Use Gibbs sampler for Bayesian causal inference

[jaryaman]

Stan can't deal with latent discrete variables. We could try using a Gibbs sampler from pymc3 in the Bayesian causal inference notebook, as suggested in Pearl's Causality.