All Times EDT
Keywords: Average treatment effect, model averaging, ensemble methods, real world evidence
In this presentation we provide an overview of ensemble-based approaches to estimating the average treatment effect from observational data emphasizing the importance of model selection and accounting for model selection uncertainty. While many researchers fall back on their preferred strategies such as specific combinations of propensity and outcome modeling, we argue that a better approach would combine multiple strategies within a single ensemble framework where each strategy contributes according to its support from the data. We illustrate these concepts with a recently proposed Frequentist Model Averaging framework (FMA, Zagar et al, 2021) and present a simulation study assessing its operating characteristics compared to using various pre-specified individual strategies or a single best strategy selected via a data-driven approach. The feasibility of the FMA framework depends critically on its implementation, such as having a user-friendly front-end and unified reports allowing for efficient evaluation of multiple individual stratgeis as well as the ensemble strategy