Although dietary exposures in the distant past are considered important in the etiology of several diseases, few studies have addressed methodological aspects of long-term (> or = 20 yrs) recall. We evaluated the reliability of retrospective self-reports about diet 20 years before the interview and consistency (between siblings) of self-reports about diet during adolescence in a population-based case-control study of stomach cancer in Sweden. Short-term reliability (a questionnaire self-administered 9-12 mos after the personal interview) of reports on diet 20 years ago by 374 control subjects showed mean correlation coefficients for 42 foods/beverages of 0.41 (Pearson) and 0.46 (Spearman) and a mean weighted kappa statistic of 0.42; for 15 nutrients/food constituents the corresponding mean values were 0.46, 0.47, and 0.42, respectively. Consistency of independent reports by siblings about their own diet during adolescence studied in 201 control-sibling pairs was modest. The mean Pearson correlation coefficient for 33 foods/beverages was 0.29, and the mean weighted kappa statistic was 0.30; at the nutrient level the means were 0.26 and 0.24, respectively. A comparison of intersibling differences between controls and stomach cancer cases revealed correlations of a similar magnitude (mean Pearson correlation for 33 foods = 0.29 for control-sibling and 0.27 for case-sibling pairs), thus contradicting differential recall. Our results imply that although reliability of self-reports about the diet in the distant past is generally lower than for the actual diet, we can use these measurements when remote time periods are of special interest in etiological epidemiologic studies.