Measurement error in biomarkers: sources, assessment, and impact on studies.

Measurement error in a biomarker refers to the error of a biomarker measure applied in a specific way to a specific population, versus the true (etiologic) exposure. In epidemiologic studies, this error includes not only laboratory error, but also errors (variations) introduced during specimen collection and storage, and due to day-to-day, month-to-month, and year-to-year within-subject variability of the biomarker. Validity and reliability studies that aim to assess the degree of biomarker error for use of a specific biomarker in epidemiologic studies must be properly designed to measure all of these sources of error. Validity studies compare the biomarker to be used in an epidemiologic study to a perfect measure in a group of subjects. The parameters used to quantify the error in a binary marker are sensitivity and specificity. For continuous biomarkers, the parameters used are bias (the mean difference between the biomarker and the true exposure) and the validity coefficient (correlation of the biomarker with the true exposure). Often a perfect measure of the exposure is not available, so reliability (repeatability) studies are conducted. These are analysed using kappa for binary biomarkers and the intraclass correlation coefficient for continuous biomarkers. Equations are given which use these parameters from validity or reliability studies to estimate the impact of nondifferential biomarker measurement error on the risk ratio in an epidemiologic study that will use the biomarker. Under nondifferential error, the attenuation of the risk ratio is towards the null and is often quite substantial, even for reasonably accurate biomarker measures. Differential biomarker error between cases and controls can bias the risk ratio in any direction and completely invalidate an epidemiologic study.