As large amounts of data become available, the need for computationally efficient data-driven methods is of great need. We study the use of Bayesian parametrization methods in setting where the likelihood is intractable but replaced with simulation-based approaches, e.g., Approximate Bayesian Computations (ABC). The simulation-based approaches increase the span of models one can consider. However, they often induce a more computationally demanding challenge.
Currently, we apply a version of ABC, Synthetic likelihoods, on a national-scale data-driven infectious bacterial spread in the Swedish cattle population. We combine real network- and measurement data with an epidemiological model inside of the high-performance simulator: SimInf. We see clear arguments from a public health perspective to further expand the methodology as it naturally promotes the use of statistical tests of suggested decisions, which point based methods does not support.
The project, in full, aims to expand the methodology into fields where data-driven methods show a viable route and to further develop the use of simulation-based Bayesian methods.