A1459
Title: Bayesian hierarchical bootstrap framework for causal subgroup estimation with a time-to-event outcome
Authors: Mengyao Shi - University of Toronto (Canada) [presenting]
Kuan Liu - University of Toronto (Canada)
Abstract: Estimating treatment effects within prespecified subgroups is a central goal of causal inference. For time-to-event outcomes, subgroup causal estimands are defined via g-computation by averaging a conditional survival model over the subgroup-specific covariate distribution. Existing Bayesian approaches typically treat these distributions independently, which can be unstable when subgroups are small or imbalanced. Right censoring further reduces information, increasing uncertainty in subgroup survival contrasts. The hierarchical Bayesian bootstrap (HBB) is extended to subgroup causal inference with right-censored outcomes. The HBB places a nonparametric hierarchical prior on subgroup covariate distributions, enabling borrowing of information across related subgroups while preserving heterogeneity. This is combined with a Bayesian accelerated failure time model to perform posterior g-computation, propagating uncertainty from both the survival model and the covariate distribution. The proposed framework stabilizes subgroup causal survival estimates in sparse settings without parametric assumptions. Simulation studies evaluate performance across varying subgroup sizes and censoring levels, and compare with Bayesian additive regression trees for heterogeneous survival effects.
