Abstract

When is a statistical dependency between two variables best explained by the supposition that one of these variables causes the other, as opposed to the supposition that there is a (possibly unmeasured) common cause acting on both variables? In this paper, we describe an approach towards model specification developed more fully in our book Discovering Cuud Structure, and illustrate its application to the aforementioned question. Briefly, the approach is to determine constraints satisfied by the variance-covariance matrix of a sample, and then to conduct a quasi-automated search for the causal specifications that will best explain those constraints, 1.

Citations