Abstract:
Hedonic models are subject to spatially correlated errors which are a symptom of omitted spatial variables, mis-specification or mismeasurement. Methods have been developed to address this problem through the use of spatial econometrics or spatial fixed effects. However, often spatial correlation is modeled without much consideration of the theoretical implications of the chosen model or treated as a nuisance to be dealt with holding little interest of its own. We discuss the limitations of current standard spatial approaches and demonstrate, both empirically and theoretically the generalized additive model as an alternative. The generalized additive model is compared with the spatial error model and the fixed effects model. We find the generalized additive model to be a solid alternative to the standard approaches, having less restrictive assumptions about the omitted spatial processes while still being able to reduce the problem of spatial autocorrelation and provide trustworthy estimates of spatial variables. However, challenges connected with spatially varying data remain. The choice of flexibility in the spatial structure of the model affects estimated parameters of some spatially varying characteristics markedly. This suggests that omitted variable bias may remain an important problem. We advocate for an increased use of sensitivity analysis to determine robustness of estimates to different models of the (omitted) spatial processes.