Speaker
Description
In the standard approach for model checking a model is falsified if the observed data is too extreme given the fitted model, i.e., the p-value is too small. This answers a binary question about whether the model should be rejected or not. However, at least for complex systems, we already know beforehand that the model is a simplified approximation, and thus actually not correct. An alternative approach is Bayesian model checking, where alternative models are generated from the model to be tested, and then Bayesian model selection is performed between these models and the original, to see if one of the alternatives are better at explaining the data. In this talk I will present a method for performing model checking in this way without the need for Monte Carlo simulations, which means the check can be performed quickly. This also allows for iterating the approach, and one can thereby quantify, using Kullback-Leibler divergence, how much the approach is able to improve on the original model.
