Gilles Kratzer: Simulation based inference in epidemic models

Classically, stochastic model's randomness vanishes for large population size. This is an issue for modelling infectious diseases that exhibit high seasonal stochasticity. In this Suceptible-Infected-Recovered (SIR) modelling context, infectious diseases are described by Partially Observed Markov Process (POMP). A framework to perform data driven inference over POMP objects will be presented with a focus on stochastic epidemic models. The class of simulation model which will be presented is an Euler multinomial with extra stochasticity added to the transition's rates of the individuals in between compartments. The inference algorithm, for this class of model, is based on a Sequential Monte Carlo (SMC) algorithm for hidden states inference and the parameter inference is based on Maximum Likelihood Estimation via Iterated Filtering (MIF) algorithm. Finally, as a proof of example, a single age strata without seasonality without extra noise is used to fit routine collected public health surveillance data (rotavirus data in Germany). Computational investigations will be presented to show that this framework works with this simple model and is suitable to infer more complicated models from epidemic data.