
ICD International Classifica- with parameters λ, ν
tion of Diseases G(λ, ϕ) Gompertz distribution
iid independent and with parameters λ, ϕ
identically distributed Γ(k, λ) gamma distribution
L Laplace transform with parameters k, λ
L likelihood function log L(ν, κ) log-logistic distribution
µ generic symbol for a with parameters ν, κ
hazard P s(α) positive stable distribution
µ0 generic symbol for a with parameter α
baseline hazard cP (γ, k, λ) compound Poisson
M generic symbol for a distribution with
cumulative hazard parameters γ, k, λ

xvii
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
Free ebooks ==> www.Ebook777.com

www.Ebook777.com
 Free ebooks ==> www.Ebook777.com

Preface

The analysis of lifetime data (or more exactly, time-to-event, event-history,

or duration data) plays an important role in medicine, epidemiology, biology,
demography, economics, engineering, actuarial science, and other fields. It has
expanded rapidly in the last three decades, with works having been published
in various disciplines in addition to statistics. But what distinguishes survival
analysis from other fields of statistics? Why does survival data need a special
statistical theory? The main problem is censoring, which means that, for some
individuals in the study population, the researcher only has the information
that the event of interest did not occur before a particular time point. To
put it plainly, a censored observation contains only partial information about
the random variable of interest. This kind of incomplete observation needs
special methods. As a consequence of censoring, survival times are usually
a mixture of discrete (censoring indicator) and continuous (event/censoring
time) data that lend themselves to a different type of analysis from that used
in the traditional discrete or continuous case. The mixture is the result of
censoring and has an important effect on data analysis. The Kaplan–Meier
estimator (Kaplan and Meier 1958) of the survival function is a major step
in the development of suitable models for such kind of data. Furthermore,
most evaluations are made conditionally on what is known at the time of the
analysis, and this changes over time. Usually, as the population under study
is changing, we only consider the individual risk to die for those who are still
alive, but this means that many standard statistical approaches cannot be
applied.
Models based on the hazard function have dominated survival analysis since
the construction of the proportional hazards model by Cox (1972). One of the
reasons this model is so popular is the ease with which technical difficulties
such as censoring and truncation are handled. This is due to the appealing
interpretation of the hazard as a risk that changes over time. Naturally, the
concept allows for the entering of covariates in order to describe their influence
and to model different levels of risk for different subgroups.
This book focuses on frailty models, a specific area in survival analysis.
The concept of frailty provides a convenient way of introducing unobserved
heterogeneity and associations into models for survival data. In its simplest
form, frailty is an unobserved random proportionality factor that modifies
the hazard function of an individual or related individuals. In essence, the
concept goes back to the work of Greenwood and Yule (1920) on ”accident
proneness” with binary data. The first univariate frailty model was suggested

xix
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
xx

by Beard (1959), considering different mortality models. The term frailty

itself was introduced by Vaupel et al. (1979) in the univariate context. Its
first applications to problems in multivariate survival analysis date from a
seminal paper by Clayton (1978).
Ordinary methods in survival analysis implicitly assume that populations
are homogenous, meaning that all individuals have the same risk of death.
However, in general, it is impossible to include all relevant risk factors, perhaps
because we have no information on individual values, which is often the case
in demography. Furthermore, we may not know all relevant risk factors, or it
is impossible to measure them without great financial costs, something that
is common in medical and biological studies. The neglect of covariates leads
to unobserved heterogeneity. That is, the population consists of subjects with
different risks. As a consequence, it is important to consider the population
as heterogeneous, i.e., as a mixture of individuals with different hazards. A
frailty model is a random effects model for time-to-event data, where the
random effect (frailty) has a multiplicative effect on the baseline hazard. It can
be used for univariate (independent) lifetimes, i.e., to adjust for unobserved
risk factors in a proportional hazards model (heterogeneity). The variability
of event-time data is split into one part that depends on covariates and is
thus theoretically predictable, and one part that is initially unpredictable,
even knowing all relevant information at that time. There are advantages in
separating these sources of variability: unobserved heterogeneity can explain
some unexpected results or give an alternative interpretation, for example,
crossing-over or leveling-off effects of hazards.
However, considering multivariate (dependent) duration times is especially
interesting. The introduction of a common random effect – frailty – is a
natural way of modeling the dependence of event times. The random effect
explains the dependence in the sense that had we known the frailty, the event
times would have been independent. In other words, because we do not know
the frailty, the lifetimes are independent conditionally on the frailty. This
approach can be used for survival times of related individuals such as twins
or family members, where independence cannot be assumed, or for recurrent
events in the same individual or for times to several events for the same
individual, such as onset of different diseases, relapse, or death (competing
risks). Different extensions of univariate frailty models to multivariate models
are possible and will be considered in this book.
The standard assumption is to use a gamma distribution for frailty, but
other distributions are also possible. The relationships between individual
and observed survival characteristics play a key role in the statistical analysis
of duration data in heterogeneous populations.
Various frailty models have been developed in the past. However, compared
with standard mixed models, frailty models pose additional difficulties in
developing inferential methods, caused by incomplete data due to censoring
and truncation. Thus, inferential methods have been less developed here than
in other mixed models.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
xxi

To keep the book to a reasonable length, some topics are discussed only
briefly, and references are given for further reading. Because the literature
on frailty models is extensive (especially in the last few years), the choice of
subject matter is difficult. The material discussed in detail is to some extent a
reflection of the author’s interest in this research field. However, my attempt
has been to present a relatively comprehensive and complete overview of the
fundamental approaches in the field of frailty models.
The present monograph is primarily aimed at the biostatistical community
with applications from biomedicine, (genetic) epidemiology, and demography.
Some efforts were also undertaken to include literature from other fields like
econometrics if interesting methodological problems are raised. The practical
use of models is a key issue in biostatistics, where the data at hand often
are motivating for the development of new models. The language of this
book is nontechnical and therefore it can be understood by nonspecialists.
Nevertheless, some experience with survival analysis is an advantage.
Acknowledgments
I would like to express my sincere thanks to everyone who supported directly
or indirectly this book; my former Ph.D. advisor Friedrich Liese (Rostock) for
awakening my interest in statistical research, Anatoli Yashin and Konstantin
Arbeev (Duke), Paul Janssen and Niel Hens (Hasselt), Catherine Legrand
(Louvain-la-Neuve), Alexander Begun (Hamburg), Nicole Giard (Gütersloh),
Kaare Christensen and Ivan Iachine (Odense), Isabella Locatelli (Lausanne),
Slobodan Zdravkovic (Kopenhagen), Samuli Ripatti (Helsinki), Oliver Kuß
(Halle), and Juni Palmgren (Stockholm) for many helpful discussions and
fruitful collaboration. I take the opportunity to thank all my colleagues at
the Institute for Medical Epidemiology, Biostatistics, and Informatics at the
Medical Faculty of the Martin Luther University Halle-Wittenberg, especially
Johannes Haerting as head of the institute, for providing an excellent research
environment. I would like to thank the Danish and Swedish Twin Registries
for making the unique twin data available. I extend my sincere thanks to the
Max Planck Institute for Demographic Research (Rostock) and its founding
director, James Vaupel, where I began to work on frailty models and where
I returned several times as a guest researcher during the last years. Special
thanks go to Luc Duchateau (Ghent) for careful reading an earlier version
of the manuscript and giving many very helpful suggestions. Furthermore, I
would like to thank my Ph.D. students Katharina Hirsch and Diana Pietzner
(Halle) for many fruitful discussions and help in preparing the manuscript.
Finally, acknowledgments go to my family. To my parents Margit and Kurt
Wienke for all their love and support. Also, to my sons Moritz, Jakob, and
Finn, showing that there is a life beyond research and, finally, to my loving
wife Kati Moeller, who always strongly encouraged me to finish this book.

Halle, June 2010 Andreas Wienke

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com

Chapter 1
Introduction

1.1 Goals and Outline

Survival analysis is one of the core research methods used in many fields such
as medicine, biology, epidemiology, demography, and engineering. Notion
survival analysis reflects the origin of the methods in medical and demographic
studies of mortality. Especially since the end of the 1970s, the empirical
analysis of event history data has become widespread by the development
of the proportional hazards model in the seminal paper by Cox (1972) and
several extensions during the last three decades. The present monograph
deals with one important direction of extensions in this field, namely frailty
models. A frailty model is a multiplicative hazard model consisting of three
components: a frailty (random effect), a baseline hazard function (parametric
or nonparametric), and a term modeling the influence of observed covariates
(fixed effects).
Only a few books exist on this subject, which contain short chapters devoted
to frailty models. Ibrahim et al. (2001) consider parametric as well as semi-
parametric shared frailty models based on a Bayesian approach. Klein and
Moeschberger (2003) consider the application of the EM algorithm in semi-
parametric shared frailty models. Aalen et al. (2008) take a process point of
view dealing with different frailty models. The present book extends in two
main directions the presentation of frailty models made in the seminal mono-
graphs by Hougaard (2000), Therneau and Grambsch (2000), and Duchateau
and Janssen (2008). First, univariate frailty models with their focus on unob-
served heterogeneity are covered in more detail compared to previous books.
In univariate models all durations describe the time to the same type of event,
and event times are considered as independent. Second, the main emphasis
is placed on correlated frailty models as natural extensions of shared frailty
models. Here, different strengths of association between clustered lifetimes
are of special interest.
One of the main problems in the application of frailty models to real data
is the limited availability of standard software in this area. Consequently,
one aim of this monograph is to show which of the models considered can
be applied to real data by using standard statistical packages such as R,
SAS, and STATA. Here, the link to generalized linear mixed models will be

1
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
2 Frailty Models in Survival Analysis

exploited. Both parametric as well as semiparametric models are considered.

Furthermore, models are fitted by the frequentist and Bayesian approaches.
Most of the Bayesian analyses are performed with WinBUGS, but the new
PROC MCMC in SAS opens new possibilities for the future.
In this first chapter we will introduce different data sets used throughout
the book to illustrate modeling techniques and practical interpretations of
the results. In Chapter 2 an introduction to basic and general concepts in
survival analysis and a definition of common terminology are given. The
topic is also covered by other books, for example, Miller (1981), Cox and
Oakes (1984), Andersen et al. (1993), Lawless (2002), Kalbfleisch and Prentice
(2002), Klein and Moeschberger (2003), Collett (2003), and Machin, Cheung,
and Parmar (2006). The recent book by Finkelstein (2008) has a special focus
on reliability but also covers a wide range of exiting topics in biostatistics and
demography. Though, it is not necessary for people acquainted with this field
to read it, it does contain notations and key results and lays the basis for
the more advanced frailty models treated in the following chapters. After
this preparatory chapter we deal with univariate frailty models (single spell
data) in Chapter 3, discussing the broad range of possible frailty distributions
with their specific features. Gamma distribution is the most often applied
frailty distribution because frailties appearing in conditional likelihood can be
integrated out, giving simple expressions of unconditional likelihood. Then,
maximization of unconditional likelihood can be used for estimation. Here,
the interpretation of frailty is as unobserved heterogeneity due to nonobserved
covariates. The focus of Chapter 4 is on the shared frailty model, which has
already been discussed in detail by other authors (Hougaard 2000, Therneau
and Grambsch 2000, Duchateau and Janssen 2008). Shared frailty models are
an important tool for analyzing multivariate (clustered) survival data. Hence,
this chapter forms the basis of the correlated frailty model and its extensions
considered in detail in Chapter 5. Advantages and limitations of the proposed
models are discussed, and simulations show the properties of the parameter
estimates for finite sample sizes. Different approaches and applications are
presented to demonstrate the flexibility of the correlated frailty approach in
modeling associations in clustered event times. Chapter 6 deals with copula
models and analyzes similarities and dissimilarities between frailty and copula
models. Chapter 7 gives an overview of different problems related to frailty
models such as tests for homogeneity, identifiability aspects, and available
software. The Appendix provides a series of technical mathematical results
and background about genetic models used throughout the book.
The present monograph does not attempt to give a complete overview of
the fast growing literature on frailty models; this would not be possible. The
treatment of the topics covered are restricted to explaining the basic ideas in
frailty modeling and statistical techniques, with focus on real data application
and interpretation of the results. In many cases different models are applied
to the same data to compare and discuss their advantages and limitations
under varying model assumptions.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Introduction 3

1.2 Examples
Different survival models are considered in this book. Most of them will
be applied to real data, mainly using examples from research fields such as
medicine, epidemiology, and demography. Survival analysis deals with the
analysis of times until the occurrence of a well defined event. The occurrence
of this event describes the transition from one state to another, for example,
occurrence of a disease is the transition from the state of being healthy to the
state of being sick. Sometimes the transition is of special interest (incidence
of the disease), and in other cases the state (prevalence of the disease) is the
target of the analysis. For such kind of analysis it is necessary to define the
time scale and a starting time point zero. In many cases the time scale is the
age of the individual. In clinical trials the starting point is often beginning
of treatment. If the focus is on the development of a disease, the time of
diagnosis is usually the starting point. In occupational cohort studies the
starting point is often the beginning of employment or unemployment.
We first consider the univariate event times, which means data with no
clustering. Such data set is given in Example 1.1, based on a prognostic study
analyzing the value of electron-beam computed tomography (EBCT) derived
calcium scores for risk stratification in symptomatic patients. Example 1.2
presents the malignant melanoma data. The models fitted to these data sets
are parametric and semiparametric proportional hazard models. The most
important goal of Chapter 3 is to analyze the effect of including unobserved
heterogeneity on regression parameter estimates.
However, the main focus of this book is on multivariate frailty models, where
event times are clustered. Example 1.3 will serve as an example of univariate
as well as multivariate data. In the last situation the cancer-diagnosing units
were considered as clusters. The cluster size differs from cluster to cluster,
which is common in multicenter clinical trials. Here, frailty can describe
center-to-center variations not explained by observed covariates. In addition
to the problem of analyzing the effect of observed covariates, an important
research problem is evaluating the dependence between event times in clusters.
In genetic studies, correlations between family members are the basis of the
analysis of heritability of specific traits, for example, the times of onset of
breast cancer or cause of death specific lifetimes. We use Danish and Swedish
twin data provided by the Danish Twin Registry at the University of Southern
Denmark in Odense and the Swedish Twin Registry at the Karolinska Institute
in Stockholm to emphasize the practical purpose of the frailty models with
fixed and small cluster sizes. In Example 1.4, cause-specific lifetimes of Danish
twins are considered. A subsample with additional covariate information is
presented in Example 1.5. Example 1.6 provides data on the age of onset
of breast cancer in Swedish twins, whereas Example 1.7 deals with current
status data. The next section provides a brief description of these data.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
4 Frailty Models in Survival Analysis

Example 1.1 Prognostic Study of the EBCT Calcium Score

EBCT-derived calcium score is a measure of coronary arteriosclerotic plaque
that can be used for more precise risk stratification in symptomatic patients
(Schmermund et al. 2004). In the study presented here, it was investigated
whether EBCT-derived calcium score can add prognostic information com-
pared with clinical information derived from risk-factor assessment, exercise
stress testing, and coronary angiography. Patients with recent (<3 months)
onset of symptoms were retrospectively identified and examined for possible
coronary artery disease (CAD) and underwent EBCT. Complete follow-up
after 42 months was available for 255 patients with mean age at baseline of
58 years, who were finally included into the study.

Table 1.1: Five patients in the EBCT study

id time status risk group calcium score age

1 42 0 2 0 70
2 42 0 1 0 59
3 42 0 1 1 74
4 14 1 4 1 70
5 42 0 1 0 50

Four clinical risk groups with increasing evidence of CAD were constructed
based on risk factor assessment, exercise stress testing, coronary angiographic
anatomy, and revascularization at baseline. The main interest was in the
occurrence of a combined event consisting of major adverse cardiac events
such as myocardial infarction, cardiac death, and revascularization. The event
was observed in 40 (16%) patients during the follow-up, the observations of
the other patients are mainly censored after 42 months at the end of the study.
The data for five patients are presented in Table 1.1.

Table 1.2: Description of the EBCT study population

covariate category absolute frequency relative frequency

risk group 1 79 31.0%

2 78 30.6%
3 42 16.5%
4 56 21.9%
calcium score < 100 150 58.8%
≥ 100 105 42.2%
age (years) 23 – 54 85 33.3%
55 – 62 79 31.0%
63 – 84 91 35.7%

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Introduction 5

The first column gives the patient specific identification number, the observed
event or censoring time in the second column is measured in months. The
covariate of main interest in this study is the CAD risk divided into four
prognostic groups (group 1 – no evidence of ischemia, ≤ 1 conventional risk
factor; group 2 – evidence of ischemia and/or ≥ 2 conventional risk factors, no
angiographic stenoses; group 3 – angiographic stenoses, no revascularization
at baseline; group 4 – early revascularization). The dichotomous covariate
calcium indicates an EBCT-derived calcium score larger than 100. There is
no clustering in this data set. One of the research questions was to examine
whether the EBCT-derived calcium score can add prognostic information
compared with the clinical information summarized in the risk groups. Age
(in years) was categorized into three groups with the youngest age group as
the reference. The covariate frequencies are given in Table 1.2.

Example 1.2 Malignant Melanoma Data

The data set contains observations of 205 patients with radical surgery for
malignant carcinoma (skin cancer) at the University Hospital of Odense in
Denmark during 1962–1977. Radical surgery means that the tumor was
completely removed including the skin within a distance of about 2.5 cm
around it. Patients were followed up until 1977 and 57 deaths from malignant
melanoma, and 14 deaths due to other causes (coded as censored event times)
were observed. The data of five patients are given in Table 1.3.

Table 1.3: Data of five patients of the

malignant melanoma study
id time status gender age thickness

1 10 0 1 76 676
2 30 0 1 56 65
3 35 0 1 41 134
4 99 0 0 71 290
5 185 1 1 52 1208

The first column provides the unique patient identification number. Variable
time measures time since surgery in months and variable status indicates
the occurrence of death caused by malignant melanoma. There are several
covariates available in the data set; for ease of presentation we will restrict
them in this application to the following three covariates: gender (0 = female,
1 = male), age at surgery (years), and tumor thickness (in 1/100 mm). This
is a univariate data set without clustering. The data was first analyzed by
Drzewiecki et al. (1980a,b) and later published and reanalyzed in more detail
by Andersen et al. (1993).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
6 Frailty Models in Survival Analysis

Example 1.3 Halluca Study

The Halle Lung Cancer (Halluca) study was a study investigating provision
of medical care to lung cancer patients in the region of Halle and Dessau in
the eastern part of Germany (Bollmann et al. 2004, Kuß et al. 2008). The
study region covers about 1.5 million inhabitants and belongs to the State
of Saxony-Anhalt. In cooperation with the regional clinical tumor registries,
all lung cancer patients in the study region were recorded from April 1996
to September 1999, and follow-up was done until September 2000. A total of
1696 lung cancer patients were observed, and survival was defined as time from
clinical, histological or cytological diagnosis to death. 1349 patients (79.5%)
died until the end of follow-up; median survival in the study population was
9.3 months. To validate and complement survival information, the data from
the Clinical Cancer Registry were compared to death certificates collected by
the local health institutions and linked to the data from the Common Cancer
Registry of Eastern Germany. Minimal follow-up time was 12 months, and
the median follow-up time 33 months. To judge the influence of prognostic
and risk factors on overall survival, five fixed-effects covariates, known to be
important in the prognosis of survival regarding lung cancer, were included.
The information for five patients is given in Table 1.4. The first column gives

Table 1.4: Data of five patients in the Halluca study

id time status unit gender age type ECOG stage

1 54.74 1 1 1 75.02 1 0 4
2 6.68 1 1 1 63.60 1 0 5
3 0.33 1 1 1 52.68 2 . .
4 24.28 1 2 1 55.14 . 0 .
5 15.46 0 2 0 79.28 2 0 1

the patient-specific id number, and the second column the survival time (in
months). The third column contains the survival status (1 = death, 0 = alive)
and the fourth column the cluster variable diagnosing unit. Lung cancer was
diagnosed in 56 different diagnosing units with numbers of patients ranging
from 1 to 392 (mean 30.3). The Halluca data is analyzed using univariate
approaches in Chapters 2 and 3. In Chapter 4 the data is treated like from
a multicenter study with the diagnosing unit as cluster variable. A cluster
effect by diagnosing unit is indicated by Figure 1.1. In multicenter studies
(with treatment center as cluster), despite the tight study protocols, often
center-to-center variation occurs, which cannot be explained by covariates.
Frailty models can be used to investigate this variation. The other columns
represent variables gender (0 = female, 1 = male), age (years), histologic type
(1 = small-cell lung cancer, 2 = non-small-cell lung cancer), ECOG status
(range 0 to 4), and UICC stage (1 = I, 2 = II, 3 = IIIa, 4 = IIIb, 5 = IV).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Introduction 7

500
400
●
● ●

●
●
300
●
●
● ●
●
days

● ●

●
200

● ●

●
● ●
100

●
0

5 10 15 20 25

diagnosing units

Figure 1.1: Median survival (95% CI) of 25 diagnosing units in Halluca

Dots indicate missing covariate values. It was initially decided to model these
as separate categories despite the dangers of this procedure. Throughout the
book, in the tables presenting the results, these missing categories are omitted.

Table 1.5: Description of the Halluca study population

covariate category absolute frequency relative frequency
gender male 1374 81.0%
female 322 19.0%
histologic type NSCLC 1183 69.8%
SCLC 366 21.5%
missing 147 8.7%
ECOG ECOG 0-2 1366 68.7%
ECOG 3-4 123 7.3%
missing 407 24.0%
stage I 185 10.9%
IIa 79 4.7%
IIb 195 11.5%
III 280 16.5%
IV 621 36.6%
missing 336 19.8%

We further explicitly omitted the primary treatment as a covariate. This

was to prevent unjustified treatment recommendations, which should only be
derived from randomized trials and not from observational studies. The study
population is described in Table 1.5. Mean age at diagnosis was 65 years.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
8 Frailty Models in Survival Analysis

Example 1.4 Danish Twins Cause-Specific Mortality Data

The Danish Twin Registry was the world’s first nation–wide twin registry,
established in 1954 by Bent Harvald and Mogens Hauge. The older part
of this population based registry includes all twins born in Denmark during
the period 1870–1910 and all like-sex pairs born between 1911 and 1930 in
which both partners survived to the age of six years. Pairs with deaths be-
fore the age of six were excluded because it turns out to be very difficult to
obtain detailed information especially about the zygosity of such twin pairs.
The birth registers from all 2200 parishes of Denmark during the relevant
calendar years were manually scrutinized to identify multiple births. After
such births were identified, a search was then carried out for the twins or,
whenever needed, for their closest relatives in regional population registers
(in operation since 1924) or other public sources, especially the archives of
probate courts and censuses. As soon as a twin was traced, a questionnaire
was sent to the twin (if she or he was alive) or to the closest relatives (if she
or he was not alive). Questions about phenotypic similarities were included
in the questionnaires to assess the zygosity by self-reported similarities. The
reliability of this method was validated by comparison with the results of
laboratory methods based on blood serum enzyme group determination in a
subgroup of twins. The misclassification rate of this method was found to
be less than 5% in the Danish twin data (Holm 1983). Similar results are
known from the Swedish Twin Registry (Cederlöf et al. 1961) The follow-up
procedure traced nearly all twins who did not die or emigrate before the age
of six years. For further, more detailed information about the construction
and the composition of the Danish Twin Registry see Hauge (1981).
The data provided by the Danish twin registry contain 8201 monozygotic
(MZ) and dizygotic (DZ) twin pairs who were born between 1 January 1870
and 31 December 1930, and who were both still alive on 1 January 1943.
As a consequence of this restriction, around two-thirds of the twin pairs born
were excluded because of the high infant mortality of this period (1870–1930).
Furthermore, twins have a higher infant mortality than singletons because of
their lower birth weight. It was necessary to exclude early deaths because it
was nearly impossible to obtain zygosity information when one or both twin
partners died at young ages. Zygosity information is crucial to the application
of the methods in twin research.

Table 1.6: Data of three Danish twin pairs

id time status pair gender zygosity cause birth

1 76.09 1 1 1 1 2 1889
2 76.02 1 1 1 1 2 1889
3 64.73 1 2 0 2 4 1908
4 94.75 1 2 0 2 1 1908
5 85.62 0 3 0 1 0 1881
6 68.61 1 3 0 1 2 1881

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Introduction 9

Observed covariates are gender, zygosity and year of birth. A total of 246 twin
pairs with incomplete information about the cause of death were excluded,
leaving a study population of 7955 twin pairs. Individuals were followed up
through 31 December 1993, and those identified as deceased after that date
have been classified here as living. Altogether, we have 1344 male MZ twin
pairs and 2411 DZ twin pairs, and 1470 female MZ twin pairs and 2730 DZ
twin pairs. In addition to the lifetimes, there is information about cause of
death for all noncensored lifetimes, that is, for all individuals in the study
population who died before 31 December 1993. For the present analysis, only
the underlying cause of death was considered.
The data for the first six twins are given in Table 1.6. The first column gives
the identification number of the individual, and the second one the survival or
censoring time (in years). The third column contains the censoring indicator
(1 = death, 0 = alive), the fourth column is the identification number of
the twin pair (cluster), and the four other columns represent the covariates
gender (0 = female, 1 = male), zygosity (1 = monozygotic, 2 = dizygotic),
cause of death (0 = alive, 1 = cancer, 2 = coronary heart disease, 3 = stroke,
4 = respiratory diseases, 5 = other), and year of birth. For more detailed
information about cause of death, gender, and zygosity of the study population
see Table 1.7.

Table 1.7: Causes of death in the Danish twin population (number of

individuals)
males females
cause of death MZ twins DZ twins MZ twins DZ twins
cancer 440 809 423 823
coronary heart disease 666 1180 548 999
stroke 161 278 186 335
respiratory diseases 143 203 89 205
other causes 336 661 330 555
all causes together 1746 3131 1576 2917
alive 942 1691 1364 2543

Information regarding death status, age at death, and cause of death was
obtained from the Central Person Register, the Danish Cancer Register, the
Danish Cause–of–Death Register, and other public registries in Denmark.
The main source for obtaining information on cause of death was the Death
Register at the National Institute of Public Health. Information about cause
of death is available from this register for individuals who died after 1942
(Juel and Helweg-Larsen 1999). Consequently, cause of death is included in
the twin register only for twins who died after this year.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
10 Frailty Models in Survival Analysis
Table 1.8: Cause of death groups by ICD number
cause of death ICD revision 6 & 7 ICD revision 8

cancer 140 – 205 140 – 209

coronary heart disease 420 410 – 414
stroke 330 – 334 430 – 439
respiratory diseases 470 – 527 460 – 519

The validity of the twin register was checked on the basis of a comparison of
information about year of death with the nationwide Danish Cancer Register.
There was around 99% agreement, although both registries were independent.
Further data corrections increased this level of agreement to almost 100%.
Cause of death was coded following the sixth, seventh, and eighth edition of
the International Classification of Diseases (ICD). Four different groups of
main causes of death are considered in the present example: cancer, coronary
heart disease (CHD), stroke, and diseases of the respiratory system. ICD
codes in three revisions of the ICD for these broad cause-of-death groups are
given in Table 1.8. Causes of death are a common example for competing
risks.

Example 1.5 Danish twins CHD mortality data with covariates

The data in this example is a subset of the cause-specific mortality data in
the foregoing example. Here the main focus is on age at death with death
caused by coronary heart disease and the influence of BMI and smoking on
this outcome.
In 1966, a questionnaire including questions about smoking, height, and
weight was mailed by the Danish Twin Registry to all Danish twins born
in the period 1890–1920 who were alive and traceable on 1 January 1966.
3709 individuals answered the questionnaire (response rate 65%). Excluded
from the study were 813 twins with nonresponding partners, four pairs with
unknown zygosity, and 212 pairs with incomplete or uncertain information on
height and weight. A total of 23 pairs were excluded because of incomplete
information about the cause of death, resulting in a total study population of
1209 complete twin pairs.
Individuals were followed from 1 January 1966 to 31 December 1993. Those
persons identified as deceased after the follow-up period are classified for our
purposes as censored. At the end of follow-up period, approximately 40% of
the twins were still alive, resulting in the right censored data. Altogether,
there were 210 male monozygotic twin pairs and 316 dizygotic twin pairs,
and 273 female monozygotic twin pairs and 410 dizygotic twin pairs. The
data for the first six twins in the study population are given in Table 1.9.
The first column provides the unique identification number of the twin, the
second column the observation time (in years). The third column contains the

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Introduction 11
Table 1.9: Data of three Danish twin pairs with covariates
id time status pair gender zygosity birth BMI smoking

1 74.23 1 1 0 2 1893 1 1
2 81.52 0 1 0 2 1893 2 2
3 72.58 1 2 0 2 1912 2 4
4 54.89 1 2 0 2 1912 1 2
5 57.31 0 3 1 1 1897 3 3
6 83.79 1 3 1 1 1897 1 4

censoring indicator (1 = death by CHD, 0 = censored (including non-CHD

deaths)), the fourth column the number of the twin pair (cluster), and the
following five columns represent the covariates gender (0 = female, 1 = male),
zygosity (1 = monozygotic, 2 = dizygotic), year of birth, body mass index
(1 if BMI < 22kg/m2 , 2 if 22kg/m2 ≤ BMI ≤ 28kg/m2, 3 if 28kg/m2 < BMI),
and smoking (1 = nonsmoker, 2 = cigarette smoker, 3 = other smoker (pipe
smoker, cigar smoker, etc.), 4 = former smoker). More detailed information
about the distribution of gender, smoking, and BMI in the study population
is given in Table 1.10.

Table 1.10: Study population (number of individuals) by gender,

BMI, and smoking
non- former other cigarette total
smokers smokers smokers smokers

males
BMI < 22 11 12 45 31 99 ( 9.4%)
BMI 22–28 78 141 333 208 760 (72.2%)
BMI > 28 24 35 91 43 193 (18.3%)
Total 113 188 469 282
(10.7%) (17.9%) (44.6%) (26.8%)
females
BMI < 22 105 46 47 99 297 (21.7%)
BMI 22–28 350 138 134 171 793 (58.1%)
BMI > 28 157 40 37 42 276 (20.2%)
Total 612 224 218 312
(44.8%) (16.4%) (16.0%) (22.8%)

Altogether, 562 CHD deaths occurred in the population during follow-up.

CHD is grouped as ICD 420 in the sixth and seventh revision, and as ICD 410–
414 in the eighth ICD revision. The same data set (with minor differences)
was used by Herskind et al. (1996a) with focus on total mortality.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
12 Frailty Models in Survival Analysis

Example 1.6 Swedish twins breast cancer data

First established in the late 1950s to study the influence of smoking and
alcohol consumption on cancer and cardiovascular diseases while controlling
for genetic propensity to disease, the Swedish Twin Registry has developed
into a unique source. Since its establishment, the registry has been expanded
and updated on several occasions, and the focus has similarly been broadened
to include most common complex diseases. Today it contains around 70,000
twin pairs born between 1886 and 1990. Nearly 400 papers have been published
based on data from the Swedish Twin Registry.
When the twin registry was initiated, the church registers from all parishes of
the period 1886–1925 were checked manually to identify all multiple births.
Between 1959 and 1961, questionnaires were sent to all twins, including a
question about phenotypic similarities between partners to assess zygosity:
‘Were you as children as alike as two peas in a pod?’ When both partners
agreed, they were defined as monozygotic twins. If both responded ‘not
alike’, they were classified as dizygotic. If the twins did not agree, or if only
one member of the pair answered the question, the pair was classified as of
unknown zygosity. This classification was compared with laboratory methods
(serological markers) and the misclassification rate was found to be very low
(Cederlöf et al. 1961), similar to the Danish data (Holm 1983).
At present, the Swedish Twin Registry contains information about three
cohorts of Swedish twins referred to as the ‘old’, ‘middle’, and ‘young’ cohort.
Here the focus is on the old cohort consisting of all like-sex pairs born between
1886 and 1925 where both members in a pair were alive and living in Sweden
in 1959. The data for the first six twins are given in Table 1.11.

Table 1.11: Data of three Swedish twin pairs

id time status pair gender zygosity birth age

1 76.09 1 1 1 1 1889 23
2 76.02 0 1 1 1 1889 26
3 64.73 0 2 0 2 1908 0
4 94.75 0 2 0 2 1908 23
5 85.62 0 3 0 1 1887 25
6 68.61 1 3 0 1 1887 29

The first column gives the identification number of the individual, and the
second column the age at diagnosis of breast cancer or censoring (in years).
The third column contains the censoring indicator (1 = onset of breast cancer,
0 = no onset of breast cancer), the fourth column the identification number of
the twin pair (cluster), and the other columns represent gender (0 = female,
1 = male), zygosity (1 = monozygotic, 2 = dizygotic), year of birth, and age
at giving first birth (in years). Zero indicates nulliparous women.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Introduction 13

The summary for the old cohort is given in Table 1.12, stratified according
to the censoring status. The event under study is the onset of breast cancer.
If a woman did not develop breast cancer or if she died from other causes
during the follow-up, the corresponding observation is censored. Age at onset
of breast cancer ranges from 36 years to 93 years.
The data set was created by merging the Swedish Twin Registry with the
Swedish Cancer Registry maintained by the National Board of Health and
Welfare. At the time of record linkage of the data used here, the Swedish
Cancer Registry contained all cases of cancer that were diagnosed during
the period 1959 through 2000, and 715 cases of breast cancer were identified
during follow-up.

Table 1.12: Breast cancer in Swedish twins (old cohort)

number of twin pairs
both censored one censored none censored total
MZ twin pairs 1767 218 18 2003
DZ twin pairs 3420 407 27 3854
total 5187 625 45 5857

In another analysis, without considering the covariate age at first birth, the
twins from the old and the middle cohort are combined. The middle cohort
comprises all twins born between 1926 and 1967 who were alive and living in
Sweden in 1970. Altogether, 1096 breast cancer cases were observed during
the follow-up until 2000 in both cohorts. The data structure is the same as in
Table 1.11 (but without covariate age). Summary statistics for both cohorts
combined are given in Table 1.13.

Table 1.13: Breast cancer in Swedish twins (old & middle cohort)
number of twin pairs
both censored one censored none censored total
MZ twin pairs 4304 335 33 4672
DZ twin pairs 7236 625 35 7896
total 11540 960 68 12568

For a comprehensive description of the Swedish Twin Registry database at

the Karolinska Institute (Stockholm) with focus on the recent data collection
efforts and a review of the principal findings that have come from the registry,
see Lichtenstein et al. (2002).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
14 Frailty Models in Survival Analysis

Example 1.7 Cross-Sectional Multisera Data on Hepatitis A and B

Viral hepatitis is a serious health problem throughout the world. To obtain a
clear picture of the prevalence of hepatitis A, B, and C, a sero-epidemiological
study was undertaken in 1993/1994 in Flanders (Belgium). From the 4058
blood samples drawn from a study population representative of the Flemish
population, the main focus is on complete cases specifically with respect to
hepatitis A and B, resulting in 3787 blood samples. These blood samples
were then tested for the presence of antibodies for the different infections and,
using prespecified cut-off values, were classified as either positive or negative.
Together with the patient’s age and the assumption of lifelong immunity,
these data constitute current status data on whether or not past infection
took place. Hepatitis C was not considered here because of its low prevalence
(less than 1%). More details about this data can be found in Beutels et al.
(1997). It would be interesting to look at the heterogeneity in acquisition of
either infection and the correlation between the acquisition of both infections.
While age-dependent seroprofiles reflect the age-specific risk of infection, the
proper assessment of heterogeneity has direct implications with respect to the
two key quantities in infectious disease epidemiology, the basic reproduction
number, and their associated critical vaccination coverage (Farrington et al.
2001). Estimating the correlation could indicate transmission through similar
routes (correlation near one) or could reflect to what extent a latent process,
such as the social or hygienic behavior of people, drives the more general
infection process. Note that the main transmission route for hepatitis A is
foodborne or feco-oral, and for hepatitis B it is sexual or bloodborne, reflecting
hygienic behavioral conduct of individuals. Moreover, joint infections caused
by more than one pathogen are an aggravating factor in disease progression
for virtually all infections and thus of interest to be quantified. The variable

Table 1.14: Data of five

patients of the hepatitis study
id time status A status B

1 6 0 0
2 17 1 1
3 35 0 1
4 56 0 0
5 12 1 1

id is the identification number of the probands, who are the clusters. Time
refers to age of the probands when the sample was taken and is given in years.
Consequently, time denotes monitoring time and not event time. Status A and
B refer to the presence of antibodies with respect to the hepatitis A and B
virus, respectively. The data of five probands are given in Table 1.14.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com

Chapter 2
Survival Analysis

2.1 Basic Concepts in Survival Analysis

This section describes basic statistical aspects of univariate survival data,
and therefore might be skipped by experienced readers. However, it contains
notation and important results that form the basis of specific points considered
in later chapters. We will use the traditional approach to survival analysis
and refer the reader interested in the approach based on counting processes
and martingale theory (Aalen 1978) to the well-known books of Fleming and
Harrington (1991), Andersen et al. (1993), Martinussen and Scheike (2006),
and Aalen et al. (2008).
First we come back to the questions from the preface: What distinguishes
survival analysis from other fields of statistics? Why does survival data need
a special statistical theory? The reason is that time is the response variable,
and time is not measured like other variables. First, survival times are usually
a mixture of discrete and continuous data that require a different type of
analysis than in the traditional discrete or continuous case. The mixture is
the result of censoring and has an important effect on data analysis. Second,
most evaluations are made conditionally on the knowledge available at the
time of the analysis, and this changes over time. Usually, as the population
under study is changing, we only consider the individual risk of death for
those who are still alive. However, this means that many standard statistical
approaches cannot be applied. This special feature of survival data is called
conditioning and is discussed in detail in the books by Hougaard (2000) and
Duchateau and Janssen (2008).
We consider a random variable T ∗ . Specifically, let T ∗ be nonnegative,
representing the time from a well-defined specific starting point until the
occurrence of an event. If the event is death, T ∗ is the survival time, which
gives the field its name. The term survival time is also used with other
events such as the onset of disease, complications after surgery, healing, or
relapse in the medical field. In demography, death as well as leaving home of
young adults, pregnancy, marriage, giving birth, migration, or divorce are of
special interest. In engineering, the event is typically the failure of a technical
unit, often consisting of different subunits. In economics, it can denote the
acceptance of jobs by the unemployed or early retirement, major investments,

15
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
16 Frailty Models in Survival Analysis

end of strikes or training programs. In consumer economics, durations are

studied until the purchase of a durable or storable product. In macroeconomic
research, the durations of business cycles are of special interest. Considering
convicted criminals, the event of recidivism is of great interest in making
conclusions about the efficiency of punishments. The occurrence of mastitis
in diary cow udders is analyzed in veterinary sciences and is an important
economic factor. In political sciences, events such as international conflicts or
cabinet transitions are considered. We will use survival time, lifetime, event
time, time to event, and duration time as synonyms.
Usually, the event time T ∗ is assumed to follow a continuous distribution,
and we will mainly focus on this case in the present book with only a few
exceptions considered later. All functions of the event time distribution are
defined over the interval [0, ∞) if not stated otherwise in the following. The
probability density function (p.d.f.) is denoted by f . The distribution of
a random variable is completely and uniquely determined by its probability
density function. Other useful functions exist that can be obtained from the
probability density function. The most important one is
Z t
∗
F (t) = P(T ≤ t) = f (s) ds,
0

the cumulative distribution function (c.d.f.) of T ∗ , where P(A) denotes the

probability that event A occurs. In survival analysis, one is more interested
in the probability of an individual to survive time t, which is given by the
survival function
Z ∞
S(t) = 1 − F (t) = P(T ∗ > t) = f (s) ds.
t

The major concept in survival analysis is the hazard function. This function
is also called (depending on the field of application) mortality rate, incidence
rate, mortality curve, failure rate, or force of mortality. The hazard function
is defined by

P(t < T ∗ ≤ t + ǫ|T ∗ > t) f (t)

µ(t) = lim = . (2.1)
ǫ→0 ǫ 1 − F (t)
The hazard function characterizes the hazard of death changing over time (or
age, depending on which time scale is used). It specifies the instantaneous
failure rate at time t, given that the individual survives until time t. It is also
called the exit rate to stress the fact that it means exit out of the state of
interest. Sometimes, it is useful to deal with the cumulative (or integrated)
hazard function
Z t
M (t) = µ(s) ds.
0

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 17

For the topic covered by this monograph, the concept of the Laplace transform
L of a random variable is crucial to inference in this research area:
Z ∞
−uT ∗
L(u) = Ee = e−ut f (t) dt.
0
All the functions f, F, S, µ, M , and L provide equivalent specifications of the
distribution of the nonnegative random variable T ∗ . We use f, F, S, µ, M and
L as generic symbols without index, and their arguments make evident which
random variable is considered.
It is easy to derive relations between the different notions; for example,
(2.1) implies that
Z t Z t
f (s)
M (t) = µ(s) ds = ds = − ln(1 − F (t))
0 0 1 − F (s)
and consequently

Rt
S(t) = 1 − F (t) = e− 0
µ(s) ds
= e−M(t) . (2.2)

This equation is the main exponential formula of survival analysis. It presents

a characterization of the distribution and survival function via the hazard
function. It turns out that the hazard function is often more convenient to deal
with compared to the density or distribution and survival function because
of its meaningful probabilistic interpretation and its simplicity in likelihood
expressions considered later on.

Example 2.1
Suppose that the random variable T ∗ follows a distribution with probability
density function
ν
f (t) = λνtν−1 e−λt , t ≥ 0, (2.3)
where λ, ν are one-dimensional nonnegative parameters. This is a Weibull
distribution, discussed in more detail in the next section. We consider it here
to illustrate the foregoing formulas. The following relations hold, starting
from the density (2.3)

ν
probability density function f (t) = λνtν−1 e−λt
ν
survival function S(t) = e−λt
hazard function µ(t) = λνtν−1
cumulative hazard function M (t) = λtν .

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
18 Frailty Models in Survival Analysis

As mentioned before, the concept of hazard function is particularly useful in

survival analysis. That is, properties of the distribution of event times are
generally discussed in terms of properties of the hazard function. It describes
the way in which the instantaneous probability of failure for an individual
changes with time. Applications often contain qualitative information about
the hazard function, which is helpful in selecting a model with good fit. For
example, there may be biological reasons to restrict the analysis to models
with increasing hazards or with hazard functions that have other well-defined
characteristics.
The hazard function can take very different shapes. It can be increasing
(for example, a Weibull distribution with shape parameter ν > 1), decreasing
(Weibull distribution with ν < 1), constant (Weibull distribution with ν = 1),
J-, U- or bell-shaped (log-normal distribution). Models with these and other
shapes of the hazard function are all useful in practice: In demography, for
example, for following humans from birth to death, a J-shaped hazard function
is often appropriate. After an initial period in which deaths result primarily
from birth defects and infant diseases, the death rate drops and remains
relatively constant until the age of 30 years or so, after which it increases again
with age. This pattern is also observed in many other populations, including
those consisting of technical items. For example, during a run-in phase, the
hazard of failure of technical items often decreases, followed by a period with
relatively constant hazard. After some time the hazard function starts to
increase again. Such pattern can be described, for example, by a U-shaped
hazard function. Models with increasing hazards are used most often. This
is because interest often centers on periods in the life of individuals in which
measurable aging takes place (for example, old ages in humans). Models with
a constant hazard function have a simple structure, as we will see in the next
section. They can be extended to piecewise constant hazard models, which
are much more flexible and easy to handle. Less common are models with
a decreasing hazard, but they are sometimes used in reliability to describe
failure times of electronic devices, at least over a fairly long initial period of
use. Furthermore, mortality rates of patients who have recently undergone a
serious surgery (for example, a transplantation) are usually declining starting
from the surgery. This is because patients experiencing complications often
die shortly after the surgery, leaving the patients without such complications
in the study population.
The main points to remember here are that the hazard function represents
an aspect of the probability distribution of a nonnegative random variable
that has a direct physical meaning, and that qualitative information about the
form of the hazard function is useful in selecting an appropriate (parametric)
model for the situation under investigation.
Furthermore, from a more practical point of view, models based on the
hazard function can easily handle censoring and truncation often occurring in
survival data. These specific and common problems of time to event data are
discussed in the following two sections in more detail.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 19

2.2 Censoring and Truncation

Censoring is what distinguishes survival analysis from other fields of statistics.
Basically, a censored observation is an incomplete observation; it contains only
partial information about event time. That means the patient is followed
up for some time, but the event does not occur during this period. It is
only known that the true event time exceeds the observed censoring time.
There are different types of censoring; here we consider type I right censoring
only because this type of censoring is the most common one in real data
applications of survival analysis.
Let T1∗ , T2∗ , . . . , Tn∗ be i.i.d. survival times with cumulative distribution
function F and let C1 , C2 , . . . , Cn be i.i.d. censoring times with cumulative
distribution function G. Throughout the book, we assume that F , and G
are absolutely continuous. Furthermore, let f and g be probability density
functions with respect to F and G. We are only able to observe the bivariate
data (T1 , ∆1 ), (T2 , ∆2 ), . . . , (Tn , ∆n ), where Ti = min{Ti∗ , Ci } denotes the
observation time and

: if Ti∗ ≤ Ci ,

1 that is, Ti is not censored
∆i =
0 : if Ti∗ > Ci , that is, Ti is censored.

Censoring is a common problem arising in many applications, for example, in

clinical trials. Here, patients may enter the study at different times; then each
is treated with one of several possible drugs or therapies. We are interested
in observing their event times, but for some patients censoring occurs in one
of the following forms:

• Loss to follow-up. The patient may move elsewhere; she or he is never

seen again.

• Drop out. The treatment may have such strong side effects that it is
necessary to stop the therapy. Or the patient may refuse to continue
the treatment.

• Termination of study. The study ends at a predefined point of time.

This type of censoring is called administrative censoring.

• Competing risks. The event of interest cannot be observed because of

the occurrence of a competing event (for example, death by accident).

In Figure 2.1 the event times of patients 2, 4, and 5 are completely observed.
The event times of patients 1 and 3 are censored because of loss to follow-up,
drop out, or competing risks. Event times of patients 6 and 7 are censored
because of termination of study.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
20 Frailty Models in Survival Analysis

patient ż - censored observations

Ƈ - event times

begin of study end of study

Figure 2.1: Event and censoring times of patients in a clinical trial.

In the subsequent part we use T ∗ and C, without subscripts, as shorthand for

all the Ti∗ and Ci lifetime and censoring variables, respectively, and denote by
H the distribution function of the observation time T = min{T ∗, C}. Then
the following relation can easily be derived:

H(t) = P(min{T ∗, C} ≤ t)
= 1 − P(min{T ∗, C} > t)
= 1 − P(T ∗ > t, C > t).
Assuming independence between event time T ∗ and censoring time C implies
the simplification

H(t) = 1 − P(T ∗ > t)P(C > t)

= 1 − (1 − P(T ∗ ≤ t))(1 − P(C ≤ t))
= 1 − (1 − F (t))(1 − G(t)).
This underlines the importance of the independence assumption regarding
event and censoring times usually made in survival analysis.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 21

Remark: The cumulative distribution function of the noncensored data

(discarding the censored observations) is not F !

P(T ≤ t, ∆ = 1) = P(T ∗ ≤ t, T ∗ ≤ C)
Z Z
= f (t∗ )g(c) dt∗ dc
t∗ ≤t,t∗ ≤c
Z  Z 
= f (t∗ ) g(c) dc dt∗
t∗ ≤t t∗ ≤c
Z
= f (t∗ )(1 − G(t∗ )) dt∗ 6= F (t) (2.4)
t∗ ≤t

Consequently, ignoring the censored event times implies biased estimates.

THEOREM 2.1
The probability density function of the survival data (T, ∆) is

δ
1−δ
f (t, δ) = f (t)(1 − G(t)) g(t)(1 − F (t)) . (2.5)

Proof: Denote by H0 and H1 subdistribution functions of the observation

time T , and by h0 and h1 their subdensities. It holds that
Z
H1 (t) = P(T ≤ t, ∆ = 1) = f (t∗ )(1 − G(t∗ )) dt∗ ,
t∗ ≤t

see (2.4). Furthermore,

H0 (t) = P(T ≤ t, ∆ = 0) = P(C ≤ t, C < T ∗ )

Z Z
= f (t∗ )g(c) dt∗ dc
c≤t,c<t∗
Z  Z  Z
∗ ∗
= g(c) f (t ) dt dc = g(c)(1 − F (c)) dc.
c≤t c<t∗ c≤t

dH0 (t) dH1 (t)

Consequently, h0 (t) = dt = g(t)(1 − F (t)), h1 (t) = dt = f (t)(1 − G(t))
and

f (t, δ) = δh1 (t) + (1 − δ)h0 (t)

= h1 (t)δ h0 (t)1−δ

δ
1−δ
= f (t)(1 − G(t)) g(t)(1 − F (t)) .

This completes the proof. 

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
22 Frailty Models in Survival Analysis

In the following paragraph the likelihood function is derived based on results

of Theorem 2.1 given earlier. Here a parametric situation is considered, which
means the distribution of survival time is assumed to be known up to a finite
dimensional unknown parameter vector θ. A parametric model for survival
data is the exponential distribution. It is very rare that survival data follow
this distribution because it is not very flexible as an one-parameter family. But
for understanding all other survival distributions, the exponential model is of
great importance. Other parametric approaches such as the Weibull model
(often used in reliability) and the Gompertz model (preferred in actuarial and
demographic settings) contain the exponential distribution as a special case.
In biostatistical applications the piecewise exponential distribution becomes
more and more popular. We will consider all these, emphasizing similarities
and differences. The distributions are studied in the most simple case of
independent and identically distributed random variables. Furthermore, semi-
parametric and nonparametric approaches will be considered later on.
Here and in the following text, in addition to the independence between
event and censoring times, we assume noninformative censoring. That means
that the censoring distribution must not depend on the parameter vector θ of
the survival distribution. Otherwise the censoring distribution would contain
information on the parameters of interest, which is, for example, the case
in the Koziol–Green Model (Koziol and Green 1976). Furthermore, we do
not aim to estimate parameters of the censoring distribution. Hence, the
terms g(t) and G(t) in the density function (2.5) become additive constants
independent of θ in the log likelihood function. These additive constants drop
from the derivative of the log-likelihood function and can therefore be dropped
from the likelihood of the data. As a consequence of this and formula (2.5),
the contribution of right-censored survival data (ti , δi ) (i = 1, . . . , n) to the
likelihood function simplifies to

Li (θ) = f (ti ; θ)δi S(ti ; θ)1−δi . (2.6)

For complete observations, the likelihood is their density, as is the case in

standard situations. Censored observations provide the information that the
unknown survival time exceeds the observed censored time. Consequently,
the second term in (2.6) reflects the probability to survive at least until ti .
Formula (2.6) can easily be expressed in terms of the hazard function
R ti
Li (θ) = µ(ti ; θ)δi e− 0 µ(s;θ) ds
.

If we consider a sample of independent lifetimes (t1 , δ1 ), . . . , (tn , δn ), the like-

lihood function of the data is
n
Y n
Y R ti
L(θ) = Li (θ) = µ(ti ; θ)δi e− 0 µ(s;θ) ds
(2.7)
i=1 i=1

because of the independence between the observed times (univariate model).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 23

To underline the importance of the common assumption about independent

censoring in survival analysis, the following example is provided, which deals
with the case of dependent censoring. A joint distribution of lifetimes and the
censoring times is assumed. It turns out that the likelihood function in the
case of dependent censoring is the product of derivatives of the joint survival
function of lifetimes and censoring times.

Example 2.2
Denote by (T, ∆), T = min{T ∗ , C}, ∆ = 1(T ∗ ≤ C) censored observations
under the assumption of dependent censoring. Let S(t∗ , c) and f (t∗ , c) be
the joint survival and probability density function of T ∗ and C, respectively.
Consequently, the subdistribution functions needed for the construction of the
likelihood function can be derived by

H1 (t) = P(T ≤ t, ∆ = 1)
= P(T ∗ ≤ t, T ∗ ≤ C)
Z Z
= f (t∗ , c) dc dt∗
{t∗ ≤t,t∗ ≤c}
Z t
=− S1 (t∗ , t∗ ) dt∗
0

with S1 (t, c) = ∂S(t,c)

∂t . This implies that the subdensity of a noncensored
(δ = 1) observation is a derivative of the subdistribution function

dH1 (t)
h1 (t) = = −S1 (t, t).
dt
Similar calculations yield the subdistribution and subdensity functions in the
case of a censored observation (δ = 0):

H0 (t) = P(T ≤ t, ∆ = 0)
= P(C ≤ t, C < T ∗ )
Z Z
= f (t∗ , c) dt∗ dc
{c≤t,c<t∗ }
Z t
=− S2 (c, c) dc
0

with

∂S(t∗ , c) dH0 (t)

S2 (t∗ , c) = and h0 (t) = = −S2 (t, t).
∂c dt
Consequently, the likelihood in case of a parametric model with θ as the
vector of unknown parameters is a composition of the subdensity functions

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
24 Frailty Models in Survival Analysis

S1 (t, t; θ) = S1 (t, t) and S2 (t, t; θ) = S2 (t, t). For a sample (t1 , δ1 ), . . . , (tn , δn )
the likelihood function can be written as

n
Y
δi
1−δi
L(θ) = − S1 (ti , ti ; θ) − S2 (ti , ti ; θ) .
i=1



In the case of independent censoring with S(t, c; θ) = 1 − F (t; θ) 1 − G(c)
this expression simplifies to (2.7).

Up to now, only the situation of right-censored event time data was considered
in the present monograph. However, in some cases, event times are only known
to lie in a specific interval. This situation especially arises when study subjects
are not under continuous observation, such as, for example, patients visiting
their doctor at predetermined times (or times that are convenient to them),
where the occurrence of the event can be diagnosed knowing that the event had
not occurred at the time of the previous visit. Another situation is inspection
times of technical equipment, where events can happen between two inspection
times. Hence, it is only known that the event occurred between two visits
or inspections but not the exact time point. This kind of censoring is called
interval censoring and was considered in detail by Sun (2006). In general, right
censoring is a special case of interval censoring and some of the methods for
right censored data can be directly, or with minor changes, applied to interval
censored data. However, most of the approaches for right-censored data are
not appropriate for interval-censored data because the censoring mechanism
behind interval censoring is much more complicated than in the case of right
censoring.
An important special case of interval-censored data are so called current
status data. The term current status data originates from applications in the
field of demography (Diamond et al. 1986). It means the observation on each
individual survival time interval includes either zero or infinity. Such kind of
data usually occur when each study subject is observed only once, and the
only available information for the event under study is whether the event has
occurred before the observation was taken. Consequently, current status data
are given in the form (T, ∆), where T denotes the monitoring time (which
is not the time when the event happens!) and ∆ is the indicator whether
the event already occurred before the monitoring or not. In the parametric
case, the likelihood function of a sample (t1 , δ1 ), . . . , (tn , δn ) with unknown
parameter vector θ to be estimated can be written in the form (Sun 2006)

n
Y
δi
L(θ) = 1 − S(ti ; θ) S(ti ; θ)1−δi , (2.8)
i=1

with components depending on whether the event already occurred before the
monitoring times (δi = 1) or not (δi = 0).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 25

In the following paragraphs we shall take truncation into account. Truncation

is, besides censoring, another important feature of event time data, requiring
specific adaptation of the parameter estimation procedures. Censoring is a
mapping of the original event times and censoring times, whereas truncation
results in conditional distributions of the observations. Here we discuss only
the most common truncation type, that is, left truncation. Furthermore, let
us assume that truncation is nonrandom. Nonrandom left-truncation occurs
when individuals are observed only from some known time after the natural
origin of the event under study. That means individuals who failed before the
truncation time in question are not recorded.
For example, the first patient in Figure 2.2 cannot be observed because
of death before the study started. That means the researcher carrying out
the study would not be aware of that observation. Patients 2, 3, and 4 are
observable because their observation time T exceed their truncation time t+
(shown in detail for patient 3 in the figure). In other words, truncation is
sampling from a conditional distribution. That means an unknown number of
subjects who experienced the event under study before the observation started
are missing in the study sample.

patient ż - censored observations

Ƈ - event times
lifetime T

truncation time t+
2

lifetime T

truncation time t+

begin of study end of study

Figure 2.2: Event times of patients in a clinical trial with truncation.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
26 Frailty Models in Survival Analysis

Denote by (T + , ∆+ ) the observed truncated data following the distribution of

the original data under the condition (T > t+ ) with t+ as known (nonrandom)
truncation time. So we get

P(T + > t, ∆+ = δ) = P(T > t, ∆ = δ|T > t+ )

P(T > t, ∆ = δ)
=
P(T > t+ )
P(T > t, ∆ = δ)
=
(1 − F (t+ ))(1 − G(t+ ))
and the density of the (nonrandom) left-truncated and independent right-
censored observations is given by the expression

δ
1−δ
f (t)(1 − G(t)) g(t)(1 − F (t))
f (t, δ, t+ ) = .
(1 − F (t+ ))(1 − G(t+ ))
As a consequence of this result, the likelihood function of univariate survival
data (t1 , δ1 , t+ + +
1 ), . . . , (tn , δn , tn ) with (nonrandom) left truncation times ti
(i = 1, . . . , n) in the parametric case with µ(t) = µ(t; θ) and θ as the vector
of parameters to be estimated can be written in the form
n
Y Z ti
µ(ti ; θ)δi exp −


L(θ) = µ(s; θ) ds .
i=1 t+
i

Here, similar to the case of no truncation the distribution of the censoring

times disappears from the likelihood function because of the independent and
noninformative censoring assumption.
Truncation, alternatively called late entry, has enormous consequences for
the approach to time to event data. It turns out that survival function
and density function are changed by truncation at time t+ , which means
by conditioning on T > t+ . This shows the advantage of the hazard function,
which, unlike the other quantities, is not changed by truncation. This is
because the hazard is already conditioned on survival time t. Therefore, it
makes no difference if conditioning is also on smaller survival times than t.
More details can be found in Turnbull (1976).
Truncation is a common problem in register data. Usually, a register is
founded at a specific time point, and data collection starts from this time
onward. Subjects who fail before this date are usually truncated, which means
they are not included in the study population. For example, Swedish twins
are only included in the Swedish Twin Registry if they were alive in 1959 when
the register was established. Information about cancer is available for Danish
people from 1943 onwards, the year in which the Danish Cancer Registry was
founded. This problem will be discussed later on in a bivariate data situation
in more detail.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 27

2.3 Parametric Models

Now we will briefly consider some probability distributions that are helpful
in the field of survival analysis. Naturally, any distribution of nonnegative
random variables can be used to describe durations. In parametric event time
models the baseline hazard is specified by a small number of one-dimensional
parameters. The survival distributions to be discussed here in more detail
are continuous with one exception – the piecewise constant hazard function.
Throughout the literature on survival analysis, certain parametric models
have been used repeatedly, such as exponential, Weibull, and Gompertz.
These distributions have closed-form expressions for survival, density, and
hazard functions. Gamma and lognormal distributions are computationally
less convenient but applied frequently. To avoid model validity issues, the
nonparametric approach, based on the Kaplan–Meier estimator (Kaplan and
Meier 1958), is usually the preferred course. However, this alternative is
often inefficient, as noted by Miller (1983). In particular, standard errors
of parameter estimates in parametric models will tend to be smaller than in
nonparametric models. However, the adequacy of the chosen distribution has
to be checked. The pros and cons of different parametric, semiparametric and
nonparametric models and methodology for statistical inference in event time
data can be found in the monographs by Kalbfleisch and Prentice (1980),
Miller (1981), Lawless (1982), Cox and Oakes (1984), Andersen et al. (1993),
Parmar and Machin (1995), Smith (2002), Klein and Moeschberger (2003),
Collett (2003), Tableman and Kim (2004), and Aalen et al. (2008). Keiding
and Andersen (2006) provide a concise overview of important topics in survival
analysis in the form of an encyclopedia.
Now we discuss some of the parametric standard failure time models for
homogeneous populations. The properties and theoretical bases of these
distributions are considered here only briefly. For a more detailed overview
of lifetime distributions, the interested reader is referred to the book by
Marshall and Olkin (2007). In this and the following sections the random
variable T denotes the time until the occurrence of the event, which we are
interested in making inferences about. Only the simplest case of independent
and identically distributed event times T1 , T2 , . . . , Tn is considered here. How
to overcome this restriction will be the topic of the following chapters. In
general, we will use capital letters for random variables and lowercase letters
for their realizations.
We will start with the exponential distribution in the next section as the
fundamental one in survival analysis even if it is relatively rare that event
time data follow this one-parameter distribution. However, the exponential
distribution helps to understand the main issues of many other typical survival
distributions.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
28 Frailty Models in Survival Analysis

2.3.1 Exponential distribution

The exponential model (T ∼ Exp(λ)) is the simplest parametric lifetime
model. It has only one single parameter λ. The model assumes a constant
risk over time, which reflects the property of the distribution appropriately
called lack of memory. The probability of failure within a particular time
interval depends only on the length but not on the location of this interval.
This means that the distribution of T − t conditional on T > t is the same as
the original distribution. In other words, it holds that

P(t < T ≤ t + ǫ|T > t) = P(T ≤ ǫ)

for any positive ǫ. As a consequence, the exponential distribution as the
only one is not influenced by the definition of time point zero. This property
makes the exponential model a poor choice for modeling human survival or
age at onset of diseases except over short time intervals. The parameter λ
can take all positive values, and the distribution with λ = 1 is called the unit
or standard exponential. Therefore, the following formulas can be derived by
some simple algebraic calculations:

probability density function f (t) = λe−λt (λ > 0)

−λt
survival function S(t) = e
hazard function µ(t) = λ
cumulative hazard function M (t) = λt
1
expectation ET =
λ
1
variance V(T ) = 2
λ
Exponential distribution was widely used in early work on the reliability of
electronic components and technical systems. The distribution of cT with
a positive constant c is again exponentially distributed with parameter λ/c.
The minimum of n independent exponential random variables with parameter
λ is still exponential with parameter nλ:

n
Y n
Y
P(min{T1 , . . . , Tn } > t) = P(Ti > t) = e−λt = e−nλt .
i=1 i=1

Exponential distribution is a special case of all the Weibull, Gompertz, gamma,

and piecewise constant distributions to be described in the following sections.
The simple form of the density implies a very easy and explicit maximum
likelihood estimate for the parameter λ, which is shown in Example 2.3.
It turns out that the maximum likelihood estimate is the reciprocal of the
number of total person years of exposure. Methods in epidemiology based on

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 29

person years in fact implicitly assume an exponential distribution. Software

for planning sample size in clinical trials also often uses the assumption of
exponential failure times.

Example 2.3
Let T1 , T2 , . . . , Tn be independent and identically distributed survival times
with cumulative distribution function F (t) = 1 − e−λt . The question arises:
How do we estimate the unknown parameter λ of interest when using the
maximum likelihood method?

First step: Under the assumption of independent censoring, the likelihood

function can be derived using (2.7) and θ = λ:

n
Y R ti
L(λ) = µ(ti ; λ)δi e− 0 µ(s;λ) ds

i=1
n
Y
= λδi e−λti .
i=1

Taking the logarithm yields

n
X n
X
log L(λ) = log(λ) δi − λ ti .
i=1 i=1
Second step: Maximization of the log-likelihood function with respect to the
unknown parameter λ:

n n
∂ log L(λ) 1X X
= δi − ti
∂λ λ i=1 i=1
n n
1X X
δi − ti = 0
λ̂ i=1 i=1
n
P
δi
λ̂ = i=1
n
P
ti
i=1

The estimator has a straightforward interpretation, as λ̂ is just the ratio of

the number of events and the total observation time.

The model is very sensitive to even a modest variation because it has only
one adjustable parameter. The inverse of parameter λ is both mean and
standard deviation. Recent works have overcome this limitation by using
more flexible distributions, which are introduced in the following sections.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
30 Frailty Models in Survival Analysis

One interesting extension of the exponential model is the piecewise constant

hazard model. In this model the hazard function is constant within some
prespecified time intervals, meaning, that for each time interval, a separate
exponential model is assumed. Consequently, results from the exponential
distribution can be used by considering each interval separately. The main
advantage of this model is its flexibility, depending on the number and length
of time intervals used in the analysis. On the other hand, each additional
interval implies an additional parameter to be estimated, which practically
restricts the number of intervals depending on the sample size. The main
disadvantage of the model is that it is not continuous. There are jumps at
each interval end. One interesting special case is the equidistant case where all
intervals are of the same length. Liu and Huang (2008) recommend the use of
piecewise constant hazard functions especially in frailty models because of the
ease of implementation. Simulations show the very good performance of this
method compared to semiparametric methods. An interesting extension of
the model is the use of piecewise linear functions. At the cost of an additional
slope parameter per time interval, the model can be extended to a continuous
model. Of course, this model is also much more difficult to fit. Further
extensions in this direction deal with splines of higher order that yield a
smooth hazard function.

2.3.2 Weibull distribution

The Weibull model was introduced by Waloddi Weibull (1939) and is a popular
generalization of the exponential model with two positive parameters. The
second parameter ν allows great flexibility to the model and different shapes
of the hazard function. The convenience of the Weibull model is due, on
the one hand, to this flexibility and, on the other hand, the simplicity of the
hazard and survival functions. The expressions of expectation and variance
contain the gamma function and are not of simple form. We abbreviate the
distribution as W (λ, ν). In the case of ν = 1, the exponential distribution is
obtained as a special case. For the Weibull distribution the following formulas
can be derived

ν
probability density function f (t) = λνtν−1 e−λt (λ > 0, ν > 0)
−λtν
survival function S(t) = e
hazard function µ(t) = λνtν−1
cumulative hazard function M (t) = λtν
1 1
expectation ET = λ− ν Γ(1 + )
ν
2 2 1

V(T ) = λ− ν Γ(1 + ) − Γ(1 + )2 ,
variance
ν ν
R∞
where Γ is the gamma function with Γ(k) = 0 sk−1 e−s ds (k > 0).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 31

5 W(2;1.25)
W(2;1.0)
W(2;0.5)

3
hazard

0
0 1 2 3 4 5
time
Figure 2.3: Weibull hazard functions with different shape parameters.

The hazard function decreases monotonously for ν < 1, is constant for ν = 1,

and it monotonously increases for ν > 1, which is illustrated by Figure 2.3.
If T ∼ W (λ, ν), then it holds that cT ∼ W (λc−ν , ν), when c is a positive
constant. Furthermore, the minimum of n i.i.d. random variables is W (nλ, ν)
distributed (minimum-stable distribution property). Weibull distribution can
also be generated as the limiting distribution of the minimum of a sample
from a continuous distribution on [0, u) for some u (0 < u < ∞). Due to
this extreme value property, Weibull distribution is appropriate for modeling
the distribution of individual time to death. Different causes of death compete
with each other, and the first one to strike will cause the death of the individual.
The Weibull model seems to be the most widely applied parametric lifetime
model. It has been used in a wide range of biostatistical problems; for example,
it has been derived theoretically for studies of time until occurrence of cancer
in laboratory animals by Pike (1966) and Peto et al. (1972). In medical
applications time to death by lung cancer is modeled by Weibull distribution
(Whitthemore and Altshuler 1976, Berry 2007). It is unknown whether it has
similar relevance for other diseases. Weibull distribution is inappropriate for
hazard functions that are unimodal or bathtubshaped. A generalization of
Weibull distribution to include such shapes was proposed by Mudholkar et al.
(1996). A comprehensive overview about the Weibull distribution is given in
Murthy et al. (2003).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
32 Frailty Models in Survival Analysis

2.3.3 Log-logistic distribution

An alternative to the foregoing Weibull distribution is log-logistic distribution.
Log-logistic distribution is fairly flexible with two parameters, denoted by
logL(υ, κ). It is one of the parametric survival-time models in which the
hazard is decreasing for κ ≤ 1 and hump-shaped for κ > 1 (Figure 2.4). The
distribution imposes the following functional forms on the density, survival,
hazard, and cumulative hazard function:

υκ(υt)κ−1
probability density function f (t) = (υ > 0, κ > 0)
(1 + (υt)κ )2
1
survival function S(t) =
1 + (υt)κ
υκ(υt)κ−1
hazard function µ(t) =
1 + (υt)κ
cumulative hazard function M (t) = ln(1 + (υt)κ )

The general shape of the hazard function of a log-logistic distribution is

very similar to that of log-normal distribution which will be introduced in a
later section.

4 logL(3;0.5)
logL(3;1)
logL(3;1.5)

3
hazard

0
0 1 2
time

Figure 2.4: Log-logistic hazard functions with different parameters.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 33

2.3.4 Gompertz distribution

In 1825 the British actuary Benjamin Gompertz made the observation that
a law of exponential progression gives a good description of large portions
of different tables of mortality for humans. The simple formula he derived
describes the exponential rise in death rates between sexual maturity and
old age. It is commonly referred to as the Gompertz equation – a formula
that remains a valuable tool in demography and other scientific disciplines.
Gompertz’s observation of a mathematical regularity in human life tables
made him believe in the presence of a law of mortality that explained common
age patterns of death. It has been widely used, especially in actuarial and
biological applications and in demography. A random variable T follows a
Gompertz distribution with parameters λ and ϕ (T ∼ G(λ, ϕ)), if one of the
following relations holds:

λ ϕt
probability density function f (t) = λeϕt e− ϕ (e −1)
(λ > 0)
λ
−ϕ (eϕt −1)
survival function S(t) = e
hazard function µ(t) = λeϕt
λ
cumulative hazard function M (t) = (eϕt − 1)
ϕ

The hazard function is increasing starting from λ at time zero (Figure 2.5). For
parameter values ϕ < 0, the hazard function is decreasing, and the cumulative
hazard converges to the constant −λ/ϕ for t → ∞ so that not all individuals
in the population experience the event under study. This situation is discussed
in more detail in so-called cure models later on in this monograph. Obviously,
exponential distribution is a special case of Gompertz distribution in the case
of ϕ = 0. The Gompertz model (Gompertz 1825) was generalized to the
Gompertz–Makeham distribution (Makeham 1860) by adding a constant c to
the hazard function

µ(t) = λeϕt + c.
Here the additional parameter c describes a nonaging aspect in the study
population that is independent of time t, whereas the Gompertz part in the
formula still represents the age-dependent aspect with an exponential form.
The parameters λ and c are not identifiable in the case of ϕ = 0; only their
sum can be estimated.
The Gompertz–Makeham distribution describes the age dynamics of human
mortality rather accurately in the age range of about 30 – 80 years. At more
advanced ages the death rates do not increase as fast as predicted by this
mortality law – a phenomenon known as the late-life mortality deceleration.
This phenomenon was one of the starting points for developing univariate
frailty models.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
34 Frailty Models in Survival Analysis

20 G(0.1;0.05)
G(0.1;0.08)
18 G(0.1;0.10)

16
14
12
hazard

10
8
6
4
2
0
0 5 10 15 20 25 30 35 40 45 50
time

Figure 2.5: Gompertz hazard functions with different parameters.

2.3.5 Log-normal distribution

In the log-normal model (T ∼ logN(m, s2 )), the natural logarithm log(T ) (or
more exactly ln(T )) of event time T is assumed to be normally distributed with
mean m and variance s2 . The product of a large number of independent and
identically distributed random variables follows asymptotically a log-normal
distribution. This is similar to the fact that a normal distribution results
if a variable is the sum of a large number of independent and identically
distributed random variables (Central Limit Theorem). The following relations
hold for log-normal distribution:

1 (log t−m)2
probability density function f (t) = √ e− 2s2
2πst
log t − m
survival function S(t) = 1 − Φ( )
s
1
φ( log t−m )
hazard function µ(t) = st s
log t−m
1 − Φ( s )
s2
expectation ET = em+ 2

2 2
variance V(T ) = e2m+s (es − 1)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 35

2 logN(0;0.5)
logN(0;1)
logN(0;1.5)
hazard

0
0 1 2 3 4 5
time
Figure 2.6: Log-normal hazard functions with different parameters.

x2
In the formulas used above φ(x) = √12π e− 2 denotes the probability density
function, and Φ(x) the cumulative distribution function, of the standard
normal distribution N (0, 1).

Log-normal distribution may be convenient to use with noncensored data,

but when this distribution is applied to censored event times computations
quickly become challenging. Unfortunately, the hazard function has a strange
form: it takes value zero at t = 0, increases to a maximum, and then decreases,
approaching zero as t heads to infinity (Figure 2.6). Because of the decreasing
form of the hazard function for older ages, the distribution seems implausible
as a lifetime model in most real situations. Nevertheless, it makes sense if
interest is focused on time periods of younger ages. Despite its unattractive
features, log-normal distribution has been widely used as failure distribution in
diverse situations such as the analysis of electrical insulation or for modeling
repair times in engineering systems. In medicine log-normal distribution is
often used for modeling latent times, for example, time from infection until
the occurrence of the first symptoms of infectious diseases (Sartwell 1966).
Around two third of the 86 examples reviewed by Kondo (1977) in this research
field appear to follow approximately a log-normal distribution.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
36 Frailty Models in Survival Analysis

2.3.6 Gamma distribution

Gamma distribution is another exciting extension of exponential distribution.
It is of limited use in survival analysis because the gamma models do not have
closed-form expressions for survival and hazard functions. Both include the
incomplete gamma integral
R x k−1 −s
s e ds
Ik (x) = 0 .
Γ(k)
Consequently, traditional maximum likelihood estimation is not simple and
requires the calculation of incomplete gamma integrals, imposing numerical
problems in parameter estimation. If variable T follows a gamma distribution
with shape parameter k and inverse scale parameter λ (T ∼ Γ(k, λ)), then
the following relations hold

λk tk−1 e−λt
probability density function f (t) = (k > 0, λ > 0)
Γ(k)
survival function S(t) = 1 − Ik (λt)
λk tk−1 e−λt
hazard function µ(t) =
(1 − Ik (λt))Γ(k)
k
expectation ET =
λ
k
variance V(T ) = 2
λ
u
Laplace transform L(u) = Ee−T u = (1 + )−k
λ
If k = 1, gamma distribution is reduced to exponential distribution. With
integer k, gamma distribution is often called a special Erlangian distribution.
It can be derived as the distribution of the waiting time until the kth emission
from a Poisson source with intensity parameter λ. Consequently, the sum of k
independent exponential variables with parameter λ has a gamma distribution
with parameters k and λ (see Example 2.4) and can be used to model lifetimes
of technical systems with repeated repairing after failure.

Example 2.4
Let T1 , T2 , . . . , Tk denote k i.i.d. random variables with Ti ∼ Exp(λ) and
introduce T by T = T1 + . . . + Tk . Then it holds that

k k
Y Y u −1 u
L(u) = Ee−T u = Ee−(T1 +...+Tk )u = Ee−Ti u = (1 + ) = (1 + )−k ,
i=1 i=1
λ λ
which is the Laplace transform of a gamma distribution with parameters k
and λ. Here T1 , T2 , . . . , Tk denote times between repairing.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 37

2.3.7 Pareto distribution

Pareto distribution was introduced by Vilfredo Pareto (1897) to explain the
distribution of income of a given population at the end of the 19th century.
The distribution is skewed and heavy-tailed with two parameters ω > 0 and
ζ > 0. The kth moment of the Pareto distribution is finite if the restriction
ζ > k applies. The hazard is monotone declining. Here we consider a shifted
version of the Pareto distribution.

ζ  ω ζ+1
probability density function f (t) = (ω > 0, ζ > 0)
ω ω+t
 ω ζ
survival function S(t) =
ω+t
ζ
hazard function µ(t) =
ω+t
ω
cumulative hazard function M (t) = −ζ log( )
ω+t
ω
expectation ET = (ζ > 1)
ζ −1
Pareto distribution can be seen as a gamma mixture of exponential distributed
lifetimes. It is often applied in areas including city population distributions,
stock price fluctuations, oil-field locations, and socioeconomic studies. Pareto
distribution is also a standard distribution for the purposes of reinsurance,
taking care of the largest claims of a portfolio. It can be extended to the
ζ
generalized Pareto distribution with hazard function µ(t) = ω+t + c. More
details about the generalized Pareto distribution can be found in Davis and
Feldstein (1979).

Example 2.5
We would like to derive the expectation of Pareto distribution under the
assumption ζ > 1.
Z ∞
ET = tf (t) dt
0
Z ∞
= S(t) dt
Z0 ∞ 
ω ζ
= dt
0 ω+t
ω  ω ζ−1 ∞
=−
ζ−1 ω+t

0
ω
=
ζ−1
Here assumption ζ > 1 is important, otherwise the expectation is infinite.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
38 Frailty Models in Survival Analysis

2.4 Estimation of Survival and Hazard Functions

In the case of parametric inference, it is necessary to make assumptions about
the distribution of failure times. In some circumstances this makes sense,
especially when additional information about the nature of the underlying
aging or disease process is available. If such information is not available,
it is common to use nonparametric models. The simplest nonparametric
estimate of a distribution function is the empirical distribution function. That
means, even in the case of a continuous distribution, the estimate is discrete.
Major steps in the development of appropriate nonparametric and semipara-
metric techniques in survival analysis with censored observations were the
introduction of the Kaplan–Meier estimator (Kaplan and Meier 1958) and its
extension to the case of truncated lifetimes (Turnbull 1976, Tsai et al. 1987).
Another important milestone in this field was the semiparametric proportional
hazards regression model (Cox 1972).
Whether to use a parametric or a nonparametric model is an important
question. An advantage of nonparametric models is their flexibility and the
resulting ability to deal with any probability distribution. However, there is a
high price to pay. First, nonparametric methods need much more data to get
reasonable results. Second, it is hard to get estimates of the hazard function,
which is often an interesting and relevant information. For this, it is necessary
to smooth out the discrete point masses of the Kaplan–Meier estimator, for
example, by kernel function smoothing. In contrast, parametric models often
allow closed-form expressions of the hazard and survival function depending
on the chosen model. Parametric models can be described by the values of a
few parameters. They often give good results even in the case of small sample
size. If the assumed model is correct, the estimation is more efficient than in
a nonparametric estimation procedure.

2.4.1 Kaplan–Meier estimator

A useful way of characterizing survival in a group of individuals is to compute
and graph the empirical survival function. If there are no censored observa-
tions in the sample, the empirical survival function at time t is the ratio of
survivors at time t and the sample size n. This step function decreases by
1
n just after each observed failure (for ease of presentation we first assume
no ties here). In practical situations, however, some of the observations are
censored. Consequently, a methodology for handling this with convenience is
required. Remember that we observe the pairs (T1 , ∆1 ), . . . , (Tn , ∆n ), where
Ti = min{Ti∗ , Ci } and ∆i = 1(Ti∗ ≤ Ci ). Let T(1) < T(2) < . . . < T(n) be the
order statistics of T1 , T2 , . . . , Tn , and with an abuse of notation define ∆(i) to
be the value of ∆ that is associated with T(i) , that is, ∆(i) = ∆j if T(i) = Tj .
Note that the ∆(1) , ∆(2) , . . . , ∆(n) are not ordered.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 39

The Kaplan–Meier estimator (sometimes called product limit estimator) was

introduced by Kaplan and Meier (1958) as
Y  δ(i) 
Ŝ(t) = 1− .
i:t(i) <t
n−i+1

Assuming a continuous event time distribution, only one death will be observed
at most at each time ti with probability one. In practice, ties can occur, for
example, because of coarse measurement. To deal with ties, let ti denote the
r distinct observation times t1 < t2 < . . . < tr with r ≤ n. Taking this into
account, the Kaplan–Meier estimator is
Y  di 
Ŝ(t) = 1−
#R(ti )
i∈R(t)

with di the number of events at time ti . R(t) denotes the set of indices of all
individuals at risk at time t, meaning all individuals alive just before t. #R(t)
denotes the number of individuals in the risk set at time t. The Kaplan–Meier
estimator is a decreasing step function, changing only at time of an event. A
problematic point is that Ŝ is not defined after the largest observation time
if the last observation is a censored one. In this case, Ŝ(t) is usually left
unspecified after the largest observation time. One consequence of this is
that the mean lifetime cannot be estimated. A solution to this problem is to
assume that the survival function is zero after the largest time, which results
obviously in a biased estimate. A better solution is to consider the median
survival time.
Figure 2.7 illustrates the Kaplan–Meier method by presenting the estimates
of the survival curves of different patient groups in the EBCT study presented
in detail in Example 1.1. Here time from baseline investigation of patients
with coronary artery disease until the occurrence of a major adverse cardiac
event is the duration of interest. Consequently, in this application the notion
of survival is used in the more general sense of being free of well defined
major cardiac events. The patients were grouped according to four levels of
evidence of coronary artery disease. This allows a comparison of the event
times in these groups. It is easy to see that group is a very strong predictor
of event time; for example, individuals presenting with the lowest evidence
level in group 1 have around 98% chance to be event free compared to only
50% chance in group 4, which is the group of highest level of disease evidence.
Furthermore, jumps can be found at each event time point in the survival
curve estimates.
Throughout the book we assume continuous survival times preventing ties.
However, in many real-data applications, the researcher will be faced with
observations at the same time point. Different methods exist to handle
ties, which can produce different results. The exact method computes the
exact conditional probability under the proportional hazards assumption for

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
40 Frailty Models in Survival Analysis

1.0 group 1
group 2
group 3
0.9 group 4

0.8
survival

0.7

0.6

0.5

0.4
0 10 20 30 40 50
months

Figure 2.7: Kaplan–Meier curves for different patient groups in the EBCT
study.

all tied events occurring before censored observations of the same or larger
time. This is equivalent to summing all terms of the marginal likelihood
that are consistent with the observed data. The method assumes that ties
are due to lack of precision in measuring survival times and computes all
possible orderings of tied event times (Kalbfleisch and Prentice 1980). The
exact method needs abundant of computer resources for large data sets with
many ties. The discrete method assumes that events occurred at exactly
the same time and computes probabilities for events occurring to a set of
observations with tied event times. It consumes less computer resources than
the exact method. The most common method to handle ties is that by Breslow
(1974) using an approximate likelihood. Another method suggested later by
Efron (1977) also uses an approximate likelihood. If ties are not extensive,
the methods by Breslow and Efron provide satisfactory approximations to
the exact method for the continuous time-scale model. In general, Efron’s
approximation gives results that are much closer to the exact method results
than Breslow’s approximation does. If there are no ties, all three methods
result in the same likelihood and yield identical estimates. The Breslow
method is the most efficient one when there are no ties. With the exception
of the R package, nearly all statistical software use the Breslow method as
default. Throughout the book we have used this method.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 41

For the variance of the Kaplan–Meier estimate, the Greenwood formula (Green-
wood 1926) given by the expression
X di
V(Ŝ(t)) = Ŝ 2 (t)
#R(ti ) − di
i∈R(t)

is commonly used. Peterson (1977) shows that the Kaplan–Meier estimator

is consistent, and Breslow and Crowley (1974) show its asymptotic normality.
Efron (1967) suggested an interesting procedure to obtain the Kaplan–Meier
estimator (called redistribute-to-the-right algorithm): Start with an ordinary
estimate of S, which assumes point mass n1 at each observed time (survival
or censoring times). Start at the first observation t(1) , go to the right to
the first censored observation, and redistribute the mass of this observation
equally among the following times. Now go to the next censored observation
and distribute its mass again among the following observed times. Treating
the other censored times in a similar manner results in the Kaplan–Meier
estimator.

2.4.2 Nelson–Aalen estimator

Consider the problem of estimating the cumulative hazard function M (t).
The Nelson–Aalen estimator was suggested by Nelson (1972) in a reliability
context and rediscovered by Aalen (1978) who derived the estimator using
modern counting processR techniques. It is an estimator of the cumulative
t
hazard function M (t) = 0 µ(s) ds and of the form
X di
M̂ (t) =
#R(ti )
i∈R(t)

with notations similar to the Kaplan–Meier estimator in the previous section.

