1242 Spatial Hedonic Models
Kestenset al. (2006). However, their feasibility in practice is typically limited due
to the severe problems of multicollinearity that follow from the interaction terms.
Continuous spatial heterogeneity as implemented in GWR has seen several appli-
cations to hedonic specifications. Fotheringham’s text contains multiple examples
of hedonic price models (Fotheringhamet al., 2002), but others have applied this
methodology as well, such as Choet al. (2006), Kestenset al. (2006), Farber and
Yeates (2006), Longet al. (2007), Bitteret al. (2007) and Wanget al. (2008). Related
applications to repeat house sales models include McMillen (2003, 2004).
Pavlov (2000) suggests that the spatial varying coefficient model forms an alter-
native approach to dealing with sub-markets that outperforms several competing
specifications in terms of cross-validation residuals. The models evaluated include
a standard linear regression and a linear regression that includes dummy variables
for zip codes, as well as a parametric model including a quadratic polynomial of
theX,Ycoordinates of the points in the data.
Other applications focus more on the parameter instability related to specific
characteristics in the hedonic specification, such as environmental quality. For
example, Choet al. (2006) investigate whether public open space is capitalized
into nearby residential property values. They use an original dataset that includes
over 22,000 single-family housing sales transactions between 1998 and 2002 in
Knox County, Tennessee. Of this total sample, 15,500 transactions were randomly
selected for analysis. GWR estimation of a hedonic specification that includes prox-
imity to water bodies and parks suggests considerable variation in the marginal
prices of both amenities along different regions of the county. The resulting local
marginal prices would be obscured in a global model that assumes a constant
marginal price across the whole region.
A number of studies have compared the performance of the parametric expansion
method to the nonparametric GWR. For example, using data on 761 single-family
properties sold between 1993 and 2001 in Quebec City, Canada, as well as infor-
mation on the profiles of buyers, Kestenset al. (2006) compare the results from a
spatial expansion method and GWR. They also suggest that introducing detailed
household-profile data into the hedonic specification helps in explaining spatial
heterogeneity while at the same time reducing spatial dependence. Household
income, previous tenure status and age of the buyer show a significant effect on
house prices.
A similar comparison is carried out by Farber and Yeates (2006), who consider
global specifications, such as a standard linear and a SAR model (both estimated
using OLS) relative to local models such as GWR. Using an adjusted version ofR^2
as the criterion, they conclude that GWR obtains the best fit. Similarly, Bitteret al.
(2007), using data for over 10,000 single-family house sales in Tucson, Arizona, find
that GWR outperforms the expansion method in terms of both explanatory power
and predictive accuracy.^4 On the other hand, Kestenset al. (2006) suggest that, in
their application, GWR and the expansion method have similar explanatory power.
Longet al. (2007) assess the difference in predictive accuracy between moving
windows regression (MWR), GWR, kriging, and moving windows kriging (MWK)