Title: On the use of scoring rules for Bayesian model selection with improper priors
Authors: Erlis Ruli - University of Padova (Italy) [presenting]
Laura Ventura - University of Padova (Italy)
Monica Musio - University of Cagliari (Italy)
Abstract: The Bayes factor (BF) is the standard model selection tool. However, it is well known that BF tends to be sensitive to the prior distributions of the models under comparison and therefore requires careful elicitation. Furthermore, the Bayes factor cannot be used with objective improper priors, because of the dependence of the marginal likelihood on the arbitrary scaling constants of the model prior densities. It has been proposed to solve this problem by replacing marginal log-likelihood by a homogeneous proper scoring rule, which is insensitive to the scaling constants. We apply and study this methodology in the context of continuous exponential family. A couple of examples are provided.