Adjusting for unmeasured confounders in pharmacoepidemiologic claims data using external information: the example of...

Large health care utilization datasets are frequently used to analyze the incidence of rare adverse events from medications. However, possible confounders are typically not measured in such datasets. We show how to assess the impact of confounding by factors not measured in Medicare claims data in a study of the association between selective COX2 inhibitors and acute myocardial infarction (MI).

Using the Medicare Current Beneficiary Survey, we assessed the association between use of selective COX2 inhibitors and 5 potential confounders not measured in Medicare claims data: body-mass index, aspirin use, smoking, income, and educational attainment. For 8,785 participants > or =65 years, we estimated the prevalence of selective COX2 inhibitor use and also of each confounder, as well as the association between drug exposure and confounders. Estimates of the confounder-disease associations from the medical literature were used to calculate the extent of residual confounding bias for each potential confounder.

Selective COX2 inhibitor users were less likely to be smokers (8% versus 10%) than nonselective NSAID users, while the prevalence of obesity was comparable (24%). Aspirin use was also balanced among all drug exposure categories. Failure to adjust for 5 potential confounders led to a small underestimation of the association between selective COX2 inhibitors and MI; comparing selective COX2 inhibitors with NSAIDs, the net bias was estimated to be -1.0% of the unknown true effect size (maximum range: -6% to 0%).

In this example of the relationship between selective COX2 inhibitors and MI, not adjusting for 5 potential confounders in Medicare claims data analyses tended to slightly underestimate the association, but is unlikely to cause important bias.