Bayesian Approaches

in Oncology Using R

and OpenBUGS

Bayesian Approaches

in Oncology Using R

and OpenBUGS

Atanu Bhattacharjee
First edition published 2021
by CRC Press
6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742
and by CRC Press
2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN
© 2021 Taylor & Francis Group, LLC
[First edition published by Willan 2008]
[Sixth edition published by Routledge 2009]
CRC Press is an imprint of Taylor & Francis Group, LLC
The right of Atanu Bhattacharjee to be identified as author of this work has been asserted by him in
accordance with sections 77 and 78 of the Copyright, Designs and Patents Act 1988.
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 repro­
duced in this publication and apologize to copyright holders if permission to publish in this form
has not been obtained. If any copyright material has not been acknowledged please write and let us
know so we may rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information
storage or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, access www.copyright.
com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA
01923, 978-750-8400. For works that are not available on CCC please contact mpkbookspermis­
[email protected]
Trademark notice: Product or corporate names may be trademarks or registered trademarks and
are used only for identification and explanation without intent to infringe.

ISBN: 978-0-367-35050-5 (hbk)

ISBN: 978-0-429-32944-9 (ebk)

Typeset in Computer Modern font

by KnowledgeWorks Global Ltd.
I dedicated this book to Dr.Balekudaru Shantha,

Ophthalmologist, Director, Sankara Nethralaya

Chennai, for preserving my vision in a critical time.

Contents

Preface xiii

Author xv

I Bayesian in Clinical Research 1

1 Introduction to R and OpenBUGS 3

1.1 Introduction to R . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 How to Install R . . . . . . . . . . . . . . . . . . . . . . . . 4

1.3 Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.4 How to Install RStudio . . . . . . . . . . . . . . . . . . . . . 6

1.5 OpenBUGS . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Sample Size Determination 13

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.2 The Sample Size Formula . . . . . . . . . . . . . . . . . . . 14

2.3 Measured Outcomes . . . . . . . . . . . . . . . . . . . . . . 15

2.3.1 Time to event endpoints . . . . . . . . . . . . . . . . 15

2.3.2 Sample size with event-driven trials . . . . . . . . . . 16

2.4 Bayesian Sample Size Determination . . . . . . . . . . . . . 18

2.5 Study Objective and Sample Size Determination . . . . . . 19

2.6 Bayesian Survival Analysis Sample Size Calculation with R 22

2.7 Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.8 Posterior Error Approach . . . . . . . . . . . . . . . . . . . 23

2.9 Different Sample Size Determination Packages in R . . . . . 23

2.9.1 BAEssd . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.9.2 BayesianPower . . . . . . . . . . . . . . . . . . . . . . 24

2.9.3 NPHMC . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.9.4 PowerTOST . . . . . . . . . . . . . . . . . . . . . . . 26

2.9.5 SampleSize4ClinicalTrials . . . . . . . . . . . . . . . . 26

2.9.6 SSRMST . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.9.7 powerSurvEpi . . . . . . . . . . . . . . . . . . . . . . 28

2.9.8 SampleSizeMeans . . . . . . . . . . . . . . . . . . . . 28

vii
viii Contents

3 Study Design-I 31

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.2 Bayesian in Early Phase Oncology Trial . . . . . . . . . . . 32

3.2.1 Rule-based designs . . . . . . . . . . . . . . . . . . . 32

3.2.1.1 3+3 design . . . . . . . . . . . . . . . . . . 33

3.2.1.2 Pharmacologically guided dose escalation . 33

3.2.1.3 Accelerated titration designs . . . . . . . . . 34

3.2.2 Model-based designs for determining the MTD . . . . 34

3.2.2.1 Continual reassessment method (CRM) . . 34

3.3 Study Design Package Using R . . . . . . . . . . . . . . . . 35

4 Study Design-II 39

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

4.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

4.2.1 Estimating treatment success . . . . . . . . . . . . . 41

4.2.2 Treatment difference testing . . . . . . . . . . . . . . 42

4.3 Illustration with Bayesian Using R . . . . . . . . . . . . . . 43

4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

5 Optimum Biological Dose Selection 47

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

5.2 Illustration with Head and Neck Cancer Data . . . . . . . . 48

5.3 Toxicity Profile Testing . . . . . . . . . . . . . . . . . . . . 48

5.4 Logistic Response Model . . . . . . . . . . . . . . . . . . . . 49

5.5 Dose Selection Algorithm . . . . . . . . . . . . . . . . . . . 51

5.6 Quadratic Logistic Design . . . . . . . . . . . . . . . . . . . 52

5.7 Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

5.8 Bayesian Dose Selection Using OpenBUGS . . . . . . . . . 57

5.9 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

5.10 Dose Finding Package Using R . . . . . . . . . . . . . . . . 60

5.10.1 DoseFinding . . . . . . . . . . . . . . . . . . . . . . . 60

II Bayesian in Time-to-Event Data Analysis 63

6 Survival Analysis 65

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

6.1.1 Kaplan-Meier estimator . . . . . . . . . . . . . . . . . 66

6.1.2 Nelson-Aalen estimator . . . . . . . . . . . . . . . . . 67

6.2 Bayesian in Survival Analysis . . . . . . . . . . . . . . . . . 68

6.2.1 Cox’s proportional hazards model . . . . . . . . . . . 68

6.2.2 Hazard ratio . . . . . . . . . . . . . . . . . . . . . . . 69

6.2.3 Partial likelihood function . . . . . . . . . . . . . . . 69

Contents ix

6.2.4 Stratified Cox model . . . . . . . . . . . . . . . . . . 70

6.2.5 Wald-score and likelihood ratio tests . . . . . . . . . 71

6.2.6 Diagnostics for Cox’s PH model . . . . . . . . . . . . 71

6.2.7 Schoenfeld residuals . . . . . . . . . . . . . . . . . . . 72

6.3 Bayesian Survival Analysis Using R . . . . . . . . . . . . . . 73

6.3.1 BayesSurvival . . . . . . . . . . . . . . . . . . . . . . 76

6.3.2 muhaz . . . . . . . . . . . . . . . . . . . . . . . . . . 76

6.3.3 Bayesian in Kaplan-meier estimator . . . . . . . . . . 76

6.3.4 Bayesian Cox proportional hazards model . . . . . . 79

7 Competing Risk Data Analysis 87

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

7.2 Competing Risk as Bivariate Random Variable . . . . . . . 88

7.3 Cumulative Incidence Rate . . . . . . . . . . . . . . . . . . 89

7.4 Competing Risk Model Using R . . . . . . . . . . . . . . . . 89

7.5 Cause-Specific Hazard Model . . . . . . . . . . . . . . . . . 91

7.6 Bayesian Information Criteria . . . . . . . . . . . . . . . . . 92

7.7 Illustration Using R . . . . . . . . . . . . . . . . . . . . . . 94

8 Frailty Data Analysis 97

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

8.2 Frailty Model . . . . . . . . . . . . . . . . . . . . . . . . . . 99

8.3 Motivating Example . . . . . . . . . . . . . . . . . . . . . . 99

8.4 Bayesian in Frailty Survival . . . . . . . . . . . . . . . . . . 104

8.4.1 Frailty modeling . . . . . . . . . . . . . . . . . . . . . 104

8.4.2 Frailty on recurrent events . . . . . . . . . . . . . . . 104

8.4.3 Generalized accelerated failure time (GAFT) frailty

model . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

8.4.4 Illustration with leukemia data using R . . . . . . . . 105

9 Relative Survival Analysis 111

9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

9.2 Relative Survival Analysis . . . . . . . . . . . . . . . . . . . 113

9.3 Data Methodology . . . . . . . . . . . . . . . . . . . . . . . 114

9.4 Illustration with R . . . . . . . . . . . . . . . . . . . . . . . 115

9.5 Piecewise Hazard Function . . . . . . . . . . . . . . . . . . . 118

9.6 Piecewise Hazard testing . . . . . . . . . . . . . . . . . . . . 118

9.7 Piecewise Hazard Function Analysis . . . . . . . . . . . . . 120

9.8 Different Relative Survival Analysis Package with R . . . . 120

9.8.1 flexrsurv . . . . . . . . . . . . . . . . . . . . . . . . . 120

9.8.2 relsurv . . . . . . . . . . . . . . . . . . . . . . . . . . 123

9.8.3 survexp.fr . . . . . . . . . . . . . . . . . . . . . . . . 124

9.9 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

x Contents

III Bayesian in Longitudinal Data Analysis 129

10 Longitudinal Data Analysis 131

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

10.2 Advantages of Longitudinal Analysis . . . . . . . . . . . . . 132

10.3 Limitation of Longitudinal Analysis . . . . . . . . . . . . . . 132

10.4 Mixed Effect Model . . . . . . . . . . . . . . . . . . . . . . . 132

10.5 Different R Packages for Longitudinal Data Analysis . . . . 133

10.6 Random-Effect Model with R . . . . . . . . . . . . . . . . . 134

10.6.1 bayeslongitudinal . . . . . . . . . . . . . . . . . . . . 134

10.7 Illustration with Quality of Life Using OpenBUGS . . . . . 136

10.8 Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

11 Missing Data Analysis 143

11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

11.2 Different Types of Missing Data . . . . . . . . . . . . . . . . 144

11.2.1 Missing completely at random (MCAR) . . . . . . . . 144

11.2.2 Missing at random (MAR) . . . . . . . . . . . . . . . 144

11.2.3 Missing not at random (NMAR) . . . . . . . . . . . . 144

11.3 Different Softwares . . . . . . . . . . . . . . . . . . . . . . . 145

11.4 Illustration with Lung Cancer Data . . . . . . . . . . . . . . 145

11.5 Different Package for Missing Data with R . . . . . . . . . . 146

11.5.1 BMTAR . . . . . . . . . . . . . . . . . . . . . . . . . 146

11.5.2 NestedCohort . . . . . . . . . . . . . . . . . . . . . . 148

11.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

12 Joint Longitudinal and Survival Analysis 153

12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

12.2 Data Methodology . . . . . . . . . . . . . . . . . . . . . . . 154

12.3 The Longitudinal Model . . . . . . . . . . . . . . . . . . . . 155

12.4 The Survival Model . . . . . . . . . . . . . . . . . . . . . . . 155

12.5 The Joint Model . . . . . . . . . . . . . . . . . . . . . . . . 156

12.6 Submodels Specification . . . . . . . . . . . . . . . . . . . . 156

12.7 Different R Package for Joint Longitudinal Model . . . . . . 165

12.7.1 joint.Cox . . . . . . . . . . . . . . . . . . . . . . . . . 165

12.7.2 JM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

13 Covariance Modeling 169

13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

13.2 Covariance Patterns . . . . . . . . . . . . . . . . . . . . . . 170

13.3 Covariance Patterns Challenges . . . . . . . . . . . . . . . . 173

13.4 Illustration with Protein-Gene Expression Time-Course Data 173

Contents xi

IV Bayesian in Diagnostics Test Statistics 183

14 Bayesian Inference in Mixed-Effect Model 185

14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

14.2 Likelihood Function with Doses and Measurement Process . 187

14.3 Linear Mixed-Effects Model . . . . . . . . . . . . . . . . . . 188

14.4 Autoregressive Linear Mixed-Effects Model . . . . . . . . . 189

14.5 Linear Mixed-Effects Model Compatible with Dose Response 189

14.6 Autoregressive Linear Mixed Effects Model Compatible with

Dose Response . . . . . . . . . . . . . . . . . . . . . . . . . 190

14.7 Generating Data . . . . . . . . . . . . . . . . . . . . . . . . 190

14.8 Computation Support . . . . . . . . . . . . . . . . . . . . . 192

14.9 Power Analysis of the Performed Models . . . . . . . . . . . 194

14.10 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194

15 Concordance Analysis 197

15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

15.2 Computational Methodology . . . . . . . . . . . . . . . . . 199

15.3 Bayes Factor . . . . . . . . . . . . . . . . . . . . . . . . . . 202

15.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

15.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206

16 High-Dimensional Data Analysis 209

16.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

16.2 Heatmap with R . . . . . . . . . . . . . . . . . . . . . . . . 210

16.3 Principal Component Analysis with R . . . . . . . . . . . . 212

16.4 Penalized Partial Log-Likelihood (PLL) . . . . . . . . . . . 213

16.5 Estimating the Predictive PLL . . . . . . . . . . . . . . . . 214

16.6 Integrated Prediction Error Curve (IPEC) . . . . . . . . . . 214

16.7 Ridge Estimators in Cox Regression . . . . . . . . . . . . . 214

16.8 High-Dimensional Data Analysis Using R . . . . . . . . . . 215

16.9 Different Packages with R . . . . . . . . . . . . . . . . . . . 222

16.9.1 BAMA . . . . . . . . . . . . . . . . . . . . . . . . . . 222

16.9.2 countgmifs . . . . . . . . . . . . . . . . . . . . . . . . 223

16.9.3 fastcox . . . . . . . . . . . . . . . . . . . . . . . . . . 223

16.9.4 HighDimOut . . . . . . . . . . . . . . . . . . . . . . . 224

16.10 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226

Bibliography 227

Index 241

Preface

Over the past decades, computational flexibility has extended the application
of the Bayesian approach as a handy method in research practice. The oncol­
ogy disease complexity and existing challenge to obtain the best therapeutic
effect has provided immense opportunity to develop the statistical methodol­
ogy. Perhaps, the acceptance of the Bayesian approach in oncology research
is relatively low. There is a vast scope to extend the analytical approach by
Bayesian in oncology research practice. This book intended to make a single
source of information on Bayesian statistical methodology for oncology re­
search to cover several dimensions like study design, sample size calculation,
time-to-event data analysis, and diagnostics implication. The idea is to overall
extending the Bayesian approach in oncology research practice.
This book is structured with four sections to extend Bayesian in oncology
research. Study design-related topics proposed to bring in Part I. The impor­
tance of statistics in the oncology domain first felt during the setup of the
required participant’s size for an oncology trial. Chapter 1 gives an introduc­
tion about R and OpenBUGS software. It is prepared to understand about
basic stuff of R and OpenBUGS.
Now Chapter 2 is explained with sample size determination. The Bayesian
counterpart to calculate the sample size determination presented. An inferior­
ity trial overrules the application of a placebo-controlled trial due to the eth­
ical constraints and availability of several molecules in the oncology domain.
Chapter 3 illustrated about conventional study design for oncology clinical tri­
als. The alternative approach of traditional design of study shown in Chapter
4. It is not always that high-dose chemotherapy will work as the best ther­
apeutic arm. Oncology is a complex disease. Sometimes low-dose frequently
used chemo regime works well to control the disease progression. The low-dose
chemotherapy is decided based on the optimum biological dose (OBD). The
Bayesian statistical methodology on low-dose chemotherapy selection proce­
dure demonstrated in Chapter 5.
Section II proposed to develop and present a different statistical approach
with real data application with time-to-event data analysis. The conventional
technique on survival analysis is a frequentist-based approach and method
with Kaplan-Meier and Cox proportional hazard models always performed.
However, there is minimal literature support about the Bayesian counterpart
on Cox proportional hazard modeling. Chapter 6 shows about Bayesian sur­
vival analysis. The competing risk, i.e., cancer patient, died due to other causes

xiii
xiv Preface

than cancer is a new issue in oncology data. The decision about the best ther­
apeutic regime in the presence of competing for risk is always challenging.
Chapter 7 is presented with Bayesian competing risk analysis to compare two
therapeutic schemes. The concept of frailty provides a convenient way of in­
troducing unobserved heterogeneity and associations into models for survival
data. Frailty is an unobserved random proportionality factor. It influences the
hazard function of an individual or related individual chapter 7 prepared on
Bayesian frailty data analysis. Similarly, Chapter 8 prepared about Bayesian
relative survival analysis. This work developed with an illustration.
In this chapter, we will show with a Bayesian analysis for right-censored
survival suitable for curable cancer site data. Bayesian inference with Markov
chain Monte Carlo (MCMC) methods will be established through model se­
lection technique and example with a real dataset. The Section III is about
statistical methodology in longitudinal and survival data analysis. Bayesian
inferences for longitudinal data analysis in oncology research is evitable. It is
attractive due to the ability to consider prior information for statistical growth
curve modeling. Chapter 10 focused on longitudinal data analysis. The mixed-
effect model illustrated with R. The Bayesian inference on missing longitudinal
data illustrated in Chapter 11. Similarly, the Bayesian approach is suitable to
handle Joint longitudinal and survival data. It helps to consider the depen­
dence between two types of responses obtained by two different methods. In
Chapter 12, the joint longitudinal and survival analysis presented. Different
prior assumptions and covariance matrix presented with posterior inference.
The work performed with OpenBUGS software. Chapter 13 has presented
with different covariance structures observed by longitudinal measurements.
The Bayesian counterpart illustrated different covariance structures.
The final section-IV is detailed on diagnosis test statistics with available
statistical methodology. The author illustrated the Bayesian method to pro­
vide inference on different inter and intra-rater agreement analyzes. Chapter
14 is present about mixed effect modeling. It is dedicated to select the spe­
cific dose in oncology. Chapter 15 is dedicated to concordance analysis by
the Bayesian approach. The CCC is a useful tool in agreement analysis. The
importance of high-dimensional data in oncology research will be explored.
The Bayesian approach in high-dimensional data is in Chapter 16. The R and
OpenBUGS codes illustrated.
Author

Atanu Bhattacharjee is an Assistant Professor at the

Section of Biostatistics, Centre for Cancer Epidemiol­
ogy, Tata Memorial Centre, India. He previously taught
Biostatistics at the Malabar Cancer Centre, Kerala, In­
dia. He completed his PhD at Gauhati University, As­
sam, on Bayesian Statistical Inference. He is an elected
member of the International Biometric Society (Indian
Region). He served as Associate Editor of BMC Medical
Research Methodology. He has published over 200 research articles in various
peer-reviewed journals.

xv
Part I

Bayesian in Clinical

Research

1
Chapter 1

Introduction to R and OpenBUGS

Abstract
This chapter starts with the history of R software development. Steps to
install R are discussed. It provides about the installation of different packages
in R.Graphical representation by R will help many users to get interested in a
graphical illustration.Steps to install R studio. Description of OpenBUGS soft­
ware and illustration to perform Markov Chain Monte Carlo in OpenBUGS.
Introduction about Bayesian using Gibbs sampling accessible. Different types
of Markov chain Monte Carlo (MCMC) approaches are outlined. OpenBUGS
model explanation, trace plot generation, data loading and posterior estimates
generation are described. The real-world examples are used to illustrate the
methods presented. It should provide some familiarity about the R process
and essentially environment on what actually works. R software works as the
integration of data calculation, graphical procedure, and data manipulation.
Command to work with R is presented. This chapter did not explain about
statistics in R. The aim is to provide R and OpenBUGS environment. It is
preferred to show the environment to work with the classical and modern sta­
tistical test to implement. Illustrations are presented by the workstations with
windows system. It will guide the users about the available facility.

1.1 Introduction to R
R is a programming language developed by Ross Ihaka and Robert
Gentleman in 1993. It is a free software for graphics and statistical computing.
The source code of R is prepared primarily in Fortran, R, and C. Although R
works are command-driven software, there are other graphical user interfaces
like RStudio. R is named partly after the first names of the first two R authors
and partly as a play on the name of S [1].
There are several dedicated R packages available in R for specific statistical
methods. However, linear modeling, nonlinear modeling, time-series analysis,
clustering, and other analysis can be performed in R itself. No special packages

3
4 Bayesian Approaches in Oncology Using R and OpenBUGS

are required to install. The R community is noted for its active contributions
in terms of packages. R users typically access the command-line interpreted.
For example, if the user types 4+4 at the command prompt and press enter,
the computer replies with 8, as shown below.

>4+4
>8

RStudio is an integrated development environment (IDE) for R. It is ba­

sically a nice front-end for R, giving you a console, a scripting window, a
graphics window, and an R workspace, among other options.

1.2 How to Install R

1. Download R from: http://cran.us.r-project.org/ (click on “Download R
for Windows” > “base”> “Download R 3.6.2 for Windows”).
2. Install R. Leave all default settings in the installation options.

1.3 Packages
The datasets and functions available in R are stored in packages. A spe­
cific package is needed to install before its contents available. The packages
available can be obtained as

> library()

To see which packages are currently loaded, use

> search()

The standard (or base) packages are defined as part of the R source code.
This contains the essential functions that allow R to work. Datasets and stan­
dard statistical and graphical functions are available. Data can be generated
from the normal distribution by ‘rnorm’ as
#Generate Data Distribution

x <­ rnorm(50)
y <­ rnorm(x)
plot(x, y)
Introduction to R and OpenBUGS 5

FIGURE 1.1: How to download R software.

#Generate Line Plot with R

chemotreated <­ c(105,112,110,90,109,115)

radiothrapytreated <­ c(120,127,132,125,127,118)
grange <­ range(0,chemotreated,radiothrapytreated)
plot(chemotreated, type=“o”, col=“blue”, ylim=grange,
axes=FALSE, ann=FALSE)
axis(1, at=1:5, lab=c(“Mon”,“Tue”,“Wed”,“Thu”,“Fri”))
axis(2, las=1, at=4*0:grange[2])
box()
lines(radiothrapytreated, type=“o”, pch=22, lty=2, col=“red”)
title(main=“Chemo and Radiation Given Patients”, col.main=“red”,
font.main=4)
title(xlab=“Days”, col.lab=rgb(0,0.5,0))
title(ylab=“Total Treated Patients”, col.lab=rgb(0,0.5,0))
legend(1,50, c(“Chemo”,“Radiation”), cex=0.8,
col=c(“blue”,“red”), pch=21:22, lty=1:2);
6 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 1.2: Simple plot with R.

Now the Plot function is useful. A graphics window will appear automati­
cally. There are thousands of functions available in R. This function can be ob­
tained from the Comprehensive R Archive Network (CRAN): https://cran.r­
project.org.

1.4 How to Install RStudio

1. Download RStudio from: http://rstudio.org/download/desktop and in­
stall it. Leave all default settings in the installation options.
2. Open RStudio.
3. Go to the “Packages” tab and click on “Install Packages”. The first time it
will prompt to choose a CRAN mirror.
4. R will download all necessary files from the server to select here. Choose
the location closest to your geographical location.
5. Install packages in Windows.
Introduction to R and OpenBUGS 7

FIGURE 1.3: Line plot with R.

1.5 OpenBUGS
BUGS is an open source statistical software package used for Bayesian in­
ference using Gibbs sampling. The statistical model is specified by merely stat­
ing the relationship between covariates and outcome variables. The MCMC
technique is used for analyzing the specified model. It can be downloaded
from: http://www.openbugs.net/w/Downloads. A setup program for Open-
BUGS on Windows computers is available. The latest version is available as
version 3.2.2. The manual is also available over there.
It is better to have knowledge of MCMC methods to work with Open-
BUGS. There are three different families of MCMC algorithms like Gibbs,
Metropolis-Hastings, and slice sampling to work with OpenBUGS. Primarily,
Gibbs sampling algorithm works to sample from the conditional distribution
of each node given for all the others in the graph. It works as a particular case
of Metropolis-Hastings algorithm.
8 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 1.4: Selection of free RStudio version.

The OpenBUGS is used to define the concise expression of the model, by

the ‘twiddles’ symbol ∼. This symbol is used to define the stochastic (proba­
bilistic) relationships. Thereafter, the left arrow ‘ < ’ followed by ‘-’ is used to
define the relationship between outcome and covariates. The model is defined
as:
Introduction to R and OpenBUGS 9

FIGURE 1.5: How to install OpenBUGS.

#Define OpenBUGS Model

model
{
for (i in 1 : Dogs) {
xa[i, 1] <­ 0; xs[i, 1] <­ 0 p[i, 1] <­ 0
for (j in 2 : Trials) {
xa[i, j] <­ sum(Y[i, 1 : j - 1])
xs[i, j] <­ j - 1 - xa[i, j]
log(p[i, j]) <­ alpha × xa[i, j] + beta × xs[i, j]
y[i, j] <­ 1 - Y[i, j]
y[i, j] ∼ dbern(p[i, j])
}
}
alpha ∼ dnorm(0, 0.00001)I(, -0.00001)
beta ∼ dnorm(0, 0.00001)I(, -0.00001)
A <­ exp(alpha)
B <­ exp(beta)
}
10 Bayesian Approaches in Oncology Using R and OpenBUGS

#Define Initial Data

list(alpha = -1, beta = -1)

Now at step 1, we select the model.

Now at step 2, we will check the model description and look it runs appropri­
ately or not.
At the next step, we should load the data.
Now at step 4, we need to select the number of chains (i.e., sets of samples to
simulate). The default is 1, but we will use two chains for this tutorial.
#Define Data

list(Dogs = 30, Trials = 25, Y = structure(.Data = c(0, 0, 1, 0, 1, 0,

1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0,
0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0,
1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 1, 1, 0, 1, 0, 1, 0, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0,
0, 0, 0, 1, 0, 1, 0, 1, 1, 0, 1, 0, 0, 0, 1, 1, 1, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 1,
0, 0, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0,
0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 0, 1, 0, 0, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 0, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 0, 1, 0, 1, 0, 0, 0, 1, 0, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0,
0, 0, 0, 1, 0, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0,
0, 0, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0,
1, 0, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 0, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 1, 0, 0, 0, 1, 1,
0, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 0, 0, 0, 1, 0, 1, 0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0,
0, 0, 0, 1, 1, 0, 0, 1, 1, 1, 0, 1, 0, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0,
0, 1, 1, 1, 1, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1),.Dim = c(30, 25)))
Introduction to R and OpenBUGS 11

#Results

mean sd MC error val2.5pc median val97.5pc

alpha 0.24 0.027 0.00 − 0.29 − 0.24
beta − 0.07 0.018 0.00 − 0.10 − 0.07

FIGURE 1.6: Trace plot example in OpenBUGS.

Chapter 2

Sample Size Determination

Abstract
The sample size for a clinical trial jointly works with study design. A
different issue like a number of sites, number of treatments also considered for
sample size determination. The appropriate response plays an essential role in
sample size calculation. Computing large sample size always wastes of time,
money and unnecessarily put patients into a risk. Similarly, sample size too
small always makes a less powerful statement. Thus, it is essential to compute
sample size wisely to decide minimal clinically meaningful effect size with
pre-specified type I error and power. It is not possible to plan a clinical trial
unless we know the required sample size. However, there is a different strategy
to perform any oncology trial. The study design and study hypothesis play a
combined role to determine the sample size. There are some great sample size
and power calculation software packages available. We presented this chapter
with some R sample size calculation packages. However, there are undoubtedly
other packages that you might consider. The R packages are illustrated toward
implementing Bayesian for sample size calculation. Theoretical explanation
about Bayesian Sample size determination is discussed. Different R packages
are used to solve the problem and all of the problems that encountered do a
better job the other on certain types of scenario.

2.1 Introduction
How many subjects do we need? How long will the study take to complete?
In survival analysis, we need to specify information regarding the censoring
mechanism and the particular survival distributions in the null and alternative
hypotheses.
First, one needs either to specify what parametric survival model to use, or
that the test will be semi-parametric, e.g., the log-rank test. It requires to
determining the number of deaths (or events) required to meet the power and
other design specifications.

13
14 Bayesian Approaches in Oncology Using R and OpenBUGS

Second, one must also provide an estimate of the number of patients that need
to entered into the trial to produce the required number of deaths. We shall
assume that the patients enter a trial over a certain accrual period of length
a, and then followed for an additional period f known as the follow-up time.
Patients still alive at the end of follow-up are censored. Sample size calcula­
tions are an essential part of study design Consider sample size requirements
early.
A well-designed trial is large enough to detect clinically important differences
between groups with high probability.
To perform sample size calculations, we need well-defined study endpoints,
hypotheses, and statistical tests. Study hypotheses decided on a clearly de­
fined endpoint and period of study.
In most RCTs, known as superiority trials, the study hypothesis presented as
a null hypothesis of no difference in the distribution of the primary endpoint
between study groups. We have an alternative hypothesis in mind, for exam­
ple: HA . The frequency of events at 30 days will differ in the two treatment
groups.
In superiority trials, we test the null hypothesis against a two-sided alterna­
tive. When we test the null hypothesis, there are two possible states of nature
and two decisions: We will perform a test that has a small probability of a
Type 1 error, usually 0.05. The power of the study is the probability that
we will reject the null hypothesis when the alternative hypothesis is true. We
want this probability to be large, typically at least 0.8.

2.2 The Sample Size Formula

2p̄ ∗ (1 − p̄)(Zα/2 + Zβ )2
n= (2.1)
Δ2
Zα/2 and Zβ are the critical values of the normal distribution, p̄ is the average
of the event rates under the alternative hypothesis, and Δ is the true difference
under HA .
For example, α = 0.05, β = 0.20, 2n = 1, 903 or n = 806.
Sample Size Determination 15

2.3 Measured Outcomes

If we are testing the equality of two treatments (T and C), and the endpoint
is a measurement, the null hypothesis is typically expressed in terms of the
difference in means

−True Means :µT = µC (2.2)

−Difference :Δ = µT − µC (2.3)
−H0 : Δ = 0, Hα : Δ > 0 orΔ < 0 (2.4)

2.3.1 Time to event endpoints

In many clinical trials, the primary endpoint is the time to an event, e.g.,
death or disease progression.
In that circumstance, the analyst will employ methods specific to the analysis
of “survival data”.
We discuss those methods briefly here, but in greater detail later in the course.
Estimate the survival distribution in each treatment group and use nonpara­
metric methods to compare them.
The most common test is known as the log-rank test.
Modern methods accommodate variable entry times and periods of follow-up.
The standard nonparametric test for comparing two distributions is the log-
rank test.
Interestingly, sample size formulas for the log-rank test are closely related to
those that apply when the times-to-event follow an exponential distribution.
With exponential survival distributions, the null hypothesis is H0 : λT = λC

A simple formula for the sample size, assuming all subjects are followed to
the event, is
n = 2(Zα/2 + Zβ )2 /[In(λC /λT )]2 (2.5)
where the values of λC and λT are given by HA . If the subjects are recruited

over some time, and the study ends when some items have not had the event,

sample size calculations are more complex

Lachin formula [2]

n = (Zα/2 + Zβ )[φ(λT ) + φ(λC )]/(λT − λC )2 (2.6)

The sample size depends on the recruitment and follow-up schedules.
16 Bayesian Approaches in Oncology Using R and OpenBUGS

2.3.2 Sample size with event-driven trials

For the logrank test, the power of the study depends on the number of
events observed. This has led to the concept of the event-driven trial.
Schoenfeld showed that, if the logrank test will be used to compare two time-
to-event distributions, the number of events required to achieve power of 1 − β
is
n = 4(zα/2 + zβ )2 /[In(λC /λT )]2 (2.7)

#Sample Size Calculation with R

library(“gsDesign”)
ss <­ nSurvival(lambda1=.2, lambda2=.1, eta = .1,
Ts = 2, Tr = .5,sided=1, alpha=.025)
ss

# R Output

Fixed design, two-arm trial with time-to-event

outcome (Lachin and Foulkes, 1986).
Study duration (fixed): Ts=2
Accrual duration (fixed): Tr=0.5
Uniform accrual: entry="unif"
Control median: log(2)/lambda1=3.5
Experimental median: log(2)/lambda2=6.9
Censoring median: log(2)/eta=6.9
Control failure rate: lambda1=0.2
Experimental failure rate: lambda2=0.1
Censoring rate: eta=0.1
Power: 100*(1-beta)=90%
Type I error (1-sided): 100*alpha=2.5%
Equal randomization: ratio=1
Sample size based on hazard ratio=0.5 (type="rr")
Sample size (computed): n=430
Events required (computed): nEvents=91

# Sample Size Calculation with R

library(“gsDesign”)
x<-gsDesign(k = 5, test.type = 2, n.fix=ss$nEvents,
nFixSurv=ss$n,delta1=log(ss$lambda2/ss$lambda1))

x
Sample Size Determination 17

#R Output

Group sequential design sample size for time-to-event outcome

with sample size 440. The analysis plan below shows events
at each analysis.
Symmetric two-sided group sequential design with
90 % power and 2.5 % Type I Error.
Spending computations assume trial stops
if a bound is crossed.
Analysis N Z Nominal p Spend
1 19 3.25 0.0006 0.0006
2 37 2.99 0.0014 0.0013
3 56 2.69 0.0036 0.0028
4 74 2.37 0.0088 0.0063
5 93 2.03 0.0214 0.0140
Total 0.0250
++ alpha spending:
Hwang-Shih-DeCani spending function with gamma = -4.
Boundary crossing probabilities and expected sample size
assume any cross stops the trial
Upper boundary (power or Type I Error)
Analysis
Theta 1 2 3 4 5 Total E{N}
0.0000 0.0006 0.0013 0.0028 0.0063 0.0140 0.025 91.5
0.3415 0.0370 0.1512 0.2647 0.2699 0.1771 0.900 66.4
Lower boundary (futility or Type II Error)
Analysis
Theta 1 2 3 4 5 Total
0.0000 6e-04 0.0013 0.0028 0.0063 0.014 0.025
0.3415 0e+00 0.0000 0.0000 0.0000 0.000 0.000

#Sample Size Calculation for the Comparison of Survival Curves Com­

parison

library(“powerSurvEpi”)
ssizeCT.default(power = 0.8,k = 1,pE=0.3707,
pC = 0.4890,RR = 0.7,alpha = 0.05)
18 Bayesian Approaches in Oncology Using R and OpenBUGS

2.4 Bayesian Sample Size Determination

It is not possible to plan a clinical trial unless we know the required sample
size. However, there is a different strategy to perform any oncology trial. The
study design and study hypothesis play a combined role to determine the
sample size. Conventionally approach is to consideration of comparison to the
existing arm. Now the interpretation can be separated as a null and alternative
hypothesis. Under the null hypothesis, it is assumed that the experimental and
exsisting arms similar. The outcome is survival. Now the survival outcome
can be formulated as an event/event-free at a specific time point. Now it can
be death or relapse. Determination of event size is required to perform any
oncology clinical trial. It may be an event number of sample size. Suppose we
have two types of treatment for head and neck cancer defined as Arm A and
Arm B. The prior information about death probability those are treated with
Arm-A is 40% after the end of 3 -years. Similarly, it is 50% after three years for
those who are treated with Arm-B. Now the null hypothesis is formulated as
H0 : PA = PB . The alternative hypothesis can be formulated as H1 : PA = � PB .
The level of significance is defined as α, and type-II error is defined as β.
The level of importance can be defined as a type-I error. The task is about
balancing the type-I and type-II error to get the optimum sample size. Suppose
the sample size for trial 1 and 2 are calculated as n1 and n2 . Both trials are
designed for the same treatment effect. Now the success of those trials is
defined as π1 and π1 respectively. The parameter π1 and π2 are measured
as x1 /n1 and x2 /n2 respectively. The number of patients alive after the end
of 3 years for both the trial are x1 and x2 respectively. Now the posterior
distribution is presented as
2
�
f (π1 , π2 |x1 , x2 , n1 , n2 ) = kf (π1 , π2 ) πixi (1 − πi )ni −xi (2.8)
i=1

The normalizing constant is prepared as k. The posterior dsitribution is de­

pendent on the prior distribution. Now π1 and π2 can calculated from beta
distributions with paramters (c1 , d1 ) and (c2 , d2 ), respectively. The equation
can be formulated as
2
�
f (π1 , π2 |x1 , x2 , n1 , n2 ) = k πixi +ci −1 (1 − πi )ni −xi +di −1 (2.9)
i=1

Now

k = [B(x1 + c1 , n1 − x1 + d1 )B(x2 + c2 , n2 − x2 + d2 )]−1 (2.10)

and B(., .) represents the beta function. The term (x1 , x2 ) can be formulated
Sample Size Determination 19

as
2
� (nx i )B(xi + ci , ni − xi + di )
i
p(x1 , x2 ) = (2.11)
i=1
B(ci , di )
The simplified form of joint posterior distribution can be formulated as

f (π1 , θ|x1 , x2 , n1 , n2 ) = kπ1x1 +x1 −1 (1 − π1 )n1 −x1 +d1 −1

(2.12)
(π1 − θ)x2 +c2 −1 (1 − π1 + θ)n2 −x2 +d2 −1

which is non-zero over the region in the plane bounded by the lines π1 =
0, π1 = 1, π1 = θ and π1 = θ + 1. The marginal posterior distribution of θ can
be formulated as
� min(θ+1,1)
f (θ|x1 , x2 , n1 , n2 ) = f (π1 , θ|x1 , x2 , n1 , n2 )dπ1 (2.13)
max(0,θ)

2.5 Study Objective and Sample Size Determination

The sample size for a clinical trial jointly works with study design. A
different issue like a number of sites, number of treatment also considered for
sample size determination. The appropriate response plays an essential role in
sample size calculation. Computing large sample size always wastes of time,
money and unnecessarily put patients into a risk. Similarly, sample size too
small always makes a less powerful statement Thus, it is essential to compute
sample size wisely to decide minimal clinically meaningful effect size with
prespecified type I error and power. Hypothesis testing and statistical test are
two crucial part in sample size calculation. Initially, the null and alternative
hypothesis need to be specified. Therefore proper statistical analysis needs
to be determined.Sample size can be specified with type-I error and power
of 1 − β, effect size, and other design related parameter. Suppose that the
null hypothesis with the probability of survival after 3/5 years of treatment
initiation is specified as p1 and p2 for arm A and arm B. The null hypothesis
is defined as
H0 : p1 ≤ p2 vs H1 : p1 > p2 (2.14)
The study is powered with alternative hypothesis as p = p1 (< p2 ). Now, given
the type I error α and power of 1 − β under the alternative H1 : p = p1 .
The effect size is ω = p2 − p1 > 0. The sample size is calculated based on
the following assumptions. The statistics U under the H0 is asymptotically
normal distributed with mean p1 and variance σ02 /n.
Let σ̂ 2 is a consistent estimator of σ02 . The null hypothesis is
√
n(U − µ0 )
Z= → N (0, 1) (2.15)
σ̂
20 Bayesian Approaches in Oncology Using R and OpenBUGS

Here Z converges to normal distribution with mean zero and variance 1.

� x
1 u2
Φ(x) = √ exp− 2 du (2.16)
−∞ 2π
Under the one-sided test, we reject the H0 if Z > z1−α Now, z1−α = 1 −
α percentile of the standard normal distribution and z1−α = Φ−1 (1 − α).
Under the alternative p1 = p2 , σ 2 converges � to σ 2 . It is assumed that Z is
asymptotically normal distributed with mean nω/σ and unit variance. The
power 1 − β is defined as an alternative p0 = p2 . It present with the equations:
√
n(U − µ0 )
1 − β = P (Z > z1−α |H1 ) = P ( > z1−α |H1 ) (2.17)
σ̂

√ √
n(U − µ0 ) nω
1 − β = P (Z > z1−α |H1 ) = P ( > z1−α − |H1 ) (2.18)
σ̂ σ
√
nω
1 − β = P (Z > z1−α |H1 ) � Φ(−z1−α + ) (2.19)
σ
√
(z1−α + z1−β ) = nω
Finally, the sample size is obtained with n by

(z1−α + z1−β )2 σ 2
n= (2.20)
ω2

#Average Coverage Criterion (ACC)

The θ pararmeter is used to make different two binomical parameters,

it can be specified by minimizing n
n
� n
�
P r{θ ∈ (a(x1 , x2 ), a(x1 , x2 ) + l)}p(x1 , x2 ) ≥ 1 − α (2.21)
x1 =0 x2 =0

where
n
� n
� � a(x1 ,x2 +l)
P r{θ ∈ (a(x1 , x2 ), a(x1 , x2 )+l)} = f (θ|x1 , x2 , n)dθ
x1 =0 x2 =0 a(x1 ,x2 )
(2.22)
The lower limit of the Highest Posterior Density (HPD) is defined
through x1 , x2 . The term probability 1 − α varies with x with HPD
interval l. It helps to define the sample with minimum size n by
� � a(x,n)+1
{ f (θ|x)dθ}f (x)dx ≥ 1 − α (2.23)
X a(x,n)
Sample Size Determination 21

#Average Length Criterion (ALC)

The minimum sample size n can be obtained by

n n
� � �
l (x1 , x2 )p(x1 , x2 ) ≤ l (2.24)
x1 =0 x2 =0

The HPD is obtained from (x1 , x2 ) by solving

�
� a(x1 ,x2 )+l (x1 ,x2 )
f (θ|x1 , x2 , n) = 1 − α (2.25)
a(x1 ,x2 )
�
where a(x1 , x2 ) and a(x1 , x2 ) + l (x1 , x2 ) are used to cover the HPD
interval. In this case the coverage probability is fixed and allows to
vary the HPD interval. Length is defined as
�
� a(x,n)+l (x,n)
f (θ|x)dθ = 1 − α (2.26)
a(x,n)

while the sample size is defined as n with minimum integer by

�
�
l (x, n)f (x)dx ≤ l (2.27)
X

It is obvious to obtain different sample size through ACC and ALC.

#Worst Outcome Criterion (WOC)

If we fixed the interval with length l and defined the coverage prob­
ability with 1 − α then the minimum sample size n can be defined
as ∗ ∗
� a(x1 ,x2 )+l
f (θ|x∗1 , x∗2 , n)dθ ≥ (1 − α) (2.28)
a(x∗ ∗
1 ,x2 )

The optimum sample size not possible to obtain by ACC and ALC.
The conservative approach to maximize the length of l and reduce the
minimum coverage probability of (1 − α) in connection to the data.
The optimum sample size is obtained by
� a(x,n)+l
inf { f (θ|x)dθ} ≥ 1 − α (2.29)
x∈X a(x,n)

It helps to avoid unnecessarily selecting large sample size for clinical

trial.
22 Bayesian Approaches in Oncology Using R and OpenBUGS

2.6 Bayesian Survival Analysis Sample Size Calculation

with R

>library(“SampleSizeProportions”)

library(“SampleSizeProportions”)
propdiff.acc(len=0.05, c1=3, d1=2, c2=2, d2=3)
[1] 2483 2483

In the above equation the length is defined as the posterior credible interval
for the two proportions difference. Now c1 , c2 are the first priors parameter
with Beta density for the first and second population. Similarly d1 , d2 are the
second priors parameter with Beta density for the first and second population.
Level of significant is defined as 0.95. Now the calculated sample size for both
the groups are (n1 , n2 ).

propdiff.alc(len=0.05, c1=3, d1=2, c2=3, d2=2)

[1] 2430 2430

propdiff.mblacc(len=0.05, c1=3, d1=2, c2=2, d2=3)

[1] 2478 2478

propdiff.mblmodwoc(len=0.05, c1=3, d1=2, c2=2, d2=3,

worst.level=0.95)
[1] 3070 3070

2.7 Illustration
Let there are two experimental arms named as Arm A and Arm B respec­
tively. Arm A is exsisting arm and we want to compare the therapeutic effect
of Arm B with reference to Arm A. Now the null hypothesis is required to for­
mulate. The null hypothesis should be linked with sucess of Arm A is similar
with Arm B. The success of Arm A is defined as probability of survival of the
patients at the end of 3 years is 38%. Similarly, for Arm B it is assumed as
61%. The prior parameters assumption for beta density on Arm A is defined as
Sample Size Determination 23

3 and 2. Similarly, for Arm B is assumed as 2 and 3. The parameteric assump­

tion with 3 and 2 provides median value as 0.61. Similarly, for arm B as 0.38.
There after we fixed the level of significance with 95%. The calculated sample
size obtained with ACC for both the arms are 2483. Similarly, we can run the
same parameteric assumption with function propdiff.alc,propdiff.mblacc and
propdiff.mblmodwoc respectively.

