We determined whether a battery of neuropsychological tests could predict who would develop Alzheimer's disease (AD) in a group of 123 memory-impaired nondemented patients. Patients were followed longitudinally for 2 years with a research battery of neuropsychological tests. After 2 years, 29 developed probable AD, and 94 did not develop dementia. We used logistic regression analyses to examine the classification accuracy of subjects' performance at entry to the study on the research battery. The logistic regression model was significant with an accuracy of 89%, sensitivity of 76%, and specificity of 94%. Two tests contributed significantly to this model: the delayed recall from the Rey Auditory Verbal Learning Test and the Mental Control subtest of the Wechsler Memory Scale. These two tests alone produced the same accuracy, sensitivity, and specificity as the larger model. These results demonstrate that probable AD can be predicted with a high degree of accuracy and with a relatively brief battery of neuropsychological tests.