Primary Submission Category: Multilevel Causal Inference
Examining Longitudinal Treatment Effects: Contrasts of Treatment Regimes Accounting for Time-varying Confounders and Clustering
Authors: Hanna Kim, Jee-Seon Kim,
Presenting Author: Hanna Kim*
In longitudinal social science research, treatments offered at different time points often share a common objective but vary in implementation. When individuals follow different treatment patterns, the effects can be conceptualized as contrasts of potential outcomes under static treatment regimes. Studying these causal estimands provides evidence on the effectiveness of competing strategies within a unified framework.
This study evaluates methods for estimating static treatment regime effects, addressing key challenges in social science applications: time-varying confounders and clustered data. For example, the impact of participating in Head Start from ages three to four on children’s vocabulary development may depend on intermediate vocabulary levels and cluster-specific variations in curricula. We compare longitudinal inverse probability of treatment weighting (l-IPTW), g-computation, and targeted maximum likelihood estimation (TMLE), adapting models to include cluster fixed effects and robust standard errors. Additionally, we integrate TMLE with the SuperLearner algorithm for flexible functional forms.
Results from a real data analysis of Head Start effectiveness and a small-scale simulation study highlight differences among approaches and provide practical guidance for identifying and interpreting longitudinal treatment effects under static regimes, a novel framework in social science research.