Detecting the causal direction using elementary statistical knowledge

Hei Chan – National Institute of Advanced Industrial Science and Technology

Manabu Kuroki – Osaka University

Japan

We attempt to find the causal direction between two variables, where one is known to causally influence the other, but not vice versa, given a data set of N samples of the two variables. Our method involves the application of elementary statistical knowledge. We first identify groups of data sets according to the strength of statistical dependencies, and thus obtaining the relationships between different data generating processes and how many roots (that is, causes) in each group. We then examine the extent the data is truncated for each variable, allowing us to consider which variable to be manipulated or to be stochastically manipulated. By combining the two steps, we judge the causal direction between two variables in each dataset.

NIPS 2008 workshop on causality