Intuitively, this expression is estimating the hazard at each distinct death
time as the ratio of the number of deaths to the number at risk. Cumulative
hazard is just the sum of all hazards at all death times up to time t, and has a
nice interpretation as the expected number of deaths in (0, t] per unit at risk.
The variance of the Nelson–Aalen estimate can be evaluated by (Aalen 1978)
X di
V(M̂ (t)) = .
(#R(ti ))2
i∈R(t)

Figure 2.8 shows Nelson–Aalen plots of the four different levels of evidence
for coronary artery disease in the EBCT study of Example 1.1. In agreement
with the Kaplan–Meier curves, patients of group 1 show a good prognosis,
whereas patients in groups with higher levels of evidence of the disease are
faced with an increasing risk of major cardiac events.
The Nelson–Aalen estimator is often used in choosing between different
parametric models. For this, one plots the estimator using a transformed

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
42 Frailty Models in Survival Analysis

0.7 group 1
group 2
group 3
0.6 group 4

0.5
cumulative hazard

0.4

0.3

0.2

0.1

0.0
0 10 20 30 40 50
months
Figure 2.8: Nelson-Aalen curves for different patient groups in the EBCT
study.

scale so that, if a given parametric model fits the data, the resulting graph
should be approximately linear. For example, a plot of M̂ (t) versus t will be
approximately linear in the case of an exponential distribution. Checking the
adequacy of a Weibull model, ln(M̂ (t)) versus ln(t) should be approximately
linear. The event times of the different groups in Figure 2.8 follow the
exponential model reasonably, with some deviations at later times in group 4.
Similar to the Kaplan–Meier plots, there are only jumps at event times. In
the case of continuous event times the functions S(·) and M (·) are related
by S(t) = e−M(t) (see (2.2)). Consequently, it seems reasonable to consider
S̄(t) = e−M̂(t) as an estimator of the survival function, which is different from
the Kaplan–Meier estimator and was suggested by Breslow in the discussion
of the paper by Cox (1972). Fleming and Harrington (1991) recommend
it as an interesting alternative and have pointed out that it has a slightly
smaller mean-squared error in some situations. Estimators for the hazard
function µ can be derived from the Nelson–Aalen estimator based on kernel
estimators. The formal derivations of the Kaplan–Meier and Nelson–Aalen
estimators and their asymptotic properties are performed in the framework of
counting processes (Aalen 1978, Fleming and Harrington 1991, Andersen et
al. 1993, Martinussen and Scheike 2006, Aalen et al. 2008). Both estimators
are asymptotically equivalent and quite close to each other, particularly when
the number of deaths is small relative to the number of individuals at risk.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 43

2.5 Regression Models

2.5.1 Proportional hazards model
The models presented so far deal with the simplest case of independent and
identically distributed variables. This implies a homogeneous population.
However, in most practical applications, the population under study is not
homogeneous. For example, individuals in epidemiological studies may differ
in age, gender, socioeconomic status, education, blood pressure, body mass
index, smoking habits, nutrition, physical activity, heart rate, family status,
genetic predisposition, and other factors. Perhaps some of these covariates are
of special interest, such as the effect of a treatment in a clinical trial, or they
are confounders which effect needs to be adjusted for in the analysis. In both
cases, we will use the notion covariate for these variables. The proportional
hazards model introduced by Cox (1972) is a regression model with event time
as dependent variable. It allows the inclusion of information about known
(observed) covariates in models of survival data in an easy way and is the
most applied model in this area.
Let µ(t|X) denote the hazard of an individual at time t with covariate vector
X′ = (X1 , . . . , Xk ). Here X′ denotes the transpose of the column vector X.
The proportional hazards model specifies that

µ(t|X) = µ0 (t)h(X) (2.9)

where µ0 (t) is the baseline hazard function and h(·) some positive function.
The model assumes a baseline hazard (risk of death or other event) that all
individuals in the study population have in common. This assumption can be
relaxed but keeps presentation easy. The parameters of primary interest are
contained in h(X) = h(β, X), often
′
h(X) = eβ X ,
with β ′ = (β1 , β2 , . . . , βk ) denoting the vector of regression parameters. In
this model, the covariates act multiplicatively on the baseline hazard, adding
further risks, as determined by the individuals’ prognostic information. This
allows the model a simple and easy interpretation. It is assumed that all
individual variation in the hazard can be characterized by a finite-dimensional
vector of observed covariates (explanatory variables, risk factors, predictive
factors, and regressors). The main idea behind it is the separation of the time
effect in the baseline hazard function on the one hand and the effect of the
covariates in an exponential term on the other. In essence, this assumption
says that the hazards of two individuals at time t is related by a proportionality
constant that does not depend on t. The simple two-sample situation is
obtained by restricting to a single (k = 1) binary covariate X in the model with
X = 0 or X = 1, depending on group membership. In this case, the method is

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
44 Frailty Models in Survival Analysis

truly nonparametric, and eβ denotes the hazard ratio between the two groups.
However, if X is continuous, a parametric form of h(·) is required. Inference
is now dependent on that parametric form but still independent of µ0 (t), and
the model is called a semiparametric model because of the parametric nature
of the covariate term and the nonparametric baseline hazard function. The
survival function given the covariates X is
β′X
S(t|X) = S0 (t)e ,
Rt
where S0 (t) = e− 0 µ0 (s) ds denotes the baseline survival function and the
components of the vector β are unknown regression parameters. That means
the survival function of an individual with covariate vector X is a power of
the baseline survival function. The class of distributions generated by this
procedure is sometimes called Lehmann alternatives.
Two different approaches are possible with the proportional hazards model.
Sometimes it occurs that covariates have skewed distributions, for example,
when only a small fraction of the individuals are exposed to the risk factor of
interest. It is also very common that a large fraction of lifetimes is censored.
Especially in large cohort studies analyzing the effect of a rare exposition
on an event, the number of exposed cases may be very small. One may
then question the validity of inference based on asymptotic results. In the
parametric case, the baseline hazard is chosen from the class of parametric
lifetime distributions. For example, starting from (2.7), the likelihood function
in the Weibull model with θ = (λ, ν) is of the form
n
Y ′
δi ν β ′ Xi
L(β, θ) = λνtν−1
i eβ Xi e−λti e (2.10)
i=1

and looks like

n ′
Y ′
δi λ ϕti
−1)eβ Xi
L(β, θ) = λeϕti eβ Xi e− ϕ (e (2.11)
i=1

with θ = (λ, ϕ) in the Gompertz model.

Example 2.6
Table 2.1 gives the results for the parametric proportional hazards model
with the Weibull, exponential, and Gompertz baseline hazard, respectively,
analyzing the Halluca data from Example 1.3.
Analysis was performed using PROC NLMIXED in SAS by maximizing the
likelihood function given (2.10) and (2.11). The likelihood in the exponential
model is obtain from (2.10) with ν = 1. Both the Weibull and the Gompertz
model provide a much better fit compared to the exponential model based
on the values of the log-likelihood function. The cost of the better fit is
an additional parameter. The exponential model is nested in the Weibull

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 45
Table 2.1: Parametric proportional hazards models for
Halluca data.
parameter Weibull exponential Gompertz

age (in years) 0.010 (0.003) 0.010 (0.003) 0.010 (0.003)

sex (females) -0.166 (0.071) -0.173 (0.071) -0.151 (0.071)
type (NSCLC) -0.136 (0.067) -0.162 (0.066) -0.114 (0.067)
ECOG (3 & 4) 0.612 (0.103) 0.619 (0.104) 0.595 (0.104)
stage II 0.455 (0.168) 0.431 (0.172) 0.418 (0.171)
stage IIIa 0.606 (0.131) 0.661 (0.130) 0.596 (0.132)
stage IIIb 0.970 (0.120) 1.043 (0.120) 0.971 (0.120)
stage IV 1.358 (0.110) 1.466 (0.109) 1.357 (0.110)
λ 0.029 (0.007) 0.018 (0.004) 0.023 (0.006)
ν 0.859 (0.019)
ϕ 0.022 (0.003)
log-likelihood -4858.672 -4883.351 -4858.707

as well as Gompertz models. Consequently, the two models can be com-

pared by means of the likelihood ratio test, which yields a test statistic
χ2 = −2(−4883.351+4858.672) = 49.358 (Weibull vs. exponential model) and
χ2 = −2(−4883.351 + 4858.707) = 49.288 (Gompertz vs. exponential model).
Both test statistics follow asymptotically an χ2 -distribution with one degree of
freedom, resulting in p-values less than 0.001. Hence, the additional parameter
in the Weibull and Gompertz distributions cause a significant improvement
in the fit to the data. The Gompertz model show a similar value of the log-
likelihood function compared to the Weibull model, with both models having
the same number of parameters.
Age (in years) is the only continuous variable in the model. An increase
in age by 10 years results, for an example, in an increased risk of e0.1 =1.1.
Females show a reduced risk of e−0.166 =0.85 compared to males (using the
results from the Weibull model). Patients with non-small-cell lung carcinoma
have a reduced risk of e−0.136 =0.87 compared to patients with small-cell lung
carcinoma. Patients with bad performance status (ECOG 3 or 4) at baseline
face a higher risk (e0.612 =1.84) compared to patients with ECOG status less or
equal to 2. We see an increasing risk with higher disease stage. For example,
the hazard ratio of patients in stage IV is e1.358 = 3.89 times higher compared
to patients with stage I who serve as reference group.

The parameter estimates depend on the parametric assumption about the

baseline hazard. This limits the applicability of the parametric proportional
hazards model because more detailed investigations about the shape of the
baseline hazard are necessary. The Weibull model is the only one that is
a proportional hazards model as well as an accelerated failure time (AFT)
model. Parameters can easily be transformed from one model to the other
and back. Similar to the proportional hazards model the AFT model allows

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
46 Frailty Models in Survival Analysis

an easy and natural interpretation of the parameter estimates and is discussed

in Section 2.5.2. However, the main focus of the present book is on extensions
of the proportional hazards model, the AFT model is considered here for
completeness only.
It turns out that the parametric approach is much more important in frailty
models than in proportional hazards models without random effects because
of additional technical estimation problems in the frailty approach. Other
choices are of course possible, but in the sequel we will mainly use the Weibull
and Gompertz baseline hazard functions in parametric models.
In the common semiparametric case (without parametric specification of the
baseline hazard function), the model is natural and sufficiently flexible to suit
′
many purposes. Since eβ X is always positive, the individual hazard µ(t|X)
is automatically nonnegative for all t and all values of β. One additional
reason for considering this model is that censoring and competing risks are
easily accommodated within this formulation. Since the likelihood contains
the unspecified hazard function, the use of the traditional maximum likelihood
method becomes problematic. Therefore, we need an adapted version of the
likelihood that contains sufficient information about the parameter vector β
but not the unspecified baseline hazard. Cox (1972, 1975) suggested using a
partial likelihood approach. In this particular case, the technical problems of
statistical inference have a simple solution if the baseline hazard is considered
an infinite-dimensional nuisance parameter. Denote the observed data with
sample size n by (ti , δi , Xi ) (i = 1, . . . , n) and assume no ties for ease of
presentation. The likelihood of the survival data is given in (2.7) and can now
be written in the form
n
Y ′
δi ′
µ0 (ti )eβ Xi exp − M0 (ti )eβ Xi .


(2.12)
i=1

Based on arguments from nonparametric maximum likelihood estimation, a

discrete version of the baseline hazard µ0 (·) is assumed, which is zero except
for the times ti at which an event takes place (similar to the Nelson–Aalen
estimator). This leads to a discrete version of the cumulative baseline hazard
function
X
M0d (t) = µ0 (tj ).
tj ≤t

Plugging this discrete cumulative baseline hazard function into the likelihood
(2.12) results in
n 
Y ′ X ′
δi
µ0 (ti )eβ Xi exp − µ0 (ti ) eβ Xj , (2.13)
i=1 j∈R(ti )

where R(t) denotes the risk set at time t containing all individuals (more
exactly their indices) that are still at risk of experiencing the event of interest

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 47

at time t. This is a modified likelihood function that can be used to find

estimates of the (discrete) baseline hazard function. Taking the logarithm
and the partial derivative with respect to µ0 (ti ) of expression (2.13) we get
1 X ′
− eβ Xj .
µ0 (ti )
j∈R(ti )

Equating these partial derivatives to zero implies

δi
µ̂0 (ti ) = P ′ . (2.14)
eβ Xj
j∈R(ti )

Plugging in this solution into expression (2.13), a likelihood function for the
regression parameters is obtained:
n  ′
Y eβ Xi δi
L(β) = P β′ X . (2.15)
i=1
e j
j∈R(ti )

This expression is called partial likelihood and does not depend on the baseline
hazard µ0 (t), which simplifies parameter estimation. It is used to estimate
the regression coefficients in the semiparametric proportional hazards model.
The second derivative of the partial likelihood function is well behaved in the
sense that the negative of the second derivative is always positive definite (or
semidefinite). This is in contrast to the behavior of general likelihoods, which
are only known to fulfill this property locally around the parameter estimate.
The second derivative of the partial likelihood can be used to evaluate the
(asymptotic) variance.
Inference for the Cox estimator is almost exclusively based on asymptotic
results (Andersen and Gill 1982). The validity of these large sample properties
have been found acceptable with moderately large sample sizes, moderate
amount of censoring, and balanced covariates. This semiparametric model
is the most often applied one in survival analysis. It is implemented in all
statistical packages, is very easy to handle, and results allow an easy and
intuitive interpretation.
Expression (2.14) forms the basis for the well-known Breslow estimator of
the cumulative hazard function
n
X 1(ti ≤ t)δi
M̂0 (t) = P β′ X .
i=1
e j
j∈R(ti )

The Breslow estimator of the cumulative hazard function can be used to

create a corresponding estimator for the survival function. It is based on
the exponential formula (2.2)
β′ X
Ŝ(t|X) = e−M̂0 (t)e .

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
48 Frailty Models in Survival Analysis

Hypothesis testing in the Cox model for a single regression parameter, usually
for testing the null hypothesis H0 : βj = β0 (with β0 as specified value, often
zero) versus the alternative hypothesis HA : βj 6= β0 , can be performed in
three different ways. The first one is a likelihood ratio test based on the
partial likelihood L(·), treating it as an ordinary likelihood function. Under
the null hypothesis, the test statistic
 L(β0 ) 
T = −2 log
L(β̂j )

is asymptotically χ2 distributed with one degree of freedom. Here β̂j denotes

the maximum likelihood estimate of regression parameter βj . Alternatively, if
several regression parameters need to be tested jointly in one test procedure,
the test statistic is again asymptotically χ2 distributed and the number of
parameters tested defines the number of degrees of freedom.
A simple alternative to the likelihood ratio test is the Wald test with test
statistic

 β̂ − β 2
j 0
T = ,
se(β̂j )

which is also asymptotically χ2 distributed with one degree of freedom under

the null hypothesis. Here, the standard error se of the parameter estimate
is found from the inverse of the second derivatives of the partial likelihood
function.
A third but less frequently used possibility for parameter testing is the score
test with test statistic

l12
T =−
l2
where

∂ log L(β)
l1 = |βj =β̂j
∂βj

and

∂ 2 log L(β)
l2 = |βj =β̂j
∂βj2

are the first- and second-order derivatives of the logarithm of the partial
likelihood function under the alternative hypothesis. Similar to the other
two tests the test statistic follows under the null hypothesis asymptotically a
χ2 distribution with one degree of freedom.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 49

Example 2.7
In the following the Cox model is applied to the data from the Halluca study.
In Table 2.2 the parameter estimates of the (semiparametric) Cox model are
given. For comparison, the parameter estimates from the Weibull proportional
hazards model are given as well. Despite the fact that the baseline hazard in

Table 2.2: Proportional hazards models for

Halluca data.
parameter Weibull PH Cox model

age (in years) 0.010 (0.003) 0.009 (0.003)

sex (females) -0.166 (0.071) -0.162 (0.071)
type (NSCLC) -0.136 (0.067) -0.125 (0.067)
ECOG (3 & 4) 0.612 (0.103) 0.610 (0.104)
stage II 0.455 (0.168) 0.455 (0.169)
stage IIIa 0.606 (0.131) 0.594 (0.131)
stage IIIb 0.970 (0.120) 0.947 (0.120)
stage IV 1.358 (0.110) 1.345 (0.110)
λ 0.029 (0.007)
ν 0.859 (0.019)

the parametric approach is modeled by a Weibull distribution, the parameter

estimates and standard errors in both models are very similar. The inter-
pretation of the parameters is similar in both models, and the hazard ratios
eβ are usually interpreted as relative risks. Estimates in the Weibull model
depend on the correctness of the assumption of the baseline hazard.

The partial likelihood depends only on the ranking of the observations, that
is, the order in which the events happen, but not on the actual times. This can
be illustrated by varying observations, keeping the ranks unchanged. Then
there is no change in the parameter estimate, but if the ranks are changed,
the estimate is changed. In this sense, the parameter estimator is a step
function and not a continuous function. This underlines the importance of
measuring event times as accurately as possible (for example, in days) to get
the ranks right and to prevent tied observations. If this is impossible because
of important measurement error in the event times (as with cancer diagnosis),
it might be preferable to use a parametric Cox model to avoid discontinuities
in the likelihood function.
Samuelsen (2003) investigates advantages and limitations of exact inference
in the proportional hazards regression model and compares it with logistic and
conditional logistic regression.
Note that covariate vector X could vary with time, but this is beyond the
scope of this book.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
50 Frailty Models in Survival Analysis

2.5.2 Accelerated failure time model

The second important regression model in survival analysis is the accelerated
failure time (AFT) model (Lawless 1982, Kalbfleisch and Prentice 2002). The
Cox model and its various generalizations are mainly used in the medical and
biostatistical field, while the AFT model is primarily applied in reliability
theory and industrial experiments and as an alternative if the proportional
hazards assumption does not hold. The hazard of the AFT regression model
can be written in the form
′ ′
µ(t|X) = µ0 (teβ X )eβ X
, (2.16)
with µ0 (·) as baseline hazard. Due to computational difficulties AFT models
are mainly used based on parametric approaches with log-normal, gamma, and
inverse Gaussian baseline hazards. In contrast to the proportional hazards
model, the AFT model can best characterized and interpreted in terms of the
survival function. Assuming a model with only one single binary covariate
X (for example, indicating treatment (X=1) and control group (X=0) in a
randomized clinical trial) for the survival function, the relation

S(t|X = 1) = S(eβ t|X = 0)

holds. The hazard function (2.16) of an individual in the control group is then
the baseline hazard µ(t|0) = µ0 (t). The factor eβ is called acceleration factor,
meaning that the probability to survive time point t in the treatment group
is similar to the probability to survive time point teβ in the control group.
This means, for example, that the median survival time of a patient in the
treatment group is eβ times that of a patient in the control group. In fact, the
same argument holds for any percentile of the event time distribution. Hence,
individuals age eβ -times faster in some biological sense in the control group
compared to the treatment group, which allows a very simple and, especially
for clinicians, attractive and natural interpretation of the model parameters.
The AFT model will not be treated in detail here in this book, but we
would like to mention that these models offer a number of advantages. In
particular, they provide a wide variety of shapes of baseline hazard functions
since the family includes unimodal hazards such as the one in the log-normal
distribution. Furthermore, the loglinear formulation of such models underlines
that the roles of regression and dispersion parameters are clearly separated
(Keiding et al. 1997).

Example 2.8
In the present analysis, PROC LIFEREG from SAS is used to apply Weibull,
exponential, and log-logistic AFT models to the Halluca lung cancer data from
Example 1.3. The results are given in Table 2.3. As already mentioned in
the paragraph about the parametric proportional hazards model, the model
with a parametric Weibull baseline hazard function is also an AFT model,

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 51
Table 2.3: Parametric AFT models for Halluca data
parameter Weibull exponential log-logistic

age (in years) -0.011 (0.003) -0.010 (0.003) -0.014 (0.004)

sex (females) 0.193 (0.083) 0.178 (0.071) 0.182 (0.092)
type (NSCLC) 0.161 (0.078) 0.159 (0.067) 0.048 (0.088)
ECOG (3 & 4) -0.705 (0.121) -0.638 (0.103) -1.210 (0.150)
stage II -0.532 (0.196) -0.488 (0.169) -0.545 (0.201)
stage IIIa -0.707 (0.152) -0.646 (0.131) -0.766 (0.156)
stage IIIb -1.131 (0.140) -1.037 (0.120) -1.205 (0.144)
stage IV -1.584 (0.128) -1.463 (0.109) -1.662 (0.130)
intercept 4.139 (0.278) 4.022 (0.239) 4.030 (0.297)
ν 0.859 (0.019)
κ 0.848 (0.020)
log-likelihood -2766.666 -2791.216 -2786.161

and parameters can be transformed from one model to the other. Denoting
the regression parameters in the AFT model by β ∗ for the parameters of the
proportional hazards model hold β = −νβ ∗ with ν from the Weibull baseline
hazard. Parameter λ is given by λ = exp(−ν × intercept). The interpretation
of the regression parameters in both models is completely different. Note
the opposite sign of the model parameters in the proportional hazards and
AFT models. The hazard ratio of patients with disease stage IV is around
e1.358 =3.89 times higher compared to patients with stage I (reference group)
in the proportional hazards model, interpreted as relative risk. In the AFT
model the expression e−1.584 = 0.21 indicates a reduction of survival time
by this factor, meaning that patients in stage group IV experience the event
around five times faster than patients in risk group I.
Similar to the parametric proportional hazards model, the Weibull model
shows a significant improvement compared to the exponential model. The
exponential AFT model is not nested in the log-logistic AFT model. Hence,
for comparison of the two models the Akaike Information Criterion (AIC)
should be used (Akaike 1974). The AIC statistic is given by the expression
-2(log-likelihood+ #parameters), where #parameters denote the number of
model parameters. Consequently, the AIC value is 8.11 for the comparison of
the exponential and log-logistic model. This favors the log-logistic compared
to the exponential model, which is not a proportional hazards model. Because
of the same number of parameters in both the Weibull as well as log-logistic
models the Weibull model performs better based on the log-likelihood. The
AFT model is more robust with respect to unobserved covariates compared
to the proportional hazards model (Hougaard et al. 1994, Hougaard 1999).
Orbe et al. (2002) compare Cox and AFT models in detail and discuss their
advantages and limitations.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
52 Frailty Models in Survival Analysis

2.6 Identifiability Problems

The assumption of independent censoring is made in numerous applications of
lifetime experiments. Of course, for a number of reasons, one can assume some
of the risks to be acting independently, namely, the end of study censoring,
patients moving to another location for reasons unrelated to the treatment,
accidental deaths, etc. But in some situations this convenient independence
assumption is of doubtful validity. Especially, different causes of death such
as cancer and heart diseases are usually not independent. Hence, a need for
methods in survival analysis in dependent censoring arises. Unfortunately,
the observed data (T, ∆) with T = min{T ∗, C} and ∆ = 1(T ∗ ≤ C) provides
not sufficient information to determine the joint distribution of the vector
(T ∗ , C). Cox (1959) discussed this identifiability problem first. It turns
out that there exist both an independent and one or even more dependent
models for (T ∗ , C) that generate the same joint distribution for (T, ∆) (Tsiatis
1975). It turns out that these ”equivalent” independent and dependent joint
distributions may have quite different marginal distributions as demonstrated
by the following simple example, which is based on a two-dimensional version
of the exponential distribution.

Example 2.9
(Tsiatis 1975) Let T ∗ and C be nonnegative random variables with joint
cumulative distribution function

∗ ∗
−µc−ϑt∗ c
H(t∗ , c) = 1 − e−λt − e−µc + e−λt .

The distribution of the observed data (T, ∆) is uniquely determined by the

two subdistributions H0 (t) = P(T < t, ∆ = 0) and H1 (t) = P(T < t, ∆ = 1).
2
H(t∗ ,c)
Denote by h(t∗ , c) = ∂ ∂t ∗ ∂c the probability density function of (T, ∆) and
∂H(t∗ ,c) ∗
∗
let Ht∗ (t , c) = ∂t∗ and Hc (t∗ , c) = ∂H(t∂c
,c)
. Then it follows for the
subdistribution functions

Z Z
H0 (t) = h(t∗ , c) dt∗ dc
{c<t,c<t∗ }
Z t
= Hc (∞, c) − Hc (c, c) dc
0
Z t
2
= (µ + ϑc)e−λc−µc−ϑc dc
0

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Survival Analysis 53

and in a similar way

Z Z
H1 (t) = h(t∗ , c) dt∗ dc
{t∗ <t,t∗ ≤c}
Z t
= Ht∗ (t∗ , ∞) − Ht∗ (t∗ , t∗ ) dt∗
0
Z t
∗
−µt∗ −ϑ(t∗ )2
= (λ + ϑt∗ )e−λt dt∗ .
0
∗
Let T̄ and C̄ be independent random variables with cumulative distribution
functions
∗
− 12 ϑ(t∗ )2 1 2
F (t∗ ) = 1 − e−λt and G(c) = 1 − e−µc− 2 ϑc .
¯ = 1(T̄ ∗ ≤ C̄).
Using the foregoing notation it holds that T̄ = min{T̄ ∗ , C̄} and ∆
Consequently,

¯ = 0)
H̄0 (t) = P(T̄ < t, ∆
Z Z
= f (t∗ )g(c) dt∗ dc
{c<t,c<t∗ }
Z t
= (1 − F (c))g(c) dc
0
Z t
2
= (µ + ϑc)e−λc−µc−ϑc dc
0

and

¯ = 1)
H̄1 (t) = P(T̄ < t, ∆
Z Z
= f (t∗ )g(c) dt∗ dc
{t∗ <t,t∗ ≤c}
Z t
= (1 − G(t∗ ))f (t∗ ) dt∗
0
Z t
∗
−µt∗ −ϑ(t∗ )2
= (λ + ϑt∗ )e−λt dt∗ .
0

That means H0 (t) = H̄0 (t) and H1 (t) = H̄1 (t). Hence, the distributions of
¯ are equal. It is impossible
the observable random variables (T, ∆) and (T̄ , ∆)
to distinguish between the dependent and the independent model based on
the observed data (T, ∆). One possibility to circumvent this problem is the
inclusion of observed covariates into the model.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com

Chapter 3
Univariate Frailty Models

This chapter focuses on the analysis of univariate data, for example, event
times of unrelated individuals. Basic survival models deal with the simplest
case of independent and identically distributed data. This is based on the
assumption that the study population is homogeneous up to some observed
covariates. Such kind of models were considered in the last chapter. However,
it is a basic observation that individuals differ greatly, for example, with
respect to the effects of a drug, a treatment, or the influence of various
explanatory variables. This heterogeneity is often referred to as variability,
and it is one of the important sources of variability in medical, epidemiological
and biological applications. The issue of this chapter is unobserved hetero-
geneity in survival analysis. This heterogeneity may be difficult to assess, but
it is nevertheless of great importance. In recent decades, a large amount of
papers on frailty models have appeared. The key idea of these models is that
individuals have different frailties, and that the most frail will die earlier than
the less frail. Consequently, systematic selection of robust individuals takes
place, which biases what is observed. When mortality rates are estimated,
one may be interested in how they change over time or age. Quite often, they
rise at the beginning of the observation period, reach a maximum, and then
decline (unimodal hazard) or level off. This, for example, is typical for death
rates of cancer patients, meaning that the longer the patient lives beyond
diagnosis and treatment, the better her or his chances of survival are. But
it is often an open question whether this reflects changes in the individual
hazard. It is likely that unimodal hazards are often the result of selection and
that they do not reflect an underlying development on the individual level.
The population hazard starts to decline simply because high-risk individuals
have already died, but the hazard of a given individual might well continue
to increase.
If covariates are observed, then they can be included in the analysis, for
example, by using the proportional hazards model. However, it is nearly
impossible to include all important risk factors, perhaps because the researcher
has little or no information on the individual level. This applies, for example,
to population studies where often the only known variables are sex and age.
Furthermore, we may not know the relevance of the risk factor or even that
the factor exists. In other cases it may be impossible to measure the risk
factor without great financial cost or time effort. In such cases it is useful to

55
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
56 Frailty Models in Survival Analysis

consider two sources of variability in duration data: variability accounted for

by observable risk factors included in the model (and therefore theoretically
predictable) and heterogeneity caused by unknown covariates and which is
thus theoretically unpredictable. It is the latter that is of specific interest
in the present chapter, and the subject of observable covariates is treated
here only for completeness. Following Hougaard (1991), there are advantages
in considering these two sources of variability separately: heterogeneity may
explain some unexpected results or gives an alternative explanation of some
results, for example, nonproportional or decreasing hazard functions. If some
individuals experience a higher risk of failure, then the remaining individuals
tend to form a more or less selected group with lower risk. An estimate of
the individual hazard rate without taking into account the unobserved frailty
will thus underestimate the hazard function to an increasingly greater extent
as time goes by. Here the reader is referred to Aalen et al. (2008).
To be aware of such selection effects, mixture models could be used. That
means the population is assumed to be a mixture of individuals with at least
partly unknown different risks. The nonobservable risks are described by the
mixture variable, which is called frailty in survival analysis. It is a random
variable that is assumed to follow some distribution. The precise nature of the
relationship between individual and population aging depends on the frailty
distribution among individuals. Different choices of distributions for the
unobserved covariates are possible. The general characteristics of the frailty
distributions are studied in Section 3.1. Of special interest is how the effect
of the frailty on the survival and hazard function can be deduced from the
Laplace transform of the frailty distribution. Section 3.2 considers the discrete
frailty model, including binary frailty, in detail. The gamma and log-normal
distributions are discussed in Sections 3.3 and 3.4. The popularity of the
log-normal frailty model stems mainly from the link with generalized mixed
models, where the standard assumption is that the random effects (which are
the log frailties) follow a normal distribution. Section 3.5 deals with inverse
Gaussian frailty. In Sections 3.6 and 3.7 the positive stable and the power
variance function (PVF) frailty models are considered. The positive stable
distribution has some interesting features such as absence of finite moments.
It transfers the proportional hazards condition from the conditional to the
unconditional model. The PVF family contains the gamma, positive stable,
and inverse Gaussian distributions as special cases. The compound Poisson
frailty model is discussed in Section 3.8. It allows a subpopulation to be
not at risk for the event under study. Section 3.9 considers distributions
based on the concept of quadratic hazard, whereas Sections 3.10 and 3.11
discuss frailty generated by the Lévy processes and log-t frailty models. Cure
models as a specific type of frailty models are discussed in Section 3.12. The
frailty approach to modeling unobserved covariates is the topic of Section 3.13.
Based on the choice of frailty distribution, the variance in frailty determines
the degree of unobserved heterogeneity and deals as an indicator for important
risk factors missing in the proportional hazards model.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 57

3.1 The Concept of Univariate Frailty

Before starting to study specific frailty distributions in detail, we collect a few
results that hold for all these distributions in general. Special focus is on the
Laplace transform of the frailty because unconditional survival and hazard
functions can be easily expressed using this Laplace transform. Hence, the
likelihood function can also be expressed by means of the Laplace transform.
This is the reason why frailty distributions with easy Laplace transforms are so
popular; they allow for traditional maximum likelihood methods in parameter
estimation.
To address the problem of unobserved heterogeneity in event times resulting
from unobserved covariates, first Beard (1959) (in a long-time unrecognized
paper) and later Vaupel et al. (1979) and Lancaster (1979) independently
suggested a random effects model for durations. Beard (1959) used the term
longevity factor rather than frailty. The purpose of introducing the random
effect was to improve the fit of mortality models in populations. Vaupel et al.
(1979) introduced the concept of frailty to the biostatistical community and
applied it to population mortality data. Lancaster (1979) dealt with times of
unemployment and introduced the model to the econometric literature, where
the model is known as the mixed proportional hazards model (MPH). First
we will consider the frailty model without observed covariates to focus on the
main aspects of univariate frailty modeling. The classical and most frequently
applied model assumes a proportional hazards structure that is conditional
on the random effect (frailty). To be more specific, the hazard function of an
individual depends on an unobservable, time-independent random variable Z.
In the multiplicative hazards framework that we will focus on throughout this
book, Z acts multiplicatively on the baseline hazard function µ0

µ(t|Z) = Zµ0 (t). (3.1)

Here, Z is considered a nonnegative random mixture variable, varying across

the population. Note that a scale factor common to all subjects in the study
population maybe absorbed into the baseline hazard function µ0 (t), so that
frailty distributions are standardized to EZ = 1 if the expectation of frailty
distribution exists. Otherwise, another standardization has to be used as will
be discussed in the positive stable model. The variance σ 2 = V(Z) (if it exists)
is interpretable as a measure of heterogeneity across the population in baseline
risk. When σ 2 is small, the values of Z are closely located around one. If σ 2
is large, then values of Z are more dispersed, inducing greater heterogeneity
in the individual hazards Zµ0 (t). Frailty changes the individual hazard and
is sometimes called liability or susceptibility in other settings. All individuals,
apart from the individual frailty variable Z, are assumed to follow the same
mortality pattern.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
58 Frailty Models in Survival Analysis

The problem is that what can be observed in a study population is not the
conditional hazard but the net result for all the individuals with different
values of the random variable Z. The individual frailty variable Z is assumed
to be constant over time throughout the book. Approaches that relaxes this
restriction are discussed in Section 7.7.
It is quite clear that a multiplicative frailty model such as (3.1) represents
a rather simplified view of how heterogeneity might act. Nevertheless, simple
mathematical models represent one way of understanding the consequences
of heterogeneity. The assumptions that the frailty is timeindependent and
that it acts multiplicatively on the underlying baseline hazard function are
arbitrary, but they have been taken as the basis for much subsequent work on
unobserved heterogeneity in survival analysis.
Only for the sake of completeness would we want to mention other cases
of frailty models that are not based on the proportional hazards assumption.
For example, the additive frailty model where frailty acts additively on the
baseline hazard function. For more details of this model, see Rocha (1996),
Silva and Amaral-Turkman (2004), and Tomazella et al. (2006). A more
general model, including the additive as well as the multiplicative frailty model
as special cases, is considered by Gupta and Gupta (2009). Proportional odds
frailty models are covered in detail by Lam et al. (2002) and Lam and Lee
(2004). Murphy et al. (1997) point out the link between proportional odds
models and frailty models. AFT frailty models are dealt with, for example,
by Anderson and Louis (1995), who use the model in bivariate survival with
parametric and nonparametric frailty distributions. Keiding et al. (1997)
focus on the effect of heterogeneity caused by omitted covariates and found the
AFT model more stable than the proportional hazards model in the presence
of heterogeneity. Klein et al. (1999) consider a normal regression model
based on log-normal frailty distribution. Pan (2001) proposes a model for
correlated failure times by modeling the error term of the AFT model with
a frailty approach. Because of instability problems in Pan’s EM algorithm,
Zhang and Peng (2007) suggest a semiparametric estimation procedure based
on M-estimates and the EM algorithm. Xu and Zhang (2010) develop a more
stable estimation procedure in the semiparametric gamma frailty AFT model.
Lambert et al. (2004) study a parametric AFT model with an additive frailty
term. Further interesting papers in this direction are Schnier et al. (2004),
Chang (2004) and Komárek et al. (2007). Sankaran and Gleeja (2008) suggest
the use of proportional reversed hazards frailty models, which are of interest
in the case of left-truncated event time data.
It is natural to introduce observed covariates into model (3.1) similar to the
Cox model by
′
µ(t|X, Z) = Zµ0 (t)eβ X (3.2)

with X = (X1 , . . . , Xk ) and β = (β1 , . . . , βk ) as covariates and regression

parameters, respectively. Consequently, a frailty model is a generalization of

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 59

the well-known proportional hazards model. The proportional hazards model

is obtained if the frailty distribution degenerates to Z=1 for all individuals.
The likelihood of the data (Ti , ∆i , Xi , Zi ) (i = 1, 2, . . . , n) is similar to the
expression (2.12) given by
n
Y ′
δi ′
Zi µ0 (ti )eβ Xi exp − Zi M0 (ti )eβ Xi ,


(3.3)
i=1

but now conditional on the unobserved frailty variables Z1 , Z2 , . . . , Zn . This

likelihood forms the basis for further parameter estimation. The random effect
(frailty) can be integrated out (in closed form or by numerical or stochastical
integration, depending on the frailty distribution) to get a likelihood function
not depending on unobserved quantities. Estimation procedures now depend
on whether a parametric or a semiparametric model is chosen. Inference for
the semiparametric frailty model with results about uniform consistency and
weak convergence are established under mild regularity assumptions for all
parameters, including the baseline hazard function, in Kosorok et al. (2004).
The paper is restricted to the case of one-parameter frailty distributions, which
covers most of the common frailty distributions used. If the model is correctly
specified, its method is efficient, achieving the semiparametric information
bound for all parameter components. Detailed results are also given for the
case of misspecified frailty distribution.
For the sake of simplicity, we will restrict our treatment to the model (3.1)
sometimes in order to focus on the main ideas of frailty models. Let S(t|Z)
denote the survival function of an individual conditional on the frailty Z:

Rt Rt
S(t|Z) = e− 0
µ(s|Z) ds
= e−Z 0
µ0 (s) ds
= e−ZM0 (t) . (3.4)
Rt
Here M0 (t) = 0 µ0 (s) ds denotes the cumulative baseline hazard function,
and equation (3.4) is a generalization of relation (2.2). Up to now, the model
has been described at the individual (conditional) level. However, data for this
individual model are not observable. Consequently, it is necessary to consider
the population level where the frailty term is integrated out. The population
survival function is the weighted mean of the conditional survival functions
with weights given by the density function of the frailty distribution. The
population survival function is obtained from the conditional survival function
S(t|Z) by integrating out the frailty. It can be viewed as the (unconditional)
survival function of an individual randomly drawn from the study population,
and corresponds to what can actually be observed:

S(t) = ES(t|Z) = Ee−ZM0 (t) = L(M0 (t)). (3.5)

This is based on relation (3.4) and underlines the important role of the Laplace
transform in frailty models. The derivatives of the Laplace transform can be
used to obtain general results about unconditional survival distribution. For

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
60 Frailty Models in Survival Analysis

example, density and hazard function of the event times and expectation and
variance of the frailty can be characterized by the Laplace transform of the
frailty distribution and their derivatives

f (t) = −µ0 (t)L′ (M0 (t))

L′ (M0 (t))
µ(t) = −µ0 (t)
L(M0 (t))
′
EZ = −L (0)
V(Z) = L′′ (0) − (L′ (0))2 , (3.6)

assuming the existence of the foregoing expressions. For example, in the case
of a positive stable distribution the moments do not exist. Here, L′ and L′′
denote the first and second derivative of the Laplace transform. Thus, if the
Laplace transform has a simple form, performing this calculation is easy. The
connection with the Laplace transform was first pointed out and exploited
by Hougaard (1984, 1986a,b). It follows that, when seeking distributions for
the frailty variable Z, it is natural to use those which have explicit Laplace
transforms. This simplifies parameter estimation.
Frailty distribution describes the frailty in the population at the start of the
follow-up. Frailty is assumed to be fixed for each individual over time, but the
composition of the population changes as time goes by. On average, the more
frail individuals die earlier. Due to this fact the frailty distribution in the
population at risk changes over time. The following theorem points on this
fact and establishes the link between the conditional and the unconditional
model.

THEOREM 3.1
(Vaupel et al. 1979) Assume a frailty model given by formula (3.1). The
f (t)
population hazard µ(t) = S(t) is generally µ(t) = E(µ(t|Z)|T > t), or more
specifically,

Z ∞
µ(t) = µ(t|z)f (z|T > t) dz
0
Z ∞
= µ0 (t) zf (z|T > t) dz, (3.7)
0

where f (z|T > t) represents the frailty density among the survivors of time
point t.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 61

Proof: Starting with relation (3.1) we get

f (t|z)
µ(t|z) = = zµ0 (t)
S(t|z)
f (t|z) = zµ0 (t)S(t|z)
f (t, z) = zµ0 (t)S(t|z)fZ (z)
Z ∞
f (t) = µ0 (t) zS(t|z)fZ (z) dz
0

with fZ as p.d.f. of the frailty distribution. Hence,

R∞
µ0 (t) 0 zS(t|z)fZ (z) dz
µ(t) = .
S(t)
Because survival of time point t implies a death time that is greater than t,
it holds that
Z ∞
f (z, T > t) = f (z, s) ds
t
Z ∞
= fZ (z) zµ0 (s)S(s|z) ds
t
= fZ (z)S(t|z)
fZ (z)S(t|z)
f (z|T > t) = . (3.8)
S(t)
This completes the proof. 
The hazard of the population is thus interpreted as the weighted mean of
individual hazards among the survivors (see formula (3.7)). The weights are
determined by the frailty distribution. Frail individuals with high values of
Z will tend to die first. For example, an individual with frailty equal to
two is twice as likely to die as a standard individual with frailty equal to
one. An individual with frailty equal to 0.5 experiences only half of the risk
to die compared Rto a standard individual. Thus, the average frailty of the
∞
surviving cohort 0 zf (z|t) dz will decline with time, or more generally, the
frailty distribution of the survivors is changing. Consequently, equation (3.7)
implies that the hazard for individuals increases more rapidly than for the
cohort to which the individuals belong. In this sense, individuals age faster
than their cohort (see Figure 3.1). An illustrative example is given in the paper
by Manton and Stallard (1981), describing the different mortality patterns of
heterogeneous populations. The frailty was introduced in order to interpret
mortality data more appropriately. The proposed model is a univariate model
where each subject owns it individual frailty value, thus allowing for individual
differences in hazard functions. The main objective of Vaupel et al. (1979) and
Manton and Stallard (1981) was to point out that the population hazard does
not reflect the hazard of individuals from that population as shown earlier.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
62 Frailty Models in Survival Analysis

This kind of leveling-off effects in population hazard functions are reported by

Vaupel et al. (1998) for different populations such as humans, med flies, fruit
flies, nematodes, yeast, and automobiles. Vaupel et al. (1979) used the frailty
approach to derive the individual hazard function based on the population
hazard function obtained from life tables. Finkelstein (2008) analyzes the
asymptotic behavior of unconditional (mixed) hazard functions. It can be
shown that, in some special cases, even for increasing conditional hazards the
unconditional hazard function asymptotically decreases to zero. Furthermore,
the asymptotic behavior of the unconditional hazard depends only on the
behavior of the frailty distribution in the neighborhood of the left end point
of its support and not on the whole frailty distribution.
Equation (3.8) gives the distribution of the frailty among the survivors at
time point t, which will be used in the EM algorithm for parameter estimation
later on. The frailty distribution of those individuals who die at time point t
is a result of Bayes’ formula:

f (z, t) f (t|z)f (z)

f (z|t) = = . (3.9)
f (t) f (t)

2
Z=2
Z=1
Z=0.5
unconditional
hazard

0
0 10 20 30 40 50
time

Figure 3.1: Conditional and unconditional hazard functions of a fictive het-

erogeneous population.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 63

To prove the existence of unobserved heterogeneity in a population, stress

experiments with laboratory animals (e.g., insects, worms) were suggested.
The idea of using data from stress experiments for testing the heterogeneity
assumption is based on the belief that the same exposure to stress produces
different survival outcomes in heterogeneous and homogeneous populations
after the exposure to stress. During the exposure, the heterogeneous study
population will experience higher selection (since the frailer individuals die
first). In the homogenous group, mortality is driven by randomness and not
by selecting the frailer individuals. Hence, after exposure to stress, mortality
of the stress group should be lower than that of the control group. Frailty
models focus only on aspects of unobserved heterogeneity, which is of course
a simplification. For example, the concept of adaptation is not included in the
frailty models considered here. Adaptation means that individuals can adapt
to stress after exposition to moderate stress. Debilitation is the opposite effect
after stress exposure. Both effects are not included in the frailty concept
because frailty in statistical modeling is usually assumed to be time constant,
excluding the time-dependent effects of adaptation or debilitation. For more
details see a series of discussion papers, for example, Khazaeli et al. (1995),
Drapeau et al. (2000), Mueller et al. (2003), Wu et al. (2006), Rose et al.
(2006), and Rockwood and Mitnitski (2007).
Under model (3.2), the unconditional hazard function can be expressed as
the expectation of the hazard function (Nielsen et al. 1992) conditional on
being at risk at time t and covariate vector X; that is,
′
µ(t|X) = µ0 (t)eβ X E(Z|T > t, X).

Taking the expectation above represents an averaging over all individuals

in the population. Since the expectation is conditional on being at risk at
time point t, it implies averaging over a subset of the original population.
Hence, relative weights for hazards with high frailty become smaller as time
goes by, corresponding to high mortality. An important implication is that
studies of human aging based on cohort mortality data may be systematically
biased or based on erroneous functional forms. The precise nature of the
relationship between individual and cohort aging depends on the distribution
of the frailty among individuals. Different choices for the frailty distribution
are possible. Some main results can be found in Vaupel and Yashin (1985).
Note that a parametric specification of the frailty distribution is a matter
of mathematical convenience, and we will follow this line. However, this
parametric specification is not necessary. Especially in the econometric field
this is a topic of research. For example, Heckman and Singer (1984a) estimate
the frailty distribution nonparametrically but under parametric assumptions
on the baseline hazard. Horowitz (1999) proposes strategies to estimate both
the frailty distribution and the baseline hazard nonparametrically. It turns
out, that it is very difficult to estimate the frailty distribution in real data
applications, see Heckman and Taber (1994), Kortram et al. (1995), and

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
64 Frailty Models in Survival Analysis

Horowitz (1999). For an extensive overview of identifiability results regarding

frailty models with unspecified frailty distribution, see van den Berg (2001).
A short summary of the most important identifiability results related to the
different univariate and multivariate frailty models considered in this book
is given in Section 7.8. However, analytical results are sparse and in many
situations simulations are the only way to deal with this often overlooked
problem.
Assigning a probability distribution to frailty implies that frailty can be
integrated out in the likelihood function. After this integration the likelihood
can be maximized in the usual way if an explicit form of it exists. Otherwise,
more sophisticated approaches like numerical integration or Markov Chain
Monte Carlo methods need to be applied. In the following sections different
frailty distributions are discussed. The most often used frailty distributions
are the gamma and the log-normal distribution. The positive stable and the
inverse Gaussian distribution are also common. Using a larger distribution
family provides for a better fit but at the cost of additional parameters and
more complex models. We will consider the power variance function (PVF)
distribution as an example of such a large mixing distribution family, which
includes the gamma and inverse Gaussian distributions as special cases. Only
very few results are published in the literature on comparing models with
different frailty distributions; much more research is needed in this important
research area. In the following we will compare different frailty distributions
and discuss their specific advantages and limitations in applications to real
data examples.
The term frailty originates from gerontology where it is used to indicate
the susceptibility to mortality and morbidity. However, there is a lack of
agreement as to what determines the medical and gerontological concept of
frailty (Rockwood 2005, Rockwood and Mitnitski 2007 and references therein).
In the fields of biostatistics and demography, frailty is interpreted as a random
effect. This statistical definition of frailty is different from the frailty used
in the fields of gerontology and medicine, and the two concepts should not
be mixed. Frailty in biostatistics and demography is usually assumed to be
constant over time for an individual (for a few exceptions, see Section 7.7) with
an interpretation as a random effect. In the medical and gerontological context
frailty is assumed to be changing, usually increasing over time. Univariate
frailty models in biostatistics are often used to model the effect of unobserved
heterogeneity (unobserved covariates), whereas in the medical context the
main objective is to find surrogate measures (specific scores, for example, with
respect to the activity of daily living) for frailty to identify frail individuals. In
contrast, determining which individuals are frail seems to be of less importance
for frailty models in biostatistics and demography.
A deeper discussion of the variety of frailty concepts from a medical and
gerontological point of view is beyond the scope of this book. For more details
see Morley et al. (2002) and references therein.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 65

3.2 Discrete Frailty Model

As a simple introduction and explanation of univariate frailty models, the
study population is considered as consisting of two subpopulations. Both
subpopulations are assumed to be homogeneous with different risks. This
model is called the binary frailty model. This could, for example, correspond
to the presence/nonpresence of a specific disease genotype. If no detailed
information on the disease genotype of each individual exists, it is impossible
to include the genotype in the analysis such as an observed covariate. Instead
we have to consider the genotype as unobserved covariate. Let us assume that
the proportion of individuals carrying the disease genotype η1 is π. Hence,
a randomly selected person from the population carries the genotype with
probability π. Naturally, this leads to an increased variation in the response
time when compared to the case where the genotype is known. Furthermore,
assume that P(Z = η1 ) = π and P(Z = η2 ) = 1 − π. In this case, it holds for
the Laplace transform L and their derivatives

L(u) = πe−uη1 + (1 − π)e−uη2

L′ (u) = −πη1 e−uη1 − (1 − π)η2 e−uη2
L′′ (u) = πη12 e−uη1 + (1 − π)η22 e−uη2 .

The expectation and variance of a binary-distributed frailty variable can now

be obtained from the derivatives of the Laplace transform (3.6)

EZ = πη1 + (1 − π)η2
V(Z) = πη12 + (1 − π)η22 − (πη1 + (1 − π)η2 )2
= π(1 − π)(η1 − η2 )2 .

The survival and density function can now be easily derived from the Laplace
transform by using (3.5)

S(t) = L(M0 (t)) = πe−η1 M0 (t) + (1 − π)e−η2 M0 (t)

and

f (t) = πη1 µ0 (t)e−η1 M0 (t) + (1 − π)η2 µ0 (t)e−η2 M0 (t) .

This implies the hazard

πη1 e−η1 M0 (t) + (1 − π)η2 e−η2 M0 (t)

µ(t) = µ0 (t). (3.10)
πe−η1 M0 (t) + (1 − π)e−η2 M0 (t)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
66 Frailty Models in Survival Analysis

0.2 subpopulation 1
mixed population
subpopulation 2
hazard

0.1

0.0
0 20 40 60 80 100
age

Figure 3.2: Mortality hazards of two populations with Gompertz hazard

µ0 (t) = 0.001e0.05t and their mixed hazard. The frailty values are η1 = 1.2
and η2 = 0.4 for the two subpopulations. At birth, the mixed population
contains 75% of subpopulation 1 and 25% of subpopulation 2. The mortality
of the mixed population is given by expression (3.10). The change in the
proportion of the two subpopulations is shown in Figure 3.3.

To fulfill the standardization condition EZ = 1 in the binary frailty model, it

holds πη1 + (1 − π)η2 = 1, yielding the relation

1 − πη1
η2 = . (3.11)
1−π
Consequently, the model contains, with π and η1 , two additional parameters
compared to the model without binary frailty. Note that the condition EZ = 1
is not necessary in the binary frailty model, but is used here to make models
comparable. Using (2.7) and highlighting the dependence of the baseline
hazard distribution on parameter vector θ, the likelihood function in the
binary frailty model with η = η1 is of the form

n  1−πη
Y πηe−ηM0 (ti ;θ) + (1 − πη)e− 1−π M0 (ti ;θ) δi
L(θ, η, π) = 1−πη µ0 (ti ; θ) (3.12)
i=1 πe−ηM0 (ti ;θ) + (1 − π)e− 1−π M0 (ti ;θ)
 1−πη

× πe−ηM0 (ti ;θ) + (1 − π)e− 1−π M0 (ti ;θ) .

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 67

Denote by π(t) the proportion of individuals with the disease genotype (sub-
population 1) at time point t (e.g., π = π(0)). Consequently, 1 − π(t) denotes
the proportion of the second subpopulation without the disease genotype.
Here, special focus is on the proportion π(t). This quantity changes over time
as a result of the selection process. Selection takes place if the lifetimes of
different subpopulations follow different mortality patterns. To calculate π(t),
we introduce the conditional survival function for individuals with the disease
genotype and without the disease genotype. Conditioning is with respect to
the event {J = j}, meaning that the individual belongs to subpopulation j
(j = 1, 2): P(T > t|J = j) = e−ηj M0 (t) . By using the Bayesian formula we
can now write the proportion of the fraction in the form
P(T > t, J = 1)
π(t) = P(J = 1|T > t) =
P(T > t)
P(T > t|J = 1)P(J = 1)
=
P(T > t|J = 1)P(J = 1) + P(T > t|J = 2)P(J = 2)
πe−η1 M0 (t)
= . (3.13)
πe−η1 M0 (t)+ (1 − π)e−η2 M0 (t)

The binary frailty is sometimes also called two-point frailty. Nickell (1979)
used this model to account for heterogeneity in unemployment spell data. The
population was divided into groups of motivated and nonmotivated searchers
of a job. Furthermore, this frailty distribution was used by Vaupel and Yashin
(1985) to discuss ideas of the heterogeneity and selection concept in detail,
and by Schumacher et al. (1987) to model heterogeneity in clinical trials. A
special case of binary frailty is dividing the population into a proportion that is
at risk and a proportion that is never at risk. The terminology to describe the
never-at-risk group varies from field to field. It includes ”long-term survivors”
(Farewell 1982) or ”cured” in epidemiology (Price and Manatunga 2001),
”nonsusceptibles” in toxicology (Pack and Morgan 1990), ”stayers” in finite
Markov transition models of occupational mobility (Blumen et al. 1955), the
”nonfecundable” in fertility models (Heckman and Walker 1990), and ”non-
recidivists” among convicted criminals (Schmidt and Witte 1989, Maller and
Zhou 2002). We will use the term cure model to describe such models and
come back to this problem later on in Section 3.12 and Section 5.8.

Example 3.1
In Table 3.1 the results are given for the parametric binary frailty model
with Weibull baseline hazard analyzing the Halluca lung cancer data. For
comparison, the results of the parametric proportional hazards model with
Weibull baseline hazard are given as well. The analysis was performed using
PROC NLMIXED in SAS by specifying the likelihood function given in (3.12),
substituting the expressions µ0 (ti ) and M0 (ti ) with their regression counter-
′ ′
parts µ0 (ti )eβ Xi and M0 (ti )eβ Xi , respectively. The two-point frailty model

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
68 Frailty Models in Survival Analysis

1.0 subpopulation 1
subpopulation 2

0.8
proportion

0.6

0.4

0.2

0.0
0 20 40 60 80 100
age

Figure 3.3: Age trajectories of the proportions of the two subpopulations

in the mixing of the two Gompertz populations described in Figure 3.2. The
proportions are given by (3.13). The first population (with higher mortality)
dies faster over time. For this reason the hazards of the mixed population and
the second subpopulation converge at older ages in Figure 3.2.

with Weibull baseline hazard provides a better fit compared to the Weibull
proportional hazards model based on the comparison of the log-likelihood (see
Table 3.1). Both mass points η1 , η2 of the frailty are very different from each
other. The model detects a very small subgroup of individuals with extremely
high risk of death. Further detailed analysis shows that this group consists of
patients in whom lung carcinoma was detected after death in the pathology
(cluster 4 in Figure 1.1). Hence, for these patients, the event time is zero.
The standard error for η2 is not given because it is not a free parameter in
the model (see (3.11)).

Obviously, the binary frailty model can be extended with no additional

efforts to the finite discrete distribution of frailty Z (k-point distribution). In
this case the hazard function is
′
µ(t|X, Z) = Zµ0 (t)eβ X
with
k
X
EZ = P(Z = ηj )ηj = 1.
j=1

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 69
Table 3.1: Parametric binary frailty model for
Halluca data
parameter Weibull PH two-point frailty

age (in years) 0.010 (0.003) 0.010 (0.003)

sex (females) -0.166 (0.071) -0.159 (0.072)
type (NSCLC) -0.136 (0.067) -0.159 (0.067)
ECOG (3 & 4) 0.612 (0.103) 0.568 (0.107)
stage II 0.455 (0.168) 0.461 (0.170)
stage IIIa 0.606 (0.131) 0.620 (0.131)
stage IIIb 0.970 (0.120) 0.990 (0.120)
stage IV 1.358 (0.110) 1.388 (0.110)
λ 0.029 (0.007) 0.200 (0.080)
ν 0.859 (0.019) 0.920 (0.022)
π 0.987 (0.034)
1−π 0.013 ( – )
η1 0.118 (0.168)
η2 66.22 ( – )
log-likelihood -4858.672 -4844.373

Similar to the binary frailty model denote by πj (t) the size of the fraction of
k
P
subpopulation j at time point t (e.g., πj (t) = 1) and πj = πj (0). We are
j=1
now interested in the behavior of πj (t). To calculate this function we introduce
the conditional survival function. Again, conditioning is with respect to the
event {J = j}, meaning that the individual belongs to the subpopulation j
(j = 1, 2, . . . , k).
Rt
Sj (t) = P(T > t|J = j) = e−ηj M0 (t) = e−ηj 0
µ0 (s)ds

Now we can write the proportion of subpopulation j on the whole population

using the Bayes’ formula in the form

πj (t) = P(J = j|T > t)

P(T > t, J = j)
=
P(T > t)
P(T > t|J = j)P(J = j)
= k
P
P(T > t|J = i)P(J = i)
i=1
πj e−ηj M0 (t)
= k
.
πi e−ηi M0 (t)
P
i=1

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
70 Frailty Models in Survival Analysis

Obviously, this is a simple generalization of relation (3.13). The proportion

of subpopulation j changes over time, calculated by
Pk
πj (t) πj ηj µ0 (t)e−ηj M0 (t) πj e−ηj M0 (t) i=1 πi ηi µ0 (t)e−ηi M0 (t)
= − Pk +
−ηi M0 (t) 2
dt −ηi M0 (t)
Pk

i=1 πi e i=1 πi e
Pk
πj µ0 (t)e−ηj M0 (t)  πi ηi e−ηi M0 (t) 
= Pk − ηj + Pi=1k
−ηi M0 (t) −ηi M0 (t)
i=1 πi e i=1 πi e
k
X

= πj (t)µ0 (t) πi ηi − ηj
i=1
= πj (t)µ0 (t)(1 − ηj ),

which holds because of EZ = ki=1 P(Z = ηi )ηi = ki=1 πi ηi = 1. Hence,

P P
the proportion of subpopulations with frailty larger than one is decreasing
because mortality is higher compared to subpopulations with smaller frailty.
In genetic studies, for example, the grouping factor is the genotype, and in
most medical studies this is not directly observable. However, under a single-
locus model with two alleles, A and a, three genotypes will occur; AA, Aa,
and aa, with allelic frequencies of π (for A) and 1 − π (for a), respectively.
Under the Hardy–Weinberg equilibrium, the expected proportions of the three
unobservable genotypes are π 2 , 2π(1 − π), and (1 − π)2 . Thus, three latent
classes with known expected proportions with parameter π to be estimated
are assumed. If alleles are codominant, then the location of the heterozygote
Aa will be midway between AA and aa. If A is dominant, then genotypes AA
and Aa will coincide, and the model collapses to a two class problem. Where
A is recessive, Aa and aa will coincide.

Example 3.2
In addition to binary frailty, a three-point as well as a four-point frailty
distribution with Weibull baseline hazard function was applied to the Halluca
data. The results are given in the third and fourth column of Table 3.2. Again,
the small subgroup with extreme high risk as in the two-point frailty model
can be found. Furthermore, the log-likelihood indicates a further improvement
in the model by additional mass points in the discrete frailty distribution. It
turns out that a five-point frailty model does not provide further significant
improvement in the fit to the Halluca data (results are not shown).

We would like to point out that, in the literature, this model with a fixed
number of discrete mass points is sometimes called a nonparametric frailty
model, which is of course a misleading name. The notion of nonparametric
(discrete) frailty should be reserved for models where the number of mass
points is a random variable, which means the number of mass points is an
additional parameter to be estimated (Heckman and Singer 1982b, dos Santos

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 71
Table 3.2: Discrete frailty models for Halluca data
parameter 2-point frailty 3-point frailty 4-point frailty

age (in years) 0.010 (0.003) 0.013 (0.004) 0.014 (0.004)

sex (females) -0.159 (0.072) -0.151 (0.086) -0.204 (0.103)
type (NSCLC) -0.159 (0.067) -0.083 (0.082) -0.160 (0.104)
ECOG (3 & 4) 0.568 (0.107) 1.034 (0.168) 1.085 (0.180)
stage II 0.461 (0.170) 0.498 (0.198) 0.592 (0.198)
stage IIIa 0.620 (0.131) 0.703 (0.151) 0.904 (0.190)
stage IIIb 0.990 (0.120) 1.096 (0.144) 1.360 (0.184)
stage IV 1.388 (0.110) 1.591 (0.131) 1.972 (0.194)
λ 0.200 (0.080) 0.201 (0.082) 0.322 (0.153)
ν 0.920 (0.022) 1.055 (0.040) 1.251 (0.080)
π1 0.987 (0.034) 0.833 (0.065) 0.649 (0.052)
π2 0.013 ( – ) 0.016 (0.004) 0.018 (0.004)
π3 0.151 ( – ) 0.239 (0.065)
π4 0.094 ( – )
η1 0.118 (0.168) 0.071 (0.024) 0.075 (0.032)
η2 66.22 ( – ) 60.07 (12.95) 55.14 (10.65)
η3 0.011 ( – ) 0.014 (0.008)
η4 0.002 ( – )
log-likelihood -4844.373 -4824.343 -4820.123

et al. 1995, Lindstrom 1996). Furthermore, the standard errors in these

models do not reflect uncertainty with respect to the true number of mass
points. It is common to use an iterative procedure, starting with one mass
point and increasing the number until the likelihood fails to show a significant
improvement. Often, two or three points of support provide a sufficient fit
(Guo and Rodriguez 1992). Macdonald (1999) applied two- and four-point
frailty distributions to model the impact of genetics on insurance problems.
Parametric proportional hazards frailty models to allow a negative binomial,
Poisson, geometric or other discrete distribution of the frailty variable are
considered recently by Caroni et al. (2010).
If it is known to which of the k subpopulations the individuals belong, then
this can be included in the model as an observed covariate (fixed effects). This
becomes problematic if the number of subpopulations and consequently the
number of model parameters is large. Another possible approach is the use of
different baseline hazards (stratified analysis). Here it is assumed that each
subpopulation j (j = 1, 2 . . . , k) has its own hazard function instead of using
′
a common baseline in the population: µ(t|X) = µ0j (t)eβ X . This approach
also is only applicable if the number of subpopulations is small compared to
the sample size in each subpopulation. More complex designs, for example,
in the context of multicenter clinical trials, are considered by Legrand et al.
(2005, 2006). Their method is based on multivariate frailty models with a
two-dimensional frailty distribution.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
72 Frailty Models in Survival Analysis

3.3 Gamma Frailty Model

The gamma distribution has been widely applied as a mixture distribution
(Greenwood and Yule 1920, Beard 1959, Vaupel et al. 1979, Congdon 1995,
dos Santos et al. 1995, Hougaard 2000, Duchateau and Janssen 2008). From
a computational and analytical point of view, it fits very well as a mixture
distribution to failure data. It is easy to derive the closed-form expressions
of unconditional survival, cumulative density, and hazard function, which is
due to the simplicity of the Laplace transform. This is also the reason why
this distribution has been used in most applications published to date. The
gamma distribution Γ(k, λ) is a flexible distribution that takes a variety of
shapes as k varies: when k = 1, it is identical to the well-known exponential
distribution; when k is large, it takes a bell-shaped form reminiscent of a
normal distribution (see Figure 3.4).
Despite these advantages, it is necessary to mention that no biological
reason exists that makes the gamma distribution more preferable than other
frailty distributions. Nearly all arguments in favor of the gamma distribution
are based on mathematical and computational aspects. The paper by Abbring
and van den Berg (2007) rationalizes the use of the gamma distribution for
1.2

σ2 = 1
σ2 = 0.5
σ2 = 0.25
1.0

σ2 = 0.125
0.8
density

0.6
0.4
0.2
0.0

0 1 2 3 4

Figure 3.4: Probability density functions of gamma distributions with ex-

pectation 1 and variances 1, 0.5, 0.25, and 0.125

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 73

frailties in time-to-event data analysis. The authors show that, in a large

class of univariate frailty models, the distribution of the frailty among the
survivors converges to a gamma distribution as time goes to infinity under
mild regularity assumptions.
Frailty cannot be negative, and the gamma distribution is, along with the
log-normal distribution, one of the most commonly used distributions to model
variables that are nonnegative. The density of a gamma-distributed random
variable is given by (see Section 2.3.6)
1 k k−1 −λz
f (z) = λ z e .
Γ(k)
Consequently, for the Laplace transform it holds that

1 k
Z
L(u) = λ e−uz z k−1 e−λz dz
Γ(k)
λk 1
Z
= k
(λ + u)k z k−1 e−(λ+u)z dz
(λ + u) Γ(k)
u
= (1 + )−k .
λ
Here, the equivalence of the second and third line is a consequence of the
integration over the density of a gamma distribution with parameters k and
λ + u. The first and second derivatives of the Laplace transform are

k u
L′ (u) = − (1 + )−k−1 (3.14)
λ λ
′′ k(k + 1) u
L (u) = (1 + )−k−2 . (3.15)
λ2 λ
Evaluating these derivatives at u = 0 implies

k
EZ =
λ
k(k + 1) k 2 k
V(Z) = − 2 = 2
λ2 λ λ
To make sure that the model is identifiable, the restriction k = λ is used
for the gamma distribution, which results in EZ = 1. Denote by σ 2 := λ1 the
variance of the frailty variable. The density of a gamma-distributed random
variable Z ∼ Γ( σ12 , σ12 ) is given by

1  1  σ12 12 −1 z

f (z) = 1 2
zσ exp − 2 (3.16)
Γ( σ2 ) σ σ
and is depicted in Figure 3.4.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
74 Frailty Models in Survival Analysis

The unconditional survival function can be derived by the Laplace transform

given in (3.5):

1
S(t) = L(M0 (t)) = 1 . (3.17)
(1 + σ2 M 0 (t)) σ
2

This implies the unconditional probability density function

µ0 (t)
f (t) = 1
+1
(1 + σ2 M 0 (t)) σ
2

and the unconditional hazard function

µ0 (t)
µ(t) = . (3.18)
1 + σ 2 M0 (t)
Furthermore, it turns out that the assumption about a gamma distribution
of the frailty at the start of the follow-up yields some useful mathematical
results, which will be discussed in detail in the following. Mainly for reasons
of convenience, frailty is assumed to be constant over time for each individual.
But because of the selection process that takes place as time goes by, the
distribution of the frailty in the population still at risk changes over time.
In the following presentation we extend the model by including Cox type
regression terms into the formulas meaning substitution of the cumulative
′
baseline hazard M0 (t) with M0 (t)eβ X . We will show that the frailty among
survivors of a specific time point is still gamma distributed but now with
new parameters depending on the original parameter σ 2 and the cumulative
′
hazard function M0 (t)eβ X . Furthermore, a similar result can be obtained for
individuals dying at a specific time point. These results are helpful for the
development of estimation procedures especially in semiparametric gamma
frailty models. Using relations (3.4), (3.8), (3.16), and (3.17), the density of
the frailty distribution among the survivors (indicated by the condition T > t)
can be written in the form

S(t|X, z)f (z)

f (z|X, T > t) =
S(t|X)
′
1 −1
z σ2 exp − σz2


exp − zM0 (t)eβ X
= 2
− 1
Γ( σ12 )σ σ2 1 + σ 2 M0 (t)eβ X σ2
′

1 β′ X

12
+ M0 (t)e σ
1
 1 β′ X


= σ2
z σ2 −1 exp − z + M 0 (t)e ,
Γ( σ12 ) σ2

which is the density of a gamma distribution with the same value of the
shape parameter σ12 as at the begin of the follow-up. The value of the second
′
parameter, however, is now given by σ12 + M0 (t)eβ X . In a similar way (3.9)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 75

can be used to calculate the frailty distribution of individuals who die at time
point t:

f (t|X, z)f (z)

f (z|X, T = t) =
f (t|X)
′
1 −1
z σ2 exp − σz2


zµ0 (t) exp − zM0 (t)eβ X
= 2
− 1 −1
Γ( σ12 )σ σ2 µ0 (t) 1 + σ 2 M0 (t)eβ X σ2
′

1

1
β ′ X σ2 +1
σ2 + M0 (t)e 1 1 ′
z σ2 exp − z( 2 + M0 (t)eβ X )


= 1 1
Γ( σ2 ) σ2 σ
1

1
β ′ X σ2 +1
σ2 + M0 (t)e 1 1 ′
z σ2 +1−1 exp − z( 2 + M0 (t)eβ X ) .


= 1
Γ( σ2 + 1) σ
′
This is a gamma density with the same scale parameter σ12 + M0 (t)eβ X as
among those surviving to t but with shape parameter σ12 + 1. In particular,
it follows that the mean frailty among the deaths at age t is

1 + σ2
E(Z|X, T = t) = (3.19)
1 + σ 2 M0 (t)eβ X
′

compared to
1
E(Z|X, T > t) = (3.20)
1 + σ 2 M0 (t)eβ X
′

among the survivors at the same age. This demonstrates the selection by
early death of the high-risk individuals. The individuals dying at time t have a
higher mean frailty compared to the survivors of this time point. Furthermore,
it holds that the frailty variance among the individuals dying at time t is

σ 2 (1 + σ 2 )
V(Z|X, T = t) = (3.21)
(1 + σ 2 M0 (t)eβ X )2
′

and

σ2
V(Z|X, T > t) = (3.22)
(1 + σ 2 M0 (t)eβ X )2
′

among the survivors. Consequently, the variance of frailty declines also over
time, so the study population becomes more homogeneous in absolute terms.
However, the coefficient of variation stays constant over time, so the population
does not become more homogeneous relative to the mean.
The univariate gamma frailty model (without covariates) was introduced by
Vaupel et al. (1979). In an earlier paper, Beard (1959) used a two-parameter
gamma frailty distribution and a Gompertz–Makeham baseline hazard.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
76 Frailty Models in Survival Analysis

Example 3.3
Let µ0 (t) = λeϕt be a Gompertz baseline hazard and frailty follows a gamma
distribution Z ∼ Γ( σ12 , σ12 ). Hence, the unconditional hazard and survival
function are given by the expressions

λeϕt
µ(t) = λ ϕt
1 + σ2 ϕ (e − 1)
λ 1
S(t) = (1 + σ 2 (eϕt − 1))− σ2
ϕ

as a consequence of (3.17) and (3.18).

Figures 3.5 and 3.6 deal with an important topic related to univariate frailty
models, the so-called crossing-over effects in the mortality hazards of different
populations. Figure 3.5 shows the logarithms of two proportional baseline
hazards, where the second population has a higher mortality. Assuming a
higher degree of heterogeneity in this population (the variance of the gamma-
distributed frailty is larger in the second population) results in the crossing-
over effect because of selection shown in Figure 3.6. Higher heterogeneity
implies a higher selection pressure in the second population.

2 population 1
population 2

-2
log(hazard)

-4

-6

-8

-10
0 20 40 60 80
time

Figure 3.5: Baseline hazard functions of two populations with Gompertz

hazard. The parameters are λ1 = 10−4 , λ2 = 4 × 10−4 , and ϕ1 = ϕ2 = 0.12.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 77

0 population 1
population 2
-1
-2
-3
log(hazard)

-4
-5
-6
-7
-8
-9
-10
0 20 40 60 80
time

Figure 3.6: Unconditional hazards of two populations with Gompertz base-

line hazard and gamma distributed frailty. The parameters of the baseline
hazards are the same as in Figure 3.5. The variance of the frailty σ 2 is as-
sumed to be one in population 1 and four in population 2.

Consequently, frailer individuals die faster in the second population than

in the first. If heterogeneity is on the same level in both populations, a
convergence of the hazards will be observed instead of crossing over. Such
effects are described in detail by Clayton and Kaldor (1985).
These phenomena were also observed, for example, by Manton and Vaupel
(1995) comparing mortality in Sweden, France, England, and Japan with that
in the United States. The authors noticed that mortality was much lower in
all countries except the United States up to the ages of around 65. After the
age of 80, the mortality was much lower in the United States compared to the
three European countries. The authors pointed out that greater heterogeneity
in social and economic status and health insurance coverage in the United
States may account for the higher mortality at younger ages in this country.
To make this more clear, consider the form of the conditional expectation for
the gamma-distributed frailty of the survivors of time t (3.20)
1
E(Z|X, T > t) =
1 + σ 2 M0 (t)eβ X
′

Rt
with M0 (t) = 0 µ0 (s) ds. Clearly, the average frailty value is a decreasing
function of time. The decrease is faster in situations with larger heterogeneity

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
78 Frailty Models in Survival Analysis

in the population, as measured by σ 2 , a higher cumulative baseline hazard

function M0 (t), and larger increases due to covariates, which are modeled by
′
the term eβ X . All these factors accelerate the ”frailty selection” that results
from unobserved heterogeneity in individual risks.
Assume a single binary covariate X with values 0 (control group) and 1
(treatment group). Frailty selection also affects the marginal hazard ratio
for treatment. Under the gamma frailty model, the marginal hazard ratio is
(compare relation (3.18))

µ(t|X = 1) 1 + σ 2 M0 (t) β
= e . (3.23)
µ(t|X = 0) 1 + σ 2 M0 (t)eβ
A critical point here is that, under the proportional hazards assumption for
the conditional model, the ratio of the treatment-specific population hazards
is not generally timeinvariant; unless σ 2 or β is zero, the population hazard
ratio (3.23) is a decreasing function. Only at time zero the population hazard
ratio is equal to the conditional hazard ratio eβ . As t increases, the ratio tends
to one. Under these conditions, the time-invariant hazard ratio obtained by
applying a simple population model is attenuated from the conditional hazard
ratio eβ . The attenuation is increased in situations with larger values of σ 2 , β
and M0 (t), representing the three factors that accelerate frailty selection.
Conversely, the attenuation is modest if any of these factors is near zero, in
which case the individual- and population-averaged parameters are relatively
close to each other. Formal expressions and approximations were obtained by
Henderson and Oman (1999) who consider the effects of fitting a population
model to conditional covariate effects. Such selection effects occur in all
univariate frailty models; they are not restricted to the gamma frailty model.
The advantage of the gamma frailty model is that, for all important quantities,
explicit formulas can be derived in an easy way.

Example 3.4
Assume a gamma-distributed frailty model with constant baseline hazard
function µ0 (t) = λ (exponential distribution). Consequently, using (3.17),
1
S(t) = (1 + σ 2 M0 (t))− σ2
1
= (1 + σ 2 λt)− σ2
1
= (1 + λt)−ζ
ζ
1
= (1 + t)−ζ
ω
ω ζ
=( ) ,
ω+t
with ζ := σ12 and ω := λζ , which is a Pareto distribution with parameters ω
and ζ. Clayton and Cuzick (1985a) considered this model.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 79

3.3.1 Parametric gamma frailty model

As in the proportional hazards model in Section 2.5.1, a parametric as well
as a semiparametric approach is possible. We will consider the parametric
case first. Based on (3.3), the unconditional likelihood function in the frailty
model is of the form
n
Y ′
δi β ′ Xi
L(β, θ|Z1 , . . . , Zn ) = Zi µ0 (ti ; θ)eβ Xi e−Zi M0 (ti ;θ)e . (3.24)
i=1

This conditional likelihood function still depends on the unobserved frailty

variables. Under the assumption of a gamma-distributed frailty, the random
terms can be integrated out using relations (3.17) and (3.18), yielding the
likelihood function
n  ′
Y µ0 (t; θ)eβ Xi δi ′ 1
L(β, θ, σ 2 ) = 2 β ′X (1 + σ 2 M0 (t; θ)eβ Xi )− σ2 . (3.25)
i=1
1 + σ M0 (t; θ)e i

Example 3.5
In Table 3.3, results are given for the parametric gamma frailty model with
Weibull baseline hazard function analyzing the Halluca data. For comparison,
the results of the parametric proportional hazards model with Weibull baseline
hazard function are given as well. The analysis was performed using PROC

Table 3.3: Parametric gamma frailty model

for Halluca data
parameter Weibull PH gamma frailty

age (in years) 0.010 (0.003) 0.012 (0.003)

sex (females) -0.166 (0.071) -0.184 (0.083)
type (NSCLC) -0.136 (0.067) -0.114 (0.078)
ECOG (3 & 4) 0.612 (0.103) 0.861 (0.150)
stage II 0.455 (0.168) 0.516 (0.188)
stage IIIa 0.606 (0.131) 0.701 (0.148)
stage IIIb 0.970 (0.120) 1.109 (0.140)
stage IV 1.358 (0.110) 1.562 (0.136)
λ 0.029 (0.007) 0.020 (0.006)
ν 0.859 (0.019) 0.951 (0.036)
σ2 0.254 (0.084)
log-likelihood -4858.672 -4853.032

NLMIXED in SAS by specifying the likelihood function given in (3.25). The

gamma frailty model with Weibull baseline provides a better fit compared

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
80 Frailty Models in Survival Analysis

to the Weibull proportional hazards model based on the value of the log-
likelihood because of the additional parameter σ 2 . The Weibull proportional
hazards model is nested in the parametric gamma frailty model with Weibull
baseline hazard. Consequently, the models can be compared by the likelihood
ratio test, which yields a test statistic χ2 = −2(−4858.672+4853.032) = 11.28.
Because the parameter value under the null hypothesis σ 2 = 0 to be tested lays
on the boundary of the parameter space, standard methods cannot be applied.
It turns out that the approximation of the likelihood ratio test statistic by
a χ2 -distribution with one degree of freedom is too conservative (Self and
Liang, 1987). Rather, a mixture of a 12 χ20 and a 21 χ21 distribution should
be used (Claeskens et al. 2008). In the present example, this results in
a highly significant p-value. Hence, the additional parameter in the frailty
model causes a significant improvement in the fit to the data. This agrees
with the small standard error of the estimate of the frailty variance, which also
supports the hypothesis σ 2 > 0. In the frailty model – similar to the Weibull
proportional hazards model – an increasing risk of death is seen for increasing
disease stage. However, the interpretation of the hazard ratios in the frailty
model is slightly different from the interpretation in the Weibull proportional
hazards model. For example, the hazard ratio of patients with disease stage
IV is e1.562 =4.77 times higher compared to patients in risk group I given the
same value of frailty. Patients in group I serve as reference group. Hazard
ratios have to be interpreted conditional on the frailty. In the proportional
hazards model with Weibull baseline, the analogous hazard ratio for stage IV
is e1.358 =3.89, meaning an increase in risk by this factor comparing patients
with disease stage IV and I with similar values of all other risk factors in the
model. Consequently, in the Weibull proportional hazards model (as well as
in the semiparametric Cox model), comparisons are made based on (more or
less hypothetical) groups where all observed covariate values are similar. In
the frailty model, comparisons are made based on groups where all observed
and unobserved covariate values are assumed to be similar.
The obtained estimates of the model parameters depend on the parametric
assumption about the Weibull baseline hazard. Furthermore, in the frailty
model, the parameter estimates depend additionally on the assumption of
gamma distribution of frailty. In the following sections the analysis is repeated
with different assumptions about frailty distribution. Typically, an absolute
increase in regression parameter estimates and their standard errors can be
observed. One exception here is variable type.
Different parametric gamma frailty models are applied to the Halluca data.
The baseline hazard function is assumed to be exponential, Gompertz and
Weibull distributed. The results are given in Table 3.4. It is easy to see that
the results depend on the parametric form of the baseline hazard function.
Nevertheless, in all cases, the estimated frailty variance is different from zero
indicating the presence of unobserved heterogeneity. Careful considerations
are necessary because the null hypothesis σ 2 = 0 is again on the boundary
of the parameter space. For further details about tests for heterogeneity

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 81
Table 3.4: Parametric gamma frailty models for Halluca data
parameter Weibull exponential Gompertz

age (in years) 0.012 (0.003) 0.013 (0.004) 0.012 (0.004)

sex (females) -0.184 (0.083) -0.192 (0.086) -0.188 (0.085)
type (NSCLC) -0.114 (0.078) -0.104 (0.082) -0.104 (0.080)
ECOG (3 & 4) 0.861 (0.150) 0.959 (0.139) 0.925 (0.160)
stage II 0.516 (0.188) 0.514 (0.195) 0.540 (0.193)
stage IIIa 0.701 (0.148) 0.726 (0.150) 0.720 (0.154)
stage IIIb 1.109 (0.140) 1.159 (0.139) 1.145 (0.151)
stage IV 1.562 (0.136) 1.630 (0.126) 1.611 (0.150)
λ 0.020 (0.006) 0.017 (0.005) 0.017 (0.005)
ν 0.951 (0.036)
ϕ -0.003 (0.007)
σ2 0.254 (0.084) 0.344 (0.053) 0.307 (0.109)
log-likelihood -4853.032 -4853.888 -4853.812

in the proportional hazards models see Section 7.5. The higher flexibility
of the Gompertz and Weibull baseline hazard does not imply a significant
improvement based on the likelihood function. If the models are not nested,
then the Akaike information criteria (AIC) can be used. Consequently, the
exponential model should be preferred among the three parametric models
presented in Table 3.4.
This links to a general problem in the field of frailty modeling. On the
one hand, parametric estimation procedures (parametric specification of the
baseline hazard) are much easier to implement compared to semiparametric
models. Furthermore, standard errors of the frailty variance estimate are easy
to obtain. On the other hand, in many practical applications the investigator
has only limited information about the form of the baseline hazard function,
which makes parametric analysis difficult. Estimations in parametric frailty
models can be performed with an SAS macro described by Liu and Yu (2008).
This approach requires a closed-form expression of the frailty density, which is,
for example, available for the gamma distribution. Otherwise, in the gamma
case, the integrals in the conditional likelihood function can be integrated
out explicitly without using the approximative method of Liu and Yu (2008).
Here the latter method was used.