2.8 Posterior Error Approach

The minimum sample size can be obtained by manipulation type-I and
type-II error through the frequentist approach. Similarly, the posterior error
rate stands to obtain a desirable sample size. The minimum sample size can
be obtained so that the posterior error rate can be controlled.It is important
to explore the posterior error rate and the relationship with the traditional
type I and type II error rate. The posterior error rate is considered as a check
and balance of Bayesian sample size calculation.

2.9 Different Sample Size Determination Packages in R

2.9.1 BAEssd
The Bayesian average error approach applicable to estimate the sample
size. The function available in BAEssd package named as “norm1KV.2sided”
is useful to
#BAEssd Package

library(“BAEssd”)
f1 <­ norm1KV.2sided(sigma=5,theta0=0,prob=0.5,mu=2,tau=.1)
ss1 <­ ssd.norm1KV.2sided(alpha=0.25,w=0.5,minn=2,maxn=200)
ss1

#Outcome on BAEssd Package

Sample Size: 34
Total Average Error: 0.2451311
Acceptable sample size determined!
plot(ss1)
plot(ss1,y=“AE1”,alpha.line=FALSE) abline(h=0.05,lty=2)
24 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 2.1: Type-I error plot on sample size.

2.9.2 BayesianPower
This package is useful to compute the sample size and power for comparing
inequality constrained hypotheses. The example of calculating the sample size
by “bayes sampsize” function is in the following text.
#Calculating Sample Size by BayesianPower Package

h1 <­ matrix(c(1,-1), nrow= 1, byrow= TRUE)

h2 <­ ‘c’
m1 <­ c(.3, 0)
m2 <­ c(0, .2)
bayes sampsize(h1, h2, m1, m2, sd1 = 1, sd2 = 1, scale = 1000,
type = “de” cutoff = .125, nsamp =100, datasets =100,
minss =30, maxss =60)
Sample Size Determination 25

#Inputs Required for BayesianPower Package

n A number. The sample size.

h1 A constraint matrix defining H1 .
h2 A constraint matrix defining H2 .
m1 A vector of expected population means under H1 .
m2 A vector of expected populations means under H2 .m2 must be
of same length as m1.
sd1 A vector of standard deviations under H1 . Must be a single
number (equal).
standard deviation under all populations), or a vector of the same
length as m1.
sd2 A vector of standard deviations under H2 . Must be a single
number (equal standard deviation under all populations), or a vector
of the same length as m2.
scale A number specifying the prior scale.
bound1 A number. The boundary above which BF12 favors H1 .
bound2 A number. The boundary below which BF12 favors H2 .
datasets A number. The number of datasets to compute the error
probabilities.
nsamp A number. The number of prior or posterior samples to
determine the fit and complexity.

#Output Obtained by Bayes Sampsize

Sample size Type 1 error Type 2 error

46.00000000 0.08000000 0.17000000
Decision error Indecision error Median BF12 under H1
0.12500000 0.00000000 0.09078999

2.9.3 NPHMC
This Sample Size Calculating package is used for the Proportional Hazards
Mixture having cure or without cure Model
#Sample Size Calculation for PH Mixture Cure Model and Standard
PH Model

library(“NPHMC”)
NPHMC(power=0.90,alpha=0.05,accrualtime=3,
followuptime=4,p=0.5,accrualdist=“uniform”,
hazardratio=2/2.5,oddsratio=2.25,pi0=0.1,
survdist=“exp”,k=1,lambda0=0.5)
26 Bayesian Approaches in Oncology Using R and OpenBUGS

#Output on NPHMC

At alpha = 0.05 and power = 0.9 :

PH Mixture Cure Model: n = 429
Standard PH Model: n = 908

2.9.4 PowerTOST
This package is used for calculating the Power and Sample Size for (Bio)
Equivalence Studies. It can be obtained by two and one-sided t-tests.

#Power and Sample Size Calculation for (Bio)Equivalence Studies

library(“PowerTOST”)
power.TOST(CV = 0.25, n = 24)

#Power Calculation for the 2 × 2 Cross-Over Design with 24 Subjects

and CV 25%
0.7391155

2.9.5 SampleSize4ClinicalTrials
This pacakge is used to calculating sample size for Means and Proportion
in Phase III Clinical trials. Different ressearch design can be performed by this
package. Research design like: (1) Testing for equality, (2) Superiority trial,
(3) Non-inferiority trial, and (4) Equivalence trial. Now performed sample size
calcuation by comparison of Means and Proportions separately.
#Sample Size Calculation by SampleSize4ClinicalTrials Package

library(“SampleSize4ClinicalTrials”)
Calculation of Means
ssc meancomp(design = 3L, ratio = 1, alpha = 0.05, power = 0.8, sd
= 0.1, theta = 0, delta = -0.05)
Treatment Control
50 50

Calculation of Proportion
ssc propcomp(design = 4L, ratio = 1, alpha = 0.05, power = 0.8, p1
= 0.75, p2 = 0.80, delta = 0.2)
Treatment Control
133 133
Sample Size Determination 27

#Inputs Required for SampleSize4ClinicalTrials Package

design=The design of the clinical trials=1L.

Testing for equality=2L.
Superiority trial=3L.
Non-inferiority trial=4L.
Equivalence trial.
ratio=The ratio between the number of subjects in the treatment arm
and that in the control arm.
alpha=Type I error rate.
power=Statistical power of the test (1-type II error rate).
p1=The true mean response rate of the treatment arm.
p2=The true mean response rate of the control arm.
delta=The prespecified superiority, non-inferiority or equivalence
margin sample.

2.9.6 SSRMST
The sample size calculation by Restricted Mean Survival Time is possible
to compute. It helps to calculate the sample size based on the difference in
Restricted Mean Survival Time.
#SSRMST I

library(“SSRMST”)
ac rate=The Accrual rate is the number of patients
per unit time. ac period=The Accrual period as the time point at
last accrual.
ac number=The Accrual number of accrual patients.
tot time=The Total study time as the time point
at last follow-up.
tau Truncation is the time point to calculate RMSTs.
shape0, shape1 Shape parameters for the Weibull
distribution in
both the control (arm0) and the
treatment (arm1).
ac rate=15
ac period=35
tot time = 510
tau = 500
scale0 = 7000
scale1 = 7000
margin = 10
a=ssrmst(ac rate=ac rate,ac period=
ac period, tot time=tot time,
tau=tau, scale0=scale0, scale1=scale1,
margin=margin, ntest=20)
print(a)
28 Bayesian Approaches in Oncology Using R and OpenBUGS

#Result of SSRMST I

Non-inferiority test

Total arm0 arm1

Sample size 524 262 262
Expected number of events 35 17 18
Power (separate)
0.2
Power (pooled)
0.2

2.9.7 powerSurvEpi

#Power Calculation for the Survival Analysis

library(“powerSurvEpi”)
ssizeEpi.default(power = 0.80, theta = 2, p = 0.39, psi = 0.505,
rho2 = 0.1322 , alpha = 0.05)
139

#Inputs Required for PowerSurvEpi Package

library(“powerSurvEpi”)
power=postulated power.
theta=postulated hazard ratio.
p=proportion of subjects taking value one for the covariate of interest.
psi=proportion of subjects died of the disease of interest.
rho2=square of the correlation between the covariate of interest and
the other covariate.
alpha=type I error rate.

2.9.8 SampleSizeMeans
This package is useful to obtain the normal means. Several functions are
available to calculate the sample size by Bayesian approach. For example,
“mu.acc” function is used to obtain the given coverage probability on average
for a posterior credible interval of fixed length for a normal mean.
Sample Size Determination 29

#SampleSizeMeans I

library(“SampleSizeMeans”).
mu.acc(len, alpha, beta, n0, level=0.95).
len is the desired fixed length of the posterior
credible interval for the mean.
alpha is the First parameter of the Gamma prior density
for the precision (reciprocal of the variance).
beta is the second parameter of the Gamma prior density
for the precision (reciprocal of the variance).
n0 is the prior sample size equivalent for the mean.
level is the desired average coverage probability of the
posterior credible interval (e.g.,0.95).
#Calculated Sample Size#
mu.acc(len=0.2, alpha=2, beta=2, n0=50)
[1] 721

Similarly, sample size can be determined by the single normal mean using
the Average Length Criterion. The function is “mu.alc”. If the variance is
known, then sample size is calculated by mixed of Bayesian and likelihood
approach with “mu.mbl.varknown” function.
#SampleSizeMeans II

library(“SampleSizeMeans”)
mu.mbl.varknown(len, lambda, level = 0.95)
len is the desired total length of the posterior.
credible interval for the mean.
lambda is the known precision(reciprocal of variance).
level is the desired coverage probability of the
posterior credible interval (e.g., 0.95).
#Calculated Sample Size#
mu.mbl.varknown(len=0.2, lambda=1/4)
[1] 1537

Other R packages like samplesizelogisticcasecontrol,coprimary,CRTSize,

EurosarcBayes,and longpower are also usefulfor sample size calculation.
These are dedicated packages for sample size calculation for the clinical trial.
Chapter 3

Study Design-I

Abstract
It is preferred to define suitable statistical models for dose-response mod­
eling. It is used as consistent by the relevant biological process with a specific
case. The natural process, cell growth and biomarker changes or covariates
with response variable is required under consideration. Mostly, dose-response
models are curve fitting practice. Curve fitting can be performed by conven­
tional or Bayesian approaches. This chapter is dedicated toward the applica­
tion of the Bayesian technique in early phase oncology trial. Oncology trial’s
drugs are considered for humans with a small number of participants. The
primary objective of any initial phase clinical trial is to explore the safety of a
drug. Therefore, finalize the effective dose of medicine by looking acceptabil­
ity on pharmacokinetics, and pharmacodynamics principle. Phase I studies
are dedicated to safe-effective treatment. Oncology drugs come with a toxic
agent. Phase I trial patients are cancer patients for those standard therapies
failed to respond. Phase I dose-escalation method is explained with Open-
BUGS. Further, the continual reassessment method is detailed by Bayesian
and illustrated with an example. The model-based designs for determining
the maximum tolerated dose is explained. Packages available in R for dose
selection modeling is detailed.

3.1 Introduction
The Bayesian technique is prominent methodology in data science. Data
science is inevitable in clinical research. So Bayesian merged tremendously
in the clinical trials well. Study design, sample size calculation, and statis­
tical analysis have emerged with the Bayesian technique. The approach to
analyzing data in Bayesian is different from the classical approach. Recent
computation advancement helped to incorporate Bayesian in clinical trials. It
also helped to upgrade data analysis practice. Oncology research is involved
with time-to-event outcomes (i.e., death). Data analysis with survival analysis

31
32 Bayesian Approaches in Oncology Using R and OpenBUGS

is not straight forward. In the presence of missing or incomplete information, it

becomes more complicated. Bayesian helps to overcome different challenges.
Recently, Bayesian approach is considered in several clinical trial study de­
sign. The adoption of probability is different in the Bayesian technique. The
evidence-based probability plays a role in Bayesian. It helps to upgrade the
proportion of uncertainty in data analysis logic. The prior information is re­
quired to incorporate. The literature review is required to upgrade the prior
information. Aim of this chapter is to a show about different study design
with Bayesian for oncology clinical trials.

3.2 Bayesian in Early Phase Oncology Trial

This chapter is dedicated toward the application of the Bayesian technique
in early phase oncology trial. Oncology trial’s drugs are considered for humans
with a small number of participants. The primary objective of any early phase
clinical trial is to explore the safety of a drug. Therefore, finalize the effective
dose of a drug by looking acceptability on pharmacokinetics, and pharmaco­
dynamics principle. Phase I studies are dedicated to the safe-effective dose.
Oncology drugs come with a toxic agent. Phase I trial patients are cancer
patients for those standard therapies failed to respond. Conventionally it is
assumed that a higher drug dose will be an effective dose. Perhaps, the high­
est dose in oncology is decided based on a maximum tolerable dose (MTD)
for cancer patients. The MTD is defined based on the dose-escalation proce­
dure. The occurrence of severe toxicities in the first cycle of cancer therapy is
defined as dose-limiting toxicities (DLTs). It is believed that the probability
of toxicity and efficacy both incline with a higher dose. Hence, the objective
is to define the MTD. Dose escalation study is performed with a pre-defined
sample size. The increment and decrement procedure are also an important
component to define MTD. The drug doses can be inclined as 10 mg, 20 mg,
and 40 mg scheme. Sometimes drug comes with the non-toxic element and
50% or 25% increment of dose schemes are carried. Conventionally, the drug
dose in phase I trial is defined by the rule-based. The model-based method
is also suitable in another scenario. We will show a different methodology to
define toxicity.

3.2.1 Rule-based designs

The rule-based design stands with prespecified dose description. In the
absence of toxicity, the escalation step is carried. Similarly, in the presence of
toxicity dose de-escalation. This rule-based procedure is performed with (I) 3+
3 design, (II) Pharmacologically guided dose escalation, and (III) Accelerated
titration design respectively.
Study Design-I 33

3.2.1.1 3+3 design

This design is free from statistical modeling. Prespecified doses are deliv­
ered bases on the Fibonacci sequence. The absence of toxicity is called the
success of a dose. The step toward a higher/lower dose is decided based on
the success/failure of a dose level. A cohort of 3 patients is considered with
the specified initial dose and thereafter proceed. The algorithm of this design
given below.

FIGURE 3.1: Flowchart of 3+3 design.

3.2.1.2 Pharmacologically guided dose escalation

This type of study useful before to carry the 3+3 design . The specific dose
to start the 3+3 design can be determined by pharmacologically guided dose
escalation (PGDE) design. The pharmacokinetic data are captured for each
patient to define the dose level. The plasma exposure concentration is deter­
mined by pre-defined concentration-time area under the ROC curve (AUC).
The 3+3 design is carried after reaching the prespecified AUC.
34 Bayesian Approaches in Oncology Using R and OpenBUGS

3.2.1.3 Accelerated titration designs

This accelerated titration designs (ATD) is an extension of the 3+3 design.
It helps to reduce exposure to toxic doses to the next patient. One patient not
achieved MTD is exposed with an escalated dose to establish the MTD. In
this design, the same patient may be exposed to multiple doses. It also helps
to treat a patient as cumulatively higher doses. However, this design only
concentrates on early toxicity, not long-term toxicity.

3.2.2 Model-based designs for determining the MTD

An alternative to the rule-based methods for finding the MTD is to assume
that there is a monotonic dose-response relationship between the dose and the
probability of DLT for patients treated at that dose. In this approach, a dose-
toxicity curve as well as the TTL are explicitly defined. The goal for the phase I
clinical trial is, through treating patients in a dose escalation fashion, to seek a
suitable quantile of the dose-toxicity curve; specifically, a dose that will induce
a probability of DLT at a specified target toxicity level. This method is most
conveniently carried out under the Bayesian framework. Simple one- or two-
parameter parametric models are often used to characterize the dose-toxicity
relationship, with the Bayesian posterior distribution used to estimate the
parameters. These designs use all the data to model the dose-toxicity curve,
and provide a credible interval for the MTD at the end of the trial.

3.2.2.1 Continual reassessment method (CRM)

The Bayesian technique is a phase I study is performed with the continual
reassessment method (CRM). The dose and toxicity measurement are linked
with statistical modeling. There are different models like the logistics model,
tangent model, and power model. The probability of dose toxicity is defined
as p(d). The model is defined as
exp(3 + ad)
Logistic: p(d) = (3.1)
1 + exp(3 + ad)
exp
Hyperbolic tangent p(d) = [(tanh(d) + 1)/2]a = [ ]a (3.2)
exp + exp(−d)
Power : p(d) = dexp(a) (3.3)
The parameter (a) is requried to estimate. The likelihood to obtaint the pos­
terior estimate of a is defined as
n
�
L(a; d, y) p(di )yi [1 − p(di )]1−yi (3.4)
i=1

Now di and yi are the dose level and toxicity outcome for patient i, and where
yi = 1 if a DLT is observed and yi = 0 if not. The algorithm to perform the
CRM design is given in the following text.
Study Design-I 35

FIGURE 3.2: Flowchart of continual reassessment design.

3.3 Study Design Package Using R

Phase I Dose-Escalation Method Using R

library(“bcrm”)
p.tox0 <­ c(0.05, 0.10, 0.20, 0.30)
dose.label <­ c(5, 10, 15, 25)
bcrm(stop = list(nmax = 28), p.tox0 = p.tox0,
dose = dose.label,ff = “power”,
prior.alpha = list(1, 1, 1), target.tox = 0.30,start = 1)
36 Bayesian Approaches in Oncology Using R and OpenBUGS

R Output

RECOMMENDED DOSE FOR PATIENTS 1 to 3 IS: 5

ENTER DOSE FOR PATIENTS 1 to 3
POSSIBLE CHOICES ARE 5 10 15 25 40 50 60
EITHER:
a) HIT 'RETURN' TO ACCEPT RECOMMENDATION
b) TYPE ANOTHER POSSIBLE DOSE AND HIT 'RETURN' TO USE
ANOTHER DOSE
c) TYPE '0' AND HIT 'RETURN' TO EXIT AND RETURN
CURRENT RESULTS)
5
ENTER TOXICITY DATA FOR PATIENT 1 (1 = TOX, 0 = NO TOX):
0
ENTER TOXICITY DATA FOR PATIENT 2 (1 = TOX, 0 = NO TOX):
0
ENTER TOXICITY DATA FOR PATIENT 3 (1 = TOX, 0 = NO TOX):
0
ENTERED VALUES:
DOSE ... 5
TOXICITIES ... 0 0 0
HIT 'RETURN' IF OK OR ANY OTHER KEY TO ENTER NEW VALUES
RECOMMENDED DOSE FOR PATIENTS 4 to 6 IS: 10
ENTER DOSE FOR PATIENTS 4 to 6
POSSIBLE CHOICES ARE 5 10 15 25 40 50 60
EITHER:
a) HIT 'RETURN' TO ACCEPT RECOMMENDATION
b) TYPE ANOTHER POSSIBLE DOSE AND HIT 'RETURN' TO USE
ANOTHER DOSE
c) TYPE '0' AND HIT 'RETURN' TO EXIT AND RETURN CURRENT
RESULTS)
Study Design-I 37

FIGURE 3.3: Phase I dose-escalation method using R.

Dose-Escalation to Decide the DLT with R

library(“bcrm”)
dose <­ c(1,2.5,5,10,15,20,25,30,40,50,75,100,150,200,250)
p.tox0 <­ c(0.010, 0.015, 0.020, 0.025, 0.030, 0.040, 0.050,
0.100, 0.170, 0.300, 0.400, 0.500, 0.650, 0.800, 0.900)
## Data from the first 5 cohorts of 18 patients
data <­ data.frame(patient=1:18, dose=rep(c(1:4, 7),
c(3, 4, 5, 4, 2)), tox=rep(0:1, c(16, 2)))
## Target toxicity level
target.tox <­ 0.30
## A 1-parameter power model is used, with standardised
doses calculated using
## the plug-in prior median
## Prior for alpha is lognormal with mean 0 (on log scale)
## and standard deviation 1.34 (on log scale)
Power.LN.bcrm <­ bcrm(stop=list(nmax=18), data=data,
p.tox0=p.tox0, dose=dose, ff=“power”,
prior.alpha=list(3, 0, 1.342 ),
target.tox=target.tox, constrain=FALSE,
sdose.calculate=“median”, pointest=“mean”)
38 Bayesian Approaches in Oncology Using R and OpenBUGS

R Output to Finalized the Dose

plot(Power.LN.bcrm)

FIGURE 3.4: Dose­escalation cancer trial of DLT.

Chapter 4

Study Design-II

Abstract
It is a difficult task to decide optimal study design for phase II study. Some­
times, it becomes more complicated while work with new molecular targeted
agents. as a result, phase III gets failed. It raises the issue that several agents
to be tested with the increasing complexity and cost. Simultaneously, the
study design is expected to be robust and efficient. This chapter is dedicated
to a phase II study design. Different study designs are considered to perform
phase II study. Similarly, different types of outcomes may arise in phase II
study as binary and continuous outcomes. Results observed with binary or
continuous are illustrated. Likewise, the sample size is another issue for the
phase II study design. Sample size determination procedure also elaborated.
Different R packages required to perform phase II study with the Bayesian
approach are explained. Example are presented with prior value, likelihood
and posterior estimates generation. This chapter will help to perform a Phase
II study in oncology domains.

4.1 Introduction
Phase II oncology trials are performed to decide a trial can be conducted
in massive or not. The primary endpoint in oncology trial is tumor size reduc­
tion. Similarly, the disease progression both is critical endpoints in Phase-II
clinical trials. Drugs used in oncology are cytostatic. The cytostatic drug de­
stroys the tumor cells and increases survival. The effect of the cytostatic drug
is measured by progression-free survival. However, the tumor progression and
tumor shrinkage both are anatomical burdens of the tumor. The tumor size re­
duction is decided based on the Response Evaluation Criteria in Solid Tumors
(RECIST). There are clinicians from academia and industry, image special­
ist, statistician and another subject expert to as RECIST Working Group.
The RECIST guideline is routinely updated. The target of Cytostatic drugs
is molecularly targeted agents (MTA). The control on MTA helps to prolong

39
40 Bayesian Approaches in Oncology Using R and OpenBUGS

the survival. The solid tumor classified as progressive disease (PD), partial
responses (PR), stable disease (SD) or complete responses (CR).The terms
PD and SD stand with treatment failure. Similarly, treatment success as CR
and PR. Jointly the PR and CR patients are defined as the objective response
rate (ORR). The proportion of patients under disease control are defined as
the disease control rate (DCR). It is better to not bring Phase II patients into
Phase III due to the occurrence of the substantial risk of toxicity or death.
The measurement of toxicity requires to take in Phase II trial design. The
combined inference about treatment success is decided by (I) change in tumor
size, (II) occurrence of new lesions, and (III) death/relapse. The treatment
failure is defined as a dichotomous endpoint ( died or not). But the reduction
in tumor size is measured as continuous (i.e., the amount of tumor shrinkage).
Perhaps, the tumor shrinkage can be classified into a dichotomous category
by predefined threshold reduction. The consolidated endpoint is measured by
continuous and binary components. The challenge is to estimate the differ­
ent probability and compare the uncertainty of the estimates by confidence
interval (CI).

FIGURE 4.1: Disease progression cycle.

Study Design-II 41

4.2 Methods
4.2.1 Estimating treatment success
Suppose that the proportion of tumor shrinkage is assumed as 20%, and
these patients are free from toxicity or death. A total of n patients are allocated
to the treatment under consideration. The measurements are taken as t1i ,t2i
and t3i . Now D1i = 1 if patients “i” fails due to other reason and D2i = 1 if
fails occurs during the two measurement. The log tumor size ratio is defined
as
z1i z2i
(y1i , y2i ) = (log( ), log( ) (4.1)
z0i z0i
Further, the tumor shrinkage model is defined with bivariate distribution
�
(y1i , y2i )T |zi0 ≈ N ((µ1i , µ2i )T , ) (4.2)
where
µ1i = α + γzi0 , µ2i = β + γzi0 (4.3)
If non-shrinkage tumor size is intermediately unmeasured, then it can be
treated as missing at random (MAR). The model can be estimated by the
linear model. The probability of non-shrinkage failure before the interim mea­
surement is presented as a probability of Di1 = 1. If non-shrinkage tumor size
is intermediately unmeasured, then it can be treated as missing at random
(MAR). The model can be estimated by the linear model. The probability of
non-shrinkage failure before the interim measurement is presented as a prob­
ability of Di1 = 1. Logistics Model is useful for both model

Logit(P (Di1 = 1)|Zi0 ) = αD1 + γD1 zi0 (4.4)

Logit(P (Di2 = 1|Di1 = 0, Zi0 , Zi1 ) = αD2 + γD2 zi1 (4.5)

�
Let the covariance matrix is . The vector parameter θ is used to represent
the tumor shrinkage and non-shrinkage model. The probability of treatment
success for the ith patients with baseline tumor size zoi is as follows:
� ∞
P (Si = 1|z0i , θ) = P (Si = 1|z0i , y1i , y2i , θ)fY1 ,Y2 (y1i , y2i ; θ)dy1i dy2i (4.6)
−∞

� log(0.7) � ∞
P (Si = 1|z0i , θ) =
−∞ −∞ (4.7)
P (D1i = 0|z0i , θ)P (D2i = 0|D1i = 0, z0i , y1i ; θ)dy1i dy2i

The pdf of the bivaraite distribution is defined as fY1 ,Y2 (y1i , y2i ; θ).
42 Bayesian Approaches in Oncology Using R and OpenBUGS

The mean of treatment success is presented as

n
� P (Si = 1|z0i , θ)
Pˆ (S = 1|θ) = (4.8)
i=1
n

and estimated as P̂ (S = 1|θ̂).

The confidence interval for l(θ) is obtained through variance of l(θ̂) as

var(l(θ̂)) ≈ (�l(θ̂))T var(θ̂)(�l(θ̂)) (4.9)

The delta method is useful to obtain the estimate.

4.2.2 Treatment difference testing

Suppose there are 2n patients allocated to two arms by n. The tumor
shrinkage is defined as
�
(yi1 , yi2 )T |ti , zi0 ∼ N ((µ11 , µ2i )T , ) (4.10)

The ith patients 1st observation is determined as yi1 . The mean response of
ith patients 1st measurement is defined as

µ1i = µ + β1 ti + γzi0 (4.11)

µ2i = µ + δ + β2 ti + γzi0 (4.12)

the treatment is defined as ti for the ith patient.
Now the tumor non-shrinkage model is defined as

Logit(P (Di1 = 1|ti , zi0 )) = αD1 + βD1 ti + γD1 zi0 (4.13)

Logit(P (Di2 = 1|Di1 = 0, ti , zi0 , zi1 )) = αD2 + βD2 ti + γD2 zi1 (4.14)

The probability of treatment sucess from shrinkage and non-shrinkage is de­

fined as P (S = 1|t, z0 , θ). It is reformulated as
2n
� P (Si = 1|t = 1, z0i , θ) − P (Si = 1|t = 0, z0i , θ)
m(θ) = (4.15)
i=1
2n

The term θ is used as vector parameter and it can be formulated as

2n
� P (Si = 1|t = 1, z0i , θ) − P (Si = 1|t = 0, z0i , θ)
m(θ) = (4.16)
i=1
2n

The estimate of m(θ) is determined as m(θ̂). Th e delta method is useful to

obtain the estimate of m(θ). The Wald statistic is performed to obtain the
test statistics.
Study Design-II 43

4.3 Illustration with Bayesian Using R

# Different Prior Distribution Assumption

library(“trialr”) # Call trialr libary in R#

library(“Rcpp”)# Call Rcpp libary in R#
priors <­ list(alpha mean = 0, alpha sd = 1,
beta mean = 0, beta sd = 1,gamma mean = 0, gamma sd = 1,
sigma mean = 0, sigma sd = 1,omega lkj eta = 1,
alpha d1 mean=0,alpha d1 sd=1,gamma d1 mean=0,gamma d1 sd=1,
alpha d2 mean=0,alpha d2 sd=1,gamma d2 mean=0,gamma d2 sd=1)
#generate data named as ”prior” #
#with different distribution assumption#
library(“trialr”)
N <­ 50 # Assigned sample size 50 to generate data#
sigma <­ 1 # seed value assumption about varaince#
delta1 <­ -0.356 # seed value assumption about delta#
mu <­ c(0.5 × delta1, delta1) # generate mean value#
Sigma = matrix(c(0.5 × sigma2 , 0.5 × sigma2 , 0.5 × sigma2 ,
sigma2 ),ncol = 2)
alphaD <­ -1.5
gammaD <­ 0
set.seed(123456)
y <­ MASS::mvrnorm(n = N, mu, Sigma)
z0 <­ runif(N, min = 5, max = 10)
z1 <­ exp(y[, 1]) × z0
z2 <­ exp(y[, 2]) × z0
d1 <­ rbinom(N, size = 1, prob = gtools::inv.logit(alphaD
+ gammaD × z0))
d2 <­ rbinom(N, size = 1, prob = gtools::inv.logit(alphaD
+ gammaD × z1))
tumor size <­ data.frame(z0, z1, z2) # Sizes in cm
non shrinkage failure <­ data.frame(d1, d2)
fit <­ stan augbin(tumor size, non shrinkage failure,
prior params=priors,model =’2t-1a’, seed = 123)
fit
44 Bayesian Approaches in Oncology Using R and OpenBUGS

Sampling for Model ‘AugBin2T1A’ with (Chain 1)

Chain 1:
Chain 1: Gradient evaluation took 0 seconds
Chain 1: 1000 transitions using 10 leapfrog steps
per transition would take 0 seconds.
Chain 1: Adjust your expectations accordingly!
Chain 1:
Chain 1:
Chain 1: Iteration: 1 / 2000 [ 0%] (Warmup)
Chain 1: Iteration: 200 / 2000 [ 10%] (Warmup)
Chain 1: Iteration: 400 / 2000 [ 20%] (Warmup)
Chain 1: Iteration: 600 / 2000 [ 30%] (Warmup)
Chain 1: Iteration: 800 / 2000 [ 40%] (Warmup)
Chain 1: Iteration: 1000 / 2000 [ 50%] (Warmup)
Chain 1: Iteration: 1001 / 2000 [ 50%] (Sampling)
Chain 1: Iteration: 1200 / 2000 [ 60%] (Sampling)
Chain 1: Iteration: 1400 / 2000 [ 70%] (Sampling)
Chain 1: Iteration: 1600 / 2000 [ 80%] (Sampling)
Chain 1: Iteration: 1800 / 2000 [ 90%] (Sampling)
Chain 1: Iteration: 2000 / 2000 [100%] (Sampling)
Chain 1:
Chain 1: Elapsed Time: 5.849 seconds (Warm-up)
Chain 1: 4.818 seconds (Sampling)
Chain 1: 10.667 seconds (Total)
Chain 1:
Study Design-II 45

Sampling for Model � AugBin2T1A� with (Chain 2)

Chain 2:
Chain 2: Gradient evaluation took 0 seconds
Chain 2: 1000 transitions using 10 leapfrog steps
per transition would take 0 seconds.
Chain 2: Adjust your expectations accordingly!
Chain 2:
Chain 2:
Chain 2: Iteration: 1 / 2000 [ 0%] (Warmup)
Chain 2: Iteration: 200 / 2000 [ 10%] (Warmup)
Chain 2: Iteration: 400 / 2000 [ 20%] (Warmup)
Chain 2: Iteration: 600 / 2000 [ 30%] (Warmup)
Chain 2: Iteration: 800 / 2000 [ 40%] (Warmup)
Chain 2: Iteration: 1000 / 2000 [ 50%] (Warmup)
Chain 2: Iteration: 1001 / 2000 [ 50%] (Sampling)
Chain 2: Iteration: 1200 / 2000 [ 60%] (Sampling)
Chain 2: Iteration: 1400 / 2000 [ 70%] (Sampling)
Chain 2: Iteration: 1600 / 2000 [ 80%] (Sampling)
Chain 2: Iteration: 1800 / 2000 [ 90%] (Sampling)
Chain 2: Iteration: 2000 / 2000 [100%] (Sampling)
Chain 2:
Chain 2: Elapsed Time: 5.702 seconds (Warm-up)
Chain 2: 14.104 seconds (Sampling)
Chain 2: 19.806 seconds (Total)
Chain 2:
46 Bayesian Approaches in Oncology Using R and OpenBUGS

Inference for Stan Model: AugBin2T1A

4 chains, each with iter=2000; warmup=1000; thin=1;

post-warmup draws per chain=1000, total post-warmup

draws=4000.
Mean SE SD 2.5% 97.5% neff R
alpha 0.07 0.01 0.44 -0.78 0.94 2008 1
beta -0.11 0.01 0.44 -0.97 0.75 2122 1
gamma -0.02 0.00 0.06 -0.14 -0.09 2135 1
Omega11 1.00 NaN 0.00 1.00 1.00 NaN NaN
Omega1,2 0.73 0.00 0.07 0.58 0.84 2733 1
Omega2,1 0.73 0.00 0.07 0.58 0.84 2733 1
Omega2,2 1.00 0.00 0.00 1.00 1.00 3974 1
sigma1 0.74 0.00 0.07 0.62 0.90 3058 1
sigma2 1.04 0.00 0.10 0.86 1.27 3097 1
alphaD1 -0.22 0.02 0.85 -1.90 1.44 2734 1
gammaD1 -0.15 0.00 0.13 -0.40 0.10 2861 1
alphaD2 -0.30 0.01 0.52 -1.33 0.72 3288 1
gammaD2 -0.08 0.00 0.06 -0.23 0.03 3021 1
For each parameter, neff is a crude measure of effective
sample size,and Rhat is the potential scale reduction
factor on split chains (at convergence, R=1$).

4.4 Discussion
The proposed methods work in combination with binary and continuous
outcomes. It is suited well in phase II cancer research. The method compatible
one decision about binary and another response about continuous outcomes.
The sample size determination is a phase II clinical trial is a significant issue.
Conventionally study design is required to calculate the sample size. It can
be specified that any Phase II study planned with 50 sample size can be
accommodated with the pre-specified methodology.
Chapter 5

Optimum Biological Dose Selection

Abstract
The biomarker is a realistic indicator to explore angiogenesis in cancer.
Recently, Subtoxic doses of chemotherapy are administered and defined as
Metronomic Chemotherapy (MC). Now set the best effective treatment in MC
is a challenging task. This chapter is elaborated about dose fixation by looking
biomarker level. The dose could help to maintain the desired level of the
biomarker as the optimum biological dose (OBD). We presented the Bayesian
algorithm to determine the OBD. The curve fitting procedure is illustrated in
OpenBUGS to determine the OBD. This work is explored by looking at the
performance of one surrogate marker toward influencing the disease outcome.
Disease outcome is defined as time-to-event outcomes. Conventional, time-to­
event methodology like Kaplan-Meier estimates are elaborated. Similarly, the
treatment outcome is classified as a dose that is suitable to control the OBD.
In this chapter, theoretical explanation about Toxicity profile testing, Logistic
Response Model and Quadratic Logistic Design are detailed with Bayesian
illustration. The illustration will help to determine the best dose by looking
at the posterior estimates of regression coefficients. Similarly, model suitability
is explained by standard deviation and model selection criteria.

5.1 Introduction
The maximum dose of the effective drug within the tolerable limit is called
the maximum tolerated dose (MTD). Conventionally, it assumed that in­
creased dose is effective. The best effective dose concluded by one level below
the tolerable highest dose [3, 4, 5]. Metronomic chemotherapy (MC) emerges
as a therapeutic option in medical oncology [6, 7, 8]. Sometimes, conventional
chemotherapy becomes resistant [8, 9, 10]. Development of chemoresistance
is a common problem [11]. The alternative of high-dose chemotherapy stands
with Metronomic chemotherapy (MC). The MC causes less severe side effects
than standard chemotherapy. The MC is marked as low-dose chemotherapy.

47
48 Bayesian Approaches in Oncology Using R and OpenBUGS

Now the determination of best effective dose in MC is a challenging task. The

predefined biomarker is considered to represent the disease status. The thresh­
old value of a biomarker is required to define the disease status. Thereafter the
task is to explore that the MC is most effective to maintain the biomarker’s
desired value. The MC is decided by the optimum biological dose (OBD)[12].
In this chapter, we will explore the Bayesian approach to decided the low dose
of chemotherapy.

5.2 Illustration with Head and Neck Cancer Data

Establishing a biomarker for cancer is a difficult task. Unless biomarker is
established, it is challenging to establish the best effective metronomic dose.
We will illustrate the methodology with simulated data. A simulation study
was carried to explore the OBD and check the MC performance. The skewed
distribution was used to obtain the head and neck cancer (HNC) data. The
circulating endothelial cell (CEC) is considered as a biomarker. The repeatedly
measured CEC is controlled within the desired level or not is considered as an
outcome of dose. The prior mean 114 and standard deviation 15 are considered
to generated CEC measurement. A total of 4 doses named as dose 1, dose 2
dose 3 and dose 4 are presented. Data is generated for a total of 220 patients.
Persons alive status is presented as 0 or 1. the solid tumor size is generated
from (mean 4.0 cm, SD 2). histological grading is measured as ( well, moderate
and poor)were generated randomly for a total of 220 patients and merged
with the excel sheet. The survival representation is provided by the Kaplan-
Meier curve on progression-free survival in Figure 5.4. Relapse is considered an
event. A total of five follow-up visits observations for CEC and tumor size were
assumed to be distributed with a normal distribution. Different time points
are assumed as t1 , t2 , t3 , t4 , t5 . Corresponding mean of CEC is assumed as 124,
128, 135, 126, and 120 respectively. Censored distribution about duration of
survival was obtained from a hazard rate of 0.0003 with approximately 45%
censored observations.

5.3 Toxicity Profile Testing

Suppose the specific dose is defined as dj from a set of doses i.e.(d1 , ..., dJ )
. Now a total of n patients are exposed with dose level j. However, a low dose
of chemotherapy is free from toxicity. However, we show a provision about
measure the toxicity profile in addition to the efficacy level.
Optimum Biological Dose Selection 49

Now the qj is presented as the toxicity of dose level dj . A total of xj

patients are observed with toxicity occurrence with a specific dose of dj . The
prior assumption about dj is obtained as

xj ∼ binom(nj , qj ) (5.1)

Probability of toxicity qj is assumed as

qj ∼ beta(a, b) (5.2)

The hyperparameters are a and b. maximum tolerable toxicity range is

φ.Patients exposed to dose level are presumed as j is xj . Posterior distri­
bution about Pr(qj > φ|ηj , xj ) is obtained to perform the model diagnostic
among different doses dj .

A = {j : P¯r(qj > φ|ηj , xj ) < Cτ , j = 1, .....J} (5.3)

The prespecified dose limit is Cτ with specific toxicity level. The term Cτ can
be defined as a tolerable limit or threshold limit. It helps to protect patients
from the occurrence of over toxicity. A small prior assumption of a, b are used
with threshold value as Beta(φ; a, b) = 1 − Cτ + δ. The limit of δ is captured
with a small positive number 0.5.

5.4 Logistic Response Model

The logistics model is defined below. The biomarker is considered as CEC.
The dose could maintain the limit of a CEC is defined as OBD.The probability
of sucess is defined as g(pj ). The corresponding dose is termed as dj . The
function f (s) is used to define the threshold value of OBD as

g(pj ) = β1 + β2 dj (5.4)
The predefined dose of OBD is determined as [c1 , c2 ] with uniform distribution
� 1
if c1 ≤ s ≤ c2 ;
f (s) = c2 −c1
0 if otherwise.
The cummulative dose is presented as pj as
dj
dj − c1
�
pj = f (s)ds = for c1 ≤ dj ≤ c2 (5.5)
c1 c2 − c1
−c1 1
It becomes β1 = c2 −c1 and β2 = c2 −c1
50 Bayesian Approaches in Oncology Using R and OpenBUGS

The OBD is defined as c1 ≤ dj ≤ c2 . It is expected that pj will distribut

normally as
� x
1 1 s−µ 2
pj = exp[− ( ) ]ds, Φ ∼ N (0, 1) (5.6)
2π −∞ 2 σ

Φ−1 (pj ) = β1 + β2 dj (5.7)

where β1 = −µ/σ and β2 = 1/σ. The link function g is adopted with
inverse cumulative Normal probability function Φ−1 .
The function of OBD is defined as
β2 exp(β1 + β2 s)
f (s) = (5.8)
[1 + exp(β1 + β2 s)]2
so � x
exp(β1 + β2 dj )
pj = f (s)ds = (5.9)
−∞ [1 + exp(β1 + β2 dj )]
The link function is considered as
pj
log( ) = β 1 + β2 dj (5.10)
1 − pj

The probability of sucess due to dose label dj is pi . The model is defined as

exp(β1 + β2 dj )
pj = (5.11)
1 + exp(β1 + β2 dj )
pj
log( ) = β 1 + β2 dj (5.12)
1 − pj
log(1 − pj ) = −log[1 + exp(β1 + β2 dj )] (5.13)
If the patients size is N then the log-likelihood function can be defined as
N
�
l= [yi (β1 + β2 dj ) − ni log[1 + exp(β1 + β2 dj )] + log(nyii ) (5.14)
i=1

δl � exp(β1 + β2 di ) �
U1 = = {yi − ni [ ]} = (yi − ni pj ) (5.15)
δβ1 1 + exp(β1 + β2 di )
δl � exp(β1 + β2 dj )
U2 = = {yi dj − ni dj [ ]} (5.16)
δβ2 1 + exp(β1 + β2 dj )
� �
U2 = (yi − ni pj ) = dj (yi − ni pj ) (5.17)
The information matrix is defined as
� � � �
n p (1 − pj ) � ni dj pj (1 − pj )
T = �� � i j
� �
�.
ni dj pj (1 − pj ) ni d2j pj (1 − pj ) �
Optimum Biological Dose Selection 51

The maximum likelihood estimates is defined as

T (m−1) bm = I (m−1) b(m−1) + U (m−1) (5.18)

The number of approximation is defined as (m) and vector of estimates is

given as b through with deviance function as,
N
� yi n i − yi
D=2 [yi log( ) + (ni − yi )log( )] (5.19)
i=1
ŷi ni − yˆi

5.5 Dose Selection Algorithm

Dose Selection Algorithm

The OBD of low-dose chemotherapy will be decided based on the fol­

lowing steps:
(I)Start to recruit a very minimal number of patients for a specific dose
level j.
(II)The minimal number of patients will be defined as l.
(III)Measure the toxicity for each dose level.
(IV) Similarly, the dose level l − 1 will be measured.
(V) Credible interval of posterior estimates will be obtained for each
dose by posterior probability distribution with Pr(β > 0|dj ).
(V) Now it may possible that the dose j +1 comes with Pr(β > 0|dj ) >
CE1 .
(VI)Decision about escalation will be concluded by looking dose level
Pr(β > 0|dj ) < CE2 probability.
(VII) Decision about de-escalation will be concluded by dose level
Pr(β > 0|dj ) < CE2 , to the dose level j − 1.
(VIII)The next step is to continue the study with a specific dose and
cover the maximum sample size.
(IX) Finally, the optimum dose will be determined by the maximum
pick of the curve reached by the dose levels.
(X) The best dose will be obtained by calibration on CE1 and CE2 .
52 Bayesian Approaches in Oncology Using R and OpenBUGS

5.6 Quadratic Logistic Design

A specific dose is effective or not can be classified into binary format. Let
us assume that the dose level is j and sucess of dose is pj . The quadratic
lonistic curve is defined as
pj
log( ) = α + βdj + γd2j , j = 1, 2, ....j (5.20)
1 − pj

Suppose there are γi patients treated with dj dose from a toal of ηj . The
likelihood function is D = {yj , ηj ; j = 1, ..., J} with
J 2
� eα+βdj +γdj 1
L(D|α, β, γ) ∝ ( α+βdj +γd2j
)γj ( α+βdj +γd2j
)ηi −γi (5.21)
j=1 1+e 1+e

The posterior estimates of the parameters will be obtained by Markov

chain Monte Carlo (MCMC) method. The α is assumed with prior assump­
tion Cauchy(0,10) distribution. Similarly, β and γ will be obtained from
Cauchy(0,2.5).

