Choosing Methods to Minimize Confounding in Observational Studies

The comparison of alternative treatments has long been a primary objective of researchers engaged in clinical investigation. However, with the commitment of $1.1 billion in support of clinical research in the American Recovery and Reinvestment Act,1 followed subsequently by the creation of the Patient Centered Outcomes Research Institute (PCORI) as part of the Affordable Care Act,2 “comparative effectiveness research” has grown to become a publicly discussed national priority.Article see p 587Such comparisons have traditionally relied heavily on the randomized, clinical trial—the gold standard for comparing two treatments. However, because of limitations in generalizability, challenges with feasibility, and typically high costs of executing such trials, coupled with the increasing availability of prospectively captured clinical and administrative data, comparative effectiveness studies using observational study designs have offered an attractive alternative, harnessing the ability of various statistical approaches to overcome nonrandomized treatment allocation. It is in this context that Venkitachalam and colleagues provide a cautionary example illustrating the challenges and limitations of these approaches. Mortality Benefit and Drug-Eluting Stents: Fact or Artifact?The authors are motivated by a consistent discrepancy seen in results of randomized trials and observational studies comparing drug-eluting stents (DES) with bare metal stents (BMS) for percutaneous coronary intervention.3 Observational studies comparing DES and BMS have typically shown a mortality difference favoring DES, findings not corroborated in randomized trials. The authors hypothesize that the discrepancy may be rooted in the incomplete adjustment for unmeasured variables that compel physicians to select a certain stent type but may also be associated with mortality—a phenomenon referred to variously as treatment selection bias or confounding by indication.4,5 They recapitulate a new set of DES versus BMS comparisons within the EVENT Registry, by first applying commonly used multivariable regression, then propensity-score matching, and finally, instrumental variable analysis. Similar to what …