A1336
Title: A numerical procedure to estimate dynamic treatment regimes from observational data using structural equation modeling
Authors: Terrence Jorgensen - University of Amsterdam (Netherlands) [presenting]
Wen Wei Loh - Maastricht University (Netherlands)
Abstract: Estimating causal effects from longitudinal observational designs is complicated by natural changes in participation (e.g., students in a multi-year mentoring program to improve academic outcomes may decide to opt-out after better attendance and behavior). Standard comparison of static vs. non-participation has limited real-world relevance, whereas a Dynamic Treatment Regime (DTR) offers greater insight into an interventions efficacy by accounting for differential participation over time. DTRs are an established framework in personalized medicine designed to investigate how individual treatment decisions can optimize intervention effects. Because conducting a randomized study to assess multiple DTRs is rarely feasible practically, how to estimate DTRs using observational longitudinal data is described. Crucially, the estimation procedure resolves the perennial challenges of treatment-dependent confounding inherent in longitudinal designs with treatment-confounder feedback. A numerical procedure is described to estimate causal effects using lavaan, a freely available open-source R package for structural equation modeling, which is widely used in psychological and education sciences. Estimating and interpreting the proposed DTR analysis is demonstrated using educational data as an example. DTRs offer the potential to spur more relevant, tailored, and impactful interventions in psychological and social science research.
