Model-Assisted Bayesian

Designs for Dose Finding

and Optimization
Methods and Applications

Bayesian adaptive designs provide a critical approach to improve the efficiency and success of
drug development that has been embraced by the US Food and Drug Administration (FDA).
This is particularly important for early phase trials as they form the basis for the development
and success of subsequent phase II and III trials.
The objective of this book is to describe the state-of-the-art model-assisted designs to facilitate
and accelerate the use of novel adaptive designs for early phase clinical trials. Model-assisted
designs possess avant-garde features where superiority meets simplicity. Model-assisted
designs enjoy exceptional performance comparable to more complicated model-based adap-
tive designs, yet their decision rules often can be pre-tabulated and included in the protocol—
making implementation as simple as conventional algorithm-based designs. An example is
the Bayesian optimal interval (BOIN) design, the first dose-finding design to receive the fit-
for-purpose designation from the FDA. This designation underscores the regulatory agency’s
support of the use of the novel adaptive design to improve drug development.
Features

• Represents the first book to provide comprehensive coverage of model-assisted

designs for various types of dose-finding and optimization clinical trials
• Describes the up-to-date theory and practice for model-assisted designs
• Presents many practical challenges, issues, and solutions arising from early-phase
clinical trials
• Illustrates with many real trial applications
• Offers numerous tips and guidance on designing dose finding and optimization trials
• Provides step-by-step illustrations of using software to design trials
• Develops a companion website (www.trialdesign.org) to provide freely available,
easy-to-use software to assist learning and implementing model-assisted designs
Written by internationally recognized research leaders who pioneered model-assisted designs
from the University of Texas MD Anderson Cancer Center, this book shows how model-
assisted designs can greatly improve the efficiency and simplify the design, conduct, and
optimization of early-phase dose-finding trials. It should therefore be a very useful practical
reference for biostatisticians, clinicians working in clinical trials, and drug regulatory profes-
sionals, as well as graduate students of biostatistics. Novel model-assisted designs showcase
the new KISS principle: Keep it simple and smart!
Chapman & Hall/CRC Biostatistics Series

Series Editors
Shein-Chung Chow, Duke University School of Medicine, USA
Byron Jones, Novartis Pharma AG, Switzerland
Jen-pei Liu, National Taiwan University, Taiwan
Karl E. Peace, Georgia Southern University, USA
Bruce W. Turnbull, Cornell University, USA

Recently Published Titles

Statistical Thinking in Clinical Trials

Michael A. Proschan
Simultaneous Global New Drug Development
Multi-Regional Clinical Trials after ICH E17
Edited by Gang Li, Bruce Binkowitz, William Wang, Hui Quan, and Josh Chen
Quantitative Methodologies and Process for Safety Monitoring and Ongoing
Benefit Risk Evaluation
Edited by William Wang, Melvin Munsaka, James Buchanan and Judy Li
Statistical Methods for Mediation, Confounding and Moderation Analysis
Using R and SAS
Qingzhao Yu and Bin Li
Hybrid Frequentist/Bayesian Power and Bayesian Power in Planning Clinical
Trials
Andrew P. Grieve
Advanced Statistics in Regulatory Critical Clinical Initiatives
Edited By Wei Zhang, Fangrong Yan, Feng Chen, Shein-Chung Chow
Medical Statistics for Cancer Studies
Trevor F. Cox
Real World Evidence in a Patient-Centric Digital Era
Edited by Kelly H. Zou, Lobna A. Salem, Amrit Ray
Data Science, AI, and Machine Learning in Pharma
Harry Yang
Model-Assisted Bayesian Designs for Dose Finding and Optimization
Methods and Applications
Ying Yuan, Ruitao Lin, and J. Jack Lee

For more information about this series, please visit: https://www.routledge.com/

Chapman--Hall-CRC-Biostatistics-Series/book-series/CHBIOSTATIS
Model-Assisted Bayesian
Designs for Dose Finding
and Optimization
Methods and Applications

Ying Yuan
The University of Texas MD Anderson Cancer Center, USA

Ruitao Lin
The University of Texas MD Anderson Cancer Center, USA

J. Jack Lee
The University of Texas MD Anderson Cancer Center, USA
First edition published 2023
by CRC Press
6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742

and by CRC Press

4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN

© 2023 Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, LLC

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

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or here-
after 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 [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.

Library of Congress Cataloging-in-Publication Data

Names: Yuan, Ying (Professor of biostatistics), author. | Lin, Ruitao,

author. | Lee, J. Jack, author.
Title: Model-assisted Bayesian designs for dose finding and optimization :
methods and applications / Ying Yuan, Ruitao Lin, J. Jack Lee.
Other titles: Chapman & Hall/CRC biostatistics series.
Description: First edition. | Boca Raton : C&Hall/CRC Press, 2023. |
Series: Chapman & Hall/CRC biostatistics series | Includes
bibliographical references and index.
Identifiers: LCCN 2022017949 (print) | LCCN 2022017950 (ebook) | ISBN
9780367146245 (hardback) | ISBN 9781032357126 (paperback) | ISBN
9780429052781 (ebook)
Subjects: MESH: Drug Development--methods | Drug Design--methods | Drug
Dosage Calculations | Models, Statistical | Bayes Theorem | Clinical
Trials, Phase I as Topic
Classification: LCC RM301.25 (print) | LCC RM301.25 (ebook) | NLM QV 745
| DDC 615.1/9--dc23/eng/20220801
LC record available at https://lccn.loc.gov/2022017949
LC ebook record available at https://lccn.loc.gov/2022017950

ISBN: 978-0-367-14624-5 (hbk)

ISBN: 978-1-032-35712-6 (pbk)
ISBN: 978-0-429-05278-1 (ebk)

DOI: 10.1201/9780429052781
Typeset in CMR10
by KnowledgeWorks Global Ltd.

Publisher’s note: This book has been prepared from camera-ready copy provided by the authors.
 To my wife, Suyu, and daughter, Selina.

Ying Yuan

To the memory of my mother, a brave cancer fighter.

Ruitao Lin

To my wife, Vei-Vei, and children, Joseph and Hope, for their

love and support.

J. Jack Lee
Contents

Preface xi

Author Biographies xiii

1 Bayesian Statistics and Adaptive Designs 1

1.1 Basics of Bayesian statistics . . . . . . . . . . . . . . . . . . 1

1.1.1 Bayes’ theorem . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Bayesian inference . . . . . . . . . . . . . . . . . . . . 5
1.2 Bayesian adaptive designs . . . . . . . . . . . . . . . . . . . . 7
1.3 Adoption of Bayesian adaptive designs . . . . . . . . . . . . 10

2 Algorithm-Based and Model-Based Dose Finding Designs 13

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Traditional 3+3 design . . . . . . . . . . . . . . . . . . . . . 16
2.3 Cohort expansion . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4 Accelerated titration design . . . . . . . . . . . . . . . . . . . 20
2.5 Continual reassessment method . . . . . . . . . . . . . . . . 21
2.6 Bayesian model averaging CRM . . . . . . . . . . . . . . . . 25
2.7 Escalation with overdose control . . . . . . . . . . . . . . . . 28
2.8 Bayesian logistic regression method . . . . . . . . . . . . . . 30
2.9 Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3 Model-Assisted Dose Finding Designs 33

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Modified toxicity probability interval design . . . . . . . . . 33
3.3 Keyboard design . . . . . . . . . . . . . . . . . . . . . . . . 35
3.4 Bayesian optimal interval (BOIN) design . . . . . . . . . . . 38
3.4.1 Trial design . . . . . . . . . . . . . . . . . . . . . . . . 38
3.4.2 Theoretical derivation . . . . . . . . . . . . . . . . . . 42
3.4.3 Specification of design parameters . . . . . . . . . . . 46
3.4.4 Statistical properties . . . . . . . . . . . . . . . . . . . 47
3.4.5 Frequently asked questions . . . . . . . . . . . . . . . 49
3.5 Operating characteristics . . . . . . . . . . . . . . . . . . . . 52
3.6 Software and case study . . . . . . . . . . . . . . . . . . . . . 57

vii
viii

4 Drug-Combination Trials 69

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Model-based designs . . . . . . . . . . . . . . . . . . . . . . . 71
4.3 Model-assisted designs . . . . . . . . . . . . . . . . . . . . . . 74
4.3.1 BOIN combination design . . . . . . . . . . . . . . . . 74
4.3.2 Keyboard combination design . . . . . . . . . . . . . . 77
4.3.3 Waterfall design . . . . . . . . . . . . . . . . . . . . . 77
4.4 Operating characteristics . . . . . . . . . . . . . . . . . . . . 82
4.5 Software and case study . . . . . . . . . . . . . . . . . . . . . 85

5 Late-Onset Toxicity 93

5.1 A common logistical problem . . . . . . . . . . . . . . . . . . 93

5.2 Late-onset toxicities . . . . . . . . . . . . . . . . . . . . . . . 94
5.3 TITE-CRM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.4 TITE-BOIN . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.4.1 Trial design . . . . . . . . . . . . . . . . . . . . . . . . 98
5.4.2 Incorporating prior information . . . . . . . . . . . . . 102
5.4.3 Statistical properties . . . . . . . . . . . . . . . . . . . 103
5.4.4 Operating characteristics . . . . . . . . . . . . . . . . 103
5.5 A unified approach using “effective” data . . . . . . . . . . . 105
5.6 TITE-keyboard and TITE-mTPI designs . . . . . . . . . . . 109
5.7 Software and case study . . . . . . . . . . . . . . . . . . . . . 112

6 Incorporating Historical Data 119

6.1 Historical data and prior information . . . . . . . . . . . . . 119

6.1.1 Incorporate prior information in CRM . . . . . . . . . 119
6.2 BOIN with Informative Prior (iBOIN) . . . . . . . . . . . . . 121
6.2.1 Trial design . . . . . . . . . . . . . . . . . . . . . . . . 121
6.2.2 Practical guidance . . . . . . . . . . . . . . . . . . . . 125
6.3 iKeyboard design . . . . . . . . . . . . . . . . . . . . . . . . 126
6.4 Operating characteristics . . . . . . . . . . . . . . . . . . . . 127
6.5 Software and case study . . . . . . . . . . . . . . . . . . . . . 128

7 Multiple Toxicity Grades 137

7.1 Multiple toxicity grades . . . . . . . . . . . . . . . . . . . . . 137

7.2 gBOIN accounting for toxicity grade . . . . . . . . . . . . . . 140
7.2.1 Trial design . . . . . . . . . . . . . . . . . . . . . . . . 140
7.2.2 Statistical derivation and properties . . . . . . . . . . 142
7.3 Multiple toxicity BOIN . . . . . . . . . . . . . . . . . . . . . 144
7.3.1 Trial design . . . . . . . . . . . . . . . . . . . . . . . . 144
7.3.2 Statistical derivation and properties . . . . . . . . . . 147
7.4 Software and illustration . . . . . . . . . . . . . . . . . . . . 149
 ix

8 Finding Optimal Biological Dose 155

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 155

8.1.1 Phase I–II design paradigm . . . . . . . . . . . . . . . 156
8.2 EffTox design . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.2.1 Efficacy–toxicity trade-off . . . . . . . . . . . . . . . . 159
8.2.2 Joint probability model for efficacy and toxicity . . . . 160
8.2.3 Admissible rules . . . . . . . . . . . . . . . . . . . . . 161
8.2.4 Dose-finding algorithm . . . . . . . . . . . . . . . . . . 161
8.3 U-BOIN design . . . . . . . . . . . . . . . . . . . . . . . . . . 162
8.3.1 Utility-based risk-benefit trade-off . . . . . . . . . . . 163
8.3.2 Statistical model . . . . . . . . . . . . . . . . . . . . . 165
8.3.3 Admissible rules . . . . . . . . . . . . . . . . . . . . . 166
8.3.4 Dose-finding algorithm . . . . . . . . . . . . . . . . . . 167
8.3.5 Operating characteristics . . . . . . . . . . . . . . . . 170
8.4 BOIN12 design . . . . . . . . . . . . . . . . . . . . . . . . . . 171
8.4.1 Utility estimation using quasi-binomial likelihood . . . 173
8.4.2 Admissible rules and dose comparison . . . . . . . . . 174
8.4.3 Dose-finding rules . . . . . . . . . . . . . . . . . . . . 176
8.4.4 Operating characteristics . . . . . . . . . . . . . . . . 180
8.4.5 Extension to more complicated endpoints . . . . . . . 181
8.5 TITE-BOIN12 design . . . . . . . . . . . . . . . . . . . . . . 183
8.6 Other model-assisted phase I/II designs . . . . . . . . . . . . 188
8.6.1 uTPI design . . . . . . . . . . . . . . . . . . . . . . . . 188
8.6.2 STEIN and BOIN-ET designs . . . . . . . . . . . . . . 190
8.7 Software and case study . . . . . . . . . . . . . . . . . . . . . 193

Bibliography 205

Index 217
Preface

Despite extensive efforts, the drug development enterprise is challenged by

prohibitively high costs and slow progress. To address these pressing issues,
one important approach embraced by the US Food and Drug Administration
(FDA) is the use of novel adaptive designs. This is particularly important
for early phase clinical trials, which serve as the gatekeepers to identifying
promising drugs and optimizing treatment regimens (e.g., dose) before launch-
ing expensive phase III trials. A major barrier to the use of these tools is that
most novel adaptive designs are difficult to understand, requiring complicated
statistical modeling, complex computation, and expensive infrastructure for
implementation.
The objective of this book is to systematically introduce and review a class
of novel adaptive designs, known as model-assisted designs, to remove this bar-
rier while increasing the use of novel adaptive designs for early phase clinical
trials. Model-assisted designs enjoy superior performance comparable to more
complicated model-based adaptive designs, yet their decision rule often can be
pre-tabulated and included in the protocol—making implementation as sim-
ple as conventional algorithm-based designs. Thus, these novel model-assisted
design features create a space where superiority meets simplicity. They ad-
here to the new KISS principle: Keep it simple and smart! A typical example
is the Bayesian optimal interval (BOIN) design, the first maximum-tolerated
dose (MTD)-finding design to receive the fit-for-purpose designation from the
FDA as a tool for drug development. This designation underscores the regu-
latory agency’s support for the use of novel adaptive design to improve drug
development.
This book is intended for biostatisticians working on the design and anal-
ysis of clinical trials, as well as clinical trialists with an interest in learning
adaptive designs. One unique feature of this book is that, besides statistical
methodology, it also provides a detailed description of each design’s software
and step-by-step guidance for clinical trial design and conduct using real-world
examples. A great amount of effort has been devoted to developing the com-
panion website www.trialdesign.org to provide freely available, easy-to-use
software, along with extensive help and documentation, to assist readers in
learning and implementing the designs in this book.
Our hope is that this book and the accompanying software will provide
a practical toolkit for practitioners to implement novel adaptive designs for
their own trials. Our overarching goal is to benefit patients by accelerating
drug development and clinical research.

xi
xii Preface

It is almost impossible to thank all of the people who have contributed

in various ways to making this book a reality. We are extremely grate-
ful to all of our students, postdocs, and colleagues who helped develop the
novel methodology and software in this book—in particular, Suyu Liu, Ph.D.;
Peter F. Thall, Ph.D.; Kenneth R. Hess, Ph.D.; Susan G. Hilsenbeck, Ph.D.;
Mark R. Gilbert, M.D.; Jing Wu, M.D.; Guosheng Yin, Ph.D.; Sumithra J.
Mandrekar, Ph.D.; Yanhong Zhou, Ph.D.; Heng Zhou, Ph.D.; Fangrong Yan,
Ph.D.; Haitao Pan, Ph.D.; Rongji Mu, Ph.D.; Kai Chen, M.S.; Ying-Wei Kuo,
M.S.; and Nan Chen, Ph.D. We would also like to express our sincere grati-
tude to our industrial or regulatory collaborators including Daniel Li, Ph.D.;
Lei Nie, Ph.D.; Sammi Tang, Ph.D.; Li Wang, Ph.D.; and Zhiying Pan, Ph.D.
for their insightful discussions and suggestions. We thank Jessica Swann for
her detailed reading and editorial assistance, and Evan Kwiatkowski, Grace
Nie, Peng Yang, Feng Tian, Xiaohan Chi, Jingyi Zhang, Mengyi Lu, and other
students for proofreading.

Ying Yuan
Ruitao Lin
J. Jack Lee
Author Biographies

Ying Yuan, Ph.D., is Bettyann Asche Murray Distinguished Professor in

Biostatistics and Deputy Chair at the Department of Biostatistics at the Uni-
versity of Texas MD Anderson Cancer Center. He has published over 100
statistical methodology papers on innovative Bayesian adaptive designs, in-
cluding early phase trials, seamless trials, biomarker-guided trials, and basket
and platform trials. The designs and software developed by Dr. Yuan’s and Dr.
J. Jack Lee’s team (www.trialdesign.org) have been widely used in medical re-
search institutes and pharmaceutical companies. The BOIN design developed
by Dr. Yuan’s team is the first oncology dose-finding design designated as a fit-
for-purpose drug development tool by FDA. Dr. Yuan is an elected Fellow of
the American Statistical Association, and is a co-author of the book Bayesian
Designs for Phase I-II Clinical Trials published by Chapman & Hall/CRC
Press.
Ruitao Lin, Ph.D., is an Assistant Professor in the Department of Biostatis-
tics at the University of Texas MD Anderson Cancer Center. Motivated by
the unmet need for the development of precision medicine, Dr. Lin has devel-
oped many innovative statistical designs to increase trial efficiency, optimize
healthcare decisions, and expedite drug development. He made substantial
contributions to generalize model-assisted designs, including BOIN, to han-
dle combination trials, late-onset toxicity, and dose optimization. Dr. Lin has
published over 40 papers in top statistical and medical journals. He currently
is an Associate Editor of Biometrical Journal, Pharmaceutical Statistics, and
Contemporary Clinical Trials.
J. Jack Lee, Ph.D., is a Professor of Biostatistics, Kenedy Foundation Chair
in Cancer Research, and Associate Vice President in Quantitative Sciences at
the University of Texas MD Anderson Cancer Center. He is an expert on the
design and analysis of Bayesian adaptive designs, platform trials, basket tri-
als, umbrella trials, master protocols, statistical computation/graphics, drug
combination studies, and biomarkers identification and validation. Dr. Lee has
also been actively participating in basic, translational, and clinical cancer re-
search in chemoprevention, immuno-oncology, and precision oncology. He is an
elected Fellow of the American Statistical Association, the Society for Clinical
Trials, and the American Association for the Advancement of Science. He is
Statistical Editor of Cancer Prevention Research and serves on the Statistical
Editorial Board of Journal of the National Cancer Institute. He has over 500
publications and is a co-author of the book Bayesian Adaptive Methods for
Clinical Trials published by Chapman & Hall/CRC Press.
xiii
1
Bayesian Statistics and Adaptive Designs

1.1 Basics of Bayesian statistics

1.1.1 Bayes’ theorem
Clinical trials are prospective studies to evaluate the health-related biomedical
effect of one or more interventions in human subjects. An intervention could be
a drug, medical device, or biologic such as a vaccine or gene therapy. Through
clinical trials, data D are collected from study participants to inform the
parameters of interest, θ, such as the treatment effect or the safety of the
intervention. There are two major paradigms to make statistical inference on
θ: frequentist and Bayesian. The frequentist paradigm assumes that θ is a fixed
unknown quantity and D are random. The inference of θ is made by modeling
the distribution of D, given θ, known as likelihood L(D|θ). The most common
inferential method is the maximum likelihood method.
In contrast, the Bayesian paradigm assumes that D are fixed, as they have
been already observed and do not change, and the unknown parameter θ is
random. In the Bayesian framework, the unknown parameter has a distribu-
tion, which is used to quantify its uncertainty. Bayesian inference is based on
the posterior distribution of θ given D. This is done through Bayes’ theorem,
which describes the conditional probability of an event A given an event B,
formally:
Pr(A ∩ B) Pr(A) Pr(B | A)
Pr(A | B) = = .
Pr(B) Pr(B)
Here, Pr(A) ≥ 0 and Pr(B) > 0 are the probabilities of observing the events
A and B, respectively. Pr(A | B) is the conditional probability of A given
B occurring, and Pr(A ∩ B) is the probability of both A and B occurring.
Applying Bayes’ theorem, the posterior distribution of θ, f (θ | D), is given by

f (θ)L(D | θ)
f (θ | D) = ,
f (D)

where f (θ) is the

R prior distribution of θ, quantifying the prior knowledge on
θ, and f (D) = f (θ)L(D | θ)dθ is the marginal likelihood of D, serving as
a normalizing constant to ensure
R that the posterior distribution is a valid
density function, such that f (θ | D)dθ = 1. As the normalizing constant

DOI: 10.1201/9780429052781-1 1
2 Model-Assisted Bayesian Designs for Dose Finding and Optimization

f (D) does not depend on θ, the (unnormalized) posterior distribution can

also be expressed as
f (θ | D) ∝ f (θ)L(D | θ). (1.1)
This simpler expression encapsulates the essence of Bayesian inference: the
posterior distribution is proportional to the product of the prior distribution
and the likelihood function. In other words, the posterior distribution is the
outcome of synthesizing the information contained in the prior distribution
and data.
Compared to frequentist approaches, the Bayesian paradigm has several
advantages that make it particularly suitable for clinical trial design. First,
as shown in (1.1), it provides a convenient and principled way to incorporate
prior information, when available, into the current clinical trial through the
specification of the prior distribution f (θ). The prior information may come
from preclinical data, results of clinical trials of the drug in different patient
populations, and expert opinions. When the prior distribution is appropri-
ately specified, Bayesian methods improve the trial efficiency and inference
accuracy; see examples below.
Second, the Bayesian paradigm provides a coherent mechanism that allows
us to continuously update our knowledge on θ and accordingly make appro-
priate clinical decisions based on accumulating data. Due to ethical consider-
ations, clinical trials are often conducted in a group sequential way—treating
patients by cohorts. Treatment decisions for later cohorts are made based on
the outcomes of early cohorts. For example, in phase I trials, if the data from
treated patients (cohorts) show that the current dose is overly toxic, we should
de-escalate the dose to treat the next cohort. In phase II trials, if the data
from treated patients indicate that the experimental drug is most likely inef-
fective or too toxic, we may terminate the trial early for futility or toxicity.
Let D1 , · · · , DK denote the data observed from cohorts 1, · · · , K. Under the
Bayesian paradigm, our knowledge on θ is updated as follows: starting from the
prior knowledge of θ, i.e., f (θ), after observing D1 , we update our knowledge
of θ and obtain the posterior knowledge of θ as f (θ | D1 ) ∝ f (θ)L(D1 | θ),
which is used to guide the treatment decision for cohort 2. As this poste-
rior represents our prior knowledge of θ before observing D2 , it serves as a
prior distribution before the next step. After observing D2 , we apply Bayes’
theorem and update our knowledge of θ again as follows:

f (θ | D1 , D2 ) ∝ f (θ|D1 )L(D2 | θ) ∝ f (θ)L(D1 | θ)L(D2 | θ).

This posterior is then used to guide the treatment decision for cohort 3. By
continuing this process, we can learn about θ successively and refine treatment
decisions for patients as the trial progresses.
Third, Bayesian methods conform to the likelihood principle, which states
that all information for making inference on the parameters is contained in
the observed data only and not in the unobserved data. Fourth, Bayesian
Bayesian Statistics and Adaptive Designs 3

modeling can naturally and appropriately address all levels of uncertainty in

obtaining the data.

Example 1.1 (Beta-binomial model) Suppose that we observe that y out of

n patients have responded favorably by taking a new drug, the observed data
D = (y, n) can be modeled by a binomial distribution, y ∼ Binomial(n, θ),
with 0 ≤ θ ≤ 1 denoting the response rate. The likelihood function of D is
 
n
L(D | θ) = θy (1 − θ)n−y .
y

Let f (θ) be the prior of θ, which takes a beta distribution, θ ∼ Beta(α, β)

with the density of
Γ(α + β) α−1
f (θ) = θ (1 − θ)β−1 ,
Γ(α)Γ(β)
where α > 0 and β > 0 are hyperparameters. Applying Bayes’ theorem, the
posterior distribution of θ is given by

f (θ | D) ∝ f (θ)L(D | θ)
∝ θα−1 (1 − θ)β−1 θy (1 − θ)n−y
∝ θα+y−1 (1 − θ)β+n−y−1 . (1.2)

Noting that (1.2) is the kernel of the beta distribution, the posterior distribu-
tion of θ follows a beta distribution

θ | D ∼ Beta(α + y, β + n − y).

Thus, α and β can be interpreted as the prior numbers of responders and

nonresponders, respectively, and (α + β) is the prior effective sample size.
To see the impact of the prior on the posterior distribution, we as-
sume (y, n) = (10, 25) and consider the following three beta priors: (a)
θ ∼ Beta(0.5, 0.5), which is a vague (or weakly informative) prior with the
prior effective sample size equal to one patient; (b) θ ∼ Beta(2, 6), a mod-
erately informative prior containing the information corresponding to eight
patients; and (c) θ ∼ Beta(150, 300), a strongly informative prior containing
the information corresponding to 450 patients. The resulting posterior dis-
tributions based on the three priors are Beta(10.5, 15.5), Beta(12, 21), and
Beta(160, 315), respectively. Figure 1.1 exhibits the prior, likelihood (unnor-
malized), and posterior densities of θ based on the Beta-binomial model. It
shows that when the prior is vague, the likelihood and posterior functions are
almost identical, and when the prior is extremely informative, as exemplified
by the prior Beta(150, 300), the prior would dominate the posterior distribu-
tion, and the observed data play an extremely limited role in the inference of
θ.
Example 1.2 (Normal-normal model with a known variance) Suppose
that in a clinical study of n patients, the outcome yi for the ith patient
4 Model-Assisted Bayesian Designs for Dose Finding and Optimization

D θa%HWD  E θa%HWD  F θa%HWD 

 /LNHOLKRRG /LNHOLKRRG /LNHOLKRRG

3ULRU 3ULRU 3ULRU
3RVWHULRU 3RVWHULRU 3RVWHULRU





'HQVLW\

'HQVLW\

'HQVLW\










                 

3UREDELOLW\RIUHVSRQVH 3UREDELOLW\RIUHVSRQVH 3UREDELOLW\RIUHVSRQVH

FIGURE 1.1: Prior, likelihood (unnormalized), and posterior densities for the
probability of response θ, based on the Beta-binomial model with different
beta priors.

follows a normal distribution yi ∼ N(µ, σ 2 ), i = 1, . . . , n, where the mean µ is

unknown and the variance σ 2 is known. Based on the data from n patients,
D = {y1 , . . . , yn }, the likelihood function is given by
n
( n
)
Y
2 1 1 X 2
L(D | µ) = f (yi | µ, σ ) = n exp − 2 (yi − µ) ,
i=1
σ (2π)n/2 2σ i=1

where f (yi | µ, σ 2 ) denotes the density of the normal distribution. Assuming

the prior distribution µ ∼ N(θ0 , τ 2 ), where θ0 and τ 2 are hyperparameters,
the posterior distribution of µ is derived as

f (µ | D) ∝ f (µ | θ0 , τ 2 )L(D | µ)
  ( n
)
1 1 X
∝ exp − 2 (µ − θ0 )2 exp − 2 (yi − µ)2
2τ 2σ i=1
  Pn 2 
 σ 2 θ0 + τ 2 i=1 yi 

 µ− 

 σ 2 + nτ 2 
∝ exp − 2 2

 2τ σ 

 
 2
σ + nτ 2 
Pn
θ0 /τ 2 + i=1 yi /σ 2
 
1
= N µ| , .
1/τ 2 + n/σ 2 1/τ 2 + n/σ 2

As can be seen, the posterior mean is the weighted average of the prior mean
and the observed sample mean, weighted by the respective variance.
In Example 1.1 or 1.2, the posterior distribution is in the same distribu-
tional family as the prior distribution, greatly facilitating the posterior com-
putation. This type of prior is called a conjugate prior. For many statistical
models, however, the conjugate prior does not exist. In these cases, the pos-
terior distribution does not have a form of standard distributions, and the
Bayesian Statistics and Adaptive Designs 5

inference on θ can be made by drawing posterior samples of θ from f (θ | D)

via Markov chain Monte Carlo (MCMC) methods, such as the Gibbs sampler
or Metropolis-Hastings algorithms. More details about MCMC methods are
covered in Bayesian textbooks, including Gelman et al. (2013), Carlin and
Louis (2008), and Robert and Casella (2013).

1.1.2 Bayesian inference

According to Bayes’ theorem, the posterior distribution contains all relevant
information we know about the parameter of interest θ in light of the data
and prior knowledge. Once the posterior distribution is available, Bayesian
inference about θ can be readily made by exploiting this distribution.

Point and interval estimations

For illustrative purposes, we assume that θ is univariate. Based on the poste-
rior distribution f (θ | D), commonly used Bayesian point estimators include:
R
• Posterior mean: θ̂ = E(θ | D) = θf (θ | D)dθ;
R θ̂
• Posterior median: θ̂ such that −∞
f (θ | D)dθ = 0.5;

• Posterior mode: θ̂ = arg maxf (θ | D).

θ

When a non-informative prior or the flat prior f (θ) ∝ 1 is used, the posterior
mode estimator is often equal or similar to the frequentist maximum likelihood
estimator.
Because Bayesian statistics treat the unknown parameter θ as a random
variable, it is straightforward to make interval inference about θ based on its
posterior distribution. Analogous to the frequentist confidence interval, the
100(1 − α)% Bayesian credible interval Cα is given by
Z
f (θ | D)dθ = 1 − α,
Cα

where α is a prespecified value such as 0.05. Unlike the frequentist confidence

interval, whose interpretation relies on repeated sampling (i.e., repeat the trial
many times), the Bayesian credible interval has a more intuitive probability
interpretation: the 100(1 − α)% Bayesian credible interval is the interval that
has a 100(1 − α)% chance to contain the true value of θ.
There are different ways to construct credible intervals. A commonly-used
one is the equal-tailed credible interval Cα = (θα/2 , θ1−α/2 ), where θα/2 and
θ1−α/2 satisfy
Z θα/2 Z θ1−α/2
f (θ | D)dθ = α/2, and f (θ | D)dθ = 1 − α/2.
−∞ −∞
6 Model-Assisted Bayesian Designs for Dose Finding and Optimization

Alternatively, the highest posterior density (HPD) interval gives the short-
est length among all credible intervals with a given α. A 100(1 − α)% HPD
interval for θ is given by

Cα = {θ : f (θ | D) ≥ πα },

where πα is chosen such that

Z
Pr(θ ∈ Cα | D) = f (θ | D)dθ = 1 − α.
θ:f (θ|D)≥πα

When the posterior function f (θ | D) is unimodal and symmetric, the equal-

tailed credible interval coincides with the HPD interval. If the posterior func-
tion has more than one mode, then it is possible that the HPD interval is a
union of several disjointed regions.
In addition, under the Bayesian paradigm, it is straightforward to esti-
mate interval probabilities that are of practical interest and importance, e.g.,
Pr(θ > φ) where φ is a pre-specified cutoff. For example, in Example 1.1, in-
vestigators may be interested in estimating the probability that the response
rate θ is greater than 0.4, or less than 0.2, which can be useful for treatment
decisions and easily computed under the Bayesian framework. However, these
quantities cannot be computed under the frequentist framework because the
unknown parameter is assumed to be fixed and does not have a distribution.

Prediction
Another strength of Bayesian statistics is its ability to make predictions on
future observations. Assuming that the future observations D̃ and the cur-
rent data D are conditionally independent given θ, the posterior predictive
distribution for D̃ is given by
Z
f (D̃ | D) = f (D̃, θ | D)dθ
Z
= f (D̃ | θ, D)f (θ | D)dθ
Z
= L(D̃ | θ)f (θ | D)dθ.

Here, f (D̃, θ | D) denotes the joint probability function of D̃ and θ given D,

and f (D̃ | θ, D) denotes the density function of D̃ given θ and D. The last
equation shows that the posterior predictive distribution of D̃ is formed by
averaging the distribution of D̃, given a value of θ, over all possible values of
θ under its posterior distribution f (θ | D).

Bayesian hypothesis testing

The cornerstone of Bayesian hypothesis testing is the Bayes factor. Consider
two candidate models (or hypotheses) H0 and H1 , and let L(D | θk , Hk )
Bayesian Statistics and Adaptive Designs 7

denote the likelihood function, characterized by model parameters θk , under

model Hk , k = 0, 1. In some applications, it is possible that the competing
models have the same set of parameters but this is not necessary. Given a prior
distribution f (θk | Hk ), the marginal likelihood given model Hk , which is the
probability of the observed data D given model Hk , is obtained by integrating
out θk from the joint distribution of (D, θk ) given Hk , that is,
Z Z
Pr(D | Hk ) = f (D, θk | Hk )dθk = L(D | θk , Hk )f (θk | Hk )dθk ,

for k = 0, 1. The Bayes factor is defined as the ratio of the marginal likelihoods
under H1 verus H0
Pr(D | H1 )
BF10 = .
Pr(D | H0 )
Let Pr(Hk ) denote the prior probability of Hk being true, the posterior prob-
ability of Hk being true is given by

Pr(D | Hk ) Pr(Hk )
Pr(Hk | D) =
Pr(D)
Pr(D | Hk ) Pr(Hk )
= .
Pr(D | H0 ) Pr(H0 ) + Pr(D | H1 ) Pr(H1 )

Then, the posterior odds in favor of H1 versus H0 , given D, is given by

Pr(H1 | D) Pr(H1 )
= BF10 .
Pr(H0 | D) Pr(H0 )

Therefore, the Bayes factor can be interpreted as a multiplication factor that

converts the prior odds of candidate models being true (i.e., Pr(H1 )/ Pr(H0 ))
into the posterior odds of candidate models being true. When the two models
are equally probable a priori with Pr(H0 ) = Pr(H1 ) = 1/2, the Bayes fac-
tor reduces to the posterior odds. In other words, the Bayes factor measures
the evidence provided by the observed data in favor of one model against the
other. A value of BF10 > 1 indicates that H1 is more likely to be true than H0 ,
given the observed data D, compared to the belief before seeing the observa-
tion. In the context of hypothesis testing, Kass and Raftery (1995) provided
interpretations of the Bayes factor on the log10 scale as a measure to quantify
the level of evidence, see Table 1.1.

1.2 Bayesian adaptive designs

Traditionally, clinical trials are often conducted in a “static” way by fixing de-
sign parameters in advance (e.g., the sample size, population to be enrolled,
8 Model-Assisted Bayesian Designs for Dose Finding and Optimization

TABLE 1.1: The interpretation of the Bayes factor as a measure to quantify

the level of evidence according to Kass and Raftery (1995).

log10 BF10 BF10 Evidence against H0

0 to 1/2 1 to 3.2 Not worth more than a bare mention
1/2 to 1 3.2 to 10 Substantial
1 to 2 10 to 100 Strong
>2 > 100 Decisive

drugs to be tested, treatment assignment, and randomization probability).

Such static trials are designed based on a set of (strong) assumptions on the
characteristics of the drug and target population (e.g., safety, treatment effect
size, and sensitive patients), which unfortunately often turn out to be incor-
rect. This leads to a high failure rate, loss in efficiency, and high development
costs (Berry, 2003).
Adaptive designs provide an important approach to addressing this is-
sue. Per the US Food and Drug Administration (FDA) guidance on Adaptive
Designs for Clinical Trials of Drugs and Biologics (FDA, 2019), an adaptive
design is defined as a clinical trial design that allows for prospectively planned
modifications on one or more aspects of the design based on accumulating data
from subjects in the trial. Examples of the modifications include adaptive dose
escalation and de-escalation decisions, futility/efficacy/safety monitoring to
early terminate or graduate a treatment arm, adding new treatment arms,
adaptive randomization based on patients’ treatment outcomes or covariates,
enriching or reducing the enrollment of a subpopulation during the trial, adap-
tive treatment allocation, sample size recalculation, seamless phase transition,
making trial go/no-go decisions in multistage trials, and adaptive estimation
of the treatment effect or toxicity rate using interim data (Zang and Lee, 2014;
Pallmann et al., 2018). These in-trial adaptations make the design more ac-
curately reflect the characteristics of the drug effect on the target population,
thus potentially improving the safety, efficiency, and success rate of clinical
trials (Chow and Chang, 2008; Mahajan and Gupta, 2010; Pallmann et al.,
2018).
While adaptive designs can be developed using either frequentist or
Bayesian approaches, the Bayesian paradigm is particularly appealing due to
its “learn-as-we-go” nature. As described previously, by repeatedly applying
Bayes’ theorem, the Bayesian paradigm seamlessly integrates accumulating
interim trial data and updates the knowledge of the parameter of interest θ
to guide treatment decisions. Compared to frequentist approaches, Bayesian
methods have several advantages (Lee and Chu, 2012; Berry et al., 2010),
including:
Bayesian Statistics and Adaptive Designs 9

1. Bayesian designs treat the parameters of interest (e.g., the treat-

ment effect) as random variables, directly modeled by well-defined
probability distributions. Trial decisions and inference are made
based on the (posterior or predictive) probability distribution of
the parameters of interest, thus they properly incorporate various
levels of uncertainty contained in the data.
2. Because of the probability-based inferential structure, Bayesian
methods directly answer the scientific questions through the use of
posterior distributions, e.g., the probability that the response rate
of the drug is greater than a certain benchmark. This facilitates
communication with physicians and provides a consistent way to
tackle more complex problems.
3. Bayesian methods support continuous learning in a learn-as-we-go
fashion, which forms the basis of adaptive clinical trial designs. As
long as new data or any relevant information is available during the
trial, the design can be automatically integrated in the posterior
inference at predefined time points for making interim decisions.
This feature greatly facilitates flexible and efficient trial adaptation
and monitoring in adaptive clinical trials, without compromising
the integrity and validity of the trial.
4. As an essential feature of Bayesian approaches, prior or external
information can be naturally synthesized into the current trial. With
a properly specified prior distribution, Bayesian adaptive designs
tend to make more accurate and efficient decisions, which has the
potential to result in a trial with less cost, a reduced sample size,
and a shorter duration.
5. By applying outcome adaptive randomization, more patients can be
treated by the putatively more effective treatment(s) based on the
interim data. For life-threatening diseases, this is a desirable fea-
ture. Before a definitive decision is made, the probability of patient
allocation to arms can be computed to maximize the overall benefit,
e.g., the overall response rate, in the trial and/or the future patient
horizon in the population while controlling the false positive and
false negative error rates in decision making.
6. The Bayesian framework provides an ideal tool for hierarchical mod-
eling. Different levels of information can be naturally incorporated
under the hierarchical model assumption, which enables information
borrowing across different patient subgroups, treatment regimes,
and time periods. Such a feature is of great importance in designing
basket, umbrella, or platform trials in the modern era of precision
medicine.
7. The “gain/loss” utility functions can be formally incorporated by
Bayesian approaches, which facilitates informed decision making in
10 Model-Assisted Bayesian Designs for Dose Finding and Optimization

a complex setting of balancing risk and benefit. Optimal treatment

or trial decisions can be obtained by maximizing the utility function
(or minimizing the loss function) in the Bayesian decision-theoretic
framework. Furthermore, the posterior predictive distribution ren-
ders the evaluation of future patient outcomes based on the current
observed data, which offers additional flexibility in adaptively treat-
ing future patients to maximize risk–benefit tradeoffs.

1.3 Adoption of Bayesian adaptive designs

Bayesian adaptive designs are gaining popularity in clinical research, espe-
cially in oncology. As one of the world’s largest premier cancer centers, The
University of Texas MD Anderson Cancer Center (MDACC) has pioneered
the application of Bayesian adaptive designs in clinical trials. Since the early
2000s, Bayesian approaches have become an indispensable approach in design-
ing clinical trials at MDACC. Biswas et al. (2009) and Tidwell et al. (2019)
reviewed a total of 1984 trials (964 and 1020 trials, respectively) registered in
the MDACC electronic protocol database from January 2000 to April 2005,
and from January 2009 to December 2013. Table 1.2 presents the numbers and
percentages of Bayesian and non-Bayesian trials by site, respectively. Bayesian
trials were defined as trials which incorporated at least one Bayesian feature
in the design. Of the 964 trials registered during 2000 to early 2005, 394 (41%)
were multicenter trials and 570 (59%) were MDACC-only trials. Among the
MDACC-only trials, 30% (169 out of 570) implemented Bayesian designs be-
tween 2000 and 2005; the percentage of Bayesian trials significantly increased
to 56% (189 out of 335) between 2009 and 2013. Such an increasing trend in
use of Bayesian designs was also observed in multicenter trials: the percentage
of Bayesian trials conducted in 2009–2013 (14%) doubled that in 2000–2005.
Figure 1.2 depicts the number of protocols by year and development phase
for the two time periods. Panels (a) and (b) of Figure 1.2 demonstrate the
increasing use of Bayesian designs over time: the percentage of trials that ap-
plied Bayesian methods grew from 16% in 2000 to 30% in 2013. Panels (c)
and (d) of Figure 1.2 show that the Bayesian designs were more commonly
adopted in early-phase trials, such as phase I, phase II, and seamless phase I–II
trials. This is expected as early-phase trials are explorative in nature, starting
with substantial unknowns. Bayesian adaptive designs allow investigators to
efficiently synthesize accumulating data and make timely adaptions and mod-
ifications to treat patients and learn about the treatment more efficiently (Lin
and Lee, 2020). Furthermore, due to less stringent regulatory requirements
for early-phase trials as compared to late-phase trials, e.g., no need a rigorous
control of type I/II errors, implementation of Bayesian designs in early-phase
clinical trials is more acceptable. This partially explains the lower percentage
Bayesian Statistics and Adaptive Designs 11

TABLE 1.2: Numbers and percentages of Bayesian and non-Bayesian trials by

site.

Site Bayesian Non-Bayesian Total

Period: Janauary 2000 to April 2005
MDACC only 169 (30%) 401 (70%) 570
Multicenter 26 (7%) 368 (93%) 394
All sites 195 (20%) 769 (80%) 964

Period: January 2009 to December 2013

MDACC only 189 (56%) 146 (44%) 335
Other sites only 0 (0%) 9 (100%) 9
Multicenter 94 (14%) 582 (86%) 676
All sites 283 (28%) 737 (72%) 1020

of multicenter trials using Bayesian designs, as shown in Table 1.2, because

multicenter trials are usually larger, late-phase trials.
Despite the encouraging trend of using Bayesian adaptive designs, some
major barriers remain. Compared to conventional designs, Bayesian adaptive
designs are often difficult to understand, require complicated statistical mod-
eling, demand complex computation and simulations to explore the operating
characteristics, and need expensive infrastructure for implementation. As a
result, in most research institutions and pharmaceutical industry, the use of
Bayesian adaptive designs is still limited. To explore the promising features of
the Bayesian adaptive designs, FDA launched the Complex Innovative Trial
Design (CID) Pilot Program in 2018. Through this initiative, FDA encour-
aged the collaboration among industry, academia, and regulatory agency to
learn and advance novel clinical trial methodology.
In the last decade, great progress has been made to break these barriers.
One direction that holds great promise and potential is the development of
model-assisted designs (Yuan et al., 2019). Model-assisted designs yield supe-
rior performance compared to conventional algorithm-based designs, and their
decision rules can be pre-tabulated and included in the protocol and thus im-
plemented as simply as conventional designs. In the remaining chapters of
this book, we will introduce and review model-assisted designs to address
the quandary of simplicity versus performance that hinders the adoption of
Bayesian adaptive designs in practice. Software and trial examples will be
provided to illustrate the methodology. We consider that novel model-assisted
12 Model-Assisted Bayesian Designs for Dose Finding and Optimization

D 1XPEHURISURWRFROVE\\HDUIURPWR E 1XPEHURISURWRFROVE\\HDUIURPWR





1RQí%D\HVLDQ 1RQí%D\HVLDQ
%D\HVLDQ %D\HVLDQ



1XPEHURISURWRFROV

1XPEHURISURWRFROV

 
 





  




     


 



    í     

<HDU <HDU

F 1XPEHURISURWRFROVE\SKDVHIURPWR G 1XPEHURISURWRFROVE\SKDVHIURPWR


1RQí%D\HVLDQ  

  1RQí%D\HVLDQ
%D\HVLDQ
%D\HVLDQ



1XPEHURISURWRFROV

1XPEHURISURWRFROV
















 



  
  
     



, ,í,, ,, ,,í,,, ,,, ,9 2WKHU , ,í,, ,, ,,í,,, ,,, ,9 2WKHU

3KDVHRIGHYHORSPHQW 3KDVHRIGHYHORSPHQW

FIGURE 1.2: Number of trial protocols and percentage of Bayesian trials

(highlighted by shaded black areas) by year and phase of development. The
number above the bar represents the total number of trial protocols, and the
percentage shows the percentage of Bayesian trials.

designs possess features where superiority meets simplicity. They adhere to

the new KISS principle: Keep it simple and smart! The overarching goal of
this book is to increase the use of Bayesian adaptive designs, thereby speeding
up and improving the success of early phase clinical trials.
2
Algorithm-Based and Model-Based Dose
Finding Designs

2.1 Introduction
Conventionally, the objective of a phase I clinical trial is to identify the highest
dose of a new drug that is acceptably safe. This dose is called the maximum
tolerated dose (MTD), typically defined as the dose with the probability of
causing a dose-limiting toxicity (DLT) closest to a prespecified target rate,
for example 20% or 30%. The adverse events (AEs) that define a DLT are
prespecified by the investigators, and often scored using the Common Ter-
minology Criteria for Adverse Events (CTCAE) from the National Cancer
Institute (NCI). The CTCAE defines the severity of the AE using a 5-grade
scale based on the general guideline: grade 1 is mild; grade 2 is moderate;
grade 3 is severe or medically significant but not immediately life-threatening;
grade 4 is life-threatening; and grade 5 is death related to AE. Typically, the
DLT is often defined as an AE of grade 3 or higher.
An implicit assumption for finding the MTD is that efficacy and toxicity
of the investigational drug both increase with the dose, and thus the MTD
is presumed to be the most efficacious dose with an acceptable probability of
causing a DLT. This dose–toxicity–efficacy monotonicity assumption is rea-
sonable for most conventional cytotoxic agents, such as chemotherapies. For
many novel molecularly targeted or immunotherapy agents, efficacy may not
monotonically increase with the dose, although toxicity generally increases
with the dose. In these cases, a more appropriate objective for dose-finding
trials is to find the optimal biological dose (OBD), which is generally defined
as the dose that has the highest desirability in terms of the efficacy-toxicity
tradeoff. For example, assume that 20 mg of a new drug produces the efficacy
probability of 0.4 and toxicity probability of 0.3, if 10 mg produces the effi-
cacy probability of 0.39 and toxicity probability of 0.1, then 10 mg is more
desirable because it yields comparable (or higher) efficacy with lower toxicity.
Chapter 8 describes designs for finding the OBD.
Due to logistical reasons (e.g., preparation and manufacture of the drug),
the set of doses to be explored in a phase I trial is often prespecified by
investigators. The lowest dose typically is specified as one-tenth of the dose
that killed 10% of rodents during pre-clinical studies, i.e, one-tenth of the LD10

DOI: 10.1201/9780429052781-2 13
14 Model-Assisted Bayesian Designs for Dose Finding and Optimization

in rodents, after adjusting for differences in body surface area. The other doses
may be specified to follow a modified Fibonacci sequence so that successive
increments are, say 100%, 67%, 50%, 40%, and 33% thereafter. Alternatively,
successive increments may be a fixed percentage, e.g., 33%, or quantity, e.g.,
50 mg.
Two central statistical issues in dose finding are
1. Dose exploration (i.e., dose escalation/de-escalation), that is, how
to explore the set of doses during the trial?
2. Dose selection, i.e., how to determine the MTD upon completion of
the trial?
One unique challenge associated with these issues is that we should consider
not only statistical efficiency, but also patient ethics. On one hand, dose ex-
ploration should proceed quickly through doses that are well below the MTD,
since these doses presumably are subtherapeutic. On the other hand, dose ex-
ploration in the phase I trial should also proceed cautiously to avoid exposing
an excessive number of patients to doses that are above the MTD.
Dose selection is critical because the dose selected as the MTD upon com-
pletion of the phase I trial will likely be used in subsequent phase II and III
trials. If the dose selected as the MTD is well below the true MTD, then a
subsequent trial may likely fail, thereby wasting enormous resources and pos-
sibly overlooking a promising new drug. If the dose selected as the MTD is
well above the true MTD, then a subsequent trial will expose patients to an
unsafe dose and likely be terminated early due to an excessive rate of DLTs.
Phase I trial designs can be generally classified as algorithm-based designs,
model-based designs, and model-assisted designs (Yuan et al., 2019). This
taxonomy is based more on the characteristics of implementation of designs,
rather than statistical theory, to facilitate practitioners to understand and
apply the designs.
Algorithm-based designs use a set of simple, pre-specified rules (or al-
gorithms) to determine dose escalation and de-escalation, without assuming
any model on the dose–toxicity relationship. Examples include the 3+3 de-
sign (Storer, 1989), the accelerated titration design (Simon et al., 1997), the
rolling-six design (Skolnik et al., 2008), the A+B design (Lin and Shih, 2001),
the biased-coin design (Durham et al., 1997) and their variations (Ivanova
et al., 2003; Stylianou and Follmann, 2004). The implementation of the de-
signs does not require a computer program or much support from statisticians.
Despite widespread criticism of the 3+3 design for poor operating character-
istics, its simplicity continues to make it one of the most widely used phase I
trial designs in practice.
In contrast to the algorithm-based designs, model-based designs utilize
prespecified parametric dose–toxicity models (e.g., the power model or logis-
tic model) to guide dose escalation and de-escalation. As information accrues
during the trial, the dose–toxicity relationship is re-evaluated by updating the
Algorithm-Based and Model-Based Dose Finding Designs 15

estimates of the model parameters and then used to guide the dose alloca-
tion for subsequent patients. A typical example of the model-based design
is the continuous reassessment method (CRM) (O’Quigley et al., 1990). Al-
though a model-based design, such as the CRM, yields better performance
than an algorithm-based design (Le Tourneau et al., 2009; Jaki et al., 2013;
van Brummelen et al., 2016), it is considered by many to be statistically and
computationally complex due to the requirement of specifying the model and
prior, as well as repeated model fitting and estimation. This leads practition-
ers to perceive dose allocations as coming from a “black box.” As a result, the
use of the model-based designs has been fairly limited in practice (Rogatko
et al., 2007).
Emerging in the last decade, model-assisted designs combine the simplicity
of algorithm-based designs and the good performance of model-based designs.
This class of designs utilizes a probability model to derive the design, similar
to the model-based designs, but their rules of dose escalation and de-escalation
can be pre-tabulated before the onset of the trial in a fashion similar to the
algorithm-based designs. Examples of model-assisted designs include the mod-
ified toxicity probability interval (mTPI) design (Ji et al., 2010), Bayesian
optimal interval (BOIN) design (Liu and Yuan, 2015; Yuan et al., 2016a),
and keyboard design (Yan et al., 2017). Due to their competitive performance
and simplicity, model-assisted designs have been increasingly used in practice
(Yuan et al., 2019).
Statistically, these three classes of designs are more or less intertwined.
For example, algorithm-based designs involve, explicitly or implicitly, a certain
probability model (e.g., the binomial or Bernoulli model). Model-based designs
often also use some algorithms/rules to guide dose escalation (e.g., no skipping
of untried doses). Model-assisted designs, from a certain perspective, might
be regarded as a hybrid of model-based and algorithm-based designs as their
decision rules are model-based but can be enumerated and implemented in a
way similar to algorithm-based designs.
In this chapter, we briefly overview some algorithm-based designs and
model-based designs to lay down the foundation for the model-assisted designs,
the focus of this book and subsequent chapters. We assume that the DLT is
scored as a binary variable (i.e., DLT/no DLT) and is quickly ascertainable
such that when a new patient is enrolled and ready for dose assignment,
the DLT outcomes have been ascertained for all patients already enrolled. In
Chapters 5 and 7, we will discuss how to design phase I trials when the DLT
cannot be ascertained quickly (i.e., late-onset) and account for toxicity grades
scored in a scale of more than two levels.
Before describing the designs, we establish some notation. Let d1 < · · · <
dJ denote the J prespecified doses of the new drug that is under investigation
in the trial. We use π(dj ), or shorthand πj when no confusion is caused, to
denote the DLT probability that corresponds to dj , and φ to denote the target
DLT probability for the MTD. We use nj to denote the number of patients who
have been assigned to dj , and yj to denote the number of DLTs observed at
dj , j = 1, . . . , J. Therefore, at a particular point during the trial, the observed
16 Model-Assisted Bayesian Designs for Dose Finding and Optimization

data are D = {Dj , j = 1, . . . , J}, where Dj = (nj , yj ) are the “local” data
observed at dose level j.

2.2 Traditional 3+3 design

The traditional 3+3 design is the most widely used algorithm-based phase I
trial design in practice. This design is transparent and easy to implement, but
has poor operating characteristics (O’Quigley and Chevret, 1991; Le Tourneau
et al., 2009) and suffers from a number of drawbacks. For example, it cannot
target a specific toxicity rate, has poor accuracy in identifying the MTD, and
tends to treat a large percentage of patients at low doses that are potentially
subtherapeutic.
The so-called 3+3 design actually is a family of designs, including numer-
ous variations. Figure 2.1 shows the dose escalation and de-escalation rules
for a commonly used version of the 3+3 design. The design sequentially treats
patients in cohorts of three, and typically defines the MTD as the highest
dose, at which no more than one out six patients has the DLT. The toxicity
outcomes in the current cohort must be observed before any patients in the
next cohort enter the trial.
Under the 3+3 design depicted in Figure 2.1(a), the first cohort is treated
at a prespecified starting dose, which is usually the lowest dose under consid-
eration in the trial, and each subsequent cohort is treated as follows:
• If 0/3 patient in the current cohort has a DLT, the next cohort is treated
at the next higher dose.
• If 1/3 patient has a DLT, the next cohort is treated at the same dose.
• If ≥ 2/3 patients have a DLT, the current dose and all higher doses are
declared to be above the MTD, and the next cohort is treated at the next
lower dose. In the case that the current dose is the lowest dose, then the
trial is terminated and no dose is selected as the MTD.
After two cohorts have been treated at the same dose, i.e., six total patients,
the design proceeds as follows:
• If ≤ 1/6 patients had a DLT, the next cohort is treated at the next higher
dose, provided that this dose has not previously been declared to be above
the MTD.
• If ≥ 2/6 patients present with a DLT, then the next lower dose is selected as
the MTD, provided that 6 patients have already been treated in that dose.
There are different versions of the 3+3 design. A more aggressive variation of
the traditional 3+3 design is to select the MTD as the highest dose at which
two or fewer of six patients present with a DLT, see Figure 2.1(b).
Algorithm-Based and Model-Based Dose Finding Designs 17

Enroll 3
paƟents at
current dose

Escalate to DeͲescalate
the next 0 DLT 1 DLT 2 or 3 DLTs to the next
higher dose lower dose

NO
Enroll 3 more
paƟents at the
same dose
Have 6 paƟents
been treated at
previous dose?

0 or 1 DLT 2 or more DLTs

ŝŶϲƉĂƟĞŶƚƐ in 6 paƟents
YES

Stop
MTD =
previous dose

(a) A 3+3 design targeting the MTD with the DLT rate ≤ 1/6

Enroll 3
patients at
current dose

Escalate to DeͲescalate
the next 0 DLT 1 DLT 2 or 3 DLTs to the next
higher dose lower dose

NO
Enroll 3 more
patients at the
same dose
Have 6 patients
been treated at
previous dose?

0 or 1 DLT in 2 DLTs in 3 or more DLTs

6 patients 6 patients In 6 patients
YES

Stop Stop
MTD = current MTD =
dose previous dose

(b) A 3+3 design targeting the MTD with the DLT rate ≤ 2/6

FIGURE 2.1: Two different versions of the 3+3 design.

18 Model-Assisted Bayesian Designs for Dose Finding and Optimization

Many numerical studies show that the 3+3 design has poor operating
characteristics, e.g., poor accuracy to identify the MTD (Ahn, 1998; Iasonos
et al., 2008; Onar-Thomas and Xiong, 2010; Zhou et al., 2018b). This can
be explained by Table 2.1 applying the design shown in Figure 2.1(a), which
displays each of the possible outcomes and the corresponding posterior sum-
maries, including the 95% credible interval of the DLT probability and the
posterior probability that the DLT probability is greater than 25%. The pos-
terior summaries are derived by assuming independent beta-binomial models
for πj at each dj , i.e.,

yj | nj , πj ∼ Binomial(nj , πj )
πj ∼ Beta(0.5, 0.5),

resulting in the posterior

iid
πj | yj , nj ∼ Beta(yj + 0.5, nj − yj + 0.5).

The 3+3 design does not target a specific DLT probability, but a range of DLT
probabilities approximately ranging from 1/6 to 1/3, with the mean of 0.25.
Thus, Prob(πj ≤ 0.25 | yj , nj ) is used to summarize the posterior evidence for
whether dj is tolerable.
As shown in Table 2.1, the widths of the 95% credible intervals, which
are mostly > 0.5, indicate that three or six patients are far from adequate for
precisely estimating the DLT probability at a particular dose. For example, for
a dose with 1/6 DLT, the 95% credible interval for the true DLT probability
is (0.019, 0.558), which is too wide to make meaningful assessment of safety.
In addition, despite that doses corresponding to the outcome 2/6 still have

TABLE 2.1: What can be learned from the traditional 3+3 design?

