Bayesian Clinical Trial Design for Distributional Outcomes with Noncompliance.

Abstract

We consider sequential monitoring in early-phase clinical trial designs with patient noncompliance. Unlike most existing methods that focus on modeling a single endpoint, we consider the entire outcome distribution and propose a Bayesian causal effect estimation method based on the Wasserstein distance. Using a principal stratification framework and random forest to classify patient compliance statuses, we quantify a distributional causal-effect magnitude among Compliers by calculating the Wasserstein distance between the barycenters of their potential outcome distributions. The proposed method also adaptively determines early trial termination based on accumulated outcomes. Simulation results show favorable operating characteristics under the scenarios investigated, with the proposed analysis remaining close to the oracle benchmark as noncompliance increases.