Consistent and fast inference in compartmental models of epidemics using Poisson Approximate Likelihoods