No. DLTs/No. Patients Prob(π ≤ 0.25 | Data)† 95% Credible Interval†

Dose is considered to be below the MTD
0/3 0.830 (0.000, 0.536)
1/6 0.652 (0.019, 0.558)
Dose is selected as the MTD
0/6 0.942 (0.000, 0.330)
1/6∗ 0.652 (0.019, 0.558)
Dose is declared to be above the MTD
2/6 0.298 (0.077, 0.714)
3/6 0.085 (0.167, 0.833)
2/3 0.058 (0.177, 0.961)
4/6 0.014 (0.286, 0.923)
3/3 0.003 (0.464, 1.000)
*When the next higher dose has been used to treat patients. † Assuming that
the prior distribution of the DLT probability is Beta(0.5, 0.5)
Algorithm-Based and Model-Based Dose Finding Designs 19

approximately 30% chance to be at or below the MTD (i.e., πj ≤ 25%), the

traditional 3+3 design will declare any dose with this outcome to be above
the MTD.
These results highlight that the fundamental problem of the 3+3 design is
not its dose escalation and de-escalation rules, which actually are reasonable
when the target DLT probability is around 25%, but that it often stops too
early at a dose lower than the MTD and unreliably claims the MTD based on
merely six patients treated at that dose. In addition, its algorithm precludes
the possibility of calibrating the sample size to achieve higher reliability and
better operating characteristics. As these issues are not related to the number
of doses, the common belief that the 3+3 design works well for a trial with
a few (e.g., two or three) doses actually is incorrect. Due to the same reason,
the extension of the 3+3 design, such as the rolling six design (Skolnik et al.,
2008), also has poor accuracy in identifying the MTD (Onar-Thomas and
Xiong, 2010; Zhao et al., 2011). Furthermore, algorithm-based designs are
rigid and inflexible when the study conduct deviates from the design. For
example, a decision cannot be rendered for the 3+3 design if two or four
patients are treated at a certain dose. This rigidity poses tangible operational
problems as clinical trials often might not be carried out exactly as designed.

2.3 Cohort expansion

A common approach to address the sparse information at the MTD, selected
by the 3+3 design, is to perform cohort expansion at the MTD. That is, treat-
ing an additional certain number of patients (e.g., 6 or 10) at the estimated
MTD to confirm safety. This approach may seem sensible, since a larger sam-
ple size provides a more reliable inference, but actually suffers from a number
of logical, scientific, and ethical flaws.
The practice of adding an expansion cohort after selecting the MTD is
based on the fallacious assumption that the MTD selected by the 3+3 design
is reliable to be the true MTD. This ignores the basic statistical principle
that any estimate computed from a small sample size is associated with large
uncertainty. Suppose that a dose is selected as the MTD, at which 1/6 patients
had DLT, the 95% credible interval (see Table 2.1) shows there is a 95% chance
that the true toxicity probability of that dose is between 0.019 and 0.558. As
a result, a wide variety of questions may arise for the different numbers of
toxicities that may be observed in the expansion cohort. Since six is a very
small sub-sample, as more patients are treated at the estimated MTD in an
expansion cohort, the additional toxicity data easily may contradict the earlier
conclusion that the selected dose is the MTD. For example, what should one
do if the first three patients in an expansion cohort of size 10 all have toxicity?
20 Model-Assisted Bayesian Designs for Dose Finding and Optimization

This would give a total of 1/6 + 3/3 = 4/9 (44%) toxicities at the MTD. The
initial 3/3 toxicities in the expansion cohort suggest that the selected MTD
is not safe. Should one treat seven more patients at the estimated MTD, as
mandated by the protocol, or violate the protocol by abandoning the MTD
and de-escalating to a lower dose? If one de-escalates, what sort of rule or
algorithm should be applied to choose a dose, or doses, for the remaining seven
patients? If one continues to treat patients at the selected MTD, and ends up
with 7/10 toxicities, for a total of 1/6 + 7/10 = 8/16 (50%), what should one
conclude? The point is that the idea of treating a fixed expansion cohort at a
chosen MTD may seem sensible, but in practice can be problematic.
In recent years, the sizes of expansion cohorts following phase I trials have
exploded, from 6 or 10 to hundreds in some protocols (Bugano et al., 2017).
What nominally is a large “phase I expansion cohort” actually is a phase II
trial, but conducted without any design, other than a specification of sample
size. This practice magnifies all of the problems described above that occur
with a small expansion cohort. It fails to use the new data in the expansion
cohort adaptively to change the MTD if appropriate, and thus fails to protect
patient safety adequately. Furthermore, expansion cohorts are often used to
provide an initial exploration of treatment efficacy in specific subgroups. How-
ever, without a statistical design, there is no provision to discontinue trials
when experimental treatments are ineffective. As a general rule of thumb, if
the size of an expansion cohort reaches to the size of a phase II study, e.g., 30,
a statistical design is required. More discussion of scientific and ethical pitfalls
of cohort expansion can be found in Yan, Thall and Yuan (2017).

2.4 Accelerated titration design

Phase I trials often include many doses that are well below the MTD. This
is because the starting dose usually is chosen conservatively based on pre-
clinical studies. As a result, a large number of patients might be treated at
subtherepeutic doses that are far below the MTD. To address this issue, Simon
et al. (1997) proposed the accelerated titration design (ATD), which includes
an initial phase of accelerated dose exploration to quickly transition from these
presumably inefficacious doses to the more promising doses. During the initial
phase of accelerated titration, patients are treated in the cohort size of one
until the first DLT is observed or two moderate (grade 2) toxicities occur.
At that point in the trial, two additional patients are treated at that dose,
and then dose exploration proceeds as in the traditional 3+3 design. Compared
to the traditional 3+3 design, on average adding the accelerated titration
leads to fewer patients at the doses below the MTD, but slightly more pa-
tients at the doses above the MTD.
Algorithm-Based and Model-Based Dose Finding Designs 21

2.5 Continual reassessment method

Model-based designs are developed to find the MTD in a more efficient and
deliberate way. This class of designs assumes a parametric model between the
dose and DLT probability to facilitate the dose exploration in a statistically
justified manner. In contrast, the traditional 3+3 design and the ATD rely on
simple ad hoc rules for dose exploration that lack clear statistical justification.
The continual reassessment method (CRM; O’Quigley et al. (1990)) is the
first example of model-based dose finding designs. The CRM assumes a para-
metric model to describe the dose–toxicity relationship. As information ac-
crues during the trial, the dose–toxicity curve is re-evaluated by updating the
estimates of the model parameters via Bayesian machinery, and then used to
guide the dose allocation for subsequent patients. The commonly used model
for the CRM is the following power model,
exp{α}
πj = q j , for j = 1, . . . , J, (2.1)

where α is the unknown parameter, and q1 < · · · < qJ are prior estimates of
the DLT probabilities, called the skeleton, at each of the dose levels, respec-
tively. Other models, such as single-parameter logistic and hyperbolic tangent
models have also been proposed for the CRM, see, Cheung (2011, Section
3.2.2). Research shows that the choice of the model has little impact on the
performance of the design. What is more important is the configuration and
calibration of the model, e.g., the skeleton and priors.
Because the sample size of phase I trials is typically small, the model
used in the CRM is often simple and parsimonious, containing one or two
parameters (Iasonos et al., 2016). For illustration, under a target DLT rate
of 0.25, Figure 2.2 depicts the family of dose–toxicity curves that the power
model covers with the skeleton (q1 , · · · , q6 ) = (0.01, 0.04, 0.12, 0.25, 0.40, 0.54)
for a trial with six doses. This skeleton is computed using the model calibration
approach of Lee and Cheung (2009) with the half-width of the indifference
intervals being 0.07 and the prior guess of the MTD being the fourth dose.
Let D = {(nj , yj ), j = 1, . . . , J} denote the accrued data with yj DLTs
in nj patients at dose level j after n patients have been treated in the trial,
PJ
n = j=1 nj . The likelihood function under (2.1) is

J h iy h i(nj −yj )
exp{α} j exp{α}
Y
L(D | α) = qj 1 − qj .
j=1

Let f (α) denote the prior distribution for α, which often is taken as N (0, 2).
Applying Bayes’ theorem, the posterior distribution for α is

L(D | α)f (α)

f (α | D) = R .
L(D | α)f (α)dα
22 Model-Assisted Bayesian Designs for Dose Finding and Optimization

 ●
●
●
●
● ●
● ●
● ●
 ●
●

●
●
●
●
●
●
●
● ●
● ● ●
●
●
● ●
 ●

●
●
●
●
●
●
●
●
● ●
●
● ●
● ●
● ● ●

 ●
● ●

●
●
●
●
●
●

'/73UREDELOLW\
● ●
● ●
● ●
● ● ●
●
 ●
● ●
●
●
●
●
● ● ●
●
● ●
●
● ● ●
● ●
 ●
● ●
●

●
●
●
●
●
●
● ● ●
● ●
●
● ●

 ●
●

●
●
●

●
●
●
●
●

●
● ●
●
● ●
● ●
●
● ●

 ● ●

●
●

●
●
●

●
●

● ●
● ● ●
●
● ● ●
●
● ●
●
 ●
●
●
●
●
●
●
●
●
●

●
● ● ●
● ● ●
● ●
● ● ●
● ● ●
● ●

 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ●
● ● ●
●
● ● ● ●
● ● ● ● ●
● ● ● ● ● ●
●
● ● ● ●
● ●
● ● ● ● ●

 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●

     
'RVH

FIGURE 2.2: Prior support of the power model with the skeleton of (0.01,
0.04, 0.12, 0.25, 0.40, 0.54) and a normal N (0, 2) prior distribution assigned
to the unknown parameter α.

The posterior mean estimate of the actual DLT probability at dose level j is
Z
exp{α}
π
bj = qj f (α | D)dα, for j = 1, . . . , J. (2.2)

Because α is the only unknown parameter, {b πj } can be calculated using adap-

tive quadrature. Alternatively, these can be estimated using a Markov chain
Monte Carlo (MCMC) sampling algorithm, but adaptive quadrature is faster,
reasonably accurate, and does not require posterior convergence monitoring.
Dose exploration in the CRM is based on the continuously updated es-
timates of the true DLT probabilities defined in (2.2), and the prespecified
target toxicity rate φ. The original CRM proposed to treat each cohort at
the dose level corresponding to the estimated DLT probability closest to the
target φ, i.e., j ∗ = arg minj {|b
πj − φ|}. In practice, for patient safety, a “no-
dose-skipping” rule is typically imposed that treats the first cohort of patients
at the pre-specified starting dose and prohibits an untried dose to be skipped
when escalating the dose (Faries, 1994; Goodman et al., 1995; Piantadosi
et al., 1998). Zhou et al. (2018b) show that allowing dose skipping in the
CRM substantially increases the risk of overdosing patients with little gain on
the accuracy of finding the MTD. Thus, in general we recommend that the
no-dose-skipping rule should always be used.
The dose-finding algorithm of the CRM is described below.
Algorithm-Based and Model-Based Dose Finding Designs 23

1. Patients in the first cohort are treated at the lowest dose d1 or the
pre-specified dose.
2. Based on the cumulated data, we obtain the posterior DLT proba-
bility estimates π̂j , and find dose level j ∗ that has a DLT probability
closest to the target φ. Let j denote the current dose level,
• if j ∗ < j, de-escalate the dose level to j − 1;
• if j ∗ > j, escalate the dose level to j + 1;
• otherwise, stay at the same level j for the next cohort of pa-
tients.
3. The trial continues until the planned maximum sample size N is
exhausted, after which the MTD is selected as the dose with π
bj
closest to the target φ, i.e.,

MTD = arg min |b

πj − φ|. (2.3)
dj ∈(d1 ,...,dJ )

Although not included in the original CRM, an early stopping rule is often
imposed in practice to guard against the situation where the lowest dose is
excessively toxic. Specifically, if

Pr(π1 > φ | D) > pcut ,

then the trial is terminated, where pcut is a prespecified large probability

cutoff, e.g., pcut = 0.90. That is, the trial is stopped for safety when the
posterior probability that the lowest dose has a DLT probability above the
target exceeds a specified threshold, pcut . In addition to the early stopping
rule, Heyd and Carlin (1999) proposed to allow the trial to stop early also
when the width of the posterior 95% probability interval becomes sufficiently
narrow for the dose selected as the MTD. Numerous studies show that the
CRM has substantially better performance than the 3+3 design (Ahn, 1998;
Garrett-Mayer, 2006; Zhou et al., 2018b).
An important practical issue associated with the CRM is how to specify
the skeleton, q1 < · · · < qJ . If α = 0, then πj = qj , for j = 1, . . . , J. Therefore,
given the normal prior f (α) = N (0, 2), the skeleton represents the set of prior
(median) estimates for the DLT probabilities of J doses, which can be elicited
from clinicians. The skeleton determines the initial structure of the model and
the family of dose–toxicity curves that the model can cover. As the decisions
of dose exploration and selection are made based on the model estimates, dif-
ferent skeletons may lead to quite different design properties, especially under
the small sample size of the phase I trials. Yin and Yuan (2009b) showed
that the selection probability of the target dose can be 40% lower under one
skeleton than that under another skeleton in finite samples. The CRM may
24 Model-Assisted Bayesian Designs for Dose Finding and Optimization

perform poorly if it is based on a skeleton such that the true DLT probabil-
ities cannot be approximately recovered by the parsimonious model defined
in (2.1). Unfortunately, because the actual DLT probabilities are unknown,
practitioners often lack adequate information to determine whether a skeleton
is reasonable or not.
To simplify the specification of the skeleton, Lee and Cheung (2009) de-
veloped an automatic approach for selecting a skeleton when reliable prior
information is lacking. Their method is based on the “indifference interval,”
which is the range of DLT probabilities that the CRM cannot distinguish from
the target φ in large samples. The indifference-interval method determines a
skeleton based on φ (the target toxicity rate), j † (the prior guess for which
dose level is the MTD), J (the number of doses under consideration in the
trial), and δ (the desired half-width of the indifference interval). For the power
model CRM, the recommended skeleton is

qj = φ, j = j†,
 
log(φ − δ) log(qj+1 )
qj = exp , j = j † − 1, . . . , 1,
log(φ + δ) (2.4)
 
log(φ + δ) log(qj−1 )
qj = exp , j = j † + 1, . . . , J.
log(φ − δ)

Pan and Yuan (2017) showed that given a fixed half-width of the indifference
interval (i.e., δ), the indifference-interval method is invariant to the prior guess
for which dose level is the MTD (i.e., j † ). That is, different values of j † result
in equivalent skeletons, where equivalent skeletons are defined as skeletons
that lead to the same likelihood under the power model (2.1). As a result, the
half-width of the indifference interval plays a much more important role than
the prior guess of the MTD location.
Lee and Cheung’s method simplifies the specification of the skeleton, but
does not resolve the sensitivity issue of the CRM pertaining to model mis-
specification. Table 2.2 shows the simulation results of the CRM with two
different skeletons, skeleton 1 = (0.070, 0.127, 0.200, 0.286, 0.377, 0.468) and
skeleton 2 = (0.012, 0.069, 0.200, 0.380, 0.560, 0.706), respectively generated
by the method of Lee and Cheung (2009) with a half-width of the indifference
interval of 0.04 and 0.08. We can see that skeleton 1 substantially outperforms
skeleton 2 in scenario 1, whereas the result is opposite in scenario 2. In other
words, a skeleton that works well in one scenario may not work well in another
scenario, and there does not exist a single “best” skeleton that outperforms
all others in every scenario. In the next section, we will describe how to solve
this sensitivity issue by specifying multiple skeletons and then using Bayesian
model averaging or selection to adaptively identify the best fitted skeleton for
robust decision making.
Algorithm-Based and Model-Based Dose Finding Designs 25

TABLE 2.2: The performance of the CRM and Bayesian model averaging
CRM (BMA-CRM) with two different skeletons generated with half-widths of
the indifference interval of 0.04 (skeleton 1) and 0.08 (skeleton 2). The target
toxicity rate is φ = 0.2 and sample size is N = 36.

Scenario 1
Dose level 1 2 3 4 5 6
True DLT rate 0.03 0.04 0.05 0.06 0.07 0.20
% sel§ 0 0.10 1.25 3.85 21.05 73.20
CRM (skeleton 1)
No. pts† 1.5 1.7 2.2 3.5 8.0 18.8
% sel 0.05 1.35 5.70 8.10 28.25 56.40
CRM (skeleton 2)
No. pts 1.5 2.2 3.4 5.2 10.1 13.4
% sel 0 0.25 2.45 5.65 21.80 69.60
BMA-CRM
No. pts 1.5 1.9 2.7 3.9 8.6 17.4

Scenario 2
Dose level 1 2 3 4 5 6
True DLT rate 0.12 0.24 0.33 0.60 0.70 0.80
% sel 35.15 43.50 10.45 0.25 0 0
CRM (skeleton 1)
No. pts 13.9 12.2 5.2 1.2 0.3 0.1
% sel 30.60 53.10 11.15 0 0 0
CRM (skeleton 2)
No. pts 12.3 15.1 6.1 1.0 0.1 0
% sel 32.30 49.85 10.05 0 0 0
BMA-CRM
No. pts 12.7 14.3 5.5 1.0 0.3 0.1
§
: Average selection percentage at each dose;
†
: Average number of patients treated at each dose.

2.6 Bayesian model averaging CRM

A major practical issue associated with the CRM is the requirement of pre-
specifying the skeleton. Because of the lack of prior toxicity information on a
new drug, physicians may have quite different opinions about which skeleton
is reasonable for their particular context. If the true DLT probabilities are not
near the prior support of the skeleton, then the CRM may have a high prob-
ability of selecting the wrong dose as the MTD. To improve robustness, Yin
and Yuan (2009b) proposed prespecifying multiple skeletons, each represent-
ing a set of prior estimates of the toxicity probabilities, and using a Bayesian
model averaging (BMA) or Bayesian model selection approach to estimate
the true DLT probabilities of the doses under consideration in the trial. Yin
26 Model-Assisted Bayesian Designs for Dose Finding and Optimization

and Yuan (2009b) showed that the Bayesian model selection approach yields
similar performance as the BMA, thus we herein focus on the BMA approach.
Through the choice of multiple skeletons, BMA provides a more robust way
to construct the dose–oxicity curve compared to the standard CRM.
The rationale behind the BMA-CRM is to use multiple skeletons to repre-
sent different dose–toxicity relationships. Each skeleton corresponds to a CRM
model described previously with a different set of q1 < · · · < qJ . As long as
one of them is close to the truth, the design will perform well, since BMA
automatically identifies and favors the best fitted model.
Specifically, let {Mk }K
k=1 denote the CRM models corresponding to the K
prespecified skeletons {q1k < · · · < qJk : k = 1, . . . , K}. Like the CRM, the
kth model in the BMA-CRM connects the actual DLT probabilities to the kth
skeleton by assuming,
exp{αk }
πjk = qjk , for j = 1, . . . , J, k = 1, . . . , K.

Let Pr(Mk ) be the prior probability that model Mk is the true model, i.e.,
the probability that the kth skeleton (q1k < · · · < qJk ) matches the true
dose–toxicity curve. If there is no preference a priori for any one skeleton
over the other skeletons, equal weights can be assigned to the different models
by setting Pr(Mk ) = 1/K, k = 1, . . . , K. When there is prior information
about the importance of each set of the prespecified toxicity probabilities, such
information can be incorporated into Pr(Mk ), k = 1, . . . , K. For example, if
one skeleton is more likely to be true, it can be assigned a higher prior model
probability. After n patients have PJ been treated and the observed data are D =
{(nj , yj ), j = 1, . . . , J}, n = j=1 nj , the likelihood function corresponding
to the kth model is
J h iyj h i(nj −yj )
exp{αk } exp{αk }
Y
L(D | αk , Mk ) = qjk 1 − qjk .
j=1

The posterior model probability for Mk is

m(D | Mk )Pr(Mk )
Pr(Mk | D) = PK ,
`=1 m(D | M` )Pr(M` )

where Z
m(D | Mk ) = L(D | αk , Mk )f (αk )dαk

is the marginal likelihood of model Mk , αk is the unknown parameter associ-

ated with the CRM power model Mk , and f (αk ) is the prior distribution of
αk under model Mk , for k = 1, . . . , K.
During the trial, conditional on the observed data, the models for each
skeleton usually yield different estimates of the toxicity probabilities. Some
of these estimates may be close to the true values, whereas others may not,
depending on how well the models fit the accumulated data. To obtain the
Algorithm-Based and Model-Based Dose Finding Designs 27

estimate of the toxicity probabilities for doses under consideration in the trial,
the BMA approach takes a weighted average across the CRM models corre-
sponding to the different skeletons, where the weight for each model reflects
how well that model fits the accumulated data relative to the other models.
The potential estimation bias caused by a misspecification of the skeleton is
averaged out, leading to a more robust design compared to the original CRM.
More precisely, the BMA estimate for the toxicity probability at each dose
level is given by
K
X
π̄j = π̂jk Pr(Mk | D), j = 1, . . . , J, (2.5)
k=1

where π
bjk is the posterior mean of the toxicity probability of dose level j under
model Mk , i.e.,
L(D | αk )f (αk )
Z
exp{α }
π̂jk = qjk k R dαk .
L(D | αk )f (αk )dαk

By assigning π̂jk a weight of Pr(Mk | D), the BMA method automatically

identifies and favors the best fitted model, and thus π̄j is always close to the
best estimate. Dose allocation and selection in the BMA-CRM is based upon
the BMA estimates {π̄j }, and follows similarly to the original CRM.
For practical use, Yin and Yuan (2009b) recommend using three skeletons
in the trial design, and these skeletons should be chosen to reflect different
possible shapes of the dose–toxicity curve with different prior locations for the
MTD. For example, the first skeleton may increase steadily with the middle
dose as the prior MTD, the second skeleton may increase quickly then slow
down with a low dose as the prior MTD, and the third skeleton may increase
slowly then speed up with a high dose as the prior MTD. The specification
of the skeletons should be collaborated with physicians to incorporate their
prior knowledge on the possible shapes of the dose–toxicity relationship.
When there is a lack of prior information, in our experience, for J > 3, the
three skeletons that correspond to (j † , δ) = {(2, 0.10), (ceiling((J + 1)/2),
0.07), (J − 1, 0.04)} using the model calibration method of Lee and Cheung
(2009) (see Section 2.5) provide a reasonable skeleton set for the BMA-CRM.
When there are J = 3 doses, the three skeletons can be generated using Lee
and Cheung’s method with (j † , δ) = {(1, 0.10), (2, 0.07), (3, 0.04)}; and when
there are only J = 2 doses, the standard CRM with a single skeleton is often
sufficient. These skeletons have prior support for a wide range of dose–toxicity
curves that are plausible in many contexts. Figure 2.3 depicts the prior support
of these three skeletons for a trial with six doses and N (0, 2) prior distribu-
tions on αk , k = 1, . . . , K. In this case, dose levels two, four, and five are the
prior expected MTDs for the first, second, and third skeletons, respectively,
but it is plausible a priori with each of the three skeletons that each dose
under consideration is actually the MTD. Alternatively, Pan and Yuan (2017)
proposed to choose skeletons by maximizing the model space covered by them,
28 Model-Assisted Bayesian Designs for Dose Finding and Optimization

 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●
●
●
●
●
●
 ● ● ●
'/73UREDELOLW\  ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●
● ● ●
●
●
 ●
● ●
●
●

'/73UREDELOLW\

'/73UREDELOLW\
● ● ●
● ●
● ● ● ● ● ●
● ● ●
● ●
● ●
● ● ● ● ● ● ●
● ● ● ●
● ●
● ● ● ● ● ● ● ●
● ● ● ● ● ● ●

 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●

●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●

 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●

●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●

 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●  ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●

 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●  ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●

 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●

●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●  ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●

 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●

 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●
●
●
●
●
●
●
●
●
● ●
● ●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
 ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●

                 
'RVH 'RVH 'RVH

FIGURE 2.3: Prior support of the proposed default set of three skeletons for
the BMA-CRM when the target DLT rate is 0.25.

and then further calibrate the skeletons to maximize the probability of correct
selection of the MTD using a set of prespecified scenarios. That calibration
method is computationally intensive, but it results in a set of three skeletons
that performs well in the specified scenarios.
Table 2.2 provides a numerical example that the BMA-CRM is more reli-
able than the CRM: the performance of the BMA-CRM is close to that of the
CRM with the better skeleton in both scenarios 1 and 2. Such robustness and
reliability are tremendously important because in practice we often prefer a
method that yields reliable performance to a method that has high variability
(i.e., performs well in one scenario but not in another scenario).

2.7 Escalation with overdose control

Babb et al. (1998) proposed another model-based design called escalation with
overdose control (EWOC). EWOC assumes a two-parameter logistic model for
the dose–toxicity curve:

logit{π(dj )} = β0 + β1 dj , j = 1, · · · , J, (2.6)

where β0 and β1 are unknown intercept and slope parameters, respectively, and
logit(x) = log{x/(1−x)}. To facilitate interpretation, EWOC re-parameterizes
the logistic model by replacing (β0 , β1 ) with (π1 , ρ), where π1 = π(d1 ) is the
DLT probability at the lowest dose under consideration in the trial, and ρ is
the MTD with logit−1 (β0 + β1 ρ) = φ. The re-parameterized logistic model is
given by
(dj − d1 )logit(φ) + (ρ − dj )logit(π1 )
logit{π(dj )} = .
ρ − d1
If dj = ρ, then logit{π(ρ)} = logit(φ), and thus ρ is the unknown MTD.
Algorithm-Based and Model-Based Dose Finding Designs 29

The prior distributions for π1 and ρ are usually taken to be independent

Uniform(0, φ) and Uniform(d1 , dJ ) distributions, respectively. That is, the low-
est dose under consideration is assumed to be below the MTD, and the MTD
is assumed to be in the investigational dose range [d1 , dJ ]. An MCMC sam-
pling algorithm can be used for posterior inference. Because of the former
model assumption, if a DLT is observed at the lowest dose, then Babb et al.
(1998) recommended stopping the trial.
The main difference between the CRM and EWOC is that EWOC explic-
itly controls the predicted proportion of patients treated with doses that are
above the MTD in its dose exploration rule. Specifically, after treating the
first cohort at the lowest reasonable dose, EWOC recommends that each sub-
sequent cohort is treated with dose d that has a predicted probability of being
above the MTD equal to a prespecified value α,

Pr(π(d) ≥ φ | D) = α.

As Pr(π(d) ≥ φ | D) = Pr(ρ ≤ d | D), the predicted proportion of patients

treated at a dose above the MTD is equal to α. Larger values of α correspond
to more aggressive dose escalation. Babb et al. (1998) recommended α = 0.25.
For patient safety, as with the CRM, the no-dose-skipping rule is often imposed
in practice. The trial continues until the prespecified maximum sample size
N is exhausted. Upon completion of the trial, EWOC selects the MTD as the
dose that minimizes the posterior expected risk
Z
L(d; ρ, α)f (ρ|D)dρ,

with respect to the asymmetric loss function

(
α(ρ − d), if d ≤ ρ,
L(d; ρ, α) =
(1 − α)(d − ρ), if d > ρ,

where f (ρ|D) is the marginal posterior function of the MTD. If α < 0.5,
then L(ρ − δ; ρ, α) < L(ρ + δ; ρ, α), and thus underdosing is preferable to
overdosing. More precisely, since L(ρ + δ; ρ, α)/L(ρ − δ; ρ, α) = (1 − α)/α,
this loss function says that the loss incurred by treating a patient with a
dose δ units above the MTD is (1 − α)/α times more than treating a patient
with a dose δ units below the MTD. Due to the use of the overdose control
rule, EWOC is more conservative and safer than the CRM, but at the cost
of sacrificing the accuracy of identifying the MTD, which is often substantial.
In addition, EWOC is subject to the influence of model misspecification in a
similar way as the CRM.
30 Model-Assisted Bayesian Designs for Dose Finding and Optimization

2.8 Bayesian logistic regression method

The Bayesian logistic regression method (BLRM) is another modification of
the CRM. The BLRM uses a similar two-parameter logistic regression model
as the EWOC, given by

logit{π(dj )} = β0 + β1 log (dj /d∗ ) , j = 1, · · · , J, (2.7)

where β0 and β1 are unknown intercept and slope parameters, dj is the raw
dosage at dose level j, and d∗ is the reference dose. Neuenschwander et al.
(2008) recommended the use of the vague bivariate normal distribution as the
prior distribution of (β0 , log(β1 )), e.g.,
   2 
µ1 σ1 ρσ1 σ2
(β0 , log(β1 )) ∼ N , .
µ2 ρσ1 σ2 σ22

For example, (µ1 , µ2 , σ1 , σ2 , ρ) = (−0.847, 0.381, 2.015, 1.207, 0).

There are two main differences between the BLRM and the CRM. First,
the BLRM requires specifying the proper dosing interval (δ1 , δ2 ), defined as
the range of the DLT probabilities regarded as acceptable, where δ1 < φ < δ2 .
For example, given the target DLT probability φ = 0.27, the proper dosing
interval (δ1 , δ2 ) may be chosen as (0.20, 0.35). The choice of (δ1 , δ2 ) depends
on characteristics of the trial (e.g., the target population, toxicity profile of
the experimental drug, and risk-benefit consideration), and could differ across
trials. Second, the BLRM imposes an overdose control rule, similar to that
of the EWOC, as follows: if the observed data suggest that there is 25%
posterior probability that the DLT rate of a dose is greater than δ2 , i.e.,
Pr(πj > δ2 |D) ≥ 0.25, then that dose is an overdose and cannot be used to
treat patients.
The BLRM is highly similar to the EWOC. Both methods use the two-
parameter logistic model with slightly different re-parameterization. The main
difference is that the BLRM specifies the target as an interval (δ1 , δ2 ), while
the EWOC uses the point target φ as the CRM. However, as the investigational
dose is discrete, this difference has little impact on operating characteristics,
in particular when setting φ = (δ1 + δ2 )/2. In other words, the two designs
will generate similar operating characteristics when the prior is properly cali-
brated.
To find the MTD, the BLRM proceeds in a similar way as the CRM. The
BLRM starts the trial by treating the first cohort of patients at the lowest
dose d1 or a prespecified dose. After each patient cohort is treated, the BLRM
updates the estimate of the dose–toxicity curve based on the accumulating
DLT data across all dose levels, and assigns the next cohort of patients to
the “optimal” dose. The “optimal” dose is defined as the dose j that sat-
isfies the overdose control condition Pr(πj > δ2 |D) < 0.25 and meanwhile
maximizes the posterior probability of the proper dosing interval (δ1 , δ2 ), i.e.,
Algorithm-Based and Model-Based Dose Finding Designs 31

Pr(πj ∈ (δ1 , δ2 )|D). For patient safety, as with the CRM, the no-dose-skipping
rule is often imposed in practice. Thus, if the estimated optimal dose is higher
than the current dose, we escalate the dose by one level; and if the estimated
optimal dose is lower than the current dose, we de-escalate the dose by one
level. The overdose control rule leads to the following safety stopping rule:
stop the trial if Pr(π1 > δ2 |D) ≥ 0.25 (i.e., the lowest dose is an overdose).

The trial continues until the prespecified maximum sample size N is ex-
hausted. Upon completion of the trial, the BLRM selects the final estimate
of the “optimal” dose as the MTD. Alternative early stopping rules can be
added to the BLRM, for example, stop the trial if the “optimal” has a large
probability (say ≥ 50%) of the proper dosing interval or a minimum number
of patients have been treated at the “optimal” dose.
Of note, given dj and the logistic model (2.7), the prior distribution of
β0 , log(β1 )) automatically determines a set of the prior estimates of π(dj ),
i.e., the skeleton. Thus, the BLRM suffers from the similar sensitivity issue to
the skeleton, or more precisely the prior specification. In addition, as noted
and elaborated by Cheung (2011) and Iasonos et al. (2016), somewhat counter-
intuitively, using the more flexible two-parameter logistic model (2.7) actually
is inferior to and leads to worse performance than the single-parameter power
model (2.1). This is also verified by extensive simulation study by Zhou et al.
(2018b), which shows that the CRM (Section 2.5) outperforms the BLRM.
Lastly, the BLRM depends on the (standardized) raw dosage dj , thus may be
decision-inconsistent. Consider two drugs with the same number of dose levels,
but different raw dosages, e.g., one drug is (5 mg, 10 mg, 15 mg, 20 mg), and
the other drug is (30 mg, 60 mg, 120 mg, 240 mg). We call a design decision-
consistent if it generates identical operating characteristics as long as the DLT
probabilities of the dose levels are the same, regardless of the raw dosages. For
example, the two drugs have the same DLT probabilities (0.05, 0.1, 0.25, 0.45)
at the four dose levels. Clearly, decision-consistency is a desirable property to
have because when the underlying data generation mechanisms are the same,
the design should yield the same operating characteristics. The BLRM, how-
ever, does not have this property because the same data may result in different
estimates when two trials have different raw dosages dj . In contrast, the CRM
and BMA-CRM (based on the power model) do not depend on raw dosages,
but only dose levels, and thus are decision-consistent.

2.9 Software
Software is not required to conduct the algorithm-based 3+3 design and the
accelerated titration design, since their dose exploration and selection rules
32 Model-Assisted Bayesian Designs for Dose Finding and Optimization

are simple and prespecified. In contrast, the model-based designs, such as

CRM/BMA-CRM, EWOC, and BLRM, rely on the real-time estimate of
dose–toxicity model to make the decision of dose escalation and de-escalation,
therefore software is necessary for trial conduct and to obtain simulation-based
operating characteristics.
Software for the CRM/BMA-CRM is available in several forms: (1) Win-
dows desktop program “CRM Suite” with intuitive graphical user interface,
freely available at MD Anderson Cancer Center Software Download Website
https://biostatistics.mdanderson.org/softwaredownload/; (2) Web
application with intuitive graphical user interface, freely available at
https://www.trialdesign.org; and (3) R packages freely available at CRAN,
such as bcrm and dfcrm.
The EWOC design can be implemented in R using the bcrm() function
from the bcrm package. Interactive software for implementing EWOC is also
freely available at https://biostatistics.csmc.edu/ewoc/index.php. The
BLRM design is available in commercial software such as East by Cytel. R
packages such as blrm for implementing BLRM can be also found at CRAN.
3
Model-Assisted Dose Finding Designs

3.1 Introduction
Model-assisted designs have emerged as an attractive approach for phase I
clinical trials that combine the simplicity of algorithm-based designs with
the superior performance of model-based designs. Model-assisted designs refer
to a class of novel designs that use a model (e.g., the binomial model) for
efficient decision making like model-based designs, while their dose escalation
and de-escalation rules can be tabulated before the onset of a trial as with
algorithm-based designs (Yuan et al., 2019). This chapter introduces several
model-assisted phase I designs that aim to find the maximum tolerated dose
(MTD), including the modified toxicity probability interval (mTPI) design (Ji
et al., 2010), keyboard design (Yan et al., 2017), and Bayesian optimal interval
(BOIN) design (Liu and Yuan, 2015; Yuan et al., 2016a).

3.2 Modified toxicity probability interval design

The mTPI design (Ji et al., 2010) starts by defining three dosing intervals:
the underdosing interval (0, δ1 ), proper dosing interval (δ1 , δ2 ), and overdosing
interval (δ2 , 1), where δ1 < φ < δ2 . For example, given the target probability
of the dose-limiting toxicity (DLT) φ = 0.2, the three intervals may be defined
as (0, 0.15), (0.15, 0.25), and (0.25, 1), respectively.
Suppose that at the current dose level j, yj of nj patients have experienced
DLT. mTPI assumes that yj follows a beta-binomial model

yj | nj , πj ∼ Binomial(nj , πj ) (3.1)
πj ∼ Beta(1, 1) ≡ Unif(0, 1).

Then, the posterior distribution of πj based on Dj = (nj , yj ) is given by

πj | Dj ∼ Beta(yj + 1, nj − yj + 1), for j = 1, . . . , J. (3.2)

DOI: 10.1201/9780429052781-3 33
34 Model-Assisted Bayesian Designs for Dose Finding and Optimization

Unlike continual reassessment method (CRM), which models toxicity

across doses using the power model (2.1) or the logistic model, mTPI mod-
els toxicity at each dose independently. This is a common feature of model-
assisted designs, rendering them computational simplicity and the ability to
tabulate all dose escalation/de-escalation rules prior to the trial conduct.
To make the decision of dose escalation/de-escalation, mTPI calculates the
unit probability mass (UPM) for each of the three intervals, defined as:

UPM1 = Pr(πj ∈ (0, δ1 ) | Dj ) /δ1 ,

UPM2 = Pr(πj ∈ (δ1 , δ2 ) | Dj ) /(δ2 − δ1 ) , and
UPM3 = Pr(πj ∈ (δ2 , 1) | Dj ) /(1 − δ2 ) .

Suppose j is the current dose level. mTPI determines the next dose as follows:
• If UPM1 = max{UPM1, UPM2, UPM3}, then escalate the dose to level j+1.
• If UPM2 = max{UPM1, UPM2, UPM3}, then stay at the current dose level
j.
• If UPM3 = max{UPM1, UPM2, UPM3}, then de-escalate the dose to level
j − 1.
Because the three UPMs can be determined for all possible outcomes Dj =
(nj , yj ), the dose escalation and de-escalation rules can be tabulated before
the trial begins, which makes mTPI easy to implement in practice.
This can be done as follows: enumerate all possible values of nj from
1 up to the maximum sample size N , and given each possible value of nj ,
enumerate all possible values of yj from 0 up to nj . Then, given a pair of
(nj , yj ), calculate three UPMs and record the resulting dose escalation/de-
escalation decision. Table 3.1 shows the dose escalation and de-escalation table
for mTPI based on a target DLT rate φ = 0.20 and the proper dosing interval
(δ1 , δ2 ) = (0.17, 0.23).

TABLE 3.1: Decision table of the mTPI design based on a target toxicity
rate φ = 0.20 and the proper dosing interval (δ1 , δ2 ) = (0.17, 0.23), nj is the
number of patients treated at dose level j, and yj is the number of DLTs
observed at dose level j.

Number of patients nj 1 2 3 4 5 6 7 8 9 10 11 12
Escalate to j + 1 if yj ≤ 0 0 0 0 0 0 0 0 1 1 1 1
De-escalate to j − 1 if yj ≥ 1 1 2 2 3 3 3 4 4 4 5 5
Eliminate levels j to J if yj ≥ NA 2 2 3 3 3 4 4 4 5 5 5
Model-Assisted Dose Finding Designs 35

The trial continues until the prespecified maximum sample size N is

reached. At that point, the MTD is selected as the dose whose isotonic
estimate of the DLT rate is closest to the target φ,

MTD = arg min |π̃j − φ|,

dj ∈(d1 ,...,dJ )

where π̃j is the estimate of πj based on isotonic regression, which can be

obtained using the pooled adjacent violators algorithm (Barlow et al., 1972)
on {π̂j , j = 1, . . . , J} with π̂j = yj /nj . Isotonic regression is the technique
of fitting a nonparametric or free-form line to observations (e.g., (yj , nj ))
such that the fitted line is non-decreasing everywhere and lies as close to the
observations as possible, see Barlow et al. (1972) for technical details.
For safety, mTPI includes a dose exclusion/safety stopping rule: if Pr(πj >
φ | Dj ) > 0.95, dose level j and higher are excluded from the trial. If the lowest
dose is excluded, the trial is stopped for safety. Such a safety monitoring rule
also can be reflected in the decision table (e.g., see the last row of Table 3.1).
One deficiency of mTPI is that its decision rule (i.e., use UPM to guide dose
exploration) lacks clear interpretation and leads to a high risk of overdosing
patients. To see the problem, consider a trial with a target DLT probability
φ = 0.20, and underdosing, proper dosing, and overdosing intervals of (0, 0.17),
(0.17, 0.23), and (0.23, 1), respectively. Suppose at a certain stage of the trial,
the observed data indicate that the posterior probabilities of the underdosing
interval, proper dosing interval, and overdosing interval are 0.01, 0.09, and 0.9,
respectively. That is, the observed data indicate that there is a 90% chance
that the current dose is overdosing patients and only a 9% chance that the
current dose is properly dosing patients. Despite such dominant evidence of
overdosing, mTPI does not de-escalate the dose, staying at the same dose
for treating the next patient cohort, as UPM1 = 0.01/0.17 = 0.059, UPM2
= 0.09/(0.23 − 0.17) = 1.5, and UPM3 = 0.9/(1 − 0.23) = 1.17.
The pathological behavior of mTPI is caused by UPM non-proportionally
weighting the evidence of underdosing, proper dosing, and overdosing due to
different lengths of the three intervals. As the overdosing interval is typically
the longest, UPM has bias in suppressing the evidence of overdosing. The
numerical studies in Section 3.5 confirm this issue.

3.3 Keyboard design

Yan et al. (2017) proposed the keyboard design as a seamless upgrade of the
mTPI design to address the latter’s overdosing issue. Unlike the mTPI design,
which divides the toxicity probabilities into three intervals (i.e., underdosing,
proper dosing and overdosing), the keyboard design defines a series of equal-
width dosing intervals (referred to as “keys”) that correspond to all potential
locations of the true toxicity rate of a particular dose. The design then uses the
key with the highest posterior probability (referred to as the “strongest key”)
36 Model-Assisted Bayesian Designs for Dose Finding and Optimization

(a) mTPI (b) Keyboard

2.0

2.0
Target key Strongest key
1.5

1.5
Density

Density
1.0

1.0
UPM2

UPM3
UPM1
0.5

0.5
0.0

0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

DLT rate DLT rate

FIGURE 3.1: Contrast between (a) the mTPI design and (b) the keyboard
design. The curves are the posterior distributions of πj . To determine the next
dose, the mTPI design compares the values of the three UPMs, whereas the
keyboard design compares the location of the strongest key with respect to
the target key.

to guide dose escalation and de-escalation. Figure 3.1 contrasts the keyboard
and mTPI designs.
The keyboard design starts by specifying a proper dosing interval I ∗ =
(δ1 , δ2 ), referred to as the “target key,” and then populates this interval toward
both sides of the target key, forming a series of keys of an equal width. The
keys span the range of 0 to 1. For example, given the proper dosing interval
or target key of (0.25, 0.35), on its left side we form two keys of width 0.1
(i.e., (0.15, 0.25) and (0.05, 0.15)); and on its right side we form six keys of
width 0.1 (i.e., (0.35, 0.45), (0.45, 0.55), (0.55, 0.65), (0.65, 0.75), (0.75, 0.85),
and (0.85, 0.95)). We denote the resulting intervals/keys as I1 , · · · , IK . As all
keys must have an equal width and be within [0, 1], some DLT probability
values at the two ends (e.g., < 0.05 or > 0.95 in the example) may not be
covered by keys. As explained in Yan et al. (2017), ignoring these “residual”
DLT rates at the two ends does not pose any issue for the decision making of
dose escalation and de-escalation. This is because the posterior distribution of
πj is unimodal, and the decision is made by the relative position of the target
key to the strongest key.
To make the decision of dose escalation and de-escalation, given the ob-
served data Dj = (nj , yj ) at the current dose level j, the keyboard design
identifies the interval Imax that has the largest posterior probability,

Imax = arg max{Pr(πj ∈ Ik | Dj )},

I1 ,··· ,IK
Model-Assisted Dose Finding Designs 37

which can be easily evaluated based on πj ’s posterior distribution given by

equation (3.2), assuming that πj follows a beta-binomial model (3.1). Imax
represents the interval that the true value of πj is most likely located, referred
to as the “strongest key.” Graphically, the strongest key is the one with the
largest area under the posterior distribution curve of πj , see panel (b) of
Figure 3.1. If the strongest key is on the left (or right) side of the target
key, then it means the observed data suggest that the current dose is most
likely to represent underdosing (or overdosing), and thus dose escalation (or
de-escalation) is needed. If the strongest key is the target key, the observed
data support that the current dose is most likely to be in the proper dosing
interval, and thus it is desirable to stay at the current dose for treating the
next patient. In contrast, UPM used by the mTPI design does not have such
an intuitive interpretation and tends to distort the evidence of overdosing, as
described previously.
Based on the keyboard design, patients in the first cohort are treated at
the lowest dose level, or the physician-specified level. Suppose j is the current
dose level. The keyboard design determines the next dose as follows:
• If the strongest key Imax is on the left side of the target key I ∗ , then escalate
the dose to level j + 1.
• If Imax is I ∗ , then stay at the current dose level j.
• If Imax is on the right side of I ∗ , then de-escalate the dose to level j − 1.
The trial continues until the maximum sample size N is reached. At that point,
the MTD is selected based on isotonic estimates π̃j as described previously.
During the trial conduct, the keyboard design imposes the dose exclu-
sion/early stopping rule such that if Pr(πj > φ | Dj ) > 0.95 and nj ≥ 3, dose
level j and higher are eliminated from the trial, and the trial is terminated if
the lowest dose is eliminated, where Pr(πj > φ |Dj ) is evaluated based on the
posterior distribution (3.2).
Similar to the mTPI design, the dose escalation and de-escalation rules of
the keyboard design can be tabulated for each possible nj = 1, · · · , N before
the trial begins, making it easy to implement in practice. For example, Table
3.2 shows the dose escalation and de-escalation boundaries of the keyboard
design based on a target toxicity rate φ = 0.20 and the proper dosing interval
(δ1 , δ2 ) = (0.17, 0.23). By contrasting Tables 3.1 and 3.2, we find that given
the same proper dosing interval the keyboard design is less aggressive than the
mTPI design. In particular, when three out of eight patients have toxicities at
a given dose level, the keyboard design recommends dose-escalation, whereas
the mTPI design suggests retaining the dose even though this dose yields an
observed DLT rate of 3/8 = 37.5% with Pr(πj > 0.20 | nj = 8, yj = 3) = 0.91.
As the location of the strongest key approximately indicates the mode
of the posterior distribution of πj , the keyboard design can be viewed as a
posterior mode-based dose-finding method. Yan et al. (2017) shows that the
keyboard design substantially outperforms mTPI. Of note, the variation of
38 Model-Assisted Bayesian Designs for Dose Finding and Optimization

TABLE 3.2: Decision table of the keyboard design based on a target toxicity
rate φ = 0.20 and the proper dosing interval (δ1 , δ2 ) = (0.17, 0.23), nj is the
number of patients treated at dose level j, and yj is the number of DLTs
observed at dose level j.

Number of patients nj 1 2 3 4 5 6 7 8 9 10 11 12
Escalate to j + 1 if yj ≤ 0 0 0 0 0 1 1 1 1 1 1 2
De-escalate to j − 1 if yj ≥ 1 1 1 1 2 2 2 2 3 3 3 3
Eliminate levels j to J if yj ≥ NA NA 2 3 3 3 4 4 4 5 5 5

mTPI, mTPI-2 (Guo and Yuan, 2017), is equivalent to the keyboard design,
but is perplexing and less transparent than the keyboard design. mTPI-2 relies
on complicated procedures, such as Occam’s razor and model selection, which
are difficult to understand and communicate with non-statisticians.

3.4 Bayesian optimal interval (BOIN) design

3.4.1 Trial design
Unlike the mTPI and keyboard designs, which require specifying the prior
distribution of πj and calculating πj ’s posterior for each of the possible val-
ues of (nj , yj ), the BOIN design is more transparent and straightforward
(Yuan et al., 2016a). Under the BOIN design, the decision of dose escala-
tion and de-escalation involves only a simple comparison of the observed
DLT rate at the current dose with a pair of fixed, prespecified dose es-
calation and de-escalation boundaries (Liu and Yuan, 2015; Yuan et al.,
2016a). In addition, it has a theoretical foundation of possessing the opti-
mal property by minimizing the decision error. Due to its transpancy and
competitive and robust performance, the US Food and Drug Administration
(FDA) granted the BOIN design the fit-for-purpose designation as a drug de-
velopment tool in 2021 (https://www.fda.gov/drugs/development-approval-
process-drugs/drug-development-tools-fit-purpose-initiative). This is the first
MTD-finding design that has received this designation.
Specifically, let π̂j = yj /nj denote the observed DLT rate at the current
dose, and λe and λd denote the optimal dose escalation and de-escalation
boundaries, respectively, which will be described later. The BOIN design is
illustrated in Figure 3.2 and described as follows:

1. Patients in the first cohort are treated at the lowest dose d1 , or the
physician-specified dose.
Model-Assisted Dose Finding Designs 39

TABLE 3.3: The escalation/de-escalation boundaries (λe , λd ) under the

BOIN design for different target toxicity rates φ, using the default under-
dosing toxicity probability φ1 = 0.6φ and overdosing toxicity probability
φ2 = 1.4φ.

Target toxicity probability φ

Boundaries 0.15 0.2 0.25 0.3 0.35 0.4
λe 0.118 0.157 0.197 0.236 0.276 0.316
λd 0.179 0.238 0.298 0.358 0.419 0.479

2. Suppose j is the current dose level; to assign a dose to the next

cohort of patients:
• if π̂j ≤ λe , escalate the dose to level j + 1.

Start
at the prespecified
starting dose

Treat a patient or a
cohort of patients

Reach
Stop the trial and Yes
the maximum
select the MTD sample size

No

≤ λe Compute !Lϊ
the DLT rate*
at the current
dose

Within (λe, λd@

Retain the current De−escalate the

Escalate the dose
dose dose

Total number of patients who experienced DLT at the current dose

* DLT rate =
Total number of patients treated at the current dose

FIGURE 3.2: BOIN design flowchart.

40 Model-Assisted Bayesian Designs for Dose Finding and Optimization

• if π̂j > λd , de-escalate the dose to level j − 1.

• otherwise, (i.e., λe < π̂j ≤ λd ), stay at the current dose level
j.
3. Repeat step 2 until the maximum sample size N is reached. At that
point, select MTD as the dose whose isotonic estimate π̃j is closest
to the target φ, where π̃j is obtained by applying the pooled adja-
cent violators algorithm (Barlow et al., 1972) to {π̂j , j = 1, . . . , J}.
During the trial conduct, the BOIN design imposes a dose elimination (or
overdose control) rule as follows:

if Pr(πj > φ | Dj ) > 0.95 and nj ≥ 3, dose levels j and higher are elimi-
nated from the trial, and the trial is terminated if the lowest dose level is
eliminated.
The value of Pr(πj > φ | Dj ) is evaluated based on the posterior distribution
(3.2). When the trial is terminated due to toxicity, no dose should be selected
as MTD.
Table 3.3 provides the optimal dose escalation and de-escalation bound-
aries (λe , λd ) for commonly-used target DLT probabilities φ. For example,
given φ = 0.3, the corresponding escalation boundary λe = 0.236 and the
de-escalation boundary λd = 0.358. That is, escalate/de-escalate/retain the
dose if 0/3 or 2/3 or 1/3 patients have DLT, respectively, given that three
patients have been treated at the current dose.
Table 3.4, or equivalently Figure 3.3, shows the discretized escalation/de-
escalation boundaries up to nj = 12 when φ = 0.3. This discretized version of
the decision table is often handy in practice to conduct the trial, which can
be easily generated using the software described later.
The hallmark of the BOIN design is its simplicity and transparency. Ar-
guably, its decision rule is even simpler than the 3+3 design; it just involves a
simple comparison between π̂j and (λe , λd ). Such simplicity renders BOIN sev-
eral important advantages over other model-assisted designs (e.g., the mTPI
and keyboard designs). From the statistical viewpoint, π̂j is the (nonpara-
metric) maximum likelihood estimate of πj , and it enjoys desirable statistical
properties such as being consistent and efficient. From a practical viewpoint,
π̂j is the most natural and intuitive estimate of πj that is accessible by non-
statisticians, making the BOIN design simpler and more transparent than the
mTPI and keyboard designs. In our experience, explaining the BOIN design
to clinicians, especially when equipped with the flowchart displayed in Figure
3.2, is easy and well received.
In addition, due to the feature that the BOIN design guarantees de-
escalating the dose when π̂j > λd , it is easy for clinicians and regulatory
agencies to assess the safety of a trial using the BOIN design. For example,
given a target DLT rate φ = 0.25, we know a priori that a phase I trial using
the BOIN design guarantees de-escalating the dose if the observed DLT rate is
Model-Assisted Dose Finding Designs 41

TABLE 3.4: The escalation/de-escalation boundaries of the BOIN design up

to 12 patients at a dose when the target toxicity rate φ = 0.3.

The number of patients treated at the current dose

Action 1 2 3 4 5 6 7 8 9 10 11 12
Escalate if no. 0 0 0 0 1 1 1 1 2 2 2 2
of DLT <=
De-escalate if no. 1 1 2 2 2 3 3 3 4 4 4 5
of DLT >=
Eliminate if no. NA NA 3 3 4 4 5 5 5 6 6 7
of DLT >=
“no. of DLT” is the number of patients with at least one DLT. When none
of the actions (i.e., escalate, de-escalate, or eliminate) is triggered, stay at
the current dose for treating the next cohort of patients. “NA” means that
a dose cannot be eliminated before treating three evaluable patients.

FIGURE 3.3: BOIN design decision table up to 12 patients at a dose when

the target toxicity rate φ = 0.3.
42 Model-Assisted Bayesian Designs for Dose Finding and Optimization

higher than λd = 0.298 (i.e., the default de-escalation boundary). Accordingly,

the BOIN design also allows users to easily calibrate the design to satisfy a
specific safety requirement mandated by regulatory agencies by choosing an
appropriate target DLT rate φ. For example, consider a phase I trial with a
new compound for which the regulatory agency mandates a dose de-escalation
for any observed toxicity rate higher than 0.25. We can easily fulfill that re-
quirement by setting the target DLT rate φ = 0.21 (based on equation (3.9)
described in the next section), under which BOIN automatically guarantees
de-escalating the dose if the observed toxicity rate π̂j > λd = 0.25. Such flex-
ibility and transparency grants the BOIN design an important advantage in
practice.
Another unique strength of BOIN is its Bayesian-frequentist dual inter-
pretation, making the design appealing to wider audiences. In contrast, the
mTPI and keyboard designs only have a Bayesian interpretation and require
specifying priors and calculating posterior distributions. As shown in the fol-
lowing sections, BOIN is derived as a Bayesian optimal design, but has fre-
quentist interpretation and optimality (i.e., under the non-informative prior,
the BOIN decision rule is equivalent to using the likelihood ratio test to deter-
mine dose escalation/de-escalation (Liu and Yuan, 2015)). That is also why
the BOIN’s decision rules (with the non-informative prior) have the appear-
ance of a classical frequentist design and only involve the observed DLT rate
π̂j , the maximum likelihood estimate of πj .
Despite its simplicity, numerical studies show that BOIN yields excel-
lent performance comparable to the model-based CRM design (Zhou et al.,
2018a,b; Ruppert and Shoben, 2018). BOIN outperforms mTPI with higher
accuracy for identifying MTD and a substantially lower risk of overdos-
ing patients, and it is simpler and more transparent than the keyboard (or
mTPI-2) design. In addition, BOIN also is more versatile; it can handle drug-
combination trials (Lin and Yin, 2017a; Zhang and Yuan, 2016), late-onset
toxicity (Yuan et al., 2018), low-grade toxicities (Lin, 2018), and toxicity and
efficacy jointly (Takeda et al., 2018; Zhou et al., 2019; Lin et al., 2020b),
and incorporate historical information and real-world evidence (Zhou et al.,
2021a). These extensions of BOIN will be covered in later chapters.
More importantly, BOIN has been widely validated in practice. It has
been used in variety of oncology trials, including trials for pediatric tumors,
adult tumors, solid tumors, and liquid tumors, see Yuan et al. (2019) for a
list of completed and ongoing trials that have implemented the BOIN design.
Additionally, BOIN has been used in non-oncology trials such as stem cell
therapy for patients with stroke (Phan et al., 2018).

