Bayesian model choice via mixture distributions with application to epidemics and population process models

We consider Bayesian model choice for the setting where the observed data are partially observed realisations of a stochastic population process. A new method for computing Bayes factors is described which avoids the need to use reversible jump approaches. The key idea is to perform inference for a hypermodel in which the competing models are components of a mixture distribution. The method itself has fairly general applicability. The methods are illustrated using simple population process models and stochastic epidemics