Targeted Learning for Adaptive Experimentation and Data-Driven Decision-Making
Skip to main content
eScholarship
Open Access Publications from the University of California

UC Berkeley

UC Berkeley Electronic Theses and Dissertations bannerUC Berkeley

Targeted Learning for Adaptive Experimentation and Data-Driven Decision-Making

Abstract

Human decision-making often unfolds as a feedback loop: actions generate observations, observations refine insight, and insight guides the next action. As data collection and decision-making have become more continuous and dynamic, this loop has increasingly been operationalized through sequential adaptive experiments that continuously update treatment allocation or other aspects of the experimental design as data accumulate. These designs can improve participant outcomes and enhance the efficiency of statistical estimation, with applications ranging from adaptive clinical trials that update randomization probabilities based on interim analyses, to digital health platforms that personalize recommendations from streaming data, to online experiments that adapt product features based on continuously accumulating user responses. However, their adaptive nature, in turn, creates complex statistical challenges in learning from accumulating data, adapting experimental strategies to different objectives, and conducting valid statistical inference under adaptive data collection. This dissertation brings together three lines of my doctoral work that develop general frameworks for formulating research questions and addressing statistical challenges in adaptive experimental design and estimation, building upon the literature on targeted maximum likelihood estimation (TMLE) for adaptive designs. Chapter 1 introduces a general TMLE-based Online Super Learner framework for the real-time evaluation and selection among multiple candidate adaptive designs under a benefit-driven objective, with an application to evaluating and utilizing surrogate outcomes in sequential adaptive experiments. Chapter 2 establishes a semiparametric efficient TMLE framework for estimating a broad class of target estimands under diverse adaptive designs, with additional implications for constructing efficiency-oriented designs. Chapter 3 further develops a general design-and-estimation framework for longitudinal adaptive experiments with time-varying treatments to optimize estimation efficiency while enabling valid and efficient estimation and inference. Together, these complementary frameworks provide a coherent methodological foundation for advancing adaptive experimentation and data-driven decision-making in real-world settings.