Misclassification and selection bias in case-control studies using an automated database.

Abstract

Automated databases are increasingly being used for pharmacoepidemiological research and fieldwork is often carried out to supplement and validate information held within them. In MEMO's case-control studies, patients are identified using computerized ICD9 diagnosis codes, the original medical records are retrieved and checked, and only patients fulfilling case inclusion criteria are used. The 20-30% of patients for whom medical records cannot be found are usually excluded. The aim is to eliminate misclassification bias. However, selection bias may be introduced if availability of medical records is associated with exposure. This investigation was therefore carried out to assess the relative importance of misclassification and selection bias. Data from four previous case-control studies, investigating the associations between NSAIDs and hospitalization for colitis, acute renal failure, appendicitis and colorectal cancer, were used. To assess misclassification, odds ratios (with 95% CI) for recent exposure to NSAIDs were compared in repeated studies that used all patients identified by ICD9 codes, with studies using validated cases only. Selection bias was assessed by comparing results in studies using patients for whom records could and could not be found. Results were plotted and the graphs indicated that misclassification bias was relatively unimportant, but that selection bias could be introduced into a study in this way.