Journal of Real Estate Finance and Economics, 27:3, 303±320, 2003
# 2003 Kluwer Academic Publishers.Manufactured in The Netherlands.
A Semiparametric Method for Valuing Residential
Locations: Application to Automated Valuation
University of Connecticut, 2100 Hillside Road, Unit 1041RE, Storrs, CT 06269-1041, USA
This paper is motivated by automated valuation systems, which would bene®t from an ability to estimate spatial
variation in location value.It develops theory for the local regression model (LRM), a semiparametric approach
to estimating a location value surface.There are two parts to the LRM: (1) an ordinary least square (OLS) model
to hold constant for interior square footage, land area, bathrooms, and other structural characteristics; and (2) a
non-parametric smoother (local polynomial regression, LPR) which calculates location value as a function of
latitude and longitude.Several methods are used to consistently estimate both parts of the model.The LRM was
®t to geocoded hedonic sales data for six towns in the suburbs of Boston, MA.The estimates yield substantial,
signi®cant and plausible spatial patterns in location values.Using the LRM as an exploratory tool, local peaks and
valleys in location value identi®ed by the model are close to points identi®ed by the tax assessor, and they are
shown to add to the explanatory power of an OLS model.Out-of-sample MSE shows that the LRM with a ®rst-
degree polynomial (local linear smoothing) is somewhat better than polynomials of degree zero or degree two.
Future applications might use degree zero (the well-known NW estimator) because this is available in popular
commercial software.The optimized LRM reduces MSE from the OLS model by between 5 percent and 11
percent while adding information on statistically signi®cant variations in location value.
Key Words: land values, neighborhood house values, property tax assessment, automated valuation, local
polynomial smoothing regression, non-parametric methods, semiparametric models
Residential hedonic models have not dealt adequately with location, typically relying on
distances from known centers of activity (the central business district (CBD); an airport) or
neighborhood dummy variables.But, the values of identical houses are thought to vary
signi®cantly over space because of unmeasured location differences.
The proposed method, the local regression model (LRM), places a latitude and
longitude grid over the area and estimates the value of a standard building lot at each knot
on the grid.The grid can be as coarse or ®ne as desired, subject to the amount of data
available.Hedonic characteristics of the house are controlled parametrically.
The proposed method obviates the need for de®ning neighborhood boundaries by using
housing transactions at known points in space (i.e., geocoded data) to estimate a
semiparametric value surface.The method is descriptive: it seeks to faithfully reproduce
the value of a standard house over space at any point in time.Neighborhood boundaries are