3.4.2 Theoretical derivation

This section describes the theoretical foundation of the BOIN design, which
provides the basis for various extensions of the design (e.g., drug combination,
late-onset toxicity, and toxicity-efficacy tradeoff) described in later chapters.
Model-Assisted Dose Finding Designs 43

The BOIN design is derived based on the optimal design theory. Let
λe (dj , nj , φ) and λd (dj , nj , φ) denote the general dose escalation and de-
escalation boundaries that are unspecified functions of the dose (i.e., dj ), the
number of patients treated (i.e., nj ), and the DLT target (i.e., φ), where
0 ≤ λe (dj , nj , φ) < λd (dj , nj , φ) < 1. Consider a class of nonparametric de-
signs Cnp :
1. Patients in the first cohort are treated at the lowest dose d1 , or the
physician-specified dose.
2. Suppose j is the current dose level; to assign a dose to the next
cohort of patients:
• if π̂j ≤ λe (dj , nj , φ), escalate the dose level to j + 1.
• if π̂j > λd (dj , nj , φ), de-escalate the dose level to j − 1.
• otherwise, i.e., λe (dj , nj , φ) < π̂j ≤ λd (dj , nj , φ), stay at the
current dose level j.
3. Repeat step 2 until the maximum sample size N is reached.
As escalation and de-escalation boundaries λe (dj , nj , φ) and λd (dj , nj , φ) can
freely vary according to dj , nj , and φ, this class of designs are extremely
broad and contain all possible nonparametric designs that do not impose
the parametric assumption on the dose-toxicity curve and make dose esca-
lation and de-escalation based on the local data Dj = (yj , nj ). The mTPI,
keyboard/mTPI-2, and BOIN designs all belong to Cnp . For notational brevity,
we use shorthands λej ≡ λe (dj , nj , φ) and λdj ≡ λd (dj , nj , φ).
The BOIN design is obtained by minimizing the probability of making
incorrect decisions of dose escalation and de-escalation within Cnp . Liu and
Yuan (2015) described two versions of the BOIN design: the local optimal
BOIN design, which is optimized based on point hypotheses, and the global
optimal BOIN design, which is optimized based on interval hypotheses. Liu
and Yuan (2015) recommended the local optimal BOIN design because of
its better performance. Thus, we here focus on the development of the local
optimal BOIN design and simply call it the BOIN design.
To proceed, define three point hypotheses:

H0j : πj = φ, H1j : πj = φ1 , H2j : πj = φ2 ,

where φ1 indicates that the dose is substantially underdosing (i.e., below

MTD) such that escalation is required, and φ2 indicates that the dose is
substantially overdosing such that de-escalation is required. Let S, E, and D
denote the decisions of stay (at the current dose), escalate, and de-escalate,
respectively. Under H0j , the correct decision is S, and incorrect decisions are
S̄ = {E, D}; under H1j , the correct decision is E, and incorrect decisions are
44 Model-Assisted Bayesian Designs for Dose Finding and Optimization

Ē = {S, D}; and under H2j , the correct decision is D, and incorrect decisions
are D̄ = {S, E}.
Under the Bayesian paradigm, we assign each of the hypotheses a prior
probability of being true, denoted as ωkj = Pr(Hkj ), k = 0, 1, 2. The prob-
ability of making an incorrect decision (the decision error rate), denoted as
α(λej , λdj ), at each of the dose assignments is given by
α(λej , λdj )
= Pr(H0j ) Pr(S̄|H0j ) + Pr(H1j ) Pr(Ē|H1j ) + Pr(H2j ) Pr(D̄|H2j )
= Pr(H0j ) Pr(yj ≤ nj λej or yj > nj λdj |H0j ) + Pr(H1j ) Pr(yj > nj λej |H1j )
+ Pr(H2j ) Pr(yj ≤ nj λdj |H2j )
= ω0j {Bin(nj λej ; nj , φ) + 1 − Bin(nj λdj ; nj , φ)} +
ω1j {1 − Bin(nj λej ; nj , φ1 )} + ω2j Bin(nj λdj ; nj , φ2 ), (3.3)
where Bin(b; n, φ) is the cumulative density function (CDF) of the binomial
distribution, with size and probability parameters n and φ evaluated at the
value b. We rewrite the decision error α(λej , λdj ) as
α(λej , λdj ) = α1 (λej ) + α2 (λdj ) + ω0j + ω1j ,
where
α1 (λej ) = ω0j Bin(nj λej ; nj , φ) − ω1j Bin(nj λej ; nj , φ1 )
α2 (λdj ) = ω2j Bin(nj λdj ; nj , φ2 ) − ω0j Bin(nj λdj ; nj , φ).
To minimize α(λej , λdj ), we can minimize α1 (λej ) and α2 (λdj ) separately with
regard to λej and λdj , respectively. As α1 (λej ) and α2 (λdj ) are symmetric,
below we consider the minimization of α1 (λej ):
α1 (λej ) = ω0j Bin(nj λej ; nj , φ) − ω1j Bin(nj λej ; nj , φ1 )
bnj λej c  
X nj
= {ω0j φy (1 − φ)nj −y − ω1j φy1 (1 − φ1 )nj −y }
y=0
y
bnj λej c   (  y  nj −y )
X nj y ω0j φ 1 − φ
= ω1j φ1 (1 − φ1 )nj −y −1 .
y=0
y ω1j φ1 1 − φ1

By the definition of the CDF, αe (λej ) = 0 when bnj λej c < 0 and α1 (λej ) = 1
when bnj λej c ≥ nj .
Assuming that y ∗ is continuous and setting
 y ∗  nj −y∗
ω0j φ 1−φ
− 1 = 0, (3.4)
ω1j φ1 1 − φ1
we obtain    
1 − φ1 ω1j
nj log + log
1−φ ω
y∗ =   0j .
φ(1 − φ1 )
log
φ1 (1 − φ)
Model-Assisted Dose Finding Designs 45
 y  nj −y
Because φ > φ1 , φφ1 1−φ
1−φ1 monotonically increases with y. It follows
  y  nj −y 
π0j φ 1−φ
that π1j φ1 1−φ1 − 1 ≥ 0 when y ≥ y ∗ , and < 0 when y < y ∗ .
Therefore, given 0 ≤ y ≤ nj , α1 (λej ) is minimized when
∗
y ∗ ≥ nj ,

 [nj −I(y = nj ), ∞) , if

nj λej ∈ [dy ∗ e − 1, by ∗ c + 1) , if 0 < y ∗ < nj ,
(−∞, I(y ∗ = 0)), y ∗ ≤ 0,

if


where I(·) is an indicator function. This leads to the solution

  ∗

I(y = n j )
 1−


 nj ,∞ , if y ∗ ≥ nj ,



  ∗
dy e − 1 by ∗ c + 1

λej ∈ nj , nj , if 0 < y ∗ < nj ,



  I(y ∗ = 0)
 
 −∞, , if y ∗ ≤ 0.


nj
(3.5)
Because of the symmetry of α1 (λej ) and α2 (λdj ), it follows that the solution
that minimizes α2 (λdj ) is given by

I(y ∗∗ = nj )
  



 1− nj , ∞ , if y ∗∗ ≥ nj ,



  ∗∗
dy e − 1 by ∗∗ c + 1

λdj ∈ nj , nj , if 0 < y ∗∗ < nj , (3.6)



I(y ∗∗ = 0)

  
 −∞, , if y ∗∗ ≤ 0,


nj

where    
1−φ ω0j
nj log + log
1 − φ2 ω
y ∗∗ =   2j .
φ2 (1 − φ)
log
φ(1 − φ2 )
As any values of λej and λdj located in the interval solutions (3.5) and
(3.6) produce the same error rate, for the purpose of designing the trial and
making decisions, a point solution located in the interval solutions is sufficient.
As dxe − 1 < x < bxc + 1, one specific “middle” point solution is
   
1 − φ1 −1 ω1j
log + nj log
1−φ ω
λ∗ej = y ∗ /nj =   0j , (3.7)
φ(1 − φ1 )
log
φ1 (1 − φ)
46 Model-Assisted Bayesian Designs for Dose Finding and Optimization
   
1−φ −1 ω0j
log + nj log
1 − φ2 ω
λ∗dj = y ∗∗ /nj =   2j . (3.8)
φ2 (1 − φ)
log
φ(1 − φ2 )
This is the solution provided by Liu and Yuan (2015). The above derivation
provides a more complete interval solution pair (3.5) and (3.6).
When the non-informative prior ω0j = ω1j = ω2j = 1/3 is used, which is
recommended for most trials, the optimal escalation boundaries become
   
1 − φ1 1−φ
log log
1−φ 1 − φ2
λe =  , λd =  , (3.9)
φ(1 − φ1 ) φ2 (1 − φ)
log log
φ1 (1 − φ) φ(1 − φ2 )

which are independent of j and nj . This results in the hallmark simplicity of

BOIN — using a pair of fixed dose escalation and de-escalation boundaries
throughout the trial to guide the dose transition.
As noted by Liu and Yuan (2022), the original publication of the BOIN
design (Liu and Yuan, 2015) had incorrectly specified the de-escalation rule
as: de-escalate the dose if π̂j ≥ λd , which should be π̂j > λd , as shown
above. That error does not affect the application of the BOIN design as it is
virtually impossible that π̂j = λd in practice. Liu and Yuan (2022) examined
the target DLT rate φ ∈ {0.1000, 0.1001, 0.1002, . . . , 0.4000} with nj = 30
(i.e., up to treating 30 patients per dose), and did not find any case that
π̂j = λd . In addition, de-escalation if π̂j ≥ λd is slightly more conservative
than de-escalation if π̂j > λd . Therefore, even if the equal sign could be taken,
there are no safety concerns that arise for the trials using the “original” rule.

3.4.3 Specification of design parameters

The specification of φ1 and φ2 are critical as they determine the dose escala-
tion and de-escalation boundaries (λe , λd ). Liu and Yuan (2015) recommended
the default values of φ1 = 0.6φ and φ2 = 1.4φ for general use, which yield
good operating characteristics as confirmed by numerous subsequent simu-
lation studies and real clinical trials. These are the default values used in
the BOIN software described later. Therefore, the BOIN design is essentially
calibration-free: once the target DLT rate φ is specified, the design (i.e., es-
calation and de-escalation boundaries λe and λd ) is determined and ready to
use.
When needed, however, the values of φ1 and φ2 can be tuned to achieve a
particular requirement of the trial at hand. For example, if more conservative
dose escalation is required, we may set φ2 = 1.2φ. The performance of BOIN
is generally robust to the specification of φ1 and φ2 , but we should refrain
from setting φ1 and φ2 too close to φ (e.g., a difference less than 20% of φ).
This is because the small sample size of phase I trials provides virtually no
Model-Assisted Dose Finding Designs 47

power to distinguish a difference smaller than that, and minimizing decision

errors under such hypotheses does not make any practical sense. For example,
given the sample size of 30 patients per dose, based on Fisher’s exact test,
we only have 7% power to distinguish a dose with a toxicity rate of 0.3 from
the other with toxicity rate of 0.35, at the one-sided significance level of 0.05,
using R function power.fisher.test( ) in the statmod package.
As a side note, φ1 and φ2 used in the BOIN design have different interpre-
tations than the proper dosing interval (δ1 , δ2 ) used in the mTPI and keyboard
designs. Specifically, φ1 and φ2 represent the DLT rates that are unacceptable
(i.e., underdosing or overdosing such that dose escalation or de-escalation is
required); whereas δ1 and δ2 represent the range of DLT probabilities that
are acceptable. For example, given that the target DLT probability φ = 0.25,
setting φ1 = 0.15 and φ2 = 0.35 means that the doses with the DLT rates
of 0.15 and 0.35 are regarded as unacceptably underdosing and overdosing,
respectively, whereas setting δ1 = 0.15 and δ2 = 0.35 means that the dose
with a DLT rate between 0.15 and 0.35 is regarded as acceptable. Thus, in
general, the value of φ1 should be smaller than that of δ1 and the value of φ2
should be greater than that of δ2 .

3.4.4 Statistical properties

Bayesian and frequentist optimality

One unique property of BOIN is that it enjoys both Bayesian and frequentist
interpretation and optimality.

Theorem 3.1 The dose escalation and de-escalation boundaries (λe , λd ) of

the BOIN design are the boundaries from where Pr(H1 |Dj ) ≥ Pr(H0 |Dj )
and Pr(H2 |Dj ) > Pr(H0 |Dj ), respectively. When the non-informative prior is
used, the boundaries are the likelihood ratio test boundaries.

In other words, the decision rule of BOIN is equivalent to the following intu-
itive Bayesian decision rule:
Pr(H1j |Dj )
• If ≥ 1 (i.e., the data indicate that H1j is equal or more likely
Pr(H0j |Dj )
to be true than H0j ), escalate the dose.
Pr(H2j |Dj )
• If > 1 (i.e., the data indicate that H2j is more likely to be true
Pr(H0j |Dj )
than H0j ), de-escalate the dose.
The proof of Theorem 3.1 is straightforward noting that the ratio in equation
(3.4) is the Pr(H1j |Dj )/Pr(H0j |Dj ).
Define L(Dj |H0j ) ∝ φyj (1 − φ)(nj −yj ) as the binomial likelihood func-
tion of the data Dj = (nj , yj ) under H0j , and similarly define L(Dj |Hkj ) as
the likelihood function under Hkj , k = 1, 2. When the non-informative prior
48 Model-Assisted Bayesian Designs for Dose Finding and Optimization

is used, the decision rule of BOIN is equivalent to the following frequentist

decision rule:
L(Dj |H1j )
• If ≥ 1 (i.e., the likelihood of Dj under H1j is equal to or greater
L(Dj |H0j )
than that under H0j ), escalate the dose.
L(Dj |H2j )
• If > 1 (i.e., the likelihood of Dj under H1j is greater than that
L(Dj |H0j )
under H0j ), de-escalate the dose.

Long-memory coherence
Coherence is a finite-sample property that describes how a phase I design be-
haves in dose escalation and de-escalation in light of observed DLT data. Che-
ung (2005) originally defined coherence as a design property by which dose es-
calation (or de-escalation) is prohibited when the most recently treated patient
experiences (or does not experience) toxicity. Liu and Yuan (2015) extended
that concept and defined two different types of coherence: short-memory co-
herence and long-memory coherence. They referred to the coherence proposed
by Cheung (2005) as short-memory coherence because it concerns the obser-
vation from only the most recently treated patient, ignoring the observations
from the patients who were previously treated. By contrast, long-memory co-
herence concerns the accumulated data observed from the most recent dose
level.
Definition (Short-memory coherence) A design is called short-memory
coherent if it never escalates the dose when the most recently treated
patient experiences DLT, and never de-escalates the dose when the most
recently treated patient does not experience DLT.
Definition (Long-memory coherence) A design is called long-memory co-
herent if it never escalates the dose when the observed DLT rate at the
current dose is higher than the target φ, and never de-escalates the dose
when the observed DLT rate at the current dose is lower than φ.
From a practical viewpoint, long-memory coherence is more relevant be-
cause when clinicians determine whether a dose escalation/de-escalation is
practically plausible, they almost always base their decision on the toxicity
data from all patients treated at the current dose, rather than only the sin-
gle patient most recently treated. This is more important considering that,
patients in phase I trials are highly heterogeneous, and the toxicity outcome
from a single patient can be spurious. For example, suppose the target DLT
rate φ = 0.3 and, at the current dose, the most recently treated patient expe-
rienced DLT but none of the nine patients previously treated at the same dose
had DLT. As the overall observed DLT rate at the current dose is 1/10, esca-
lating the dose should not be regarded as an inappropriate action, although
it violates short-memory coherence.
Model-Assisted Dose Finding Designs 49

Theorem 3.2 The BOIN design based on the non-informative prior with
ω0j = ω1j = ω2j = 1/3 is long-memory coherent.

The proof of Theorem 3.2 is straightforward based on the equation (3.9). Be-
sides the above finite-sample properties, Liu and Yuan (2015) showed that,
assuming the existence of the target dose (i.e., at least a dose located in
(λe , λd )), the BOIN design also has the following desirable large-sample prop-
erty.
Theorem 3.3 As the number of patients goes to infinity, the dose assignment
and the selection of MTD under the BOIN design converge almost surely to
dose level j ∗ if dose level j ∗ is the only dose satisfying πj ∗ ∈ (λe , λd ). If there
are multiple dose levels in (λe , λd ), the design will converge almost surely to
one of these levels.
As a side note, one might be concerned that BOIN converges to the “stay”
interval (λe , λd ), rather than the target φ. Actually, this is not a concern. As
noted by Zhou et al. (2021b), under large samples it is more √ appropriate to
use the local
√ alternative hypotheses H 1 : π j = φ − ∆ 1 / n and H2 : πj =
φ + ∆2 / n, where ∆1 and ∆2 are constant, rather than fixed alternatives
H1 : πj = φ1 and H2 : πj = φ2 as used in the above theorem. Then, as
(λe , λd ) converges to φ, the dose assignment and the selection of MTD under
the BOIN design naturally converge almost surely to φ.

3.4.5 Frequently asked questions

In this section, we describe several frequently asked questions with answers to
provide more insights on BOIN and model-assisted designs.

1. Does the BOIN decision rule account for the variance of π̂j (or equivalently,
the sample size nj )?

The answer is yes. The simplicity of the BOIN design (i.e., making decisions
by comparing π̂j with λe and λd ) might lead one to think that the BOIN
decision rule does not consider the variance of π̂j (or equivalently, the sample
size nj ). This, however, is not true. As shown by equation (3.3), the derivation
and minimization of the decision error α depends on the sampling distribution
of π̂j , thus it directly accounts for the uncertainty of π̂j . The optimal decision
boundaries independent of nj should not be mistakenly regarded as ignoring
nj .
To help readers to understand this point, consider an experiment of draw-
ing balls with replacement from a bag of red and black balls. There are a total
of 9 balls in the bag, but we do not know if there are more red or black balls.
The objective is to determine if there are more red or black balls. The exper-
iment is to randomly draw a ball from the bag, record the color, put it back,
and repeat. Clearly, no matter whether we do the experiment 3 or 30 times, as
50 Model-Assisted Bayesian Designs for Dose Finding and Optimization

long as we see more red balls, the best decision is to claim that there are more
red balls. The only difference is that the decision based on 30 experiments
has a smaller decision error, although both minimize the decision error. This
is exactly how the Bayes classifier works, which optimizes and minimizes the
Bayes error rate (Berger, 2013).

2. Is it reasonable to assume the non-informative prior ω0j = ω1j = ω2j = 1/3

for all doses d1 < · · · < dJ ?

The answer is yes. This seems counterintuitive: as toxicity monotonically in-

creases with the dose, it seems natural to assume that high doses have higher
prior probabilities of overdosing (e.g, ω21 < · · · < ω2J ). The reason that it
is reasonable to assume non-informative prior ω0j = ω1j = ω2j = 1/3 hinges
on the fundamental characteristic of dose finding: it is a sequential decision-
making process of testing the dose in the one-at-a-time fashion from low to
high doses. We escalate to the next higher dose only when interim data show
that the current dose is safe. For instance, when data (e.g., 0/3 DLT) show
that d1 is safe, we are ready to escalate to d2 . At that moment, although we
know d2 is more toxic than d1 , we do not know by how much. Compared to
d1 , a priori, it is not clear that d2 is just slightly more toxic and still deemed
safe, or more toxic and close to the MTD, or substantially more toxic and
thus above the MTD. In other words, during the trial, for each new dose to be
escalated to for testing, a priori, we do not know if it is underdosing, proper
dosing, or overdosing. Thus, the most sensible approach is to assume that dj ,
j = 1, · · · , J, has an equal probability of being underdosing, proper dosing
and overdosing (i.e., ω0j = ω1j = ω2j = 1/3). Actually, the dose escalation
rule itself (i.e., escalate when the current dose is safe) already (implicitly)
accounts for the prior information that a higher dose is expected to be more
toxic. There is no need to do it again by imposing an informative prior. Liu and
Yuan (2015) evaluated the use of an informative prior (e.g., ω21 < · · · < ω2J )
and found that overall it does not improve performance. Of note, when the
trial is completed, to evaluate the dose-toxicity relationship, BOIN does ac-
count for the monotonicity to estimate πj by isotonic regression.

3. BOIN makes dose escalation/de-escalation decisions based only on the lo-

cal data at the current dose. Does this seemly “myoptic” approach cause any
notable loss of efficiency?

The answer is no. For the purpose of dose finding, the loss of efficiency due to
the use of local data is minimal, and mostly ignorable. This is because unlike
most statistical inferential procedures, dose finding is a sequential decision-
making process, escalating from low doses to high doses. Suppose that the
current dose level is j. In order to reach j, the data observed previously at lower
doses (i.e., < j) must indicate that these doses are safe and substantially lower
than the MTD (e.g., 0/3 DLT). Thus, these data provide little information
to determine whether the current dose j is below, equal (or sufficiently close)
Model-Assisted Dose Finding Designs 51

to, or above the MTD to make the decision of dose escalation/de-escalation.

Similarly, in order to reach j, the data observed previously at higher doses
(i.e., > j), if exist, must indicate that these doses are substantially higher
than the MTD (e.g., 2/3 DLT). Thus, borrowing information from these data
helps little to determine if the current dose j is below, equal (or sufficiently
close) to, or above the MTD. In other words, the information used to evaluate
the safety of dose j and make the decision of dose escalation/de-escalation
is mostly provided by the local data observed at dose j. This explains why
in general BOIN yields comparable performance as CRM that uses a dose-
toxicity model to borrow information across doses, as demonstrated in the
next section. Lin and Yuan (2019) showed that incorporating the data other
than the current dose into BOIN provides virtually no efficacy gain.
Although BOIN uses local data to make the decisions of dose escalation/de-
escalation, it indeed uses all data to select the MTD based on nonparametric
isotonic regression. As noted by Liu and Yuan (2015), dose transition and
MTD selection actually are two independent components of a dose finding
study. When the trial is completed, any method (e.g., a logistic model), can
also be used with BOIN to select the MTD. The parametric model yields
better MTD selection when the model is correctly specified, but may perform
poorly when the model is misspecified. Overall, isotonic regression provides a
robust and competitive approach.

4. What is the relationship between the keyboard/mTPI-2 design and BOIN?

Which one is better?

The keyboard/mTPI-2 design can be regarded as a special convoluted version

of BOIN. Keyboard/mTPI-2 makes decisions of dose escalation/de-escalation
based on the location of the strongest key (i.e., the interval with the largest
posterior probability), with respect to the target key/interval (δ1 , δ2 ). As each
key is of equal length and the beta distribution is unimodal, this is virtually
equivalent to pinpoint where the mode of f (πj |Dj ) is, with respect to (δ1 , δ2 ).
It is not exactly equivalent because the beta distribution is asymmetric. Ac-
cording to equation (3.2), when the unform prior πj ∼ Beta(1, 1) is assumed,
the mode of f (πj |Dj ) is simply π̂j = yj /nj . As a result, the dose escalation/de-
escalation/stay decision of keyboard/mTPI-2 boils down to whether π̂j ≤ δ1 ,
π̂j > δ2 , or π̂j ∈ (δ1 , δ2 ], which essentially is the BOIN decision rule. This
intrinsic link explains why keyboard/mTPI-2 often has similar performance
as BOIN (see the next section for a numerical study).
There are, however, several important differences between keyboard/mTPI-
2 and BOIN. First, BOIN is more transparent and straightforward, as de-
scribed previously. Second, (δ1 , δ2 ) in keyboard/mTPI-2 (when the uniform
prior Beta(1, 1) is assumed) plays the same role as dose escalation/de-
escalation boundaries (λe , λd ) in BOIN. However, the latter is optimized
to minimize the decision error, while the former is subjectively specified
and lack of clear statistical properties and assurance. Third, the operating
52 Model-Assisted Bayesian Designs for Dose Finding and Optimization

characteristics of keyboard/mTPI-2 critically rely on the prior assumption

that πj ∼ Beta(1, 1). Under the beta-binomial model, this prior assumes a
prior sample size of two patients and the prior DLT probability of 0.5 for each
dose, which may be overly informative (given that some doses often only treat
3 patients) and unrealistically toxic. Often, one may prefer a vague prior such
as Beta(0.02, 0.08), which is equivalent to a prior sample size of 0.1 patient
with a prior DLT probability of 0.2, to allow observed data to drive decisions.
Unfortunately, when this more reasonable vague prior is used, the operating
characteristics of keyboard/mTPI-2 become elusive (e.g., resulting in overly
aggressive dose escalation) (Zhou et al., 2021a). In this case, the mode of
f (πj |Dj ) is not π̂j anymore, and the decision rule of keyboard/mTPI-2 di-
verges from the BOIN decision rule. In contrast, BOIN does not require the
prior assumption πj ∼ Beta(1, 1). It only assumes that for any new dose to
be tested, it has an equal probability of being below, equal (or sufficiently
close) to, or above the MTD, see the answer to question 3 for why it is a
reasonable prior assumption of equipoise. BOIN’s dose elimination rule em-
ploys a beta-binomial model with the uniform prior π ∼ Beta(1, 1), see Section
3.4.1. However, for the purpose of dose elimination (or overdose control), using
Beta(1, 1) prior is not critical. When another prior is used (e.g., Beta(0.02,
0.08)), the overdose control probability cutoff 0.95 can be slightly adjusted
(e.g., to 0.93) to obtain same operating characteristics. Therefore, BOIN is a
preferred, simpler, and more robust choice of design.

3.5 Operating characteristics

Zhou et al. (2018b) conducted comprehensive Monto Carlo studies to compare
the operating characteristics of various phase I designs, including the 3+3 de-
sign, three model-based designs (CRM, EWOC, and BLRM), and three model-
assisted designs (mTPI, keyboard design, and BOIN). In addition, variations
of the CRM and BLRM designs, including CRM-DS (allowing dose skipping)
and BLRM-NOC (without overdose control), were also investigated. The tar-
get DLT probability is φ = 0.25, with J = 6 dose levels and a maximum sample
size of N = 36. The starting dose level is j = 1. Patients are treated in cohorts
of size three. The proper dosing interval is (δ1 , δ2 ) = (φ−0.05, φ+0.05) for the
mTPI, keyboard, and BLRM designs, and the BOIN design uses φ1 = 0.6φ
and φ2 = 1.4φ as recommended by these designs. The 3+3 design often com-
pletes before reaching the prespecified maximum sample size (e.g., when 2/3
or 2/6 had DLT). For fair comparisons, after the 3+3 design selects MTD, an
expansion cohort is treated at MTD to reach a total sample size of 36. The
detailed configurations for these designs and more simulation results (e.g.,
φ = 0.2 or 0.3, or cohort size = 1) can be found in Zhou et al. (2018b).
To avoid cherry-picking and inadvertent selection biases, 1000 true dose–
toxicity scenarios (or curves) were randomly generated using the pseudo-
uniform algorithm (Clertant and O’Quigley, 2017) for comparison. Given a
Model-Assisted Dose Finding Designs 53

FIGURE 3.4: Panel (a): 25 randomly selected dose–toxicity curves with six
picked curves showing different shapes; Panel (b): Distribution of the DLT
probabilities by dose level from the 1000 randomly generated scenarios.

target DLT rate φ and J dose levels, the random scenarios were generated as
follows:
1. Select one of the J dose levels as the MTD with equal probabilities.
2. Sample M ∼ Beta(max{J − j, 0.5}, 1), where j denotes the selected
MTD level, and set an upper bound B = φ + (1 − φ)M for the
toxicity probabilities.
3. Repeatedly sample J toxicity probabilities uniformly on [0, B] until
these correspond to a scenario in which dose level j is MTD.
Figure 3.4 panels (a) and (b) show 25 randomly selected scenarios and dis-
tributions of the DLT probabilities by dose level from the 1000 scenarios,
respectively. It can be seen that the simulated dose–toxicity curves cover var-
ious shapes and a wide range of toxicity probabilities. The algorithm above
guarantees that the generated dose–toxicity curves are monotonically increas-
ing (i.e., higher doses have higher toxicity rates). For each scenario, 2000 trials
were simulated.

Performance metrics

• Accuracy
A1. The percentage of correct selection (PCS), defined as the percentage of
simulated trials in which the target dose is correctly selected as MTD. When
54 Model-Assisted Bayesian Designs for Dose Finding and Optimization

all the dose levels are above MTD (i.e., the DLT probability of the lowest
dose > φ + 0.1), PCS is defined as the percentage of early termination of
trials.
A2. The average percentage of patients who are assigned to MTD across
the simulated trials. When all the dose levels are above MTD (i.e., the DLT
probability of the lowest dose > φ + 0.1), the average percentage of patients
not enrolled into the trial is used for this metric.

• Safety
B1. The percentage of simulated trials in which a toxic dose with the true
DLT probability > 33% is selected as MTD when the target φ = 25%.
B2. The average percentage of patients assigned to the toxic doses with true
DLT probabilities > 33% when the target φ = 25%.
• Reliability
C1. The risk of overdosing, defined as the percentage of simulated trials with
more than 50% of patients treated at doses above MTD.
C2. The risk of poor allocation, defined as the percentage of simulated trials
in which fewer than six patients are treated at MTD.
C3. The risk of irrational dose assignment, defined the percentage of times
that the design fails to de-escalate the dose when 2/3 or > 3/6 patients had
DLTs at a dose.
Reliability metrics C1 to C3 measure the likelihood of a design demonstrat-
ing problematic behaviors (e.g., treating 50% or more patients at toxic doses,
or fewer than six patients at MTD) that have severe clinical consequences.
These metrics are of great practical importance, but unfortunately are often
overlooked in the literature. The reliability metrics are not covered by other
metrics. For example, the percentage of patients overdosed (i.e., metric B2)
does not cover the risk of overdosing (i.e., metric C1). Two designs can have
similar percentages of patients overdosed, but rather different risks of over-
dosing 50% of the patients. Statistically, metric B2 measures the mean of
overdosing, while metric C1 measures the tail probability of overdosing.

Results
Accuracy Panels A1 and A2 in Figure 3.5 show distributions of the PCS
and the average percentages of patients treated at MTD, respectively, for the
investigational designs relative to the 3+3 design across 1000 scenarios. That
is, the values displayed in the figure are the difference between those of a
specified design and the reference (i.e., 3+3 design). For example, PCS = 0
means that the design has the same PCS as the 3+3 design. As each dose–
toxicity scenario generates a value of the performance metric (e.g., PCS), there
are a total of 1000 values for each of the metrics across the 1000 scenarios.
The boxplot reflects the distribution of the metric across the 1000 scenarios. In
terms of the accuracy of correctly selecting the MTD, the CRM, mTPI, BOIN,
Model-Assisted Dose Finding Designs 55

and keyboard designs are comparable and substantially outperform the 3+3
design. The BLRM and EWOC designs perform the worst, with the average
PCS similar to that of the 3+3 design. The EWOC design also has the largest
variation in PCS. The results for the number of patients treated at MTD are
similar to those for PCS. The CRM, mTPI, BOIN, and keyboard designs are
generally comparable and substantially outperform the 3+3 design. The mTPI
and CRM designs allocate slightly more patients to MTD than the BOIN and
keyboard designs, but the latter two designs are less variable, as shown by
the shorter boxes in the box plot. mTPI is less robust than the BOIN and
keyboard designs. For example, when the target φ = 20%, mTPI has notably
lower PCS than the BOIN and keyboard designs, see Zhou et al. (2018b) for
details.

FIGURE 3.5: Accuracy and safety of the eight designs with respect to the 3+3
design, including (A1) percentage of correct selection of MTD, (A2) percentage
of patients treated at MTD, (B1) percentage of selecting a dose with the DLT
probability ≥ 33% as MTD, and (B2) percentage of patients treated at doses
with DLT probabilities ≥ 33%. For (A1) and (A2), a larger value indicates
better performance; a positive value means that the design outperforms the
3+3 design. For (B1) and (B2), a smaller value indicates better performance;
a negative value means that the design outperforms the 3+3 design.
56 Model-Assisted Bayesian Designs for Dose Finding and Optimization

Safety As shown in Figure 3.5 panel B1, the CRM, mTPI, BOIN, and key-
board designs are comparable in terms of the percentage of selecting a toxic
dose (with a DLT probability ≥ 33%) as MTD, but CRM and mTPI are
slightly more variable than the BOIN and keyboard designs. The BLRM and
EWOC designs are the most conservative and least likely to select a toxic dose
as MTD. In terms of the percentage of patients treated at a toxic dose with
a DLT probability ≥ 33%, on average the CRM, mTPI, BOIN, and keyboard
designs are comparable, but BOIN and keyboard show smaller variations (Fig-
ure 3.5 panel B2).
Reliability In terms of the risk of overdosing 50% or more of the patients
(Figure 3.6, panel C1), the BLRM, BOIN, and keyboard designs perform the
best. The performances of the CRM and mTPI designs are similar and rank

FIGURE 3.6: Reliability of the eight designs with respect to the 3+3 design,
including (C1) risk of overdosing 50% or more patients, (C2) risk of treating
< 6 patients at the MTD, and (C3) risk of irrational dose assignments. A
smaller value indicates better performance; negative value means that the
design outperforms the 3+3 design.
Model-Assisted Dose Finding Designs 57

in between the performances of these other designs. The EWOC design has
a similar averaged risk of overdosing patients as the BOIN and keyboard
designs, but is much more variable. Of note, the CRM, mTPI, BOIN, and
keyboard designs, on average, overdose similar percentages of patients (Figure
3.6 panel B2), but have different risks of overdosing 50% or more of the patients
(Figure 3.6, panel C1). This indicates that the risk of overdosing (50% or more
patients) and the average percentage of patients overdosed indeed measure
different aspects of a design, and it is thus important to consider both metrics
when evaluating a design. In terms of the risk of poor allocation (i.e., treating
fewer than six patients at the MTD, see Figure 3.6, panel C2), BLRM and
EWOC perform the worst, with a significantly higher risk than the other
designs. The CRM, BOIN, and keyboard designs have comparable risks of
poor allocation.
In terms of the risk of irrational dose assignment (Figure 3.6, panel C3), the
model-assisted designs outperform the model-based designs. The model-based
designs (i.e., CRM, BLRM, and EWOC) have an 8% to 55% chance of failing
to de-escalate the dose when ≥ 2/3 or ≥ 3/6 patients have DLTs, whereas such
irrational dose assignments never occur in the mTPI, BOIN, and keyboard
designs. The model-based designs rely on the assumed model to make the
decision of dose assignment. When the model is misspecified, the estimates
can be biased and thus irrational dose assignment arises. The model-assisted
designs are free of that issue because they do not impose any model assumption
on the dose–toxicity curve. For example, by its dose escalation/de-escalation
rule, the BOIN design guarantees de-escalating the dose if the observed DLT
rate at the current dose is higher than 29.8%, given the target DLT rate of
25%.
In summary, the model-assisted designs (e.g., the BOIN and keyboard
designs) substantially outperform the algorithm-based 3+3 design in the ac-
curacy of identifying MTD and allocating patients to MTD. They produce
competitive accuracy and safety comparable to the model-based designs (e.g.,
CRM), but are much simpler and more transparent. In addition, the model-
assisted designs are more robust, and avoid the irrational dose assignment of
the model-based designs due to model misspecification. Among the model-
assisted designs, BOIN stands out. It has similar operating characteristics as
the keyboard design, but is simpler, more flexible, and transparent. The mTPI
design is not recommended due to the poor reliability and safety concerns.

3.6 Software and case study

Software
Software for the BOIN design is available in several different forms: (1)
Web application “BOIN Suite” with intuitive graphical user interface, freely
58 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 3.7: The launchpad of the BOIN Suite, freely available at

http://www.trialdesign.org.

available at http://www.trialdesign.org. Figure 3.7 shows the launchpad

of BOIN Suite, which is a software platform to handle various types of early-
phase clinical trials based on the unified family of the BOIN design and its
extensions. Figure 3.8 shows the decision tree to assist users to choose an ap-
propriate BOIN design module for different types of trials; (2) R package BOIN
freely available at CRAN; and (3) Windows desktop program BOIN Suite, freely
available at the MD Anderson Cancer Center Software Download Website
https://biostatistics.mdanderson.org/softwaredownload/. We recom-
mend the web application as it is most frequently updated.
The keyboard design can be implemented using the web applica-
tion “Keyboard Suite,” freely available at http://www.trialdesign.org.
The R script for implementing the mTPI design can be downloaded
from the MD Anderson Cancer Center Software Download Website
https://biostatistics.mdanderson.org/softwaredownload/.

Case study
Solid Tumor Dose Finding Trial The objective of this phase I trial
(ClinicalTrials.gov Identifier: NCT03725436) was to determine MTD for the
Model-Assisted Dose Finding Designs 59

FIGURE 3.8: The decision tree of the BOIN Suite software to assist users to
choose an appropriate BOIN design module.

MDM2/MDMX inhibitor ALRN-6924 in combination with paclitaxel in adult

patients with advanced or metastatic solid tumors with wild-type (WT) TP53.
The trial employed the BOIN design to find MTD. Five doses of ALRN-6924
were investigated with the target DLT probability φ = 25%.

We use this trial as an example to illustrate the use of the BOIN software,
and provide guidance to address some common design issues in phase I trials.
The Keyboard Suite web application shares a similar user interface as the
BOIN Suite, thus we here focus on the latter. After selecting and launching
the “BOIN/iBOIN” module from the BOIN Suite launchpad, we design the
trial using the following three steps:

Step 1: Enter trial parameters

Doses and Sample Size As shown in Figure 3.9, the number of doses under
investigation is five; the starting dose level is one. For trials where the lowest
60 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 3.9: Specify doses, sample size, and convergence stopping rule.

dose is believed to be safe, starting from a slightly higher dose level (e.g.,
two) may reduce the sample size as it allows the design to reach MTD sooner.
However, due to limited knowledge on the safety of the new drug, in general
it is not recommended to start from a high dose level (e.g., four).
The cohort size for the trial is three and the number of cohorts is 10, with
a total sample size of 30. As a rule of thumb, we recommend the maximum
sample size N = 6 × J (i.e., the maximum sample size of the 3+3 design) as
the total sample size, where J is the number of doses. This sample size gener-
ally yields reasonable operating characteristics (e.g., 50–70% correct selection
percentage of the true MTD). To reduce the sample size, it is often useful
to use the “convergence” stopping rule: stop the trial early when m patients
have been assigned to a dose and the decision is to stay at that dose. This
stopping criterion suggests that the dose finding approximately converges to
MTD, thus the trial can be stopped. We recommend m = 9 or larger. In this
trial, m = 12 is used. Because of the early stopping rule, the actual sample
size used in the trial is often smaller than the prespecified maximum sample
size N . The saving depends on the true dose–toxicity scenario and can be
evaluated using simulation. Usually, the savings in the sample size is more
prominent when the true MTD is near the starting dose.
The choice of the cohort size should be based on considerations of de-
sign performance and logistic complexity. Given a fixed total sample size, the
use of large cohort sizes will result in fewer cohorts. This reduces the logisti-
cal burden and requires fewer dose escalation/de-escalation decisions for trial
conduct. As a tradeoff, using a large cohort size often reduces the accuracy
of identifying the MTD and evaluating trial safety, because trials based on
a smaller number of cohorts tend to be less adaptive. For example, given a
Model-Assisted Dose Finding Designs 61

total sample size of 30 patients, using a cohort size of six patients results in
a total of five cohorts. This means that during the trial, we only have four
chances to escalate/de-escalate the dose. If MTD is dose level five or six, and
if the dose escalation starts from dose level one and no dose skipping is per-
mitted, then there is a high likelihood that we will not be able to reach MTD
because many patients are treated in lower doses. In addition, using a co-
hort size of six may expose up to six patients to an overly toxic dose before
dose de-escalation is made. In contrast, using small cohort sizes, such as one
or two patients per cohort, renders the trial more freedom to move up and
down between doses and be more responsive/adaptive to the observed data.
This is logistically more complicated, however, because both data and dose
escalation/de-escalation decisions need to be updated more frequently. In ad-
dition, it prolongs the trial duration as new patients cannot be enrolled until
the patient in the previous cohort has completed their DLT assessment. As a
result, the most commonly-used cohort size in practice is two to four patients
per cohort.

Target and Accelerated Titration As shown in Figure 3.10, the target DLT
probability is φ = 0.25, which should be elicited from clinicians. The target
φ can be adjusted if a specific safety requirement is desirable. For example,
if it is desirable to de-escalate the dose when the DLT rate is > 30%, then
φ = 25% is an appropriate target with the de-escalation boundary λd =
0.298. As described previously, such simple mapping between the target and
escalation/de-escalation boundaries is a unique and important advantage of
BOIN.
Given φ, although the software allows users to specify φ1 and φ2 (see Sec-
tions 3.4.2 and 3.4.3 for their definition and interpretation) by unchecking the

FIGURE 3.10: Specify the target, accelerated titration, and 3+3 design run-in
(available only when the target φ = 0.25).
62 Model-Assisted Bayesian Designs for Dose Finding and Optimization

box “X  use the default alternatives to minimize decision error” under the tar-
get toxicity probability field, we highly recommend using the default values
(i.e., φ1 = 0.6φ and φ2 = 1.4φ) provided by the software. These default values
have been shown to produce highly robust and desirable operating character-
istics. Nevertheless, when necessary, the values of φ1 and φ2 can be calibrated
to satisfy certain trial design goals. For example, if we do not want to modify
φ and prefer a more conservative design, we may set φ2 = 1.2φ to obtain a
lower de-escalation boundary λd . To this end, it is important to distinguish
(φ1 , φ2 ) from (λe , λd ), where φ1 and φ2 are the toxicity probabilities used to
minimize the decision errors, and λe and λd are the decision boundaries actu-
ally used to determine dose escalation and de-escalation. A practical way to
judge if (φ1 , φ2 ) specified by users is appropriate or not is to examine whether
the resulting (λe , λd ) makes practical and clinical sense. For example, given
φ = 0.25, if we set (φ1 , φ2 ) = (0.9φ, 1.1φ), the resulting (λe , λd )=(0.237, 0.262)
makes little sense because λe and λd are too close and nearly indistinguishable
under small sample sizes. The fundamental issue here is that it is not mean-
ingful to minimize the decision error for the hypotheses (i.e., φ vs. φ1 = 0.9φ
vs. φ2 = 1.1φ) given that we have no power to distinguish. See Section 3.4.3
for more discussion.
In addition to the “convergence” stopping rule described above, another
useful approach offered by the BOIN software to reducing the sample size is to
conduct the accelerated titration (Simon et al., 1997) before treating patients
in cohorts of three (Figure 3.10). During the accelerated titration, we treat
patients in cohorts of one, and we continue escalating the dose in the one-
patient-per-dose-level fashion until any of the following events are observed:
(i) the first instance of DLT, (ii) the second instance of moderate (grade 2)
toxicity, or (iii) the highest dose level is reached. At that point, the titration
ends. We add two more patients to the current dose level, and hereafter switch
to the cohort size of three.
In the simulation, however, the software considers only (i) and (iii), ignor-
ing (ii), due to two practical considerations. First, incorporating (ii) in the
simulation requires users to specify the probability of grade 2 toxicity at each
dose level, which is cumbersome and brings substantial noise and uncertainty.
Second, the main concern of using the accelerated titration is that it may
make the design risky. The design ignoring (ii) is more aggressive than its
counterpart considering (ii), thus if the operating characteristics of the former
is satisfactory, the trial with (ii) is safer.
As the accelerated titration generally leads to more aggressive dose esca-
lation, it should be used only when there is sufficient evidence that low doses
are most likely underdosing. When the accelerated titration is used, based on
our experience, a reasonable rule of thumb for the maximum sample size is
N = j ∗ − 1 + 6(J − j ∗ + 1), where j ∗ is the dose level where the first DLT
is expected to occur. For example, if we expect that the first three doses are
very safe and the first DLT may occur in dose level j ∗ = 4, then the maximum
required sample size N could be reduced to 15 when there are five dose levels.
Model-Assisted Dose Finding Designs 63

When the target φ = 0.25, the software provides an option “Apply the
3+3 design run-in” to embed the 3+3 design rule into the BOIN design. The
rationale for this option is that when φ = 0.25, the default BOIN de-escalation
boundary is λd = 0.298, which means de-escalating the dose when 1/3 or
2/6 patients have DLT. Due to the influence of the conventional 3+3 design,
in some cases, investigators prefer that the design stays at the current dose
when 1/3 patients have DLT. The 3+3 run-in option enables that and en-
forces the design to stay at the current dose when 1/3 DLT. This option is
added mainly based on practical consideration, not statistical consideration.
Actually, the 3+3 design rule—stay when 1/3 DLT, but de-escalate when 2/6
DLT—is not self-consistent, see Section 3.4.4 for the explanation of why using
the same fixed de-escalation boundary is optimal. Nevertheless, as the modifi-
cation only occurs when the cumulative number of patients is three, activating
the option generally has minor impact on the design operating characteristics.

Overdose Control This panel (see Figure 3.11) specifies the overdose con-
trol rule, described in Section 3.4.1. That is, if Pr(πj > φ | Dj ) > 0.95 and
nj ≥ 3, dose level j and higher are eliminated from the trial, and the trial is
terminated if the lowest dose level is eliminated. When the trial is terminated
due to toxicity, no dose should be selected as MTD. In general, we recommend
using the default probability cutoff 0.95. A smaller value (e.g., 0.9), results in
stronger overdose control, but at the cost of reducing the probability of cor-
rectly identifying MTD. This is because, in order to correctly identify MTD,
it is imperative to explore the doses sufficiently to learn their toxicity profile.
In some trial settings, under the null case that all doses are overly toxic
(i.e., the lowest/first dose is above the MTD), the probability of early trial
termination may not be as high as we desire (e.g., > 70%). That is simply

FIGURE 3.11: Specify the overdose control rules.

64 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 3.12: Adjust the overdose control cutoff.

because the small sample size cannot provide enough power to distinguish
whether a dose is overly toxic or not (e.g., 40% vs. 30%). To achieve a high
early termination probability (when the first dose is overly toxic), we can
activate the first option “ Check to impose a more stringent safety stopping
rule on the lowest dose” (see Figure 3.12). This option makes the lowest/first
dose more likely to be eliminated by lowering the probability cutoff by δ. The
default value δ = 0.05 is generally recommended and produces a good balance
between the safety and the accuracy to identify MTD. A large value of δ (e.g.,
0.1) increases the early termination probability (when all doses are toxic), but
at the cost of reducing the probability of correctly identifying MTD when it
is the lowest dose.
At the end of the trial, the BOIN design selects MTD as the dose whose
isotonic estimate of the DLT probability is closet to φ. In some cases, it may
be desirable to require that the DLT probability estimate of MTD be lower
than the de-escalation boundary λd . This can be done by activating the option
at the bottom of the “Overdose Control” panel (i.e., by checking the box “
Check to ensure p̂M T D < de-escalation boundary.”)
After completing the specification of trial parameters, the decision table
will be generated by clicking the “Get Decision Table” button. The decision
table automatically will be included in the protocol template in Step 3, but it
can also be saved as a separate csv, Excel, or pdf file in this step if needed.

Step 2: Run simulation

Operating Characteristics This step generates the operating characteristics
of the design through simulation, see Figure 3.13. The scenarios used for sim-
ulation should cover various possible clinical scenarios (e.g., MTD is located
Model-Assisted Dose Finding Designs 65

FIGURE 3.13: Simulate operating characteristics of the design.

at different dose levels). To facilitate the generation of the operating charac-

teristics of the design, the software automatically provides a set of randomly
generated scenarios with various MTD locations, which are often adequate for
most trials. Depending on the application, users can add or remove scenarios.
The software also provides an option to include the 3+3 design as a com-
parator to facilitate the communication with clinicians who are more familiar
with the conventional design. Figure 3.14 shows the operating characteristics
of BOIN for the solid tumor trial. The simulation results will be automatically
66 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 3.14: Operating characteristics of the BOIN design for the solid
tumor trial.

included as a table in the protocol template in the next step, but can also be
saved as a separate csv or Excel file if needed.

Step 3: Generate protocol template

Protocol Preparation The BOIN software generates sample texts and a
protocol template (including the simulation results in Step 2) to facilitate the
protocol write-up. The protocol template can be downloaded in various for-
mats (see Figure 3.15). Use of this module requires the completion of Steps 1
and 2. Once the protocol is approved by regulatory bodies (e.g., Institutional
Review Board), we follow the design decision table to conduct the trial and
Model-Assisted Dose Finding Designs 67

FIGURE 3.15: Download protocol templates.

make adaptive decisions (e.g., dose escalation/stay/de-escalation). When the

trial completes, use the “Select MTD” module to select MTD. The software
outputs the recommended MTD, the posterior estimate of DLT probability
at each dose, and a 95% credible interval (see Figure 3.16).

FIGURE 3.16: Estimate and identify MTD at the completion of the solid
tumor trial based on a hypothetical dataset.
4
Drug-Combination Trials

4.1 Introduction
Drug combination therapy provides an effective approach to improving treat-
ment efficacy and overcoming most cancers’ resistance to monotherapy. The
objectives of using drug combinations are to induce an additive or a syner-
gistic treatment effect, increase the joint dose intensity with non-overlapping
toxicities, and target various tumor cell susceptibilities and disease pathways.
Despite the enormous importance of combination therapies, statistical designs
currently used for dose finding in phase I trials of combination therapies are
grossly inefficient and rudimentary—most combination trials have used the
conventional 3+3 design (Riviere et al., 2015a). The objective of this chapter
is to address the challenges and clarify misconceptions in designing combina-
tion therapy trials and introduce more efficient designs, in particular model-
assisted designs, for dose finding in phase I drug-combination trials.
In general, drug combinations may involve one of the followings: two or
more previously marketed drugs or biologics, two or more new molecular en-
tities, or a mix of previously marketed drugs or biologics and new molecular
entities. According to the US Food and Drug Administration (FDA) guide-
lines (FDA, 2006, 2013), prior to testing a new drug combination in human
beings, extensive preclinical studies are required to demonstrate the biologi-
cal rationale for the combination and to assess the safety of the combination
(FDA, 2006). When such data are not available or indicate safety concerns
for the combination, additional toxicology studies are required to address the
concerns. Sometimes drug-combination trials may involve two or more new in-
vestigational drugs that have not been previously studied for any indication.
In such cases, additional considerations are needed for the co-development of
the new investigational drugs for use in combination (FDA, 2013). In drug-
combination trials, it is useful to test multiple doses of each drug to identify the
optimal dose combination in terms of risks and benefits (FDA, 2013). Com-
pared to single-agent trials, drug-combination trials have a higher dimension
for the dose searching space, leading to several unique challenges.
The major challenge in designing combination trials is that the dose com-
binations under investigation are only partially ordered by the dose-limiting
toxicity (DLT) probability). This is in contrast with monotherapy trials, for
which the doses under investigation are fully ordered by the DLT probability

DOI: 10.1201/9780429052781-4 69
70 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 4.1: Partial ordering (left) and toxicity contours (right) for drug
combinations.

(i.e., the higher the dose, the greater the DLT probability). Consider a trial
combining J doses of agent A, denoted as A1 < A2 < · · · < AJ , and K doses of
agent B, denoted as B1 < B2 < · · · < BK . Let Aj Bk denote the combination
of Aj and Bk . Often it is reasonable to assume that when the dose of agent
A is held constant, the DLT probability for the combination increases in the
dose of agent B, and vice versa. As shown in the left panel of Figure 4.1, the
rows and columns of the dose combination matrix are partially ordered, with
the DLT probability increasing in the dose of the corresponding agent when
the dose of the other agent is fixed. However, in other directions of the dose
matrix (e.g., along the diagonals from the upper left corner to the lower right
corner), the ordering is not clear due to unknown drug-drug interactions. For
example, a priori we do not know whether the DLT probability for A2 B2 is
higher than the DLT probability for A1 B3 or A3 B1 .
The partial ordering of dose combinations has several implications. First,
monotherapy designs for finding the maximum tolerated dose (MTD), de-
scribed in previous chapters, cannot directly be used for finding MTD of a
drug combination. The second important implication is that there is not just
a single MTD. Rather, as depicted in the right panel of Figure 4.1, there is an
MTD contour in the two-dimensional dose space. Therefore, multiple MTDs
may exist in the J ×K dose matrix. When designing a drug-combination trial,
one must decide whether to look for a single MTD or multiple MTDs. In some
settings, it can be advantageous to find multiple MTDs so that we can further
study which one yields the highest synergistic treatment effect. We begin this
Drug-Combination Trials 71

chapter with designs that look for a single MTD, and we finish with designs
that target multiple MTDs.
Before describing the designs, we establish some notation. Let πjk denote
the DLT probability for dose combination Aj Bk , and let φ denote the target
DLT probability for MTD. We use njk to denote the number of patients who
have been assigned to Aj Bk , and yjk to denote the number of DLTs observed
at Aj Bk , j = 1, . . . , J and k = 1, . . . , K. Therefore, at a particular point
during the trial, the observed data are D = {Djk , j = 1, . . . , J, k = 1, . . . , K},
where Djk = (njk , yjk ) are the data observed at Aj Bk .

4.2 Model-based designs

Numerous model-based designs have been proposed to find a single MTD for
combination trials. Thall et al. (2003) proposed a Bayesian drug-combination
dose-finding method based on a six-parameter model, assuming that doses
are continuous and can be freely changed during the trial. Yin and Yuan
(2009a) and Yin and Yuan (2009c) proposed Bayesian dose-finding designs
based on a copula-type model and latent contingency tables, respectively.
Braun and Wang (2010) developed a dose-finding method based on a Bayesian
hierarchical model. Wages et al. (2011) proposed the partial ordering contin-
uous reassessment method (POCRM) by reducing the two-dimensional dose-
searching space to a one-dimensional searching line based on partial ordering
of the dose combinations of two drugs. Braun and Jia (2013) generalized CRM
to handle drug-combination trials. Riviere et al. (2014) proposed a Bayesian
dose-finding design based on the logistic model. Cai et al. (2014) and Riv-
iere et al. (2015b) proposed Bayesian adaptive designs for drug-combination
trials involving molecularly targeted agents. Neuenschwander et al. (2015)
proposed a drug-combination design based on the Bayesian logistic regression
model (BLRM) with an overdose control rule. Albeit different, most of these
designs adopt a common dose-finding strategy:
1. Devise a model to describe the dose–toxicity surface.
2. Based on the accumulating data, continuously update the model
estimate, and make the decision of dose assignment for the incoming
new patient, typically by assigning the new patient to the dose
whose estimated DLT probability is closest to the target φ.
3. Stop the trial when the maximum sample size is reached, and select
the MTD based on the estimates of the DLT probabilities.
In what follows, we use the copula-type design (Yin and Yuan, 2009a)
as an example to illustrate the core dose-finding strategy of the model-based
designs. Let pj be the prespecified toxicity probability corresponding to dose
72 Model-Assisted Bayesian Designs for Dose Finding and Optimization

Aj of agent A, p1 < · · · < pJ , and qk be that of dose Bk of agent B, q1 <

· · · < qK . Before two drugs are combined, typically they have been individually
studied in phase I trials. Thus, the values of pj ’s and qk ’s can be elicited from
clinicians, based on the historical data from previous single-agent trials. The
DLT probability of Aj Bk can be modeled using the copula-type regression
model (Yin and Yuan, 2009a; Yin and Lin, 2015) as follows:

πjk = 1 − {(1 − pα
j)
−γ
+ (1 − qkβ )−γ − 1}−1/γ , (4.1)

for j = 1, . . . , J and k = 1, . . . , K, where α, β, γ > 0 are unknown model

parameters. An attractive feature of model (4.1) is that if only one drug is
tested, it reduces to the power model used in CRM. For example, if there is
only drug A (i.e., qk = 0), the model reduces to

πj = p α
j, j = 1, . . . , J.

Therefore, the copula-type regression model can be viewed as a generalization

of the single-agent CRM to drug-combination trials. This unique feature al-
lows us to easily incorporate prior information of single-agent trials into the
combination trials by specifying pj ’s and qk ’s, which are known as “skeletons”
in the CRM.
Alternatively, one can also use the logistic regression model for πjk (Riviere
et al., 2014),
logit(πjk ) = β0 + β1 uj + β2 vk + β3 uj vk , (4.2)
where β0 , β1 , β2 , and β3 are unknown parameters, with β1 > 0 and β2 > 0 to
ensure that the toxicity probability is increasing with the increasing dose level
of each agent alone, and for all k, β1 + β3 vk > 0 and for all j, β2 + β3 uj > 0
to ensure that the toxicity probability is increasing with the increasing dose
levels of both agents together. The covariates uj and vk are the standardized
doses of Aj and Bk , respectively, defined as uj = log {pj /(1 − pj )} and vk =
log {qk /(1 − qk )}, to incorporate the prior information on single-agent toxicity
probabilities.
Denote θ as the set of unknown parameters of the dose–toxicity model.
For example, θ = (α, β, γ)T in the copula-type regression model, and θ =
(β0 , β1 , β2 , β3 )T in the logistic regression model. Suppose that at a certain
stage of the trial, among njk patients treated at Aj Bk , yjk subjects have
experienced DLT. The likelihood given the observed data D is
J Y
K
y
Y
L(D | θ) ∝ πjkjk (1 − πjk )njk −yjk .
j=1 k=1

The posterior distribution of θ is given by

f (θ | D) ∝ L(D | θ)f (θ),

Drug-Combination Trials 73

where f (θ) is the prior distribution of θ. For example, f (θ) = f (α)f (β)f (γ)
in the copula-type regression model, where f (α), f (β), and f (γ) denote inde-
pendent, vague gamma prior distributions with mean one and large variances
for α, β, and γ, respectively.
Based on the posterior distribution of the model parameters, the dose-
finding algorithm can be described as follows:
1. The first cohort of patients is treated at the lowest dose combination
A1 B1 .
2. Suppose the current dose combination is Aj Bk , to determine the
dose for the next cohort, consider the following:
(i) If Pr(πjk < φ|D) > ce , where ce is the fixed probability cut-
off for dose escalation, the dose for the next cohort of pa-
tients moves to an adjacent dose combination chosen from
{Aj+1 Bk , Aj+1 Bk−1 , Aj−1 Bk+1 , Aj Bk+1 }, which has a DLT
probability higher than the current doses and closest to φ. If
the current dose combination is AJ BK , the dose stays at the
same dose combination.
(ii) If Pr(πjk < φ|D) < cd , where cd is the fixed probability
cutoff for dose de-escalation, the dose moves to an adjacent
dose combination chosen from {Aj−1 Bk , Aj−1 Bk+1 , Aj+1 Bk−1 ,
Aj Bk−1 }, which has a DLT probability lower than the current
doses and closest to φ. If the current dose combination is A1 B1 ,
the trial is terminated.
(iii) Otherwise, the next cohort of patients continues to be treated
at the current dose combination.
3. Repeat Step 2 until the maximum sample size is reached, and select
MTD as the dose whose estimate of DLT probability is closest to
φ.
The model-based designs perform reasonably well, but for several reasons
they are rarely used in practice. First, these designs are statistically and com-
putationally complicated, leading many practitioners to perceive that deci-
sions of dose allocation arise from a “black box,” which limits its application
in practice. Secondly, robustness is another potential issue for the model-based
drug-combination designs. Since these designs use a strategy akin to CRM, one
might expect them to share the similar robustness (e.g., consistent under mis-
specified models (Shen and O’Quigley, 1996)). Unfortunately, that generally
is not the case. The consistency of CRM under misspecified models requires
several assumptions (Shen and O’Quigley, 1996). A critical one is monotonic-
ity (i.e., the DLT probability monotonically increases with the dose), which
does not hold for drug combinations. Based on our experience, model-based
drug-combination trial designs are substantially more delicate, and it is not
difficult to find scenarios where such designs do not perform well.
74 Model-Assisted Bayesian Designs for Dose Finding and Optimization

4.3 Model-assisted designs

Model-assisted designs have emerged as an attractive approach for phase I
drug-combination trials. These designs yield competitive performance similar
to that of model-based designs, but are much simpler to implement and more
robust without assuming any parametric model on the dose–toxicity surface.
This section introduces two model-assisted designs: the BOIN combination
design and the keyboard combination design.

4.3.1 BOIN combination design

The BOIN combination design (Lin and Yin, 2017a) uses the same dose
escalation/de-escalation rule as the BOIN single-agent design, described in
Section 3.4. Let π̂jk = yjk /njk denote the observed DLT rate at Aj Bk , and λe
and λd denote the optimal dose escalation and de-escalation boundaries, re-
spectively, as defined in Section 3.4. The BOIN dose escalation/de-escalation
rule is: if π̂jk ≤ λe , escalate the dose; if π̂jk > λd , de-escalate the dose; other-
wise, retain the current dose.
For combination trials, the new challenge is that there is more than one
option for dose escalation: we can escalate the dose of A or the dose of B.
Similarly, when we decide to de-escalate the dose, we can de-escalate the
dose of A or the dose of B. The BOIN combination design uses the Bayesian
posterior probability of πjk ∈ (λe , λd ) to make the decision. As the posterior
distribution of πjk (i.e., beta distribution) is continuous, πjk ∈ (λe , λd ) =
πjk ∈ (λe , λd ] and thus we use the former throughout. Specifically, define
admissible dose escalation and de-escalation sets as

AE = {Aj+1 Bk , Aj Bk+1 } and AD = {Aj−1 Bk , Aj Bk−1 }.

