An efficient design for verifying disease outcome status in large cohorts with rare exposures and low disease rates

Cohort studies require the use of large samples when the risk of the event is very low. Databases that are large and population-based, such as Medicaid files, are frequently used for cohort studies, since they provide access to the large samples required for adequate statistical power at a relatively affordable cost. Epidemiologic studies using these databases typically require verification of reported diagnoses, however, because of the potential for errors in disease reporting. When exposure prevalence is also low, as in many pharmacoepidemiologic investigations of drug toxicity, there are few exposed cases compared to the number of unexposed cases. Verification of all unexposed presumptive cases through medical records is costly. We investigate the statistical efficiency of a design in which all exposed cases but only a subsample of the unexposed cases are verified. We show that good efficiency can usually be achieved with a small subsample of unexposed cases. Published in 1999 by John Wiley & Sons, Ltd.