Example 3.6
In Table 3.5 the results are given for the parametric gamma frailty model
with exponential baseline hazard analyzing the EBCT data. For comparison,
the results of the parametric proportional hazards model with exponential
baseline hazard function are given as well. The analysis was performed using
PROC NLMIXED in SAS by specifying the likelihood function given in (3.25).
The gamma frailty model with exponential baseline hazard provides a better

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
82 Frailty Models in Survival Analysis
Table 3.5: Parametric gamma frailty model for the
EBCT data
parameter gamma frailty exponential PH

risk group II 0.23 (0.94) 0.23 (0.92)

risk group III 1.31 (0.86) 1.28 (0.83)
risk group IV 2.73 (0.82) 2.53 (0.78)
calcium (> 100) 1.60 (0.57) 1.49 (0.53)
age (55–62) 0.36 (0.51) 0.46 (0.42)
age (63–84) 0.03 (0.49) 0.15 (0.41)
λ 3.4e-4 (2.8e-4) 3.2e-4 (2.6e-4)
σ2 0.73 (0.69)
log-likelihood -219.622 -220.294

fit compared to the exponential proportional hazards model based on the

values of log-likelihood because of the additional parameter σ 2 in the frailty
model. The exponential proportional hazards model is nested in the gamma
frailty model. Consequently, the models can be compared by the likelihood
ratio test, which yields a test statistic χ2 = −2(−220.294 + 219.622) = 1.324.
Because of the nonstandard test situation, a mixture of a 12 χ20 and a 12 χ21
distribution should be used (Claeskens et al. 2008), which results in a p-value
of p = 0.10. Hence, the additional parameter σ 2 in the frailty model does not
cause a significant improvement in the fit of the data. This agrees with the
large standard error of σ 2 , which does not support the hypothesis σ 2 > 0. In
both models we see an increasing risk for major CHD events with the higher
risk group. Again, the interpretation of the hazard ratios in both models
is different. For example, the hazard ratio of patients in risk group IV is
e2.73 =15.3 times higher compared to patients in risk group I given the same
value of frailty with patients in group I serving as reference group. Hazard
ratios have to be interpreted conditional on unobserved frailty. However, this
is in some sense similar to the traditional Cox model. In the Cox model the
relative risk of a patient from risk group IV compared to a patient from risk
group I is conditional on that all other risk factors (included in the model)
are the same for both. In the frailty model the comparison is made under
the condition that all observed, and additionally also the unobserved, factors
(frailty) are equal. Furthermore, patients with calcium values higher than
100 show a e1.6 = 5 times higher risk for an event compared to patients
with a calcium score less than 100 given the same value of frailty. Age does
not play any role as a predictor. Individuals of age 23 – 54 serve as the
reference group. The parameter estimates depend on the assumption about
the exponential baseline hazard and, in the frailty model, on the assumption
of gamma distribution of frailty. Typically, an increase in the parameter
estimates away from zero as well as an increase in the standard errors can be
observed. Exceptions are the estimates for the two age groups.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 83

In Table 3.6 parametric gamma frailty models with Weibull, exponential, and
Gompertz baseline hazard are compared. The gamma frailty model with

Table 3.6: Parametric gamma frailty models for the EBCT data
parameter Weibull exponential Gompertz

risk group II 0.14 (1.05) 0.23 (0.94) 0.21 (0.96)

risk group III 1.62 (1.01) 1.31 (0.86) 1.36 (0.89)
risk group IV 4.29 (1.19) 2.73 (0.82) 3.06 (1.00)
calcium (> 100) 2.19 (0.73) 1.60 (0.57) 1.77 (0.65)
age (55–62) 0.16 (0.71) 0.36 (0.51) 0.27 (0.60)
age (63–84) -0.34 (0.72) 0.03 (0.49) -0.10 (0.59)
λ 4.9e-6 (1.0e-5) 3.4e-4 (2.8e-4) 2.2e-4 (2.4e-4)
ν 2.14 (0.50)
ϕ 0.02 (0.03)
σ2 4.45 (2.01) 0.73 (0.69) 1.60 (1.57)
log-likelihood -215.035 -219.622 -219.402

Weibull baseline hazard function shows the best fit to the data based on the
log-likelihood value. The variance of the frailty is significantly larger than
zero in this model, whereas in the models with exponential and Gompertz
baseline no significant unobserved heterogeneity can be detected. This makes
the main drawback of parametric frailty models obvious. The interpretation of
the results (here especially the question if unobserved heterogeneity is present
or not) depends one the choice of the baseline hazard. Usually, the researcher
has no additional prior information about the form of the baseline distribution.
Consequently, the decision for the best-fitting model should be based on the
likelihood ratio test if the models are nested. Here the nonstandard test
situation for testing the hypothesis σ 2 > 0 versus σ 2 = 0 has been taken into
account. If the models are not nested, then the Akaike information criteria
(AIC) should be used.

The unobserved frailty Zi (i = 1, . . . , n) for each patient can be estimated

based on the distribution of the frailty conditional on event time and status
given in (3.19) and (3.20) by the expression

1/σ̂ 2 + δi
Ẑi = ′ . (3.26)
1/σ̂ 2 + M0 (ti ; θ̂)eβ̂ Xi

Here σ̂ 2 denotes the estimate of the frailty variance, θ̂ the vector of parameter
estimates of the cumulative baseline hazard, and β̂ the vector of estimated
regression coefficients. A parametric gamma frailty model with exponential
baseline hazard is considered, for example, by Tomazella et al. (2008).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
84 Frailty Models in Survival Analysis

3.3.2 Semiparametric gamma frailty model

In the semiparametric frailty model, no assumption about the form of the
baseline hazard function is made. This requires new estimation strategies
compared to the parametric model. In the last section we considered the
extension of the parametric proportional hazards model to the parametric
gamma frailty model. In the following the extension of the semiparametric
proportional hazards (Cox) model to the semiparametric gamma frailty model
will be discussed. The likelihood function looks similar to the parametric
model (3.25), but now the baseline hazard µ0 (t) is treated as a nuisance. In
the following the EM algorithm (expectation-maximization algorithm) will be
explained, which allows parameter estimation in the semiparametric gamma
frailty model. This algorithm was suggested by Dempster et al. (1977) and is
often used in the presence of unobserved data. It was adopted for parameter
estimation in frailty models first by Nielsen et. al (1992), Klein (1992), and
Guo and Rodriguez (1992). The EM algorithm iterates between two steps.
In the first, one estimates the expectations of the unobserved frailties based
on observed data, and current estimates are obtained. These estimates are
now used in the maximization step to obtain new parameter estimates given
the estimated frailties. In the gamma frailty model, closed-form expressions
exist for the conditional expectations of the frailties in the expectation step.
Furthermore, applying the partial likelihood approach with the estimated
frailties as offset terms in the M step is easy to perform, which makes the
EM algorithm useful.
First we consider the full likelihood with the frailty variables assumed to
be observed random variables similar to the event times. This full likelihood
results from the joint density of (ti , δi , Zi ) (i = 1, . . . , n) and is of the form

n
Y
L(β, σ 2 |Z) = f (ti , δi , Zi ; β, σ 2 )
i=1
Yn n
Y
= f (ti , δi , Zi ; β) f (Zi ; σ 2 )
i=1 i=1
= L1 (β|Z)L2 (σ 2 |Z) (3.27)

with Z = (Z1 , . . . , Zn ) and

n
Y ′
δi β ′ Xi
L1 (β|Z) = Zi µ0 (ti )eβ Xi e−Zi M0 (ti )e (3.28)
i=1

from relation (3.25), which is the likelihood function of the observed event
times conditional on the frailties.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 85

The second term is given by the probability density of the frailty variables:
n
Y
L2 (σ 2 |Z) = f (Zi ; σ 2 ).
i=1

If the frailties Zi were known, the regression parameters β could be estimated

′
by the partial likelihood method rewriting the terms Zi eβ Xi in (3.28) in the
′
form eβ Xi +log(Zi ) , using the log(Zi ) as fixed offset values. Consequently,
the expectation step is needed to get estimates of the frailty values. These
estimates are used instead of the unknown frailties in the maximization step
to obtain the estimates for the regression parameters.

Maximization-Step: Extending the partial likelihood idea in the Cox model

described in Section 2.5.1, the partial likelihood for the regression parameters
in the frailty model is similar to equation (2.15):
n  ′
Y eβ Xi +log(Zi ) δi
L(β|Z) = ′ . (3.29)
Zj eβ Xj
P
i=1
j∈R(ti )

The unknown random variables Zi and log(Zi ) are now substituted by their
current expected values (at iteration step k) E(k) (Zi ) and E(k) (log Zi )

n
X h  X i
′
log L(β, σ 2 ) = δi β ′ Xi + E(k) (log(Zi )) − log E(k) (Zj )eβ Xj .
i=1 j∈R(ti )

From this expression, new estimates β(k) can be obtained using standard soft-
2
ware. A new estimate of the frailty parameter σ(k) is derived by maximization
of L2 (σ 2 |Z) also replacing the unknown variables Zi by their current expected
values at iteration step k.

Expectation-Step: Similar to the parametric gamma frailty model (3.26),

the unobserved frailty Zi (i = 1, . . . , n) of each individual can be estimated
by expression
2
1/σ(k) + δi
E(k+1) (Zi ) = β ′(k) Xi
. (3.30)
2 +M
1/σ(k) (k) (ti )e

Here M(k) (·) is a nonparametric estimator of the cumulative baseline hazard

based on the current parameter estimates at iteration step k, for example, the
Nelson–Aalen estimator
X δi
M(k) (t) = P β ′(k) Xj
.
i:ti <t E(k) (Zj )e
j∈R(ti )

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
86 Frailty Models in Survival Analysis

Furthermore, as the conditional distribution of Z is gamma, log(Z) has a

log-gamma distribution with expectation
′
2 2
+ M(k) (ti )eβ(k) Xi


E(k+1) (log(Zi )) = ψ(1/σ(k) + δi ) − log 1/σ(k)

with the digamma function ψ(x) = Γ′ (x)/Γ(x).

Barker and Henderson (2005) investigate the performance of the EM algo-
rithm in the univariate gamma frailty model. In their extensive simulation
study they found a large underestimation of the frailty variance similar (or
even much stronger) compared to the results reported by Nielsen et al. (1992)
in the shared gamma frailty model. The corresponding regression parameters
are underestimated in absolute terms. To correct for this underestimation, an
adaption of the EM algorithm is suggested by Barker and Henderson (2005),
substituting the nonparametric Breslow estimate of the hazard function by a
local likelihood formulation for the baseline hazard that allows the survival
times themselves (instead of their ranks) to enter the estimation terms.

Example 3.7
To account for heterogeneity within the failure times, a semiparametric shared
gamma frailty model was applied to the EBCT data. Results are given in
Table 3.7. The typical inflation of the effect of the covariates is seen, combined

Table 3.7: Semiparametric gamma frailty

model applied to the EBCT data
parameter Cox model gamma frailty

risk group II 0.23 (0.92) 0.19 (0.96)

risk group III 1.27 (0.83) 1.32 (0.89)
risk group IV 2.49 (0.78) 3.16 (1.01)
calcium (> 100) 1.48 (0.53) 1.87 (0.66)
age (55–62) 0.43 (0.42) 0.30 (0.60)
age (63–84) 0.10 (0.40) -0.22 (0.61)
σ2 1.95 (1.64)

with larger standard errors. One exception is the covariate risk group II, whose
estimate lowers from 0.23 (0.92) in the Cox model to 0.19 (0.96) in the gamma
frailty model. The variance of the frailty is large but not significantly different
from zero when taking into account the size of the standard error. The simple
Cox model seems to be appropriate to model the data compared to the more
complex gamma frailty model. The SAS macro SPGAM written by Hien Vu
(Vu and Knuiman 2002) was used for analysis. It is based on a modified
version of the EM algorithm. The above-mentioned correction by Barker and
Henderson (2005) is not included in this macro.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 87

The second important estimation approach to the semiparametric gamma

frailty model is the penalized partial likelihood (PPL) method. Introduced
by McGilchrist and Aisbett (1991) for the log-normal frailty model it was
described in detail by Therneau et al. (2003). It leads to the same estimates
as the EM algorithm in the gamma frailty model (Therneau and Grambsch
2000, Duchateau and Janssen 2008), but this does not hold for other frailty
distributions. For the following presentation it is convenient to rewrite the
frailty model by

′ ′ ′
µ(t|X, Z) = Zµ0 (t)eβ X
= µ0 (t)eβ X+ln(Z) = µ0 (t)eβ X+W
. (3.31)

We will call Z a frailty, and W = ln(Z) a random effect, to distinguish between

both. In the penalized partial likelihood approach, the second part of the
likelihood in (3.27) is considered to be a penalty, implying large penalties to
the likelihood if the value of the random effect is far away from its expectation
value. Considering the random effects as another set of parameters in the first
part of the likelihood, the penalized partial likelihood can be written as the
product of the partial likelihood (2.15) with W = (W1 , . . . , Wn ) as additional
unknown parameters and the density of the random effects
n  ′
Y eβ Xi +Wi δi
2
Lppl (β, σ |W) = P β′ X +W f (Wi ; σ 2 ). (3.32)
i=1
e j j
j∈R(ti )

Consequently, the logarithm of the penalized partial likelihood can be split

into two parts, where the first part does not depend on σ 2 and the second
part does not depend on β:

log Lppl (β, σ 2 |W) = log Lpart (β|W) + log Lpen (σ 2 |W) (3.33)
with
n
Y
Lpen (σ 2 |W) = f (Wi ; σ 2 )
i=1

and the density

2 w 2
(ew )1/σ e−e /σ
2
f (w; σ ) = 2 1/σ2 .
(σ ) Γ(1/σ 2 )
This results in a penalty term of the log-likelihood
n
1 X
log Lpen (σ 2 |W) = − (Wi − eWi ).
σ 2 i=1
The maximization of the penalized partial likelihood consists now of an inner
and an outer loop. The inner loop estimates β and W by maximization of

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
88 Frailty Models in Survival Analysis

log Lppl (β, σ 2 |W) based on a provisional value of σ 2 . In the outer loop, σ 2
is obtained as maximization of a profiled version of the marginal likelihood.
This version is obtained as follows: For a specific value of σ 2 , estimates of
β and W are obtained by maximizing the penalized partial likelihood given
σ 2 . Based on these estimates, other estimates for the cumulative hazard
function can be found using the Nelson–Aalen estimator similar to the EM
algorithm. As in the EM algorithm, this profile marginal likelihood for σ 2 is
used to find an estimate for the frailty variance. The procedure is iterated
until convergence. More details about the EM algorithm and the penalized
partial likelihood approach in the semiparametric gamma frailty can be found
in Duchateau and Janssen (2008). The interested reader is also referred to
this book with respect to a detailed derivation of the fact that the estimates
in both approaches coincide in the gamma frailty model. This is not true in
general for other frailty distributions. In the following example the penalized
partial likelihood approach is applied.

Example 3.8
We reanalyze the Halluca lung cancer data again. A semiparametric gamma
frailty model is applied in Table 3.8. Results of the Cox model are given
for comparisons. As expected, the effect of the covariates is smaller (the

Table 3.8: Semiparametric gamma frailty

model for Halluca data
parameter Cox model gamma frailty

age (in years) 0.009 (0.003) 0.013 (0.004)

sex (females) -0.162 (0.071) -0.197 (0.091)
type (NSCLC) -0.125 (0.067) -0.096 (0.086)
ECOG (3 & 4) 0.610 (0.104) 1.054 (0.145)
stage II 0.455 (0.169) 0.578 (0.203)
stage IIIa 0.594 (0.131) 0.780 (0.157)
stage IIIb 0.947 (0.120) 1.232 (0.146)
stage IV 1.345 (0.110) 1.746 (0.133)
σ2 0.460 ( )

only exception is covariate type) in the Cox model when not corrected for
unobserved heterogeneity in the study population. The gamma frailty model
is able to account (at least in part) for this unobserved heterogeneity. Note
that the standard errors also increase in the gamma frailty model. Estimation
is based on the penalized partial likelihood approach and performed by the
R package coxph. This procedure (adapted from the Splus macro by Thierry
Therneau) provides no standard error of the frailty variance estimate.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 89

The importance of the penalized partial likelihood approach stems especially

from the fact that it can also be used to fit log-normal frailty models considered
in the next section. Furthermore, it is much faster than the EM algorithm.
The major disadvantage is that it is much more complicated to obtain a valid
estimate of the standard error of the frailty variance σ 2 . One solution to obtain
the standard error of the frailty variance estimate are bootstrap resampling
methods. Because of the speed of the penalized partial likelihood estimation
procedure compared to the EM algorithm, bootstrap methods are a realistic
alternative. Unfortunately, such kind of methods are not integrated in the
coxph package in R, which limits its applicability.
Parner (1998) and Therneau et al. (2003) show the exact link between the
gamma frailty model and penalized estimation. In the case of log-normal-
distributed frailty approximation methods of more complex frailty models are
given by Ripatti and Palmgren (2000).
Rondeau et al. (2003) use a different approach and penalize the hazard
function, while Therneau et al. (2003) penalize the frailties. A semiparametric
gamma frailty model is applied to study the effect of aluminium on the risk
of dementia in a large cohort. The aim of the paper is to propose a smooth
baseline hazard estimator by maximizing the penalized log-likelihood
Z ∞
2 2
log Lppl (β, σ ) = log L(β, σ ) − κ µ′′0 (t)2 dt, (3.34)
0

where log L(β, σ 2 ) is the full log-likelihood in the gamma frailty model and
κ is a smoothing parameter. The integral represents a trade-off between the
faithfulness to the data, as represented by the term log L, and smoothness
of the solution, as represented by the squared norm of the second derivative
of the baseline hazard in the penalty term. For large κ, the integral will be
forced toward zero, and the baseline hazard will approach a linear function. If
κ is small, then the main contribution to log Lppl will be log L and the curve
estimate will track the data closely, but will be more irregular.
Other approaches such as the Markov Chain Monte Carlo (MCMC) method
have also been suggested for estimation in the semiparametric gamma frailty
model. The new PROC MCMC routine in SAS offers flexible alternatives
compared to the traditional WinBugs software often used for MCMC analyses.
Another choice for estimation in frailty models is Gaussian quadrature. It
approximates the integral of a parametric function with respect to a frailty
density by a weighted sum over predefined abscissas for the frailties. However,
all current implementations make a parametric choice of the baseline hazard
function (especially exponential, piecewise constant), see Yamaguchi et al.
(2002), Littell et al. (2006), Nelson et al. (2006), Liu and Huang (2008),
and Liu and Yu (2008). However, a piecewise constant hazard function with
raising number of pieces can be used to approximate semiparametric models.
The SAS routine PROC NLMIXED provides a powerful tool for adaptive as
well as nonadaptive Gaussian quadrature.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
90 Frailty Models in Survival Analysis

3.3.3 Gamma frailty model for current status data

The case of current status data in frailty models is considered in this section.
The likelihood function for current status data is given by formula (2.8):
n
Y
δi
L(θ) = 1 − S(ti ; θ) S(ti ; θ)1−δi .
i=1

Plugging in the survival function of the gamma frailty model with Gompertz
baseline from Example 3.3 and θ = (λ, ϕ), the likelihood function becomes
n 
Y λ
− 1 δi λ
− 1−δ2 i
L(λ, ϕ, σ 2 ) = 1 − 1 + σ 2 (eϕti − 1) σ2 1 + σ 2 (eϕti − 1) σ
.
i=1
ϕ ϕ

Example 3.9
The hepatitis A and B current status data from Example 1.7 are analyzed
in detail. A univariate parametric gamma frailty model with the Gompertz
baseline hazard function as described earlier is fitted to the hepatitis data.
Age-dependent prevalences of hepatitis A and B are shown in Figures 3.7 and
3.8, respectively. Results of the analysis are given in Table 3.9. The univariate
frailty model treats time to infection by hepatitis A and B, respectively, as
independent times for each individual. It turns out that the estimation of the
frailty variance regarding hepatitis B is difficult, because the likelihood is very
flat with respect to σ 2 . This is also indicated by a large standard error of σ 2 .
A reason for this variability is the smaller prevalence of hepatitis B compared
to hepatitis A.
0.8
prevalence

0.6
0.4
0.2

0 20 40 60 80

age

Figure 3.7: Age-specific prevalence of hepatitis A in Flanders 1994/95. The

circles indicate the size of the respective age group on which the prevalence
estimates are based.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 91

Another, maybe more important reason is the irregularity in the relationship

between age and prevalence of hepatitis B, which can be seen in Figure 3.8.
Results depend on the choice of the baseline. Here a Gompertz model for the
baseline hazard was assumed, other distributions are possible. The likelihood
function was maximized by using PROC NLMIXED in SAS.

Table 3.9: Gamma frailty model for

hepatitis A and B current status data
parameter hepatitis A hepatitis B

σ2 2.018 (0.510) 10.037 (8.821)

λ 0.008 (0.001) 0.002 (0.001)
ϕ 0.085 (0.016) 0.015 (0.142)

The parameter σ 2 provides a nice interpretation as a measures of population

heterogeneity in the susceptibility to hepatitis A and B, respectively. This
parameter provides important implications for further programs to prevent
infections as the critical vaccination coverage is higher for more heterogeneous
populations. More details about this data are given in Hens et al. (2009).

Another possibility to analyze interval-censored data in a frailty model is

provided by the R package developed by Henschel et al. (2009).
0.00 0.05 0.10 0.15 0.20
prevalence

0 20 40 60 80

age
Figure 3.8: Age-specific prevalence of hepatitis B in Flanders 1994/95. The
circles indicate the size of the respective age group on which the prevalence
estimates are based.

© 2011 by Taylor and Francis Group, LLC www.Ebook777.com

 Free ebooks ==> www.Ebook777.com
92 Frailty Models in Survival Analysis

3.3.4 Extensions of the gamma frailty model

An interesting extension of the univariate gamma frailty model was suggested
by Semenchenko et al. (2004a,b). Consider the problem of comparing control
and treatment group in animal experiments or clinical trials. The distinctive
feature of the model compared to other frailty models is that it assumes
that treatment influences parameters of both frailty distribution and baseline
hazard function. The survival function of the control group is assumed to be

− 1
S1 (t) = 1 + σ 2 M01 (t) σ2
with cumulative baseline hazard function M01 (t). The survival function of
the treatment group can be written as
 λ
− γσ1 2
S2 (t) = 1 + rγσ 2 M01 (t) + (eϕt − 1) .
ϕ
One can see that the model has five unknown parameters λ, ϕ, r, γ, and σ 2 .
Parameter σ 2 indicates the presence of heterogeneity in control and treatment
group. Parameters λ and ϕ describe changes in the baseline hazard caused
by treatment. The model is based on the relation

λ ϕt
M02 (t) = M01 (t) + (e − 1),
ϕ
where M02 (t) denotes the cumulative baseline hazard of the treatment group.
Frailty in the treatment group is now gamma distributed with mean r and
variance γσ 2 . Hence, changes in the frailty distribution caused by treatment
are given by parameters r and γ. Parameter r < 1 shows an increase in
the average robustness (decrease in mean frailty), while r > 1 indicates an
accumulation of frail individuals in the treatment group. A parameter value
γ 6= 1 signalizes an increase (γ > 1) or decrease (γ < 1) in the heterogeneity
by the treatment. In the above expression, σ 2 M01 (t) can be substituted by
2
S1 (t)−σ − 1 resulting in
 2 λ − 1 2
S2 (t) = 1 + rγ(S1 (t)−σ − 1) + rγ (eϕt − 1)
γσ
.
ϕ
The nonparametric Kaplan–Meier estimator can now be used to estimate
the unknown survival function S1 (t) of the control group. Consequently, the
suggested model is semiparametrically in this sense.
Another extension of the gamma frailty model was recently introduced by
Barker and Henderson (2004). The authors claim that it is likely that only a
part of the covariates used in a Cox regression fulfill the proportional hazards
assumption. To circumvent this problem they introduce a mixed model that
allows for both proportional and converging hazards. Two kinds of covariates
are considered, X1 and X2 . The usual gamma frailty model with covariates
′
X2 has a hazard of the form µ(t) = Zµ0 (t)eβ2 X2 , but the frailty distribution

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 93
β ′ X1
is assumed to depend on covariates X1 , namely, Z ∼ Γ( e σ12 , σ12 ). Then the
unconditional survival function is obtained as
Z ∞
β ′ X2
S(t|X1 , X2 ) = e−zM0 (t)e 2 f (z|X1 ) dz
0
′
 − eβ12X1
2 β ′2 X2 σ
= 1 + σ M0 (t)e . (3.35)

When the frailty variance σ 2 is zero, the model reduces to the standard Cox
proportional hazards model with covariates {X1 , X2 }. If β 2 = 0, the model
reduces to the proportional hazards model with covariates X1 , whereas if
β 1 = 0 holds, it is clear from equation (3.35) that the suggested mixed
model reduces to the univariate gamma frailty model with covariates X2 .
The baseline hazard function in this model is left unspecified.
In the following we a nontypical extension of the univariate gamma frailty
model is presented to demonstrate the strengths of this approach. Consider a
bivariate outcome (T, L) in the following analysis, where the first component
T is an event time, but the second one L is a nonnegative integer number.
For example, T could be the lifetime of a woman, and L the number of their
children. The aim of such an approach is to analyze both outcomes in parallel,
assuming a dependence between lifetime and number of children caused by
unobserved underlying factors influencing lifetime as well as fecundability.
The difference to a Cox regression model with number of children as covariate
is the causal relationship. If number of children is modeled as an independent
variable, and lifetime as a dependent variable, a causal relationship between
number of children and lifetime is assumed. The model presented in this
paragraph assumes an association between both variables, but no assumption
about the direction of a causal relationship is made. Let the expression
S(t|Z) = e−ZM0 (t)
denote the conditional survival function of lifetime T , conditional on the
frailty (random effect) Z. Furthermore, let p > 0, and the integer num-
ber L be conditional Poisson distributed, conditional on the random effect Z
with parameter pZ:
(pZ)l −pZ
P(L = l|Z) = e .
l!
The dependence between T and L is based on the unobserved random effect
Z. Let Z be gamma distributed with expectation one and variance σ 2 . Then
the model can be derived in the following way:
P(T > t, L = l) = EP(T > t, L = l|Z)
= ES(t|Z)P(L = l|Z)
(pZ)l −pZ
= Ee−ZM0 (t) e
l!
1
Z
= ezM0 (t) e−pz (pz)l λk z k−1 e−λz dz.
Γ(k)l!

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
94 Frailty Models in Survival Analysis

Consequently,

λk Γ(k + l)pl (λ + M0 (t) + p)k+l z k+l−1 e−(λ+M0 (t)+p)z dz

R
P(T > t, L = l) =
Γ(k)l!(λ + M (t) + 1)k+l Γ(k + l)
λk Γ(k + l)pl
=
Γ(k)l!(λ + M0 (t) + p)k+l
 
l+k−1 λ p
= ( )k ( )l
k−1 λ + M0 (t) + p λ + M0 (t) + p

Now we consider the two marginal distributions in more detail. The event
time distribution is obtained by

∞
X
S(t) = P(T > t, L = l)
l=0
∞  
X l+k−1 λ p
= ( )k ( )l
k−1 λ + M0 (t) + p λ + M0 (t) + p
l=0
∞  
1 −k
X l+k−1 λ + M0 (t) k p
= (1 + M0 (t)) ( ) ( )l
λ k−1 λ + M0 (t) + p λ + M0 (t) + p
l=0
1
= (1 + M0 (t))−k ,
λ
which results because the infinite sum is the probability function of a negative
λ+M0 (t)
binomial distribution with parameters k and λ+M 0 (t)+p
. If now the restriction
k = λ = σ12 is used, we obtain the unconditional survival function with gamma
frailty in (3.17).
The count data distribution can be derived from the joint distribution by
 
l+k−1 λ k p l
P(L = l) = P(T > 0, L = l) = ( ) ( ),
k−1 λ+p λ+p
λ
which is a negative binomial distribution with parameters k and λ+p . Using
again parameterization k = λ = σ12 results in

l + σ12 − 1   12 pσ 2 l
 
1 σ
P(L = l) = 1 ( ),
σ2 − 1 1 + pσ 2 1 + pσ 2
1
a negative binomial distribution with parameters σ12 and 1+pσ 2 , implying

expectation p and variance p(1+pσ 2 ). The parameters p and σ 2 give the model
enough flexibility to fit the marginal distribution of the number of children
in the population. Furthermore, σ 2 can be interpreted in some sense as an
association parameter. Large values speak in favor of a strong association
between lifetime and number of children, whereas small values do not.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 95

The negative binomial distribution has been widely used to model count data.
One of the first applications was used by Greenwood and Yule (1920) to
model the number of accidents of women working in munition shell production
during World War I in England. Up to this time the Poisson distribution was
considered as a universal count data distribution based on the book by von
Bortkiewicz (1898). Greenwood and his colleagues observed discrepancies
between the observed and expected numbers of accidents and suggested an
extension of the Poisson distribution to the negative binomial distribution
by assuming the Poisson parameter to be a random mixture variable that
follows a gamma distribution. Consequently, heterogeneity was assumed in
the population. This was the starting point for a series of accident studies
based on the theory of accident proneness. The idea was that it should be
able to reduce the number of accidents by identifying the most accident-prone
individuals with the help of such heterogeneity models.
There are many applications of the univariate gamma frailty model in the
literature. For the first time the model was used by Beard (1959). Lancaster
(1979) suggested this model for the duration of unemployment spells, and
Vaupel et al. (1979) used it to calculate individual hazard functions and to
adjust life tables in the case of heterogeneous populations. Manton et al.
(1981) compared the mortality experience of heterogeneous populations, and
Manton and Stallard (1981) explained the black/white mortality crossing-over
effects observed in the United States. Manton et al. (1986) compared the
inverse normal and the gamma models, together with Gompertz and Weibull
baseline hazards, in a study of survival at advanced ages, based on data from
U.S. Medicare insurance. Aalen (1987) studied the expulsion of intrauterine
contraceptive devices. Ellermann et al. (1992) used a parametric model to
analyze recidivism among criminals. Andersen et al. (1993) applied the model
to the malignant melanoma data from Example 1.2. Jones (1998) used a
gamma–Gompertz model for analyzing the impact of selective lapsation on
mortality in life insurance. Jeong et al. (2003) used a gamma frailty to model
long-term follow-up survival data from breast cancer clinical trials when the
treatment effect diminishes over time as an alternative to the proportional
hazards model. Solid cancer incidence data from atomic bomb survivors in
Japan were used by Izumi and Ohtaki (2004) to check two different hypotheses
about individual hazard functions. Balakrishnan and Peng (2006) consider an
extension of the gamma frailty model which also includes the log-normal frailty
model considered in the following section. The use of gamma-distributed
frailty in univariate models is supported by the results of Abbring and van
den Berg (2007), who showed that, under some regularity assumptions, frailty
among survivors converges against a gamma distribution even if the original
frailty distribution is not a gamma distribution.
Parametric gamma frailty models are implemented in STATA and in the
SAS macro PGAM by Vu. Furthermore, NLMIXED in SAS can be used
based on Liu and Yu (2008). Semiparametric versions are included in R/S
plus (Therneau) and SAS macros by Vu (SPGAM) and Klein (GAMFRAIL).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
96 Frailty Models in Survival Analysis

3.4 Log-normal Frailty Model

The log-normal distribution together with the gamma distribution is the most
important and often-used frailty distribution. Increasing computer power
helps to overcome the restrictions of this model, which needs the solution
of numerical integrals because of the lack of closed-form expressions for the
marginal survival function and the resulting terms in the likelihood. The
strong link to random effects/mixed models makes this model very attractive.
As it will be shown in the subsequent chapters, log-normal frailty models are
especially useful in modeling dependence structures in multivariate problems.
Their main advantage is the large flexibility in the dependence structures by
means of multivariate normal distribution. However, log-normal distribution
has also been applied in univariate cases, for example, by Flinn and Heckman
(1982). Log-normal distribution is in fact very close to inverse Gaussian
distribution.
Two variants of the log-normal frailty model exist. We assume a normally
distributed random variable W ∼ N(m, s2 ) to generate frailty as Z = eW ,
with N denoting the normal distribution. The parameters of this normal
distribution are some functions of the frailty parameters µ and σ 2 (see, for
example, Hutchinson and Lai 1991)

s2
µ = EZ = em+ 2 (3.36)
2 2
σ 2 = V(Z) = e2m+s (es − 1). (3.37)

Simple calculations show that under the constraint µ = 1 this results in

1
m = E ln(Z) = − s2
2
s2 = V(ln(Z)) = ln(1 + σ 2 ) .

First, following the historical definition of frailty originally introduced in the

field of demography (Vaupel et al. 1979), one can use the restriction µ = 1.
Second, one can use the restriction m = 0. This means that the logarithm of
the frailty has a mean of zero. In this case, a ”standard”
individual has the
logarithm of the hazard rate, which is equal to ln µ0 (t) . Any individual in
a population has the logarithm of the hazard rate distorted by some random
variable W =
ln(Z). This value is added to the ”true” logarithm of the hazard
rate ln µ0 (t) to provide the logarithm of hazard rate of the individual. In
this interpretation, it is natural to assume that the distortions W have a
normal distribution with an expectation of zero.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 97

2.2
s=0.5
2.0 s=1.0
s=2.0
1.8
1.6
1.4
density

1.2
1.0
0.8
0.6
0.4
0.2
0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0

Figure 3.9: Probability density functions of log-normal distributions with

parameter values m = 0 and s2 = 0.25, 1, and 4.

The density of the frailty distribution based on the parameters m and s2 is

given by the expression
1 (ln(z))2
f (z) = √ e− 2s2 ,
2πsz
where the form of the density for different parameters is shown in Figure 3.9.
In this model, a ”standard” individual has the hazard rate µ0 (t). Individual
i has the hazard rate of a ”standard” individual multiplied by the frailty Zi .
This restriction on µ means that the average frailty in a population equals 1 (at
the beginning of the follow-up). This makes the model comparable with other
frailty models considered in the preceding sections. Despite this advantage
of comparability of frailty variances in different models, the parameterization
based on EW = m = 0 is much more common in the literature and standard
in statistical packages. The reason is that the normally distributed random
effect acts on the linear predictor, which suits nicely generalized linear models
and the assumption of normality of the linear predictor. To take advantage
of the theoretical research results and software solutions provided in the fast-
growing field of mixed models, we will use this approach in the following as
well. Furthermore, interpretation of the log-normal frailty model as a random
effects model with normally distributed random effects makes this model class
more convenient for application by researchers not especially trained in the
field of survival analysis/frailty models.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
98 Frailty Models in Survival Analysis

3.4.1 Parametric log-normal frailty model

Unfortunately, no explicit form of unconditional likelihood exists because
of the intractable Laplace transform. The likelihood of the data (ti , δi , Xi )
(i = 1, . . . , n) is of the form

n Z
Y ′
δi β ′ Xi +wi
L(β, θ, s2 ) = µ0 (ti ; θ)eβ Xi +wi e−M0 (ti ;θ)e dΦ(wi ) (3.38)
i=1

with Φ as cumulative distribution function of the normal distribution with

mean zero and variance s2 . Consequently, more sophisticated estimation
strategies are required, for example, numerical integration in the maximum
likelihood approach or MCMC methods. In the parametric case, numerical
integration of the normally distributed random effects based on the Laplace
approximation or (adaptive/nonadaptive) Gaussian quadrature can be used.
We follow this approach here by making three different assumptions about
the baseline hazard in the Halluca study. The results are given in Table 3.10.

Table 3.10: Parametric log-normal frailty models for Halluca

lung cancer data
parameter Weibull exponential Gompertz

age (in years) 0.013 (0.004) 0.013 (0.004) 0.013 (0.004)

sex (females) -0.192 (0.089) -0.198 (0.088) -0.203 (0.091)
type (NSCLC) -0.126 (0.084) -0.120 (0.083) -0.132 (0.086)
ECOG (3 & 4) 0.938 (0.149) 0.900 (0.136) 0.958 (0.146)
stage II 0.545 (0.203) 0.533 (0.200) 0.562 (0.207)
stage IIIa 0.738 (0.159) 0.742 (0.155) 0.769 (0.163)
stage IIIb 1.197 (0.152) 1.181 (0.143) 1.229 (0.156)
stage IV 1.669 (0.147) 1.654 (0.130) 1.713 (0.149)
s2 0.516 (0.134) 0.475 (0.073) 0.578 (0.134)
λ 0.013 (0.005) 0.014 (0.004) 0.012 (0.004)
ν 1.013 (0.041)
ϕ 0.006 (0.006)
log-likelihood -4848.24049 -4848.26798 -4847.85248

The estimates of the regression parameters are very similar to each other,
indicating robustness of the analysis with respect to the choice of the baseline
hazard. Small differences occur between the estimates of the random effects
variance. The likelihood ratio test indicates that the simple exponential model
is significantly better compared to the more complex Gompertz or Weibull
model. Analysis was performed by SAS procedure NLMIXED. The variance
s2 of the random effect differs significantly from zero in all three models. It
cannot be directly compared with frailty variance σ 2 from the gamma model.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 99

3.4.2 Semiparametric log-normal frailty model

If the form of the underlying baseline hazard function is known, the parametric
approach is efficient. However, in most practical applications, this information
is not available, and semiparametric models are preferred. If no assumptions
about the form of the baseline hazard function are made, parameter estimation
becomes much more difficult. One way to solve the estimation problem is the
penalized partial likelihood approach, already considered in the case of the
gamma frailty model. Similar to the gamma case, the random effects are
considered as another set of parameters in the first part of the likelihood.
Note that, due to (3.31), the relation EW = 0 is used to keep the model
identifiable instead of EZ = 1. Furthermore, the model is parameterized in
terms of the variance of the random effects V(W ) = s2 , which is different from
the gamma model. Based on (3.32), the logarithm of the penalized partial
likelihood can be split in a partial likelihood Lpart and a penalty term Lpen
and written in the form

log Lppl (β, s2 |W) = log Lpart (β|W) + log Lpen (s2 |W)
with W = (W1 , . . . , Wn ),
n
Y
Lpen (s2 |W) = f (Wi ; s2 )
i=1

and the density of a normal distribution with mean zero and variance s2

1 w2
f (w; s2 ) = √ e− 2s2 .
2πs2
This results in a penalty term of the log-likelihood
n
1 X Wi2
log Lpen (s2 |W) = − + ln(2πs2 ) .


2 i=1 s 2

The maximization of the penalized partial likelihood consists again of an

inner and an outer loop. The inner loop is similar to the gamma frailty
model and estimates β and W by a Newton–Raphson procedure to maximize
log Lppl (β, s2 |W) based on a provisional value of s2 (best linear unbiased
predictor – BLUB). The outer loop is different from the gamma model. The
restricted maximum likelihood estimator for s2 is based on the BLUBs. Then
the procedure is iterated until convergence occurs. The asymptotic variance
for the estimates of the regression parameters can be found in McGilchrist and
Aisbett (1991). Asymptotic variance of the random effects variance estimate
is given in McGilchrist (1993). The more technical details of the algorithm can
be found elsewhere (Therneau and Grambsch 2000, Duchateau and Janssen
2008) and are therefore omitted here.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
100 Frailty Models in Survival Analysis

Example 3.10
The penalized partial likelihood approach is illustrated with an application to
the Halluca data. In Table 3.11 the results of a semiparametric log-normal
frailty model are given with the Cox model for comparison. We see the typical
pattern of stronger covariate effects (estimates of the regression parameters
are further away from zero) with variable type as an exception. Furthermore,
standard errors for the regression coefficients are larger in the frailty model.
Estimation was performed by means of the R procedure COXPH based on
a penalized partial likelihood approach. This procedure does not provide
a standard error for the frailty variance estimate. Comparing the results

Table 3.11: Cox and semiparametric log-normal

frailty model for Halluca lung cancer data
parameter Cox model log-normal frailty

age (in years) 0.009 (0.003) 0.012 (0.003)

sex (females) -0.162 (0.071) -0.194 (0.088)
type (NSCLC) -0.125 (0.067) -0.112 (0.083)
ECOG (3 & 4) 0.610 (0.104) 0.924 (0.134)
stage II 0.455 (0.169) 0.552 (0.198)
stage IIIa 0.594 (0.131) 0.742 (0.154)
stage IIIb 0.947 (0.120) 1.177 (0.142)
stage IV 1.345 (0.110) 1.667 (0.130)
s2 0.498 ( )

of the log-normal frailty model and the gamma frailty model in Table 3.8,
one has to note that frailty in the log-normal model with a parameterization
based on EW = m = 0 is not standardized to EZ = µ = 1. Despite this
fact the estimates for the regression coefficients are similar in both models.
Further caution is needed when variances in the two models are compared.
The parameter σ 2 describes the variance of the frailty term Z in the gamma
model, whereas s2 denotes the variance of the random effect W = ln(Z) in
the log-normal frailty model. Both quantities cannot be directly compared.
However, both estimates are similar (σ 2 = 0.46 compared to s2 = 0.498).
In general, the variance of the random effect W in a log-normal frailty model
is larger than the variance of the respective frailty Z = eW (in the same
model), and the difference becomes more pronounced the larger the frailty
variance is. In situations with reasonable small frailty variance, there is no
big difference between σ 2 and s2 (Duchateau and Janssen 2008). The results
show robustness regarding the frailty distribution. Similar to the gamma
model, the results from parametric and semiparametric models support the
robustness of parametric frailty analysis regarding the baseline hazard.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 101

3.5 Inverse Gaussian Frailty Model

The inverse Gaussian (inverse normal) distribution was introduced as a frailty
distribution alternative to the gamma distribution by Hougaard (1984) and
was used, for example, by Manton et al. (1986), Klein et al. (1992), Keiding et
al. (1997), Price and Manatunga (2001), Economou and Caroni (2005), Kheiri
et al. (2007), and Duchateau and Janssen (2008). Similar to the gamma frailty
model, simple closed-form expressions exist for the unconditional survival and
hazard functions, which makes the model attractive. The probability density
function of an inverse Gaussian distributed random variable with parameters
µ > 0 and λ > 0 is given by

√
λ λ
(z − µ)2 .


f (z) = √ exp −
2πz 3 2µ2 z

The form of the density function for different parameter values is depicted in
Figure 3.10.

1.6

1.4

1.2

1.0
density

0.8

0.6

0.4

0.2

0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0

Figure 3.10: Probability density functions of inverse Gauss distributions

with expectation 1 and variances 0.5, 1, and 2.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
102 Frailty Models in Survival Analysis

The density given above is the starting point to derive the Laplace transform
of the inverse Gaussian distribution:

L(u) = Ee−uZ
Z √
λ −uz λ
exp − 2 (z − µ)2 dz


= √ e
2πz 3 2µ z
Z √
λ (λ + 2µ2 u)z 2 − 2µλz + λµ2

= √ exp − dz
2πz 3 2µ2 z
√
λ z λ + 2µ2 u λ λ

Z
= √ exp − 2
+ − dz
2πz 3 2 µ µ 2z
q
2
λ 1 + 2µλ u λ

= exp − +
µ µ
Z √
q
2
λ λz 1 + λ 2µ2 u λ 1 + 2µλ u λ

× √ exp − 2
+ − dz (3.39)
2πz 3 2 µ µ 2z

Because of the relation

q
2 2µ2 u
λz 1 + 2µλ u λ 1+ λ λ λ µ
− + − =− µ2
(z − q )2 ,
2 µ2 µ 2z 2 2 z 2µ2 u
1+ λ
1+ 2µλ u

the expression

√
q
2
λ  2µ2 u
λz 1 + λ λ 1 + 2µλ u λ
√ exp − + − =
2πz 3 2 µ2 µ 2z
√
λ  λ µ
2 
√ exp − 2 z − 2
2πz 3 2 µ2µ2 u z 1 + 2µλ u
1+ λ

in (3.39) is the density of an inverse Gaussian distribution with parameters

q µ and λ. Thus, the integral equals one. As a consequence, the Laplace
2µ2 u
1+ λ
transform of an inverse Gaussian distributed random variable simplifies to the
expression

q
2µ2 u
λ 1+
r

λ λ λ 2µ2 u

L(u) = exp − + = exp 1− 1+ . (3.40)
µ µ µ λ

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 103

The first and second derivatives of the Laplace transform are given by
r
′ µ λ 2µ2 u

L (u) = − q exp 1− 1+
2 µ λ
1 + 2µλ u
and

r
′′ µ3 λ 2µ2 u

L (u) = 2µ2 u 3/2
exp 1− 1+
λ(1 + µ λ
λ )
r
µ2 λ 2µ2 u

+ 2µ2 u
exp 1− 1+ .
1+ µ λ
λ

Expectation and variance of the frailty distribution can now be calculated by

evaluating the derivatives at u = 0:

EZ = −L′ (0) = µ

µ3
V(Z) = L′′ (0) − (L′ (0))2 = .
λ
Taking EZ = µ = 1 and V(Z) = σ 2 = 1/λ results in the following simplified
Laplace transform:
1
√
1+2σ2 u)
L(u) = e σ2 (1− .
Hence, the unconditional survival and hazard function can be written in the
form
1
√ 2
S(t) = e σ2 (1− 1+2σ M0 (t))
and

µ0 (t)
µ(t) = .
(1 + 2σ 2 M0 (t))1/2
′
In the following we include regression terms eβ X into the model. The aim is
to evaluate the effect of unobserved heterogeneity on the regression parameters
in the inverse Gaussian frailty model. We consider the marginal hazards of
two individuals with only one binary covariate in the model. The hazard
ratio for two individuals with covariate values one (meaning, for example,
experimental treatment) and zero (for example, placebo treatment) is of the
form

µ(t|X = 1) (1 + 2σ 2 M0 (t))1/2 β
= e .
µ(t|X = 0) (1 + 2σ 2 M0 (t)eβ )1/2

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
104 Frailty Models in Survival Analysis

This hazard ratio is eβ at time point t = 0 and converges towards eβ/2 for
t → ∞. Similar to the gamma frailty model, this marginal model is no longer
a proportional hazards model. The effect of the covariate weakens over time.
The time-independent hazard ratio obtained from the conditional model is
attenuated from the conditional hazard ratio eβ . The influence of unobserved
heterogeneity is stronger in situations with larger values of σ 2 , β, or M0 (t),
representing the three factors that accelerate selection.
The density of the frailty distribution among the survivors at time point t
can be written in the form

S(t|X, z)f (z)

f (z|X, T > t) =
S(t|X)
2

exp − zM0 (t)eβ X exp − (z−1)
′

2σ2 z
= √ p
2π 2 z 3 exp σ12 (1 − 1 + 2σ 2 M0 (t)eβ X )
′

2 2 β′ X
exp − z (1+2σ M2σ 0 (t)e
2z
)−2z+1

= √ p
2π 2 z 3 exp σ12 (1 − 1 + 2σ 2 M0 (t)eβ X )
′


′
2
1  z − (1 + 2σ 2 M0 (t)eβ X )−1/2 
= √ exp − 2σ2 z
.
2π 2 z 3 1+2σ2 M0 (t)eβ ′ X

This is the density of an inverse Gaussian-distributed frailty variable with

expectation

1
E(Z|X, T > t) = p
1 + σ 2 M0 (t)eβ
′X

among the survivors at time t. This demonstrates selection by early death of

the high-risk individuals, for example, individuals with high values of frailty
Z and high values of observed risk factors. The variance of frailty among the
survivors at time point t is given by

σ2
V(Z|X, T > t) = .
(1 + σ2 M0 (t)e
β ′ X )2

The frailty distribution of those who die at time t is a ”generalized” in-

verse Gaussian one, which is a much more complicated distribution, and is
not considered in detail here. Parametric versions of the inverse Gaussian
frailty model are available in the STATA package. Possible choices for the
baseline hazard in STATA are exponential, Weibull, Gompertz, log-logistic,
log-normal, and gamma distributions.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 105

3.6 Positive Stable Frailty Model

A distribution is called stable if the normalized sum of n independent random
variables from this distribution has the same distribution as a scale factor1
multiplied by a single random variable. The normalization is given by n γ ,
where the parameter γ is from the interval (0, 2] and called a characteristic
exponent. To ensure a distribution on the positive numbers, we restrict here
to the case of positive stable distributions, characterized by γ ∈ (0, 1]. The
probability density function of such a one-parameter positive stable random
distribution is given by (Feller 1971)
∞
1X Γ(κγ + 1) −κγ−1
f (z) = (−1)κ+1 z sin(κγπ)
π κ=1 κ!

with z ≥ 0 and 0 < γ ≤ 1. This expression is a power series, converging fast for
large values of z, and slow for small values of z. In the special case of γ = 1, the
frailty distribution becomes degenerated at the point mass Z = 1. Although
the probability density function of a random variable with positive stable
distribution can only be represented by infinite series, the Laplace transform
has a very simple form. This makes the distribution especially attractive as a
frailty distribution because many characteristics of the unconditional survival
distribution of the event times follow from the Laplace transform and can
easily be deduced:
γ
L(u) = e−u . (3.41)
All moments of this distribution are infinite. Consequently, the expectation
of the frailty is infinite, and variance does not exists. In the first view this
seems to be a disadvantage because an infinite expectation is more difficult
to work with. However, infinite expectation was one of the main reasons why
this frailty distribution was introduced. The first derivative of the Laplace
transform is used to show this fact for the expectation (Duchateau and Janssen
2008)
γ
′ e−u
lim L (u) = −γ lim 1−γ = −∞,
u→+0 u→+0 u

where lim denotes the right limit. Here the comparison of the population
u→+0
hazard function with the conditional hazard function for a subject with frailty
value one makes no sense because the expectation of the frailty variable does
not exist. Consequently, an individual with frailty equal to one cannot serve
as a standard individual as in other frailty models with finite expectation of
the frailty. This infinite expectation result is also important with respect to
identifiability issues treated by Elbers and Ridder (1982). They found that a

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
106 Frailty Models in Survival Analysis

finite mean of the frailty distribution is one condition (among others) for the
identifiability of univariate frailty models. This was the main reason why the
positive stable distribution was introduced as a frailty distribution. Especially
in bivariate/multivariate applications, much attention is given to overcoming
confounding problems in shared (gamma) frailty models, considered later in
this book (for more details see Section 4.9). Using the Laplace transform of
a positive stable frailty variable given earlier, the unconditional survival and
density functions are
γ
S(t) = e−M0 (t) (3.42)

and
γ
f (t) = γµ0 (t)M0 (t)γ−1 e−M0 (t) ,

respectively, resulting in the unconditional hazard function

µ(t) = γµ0 (t)M0 (t)γ−1 . (3.43)

The positive stable model implies some interesting features, which are given
in the following examples. The most important one is that the positive stable
distribution is the only frailty distribution that preserves the proportional
hazards condition in unconditional hazards after integrating out the frailty.
In the following we introduce observed covariates to the model. In the case of
only one binary 0-1 covariate, the ratio of two hazards with different individual
covariate values is

µ(t|X = 1)
= eγβ .
µ(t|X = 0)
The two marginal hazards are still proportional but with proportionality
factor eγβ instead of eβ . This shows that parameter estimates are biased
in a proportional hazards model if relevant covariates are not included. Since
0 < γ < 1, the population hazard ratio will typically be closer to one. The
more the parameter γ deviates from one, the more the population hazard ratio
will deviate from the conditional hazard ratio.

Example 3.11
Let M0 (t) = λtν (Weibull) and assume that the frailty variable is positive
stable distributed with parameter γ. Then the resulting survival function is
given by
γ γ γν
S(t) = e−M0 (t) = e−λ t
.

This is again a Weibull distribution but with parameters λγ and γν.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 107

Example 3.12
In the following, the positive stable frailty model is applied to the Halluca
data. The survival and hazard function, respectively, are given by (3.42) and
(3.43) for a parametric analysis. Results can be found in Table 3.12. The

Table 3.12: Parametric positive stable frailty models for

Halluca lung cancer data with different baseline hazard functions
parameter Weibull exponential Gompertz

age (in years) 0.011 (0.001) 0.011 (0.001) 0.010 (0.003)

sex (females) -0.192 (0.085) -0.193 (0.083) -0.176 (0.078)
type (NSCLC) -0.162 (0.078) -0.161 (0.077) -0.139 (0.073)
ECOG (3 & 4) 0.708 (0.121) 0.705 (0.121) 0.663 (0.114)
stage II 0.535 (0.193) 0.532 (0.196) 0.489 (0.184)
stage IIIa 0.702 (0.155) 0.707 (0.152) 0.649 (0.144)
stage IIIb 1.134 (0.146) 1.131 (0.140) 1.038 (0.135)
stage IV 1.576 (0.124) 1.584 (0.128) 1.457 (0.129)
γ 0.863 (0.018) 0.859 (0.019) 0.919 (0.031)
λ 0.017 (0.004) 0.016 (0.004) 0.020 (0.005)
ν 0.999 (0.042)
ϕ -0.013 (0.005)
log-likelihood -4858.66462 -4858.66470 -4855.45463

estimates of the regression parameters are very similar, indicating robustness

of the analysis with respect to the choice of the baseline hazard function. The
values of the log-likelihood function indicate that the Weibull and exponential
models provide a similar fit to the data. The Gompertz model fits the data
significantly better than the two other models. In the positive stable model no
moments of frailty exist. Consequently, the model parameter γ ∈ (0, 1] has no
interpretation as frailty variance. Despite this fact, γ can be used as a measure
of heterogeneity in the population. For γ = 1, there is no heterogeneity in
the population. The smaller the values of γ, the larger the heterogeneity. In
all three models the parameter estimates for γ are significantly different from
one. The foregoing analysis was performed by maximizing the parametric
log-likelihood function using the SAS procedure NLMIXED.

The positive stable frailty distribution was introduced as a frailty distribution

by Hougaard (1986b) and applied by Wang et al. (1995), Lam and Kuk
(1997), Qiou et al. (1999), and Manatunga and Oakes (1999). Fine et al.
(2003) and Martinussen and Pipper (2005) recently suggested new estimation
procedures in the shared positive stable frailty model. It was further extended
by Hougaard’s power variance function distribution (Hougaard, 1986a) and
Aalen’s compound Poisson distribution (Aalen 1988, 1992).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
108 Frailty Models in Survival Analysis

3.7 PVF Frailty Model

A generalized family of frailty distributions that includes gamma, inverse
Gaussian, and positive stable distributions is the family of power variance
function distributions suggested by Tweedy (1984) and derived independently
by Hougaard (1986a). This is a three-parameter family with parameters
µ > 0, λ > 0 and 0 < γ ≤ 1. The probability density function is given
by

z 1
f (z) = e−λ(1−γ)( µ − γ )
∞
1X (λ(1 − γ))κ(1−γ) µκγ Γ(κγ + 1) −κγ−1
× (−1)k+1 z sin(κγπ)
π κ=1 γκ κ!

The derivation of the Laplace transform from this density function is difficult
and therefore omitted here. The interested reader is referred to Aalen (1992).
The Laplace transform is
λ(1−γ) µu


1−(1+ λ(1−γ) )γ
L(u) = e γ . (3.44)
The first and second derivatives are given by
µu λ(1−γ) µu


1−(1+ λ(1−γ) )γ
L′ (u) = −µ(1 + )γ−1 e− γ
λ(1 − γ)
and

µ2 µu λ(1−γ) µu


)γ
′′
L (u) = (1 + )γ−2 e− γ 1−(1+ λ(1−γ)
λ λ(1 − γ)
µu λ(1−γ) µu


)γ
+ µ2 (1 + )2γ−2 e− γ 1−(1+ λ(1−γ)
λ(1 − γ)

Similar to the gamma model, expectation and variance of a PVF-distributed

random variable Z are derived with the help of the first and second derivatives
of the Laplace transform:

EZ = −L′ (0) = µ (3.45)

µ2
V(Z) = L′′ (0) − (L′ (0))2 =
. (3.46)
λ
The positive stable distribution can be obtained as a special case of the PVF
distribution. To show this fact, some asymptotic considerations are necessary.
Following the presentation in the book by Duchateau and Janssen (2008), we

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 109

rewrite the logarithm of the Laplace transform of a PVF-distributed random

variable in the following way:

λ(1 − γ)   µu γ 
log L(u) = 1− 1+
γ λ(1 − γ)
λ(1 − γ)   µ γ  λ(1 − γ) γ 
= 1− +u
γ λ(1 − γ) µ
λ(1 − γ)  µ γ  λ(1 − γ) γ  λ(1 − γ) γ 
= − +u (3.47)
γ λ(1 − γ) µ µ

Now it is necessary that µλ converges to zero in an appropriate way by µ

converging toward infinity and λ converging toward zero. Assuming the
convergence

µγ γ
γ−1
−−−−−−−→ ,
λ µ→∞,λ→0 (1 − γ)1−γ

the first factor in the expression (3.47) converges against one, and the second
one goes to −uγ . This is the logarithm of the Laplace transform of a positive
stable-distributed random variable.
Using the standard assumption in frailty models that EZ = µ = 1, and
introducing the frailty variance as a model parameter by the relationship
V(Z) = λ1 := σ 2 , the Laplace transform (3.44) of a PVF random variable
becomes


1−γ σ2 u γ
1−(1+ 1−γ )
L(u) = e γσ2 .
This implies the unconditional survival and hazard function in the PVF frailty
model
σ2 M0 (t) γ
1−γ


2 1−(1+ )
S(t) = e γσ 1−γ

and

µ0 (t)
µ(t) = σ2
.
(1 + 1−γ
1−γ M0 (t))

If γ = 0, the gamma frailty model with frailty distribution Γ(1/σ 2 , 1/σ 2 ) is

easily obtained from the foregoing hazard function by comparison with (3.18).
When comparing the Laplace transforms, the gamma distribution is derived
as a limiting case after application of the rule by L’Hospital. For γ = 0.5, the
inverse Gaussian distribution is obtained.
To quantify the effect of unobserved heterogeneity in the PVF frailty model,
observed covariates are introduced into the model. Similar to the other frailty
models, we assume again for ease of presentation that there exists only one

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
110 Frailty Models in Survival Analysis

single binary covariate with possible individual values zero and one in the
model. Then the ratio of the marginal hazards becomes

2
σ
µ(t|X = 1) (1 + 1−γ M0 (t))1−γ β
= σ 2 e ,
µ(t|X = 0) (1 + 1−γ M0 (t)eβ )1−γ

which is eβ at the beginning of the follow-up t = 0 and converges towards eγβ

as time goes to infinity. This is the same value as the ratio of the marginal
hazards in the positive stable frailty model. However, in contrast to the
positive stable frailty model, the marginal hazards are not parallel in the
PVF frailty model.
To calculate the frailty distribution among the survivors at a specific time
point, we restrict again to the case µ = 1 and drop the observed covariates to
keep the formulas simple:

S(t|z)f (z)
f (z|T > t) =
S(t)


exp − zM0 (t) 1−γ
(z− γ1 )
= 1−γ σ2 M0 (t) γ


e− σ2

exp γσ2 1 − (1 + 1−γ )

∞
1 X (1 − γ)κ(1−γ) Γ(κγ + 1) −κγ−1
× (−1)κ+1 z sin(κγπ)
π κ=1
σ 2κ(1−γ) γ κ κ!


z
−(1−γ)
− γ1
=e σ2 A−γ Aγ−1

∞
1X (1 − γ)κ(1−γ) Aκγ(γ−1) Γ(κγ + 1) −κγ−1
× (−1)κ+1 z sin(κγπ)
π κ=1 (σ 2 A−γ )κ(1−γ) γ κ κ!

This is the density of a PVF-distributed random variable with parameters

2
A = 1 + σ 1−γ M0 (t)
(instead of µ), σ 2 A−γ (instead of σ 2 ) and γ.
On the one hand, the PVF frailty model is easily tractable because of closed-
form expressions for the marginal survival function. On the other hand, the
model is flexible in the sense that it contains many other interesting frailty
models as special cases. That makes the model attractive for applications, and
it is often used. Crowder (1989) studied the PVF frailty model to generate a
multivariate distribution with Weibull marginals and one parameter governing
association, which may be positive or negative. Hougaard et al. (1992) applied
the model to lifetimes of Danish twins. Mallick and Ravishanker (2004) used a
PVF frailty distribution and a Weibull baseline hazard. Inference was carried
out in the Bayesian framework using Markov Chain Monte Carlo techniques.
The model was extended regarding the baseline hazard function by Mallick
and Ravishanker (2006).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 111

3.8 Compound Poisson Frailty Model

The compound Poisson distribution was introduced by Aalen (1988, 1992)
as a frailty distribution. The probability density function is similar to the
density of the PVF distribution
z 1
f (z) = e−λ(1−γ)( µ − γ ) (3.48)
∞
1 X (λ(1 − γ))κ(1−γ) µκγ Γ(κγ + 1)
× (−z)−κγ−1 sin(κγπ),
π κ=1 γκ κ!

with parameters µ > 0, λ > 0 and (different from the PVF distribution) γ < 0.
An interesting property of the model is that it allows for a subgroup of zero
frailty, which never experiences the event of interest. In the case of total
mortality, this is impossible because nobody is immortal. But this model is
relevant to medicine and demography, for example, when considering cause-
specific mortality or the occurrence of a disease. Despite the fact that the
density of the continuous part is only given as an infinite series that has to be
calculated numerically, the distribution is mathematically convenient. It may
also be seen as a natural choice. The distribution can be constructed as the
sum of a Poisson-distributed number of independent and identically gamma-
distributed random variables. This can be viewed as a hit model where each
individual experiences a random number of hits, each of random size. More
formally, the frailty Z of individuals in the study population is

V1 + V2 + . . . + VN : if N > 0,
Z= (3.49)
0 : if N = 0,

where N is Poisson distributed with expectation ρ. The variables V1 , V2 , . . .

are independent and gamma distributed with Vi ∼ Γ(k, λ). Furthermore, N
is assumed to be independent of V1 , V2 , . . . Laplace transforms of gamma and
−u
Poisson distributions are given by LV (u) = (1 + λu )−k and LN (u) = e−ρ+ρe ,
respectively. The Laplace transform of the frailty can now be derived by

L(u) = Ee−uZ
∞
X
E e−u(V1 +...+VN ) |N = κ P(N = κ)


= P(N = 0) +
κ=1
∞ Y
κ
X u −k ρκ −ρ
= e−ρ + (1 + ) e
κ=1 i=1
λ κ!
∞
X (ρ(1 + uλ )−k )κ
= e−ρ
κ=0
κ!
u −k
= e−ρ(1−(1+ λ ) )
. (3.50)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
112 Frailty Models in Survival Analysis

In the following we will switch to another, more convenient parameterization,

given by
kλγ
ρ=− , λ = λ, k = −γ.
γ
The Laplace transform of a compound Poisson-distributed random variable
then becomes
 kλγ u
{1 − (1 + )γ } .

L(u) = exp
γ λ
The first and second derivatives of the Laplace transform are

u γ−1 kλγ u
L′ (u) = −kλγ−1 (1 + ) exp[ {1 − (1 + )γ }]
λ γ λ
γ
u kλ u
L′′ (u) = −kλγ−2 (γ − 1)(1 + )γ−2 exp[ {1 − (1 + )γ }]
λ γ λ
γ
u kλ u
+ k 2 λ2γ−2 (1 + )2γ−2 exp[ {1 − (1 + )γ }].
λ γ λ

Expectation and variance of frailty distribution can now be calculated by

evaluating the derivatives at u = 0

EZ = kλγ−1 (3.51)
γ−2
V(Z) = k(1 − γ)λ . (3.52)
1−γ
Using the standard relation EZ = 1 and σ 2 = λ , the Laplace transform
becomes
1 − γ σ 2 u γ

L(u) = exp 2
1 − (1 + ) . (3.53)
γσ 1−γ
This Laplace transform implies the marginal survival and hazard function in
the case of a compound Poisson frailty model given by
σ2 M0 (t) γ
1−γ


2 1−(1+ )
S(t) = e γσ 1−γ
(3.54)
and
µ0 (t)
µ(t) = σ2
.
(1 + 1−γ
1−γ M0 (t))

Because the Laplace transforms of the PVF distribution and the compound
Poisson distribution are equal (except for the range of parameter γ, which can
be negative in the compound Poisson model), the properties obtained from the
Laplace transform are the same. Consequently, the population survival and

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 113

hazard functions of the compound Poisson frailty model coincides with the
respective functions in the PVF frailty model. Also, the ratio of the marginal
hazards is similar to the PVF model when introducing a single binary observed
covariate into the model

σ2

1−γ
µ(t|X = 1) 1 + 1−γ M0 (t) β
=
1−γ e ,
µ(t|X = 0) σ2
1 + 1−γ M0 (t)e β

which is eβ at t = 0 and converges toward eγβ for t → ∞.

Using the frailty representation (3.49), it is easy to derive the density of the
continuous part of the distribution following the presentation in Duchateau
and Janssen (2008). Denote by f (z|N = κ) the probability density function
of the sum V1 + . . . + Vκ . Because V1 , . . . , Vκ are i.i.d. random variables with
gamma distribution Γ(k, λ),
s −κk
Ee−(V1 +...+Vκ )s = 1 + ) ,
λ
meaning V1 + . . . + Vκ ∼ Γ(κk, λ). Hence, the density can be written as
∞
X
f (z) = f (z|N = κ)P(N = κ)
κ=1
∞
X λκk z κk−1 e−λz ρκ −ρ
= e
κ=1
Γ(κk) κ!
∞
X λκk ρκ κk−1
= e−ρ−λz z
κ=1
Γ(κk)κ!
1−γ −κγ
∞
(− 1−γ κ


1−γ
− 1−γ z µσ2 γσ2 )
X
= e γσ2 µσ2 z −κγ−1
κ=1
Γ(−κγ)κ!
∞ 1−γ κ(1−γ) κγ


− 1−γ ( z
− γ1 )
X 2 µ
=e σ 2 µ σ
κ
(−z)−κγ−1
κ=1
γ Γ(−κγ)κ!
∞ 1−γ κ(1−γ) κγ


− 1−γ z
2 (µ−γ )
1 1 X
σ2 µ Γ(κγ + 1)
=e σ (−z)−κγ−1 sin(κγπ),
π κ=1 γκ κ!

where the last equation holds because of the relation

1 1
= Γ(κγ + 1) sin(κγπ).
Γ(−κγ) π

This is the density given in (3.48) with λ = σ12 . It should be noted that
the integral of µ(t) over [0, ∞) is finite if γ < 0. Consequently, the survival
function is incomplete because a fraction of individuals has zero frailty and

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
114 Frailty Models in Survival Analysis

will never experience the event under study. The size of the nonsusceptible
fraction is given by


1−γ σ2 1−γ
1−(1+ 1−γ M0 (t))γ
S(t) = e γσ2 −−−→ e γσ2 .
t→∞

The parameter γ divides the class of distributions in two major subfamilies:

For γ ≥ 0, the distribution is a PVF distribution. The extension to γ < 0 was
suggested by Aalen (1988) and yields the compound Poisson distribution. The
two subclasses are separated by the gamma distribution (γ = 0). Note that a
different parameterization was applied by Aalen. The notation cP (γ, k, λ) is
used for a compound Poisson distribution.
It is necessary to mention that a connection exists between the probability
of susceptibility and the frailty distribution in the compound Poisson frailty
model. A model relaxing this assumption is considered in Section 3.12.
The model was used by Aalen (1992) to model the incidence of marriage
of women born in Denmark. Marriage is an example of an event that not
everybody experiences. A certain percentage of individuals never marry, and
models of marriage incidence have to account for this. The fact that the
observed incidence peaks around age 23 and becomes rather low after age 30
is therefore interpreted to be a selection phenomenon due to heterogeneity.
This means that those who are most prone to marriage will marry quite early,
and the reminder will have a lower tendency to marry.
A second interesting application of Aalen (1992) deals with fertility data
in Norwegian women. It is a well-known fact that around 5% to 10% of
all couples are unable to conceive children. Naturally, P(Z = 0) (where Z
denotes the frailty to conceive a child) will be the probability of infertility, and
the variation of Z expresses the varying fecundabilities among fertile couples.
Hougaard et al. (1994) applied the model to diabetic nephropathy onset
data, a serious complication for insulin-dependent diabetic patients. Data are
interval-censored because measurement of protein takes place only at selected
time points. Less than 50% of diabetics with long-term survival develop
nephropathy, suggesting that only some but not all patients are susceptible.
Aalen and Tretli (1999) applied the compound Poisson model to testicular
cancer. Testicular cancer has two striking epidemiological features. First,
its incidence has increased rapidly over the past few decades. Second, the
incidence is greatest among younger men, and then declines from a certain
age. The idea of the model is that a subgroup of men is particularly susceptible
to testicular cancer, which results in selection over time. The model is fit to
incidence data from the Norwegian Cancer Registry collected between 1953
and 1993. This work is being continued by Moger et al. (2004a).
Haukka et al. (2003) applied the model to schizophrenia data from the
Finnish population born between 1950 and 1968. They concluded that only
a small part of the population is susceptible to schizophrenia and found
increasing individual risk with higher age among the susceptible part of the
study population.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 115

Example 3.13

To illustrate the many interesting features of the compound Poisson frailty

model, the malignant melanoma data from Example 1.2 is considered. Here
we apply three parametric models with Gompertz baseline to this data: the
proportional hazards model, the gamma frailty model, and the compound
Poisson/PVF frailty model. The results are given in Table 3.13.

Table 3.13: PH, gamma frailty, and compound Poisson frailty

model applied to the malignant melanoma data
parameter Gompertz model gamma frailty cP/PVF frailty

age 0.014 (0.009) 0.011 (0.011) -0.001 (0.015)

gender -0.528 (0.266) -0.729 (0.388) -0.886 (0.509)
thickness 0.144 (0.033) 0.241 (0.114) 0.286 (0.150)
λ 5.5e-5 (3.2e-5) 4.6e-5 (3.6e-5) 4.0e-5 (4.7e-5)
ϕ -1.9e-5 (1.4e-4) 3.7e-7 (4.2e-7) 1.4e-6 (1.1e-6)
σ2 2.064 (2.029) 3.823 (3.151)
γ -0.822 (1.153)
log-likelihood -554.638 -554.153 -553.666

The proportional hazards model with Gompertz baseline hazard suggests that
women face a lower risk of death caused by malignant melanoma compared
to men. Increasing age and tumor size result in higher mortality. The gamma
frailty model implies an increase in the effect of the covariates gender and
thickness, whereas the effect of age is decreased. Standard errors are also
increased. The heterogeneity estimate σ 2 is not significantly greater than
zero. The compound Poisson/PVF frailty model extends the gamma frailty
model. Estimates of the regression parameters are different from the estimates
in the gamma frailty model. The compound Poisson model shows only a
nonsignificant improvement in the fit compared to the gamma frailty model
(γ = 0), which is also supported by the large standard error of the estimate
of parameter γ. Comparing all three models, the proportional hazards model
with Gompertz baseline shows the best fit with respect to the likelihood ratio
test. The estimate of γ is negative, implying that this is a compound Poisson
frailty model. Consequently, some patients will never die from this disease.
This is possible because only death by malignant melanoma was studied and
not total mortality. That means a part of the population seems to be protected
against mortality from malignant melanoma. The estimated proportion with
1−γ
zero risk is around e γσ2 = 56%. A possible conclusion would be that this
is a cure or nonsusceptible fraction. But caution is necessary, especially in
the present case where the compound Poisson frailty model does not show a
significant better fit compared to the gamma frailty model.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
116 Frailty Models in Survival Analysis

3.9 Quadratic Hazard Frailty Model

Making appropriate distributional assumptions about frailty, it is often possible
to find a complete analytical form of the likelihood function. One disadvantage
of the log-normal frailty model considered in Section 3.4 the absence of explicit
expressions for the survival and hazard functions. To avoid this problem and
to combine the flexibility of the normal distribution with advantages of frailty
models that provide an analytic survival function, Aalen (1987) and Yashin
and Iachine (1996) suggested the quadratic hazard frailty model. Their model
is based on the idea of quadratic conditional hazard and normal distribution
of the random effects. Assume Z = W 2 and let W be a normally distributed
random variable with W ∼ N (m, s2 ). The model is given in terms of the
conditional hazard function (without covariates):

µ(t|Z) = Zµ0 (t) = W 2 µ0 (t).

The Laplace transform can be calculated in the following way:

L(u) = E exp(−uW 2 )
Z∞  (w − m)2 
1
= √ exp(−w2 u) exp − dw
2πs 2s2
−∞
Z∞  2s2 w2 u + w2 − 2mw + m2 
1
= √ exp − dw
2πs 2s2
−∞
Z∞ w2 − 2mw
+ m2 m2 m2
1 (1+2s2 u)2 + 1+2s2 u −
 
1+2s2 u (1+2s2 u)2
= √ exp − s2
dw
2πs 2 1+2s 2u
−∞
m2 m2
1 1+2s2 u − (1+2s2 u)2
 
= √ exp − s2
1 + 2s2 u 2 1+2s 2u

Z∞  (w − m 2 )2 
1 1+2s u
× √ s
exp − s2
dw
2π 1+2s2 u
√ 2 1+2s 2u
−∞
1  m2 u 
= √ exp − , (3.55)
1 + 2s2 u 1 + 2s2 u

where the last equation holds because the integral equals to one as it is over
the density of a normal distribution. It is easy to see that, in case of m = 0,
the frailty Z = W 2 is gamma distributed with shape parameter k = 1/2 and
form parameter λ = 1/2s2.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 117

The first and second derivatives need some algebra.

m2 u m2 u
3 − 5 −
L′ (u) = −s2 (1 + 2s2 u)− 2 e 1+2s2 u − m2 (1 + 2s2 u)− 2 e 1+2s2 u

m u 2 m u 2
7 − 1+2s 5 − 1+2s
L′′ (u) = 2m2 s2 (1 + 2s2 u)− 2 e 2u
+ 3s4 (1 + 2s2 u)− 2 e 2u

m2 u m2 u
7 − 7 −
= 3m2 s2 (1 + 2s2 u)− 2 e 1+2s2 u + m2 s2 (1 + 2s2 u)− 2 e 1+2s2 u

m u 2
11 − 1+2s
= m4 (1 + 2s2 u)− 2 e 2u

Expectation and variance of the frailty Z = W 2 are then given by

EZ = −L′ (0) = m2 + s2
and

V(Z) = L′′ (0) − (L′ (0))2

= 3s4 + 6m2 s2 + m4 − (s2 + m2 )2
= 2s4 + 4m2 s2 .
The unconditional survival function is

1  m2 M0 (t) 
S(t) = p exp − ,
1 + 2s2 M0 (t) 1 + 2s2 M0 (t)
resulting in the unconditional hazard function
 m2 s2 
m(t) = m0 (t) + .
(1 + 2s2 M0 (t))2 1 + 2s2 M0 (t)
One nice feature of this model is the Gaussian property of the conditional
distribution P(W ≤ w|T > t):

P(T > t, W ≤ w)
P(W ≤ w|T > t) =
S(t)
Z w
1
= P(T > t|s)fW (s) ds
S(t) −∞
Z w
1 1 (s − m)2

= √ exp(−s2 M0 (t)) exp − ds
S(t) 2πs −∞ 2s2
 s2 − 2ms m2
1+2s2 M0 (t) + (1+2s2 M0 (t))2
Z w
1 1 
= √ exp − 2 ds
S(t) 2πs −∞ 2 1+2ss2 M0 (t)
 (s − m 2
1+2s2 M0 (t) )
Z w
1
= √ s
exp − 2 ds.
2π √ 2
−∞ 2 1+2ss2 M0 (t)
1+2s M0 (t)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
118 Frailty Models in Survival Analysis

Consequently, W (square root of frailty) among survivors of age t is normally

distributed with parameters

m s2

W |(T > t) ∼ N 2
, 2
.
1 + 2s M0 (t) 1 + 2s M0 (t)
The common constraint EZ = EW 2 = 1 implies the relation s2 + m2 = 1,
which restricts possible values of m and s2 and limits the applicability of the
model to real-life problems. Substituting m2 by the expression 1 − s2 (and
keeping in mind s2 ≤ 1) results in the following unconditional survival and
hazard functions:
(1−s2 )M0 (t)
1 − 1+2s2 M0 (t)
S(t) = p e (3.56)
1 + 2s2 M0 (t)
and

1 + 2s4 M0 (t)
µ(t) = µ0 (t) . (3.57)
(1 + 2s2 M0 (t))2
To quantify the effect of unobserved heterogeneity in the quadratic hazard
frailty model, observed covariates are introduced into the model. Considering
only one single binary 0-1 covariate in the model, the ratio of the marginal
hazards becomes

µ(t|X = 1) 1 + 2s4 M0 (t)eβ (1 + 2s2 M0 (t))2 β

= e ,
µ(t|X = 0) (1 + 2s2 M0 (t)eβ )2 1 + 2s4 M0 (t)
which is eβ at t = 0 and converges towards eβ as time goes to infinity. Hence,
the marginal hazards are not time independent as in the positive stable frailty
model, but the ratio of the marginal hazards converges against the ratio of the
conditional hazards. This means that the marginal and conditional hazard
ratios coincide at time point zero, are becoming more and more different
during the first part of the follow-up and finally converge again as time goes
to infinity.
It is necessary to keep in mind that s2 is the variance of the random effect W
but not the variance of the frailty term Z = W 2 . To underline this difference,
s2 is used as the variance parameter for the random effect, whereas σ 2 is used
as the variance parameter for the frailty term. Frailty variance can easily be
calculated by using the 4th moments of normal distributed random variables
and the restriction m2 = 1 − s2 introduced above for identifiability reasons

V(Z) = EW 4 − (EW 2 )2
= 2s4 + 4m2 s2
= 2s2 (2 − s2 ). (3.58)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 119

As a consequence, it holds that 0 ≤ V (Z) ≤ 2. This means that the amount

of unobserved heterogeneity is restricted in this model by an upper bound.
Furthermore, it is easy to see that, in the limiting case s2 → 0, the original Cox
proportional hazards model without unobserved heterogeneity is obtained.

Example 3.14
The quadratic hazard frailty model is applied to the Halluca lung cancer data
from Example 1.3. The hazard and survival functions have an explicit form in
this model (see (3.56) and (3.57)) and can be used in the likelihood function
for parametric analysis. Similar to other frailty models considered before a
Weibull, exponential, and Gompertz baseline hazard function is used, and the
results of the analysis are given in Table 3.14.

Table 3.14: Parametric quadratic hazard frailty models with

different baseline hazard functions applied to the Halluca lung
cancer data
parameter Weibull exponential Gompertz

age (in years) 0.012 (0.002) 0.013 (0.004) 0.013 (0.004)

sex (females) -0.183 (0.084) -0.188 (0.086) -0.188 (0.086)
type (NSCLC) -0.104 (0.080) -0.101 (0.082) -0.098 (0.082)
ECOG (3 & 4) 0.943 (0.160) 1.007 (0.145) 1.012 (0.145)
stage II 0.522 (0.190) 0.540 (0.194) 0.541 (0.194)
stage IIIa 0.707 (0.149) 0.738 (0.150) 0.738 (0.150)
stage IIIb 1.119 (0.140) 1.159 (0.138) 1.159 (0.138)
stage IV 1.584 (0.134) 1.642 (0.126) 1.643 (0.126)
s2 0.065 (0.017) 0.078 (0.011) 0.078 (0.011)
λ 0.018 (0.006) 0.016 (0.004) 0.016 (0.004)
ν 0.964 (0.035)
ϕ 1e-8 (0.001)
log-likelihood -4851.86641 -4852.41278 -4852.41104

Estimates of the regression parameters are very similar in the three models,
indicating robustness of the analysis with respect to the choice of the baseline
hazard function. The values of the log-likelihood function indicate that the
Weibull and Gompertz models do not fit the data significantly better than the
exponential model. Only small differences occur between the estimates of the
random effects variance. In general, the variance of the random effects is small
(but significantly away from zero in all three models). Using formula (3.58),
the variance of the frailty term σ 2 is around 0.3. Analysis was performed by
maximizing the parametric log-likelihood function using the SAS procedure
NLMIXED.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
120 Frailty Models in Survival Analysis

3.10 Lévy Frailty Models

A common problem is statistics is the comparison of groups with respect
to their event time outcome. When applying a frailty model one usually
assumes the same frailty distribution for all groups and that the individuals
in one group have a proportionally higher conditional hazard than in the other
groups. Integrating out the frailty and calculating the unconditional hazards,
one finds that these hazards are no longer proportional (with one exception
– the positive stable distribution). Aalen and Hjort (2002) suggest the use of
different frailty distributions in the groups. One way of constructing this is
to think of a damage process that has generated frailty in the groups. This
process is assumed to have run its course before the current observations
start. In high-risk groups, this process has been running for a longer time, or
at a higher rate, thereby yielding greater frailty than in low-risk groups. The
use of frailty distributions based on Lévy processes implies proportionality of
the population hazards, which is of course a very convenient feature of this
approach.
Aalen and Hjort (2002) and Aalen et al. (2008) consider frailty distributions
which are generated by a nonnegative Lévy process D = {D(s), s ≥ 0},
that is a random process with nonnegative, independent, time-homogeneous
increments. The Laplace transform of such a damage process is given by the
Lévy–Chinchin formula

L(u; s) = e−sψ(u) .
This is a valid Laplace transform for all nonnegative s when function ψ is a
characteristic exponent of the Lévy process. Here s denotes the time which
the frailty generating Lévy process has been running prior to the follow-up.
Larger values of s indicate higher risks because of a longer running damage
process. More details about Laplace exponents and Lévy processes can be
found in Bertoin (1996). The family of Lévy processes contains a number of
interesting special cases such as compound Poisson processes, PVF processes,
gamma processes, stable processes, etc. The survival function in this model
is given by

S(t) = e−sψ(M0 (t))

and the unconditional hazard function is

µ(t) = sµ0 (t)ψ ′ M0 (t) .

One can see that s is a proportionality parameter and there is a natural link to
the proportional hazards model. Of course, the special Laplace transform of a
Lévy process produces this result. Aalen and Hjort (2002) present a number
of examples of frailty models based on Lévy processes. The majority of the

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 121

frailty distributions applied in practice follow from them. The connection to

the compound Poisson frailty model in the parameterization (3.50) used by
Aalen and Hjort (2002) is obtained by ψ(u) = ρ 1 − (1 + λu )−k , resulting in

the Laplace transform

L(u; s) = e−sρ(1−(1+ λ ) ) .
u −k

Here it is clearly to see that the additional parameter s influences only ρ from
the original model which is the parameter of the Poisson distribution of the
number of damaging hits. The parameters k and λ describing the gamma
distribution of the hits are not influenced by s. It is necessary to note that
the compound Poisson frailty distribution in this model depends on s and is
therefore no longer restricted to expectation one. In the gamma frailty model
as another example the characteristic exponent is of the form
u
ψ(u) = k ln(1 + ) (3.59)
λ
which gives
u −sk
L(u; s) = (1 +
) .
λ
This is the Laplace transform of a gamma distribution Γ(sk, λ). Consequently,
s influences only the shape parameter of the gamma distribution, but not the
form parameter. Again, the restriction of zero expectation does no longer hold
for different values of s if the other two parameters are fixed. For different
values of s proportional hazards are obtained, which is clearly a nice and
interesting feature of this approach, especially when event times of different
groups need to be compared.

3.11 Log-t Frailty Model

The logarithm of a t-distribution is used as a frailty distribution in very few
cases in the literature. Matsuyama et al. (1998) used this approach in a multi-
center cancer clinical trial to compare the effect of institutions. Sahu and Dey
(2004) applied the model to the kidney infection data by McGilchrist and
Aisbett (1991) and rat litter tumor data by Mantel, Bohidar, and Ciminera
(1977). Sargent (1998) applied the model also to rat tumor data. All papers
used a Bayesian approach, which is rather complex and does not allow fitting
analytic methods. Consequently, MCMC methods were used to estimate the
model parameters. The R package COXPH as alternative includes the log-t
distribution as an option for frailty (beside gamma and log-normal frailty
distribution). But the author does not have good experience with this model
because several example data could not be fitted due to numerical problems
in the R package.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
122 Frailty Models in Survival Analysis

3.12 Univariate Frailty Cure Models

The Cox model is commonly used in the analysis of survival time data. An
often unstated assumption of this model (and of frailty models with frailty
distributions without mass point zero) is that all individuals are susceptible to
experiencing the event of interest. Consequently, if follow-up is long enough,
all individuals will experience the event. However, in some situations it
makes sense to assume the existence of a fraction of individuals who are not
expected to experience the event. These individuals are nonsusceptible or
cured, for example, by a genetic predisposition or a vaccination. For instance,
researchers may be interested in analyzing the recurrence of a disease. Many
individuals may never experience a recurrence of that disease; therefore, a
fraction in the study population exists that has recovered from disease and is
now protected. Historically, cure models were first presented by Boag (1949)
and Berkson and Gage (1952), and have been utilized to estimate the cured
(or nonsusceptible) fraction. Cure models are survival models that allow for
a cured fraction of individuals. These models extend the understanding of
time-to-event data by allowing for the formulation of more accurate and infor-
mative conclusions than previously made. These conclusions would otherwise
be unobtainable from an analysis that fails to account for a cured fraction
in the population. If a cured fraction component is not present, the analysis
is reduced to standard approaches of survival analysis. A key characteristic
of data modeled by cure models is that there needs to be sufficient follow-up
time to determine that a cure fraction is present in the study population. For
the proportion of individuals in the population experiencing the event under
study, there exist many ways to model the distribution of their event times.
For example, Peng et al. (1998) used the generalized F distribution, whereas
Kuk and Chen (1992) and Sy and Taylor (2000) applied a proportional hazards
approach to model event times.
Cure models assume that individuals prone to experience the event are
homogeneous in risk. This section deals with extensions of cure models in
order to allow for heterogeneity among the fraction under risk by using frailty
models. Or, depending on the point of view, it deals with extensions of frailty
models to allow for a cured fraction in the study population. In this case, the
distribution of frailty is a combination of discrete and continuous distributions.
Spilerman (1972) considered the ”spiked-gamma” as an example of such a
distribution.
In cure models, the population is divided into two subpopulations. An
individual is either cured with probability 1 − φ, or has a proper survival
function S(t), with probability φ. Individuals regarded as cured will never
experience the event of interest, and their survival time will be defined as
infinity. Therefore, hazard and survival functions of cured individuals are set
to zero and one, respectively, for all finite values of t. A model of survival

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 123

times that incorporates a cured fraction is given by Berkson and Gage (1952):

S ∗ (t) = (1 − φ) + φS(t).

