High-dimensional causal inference often founders because conditioning on all variables is impractical or statistically inefficient. The authors introduce structural causal bottleneck models (SCBMs), a class of structural causal models in which causal effects between high-dimensional variables are mediated by low-dimensional summary statistics, or bottlenecks. This gives a flexible framework for mechanism-specific, target-dependent dimension reduction while remaining estimable with standard learning algorithms.

SCBMs come with theoretical guarantees. The authors prove that bottleneck variables are identifiable from observational data up to bijection, and they confirm this experimentally. In linear SCBMs, they examine three standard adjustment strategies. For instrumental variables, they show that naive two-stage least squares suffers a population-level non-identifiability when all variables are high-dimensional, and that bottleneck representations restore identification. For mediation and backdoor adjustment, conditioning on bottleneck variables yields valid effect estimates and provides explicit finite-sample efficiency gains over conditioning on the full high-dimensional variables.

The authors position SCBMs as a principled alternative to existing causal dimension reduction frameworks such as causal representation learning and causal abstractions. The paper, which appears at NeurIPS 2026, is available on arXiv.