TABLE 5.1: Median estimates of survival among different doses

Doses Number at risk Events Median 0.95LCL 0.95UCL
d1 249 139 50 43 58
d2 246 130 53 49 64
d3 246 135 49 41 57
d4 236 122 48 42 57

TABLE 5.2: Posterior estimates of the models through measurement of sur­

rogate marker
Mode l Parameter Posterior Mean SD HPD DIC
β1 332.2 74.81 (216.9,450.1)
Model1 β2 -8.182 6.47 (18.56,1.942) 253.2
σ1 67.19 12.21 (48.25,95.6)
τ1 0.00024 0.00008 (0.0001,0.0004)
α 679.0 2621.0 (-604.6, 10660.0)
β -74.28 502.5 (-1983.0, 170.8)
Model 2 γ 2.923 22.45 (-7.975, 87.25) 226.4
σ2 74.66 28.77 (47.66, 179.7)
τ2 0.00021 0.00010 (0.00025,0.00043)
Optimum Biological Dose Selection 53

TABLE 5.3: Performance of methods through the measurement of surrogate

marker
Dose
Scenario Characteristics d1 d2 d3 d4 Overtoxic
Scenario 1: Selected Probability 0.01 0.10 0.20 0.69 0
Proportion of Patients 0.08 0.19 0.1 0.63
Scenario 2: Selected Probability 0.12 0.15 0.23 0.50 0
Proportion of Patients 0.18 0.22 0.09 0.51
Scenario 3: Selected Probability 0.21 0.27 0.15 0.37 0.01
Proportion of Patients 0.35 0.29 0.04 0.32
Scenario 4: Selected Probability 0.29 0.43 0.18 0.10 0
Proportion of Patients 0.49 0.31 0.1 0.10

5.7 Illustration
In this illustration, two models are provided. Results are obtained through
posterior mean and standard deviation. A different scenario is prepared, and
those are given in Table 5.3. It provides the scope to consider the doses and
toxicity occurrences. The optimal dose can be obtained from this table. Since

FIGURE 5.1: Different doses and biomarker value changes.

54 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 5.2: Biomarker changes over the visits.

the work is about low-dose chemotherapy, then the appearance of over toxicity
is very minimal.
This work is presented by a simulation study. The posterior probability of
efficacy and safety is presented as E and S, respectively. Suppose the marginal
posterior probability of the associated parameters are defined as (σ2 or τ2 ) for
S and E respectively.
The parameter S and E is linked as Pr(S = 1) = V ∗ Pr(E = 1). Now
V = 1 shows that OBD through a biomarker is achieved. If V > 1, then the
OBD of a biomarker is not achieved. Now the term σ and τ represents the
relation between S and E. if there is no relation between σ1 and τ1 then the
efficacy E could be ignored and the corresponding dose may be ignored. Now
Logistic response model is defined as Model1. Similarly, Quadratic Logistic
design is defined as Model 2.The additive parameters σ and τ for Model 1 is
defined as σ1 and τ1 respectively. Similarly, Model 2 is defined as σ2 and τ2 . In
this illustration a total of 20,000 iterations with 1,000 burn-in is carried out
with different fixed value of σ1 , τ1 , σ2 and τ2 . The simulation is performed to
obtain the convergence. The convergence is obtained by observing the trace
Optimum Biological Dose Selection 55

FIGURE 5.3: Mean Biomarker changes over the visits.

plots for each parameter of the model. The models are compared by Decision
Information Criteria(DIC) value.
56 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 5.4: Dose-wise survival curve comparison.

Optimum Biological Dose Selection 57

5.8 Bayesian Dose Selection Using OpenBUGS

# OpenBUGS code 1

Model1
model
{
for (i in 1:20)
{
mu.dmf[i] <­ beta1 + beta2 × fl[i]
dmf[i] ∼ dnorm(mu.dmf[i],tau)
}
beta1 ∼ dnorm(0.0,0.000001)
beta2 ∼ dnorm(0.0,0.000001)
sigma1 ∼ dunif(0,400)
tau1 <­ 1/(sigma1×sigma1)
for (i in 1:20)
{ # calculate residuals
residual[i] <­ dmf[i] - mu.dmf[i]
}
pred.mean.1.7 <­ beta1 + beta2×1.7
pred.ind.1.7 ∼ dnorm(pred.mean.1.7, tau1)
}
58 Bayesian Approaches in Oncology Using R and OpenBUGS

# OpenBUGS code 2

Model2
model {
for (i in 1:20)
{
mu.dmf[i] <­ alpha + beta×fl[i] + gamma×fl[i]×fl[i]
# regression equation
dmf[i] ∼ dnorm(mu.dmf[i],tau)
# distribution individual values
}
alpha ∼ dnorm(0.0, 0.000001)
beta ∼ dnorm(0.0, 0.000001)
gamma ∼ dnorm(0.0, 0.000001)
sigma2 ∼ dunif(0,200)
tau2 <­ 1/(sigma2×sigma2)
for (i in 1:20)
{ # calculate residuals
residual[i] <­ dmf[i]-mu.dmf[i]
}
pred.mean.1.7 <­ alpha + beta1×1.7 + beta2×1.7×1.7
# mean prediction for fl=1.7
pred.ind.1.7 ∼ dnorm(pred.mean.1.7, tau2)
# individual pred for fl=1.7
}
Optimum Biological Dose Selection 59

# OpenBUGS code 2

dmf[] fl[]
159 15
168 7.5
201 15
206 15
209 15
212 15
230 12.5
265 12.5
270 9
285 7.5
290 7.5
195 9
201 9
236 7.5
191 12.5
225 12.5
232 7.5
265 12.5
396 9
412 9
END
list(dmf=c( 159,168,201,206,209,212,230,265,
270,285,290,195,201,236,191,225,232,265,396,
412),
fl=c(15,7.5,15,15,15,15,12.5,12.5,9,7.5,
7.5,9,9,7.5,12.5,12.5,7.5,12.5,9,9))}

5.9 Discussion
The MC can be classified as next-generation chemotherapy [13]. It is possi­
ble to explore the performance of one surrogate marker toward influencing the
disease outcome. Similarly, the treatment outcome can be classified as a dose
best to control the OBD. Perhaps, it is difficult to establish the biomarker or
surrogate marker.
60 Bayesian Approaches in Oncology Using R and OpenBUGS

5.10 Dose Finding Package Using R

5.10.1 DoseFinding
#DoseFinding package with R for samplesize determination

install.packages(“DoseFinding”)
library(“DoseFinding”)
doses <­ c(0, 10, 25, 50, 100, 150)
fmodels <­ Mods(linear = NULL, emax = 25,
logistic = c(50, 10.88111), exponential = 85,
betaMod = rbind(c(0.33, 2.31), c(1.39, 1.39)),
linInt = rbind(c(0, 1, 1, 1, 1),
c(0, 0, 1, 1, 0.8)),
doses=doses, placEff = 0, maxEff = 0.4,
addArgs=list(scal=200))
## Calculate doses giving an improvement of 0.3 over placebo
TD(fmodels, Delta=0.3)
## discrete version
TD(fmodels, Delta=0.3, TDtype =“discrete”, doses=doses)
## doses giving 50% of the maximum effect
ED(fmodels, p=0.5)
ED(fmodels, p=0.5, EDtype = “discrete”, doses=doses)
plot(fmodels, plotTD = TRUE, Delta = 0.3)
Optimum Biological Dose Selection 61

FIGURE 5.5: Dose determination by different model.

Part II

Bayesian in Time-to-Event

Data Analysis

63
Chapter 6

Survival Analysis

Abstract
The survival analysis is widely adopted statistical analysis in oncology.
Commonly, the Kaplan-Meier estimate and Cox proportional hazard models
are performed in survival analysis. The Kaplan-Meier showed by assuming
the censoring mechanism is non-informative. However, it gets violated some
times. The informative censoring procedure using a Bayesian framework is
suitable to overcome this violation. This chapter is dedicated to Bayesian sur­
vival analysis. Real-life data analysis is illustrated with OpenBUGS software.
Similarly, survival analysis with R by Kaplan-Meier method, Cox PH model,
Schoenfeld Residuals and other tools are theoretically explained.

6.1 Introduction
The objective of the survival analysis is to estimate and interpret hazard
functions from survival data. Secondly, compare the hazard functions. Thirdly,
explore the relationship of explanatory variables to survival time. Hazard func­
tion created from several groups.
Survival analysis comes with an analytical problem as censoring. Censoring
comes while we have some information about individual survival time, but we
do not have actual survival time.
The censoring data occurred if the event may not occur before the study end,
the person is lost to follow-up during the study ends, and a person withdraws
from the study due to other causes of death.
If the real survival time is equal to or greater than the observed survival time
then it is defined as Right censoring.
Similarly, if the real survival time is less than or equal to the observed survival
time, then it is defined as Left censoring.
Conventionally, the survival analysis is performed with hazard function. The
hazard function h(t) is defined as the instantaneous potential per unit time
for the event to occur, given that the individual has survived up to time t.

65
66 Bayesian Approaches in Oncology Using R and OpenBUGS

6.1.1 Kaplan-Meier estimator

The survival data having censored data can be estimated by Kaplan-Meier
estimator. It is a nonparametric estimator. The survival curves are compared
by the Kaplan-Meier estimator. It is obtained as a product limit formula. The
Kaplan-Meier estimator for the survival distribution function is estimated as
� dj
Ŝ(t) = (1 − ) (6.1)
rj
tj ≤t

dj is the person died till time tj .

Now rj is the persons at risk before time tj . The variance of the Kaplan-
Meier estimator is obtained as
� dj
σ̂ 2 (t) = Ŝ(t)2 (6.2)
rj (rj − dj )
tj ≤t

In absence of censoring, the term Ŝ(t)[1 − Ŝ(t)]/n becomes to standard bino­

mial variance estimator. The function to obtain Kaplan-Meier estimates in R
is given below. The package “survival” is used to obtain the estimates.
Kaplan-Meier Using R

library(“survival”)
data(lung)
KM.result<-survfit(Surv(time, status) ∼ 1, data=lung)
KM.result

Call: survfit(formula = Surv(time, status) ~ 1, data = lung)

n events median 0.95LCL 0.95UCL

228 165 310 285 363

Kaplan-Meier Plot R

library(“survival”)
library(“ranger”)
library(“ggplot2”)
library(“dplyr”)
library(“ggfortify”)
autoplot(“KM.result”)
Survival Analysis 67

FIGURE 6.1: Kaplan-Meier plot R.

6.1.2 Nelson-Aalen estimator

Nelson-Aalen estimator is a nonparametric estimator. It is used to esti­
mate the cumulative hazard rate function from censored survival data. Here,
no distributional assumptions are required. It is more appropriate for data
exploration through graphically. It is also known as: (1) Altshuler estimator,
(2) Empirical cumulative hazard estimator, (3) Aalen-Nelson estimator. The
Nelson-Aalen estimator for the cumulaive hazard rate function takes the form
� dj
 = (6.3)
rj
tj ≤t

Here, rj is the number of individual at risk (not censored and alive). dj is

the person died till time tj . This estimator is a increasing right-continous step
d
function with increments of rjj at the observed faliure times.
Nelson-Aalen estimator performs better when the sample size is small.
It is asymptotically equivalent to Product Limit estimator and is also the
nonparametric maximum likelihood estimator. It is commonly used to check
the parametric model assumption and get a crude estimation of the hazard
function.
68 Bayesian Approaches in Oncology Using R and OpenBUGS

Nelson-Aalen Estimator Plot Using R

library(“mice”)
library(“MASS”)
lung$status <­ 2
ch <­ nelsonaalen(lung, time, status)
plot(x = lung$time y = ch,ylab=‘Cumulative hazard’, xlab=‘Time’)

FIGURE 6.2: Nelson-Aalen estimator plot obtained by R.

6.2 Bayesian in Survival Analysis

Bayesian statistics comes as an alternative to the conventional approach.
It helps to design and to analyze oncology data. Time-to-event outcomes are
common in oncology research. Several complex issues in a time-to-event out­
come can be handled with the Bayesian approach. In this chapter, we will
show the Bayesian extension of survival in simple formation.

6.2.1 Cox’s proportional hazards model

The Cox model specifies the hazard for individual i as
hi (t) = h0 (t)eXi (t)β (6.4)
where h0 is an unspecified positive function of time called the baseline hazard,
and β is a p×1 column vector of coefficients. The covaraites is defined by Xi (t).
Survival Analysis 69

6.2.2 Hazard ratio

The hazard ratio for two subjects with fixed covariate vectors Xi and Xj
are
hi (t) h0 (t)eXi β eXi β

= X β
= Xj β (6.5)

hj (t) h0 (t)e j e
The term
hi (t)
hj (t)
is constant over time and the model is known as the proportional hazards
model.

6.2.3 Partial likelihood function

Partial likelihood function is
n �
� Yi (t)ri (β, t)
P L(β) = {� }dNi (t) (6.6)
i=1 j Y j (t)rj (β, t)
t≥0

The log partial likelihood is presented as

�n � ∞ �
l(β) = [Yi (t)Xi (t)β − log( Yj (t)rj (t))]dNi (t) (6.7)
i=1 0 j

The log partial likelihood with respect to β gives the p × 1 score vector,
U (β):
n �
� ∞
U (β) = [Xi (s) − x̄(β, s)]dNi (s) (6.8)
i=1 0

and �
Yi (s)ri (S)Xi (s)
x̄(β, s) =
� (6.9)
Yi (s)ri (s)
with Yi (s)ri (s) as the weights. The maximum partial likelihood estimator is
found by solving the partial likelihood equation:
U (β̂) = 0 (6.10)
Now β̂ is consistent and asymptotically normally distributed with mean β
with variance {EI(β)}−1
Cox PH using R

library(“survival”)
data(lung)
a<-coxph(Surv(time, status) ∼ ph.ecog,data=lung)
summary(a)
70 Bayesian Approaches in Oncology Using R and OpenBUGS

#R Output

coxph(formula = Surv(time, status) ~ ph.ecog, data = lung)

n= 227, number of events= 164

(1 observation deleted due to missingness)

coef exp(coef) se(coef) z Pr(>|z|)

ph.ecog 0.4759 1.6095 0.1134 4.198 2.69e-05 ***
---
Signif. codes:0 '***' 0.001 '**'0.01'*'0.05'.'0.1' '1

exp(coef) exp(-coef) lower .95 upper .95

ph.ecog 1.61 0.6213 1.289 2.01

Concordance= 0.604 (se = 0.024 )

Likelihood ratio test= 17.57 on 1 df, p=3e-05
Wald test = 17.62 on 1 df, p=3e-05
Score (logrank) test = 17.89 on 1 df, p=2e-05

6.2.4 Stratified Cox model

Let that subjects i = 1, ...., n1 are in stratum 1, subjects nl+1 , ..., nl + n2
are in stratum 2, and so on. The hazard for an individual i, who belongs to
stratum k is
λk (t)eXi β (6.11)
The overall loglikelihood �
l(β) = lk (β) (6.12)
k=1
�
The
� score vector and information matrix are U (β) = Uk (β) and I(β) =
Ik (β) respectively.
Stratified Cox Model Using R

library(“survival”)
data(lung)
b<-coxph(Surv(time, status) ∼ sex + strata(ph.ecog),data=lung)
summary(b)
Survival Analysis 71

#R Output

coxph(formula = Surv(time, status) ~ sex + strata(ph.ecog),

data = lung)

n= 227, number of events= 164

(1 observation deleted due to missingness)

coef exp(coef) se(coef) z Pr(>|z|)

sex -0.5544 0.5744 0.1706 -3.25 0.00115 **
---
Signif. codes:0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

exp(coef) exp(-coef) lower .95 upper .95

sex 0.5744 1.741 0.4112 0.8025

Concordance= 0.585 (se = 0.022 )

Likelihood ratio test= 11.09 on 1 df, p=9e-04
Wald test = 10.56 on 1 df, p=0.001
Score (logrank) test = 10.81 on 1 df, p=0.001

6.2.5 Wald-score and likelihood ratio tests

The Wald-score and likelihood ratio are useful for the Cox partial likeli­
hood to test hypotheses about β. The likelihood ratio test is 2(l(β̂) − l(β (0) ),
twice the difference in the log partial likelihood at the initial and final esti­
mates of β̂.
�
The Wald test is (β̂ − β (0) ) Iˆ(β̂ − β (0) ), where Iˆ = I(β̂) is the estimated in­
formation matrix at the solution.
� −1
The efficient score test statistic U (β (0) )I(β (0) U (β (0) ) can be computed us­
ing only the first iteration of the Newton-Raphson algorithm.

6.2.6 Diagnostics for Cox’s PH model

A large number of diagnostic methods for checking have been proposed
over the year for Cox’ proportional hazard (PGH) regression model, nut they
are still not used as widely as they should be. The PH model [Cox (1972)]
specifies
�
h(t; x) = h0 (t)eβ x (6.13)
for the hazard function for the failure time T associated with a p-dimensional
vector of covariates x, given an unknown underlying hazard function h0 and a
72 Bayesian Approaches in Oncology Using R and OpenBUGS

vector of coefficients β which must be estimated. The hazard function of the

PH model is defined as,
�
S(t; x) = exp{−Δ0 (t)eβ x } (6.14)

The cumulative hazard function Δ0 corresponding to h0 .

�
In{−InS(t; x)} = InΔ0 + β x (6.15)
of the PH model is defined as,
�
S(t; x) = exp{−Δ0 (t)eβ x } (6.16)

Any survival function S(t; x) plotted on the complementary log-log scale, dif­

�
fers from InΔ0 by the constant amount β x throughout its duration.

Hence there are two functions. S(t; x1 ) and S(t; x2 ) for different values of the

covariate vector x, will be parallel.

This gives the basic means of inspecting the PH assumption. Reliable esti­

mates of Ŝ is important to take decision about plot.

A large amount of data available for each chosen x is significant to make a

decision about the plot.

Diagnostics for Cox’s PH Model Using R

library(“survival”)
data(lung)
res.cox<-coxph(Surv(time, status) ∼ sex,data=lung)
res.cox
test.ph <­ cox.zph(res.cox)
test.ph

#R Output

chisq df p
sex 2.86 1 0.091
GLOBAL 2.86 1 0.091

6.2.7 Schoenfeld residuals

Residuals from a fitted model are most easily understood when they ex­
press in some way the difference between observed data values and the corre­
sponding fitted (predicted) values.
Large amount of data available for each chosen x is very important to take
decision about plot. At the time i, of death or failure of unit or individual i,
there is a risk set Ri ; consisting of the individuals who have not yet failed
Survival Analysis 73

(or been censored) so were at risk of failure at this time. The conditional
probability that i failed, given that there was a failure at tl is given by
�
hi (ti ) eβ x
pi = � =� (6.17)
β � xj
j�Ri hj (ti ) j�Ri e

The expected value E(X|Ri ) is

�
xk eβ xk
�
� k�Ri
E(X |Ri ) = x k pk = � � (6.18)
k�Ri j�Ri eβ xj

The vector is defined as

ri = xi − E(X|Ri ) (6.19)
The estimate of schoenfeld residual is
r̂i = xi − Ê(X|Ri ) (6.20)
The Schoenfeld residual is closely related to the estimation procedure of Cox’s
PH model.
Graphical Test for Proportional Hazards Using R

library(“survival”)
fit <­ coxph(Surv(time,status) ∼ sex+factor(ph.ecog)+age,
data=lung,x=TRUE)
plot(cox.zph(fit))

6.3 Bayesian Survival Analysis Using R

Bayesian Survival Analysis Using R

library(“survival”)
library(“BMA”)
data(veteran)
veteran[1:10,]
test.bic.surv<­ bic.surv(Surv(time,status) ∼ ., data = veteran,
factor.type = TRUE)
summary(test.bic.surv, conditional=FALSE, digits=2)
plot(test.bic.surv)
imageplot.bma(test.bic.surv)
74 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 6.3: Beta(t) for age is obtained by imageplot.bme.

#R Script to Show Data

trt celltype time status karno diagtime age prior

1 1 squamous 72 1 60 7 69 0
2 1 squamous 411 1 70 5 64 10
3 1 squamous 228 1 60 3 38 0
4 1 squamous 126 1 60 9 63 10
5 1 squamous 118 1 70 11 65 10
6 1 squamous 10 1 20 5 49 0
7 1 squamous 82 1 40 10 69 10
8 1 squamous 110 1 80 29 68 0
9 1 squamous 314 1 50 18 43 0
10 1 squamous 100 0 70 6 70 0
Survival Analysis 75

#R Output

bic.surv.formula(f = Surv(time, status) ~ ., data = veteran,

factor.type = TRUE)
6 models were selected
Best 5 models (cumulative posterior probability = 0.95 ):
p!=0 EV SD model 1 model 2
trt 11.3 0.02909 0.1058 . .
celltype 84.6
.smallcell 0.61564 0.3538 0.7121 .
.adeno 0.97619 0.4965 1.1508 .
.large 0.28260 0.2830 0.3251 .
karno 100.0 -0.03135 0.0052 -0.0309 -0.03
diagtime 5.4 0.00018 0.0020 . .
age 6.1 -0.00036 0.0026 . .
prior 5.6 0.00058 0.0054 . .
nVar 2 1
BIC -39.3626 -36.77
post prob 0.562 0.15
p!=0 EV SD model 3
trt 11.3 0.02909 0.1058 0.2573
celltype 84.6
.smallcell 0.61564 0.3538 0.8196
.adeno 0.97619 0.4965 1.1477
.large 0.28260 0.2830 0.3930
karno 100.0 -0.03135 0.0052 -0.0311
diagtime 5.4 0.00018 0.0020 .
age 6.1 -0.00036 0.0026 .
prior 5.6 0.00058 0.0054 .
nVar 3
BIC -36.1558
post prob 0.113
model 4 model 5
trt . .
celltype
.smallcell 0.7208 0.7264
.adeno 1.1643 1.1765
.large 0.3215 0.3276
karno -0.0318 -0.0311
diagtime . .
age -0.0059 .
prior . 0.0103
nVar 3 3
BIC -34.9292 -34.7553
post prob 0.061 0.056
76 Bayesian Approaches in Oncology Using R and OpenBUGS

6.3.1 BayesSurvival
PlotBayesSurv to Create Posterior Mean with Credible Band for the
Survival Function or Cumulative Hazard.

library(“BayesSurvival”)
library(“simsurv”)
library(“ggplot2”)
hazard.true <­ function(t,x, betas, ...)
1.2×(5×(t+0.05)3 -10 × (t+0.05)2 +5×(t+0.05))+0.7
sim.df <­ data.frame(id = 1:1000)
df <­ simsurv(x = sim.df, maxt = 1, hazard = hazard.true)
bs <- BayesSurv(df, “eventtime”, “status”)
PlotBayesSurv(bs, object = “survival”)
cumhaz.plot + labs(title = “Cumulative hazard”)

6.3.2 muhaz

Cumulative Hazard Plot with R

library(“muhaz”)
library(“survival”)
data(lung)
attach(lung)
fit <­ pehaz(time,status)
plot(fit)
fit <­ muhaz(time,status)
summary(fit)
plot(fit)

6.3.3 Bayesian in Kaplan-meier estimator

The Kaplan-Meier estimator stands as a product-limit method. The poste­
rior distribution of the parameter is estimated through the Bayesian approach.
It is estimated as the probability for the various ordered of censoring and re­
currence time. It is the binomial distribution, for the ith ordered distribution
by
d[i] ∼ binomoial(q[i[, R[i]) (6.21)
Survival Analysis 77

Now q[i] represent the probability of recurrence and R[i] as the number at risk
at the beginning of the ith ordered time. A beta prior for the q[i] is suitable
to estimate as
q[i] ∼ beta(0.01, 0.01) (6.22)
It used to estimate the posterior distribution for the recurrence probability.
Now the recurrence probability is estimated as

P [T > t(i) |T ≥ t(i) ]for i=1,2, .,m (6.23)

The number of ordered time/censored time is represented by m. Now at time

point i, the number of persons at risk is defined as

r[i] = r[i − 1] − c[i − 1] − d[i − 1] (6.24)

for i = 2, 3, ., m, where m is the number of distinct ordered times. Like death,

the censored can also be defined as

c[i] ∼ binomial(qc[i], R[i]) (6.25)

and
qc[i] ∼ beta(0.01, 0.01) (6.26)
Now distribution assumptions about death and censoring jointly specify the
number at risk. Now the task is to generate the posterior distribution of S(t(i) )
as the conditional probabilities

P [T > t(i) |T ≥ t(i) ] (6.27)

to obtain the Kaplan-Meier estimate. Suppose patients treated with “weekly

cisplatin chemotherapy” is defined as arm=0 and “3-weekly cisplatin
chemotherapy” as arm=1. The given below OpenBUGS code is useful to ob­
tain the Kaplan-Meier as
78 Bayesian Approaches in Oncology Using R and OpenBUGS

Bayesian Kaplan-Meier with OpenBUGS

model;
{
for (i in 1:m1){ d1[i] ∼ dbin(q1[i],r1[i])}
for (i in 1:m1){ ce1[i] ∼ dbin(qc1[i],r1[i])}
for (i in 1:m1){ q1[i] ∼ dbeta (.01,.01)}
for (i in 1:m1){ qc1[i] ∼ dbeta (.01,.01)}
for(i in 1:m1){ p[i]<-(r[i]-d[i])/r[i]}
for (i in 1:m1){ p1[i]<-1-q1[i]}
r1[1]<­ 35
for(i in 2:m1){ r1[i]<-r1[i-1]-d1[i-1]-ce1[i-1]}
for (i in 2:m1){ s1[i]<-s1[i-1]× p1[i]}
s1[1]<-p1[1]
for (i in 1:m2){ d2[i]∼ dbin(q2[i],r2[i])}
for (i in 1:m2){ ce2[i]∼ dbin(qc2[i],r2[i])}
for (i in 1:m2){ q2[i]∼ dbeta (.01,.01)}
for (i in 1:m2){ qc2[i]∼ dbeta (.01,.01)}
for(i in 1:m2){ p2[i]<-(r2[i]-d2[i])/r2[i]}
for (i in 1:m2){ p2[i]<-1-q2[i]}
r2[1]<­ 19
for(i in 2:m2){ r2[i]<-r2[i-1]-d2[i-1]-ce2[i-1]}
for (i in 2:m2){ s2[i]<-s2[i-1] × p2[i]}
s2[1]<-p2[1]
}
#call the data #
list(m1 = 9,d1 = c(3,2,1,1,2,1,1,1,1),ce1 = c(1,1,2,0,3,0,5,6,4),
m2 = 10,d2 = c(2,2,1,2,2,4,2,2,1,1),
ce2 = c(0,0,0,0,0,0,0,0,0,0))
#Define the initial data#
list(q1 = c(.5,.5,.5,.5,.5,.5,.5,.5,.5),
q2 = c(.5,.5,.5,.5,.5,.5,.5,.5,.5,.5))

Now we performed the analysis with 20,000 observations as simulation. A

burn-in of 1,000 and a refresh of 100 is considered. The conditional distribution
for both the arm is presented below. The recurrence probability is defined as
S[i] = S(t(i) ) (6.28)
Similarly, the conditional probabilty is defined as
P [i] = P [T > t(i) |T ≥ t(i) ] (6.29)
The link between condition probability and recurrence probability is
S[i] = S[i − 1]P [i] (6.30)
The curve between Ŝ1 (t) and Ŝ2 (t) are useful for survival comparison between
the arms.
Survival Analysis 79

TABLE 6.1: Posterior estimates survival estimates for Arm=1

Parameter Mean SD val2.5pc median val97.5pc

s1[1] 0.91 0.04 0.80 0.92 0.98
s1[2] 0.85 0.05 0.72 0.86 0.94
s1[3] 0.82 0.06 0.68 0.82 0.93
s1[4] 0.79 0.06 0.64 0.79 0.90
s1[5] 0.72 0.07 0.56 0.72 0.86
s1[6] 0.68 0.08 0.51 0.69 0.83
s1[7] 0.64 0.08 0.47 0.65 0.80
s1[8] 0.59 0.09 0.40 0.59 0.76
s1[9] 0.47 0.12 0.22 0.48 0.69
p1[1] 0.91 0.04 0.80 0.92 0.98
p1[2] 0.93 0.04 0.83 0.94 0.99
p1[3] 0.96 0.03 0.87 0.97 0.99
p1[4] 0.96 0.03 0.86 0.97 0.99
p1[5] 0.91 0.05 0.77 0.92 0.98
p1[6] 0.94 0.05 0.81 0.96 0.99
p1[7] 0.94 0.05 0.80 0.95 0.99
p1[8] 0.91 0.07 0.71 0.93 0.99
p1[9] 0.79 0.16 0.39 0.83 0.99

6.3.4 Bayesian Cox proportional hazards model

The Cox proportional hazards model is defined as
�i=p
h(t, X) = h0 (t)exp( βi Xi ) (6.31)
i=1

The regression parameter is defined as Xi . Similarly, the baseline hazard func­

tion is h0 (t). Now covariate is defined as Xi . The baseline hazard function is
presented with time, but the covariates are not the functions of t. Further, at
the time t the hazard function is presented as
limΔt→∞ P (t ≤ T < t + Δt|T ≥ t)
h(t) = (6.32)
Δt
Now the hazard function can be defined with S(t) to obtain the Bayesian
inference
[dS(t)/dt]
h(t) = (6.33)
S(t)
The survival function is
S(t) = P (T > t) (6.34)
The hazard ratio is presented as
h(t, X ∗ )
HR = (6.35)
h(t, X)
80 Bayesian Approaches in Oncology Using R and OpenBUGS

TABLE 6.2: Posterior estimates survival estimates for Arm=2

Parameter Mean SD val2.5pc median val97.5pc

s2[1] 0.89 0.06 0.72 0.90 0.98
s2[2] 0.78 0.09 0.58 0.79 0.93
s2[3] 0.73 0.09 0.52 0.74 0.89
s2[4] 0.63 0.10 0.40 0.63 0.82
s2[5] 0.52 0.11 0.30 0.52 0.74
s2[6] 0.31 0.10 0.13 0.30 0.53
s2[7] 0.20 0.09 0.06 0.19 0.41
s2[8] 0.10 0.06 0.01 0.091 0.27
s2[9] 0.05 0.04 0.00 0.037 0.18
s2[10] 0.00 0.00 0 0 0.00
p2[1] 0.89 0.06 0.72 0.90 0.98
p2[2] 0.88 0.07 0.69 0.89 0.98
p2[3] 0.93 0.06 0.76 0.95 0.99
p2[4] 0.85 0.09 0.63 0.87 0.98
p2[5] 0.83 0.10 0.58 0.85 0.97
p2[6] 0.59 0.14 0.29 0.60 0.86
p2[7] 0.66 0.17 0.28 0.68 0.94
p2[8] 0.50 0.22 0.09 0.50 0.90
p2[9] 0.50 0.28 0.02 0.50 0.97
p2[10] 0.01 0.07 0.00 0 0.08

Now the same covariate X can be separated as X ∗ and X. The term X ∗ shows
the one category of a variable X. Similarly, X defined another category. The
HR is defined as
i=p
�
HR = exp( βi (Xi∗ − Xi )) (6.36)
i=1

Now the Bayesian specify the prior distribution of β. It used to obtain the
hazard ratio. If X = 0 and X ∗ = 1 then hazard ratio formed as

HR(β) = exp(β) (6.37)

The OpenBUGS codes is provided to compute the hazard ratio through

Bayesian approach.

TABLE 6.3: Posterior estimates Hazard Ratio for Arm=2 with respect to
Arm=1
Parameter Mean SD val2.5pc median val97.5pc
HR 4.763 2.645 1.375 4.022 11.52
Survival Analysis 81

Bayesian Cox PH with OpenBUGS

model
{
for(i in 1:N) {
for(j in 1:T) {
Y[i,j] <­ step(obs.t[i]-t[j]+eps)
dN[i, j] <­ Y[i, j] × step(t[j + 1]-obs.t[i]-eps) × fail[i]
}}
for(j in 1:T) {
for(i in 1:N) {
dN[i, j] ∼ dpois(Idt[i, j])
Idt[i, j] <­ Y[i, j] × exp(beta × x1[i]) × dL0[j]
}
dL0[j] ∼ dgamma(mu[j], c)
mu[j] <­ dL0.star[j] × c
S.treat[j] <­ pow(exp(-sum(dL0[1 : j]))
exp(beta × -0.5))
S.placebo[j] <­ pow(exp(-sum(dL0[1 : j]))
exp(beta × 0.5))
}
c <­ 0.001
r <­ 0.1
for (j in 1 : T) {
dL0.star[j] <­ r × (t[j + 1]-t[j])
}
beta ∼ dnorm(0.0,0.000001)
HR<-exp(beta)
}
list(N = 44,T = 16,eps=1.0E-20,obs.t = c(1, 1, 2, 2, 3, 4, 4, 5, 5,
8, 8, 8, 8, 11,11,12,12,12,15,15,17,22,23,6,6, 6, 6, 7, 9, 10,
10,11,13,16,17,19,20,22,23,25,32,32,34,35),
fail=c(1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,
1, 1, 1, 1, 1, 1,1,0, 1, 0, 1, 0, 0, 1, 1, 0,
0, 0, 1, 1, 0, 0, 0, 0, 0),
x1 = c(0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5,
0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5,0.5,0.5, 0.5, 0.5,
0.5, -0.5,-0.5, -0.5,-0.5, -0.5,-0.5,-0.5, -0.5,-0.5,-0.5,
-0.5,-0.5,-0.5,-0.5,-0.5,-0.5,-0.5,-0.5, -0.5, -0.5, -0.5),
t = c(1, 2, 3, 4, 5, 6, 7, 8, 10, 11, 12, 13, 15, 16,
17, 22, 23))
list(beta = 0.0,dL0 = c(1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,
1.0,1.0,1.0,1.0,1.0,1.0,1.0, 1.0))
82 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 6.4: Posterior estimates convergence plot.

Survival Analysis 83

FIGURE 6.5: Bayesian model average.

FIGURE 6.6: Posterior estimate plot of survival function.

84 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 6.7: Posterior estimate plot of cumulative hazard.

FIGURE 6.8: Estimates piecewise exponential hazard function.

Survival Analysis 85

FIGURE 6.9: Posterior estimate plot of cumulative hazard.

FIGURE 6.10: Survival curve comparison.

Chapter 7

Competing Risk Data Analysis

Abstract
Survival analysis data comes with multiple and mutually exclusive events.
The application of competing risk modeling helps to account different causes
of death. There are several techniques to work with competing risk model­
ing to prioritize the specific cause. Several techniques are available to work
with the cause-specific framework, mixture models, sub-distribution hazard,
and method of pseudo-observations. This chapter is dedicated toward differ­
ent methodology to handle cause-specific model. Details review of competing
risk model is available. Theory about competing Risk as Bivariate Random
Variable, Cumulative Incidence Rate, and Cause-specific hazard model are
discussed. The illustration to perform competing risk model with R and Open-
BUGS are elaborated. This chapter will be helpful to perform a competing
risk model in oncology or time-to-event data.

7.1 Introduction
The occurrence of Competing risk is prevalent in oncology. The reason
for death for the treated cancer patients may be different, and it is defined as
‘Competing Risk’. Now, the reason of death may be a local recurrence, distant
metastases, occurrence of a secondary tumor. There are different strategies to
deal with competing risks with survival analysis. The conventional survival
suggests pursuing with cause-specific death modeling. It occurs when there
are at least two possible ways that a person can fail, but only one such fail­
ure type can occur. The situation in which an individual can experience more
than one type of event [14, 15].The failure to achieve independence between
the time to an event and the censoring mechanism [16, 17].The concept of
competing risks as the situation where one type of event ’either precludes the
occurrence of another event under investigation or fundamentally alters the
probability of occurrence of this other event [18].
Suppose, T is a continuous positive random variable provides survival time.

87
88 Bayesian Approaches in Oncology Using R and OpenBUGS

The probability density function (pdf) is defined as f (t). The cumulative dis­
tribution function (cdf) is defined as

F (t) = P r{T ≤ t} (7.1)

Now the survival function of the probability of being alive at t. The survival
function of the probability of being alive at t is presented as.

S(t) = P r{T > t} (7.2)

The hazard function is defined as

f (t)
λ(t) = [ ] (7.3)
S(t)

The cumulative hazard function is defined as

� t
Δ(t) = λ(µ)dµ (7.4)
0

The survival function at time t is presented as

S(t) = exp{−Δ(t)} (7.5)

7.2 Competing Risk as Bivariate Random Variable

Let Censoring variable(C) and Time variable(T).Survival data are usually
presented as a bivariate random variable or pair (C, T ).If C = 1 then the first
member of the pair is defined by T . The term T is time at which the event
occurred.
If C = 0 then T is the time at which the observation was censored. Further,
C can be extended into 0 and i.
Now C = i, i = 1, ...p i represents the type of the first failure/event observed
and T is the time at which the event of type i occurred. Now C = 0, If the
observation is censored C = i, i = 1, ...p otherwise.
For example, if T = i, i = 1, ...p, i is the sepecific event and Ti is the duration
of specific event. T = i, i = 1, 2, 3, i = 1, and T1 is time to local relapse.
i = 2, and T2 is time to distant relapse.
i = 3, and T3 is time to death.
when C = i ,then T = min(T1 , T2 , T3 ) otherwie C = 0.
Competing Risk Data Analysis 89

7.3 Cumulative Incidence Rate

If it confirmed that if C =� 0 then cumulative incidence function (CIF)
required to present. Suppose the total number of observations are defined as
N . However, if among the N observations some did not yet experience any of
the p types of events, then censored observations are present in the dataset.
The CIF for any event of type i(i = 1, 2, ....p) is defined as the joint probability

Fi (t) = P (T ≤ t, C = i) (7.6)

It is the probability that an event of any type occurs at or before time t.

Hence,
p
� p
�
F (t) = P (T ≤ t) = P (T ≤ t, C = i) = Fi (t) (7.7)
i=1 i=1

The probability that an event i does not occur by time t can be denoted as

Si (t) = P (T > t, C = i) (7.8)

When the competing risks are not present the overall distribution function
spans the interval [0, 1].

7.4 Competing Risk Model Using R

A step-by-step procedure used to run R script to generate compting risk
data, competing risk plot is detailed below.
Competing Risk Plot Using Timepoints

library(“cmprsk”)
library(“survival”)
os <­ rexp(100)
death<-sample(1:2,100,replace=TRUE)
arm<­ factor(sample(1:2,100,replace=TRUE), 1:2,c(‘arm1’,‘arm2’))
competigndata<-cuminc(os,death,arm)
plot(competigndata,lty=1,color=1:4)
fit=cuminc(os,death)
fit
90 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 7.1: Armwise cumulative hazard comparison.

#R Output

Estimates and Variances:

$est
1 2 3 4 5
1 1 0.30 0.41 0.43 0.44 0.46
1 2 0.34 0.47 0.52 0.52 0.53

$var
1 2 3 4 5
1 1 0.002135 0.0024856 0.0025274 0.0025562 0.0026351
1 2 0.002282 0.0025567 0.0025940 0.0025940 0.0026136

#The CIF Estimates at 1,2,3 and 4 Yearss is Obtained by Function

Using Timepoints

timepoints(fit,times=c(1,2,3,4))
Competing Risk Data Analysis 91

#R Output

$est
1 2 3 4
1 1 0.30 0.41 0.43 0.44
1 2 0.34 0.47 0.52 0.52

$var
1 2 3 4
1 1 0.00213530 0.002485680 0.002527465 0.002556282
1 2 0.00228206 0.002556745 0.002594095 0.002594095

7.5 Cause-Specific Hazard Model

Survival analysis data comes with multiple and mutually exclusive events.
The application of competing risk modeling helps to account different causes
of death. There are several techniques to work with competing risk modeling
to prioritize the specific cause. There are several techniques to work cause-
specific framework [14], mixture models [19], sub-distribution hazard [20], and
method of pseudo-observations [21]. Details review of competing risk model is
available [22]. Other techniques like machine learning through random forest
method [23] are widely adopted in competing risk modeling [23, 24]. Suppose
there are K coupled. The cumulative incidence functions Fk (t). First-derivate
can be presented as
K
dFk (t) �
= (1 − Fk� (t))λk (t) (7.9)
dt �
k =1
Now λk (t)(≥ 0, ∀t > 0) shows that kth cause-specific hazard function. The
equation can be reformulated as ∀k:
K K K
� dFk (t) � �
= (1 − Fk� (t)) λk (t) (7.10)
dt
k=1 k� =1 k=1
dE(t) �
=− λk (t) (7.11)
E(t)
k
Now the event-free probability function is presented as E(t),
K
�
E(t) = 1 − Fk (t) (7.12)
k=1

It can be obtained as
K
�
E(t) = Sk (t) (7.13)
k=1
92 Bayesian Approaches in Oncology Using R and OpenBUGS

If Sk (t) stands with unadjusted survival function. let the causes are k. It is
defined as � t
� �
Sk (t) = exp(− λk (t )dt ) (7.14)
t� =0
It can be obtained as
0 ≤ Sk (t) ≤ 1, ∀k = 1, ..., K, t ≥ 0 (7.15)
It is obtained as Sk (0) = 1. The one-dimensional integration can be obtained
as
� t K
� � � �
Fk (t) = ( Sk (t ))λk (t )dt (7.16)
t� =0
k� =1
It can be reformed as Weibull survival model as
λk (t) = αk γk tαk −1 (7.17)

Sk (t) = exp(−γk tαk ) (7.18)

It leads to cumulative incidence functions(CIF) with two competing risks (K =
2).
� t
α1 +µα2
F1 (t) = −α1 γ1 µα1 −1 eγ1 µ (7.19)
0
Bayesian inference with Markov chain Monte Carlo (MCMC) useful to obtain
the inference.

