Statistical methods for time-varying adaptive interventions

Abstract

The growing demand for personalized medicine has led to the popularity of sequential multiple assignment randomized trials (SMARTs) in developing and evaluating adaptive interventions (AIs). AIs are sequences of decision rules that guide whether, when, and how to adapt an intervention for a patient. A SMART design is longitudinal and sequential in nature, as participants may progress through multiple treatment stages: they are initially randomized to a treatment and may be subsequently re-randomized to adapted or additional treatments based on their response to a prior treatment or other measured characteristics. This dissertation addresses two key challenges when developing and evaluating AIs. The first challenge is handling unique missing data issues in SMARTs that arise from its complex sequential design, where earlier responses (and missing data) could directly impact subsequent treatment pathways. Although multiple imputation (MI) is a widely used missing data method that can facilitate unbiased estimation in longitudinal studies, it relies on the untestable assumption that data are missing at random (MAR). It is unclear how violations of MAR affect inference with MI about adaptive interventions in a SMART setting, eliciting the need for sensitivity analyses to evaluate the robustness of conclusions to deviations from MAR. We highlight challenges in conducting sensitivity analyses in SMARTs, evaluate the performance of MI under varying SMART-specific departures from MAR, and propose a tailored sensitivity analysis framework with metrics to synthesize the SMART sensitivity analysis for count and continuous outcomes. The second challenge lies in the resource-intensive process of identifying optimal AIs and then testing them against control or non-personalized conditions, which traditionally requires two separate studies. We propose two cost-effective and potentially less time-consuming alternatives to invesitgate personalized treatments. First, a group sequential method that combines the process of selecting the best AI and confirming it is superior to a control treatment within a single SMART, and also allows for early termination of the study if there is overwhelming evidence of the AI's superiority. Second, we apply a causal inference method (g-computation) to evaluate hypothetical interventions via simulation using existing observational data, offering a cost-effective alternative that can generate preliminary evidence about personalized treatments before initiating large-scale experimental studies. The methods in this dissertation were motivated by and are illustrated using data from two applications: a SMART aimed at reducing alcohol-related outcomes for college students and an observational cohort study examining mental health and financial outcomes in older adults from the Rush Alzheimer’s Center.