B0200
Title: Non-stationary distributional regression methods for historical climate analysis
Authors: Finn Lindgren - University of Edinburgh (United Kingdom) [presenting]
Abstract: Traditional approaches for spatial and spatiotemporal analysis have typically had to assume different degrees of stationarity, both in space and time, for computational convenience. However, in reality, there is both spatial and temporal non-stationarity in the systematic and random behaviour of weather and climate. In addition, while some quantities may be well approximated with Gaussian distributions, others, such as the diurnal temperature range, are clearly non-Gaussian. Approaches to dealing with distributional non-stationarity include spatiotemporal quantile regression and semi-parametric spatiotemporal modelling. Methods that can be implemented in a Bayesian context are discussed through the inlabru and INLA R packages, allowing estimation of spatially and seasonally varying distributions, while keeping computationally efficient Gaussian random fields as core components of a hierarchical model.