7.6 Bayesian Information Criteria

Different models can be formulated and thereafter models can be compared
with Bayesian information criterion (BIC) [25]. It is an evaluation criterion
in terms of the posterior probability [26]. Suppose there are M1 , M2 , ..., Mr
types of model and assigned as r candidate models. Let the parametric dis­
tribution is defined as fi (x|θi )(θi ∈ Θi ∈ Rki ) The prior distribution πi (θi )
Let M1 , M2 , ..., Mr be r candidate models, and assume that each model of
the ki -dimensional parameter vector θi . Now the number of observations are
xn = {x1 , ..., xn }. The probability of xn is defined as
�
pi (xn ) = fi (xn |θi )πi (θi )dθi (7.20)

The posterior probability of the ith can be defined as P (Mi ).

pj (xn )P (Mj )
P (Mi |xn ) = �r , i = 1, 2, ...., r (7.21)
j=1 pj (xn )P (Mj )
Competing Risk Data Analysis 93

Posterior probability represents ith model when xn are observed. The model is
selected from r models. The process is to consider the model that comes with
the highest probability. The model maximize the numerator pi (xn )P (Mi ) is
defined as best model.
It is assumed that the prior probability P (Mi ) are similar in all the models.
It maximizes the marginal likelihood through pi (xn ) among the data selected.
P (M1 |xn ) p1 (xn ) P (M1 )
= (7.22)
P (M2 |xn ) p2 (xn ) P (M2 )
The Bayes factor is defined as
p1 (xn ) f1 (xn |θ1 )π1 (θ1 )dθ1
B12 = = (7.23)
p2 (xn ) f1 (xn |θ1 )π2 (θ2 )dθ2

Fine and Gray Test Using R

data(“mgus2”)
head(mgus2)
etime <­ with(mgus2, ifelse(pstat==0, futime, ptime))
event <­ with(mgus2, ifelse(pstat==0, 2× death, 1))
event <­ factor(event, 0:2, labels=c(‘censor’, ‘pcm’, ‘death’))
pdata <­ finegray(Surv(etime, event) ∼ ., data=mgus2)
fgfit <­ coxph(Surv(fgstart, fgstop, fgstatus) ∼ age,weight=fgwt,
data=pdata)
summary(fgfit)

#R Output

Call:
coxph(formula = Surv(fgstart, fgstop, fgstatus) ~ age,
data = pdata,weights = fgwt)

n= 41775, number of events= 115

coef exp(coef) se(coef) z Pr(>|z|)

age -0.016617 0.983521 0.007008 -2.371 0.0177 *
---
Signif.codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
exp(coef) exp(-coef) lower .95 upper .95
age 0.9835 1.017 0.9701 0.9971

Concordance= 0.539 (se = 0.024 )

Likelihood ratio test= 5.35 on 1 df, p=0.02
Wald test = 5.62 on 1 df, p=0.02
Score (logrank) test = 5.63 on 1 df, p=0.02
94 Bayesian Approaches in Oncology Using R and OpenBUGS

7.7 Illustration Using R

Illustration Using R

library(“survival”)
library(“BMA”)
os <­ rexp(100)
death<-sample(1:2,100,replace=TRUE)
arm<-rbinom(100, 1,.5)
age<-rnorm(100,50,10)
diagtime<-rnorm(100,5,1)
cdata<-data.frame(os,death,arm,age,diagtime)
test.bic.surv<­ bic.surv(Surv(os,death) ∼., data =cdata)
summary(test.bic.surv, conditional=FALSE, digits=2)
imageplot.bma(test.bic.surv)

#R Output

Call:
bic.surv.formula(f = Surv(os, death) ~ .,
data = cdata)

Six models were selected

Best 5 models (cumulative posterior probability=0.96):

p!=0 EV SD model 1 model 2

arm 12.9 -0.0243 0.120 . .
age 36.1 -0.0072 0.012 . -0.019
diagtime 11.9 0.0101 0.065 . .

nVar 0 1
BIC 0.000 1.179
post prob 0.484 0.268

p!=0 EV SD model 3 model 4 model 5

arm 12.9 -0.0243 0.120 . -0.125 -0.269
age 36.1 -0.0072 0.012 . . -0.023
diagtime 11.9 0.0101 0.065 0.113 . .

nVar 1 1 2
BIC 3.541 3.808 4.290
post prob 0.082 0.072 0.057
Competing Risk Data Analysis 95

FIGURE 7.2: Comparison of Bayesian Moving average.

Chapter 8

Frailty Data Analysis

Abstract
Disease progression by recurrence, metastases are common in cancer. How­
ever, disease progression is dependent on individual-level effect. Several un­
known factors are contributing to disease progression. Survival analysis of
disease progression is useful. It makes it essential to accommodate several
contributing factors on disease progression by frailty model. Frailty model is
a branch of survival analysis.
This chapter is dedicated to frailty modelling. There are several techniques
to work with frailty data to make survival analysis robust. Different method­
ology to handle frailty data analysis is presented. Data analysis with frailty
model is prepared with R. Further, Bayesian inference of frailty modelling
is presented. There are several techniques to work with frailty data to make
survival analysis robust. This chapter is dedicated to different methodology
to handle frailty data analysis. Details review of frailty data analysis is avail­
able. Theory about frailty data analysis is discussed. Example to handle frailty
data analysis is illustrated with R. The Bayesian counterpart of Frailty data
analysis also explained and presented by examples.

8.1 Introduction
A cancer patient may be observed with several events after or during
the first treatment. For example, the patient opted for the head, and cancer
may appear with disease recurrence or metastases disease. Cancer patients
may experience recurrent even in the superficial tumor. The entire chance of
recurrence is dependent on patient-level influence or can be defined as the
individual-level effect. There could be several unknown factors that are con­
tributing to disease appearance. The data handling technique on the recurrent
event is conventionally known as survival analysis. However, the inclusion of
several factors contribution at the individual level presents the extension of
conventional survival model as a frailty model. One patient is defined as frail

97
98 Bayesian Approaches in Oncology Using R and OpenBUGS

if he is having a higher/lower of exposure. Due to this, he is having a differ­

ent disease relapse rate than others. Comprehensively, a different individual
has different disease relapse rate. It introduces the term frailty. The frailty
term is utilized as different individuals are at different risks even though it
is assumed that they are at the same risk at treatment initiation. It becomes
interesting while the appearance of other factors like age, sex, district are con­
sidered. Frailty term is used to serve as an unobservable random effect shared
by subjects having similar risk in the analysis. Inclusion of random effect part
help to describe the excess risk of mortality rate. A frailty data analysis is an
extension of the survival model. The random effect component is attached to
the conventional survival data model. However, conventional survival model
always considered the random effect component. One extended random effect
component is joined in the survival model and defined as a frailty model. The
random effect can be utilized in the univariate data to make it more flexible. It
becomes attractive in multivariate modeling. In oncology, multiple recurrent
events may occur. The application of frailty model becomes attractive in case
of multiple recurrent event modeling. Suppose the hazard function is defined
as
Y µ(t) (8.1)
The term Y is unknown parameter and hazard function is µ(t). Y is useful
toward random effect modeling. There are several unknown parameters, and
those can be pushed as random components. It helps to make unknown values
by integrating the random component. Now random effects may be reflected as
the effect of unobserved covariates. Now the term K can be assumed as Poisson
distribution with mean Y k. It can be conditioned with Y having parameter k.
It can be formulated as Y with gamma distribution. It can be integrated out
by the relevant time period. Survival data can not be avoided in oncology. It is
the most common dataset. Medical oncology treated patients are followed for
the primary outcome of a study as death or relapse. However, the censoring
of the event is natural phenomena. Censoring becomes right while the time
of interest is not measured, but the lower bound is observed. Handling right
censoring data is different and challenging [27]. Widely used method for sur­
vival data is proportional hazard model. However, generated inference from
the Cox proportional hazard model is applicable as identically independently
distributed individuals. Perhaps, individuals may be exposed to different risk,
even after maintaining other risk factors. Sometimes, patients may belong to
different cluster or regions. It may be assumed that the individual belongs
to the same cluster behave similarly. The intention of the multicentric clini­
cal trial is about to prioritize the cluster-specific inference. The frailty model
helps to account the heterogeneity at the baseline level. It is an extension of
the proportional hazard model, while hazard function is dependent on un­
observable random quantity. It is called as frailty. Distributional form of the
frailty model can be assumed-the distribution of the frailties considered as
gamma frailty. Gamma frailty model is widely used due to mathematical ac­
ceptability. The linear model is the extention of the generalized linear mixed
Frailty Data Analysis 99

effect model by lognormal frailty [28]. There are different multivariate forma­
tion by frailty data distribution [29]. The relation between baseline to time
to event plays a crucial role. Different distribution formation is required to
approach different frailty distribution [30]. Generally, the gamma distribution
is a widely accepted choice for frailty distribution. The widely used R package
is survival. Function available coxph() in survival package is useful to handle
frailties as a semi-parametric model. Similarly, frailtypack package [31] pro­
vides the gamma frailty models as a semi-parametric estimation. The Weibull
is the parametric choice. It can also be used in frailtypack package. The coxme
[32] and phmm [33] functions also work in a similar direction with lognormal
frailty model. But a large choice of parametric distributions is available in
parfm [34], an R package. The gamma, inverse Gaussian, lognormal, Weibull,
Gompertz, log-logistic are available in parfm. The parametric distributional
estimation is performed by marginal log-likelihood maximization. Frailty dis­
tribution data is dependent with individual level variation.

8.2 Frailty Model

The frailty model is defined in terms of the conditional hazard Suppose
the hazard function is assumed as

hij (t|ui ) = h0 (t)ui exp(xTij β) (8.2)

with i ∈ I = {1, ....., G} and j ∈ Ji = {1, ....., ni }. Now the h0 (.) is

considered as the baseline hazard function. The term ui is assumed as frailty
term for the covariates i, xij . The regression coefficient is defined as β. It is
possible to obatain the univariate frailty model for ni subjects [35]. Otherwise
it is defindd as shared frailty model [28, 29]. Now the baseline hazard can
be defined through the parametric approach. The term Ψ is formulated with
the regression coefficient the parfm package available in Weibull, exponential,
lognormal and log-logistic to serve the parametric distribution.

8.3 Motivating Example

Recently, a total of 50,381 patients data of multidisciplinary hepatocellu­
lar carcinoma (HCC) is analyzed through frailty modeling [36]. The type of
hospital is considered in this illustration. The relation between the overall risk
of death and type of hospital is figured through frailty. They used the mul­
tivariable Cox shared frailty modeling. Similarly, to explore the distance to
100 Bayesian Approaches in Oncology Using R and OpenBUGS

travel to get treatment among head and neck cancer patients are established
by Cox shared frailty model [37]. The overall survival(OS) was evaluated.
#Example1

time cens xcoord ycoord age sex wbc tpi district

24 1 1 0.4123484 0.4233738 44 1 281.0 4.87 1
62 3 1 0.3925028 0.4531422 72 1 0.0 7.10 1
68 4 1 0.4167585 0.4520397 68 0 0.0 5.12 1
128 9 1 0.4244763 0.4123484 61 1 0.0 2.90 1
129 9 1 0.4145535 0.4520397 26 1 0.0 6.72 1
163 15 1 0.4013230 0.4785006 67 1 27.9 1.50 1

# Output1

Generalized accelerated failure time frailty model:

Call:
frailtyGAFT(formula = Surv(time, cens) ~ age + sex
+wbc+tpi+baseline(age,sex,wbc,tpi)+frailtyprior("car",
district), data = d, mcmc = mcmc, prior = prior,
Proximity = E)

Posterior means for regression coefficients:

Mean Median Std 95%HPD-Low 95%HPD-Upp
Intercept 8.58 8.60 0.28 7.98 9.14
age -0.05 -0.05 0.00 -0.05 -0.04
sex -0.26 -0.28 0.16 -0.53 0.06
wbc -0.00 -0.00 0.00 -0.00 -0.00
tpi -0.06 -0.06 0.01 -0.09 -0.02
Bayes factors for LDTFP covariate effects:
intercept age sex wbc tpi overall normality
130.95 11.62 1.26 14.01 0.48 7.88 994.86
Frailty Data Analysis 101

#Kidney Data Fit with Gamma Frailty Distribution

library(“parfm”)
head(kidney)
kidney$ sex <­ kidney$sex - 1
mod <­ parfm(Surv(time, status) ∼ sex + age, cluster=‘id’,
data=kidney, dist=‘exponential’, frailty=‘gamma’)
mod
ci.parfm(mod, level=0.05)[‘sex’,]
u <­ predict(mod)
plot(u, sort=‘i’)

#R Output on Gamma Distribution Fit

Frailty distribution: gamma

Baseline hazard distribution: Exponential
Loglikelihood: -333.248

ESTIMATE SE p-val
theta 0.301 0.156
lambda 0.025 0.014
sex -1.485 0.396 <.001 ***
age 0.005 0.011 0.657
---
Signif.codes:0 '***' 0.001 '**' 0.01'*' 0.05 '.' 0.1 ' '1

Kendall's Tau: 0.131

# R Output Confidence Interval Fit

low up
0.104 0.492
102 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 8.1: Gamma frailty model prediction.

#Kidney Data Fit with Different Distribution

kidney.parfm <­ select.parfm(Surv(time, status) ∼ sex + age,

cluster=‘id’, data=kidney,dist=c(‘exponential’,‘weibull’,‘gompertz’,
‘loglogistic’,‘lognormal’),frailty=c(‘gamma’,‘ingau’,‘possta’))
kidney.parfm
plot(kidney.parfm)
Frailty Data Analysis 103

# R Output with Parametric Frailty Models

### - Parametric frailty models - ###

Progress status:
'ok' = converged
'nc' = not converged

Frailty
Baseline gamma invGau posSta
exponential.........ok......ok......ok....
Weibull.............ok......ok......ok....
Gompertz............ok......nc......ok....
loglogistic.........ok......ok......ok....
lognormal...........ok......ok......ok....

FIGURE 8.2: Estimates through different distribution.

104 Bayesian Approaches in Oncology Using R and OpenBUGS

8.4 Bayesian in Frailty Survival

The Bayesian counterpart of frailty survival model becomes interesting for
researchers. Bayesian is applicable in different clinical; trial study design. Sim­
ilarly, it is applicable to survival analysis. Bayesian works with prior distribu­
tion, design, monitoring, and data analysis. Perhaps, there is some limitation
about the application of Bayesian toward prior distribution selection.

8.4.1 Frailty modeling

Frailty model can be adopted in any distribution if Y ≥ 0. The univariate
form is mentioned in equation 8.1. The parametric expression can be defined
as µ(t). It helps to make the regression model with the Cox proportional model
as �
Y exp(β z)µ0 (t) (8.3)
The covariate is presented as z. The regression coefficient is β. If z = 0 then the
hazard function becomes to µ0 (t). It is possible to make µ0 (t) as parametric
or nonparametric. If covariate is absent then the conditional survival function
becomes
S(t|Y ) = exp(−Y M (t)) (8.4)
�t
The integrated conditional hazard is formulated as M (t) = 0 µ(u)dµ. The
term Y may become independent and not observed. In this context, it may
be integrated out as expression given below.
� ∞
S(t) = (−yM (T ))g(y)dy (8.5)
0
It is only possible while the frailty is continuous distribution.

8.4.2 Frailty on recurrent events

The frailty is assumed as a random effect for the subject. No individual
effect is considered in this context. The conditional Poisson model is consid­
ered. It provides scope about any time point about the hazard of a subject
about the appearance of a new event. The Poisson distribution is measured
by Y µ(t). This model is defined as the time since the start of the process.
Suppose the entire interval is 0 to t and the total number of events is K. The
mean is Y M (t). The expression of probability term becomes

Pr(K = k) = (−1)k {M (t)}k L(k) (M (t))/k! (8.6)

Frailty Data Analysis 105

8.4.3 Generalized accelerated failure time (GAFT) frailty

model
The accelerated failure time(AFT) model can be easily extended with
frailty model due to the linear formation. Suppose the baseline survival func­
tion is defined as S0 (t) with q-dimensional vector zij . It is usually a subset of
xij . The model can be formulated with GAFT as
T
SXij (t) = S0,zij (e−xij β−vi t ) (8.7)

or equivalently,
yij = log(tij ) = xTij β + vi + �ij (8.8)
Now xij = (1, xTij )T provides the intercept value. The corresponding vector is
formulated as β = (β0 , β T )T . The heteroscedastic error can be estimated as
�ij . It is independent of vi , and P (eβ0 +�ij > t|zij ) = S0,zij (t). It is defined as

�ij |Gzij ∼ Gzij (8.9)

Now the term Gz is a probability measure defined on R for every z ∈ X. It is

defined as a model for the entire collection of probability as GX = {Gz : z ∈
X}. It allows us about smoothly change with the covariates z.
It can be formulated as

β ∼ Np+1 (m0 , S0 )

Gz |α, σ 2 ∼ LDTFPL (α, σ 2 ), α ∼ Γ(a0 , b0 ), σ −2 ∼ Γ(aσ , aσ )

8.4.4 Illustration with leukemia data using R

A total of seven R packages are required to work Bayesian frailty model in
R. Packages are“coda”,“survival”,“spBayesSurv”,“fields”,“BayesX”,“R2BayesX”
and “spBayesSurv” respectively. and
#Leukemia Data Using R

library(“coda”)
library(“survival”)
library(“spBayesSurv”)
library(“fields”)
library(“BayesX”)
library(“R2BayesX”)
library(“spBayesSurv”)
106 Bayesian Approaches in Oncology Using R and OpenBUGS

A total of n=1,043 patients leukemia data is considered. The dataset is

available as LeukSurv data in “spBayesSurv” package. This illustration is
about exploring the subject-specific variation after adjusting the factors like
sex, white blood cell, Townsend score (TPI) and sex.
Now patients have belonged to a different district. It may be possible
that subjective variation has occurred due to a native of the patients. In this
context, the geostatistical and area wise models are fitted. Dataset is presented
as
#Leukemia Data Illustration Using R

data(“LeukSurv”)
d <- LeukSurv[order(LeukSurv$district), ] head(d)

The outcome looks like

# R Output to Check Dataset

time cens xcoord ycoord age sex wbc tpi district

24 1 1 0.4123484 0.4233738 44 1 281.0 4.87 1
62 3 1 0.3925028 0.4531422 72 1 0.0 7.10 1
68 4 1 0.4167585 0.4520397 68 0 0.0 5.12 1
128 9 1 0.4244763 0.4123484 61 1 0.0 2.90 1
129 9 1 0.4145535 0.4520397 26 1 0.0 6.72 1
163 15 1 0.4013230 0.4785006 67 1 27.9 1.50 1

Now another dataset“newngland”is called that is available in spBayesSurv.

#Newngland Data Illustration Using R

Newngland <­ read.bnd(system.file(“otherdata/newngland.bnd”,

+package = “spBayesSurv”))
adj.mat <­ bnd2gra(Newngland)
E <­ diag(diag(adj.mat)) - as.matrix(adj.mat)

Thereafter district are sorted from the LeukSurv dataset. Finally, the
adjacency matrix E is obtained as Next task is to assign a function to
run the MCMC algorithm. A total of 5,000 burn-in scans are planned with
nburn=5000. A final chain of 2,000 is fixed as nsave=2000. The thinning in­
terval to run the MCMC is defined by nskip=4. A total of scans to be saved
is defined as nsave. The predefined number of scans to be displayed on the
screen is called by ndisplay.
Frailty Data Analysis 107

#Simulation Size Specification

set.seed(1)
mcmc <- list(nburn = 5000, nsave = 2000, nskip = 4, ndisplay = 1000)

The prior information is assigned as prior. It is performed as transformed

Bernstein polynomial (TBP) prior [38].
#Prior Information Specification

prior <­ list(maxL = 4, a0 = 5, b0 = 1)

ptm <­ proc.time()

Finally, the function is used to fit a generalized accelerated failure time

frailty model. The function is frailtyGAFT available in spBayesSurv [39]. This
function is useful for frailty modeling on cluster level time-to-event data. The
outcome is assigned to res1. The variables like age, sex, wbc, and tpi are
considered as covariates on time-to-event data modeling.
#Model Specification

res1 <­ frailtyGAFT(formula = Surv(time, cens) ∼

age+sex+wbc+tpi+
baseline(age, sex, wbc, tpi) + frailtyprior(“car” district),
data = d, mcmc = mcmc, prior = prior, Proximity = E)
(sfit1 <­ summary(res1))
frailtyGAFT(formula = Surv(time, cens) ∼ age + sex + wbc + tpi +
baseline(age, sex, wbc, tpi) + frailtyprior(“car”, district),
data = d, mcmc = mcmc, prior = prior, Proximity = E)
proc.time() - ptm

The posterior estimates for age, sex, WBC and tpi are obtained as 8.58,
−0.05, −0.26, −0.00 and −0.06 respectively. The estimates obtained through
Bayesian is presented with posterior estimates.
108 Bayesian Approaches in Oncology Using R and OpenBUGS

#R Output

Generalized accelerated failure time frailty model:

Call:
frailtyGAFT(formula = Surv(time, cens) ~ age + sex
+wbc+tpi+baseline(age,sex,wbc,tpi)+frailtyprior("car",
district), data = d, mcmc = mcmc, prior = prior,
Proximity = E)

Posterior means for regression coefficients:

Mean Median Std 95%HPD-Low 95%HPD-Upp
Intercept 8.58 8.60 0.28 7.98 9.14
age -0.05 -0.05 0.00 -0.05 -0.04
sex -0.26 -0.28 0.16 -0.53 0.06
wbc -0.00 -0.00 0.00 -0.00 -0.00
tpi -0.06 -0.06 0.01 -0.09 -0.02
Bayes factors for LDTFP covariate effects:
intercept age sex wbc tpi overall normality
130.95 11.62 1.26 14.01 0.48 7.88 994.86
Frailty Data Analysis 109

The traceplot for age covariates can be generated through

#Model Diagnostics

traceplot(mcmc(res1beta[1, ]))

FIGURE 8.3: Trace plot of estimates.

Chapter 9

Relative Survival Analysis

Abstract
Relative survival is helpful to understand the more deaths due to cancer
in comparison to the general population. It helps to determine the deaths due
to cancer by eliminating other causes of mortality. However, it requires to
define the causes of death. Even cancer patients may die due to other causes.
The cancer registry data always influenced by other causes of mortality. It is
required to define the causes of death rigorously. The application of relative
survival is minimal only on cancer have low survival rates. Now the relative
survival analysis by Bayesian counterpart is presented in the chapter. Cancer
clinical trial data example is shown in this chapter. The definition of piece­
wise hazard function is presented. Further, the step to compute the piecewise
hazard value is performed. This chapter will help to provide information that
is difficult to address by conventional analysis. The computation of relative
survival is presented by OpenBUG software, as a Bayesian counterpart. The
methodology of relative survival by the multiplicative and additive model is
presented.

9.1 Introduction
Survival analysis is the only tool to understand cancer disease severity. It
is very much required to understand different progress in cancer care by apply­
ing the survival analysis. Now the cancer survival can be presented in different
forms. Now the net cancer survival separates the impact of a cancer diagnosis
on different survival, and it is useful statistics to explain the cancer prognosis.
Net cancer survival defines the probability of surviving a cancer diagnosis in
the absence of competing causes of death. Net survival is most often used by
two technique, i.e., competing for risk and relative survival. Now, in this chap­
ter, we will show the work with relative survival. The application of relative
survival becomes useful in the National Cancer Institute’s Surveillance, Epi­
demiology and End Results (SEER) the program. Relative survival estimates
the percent of persons surviving by death adjusted for expected deaths by life
table data. Now relative survival estimates the life expectancies for the US

111
112 Bayesian Approaches in Oncology Using R and OpenBUGS

populations based on current age. In contrast, the expected survival of a pop­

ulation does not give us accurate information about the expected survival of a
population of patients with a cancer diagnosis. Now the life tables cannot be
generalized to all populations, and relative survival analysis is an alternative.
However, it is required to have a reference population to perform relative sur­
vival. However, while the reference population is not available, the competing
risk analysis is possible to apply. Relative survival used to compare the death
of cancer patients in comparison to the general population. The documenta­
tion about the cause of death among cancer patients plays an important role.
Thereafter, the comparison of the patients’ death due to cancer is compared
with the population. This helps to consider population-based studies where
a large number of patients are followed through a longitudinal study. It is
also important for clinical trials where more clinical information is present.
This is a standard methodology in registry data analysis for cancer patients.
The approach is about coupling the measured death information with general
mortality data. It provides to obtain the disease-specific hazard. Perhaps, the
practical application of relative survival analysis is relatively complex. Be­
cause the hazard rate also changes for different age and time. Commonly the
population-based studies involved a large number of population with long-term
follow-up. However, it is difficult to capture all types of clinical information
for population-based long-term follow-up studies. It is difficult to capture all
cause of death for a long-time follow-up study. The information about relapse,
remission and progression are difficult to obtain in population-based studies.
The cause of death is often unreliable or unavailable. Suppose a bone metas­
tasis has occurred to a lung cancer patient. The cause of death is documented
as bone cancer. The primary cancer is lung, and cause of death should be lung
cancer. Sometimes cancer patients die due to chemotherapy-related severe tox­
icity. It is difficult to find the cause of death as specific cancer or drug toxicity.
Even the actual information about the cause of death is present. It is often
challenging to classify the cause of death due to cancer or not, the widely ac­
cepted approach to classifying the cause of death as due to cancer or not. The
application of relative survival is not new [7]. It is used a long time in cancer
registry data. The widely accepted approach is non-parametric survival curve
comparison [40]. Recently, methodological extensions are carried in terms of
the estimator [41, 42]. The net survival becomes an attractive choice as non­
parametric estimator [43]. The detailed about the non-parametric procedure
is well-documented [44].
Relative Survival Analysis 113

9.2 Relative Survival Analysis

#Definition of Relative Survival

The relative survival is defined as

SO (t)
SR (t) = (9.1)
SP (t)

This ratio helps to make a comparison of patients from the general popula­
tion. Survival of the patients becomes very poor if the ratio is lesser than one.
It is not obvious that the survival curve will be monotonously increasing [45].
The ratio can be used to compare the survival measures between two cohorts
adjusted by different demographic variable. It is assumed that the overall haz­
ard of each cancer patients can be defined as λOi . It can be split into two parts
as a hazard due to cancer(λEi ) and hazard present in the general population
(λP i ). There is a gap between methodological development and practical ap­
plication. Some methodological development is required on ad-hoc changes in
the existing function. The relative survival analysis can be performed in R.
The “resurv” is suitable enough to perform the different relative survival anal­
ysis. This function is useful toward importing the population tables through
regression analysis. Futher, the individual level survival can be defined as
� t �t
exp{− 0 λOi (u)du}
SEi (t) = exp{− λEi (u)du} = �t (9.2)
0 exp{− 0 λPi (u)du}

Suppose a cohort size is N . The marginal relative survival ratio can be defined
as
N
1 �
SE (t) = SE (t) (9.3)
N i=1 i
Now the net survival is defined as
#Definition of Net Survival

1
�N 1
�N
N i=1 SOi (t) N i=1 SOi (t)
SR (t) = 1
�N ; SE (t) = 1
�N (9.4)
N i=1 SPi (t) N i=1 SPi (t)
114 Bayesian Approaches in Oncology Using R and OpenBUGS

9.3 Data Methodology

Data of this illustrated work is obtained from Tata Memorial Hospital
among patients treated between January 2012 and December 2015. Patients
having Stage III and Stage IV treated with TKI were included in this study.
The proven pathological disease like ECOG, PS indicator a 3-4 and previously
untreated patients were considered in this study. A total of 630 patients (396
males, 234 females), with ages ranging from 2 to 87 (mean 60), diagnosed
between 2011 and 2015 and followed until 2017 were considered in this work.
The data consist of times in days from date of registration to death. Predic­
tor variables include the Stage of the disease at diagnosis (1-4), the year of
diagnosis (Yr), and the age at diagnosis (Age). The predictor variables are
EGFR, Age, Stage, Diabetes and metastatic present or not. We are interested
in comparing the survival between EGFR positive and negative.

FIGURE 9.1: Survival comparison between EGFR postive and EGFR neg­
ative.
Relative Survival Analysis 115

FIGURE 9.2: Cumulative survial curve between EGFR postive and EGFR
negative.

9.4 Illustration with R

The individual expected survival of a cohort could be obtained through Rel­

ative Survival. The individual expected survival is required to make toward
comparison on age and sex composition. The R package relsurv is useful to
obtain relative survival based on population table. The population contains
expected death rates from the calendar year, age, and sex.
#Relaive Survival by Survexp Function

library(“survival”)
haz <­ survexp(OS∼1,data=mydata, rmap = list(year=year,
age=age1), method=‘individual.h’)

This function is useful to obtain each individual’s survival experience. For

each individual, the total hazard experienced up to their observed death time
or last follow-up time. This probability is used as a rescaled time value in the
models.
116 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 9.3: Survial comparison between EGFR postive and EGFR nega­
tive.

#Relaive Survival Using R

glm(mydata $ death ∼ 1 + offset(log(haz)), family=poisson)

glm(mydata$ death ∼mydata$EGFR+offset(log(haz)),family=poisson)

In the first model, a statistical test is performed. The intercept is consid­

ered as 0. It defined with the log-rank test of whether the observed group of
the subject has equal survival to the baseline population or not. The second
model is considered for an effect of variable x after adjusting with reference
to sex and age. Different variables are combined in the rate table. It is useful
to have a different dataset, to deal with the ramp. The rate table served to
compute survexp.us and thereafter expected survival with reference to age and
sex. The cohort group comparison is performed by the overall survival curve.
The visual conclusion can be obtained through the survival curve. Survival
curves can be compared through the “exact method”.
Relative Survival Analysis 117

#Relative Survival Analysis

library(“survival”)
haz <­ survexp(OS∼1, data=mydata,
rmap = list(year=year, age=age1),method=“individual.h”)
glm(mydata$death∼1 + offset(log(haz)), family=poisson)
glm(mydata$death ∼mydata$EGFR + offset(log(haz)),
family=poisson)
pfit<-coxph(Surv(OS,death>0) EGFR+Stage+Diabetes+
metapresent,data=mydata)
plot(survfit(Surv(OS,death>0)∼EGFR, data=mydata),
ylab=“Survival probability”,xlab=“Survival in Days”,
main=“Relative Surival of having EGFR positive”,
mark.time=FALSE)
lines(survexp(∼EGFR, ratetable=pfit, data=mydata),
col=‘purple’)

# R output with Realtive Survival Analysis(Model 1)

Call:glm(formula = mydata$death~1+offset(log(haz)),
family = poisson)

Coefficients:
(Intercept)
3.474

Degrees of Freedom: 629 Total (i.e. Null); 629 Residual

Null Deviance: 453.9
Residual Deviance: 453.9 AIC: 1086

# R Output with Realtive Survival Analysis(Model 2)

Call:glm(formula=mydata$death~mydata$EGFR+offset(log(haz)),
family = poisson)

Coefficients:
(Intercept) mydata$EGFR
3.49530 -0.04192

Degrees of Freedom: 629 Total (i.e. Null); 628 Residual

Null Deviance: 453.9
Residual Deviance: 453.8 AIC: 1088
118 Bayesian Approaches in Oncology Using R and OpenBUGS

9.5 Piecewise Hazard Function

The survival quantification can be carried by the cumulative risk of sur­
vival. It is reliable toward treatment effect comparison [46, 47]. The goal is
to quantify the survival benefit in a specific time interval. Suppose the entire
time interval is [0, τk ]. Thereafter it canbe split into τ1 −τ0 , τ2 −τ1 , ..., τk −τk−1
and 0 = τ0 < τ1 < ........ < τk . Similarly, the hazard is formulated as
0 = g0 < g1 < ........ < gk and g0 = 1. Piecewise constant function is
k
�
h(t) = h0 τl Il (t) with Il (t) = 1 if τl ≤ t ≤ τl+1 , otherwise Il (t) = 0 (9.5)
l=0

Survival function can be defined as

S(t) = exp(−H(t)) (9.6)

Cumuative hazard function is

� t m
� � t
H(t) = h(s)ds = h0 gl Il (s)ds (9.7)
0 l=0 0

9.6 Piecewise Hazard testing

Since we have several hazard functions value at a different point, it is useful
to compare different hazard, and it comes under multiple testing problem. Sup­
pose the term X1 , .....Xn gives independently identically distributed survival
times, and it is defined with C1 , .....Cn censoring time. It is assumed that as
independently distributed of X. Pair of observation are (Ti , δi ), i = 1, 2, .....n,
with minimal time point Ti = min(Xi , Ci ) and i = 1. If δi = 1 ifXi ≤ Ci and
zero otherwise. Now it is presented as 0 < τ0 < τ1 < ..... < τk = ∞. Let us
assume that a total of change point is k, and the number of change points in
the model is αj . The value of the hazard function between the time point is
τj−1 and τj .
The log-likelihood function is presented as
k+1
�
logL(α1 , ...., αk+1 , τ1 , ....τk ) = [X(τj ) − X(τj−1 )] log αj −
j=1
(9.8)
n k+1
� �
αj (Ti Λτj − Ti Λτj−1 )
i=1 j=1
Relative Survival Analysis 119
�n
Now X(t) = i=1 I(T ≤ t, δi = 1) is presented as the number of death
measured till time t. It is possible to make τj , j = 1, ...k as fixed and maximize
it as τj , j = 1, ...k + 1. The function can be formulated as

(X(τj ) − X(τj−1 ))
τj = �n (9.9)
(T
i=1 i Λτ j − Ti Λτj−1 )I(T > τj−1 )

The likelihood for τj is expressed as

k+1
� (X (τj ) − X (τj −1 ))
l(τ1 , τ2 , ...., τk ) = {X(τj −X(τj−1 )}log �n
j=1 i=1 i τj − Ti Λτj −1 )I (T > τj−1 )
(T Λ
(9.10)
It is maximize with l(τ1 , τ2 , ...., τk ) with respect to τj .j = 1, ....., k and
insert the obtained values back to τ̂j , j = 1, ....., k + 1 for MLEs of αj . Now
it is required to test the changes of τj . The hypothesis can be performed as
H0 : τj1 − τj2 = 0. Now different factors can be considered to prepare τj . Now
τj1 and τj2 are independent in nature.
It is explored that τj1 and τj2 are independent in nature [48]. The Wald-test
is performed with
(τ̂k−1 − τ̂k )2
Xw = (9.11)
var(τ̂k−1 − τ̂k )
It can be formulated with chi-quare test statistics with one degree of freedom
under null hypothesis. The R function is defined as WaldTest(), to perform
the test statistics.
#R Code to Run Relative Survival

WaldTest= function (L)

{
WaldTest = numeric (3)
names (WaldTest) = c(“W”,“df”,“value”)
r = dim (L)[1]
W = ((tau1-tau2)2 )/v
W = as.numeric(W)
pval=1-pchisq(W,1)
WaldTest[1] = W; WaldTest[2] = r; WaldTest[3]=pval
WaldTest
}#End function WaldTest
LL = rbind(c(1, −1)); LLthetahat = c(1, 1)
120 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 9.4: Piecewise hazard rate estimates comparison for different sur­
vival duration.

9.7 Piecewise Hazard Function Analysis

The prostate cancer data obtained from the Surveillance, Epidemiology,
and End Results (SEER) Program (www.seer.cancer.gov) Public-Use Data
(1973–2015), with a follow-up until December 2015. Dataset provides the in­
cidence and survival information of those patients. Causes of death among
patients are included in this dataset. cause of death due to other reasons are
excluded from the dataset. Now patients died to prostrate cancer and censored
are included.

9.8 Different Relative Survival Analysis Package with R

9.8.1 flexrsurv
This package is useful to allow to model non-linear and non-proportional
effects using splines. It includes both non proportional and non linear effects.
Relative Survival Analysis 121

#R Package Flexsurv for Relative Survival

library(“flexrsurv”)
library(“relsurv”)
data(rdata)
data(slopop)
rdata$iage<-findInterval(rdata$age ×365.24+rdata$time,
attr(slopop,“cutpoints”)[[1]])
data$iyear<-findInterval(rdata$year+rdata$time,
attr(slopop, “cutpoints”)[[2]])
therate <­ rep(-1, dim(rdata)[1])
for( i in 1:dim(rdata)[1]){
therate[i] <­ slopop[rdataiage[i], rdataiyear[i],
rdata$sex[i]] } rdata$slorate <­ therate
rdata$sex01 <­ rdata$sex -1
fit <­ flexrsurv(Surv(time,cens) ∼ sex01+
NLL(age, Knots=60, Degree=3,
Boundary.knots = c(24, 95)),
rate=slorate, data=rdata,
knots.Bh=1850, # one interior knot at 5 years
degree.Bh=3,
Max T=5400,
Spline =“b-spline”,
initbyglm=TRUE,
initbands=seq(0, 5400, 100),
int meth=“BANDS”,
bands=seq(0, 5400, 50) )
summary(fit)
122 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 9.5: Different age wise cancer diagnosis, prostate cancer case dis­
tribution.

#Outcome of Flexrsurv Function

Coefficients:
Estimate Std. Error z value Pr(>|z|)
Baseline hazard:1 -9.3669 1.9478 -4.809 1.52e-06 ***
Baseline hazard:2 -10.8062 1.9569 -5.522 3.35e-08 ***
Baseline hazard:3 -10.6876 2.1408 -4.992 5.96e-07 ***
Baseline hazard:4 -10.7724 2.2308 -4.829 1.37e-06 ***
Baseline hazard:5 -10.5813 2.6117 -4.051 5.09e-05 ***
sex01 0.6852 0.1858 3.687 0.000227 ***
NLL(age):1 -0.6171 2.7501 -0.224 0.822446
NLL(age):2 1.6521 1.6531 0.999 0.317608
NLL(age):3 0.9809 2.5093 0.391 0.695865
NLL(age):4 2.3640 2.4392 0.969 0.332446
—
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Log-likelihood: -4781.327
Relative Survival Analysis 123

9.8.2 relsurv
The relsurv package is useful to analyzing the relative survival data. It in­
cludes relative survival ratio, crude mortality, multiplicative regression model.
#R Package “Relsurv” for Relative Survival

library(“relsurv”)
data(slopop)
data(rdata)
rs.surv(Surv(time,cens)∼sex,rmap=list(age=age×365.241),
ratetable=slopop,data=rdata)
d <­ rs.surv.rsadd(fit,newdata=data.frame(sex=1,
age=65,year=as.date(“1Jul1982”))
fit <­ rsadd(Surv(time,cens) sex+age+year,
rmap=list(age=age×365.241),
ratetable=slopop,data=rdata,int=c(0:10,15))
plot(d,xscale=365.241)

FIGURE 9.6: R package “relsurv” on r Relative Survival.

124 Bayesian Approaches in Oncology Using R and OpenBUGS

9.8.3 survexp.fr
This package is useful to obtain the Relative survival. Function
“survexp plot” avalible in “survexp.fr” is helpful to obtain the relative survival
plot. The illustration given below.
#R Package “Survexp.fr” is Useful for Relative Survival

library(“survexp.fr”)
attach(data.example)
survexp plot(futime, status, age, sex, entry date)

FIGURE 9.7: Survival plot obtained by survexp function.

Relative Survival Analysis 125

TABLE 9.1: Piecewise hazard ratio estimates in different survival intervals

in months
Interval τ1 τ2 Hazard 95%Confidence Wald P
Ratio Interval -Statistics -value
1 0 12 0.94 (0.94,0.94) 0.000 0.989
2 13 24 0.94 (0.94,0.94) 0.180 0.671
3 25 36 0.95 (0.94,0.95) 0.002 0.964
4 37 48 0.94 (0.94,0.95) 0.001 0.979
5 49 60 0.95 (0.94,0.95) 0.000 1
6 61 72 0.95 (0.94,0.95) 0.003 0.96
7 73 84 0.94 (0.94,0.95) 0.120 0.729
8 85 96 0.95 (0.95,0.95) 0.001 0.974
9 97 108 0.95 (0.95,0.96) 0.000 0.985
10 109 120 0.95 (0.95,0.96) 0.018 0.893
11 121 132 0.95 (0.94,0.95) 0.000 0.988
12 133 144 0.95 (0.94,0.95) 0.021 0.883
13 145 156 0.95 (0.95,0.96) 0.001 0.975
14 157 168 0.95 (0.95,0.96) 0.010 0.918
15 169 180 0.95 (0.94,0.95) 0.005 0.943
16 181 192 0.95 (0.95,0.96) 0.003 0.958
17 193 204 0.96 (0.95,0.97) 0.002 0.965
18 205 216 0.95 (0.95,0.96) 0.001 0.977
19 217 228 0.95 (0.94,0.96) 0.005 0.943
20 229 240 0.95 (0.94,0.96) 0.043 0.836
21 241 252 0.96 (0.95,0.97) 0.001 0.973
22 253 264 0.96 (0.95,0.97) 0.007 0.932
23 265 276 0.97 (0.95,0.98) 0.002 0.964
24 277 288 0.97 (0.96,0.99) 0.004 0.949
25 289 300 0.96 (0.94,0.98) 0.000 0.997
26 301 312 0.96 (0.93,0.99) 0.000 0.987
27 313 324 0.96 (0.92,0.99) 0.023 0.878
28 325 336 0.98 (0.95,1.00) 0.002 0.967
29 337 348 0.96 (0.94,0.99) 0.010 0.922
30 349 360 1.00 (0.93,1.09) 0.051 0.821
31 361 372 0.96 (0.90,1.02) 0.000 0.099
32 373 384 0.95 (0.85,1.07) 0.041 0.839
33 385 396 0.88 (0.75,1.04) 0.031 0.861
34 397 408 0.96 (0.82,1.12) 2.430 0.119
35 409 420 0.93 (0.81,1.07) 0.005 0.938
36 421 432 1.00 (0.84,1.19) Inf 0
37 433 444 0.93 (0.79,1.09) 0.008 0.926
38 445 456 1.03 (0.93,1.15) 0.007 0.93
39 457 468 0.87 (0.57,1.34) inf 0
126 Bayesian Approaches in Oncology Using R and OpenBUGS

TABLE 9.2: Piecewise hazard ratio estimates in different survival intervals

in months for grade I
Interval τ1 τ2 Hazard 95%Confidence Wald P
Ratio Interval -Statistics -value
1 0 12 0.94 (0.93,0.94) 0.01 0.93
2 13 24 0.91 (0.88,0.95) 0 0.98
3 25 36 0.91 (0.88,0.95) 0.71 0.4
4 37 48 0.92 (0.89,0.95) 0.15 0.7
5 49 60 0.88 (0.84,0.92) 0.09 0.77
6 61 72 0.89 (0.82,0.97) 0.09 0.76
7 73 84 0.9 (0.81,0.99) 0 0.99
8 85 96 0.9 (0.83,0.97) 0.01 0.92
9 97 108 0.9 (0.83,0.98) 0.12 0.73
10 109 120 0.88 (0.83,0.93) 0.44 0.51
11 121 132 0.93 (0.88,0.98) 0.04 0.84
12 133 144 0.91 (0.86,0.95) 0.01 0.91
13 145 156 0.92 (0.88,0.95) 0.02 0.89
14 157 168 0.9 (0.87,0.94) 0.04 0.89
15 169 180 0.91 (0.88,0.94) 0.02 0.85
16 181 192 0.93 (0.91,0.96) 0.61 0.88
17 193 204 0.95 (0.92,0.98) 0 0.44
18 205 216 0.95 (0.92,0.98) 0 0
20 229 240 0.93 (0.90,0.95) 0.1 0.33
21 241 252 0.97 (0.95,1.00) 0.02 0.75
22 253 264 0.94 (0.92,0.97) 0.08 0.88
23 265 276 0.98 (0.95,1.01) 0 0.77
24 277 288 0.96 (0.93,1.00) 0.02 0.96
25 289 300 0.98 (0.94,1.02) 0 0.89
26 301 312 0.96 (0.92,1.00) 0.01 0.95
27 313 324 0.98 (0.93,1.04) 0.05 0.94
28 325 336 0.95 (0.89,1.01) 0 0.82
29 337 348 0.95 (0.86,1.05) 0.02 0.1
30 349 360 0.99 (0.91,1.08) 0 0.99
31 361 372 1 (0.90,1.11) 0.03 0.85
32 373 384 0.94 (0.81,1.09) 0.03 0.85
33 385 396 0.84 (0.66,1.06) 0.03 0.86
34 397 408 0.97 (0.77,1.22) 0.03 0.86
35 409 420 0.87 (0.71,1.07) 0.02 0.88
36 421 432 1.11 (0.60,2.05) Inf 0
37 433 444 0.21 (0.00,0.38) 0.1 0.75
38 445 456 0.88 (0.41,1.90) 0.17 0.68
Relative Survival Analysis 127

9.9 Discussion
Relative survival analysis can help us to provide information that is difficult
to address by conventional analysis. It is really useful to adopt a conventional
analysis of a better therapeutic effect comparison. We illustrate the application
through OpenBUG software. It is performed as a Bayesian counterpart. The
methodology of relative survival by the multiplicative and additive model is
well-elaborated [49]. The application of relative survival at an individual level
is also documented [50]. The application of relative survival is very limited
only on cancer have low survival rates. A good explanation about relative and
net survival is well-documented [45]. Cause-specific survival can be used to­
ward median duration survival estimates. However, it is applicable in relative
survival. The hazard function estimates are developed through predictive scor­
ing. The treatment effect can be compared through different age interval by
piecewise hazard. It is also applicable in health policy implementation. This
chapter provides the importance of relative survival, and piecewise estimates
of hazard function are feasible to use to develop prediction score as well. It
will provide us with another dimension about the establishment of therapeutic
effect. It may be important toward health policy decision. This work will help
to understand the hazard function at a different time point interval.
Part III

Bayesian in Longitudinal

Data Analysis

129
Chapter 10

Longitudinal Data Analysis

Abstract
To describe different types of statistical tool and analytical approach for
longitudinal data analysis in oncology. A different aspect of longitudinal study
design is presented. The Bayesian counterpart to perform longitudinal data
analysis in oncology is furnished. The Bayesian estimation of a balanced lon­
gitudinal model with ARMA(1) the structure is provided. Another covariance
structure like compound symmetry is explained with R packages. Finally, A
study is presented to compare the Quality of Life in palliative patients treated
with two chemotherapeutic arms. It provides the best arm to maintain bet­
ter QoL among treated patients. The observations are taken repeatedly for
each patient during treatment. The application of the mixed effect model is
implemented.

