Nelson's (1976) modified version of the Wisconsin Card Sorting Test (mWCST) was administered to 23 patients with mild dementia of the Alzheimer type (DAT), 33 moderate DAT patients, 31 severe DAT patients, and 75 demographically matched normal control (NC) subjects. DAT patients attained fewer categories and committed significantly more perseverative errors than NC subjects. None of the DAT patient subgroups differed in the number of perseverative errors committed, but the severe DAT patients attained significantly fewer categories than either the mild or moderate DAT patients. Receiver operating characteristic [ROC] curves were plotted to determine which among several of the measures provided the best sensitivity (i.e., correctly identified patients) and specificity (i.e., correctly identified normal controls). The number of categories attained was found to be the best discriminative measure for the entire sample of DAT patients, whereas the number of perseverative errors provided the best discriminability in mild DAT patients. Findings suggest that Nelson's original modification of the WCST is sensitive in the early detection of DAT.