Finding a causal ordering via independent component analysis

The application of independent component analysis to discovery of a causal ordering between observed variables is studied. Path analysis is a widely-used method for causal analysis. It is of confirmatory nature and can provide statistical tests for assumed causal relations based on comparison of the implied covariance matrix with a sample covariance. However, it is based on the assumption of normality and only uses the covariance structure, which is why it has several problems, for example, one cannot find the causal direction between two variables if only those two variables are observed because the two models to be compared are equivalent to each other. A new statistical method for discovery of a causal ordering using non-normality of observed variables is developed to provide a partial solution to the problem.