10.1 Introduction
In oncology, conventionally the endpoints are considered as a clinical out­
come as clinical efficacy. Conventionally, clinical efficacy is defined by overall
survival (OS). Similarly, biological efficacy also considered through disease-free
survival (DFS). The OS is the gold standard in any oncology trial. Perhaps, it
is difficult to consider for any clinical benefits. The therapeutic effect becomes
more attractive if it can provide a better quality of life (QoL). The QoL always
measured in a repeated manner through patients experience. The QoL mea­
sured as a standard questionnaire. After that, the comparison of QoL between
the treatment arm provides the best effective treatment. The QoL measures
repeated outcomes over time. In a conventional randomized controlled trial,
the baseline QoL assessed and followed up longitudinally. In repeated QoL,
the same individual subjects observed, and observations are correlated. The
traditional statistical approach for analyzing such data is not suitable enough.
In this context, this chapter dedicated to mixed effect modeling to Analysis
of QoL data. The Linear mixed effect is a method to deal with repeatedly

131
132 Bayesian Approaches in Oncology Using R and OpenBUGS

measured QoL. Currently, the health-related quality of life (HRQoL) becomes

an important parameter. Longitudinal QoL data are widely adopted to explore
cancer progression. It shows that whether the change measured over time is
related to socio-economic factors, age at onset, education and other genetic
factors. However, the change of QoL over time across individuals is logical.
This data analysis is not straightforward. The well-appreciated method for
longitudinal data analysis is random or mixed effect modeling. There are sev­
eral biostatistical application of longitudinal data. Both in epidemiological and
clinical filed. The term longitudinal defined as repeatedly measured observa­
tions of the same individual. The focus of this work is to determine whether
changes over time occurred or not. The focus of this work is about Bayesian
methods for longitudinal data analysis through the mixed effect model.

10.2 Advantages of Longitudinal Analysis

(I) Prospective ascertainment of exposure.
(II) Separation of time effects.
(III) Incident events are recorded.
(IV) Measurement of individual change in outcomes.

10.3 Limitation of Longitudinal Analysis

(I) Participant follow-up.
(II) Correlated data.
(III) Time-varying covariates.

10.4 Mixed Effect Model

The relation between outcome and associated factors are explored by sta­
tistical modeling. The widely adopted statistical model is the linear model.
Further, this linear model can be separated into fixed and random effect. The
parameters linked with an entire population is considered as a fixed effect.
However, the parameters changes individually are considered a random ef­
fect. A model consisted of the fixed and random parameter is defined as the
mixed-effect model. Suppose that patients QOL is measured repeatedly. The
Longitudinal Data Analysis 133

repeated measurements are defined as yij . The ith individuals jth time point
measurement is defined as yij . Now if the maximum number of visits are de­
fined as m then j = 1, ..., mi . Similarly, a total of n sample data is defined as
i = 1, ...., n. The mean parameter linked with each measurement is µij . Now
the term µij changes from different situations. For example, the presence of
two time period (j = 1, 2) says the estimation of µi1 , µi2 will vary. The simple
relation can be obtained as µij = αµi,j−1 .
The model can be settled as

yij |µij ∼ N(µij , σ 2 ) (10.1)

µij = xTij β + zij

T
bi (10.2)
The term xT
ij β
is a linear predictor. The random effect component is zij T
bi
for the ith individuals jth time point measurement. The covariance matrix is
presented as Vi .
In a Bayesian framework, the prior distributions required to assigned. The
prior dustrubution for β and bi is assumed. It is feasible to assign bi as N (0, τ 2 ).
Further the prior value τ 2 can be assigned as IG(1, 0.0260). The likelihood can
be written as �
p(Yij |tij , bi ) (10.3)
i,j

Now it can be assumed that p(Yij |tij , bi ) = N(µij , σ 2 ) and

µij = β0 + β1 .tij + b0,i (10.4)

It is required to assign the prior distributions for bi , β0 , β1 , and σ 2 . The pos­

terior distribution is presented as
� �
p(Yij |tij , bi , σ 2 , τ 2 ) π(bi , τ 2 )π(β0 )π(β1 )π(σ 2 )π(τ 2 ) (10.5)
i,j i

10.5 Different R Packages for Longitudinal Data Analy­

sis
Longitudinal data always measured to understand disease progression and
process. However, it depends on the number of longitudinal observations tu­
mour progression measured by growth or recurrence. The “growcurves” pack­
age available in R is suitable to perform the hierarchical Bayesian modeling
for dependence among subjects, by continued and combined therapy. It works
by Dirichlet process before random-effect modeling. It brings the information
to estimate as flexible and non-parametric. A different statistical package like
134 Bayesian Approaches in Oncology Using R and OpenBUGS

MLwiN is useful for parametric hierarchical modeling. DP prior one random

effect model is applicable by “DPpackage” . Challenges to work with missing
value, explore different quality criterion and the graphical interface is nicely
handled by “kml” package. The likelihood clustering distance metric performed
by “kmlcov” package works through the generalized linear model the recursive
separation of linear and nonlinear mixed effect performed by “longRPart2”
package.

10.6 Random-Effect Model with R

A dedicated team for R (R Core Team 2019) provides the R package lme4
to work with linear mixed effect model [51]. The function is known as lme4.
The generalized linear mixed models and nonlinear mixed models are possible
to work with lme4 function. The term mixed is used to accommodate the
fixed and random effect. In this chapter, we discussed Bayesian application
with linear mixed effect model.
The alterative of lme4 is nlme that is available in R for mixed-effect model­
ing (Pinheiro, Bates, DebRoy, Sarkar, and R Core Team 2015). However, lme4
is efficient toward the complex situation. It is simple to perform on random
effect models. It is easy to perform by likelihood function on random effect.
The package “nlme” is compatible to work with residuals and the residuals
generated due to heteroscedasticity and autocorrelation structure.
The packages used to work with specific functions in R. These functions
developed with logic, statistics, probability, and mathematical programming.
Few essential and basic functions automatically installed in R. The specific
function can only work through the installation of the package. There are sev­
eral thousand packages available in R to work with the specialized problem.
The list of packages is available in http://cran.r-project.org. These packages
can also be downloaded from: http://cran.r-project.org. In R console this pack­
ages can also be installed by using the packages pulldown menu in R. The lme4
package is useful to work with mixed effect model. A drop-down menu can in­
stall this package. After that, this package is required to load by running the
library( lme4) in R console.

10.6.1 bayeslongitudinal
This function is used to obtain the longitudinal regression model by
Bayesian methodology. There are two functions, i.e., mhsc and mharma11.
The mharma11 is used for autoregressive modeling and mhsc for compound
symmetry structure. Both the function is illustrated below:
Longitudinal Data Analysis 135

# Bayesian Estimation of a Balanced Longitudinal Model with

ARMA(1) Structure

library(“bayeslongitudinal”)
attach(Dental)
Y=as.vector(distance)
X=as.matrix(cbind(1,age))
mharma11(Y,X,27,4,c(1,1),0.5,0.5,1,1,1,1,500,50)
mharma11(Data, Matriz, individuos, tiempos, betai, rhoi, gammai,
beta1i, beta2i, beta1j, beta2j, iteraciones, burn).
Data A=vector with the observations of the response variable.
Matriz=The model design matrix.
individuos=A numerical value indicating the number of individuals
in the study.
tiempos=A numerical value indicating the number of times observa­
tions were repeated.
betai=A vector with the initial values of the vector of regressors.
rhoi= A numerical value with the initial value of the correlation for
rho.
gammai= A numerical value with the initial value of the correlation
for phi.
beta1i= A numerical value with the shape parameter of a beta apriori
distribution of rho.
beta2i= A numerical value with the scaling parameter of a beta
apriori distribution of
rho.
beta1j= A numerical value with the shape parameter of a beta apriori
distribution of phi.
beta2j= A numerical value with the scaling parameter of a beta
apriori distribution of phi.
iteraciones= A numerical value with the number of iterations that
will be applied the algorithm MCMC.
burn= Number of iterations that are discarded from the chain.
136 Bayesian Approaches in Oncology Using R and OpenBUGS

# Bayesian Estimation of a Balanced Longitudinal Model with Com-

pound Symmetry Structure

library(“bayeslongitudinal”)
attach(Dental)
Y=as.vector(distance)
X=as.matrix(cbind(1,age))
mhsc(Y,X,27,4,c(1,1),0.5,1,1,500,50)
mhsc(Data, Matriz, individuos, tiempos, betai, rhoi, beta1i, beta2i,
iteraciones, burn)
Data A=vector with the observations of the response variable.
Matriz=The model design matrix.
individuos=A numerical value indicating the number of individuals
in the study.
tiempos=A numerical value indicating the number of times observa­
tions were repeated.
betai=A vector with the initial values of the vector of regressors.
rhoi=A numerical value with the initial value of the correlation.
beta1i=A numerical value with the shape parameter of a beta apriori
distribution of rho.
beta2i=A numerical value with the scaling parameter of a beta
apriori distribution of
rho=iteraciones numerical value with the number of iterations that
will be applied the algorithm.
MCMC=burn Number of iterations that are discarded from the chain.

10.7 Illustration with Quality of Life Using OpenBUGS

A study was performed in India to compare the Quality of Life in palliative
patients treated with two chemotherapeutic arms. The study about compares
the best arm to maintain better QoL among treated patients. The observa­
tions are taken repeatedly for each patient during treatment. This study has
required the application of the mixed effect model. The data illustrated in Ta­
ble 10.1. the overall data trend plotted through the spaghetti plot. Spaghetti
plot is a widely used tool in longitudinal data plotting. The R package “ggplot”
can be used to plot in R.
Longitudinal Data Analysis 137

FIGURE 10.1: Bayesian estimation with compound symmetry structure.

library(“ggplot2”)
mydata<-data.frame(ID,painscore,visit,Arm)
p <­ ggplot(data = mydata, aes(x = time,y=painscore,group=ID))
p + geomline()
138 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 10.2: Bayesian estimation with ARMA(1) structure.

# Pain Score Illustrated Data

ID painscore Visit Arm ID painscore Visit Arm

1 13.18 1 A 4 20.75 1 A
1 17.32 2 A 4 30.74 2 A
1 29.55 3 A 4 23.24 3 A
1 34.53 4 A 4 34.17 4 A
1 23.78 5 A 4 24.93 5 A
1 20.18 6 A 4 32.94 6 A
1 30.63 7 A 4 23.51 7 A
2 27.46 1 B 5 16.90 1 B
2 15.63 2 B 5 24.23 2 B
2 26.45 3 B 5 20.03 2 B
2 14.62 4 B 5 27.03 4 B
2 29.15 5 B 5 20.09 5 B
2 29.53 6 B 5 29.49 6 B
2 35.14 7 B 5 24.23 7 B
3 39.10 1 B 6 18.62 1 B
3 20.98 2 B 6 25.08 2 B
3 09.62 3 B 6 16.80 2 B
3 29.98 4 B 6 17.73 4 B
3 23.60 5 B 6 20.02 5 B
3 30.95 6 B 6 11.96 6 B
3 27.28 7 B 6 39.44 7 B
Longitudinal Data Analysis 139

FIGURE 10.3: Longitudinal profile plot obtained by Spaghetti plot.

# Bayesian Mixed Effect Model in OpenBUGS

model {
for (i in 1:n) {
for (j in 1:K) {
mu[i,j] <­ m + a[i]
y[i,j]∼ dnorm( mu[i,j], tau )
}
a[i] ∼ dnorm( 0, tau.a)
}
m∼ dnorm(0.0,0.001) #Define prior distribution of m
tau∼dgamma(0.001,0.001)#Define prior distribution of tau
tau.a∼dgamma(0.001,0.001)#Define prior distribution of tau.a
s2<­ 1/tau # Define prior value for s2
s2.a<­ 1/tau.a # Define prior value for s2.a
ts2<­ s2+s2.a # Define value for ts2
cor<­ s2.a/ts2 # calculate correaltion coefficient
s<-sqrt(s2)
s.a<­ sqrt(s2.a)
for(i in 1:n) {
for(j in 1:K){
res[i,j]<­ y[i,j]-mu[i,j]
}}
R2<-1-pow(sd(res[1:n,1:K])/sd(y[1:n,1:K]),2)
}
140 Bayesian Approaches in Oncology Using R and OpenBUGS

# Pain Score Illustrated Data

DATA
list(n=14, K=7,y=structure(.Data=c(13.18,17.32,29.55,34.53,
23.78,20.18,30.63,27.46,15.63,26.45,14.62,29.15,29.53,35.14,
39.10,20.98,9.62,29.44,23.60,30.95,27.28,20.75,30.74,23.24,
34.17,24.93,32.94,23.51,16.90,24.23,2.03,27.03,20.09,29.49,
24.23,18.62,25.08,16.80,17.73,20.02,11.96,39.44,15.59,21.12,
13.63,33.99,12.74,26.82,20.28,27.95,13.01,27.19,34.90,38.22,
22.25,36.68,12.98,15.71,23.14,21.09,30.17,30.09,10.30,
36.75,15.31,16.35,21.38,18.44,11.39,17.33,28.60,30.38,
33.50,31.80,22.62,19.50,22.95,28.68,29.28,18.75,24.31,
22.89,14.33,25.12,22.60,12.62,15.89,19.03,18.36,25.80,
8.39,25.17,22.21,45.62,7.82,17.88,19.02,20.87)
,.Dim = c(14, 7)))
INITS
list(m=0.0, tau=1.0)

10.8 Result
The application of longitudinal data in oncology research arises several
times [52, 53, 54, 55, 56, 57, 54]. In epidemiological and clinical setup, both
the case it is possible to occur. The mixed-effect model is used to show whether
the changes in measurement over time individual level or group level as well.
If the group level measurement difference occurs, then it is significant or not.
The simple application of a mixed effect model is about a group of patients
comparison between two time periods. The Bayesian method in longitudinal
data presented in this chapter. The posterior mean of changes in pain score ob­
tained with 23.11. The results also present individual variation and group-wise
variation. The Analysis executed with 20,000 observations for the simulation,
with the burn-in of 1,000 and a refresh of 100. The posterior distributions of
within- and between-subject variabilities are observed through (σa2 and σ 2 , re­
spectively) respectively. The trace plots for all these parameters given below.
The OpenBUGS gives the trace plots. It shows the variation result for each
parameter during the simulation.
Longitudinal Data Analysis 141

TABLE 10.1: Posterior estimates obtained are presented

Parameter Posterior mean SD 2.5% 97.5%
µ 23.11 1.24 21.33 24.92
cor 0.02 0.05 0.00 0.15
τ 0.015 0.00 0.01 0.02
τa 95.15 275.6 0.09 917.30
σa 0.86 1.36 0.03 3.312
σa2 2.61 26.15 0.00 10.97
σ 2 + σa2 3.47 27.51 0.03 14.282
σ 8.08 0.60 7.00 9.34

FIGURE 10.4: Trace plot of regression estimates obtained by 20,000 itera­

tions.
Chapter 11

Missing Data Analysis

Abstract
The occurrence of missing data is common in any experimental field of
research. Oncology research is not exceptional. Patients follow-up visits often
occur. The follow-up visits raise the chance of appearance of missing observa­
tions. Missing observation seriously affects statistical inference. This chapter
is about presenting a Bayesian technique to handle the missing data analy­
sis. The objective is to obtain the best statistical inference in the presence of
missing data.

11.1 Introduction
There is a different definition of missing data. Suppose the ith individuals
jth time point measurement is defined as Yij . Now i could take a value from
1 to N. N is the sample size. Similarly, j can take any value from 0 to a
maximum number of visits. Now another term Rij as an indicator to represent
the missing observation. If Yij is missing then Rij = 0. If Yij is present then
Rij = 1. The term Rij can be defined as the probability

P (R|Yobs , Ymiss , η) (11.1)

η is the unknow parameter linked with missing.

Yobs are the portion of the n × n matrix of Y that are present.

Ymiss are the portion of the n × n matrix of Y that are missing.

143
144 Bayesian Approaches in Oncology Using R and OpenBUGS

11.2 Different Types of Missing Data

11.2.1 Missing completely at random (MCAR)
The simplest form to define the type of missing data is called as missing
completely at random (MCAR).

P (R|Yobs , Ymiss , η) = P (R|η) (11.2)

While the probability of missing data is independent of the observed data. If

the missing data is unrelated to the observed data, for example, the patients
missed his follow-up visit due to a hurricane of that day. The type of missing
will be defined as MCAR.
However, if the same person missed the visit due to severe pain in his body.
The patient is a subject of treatment outcome. Pain may occur due to dose
toxicity. The missing data due to pain appearance will not be considered as
MCAR. This MCAR do not affect the biasedness of the analysis. The power
of the analysis indeed reduced, but it does not incline the bias in the dataset.

11.2.2 Missing at random (MAR)

The alternative form of MCAR is defined as missing at random (MAR).

P (R|Yobs , Ymiss , η) = P (R|Yobs , R|η) (11.3)

The probability of the missing information is dependent only on the observed

data. Data imputation technique is required to perform to overcome the miss­
ing information. Simple regression technique is useful to overcome the missing
observation. It can be considered as ignorable missing.

11.2.3 Missing not at random (NMAR)

The third alternative is defined as missing not at random (NMAR). For
example, a person having severe chemo response toxicity is less likely to fulfil
the quality of life (QOL) questionnaires than others. This type of missing data
is known as NMAR. The mean response of QOL obtained from missing data
will be biased enough.

P (R|Yobs θ, η) = P (R|Yobs , η)P (Yobs |θ) (11.4)

If θ represents the vector of parameters for the distribution of Y . It shows

that the inference for θ is based on the likelihood function, and it depends on
the observed data. The likelihood is defined as

L(θ|Yobs ) ∝ P (Yobs |θ) (11.5)

Missing Data Analysis 145

The Bayesian approach helps in this context. The likelihood is multiplied with
the prior value of θ. The application of Bayes helped to obtain the posterior
value of θ. Bayes’s theorem is the posterior density of theta. Now the posterior
estimate of θ can be obtained as

π(θ|Y ) = cL(Y |θ)π(θ) (11.6)

The term c is used as normalizing constant and the likelihood function is

presented as
L(Y |θ) = f (Y |θ) (11.7)
The density is measured by f . The predicted density of the future observations
is Z �
g(z|y) ∝ f (z|θ)π(θ|y)dθ (11.8)
Ω
The integration is taken with respect to the parameter space Ω. The Bayesian
method for missing data is helping to identify distributions for all the param­
eters. The parameters are assumed as the prior distribution. Both response
and covariate can be imputed through the Bayesian method. The imputation
is performed through prior distribution. Bayesian methods are free to impute
data without any statistical inference. There are different method exists like
Full Bayesian, maximum likelihood through Bayesian technique.

11.3 Different Softwares

There are different software to handle missing data in oncology research.
The PROC MI is available is SAS to work with multiple imputation tech­
niques for missing data. Perhaps, the normally distributed data is suitable to
work with SAS. The right-censored data is not suitable to imputed through
PROC MI. The OpenBUGS is an application software for the Bayesian anal­
ysis of complex statistical models. It works through Markov chain Monte
Carlo (MCMC) methods. The missing data imputation technique can be
easily performed in OpenBUGS. This software also works well with R. The
R2OpenBUGS or BRugs packages are available in R to jointly work for further
analysis.

11.4 Illustration with Lung Cancer Data

We consider data of Head and neck cancer conducted by the Tata Memo­
rial Centre. The results of this study were reported by Noronha [58]. The
palliative chemotherapy patients were selected in this study. Cetuximab is a
costly chemotherapeutic drug. Only a total of 30 patients could afford this
146 Bayesian Approaches in Oncology Using R and OpenBUGS

drug. The quality of life (QOL) data is presented among those patients. Sim­
ilarly, another group of the cohort (n=40) is treated with cisplatin therapy.
Data is filled with European Organisation for Research and Treatment of Can­
cer(EORTC) QOL questionnaire. The primary objective of this study was to
compare the QOL between Cisplatin and Cetuximab arm. Presence of missing
data is handled by Bayesian methodology. The imputation technique is per­
formed to obtain the missing data. The comparison between treatment effect
on repeated measurements is presented in Chapter 9. This chapter is restricted
only on missing data handling technique. Data is presented below.

11.5 Different Package for Missing Data with R

11.5.1 BMTAR
This package is used for Bayesian approach for multivariate threshold au­
toregressive (MTAR) models with missing data using Markov chain Monte
Carlo methods. One example is given in the following text.
Autoplot on Missing Data by BMTAR Package

library(“ggplot2”)
library(“BMTAR”)
data(missingest)
autoplot.regime missing(missingest,1)
library(ggplot2)
data(datasim)
data = datasim$Sim$Zt
parameters=list(l=1,orders = list(pj = 1))
initial=mtarinipars(tsregime obj = tsregime(data),
list model = list(pars = parameters))
estim1=mtarns(ini obj = initial,niter = 500,chain = TRUE)
autoplot.regime model(estim1,2)
autoplot.regime model(estim1,2)
autoplot.regime model(estim1,3)
autoplot.regime model(estim1,5)
data(datasim)
yt=datasim$Sim
Yt=yt$Yt
Zt=yt$Zt
data=tsregime(Yt,Zt)
autoplot.tsregime(data)
Missing Data Analysis 147

FIGURE 11.1: regime model object ggplot for the outputs on the function
outputs.

FIGURE 11.2: tsregime object ggplot for the outputs.

148 Bayesian Approaches in Oncology Using R and OpenBUGS

11.5.2 NestedCohort
The NestedCohort is dedicated for survial analysis if the covaraites are
missing. The nested cohort analysis is performed. One example to create sur­
vival plot is provided below. Secondly, the application of Cox PH by Nestec-
Cohort package is given below.
Survival Analysis with Missing Covariate

library(“NestedCohort”)
data(zinc)
mod <­ nested.km(survfitformula=“Surv(futime01,ec01==1)∼ zn­
quartiles”,
samplingmod=“ec01×basehist”,data=zinc)
plot(mod,ymin=.6,xlab=“Time (Days)”,ylab=”Survival”,
main=“Survival by Quartile of Zinc”,lty=1:4,)
legend(2000,0.7,c(“Q1”,“Q2”,“Q3”,“Q4”),lty=1:4)
coxmod <­ nested.coxph(coxformula=“Surv(futime01,ec01==1)∼
sex+agepill+smoke+drink+mildysp+moddysp+sevdysp+anyhist+
zncent”,samplingmod=“ec01×basehist”,data=zinc)
summary(coxmod)
mod <­ nested.stdsurv(outcome=“Surv(futime01,ec01==1)”,
exposures=“znquartiles”,
confounders=“sex+agestr+smoke+drink+mildysp+moddysp+
sevdysp+anyhist”
samplingmod=“ec01×basehist”,exposureofinterest=“Q4”,plot=TRUE,
main=“Time to Esophageal Cancer by Quartiles of Zinc”,data=zinc)

FIGURE 11.3: Survival plot by NestedCohort package.

Missing Data Analysis 149

FIGURE 11.4: Standarized survival plot.

Prepare Dataset

Measurment on pain score for first Arm for 14 patients at

four vistis are obtained as
Painscore1<-c(13.18,17.32,29.55,34.53,27.46,15.63,NA,NA,
39.10,20.98,9.62,29.44,20.75,30.74,23.24,34.17,
16.90,24.23,NA,NA,18.62,25.08,16.80,17.73,
15.59,21.12,NA,NA,27.95,13.01,27.19,34.90,
12.98,15.71,NA,NA,36.75,15.31,16.35,21.38,
28.60,30.38,33.50,31.80,28.68,29.28,NA,NA,
22.60,12.62,15.89,19.03,25.17,22.21,NA,NA)
visit<-c(1,1,1,1,1,1,1,1,1,1,1,1,1,1,
2,2,2,2,2,2,2,2,2,2,2,2,2,2,3,3,3,3,3,
3,3,3,3,3,3,3,3,3,4,4,4,4,4,4,4,4,4,4,4,4,4,4)
Similarly, for second arm is presented as
Painscore2<-c(13.18,17.32,29.55,34.53,27.46,15.63,NA,NA,
39.10,20.98,9.62,29.44,20.75,30.74,23.24,34.17,
16.90,24.23,NA,NA,18.62,25.08,16.80,17.73,
15.59,21.12,NA,NA,27.95,13.01,27.19,34.90,
12.98,15.71,NA,NA,36.75,15.31,16.35,21.38,
28.60,30.38,33.50,31.80,28.68,29.28,NA,NA,
22.60,12.62,15.89,19.03,25.17,22.21,NA,NA)
visit<-c(1,1,1,1,1,1,1,1,1,1,1,1,1,1,
2,2,2,2,2,2,2,2,2,2,2,2,2,2,3,3,3,3,3,
3,3,3,3,3,3,3,3,3,4,4,4,4,4,4,4,4,4,4,4,4,4,4)
150 Bayesian Approaches in Oncology Using R and OpenBUGS

Phase I Dose-Escalation Method Using OpenBUG 1

model;
{
beta1 ∼ dnorm(0.0, 0.001)
beta2 ∼ dnorm(0.0, 0.001)
for(i in 1:N)for (j in 2:M)Y[i,j]∼ dnorm(m[i,j],T)
for(i in 1:N)for (j in 2:M)m[i,j]<-beta1 ×(1-rho)
+beta2×(visit[j]+rho×Y[i,j-1]-rho×visit[j-1])
for(i in 1:N)Y[i,1]∼ dnorm(mu[i,1],T)
rho ∼ dbeta(1,1)
for(i in 1:N)mu[i,1]<-beta1+beta2×visit[1]
sigma<-1/T
T ∼ dgamma(.01,.01)
}
list(N = 14,M= 4,Y = structure(.Data = c(
13.18,17.32,29.55,34.53,27.46,15.63,NA,NA,39.10,20.98,
9.62,29.44,20.75,30.74,23.24,34.17,16.90,24.23,NA,NA,
18.62,25.08,16.80,17.73,15.59,21.12,NA,NA,27.95,13.01,
27.19,34.90,12.98,15.71,NA,NA,36.75,15.31,16.35,21.38,
28.60,30.38,33.50,31.80,28.68,29.28,NA,NA,22.60,12.62,
15.89,19.03,25.17,22.21,NA,NA),. Dim = c(14,4)),
visit = c(1,2,3,4))
# initial values
list(beta1 = 0,beta2 = 0,rho =.5,T= 1)

Similarly, code can be performed for arm two.

Missing Data Analysis 151

Phase I Dose-Escalation Method Using OpenBUG 2

model;
{
beta1 ∼ dnorm(0.0, 0.001)
beta2 ∼ dnorm(0.0, 0.001)
for(i in 1:N1){for (j in 2:M1){Y[i,j]∼dnorm(mu[i,j],tau)}}
for(i in 1:N1){for (j in 2:M1){mu[i,j]<-beta1∗ (1-rho)
+beta2∗ (age[j]-rho∗ age[j-1])+rho∗ Y[i,j-1]}}
for(i in 1:N1)Y[i,1]{dnorm(mu[i,1],tau)}
rho ∼ dbeta(1,1)
for(i in 1:N1){mu[i,1]<-beta1+beta2∗ age[1]}
tau ∼ dgamma(.01,.01)
sigma<-1/tau
}
list(N1 = 17,M1 =4,Y = structure(.Data = c(10.18,14.32,26.55,31.53,
24.46,NA,23.45,11.62,
36.10,NA,6.62,26.44,
17.75,27.74,20.24,31.17,
13.90,21.23,NA,24.03,
15.62,22.08,13.80,14.73,
12.59,NA,10.63,30.99,
24.95,10.01,24.19,31.90,
9.98,NA,20.14,18.09,
33.75,12.31,13.35,18.38,
25.60,NA,NA,28.80,
25.68,26.28,15.75,21.31,
19.60,9.62,12.89,16.03,
22.17,NA,NA,NA,
24.61,22.87,17.00,17.87,
36.14,24.20,26.92,24.21,
16.85,16.43,22.93,41.27),.Dim = c(17,4)),
age = c(2,4,14,16))
# initial values
list(beta1 = 0,beta2 = 0,rho =.5,tau = 1)

11.6 Conclusion
Repeatedly measured immune response and QOL is typical in oncology
research. Longitudinal modeling is required in repeatedly measured oncology
152 Bayesian Approaches in Oncology Using R and OpenBUGS

research. The challenge of longitudinal modeling is the presence of missing

data. Unfortunately, the presence of missing observation in response and co­
variates are frequent. Different missing value techniques are presented through
MCAR, NMAR, and MAR. The MAR is presented; the likelihood function
is explored. Finally, the missing values imputed. The Bayesian inference is
applied, and the posterior estimate is obtained the Bayesian works through a
combination of likelihood, prior and posterior estimates. The posterior esti­
mate is presented in Table 11.1. The estimates of the posterior mean, standard
deviation, and 95% credible intervals are presented in Table 11.1.

TABLE 11.1: Posterior estimate in presence of missing data

Intercept Posterior Mean SD MC Error 2.5% Median 97.5%

β1 19.61 7.16 0.57 0.68 21.7 26.58
β2 1.56 2.11 0.171 -0.15 0.86 8.07
rho 0.16 0.13 0.00 0.00 0.14 0.47
T 0.01 0.00 0.00 0.00 0.01 0.025
Chapter 12

Joint Longitudinal and Survival

Analysis

Abstract
This chapter is dedicated about joint modeling toward analyzing the joint
modeling approach. It is possible to work separately for longitudinal data and
survival data. However, we will explore the joint model to handle both types
of data. The Cox PH hazard model is capable of working with survival data.
However, we can extend to incorporate time-dependent variables. The work is
about extending the longitudinal measurement as a time-dependent covariate.
A conventional approach like partial likelihood is not suitable enough in this
context. Joint modeling is attempted to estimate the same parameters present
in the two or more models.

12.1 Introduction
The survival outcomes are handled by the Cox model. Similarly, the linear
mixed effect model is suitable to work with longitudinal measurements. The
joint model plays the role of longitudinal covariate and survival outcomes.
Present relation between survival and longitudinal model can be worked in
two ways one by using trajectory obtained from the longitudinal observations
in the hazard function of the death event. The shared random effects are
also useful in this context [59]. Models can be estimated by the likelihood
or Bayesian method. R packages to fit joint models are joinR, JMbayes, JM
[60, 61, 62, 63, 64]. The Joint modeling was first considered in AIDS research
[65, 66]. The primary interest was about survival analysis in the presence
of time-dependent covariates. It was measured repeatedly with measurement
error. However, drop out presence is obvious [67, 68]. Joint modeling provides
less bias and more efficient estimates of the treatment effect. It is considered
as most power treatment comparison method [67].

153
154 Bayesian Approaches in Oncology Using R and OpenBUGS

The joint model through consideration of dropout is presented as

� �
p(yi0 , di |xi , zi , wi , θ, Ψ) = p(yi |xi zi , θ, bi )p(di |wi , Ψ, bi )dyim dbi
� �
= p(yi |xi zi , θ, bi ) (12.1)
�
= p(yi0 |xi , zi , θ, bi )p(di |wi , Ψ, bi )dbi

The term di shows the time of dropout, e.g., di = 3 as ri = (1, 1, 0, 0, ...). The
separate model for longitudinal data works with mixed effect model. The time-
to-event model works for survival data. In time-dependent survival model,
the time-dependent covariates are external because the value of the covariate
at point t is not affected by the occurrence of an event at any time point
u, t > u. Perhaps, it is not same in longitudinal setup. It is inherited property
by the subject those are involved in the failure process. Therefore, the joint
model is required to perform. Different software is capable of working with
longitudinal and survival model separately. These are available for several
years. For example, most advanced statistical software in R (R Development
Core Team 2019) several packages are available to work with mixed effect
modeling. Packages are nlme [69] and package lme4 [70]. Survival analysis
packages are available as survival [71]. Finally, the packages available for joint
modeling as well. In this work, we explored the Bayesian counterpart of the
joint model through OpenBUGS.

12.2 Data Methodology

Dataset is obtained from: Gene Expression Omnibus (GEO) database
(https://www.ncbi.nlm.nih.gov.in/geo/). The data accession number is GSE
65622.This illustration is obtained from a published dataset. Initially, it was
published in April 2016 and thereafter in August 2016. All patients were
treated with neoadjuvant chemotherapy (NACT) and followed by chemoradio­
therapy (CRT). Treated patients were evaluated after four weeks after CRT
completion. A total of 80 patients was considered for this analysis. Serum
values were collected at baseline(visit=1), at post-NACT(visit=2), end of
CRT(visit=3), post-CRT (visit=4), and at post-CRT(visit=5). The anti-body
was obtained by AHH-BLG-1; RayBiotech Inc. A total of 507 proteins were
measured. However, in this work, a total of 4 protein value were considered
at different visits. Different gene filtration technique is performed to obtain
a total of 4 protein. In addition, the value of time-to-event, i.e., death and
duration of survival, is defined under the survival analysis context. The ob­
jective of this study is to obtain the effect of proteins according to their effect
on cancer progression.
Now a separate discussion of survival and longitudinal is given below.
Joint Longitudinal and Survival Analysis 155

12.3 The Longitudinal Model

Suppose the longitudinal model is defined as

yij |µij ∼ N (µij , σ 2 ) (12.2)

and
µij β = xTij β + zij
T
bi (12.3)
The times of observation for the ith individual is presented as sij . It is defined
as

xTij β = β10 + β11 sij + β12 sij × drugi + β13 genderi + β14 previ + β15 stratumi
(12.4)
T
zij bi = W1i (sij ) (12.5)
It is formaulated as
W1i (sij ) = b0i + b1i sij (12.6)
Now there is two component. One is random effect and another is random slop
for time. It is presented with bivariate normal prior distribution of covariance
G, i.e., bi ∼ N (0, G). The noninformative prior is considered with multivariate
normal distribution formation. The inverse gamma distribution is defined for
σ 2 ∼ IG(0.1, 0.1). There could be a different prior choice of the inverse gamma
distribution.

12.4 The Survival Model

Suppose the baseline distribution about the occurrence of death is general
Weibull distribution. The duration of the interval is defined as ti inline with the
Cox proportional hazard model. Now Weibull distribution can be formulated
with two parameters by scale φ and mean level Ψ parameter as ,

ti ∼ Weibull(φ, Ψi ) (12.7)

Now φ = 1 follows the exponential model with constand hazard model in

followup period. It can be specified as

log(Ψ) ∼ Weibull(φ, Ψi ) (12.8)

The linear predictor is formulated as

xT2i β2 = β20 + β21 drugi + β22 genderi + β23 prev + β24 stratumi (12.9)
156 Bayesian Approaches in Oncology Using R and OpenBUGS

Finally, the random effect is formualted as

W2i = γ0 b0i + γ1 b1i (12.10)

Regession parameters are assumed to have non-informative prior. Weibull pa­

rameter φ is defined as Gamma(1,1) prior.

12.5 The Joint Model

The objective of the joint model is to link between two models. The first
step includes all similar variable in both the model. Both survival and protein
value is related to treatment, trgscore ,crcargrade, tstage and yptstage. Next
step is to define the common variable into the random effect term. It is assumed
that b0i and b1i appears in both the model. The scaled parameter in the
survival model is defined as γ0 and γ1 .

12.6 Submodels Specification

Suppose Ti is defined as observed survival time for the ith subject (i =
1, ..., n). The minimum value between Ti and censoring time Ci . Now Ti =
min(Ti∗ , Ci ).
δi = I(Ti∗ ≤ Ci ), where I() (12.11)
Now I() is the indicator function. It is 1 if Ti∗ ≤ Ci and 0 otherwise. Data
measured toward a time-to-event outcome is generated as {(Ti δi ), i = 1, ...n}.
Now the longitudinal response is presented as yi (t) for the ith subject t time
point measurement. Suppose the longitudinal data defined as

yij = {yi (tij ), j = 1, ..., ni } (12.12)

The objective is to link the true and unobserved value of longitudinal data
at time point t. The unobserved value is defined as Mi (t) with the outcome
event Ti∗ .

hi (t|Mi (t), wi ) = limdt→0 Pr{t ≤ Ti∗ < t + dt|Ti∗ ≥ t, Mi (t), wi }/dt (12.13)

hi (t|Mi (t), wi ) = h0 (t)exp{γ T wi + αmi (t)} (12.14)

Now mI (t) = {MI (U ), 0 ≤ U < T }. The history of the true unobserved
longitudinal process till the observation t, is defined as h0 (). The term h0 ()
is considered as baseline risk function. Now the vector of baseline covariate
Joint Longitudinal and Survival Analysis 157

TABLE 12.1: Estimates obtained through survival package in R on survival

outcome
n Events Median 0.95LCL 0.95UCL

Trgscore=1 32 3 NA 2668 NA
Trgscore=2 148 18 NA 1445 NA
coef exp(coef) se(coef) z p-value
Trgscore 0.4705 1.6008 0.6267 0.751 0.453

is presented by wi . The corresponding regression coefficients are γ. Now the

parameter α is used to define the longitudinal toward the risk of death. It is
required to incorporate the time-dependent covariates Suppose s0 define the
absolute continuous baseline survival function. It is defined as
�� ∗ �
T
T
exp{γ ω + αm(s)}ds ∼ S0 (12.15)
0

It can be expressed as

hi (t|Mi (t), ωi ) = h0 {Vi (t)} + αmi (s)}, (12.16)

and
� t
Vi (t) = exp{γ T ωi + αmi (s)}ds (12.17)
0

Further, the baseline hazard rate can be formulated as h0 (.) with a spe­
cific distribution. Now the entire covariate history is defined as Mi (t) to
influence the subject-specific risk factor. Now h0 () can be evaluated as a
subject-specific risk by Vi (t)). Covriate infulence Mi (t) can be formualted
as Si {t|Mi (t)} = S0 Vi (t). Thereafter the survival models presented as mi (t)
to define the value of longitudinal covariate at single time point t. But the lon­
gitudinal information is collected repeatedly and covariate or risk developed
by longitudinal trajectory Mi (t) as

yij = {yi (tij ), j = 1, ..., ni (12.18)

yi (t) = mi (t) + �i (t)

(12.19)
= xTi (t)β + ziT (t)bi + �i (t), �i (t) ∼ N (0, σ 2 )

The fixed effect parameter presented by β. Now the random effect is xi (t)[72,
73].
158 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 12.1: Kaplan-Meier curve on Tumor Regression Grading.

FIGURE 12.2: Linear mixed effect plot on CLC on Tumor Regression Grad­
ing.
Joint Longitudinal and Survival Analysis 159

FIGURE 12.3: Linear mixed effect plot on GREMLIN on Tumor Regression

Grading.

FIGURE 12.4: Linear mixed effect plot on MMP19 on Tumor Regression

Grading.
160 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 12.5: Linear mixed effect plot on TGFbetaRIII on Tumor Regres­

sion Grading.

FIGURE 12.6: Traceplots on 20,000 iterations on β1 .

Joint Longitudinal and Survival Analysis 161

FIGURE 12.7: Traceplots on 20,000 iterations on β2 .

