In recent years, the machine learning community has become more interested in causal inference. Identifiability criteria for causality between pairs of variables, and theoretical arguments employing Kolmogorov Complexity have been used to motivate the search for minimal causal networks. However, most practically useful research has been limited to the case of distinguishing the direction of causality between only two variables, under the assumption that one of the directions is correct.
Since confounding due to other potentially unobserved factors plagues nearly all empirical investigations, in this project I used a Minimum Description Length (MDL) approach to find out to what extent we can distinguish direct causal effects from those which are due to effects by a common factor – observed or unobserved – on both observed variables.