Challenges and Importance of Finding Hidden Confounders When Conducting Comparative Effectiveness Studies Using Registry Data

In this issue of Circulation , Yeh and coworkers report the results of a study designed to provide insight into the frequency and impact of coronary artery bypass grafting surgery (CABG) surgical turn-down, that is, how often patients with an indication for CABG were declared unsuitable for it by a surgeon, and how that decision affected outcomes1. The analyzed data were extracted from the Partners Long-Term Outcomes Database that aligns its variables and definitions with the National Cardiovascular Data Registry (NCDR) CathPCI Registry and the Society of Thoracic Surgeons (STS) Adult Cardiac Surgery Database. Looking at more than 1000 patients undergoing non-emergent percutaneous coronary intervention (PCI) over a 3 year period at two Boston hospitals, these investigators found that 22% had been deemed ineligible for CABG, usually because of poor bypass target vessels, advanced age or renal insufficiency. CABG ineligible patients were sicker than CABG eligible patients, and they had higher mortality rates: CABG ineligibility was independently associated with a roughly 6-fold increase in adjusted in-hospital death, and about a 3-fold increase in adjusted late death compared with CABG eligible patients. CABG ineligible patients had more complex PCI procedures (more lesions, more stents) but fewer vessels were treated compared with CABG eligible patients despite similar burdens of 3 vessel disease, suggesting the ineligible patients may have had less complete revascularization and hence greater residual ischemic burden after PCI. Inserting CABG ineligibility as an independent term into the NCDR PCI mortality risk model improved the predictive capability of the model significantly. The authors concluded that CABG ineligibility occurs frequently enough to be important to practice statistics, and appears to impact outcomes powerfully, but they note that it has not been incorporated into current risk adjustment models.