TABLE 12.2: Estimates obtained through nlme package in R

Outcome Parameter Value Std.Error DF t-value p-value

CLC Intercept 136.27 58.54 81 2.32 0.02

Trgscore 20.17 31.26 41 0.64 0.52
time -3.05 2.32 81 -1.31 0.19
GREMLIN Intercept 3133.90 905.92 130 3.45 0.00
Trgscore -450.20 483.01 43 -0.93 0.35
time 35.10 38.70 130 0.90 0.36
MMP19 Intercept 738.42 245.00 111 3.01 0.001
Trgscore -149.77 129.36 39 -1.15 0.25
time -1.13 6.53 111 -0.17 0.86
TGFbetaRIII Intercept 33.82 11.50 59 2.93 0.00
Trgscore 4.05 6.08 34 0.66 0.50
time -0.46 0.52 59 -0.87 0.38
162 Bayesian Approaches in Oncology Using R and OpenBUGS

TABLE 12.3: Model selection criteria obtained by nlme package in R

Outcome AIC BIC logLik

CLC 1363.58 1377.60 -676.79

GREMLIN 2927.24 2943.00 -1458.62
MMP19 2018.83 2033.88 -1004.41
TGFbetaRIII 727.74 740.40 -358.87

TABLE 12.4: Estimates obtained through OpenBUGS code by 20,000 iter­

ation
Parameter Mean SD 95% Credible Interval Median
β11 14.16 3.33 ( 7.33 , 20.79 ) 14.15
β12 0.04 0.16 ( -0.27 , 0.38 ) 0.03
β13 -0.11 0.22 ( -0.59 , 0.30 ) -0.10
β14 0.93 1.74 ( -2.62 , 4.46 ) 0.94
β15 0.53 0.90 ( -1.27 , 2.39 ) 0.53
β16 0.54 1.89 ( -3.08 , 4.47 ) 0.51
β21 -38.18 9.42 ( -49.18 , -15.14 ) -40.98
β22 -0.14 0.44 ( -1.01 , 0.73 ) -0.14
β23 -0.43 0.42 ( -1.27 , 0.39 ) -0.44
β24 -0.19 0.20 ( -0.60 , 0.22 ) -0.19
β25 -0.82 0.48 ( -1.76 , 0.13 ) -0.83
tauz 0.025 0.00 ( 0.01 , 0.04 ) 0.02133
r1 -0.02 1.41 ( -3.26 , 2.89 ) -0.00
r2 -0.32 3.36 ( -8.83 , 6.54 ) -0.20
Joint Longitudinal and Survival Analysis 163

OpenBUGS Code on Joint Longitudinal and Survival Model

model {
for (i in 1:N) {
for (j in 1:M) {
Y[i, j] ∼ dnorm(muy[i, j], tauz)
muy[i, j]<­ beta1[1]+beta1[2] ×
t[j]+beta1[3] × t[j]× trgscore[i]+beta1[4]× ctcaegrade[i]
+beta1[5]× tstage[i]+beta1[6] × yptstage[i]+U[i,1]+ U[i,2] × t[j] }
for (j in 1: M) {
yp1[i,j] ∼ dnorm(muy[i, j], tauz)
r12[i,j]<-Y[i,j]-yp1[i,j]
sqr1[i,j]<-r12[i,j]× r12[i,j]
}
surt[i] ∼ dweib(p,mut[i]) I(surt.cen[i],)
log(mut[i])<-beta2[1]+beta2[2]*trgscore[i]+
beta2[3]× ctcaegrade[i]+beta2[4]× tstage[i]+
beta2[5]× yptstage[i]+r1*U[i,1]+r2 × U[i, 2]
U[i,1:2] ∼ dmnorm(U0[],tau[,])
surP[i] ∼ dweib(p,mut[i]) I(surt.cen[i],)
surR[i]<-surt[i]-surP[i]
sqr2[i]<-surR[i]× surR[i]
}
mspe1<-mean(sqr1[,])
mspe2<-mean(sqr2[])
p ∼ dgamma(1,1) # p <­ 1
sigmaz<-1/tauz
tau[1:2,1:2]<-inverse(sigma[,])
sigmab1 ∼ dunif(0,100) # priors
sigmab2 ∼ dunif(0,100)
cor ∼ dunif(-1,1)
sigma[1,1] <­ pow(sigmab1,2)
sigma[2,2] <­ pow(sigmab2,2)
sigma[1,2] <­ cor ×sigmab1∗ sigmab2
sigma[2,1] <­ sigma[1,2]
beta1[1:6] ∼ dmnorm(betamu1[],Sigma1[,])
tauz ∼ dgamma(0.1, 0.1)
beta2[1:5] ∼ dmnorm(betamu2[],Sigma2[,])
r1 ∼ dnorm(0, 0.01)
r2 ∼ dnorm(0, 0.01)
164 Bayesian Approaches in Oncology Using R and OpenBUGS

#Importing Data

list(N=45,M=4,betamu1=c(0,0,0,0,0,0),
betamu2=c(0,0,0,0,0),Sigma1=structure(.Data=
c(0.01,0,0,0,0,0,0,0.01,0,0,0,0,0,0,0.01,0,0,0,0,0,0,
0.01,0,0,0,0,0,0,0.01,0,0,0,0,0,0,0.01),.Dim=c(6,6)),
Sigma2=structure(.Data=c(0.01,0,0,0,0,0,0.01,0,0,0,
0,0,0.01,0,0,0,0,0,0.01,0,0,0,0,0,0.01),.Dim=c(5,5)),
U0=c(0,0),R=structure(.Data=c(100, 0, 0, 100),
.Dim=c(2,2)),t=c(0,2,6,12),Y=structure
(.Data=c(17.53,25.60,19.85,17.82,25.68,17.23,25.30,
18.30,16.26,18.01,19.47,19.98,18.00,15.96,12.57,20.46,
19.68,13.23,12.93,8.49,8.19,7.03,8.14,7.84,14.67,17.61,
17.31,15.30,6.422,5.22,6.51,7.85,7.55,4.55,13.89,13.59,
15.00,18.34,19.95,18.92,14.81,16.40,16.10,12.77,8.32,
8.02,6.82,6.52,18.62,10.67,10.37,14.46,12.97,13.81,
12.98,10.18,9.88,12.68,13,12.70,16.34,13.94,14.56,
14.26,9.79,11.00,9.89,9.70,40.46,40.16,41.43,32.59,
11.33,8.61,10.07,12.32,8.39,10.74,10.44,12.02,8.09,
25.24,18.73,22.05,21.75,24.94,8.75,8.28,15.90,15.60,
8.45,8.93,55.56,28.12,44.68,30.08,21.51,19.15,20.05,
29.78,31.48,34.47,28.01,34.69,15.85,17.00,16.8115.30,
15.55,16.70,16.51,18.69,8.45,7.07,15.60,12.41,11.17,
8.63,10.97,27.82,44.38,11.82,12.62,11.18,7.24,7.63,
7.257,8.23,19.38,20.87,19.83,16.70,4.85,15.00,6.63,
6.37,16.98,18.39,16.67,19.19,20.60,23.51,19.36,19.65,
13.98,12.75,13.55,9.75,17.61,20.68,6.94,22.06,7.20,
7.93,6.68,8.74,16.82,16.73,15.96,15.84,3.73,33.37,
12.31,22.89,16.37,18.89,20.30,23.21,19.94,6.77,26.25,
12.45,13.25,9.45,17.31,20.38,32.26,23.80,18.68,35.64),
.Dim=c(45,4)),surt=c(2752,NA,2731,2731,2736,2719,NA,
2695,2668,NA,2645,2611,2603,2536,2492,2467,NA,NA,2393,
NA,NA,2352,2327,2255,2240,NA,2234,2199,2164,2164,NA,NA,
2088,2059,2031,NA,2010,1982,1961,1919,NA,NA,NA,1485,NA),
surt.cen=c(0,319,0,0,0,0,19,0,0,2668,0,0,0,0,0,0,1110,
2432,0,21,493,0,0,0,0,614,0,0,0,0,1067,692,0,0,0,246,0,
0,0,0,160,200,365,0,1445),trgscore=c(0,1,1,0,1,0,0,1,0,
1,0,1,0,1,1,0,0,0,1,0,1,1,1,0,0,1,0,1,1,0,0,1,1,0,1,0,0,
1,0,1,0,0,0,1,0),ctcaegrade=c(1,1,1,0,0,0,0,0,1,1,1,1,0,
1,1,1,1,1,1,1,0,1,1,1,1,0,0,0,1,1,1,0,0,0,0,0,0,0,1,0,0,
1,0,0,0),tstage=c(4,1,2,3,1,3,2,4,1,2,4,4,4,2,1,3,3,3,2,
1,2,2,4,4,1,4,2,1,1,2,2,2,1,4,4,4,4,4,4,2,2,4,3,3,1),
yptstage=c(1,0,1,1,1,1,1,0,1,1,1,0,0,0,1,1,0,1,1,1,1,0,
0,0,1,0,0,1,1,,1,1,1,1,0,1,1,0,0,0,1,0,0,0,1))
Joint Longitudinal and Survival Analysis 165

#Importing Prior Initial Value

list(beta1=c(0,0,0,0,0,0),
tauz=1,beta2=c(0,0,0,0,0),r1=0,r2=0,U=structure(.Data=c(0,
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
0,0,0,0,0),.Dim=c(45,2)))

12.7 Different R Package for Joint Longitudinal Model

12.7.1 joint.Cox
This package is useful for Joint Frailty-Copula Models for Tumor Progres­
sion and Death in Meta-Analysis. It used to fit the survival data and perform
dynamic prediction under joint frailty-copula models for tumor progression
and death.

FIGURE 12.8: jointCox estimates on log-likelihood.

166 Bayesian Approaches in Oncology Using R and OpenBUGS

#R Code on joint.Cox Function

library(“joint.Cox”)
data(dataOvarian)
t.event=dataOvarian$t.event
event=dataOvarian$event
t.death=dataOvarian$t.death
death=dataOvarian$death
Z1=dataOvarian$CXCL12
group=dataOvarian$group
alpha given=0
kappa grid=seq(10,1e+17,length=30)
set.seed(1)
jointCox.indep.reg(t.event=t.event,event=event,
t.death=t.death,death=death,Z1=Z1,Z2=Z1,
group=group,alpha=alpha given,kappa1=kappa grid,
kappa2=kappa grid,LCV.plot=TRUE,Adj=500)

12.7.2 JM
This JM package is useful for Joint Modeling of Longitudinal and Survival
Data.
#R Code on JM Function

library(“JM”)
fitLME <­ lme(log(serBilir) ∼ drug × year,
random = ∼ 1 | id, data = pbc2)
fitSURV <­ survreg(Surv(years, status2) ∼ drug,
data = pbc2.id, x = TRUE)
fitJOINT <­ jointModel(fitLME, fitSURV, timeVar =“year”)
plot(fitJOINT, 3, add.KM = TRUE, col =“red”,lwd = 2)
par(mfrow = c(2, 2))
plot(fitJOINT)
Joint Longitudinal and Survival Analysis 167

FIGURE 12.9: JM package estimate of residuals.

Chapter 13

Covariance modeling

Abstract
The mixed-effect model is useful to analyze the longitudinal data. Conven­
tionally, the patient-specific changes are considered as random. The random
effect model is supposed to perform the analysis. However, the random effect
model found that the correlation between all the observations is the same. It
assumed that the correlation coefficient is the same for all the patients who
are the same as well. Nevertheless, these assumptions are not correct always.
Correlation always depends on the time gap between the observations. It is pa­
tients specific. It is better to use a random coefficient model while the response
variable is linked as the time of interest. It suits to allow the random slopes
varies between the patients. Different covariance structures are detailed. Co­
variance structure handling techniques are illustrated with R and OpenBUG
software.

13.1 Introduction
It is assumed that the patients will have different covariance, and the ran­
dom effect model can not capture it. The covariance structure can be defined
separately from random effects. It helps to consider the specific pattern of
covariance for each visit. Sometimes, visit wise observations are considered as
same for patients. Now for a clinical trial, a specific pattern across the periods
could be defined same between the observations appeared on the same pa­
tients. We do define the covariance pattern by matrix R. This matrix provides
that the observations are correlated. It is presented as
⎛ ⎞
R1 0 0 0 0
⎜ 0 R2 0 0 0⎟
⎜ ⎟
R=⎜ ⎜ 0 0 R 3 0 0 ⎟
⎟ (13.1)
⎝0 0 0 R4 0 ⎠
0 0 0 0 R5

169
170 Bayesian Approaches in Oncology Using R and OpenBUGS

13.2 Covariance Patterns

The mixed-effect model is suitable to work with different types of co­
variance pattern. The pattern of covariance dependent on fixed times. It is
easy to define the covariance pattern, while observations are equally spaced.
Sometimes, the time intervals are not equal, and covariance patterns become
challenging. The commercial software mixed effect modeling code can han­
dle different types of covariance pattern. Suppose there are four-time points
observations. Ri matrices define the covariance pattern. The general trend is
defined as (i). The variance of the responses can be defined as σi2 . It may
vary from time to time. With the covariance θjk , between the pair of period
j and k. Now the first-order autoregressive model defined as covariance de­
clined exponentially from time point j to k as θjk = ρj−k σ 2 . It considered that
the periods are equally spaced. It defined as the ’natural’ model. Sometimes,
observations are not equally spaced. For example, after delivering chemother­
apeutic drug patients may not be observed with equal time points. It may
be required to keep them rapidly with unequal time intervals. Different types
of covariance patterns are like (I) General, (II) First Order, (III) Compound
symmetry, and (IV) Toeplitz.
(I) General
⎛ 2 ⎞
σ1 θ12 θ13 θ14
⎜θ12 σ22 θ23 θ24 ⎟
R=⎜ ⎝θ13 θ23 σ32 θ34 ⎠
⎟ (13.2)
2
θ14 θ24 θ34 σ4
(II) First Order

σ12
⎛ ⎞
θ12 θ13 θ14
⎜θ12 σ22 θ23 θ24 ⎟
R=⎜ ⎟ (13.3)
⎝θ13 θ23 σ32 θ34 ⎠
θ14 θ24 θ34 σ42
(III) Compound symmetry

σ12
⎛ ⎞
θ12 θ13 θ14
⎜θ12 σ22 θ23 θ24 ⎟
R=⎜ ⎟ (13.4)
⎝θ13 θ23 σ32 θ34 ⎠
θ14 θ24 θ34 σ42
(IV) Toeplitz

σ2
⎛ ⎞
θ1 θ2 θ3
⎜ θ1 σ2 θ1 θ2 ⎟
R=⎜ ⎟ (13.5)
⎝θ13 θ23 σ32 θ34 ⎠
θ14 θ24 θ34 σ42
Covariance modeling 171

There are two methods to handle the covariance pattern for repeatedly mea­
sure data: (I) Random effect model, (II) Multivariate normal distribution. The
random effect model can address the compound symmetry pattern. However, it
is required to perform the sphericity test before assuming the compound sym­
metry model. In case the sphericity test failed then it is necessary to achieve
multivariate normal distribution. The multivariate normal distribution can
consider the general covariance pattern. The multivariate normal distribution
can address several covariance patterns. Perhaps, missing data inflated data
can not be dealing with the multivariate normal distribution. Multiple impu­
tation techniques are needed to overcome the missing values, and after that,
covariance patterns can be taken care of suitable modeling. The suitable co­
variance pattern is general covariance if the measurements differ between the
time points. Different variance for repeatedly measured data can be addressed
by (V) Heterogeneous uncorrelated, (VI) Heterogeneous Compound Symme­
try, (VII) Heterogeneous first-order autoregressive, and (VIII) Heterogeneous
Toeplitz. (V) Heterogeneous uncorrelated
⎛ 2 ⎞
σ1 0 0 0
⎜ 0 σ22 0 0⎟
Ri = ⎜⎝0 2
⎟ (13.6)
0 σ3 0 ⎠
0 0 0 σ42
(VI)Heterogeneous Compound Symmetry.

σ12 ρσ1 σ2 ρ2 σ1 σ3 ρ3 σ1 σ4
⎛ ⎞
⎜ ρσ1 σ2 σ22 ρσ2 σ3 ρ2 σ2 σ4 ⎟
Ri = ⎜
⎝ρ2 σ1 σ3 ρσ2 σ3
⎟ (13.7)
σ32 ρσ3 σ4 ⎠
2
ρσ1 σ4 ρ σ2 σ4 ρσ3 σ4 σ42
(VII) Heterogeneous First-Order Autoregressive.

σ12 ρσ1 σ2 ρ2 σ1 σ3 ρ3 σ1 σ4
⎛ ⎞
⎜ ρσ1 σ2 σ22 ρ2 σ2 σ3 ρ2 σ2 σ4 ⎟
Ri = ⎜ 2
⎟ (13.8)
⎝ρ σ1 σ3 ρσ2 σ3 σ32 ρσ3 σ4 ⎠
3 2
ρ σ1 σ4 ρ σ2 σ4 ρ1 σ3 σ4 σ42
(VIII) Heterogeneous Toeplitz

σ12
⎛ ⎞
ρ1 σ1 σ2 ρ2 σ1 σ3 ρ3 σ1 σ4
⎜ρ1 σ1 σ2 σ22 ρ1 σ2 σ3 ρ2 σ2 σ4 ⎟
Ri = ⎜ ⎟ (13.9)
⎝ρ2 σ1 σ3 ρ1 σ2 σ3 σ32 ρ1 σ3 σ4 ⎠
ρ3 σ1 σ4 ρ2 σ2 σ4 ρ1 σ3 σ4 σ42
The covariance structure sometimes changes differently between the thera­
peutic arms. It may be possible that the measurements are more correlated
172 Bayesian Approaches in Oncology Using R and OpenBUGS

among arm A than arm B. The covariance matrix formulated as

⎛ 2 ⎞
σA θA θA 0 0 0 0 0 0
⎜ θA 2
⎜ σA θA 0 0 0 0 0 0⎟ ⎟
⎜ θA 2
⎜ θA σA 0 0 0 0 0 0⎟ ⎟
⎜0 2
⎜ 0 0 σB θB θB 0 0 0⎟ ⎟
2
R=⎜ ⎜0 0 0 θB σB θB 0
2
0 0⎟ ⎟ (13.10)
⎜0 0 0 θB θB σB 0 0 0⎟
⎜ 2
⎟
⎜0 0 0 0 0 0 σA θA θA ⎟
⎜ 2
⎟
⎝0 0 0 0 0 0 θA σA θA ⎠
2
0 0 0 0 0 0 θ A θ A σA
It may possible that the covariance pattern is negligible in nature. Some time
several repeated measurements make the correlation structure poor. The Ri
matrix by setting the ’band’ correlation helps to work with the correlation
pattern. The band with size 3 is provided below. This practice reduces the
correlation parameter into minimal size.
⎛ 2 ⎞
σ1 θ12 θ13 0
⎜θ12 σ22 θ23 θ24 ⎟
Ri = ⎜ 2
⎟ (13.11)
⎝θ13 σ23 σ32 θ34 ⎠
0 θ24 θ34 σ42
There are different covariance patterns available. It is difficult to choose the
appropriate one. The approach is to select the most suitable one that fits with
the data. It also generated with standard error with fixed effect. Sometimes the
presence of more covariance parameters may increase the chance of overfitting.
Likelihood testing is useful to explore the amount of overfitting. Different types
of covariance patterns are available. It is not easy to select an appropriate one.
It is suitable to define the best covariance by fitting the data. The standard
error for the fixed effects model estimate is useful to understand the covariance
pattern. The model diagnostic criteria are useful to test the best covariance
structure. It is defined as
AIC = log(L) − q (13.12)
Now the covariance pattern is defined by q. Suppose the fixed effect is p
represent the number of observations. The total number of observations are
N . It is presented as
SIC = log(L) − (qlog(N − p))/2 (13.13)
Models presented with the lower value of AIC and SIC define as better fits.
This model selection criterion is available in the R software package ’nlme’.
Alternatively, the likelihood ratio test is useful to understand the covariance
parameters are important or not. The likelihood ratio test of the statistics is
defined as
2(log(L1 ) − log(L2 )) ∼ χ2
DF (13.14)
Now the difference in the number of covariance parameter is fitted by DF.
Covariance modeling 173

13.3 Covariance Patterns Challenges

It is always difficult to perform a test to understand a large covariance
structure. The best step would be to start with compound symmetry or first-
order autoregressive modeling. The task is to look about intricate patterns
present or not. The appearance of the sophisticated design is fixed by likeli­
hood value. However, it is possible to have a complex model while the dataset is
large enough. However, large datasets always required to understand intricate
covariance patterns. Now the covariance pattern is not of intrinsic interest.
The compound symmetry pattern may be adequate to estimate the treatment
effect and standard errors. Now model checking is required to explore the pat­
tern. Now if the difference is small, then the compound symmetry pattern can
be used with reasonable confidence.

13.4 Illustration with Protein-Gene Expression Time-

Course Data
#Running the Function as Gamma Distribution

library(“blme”)
data(“sleepstudy”, package=“lme4”)
fm1<-blmer(Reaction ∼ Days+(0+Days|Subject),sleepstudy,
cov.prior=gamma)
summary(fm1)
174 Bayesian Approaches in Oncology Using R and OpenBUGS

# R Output on Gamma Distribution

Cov prior : Subject ~ gamma(shape = 2.5, rate = 0,

posterior.scale = sd, common.scale = TRUE)
Prior dev : 3.9858

Linear mixed model fit by REML ['blmerMod']

Formula: Reaction ~ Days + (0 + Days | Subject)
Data: sleep study

REML criterion at convergence: 1766.6

Scaled residuals:
Min 1Q Median 3Q Max
-3.5411 -0.5565 0.0534 0.6239 4.6236

Random effects:
Groups Name Variance Std. Dev.
Subject Days 58.49 7.648
Residual 833.79 28.875
Number of obs: 180, groups: Subject, 18

Fixed effects:
Estimate Std. Error t value
(Intercept) 251.405 4.000 62.847
Days 10.467 1.952 5.362

Correlation of Fixed Effects:

(Intr)
Days -0.324
Covariance modeling 175

#Running the Function as Gamma Distribution

fm2 <­ blmer(Reaction ∼ Days + (0 + Days|Subject),

sleepstudy,cov.prior = gamma(shape = 2, rate = 0.5,
posterior.scale =‘sd’))
summary(fm2)

# Gamma Prior with Different Rate and Scale

Cov prior : Subject ~ gamma(shape = 2, rate = 0.5,

posterior.scale = sd, common.scale = TRUE)
Prior dev : 5.7375

Linear mixed model fit by REML ['blmerMod']

Formula: Reaction ~ Days + (0 + Days | Subject)
Data: sleep study

REML criterion at convergence: 1766.6

Scaled residuals:
Min 1Q Median 3Q Max
-3.5281 -0.5575 0.0533 0.6242 4.6146

Random effects:
Groups Name Variance Std. Dev.
Subject Days 55.9 7.477
Residual 837.3 28.935
Number of obs: 180, groups: Subject, 18

Fixed effects:
Estimate Std. Error t value
(Intercept) 251.405 4.009 62.717
Days 10.467 1.916 5.464

Correlation of Fixed Effects:

(Intr)
Days -0.330
176 Bayesian Approaches in Oncology Using R and OpenBUGS

#Prior as Inverse Gamma Distribution

fm3 <­ blmer( Reaction ∼ Days + (0 + Days | Subject),

sleepstudy,cov.prior =invgamma(shape = 0, scale = 0,
posterior.scale =‘sd’))
summary(fm3)

#Inverse Gamma Prior with Different Rate and Scale

Cov prior : Subject ~ gamma(shape = 0, scale = 0,

posterior.scale = sd, common.scale = TRUE)
Prior dev : -Inf

Linear mixed model fit by REML ['blmerMod']

Formula: Reaction ~ Days + (0 + Days | Subject)
Data: sleep study

REML criterion at convergence: 1893.7

Scaled residuals:
Min 1Q Median 3Q Max
-2.3231 -0.5760 0.0324 0.5479 2.9331

Random effects:
Groups Name Variance Std. Dev.
Subject Days 0 0.00
Residual 2277 47.71
Number of obs: 180, groups: Subject, 18

Fixed effects:
Estimate Std. Error t value
(Intercept) 251.405 6.610 38.033
Days 10.467 1.238 8.454

Correlation of Fixed Effects:

(Intr)
Days -0.843
Covariance modeling 177

#Prior as Wishart Distribution

fm4 <­ blmer(Reaction ∼ Days + (1 + Days|Subject),

sleepstudy,cov.prior = wishart)
summary(fm4)

#Wishart Prior Distributionn

Cov prior : Subject ~ wishart(df = 4.5, scale = Inf,

posterior.scale = cov, common.scale = TRUE)
Prior dev : 4.1552

Linear mixed model fit by REML ['blmerMod']

Formula: Reaction ~ Days + (1 + Days | Subject)
Data: sleep study

REML criterion at convergence: 1743.8

Scaled residuals:
Min 1Q Median 3Q Max
-4.0094 -0.4665 0.0167 0.4698 5.2207

Random effects:
Groups Name Variance Std.Dev. Corr
Subject (Intercept) 646.16 25.420
Days 40.33 6.351 0.00
Residual 644.93 25.396
Number of obs: 180, groups: Subject, 18

Fixed effects:
Estimate Std. Error t value
(Intercept) 251.405 6.948 36.18
Days 10.467 1.636 6.40

Correlation of Fixed Effects:

(Intr)
Days -0.170
178 Bayesian Approaches in Oncology Using R and OpenBUGS

#Prior as Invwishart Distribution

(fm5 <­ blmer(Reaction ∼ Days + (1 + Days|Subject),

sleepstudy,cov.prior = invwishart(df=
5, scale = diag(0.5, 2)))
summary(fm5)

#Inverse Wishart Prior Distribution

Cov prior : Subject ~ invwishart(df = 5, scale = c(0.5

, 0, 0, 0.5),
posterior.scale = cov, common.scale = TRUE)
Prior dev : 1.7533

Linear mixed model fit by REML ['blmerMod']

Formula: Reaction ~ Days + (1 + Days | Subject)
Data: sleep study

REML criterion at convergence: 1743.7

Scaled residuals:
Min 1Q Median 3Q Max
-3.9930 -0.4627 0.0128 0.4683 5.2088

Random effects:
Groups Name Variance Std.Dev. Corr
Subject (Intercept) 634.52 25.190
Days 38.59 6.212 0.04
Residual 647.68 25.450
Number of obs: 180, groups: Subject, 18

Fixed effects:
Estimate Std. Error t value
(Intercept) 251.405 6.905 36.408
Days 10.467 1.606 6.517

Correlation of Fixed Effects:

(Intr)
Days -0.145
Covariance modeling 179

#Prior as PenaltyFn

penaltyFn <­ function(sigma)

dcauchy(sigma, 0, 10, log = TRUE)
fm6 <­ blmer(Reaction ∼ Days + (0 + Days|Subject),
sleepstudy,cov.prior = custom(penaltyFn,
chol = TRUE, scale = “log”))
summary(fm6)

# Prior as PenaltyFn

Cov prior : Subject ~ custom(fn = penaltyFn, chol = TRUE,

scale = log, common.scale = TRUE)
Prior dev : 6.8959

Linear mixed model fit by REML ['blmerMod']

Formula: Reaction ~ Days + (0 + Days | Subject)
Data: sleep study

REML criterion at convergence: 1766.5

Scaled residuals:
Min 1Q Median 3Q Max
-3.5104 -0.5588 0.0541 0.6244 4.6022

Random effects:
Groups Name Variance Std. Dev.
Subject Days 52.7 7.26
Residual 842.0 29.02
Number of obs: 180, groups: Subject, 18

Fixed effects:
Estimate Std. Error t value
(Intercept) 251.405 4.020 62.539
Days 10.467 1.869 5.599

Correlation of Fixed Effects:

(Intr)
Days -0.340
180 Bayesian Approaches in Oncology Using R and OpenBUGS

#Prior as Normal Distribution as Fixed Effect

fm7 <­ blmer(Reaction ∼ Days + (1 + Days|Subject),

sleepstudy,cov.prior = NULL,fixef.prior = normal)
summary(fm7)

#Normal Distribution with Fixed Effect as Prior Distribution

Fixef prior: normal(sd = c(10, 2.5), corr = 0,

common.scale = TRUE)
Prior dev : 24.0661

Linear mixed model fit by REML ['blmerMod']

Formula: Reaction ~ Days + (1 + Days | Subject)
Data: sleep study

REML criterion at convergence: 1743.6

Scaled residuals:
Min 1Q Median 3Q Max
-3.9688 -0.4628 0.0222 0.4674 5.1974

Random effects:
Groups Name Variance Std.Dev. Corr
Subject (Intercept) 613.67 24.772
Days 35.13 5.927 0.06
Residual 650.42 25.503
Number of obs: 180, groups: Subject, 18

Fixed effects:
Estimate Std. Error t value
(Intercept) 251.229 6.822 36.825
Days 10.467 1.545 6.773

Correlation of Fixed Effects:

(Intr)
Days -0.137
Covariance modeling 181

#Prior as Normal Distribution

fm8 <­ blmer(Reaction ∼ Days + (1 + Days|Subject),

sleepstudy,cov.prior = NULL,fixef.prior=
normal(cov = diag(0.5, 2), common.scale = FALSE))
summary(fm8)

#Normal Prior Distribution

Fixef prior: normal(sd = c(0.7071, 0.7071),

corr = 0, common.scale = FALSE)
Prior dev : 2.2959

Linear mixed model fit by REML ['blmerMod']

Formula: Reaction ~ Days + (1 + Days | Subject)
Data: sleep study

REML criterion at convergence: 1830.7

Scaled residuals:
Min 1Q Median 3Q Max
-3.9442 -0.4592 -0.0049 0.4604 5.1673

Random effects:
Groups Name Variance Std.Dev. Corr
Subject (Intercept) 63677.9 252.34
Days 141.6 11.90 0.88
Residual 655.0 25.59
Number of obs: 180, groups: Subject, 18

Fixed effects:
Estimate Std. Error t value
(Intercept) 0.03354 0.70696 0.047
Days 0.04581 0.64477 0.071

Correlation of Fixed Effects:

(Intr)
Days 0.007
Part IV

Bayesian in Diagnostics

Test Statistics

183
Chapter 14

Bayesian Inference in Mixed-Effect

Model

Abstract
It becomes an attractive choice to get a higher blockade of cancer cell pro­
gression by Molecular Targeted Agents (MTA). The statistical methodology
for MTA has required attention. The literature on early phase dose-finding
studies with MTA is minimal. The dose-finding design in the MTA setting is
attempted by continuously updating the influence of the biological agent. In
this chapter, the Bayesian approach is presented to perform the linear mixed
effect models. The model has performed with the Markov chain Monte Carlo
(MCMC) method. It provides an algorithm to assign different possible doses
for each subject. It helps to define the optimum treatment on MTA value.
It is potent enough to accumulate the doses information by capturing their
variability, an overlooked area in the conventional model. A simulation study
is performed to create dose values. This work is suitable for Phase-I dose with­
out the MTD or DLT. The intention is to look for specific biomarker value
by contributing effect of different dose. The illustration in this work will help
to build a dose-response curve to detect the best effective treatment through
MTA in a group of patients.

14.1 Introduction
The likelihood estimation is one way to perform a linear mixed effect
model. However, another alternative is Bayesian inference. Earlier, it was chal­
lenging to achieve Bayesian mixed-effect model through R. However, R with
“glm” and “lmer” package becomes a useful tool to perform a mixed-effect
model with R. But these two packages helpful to work with likelihood-based
models. It is a fact that Bayesian is not a replacement of maximum likelihood
estimation. However, Bayesian is easy to perform in recent years. It requires
to the known fundamental concept of Bayesian before application. Hypothesis

185
186 Bayesian Approaches in Oncology Using R and OpenBUGS

testing is not the same in Bayesian inference. The likelihood-based method

obtained by “lme4” for the parameter θ and data y looks by Bayesian theorem

p(θ|y)∞l(θ|y)p(θ) (14.1)

The term p(θ|y) used to define the posterior probability. We can only un­
derstand the parameter from the data. It can give us predictions. The prior
probability is p(θ). The specification of the prior is the challenging task to
explore Bayesian modeling. The conventional dose-finding trial is based on
controlling the tolerable dose. Now the cytotoxic agents are tested about the
maximum tolerated dose (MTD). Now it is anticipated that drug response
will be higher through high doses. However, the toxicity will also increase
with increased dose. Similarly, the probability of toxicity will also incline. The
method used to identify the best effective dose is well known as rule-base
as a 3 + 3 design. Now the model-based design can be classified as Contin­
ual Reassessment Method (CRM) [74, 75], and modified CRM [76, 77, 78],
Time-to-Event CRM (TITE-CRM) [79] and Escalation with Overdose Con­
trol (EWOC) design [80]. Now the challenge is that a small sample size needs
to be lesser to take any decision about drug-dose. The maximum tolerable dose
that any patient can tolerate is defined as Dose Limiting Toxicity(DLT). In
this continuation, it is required to promote the model-based designs [81, 82].
But making any decision about effective dose by looking MTD is not enough
[83, 84, 85, 86]. The drugs or other essences promote tumor progression [87].
It becomes an attractive choice to get a higher blockade of cancer cell progres­
sion by molecular targeted agents (MTA) .The transduction pathway by the
bypassing mechanism [88]. The statistical methodology for MTA is required to
apply. The literature on early phase dose-finding studies with MTA is minimal.
The dose-finding design in the MTA setting is attempted by continuously up­
dating the influence of the biological agent [89]. Recently, there is an attempt
to compare the dose-efficacy by curves [90]. The dose-toxicity modeling can
be performed by change-point analysis [91] is performed to estimate the best
effective dose in the MTA setup. Several methodologies are about balancing
between dose-efficacy and dose-toxicity. However, the preliminary assumption
that the increase of dose will be more efficient for treatment. Alternatively, it
is also observed that increasing dose can reduce the dose efficiency [89, 91]. It
is aimed to develop a novel dose-finding model of MTA. It is required to the
point that best effective dose in MTA or better called the optimum biological
dose (OBD). There are computationally possibilities to work with Bayesian
inference by a linear mixed effect model [92, 93, 94] to get the OBD. Ini­
tial attempts motivate us to propose a model with a mixed effect model set
up to deal with OBD. The dose-Response model is performed. Different dose
levels are defined as covariates on treatment efficacy. The autoregressive [95]
and dynamic linear mixed effect modeling [96] are considered. In contrast, a
flexible-dose regime is presented to treated patients to obtain the best effec­
tive dose. The non-linear varying coefficient is considered to run the linear
mixed-effect model. The best effective dose is obtained by a linear setting
Bayesian Inference in Mixed-Effect Model 187

[97]. Similarly, a mixed-effect model with a Gaussian process functional form

is adopted. The non-parametric Gaussian process component addresses the
fixed effect and parametric Gaussian as the random effect component [98].
The challenge is to link between earlier time point measurements on the cur­
rent dose level. The index value can be used to model the dependence between
response with time-varying predictor values [99] or by the coefficient of a re­
gression model as time-dependent functions [100]. The functional form to serve
the random effect mixed effect model is initiated [101]. The Gaussian process
regression model is obtained to estimate the functional form of random effect
[102, 103]. In this chapter, the Bayesian approach is presented to perform the
linear mixed effect models. The model has performed with the Markov chain
Monte Carlo (MCMC) method. Mixture distribution is adopted with point
mass and continuous distribution to select the variables as fixed or random
effects. The MCMC is used to obtain the posterior estimates. The primary
aim of dose-response studies is to classify the individual level response change
for different doses or agents. Based on this work we can build a dose-response
curve to detect the best effective dose through MTA in a group of patients.

14.2 Likelihood Function with Doses and Measurement

Process
The repeatedly measured continuous variable for each patients and corre­
sponding dose is defined as
(h)
Yi,t = (Yi,0 , Yi,1 , Yi,2 , ..., Yi,t )T (14.2)

The response value measured as the vector of responses from 0 to t for the
individuals (i = 1, ..., N ). The prior to baseline measurements are defined by
(h)
superscript (h) as Yi,t and baseline measurements by Yi,0 and the k-th time
point measurements are defined as Yi,k . Further, the administered doses are
defined as
(d)
xi,t = (Xi,1 , Xi,2 , ..., Xi,t )T (14.3)
for the time point from 1 to t. The doses are defined as time dependent
covariates. The doses administered between time points t − 1 and t are defined
as Xi,t . The time of the last measurement for the ith subject is defined as Ti .
The likelihood function is defined as
(h) (d)
L(θ, ψ) = f (Yi,Ti , Xi,Ti |Zi , θ, ψ) (14.4)

The parameter θ and ψ are adopted to define the measurement process and
dose process, respectively. The time independent covariates are defined by Zi ,
188 Bayesian Approaches in Oncology Using R and OpenBUGS

with the measurement process

Ti
(h) (d)
�
L(θ) = f (Yi,0 |Zi , θ) f (Yi,t |Yi,t−1 , Xi,t , Zi , θ) (14.5)
t=1

The likelihood for the dose process is generated as

Ti
(h) (d)
�
L(ψ) = f (Xi,1 |Yi,0 , Zi , ψ) f (Xi,t |Yi,t−1 , Xi,t−1 , Zi , ψ) (14.6)
t=2

The compact likelihood function is prepared as

L(θ, ψ) = L(θ) × L(ψ) (14.7)
The factorization process is separated as f (a, b) = f (a|b)f (b).
(h) (d)
f (Yi,Ti , Xi,Ti |Zi , θ, ψ) = f (Yi,0 , ..., Yi,Ti , Xi,1 , ....Xi,Ti |Zi , θ, ψ) (14.8)

(h) (d) (h) (d) (h) (d)

f (Yi,Ti , Xi,Ti |Zi , θ, ψ) = f (Yi,Ti |Yi,Ti −1 , Xi,Ti , Zi , θ) × f (Yi,Ti −1 , Xi,Ti |Zi , θ, ψ)
(14.9)
and so on.

14.3 Linear Mixed-Effects Model

The linear mixed-effect model can jointly handle fixed and random effect
models [104, 105]. Different assumptions are required. There is shallow adop­
tion of the mixed effect model for non-availability of user-friendly software.
However, it is adopted primarily by open-source software R [106, 107]. An­
other reason for avoiding the linear mixed-effect models is the presences of
complexity in comparison to the regression models. It requires a high level
of computational expertise [108]. The mathematical explanation is described
below about linear mixed-effect model.
The response variable is defined with vector Yi = (Yi,0 , Yi,1 , Yi,2 , ..., Yi,Ti )T for
the i-th (i = 1, ..., N ) subject measured at 0 to Ti . The linear mixed effect
model is defined as
Yi = Xi β + Zi bi + �i (14.10)
The β is a p × 1 vector with unknown-fixed effect, and Xi is matrix with
(Ti + 1) × p for fixed effect and bi is a q × 1 vector with unknown random-effect
parameters. The term Zi is a known, (Ti + 1) × q stands for design matrix
of random effects and �i stands for random errors for the vector with time
(Ti + 1) × 1. It is assumed that the term �i and bi follows the i.i.d with mean
zero and covariance matrix Ri and G, respectively.
Bayesian Inference in Mixed-Effect Model 189

14.4 Autoregressive Linear Mixed-Effects Model

The response is free with toxicity effects. The novelty of our method is to
decide the best effective dose for toxicity free OBD measurements. The para­
metric assumptions are not required. The auto-regressive covariance structures
are defined to declare the optimum dose. The conventional method is decided
based on DLT, and no patients are permitted to prescribe more than DLT once
if it is archived. However, our model is free from DLT module. Best doses are
decided from available all lower doses. Our proposed method provides scope
about intra-patient dose reductions at any time. It is an alternative about
the administration of multiple doses and therefore, the consequence of DLT
[109, 110]. The autoregressive linear model in this direction is defined as
Yi = ρFi Yi + Xi β + Zi bi + �i (14.11)
The time expresser matrix is (Ti + 1) × (Ti + 1). The elements below the
diagonal are 1 and above are 0. The earlier time point response is defined as
Fi Yi = (0, Yi,0 , Yi,1 , ..., Yi,Ti )T . The regression coefficient linked with earlier
response to post responses is ρ. The unconditional response vector is defined
as
Yi = (Ii − ρFi )−1 (Xi β + Zi bi + �i ) (14.12)
Here, Ii is a (Ti + 1) × (Ti + 1) identity matrix. The linear mixed effect model
extended with autoregressive model is prepared as
Yi = Xi∗ β ∗ + Zi bi + �i (14.13)
∗ T
where β = (β , ρ) and T
Xi∗ = (Xi , Fi Yi ).

14.5 Linear Mixed-Effects Model Compatible with Dose

Response
The mixed effect model compitable with repeatedly measured doses are
defined as
Yi,t = (βint + bint i ) + (βdose + bdose,i )Xi,t + �i,t (14.14)
The responses are measured with steady states by Yi,t and Xi,t for the re­
sponse levels and doses for i-th subject respectively. The fixed effect is defined
as β = (βint , βdose )T and the random-effects bi = (bint i , bdose i )T with
the intercept and the dose effect. Further, it is assumed that bi follows the
normal distribution with mean zero and covariance structure Gss . It is free
with baseline measurement and error term is defined by �i,t . The error term
�i,t follows mean zero and variance σ 2 .
190 Bayesian Approaches in Oncology Using R and OpenBUGS

14.6 Autoregressive Linear Mixed Effects Model Com­

patible with Dose Response
The autoregressive model is defined with dose modification by

Yi,t = βbase + bbase i + �i,t (t = 0) (14.15)

Yi,t = ρYi,t−1 + (βint + bint i ) + (βdose + bdose i )Xi,t + �i,t − ρ�i,t−1 (t > 0)
(14.16)
The term Yi,t and Xi,t are the representative of response and dose con­
tribution for the i-th subject at time point t, respectively. The fixed-
effect determined by β = (βbase , βint , βdose )T and random-effects by bi =
(bbase i , bint i , bdose i )T for the baseline, follow-up intercept, and follow-up
dose effect, respectively. It is assumed that the random-effect, i.e., bi , follows
the normal distribution with mean zero and covariance structure, i.e., unstruc­
tured as G. The assumption of �i considered as N (0, σ 2 ). The autoregressive
error is captured through �i . Further, the marginal representation of model
for (t > 0) is represented as
t
�
Yi,t = ρt (βbase +bbase i + ρt−j {(βint +bint i )+(βdose +bdose i )Xi,j +�i,j }
j=1
(14.17)
The dose effect at time point s on the response at time t(t ≥ s) is defined as
ρt−s (βdose + bdose i )Xi,s . It will decline while the time progressed. However, if
the dose X provided repeatedly with time, then the asymptotes for the popu­
lation mean and the ith subject is defined as (1 − ρ)−1 (βint + βdose X)and(1 −
ρ)−1 (βint + βdose + (βdose X + bdosei X) respectively.

14.7 Generating Data

There is no consensus on the most appropriate methodology on MTA re­
sponse for Metronomic Chemotherapy(MC). Therefore, a simulation study
was performed to assess the OBD about the performance of an MC. Datasets
were generated to resemble the skewed distributions seen in a motivating ex­
ample. The results obtained through the linear mixed effect model and au­
toregressive models are given in Table 14.2 and Table 14.3, respectively. A
total of 25 patients with different doses of treatments are presented. All of
the patients were exposed to four different visits to measure their MTA value.
The doses selected as Dose=15 mg/m2 , Dose=12.5 mg/m2 , Dose=10 mg/m2 ,
Dose=7.5 mg/m2 and Dose=5 mg/m2 respectively. The simulation techniques
Bayesian Inference in Mixed-Effect Model 191

were performed. Dose-effect relations are established by the MCMC iterations.

Two OpenBUGS programs were written for the linear mixed effect model and
autoregressive effect separately. A total of 20,000 iterations were used to gen­
erate convergence statistics. The credible intervals were given in Table 14.2
and Table 14.4, respectively.

TABLE 14.1: Different doses and corresponding mean(sd) MTA response

observed through simulated data.
Dose Time1 Time2 Time3 Time4
2
15 mg/m 9.60(3.08) 9.11(1.40) 8.28(1.36) 10.47(1.22)
12.5 mg/m2 15.65(1.69) 14.24(2.21) 14.90(1.68) 16.15(0.51)
10 mg/m2 19.35(2.70) 19.42(1.43) 20.03(1.55) 19.38(3.23)
7.5 mg/m2 13.37(2.09) 14.24(2.73) 14.07(1.45) 15.38(2.87)
5 mg/m2 10.79(1.71) 9.03(1.82) 10.24(1.57) 9.59(1.53)

TABLE 14.2: Posterior estimates generated through different models

through linear mixed-effect model on MTA response.
Doses Parameter Posterior Mean(SD) 95% HPD
2
15 mg/m β1 (Intercept) 9.36(0.81) (7.81,10.93)
β2 0.72(0.51) (-0.30,1.708)
σ12 0.85(2.31) (0.13,3.50)
12.5 mg/m2 β3 (Intercept) 14.91(0.79) (13.34,16.45)
β4 -0.046(0.52) (-1.10,1.00)
σ22 0.67(0.97) (0.12,2.60)
10mg/m2 β5 (Intercept) 20.47(0.84) (18.90,22.07)
β6 -0.43(0.52) (-1.40,0.55)
σ32 0.82(1.39) (0.13,3.48)
7.5 mg/m2 β7 (Intercept) 14.30(0.84) (12.76,15.85)
β8 0.0912(0.51) (-0.91,1.10)
σ42 0.79(1.08) (0.14,2.99)
5 mg/m2 β9 (Intercept) 11.23(0.79) (9.66,12.81)
β10 -0.76(0.52) (-1.77,0.22)
σ52 0.89(3.21) (0.13,3.39)
192 Bayesian Approaches in Oncology Using R and OpenBUGS

TABLE 14.3: Model selection criteria of the linear mixed-effect model.

