Deconfounding Scores and Representation Learning for Causal Effect Estimation with Weak Overlap

A team of researchers from leading institutions, including Oscar Clivio and Alexander D'Amour, has published a significant paper on a core challenge in causal machine learning: the 'weak overlap' problem. In causal effect estimation, 'overlap' (or positivity) is a key assumption requiring that all types of individuals have a chance of receiving any treatment. In high-dimensional data, this assumption often breaks down, causing popular estimators to become brittle and produce high-variance results. The authors propose a novel solution by introducing a general class of feature representations they call 'deconfounding scores,' which preserve the ability to identify causal effects while improving overlap. This framework elegantly unifies two classic tools—propensity scores and prognostic scores—as special cases.

The researchers formalize the search for a better representation as minimizing an 'overlap divergence' under a deconfounding constraint. Crucially, they derive closed-form expressions for these scores under a broad family of generalized linear models with Gaussian features. Their theoretical analysis reveals that within this class, prognostic scores—which summarize an individual's expected outcome without treatment—are the optimal choice for maximizing overlap. This provides a rigorous justification for a specific modeling approach in challenging high-dimensional settings. The paper, which will appear at the AISTATS 2026 conference, includes extensive experiments to empirically validate their findings, marking a step toward more robust and trustworthy causal inference from observational data.