When the BOIN rule says escalate, we escalate to the dose combination that
belongs to AE and has the highest value of Pr(πjk ∈ (λe , λd )|Djk ); and when
the BOIN rule says de-escalate, we de-escalate to the dose combination that
belongs to AD and has the highest value of Pr(πjk ∈ (λe , λd )|Djk ). That is,
we always move toward the dose that is most likely to be in the acceptable
(or “stay”) interval (λe , λd ). The value of Pr(πjk ∈ (λe , λd )|Djk ) can be easily
evaluated based on the beta-binomial model

yjk | njk , πjk ∼ Binomial(njk , πjk ) (4.3)

πjk ∼ Beta(1, 1) ≡ Unif(0, 1)

with the posterior πjk | Djk ∼ Beta(yjk + 1, njk − yjk + 1). The BOIN combi-
nation design is summarized in Table 4.1.
Because Pr{πjk ∈ (λe , λd )|Djk } can be pre-determined for all possible
outcomes Djk = (njk , yjk ), the dose escalation and de-escalation rule in Step
Drug-Combination Trials 75

TABLE 4.1: BOIN drug-combination design.

(a) Patients in the first cohort are treated at the lowest dose combina-
tion A1 B1 or a prespecified dose combination.
(b) Suppose the current cohort is treated at dose combination Aj Bk ;
to assign a dose to the next cohort of patients:
• If π̂jk ≤ λe , we escalate the dose to the combination that
belongs to AE and has the largest value of Pr{πj 0 k0 ∈
(λe , λd )|Dj 0 k0 }. If the current dose combination is AJ BK , then
we retain this dose for treating the next cohort of patients.
• If π̂jk > λd , we de-escalate the dose to the combination
that belongs to AD and has the largest value of Pr{πj 0 k0 ∈
(λe , λd )|Dj 0 k0 }. If the current dose combination is A1 B1 , then
we retain this dose for treating the next cohort of patients.
• Otherwise, if λe < π̂jk ≤ λd , then the dose stays at the same
combination Aj Bk .
(c) Repeat Step (b) until the maximum sample size N is reached, and
select MTD as the dose combination whose isotonic estimate (Bril
et al., 1984b) of the DLT probability is closest to φ.

(b) can be tabulated before the trial begins, which makes the BOIN combi-
nation design easy to implement in practice. One important characteristic of
the dose transition rule of the BOIN combination design is that, to make the
decision, what really is needed is the ordering of Pr{πjk ∈ (λe , λd )|Djk } for
doses within AE and AD , not their absolute values. Thus, for the purpose
of decision making, we only need to tabulate the rank of each possible Djk
according to the value of Pr{πjk ∈ (λe , λd )|Djk }. This greatly simplifies the
decision table. We refer to the rank of a dose with Djk as the desirability score
of that dose. Table 4.2 shows the desirability score for njk up to 12, with the
cohort size of 3 and the target DLT probability φ = 0.3. A larger value in-
dicates a more desirable dose with a larger value of Pr{πjk ∈ (λe , λd )|Djk }.
To conduct the trial, there is no need for any model fitting or complicated
calculation (as required by model-based designs). Users simply look up the
desirability score table to make the dose escalation/de-escalation decision.
To illustrate the use of the decision table, suppose that at the current dose
A1 B1 , we observed π̂11 = 1/6 < λe = 0.236 and thus need to escalate the
dose. Assume that at this point, the observed data at A2 B1 and A1 B2 are
D21 = (0, 0) and D12 = (3, 1), respectively. To determine which dose the trial
should be escalated to, we simply look up Table 4.2 and identify that the
desirability scores of A2 B1 and A1 B2 are 25 and 40, respectively. As A1 B2
76 Model-Assisted Bayesian Designs for Dose Finding and Optimization

TABLE 4.2: Desirability score table for the BOIN combination design with
the target φ = 0.3. A larger value indicates a higher desirability.

No. No. Des. No. No. Des. No. No. Des.

Pts DLTs Score Pts DLTs Score Pts DLTs Score
0 0 25 6 3 34 12 0 11
3 0 28 6 ≥4 E 12 1 22
3 1 40 9 0 14 12 2 45
3 2 24 9 1 32 12 3 61
3 ≥3 E 9 2 53 12 4 62
6 0 19 9 3 57 12 5 48
6 1 42 9 4 41 12 6 30
6 2 49 9 ≥5 E 12 ≥7 E
Note: “No. Pts” is the total number of patients treated, “No. DLTs” is the
number of patients who experienced DLT, “Des. Score” is the desirability
score. “E” indicates that the dose level is too toxic and should be eliminated
from the trial. If both dose levels should be eliminated, then the trial should
stay at the current dose level.

has a higher desirability score, the next cohort of patients will be treated at
A1 B2 . In the case that A2 B1 and A1 B2 have the same desirability scores (e.g.,
both doses have not yet been used to treat any patients, or have been used to
treat patients and generated same data), we can choose one dose randomly or
based on other clinical considerations. The BOIN combination decision table,
such as Table 4.2, can be generated using the software described later, and
included in the trial protocol for trial conduct.
Like the BOIN single-agent design, during the trial conduct the BOIN
combination design imposes the dose elimination/early stopping rule such
that: if Pr(πjk > φ | Djk ) > 0.95 and njk ≥ 3, dose combination Aj Bk and
the higher combinations (i.e., {Aj 0 Bk0 , j 0 ≥ j, k 0 ≥ k}), are eliminated from the
trial, and the trial is terminated if the lowest dose combination is eliminated,
where Pr(π11 > φ | D11 ) is evaluated based on the posterior distribution
(3.2). The letter “E” in Table 4.2 reflects this dose elimination rule such that
a dose with a desirability score of “E” is excessively toxic and should not be
considered in the admissible escalation/de-escalation set.
The BOIN combination design employs the same optimal escalation and
de-escalation boundaries as the BOIN single-agent design to guide the dose
transition, thus it inherits the latter’s desirable statistical and operational
properties. It is simple, transparent, easy-to-calibrate, and efficient to identify
MTD. In addition, because no parametric assumption is made on the dose–
toxicity surface, the BOIN combination design is more robust than model-
based designs. Extensive simulation studies have demonstrated that the BOIN
combination design yields competitive performance comparable to more com-
plicated model-based designs, see Section 4.4 for more details.
Drug-Combination Trials 77

4.3.2 Keyboard combination design

The keyboard combination design (Pan et al., 2020) provides another model-
assisted design for finding MTD for drug-combination trials. This design uses
the same dose escalation/de-escalation rule as the keyboard single-agent de-
sign, described in Section 3.3 (e.g., Figure 3.1). That is, let I ∗ = (δ1 , δ2 ) denote
the target key (i.e., the pre-specified proper dosing interval), and Imax denote
the strongest key (i.e., the interval with the highest posterior probability, given
Djk ). If Imax is located on the left side of I ∗ (denoted as Imax ≺ I ∗ ), escalate
the dose; if Imax is located on the right side of I ∗ (denoted as Imax  I ∗ ),
de-escalate the dose; if Imax is I ∗ (denoted Imax ≡ I ∗ ), retain the current
dose.
Again, for combination trials, the new challenge is that there are two op-
tions for dose escalation and de-escalation: escalate the dose of A or the dose
of B when Aj Bk is deemed an underdose, or de-escalate the dose of A or
the dose of B when Aj Bk is deemed excessively toxic. The keyboard com-
bination design adopts a similar strategy to the BOIN combination design,
using Pr{πjk ∈ I ∗ |Djk } to determine which drug’s dose should be esca-
lated or de-escalated. That is, when the dose escalation is needed, we es-
calate to the dose combination that belongs to AE and has the highest value
of Pr(πjk ∈ I ∗ |Djk ); and when the dose de-escalation is needed, we esca-
late to the dose combination that belongs to AD and has the highest value
of Pr(πjk ∈ I ∗ |Djk ). The keyboard combination design is summarized in
Table 4.3.

4.3.3 Waterfall design

The BOIN and keyboard combination designs described above focus on finding
a single MTD. For some combination trials, it is of interest to find multiple
MTDs (i.e., the MTD contour) from the dose matrix. These MTDs can be
further evaluated in subsequent phases of cohort expansion or phase II trials
to identify the optimal combination that generates the highest synergistic
treatment effect. This section introduces a model-assisted design, the waterfall
design (Zhang and Yuan, 2016), which can be used to find the MTD contour.
Finding the MTD contour is substantially more challenging than finding
a single MTD. This is because, in order to find all MTDs in the dose matrix,
we must explore the whole dose matrix using the limited sample size that
is a key characteristic of phase I trials. If we do not explore the whole dose
matrix, we risk missing some MTDs. To address this challenge, the waterfall
design employs the divide-and-conquer strategy: it partitions the dose matrix
into blocks, and then it applies the BOIN single-agent design to each of the
blocks to find all MTDs. Each block is known as a “subtrial,” and the doses
within a block must be fully ordered in toxicity. In other words, the divide-
and-conquer strategy reduces the two-dimensional searching space to several
78 Model-Assisted Bayesian Designs for Dose Finding and Optimization

TABLE 4.3: Keyboard drug-combination design.

(a) Patients in the first cohort are treated at the lowest dose combina-
tion A1 B1 or a prespecified dose combination.
(b) Suppose the current cohort is treated at dose combination Aj Bk ,
given the observed data Djk = (njk , yjk ), we identify the strongest
key Imax based on the posterior distribution of πjk and assign a
dose to the next cohort of patients as follows:

• If Imax ≺ I ∗ , we escalate the dose to the combination that

belongs to AE and has the largest value of Pr{πj 0 k0 ∈ I ∗ |Dj 0 k 0 }.
If the current dose combination is AJ BK , then we retain this
dose for treating the next cohort of patients.
• If Imax  I ∗ , we de-escalate the dose to the combination
that belongs to AD and has the largest value of Pr{πj 0 k0 ∈
I ∗ |Dj 0 k 0 }. If the current dose combination is A1 B1 , then we
retain this dose for treating the next cohort of patients.
• Otherwise, if Imax ≡ I ∗ , then the dose stays at the same com-
bination Aj Bk .
(c) Repeat Step (b) continued until the maximum sample size N is
reached, and select the MTD as the dose combination whose isotonic
estimate of the DLT probability is closest to φ.

one-dimensional searching lines (i.e., subtrials) such that the BOIN single-
agent design can be directly applied.
As illustrated in Figure 4.2, the waterfall design partitions the J × K
dose matrix into J subtrials (or blocks), within which the doses are fully
ordered. Without loss of generality, we assume that J ≤ K. These subtrials
are conducted sequentially from the top of the matrix to the bottom, which is
where the design gets its waterfall name. The goal of the waterfall design is to
find the MTD contour, which is equivalent to finding MTD in each row of the
dose matrix, if one exists. The waterfall design can be described as follows:
1. Divide the J ×K dose matrix into J subtrials SJ , · · · , S1 , according
to the dose level of drug A:

SJ = {A1 B1 , A2 B1 , · · · , AJ B1 , AJ B2 , · · · , AJ BK },
SJ−1 = {AJ−1 B2 , AJ−1 B3 , · · · , AJ−1 BK },
SJ−2 = {AJ−2 B2 , AJ−2 B3 , · · · , AJ−2 BK },
···
S1 = {A1 B2 , A1 B3 , · · · , A1 BK }.
Drug-Combination Trials 79

 ● ● ● ● ●
(a)  ● ● ● ● ●

 ● ● ● ● ●

 ● ● ● ● ●
(b)  ● ● ● ● ●

 ● ● ● ● ●
Drug A

 ● ● ● ● ●
(c)  ● ● ● ● ●

 ● ● ● ● ●

 ● ● ● ● ●
(d)  ● ● ● ● ●

 ● ● ● ● ●

    

Drug B

FIGURE 4.2: Illustration of the waterfall design for a 3 × 5 dose matrix. The
outlined doses in each panel form a subtrial, and the asterisk denotes the
candidate MTD. As shown in panel (a), the first subtrial starts with dose
combination A1 B1 . After the first subtrial identifies A3 B2 as the candidate
MTD, the second subtrial starts with dose combination A2 B3 (see panel (b)).
After the second subtrial identifies A2 B4 as the candidate MTD, the third
subtrial starts with dose combination A1 B5 (see panel (c)). After all subtrials
complete, the MTD in each row of the dose matrix is selected based on the
data from all subtrials, as shown in panel (d).

To ensure the trial starts at the lowest dose, subtrial SJ includes

lead-in dose combinations A1 B1 , A2 B1 , · · · , AJ B1 (i.e., the first col-
umn of the dose matrix, see panel (a) in Figure 4.2). Within each
subtrial, the dose combinations are fully ordered with monotonically
increasing DLT probabilities.
2. Conduct the subtrials sequentially using the BOIN design (or an-
other single-agent dose-finding method) as follows:
(i) Conduct subtrial SJ , starting from the lowest dose combina-
tion A1 B1 , to find the J-th candidate MTD. We call the dose
selected by a particular subtrial the “candidate MTD” to high-
light that it may not be selected as an MTD upon completion
of the trial, which will be based on the data collected from all
80 Model-Assisted Bayesian Designs for Dose Finding and Optimization

the subtrials. The candidate MTD is used to determine which

subtrial to conduct next and which combination to assign first.
(ii) If the current subtrial Sj , j = J, . . . , 2 selects Aj ∗ Bk∗ as the
candidate MTD, then conduct subtrial Sj ∗ −1 beginning with
Aj ∗ −1 Bk∗ +1 . That is, the next subtrial corresponds to the dose
of drug A one level below the candidate MTD from the previ-
ous subtrial. Upon completion of subtrial Sj ∗ −1 , and selection
of another candidate MTD, this rule is applied similarly to de-
termine the next subtrial and its starting dose (e.g., see Figure
4.2).
(iii) Repeat Step (ii) until subtrial S1 completes.
3. Estimate the DLT probabilities (i.e., πjk ), based on the data from
all the subtrials using the two-dimensional isotonic regression (Bril
et al., 1984b). For each row of the dose matrix, select MTD as the
dose combination with an estimated DLT probability closest to the
target, unless every dose combination in this row is too toxic.
In Step 2 (ii), the reason subtrial Sj ∗ −1 starts with dose Aj ∗ −1 Bk∗ +1 rather
than the lowest dose in this subtrial (i.e., Aj ∗ −1 B2 ) is that—assuming toxicity
monotonically increases in each dose and the candidate MTD in the previous
subtrial is an actual MTD—Aj ∗ −1 Bk∗ +1 is the lowest dose that could be MTD
in subtrial Sj ∗ −1 . Therefore, starting from Aj ∗ −1 Bk∗ +1 allows the design to
quickly find MTD in subtrial Sj ∗ −1 . Using Figure 4.2 as an example, dose
combination A3 B2 is the candidate MTD identified in the first subtrial S3 ,
and thus, the second subtrial S2 starts with A2 B3 . It is illogical for subtrial
S2 to start with the lowest dose A2 B2 , since the partial ordering implies that
A2 B2 is farther below the MTD contour. Starting S2 with A2 B2 in this exam-
ple would expose excessive numbers of patients to putatively subtherapeutic
dose combinations, thereby wasting the already limited patient resources. The
waterfall design is efficient because it uses the results from the current subtrial
to inform the design of subsequent subtrials.
Because the dose combinations in each subtrial are fully ordered, applying
the BOIN single-agent design to them is straightforward. The main challenge
is how to determine when the current subtrial should be terminated. One
simple approach is to prespecify a maximum sample size for each subtrial.
Once the maximum sample size is reached, we stop the subtrial, determine its
candidate MTD, and initiate the next subtrial. Although this approach works
well for single-agent dose-finding trials, it is inefficient for multiple subtrials.
This is because—depending on the number of dose combinations between the
starting dose combination and the actual MTD, as well as the shape of the
dose–toxicity curve—often each subtrial requires different numbers of patients
to identify MTD reliably. Based on this consideration, Zhang and Yuan (2016)
proposed and recommended the following “convergence” stopping rule for the
subtrials:
If the number of patients treated at the current dose combination reaches
or exceeds a prespecified number, say m, then stop the subtrial, select the
candidate MTD, and initiate the next subtrial.
Drug-Combination Trials 81

The rationale for the stopping rule is that when the patient allocation concen-
trates at a particular dose combination, this indicates that the dose-finding
algorithm likely has converged on the MTD, so we should stop the subtrial
and select the current dose combination as the candidate MTD. This stop-
ping rule automatically adjusts the sample size of each subtrial to reflect the
difficulty of the dose finding (e.g., the number of dose combinations between
the starting dose combination and the actual MTD, and the shape of the
dose–toxicity curve). In addition, this stopping rule ensures that a certain
number of patients are treated at the candidate MTD, which is achieved in
single-agent dose-finding designs using cohort expansion after selecting MTD.
Setting m ≥ 9 (m = 12 is preferred) usually ensures reasonable operating
characteristics.
Although the above stopping rule provides an automatic, reasonable way
to determine the sample size for a particular subtrial, in some cases, it is
advantageous to prespecify a maximum sample size for each subtrial as well.
This can be done by adding an extra stopping rule:
If the number of patients treated in subtrial Sj reaches or exceeds nmax
j ,
where nmax
j is the prespecified maximum sample size for the subtrial Sj,
then stop the subtrial, select the candidate MTD, and initiate the next
subtrial.
We recommended setting nmax j between 4×(the number of doses in subtrial
Sj ) and 6×(the number of doses in subtrial Sj ) for j = 1, . . . , J. For example,
a trial with a 3 × 5 dose matrix, like the trial depicted in Figure 4.2, consists
of a first subtrial with seven doses, and second and third subtrials each with
four doses. We may set nmaxj = 28, 16, and 16 for three subtrials, respectively,
and thus a maximum of 60 patients for the trial. Although a maximum sample
size of 60 patients may seem large, because there are 15 dose combinations,
60 patients actually is not a very large sample size. Consider a single-agent
dose-finding trial with 15 doses, the maximum sample size under the 3+3
design is 6 × 15 = 90 patients. We recommend using computer simulations
to calibrate m and nmaxj , thereby ensuring the design has desirable operating
characteristics. This simulation-based calibration can be carried out with the
software described in the next section.
Lastly, the partition of the dose matrix is not unique. Any partition can
be used as long as it is clinically sound and the doses within each block are
fully ordered. For example, we may use each row of the dose matrix as a
subtrial. In addition to the advantage of being simple and transparent, as
shown in the next section, the waterfall design also has competitive and often
better performance than the more complicated model-based designs, such as
the design based on the product of independent beta probabilities (PIPE)
(Mander and Sweeting, 2015).
82 Model-Assisted Bayesian Designs for Dose Finding and Optimization

TABLE 4.4: Average computation time (minutes) needed to simulate 2000

trials for each design to study drug combinations.

Combinations BLRM POCRM Copula BOIN PIPE Waterfall

2×4 192.7 12.2 140.0 0.06 2.4 0.9
3×5 456.7 16.2 302.0 0.07 9.5 0.9
4×4 498.6 16.9 352.1 0.08 11.8 1.0

4.4 Operating characteristics

Liu et al. (2022) performed a comprehensive simulation study to compare the
operating characteristics of model-based designs (including the copula design,
BLRM, and POCRM) with the model-assisted BOIN combination design.
The target DLT probability is φ = 25%. To avoid cherry-picking scenarios,
this study randomly generated 3000 scenarios for each of the 2 × 4, 3 × 5, and
4 × 4 dose combination matrices, and simulated 500 trials under each scenario
for each design. The sample size N = 27 for 2 × 4 dose combinations, and
N = 48 for 3 × 5, and 4 × 4 dose combinations, with the cohort size of three.
Detailed simulation configurations can be found in Liu et al. (2022).
Figure 4.3 shows the percentage of correct selection of MTD (i.e., a metric
for accuracy) and the percentage of patients treated at the toxic doses with
true DLT probabilities ≥ 33% (i.e., a metric for safety). Compared to the
model-based designs, the BOIN combination design yields the highest percent-
age of correct selection of MTD, and comparable or better safety. In addition,
because of the simplicity of the BOIN combination design, its computational
time is just a fraction of that of the model-based designs (see Table 4.4). More
importantly, it is significantly simpler to implement the BOIN combination
design in practice by using its decision table as described in Section 4.3.1.
Liu et al. (2022) also compared the waterfall design with the model-based
PIPE design for finding the MTD contour based on 3000 random scenarios of
the 2 × 4 combination trial with 36 patients, a 3 × 5 combination trial with
48 patients, and a 4 × 4 combination trial with 48 patients. The results (see
Figure 4.4) support the superior performance of the model-assisted design.
Drug-Combination Trials 83

Percentage of correct selection

100
(i) (ii) (iii)

80
60
Percentage

41.3 46.8 43.7

41.1 43.2 40.1 40.7
41.6
38.4
37.8
40

37.8
34.6
25.8
22.6
21.0
20
0

3+3

BLRM

POCRM

Copula

BOIN

3+3

BLRM

POCRM

Copula

BOIN

3+3

BLRM

POCRM

Copula

BOIN
Percentage of patients treated at overdoses
60

(i) (ii) (iii)

50
40
Percentage

30

23.1

19.5 19.3 19.9 18.8

20

19.8
15.5 19.9 20.1 18.6
15.1
13.7
13.4
12.2
12.3
10
0

3+3

BLRM

POCRM

Copula

BOIN

3+3

BLRM

POCRM

Copula

BOIN

3+3

BLRM

POCRM

Copula

BOIN

FIGURE 4.3: Simulation results of designs for finding one MTD based on 3000
random scenarios of (i) the 2×4 combination trial with 27 patients, (ii) the
3×5 combination trial with 48 patients, and (iii) the 4×4 combination trial
with 48 patients.
84 Model-Assisted Bayesian Designs for Dose Finding and Optimization

Percentage of selecting at least one MTD

(i) (ii) (iii)

100
80

64.7
63.2
62.0
Percentage

60

52.5 52.1
49.0
40
20
0

PIPE

Waterfall

PIPE

Waterfall

PIPE

Waterfall
Percentage of patients treated at overdoses

(i) (ii) (iii)

60
50
40
Percentage

30

26.6
25.4
20

16.3 15.9 16.3

14.9
10
0

PIPE

Waterfall

PIPE

Waterfall

PIPE

Waterfall

FIGURE 4.4: Simulation results of the PIPE and waterfall designs for finding
the MTD contour based on 3000 random scenarios of (i) the 2×4 combination
trial with 27 patients, (ii) the 3×5 combination trial with 48 patients, and
(iii) the 4×4 combination trial with 48 patients.
Drug-Combination Trials 85

4.5 Software and case study

Software
The keyboard combination design can be implemented using the web appli-
cation “Keyboard Suite,” freely available at http://www.trialdesign.org.
Software for the BOIN combination design is available in several differ-
ent forms: (1) the web application “BOIN Suite” with intuitive graphical
user interface, freely available at http://www.trialdesign.org (Figure 4.5
shows the launchpad of BOIN Suite); (2) the R package BOIN freely avail-
able at CRAN; and (3) the Windows desktop program “BOIN Suite,” freely
available at the MD Anderson Cancer Center Software Download Website
https://biostatistics.mdanderson.org/softwaredownload/. We recom-
mend the web application as it is most frequently updated.

Case study
Acute Myeloid Leukemia (AML) Trial The objective of this phase I
trial (ClinicalTrials.gov Identifier: NCT03600155) was to determine MTD for
nivolumab and ipilimumab alone and in combination in patients with high risk
or refractory/relapsed AML and myelodysplastic syndrome (MDS) following
allogeneic stem cell transplantation. The trial consisted of three parallel arms:
nivolumab alone, ipilimumab alone, and the combination of nivolumab and ip-
ilimumab. We here focus on the combination arm, which employed the BOIN
combination design to find MTD. Three doses of ipilimumab and two doses
of nivolumab were investigated (i.e., six combinations) with the target DLT
probability φ = 30%.

We use this trial as an example to illustrate the use of the BOIN web ap-
plication to design drug-combination trials. After selecting and launching the
“BOIN Comb” module from the BOIN Suite launchpad (see Figure 4.5), the
trial can be designed using the following three steps:
Step 1: Enter trial parameters
Doses and Target As shown in Figure 4.6, drug A (nivolumab) has two
doses and drug B (ipilimumab) has three doses. The starting dose is A1 B1 .
For trials where the lowest dose is believed to be safe, starting from a slightly
higher dose level (e.g., A1 B2 ) may save the sample size as it allows the design
to reach the MTD sooner.
The target DLT probability is φ = 0.3, which should be elicited from clini-
cians. The target φ can be adjusted if a specific safety requirement is desirable,
see Section 3.6 for more guidance on the specification of φ1 and φ2 . Depending
on the trial objective, users can select to find a single MTD (using the BOIN
combination design) or the MTD contour (using the waterfall design). In gen-
eral, finding the MTD contour requires a larger sample size than find a single
MTD. This is because finding the MTD contour requires a more thorough
86 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 4.5: Launchpad of BOIN Suite.

exploration of the dose matrix. When target φ = 0.25, the software provides
an option called “Apply the 3+3 design run-in” to embed the 3+3 design rule
into the BOIN decision table, see Section 3.6 for details.

Sample Size and Cohort Size The cohort size for the trial is three and
the number of cohorts is 10, with a total sample size of 30 (Figure 4.7). As
a rule of thumb, for a J × K dose matrix, we recommend the maximum
sample size N ∈ [4 × J × K, 6 × J × K], where 6 × J × K corresponds to the
maximum sample size of the 3+3 design for J × K doses. This sample size
generally yields reasonable operating characteristics. In the AML trial, J = 2
and K = 3, thus, the recommended sample size is 24 to 36. To reduce the
actual sample size, it is often useful to use the “convergence” stopping rule:
stop the trial early when m patients have been assigned to a dose and the
decision is to stay at that dose. This stopping criterion suggests that the dose
finding approximately converges to the MTD, thus the trial can be stopped.
We recommend m = 9 or larger. In this trial, m = 12 is used. Because of the
early stopping rule, the actual sample size used in the trial is often smaller
than N . The saving depends on the true dose–toxicity relationship and can
be evaluated using simulation.
Another useful approach to reducing the sample size is to conduct the ac-
celerated titration before treating patients in cohorts of three (i.e., checking
Drug-Combination Trials 87

FIGURE 4.6: Specify doses, target, and accelerated titration.

“Yes” for “Perform accelerated titration” in Figure 4.6). During the acceler-
ated titration, we treat patients in cohorts of one, and we continue escalating
the dose in the one-patient-per-dose-level fashion until any of the following
events is observed: (i) the first instance of DLT, (ii) the second instance of

FIGURE 4.7: Specify the sample size, cohort size, and convergence stopping
rule.
88 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 4.8: Specify the sample size of subtrials for the waterfall design.

moderate (grade 2) toxicity, or (iii) the highest dose level is reached. At that
point, the titration ends. We add two more patients to the current dose level,
and hereafter switch to the cohort size of three. During the titration, the drug
to be escalated (A or B) can be randomly selected, which is adopted by the
software for simulation, or chosen based on clinical consideration. In addition,
as described in Section 3.6, in the simulation, the software considers only (i)
and (iii), ignoring (ii), due to practical considerations described previously.
As the accelerated titration leads to more aggressive dose escalation, it
should be used only when there is sufficient evidence that low doses are most
likely underdosing. When the accelerated titration is used, the sample size can
be chosen using simulation to obtain desirable operating characteristics.
In the case that the trial objective is to find the MTD contour, we need
to specify the maximum sample size nmax j for each subtrial (Figure 4.8). As
discussed in Section 4.3.3, we recommend setting nmax j between 4×(the num-
ber of doses in subtrial Sj ) and 6×(the number of doses in subtrial Sj ), for
j = 1, . . . , J. For this trial with a 2 × 3 dose matrix, the first and second
subtrials have four and two doses, respectively. We may set nmax j = 24 and
12 (i.e., eight and four cohorts), respectively.

Overdose Control This panel (Figure 4.9) specifies the overdose control rule,
described in Section 4.3.1. That is, if Pr(πjk > φ | Djk ) > pE and njk ≥ 3,
dose combination Aj Bk and higher dose combinations are eliminated from the
trial, and the trial is terminated if the lowest dose combination is eliminated.
When the trial is terminated due to toxicity, no dose combination should be
selected as MTD. In general, we recommend to use the default probability
cutoff pE = 0.95. A smaller value (e.g., 0.9), results in stronger overdose
control, but at the cost of reducing the probability of correctly identifying
Drug-Combination Trials 89

FIGURE 4.9: Specify the overdose control rules.

MTD. This is because, in order to correctly identify MTD, it is imperative to

explore the doses sufficiently to learn their toxicity profile.
In some trial settings, under the null case that all doses are overly toxic
(i.e., the lowest/first dose combination is above MTD), the probability of early
trial termination may not be as high as we desire (e.g., > 70%). That is simply
because the small sample size cannot provide enough power to distinguish
whether a dose is overly toxic or not (e.g., 40% vs. 30%). To achieve a high
early termination probability (when the first dose is overly toxic), we can
activate the first option “ Check to impose a more stringent safety stopping
rule” (see Figure 4.10). This option allows the user to impose a more strict dose

FIGURE 4.10: Adjust the overdose control cutoff.

90 Model-Assisted Bayesian Designs for Dose Finding and Optimization

elimination rule that is more likely to eliminate the lowest/first dose, and thus
terminate the trial, by shifting the probability cutoff by δ. The default value
δ = 0.05 is generally recommended and produces a good balance between the
safety and accuracy to identify MTD. A large value of δ (e.g., 0.1) increases
the early termination probability (when all doses are toxic), but at the cost of
reducing the probability of correctly identifying MTD (when one of the doses
is MTD).
At the end of the trial, both the BOIN combination and waterfall designs
select MTD as the dose whose isotonic estimate of DLT probability is closet
to φ, globally or within each row of the dose matrix. In some cases, it may be
desirable to require the DLT probability estimate of MTD to be lower than
the de-escalation boundary λd . This can be done by activating the option at
the bottom of the “Overdose Control” panel (Figure 4.9).
After completing the specification of trial parameters, the decision table
will be generated by clicking the “Get Decision Table” button. The decision
table and the desirability score table will be automatically included in the
protocol template in Step 3, but they also can be saved as a separate csv,
Excel, or pdf file in this step if needed.

Step 2: Run simulation

Operating Characteristics This step generates the operating characteristics
of the BOIN combination design or the waterfall design through simulation,
see Figure 4.11. The simulation scenarios should cover various possible clinical
scenarios (e.g., MTDs are located at different dose combinations). To facili-
tate the generation of the operating characteristics of the design, the software
automatically provides a set of randomly generated scenarios with various
MTD locations, which are often adequate for most trials. Depending on the
application, users can add or remove scenarios. The simulation results will be
automatically included as a table in the protocol template in the next step,
but can also be saved as a separate csv or Excel file if needed.

Step 3: Generate protocol template

Protocol Preparation The BOIN combination software generates sample
texts and a protocol template to facilitate the protocol write-up. The protocol
template can be downloaded in various formats (see Figure 4.12). Use of this
module requires the completion of Steps 1 and 2. Once the protocol is approved
by regulatory bodies (e.g., Institutional Review Board), we follow the design
decision table as well as the desirability score table to conduct the trial and
make adaptive decisions (e.g., dose escalation/stay/de-escalation). When the
trial completes, use the “Select MTD” module to select MTD.
Drug-Combination Trials 91

FIGURE 4.11: Simulate operating characteristics of the design.

FIGURE 4.12: Download protocol templates.

5
Late-Onset Toxicity

5.1 A common logistical problem

The paradigm for phase I clinical trial design was initially established in the
era of cytotoxic chemotherapies, for which toxicities were often acute and as-
certainable in the first cycle of therapy. Over the past decade, non-cytotoxic
therapies such as molecularly targeted therapies and immunotherapies have
entered the clinic. Toxicity associated with these agents is often late onset
(Postel-Vinay et al., 2011; Weber et al., 2015), as is that associated with
conventional radiotherapy, which may occur several months post-treatment.
To account for late-onset toxicity, it is imperative to use a relatively long
toxicity assessment window (e.g., over multiple treatment cycles) to define
the dose-limiting toxicity (DLT) such that all DLTs relevant to the dose
escalation/de-escalation and determination of the maximum tolerated dose
(MTD) are captured. This, however, causes a common logistical problem
with sequentially outcome-adaptive dose-finding designs, which require the
outcomes of previously treated patients to be observed soon enough to apply
design’s decision rules to choose treatments for new patients. For example, if
the DLT takes up to eight weeks to evaluate and the accrual rate is one pa-
tient/week, then on average, seven new patients will be accrued while waiting
to evaluate the outcome of the first patient. The question is: how can new
patients receive timely treatment when the previous patients’ outcomes are
pending?
The same logistical difficulty arises with rapid accrual. Suppose that DLT
of a new agent can be assessed in the first 28-day cycle; if the accrual rate
is eight patients/28 days, then on average seven new patients will be accured
while waiting to evaluate the first patient’s outcome, and the clinicians must
determine how to provide them with timely treatment.
This problem, often known as the late-onset (or delayed) outcome problem,
persists throughout the trial. Several empirical approaches have been used
in practice to alleviate this issue, but they suffer from various flaws. One
approach is to turn away new patients and treat them off protocol. This may
be less desirable than giving patients the experimental regimen, or impossible
if no alternative treatment exists.
Another approach is to give all new patients the dose or treatment
that is optimal based on the most recent data. This can have disastrous

DOI: 10.1201/9780429052781-5 93
94 Model-Assisted Bayesian Designs for Dose Finding and Optimization

consequences if the most recently chosen dose later turns out to be overly
toxic. For example, suppose that the current dose level j has a total of
three patients, but all of them are waiting for toxicity evaluation. Motivated
by the desire to complete the trial quickly and based on the most recent
data, the dose-assignment decision for a new cohort of three patients is to
escalate the dose to j + 1, as none of the three patients have the DLT so far.
If it turns out that all three patients at dose level j experience severe late-
onset toxicities, then the three new patients already have been treated at an
excessively toxic dose.
A third approach is to suspend accrual after each cohort and wait until the
DLT data for the already accrued patients have cleared before enrolling the
next new cohort. This approach of repeatedly interrupting accrual, however, is
highly undesirable and often infeasible in practice. It delays treatment for new
patients and slows down the trial. Additionally, this approach consequently
increases the drug development duration, especially when the accrual is hard
as is the case in rare diseases.
In this chapter, we describe several novel phase I designs to handle late-
onset toxicities. Compared to the above empirical approaches, these designs
are statistically and scientifically rigorous and yield substantially better op-
erating characteristics. In what follows, we first characterize the implications
of late-onset toxicities, and then introduce and compare model-based designs
and model-assisted designs.

5.2 Late-onset toxicities

In a phase I clinical trial, patients are sequentially treated, and each patient is
followed for a fixed period of time (i.e., DLT assessment window), say (0, τ ), to
assess the toxicity. For the ith patient, let xi denote the binary DLT outcome,
(
1, if a DLT is observed within (0, τ ),
xi =
0, if a DLT is not observed within (0, τ ),

and let ti denote the time to DLT. For subjects who have experienced DLT,
we have 0 ≤ ti ≤ τ ; for those who do not experience DLT during the DLT
assessment window, we set ti = ∞. τ should be elicited from clinicians and
large enough to capture all DLTs relevant to MTD determination. For many
chemotherapies, τ is often taken as the first cycle of the therapy (e.g., 21 or
28 days), whereas for agents expected to induce late-onset toxicity (e.g., some
targeted or immunotherapy agents), τ can be several months or longer after
treatment.
At an interim decision time, let ui (0 ≤ ui ≤ τ ) denote the actual follow-
up time for patient i. If the patient has finished the DLT assessment, then
Late-Onset Toxicity 95

ui = τ . Let δi indicate whether the toxicity outcome xi has been ascertained

(i.e., δi = 1) or is still pending (i.e., δi = 0) by the interim decision time. Then
it follows that (
1, if ti ≤ ui or ui = τ < ti ,
δi = (5.1)
0, if ti > ui and ui < τ.
That is, xi is unobserved for patients who have not yet experienced toxicity
(ti > ui ) and have not been fully followed up (ui < τ ); and xi is observed
when patients either have experienced DLT (ti ≤ ui ) or have completed the
entire assessment (ui = τ ) without experiencing DLT.
In general, missing data are more likely to occur for patients who would
not experience DLT in (0, τ ). This phenomenon is illustrated in Figure 5.1.
The following theorem holds:
Theorem 5.1 The missing data induced by late-onset toxicity or fast accrual
are nonignorable with Pr(δi = 0|xi = 0) > Pr(δi = 0|xi = 1).
Proof: Considering that each patient is fully followed up to an assessment
period τ , if ti > τ , then xi = 0 and if ti ≤ τ , then xi = 1. For a patient who
will not experience DLT within the DLT assessment window, the probability

Patient 2
(x2=0)

Patient 1
(x1=1)

0 1 2 3 4

Time (month)

FIGURE 5.1: Illustration of missing outcomes when the toxicity assessment

window is three months and accrual rate is one patient per month. For each
patient, the horizontal line segment represents the three-month assessment
period, during which occurrence of DLT is indicated by a cross. The outcome
of patient 1 is missing at month 1, but observed between months 1 and 2. The
outcome of patient 2 is missing between months 1 and 4.
96 Model-Assisted Bayesian Designs for Dose Finding and Optimization

that his/her toxicity outcome will be missing is given by

Pr(δi = 0 | xi = 0) = Pr(ti > ui , ui < τ | xi = 0)

= Pr(ui < τ | xi = 0) Pr(ti > ui | ui < τ, xi = 0)
= Pr(ui < τ | xi = 0) Pr(ti > ui | ui < τ, ti > τ )
= Pr(ui < τ ),

where the last equality follows because ti and ui are independent, and Pr(ti >
ui | ui < τ, ti > τ ) = 1. Similarly, for a patient who will experience DLT, the
probability that his/her toxicity outcome will be missing is given by

Pr(δi = 0 | xi = 1) = Pr(ti > ui , ui < τ | xi = 1)

= Pr(ui < τ | xi = 1) Pr(ti > ui | ui < τ, xi = 1)
= Pr(ui < τ ) Pr(ti > ui | ui < τ, ti ≤ τ ).

Because of Pr(ti > ui | ui < τ, ti ≤ τ ) < 1, it follows that

Pr(δi = 0|xi = 0) > Pr(δi = 0|xi = 1).

Therefore, the missing data are more likely to occur for those patients who
would not experience DLT in the follow-up period. Because the missing data
are nonignorable, the empirical approach that discards the missing data and
makes dose-assignment decisions solely based on the observed toxicity data
leads to biased inference and poor operating characteristics (Little and Rubin,
2014).

5.3 TITE-CRM
A number of model-based designs have been proposed to address late-onset
toxicities. Based on the framework of the continual reassessment method
(CRM), Cheung and Chappell (2000) proposed the time-to-event CRM
(TITE-CRM) to incorporate pending patients’ follow-up times into toxicity
evaluation via a weighting scheme. Bekele et al. (2007) proposed to monitor
late-onset toxicities using predicted risks. Regarding the late-onset toxicity is-
sue as a missing data problem, Yuan and Yin (2011b) utilized the expectation–
maximization algorithm to estimate the dose toxicity probabilities based on
the incomplete data to direct dose assignment, and Liu et al. (2013) proposed
the Bayesian data-augmentation CRM (DA-CRM) to handle pending toxic-
ity data. Yin et al. (2013) imputed the missing toxicity data based on the
Kaplan–Meier estimate. In what follows, we use TITE-CRM as an example
to illustrate the model-based approach.
Late-Onset Toxicity 97

The idea of TITE-CRM is to weigh pending patients by their follow-up

proportions when evaluating the likelihood of the CRM model. Recall that
CRM assumes a parametric dose–toxicity model
exp{α}
πj = q j , for j = 1, . . . , J,
where α is the unknown parameter, and q1 < · · · < qJ are prior estimates
of the DLT probabilities, called the skeleton. Let x̃i be the actual observed
data at the interim decision-making time, indicating whether patient i has
experienced DLT (x̃i = 1) or not yet (x̃i = 0) at the moment of the interim
decision. It is important to distinguish x̃i from xi (i.e., the final data when
patients complete their DLT assessment). Clearly, x̃i = 1 implies xi = 1,
but when x̃i = 0, xi can be either 0 or 1 due to late-onset toxicities. Let
d[i] = 1, . . . , J denote the dose level that patient i received. TITE-CRM uses
the following weighted likelihood to estimate the toxicity probabilities of the
doses:
n h iI(x̃i =1) h iI(x̃i =0)
exp{α} exp{α}
Y
L(D | α) = ωi qd[i] 1 − ωi qd[i] , (5.2)
i=1

where ωi = ui /τ is the weight and I(·) is an indicator function. Under this

weighting scheme, patients who completed the DLT assessment receive a full
credit of ωi = 1, whereas pending patients receive a fractional credit that is
proportional to their follow-up time ui . Combining this weighted likelihood
with a prior of α, we obtain the posterior estimate π̂j and the dose escalation
and de-escalation decision follows the rule described in Section 2.5.
The specification of ωi = ui /τ is based on the assumption that the time-
to-DLT is uniformly distributed over (0, τ ). In equation 5.2, the individual
likelihood for a patient who has experienced DLT before ui is
Pr(ti ≤ ui ) = Pr(ti ≤ ui | ti ≤ τ ) Pr(ti ≤ τ )
exp{α}
= ωi Pr(xi = 1) = ωi qd[i] ,
and that for a pending patient with ti > ui is
exp{α}
Pr(ti > ui ) = 1 − Pr(ti ≤ ui ) = 1 − ωi qd[i] .

Cheung and Chappell (2000) considered other more complicated choices of ωi ,

and found that the simple uniform weighting scheme ωi = ui /τ is remarkably
robust and works well.
The weighted likelihood used by TITE-CRM is a pseudo-likelihood in the
sense that it equals the exact likelihood only when the time to toxicity ti is
uniformly distributed over (0, τ ). A more efficient design is DA-CRM (Liu
et al., 2013), which uses the true data likelihood for estimation and decision
making. DA-CRM models the time-to-DLT ti using a flexible piecewise expo-
nential model and uses it to impute the missing xi based on Bayesian data
augmentation. It has been demonstrated that the DA-CRM can treat patients
safely and also select MTD with a high probability.
98 Model-Assisted Bayesian Designs for Dose Finding and Optimization

5.4 TITE-BOIN
The time-to-event Bayesian optimal interval (TITE-BOIN) design is an ex-
tension of the BOIN design (Section 3.4) to address the issue of late-onset
toxicities (Yuan et al., 2018). As a model-assisted design, TITE-BOIN is sim-
ple to implement and its decision rule can be pre-tabulated and included in
the trial protocol. In addition, it yields outstanding performance comparable
to the more complicated model-based TITE-CRM.

5.4.1 Trial design

Suppose that at an interim decision time, a total of nj patients have been
enrolled at the current dose, among which rj patients have completed the
DLT assessment and their DLT data xi are known, but cj = nj − rj patients
have not completed the DLT assessment and their DLT data xi are pending.
Let Oj denote the set of patients whose DLT data are known (i.e., observed
with δi = 1) and Mj denote the set of patients whose DLT data are pending
(i.e., missing with δi = 0) at the current dose level j. The estimate of the
toxicity rate πj is given by
P P
i∈Oj xi + i∈Mj xi
π̂j = .
nj
Since the DLT data of patients from the pending set Mj are unknown, we
cannot obtain the value of π̂j . Without π̂j , the dose-assignment rules of BOIN
cannot be applied.
TITE-BOIN addresses this issue by imputing the unobserved xi . Assuming
that the time to DLT follows a uniform distribution over (0, τ ), the expected
value of xi , i ∈ Mj , for a pending patient treated at the current dose j with
follow-up time uj is given by
x̂i = E(xi | ti > ui )
= Pr(xi = 1 | ti > ui )
Pr(xi = 1) Pr(ti > ui | xi = 1)
=
Pr(xi = 1) Pr(ti > ui | xi = 1) + Pr(xi = 0) Pr(ti > ui | xi = 0)
πj (1 − ui /τ )
=
πj (1 − ui /τ ) + (1 − πj )
πj (1 − ui /τ )
≈ .
1 − πj
Here, the approximation in the last line is appropriate as πj and thus πi (1 −
ui /τ ) is typically smaller than 1 − πj . In addition, this approximation slightly
inflates the expected value of xi , which reduces the chance of aggressive dose
escalation potentially caused by late-onset toxicity.
Late-Onset Toxicity 99

By imputing the unobserved xi , i ∈ Mj , with its expected value x̂i , the

toxicity rate estimate π̂j is given by
P P
i∈Oj xi + i∈Mj x̂i
π̂j =
nj
πj
ỹj + (cj − STFTj )
1 − πj
= , (5.3)
nj
P
where ỹj = i∈Oj xi is the number of DLTs observed so far, STFTj =
P
i∈Mj u i /τ is the standardized total follow-up time (STFT) for the pend-
ing patients at the dose level j, j = 1, . . . , J. For example, given that the
DLT assessment window τ = 3 months and, at the current dose, three pend-
ing patients have been followed 1, 1.6, and 2.5 months, respectively, the
STFTj = (1 + 1.6 + 2.5)/3 = 1.7. STFT plays a key role in TITE-BOIN
for making dose-assignment decisions.
In the statistical literature, the above approach is known as single mean
imputation (Little and Rubin, 2014). One drawback of single mean imputa-
tion is that, although it provides an unbiased and consistent point estimate,
the resulting variance estimate is biased because of ignoring the imputation
uncertainty. In our case, this is not a concern as the decision rules of BOIN
only rely on the point estimate of πj . After the single mean imputation, π̂j is
a valid point estimate of πj .
Equation (5.3) involves an unknown value πj . Yuan et al. (2018) replaced it
with its Bayesian posterior mean estimate π̃j based on the observed data. As-
suming the beta-binomial model with the prior πj ∼ Beta(α, β), the Bayesian
posterior mean estimate is given by π̃j = (ỹj + α)/(rj + α + β). A vague
beta prior with α = φ/2 and β = 1 − α is recommended such that the prior
corresponds to an effective prior sample size of 1 with the prior mean φ/2. As
described in Section 5.2, π̃j is based on the observed data, and thus overesti-
mates πj . This is actually beneficial because it slightly inflates π̂j , reducing the
chance of aggressive dose escalation potentially caused by late-onset toxicity.
One important property of equation (5.3) is that π̂j is a monotonically
decreasing function of STFT. Therefore, determining whether π̂j crosses the
BOIN escalation and de-escalation boundaries λe and λd (i.e., π̂j ≤ λe or
π̂j > λd ) is equivalent to determining whether STFTj ≥ ∆e or STFTj < ∆d ,
with
     
1 − π̃j ỹj ỹj
∆e = cj − (nj λe − ỹj ) I < φ + ∞I ≥φ
π̃j nj nj
     
1 − π̃j ỹj ỹj
∆d = cj − (nj λd − ỹj ) I > φ − ∞I ≤φ ,
π̃j nj nj
   
ỹ ỹ
for j = 1, . . . , J. Here, the indicator functions I njj < φ and I njj > φ are
imposed to ensure that TITE-BOIN has the similar long-memory coherence
100 Model-Assisted Bayesian Designs for Dose Finding and Optimization

property as BOIN (i.e., the dose is never escalated/de-escalated if the observed

DLT rate ỹj /nj is greater/smaller than the target DLT rate φ). As a result,
the TITE-BOIN dose escalation/de-escalation rule becomes the following:
• If STFTj ≥ ∆e , escalate the dose to the next higher level.
• If STFTj < ∆d , de-escalate the dose to the next lower level.
• Otherwise, stay at the current dose level.

Because the formulas for ∆e and ∆d are functions of summary statistics such
as nj , ỹj , cj , and STFTj , this dose escalation/de-escalation rule can be tabu-
lated prior to trial conduct. As a result, TITE-BOIN inherits the transparency
and simplicity of the standard BOIN design and does not require repeated,
complicated model fitting after treating each cohort/patient.
Table 5.1 shows the TITE-BOIN decision rule with a cohort size of three
and a target DLT rate of 0.2. To conduct the trial, we only need to count
the number of patients at the current dose, the number of patients who expe-
rienced DLT, the number of pending patients, and STFT, and then look up
the table to determine the dose escalation/de-escalation. Suppose that three
patients have been treated at the current dose, one of them had DLT. We
de-escalate the dose regardless of STFT. Consider another case where nine
patients have been cumulatively treated at the current dose, one patient had
DLT, and four patients have DLT data pending. To treat the next cohort of
patients, if STFT of the four pending patients is greater than 2.15, we escalate
the dose; otherwise, we retain the current dose. Table 5.1 assumes a cohort
size of three, but TITE-BOIN allows any prespecified cohort size, and the
corresponding decision table (i.e., similar to Table 5.1 but with more rows)
can be easily generated using the software described later.
In principle, TITE-BOIN supports continuous accrual and allows for real-
time dose assignment whenever a new patient arrives. To avoid risky decisions
caused by sparse data, an accrual suspension rule is imposed:

If at the current dose, more than 50% of the patients’ DLT outcomes are
pending, suspend the accrual to wait for more data to become available.
This rule corresponds to “Suspend accrual” in Table 5.1. In addition, the same
overdose control/safety stopping rule as BOIN is also employed:

If Pr(πj > φ | Dj ) > 0.95 and nj ≥ 3, dose level j and higher are elimi-
nated from the trial, and the trial is terminated if the lowest dose level is
eliminated.
This overdose control rule corresponds to the decision “Y&Elim,” representing
“Yes & Eliminate,” under the column entitled “De-escalate” in Table 5.1.
Late-Onset Toxicity 101

TABLE 5.1: Dose escalation and de-escalation rule for TITE-BOIN with a
target DLT rate of 0.2 and a cohort size of three, up to 12 patients.

No. No. No. data STFT

treated DLTs pending Escalate Stay De-escalate
3 0 ≤1 Y
3 0 ≥2 Suspend accrual
3 1 ≤2 Y
3 ≥2 ≤1 Y&Elim
6 0 ≤3 Y
6 0 ≥4 Suspend accrual
6 1 ≤3 Y
6 1 ≥4 Suspend accrual
6 2 ≤4 Y
6 ≥3 ≤3 Y&Elim
9 0 ≤4 Y
9 0 ≥5 Suspend accrual
9 1 ≤2 Y
9 1 3 ≥ 0.77 < 0.77
9 1 4 ≥ 2.15 < 2.15
9 1 ≥5 Suspend accrual
9 2 0 Y
9 2 1 > 0.52 ≤ 0.52
9 2 2 > 1.59 ≤ 1.59
9 2 3 > 2.66 ≤ 2.66
9 2 4 > 3.73 ≤ 3.73
9 2 ≥5 Suspend accrual
9 3 ≤6 Y
9 ≥4 ≤5 Y&Elim
12 0 ≤6 Y
12 0 ≥7 Suspend accrual
12 1 ≤5 Y
12 1 6 ≥ 1.24 < 1.24
12 1 ≥7 Suspend accrual
12 2 ≤6 Y
12 2 ≥7 Suspend accrual
12 3, 4 ≤9 Y
12 ≥5 ≤7 Y&Elim
Note: “No. treated” is the total number of patients treated at the current dose level, “No.
DLTs” is the number of patients who experienced DLT at the current dose level, “No. data
pending” denotes that number of patients whose DLT data are pending at the current dose
level, “STFT” is the standardized total follow-up time for the patients with data pending,
defined as the total follow-up time for the patients with data pending divided by the
length of the DLT assessment window. “Y” represents “Yes,” and “Y&Elim” represents
“Yes & Eliminate.” When a dose is eliminated, all higher doses should also be eliminated.
102 Model-Assisted Bayesian Designs for Dose Finding and Optimization

5.4.2 Incorporating prior information

In some trials, prior information is available on the distribution of ti (i.e.,
time to DLT). For example, for a certain drug, we may know a priori that the
DLT is more likely to occur in the later part of the DLT assessment window
(0, τ ). Such prior information can be easily incorporated into TITE-BOIN by
assuming ti follows a piecewise uniform prior distribution, which partitions
(0, τ ) into several intervals and specifies a uniform distribution within each
interval. By increasing the number of partitions, the piecewise uniform dis-
tribution can approximate any shape of the time-to-toxicity distribution. For
many applications, three partitions (i.e., (0, τ /3), (τ /3, 2τ /3), and (2τ /3, 1)),
are often adequate for practical use. Let ν1 , ν2 , and ν3 (ν1 + ν2 + ν3 = 1)
be the prior probabilities that the DLT would occur at the three intervals,
respectively. Let h0 = 0, h1 = τ /3, h2 = 2τ /3, and h3 = τ , and define

τ,
 ui > hk
ũik = ui − hk−1 , ui ∈ (hk−1 , hk ], k = 1, 2, 3.

0, otherwise


Then, the conditional probability Pr(ti > ui | xi = 1) is given by

3
X
Pr(ti > ui | xi = 1) = 1 − 3νk ũik /τ.
k=1

Using similar algebra and approximation, we have

πj
ỹj + (cj − WSTFTj )
1 − πj
π̂j = , (5.4)
nj
where WSTFT is the weighted STFT, given by
3
X X
WSTFTj = 3νk ũik /τ.
i∈Mj k=1

When ν1 = ν2 = ν3 = 1/3 (i.e., assigning an equal weight over the DLT

assessment window), WSTFT reduces to STFT. In other words, equation
(5.3) using the uniform weight is a special case of equation (5.4).
One attractive feature of TITE-BOIN is that using a non-uniform (in-
formative) prior for ti does not alter its decision table, such as Table 5.1,
generated under the non-informative uniform prior. That is, the TITE-BOIN
decision table is invariant to the prior distribution of ti . This is because equa-
tion (5.4) has the exact same form as equation (5.3), except that STFTj
is replaced by WSTFTj . In other words, Table 5.1 (generated by using the
non-informative uniform prior) can be used to make the decision of dose
escalation/de-escalation, with STFTj replaced by WSTFTj , when an infor-
mative prior of ti is used.
Late-Onset Toxicity 103

5.4.3 Statistical properties

Another desirable feature of TITE-BOIN is that its decision rule is invariant to
the length of the assessment window τ , because STFT has been standardized
by the latter. This means that given a target DLT rate φ, the same decision
table can be used to guide dose escalation and de-escalation, regardless of
the value of τ . For example, Table 5.1 can be used for any trial with φ = 0.2,
regardless of the length of the assessment window. This is practically appealing
and greatly simplifies trial protocol preparation. In practice the assessment
window varies from one trial to another, depending on the experimental drug
and patient population, whereas the target DLT rate is often taken as 0.2,
0.25, or 0.3. Another attractive feature of TITE-BOIN is that when there is
no pending DLT data, it seamlessly reduces to the BOIN design.
When dealing with late-onset toxicities, the top concern is patient safety,
as the patients with pending outcome data who have not experienced DLT at
the interim decision time may yet experience DLT late in the follow-up pe-
riod. Any reasonable design that handles late-onset toxicity should take that
fact into account in its decision making, which can be described by the mono-
tonicity property. Let Dj = {(x̃i , ti ), i = 1, . . . , nj } be the actually observed
interim data at dose level j, and let Djs denote the “cross-sectional” interim
data obtained by setting xi = x̃i (i.e., treating the patients’ temporary DLT
outcomes at the interim time as their final DLT outcomes at the end of the
assessment window). Because of late-onset toxicity, for pending patients, x̃i
(i.e., the DLT status observed at the interim) is not necessarily equal to xi
(i.e., the DLT status at the end of assessment window). More precisely, x̃i = 1
implies xi = 1, but when x̃i = 0, xi can be either 0 or 1. Let a(Dj ) = −1, 0
and 1 denote the decisions of dose de-escalation, retaining the current dose,
and dose escalation, respectively, based on the data Dj .

Definition (Monotonicity) A dose-finding design is monotonic if a(Dj ) ≤

a(Djs ) for j = 1, . . . , J.

Monotonicity indicates that the decision of dose transition based on the

observed data Dj should be less aggressive than that based on Djs . This is a
property that any reasonable design should obey to reflect that patients who
have not experienced DLT by the interim decision time may yet experience
DLT late in the follow-up period. The TITE-BOIN design has this property
because x̂i ≥ x̃i , thus leading to π̂j > ỹj /nj , where ỹj /nj is the observed
toxicity rate based on the “cross-sectional” data Djs .

Theorem 5.2 The TITE-BOIN design is monotonic.

5.4.4 Operating characteristics

Yuan et al. (2018) compared the operating characteristics of the TITE-BOIN
design, the 3+3 design, the rolling six design (Skolnik et al., 2008), and the
104 Model-Assisted Bayesian Designs for Dose Finding and Optimization

TABLE 5.2: Dose–toxicity scenarios used in the simulation study. The target
dose (i.e., MTD) is bolded.

Dose Level
Scenario 1 2 3 4 5 6 7
1 0.3 0.4 0.5 0.6 0.7 0.8 0.9
2 0.14 0.3 0.39 0.48 0.56 0.64 0.7
3 0.07 0.23 0.41 0.49 0.62 0.68 0.73
4 0.05 0.15 0.3 0.4 0.5 0.6 0.7
5 0.05 0.12 0.2 0.3 0.38 0.49 0.56
6 0.01 0.04 0.08 0.15 0.3 0.36 0.43
7 0.02 0.04 0.08 0.1 0.2 0.3 0.4
8 0.01 0.03 0.05 0.07 0.09 0.3 0.5

TITE-CRM design using simulation studies. Table 5.2 shows eight scenarios
used to simulate trial data. The target DLT probability is 30%. The DLT
assessment window is three months, the accrual rate is two patients/month.
The time to DLT is sampled from a Weibull distribution, with 50% of DLTs
occurring in the second half of the assessment window. The maximum sample
size is 36 patients, and patients are treated in cohorts of three. Because the
3+3 and rolling six designs often stopped the trial early (e.g., when two of
three patients experienced DLT) before reaching 36 patients, in these cases
the remaining patients are treated at the selected “MTD” as the cohort ex-
pansion, such that the four designs have comparable sample sizes. For the
3+3 design, a new cohort is enrolled only when the previous cohort’s DLT
data are cleared. More details about the simulation study can be found in
Yuan et al. (2018).
Figure 5.2 shows the relative performance of the rolling six, TITE-BOIN
and TITE-CRM designs against the performance of the 3+3 design, includ-
ing the differences in (a) the percentage of correct selection of MTD, (b) the
percentage of patients overdosed (i.e., treated at doses above MTD), (c) the
percentage of patients underdosed (i.e., treated at doses below the MTD), and
(d) the average trial duration. TITE-BOIN is more efficient and has higher
percentages of correct selection of MTD than the algorithm-based design (i.e.,
the rolling six design) because of using the follow-up times of pending patients
to determine dose escalation and de-escalation. Compared to the model-based
TITE-CRM, TITE-BOIN has comparable accuracy to identify MTD, but it
is safer and much simpler. Compared to the 3+3 design, TITE-BOIN dra-
matically shortens the trial duration owing to its ability to make real-time
decisions in the presence of pending patients.
Late-Onset Toxicity 105

(a) Percentage of correct selection (b) Percentage of patients overdosed

30
30

20
Percentage

Percentage
20

10
10

0
0

1 2 3 4 5 6 7 8 Average 1 2 3 4 5 6 7 8 Average
Scenario Scenario

(c) Percentage of patients underdosed (d) Average trial duration

0
10

0 −5
Percentage

Month
−10 −10

−20
−15

−30
−20

1 2 3 4 5 6 7 8 Average 1 2 3 4 5 6 7 8 Average
Scenario Scenario

Method R6 TITE−BOIN TITE−CRM

FIGURE 5.2: The relative performance of the rolling six (R6), TITE-BOIN,
and TITE-CRM designs against the performance of the 3+3 design, including
the differences in (a) the percentage of correct selection of the MTD, (b) the
percentage of patients overdosed (i.e., treated at doses above the MTD), (c)
the percentage of patients underdosed (i.e., treated at doses below the MTD),
and (d) the average trial duration.

