Zero-Inflated models in Stan

[betanalpha]

@jsocolar

The simplifications of the likelihood from @betanalpha’s post above will improve computational efficiency and should be used if possible. However, the are not applicable in all cases. Namely, if you want to model lambda as varying across observations depending on covariates, then you cannot convert the Bernoulli likelihood to a Binomial likelihood.

To be clear in this case there is a similar decomposition even after covariates are introduced, only instead of having independent Bernoulli and baseline models one gets a logistic regression along with the baseline model.