Latent Variable Modeling for Integrating Output from Multiple Climate Models

Numerical models of atmosphere–ocean circulation are widely used to understand past climate and to project future climate change. Although the same laws of physics, chemistry, and fluid dynamics govern any general circulation model, each model’s formulations and parameterizations are different, yielding different projections. Notwithstanding, models within an ensemble will have varying degrees of similarity for different outputs of interest. Multi-model ensembles have been used to increase forecast skill by using simple or weighted averages where weights have been obtained by considering factors such as estimated model bias and consensus with other models (Giorgi and Mearns, J. Clim. 15:1141–1158, 2002, Geophys. Res. Lett. 30:1629–1632, 2003; Tebaldi et al., Geophys. Res. Lett. 31:L24213, 2004, J. Clim. 18:1524–1540, 2005). This paper considers an alternative view of multi-model ensembles. For use with the North American Regional Climate Change Assessment Program (NARCCAP), multivariate statistical models are employed to characterize modes of similarity within the members of an ensemble. Specifically, we propose a spatially-correlated latent variable model which facilitates the exploration of when, where, and how regional climate models are similar, and what factors best predict observed locations of model convergence.