IGC-Net for conditional average potential outcome estimation over time

TL;DR

The paper tackles estimating conditional average potential outcomes (CAPOs) over time from observational data amidst time-varying confounding. It introduces Iterative G-computation Network (IGC-Net), an end-to-end neural architecture that embeds regression-based iterative G-computation to adjust for time-varying confounders without relying on high-variance IPW or full distribution estimation. The model combines a transformer-based backbone with a sequence of G-computation heads to perform recursive conditional expectations, enabling CAPO estimation via a single regression-based objective. Empirical results on synthetic, semi-synthetic, and real-world data show superior accuracy and robustness, suggesting significant potential for personalized decision-making from electronic health records and related longitudinal data.

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

Figures (9)

• Figure 1: Iterative G-computation network. Neural end-to-end architecture of our iterative G-computation network.
• Figure 2: How our IGC-Net performs G-computation to adjust for time-varying confounding.
• Figure 3: Ablations.IGC-LSTM has competitive performance, while the biased transformer without proper adjustments is inferior.
• Figure 4: During inference, future time-varying confounders are unobserved (here: $(X_{t+1},Y_{t+1})$). In order to estimate CAPOs for an interventional treatment sequence without time-varying confounding bias, proper causal adjustments such as G-computation are required.
• Figure 5: Tumor growth data: We report previous results of the baselines with the new ablations: IGC-LSTM and BT. $\rightarrow$ Notably, our IGC-LSTM has competitive performance, while BT suffers from significant bias. Our IGC-transformer remains the strongest method.

Theorems & Definitions (12)

• Proposition 1
• proof
• Proposition 2
• proof
• Proposition 3
• proof
• Proposition 1
• proof
• Proposition 2
• proof