Doses D̄ D̂ DIC pD
2
15 mg/m 83.14 77.84 88.45 5.301
12.5 mg/m2 82.96 77.5 88.41 5.451
10mg/m2 72.86 67.45 78.27 5.408
7.5 mg/m2 81.17 75.69 86.65 5.48
5 mg/m2 83.06 77.74 88.37 5.311

TABLE 14.4: Posterior estimates generated through different models

through autoregressive model on MTA response.
Doses Parameter Posterior Mean(SD) 95% HPD
2
15 mg/m β1 (Intercept) 4.84(6.48) (-9.98,10.96)
β2 1.70(1.76) (-0.13,5.96)
σ62 5.60(2.52) (2.57,11.99)
12.5 mg/m2 β3 (Intercept) 9.65(23.48) (-35.53,52)
β4 4.81(23.47) (-37.62,50)
σ72 2.06(4.80) (0.36,8.50)
10mg/m2 β5 (Intercept) 19.73(6.38) (3.46,26.87)
β6 -1.72(1.86) (-3.97,3.59)
σ82 46.89(27.91) (11.84,114)
7.5 mg/m2 β7 (Intercept) 11.87(4.30) (0.00,16.62)
β8 0.60(1.12) (-0.85,3.81)
σ92 6.12(5.71) (2.158,23.92)
5 mg/m2 β9 (Intercept) 11.26(2.86) (3.42,14.83)
β10 -0.52(0.90) (-1.73,1.88)
2
σ10 26.25(30.38) (2.66,101.6)

14.8 Computation Support

This work is illustrated on MTA measurements. MTA measured as gene
expressions occurred to measured as high-dimensional data. So it is better to
present any statistical methods that can be handled with the high-dimensional
setup as well. The computational flexibility toward high-dimensional data
analysis is limited. Markov chain Monte Carlo (MCMC) algorithms allow com­
putation flexibility to perform the linear mixed model [111]. There are several
attempts in last two decades to computationally enrich the power of compu­
tation. There are dedicated software like MCMCglmm [112, 113, 114, 115],
WinBUGS [116, 117]. However, the challenges faced by slow convergence in
high-dimensional correlated data [114, 118]). The conjugate priors are only
Bayesian Inference in Mixed-Effect Model 193

TABLE 14.5: Model selection criteria of the autoregressive model on MTA

response.

Doses D̄ D̂ DIC pD
2
15 mg/m 85.7 83.68 87.73 2.02
12.5 mg/m2 93.75 91.45 96.05 2.29
10mg/m2 20.78 20.25 21.31 0.53
7.5 mg/m2 83.28 81.13 85.42 2.14
5 mg/m2 84.67 83.39 85.94 1.27

available for Gibbs sampling to the likelihood of a parameter [114]. It is easy

to work Stan software by Hamiltonian Monte Carlo simulation [118, 119]. It
converges very fast, particularly for high-dimensional data [120]. The compu­
tational flexibility is obtained by collectively working with R, OpenBUGS and
Stan [113]. There are dedicated available packages like “brms” and “blme” in
R.

FIGURE 14.1: OBD achieved through different dose labels.

194 Bayesian Approaches in Oncology Using R and OpenBUGS

14.9 Power Analysis of the Performed Models

If a study is underpowered, then there is a possibility about the waste of
resources and the real effects may not be generated [110, 121]. Similarly, a
large study comes with overpowered and expensive than a real requirement
[122]. A sample size calculation is required to fulfil the research queries. There
are several dedicated packages those available with R to work with mixed
effect model and power analysis [123, 124, 125]). There is also the provision
to perform with power analysis of mixed effect by specifying the fixed and
random effect component [122]. The power analysis is obtained by “SIMR”
and “LME4” packages [106]. A total of 5 different doses were selected to gen­
erate the simulated data. The doses were 15mg/m2 , 12.5mg/m2 , 15mg/m2 ,
10mg/m2 , 7.5mg/m2 and 5mg/m2 respectively.The MTA value as a continu­
ous variable for 25 patients was generated, with 4-time points/visits. The MTA
was assumed to follow the normal distribution with mean ten and standard
deviation 2. The age interval was obtained from 21 to 45 years. The power
analysis was performed with “powers” function available in SIMR package with
R. A total of 10 times simulations were carried out to obtain the result. A 95%
confidence interval provides the power of predictor with 65%[50.32%, 80.85%]
range. The power of a selected linear mixed effect model is relatively robust.

14.10 Discussion
Our proposed method provides an algorithm to assign different possible
doses for each subject. It helps to define the optimum dose on MTA value.
It is potent enough to accumulate the doses information by capturing their
variability, an overlooked area in the conventional model. A simulation study
is performed to create dose values. This work is suitable for Phase-I dose with­
out the MTD or DLT. The intention is to look for specific biomarker value by
contributing effect of different doses [126]. The continual reassessment method
is found suitable to work with OBD. It is detected by highest toxicity label
[127]. The OBD detection technique is performed through two-stage clinical
trial design [128, 129]. Further, the dose-finding design is performed with OBD
by adaptive trial on molecularly targeted agents. The low-dose chemotherapy
is found suitable to suppress the tumor vessel growth and controlling the dam­
age in vascular endothelial cells [130]. The semiparametric approach proposed
to detect the whole dose levels of MC by OBD [131]. In survival analysis
with MCMC iteration method, the possible threshold value for OBD serves
the same challenge [132].The high dose of chemotherapy with cisplatin can
produce a more toxic effect [130]. The targeted MC on molecularly targeted
Bayesian Inference in Mixed-Effect Model 195

agent works to control the vascular endothelial cells [133]. It is mentioned

that new vessels generated in long-run for patients treated with the maximum
tolerated dose(MTD) by angiogenesis [134]. The MTD with conventional high
dose chemotherapy fails to control the growth of tumor cells by neovascu­
larization. This drug could cause tumor neovascularity [121]. However, since
the last decade, the serum creatinine is emerged as a promising surrogate
marker to assess the tumor anti-angiogenesis effect [135, 136, 137, 138]. The
low-dose chemotherapy is administered frequently into continuous manner to
selectively suppress the vessel growth in tumor tissue and repair the damage
in vascular endothelial cells (VECs) [130]. This proposed model is performed
by dose-response modeling by avoiding the toxicity label in a clinical trial.
The simulation study shows that the proposed model works relatively well. It
is better than the parametric model. Because in MC trial, the occurrence of
toxicity is limited.
#Model 1

model;
{
#prior distribution for beta
for(i in 1:10){beta[i]∼ dnorm(0,.0001) } #Dose=X
for(i in 1:N1){for(j in 1:M1)
{Y1[i,j] ∼ dnorm(mu1[i,j],tau)}
for(i in 1:N1){for(j in 1:M1)
{
mu1[i,j]<­ beta[1]+beta[2] ×time1[j]+
b[i,1]+b[i,2]×time1[j]}}
for (i in 1:N5){for (j in 1:M5)mu5[i,j]<­ beta[9]
+beta[10]×time4[j]+b[i,9]+b[i,10]×time5[j]}
for(i in 1:N1)b[i,1:10] ∼ dmnorm(vu[],Omega[,])}
for (
i in 1:10){vu[i]<-0
Omega[1:10,1:10]{ dwish(R[,],10)
Sigma[1:10,1:10]<-inverse(Omega[,])
tau ∼ dgamma(.00001,.00001)
Var<-1/tau
}
196 Bayesian Approaches in Oncology Using R and OpenBUGS

#Model 2

model;
{
beta[i] ∼ dnorm(0.0, 0.001) # Dose=X
for(i in 1:N1){ for (j in 2:M1)Y1[i,j] ∼ dnorm(mu1[i,j],tau) }}
for(i in 1:N1){for (j in 2:M1)
mu1[i,j]<-beta1[i]×(1-rho)+beta2[i]×(time1[j]­
rho ×time1[j-1])+rho×Y1[i,j-1]}}
for(i in 1:N1){ Y1[i,1] ∼ dnorm(mu1[i,1],tau)}
for(i in 1:N1){ mu1[i,1]<-beta1+beta2 ×time1[1]}
tau ∼ dgamma(.001,.001)
rho ∼ dbeta(1,1)
sigma<-1/tau
}
Chapter 15

Concordance Analysis

Abstract
One of the parameter to promote cancer research is dependent on diagnos­
tics technique expansion. However, every time new diagnostic technique need
to test in presences of an actual diagnostic test. Several times disease measured
by continuous measurement. Now concordance correlation coefficient (CCC)
is one of the tools in diagnostics test statistics. It performs as a tool in agree­
ment analysis. It supports continuous measurement. The continuous variable,
i.e., tumor size measured or observed with a different diagnostic procedure,
can be computed by CCC. This chapter shows the Bayesian extension of the
CCC for continuous data. Now Bayes factor is illustrated by real-life data to
define the best diagnostics tool. The approach illustrated in this work provides
the researchers with an opportunity to find out the most appropriate model
for specific data and apply CCC to fulfil the desired hypothesis.

15.1 Introduction
One of the significant proportion of oncology research is devoted to diag­
nostic research. The risk of misdiagnosis always stands in oncology. The diag­
nosis is not easy and straightforward. The complex diagnosis and differential
diagnosis needs to be similar to define conclusive remark about the presence
of the disease condition. Sometimes physicians make a diagnosis error. Some­
times two physicians disagree on a diagnosis among them. However, they look
at the same CT scan or MRI report. It becomes confusing while compare CT
scans with MRI reports. Alternatively, two physicians independently comment
same or different diagnosis tets. The disease status detection is not an easy
task. There is always a dilemma to define cancer staging. Due to rapid disease
progression and disease complexity, it is not an easy task to detect tumor stag­
ing. Now it requires to make a conclusive remark about diagnosis test. The
agreement comes while we define the diagnosis status. Since the tumor mea­
sured into size or dimension, so the tumor measurement data generated as a

197
198 Bayesian Approaches in Oncology Using R and OpenBUGS

continuous variable. The continuous variable, i.e., tumor size measured or ob­
served with a different diagnostic procedure can be concluded by concordance
correlation, in this chapter, the concordance correlation illustrated with tumor
size and tumor volume measurements. While the measured outcome defined
with the binary variable, then it is presented with kappa statistics. Sometimes,
the weighted kappa also useful. The ‘inter-rater agreement’ and ‘inter-device
agreement’ required to quantifying the extent of agreement between two or
more examiners or devices is of primary interest.Widely known as the intr­
aclass correlation measurement is a choice [139, 140, 141]. The application
of the within-subject coefficient of variation found attractive in this context
[142]. Now the Inter-rater agreement is important, for example, when different
raters evaluate the severity of a specific disease during a clinical trial, and the
reliability of each of their subjective evaluation is to determine by measuring
agreement among the raters [143].
#Concordance Correlation Coefficient

The concordance correlation coefficient ρc defined as

2ρσx σy
ρc = (15.1)
σx2 + σy2
+ (µx − µy )2

Now, µx and µy are the average of the two continuous variables. The
term σx2 and σy2 are the corresponding variance. Now ρ is the correlation
coefficient between the two variables.

#Concordance Correlation Coefficient for Paired Data

Similarly, it can be calculated from the paired data. If x and y variables

are observed into paired manner, then the concordance correlation co­
efficient becomes
2xy
ρ̂c = 2 2 (15.2)
x + y + (x̄ − ȳ)2
The pair of observations is n.

It is always required for new diagnosis procedure to diagnose cancer more

accurately. Now the challenge is to establish the accuracy of the new diagnostic
tool in presences of the existing one. The concordance correlation coefficient is
useful to quantify the agreement between two raters measured independently
on the same subject. A new diagnostic tool promotion becomes challenging in
the presence of an available tool. Now new tool needs to proof as authoritative
or as equal in the presence of diagnostic tool. Now the different diagnostic tool
is required to be compared to explore the agreement on measurement. Unless
the new diagnostic tools are performed equal or better, we can not use in
regular news. In this context, the concordance correlation coefficient becomes
useful. There are different tools available on continuous data like the coefficient
Concordance Analysis 199

of variation, t-test, intra-class correlation, Pearson correlation coefficient, and

least square analysis. However, these are not useful to work toward agreement
analysis. There are a few limitations. The Intra-class correlation coefficient as­
sumed that the readings between two observers are interchangeable. It is not
robust enough. Now the least-squares analysis is not suitable enough for the
null hypothesis in the presence of a residual error that is very large or small
[144]. Similarly, the Pearson correlation is also not enough in this context. Now
the concordance correlation coefficient (CCC) supports for differentiating the
concordance measurement [145]. The computation of the Z-transformation is
difficult. The methodological work supports that CCC provides similar to the
value of the Pearson correlation coefficient if the mean and variance of the
measurement of interest [144]. There is an illustration of the CCC [146]. It is
based on the covariance-based index. Now the stratified CCC [147] and the
generalized estimating equation is established [148]. The size of the tumor is
mainly concerned in any regular or experimental cancer therapy. Determina­
tion of tumor size by diagnostic procedures is an important key indicator for
any therapeutic success. Recently, several types of diagnostic procedures are
available for tumor size detection like computed tomography (CT) and mag­
netic resonance imaging (MRI) with advanced technology. Generally, pre and
post-therapy tumor size provide therapeutic results in any specific direction.
It is ideal that the same diagnostic procedure should be applied to detect the
tumor size before and after a therapeutic effect to reduce the diagnostic test­
ing variation. Simultaneously, same way interpretation about tumor size needs
to be performed. Broadly, two types of approaches are available for detection
of tumor size, either through tumor volume or by the maximum area covered
by the tumor. This chapter is about proposing the Bayesian counterpart to
compute CCC for continuous data.

15.2 Computational Methodology

This data is obtained through a brain cancer diagnosis. A total of 35
patients tumor size and volume is considered in this work. Now tumor volume
is measured with MRI and tumor size by CR scan. The measurements are
obtained at two-time points, i.e., pre and post-surgery about tumor volume
and size. In this work, the Bayesian counterpart of the concordance correlation
coefficient is illustrated in this chapter. Suppose the ith reading is defined for
N subjects in a study. The response vector is defined as Yi (i = 1, 2, ...N, i ≤ I)
for I × 1 vectors that contained the I readings. Now, different methods is
� �
defined as Yi and Yi� Yi� (1 ≤ I, i ≤ I, I �= i ) respectively. The expected
square different is E[(Yi − Yi� )2 ] defined as CCC as scaled between -1 and 1.
200 Bayesian Approaches in Oncology Using R and OpenBUGS

The CCC is defined as

E[(Yi − Yi� )2 ] 2σii�
ρCCC = 1 − = 2 � (15.3)
σi2 + σi2� + (µi − µi� )2 σi + σi2� + (µi − µi )2

Now E(Yi ) = µi , E(Yi� ) = µi� , σi2 = var(Yi ), σi2� = var(Yi� ),

σii� = cov(Yi , Yi� ) = σi σi� ρii� . Now the terms Sii� , s2i , s2i� , Y¯i and Y¯i� are con­
sidered as unbiased estimates of σij , σi2 , σi2� , µi and µi� respectively. Now the
expected of this estimates for k types of different methods is limited with
k = 2 for two scanner is
k −1 �k k−1 k
2 � 2 � �
E( Sii� ) = σii� (15.4)
k(k − 1) i=1 � k(k − 1) i=1 �
i =i+1 i =i+1

k k
1� 2 1� 2
E( Si ) = σ (15.5)
k i=1 k i=1 i
k−1 k k−1 k
1 � � 1 � � σ2
E( (Ȳi −Ȳi� )2 ) = (µi −µi� )2 + e (15.6)
k(k − 1) i=1 � k(k − 1) i=1 � n
i =i+1 i =i+1
�k−1 �k
Now, i=1 i� =i+1 (Y¯i − Ȳi� )2 is a biased estimator of
�k−1 �k 2
i� =i+1 (µi − µi ) .
�
i=1
The corresponding unbiased estimator is defined as

k−1 k
� � k(k − 1) 2
(Ȳi − Ȳi� )2 − σe = W − Z (15.7a)
i=1 i� =i+1
n
k−1
� k
�
W = (Ȳi − Ȳi� )2 (15.7b)
�
i=1 i =i+1
k−1 k
k(k − 1) � �
Z= (Si2 + Si2� − 2Sii� )2
n i=1 � i =i+1
(15.7c)
Finally, unbiased estimate of ρCCC is defined as ρ̄CCC

�k−1 �k
2 i=1 i� =i+1 Sii�
ρ̄CCC = (15.8)
P −Q
where
k
� k−1
� k
�
P = (k − 1) Si2 + (Ȳi − Ȳi� )2 (15.9)
i=1 i=1 i =i+1�
Concordance Analysis 201

k−1 k
k(k − 1) � �
Q= (Si2 + Si2� − 2Sii� ) (15.10)
n i � i =k+1
�
Suppose i and i represents different reading for the jth patients. The
method number assigned by k = 2, i.e., CT Scanner and MRI Scanner. The
responses are defined as Yij and Yi� j of the same patients � j � . Now the linear
model for two scanners are presented as

Yii� j = θ + βY + βi Yi + βi� Yi� + �ii� j (15.11)

Now the continuous outcome of the measured variable is represented as

Yii� j . The response measured for the jth individual, ith observation by kth
method. The overall mean is presented as θ with mean value tumor size Y as
fixed effect. The random effects are presented as Yi and Yi� as random effects.
The error term is presented as �ii� j ∼ N (0, τ ). Now the term �ii� j is stands
for precision for jth individual. The regression parameter of the fixed effect is
presented with N(0.0.0001) and random effects as Gamma(0.0.0001). Further,
the fixed effect variance, random effect variance and error terms are combined
with equation (15.7) as
k−1 k
2 � �
σf2 ixed = σij (15.12)
k(k − 1) � �
i +1 i =i+1

k−1 k
2 1 � �
σrandom = (µi − µi� )2 (15.13)
k(k − 1) i=1 �
i =i+1

and
k−1 k
2 2 � � 1 2
σerror = (σ + σi2� − 2σii� )2
k(k − 1) i=1 � 2 i
i =i+1
k k−1 k
1 � 2 � �
= σi2 − σ � (15.14)
k i=1
k(k − 1) i=1 i=i+1 ii

The CCC is defined as

�k−1 �k
σf2 ixed i=1 i� =i+1
ρCCC = 2 2 2
= �k �k
σf ixed + σrandom + σerror (k − 1) i=1 i� =i+1 (µi − µi� )2
(15.15)
The term θ is constant σf2 ixed = τ1f , σrandom
2
= τ1r and σerror
2
= τ1e . To
formulate the Bayesian analysis, we assign prior on the parameters, τf , τr , τe
as follows:
θ ∼ dnorm(0, 0.001) (15.16)
202 Bayesian Approaches in Oncology Using R and OpenBUGS

τf ∼ dgamma(0.0001, 0.0001) (15.17)

τr ∼ dnorm(0.0001, 0.0001) and (15.18)

τe ∼ dgamma(0.0001, 0.0001) (15.19)

15.3 Bayes Factor

Where dnorm stands for normal distribution, dgamma denotes the Gamma
distribution. The parameters of the distribution considered as non-informative
prior and fixed as 0 and 0.001 for normal distribution and 0.0001 and 0.0001
for Gamma distribution. The Bayesian factor is applied through JZS for Con­
cordance Correlation in regression line [148]. The regression coefficient β is
permitted to the application of JZS prior. The CCC, Intercept (θ), regres­
sion coefficients and error term (�ii� j ) are detailed in equation (15.7). Let the
equation (15.7) further been separated into Model (M1 ) and Model (M0 ) by

M1 : Y = θ + βX + �ii� j (15.20)

M0 : Y = θ + �ii� j (15.21)
The model (M1 ) states the presence of CCC and absence of it by model (M0 ).
Now, the Bayes Factor through JZS is defined [148, 149, 150] as,
∞
(n/2)1/2
�
(n−1)
BF10 = × (1 + g)(n−2)/2 × [1 + (1 − r2 )g]− 2 (15.22)
τ (1/2) 0

p(Y (M1 )
BF10 = (15.23)
p(Y (M0 )
If the value of BF10 becomes more than 1, it states about the presences
of CCC otherwise not. The statistical test can be performed with two Hy­
potheses: the Null Hypothesis, H0 as given in model (M0 ) and the alternative
hypothesis H1 or (M1 ). The prior probability of null hypothesis is assigned as
p(M0 ) and alternative as p(M1 ). Therefore, Baye’s theorem is applied to the
observed data to compute the posterior probability of the Hypothesis. The
appearance of the posterior probability of alternative Hypothesis is computed
as
p(Y |M1 )p(M1 )
p(M1 /Y ) = (15.24)
p(Y |M1 )p(M1 ) + p(Y |M0 )p(M0 )
Concordance Analysis 203

The term P (Y |M1 ) is the marginal likelihood of the data for alternative hy­
potheses. Further, the marginal likelihood is calculated as
� ∞
p(Y |M1 ) = p(Y |θ, M1 )p(θ|H1 )dθ (15.25)
θ

Bayes Factor is used to compute the appearance of P (M1 |Y ) in comparison

to P (M0 |Y ) [151]:

p(M1 |Y ) p(M1 )
= BF10 × (15.26)
p(M0 |Y ) p(M0 )

15.4 Results
Initially, we performed a descriptive data analysis. The classical test car­
ried on CCC between tumor size measurements. Under the null hypothesis is
assumed that the concordance correlation coefficient is zero as ρccc = 0. The
function named as cccUst is available in the “cccrm” package of R. Similarly,
CCC is performed on tumor volume. Demographic profile of the patients pro­
vided in Table 15.1. The Bayesian posterior estimate of CCC detailed in Table
15.2 and Table 15.3. In Bayesian, the same models selected to fit the correla­
tion coefficient to measure the association. It implies that the precision of the
methods is moderate. The correlation coefficients are not very much different
from calculated classical CCC detailed in Table 15.3. It can be concluded by

TABLE 15.1: Posterior estimates obtained are presented

Sex Pre Surgery (Max,Min) Mean Post Surgery Mean
(Max,Min) (SD) (Max,Min) (SD)
Male (20.83,6.08) 12.69(3.64) (15.22,0.95) 8.33(3.42)
Female (20.14,10.23) 13.48(3.95 (14.60,1.69) 9.38(2.81)

TABLE 15.2: Posterior Estimates of concordance correlation coefficients on

tumor size
Parameter Mean SD 2.5% Median 97.5%
ρccc 0.67 0.23 0.13 0.65 0.93
2
σw 2.52 0.31 2.00 2.50 3.21
σa2 84.22 1960.00 0.56 5.23 329.61
σb2 0.01 0.024 0.00 0.00 0.06
σd2 0.01 0.030 0.00 0.00 0.09
θ 3.63 2.80 -4.10 3.62 9.11
204 Bayesian Approaches in Oncology Using R and OpenBUGS

TABLE 15.3: Table 15.3:Posterior estimates of concordance correlation co­

efficients on tumor volume
Parameter Mean SD 2.5% Median 97.5%
ρccc 0.75 0.13 0.09 0.45 0.90
2
σw 11.76 1.43 9.27 11.64 14.88
σa2 259.50 5528.00 1.36 13.74 992.10
σb2 0.01 0.07 0.001 0.00 0.16
σd2 0.01 0.03 0.00 0.00 0.07
θ 10.29 5.09 -3.84 10.63 19.95

FIGURE 15.1: Tumor volume changes in both gender.

having high estimates for accuracy. The estimates obtained through Bayesian
methodology is presented in Table 15.3 and Table 15.4. The same models are
used to obtain the correlation coefficient to obtain the association. The result
shows that the precision is moderate enough. Now the correlation coefficients
are not different from calculated conventional CCC in Table 15.3 with high
accuracy value. In Model 1, the posterior estimate of the concordance correla­
tion coefficient on tumor size is obtained as 0.69. The SD is obtained as 0.23.
The HPD is obtained as (0.13,0.95). Similarly, the posterior estimate on tumor
volume is measured as 0.75(0.13). The HPD in a similar context is measured
as (0.09,0.90). It shows that the confidence interval of the estimates is closed.
Finally, the Bayes factor is measured as 9.53.
Concordance Analysis 205

FIGURE 15.2: Tumor size changes by gender.

FIGURE 15.3: Tumor volume changes by both the method.

206 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 15.4: Tumor size changes by both the method.

15.5 Conclusion
The CCC is an available tool to work on continuously measured variables
for agreement analysis. The presence of covariance, variance, and mean easily
handled by CCC [146, 147]. Perhaps, it may be generated with biased esti­
mates or in the presence of error [152]. The marginal modeling found attractive
to reduce bias value [144]. There is an application of CCC by stratified model­
ing [153]. The wider extension of CCC is known as the generalized concordance
correlation coefficient. The variance component mechanism is adopted for gen­
eralized concordance correlation by [139]. There is a robust outcome observed
through the application of CCC [153]. It is also found useful in time-to-event
data analysis. In oncology, the time-to-event measurements are nonignorable.
The CCC found attractive for censoring data [154, 145]. The extension also
performed for longitudinal measurements [151]. There is an attempt by the
linear mixed effect model by using the generalized CCC (GCCC). The com­
putational algorithm is developed to boost up the Bayesian methodology in
CCC. Initially, the data fitted by GLMM and later on the Bayesian CCC is
considered in this direction. It provides scope to extend it by different types
of hypothesis testing. The generation of the p-value can be overlooked in this
Concordance Analysis 207

direction. The methodology on Bayesian ICC is well established [155]. As a

choice, the beta bionomical model defined. It helped to generate credible in­
tervals [156]. In this chapter, the model generated to work with MCMC. Now
the technique is simple and suitable to handle continuous data. The intention
is to make the Bayesian counterpart of CCC. The work is about exploring the
Bayesian for CCC. Tumor volume and tumor size measurements are by CT
and MRI scanner. Now the measurement obtained by MRI and CT is highly
concordant. Even sometimes they replicate to each other-similarly, several di­
agnostic issues faced in oncology research. Now the Bayesian is functional for
strong evidence on test statistics about the variables. Now the Bayes factor is
used for computing the CCC. This approach promotes the different dimension
of the hypothesis testing.
Chapter 16

High-Dimensional Data Analysis

Abstract
Currently, the proliferation of knowledge of cancer promotes cancer man­
agement by gene therapy. High-dimensional data measure genomic informa­
tion of cancer patients. Motivated by these essential applications in cancer
research, there has been a dramatic growth in the development of statisti­
cal methodology in the analysis of high-dimensional data, mainly related to
regression model selection, estimation and prediction. The high-dimensional
data is useful to explore the cancer progression, especially for disease man­
agement. However, the analyst faces difficulties to deal with high-dimensional
data where the feature dimension p grows exponentially. This chapter is ded­
icated to shows the Bayesian approach in high-dimensional data. Variable
selection in high-dimensional data is one of the challenges. Bayesian variable
selection strategy is presented in this work. Different R packages for variable
selection steps are illustrated. This chapter will help to consider Bayesian in
high-dimensional variable selection.

16.1 Introduction
The base sequence procedure is now well developed. A large amount of ge­
netic data is available publically. This large amount of data challenged us for
the development of analytical tools for analyzing such accumulated data. It
is essential to analyze such extensive genetic data by advanced computational
methodology coupled with statistical techniques for processing genetic data.
Similarly, microarray also provides gene expression information. Commonly,
ten of thousands of variables obtained by a single experiment. Dataset with
this large number of variables are known as high-dimensional data. Earlier, this
used to measure gene expression in serum or tissue. Currently, it used for DNA
methylation expression. Tremendous progress in a microarray experiment ob­
served. Similar, growth in the statistical analysis method followed. Primarily,
the gene effect classification is the main challenge in high-dimensional data

209
210 Bayesian Approaches in Oncology Using R and OpenBUGS

analysis. Filter out a few variables from ten of thousands of variables is the task
of the gene classification. The conventional approach for statistical methodol­
ogy is known as an unsupervised approach. But currently, the direction shifted
from unsupervised to supervised approach. The supervised approach help to
define the characteristics (Y ) to gene expression data (X).
The objective of high-dimensional data analysis is to make clusters be­
tween the observations. The preparation of clustering is useful to merge the
representation into a similar group of individuals. It is required to take a sim­
ilar type of treatment decision for the same cluster. There are different types
of clustering methods, and we will explore through R.However, the statisti­
cal model is useful to predict an individual for specific clustering. Mainly, it
is known as a predictive model for survival data. The method is useful, like
penalized partial log-likelihood (PLL) for the estimation of the regression co­
efficients. The widely adopted method is lasso [157, 158]. The lasso defines the
L1 penalty function. The amount of variations of the penalization is presented
as SCAD [159, 160]. Other methods like the elastic net [161] and Dantzig
selector[162]. The dimension reduction is also useful and can define by p ≥ n.
Methods like tree-based methods, as random survival forests [163] now well
explored. The conventional approach stands with a Cox proportional hazards
model. It works with time-to-event data. The procedure is to use the predic­
tive PLL and after that, the cross-validation. It helps to variable selections
from a ten of thousand of variable [164]. As an alternative to the PLL, the
prediction error curve works well [165]. It helps by identifying the probabilities
of survival from a risk prediction model. The prediction of the error curve is
obtained by the expected squared difference of the risk prediction with actual
event status by Brier score at a given time point. In a Cox regression model,
there is a possibility to obtain unstable estimated regression coefficients. But
stability can be achieved by maximizing the penalized partial log-likelihood.
A penalty function of the regression coefficients subtracted from the partial
log-likelihood. It helps to choose the optimal weight of the penalty function
through maximizing the predictive value of the model [165].

16.2 Heatmap with R

#R Code to Prepare Heatmap

df<-read.csv(“import the data”,header=TRUE)

geneExpmatrix <­ as.matrix(df[2:25])
head(geneExpmatrix)
heatmap.2(geneExpmatrix)
High-Dimensional Data Analysis 211

FIGURE 16.1: Heatmap in high-dimensional data.

#R Code for Stepwise Approach Tree Cutting

hr<-hclust(as.dist(1-cor(t(geneExpmatrix),method=“pearson”)),
method=“complete”)
hc<-hclust(as.dist(1-cor(geneExpmatrix,method=“spearman”)),
method=“complete”)
mycl<-cutree(hr, h=max(hr$height)/1.5)
mycolhc<-rainbow(length(unique(mycl)),start=0.1,end=0.9)
mycolhc<-mycolhc[as.vector(mycl)]
mycol<-colorpanel(40,“darkblue”,“yellow”,“white”)
heatmap.2(geneExpmatrix,Rowv=as.dendrogram(hr),
Colv=as.dendrogram(hc),col=mycol,scale=“row”,density.info=
“none”,trace=“none”RowSideColors=mycolhc)
212 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 16.2: Stepwise approach tree cutting with R.

16.3 Principal Component Analysis with R

In this procedure, the first principal component axis and thereafter second
components are obtained by looking at the eigenvalue between different vari­
ables. In this example, we illustrate the analysis of the principal component
analysis (PCA) as a data reduction technique to simplify reducing the multi­
dimensional data sets of 2 or 3 dimensions by plotting the purposes through
variance analysis. Initially, the first principal component axis and thereafter
second components are obtained by exploring the eigenvalue between different
variables. The same
#R Code for Principal Component Analysis

library(“gplots”)
df<-read.csv(“import the data from hard drive.csv”, header=TRUE)
geneExpmatrix <­ as.matrix(df[2:25])
pca <­ prcomp(geneExpmatrix, scale=T)
summary(pca)
plot(pca$x, pch=20, col=“blue”, type=“n”)
text(pca$x, rownames(pca$x), cex=0.8)
High-Dimensional Data Analysis 213

FIGURE 16.3: Principle component analysis with R.

16.4 Penalized Partial Log-Likelihood (PLL)

The penalized partial log-likelihood (PLL) is an extension of the Cox pro­
portional hazards model
�
λ(t|Zj ) = λ0 (t) exp(Zi β) (16.1)

for the hazard λ(t|Zi ), i.e., the instantaneous risk of having an event at time
t, given the covariate information in Zi , for individual i. The λ0 (t) is used for
�
baseline hazard function. Suppose the parameter vector beta = (β1 , ..., βp ) is
defined by maximizing the PLL with
n n
� � � �
J(β) = δi (zi β − log( J(tj ≥ ti )exp(zj β)) (16.2)
i=1 j=1

It is used as indicator function J. Now, J(A) = 1, if A is true, and J(A) = 0

otherwise. Further, the direct maximization of the PLL is not permissible if the
number of covariates p is larger than the number of observations n. There are
several approaches that are developed by random survival forests [166]. The
PLL is useful to consider the performance of new data after model complexity
is settled.
214 Bayesian Approaches in Oncology Using R and OpenBUGS

16.5 Estimating the Predictive PLL

The prediction performance of PLL is obtained by leave-one-out cross-
validation procedure [166]. The 10-fold cross-validation is prescribed [167].
Now the B cross-validation splots the PLL into a new dataset by predictive
PLL and it is estimated as
B
� � � � �
Jpred (β̂) = δi (zi β̂b − log( J(tj ≥ ti )exp(zj β̂b− )) (16.3)
b=1 i:xi ∈xb j:xj ∈xb

Now the estimated vector of coefficients are fitted by β̂b− after leaving the bth
part of the data out. Now, xb is the bth part of the data with b = 1, ..., B.

16.6 Integrated Prediction Error Curve (IPEC)

The risk prediction model is defined from the fitted Cox proportional haz­
ard model up to time t, given the covariate information zi is
�
r̂(t|zi ) = exp(−Δ̂0 (t)exp(zi β̂)) (16.4)

The Breslow estimator Δ̂0 is used to define the baseline hazard

� t
Δ0 (t) = λ0 (s)ds (16.5)
0

The time-dependent measure of the prediction error can be defined by Brier

score as

BS(t, r̂) = E(Y (t) − τ̂ (t|zi ))2 (16.6)

The true event status is defined as Y (t) = J(T > t) at time t. The predicted
event status is formulated as τ̂ (t|zi ) at that time.

16.7 Ridge Estimators in Cox Regression

Suppose the dimension is defined as p ≥≥ n. It is possible to handle with
the Cox model. The sample size is n. Further, the follow-up time is defined
as t1 , t2 , ...tn . The death event indicator is defined as d1 , .....dn . If di = 1 then
event occured otherwise not. Now for the individual’s the dimension is p as an
explanatory variable. It can be presented as p-vectors X1 ; :::; Xn , which are
High-Dimensional Data Analysis 215

the rows of the design matrix X. The proportional hazards (PH) model can
be defined as
h(t|Xi ) = h0 (t)exp(XiT β) (16.7)
The vector of the regression coefficients R can be defined by maximizing the
partial log-likelihood as
n
� �
pl(β) = di (XiT β) − In( exp(XjT β)) (16.8)
i=1 tj ≥ti

Suppose the baseline hazard function is defined as h0 (t). Suppose the event
point is defined as ti with di = 1. Further, the Breslow estimator can be
formulated as �
ĥ0 (ti ) = 1/ exp(XjT β̂) (16.9)
tj ≥tj

Further, the total likelihood as combination of Breslow estimator and the

partial likelihood is defined as
n
�
l(β, h0 ) = [−exp(XiT β)H0 (ti ) + di ((h0 (ti )) + XiT β)] (16.10)
i=1

with
�
H0 (t) = h0 (S) (16.11)
s≤t

This model with p >> n helps to fit large coefficients as |β̂k | → ∞. It is

perfect to fit as pl(β̂) = 0), to degenerate baseline hazard and filter out no
useful prediction model. This problem can be handled by principle component
analysis by using the dimension reduction method.
lpen (β, h0 ) = l(β, h0 ) − 0.5λβ T β(λ ≥ 0) (16.12)
This is equivalent to putting the same penalty on the partial log-likelihood
plpen (β) = pl(β) − 0.5λβ T β (16.13)
Ridge regression works as a trad-off between bias and variance. It shows
that there is an advantage in terms of the mean squared error (MSE) because
(δMSE/δλ) < 0 at λ = 0.

16.8 High-Dimensional Data Analysis Using R

In this section, we will illustrate about Bayesian Variable Selection in High-
Dimensional Settings by Nonlocal Priors. The BVSNLP package available in
216 Bayesian Approaches in Oncology Using R and OpenBUGS

R will be used to make the illustration. The Bayesian approach is available

in this package for variable selection from high-dimensional data. It helps
through the exploits mixture of point masses at zero and nonlocal priors to
improve the performance of variable selection and coefficient estimation. The
and product-moment (pMOM)and product inverse moment (piMOM) nonlo­
cal priors are available for analysis. However, the binary response and survival
duration is used in this package. It is possible to perform Bayesian variable
selection method by Highest Posterior Probability Model (HPPM), Median
Probability Model (MPM) and posterior inclusion probability for each of the
covariates in the model. The data preparation steps are provided in the “Pa­
rameters Initialization” section. Different covariates are survival duration, and
death status is given in this section. The cox − bvs function is useful to de­
liver the model. Finally, the number of model performed can be observed by
length(coxout$hashkey ) function. Delivraible selected model is obtained by
which (coxout$max-model > 0). The idea about unnormalized probability of
the selected model is obtained by (coxout$max-model > 0).
#Parameters Initialization

library(“BVSNLP”)
n <­ 100
p <­ 40
set.seed(123)
Sigma <­ diag(p)
full <­ matrix(c(rep(0.5, p∗ p)), ncol=p)
Sigma <­ full + 0.5∗ Sigma
cholS <­ chol(Sigma)
Beta <­ c(-1.8, 1.2, -1.7, 1.4, -1.4, 1.3)
X = matrix(rnorm(n×p), ncol=p)
X = X X <­ scale(X)
beta <­ numeric(p)
beta[c(1:length(Beta))] <­ Beta
XB <­ X
surtimes < −rexp(n,exp(XB))
censtimes <­ rexp(n,0.2)
times <­ pmin(surtimes,censtimes)
status <­ as.numeric(surtimes <= censtimes)
exmat <­ cbind(times,status,X)
L <­ 10; J <­ 10
d <­ 2 ∗ ceiling(log(p))
temps <­ seq(3, 1, length.out = L)
tau <­ 0.5; r <­ 1; a <­ 6; b <­ p-a
nlptype <­ 0
cur cols <­ c(1,2,3)
nf <­ 0
High-Dimensional Data Analysis 217

# Runthe Function

coxout <­ cox_bvs(exmat,cur_cols,nf,tau,r,nlptype,

a,b,d,L,J,temps)
summary(coxout)

#Output with Function

Length Class Mode

max_model 40 -none­ numeric
hash_key 156 -none­ numeric
max_prob 1 -none­ numeric
all_probs 156 -none­ numeric
vis_covs_list 156 -none­ list

#The Number of Visited Model for This Specific Run

length(coxouthash_key)

#Specify Model Length

length(coxout$hash_key)
[1] 156

#The Selected Model

which(coxoutmax_model}>0)

#Number of Model Selected

which(coxout$max_model>0)
[1] 1 2 3 4 5 6

#The Unnormalized Probability of the Selected Model

coxoutmax_prob
[1] -247.6858
218 Bayesian Approaches in Oncology Using R and OpenBUGS

#Bayesian Variable Selection in High-Dimensional Settings Using

library(“BVSNLP”)
n <­ 200
p <­ 40
set.seed(123)
Sigma <­ diag(p)
full <­ matrix(c(rep(0.5, p×p)), ncol=p)
Sigma <­ full + 0.5×Sigma
cholS <­ chol(Sigma)
Beta <­ c(-1.7,1.8,2.5)
X <­ matrix(rnorm(n×p), ncol=p)
X <­ X×cholS
colnames(X) <­ c(paste(“gene ”,c(1:p),sep=“ ”))
beta <­ numeric(p)
beta[c(1:length(Beta))]<-Beta
Xout <- PreProcess(X)
X <- Xout$X
XB <­ X×beta
probs <­ as.vector(exp(XB)/(1+exp(XB)))
y <­ rbinom(n,1,probs)
X <­ as.data.frame(X)
train idx <­ sample(1:n,0.×n)
test idx <­ setdiff(1:n,train idx)
X train <­ X[train idx,]
y train <­ y[train idx]
bout <­ bvs(X train, y train, prep=FALSE,
family=“logistic”,
mod prior =“beta”,niter=50)
BMAout<-predBMA(bout,X,y,prep=FALSE,logT=FALSE,
train idx=train idx,test idx=test idx,
family=“logistic”)

#AUC for the Prediction

BMAout$auc
0.6953545
High-Dimensional Data Analysis 219

#BMA Prediction

$auc
[1] 0.6953545

$roc_curve
TPR FPR
[1,] 1.00000000 1.00000000
[2,] 0.89248140 0.98816334
[3,] 0.83054193 0.97708525
[4,] 0.83054193 0.94737135
[5,] 0.77365963 0.94283795
[6,] 0.73734265 0.94283795
[7,] 0.70102566 0.93996391
[8,] 0.66470867 0.93708988
[9,] 0.62839169 0.93708988
[10,] 0.60986843 0.88243937
[11,] 0.54350027 0.88019476
[12,] 0.54075684 0.85048086
[13,] 0.54075684 0.82076696
[14,] 0.50443986 0.82076696
[15,] 0.49018789 0.76967992
[16,] 0.47424697 0.72165646
[17,] 0.47424697 0.69194256
[18,] 0.46870614 0.66222866
[19,] 0.43238916 0.66222866
[20,] 0.39607217 0.66222866
[21,] 0.35975518 0.66222866
[22,] 0.35426833 0.62798136
[23,] 0.31795135 0.62798136
[24,] 0.30924020 0.56879746
[25,] 0.27292322 0.56167016
[26,] 0.27292322 0.53195626
[27,] 0.26252634 0.48879049
[28,] 0.25072939 0.44575389
[29,] 0.20422736 0.43512925
[30,] 0.20422736 0.39891747
[31,] 0.20422736 0.36920357
[32,] 0.20422736 0.32873852
[33,] 0.15120017 0.31770459
[34,] 0.13694820 0.27798936
[35,] 0.12290807 0.22769629
[36,] 0.08307837 0.21862949
[37,] 0.03847714 0.21070557
[38,] 0.02744947 0.15743430
[39,] 0.02169358 0.12772040
[40,] 0.01818087 0.09163778
[41,] 0.01298243 0.04376265
[42,] 0.00000000 0.00000000
220 Bayesian Approaches in Oncology Using R and OpenBUGS

#Plotting ROC Curve

roc <- BMAout$roc_curve

plot(roc)

FIGURE 16.4: True positive rate and false positive rate relation.

#Install Packages with R for High-Dimensional Survival Analysis

install.packages(“glmnet”)
install.packages(“Matrix”)
install.packages(“tidyverse”)
install.packages(“broom”)
install.packages(“survival”)

#Load Packages

library(“glmnet”)
library(“Matrix”)
library(“tidyverse”)
library(“broom”)
library(“survival”)
High-Dimensional Data Analysis 221

#Generate toy Survival Data

time<-rnorm(100,54,10) #Survival Duration data

status<-rbinom(100,1,.5)} # Survival status data
x.1<-rnorm(100,4,.2) } # x.1 to x.10 are ten variables
x.2<-rnorm(100,5,.18)
x.3<-rnorm(100,6,.29)
x.4<-rnorm(100,7,.26)
x.5<-rnorm(100,8,.22)
x.6<-rnorm(100,9,.29)
x.7<-rnorm(100,10,.26)
x.8<-rnorm(100,11,.12)
x.9<-rnorm(100,12,.22)
x.10<-rnorm(100,13,.3))

Reformed the x.1 to x.10 Variables into Matrix

fdata<-data.frame(x.1,x.2,x.3,x.4,x.5,x.6,x.7,x.8,x.9,x.10)
x <­ fdata %>% select(x.1,x.2,x.3,x.4,x.5,x.6,x.7,x.8,x.9
,x.10)
%>% data.matrix()

#Output1

cv.fit2 =cv.glmnet(x,Surv(time,status),alpha = 1, family =

“cox”,maxit =1000) plot(cv.fit2)

#Output2

log(cv.fit2$lambda.min)
-2.229185
cv.fit2$lambda.min
0.1076161

#Output3

est.coef = coef(cv.fit2, s = cv.fit2$lambda.min)

active.k = which(est.coef != 0)
active.k
integer(0)

#Output4

active.k.vals = est.coef[active.k]
active.k.vals
numeric(0)
222 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 16.5: Partial likelihood estimates.