5.5 A unified approach using “effective” data

We have discussed the use of mean imputation to address the late-onset tox-
icity for the BOIN design. That approach, however, cannot be applied to
other model-assisted designs such as the mTPI and keyboard designs because
these designs make decisions based on the posterior distribution of πj . In
this section, we introduce a more general approach that is applicable to all
model-assisted designs (Lin and Yuan, 2020).
In the presence of late-onset toxicity, the interim patients can be divided
into two subgroups: the patients with ascertained toxicity outcomes (i.e., δi =
1), and the patients with pending outcomes (i.e., δi = 0). For a patient with
δi = 1, we have x̃i = xi . Assuming that patient i received dose level j, the
106 Model-Assisted Bayesian Designs for Dose Finding and Optimization

likelihood is given by

Pr(x̃i = xi , δi = 1) = πjxi (1 − πj )1−xi . (5.5)

For a patient with δi = 0, xi has not been ascertained yet with DLT outcome
pending, and his/her actual observed outcome is x̃i = 0. These pending pa-
tients are a mixture of two subgroups: patients who will not experience DLT
(i.e., xi = 0), and patients who will experience DLT (i.e., xi = 1) but have
not experienced it yet by the interim decision time (i.e., ui < ti ). Therefore,
the likelihood for a pending patient is given by
Pr(x̃i = 0, δi = 0)
= Pr(xi = 0) Pr(x̃i = 0, δi = 0|xi = 0) + Pr(xi = 1) Pr(x̃i = 0, δi = 0 | xi = 1)
= Pr(xi = 0) + Pr(xi = 1) Pr(ti > ui | xi = 1)
= 1 − πj + πj {1 − Pr(ti ≤ ui | xi = 1)}
= 1 − π j ωi ,

where ωi = Pr(ti ≤ ui | xi = 1) can be treated as a weight function, which

will be discussed later. Of note, for pending patients, x̃i only takes a value of
0 because once x̃i = 1 (i.e., the patient experiences the DLT), xi is observed.
Therefore, given the interim data Dj observed at dose level j, the joint
likelihood function is given by
nj
Y
L(Dj | πj ) ∝ πjδi xi (1 − πj )δi (1−xi ) (1 − ωi πj )1−δi
i=1
nj
ỹ
Y
= πj j (1 − πj )mj (1 − ωi πj )1−δi , (5.6)
i=1
Pnj
where ỹj = i=1 δi xi is the
Pnjnumber of patients who experienced DLT by the
interim time, and mj = i=1 δi (1 − xi ) = rj − ỹj is the number of patients
who have completed the assessment without experiencing DLT.
Let f (πj ) denote the prior distribution for πj (e.g., f (πj ) = Beta(1, 1)).
The posterior distribution f (πj | Dj ) is given by

f (πj | Dj ) ∝ f (πj )L(Dj | πj ).

Recall that without late-onset toxicity, the posterior f (πj | Dj ) follows a

beta distribution, see equation (3.2). In the presence of pending DLT data,
f (πj | Dj ) however does not have a standard distributional form. Although
f (πj | Dj ) can be sampled using the Markov chain Monte Carlo method, it
prohibits the enumeration of the decision rule and thus destroys the simplicity
of model-assisted designs.
To circumvent this issue and maintain the simplicity of the model-assisted
designs, Lin and Yuan (2020) proposed to approximate the last term in (5.6)
as follows,
(1 − ωi πj )1−δi ≈ (1 − πj )ωi (1−δi ) . (5.7)
Late-Onset Toxicity 107

When δi = 0, this is a first-order Taylor expansion approximation, noting that

the Taylor expansion of (1 − πj )ωi at πj = 0 is

(1 − πj )ωi = 1 − ωi πj + ωi (ωi − 1)πj2 + · · · .

When δi = 1, the approximation actually is exact as (1 − ωi πj )1−δi = (1 −

πj )1−δi = 1. Thus, the likelihood (5.6) is approximated as
ỹ
L(Dj | πj ) ∝ πj j (1 − πj )ñj −ỹj , (5.8)

where
nj
X
ñj = ỹj + m̃j , m̃j = mj + (1 − δi )ωi . (5.9)
i=1

This approximation is simple but extremely powerful. It converts the non-

regular likelihood (5.6) into a standard binomial likelihood arising from “ef-
fective” binomial data D̃j = (ñj , ỹj ), where ñj = ỹj + m̃j is the effective
sample size, and m̃j is the effective number of patients who have not ex-
perienced DLT. Because the model-assisted designs in previous chapters are
based on the binomial likelihood, these designs can be seamlessly extended
to accommodate pending DLT data using the approximated likelihood (5.8),
as demonstrated in the next section. In addition, because the approximated
likelihood (5.8) depends on the aggregated value of the ωi ’s, rather than the
individual value of ωi , the approximation renders it possible to enumerate the
decision rules for the resulting design, maintaining the most important feature
of the model-assisted designs.
The following theorem establishes that (5.8) provides an accurate approx-
imation of the exact likelihood (5.6).
Theorem 5.3 Let l(ωi , δi , πj ) = (1 − ωi πj )1−δi and ˜l(ωi , δi , πj ) = (1 −
πj )ωi (1−δi ) . The approximation error is bounded by

d(ωi , δi , πj ) = |l(ωi , δi , πj ) − ˜l(ωi , δi , πj )| ≤ (1 − βj πj ) − (1 − πj )βj ,

where βj = log{−πj / log(1 − πj )}/ log(1 − πj ). Specifically, for any πj ≤ 0.4,

the approximation error d(ωi , δi , πj ) < 0.0255.
The proof of Theorem 5.3 is provided in Lin and Yuan (2020). To illustrate
the accuracy of the approximation, consider two examples: (a) nj = 5 patients
have been treated and only mj = 2 patients have finished the assessment
without any DLT, and the weights ωi for the remaining three patients are
0.3, 0.4, and 0.5, respectively, leading to m̃j = 2 + 0.3 + 0.4 + 0.5 = 3.2; (b)
nj = 12 patients have been treated, ỹj = 1 DLT has been observed, mj = 4
patients have finished the assessment without any DLT, and the weights ωi
for the remaining seven patients are 0.1, 0.2, . . . , 0.7, respectively, leading to
m̃j = 4 + 0.1 + · · · + 0.7 = 6.8. Given the prior πj ∼ Unif(0, 1), Figure 5.3
shows the near coincidence between the exact and approximated posterior
distribution functions, indicating that the approximation is quite accurate.
108 Model-Assisted Bayesian Designs for Dose Finding and Optimization

(a) (nj, ~ ~ )=(5, 0, 2, 3.2)

y j, mj, m (b) (nj, ~ ~ )=(12, 1, 4, 7.8)
y j, mj, m
j j

1.0

0.05
True True
Approximated Approximated
0.8

0.04
Posterior distribution

Posterior distribution
0.6

0.03
0.02
0.4

0.01
0.2

Ik*−1 Ik* Ik*+1 Ik*−1 Ik* Ik*+1

0.00
0.0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
πj πj

FIGURE 5.3: The exact and approximated posterior functions based on the
observed data: (a) nj = 5 patients have been treated and only mj = 2 pa-
tients have finished the assessment without any DLT, and the weights for the
remaining three patients are 0.3, 0.4, and 0.5; (b) nj = 12 patients have been
treated, ỹj = 1 DLT has been observed, mj = 4 patients have finished the as-
sessment without any DLT, and the weights for the remaining seven patients
are 0.1, 0.2, . . . , 0.7. The prior distribution of πj is πj ∼ Unif(0, 1); ỹj and m̃j
represent the “effective” numbers of patients with DLT and patients without
DLT, respectively, by the interim time γ; Ik∗ represents the target key in the
keyboard design.

We now discuss how to specify weight ωi that appears in the approximated

likelihood (5.8). By construction, the weight ωi adjusts for the fact that the
DLT outcome for the pending patient i has not been ascertained yet. Following
TITE-CRM and TITE-BOIN, the simplest way is to assume that the time to
toxicity ti is uniformly distributed over the assessment period (0, τ ), leading
to
ωi = Pr(ti < ui | xi = 1) = ui /τ. (5.10)
As a result, ωi can be interpreted as the follow-up proportion that patient i
has finished. Based on this uniform weight, the effective sample size ñj in the
approximated likelihood (5.8) can be easily calculated as
Total follow-up time for pending patients at dose j
ñj =
Length of assessment window
+ No. of non-pending patients at dose j.
= STFTj + rj

Although the uniform scheme seems very restrictive, it yields remarkably ro-
bust performance. This was also observed in TITE-CRM and TITE-BOIN.
As described in section 5.4.2, a more flexible weighting scheme is to assume
that ti follows a piecewise uniform distribution, which partitions (0, τ ) into
Late-Onset Toxicity 109

several intervals and assumes a uniform distribution within each interval. This
approach is useful to incorporate prior information on ti . For ease of exposi-
tion, the assessment window (0, τ ) is partitioned into the initial part (0, τ /3),
the middle part (τ /3, 2τ /3), and the final part (2τ /3, τ ), and it assumes that
ti is uniformly distributed in each interval. Let (ν1 , ν2 , ν3 ) be the prior proba-
bility that the DLT would occur at the three parts of the assessment window,
where ν1 + ν2 + ν3 = 1. For example, prior data may suggest that the DLT
is more likely to occur late in the assessment window, in which case we can
choose ν3 > ν2 > ν1 . It then follows that

3ν1 ui /τ,
 ui ∈ (0, τ /3),
ωi = Pr(ti < ui | xi = 1) = ν1 − ν2 + 3ν2 ui /τ, ui ∈ (τ /3, 2τ /3) ,

ν1 + ν2 − 2ν3 + 3ν3 ui /τ, ui ∈ (2τ /3, τ ),


where ωi can be interpreted as the weighted follow-up probability that patient

i has completed the assessment time. Lin and Yuan (2020) also discussed a
more complicated data-driven adaptive weighting scheme, however it only
provides negligible improvement over the above two simpler schemes.

5.6 TITE-keyboard and TITE-mTPI designs

Application of the above effective data approach to the keyboard design is
straightforward. We refer to the resulting design based on the approximated
likelihood (5.8) as the TITE-keyboard design. The decision rule of the TITE-
keyboard design is almost the same as that of the keyboard design. The only
difference is that to make decisions of dose escalation and de-escalation, the
complete data Dj = (nj , yj ), which are not observable when some DLT data
are pending, are replaced by the effective binomial data D̃j = (ñj , ỹj ) for
calculating the posterior distribution of πj and identifying the strongest key.
Once the strongest key is identified, the same dose escalation/de-escalation
rule is used to guide the dose transition.
Compared to the model-based TITE-CRM, the most appealing feature of
TITE-keyboard is that its dose transition rule can be tabulated before the
trial begins; see Table 5.3 as the decision table with the target DLT rate
φ = 0.3. To conduct the trial, there is no need for real-time model fitting,
investigators only need to count the number of patients
Pnj with DLTs (i.e., ỹj ),
the number of patients with data pending (i.e., cj = i=1 (1−δi )), the effective
number of patients without DLT (i.e., m̃j ), and then use the decision table
to determine the dose assignment for the next new cohort. Like the TITE-
BOIN design, another feature of the TITE-keyboard design is that its decision
110 Model-Assisted Bayesian Designs for Dose Finding and Optimization

TABLE 5.3: Dose escalation and de-escalation boundaries for TITE-keyboard

with a target DLT probability of 0.3 and cohort size of three, up to 12 patients.

nj ỹj cj Escalation Stay De-escalation

3 0 ≤1 Y
3 0 ≥2 Suspend accrual
3 1 0 Y
3 1 [1, 2] m̃j > 1.88 m̃j ≤ 1.88
3 2 ≤1 Y
3 3 0 Y&Elim
6 0 ≤6 Y
6 1 ≤1 Y
6 1 [2, 3] m̃j ≥ 3.07 m̃j < 3.07
6 1 [4, 5] m̃j ≥ 3.07 1.88 < m̃j > 3.07 m̃j ≤ 1.88
6 2 0 Y
6 2 [1, 4] m̃j > 3.75 m̃j ≤ 3.75
6 3 ≤3 Y
6 4 ≤2 Y&Elim
9 0 ≤9 Y
9 1 ≤4 Y
9 1 [5, 6] m̃j ≥ 3.07 m̃j < 3.07
9 1 [7, 8] m̃j ≤ 3.07 1.88 < m̃j < 3.08 m̃j ≤ 1.88
9 2 0 Y
9 2 [1, 3] m̃j ≥ 6.15 m̃j < 6.15
9 2 [4, 7] m̃j ≥ 6.15 3.75 < m̃j < 6.15 m̃j ≤ 3.75
9 3 0 Y
9 3 [1, 6] m̃j > 5.63 m̃j ≤ 5.63
9 4 ≤5 Y
9 5 ≤4 Y&Elim
12 0 ≤ 12 Y
12 1 ≤7 Y
12 1 [8, 9] m̃j ≥ 3.07 m̃j < 3.07
12 1 [10, 11] m̃j ≥ 3.07 1.88 < m̃j < 3.08 m̃j ≤ 1.88
12 2 ≤3 Y
12 2 [4, 6] m̃j ≥ 6.15 m̃j < 6.15
12 2 [7, 10] m̃j ≥ 6.15 3.75 < m̃j < 6.15 m̃j ≤ 3.75
12 3 ≤3 Y
12 3 [4, 9] m̃j > 5.63 m̃j ≤ 5.63
12 4 0 Y
12 4 [1, 8] m̃j > 7.50 m̃j ≤ 7.50
12 5,6 ≤7 Y
12 7 ≤5 Y&Elim
Note: nj is the number of patients at dose level j, ỹj is the number of DLTs observed by
Pnj
the decision time, c̃j = i=1 (1 − δi ) is the number of patients who have data pending,
and m̃j is the effective number of patients without any DLT. Dose escalation is not
allowed if fewer than two patients at dose level j have finished the assessment. “Y”
represents “Yes,” and “Y&Elim” represents “Yes & Eliminate.” When a dose is
eliminated, all higher doses should also be eliminated.
Late-Onset Toxicity 111

table does not depend on the weighting scheme or the length of the DLT
assessment window. Table 5.3 applies no matter which of the aforementioned
three weighting schemes is used and no matter the length of the assessment
window. This is because the likelihood (5.8) depends on m̃j , which is the
effective number of patients without DLT. Moreover, when all the pending
DLT data become available, we have (ñj , ỹj ) = (nj , yj ). As a result, the TITE-
keyboard design reduces to the standard keyboard design in a seamless way.
For patient safety, Lin and Yuan (2020) required that dose escalation is not
allowed until at least two patients have completed the DLT assessment at the
current dose level. In addition, an overdose control/stopping rule is imposed:
at any time during the trial if any dose j satisfies Pr(πj > φ | nj , ỹj ) > η
and nj ≥ 3, then that dose and any higher doses are regarded as overly toxic
and should be eliminated from the trial, and the dose is de-escalated to level
j − 1 for the next cohort of patients, where η is the prespecified elimination
cutoff, say η = 0.95. If the lowest dose level is eliminated, the trial should be
terminated early. Table 5.3 also reflects such safety and overdose control rules.
The TITE-keyboard design has some desirable statistical properties. First,
similar to the TITE-BOIN design, the TITE-keyboard design also enjoys the
monotonicity property.
Theorem 5.4 The TITE-keyboard design is monotonic.
Second, the TITE-keyboard design is long-memory coherent. The proofs of
Theorems 5.4 and 5.5 are given in Lin and Yuan (2020).

Theorem 5.5 The TITE-keyboard design is long-memory coherent in the

sense that if the empirical toxicity rate π̃j = ỹj /ñj at the current dose is
greater (or less) than the target toxicity rate, the design will not escalate (nor
de-escalate) the dose, with probability one.

The effective data approach is general. It is also applicable to BOIN. We

compute π̃j = ỹj /ñj using the observed effective data, then the BOIN dose
escalation/de-escalation rule can be applied. The effective data approach also
can be directly applied to the mTPI design. The only change needed is to
replace the complete data (nj , yj ), potentially unobserved due to late-onset
toxicity or fast accrual, with the observed effective data (ñj , ỹj ) in (3.2) when
calculating the unit probability mass (UPM). Once the three UPMs are de-
termined, the decision rules remain the same. We refer to the resulting design
as the TITE-mTPI design.
Numerical study shows that TITE-keyboard has excellent performance
similar to TITE-BOIN and TITE-CRM. However, TITE-mTPI does not per-
form as well as the other designs, and it has a high risk of overdosing patients
because of the deficiency of mTPI (Section 3.2), see Lin and Yuan (2020) for
details of the numerical study.
112 Model-Assisted Bayesian Designs for Dose Finding and Optimization

5.7 Software and case study

Software
The web application for the TITE-BOIN design is available as a module of the
BOIN Suite at http://www.trialdesign.org. Figure 5.4 shows the launch-
pad of BOIN Suite. The TITE-keyboard design also can be implemented using
the web-based application at http://www.trialdesign.org. The two graph-
ical user interface-based software programs allow users to generate the dose-
assignment decision table, conduct simulations, obtain the operating char-
acteristics of the design, and generate a trial design template for protocol
preparation. The R codes for other time-to-event model-assisted designs are
available on Github (https://github.com/ruitaolin/TITE-MAD).

Case study
Recurrent or High Grade Gynecologic Cancer Trial The objective of this
phase I trial (ClinicalTrials.gov Identifier: NCT03508570) is to determine

FIGURE 5.4: The launchpad of web application BOIN Suite available at

http://www.trialdesign.org.
Late-Onset Toxicity 113

MTD and the recommended phase II dose (RP2D) of intraperitoneal

nivolumab in combination with ipilimumab in treating patients with recur-
rent or high-grade gynecologic cancer. As late-onset toxicities are anticipated
for this combination immunotherapy, the DLT assessment window is set as
12 weeks and the TITE-BOIN design is employed to conduct the trial. The
maximum sample size is 24 and patients are treated in a cohort size of three.
The expected accrual rate is two patients per month. Four doses of ipilimumab
will be investigated in combination with a fixed dose of nivolumab, and the
target DLT probability is φ = 30%.

We use this trial as an example to illustrate how to use TITE-BOIN to de-

sign phase I trials with late-onset toxicities. After selecting and launching the
TITE-BOIN module from the BOIN Suite launchpad (Figure 5.4), the trial is
desgined using the following three steps:
Step 1: Enter trial parameters
Doses and Sample Size As shown in Figure 5.5, the number of doses under
investigation is four, starting dose level is one. For trials where the lowest dose
is believed to be safe, starting from a slightly higher dose level (e.g., two) may
save the sample size as it allows the design to reach MTD sooner. However,
due to limited knowledge on the safety of the new drug, in general it is not
recommended to start from a high dose level (e.g., four).
The cohort size for the trial is three and the number of cohorts is eight,
with a total sample size of 24. When m = 12 patients have been assigned to
a dose and the decision is to stay at that dose, the trial will be stopped early

FIGURE 5.5: Specify doses, sample size, and convergence stopping rule.
114 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 5.6: Specify the target, DLT assessment window, and accrual rate.

even if the maximum sample size of 24 is not reached. Because the TITE-
BOIN design is a generalization of BOIN, one can follow the rule of BOIN
to specify the maximum sample size and the convergence criterion for early
stopping, see more details in Section 3.6.

Target and Accrual Rate As shown in Figure 5.6, the target DLT proba-
bility is φ = 0.3, and the default values for (φ1 , φ2 ) are used to derive the
optimal dose escalation and de-escalation boundaries. Section 3.6 provides
some guidance on how to adjust the values for φ, φ1 , and φ2 if a specific
safety requirement is desirable.
TITE-BOIN requires the specification of the DLT assessment window
(τ = 2.8 months) and the accrual rate (two patients/month). The decision
rule of TITE-BOIN does not depend on the accrual rate. The accrual rate
is used to simulate the operating characteristics of the design. By default,
TITE-BOIN assumes a uniform prior for the time to toxicity (i.e., the time to
toxicity is uniformly distributed over the assessment window). As described
previously, this assumption seems strong, but the design is remarkably ro-
bust to the violation of this assumption. Therefore, in general, we recommend
using this default prior. In the case that there is reliable prior information
on the distribution of the time to toxicity, we can incorporate that prior in-
formation by specifying prior DLT probabilities over the trimesters of the
assessment window (see Section 5.4.2). For example, Figure 5.7 sets the prior
DLT probabilities as 0.3, 0.5, and 0.2 for the trimesters of the assessment win-
dow to reflect that DLT is more likely to occur in the middle of the assessment
window.
Late-Onset Toxicity 115

FIGURE 5.7: Specify an informative prior for the time to toxicity.

Overdose Control This panel (see Figure 5.8) specifies the overdose control
rule, described in Section 5.4. That is, if Pr(πj > φ | Dj ) > 0.95 and nj ≥ 3,
dose level j and higher are eliminated from the trial, and the trial is termi-
nated if the lowest dose level is eliminated. When the trial is terminated due
to toxicity, no dose should be selected as MTD. see Section 3.6 for a detailed
discussion on how to specify overdose control through this panel to accommo-
date various trial objectives.

FIGURE 5.8: Specify the overdose control rules.

116 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 5.9: Adjust the cutoff of the accrual suspension rule.

TITE-BOIN additionally has an accrual suspension rule. To avoid risky

decisions caused by sparse data, TITE-BOIN suspends the accrual by default
if more than γ = 50% patient’s DLT outcomes are pending. This default value
generally works well. For some trials, if desirable, γ may be calibrated using
simulation. A smaller value of γ strengthens safety and design reliability, but
prolongs the trial duration. The software allows users to specify the value of
γ (Figure 5.9), but requires γ ≤ 65% for safety.
After completing the specification of trial parameters, the decision table
will be generated by clicking the “Get Decision Table” button. The decision
table will be automatically included in the protocol template in Step 3, but
can also be saved as a separate csv, Excel, or pdf file in this step, if needed.

Step 2: Run simulation

Operating Characteristics This step generates the operating characteristics
of the design through simulation. The scenarios (i.e., the true DLT probability
at each dose, as shown in the lower panel of Figure 5.10) used for simulation
should cover various possible clinical scenarios (e.g., MTD located at different
dose levels). To facilitate the generation of the operating characteristics of
the design, the software automatically provides a set of randomly generated
scenarios with various MTD locations, which are often adequate for most tri-
als. Depending on the application, users can add or remove scenarios. The
software provides three distributions to simulate the time to toxicity: the
Late-Onset Toxicity 117

FIGURE 5.10: Simulate operating characteristics of the design.

uniform, Weibull, or log-logistic distribution (see Figure 5.10). When the

Weibull or log-logistic distribution is selected, users should additionally spec-
ify the probability that DLT occurs in the later half of the assessment win-
dow, which quantifies the degree of late-onset. With these specifications, the
118 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 5.11: Download protocol templates.

parameters of the time-to-toxicity distributions can be determined and used to

simulate the event times. The software also provides an option to include the
rolling six design for comparison to facilitate communication with clinicians
who are more familiar with the conventional design. The simulation results
will be automatically included as a table in the protocol template in the next
step, but they can also be saved as a separate csv or Excel file, if needed.

Step 3: Generate protocol template

Protocol Preparation The TITE-BOIN software generates sample texts and
a protocol template to facilitate the protocol write-up. The protocol template
can be downloaded in various formats (see Figure 5.11). To use this mod-
ule requires the completion of Steps 1 and 2. Once the protocol is approved
by regulatory bodies (e.g., Institutional Review Board), we follow the design
decision table to conduct the trial and make adaptive decisions (e.g., dose
escalation/stay/de-escalation).
To facilitate the trial conduct, the software provides the STFT Calculator
to compute the key parameter (i.e., the standardized total follow-up time),
for determining dose escalation/de-escalation. After users enter the follow-up
times for pending patients and the DLT assessment window, the calculator
outputs STFT. When the trial completes, use the Select MTD tab to select
MTD.
6
Incorporating Historical Data

6.1 Historical data and prior information

Incorporating historical data or real-world data (RWD) has great potential
to improve the efficiency of phase I clinical trials and to accelerate drug de-
velopment. The US Food and Drug Administration (FDA) released a draft of
guidelines for submitting documents using RWD or evidence to the FDA for
drugs and biologics (FDA, 2019).
When designing phase I trials, prior information is often available from
previous studies. For example, the drug to be investigated has been studied
previously in other indications, or similar drugs belonging to the same class
have been studied in earlier phase I trials (Zohar et al., 2011). Phase I bridging
trials are another example. These trials extend a drug from one ethnic group
(e.g., Caucasian) to another (e.g., Asian) (Liu et al., 2015) or from adult
patients to pediatric patients (Petit et al., 2018). In such cases, dose–toxicity
data from the original trial in one ethnic group or adult patients can be used
to inform the design of the subsequent bridging trials, see for example Liu
et al. (2015), Morita (2011), and Li and Yuan (2020).
In this chapter, we first describe how to incorporate prior information in
the model-based continual reassessment method (CRM) based on the intu-
itive concepts of “skeleton” and prior effective sample size (PESS), and then
describe how the similar approach can be used for model-assisted designs,
including the BOIN and keyboard designs.

6.1.1 Incorporate prior information in CRM

Let πj denote the true DLT probability of the jth dose of the J doses under
investigation, j = 1, . . . , J. We assume that the historical data or RWD have
been summarized in the form of the prior estimate of πj , denoted as qj , j =
1, · · · , J. The prior estimate qj can be obtained by fitting a statistical model
(e.g., a logistic model or nonparametric model discussed in Liu et al. (2015)) to
the historical data, or provided by clinicians based on their clinical experience.
It is straightforward to incorporate the prior information into CRM
through the following power model:
exp{α}
πj = q j , for j = 1, . . . , J, (6.1)

DOI: 10.1201/9780429052781-6 119

120 Model-Assisted Bayesian Designs for Dose Finding and Optimization

where α is the unknown parameter. Recall that q1 < · · · < qJ are known as
the skeleton, see Section 2.5. Under the Bayesian paradigm, we assign α a
normal prior f (α) = N(0, σ 2 ), where σ 2 is a prespecified hyperparameter.
This model a priori centers the dose–toxicity curve (π1 , · · · , πJ ) around
the skeleton (q1 , · · · , qJ ). The value of σ 2 controls the amount of informa-
tion borrowed from historical data via the skeleton. A smaller value leads to
stronger borrowing. If σ 2 = 0, the prior completely dominates the observed
data with πj ≡ qj regardless of the observed data.
To make the decision of dose escalation and de-escalation, CRM up-
dates the posterior estimate of πj based on the observed interim data D =
{(nj , yj ), j = 1, . . . , J}:

L(D | α)f (α)

Z
exp(α)
π̂j = qj R dα,
L(D | α)f (α)dα
QJ n exp(α) oyj n onj −yj
exp(α)
where L(D | α) = j=1 qj 1 − qj is the likelihood func-
tion. Then, CRM assigns the next cohort of patients at the dose whose π̂j is
closest to the target toxicity rate φ. In practice, we typically impose safety
rules, such as starting at the lowest dose level and no dose skipping during
dose transition, see Section 2.5 for details.
To choose an appropriate σ 2 , it is of great importance to quantify how
much information is borrowed from historical data with the prior f (α) =
N(0, σ 2 ). Zhou et al. (2021a) proposed a simple and intuitive approach to
formally quantify the information borrowed through the skeleton using the
concept of PESS. The strategy is to approximate the prior distribution of
πj , induced by model (6.1) and f (α), with a beta distribution by matching
the first and second moments. Given skeleton (q1 , · · · , qJ ) and prior f (α), the
mean µj and variance τj2 of πj is given by
Z Z
µj = πj f (πj )dπj , τj = πj2 f (πj )dπj − µ2j ,
2

where f (πj ) is the prior distribution of πj induced by the prior distribution

of f (α) = N(0, σ 2 ) and the model (6.1), given by
 h  i2 
log(πj )
1

 log log(qj )

 1
f (πj ) = √ exp − 2
.
2πσ 
 2σ  πj log(πj )


Matching the first and second moments, f (πj ) is approximated by Beta(aj , bj ),

where
µ2j (1 − µj ) aj (1 − µj )
aj = − µj , bj = . (6.2)
τj2 µj
Incorporating Historical Data 121

As yj ∼ Binomial(nj , πj ), the posterior of πj is f (πj |D) = Beta(aj +

yj , bj + nj − yj ), therefore the PESS of f (πj ) is aj + bj . An alternative more
sophisticated method of calculating PESS is provided by Morita et al. (2008).
This reveals a property of CRM: once the prior f (α) is specified, PESS
for each dose is automatically determined because πj is a function of α. For
example, given skeleton (q1 , · · · , q5 ) = (0.10, 0.19, 0.30, 0.42, 0.54) and prior
f (α) = N(0, 0.72), PESS is (3, 3, 3, 3.1, 3.4) for the five doses. Since the dose–
toxicity model of CRM usually has degrees of freedom smaller than the number
of doses under investigation, CRM does not allow users to specify dose-specific
prior information or PESS. However, in practice, we often have an unequal
amount of prior information for different doses. For example, we often have
more data at the doses that are below and around MTD from historical phase
I trials. In this case, it is highly desirable to be able to specify a different PESS
for each unique dose according to the historical data. Most model-based phase
I designs, e.g., the designs based on escalation with overdose control (Babb
et al., 1998) or the Bayesian logistic regression model (Neuenschwander et al.,
2008), share this limitation.

6.2 BOIN with Informative Prior (iBOIN)

6.2.1 Trial design
We now discuss how to use the skeleton, coupled with PESS, to incorporate
prior information into the BOIN design. Let π̂j = yj /nj denote the observed
DLT rate at the current dose, and λej and λdj denote the dose escalation and
de-escalation boundaries, respectively. As shown in Section 3.4.2 and using
the same notations, BOIN’s optimal λej and λdj for the observed DLT rate
π̂j are given by
   
1 − φ1 ω1j 

−1

 log + n j log 
 1−φ ω 
λej = max 0,   0j , (6.3)
 φ(1 − φ1 ) 

 log 

φ1 (1 − φ)
   
1−φ ω0j 
 

 log + n−1j log

 1 − φ2 ω 
λdj = min 1,   2j , (6.4)
 φ2 (1 − φ) 

 log 

φ(1 − φ2 )

with the constraint λej ≤ λdj . When the non-informative prior ω0j = ω1j =
ω2j = 1/3 is used, which is recommended for most trials, the optimal
122 Model-Assisted Bayesian Designs for Dose Finding and Optimization

escalation and de-escalation boundaries become

   
1 − φ1 1−φ
log log
1−φ 1 − φ2
λe ≡ λej =  , λd ≡ λdj =  ,
φ(1 − φ1 ) φ2 (1 − φ)
log log
φ1 (1 − φ) φ(1 − φ2 )
which are independent of j and nj . These are boundaries used in the standard
BOIN, see Section 3.4. It can be shown that λe ≤ λd .
When historical data (or reliable prior information) are available, they can
be incorporated into BOIN based on the following steps:
1. Estimate skeleton (q1 , · · · , qJ ) from the historical data and elicit
corresponding PESS (n01 , · · · , n0J ), where n0j is the desirable PESS
for dose level j, j = 1, · · · , J.
2. Determine the informative prior for Hk , i.e., ωkj , j = 0, 1, 2, as
n0j
ϕxk (1 − ϕk )n0j −x
 
X n0j x
ωkj = P2 qj (1 − qj )n0j −x , (6.5)
x=0 k 0 =0 ϕx
k 0 (1 − ϕk 0)
n0j −x x

where ϕ0 = φ, ϕ1 = φ1 , and ϕ2 = φ2 .
3. Calculate (λej , λdj ) using formulas (6.3) and (6.4). In the rare case
that formulas (6.3) and (6.4) result in a solution λej > λdj , then de-
termine the optimal (λej , λdj ) using a numerical search to minimize
the decision error.
The derivation of ωkj in Step 2 is as below. For dose j, the predictive proba-
bility of Hkj based on the prior DLT rate qj and PESS nj can be expressed
as

ωkj = Pr(Hkj | n0j , qj )

n0j
X
= Pr(Hkj | x) Pr(x | n0j , qj )
x=0
n0j
X Pr(x | Hkj ) Pr(Hk )
= P2 Pr(x | n0j , qj )
x=0 Pr(x | Hk0 ) Pr(Hk0 )
k0 =0
n0j
ϕxk (1 − ϕk )n0j −x
 
X n0j x
= P2 qj (1 − qj )n0j −x .
x=0 0
k =0 k ϕ x
0 (1 − ϕk
n
0 ) 0j
−x x

The last equality follows by assuming Pr(H0 ) = Pr(H1 ) = Pr(H2 ) = 1/3.

In Step 3, under certain extreme specifications of the prior information
(e.g., extremely informative prior), it is possible that the λej > λdj based
on formulas (6.3) and (6.4). In this case, the escalation and de-escalation
boundaries (λej , λdj ) should be re-determined using a numerical search to
minimize the decision error, given by equation (3.3) in Section 3.4, under the
Incorporating Historical Data 123

constraint of λej ≤ λdj . This numerical search is fast and straightforward

given that both the possible numbers of patients treated and DLTs at a dose
are limited (e.g., < 20).
We refer to the resulting design (with informative prior) based on (λej , λdj )
as iBOIN. Because of the incorporation of the informative prior information,
the escalation and de-escalation boundaries λej and λdj of iBOIN depend
on the dose level j, as well as nj . When (n01 , · · · , n0J ) = 0, the value of
ωkj becomes 1/3 using the above equation, as a result iBOIN reduces to the
standard BOIN design in this case.
Figure 6.1 contrasts the boundaries under a non-informative prior and
those under an informative prior for a trial with five doses and an elicited
skeleton (0.10, 0.19, 0.30, 0.42, 0.54) when the target DLT probability is 0.3
and PESS is 3. For example, because the prior information says that the
lowest dose is below the true MTD (with the prior DLT probability of 0.10),
its escalation boundary λej is higher than that of the non-informative prior
to encourage dose escalation. On the contrary, because the prior information
says the highest dose is above MTD (with the prior DLT probability of 0.54),
its de-escalation boundary λdj is lower than that of the non-informative prior
to encourage dose de-escalation.
Compared to CRM/EWOC/BLRM, iBOIN is more flexible and allows
users to accurately incorporate prior information by specifying PESS for each
dose. For example, given a phase I trial with five doses, if historical data
provide more information on the first two doses than the last two doses and
most information on dose level 3, we could specify the five doses’ PESS as

boundaries using non-informative prior boundaries using informative prior

2 4 6 8 2 4 6 8

q1 q2 q3 q4 q5
1.0

0.8
Probability

0.6

0.4

0.2

0.0

2 4 6 8 2 4 6 8 2 4 6 8

Number of patients treated at current dose (nj)

FIGURE 6.1: Dose escalation and de-escalation boundaries under the

non-informative prior and an informative prior with the skeleton
(q1 , q2 , q3 , q4 , q5 ) = (0.10, 0.19, 0.30, 0.42, 0.54), PESS = 3, and φ = 0.3.
124 Model-Assisted Bayesian Designs for Dose Finding and Optimization

(3, 3, 6, 1, 1) to reflect that. As described previously, this is not possible

under CRM.
The other advantage of iBOIN, as a model-assisted design, is that the dose
escalation and de-escalation rule can be pre-tabulated and included in the trial
protocol. Table 6.1 shows the decision table of iBOIN with the skeleton (0.10,
0.19, 0.30, 0.42, 0.54) and PESS n01 = · · · = n05 = 3. This decision table
is equivalent to the rule based on λej and λdj , but easier to use in practice.
Users only need to identify the row corresponding to the current dose level,
and then they can use the boundaries listed in that row to easily make the
decision of dose escalation and de-escalation.
In summary, the dose-finding rules of the iBOIN design can be described
as follows:
1. Patients in the first cohort are treated at the lowest dose d1 , or the
physician-specified dose.
2. Given data (nj , yj ) observed at the current dose level j, make the
decision of escalation/de-escalation according to the iBOIN decision
table (e.g., Table 6.1) for treating the next cohort of patients.
3. Repeat Step 2 until the prespecified maximum sample size is
reached, and then select MTD as the dose whose isotonic estimate
of πj is closest to φ.
For the purpose of overdose control, following BOIN, the iBOIN design
imposes a dose elimination rule: if Pr(πj > φ | nj , yj ) > 0.95 and nj ≥ 3, dose

TABLE 6.1: iBOIN decision boundaries up to 30 patients with a cohort size of

three, given the skeleton (q1 , · · · , q5 ) = (0.10, 0.19, 0.30, 0.42, 0.54) and PESS
n01 = · · · = n05 = 3. The target DLT probability φ = 0.3.

Dose No. of patients at current dose

level Action∗ 3 6 9 12 15 18 21 24 27 30
Escalate if no. of DLT ≤ 1 1 2 3 4 4 5 6 6 7
1
De-escalate if no. of DLT ≥ 2 3 4 5 7 8 9 10 11 12
Escalate if no. of DLT ≤ 0 1 2 3 3 4 5 5 6 7
2
De-escalate if no. of DLT ≥ 2 3 4 5 6 7 8 9 11 12
Escalate if no. of DLT ≤ 0 1 2 2 3 4 4 5 6 7
3
De-escalate if no. of DLT ≥ 2 3 4 5 6 7 8 9 10 11
Escalate if no. of DLT≤ 0 1 1 2 3 3 4 5 6 6
4
De-escalate if no. of DLT ≥ 1 2 3 4 6 7 8 9 10 11
Escalate if no. of DLT ≤ 0 0 1 2 2 3 4 5 5 6
5
De-escalate if no. of DLT ≥ 1 2 3 4 5 6 7 8 10 11
*When neither “Escalate” nor “De-escalate” is triggered, stay at the
current dose for treating the next cohort of patients.
Incorporating Historical Data 125

level j and higher are eliminated from the trial, where Pr(πj > φ | nj , yj ) is
evaluated based on the beta-binomial model with the Unif(0, 1) prior. As the
objective of the dose elimination rule is to protect patients from excessively
toxic doses, it is sensible to use the uniform prior to evaluate this rule to avoid
potential bias due to misspecification of the prior. As discussed in Section
3.4.5, if desirable, a vague prior such as Beta(0.02, 0.08) with the PESS = 0.1
can be also used to obtain similar operating characteristics on dose elimination
by slightly adjusting the overdose control probability cutoff 0.95 (e.g., to 0.93).
The trial is terminated if the lowest dose level is eliminated.

6.2.2 Practical guidance

Choosing PESS
As the skeleton is typically estimated based on the history data, when
applying iBOIN, the key question is: “How to specify PESS?” PESS should
be chosen to reflect the appropriate amount of prior information to be incor-
porated, which depends on the reliability of the prior information and varies
from trial to trial. When there is strong evidence that the prior is most likely
to be specified correctly, it is appropriate to use a large PESS to borrow more
information; when there is a great amount of uncertainty regarding whether
the prior is most likely correctly specified, we may use a small PESS to avoid
bias.
In practice, there is often sizable uncertainty on the reliability of the prior
information. Thus, PESS should be chosen carefully to achieve an appropriate
balance between design performance and robustness. Using a large PESS im-
proves the design performance (i.e., the accuracy to identify MTD) when the
prior is correctly specified, but may lead to a substantial loss of performance
when the prior is grossly misspecified. Based on numerical studies, Zhou et al.
(2021a) recommended PESS ∈ [1/3(N/J), 1/2(N/J)] as the default value that
improves trial performance while maintaining reasonable robustness. For ex-
ample, when J = 5 and N = 30, the recommended value for PESS is n0j = 2
or 3 (i.e., across five doses, the total PESS is 10 or 15). The value of n0j can
be further calibrated by simulation using the software described in Section 6.5.

Robust prior
In general, iBOIN is robust to the misspecification of prior information
(i.e., prior-data conflict), especially when PESS is chosen as described above.
However, when the informative prior is grossly misspecified (e.g., the prior
estimate of MTD and the true MTD differ by three dose levels), it may com-
promise the accuracy of identifying MTD. To further strengthen the robust-
ness of iBOIN, Zhou et al. (2021a) proposed a robust prior, which is easy to
implement and yields superior operating characteristics.
Given the elicited skeleton (q1 , · · · , qJ ) with dose level j ∗ as the prior
estimate of MTD (i.e., qj ∗ = φ), the robust prior is the same as the prior
described above when j ∗ < J/2, but modifies PESS to (n01 , · · · , n0j ∗ , 0, · · · , 0)
when j ∗ ≥ J/2. In other words, when prior MTD j ∗ ≥ J/2, the robust prior
126 Model-Assisted Bayesian Designs for Dose Finding and Optimization

uses informative prior information for the dose up to the prior MTD estimate,
and after that it uses the non-informative prior.
The rationale of this robust prior is that the dose finding is a sequential
process of allocating patients from low doses to high doses. Thus, by the time
that dose finding reaches high doses, there is an extremely limited sample
size remaining to override the prior if it is misspecified. The robust prior
modifies the prior of high doses to be non-informative to facilitate overriding
the prior when the data conflict with the prior, thus alleviating the impact of
prior misspecification. This prior is particularly useful when there is a great
amount of uncertainty regarding the prior information. Zhou et al. (2021a)
also considered another robust prior, which is a mixture of informative and
non-informative priors. This mixture robust prior is conceptually appealing,
but more complicated and does not perform as well as the simple robust prior
described above.

6.3 iKeyboard design

It is straightforward to incorporate the prior information into the keyboard
design using the skeleton and PESS. As described in Section 3.3, the keyboard
design assumes a beta-binomial model,
yj | nj , πj ∼ Binomial(nj , πj ),
πj ∼ Beta(aj , bj ), (6.6)
where aj and bj are hyperparameters. Let Dj = (nj , yj ), the posterior distri-
bution of πj arises as
πj | Dj ∼ Beta(yj + aj , nj − yj + bj ), for j = 1, . . . , J. (6.7)
The keyboard design divides the toxicity probability line (0, 1) into a series
of equal width intervals, and makes the decision of dose escalation and de-
escalation by comparing the location of the strongest key Imax , defined as the
interval with the highest posterior probability, with the target key I ∗ = (δ1 , δ2 )
(i.e., the pre-specified target dosing interval). If Imax is located on the left side
of I ∗ (denoted as Imax ≺ Itarget ), escalate the dose; if Imax is located on the
right side of I ∗ (denoted as Imax  Itarget ), de-escalate the dose; if Imax is
I ∗ (denoted Imax ≡ Itarget ), retain the current dose. See Section 3.3 for more
details.
In the beta-binomial model (6.6), aj + bj can be interpreted as PESS.
Thus, the prior information can be incorporated into the keyboard design as
follows:
1. Estimate skeleton (q1 , · · · , qJ ) from the historical data and elicit
corresponding PESS (n01 , · · · , n0J ), where n0j is the desirable PESS
for dose level j, j = 1, · · · , J.
Incorporating Historical Data 127

2. Determine hyperparameters aj and bj in the beta prior (6.6) as

follows:

aj = n0j qj ; bj = n0j (1 − qj ), j = 1, · · · , J.

3. Make dose escalation and de-escalation based on the resulting pos-

terior given by equation (6.7).
We refer to the keyboard design with an informative prior as the iKeyboard de-
sign. Given a fixed maximum sample size, all possible outcomes Dj = (nj , yj )
can be enumerated. For each possible outcome, the posterior distribution
f (πj |Dj ) can be calculated. Therefore, the dose escalation/de-escalation rule
of iKeyboard can be tabulated, similar to iBOIN. The above approach is di-
rectly applicable to the mTPI design for incorporating prior information.

6.4 Operating characteristics

In this section, we briefly describe a simulation study to evaluate the operating
characteristics of the iCRM (i.e., CRM with informative prior), iBOIN, and
iKeyboard designs. A comprehensive simulation study (including comparison
using random scenarios) is provided by Zhou et al. (2021a). Consider J = 5
doses and the target DLT probability φ = 0.3. The maximum sample size is
N = 30 with a cohort size of three. For iBOIN and iKeyboard, we set PESS
n0j = 3 for j = 1, · · · , 5; and for iCRM, the prior is chosen such that PESS at
the prior MTD is three. All the designs use the same skeletons (i.e., the prior
DLT probabilities), which are provided in Table 6.2. The counterparts of the
designs with a non-informative prior (denoted as CRM, BOIN, and keyboard)
are included for comparison.
Table 6.3 shows the results, including (1) percentage of correct selection
(PCS), defined as the percentage of simulated trials in which MTD is correctly
identified; (2) percentage of patients treated at MTD; (3) percentage of pa-
tients treated above MTD; (4) risk of overdosing, defined as the percentage of
simulated trials that assigned 50% or more patients to doses above MTD; and
(5) risk of poor allocation, defined as the percentage of simulated trials that
assigned fewer than six patients to MTD. As noted by Zhou et al. (2018b),
metrics (4) to (5) measure the reliability of the design, i.e., the likelihood of a
design demonstrating extremely problematic behaviors (e.g., treating 50% or
more patients at toxic doses, or fewer than six patients at MTD), which are
of great practical importance. Note that the percentage of patients overdosed
(i.e., metric (3)) does not completely capture the risk of overdosing (i.e., met-
ric (4)). Two designs can have a similar percentage of patients overdosed, but
rather different risks of overdosing 50% of the patients.
128 Model-Assisted Bayesian Designs for Dose Finding and Optimization

TABLE 6.2: Four dose–toxicity scenarios with target DLT probability φ =

0.30. The prior MTDs are correctly specified in Scenarios 1 and 2 and mis-
specified in Scenarios 3 and 4.

Dose level
1 2 3 4 5
Scenario 1
True Pr(DLT) 0.30 0.42 0.50 0.60 0.65
Prior Pr(DLT) 0.30 0.42 0.54 0.64 0.73
Scenario 2
True Pr(DLT) 0.15 0.27 0.40 0.50 0.65
Prior Pr(DLT) 0.19 0.30 0.42 0.54 0.64
Scenario 3
True Pr(DLT) 0.09 0.12 0.15 0.30 0.45
Prior Pr(DLT) 0.01 0.04 0.10 0.19 0.30
Scenario 4
True Pr(DLT) 0.08 0.15 0.31 0.45 0.55
Prior Pr(DLT) 0.19 0.30 0.42 0.54 0.64

The simulation results show that (i) incorporating prior information im-
proves the accuracy of identifying MTD for all designs; (ii) iBOIN and iCRM
have similar performance in identifying MTD and allocating patients to MTD,
but iBOIN has a lower risk of overdosing patients and poor allocation; (iii)
Compared to iBOIN, iKeyboard has a larger variation in identifying MTD
(e.g., performs best in scenario 3, but worst in scenario 4), and a higher risk
of overdosing patients and poor allocation.

6.5 Software and case study

The web application for iBOIN is available as a module of the “BOIN Suite”
at http://www.trialdesign.org. Figure 6.2 shows the launchpad of BOIN
Suite.

Non-small Cell Lung Cancer (NSCLC) Cancer Trial The objective of

this phase I trial (ClinicalTrials.gov Identifier: NCT04479306) is to determine
MTD and RP2D of alisertib in combination with osimertinib in treating pa-
tients with EGFR mutated stage IIIB or IV NSCLC. Toxicity will be evaluated
according to the National Cancer Institute Common Terminology Criteria for
Adverse Events (NCI CTCAE) version 5.0. DLT will be assessed during the
first 28-day cycle of combination therapy. The maximum sample size is 12 and
patients are treated in a cohort size of three. Three doses of alisertib will be
Incorporating Historical Data 129

investigated in combination with osimertinib, and the target DLT probability

is φ = 30%.

We use this trial as an example to illustrate how to use iBOIN to leverage the
prior information to improve the efficacy of phase I trials. After selecting and

TABLE 6.3: Operating characteristics of iCRM, iBOIN, and iKeyboard in

comparison with their counterparts with non-informative priors. PCS is the
percentage of correction selection, ROD is the risk of overdosing, and RPA is
the risk of poor allocation.

% Pts % Pts
Design PCS at MTD >MTD ROD RPA
Scenario 1
CRM 54.8 59.9 27.8 23.2 12.2
iCRM 63.1 65.2 24.9 19.4 9.8
BOIN 59.2 59.6 29.0 23.6 10.2
iBOIN 64.2 66.2 22.4 12.8 4.5
Keyboard 59.2 59.3 29.3 23.6 10.2
iKeyboard 64.2 50.7 39.6 34.2 17.8
Scenario 2
CRM 51.6 36.1 7.7 29.5 25.2
iCRM 53.3 42.4 5.6 23.7 16.8
BOIN 50.6 41.1 6.0 23.0 17.1
iBOIN 57.8 47.6 3.7 10.4 8.6
Keyboard 50.2 41.1 6.0 23.0 16.7
iKeyboard 59.6 37.8 6.6 35.1 23.6
Scenario 3
CRM 50.7 29.9 14.6 9.3 32.9
iCRM 57.3 33.8 17.0 13.0 28.2
BOIN 51.5 28.6 13.1 1.2 24.6
iBOIN 58.6 35.5 18.4 3.8 11.8
Keyboard 52.1 8.6 13.1 1.2 24.6
iKeyboard 59.5 9.2 27.9 11.2 17.9
Scenario 4
CRM 58.0 38.1 21.7 17.3 21.8
iCRM 59.8 38.0 18.3 13.4 21.3
BOIN 52.3 35.6 17.0 7.9 19.2
iBOIN 61.6 33.0 10.9 2.2 14.8
Keyboard 52.4 35.7 17.1 7.9 18.9
iKeyboard 56.4 45.4 15.3 3.6 6.8
130 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 6.2: The launchpad of web application BOIN Suite available at

http://www.trialdesign.org.

launching the “iBOIN” module from the BOIN Suite launchpad, we design
the trial by taking the following three steps:
Step 1: Enter trial parameters
Doses, Sample Size, and Target As shown in Figure 6.3, the number of
doses under investigation is three, starting dose level is one. The starting dose
level can be adjusted according to prior information, e.g., one level below the
prior MTD estimated using historical data. For example, in some trials, the
appropriate starting dose level may be two or even higher.
The cohort size for the trial is three and the number of cohorts is four, with
a total sample size of 12. As a rule of thumb, we recommend the maximum
sample size N = 6 × J (i.e., the maximum sample size of the 3+3 design) as
the total sample size, where J is the number of doses. This trial uses a smaller
sample size due to the limited accrual and the intention to shorten the trial
duration. In this case, leveraging available prior information provides a useful
approach to compensate the small sample size.
To reduce the sample size, it is often useful to use the “convergence”
stopping rule: stop the trial early when m patients have been assigned to a
dose and the decision is to stay at that dose. This stopping criterion suggests
that the dose finding approximately converges to MTD, thus the trial can
be stopped. We recommend m = 9 or larger. In this trial, m = 9 is used.
Because of the early-stopping rule, the actual sample size used in the trial
is often smaller than N . The saving in the sample size depends on the true
Incorporating Historical Data 131

FIGURE 6.3: Specify doses, sample size, target, convergence stopping rule,
and accelerated titration.

dose–toxicity scenario and can be evaluated using simulation. In general, a

larger sample size saving is expected when the starting dose is close to the
target.
The target DLT probability is φ = 0.3, which should be elicited from clini-
cians. If a specific safety requirement is desirable, the values for φ, φ1 , and φ2
can be further adjusted. In addition, as described previously, the accelerated
titration option is another tool to reduce the sample size, see Section 3.6 for
details.

Overdose Control This panel (see Figure 6.4) specifies the overdose control
rule, described in Section 5.4. That is, if Pr(πj > φ | Dj ) > 0.95 and nj ≥ 3,
dose level j and higher are eliminated from the trial, and the trial is termi-
nated if the lowest dose level is eliminated. When the trial is terminated due
to toxicity, no dose should be selected as MTD. In general, we recommend
to use the default probability cutoff 0.95. A smaller value, e.g., 0.9, results in
132 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 6.4: Specify the overdose control rules.

stronger overdose control, but at the cost of reducing the probability of cor-
rectly identifying MTD. This is because, in order to correctly identify MTD,
it is imperative to explore the doses sufficiently to learn their toxicity profile.
The option “ Check to impose a more stringent safety stopping rule on the
lowest dose” can be used to increase the early stopping probability when all
doses are toxic; see Section 3.6 for a detailed discussion on how to use this
option.

Prior and PESS This panel (see Figure 6.5) specifies the prior estimate of
toxicity probability at each dose (i.e., the skeleton estimated based on the his-
torical data) and PESS. Based on the historical data on alisertib and osimer-
tinib, the combination is expected to be safe with little overlapped toxicities,

FIGURE 6.5: Specify the skeleton and PESS.

Incorporating Historical Data 133

and thus the skeleton is set as (0.05, 0.12, 0.25) for the three doses. Following
the guidance described in Section 6.2.2, PESS is set as two for the three doses,
leading to a prior being informative without dominating the trial data. For
many trials, it is desirable to use the robust prior to obtain extra robustness.
This can be done by checking the corresponding activation box. For this trial,
the robust prior is identical to the informative prior, as the prior estimate of
MTD is the last dose.
After completing the specification of trial parameters, the decision table
will be generated by clicking the “Get Decision Table” button. The decision
table will be automatically included in the protocol template in Step 3, but
can also be saved as a separate csv, Excel or pdf file in this step if needed.
The resulting dose escalation and de-escalation boundaries are shown in
Table 6.4. As the prior information suggests that dose level one is far be-
low MTD, the escalation boundary of dose level one (e.g., escalate if ≤ 1/3
DLT) is higher than that of dose levels two and three (e.g., escalate if ≤ 0/3
DLT), making it easier to escalate at dose level 1. When more patients are
treated, the trial data start to dominate the prior, and thus the escalation
and de-escalation boundaries become the same for three doses when six or
nine patients are treated at a dose. These boundaries eventually shrink to the
standard BOIN boundaries when the sample size is sufficiently large.

Step 2: Run simulation

Operating Characteristics This step generates the operating characteristics
of the design through simulation, see Figure 6.6. The scenarios used for sim-

TABLE 6.4: iBOIN decision boundaries for the lung cancer trial, given the
skeleton (q1 , · · · , q5 ) = (0.05, 0.12, 0.25) and PESS n01 = n02 = n03 = 2. The
target DLT probability φ = 0.3.

Dose No. of patients at current dose

∗
level Action 3 6 9
Escalate if No. of DLT ≤ 1 1 2
1 De-escalate if No. of DLT ≥ 2 3 4
Eliminate if No. of DLT ≥ 3 4 5
Escalate if No. of DLT ≤ 0 1 2
2 De-escalate if No. of DLT ≥ 2 3 4
Eliminate if No. of DLT ≥ 3 4 5
Escalate if No. of DLT ≤ 0 1 2
3 De-escalate if No. of DLT ≥ 2 3 4
Eliminate if No. of DLT ≥ 3 4 5
*When neither “Escalate” nor “De-escalate” is triggered, stay at the
current dose for treating the next cohort of patients.
134 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 6.6: Simulate operating characteristics of the design.

ulation should cover various possible clinical scenarios, e.g., MTD is located
at different dose levels. To facilitate the generation of the operating charac-
teristics of the design, the software automatically provides a set of randomly
generated scenarios with various MTD locations, which are often adequate
for most trials. Depending on the application, users can add or remove sce-
narios. The software also provides an option to include the 3+3 design as a
comparator to facilitate communication with clinicians who are more familiar
with the conventional design. Table 6.5 shows the operating characteristics of
iBOIN compared with those of BOIN using the non-informative prior. When
the MTD prior estimate is correctly specified, iBOIN improves the percent-
age of correct selection from 61.4% to 72.5%. The simulation results will be
Incorporating Historical Data 135

TABLE 6.5: Operating characteristics of the iBOIN design for the lung cancer
trial in comparison with the BOIN with the non-informative prior. The dose
in bold is MTD.

Dose level
Design 1 2 3
Pr(DLT) 0.05 0.13 0.3
BOIN Selection (%) 1.5 37.1 61.4
% of Patients 30.6 38.3 31.0
iBOIN Selection (%) 1.4 26.1 72.5
% of Patients 27.2 39.2 33.6

Pr(DLT) 0.13 0.3 0.44

BOIN Selection (%) 23.8 53.3 22.5
% of Patients 43.1 43.1 13.8
iBOIN Selection (%) 15.0 53.5 30.9
% of Patients 36.3 47.1 16.6

Pr(DLT) 0.3 0.48 0.67

BOIN Selection (%) 60.2 23.4 2.9
% of Patients 68.5 28.7 2.8
iBOIN Selection (%) 54.0 31.3 3.9
% of Patients 56.7 38.8 4.5

automatically included as a table in the protocol template in the next step,

but can also be saved as a separate csv or Excel file if needed.

Step 3: Generate protocol template

Protocol Preparation The iBOIN software generates sample texts and a pro-
tocol template to facilitate the protocol write-up. The protocol template can
be downloaded in various formats (see Figure 6.7). Use of this module requires
the completion of Steps 1 and 2. Once the protocol is approved by regulatory

FIGURE 6.7: Download protocol templates.

