Sample size requirements for calibration studies of dietary intake measurements in prospective cohort investigations.

Pooling data from multiple cohort studies on diet and cancer has the advantage that it allows detailed cross-validation of relative risk estimates between different study populations. If there is reasonable agreement between cohort-specific relative risk estimates, a more powerful pooled summary estimate can be obtained. A complication, however, is that in different cohorts, relative risk estimates may be biased to a different degree as a result of errors in the baseline assessments of habitual dietary intake levels. Such divergent biases can be adjusted for by means of "calibration" studies, using standardized reference measurements obtained in a subgroup of each cohort. These adjustments entail a cost, however, in terms of an increase in the confidence interval of relative risk estimates within each cohort separately. In this paper, the authors evaluate the possible magnitude of such intracohort losses in precision and discuss the approximate sample size required to have a sufficient level of accuracy in dietary calibration studies to adjust for bias.