Abstract

The field of dietary assessment provides numerous statistical challenges. The great majority of large prospective studies of diet and disease have been using a food frequency questionnaire (FFQ) to assess usual (i.e., long-term average) dietary intake. It is now well appreciated that such questionnaires involve substantial measurement error, random and systematic, leading to distorted effects of diet and incorrect statistical tests of diet-disease relationships. To correct for this error, many cohorts include calibration sub-studies in which more precise short-term dietary instruments, such as biomarkers, multiple 24-hour dietary recalls (24HRs), or food records, are administered as reference instruments. Thus, usual dietary intake is assessed by those instruments with considerable measurement error. A popular method for correcting for measurement error of FFQ -reported intakes, regression calibration, has been developed under the assumption that although reference observations may contain within-person random error, they provide unbiased measurements. Application of this method to foods that are not consumed every day by everyone (episodically consumed foods) is problematic, since short-term reference instruments usually include a substantial proportion of subjects with zero intakes, leading to observations with excess zeros. In addition, it is often preferable to analyze the amounts of dietary components relative to the amount of total energy intake to take into account dietary composition. We present the recently developed bivariate model for short-term reference observations on food and total energy intakes. The new model allows a rigorous application of the regression calibration to the analysis of energy-adjusted associations between foods/nutrients and disease using repeat unbiased short-term reference measurements in a calibration sub-study. We exemplify the method by applying it to data from the US NIH -AARP Diet and Health Study. We also use simulations to compare the newly developed methodology with more conventional applications of regression calibration that do not take excess zeros in reference measurements into account.