B1891
Title: Consistent model selection and elicited prior information: The posterior equivalence principle for chain event graphs
Authors: Peter Strong - University of Warwick (United Kingdom) [presenting]
Jim Smith - Warwick University (United Kingdom)
Abstract: When using elicited prior parameter conditional distributions, current Bayesian model selection techniques can lead to a lack of consistency between how the elicited information and the data are treated. We propose a solution: the posterior equivalence principle. This satisfies the condition where, when performing model selection with the same set of models and the same parameter posterior distribution for each model, the distribution over the set of models should be the same. We demonstrate how we can satisfy this condition for the Bayesian Dirichlet score on Chain Event Graphs by setting a structural prior.
