Substantial improvements to the performance of discriminating explosions from earthquakes have been demonstrated by correcting regional seismic phase amplitudes or their ratios (e.g., Pn/Sn and Pn/Lg) for variations in propagation path and station effects. One of the most promising methods for estimating the corrections and their uncertainties is a form of Bayesian kriging. However, there are valid concerns regarding the robustness of the kriged amplitude corrections to aberrant calibration data, that is, that regional phase amplitudes from earthquakes with anomalous mechanisms, depths, and/or paths could significantly bias the amplitude correction grids. The methodology of Bottone et al. (2002) will properly account for such situations, unless the residual variance is correlated. To first demonstrate the problem, the approach of Bottone et al. (2002) is applied to explosions and clusters of earthquakes with distinct focal mechanisms at the Nevada Test Site (NTS). The results indicate that serious discrimination errors (i.e., miscategorizations of explosions as earthquakes) can occur. In the case of highly correlated clusters, the localized reference data are weighted too high in estimating the corrections and the posterior calibration uncertainties become too small. To address this issue, an extension of the kriging methodology, incorporating an additional correlation length to treat correlated focal mechanisms, is presented and tested on the NTS data sets and on large samples of explosions and earthquakes worldwide. Application to the worldwide data sets indicates that the power of the hypothesis test to reject earthquakes as explosions is marginally reduced, without miscategorizing any known explosions and providing a robust treatment of clustered data.