Selection of controls in database case-control studies: glucocorticoids and the risk of glaucoma.

In case-control studies conducted using computerized databases, controls are often selected as a random sample from the base population. This representative choice of controls is intended to guard against selection bias. We show, using data from a database case-control study, that such a definition of controls may also lead to selection bias under two conditions: (1) if the target disease has a prolonged asymptomatic clinical course with its detection depending on a specific physical examination and (2) if exposed patients have a higher likelihood of having the disease detected than unexposed patients. The extent of the bias that could result from the use of randomly selected controls was investigated in the context of a case-control study of the risk of ocular hypertension or glaucoma associated with the use of glucocorticoids, conducted using the Quebec universal health insurance computerized databases. This article also illustrates that a computerized database can be useful to empirically explore opportunities for bias.