Bayesian Estimation of Causal Direction in Acyclic Structural Equation Models with Individual-specific Confounder Variables and Non-Gaussian Distributions

摘要

Several existing methods have been shown to consistentlyestimate causal direction assuming linear or some form ofnonlinear relationship and no latent confounders. However, theestimation results could be distorted if either assumption isviolated. We develop an approach to determining the possiblecausal direction between two observed variables when latentconfounding variables are present. We first propose a new linearnon-Gaussian acyclic structural equation model with individual-specific effects that are sometimes the source of confounding.Thus, modeling individual-specific effects as latent variablesallows latent confounding to be considered. We then propose anempirical Bayesian approach for estimating possible causaldirection using the new model. We demonstrate the effectivenessof our method using artificial and real-world data. color="gray">