16.9 Different Packages with R

16.9.1 BAMA
#R Code for Shrinkage Estimator on High-Dimensional Datadd

install.packages(“bama”)
library(bama) # call the library bma#
Y <­ bama.data$y # generate continous variable of size 1000#
A <­ bama.data$ a # generate continous variable of size 1000#
M <­ as.matrix(bama.data[, paste0(“m”, 1:100)],
nrow(bama.data))
#Create the Matrix of having M#
C1 <­ matrix(1, 1000, 1) #Create one covariate of C1#
C2 <­ matrix(1, 1000, 1)#Create one covariate of C2#
beta.m <­ rep(0, 100)
alpha.a <­ rep(0, 100)
set.seed(12345)
#bama function is used filter influencing variable in the regression#
out<-bama(Y,A,M,C1,C2,beta.m,alpha.a,burnin=1000,
ndraws = 100)
summary <­ summary(out)
head(summary)
High-Dimensional Data Analysis 223

16.9.2 countgmifs
This package is useful to fit negative binomial nd Poisson regression by
stagewise manner. It is illustrated in the following text:
#R Code on Countgmifsp

install.packages(“countgmifs”)
library(“countgmifs”)
set.seed(10)
n <­ 50 # Sample size of the data
p <­ 500 #Covariates required for the data
intercept<­ .5
#generate parameter for 500 covariates#
beta<­ c(log(1.5), log(1.5),
-log(1.5), -log(1.5),
-log(1.5), rep(0,495))
alpha<­ 0.5 # Intercept
x<­ matrix(rnorm(n×p,0,1), nrow=n, ncol=p,
byrow=TRUE) #Covariate values
colnames(x)<­ paste(“Var”,1:p, sep=“ ”)
mu<­ exp(intercept + crossprod(t(x),beta))
y<­ rnbinom(n=n, size=1/alpha ,mu=mu) # Discrete response
data<­ data.frame(y,x)
nb<-countgmifs(y 1 , data=data, offset=NULL,
x=x, epsilon=0.01, tol=0.001,
scale=TRUE, verbose=FALSE)
coef.AIC<-coef(nb, model.select=“AIC”)
coef.AIC[coef.AIC!=0]
predict(nb, model.select=“AIC”)
plot(predict(nb, model.select=“AIC”), y)
plot(nb)

16.9.3 fastcox
This R package is used for Lasso and Elastic-Net Penalized Cox’s Regres­
sion in High Dimensions Models by the Cocktail Algorithm. In this package,
the majorization-minimization principle is applied of the elastic net penalized
Cox’s proportional hazard model. It is illustrated on the following page.
224 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 16.6: Model selection by AIC.

#R Code on Fastcox

install.packages(“fastcox”)
library(“fastcox”)
data(FHT)
m1<-cocktail(x=FHT$x,y=FHT$y,d=FHT$status,alpha=0.5)
predict(m1,type=“nonzero”)
plot(m1)
plot(m1,xvar=“lambda”,label=TRUE)
plot(m1,color=TRUE)

16.9.4 HighDimOut
The HighDimOut package is useful for Outlier Detection Algorithms in
High-Dimensional Data. It is helpful for outlier detection by unification
scheme. One example of running this package is detailed.
High­Dimensional Data Analysis 225

FIGURE 16.7: plot predicted values.

#R Code to Run HighDimOut

library(“HighDimOut”)
library(“ggplot2”)
res.ABOD <­ Func.ABOD(data=TestData[,1:2], basic=FALSE,
perc=0.2)
data.temp <­ TestData[,1:2]
data.temp$Ind <­ NA
data.temp[order(res.ABOD,
decreasing = FALSE)[1:10],“Ind”]
<­ “Outlier”
data.temp[is.na(data.temp$Ind),“Ind”] <­ “Inlier”
data.temp$Ind <­ factor(data.temp$Ind)
ggplot(data = data.temp) + geom point(aes(x = x, y = y,
color=Ind, shape=Ind))
226 Bayesian Approaches in Oncology Using R and OpenBUGS

FIGURE 16.8: prediction of L1 norm and log lambda.

16.10 Discussion
Data science is an emerging field in oncology. Due to the availability of huge
gene data, it can not be avoided. Always analyst faces difficulties to deal with
high-dimensional data where the feature dimension p grows exponentially. The
sample size is fixed at n, but log(p) = 0(nα )for some α ∈ (0, 1/2). Currently,
modern technology, like microarray analysis and next-generation sequencing
data is available to generate genomics data. Large amounts of data containing
more than thousands of variables are available. It is common to collect gene
expressions from p > 20,000 genomics studies. These genomics data are having
a large sample size and a limited size p parameter. The large size gene data
analysis is a challenge. Consideration of not relevant features in the statistical
literature indeed may provide undesirable computing challenges. The chal­
lenges are handled by variable screening and selection techniques. Among all
methodological development, the penalized approach is favored by researchers
with K-fold cross-validation.
Bibliography

[1] Shashi Kant Mishra and Bhagwat Ram. Introduction to Unconstrained

Optimization with R. Springer Nature, 2019.
[2] John M Lachin and Mary A Foulkes. Evaluation of sample size and
power for analyses of survival with allowance for nonuniform patient
entry, losses to follow-up, noncompliance, and stratification. Biometrics,
pages 507–519, 1986.
[3] Evangelos Briasoulis, Periklis Pappas, Christian Puozzo, Christos To-
lis, George Fountzilas, Urania Dafni, Marios Marselos, and Nicholas
Pavlidis. Dose-ranging study of metronomic oral vinorelbine in patients
with advanced refractory cancer. Clinical cancer research, 15(20):6454–
6461, 2009.
[4] Vincent T DeVita and Edward Chu. A history of cancer chemotherapy.
Cancer research, 68(21):8643–8653, 2008.
[5] Howard E Skipper, Frank M Schabel Jr, L Bruce Mellett, John A Mont­
gomery, Lee J Wilkoff, Harris H Lloyd, and R Wallace Brockman. Im­
plications of biochemical, cytokinetic, pharmacologic, and toxicologic
relationships in the design of optimal therapeutic schedules. Cancer
chemother Rep, 54(6):431–450, 1970.
[6] Pan Pantziarka, Lisa Hutchinson, Nicolas André, Sébastien Benzekry,
Francesco Bertolini, Atanu Bhattacharjee, Shubhada Chiplunkar, Dan G
Duda, Vikram Gota, Sudeep Gupta, et al. Next generation metronomic
chemotherapy-report from the fifth biennial international metronomic
and anti-angiogenic therapy meeting, 6–8 may 2016, mumbai. ecancer­
medicalscience, 10, 2016.
[7] Fred Ederer. The relative survival rate: a statistical methodology. NCI
monograph, 6:101–121, 1961.
[8] Francesca De Felice, Daniela Musio, and Vincenzo Tombolini. Head and
neck cancer: metronomic chemotherapy. BMC cancer, 15(1):677, 2015.
[9] Peter Schmid, Walter Schippinger, Thorsten Nitsch, Gerdt Huebner,
Volker Heilmann, Wolfgang Schultze, Hubert Hausmaninger, Man­
fred Wischnewsky, and Kurt Possinger. Up-front tandem high-dose

227
228 Bibliography

chemotherapy compared with standard chemotherapy with doxorubicin

and paclitaxel in metastatic breast cancer: results of a randomized trial.
Journal of clinical oncology, 23(3):432–440, 2005.
[10] Douglas Hanahan, Gabriele Bergers, and Emily Bergsland. Less is more,
regularly: metronomic dosing of cytotoxic drugs can target tumor angio­
genesis in mice. The Journal of clinical investigation, 105(8):1045–1047,
2000.
[11] Leonard B Saltz. Progress in cancer care: the hope, the hype, and
the gap between reality and perception. Journal of clinical oncology,
26(31):5020–5021, 2008.
[12] YA Luqmani. Mechanisms of drug resistance in cancer chemotherapy.
Medical principles and practice, 14(Suppl. 1):35–48, 2005.
[13] RS Kerbel, G Klement, KI Pritchard, and B Kamen. Continuous low-
dose anti-angiogenic/metronomic chemotherapy: from the research lab­
oratory into the oncologyclinic, 2002.
[14] Ross L Prentice, John D Kalbfleisch, Arthur V Peterson Jr, Nancy
Flournoy, Vernon T Farewell, and Norman E Breslow. The analysis
of failure times in the presence of competing risks. Biometrics, pages
541–554, 1978.
[15] John D Kalbfleisch and Ross L Prentice. The statistical analysis of
failure time data, volume 360. John Wiley & Sons, 2011.
[16] Rebecca Gelman, Richard Gelber, I Craig Henderson, C Norman Cole­
man, and Jay R Harris. Improved methodology for analyzing local and
distant recurrence. Journal of clinical oncology, 8(3):548–555, 1990.
[17] Richard J Caplan, Thomas F Pajak, and James D Cox. Analysis of the
probability and risk of cause-specific failure. International journal of
radiation oncology* biology* physics, 29(5):1183–1186, 1994.
[18] Ted A Gooley, Wendy Leisenring, John Crowley, and Barry E Storer.
Estimation of failure probabilities in the presence of competing risks:
new representations of old estimators. Statistics in medicine, 18(6):695–
706, 1999.
[19] Martin G Larson and Gregg E Dinse. A mixture model for the regression
analysis of competing risks data. Journal of the royal statistical society:
series C (Applied Statistics), 34(3):201–211, 1985.
[20] Jason P Fine and Robert J Gray. A proportional hazards model for the
subdistribution of a competing risk. Journal of the american statistical
association, 94(446):496–509, 1999.
Bibliography 229

[21] Per Kragh Andersen, John P Klein, and Susanne Rosthøj. Generalised
linear models for correlated pseudo-observations, with applications to
multi-state models. Biometrika, 90(1):15–27, 2003.
[22] Bernhard Haller, Georg Schmidt, and Kurt Ulm. Applying competing
risks regression models: an overview. Lifetime data analysis, 19(1):33–
58, 2013.
[23] Leo Breiman. Random forests. Machine learning, 45(1):5–32, 2001.
[24] Hemant Ishwaran, Thomas A Gerds, Udaya B Kogalur, Richard D
Moore, Stephen J Gange, and Bryan M Lau. Random survival forests
for competing risks. Biostatistics, 15(4):757–773, 2014.
[25] G Schwarz. Estimating the dimension of a model. annals of statistics, 6
(2), 461–464, 1978.
[26] H Akaike. hon entropy maximization principle. iin pr krishnarah (ed.),
applications of statistics. Amsterdam: NorthjHolland, 1977.
[27] John P Klein and Melvin L Moeschberger. Survival analysis: techniques
for censored and truncated data. Springer Science & Business Media,
2006.
[28] Luc Duchateau and Paul Janssen. The frailty model. Springer science
& business media, 2007.
[29] Philip Hougaard. Shared frailty models. In Analysis of multivariate
survival data, pages 215–262. Springer, 2000.
[30] José Cortiñas Abrahantes, Catherine Legrand, Tomasz Burzykowski,
Paul Janssen, Vincent Ducrocq, and Luc Duchateau. Comparison of
different estimation procedures for proportional hazards model with ran­
dom effects. Computational statistics & data analysis, 51(8):3913–3930,
2007.
[31] Virginie Rondeau, Yassin Mazroui, and Juan R Gonzalez. frailtypack:
an R package for the analysis of correlated survival data with frailty
models using penalized likelihood estimation or parametrical estimation.
Journal of statistical software, 47(4):1–28, 2012.
[32] Terry Therneau. coxme: mixed effects cox models. R package version
2.2-3. Vienna, Austria: R Foundation for Statistical Computing, 2012.
[33] Il Do Ha, Maengseok Noh, and Youngjo Lee. frailtyhl: A package for
fitting frailty models with h-likelihood. The R Journal, 4(2):28–36, 2012.
[34] Marco Munda, Federico Rotolo, Catherine Legrand, et al. Parfm: para­
metric frailty models in uppercaseR. Journal of statistical software,
51(11):1–20, 2012.
230 Bibliography

[35] Andreas Wienke. Frailty models in survival analysis. CRC press, 2010.
[36] Ronald T Cotton, Hop S Tran Cao, Abbas A Rana, Yvonne H Sada,
David A Axelrod, John A Goss, Mark A Wilson, Steven A Curley, and
Nader N Massarweh. Impact of the treating hospital on care outcomes
for hepatocellular carcinoma. Hepatology, 68(5):1879–1889, 2018.
[37] Evan M Graboyes, Mark A Ellis, Hong Li, John M Kaczmar, Anand K
Sharma, Eric J Lentsch, Terry A Day, and Chanita Hughes Halbert.
Racial and ethnic disparities in travel for head and neck cancer treatment
and the impact of travel distance on survival. Cancer, 124(15):3181–
3191, 2018.
[38] Yuhui Chen, Timothy Hanson, and Jiajia Zhang. Accelerated hazards
model based on parametric families generalized with bernstein polyno­
mials. Biometrics, 70(1):192–201, 2014.
[39] Haiming Zhou and Timothy Hanson. A unified framework for fitting
bayesian semiparametric models to arbitrarily censored survival data,
including spatially referenced data. Journal of the american statistical
association, 113(522):571–581, 2018.
[40] T Hakulinen and L Tenkanen. Regression analysis of relative survival
rates. Journal of the royal statistical society: series C (applied statistics),
36(3):309–317, 1987.
[41] Arun Pokhrel and Timo Hakulinen. Age-standardisation of relative sur­
vival ratios of cancer patients in a comparison between countries, genders
and time periods. European journal of cancer, 45(4):642–647, 2009.
[42] PC Lambert, PW Dickman, CP Nelson, and P Royston. Estimating the
crude probability of death due to cancer and other causes using relative
survival models. Statistics in medicine, 29(7-8):885–895, 2010.
[43] Klemen Pavlič and Maja Pohar Perme. On comparison of net survival
curves. BMC medical research methodology, 17(1):79, 2017.
[44] Klemen Pavlič and Maja Pohar Perme. Using pseudo-observations for
estimation in relative survival. Biostatistics, 20(3):384–399, 2019.
[45] Maja Pohar Perme, Janez Stare, and Jacques Estève. On estimation in
relative survival. Biometrics, 68(1):113–120, 2012.
[46] Paola Rebora, Stefania Galimberti, and Maria Grazia Valsecchi. Us­
ing multiple timescale models for the evaluation of a time-dependent
treatment. Statistics in medicine, 34(28):3648–3660, 2015.
[47] Margaret Sullivan Pepe and Motomi Mori. Kaplan-meier, marginal or
conditional probability curves in summarizing competing risks failure
time data? Statistics in medicine, 12(8):737–751, 1993.
Bibliography 231

[48] Yi-Ching Yao. Maximum likelihood estimation in hazard rate models

with a change-point. Communications in statistics-theory and methods,
15(8):2455–2466, 1986.
[49] JD Buckley. Additive and multiplicative models for relative survival
rates. Biometrics, pages 51–62, 1984.
[50] Janez Stare, Robin Henderson, and Maja Pohar. An individual measure
of relative survival. Journal of the royal statistical society: series C
(applied statistics), 54(1):115–126, 2005.
[51] Douglas Bates, Martin Maechler, Ben Bolker, Steven Walker, Rune
Haubo Bojesen Christensen, Henrik Singmann, Bin Dai, Gabor
Grothendieck, Peter Green, and Maintainer Ben Bolker. Package lme4.
Convergence, 12(1):2, 2015.
[52] David Clayton. Some approaches to the analysis of recurrent event data.
Statistical methods in medical research, 3(3):244–262, 1994.
[53] Hemant Ishwaran and Lancelot F James. Computational methods for
multiplicative intensity models using weighted gamma processes: propor­
tional hazards, marked point processes, and panel count data. Journal
of the american statistical association, 99(465):175–190, 2004.
[54] Alan E Gelfand and Adrian FM Smith. Sampling-based approaches
to calculating marginal densities. Journal of the american statistical
association, 85(410):398–409, 1990.
[55] Bradley P Carlin, Alan E Gelfand, and Adrian FM Smith. Hierarchical
bayesian analysis of changepoint problems. Journal of the royal statis­
tical society: series C (applied statistics), 41(2):389–405, 1992.
[56] Nicholas Lange, Bradley P Carlin, and Alan E Gelfand. Hierarchi­
cal bayes models for the progression of hiv infection using longitudi­
nal cd4 t-cell numbers. Journal of the american statistical association,
87(419):615–626, 1992.
[57] Siddhartha Chib and Bradley P Carlin. On mcmc sampling in hierar­
chical longitudinal models. Statistics and computing, 9(1):17–26, 1999.
[58] Vanita Noronha, Vijay M Patil, Amit Joshi, Atanu Bhattacharjee,
Davinder Paul, Sachin Dhumal, Shashikant Juvekar, Supreeta Arya, and
Kumar Prabhash. A tertiary care experience with paclitaxel and cetux­
imab as palliative chemotherapy in platinum sensitive and nonsensitive
in head and neck cancers. South asian journal of cancer, 6(1):11, 2017.
[59] Shanshan Li. Joint modeling of recurrent event processes and intermit­
tently observed time-varying binary covariate processes. Lifetime data
analysis, 22(1):145–160, 2016.
232 Bibliography

[60] Dimitris Rizopoulos. Jm: An R package for the joint modelling of longi­
tudinal and time-to-event data. Journal of statistical software (Online),
35(9):1–33, 2010.
[61] Dimitris Rizopoulos. The R package jmbayes for fitting joint models
for longitudinal and time-to-event data using mcmc. arXiv preprint
arXiv:1404.7625, 2014.
[62] Pete Philipson, Peter Diggle, Ines Sousa, Ruwanthi Kolamunnage-Dona,
Paula Williamson, and Robin Henderson. joiner: Joint modelling of
repeated measurements and time-to-event data. 2012.
[63] Cécile Proust-Lima and Benoit Liquet. Lcmm: an R package for esti­
mation of latent class mixed models and joint latent class models. In
The R User Conference, useR! 2011 August 16-18 2011 University of
Warwick, Coventry, UK, page 66. Citeseer, 2011.
[64] Menggang Yu, Ngayee J Law, Jeremy MG Taylor, and Howard M San­
dler. Joint longitudinal-survival-cure models and their application to
prostate cancer. Statistica Sinica, pages 835–862, 2004.
[65] Anastasios A Tsiatis, Victor Degruttola, and Michael S Wulfsohn. Mod­
eling the relationship of survival to longitudinal data measured with er­
ror. applications to survival and cd4 counts in patients with aids. Journal
of the american statistical association, 90(429):27–37, 1995.
[66] Michael S Wulfsohn and Anastasios A Tsiatis. A joint model for survival
and longitudinal data measured with error. Biometrics, pages 330–339,
1997.
[67] Geert Verbeke. Linear mixed models for longitudinal data. In Linear
mixed models in practice, pages 63–153. Springer, 1997.
[68] Dimitris Rizopoulos, Geert Verbeke, and Geert Molenberghs. Shared
parameter models under random effects misspecification. Biometrika,
95(1):63–74, 2008.
[69] Jose Pinheiro, Douglas Bates, Saikat DebRoy, Deepayan Sarkar, et al.
Linear and nonlinear mixed effects models. R package version, 3:57,
2007.
[70] D Bates, M Maechler, and B Bolker. The lme4 package, uppercaseR
package, version 2. 2007.
[71] Terry M Therneau and Thomas Lumley. Package survival. Survival
analysis published on CRAN, 2:3, 2014.
[72] Jimin Ding and Jane-Ling Wang. Modeling longitudinal data with non­
parametric multiplicative random effects jointly with survival data. Bio­
metrics, 64(2):546–556, 2008.
Bibliography 233

[73] Elizabeth R Brown, Joseph G Ibrahim, and Victor DeGruttola. A flex­

ible b-spline model for multiple longitudinal biomarkers and survival.
Biometrics, 61(1):64–73, 2005.
[74] Stephanie Green, Jacqueline Benedetti, Angela Smith, and John Crow­
ley. Clinical trials in oncology, volume 28. CRC press, 2012.
[75] John O’Quigley, Margaret Pepe, and Lloyd Fisher. Continual reassess­
ment method: a practical design for phase 1 clinical trials in cancer.
Biometrics, pages 33–48, 1990.
[76] Steven N Goodman, Marianna L Zahurak, and Steven Piantadosi. Some
practical improvements in the continual reassessment method for phase
i studies. Statistics in medicine, 14(11):1149–1161, 1995.
[77] S Piantadosi, JD Fisher, and S Grossman. Practical implementation of a
modified continual reassessment method for dose-finding trials. Cancer
chemotherapy and pharmacology, 41(6):429–436, 1998.
[78] Z Yuan, R Chappell, and H Bailey. The continual reassessment method
for multiple toxicity grades: a bayesian quasi-likelihood approach. Bio­
metrics, 63(1):173–179, 2007.
[79] Ying Kuen Cheung and Rick Chappell. Sequential designs for phase
I clinical trials with late-onset toxicities. Biometrics, 56(4):1177–1182,
2000.
[80] James Babb, André Rogatko, and Shelemyahu Zacks. Cancer phase-I
clinical trials: efficient dose escalation with overdose control. Statistics
in medicine, 17(10):1103–1120, 1998.
[81] André Rogatko, David Schoeneck, William Jonas, Mourad Tighiouart,
Fadlo R Khuri, and Alan Porter. Translation of innovative designs into
phase- uppercaseI trials. Journal of clinical oncology, 25(31):4982–4986,
2007.
[82] Ying Yuan, Beibei Guo, Mark Munsell, Karen Lu, and Amir Jazaeri.
Midas: a practical bayesian design for platform trials with molecularly
targeted agents. Statistics in medicine, 35(22):3892–3906, 2016.
[83] Victor Moreno Garcı́a, David Olmos, Carlos Gomez-Roca, Philippe A
Cassier, Rafael Morales-Barrera, Gianluca Del Conte, Elisa Gallerani,
Andre T Brunetto, Patrick Schöffski, Silvia Marsoni, et al. Dose–
response relationship in phase I clinical trials: a european drug devel­
opment network (eddn) collaboration study. Clinical cancer research,
20(22):5663–5671, 2014.
[84] Stephen A Cannistra. Challenges and pitfalls of combining targeted
agents in phase I studies, 2008.
234 Bibliography

[85] Wendy R Parulekar and Elizabeth A Eisenhauer. Phase I trial design

for solid tumor studies of targeted, non-cytotoxic agents: theory and
practice. Journal of the national cancer institute, 96(13):990–997, 2004.
[86] Edward L Korn, Susan G Arbuck, James M Pluda, Richard Simon,
Richard S Kaplan, and Michaele C Christian. Clinical trial designs
for cytostatic agents: are new approaches needed? Journal of clinical
oncology, 19(1):265–272, 2001.
[87] Feifei Li, Changqi Zhao, and Lili Wang. Molecular-targeted agents com­
bination therapy for cancer: developments and potentials. International
journal of cancer, 134(6):1257–1269, 2014.
[88] J-C Soria, JY Blay, JP Spano, X Pivot, Y Coscas, and D Khayat. Added
value of molecular targeted agents in oncology. Annals of oncology,
22(8):1703–1716, 2011.
[89] Sumithra J Mandrekar, Yue Cui, and Daniel J Sargent. An adaptive
phase uppercaseI design for identifying a biologically optimal dose for
dual agent drug combinations. Statistics in medicine, 26(11):2317–2330,
2007.
[90] Akihiro Hirakawa. An adaptive dose-finding approach for correlated
bivariate binary and continuous outcomes in phase I oncology trials.
Statistics in medicine, 31(6):516–532, 2012.
[91] Chunyan Cai, Ying Yuan, and Yuan Ji. A bayesian dose finding design
for oncology clinical trials of combinational biological agents. Journal of
the royal statistical society: series C (applied statistics), 63(1):159–173,
2014.
[92] David B Dunson and Brian Neelon. Bayesian inference on order-
constrained parameters in generalized linear models. Biometrics,
59(2):286–295, 2003.
[93] Laura H Gunn and David B Dunson. A transformation approach for in­
corporating monotone or unimodal constraints. Biostatistics, 6(3):434–
449, 2005.
[94] Yisheng Li, B Nebiyou Bekele, Yuan Ji, and John D Cook. Dose–
schedule finding in phase uppercaseI/IIclinical trials using a bayesian
isotonic transformation. Statistics in medicine, 27(24):4895–4913, 2008.
[95] Ikuko Funatogawa, Takashi Funatogawa, and Yasuo Ohashi. An autore­
gressive linear mixed effects model for the analysis of longitudinal data
which show profiles approaching asymptotes. Statistics in medicine,
26(9):2113–2130, 2007.
Bibliography 235

[96] Xu Xu, Min Yuan, and Partha Nandy. Analysis of dose–response in flex­
ible dose titration clinical studies. Pharmaceutical statistics, 11(4):280–
286, 2012.
[97] Toshihiro Misumi and Sadanori Konishi. Mixed effects historical
varying-coefficient model for evaluating dose–response in flexible dose
trials. Journal of the royal statistical society: series C (applied statis­
tics), 65(2):331–344, 2016.
[98] JQ Shi, B Wang, EJ Will, and RM West. Mixed-effects gaussian process
functional regression models with application to dose–response curve
prediction. Statistics in medicine, 31(26):3165–3177, 2012.
[99] Damla Sent¨
¸ urk and Hans-Georg M¨ uller. Functional varying coefficient
models for longitudinal data. Journal of the american statistical associ­
ation, 105(491):1256–1264, 2010.
[100] Trevor Hastie and Robert Tibshirani. Varying-coefficient models. Jour­
nal of the royal statistical society: series B (methodological), 55(4):757–
779, 1993.
[101] Wensheng Guo. Functional mixed effects models. Biometrics, 58(1):121–
128, 2002.
[102] Jian Qing Shi and Taeryon Choi. Gaussian process regression analysis
for functional data. CRC Press, 2011.
[103] Hulin Wu and Jin-Ting Zhang. Nonparametric regression methods for
longitudinal data analysis: mixed-effects modeling approaches, volume
515. John Wiley & Sons, 2006.
[104] Harvey Goldstein. Efficient statistical modelling of longitudinal data.
Annals of human biology, 13(2):129–141, 1986.
[105] Nan M Laird and James H Ware. Random-effects models for longitudinal
data. Biometrics, pages 963–974, 1982.
[106] Douglas Bates, Martin Mächler, Ben Bolker, and Steve Walker. Fitting
linear mixed-effects models using lme4. arXiv preprint arXiv:1406.5823,
2014.
[107] Jose Pinheiro, Douglas Bates, Saikat DebRoy, Deepayan Sarkar, and
R Core Team. nlme: Linear and nonlinear mixed effects models. R
package version, 3(2), 2015.
[108] Elena N Naumova, Aviva Must, and Nan M Laird. Tutorial in biostatis­
tics: evaluating the impact of critical periods in longitudinal studies of
growth using piecewise mixed effects models. International journal of
epidemiology, 30(6):1332–1341, 2001.
236 Bibliography

[109] Thomas M Braun, Zheng Yuan, and Peter F Thall. Determin­

ing a maximum-tolerated schedule of a cytotoxic agent. Biometrics,
61(2):335–343, 2005.
[110] Changying A Liu and Thomas M Braun. Parametric non-mixture cure
models for schedule finding of therapeutic agents. Journal of the Royal
Statistical Society: Series C (Applied Statistics), 58(2):225–236, 2009.
[111] Youyi Fong, Håvard Rue, and Jon Wakefield. Bayesian inference for
generalized linear mixed models. Biostatistics, 11(3):397–412, 2010.
[112] Jarrod D Hadfield et al. Mcmc methods for multi-response generalized
linear mixed models: the mcmcglmm r package. Journal of statistical
software, 33(2):1–22, 2010.
[113] Bob Carpenter, Andrew Gelman, Matthew D Hoffman, Daniel Lee, Ben
Goodrich, Michael Betancourt, Marcus Brubaker, Jiqiang Guo, Peter Li,
and Allen Riddell. Stan: A probabilistic programming language. Journal
of statistical software, 76(1), 2017.
[114] Andrew Gelman et al. Prior distributions for variance parameters in hi­
erarchical models (comment on article by browne and draper). Bayesian
analysis, 1(3):515–534, 2006.
[115] David J Lunn, Andrew Thomas, Nicky Best, and David Spiegelhalter.
Winbugs-a bayesian modelling framework: concepts, structure, and ex­
tensibility. Statistics and computing, 10(4):325–337, 2000.
[116] Martyn Plummer. Jags: Just another gibbs sampler, 2004.
[117] Paul C Lambert, Alex J Sutton, Paul R Burton, Keith R Abrams, and
David R Jones. How vague is vague? a simulation study of the impact
of the use of vague prior distributions in mcmc using winbugs. Statistics
in medicine, 24(15):2401–2428, 2005.
[118] Radford M Neal et al. Mcmc using hamiltonian dynamics. Handbook of
Markov chain Monte Carlo, 2(11):2, 2011.
[119] Simon Duane, Anthony D Kennedy, Brian J Pendleton, and Duncan
Roweth. Hybrid monte carlo. Physics letters B, 195(2):216–222, 1987.
[120] Matthew D Hoffman and Andrew Gelman. The no-u-turn sampler:
adaptively setting path lengths in hamiltonian monte carlo. Journal
of machine learning research, 15(1):1593–1623, 2014.
[121] Irina Sousa Moreira, Pedro Alexandrino Fernandes, and Maria
Joao Ramos. Vascular endothelial growth factor (vegf) inhibition-a crit­
ical review. Anti-cancer agents in medicinal chemistry (formerly current
medicinal chemistry-anti-cancer agents), 7(2):223–245, 2007.
Bibliography 237

[122] Paul CD Johnson, Sarah JE Barry, Heather M Ferguson, and Pie Müller.
Power analysis for generalized linear mixed models in ecology and evo­
lution. Methods in ecology and evolution, 6(2):133–142, 2015.
[123] Julien GA Martin, Daniel H Nussey, Alastair J Wilson, and Denis Réale.
Measuring individual differences in reaction norms in field and experi­
mental studies: a power analysis of random regression models. Methods
in ecology and evolution, 2(4):362–374, 2011.
[124] Xianggui Qu. Linear mixed-effects models using R: A step-by-step ap­
proach, 2014.
[125] Nicholas G Reich, Jessica A Myers, Daniel Obeng, Aaron M Milstone,
and Trish M Perl. Empirical power and sample size calculations for
cluster-randomized and cluster-randomized crossover studies. PLoS one,
7(4), 2012.
[126] Christophe Le Tourneau, J Jack Lee, and Lillian L Siu. Dose escalation
methods in phase I cancer clinical trials. JNCI: Journal of the national
cancer institute, 101(10):708–720, 2009.
[127] Wei Zhang, Daniel J Sargent, and Sumithra Mandrekar. An adaptive
dose-finding design incorporating both toxicity and efficacy. Statistics
in medicine, 25(14):2365–2383, 2006.
[128] Mei-Yin Polley and Ying Kuen Cheung. Two-stage designs for dose-
finding trials with a biologic endpoint using stepwise tests. Biometrics,
64(1):232–241, 2008.
[129] Yong Zang, J Jack Lee, and Ying Yuan. Adaptive designs for identifying
optimal biological dose for molecularly targeted agents. Clinical trials,
11(3):319–327, 2014.
[130] Fang-Zhen Shen, Jing Wang, Jun Liang, Kun Mu, Ji-Yuan Hou, and
Yan-Tao Wang. Low-dose metronomic chemotherapy with cisplatin: can
it suppress angiogenesis in h22 hepatocarcinoma cells? International
journal of experimental pathology, 91(1):10–16, 2010.
[131] Atanu Bhattacharjee and Vijay M Patil. Determining an optimum bi­
ological dose of a metronomic chemotherapy. Journal of data science,
15(1):77–94, 2017.
[132] Atanu Bhattacharjee and Vijay M Patil. Time-dependent area under
the roc curve for optimum biological dose detection. Turkiye klinikleri
journal of biostatistics, 8(2), 2016.
[133] Yuval Shaked, Alessia Ciarrocchi, Marcela Franco, Christina R Lee,
Shan Man, Alison M Cheung, Daniel J Hicklin, David Chaplin, F Stuart
238 Bibliography

Foster, Robert Benezra, et al. Therapy-induced acute recruitment of cir­

culating endothelial progenitor cells to tumors. Science, 313(5794):1785–
1787, 2006.
[134] Q-B Lu. Molecular reaction mechanisms of combination treatments of
low-dose cisplatin with radiotherapy and photodynamic therapy. Jour­
nal of medicinal chemistry, 50(11):2601–2604, 2007.
[135] D Malka, V Boige, N Jacques, N Vimond, A Adenis, E Boucher,
JY Pierga, T Conroy, B Chauffert, E François, et al. Clinical value
of circulating endothelial cell levels in metastatic colorectal cancer pa­
tients treated with first-line chemotherapy and bevacizumab. Annals of
oncology, 23(4):919–927, 2012.
[136] Adrian M Jubb, Adam J Oates, Scott Holden, and Hartmut Koeppen.
Predicting benefit from anti-angiogenic agents in malignancy. Nature
reviews cancer, 6(8):626–635, 2006.
[137] Rupal S Bhatt, Pankaj Seth, and Vikas P Sukhatme. Biomarkers for
monitoring antiangiogenic therapy. Clinical cancer research, 13(2):777s–
780s, 2007.
[138] Francesco Bertolini, Yuval Shaked, Patrizia Mancuso, and Robert S Ker­
bel. The multifaceted circulating endothelial cell in cancer: towards
marker and target identification. Nature reviews cancer, 6(11):835–845,
2006.
[139] Josep L Carrasco and Lluis Jover. Estimating the generalized concor­
dance correlation coefficient through variance components. Biometrics,
59(4):849–858, 2003.
[140] Jacob Cohen. Weighted kappa: nominal scale agreement provision for
scaled disagreement or partial credit. Psychological bulletin, 70(4):213,
1968.
[141] Jacob Cohen. A coefficient of agreement for nominal scales. Educational
and psychological measurement, 20(1):37–46, 1960.
[142] Allan Donner and John J Koval. The estimation of intraclass correlation
in the analysis of family data. Biometrics, pages 19–25, 1980.
[143] Ying Guo and Amita K Manatunga. Nonparametric estimation of the
concordance correlation coefficient under univariate censoring. Biomet­
rics, 63(1):164–172, 2007.
[144] Huiman X Barnhart and John M Williamson. Modeling concordance
correlation via gee to evaluate reproducibility. Biometrics, 57(3):931–
940, 2001.
Bibliography 239

[145] S Duggirala, H Larsen, RW Taylor, S Hanson, and CB Nemeroff. Pmcid:

Pmc1381327.
[146] I Lawrence and Kuei Lin. Assay validation using the concordance cor­
relation coefficient. Biometrics, pages 599–604, 1992.
[147] I Lawrence and Kuei Lin. A concordance correlation coefficient to eval­
uate reproducibility. Biometrics, pages 255–268, 1989.
[148] Feng Liang, Rui Paulo, German Molina, Merlise A Clyde, and Jim O
Berger. Mixtures of g priors for bayesian variable selection. Journal of
the american statistical association, 103(481):410–423, 2008.
[149] Jeffrey N Rouder, Paul L Speckman, Dongchu Sun, Richard D Morey,
and Geoffrey Iverson. Bayesian t tests for accepting and rejecting the
null hypothesis. Psychonomic bulletin & review, 16(2):225–237, 2009.
[150] Jeffrey N Rouder, Richard D Morey, Paul L Speckman, and Jordan M
Province. Default bayes factors for anova designs. Journal of mathe­
matical psychology, 56(5):356–374, 2012.
[151] Ruud Wetzels, Don van Ravenzwaaij, and Eric-Jan Wagenmakers.
Bayesian analysis. The encyclopedia of clinical psychology, pages 1–11,
2014.
[152] Josep L Carrasco, Lluis Jover, Tonya S King, and Vernon M Chinchilli.
Comparison of concordance correlation coefficient estimating approaches
with skewed data. Journal of biopharmaceutical statistics, 17(4):673–
684, 2007.
[153] Tonya S King and Vernon M Chinchilli. A generalized concordance
correlation coefficient for continuous and categorical data. Statistics in
medicine, 20(14):2131–2147, 2001.
[154] Xinhua Liu, Yunling Du, Jeanne Teresi, and Deborah S Hasin. Concor­
dance correlation in the measurements of time to event. Statistics in
medicine, 24(9):1409–1420, 2005.
[155] Weining Zhao Robieson. On weighted kappa and concordance correla­
tion coefficient. Ph.D. thesis, 2000.
[156] Rebecca M Turner, Rumana Z Omar, and Simon G Thompson. Con­
structing intervals for the intracluster correlation coefficient using
bayesian modelling, and application in cluster randomized trials. Statis­
tics in medicine, 25(9):1443–1456, 2006.
[157] Robert Tibshirani. Regression shrinkage and selection via the lasso.
Journal of the royal statistical society: series B (Methodological),
58(1):267–288, 1996.
240 Bibliography

[158] Robert Tibshirani. The lasso method for variable selection in the cox
model. Statistics in medicine, 16(4):385–395, 1997.
[159] Jianqing Fan and Runze Li. Variable selection via nonconcave penalized
likelihood and its oracle properties. Journal of the american statistical
Association, 96(456):1348–1360, 2001.
[160] Jianqing Fan and Runze Li. Variable selection for cox’s proportional
hazards model and frailty model. Annals of statistics, pages 74–99,
2002.
[161] Hui Zou and Trevor Hastie. Regularization and variable selection via the
elastic net. Journal of the royal statistical society: series B (statistical
methodology), 67(2):301–320, 2005.
[162] Emmanuel Candes, Terence Tao, et al. The dantzig selector: Statisti­
cal estimation when p is much larger than n. The annals of statistics,
35(6):2313–2351, 2007.
[163] Hemant Ishwaran, Udaya B Kogalur, Eugene H Blackstone, Michael S
Lauer, et al. Random survival forests. The annals of applied statistics,
2(3):841–860, 2008.
[164] Pierre JM Verweij and Hans C Van Houwelingen. Cross-validation in
survival analysis. Statistics in medicine, 12(24):2305–2314, 1993.
[165] Erika Graf, Claudia Schmoor, Willi Sauerbrei, and Martin Schumacher.
Assessment and comparison of prognostic classification schemes for sur­
vival data. Statistics in medicine, 18(17-18):2529–2545, 1999.
[166] Pierre JM Verweij and Hans C Van Houwelingen. Penalized likelihood
in cox regression. Statistics in medicine, 13(23-24):2427–2436, 1994.
[167] Hege M Bøvelstad, Ståle Nygård, Hege L Størvold, Magne Aldrin, Ør­
nulf Borgan, Arnoldo Frigessi, and Ole Christian Lingjærde. Predicting
survival from microarray data—a comparative study. Bioinformatics,
23(16):2080–2087, 2007.
Index

3+3 design, 33
covariance, 169

covariance matrix, 41

Aalen-Nelson estimator, 67
Covariance pattern, 170

ACC, 21
Coverage probability, 28

Accelerated failure time, 105

Cox proportional hazards, 213

Accelerated titration design, 32

Credible band, 76

ALC, 21
CRTSize package, 29

Altshuler estimator, 67
CT scan, 199

Area under the ROC curve (AUC),

Cumulative hazard rate function, 67

33
Cumulative incidence function, 89

Asymptotically, 20

Autoregressive, 186
Data science, 226

Diagnostics for Cox’s PH model , 72

Bayesian factor, 202

Diagnostics test, 197

Bayesian frailty survival, 104

DNA methylation expression, 209

Bayesian model average, 75

Dose escalation step dose, 32

Bayesian sample size calculation, 23

Dose selection algorithm, 51

Bayeslongitudinal, 134
Dose-limiting toxicities (DLTs), 32

BayesX package, 105

Dose-limiting toxicities(DLTs), 189

Breslow estimator, 215

DPpackage, 134

BVSNLP package, 215

Dynamic linear mixed effect model,

186

Circulating endothelial cell, 48

Clinical trial, 18, 19

Elastic-Net, 223

coda package, 105

Equivalence trial, 26

Competing risk, 87
EurosarcBayes package, 29

Competing risk plot, 89

Compound symmetry, 170

fields package, 105

Comprehensive R archive network

First order, 170

(CRAN), 6
Fixed effect model, 188

Concordance correlation
flexsurv package, 120

coefficient(CCC), 199
Frailty data analysis, 97

Conditonal probability, 78
Frailty on recurrent events, 104

Confidence interval, 204

Gene expression omnibus, 154

Continual reassessment method

General covariance structure, 170

(CRM), 34

Gibbs, 7

coprimary package, 29

glm, 185

241
242 Index

Hazard function, 65
Mixed effect model, 187

Heterogeneous compound symmetry,

Mixed-effect model, 169

171
Molecular targeted agents(MTA),

Heterogeneous first-order
186

autoregressive, 171
MRI, 199

Heterogeneous toeplitz, 171

Multivariate normal distribution, 171

Heterogeneous uncorrelated, 171

High dimensional data analysis, 210

Net survival, 111

HighDimOut package, 224

Non-inferiority trial, 26

Highest posterior density (HPD), 20

Non-shrinkage failure, 41

Hypothesis, 14

Oncology trial, 18, 32

JM package, 166
OpenBUGS, 7

Joint modeling, 153

Optimum, 21

JZS prior, 202

Optimum biological dose (OBD), 186

Overall survival, 131

K-fold cross-validation, 226

Kaplan-Meier estimator, 66
Partial log-likelihood, 215

Penalized cox’s regression, 223

L1 penalty function, 210

Pharmacologically guided dose

LASSO, 210
escalation(PGDE) design,

Linear mixed effect, 188

33

lme4, 186
Phase I trial, 32

lmer, 185
Phase II, 39

Log-Rank, 13
Piecewise Hazard, 118

Logistics model, 41
Piecewise hazard, 118

Longitudinal data, 169

Posterior credible interval, 28

Longitudinal data analysis, 131

Posterior distribution, 18

longpower package, 29
Posterior error rate, 23

Posterior inclusion probability, 216

Marginal posterior distribution, 19

Power, 19

Maximum partial likelihood, 69

Principal component analysis, 212

Maximum tolerable dose (MTD), 32,

Prior distribution, 18

195
Product inverse moment (piMOM),

MCMC (Markov chain Monte

216

Carlo), 7
Product limit, 66

MCMCglmm, 192
Product-moment (pMOM), 216

Median probability model (MPM),

Proportional hazards, 215

216

Metronomic chemotheapy, 190

QOL, 131

Metropolis, 7
Quadratic logistic design, 52

Microarray, 209

Missing at random, 144

R, 3

Missing at random (MAR), 41

R packages, 3

Missing data analysis, 143

R2BayesX package, 105

Missing not at random, 144

Random effect model, 188

Index 243

RCTs, 14
Relative survival, 111, 120
Right censoring, 65
Risk, 19
Rule-based design, 32

Sample size, 14, 18

samplesizelogisticcasecontrol
pacakge, 29
SIMR package, 194
spBayesSurv, 106
spBayesSurv package, 105
Superiority trial, 26
Surveillance, epidemiology and end
results, 111
Survival, 18
Survival analysis, 13
survival package, 66, 105

Time course data, 173

Time-dependent survival model, 154
Time-to-event, 68
Toeplitz, 170
Toxicity profile testing, 48
Type-I error, 18
Type-II error, 18

Vascular endothelial cells (VECs),

195

Wald score, 71
Wald statistics, 42
Walds test, 119