Deep Doubly Debiased Longitudinal Effect Estimation with ICE G-Computation

TL;DR

This paper introduces D3-Net, a novel framework that enhances longitudinal treatment effect estimation by reducing bias and variance through a combination of pseudo-outcomes, multi-task learning, and targeted debiasing, outperforming existing methods.

Contribution

The paper proposes D3-Net, which mitigates error propagation in ICE-based estimators using SDR pseudo-outcomes, multi-task Transformers, and a final targeted correction for robust effect estimation.

Findings

D3-Net significantly reduces bias and variance in effect estimates.

The method outperforms existing ICE-based estimators across various scenarios.

Robustness is demonstrated under different horizons and confounding conditions.

Abstract

Estimating longitudinal treatment effects is essential for sequential decision-making but is challenging due to treatment-confounder feedback. While Iterative Conditional Expectation (ICE) G-computation offers a principled approach, its recursive structure suffers from error propagation, corrupting the learned outcome regression models. We propose D3-Net, a framework that mitigates error propagation in ICE training and then applies a robust final correction. First, to interrupt error propagation during learning, we train the ICE sequence using Sequential Doubly Robust (SDR) pseudo-outcomes, which provide bias-corrected targets for each regression. Second, we employ a multi-task Transformer with a covariate simulator head for auxiliary supervision, regularizing representations against corruption by noisy pseudo-outcomes, and a target network to stabilize training dynamics. For the final estimate, we discard the SDR correction and instead use the uncorrected nuisance models to perform Longitudinal Targeted Minimum Loss-Based Estimation (LTMLE) on the original outcomes. This second-stage, targeted debiasing ensures robustness and optimal finite-sample properties. Comprehensive experiments demonstrate that our model, D3-Net, robustly reduces bias and variance across different horizons, counterfactuals, and time-varying confoundings, compared to existing state-of-the-art ICE-based estimators.