The statistical power of prospective studies on diet in relation to chronic disease risk can be improved by maximizing the variation in true intake levels actually distinguished--or 'predicted'--by dietary questionnaire assessments collected at baseline. This can be achieved by 1) developing a questionnaire method that provides measurements with the smallest possible random errors, thus maximizing the correlation of measured with true habitual intake levels; and 2) increasing the between-person variation in true dietary intake levels when combining multiple cohorts in populations with diverse consumption patterns. The first approach implies that, during the development or selection of the questionnaire method, correlations between measurements and true intake levels can be monitored; the second approach requires adjustment for between-centre differences in over--or underestimation of dietary questionnaire measurements. Besides optimizing the statistical power, it is important that the magnitude of the predicted variation in true intake level is estimated accurately, so as to allow unbiased estimations of relative risks. To meet these various objectives, substudies must be conducted for the 'validation' or 'calibration' of dietary questionnaire assessments, by comparison with additional measurements that have independent sources of error. This paper reviews the methodological considerations underlying the design and implementation of such substudies in the EPIC project, a collaborative multicentre study in nine Western European countries.