External Adjustment Sensitivity Analysis for Unmeasured Confounding: An Application to Coronary Stent Outcomes, Pennsylvania 2004-2008

Background Assessing the real-world comparative effectiveness of common interventions is challenged by unmeasured confounding. Objective To determine whether the mortality benefit shown for drug-eluting stents (DES) over bare metal stents (BMS) in observational studies persists after controls for/tests for confounding. Data Sources/Study Setting Retrospective observational study involving 38,019 patients, 65years or older admitted for an index percutaneous coronary intervention receiving DES or BMS in Pennsylvania in 20042005 followed up for death through 3years. Study Design Analysis was at the patient level. Mortality was analyzed with Cox proportional hazards models allowing for stratification by disease severity or DES use propensity, accounting for clustering of patients. Instrumental variables analysis used lagged physician stent usage to proxy for the focal stent type decision. A method originating in work by Cornfield and others in 1954 and popularized by Greenland in 1996 was used to assess robustness to confounding. Principal Findings DES was associated with a significantly lower adjusted risk of death at 3years in Cox and in instrumented analyses. An implausibly strong hypothetical unobserved confounder would be required to fully explain these results. Conclusions Confounding by indication can bias observational studies. No strong evidence of such selection biases was found in the reduced risk of death among elderly patients receiving DES instead of BMS in a Pennsylvanian state-wide population.