1st Edition
Dynamic Treatment Regimes Statistical Methods for Precision Medicine
Dynamic Treatment Regimes: Statistical Methods for Precision Medicine provides a comprehensive introduction to statistical methodology for the evaluation and discovery of dynamic treatment regimes from data. Researchers and graduate students in statistics, data science, and related quantitative disciplines with a background in probability and statistical inference and popular statistical modeling techniques will be prepared for further study of this rapidly evolving field.
A dynamic treatment regime is a set of sequential decision rules, each corresponding to a key decision point in a disease or disorder process, where each rule takes as input patient information and returns the treatment option he or she should receive. Thus, a treatment regime formalizes how a clinician synthesizes patient information and selects treatments in practice. Treatment regimes are of obvious relevance to precision medicine, which involves tailoring treatment selection to patient characteristics in an evidence-based way. Of critical importance to precision medicine is estimation of an optimal treatment regime, one that, if used to select treatments for the patient population, would lead to the most beneficial outcome on average. Key methods for estimation of an optimal treatment regime from data are motivated and described in detail. A dedicated companion website presents full accounts of application of the methods using a comprehensive R package developed by the authors.
The authors’ website www.dtr-book.com includes updates, corrections, new papers, and links to useful websites.
Preface
1. Introduction
What is a Dynamic Treatment Regime?
Motivating Examples
Treatment of Acute Leukemias
Interventions for Children with ADHD
Treatment of HIV Infection
The Meaning of \Dynamic"
Basic Framework 8
Definition of a Dynamic Treatment Regime
Data for Dynamic Treatment Regimes
Outline of this Book
2. Preliminaries
Introduction
Point Exposure Studies
Potential Outcomes and Causal Inference
Potential Outcomes
Randomized Studies
Observational Studies
Estimation of Causal E ects via Outcome Regression
Review of M-estimation
Estimation of Causal E ects via the Propensity Score
The Propensity Score
Propensity Score Stratification
Inverse Probability Weighting
Doubly Robust Estimation of Causal E ects
Application
3. Single Decision Treatment Regimes: Fundamentals
Introduction
Treatment Regimes for a Single Decision Point
Class of All Possible Treatment Regimes
Potential Outcomes Framework
Value of a Treatment Regime
Estimation of the Value of a Fixed Regime
Outcome Regression Estimator
Inverse Probability Weighted Estimator
Augmented Inverse Probability Weighted Estimator
Characterization of an Optimal Regime
Estimation of an Optimal Regime
Regression-based Estimation
Estimation via A-learning
Value Search Estimation
Implementation and Practical Performance
More Than Two Treatment Options
Application
4. Single Decision Treatment Regimes: Additional Methods
Introduction
Optimal Regimes from a Classification Perspective
Generic Classification Problem
Classification Analogy
Outcome Weighted Learning
Interpretable Treatment Regimes Via Decision Lists
Additional Approaches
Application
5. Multiple Decision Treatment Regimes: Overview
Introduction
Multiple Decision Treatment Regimes
Statistical Framework
Potential Outcomes for K Decisions
Data
Identifiability Assumptions
The g-Computation Algorithm
Estimation of the Value of a Fixed Regime
Estimation via g-Computation
Inverse Probability Weighted Estimator
Characterization of an Optimal Regime
Estimation of an Optimal Regime
Q-learning
Value Search Estimation
Backward Iterative Implementation of Value Search Estimation
Implementation and Practical Performance
Application
6. Multiple Decision Treatment Regimes: Formal Framework
Introduction
Statistical Framework
Potential Outcomes for K Decisions
Feasible Sets and Classes of Treatment Regimes
Potential Outcomes for a Fixed K-Decision Regime
Identifiability Assumptions
The g-Computation Algorithm
Estimation of the Value a Fixed Regime
Estimation via g-Computation
Regression-Based Estimation
Inverse Probability Weighted Estimator
Augmented Inverse Probability Weighted Estimator
Estimation via Marginal Structural Models
Application
7. Optimal Multiple Decision Treatment Regimes
Introduction
Characterization of an Optimal Regime
Specific Regimes
