Causal Foundation Models for Partial Identification Introduced

The field of causal inference often grapples with the challenge of identifying causal effects from observational data, especially when unobserved confounding factors are present. This research introduces a new paradigm: causal foundation models designed for partial causal identification. These models aim to bound the effects of interventions and counterfactuals even when full identification is not possible. The core idea involves defining a canonical prior with full support over the space of structural causal models, specifically those with discrete observables. By establishing this prior, the complex problem of bounding counterfactuals is reframed. It becomes a task of learning distributions over functions that can map observed data, along with any prior structural assumptions, directly to the causal query of interest. This approach extends the concept of causal foundational modeling to scenarios where causal effects are only partially identifiable. This means that instead of a single, precise causal effect, the model can provide a range of values that are equally compatible with the available data and initial structural assumptions, offering a more realistic and robust assessment in the presence of unobserved confounders.