Background: Recently, there has been increased interest in leveraging observational studies for comparative effectiveness research. Without robust and valid risk adjustment, however, findings from these nonrandomized studies may remain biased. Previous studies examining long-term mortality with drug-eluting stents (DESs) have demonstrated discordant results between randomized trials and observational studies. To examine the impact of treatment selection bias on these findings, we used data from a prospective percutaneous coronary intervention (PCI) registry (EVENT [Evaluation of Drug Eluting Stents and Ischemic Events]) to compare clinical outcomes between DESs and bare metal stents (BMSs) using conventional (multivariable regression and propensity matching) and novel (instrumental variable analysis) risk-adjustment techniques.
Methods and results: The study population consisted of 9266 patients who underwent nonemergent PCI with stent placement at 55 US centers between 2004 and 2007. All-cause mortality and target lesion revascularization (TLR) were assessed prospectively over 1 year of follow-up. Overall, 8171 patients (88%) received DES, but this proportion substantially differed by treatment year (93% in 2004-2006 and 73% in 2007; P<0.001). One-year rates of death and TLR were significantly lower with DES versus BMS (death, 2.5% versus 5.6%; TLR, 4.2% versus 6.9%; P<0.001 for both), findings that persisted in both multivariable-adjusted and propensity-matched analyses. In contrast, instrumental variable analysis, using enrollment period (2004-2006 versus 2007) as the instrument, demonstrated no significant difference in 1-year mortality (predicted absolute difference, 2.0%; 95% CI, -1.8% to 5.7%; P=0.30) and a strong trend toward reduced TLR with DES use (predicted absolute difference, -4.2%; 95% CI, -8.8% to 0.4%; P=0.07).
Conclusions: Among unselected PCI patients in contemporary practice, DES use tended to be associated with a consistent reduction in TLR regardless of risk-adjustment method but showed discordant effects on mortality with conventional risk adjustment compared with instrumentable variable analysis. These findings underscore the limitations of standard risk-adjustment methods to adequately address treatment selection bias in nonrandomized studies and have important implications for comparative effectiveness research using observational data.