B0915
Title: Intuitive prior specification for spatiotemporal models in ecology
Authors: Luisa Ferrari - University of Modena & Reggio Emilia (Italy) [presenting]
Massimo Ventrucci - Department of Statistical Sciences, University of Bologna (Italy)
Alex Laini - University of Turin (Italy)
Abstract: The use of Bayesian GLMMs in ecology has become very popular, as they allow the inclusion of potential spatial and temporal dependencies which often characterize this type of data. However, the fundamental prior specification step is usually completely overlooked and the traditional choice of independent vague priors is naively adopted. This is specifically detrimental for variance parameters since the popular inverse Gamma distribution has been found to perform poorly. The hierarchical variance decomposition (HD) is a newly developed framework, based on the reparametrization of the variance parameters according to a decomposition tree. Once the user has defined a tree, the focus is no more on single variance parameters but on more intuitive parameters representing the proportions of variance explained by the random effects. This method facilitates the inclusion of prior information in the specification and has also been shown to perform better than the traditional approach. The aim is to study the application of the HD in ecology, as its intuitiveness greatly facilitates the introduction of experts' beliefs and raises awareness about the importance of a thorough prior specification. A particular model is presented for georeferenced data, highly common in ecology, which further eases the interpretability of this approach. Finally, general guidelines for the design of the tree are outlined, useful for a vast range of ecological applications.
