Deep Doubly Debiased Longitudinal Effect Estimation with ICE G-Computation
Researchers have introduced D3-Net, a novel framework designed to improve the estimation of longitudinal treatment effects, particularly in scenarios with time-varying confounders. The method addresses error propagation inherent in existing Iterative Conditional Expectation (ICE) G-computation techniques by employing Sequential Doubly Robust (SDR) pseudo-outcomes during training. Additionally, D3-Net incorporates a multi-task transformer with auxiliary supervision and a target network to stabilize learning. The final estimation uses Longitudinal Targeted Minimum Loss-Based Estimation (LTMLE) for enhanced robustness and optimal finite-sample properties, demonstrating superior performance over current state-of-the-art ICE-based estimators in comprehensive experiments. AI
IMPACT Introduces a novel framework to improve the accuracy and robustness of longitudinal effect estimation in machine learning models.