Reproducibility of a food frequency questionnaire and stability of dietary habits determined from five annually repeated measurements

Abstract

Objectives: In studies on diet and cancer, diet assessment should address long-term intake. Therefore, the authors determined the 5-year reproducibility of a self-administered 150-item food frequency questionnaire (FFQ) used to assess dietary habits in the Netherlands Cohort Study on diet and cancer. Since the FFQ was repeated more than once, the pure test-retest reliability of the FFQ could be distinguished from the intra-individual change in nutrient intake over time. These results were furthermore used to investigate the measurement error structure of the FFQ.

Design: After baseline administration in 1986, the FFQ was annually repeated from 1987 to 1991 in independent random samples of the cohort (n = 400). Pearson correlation coefficients (r) between baseline and repeated measurements of nutrient intake, calculated for each time interval, were regressed on time interval to provide separate estimates of the test-retest correlation (intercept of regression line) and of the decline in correlation over time (slope). The proportion of correlated measurement error was derived from combining the test-retest results with those from a validation study, in which the FFQ was validated against three 3-day diet records.

Results and conclusions: Response was stable at 82%. The test-retest r ranged from 0.42 for selenium intake to 0.90 for alcohol intake. The slopes of the regression lines were relatively flat, but negative for most nutrients; on average, the decline in r amounted to 0.07 after 5 years, indicating that the potential of a single FFQ measurement to rank subjects according to nutrient intake dropped only slightly over time. This is important for studies on cancer since a long induction period may be involved. It was furthermore shown that the proportion of within-subject (error) variance of the FFQ method that could be attributed to correlated error ranged from 0 to 50%. This finding confirms that a reliability study may underestimate the measurement error of a method.

Sponsorship: Dutch Cancer Society (grants CIVO 86-1 and CIVO 90-3).