Calibration of seismic nuclear test monitoring stations relies on a diverse reference-event database. The reference-event covariance structure must be characterized and the covariance propagated to subsequent processes and calibration products. In seismic location one of the first steps in the calibration process is declustering, in which closely spaced epicenters are combined to reduce redundant data and random observational errors. We formulate a new declustering procedure that accounts for correlated and uncorrelated components of reference-event error, producing a declustered data set that tracks raw reference-event uncertainties. Declustering behavior is demonstrated using example data sets.