Events

Understanding causal relationships among variables is critical for effective decision-making in diverse fields such as economics, healthcare, social sciences, and business analytics. While numerous methods for causal discovery have been proposed, existing approaches typically yield single-point estimates or equivalence classes, often neglecting the quantification of uncertainty. In practice, however, it is crucial to understand the reliability of inferred causal structures, especially in high-stakes scenarios where incorrect causal inferences can lead to misguided interventions.

In this talk, we introduce a novel approach to quantify uncertainty in causal discovery by constructing confidence sets for causal orderings. Recognizing that establishing a causal ordering among variables is fundamental yet challenging, our methodology focuses on generating sets of causal orderings that remain plausible given observed data. Specifically, we consider identifiable structural equation models with additive errors and employ a residual bootstrap procedure designed to assess the goodness-of-fit of candidate causal orderings. Our proposed method provides asymptotically valid confidence sets, enabling practitioners to explicitly incorporate model uncertainty into their causal inference process.

We demonstrate how these confidence sets not only allow researchers to identify ancestral relationships with specified confidence but also facilitate constructing confidence intervals for causal effects that inherently reflect underlying model uncertainty.

Joint work with Y. Samuel Wang and Mathias Drton.   https://arxiv.org/abs/2305.14506