Deep Doubly Debiased Longitudinal Effect Estimation with ICE G-Computation

2026-02-16 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

The D³-Net framework addresses challenges in estimating longitudinal treatment effects, particularly error propagation in Iterative Conditional Expectation (ICE) G-computation, which is crucial for sequential clinical decision-making. This novel approach mitigates error propagation during training by using Sequential Doubly Robust (SDR) pseudo-outcomes, providing bias-corrected targets for each regression. It employs a multi-task Transformer architecture with a covariate simulator head for auxiliary supervision and a target network to stabilize training. For the final estimate, D³-Net discards the SDR correction and applies Longitudinal Targeted Minimum Loss-Based Estimation (LTMLE) to the original outcomes, ensuring robustness and optimal finite-sample properties. Experiments on semi-synthetic MIMIC-III and MIMIC-IV datasets demonstrate that 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.

Key takeaway

For research scientists developing causal inference models for longitudinal data, D³-Net offers a robust solution to the persistent problem of error propagation in deep ICE G-computation. You should consider integrating SDR-based training with LTMLE targeting, along with architectural stabilizers like target networks and auxiliary covariate simulators, to achieve lower bias and variance in your estimations, especially in complex, long-horizon scenarios with strong treatment-confounder feedback.

Key insights

D³-Net combines SDR-based training with LTMLE targeting to robustly estimate longitudinal treatment effects, mitigating error propagation.

Principles

• Sequential Doubly Robust (SDR) pseudo-outcomes reduce error propagation in recursive training.
• LTMLE provides robust, statistically efficient final debiasing.
• Auxiliary supervision and target networks stabilize deep learning for causal inference.

Method

D³-Net trains a multi-task Transformer using SDR pseudo-outcomes, a covariate simulator head, and a target network, then applies LTMLE for final debiasing.

In practice

• Use SDR pseudo-outcomes to correct bias during recursive outcome regression training.
• Implement a multi-task Transformer with a covariate simulator for representation regularization.
• Apply LTMLE as a final debiasing step to improve finite-sample stability.

Topics

• Longitudinal Causal Inference
• G-computation
• Doubly Robust Estimation
• Transformer Networks
• Clinical Decision Support

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• Causal Decomposition Analysis with Synergistic Interventions: A Triply-Robust Machine Learning Approach to Addressing Multiple Dimensions of Social Disparities — An extended causal decomposition analysis evaluates synergistic effects of multi-domain interventions using a triply robust machine learning estimator.
• A Hitchhiker's Guide to Poisson Gradient Estimation — Eat${}_{ extsf{cubic}}$ offers a robust, unbiased method for Poisson gradient estimation by using cubic smoothstep.
• Prediction-Powered Causal Inference by Automatic Debiased Machine Learning and Semi-Supervised Riesz Regression — Prediction-Powered Causal Inference (PPCI) uses unlabeled data to significantly reduce asymptotic variance in causal and structural parameter estimation.
• Estimate Collapsibility of Causal Effects in Completed Partial DAGs via Strong d-Convex Hulls — A collapsible method estimates causal effects in CPDAGs, maintaining consistency before and after marginalization.
• Three-in-One World Model: Energy-Based Consistency, Prediction, and Counterfactual Inference for Marketing Intervention — A DBM-based world model with task-specific adapters unifies prediction, consistency, and counterfactual inference for marketing interventions.
• Uncertainty-Aware Longitudinal Forecasting of Alzheimer's Disease Progression Using Deep Learning — A probabilistic deep learning framework forecasts Alzheimer's disease progression with quantified uncertainty, outperforming existing baselines.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.