136 Model-Assisted Bayesian Designs for Dose Finding and Optimization

bodies (e.g., Institutional Review Board), we follow the design decision table to
conduct the trial and make adaptive decisions (e.g., dose escalation/stay/de-
escalation). At the completion of the trial, the “Select MTD” function can be
used to determine MTD based on the observed trial data.
7
Multiple Toxicity Grades

7.1 Multiple toxicity grades

The landscape of oncology drug development has recently changed with the
emergence of molecularly targeted agents and immunotherapies. These new
therapeutic agents appear more likely to induce multiple low- or moderate-
grade toxicities rather than the dose-limiting toxicity (DLT) (Brahmer et al.,
2010; Le Tourneau et al., 2010; Penel et al., 2011). To accurately account for
the side effects of the agents, it is important to incorporate the grade of toxi-
city into dose finding and decision making. The widely used National Cancer
Institute Common Terminology Criteria for Adverse Events (NCI CTCAE)
scores toxicity as a five-grade ordinal variable. Its general guidelines use grade
0 for no toxicity, grade 1 for mild toxicity, grade 2 for moderate toxicity,
grade 3 for severe toxicity, grade 4 for life-threatening toxicity, and grade 5
for toxicity-related death.
The designs introduced so far, including the 3+3, continual reassessment
method (CRM), and Bayesian optimal interval (BOIN) designs, all assume
a single binary outcome: DLT (e.g., grade 3 or higher toxicity) or not. This
dichotomization disregards low-grade adverse events, and it thus may be insuf-
ficient to quantify the safety of the targeted agents and immunotherapies that
cause a high frequency of low- or moderate-grade toxicities with little DLT.
The aggregate effect of multiple lower- or moderate-grade toxicity events, how-
ever, often leads to dose interruption, reduction, and discontinuation, which
reduces the efficacy of the drug in the real-world setting.
Three approaches have been proposed to incorporate toxicity grades into
dose finding, including total toxicity burden, toxicity score, and multiple tox-
icities.
(1)Total toxicity burden
One general approach is to assign a severity weight to each grade and type
of toxicity event, and then combine the weights as a composite score. Bekele
and Thall (2004) proposed the total toxicity burden (TTB) as the arithmetic
sum of different grades and types of toxicity, weighted by the severity weights
elicited from clinicians to reflect the importance among the toxicities that the
clinicians had identified. For example, Table 7.1 shows the severity weights
for the soft tissue sarcoma phase I trial described by Bekele and Thall (2004).
If a patient experienced a grade 3 myelosuppression without fever, grade 3

DOI: 10.1201/9780429052781-7 137

138 Model-Assisted Bayesian Designs for Dose Finding and Optimization

TABLE 7.1: Toxicities and severity weights in the sarcoma trial, reproduced
from Bekele and Thall (2004).

Type of toxicity Grade Severity weight

1 Myelosuppression without fever 3 1.0
4 1.5
Myelosuppression with fever 3 5.0
4 6.0
2 Dermatitis 3 2.5
4 6.0
3 Liver 2 2.0
3 3.0
4 6.0
4 Nausea/vomiting 3 1.5
4 2.0
5 Fatigue 3 0.5
4 1.0

dermatitis, and grade 3 nausea/vomiting, the TTB for that patient is then
calculated as TTB = 1 + 2.5 + 1.5 = 5. The maximum tolerated dose (MTD)
is defined as the dose level where the true TTB value is closest to the target
TTB value specified by the oncologists. To identify the MTD, Bekele and Thall
(2004) built a joint model for the toxicities under the Bayesian framework.
At each decision-making time, the dose that has the posterior expected TTB
closest to the target TTB is selected as the next level.
Similarly, Chen et al. (2010) proposed to map toxicity grades into a quasi-
continuous or continuous endpoint using the normalized equivalent toxicity
score (NETS). Lee et al. (2009) proposed the toxicity burden score (TBS)
to summarize toxicity using a weighted sum, where the severity weights were
estimated via regression based on historical data. Ezzalfani et al. (2013) pro-
posed another flexible toxicity endpoint, called the total toxicity profile (TTP),
which is computed as the Euclidean norm of the severity weights. Although
these composite scores (TTB, NETS, TBS, and TTP) are not strictly continu-
ous, in practice, after appropriate transformation, they can be approximately
regarded as a normally distributed endpoint.

(2) Toxicity score

Several methods have been proposed to “convert” toxicity grades to numeric
scores that reflect their relative severity in the unit of DLT. For example, we
may define a grade 3 toxicity as equivalent to a DLT, a grade 2 toxicity as
equivalent to a 0.5 DLT, and a grade 4 toxicity as equivalent to 2 DLTs. This
is the “equivalent toxicity score” (ETS) approach proposed by Yuan et al.
(2007). Although the (normalized) ETS is a fractional event and does not
really follow a Bernoulli distribution, Yuan et al. (2007) showed that it can
Multiple Toxicity Grades 139

be simply treated as a binary endpoint (more precisely, a quasi-binary end-

point) and modeled using the quasi-Bernoulli likelihood. Yuan et al. (2007)
implemented this quasi-Bernoulli likelihood in the CRM framework, leading
to a quasi-CRM approach.

(3) Multiple toxicities

The aforementioned approaches collapse different types and/or grades of toxi-
city into a single numerical index (e.g., TTB or TTP) to summarize the overall
severity of multiple toxicities. An alternative approach is to retain the origi-
nal scale of the toxicity grade and use multiple toxicity constraints to quantify
the safety of the drug (Lee et al., 2010). For example, the target dose may be
defined as the dose with the grade 2 toxicity rate ≤ 40%, and the grade 3 and
higher toxicity rate ≤ 30%. Lee et al. (2010) extended CRM to account for
multiple toxicity constraints (referred to as the MC-CRM design hereafter)
based on the proportional odds model. This MC-CRM design was further ex-
tended to trials with late-onset outcomes (Lee et al., 2017).

Among the above three approaches, we generally favor approaches 2 and

3. The difficulty of approach 1 is that it is challenging to define the target
TTB (NETS, TBS, and TTP) in practice, as these composite scores lack in-
tuitive clinical interpretation. For example, TTB = 3 does not have direct
interpretation in terms of the toxicity rates in grades and types. In addition,
different toxicity profiles may result in the same TTB, but have dramatically
different clinical implications. Bekele and Thall (2004) and Ezzalfani et al.
(2013) provided some guidance to facilitate the elicitation of the target, but
specifying the target remains complicated and extremely challenging in prac-
tice. In contrast the elicitation of the target for approaches 2 and 3 is similar
to that of conventional DLT-based dose finding designs.
The majority of designs accounting for toxicity grades and types are model-
based. Albeit using different metrics to summarize toxicity, these designs all
employ a similar strategy to CRM. That is, they assume a model to describe
the relationship between toxicity grade/type and the dose; then, based on the
accumulating data, they continuously update the model estimate and make
the decision of dose assignment for the incoming new patient (typically by
assigning the new patient to the dose for which the estimate of the toxicity
metric is closest to the target) until the stopping rule is satisfied (e.g., the
maximum sample size is reached). In order to account for toxicity grade and
type, the model used by the designs is significantly more complex than that
of CRM, e.g., the multinomial model for multiple-level toxicity grade (Lee
et al., 2010), and the joint ordinal model to account for the correlation among
toxicity types (Bekele and Thall, 2004). This makes these designs difficult to
understand and implement, and thus they are rarely used in practice.
In this chapter, we describe two model-assisted designs that are transpar-
ent and easy to implement. The first one is the generalized BOIN (gBOIN)
design, which accommodates continuous and quasi-binary toxicity endpoints,
140 Model-Assisted Bayesian Designs for Dose Finding and Optimization

as well as the standard binary endpoint (Mu et al., 2018). The second method
is the multiple-toxicity BOIN (MT-BOIN) design based on multiple toxicity
constraints (Lin, 2018). Unlike model-based designs, the decision rules of the
gBOIN and MT-BOIN designs can be tabulated prior to the onset of the trial,
making the trial conduct simple and straightforward, similar to the standard
BOIN design (with a binary endpoint).

7.2 gBOIN accounting for toxicity grade

7.2.1 Trial design
Consider a phase I trial with J prespecified doses, d1 < · · · < dJ . Let y denote
the toxicity endpoint of interest, either a binary or quasi-binary outcome (e.g.,
DLT or ETS) or a continuous outcome (e.g., TTB, TBS, or TTP), belonging
to the exponential family of distributions,

f (y|dj ) = h(y) exp{η(θj )T (y) − A(θj )}, (7.1)

where h(·), T (·), η(·), and A(·) are known functions, and θj is a distributional
parameter that can be scalar or vector, depending on the distribution of y
at dose dj . The exponential family of distributions includes many commonly
used distributions, such as normal, binomial, multinomial, Poisson, gamma,
and beta distributions. For example, define µ = E(y) and µj = E(y|dj ), then
• y follows a Bernoulli distribution if θj = µj , η(θj ) = log{µj /(1 − µj )},
A(θj ) = − log(1 − µj ) , T (y) = y, and h(y) = 1.
• y follows a normal distribution if θj = µj , η(θj ) = µj /σj2 , A(θj ) = µ2j /(2σj2 ),
T (y) = y, and h(y) = (2πσj2 )−1/2 exp{−y 2 /(2σj2 )}.
Let φ denote the target value of µ for dose finding. For binary or quasi-
binary toxicity endpoints (DLT or ETS), φ is simply the target DLT proba-
bility; for continuous endpoints (e.g., the TTB, TBS, or TTP), φ is the target
value of the TTB, TBS, or TTP. Let Dj = (y1 , · · · , ynj ) denote the observed
toxicity data from nj patients treated at dose dj , and define the corresponding
sample mean
nj
X
µ̂j = yi /nj .
i=1

For a binary or quasi-binary toxicity endpoint (DLT or ETS), µ̂j is the ob-
served toxicity rate at dose level j; and for continuous endpoints such as the
TTB or TTP, µ̂j is the sample mean of the observed TTB or TTP at dose
level j.
Multiple Toxicity Grades 141

The gBOIN design shares a similarly concise dose escalation/de-escalation

rule as the standard BOIN design (with a binary endpoint). Let λe and λd
denote the optimal dose escalation/de-escalation boundaries. The gBOIN is
summarized as follows:
1. Patients in the first cohort are treated at the lowest dose d1 , or the
physician-specified dose.
2. Suppose j is the current dose level. To assign a dose to the next
cohort of patients:
• If µ̂j ≤ λe , escalate the dose to level j + 1.
• If µ̂j > λd , de-escalate the dose to level j − 1.
• Otherwise, (i.e., λe < µ̂j ≤ λd ), stay at the current dose level
j.
3. Repeat Step 2 until the maximum sample size N is reached. At
that point, select MTD as the dose whose isotonic estimate µ̃j is
closest to target φ, where {µ̃j , j = 1, . . . , J} is obtained by applying
the pooled adjacent violators algorithm (Barlow et al., 1972) to
{µ̂j , j = 1, . . . , J}.
Table 7.2 provides examples of the values of (λe , λd ) for different target values
of φ. The general formula to calculate (λe , λd ) is described in the next section.
For patient safety, the following overdose control rule is imposed when
using the gBOIN design:
If Pr (µj > φ | Dj ) > 0.95 and nj ≥ 3, dose level j and higher are elimi-
nated from the trial, and the trial is terminated with no dose being selected
as MTD if the lowest dose level is eliminated.

TABLE 7.2: Dose escalation and de-escalation boundaries (λe , λd ) for

Bernoulli, quasi-Bernoulli, and continuous toxicity endpoints.

Bernoulli or quasi-Bernoulli endpoint (e.g., DLT or ETS)

Target toxicity probability φ
0.10 0.15 0.20 0.25 0.30 0.35 0.40
λe 0.078 0.118 0.157 0.197 0.236 0.276 0.316
λd 0.119 0.179 0.238 0.298 0.358 0.419 0.479

Continuous endpoint (e.g., TTB, TBS or TTP)

Target toxicity value φ
0.25 0.5 0.75 1.0 1.5 2.0 3.0
λe 0.2 0.4 0.6 0.8 1.2 1.6 1.4
λd 0.3 0.6 0.9 1.2 1.8 2.4 3.6
*Under the default setting φ1 = 0.6φ and φ2 = 1.4φ.
142 Model-Assisted Bayesian Designs for Dose Finding and Optimization

The posterior probability Pr (µj > φ | Dj ) > 0.95 can be evaluated on the ba-
sis of a beta-binomial model for the binary or quasi-binary endpoint, assum-
ing µj follows a vague beta prior, e.g., µj ∼ Beta (1, 1), leading to a posterior
µj |yj , nj ∼ Beta (yj + 1, nj − yj + 1). For normal endpoint y with mean µj
and variance σj2 , assuming noninformative prior (µ, σj2 ) ∝ σj−2 , the marginal
posterior distribution of µj follows a student’s t distribution
Pnj with degrees of
freedom (nj − 1), mean µ̂j and scale parameter n−1 j
2
i=1 (yi − µ̂j ) .

7.2.2 Statistical derivation and properties

The derivation of the gBOIN design is similar to that of the BOIN design,
described in Section 3.4.2. To proceed, a class of nonparametric designs with
dose escalation and de-escalation boundaries (λe (dj , nj , φ), λd (dj , nj , φ)) are
introduced as follows:
1. Patients in the first cohort are treated at the lowest dose level or
the physician-specified level.
2. Suppose j is the current dose level. To assign a dose to the next
cohort of patients, consider the following:
• If µ̂j ≤ λe (dj , nj , φ), escalate the dose level to j + 1.
• If µ̂j > λd (dj , nj , φ), de-escalate the dose level to j − 1.
• Otherwise, i.e., λe (dj , nj , φ) < µ̂j ≤ λd (dj , nj , φ), stay at the
current dose level j.
3. Continue to repeat Step 2 until the maximum sample size N is
reached.
As the dose escalation and de-escalation boundaries λe (dj , nj , φ) and
λd (dj , nj , φ) are unspecified functions of dose dj and nj (i.e., the number
of patients who have been treated at dose level j), this family of designs in-
cludes all possible nonparametric designs that do not impose a parametric
model on the dose–toxicity curve and use the local data observed from the
current dose level in decision making.
The gBOIN design is obtained by choosing the values of λe (dj , nj , φ) and
λd (dj , nj , φ) that minimize the probability of making incorrect decisions on
dose escalation and de-escalation. Toward that goal, following the approach
of Liu and Yuan (2015), Mu et al. (2018) considered the following three point
hypotheses:

H0j : µj = φ, H1j : µj = φ1 , H2j : µj = φ2 ,

where φ1 indicates that the dose is substantially underdosing (i.e., below

MTD) such that escalation is required, and φ2 indicates that the dose is
substantially overdosing such that de-escalation is required. Let S, E, and D
Multiple Toxicity Grades 143

denote stay (at the current dose), escalation, and de-escalation, respectively.
Under H0j , the correct decision is S, and incorrect decisions are S̄ = {E, D};
under H1j , the correct decision is E, and incorrect decisions are Ē = {S, D};
and under H2j , the correct decision is D, and incorrect decisions are D̄ =
{S, E}.
Taking a non-informative approach, the investigational dose dj is assumed
a priori to have an equal chance of being below, equal to, or above the target,
i.e., Pr(H0j ) = Pr(H1j ) = Pr(H2j ) = 1/3. Then the probability of making an
incorrect decision, denoted by α, is given by



α = Pr(H0j ) Pr(S|H0j ) + Pr (H1j ) Pr E |H1j + Pr (H2j ) Pr D|H2j
1 1
= Pr{µ̂j ≤ λe (dj , nj , φ) or µ̂j > λd (dj , nj , φ)} + Pr{µ̂j > λe (dj , nj , φ)}
3 3
1
+ Pr{µ̂j ≤ λd (dj , nj , φ)}. (7.2)
3
Mu et al. (2018) derived the optimal values of λe (dj , nj , φ) and λd (dj , nj , φ)
that minimize the decision error (7.2).
Theorem 7.1 Let ϑk denote the model parameters under Hkj , k = 0, 1, 2.
The probability of making an incorrect decision is minimized by
A(ϑ1 ) − A(ϑ0 ) A(ϑ2 ) − A(ϑ0 )
λe = , λd = , (7.3)
η(ϑ1 ) − η(ϑ0 ) η(ϑ2 ) − η(ϑ0 )

given that the non-informative prior Pr(H0j ) = Pr(H1j ) = Pr(H2j ) = 1/3 is

used.
Specifically, when y is a binary or quasi-binary toxicity endpoint (DLT or
ETS), we have ϑk = φk , A(ϑk ) = − log(1 − φk ), η(ϑk ) = log{φk /(1 − φk )},
with φ0 ≡ φ. Then,
   
1 − φ1 1−φ
log log
1−φ 1 − φ2
λe =  , λd =  , (7.4)
φ(1 − φ1 ) φ2 (1 − φ)
log log
φ1 (1 − φ) φ(1 − φ2 )

which are the same as the boundaries of the BOIN design (Liu and Yuan,
2015) for a standard binary toxicity endpoint.
When y is a continuous endpoint (e.g., TTB, NETS, TBS, and TTP)
following a normal distribution, we have ϑk = φk , A(ϑk ) = φ2k /(2σj2 ), η(ϑk ) =
φk /σj2 . Then,

φ + φ1 φ + φ2
λe = , λd = . (7.5)
2 2
Therefore, gBOIN generalizes the BOIN design to embrace various types of
endpoints.
Table 7.2 shows examples of the values of (λe , λd ) for different target values
of φ with φ1 = 0.6φ and φ2 = 1.4φ. It is remarkable that, regardless the type
144 Model-Assisted Bayesian Designs for Dose Finding and Optimization

of endpoint, the optimal dose escalation and de-escalation boundaries (λe , λd )

are independent of dj and nj , which means that the same pair of boundaries
can be used throughout the trial no matter which dose is the current dose or
how many patients have been treated at the current dose. This feature makes
the gBOIN design simple to implement in practice.
Similar to the BOIN design, the gBOIN design has the following desirable
finite-sample and large-sample properties:

Theorem 7.2 The gBOIN design is long-term memory coherent in the sense
that the design will never escalate the dose when µ̂j > φ, and it will never
de-escalate the dose when µ̂j < φ.

Theorem 7.3 As the number of patients goes to infinity, the dose assignment
and the selection of MTD under the gBOIN design converge almost surely to
dose level j ∗ if dose level j ∗ is the only dose satisfying µj ∗ ∈ (λe , λd ). If there
are multiple dose levels in (λe , λd ), the design will converge almost surely to
one of these levels.

7.3 Multiple toxicity BOIN

The gBOIN design requires specifying a single toxicity endpoint (e.g., ETS,
TTB, or TBS) to summarize the toxicity profile of the dose. In some cases,
investigators prefer to retain the original scale of the toxicity grade (e.g., as
scored by CTCAE) and use multiple toxicity constraints to identify MTD. For
example, if the investigators are interested in finding MTD as the highest dose
with Pr(grade 2 toxicity) ≤ 40% and Pr(grade 3 and higher toxicity) ≤ 30%.
The multiple-toxicity BOIN (MT-BOIN) design (Lin, 2018) can be used to
achieve this goal.

7.3.1 Trial design

Consider a phase I trial with two binary toxicity endpoints of interest, Y1 and
Y2 , where, for example, Y1 = 1 denotes the occurrence of grade 2 toxicity, and
Y2 = 1 denotes the occurrence of grade 3 or higher toxicity. Let πlj = Pr(Yl =
1|dj ) denote the toxicity probability of Yl at dose dj , with πl1 < · · · < πlJ ,
l = 1, 2, dj ∈ {d1 , . . . , dJ }. Lin (2018) discussed MT-BOIN designs for two
different structures of Y1 and Y2 , i.e., Y1 and Y2 are nested or non-nested. An
example of Y2 nested in Y1 is that Y1 denotes grade 2 or higher toxicity, and
Y2 denotes grade 3 or higher toxicity. Here, we focus on the case that Y1 and
Y2 are not nested, e.g., Y1 denotes grade 2 toxicity, and Y2 denotes grade 3
or higher toxicity. In many cases, nested endpoints can be transformed into
non-nested endpoints.
Let φ(l) denote the target toxicity probability of Yl , l = 1, 2. In the example
above, φ(1) = 40% (for Y1 : grade 2 toxicity) and φ(2) = 30% (for Y2 : grade 3
Multiple Toxicity Grades 145

TABLE 7.3: The escalation/de-escalation boundaries (λe , λd ) under the BOIN

design for different target toxicity rates φ, using the default underdosing tox-
icity probability φ1 = 0.6φ and overdosing toxicity probability φ2 = 1.4φ.

Target toxicity probability φ

Boundaries 0.15 0.2 0.25 0.3 0.35 0.4
λe 0.118 0.157 0.197 0.236 0.276 0.316
λd 0.179 0.238 0.298 0.358 0.419 0.479

or higher toxicity). With the two toxicity constraints, MTD (i.e., target dose)
is defined as
dj † = min{dj1∗ , dj2∗ }, (7.6)
where
djl∗ = arg min |πlj − φ(l) |, l = 1, 2.
dj ∈{d1 ,...,dJ }

Here, djl∗ is the dose that has the toxicity probability closest to the target
toxicity probability with respect to Yl . Depending on trial objectives, other
definitions of MTD can also be used when appropriate.
Suppose that at dose dj , a total of ylj out of nj patients have experienced
the toxicity event associated with Yl . The observed toxicity rate of Yl at dj is

π̂lj = ylj /nj , l = 1, 2, j = 1, . . . , J.

(1) (1)
Let (λe , λd ) denote the optimal dose escalation/de-escalation boundaries
corresponding to Y1 , which are the same as those of the standard BOIN design
(2) (2)
with Y1 as the endpoint and φ(1) as the target. Similarly, let (λe , λd ) denote
the dose escalation/de-escalation boundaries corresponding to Y2 , which are
the same as those of the standard BOIN design with Y2 as the endpoint and
φ(2) as the target. Table 7.3 provides the standard BOIN dose escalation/de-
escalation boundaries for different targets, see Section 3.4 for more details.
(1) (1) (2) (2)
In the above trial example, (λe , λd ) = (0.316, 0.479) and (λe , λd ) =
(0.236, 0.358).
The MT-BOIN design has a similarly concise dose escalation/de-escalation
rule as the standard BOIN design (with a binary endpoint), by simply com-
(l) (l)
paring the observed toxicity rate π̂lj with λe and λd , l = 1, 2. The MT-BOIN
design is illustrated in Figure 7.1 and described as follows:
1. Patients in the first cohort are treated at the lowest dose d1 , or the
physician-specified dose.
2. Suppose j is the current dose level. To assign a dose to the next
cohort of patients:
(1) (2)
• If π̂1j ≤ λe and π̂2j ≤ λe , escalate the dose to level j + 1.
146 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 7.1: Dose escalation/retainment/de-escalation regions of the MT-

(l) (l)
BOIN design with two toxicity endpoints Y1 and Y2 , where λe and λd are
standard BOIN dose escalation and de-escalation boundaries corresponding
to Yl , l = 1, 2.

(1) (2)
• If π̂1j > λd or π̂2j > λd , de-escalate the dose to level j − 1.
(1) (2) (2) (1)
• Otherwise, i.e., π̂1j ≤ λd , λe < π̂2j ≤ λd or λe < π̂1j ≤
(1) (2)
λd , π̂2j ≤ λd , stay at the current dose level j.
3. Repeat Step 2 until the maximum sample size N is reached. At that
point, perform the isotonic regression (Barlow et al., 1972) to the
estimated toxicity rates {π̂lj , j = 1, . . . , J}, so that the isotonically
transformed estimator {π̃lj , j = 1, . . . , J} satisfies the monotonic
constraint, and then select the overall MTD ĵ † as

dĵ † = min{dĵ ∗ , dĵ ∗ },

1 2

where dĵ ∗ = arg min|π̃lj − φ(l) |, and the set N = {dj : nj > 0}
l
dj ∈N
contains all the doses that have been tested in the trial.
For the purpose of overdose control, the MT-BOIN design imposes a dose
elimination rule similar to BOIN as follows:
Multiple Toxicity Grades 147

If any of Pr(πlj > φ(l) | Dj ) > 0.95, l = 1, 2, and nj ≥ 3, dose level j

and higher are eliminated from the trial, and the trial is terminated if the
lowest dose level is eliminated.

The value of Pr(πlj > φ | Dj ) is evaluated based on the beta-binomial model

(see Section 3.4). When the trial is terminated due to toxicity, no dose should
be selected as MTD.

7.3.2 Statistical derivation and properties

(l)
Let φ1 < φ(l) denote the highest probability considered to be subtherapeutic
(l)
such that dose escalation is needed, and let φ2 > φ(l) be the lowest value
deemed excessively toxic such that dose de-escalation is warranted, l = 1, 2.
Following Liu and Yuan (2015), three hypotheses are considered for the MT-
BOIN design with non-nested outcomes,

H0j : (π1j , π2j ) ∈ Θ0j , H1j : (π1j , π2j ) ∈ Θ1j , H2j : (π1j , π2j ) ∈ Θ2j ,
(7.7)
where the parameter set Θij , i = 0, 1, 2 is defined as
n o
(2) (1)
Θ0j = (φ(1) , φ(2) ), (φ(1) , φ1 ), (φ1 , φ(2) ) ,
n o
(1) (2)
Θ1j = (φ1 , φ1 ) ,
n o
(1) (2) (1) (1) (2) (2) (1) (2)
Θ2j = (φ2 , φ2 ), (φ2 , φ(2) ), (φ2 , φ1 ), (φ(1) , φ2 ), (φ1 , φ2 ) .

Unlike the standard BOIN design that only involves three point hypotheses,
each hypothesis Hij , i = 0, 1, 2 of the proposed MT-BOIN enjoys a hierar-
chical structure and includes a set of multiple point sub-hypotheses, which
is thus composite. In particular, Θ1j indicates that both toxicity rates for
Y1 and Y2 at dose level j are overly small, hence, the hypothesis H1j cor-
responds to the subtherapeutic case. Θ2j is made up by the combinations
that at least one toxicity rate is excessively high, then H2j means that dose
level j is not safe. Similarly, H0j indicates that the dose is within the target
zone. Note that under H0j , the dose level with the joint toxicity probabilities
(2) (1)
(π1j , π2j ) = (φ(1) , φ1 ) (or (π1j , π2j ) = (φ1 , φ(2) )) is treated as proper dosing
instead of underdosing. This is because the monotone increasing dose–toxicity
relationship implies that φ(1) < π1,j+1 (or φ(2) < π2,j+1 ); as a result, dose level
j + 1 might be overly toxic in terms of toxicity Y1 (or Y2 ).
Let Oj = (y1j , y2j , nj ) be the local data observed at dose level j. De-
note the prior model probability Pr(Hij ) by ωij . For simplicity, the uniform
prior ωij = 1/3 is considered, i = 0, 1, 2. The correct dose-assignment de-
cisions under H0j , H1j , and H2j are S, E, and D, respectively. To account
for the composite nature of the three hypotheses, Lin (2018) defined αmax ,
the maximum probability of making incorrect decisions with multiple toxicity
148 Model-Assisted Bayesian Designs for Dose Finding and Optimization

outcomes, as follows:
αmax = ω1j Pr(S or D | H1j ) + ω0j Pr(D or E | H0j ) + ω2j Pr(E or S | H2j )
nj nj 
X X
= ω1j I(π̂1j > λ(1)
e or π̂2j > λ(2)
e ) max f (Oj ; π1j , π2j )
(π1j ,π2j )∈Θ1j
y1j =0 y2j =0

+ω0j I(π̂1j ≤ λ(1)

e and π̂2j ≤ λ(2)
e ) max f (Oj ; π1j , π2j )
(π1j ,π2j )∈Θ0j
(1) (2)
+ω0j I(π̂1j > λd or π̂2j > λd ) max f (Oj ; π1j , π2j )
(π1j ,π2j )∈Θ0j

(1) (2)
+ω2j I(π̂1j ≤ λd and π̂2j ≤ λd ) max f (Oj ; π1j , π2j ) ,
(π1j ,π2j )∈Θ2j

where f (Oj ; π1j , π2j ) is the joint binomial likelihood function for Y1 and Y2 .
Due to the small sample size of the phase I dose-finding trials, the incorpo-
ration of the correlation between the two toxicity outcomes does not neces-
sarily improve the performance of the design. Therefore, the joint likelihood
f (Oj ; π1j , π2j ) can be simply treated as the product of independent binomial
likelihood functions. Note that even when the two outcomes are correlated,
the marginal estimate of the toxicity rate under the working independence
assumption is still consistent.
Based on the minimax theory, Lin (2018) obtained explicit expressions
(l) (l)
of optimal interval boundaries λe (dj , nj , φ(l) ) and λd (dj , nj , φ(l) ) for non-
nested toxicities by minimizing αmax .
Theorem 7.4 If uniform prior model probabilities are considered, the optimal
interval boundaries obtained by minimizing the maximum incorrect probability
(7.8) for the toxicity Yl are exactly the same as those of standard BOIN by
treating Yl alone, that is, for l = 1, 2,
! !
(l)
1 − φ1 1 − φ(l)
log log
1 − φ(l) (l) 1 − φ2
(l)
(l)
λe = (l)
! , λd =
(l)
!. (7.8)
φ(l) (1 − φ1 ) φ2 (1 − φ(l) )
log (l)
log (l)
φ1 (1 − φ(l) ) φ(l) (1 − φ2 )

The proof of Theorem 7.4 is provided in Lin (2018). Remarkably, The-

orem 7.4 shows that the optimal boundaries of MT-BOIN coincide exactly
with those of standard BOIN based on a single toxicity outcome and are in-
dependent of dj and nj . Because the optimal interval boundaries from (7.8)
minimize the maximum probability of incorrect decisions, MT-BOIN enjoys
optimality within the aforementioned class of local-data designs for multiple
toxicities.
Multiple Toxicity Grades 149

7.4 Software and illustration

Software for the gBOIN and MT-BOIN designs are available as a module of
“BOIN Suite” at http://www.trialdesign.org. In what follows, we use the
solid tumor trial introduced in Section 3.6 to illustrate the use of gBOIN to
design the trial. Recall that the objective of the trial is to determine MTD
for the MDM2/MDMX inhibitor ALRN-6924 in combination with paclitaxel
in adult patients with advanced or metastatic solid tumors. Five doses of
ALRN-6924 are investigated with the target DLT probability φ = 25%. For the
purpose of illustration, we here consider an alternative target DLT probability
φ = 30%. For MT-BOIN, design parameters and setup are similar to these of
the standard BOIN design, see Section 3.6 for details.
Suppose that investigators anticipate that the agent may induce a high
percentage of low grade toxicities that may cause substantial dose reduction
and discontinuation. Thus, the investigators prefer to account for low grade
toxicities in the decisions of dose escalation/de-escalation and MTD determi-
nation. After selecting and launching the “gBOIN/MT-BOIN” module from
the BOIN Suite launchpad, the trial can be designed using the following three
steps:
Step 1: Enter trial parameters
Doses and Sample Size As shown in Figure 7.2, the number of doses under
investigation is five; the starting dose level is one. The cohort size for the trial
is three and the number of cohorts is 10, with a total sample size of 30. As a
rule of thumb, we recommend the maximum sample size N = 6 × J (i.e., the

FIGURE 7.2: Specify doses, sample size, and convergence stopping rule.
150 Model-Assisted Bayesian Designs for Dose Finding and Optimization

maximum sample size of the 3+3 design) as the total sample size, where J is
the number of doses. This sample size generally yields reasonable operating
characteristics (e.g., 50-70% correct selection percentage of the true MTD).
Section 3.6 provides more guidance on the determination of cohort size.
To reduce the sample size, it is often useful to use the “convergence”
stopping rule: stop the trial early when m patients have been assigned to a
dose and the decision is to stay at that dose. This stopping criterion suggests
that the dose finding approximately converges to MTD, thus the trial can be
stopped. We recommend m = 9 or larger. In this trial, m = 12 is used. Because
of the early stopping rule, the actual sample size used in the trial is often
smaller than the prespecified maximum sample size N . The saving depends
on the true dose–toxicity scenario and can be evaluated using simulation.
Usually, the saving in the sample size is more prominent when the true MTD
is near the starting dose.
Target and Equivalent DLT As shown in Figure 7.3, the target DLT proba-
bility is φ = 0.3. When appropriate, the accelerated titration can be chosen to
speed up dose escalation and reduce the total sample size. See Section 3.6 for
more discussion on choosing the target and pros and cons of the accelerated
titration.
For gBOIN, the new design parameters are “Equivalent Number of DLT,”
which maps toxicity grades to numeric scores that reflect their relative severity
in the unit of DLT. This is the “Toxicity score” approach discussed in Section
7.1. For example, in Figure 7.3, grade 3 or higher toxicity is defined as a DLT;
grade 2 toxicity is regarded as equivalent to 0.5 DLT, while grade 1 toxicity is
regarded as acceptable and should not affect decisions of dose transition and
MTD determination. The toxicity score should be elicited from clinicians and
reflects the relative severity of different toxicity grades.

FIGURE 7.3: Specify the target and equivalent DLT scores.

Multiple Toxicity Grades 151

FIGURE 7.4: Specify the overdose control rules.

Overdose Control This panel (see Figure 7.4) specifies the overdose con-
trol rule, and it is the same as that of standard BOIN. Section 3.6 provides
guidance on how to set up the parameters. After completing the specification
of trial parameters, the decision table will be generated by clicking the “Get
Decision Table” button. The decision table will be automatically included in
the protocol template in Step 3, but can also be saved as a separate csv, Excel
and pdf file in this step if needed.

Step 2: Run simulation

Operating Characteristics This step generates the operating characteristics
of the design through simulation. As gBOIN considers multiple toxicity grades,
the construction of scenarios is more complicated. Users need to provide the
probability of each toxicity grade for each dose, see Figure 7.5. The scenarios
used for simulation should cover various possible clinical scenarios, e.g., MTD
is located at different dose levels. To facilitate the generation of the operat-
ing characteristics of the design, the software automatically provides a set of
randomly generated scenarios with various MTD locations, which are often
adequate for most trials. Depending on the application, users can modify, add
or remove scenarios.
Figure 7.6 shows the operating characteristics of gBOIN for the solid tu-
mor trial. The simulation results will be automatically included as a table in
the protocol template in the next step, but can also be saved as a separate
csv or Excel file if needed.

Step 3: Generate protocol template

Protocol Preparation The gBOIN software generates sample texts and a
protocol template (including the simulation results in Step 2) to facilitate
152 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 7.5: Simulate operating characteristics of the design.

the protocol write-up. The protocol template can be downloaded in various

formats (e.g., html or Word file). Use of this module requires the comple-
tion of Steps 1 and 2. Once the protocol is approved by regulatory bodies
(e.g., Institutional Review Board), we follow the design decision table to
conduct the trial and make adaptive decisions (e.g., dose escalation/stay/
de-escalation).
When the trial is completed, use the “Select MTD” module to select MTD.
The software outputs the recommended MTD, the posterior estimate of DLT
Multiple Toxicity Grades 153

FIGURE 7.6: Operating characteristics of the gBOIN design for the solid
tumor trial.

probability at each dose, and a 95% credible interval (see Figure 7.7). Given
the data entered, dose level three is selected as MTD. In contrast, if we
considered only DLT, ignoring grade 2 toxicities, then dose level four or five
would be selected as MTD.

FIGURE 7.7: MTD estimation at the completion of the solid tumor trial based
on a hypothetical dataset.
8
Finding Optimal Biological Dose

8.1 Introduction
Conventionally, the primary objective of phase I oncology trials is to establish
the maximum tolerated dose (MTD). This more-is-better paradigm is based
on the monotonicity assumption that a higher dose leads to higher toxicity and
also higher efficacy, which typically holds for conventional chemotherapies.
The advent of novel targeted therapy and immunotherapy, such as check-
point inhibitors and chimeric antigen receptor (CAR) T-cell therapy, has rev-
olutionized cancer treatment. For these novel therapies, although toxicity typ-
ically increases with the dose, efficacy may plateau or even decrease at high
doses (Cook et al., 2015; Sachs et al., 2016; Shah et al., 2021). In addition,
some targeted therapies demonstrate minimal toxicity in the therapeutic dose
range, making MTD unlikely to be reached (Mathijssen et al., 2014). In these
cases, the conventional more-is-better paradigm, ignoring efficacy data, does
not depict the underlying setting and may result in undesirable consequences.
(Shah et al., 2021). As a result, FDA Oncology Center of Excellence recently
initiated Project Optimus “to reform the dose optimization and dose selection
paradigm in oncology drug development” (FDA, 2022).
To illustrate the issue, consider five doses of a targeted or immunother-
apy agent, d1 < · · · < d5 . Suppose that the true toxicity probabilities for the
five doses are (πT 1 , . . . , πT 5 ) = (0.05, 0.12, 0.27, 0.35, 0.50). If the true efficacy
probabilities are (πE1 , . . . , πE5 ) = (0.20, 0.35, 0.36, 0.37, 0.38), then efficacy
reaches a plateau of about 0.35 at dose d2 . All dose-finding methods consider-
ing toxicity only and with a target toxicity probability of 0.3 or 0.25, are most
likely to select d3 as MTD and use it for cohort expansion or a subsequent
phase II trial. However, d2 is obviously more desirable than MTD (i.e., d3 )
with much lower toxicity and virtually identical efficacy. Any “toxicity-only”
phase I method cannot determine this because it ignores efficacy.
Furthermore, suppose that the toxicity rates are the same but the true effi-
cacy probabilities are (πE1 , . . . , πE5 ) = (0.01, 0.05, 0.30, 0.60, 0.60). Escalating
from d3 to d4 only increases the toxicity probability from 0.27 to 0.35, but
doubles the efficacy probability from 0.3 to 0.6. Often this small increase in
toxicity may be considered as a reasonable trade-off for the large increase in
efficacy by choosing d4 rather than d3 , however, toxicity-only methods cannot
determine this.

DOI: 10.1201/9780429052781-8 155

156 Model-Assisted Bayesian Designs for Dose Finding and Optimization

Lastly, if the agent is ineffective for all doses, with true efficacy probabilities
(πE1 , . . . , πE5 ) = (0.00, 0.01, 0.01, 0.02, 0.02), the best decision is to not choose
any dose, but the toxicity-based methods still are most likely to choose d3 .
These examples demonstrate the deficiency of choosing a “best” dose based
only on toxicity (e.g., MTD), ignoring efficacy, which may lead to the failure
of subsequent phase II or III trials.

8.1.1 Phase I–II design paradigm

The aforementioned issues can be avoided by adopting the phase I–II design
paradigm, which considers both toxicity and efficacy to adaptively allocate
patients and identify the optimal biological dose (OBD). In general, the OBD
is defined as the dose that optimizes a certain metric of the risk-benefit trade-
off. Figure 8.1 shows the schema of phase I–II designs, which consist of the
following elements:
• Toxicity and efficacy outcomes that characterize potential risks and
benefits of the treatment being investigated. Examples of toxicity outcome
include dose limiting toxicity (DLT) or tolerability rates, and examples of ef-
ficacy outcome include tumor responses, pharmacodynamic (PD) biomarkers
measuring biological activities of the drug (e.g., inhibition of an oncogene
pathway or proliferation of T cells), or surrogate efficacy endpoints (e.g.,
ctDNA, minimal residual disease). In this chapter, when describing designs,
we use the toxicity rate and efficacy rate to generically represent the risk and
benefit of the treatment, respectively, which should be customized based on
the trial under consideration.
• Risk–benefit trade-off criterion that characterizes and quantifies the
trade-off between efficacy and toxicity for each dose of the drug. Specific
example includes the trade-off based on marginal toxicity and efficacy prob-
abilities (Section 8.2.1) or utility based on the values of toxicity and efficacy
endpoints (Section 8.3.1).
• Optimal biological dose (OBD) that maximizes the predefined risk–
benefit trade-off criterion among all considered doses.
• Statistical model that describes the relationship between dose, toxicity,
and efficacy. Examples include the Gumbel-Morgenstern copula model (Sec-
tion 8.2.2) and multinomial model for bivariate binary toxicity and efficacy
endpoints (Section 8.3.2).
• Adaptive decision rule that determines the best dose for the next cohort,
based on the (dose, toxicity, efficacy) data observed from previous patients.
Typically, the next cohort is assigned to the dose that has the highest esti-
mate of the risk-benefit trade-off, based on the interim data.
• Admissibility rules that protect patients in the trial from unacceptably
toxic or inefficacious doses. Admissibility rules are used to define admissible
doses that satisfy certain prespecified toxicity and efficacy requirements.
Finding Optimal Biological Dose 157

Upon identification, only admissible doses can be used to treat patients.

Examples will be provided in the next section.
• Stopping rule that terminates the trial early if all the doses being consid-
ered are unacceptably toxic or inefficacious.
• Selection rule that identifies OBD for subsequent studies if the trial is not
terminated early. Typically, OBD is selected as the dose that is admissible
and has the most favorable estimate of the risk-benefit trade-off.
A number of phase I–II designs have been proposed in recent years. The
majority of them are model-based. These model-based designs consider differ-
ent settings (e.g., different types of endpoints, statistical models, risk-benefit
trade-off criteria, and decision rules), but share a similar design strategy to
that outlined in Figure 8.1. Thall and Cook (2004) introduced a toxicity–
efficacy trade-off contour for dose finding based on the Gumbel-Morgenstern
copula model (hereafter referred to as the EffTox design). Bekele and Shen
(2005) considered jointly modeling a binary toxicity outcome and a continuous
efficacy outcome in phase I–II dose-finding oncology trials. Hunsberger et al.
(2005) proposed a slope-sign design to guide dose escalation. Yin et al. (2006)
parameterized both the toxicity and efficacy probabilities based on multivari-
ate normal random variables, and they applied the global cross-ratio model
(Dale, 1986) to jointly quantify the toxicity and efficacy outcomes. Zhang et al.
(2006) developed a flexible continuation-ratio model to select the optimal dose
at each decision-making time. Zang et al. (2014) used the logistic model to
depict the local dose–efficacy relationship around the current dose. Liu and
Johnson (2016) proposed a nonparametric design using a flexible Bayesian dy-
namic model for monotone dose–response curves. Riviere et al. (2018) utilized
adaptive randomization to determine the plateau in phase I–II dose-finding
designs. More sophisticated phase I–II designs were developed to handle more
complex early-phase trials; for example, drug-combination trials (Yuan and

Treat the first

cohort of patients
at the prespecified
starting dose

Collect both
efficacy and
toxicity data

Select the optimal Fit dose-efficacy

dose for the next and dose-toxicity
cohort of patients models

NO

Stop the trial and YES Is maximum Are all doses YES Terminate the
select the optimal sample size overly toxic or/
trial early
dose reached? and futile?

Update dose NO
acceptability and
dose desirability by
including the most
recent data

FIGURE 8.1: Diagram of phase I–II trial design.

158 Model-Assisted Bayesian Designs for Dose Finding and Optimization

Yin, 2011a; Cai et al., 2014; Guo and Li, 2015), dose–schedule optimization
(Guo et al., 2016; Lin et al., 2020a, 2021), ordinal toxicity and efficacy out-
comes (Houede et al., 2010; Lee et al., 2016), personalized dose-finding based
on biomarkers (Guo and Yuan, 2017), and immunotherapy considering tox-
icity, efficacy and immune response (Liu et al., 2018), among others. For a
comprehensive review on model-based phase I–II designs, see Mandrekar et al.
(2010) and Yuan et al. (2016b).
These model-based phase I–II designs are statistically complicated and
computationally intensive. To account for toxicity and efficacy, the model
used by the designs is significantly more complex than that of the conven-
tional phase I trial design (e.g., CRM), as illustrated in the next section.
Highly complicated and structured parametric models also make the design
susceptible to model misspecification. As a result, the model-based phase I-
II designs are often regarded by practitioners as difficult to understand and
implement, severely limiting their use in practice.
Model-assisted designs have been proposed to simplify the implementation
of phase I–II trials, while yielding the performance comparable to or even bet-
ter than model-based designs. As model-assisted designs do not assume any
parametric dose-toxicity and -efficacy curves, they are also more robust than
model-based designs. Lin and Yin (2017b) developed a toxicity-efficacy inter-
val based phase I–II design (called STEIN) using the pool-adjacent-violators
algorithm and model averaging. Li et al. (2017) proposed a toxicity and effi-
cacy probability interval design that separately models toxicity and efficacy.
Takeda et al. (2018) presented a Bayesian optimal interval design that ac-
commodates both efficacy and toxicity endpoints, known as BOIN-ET. Zhou
et al. (2019) developed a two-stage utility-based Bayesian optimal interval (U-
BOIN) design using a Dirichlet-multinomial model to jointly model toxicity
and efficacy. Lin et al. (2020b) adopted a quasi-binomial likelihood approach
to model the observed utility directly, and developed the BOIN12 design. Shi
et al. (2021) developed a utility-based toxicity probability interval (uTPI)
design as an extension of the toxicity-only keyboard design (Yan et al., 2017).
In the following sections, we first introduce the EffTox design as an example
to illustrate the characteristics of model-based designs, and then we describe
several model-assisted phase I–II designs, including the BOIN12, U-BOIN, and
uTPI designs. Other model-assisted designs such as STEIN and BOIN-ET will
be briefly discussed at the end of this chapter.

8.2 EffTox design

The EffTox design is a model-based phase I–II design. It considers a binary
toxicity endpoint yT and a binary efficacy endpoint yE , with yT = 1 and yE =
1 denoting the occurrence of toxicity and efficacy events, respectively. Let d1 <
· · · < dJ denote J prespecified doses. We assume that the (marginal) toxicity
Finding Optimal Biological Dose 159

probability increases with the dose, i.e., πT 1 < · · · < πT J . The (marginal)
efficacy probability πE1 , · · · , πEJ , however, does not necessarily increase with
the dose, which may plateau or even decrease at higher doses. The EffTox
design uses an efficacy–toxicity trade-off contour as the criterion to define and
select OBD (Thall and Cook, 2004).

8.2.1 Efficacy–toxicity trade-off

To construct trade-off contours, we first elicit from clinicians three equally
(1) (2) (3) (3)
desirable efficacy–toxicity probability pairs (πE , 0), (1, πT ), and (πE , πT ),
(1) (3) (3) (2) (1)
subject to the constraints πE < πE and πT < πT . For example, (πE , 0) =
(2) (3) (3)
(0.25, 0), (1, πT ) = (1, 0.60), and (πE , πT ) = (0.40, 0.15) means that a dose
with an efficacy rate of 25% and no toxicity, a dose with 100% efficacy and a
toxicity rate of 60%, and a dose with an efficacy rate of 40% and a toxicity
rate of 15% are considered equally desirable.
Based on the three equally desirable efficacy–toxicity probability pairs, an
efficacy–toxicity trade-off (or utility) function is defined as a function of the
marginal toxicity and efficacy probabilities
( !r !r )1/r
πE − 1 πT − 0
ψ(πE , πT ) = 1 − (1)
+ (2)
, (8.1)
πE − 1 πT − 0

where r > 0 controls the degree of the curvature of the trade-off contours
(Thall and Cook, 2004). The value of r is determined by solving the equation
(3) (3) (1) (2)
ψ(πE , πT ) = ψ(πE , 0) = ψ(1, πT ).

Once the value of r is determined, the family of efficacy–toxicity trade-off

contours is obtained by the utility function (8.1). Specifically, for a given
utility value δ, the efficacy–toxicity trade-off contour is defined as

Cδ = {(πE , πT ) : ψ(πE , πT ) = δ} .

That is, all (πE , πT ) pairs in Cδ have the same utility value δ.
Figure 8.2 gives an example for the efficacy–toxicity trade-off contours
based on the function (8.1). The utility is standardized between 0 and 1, with
the right bottom corner representing the most desirable case with a utility
of 1, where efficacy is certain and toxicity is impossible, and the left upper
corner representing the least desirable case with a utility of 0, where efficacy
is impossible and toxicity is certain. All (πE , πT ) pairs on each contour are
equally desirable. The utilities of the contours increase moving from upper
left to lower right, as πT becomes smaller and πE becomes larger.
160 Model-Assisted Bayesian Designs for Dose Finding and Optimization

Toxicity-efficacy Trade-off Contour

1.0
0.15
0.3

0.8
0.45
Probability of toxicity

0.6
0.6

0.4 d1 0.75
d3

0.2
0.9
d2

0.0

0.0 0.2 0.4 0.6 0.8 1.0

Probability of efficacy

FIGURE 8.2: Toxicity–efficacy trade-off based on the utility function (8.1),

where the lines represent trade-off contours, and the numbers on the contours
indicate the desirability of the contours. The dose d2 is more desirable than
d1 and equally desirable as d3 .

8.2.2 Joint probability model for efficacy and toxicity

The EffTox design uses the standardized dose xj in its efficacy-toxicity model,
defined as J
j  =1 log(dj )

xj = log(dj ) − .
J
For example, a trial of 4 doses, (d1 , d2 , d3 , d4 ) = (200, 300, 400, 600) mg, would
have (x1 , x2 , x3 , x4 ) = (−0.55, −0.14, 0.14, 0.55).
The marginal probabilities of efficacy and toxicity at dose dj are modeled
using a logistic model,

πqj = logit−1 {ηqj }, q = E, T and j = 1, . . . , J,

where

ηEj = µE + βE1 xj + βE2 x2j

ηT j = µT + β T x j ,
Finding Optimal Biological Dose 161

with βT > 0 to ensure that the marginal toxicity probability πT j increases with
the dose. There is no restriction on βE1 and βE2 , so that the dose–efficacy
curve can take various shapes. To induce correlation between efficacy and
toxicity, a Gumbel-Morgenstern copula is used to model the joint distribution
of yE and yT ,
yE
f (yT , yE , dj ; θ) = πEj {1 − πEj }1−yE πTyTj {1 − πT j }1−yT +
eψ − 1
 
yT +yE
+(−1) πEj {1 − πEj }πT j {1 − πT j } ,
eψ + 1
where θ = (µE , βE1 , βE2 , µT , βT , ψ) is the vector of unknown parameters,
and ψ is a real-valued association parameter. Denoting the data of the first
n patients in the trial by Dn = {yT i , yEi , d[i] }ni=1 with d[i] ∈ {d1 , . . . , dJ }
indicating the dose that patient i has received, the joint likelihood is given by
n
Y
L(Dn | θ) = f (yT i , yEi , d[i] ; θ).
i=1

Priors on the model parameters are assumed to be normally distributed, with

hyperparameter means determined from elicited means of πqj for each (q, j)
combination and hyperparameter variances calibrated to obtain a given spec-
ified prior effective sample size, q = E, T and j = 1, . . . , J. Additional details
are given in Thall and Cook (2004).

8.2.3 Admissible rules

Let φT be an upper limit of πT j and φE be a lower limit of πEj for a dose to
be acceptable, respectively. Values of φE and φT should be determined by the
clinical investigators based on the trial under consideration, and often vary
from one trial to another. For example, φT = 0.3 and φE = 0.2 means that
a dose is regarded as acceptable only when its toxicity probability ≤ 0.3 and
efficacy probability ≥ 0.2.
Formally, a dose dj is defined as acceptable (or admissible) when it satisfies
the following two criteria:

(Safety criterion) Pr(πT j > φT | Dn ) ≤ cT ,

(Efficacy criterion) Pr(πEj < φE | Dn ) ≤ cE ,
where cE and cT are a large probability threshold such as 0.95. Let A(Dn )
denote the set of admissible doses that satisfy the above safety and efficacy
requirements. During the trial, A(Dn ) should be continually updated based
on Dn to guide dose escalation and de-escalation. Only doses in A(Dn ) can
be used to treat patients.

8.2.4 Dose-finding algorithm

The dose-finding algorithm of the EffTox design, including the adaptive dose
assignment, stopping, and selection rules, is given as follows:
162 Model-Assisted Bayesian Designs for Dose Finding and Optimization

(1) The first cohort of patients are treated at the lowest dose, or the
starting dose specified by the clinicians.
(2) For each subsequent cohort after the first cohort, we fit the dose-toxicity
and dose-efficacy models based on the observed data Dn , and obtain the ad-
missible dose set Ã(Dn ) = A(Dn ) ∪ {dj }, where dj is the lowest untried dose
that has acceptable toxicity. If there is no untried dose or the untried doses
are all overly toxic, then Ã(Dn ) = A(Dn ).
(3) If Ã(Dn ) is empty, the trial should be terminated and no dose is
selected; otherwise, the next cohort is treated at the most desirable dose
dj ∈ Ã(Dn ) that has the largest estimate of the efficacy-toxicity trade-off
ψ(πEj , πT j ), subject to the constraint that no untried dose may be skipped
when escalating. Repeat Steps (2) and (3) if the trial is not terminated.
(4) Stop the trial when the maximum sample size N is reached. If A(DN )
is not empty, select the OBD as the dose dj ∗ ∈ A(DN ) that maximizes the
efficacy-toxicity trade-off ψ(πEj ∗ , πT j ∗ ).

Compared to toxicity-only dose-finding designs such as CRM, by con-

sidering the efficacy-toxicity trade-off, the EffTox design better reflects the
risk-benefit trade-off underlying medical decisions in practice and identifies
the OBD rather than MTD. In addition, the EffTox design uses more data
(both toxicity and efficacy data) to make adaptive decisions, and thus is more
efficient. The challenges are that the EffTox design is complicated to imple-
ment due to complex model (including prior) specification and estimation. In
addition, the EffTox design appears sensitive to model misspecification (see
Section 8.3.5 and 8.4.4). Although we only cover the EffTox design here, the
other model-based phase I–II designs have similar structures as the EffTox
design, and share similar challenges.
In the following sections, we introduce several model-assisted phase I-II
designs that are easy to implement and that yield performance comparable
or better than model-based designs such as EffTox. To use the model-assisted
designs, no complicated model fitting is needed, and clinicians can simply
look up the decision table to determine the dose for treating the next cohort
of patients. In addition, these designs do not assume any parametric model on
the dose-toxicity and dose-efficacy curves, and thus are not sensitive to model
misspecification.

8.3 U-BOIN design

U-BOIN extends the BOIN design to identify OBD for phase I-II trials (Zhou
et al., 2019). It considers a binary toxicity endpoint yT and a binary efficacy
endpoint yE , where yE can be tumor response or PD biomarkers measuring
Finding Optimal Biological Dose 163

biological activities of the drug (e.g., inhibition of an oncogene pathway, gene

expression, or proliferation of T cells).

8.3.1 Utility-based risk-benefit trade-off

U-BOIN uses utility to quantify the risk-benefit trade-off, based on possible
outcomes that a patient can have. Specifically, given binary yT and yE , there
are four possible outcomes for any patient in the trial: (yT , yE ) = (0, 1) =
(no toxicity, efficacy); (0, 0) = (no toxicity, no efficacy); (1, 1) = (toxicity,
efficacy); and (1, 0) = (toxicity, no efficacy) (Table 8.1). The joint outcome
(no toxicity, efficacy) is the most desirable, whereas (toxicity, no efficacy) is the
least desirable, and the other two are in between. We quantify the desirability
of each outcome by assigning it a utility (score), which should be elicited
from physicians to reflect the risk-benefit trade-off underlying their medical
decisions.
One way to elicit the utility is given as follows: fix the score of the most
desirable outcome (no toxicity, efficacy) as υ01 = 100 and the score of the least
desirable outcome (toxicity, no efficacy) as υ10 = 0, and then ask clinicians
to use them as a reference to specify scores υ00 and υ11 for the other two
outcomes, i.e., (no toxicity, no efficacy) and (toxicity, efficacy), respectively to
quantify the risk-benefit trade-off under each outcome. υ00 and υ11 should be
non-negative and ∈ [0, 100]. Table 8.1 lists an example of an elicited utility
table. The specification of υ00 and υ11 needs to reflect the clinical desirability
of the corresponding outcomes: setting υ11 ≥ υ00 means that having a re-
sponse is more important than not having a toxicity, and patients are willing
to tolerate toxicity in exchange for efficacy; otherwise, υ11 < υ00 should be
used. From the dose-finding perspective, setting υ11 ≥ υ00 is usually more
meaningful and leads to a more thorough exploration of the dose space than
setting υ11 < υ00 .
Let (p01 , p00 , p11 , p10 ) denote the probabilities of observing the four pos-
sible toxicity–efficacy outcomes at a certain dose d. The probability vector
(p01 , p00 , p11 , p10 ) typically varies across doses, i.e., as a function of d. For ease
of exposition, we suppress the argument d. Averaging over the four possible
outcomes, the desirability (or mean utility) of dose d is

u = p01 υ01 + p00 υ00 + p11 υ11 + p10 υ10 . (8.2)

TABLE 8.1: Utility table for binary toxicity and efficacy endpoints yT and
yE .

