Abstract:
|
Estimating multivariate spatial models, such as the multivariate Matern, is challenging due to large data sizes, which slows likelihood evaluations, and large numbers of model parameters, which makes optimization difficult. In principle, though, having large datasets should make statistical estimation easier. We study the implementation of multivariate spatial models in the GpGp R package, which uses Vecchia's approximation to the likelihood and a Fisher scoring algorithm for optimization, as well as parallelization to make use of multi-core and multi-threaded computing environments. The approximations in GpGp involve selecting an ordering of the data points and choices of conditioning sets. We study several different choices for ordering and conditioning sets, and we validate the choices on prediction problems.
|