Yewen Chen

Yewen Chen

he/him

Estimating direct and indirect treatment effects and ranking treatment regimens in spacetime-dependent cluster randomized trials: a Schistosomiasis case study

Accurate estimation of treatment effects and reliable ranking of treatment regimens are essential to guide optimal treatment assignments in cluster randomized trials (CRTs). However, these efforts are often hindered by high bias and large variance. The high bias can arise from indirect effects (IEs) or time-varying direct effects (DEs), while the large variance results primarily from inadequate consideration of intrinsic dependencies. Motivated by the need to identify optimal treatment regimens in the Schistosomiasis CRT, this study addresses the methodological challenges introduced by sequential treatments and spatiotemporal dependencies. We identify optimal treatment regimens by evaluating differences through two components: DEs from the most recent treatment and IEs from historical treatment trajectories. To efficiently estimate DEs and IEs, we develop a Joint Spatiotemporal Varying Coefficient (JSTVC) model. JSTVC accounts for spatiotemporal dependencies and regional heterogeneities, while also capturing spatial anisotropic patterns in schistosomiasis transmission. To support scalable inference under complex dependent structures, we extend a hybrid computational method that integrates Variational Bayes with ensemble-based techniques. This facilitates both statistical inference and rigorous comparisons across regimens. Numerical and graphical comparisons demonstrate that JSTVC accurately captures random effects and outperforms competing methods that ignore IEs or spatiotemporal dependencies in estimating treatment effects and ranking regimens. The proposed methodology provides a broadly applicable framework for modeling complex dependencies in randomized experiments, especially in those involving multiple sequential interventions.

Figure 1: Ranking treatment sequences across methods: (A) The proposed JSTVC; (B) JSTVC without the spatiotemporal random effect; (C) JSTVC which mixed DEs and IEs and did not decompose the ATE; and (D) Ranking based on the average reduction in outcomes from Year 1 to Year 5.
Figure 2: Spatiotemporal patterns of schistosomiasis prevalence based on the recovered random effects.

Recommended citation: Chen Y, Wen X, Luo F, Yang Y, and Shen Y. (2025). "Estimating direct and indirect treatment effects and ranking treatment regimens in spacetime-dependent cluster randomized trials: a Schistosomiasis case study." Revised.