Classification and Regression Trees (CART) is a popular statistical technique for predictions that uses binary recursive partitioning to split out data. When dealing with spatial data, issues may emerge due to cross-sectional depen- dence. This paper contributes to the current literature by introducing an alternative CART algorithm for spatial data. The idea is that predictive performance can be improved by introducing spatial information in the algorithm. Results from an empirical application are reported, and the predictions obtained by our algorithm are compared with those obtained with standard CART.