Food frequency reports in 1967-1969 were compared to frequency reports of the same foods asked retrospectively in 1982-1983 and 1967-1969 for 1184 respondents aged 45-64 years in the Tecumseh Community Health Study. The kappa statistic for concordance of the retrospective and baseline reports was used as a summary measure of the individual's ability to reproduce his or her earlier diet report. Reproducibility was estimated for total diet, represented by 83 foods, and for 9 subsets of foods of epidemiologic interest. In bivariate and multivariate analyses, reproducibility was strongly related to stability of diet; those whose diets changed least over the 15-year period had greatest diet reproducibility. Greater total diet reproducibility was also found among men with higher education, among women of less than 110% desirable weight reporting no special diet and among women reporting no medications. Consistent with current models of memory, the retrospective report of diet was strongly related to the current report of diet. Agreement between the retrospective and baseline diet reports was greater than agreement between the current and baseline diet reports. This indicates that, as a proxy for past diet, the retrospective report of diet is superior to the current report. Similar relationships were found for the 9 subset of foods.