Interpretability versus out-of-sample prediction performance in spatial hedonic models

The presence of spatial correlation in the error terms is a well-known concern in the estimation of hedonic real estate valuation models. Several methods have been devised to address this issue in order to improve precision of parameter estimates and model predictions. When the hedonic model is used for valuation of externalities or amenities, such as noise or green spaces, focus is on reducing omitted variable bias on parameter estimates. In this context, interpretability of the spatial component is paramount. In the context of appraisal or automatic valuation, focus is primarily on out-of-sample prediction performance. This dichotomy of objectives implies a potential trade-off between interpretability of parameter estimates and out-of-sample prediction accuracy. Here, we compare four methods in terms of their suitability for amenity valuation and real estate valuation respectively. These include an aspatial OLS model, the fixed spatial effects OLS model, the spatial error model and a spatial generalized additive model utilizing a soap film smoother over geographic coordinates. Each model is estimated under differing spatial specifications, utilizing a rich, Danish dataset, and the trade-off between predictive accuracy and stability of parameter estimates for a spatial variable of interest is studied.