Abstract

We propose a formalization of the three-tier causal hierarchy of association,intervention, and counterfactuals as a series of probabilistic logicallanguages. Our languages are of strictly increasing expressivity, the firstcapable of expressing quantitative probabilistic reasoning---includingconditional independence and Bayesian inference---the second encodingdo-calculus reasoning for causal effects, and the third capturing a fullyexpressive do-calculus for arbitrary counterfactual queries. We give acorresponding series of finitary axiomatizations complete over both structuralcausal models and probabilistic programs, and show that satisfiability andvalidity for each language are decidable in polynomial space.

 

Quick Read (beta)