Bridging deconfounding scores and causal deep learning representations

Determine a principled methodology to integrate deconfounding scores—which guarantee zero confounding bias and directly control overlap via the overlap divergence—with flexible neural‑network representations used in causal deep learning, thereby achieving representation learning that preserves identification and improves overlap.

Background

The authors focus on one‑dimensional linear deconfounding scores to enable analytic characterization and guarantees on overlap. In contrast, causal deep learning methods learn flexible multivariate representations (e.g., via neural networks) that can capture rich prognostic/balancing information but typically lack explicit zero‑confounding‑bias guarantees.

The paper notes the importance of unifying these approaches so that learned representations remain flexible while preserving the deconfounding property and improving overlap as quantified by the overlap divergence.