Each of 430 subjects received a diagnosis using two independent methods: a test-based quantitative paradigm and a semistructured neurological examination by a physician. The paradigm diagnosis was based on a battery of tests that assessed short- and long-term verbal memory and short-term nonverbal memory, orientation, construction, abstract reasoning, and language. The subjects came from a community in Manhattan County, in New York City, and were characterized by diversity with respect to both ethnicity (29.1% black, 33.4% Hispanic) and educational level (23.5% with 6 or fewer years of education, 25.6% college educated). Based on the paradigm, 10.5% of subjects received diagnoses of dementia, 29.1% of cognitive impairment, and 60.5% of normal. Based on the physician's diagnosis, 9.8% were demented, 21.6% cognitively impaired, and 68.6% normal. There was agreement between the two diagnostic methods for 71.8% of subjects. Diagnostic disagreement (n = 121) was in most cases between normal and cognitively impaired (71.0%) or between cognitively impaired and demented (21.5%). There were only nine cases (7.5%) in which a subject was judged demented by one method and normal by the other. The reliability of each method with respect to the other was moderate (intraclass correlation coefficient, .62), while the reliability of a composite diagnosis based on both methods was much higher (.77). The paradigm was more likely than the physician to give the diagnosis of dementia to patients with low educational levels. The physician's diagnosis was strongly influenced by measures of functioning and by the mental status test administered in the semistructured neurological examination. Race and diagnosis were not related when the effect of education was controlled.(ABSTRACT TRUNCATED AT 250 WORDS)