Overlap-weighted orthogonal meta-learner for treatment effect estimation over time

Core task: heterogeneous treatment effect estimation in time-varying settings. This field addresses the challenge of identifying how treatment effects differ across individuals when both treatments and confounders evolve over time. The taxonomy reveals a rich methodological landscape organized around data structures and modeling paradigms. Several branches focus on specific temporal data types: Longitudinal Observational Data with Time-Dependent Confounding tackles sequential confounding using methods like g-computation and marginal structural models (Time Varying Confounders[9], Variational Temporal Deconfounder[16]); Time-to-Event and Survival Outcomes extends HTE estimation to censored data (Survite[14], Competing Events HTE[1]); Panel Data and Difference-in-Differences Designs exploits repeated cross-sections or panel structure for policy evaluation (DiD Time Varying[7], TWFE Comparison[31]). Other branches emphasize decision-making frameworks: Dynamic Treatment Regimes and Sequential Decision Rules optimize multi-stage treatment policies (Dynamic Treatment Regimes[48], Optimal Dynamic Regimes[13]); Mobile Health and N-of-1 Trials address personalized interventions in digital health contexts (N of 1 Trials[42], Mobile Device Causal[27]). Meanwhile, Meta-Learners and Outcome Modeling Approaches adapt flexible machine learning architectures to time-varying settings, and Real-World Evidence branches connect methods to electronic health records and pragmatic trials (Patient Data Treatment Effects[2], Real World Evidence[28]). A central tension across these branches involves balancing flexible modeling of temporal dynamics against the need for robust causal identification under time-dependent confounding. Many works in the longitudinal observational branch develop neural or variational approaches to capture complex confounder trajectories (Copula CNN LSTM[3], Neural SDEs Trajectories[21]), while panel-based methods leverage fixed effects and parallel trends assumptions for identification (Interactive Fixed Effects[10], High Dimensional Panels[22]). The original paper, Overlap Weighted Orthogonal[0], sits within the Meta-Learners and Outcome Modeling branch, specifically focusing on overlap-weighted and orthogonal estimators. This positions it alongside flexible outcome-modeling strategies that adapt cross-sectional meta-learner frameworks to temporal data, emphasizing efficient estimation and robustness to model misspecification. Compared to works like ODE Longitudinal HTE[35] that embed mechanistic differential equation models or Temporal Uplift Modeling[50] that target marketing applications, Overlap Weighted Orthogonal[0] appears to prioritize statistical efficiency and doubly-robust properties in settings where overlap and covariate balance vary over time, bridging classical semiparametric theory with modern machine learning tools for time-varying HTE estimation.