Geostatistical mapping of real estate prices: An empirical comparison of kriging and cokriging

Small-sized housing samples and price predictions at nonobserved locations require
geostatistical approaches, particularly the kriging estimator. Nevertheless, geostatistics
has thus far received little attention in real estate economics. The article’s objective is
to empirically compare the prediction accuracy of univariate kriging variants, namely
detrended kriging (DK) and universal kriging (UK), and multivariate extensions,
including detrended cokriging (DCK) and universal cokriging (UCK). Both latter
methods consider structural and neighborhood characteristics as auxiliary variables.
While the price surfaces of DK and UK show nearly identical cross-validated accuracies,
the cross-validation-based prediction accuracy of DCK and UCK differ in favor
of the latter. If real estate agencies are faced with a univariate sample of property
prices, either DK or UK can be used, while in the multivariate case, UCK is recommended,
although numerically more complex.