Tutorial: Strategies addressing detection bias were reviewed and implemented for investigating the statins-diabetes association

Objectives Literature on specific analytic methods for addressing detection bias is fragmented. We illustrate some analytic strategies to account for detection bias. Study Design and Setting Several tools addressing detection bias are described, namely (1) sensitivity analysis, (2) conditioning on outcome detectability, and (3) use of negative controls. These tools are applied in a population-based cohort study on the association between adherence to statins and start of antidiabetic therapy (as proxy of type 2 diabetes mellitus onset). Results Compared with patients on very low adherence to statins, those with high adherence had hazard ratio (HR) for diabetes of 1.53 (95% confidence interval: 1.44, 1.64). The observed association was potentially affected by detection bias because long-term exposure to statins implies a more regular use of primary care services, triggering the search for diabetes. Nevertheless, from the considered tools, (1) we showed that the HR for diabetes risk decreased to 1.28 if diabetes detection was assumed to be 20% more likely in highly adherent patients; (2) an increased risk of diabetes was found among patients with no specialist visits during the first year of follow-up; (3) no association was found between adherence to bisphosphonates (negative exposure) and diabetes nor between adherence to statins and initiation of antihypertensive pharmacotherapy (negative outcome). Conclusion Implementation of analytic strategies for addressing detection bias is advisable whenever this is suspected. As illustrated, several methods could be considered. Their implementation suggested that detection bias had a limited impact in our application.