Evaluation and testing of non-nested specifications of spatial econometric models

Spatial econometric models have generally non-nested specifications if they are based on different spatial setups (connectivity and weight matrices). For maximum likelihood estimators and non-nested models, the usual tests (likelihood ratio, Wald, et.c) gnerally cannot be used for model selection and/or testing. This article provides a structured discussion on estimation and evaluation (selection and/or testing) of non-nested spatial models. The distinction between model selection and model testing is important. While model selection algorithms approach all models symmetrically, the null and alternative models are treated differently for testing purposes. The empirical part of this paper provides an illustrative application of the evaluation methods discussed. Emphasis is given to models estimated by maximum likelihood approaches and to Vuong's test, which is derived from the Kullback-Leibler information criterion.