IGC-Net for conditional average potential outcome estimation over time

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 our IGC-Net across various experiments. In sum, this work represents a significant step towards personalized decision-making from electronic health records.

💡 Research Summary

The paper tackles the problem of estimating conditional average potential outcomes (CAPOs) over time from observational data, a task that is central to personalized medicine but is complicated by time‑varying confounding. Existing neural approaches fall into two categories. The first group (e.g., CRN, Causal Transformer, TE‑CDE) relies on balanced representations to mitigate confounding, but balancing was originally designed to reduce variance, not bias, and therefore does not guarantee unbiased CAPO estimates. The second group (e.g., RMSN, G‑Net, G‑Transformer) performs explicit causal adjustments, yet each suffers from a critical drawback: RMSN uses inverse propensity weighting (IPW), which leads to exploding variance when propensity scores approach zero, especially for multi‑step ahead predictions; G‑Net and G‑Transformer employ G‑computation but require estimating the full joint distribution of all future time‑varying covariates and outcomes, which is computationally prohibitive and introduces Monte‑Carlo sampling error.

To overcome these limitations, the authors propose the Iterative G‑Computation Network (IGC‑Net), a novel end‑to‑end neural architecture that integrates G‑computation through a regression‑based iterative scheme. The key insight is to rewrite the nested expectations of the G‑computation formula as a recursion of conditional expectations. For a horizon τ, pseudo‑outcomes (G^{\bar a}{t+\delta}) are defined for (\delta = 0,\dots,\tau-1) with (G^{\bar a}{t+\tau}=Y_{t+\tau}). Starting from the last step, the model learns a regression function (g^{\bar a}_{t+\delta}) that predicts the pseudo‑outcome at step (\delta) given the observed history and the planned treatment sequence up to that point. This yields a sequence of low‑dimensional regression problems rather than a high‑dimensional density estimation.

Training proceeds in an alternating “generation‑learning” loop. In the generation (A) phase the current network parameters are used to predict the pseudo‑outcomes for all intermediate steps. In the learning (B) phase these predictions, together with the observed final outcome, are regressed on the history to update the conditional expectation functions. Because both phases are implemented within a single computational graph, gradients flow through the entire recursion, ensuring that information is shared across time steps and that the estimator converges to the true CAPO (Proposition 1). Proposition 2 proves that the procedure yields a proper adjustment for time‑varying confounding, while Proposition 3 demonstrates that the variance of IGC‑Net’s estimator is uniformly lower than that of IPW‑based RMSN.

Architecturally, IGC‑Net consists of a neural backbone (a modified transformer encoder) that processes the longitudinal covariate and treatment history, followed by treatment‑conditioned regression heads at each time step. A “blocked‑gradient” technique isolates the pseudo‑outcome generation from the parameter update to stabilize training. The model does not require Monte‑Carlo sampling or inverse propensity scores, eliminating the two major sources of instability in prior work.

Empirical evaluation includes (1) synthetic simulations where propensity scores can be arbitrarily close to zero, and (2) a real‑world electronic health record (EHR) dataset of patients with severe diabetes undergoing medication regimes. In synthetic experiments IGC‑Net achieves substantially lower mean‑squared error and higher policy value than RMSN, G‑Net, CRN, and TE‑CDE, even under extreme confounding. In the EHR study, IGC‑Net reduces mean absolute error by more than 15 % relative to the best competing method and yields more accurate individualized treatment effect predictions, as measured by held‑out outcome trajectories and simulated treatment policies.

In summary, IGC‑Net offers three decisive advantages: (1) it avoids the variance explosion of inverse propensity weighting, (2) it sidesteps the computational and statistical burdens of full distribution estimation and Monte‑Carlo sampling, and (3) it delivers an end‑to‑end trainable system that properly adjusts for time‑varying confounders. The work therefore represents a significant methodological advance for causal inference in longitudinal health data and provides a practical tool for clinicians seeking data‑driven, personalized treatment recommendations.