Longini and Halloran (1996) have proposed frailty cure (cure-mixture) models
that extends a model by Farewell (1977). In their model the frailty variable
has point mass at zero with probability 1− φ while heterogeneity among those
experiencing the event of interest is modeled via a continuous distribution with
probability φ. In the gamma frailty cure model, the survival function S of the
susceptible individuals is substituted by the marginal survival function in the
gamma frailty model (3.17):
2
S ∗ (t) = (1 − φ) + φ(1 + σ 2 M0 (t))−1/σ . (3.60)

The idea behind this model is similar but not equivalent to the compound
Poisson frailty model suggested by Aalen (1988, 1992). In the model by
Longini and Halloran (1996), the shape of the frailty distribution is assumed
to be independent of the size of the cure fraction, while in the compound
Poisson frailty model there exists a connection between the shape of the frailty
density and the probability of susceptibility.
An interesting work about cure models is that of Maller and Zhou (1996).
Price and Manatunga (2001) give a comprehensive introduction to the area
and apply different cure, frailty, and frailty cure models to leukemia remission
data. They conclude that frailty models are useful in modeling data with a
cured fraction and found that the gamma frailty cure model provides a better
fit to their remission data compared to the standard cure model.
The next example provides an extension of the foregoing model to include
censored observations. Consider two types of expressions for a disease: the
incidence and the age of disease onset. Risk models for overall susceptibility
(lifetime risk) that consider only the first expression by treating the disease
as a binary trait of being affected or not can give wrong results. The reason
is, that for individuals without the disease, due to censoring, it is often not
known whether they will eventually develop the disease. On the other hand,
models from survival analysis typically assume that everyone has the same
susceptibility to the disease and will eventually be effected if followed up for
a sufficiently long period of time. It is possible that these models do not
accurately describe the disease risk factors. In models dealing with both
types of expressions, the effect of a covariate can act on either the overall
susceptibility or the age at onset or both. Such a model is considered later in
(3.63). The following model is a special case of the binary frailty model. Here,
one of the two point masses is located at zero, resulting in a cure fraction with
hazard zero with respect to the event of interest.
The application of mixture models for joint modeling of the overall risk of a
disease and the age-at-onset distribution of the individuals at risk is popular
(Farewell 1977, Kuk and Chen 1992, Lam et al. 2005). We define an individual

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
124 Frailty Models in Survival Analysis

to be susceptible if she or he will eventually develop the disease after being

followed for a sufficiently long time. Define

1 : if the individual is susceptible
Y = (3.61)
0 : if not

and let T denote the age at onset when Y = 1. With the foregoing concept,
let φ = P(Y = 1) and S(t) = P(T > t|Y = 1) describe the distribution of Y
and of the failure time T . It is impossible to observe Y , but it is possible to
observe whether or not a subject has experienced the event during the period
of its follow-up time.

• For those individuals that experience the event, it holds that ∆ = 1.

Obviously, if Y = 1, then the survival function for the uncensored ob-
servations is of the form P(Y = 1)P(T ≤ C|Y = 1) = φ(1 − S(C)),
where C denotes the censoring time.

• For the other observations, a failure is not observed (∆ = 0). This may
occur either because Y = 0 or because the observation is really censored.
Therefore, P(Y = 0) + P(Y = 1)P(T > C|Y = 1) = (1 − φ) + φS(C).

Combining these results, the likelihood function takes the form

n
Y
δi
1−δi
L(θ, φ) = φf (ti ; θ) 1 − φ + φS(ti ; θ) . (3.62)
i=1

This is a generalization of (2.6) and (2.7) because these relations are easily
obtained with φ = 1.

Example 3.15
Three different frailty models are applied to age at onset of breast cancer in
11,714 Swedish female twins described in Example 1.6. The results are given
in Table 3.15.

Table 3.15: Frailty and frailty cure models for the breast cancer
data of Swedish twins
parameter gamma gamma cure compound Poisson

σ2 59.060 (7.677) 9.182 (11.14) 45.079 (15.42)

γ 0 -0.089 (0.133)
φ 1.000 0.221 (0.174) 0.237 ( )
λ 2.4e-7 (1.6e-7) 1.8e-6 (2.6e-6) 3.9e-7 (3.1e-7)
ϕ 0.162 (0.013) 0.150 (0.017) 0.150 (0.017)
log-likelihood -5236.5914 -5236.2517 -5236.2751

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 125

In the first column of Table 3.15, the gamma frailty model is applied to account
for heterogeneity in the study population. This model is described in detail
in Section 3.3. The model ignores the existence of a cure fraction, and the
parameter φ = 1 indicates that all individuals are (more or less) susceptible.
This model is extended to the gamma frailty cure model (second column, see
(3.60)), which allows for a fraction of nonsusceptible individuals, for example,
patients who are not at risk of suffering from breast cancer. The third model is
the compound Poisson frailty model (Section 3.8). The size of the susceptible
fraction is calculated by φ = 1 − exp( 1−γ
γσ2 ). That is why no standard error is
given for this quantity. All models are parametric models using a Gompertz
baseline hazard function with parameters λ and ϕ.
In the gamma frailty model, the size of the susceptible fraction is 100%
per definition as in classical survival analysis. The gamma frailty cure model
gives an estimate of 22.1% of the size of the susceptible fraction, for example,
only 22.1% of all women are at risk for breast cancer. The respective number
calculated from the compound Poisson frailty model is similar, with 23.7%.
The gamma frailty model is a special case of the gamma frailty cure model
(when φ = 1) and the compound Poisson frailty model (when γ = 0). These
estimates are in good agreement with results found by Chatterjee and Shih
(2001) and Wienke et al. (2003a) in bivariate analyses and estimates of a
lifetime risk of breast cancer of around 8–12% (Feuer et al. 1993, Rosenthal
and Puck 1999, Ries et al. 1999) in current Western populations. The latter
numbers give a lower boundary for the size of the susceptible fraction. A
huge estimate of σ 2 in the gamma model indicates the existence of large
heterogeneity in the study population, which is at least partially accounted
for by the introduction of a nonsusceptible fraction in the gamma frailty cure
model, where the heterogeneity is smaller, but still large. In the compound
Poisson frailty model, the estimate of γ is negative, indicating the existence of
a nonsusceptible fraction. However, differences between the three models are
not significant in terms of the log-likelihood function. This speaks in favor of
the most simple gamma frailty model. To compare the nested gamma model
(γ = 0) with the compound Poisson model standard test theory can be applied
because the value γ = 0 is no longer on the boundary of the parameter space.
The fit of the compound Poisson frailty model to the Swedish breast cancer
data is demonstrated in Figure 3.11.
It is necessary to mention that the correlation between the ages at onset of
breast cancer in the twin pairs is neglected in these univariate models. We
will reanalyze the data using a bivariate model later on to account for the
correlated observation times.

Model (3.60) was extended to the mixture cure frailty model with gamma
frailty by Peng and Zhang (2008a):

′ 2
S ∗ (t|X1 , X2 ) = (1 − φ(X2 )) + φ(X2 )(1 + σ 2 M0 (t)eβ1 X1 )−1/σ . (3.63)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
126 Frailty Models in Survival Analysis

1.00 Kaplan Meier

compound Poisson Model

0.98
survivor function

0.96

0.94

0.92

0.90
35 45 55 65 75 85 95
age

Figure 3.11: Kaplan–Meier estimator and compound Poisson model for

Swedish breast cancer data.

Here X2 denotes the vector of covariates that may have effects on the cure
rate. The vector X1 represents the covariates that influence the event times
of those individuals who are susceptible to the event. Without frailty, the
model reduces to the proportional hazards model with cure fraction. Without
cure fraction, the model becomes the gamma frailty model considered earlier.
Without both cure fraction and the frailty the common proportional hazards
model is obtained. If the probability of cure depends on observable covariates,
a common way to introduce this into the model is logistic regression (Farewell
1982)
′
eβ2 X2
φ(X2 ) = ′
1 + eβ2 X2
with β2 as the vector of regression parameters to be estimated. Other link
functions such as the probit link and the log–log link can also be used.
Cure models suffer from an inherent identifiability problem due to right-
censored observations. The event under study has not occurred at the end of
the follow-up either because the person is nonsusceptible or because the person
is susceptible, but follow-up was not long enough to observe the event. A key
component of identifiability is the presence of covariates in both components of
the mixture model. In fact, Li et al. (2001) show that, if the baseline hazard
is nonparametric and independent of covariates, and if the cure fraction is

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 127

taken to be constant (independent of covariates), the mixture cure survival

model is not identifiable. The situation becomes less restricted in models with
parametric baseline hazard functions. There exists a series of results regarding
identifiability in competing risk models, which are analogous to identifiability
in mixture cure models (Heckman and Honoré 1989, Abbring and van den
Berg 2003, Peng and Zhang 2008b). All of these results are based on the
presence of observed covariates in the models.
As a consequence of these identifiability results, estimating the proportion
of nonsusceptibles is not easy. However, it is tempting to try it, especially
when the Kaplan–Meier curves appear to level off above zero. For a detailed
discussion of pros and cons of univariate frailty models as cure models, see
Schumacher (1989). The identifiability problem grows with increasing amount
of censoring but is reduced by the parametric modeling of the baseline hazard.
In cure models with fixed censoring times (caused by end of study), censoring
is no longer noninformative. The proportion of censored observations contains
important information about the model parameters; for example, in the case
of no censoring, it holds that φ = 1.
Further interesting examples of cure models based on a frailty approach are
considered in Price and Manatunga (2001). An application of the model to
the analysis of colorectal cancer incidence data in Norway can be found in
Svensson et al. (2006).

3.13 Missing Covariates in PH Models

In most practical situations it is impossible to include all relevant covariates
into the model. Sometimes researchers are not able to measure an important
prognostic factor because of time or financial restrictions. In other situations
there might exist important prognostic factors that are unknown yet. For
example, these could be genetic factors as researcher usually do not know all
genes having a relevant influence on the event time. In other cases, maybe the
information about covariates is lost or incomplete. This consideration is true
for all regression models. The standard approach is to ignore such kind of
variables, which means that the unobserved heterogeneity is absorbed by the
error term. The consequence of omitted covariates in linear regression models
is known: parameter estimates are unbiased unless the omitted risk factors are
correlated with the included observed covariates. In hazard regression models,
even this result does not hold. The omission of influential covariates implies
biased parameter estimates. The main goal of univariate frailty models is
therefore to reduce this bias. We will consider a proportional hazards model
with vectors of observed and unobserved covariates X1 and X2 , respectively:
′ ′
µ(t|X1 , X2 ) = µ0 (t)eβ1 X1 +β2 X2 . (3.64)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
128 Frailty Models in Survival Analysis

Here β2 denotes the vector of regression parameters of unobserved covariates.

Because the vector X2 is unknown, the model cannot be used in practice.
′
That is why it is assumed that the expression Z = eβ2 X2 is a random
unobservable variable, combining the effect of all unobserved covariates in
the frailty Z. For convenience it is assumed that the distribution of Z is
independent of X1 and time t. Both assumptions are reasonable because,
in the classical proportional hazards models with observed covariates, it is
also assumed that there is no interaction between the covariates and that the
covariates are time independent. Now all that matters is the distribution of
frailty, discussed throughout the previous sections. As shown there, for some
frailty distributions, the (unconditional) hazard function can be given in an
explicit form.
Many biostatistical studies dealing with univariate survival have focused on
evaluating the effect of omitted covariates on the estimates of the regression
parameters β, see, for example, Gail et al. (1984), Hougaard et al. (1986a),
Schumacher et al. (1987), Schumacher (1989), Bretagnolle and Huber-Carol
(1988), Chamberlain (1992), Schmoor and Schumacher (1997), Keiding et al.
(1997), and Oakes and Jeong (1998). The present section deals with the
problem of unobserved covariates. Modeling of this unobserved heterogeneity
was the main reason for the introduction of univariate frailty models by Vaupel
et al. (1979) and Lancaster (1979) to the scientific community. The main aim
is to reduce the bias caused by unobserved covariates in the proportional
hazards model.
We will consider the analytical result by Gail et al. (1984) a little bit
more in detail. If the baseline hazard is assumed to be known (parametric
model), if frailty is erroneously ignored in the model, and if there is no
censoring, then the parameter estimates β are consistently estimated by the
classical maximum likelihood method. Unfortunately, this nice result does
not generalize to more realistic settings. In general, hazard models (including
censored observations) neglecting a subset of the important covariates leads
to biased estimates of both regression coefficients and the hazard rate. The
reason for such bias is the time-dependent hazard function, a unique element
in survival analysis. To understand this result, we consider the example of a
population initially consisting of two subpopulations, a high-risk group and
a low-risk group, each subject to a constant risk binary frailty model. If this
heterogeneity is not accounted for and a common hazard is assumed for the
whole population, the estimated hazard will not be a simple average of the
hazards of the two groups as depicted in Figure 3.2. It will increase less
fast over time (or even level off or decrease) compared to the hazards of the
two subpopulations because of the selection effect, which drops out high-risk
individuals early.
Consequently, in a proportional hazards model with observed covariates, an
omitted covariate may result in biased parameter estimates for the remaining
factors. This even holds when the distribution of the omitted characteristics is
initially the same in all categories of included covariates. The problem is that

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Univariate Frailty Models 129

the distribution of characteristics cannot remain the same across categories

over time. This results because high-risk individuals with certain values
of the omitted covariate will die earlier, and so the distribution of omitted
characteristics within each category of an included covariate must also change
over time. As a consequence, all parameter estimates must be contaminated
by the effect of the omitted risk factor. This is quite disturbing since one can
never know the true model.
Bretagnolle and Huber-Carol (1988) predict the sign of the bias in the
two following cases: first, when there is only one covariate left in the model,
whatever the number of omitted covariates is. Then the effect on the survival
of the covariate under study is always underestimated. Second, in case of
several covariates remaining in the analyzed model, the authors prove that
the same result of underestimation holds for each of them at least up to some
fixed time, which, in practical cases, is reasonably long. The asymptotic bias
resulting from such omissions is not negligible, as shown by simulations. In
the case of left-truncated and right-censored data, it may even happen that
the estimates cross zero. That means a beneficial covariate (β < 0) results
through the biased estimate in an unfavorable one (β > 0). Nevertheless, this
crossing zero phenomenon does not happen with only right censoring if one
assumes that censoring is independent of the included covariates.
Univariate frailty models are one (but not the only) way to account for
the effect of omitted covariates in a proportional hazards model. For reasons
of convenience, analysts frequently choose parametric representations of the
frailty distribution that are mathematically tractable. We follow this line
throughout the book. With such parametric assumptions about heterogeneity,
the hazard function can be represented parametrically or nonparametrically.
There could be the problem of misspecifying the frailty distribution. Heckman
and Singer (1982b) demonstrate, however, that for a given parametric baseline
hazard, results can be very sensitive to the choice of the parametric form of
the frailty distribution even when a flexible form is chosen (see also Heckman
and Singer 1984a, and Keiding et al. 1997). For this purpose, Heckman
and Singer (1982b) use simple parametric distribution families (log-normal,
normal, and gamma distribution) for the random effects to investigate this
sensitivity. The authors propose a discrete nonparametric frailty distribution
and show through simulations that the regression parameters can be estimated
with great precision. The frailty distribution is approximated by a discrete
mixture with a finite number of mass points where the number, location, and
probability mass associated with each point are to be estimated (Section 3.2).
The results of Heckman and Singer (1982b) indicate that the distribution of
heterogeneity cannot be approximated well. Thus their procedure corrects for
heterogeneity without clearly identifying its exact distribution.
It should be noted that Heckman and Singer’s (1982b) arguments against
parametric assumptions about the frailty distribution are points of discussion
and controversies by authors who claim that Heckman and Singer (1982b)
misspecified the duration dependence in the example data used to demonstrate

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
130 Frailty Models in Survival Analysis

the potentials of their nonparametric approach. Newer studies conclude that

parameter estimates are not that sensitive with respect to the choice of frailty
distribution (see Manton et al. 1986, Klein et al. 1992, Guo and Rodriguez
1992, Guo 1993).
To get any results at all with a nonparametric frailty distribution, one must
impose a parametric form on the baseline hazard function like the Weibull
or Gompertz distribution. What is not mentioned in Heckman and Singer
(1982b) but is demonstrated, though not emphasized, by Manton et al. (1981)
is that results can also be sensitive to the functional form of the assumed
baseline hazard. Trussell and Richards (1985) show that parameter estimates
could be very sensitive with respect to the assumed parametric form of the
baseline hazard function when using the nonparametric frailty distribution
suggested by Heckman and Singer (1982b).
Henderson and Oman (1999) intend to quantify the bias that may occur
in estimated covariate effects. They fit marginal distributions when frailty is
present in the survival data but ignored in a wrongly specified proportional
hazards analysis. The authors investigate this problem for positive stable
frailty models. In general, however, no explicit solution was found. Henderson
and Oman (1999) used first-order approximations to estimate the bias in
different situations. It turns out that the amount of bias depends on the
form of frailty distribution, but to the first-order approximation the bias is
independent of the covariate distribution. Censoring in the Koziol–Green
model (Koziol and Green 1976), where the censoring survival is a power of
the baseline survival function, and type I censoring show that the asymptotic
bias is reduced by censoring, particularly with log-normal frailty. For the
very specific situation of noncensored survival times and gamma-distributed
frailty with variance parameter σ 2 , the following approximation was derived
by Henderson and Oman (1999)

β = (1 + σ 2 )β ∗ , (3.65)
∗
where β, β denote the unbiased and biased parameters, respectively, in the
Cox model, assuming that the covariates are centered. Consequently, the
bias due to omitted covariates becomes greater with increasing heterogeneity
(frailty variance), and relation (3.65) could be used to compensate for the
frailty effect.
Congdon (1995) investigates the influence of different frailty distributions
(gamma, inverse Gaussian, stable, and binary) on the analysis of cause-specific
and total mortality data from the London area during the years 1988 – 1990
by using Weibull and Gompertz baseline hazard functions.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com

Chapter 4
Shared Frailty Models

So far we have focused on the frailty model as a way of dealing with possible
heterogeneity due to unobserved covariates. This is the main interpretation
of frailty in the application to univariate time-to-event data. As discussed in
the last chapter, this results in selection over time; for example, this is shown
as leveling-off or crossing-over effects in population hazards. The concept of
frailty introduced by Vaupel et al. (1979) to biostatistics and by Lancaster
(1979) to the econometric literature originates from this kind of models.
Another, completely different, aspect of this approach is to use the frailty
term to model associations between event times, which goes back to the work
of Clayton (1978). Implicitly, most of the statistical models and methods for
failure time data, and here especially the Cox proportional hazards model,
were developed under the assumption that the observations from different
subjects are statistically independent of each other. While this is sensible in
many applications, it has become obvious that this assumption does not hold
in certain situations that are more common as originally thought.
In the following section the two main approaches in multivariate survival
analysis – marginal and frailty models – are outlined and compared, followed
by a description and discussion of the shared frailty model in the subsequent
sections. The shared frailty model is a mixture model because the common
risk in each cluster (the frailty Z) is assumed to be random. The model
assumes that all event times in a cluster are independent given the frailty
variables. In other words, it is a conditional independence model where the
frailty is common to all individuals in a cluster and therefore responsible for
creating dependence between event times. This is the reason for the concept
of shared frailty. A shared frailty model can be considered as a mixed (random
effects) model in survival analysis with group variation (frailty) and individual
variation described by the hazard function. In contrast, mixed models show
a more symmetric handling of these two sources of variation. Because of the
censored observations the Cox model and the frailty models do not belong
to the class of generalized linear mixed models. It is assumed that there is
independence between the observations from different clusters. If the variation
of the frailty variable is zero, this implies independence between event times
in the clusters; otherwise, there is positive dependence between event times.
A more detailed presentation of shared frailty models can be found in the
excellent book by Duchateau and Janssen (2008).

131
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
132 Frailty Models in Survival Analysis

4.1 Marginal versus Frailty Model

Multivariate survival analysis may provide an effective tool for analyzing
information from multiple/recurrent events in situations where independence
between event times cannot be assumed. The following three examples given
in Liang et al. (1995) may serve to illustrate different problems in multivariate
event-time data analysis.

• Diabetic retinopathy is one of the most important causes of visual loss

and blindness. Given the relatively high prevalence of this disease, new
treatments to delay the onset of severe visual loss are critical. In 1971,
the Diabetic Retinopathy Study was initiated to study the effectiveness
of laser photocoagulation. This randomized, controlled clinical trial
involved more than 1700 patients enrolled at 15 medical centers in the
United States. Patients with diabetic retinopathy and visual acuity
of 20/100 or higher in both eyes were eligible to participate in the
study. The design protocol randomly selected one of each participant’s
eyes that was treated, and the other eye was observed without having
received treatment. The event was the occurrence of visual acuity less
than 5/200 at two consecutively completed follow-ups of four months.
This design differs from conventional trials in that each patient served as
his own control. Consequently, each patient contributed two correlated
observations to the analysis, one from each eye.

• Family studies are critical in assessing the role which genetics play in the
disease process. Statistical methods such as variance component models
and path analysis have been developed and adopted to analyze family
data with quantitative traits such as cholesterol. Variance component
models aim to measure the extent to which the total variation in the
quantitative trait is due to a correlation between relatives, as well as the
degree of correlation among full siblings and other relatives. When the
considered trait is onset age of a disease, conventional methods are not
appropriate. This is mainly due to censoring/truncation of this variable
but also to the need for a measure of within-family correlation that
incorporates time, a feature not shared by conventional measures such
as the correlation coefficient.

• The trial on gamma interferon as described in Fleming and Harrington

(1991) is a typical example of recurrent events. In their study, patients
with chronic granulomatous disease (CGD) were randomly assigned to
either gamma interferon or placebo, with the event of interest being the
diagnosis of a serious infection. It is clear that infections can reoccur.
Thus both the time and the number of infection occurrences may be
informative about the effectiveness of the treatment.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 133

These three examples share an important feature, namely, that the failure
times for observations from the same cluster correlate with one another. In
Examples one and three the cluster is an individual, in Example two it is a
family. In these examples, the cluster sizes are small relative to the number
of clusters. These three examples, however, differ from one another in terms
of their scientific objectives. The main objective of examples one and three
is to examine the effectiveness of a new treatment, which presumably can be
characterized through regression modeling. The within-cluster correlation in
these examples is usually of secondary interest, although ignoring it could lead
to erroneous conclusions. The within-family association, for example, two is
of primary interest, although regression adjustment for each related subject
is critical in order to minimize the potential that the observed association is
mainly due to environmental factors shared by family members. Furthermore,
the mechanisms behind within-cluster associations may vary so that different
statistical models to describe the associations may be needed. It is clear that
the mechanism leading to the correlation between two fellow eyes is different
from that attributed to the correlation of observations measured over time
from the same eye.
Another typical example with clustered event times are multicenter clinical
trials with event-time outcome. Here the treatment centers are the clusters.
Event times of patients in a cluster are assumed to be correlated because of
center-specific differences in treatment, care, and diagnosis of the patients, or
in the composition of the center-specific study populations. The lung cancer
data from the Halluca study provide such an example. Dependent on the
prevalence of the disorder of interest, there are sometimes many centers with
only a few patients (rare diseases) or a few centers with a large number of
patients each (common diseases).
Available statistical models fall into two broad classes – marginal and frailty
models (Wei and Glidden 1997). Marginal methods specify models for the
effect of covariates on the hazards of the individuals (the margins) under the
working independent assumption. That means, for the point estimates of the
parameters, independence between the event times in a cluster is assumed.
To take into account the fact that observed event times are correlated, the
variance estimates need to be adjusted but without the need for explicitly
modeling the correlation (Wei et al. 1989, Lee et al. 1992, Cai and Prentice
1995). The association between the events is considered a nuisance parameter.
The marginal baseline hazards can be modeled differently (Wei et al. 1989)
or with a common functional form (Lee et al. 1992). As with the analysis
of longitudinal data, regression parameters are estimated from generalized
estimating equations, and the corresponding variance–covariance estimators
are corrected properly to account for the dependence structure. Yu and Peng
(2008) suggested a marginal model with cure fraction where both the margins
as well as the cure fraction may depend on observed covariates. An excellent
and detailed review of the robust and well-developed marginal approach is
given in Lin (1994).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
134 Frailty Models in Survival Analysis

The other commonly used and general approach to the problem of modeling
multivariate data is the specification of independence among observed data
items conditional on a set of unobserved or latent variables (random effects).
A multivariate model for the observed data is then induced by averaging over
an assumed distribution for the latent variables. The dependence structure
in the multivariate model arises when dependent latent variables enter into
the conditional models for multiple observed data items, and the dependence
parameters may often be interpreted as variance components. Frailty models
for multivariate survival data are derived under a conditional independence
assumption by specifying latent variables that act multiplicatively on the
baseline hazard function. Different assumptions about the distribution of
the random effects can create different dependence structures in the cluster
as discussed in detail for the shared frailty model in the book by Duchateau
and Janssen (2008). This popular approach with its strong link to the field of
generalized linear mixed models is one of the building blocks for the present
monograph.
Two main differences between the two approaches should be kept in mind.
First, the interpretation of regression parameters in both models is different.
In the marginal model, the parameters describe the population level relative
risk, whereas in the frailty model, the parameters have a cluster-level relative
risk interpretation. That is, in the latter approach, the hazard ratio refers to
comparisons within clusters where individuals share the same frailty. In the
marginal approach, the hazard ratio refers to comparisons between individuals
randomly drawn from the total study population, independent of which cluster
the individuals belong to. As such, the estimates are not expected to be the
same in marginal and frailty models since they estimate different quantities,
unless the within-cluster correlation is zero (meaning that clustering does not
play any role) and both models comply with the Cox model for independent
data.
Second, the marginal approach, although effective, can predict only marginal
survival probabilities of single individuals. In contrast, in the frailty approach
it is possible to perform joint prediction of survival for individuals from the
same cluster. Frailty models also allow the prediction of survival based on the
current status of the other individuals in the cluster. Such kind of analysis
is of special interest in small clusters, for example, in event times of sibships.
Practical comparisons of both methods show that the confidence intervals for
regression parameter estimates are smaller in frailty models compared to that
in the marginal approach (Chuang and Cai 2006). This is based on the fact
that the frailty approach makes stronger model assumptions by specifying the
correlation structure.
An interesting approach occupying an intermediate position between the
two aforementioned models to allow estimation both regression coefficients
with traditional interpretation as well as correlations between event times is
described by Mahé and Chevret (1999).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 135

4.2 The Concept of Shared Frailty

In this section we would like to focus on multivariate event-time models
with dependent hazards as described earlier. This concept provides multi-
variate extensions of the traditional univariate frailty model (Vaupel et al.
1979, Lancaster 1979), and it allows mutual dependence of clustered event
times to be taken into account in the analysis of event-time data. Survival
models for dependent event times are especially useful because they allow
more sophisticated questions about the nature of aging, disease, disability,
and the mortality processes to be addressed. Such dependence occurs, for
example, in event times of related individuals (e.g., family members) or in
duration times of recurrent events (e.g., asthma attacks).
One important and mostly used approach in this field is the shared frailty
concept. The hazard model for each individual in this approach, however,
looks exactly the same as in the standard univariate frailty model considered in
the last chapter. The only, but important, difference is that, in a shared frailty
model, frailty is defined as a measure of the relative risk that individuals in a
group share. Thus the frailty variable is associated with groups of individuals
rather than individuals as such.
The shared frailty approach assumes that all failure times in a cluster are
conditionally independent given the frailties. The value of the frailty term is
constant over time and common to all individuals in the cluster, and thus it
is responsible for creating dependence between event times in a cluster. This
dependence is always positive in shared frailty models. Originally, the model
was introduced to the literature by David Clayton and was considered in the
bivariate case without using the notion of frailty (Clayton 1978), modeling
the event times of sons and their fathers. The shared frailty model dominates
the literature on multivariate frailty models and was extensively studied in
the monographs by Hougaard (2000), Therneau and Grambsch (2000), and
Duchateau and Janssen (2008). That is the reason why we will treat this
model only in brief to develop the main ideas of multivariate frailty models,
but without going into much details here. This lays the basis for extensions
of the shared frailty model in different directions. These extensions are one of
the main tasks of the present monograph and will be considered in the next
chapter.
In the special case of degenerated frailty Z (meaning the variable Z is
a nonrandom constant), no dependence exists between the lifetimes within
a group. In this situation, the shared frailty model reduces to the case of
independent event times without unobserved heterogeneity, and that means
the proportional hazards model. The univariate frailty model considered in
the last chapter is a special case of the shared frailty model with cluster size
one. In this case the interpretation of the model is different from that of
the multivariate frailty model. In the univariate case, modeling of unobserved

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
136 Frailty Models in Survival Analysis

heterogeneity caused by unobserved covariates is the main goal of the analysis.

In the multivariate approach, modeling of or adjusting for the correlation in
the regression models is the main focus. This conceptional difference is the
main reason for treating the univariate and the shared frailty model in separate
chapters in the present monograph.
Event times from different clusters are considered to be independent. The
number of observations in a cluster is assumed to be known. Twin studies
provide typical examples of bivariate event data. Another example are the
times to failure for several similar human organs, like time to blindness of the
right and the left eye as in studies on diabetic retinopathy. In family studies
one is often faced with small and unequal cluster sizes. Event time data from
multi-center clinical trials are an example for moderate to large and unequal
cluster sizes.
In all cases, a probability distribution is assigned to the frailty variable for
practical reasons. Some results exist, especially in the econometric literature,
on mixed proportional hazards (frailty) models deriving specific conditions
that make models identifiable even in case of a nonspecified frailty distribution.
For a comprehensive overview in this field, see van den Berg (2001). We
will not consider this situation here. Frailty distributions were discussed
in detail in Chapter 3. Assuming a random frailty means that we have to
integrate the frailty out of the likelihood expression or use more sophisticated
estimation approaches to obtain parameter estimates. In the first case, almost
all calculations can be made based on the Laplace transform of the respective
frailty distribution, and in the second case, MCMC or numerical integration
techniques are applied. In the examples in the following sections we will
(similar to the last chapter) focus on parameter estimation strategies that are
included in easily accessible statistical packages (R, S plus, SAS, STATA) to
allow also nonspecialists in this field to apply the discussed models without
extensive programming. Despite the fact described earlier that univariate
and multivariate frailty models differ widely, software routines analyzing data
wit frailty models can usually deal with both types, the univariate and the
shared frailty models. The routine only needs different cluster variables in the
data sets, indicating individuals (cluster size one) or groups of individuals as
cluster.
In the following paragraph the basic definition of the model is given. A
shared frailty model in survival analysis is defined as follows. Suppose there
are n clusters and that cluster i has ni observations and associates with the
unobserved frailty Zi (1 ≤ i ≤ n). The vector Xij (1 ≤ i ≤ n, 1 ≤ j ≤ ni )
contains the covariate information of the event time Tij of the j th observation
in the ith cluster. Conditional on the frailty term Zi , the survival times in
cluster i (1 ≤ i ≤ n) are assumed to be independent and their hazard functions
to be of the form

′
µ(t|Xij , Zi ) = Zi µ0 (t)eβ Xij
,

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 137

where µ0 (t) denotes the baseline hazard function, and β is a vector of fixed-
effect parameters to be estimated. The frailties Zi (i = 1, . . . , n) are assumed
to be independently and identically distributed random variables with density
function f (z). The frailty density depends on unknown parameters to be
estimated. Similar to the univariate case a semiparametric shared frailty
model is one with a nonparametric baseline hazard µ0 (t). Furthermore,
please note that, with the foregoing notation, the case of event times from
clustered individuals as well as the case of recurrent event times in individuals
is covered. However, in the analysis of recurrent event-time data, there may be
specific information about the ordering of the events, and the time scale used
needs careful consideration. For specific aspects with shared frailty models in
recurrent event times, see Chapter 9 of Hougaard (2000) and Duchateau et
al. (2003).
The main assumption of a shared frailty model is that all individuals in
cluster i share the same value of frailty Zi (i = 1, . . . , n), and this is why
the model is called the shared frailty model. The lifetimes are assumed to be
conditionally independent with respect to the shared (common) frailty. This
shared frailty is the cause of dependence between lifetimes within the clusters.
We derive in the following paragraphs the quantities based on this conditional
formulation. Independence of the lifetimes within the clusters corresponds to
a degenerate frailty distribution (no variability in Zi ). In all other cases, the
dependence is positive. It is assumed that there is independence between event
times from different clusters. If condition P(Zi > 0) = 1 holds, the shared
frailty model leads to absolute continuous distributions and thus cannot model
dependence due to common events. Consequently, it is not appropriate for
event-related dependence (shock models) because an event in one individual
is not relevant to the partner; it only changes the information available on
the frailty. One exception to this general assumption is the shared compound
Poisson frailty model, which does not fulfill the condition P(Zi > 0) = 1. In
this specific model, it is possible that all individuals in a cluster share the
frailty value zero (with positive probability) and are nonsusceptible to the
event of interest (cure model).
Using an argument similar to that in equation (3.5), we can derive the
joint conditional multivariate survival function for the individuals in the ith
cluster. Conditional on frailty Zi which is shared by all individuals in cluster
i, it holds that

S(ti1 , . . . , tini |Xi , Zi ) = S(ti1 |Xi1 , Zi ) . . . S(tini |Xini , Zi )

Xni
′
M0 (tij )eβ Xij


= exp − Zi (4.1)
j=1

Rt
where M0 (t) = 0 µ0 (s) ds denotes the cumulative baseline hazard function
and Xi = (Xi1 , . . . , Xini ) is the covariate matrix of the individuals in the ith

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
138 Frailty Models in Survival Analysis

cluster. This is the starting point to derive the unconditional joint survival
function. Averaging expression (4.1) with respect to the frailty Zi gives the
marginal survival function:

S(ti1 , . . . , tini |Xi ) = ES(ti1 , . . . , tini |Xi , Zi )

Xni
′
M0 (tij )eβ Xij


= E exp − Zi
j=1
ni
X ′
M0 (tij )eβ Xij ,


=L
j=1

where L denotes the Laplace transform of the frailty variable. Thus, the
multivariate survival function is expressed as the Laplace transform of the
frailty distribution, evaluated at the cumulative baseline hazard. The joint
survival function for all event-time data is now the product of the survival
functions of all the clusters because of the assumption about independence
between clusters

n
Y ni
X ′
M0 (tij )eβ Xij .


S(t11 , . . . , tnnn |X1 , . . . , Xn ) = L
i=1 j=1

The univariate unconditional survival functions can be expressed by means of

the Laplace transform

S(tij |Xij ) = ES(tij |Xij , Zi )

′
= E exp − Zi M0 (tij )eβ Xij


′
= L M0 (tij )eβ Xij .

Denote by L−1 the inverse function of the Laplace transform L. Consequently,

M0 (tij )eβ Xij = L−1 (S(tij |Xij )), and the unconditional survival function of
′

the ith cluster is given by

S(ti1 , . . . , tini |Xi ) = L L−1 S(ti1 |Xi1 ) + . . . + L−1 S(tini |Xini ) ,

the Archimedian copula family of Genest and MacKay (1986). There is a

strong link between frailty models and copula models. For an illustrative
discussion in this field we refer to Goethals et al. (2008), Duchateau and
Janssen (2008), and Chapter 6 of this monograph.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 139

4.3 Shared Gamma Frailty Model

Similar to the univariate case in the last chapter, the standard assumption
about frailty in shared frailty models is that it follows a gamma distribution.
The main reason for the popularity of the gamma distribution is their nice
mathematical properties, especially the simple form of the Laplace transform.
In the shared frailty model, another aspect has to be considered additionally
with respect to frailty distribution. Each frailty distribution implies a specific
form of dependence between event times in clusters. For example, the gamma
distribution models late dependence in shared frailty models (Duchateau and
Janssen 2008). Assuming for frailty a gamma distribution with expectation
one and variance σ 2 , averaging equation (4.1) with respect to the cluster-
specific random variable Zi produces the multivariate survival function for
the ith cluster

ni
X 
′
S(ti1 , . . . , tini |Xi ) = L M0 (tij )eβ Xij

j=1
 ni − 1
σ2
X ′
= 1 + σ2 M0 (tij )eβ Xij
j=1
ni
X − 12
2
S(tij |Xij )−σ − (ni − 1)
σ
= , (4.2)
j=1

where the last relation is a consequence of equation (3.17). It should be noted

that the concept of shared frailty is different from that of individual frailty
introduced by Vaupel et al. (1979) and Lancaster (1979) in their analysis
of univariate duration data. This difference has gone largely unrecognized,
perhaps because of the superficial similarity of the individual hazards in the
two approaches. Frailty in the shared frailty model is only a part of individual
frailty, capturing only the components of frailty shared by all individuals
in a cluster. Noncommon unobserved risk factors are not included in the
shared frailty model and require extensions of the shared gamma frailty model
considered in the next chapter.
In some situations we will pay special attention to the case of bivariate
frailty models (ni = 2, i = 1, . . . , n). Examples of bivariate data, based on
event times from Danish and Swedish twins, are analyzed and discussed in
detail in the present and the following chapters. In the bivariate case, the
survival function (4.2) simplifies to

2 2
− 1
S(t1 , t2 ) = S1 (t1 )−σ + S2 (t2 )−σ − 1 σ2
, (4.3)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
140 Frailty Models in Survival Analysis

where we drop the cluster index i in the following for ease of presentation
but assume a single binary covariate X representing two strata under the
assumption of strata-specific baseline hazard functions. This results in two
different marginal survival functions for the partners in each pair. In this
situation it holds that S(t|X) = S1 (t) if X = 1 (strata 1) and S(t|X) = S2 (t)
if X = 2 (strata 2).
The correlations between lifetimes of randomly selected pairs are always
the same in the shared frailty model, implying a symmetric situation. This
symmetry (which is equivalent to the compound symmetry assumption in
linear mixed models) makes the model less useful for modeling correlations
in family studies with groups of different relatives (mother–father, mother–
daughter, grandfather–son, brother–brother, etc.), which are, for example, of
special interest in genetic studies. In the following we will distinguish between
the parametric and the semiparametric shared gamma frailty models.

4.3.1 Parametric shared gamma frailty model

Different parametric forms can be assumed for the baseline hazard function.
In most examples, a Gompertz or Weibull baseline hazard is assumed, but
the theory is developed in general. In the parametric shared gamma frailty
model, the unconditional likelihood function can be easily derived. Similar
to the univariate case, the regression parameters β, the parameter vector θ
of the baseline hazard function µ0 , and the variance of the frailty σ 2 are the
parameters to be estimated. Consider first the conditional likelihood in the
case of n clusters of size ni (i = 1, . . . , n):

n Z ni
∞Y
Y ′
δij β ′ Xij
L(β, θ, σ 2 ) = zi µ0 (tij ; θ)eβ Xij e−zi M0 (tij ;θ)e f (zi ; σ 2 )dzi
i=1 0 j=1

with
zi
1/σ2 −1
zi2 e− σ 2
f (zi ; σ ) = 2/σ
σ 2 Γ(1/σ 2 )
denoting the density function of a gamma distribution with expectation one
and variance σ 2 . A closed form solution to this integral exists. Using the
simplification
ni
X ′
yi = 1/σ 2 + M0 (tij ; θ)eβ Xij
j=1
the foregoing likelihood can be written in the form

β ′ Xij δij
n Qni

∞
j=1 µ0 (tij ; θ)e
Z
Y 2
2
L(β, θ, σ ) = 1/σ2 +di 2/σ2
(yi zi )1/σ +di −1 −yi zi
e yi d(zi ),
i=1 yi σ Γ(1/σ 2 ) 0

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 141
Pni
where the expression di = j=1 δij denotes the number of observed events in
cluster i (i = 1, . . . , n). Now the integral can be solved giving

n Qni ′
δij
2
Y Γ(1/σ 2 + di ) j=1 µ0 (tij ; θ)eβ Xij
L(β, θ, σ ) =
1/σ2 +di .
1/σ 2 + nj=1 σ 2/σ2 Γ(1/σ 2 )
P i ′
i=1 M0 (tij ; θ)eβ Xij

By taking the natural logarithm of this expression we obtain the well-known

unconditional log-likelihood function of the shared gamma frailty model (Klein
1992, Duchateau and Janssen 2008):

n h
X
log L(β, θ, σ 2 ) = di log σ 2 + log Γ(1/σ 2 + di ) − log Γ(1/σ 2 )
i=1
ni
X ′
2 2
− (1/σ + di ) log(1 + σ M0 (tij ; θ)eβ Xij
)
j=1
ni
X i
+ δij (β′ Xij + log µ0 (tij ; θ)) . (4.4)
j=1

In the parametric shared gamma frailty model with observed covariates, the
unobserved frailty Zi (i = 1, . . . , n) in each cluster can be estimated (Nielsen
et al. 1992) by the expression

1/σ̂ 2 + nj=1
P i
δij
Ẑi = n ′ . (4.5)
1/σ̂ 2 + j=1 M0 (tij ; θ̂)eβ̂ Xij
P i

Here σ̂ 2 is the estimate of the frailty variance, θ̂ the vector of the estimated
parameters of the cumulative baseline hazard function, and β̂ the vector of
estimated regression coefficients. This is a natural extension of formula (3.26)
in the univariate case.

Example 4.1
We consider the Halluca study from Example 1.3 and apply three different
parametric shared gamma frailty models with exponential, Gompertz, and
Weibull baseline hazards (Table 4.1). It turns out that the exponential hazard
function is not flexible enough, the Gompertz and Weibull model show a
significant better fit to the data with respect to the likelihood ratio test (the
exponential model is nested in the Gompertz as well as the Weibull model).
The Weibull model shows a slightly better fit compared to the Gompertz
model, but parameter estimates and standard errors are similar in both models
and also closed to the values obtained in the semiparametric shared gamma
frailty model (Table 4.2) considered in the next section.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
142 Frailty Models in Survival Analysis
Table 4.1: Parametric shared gamma frailty models of the
Halluca lung cancer data
parameter Weibull exponential Gompertz

age (in years) 0.009 (0.003) 0.009 (0.003) 0.009 (0.003)

sex (females) -0.162 (0.073) -0.172 (0.073) -0.160 (0.073)
type (NSCLC) -0.107 (0.068) -0.119 (0.068) -0.098 (0.068)
ECOG (3 & 4) 0.595 (0.108) 0.627 (0.108) 0.610 (0.108)
stage II 0.450 (0.170) 0.476 (0.170) 0.449 (0.170)
stage IIIa 0.613 (0.133) 0.651 (0.133) 0.606 (0.133)
stage IIIb 0.978 (0.123) 1.043 (0.123) 0.961 (0.123)
stage IV 1.324 (0.113) 1.413 (0.112) 1.312 (0.113)
λ 0.033 (0.009) 0.023 (0.006) 0.029 (0.008)
ν 0.878 (0.020)
ϕ -0.018 (0.003)
σ2 0.318 (0.116) 0.403 (0.134) 0.305 (0.113)
log-likelihood -4891.1150 -4908.4594 -4891.8328

In all three models we see an increase in risk with higher age, ECOG status 3 or
4 (compared to ECOG status 0 – 2), and higher stage of the disease (reference
group is stage I). Females and patients with a non-small cell lung carcinoma
experience a lower risk. The estimates of the regression and frailty parameters
depend on the parametric assumption regarding the baseline hazard. This
limits the applicability of the parametric shared gamma frailty model because
more detailed investigations about the shape of the baseline hazard function
are necessary. In most practical cases, there is no additional information
about the form of the baseline hazard function, which makes the parametric
approach questionable. Analysis is performed using SAS PROC NLMIXED
based on the suggestion by Liu and Yu (2008).

An interesting application of the shared gamma frailty model with piecewise

constant baseline hazard function to the lifetimes of Danish adoptees and their
biological and adoption parents is presented in Petersen et al. (2010).

4.3.2 Semiparametric shared gamma frailty model

In the last section the parametric shared gamma frailty model was discussed
in detail. Here we consider the semiparametric shared gamma frailty model.
This model is a standard tool to analyze multivariate survival data. As already
discussed, in many practical situations there is little or no information about
the form of the baseline hazard function. Consequently, if the sample size is
large enough, a semiparametric approach is often preferred because it does
not rely on assumptions that are difficult to verify. The log-likelihood in the
semiparametric shared gamma frailty model looks similar to (4.4), but the
vector of model parameters to be estimated contains only the regression and

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 143

frailty parameters, and the baseline hazard function is treated as an infinite

dimensional nuisance parameter. For such a model, direct maximization of the
marginal likelihood is no longer possible, and more sophisticated estimation
procedures are necessary.
Semiparametric proportional hazards models without frailty are analyzed
by means of the partial likelihood method (Cox 1972, 1975). To adapt
this approach to proportional hazard models with frailty, the EM algorithm
(Expectation-Maximization algorithm) can be used.
The algorithm was suggested by Dempster et al. (1977) and is typically
used in the presence of unobserved data. The EM algorithm iterates between
two steps. In the first one, an estimate (the expectation) of the unobserved
frailties based on observed data and current parameter estimates is obtained.
These estimates are used in the maximization step to obtain new parameter
estimates given the estimated frailties. In the gamma frailty model, closed-
form expressions exists for the conditional expectations of the frailties in the
E step, which is very convenient for the estimation procedure. Furthermore,
applying the partial likelihood concept with the estimated frailties as off-
set terms in the M step is easy to perform, which makes the EM algorithm
useful. The EM algorithm is described in more detail for the univariate gamma
frailty model in Section 3.3.2, and for the shared gamma frailty in the book
by Duchateau and Janssen (2008). Therefore we omit further details of the
algorithm here.
Similar to the parametric shared gamma frailty model, the unobserved
frailty Zi , (i = 1, . . . , n) in each cluster can be estimated (Nielsen et al. 1992)
by the expression
Pni
1/σ̂ 2 + j=1 δij
Ẑi = Pni ′ . (4.6)
2
1/σ̂ + j=1 M̂0 (tij )eβ̂ Xij

Here the difference from relation (4.5) lies in the expression M̂0 , which is now
a nonparametric estimator of the cumulative baseline hazard (for example,
the Nelson–Aalen estimator). This approach was used by Carvalho et al.
(2003) to investigate the quality of dialysis centers in Brazil represented
by the unobserved frailty values assigned to each center. This underlines
the possibility of using frailty models to rank treatment centers or other
institutions with respect to their estimated frailties.

Example 4.2
We consider the Halluca data again and apply the semiparametric shared
gamma frailty model. Patients in the study are clustered by the center where
the lung carcinoma was diagnosed. Altogether, there were 53 diagnosing
centers representing the clusters. A univariate and a shared gamma frailty
model is applied to the data in Table 4.2, where in the univariate approach
the clustering of the data is not taken into account.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
144 Frailty Models in Survival Analysis
Table 4.2: Univariate and shared gamma frailty
model of the Halluca lung cancer data
parameter univariate frailty shared frailty

age (in years) 0.013 (0.004) 0.009 (0.003)

sex (females) -0.197 (0.091) -0.158 (0.073)
type (NSCLC) -0.096 (0.086) -0.097 (0.068)
ECOG (3 & 4) 1.054 (0.145) 0.600 (0.108)
stage II 0.578 (0.203) 0.450 (0.170)
stage IIIa 0.780 (0.157) 0.600 (0.133)
stage IIIb 1.232 (0.146) 0.957 (0.123)
stage IV 1.746 (0.133) 1.314 (0.113)
σ2 0.460 ( ) 0.311 ( )

In general, the effect of the covariates is weakened (an exception from this
general rule is the covariate type) when applying the shared gamma frailty
model instead of the univariate gamma frailty model. Please note that the
interpretation of the parameter estimates is different in both models. In the
univariate model the estimates of the regression parameters are adjusted for
the unobserved heterogeneity in the population and only conditional given
the same frailty. Frailty variance is interpreted as a measure of unobserved
heterogeneity in the study population. In the shared gamma frailty model
the parameters are adjusted for the correlation in the clusters, and the frailty
variance is interpreted as a measure of the correlation between the lifetimes in
the clusters. Again, the parameter estimates are conditional given the same
frailty (or, which is equivalent in the shared frailty model, conditional on
being from the same cluster). It is important to note these subtle differences
between the models. Unfortunately, in the literature about frailty models,
sometimes this difference is overlooked and a clear distinction between these
two classes of models is missed, leading to wrong interpretations of the results.
In both models, estimates are obtained by means of the COXPH function
of the statistical package R, which does not provide estimates for the standard
error of frailty variance.

The asymptotic properties of the maximum likelihood estimates in the non-

parametric shared gamma frailty model are well established. Murphy shows
consistency (Murphy 1994) and asymptotic normality (Murphy 1995) in the
model without observed covariates. These results were generalized to the
correlated gamma frailty model (which includes the shared gamma frailty
as a special case and will be considered in detail in the next chapter) with
covariates by Parner (1998). Gorfine et al. (2006) suggest new estimates
and provide large sample theory for the general semiparametric shared frailty
model with covariates. Their method needs only a finite first moment of the
frailty distribution, but also makes use of the specific frailty distribution.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 145

Regarding goodness-of-fit, Shih and Louis (1995b) introduced a graphical

method for assessing the assumption about the gamma distribution of the
frailty when the baseline hazard function is parametric and in the absence
of covariates. This work was continued by Shih (1998), introducing a simple
test statistic for a goodness-of-fit test in the shared gamma frailty model,
allowing for censored observations, but in the absence of observed covariates.
Fine and Jiang (2000) suggest a goodness-of-fit test that allows for observed
covariates, which are included via semiparametric accelerated failure time
regression models. Glidden (1999) suggests a test for the semiparametric
gamma frailty model without covariates. A graphical as well as numerical
method for checking the adequacy of the gamma distribution in a shared
frailty model is suggested by Cui and Sun (2004). Their test is based on
the posterior expectation of frailties given the observable data over time,
extending the work by Glidden (1999). Viswanathan and Manatunga (2001)
use kernel regression smoothing in order to distinguish between gamma and
positive stable frailty distribution. Papers by Chen and Bandeen-Roche (2005),
Glidden (2007), and Economou and Caroni (2008) contain additional material
about graphical procedures checking the dependence structure in shared frailty
models. These are important results that can be used for the decision about
the frailty distribution to be applied to the data in specific situations.
The choice of the gamma distribution as a frailty distribution is supported
in addition by asymptotic results obtained by Abbring and van den Berg
(2007). Their main results focus on univariate models but are also proven in
a specific bivariate model.

4.3.3 Shared gamma frailty model for current status data

In this section the parametric shared gamma frailty model will be applied
to current status data. The parametric univariate gamma frailty model for
current status data was already considered in Section 3.3.3. In the bivariate
case with unconditional survival function S(t1 , t2 ; θ), the likelihood function
contains four components, depending on the censoring status of the observed
data:

Li (θ) = δ1 δ2 1 − S1 (t1 ; θ) − S2 (t2 ; θ) + S(t1 , t2 ; θ)


+ δ1 (1 − δ2 ) S2 (t2 ; θ) − S(t1 , t2 ; θ)


+ (1 − δ1 )δ2 S1 (t1 ; θ) − S(t1 , t2 ; θ)
+ (1 − δ1 )(1 − δ2 )S(t1 , t2 ; θ)

with S1 (t1 ; θ) = S(t1 , 0; θ) and S2 (t2 ; θ) = S(0, t2 ; θ) as marginal survival

functions. One interesting research problem deals with the question how much
information is lost when current status data are observed instead of (right-
censored) lifetimes in the usual way. To answer this question, simulations are

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
146 Frailty Models in Survival Analysis

performed based on the bivariate shared gamma frailty model under different
censoring assumptions.
Data are generated following the bivariate shared gamma frailty model with
frailty variance V(Z) = σ 2 = 1, expectation EZ = 1, and Gompertz baseline.
One thousand samples of sample size 1000 are generated. We consider two
scenarios as follows. In the first scenario, the censoring/monitoring times
are assumed to be independent. Consequently, current status data contain
information about the two-dimensional distribution of the event times. In the
second scenario, both censoring/monitoring times are assumed to be equal.
This is a typical situation in current status data, where blood samples are
used to test for the presence of different infections.

Scenario 1: Both censoring/monitoring times are independent.

Table 4.3: Simulation (Scenario 1) of the shared

gamma frailty model with current status data
parameter true current status right censored

σ2 1.000 1.003 (0.076) 1.002 (0.054)

λ1 0.100 0.101 (0.015) 0.100 (0.007)
ϕ1 0.120 0.121 (0.026) 0.121 (0.015)
λ2 0.100 0.100 (0.014) 0.100 (0.008)
ϕ2 0.120 0.121 (0.025) 0.121 (0.015)

Each data set is analyzed under two different censoring mechanisms. In

the case of right censoring (last column in Table 4.3), (T1 , T2 , ∆1 , ∆2 ) are
assumed to represent common right-censored lifetimes. The second analysis
(third column in Table 4.3) assumes that the data represent current status
data, which means monitoring times are observed instead of event times.
The estimates based on right-censored data are very similar to those based
on current status data (and very close to the true values). The standard
errors are larger than in the case of the current status data, but in general
the information loss when switching from right-censored event time data to
current status data is moderate in this scenario.

Scenario 2: Both censoring/monitoring times are equal.

Again, each data set is analyzed under two different censoring mechanisms,
once in the case of right-censored lifetimes and once data are assumed to
represent current status data. The parameter estimates based on the right-
censored data are very similar to the estimates based on current status data
(and very close to the true parameter values). The standard errors are
larger with the current status data, but in general the information loss when

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 147
Table 4.4: Simulation (Scenario 2) of the shared
gamma frailty model with current status data
parameter true current status right censored

σ2 1.000 0.996 (0.072) 0.998 (0.052)

λ1 0.100 0.101 (0.015) 0.100 (0.008)
ϕ1 0.120 0.120 (0.026) 0.120 (0.015)
λ2 0.100 0.100 (0.015) 0.100 (0.008)
ϕ2 0.120 0.121 (0.026) 0.121 (0.014)

switching from right-censored event times to current status data is similarly

compared to Scenario 1. These simulations are undertaken to mimic the data
considered in the following example.

Example 4.3
The hepatitis A and B current status data from Example 1.7 are analyzed in a
bivariate model. In the second column of Table 4.5, the results from a shared
gamma frailty model, and in the third column from the univariate frailty
model (compare Table 3.9), are presented. The univariate frailty models treat

Table 4.5: Hepatitis A and B current status

data analyzed with the shared gamma frailty model
parameter shared frailty univariate frailty

σ12 0.522 (0.028) 2.018 (0.510)

λ1 0.012 (0.001) 0.008 (0.001)
ϕ1 0.037 (0.002) 0.085 (0.016)
σ22 0.522 (0.028) 10.037 (8.821)
λ2 0.002 (0.001) 0.002 (0.001)
ϕ2 -0.001 (0.007) 0.015 (0.142)
log-likelihood -2843.5100 -2841.4151

time to infection by hepatitis A and B, respectively, as independent times. It

allows for different variances. The shared gamma frailty model is restricted
to similar variances. Because of this constraint, the bivariate frailty model
provides no better fit compared to the univariate model. This is based on
the comparison of the loglikelihoods keeping in mind that the shared frailty
model contains one parameter less. More details including the SAS code can
be found in Hens et al. (2009).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
148 Frailty Models in Survival Analysis

4.4 Shared Log-normal Frailty Model

Together with the gamma distribution, the log-normal distribution is the most
important frailty distribution. The gamma distribution has a long tradition
as a frailty distribution for the last 30 years since the introduction of the
frailty model by Vaupel et al. (1979) and Lancaster (1979). The reason for
the popularity of the gamma distribution is its mathematical convenience,
especially due to the explicit representation of the survival function based on
the simple Laplace transform. This yields simple expressions of the likelihood
function, which are important for ML parameter estimation.
However, normally distributed random effects allow much more flexibility,
especially in modeling multivariate correlation structures. This point of view
becomes more pronounced in the next chapter. The development of new
user-friendly parameter estimation procedures, for example, in SAS (MCMC,
adaptive Gaussian quadrature techniques) in connection with a huge increase
in computational power, this flexibility of normal distributed random effects
outweighs its disadvantage of no explicit solution to the marginal survival
function including the normal integral.

Example 4.4
Analogous to the last section with gamma-distributed frailty, the Halluca
data set is analyzed, starting with parametric frailty models based on three
different baseline hazard functions.

Table 4.6: Parametric shared log-normal frailty model of the

Halluca lung cancer data
parameter Weibull exponential Gompertz

age (in years) 0.009 (0.003) 0.009 (0.003) 0.009 (0.003)

sex (females) -0.163 (0.073) -0.173 (0.073) -0.160 (0.073)
type (NSCLC) -0.106 (0.068) -0.119 (0.068) -0.097 (0.068)
ECOG (3 & 4) 0.601 (0.108) 0.633 (0.108) 0.613 (0.108)
stage II 0.448 (0.170) 0.475 (0.170) 0.440 (0.170)
stage IIIa 0.612 (0.133) 0.648 (0.133) 0.604 (0.133)
stage IIIb 0.977 (0.123) 1.040 (0.123) 0.964 (0.123)
stage IV 1.324 (0.113) 1.412 (0.112) 1.314 (0.113)
λ 0.029 (0.008) 0.019 (0.005) 0.025 (0.007)
ν 0.880 (0.020)
ϕ 0.025 (0.007)
σ2 0.305 (0.105) 0.376 (0.125) 0.297 (0.103)
log-likelihood -4836.43904 -4853.51499 -4837.138

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 149

Analysis is performed using SAS PROC NLMIXED, which allows writing

down of the conditional likelihood function, including the random effects. By
means of adaptive Gaussian quadrature techniques, the likelihood function
is integrated out and maximized. This is a very user-friendly approach that
allows for flexible modeling, especially in more complex models. It is easy to
see that the one-parameter exponential baseline hazard is not flexible enough
to model the mortality pattern accurately. The two-parameter Weibull and
Gompertz models show a much better fit based on the likelihood function
with slight advantage for the Weibull model. Parameter estimates in the
Weibull and Gompertz model are more similar to each other compared to the
exponential model. The exponential model seems to overestimate the frailty
variance, resulting in stronger fixed effects estimates.

Example 4.5
We perform a semiparametric analysis similar to Example 4.2 but with log-
normal-distributed frailty. The results given in Table 4.7 are very close to the
estimates obtained in the semiparametric shared gamma frailty model. This
underlines the robustness of the method against misspecification of the

Table 4.7: Univariate and shared log-normal

frailty model of the Halluca lung cancer data
parameter univariate frailty shared frailty

age (in years) 0.012 (0.003) 0.008 (0.003)

sex (females) -0.194 (0.088) -0.160 (0.072)
type (NSCLC) -0.112 (0.083) -0.097 (0.068)
ECOG (3 & 4) 0.924 (0.134) 0.605 (0.108)
stage II 0.552 (0.198) 0.450 (0.170)
stage IIIa 0.742 (0.154) 0.599 (0.133)
stage IIIb 1.177 (0.142) 0.956 (0.123)
stage IV 1.667 (0.130) 1.314 (0.113)
σ2 0.498 ( ) 0.283 ( )

frailty distribution. Parameter estimates are obtained by using the routine

COXPH in the R package which does not provide standard errors for the
frailty variance estimates.

The shared log-normal model was applied by McGilchrist and Aisbett (1991),
McGilchrist (1993), Gustafson (1997), and Bellamy et al. (2004). In the latter
paper, the case of interval-censored data is considered with underlying Weibull
baseline hazard function. The shared log-normal frailty model was extended
to allow for heterogeneity in the frailty distribution (dispersed frailty) by Lee
and Lee (2003).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
150 Frailty Models in Survival Analysis

4.5 Shared Positive Stable Frailty Model

The shared positive stable frailty model is of special interest for two reasons.
First, similar to the gamma frailty model, the simple form of the Laplace
transform allows simple maximum likelihood estimation in the parametric
model. Second, there exists no moments of this distribution. This was the
reason why Hougaard (1986a) introduced this distribution in the shared frailty
approach. Because of the infinite mean, the univariate positive stable frailty
model is not identifiable. This makes sense because, in shared frailty models,
the frailty parameter is interpreted as an association parameter, that should
not be identifiable from univariate observations. The shared positive stable
frailty model was discussed in detail by Hougaard (2000) and Duchateau and
Janssen (2008). The Laplace transform of a positive stable random variable
γ
with parameter γ is given in (3.41) by L(u) = e−u . Starting from this point
and averaging with respect to the cluster-specific random variable Zi produces
the multivariate survival function for the ith cluster in the form
ni
X 
′
S(ti1 , . . . , tini |Xi ) = L M0 (tij )eβ Xij
j=1
ni
h X γ i
′
= exp M0 (tij )eβ Xij
j=1
ni
h X
1/γ γ i
= exp − ln S(tij |Xij )
j=1

where the last relation is a consequence of equation (3.42). Focusing on frailty

models with fixed cluster size of two (ni = 2, i = 1, . . . , n), the joint survival
function simplifies to
h
1/γ
1/γ γ i
S(t1 , t2 ) = exp − ln S1 (t1 ) + − ln S2 (t2 ) , (4.7)

where the cluster index i is dropped but a single dichotomous covariate X

is assumed representing two strata-specific baseline hazard functions. This
results in two different marginal survival functions for the observations in
each pair.
A bivariate parametric shared positive stable frailty model with Weibull
baseline hazard function was used by Manatunga and Oakes (1999). The
authors applied the model to the Diabetic Retinopathy Study, which examined
the effectiveness of laser photocoagulation in delaying the onset of blindness
in patients with diabetic retinopathy. One eye of each patient was randomly
selected for photocoagulation treatment, and the other eye was left untreated.
The positive stable frailty allows for proportional hazards in both the marginal
and the conditional model.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 151

4.6 Shared Compound Poisson/PVF Frailty Model

The compound Poisson distribution is a convenient frailty distribution because
it includes the gamma and the inverse normal distribution, two important
special cases. As already discussed in the univariate model, it allows a part
of the study population to be not susceptible with respect to the event under
investigation. And last but not least, its explicit form of the Laplace transform
qualifies it also for use in a shared frailty model, allowing simple maximum
likelihood estimation in the parametric model. The shared compound Poisson
and PVF frailty model was considered, for example, by Hougaard (2000) and
Duchateau and Janssen (2008). Because the PVF model is a special case of
the compound Poisson model just with a restriction of the parameter space for
the parameter γ, the shared PVF frailty model is not considered separately
here. All formulas are similar in both cases.
The Laplace transform of a compound Poisson-distributed random variable
with mean one, variance σ 2 , and parameter γ is given by (3.53)
1 − γ σ 2 u γ

L(u) = exp 1 − (1 + ) .
γσ 2 1−γ
Assuming a compound Poisson distribution for the frailty, averaging with
respect to the cluster-specific random variable Zi produces the multivariate
survival function for the ith cluster
S(ti1 , . . . , tini |Xi ) (4.8)
X ni 
′
=L M0 (tij )eβ Xij
j=1
ni
h1 − γ  σ2 X ′
γ i
= exp 1− 1+ M0 (tij )eβ Xij
γσ 2 1 − γ j=1
ni
h1 − γ  X γσ 2
1/γ
γ i
= exp 1− 1+ 1− ln S(tij |Xij ) − ni
γσ 2 j=1
1−γ

where the last relation is a consequence of equation (3.54). Similar to the

gamma model, we focus on bivariate frailty models (ni = 2, i = 1, . . . , n),
where the joint survival function simplifies to
S(t1 , t2 ) =
h1 − γ  γσ 2
1/γ γσ 2
1/γ
γ i
exp 1− 1− ln S1 (t1 ) + 1− ln S2 (t2 ) −1 ,
γσ 2 1−γ 1−γ
where cluster index i is dropped, but a single binary covariate X is assumed
representing two strata-specific baseline hazard functions. This results in two
different marginal survival functions for the partners in each pair.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
152 Frailty Models in Survival Analysis

4.7 Shared Frailty Models More General

One of the major drawbacks of the shared frailty approach is the fact that
only positive association can be modeled. To overcome this strong limitation,
Gordon (2000) suggests the reciprocal frailty model with negatively dependent
competing risks. Negative dependence enters by means of individual-specific
frailty that simultaneously increases the hazard for one risk while decreasing
the hazard for the second. He introduced the model by assuming that the
individual-specific hazard function for one risk is directly proportional to the
heterogeneity term, while that of the second is inversely proportional, that
is, µ1 (t1 |Z) = Zµ01 (t1 ) and µ2 (t2 |Z) = Z1 µ02 (t2 ). The conditional bivariate
survival function is
n M02 (t2 ) o
S(t1 , t2 |Z) = exp − ZM01 (t1 ) − .
Z
Integrating with respect to the frailty distribution (here assumed to be a
gamma distribution Γ(λ, λ)), we obtain the unconditional bivariate survival
function

2λλ h M02 (t2 ) λ/2 i

S(t1 , t2 ) = Kλ (D(t1 , t2 )) ,
Γ(λ) M01 (t1 ) + λ
where Kλ (·) is the order λ modified Bessel function of the second kind, and
p
D(t1 , t2 ) = 2 M01 (t1 )M02 (t2 ) + λM02 (t2 ).
Because of the complicated form of the unconditional survival function, this
model was not further investigated in the literature.
An alternative for multivariate event times with dependence structures that
are more general than in the symmetric shared frailty model is given by
Hougaard (1986b). He considered the lifetimes of a monozygotic twin pair
and a third sibling with equal marginal distributions. Let Z be common in
the family, and let Y1 , Y2 be terms shared by individuals with identical genes.
Then, conditionally on Z, Y1 , Y2 , the hazard of each of the twins is ZY1 µ0 (t),
and for the third individual it is ZY2 µ0 (t). Consequently,

S(t1 , t2 , t3 |Z, Y1 , Y2 ) = P(T1 > t1 , T2 > t2 , T3 > t3 |Z, Y1 , Y2 )

= e−ZY1 (M0 (t1 )+M0 (t2 ))−ZY2 M0 (t3 ) .

If Y1 and Y2 are independent stable distributed with parameter γ, and Z γ is

stable distributed with parameter κ, they can be integrated out by using the
method of conditional means, giving
γ
+M0 (t3 )γ )κ
S(t1 , t2 , t3 ) = e−((M0 (t1 )+M0 (t2 )) .

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 153

This model allows for a stronger association between the lifetimes of the twins
compared to the association between a twin and the third sibling in the family.
Such kind of nonsymmetric correlation structure is considered in more detail
in Chapter 5.
A shared frailty model with log-skew-t distribution of the frailty (including
the log-normal distribution along with many other heavy-tailed distributions,
such as log-Cauchy or log t-distribution) is considered by Sahu and Dey (2004).
A shared gamma frailty model which is based on a Box–Cox transformation
is investigated by Yin and Ibrahim (2005).
A completely different approach to bivariate frailty modeling was used by
Bandyopadhyay and Basu (1990), and Gupta and Gupta (1990). The key
idea of their (more specific) model is a bivariate hazard model in the form

µ(t1 , t2 |Z) = Zµ0 (t1 , t2 ).

Here the two lifetimes are not conditionally independent given the frailty.
There are two types of dependencies. Dependence is caused by the bivariate
hazard model µ0 (t1 , t2 ) as well as by the shared frailty Z, which has to be
integrated out. Similar models based on different bivariate extensions of the
Weibull distribution are considered by Hanagal (2009). The advantage of such
models is that dependence between lifetimes is still present even if frailty as
a measure of heterogeneity is absent. Such models could be an alternative to
the frailty models considered here in this monograph. As opposed to these
models, multivariate frailty models simplify to univariate models (without
frailty) if the frailty variance approaches zero. Consequently, correlation and
unobserved heterogeneity are not completely separated in multivariate frailty
models. Without unobserved heterogeneity, indicated by σ 2 = 0, there exists
no correlation between lifetimes. This problem will be considered in more
detail in Section 5.9.

4.8 Dependence measures

It is difficult to compare the degree of dependence between different frailty
models. Often the variance of the frailty is used as a dependence measure, but
this measure is useless in case of frailty distributions without existing second
moments such as the positive stable distribution. Correlation coefficients
(with respect to the observed event times) are tricky to obtain because of
censored observations. Furthermore, they are often dependent on the time
scale, which makes them difficult to compare. Kendall’s τ can be used to
quantify dependence because it is independent of transformations on the time
scale and the frailty model used. It is a rank-based dependence measure. We
consider bivariate data, for example, the times to blindness of n pairs of eyes.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
154 Frailty Models in Survival Analysis

Then, the uncensored data is given by (T11 , T12 ), (T21 , T22 ), . . . , (Tn1 , Tn2 ),
where the first index denotes the cluster (patient), and the second index
denotes the eye. Here τ is defined as

τ = E sign{(Ti1 − Tj1 )(Ti2 − Tj2 )} , (4.9)
with sign{t} equal to -1 for negative values, 1 for positive values, and 0 if t = 0
(i, j = 1, . . . , n; i 6= j). Another formula is based on the Laplace transform:

Z
τ =4 sL′′ (s)L(s) ds − 1. (4.10)

To keep the notation simple, here and in the following all integrals are defined
over the interval [0, ∞) because lifetimes as well as frailties are nonnegative
variables. In the next paragraph we will derive relation (4.10). Starting from
definition (4.9), the conditional probability Ti1 < Tj1 (i, j = 1, . . . , n; i 6= j)
given the frailty vector (Zi , Zj ) is calculated

RR
ti <tj
f (ti , tj , Zi , Zj ) dti dtj
P(Ti1 < Tj1 |Zi , Zj ) =
f (Zi , Zj )
RR
ti <tj
f (ti , Zj )f (ti , Zj ) dti dtj
=
f (Zi )f (Zj )
R
f (t, Z1 )S(t, Z2 ) dt
=
f (Z1 )f (Z2 )
Z
= f (t|Zi )S(t|Zj ) dt
Z
= µ(t|Zi )S(t|Zi )S(t|Zj ) dt
Z
= Zi µ0 (t)e−(Zi +Zj )M0 (t) dt
Zi
= . (4.11)
Zi + Zj
For the last relation it is necessary that M0 (t) be a nondegenerated cumulative
hazard converging toward infinity as time approaches infinity. Hence, the
difference of the event times Ti1 − Tj1 is negative with probability ZiZ+Z i
j
,
j Z
and positive with probability Zi +Z j
(conditional on (Zi , Zj )). Because the
function sign{·} can take on only the two values minus and plus one (because
of the condition i 6= j zero can only occur with probability zero), it holds that

Zj − Zi
E (sign{(Ti1 − Tj1 )})|Zi , Zj = ,
Zi + Zj

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 155

where the expectation is conditional on (Zi , Zj ). An equivalent relationship

holds for the second observations in the pairs:

Zj − Zi
E (sign{(Ti2 − Tj2 )})|Zi , Zj = .
Zi + Zj

The lifetimes are independent given the frailties, which means

τ = E sign{(Ti1 − Tj1 )(Ti2 − Tj2 )}



= E E (sign{(Ti1 − Tj1 )})|Zi , Zj E (sign{(Ti2 − Tj2 )})|Zi , Zj
 Z − Z 2
j i
=E .
Zi + Zj

R∞
In the following we consider the integral 0
se−sx ds. Using integration by
parts it holds that

Z
s ∞ 1 Z
se−sx ds = − e−sx  + e−sx ds

x 0 x
1
Z
= e−sx ds
x
−1 ∞
= 2 e−sx 

x 0
1
= 2.
x
Applying this relation yields

Zj − Zi
2
Z Z Z
τ =E = se−s(zi +zj ) (zj − zi )2 f (zi )f (zj ) dzj dzi ds.
Zi + Zj

Using

Z Z
e−szj (zj − zi )2 f (zj ) dzj = zj2 e−szj f (zj ) dzj
Z
− 2zi zj e−szj f (zj ) dzj
Z
+ zi2 e−szj f (zj ) dzj

= L′′ (s) + 2zi L′ (s) + zi2 L(s)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
156 Frailty Models in Survival Analysis

we obtain
Z Z Z
τ = se−s(zi +zj ) (zj − zi )2 f (zi )f (zj ) dzj dzi ds
Z Z
= sL′′ (s)e−szi f (zi ) dzi ds
Z Z
+2 sL′ (s)zi e−szi f (zi ) dzi ds
Z Z
+ sL(s)zi2 e−szi f (zi ) dzi ds
Z Z Z Z
= sL′′ (s)L(s) ds − 2 s(L′ (s))2 ds + sL(s)L′′ (s) ds
Z Z
= 2 sL′′ (s)L(s) ds − 2 s(L′ (s))2 ds. (4.12)

As a consequence of integration by parts it holds that

Z ∞ Z
sL′′ (s)L(s) ds = L′ (s)sL(s) − L′ (s)(L(s) + sL′ (s)) ds

0
Z Z
= − s(L (s)) ds − L′ (s)L(s) ds.
′ 2
(4.13)

Using integration by parts again we see that

Z ∞ Z
L′ (s)L(s) ds = (L(s))2  − L′ (s)L(s) ds,

0

resulting in

Z
2 L′ (s)L(s) ds = −1. (4.14)

Combining (4.13) and (4.14) implies

1
Z Z
s(L′ (s))2 ds = − sL′′ (s)L(s) ds + .
2

Substituting this expression in (4.12) yields relation (4.10). This can be

used to calculate Kendall’s τ for shared frailty models with different frailty
distributions. In the case of the shared gamma frailty model, by using (3.14)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 157

and (3.15) with λ = k = 1/σ 2 , it holds that

Z Z
1 1
sL (s)L(s) ds = s(1 + σ 2 )(1 + σ 2 s)− σ2 −2 (1 + σ 2 s)− σ2 ds
′′

Z
2
= (1 + σ 2 ) s(1 + σ 2 s)− σ2 −2 ds

1 + σ2
Z
2
= 2
(1 + σ 2 s)− σ2 −1 ds
2+σ
1 + σ2 2 − σ22 
∞
= (1 + σ s)
−2(2 + σ 2 )

0

1 + σ2
= .
2(2 + σ 2 )

Consequently,

σ2
Z
τ =4 sL′′ (s)L(s) ds − 1 = .
σ2 + 2
For the shared positive stable frailty model, Kendall’s τ can be also easily
calculated using (3.41)

Z Z
γ γ
sL′′ (s)L(s) ds = −γ(γ − 1)sγ−1 e−2s + γ 2 s2γ−1 e−2s ds. (4.15)

The first expression of (4.15) gives

γ−1
Z
γ
−γ(γ − 1)sγ−1 e−2s ds = − ,
2
and using integration by parts it holds for the second expression of (4.15) that

1 γ −2sγ ∞ 1 2 γ
Z Z
2 2γ−1 −2sγ γ
γ s e ds = − γs e  + γ sγ−1 e−2s = .
2 0 2 4
Finally,

Z
τ =4 sL′′ (s)L(s) ds − 1
γ − 1 γ

=4 − + −1
2 4
= 1 − γ.

