Incorporating Geospatial Data into House Price Indexes: A Hedonic Imputation Approach with Splines

We estimate a hedonic model of the housing market that includes a spline surface defined on geospatial data (i.e., the longitudes and latitudes of individual dwellings). House price indexes are then obtained by imputing prices for individual dwellings from the hedonic model and then inserting them into the Fisher price index formula. Using data for Sydney, Australia we compare the performance of four models: (i) generalized additive model (GAM) with a geospatial spline, (ii) GAM with postcode dummies, (iii) semilog with geospatial spline, (iv) semilog with postcode dummies. Our results clearly confirm the superiority of geospatial-splines, both in terms of the deviation between actual and imputed prices and in the case of repeat-sales between actual and imputed price relatives. Furthermore, the use of geospatial splines significantly affects the resulting house price indexes. The cumulative increase in our Fisher price indexes is between 8 and 30 percent higher (depending on the functional form and data sample) over the 2001 to 2011 period when a geospatial spline is used. This difference can be attributed to the failure of postcode dummies to fully adjust for omitted locational characteristics. Price indexes generated using geospatial splines are also more robust to changes in the functional form of the hedonics model and relatively immune to sample selection bias. (JEL. C43; E01; E31; R31)