Characterization in Terms of Potential Outcomes
Justification
Characterization in Terms of Observed Data
Optimal \Midstream" Regimes
Estimation of an Optimal Regime
Q-learning
A-learning
Value Search Estimation
Backward Iterative Estimation
Classification Perspective
Interpretable Regimes via Decision Lists
Estimation via Marginal Structural Models
Additional Approaches
Implementation and Practical Performance
Application
8. Regimes Based on Time-to-Event Outcomes
Introduction
Single Decision Treatment Regimes
Statistical Framework
Outcome Regression Estimators
Inverse Probability of Censoring Regression Estimators
Inverse Probability Weighted and Value Search Estimators
Discussion
Multiple Decision Treatment Regimes
Multiple Decision Regimes
Statistical Framework
Estimation of the Value of a Fixed Regime
Characterization of an Optimal Regime
Estimation of an Optimal Regime
Discussion
Application
Technical Details
9. Sequential Multiple Assignment Randomized Trials
Introduction
Design Considerations
Basic SMART Framework, K = 2
Critical Decision Points
Feasible Treatment Options
Interim Outcomes, Randomization, and Stratification
Other Candidate Designs
Power and Sample Size for Simple Comparisons
Comparing Response Rates
Comparing Fixed Regimes
Power and Sample Size for More Complex Comparisons
Marginalizing Versus Maximizing
Marginalizing Over the Second Stage
Marginalizing With Respect to Standard of Care
Maximizing Over the Second Stage
Power and Sample Size for Optimal Treatment Regimes
Normality-based Sample Size Procedure
Projection-based Sample Size Procedure
Extensions and Further Reading
10. Statistical Inference
Introduction
Nonsmoothness and Statistical Inference
Inference for Single Decision Regimes
Inference on Model Parameters
Inference on the Value
Inference for Multiple Decision Regimes
Q-learning
Value Search Estimation with Convex Surrogates
g-Computation
Discussion
11. Additional Topics
Biography
Anastasios Tsiatis is Gertrude M. Cox Distinguished Professor Emeritus, Marie Davidian is J. Stuart Hunter Distinguished Professor, Shannon Holloway is Senior Research Scholar, and Eric Laber is Goodnight Distinguished Professor, all in the Department of Statistics at North Carolina State University. They have published extensively and are internationally-recognized authorities on methodology for dynamic treatment regimes.
"Biostatisticians, those that are professional as well as masters level and PhD level, will find this book useful. It is written by well-known experts who have incredible track records in this field, both methodologically and in designing and implementing/analyzing SMARTs and observational studies to uncover optimal dynamic treatment regimes. The text is rigorous in its statistical definitions and theorems. It is a comprehensive text on the area of dynamic treatment regimes and SMART design. Both those familiar with this area and those new to the area will learn something. They offer some interesting uses of the SMART design (e.g., dose finding and extending beyond 2 stages), that you cannot find in current manuscripts."
~Kelley Kidwell, University of Michigan"The book will serve as an excellent reference and textbook. I expect I will use the book in my own class, once it is available. Besides being a comprehensive treatment of dynamic treatment regimes, the revision/re-introduction to causal inference, potential outcomes, M-estimators, propensity scores, and related issues is extremely useful."
~Daniel Lizotte, The University of Western Ontario"Statisticians/biostatisticians directly involved in planning SMARTs would likely find this material useful, as they would have to adapt or extend these methods to particular trials being planned. Also, academic statisticians aiming to get into this field of methodological research would likely find the material as a useful summary of the already extensive literature; however, as the field is fast-moving, this material only serves as a starting point. The authors are not providing a cookbook-style guide to planning a variety of different kind of SMARTs, they provide examples, and enough theoretical background rigorously presented to get started in the area."
~Olli Saarela, University of Toronto"(Chapters 2-4) are very nice indeed: well written, well structured, informative and interesting. Congratulations to the authors… I went to the website, which is beautiful. The authors have put lots of effort into this.
~Robin Henderson, Newcastle University