PhD Candidate University of Illinois Chicago Chicago, United States
Background: In pharmacoepidemiologic studies using databases, outcomes of interest are commonly ascertained using an algorithm, which can be multi-component. When the components occur at different follow-up times, an inappropriate choice of the event date may result in misclassification of at-risk time and lead to bias.
Objectives: To examine the potential bias on the targeted treatment effect arising from inappropriate anchor of the outcome event date.
Methods: Using data from Merative MarketScan databases, we estimated the treatment effects in 2 empirical examples where outcomes are believed not to be associated with the treatments based on existing literature: 1) risk of Crohn’s disease (CD), and 2) completion of 2-dose recombinant zoster vaccination series (RZV) associated with first-line antihypertensive treatment (hydrochlorothiazide [HCTZ] vs. angiotensin-converting enzyme inhibitor [ACEI] monotherapy). Both outcomes were identified using 2-component claims-based algorithms that were previously validated or commonly used. We estimated the intention-to-treat effects using appropriate and inappropriate anchors of the outcome event date and compared the results to each other and the expected null effect. We considered the first outcome component date in the CD example and the second component date in the RZV example as the appropriate anchor. Sensitivity analyses included varying outcome time intervals.
Results: We identified 1,410,378 HCTZ and 2,512,429 ACEI new users in the CD example, and 223,085 HCTZ and 292,266 ACEI new users in the RZV example. The incidence rates were low using both appropriate and inappropriate anchors (CD: HCTZ vs. ACEI: ~30 and 28 per 100,000 person-years; RZV: HCTZ vs. ACEI: ~5,200 and 5,000 per 100,000 person-years). In the CD example, both anchors yielded identical results that were consistent with the expected null effect (adjusted hazard ratio [HR]: 0.97, 95% CI: 0.90-1.06). In the RZV example, the appropriate and inappropriate anchors yielded similar results (HR, 95% CI: appropriate anchor, 0.94, 0.92-0.96; inappropriate anchor, 0.93, 0.91-0.95), which both deviated slightly from the expected null effect, likely attributable to residual confounding by race/ethnicity, that is unavailable in the database. Similar trends were observed when using different outcome time intervals in both examples.
Conclusions: Inappropriate anchor of outcome event date does not seem to adversely impact treatment estimates when the outcomes of interest are rare. The magnitude of impact remains unclear in the setting with common outcomes, which warrants further investigation.
