A0595
Title: Comparing TMLE variants: Cross-fit and cross-validated approaches for robust causal effect estimation
Authors: Mohammad Ehsanul Karim - The University of British Columbia (Canada) [presenting]
Rainie Fu - The University of British Columbia (Canada)
Abstract: Targeted maximum likelihood estimation (TMLE) is widely used for causal effect estimation, offering double robustness and efficiency. Recent variants of single cross-fit TMLE (SCTMLE), double cross-fit TMLE (DCTMLE), CVqTMLE (cross-validation for the outcome model), and full CVTMLE enable better adaptation to machine learning models beyond the Donsker class. These methods are systematically compared, with and without repeated sample splitting. Vanilla TMLE is used as a baseline to evaluate statistical performance and computational trade-offs for estimating the average treatment effect in complex confounding settings. Using simulations and a real-world case study, bias, MSE, coverage, and standard error estimation are assessed. All variants were robust, though performance varied. SCTMLE had the lowest bias but higher relative error. DCTMLE minimized MSE and was strong across metrics. CVqTMLE was moderate across the board. Full CVTMLE excelled in coverage and standard error estimation. Repetition improved results, stabilizing after 2530 replicates. DCTMLE is recommended for minimizing error, full CVTMLE for uncertainty quantification, and CVqTMLE as a resource-efficient compromise. Method choice should align with study-specific priorities in reliable causal inference.
