Palgrave Handbook of Econometrics: Applied Econometrics

(Grace) #1
Luc Anselin and Nancy Lozano-Gracia 1237

out to eliminate the effects of spatial autocorrelation, which reduces the ultimate
sample size to 51,110 observations. The WTP estimates obtained from the differ-
ent models ranged between $15,719 and $34,154 for the SAR model and between
$15,639 and $30,489 for the OLS-based models. Beronet al.(2004) suggest that,
even though a spatial model for the first stage of the hedonic model provides a
statistically superior specification, it does not reduce the variability seen in esti-
mated benefits derived from different model specifications. While it is clear that
both spatial heterogeneity and spatial dependence violate the assumption of spher-
ical error terms, the implications for the empirical results of the hedonic model of
taking these misspecifications into account may vary from one case to the other.
The wide range seen in the estimated WTP for reductions in air quality calls
for further research in this area. Additional insight remains to be gained into the
sensitivity of the WTP to different specifications. In addition, many studies do not
obtain standard errors for the calculated welfare effects, which makes meaningful
comparisons difficult.
Further attention to non-marginal changes in environmental quality is given in
Brasington and Hite (2005), who found that the demand for environmental quality
is considerably more inelastic in spatial models than non-spatial models suggest.
They use house sales data in Ohio for 1991, aggregated to the census block group
(CBG) level, to estimate a hedonic specification that considers the presence of
spatial autocorrelation in both first and second stages of the model.^3 The original
number of house transactions considered is 44,255. However, since the study is
carried out at the CBG level, the effective sample size consists of the number of
CBG in the study, 5,051. Brasington and Hite estimate a model that includes both
a spatial lag and a spatially correlated error term, with the environmental variable
measured as the distance to the nearest hazard.^3
They show that the explanatory power of the spatial model is higher and, most
importantly, this model suggests a more inelastic demand function than the other
specifications. For example, using the non-spatial model, the estimated consumer
surplus loss from a decrease of half a mile in the median distance to a hazard is
$2,276 per household. In contrast, the spatial model gives an estimate of $3,278.
These results suggest that ignoring spatial characteristics may lead to underestimat-
ing the consumer surplus loss of a reduction in environmental quality. However,
since no standard errors are reported, it is difficult to assess the significance of the
difference in estimates.
Other recent examples of spatial econometric hedonic applications include
Anselin and Le Gallo (2006), Donovanet al. (2007), Huiet al., (2007), Munroe
(2007), Anselin and Lozano-Gracia (2008) and Anselinet al. (2008), among others.
Anselin and Le Gallo (2006) point out that the spatial interpolation method used
to create some of the explanatory variables included in hedonic models determines
to a great extent the estimates of marginal prices. They recommend the kriging
interpolator as the preferred method. Donovanet al. (2007) use a spatial lag model
to estimate the effect of wildfire risk on house prices. This study follows an inter-
esting approach to define the weights matrix, in which a correlogram is used to
determine the extent of spatial correlation. They conclude that ignoring spatial

Free download pdf