The Impact of Wind Power Projects on Residential Property Values ...

un with White’s heteroskedasticity correction to the parameter estimates’ standard errors (which
will not adversely influence the errors if they are homoskedastic).
Serial Autocorrelation: A standard formal test for the presence <strong>of</strong> serial autocorrelation in the
error term is the Durbin-Watson statistic (Durbin and Watson, 1951). Applying this test as
proposed by Durbin and Watson to the full panel dataset was problematic because the test looks
at the error structure based on the order that observations are included in the statistical regression
model. Any ordering choice over the entire panel data set invariably involves mixing home
transactions from various study areas. Ideally, one would segment the data by study area for
purposes <strong>of</strong> calculating this test, but that method was not easily implemented with the statistical
s<strong>of</strong>tware package used for this analysis (i.e., SAS). Instead, study area specific regression
models were run with the data chronologically ordered in each to produce twelve different
Durbin-Watson statistics, one for each study area specific model. <strong>The</strong> Durbin-Watson test
statistics ranged from 1.98–2.16, which are all within the acceptable range. 126 Given that serial
autocorrelation was not found to be a significant concern for each study area specific model, it is
assumed that the same holds for the full dataset used in the analysis presented in this report.
Spatial Autocorrelation: It is well known that the sales price <strong>of</strong> a home can be systematically
influenced by the sales prices <strong>of</strong> those homes that have sold nearby (Dubin, 1998; LeSage, 1999).
Both the seller and the buyer use information from comparable surrounding sales to inform them
<strong>of</strong> the appropriate transaction price, and nearby homes <strong>of</strong>ten experience similar amenities and
disamenities. <strong>The</strong>refore, the price for any single home is likely to be weakly dependent <strong>of</strong> the
prices <strong>of</strong> homes in close temporal and spatial proximity. This lack <strong>of</strong> independence <strong>of</strong> home sale
prices could bias the hedonic results (Dubin, 1998; LeSage, 1999), if not adequately addressed.
A number <strong>of</strong> techniques are available to address this concern (Case et al., 2004; Espey et al.,
2007), but because <strong>of</strong> the large sample and computing limits, a variation <strong>of</strong> the Spatial Auto
Regressive Model (SAR) was chosen (Espey et al., 2007).
Specifically, an independent variable is included in the models: the predicted values <strong>of</strong> the
weighted nearest neighbor’s natural log <strong>of</strong> sales price in 1996 dollars. 127 To construct this vector
<strong>of</strong> predicted prices, an auxiliary regression is developed using the spatially weighted average
natural log <strong>of</strong> sales price in 1996 dollars as the independent variable and the spatially weighted
average set <strong>of</strong> home characteristics as the dependent variables. This regression was used to
produce the predicted weighted nearest neighbor’s natural log <strong>of</strong> sales price in 1996 dollars that
is then included in the Base and Alternative Models. This process required the following steps:
1) Selecting the neighbors for inclusion in the calculation;
2) Calculating a weighted sales price from these neighbors’ transactions;
3) Selecting and calculating the weighted neighbors home characteristics; and
4) Forecasting the weighted average neighbor’s sales price.
Selecting the neighbors: To select the neighbors whose home transactions would most
likely have affected the sales price <strong>of</strong> the subject home under review, all <strong>of</strong> the homes that
126 <strong>The</strong> critical values for the models were between 1.89 and 2.53, assuming 5% significance, greater than 20
variables, and more than 200 cases (Gujarati, 2003).
127 <strong>The</strong> predicted value was used, instead <strong>of</strong> the actual value, to help correct for simultaneity or endogeneity
problems that might otherwise exist.
136