The Augmented Synthetic Control Method with Interference and Randomization-Based Covariance Analysis for Restricted Mean Survival Time Estimates

Estimates of the effects of policies to control infectious diseases are critical for informing public health decisions. Such policies are often deployed within a single unit (e.g., country), but quantifying their impact is difficult since the unit's outcome without treatment is unobservable. Since policies are typically not randomly assigned, their impact is usually evaluated using statistical analysis of observational data collected over time from the treated unit and comparable control units. The Synthetic Control Method (SCM) provides interpretable estimates of policy effects on single units in these settings, and the Augmented SCM (ASCM) modifies SCM for broader applicability. Despite their utility, SCMs are underutilized in public health and biomedical research. A barrier to using SCMs, particularly in infectious disease research, is the assumption of no interference - when one unit’s exposure affects another’s outcome. While interest may center on estimating an intervention’s direct effect, estimating spillover effects on other units may be essential to informing policy. The first part of this dissertation compares SCM and ASCM through analyses of an antimalarial campaign in Magude, Mozambique. Then, ASCM is extended via a stratified control framework to estimate direct and spillover effects when neighboring units receive treatment and interference may be present. This method is applied to assess localized COVID-19 lockdown effects in Chile. Simulations illustrate improved bias compared to standard ASCM under various data-generating processes. This method can improve analysis of interventions targeting disease spread but is broadly applicable to estimating policy effects when interference may be present. The second part of this dissertation develops randomization-based methods for covariate adjustment of restricted mean survival time (RMST) comparisons in randomized controlled trials. Existing covariate-adjusted RMST methods rely on model-based assumptions that may not be compatible with survival data complexity. Treatment differences in RMST are estimated using randomization-based ANCOVA (RB-ANCOVA) for categorized time-to-event (TTE) data, constraining covariate mean differences to zero. Confidence intervals improve precision over unadjusted estimates. The methodology is extended to hypothesis testing under the strong null of no difference between treatments for each participant to improve power and Type I error control. Extensions to continuous TTE data are also considered.