Efficacy (yE )
Toxicity (yT ) Yes (= 1) No (= 0)
No (= 0) υ01 = 100 υ00 = 40
Yes (= 1) υ11 = 60 υ10 = 0
164 Model-Assisted Bayesian Designs for Dose Finding and Optimization

A higher value of u indicates a higher dose desirability in terms of the risk-

benefit trade-off.
Revisit the example in Section 8.1 and assume the utility in Table
8.1, when the true toxicity and efficacy probabilities are (πT 1 , . . . , πT 5 ) =
(0.05, 0.12, 0.27, 0.35, 0.50) and (πE1 , . . . , πE5 ) = (0.20, 0.35, 0.36, 0.37, 0.38),
respectively, the desirability of d3 is u3 = 50.8, and the desirability of d2
is u2 = 56.2. By using the utility approach, we naturally account for the
risk–benefit trade-off and correctly identify that d2 is more desirable than d3 .
Similarly, if (πE1 , . . . , πE5 ) = (0.01, 0.05, 0.30, 0.60, 0.60), the desirability of d4
is u4 = 62.0, which is higher than u3 = 47.2 of dose d3 .
In our experience, clinicians can quickly understand what the utility means
and provide utility scores, since the framework resembles clinical practice and
the utility scores reflect actual clinical judgment. After completing this pro-
cess, simulation should be performed to verify the operating characteristics of
the design. In some cases, the simulation results may motivate slight modifi-
cation of some of the numerical utility values.
One possible criticism for using the utility values is that they require sub-
jective input. However, we are inclined to view this as a strength rather than
a weakness. The process of specifying the utility requires clinicians to care-
fully consider the potential risks and benefits of the treatment that underlie
their clinical decision making in a more formal way and incorporate that into
the trial. In addition, our simulation study and previous studies (Guo and
Yuan, 2017; Liu et al., 2018; Murray et al., 2018; Zhou et al., 2019; Lin et al.,
2020b) show that the design is generally not sensitive to small differences in
the numerical values of the utility as long as the values reflect a similar trend.
Compared to the toxicity-efficacy trade-off contour approach used by the
EffTox assign, the utility approach has several advantages. First, it is simpler
and only requires elicitation of the utility score of element outcomes that have
transparent clinical interpretation. In contrast, the process of constructing
the toxicity-efficacy trade-off contour is complicated. More problematically, it
assumes that the contour must follow a specific function form, equation (8.1),
which lacks clinical interpretation and justification, and is likely to be false.
In our experience, it is often much easier for clinicians to understand and
quantify the risk-benefit trade-off based on patient outcomes than based on
toxicity and efficacy probabilities.
Second, the utility approach is highly flexible and includes other risk-
benefit trade-off methods as special cases (Zhou et al., 2019; Lin et al., 2020b).
For example, by imposing the constraint υ00 + υ11 = 100, we have

u = 100p01 + υ00 p00 + υ11 p11 + 0p10

= (υ00 + υ11 )p01 + υ00 p00 + υ11 p11
= υ00 (p01 + p00 ) + υ11 (p01 + p11 )
 
υ00
= υ11 πE + (1 − πT ) ,
υ11
Finding Optimal Biological Dose 165

where πT = p11 + p10 and πE = p01 + p11 are the marginal probabilities of
toxicity and efficacy at dose d, respectively. Therefore, the risk-benefit trade-
off approach based on the marginal toxicity and efficacy probabilities,

u0 = πE − wπT , (8.3)

is a special case of (8.2) with w = υ00 /υ11 and υ00 + υ11 = 100. For example,
the utility shown in Table 8.1 satisfies υ00 +υ11 = 100, and thus it is equivalent
to the simplified trade-off

u0 = πE − 2/3πT . (8.4)

This toxicity–efficacy trade-off means that a 1% toxicity increase in conjunc-

ture with a w% increase of efficacy does not change the desirability of the
treatment. In other words, the weight w = υ00 /υ11 implicitly determines the
relative importance between toxicity and efficacy.
Another important special case of the utility approach is to set υ00 = 0
and υ11 = 100 to turn off the toxicity-efficacy trade-off. Under this utility, we
favor the dose with the highest efficacy, which is appropriate for trials aiming
to identify the dose yielding the highest efficacy. The safety of the dose is
safeguarded by the admissible rules described below.
Third, the utility approach is highly scalable. It is directly applicable
to categorical yT and yE with more than two levels. Table 8.2 provides
an example of the utility approach for a trinary toxicity endpoint (mi-
nor/moderate/severe) and a trinary efficacy endpoint (progressive disease
(PD)/stable disease (SD)/partial response or complete response (PR/CR)).
In addition, Lin et al. (2020b) shows that it is similarly straightforward to
accommodate more than two endpoints. In contrast, it is not clear how to
apply the toxicity-efficacy trade-off contour approach to these cases.

8.3.2 Statistical model

Let D = (y01 , y00 , y11 , y10 ) denote the observed data at dose d, where
(y01 , y00 , y11 , y10 ) denote the numbers of patients having outcomes (no tox-
icity, efficacy), (no toxicity, no efficacy), (toxicity, efficacy), and (toxicity, no

TABLE 8.2: Utility table for trinary efficacy and toxicity endpoints

Efficacy
Toxicity CR/PR SD PD
Minor 100 60 35
Moderate 65 30 25
Severe 30 15 0
PD: progressive disease, SD: stable disease, PR: partial response, CR:
complete response.
166 Model-Assisted Bayesian Designs for Dose Finding and Optimization

efficacy), respectively. Here, n = y01 + y00 + y11 + y10 is the number of patients
treated at dose d. The numbers of patients who have experienced toxicity and
efficacy are nT = y11 + y10 and nE = y01 + y11 , respectively. Subscript j (i.e.,
dose level) is suppressed in this and next sections for notational brevity.
Under the Bayesian paradigm, the U-BOIN design assumes a Dirichlet-
multinomial model,

(y01 , y00 , y11 , y10 ) ∼ Multinomial(n; p01, p00 , p11 , p10 )

(p01, p00 , p11 , p10 ) ∼ Dirichlet(α01 , α00 , α11 , α10 )

where (α01 , α00 , α11 , α10 ) are hyperparameters, representing the prior numbers
of events for four outcomes, with the total prior sample size n0 = α01 + α00 +
α11 + α10 . The posterior distribution of (p01, p00 , p11 , p10 ) based on D is

(p01, p00 , p11 , p10 ) | D ∼ Dirichlet(α01 + y01 , α00 + y00 , α11 + y11 , α10 + y10 ).

Therefore, the posterior mean estimate of u is given by

u = p̂01 υ01 + p̂00 υ00 + p̂11 υ11 + p̂10 υ10 , (8.5)

where p̂ab = (αab + yab )/(n + n0 ) is the posterior mean estimate of pab , a, b ∈
(0, 1).
It follows that the marginal posterior distributions of πT and πE follow
the beta distribution,

πT | D ∼ Beta(αT + nT , βT + n − nT ), (8.6)
πE | D ∼ Beta(αE + nE , βE + n − nE ),

where the hyperparameters αT = α11 + α10 , αE = α11 + α01 , βT = α01 +

α00 , βE = α10 + α00 . For example, specifying αT = αE = βT = βE = 1 (e.g.,
α01 = α00 = α11 = α10 = 0.5) leads to Unif(0, 1) prior distributions for πT
and πE .

8.3.3 Admissible rules

The U-BOIN design uses similar admissible rules as those of the EffTox design.
Let φT and φE be the clinician-specified toxicity upper limit and efficacy lower
limit, respectively. Generally, φE can take the value of the target response rate
specified for a standard phase II trial. Because U-BOIN considers the toxicity–
efficacy trade-off, the value of φT should be set slightly higher (e.g, 0.05) than
the target toxicity rate used in conventional toxicity-based phase I designs. For
example, if 25% is an appropriate target toxicity rate that the conventional
phase I design used, then φT = 30% is a reasonable choice for U-BOIN.
Given the observed interim data D, a dose dj is defined as acceptable (or
admissible) when it satisfies both of the following two criteria:

(Safety criterion) Pr(πT j > φT | D) ≤ cT , (8.7)

(Efficacy criterion) Pr(πEj < φE | D) ≤ cE , (8.8)
Finding Optimal Biological Dose 167

where cE and cT is a probability threshold. In general, we recommend

cT = 0.95 and cE = 0.90 as the default, which can be calibrated by simu-
lation. These cutoffs seem high, but actually are appropriate as their purpose
is to rule out excessively toxic and ineffective doses. Among admissible doses,
the dose assignment rule will allocate patients to the most desirable dose. In
other words, even if the admissible dose set includes some doses that are not
particularly safe or efficacious, the design likely will not assign patients to
these suboptimal doses. Due to the large uncertainty of small sample sizes,
using small values for cT and cE will inadvertently eliminate the doses that are
actually admissible and thus affect the operating characteristics of the design.
Evaluation of the admissibility of a dose is straightforward under the
U-BOIN design because f (πT j | D) and f (πEj | D) follow beta distribu-
tions given by (8.6). In contrast, evaluation of the admissibility of a dose is
much more complicated in the EffTox design and requires fitting the com-
plex Gumbel-Morgenstern copula model using a Markov chain Monte Carlo
method. Let A(D) denote the set of admissible doses. During the trial, only
doses in A(D) can be used to treat patients.

8.3.4 Dose-finding algorithm

The U-BOIN design consists of two seamless, connected stages (Figure 8.3).
The objective of Stage I is to quickly explore the dose space to identify a set
of admissible doses that are reasonably efficacious and safe for Stage II. The
objective of Stage II is to optimize the dose based on the utility. In Stage
I, we conduct dose escalation based on the BOIN design using only yT , but
yE is also collected and will be used for decision making in Stage II. Given
the exploratory nature of Stage I, if yT has more than two categories, we
dichotomize it as DLT/no-DLT to facilitate the exploration of the dose space.
This is in line with clinical practice and serves well for the purpose of Stage I.
Consider a phase I-II trial with J doses, d1 < · · · < dJ , let π̂T j = nT j /nj
be the observed toxicity rate of dose dj . Based on the BOIN’s dose escalation
and de-escalation boundaries (λe , λd ), the dose-finding algorithm of U-BOIN
in Stage I proceeds as follows:
A1 Patients in the first cohort are treated at the lowest dose, or a
physician-specified dose.
A2 Suppose j is the current dose level. To assign a dose to the next
cohort of patients:
• If π̂T j ≤ λe , escalate the dose to level j + 1.
• If π̂T j > λd , de-escalate the dose to level j − 1.
• Otherwise, i.e., λe < π̂T j ≤ λd , stay at the current dose level
j.
168 Model-Assisted Bayesian Designs for Dose Finding and Optimization

Start at the
pre-specified
starting dose

Compute the toxicity rate at

the current dose 𝜋$ #$
≤ 𝜆% > 𝜆!
within (𝜆% , 𝜆! ]

Stage I Escalate the Stay at the De-escalate

dose current dose the dose

Treat the next cohort of patients

No
Does the number of patients on
any dose reach 𝑠" ?
Yes
Based on data in both Stages I and II, Stop the
Yes trial & no
estimate dose utility, and
determine admissible set 𝒜(𝐷) dose
selected
No
Yes Escalate
The toxicity rate at the highest
tried dose 𝜋$ #$∗ ≤ 𝜆% ? the dose to
𝑗∗ + 1
No
Stage II
Treat the next cohort of patients at the
admissible dose with the largest
estimated utility

No Does the number of patients Yes

on any dose reach 𝑠' or Select OBD
the total sample size reach 𝑁?

FIGURE 8.3: Diagram of the utility-based Bayesian optimal interval (U-

BOIN) design with the pick-the-winner strategy for patient allocation in
stage II.

A3 Repeat Step A2 until the number of patients treated on one of the

doses reaches s1 , and then move to Stage II. Zhou et al. (2019)
recommended s1 = 12 as the default value, while s1 = 9 to 15
generally yields good operating characteristics.
Stage II proceeds as follows:
B1 Let j ∗ denote the highest dose level that has been tried. If π̂T j ∗ ≤ λe
and j ∗ is not the highest dose in the trial, escalate the dose to level
j ∗ + 1 for treating the next cohort of patients; otherwise, proceed
to Step B2.
Finding Optimal Biological Dose 169

B2 Update the admissible dose set A(D) based on the interim data D.
If no dose is admissible, terminate the trial and no dose should be
selected as OBD. Otherwise, use one of the following strategies to
allocate the next cohort of patients:
(i) Equally randomize the next cohort of patients to a dose in
A(D).
(ii) Assign the next cohort of patients to the admissible dose that
has the largest posterior mean utility û.
(iii) Adaptively randomize the next cohort of patients to a dose in
A(D) with a probability proportional to its û.
B3 Repeat Steps B1 and B2 until the maximum sample size N is ex-
hausted or the number of patients treated at one of the doses reaches
s2 (> s1 ), and then select the OBD as the admissible dose (i.e.,
∈ A(D)) that has the largest posterior mean utility.
In Stage I, following the BOIN design, we impose an overdose control rule as
follows: if Pr(πT j > φT | Dj ) > 0.95 and nj ≥ 3, dose level j and higher are
eliminated from the trial, and the trial is terminated if the lowest dose level
is eliminated, where Pr(πT j > φT | Dj ) > 0.95 is evaluated based a beta-
binomial model with the uniform prior. Once the trial moves on to Stage II,
this overdose control rule is seamlessly merged as the safety criterion of the
inadmissible rule defined in Section 8.3.3.
For Stage II Step B1, the reason that we perform dose escalation when
π̂T j ∗ ≤ λe is to allow the trial to continue exploring the dose space, given
that the highest tried dose is safe, to reduce the risk of being stuck at a local
suboptimal dose due to large variation caused by small sample size.
In Stage II Step B2, the three strategies generally yield similar performance
in identifying the OBD, but have a different emphasis on patient allocation.
Determining which one to use should be based on specific trial considerations.
Strategy (i) (i.e., equal randomization) is the easiest to implement and al-
lows more uniform learning of admissible doses. In addition, equal randomiza-
tion does not require efficacy readout, avoiding the common logistic difficulty
that the efficacy endpoint may take a relatively long time to be ascertained.
Also, equal randomization is useful to balance confounders across admissible
doses to facilitate reliable identification of OBD particularly when patients
are highly heterogenous (e.g., all-comer trials). Equal randomization can be
modified to a fixed-ratio randomization. For example, based on the number of
patients allocated at Stage I, we choose a fixed randomization ratio such that
at the end of Stage II, the sample size at each admissible dose is expected to
be the same.
In comparison, Strategy (ii), i.e., pick the winner, aims to maximize patient
benefit by assigning the next cohort to the currently estimated optimal dose.
This approach tends to assign more patients to OBD than Strategy (i). As
a trade-off, in certain scenarios it may have a slightly lower probability of
170 Model-Assisted Bayesian Designs for Dose Finding and Optimization

selecting OBD as it may be stuck at local suboptimal doses, especially when

the sample size is small. This issue, however, is generally minor. Due the small
sample size, the estimated optimal dose often associates high uncertainty, and
varies when data accumulate. Thus, Strategy (ii) actually behaves rather like
adaptive randomization, i.e., Strategy (iii), although the latter exibits slightly
more randomness by design. Compared to Strategy (iii), Strategy (ii) has
similar performance, but is easier to understand and implement as described
below. Thus, we generally recommend Strategy (i) or (ii). See Zhou et al.
(2019) for more details and the numerical comparison of the three patient
allocation strategies.
Compared to the EffTox design, one prominent advantage of U-BOIN is
that its dose assignment rules can be pretabulated in decision tables and
included in the trial protocol before the trial starts. To conduct the trial, no
complicated calculation is needed. The investigator can simply use the decision
tables to determine dose assignment. Specifically, Stage I and Stage II Step
B1 are guided by the BOIN design decision table (see Section 3.4 for details),
and Steps B2 and B3 can be implemented based on the pre-calculated utility
table. Table 8.3 provides the utility table when the number of patients at a
dose is three or six, with the elicited utility scores in Table 8.1, φT = 0.3,
φE = 0.3, cT = 0.95, and cE = 0.9. A similar utility table can be generated
for any given n using the U-BOIN software described later. To illustrate the
use of the table, suppose at an interim of the trial, at d1 , three patients were
treated with no toxicity and one efficacy, and at d2 , six patients were treated
with two toxicity and two efficacy. By looking up the table, the estimated
utility of d1 is 57.5 and the estimated utility of d2 is 47.1. Thus, if we use
Strategy (ii) to allocate patients, the next cohort should be treated at dose
level one as it has higher desirability.
Of note, in Table 8.3, û depends only on nT and nE because the utility
score used (Table 8.1) satisfies υ00 + υ11 = 100. In this case, as described
previously, u is a function of πT and πE , see equation (8.3), thus û depends
only on nT and nE . In the general case that υ00 + υ11 6= 100, û is a function
of y10 , y00 , y01 , and y11 . The utility table still can be constructed, but with
more entries because of more possible data patterns given n.

8.3.5 Operating characteristics

Zhou et al. (2019) conducted a simulation study to compare the operating
characteristics of U-BOIN to the EffTox design. The simulation considered
J = 5 doses, and the total sample size N = 54 patients with s1 = 12 (i.e.,
Stage I stopping cutoff). The lower limit for efficacy was φE = 0.2 and the
upper limit for toxicity was φT = 0.30. For the admissible rules (i.e., equations
(8.7) and (8.8)), probability cutoffs were set as cT = 0.95 and cE = 0.9.
Table 8.4 shows a part of the simulation results, see Zhou et al. (2019) for
details. In general, U-BOIN outperformed EffTox with higher OBD selection
percentages and often allocated more patients to OBD. EffTox performed
Finding Optimal Biological Dose 171

TABLE 8.3: Estimated mean utility (i.e., û) given the number of patients
treated at a dose is 3 or 6 with the elicited utility scores in Table 8.1, toxicity
upper limit φT = 0.3 and efficacy lower limit φE = 0.3.

nT nE û nT nE û
n=3
0 0 42.5 1 3 77.5
0 1 57.5 2 0 22.5
0 2 72.5 2 1 37.5
0 3 87.5 2 2 52.5
1 0 32.5 2 3 67.5
1 1 47.5 >2 Any 0
1 2 62.5
n=6
0 0 41.4 2 1 38.6
0 <1 0 2 1 38.6
0 1 50 2 2 47.1
0 2 58.6 2 3 55.7
0 3 67.1 2 4 64.3
0 4 75.7 2 5 72.9
0 5 84.3 2 6 81.4
0 6 92.9 3 <1 0
1 <1 0 3 1 32.9
1 1 44.3 3 2 41.4
1 2 52.9 3 3 50
1 3 61.4 3 4 58.6
1 4 70 3 5 67.1
1 5 78.6 3 6 75.7
1 6 87.1 >3 Any 0
2 <1 0
Note: n is the number of the patients at a specific dose, nT and nE are
respectively the numbers of toxicity and efficacy at that dose. “0” denotes
that the dose should be eliminated as not admissible because of high toxicity
or low efficacy.

well in scenario 2, but (relatively) poorly in scenarios 1 and 3, showing its

sensitivity to model misspecification, i.e., how well the true dose-toxicity or
dose-efficacy relationship can be approximated by the model assumed.

8.4 BOIN12 design

BOIN12 is another efficient model-assisted design to find OBD. BOIN12 takes
the single-stage approach, which uses toxicity and efficacy jointly through-
out the trial to determine dose escalation/de-escalation. This differs from
172 Model-Assisted Bayesian Designs for Dose Finding and Optimization

TABLE 8.4: Comparison of U-BOIN with EffTox, including the selection per-
centage (Sel %) and the average number of patients treated at each dose (No.
of pts). The optimal biological dose (OBD) is bolded.

Dose Level

Design 1 2 3 4 5
Scenario 1
πT 0.02 0.15 0.30 0.45 0.60
πE 0.20 0.65 0.65 0.65 0.65
u 43.0 69.0 63.0 56.0 50.0
EffTox Sel % 2.0 50.0 45.0 2.0 0.0
No. of pts 4.3 22.7 23.8 2.8 0.4
U-BOIN Sel % 1.7 72.9 22.4 2.8 0.0
No. of pts 6.2 29.9 13.8 3.5 0.5
Scenario 2
πT 0.03 0.08 0.15 0.28 0.40
πE 0.10 0.22 0.60 0.60 0.60
u 36.0 43.0 66.0 60.0 55.0
EffTox Sel % 0.0 4.0 60.0 29.0 7.0
No. of pts 3.4 4.9 26.4 14.2 5.1
U-BOIN Sel % 1.1 3.2 65.7 24.9 4.3
No. of pts 4.9 7.5 24.4 12.7 4.3
Scenario 3
πT 0.05 0.07 0.10 0.12 0.16
πE 0.35 0.45 0.50 0.55 0.75
u 53.0 59.0 61.0 64.0 75.0
EffTox Sel % 10.0 12.0 26.0 24.0 29.0
No. of pts 8.6 7.9 14.2 11.1 12.1
U-BOIN Sel % 5.9 11.7 13.1 13.6 55.7
No. of pts 7.0 8.8 9.0 9.1 20.1
Note: πT and πE are the toxicity probability and efficacy probability,
respectively. u is the mean utility or desirability.

U-BOIN’s two-stage approach, which first performs toxicity-based dose es-

calation, and then switches to dose optimization jointly based on toxicity and
efficacy. As a result, U-BOIN and BOIN12 demonstrate different design char-
acteristics and are suitable for different trial objectives.
U-BOIN is more appropriate for the case where both identification of OBD
and evaluation of the dose-toxicity profile of the drug are of interest, whereas
BOIN12 is a better choice when the primary goal is to identify OBD. For
the purpose of finding the OBD, BOIN12 is more efficient and often requires
a smaller sample size than U-BOIN. For example, for drugs with a low, flat
dose-toxicity curve and an increase-then-plateau dose-efficacy curve, U-BOIN
will first escalate all the way up to the highest dose, and then turn around
Finding Optimal Biological Dose 173

to search for OBD. In contrast, BOIN12 will quickly reach and stay at OBD
because it considers both efficacy and toxicity from the beginning of the trial.
Compared to BOIN12, U-BOIN has some operational advantages. BOIN12
requires that the efficacy endpoint must be scored quickly enough so that the
dose assignment decision is timely for each new cohort. U-BOIN is lenient
on this requirement as its first stage often “buys” sufficient time to evaluate
the efficacy endpoint, especially when the equal or fixed-ratio randomization
is employed in Stage II of U-BOIN. Section 8.5 introduces an extension of
BOIN12 that is able to handle delayed efficacy or/and toxicity endpoints.

8.4.1 Utility estimation using quasi-binomial likelihood

Similar to U-BOIN, BOIN12 uses utility to quantify the risk-benefit trade-
off and desirability of each dose. As described in Section 8.3.1, the utility
approach has the advantage of being easy to understand and implement, and
it is highly flexible and scalable to accommodate various types of endpoints.
We adopt the same notation as the previous section and focus on a binary
toxicity endpoint yT and binary efficacy endpoint yE .
Let (p01 , p00 , p11 , p10 ) denote the probabilities of observing the four possi-
ble values of (yT , yE ), i.e., (0, 1), (0, 0), (1, 1), and (1, 0), at a certain dose d.
For notational brevity, we suppress that (p01 , p00 , p11 , p10 ) is a function of d.
Let (υ01 , υ00 , υ11 , υ10 ) denote the utility scores, which are elicited from clin-
icians to quantify the desirability of the four possible outcomes of (yT , yE ),
see Table 8.1 for an example. Similar to Section 8.3.1, we set the utility of the
most desirable outcome υ01 = 100 to fix the scale of the score. The desirability
(or mean utility) of dose d is given by

u = p01 υ01 + p00 υ00 + p11 υ11 + p10 υ10 . (8.9)

A higher value of u indicates a higher dose desirability in terms of the risk-

benefit trade-off.
The BOIN12 design makes decisions of dose escalation/de-escalation based
on f (u | D), the posterior distribution of u given the interim data D. A straight-
forward approach is to derive the posterior f (u | D) based on the definition
(8.9) and the Dirichlet-multinomial model, described in Section 8.3.2.
Lin et al. (2020b) proposed a more straightforward approach by di-
rectly modeling u based on the quasi-binomial likelihood theory (Papke and
Wooldridge, 1996; Yuan et al., 2007). Define the standardized desirability
ũ = u/100. Because ũ ∈ [0, 1] and takes a form of the weighted average of
(p01 , p00 , p11 , p10 ), it can be viewed as a probability and modeled using the
binomial distribution with “quasi-binomial” data (x, n), where
100y01 + υ00 y00 + υ11 y11
x= . (8.10)
100
174 Model-Assisted Bayesian Designs for Dose Finding and Optimization

This quasi-number x can be interpreted as the number of “events” ob-

served from n patients treated at dose d, given the event probability ũ. This
interpretation follows because E(x) = n × ũ. Unlike the standard binomial
data, x is a real value that can take any value between 0 and n instead of
integers only, and thus it is known as quasi-binomial. When υ00 + υ11 = 100,
the expression of x can be further simplified to

υ11 nE + υ00 (n − nT )
x= .
100
The quasi-binomial likelihood of the unknown ũ based on the interim data D
is
n−x
L(D | ũ) ∝ ũx (1 − ũ) .
Under the Bayesian framework, assigning ũ a Beta prior, i.e., ũ ∼
Beta(αu , βu ), the posterior distribution of ũ arises as

ũ | D ∼ Beta(αu + x, βu + n − x). (8.11)

By default, the non-informative uniform prior distribution with αu = βu = 1

is used. A similar quasi-binomial approach was used by Yuan et al. (2007) to
model the normalized equivalent toxicity score that maps toxicity grades into
a fraction of dose limiting toxicity under the parametric CRM model.

8.4.2 Admissible rules and dose comparison

BOIN12 uses the same admissible rules as those of U-BOIN, described in Sec-
tion 8.3.3. Briefly, let φT and φE denote the toxicity upper limit and efficacy
lower limit, respectively, specified by clinicians. Generally, φE may take the
value of the target response rate specified for a standard phase II trial. The
value of φT should be set slightly higher (e.g., 0.05) than the target toxicity
rate used in conventional toxicity-based phase I designs to provide additional
space for toxicity–efficacy trade-off. For example, if 30% is an appropriate tar-
get toxicity rate that the conventional phase I design used, then φT = 35% is
a reasonable choice for BOIN12.
Given the interim data D, a dose dj is defined as acceptable (or admissi-
ble) when it satisfies both of the following two criteria:

(Safety criterion) Pr(πT j > φT | D) ≤ cT ,

(Efficacy criterion) Pr(πEj < φE | D) ≤ cE ,
where cE and cT are probability thresholds. These two posterior quantities
can be evaluated based on the posterior of πT and πE , given by equation
(8.6). Denote the set of admissible doses by A(D), which should be updated
with interim data D. Only doses in A(D) can be used to treat patients. In
general, we recommend cT = 0.95 and cE = 0.90 as the default, which can be
Finding Optimal Biological Dose 175

further calibrated by simulation, see Section 8.3.3 for more discussion on the
choice of cT and cE .
In BOIN12, the posterior probability PPj = Pr(uj > ub | Dj ) is used to
determine the most desirable dose within A(D) for treating a new cohort of
patients, where Dj is the data observed at dj . Here, ub is a prespecified utility
benchmark used to evaluate dose desirability. Between two admissible doses
dj and dj 0 ,
• If PPj > PPj 0 , then dj is more desirable than dj 0 .
• If PPj < PPj 0 , then dj is less desirable than dj 0 .
• If PPj = PPj 0 , then dj and dj 0 are equally desirable.
In principle, the third case occurs if and only if Dj = Dj 0 . The posterior prob-
ability PPj automatically accounts for the uncertainty in estimating u, leading
to more reasonable decision making of dose escalation and de-escalation. In
contrast, most existing designs use the point estimate û = x/n to choose the
optimal dose, which ignores the uncertainty of the estimate. To see the im-
portance of accounting for estimation uncertainty, suppose dose dj has been
used to treat two patients and ûj = 60, and dose dj 0 has been used to treat 10
patients and ûj 0 = 59. Although ûj > ûj 0 , because ûj 0 is much more reliable, a
better decision actually is to choose dj 0 for treating the next cohort of patients
as stipulated by the posterior probability approach.
Evaluating PPj requires specifying a benchmark ub for comparison. The
value of ub can be elicited by clinicians to reflect their expectation. Lin et al.
(2020b) recommended the following default value of ub that yields desirable
operating characteristics in a variety of scenarios. Given φT and φE and as-
suming independence between toxicity and efficacy, the highest utility that is
deemed undesirable is given by

u = 100φE (1 − φT ) + υ00 (1 − φT )(1 − φE ) + υ11 φT φE .

The recommended default value of ub is

ub = (100 + u)/2,

representing that a utility value lies in the middle between u and the maximum
utility 100. Alternatively, one can also use the weighted average as the choice
for ub , i.e., ub = ω100 + (1 − ω)u, where ω ∈ [0, 1] is a non-negative weight.
In general, the weight ω controls the aggressiveness of the trial: the larger the
value of ω, the more aggressive the dose exploration.
One prominent feature of BOIN12 is that the desirability can be pre-
calculated and included in the trial protocol, due to the use of the quasi-
binomial likelihood. Specifically, given the maximum sample size N , all
possible outcome combinations (n, nT , nE ) (or (y01 , y00 , y11 , y10 )) can be enu-
merated, and the corresponding posterior probabilities PPj can be computed
based on (8.11). By sorting all the possible values of PPj from the smallest
to the largest, we can assign the ordered value from 0, 1, . . . to each possible
176 Model-Assisted Bayesian Designs for Dose Finding and Optimization

outcome combination. These ordered values are called rank-based desirability

scores (RDS). RDS allows for a simpler presentation of the desirability of a
dose using integers. Comparing the posterior probabilities PPj between two
doses is equivalent to comparing their corresponding RDS; therefore, RDS
can be used as a convenient way to identify a more desirable dose. For ex-
ample, based on the utility in Table 8.1, Table 8.5 lists RDS with a cohort
size of three to nine patients treated at a dose, with the highest acceptable
toxicity probability φT = 0.35 and the lowest acceptable efficacy probability
φE = 0.25.
During the trial, the desirability of a dose can be simply determined as
follows: count the number of patients treated at that dose n, the number
of patients who experienced toxicity nT , and the number of patients who
experienced efficacy nE (and the number of patients who experienced efficacy
without toxicity when υ00 + υ11 6= 100), then look up Table 8.5 to determine
its RDS.
For example, suppose at a certain point of the trial, the numbers of pa-
tients treated at the first three doses are three, six, and three; the numbers
of toxicities are zero, one, and two; and the numbers of efficacy outcomes are
zero, three, and one. By looking up Table 8.5, the RDS of the first three doses
are 35, 56, and 31, respectively. Because dose d2 has the highest RDS, it is
more desirable than d1 and d3 . Table 8.5 assumes a cohort size of three, but
the BOIN12 design is flexible and can allow any prespecified cohort size. The
RDS table can be easily generated using the BOIN12 companion software.
It is worth noting the difference between BOIN12 and U-BOIN. U-BOIN
uses the posterior mean of u (i.e., û) to determine the most desirable dose
within A(D) for treating a new cohort of patients. The utility table of U-
BOIN (e.g., Table 8.3) lists the value of û for each possible interim data, not
the rank of û. Nevertheless, in principle, for U-BOIN, a RDS table similar to
that of BOIN12 can also be developed by replacing the value of û with its
rank among all possible values.

8.4.3 Dose-finding rules

Suppose that J dose levels are investigated, d1 < · · · < dJ . We use subscript
j to indicate variables associated with dose dj . Let π̂T j = nT j /nj denote the
observed toxicity rate at dose dj , and (λe , λd ) denote the dose escalation and
de-escalation boundaries from the standard BOIN design. The dose-finding
rule of the BOIN12 design is illustrated in Figure 8.4 and described as follows:
1. Patients in the first cohort are treated at the lowest dose d1 , or the
physician-specified dose.
2. Based on the interim data D, we obtain RDS for each dose using
the RDS table (e.g., Table 8.5). Suppose j is the current dose level.
To assign a dose to the next cohort of patients,
Finding Optimal Biological Dose 177

TABLE 8.5: Rank-based desirability score (RDS) table for the BOIN12 design.

n nT nE RDS n nT nE RDS n nT nE RDS

0 0 0 60 6 3 2 22 9 2 4 45
3 0 0 35 6 3 3 38 9 2 5 58
3 0 1 55 6 3 4 51 9 2 6 70
3 0 2 76 6 3 5 67 9 2 7 83
3 0 3 91 6 3 6 81 9 2 8 92
3 1 0 24 6 4 0 1 9 2 9 98
3 1 1 44 6 4 1 6 9 3 0 E
3 1 2 63 6 4 2 15 9 3 1 7
3 1 3 80 6 4 3 27 9 3 2 14
3 2 0 13 6 4 4 42 9 3 3 25
3 2 1 31 6 4 5 56 9 3 4 36
3 2 2 48 6 4 6 72 9 3 5 49
3 2 3 69 6 ≥ 5 Any E 9 3 6 61
3 ≥ 3 Any E 9 0 0 E 9 3 7 74
6 0 0 22 9 0 1 25 9 3 8 85
6 0 1 38 9 0 2 36 9 3 9 94
6 0 2 51 9 0 3 49 9 4 0 E
6 0 3 67 9 0 4 61 9 4 1 3
6 0 4 81 9 0 5 74 9 4 2 9
6 0 5 93 9 0 6 85 9 4 3 17
6 0 6 100 9 0 7 94 9 4 4 29
6 1 0 15 9 0 8 99 9 4 5 40
6 1 1 27 9 0 9 102 9 4 6 53
6 1 2 42 9 1 0 E 9 4 7 65
6 1 3 56 9 1 1 17 9 4 8 78
6 1 4 72 9 1 2 29 9 4 9 88
6 1 5 87 9 1 3 40 9 5 0 E
6 1 6 96 9 1 4 53 9 5 1 2
6 2 0 8 9 1 5 65 9 5 2 5
6 2 1 19 9 1 6 78 9 5 3 10
6 2 2 34 9 1 7 88 9 5 4 20
6 2 3 47 9 1 8 97 9 5 5 32
6 2 4 64 9 1 9 101 9 5 6 45
6 2 5 77 9 2 0 E 9 5 7 58
6 2 6 90 9 2 1 10 9 5 8 70
6 3 0 4 9 2 2 20 9 5 9 83
6 3 1 12 9 2 3 32 9 ≥ 6 Any E
Note: n is the number of the patients at a specific dose, nT and nE are respectively the
numbers of toxicity and efficacy at that dose. “E” denotes that the dose should be
eliminated because it does not satisfy the safety and efficacy admissible criteria (i.e., not
admissible because of high toxicity or low efficacy).
178 Model-Assisted Bayesian Designs for Dose Finding and Optimization

Start at the lowest dose or

prespecified starting dose

Treat a patient or a
cohort of patients

Stop the trial and Yes Reach the maximum

select OBD sample size?

No

> 𝜆! ≤ 𝜆"
Compute the DLT rate
at current level j

Within (𝜆" , 𝜆! ]

< 𝑁∗
Count the number of
patients at level j

≥ 𝑁∗

Choose a level from Choose a level from

De-escalate to
{𝑗 − 1, 𝑗} using the {𝑗 − 1, 𝑗, 𝑗 + 1} using the
level 𝑗 − 1
desirability score table desirability score table

Target toxicity rate 𝝓𝑻 0.20 0.25 0.30 0.35 0.40

Escalation boundary 𝜆" 0.157 0.197 0.236 0.276 0.316

De-escalation boundary 𝜆# 0.238 0.298 0.359 0.419 0.480

FIGURE 8.4: The schema of the Bayesian optimal interval phase I–II
(BOIN12) design, where (λe , λd ) are a pair of optimized dose escalation and
de-escalation boundaries adopted from the BOIN design, and N ∗ is a pre-
specified sample size cutoff (e.g., N ∗ = 6). DLT, dose-limiting toxicity; OBD,
optimal biologic dose.

(a) If π̂T j > λd , de-escalate the dose to level j − 1.

(b) If π̂T j > λe and nj ≥ N ∗ , say N ∗ = 6, choose the level from
{j − 1, j} that has a higher RDS.
(c) Otherwise, if λe < π̂T j ≤ λd and nj < N ∗ , or π̂T j ≤ λe ,
choose the level from {j − 1, j, j + 1} that has the highest RDS.
3. Repeat Step 2 until the maximum sample size N is reached.
Finding Optimal Biological Dose 179

Step 2(a) says that if the observed toxicity rate is high (i.e., π̂T j > λd ),
we should de-escalate the dose. Step 2(b) says that if there are sufficient data
(nj ≥ N ∗ ) to support that the toxicity probability of the current dose dj
is moderate (i.e., λe < π̂T j ≤ λd ), escalating the dose to dj+1 may cause
overdosing. We then examine whether the adjacent lower dose dj−1 provides
a better treatment benefit (i.e., desirability) than the current dose dj . If so,
we treat the next cohort of patients at dose level j − 1; otherwise stay at
the current level j. Rule 2(c) says that if the observed toxicity rate at the
current dose dj is low (π̂T j ≤ λe ) or moderate ( λe < π̂T j ≤ λd ) but with high
uncertainty (n < N ∗ ), then we examine whether the adjacent higher dose dj+1
or the lower dose dj−1 can provide a better treatment benefit (i.e., desirability)
than the current dose dj . The next cohort of patients will be treated at the
dose with the highest desirability among the adjacent doses. In Step 2(b), the
higher dose dj+1 is not considered due to the fact that, given that there is
substantial evidence (i.e., nj ≥ N ∗ ) that dj is already close to the de-escalation
boundary λd and dj+1 is likely to be overly toxic. Therefore, N ∗ is a cutoff
for sufficiency of the data, and a larger value of N ∗ encourages more active
exploration of new doses. By default, Lin et al. (2020b) recommended N ∗ = 6.
When needed, N ∗ can be further calibrated by simulation to obtain certain
design properties, e.g., using a larger N ∗ encourages faster dose exploration.
During the trial, only admissible doses can be given to incoming patients,
and doses that are not admissible should be eliminated. In other words, the
decision rule in Step 2 should be applied only to the doses in A(D). For
example, in Step 2(c), if only dj−1 , dj ∈ A(D) and dj+1 ∈ / A(D), then when
we apply the rule 2(c), we should choose the level from {j − 1, j} that has the
highest RDS to treat the next cohort of patients. At any time, if all doses are
eliminated, the trial should be stopped with no dose selected as OBD.
For ethical considerations, BOIN12 always assigns the next cohort of pa-
tients to the dose with the highest estimate of desirability. With a small sample
size, this myopic approach may cause the dose finding to get stuck at a locally
optimal dose. To alleviate that issue, a dose exploration rule can be imposed:
treat the next cohort of patients at the next higher dose if the following three
conditions are all satisfied:
• The number of patients treated at the current dose dj is greater than eight.
• The observed toxicity rate π̂T j is less than the de-escalation boundary λd .
• The next higher dose has never been used for treating patients.
When the trial completes, the final OBD is selected based on a two-step
procedure: in the first step, we select MTD based on the target toxicity rate
φT so that any dose levels above the selected MTD are deemed overly toxic; in
the second step, we then choose the dose level that has the highest desirability
among the doses that are not higher than MTD. Specifically, at the end of the
trial, we first obtain the observed marginal toxicity rates π̂T j for each dose
level j = 1, . . . , J. To borrow information across dose levels, an isotonic re-
gression is performed on {π̂T j } through the pool-adjacent-violators algorithm
180 Model-Assisted Bayesian Designs for Dose Finding and Optimization

(Bril et al., 1984a), so that the isotonically transformed estimate π̃T j mono-
tonically increases with the dose. MTD, dMTD , is then selected as the dose
that has the estimated toxicity rate closest to the target toxicity rate φT , that
is,
dMTD = arg min |π̃T j − φT |. (8.12)
dj ∈{d1 ,...,dJ }

Next, we obtain the posterior estimate of the mean utility at each dose level
j = 1, . . . , J as follows:
xj + αu
ūj = . (8.13)
nj + αu + βu
The final OBD, dOBD , is then selected as the admissible dose that does not
exceed the estimated MTD, dMTD , and also yields the highest estimated mean
utility, that is,
dOBD = arg max {ūj } . (8.14)
dj ≤dMTD , dj ∈A(D)

When there is a tie, select the lower dose level.

8.4.4 Operating characteristics

Lin et al. (2020b) carried out a simulation study to compare the operating
characteristics of BOIN12, the model-based EffTox design, the toxicity and ef-
ficacy probability interval (TEPI) design (Li et al., 2017), and the conventional
design consisting of a dose escalation phase based on the 3+3 design followed
by a cohort expansion at the identified MTD (referred to as 3+3+CE). Five
scenarios with the marginal toxicity and efficacy rates of each of the five dose
levels considered in the simulation study are provided in Table 8.6. The high-
est acceptable toxicity probability is φT = 0.35, and the lowest acceptable
efficacy probability is φE = 0.25. The utility in Table 8.1 is used to define
the toxicity–efficacy trade-off. Patients are treated in cohorts of three, with a
maximum sample size of 36. The Simon’s two-stage design (Simon, 1989) is
used to monitor efficacy in the cohort expansion for 3+3+CE.
Figure 8.5 presents the simulation results based on 10,000 simulated trials
that quantify the operating characteristics of the designs, including the per-
centage of correct selection of OBD, the number of patients treated at OBD,
the number of patients treated at the overdoses (i.e., the doses with toxicity
rates > 0.35), and risk of poor allocation (i.e., the percentage of trials that
allocate < 36/5 = 7.2 patients to OBD). In general, BOIN12 has the best
overall performance among the four designs. It has the highest average per-
centage of correct selection of OBD and allocates the largest average number
of patients to OBD. In addition, BOIN12 is safer and more reliable than other
designs, as demonstrated by having amongst the smallest number of patients
treated at toxic doses and the lowest risk of poor allocation.
As expected, the 3 + 3 + CE design has the worst overall performance
because that design ignores the toxicity-efficacy trade-off and does not target
OBD. The performance of the EffTox design varies dramatically from one
scenario to another, depending on how well the assumed model reflects the true
Finding Optimal Biological Dose 181

TABLE 8.6: Five scenarios with the marginal toxicity and efficacy probabil-
ities, as well as mean utilities (πT , πE , u) in the simulation study. The mean
utility is based on the 2 × 2 utility table given in Table 8.1. The OBD that
maximizes the utility is bolded.

Dose level
Scenario
1 2 3 4 5
1 (πT , πE ) (0.05, 0.20) (0.12, 0.35) (0.27, 0.36) (0.35, 0.37) (0.50, 0.38)
u 50.0 56.2 50.8 48.2 42.8
2 (πT , πE ) (0.05, 0.01) (0.12, 0.05) (0.27, 0.30) (0.35, 0.60) (0.50, 0.60)
u 38.6 38.2 47.2 62.0 56.0
3 (πT , πE ) (0.03, 0.05) (0.05, 0.10) (0.20, 0.50) (0.22, 0.68) (0.45, 0.70)
u 41.8 44.0 62.0 72.0 64.0
4 (πT , πE ) (0.02, 0.05) (0.05, 0.15) (0.10, 0.40) (0.20, 0.40) (0.30, 0.40)
u 42.2 47.0 60.0 56.0 52.0
5 (πT , πE ) (0.15, 0.25) (0.20, 0.55) (0.25, 0.40) (0.35, 0.30) (0.45, 0.20)
u 49.0 65.0 54.0 44.0 34.0

dose–toxicity or dose–efficacy relationship. Compared with BOIN12, TEPI has

uniformly poor accuracy to identify OBD, a higher (i.e., more than doubled)
likelihood of overdosing patients, and a higher (i.e., almost doubled) risk of
poor allocation.

8.4.5 Extension to more complicated endpoints

One advantage of the quasi-binomial approach introduced in Section 8.4.1 is
that it can be easily generalized to accommodate different utility (or desir-
ability) functions. Let µ be any desirability function that can be calculated
based on the data D(d) observed at a specific dose d, let µmax and µmin be
the upper and lower limits of µ, respectively. Then the “quasi-binomial” data
(x̃, n) can be defined as follows:

n(µ − µmin )
x̃ = .
µmax − µmin

By substituting (x̃, n) to equation (8.11), the posterior desirability distribution

can be obtained and incorporated into the BOIN12 design.
As an illustration, we discuss the BOIN12 design when toxicity and efficacy
endpoints have more than two levels. Assume that the efficacy outcome has
three levels: complete response (CR), partial response (PR), and no response
(NR); and the toxicity outcome also has three levels: no/minor, moderate,
and severe. In immune checkpoint inhibitor trials, a portion of patients often
experience severe immune-related adverse events (irAEs), which may need
182 Model-Assisted Bayesian Designs for Dose Finding and Optimization

(a) Selection percentage of OBD

80
Selection percentage 70.2
62.7
58.6
60
55.9
47.9 49.6 50.5
47.0
43.1 41.0 42.0 43.3
38.0
40

34.1 34.0
28.6 26.6 28.4
22.9
20

13.2
0

1 2 3 4 5
Scenario
(b) Number of patients at OBD 17.0
Number of patients
15

12.5 12.9 12.9

12.2 12.0
11.3 10.9
10.2 10.3 9.8
9.4 9.4
10

9.0 8.4
8.1 7.6 8.1

5.1 5.6
5
0

1 2 3 4 5
Scenario
(c) Number of patients at overdoses 3+3+CE EffTox TEPI BOIN12
Number of patients
15

11.0
9.6
10

8.7

4.1 4.0
5

3.8 3.6
3.2 3.2 2.8
2.1 2.4
1.4 1.4 1.1 1.2
0.0 0.0 0.0 0.0
0

1 2 3 4 5
Scenario
(d) Risk of poor allocation
86.8
79.3
80

77.0
72.8 71.2
Percentage

66.3 64.3 65.9

60.1
60

57.1 55.8
47.5 50.3
45.6 43.9 41.2
40

32.4 33.1
29.3
20

16.3
0

1 2 3 4 5
Scenario

FIGURE 8.5: Percentages of correct selection of OBD, number of patients

treated at OBD, number of patients treated at overdoses, and risk of poor
allocations for the 3+3+CE, EffTox, TEPI, and BOIN12 designs, computed
based on 10,000 simulated trials under each of the five scenarios shown in
Table 8.6. The overdoses are defined as the doses with toxicity rates greater
than 0.35. The maximum sample size is 36. Risk of poor allocation is the
percentage of trials than allocates less than 36/5 = 7.2 patients to OBD.

separate consideration. In this case, the “severe” level of the toxicity outcome
can be used to indicate such irAEs. This results in nine possible outcomes,
denoted as outcome 1, . . . , 9, and the corresponding utility table is given in
Table 8.7.
Here, υ1 denotes the utility of the most desirable outcome (i.e., no/minor
toxicity and CR), thus setting υ1 = 100; and υ9 denotes the utility of the most
Finding Optimal Biological Dose 183

TABLE 8.7: Utility table for three-level efficacy and toxicity endpoints.

Efficacy
Toxicity CR PR NR
No/minor υ1 = 100 υ2 υ3
Moderate υ4 υ5 υ6
Severe υ7 υ8 υ9 = 0

undesirable outcome (i.e., severe toxicity and NR), thus setting υ9 = 0. Using
these two extreme cases as the reference, we elicit from clinicians the utility
scores of the other seven possible outcomes to reflect their clinical desirability.
For example, to reflect that irAE is highly undesirable, we can assign low
scores to υ7 and υ8 , e.g., υ7 = 30 and υ8 = 10.
Let pl denote the probability of observing the lth outcome l at dose d,
l = 1, . . . , 9. The desirability (or mean utility) of dose d is given by
9
X
µ= pl υl ,
l=1

and the standardized desirability is µ̃ = (µ−µmin )/(µmax −µmin ) = µ/100. Let

yl denote the number of patients experiencing outcome l, the “quasi-binomial”
outcome x̃ at dose d is
X 9
x̃ = yl υl /100.
l=1

As a result, the quasi-binomial modelling of the utility discussed in Section

8.4.1, and also the BOIN12 dose-finding rule, can be directly applied for such
a trial with three-level toxicity and three-level efficacy endpoints.

8.5 TITE-BOIN12 design

One practical issue encountered when applying BOIN12, especially in trials of
targeted or immunotherapy agents, is that of possible late-onset efficacy and
toxicity. As a motivating example, consider a phase I/II renal cell carcinoma
trial. The objective is to determine OBD of a novel targeted agent, combined
with nivolumab at a fixed dose of 3 mg/kg daily for two weeks, in patients
with metastatic renal cell carcinoma. Toxicity is defined as a grade 3 or 4 liver,
lung, gastrointestinal, or endocrine toxicity, or myelosuppression, within 45
days from the start of therapy according to Common Terminology Criteria for
Adverse Events (CTCAE) version 5. Efficacy is defined as achieving complete
response or partial response within 60 days. The expected accrual rate is three
patients per month. The challenge of designing this trial is that, on average,
184 Model-Assisted Bayesian Designs for Dose Finding and Optimization

six (or four) new patients will be enrolled by the time the previous cohort of
patients complete their efficacy (or toxicity) evaluation. The question is how
to treat these new patients in a timely fashion, given that the outcomes of
some of the previously treated patients are still pending.
Formally, let τT and τE denote the lengths of the assessment windows
for yT and yE , respectively. The value of τT and τE should be elicited from
clinicians and wide enough to capture all necessary toxicity and efficacy events
relevant to OBD determination. For example, τT may be the first cycle of the
therapy (e.g., 21 or 28 days), while τE may be also the first cycle of the therapy
(e.g., pharmacodynamic (PD) biomarkers), or several months after treatment
(e.g., tumor response). Let tT and tE denote the times to toxicity and efficacy,
respectively. For patients who will not experience toxicity or efficacy, define
tT = ∞ and tE = ∞. The relation between (yT , yE ) and (tT , tE ) is given as
follows:
yT = 1{tT ≤ τT }, and yE = 1{tE ≤ τE },
where 1{·} denotes the indicator function, which takes a value of 1 if the event
specified in the braces occurs and zero otherwise.
Suppose that n patients have been accrued and treated at dose d, we
use the subscript i to indicate the data for the ith patient, i = 1, . . . , n.
The relationship between the multinomial variable (y01 , y00 , y11 , y10 ) and the
individual patient-level data is
Pn Pn
y01 = i=1 1 {(yT i , yEi ) = (0, 1)} = i=1 1 {tT i > τT , tEi ≤ τE } ,
Pn Pn
y00 = i=1 1 {(yT i , yEi ) = (0, 0)} = i=1 1 {tT i > τT , tEi > τE } ,
Pn Pn
y11 = i=1 1 {(yT i , yEi ) = (1, 1)} = i=1 1 {tT i ≤ τT , tEi ≤ τE } ,
Pn Pn
y10 = i=1 1 {(yT i , yEi ) = (1, 0)} = i=1 1 {tT i ≤ τT , tEi > τE } .

Denote tU as the follow-up time. In cases of late-onset toxicity or efficacy, the

issue is that the time to toxicity or efficacy may not be observed as fast as
the patients’ interarrival time. As a result, tU may censor tT or tE or both
of them. In other words, when tU is smaller than tT or tE , yT or yE may be
unobservable. Section 5.2 provides statistical properties of these missing data,
and why we cannot simply ignore them in decision making.
As an extension of BOIN12, the TITE-BOIN12 design addresses the late-
onset issue by using the follow-up times of pending patients to facilitate real-
time decision making to choose the optimal dose for new patients (Zhou et al.,
2022). More precisely, by treating the outcome data of the pending patients as
missing, two approaches, based on different imputation techniques, have been
proposed: Bayesian data augmentation (BDA) (Little and Rubin, 2014) and
mean imputation based on the approximate likelihood (Lin and Yuan, 2020).

Bayesian data augmentation

BDA iterates between two steps: the imputation (I) step, in which the missing
yT and yE are imputed from their conditional posteriors, and the posterior
Finding Optimal Biological Dose 185

(P) step, in which the posterior samples of unknown parameters including

(p01 , p00 , p11 , p10 ) and ũ are simulated based on the imputed data. As the P
step is the same as BOIN12, we here focus on the I step, which involves the
conditional posteriors of missing yT and yE .
In the presence of late-onset toxicity or efficacy, there are three possible
missing patterns: (1) both yT and yE are missing, (2) only yT is missing, and
(3) only yE is missing. Define δT = 1{tU < τT ∩ tU < tT } as the missing
indicator for toxicity, which takes a value of 1 if yT is missing. Similarly,
define δE = 1{tU < τE ∩ tU < tE } as the missing indicator for efficacy. The
imputation steps for the three missing patterns are described as follows:
• When both yT and yE are missing, we impute (yT , yE ) by drawing a random
sample from the multinomial distribution with individual joint probabilities
given by
pab Sab
Pr {(yT , yE ) = (a, b) | (δT , δE ) = (1, 1)} = P1 P1 ,
a0 =0 b0 =0 pa0 b0 Sa0 b0

for a, b ∈ {0, 1}, where Sab = Pr {tT > tU , tE > tU | (yT , yE ) = (a, b)}. As-
suming “working” independence between tT and tE , we have

Sab = Pr(tT > tU | yT = a) Pr(tE > tU | yE = b).

Let wq = Pr(tq ≤ tU | yq = 1) denote a weight, which accounts for the

partial information from patients who have not completed the assessment of
yq , q ∈ {T, E}. The specification of wq will be discussed below. As Pr(tq >
tU | yq = 0) = 1, q ∈ {T, E}, we have S00 = 1, S01 = 1 − wE , S10 = 1 − wT ,
and S11 = (1 − wT )(1 − wE ).
• When yT is observed but yE is missing, we draw the missing value of yE
from a Bernoulli distribution with the response probability given by
 yT  1−yT
p11 (1 − wE ) p01 (1 − wE )
Pr(yE = 1 | yT , δE = 1) = .
p10 + p11 (1 − wE ) p00 + p01 (1 − wE )

• When yE is observed but yT is missing, we draw missing yT from a Bernoulli

distribution with the toxicity probability given by
 yE  1−yE
p11 (1 − wT ) p10 (1 − wT )
Pr(yT = 1 | yE , δT = 1) = .
p01 + p11 (1 − wT ) p00 + p10 (1 − wT )

Derivation for the above conditional probabilities and more details of the BDA
procedure are provided in Zhou et al. (2022). Following Cheung and Chappell
(2000) and Yuan et al. (2018), the TITE-BOIN12 design by default assumes
that the time to toxicity and time to efficacy are uniformly distributed over the
respective assessment windows. As a result, wT = tU /τT , and wE = tU /τE .
This assumption seems strong, but it is very robust for the purpose of dose
finding (Cheung and Chappell, 2000; Yuan et al., 2018; Zhou et al., 2022).
186 Model-Assisted Bayesian Designs for Dose Finding and Optimization

Mean imputation based on the approximated likelihood

One limitation of BDA is that it is computationally intensive. A second ap-
proach is to use mean imputation to deal with the missing data through the use
of the approximated likelihood approach by Lin and Yuan (2020). The key ob-
servation is that, in the presence of pending outcomes, the reason that BOIN12
cannot be directly applied is that the quasi-number of events x given by equa-
tion (8.10) cannot be calculated. Zhou et al. (2022) proposed to approximate x,
and thus the quasi-binomial likelihood, by replacing the pending/missing val-
ues of yT and yE with their expectations E(yT | δT = 1) = Pr(yT = 1 | δT = 1)
and E(yE | δE = 1) = Pr(yE = 1 | δE = 1), respectively.
Depending on the values of δT and δE , patients can be divided into four
types, i.e., (δT , δE ) = (0, 0), (1, 0), (0, 1), (1, 1). The quasi-number of events x
can be approximated as follows,
1 1
( n
1 XX X
x = υab 1 {yT i = a} 1 {yEi = b} (1 − δT i )(1 − δEi )
100 a=0 i=1
b=0
n
X
+υab 1 {yT i = a} Pr(yEi = b | δEi = 1)(1 − δT i )δEi
i=1
Xn
+υab Pr(yT i = a | δT i = 1)1 {yEi = b} δT i (1 − δEi )
i=1
n
)
X
+υab Pr(yT i = a | δT i = 1) Pr(yEi = b | δEi = 1)δT i δEi .
i=1

In this equation, the first term corresponds to those patients whose yT and
yE are both observed (i.e., (δT , δE ) = (0, 0)), thus their contribution to the
quasi-number of events x can be directly computed based on the observed
data. The last three terms correspond to the patients with at least one of yT
and yE pending, i.e., δT + δE > 0, and involve the mean imputation of yT
and yE respectively by Pr(yT = a | δT = 1) and Pr(yE = 1 | δE = b), a, b =
0, 1, based on the assumption that yT and yE are “working” independent.
Technically speaking, the correlation between efficacy and toxicity can be
modeled. The small sample sizes of phase I/II trials, however, provide limited
information to estimate the correlation parameter reliably. The independence
assumption for pending outcomes seems strong, but it makes the method
simple and computationally fast. Zhou et al. (2022) showed by numerical
studies that the TITE-BOIN12 design is remarkably robust to the violation
of this assumption. This may be because the assumption is only used for the
patients who have pending data. For patients with both yT and yE observed,
the independence assumption is not made. When the trial progresses, the
percentage of observed data increases, limiting the impact of the violation of
the independence assumption.
By assuming that given yq = 1, the time-to-event outcome tq is a uniform
random variable over (0, τq ), we have Pr(δq = 1 | yq = 1) = Pr(tq > tU | yq =
Finding Optimal Biological Dose 187

1) = 1 − tU /τq , and thus

Pr(yq = 1 | δq = 1)
Pr(δq = 1 | yq = 1) Pr(yq = 1)
=
Pr(δq = 0 | yq = 0) Pr(yq = 0) + Pr(δq = 1 | yq = 1) Pr(yq = 1)
Pr(δq = 1 | yq = 1) Pr(yq = 1)
=
Pr(yq = 0) + Pr(δq = 1 | yq = 1) Pr(yq = 1)
πq (1 − tU /tq )
= ,
1 − πq tU /tq

for q = T, E. As a result, Pr(yq = 0 | δq = 1) = 1 − Pr(yq = 1 | δq = 1) =

(1 − πq )/(1 − πq tU /tq ).
To further calculate the unknown πq involved in the above equation, the
approximated likelihood approach by Lin and Yuan (2020) is adopted. As
derived in Section 5.5, the marginal likelihood function for a patient with
y
outcome q observed is πq q (1 − πq )1−yq , and that for a patient with outcome
q pending is 1 − πq wq , where wq = Pr(tq ≤ tU | yq = 1) = tU /τq . Based on
the approximation 1 − πq wq ≈ (1 − πq )wq , the (marginal) overall likelihood
function for πq based on the data D(d) observed at a specific dose d is given
by
n
Y 1−δqi δqi
πqyqi (1 − πq )1−yqi

L(D(d) | πq ) = (1 − πq wqi )
i=1
Yn
δqi wqi
πqyqi (1 − πq )1−yqi (1 − πq )

≈
i=1
= πqñq (1 − πq )m̃q ,
Pn
where ñq = i=1 (1−δqi )yqi is the number
Pn of outcomes for endpoint
Pn q observed
so far by the interim time, and m̃q = i=1 (1 − δqi )(1 − yqi ) + i=1 δqi wqi is
the “effective” number of patients who do have endpoint q, q = T, E.
Various appropriate estimates of the unknown πq can be easily obtained
based on the above approximated likelihood function. For example, the most
straightforward one is the maximum likelihood estimator π̂q = ñq /(ñq +
m̃q ), q = T, E. By plugging π̂q into the expressions of Pr(yq = 1 | δq = 1)
and Pr(yq = 0 | δq = 1), the quasi-number of events x can be calculated
accordingly. Then, the method of BOIN12 can be directly applied to obtain
the posterior of ũ as in equation (8.11) for decision making.
The dose-finding algorithm for TITE-BOIN12, using either BDA or ap-
proximated likelihood methods, is the same as that for BOIN12. At each
interim decision, TITE-BOIN12 needs to update the admissible set A(D), the
estimate for the marginal toxicity probability π̂T at the current dose, and the
desirability PPj for each dose considered in the trial. In addition, to avoid
risky decisions caused by sparse data, TITE-BOIN12 imposes the following
accrual suspension rule: if more than 50% of the patients have pending DLT
188 Model-Assisted Bayesian Designs for Dose Finding and Optimization

or efficacy outcomes at the current dose, suspend the accrual to wait for more
data to become available.
Of note, another appealing feature of TITE-BOIN12 is that it naturally
accommodates the case that some patients may be evaluable for toxicity, but
not evaluable for efficacy (e.g., patients are off treatment due to toxicity).
These patients can be regarded as yT are observed and yE are (permanently)
pending, and thus directly incorporated into the utility estimation and deci-
sion making.

