B1214
Title: Causal discovery with non-Gaussian data and latent variables
Authors: Y Samuel Wang - Cornell University (United States) [presenting]
Abstract: Estimating causal structure is considered from multivariate observational data, possibly with latent confounding. Specifically, we assume the data is generated by a linear structural equation model with non-Gaussian errors. We show that if the true structure corresponds to a bow-free acyclic path diagram, then the exact causal structure--not just an equivalence class--can be identified. In particular, we propose a constraint based algorithm, BANG (Bow-free Acyclic Non-Gaussian) which consistently estimates the underlying graph. If the maximum in-degree of the graph is bounded, then the number of tests required by BANG is bounded by a polynomial of the number of observed variables.
