A Gaussian Markov Random Fields Approach to Integrating Multiple Climate Model Output

• Activity Number: 456
• Type: Topic Contributed
• Date/Time: Wednesday, August 3, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics and the Environment
• Author(s): Shannon Neely*+ and William F. Christensen and Stephan Sain
• Companies: Brigham Young University and Brigham Young University and National Center for Atmospheric Research
• Keywords: Gaussian Markov Random Fields ; Regional Climate Models ; NARCCAP ; Factor Analysis ; Spatial Prediction

Abstract

Multi-model ensembles are commonly used in climate research to help quantify and understand single model uncertainties. There have been various published methods for creating these ensembles often by taking simple or weighted averages within a Bayesian hierarchical model framework. Christensen and Sain (Math Geosci, 2011) consider an alternative approach to multi-model ensembles using a spatially correlated latent variable model. Rather than providing a weighted average, their approach highlights similarities among model ensemble members and provides feedback about which models differ most from the group and why. The current work extends and improves their method with a model that uses a single smoothed factor analysis rather than running an independent factor analysis at each site and smoothing afterward with heterogeneous variance measurement-error-filtered kriging. We use Gaussian Markov Random Fields (GMRF) in conjunction with spatial factor analysis. We apply our method to the North American Regional Climate Change Assessment Program (NARCCAP). Our model still allows assessment of model similarities while incorporating a more realistic smoothing method.