Unfortunately, for most other frailty distributions, Kendall’s τ does not have
such an explicit and simple form.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
158 Frailty Models in Survival Analysis

4.9 Limitations of the Shared Frailty Model

Shared frailty explains correlations within clusters. A cluster can consist of
individuals from the same group, such as a family, litter, clinic, community;
or of multiple or recurrent events from the same individual. However, it does
have some limitations. In the following, we use some of the arguments already
discussed by Xue and Brookmeyer (1996).
First, the concept of shared frailty forces the unobserved factors to be the
same within the cluster, which is often not appropriate. For example, in
general, it may be inappropriate to assume that both partners in a twin pair
or all relatives in a family share all of their unobserved risk factors.
Second, in most cases, shared frailty will only induce positive associations
within clusters (for exceptions see Joe 1993). However, in some situations the
survival times for subjects within the same cluster are negatively associated.
For example, growth rates of animals living in the same cage with limited
food are probably negatively associated. Another example are transplantation
studies, which found out that generally the longer an individual must wait for a
transplantation, the shorter the chances of survival after the transplantation.
Thus, the waiting and the survival time may be negatively associated. As
another example, suppose a patient is repeatedly admitted to hospital for the
same disease. The sicker the patient is, the higher the risk of readmission,
and the lower the ‘risk’ (chance) of discharge. Therefore, the duration of the
stays inside the hospital, on the one hand, and outside the hospital, on the
other hand, are expected to be negatively associated. An additional example
is provided by competing risk scenarios, where the reduction of the risk of
dying from one disease increases the risk of dying from another disease.
Third, the dependence between survival times within the cluster is based
on marginal distributions of survival times. To see this, when covariates are
present in a proportional hazards model with a gamma-distributed frailty,
the dependence parameter and the population heterogeneity are confounded
(Clayton and Cuzick 1985b), meaning that the joint event time distribution
can be identified from the marginal distributions (Hougaard 1986a). Elbers
and Ridder (1982) show that this problem exists for any frailty distribution
with a finite mean. However, ”shared frailty” in multivariate models differs
from ”individual frailty” used in the case of univariate data. Originally, this
difference in the notions of frailty was not clearly understood. But it is worth
noting that the value of σ 2 estimated from univariate data may, in fact, has
nothing to do with association. Indeed, consider two hypothetical bivariate
data sets with different associations between lifespans of related individuals,
for example, the lifetimes of men and their sons, and between the same men
and their grandsons. In both situations the value of the variance parameter σ 2
will be the same estimated from the data on the group of men (grandfathers),
despite the fact that associations between lifespans of grandfathers and their

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Shared Frailty Models 159

grandsons, and grandfathers and their sons are different. The use of other
association measure such as Kendall’s τ does not really help in this situation
because it is a function of the frailty variance.
This last, and maybe most important, limitation of shared frailty models is
a consequence of identifiability of the univariate frailty model with observed
covariates. Hence, it is an inherent feature to all shared frailty models with
a finite mean of frailty distribution. To overcome this problem, Hougaard
(1986a, 1987) suggests the shared positive stable frailty model. In this case
the univariate model with observed covariates is not identifiable because the
mean of the positive stable distribution is infinite. So one can expect more
flexibility from a shared frailty model with positive stable distribution than
from models with gamma frailty. The bivariate survival function in the shared
positive stable frailty model is given by (4.7)
h
1/γ
1/γ γ i
S(t1 , t2 ) = exp − ln S1 (t1 ) + − ln S2 (t2 ) . (4.16)
When single covariates Xi (i = 1, 2) are observed, and the conditional hazard
is (3.2), then the univariate survival functions are
γ ∗
(t)eγβi Xi = exp − Mi∗ (t)eβi Xi ,



Si (t) = exp − M0i
γ
where βi∗ = γβi and Mi∗ (t) = M0i (t). Using a shared positive stable frailty
model it is possible to estimate both association parameter γ and regression
coefficients βi from data on related individuals. One problem remains yet
unsolved: the interpretation of regression parameters βi . To illustrate this
problem, let β = β1 = β2 for two groups of relatives, for example, MZ and
DZ twins. It is clear that the association parameter γ is different for MZ and
DZ twins. Hence the values of parameter β ∗ = γβ and M ∗ (t) in equation
(4.16) are also different for MZ and DZ twins, which contradicts the natural
assumption that the survival of these individuals conditioned on observed
covariates follow the same survival model. If the parameters β ∗ = γβ are
assumed to be the same for MZ and DZ twins, then the parameters β and
the baseline hazards µ0 (t) should be different for these pairs of individuals,
which creates a problem for the interpretation of the conditional hazard (3.2)
for this model.
The problem can be described in a more general way. The main feature of
the shared frailty approach is its symmetry. In multicenter clinical trials it
is reasonable to assume a symmetric relationship between all possible pairs
of patients in a study center because patients in a center are exchangeable.
It makes no sense to assume a nonsymmetric correlation structure inside the
centers. The situation changes dramatically when considering family studies
with relatives of different relationship. Applying a shared frailty model would
imply the same relationship (correlation) between any pairs of relatives in
the family. This is, of course, not a reasonable assumption. Furthermore, it
contradicts the assumption of different relationships in families, which is the
basis for genetic studies.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
160 Frailty Models in Survival Analysis

To avoid such methodological problems, correlated frailty models have been

developed for the analysis of multivariate failure time data. We will focus here
mainly on bivariate models because generalizations of correlated frailty models
to the multivariate case are not straightforward, and these will be discussed
later on in more detail in some specific situations. Bivariate correlated frailty
models have two associated random variables (Z1 , Z2 ) that characterize the
frailty effect for each cluster. One random variable, for example, is assigned to
individual 1 and one to individual 2 so that they are no longer constrained by
having a common frailty. The two random variables (Z1 , Z2 ) are associated
and jointly distributed; therefore, knowing one of them does not necessarily
imply the other. Also, the two variables may be negatively associated, and
this would induce a negative association between survival times. Similar to
the univariate case as well as to shared frailty, the choice of frailty distribution
is important for the modeling and will be considered in more detail later on.
Between shared and correlated frailty models (sometimes called univariate
and bivariate frailty models with respect to the dimension of the frailty, but
both models are considered here to deal with bivariate time-to-event data),
there is an intermediate approach in which two variables are assigned to each
cluster (of size two) to count for heterogeneity but are generated from one
common random variable. An example are bivariate survival times (T1 , T2 )
with Z1 to account for the heterogeneity of T1 and Z2 to account for T2 , where
Z1 and Z2 are defined by

Z1 = eα1 W and Z2 = eα2 W , (4.17)

where W is a random variable, and α1 and α2 are parameters to be estimated.
This model has been used frequently (Flinn and Heckman 1982, Clayton and
Cuzick 1985b, Heckman and Walker 1990, Huang and Wolfe 2002) and is
named one-factor loading specification in the econometric literature (van den
Berg 2001). This approach is more flexible than that of assuming shared
frailty for T1 and T2 , and to some extent it allows for a negative association
by allowing different signs of α1 and α2 . However, it still imposes a linkage
between the variance and the correlation, and also the expectation and the
variance of the frailties. The overall hazard for the population (rather than
the individual hazard), in consequence, is forced to move between restricted
ranges. Therefore, Lindeboom and van den Berg (1994) conclude that it is
hazardous to estimate bivariate survival models where the mixing distribution
is parameterized univariately because a univariate random variable may not be
able to account for both the dependence of the survival times and the change
in the composition of the sample due to unobserved heterogeneity caused by
covariates not being included in the analysis. Clearly, such kinds of problems
do not arise in a genuine bivariate approach in which the dependence between
T1 and T2 can be changed without changing the marginal distributions of T1
and T2 . This will be the topic of the following chapter.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com

Chapter 5
Correlated Frailty Models

In the last chapter we focused on the shared frailty model as a way of modeling
associations between event times. The concept of shared frailty goes back to
Clayton (1978), and was extensively studied in books by Hougaard (2000),
and Duchateau and Janssen (2008). The shared frailty approach has proven
to be a useful and popular extension of the Cox model when observations
from subjects are not statistically independent of each other. As discussed in
Section 4.9, the shared frailty model is especially useful for clustered event
time data when the correlation structure in the clusters is symmetric and not
of special interest as, for example, in multicenter clinical trials.
Similar to the univariate and shared frailty model, the correlated frailty
model is a mixture model because frailty for each individual is assumed to be
random. This is also known as the concept of random hazards. The model is
based on the assumption that event times in a cluster are independent, given
the vector of frailties. In other words, it is a conditional independence model
where the frailty variables are correlated but not necessarily common for all
individuals in a cluster, and therefore, responsible for dependence between
event times. The shared frailty model is a special case of the correlated
frailty model with correlations between the frailties equal to one. A correlated
frailty model can be considered as a mixed (random effects) model in survival
analysis, with group and individual variation both included in the distribution
of the frailty vector. Therefore, correlated frailty models contain association
characteristics of frailty (correlation coefficients) among other parameters,
which makes them especially convenient for genetic studies of relatives with
event-time outcome. In particular, questions about the role of genes and
environment in susceptibility to diseases and death can be addressed. Here,
the correlation between frailties is of key interest. This is different from many
other applications where correlation is treated as a nuisance to be adjusted
for in the model, and not of special interest. Consequently, we will focus,
the applications of the correlated frailty model, mainly on twin and family
event-time data. In such kinds of applications, the researcher is usually faced
with small cluster sizes. Special focus is on the bivariate case.
It is assumed that there is independence between observations from different
clusters. If variances of the frailty variables are zero, this implies independence
between event times in the clusters. The structure of the model makes them
convenient for application of the EM algorithm, allowing for semiparametric

161
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
162 Frailty Models in Survival Analysis

estimation. The model can also be used to model dependent competing

risks. Whereas the correlated gamma frailty model can only cover positive
dependence between event times, the log-normal correlated frailty model offers
a wider range of dependence between event times in the clusters. Hence, this
approach has become a very attractive tool, with a flexibility in modeling
correlation structures known from mixed models.
For the popular shared frailty model, many theoretical and practical results
exist in the literature on frailty models. The model is implemented in many
statistical packages, and, therefore accessible even to nonspecialists. For
the correlated frailty model, there exist only a few papers, and software is
practically not available. This has changed with the reformulation of the
frailty model as a random effects/mixed model and with the application of
recent results already achieved in this research area. Furthermore, during
the last few years, powerful and user-friendly software solutions have become
available for mixed models with normal distributed random effects. This
opens new avenues for research in the field of correlated frailty models which
is the topic of the present chapter. The similarity to mixed models is of special
importance for correlated frailty models and will be exploited throughout this
chapter.
Some history exists behind the development of correlated frailty models.
Aalen (1987) discusses multivariate mixing distributions applied on a Markov
chain. Marshall and Olkin (1988) discussed various multivariate correlated
frailty distributions. Their work was continued by Yashin et al. (1993, 1995),
Pickles et al. (1994), Yashin and Iachine (1997, 1999a,b), Petersen (1998),
Iachine et al. (1998), and Iachine (2002), who mainly considered the bivariate
correlated gamma frailty model and applied it to twin survival data. This
approach forms the starting point for different extensions discussed in more
detail in this monograph later on. Some of the results are already published
in Wienke et al. (2000, 2001, 2002, 2003a,b, 2004, 2005a,b, 2006a,b, 2010).
Another approach with an interesting real data example can be found in van
den Berg et al. (2008), who combine twin data with economic information
to analyze the influence of the economic situation at year of birth on later
developments of cause-specific mortality. Xue and Brookmeyer (1996) deal
with the bivariate correlated log-normal frailty model, developed in more
general cases by Ripatti and Palmgren (2000), and Vaida and Xu (2000).
In the following section the main ideas of the correlated frailty concept
are outlined, independent of frailty distribution. This more general part is
followed by a detailed description and discussion of the correlated gamma
frailty model in Section 5.2. The correlated log-normal frailty model as a very
flexible model is considered in Section 5.3. A MCMC estimation procedure
for the correlated log-normal frailty model is presented in Section 5.4. The
newly suggested correlated compound Poisson frailty in Section 5.5 contains
the correlated gamma and the correlated inverse Gaussian frailty model as
special cases. It allows single individuals in the cluster to be at no risk (frailty
equal to zero) for the event of interest.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 163

5.1 The Concept of Correlated Frailty

The correlated frailty model is the second important concept in the area of
multivariate frailty models. It is a natural extension of the shared frailty
approach on the one hand, and of the univariate frailty model on the other.
In the correlated frailty model, the frailties of individuals in a cluster are
correlated but not necessarily shared. It enables the inclusion of additional
correlation parameters, which then allows the addressing of questions about
associations between event times. Furthermore, associations are no longer
forced to be the same for all pairs of individuals in a cluster. This makes
the model especially appropriate for situations where the association between
event times is of special interest, for example, genetic studies of event times
in families. We will come back to such applications later on. For introductory
reasons we will discuss bivariate models first. The conditional survival function
in the bivariate case (here without observed covariates) looks like

S(t1 , t2 |Z1 , Z2 ) = S1 (t1 |Z1 )S2 (t2 |Z2 ) = e−Z1 M01 (t1 ) e−Z2 M02 (t2 ) ,

where Z1 and Z2 are two correlated frailties. The distribution of the random
vector (Z1 , Z2 ) needs to be specified and determines the association structure
of the event times in the model.
Consider some bivariate event times – for example, the lifetimes of twins, or
age at onset of a disease in spouses, time to blindness in the left and right eye,
or time to failure in the left and right kidney of patients. In the (bivariate)
correlated frailty model, the frailty of each individual in a pair is defined by
a measure of relative risk, that is, exactly as it was defined in the univariate
case. For two individuals in a pair, frailties are not necessarily the same, as
they are in the shared frailty model. We are assuming that the frailties are
acting multiplicatively on the baseline hazard function (proportional hazards
model) and that the observations in a pair are conditionally independent,
given the frailties. Hence, the hazard of the individual j (j = 1, 2) in pair
i (i = 1, . . . , n) has the form

′
µ(t|Xij , Zij ) = Zij µ0j (t)eβ Xij , (5.1)

where t denotes age or time, Xij is a vector of observed covariates, β is a

vector of regression parameters describing the effect of the covariates Xij ,
µ0j (·) are baseline hazard functions, and Zij are frailties. Bivariate correlated
frailty models are characterized by the joint distribution of a two-dimensional
vector of frailties (Zi1 , Zi2 ). If the two frailties are independent, the resulting
lifetimes are independent, and no clustering is present in the model. If the
two frailties are equal, the shared frailty model is obtained as a special case
of the correlated frailty model with correlation one between the frailties.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
164 Frailty Models in Survival Analysis

Any method in this context is based on likelihood functions. In order to derive

a marginal likelihood function, the assumption of conditional independence
of lifespans, given the frailty, is used. Let δij be a censoring indicator for
individual j (j = 1, 2) in pair i (i = 1, . . . , n). Indicator δij is 1 if the
individual has experienced the event of interest, and 0 otherwise. According
to (5.1), the conditional survival function of the jth individual in the ith pair
is

β ′ Xij
S(t|Xij , Zij ) = eZij M0j (t)e ,

with M0j (t) denoting the cumulative baseline hazard function. Here and
in the following, S is used as a generic symbol for a survival function. The
contribution of individual j (j = 1, 2) in pair i (i = 1, . . . , n) to the conditional
likelihood is given by

 ′
δij β ′ Xij
Zij µ0j (tij )eβ Xij eZij M0j (tij )e ,

where tij stands for observation time of individual j from pair i. Assuming
the conditional independence of lifespans, given the frailty, and integrating
out the frailty, we obtain the marginal likelihood function

n ZZ  δi1
Y ′ β ′ Xi1
zi1 µ01 (ti1 )eβ Xi1 ezi1 M01 (ti1 )e (5.2)
i=1
R+ ×R+
 ′
δi2 β ′ Xi2
× zi2 µ02 (ti2 )eβ Xi2 ezi2 M02 (ti2 )e f (zi1 , zi2 ) dzi1 dzi2 ,

where f (·, ·) is the probability density function of the corresponding frailty

distribution. All these formulas can be easily extended to the multivariate
case, but need a specification of the correlation structure between individuals
in a cluster in terms of the multivariate density function, which complicates
analysis. An example is investigations dealing with family data. Here, the
correlation structure is given by the degree of kinship. This remains relatively
easy if only kernel families are considered (father–mother–child, for example),
but becomes much more difficult if relatives of any degree are included such as
half-sibs, grand–parents, nieces, and nephews, etc. Furthermore, the varying
family size may also causes additional nontrivial technical problems in the
estimation procedure.
Similar to univariate and shared frailty models, the choice of the frailty
distribution is of great importance for modeling specific features of the joint
survival distribution. Next, correlated frailty models with different frailty
distribution are considered.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 165

5.2 Correlated Gamma Frailty Model

This model was introduced by Yashin et al. (1993, 1995) and applied to
related lifetimes in many different settings, for example, by Pickles et al.
(1994), Yashin and Iachine (1995a,b, 1997, 1999a,b), Yashin et al. (1996),
Zahl (1997), Iachine et al. (1998), Iachine (2002), Petersen (1998), Wienke et
al. (2000, 2001, 2002, 2003a,b, 2004, 2005a), Zdravkovic et al. (2002, 2004),
Kheiri et al. (2005), and van den Berg et al. (2008). The model has a very
convenient representation of the survival function in closed form expressions,
which allows nice interpretation of the model parameters. Here we restrict
again to the bivariate model first.
Let k0 , k1 , k2 be some nonnegative real-valued numbers. Set λ1 = k0 + k1
and λ2 = k0 + k2 . Let Y0 , Y1 , Y2 be independent, gamma-distributed random
variables with parameters Y0 ∼ Γ(k0 , λ0 ), Y1 ∼ Γ(k1 , λ1 ), and Y2 ∼ Γ(k2 , λ2 ).
Consequently,

λ0
Z1 = Y0 + Y1 ∼ Γ(k0 + k1 , λ1 ) (5.3)
λ1

λ0
Z2 = Y0 + Y2 ∼ Γ(k0 + k2 , λ2 ) (5.4)
λ2
1 1
and EZ1 = EZ2 = 1, V(Z1 ) = λ1 := σ12 , V(Z2 ) = λ2 := σ22 . Then the
following relation holds

λ0 λ0
cov(Z1 , Z2 ) = cov( Y0 + Y1 , Y0 + Y2 )
λ1 λ2
λ20
= V(Y0 )
λ1 λ2
λ20 k0
=
λ1 λ2 λ20
k0
=
(k0 + k1 )(k0 + k2 )

This leads to the correlation

cov(Z1 , Z2 ) k0
ρ= p =p .
V(Z1 )V(Z2 ) (k0 + k1 )(k0 + k2 )
1
Consequently, because of the relation k0 + ki = λi = σi2
(i = 1, 2), it holds
σ
1− σ i ρ
ρ 1 j
that k0 = σ1 σ2 and ki = σi2
− k0 = σi2
(i, j = 1, 2; i 6= j). Now we can

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
166 Frailty Models in Survival Analysis

derive the unconditional survival function, applying the Laplace transform of

gamma-distributed random variables. Hence,

S(t1 , t2 ) = ES(t1 , t2 |Z1 , Z2 )

= ES1 (t1 |Z1 )S2 (t2 |Z2 )
= Ee−Z1 M01 (t1 ) e−Z2 M02 (t2 )
λ0 λ0
= Ee−( λ1 Y0 +Y1 )M01 (t1 ) e−( λ2 Y0 +Y2 )M02 (t2 )
λ λ
−Y0 ( λ0 M01 (t1 )+ λ0 M02 (t2 ))−Y1 M01 (t1 )−Y2 M02 (t2 )
= Ee 1 2

1 λ0 λ0
= (1 + ( M01 (t1 ) + M02 (t2 )))−k0
λ0 λ1 λ2
1 1
× (1 + M01 (t1 ))−k1 (1 + M02 (t2 ))−k2
λ1 λ2
−ρ
= (1 + σ12 M01 (t1 ) + σ22 M02 (t2 )) σ1 σ2
σ1 σ2
−1+ ρ −1+ ρ
σ2 σ1

× (1 + σ12 M01 (t1 )) σ2

1 (1 + σ22 M02 (t2 )) σ2
2 , (5.5)

which results in the following representation of the correlated gamma frailty

model
σ σ
1− σ1 ρ 1− σ2 ρ
S1 (t1 ) 2 S2 (t2 ) 1
S(t1 , t2 ) = ρ , (5.6)
(S1 (t1 ) −σ12 + S2 (t2 ) −σ22 − 1) σ1 σ2
using independence of the gamma-distributed variables Y0 , Y1 , Y2 and (3.17).
This representation is called here the copula representation, because it re-
veals that the correlated gamma frailty model can be considered as a copula,
but this copula is not an Archimedian copula as in the shared gamma frailty
model. Copula representations allow separation of the marginal part of the
model (univariate distributions) from the correlation structure of the model.
Besides the fact that frailty models and copulas look very similar, it is im-
portant to note that there are also differences between both approaches that
are often overlooked. For more details regarding this exciting aspect, we refer
the interested reader to the paper by Goethals et al. (2008) and Chapter 6 in
this monograph.
The possible range of the correlation between frailties depends on the values
of σ1 and σ2 :
σ1 σ2
0 ≤ ρ ≤ min{ , }.
σ2 σ1
Hence, if σ1 6= σ2 , the correlation between the frailties is always less than one.
This property can be a serious limitation, especially when the values of σ1
and σ2 differ strongly.
Partial derivatives of the bivariate survival function (which are necessary
for the likelihood) are given in Appendix A.2. The log-likelihood contribution

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 167

of the ith pair in the parametric case without censoring, truncation, and
observed covariates is given by

ρ 2 ρ 2
log Li (θ, σ 2 , ρ) = ln(S1 (ti1 ; θ)−σ1 ) + ln(S2 (ti2 ; θ)−σ2 )
σ1 σ2 σ1 σ2
ρ 2 2
−( + 2) ln(S1 (ti1 ; θ)−σ1 + S2 (ti2 ; θ)−σ2 − 1)
σ1 σ2
 σ1 σ2 σ2 2
+ ln 1 − ρ − ρ + ρ2 + (1 − ρ)S1 (ti1 ; θ)−2σ1
σ2 σ1 σ1
σ1 2
+ (1 − ρ)S2 (ti2 ; θ)−2σ2
σ2
σ1 σ2 2 2
+ (2 − ρ − ρ + σ1 σ2 ρ + ρ2 )S1 (ti1 ; θ)−σ1 S2 (ti2 ; θ)−σ2
σ2 σ1
σ1 σ2 2
+ (−2 + ρ + 2 ρ − ρ2 )S1 (ti1 ; θ)−σ1
σ2 σ1
σ2 σ1 2

+ (−2 + ρ + 2 ρ − ρ2 )S2 (ti2 ; θ)−σ2 .
σ1 σ2

In the following we apply the correlated gamma frailty model to cause-specific

mortality data of Danish twins, described in Example 1.4. The assumption of
conditional independence plays an important role in genetic analysis of event
times. Under this assumption, the frailty variable becomes the only carrier of
genetic influence on durations. It means that the underlying hazard represents
only nongenetic influence on the event time. Univariate frailty models do not
have such a property.

Example 5.1
Age at death by cancer (all types of cancer combined) and age at death by
stroke in twin pairs are the event times of interest. Because of the symmetric
structure of the twin data, we use the simplifications S(t) = S1 (t) = S2 (t)
and σ 2 = σ12 = σ22 . A parametric approach with a Gompertz baseline hazard
µ0 (t) = λeϕt is fitted, resulting in the univariate marginal survival function

λ 1
S(t) = (1 + σ 2 (eϕt − 1))− σ2 .
ϕ
The data is right censored, and left truncated, which has to be included into
the likelihood. The results of the maximum likelihood parameter estimation
procedure are given in Table 5.1.
For each cause of death, a separate analysis of the same data was performed,
treating all other causes of death as independent censored observations. To
allow a combined analysis of monozygotic (MZ) and dizygotic (DZ) twins, we
include two correlation coefficients into the model, ρMZ and ρDZ , respectively.
These correlations between MZ and DZ twins provide important information
about genetic and environmental influences on frailty (susceptibility) within

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
168 Frailty Models in Survival Analysis
Table 5.1: Parameter estimates in the correlated gamma frailty model
applied to Danish twin pairs
males females
cancer stroke cancer stroke

σ2 3.388 (0.572) 5.146 (1.373) 7.194 (1.197) 1.794 (0.793)

ρMZ 0.350 (0.088) 0.706 (0.160) 0.245 (0.065) 0.688 (0.316)
ρDZ 0.194 (0.068) 0.208 (0.119) 0.076 (0.041) 0.407 (0.254)
λ 4.4e-6 (1.5e-6) 5.8e-8 (3.9e-8) 7.3e-6 (2.8e-6) 1.0e-7 (5.6e-8)
ϕ 0.118 (0.006) 0.160 (0.010) 0.111 (0.007) 0.144 (0.008)

individuals (see Appendix A.6). For both causes of death and both sexes, the
estimates of correlation of frailty for MZ twins (ρMZ ) are higher than that for
DZ twins (ρDZ ). These higher correlations in MZ twins compared to DZ twins
indicate the importance of genetic factors involved in susceptibility (frailty) to
cause-specific mortality. Estimates of correlations range between 0.08 (cancer,
DZ female twins) and 0.71 (stroke, MZ male twins). The differences in the
estimates of the frailty variance σ 2 indicate different levels of heterogeneity.
The shared gamma frailty model ρ = 1 is sometimes applied to twin data, but
results in different estimates of σ 2 for MZ and DZ twins (Hougaard 2000).
This contradicts the observation that univariate (marginal) lifetimes of MZ
and DZ twins are similar.
Parameter estimates were obtained using a selfwritten GAUSS routine. The
routine is based on constrained maximum likelihood estimation, maximizing
the parametric likelihood function under user-specified constraints on the
parameters. The likelihood function for left-truncated and right-censored
bivariate data is given in Appendix A.1 and A.2. We will reanalyze the data
again in the following example with the emphasis on obtaining quantitative
results about the heritability of cause-specific mortality.

Example 5.2
This example deals with an extension of the above correlated gamma frailty
model which allows the integration of genetic models (see Appendix A.6).
In this approach, the correlations between the frailties of family members are
substituted by their respective variance components. In the present case of MZ
and DZ twins, relation (A.8) can be used. In ACE models containing additive
genetic factors (A), common environment (C), and unique environment (E),
the corresponding variance components are denoted by a2 , c2 , and e2 . Using
this reparameterization of the model, the data from Example 5.1 are analyzed
again in Table 5.2. The main interest is in the estimation of the heritability
(a2 ), which quantifies the importance of genetic factors on a trait. Here,
the trait of interest is susceptibility to death caused by cancer and stroke,
respectively. Again, a Gompertz baseline hazard is used.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 169
Table 5.2: Heritability of cause of death. Parameter estimates in the
correlated gamma frailty model applied to Danish twin pairs
males females
cancer stroke cancer stroke

σ2 3.388 (0.572) 4.956 (1.360) 7.233 (1.198) 1.791 (0.796)

a2 0.311 (0.215) 0.635 (0.150) 0.211 (0.052) 0.564 (0.678)
c2 0.038 (0.158) 0.000 ( - ) 0.000 ( - ) 0.126 (0.509)
e2 0.650 (0.088) 0.365 (0.150) 0.789 (0.052) 0.311 (0.316)
λ 4.4e-6 (1.5e-6) 6.3e-8 (4.1e-8) 7.2e-6 (2.8e-6) 1.0e-7 (6.0e-8)
ϕ 0.118 (0.006) 0.159 (0.010) 0.111 (0.007) 0.144 (0.008)

Heritability estimates differ by gender and cause of death, ranging from 0.21
(cancer in females) to 0.63 (stroke in males). At least moderate heritability
is usually the basis for further linkage and association studies to identify risk
genes.
This approach from event history analysis permits accounting for censoring
and truncation present in the data. The method is combined with variance
component models from genetic epidemiology. For such a combined analysis,
we apply the correlated gamma frailty model, which takes into account the
dependence of lifespans of relatives. This enables the estimation of the effect
of genetic factors in susceptibility to cancer and stroke. The approach allows
the combination of data on age of death and cause of death and for overcoming
the drawbacks of the traditional concordance analysis usually applied in twin
studies with time-to-event data. For each twin partner two competing risks of
latent times (with respect to death due to cancer or stroke, and with respect to
death due to all other diseases including censoring) are modeled. In addition,
we assume that these competing risks are independent.

In the following, an adaption of the bivariate correlated gamma frailty

model to a three-dimensional approach in genetics is considered. This model
was introduced by Korsgaard and Andersen (1998). It is based on the idea of
correlated gamma frailty but with pre-specified correlations between family
members. In one of their examples, the authors consider families of size three
with a father (F), a mother (M), and one child (C). Let V1 represent the
part of the father’s genome affecting frailty that is transmitted to the child,
and V2 the corresponding part of the father’s genome not transmitted to
the child. Similarly, V3 represents the part of the mother’s genome affecting
frailty transmitted to the child, and V4 the corresponding part of the mother’s
genome not transmitted to the child. Assuming that the father and mother
are unrelated, additive genetic gamma frailties for the father, ZF , the mother,
ZM , and the child, ZC , are given by

ZF = V1 + V2 ZM = V3 + V4 ZC = V1 + V3 .

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
170 Frailty Models in Survival Analysis

Here, it is necessary to mention that the interpretation of frailty in the present

model is different from the interpretation of frailty in the foregoing correlated
frailty model applied to twin data to estimate heritability of specific traits.
In the Korsgaard and Andersen model, frailty only represents genetic factors,
whereas in the correlated frailty model used in Example 5.2, frailty represents
genetic as well as environmental factors.
Assuming that the Vi ’s are i.i.d., with Vi ∼ Γ( λ2 , λ), so that E(Vi ) = 21 and
1
V(Vi ) = 2λ , then E(ZF ) = E(ZM ) = E(ZC ) = 1 and

1
 
1 0
1 2
1
V(ZF , ZM , ZC ) =  0 1 2
.
λ 1 1
2 2 1

The correlation between the frailty of the father ZF or the mother ZF and
frailty of the child ZC is one half due to the fact that the two individuals share
half of their genes. The frailties of the father and mother in each family are
assumed to be independent. It is assumed that the baseline hazard functions
were unknown (semiparametric model). In the bivariate cases (father – child
or mother – child), the genetic frailty model is a correlated frailty model with
correlation equals to one half. This can be seen from the trivariate survival
function of the model (with reparameterization σ 2 = 1/λ)

S(t1 , t2 , t3 ) = Ee−ZF M1 (t1 ) e−ZM M2 (t2 ) e−ZC M3 (t3 )

= Ee−(V1 +V2 )M01 (t1 )−(V3 +V4 )M02 (t2 )−(V1 +V3 )M3 (t3 )
= Ee−V1 (M01 (t1 )+M3 (t3 )) e−V3 (M02 (t2 )+M3 (t3 )) e−V2 M01 (t1 ) e−V4 M02 (t2 )
1 λ 1 λ
= (1 + (M01 (t1 ) + M3 (t3 )))− 2 (1 + (M02 (t2 ) + M3 (t3 )))− 2
λ λ
1 λ 1 λ
× (1 + M01 (t1 ))− 2 (1 + M02 (t2 ))− 2
λ λ
2 2
− 1
= S1 (t1 )−σ + S3 (t3 )−σ − 1 2σ2
2 2
− 1 1 1
× S2 (t2 )−σ + S3 (t3 )−σ − 1 2σ2 S1 (t1 ) 2 S2 (t2 ) 2

using the Laplace transform of a gamma distributed random variable (3.17).

Obviously, the following relations hold:

S(t1 , t2 , 0) = S1 (t1 )S2 (t2 )

2 2
− 1 1 1
S(t1 , 0, t3 ) = S1 (t1 )−σ + S3 (t3 )−σ − 1 2σ2
S1 (t1 ) 2 S3 (t3 ) 2
2 2
− 1 1 1
S(0, t2 , t3 ) = S2 (t2 )−σ + S3 (t3 )−σ − 1 2σ2
S2 (t2 ) 2 S3 (t3 ) 2 .

The first equation reflects the fact that the frailties of the parents are assumed
to be uncorrelated. This results in independent event times. The second and

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 171

third function, respectively, are survival functions of the correlated gamma

frailty model with correlation ρ = 0.5 (see equation (5.6)). Korsgaard and
Andersen (1998) apply their model to data from the Danish adoptee register.
Each of the 792 families, as well as the father and mother of each family, are
assumed to be unrelated. It seems reasonable to assume that no environmental
correlations between lifetimes are present. No environmental correlations exist
between the child and its biological father and mother because the child is
adopted very soon after birth. It is therefore very likely that the parents did
not live together, and thus, it is reasonable to assume that the environmental
correlation is absent between the lifetimes of the parents. The (bivariate)
correlated gamma frailty model was used to perform a goodness-of-fit test of
the (trivariate) additive genetic gamma frailty model. Further extensions by
including genetic marker information can be found in a series of papers by Li
(1999, 2002), Li and Zhong (2002), Zhong and Li (2004), and Jonker et al.
(2009). A more general multivariate correlated gamma frailty is considered
at the end of this section.
It is necessary to keep in mind that parameter ρ describes the correlation
between the frailties in a pair and not the correlation of the related lifetimes.
Lindeboom and van den Berg (1994) analyzed the relationship of correlations
between frailties and between lifetimes. They derive explicit expressions for
the correlation between the survival times, and examined the properties of
this correlation in the special case of a constant baseline hazard function.
Unfortunately, in general situations, such explicit results are not available.
One possibility would be to simulate noncensored event-time data following
the estimated model and using it to estimate the correlation between the event
times directly.
Up to now only the case without observed covariates was considered. This
situation is common in demographic applications with large register data. In
many demographic studies, only information about gender and year of birth is
available. However, especially in medical applications, covariates are usually
available and need to be included into the model. This can be done in a natural
way similar to the univariate and shared frailty model. The correlated gamma
frailty model with observed covariates is a simple extension of (5.6):

σ1 σ2
S1 (t1 |X1 )1− σ2 ρ S2 (t2 |X2 )1− σ1 ρ
S(t1 , t2 |X1 , X2 ) = 2 2
ρ ,
(S1 (t1 |X1 )−σ1 + S2 (t2 |X2 )−σ2 − 1) σ1 σ2

where X1 and X2 are the covariate vectors of the partners in the pair. Again,
the marginal survival functions are assumed to be identical in the following
applications to twin data by S(t|X) = S1 (t|X) = S2 (t|X) and σ 2 = σ12 = σ22
yielding:

 ′
− 12
S(t|X) = 1 + σ 2 M0 (t)eβ X
σ
. (5.7)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
172 Frailty Models in Survival Analysis

Example 5.3
In the following we present an application of the bivariate correlated gamma
frailty model with observed covariates to the lifetimes of MZ and DZ Danish
twins with respect to coronary heart disease (CHD). We analyze the influence
of smoking, body mass index (BMI), gender, and birth year (using the variable
transformation birth year minus 1890 because the oldest twins are born in
1890) on the susceptibility to death by CHD.

Table 5.3: Correlated gamma frailty model with covariates

parameter Weibull exponential Gompertz

σ2 0.467 (0.196) 0.107 (0.032) 0.622 (0.214)

ρMZ 1.000 ( - ) 1.000 ( - ) 1.000 ( - )
ρDZ 0.884 (0.428) 0.889 (0.425) 0.689 (0.306)
λ 1.7e-20 (4.4e-20) 0.397 (0.213) 3.5e-6 (2.4e-6)
ν 10.362 (0.590)
ϕ 0.131 (0.008)
cigarette smokers 0.451 (0.144) 0.337 (0.150) 0.465 (0.149)
other smokers 0.167 (0.139) 0.142 (0.147) 0.184 (0.144)
former smokers 0.168 (0.147) 0.162 (0.154) 0.186 (0.152)
BMI < 22 0.213 (0.139) 0.155 (0.147) 0.228 (0.144)
BMI > 28 0.073 (0.122) 0.014 (0.133) 0.081 (0.127)
gender (female) -0.734 (0.120) -0.728 (0.149) -0.783 (0.126)
birthyear -0.022 (0.008) -0.136 (0.016) -0.022 (0.009)
likelihood -2733.98424 -2929.10475 -2729.89782

Parametric models with different cumulative baseline hazards M0 (t) in (5.7)

were fitted. The results are given in Table 5.3. The Gompertz model shows
the best fit to the data (based on the likelihood) compared to the other two
models and should be preferred. The correlation for MZ twins (ρMZ = 1) is
on the boundary of the parameter space, whereas DZ twins show a smaller
correlation (ρDZ = 0.689). Cigarette smokers have a significant higher risk
to die from CHD, with β = 0.465 (0.149), compared to nonsmokers. Other
smokers and former smokers show a small but nonsignificant risk increase
compared to nonsmokers. Twins with small or high BMI have a slightly worse
prognosis compared to twins in the reference category with BMI between 22
and 28 kg/m2 . Women show a significant better survival with β = −0.783
(0.126) compared to males, and later birth is related to a decreasing risk.

An interesting extension of the bivariate correlated gamma frailty model

is given in a recent paper by van den Berg et al. (2008). Here, information
about the economic situation at birth time and the seasonality of birth is
included into the analysis of lifetimes of Danish twins with special interest in
mortality caused by cardiovascular diseases.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 173

The bivariate correlated gamma frailty model can be extended to the multi-
variate case with p related lifespans, which results in the gamma-distributed
frailty case in the multivariate survival times
p p
Y
1−ρ
X 2
−ρ/σ2
S(t1 , . . . , tp ) = Si (ti ) Si (ti )−σ − p + 1 .
i=1 i=1

In this model, identical distributions are used for the frailties of the subjects
in a cluster, and the correlations between frailties in a cluster are unique and
given by ρ. Further extensions of this model with different frailty variances
can be found in Yashin and Iachine (1999a), but need additional requirements,
which are rather restrictive for real applications. A major disadvantage of the
model is that it becomes very complex with increasing cluster size. In the fore-
going twin example the likelihood consists of four terms caused by the different
possibilities for censoring plus one term for truncation (see Appendix (A.2)).
With cluster size three, the number of terms increases to nine; with cluster
size four, 17 terms have to be calculated. These terms include derivatives of
the survival function of the order of the cluster size. This limits multivariate
extensions of the correlated gamma frailty model to cluster sizes of three or
at most four. This inflexibility was the main reason why this model was
not considered further in the literature, for example, to model family data.
For such kind of applications (cluster size larger than two), the correlated
log-normal frailty model considered in the following chapter is much more
appropriate because of the flexibility of multivariate normal distribution.
In the applications of the correlated gamma frailty model presented here, a
parametric approach is used. Similar to shared frailty models, semiparametric
approaches are sometimes preferable to prevent assumptions about the form
of the baseline hazard function. For the bivariate correlated gamma frailty
model, an EM algorithm for a semiparametric approach was suggested by
Iachine (1995), but software is not publicly available for this model. Further
developments in this direction are restricted by the problems with extensions
of the bivariate correlated gamma frailty model to the larger cluster sizes
discussed earlier. The main problem in comparison with the shared frailty
model is that the number of frailties per cluster to be estimated in the EM
algorithm is determined by the number of observations in the cluster. This
causes severe problems with larger cluster size.
Parner (1998) extends results from Murphy (1994,1995) from the shared
gama frailty model and shows consistency and asymptotic normality of the
nonparametric maximum likelihood estimator in the multivariate correlated
gamma frailty model with observed covariates.
The correlated gamma frailty model can also be applied to current status
data. In the case of bivariate event times the likelihood function is given by
(4.7). Information about identifiability, consistency, and asymptotic normality
in the correlated gamma frailty model with current status data is given in the
paper by Chang et al. (2007). In the following results of a small simulation

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
174 Frailty Models in Survival Analysis

study are presented to check the performance of the estimation procedure.

One thousand samples of 1000 pairs each are generated using the bivariate
correlated gamma frailty model with frailty variance V(Z) = σ 2 , expectation
EZ = 1, and Gompertz baseline. We consider the two scenarios similar to
that of the shared gamma frailty model in Section 4.3.3.
Scenario 1: Both censoring/monitoring times c1 , c2 are independent.

Table 5.4: Simulation (Scenario 1) of correlated

gamma frailty (ρ = 0.5) with current status data
parameter true current status right censored

σ2 1.000 1.442 (1.332) 1.030 (0.379)

ρ 0.500 0.636 (0.290) 0.596 (0.241)
λ1 0.100 0.105 (0.136) 0.099 (0.008)
ϕ1 0.120 0.436 (0.990) 0.143 (0.103)
λ2 0.100 0.112 (0.134) 0.100 (0.008)
ϕ2 0.120 0.435 (0.981) 0.143 (0.102)

Each data set is analyzed under two different censoring mechanisms – once
as right-censored lifetimes, and once as current status data. The correlated
gamma frailty model is much more complex than the shared gamma frailty
model because of the additional correlation parameter ρ in the model. In this
model the parameter estimates based on right-censored data are much closer
to the true parameter values and with much smaller standard errors compared
to the estimates based on the current status data. Especially, the parameter
estimates for ϕ1 , ϕ2 are strongly biased upward with the current status data,
as are the estimates of σ 2 and ρ. Consequently, switching from right-censored
lifetime data to current status data causes an important information loss.
However, further simulations with increasing sample size (results are not
shown) imply that parameters can be estimated consistently based on event-
time data as well as current status data.
Scenario 2: Both censoring/monitoring times are equal.
We consider this situation in detail here because it holds for the real data
example considered later. Correlation coefficients of ρ = 0.25, 0.5, 0.75, and 1
are simulated. The results are similar to that of the situation of independent
censoring. For the current status data, the upwards bias in the parameter
estimates of ϕ1 , ϕ2 , σ 2 is more pronounced; also, the standard errors are larger.
Obviously, the upwards bias becomes less pronounced for larger values of the
correlation ρ (see Tables 5.5-5.8). The correlated gamma frailty model with
parametric baseline hazard is identifiable with current status data and gives
better results with larger sample size (results are not shown).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 175

Table 5.5: Simulation Scenario 2 with ρ = 0.25

parameter true current status right censored

σ2 1.000 1.904 (2.140) 1.249 (0.978)

ρ 0.250 0.378 (0.326) 0.382 (0.309)
λ1 0.100 0.168 (0.882) 0.099 (0.012)
ϕ1 0.120 0.912 (2.175) 0.286 (0.724)
λ2 0.100 0.120 (0.465) 0.100 (0.023)
ϕ2 0.120 0.918 (2.208) 0.285 (0.741)

Table 5.6: Simulation Scenario 2 with ρ = 0.5

parameter true current status right censored

σ2 1.000 1.489 (1.592) 1.020 (0.346)

ρ 0.500 0.606 (0.292) 0.585 (0.230)
λ1 0.100 0.108 (0.141) 0.100 (0.008)
ϕ1 0.120 0.531 (1.799) 0.138 (0.091)
λ2 0.100 0.100 (0.052) 0.100 (0.008)
ϕ2 0.120 0.539 (1.865) 0.138 (0.092)

Table 5.7: Simulation Scenario 2 with ρ = 0.75

parameter true current status right censored

σ2 1.000 1.228 (0.805) 1.023 (0.201)

ρ 0.750 0.764 (0.212) 0.769 (0.156)
λ1 0.100 0.102 (0.078) 0.100 (0.008)
ϕ1 0.120 0.252 (0.623) 0.131 (0.050)
λ2 0.100 0.101 (0.038) 0.100 (0.008)
ϕ2 0.120 0.250 (0.600) 0.131 (0.050)

Table 5.8: Simulation Scenario 2 with ρ = 1

parameter true current status right censored

σ2 1.000 1.164 (0.320) 1.061 (0.099)

ρ 1.000 0.916 (0.110) 0.950 (0.071)
λ1 0.100 0.099 (0.016) 0.100 (0.008)
ϕ1 0.120 0.173 (0.185) 0.136 (0.027)
λ2 0.100 0.102 (0.127) 0.099 (0.008)
ϕ2 0.120 0.174 (0.173) 0.136 (0.027)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
176 Frailty Models in Survival Analysis

Example 5.4
Here, the hepatitis A and B current status data from Example 1.7 are analyzed
in detail. In the second column of Table 5.9, a correlated gamma frailty model
with different frailty variances σ12 and σ22 for hepatitis A and B, respectively,
is presented (correlated I). In the third column, a model with equal variances
σ12 = σ22 is used (correlated II). The fourth column contains a shared gamma
frailty model (ρ = 1), and column five contains the univariate frailty model
(ρ = 0). There is no reason to assume per se that the frailty variance with

Table 5.9: Analysis of hepatitis A and B current status data with the
correlated gamma frailty model
parameter correlated I correlated II shared univariate

σ12 2.668 (0.572) 2.600 (0.553) 0.522 (0.028) 2.018 (0.510)

σ22 1.472 (2.528) 2.600 (0.553) 0.522 (0.028) 10.037 (8.821)
ρ 0.653 (1.938) 0.485 (0.038) 1.000 ( – ) 0.000 ( – )
λ1 0.007 (0.001) 0.007 (0.001) 0.012 (0.001) 0.008 (0.001)
ϕ1 0.105 (0.017) 0.103 (0.016) 0.037 (0.002) 0.085 (0.016)
λ2 0.002 (0.001) 0.002 (0.001) 0.002 (0.001) 0.002 (0.001)
ϕ2 0.002 (0.032) 0.002 (0.007) -0.001 (0.007) 0.015 (0.142)
log-likelihood -2826.7159 -2826.7918 -2843.5100 -2841.4151

respect to hepatitis A is similar to the frailty with respect to hepatitis B.

That is why a correlated frailty model with different variances is applied,
but the model with equal variances is supported by the log-likelihood. A
further simplification of the model to the shared gamma frailty model implies
a significant worsening in the fit (p < 0.0001). The univariate frailty models
treats time-to-infection by hepatitis A and B, respectively, as independent. It
allows for different variances and shows a log-likelihood similar to the one in
the shared frailty model. A Gompertz baseline hazard was used in all models.
Besides the good fit of the correlated gamma frailty model compared to
the shared frailty model, it provides a nice interpretation of the parameters.
Here, σ12 and σ22 are measures of population heterogeneity in the susceptibility
to hepatitis A and B, respectively. These parameters yield implications for
further programs to prevent infections, as the critical vaccination coverage
is higher for more heterogeneous populations. Furthermore, the parameter
ρ provides a correlation measure. The analysis of bivariate infection data
allows study of the association between the acquisition of both infections.
The correlation indicates similar transmission routes (correlation near one) or
reflects to what extent a latent process, such as the social or hygienic behavior
of people, drives the more general infection process (Hens et al. 2009).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 177

5.3 Correlated Log-normal Frailty Model

The correlated log-normal frailty model is much more flexible than the gamma
model because it is not based on the additive composition of the frailties as
used in (5.3) and (5.4). However, the log-normal distribution does not allow us
to integrate out the frailties with an explicit representation of the likelihood.
This requires more sophisticated estimation strategies. The distribution can
be obtained by assuming a bivariate normal distribution on the logarithm of
the frailty vector,
   2
s rs2
    
ln Z1 W1 m
= ∼N , ,
ln Z2 W2 m rs2 s2
with N denoting the bivariate normal distribution whose parameters are some
functions of the frailty parameters σ 2 and ρ, similar to the univariate log-
normal frailty model in (3.36) and (3.37):

s2
µ = EZj = em+ 2

2 2
σ 2 = V(Zj ) = e2m+s (es − 1)
2
ers − 1
ρ = corr(Z1 , Z2 ) = s2 .
e −1
Again, the restriction m = 0 is used to guarantee identifiability of the model
parameters. This means that the random effects Wj (j = 1, 2) (log frailties)
have an expectation of zero.
Xue and Brookmeyer (1996) were the first authors to consider the correlated
log-normal frailty model and applied it to mental health data to evaluate
the health policy effects for inpatient psychiatric care. Yau and McGilchrist
(1997) used a generalized linear mixed model approach to analyze data from
litter-matched tumorigenesis experiments in rats with help of the correlated
log-normal model. Cook et al. (1999) used a correlated log-normal frailty
in two-state mixed renewal processes for chronic diseases. Other examples
can be found in Ripatti and Palmgren (2000) and Ripatti et al. (2002).
Based on correlated log-normal frailty models, Pankratz et al. (2005) perform
genetic analysis of age at onset in breast cancer in a large familial cohort.
Their method is based on a Laplace approximation similar to the approach
by Ripatti and Palmgren (2000).

Example 5.5
The correlated log-normal frailty model is applied to the mortality of Danish
twins. The duration of interest is lifetime with respect to CHD; that means
lifetimes with respect to other causes of death are considered as censoring

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
178 Frailty Models in Survival Analysis
Table 5.10: Correlated log-normal frailty model with
covariates
parameter without covariates with covariates

s2 6.520 (3.263) 2.394 (1.210)

ρMZ 0.509 (0.108) 0.575 (0.190)
ρDZ 0.359 (0.097) 0.426 (0.141)
λ 2.8e-11 (1.1e-10) 9.6e-9 (1.9e-8)
ϕ 0.249 (0.045) 0.195 (0.024)
cigarette smokers 0.556 (0.192)
other smokers 0.265 (0.182)
former smokers 0.247 (0.192)
BMI < 22 0.312 (0.184)
BMI > 28 0.149 (0.161)
gender (female) -0.966 (0.194)
birthyear 0.006 (0.010)
log-likelihood -2797.45474 -2761.61131

times. The results in Table 5.10 are difficult to compare with Example 5.3,
where a correlated gamma frailty model was applied. The variances σ 2 of the
gamma and s2 of the log-normal model are not directly comparable. The same
problem holds for the correlations in both models. The analysis indicates a
nonsignificant mortality increase for twins with BMI less than 22 kg/m2 with
β = 0.312 (0.184) and BMI more than 28 kg/m2 with β = 0.149 (0.161),
respectively. The reference group are twins with BMI between 22 kg/m2 and
28 kg/m2 . Cigarette smoking shows a significant influence on CHD mortality
with β = 0.556 (0.192), whereas pipe/cigar smokers β = 0.32 (0.23) and former
smokers β = 0.48 (0.26) experience only a nonsignificant increase in risk. The
reference group are the nonsmokers. Females have a significant better survival
β = −0.966 (0.194) compared to males. The birth year shows no significant
association with CHD death in this sample. Analysis was performed with
SAS NLMIXED.
In twin studies, the main focus is on the correlations between MZ and
DZ twins and not on the covariates. The effect of covariates may be better
evaluated in studies of singletons. Correlations are of special interest for
the quantification of the influence of genetic and environmental factors. A
comparison between models without and with covariates allows conclusions
about the nature of the covariates (genetic vs. environment). The inclusion
of the covariates reduces the frailty variance. This is expected because these
covariates are included in the frailty in the model without covariates. By
explicit modeling as observed covariates in the analysis, these covariates are no
longer included in the frailty (as a proxy for unobserved covariates). It implies
the reduction in the frailty variance from 6.520 (3.263) to 2.394 (1.210). In
contrast, the correlations change only slightly between the two models. This
point is further discussed in the next example.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 179

Example 5.6
In this example, an extension of the correlated log-normal frailty model is
presented, which allows the integration of genetic models (see Appendix A.6).
In this approach the correlations between family members are substituted by
the respective variance components. In the case of MZ and DZ twins relation
(A.8) can be used. In an ACE model including additive genetic factors (A),
common environment (C), and unique environment (E), the corresponding
variance components are denoted by a2 , c2 , and e2 . This reparameterization
is used to analyze the data from Example 5.3 again. The main interest is
in the estimation of the heritability (a2 ), which quantifies the importance of
genetic factors on a trait. Here, the trait of interest is lifetime with respect to
CHD. Again, a Gompertz baseline hazard is used (semiparametric approach).

Table 5.11: Correlated log-normal frailty model with

covariates
parameter without covariates with covariates

s2 7.875 (3.505) 2.500 (1.182)

a2 0.295 (0.261) 0.464 (0.365)
c2 0.223 (0.198) 0.151 (0.276)
e2 0.482 (0.097) 0.384 (0.171)
λ 5.9e-12 (2.3e-11) 8.5e-9 (1.6e-8)
ϕ 0.267 (0.045) 0.197 (0.023)
cigarette smokers 0.595 (0.196)
other smokers 0.249 (0.183)
former smokers 0.247 (0.193)
BMI < 22 0.257 (0.184)
BMI > 28 0.146 (0.162)
gender (female) -0.996 (0.194)
birthyear 0.006 (0.010)
log-likelihood -2797.41765 -2761.36539

Former studies from the Danish (Harvald and Hauge 1970) and Swedish (de
Faire 1975, Marenberg et al. 1994) twin registries found a genetic component
in the susceptibility to death from coronary heart disease. The advantage
of the approach presented here is the combination of methods from survival
analysis with methods from genetic epidemiology. This allows the estimation
of the effect of genetic factors in susceptibility to CHD and the evaluation as
to what extent smoking, BMI, and susceptibility to CHD are all influenced by
common genetic factors. For each individual, two independent, underlying,
competing risks of latent times – lifetime with respect to death due to CHD
and lifetime with respect to death due to all other causes including censoring
events – are assumed.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
180 Frailty Models in Survival Analysis

Both, smoking and BMI, are influenced by genes with heritability estimates
in the range 0.35 − 0.75 (smoking) and 0.5 − 0.8 (BMI) (Bouchard 1994,
Heath and Madden 1995, Herskind et al. 1996b). However, whether common
genes influence these phenotypic traits as well as susceptibility to CHD is
an open question. In a different situation Fisher (1958) suggested that the
association between smoking and lung cancer is spurious and reflects only the
circumstance that the same genes influence both smoking habits and lung
cancer. This was the starting point for a long debate on genetic confounding.
The main result of the present analysis is that the inclusion of smoking
and BMI do not cause any substantial decrease in the heritability estimate.
Consequently, no evidence was found for common genetic factors acting on
smoking and susceptibility to CHD or BMI, and susceptibility to CHD. This
study confirms the earlier finding that the genetic influence on susceptibility to
CHD is not mediated through genetic influence on smoking and BMI. Similar
results were found by Zdravkovic et al. (2004) in Swedish twins. The approach
here is different from the common study design in many medical applications.
Usually the investigator is interested in the effect of covariates on the outcome,
adjusted for the correlation in clustered observations, treating the correlation
as a nuisance parameter. In genetic epidemiology, the main interest is in the
correlations, adjusted for the effect of covariates.
As expected, the inclusion of covariates decreases the heterogeneity in the
study population, which can be seen in the decline of the frailty variance from
s2 = 7.875 (3.505) to s2 = 2.500 (1.182). This underlines that frailty depends
on the model, and it describes factors not included in the model.
When covariates are included in the model, the relative importance of
environmental factors is reduced, leading to an increase in the heritability
estimate observed in the present analysis. This is a result of the heritability
coefficient as a variance proportion. The variance of the random effects can
be decomposed as follows: s2 = s2genes +s2environment . Focusing on heritability
only one does not know whether the heritability increases due to an increase
in the genetic variance or due to a decrease in the environmental variance.
In the present case, the increase was largely due to the latter. By including
smoking and BMI, the genetic variance was reduced from s2genes = 2.32 to
s2genes = 1.16. The reduction of the environmental heterogeneity is more
pronounced, for example, from s2environment = 5.55 to s2environment = 1.34.
This result underlines that both factors primarily represent environmental
sources of variation for age at CHD death, despite the role that genetics
may play for the specific factors. A comprehensive discussion of variance
components can be found in Hopper (1993). Genetic confounding would lead
to a decrease in heritability estimates after the inclusion of observed covariates
because, in that case, genetic factors contribute predominantly to the observed
covariates rather than to the unobserved covariates included in the frailty.
Our results are similar to the results obtained in a study of Swedish twins
(Zdravkovic et al. 2004). More details in a slightly different analysis of the
same data are given in Wienke et al. (2005a).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 181

5.4 MCMC Methods for the Correlated Log-normal

Frailty Model
Any distribution of nonnegative random variables can be adapted to model
frailty. The gamma distribution has been widely applied in different fields
(Sections 3.3 and 5.2). The gamma distribution is a very convenient choice
from a mathematical point of view because of the simplicity of the Laplace
transform, which allows for the use of maximum likelihood procedures in
the parameter estimation. Another possibility is to assume that frailty is
log-normally distributed (Sections 3.4 and 5.3). The log-normal approach is
much more flexible than the gamma model in creating correlated but distinct
frailties as required in the correlated frailty model. Unfortunately, with a log-
normal assumption, it is impossible to derive the marginal likelihood function
in an explicit form, and parameter estimation has to be performed with the
help of more sophisticated estimation strategies.
There are four main methods for parameter estimation in frailty models:
maximum likelihood methods (only applicable in parametric models), the
EM algorithm, penalized partial likelihood, and Markov Chain Monte Carlo
(MCMC) methods. Yashin et al. (1995) and Wienke et al. (2003a,b) applied
procedures based on maximum likelihood methods in the gamma context,
where an explicit representation of the likelihood function is available. The
maximum likelihood procedure has also been adopted in log-normal frailty
models by different numerical approximation algorithms (McGilchrist and
Aisbett 1991, McGilchrist 1993, Lillard 1993, Lillard et al. 1995, Ripatti
and Palmgren 2000, Ripatti et al. 2002). For example, such routines are
implemented in the aML software package (Lillard and Panis 2000) and in
newer versions of SAS.
In the present section, based on Locatelli (2003) and Locatelli et al. (2004),
we consider an example of how to apply MCMC methods to the parametric
bivariate correlated log-normal frailty model with Gompertz baseline hazard,
as described in Section 5.3, with parameterization EZ = 1 and σ12 = σ22 = σ 2
because of the symmetry of the twin data, which is used as an illustrative
example here. A similar analysis can be found in Kheiri et al. (2005), who
use a piecewise constant baseline hazard function.
Bayesian MCMC methods have been developed by Clayton (1991), and
based on this work, implemented in WinBUGS by Spiegelhalter et al. (1996).
Further applications of Bayesian methods, especially in shared frailty models
are given by Sahu et al. (1997) and Sinha and Dey (1997) but also in the case
of correlated frailty models by Xue and Ding (1999). The Bayesian approach
is in fact natural when dealing with conditionally independent observations
and when working with hierarchical models, with the frailty variables at an
intermediate stage between the observations and the hyperparameters. In the
Bayesian context, the frailty distribution represents a ”prior” of the model,

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
182 Frailty Models in Survival Analysis

and its parameters (hyperparameters) are considered as random variables

following some noninformative distribution, meaning a distribution with huge
variance. The MCMC method consists of generating a set of Markov chains
whose joint stationary distribution corresponds to the joint posterior of the
model, this one being, in the Bayesian framework, the distribution of the
random parameters given the observed data. In most hierarchical models,
the posterior distribution is often very difficult to work with and hardly
ever possible to integrate out in order to find the marginal posterior of each
of the random parameters. By MCMC methods we can circumvent this
problem. The posterior of each parameter is approximated by the empirical
distribution of the values of the corresponding Markov chain, and empirical
summary statistics calculated along each chain can be used to make inferences
about the true value of the corresponding parameter (see Gilks et al. 1996).
The Gibbs sampling (Geman and Geman 1984) is one of the algorithms
that have been created in order to obtain Markov chains with the desired
stationary distribution. The basic idea behind the Gibbs sampling is to
successively sample from the conditional distribution of each random node,
given all the other nodes in the model. These distributions are known as
”full conditional distributions”. It can be shown that, under broad conditions,
this process eventually provides samples from the joint posterior distribution
of the unknown quantities.
In this section, MCMC methods have been adopted to estimate correlated
log-normal frailty models. Calculations are performed within the software
WinBUGS (Spiegelhalter et al. 1999). This is a package that enables the
researcher to solve Bayesian hierarchical models, essentially using the Gibbs
sampling algorithm. The correlated log-normal frailty model applied here can
be represented as a Bayesian hierarchical three–level model in the following
way:

1. Likelihood function:
n Y
2
Y Wij λ ϕtij
L(λ, ϕ|W) = (eWij λeϕtij )δij e−e ϕ (e −1)

i=1 j=1

2. Priors:

− 21 ln(σ 2 + 1) ln(σ 2 + 1) ln(ρσ 2 + 1)

     
Wi1
(i) ∼N ,
Wi2 − 21 ln(σ 2 + 1) ln(ρσ 2 + 1) ln(σ 2 + 1)
(ii) λ ∼ Γ(0.01, 0.01)
(iii) ϕ ∼ Γ(0.01, 0.01)

3. Hyperpriors:

(i) σ 2 ∼ Γ(0.01, 0.01) (ii) ρ ∼ U (−1, 1)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 183

with W = (W11 , W12 , W21 , W22 , ..., Wn1 , Wn2 ). Γ and U denote the gamma
and uniform distribution, respectively, and λ and ϕ are parameters of the
Gompertz baseline hazard. The prior (i), assigned to the vector (Wi1 , Wi2 ), is
chosen in order to have a vector of log-normal frailties (Zi1 , Zi2 ) = (eWi1 , eWi2 )
with mean equal to one:
   2
σ ρσ 2
  
Zi1 1
∼ ln N , .
Zi2 1 ρσ 2 σ 2

Finally, noninformative priors are assigned to the parameters of the Gompertz

curve and to the frailty parameters. The full conditional distributions can be
obtained because they are proportional to the joint distribution of all the
random quantities of the model. In our case, this joint distribution takes the
form
n hY
Y 2 i
π(t, δ, W, λ, ϕ, σ 2 , ρ) = L(λ, ϕ|W) π(Wij |σ 2 , ρ) π(λ)π(ϕ)π(σ 2 )π(ρ),
i=1 j=1

where π(·) indicates the density function of the corresponding argument.

Gibbs sampling is an algorithm to generate a sequence of samples from the
joint probability distribution of two or more random variables. The purpose
of such a sequence is to approximate the unknown joint distribution. The
Gibbs sampling algorithm generates an instance from the distribution of each
variable in turn, conditional on the current values of the other variables. It
can be shown that the sequence of samples constitutes a Markov chain, and
the stationary distribution of that Markov chain is just the sought-after joint
distribution.
A slice-sampler algorithm is used in WinBUGS for nonlog-concave densities
defined on a restricted range (Neal 1997). It has an adaptive phase of 500
iterations, which are discarded from all summary statistics. A Metropolis
within Gibbs algorithm based on a symmetric normal proposal distribution
is applied in the case of nonlog-concave densities defined on an unrestricted
range (Metropolis et al. 1953, Hastings 1970, Besag and Green 1993).

Example 5.7
The results of applying the correlated log-normal frailty model to the Swedish
breast cancer data using MCMC methods as described before are presented
in Table 5.12. The study population consists of 12,568 female twin pairs
of the old and middle cohort. Estimated parameters include the Gompertz
parameters λ and ϕ and the variance of the frailty distribution σ 2 , which
can be seen as the extent of population heterogeneity with respect to age at
onset of breast cancer, and estimates of the correlation coefficient for both
MZ twins (ρMZ ) and DZ twins (ρDZ ). Two estimates for each parameter
are given in terms of the mean and the median of the correspondent Markov
chain. In all cases, both values are very close to each other. This means

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
184 Frailty Models in Survival Analysis
Table 5.12: Correlated log-normal frailty model
parameter mean median sdv MC error CSRF

σ2 45.19 41.50 17.05 0.824 1.055

ρMZ 0.311 0.299 0.046 0.005 1.008
ρDZ 0.104 0.097 0.108 0.002 1.005
λ 2.54e-5 2.52e-5 3.24e-6 7.92e-8 1.002
ϕ 0.072 0.072 0.003 8.94e-5 1.006

that empirical estimates of the marginal posterior densities (kernel density

estimates) are approximately symmetric. For each parameter, the sample
standard deviation and an estimate of the standard error of the mean are also
given. This one is obtained following Roberts’ (1996) batch means method.
In the last column, the value of the corrected scale reduction factor (CSRF)
for each parameter is reported. This value corresponds to the Gelman–Rubin
convergence statistic (Gelman and Rubin 1992), as modified by Brooks and
Gelman (1998), and is based on a comparison of the inter- and intra-chain
variance for each variable. When values of this diagnostic are approximately
equal to one, the sample must have arisen from the stationary distribution. In
this case, descriptive statistics are valid estimates of the unknown parameters.
According to the foregoing model, the population is largely heterogenous
(σ 2 ) in terms of susceptibility toward breast cancer. The estimated correlation
between frailties is larger for MZ than for DZ twins. This means that MZ
individuals who genetically are more similar than DZ twins also present a
larger correlation in terms of frailty toward breast cancer. This finding suggests
that there is a genetic influence on breast cancer propensity.

The WinBUGS package proved to be useful and flexible enough to estimate

correlated frailty models. With this software it is easy to modify the hypothesis
on frailty distribution. Different assumptions about frailty distribution and
the shape of the baseline hazard can be compared using an information
criterion (DIC) which was introduced by Spiegelhalter et al. (2002) in order
to compare Bayesian models in terms of adequacy and complexity.
A disadvantage of Bayesian methods is the time required for parameter
estimation. In fact, we are working with models that include a very large
number of parameters, especially when dealing with large data sets. This
means that every MCMC algorithm that updates parameters one by one (like
the Gibbs Sampling used in WinBUGS) is very time consuming. To solve this
problem, an algorithm that enables the updating of parameters all together
(or groups of parameters) at the same time should be adopted. The MCMC
procedure in SAS 9.2. opens new avenues in this direction and provides a very
user-friendly environment in the SAS package. This makes MCMC methods
easier available for a large community of researchers, who are not familiar
with the very specific WinBUGS package.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 185

5.5 Correlated Compound Poisson Frailty Model

The model is an extension of the correlated PVF frailty model (sometimes
called three-parameter frailty model) suggested by Yashin et al. (1999a) and
Iachine (2002). It is based on a bivariate extension of the compound Poisson
frailty model introduced to univariate survival analysis by Aalen (1988, 1992).
It is also related to the correlated gamma frailty cure model in Section 5.8,
which allows for a nonsusceptible fraction in the study population, as well
as the compound Poisson frailty models with a random scale considered by
Moger et al. (2004b) and Moger and Aalen (2005).
Let be k0 , k1 nonnegative variables, and Y0 , Y1 , Y2 independently compound
Poisson-distributed random variables with distributions Y0 ∼ cP (γ, k0 , λ),
Y1 ∼ cP (γ, k1 , λ), and Y2 ∼ cP (γ, k1 , λ). Consequently, using an additive
structure for the frailties similar to (5.3), it holds that

Z1 = Y0 + Y1 ∼ cP (γ, k0 + k1 , λ)

Z2 = Y0 + Y2 ∼ cP (γ, k0 + k1 , λ).
For simplicity, we shall present only the symmetric case here, where the two
lifetimes are interchangeable. Consequently, the following assumptions are
made in the model:

EZ1 = EZ2 = 1, V(Z1 ) = V(Z2 ) = σ 2 .

This implies (compare relations (3.51) and (3.52)) that (k0 + k1 )λγ−1 = 1
and (k0 + k1 )(1 − γ)λγ−2 = σ 2 . Consequently, (k0 + k1 )λγ−2 = 1/λ, and
σ2
(k0 + k1 )λγ−2 = 1−γ = λ1 . Hence, λ = 1−γ
σ2 , which results in

1−γ
(k0 + k1 )λγ = λ = . (5.8)
σ2
It holds

cov(Z1 , Z2 ) = cov(Y0 + Y1 , Y0 + Y2 ) = V(Y0 ) = k0 (1 − γ)λγ−2 .

This leads to the correlation

cov(Z1 , Z2 )
ρ= p
V(Z1 )V(Z2 )
k0 (1 − γ)λγ−2
=
(k0 + k1 )(1 − γ)λγ−2
k0
= . (5.9)
k0 + k1

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
186 Frailty Models in Survival Analysis

Consequently, because of (5.8) and (5.9),

k0 1−γ
k0 λγ = (k0 + k1 )λγ = ρ 2 . (5.10)
k0 + k1 σ
Now we can derive the unconditional model, applying the Laplace transform
of compound Poisson-distributed random variables (3.53). Hence,

S(t1 , t2 ) = ES(t1 , t2 |Z1 , Z2 )

= ES(t1 |Z1 )S(t2 |Z2 )
= Ee−Z1 M0 (t1 ) e−Z2 M0 (t2 )
= Ee−(Y0 +Y1 )M0 (t1 ) e−(Y0 +Y2 )M0 (t2 )
= Ee−Y0 (M0 (t1 )+M0 (t2 ))−Y1 M0 (t1 )−Y2 M0 (t2 )
k0 γ γ
= e− γ ((λ+M0 (t1 )+M0 (t2 )) −λ )
k1 k1
((λ+M0 (t1 ))γ −λγ ) − ((λ+M0 (t2 ))γ −λγ )
× e− γ e γ . (5.11)

The three terms are considered in detail. For the marginal survival function,
it holds that
k0 +k1
((λ+M0 (t))γ −λγ )
S(t) = e− γ , (5.12)

which implies

γ
λ + M0 (t) = (λγ − ln S(t))1/γ . (5.13)
k0 + k1
Hence, using (5.9) and (5.12),

k1 k1 k0 +k1
((λ+M0 (t))γ −λγ ) ((λ+M0 (t))γ −λγ )
e− γ = e− k0 +k1 γ
k +k γ γ
−(1−ρ) 0 γ 1
((λ+M0 (t)) −λ )
=e
k0 +k1 γ γ
1−ρ
= e− γ ((λ+M0 (t)) −λ )
= S(t)1−ρ .

The first of the three terms in (5.11) can be rewritten because of (5.13):

k0
((λ+M0 (t1 )+M0 (t2 ))γ −λγ )
e− γ
k0
((λ+M0 (t1 )+λ+M0 (t2 )−λ)γ −λγ )
= e− γ
k0 γ γ
(((λγ − k ln S(t1 ))1/γ +(λγ − k ln S(t2 ))1/γ −λ)γ −λγ )
= e− γ 0 +k1 0 +k1
k0 λγ γ γ
(1−((1− (k γ ln S(t1 ))1/γ +(1− (k γ ln S(t2 ))1/γ −1)γ )
=e γ 0 +k1 )λ 0 +k1 )λ ,

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 187

which results because of (5.8) and (5.10) in the following representation of the
correlated model

S(t1 , t2 ) = S(t1 )1−ρ S(t2 )1−ρ (5.14)

2 2
 ρ(1 − γ)  γσ 1
γ + (1 −
γσ 1
γ − 1

γ 
× exp 1 − (1 − ln S(t 1 )) ln S(t 2 )) .
γσ 2 1−γ 1−γ
This model includes both the gamma and the inverse Gaussian correlated
frailty model as a special case; and the gamma model corresponds to γ = 0,
and the inverse Gaussian model has γ = 0.5. The correlated inverse Gaussian
model was considered by Zahl (1994) analyzing excess hazards of cancer
patients, and by Kheiri et al. (2007) using a piecewise constant baseline
hazard and a Bayesian approach. For γ ≥ 0, the correlated PVF frailty model
is obtained, which was introduced by Yashin et al. (1999a). For γ < 0, the
correlated compound Poisson frailty model is obtained, including a fraction
of individuals in the population with zero frailty.

Example 5.8
The model is applied to breast cancer incidence of MZ and DZ female Swedish
twins born 1886 – 1925 (Wienke et al. 2006a, 2010), described in detail in
Example 1.6. A parametric model with Gompertz baseline hazard function
is used with parameters λ and ϕ. The data is left truncated, because only
breast cancer cases are included that occurred after the foundation of the
cancer register. This truncation is adjusted for in the analysis. The results
are given in Table 5.13.

Table 5.13: Analysis of time to breast cancer in Swedish twins

parameter gamma inverse Gaussian compound Poisson

σ2 32.78 (7.776) 17.62 (14.22) 15.94 (8.273)

ρMZ 0.154 (0.052) 0.343 (0.131) 0.154 (0.053)
ρDZ 0.126 (0.040) 0.299 (0.106) 0.130 (0.041)
γ 0( - ) 0.5 ( - ) -0.617 (0.905)
λ 1.3e-5 (1.0e-5) 1.8e-4 (5.7e-5) 2.5e-5 (1.9e-5)
ϕ 0.099 (0.016) 0.047 (0.008) 0.080 (0.016)
susceptible 100% 100% 15.2%
log-likelihood -5122.31549 -5130.58558 -5121.23195

The model in the first column is the correlated gamma frailty model (5.2).
Parameter σ 2 is a measure of heterogeneity, which is estimated to be large in
this data, and all individuals of the population are assumed to be susceptible
to breast cancer. The correlated gamma frailty model is a special case of the
correlated compound Poisson/PVF frailty model with γ = 0.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
188 Frailty Models in Survival Analysis

In the second model, inverse Gaussian distributed frailty is used to account

for heterogeneity in the population and to model dependence between times
to onset of breast cancer in twin pairs. Parameter estimates are different from
those in the gamma case with smaller heterogeneity and larger correlations.
As in the gamma model, all women are assumed to be susceptible to breast
cancer. The correlated inverse Gaussian frailty model is a special case of the
correlated compound Poisson/PVF frailty model with γ = 0.5.
The most interesting parameter in the compound Poisson model is γ, which
is negative γ = −0.617 (0.905) and indicates the existence of a fraction
of individuals nonsusceptible to breast cancer. The size of the susceptible
1−γ
fraction is calculated by 1 − e γσ2 and is around 15.2%. This estimate is
slightly smaller than the value of 22% found by Chatterjee and Shih (2001) in
a study population that is completely different, and in Section 5.8 applying
a correlated gamma frailty cure model to the same data set. The compound
Poisson model fits the data better than the two submodels but at the cost of
an additional parameter. Compared to the gamma model, the improvement
is not significant based on the likelihood ratio test.
Additionally, the size of the susceptible fraction in the correlated compound
Poisson frailty model is not that far from the range of the figures obtained
by Farewell et al. (1977) for different combinations of four risk factors. The
authors found that, if none of the risk factors is present, the fraction of females
susceptible to breast cancer is around 1.5%. If all risk factors are present, the
estimate increases to 27.2%.
In all models, correlations in MZ pairs are higher than in DZ pairs, but
the differences are small. This is in line with the well-known fact that the
influence of genetic factors on susceptibility to breast cancer is small (5–10%).
A disadvantage of the compound Poisson/PVF model is that it is not able to
handle negative dependencies.
The correlated compound Poisson/PVF frailty model offers a very elegant
approach to integrating the concept of cure models into frailty modeling. The
likelihood function is explicitly available in a very simple form, which is the
most important advantage of the model compared to the model suggested
by Moger and Aalen (2005). Popular frailty models such as the correlated
and shared gamma model (γ = 0) and inverse Gaussian model (γ = 0.5) are
included in this model family, which provides a great flexibility to the model.
Because of the extension to negative values of γ, the gamma distribution
(γ = 0) as one of the most popular frailty distributions is no longer on the
border of the parameter space. Consequently, standard tests can be applied to
test hypotheses about the frailty distribution (for example, H0 : γ = 0 versus
HA : γ 6= 0). Simulation studies show a good performance of the parameter
estimates in this model with nearly no bias.

Another bivariate compound Poisson frailty model with Weibull baseline

hazard functions is suggested by Hanagal (2010).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 189

5.6 Correlated Quadratic Hazard Frailty Model

As already shown in the foregoing sections, the correlated frailty model is a
powerful tool in the analysis of bivariate survival data. The unconditional
survival function has an explicit form in frailty models based on the gamma
or the more general compound Poisson distribution. Otherwise, multivariate
event time models with gamma or compound Poisson distributed frailty are
less flexible in specifying multivariate correlations needed in the analysis of
family data compared to frailty models based on the (log)normal distribution.
To join the flexibility of the multivariate normal distribution in the analysis of
event time data on related individuals with the benefits of frailty distributions
allowing for an explicit representation of the unconditional survival and finally
the likelihood function, the correlated quadratic hazard model considered by
Aalen (1987), Iachine and Yashin (1999), and Iachine (2002) can be used. The
conditional bivariate survival function in this model is given by

S(t1 , t2 |Z1 , Z2 ) = S(t1 |Z1 )S(t2 |Z2 )

= e−Z1 M0 (t1 ) e−Z2 M0 (t2 )
2 2
= e−W1 M0 (t1 ) e−W2 M0 (t2 ) ,
with Zj = Wj2 (j = 1, 2). A bivariate normal distribution of the random effects
(W1 , W2 ) with parameters m1 , m2 , s21 , s22 , and r is assumed. The unconditional
bivariate survival function is

S(t1 , t2 ) = ES(t1 , t2 |Z1 , Z2 )

1
ZZ
2 2
= √ e−w1 M0 (t1 ) e−w2 M0 (t2 )
2πs1 s2 1 − r 2
(w1 −m1 )2 (w −m )2
(w1 −m1 )(w2 −m2 )


1
− (1−r 2) −r + 2 22
2s2 s1 s2
×e 1
2s
2 dw1 dw2 .
It turns out that this integral has an explicit solution but requires some
lengthy calculations. Therefore, the calculation of the unconditional survival
function has been moved to the Appendix A.4. The survival function is

S(t1 , t2 ) (5.15)
1
=p
1 + 2s21 M0 (t1 ) + 2s22 M0 (t2 ) + 4(1 − r2 )s21 s22 M0 (t1 )M0 (t2 )
n m21 M0 (t1 )(1 + 2s22 M0 (t2 )) + m22 M0 (t2 )(1 + 2s21 M0 (t1 )) o
× exp −
1 + 2s21 M0 (t1 ) + 2s22 M0 (t2 ) + 4(1 − r2 )s21 s22 M0 (t1 )M0 (t2 )
n 4rm1 m2 s1 s2 M0 (t1 )M0 (t2 ) o
× exp .
1 + 2s21 M0 (t1 ) + 2s22 M0 (t2 ) + 4(1 − r2 )s21 s22 M0 (t1 )M0 (t2 )

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
190 Frailty Models in Survival Analysis

In the case of independence (r = 0), this simplifies to the expressions in the

univariate case (compare with (3.55)):

−M0 (t1 )m2 −M0 (t2 )m2

1 1+2s2
1
1 1+2s2
2
S(t1 , t2 ) = p e 1 M0 (t1 ) p e 2 M0 (t2 ) .
1 + 2s21 M0 (t1 ) 1 + 2s22 M0 (t2 )

For identifiability reasons, it is necessary to place restrictions on the parameters

of the frailty distribution, for example, s21 = s22 = 1 as considered by Aalen
(1987) and Iachine and Yashin (1999). The latter authors also provide a
semiparametric EM algorithm for parameter estimation in the model. The
flexibility of the multivariate normal distribution in modeling correlations can
be used, but even in this case the range of the correlation between W12 and
W22 depends on m1 and m2 and does not cover the full range between -1
and 1. Especially, the modeling of negative correlations is still problematic.
For more details in a more general case, see the paper by Aalen (1987). For
mathematical convenience, here we assume m1 = m2 = 0. This restricts the
correlation between the frailties to be nonnegative, and is different from the
usual convention EZ1 = EZ2 = 1, which would lead to very complicated
expressions. Furthermore, the restriction m1 = m2 = 0 implies a gamma
distribution with Γ( 12 , 2s12 ) (j = 1, 2) for the margins of the frailties. In this
j
case, the unconditional survival function simplifies to
1
S(t1 , t2 ) = p .
1+ 2s21 M0 (t1 ) + 2s22 M0 (t2 ) + 4(1 − r2 )s21 s22 M0 (t1 )M0 (t2 )

We are now interested in the correlation between the frailties Z1 = W12 and
Z2 = W22 . First, we need the mixed second moment EW12 W22 . To calculate
this moment, we rewrite the density of the two-dimensional normal distribu-
tion of the random effects:

EW12 W22
w12 w22

w1 w2
1
ZZ 1
− −r +
w12 w22 e (1−r ) 2s2 2s2
2 s1 s2
= √ 1 2 dw1 dw2
2πs1 s2 1 − r2
2
w1

w22
rw2
1
ZZ 1
− 2(1−r −2 s w1 −
w12 w22 e
2) s2 1 s2 2(1−r2 )s2
= √ 1 e 2 dw1 dw2
2πs1 s2 1 − r2
2
w1

w22
rw2
1
Z Z 1
2 − −2 s w1 −
w12 e 2(1−r ) s2 2(1−r2 )s2
2 1 s2
= w2 √ 1 dw1 e 2 dw2
2πs1 s2 1 − r 2
Z w22
−
= w22 I(w2 )e 2(1−r )s2 dw2
2 2
(5.16)

with 2

w1 rw2
1
Z 1
− −2 s w1
w12 e 2(1−r )
2 s2 1 s2
I(w2 ) = √ 1 dw1 .
2πs1 s2 1 − r2

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 191

In the following, we first calculate the inner integral I(w2 ) and then the outer
integral in a second step:

Z∞ 2
w1 rw2


1 − 1
−2 s w1
I(w2 ) = √ √ w12 e 2(1−r2 ) s2
1 1 s2
dw1
2πs1 1 − r2
−∞
Z∞ 2 −2r s1 w w
w1
s2 2 1
1 −
= √ √ w12 e 2(1−r2 )s2
1 dw1
2πs1 1 − r2
−∞
s2
Z∞ s
(w1 −r 1 w2 )2 −r2 1 w2
2
s2 s2
1 − − 2
= √ √ w12 e 2(1−r2 )s2
1 dw1 e 2(1−r2 )s2
1
2πs1 1 − r2
−∞
2 2 2
s21 2
− 2(1−r
−r s1 w2

= (1 − r2 )s21 + r2 w e 2 )s2 s2
1 2 .
s22 2
Here, the last equation holds because the integral is the expectation of a
random variable W 2 , with W following a normal distribution with expectation
r ss12 w2 and variance (1 − r2 )s21 . Now we are solving the second integral with
respect to w2

Z∞ 2 −r2 s2 2
(1 − r2 )s21 −
w2
− 1 w2
EW12 W22 = √ w22 e 2(1−r2 )s2
2 e 2(1−r2 )s2 s2
1 2 dw2
2πs2
−∞
2
s
r2 s12 Z∞ w22
−
+ √ 2
w24 e 2(1−r2 )s2
2 dw2
2πs2
−∞
Z∞ s2 ∞
(1 − r2 )s21 − 22
w2 r2 s12 Z w22
4 − 2s22
= √ w22 e 2s2 dw2 + √ 2
w2 e dw2
2πs2 2πs2
−∞ −∞
s2
= (1 − r2 )s21 EW22 + r2 12 EW24
s2
2
s
= (1 − r2 )s21 s22 + r2 21 3s42
s2
= (1 + 2r2 )s21 s22 .

Using the relations V(Wj2 ) = EWj4 − (EWj2 )2 = 3s4j − s4j = 2s4j (j = 1, 2), we
can now calculate the correlation between Z1 = W12 and Z2 = W22

EW12 W22 − EW12 EW22 (1 + 2r2 )s21 s22 − s21 s22

corr(W12 , W22 ) = p = p = r2 .
V(W12 )V(W22 ) 4s41 s42

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
192 Frailty Models in Survival Analysis

Consequently, the correlation between the frailties Z1 = W12 and Z2 = W22

in the pairs is just the square of the correlation between the random effects
W1 , W2 . Consequently, the correlation in this model is always nonnegative,
even if the correlation between W1 and W2 is negative. Negative correlations
between the frailties are possible if restrictions other than m1 = m2 = 0 are
used in the model. This was considered in Aalen (1987).
We are now interested in a copula representation of the unconditional
survival function similar to the correlated gamma and compound Poisson
frailty model. The univariate survival functions are given by the relation
1
Sj (t) = √ 2
(j = 1, 2). Substituting 2s21 M0 (t1 ) and 2s22 M0 (t2 ) into
1+2sj M0 (t)
the expression

1
S(t1 , t2 ) = p
1+ 2s21 M0 (t1 ) + 2s22 M0 (t2 ) + 4(1 − r2 )s21 s22 M0 (t1 )M0 (t2 )

results in

1
S(t1 , t2 ) = q . (5.17)
S1−2 (t1 )S2−2 (t2 ) − r2 (S1−2 (t1 ) − 1)(S2−2 (t2 ) − 1)

This is a very interesting relation because the bivariate copula depends only
on the correlation parameter, and not on the variances of the frailties. This
is different compared to the correlated gamma and compound Poisson frailty
model. From the interpretational point of view, this is an advantage of the
model because only the correlation parameter influences the bivariate copula,
whereas the frailty variances as heterogeneity parameters influence only the
marginal distributions.
The model contains an interesting special case. Consider the case of a
shared quadratic hazard frailty model (r = 1, s2 = s21 = s22 ). In that case, the
bivariate copula representation is of the form

S(t1 , t2 ) = (S1−2 (t1 ) + S2−2 (t2 ) − 1)−1/2 . (5.18)

This is the copula representation in the shared gamma frailty model with
σ 2 = 2 (see (4.3)). Both frailty models are equal if s2 = 2 holds. In general,
the marginal survival functions in (5.18) depend on s2 , a parameter that can
vary freely and influences the marginal distributions. Interestingly, in this
shared quadratic hazard frailty model, the copula representation of the life-
times is independent of the variance of the random effects s2 . This underlines
the fact that it is misleading to use the frailty variance as an association
measure, as is often done in shared frailty models.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 193

5.7 Other Correlated Frailty Models

A more general correlated frailty model can be introduced with the help of
frailties generated by nonnegative Lévy processes. Aalen and Hjort (2002)
considered such frailty distributions in the univariate setting, see Section 3.10.
Let Z1 = D(u0 ) + D1 (u) and Z2 = D(u0 ) + D2 (u) be frailties with respect
to two correlated lifetimes. Here it is assumed that there were three damage
processes running prior to the follow-up. The first process describes damage
shared by both partners in a pair and is running until time point u0 . D1 (u) and
D2 (u) are two independent and identically distributed versions of the process
D covering the unique damage experience of each partner. Both processes
are assumed to run until time u in parallel to the shared process D(u0 ). The
univariate unconditional survival function is given by

S(t) = ES(t|Zj ) = Ee−(D(u0 )+Dj (u))M0 (t) = e−(u0 +u)ψ(M0 (t)) ,

where ψ denotes the characteristic exponent of the Lévy process (j = 1, 2).
Consequently, after some simple algebra, the cumulative baseline hazard can
be represented in the form
 − ln S(t) 
M0 (t) = ψ −1 .
u0 + u
Hence,

S(t1 , t2 ) = Ee−D(u0 )(M0 (t1 )+M0 (t2 ))−D1 (u)M0 (t1 )−D2 (u)M0 (t2 )
= e−u0 ψ(M0 (t1 )+M0 (t2 )) e−uψ(M0 (t1 )) e−uψ(M0 (t2 ))


−1 − ln S(t1 ) −1 − ln S(t2 ) u u
= e−u0 ψ ψ ( u0 +u )+ψ ( u0 +u ) S(t1 ) u0 +u S(t2 ) u0 +u .
Using the variance formula for Lévy processes V(D(u)) = uV(D(1) − D(0))
assuming finite second moments of D(0) and D(1) and because of the inde-
pendence of D(u0 ), D1 (u) and D2 (u) it holds for the correlation between the
two frailty variables

ρ = corr(Z1 , Z2 )


cov D(u0 ) + D1 (u), D(u0 ) + D2 (u)
=q


V D(u0 ) + D1 (u) V D(u0 ) + D2 (u)


V D(u0 )
=

V D(u0 ) + D1 (u)
u0
= . (5.19)
u0 + u

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
194 Frailty Models in Survival Analysis

Applying relation (5.19) we get the final form in the general correlated frailty
model

− ln S(t1 ) − ln S(t )


1−ρ
−1
1−ρ −u0 ψ ψ ( )+ψ −1 ( u +u2 )
S(t1 , t2 ) = S(t1 ) S(t2 ) e u0 +u 0 (5.20)

The correlated gamma frailty model can be obtained from (5.20) as a special
u
−1 u
exponent from (3.59) ψ(u) = k ln(1+ λ )
case by substituting the characteristic
and its inverse ψ (u) = λ e − 1
k

− ln S(t1 ) − ln S(t )


1−ρ
−1
1−ρ −u0 ψ ψ ( )+ψ −1 ( u +u2 )
S(t1 , t2 ) = S(t1 ) S(t2 ) e u0 +u 0
“  ”
−1 − ln S(t1 ) −1 − ln S(t2 )
1−ρ 1−ρ −u0 k ln 1+1/λ ψ ( u0 +u )+ψ ( u0 +u )
= S(t1 ) S(t2 ) e
1−ρ 1−ρ − k(u 1+u) − k(u 1+u)
−u0 k
= S(t1 ) S(t2 ) S(t1 ) 0 + S(t2 ) 0 −1 .

Applying the standard constraints EZ1 = EZ2 = 1, reparameterization

λ = k(u0 + u) = 1/σ 2 , and (5.19) results in the bivariate survival function
of the symmetric correlated gamma frailty model
2 2
−ρ/σ2
S(t1 , t2 ) = S(t1 )1−ρ S(t2 )1−ρ S(t1 )−σ + S(t2 )−σ − 1 .
The correlated compound Poisson frailty model from Section 5.5 can be
obtained in a similar way. Another special case is the shared positive stable
frailty model, considered by Hougaard (1986a,b, 1995), Lam and Kuk (1997),
and Qiou et al. (1999). Mallick et al. (2008) suggest a bivariate positive
stable frailty model, but the correlation structure between the frailties in this
model is complicated.
A different approach is suggested by Henderson and Shimakura (2003).
They used their correlated frailty model with gamma-distributed frailty to
model longitudinal count data instead of event time data. Their main idea is
the derivation of a multivariate frailty with gamma distributed marginals and
a suitable correlation matrix allowing the association between event counts
in different time interval to depend on the distance between the intervals.
Applying their concept of gamma-distributed frailty to bivariate lifetimes,
the survival function is

h 2 2 2 2
i−1/σ2
S(t1 , t2 ) = S(t1 )−σ S(t2 )−σ − ρ S(t1 )−σ − 1 S(t2 )−σ − 1


,

which is another correlated gamma frailty model, but not based on the additive
decomposition of the gamma-distributed frailties as the model in Section 5.2.
In the case of σ 2 = 2 the survival function of the correlated quadratic hazard
model in (5.17) is obtained. Furthermore, the shared gamma frailty model
from Section 4.3 is also a special case of the model introduced by Henderson
and Shimakura (2003) with ρ = 1.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 195

5.8 Bivariate Frailty Cure Models

A bivariate frailty cure model for familiar association in onset of diseases
was established by Chatterjee and Shih (2001), which is an extension of the
univariate frailty cure model in (3.62). For a pair of individuals we define

1 : if the j th individual is susceptible


Yj = (5.21)
0 : otherwise,

and let Tj denote the age at onset of the disease for the jth individual when
Yj = 1 (j = 1, 2). Then the bivariate survival function is of the form

S ∗ (t1 , t2 ) = φ00 + φ10 S(t1 ) + φ01 S(t2 ) + φ11 S(t1 , t2 ), (5.22)

with the relations φ11 = P(Y1 = 1, Y2 = 1), φ10 = P(Y1 = 1, Y2 = 0),
φ01 = P(Y1 = 0, Y2 = 1), and φ00 = P(Y1 = 0, Y2 = 0). S(t1 , t2 ) denotes
the bivariate survival function of pairs with both individuals susceptible to
the event under study. S(t1 ) = S(t1 , 0) and S(t2 ) = S(0, t2 ) are the (here for
ease of presentation as equal assumed) marginal survival functions. Plugging
this survival function into the general bivariate likelihood (A.1), the likelihood
function of the ith pair in the parametric model S(t1 , t2 ) = S(t1 , t2 ; θ) yields

Li (θ) = δi1 δi2 φ11 Sti1 ,ti2 (ti1 , ti2 ; θ)

+ δi1 (1 − δi2 ) φ11 Sti1 (ti1 , ti2 ; θ) + φ10 Sti1 (ti1 ; θ)


+ (1 − δi1 )δi2 φ11 Sti2 (ti1 , ti2 ; θ) + φ01 Sti2 (ti2 ; θ)


+ (1 − δ1 )(1 − δ2 ) φ11 S(t1 , t2 ; θ) + φ10 S(t1 ; θ) + φ01 S(t2 ; θ) + φ00 ,

This bivariate survival is often given as a copula. Chatterjee and Shih (2001)
considered three different copulas: the shared gamma frailty model (Claytons
model), Frank’s copula, and Hougaard’s shared positive stable frailty model.
For more details about copulas see Chapter 6. Chatterjee and Shih (2001)
applied a two-step estimation procedure to breast cancer using the kinship
data from the Washington Ashkenazi Study, but by ignoring the dependency
among different pairs within the same family. Their model can be extended
by substituting the shared gamma frailty model by the symmetric correlated
gamma frailty model (Wienke et al. 2003a)
2 2 ρ
S(t1 , t2 ) = S(t1 )1−ρ S(t2 )1−ρ (S(t1 )−σ + S(t2 )−σ − 1) σ2 .
In the following analysis a parametric model with Gompertz baseline hazard
was used, for example,
λ
− 1
S(t) = S(0, t) = S(t, 0) = 1 + σ 2 (eϕt − 1) σ2 .
ϕ

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
196 Frailty Models in Survival Analysis

Example 5.9

The model was applied to the breast cancer incidence data of MZ and DZ
Swedish female twin pairs from the old cohort of the Swedish Twin Registry.
The results are given in Table 5.14.

