A Bayesian Hierarchical framework for identifying regional hydroclimate trends or climate effects from continental or global data

Abstract:

Hierarchical Bayesian models are useful for modeling hydroclimatic trends and teleconnections with a formal approach to characterizing and reducing estimation uncertainties. A challenge to the application of these models to large areas is that the response can be spatially heterogeneous, and the choice of a local spatial covariance model and a large scale spatial trend model in the parameters of the Bayesian regression may not be intuitively obvious. We consider a multilevel modeling structure for exploring homogeneity of response in such data sets, through a multi-component mixture model. The approach allows the reduction of uncertainties through partial pooling of parameters across automatically chosen subsets of the data. Applications to a synthetic data set and to extreme precipitation data for the continental USA from the HADEX2 data set is presented considering trends and selected climate indices as potential predictors. The effect of changing the number of components in the mixture is demonstrated through the changing spatial membership and trends in the data.