Causal discovery : finding common latent causes

In this project, we explore Causal Discovery with the presence of latent variables. We describe a method of testing covariance structures in order to distinguish causal from effect indicators in structural equation models. This is done by implementing a statistical hypothesis testing on simulated data. We analyse and compare different test statistics, including a bootstrap technique. We introduce the FOFC algorithm that checks quartets of measured variables in order to cluster them by their common cause, based on these statistical tests. Finally, we discuss how this could be used in more elaborate causal discovery algorithms such as the Copula PC.