1236 Spatial Hedonic Models
(2008) and Anselin and Lozano-Gracia (2008), among others. A comparative per-
spective is offered in Bell and Bockstael (2000), using about 1,000 observations on
parcel data for Anne Arundel County, Maryland. Overall, the average difference
between estimates for the characteristics based on ML and GM methods is less
than 5%. However, the estimates for the (error) spatial autoregressive coefficient
differ between 10% and 29%.
In general, studies using a spatial econometric approach show significant differ-
ences in the estimates of marginal prices, in particular when employing a spatial
lag specification. For example, Pace and Gilley (1997) find that the simultaneous
spatial autoregressive (SAR) model ML estimator obtains a much better fit than
the OLS estimator using data from 506 sales in Boston SMSA (Standard Metropoli-
tan Statistical Area). Pace and Gilley (1998) also compare a SAR with OLS and
a grid adjustment model and conclude that, by going from the OLS to the SAR
specification, the estimated residuals fall by 44%.
For a spatial error model, the issue is not consistency of the estimates, but pre-
cision. Here again, a spatial approach seems to pay off. For example, Legget and
Bockstael (2000), using data on coastal properties from Ann Arundel County in
Maryland, suggest that the significance of the coefficients improves considerably.
This allows them to confirm the relationship between residential prices and water
quality with more confidence relative to the estimates obtained from OLS.
An important sub-set of empirical studies focuses on the way in which environ-
mental quality, and air quality in particular, becomes capitalized into the house
price. Considerable differences between the results of spatial and non-spatial esti-
mates are observed in these studies as well. For example, Kimet al. (2003) found
that the marginal price for air quality estimated using spatial 2SLS for a spatial lag
model was half the size of the estimate obtained through OLS. Using a survey of
609 owner-occupied households in Seoul, Korea, they estimated a hedonic price
equation in which air pollution is introduced asNOxandSO 2 , obtained from read-
ings from 20 monitoring stations. The air pollution measures were interpolated to
allocate a value to each of 78 residential sub-districts. An important contribution
made in Kimet al. (2003) is to spell out the estimation of the marginal benefit in
a spatial lag model. They note that it does not only include the direct effect seen
in traditional OLS applications, but also a so-calledspatial multipliereffect that
captures the “induced effects of a neighborhood’s housing characteristics change.”
Beronet al. (2004) go one step further and consider the welfare effects of non-
marginal changes in air pollution by estimating the second stage of the hedonic
model. They compare the welfare estimates from a 10% reduction in air pollution
between a standard regression using OLS and SAR-based models. Using single-
family home sales records for four counties in the South Coast Air Basin (Los
Angeles, Orange, San Bernardino and Riverside Counties), they estimate a semi-
log form of the hedonic price equation for six different years (1980, 1983, 1986,
1989, 1992 and 1995). Air quality is measured as the annual average of PM10 (air-
borne particulate matter resulting from the burning of fossil fuels, such as petrol
in cars) at each of 40 monitoring stations, and interpolated using the geostatis-
tical kriging technique. Interestingly, an additional random resampling is carried