Table 5.14: Correlated gamma frailty models with and without

cure fraction applied to breast cancer in Swedish twins
gamma frailty gamma frailty gamma frailty
parameter without cure with cure1 with cure2

σ2 32.78 (7.759) 4.438 (1.709) 2.508 (3.032)

ρMZ 0.154 (0.052) 1.000 ( - ) 1.000 ( - )
ρDZ 0.126 (0.040) 0.932 (0.361) 0.960 (0.454)
λ 1.3e-5 (5.2e-6) 7.7e-5 (4.8e-5) 1.2e-4 (1.2e-4)
ϕ 0.099 (0.016) 0.091 (0.012) 0.086 (0.015)
φ11 1.000 ( - ) 0.050 ( - ) 0.038 (0.021)
φ10 0.000 ( - ) 0.173 ( - ) 0.134 (0.059)
φ00 0.000 ( - ) 0.604 ( - ) 0.695 (0.138)
φ 1.000 ( - ) 0.223 (0.046) 0.1723 ( - )
log-likelihood -5122.31549 -5120.17769 -5120.06055
1 constrained by φ11 = φ2 , φ10 = φ01 = φ(1 − φ), φ00 = (1 − φ)2
2 constrained by φ10 = φ01 , φ11 + φ10 + φ01 + φ00 = 1
3 calculated as φ = φ11 + φ10

We consider two different cases of cure models. In the first case, it is assumed
that the susceptible status of the individuals in a pair is independent, for
example, P(Y1 = p1 , Y2 = p2 ) = P(Y1 = p1 )P(Y2 = p2 ) with p1 , p2 ∈ {0, 1}.
The size of the cure fraction is uniquely described by the univariate probability
φ = P(Y1 = 1) = P(Y2 = 1), which results in φ10 = φ01 = φ(1 − φ),
φ11 = φ2 , and φ00 = (1 − φ)2 . In the second case, which is an extension
of the first one, the restriction of independence between the susceptibility
status of the partners in a pair is relaxed and substituted by the weaker
constraints φ10 = φ01 , φ11 + φ10 + φ01 + φ00 = 1. When comparing the
likelihoods, it turns out that the cure model with an independent susceptible
status of the partners shows a significantly better fit compared to the model
without a cure fraction. The more complicated cure model without assuming
independence between the susceptible status of the twin partners shows no
significant improvement compared to the cure model assuming independence.
Interestingly, the estimate of the size of a susceptible fraction (due to breast
cancer) with φ = 0.223 (0.046) is close to the estimate φ = 0.22 (0.0093) in the
shared gamma frailty model found by Chatterjee and Shih (2001) in a study
population that is completely different from the one used here. Nevertheless,
the estimates of the susceptible fraction in both models in Table 5.14 are

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 197

in the range of the results obtained by Farewell et al. (1977) for different
combinations of four risk factors. If none of the risk factors is present, the
susceptible fraction is around 0.015; if all risk factors are present, the estimate
increases to 0.272.
A simulation study was performed to check the properties of the estimates
in the proposed gamma frailty model. All simulations involve generating
gamma-distributed frailties, bivariate lifetimes, censoring times, as well as
the inclusion of a cured fraction in the study population. A total of 5000
twin pairs are simulated in each data set. Samples are generated to mimic
the structure of the data analyzed in Table 5.14:

• Generate frailty variables using gamma-distributed random variables.

λ ϕt
• Generate lifetimes given the frailties using S(t|Z) = e−Z ϕ (e −1)
.
• Define cured individuals by using a random variable.
• Birth years are generated by using a uniform distribution on [1886,1925].
• Censored lifetimes are generated by using year 2000 as end of study.

The data sets are simulated assuming dependence between the susceptibility
status of the partners (second column in Table 5.14), but in the estimation
procedure, the more general model with independent susceptibility status was
applied (third column in Table 5.14). There were 1000 data sets simulated.
The mean of the parameter estimates and their standard errors are presented
in Table 5.15, in comparison with the true values. There appears to be only
moderate bias in the parameter estimates, and the overall performance is very
good. More information about this study is given in Wienke et al. (2003a).

Table 5.15: Parameter estimation in

the simulation study
parameter true mean s.e.

σ2 4.000 4.461 2.117

ρMZ 0.800 0.751 0.250
ρDZ 0.600 0.570 0.278
λ 1.00e-5 1.23e-5 1.14e-5
ϕ 0.120 0.121 0.022
φ11 0.040 0.042 0.014
φ10 0.160 0.161 0.029
φ00 0.640 0.619 0.094

Another approach to include a cure fraction into a bivariate frailty model is

the correlated compound Poisson model in Section 5.5.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
198 Frailty Models in Survival Analysis

5.9 Comparison of Different Estimation Strategies

In this section, we examine bivariate correlated frailty models, especially the
behavior of parameter estimates when using different estimation strategies.
We consider three different correlated frailty models: the gamma model and
two versions of the log-normal model. The traditional maximum likelihood
estimation procedure in the gamma case with explicitly available likelihood
function is compared with maximum likelihood methods based on numerical
integration and a Bayesian approach using MCMC methods by means of a
comprehensive simulation study. We investigate the correlation between the
two parameter estimates of the variance and the correlation of the frailties in
the bivariate correlated frailty model with equal variances, and analyze it in
detail.
Throughout this section we will refer to the correlated gamma frailty model
as Model 1. The second frailty model considered in detail in this section is
the log-normal model. Two variants of the log-normal model are analyzed.
We assume a normally distributed random variable W to generate frailty
Z = eW . The variants of the model are given by the constraints EW = 0
(Model 2) and EZ = 1 (Model 3). Unfortunately, no explicit form of the
unconditional likelihood exists in log-normal frailty models. Consequently,
estimation strategies based on numerical integration of the random effects in
the maximum likelihood approach are required.
To check whether the dependence between the variance and correlation
parameters is related to the estimation strategy or the choice of the frailty
distribution, we use the bivariate models mentioned above and apply three
different estimation strategies. First, we perform a traditional maximum
likelihood estimation procedure (only possible in the gamma model); second,
we use a maximum likelihood approach based on numerical integration; and
finally, we apply MCMC methods as a Bayesian approach.
Parameter estimation in the gamma model is straightforward. The frailty
term can be integrated out, and an explicit representation of the unconditional
bivariate survival function (5.5) exists to calculate the likelihood function.
Unfortunately, the integrals in expression (5.2) have no explicit solution in
the log-normal model. Several estimation methods for bivariate log-normal
frailty models in consequence have been suggested within a non-Bayesian
framework. Various modifications of the maximum likelihood procedure are
applicable. Ripatti and Palmgren (2000) derived an estimating algorithm
based on the penalized partial likelihood (PPL). Xue and Brookmeyer (1996)
suggested a modified EM algorithm for log-normal frailty models. Sastry
(1997) developed the modified EM algorithm for the multiplicative two-level
gamma frailty model. The same method can be applied to bivariate log-
normal frailty models. Ripatti et al. (2002) present yet another method for
EM-like algorithms in log-normal frailty models.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 199

In the present section we use numerical integration procedures. Integrals

over univariate and multivariate normal distributions can be approximated in
different ways. One possibility is to use Gauss–Hermite quadratures (Naylor
and Smith 1982, Smith et al. 1987). Similar ideas are employed in various
applications of mixed models in survival analysis (Lillard 1993, Lillard et al.
1995, Panis and Lillard 1995). The methods are implemented in the aML
software package (Lillard and Panis 2000).
Several examples on the application of Bayesian methods to multivariate
frailty models exist. Bolstad and Manda (2001) considered the gamma frailty
model. Gibbs’ sampler for the bivariate log-normal frailty model with an
exponential baseline hazard is given in Xue and Ding (1999). Korsgaard et
al. (1998) present a Bayesian inference in the log-normal frailty model with
a semiparametric baseline hazard.
In the Bayesian framework the correlated frailty model takes the form of a
hierarchical model, with the frailty variables at an intermediate stage between
observations and hyperparameters. The conditional likelihood represents the
first level of the model and is given by
2 
n Y δij
Y ′ β ′ Xij
L(β, λ, ϕ|Z) = Zij µ0 (tij )eβ Xij e−Zij M0 (tij )e ,
i=1 j=1

with Z = (Z11 , Z12 , Z21 , Z22 , . . . , Zn1 , Zn2 ) and λ, ϕ being the parameters of
the Gompertz baseline hazard function.
By definition of the model, the vector of frailties (Zi1 , Zi2 ) (i = 1, . . . , n)
in each pair is assumed to follow a bivariate log-normal distribution, with
variances σ12 = σ22 = σ 2 and correlation coefficient ρ. The parameters of the
baseline hazard, regression coefficients β, σ 2 , and ρ (the latter two being the
hyperparameters) are assumed to follow a noninformative distribution. We
adopt uniform priors over the intervals [1e-7, 0.005], [0.05, 0.15] and [-1,1]
for λ, ϕ and ρ, respectively. Furthermore, log-normal priors with mean 0.5
and variance 0.25 for σ 2 and multivariate normal priors for β are used. More
details about the MCMC approach can be found in Section 5.4.
We estimated Model 1 following a maximization procedure, and Models 2
and 3 using a numerical integration procedure (Gauss–Hermite quadrature).
MCMC methods are employed in all three models. We generated data sets
with different frailty distributions. First, we used σ 2 = 1 and ρ = 0.7. Second,
we used parameters σ 2 = 0.3 and ρ = 0.2. In both cases, λ = 0.003, ϕ = 0.07,
β ′ = (β1 , β2 ), β1 = 0.1, and β2 = −0.2. Two types of covariates were used.
One covariate was generated as binary variable

n

1 if i ≤ 2
Xij1 = n
0 if i > 2,

and the other as normal distributed variable Xij2 ∼ N (0, 1). Consequently,
the first covariate is pair-specific, whereas the second covariate is individual

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
200 Frailty Models in Survival Analysis

specific. We used sample sizes of 500 and 5000 pairs and simulated 500 data
sets in each case (50 data sets only for Bayesian methods because of time
constraints). Only the case of complete event times (meaning no censoring)
was considered here. Results are shown in Tables 5.16 – 5.18.

Table 5.16: Model 1 with two covariates, 500 simulated data sets (Bayes
50 data sets)
sample
method size λ ϕ σ2 ρ β1 β2
true 3.00e-3 0.070 0.300 0.200 0.100 -0.200
ML 500 3.01e-3 0.070 0.294 0.222 0.105 -0.198
(3.70e-4) (0.004) (0.087) (0.194) (0.082) (0.042)
ML 5000 3.01e-3 0.070 0.299 0.197 0.099 -0.200
(1.25e-4) (0.001) (0.027) (0.069) (0.026) (0.013)
Bayes 5000 3.04e-3 0.070 0.292 0.208 0.093 -0.196
(1.16e-4) (0.001) (0.024) (0.059) (0.027) (0.013)
true 3.00e-3 0.070 1.000 0.700 0.100 -0.200
ML 500 2.99e-3 0.070 1.001 0.699 0.107 -0.202
(4.03e-4) (0.005) (0.141) (0.080) (0.105) (0.054)
ML 5000 3.01e-3 0.070 1.000 0.700 0.098 -0.199
(1.34e-4) (0.002) (0.044) (0.023) (0.034) (0.017)
ML - maximum likelihood estimation
Bayes - MCMC estimation

The three models show the same pattern. As expected, the estimations for
the larger sample size are far more accurate. The most striking effect is the
strong negative correlation between estimates of ρ and σ 2 , independently of
the model and the estimation procedure (see Table 5.19). Similar simulations
were performed in the more general model with two different frailty variances
(one for each of the two individuals). The results are similar to those with
one frailty variance and are therefore omitted here.
As Bayesian methods are very time consuming, only 50 data sets with 5000
pairs each were generated for this analysis method. We run two parallel chains
from different starting points and considered the first 4000 iterations for each
chain as ”burn-in”.
The simulated values of parameters of random effects have auto-correlations
close to unity in our case. The convergence of the Markov chain is very slow.
Altogether, from 10000 up to 60000 iterations per chain were generated after
the ”burn-in” phase for each data set. The values of the Gelman–Rubin
statistics are quite close to one (see Section 5.4), indicating convergence of
the chains.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 201
Table 5.17: Model 2 with two covariates, 500 simulated data sets (Bayes
50 data sets)
sample
method size λ ϕ σ2 ρ β1 β2
true 3.00e-3 0.070 0.300 0.200 0.100 -0.200
ML* 500 3.02e-3 0.071 0.372 0.237 0.097 -0.203
(8.23e-4) (0.008) (0.378) (0.352) (0.080) (0.045)
ML* 5000 2.99e-3 0.070 0.308 0.204 0.099 -0.200
(2.04e-4) (0.002) (0.077) (0.083) (0.025) (0.013)
Bayes 5000 3.04e-3 0.070 0.300 0.218 0.095 -0.199
(2.01e-4) (0.002) (0.067) (0.089) (0.022) (0.012)
true 3.00e-3 0.070 1.000 0.700 0.100 -0.200
ML* 500 2.81e-3 0.075 1.283 0.689 0.113 -0.211
(1.06e-3) (0.014) (0.731) (0.186) (0.118) (0.059)
ML* 5000 3.07e-3 0.069 0.977 0.720 0.098 -0.199
(3.44e-4) (0.004) (0.173) (0.072) (0.034) (0.017)
Bayes 5000 3.12e-3 0.069 0.981 0.726 0.091 -0.198
(3.93e-4) (0.004) (0.193) (0.074) (0.031) (0.015)
ML* - maximum likelihood estimation with numerical integration
Bayes - MCMC estimation

Because of their simplicity and strengths, multivariate frailty models have

become very popular over the last two decades. A wide range of papers
has been published, dealing with different structures of multivariate models
(shared vs. correlated frailty models), different distributions of frailty (gamma,
log-normal, stable, PVF, etc.), different assumptions about the baseline hazard
(parametric vs. semi-parametric models), and different estimation strategies
(traditional maximum likelihood procedures, maximum likelihood procedures
based on numerical integration, EM algorithm, MCMC methods). Having
long time experience with correlated frailty models in real data applications,
we recognized a striking negative correlation between the variance and the
correlation estimates.
First, we tested whether the correlation of the estimates depends on the
distribution of frailty. We used three very often used frailty distributions to
answer this question: the gamma distribution and two different variants of
the log-normal distribution (Models 1 – 3).
Second, we tested whether the observed effect was caused by the estimation
strategy. This is why we used three different estimation strategies: traditional
maximum likelihood estimation (using a self-written GAUSS code), maximum
likelihood estimation based on numerical integration (using routines in aML
and a self-written code), and MCMC methods in WinBUGS.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
202 Frailty Models in Survival Analysis
Table 5.18: Model 3 with two covariates, 500 simulated data sets (Bayes
50 data sets)
sample
method size λ ϕ σ2 ρ β1 β2
true 3.00e-3 0.070 0.300 0.200 0.100 -0.200
ML* 500 3.01e-3 0.071 0.361 0.243 0.095 -0.204
(4.22e-4) (0.007) (0.297) (0.358) (0.078) (0.043)
ML* 5000 2.99e-3 0.070 0.308 0.204 0.099 -0.200
(1.38e-4) (0.002) (0.075) (0.082) (0.025) (0.013)
Bayes 5000 3.03e-3 0.070 0.302 0.218 0.095 -0.199
(1.36e-4) (0.002) (0.070) (0.092) (0.022) (0.012)
true 3.00e-3 0.070 1.000 0.700 0.100 -0.200
ML* 500 3.00e-3 0.075 1.323 0.683 0.107 -0.212
(4.35e-4) (0.015) (0.998) (0.160) (0.117) (0.064)
ML* 5000 3.00e-3 0.070 1.022 0.701 0.099 -0.201
(1.46e-4) (0.004) (0.179) (0.067) (0.034) (0.018)
Bayes 5000 3.02e-3 0.070 1.000 0.713 0.102 -0.199
(1.30e-4) (0.003) (0.134) (0.058) (0.034) (0.015)
ML* - maximum likelihood estimation with numerical integration

The results of the simulation study are very clear. The observed effect is
stable over different frailty distributions and different estimation strategies.
Moreover, different choices of parameters and sample sizes did not change the
correlation.
The present study is limited to parametric correlated frailty models. An
open question remains whether the observed negative correlation between
the parameter estimates is also present in semiparametric correlated frailty
models. This is a topic for future research.
A high correlation of parameter estimates could be a sign of identifiability
problems in the model. Correlated frailty models were investigated in order
to overcome the problems of the shared frailty models, which provide only
one parameter to model variance and correlation. One idea was to include
observed covariates into the models to improve identifiability characteristics.
This is why all models were run both with and without observed covariates.
The results in both cases are very similar. Consequently, we dropped results
for models without observed covariates. Two types of covariates were used,
one dichotomous and one continuous. However, no effect of the covariates on
the correlation between the parameter estimates was detected.
Regarding identifiability, note that heterogeneity (variance of frailty) and
correlation between frailties (implying dependence of lifetimes) are correlated
in a frailty model based on conditional independence. To see this, assume
that the variance of the frailty tends to zero. Obviously, this implies zero

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 203
Table 5.19: Models 1–3 with two covariates, 500 simulated
data sets (Bayes 50 data sets)
model method sample size corr(ρ, σ 2 ) parameters

1 ML 500 -0.229∗∗ σ 2 = 0.3, ρ = 0.2

1 ML 5000 -0.227∗∗
1 Bayes 5000 -0.331∗
1 ML 500 -0.431∗∗ σ 2 = 1, ρ = 0.7
1 ML 5000 -0.396∗∗
2 ML* 500 -0.241∗∗ σ 2 = 0.3, ρ = 0.2
2 ML* 5000 -0.414∗∗
2 Bayes 5000 -0.382∗∗
2 ML* 500 -0.789∗∗ σ 2 = 1, ρ = 0.7
2 ML* 5000 -0.875∗∗
2 Bayes 5000 -0.921∗∗

3 ML* 500 -0.242∗∗ σ 2 = 0.3, ρ = 0.2

3 ML* 5000 -0.409∗∗
3 Bayes 5000 -0.438∗∗

3 ML* 500 -0.717∗∗ σ 2 = 1, ρ = 0.7

3 ML* 5000 -0.862∗∗
3 Bayes 5000 -0.854∗∗
∗p < 0.05, ∗∗ p < 0.01
ML - maximum likelihood estimation
ML* - maximum likelihood estimation with numerical integration
Bayes - MCMC estimation

correlation. The assumption about the conditional independence of lifetimes

given the frailty could be the reason for the correlation linking the estimates of
variance (heterogeneity) and correlation between frailties. This would explain
why the observed effect is stable over the considered models and different
estimation procedures.
The conclusion to draw is that researchers should be cautious, and be aware
of the problem presented in applying correlated frailty models. Nevertheless,
this study shows that the models perform well and that there is nearly no
bias in the parameter estimates. This supports the use of correlated frailty
models for obtaining accurate parameter estimates and useful interpretations
of these estimates. More detailed information about this study can be found
in Wienke et al. (2005b).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
204 Frailty Models in Survival Analysis

5.10 Dependent Competing Risks in Frailty Models

In Section 5.2, cause-specific mortality data was analyzed using the correlated
gamma frailty model, assuming independence between causes of death in a
competing risk scenario. The assumption about independence between the
competing risks allows treating other causes of death as censored observations,
which makes analysis very easy. In this section we sketch an approach to
removing this limitation. The model allows testing of the hypothesis on
dependence between competing risks. To simplify description, we consider
models limited to two competing risks (death by cause of interest and by
other causes), but the model can easily be extended to the case of multiple
competing risks.
Estimating correlations of durations is difficult because of censored data,
which complicates the statistical analysis far more than does complete data.
Using a bivariate survival model to estimate correlations among lifetimes can
solve this problem. The correlated gamma frailty model can be used to fit
the lifetime data and provide a specific parameter for the correlation among
frailties. It was already used to analyze cause-specific mortality in twins
under the assumption of independence between competing risks in Examples
5.1–5.3. However, the assumption of independence between competing risks
is often questionable. Typically, in clinical and epidemiological studies, two
different types of competing events (censoring) occur. The observations of
certain individuals are censored because they are still alive at the end of the
study. Other individuals drop from the follow-up for reasons not associated
with the disease under study but through life events beyond the control of the
researcher, such as migration.
If censoring can be assumed noninformative in regard to all different causes,
then the traditional model applies, with the censoring times taken as the
minimum of the hypothetical censoring times arising from the different causes
of censoring. However, the situation becomes much more difficult if censoring
arising from at least one of the competing risks needs to be assumed to be
informative. Informative censoring means that the lifetimes and censoring
times are dependent or, in case of their independence, the distribution of
censoring times contains parameters of lifetime distribution.
Various approaches have been proposed to account for informative censoring.
Link (1989) proposes a model for informative censoring in which censoring
only occurs in a subpopulation defined by the frailty distribution. Emoto and
Matthews (1990) assume a bivariate Weibull model for failure and censoring
times. Zheng and Klein (1995) use a copula to study dependent competing
risks. Carling and Jacobson (1995) modeled unemployment duration in a
dependent competing risk framework by means of three different models.
Their paper also contains considerations of the identifiability aspects of the
models. Lin et al. (1996) use data collected after nonfatal failure events

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 205

to model dependent censoring by death or selective patient withdrawal. Lee

and Wolfe (1998) proposed a test for independent censoring. Their method
involves further follow-up of a subpopulation of lost-to-follow-up censored
subjects. Asymptotic results in a specific situation are obtained by Wienke
(1998). All these methods deal with uncorrelated subjects only. The method
described in this section deals with correlated subjects and was published first
in an analysis about genetic factors influencing CHD by Wienke et al. (2002).
Essentially, it understands correlated subjects as partial replicates for each
other, and hence gains identifiability for informative censoring.
Two types of censoring are considered, one noninformative and the other
∗ ∗
informative. Let (Ti1 , Yi1 , Ci1 , Ti2 , Yi2 , Ci2 ) (i = 1, . . . , n) be independent and
identically distributed vectors of nonnegative random variables. The variables
∗ ∗
(Ti1 , Ti2 ) denote the nonobservable cause-specific lifetimes with respect to the
cause of death under investigation. The (Yi1 , Yi2 ) are informative censoring
times (for example, lifetimes with respect to causes of death not of special
interest) and (Ci1 , Ci2 ) are noninformative censoring times. The event times
Tij = min{Tij∗ , Yij , Cij } (i = 1, . . . , n; j = 1, 2) are observed and

 1 : if Tij∗ ≤ min{Cij , Yij }



∆ij = 0 : if Cij < min{Tij∗ , Yij } (5.23)

−1 : if Yij < min{Tij∗ , Cij },


where ∆ij = 1 indicates no censoring, ∆ij = 0 noninformative censoring,

and ∆ij = −1 informative censoring. Now we derive the four-dimensional
survival function. The vector (T1∗ , Y1 , T2∗ , Y2 ) is used as a shorthand for
∗ ∗
(Ti1 , Yi1 , Ti2 , Yi2 ) (i = 1, . . . , n). Let (T1∗ , Y1 , T2∗ , Y2 ) and (Z1 , Z2 , Z3 , Z4 ) be
the lifetimes and informative censoring times, and the frailties of two partners
with respect to two causes of death. Their individual hazards are given by
the proportional hazards model

T1∗ ∼ µ1 (t1 |Z1 ) = Z1 µ01 (t1 ) T2∗ ∼ µ1 (t2 |Z3 ) = Z3 µ01 (t2 )
Y1 ∼ µ2 (y1 |Z2 ) = Z2 µ02 (y1 ) Y2 ∼ µ2 (y2 |Z4 ) = Z4 µ02 (y2 ), (5.24)

where T ∼ µ means that µ is the hazard function of T . Hence, the baseline

hazard of the lifetime of the first (T1∗ ) and second partner (T2∗ ) with respect
to the first cause of death are assumed to be equal (denoted by µ1 ). The
same statement holds for the lifetime of the first (Y1 ) and second partner (Y2 )
regarding to the second cause of death (µ2 ). The variables T1∗ , Y1 , T2∗ , and Y2
are assumed to be independent, given the vector of frailties (Z1 , Z2 , Z3 , Z4 ).
Let V1 , V8 ∼ Γ(k1 , λ0 ), V2 ∼ Γ(k2 , λ1 ), V3 ∼ Γ(k3 , λ2 ), V4 , V7 ∼ Γ(k4 , λ2 ),
V5 , V6 ∼ Γ(k5 , λ1 ) be independent gamma-distributed random variables with
k1 + k2 + k5 := λ1 = σ12 and k1 + k3 + k4 := λ2 = σ12 . The frailties and their
1 2
correlations are depicted in Figure 5.1.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
206 Frailty Models in Survival Analysis

first twin second twin

ρ1
first cause of death Z1 = λλ01 V1 + V5 + V2 V2 + V6 + λλ10 V8 = Z3

ρ ρ

ρ2
second cause of death Z2 = λλ02 V1 + V4 + V5 V3 + V7 + λλ20 V8 = Z4

Figure 5.1: Cause-specific frailties and their correlations in a twin pair.

Z1 , Z3 denote frailties with respect to the main cause of death (the cause
under study) and Z2 , Z4 are the frailties with respect to the second cause of
death. The parameters ρ1 , ρ2 , and ρ describe correlations between the frailties:
ρ1 = corr(Z1 , Z3 ), ρ2 = corr(Z2 , Z4 ), and ρ = corr(Z1 , Z2 ) = corr(Z3 , Z4 ).
The four-dimensional survival function is obtained by integrating out the
conditional lifetimes with respect to the frailty distribution by using (5.24)
and applying the Laplace transform of gamma-distributed random variables
(see Appendix A.5):

S(t1 , y1 , t2 , y2 ) (5.25)
Z1 Z2 Z3 Z4
= ES1 (t1 ) S2 (y1 ) S1 (t2 ) S2 (y2 )
2 2
− ρ12 2 2
− ρ22
= S1 (t1 )−σ1 + S1 (t2 )−σ1 − 1 σ1 S2 (y1 )−σ2 + S2 (y2 )−σ2 − 1 σ2
2 2
− ρ 2 2
− ρ
× S1 (t1 )−σ1 + S2 (y1 )−σ2 − 1 σ1 σ2 S1 (t2 )−σ1 + S2 (y2 )−σ2 − 1 σ1 σ2
σ σ σ σ
1−ρ1 − σ1 ρ 1−ρ1 − σ1 ρ 1−ρ2 − σ2 ρ 1−ρ2 − σ2 ρ
× S1 (t1 ) 2 S1 (t2 ) 2 S2 (y1 ) 1 S2 (y2 ) 1

with the constraint 0 ≤ ρ ≤ min{ σσ12 (1−ρ1 ), σσ12 (1−ρ2 )}. Parameter ρ1 denotes
the correlation between Z1 and Z3 . This parameter measures the correlation
between frailties of partners in a pair with respect to the cause of death
under investigation and is important for genetic analysis of susceptibility to
cause-specific mortality. Parameter ρ2 models the correlation between frailties
regarding all other causes of death (combined to the second cause of death or,
more general, informative censoring). Parameter ρ describes the association
between the unobservable cause-specific lifetimes in each individual. This
parameter allows to test the hypothesis of dependence between competing
risks. S1 and S2 denote the marginal survival functions regarding the first
and second cause of death. The likelihood function of the model can be found
in Appendix A.5.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Correlated Frailty Models 207

It is of interest to consider two special cases included in the above model to

get deeper insight into the structure. Fixing ρ = 0 results in two separate
correlated gamma frailty models for the two causes of death:
ρ1
2 2 −
S(t1 , y1 , t2 , y2 ) = S1 (t1 )1−ρ1 S1 (t2 )1−ρ1 (S1 (t1 )−σ1 + S1 (t2 )−σ1 − 1)
2
σ1

ρ2
2 2 −
× S2 (y1 )1−ρ2 S2 (y2 )1−ρ2 (S2 (y1 )−σ2 + S2 (y2 )−σ2 − 1) σ2
2 .
If ρ1 = 0 and ρ2 = 0 holds (unrelated individuals), the model simplifies to
σ σ
1− σ1 ρ 1− σ2 ρ 2 2 − σ ρσ
S(t1 , y1 , t2 , y2 ) = S1 (t1 ) 2 S2 (y1 ) 1 (S1 (t1 )−σ1 + S2 (y1 )−σ2 − 1) 1 2

σ σ ρ
1− σ1 ρ 1− σ2 ρ −σ12 −σ22
× S1 (t2 ) 2 S2 (y2 ) 1 (S1 (t2 ) + S2 (y2 ) − 1)− σ1 σ2 .
Using cause-specific mortality data of relatives (for example, twins), it is
possible to solve nonidentifiability problems in univariate censored lifetimes
as investigated by Tsiatis (1975). The model enables dependencies between
competing risks and allows to test for such dependencies.
The consistency and asymptotic normality of the estimators in this model
are not proofed yet, but simulation results (not shown here) indicate the
asymptotic validity of the proposed model.
The correlation coefficients between the frailties are always nonnegative,
which is clearly a limitation of the proposed model. This restriction poses no
problem when analyzing the lifetimes of relatives, where a positive association
between lifetimes seems reasonable. However, it is not clear that the same
holds for the competing risks in an individual. On the one hand, many major
diseases have risk factors in common, and consequently, the presence of any
one of these risk factors will increase the risk of death with respect to all
diseases. On the other hand, everyone dies eventually, so it is only logical
that if the risk of death from one cause is decreased, then the risk from
another cause is increased. Furthermore, the parameter ρ is only identifiable
in a ”real” multivariate case (cluster size larger than two). Having pairs
of unrelated individuals (e.g., ρ1 = ρ2 = 0), implying the univariate case,
makes the parameter ρ nonidentifiable. The nature of dependencies among
competing risks deserves further study.
A similar model for current state data was established in Giard (2001) and
Giard et al. (2002). A more general approach compared to model (5.25)
was investigated by Bandeen-Roche and Liang (1996). The difference in the
models is in the observed data. In the model just shown only the minimum
of two competing lifetimes in each individual is observed, whereas Bandeen-
Roche and Liang (1996) assume that all lifetimes are observable. Another
four-dimensional correlated gamma frailty model is proposed by Jonker and
Boomsma (2010).
Huang and Wolfe (2002) based their model on log-normal frailty and (4.17),
which means a model somewhere between shared and correlated frailty models.
In a recent paper, Huang et al. (2004) suggested a test procedure to test the
hypothesis of dependence between survival and censoring times in their model.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com

Chapter 6
Copula Models

Copula models are an attractive possibility to model clustered survival data.

However, the application of such models requires unique cluster size for all
clusters in the study population. This situation is illustrated in the twin
data examples. Here the cluster size is always two. Other bivariate examples
are event times in paired organs such as eyes and kidneys. Duchateau and
Janssen (2008) give an interesting four-dimensional example with udder data.
There are exactly four udder quarters for each cow, which is considered to
be the cluster. Researcher are interested in the time until the occurrence of
an infection. Typical examples of unique cluster size are family studies with
fixed family size, for example, studies with kernel families of size three (father,
mother, child).
The copula C is a function defined on [0, 1]n and taking values in [0, 1]. The
copula establishes the link between the marginal survival functions to generate
the joint survival function. Let Sj (tj ) (j = 1, . . . , n) denote the marginal
survival functions in clusters of size n. Then the joint survival function is
given by S(t1 , . . . , tn ) = C(S1 (t1 ), . . . , Sn (tn )). The existence of the copula C
follows from the theorem by Sklar (1959). For continuous marginal survival
functions, a unique copula exists. For a comprehensive overview of copula
models see Nelsen (2006). For survival data, the family of copulas is often
restricted to the class of Archimedian copulas, described in detail by Genest
and MacKay (1986):


S(t1 , . . . , tn ) = p q(S1 (t1 )) + . . . + q(Sn (tn )) ,
where p is a decreasing function defined on nonnegative numbers taking values
in [0, 1] and p(0) = 0. Furthermore, it is assumed that p(t) has a positive
second derivative for all t > 0 and q denotes the inverse of p. A detailed
discussion about the relation between Archimedian copulas and the shared
frailty model can be found in Goethals et al. (2008) and Duchateau and
Janssen (2008). In the following we will restrict our discussion to the bivariate
case for ease of presentation. Higher-dimensional extensions are possible but
they extend the length of the formulas. The copula of the shared gamma
frailty model (the Clayton copula) will serve as a specific example to explain
the basic ideas in Section 6.1. It will be helpful to derive the correlated gamma
frailty copula in Section 6.2. Section 6.3 deals with a general correlated frailty
copula and Section 6.4 considers the cross-ratio function.

209
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
210 Frailty Models in Survival Analysis

6.1 Shared Gamma Frailty Copula

Clayton (1978), Cox and Oakes (1984), and Yashin and Iachine (1999a)
pointed out that the bivariate survival function in the shared gamma frailty
model (4.3) can also be derived using a radically different approach. In the
next paragraph we will outline this approach and discuss its interpretational
consequences in more detail.
Denote two possibly dependent event times by T1 , T2 , and let the expression
S(t1 , t2 ) = P(T1 > t1 , T2 > t2 ) be their bivariate joint survival function that
is absolutely continuous with margins S1 (t1 ) = S(t1 , 0) and S2 (t2 ) = S(0, t2 ).
Consequently, the conditional survival function of T1 given T2 > t2 is

S(t1 , t2 )
S(t1 |T2 > t2 ) =
S2 (t2 )
and that of T1 given T2 = t2 is
∂S(t1 ,t2 )
∂t2
S(t1 |T2 = t2 ) = ∂S2 (t2 )
.
∂t2

The respective conditional hazard functions are important for the following
′
considerations. Using the relation µ(t) = − SS(t)
(t)
implies

∂
µ(t1 |T2 > t2 ) = − ln(S(t1 , t2 )) (6.1)
∂t1
and

∂ ∂

µ(t1 |T2 = t2 ) = − ln − S(t1 , t2 ) . (6.2)
∂t1 ∂t2
These hazards describe the risk of failure at age t1 for the first individual,
given the information about the event status of the second individual in the
pair. The first hazard (6.1) uses the condition {T2 > t2 }, and the second one
(6.2) is conditional on {T2 = t2 }. The deviation of the ratio of these hazards
from one was used by Oakes (1989) as a measure of mutual dependence of
the respective marginal lifetimes. The survival function of the shared gamma
frailty model can now be obtained from the following relation between the
above hazards

µ(t1 |T2 = t2 ) = (1 + σ 2 )µ(t1 |T2 > t2 ), (6.3)

where σ 2 denotes a nonnegative constant (without interpretation as a frailty
variance). The advantage of model (6.3) is the fancy interpretation of the
factor (1+σ 2 ) as the relative risk associated with a nonsurviving partner. This

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Copula Models 211

risk is always greater or equal to one because of the nonnegative dependence

in the shared frailty model. The main advantage of this model is that the
increase in risk with a nonsurviving partner is constant over time. Obviously,
formula (6.3) is equivalent to a similar condition with the roles of T1 and
T2 being interchanged. This relation defines the bivariate survival function
(4.3) uniquely (but not the marginal distributions). For derivation we use
definitions (6.1) and (6.2) and integrate both sides of the equation (Cox and
Oakes 1984):

µ(t1 |T2 = t2 ) = (1 + σ 2 )µ(t1 |T2 > t2 )

∂ −∂ ∂
S(t1 , t2 ) = (1 + σ 2 )


ln ln(S(t1 , t2 ))
∂t1 ∂t2 ∂t1
t1 Z t1
∂ −∂ ∂
Z
S(t, t2 ) dt = (1 + σ 2 )


ln ln(S(t, t2 )) dt.
0 ∂t ∂t2 0 ∂t

Solving the integral and rearranging the terms, it holds that

−∂ −∂
S2 (t2 ) = (1 + σ 2 ) ln(S(t1 , t2 )) − ln(S2 (t2 ))




ln S(t1 , t2 ) − ln
∂t2 ∂t2
−∂ 2 −∂ 2
S(t1 , t2 ) − ln(S(t1 , t2 )1+σ ) = ln S2 (t2 ) − ln(S2 (t2 )1+σ )



ln
∂t2 ∂t2
∂ ∂
∂t2 S(t ,
1 2t ) ∂t2 S2 (t2 )
2 = .
S(t1 , t2 )1+σ S2 (t2 )1+σ2

Integrating again implies the bivariate survival function

t2 ∂ t2 ∂
∂t S(t1 , t) ∂t S2 (t)
Z Z
dt = dt
0 S(t1 , t)1+σ2 0 S2 (t)1+σ2
2 2 2
S(t1 , t2 )−σ − S1 (t1 )−σ = S2 (t2 )−σ − 1
2 2
− 1
S(t1 , t2 ) = S1 (t1 )−σ + S2 (t2 )−σ − 1 σ2
.

Thus, there are two ways of deriving formula (4.3) based on two radically
different concepts: one uses the assumption (6.3), concerning proportionality
of conditional hazards; the other uses the concept of gamma-distributed shared
frailty. In the literature it has often been claimed that particular copulas can
be deduced from frailty models by choosing the appropriate frailty distribution.
This happens because the survival functions are similar in both approaches,
at least at first sight. It is necessary to note that there exists one important
difference between the survival functions resulting from both approaches. In
the shared gamma frailty model, transformation (3.17) is used to generate
the bivariate survival function. Consequently, the marginal survival functions

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
212 Frailty Models in Survival Analysis

S1 and S2 include the frailty parameter σ 2 , which is not the case in the
approach using proportional conditional hazards (6.3). The latter model is
a copula model. Consequently, it is necessary to make assumptions about
the data generation mechanism. This is important for the interpretation of
the data analysis; for more details see Goethals et al. (2008). The bivariate
representation (6.3) was extended to the multivariate setting by Guo and
Rodriguez (1992) and Guo (1993).
Fitting copula models is often performed by using a two-stage procedure
(Shih and Louis 1995a, Glidden 2000, Andersen 2005). In the first stage the
marginal survival functions are estimated without taking into account the
clustering of the data. This can be done in a parametric or nonparametric
way. A semiparametric approach is also possible by assuming a Cox model and
therefore the inclusion of covariates (Glidden 2000, Andersen 2005). It turns
out that these estimates are consistent (Spiekerman and Lin 1998) and can be
used in the second step to estimate the copula parameters. Alternatively, the
likelihood can also be maximized in a one-step procedure. In the next section
we consider the copula of the correlated gamma frailty model to illustrate the
difference between the copula and the frailty approach.

6.2 Correlated Gamma Frailty Copula

Starting from the conditional frailty model, the hazard of individual j in the
ith pair is given by

µ(t|Zij ) = Zij µ0j (t),

omitting covariates for ease of presentation. The joint conditional survival
function is then given by

S(t1 , t2 |Zi1 , Zi2 ) = e−Zi1 M01 (t1 )−Zi2 M02 (t2 ) .

The joint survival function is now obtained by integrating out the random
frailties
Z
S(t1 , t2 ) = e−z1 M01 (t1 )−z2 M02 (t2 ) f (z1 , z2 ) dz1 dz2

with respect to their density f (z1 , z2 ). This integral can be solved analytically,
and the joint survival function of the correlated gamma frailty model is now
(see 5.6)
σ1 σ2
S1 (t1 )1− σ2 ρ S2 (t2 )1− σ1 ρ
S(t1 , t2 ) = 2 2
ρ . (6.4)
(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1) σ1 σ2

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Copula Models 213

The joint survival function in the copula model looks completely the same. To
understand the difference between the correlated gamma frailty model and the
related copula, it is necessary to keep in mind how formula (6.4) was derived.
1
− 2
The formula is obtained by substituting the expressions (1 + σj2 M0j (tj )) σj
(j = 1, 2) by the marginal survival functions Sj (tj ) in (5.5) using (3.17).
Consequently, in the frailty model, the survival functions Sj (tj ) depend on
the variance of the frailties σj2 . This is often overlooked when dealing with
representation (6.4). In the copula approach the derivation in (5.5) is no
longer of interest, formula (5.6) is just a bivariate survival function without
any frailty interpretation. It should be noted that the copula in (6.4) is not an
Archimedian copula in contrast to the Clayton copula, which is obtained as a
special case with ρ = 1. Here, the marginal survival functions are completely
unrestricted and can follow any distribution. That is the reason why a two-
step procedure can be applied by estimating the marginal survival functions
in the first step. The assumption that the marginal survival functions depend
on the frailty variances is no longer used. As a consequence, in frailty models,
such a two-step estimation procedure is impossible. In the frailty model the
choice of the frailty distribution determines the functional form of the copula
as well as the functional form of the marginal distributions, which is different
compared to the copula model. It is important to understand the difference
between the two approaches because the results and their interpretation will
not be the same. We illustrate this problem in the following example by
reanalyzing the data of male Danish twins in Example 5.1 with respect to
cancer.

Example 6.1
Because of the symmetry in the twin data, the restrictions S(t) = S1 (t) = S2 (t)
and σ 2 = σ12 = σ22 are imposed. A parametric approach with a Gompertz
baseline hazard function is considered. In the case of the gamma frailty models
this results in

λ 1
S(t) = (1 + σ 2 (eϕt − 1))− σ2 .
ϕ
In the copula approach, no restrictions about the form of the marginal survival
functions exist. This can, for example, be incorporated into the model by the
specification

λ 1
S(t) = (1 + s2 (eϕt − 1))− s2 ,
ϕ
where s2 denotes a new parameter of the marginal distribution. In some
sense, s2 is interpretable as the variance of a univariate gamma-distributed
frailty. The correlated gamma frailty model is characterized by the condition
σ 2 = s2 . Consequently, using this parameterization, the frailty model is a

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
214 Frailty Models in Survival Analysis

special case of the copula model. Additionally, the semiparametric copula

model is also applied to the data. The results of the maximum likelihood
parameter estimation in all three models are given in Table 6.1.

Table 6.1: Correlated gamma frailty and copula models

applied to lifetimes of Danish male twins with respect to cancer
parameter frailty model copula model semip. copula

σ2 3.388 (0.572) 5.911 (3.154) 5.707 (3.019)

ρMZ 0.350 (0.088) 0.275 (0.090) 0.269 (0.091)
ρDZ 0.194 (0.068) 0.148 (0.061) 0.148 (0.062)
λ 4.4e-6 (1.5e-6) 4.7e-6 (1.6e-6)
ϕ 0.118 (0.006) 0.117 (0.006)
s2 3.264 (0.579)
log-likelihood -14283.1939 -14282.3678

The difference between the models in the first two columns is obvious. In the
copula model, σ 2 has no longer an interpretation as a frailty variance, and
ρ is no longer interpretable as a correlation between frailties. Both ρ and
σ 2 are parameters of the copula without any relation to random effects. In
contrast, the parameter s2 has an interpretation as variance of the univariate
frailty in the marginal distributions. Consequently, the foregoing approach can
be used to test whether the frailty model fits the data. If both parameters
σ 2 and s2 are similar, the frailty model provides a good fit. If there is a
significant difference between the two parameters, this contradicts the frailty
model and speaks in favor of the less restrictive copula model. In the situation
considered here, the likelihood ratio test indicates no significant difference
between the models, preferring the more simple frailty model. This fact is also
supported by the large standard error of σ 2 in the copula model. Interestingly,
semiparametric analysis (last column) reveals similar results compared to the
parametric copula model. This speaks in favor of the hypothesis that the
parametric analysis is not sensitive to the choice of the hazard function. In
the semiparametric copula model, a one-step estimation procedure is used,
where the Kaplan–Meier–estimator is plugged into the likelihood function
to substitute the unknown marginal survival functions. Unfortunately, this
approach is not possible in the frailty model, where the more demanding EM
algorithm is needed to analyze the semiparametric model. The reason for
this difference is the already mentioned dependence of the marginal survival
functions on the frailty variance in the frailty model, which is not present in
the copula approach.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Copula Models 215

6.3 General Correlated Frailty Copulas

The question arises whether the correlated gamma frailty copula can also be
derived without a frailty approach. Yashin and Iachine (1999a,b) gave such
a derivation of this copula using a transformation of a well-known bivariate
survival function. It turns out that this concept can also be used to generate
copulas related to other correlated frailty models. Additionally, the authors
extended the copula of the correlated gamma frailty model (6.4) to the case
ρ < 0, and conditions are derived that guarantee negative associations of
lifespans. Let T1 , T2 be two dependent failure times, and S(t1 , t2 ) be their
respective bivariate survival function. Any such function can be written in
the form

S(t1 , t2 ) = S1 (t1 )S2 (t2 )eA(t1 ,t2 ) , (6.5)

where A(t1 , t2 ) = ln S1S(t 1 ,t2 )

(t1 )S2 (t2 ) , and Sj (tj ) (j = 1, 2) are the marginal
univariate survival functions. If the bivariate distribution of the survival times
(T1 , T2 ) is absolutely continuous, the above-introduced function A(t1 , t2 ) can
Rt Rt
be given using the representation A(t1 , t2 ) = 0 1 0 2 φ(u, v) du dv (Anderson
et al. 1992). Expression (6.5) is called an exponential representation of a
bivariate survival function with association function φ. It turns out (and will
be proved in the following theorem) that the expression

S̃(t1 , t2 ) = S1 (t1 )S2 (t2 )eρA(t1 ,t2 )

S(t1 ,t2 )


ρ ln
= S1 (t1 )S2 (t2 )e S1 (t1 )S2 (t2 )

1−ρ
= S1 (t1 )S2 (t2 ) S(t1 , t2 )ρ (6.6)

determines for 0 ≤ ρ ≤ 1 a bivariate survival function with exactly the same

marginal distributions as in the original model S(t1 , t2 ). The new bivariate
survival function S̃(t1 , t2 ) is a geometric mean of two survival functions.
The first survival function corresponds to the independent case S1 (t1 )S2 (t2 ),
whereas the second deals with dependent survival times given by S(t1 , t2 ).
The parameter ρ may therefore be considered as an association describing the
location of S̃(t1 , t2 ) between the survival functions S1 (t1 )S2 (t2 ) and S(t1 , t2 ).
Relation (6.6) is of great practical relevance. In general, each copula can be
extended in this simple way while keeping the same marginal distributions.
Here we are especially interested in the fact that each shared frailty copula
can be generalized in a very simple way to a related correlated copula. This
opens new avenues in the modeling of dependent event times by the copula
approach.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
216 Frailty Models in Survival Analysis

THEOREM 6.1
(Yashin and Iachine 1999a) Let S(t1 , t2 ) be the joint survival function of T1
and T2 given by formula (6.5) with marginals S1 (t1 ) and S2 (t2 ). Furthermore,
Rt Rt
let A(t1 , t2 ) = 0 1 0 2 φ(u, v)dudv, with φ(u, v) ≥ 0 for all u ≥ 0, v ≥ 0. Then,
for any 0 ≤ ρ ≤ 1,

S̃(t1 , t2 ) = S1 (t1 )S2 (t2 )eρA(t1 ,t2 )

determines a survival function for survival times T̃1 and T̃2 . The marginal
distributions of Tj and T̃j (j = 1, 2) are identical.

Proof: It follows from (6.6) that S̃(t1 , 0) = S1 (t1 ), and S̃(0, t2 ) = S2 (t2 ), and
that S̃(t1 , ∞) = S̃(∞, t2 ) = S̃(∞, ∞) = 0. Hence, Ã(t1 , t2 ) = ρA(t1 , t2 ). To
2
complete the proof, it is enough to show that S̃t1 t2 (t1 , t2 ) = ∂ ∂t S̃(t1 ,t2 )
1 ∂t2
≥ 0.
Using relation (6.6) we obtain

S̃t1 t2 (t1 , t2 ) (6.7)

= S1′ (t1 )S2′ (t2 )eρA(t1 ,t2 )
+ S1′ (t1 )S2 (t2 )ρAt2 (t1 , t2 )eρA(t1 ,t2 )
+ S1 (t1 )S2′ (t2 )ρAt1 (t1 , t2 )eρA(t1 ,t2 )
+ S1 (t1 )S2 (t2 )ρAt1 t2 (t1 , t2 )eρA(t1 ,t2 )
+ S1 (t1 )S2 (t2 )ρ2 At1 (t1 , t2 )At2 (t1 , t2 )eρA(t1 ,t2 )

= S̃(t1 , t2 ) µ1 (t1 )µ2 (t2 ) − µ1 (t1 )ρAt2 (t1 , t2 )
− µ2 (t2 )ρAt1 (t1 , t2 ) + ρAt1 t2 (t1 , t2 ) + ρ2 At1 (t1 , t2 )At2 (t1 , t2 )

 
= S̃(t1 , t2 ) (ρAt1 (t1 , t2 ) − µ1 (t1 ))(ρAt2 (t1 , t2 ) − µ2 (t2 )) + ρAt1 t2 (t1 , t2 ) ,
S ′ (t )
where µi (ti ) = − Sii (tii ) . Starting from relation A(t1 , t2 ) = ln S1S(t 1 ,t2 )
(t1 )S2 (t2 ) , the
following holds:

St1 (t1 , t2 ) S1′ (t1 )

At1 (t1 , t2 ) = − = µ1 (t1 ) − µ1 (t1 , t2 )
S(t1 , t2 ) S1 (t1 )
and At2 (t1 , t2 ) = µ2 (t2 ) − µ2 (t1 , t2 ). Here, µi (t1 , t2 ) denotes the hazard of
individual i given the information that {Tj > tj } (i, j = 1, 2; i 6= j). Taking
into account that At1 t2 (t1 , t2 ) = φ(t1 , t2 ), we can rewrite (6.7) in the form


S̃t1 t2 (t1 , t2 ) = S̃(t1 , t2 ) ((ρ − 1)µ1 (t1 ) − ρµ1 (t1 , t2 ))

× ((ρ − 1)µ2 (t2 ) − ρµ2 (t1 , t2 )) + ρφ(t1 , t2 ) (6.8)

which is always nonnegative, and this completes the proof. 

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Copula Models 217

If S(t1 , t2 ) is be given by (4.2) (Clayton copula, copula of the shared gamma

frailty model), then S̃(t1 , t2 ) is obtained as the survival function of the copula
of the correlated gamma frailty model (6.4) with σ12 = σ22 :

S1 (t1 )1−ρ S2 (t2 )1−ρ

S̃(t1 , t2 ) = ρ .
(S1 (t1 )−σ2 + S2 (t2 )−σ2 − 1) σ2

Consequently, the copula of the correlated gamma frailty model can be derived
from the copula of the shared gamma frailty model without any interpretation
related to frailty. The association function φ(t1 , t2 ) in this case is of the form

2 2
σ 2 µ1 (t1 )µ2 (t2 )S1 (t1 )−σ S2 (t2 )−σ
φ(t1 , t2 ) = . (6.9)
(S1 (t1 )−σ2 + S2 (t2 )−σ2 − 1)2

In the positive stable model, a similar extension can be performed starting

from the copula of the shared positive frailty model (4.7):

h 
1/γ
1/γ γ i
S̃(t1 , t2 ) = S1 (t1 )1−ρ S2 (t2 )1−ρ exp ρ − ln S1 (t1 ) + − ln S2 (t2 ) .

It should be noted that a correlated positive stable frailty model similar to the
other correlated frailty models with explicit joint survival function does not
exist, because the moments of the positive stable distribution are all infinite.
Consequently, no correlation between the frailties in a cluster exist, which
makes the interpretation of the parameter ρ more difficult. The association
function becomes

1−γ 
1
1
γ−2
φ(t1 , t2 ) = µ1 (t1 )µ2 (t2 ) − ln S1 (t1 ) γ + − ln S2 (t2 ) γ
γ

1 −1
1 −1
× − ln S1 (t1 ) γ − ln S2 (t2 ) γ .

The same extension holds also in the more general case of the compound
Poisson model, including the gamma and the inverse Gaussian model as
special cases. Starting with the copula of the shared compound Poisson
frailty model (4.8), the copula of the correlated compound Poisson frailty
model (5.14) is obtained by applying relation (6.6):

S̃(t1 , t2 ) = S1 (t1 )1−ρ S2 (t2 )1−ρ

h ρ(1 − γ)  γσ 2 1
γ + (1 −
γσ 2 1
γ − 1

γ i
× exp 1 − (1 − ln S (t
1 1 )) ln S (t
2 2 ))
γσ 2 1−γ 1−γ

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
218 Frailty Models in Survival Analysis

and association function

φ(t1 , t2 )
 γσ 2
1 γσ 2
1 γ−2
= σ 2 µ1 (t1 )µ2 (t2 ) 1 − ln S1 (t1 ) γ + 1 − ln S2 (t2 ) γ − 1
1−γ 1−γ
γσ 2
γ1 −1 γσ 2
γ1 −1
× 1− ln S1 (t1 ) 1− ln S2 (t2 ) .
1−γ 1−γ

The following theorem extends the general copula model considered previously
to the case of negative ρ, resulting in negative association between survival
times in a cluster.

THEOREM 6.2
(Yashin and Iachine 1999a) Let the conditions of Theorem 6.1 hold. Then,
for any ρ satisfying
 µ (t )µ (t ) 
1 1 2 2
max − < ρ < 0, (6.10)
t1 ,t2 φ(t1 , t2 )
the function S̃(t1 , t2 ) given by (6.6) determines a bivariate distribution of
negatively correlated survival times.

Proof: Note that for any ρ < 0, we have

(ρ − 1)µ1 (t1 ) − ρµ1 (t1 , t2 ) (ρ − 1)µ2 (t2 ) − ρµ2 (t1 , t2 ) ≥ µ1 (t1 )µ2 (t2 ) ≥ 0

because of µi (ti ) ≥ µi (t1 , t2 ) due to φ(t1 , t2 ) ≥ 0, and hence, (6.8) yields

 
S̃t1 t2 (t1 , t2 ) ≥ S̃(t1 , t2 ) µ1 (t1 )µ2 (t2 ) + ρφ(t1 , t2 )

Thus, S̃t1 t2 (t1 , t2 ) is nonnegative when ρ satisfies (6.10). The covariance of

T̃1 , T̃2 is

Z ∞ Z ∞ Z ∞ Z ∞
cov(T̃1 , T̃2 ) = S̃(u, v) du dv − S1 (u)S2 (v) du dv.
0 0 0 0

Using representation (6.6) for S̃(t1 , t2 ), we get

Z ∞Z ∞
cov(T̃1 , T̃2 ) = S1 (u)S2 (v)(eρA(u,v) − 1) du dv.
0 0

So, if A(u, v) > 0, then the sign of cov(T̃1 , T̃2 ) coincides with the sign of ρ.
This completes the proof. 

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Copula Models 219

The extension of the copula model to negative values of ρ is important for

several reasons. For example, negatively correlated lifespans are of interest
in the situation of dependent competing risks such as competing causes of
death in medical applications. Another aspect is testing of hypotheses. When
ρ is considered as an association parameter (similar to the frailty model),
researchers are frequently interested in testing the null hypothesis H0 : ρ = 0
versus the alternative HA : ρ > 0 about the independence of the lifetimes.
By extending the model to ρ < 0, the point ρ = 0 becomes an internal
point of the parameter space, and standard test theory for the likelihood
ratio test can be applied, simplifying the analysis. In the following we consider
condition (6.10) in more detail for the correlated gamma, positive stable, and
the compound Poisson/PVF copulas. For the correlated gamma frailty copula
model, relation (6.10) becomes

2 2
2 
 µ1 (t1 )µ2 (t2 )   S1 (t1 )−σ + S2 (t2 )−σ − 1 −1
max − = max − =
t1 ,t2 φ(t1 , t2 ) t1 ,t2 σ 2 S1 (t1 )−σ2 S2 (t2 )−σ2 σ2

by using (6.9). If σ 2 is smaller than one, the parameter ρ can take on all
values between -1 and 1. In the correlated positive stable copula, it turns out
that

 µ (t )µ (t ) 
1 1 2 2
max −
t1 ,t2 φ(t1 , t2 )

1
1 2−γ
 γ − ln S1 (t1 ) γ + − ln S2 (t2 ) γ 
= max −
1 −1
1 −1 = 0,
t1 ,t2
(1 − γ) − ln S1 (t1 ) γ − ln S2 (t2 ) γ

meaning that the correlated positive stable copula cannot be extended to

negative values of ρ. In the general compound Poisson frailty model, it holds
that

 µ (t )µ (t ) 
1 1 2 2
max −
t1 ,t2 φ(t1 , t2 )

1
1 2−γ
γσ2 γσ2
 1 − 1−γ ln S1 (t1 ) γ + 1 − 1−γ ln S2 (t2 ) γ − 1  −1
= max − 1 1 = 2,
t1 ,t2 γσ2
σ 2 1 − 1−γ

−1
ln S1 (t1 ) γ γσ2
1 − 1−γ

−1
ln S2 (t2 ) γ σ

giving the same result as in the correlated gamma copula model.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
220 Frailty Models in Survival Analysis

6.4 Cross-Ratio Function

A measure of local dependence in dependent event times describing changes
over time is the cross-ratio function introduced by Clayton (1978)

µ(t1 |T2 = t2 )
θ(t1 , t2 ) = .
µ(t1 |T2 > t2 )
This expression has a nice interpretation. For twins considered as example
through the last chapter, the ratio compares the hazard of the first partner
to experience the event at time t1 , given that the second partner experiences
the event at time t2 , to the hazard of the first partner at time t1 , given that
the second partner experience the event later than time t2 . It is often helpful
to write the cross-ratio function in terms of the survival function and their
derivatives (Oakes 1989):

S(t1 , t2 )St1 t2 (t1 , t2 )

θ(t1 , t2 ) = . (6.11)
St1 (t1 , t2 )St2 (t1 , t2 )
It should be noted that the cross-ratio function in the shared frailty model
depends on t1 and t2 only through the joint survival function S(t1 , t2 ) because
it is an Archimedean copula (Duchateau and Janssen 2008). The copula of
the correlated gamma frailty model is not an Archimedean copula and their
cross-ratio function can therefore not be represented as a function of S(t1 , t2 ).
In the following we calculate the cross-ratio for the correlated gamma copula.
The derivatives needed are provided in Appendix A.2. First we consider the
numerator.

S(t1 , t2 )St1 t2 (t1 , t2 )

2 2 ρ
= S(t1 )1−ρ S(t2 )1−ρ (S(t1 )−σ + S(t2 )−σ − 1)− σ2
2 2 ρ
× S(t1 )−ρ f (t1 )S(t2 )−ρ f (t2 )(S(t1 )−σ + S(t2 )−σ − 1)− σ2 −2
 2 2
× (1 − ρ)2 (S(t1 )−σ + S(t2 )−σ − 1)2
2 2 2
+ (1 − ρ)ρS(t2 )−σ (S(t1 )−σ + S(t2 )−σ − 1)
2 2 2 2 2

+ (1 − ρ)ρS(t1 )−σ (S(t1 )−σ + S(t2 )−σ − 1) + ρ(ρ + σ 2 )S(t1 )−σ S(t2 )−σ
2 2 ρ
= S(t1 )1−ρ S(t2 )1−ρ (S(t1 )−σ + S(t2 )−σ − 1)− σ2
2 2 ρ
× S(t1 )−ρ f (t1 )S(t2 )−ρ f (t2 )(S(t1 )−σ + S(t2 )−σ − 1)− σ2 −2
 2 2 2

× (1 − ρ)(S(t1 )−σ + S(t2 )−σ − 1) + ρS(t1 )−σ
2 2 2
2 2

× (1 − ρ)(S(t1 )−σ + S(t2 )−σ − 1) + ρS(t2 )−σ + ρσ 2 S(t1 )−σ S(t2 )−σ .

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Copula Models 221

The denominator of the cross-ratio function is given by

2 2 ρ
St1 (t1 , t2 )St2 (t1 , t2 ) = S(t1 )−ρ f (t1 )S(t2 )1−ρ (S(t1 )−σ + S(t2 )−σ − 1)− σ2 −1
2 2
× − S(t1 )−σ − (1 − ρ)S(t2 )−σ + (1 − ρ)


2 2 ρ
× S(t1 )1−ρ S(t2 )−ρ f (t2 )(S(t1 )−σ + S(t2 )−σ − 1)− σ2 −1
2 2
× − (1 − ρ)S(t1 )−σ − S(t2 )−σ + (1 − ρ)


2 2 ρ
= S(t1 )−ρ f (t1 )S(t2 )1−ρ (S(t1 )−σ + S(t2 )−σ − 1)− σ2 −1
2 2 ρ
× S(t1 )1−ρ S(t2 )−ρ f (t2 )(S(t1 )−σ + S(t2 )−σ − 1)− σ2 −1
2 2 2

× (1 − ρ)(S(t1 )−σ + S(t2 )−σ − 1) + ρS(t1 )−σ
2 2 2

× (1 − ρ)(S(t1 )−σ + S(t2 )−σ − 1) + ρS(t2 )−σ .

Using the preceding expressions the cross-ratio function (6.11) of the correlated
gamma copula can be obtained as

θ(t1 , t2 ) =
2 2
ρσ 2 S(t1 )−σ S(t2 )−σ
1+
.
(1 − ρ)(S(t2 )−σ2 − 1) + S(t1 )−σ2 (1 − ρ)(S(t1 )−σ2 − 1) + S(t2 )−σ2

In the following we consider some interesting special cases. At the beginning

of the follow-up for both individuals, the cross-ratio function simplifies to the
expression θ(0, 0) = 1 + ρσ 2 . In the case of the shared gamma copula it holds
ρ = 1, and we obtain the well-known time-independent cross-ratio function
θ(t1 , t2 ) = 1 + σ 2 (compare with relation (6.3)). If times converge to infinity,
it holds that

lim θ(t1 , t2 ) = lim θ(t1 , t2 ) = 1

t1 →∞ t2 →∞

and

ρσ 2
lim θ(t1 , t2 ) = 1 + .
t1 ,t2 →∞ (2 − ρ)2

For ρ = 0, the two lifetimes are independent, indicated by θ(t1 , t2 ) = 1.

The cross-ratio function of the correlated compound Poisson copula can be
obtained in a similar way (the derivation is not shown here) by using the
γσ2 γσ2
shorthands a(t1 ) = 1 − 1−γ ln S(t1 ) and a(t2 ) = 1 − 1−γ ln S(t2 ):

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
222 Frailty Models in Survival Analysis

θ(t1 , t2 ) =
ρσ 2 a(t1 )1/γ−1 (a(t1 )1/γ + a(t2 )1/γ − 1)γ
1+
(1 − ρ)(a(t1 )1/γ + a(t2 )1/γ − 1) + ρa(t1 )1/γ−1 (a(t1 )1/γ + a(t2 )1/γ − 1)γ
a(t2 )1/γ−1
× .
(1 − ρ)(a(t1 )1/γ + a(t2 )1/γ − 1) + ρa(t2 )1/γ−1 (a(t1 )1/γ + a(t2 )1/γ − 1)γ

With the shared compound Poisson copula obtained by ρ = 1, this expression

simplifies to

σ2
θ(t1 , t2 ) = 1 +
(a(t1 )1/γ + a(t2 )1/γ − 1)γ
(see Duchateau and Janssen 2008). The gamma model is obtained as limiting
1/γ−1 2 1/γ 2
case for γ → 0. It turns out that aj → S(tj )−σ and aj → S(tj )−σ
(j = 1, 2). Furthermore, it holds (a(t1 )1/γ +a(t2 )1/γ −1)γ → 1. Consequently,

θ(t1 , t2 )
ρσ 2 a(t1 )1/γ−1 (a(t1 )1/γ + a(t2 )1/γ − 1)γ
=1+
(1 − ρ)(a(t1 )1/γ + a(t2 )1/γ − 1) + ρa(t1 )1/γ−1 (a(t1 )1/γ + a(t2 )1/γ − 1)γ
a(t2 )1/γ−1
×
(1 − ρ)(a(t1 )1/γ + a(t2 )1/γ − 1) + ρa(t2 )1/γ−1 (a(t1 )1/γ + a(t2 )1/γ − 1)γ
2
ρσ 2 S(t1 )−σ
→1+ 2
(1 − ρ)(S(t1 )−σ + S(t2 )−σ2 − 1) + ρS(t1 )−σ2
2
S(t2 )−σ
×
(1 − ρ)(S(t1 )−σ2 + S(t2 )−σ2 − 1) + ρS(t2 )−σ2
2 2
ρσ 2 S(t1 )−σ S(t2 )−σ
=1+
(1 − ρ)(S(t2 )−σ2 − 1) + S(t1 )−σ2 (1 − ρ)(S(t1 )−σ2 − 1) + S(t2 )−σ2

The cross-ratio function at the beginning of the follow-up for both individuals
is θ(0, 0) = 1 + ρσ 2 . In the case ρ = 0, the two lifetimes are independent.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com

Chapter 7
Different Aspects of Frailty
Modeling

This chapter presents a nonsystematic collection of problems related to very

different types of mainly nonstandard frailty models in various applications
not covered in the preceding chapters. Most of the problems are not treated
in detail here; they are just mentioned and shortly discussed. Additionally,
references are given for further reading.

7.1 Dependence and Interaction between Frailty and

Observed Covariates
Nearly all contributions and applications on frailty models in a regression
setting like (3.2) consider the unobserved frailty variable Z as independent of
the observed explanatory variables X. This assumption is usually made only
for mathematical convenience. Of course, in some cases this independence
assumption could be inappropriate. In the case of competing risk models,
di Serio (1997) analyzed this question to explain the effect of unexpected
protectivity. This problem occurs when a covariate shows a protective impact
not expected from a medical point of view. A simulation study was conducted
by di Serio (1997) in a two competing risks scenario with two frailties. Each
frailty is related to one of the competing events. It turns out that dependence
between the two frailties alone does not cause false protectivity; conversely,
false protectivity may occur according to the magnitude and the sign of the
dependence between frailty and the observed covariate. Here, dependence
between the two frailties implies dependence of the two competing event times,
resulting in a univariate lifetime model with dependence between lifetimes and
censoring times. The problem is that an assumption about the distribution of
frailty is not enough since what we need are actually nontestable assumptions
about the direction and the intensity of the dependence between frailty and the
observable covariates, which makes the data analysis quite arbitrary. Thus,
survival models where the observed covariates depend on unobserved random
effects may not be identifiable from univariate survival data without additional
assumptions.

223
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
224 Frailty Models in Survival Analysis

It turns out that such models can be identified from bivariate (multivariate)
data when they are embedded into a frailty model by using the relation

σ(X) = σeγX ,

where γ is a parameter to be estimated, and σ 2 denotes the variance of the

frailty Z. Thus, the distribution of Z (or more specifically, its variance)
depends on the observed covariate values X. An example of such models in
the multivariate case can be found in Wassell and Moeschberger (1993), who
have studied the impact of interventions in the Framingham Heart Study by
introducing their concept of modified gamma frailty. Yashin et al. (1999b)
have investigated the heritability of susceptibility to death after accounting for
the dependence between frailty and observed covariates (BMI and smoking)
in Danish twins. Noh et al. (2006) suggested the notion of dispersed frailty for
such kind of models and applied them to the well-known kidney infection data
from McGilchrist and Aisbett (1991) to verify the hypothesis of heterogeneity
in frailty distribution in the study population.
Several authors have considered the problem of unobserved heterogeneity
caused by omitted covariates in randomized clinical trials. However, usually
no interaction between the omitted covariate and the binary treatment variable
is allowed. Exceptions are the papers by Li et al. (2002) and Xu (2004).
Li et al. (2002) considered asymptotic relative efficiency and power of
the log-rank test in models with an interaction between normally distributed
random effects Wi and a binary covariate Xi (for example, representing the
treatment effect in a randomized trial) in a univariate model:

µ(t|Xi , Wi ) = µ0 (t)eβXi +Wi +αWi Xi . (7.1)

The interaction of the frailty term with observed covariates is incorporated

on the same scale as the fixed effects. This model is flexible in the sense that
the test on interaction can be based on the parameter α, whereas the test on
unobserved heterogeneity is based on the variance parameter of the random
effects.
Xu (2004) discussed a multivariate survival model with normal distributed
random effects. As an example, for a multicenter trial, if a binary covariate
Xij indicates one of two treatment assignments for individual j in center i,
the following model can be used to capture center-specific treatment effects:

µ(t|Xij , Wi ) = µ0 (t)eβXij +Wi Xij , (7.2)

with Wi denoting the additional random treatment effects specific to the ith
center, assumed to follow a normal distribution with zero expectation and
some unknown variance. The proposed MCEM algorithm allows estimating
the random effects, which could be, for example, used to rank the centers with
respect to their treatment success.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Different Aspects of Frailty Modeling 225

Some research efforts have been undertaken to study extensions of the shared
frailty approach to more general models allowing, for example, for interaction
between treatment and center in a multicenter clinical trial (Yamaguchi and
Ohashi 1999, Vaida and Xu 2000, Legrand et al. 2005, 2006, Massonnet et
al. 2008) or interaction between treatment and study in a meta-analysis with
event time outcome (Rondeau et al. 2008). The hazard function in this model
is
′
Xij +W0i +W1i Xij1
µ(t|Xij , Wi ) = µ0 (t)eβ , (7.3)
with Wi′ = (W0i , W1i ), where W0i denotes the shared random effect and
W1i Xij1 the random interaction term with covariate Xij1 , which could be, for
example, a binary treatment variable. Here, the covariate Xij1 is usually a
component of the vector Xij . The random effects W0i and W1i are assumed
to follow a bivariate Gaussian distribution with
     2 
W0i 0 σ0 ρσ0 σ1
∼N , .
W1i 0 ρσ0 σ1 σ12
There are two sources of heterogeneity between clusters included in the model.
First, heterogeneity between clusters may arises from the treatment itself,
meaning the treatment is stronger in some clusters than in others. Such
treatment by center interaction may reflect unobserved differences in patient
characteristics and in implementation of the study protocol. This is accounted
for by the random interaction between treatment and cluster. Second, the
variation in outcomes between clusters may be attributed to differences in the
baseline hazards reflecting, for example, differences in medical practices or in
patient populations. Such kind of heterogeneity can be accounted for by the
shared frailty term in the random effects model. A similar model based on
the AFT approach can be found in Komárek et al. (2007).
This model was considered with independent random effects by several
authors. It can be estimated by using an extension of the REML approach
suggested by McGilchrist (1993) to accommodate for two random effects
(Yamaguchi and Ohashi 1999). Vaida and Xu (2000) used a Monte Carlo
EM algorithm with MCMC sampling in the E step. Legrand et al. (2005,
2006) proposed a Bayesian approach.
Ripatti and Palmgren (2000) used the penalized partial likelihood approach
combined with a Laplace approximation in a model with dependent random
effects. Massonnet et al. (2008) applied a transformation technique to estimate
the parameters in this model. Rondeau et al. (2008) suggested a penalized
marginal likelihood method, resulting in smoothed estimates of the hazard
function. Allowing for a possible correlation between the two random effects,
the heterogeneity of the baseline hazard between clusters may depend on the
heterogeneity of the treatment effect. This seems to be reasonable because the
treatment effect is usually expected to be larger in clusters with high baseline
hazard, and vice versa.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
226 Frailty Models in Survival Analysis

7.2 Cox Model with General Gaussian Random Effects

A very general and flexible survival model with Gaussian random effects is
considered by Vaida and Xu (2000), and Ripatti and Palmgren (2000). The
model is given in terms of the conditional hazard

′ ′
Xij +Yij
µ(t|Xij , Wi ) = µ0 (t)eβ Wi
, (7.4)

where Xij and Yij denote the covariate vectors for the fixed and random
effects. Wi is the random effect from the ith cluster. The random effects are
assumed to follow a Gaussian distribution with expectation vector zero and
some covariance matrix. Vaida and Xu (2000) considered the case of diagonal
covariance matrices, resulting in independent random effects. In (7.4) Yij is
often a subset of Xij , apart from possibly an ”1” which represents the cluster
effect on the baseline. For a better understanding, assume for a moment that
′
Yij = (1, Xij ) and Wi′ = (W0i , W1i ). Each component of W1i represents
the interaction between the cluster and the respective covariate, while the
corresponding element of the parameter vector β represents the main effect
of the covariate. W0i is the random effect shared by all individuals in the ith
cluster. Vaida and Xu (2000) obtain maximum likelihood estimates of the
regression parameters, the variances of the random effects, and the baseline
hazard function via a modified EM algorithm. In the E step, MCMC methods
are used for the calculation of conditional expectations of functions of the
random effects. Ripatti and Palmgren (2000) used Laplace approximation
of the likelihood function in combination with a penalized partial likelihood
approach, where the marginal distribution of the random effects determines
the penalty term.
The flexibility of the above model is demonstrated best by discussing some
of its interesting submodels. The univariate model by Li et al. (2002) given
in (7.1) with parameter specification α = 1 is a special case of this model with
′
cluster size one, Yij = (1, Xi ), Wi′ = (Wi , Wi ), and Xij = Xi as a binary
covariate. The univariate log-normal frailty model with general covariates is
obtained with cluster size one, Yij = 1, and a one-dimensional random effect
Wi . Furthermore, the model considered by Xu (2004) and specified in (7.2)
is a special case with Yij = Xij , Wi as a one-dimensional random effect,
and Xij as a binary variable. The above model simplifies to the shared log-
normal frailty model with Yij = 1 and a one-dimensional random effect Wi .
Furthermore, the interaction model (7.3) considered in the last section is a
′
submodel via Yij = (1, Xij1 ) and Wi′ = (W0i , W1i ).
In the following section, nested (hierarchical) frailty models are considered.
It is necessary to note that nested frailty models in general are not included
in model class (7.4).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Different Aspects of Frailty Modeling 227

7.3 Nested Frailty Models

The shared frailty model is often used to model clustered survival data when
the compound symmetry assumption seems appropriate. Multilevel clustered
failure time data arise when the clustering of data occurs at more than one
level. In studies about the relationship between environmental factors such
as air pollution and mortality, for example, individuals are grouped in cities,
and furthermore, in subclusters like districts. The hypothesis of independent
observations is questionable in this setting, so a flexible survival model with
two nested random effects at city and district level seems appropriate. Other
fields of application are multicenter studies with recurrent events, where center
is one cluster level and patients form the nested subclusters. Ignoring the
additional subgroups may invalidate the results of data analysis. Thus, nested
frailty models are of interest in problems in which data are naturally clustered
at different hierarchical levels (e.g., families in different cities), or generated
by a hierarchical sampling design (e.g., districts in cities). Sastry (1997)
suggested a nested frailty model with two hierarchical levels

′
µ(t|Xijk , Zi , Zij ) = Zi Zij µ0 (t)eβ Xijk
, (7.5)

where Xijk denotes the covariate vector of the kth individual in the jth sub-
cluster belonging to the ith cluster. Frailty Zi denotes the cluster-specific
random effect, and Zij is the subcluster-specific random effect. Both Zi and
Zij are independent gamma-distributed random variables with mean one and
variances σ12 and σ22 , respectively.
Nested frailty models account for the hierarchical clustering by including
cluster-level-specific random effects. However, the extension of the estimation
procedures of the shared frailty model to the case of two or more frailty
terms is difficult. Sastry (1997) first adapted the EM algorithm to the nested
frailty model. Manda (2001) used Bayesian methods in a nested gamma
frailty model with piecewise constant baseline hazard. Rondeau et al. (2006)
applied a semiparametric maximum penalized likelihood estimation procedure
to model (7.5), which provides as a by-product a smooth estimator of the
hazard function.
More complicated nested frailty models allowing for dependence between
the cluster and subcluster-specific frailties are considered by Ma et al. (2003)
using a Poisson modeling approach, and by Shih and Lu (2009). Yau (2001)
suggested a nested log-normal frailty model of the type

′
µ(t|Xijk , Wi , Wij ) = µ0 (t)eβ Xijk +Wi +Wij (7.6)

with similar notation as above. Wi , Wij are independent normally distributed

random effects with mean zero and variances s21 and s22 , respectively.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
228 Frailty Models in Survival Analysis

7.4 Recurrent Event Time Data

Up to now, in all cases were parallel event times considered. However, one
important field for application of frailty models is the analysis of recurrent
event time data. A common feature of recurrent event time data is that
the events are naturally ordered and occur in a certain sequence over time.
Reviews of models for recurrence data appeared recently in Cook und Lawless
(2007) and Finkelstein (2008) with links to the field of reliability. However,
there is a growing interest in models for recurrence data in other fields also.
Medical examples of such data include cancer recurrences, asthma attacks, in-
fections, admissions to hospital, and headaches. To analyze such type of data,
many researchers estimate and compare event rates using the chi-square test
or event times until the occurrence of the first event, or the overall event time
using the Cox model. Such conventional methods are inefficient because they
use only parts of the available information in the data. If data have repeated
events with censored event times in longitudinal studies, complex analytic
approaches are needed to obtain accurate estimates and efficient inferences.
These models allow the use of all available information to accurately estimate
the relative risk of recurrence. We will not go into much detail, but would like
to mention the most common approaches. Therneau and Grambsch (2000)
distinguishes three categories: independent increment, marginal, and condi-
tional models. In all models, observations from one individual are grouped to
one cluster. The main difference between models is the definition of risk sets
and stratification of the baseline hazard function.
Andersen et al. (1982) proposed a model assuming independent increments.
Here, individuals are assumed to be under risk for all their events until
censoring. The baseline hazard is not stratified. This means that the risk of
an event is unaffected by the occurrence of any previous event. The numbers
of events in disjoint time interval are independent. For a parametric baseline
hazard, this model is equivalent to the nonhomogeneous Poisson process. As
an alternative, the gap times between events can be considered. Assuming
that these times are independent and identically distributed, and the baseline
hazard is parametric, the model reduces to a renewal process with covariates.
The well-known marginal model was introduced by Wei et al. (1989). All
individuals are assumed to be under risk for a future event until the occurrence
of this event or censoring. The baseline hazards may be stratified with respect
to the event number. To account for correlation between the observed events,
the variance is adjusted without explicitly modeling the correlation. As with
the analysis of longitudinal data, regression parameters are estimated from
generalized estimating equations, and the corresponding variance–covariance
estimators are corrected properly to account for the dependence structure.
An excellent review of the robust and well-developed marginal approach is
given in Lin (1994).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Different Aspects of Frailty Modeling 229

The model introduced by Prentice et al. (1981) measures the conditional risk
of experiencing an event. An individual is only at risk for an event from the
occurrence of the previous event until the occurrence of this event or censoring.
Therefore, an event-specific baseline hazard is assumed. In general, the model
by Prentice et al. (1981) is approximating the real situation best.
In the models by Prentice et al. (1981) and Wei et al. (1989), in-subject
correlation is accounted for through stratification of the baseline hazard. In
the Andersen–Gill model, independent increments are assumed. All models
can be expanded to allow for between-subject correlation. As proposed by
McGilchrist (1991), a subject-specific frailty term can be used in the hazard.
For all observations belonging to one subject, the same frailty term is used.
The frailty variables of different subjects are independent realizations of a
common frailty distribution.
Duchateau et al. (2003) applied different parametric and nonparametric
models, both with and without frailty, to recurrent asthma events from an
asthma prevention trial in young children. One problem, however, is the
decision about the time scale to use. In the gap-time approach, only the time
until the occurrence of the last event is considered. An alternative time scale
is total time, measuring the time from the beginning of follow-up, disregarding
other events having occurred meanwhile.
Frailty models specially designed for recurrence data are considered in detail
by Aalen et al. (1995), McGilchrist and Yau (1996), Yau and McGilchrist
(1998), and Manda and Meyer (2005). An overview in the field is given in
Duchateau et al. (2003) and Lim et al. (2007).
The observation of recurrent event times could usually be terminated by loss
to follow-up, end of study, or a terminal event such as death. In the analysis
of recurrent even times discussed above, the assumption of noninformative
censoring of the recurrent event process by death is made, which can be
violated. For example, the recurrence of serious events, such as tumors or
opportunistic infections, is often associated with an increased risk of death.
To circumvent this problem, joint frailty models provide a useful alternative.
Here, the recurrent event hazard and the death hazard are modeled separately
but share some random effect, causing dependence between the recurrent event
times and the lifetime. The model can be specified by the hazard functions
′
µ1 (t|Xi1 , Zi ) = Zi µ01 (t)eβ1 Xi1 (7.7)

µ2 (t|Xi2 , Zi ) = Ziγ µ02 (t)eβ2 Xi2 ,

′
(7.8)
where µ1 denotes the hazard of the recurrent events and µ2 is the hazard of
the terminal event. When γ = 0, then µ2 does not depend on the frailty
and is independent of the recurrent event times. If the frailty distribution
is degenerated, there is no unobserved heterogeneity in the data, and the
recurrent event times become independent. Liu et al. (2004) and Rondeau et
al. (2007) assume a gamma distribution of the frailty.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
230 Frailty Models in Survival Analysis

7.5 Tests for Heterogeneity

In the practical application of methods from survival analysis, the researcher
is usually confronted with the question of whether unobserved heterogeneity is
present in the data or not. For this problem a number of tests of heterogeneity
have been proposed during the last 15 years. The relevant testing problem
for the null hypothesis of no unobserved heterogeneity is H0 : σ 2 = 0 versus
HA : σ 2 > 0, where σ 2 denotes the frailty variance (unobserved heterogeneity).
For ease of presentation, the case of infinite moments of the frailty distribution
(positive stable distribution) is excluded here. In this situation, a one-sided
testing hypothesis is natural. The main problem is that the heterogeneity
parameter lays on the boundary of its parameter space under H0 , and the
classical likelihood ratio asymptotic distribution theory is no longer valid. It
turns out that, in many cases, the likelihood ratio statistic has an asymptotic
distribution under the null hypothesis, which is an equal mixture of a point
mass at zero and a chi-square distribution with one degree of freedom, often
denoted by 0.5 χ20 + 0.5 χ21 . Rigourously established limiting distributions of
the likelihood ratio test statistic are provided by Maller and Zhou (2003) and
Claeskens et al. (2008). Other generally more applied results about tests for
heterogeneity in survival data can be found in Oakes (1982), Blossfeld and
Hamerle (1989, 1992), Hamerle (1992), Commenges and Andersen (1995),
Gray (1995), Commenges and Jacqmin-Gadda (1997), Verweij et al. (1998),
Bordes and Commenges (2001), Dunson and Chen (2004), and Dominicus et
al. (2006).
Andersen et al. (1999) used the test suggested by Commenges and Andersen
(1995) and found advantages of the mixed model when compared with the
fixed-effects model, especially in the case where it is known to which subgroup
the individuals belong if the number of subgroups in the total population is
large and the number of individuals in each subgroup is only moderate. They
applied the test to a multicenter clinical trial with survival outcome.
A test procedure for proportional hazards in the presence of unobserved
heterogeneity is suggested by McCall (1994). Kimber and Zhu (1999) consider
the situation where the conditional lifetimes are Weibull, and frailty follows
a (shared) positive stable model. For this very specific situation, simple tests
for frailty are proposed. Kimber (1996) already considered the case of gamma-
distributed frailty and Weibull baseline.
A pseudo likelihood ratio test is proposed by Glidden (2000) and Andersen
(2005) to test the assumption regarding heterogeneity distribution. In PVF
frailty models, the gamma frailty model and the positive stable frailty model
are at the boundary of the parameter space. Therefore, as discussed in
Andersen (2005), the classical likelihood ratio test is not valid. Andersen
(2005) gives the asymptotic distribution for the pseudo likelihood ratio test
(but without rigorous proof).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Different Aspects of Frailty Modeling 231

7.6 Log-Rank Test in Frailty Models

In analyzing time-to-event data of two (or more) groups with the log-rank
test, a proportional hazards model is usually assumed. Nevertheless, this
hypothesis is no longer valid in the case of unobserved covariates. Omission
of a balanced covariate from a proportional hazards model generally leads to a
model with nonproportional hazards. As a consequence, the simple log-rank
test is no longer optimal. A general treatment of this problem is given in
Oakes and Jeong (1998). The authors establish connections between theories
of weighted log-rank tests and of frailty models, and analyze parametric as
well as nonparametric models. Furthermore, the authors consider different
frailty distributions: positive stable, inverse Gaussian, gamma, and two-point
distribution and the models are extended to include random censoring. The
optimal limiting weight function is
 L′′ (M (t)) L′ (M (t)) 
0 0
w(t) = 1 + M0 (t) − ,
L′ (M0 (t)) L(M0 (t))
where L′ and L′′ are the first and second derivative of the Laplace transform
L of the frailty distribution. For example, in the gamma frailty model, Oakes
and Jeong (1998) derived the expression
w(t) = (1 + σ 2 M0 (t))−1
(corrected for an error in their formula and translated to our notation). The
asymptotic relative efficiency of the simple log-rank test and of the optimally
weighted log-rank test relative to the adjusted test that would be used if
the covariate values were known are given in terms of the Laplace transform
of the frailty distributions. These frailties represent the effect of the missing
covariates. Two main conclusions are derived by Oakes and Jeong (1998) from
their analysis. First, the loss of efficiency from omitting a covariate is generally
more important than the additional loss due to the misspecification of the
resulting nonproportional hazards model. Furthermore, the loss of efficiency
increases in general with higher variance of the frailty distribution. However,
binary frailty distributions can provide exceptions to this rule, reflecting
the fact that, for a very extreme frailty distribution, one can identify the
values of the frailties from the ordering of the failures; the individuals who
experience the event of interest first are almost certain to have the higher
(of the only two possible) values of frailty. Consequently, the knowledge
of the values of the frailties (which means the values of missing covariates)
becomes less important. Second, censoring tends to remove the late events
and, consequently, reduces the loss of efficiency in the simple log-rank test
compared to the adjusted log-rank test. The above results depend on the
knowledge of the usually unknown frailty variance σ 2 ; consequently, their
practical applicability is limited.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
232 Frailty Models in Survival Analysis

Jeong (2003) extends the foregoing results to visualize and quantify the loss of
efficiency of the log-rank test when a dependence structure between survival
and censoring times is being ignored. The assumed dependence structure is
based on a correlated gamma frailty model (Section 5.2). In the given situation
the loss of efficiency is minimal under the proportional hazards model, even
when the correlation between potential survival and censoring times is strong,
unless the dependent censoring causes a severe nonproportionality.
Broe̋t et al. (1999) suggest a more practical approach for taking into account
unobserved covariates in a weighted log-rank test. The construction of the test
is based on a gamma distribution of the frailty. For the frailty parameter σ 2 ,
a range is assumed. Simulations investigate the power of the test for different
frailty distributions.
Rank tests for clustered event time data when dependent subunits are
randomized are considered by Jeong and Jung (2006).
Li et al. (2002) arrive at the conclusion that the log-rank test is nearly
fully efficient relative to the optimally weighted log-rank test if the unobserved
heterogeneity does not interact with the binary treatment variable. This is
not the case if frailty interacts with treatment. Such situation could happen,
for example, if unobserved heterogeneity is caused by genetic factors, which
interact with treatment.

