Nonlinear Causal Discovery in the Presence of Unobserved Variables

Summary

• Learning (nonlinear) causal relations amongst observed variables when there are hidden/latent variables present

• Objective: learn the maximal ancestral graph (MAG), where variables are connected by directed edges (X -> Y) or bi-directed edges (X <-> Y) that indicate causal relations and confounding relations, respectively

• Involves residual independence testing, log-likelihood estimation, and model selection methods to determine (1) the existence of confounders and (2) causal relations, when possible