Prediction Intervals for Time-varying Individual Treatment Effects in Micro-randomized Trials

Abstract

Background: Accurately estimating individual treatment effects (ITEs) across multiple decision points is crucial for personalized decision-making in dynamic environments such as healthcare, adaptive interventions, and financial forecasting. Previous work has primarily focused on constructing prediction bands for ITEs using cross-sectional data or predicting non-causal longitudinal estimands, both relying on assumptions of exchangeability. Objectives: Our work bridges these two approaches by proposing a novel method for constructing prediction intervals for time-varying ITEs without the need for exchangeability, a key limitation in much of the prior literature. Methods: Using conformal inference techniques, our method ensures valid marginal coverage with minimal assumptions about the data distribution, making it broadly applicable across various decision-making contexts. We particularly focus on its application to micro-randomized trials (MRTs), where participants are randomized at multiple decision points, and treatment effects can vary significantly over time and across individuals. The ability to quantify uncertainty for ITEs at each decision point is critical in MRTs because decisions are sequential and individualized, requiring precise and reliable measures of treatment effects as they evolve. Results: Simulations emulating MRTs reinforce our theoretical claims, demonstrating that our method effectively captures uncertainty in ITEs under time-varying treatment conditions. Additionally, we validate the practical utility of our approach by applying it to the Intern Health Study dataset, showcasing its potential for real-world applications. Conclusions: This work enhances causal inference and predictive modeling, offering a robust tool for improving the reliability and interpretability of ITE estimates in complex, multi-stage decision processes.