Primary Submission Category: Weighting
To Balance Covariates for Time-varying Treatment: A Unification for Methods
Authors: Yige Li, José Zubizarreta,
Presenting Author: Yige Li*
Due to their complexity, ubiquity, and relevance in practice, particularly in the medical sciences, longitudinal observational studies of treatment effects are one of the active frontiers in causal inference. Here, three fundamental but seemingly unrelated methods for estimating the effects of time-varying treatments are the g-computation formula, inverse probability treatment weighting (IPTW), and augmented IPTW (AIPTW). In this paper, we offer a new interpretation and diagnostics for these methods. In particular, we present a unifying framework from the standpoint of covariate adjustment or balance. We show how these methods connect and differ in finite samples, homologating them as weighting optimization methods. From this, we propose new diagnostics for longitudinal studies of treatment effects and discuss an alternative weighting approach that directly targets them. In a simulation study, we show that the proposed approach is as efficient as the g-computation formula estimator and is robust to outcome model misspecification under regularity assumptions. In a case study, we analyze the time-varying effects of drug-taking behaviors on callous-unemotional trait changes among justice-involved male adolescents.