Regional climate models (RCMs) are important tools to study local climate behavior. The North American Regional Climate Change Assessment Program (NARCCAP) is an international program
designed to provide high-resolution, limate-model output for the United States, Canada, and northern Mexico. RCMs use physical relationships, in subtly different ways, to downscale
information from a coarse-resolution global climate model. Although all RCMs generally capture the large-scale spatial variation from coast to coast, from south to north, and from
low to high elevations similarly, their outputs can differ substantially in some regions. In this talk, I analyze the 20-year-average Boreal winter temperatures from each of six RCMs that were run in Phase I of NARCCAP (where all RCMs are
driven by the boundary conditions provided by the NCEP/DOE Reanalysis II data). A Bayesian hierarchical model (BHM) is built, which includes a spatial meta-analysis model that allows
a consensus regional climate to be defined from the ensemble of RCMs. In the BHM, each RCM's "vote" for the consensus climate is "counted" differently. An MCMC implementation enables
posterior inference on the "unknowns," including the large-scale fixed effects and the small-scale random effects in each RCM and in the consensus climate. Posterior inference on the resulting spatial random fields allows a visual comparison of the RCMs and the consensus climate. Because the BHM uses a fixed-rank model, Bayesian computation on the large datasets from the six RCMs is feasible. Additionally, the model has a
spatial covariance structure that can capture the
nonstationarities expected over a region with very heterogeneous physical geography. This talk is based on research with Emily Kang (SAMSI) and Steve Sain (NCAR).
Copyright. All Rights Reserved.