To examine the validity of criteria-based (clinical) diagnosis of Alzheimer's disease (AD), 4 physicians experienced in the evaluation of dementia patients applied 3 sets of diagnostic criteria to each of 62 patients based on standardized medical record information. Diagnostic outcome was validated by neuropathologic examination (completed previously) for all (43) demented patients and 4 nondemented patients and by follow-up in the remainder (15) with no dementia. Raters were blind to the composition of the study group as well as to the clinical and pathologic diagnoses. We evaluated 3 diagnostic criteria sets for AD: the American Psychiatric Association diagnostic criteria from the Diagnostic and Statistical Manual (DSM-III), the NINCDS-ADRDA Work Group criteria for the diagnosis of Alzheimer's disease (NINCDS), and the Eisdorfer and Cohen research diagnostic criteria for primary neuronal degeneration (ECRDC). ECRDC had the highest specificity (0.88) but also the greatest odds of false-negative diagnosis (LRneg = 0.61, sensitivity = 0.46). NINCDS had the best sensitivity (0.92, specificity = 0.65), and DSM-III showed intermediate values (sensitivity = 0.76, specificity = 0.80). We conclude that the investigator or clinician who wishes to ensure that patients classified as AD are more likely to be AD should choose DSM-III, whereas the investigator who wishes to include the greatest number of AD cases, seldom assigning a diagnosis of no AD to a true case, should choose NINCDS.