Spatial Aggregation and Prediction in the Hedonic Model

Using a data set of Wake County, North Carolina, property sales for the period 1992-2000, this study provides evidence as to the acceptability of spatial aggregation in hedonic property value models. Both statistical tests and tests based upon prediction errors are performed in order to identify the circumstances under which aggregation is statistically acceptable or acceptable from a practical standpoint. This study makes extensive use of spatial econometric techniques in order to control for the spatial correlation problems which exist in models where location matters, and discusses the importance of specification and functional form as determinants of both the acceptability of aggregation and predictive power. Since multiple specifications and types of models are estimated, this study also provides guidance as to the type of model or specification providing the best performance when used to estimate hedonic property value models.The primary finding of this study is that while statistical tests typically reject aggregation, the effects of aggregation upon prediction errors is negligible. We would typically expect less than a 2000 dollar increase in mean absolute prediction error from aggregating the entire county, while in several cases the out-of-sample predictions would be improved. Further, in many cases aggregation yields more plausible coefficient values, especially for less important determinants of property values. These results may indicate that aggregation is preferable to extensive disaggregation when conducting hedonic property values studies, especially if one is concerned with the coefficient estimates. I also find that a spatial error model is typically preferred over OLS and Box-Cox alternatives, even when those alternatives include additional variables describing the locational characteristics of the properties and the spatial error model does not.