Non‐Gaussian Structural Equation Models for Causal Discovery
Shohei Shimizu · Wiley series in probability and statistics · 2016
The utilization of non-Gaussianity in estimating structural equation models (SEMs) is useful for causal discovery, because a wider variety of causal structures can be estimated in this manner than through classical methods. This chapter discusses the three fundamental Linear Non-Gaussian Acyclic Model (LiNGAM) models to elaborate the concepts and methods underlying the non-Gaussian causal discovery approach. These models include Basic LiNGAM, LiNGAM for time series and LiNGAM with latent common causes. The concepts and methods involved in the signal processing method known as independent component analysis (ICA) are closely related to those of LiNGAM. The chapter provides a brief overview of ICA model which represents a data-generation process in which latent independent source signals are linearly mixed with one another to generate observed signals. Methods for analyzing causal relations between discrete variables should be more extensively studied, as many variables in the social sciences are categorical.