8.6 Other model-assisted phase I/II designs

8.6.1 uTPI design
Similar to the BOIN12 design, the uTPI design (Shi et al., 2021) also adopts
the Dirichlet-multinomial and quasi-binomial approaches to model the joint
efficacy–toxicity and quasi-binomial utility outcomes, respectively. Unlike
BOIN12, whose decisions are made based on the point estimate of the toxicity
rate and the posterior probability distribution of the utility, the uTPI design
is a seamless extension of the keyboard design (a.k.a., the m-TPI2 design),
and it uses the toxicity and utility intervals in determining the optimal dose
for the incoming patients.
Specifically, following the idea of the keyboard design, the uTPI design
starts by partitioning the support of the probability of toxicity (0, 1) into a
series of intervals of equal width , which results in KT toxicity intervals,
IT,k , k = 1, . . . , KT . Similarly, the utility support (0, 100) also can be par-
titioned into KU intervals, IU,k , k = 1, . . . , KU , based on an equal width δ.
The parameter  (or δ) denotes the indifference margin for toxicity (or util-
ity), which means that any two dose levels whose toxicity probabilities (or
mean utilities) lying within the same toxicity (or utility) interval are treated
indifferently in terms of toxicity or mean utility. The reason for specifying the
indifference intervals is that the sample size in early-phase dose-finding trials
is typically small, and thus it is difficult to differentiate two dose levels that
have similar toxicity probabilities or mean utilities.
For illustration, we take  = 0.1 and δ = 10, leading to 10 toxic-
ity intervals IT,1 = (0, 0.1), . . . , IT,10 = (0.9, 1) and 10 utility intervals
IU,1 = (0, 10), . . . , IT,10 = (90, 100). By combining the toxicity and utility
intervals, the equal-width “keys” in a one-dimensional keyboard design are
extended to equal-area “squares” on a two-dimensional chessboard, as shown
in Figure 8.6.
Mimicking the keyboard design, the strongest toxicity interval that pos-
sesses the largest posterior probability can be identified based on data D(d)
Finding Optimal Biological Dose 189

C
Desirability increases

Toxicity probability increases

FIGURE 8.6: Dose-finding rule of the uTPI design based on a chessboard.

The shaded interval corresponds to the target toxicity probability interval k ∗ .
(1) If C is the current dose (or if A is the current dose and n(A) < N ∗ = 9),
and suppose that B is the next lower dose and D is the next higher dose,
we will allocate the next cohort of patients to the dose among {C, B, D} (or
{A, B, D}) that has the highest utility interval. In this case, we will select
dose D for the next cohort of patients. (2) If A is the current dose and n(A) ≥
N ∗ = 9, and suppose that B is the next lower dose and D is the next higher
dose, we will allocate the next cohort of patients to the dose among {A, B}
that has the highest utility interval. In this example, dose A will be selected
for treating the next cohort of patients.

observed at dose d,

kdT = arg max {Pr (πT (d) ∈ IT,k | D(d))} ,

k=1,...,KT

where the posterior probability distribution of πT (d) is derived under the

beta-binomial model (8.6). Similarly, the strongest utility interval that has
the largest posterior probability can be obtained by the quasi-beta-binomial
model (8.11),

kdU = arg max {Pr (πU (d) ∈ IU,k | D(d))} .

k=1,...,KU

The collection of strongest toxicity and desirability intervals (kdT , kdU ) can
be treated as the vector of sufficient statistics of the uTPI design in determin-
ing dose escalation or de-escalation. Specifically, let k ∗ denote the location of
190 Model-Assisted Bayesian Designs for Dose Finding and Optimization

the toxicity interval that contains the upper limit of the toxicity rate φT . In
the aforementioned example, k ∗ = 4 when φT = 0.35. As demonstrated in
Figure 8.6, the dose-finding rule of uTPI proceeds as follows:
1. Patients in the first cohort are treated at the lowest dose d1 , or the
physician-specified dose.
2. Based on the observed data, we obtain the strongest toxicity and
desirability intervals for each dose. Suppose j is the current dose
level. To assign a dose to the next cohort of patients,
(a) If kdTj > k ∗ , de-escalate the dose to level j − 1.
(b) If kdTj = k ∗ and nj ≥ N ∗ , choose the level from {j − 1, j} that
has a larger utility interval kdU .
(c) Otherwise, if kdTj < k ∗ or kdTj = k ∗ and nj < N ∗ , choose the
level from {j − 1, j, j + 1} that has the largest utility interval
kdU .
3. Repeat Step 2 until the maximum sample size N is reached, and
then use the OBD selection rule described in Section 8.4.3 to select
the OBD.
During the trial conduct, uTPI only treats the patients using doses in
the admissible set A(D) as defined in BOIN12 and U-BOIN. If no dose is
admissible, the trial should be terminated early.
Similar to BOIN12 and U-BOIN designs, the uTPI design also has a concise
decision structure such that a dose-assignment decision table can be calculated
before the trial starts and can be used throughout the trial, which simplifies
its practical implementation. More details about the implementation of the
uTPI design can be found in Shi et al. (2021).

8.6.2 STEIN and BOIN-ET designs

In some trials, OBD is defined as the lowest safe dose that has the maximum
efficacy, instead of the safe dose that has the highest utility. In general, the
designs described previously can be readily adapted to these trials simply by
properly specifying the utility functions, e.g., setting υ11 = 100 and υ00 = 0 in
the BOIN12, U-BOIN, and uTPI designs. Without using utility, some model-
assisted designs have been proposed to directly find the OBD that has the
maximum efficacy.
Lin and Yin (2017b) proposed the STEIN design on the basis of the stan-
dard BOIN design. Besides using BOIN’s toxicity-based rule to determine
whether a dose is safe or not, the STEIN design additionally uses the hy-
potheses stated next to determine whether the efficacy at a dose is acceptable
or not. Specifically, at each dose dj , two point hypotheses are defined:
E E
H0j : πE (dj ) = ψ1 versus H1j : πE (dj ) = ψ2 ,
 Finding Optimal Biological Dose 191

where ψ1 denotes a clinically uninteresting response rate, and ψ2 denotes a

clinically desired response rate. Let π̂E (dj ) = nE (dj )/n(dj ) be the observed
E
efficacy rate at dose dj . The null hypothesis H0j is rejected and dose dj is
considered to be efficacious if π̂E (dj ) ≥ δ, where δ is the boundary for the
observed efficacy rate. For ease of exposition, we slightly abuse the notation
and use expanded notations such as π̂E (dj ), nE (dj ) and n(dj ) to represent
the same things as π̂Ej , nEj and nj used previously.
Based on a similar procedure of BOIN, the optimal boundary of the ob-
served efficacy rate can be obtained,
 
1 − ψ1
log
1 − ψ2
δ=  ,
ψ2 (1 − ψ1 )
log
ψ1 (1 − ψ2 )
E E
where non-informative prior is assumed with Pr(H0j ) = Pr(H1j ) = 1/2.
By partitioning the two-dimensional efficacy–toxicity space into several re-
gions (Figure 8.7), the dose-finding rule of STEIN can be described as follows:
1. Patients in the first cohort are treated at the lowest dose d1 , or the
physician-specified dose.
2. Suppose j is the current dose level. To assign a dose to the next
cohort of patients:
(a) If π̂T (dj ) > λd , de-escalate the dose to level j − 1.
(b) If π̂T (dj ) ≤ λd and π̂E (dj ) ≥ δ, stay at the current dose.
(c) Otherwise, choose the dose that has the largest posterior prob-
ability of Pr (πE (d) > δ | D(d)), which can be obtained based
on the beta-binomial model (8.6).
3. Repeat Step 2 until the maximum sample size N is reached.
To use the STEIN design in practice, Lin and Yin (2017b) recommended
setting φ1 = 0.75φ and φ2 = 1.25φ to obtain BOIN’s escalation/de-escalation
boundaries, and they found that ψ1 ∈ (0.2, 0.4) and ψ2 ∈ (0.6, 0.8) are suitable
for most practical scenarios. For example, when φ = 0.3, ψ1 = 0.3, and ψ2 =
0.8, the values of (λe , λd , δ) are (0.26, 0.34, 0.56). When appropriate, these
design parameters can be further calibrated by simulation.
Similar to the STEIN design, the BOIN-ET design is another extension of
the BOIN design to identify OBD based on both efficacy and toxicity outcomes
(Takeda et al., 2018). Suppose the current dose level is j, according to Figure
8.8, the next dose assignment of BOIN-ET is as follows:

(1) If π̂T (dj ) ≤ λe and π̂E (dj ) ≤ δ, then the current dose is safe but not effective,
and the next dose should be escalated to level j + 1.
(2) If π̂T (dj ) > λd , then the current dose is too toxic and the next dose should
be de-escalated to level j − 1.
192 Model-Assisted Bayesian Designs for Dose Finding and Optimization
1

Promising
region

Efficacy probability
𝜑
Inadmissible
region

Exploratory
region

0
0 𝜆! 𝜙 𝜆 " 1
Toxicity probability

FIGURE 8.7: Partitioned regions under the STEIN design. If the pair of the
observed toxicity and efficacy probabilities at the current dose level lies inside
the promising region, the current dose is retained. If the pair lies inside the
exploratory region, then more dose levels should be explored. Specifically, the
lighter color subregion indicates that the current dose is safe and thus dose
escalation is warranted. The dark color subregion indicates that the toxicity
probability is close to the target, and thus the decision of dose escalation/de-
escalation depends on the observed toxicity and efficacy data jointly. If the
pair lies inside the inadmissible region, the current dose is too toxic and thus
dose de-escalation is needed.

Stay Stay De-escalate

Efficacy probability

Escalate/
Escalate Stay/ De-escalate
De-escalate

0 𝜆! 𝜆" 1

Toxicity probability

FIGURE 8.8: Dose allocation rules for the BOIN-ET design considering both
efficacy and toxicity.
 Finding Optimal Biological Dose 193

(3) If π̂T (dj ) < λd and π̂E (dj ) > δ, then the current dose is desirable in terms
of both efficacy and safety, then the next dose should not be changed.
(4) If λe < π̂T (dj ) ≤ λd and π̂E (dj ) ≤ δ, then all possible decisions including
escalation, stay, and de-escalation are considered. Typically, set the admis-
sible set Aj = {j − 1, j, j + 1}, the next dose is determined according to the
following rules:
(a) If dose level j + 1 is untried, then the next dose is escalated to level
j + 1.
(b) Otherwise, the next dose is selected as the one with the maximum ob-
served efficacy rate. If there is a tie having the maximum observed effi-
cacy rate, then the next dose is randomly chosen from the tied doses.
To determine the optimal values of (λe , λd , δ), Takeda et al. (2018) con-
sidered the following six hypotheses at dose dj :

H1j : πT (dj ) = φ1 , πE (dj ) = ψ1 , H2j : πT (dj ) = φ1 , πE (dj ) = ψ2 ,

H3j : πT (dj ) = φ, πE (dj ) = ψ1 , H4j : πT (dj ) = φ, πE (dj ) = ψ2 ,
H5j : πT (dj ) = φ2 , πE (dj ) = ψ1 , H6j : πT (dj ) = φ2 , πE (dj ) = ψ2 ,
where the design parameters (φ, φ1 , φ2 , ψ1 , ψ2 ) have similar interpretations
as in the STEIN design. By treating efficacy and toxicity jointly, the BOIN-
ET design finds the optimal values that minimize the probability of incorrect
decisions using a numerical grid search. More details can be found in Takeda
et al. (2018).
To implement the BOIN-ET design, Takeda et al. (2018) recommended set-
ting φ1 = 0.1φ, φ2 = 1.4φ, ψ1 = 0.6ψ2 , and assuming π1j = · · · = π6j = 1/6.
For illustration, consider the target toxicity and efficacy rates are (φ, ψ2 ) =
(0.3, 0.6), the optimal values of (λe , λd , δ) are calculated to be (0.14, 0.35, 0.48).
To accommodate potentially late-onset toxicity or efficacy, Takeda et al. (2020)
extended BOIN-ET to TITE-BOIN-ET along a similar line as TITE-BOIN12
using the approximated likelihood approach.

8.7 Software and case study

Software
The BOIN12, TITE-BOIN12, and U-BOIN designs can be implemented
using web applications provided at http://www.trialdesign.org. The
software allows users to generate the decision table, evaluate operat-
ing characteristics of the design, and generate the trial design template
for protocol preparation. The software for the EffTox design is freely
available at the MD Anderson Cancer Center Software Download Web-
site https://biostatistics.mdanderson.org/softwaredownload/. The R
194 Model-Assisted Bayesian Designs for Dose Finding and Optimization

code for simulating the uTPI design is available at https://github.com/

ruitaolin/uTPI.

Case study
Platinum-Refractory Oral Cancer Trial The objective of this phase I–II
trial (Clinical Trials Registry-India identifier: CTRI/2019/01/016837) is to
determine OBD for methotrexate when given along with erlotinib and cele-
coxib in treating patients with platinum-resistant or early-failure squamous
cell carcinoma of the oral cavity. DLT is scored based on the Common Termi-
nology Criteria for Adverse Events (version 4.03), and the efficacy endpoint
used in monitoring is the clinical benefit rate at two months. Five doses of
methotrexate (i.e., 3, 6, 9, 12, and 15 mg/m2 ) are investigated in combination
with 150 mg erlotinib and 200 mg celecoxib.
Additional examples include a CD33 CAR T-cell therapy trial in pa-
tients with relapsed or refractory acute myeloid leukemia (ClinicalTrials.gov
Identifier: NCT04835519, based on BOIN12), and a trial that evaluates a
plant extract to treat breast cancer patients (ClinicalTrials.gov Identifier:
NCT05007444, based on U-BOIN).
We use the oral cancer trial as an example to illustrate the use of the
BOIN12 app to design phase I-II trials. The BOIN12 app can be selected and
launched from the BOIN Suite launchpad (Figure 8.9). The app includes six

FIGURE 8.9: The launchpad of web application “BOIN Suite” available at

http://www.trialdesign.org.
Finding Optimal Biological Dose 195

FIGURE 8.10: Specify doses and the sample size.

functional tabs and has the capability to implement the three main tasks: trial
design, trial conduct, and OBD determination. We design the trial using the
following three steps:

Step 1: Enter trial parameters

Doses and Sample Size As shown in Figure 8.10, the number of doses un-
der investigation for methotrexate is five, and the starting dose level is 1. The
total sample size is 36 and patients are treated in the cohort size of three,
which results in a total of 12 cohorts. We recommend the maximum sample
size N should be no less than 6 × J (i.e., the maximum sample size of the
3+3 design), where J is the number of doses. To reduce the sample size, the
“convergence” stopping rule is implemented in BOIN12, that is, when m pa-
tients have been treated by the current dose and the decision is to stay, the
trial should be early stopped before the exhaustion of the maximum sample
size. This stopping criterion indicates that the trial can be stopped when the
dose finding approximately converges to OBD. In this trial, m = 12 is used.
Because of this early stopping rule, the actual sample size used in the trial is
often smaller than N . The saving in sample size depends on the underlying
dose–toxicity or dose–efficacy scenario and can be evaluated through simula-
tion studies.
196 Model-Assisted Bayesian Designs for Dose Finding and Optimization

(a) Approach based on the utility score.

(b) Approach based on toxicity and efficacy probabilities.

FIGURE 8.11: Two approaches to specify the risk–benefit trade-off based on

(a) utility scores for possible outcomes or (b) marginal toxicity and efficacy
probabilities.

Risk-benefit Trade-off Criteria In Figure 8.11 (a), the utility is used to

define the risk-benefit trade-off (or desirability), where the outcome (toxic-
ity, efficacy) has a slightly higher utility score than (no toxicity, no efficacy).
Utility should be elicited from physicians to reflect the risk-benefit trade-off
underlying their clinical decision making. After specifying the utility, simu-
lation should be performed to evaluate the operating characteristics of the
design. In some cases, the simulation results may motivate slight modification
of some of the numerical utility values, although such modification typically
has little or no effect on the design’s operating characteristics. Given the
Finding Optimal Biological Dose 197

utility table, the desirability of a dose is characterized by the mean utility

(8.9), and a higher value of the mean utility means the dose is more desirable.
The software provides another option to specify the risk-benefit trade-
off based on the marginal toxicity and efficacy probabilities, see Figure 8.11
(b). As described in Section 8.3.1, this approach actually is a special case of
the utility approach with υ00 + υ11 = 100. For example, the utility specified
in Figure 8.11 (a) is equivalent to specifying the toxicity-efficacy trade-off as
Pr(efficacy)-2/3Pr(toxicity). In other words, the risk-benefit trade-off specified
in Figure 8.11 (a) and (b) are equivalent.
In some trials, the objective is to find the dose that is safe and has the
highest efficacy rate. This can be implemented by (1) setting υ11 = 100 and
υ00 = 0 in the utility table, or (2) set w = 0 in the toxicity–efficacy probability
trade-off function. In both cases, the safety condition is imposed through the
Admissible Criteria, as discussed next.
Admissible Criteria To safeguard patients from toxic and/or futile doses,
two dose acceptability criteria discussed in Section 8.4.2 are used by BOIN12
to determine which doses may be used to treat patients. According to Figure
8.12, the highest acceptable toxicity probability is φT = 0.35, and the lowest
acceptable efficacy probability is φE = 0.25. Generally, φE can take the value
of the target response rate specified for a standard phase II trial. Because U-
BOIN considers the toxicity–efficacy trade-off, the value of φT should be set
slightly higher (e.g., 0.05) than the target toxicity rate used in conventional
toxicity-based phase I designs. For example, if 30% is an appropriate target
toxicity rate that the conventional phase I design used, then φT = 0.35 is a
reasonable choice for U-BOIN.
For the toxicity and efficacy probability cutoffs, we recommend cT = 0.95
and cE = 0.90, which seem high, but actually are appropriate as their purpose

FIGURE 8.12: Admissible criteria of BOIN12 design.

198 Model-Assisted Bayesian Designs for Dose Finding and Optimization

is to rule out excessively toxic and ineffective doses. Among admissible doses,
the dose assignment rule will allocate patients to the most desirable dose. In
other words, even if the admissible dose set includes some doses that are not
particularly safe or efficacious, the design will not assign patients to these
suboptimal doses. Due to the large uncertainty of small sample sizes, using
small values for cT and cE will inadvertently eliminate the doses that are
actually admissible, and thus affect the operating characteristics of the design.
If a dose is inadmissible due to violation of the safety criterion, this dose and
its higher doses are considered inadmissible. During the trial, only admissible
doses can be used to treat patients, and the doses that are not admissible
should be eliminated from the trial.
After the completion of the specification of design parameters, the de-
sign flowchart (Figure 8.13) and decision tables (Figure 8.14) can be gen-
erated by clicking the “Get Decision Table” button (see Figure 8.12). In
BOIN12, two decision tables will be generated: the first one summarizes the

FIGURE 8.13: The flowchart of the BOIN12 design generated by the BOIN12
shiny app.
Finding Optimal Biological Dose 199

FIGURE 8.14: The escalation/de-escalation boundary table (upper panel) and

rank-based desirability score table (lower panel) of the BOIN12 design gener-
ated by the BOIN12 shiny app.

escalation/de-escalation boundaries for toxicity monitoring, and the second

one is the RDS table that defines the desirability of each dose. The decision
tables will be automatically included in the protocol template in Step 3, but
can also be saved as separate csv, Excel, or pdf files in this step when needed.

Step 2: Run simulation

Operating Characteristics This step generates the operating characteristics
of the design through simulation, see Figure 8.15. Scenarios used for simulation
should cover various possible clinical scenarios, e.g., OBD located at different
dose levels. The software uses the latent bivariate normal random variables to
generate the joint toxicity and efficacy outcomes. Users need to provide the
true marginal toxicity rate and efficacy rate, as well as the correlation between
toxicity and efficacy to simulate data. Here, the correlation corresponds to the
correlation parameter in the covariance matrix of the latent bivariate normal
200 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 8.15: Simulate the operating characteristics the BOIN12 design.

distribution. The simulation results (shown in Table 8.8) will be automatically

included as a table in the protocol template in the next step, but can also be
saved as a separate csv or Excel file if needed.
Step 3: Generate protocol template
Protocol Preparation The BOIN12 software generates sample texts and a
protocol template to facilitate the protocol write-up. The protocol template
can be downloaded in various formats (see Figure 8.16). Use of this mod-
ule requires the completion of Steps 1 and 2. Once the protocol is approved
by regulatory bodies (e.g., Institutional Review Board), we follow the design
Finding Optimal Biological Dose 201

TABLE 8.8: Simulation results generated by the BOIN12 shiny app. The
values corresponding to OBD are in boldface.

Dose level
Avg. N Stop %
1 2 3 4 5
Scenario 1
Pr(DLT) 0.03 0.17 0.35 0.50 0.65
Pr(Efficacy) 0.12 0.39 0.43 0.55 0.65
Mean utility 46.0 56.6 51.8 53.0 53.0
No. patients treated 5.2 8.8 6.4 2.8 0.6 23.8
Selection % 16.5 53.5 24.0 5.5 0.0 0.5
Scenario 2
Pr(DLT) 0.01 0.08 0.10 0.12 0.35
Pr(Efficacy) 0.05 0.25 0.30 0.60 0.60
Mean utility 42.6 51.8 54.0 71.2 62.0
No. patients treated 3.4 4.8 4.7 9.0 4.2 26.1
Selection % 1.5 14.5 6.5 63.5 14.0 0.0

decision table included in the protocol to conduct the trial and make adaptive
decisions (e.g., dose escalation/stay/de-escalation). Alternatively, users can
use the Trial Conduct tab to determine the dose for next cohort of patients.
In the latter approach, users can upload trial data to the app to obtain the
recommended dose for the next cohort of patients. Summary statistics for the
interim data are also provided by the app.
After the trial completes accrual and has all patients’ outcomes evaluated,
users can use the OBD determination tab to identify OBD. After uploading
the trial data with the provided csv template, users can obtain OBD on the

FIGURE 8.16: Download protocol templates of the BOIN12 design.

202 Model-Assisted Bayesian Designs for Dose Finding and Optimization

FIGURE 8.17: Dose and sample size of the U-BOIN design.

right side of the app, as well as various estimates, including the estimates for
joint toxicity-efficacy probabilities, marginal toxicity/efficacy probability, and
utility of each dose that has been used to treat patients.

Software for U-BOIN and TITE-BOIN12

The software for U-BOIN and TITE-BOIN12 has the same layout as BOIN12.
The specification of risk-benefit trade-off and admissible criteria follow the
same principle as described above. For U-BOIN, a few extra design param-
eters need to be specified, which are mostly related to the first stage of the
design that focuses on the dose exploration based on toxicity using the stan-
dard BOIN design. As shown in Figure 8.17, the design parameters are similar
to those of BOIN, e.g., apply the 3+3 design run-in and accelerated titration.
Section 3.6 provides detailed discussion on how to choose these design param-
eters. For the Stage I “convergence” stopping rule, we recommend m1 ≥ 9. In
U-BOIN, patients treated in Stage I roll over to Stage II, thus the Stage II
“convergence” early stopping cutoff m2 should be greater than m1 . Roughly
speaking, m2 − m1 represents the maximum number of additional patients we
may expand on each of admissible doses identified at the end of Stage I to
optimize the dose. For example, in Figure 8.17, m2 = 24 means treating up to
12 additional patients at OBD and/or admissible doses. The exact meaning of
Finding Optimal Biological Dose 203

m2 − m1 depends on which strategy we use to assign patients in Stage II (e.g.,

pick the winner, adaptive randomization, or equal randomization). Simulation
should be performed to validate and calibrate the values of m1 and m2 to ob-
tain desirable operating characteristics. As described previously, identification
of OBD (or optimizing the dose in general) is substantially more challenging
than identification of MTD. Using an excessively small m1 and m2 leads to
very low power to identify OBD.
Bibliography

Ahn, C. (1998). An evaluation of phase I cancer clinical trial designs. Statistics

in Medicine, 17(14):1537–1549.

Babb, J., Rogatko, A., and Zacks, S. (1998). Cancer phase I clinical tri-
als: efficient dose escalation with overdose control. Statistics in Medicine,
17(10):1103–1120.
Barlow, R. E., Bartholomew, D. J., Bremner, J. M., and Brunk, H. D. (1972).
Statistical Inference under Order Restrictions; The Theory and Application
of Isotonic Regression. Wiley, New York, NY.
Bekele, B. N., Ji, Y., Shen, Y., and Thall, P. F. (2007). Monitoring late-onset
toxicities in phase I trials using predicted risks. Biostatistics, 9(3):442–457.
Bekele, B. N. and Shen, Y. (2005). A Bayesian approach to jointly mod-
eling toxicity and biomarker expression in a phase I/II dose-finding trial.
Biometrics, 61(2):343–354.
Bekele, B. N. and Thall, P. F. (2004). Dose-finding based on multiple tox-
icities in a soft tissue sarcoma trial. Journal of the American Statistical
Association, 99(465):26–35.
Berger, J. O. (2013). Statistical Decision Theory and Bayesian Analysis.
Springer Science & Business Media.
Berry, D. A. (2003). Statistical innovations in cancer research. Cancer
Medicine, 6:465–478.

Berry, S. M., Carlin, B. P., Lee, J. J., and Muller, P. (2010). Bayesian Adaptive
Methods for Clinical Trials. CRC press.
Biswas, S., Liu, D. D., Lee, J. J., and Berry, D. A. (2009). Bayesian clinical
trials at the University of Texas M. D. Anderson Cancer Center. Clinical
Trials, 6(3):205–216.
Brahmer, J. R., Drake, C. G., Wollner, I., Powderly, J. D., Picus, J., Sharfman,
W. H., Stankevich, E., Pons, A., Salay, T. M., McMiller, T. L., et al. (2010).
Phase I study of single-agent anti–programmed death-1 (MDX-1106) in re-
fractory solid tumors: safety, clinical activity, pharmacodynamics, and im-
munologic correlates. Journal of Clinical Oncology, 28:3167–3175.

205
206 Bibliography

Braun, T. M. and Jia, N. (2013). A generalized continual reassessment method

for two-agent phase I trials. Statistics in Biopharmaceutical Research,
5(2):105–115.
Braun, T. M. and Wang, S. (2010). A hierarchical Bayesian design for phase
I trials of novel combinations of cancer therapeutic agents. Biometrics,
66(3):805–812.
Bril, G., Dykstra, R., Pillers, C., and Robertson, T. (1984a). Algorithm AS
206: isotonic regression in two independent variables. Journal of the Royal
Statistical Society. Series C (Applied Statistics), 33(3):352–357.
Bril, G., Dykstra, R., Pillers, C., and Robertson, T. (1984b). Isotonic regres-
sion in two independent variables. Journal of the Royal Statistical Society:
Series C (Applied Statistics), 33:352–358.
Bugano, D. D., Hess, K., Jardim, D. L., Zer, A., Meric-Bernstam, F., Siu,
L. L., Razak, A. R., and Hong, D. S. (2017). Use of expansion cohorts in
phase I trials and probability of success in phase II for 381 anticancer drugs.
Clinical Cancer Research, 23(15):4020–4026.
Cai, C., Yuan, Y., and Ji, Y. (2014). 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.
Carlin, B. P. and Louis, T. A. (2008). Bayesian Methods for Data Analysis.
CRC Press.
Chen, Z., Krailo, M., Azen, S., and Tighiouart, M. (2010). A novel toxicity
scoring system treating toxicity response as a quasi-continuous variable in
phase I clinical trials. Contemporary Clinical Trials, 31:473–482.
Cheung, Y. K. (2005). Coherence principles in dose-finding studies.
Biometrika, 92(4):863–873.
Cheung, Y. K. (2011). Dose Finding by the Continual Reassessment Method.
Chapman and Hall/CRC, Boca Raton, FL.
Cheung, Y. K. and Chappell, R. (2000). Sequential designs for phase I clinical
trials with late-onset toxicities. Biometrics, 56(4):1177–1182.
Chow, S.-C. and Chang, M. (2008). Adaptive design methods in clinical trials–
a review. Orphanet Journal of Rare Diseases, 3(1):1–13.
Clertant, M. and O’Quigley, J. (2017). Semiparametric dose finding methods.
Journal of the Royal Statistical Society Series B, 79(5):1487–1508.
Cook, N., Hansen, A. R., Siu, L. L., and Razak, A. R. A. (2015). Early phase
clinical trials to identify optimal dosing and safety. Molecular Oncology,
9(5):997–1007.
Bibliography 207

Dale, J. R. (1986). Global cross-ratio models for bivariate, discrete, ordered

responses. Biometrics, 42(4):909–917.
Durham, S. D., Flournoy, N., and Rosenberger, W. F. (1997). A random walk
rule for phase I clinical trials. Biometrics, 53:745–760.
Ezzalfani, M., Zohar, S., Qin, R., Mandrekar, S. J., and Deley, M.-C. L. (2013).
Dose-finding designs using a novel quasi-continuous endpoint for multiple
toxicities. Statistics in Medicine, 32:2728–2746.
Faries, D. (1994). Practical modifications of the continual reassessment
method for phase I cancer clinical trials. Journal of Biopharmaceutical
Statistics, 4(2):147–164.
FDA (2006). Nonclinical safety evaluation of drug or biologic combinations.
FDA Guidance for Industry.
FDA (2013). Codevelopment of two or more new investigational drugs for use
in combination. FDA Guidance for Industry.
FDA (2019). Submitting documents using real-world data and real-world evi-
dence to FDA for drugs and biologics guidance for industry. FDA Guidance
for Industry.
FDA (2022). Project Optimus, https://www.fda.gov/about-fda/oncology-
center-excellence/project-optimus
Garrett-Mayer, E. (2006). The continual reassessment method for dose-finding
studies: a tutorial. Clinical Trials, 3(1):57–71.
Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., and Rubin,
D. B. (2013). Bayesian Data Analysis. Chapman and Hall/CRC.
Goodman, S., Zahurak, M., and Piantadosi, S. (1995). Some pratical improve-
ments in the continual reassessment method for phase I studies. Statistics
in Medicine, 14:1149–1161.
Guo, B. and Li, Y. (2015). Bayesian dose-finding designs for combination of
molecularly targeted agents assuming partial stochastic ordering. Statistics
in Medicine, 34(5):859–875.
Guo, B., Li, Y., and Yuan, Y. (2016). A dose–schedule finding design for
phase I–II clinical trials. Journal of the Royal Statistical Society: Series C
(Applied Statistics), 65(2):259–272.
Guo, B. and Yuan, Y. (2017). Bayesian phase I/II biomarker-based dose
finding for precision medicine with molecularly targeted agents. Journal of
the American Statistical Association, 112(518):508–520.
Heyd, J. M. and Carlin, B. P. (1999). Adaptive design improvements in the
continual reassessment method for phase I studies. Statistics in Medicine,
18(11):1307–1321.
208 Bibliography

Houede, N., Thall, P. F., Nguyen, H., Paoletti, X., and Kramar, A. (2010).
Utility-based optimization of combination therapy using ordinal toxicity
and efficacy in phase I/II trials. Biometrics, 66(2):532–540.
Hunsberger, S., Rubinstein, L. V., Dancey, J., and Korn, E. L. (2005). Dose
escalation trial designs based on a molecularly targeted endpoint. Statistics
in Medicine, 24(14):2171–2181.
Iasonos, A., Wages, N. A., Conaway, M. R., Cheung, K., Yuan, Y., and
O’Quigley, J. (2016). Dimension of model parameter space and operat-
ing characteristics in adaptive dose-finding studies. Statistics in Medicine,
35(21):3760–3775.
Iasonos, A., Wilton, A. S., Riedel, E. R., Seshan, V. E., and Spriggs, D. R.
(2008). A comprehensive comparison of the continual reassessment method
to the standard 3+ 3 dose escalation scheme in phase I dose-finding studies.
Clinical Trials, 5(5):465–477.
Ivanova, A., Montazer-Haghighi, A., Mohanty, S. G., and Durham, S. D.
(2003). Improved up-and-down designs for phase I trials. Statistics in
Medicine, 22(1):69–82.
Jaki, T., Clive, S., and Weir, C. J. (2013). Principles of dose finding studies
in cancer: a comparison of trial designs. Cancer Chemotherapy and Phar-
macology, 71(5):1107–1114.
Ji, Y., Liu, P., Li, Y., and Nebiyou Bekele, B. (2010). A modified toxicity
probability interval method for dose-finding trials. Clinical Trials, 7(6):653–
663.
Kass, R. E. and Raftery, A. E. (1995). Bayes factors. Journal of the American
Statistical Association, 90(430):773–795.
Le Tourneau, C., Diéras, V., Tresca, P., Cacheux, W., and Paoletti, X. (2010).
Current challenges for the early clinical development of anticancer drugs in
the era of molecularly targeted agents. Targeted Oncology, 5:65–72.
Le Tourneau, C., Lee, J. J., and Siu, L. L. (2009). Dose escalation methods
in phase I cancer clinical trials. JNCI: Journal of the National Cancer
Institute, 101(10):708–720.
Lee, J., Thall, P. F., Ji, Y., and Müller, P. (2016). A decision-theoretic phase
I–II design for ordinal outcomes in two cycles. Biostatistics, 17(2):304–319.
Lee, J. J. and Chu, C. T. (2012). Bayesian clinical trials in action. Statistics
in Medicine, 31(25):2955–2972.
Lee, S., Hershman, D., Martin, P., Leonard, J., and Cheung, K. (2009). Vali-
dation of toxicity burden score for use in phase I clinical trials. Journal of
Clinical Oncology, 27(15 suppl):2514–2514.
Bibliography 209

Lee, S. M., Cheng, B., and Cheung, Y. K. (2010). Continual reassessment

method with multiple toxicity constraints. Biostatistics, 12(2):386–398.
Lee, S. M. and Cheung, Y. K. (2009). Model calibration in the continual
reassessment method. Clinical Trials, 6(3):227–238.
Lee, S. M., Ursino, M., Cheung, Y. K., and Zohar, S. (2017). Dose-finding
designs for cumulative toxicities using multiple constraints. Biostatistics,
20(1):17–29.
Li, D. H., Whitmore, J. B., Guo, W., and Ji, Y. (2017). Toxicity and efficacy
probability interval design for phase I adoptive cell therapy dose-finding
clinical trials. Clinical Cancer Research, 23(1):13–20.
Li, Y. and Yuan, Y. (2020). PA-CRM: a continuous reassessment method for
pediatric phase I oncology trials with concurrent adult trials. Biometrics,
76(4):1364–1373.
Lin, R. (2018). Bayesian optimal interval design with multiple toxicity con-
straints. Biometrics, 74(4):1320–1330.
Lin, R. and Lee, J. J. (2020). Novel Bayesian adaptive designs and their
applications in cancer clinical trials. In Computational and Methodological
Statistics and Biostatistics, 395–426. Springer.
Lin, R., Thall, P. F., and Yuan, Y. (2020a). An adaptive trial design to opti-
mize dose-schedule regimes with delayed outcomes. Biometrics, 76(1):304–
315.
Lin, R., Thall, P. F., and Yuan, Y. (2021). A phase I–II basket trial design
to optimize dose-schedule regimes based on delayed outcomes. Bayesian
Analysis, 16(1):179–202.
Lin, R. and Yin, G. (2017a). Bayesian optimal interval design for dose find-
ing in drug-combination trials. Statistical Methods in Medical Research,
26(5):2155–2167.
Lin, R. and Yin, G. (2017b). STEIN: a simple toxicity and efficacy inter-
val design for seamless phase I/II clinical trials. Statistics in Medicine,
36(26):4106–4120.
Lin, R. and Yuan, Y. (2019). On the relative efficiency of model-assisted
designs: a conditional approach. Journal of Biopharmaceutical Statistics,
29(4):648–662.
Lin, R. and Yuan, Y. (2020). Time-to-event model-assisted designs for dose-
finding trials with delayed toxicity. Biostatistics, 21(4):807–824.
Lin, R., Zhou, Y., Yan, F., Li, D., and Yuan, Y. (2020b). BOIN12: Bayesian
optimal interval phase I/II trial design for utility-based dose finding in im-
munotherapy and targeted therapies. JCO Precision Oncology, 4:1393–1402.
210 Bibliography

Lin, Y. and Shih, W. J. (2001). Statistical properties of the traditional

algorithm-based designs for phase I cancer clinical trials. Biostatistics,
2(2):203–215.
Little, R. J. and Rubin, D. B. (2014). Statistical Analysis with Missing Data,
volume 333. John Wiley & Sons.
Liu, R., Yuan, Y., Sen, S., Yang, X., Jiang, Q., Li, X., Lu, C., Gonen, M.,
Tian, H., Zhou, H., Lin, R., and Marchenko, O. (2022). Accuracy and safety
of novel designs for phase I drug-combination oncology trials. Statistics in
Biopharmaceutical Research, 14(3):270–282.
Liu, S., Guo, B., and Yuan, Y. (2018). A Bayesian phase I/II trial de-
sign for immunotherapy. Journal of the American Statistical Association,
113(523):1016–1027.
Liu, S. and Johnson, V. E. (2016). A robust Bayesian dose-finding design for
phase I/II clinical trials. Biostatistics, 17(2):249–263.
Liu, S., Pan, H., Xia, J., Huang, Q., and Yuan, Y. (2015). Bridging continual
reassessment method for phase I clinical trials in different ethnic popula-
tions. Statistics in Medicine, 34(10):1681–1694.
Liu, S., Yin, G., and Yuan, Y. (2013). Bayesian data augmentation dose find-
ing with continual reassessment method and delayed toxicity. The Annals
of Applied Statistics, 7(4):1837.
Liu, S. and Yuan, Y. (2015). Bayesian optimal interval designs for phase I
clinical trials. Journal of the Royal Statistical Society: Series C (Applied
Statistics), 64(3):507–523.
Liu, S. and Yuan, Y. (2022). Bayesian optimal interval designs for phase I
clinical trials. Journal of the Royal Statistical Society: Series C (Applied
Statistics), 71(2):491–492.
Mahajan, R. and Gupta, K. (2010). Adaptive design clinical trials: method-
ology, challenges and prospect. Indian Journal of Pharmacology, 42(4):201.
Mander, A. P. and Sweeting, M. J. (2015). A product of independent beta
probabilities dose escalation design for dual-agent phase I trials. Statistics
in Medicine, 34(8):1261–1276.
Mandrekar, S. J., Qin, R., and Sargent, D. J. (2010). Model-based phase I
designs incorporating toxicity and efficacy for single and dual agent drug
combinations: methods and challenges. Statistics in Medicine, 29(10):1077–
1083.
Mathijssen, R. H., Sparreboom, A., and Verweij, J. (2014). Determining the
optimal dose in the development of anticancer agents. Nature Reviews Clin-
ical Oncology, 11(5):272.
Bibliography 211

Morita S, Thall PF, Müller P. (2008) Determining the effective sample size of
a parametric prior. Biometrics, 64(2):595–602.

Morita, S. (2011). Application of the continual reassessment method to a

phase I dose-finding trial in Japanese patients: east meets west. Statistics
in Medicine, 30(17):2090–2097.
Mu, R., Yuan, Y., Xu, J., Mandrekar, S. J., and Yin, J. (2018). gBOIN: a uni-
fied model-assisted phase I trial design accounting for toxicity grades, and
binary or continuous end points. Journal of the Royal Statistical Society:
Series C, 68(2):289–308.
Murray, T. A., Yuan, Y., Thall, P. F., Elizondo, J. H., and Hofstetter, W. L.
(2018). A utility-based design for randomized comparative trials with ordi-
nal outcomes and prognostic subgroups, Biometrics, 74(3):1095–1103.
Neuenschwander, B., Branson, M., and Gsponer, T. (2008). Critical aspects
of the Bayesian approach to phase I cancer trials. Statistics in Medicine,
27(13):2420–2439.
Neuenschwander, B., Matano, A., Tang, Z., Roychoudhury, S., Wandel, S., and
Bailey, S. (2015). Bayesian industry approach to phase I combination trials
in oncology. Statistical Methods in Drug Combination Studies, 6:95–135.
Onar-Thomas, A. and Xiong, Z. (2010). A simulation-based comparison of
the traditional method, Rolling-6 design and a frequentist version of the
continual reassessment method with special attention to trial duration in
pediatric phase I oncology trials. Contemporary Clinical Trials, 31(3):259–
270.
O’Quigley, J. and Chevret, S. (1991). Methods for dose finding studies in
cancer clinical trials: a review and results of a monte carlo study. Statistics
in Medicine, 10(11):1647–1664.

O’Quigley, J., Pepe, M., and Fisher, L. (1990). Continual reassessment

method: a practical design for phase 1 clinical trials in cancer. Biomet-
rics, 46(1):33–48.
Pallmann, P., Bedding, A. W., Choodari-Oskooei, B., Dimairo, M., Flight,
L., Hampson, L. V., Holmes, J., Mander, A. P., Sydes, M. R., Villar, S. S.,
et al. (2018). Adaptive designs in clinical trials: why use them, and how to
run and report them. BMC Medicine, 16(1):1–15.
Pan, H., Lin, R., Zhou, Y., and Yuan, Y. (2020). Keyboard design for phase
I drug-combination trials. Contemporary Clinical Trials, 92:105972.

Pan, H. and Yuan, Y. (2017). A default method to specify skeletons for

Bayesian model averaging continual reassessment method for phase I clinical
trials. Statistics in Medicine, 36(2):266–279.
212 Bibliography

Papke, L. E. and Wooldridge, J. M. (1996). Econometric methods for frac-

tional response variables with an application to 401 (k) plan participation
rates. Journal of Applied Econometrics, 11(6):619–632.
Penel, N., Adenis, A., Clisant, S., and Bonneterre, J. (2011). Nature and
subjectivity of dose-limiting toxicities in contemporary phase I trials: com-
parison of cytotoxic versus non-cytotoxic drugs. Investigational New Drugs,
29:1414–1419.

Petit, C., Samson, A., Morita, S., Ursino, M., Guedj, J., Jullien, V., Comets,
E., and Zohar, S. (2018). Unified approach for extrapolation and bridging of
adult information in early-phase dose-finding paediatric studies. Statistical
Methods in Medical Research, 27(6):1860–1877.
Phan, T. G., Ma, H., Lim, R., Sobey, C. G., and Wallace, E. M. (2018). Phase
1 trial of amnion cell therapy for ischemic stroke. Frontiers in Neurology,
9:198.
Piantadosi, S., Fisher, J. D., and Grossman, S. (1998). Practical implemen-
tation of a modified continual reassessment method for dose-finding trials.
Cancer Chemotherapy and Pharmacology, 41(6):429–436.
Postel-Vinay, S., Gomez-Roca, C., Molife, L. R., Anghan, B., Levy, A., Judson,
I., De Bono, J., Soria, J.-C., Kaye, S., and Paoletti, X. (2011). Phase I
trials of molecularly targeted agents: should we pay more attention to late
toxicities. Journal of Clinical Oncology, 29(13):1728–1735.
Riviere, M.-K., Yuan, Y., Dubois, F., and Zohar, S. (2014). A Bayesian
dose-finding design for drug combination clinical trials based on the logistic
model. Pharmaceutical Statistics, 13(4):247–257.
Riviere, M.-K., Dubois, F., and Zohar, S. (2015a). Competing designs for drug
combination in phase I dose-finding clinical trials. Statistics in Medicine,
34(1):1–12.
Riviere, M.-K., Yuan, Y., Dubois, F., and Zohar, S. (2015b). A Bayesian
dose finding design for clinical trials combining a cytotoxic agent with a
molecularly targeted agent. Journal of the Royal Statistical Society: Series
C (Applied Statistics), 64(1):215–229.

Riviere, M.-K., Yuan, Y., Jourdan, J.-H., Dubois, F., and Zohar, S. (2018).
Phase I/II dose-finding design for molecularly targeted agent: plateau de-
termination using adaptive randomization. Statistical Methods in Medical
Research, 27(2):466–479.
Robert, C. and Casella, G. (2013). Monte Carlo Statistical Methods. Springer
Science & Business Media.
Bibliography 213

Rogatko, A., Schoeneck, D., Jonas, W., Tighiouart, M., Khuri, F. R., and
Porter, A. (2007). Translation of innovative designs into phase I trials.
Journal of Clinical Oncology, 25(31):4982–4986.
Ruppert, A. S. and Shoben, A. B. (2018). Overall success rate of a safe and
efficacious drug: results using six phase 1 designs, each followed by stan-
dard phase 2 and 3 designs. Contemporary Clinical Trials Communications,
12:40–50.
Sachs, J. R., Mayawala, K., Gadamsetty, S., Kang, S. P., and de Alwis, D. P.
(2016). Optimal dosing for targeted therapies in oncology: drug development
cases leading by example. Clinical Cancer Research, 22(6):1318–1324.
Shen, L. Z. and O’Quigley, J. (1996). Consistency of continual reassessment
method under model misspecification. Biometrika, 83(2):395.
Shi, H., Cao, J., Yuan, Y., and Lin, R. (2021). uTPI: a utility-based toxicity
probability interval design for phase I/II dose-finding trials. Statistics in
Medicine, 40(11):2626–2649.
Simon, R. (1989). Optimal two-stage designs for phase II clinical trials. Con-
trolled Clinical Trials, 10(1):1–10.
Simon, R., Freidlin, B., Rubinstein, L., Arbuck, S. G., Collins, J., and Chris-
tian, M. C. (1997). Accelerated titration designs for phase I clinical trials
in oncology. Journal of the National Cancer Institute, 89(15):1138–1147.
Skolnik, J. M., Barrett, J. S., Jayaraman, B., Patel, D., and Adamson, P. C.
(2008). Shortening the timeline of pediatric phase I trials: the rolling six
design. Journal of Clinical Oncology, 26(2):190–195.
Storer, B. E. (1989). Design and analysis of phase I clinical trials. Biometrics,
45(3):925–937.
Stylianou, M. and Follmann, D. A. (2004). The accelerated biased coin up-
and-down design in phase I trials. Journal of Biopharmaceutical Statistics,
14(1):249–260.

Takeda, K., Taguri, M., and Morita, S. (2018). BOIN-ET: Bayesian optimal
interval design for dose finding based on both efficacy and toxicity outcomes.
Pharmaceutical Statistics, 17(4):383–395.
Takeda K., Morita S., and Taguri M. (2020). TITE-BOIN-ET: Time-to-event
Bayesian optimal interval design to accelerate dose-finding based on both
efficacy and toxicity outcomes. Pharmaceutical Statistics, 19(3):335–349.
Thall, P. F. and Cook, J. D. (2004). Dose-finding based on efficacy–toxicity
trade-offs. Biometrics, 60(3):684–693.
214 Bibliography

Thall, P. F., Millikan, R. E., Mueller, P., and Lee, S.-J. (2003). Dose-finding
with two agents in phase I oncology trials. Biometrics, 59(3):487–496.

Tidwell, R. S. S., Peng, S. A., Chen, M., Liu, D. D., Yuan, Y., and Lee, J. J.
(2019). Bayesian clinical trials at the University of Texas MD Anderson
Cancer Center: an update. Clinical Trials, 16(6):645–656.
van Brummelen, E. M., Huitema, A. D., van Werkhoven, E., Beijnen, J. H.,
and Schellens, J. H. (2016). The performance of model-based versus rule-
based phase I clinical trials in oncology. Journal of Pharmacokinetics and
Pharmacodynamics, 43(3):235–242.
Wages, N. A., Conaway, M. R., and O’Quigley, J. (2011). Dose-finding design
for multi-drug combinations. Clinical Trials, 8(4):380–389.
Weber, J. S., Yang, J. C., Atkins, M. B., and Disis, M. L. (2015). Toxici-
ties of immunotherapy for the practitioner. Journal of Clinical Oncology,
33(18):2092.
Yan, F., Mandrekar, S. J., and Yuan, Y. (2017). Keyboard: a novel Bayesian
toxicity probability interval design for phase I clinical trials. Clinical Cancer
Research, 23(15):3994–4003.
Yin, G., Li, Y., and Ji, Y. (2006). Bayesian dose-finding in phase I/II clinical
trials using toxicity and efficacy odds ratios. Biometrics, 62(3):777–787.
Yin, G. and Lin, R. (2015). Comments on ‘competing designs for drug com-
bination in phase I dose-finding clinical trials’ by M-K. Riviere, F. Dubois,
and S. Zohar. Statistics in Medicine, 34(1):13–17.
Yin, G. and Yuan, Y. (2009a). Bayesian dose finding in oncology for drug
combinations by copula regression. Journal of the Royal Statistical Society:
Series C (Applied Statistics), 58(2):211–224.
Yin, G. and Yuan, Y. (2009b). Bayesian model averaging continual reassess-
ment method in phase I clinical trials. Journal of the American Statistical
Association, 104(487):954–968.
Yin, G. and Yuan, Y. (2009c). A latent contingency table approach to dose
finding for combinations of two agents. Biometrics, 65(3):866–875.

Yin, G., Zheng, S., and Xu, J. (2013). Fractional dose-finding methods with
late-onset toxicity in phase I clinical trials. Journal of Biopharmaceutical
Statistics, 23(4):856–870.
Yuan, Y., Hess, K. R., Hilsenbeck, S. G., and Gilbert, M. R. (2016a). Bayesian
optimal interval design: a simple and well-performing design for phase I
oncology trials. Clinical Cancer Research, 22(17):4291–4301.
Bibliography 215

Yuan, Y., Lee, J. J., and Hilsenbeck, S. G. (2019). Model-assisted designs

for early-phase clinical trials: simplicity meets superiority. JCO Precision
Oncology, 3:1–12.
Yuan, Y., Lin, R., Li, D., Nie, L., and Warren, K. E. (2018). Time-to-event
Bayesian optimal interval design to accelerate phase I trials. Clinical Cancer
Research, 24(20):4921–4930.

Yuan, Y., Nguyen, H. Q., and Thall, P. F. (2016b). Bayesian Designs for
Phase I–II Clinical Trials. CRC Press.
Yuan, Y. and Yin, G. (2011a). Bayesian phase I/II adaptively randomized
oncology trials with combined drugs. The Annals of Applied Statistics,
5(2A):924.
Yuan, Y. and Yin, G. (2011b). Robust EM continual reassessment method
in oncology dose finding. Journal of the American Statistical Association,
106(495):818–831.
Yuan, Z., Chappell, R., and Bailey, H. (2007). The continual reassessment
method for multiple toxicity grades: a Bayesian quasi-likelihood approach.
Biometrics, 63(1):173–179.
Zang, Y. and Lee, J. J. (2014). Adaptive clinical trial designs in oncology.
Chinese Clinical Oncology, 3(4):49.
Zang, Y., Lee, J. J., and Yuan, Y. (2014). Adaptive designs for identifying
optimal biological dose for molecularly targeted agents. Clinical Trials,
11(3):319–327.
Zhang, L. and Yuan, Y. (2016). A practical Bayesian design to identify the
maximum tolerated dose contour for drug combination trials. Statistics in
Medicine, 35(27):4924–4936.
Zhang, W., Sargent, D. J., and Mandrekar, S. (2006). An adaptive dose-finding
design incorporating both toxicity and efficacy. Statistics in Medicine,
25(14):2365–2383.
Zhao, L., Lee, J., Mody, R., and Braun, T. M. (2011). The superiority of
the time-to-event continual reassessment method to the rolling six design
in pediatric oncology phase I trials. Clinical Trials, 8(4):361–369.
Zhou, H., Murray, T. A., Pan, H., and Yuan, Y. (2018a). Comparative re-
view of novel model-assisted designs for phase I clinical trials. Statistics in
Medicine, 37(14):2208–2222.
Zhou, H., Yuan, Y., and Nie, L. (2018b). Accuracy, safety, and reliability of
novel phase I trial designs. Clinical Cancer Research, 24(18):4357–4364.
216 Bibliography

Zhou, Y., Lee, J. J., Wang, S., Bailey, S., and Yuan, Y. (2021a). Incorporating
historical information to improve phase I clinical trials. Pharmaceutical
Statistics, 20(6):1017–1034.
Zhou, Y., Lee, J. J., and Yuan, Y. (2019). A utility-based Bayesian optimal
interval (U-BOIN) phase I/II design to identify the optimal biological dose
for targeted and immune therapies. Statistics in Medicine, 38(28):S5299–
S5316.

Zhou, Y., Li, R., Yan, F., Lee, J. J., and Yuan, Y. (2021b). A comparative
study of Bayesian optimal interval (BOIN) design with interval 3 + 3 (i3+
3) design for phase I oncology dose-finding trials. Statistics in Biopharma-
ceutical Research, 13(2):147–155.
Zhou, Y., Lin, R., Lee, J. J., Li, D., Wang, L., Li, R., and Yuan, Y. (2022).
TITE-BOIN12: a Bayesian phase I/II trial design to find the optimal bi-
ological dose with late-onset toxicity and efficacy. Statistics in Medicine,
41(11):1918–1931.
Zohar, S., Katsahian, S., and O’Quigley, J. (2011). An approach to meta-
analysis of dose-finding studies. Statistics in Medicine, 30(17):2109–2116.
Index

3+3 design, 16 BOIN-ET design, 190

See also dose optimization
accelerated titration design, 20 BOIN12 design, 171
admissible rules, 161, 166, 174 case study, 194
algorithm-based designs, 14 decision table, 175
3+3 design, 16 operating characteristics, 180
accelerated titration design, 20 quasi-binomial likelihood, 173
software, 193
Bayes factor, 7 See also dose optimization
Bayes’ theorem, 1
Bayesian adaptive designs, 7, 10 coherence, 48
Bayesian inference, 5 cohort expansion, 19
Bayesian hypothesis testing, 6 combination trials, 69
Bayesian prediction, 6 comparative study, 82
credible interval, 5 model-assisted designs, 74, 77
posterior mean, 5 model-based designs, 71
Bayesian logistic regression method MTD contour, 70
(BLRM), 30 partial order, 69
Bayesian model averaging CRM (BMA- software, 85
CRM), 25 conjugate prior, 4
Bayesian optimal interval (BOIN) de- continual reassessment method (CRM),
sign, 38 21
Bayesian-frequentist dual inter- dose-finding algorithm, 22
pretation, 42 likelihood and posterior, 21
case study, 58 power model, 21
comparative study, 52 sensitivity, 24
derivation, 42 skeleton, 21, 23
design parameters, 46
frequently asked questions, 49 data-augmentation CRM (DA-CRM),
software, 57 97
statistical properties, 47 Dirichlet-multinomial model, 165
beta-binomial model, 3 dose exploration, 14
BOIN combination design, 74 dose optimization, 156
case study, 85 admissible rules, 161
decision table, 75 comparative study, 170, 180
operating characteristics, 82 efficacy-toxicity trade-off, 159,
software, 85 163

217
218 Index

model-assisted designs, 162, 171, model-based designs, 96

183, 188, 190 nonignorable missing data, 95
model-based designs, 158 See also TITE-BOIN12 design
dose selection, 14
dose-limiting toxicity (DLT), 13 maximum tolerated dose (MTD), 13
drug combination, 69 model-assisted designs, 15, 33
See also combination trials model-based designs, 14
modified toxicity probability interval
efficacy-toxicity trade-off contour, 159 (mTPI) design, 33
escalation with overdose control, 28 monotonicity, 103
multiple toxicities, 139
generalized BOIN (gBOIN), 140 multiple-toxicity BOIN (MT-BOIN),
case study, 149 144
decision boundaries, 143 derivation and properties, 147
derivation and properties, 142 software, 149
software, 149
normal-normal model, 3
historical data, 119
See also incorporating historical optimal biological dose (OBD), 13
data See also dose optimization

incorporating historical data, 119 phase I–II design, 156

comparative study, 127 See also dose optimization
model-assisted designs, 121, 126 prior effective sample size (PESS),
model-based designs, 119 120
incorporating toxicity grades, 137 prior information, 119
model-assisted designs, 140, 144 See also incorporating historical
informative BOIN (iBOIN), 121 data
case study, 128
decision table, 124 quasi-binomial likelihood, 173
operating characteristics, 127
rank-based desirability scores (RDS),
practical guidance, 125
176
robust prior, 125
real-world data, 119
software, 128
See also incorporating historical
informative keyboard (iKeyboard) de-
data
sign, 126
STEIN design, 190
keyboard combination design, 77
See also dose optimization
keyboard design, 35
KISS principle, 12 time-to-event BOIN (TITE-BOIN)
design, 98
late-onset toxicity, 93
case study, 112
a unified approach, 105
decision table, 100
comparative study, 103
incorporating prior information,
logistical difficulty, 93 102
model-assisted designs, 98, 109 operating characteristics, 103
Index 219

software, 112 U-BOIN design, 162

standardized total follow-up time case study, 194
(STFT), 99 operating characteristics, 170
statistical properties, 103 software, 193
time-to-event CRM (TITE-CRM), 96 utility-based risk-benefit trade-
TITE-BOIN12 design, 183 off, 163
case study, 194 See also dose optimization
software, 193 utility-based risk-benefit trade-off, 163
See also dose optimization uTPI design, 188
TITE-keyboard design, 109 See also dose optimization
TITE-mTPI design, 109
total toxicity burden, 137 waterfall design, 77
toxicity grade, 137 divide-and-conquer, 77
See also incorporating toxicity subtrial, 77
grades
toxicity score, 138