This paper is concerned with overcoming problems caused by measurement error in multi-centre studies of diet and disease. Measurement error causes differential bias, so the information on the diet-disease relationship from different cohorts is not directly comparable. Hence this information needs to be calibrated before the data are combined in an eventual meta-analysis. We consider the design of calibration substudies. We distinguish two forms of information from a mutti-centre cohort study. The first is subject-level information, which comes from the variation of disease rate within cohorts. The second is cohort-level information, which comes from the variation of disease rate between cohorts. Tha requirements of the calibration study are different for these two forms of information. Calibration is carried out by remeasuring diet in a subsample of each cohort using a standardized reference method. This reference measurement should yield unbiased estimates of habitual intake. Using a criterion of efficiency, relative to a perfectly calibrated study, we show that the sample size should be a multiple of the expected number of cases in each cohort. To control for confounding, each cohort should be stratified and the ratio of sample size to number of cases should be constant within strata. Since the required sample size is related not to the size of the cohort, but to the eventual number of cases of disease, calibration samples need not be prohibitively large. They should, however, be concentrated on those parts of the cohort, such as older age groups, which will yield most cases.