Causal identifiability via Chain Event Graphs

We present the Chain Event Graph (CEG) as a complementary graphical model to the Causal Bayesian Network for the representation and analysis of causally manipulated asymmetric problems. Our focus is on causal identifiability — finding conditions for when the effects of a manipulation can be estimated from a subset of events observable in the unmanipulated system. CEG analogues of Pearlʼs Back Door and Front Door theorems are presented, applicable to the class of singular manipulations, which includes both Pearlʼs basic Do intervention and the class of functional manipulations possible on Bayesian Networks. These theorems are shown to be more flexible than their Bayesian Network counterparts, both in the types of manipulation to which they can be applied, and in the nature of the conditioning sets which can be used.