Identification of dynamic treatment effects when treatment histories are partially observed

2025-01-08

AI Summary

This paper addresses the identification and estimation of dynamic causal effects under partial treatment history missingness, particularly when outcomes depend on the complete treatment trajectory—accommodating fixed, absorbing, sequential, or synchronous treatment mechanisms. We propose a robust identification framework that achieves unbiased path-dependent effect estimation for heterogeneous subpopulations, even if any one of the outcome model, propensity score model, or missingness mechanism model is misspecified. Our method integrates doubly robust estimation, inverse probability weighting, and missing-data modeling to substantially enhance robustness to treatment history missingness. Simulation studies demonstrate negligible bias under various misspecification scenarios. Empirical analysis of 2022 U.S. election data reveals that voters in counties with high COVID-19 incidence during 2020–2021 exhibited a statistically significant decline in turnout, illustrating the practical relevance and reliability of our approach.

Technology Category

• Reasoning under Uncertainty: Causality
• Machine Learning: Causal Learning
• Humans and AI: Voting

Application Category

• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

This paper proposes a class of methods for identifying and estimating dynamic treatment effects when outcomes depend on the entire treatment path and treatment histories are only partially observed. We advocate for the approach which we refer to as `robust' that identifies path-dependent treatment effects for different mover subpopulations under misspecification of any one of three models involved (outcome, propensity score, or missing data models). Our approach can handle fixed, absorbing, sequential, or simultaneous treatment regimes where missing treatment histories may obfuscate identification of causal effects. Numerical experiments demonstrate how the proposed estimator compares to traditional complete-case methods. We find that the missingness-adjusted estimates have negligible bias compared to their complete-case counterparts. As an illustration, we apply the proposed class of adjustment methods to estimate dynamic effects of COVID-19 on voter turnout in the 2022 U.S. general elections. We find that counties that experienced above-average number of cases in 2020 and 2021 had a statistically significant reduction in voter turnout compared to those that did not.

Problem

• Identify path-dependent treatment effects with partial observation
• Develop robust estimator for missing treatment histories
• Improve accuracy in inference under missingness and misspecification

Innovation

• Robust estimator for missing treatment histories
• Combines outcome, propensity, missing models
• Accurate under missingness and misspecification