Bandwidth Selection for Spatial Hac and Other Robust Covariance Estimators

This research note documents estimation procedures and results for an empirical investigation of the performance of the recently developed spatial, heteroskedasticity and autocorrelation consistent (HAC) covariance estimator calibrated with different kernel bandwidths. The empirical example is concerned with a hedonic price model for residential property values. The first bandwidth approach varies an a priori determined plug-in bandwidth criterion. The second method is a data driven cross-validation approach to determine the optimal neighborhood. The third approach uses a robust semivariogram to determine the range over which residuals are spatially correlated. Inference becomes more conservative as the plug-in bandwidth is increased. The data-driven approaches prove valuable because they are capable of identifying the optimal spatial range, which can subsequently be used to inform the choice of an appropriate bandwidth value. In our empirical example, pertaining to a standard spatial model and ditto dataset, the results of the data driven procedures can only be reconciled with relatively high plug-in values (n or n). The results for the semivariogram and the cross-validation approaches are very similar which, given its computational simplicity, gives the semivariogram approach an edge over the more flexible cross-validation approach.