The aim of this paper is to introduce a new model selection mechanism for cross sectional spatial models. This method is more flexible than the approach proposed by Florax et al. (2003) since it controls for spatial dependence as well as for spatial heterogeneity. In particular, Bayesian and Maximum-Likelihood (ML) estimation methods are employed for model selection. Furthermore, higher order spatial influence is considered. The proposed method is then used to identify knowledge spillovers from German NUTS-2 regional data. One key result of the study is that spatial heterogeneity matters. Thus, robust estimation can be achieved by controlling for both phenomena.