SOURCES OF VARIANCE IN 24-HOUR DIETARY RECALL DATA - IMPLICATIONS FOR NUTRITION STUDY DESIGN AND INTERPRETATION

The National Heart Lung Blood Institute Nutrition Data System was examined with regard to sources of variance. A Graeco-Latin square design involving 30 male and 30 female subjects, each interviewed six times, three trained interviewers and three interview days was used. Through factorial analysis demonstrated that there was no difference between interviews and no training effect associated with the sequence of interviews. Conversely, there was a strong sex difference in absolute nutrient intake, and in females, but not males, a strong day of the week effect. Both of these disappeared when nutrient concentrations (in proportion to energy) were examined. In a direct comparison of centrally computed and hand calculated nutrient vectors from 60 recalls, it was found that data handling contributed a small component of variance for most nutrients studied. This variance was larger in the case of the polyunsaturated fatty acids and the P:S ratio. When the ratio of interindividual (true between subject variation in 'usual' intake) coefficient of variation over intraindividual (day by day variation in intake, day of the week variation in females, any methodological errors) coefficient of variation was calculated, it was found to be between 1 and 1.2 for absolute intakes of energy, protein, total carbohydrate, total fat, and monounsaturated fatty acids, and to rise to 1.7 to 2.1 for polyunsaturated fatty acids and cholesterol. When the nutrients were expressed in proportion to energy, total variance tended to fall, interindividual variance fell and intraindividual variance was affected to a much lesser degree. The ratio rose sharply. The presence of intraindividual variation will bias estimates of correlation coefficients and of regression slopes toward 0 and will result in misclassificationn of subjects into ranges of usual dietary intakes. Erroneous interpretations may be placed on negative findings in epidemiological studies. This problem is not eliminated by increasing group size. Knowledge of the partitioning of variance may be of considerable benefit in examining relationships between variables as in diet-lipid-heart disease studies and in the design of future studies.