7.7 Time-Dependent Frailty Models

Standard frailty models assume that the individual frailty Z is determined
at time zero (begin of follow-up) and follows the individual throughout the
rest of the life, resulting in the individual’s susceptibility (frailty) later in life
being perfectly correlated with that at the beginning. It is clearly of interest
to overcome this restrictive assumption of fixed frailty in order to develop
more flexible models. However, it turns out that this is a difficult problem.
There exist different approaches dealing with time-dependent frailty by
considering the individual hazard as a stochastic process. This provides
individual flexibility. However, since what is observable is the average hazard
function, the problem is to find models that show a tractable mathematical
connection between the individual hazard function and the average one. One
attempt to tackle this problem was suggested by Yashin and Manton (1997),
defining the hazard function as a diffusion process by means of a stochastic
differential equation. By using Ornstein–Uhlenbeck-type processes, such kind
of models reflect the fact that many biological parameters tend to stabilize
around certain values, which is called homeostasis. The quadratic hazard
frailty model in Section 3.9 is a special case of this model with fixed frailty.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Different Aspects of Frailty Modeling 233

In a second approach, McGilchrist and Yau (1996), and Yau and McGilchrist
(1998), focus on recurrent event time data. Often interrecurrence times close
to each other on the time scale are highly associated, while times that are
further apart from each other on the time scale are less correlated. To model
such kinds of serial dependence, a dynamic frailty model can be constructed
by assuming that the frailties of subsequent time intervals follow an autore-
gressive process of order one, denoted by AR(1). Yau and McGilchrist (1998)
adopted AR(1) frailty models to analyze chronic granulomatous disease data.
A similar but simpler idea without using an AR(1) process was suggested by
Wintrebert et al. (2004). A Bayesian approach to the problem can be found
in the paper by Manda and Meyer (2005).
The third approach is based on Lévy processes (Gjessing et al. 2003,
Aalen et al. 2008). Like diffusion processes, hazard functions driven by Lévy
processes also yield some degree of tractability. It is possible to get explicit
formulas for the relationship between conditional and unconditional hazard,
and the frailty and hazard of survivors may be estimated. A basic difference
between this and the first approach is the jump nature of the Lévy processes.
However, it could well be imagined that the individual hazard may increase
in jumps, for example, with the onset of an acute disease. The model is an
extension of Lévy frailty models with fixed frailty in Aalen and Hjort (2002),
discussed in Section 3.10.

Example 7.1
Gamma processes: Let the value of the process at time point u be gamma
distributed with shape parameter ku and scale parameter λ. Then it holds
that

s −ku


L(s; u) = 1 + = e−ku ln(λ+s)−ln λ , k, λ > 0,
λ


with ψ(s) = k ln(λ + s) − ln λ as characteristic exponent (Section 3.6).

Example 7.2
Standard compound Poisson processes: A Poisson process of rate ρ is
running on time scale u, and to each jump there is a gamma random variable,
independent of the past, with shape parameter k and scale parameter λ. The
compound Poisson process is the sum of the gamma variables up to time u.
The Laplace transform of the random value of the process at time u is given
by


k
(λ+s)γ −λγ
L(s; u) = e−u γ , γ < 0.
k
(λ + s)γ − λγ .


Hence, the characteristic exponent is given by ψ(s) = γ

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
234 Frailty Models in Survival Analysis

first individual second individual

first time interval Zi11 = Yi11 + Yi Yi + Yi21 = Zi21

second time interval Zi12 = Yi12 + Yi Yi + Yi22 = Zi22

Figure 7.1: Time-varying frailties of two relatives from family i in two time
intervals in the Paik model.

Example 7.3
Stable processes: The Laplace transform of a stable distributed random
k γ
variable takes the form L(s; u) = e−u γ s . Hence, ψ(s) = γk sγ .

Example 7.4
PVF processes: The power variance function distributions constitute a
general class of distributions. A class of Lévy processes may be defined from
the PVF distributions by ψ(s) = kγ (λ + s)γ − λγ (γ > 0). The special case

γ = 0 is defined by continuity and gives the gamma process. For γ < 0, the
model yields the standard compound Poisson process.

Paik et al. (1994) suggest another extension of the correlated gamma frailty
model. They consider the case of clustered individuals but with varying
frailties over time. For this purpose, the time scale is divided into intervals.
The approach uses a decomposition of the frailty in each interval into a shared
and a unique part, for example, Zijl = Yi + Yijl with independent random
variables Yi ∼ Γ(k1 , λ1 ) and Yijl ∼ Γ(k2 , λ2 ). In the correlated gamma frailty
model, the assumption λ1 = λ2 is essential because it yields the gamma
distribution of Z1 and Z2 . Paik et al. (1994) allow the λ’s to differ which,
on the one hand, complicates the computation of the likelihood function, but
on the other hand, it generates a more flexible model. In both models the
expectation of the frailties is restricted to one. The main feature of the model
in Paik et al. (1994) is to allow the frailty (and, consequently, the dependence
function) to vary over time intervals. Restricting the model to two related
individuals with two recurrent observation times each gives the structure of
frailties shown in Figure 7.1.
Note that Paik et al. (1994) considered a noncompeting risk situation,
which means it is assumed that all lifetimes are observable (with independent
censoring). Different extensions of the Paik model are discussed in Wintrebert
et al. (2004) and Wintrebert (2007).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Different Aspects of Frailty Modeling 235

7.8 Identifiability of Frailty Models

One of the aims of frailty modeling is to study the properties of the hazard
function and the frailty distribution in real-life applications. Such analysis
is possible if the frailty distribution, as well as the hazard function, are
identifiable from the survival data. There is a substantial literature on the
identification of frailty models. Especially in the econometric literature (where
these models are named mixed proportional hazards models), the problem of
identifiability is considered in detail for univariate as well as multivariate
models.

7.8.1 Univariate frailty models

It is useful to consider the frailty model in a more general form based on the
general proportional hazards model (2.9):

µ(t|X, Z) = Zµ0 (t)h(X).

It is possible to consider this model in a completely nonparametric form,
for example, without any distributional assumptions about the cumulative
baseline hazard M0 , the form of the function h, and the distribution of
the frailty term Z (for example, given by their distribution function FZ ).
We would like to point out that this situation is more general than the
situation considered in the present monograph, where usually the function
′
h is specified by h(X) = eβ X and distributional assumptions are made
about the frailty distribution. So the literature about identifiability of frailty
models is (at least partially) concerned with nonparametric identifiability. It
is useful to determine identifiability as a property of the mapping from the
three determinants (M0 , h, FZ ), given their domain, to the data. The frailty
model is now specified by the unique mapping from the domain to the data.
The model is identifiable if this mapping has an inverse, meaning for given
data there exists a unique set of functions (M0 , h, FZ ) in the domain that is
able to generate these data. We will not go into much detail here, but refer
the reader interested in identifiability results to Elbers and Ridder (1982),
who were the first to prove nonparametric identifiability of frailty models.
Their proof is not constructive. Constructive identification proofs are more
attractive because they express the underlying functions (M0 , h, FZ ) directly
in terms of observable quantities and therefore suggest an estimation method.
Constructive proofs are provided by Heckman and Singer (1984b), Melino
and Sueyoshi (1990), Ridder (1990), Heckman (1991), Heckman and Taber
(1994), Kortram et al. (1995), and Horowitz (1999) under different model
assumptions. For an overview, see van den Berg (2001). The most important
assumption here is EZ < ∞. In the case of nonidentifiability, the adoption
of a model that is observationally equivalent to (but different from) the true

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
236 Frailty Models in Survival Analysis

model leads to biased parameter estimates. This is a difficult problem because

it is usually hard to make justified assumptions about the frailty distribution
based on the observed survival data.
The fundamental identification problem was addressed by Ridder (1990)
in detail. He showed that, for every frailty model with EZ < ∞, there
exists an observational equivalent model with EZ = ∞. Therefore, the
assumption EZ < ∞ imposes identifiability on a class of models that are
not identifiable. The identifiability of the frailty distribution is a consequence
of the nonproportionality of the unconditional hazards (Hougaard 1991, van
den Berg 1992, Keiding 1998). Heckman and Taber (1994) and Kortram et
al. (1995) show that the mapping from the data generating process to the
data is not continuous, which results in the problem that two different frailty
models can generate very similar data. Furthermore, convergence rates are
sometimes very small, for example, convergence rates for the estimation of
the baseline hazard λ and regression parameters β can be at most almost
equal to n−2/5 under some regularity assumptions. For the estimation of the
frailty distribution, the rate of convergence is extremely small with (log n)−2
(Horowitz 1999). Consequently, based on sample sizes occurring in practical
applications, the estimation in this nonparametric univariate frailty model
is often very unstable, and interpretations based on such results should be
performed with extreme caution.
One solution to the foregoing identifiability problem is to impose parametric
assumptions about the frailty distribution as done in the present monograph.
Let us now return to this situation, and let S(t) represent some survival
function. It turns out that S(t) can always be represented as a marginal
survival function in some model with unobserved frailty Z and cumulative
baseline hazard M0 (t). In other words, univariate frailty models are not
identifiable if only the marginal distribution of the failure times (without
any observed covariates) are known. Lancaster and Nickell (1980) discussed
this identifiability problem in detail. Indeed, let L be the Laplace transform
of some arbitrary distribution of the nonnegative random variable Z, and let
L−1 be the inverse function of L. For any 0 ≤ t < ∞, the cumulative hazard
is defined as

M (t) = L−1 (S(t)).

To prove that M (t) is a cumulative hazard function of a survival distribution,

it is sufficient to show that M (t) is a nondecreasing function. This follows from
the fact that both S(t) and L−1 (t) are nonincreasing functions of t. Thus, for
any given survival function S(t) and probability distribution on the positive
half-line (given by Laplace transform L), one can construct a frailty model
with marginal survival function S(t). The nonidentifiability aspect of specific
distributions of frailty was discussed by Heckman and Singer (1982a, 1984a)
and Hoem (1990), among others. Elbers and Ridder (1982) show that if the
frailty distribution has a finite mean, then the presence of covariates makes the

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Different Aspects of Frailty Modeling 237

univariate proportional hazards frailty model identifiable. Other conditions

are given by Heckman and Honoré (1989). These findings stimulated the use
of frailty models with observed covariates (regression models) in the analysis
of univariate survival data (Andersen et al. 1993).
In the more general case of a frailty cure model (modeling a cure fraction
and unobserved heterogeneity together), it holds that

′
S ∗ (t|X1 , X2 ) = 1 − φ(X2 ) + φ(X2 )L(M0 (t)eβ X1
). (7.9)

Peng and Zhang (2008b) establish mild conditions that imply identifiability of
the model. Here, L again denotes the Laplace transform of a frailty variable.
Two kinds of covariates are considered: X2 influences the probability to be
cured, whereas X1 influences the event time in susceptible individuals. The
identifiability of the frailty model is obtained as a special case but with a
much simpler proof compared to the one in Elbers and Ridder (1982). The
above frailty cure model with covariates in the cure fraction and the survival
function of the noncured population is identifiable if the cure fraction φ(X2 ) is
nonconstant and some technical assumptions such a finite mean of the frailty
distribution are fulfilled. It turns out that the case of identical covariates
X1 = X2 needs additional assumptions and is therefore treated separately by
Peng and Zhang (2008b).

7.8.2 Multivariate frailty models

Bivariate (as well as multivariate) data present an opportunity to identify
both the frailty distribution as well as the baseline hazard under much weaker
assumptions than in the univariate case. For example, all shared frailty
models are identifiable from bivariate data without additional information
such as observed covariates or parametric assumptions about the baseline
hazard as shown by Honoré (1993). Furthermore, the baseline hazard may
depend on observed covariates X in an unspecified way, and frailty Z and
covariates X may be dependent. It turns out that the identifiability property
holds for a broader class of frailty models, including correlated frailty models.
Furthermore, the condition of a finite mean required for identifiability in a
univariate frailty model with covariates by Elbers and Ridder (1982) is not
needed in this case. For more details of this situation, see Yashin and Iachine
(1999b). Identifiability aspects of the correlated frailty model were analyzed
in more detail by Iachine and Yashin (1998), and were further extended by
Iachine (2004), with the latter paper discussing an interesting counterexample
based on the positive stable frailty model.
Nielsen et al. (1992, end of Section 4) mentioned in the shared gamma
frailty model the need for a positive probability of at least two failures per
cluster to ensure identifiability.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
238 Frailty Models in Survival Analysis

7.9 Applications of Frailty Models

This section provides a collection of examples from the literature dealing with
interesting frailty models usually not covered by the approaches considered
in the foregoing chapters. The main emphasis is to give an overview of the
variety of applications of frailty models in the solution of different statistical
problems. The presentation of the material here is kept very short, and the
reader is referred to related publications.
A model combining relative survival and frailty is given in Goeman et al.
(2004). The authors look at the excess hazard for a specific disease when a
concurrent disease is present. The main idea of this model is a hazard function
of the form

µ(t|Z) = µ0 (t) + Z[µ1 (t) + µ2 (t)].

Here, µ0 is a background hazard function from the risk of death by other causes
assumed to be known (the general population mortality statistics if the disease
under study contributes only a small part to overall mortality), µ1 describes
the excess hazard of the concurrent disease (assumed to be known from the
literature), and µ2 is the excess hazard caused by the primary disease. Results
are shown to be independent of the chosen frailty distribution if heterogeneity
is small. The model is applied to data on patients with head/neck tumors
using information on previous tumors.
Zahl (1997) considers a similar problem by applying different versions of the
shared and correlated gamma frailty model to cause-specific mortality data in
Norway in order to model the excess hazard of cancer mortality. The study has
the drawback that, because of a small number of cases, a reasonable estimation
of the parameters was not possible. Nevertheless, the paper demonstrates
the advantage of correlated frailty models in cause-specific mortality data
(competing risks).
O’Quigley and Stare (2002) discuss some of the fundamental concepts of
univariate and shared frailty models. The authors claimed that univariate
frailty models are of limited practical value. Any frailty model S(t|X, Z) is
equivalent to a usually nonproportional hazards model S ∗ (t|X), in which the
individual effects disappear:
Z ∞
S ∗ (t|X) = S(t|X, z) fZ (z) dz
0

which would correspond to a homogeneous population. Keeping this in mind,

any assertion about unobserved heterogeneity must be entirely arbitrary. The
assertion is fully determined by an arbitrary model choice. The statistician
who chooses S(t|X, Z) will conclude that heterogeneity is clearly evident.
The statistician who decides for S ∗ (t|X) will come to the opposite conclusion.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Different Aspects of Frailty Modeling 239

Consequently, both conclusions are arbitrary; the model we choose determines

our conclusion. The data themselves will not allow us to decide which model is
the true one. Decisions about the two situations cannot be made by looking
at the data without making further strong, parametric assumptions about
the baseline hazard or the frailty distribution. A test for the presence of
frailties is equivalent to a test of fit against time-varying covariate effects.
Unobserved heterogeneity cannot be distinguished from homogeneity under a
nonproportional hazards model. Consequently, O’Quigley and Stare (2002)
conclude that univariate frailty models cannot be used to incorporate the
effects of unobserved covariates without additional information.
In contradiction to the univariate case, the authors state that, in the multi-
variate case, frailty models can be used to advantage; especially for moderate
to large numbers of small families (up to five members), the efficiency gains of
frailty models could be important. For the analysis of data on large families,
the authors recommend the use of stratified proportional hazards models.
This is possible under the assumption of shared frailties in families, which
implies proportional hazards even when the frailty is unknown.
Warwick et al. (2003) used a modified gamma frailty model with piecewise
constant baseline hazards to model time-dependent effects on survival in
breast cancer. It was shown that the effects of tumor size, lymph node
status, and histologic grade are important predictors still after 20 years of
follow-up, but the effects of these risk factors progressively diminish with
time from prognosis. Frailty models are one possibility to allow for such kind
of attenuation of prognostic effects.
Dos Santos et al. (1995) applied different shared frailty models with Weibull
baseline hazard to the recurrence of breast cancer. The data set contains
917 women treated for breast cancer between 1980 and 1985 at the Christie
Hospital in Manchester (U.K.) and their subsequent monitoring until 1991.
The authors used two explanatory variables: age and stage of the disease at
the initial treatment. The outcome of interest is the duration to recurrence
after treatment. In modeling repeated durations, the covariate state (a factor
with tree levels) was included. Death from other causes was treated as
right censoring and therefore assumed to be noninformative. Four parametric
models are considered in the analysis: a Cox model and three frailty models
with gamma, log-normal, and four-point frailty. There is a considerable
increase in the log-likelihood when a frailty term is included in the model.
Although formal likelihood ratio tests are not appropriate, the discrete four-
point frailty model shows a better fit to the data compared to the gamma and
log-normal frailty models, respectively.
Guo and Rodriguez (1992) and Guo (1993) applied a shared frailty model
with observed covariates to child survival data in Guatemala, using gamma
and binary frailty distributions and piecewise constant hazards. The authors
compare their results with an analysis based on a proportional hazards model
with piecewise constant hazard function. In all three models, the estimates of
the observed covariates effects are remarkably stable.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
240 Frailty Models in Survival Analysis

Moreno (1994) analyzed the influence of frailty distribution on the association

between survival times under gamma, inverse Gaussian, discrete k-point,
and Poisson distributions. He shows that the pattern and strength of this
association depends on the specification of the frailty distribution.
Banerjee et al. (2003) applied a spatially correlated frailty model to infant
mortality data in Minnesota. Each county in Minnesota forms a cluster,
with common frailty for all observations in a cluster (shared frailty model).
Additionally, the frailties of different clusters are not independent as in the
classical frailty models. A correlation matrix for the frailties is used to model
spatial correlations between clusters (correlated frailty model). Other spatial
frailty models are investigated, for example, by Banerjee and Carlin (2003)
and Jin and Carlin (2005).
In dental research, researchers are often faced to the problem of clustered
event times. However, applications of frailty models to patient-oriented dental
research are sparse. Kalwitzki et al. (2002) considered times until occurrence
of caries. Sealant retention was analyzed by Morgan et al. (2002) using a
Poisson frailty model. Chuang et al. (2005) and Chuang and Cai (2006)
evaluated lifetimes of dental implants. Wong et al. (2006) applied a multi-
level log-normal frailty model to lifetimes of treatment restorations of carious
lesions of Chinese school children.
More general censoring and truncation patterns in univariate frailty models
than left truncation and right censoring, as considered in the present book,
are treated by Huber-Carol and Vonta (2004).

7.10 Software for Frailty Models

The applicability of newly developed statistical methods depend strongly on
their availability in common statistical packages. The proportional hazards
model is provided by all standard statistical packages, and the regression
parameters in this model can be estimated consistently by maximization of a
partial likelihood function (Cox 1975) that does not depend on the baseline
hazard. The baseline hazard can be estimated nonparametrically. This is one
of the great advantages of the proportional hazards model, but it does not
carry over to frailty models. Here, more sophisticated estimation procedures
are needed. Unfortunately frailty models are not standard tools provided
by large standard packages such as SPSS and SAS as a standard routine.
There exists different procedures freely available, such as COXPH, COXME,
PHMM, FRAILTYPACK, SPGAM, or SPNL3. This makes it difficult to
find the appropriate tool for the specific problem considered. Advantages
and limitations of the procedures are discussed, and a short overview of the
software for different frailty models is given in this section.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Different Aspects of Frailty Modeling 241

7.10.1 R packages
R is a free software environment for statistical computing and graphics. It
runs on a wide variety of UNIX platforms, Windows, and MacOS and has
become more and more attractive to statisticians as well for nonstatisticians.
The software can be assessed on http://cran.R-project.org.
FRAILTYPACK: This package by Rondeau and González (2005) was
mainly extended by Rondeau et al. (2010). It fits semiparametric gamma
frailty models. It also provides parameter estimation in more complex settings
such as the interaction model (7.3) with two dependent Gaussian random
effects. Also, the nested frailty model (7.5) with two gamma-distributed
frailties can be fitted. Furthermore, the joint modeling of recurrent events and
a terminal event by a shared gamma frailty (7.7-7.8) is implemented in the
package. It accommodates left-truncated and right-censored data. Stratified
analysis with two strata is possible. FRAILTYPACK uses the penalization
of the hazard function described by (3.34), applying the robust Marquardt
algorithm, which is a combination between a Newton–Raphson and a steepest-
descent algorithm. As a consequence of this estimation procedure, a smooth
estimate of the hazard function is provided.
COXME: The procedure COXME was created by Therneau and works
with the penalized partial likelihood algorithm. It provides estimation in
the general Cox proportional hazards model with Gaussian random effects
described by model (7.4). Furthermore, COXME is able to handle nested
frailty models (7.5). The package was originally a function in the kinship
package in recognition of the fact that it was primarily targeted toward genetic
problems. COXME was also distributed as part of the base survival package in
S plus. The current version from 2009 is more broad in its capabilities. It turns
out to be useful to modify the convergence criteria eps from 10−6 to 10−9 and
the maximum iteration number for the partial likelihood procedure iter.max
from 20 to 40. This increases the precision of the estimation procedure to a
reasonable size but at the cost of computer time.
SURVIVAL: This package was also written by Therneau for survival data.
One function of the package is COXPH. It fits the proportional hazards
model and is able to deal with time-dependent covariates and strata, multiple
events per subject, and other extensions. One of these extensions include
univariate and shared frailty models. The syntax is similar to that of the
COXME procedure. The frailty distributions gamma, log-t, and log-normal
are supported (Therneau et al. 2003). Similar to the COXME procedure,
modification of the parameters eps and iter.max is recommended.
PHMM: This package steams from Donohue and Ronghui in 2009/2010.
It fits the proportional hazards model incorporating Gaussian random effects
given in (7.4). The estimation procedure is based on the EM algorithm using
Markov Chain Monte Carlo methods in the E step. The software provides no
convergence criterion; the number of iterations has to be fixed.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
242 Frailty Models in Survival Analysis

To make R and SAS procedures comparable, the option ties = ”breslow” for
the handling of ties should be used in all R procedures. Furthermore, despite
the wide range of models covered, the R packages do not provide standard
errors of the random effects variance estimates, which is a major drawback.
An exception is the recent version of FRAILTYPACK.

7.10.2 SAS packages

In SAS there are only a few macros dealing with frailty models. It is important
to note that the SAS macros provide standard errors of the frailty variance
estimates in contrast to most of the R procedures.
GAMFRAIL: This macro by Klein uses the EM algorithm (Klein 1992)
to fit semiparametric gamma frailty models and can be downloaded from
http://www.mcw.edu/biostatistics/Research/Software.htm. It suffers from a
very slow estimation procedure. The handling of the macro is not very user-
friendly; for example, the data need to be arranged in a fixed order.
PS-FRAIL: This macro steams from Shu and Klein (1999) and can be
obtained from http://www.mcw.edu/biostatistics/Research/Software.htm. It
provides estimation of the semiparametric positive stable frailty model.
SPGAM: This macro implements the ML-EM algorithm for the semi-
parametric shared gamma frailty models, described by Vu (2003). The newer
version allows for left censoring in the data. Additionally, there exists a version
for estimation in the parametric shared gamma frailty model, which is called
PGAM. The macros are available on http://lib.stat.cmu.edu.
SPNL: This is a macro to fit the semiparametric shared log-normal frailty
models with left censoring, using the ML-EM algorithm described in Vu
(2004). The syntax is similar to the SPGAM macro; just the argument N
is added. N denotes the integer that controls the precision of the integrals
involved inside the macro. It is available on http://lib.stat.cmu.edu.
A macro providing a score test for unobserved heterogeneity in event times
is available on http://www.mcw.edu/biostatistics/Research/Software.htm.

7.10.3 STATA package

STATA is so far the only large commercial statistical package that supports
frailty analysis. It is able to handle gamma and log-normal frailty models.
However, analysis is restricted to parametric models. The package provides
exponential, Weibull, Gompertz, log-normal, log-logistic, and generalized
gamma distribution of the baseline hazard function.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com

Appendix A
Appendix

The aim of this appendix is to give some statistical key results that are
important for the field of frailty models. These results are more technical
and, for that reason, moved to the appendix to allow a fluent reading of
the main parts of the book. The next part deals with bivariate event-time
data. The bivariate case is used here to illustrate the main ideas of correlated
frailty models in this simple situation. It turns out that the likelihood of
the survival data can be represented in terms of the survival function and its
partial derivatives. A model specification in terms of the survival function
eases the computational burden of handling censoring and truncation.

A.1 Bivariate Lifetime Models

In Section 2.2, the likelihood function for univariate right-censored and left-
truncated data was already derived. The present section deals with right
censoring and left truncation in bivariate failure data. For this purpose, let
(T1 , T2 , ∆1 , ∆2 ) be the censored bivariate observations with Tj = min{Tj∗ , Cj },
∆j = 1(Tj∗ ≤ Cj ), (j = 1, 2) and denoting the bivariate survival function
of (T1∗ , T2∗ ) by S(t∗1 , t∗2 ). Furthermore, denote the density of (T1∗ , T2∗ ) and
(C1 , C2 ), respectively, by f (t∗1 , t∗2 ) and g(c1 , c2 ). Assume that the lifetimes
(T1∗ , T2∗ ) and censoring times (C1 , C2 ) are independent.

H(t1 , t2 , 1, 1) = P(T1 > t1 , T2 > t2 , ∆1 = 1, ∆2 = 1)

= P(T1∗ > t1 , T2∗ > t2 , T1∗ ≤ C1 , T2∗ ≤ C2 )
ZZZZ
= f (t∗1 , t∗2 )g(c1 , c2 ) dt∗1 dt∗2 dc1 dc2
{t∗
1 >t1 ,t2 >t2 ,t1 ≤c1 ,t2 ≤c2 }
∗ ∗ ∗

Z ∞Z ∞ Z ∞ Z ∞ 
∗ ∗
= f (t1 , t2 ) g(c1 , c2 ) dc1 dc2 dt∗1 dt∗2
t2 t1 t∗2 t∗
1

Consequently, the subdensity for noncensored pairs of lifetimes is

Z ∞Z ∞
h(t1 , t2 , 1, 1) = Ht1 ,t2 (t1 , t2 , 1, 1) = f (t1 , t2 ) g(c1 , c2 ) dc1 dc2 .
t2 t1

243
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
244 Frailty Models in Survival Analysis

The calculation of the other subdensities is analogous:

H(t1 , t2 , 1, 0) = P(T1 > t1 , T2 > t2 , ∆1 = 1, ∆2 = 0)

= P(T1∗ > t1 , C2 > t2 , T1∗ ≤ C1 , T2∗ > C2 )
ZZZZ
= f (t∗1 , t∗2 )g(c1 , c2 ) dt∗1 dt∗2 dc1 dc2
{t∗
1 >t1 ,c2 >t2 ,t1 ≤c1 ,t2 >c2 }
∗ ∗

Z ∞ Z ∞ Z ∞ Z ∞ 
= f (t∗1 , t∗2 ) dt∗2 g(c1 , c2 ) dc1 dt∗1 dc2 .
t2 t1 c2 t∗
1
R∞
Because of −St1 (t1 , t2 ) = t2f (t1 , t∗2 ) dt∗2 it holds that
Z ∞
h(t1 , t2 , 1, 0) = Ht1 ,t2 (t1 , t2 , 1, 0) = −St1 (t1 , t2 ) g(c1 , t2 ) dc1 .
t1

H(t1 , t2 , 0, 1) = P(T1 > t1 , T2 > t2 , ∆1 = 0, ∆2 = 1)

= P(C1 > t1 , T2∗ > t2 , T1∗ > C1 , T2∗ ≤ C2 )
ZZZZ
= f (t∗1 , t∗2 )g(c1 , c2 ) dt∗1 dt∗2 dc1 dc2
{c1 >t1 ,t∗
2 >t2 ,t1 >c1 ,t2 ≤c2 }
∗ ∗

Z ∞Z ∞Z ∞ Z ∞ 
= f (t∗1 , t∗2 ) dt∗1 g(c1 , c2 ) dc2 dc1 dt∗2 .
t2 t1 c1 t∗
2

Hence,
Z ∞
h(t1 , t2 , 0, 1) = Ht1 ,t2 (t1 , t2 , 0, 1) = −St2 (t1 , t2 ) g(t1 , c2 ) dc2 .
t2

H(t1 , t2 , 0, 0) = P(T1 > t1 , T2 > t2 , ∆1 = 0, ∆2 = 0)

= P(C1 > t1 , C2 > t2 , T1∗ > C1 , T2∗ > C2 )
ZZZZ
= f (t∗1 , t∗2 )g(c1 , c2 ) dt∗1 dt∗2 dc1 dc2
{c1 >t1 ,c2 >t2 ,t∗
1 >c1 ,t2 >c2 }
∗

Z ∞ Z ∞ Z ∞ Z ∞ 
= f (t∗1 , t∗2 ) dt∗1 dt∗2 g(c1 , c2 ) dc1 dc2 .
t2 t1 c2 c1

Consequently,

h(t1 , t2 , 0, 0) = Ht1 ,t2 (t1 , t2 , 0, 1) = S(t1 , t2 )g(t1 , t2 ).

Hence, the likelihood function under independent and noninformative right

censoring can be expressed in terms of the bivariate survival function by

St1 t2 (t1 , t2 )δ1 δ2 St1 (t1 , t2 )δ1 (1−δ2 ) St2 (t1 , t2 )(1−δ1 )δ2 S(t1 , t2 )(1−δ1 )(1−δ2 ) . (A.1)

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Appendix 245

Now the case of nonrandom left truncation will be considered. Observations

are made conditional on the fact that both survival times are larger than the
respective truncation times:

P(T1+ > t1 , T2+ > t2 , ∆1 = δ1 , ∆2 = δ2 )

= P(T1 > t1 , T2 > t2 , ∆1 = δ1 , ∆2 = δ2 |T1 > t+ +
1 , T 2 > t2 )
P(T1 > t1 , T2 > t2 , ∆1 = δ1 , ∆2 = δ2 )
= .
P(T1 > t+ +
1 , T 2 > t2 )

Finally, with a sample of n pairs, we obtain the likelihood

n
Y
Sti1 ti2 (ti1 , ti2 )δi1 δi2 Sti1 (ti1 , ti2 )δi1 (1−δi2 ) Sti2 (ti1 , ti2 )(1−δi1 )δi2 (A.2)
i=1

×S(ti1 , ti2 )(1−δi1 )(1−δi2 ) S(t+ + −1

i1 , ti2 ) .

A.2 Correlated Gamma Frailty Model

From the previous section it follows that the likelihood of bivariate event-time
data depends mainly on the derivatives from the survival function. Here,
in this section, these derivatives for the correlated gamma frailty model are
calculated, forming the basis for parameter estimation in this model. The
main ideas from the bivariate correlated gamma frailty model are used later
on for other correlated frailty models, for example, the correlated compound
Poisson frailty model, in the next section. The survival function in the
bivariate correlated gamma frailty model is given by (5.6):

σ σ ρ
1− σ1 ρ 1− σ2 ρ 2 2 −σ
S(t1 , t2 ) = S1 (t1 ) 2 S2 (t2 ) 1 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1) 1 σ2 .

Hence, the densities are of the form

St1 (t1 , t2 ) =
σ1 σ
− 1ρ
σ
1− 2 ρ 2 2 − ρ
− (1 − ρ)S1 (t1 ) σ2 f1 (t1 )S2 (t2 ) σ1 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1) σ1 σ2
σ2
σ1 σ
− 1 2 σ2 2 2 ρ
− ρS1 (t1 ) σ2 ρ−σ1 f1 (t1 )S2 (t2 )1− σ1 ρ (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)− σ1 σ2 −1
σ2
σ σ2
− σ1 ρ 1− ρ 2 2 − σ ρσ −1
= S1 (t1 ) f1 (t1 )S2 (t2 ) σ1 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)
2 1 2

2 σ1 2 σ1

× − S1 (t1 )−σ1 − (1 − ρ)S2 (t2 )−σ2 + (1 − ρ) ,
σ2 σ2

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
246 Frailty Models in Survival Analysis

St2 (t1 , t2 ) =
σ1 σ2 σ2 2 2 ρ
− S1 (t1 )1− σ2 ρ (1 − ρ)S2 (t2 )− σ1 ρ f2 (t2 )(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)− σ1 σ2
σ1
σ1 σ2 σ2 2 2 2 ρ
− S1 (t1 )1− σ2 ρ ρS2 (t2 )− σ1 ρ−σ2 f2 (t2 )(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)− σ1 σ2 −1
σ1
σ1 σ2 2 2 ρ
= S1 (t1 )1− σ2 ρ S2 (t2 )− σ1 ρ f2 (t2 )(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)− σ1 σ2 −1
σ2 2 2 σ2

× − (1 − ρ)S1 (t1 )−σ1 − S2 (t2 )−σ2 + (1 − ρ) ,
σ1 σ1
and finally, the second derivative

St1 t2 (t1 , t2 )
σ1 σ2 σ
− 1ρ
σ
− 2ρ
= (1 − ρ)(1 − ρ)S1 (t1 ) σ2 f1 (t1 )S2 (t2 ) σ1 f2 (t2 )
σ2 σ1
2 2 ρ
−
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1) σ1 σ2
σ1 σ2 σ
− 1ρ
σ
− 2 ρ−σ22
+ (1 − ρ) ρS1 (t1 ) σ2 f1 (t1 )S2 (t2 ) σ1 f2 (t2 )
σ2 σ1
2 2 ρ
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)− σ1 σ2 −1
σ2 σ1 σ
− 1 2 σ2
+ (1 − ρ) ρS1 (t1 ) σ2 ρ−σ1 f1 (t1 )S2 (t2 )− σ1 ρ f2 (t2 )
σ1 σ2
2 2 ρ
−1
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)− σ1 σ2
σ1 2 σ2 2
+ ρ(ρ + σ1 σ2 )S1 (t1 )− σ2 ρ−σ1 f1 (t1 )S2 (t2 )− σ1 ρ−σ2 f2 (t2 )
2 2 − σ ρσ −2
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1) 1 2
σ σ ρ
− σ1 ρ − σ2 ρ 2 2
= S1 (t1 ) 2 f1 (t1 )S2 (t2 ) 1 f2 (t2 )(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)− σ1 σ2 −2
 σ1 σ2 2 2
× (1 − ρ)(1 − ρ)(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)2
σ2 σ1
σ1 σ2 2 2 2
+ (1 − ρ) ρS2 (t2 )−σ2 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)
σ2 σ1
σ2 σ1 2 2 2
+ (1 − ρ) ρS1 (t1 )−σ1 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)
σ1 σ2
2 2

+ ρ(ρ + σ1 σ2 )S1 (t1 )−σ1 S2 (t2 )−σ2 .

Plugging these expressions into formula (A.2) gives the likelihood function of
the correlated gamma frailty model. It is easy to see that, even in the simple
bivariate case, the likelihood becomes formidable because of the necessary
derivatives. The problem increases with larger cluster size. Furthermore, the
maximum likelihood method is only applicable in the parametric case with full
specification of the baseline hazard function. Therefore, more sophisticated
strategies for parameter estimation are necessary, especially in the case of
cluster size larger than two.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Appendix 247

A.3 Correlated Compound Poisson Frailty Model

The correlated compound Poisson frailty model is a natural extension of the
correlated gamma frailty model. Similar to the gamma model, we start with
the bivariate marginal survival function, which is given for the compound
Poisson frailty model by (5.14):

S(t1 , t2 ) = S(t1 )1−ρ S(t2 )1−ρ

 ρ(1 − γ)  γσ 2 1
γ + (1 −
γσ 2 1
γ − 1

γ 
× exp 1 − (1 − ln S(t 1 )) ln S(t 2 )) .
γσ 2 1−γ 1−γ

The derivatives with respect to the first and second event time can be easily
calculated by

St1 (t1 , t2 ) =
− (1 − ρ)f (t1 )S(t1 )−ρ S(t2 )1−ρ
 ρ(1 − γ)  γσ 2 1 γσ 2 1
γ 
× exp 1 − (1 − ln S(t 1 )) γ + (1 − ln S(t 2 )) γ − 1
γσ 2 1−γ 1−γ
2
γσ 1
− ρf (t1 )S(t1 )−ρ S(t2 )1−ρ (1 − ln S(t1 )) γ −1
1−γ
 ρ(1 − γ)  γσ 2 1 γσ 2 1 
× exp 2
1 − (1 − ln S(t1 )) γ + (1 − ln S(t2 )) γ − 1)γ
γσ 1−γ 1−γ
2 2
γσ 1 γσ 1
γ−1
× (1 − ln S(t1 )) γ + (1 − ln S(t2 )) γ − 1
1−γ 1−γ

and

St2 (t1 , t2 ) =
− (1 − ρ)f (t2 )S(t1 )1−ρ S(t2 )−ρ
 ρ(1 − γ)  γσ 2 1
γ + (1 −
γσ 2 1
γ − 1

γ 
× exp 1 − (1 − ln S(t 1 )) ln S(t 2 ))
γσ 2 1−γ 1−γ
2
γσ 1
− ρf (t2 )S(t1 )1−ρ S(t2 )−ρ (1 − ln S(t2 )) γ −1
1−γ
 ρ(1 − γ)  γσ 2 1 γσ 2 1
γ 
× exp 2
1 − (1 − ln S(t1 )) γ + (1 − ln S(t2 )) γ − 1
γσ 1−γ 1−γ
2 2
γσ 1 γσ 1
γ−1
× (1 − ln S(t1 )) γ + (1 − ln S(t2 )) γ − 1 .
1−γ 1−γ

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
248 Frailty Models in Survival Analysis

The derivative with respect to both event times is given by

St1 t2 (t1 , t2 ) =
(1 − ρ)2 f (t1 )S(t1 )−ρ f (t2 )S(t2 )−ρ
 ρ(1 − γ)  γσ 2 1
γ + (1 −
γσ 2 1
γ 
× exp 1 − (1 − ln S(t 1 )) ln S(t2 )) γ −1
γσ 2 1−γ 1−γ
γσ 2 1
+ ρ(1 − ρ)f (t1 )S(t1 )−ρ f (t2 )S(t2 )−ρ (1 − ln S(t1 )) γ −1
1−γ
 ρ(1 − γ)  γσ 2 1 γσ 2 1
γ 
× exp 2
1 − (1 − ln S(t1 )) γ + (1 − ln S(t2 )) γ −1
γσ 1−γ 1−γ
γσ 2 1 γσ 2 1
γ−1
× (1 − ln S(t1 )) γ + (1 − ln S(t2 )) γ − 1
1−γ 1−γ
γσ 2 1
+ ρ(1 − ρ)f (t1 )S(t1 )−ρ f (t2 )S(t2 )−ρ (1 − ln S(t2 )) γ −1
1−γ
 ρ(1 − γ)  γσ 2 1 γσ 2 1
γ 
× exp 2
1 − (1 − ln S(t 1 )) γ + (1 − ln S(t2 )) γ −1
γσ 1−γ 1−γ
γσ 2 1 γσ 2 1
γ−1
× (1 − ln S(t1 )) γ + (1 − ln S(t2 )) γ − 1
1−γ 1−γ
+ ρ2 f (t1 )S(t1 )−ρ f (t2 )S(t2 )−ρ
 ρ(1 − γ)  γσ 2 1 γσ 2 1
γ 
× exp 2
1 − (1 − ln S(t 1 )) γ + (1 − ln S(t2 )) γ −1
γσ 1−γ 1−γ
γσ 2 1 γσ 2 1
2γ−2
× (1 − ln S(t1 )) γ + (1 − ln S(t2 )) γ − 1
1−γ 1−γ
γσ 2 1 γσ 2 1
× (1 − ln S(t1 )) γ −1 (1 − ln S(t2 )) γ −1
1−γ 1−γ
+ ρσ 2 f (t1 )S(t1 )−ρ f (t2 )S(t2 )−ρ
 ρ(1 − γ)  γσ 2 1 γσ 2 1
γ 
× exp 1 − (1 − ln S(t 1 )) γ + (1 − ln S(t2 )) γ −1
γσ 2 1−γ 1−γ
γσ 2 1 γσ 2 1
γ−2
× (1 − ln S(t1 )) γ + (1 − ln S(t2 )) γ − 1
1−γ 1−γ
γσ 2 1 γσ 2 1
× (1 − ln S(t1 )) γ −1 (1 − ln S(t2 )) γ −1 .
1−γ 1−γ

Substituting these expressions in (A.2) yields the likelihood function of the

bivariate correlated compound Poisson frailty model. The model was used in
Section 5.5 to analyze the ages at onset of breast cancer in Swedish female
twin pairs. As discussed there in more detail, the correlated gamma and the
correlated inverse Gaussian frailty model are special cases of the compound
Poisson frailty model.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Appendix 249

A.4 Correlated Quadratic Hazard Frailty Model

In the following, the unconditional bivariate survival function in the quadratic
hazard frailty model (5.15) is derived by integrating the conditional survival
function first with respect to the random effects using the same form of the
two-dimensional normal density as in (5.16).

S(t1 , t2 ) = ES(t1 , t2 |W1 , W2 )

1
ZZ
= √ S(t1 , t2 |w1 , w2 )
2πs1 s2 1 − r2
(w1 −m1 )2 (w −m )2
(w1 −m1 )(w2 −m2 )


−1
−r + 2 22
2 2s2 s1 s2
× e 1−r 1
2s
2 dw1 dw2
1
ZZ
2 2
= √ e−w1 M0 (t1 ) e−w2 M0 (t2 )
2πs1 s2 1 − r2
(w1 −m1 )2 w2 −m

−1 (w −m )2 w2 −m2


−1
−r 2 w1 2 2 +r m1
1−r2 2s2 s1 s2 1−r2 2s2 s1 s2
×e 1 e 2 dw1 dw2
(w2 −m2 )2 w −m


1
Z
2
−1
+r s2 s 2 m1
= √ e−w2 M0 (t2 ) I(w2 )e 1−r2 2s2
2 1 2
dw2
2πs2

with

(w1 −m1 )2 w2 −m2

1
Z −1
−w12 M0 (t1 ) 1−r2
−r w1
2s2 s1 s2
I(w2 ) = √ √ e e 1 dw1 .
2πs1 1 − r2

In the next paragraph, we first calculate the integral I(w2 ):

I(w2 )
(w1 −m1 )2 w2 −m2


1
Z −1
−w12 M0 (t1 ) −r w1
1−r2 2s2 s1 s2
=√ √ e e 1 dw1
2πs1 1 − r2
(1+2s2 2 2 2
1 (1−r )M0 (t1 ))s2 w1 −2m1 s2 w1 −2r(w2 −m2 )s1 w1 +m1 s2
1
Z
−
2s2 (1−r2 )s2
=√ √ e 1 dw1
2πs1 1 − r2
s 2
m1 +r(w2 −m2 ) s1 m21

√ 1
1+2s21 (1−r 2 )M0 (t1 )
n 1+2s21 (1−r 2 )M0 (t1 )
2
− 1+2s21 (1−r 2 )M0 (t1 )
o
= √ √
2
exp s2 (1−r 2 )
√ 2πs 2
1 1−r
2 1+2s2 (1−r
1
2 )M (t )
1+2s1 (1−r 2 )M0 (t1 ) 1 0 1

s1 s1 2
m +r(w −m ) m +r(w −m 2) s

1 2 2 1 2
Z n w12 − 2 1+2s2 (1−r2 )M (t s2
w 1 + 2 2
2
0 1) 1+2s1 (1−r )M0 (t1 )
o
1
× exp − 2
s1 (1−r )2 dw1
2 1+2s2 (1−r 2 )M (t )
0 1 1

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
250 Frailty Models in Survival Analysis

Because of
s
m1 +r(w2 −m2 ) s1
2
1
Z n w1 − 2
1+2s21 (1−r 2 )M0 (t1 )
o
√ √
2
exp − s21 (1−r 2 )
dw1 = 1,
√ 2πs 1 1−r
2
1+2s21 (1−r 2 )M0 (t1 ) 1+2s21 (1−r 2 )M0 (t1 )

it holds that

2
−2s2 2 2 s1 2 2 s1
1 (1−r )M0 (t1 )m1 +2rm1 (w2 −m2 ) s2 +r (w2 −m2 ) s2
1 2s2 2 2 2
2
I(w2 ) = p e 1 (1−r )(1+2s1 (1−r )M0 (t1 ))
1 + 2s21 (1 − r2 )M0 (t1 )
s s2
2rm1 (w2 −m2 ) 1 +r2 (w2 −m2 )2 1
−M0 (t1 )m2 s2 s2
1 1
1+2s2 (1−r2 )M0 (t1 ) 2s2 (1−r2 )(1+2s2 (1−r2 )M0 (t1 ))
2
= p e 1 e 1 1
1 + 2s21 (1 − r2 )M0 (t1 )
= I1 I2 I3 (w2 ).
Now it is necessary to integrate the conditional survival function with respect
to the second random effect. Considering the foregoing three terms in more
detail, it turns out that the first two terms I1 and I2 are independent of w2 .
The third term I3 (w2 ) contains w2 and has to be included in the integral with
respect to w2 . Consequently, it holds that

S(t1 , t2 )
(w2 −m2 )2 w2 −m2


I1 I2
Z −1
2 +r m1
e−w2 M0 (t2 ) e 1−r 2s2
2 s1 s2
= √ 2 I3 (w2 ) dw2
2πs2
(w2 −m2 )2 w2 −m2


I1 I2
Z −1
2 +r m1
e−w2 M0 (t2 ) e 1−r 2s2
2 s1 s2
= √ 2
2πs2
s s2
2rm1 (w2 −m2 ) 1 +r2 (w2 −m2 )2 1
s2 s2
2
2s2 (1−r2 )(1+2s2 (1−r2 )M0 (t1 ))
×e 1 1 dw2
(w −m2 )2 s2 2 2
1 (1+2s1 (1−r )M0 (t1 ))
I1 I2
Z
2 − 22
e−w2 M0 (t2 ) e 2s1 (1−r2 )s2 2 2
=√ 2 (1+2s1 (1−r )M0 (t1 ))
2πs2
2rm1 s1 s2 (1+2s2 2 2 2 2
1 (1−r )M0 (t1 ))(w2 −m2 )−2rm1 s1 s2 (w2 −m2 )−r (w2 −m2 ) s1
−
2s2 (1−r2 )s2 (1+2s2 (1−r2 )M0 (t1 ))
×e 1 2 1 dw2
(1+2s2 2 2 2 2 2
1 M0 (t1 ))(w2 −m2 ) +2s2 (1+2s1 (1−r )M0 (t1 ))M0 (t2 )w2
I1 I2
Z
−
2s2 2 (1−r2 )M (t ))
=√ e 2 (1+2s 1 0 1
2πs2
4rm1 s1 s2 M0 (t1 )(w2 −m2 )
−
2s2 (1+2s2 (1−r2 )M0 (t1 ))
×e 2 1 dw2
(1+2s2 2 2 2 2
1 M0 (t1 )+2s2 (1+2s1 (1−r )M0 (t1 ))M0 (t2 ))w2
I1 I2
Z
−
2s2 2 2
= √ e 2 (1+2s1 (1−r )M0 (t1 )) dw2
2πs2
−2((1+2s2
1 M0 (t1 ))m2 −2rm1 s1 s2 M0 (t1 ))w2 (1+2s2 2
1 M0 (t1 ))m2 −4rm1 m2 s1 s2 M0 (t1 )
− −
2s2 (1+2s2 (1−r2 )M0 (t1 )) 2s2 (1+2s2 (1−r2 )M0 (t1 ))
×e 2 1 dw2 e 2 1 .

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Appendix 251

Substituting the last factor by the shorthand

(1+2s2 2
1 M0 (t1 ))m2 −4rm1 m2 s1 s2 M0 (t1 )
−
2s2 2 (1−r2 )M (t ))
I4 = e 2 (1+2s 1 0 1 ,

then

S(t1 , t2 )
(1+2s2 2 2 2 2
1 M0 (t1 )+2s2 (1+2s1 (1−r )M0 (t1 ))M0 (t2 ))w2
I1 I2 I4
Z
−
2s2 (1+2s2 (1−r2 )M0 (t1 ))
= √ e 2 1 dw2
2πs2
−2((1+2s2
1 M0 (t1 ))m2 −2rm1 s1 s2 M0 (t1 ))w2
−
2s2 2 2
×e 2 (1+2s1 (1−r )M0 (t1 )) dw2
2((1+2s21 M0 (t1 ))m2 −2rm1 s1 s2 M0 (t1 ))w2
I1 I2 I4
Z n w22 − 1+2s21 M0 (t1 )+2s22 M0 (t2 )+4s21 s22 (1−r 2 )M0 (t1 )M0 (t2 )
o
= √ exp − 2s22 (1+2s21 (1−r 2 )M0 (t1 ))
dw2
2πs2
1+2s21 M0 (t1 )+2s22 M0 (t2 )+4s21 s22 (1−r 2 )M0 (t1 )M0 (t2 )
p
1 + 2s21 (1 − r2 )M0 (t1 )
= I1 I2 I4 p
1 + 2s21 M0 (t1 ) + 2s22 M0 (t2 ) + 4s21 s22 (1 − r2 )M0 (t1 )M0 (t2 )
(1+2s21 M0 (t1 ))m2 −2rm1 s1 s2 M0 (t1 )
2
n −
1+2s21 M0 (t1 )+2s22 M0 (t2 )+4s21 s22 (1−r 2 )M0 (t1 )M0 (t2 )
o
× exp − 2 2 2
2s (1+2s (1−r )M (t ))
.
2 1 0 1
1+2s21 M0 (t1 )+2s22 M0 (t2 )+4s21 s22 (1−r 2 )M0 (t1 )M0 (t2 )

Here I1 cancels out. Furthermore, we introduce

1
I5 = p ,
1+ 2s21 M0 (t1 ) + 2s22 M0 (t2 ) + 4s21 s22 (1 − r2 )M0 (t1 )M0 (t2 )

which implies that

S(t1 , t2 )
−((1+2s2 1 M0 (t1 ))m2 −2rm1 s1 s2 M0 (t1 ))
2
−
2s2 (1+2s2 (1−r2 )M0 (t1 ))(1+2s2 M0 (t1 )+2s2 M0 (t2 )+4s2 s2 (1−r2 )M0 (t1 )M0 (t2 ))
= I2 I5 e 2 1 1 2 1 2

((1+2s2 2 2 2 2 2 2
1 M0 (t1 ))m2 −4rm1 m2 s1 s2 M0 (t1 ))(1+2s1 M0 (t1 )+2s2 M0 (t2 )+4s1 s2 (1−r )M0 (t1 )M0 (t2 ))
−
2s2 2 2 2 2 2 2 2
×e 2 (1+2s1 (1−r )M0 (t1 ))(1+2s1 M0 (t1 )+2s2 M0 (t2 )+4s1 s2 (1−r )M0 (t1 )M0 (t2 ))

−2r2 m2 2 2 2 2 2 2
1 s1 M0 (t1 )+m2 M0 (t2 )(1+2s1 M0 (t1 ))(1+2s1 (1−r )M0 (t1 ))
−
(1+2s2 2 2 2 2 2 2
= I2 I5 e 1 (1−r )M0 (t1 ))(1+2s1 M0 (t1 )+2s2 M0 (t2 )+4s1 s2 (1−r )M0 (t1 )M0 (t2 ))

−4rm1 m2 s1 s2 M0 (t1 )M0 (t2 )(1+2s2 2

1 (1−r )M0 (t1 ))
−
(1+2s2 (1−r2 )M0 (t1 ))(1+2s2 M0 (t1 )+2s2 M0 (t2 )+4s2 s2 (1−r2 )M0 (t1 )M0 (t2 ))
×e 1 1 2 1 2 .

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
252 Frailty Models in Survival Analysis

Substituting the term I2 back, it holds that

S(t1 , t2 )
m2 2 2 2 2 2 2 2 2 2
1 M0 (t1 )(1+2s1 M0 (t1 )+2s2 M0 (t2 )+4s1 s2 (1−r )M0 (t1 )M0 (t2 ))−2r m1 s1 M0 (t1 )
−
(1+2(1−r2 )s2 2 2 2 2 2
= I5 e 1 M0 (t1 ))(1+2s1 M0 (t1 )+2s2 M0 (t2 )+4s1 s2 (1−r )M0 (t1 )M0 (t2 ))

m2 2
2 M0 (t2 )(1+2s1 M0 (t1 ))−4rm1 m2 s1 s2 M0 (t1 )M0 (t2 )
−
1+2s2 2 2 2 2
×e 1 M0 (t1 )+2s2 M0 (t2 )+4s1 s2 (1−r )M0 (t1 )M0 (t2 )

1
=p
1+ 2s21 M0 (t1 ) + 2s22 M0 (t2 ) + 4s21 s22 (1 − r2 )M0 (t1 )M0 (t2 )
m2 M (t )(1+2s2 2 2
1 M0 (t1 ))−4rm1 m2 s1 s2 M0 (t1 )M0 (t2 )+m1 M0 (t1 )(1+2s2 M0 (t2 ))
− 2 0 2
(1+2s2 2 2 2 2
×e 1 M0 (t1 )+2s2 M0 (t2 )+4s1 s2 (1−r )M0 (t1 )M0 (t2 )) .
In the case of independence r = 0, this results in
−M0 (t1 )m1 2 −M0 (t2 )m2 2
1 1+2s2 M0 (t1 )
1 1+2s2 M0 (t2 )
p e 1 p e 2 .
1 + 2s21 M0 (t1 ) 1 + 2s22 M0 (t2 )
Now we are considering the case of m1 = m2 = 0 and s21 = s22 = s22 , but r 6= 0.
For this special case, it is straightforward to derive the log-likelihood function,
requiring the first and second derivatives of the unconditional survival function
S(t1 , t2 ). Using the relation S(t) = (1 + 2s2 M0 (t))−1/2 , the bivariate survival
function becomes, in the copula form,

− 1
S(t1 , t2 ) = S(t1 )−2 S(t2 )−2 − r2 (S(t1 )−2 − 1)(S(t2 )−2 − 1) 2

− 3
St1 (t1 , t2 ) = − S(t1 )−2 S(t2 )−2 − r2 (S(t1 )−2 − 1)(S(t2 )−2 − 1) 2
× f (t1 )S(t1 )−3 ((1 − r2 )S(t2 )−2 + r2 )

− 3
St2 (t1 , t2 ) = − S(t1 )−2 S(t2 )−2 − r2 (S(t1 )−2 − 1)(S(t2 )−2 − 1) 2
× f (t2 )S(t2 )−3 ((1 − r2 )S(t1 )−2 + r2 ).
For the second derivative, it holds that
St1 t2 (t1 , t2 )

− 5
= 3 S(t1 )−2 S(t2 )−2 − r2 (S(t1 )−2 − 1)(S(t2 )−2 − 1) 2
× f (t1 )S(t1 )−3 f (t2 )S(t2 )−3 ((1 − r2 )S(t1 )−2 + r2 )((1 − r2 )S(t2 )−2 + r2 )

− 3
− 2(1 − r2 ) S(t1 )−2 S(t2 )−2 − r2 (S(t1 )−2 − 1)(S(t2 )−2 − 1) 2
× f (t1 )S(t1 )−3 f (t2 )S(t2 )−3

− 5
= S(t1 )−2 S(t2 )−2 − r2 (S(t1 )−2 − 1)(S(t2 )−2 − 1) 2 f (t1 )S(t1 )−3
 
× f (t2 )S(t2 )−3 (1 − r2 )S(t1 )−2 + r2 (1 − r2 )S(t2 )−2 + r2 + 2r2 .

This expression can now be used in the bivariate likelihood function (A.2).

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Appendix 253

A.5 Dependent Competing Risks Model

Here a specific four-dimensional correlated gamma frailty model is derived
to model dependent competing risks in twin survival data. The following
relations are used in the calculations: Let V1 , V8 ∼ Γ(k1 , λ0 ), V2 ∼ Γ(k2 , λ1 ),
V3 ∼ Γ(k3 , λ2 ), V4 , V7 ∼ Γ(k4 , λ2 ), V5 , V6 ∼ Γ(k5 , λ1 ) be independent gamma-
distributed random variables with parameters k1 + k2 + k5 := λ1 = σ12 and
1
1
k1 + k3 + k4 := λ2 = σ22
. Hence, it holds that EZ1 = EZ2 = EZ3 = EZ4 = 1,
1 2 1
V(Z1 ) = V(Z3 ) = k1 +k2 +k5 = σ1 , and V(Z2 ) = V(Z4 ) = k1 +k3 +k4 = σ22 . It
holds that

λ0 λ0 k2
cov(Z1 , Z3 ) = cov( V1 + V2 + V5 , V2 + V6 + V8 ) = V(V2 ) = 2 .
λ1 λ1 λ1
Consequently, the correlation coefficient between Z1 and Z3 is

cov(Z1 , Z3 ) k2
ρ1 = corr(Z1 , Z3 ) = p = = k2 σ12 . (A.3)
V(Z1 )V(Z3 ) λ1
Here, ρ1 is the correlation between the frailties of twins with respect to the first
risk. By similar calculations it is easy to obtain ρ2 = corr(Z2 , Z4 ) = k3 σ22 ,
which describes the correlation between frailties of twin partners regarding the
second competing risk. In the following we calculate the correlation coefficient
ρ = corr(Z1 , Z2 ) = corr(Z3 , Z4 ), describing the correlation between the
frailty terms of the two competing risks. It holds that

λ0 λ0 λ2 k1
cov(Z1 , Z2 ) = cov( V1 + V2 + V5 , V1 + V3 + V4 ) = 0 V(V1 ) = .
λ1 λ2 λ1 λ2 λ1 λ2
The correlation coefficient between Z1 and Z2 (and between Z3 and Z4 ) is
given by

cov(Z1 , Z2 ) k1
ρ = corr(Z1 , Z2 ) = p =√ = k1 σ1 σ2 . (A.4)
V(Z1 )V(Z2 ) λ1 λ2
1 1
Consequently, k1 + k2 + k5 = σ12
, k1 + k3 + k4 = σ22
, and relations (A.3) and
(A.4) result in
1 1 ρ1 ρ
k5 = 2 − k2 − k1 = 2 − 2 −
σ1 σ1 σ1 σ1 σ2
and
1 1 ρ2 ρ
k4 = − k3 − k4 = 2 − 2 − .
σ22 σ2 σ2 σ1 σ2

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
254 Frailty Models in Survival Analysis

For gamma-distributed random variables Y ∼ Γ(k, λ), the Laplace transform

is given by Ee−sY = (1 + λs )−k . Now we are in the state to derive the survival
function in (5.25):

S(t1 , y1 , t2 , y2 )
= ES1 (t1 )Z1 S2 (y1 )Z2 S1 (t2 )Z3 S2 (y2 )Z4
λ0 λ0
= Ee−V1 ( λ1 M01 (t1 )+ λ2 M02 (y1 )) e−V2 (M01 (t1 )+M01 (t2 ))
λ λ
−V8 ( λ0 M01 (t2 )+ λ0 M02 (y2 ))
× e−V3 (M02 (y1 )+M02 (y2 )) e 1 2

−V4 M02 (y1 ) −V5 M01 (t1 ) −V6 M01 (t2 ) −V7 M02 (y2 )
×e e e e
 M01 (t1 ) M02 (y1 ) −k 1
 M01 (t1 ) M01 (t2 ) −k2
= 1+ + 1+ +
λ1 λ2 λ1 λ1
 M02 (y1 ) M02 (y2 ) −k3  M01 (t2 ) M02 (y2 ) −k1
× 1+ + 1+ +
λ2 λ2 λ1 λ2
 M02 (y1 ) −k4  M01 (t1 ) −k5  M01 (t2 ) −k5  M02 (y2 ) −k4
× 1+ 1+ 1+ 1+
λ2 λ1 λ1 λ2
ρ1 ρ2
2 2 − 2 2 −
= S1 (t1 )−σ1 + S1 (t2 )−σ1 − 1 σ1 S2 (y1 )−σ2 + S2 (y2 )−σ2 − 1 σ2
2 2



2 2
− ρ 2 2
− ρ
× S1 (t1 )−σ1 + S2 (y1 )−σ2 − 1 σ1 σ2 S1 (t2 )−σ1 + S2 (y2 )−σ2 − 1 σ1 σ2
σ σ σ σ
1−ρ2 − σ2 ρ 1−ρ1 − σ1 ρ 1−ρ1 − σ1 ρ 1−ρ2 − σ2 ρ
× S2 (y1 ) 1 S1 (t1 ) 2 S1 (t2 ) 2 S2 (y2 ) 1 .
To derive the likelihood function of the bivariate dependent competing risk
data, it is necessary to keep in mind that the above four-dimensional model
is not fully observable. Only the minimum of the three competing risk times
(tj , yj , cj ) is observable for twin j (j = 1, 2). Because of (5.23), the likelihood
contribution of the truncated data takes the form


1(δ1 = 1, δ2 = 1)St1 ,t2 (t1 , t1 , t2 , t2 )
+ 1(δ1 = 1, δ2 = 0)St1 (t1 , t1 , c2 , c2 )
+ 1(δ1 = 0, δ2 = 1)St2 (c1 , c1 , t2 , t2 )
+ 1(δ1 = 0, δ2 = 0)S(c1 , c1 , c2 , c2 )
+ 1(δ1 = −1, δ2 = −1)Sy1 y2 (y1 , y1 , y2 , y2 )
+ 1(δ1 = −1, δ2 = 0)Sy1 (y1 , y1 , c2 , c2 )
+ 1(δ1 = 0, δ2 = −1)Sy2 (c1 , c1 , y2 , y2 )
+ 1(δ1 = 1, δ2 = −1)St1 y2 (t1 , t1 , y2 , y2 )

+ 1(δ1 = −1, δ2 = 1)Sy1 t2 (y1 , y1 , t2 , t2 ) /S(t+ , t+ , t+ , t+ ),

where t+ denotes the truncation time. In the following the derivatives of the
survival function needed for the likelihood expression are calculated.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Appendix 255

St1 (t1 , y1 , t2 , y2 )
ρ1 ρ2
2 2 − −1 2 2 −
= −ρ1 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1) (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)
2
σ1 2
σ2

2 2 ρ 2 2 ρ
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)− σ1 σ2 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)− σ1 σ2
σ1 2 σ1 σ2 σ2
× S1 (t1 )1−ρ1 − σ2 ρ−σ1 µ1 (t1 )S1 (y1 )1−ρ1 − σ2 ρ S2 (t2 )1−ρ2 − σ1 ρ S2 (y2 )1−ρ2 − σ1 ρ
ρ ρ
σ1 2 2 − 12 2 2 − 22
− ρ(S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1) σ1 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1) σ2
σ2
2 2 − σ ρσ −1 2 2 −σ ρ
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1) 1 2 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1) 1 σ2

σ σ σ σ
1−ρ1 − σ1 ρ−σ12 1−ρ1 − σ1 ρ 1−ρ2 − σ2 ρ 1−ρ2 − σ2 ρ
× S1 (t1 ) 2 µ1 (t1 )S1 (y1 ) 2 S2 (t2 ) 1 S2 (y2 ) 1

ρ
σ1 2 2 − 12
− (1 − ρ1 − ρ)σ12 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1) σ1

σ2
ρ2
2 2 − 2 2 ρ
× (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1) (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)− σ1 σ2
2
σ2

2 2 ρ
× (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)− σ1 σ2
σ1 σ1 σ2 σ2
ρ 1−ρ − ρ 1−ρ − ρ 1−ρ − ρ
× S1 (t1 )1−ρ1 − σ2 µ1 (t1 )S1 (y1 ) 1 σ2 S2 (t2 ) 2 σ1 S2 (y2 ) 2 σ1

Sy1 (t1 , y1 , t2 , y2 )
ρ1 ρ2
2 2 − −1 2 2 −
= −ρ1 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1) σ2
1 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1) σ2
2

2 2 − σ ρσ 2 2 −σ ρ
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1) 1 2 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1) 1 σ2

σ1 σ1 2 σ2 σ2
× S1 (t1 )1−ρ1 − σ2 ρ S1 (y1 )1−ρ1 − σ2 ρ−σ1 µ1 (y1 )S2 (t2 )1−ρ2 − σ1 ρ S2 (y2 )1−ρ2 − σ1 ρ
ρ ρ
σ1 2 2 − 12 2 2 − 22
− ρ(S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1) σ1 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1) σ2
σ2
2 2 ρ 2 2 ρ
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)− σ1 σ2 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)− σ1 σ2 −1
σ σ σ σ
1−ρ1 − σ1 ρ 1−ρ1 − σ1 ρ−σ12 1−ρ2 − σ2 ρ 1−ρ2 − σ2 ρ
× S1 (t1 ) 2 S1 (y1 ) 2 µ1 (y1 )S2 (t2 ) 1 S2 (y2 ) 1

ρ
σ1 2 2 − 12
− (1 − ρ1 − ρ)σ12 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1) σ
1
σ2
ρ2
2 2 − 2 2 − σ ρσ
× (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1) σ2
2 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1) 1 2

−σ12 −σ22 − σ ρσ
× (S1 (y1 ) + S2 (y2 ) − 1) 1 2

σ σ σ σ
1−ρ1 − σ1 ρ 1−ρ1 − σ1 ρ 1−ρ2 − σ2 ρ 1−ρ2 − σ2 ρ
× S1 (t1 ) 2 S1 (y1 ) 2 µ1 (y1 )S2 (t2 ) 1 S2 (y2 ) 1

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
256 Frailty Models in Survival Analysis

St2 (t1 , y1 , t2 , y2 )
ρ1 ρ2
2 2 − 2 2 − −1
= −ρ2 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1) σ2
1 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1) σ2
2

ρ
2 2 − σ ρσ 2 2 −σ
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1) 1 2 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1) 1 σ2

σ1 σ1 σ2 2 σ2
× S1 (t1 )1−ρ1 − σ2 ρ S1 (y1 )1−ρ1 − σ2 ρ S2 (t2 )1−ρ2 − σ1 ρ−σ2 µ2 (t2 )S2 (y2 )1−ρ2 − σ1 ρ
ρ ρ
σ2 2 2 − 12 2 2 − 22
− ρ(S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1) σ1 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1) σ2
σ1
2 2 ρ 2 2 ρ
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)− σ1 σ2 −1 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)− σ1 σ2
σ1 σ1 σ2 2 σ2
× S1 (t1 )1−ρ1 − σ2 ρ S1 (y1 )1−ρ1 − σ2 ρ S2 (t2 )1−ρ2 − σ1 ρ−σ2 µ2 (t2 )S2 (y2 )1−ρ2 − σ1 ρ
ρ
σ2 2 2 − 12
− (1 − ρ2 − ρ)σ22 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1) σ1
σ1
ρ2
2 2 − 2 2 − σ ρσ
× (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1) σ2
2 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1) 1 2

−σ12 −σ22 − σ ρσ
× (S1 (y1 ) + S2 (y2 ) − 1) 1 2

σ σ σ2 σ2
1−ρ1 − σ1 ρ 1−ρ1 − σ1 ρ
× S1 (t1 ) 2 S1 (y1 ) 2 S2 (t2 )1−ρ2 − σ1 ρ µ2 (t2 )S2 (y2 )1−ρ2 − σ1 ρ

Sy2 (t1 , y1 , t2 , y2 )
ρ1 ρ2
2 2 − 2 2 − −1
= −ρ2 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1) (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)
2
σ1 2
σ2

2 2 ρ 2 2 ρ
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)− σ1 σ2 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)− σ1 σ2
σ1 σ1 σ2 σ2 2
1−ρ −
× S1 (t1 ) 1 σ2 ρ S1 (y1 )1−ρ1 − σ2 ρ S2 (t2 )1−ρ2 − σ1 ρ S2 (y2 )1−ρ2 − σ1 ρ−σ2 µ2 (y2 )
ρ ρ
σ1 2 2 − 12 2 2 − 22
− ρ(S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1) σ1 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1) σ2
σ2
ρ
2 2 − σ ρσ 2 2 −σ −1
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1) 1 2 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1) 1 σ2

σ1 σ1 σ2 σ2 2
× S1 (t1 )1−ρ1 − σ2 ρ S1 (y1 )1−ρ1 − σ2 ρ S2 (t2 )1−ρ2 − σ1 ρ S2 (y2 )1−ρ2 − σ1 ρ−σ2 µ2 (y2 )
ρ
σ2 2 2 − 12
− (1 − ρ2 − ρ)σ22 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1) σ1
σ1
ρ2
2 2 − 2 2 ρ
× (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1) (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)− σ1 σ2
2
σ2

2 2 − σ ρσ
× (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1) 1 2

σ σ σ2 σ2
1−ρ1 − σ1 ρ 1−ρ1 − σ1 ρ
× S1 (t1 ) 2 S1 (y1 ) 2 S2 (t2 )1−ρ2 − σ1 ρ S2 (y2 )1−ρ2 − σ1 ρ µ2 (y2 )

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com

St1 ,y1 (t1 , y1 , t2 , y2 )

2 2 2 2 2 2
= µ1 (t1 )µ1 (y1 )(S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)−ρ1 /σ1 −2 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)−ρ2 /σ2
2 2 2 2
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)−ρ/σ1 σ2 −1 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)−ρ/σ1 σ2 −1
σ1 σ1 σ2 σ2
1−ρ − ρ 1−ρ − ρ 1−ρ − ρ 1−ρ − ρ
× S1 (t1 ) 1 σ2 S1 (y1 ) 1 σ2 S2 (t2 ) 2 σ1 S2 (y2 ) 2 σ1
h σ2 2 2 2 2
× ρ2 12 S1 (t1 )−σ1 S1 (y1 )−σ1 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)2
σ2
σ1 2 2 2 2 2 2
+ ρρ1 S1 (t1 )−σ1 S1 (y1 )−σ1 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)
σ2
σ1 σ1 2 2 2 2 2
+ ρ(1 − ρ1 − ρ) S1 (y1 )−σ1 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)2 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)

Appendix
σ2 σ2
σ1 2 2 2 2 2 2
+ ρρ1 S1 (t1 )−σ1 S1 (y1 )−σ1 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ2
σ1 σ1 2 2 2 2 2
+ ρ(1 − ρ1 − ρ) S1 (t1 )−σ1 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)2 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ2 σ2
2 2 2 2 2 2
+ ρ1 (ρ1 + σ12 )S1 (t1 )−σ1 S1 (y1 )−σ1 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ1 2 2 2
+ ρ1 (1 − ρ1 − ρ)S1 (t1 )−σ1 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)
σ2
2 2 2 2
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ1 2 2 2
+ ρ1 (1 − ρ1 − ρ)S1 (y1 )−σ1 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)
σ2
2 2 2 2
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ1 2 2 2 2 2 2
i
+ (1 − ρ1 − ρ)2 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)2 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ2

257
© 2011 by Taylor and Francis Group, LLC
www.Ebook777.com
 Free ebooks ==> www.Ebook777.com
St1 ,y2 (t1 , y1 , t2 , y2 )

258
2 2 2 2 2 2
= µ1 (t1 )µ2 (y2 )(S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)−ρ1 /σ1 −1 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)−ρ2 /σ2 −1
2 2 2 2
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)−ρ/σ1 σ2 −1 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)−ρ/σ1 σ2 −1
σ1 σ1 σ2 σ2
1−ρ − ρ 1−ρ − ρ 1−ρ − ρ 1−ρ − ρ
× S1 (t1 ) 1 σ2 S1 (y1 ) 1 σ2 S2 (t2 ) 2 σ1 S2 (y2 ) 2 σ1
h σ1 2 2 2 2 2 2
× ρρ2 S1 (t1 )−σ1 S2 (y2 )−σ2 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ2
2 2 2 2 2 2
+ ρ1 ρ2 S1 (t1 )−σ1 S2 (y2 )−σ2 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)

Frailty Models in Survival Analysis

σ1 2 2 2 2 2 2 2
+ ρ2 (1 − ρ1 − ρ)S2 (y2 )−σ2 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ2
2 2 2 2 2 2
+ ρ2 S1 (t1 )−σ1 S2 (y2 )−σ2 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)(S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)
σ2 2 2 2 2 2 2
+ ρρ1 S1 (t1 )−σ1 S2 (y2 )−σ2 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)
σ1
σ2 σ1 2 2 2
+ ρ(1 − ρ2 − ρ) S1 (t1 )−σ1 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)
σ1 σ2
2 2 2 2
× (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ1 σ2 2 2 2
+ ρ(1 − ρ1 − ρ) S2 (y2 )−σ2 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)
σ2 σ1
2 2 2 2
× (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)
σ2 2 2 2
+ ρ1 (1 − ρ2 − ρ)S1 (t1 )−σ1 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)
σ1
2 2 2 2
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ1 σ2 2 2 2 2
+ (1 − ρ1 − ρ)(1 − ρ2 − ρ)(S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)(S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)
σ2 σ1
2 2 2 2
i
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)

© 2011 by Taylor and Francis Group, LLC

www.Ebook777.com
 Free ebooks ==> www.Ebook777.com
St2 ,y1 (t1 , y1 , t2 , y2 )
2 2 2 2 2 2
= µ2 (t2 )µ1 (y1 )(S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)−ρ1 /σ1 −1 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)−ρ2 /σ2 −1
2 2 2 2
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)−ρ/σ1 σ2 −1 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)−ρ/σ1 σ2 −1
σ1 σ1 σ2 σ2
1−ρ − ρ 1−ρ − ρ 1−ρ − ρ 1−ρ − ρ
× S1 (t1 ) 1 σ2 S1 (y1 ) 1 σ2 S2 (t2 ) 2 σ1 S2 (y2 ) 2 σ1
h σ1 2 2 2 2 2 2
× ρρ2 S2 (t2 )−σ2 S1 (y1 )−σ1 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)
σ2
2 2 2 2 2 2
+ ρ1 ρ2 S2 (t2 )−σ2 S1 (y1 )−σ1 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ1 2 2 2
+ ρ2 (1 − ρ1 − ρ)S2 (t2 )−σ2 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)
σ2
2 2 2 2
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)

Appendix
2 2 2 2 2 2
+ ρ2 S2 (t2 )−σ2 S1 (y1 )−σ1 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)(S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)
σ2 σ1 2 2 2
+ ρ(1 − ρ2 − ρ) S1 (y1 )−σ1 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)
σ1 σ2
2 2 2 2
× (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)
σ2 2 2 2 2 2 2
+ ρρ1 S2 (t2 )−σ2 S1 (y1 )−σ1 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ1
σ1 σ2 2 2 2
+ ρ(1 − ρ1 − ρ) S2 (t2 )−σ2 (S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)
σ2 σ1
2 2 2 2
× (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ2 2 2 2 2 2 2 2
+ ρ1 (1 − ρ2 − ρ)S1 (y1 )−σ1 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ1
σ1 σ2 2 2 2 2
+ (1 − ρ1 − ρ)(1 − ρ2 − ρ)(S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)(S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)
σ2 σ1

259
2 2 2 2
i
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)

© 2011 by Taylor and Francis Group, LLC

www.Ebook777.com
 Free ebooks ==> www.Ebook777.com

260
St2 ,y2 (t1 , y1 , t2 , y2 )
2 2 2 2 2 2
= µ2 (t2 )µ2 (y2 )(S1 (t1 )−σ1 + S1 (y1 )−σ1 − 1)−ρ1 /σ1 S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)−ρ2 /σ2 −2
2 2 2 2
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)−ρ/σ1 σ2 −1 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)−ρ/σ1 σ2 −1
σ1 σ1 σ2 σ2
1−ρ − ρ 1−ρ − ρ 1−ρ − ρ 1−ρ − ρ
× S1 (t1 ) 1 σ2 S1 (y1 ) 1 σ2 S2 (t2 ) 2 σ1 S2 (y2 ) 2 σ1
h σ2 2 2 2 2
× ρ2 22 S2 (t2 )−σ2 S2 (y2 )−σ2 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)2
σ1

Frailty Models in Survival Analysis

σ2 2 2 2 2 2 2
+ ρρ2 S2 (t2 )−σ2 S2 (y2 )−σ2 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)(S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)
σ1
σ2 σ2 2 2 2 2 2
+ ρ(1 − ρ2 − ρ) S2 (y2 )−σ2 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)2 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)
σ1 σ1
σ2 2 2 2 2 2 2
+ ρρ2 S2 (t2 )−σ2 S2 (y2 )−σ2 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ1
σ2 σ2 2 2 2 2 2
+ ρ(1 − ρ2 − ρ) S2 (t2 )−σ2 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)2 (S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ1 σ1
2 2 2 2 2 2
+ ρ2 (σ22 + ρ2 )S2 (t2 )−σ2 S2 (y2 )−σ2 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ2 2 2 2
+ ρ2 (1 − ρ2 − ρ)S2 (t2 )−σ2 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)
σ1
2 2 2 2
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ2 2 2 2
+ ρ2 (1 − ρ2 − ρ)S2 (y2 )−σ2 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)
σ1
2 2 2 2
× (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ2 2 2 2 2 2 2
i
+ (1 − ρ2 − ρ)2 (S2 (t2 )−σ2 + S2 (y2 )−σ2 − 1)2 (S1 (t1 )−σ1 + S2 (t2 )−σ2 − 1)(S1 (y1 )−σ1 + S2 (y2 )−σ2 − 1)
σ1

© 2011 by Taylor and Francis Group, LLC

www.Ebook777.com
 Free ebooks ==> www.Ebook777.com
Appendix 261

A.6 Quantitative Genetics

One aim of genetic analysis is to determine whether and to which extent
genetic variation may account for the variation of a specific phenotype (e.g.,
lifetime, susceptibility to disease). To address this question, measurements
of the phenotype must be combined with genetic information. In studies of
related individuals (families, twins, litter, etc.), genetic information is usually
available as pedigree. Based on methods of quantitative genetics developed
by Falconer (1990) and Neale and Cardon (1992), the relative role of genetic
factors is determined by estimating heritability.
Twin studies are one of the most widely used methods for quantifying the
influence of genetic and environmental factors on specific diseases. The classic
twin method, attributed to Francis Galton (1887), argues that, whereas in MZ
twins the differences in a pair must be due to environmental differences, for
DZ twins it includes effects associated with their genetic difference. Through
making a critical assumption that differences with respect to environmental
factors between identical pairs are of the same (average) strength as those
between DZ twins (equal environment assumption), inference can be made
about the influence of genetic factors simply by comparing MZ and DZ pair
correlations. In the case of binary traits (where the disease is either present or
not), concordance analysis provides a powerful and widely accepted method
in genetic epidemiology. Concordance rates are easy to calculate and allow
for a clear interpretation (McGue 1993, Gatz et al. 2000). In many practical
applications, time-to-event data (time of onset of disease, age at death) are
available, but usually in a censored form. Censoring of bivariate observations
can be a complex problem, as either or both individuals of a pair may be
subject to censoring, and the censoring times need not be the same for both
individuals. Furthermore, observed covariates are available in many cases.
Unfortunately, it is difficult to manage censored lifetime data and covariates
within the context of concordance analysis. A large part of the motivation
for the methodology of this paragraph is exploring the potential for censored
data and the inclusion of measured covariates.
Typical models of quantitative genetics can be incorporated with ease into
the correlated frailty model described earlier. Models of quantitative genetics
(Falconer 1990) are based on the decomposition of a phenotypic trait in a
sum of different components, that are assumed to be independent. Using this
approach, it is possible to estimate the proportion of the total variability of
the phenotype that is related to genetic factors. In particular, a heritability
estimate can be calculated for human longevity by identifying the phenotype
with the lifespan variable (McGue et al. 1993). Usually, heritability is defined
as the percentage of variation of the trait explained by the variation of genetic
factors.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
262 Frailty Models in Survival Analysis

Yashin and Iachine (1995a) suggested an approach based on the frailty Z

instead of the lifespan T . It is interesting to find out the relative importance
of genes and the environment in determining individual susceptibility toward
mortality (overall or cause specific). An advantage of this approach is that,
through the additive decomposition of the frailty term into a genetic and an
environmental component, one can obtain a competing risk structure for the
respective survival model. In other words, observed mortality is represented as
a sum of two terms: one depends on genetic, and the other on environmental,
parameters, both estimated from bivariate data. In more detail, let the frailty
be represented by

Z = genes + environment = A + D + I + C + E, (A.5)

where A represents additive genetic effects, D corresponds to genetic effects

caused by dominance, I denotes epistatic genetic effects, and C and E stand
for shared and nonshared environmental effects, respectively. All factors
are assumed to be independent. The associated variance proportions of the
components are defined as follows:

V (A) 2 V (D) 2 V (I) 2 V (C) 2 V (E)

a2 = , d = , i = , c = , e = .
V (Z) V (Z) V (Z) V (Z) V (Z)

The additive decomposition of the frailty variance and the correlation between
co-twins’ frailty hold that

1 = a2 + d2 + i2 + c2 + e2 (A.6)
ρ = ρ1 a2 + ρ2 d2 + ρ3 i2 + ρ4 c2 + ρ5 e2 , (A.7)
2 2 2 2 2
where lowercase letters a , d , i , c , e indicate the proportions of the total
variance σ 2 associated with the correspondent components of frailty, and ρi
(i = 1, ..., 5) are correlations between respective components within a twin
pair. In this case, broad sense heritability can be expressed as

H 2 = a2 + d2 + i2 ,
where the term a2 denotes small sense heritability. Standard assumptions
of quantitative genetics models specify different values of ρi (i = 1, ..., 5) for
monozygotic and dizygotic twins. In the case of monozygotic twins ρi = 1
(i = 1, ..., 4) and ρ5 = 0, while for dizygotic twins ρ1 = 0.5, ρ2 = 0.25,
ρ3 = m, ρ4 = 1, ρ5 = 0, and 0 ≤ m ≤ 0.25 is an unknown parameter.
Not all parameters of the genetic decomposition of frailty can be estimated
simultaneously, even under the assumption of no epistasis (i2 = 0). In this
case, it is only possible to conclude that the true heritability H 2 is in the
interval (Iachine 2002)
4
(ρMZ − ρDZ ) ≤ H 2 ≤ min{ρMZ , 2(ρMZ − ρDZ )}.
3

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
Appendix 263

The width of the interval is at most ρMZ −ρDZ , which represents the potential
error in heritability estimation using only the correlation coefficients of MZ
and DZ twins. The model in fact reduces to three equations (two relations
(A.7) for monozygotic and dizygotic twins each, and one constraint (A.6)),
allowing estimation of no more than three parameters at the same time. One
possibility is to consider an ACE (additive genetic – common environment –
uncommon environment) model. In this case, equations (A.6) and (A.7) lead
to

1 = a 2 + c2 + e 2
ρMZ = a2 + c2 (A.8)
ρDZ = 0.5a2 + c2 .

