Deep Doubly Debiased Longitudinal Effect Estimation with ICE G-Computation

A new framework called D3-Net aims to reduce error propagation in a statistical method used for estimating longitudinal treatment effects, according to a paper posted on the arXiv preprint server [1]. The approach modifies Iterative Conditional Expectation G-computation, a technique that can be undermined when errors compound across its sequential steps [2]. The paper, titled "Deep Doubly Debiased Longitudinal Effect Estimation with ICE G-Computation," was submitted by Wenxin Chen and first appeared on arXiv on 12 February 2026, with a revised version uploaded on 12 June 2026 [1]. The initial submission was 905 KB, and the updated version is 1,474 KB [1]. arXiv, which began operating in 1991, is an open-access repository for preprints that are moderated but not peer-reviewed, and it now receives roughly 24,000 submissions per month [6]. The core challenge the work addresses is treatment-confounder feedback in sequential decision-making. ICE G-computation is a principled method for this setting, but its recursive structure means that errors in one regression model can corrupt subsequent models [2]. D3-Net interrupts this cycle by training the ICE sequence using Sequential Doubly Robust pseudo-outcomes, which provide bias-corrected targets for each regression step [2]. The framework also incorporates a multi-task transformer with a covariate simulator head for auxiliary supervision and a target network to stabilize training [2]. For the final treatment effect estimate, the authors discard the Sequential Doubly Robust correction. Instead, they apply Longitudinal Targeted Minimum Loss-Based Estimation to the original outcomes using the uncorrected nuisance models, a second-stage targeted debiasing step intended to ensure robustness and optimal finite-sample properties [2]. The authors report that in comprehensive experiments, D3-Net reduced both bias and variance across different horizons, counterfactuals, and time-varying confounding scenarios when compared to existing ICE-based estimators [2]. The paper is available through arXiv's standard interface, which also offers community-developed tools such as Bibliographic Explorer and Connected Papers for navigating related research [5].