IGC-Net for conditional average potential outcome estimation over time
Konstantin Hess, Dennis Frauen, Valentyn Melnychuk, Stefan Feuerriegel
TL;DR
This paper introduces IGC-Net, a neural network model that accurately estimates time-varying potential outcomes by properly adjusting for confounders, advancing personalized medicine decision-making.
Contribution
The paper presents the first neural model to perform fully regression-based iterative G-computation for time-varying potential outcome estimation.
Findings
IGC-Net outperforms existing methods in experiments
Effectively adjusts for time-varying confounding
Enhances personalized decision-making in medicine
Abstract
Estimating potential outcomes for treatments over time based on observational data is important for personalized decision-making in medicine. However, many existing methods for this task fail to properly adjust for time-varying confounding and thus yield biased estimates. There are only a few neural methods with proper adjustments, but these have inherent limitations (e.g., division by propensity scores that are often close to zero), which result in poor performance. As a remedy, we introduce the iterative G-computation network (IGC-Net). Our IGC-Net is a novel, neural end-to-end model which adjusts for time-varying confounding in order to estimate conditional average potential outcomes (CAPOs) over time. Specifically, our IGC-Net is the first neural model to perform fully regression-based iterative G-computation for CAPOs in the time-varying setting. We evaluate the effectiveness of…
Peer Reviews
Decision·ICLR 2026 Poster
- The authors identified the limitations of existing work and proposed a new framework to address them. - Experiments are thoroughly conducted on both synthetic and real-world datasets. - Code and data are available for review.
- Missing important baselines. The paper omits several recent and competitive methods for time-varying counterfactual prediction, such as CGM [1], State-Space Counterfactual Models [2], and G-Transformer [3], that would provide a fairer and more rigorous empirical comparison. - The experiments implicitly assume correctly specified models and strong sequential ignorability and positivity. There is no investigation into the method’s robustness under violations of overlap or unmeasured confounding
- The paper proposes a method for treatment effect estimation over time, adjusting time-varying confounders. - The paper provides a theoretical foundation for the model. - An implementation code is available for review.
- The core idea is more likely to be an engineering adaptation of the G-computation. The manuscript does not convincingly differentiate its contribution and demonstrate conceptual advances from that method. - The motivation behind the study problem (treatment effect estimation over time) needs to be further clarified, particularly in relation to its real-world applicability. It is not evident whether the problem is testable, whether the authors have empirically validated it in real applications
Some of the key strengths of the paper are as follows: - While G-computation has already been adapted to deep learning models in literature, the authors aim to circumvent the problem of full distribution estimation, by proposing an iterative regression-based approach. Using this idea of iterated expectations the authors are able to avoid costly monte carlo sampling and problems of high-dimensional overfitting that arises in presented literature - Commendably, the authors provide a comprehensive
While the paper has made a commendable efforts, there are several key issues with the paper - First, the authors haven't considered some of the latest literature on this topic. For example, G-transformer [1] was already proposed by the authors of G-net as a signficant improvement by adapting the backbone from RNN to a transformer model. A similar effort has also been reported in [2]. Another paper that aims to capture the uncertainty in G-computation has been proposed in [3] - Second, while the
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsRisk and Safety Analysis · Fault Detection and Control Systems