This system can be integrated into the correlated frailty model giving place
to a reparameterization of the original model. The only difference is that, if
we are interested in estimating the parameters of a genetic model, data for
monozygotic and dizygotic twins have to be analyzed simultaneously and a
likelihood function for combined data has to be drawn. Equally, other genetic
models can be obtained combining no more than three components of frailty.
It is necessary to note that heritability estimation requires assumptions
that are often difficult to verify in practice. For example, the trait must be
represented as an additive combination of uncorrelated environmental and
genetic factors, and the variances of phenotypic traits associated with related
individuals must be the same. The classical twin method is based on the
assumption that MZ and DZ twins have the same correlation with respect
to environmental factors (equal environment assumption). This assumption
is necessary for the identifiability of heritability, that is, so as to be able to
interpret the difference in concordance between MZ and DZ twins as being
explained in full by their difference in genetic concordance. However, without
doubt, the assumption is also sometimes questionable: MZ twins are generally
treated the same by their parents to a much greater extent than DZ twins.
This implies an overestimation of heritability, especially for behavioral traits.
This does not decrease the statistical attractiveness of this direction of
research. However, the interpretation of heritability estimates must be used
with care as pointed out by Feldman and Lewontin (1975).
After the age of six, death rates for Danish twins born between 1870
and 1900 are almost the same as those for the same cohorts of the Danish
population. The distributions of age at death for monozygotic twins are close
to those of dizygotic twins for both sexes (Christensen et al. 1995). Recent
papers dealing with twin cohorts born during the period 1870 – 1930 found
similar mortality patterns for Danish twins and the general Danish population
with respect to CHD (Wienke et al. 2001, Christensen et al. 2001). This
similarity suggests that it is possible to generalize genetic results from survival
analysis of twins to the total population with respect to mortality due to CHD.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
Free ebooks ==> www.Ebook777.com

www.Ebook777.com
 Free ebooks ==> www.Ebook777.com

References

Aalen, O.O. (1978) Nonparametric inference for a family of counting pro-

cesses. Annals of Statistics 6, 701–726.
Aalen, O.O. (1987) Two examples of modelling heterogeneity in survival
analysis. Scandinavian Journal of Statistics 14, 19–25.
Aalen, O.O. (1988) Heterogeneity in survival analysis. Statistics in Medicine
7, 1121–1137.
Aalen, O.O. (1992) Modelling heterogeneity in survival analysis by the com-
pound Poisson distribution. Annals of Applied Probability 4, 951–972.
Aalen, O.O., Bjertness, E., Sønju, T. (1995) Analysis of dependent survival
data applied to lifetimes of amalgam fillings. Statistics in Medicine 14,
1819–1829.
Aalen, O.O., Tretli, S. (1999) Analyzing incidence of testis cancer by means
of a frailty model. Cancer Causes and Control 10, 285–292.
Aalen, O., Hjort, N.L. (2002) Frailty models that yield proportional hazards.
Statistics & Probability Letters 58, 335–342.
Aalen, O.O., Borgan, O., Gjessing, H.K. (2008) Survival and Event History
Analysis: A Process Point of View. Springer, Berlin.
Abbring, J., van den Berg, G.J. (2003) The identifiability of the mixed pro-
portional hazards competing risks model. Journal of the Royal Statisti-
cal Society (B) 65, 701–710.
Abbring, J., van den Berg, G.J. (2007) The unobserved heterogeneity distri-
bution in duration analysis. Biometrika 94, 87–99.
Akaike, H. (1974) A new look at the statistical model identification. IEEE
Transactions on Automatic Control 19, 716–723.
Andersen, E.W. (2005) Two-stage estimation in copula models used in family
studies. Lifetime Data Analysis 11, 333–350.
Andersen, P.K., Gill, R.D. (1982) Cox’s regression model for counting pro-
cesses: a large sample study. Annals of Statistics 10, 1100–1120.
Andersen, P.K., Borgan, O., Gill, R.D., Keiding, N. (1993) Statistical Models
based on Counting Processes. Springer, New York.

265
www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
266 Frailty Models in Survival Analysis

Andersen, P.K., Klein, J.P., Zhang, M.-J. (1999) Testing for centre effects
in multi-centre survival studies: a Monte Carlo comparison of fixed and
random effects tests. Statistics in Medicine 18, 1489–1500.

Anderson, J.E., Louis, T.A. (1995) Survival analysis using a scale change
random effects model. Journal of the American Statistical Association
90, 669–679.

Anderson, J.E., Louis, T.A., Holm, N.V., Harvald, B. (1992) Time dependent
association measures for bivariate survival distributions. Journal of the
American Statistical Association 87, 641–650.

Balakrishnan, N., Peng, Y. (2006) Generalized gamma frailty model. Statis-

tics in Medicine 25, 2797–2816.

Bandeen-Roche, K.J., Liang, K-Y. (1996) Modelling failure-time associations

in data with multiple levels of clustering. Biometrika 83, 29–39.

Bandyopadhyay, D., Basu, A. (1990) On a generalization of a model by

Lindley and Singpurwalla. Advances in Applied Probability 22, 498–500.

Banerjee, S., Wall, M.M., Carlin, B.P. (2003) Frailty modeling for spatially
correlated survival data, with application to infant mortality in Min-
nesota. Biostatistics 4, 123–142.

Banerjee, S., Carlin, B.P. (2003) Semiparametric spatio-temporal frailty

modeling. Environmetrics 14, 523–535.

Barker, P., Henderson, R. (2004) Modelling converging hazards in survival

analysis. Lifetime Data Analysis 10, 263–281.

Barker, P., Henderson, R. (2005) Small sample bias in the gamma frailty
model for univariate survival. Lifetime Data Analysis 11, 265–284.

Beard, R.E. (1959) Note on some mathematical mortality models. In: The
Lifespan of Animals. G.E.W. Wolstenholme, M.O’Conner (eds.), Ciba
Foundation Colloquium on Ageing, Little, Brown, Boston, 302–311.

Bellamy, S., Li, Y., Ryan, L., Lipsitz, S., Canner, M., Wright, R. (2004)
Analysis of clustered and interval censored data from a community-
based study in asthma. Statistics in Medicine 23, 3607–3621.

Berkson, J., Gage, R. (1952) Survival curve for cancer patients following
treatment. Journal of the American Statistical Association 47, 501–515.

Berry, G. (2007) Relative risk and acceleration in lung cancer. Statistics in

Medicine 26, 3511–3517.

Bertoin, J. (1996) Lévy Processes. Cambridge University Press, Cambridge.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 267

Besag, J., Green, P. (1993) Spatial statistics and Bayesian computation.

Journal of the Royal Statistical Society (B) 55, 25–37.
Beutels, M., van Damme, P., Aelvoet, W., Desmyter, J., Dondeyne, F.,
Goilav, C., Mak, R, Muylle, L, Pierard, D, Stroobant, A, van Loock, F.,
Waumans, P., Vranckx, R. (1997) Prevalence of hepatitis A, B and C in
the Flemish population. European Journal of Epidemiology 13, 275–80.

Blumen, I., Kogan, M., McCarthy, P.J. (1955) The Industrial Mobility of
Labor as a Probability Process. Cornell University Press, New York.
Blossfeld, H.P., Hamerle, A. (1989) Unobserved heterogeneity in hazard rate
models – a test and an illustration from a study of career mobility.
Quality & Quantity 23, 129–141.

Blossfeld, H.P., Hamerle, A. (1992) Unobserved heterogeneity in event his-

tory models. Quality & Quantity 26, 157–168.
Boag, J.W. (1949) Maximum likelihood estimates of the proportion of pa-
tients cured by cancer therapy. Journal of the Royal Statistical Society
(B) 11, 15–44.
Bollmann, A., Blankenburg, T., Haerting, J., Kuss, O., Schütte, W., Dunst,
J., Neef, H. (2004) Survival of patients in clinical stages I-IIIb of non-
small-cell lung cancer treated with radiation therapy alone. Results of
a population-based study in Southern Saxony-Anhalt. Strahlentherapie
& Onkologie 180, 488 - 496.
Bolstad, W.M., Manda, S.O. (2001) Investigating child mortality in Malawi
using family and community random effects: a Bayesian analysis. Jour-
nal of the American Statistical Association 96, 12–19.
Bordes, L., Commenges, D. (2001) Asymptotics for homogeneity tests based
on a multivariate random effects proportional hazards model. Journal
of Multivariate Analysis 78, 83–102.
Bouchard, C. (1994) Genetics of obesity: overview and research directions.
In: The Genetics of Obesity. C. Bouchard (ed.), CRC Press, Boca
Raton, FL, pp. 223–233.
Box, G. (1976) Science and Statistics. Journal of the American Statistical
Association 71, 791–802.
Breslow, N. (1974) Covariance analysis of censored survival data. Biometrics
30, 89–99.
Breslow, N., Crowley, J. (1974) A large sample study of the life table and
product limit estimates under random censorship. Annals of Statistics
2, 437–453.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
268 Frailty Models in Survival Analysis

Bretagnolle, J., Huber-Carol, C. (1988) Effects of omitting covariates in

Cox’s model for survival data. Scandinavian Journal of Statistics 15,
125–138.

Broe̋t, P., Moreau, T., Lellouch, J., Asselain, B. (1999) Unobserved covari-
ates in the two-sample comparison of survival times: a maxmin efficiency
robust test. Statistics in Medicine 18, 1791–1800.

Brooks, S.P., Gelman A. (1998) Alternative methods for monitoring conver-

gence of iterative simulations. Journal of Computational and Graphical
Statistics 7, 434–455.

Cai, J., Prentice, R.L. (1995) Estimating equations for hazard ratio para-
meters based on correlated failure time data. Biometrika 82, 151–164.

Carling, K., Jacobson, T. (1995) Modeling unemployment duration in a

dependent competing risks framework: identification and estimation.
Lifetime Data Analysis 1, 111 - 122

Caroni, C., Crowder, M., Kimber, A. (2010) Proportional hazards models

with discrete frailty. Lifetime Data Analysis 16, 374–384.

Carvalho, M., Henderson, R., Shimakura, S., Sousa, I.P.S.C. (2003) Survival
of hemodialysis patients: modeling differences in risk of dialysis centers.
International Journal for Quality in Health Care 15, 189–196.

Cederlöf, R., Friberg, L., Jonsson, E., Kaij, L. (1961) Studies on similarity
diagnosis in twins with the aid of mailed questionnaires. Acta Genetica
11, 338–362.

Chamberlain, G. (1992) Heterogeneity, omitted variable bias, and duration

dependence. In: Longitudinal Analysis of Labour Market Data. J. Heck-
man, B. Singer (eds.) Cambridge University Press, pp. 3–38.

Chang, S.-H. (2004) Estimating marginal effects in accelerated failure time

models for serial sojurn times among repeated events. Lifetime Data
Analysis 10, 175–190.

Chang, I.-S., Wen, C.-C., Wu, Y.-J. (2007) A profile likelihood theory for
the correlated gamma-frailty model with current status family data.
Statistica Sinica 17, 1023–1046.

Chatterjee, N., Shih, J. (2001) A bivariate cure-mixture approach for mo-

deling familial association in diseases. Biometrics 57, 779–786.

Chen, M.-C., Bandeen-Roche, K. (2005) A diagnostic for association in bi-

variate survival models. Lifetime Data Analysis 11, 245–264.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 269

Christensen, K., Vaupel, J., Holm, N., Yashin, A. (1995) Mortality among
twins after age 6: fetal origins hypothesis versus twin method. British
Medical Journal 310, 432–436.

Christensen, K., Wienke, A., Skytthe, A., Holm, N., Vaupel, J., Yashin, A.
(2001) Cardiovascular mortality in twins and fetal origins hypothesis.
Twin Research 4, 344–349.

Chuang, S.K., Cai, T., Douglass, C.W., Wei, L.J., Dodson, T.B. (2005)
Frailty approach for the analysis of clustered failure time observations
in dental research. Journal of Dental Research 84, 54–58.

Chuang, S.K., Cai, T. (2006) Predicting clustered dental implant survival

using frailty methods. Journal of Dental Research 85, 1147–1151.

Claeskens, G., Nguti, R., Janssen, P. (2008) One-sided tests in shared frailty
models. Test 17, 69–82.

Clayton, D.G. (1978) A model for association in bivariate life tables and
its application in epidemiological studies of familial tendency in chronic
disease incidence. Biometrika 65, 141–151.

Clayton, D., Cuzick, J. (1985a) The semi-parametric Pareto model for re-
gression analysis of survival times. Proceedings of the 45th Session of
the International Statistical Institute 23, 1–18.

Clayton, D., Cuzick, J. (1985b) Multivariate generalizations of the propor-

tional hazards model. Journal of the Royal Statistical Society (A) 148,
82–117.

Clayton, D., Kaldor, J.M. (1985) Heterogeneity models as an alternative to

proportional hazards in cohort studies. Bulletin of the International
Statistical Institute 51, 1–16.

Clayton, D. (1991) A Monte Carlo method for Bayesian inference in frailty

models. Biometrics 47, 467–485.

Collett, D. (2003) Modelling Survival Data in Medical Research. Chapman

& Hall/CRC, London.

Commenges, D., Andersen P.K. (1995) Score test of homogeneity for survival
data. Lifetime Data Analysis 1, 145–160.

Commenges, D., Jacmin-Gadda, H. (1997) Generalized score test of homo-

geneity based on correlated random effects models. Journal of the Royal
Statistical Society (B) 59, 157–171.

Congdon, P. (1995) Modelling frailty in area mortality. Statistics in Medicine

14, 1859–1874.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
270 Frailty Models in Survival Analysis

Cook, R., Ng, E., Mukherjee, J., Vaughan, D. (1999) Two-state mixed re-
newal processes for chronic disease. Statistics in Medicine 18, 175–188.
Cook, R.J., Lawless, J.F. (2007) The Statistical Analysis of Recurrent Events.
Springer, Berlin.
Cox, D.R. (1959) Analysis of exponentially distributed life-times with two
types of failure. Journal of the Royal Statistical Society (B) 21, 411–421.
Cox, D.R. (1972) Regression models and life-tables. Journal of the Royal
Statistical Society (B) 34, 187–220.
Cox, D.R. (1975) Partial likelihood. Biometrika 62, 269–276.
Cox, D.R., Oakes, D. (1984) Analysis of Survival Data. Chapman & Hall,
London.
Crowder, M. (1989) A multivariate distribution with Weibull connections.
Journal of the Royal Statistical Society (B) 51, 93–107.
Cui, S., Sun, Y. (2004) Checking for the gamma frailty distribution under
the marginal proportional hazards frailty model. Statistica Sinica 14,
249–267.
Davis, H.T., Feldstein, M. (1979) The generalized Pareto law as a model for
progressively censored survival data. Biometrika 66, 299–306.
de Faire, U., Friberg, L., Lundman, T. (1975) Concordance for mortality
with special reference to ischaemic heart and cerebrovascular disease: a
study on the Swedish Twin Registry. Preventive Medicine 4, 509–517.
Dempster, A.P., Laird, N.M., Rubin, D.B. (1977) Maximum likelihood from
incomplete data via the EM algorithm. Journal of the Royal Statistical
Society (B) 39, 1–38.
Diamond, I.D., McDonald, J.W., Shah, I.H. (1986) Proportional hazards
models for current status data - application to the study of differentials
in age at weaning in Pakistan. Demography 23, 607–620.
di Serio, C. (1997) The protective impact of a covariate on competing failures
with an example from bone marrow transplantation. Lifetime Data
Analysis 3, 99–122.
Dominicus, A., Skrondal, A., Gjessing, H., Pedersen, N, Palmgren, J. (2006)
Likelihood ratio tests in behavioral genetics: problems and solutions.
Behavior Genetics 36, 331–340.
dos Santos, D., Davies, R, Francis, B. (1995) Nonparametric hazard ver-
sus nonparametric frailty distribution in modelling recurrence of breast
cancer. Journal of Statistical Planning and Inference 47, 111–127.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 271

Drapeau, M.D., Gassa, E.K., Simisona, M.D., Muellera, L.D., Rosea, M.R.
(2000) Testing the heterogeneity theory of late-life mortality plateaus
by using cohorts of Drosophila melanogaster. Experimental Gerontology
35, 71–84.
Drzewiecki, K.T., Ladefoged, C., Christensen, H.E. (1980a) Biopsy and prog-
nosis for cutaneous malignant melanomas in clinical stage I. Scandina-
vian Journal of Plastic and Reconstructive Surgery and Hand Surgery
14, 141–144.
Drzewiecki, K.T., Christensen, H.E., Ladefoged, C., Poulsen, H. (1980b)
Clinical course of cutaneous malignant melanoma related to histopatho-
logical criteria of primary tumour. Scandinavian Journal of Plastic and
Reconstructive Surgery and Hand Surgery 14, 229–234.
Duchateau, L., Janssen, P., Lindsey, P., Legrand, C., Nguti, R., Sylvester, R.
(2002) The shared frailty model and the power for heterogeneity tests in
multicenter trials. Computational Statistics & Data Analysis 40, 603–20.
Duchateau, L., Janssen, P., Kezic, I., Fortpied, C. (2003) Evolution of recur-
rent asthma event rate over time in frailty models. Journal of the Royal
Statistical Society (B) 52, 355–363.
Duchateau, L., Janssen, P. (2004) Penalized partial likelihood for frailties
and smoothing splines in time to first insemination models for dairy
cows. Biometrics 60, 608–614.
Duchateau, L., Janssen, P. (2008) The Frailty Model. Springer, New York.
Dunson, D.B., Chen, Z. (2004) Selecting factors predictive of heterogeneity
in multivariate event time data. Biometrics 60, 352–358.
Economou, P., Caroni, C. (2005) Graphical tests for the assumption of
gamma and inverse Gaussian frailty distributions. Lifetime Data Ana-
lysis 11, 565–582.
Economou, P., Caroni, C. (2008) Graphical tests for the frailty distribution
in the shared frailty model. Communications in Statistics - Simulations
and Computation 37, 978–992.
Efron, B. (1967) The two sample problem with censored data. Proceedings of
the 5th Berkeley Symposium on Mathematical Statistics and Probability,
pp. 831–853. University California Press
Efron, B. (1977) The efficiency of Cox’s likelihood function for censored data.
Journal of the American Statistical Association 72, 557–565.
Elbers, C., Ridder, G. (1982) True and spurious duration dependence: the
identifiability of the proportional hazard model. Review of Economic
Studies XLIX, 403–409.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
272 Frailty Models in Survival Analysis

Ellermann, R., Sullo, P., Tien, J.M. (1992) An alternative approach to model-
ing recidivism using quantile residual life functions. Operations Research
40, 485–504.

Emoto, S., Matthews, P. (1990) A Weibull model for dependent censoring.

Annals of Statistics 18, 1556–1577.

Falconer, D.S. (1990) Introduction to Quantitative Genetics. Longman Group,

New York.

Farewell, V.T. (1977) A model for a binary variable with time-censored ob-
servations. Biometrika 64, 43–46.

Farewell, V.T., Math, B., Math, M. (1977) The combined effect of breast
cancer risk factors. Cancer 40, 931–936.

Farewell, V.T. (1982) The use of mixture models for the analysis of survival
data with long-term survivors. Biometrics 38, 1041–1046.

Farrington, C., Kanaan, M., Gay, N. (2001) Estimation of the basic reproduc-
tion number for infectious diseases from age-stratified serological survey
data. Applied Statistics 50, 251–292.

Feldman, M., Lewontin, R. (1975) The heritability hang-up. Science 190,

1163–1168.

Feller, W. (1971) An Introduction to Probability Theory and its Applications.

John Wiley and Sons, New York.

Feuer, E.J., Wun, L.-M., Boring, C.C., Flanders, W.D., Timmel, M.J., Tong,
T. (1993) The lifetime risk of developing breast cancer. Journal of the
National Cancer Institute 85, 892–897.

Fine, J., Jiang, H. (2000) On association in a copula with time transforma-

tions. Biometrika 87, 559–571.

Fine, J., Glidden, D., Lee, K. (2003) A simple estimator for a shared frailty
regression model. Journal of the Royal Statistical Society (B) 65, 317–29.

Finkelstein, M. (2008) Failure Rate Modeling for Reliability and Risk. Springer,
New York.

Fisher R.A. (1958) Cancer and smoking. Nature 182, 596.

Fleming, T., Harrington, D. (1991) Counting Processes and Survival Analysis.

Wiley & Sons, Chichester.

Flinn, C., Heckman, J. (1982) New methods for analyzing structural models
of labor force dynamics. Journal of Econometrics 18, 115–168.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 273

Gail, M.H., Wieand, S., Piantadosi, S. (1984) Biased estimates of treatment

effect in randomized experiments with nonlinear regressions and omitted
covariates. Biometrika 71, 431–444.

Galton, F. (1887) Natural Inheritance. Macmillan, London.

Gatz, M., Pedersen, N.L., Crowe, M., Fiske, A. (2000) Defining discordance
in twin studies of risk and protective factors for late life disorders. Twin
Research 3, 159–164.

Gelman, A., Rubin, D.B. (1992) Inference from iterative simulation using
multiple sequences. Statistical Science 7, 457–511.

Geman, S., Geman, D. (1984) Stochastic relaxation, Gibbs distributions,

and the Bayesian restauration of images. IEEE Transactions on Pattern
Analysis and Machine Intelligence 6, 721–741.

Genest, C., MacKay, J. (1986) The joy of copulas: bivariate distributions

with uniform marginals. The American Statistician 40, 280–283.

Giard, N. (2001) A Four-Dimensional Correlated Frailty Model. GCA-Verlag,

Herdecke (Ph.D thesis).

Giard, N., Lichtenstein, P., Yashin, A. (2002) A multistate model for genetic
analysis of the ageing process. Statistics in Medicine 21, 2511–2526.

Gilks, W.R., Richardson, S., Spiegelhalter, D.G. (1996) Markov Chain Monte
Carlo in Practice. Chapman and Hall, London.

Gjessing, H.K., Aalen, O.O., Hjort, N.L. (2003) Frailty models based on
Lévy processes. Advances in Applied Probability 35, 532–550.

Glidden, D.V. (1999) Checking the adequacy of the gamma frailty model for
multivariate failure times. Biometrika 86, 381–393.

Glidden, D.V. (2000) A two-stage estimator of the dependence parameter

for the Clayton–Oakes model. Lifetime Data Analysis 6, 141–156.

Glidden, D.V. (2007) Pairwise dependence diagnostics for clustered failure

time data. Biometrika 94, 371–385.

Goeman, J.J., Le Cessie, S., Baatenburg de Jong, R.J., van de Geer, S.A.
(2004) Predicting survival using disease history: a model combining
relative survival and frailty. Statistica Neerlandica 58, 21–34.

Goethals, K., Janssen, P., Duchateau, L. (2008) Frailty models and copulas:
similarities and differences. Journal of Applied Statistics 35, 1071–1079.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
274 Frailty Models in Survival Analysis

Gompertz, B. (1825) On the nature of the function expressive of the law

of human mortality, and on a new mode of determining the value of
life contingencies. Philosophical Transactions of the Royal Society of
London 115, 513–585.

Gordon, S.C. (2000) A frailty model of negatively dependent competing risks.

Unpublished Draft.

Gorfine, M., Zucker, D., Hsu, L. (2006) Prospective survival analysis with
a general semiparametric shared frailty model: a pseudo full likelihood
approach. Biometrika 93, 735–741.

Gray, R.J. (1995) Tests for variation over groups in survival data. Journal
of the American Statistical Association 90, 198–203.

Greenwood, M., Yule, G.U. (1920) An inquiry into the nature of frequency
distributions representative of multiple happenings with particular ref-
erence to the occurrence of multiple attacks of disease or of repeated
accidents. Journal of the Royal Statistical Society 83, 255–279.

Greenwood, M. (1926) The natural duration of cancer. Reports on Public

Health and Medical Subjects. Her Majesty’s Stationery Office 33, 1–26.

Guo, G., Rodriguez, G. (1992) Estimating a multivariate proportional haz-

ards model for clustered data using the EM algorithm, with an applica-
tion to child survival in Guatemala. Journal of the American Statistical
Association 87, 969–976.

Guo, G. (1993) Use of sibling data to estimate family mortality effects in

Guatemala. Demography 30, 15–32.

Gupta, P. L., Gupta, R.D. (1990) A bivariate random environmental stress

model. Advances in Applied Probability 22, 501–503.

Gupta, R. C., Gupta, R.D. (2009) General frailty model and stochastic or-
derings. Journal of Statistical Planning and Inference 139, 3277–3287.

Gustafson, P. (1997) Large hierarchical Baysian analysis of multivariate sur-

vival data. Biometrics 53, 230–242.

Hamerle, A. (1992) A test for detecting overdispersion in multiple spell failure

time models. Biometrical Journal 34, 81–90.

Hanagal, D. (2009) Weibull extension of bivariate exponential regression

model with different frailty distributions. Statistical Papers 50, 29–49.

Hanagal, D. (2010) Modeling heterogeneity for bivariate survival data by the

compound Poisson distribution. Model Assisted Statistics and Applica-
tions 5, 1–9.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 275

Hastings, W.K. (1970) Monte Carlo sampling methods using Markov Chains
and their applications. Biometrika 57, 97–109.
Harvald, B., Hauge, M. (1970) Coronary occlusion in twins. Acta Geneticae
Medicae et Gemellologiae 19, 248–250.
Hauge, M. (1981) The Danish twin register. In: Prospective Longitudinal
Research. S. Mednich, A. Baert, B. Bachmann (eds.), Oxford Medical
Publisher, Oxford, pp. 217–222.
Haukka, J., Suvisaari, J., Lönnqvist, J. (2003) Increasing age does not de-
crease risk of schizophrenia up to age 40. Schizophrenia Research 61,
105–110.
Heath, A., Madden, P. (1995) Genetic influences on smoking behavior. In:
Behavior: Genetic Approaches in Behavioral Medicine. J.R. Turner,
L.R. Cardon, J.K. Hewitt (eds.), Plenum Press, New York, 45–63.
Heckman, J.J., Honoré B.E. (1989) The identifiability of the competing risks
model. Biometrika 76, 325–330.
Heckman, J.J., Singer, B. (1982a) The identification problem in econometric
models for duration data. In: Advances in Econometrics. W. Hilden-
brandt (ed.), Cambridge University Press, Cambridge, pp. 39–77.
Heckman, J.J., Singer, B. (1982b) Population heterogeneity in demographic
models. In: Multidimensional Mathematical Demography. K. Land, A.
Rogers (eds.), Academic Press, New York.
Heckman, J.J., Singer, B. (1984a) A method for minimizing the impact
of distributional assumptions in econometric models for duration data.
Econometrica 52, 271–320.
Heckman, J.J., Singer, B. (1984b) The identifiability of the proportional
hazard model. Review of Economic Studies 51, 231–241.
Heckman, J.J., Walker J.R. (1990) Estimating fecundability from data on
waiting-times to 1st conception. Journal of the American Statistical
Association 85, 283–294.
Heckman, J.J. (1991) Identifying the hand of the past: distinguishing state
dependence from heterogeneity. American Economic Review 81, 71–79.
Heckman, J.J., Taber, C.R. (1994) Econometric mixture models and more
general models for unobservables in duration analysis. Statistical Methods
in Medical Research 3, 279–302.
Henderson, R., Oman, P. (1999) Effect of frailty on marginal regression esti-
mates in survival analysis. Journal of the Royal Statistical Society (B)
61, 367–379.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
276 Frailty Models in Survival Analysis

Henderson, R., Shimakura, S. (2003) A serially correlated gamma frailty

model for longitudinal count data. Biometrika 90, 355–366.
Hens, N., Wienke, A., Aerts, M., Molenberghs, G. (2009) The correlated
and shared gamma frailty model for bivariate current status data: an
illustration for cross sectional serological data. Statistics in Medicine
28, 2785–2800.
Henschel, V., Engel, J., Hölzel, D., Mansmann, U. (2009) A semiparametric
Bayesian proportional hazards model for interval censored data with
frailty effects. BMC Medical Research Methodology 9, article 9.
Herskind, A.M., McGue, M., Iachine, I.A., Holm, N., Sørensen, T.I.A.,
Harvald, B., Vaupel, J.W. (1996a) Untangling genetic influences on
smoking, body mass index and longevity: a multivariate study of 2464
Danish twins followed for 28 years. Human Genetics 98, 467–475.
Herskind, A.M., McGue, M., Sørensen, T.I.A., Harvald, B. (1996b) Sex and
age specific assessment of genetic and environmental influences on body
mass index in twins. International Journal of Obesity 20, 106–113.
Hoem, J.M. (1990) Identifiability in hazard models with unobserved hetero-
geneity: the compatibility of two apparently contradictionary results.
Theoretical Population Biology 37, 124–128.
Holm, N.V. (1983) The use of twin studies to investigate causes of diseases
with complex etiology with a focus on cancer. Ph.D. thesis (in Danish),
Odense University, Odense.
Honoré, B. (1993) Identification results for duration models with multiple
spells. Review of Economic Studies 60, 241–246.
Hopper, J.L. (1993) Variance components for statistical genetics: applica-
tions in medical research to characteristics related to human diseases
and health. Statistical Methods in Medical Research 2, 199–223.
Horowitz, J.L. (1999) Semiparametric estimation of a proportional hazard
model with unobserved heterogeneity. Econometrica 67, 1001–1028.
Hougaard, P. (1984) Life table methods for heterogeneous populations. Bio-
metrika 71, 75–83.
Hougaard, P. (1986a) Survival models for heterogeneous populations derived
from stable distributions. Biometrika 73, 387–396.
Hougaard, P. (1986b) A class of multivariate failure time distributions. Bio-
metrika 73, 671–678.
Hougaard, P. (1987) Modeling multivariate survival. Scandinavian Journal
of Statistics 14, 291–304.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 277

Hougaard, P. (1991) Modeling heterogeneity in survival data. Journal of

Applied Probability 28, 695–701.

Hougaard, P., Harvald, B., Holm, N.V. (1992) Measuring the similarities
between the lifetimes of adult Danish twins born 1881 – 1930. Journal
of the American Statistical Association 87, 17–24.

Hougaard, P., Myglegaard, P., Borch-Johnsen, K. (1994) Heterogeneity mod-

els of disease susceptibility, with application to diabetic nephropathy.
Biometrics 50, 1178–1188.

Hougaard, P. (1995) Frailty models for survival data. Lifetime Data Analysis
1, 255–273.

Hougaard, P. (1999) Fundamentals of survival data. Biometrics 55, 13–22.

Hougaard, P. (2000) Analysis of Multivariate Survival Data. Springer, New

York.

Huang, X., Wolfe, R.A. (2002) A frailty model for informative censoring.
Biometrics 58, 510–520.

Huang, X., Wolfe, R.A., Hu, C. (2004) A test for informative censoring in
clustered survival data. Statistics in Medicine 23, 2089–2107.

Huber-Carol, C., Vonta, I. (2004) Frailty models for arbitrarily censored and
truncated data. Lifetime Data Analysis 10, 369–388.

Hutchinson, T.P., Lai, C.D. (1991) The Engineering Statistician’s Guide to

Continuous Bivariate Distributions. Rumsby, Adelaide.

Iachine, I.A. (1995) Parameter Estimation in the Bivariate Correlated Frailty

Model with observed Covariates via the EM-Algorithm. Research Re-
port, Population Studies of Ageing 16, Odense University.

Iachine, I.A. (2002) The Use of Twin and Family Survival Data in the
Population Studies of Aging: Statistical Methods Based on Multivariate
Survival Models. Ph.D. Thesis. Monograph 8, Department of Statistics
and Demography, University of Southern Denmark.

Iachine, I. (2004) Identifiability of Bivariate Frailty Models. Preprint 5,

Department of Statistics, University of Southern Denmark, Odense.

Iachine, I., Holm, N., Harris, J., Begun, A., Iachina, M., Laitinen, M.,
Kaprio, J., Yashin, A. (1998) How heritable is individual susceptibility
to death? The results of an analysis of survival data on Danish, Swedish
and Finnish twins. Twin Research 1, 196–205.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
278 Frailty Models in Survival Analysis

Iachine, I.A., Yashin, A.I. (1998) Identifiability of Bivariate Frailty Models

based on Additive Independent Components. Research Report 8, De-
partment of Statistics and Demography, University of Southern Den-
mark, Odense.
Ibrahim, J.G., Chen, M.-H., Sinha, D. (2001) Bayesian Survival Analysis.
Springer, New York.
Izumi, S., Ohtaki, M. (2004) Aspects of the Armitage-Doll gamma frailty
model for cancer incidence data. Environmetrics 15, 209–218.
Jeong, J. (2003) Efficiency of log-rank tests under dependent censoring.
Communications in Statistics, Theory and Methods 32, 1197–1211.
Jeong, J., Jung, S., Wieand, S. (2003) A parametric model for long-term
follow-up data from phase III breast cancer clinical trials. Statistics in
Medicine 22, 339–352.
Jeong, J., Jung, S. (2006) Rank tests for clustered survival data when de-
pendent subunits are randomized. Statistics in Medicine 25, 361–373.
Jin, X., Carlin, B. (2005) Multivariate parametric spatiotemporal models for
county level breast cancer survival data. Lifetime Data Analysis 11, 5-27
Joe, H. (1993) Parametric families of multivariate distributions with given
margins. Journal of Multivariate Analysis 46, 262–282.
Jones, B.L. (1998) A model for analyzing the impact of selective lapsation
on mortality. North American Actuarial Journal 2, 79–86.
Jonker, M., Bhulai, S., Boomsma, D., Ligthart, R., Posthuma, D., van der
Vaart, A. (2009) Gamma frailty model for linkage analysis with appli-
cation to interval-censored migraine data. Biostatistics 10, 187–200.
Jonker, M., Boomsma, D. (2010) A frailty model for interval censored family
survival data, applied to the age at onset of non-physical problems.
Lifetime Data Analysis 16, 299–315.
Juel, K., Helweg-Larsen, K. (1999) The Danish registers of causes of death.
Danish Medical Bulletin 46, 354–357.
Kalbfleisch, J., Prentice, R. (2002) The Statistical Analysis of Failure Time
Data. Wiley, New York.
Kalwitzki, M, Weiger, R., Axmann-Krcmar, D., Rosendahl, R. (2002) Caries
risk analysis: considering caries as an individual time-dependent pro-
cess. International Journal of Paediatric Dentistry 12, 132–142.
Kaplan, E., Meier, P. (1958) Nonparametric estimation from incomplete ob-
servations. Journal of the American Statistical Association 53, 457–481.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 279

Keiding, N., Andersen, P., Klein, J. (1997) The role of frailty models and ac-
celerated failure time models in describing heterogeneity due to omitted
covariates. Statistics in Medicine 16, 215–224.
Keiding, N. (1998) Selection effects and nonproportional hazards in survival
models and models for repeated events. In: Proceedings of the 19th
International Biometric Conference, Cape Town, pp. 241–250.
Keiding, N., Andersen, P. (eds.) (2006) Survival and Event History Analysis.
Wiley, New York.
Khazaeli, A., Xiu, L., Curtsinger, J.W. (1995) Stress experiments as a means
of investigating age-specific mortality in Drosophila melanogaster.
Experimental Gerontology 30, 177–184.
Kheiri, S., Meshkani, M.R., Faghihzadeh, S. (2005) A correlated frailty model
for analysing risk factors in bilateral corneal graft rejection for Kerato-
conus: a Bayesian approach. Statistics in Medicine 24, 2681–2693.
Kheiri, S., Kimber, A., Meshkani, M.R. (2007) Bayesian analysis of an inverse
Gaussian correlated frailty model. Computational Statistics and Data
Analysis 51, 5317–5326.
Kimber, A.C. (1996) A Weibull-based score test for heterogeneity. Lifetime
Data Analysis 2, 63–71.
Kimber, A.C., Zhu, C.Q. (1999) Diagnostics for a Weibull frailty model.
In: Statistical Inference and Design of Experiments. U.J. Dixit, M.R.
Satam (eds.) Narosa Publishing House, New Delhi, pp. 36–46.
Klein, J.P., Moeschberger, M.L. (2003) Survival Analysis - Techniques for
Censored and Truncated Data. Springer, New York.
Klein, J.P. (1992) Semiparametric estimation of random effects using the
Cox model based on the EM algorithm. Biometrics 48, 795–806.
Klein, J.P., Moeschberger, M., Li, Y., Wang, S. (1992) Estimating random
effects in the Framingham Heart Study. In: Survival Analysis: State of
the Art. J. Klein, P. Goel (eds.), Kluwer, Dordrecht, pp. 99–120.
Klein, J.P., Pelz, C., Zhang, M.-J. (1999) Random effects for censored data
by a multivariate normal regression model. Biometrics 55, 497–506.
Komárek, A., Lesaffre, E., Legrand, C. (2007) Baseline and treatment effect
heterogeneity for survival times between centers using a random effects
accelerated failure time model with flexible error distribution. Statistics
in Medicine 26, 5457–5472.
Kondo, K. (1977) The log-normal distribution of the incubation time of
exogenous diseases. Japanese Journal of Human Genetics 21, 217–237.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
280 Frailty Models in Survival Analysis

Korsgaard, I.R., Andersen, A.H. (1998) The additive genetic gamma frailty
model. Scandinavian Journal of Statistics 25, 255–270.
Korsgaard, I.R., Madsen, P., Jensen, J. (1998) Bayesian inference in the
semiparametric log normal frailty model using Gibbs sampling.
Genetics, Selection, Evolution 30, 241–256.
Kortram, R. A., Lenstra, A. J., Ridder, G., van Rooij, A. C. M. (1995)
Constructive identification of the mixed proportional hazards model.
Statistica Neerlandica 49, 269–281.
Kosorok, M.R., Lee, B.L., Fine, J.P. (2004) Semiparametric inference for
proportional hazards frailty regression models. Annals of Statistics 32,
1448–1491.
Koziol, J.A., Green, S.B. (1976) A Cramer–von Mises statistic for randomly
censored data. Biometrika 63, 465–474.
Kuk, A.Y.C., Chen, C.-H. (1992) A mixture model combining logistic re-
gression with proportional hazards regression. Biometrika 79, 531–541.
Kuß, O., Blankenburg, T., Haerting, J. (2008) A relative survival model for
clustered responses. Biometrical Journal 50, 408–418.
Lam, K., Kuk, Y. (1997) A marginal approach to estimation in frailty models.
Journal of the American Statistical Association 92, 985–990.
Lam, K.F., Lee, Y.W., Leung, T.L. (2002) Modeling multivariate survival
data by a semiparametric random effects proportional odds model.
Biometrics 58, 316–323.
Lam, K., Lee, Y. (2004) Merits of modelling multivariate survival data using
random effects proportional odds model.Biometrical Journal 46, 331–42.
Lam, K., Fong, D., Tang, O. (2005) Estimating the proportion of cured
patients in a censored sample. Statistics in Medicine 24, 1865–1879.
Lambert, P., Collett, D., Kimber, A., Johnson, R. (2004) Parametric accel-
erated failure time models with random effects and an application to
kidney transplant survival. Statistics in Medicine 23, 3177–3192.
Lancaster, T. (1979) Econometric methods for the duration of unemploy-
ment. Econometrica 47, 939–956.
Lancaster, T., Nickell, S. (1980) The analysis of re-employment probabilities
for the unemployed. Journal of the Royal Statistical Society (A) 143,
141 –165
Lawless, J.F. (2002) Statistical Models and Methods for Lifetime Data. Wiley,
New York.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 281

Lee, S., Lee, S. (2003) Testing heterogeneity for frailty distribution in shared
frailty model. Communications in Statistics, Theory and Methods 32,
2245–2253.

Lee, E.W., Wei, L.J., Amato, D.A. (1992) Cox-type regression analysis for
large numbers of small groups of correlated failure time observations.
In: Survival Analysis: State of the Art. J. Klein, P. Goel (eds.), Kluwer
Academic Publishers, Dordrecht, pp. 237–247.

Lee, S., Wolfe, R.A. (1998) A simple test for independent censoring under
the proportional hazards model. Biometrics 54, 1176–1182.

Legrand, C., Ducrocq, V., Janssen, P., Sylvester, R., Duchateau, L. (2005)
A Bayesian approach to jointly estimate centre and treatment by centre
heterogeneity in a proportional hazards model. Statistics in Medicine
24, 3789–3804.

Legrand, C., Duchateau, L., Sylvester, R., Janssen, P., van der Hage, J.,
van der Velde, C.J.H., Therasse, P. (2006) Heterogeneity in disease free
survival between centers: lessions learned from an EORTC breast cancer
trial. Clinical Trials 3, 10–18.

Legrand, C., Duchateau, L., Janssen, P., Ducrocq, V., Sylvester, R. (2008)
Validation of prognostic indices using the frailty model. Lifetime Data
Analysis 15, 59–78.

Li, H. (1999) The additive genetic gamma frailty model for linkage analysis
of age-of-onset variation. Annals of Human Genetics 63, 455–468.

Li, H. (2002) An additive genetic gamma frailty model for linkage analysis
of diseases with variable age of onset using nuclear families. Lifetime
Data Analysis 8, 315–334.

Li, H., Zhong, X. (2002) Multivariate survival models induced by genetic

frailties, with application to linkage analysis. Biostatistics 3, 57–75.

Li, C.-S., Taylor, J.M.G., Sy, J.P. (2001) Identifiability of cure models.
Statistics & Probability Letters 54, 389–395.

Li, Y., Betensky, R., Louis, D., Cairncross, J. (2002) The use of frailty hazard
models for unrecognized heterogeneity that interacts with treatment:
considerations of efficiency and power. Biometrics 58, 232–236.

Liang, K.-Y., Self, S., Bandeen-Roche, K., Zeger, S. (1995) Some recent
developments for regression analysis of multivariate failure time data.
Lifetime Data Analysis 1, 403–415.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
282 Frailty Models in Survival Analysis

Lichtenstein, P., de Faire, U., Floderus, B., Svartengren, M., Svedberg, P.,
Pedersen, N.L. (2002) The Swedish Twin Registry: a unique resource
for clinical, epidemiological and genetic studies. Journal of Internal
Medicine 252, 184–205.

Lillard, L.A. (1993) Simultaneous equations for hazards: marriage duration

and fertility timing. Journal of Econometrics 56, 189–217.

Lillard, L.A., Brian, M.J., Waite, M.J. (1995) Premarital cohabitation and
subsequent marital dissolution: a matter of self-selection? Demography
32, 437–457.

Lillard, L.A., Panis, C. (2000) aML User’s Guide and Reference Manual.
Econ-Ware, Los Angeles, CA.

Lim, H.J., Liu, J., Melzer-Lange, M. (2007) Comparison of methods for ana-
lyzing recurrent events data: application to the emergency department
visits of firearm victims. Accident Analysis and Prevention 39, 290–299.

Lin, D.Y. (1994) Cox regression analysis of multivariate failure time data:
the marginal approach. Statistics in Medicine 13, 2233–2247.

Lin, D., Robins, J., Wei, L. (1996) Comparing two failure time distributions
in the presence of depending censoring. Biometrika 83, 381–393.

Lindeboom, M., Van Den Berg, G.J. (1994) Heterogeneity in models for
bivariate survival: the importance of the mixing distribution. Journal
of the Royal Statistical Society (B) 56, 49–60.

Lindstrom, D.P. (1996) Economic opportunity in Mexico and return migra-

tion from the United States. Demography 33, 357–374.

Link, W.A. (1989) A model for informative censoring. Journal of the Ameri-
can Statistical Association 84, 749–752.

Littell, R.C., Milliken, G., Stroup, W., Wolfinger, R.D., Schabenberger, O.

(2006) SAS for Mixed Models (2nd edition). SAS Publishing, Cary, NC.

Liu, L., Wolfe, R.A., Huang, X. (2004) Shared frailty models for recurrent
events and a terminal event. Biometrics 60, 747–756.

Liu, L., Huang, X. (2008) The use of Gaussian quadrature for estimation in
frailty proportional hazard models. Statistics in Medicine 27, 2665–83.

Liu, L., Yu, Z. (2008) A likelihood reformulation method in non-normal

random effects models. Statistics in Medicine 27, 3105–3124.

Locatelli, I. (2003) Frailty models for twins duration data: a Bayesian ap-
proach. Ph.D. thesis, University Luigi Bocconi, Milano.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 283

Locatelli, I., Lichtenstein, P., Yashin, A.I. (2004) The heritability of breast
cancer: a Bayesian correlated frailty model applied to Swedish twins
data. Twin Research 7, 182–191.

Longini, I.M., Halloran, M.E. (1996) A frailty mixture model for estimating
vaccine efficacy. Applied Statistics 45, 165–173.

Ma, R., Krewski, D., Burnett, R.T. (2003) Random effects Cox models: a
Poisson modelling approach. Biometrika 90, 157–169.

Macdonald, A.S. (1999) Modeling the impact of genetics on insurance. North

American Actuarial Journal 3, 83–105.

Machin, D., Cheung, Y.B., Parmar, M. (2006) Survival Analysis: a Practical

Approach. Wiley, Chichester.

Mahé, C., Chevret, S. (1999) Estimating regression parameters and degree of

dependence for multivariate failure time data. Biometrics 55, 1078–84.

Makeham, W.M. (1860) On the law of mortality and the construction of

annuity tables. Journal of the Institute of Actuaries 8, 301–310.

Maller, R.A., Zhou, X. (1996) Survival Analysis with Long-Term Survivors.

Wiley & Sons, Chichester.

Maller, R.A., Zhou, X. (2002) Analysis of parametric models for competing

risks. Statistica Sinica 12, 725–750.

Maller, R.A., Zhou, X. (2003) Testing for individual heterogeneity in param-

eter models for event-history data. Mathematical Methods of Statistics
12, 276–304.

Mallick, M., Ravishanker, N. (2004) Multivariate Survival Analysis with

PVF Frailty Models. In: Advances in Ranking and Selection, Multiple
Comparisons, and Reliability Methodology and Applications. N. Bala-
krishnan, H.N. Nagaraja, N. Kannan (eds.), Birkhäuser, pp. 369–384.

Mallick, M., Ravishanker, N. (2006) PVF frailty models with a flexible base-
line hazard. International Journal of Statistics and Systems 1, 57–80.

Mallick, M., Ravishanker, N., Kannan, N. (2008) Bivariate positive stable

frailty models. Statistics and Probability Letters 78, 2371–2377.

Manatunga, A.K., Oakes, D. (1999) Parametric analysis of matched pair

survival data. Lifetime Data Analysis 5, 371–387.

Manda, S.O.M. (2001) A comparison of methods for analysing a nested frailty

model to child survival in Malawi. Australian and New Zealand Journal
of Statistics 43, 7–16.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
284 Frailty Models in Survival Analysis

Manda, S.O.M., Meyer, R. (2005) Bayesian inference for recurrent event data
using time-dependent frailty. Statistics in Medicine 24, 1263–1274.

Mantel, N., Bohidar, N.R., Ciminera, J.L. (1977) Mantel–Haenszel analyses

of litter-matched time-to-response data, with modifications for recovery
of interlitter information. Cancer Research 37, 3863–3868.

Manton, K., Stallard, E. (1981) Methods for evaluating the heterogeneity of

aging processes in human populations using vital statistics data: Ex-
plaining the black/white mortality crossover by a model of mortality
selection. Human Biology 53, 47–67.

Manton, K., Stallard, E., Vaupel, J. (1981) Methods for comparing mortality
experience of heterogeneous populations. Demography 18, 389–410.

Manton, K., Stallard, E., Vaupel, J. (1986) Alternative models for hetero-
geneity of mortality risks among the aged. Journal of the American
Statistical Association 81, 635–644.

Manton, K.G., Vaupel, J.W. (1995) Survival after the age of 80 in the United
States, Sweden, France, England, and Japan. New England Journal of
Medicine 333, 1232–1235.

Marenberg, M.E., Risch, N., Berkman, L.F., Floderus, B., de Faire, U. (1994)
Genetic susceptibility to death from coronary heart disease in a study
of twins. New England Journal of Medicine 330, 1041–1046.

Marshall, A. W., Olkin, I. (1988) Families of multivariate distributions.

Journal of the American Statistical Association 83, 834–841.

Marshall, A.W., Olkin, I. (2007) Life Distributions. Springer, Berlin.

Martinussen, T., Pipper, C.B. (2005) Estimation in the positive stable shared
frailty proportional hazards model. Lifetime Data Analysis 11, 99–115.

Martinussen, T., Scheike, T. (2006) Dynamic Regression Models for Survival

Data. Springer, Berlin.

Massonnet, G., Janssen, P., Burzykowski, T. (2008) Fitting conditional sur-

vival models to meta-analytic data by using a transformation toward
mixed-effects models. Biometrics 64, 834–842.

Matsuyama, Y., Sakamoto, J., Ohashi, Y. (1998) A Bayesian hierarchical

survival model for the institutional effects in a multi-centre cancer clin-
ical trial. Statistics in Medicine 17, 1893–1908.

McCall, B.P. (1994) Testing the proportional hazards assumption in the

presence of unmeasured heterogeneity. Journal of Applied Econometrics
9, 321–334.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 285

McGilchrist, C.A., Aisbett, C.W. (1991) Regression with frailty in survival

analysis. Biometrics 47, 461–466.

McGilchrist, C.A. (1993) REML estimation for survival models with frailty.
Biometrics 49, 221–225.

McGilchrist, C., Yau, K. (1996) Survival analysis with time dependent frailty
using a longitudinal model. Australian Journal of Statistics 38, 53–60.

McGue, M., Vaupel, J.W., Holm, N., Harvald, B. (1993) Longevity is mo-
derately heritable in a sample of Danish twins born 1870-1880. Journal
of Gerontology: Biological Sciences B 48, B237–B244.

Melino, A., Sueyoshi, G.T. (1990) A simple approach to the identifiability of

the proportional hazards model. Economics Letters 33, 63–68.

Metropolis, N., Rosenbluth, A., Rosenbluth, M., Teller A., Teller, E. (1953)
Equations of state calculations by fast computing machines. Journal of
Chemical Physics 21, 1087–1091.

Miller, R.G. (1981) Survival Analysis. Wiley & Sons, New York.

Miller, R.G. (1983) What price Kaplan–Meier? Biometrics 39, 1077–1081.

Moger, T.A., Aalen, O.O., Halvorsen, T.O., Storm, H.H., Tretli, S. (2004a)
Frailty modelling of testicular cancer incidence using Scandinavian data.
Biostatistics 5, 1–14.

Moger, T.A., Aalen, O.O., Heimdal, K., Gjessing, H.K. (2004b) Analysis of
testicular cancer data using a frailty model with familial dependence.
Statistics in Medicine 23, 617–632.

Moger, T.A., Aalen, O.O. (2005) A distribution for multivariate frailty based
on the compound Poisson distribution with random scale. Lifetime Data
Analysis 11, 41–59.

Moreno, L. (1994) Frailty selection in bivariate survival models: a cautionary

note. Mathematical Population Studies 4, 225–233.

Morgan, M.V., Adams, G.G., Campain, A.C., Wright, F.A.C. (2005) Assess-
ing sealant retention using a Poisson frailty model. Community Dental
Health 22, 237–245.

Morley, E., Perry, H. M., Miller, D. K. (2002) Something about frailty.

Journal of Gerontology: Medical Sciences 57A, M698–M704.

Mudholkar, G.S., Srivastava, D.K., Kollia, G. (1996) A generalization of the

Weibull distribution with application to the analysis of survival data.
Journal of the American Statistical Association 91, 1575–1583.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
286 Frailty Models in Survival Analysis

Mueller, L.D., Drapeau, M.D., Adams, C.S., Hammerle, C.W., Doyal, K.M.,
Jazayeri, A.J., Ly, T., Beguwala, S.A., Mamidi, A.R., Rose, M.R. (2003)
Statistical tests of demographic heterogeneity theories. Experimental
Gerontology 38, 373–386.

Murphy, S.A. (1994) Consistency in a proportional hazards models incor-

porating a random effect. Annals of Statistics 22, 712–731.

Murphy, S.A. (1995) Asymptotic theory for the frailty model. Annals of
Statistics 23, 182–198.

Murphy, S.A., Rossini, A., van der Vaart, A.W. (1997) Maximum likelihood
estimation in the proportional odds model. Journal of the American
Statistical Association 92, 968–976.

Murthy, D.N.P., Xie, M., Jiang, R. (2003) Weibull Models. Wiley, New York.

Naylor, J., Smith, A. (1982) Applications of a method for the efficient com-
putation of posterior distribution. Applied Statistics 31, 214–225.

Neal, R. (1997) Markov chain Monte Carlo methods based on ‘slicing’ the
density function. Technical Report 9722, Department of Statistics,
University of Toronto, Canada.

Neale, M.C., Cardon, L.R. (1992) Methodology for Genetic Studies of Twins
and Families. Kluwer, Dordrecht.

Nelsen, R.B. (2006) An Introduction to Copulas. Springer, New York.

Nelson, W. (1972) Theory and applications of hazard plotting for censored

failure data. Technometrics 14, 945–965.

Nelson, K.P., Lipsitz, S.R., Fitzmaurice, G.M., Ibrahim, J., Parzen, M.,
Strawderman, R. (2006) Use of the probability integral transformation
to fit nonlinear mixed-effects models with nonnormal random effects.
Journal of Computational and Graphical Statistics 15, 39–57.

Nickell, S. (1979) Estimating the probability of leaving unemployment.

Econometrica 47, 1249–1266.

Nielsen, G.G., Gill, R.D., Andersen, P.K., Sørensen, T.I.A. (1992) A counting
process approach to maximum likelihood estimation in frailty models.
Scandinavian Journal of Statistics 19, 25–43.

Noh, M., Ha, I.D., Lee, Y. (2006) Dispersion frailty models and HGLMs.
Statistics in Medicine 25, 1341–1354.

Oakes, D. (1982) A concordance test for independence in the presence of

censoring. Biometrics 38, 451–455.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 287

Oakes, D. (1989) Bivariate survival models induced by frailties. Journal of

the American Statistical Association 84, 487–493.

Oakes, D., Jeong, J.-H. (1998) Frailty models and rank tests. Lifetime Data
Analysis 4, 209–228.

O’Quigley, J., Stare, J. (2002) Proportional hazards models with frailties

and random effects. Statistics in Medicine 21, 3219–3233.

Orbe, J., Ferreira, E., Nùñez-Antòn, V. (2002) Comparing proportional haz-

ards and accelerated failure time models for survival analysis. Statistics
in Medicine 21, 3493–3510.

Panis, C., Lillard, L. (1995) Child mortality in Malaysia: explaining ethnic

differences and the recent decline. Population Studies 49, 463–479.

Pack, S.E., Morgan, B.J.T. (1990) A mixture model for interval-censored

time-to-response quantal assay data. Biometrics 46, 749–757.

Paik, M.C., Tsai, W.-Y., Ottman, R. (1994) Multivariate survival analysis

using piecewise gamma frailty. Biometrics 50, 975–988.

Pan, W. (2001) Using frailties in the accelerated failure time model. Lifetime
Data Analysis 7, 55–64.

Pankratz, V.S., de Andrade, M., Therneau, T. M. (2005) Random-effects

Cox proportional hazards model: general variance components methods
for time-to-event data. Genetic Epidemiology 28, 97–109.

Pareto, V. (1897) Cours d’Economie Politique. Rouge, Paris.

Parmar, M., Machin, D. (1995) Survival Analysis: A Practical Approach.

Wiley & Sons, New York.

Parner, E. (1998) Asymptotic theory for the correlated gamma-frailty model.

Annals of Statistics 26, 183–214.

Peng, Y., Dear, K.B.G., Denham, J.W. (1998) A generalized F mixture

model for cure rate estimation. Statistics in Medicine 17, 813–830.

Peng, Y., Zhang, J. (2008a) Estimation method of the semiparametric mix-

ture cure gamma frailty model. Statistics in Medicine 27, 5177–5194.

Peng, Y., Zhang, J. (2008b) Identifiability of a mixture cure frailty model.

Statistics and Probability Letters 78, 2604–2608.

Petersen, J. H. (1998) An additive frailty model for correlated lifetimes.

Biometrics 54, 646–661.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
288 Frailty Models in Survival Analysis

Petersen, L., Sørensen, T.I.A., Andersen, P.K. (2010) A shared frailty model
for case-cohort samples: Parent and offspring relations in an adoption
study. Statistics in Medicine 29, 924–931.

Peterson, A.V. (1977) Expressing the Kaplan–Meier estimator as a function

of empirical subsurvival functions. Journal of the American Statistical
Association 72, 854–858.

Peto, R., Lee, P.N., Paige, W.S. (1972) Statistical analysis of the bioassay
of continuous carcinogens. British Journal of Cancer 26, 258–261.

Pickles, A., Crouchley, R., Simonoff, E., Eaves, L., Meyer, J., Rutter, M.,
Hewitt, J., Silberg, J. (1994) Survival models for developmental genetic
data: age of onset of puberty and antisocial behavior in twins. Genetic
Epidemiology 11, 155–170.

Pike, M.C. (1966) A method of analysis of a certain class of experiments in

carcinogesis. Biometrics 22, 142–161.

Prentice, R.L., Williams, B.J., Peterson, A.V. (1981) On the regression anal-
ysis of multivariate failure time data. Biometrika 68, 373–379.

Price, D.L., Manatunga, A.K. (2001) Modelling survival data with a cured
fraction using frailty models. Statistics in Medicine 20, 1515–1527.

Qiou, Z., Ravishanker, N., Dey, D. (1999) Multivariate survival analysis with
positive stable frailties. Biometrics 55, 637–644.

Ridder, G. (1990) The non-parametric identification of generalized acceler-

ated failure-time models. Review of Economic Studies 57, 167–182.

Ries, L.A.G., Kosary, C.L., Hankey, B.F. (eds.) (1999) SEER Cancer Statis-
tics Review 1973-1999. National Cancer Institute, Bethesda, MD.

Ripatti, S., Palmgren, J. (2000) Estimation of multivariate frailty models

using penalised partial likelihood. Biometrics 56, 1016–1022.

Ripatti, S., Larsen, K., Palmgren, J. (2002) Maximum likelihood inference

for multivariate frailty models using an automated MCEM algorithm.
Lifetime Data Analysis 8, 349–360.

Roberts, G.O. (1996) Markov chain concepts related to sampling algorithms.

In: Markov Chain Monte Carlo in Practice. W.R. Gilks, S. Richardson,
D.J. Spiegelhalter (eds.), Chapman & Hall, London.

Rocha, C.S. (1996) Survival models for heterogeneity using the non-central
chi-squared distribution with zero degrees of freedom. In: Lifetime
Data: Models in Reliability and Survival Analysis. Jewell, N. et al.
(eds.) Kluwer Academic Publishers, Dordrecht, pp. 275–279.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 289

Rockwood, K. (2005) Frailty and its definition: a worthy challenge. Journal

of the American Geriatric Society 53, 1069–1070.

Rockwood, K., Mitnitski, A. (2007) Frailty in relation to the accumulation

of deficits. Journal of Gerontology (A) 62, 722–727.

Rondeau, V., Commenges, D., Joly, P. (2003) Maximum penalized likelihood

estimation in a gamma-frailty model. Lifetime Data Analysis 9, 139–53.

Rondeau, V., Gonzales, J. (2005) frailtypack: a computer program for the

analysis of correlated failure time data using penalized likelihood esti-
mation. Computer Methods and Programs in Biomedicine 80, 154–164.

Rondeau, V., Filleul, L., Joly, P. (2006) Nested frailty model using maximum
penalized likelihood estimation. Statistics in Medicine 25, 4036–4052.

Rondeau, V., Mathoulin-Pelissier, S., Jacqmin-Gadda, H., Brouste, V., Sou-

beyran, P. (2007) Joint frailty models for recurrent events and death
using maximum penalized likelihood estimation: application on cancer
events. Biostatistics 8, 708–721.

Rondeau, V., Michiels, S., Liquet, B., Pignon, J.P. (2008) Investigating trial
and treatment heterogeneity in an individual patient data meta-analysis
of survival data by means of the penalized maximum likelihood ap-
proach. Statistics in Medicine 27, 1894–1910.

Rondeau, V., Mazroui, Y., Gonzales, J. (2010) FRAILTYPACK: An R pack-

age for the analysis of correlated survival data with frailty models using
penalized likelihood estimation. Journal of Statistical Software, forth-
coming.

Rose, M.R., Rauser, C.L., Mueller, L.D., Benford, G. (2006) A revolution

for aging research. Biogerontology 7, 269–277.

Rosenthal, T.C., Puck, S.M. (1999) Screening for genetic risk of breast can-
cer. American Family Physician 59, 99–104.

Sahu, S.K., Dey, D.K., Aslanidou, H., Sinha, D. (1997) A Weibull regression
model with gamma frailties for multivariate survival data. Lifetime Data
Analysis 3, 123–137.

Sahu, S.K., Dey, D.K. (2004) On a Bayesian multivariate survival model with
skewed frailty. In: Skew-Elliptical Distributions and Their Applications:
A Journey Beyond Normality. M. Genton (ed.), Chapman & Hall/CRC,
Boca Raton, FL, pp. 321–338.

Samuelsen, S. (2003) Exact inference in the proportional hazard model: pos-

sibilities and limitations. Lifetime Data Analysis 9, 239–260.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
290 Frailty Models in Survival Analysis

Sankaran, P.G., Gleeja, V.L. (2008) Proportional reversed hazard and frailty
models. Metrika 68, 333–342.

Sargent, D. (1998) A general framework for random effects survival analysis

in the Cox proportional hazards setting. Biometrics 54, 1486–1497.

Sartwell, P.E. (1966) The incubation period and the dynamics of infectious
disease. American Journal of Epidemiology 83, 204–216.

Sastry, N. (1997) A nested frailty model for survival data, with an applica-
tion to the study of child survival in northeast Brazil. Journal of the
American Statistical Association 92, 426–435.

Schmermund, A., Stang, A., Mőhlenkamp, S., Eggebrecht, H., Baumgart,

D., Gilbert, V., Grőnemeyer, D., Seibel, R., Erbel, R. (2004) Prognostic
value of electron-beam computed tomography-derived coronary calcium
scores compared with clinical parameters in patients evaluated for coro-
nary artery disease. Zeitschrift fűr Kardiologie 93, 696–705.

Schmidt, P., Witte, A. (1989) Predicting criminal recidivism using split

population survival time models. Journal of Econometrics 40, 141–159.

Schmoor, C., Schumacher, M. (1997) The effect of covariate omission and

categorization when analyzing randomized trials with the Cox model.
Statistics in Medicine 16, 225–237.

Schnier, C., Hielm, S., Saloniemi, H.S. (2004) Comparison of the breeding
performance of cows in cold and warm loose-housing systems in Finland.
Preventive Veterinary Medicine 62, 135–151.

Schumacher, M., Olschewski, M., Schmoor, C. (1987) The impact of hetero-

geneity on the comparison of survival times. Statistics in Medicine 6,
773–784.

Schumacher, M. (1989) Frailty-Modelle und ihre Anwendung in der Medizin.

Biometrie und Informatik in Medizin und Biologie 20, 89–100.

Self, S.G., Liang, K.L. (1987) Asymptotic properties of maximum likelihood

estimators and likelihood ratio test under nonstandard conditions.
Journal of the American Statistical Association 82, 605–610

Semenchenko, G., Khazaeli, A., Curtsinger, J., Yashin, A. (2004a) Stress

resistence declines with age: analysis of data from a survival experiment
with drosophila melanogaster. Biogerontology 5, 17–30.

Semenchenko, G., Anisimov, V., Yashin, A. (2004b) Stressors and anti-

stressors: how do they influence life span in HER-2/neu transgenic mice?
Experimental Gerontology 39, 1499–1511.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 291

Shih, J.H., Louis, T.A. (1995a) Inferences on the association parameter in

copula models for bivariate survival data. Biometrics 51, 1384–1399.
Shih, J.H., Louis, T.A. (1995b) Assessing gamma frailty models for clustered
failure time data. Lifetime Data Analysis 1, 205–220.
Shih, J.H. (1998) A goodness-of-fit test for association in a bivariate survival
model. Biometrika 85, 189–200.
Shih, J.H., Lu, S.-E. (2009) Semiparametric estimation of a nested random
effects model for the analysis of multi-level clustered failure time data.
Computational Statistics and Data Analysis 53, 3864–3871.
Shu, Y., Klein, J.P. (1999) A SAS macro for the positive stable frailty model.
American Statistical Association Proceedings of the Statistical Comput-
ing Section 47–52.
Silva, G.L., Amaral-Turkman, M.A. (2004) Bayesian analysis of an additive
survival model with frailty. Communications in Statistics – Theory and
Methods 33, 2517–2533.
Sinha, D., Dey, K. (1997) Semiparametric Bayesian analysis of survival data.
Journal of the American Statistical Association 92, 1195–1212.
Sklar, A. (1959) Fonctions de répartition à n dimensions et leurs marges.
Publications de l’Institut Statistique de L’Université de Paris 8, 229–31.
Smith, P.J. (2002) Analysis of Failure and Survival Data. Chapman &
Hall/CRC, London.
Smith, A.F.M., Skene, A.M., Shaw, J.E.H., Naylor, J.C. (1987) Progress
with numerical and graphical methods for practical Bayesian statistics.
The Statistician 36, 75–82.
Spiegelhalter, D.J., Thomas, A., Best, N.G., Gilks, W. (1996) BUGS Exam-
ples, Volume 1, Version 0.5.
Spiegelhalter, D.J., Thomas, A., Best, N.G. (1999) WinBUGS Version 1.2.
User Manual.
Spiegelhalter, D.J., Best, N.G., Carlin, B.P., van der Linde, A. (2002) Bayesian
measures of model complexity and fit. Journal of the Royal Statistical
Society (B) 64, 583–639.
Spiekerman, C.F., Lin, D.Y. (1998) Marginal regression models for multivari-
ate failure time data. Journal of the American Statistical Association
93, 1164–1175.
Spilerman, S. (1972) Extensions of the mover–stayer model. American
Journal of Sociology 78, 599–626.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
292 Frailty Models in Survival Analysis

Sun, J. (2006) The Statistical Analysis of Interval-censored Failure Time

Data. Springer, New York.
Svensson, E., Moger, T.A., Tretli, S., Aalen, O.O., Grotmol, T. (2006) Frailty
modelling of colorectal cancer incidence in Norway: indications that
individual heterogeneity in risk is related to birth cohort. European
Journal of Epidemiology 21, 587–593.
Sy, J.P., Taylor, J.M. (2000) Estimation in a Cox proportional hazards cure
model. Biometrics 56, 227–236.
Tableman, M., Kim, J.S. (2004) Survival Analysis using S: Analysis of Time-
to-event Data. Chapman & Hall/CRC, London.
Therneau, T.M., Grambsch, P.M. (2000) Modeling Survival Data. Springer,
New York.
Therneau, T.M., Grambsch, P.M., Pankratz, V.S. (2003) Penalized survival
models and frailty. Journal of Computational and Graphical Statistics
12, 156–175.
Tomazella, V., Louzada-Neto, F., da Silva, G.L. (2006) Bayesian modeling
of recurrent events data with an additive gamma frailty distribution
and a homogeneous Poisson process. Journal of Statistical Theory and
Applications 5, 417–429.
Tomazella, V., Martins, C., Bernardo, J. (2008) Inference on the univariate
frailty model: a Bayesian reference analysis approach. Bayesian Infer-
ence and Maximum Entropy Methods in Science and Engineering 1073,
340–347.
Trussell, J., Richards, T. (1985) Correcting for unmeasured heterogeneity
in hazard models using the Heckman–Singer procedure. Sociological
Methodology 15, 242–276.
Tsai, W., Jewell, N., Wang, M. (1987) A note on the product-limit estimator
under right censoring and left truncation. Biometrika 74, 883–886.
Tsiatis, A. (1975) A nonidentifiability aspect of the problem of competing
risks. Proceedings of the National Academy of Sciences 72, 20–22.
Turnbull, B.W. (1976) The empirical distribution function with arbitrarily
grouped, censored and truncated data. Journal of the Royal Statistical
Society (B) 38, 290–295.
Tweedy, M. (1984) An index which distinguishes between some important
exponential families. In: Statistics: Applications and New Directions.
Proceedings of the Indian Statistical Institute Golden Jubilee Interna-
tional Conference. J. Ghosh, J. Roy (eds.), Indian Statistical Institute,
Calcutta, pp. 579–604.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 293

Vaida, F., Xu, R. (2000) Proportional hazards model with random effects.
Statistics in Medicine 19, 3309–3324.
van den Berg, G.J. (1992) Nonparametric tests for unobserved heterogeneity
in duration models. Working Paper, Free University of Amsterdam,
Amsterdam.
van den Berg, G.J. (2001) Duration models: specification, identification,
and multiple durations. In: Handbook of Econometrics. (Volume V)
J.J. Heckman, E. Leamer (eds.) North-Holland, Amsterdam.
van den Berg, G.J., Doblhammer-Reiter, G., Christensen, K. (2008) Being
born under adverse economic conditions leads to a higher cardiovascular
mortality rate later in life – evidence based on individuals born at dif-
ferent stages of the business cycle. Discussion paper 3635. Forschungs-
institut zur Zukunft der Arbeit.
Vaupel, J.W., Carey, J.R., Christensen, K., Johnson, T.E., Yashin, A.I.,
Holm, N.V., Iachine, I.A., Kannisto, V., Khazaeli, A.A., Liedo, P.,
Longo, V.D., Zeng, Y., Manton, K.G., Curtsinger, J.W. (1998). Biode-
mographic trajectories of longevity. Science 280, 855–860.
Vaupel, J., Manton, K., Stallard, E. (1979) The impact of heterogeneity in
individual frailty on the dynamics of mortality. Demography 16, 439–54.
Vaupel, J.W., Yashin, A.I. (1985) Heterogeneity’s ruses: some surprising
effects of selection on population dynamics. The American Statistician
39, 176–185.
Verweij, P.J., van Houwelingen, H.C., Stijnen, T. (1998) A goodness-of-fit
test for Cox’s proportional hazards model based on martingale residuals.
Biometrics 54, 1517–1526.
Viswanathan, B., Manatunga, A. (2001) Diagnostic plots for assessing the
frailty distribution in multivariate survival data. Lifetime Data Analysis
7, 143–155.
von Bortkiewicz, L. (1898) Das Gesetz der kleinen Zahlen. Teubner, Leipzig.
Vu, H.T.V., Knuiman, M.W. (2002) A hybrid ML-EM algorithm for calcula-
tion of maximum likelihood estimates in semiparametric shared frailty
models. Computational Statistics and Data Analysis 40, 173–187.
Vu, H.T.V. (2003) Parametric and semiparametric conditional shared gamma
frailty models with events before study entry. Communications in Statis-
tics, Simulation and Computation 32, 1223–1248.
Vu, H.T.V. (2004) Estimation in semiparametric conditional shared frailty
models with events before study entry. Computational Statistics and
Data Analysis 45, 621–637.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
294 Frailty Models in Survival Analysis

Wang, S.T., Klein, J.P., Moeschberger, M.L. (1995) Semi-parametric estima-

tion of covariate effects using the positive stable frailty model. Applied
Stochastic Models and Data Analysis 11, 121–133.

Warwick, J., Tabar, L., Vitak, B., Duffy, S.W. (2004) Time-dependent effects
on survival in breast carcinoma. Cancer 100, 1331–1336.

Wassell, J.T., Moeschberger, M.L. (1993) A bivariate survival model with

modified gamma frailty for assessing the impact of interventions.
Statistics in Medicine 12, 241–248.

Wei, L., Glidden, D. (1997) An overview of statistical methods for multiple

failure time data in clinical trials. Statistics in Medicine 16, 831–839.

Wei, L.J., Lin, D.Y., Weissfeld, L. (1989) Regression analysis of multivari-

ate incomplete failure time data by modeling marginal distributions.
Journal of the American Statistical Association 84, 1065–1073.

Weibull, W. (1939) A statistical theory of the strength of materials. Ingenior

Ventenskaps Akademien Handlinger 151. Generalstabens Litografiska
Anstalts Förlag, Stockholm.

Whittemore, A., Altshuler, B. (1976) Lung cancer incidence in cigarette

smokers: further analysis of Doll and Hill’s data for British physicians.
Biometrika 32, 805–816.

Wienke, A. (1998) An asymptotically optimal adaptive selection procedure in

a proportional hazards model with conditionally independent censoring.
Biometrical Journal 40, 963–978.

Wienke, A., Christensen, K., Holm, N., Yashin, A. (2000) Heritability of

death from respiratory diseases: an analysis of Danish twin survival
data using a correlated frailty model. In: Medical Infobahn for Europe.
A. Hasman et al. (eds.), IOS Press, Amsterdam, pp. 407–411.

Wienke, A., Holm, N., Skytthe, A., Yashin, A.I. (2001) The heritability of
mortality due to heart diseases: a correlated frailty model applied to
Danish twins. Twin Research 4, 266–274.

Wienke, A., Christensen, K., Skytthe, A., Yashin, A.I. (2002) Genetic anal-
ysis of cause of death in a mixture model with bivariate lifetime data.
Statistical Modelling 2, 89–102.

Wienke, A., Lichtenstein, P., Yashin, A.I. (2003a) A bivariate frailty model
with a cure fraction for modeling familial correlations in diseases.
Biometrics 59, 1178–1183.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 295

Wienke, A., Holm, N., Christensen, K., Skytthe, A., Vaupel, J., Yashin,
A.I. (2003b) The heritability of cause-specific mortality: a correlated
gamma-frailty model applied to mortality due to respiratory diseases in
Danish twins born 1870 – 1930. Statistics in Medicine 22, 3873–3887.

Wienke, A. (2004) Die Vererbbarkeit der Todesursache: ein correlated-frailty

Modell angewendet auf Dänische Zwillinge, geboren 1870 – 1930. In:
Lebenserwartung und Mortalität. R. Scholz, J. Flöthmann (eds.), Mate-
rialien zur Bevölkerungswissenschaft 111, Wiesbaden, pp. 81–98.

Wienke, A., Herskind, A., Christensen, K., Skytthe, A., Yashin, A. (2005a)
The heritability of CHD mortality in Danish twins after controlling for
smoking and BMI. Twin Research and Human Genetics 8, 53–59.

Wienke, A., Arbeev, K., Locatelli, I., Yashin, A.I. (2005b) A comparison
of different correlated frailty models and estimation strategies.
Mathematical Biosciences 198, 1–13.

Wienke, A., Locatelli, I., Yashin, I. (2006a) The modelling of a cure fraction
in bivariate time-to-event data. Austrian Journal of Statistics 35, 67–76.

Wienke, A., Lichtenstein, P., Czene, K., Yashin, A.I. (2006b) The role of
correlated frailty models in studies of human health, ageing and longe-
vity. In: Applications to Cancer and AIDS Studies, Genome Sequence
Analysis, and Survival Analysis. N. Balakrishnan, J. Auget, M. Mesbah,
G. Molenberghs (eds.), Birkhäuser, Boston, pp. 151–166.

Wienke, A., Ripatti, S., Palmgren, J., Yashin, A.I. (2010) A bivariate survival
model with compound Poisson frailty. Statistics in Medicine 29, 275–83.

Wintrebert, C.M.A., Putter, H., Zwinderman, A.H., van Houwelingen, J.C.

(2004)Centre-effect on survival after bone marrow transplantation:appli-
cation of time-dependent frailty models. Biometrical Journal 46, 512–25.

Wintrebert, C.M.A. (2007) Statistical modelling of repeated and multivariate

survival data. PhD thesis, Leiden University Medical Center, Leiden.

Wong, M.C.M., Lam, K.F., Lo, E.C.M. (2006) Multilevel modelling of clus-
tered grouped survival data using Cox regression model: an application
to ART dental restorations. Statistics in Medicine 25, 447–457.

Wu, D., Rea, S.L., Yashin, A.I., Johnson, T.E. (2006) Visualizing hidden het-
erogeneity in isogenic populations of C. elegans. Experimental Geron-
tology 41, 261–270.

Xu, R. (2004) Proportional hazards mixed models: a review with application

to twin models. Advances in Methodology and Statistics 1, 205–212.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
296 Frailty Models in Survival Analysis

Xu, L., Zhang, J. (2010) An EM-like algorithm for the semiparametric ac-
celerated failure time gamma frailty model. Computational Statistics
and Data Analysis 54, 1467–1474.
Xue, X., Brookmeyer, R. (1996) Bivariate frailty model for the analysis of
multivariate survival time. Lifetime Data Analysis 2, 277–290.
Xue, X., Ding, Y. (1999) Assessing heterogeneity and correlation of paired
failure times with the bivariate frailty model. Statistics in Medicine 18,
907–918.
Yamaguchi, T., Ohashi, Y. (1999) Investigating centre effects in a multi-
centre clinical trial of superficial bladder cancer. Statistics in Medicine
18, 1961–1971
Yamaguchi, T., Ohashi, Y., Matsuyama, Y. (2002) Proportional hazard mod-
els with random effects to examine centre effects in multicentre cancer
clinical trials. Statistical Methods in Medical Research 11, 221–236.
Yashin, A.I., Begun, A., Iachine, I.A. (1999a) Genetic factors in susceptibility
to death: a comparative analysis of bivariate survival models. Journal
of Epidemiology and Biostatistics 4, 53–60.
Yashin, A.I., Herskind, A.-M., Begun, A.Z., Iachine, I.A. (1999b) Survival
models for genetic analysis with random effects depending on observed
covariates. Unpublished draft.
Yashin, A.I., Iachine, I.A. (1995a) Genetic analysis of durations: Correlated
frailty model applied to survival of Danish twins. Genetic Epidemiology
12, 529–538.
Yashin, A.I., Iachine, I.A. (1995b) Survival of related individuals: an
extension of some fundamental results of heterogeneity analysis.
Mathematical Population Studies 5, 321–39.
Yashin, A.I., Iachine, I.A. (1996) Random effect models of bivariate survival:
quadratic hazard as a new alternative. In: Transactions of Symposium
i Anvendt Statistik. G. Kristensen (ed.), Institute of Economy, Odense
University, pp. 87–101.
Yashin, A.I., Iachine, I.A. (1997) How frailty models can be used for evaluat-
ing longevity limits: taking advantage of an interdisciplinary approach.
Demography 34, 31–48.
Yashin, A.I., Iachine, I. (1999a) Dependent hazards in multivariate survival
problems. Journal of Multivariate Analysis 71, 241–261.
Yashin, A.I., Iachine, I. (1999b) What difference does the dependence between
durations make? Insights for population studies of aging. Lifetime Data
Analysis 5, 5–22.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
References 297

Yashin, A.I., Manton, K.G. (1997) Effects of unobserved and partially ob-
served covariate processes on system failure: a review of models and
estimation strategies. Statistical Science 12, 20–34.

Yashin, A.I., Manton, K.G., Iachine, I.A. (1996) Genetic and environmental
factors in duration studies: multivariate frailty models and estimation
strategies. Journal of Epidemiology and Biostatistics 1, 115–120.

Yashin, A.I., Vaupel, J.W., Iachine, I.A. (1993) Correlated individual frailty:
an advantageous approach to survival analysis of bivariate data.
Working Paper: Population Studies of Aging 7, Odense University.

Yashin, A.I., Vaupel, J.W., Iachine, I.A. (1995) Correlated individual frailty:
an advantageous approach to survival analysis of bivariate data.
Mathematical Population Studies 5, 145–159.

Yau, K., McGilchrist, C. (1997) Use of generalised linear mixed models for
the analysis of clustered survival data. Biometrical Journal 39, 3–11.

Yau, K.K.W., McGilchrist, C.A. (1998) ML and REML estimation in survival

analysis with time dependent correlated frailty. Statistics in Medicine
17, 1201–1213.

Yau, K.K.W. (2001) Multilevel models for survival analysis with random
effects. Biometrics 57, 96–102.

Yin, G., Ibrahim, J.G. (2005) A class of Bayesian shared gamma frailty
models with multivariate failure time data. Biometrics 61, 208–216.

Yu, B., Peng, Y. (2008) Mixture cure models for multivariate survival data.
Computational Statistics and Data Analysis 52, 1524–1532.

Zahl, P.-H. (1994) Correlated frailty models - modelling of unobserved cor-

related risks of death. Norwegian Journal of Epidemiology 4, 90–94.

Zahl, P. (1997) Frailty modelling for the excess hazard. Statistics in Medicine
16, 1573–1585.

Zdravkovic, S., Wienke, A., Pedersen, N.L., Marenberg, M.E., Yashin, A.I.,
de Faire, U. (2002) Heritability of death from coronary heart disease: a
36 years follow-up of 20,966 Swedish twins. Journal of Internal Medicine
252, 247–254.

Zdravkovic, S., Wienke, A., Pedersen, N.L., Marenberg, M.E., Yashin, A.I.,
de Faire, U. (2004) Genetic influences on CHD-death and the impact of
known risk factors: comparison of two frailty models. Behavior Genetics
34, 585–591.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC
 Free ebooks ==> www.Ebook777.com
298 Frailty Models in Survival Analysis

Zhang, J., Peng, Y. (2007) An alternative estimation method for the ac-
celerated failure time frailty model. Computational Statistics and Data
Analysis 51, 4413–4423.
Zheng, M., Klein, J.P. (1995) Estimates of marginal survival for dependent
competing risks based on assumed copula. Biometrika 82, 127–138.
Zhong, X., Li, H. (2004) Score tests of genetic association in the presence of
linkage based on the additive genetic gamma frailty model. Biostatistics
5, 307–327.

www.Ebook777.com
© 2011 by Taylor and Francis Group, LLC