A0346
Title: Learning causal directionality via a directed entropy-information metric
Authors: Peter Song - University of Michigan (United States) [presenting]
Soumik Purkayastha - University of Michigan (United States)
Abstract: Interest in direction of dependence in statistical literature has been rising, as such knowledge is critical for directed graphs used to study causality and mediation. Studying asymmetric dependence between two variables may help confirm a hypothesized causal relationship. While most standard statistical models can estimate the magnitude of associations, they cannot distinguish between effect and direction of effect as they implicitly assume a symmetric dependence between two variables. We posit a new measure called the directed mutual information (DMI) in which mutual information and conditional entropy are used to study both association and directed dependence. We establish key large-sample properties for the DMI and develop algorithms to test for independence as well as quantify directed dependence. The proposed method is evaluated by simulation studies and applied to a real-world example.
