In a preliminary effort to improve the early diagnosis of dementia, we developed a regression-based method for estimating premorbid intelligence measured by the ability to read irregular words from the American version of the Nelson Adult Reading Test (AMNART). Using errors on the AMNART and years of education, a model for predicting current verbal intelligence (VIQ) was developed in a sample of nondemented elderly. Double cross validation showed that the model had high accuracy and stability in estimating current VIQ in nondemented subjects. The model was then used to estimate premorbid VIQ in mildly demented subjects. Estimated premorbid IQ exceeded current IQ by at least 10 points and did not differ from that of nondemented subjects. Less than 10% of nondemented elderly had discrepancies that were as large. If intellectual decline predicts future functional loss and can be reliably measured using cross-sectional data, the requirement of functional impairment may be an unnecessary barrier to the early diagnosis of dementia.