Spatial Modeling of House Prices Using Normalized Distance-Weighted Sums of Stationary Processes

Hedonic models are used almost universally for modeling house prices. Recognizing the importance of location, the past decade has seen increasing effort to introduce spatial considerations into such modeling. When spatial process models are used to capture association between locations, isotropic specifications have been used almost exclusively despite the fact that they seem unlikely to be appropriate in practice. The contribution of this article is to offer a novel, flexible, and computationally tractable class of non-stationary models. We accomplish this using suitably normalized distance-weighted sums of stationary processes. The number of component processes used reflects the flexibility required to adequately explain the spatial residuals in the model. A flexible nugget (or pure error) process is also introduced and is needed to capture the nonspatial idiosyncrasies of house sale transactions. The models are fitted within a Bayesian framework requiring demanding computation but yielding full and exact inference through the posterior distributions of the model unknowns. A dataset of 656 single-family home sales in Stockton, California, provide an illustration.

Duke Scholars

Author Alan E. Gelfand Statistical Science