The use of food frequency questionnaires for measuring dietary intake has become widespread in epidemiologic studies. It has been suggested that inquiring about a person's usual serving size of each food, in addition to the frequency of consumption, will improve the accuracy of this method. This approach implies that individuals characteristically eat a specific amount of any particular food, and that this amount can be reported with reasonable accuracy. To investigate the variability of portion sizes, the authors analyzed data for 68 commonly consumed foods, based on four one-week weighed diet histories recorded by 194 Boston-area women aged 34-59 years during 1980 and 1981. For each food, total population variance in portion size was partitioned into within-person (intraindividual) and between-person (interindividual) components. For all but seven food items (yogurt, liver, mixed vegetables, watermelon, pancakes/waffles, cold cereal, and cooked cereal) the within-person variance in portion size exceeded the between-person variance. The mean of the within-person to between-person variance ratios, after exclusion of two outlying foods, was 3.4 for untransformed portion sizes, and 3.2 after portion sizes were loge-transformed. Foods with a high within-person variance also tended to have a high between-person variance. The dominance of within-person variance in portion sizes suggests that the concept of usual portion size is complex, and that subjects may experience substantial difficulty in specifying their "usual" portion size. The smaller contribution of between-person variance to the total variance in portion size suggests that specification of a standard portion size by the investigator may not introduce a large error in the estimation of food and nutrient intake.