Conference is now cancelled

Spatial model selection and weighting

Evangelos Evangelou (University of Bath)

A typical problem faced by practitioners analysing spatial data is the choice among different spatial correlation families and the choice of the link function. An established measure of model comparison and weighting is AIC. However, calculation of AIC is not straightforward for Generalised Linear Spatial Models as the likelihood is intractable. In this talk I will address the problem of spatial model selection using generalised importance sampling (GIS). I will demonstrate how the GIS estimator can be used to approximate the AIC by evaluating the Bayes factors between the candidate models.

The approximated AIC can be used for model selection and weighting, thus providing an ensemble prediction method. I will demonstrate that the ensemble prediction outperforms the prediction from a single model in two examples: a Poisson and a binomial spatial generalised linear mixed model.