Nutrient intakes from a food frequency questionnaire are usually calculated as the product of frequency of intake and nutrient composition of the food, summed over the food items. This involves assumptions about the accuracy of recording, food composition data, stability during storage and preparation, and bioavailability. This usual method of calculation was compared with one using empirical weights derived by multivariate linear regression. Food intakes reported on a food frequency questionnaire by Boston, Massachusetts, area subjects in 1982-1985 were used to predict plasma levels of beta-carotene among 370 male and female nonsmokers and plasma levels of alpha-tocopherol among 339 male and female nonusers of vitamin supplements. Nutrient intake computed using empirical weights yielded a significant correlation with plasma beta-carotene (r = 0.43, p = 0.0001), similar to the correlation using nutrient intake calculated from food composition tables (r = 0.38, p = 0.0001). However, the use of empirical weights significantly improved the correlation of vitamin E intake with plasma alpha-tocopherol levels (r = 0.32, p = 0.0001), compared with the weak correlation obtained using the food composition table method to calculate intake (r = 0.16). The results support the validity of published food composition data used to compute carotenoid intake and illustrate the potential utility of empirically derived weights for foods to predict plasma levels of some nutrients.