Reliability of eight-year diet recall in cancer cases and controls.

We addressed three questions concerning diet recall in a population of 181 incident cancer cases diagnosed between 1976 and 1984 in the Adventist Health Study, and 225 controls randomly selected from the same population after removing cancer cases: (1) Are recalls of past dietary habits reliable? (2) Does recall ability differ between cancer cases and controls? and (3) Are current or retrospectively recalled reports the best estimator of past dietary practices? Three sets of dietary data were compared using a 35-item nonquantitative food frequency questionnaire: initial reports in 1976, recalled reports obtained retrospectively in 1984, and current reports for 1984. Recall ability was evaluated for individual foods and for all foods combined by comparing recall error scores summing the absolute differences between initial and recalled frequencies. Means and medians for all three food groups were similar for cases and controls. The Spearman rank-order correlations between pairs of reports (initial/recalled, initial/current, and recalled/current) averaged 0.48, 0.41, and 0.62, respectively. A crude difference of 2.0 between cases and controls (p less than 0.05) in the recall error score indicated that cases on the average recalled two foods one frequency category closer to the initial estimate compared with controls. The case-control difference decreased to a nonsignificant 0.4 (p = 0.07) in multivariate analysis that conditioned on dietary changes. On the average, recalled reports estimated initial reports one frequency category closer than did current reports for three foods (p less than 0.001), primarily because of changes in dietary habits.