The application of different correlation coefficients to assess the reproducibility of a food frequency questionnaire

We computed four different correlation coefficients to investigate the degree of reproducibility of the weekly consumption of 77 food items or groups of foods and of seven summary questions from a food frequency questionnaire developed in Italy for a case-control study on cancers of the breast and digestive tract. The questionnaire had been administered twice to 452 Italian men and women. These included Pearson correlation coefficients (a) using the weekly frequencies of consumption without any transformation (P1); (b) after applying the transformation log (x + 1) (P2); (c) after applying the transformation log (x + 0.01) (P3); and (d) the Spearman correlation coefficient (SP). The mean values were 0.55 for P1, 0.59 for P2, 0.56 for P3 and 0.59 for SP. All coefficients were positively correlated, although to variable extents: the Spearman correlation coefficient between P2 and SP was 0.92, and that between P1 and P3 was 0.53. Differences between the four coefficients were more marked for food items with a lower kappa statistic and lower intraclass correlation, ie for those items with more severe reproducibility problems. Thus, a single correlation coefficient may not be enough to detect zones of the distribution of a food item where misclassification problems are more severe. The correlation coefficients used to investigate reproducibility should therefore be chosen on the basis of subsequent data analyses.