Paul Sampson,
Department of Statistics, 
University of Washington, 
Seattle

Title:
Revisiting the spatial deformation model for nonstationary spatial covariance: proposals for new methods and applications

Abstract:
It has long been recognized (at least since the 1980's) that stationary,
isotropic spatial correlation models are inadequate for describing most spatial
and spatio-temporal environmental processes due to local effects of factors like
topography and meteorology. This talk provides a brief review of a few of the
many proposals for modeling nonstationary spatial correlation, focusing on
Sampson and Guttorp's spatial deformation model, Higdon's process or kernel
convolution model, and Fuentes' model using weighted combinations of stationary
processes. Reich has recently introduced a generalization of this latter model
that yields covariate-dependent covariance functions. None of these models has
been made available in truly useful software.

We review aspects of the thin-plate spline deformation framework of Sampson and
Guttorp's model and then sketch ideas to make this approach more flexible and
computationally easier to fit to data. We present a number of approaches for
incorporating covariates into this modeling framework.