Primary Submission Category: Heterogeneous Treatment Effects
CEEClust: A Bayesian Heterogeneous Time-Varying Causal Effect Model for Micro-Randomized Trials
Authors: Brody Erlandson, Tianchen Qian, Matt Koslovsky, Ander Wilson,
Presenting Author: Brody Erlandson*
Micro-randomized trials (MRTs) are designed for developing and optimizing mobile health interventions. A primary estimand in MRT research is the causal excursion effect (CEE), which estimates the difference in outcomes that would result from following one excursion policy versus another over a given time period. Existing research has largely focused on population-level CEEs. However individuals often have different response dynamics, including variation in response time, effect magnitude, and adherence to the intervention. Capturing this heterogeneity is essential for understanding underlying behavioral mechanisms and to help inform the design of more adaptive and personalized interventions. We propose a new CEE model in the Bayesian paradigm that incorporates nonparametric priors to learn latent subgroups of individuals with similar time-varying treatment effects. Additionally, the proposed method provides uncertainty quantification for the population-level and cluster-specific CEE. Although the primary motivation are mobile health interventions, the proposed method could be applied with minimal modification to other large decision point time-varying treatment settings in which the CEE is of interest; such as, estimating the causal effect of heat warnings on reducing heat-related hospital visits. Lastly, we conduct simulation studies to evaluate the proposed method’s ability to recover heterogeneous response patterns and apply our approach to data from the HeartSteps MRT.