In the absence of randomized treatment assignment, comparative effectiveness studies of different health interventions are plagued by selection bias. When multiple treatment options are available, the choice of intervention is often influenced by health system, physician, and patient characteristics that are correlated with health outcomes. For instance, Black patients are less likely to receive major surgical procedures than White patients in the US, 1 and it is well-established that Black race (which is correlated with social and structural determinants of health, including systemic racism) is associated with higher morbidity and mortality. 2 In JAMA Network Open, Howard and colleagues 3 use instrumental variable (IV) analysis as a method to control for selection bias in the comparative effectiveness of 2 bariatric surgery procedures (laparoscopic sleeve gastrectomy and laparoscopic gastric bypass) among Medicare beneficiaries. But, as the authors note, "although effective in their ability to reduce bias, instrumental variables can be difficult to identify."
Key Assumptions of Instrumental Variable AnalysisThe goal is to find an instrument (ie, variable) that randomly predicts whether a person receives one treatment option vs another. The instrument must impact the outcome only through the treatment assignment, and it cannot be related to any patient, physician, or system characteristics that are associated with the outcome. 4 While these IV assumptions are often not empirically testable, researchers implementing the IV method (and readers of IV studies) should think carefully about variables that are potentially correlated with both the IV and the study outcome (ie, IV-outcome confounders), which could bias results and lead to incorrect conclusions. 4 Potential IV-outcome confounders can be identified in the study data, or by searching the literature for variables that are correlated with the instrument and exploring whether those variables are also associated with the study outcome.
Potential Bias in Geographical Variation Instrumental VariablesHoward et al 3 use state variation in bariatric procedure rates in the prior year to project whether a patient seeking care in the study year receives sleeve gastrectomy (vs gastric bypass). Geographical variation in procedure rates is 1 of the 4 most commonly used IVs in comparative effectiveness research, but prior research has shown that this instrument is likely biased by IV-outcome confounders. 4 Geographical variation is strongly associated with variables that are correlated with health outcomes, such as race and ethnicity, urban vs rural status, and access to other health interventions. 4 In fact, Howard et al 3 found that Black patients undergoing bariatric procedures were more likely to live in areas with higher sleeve gastrectomy rates (and were therefore assigned to the sleeve gastrectomy treatment group by the IV) than White patients. A 2019 JAMA Surgery study 5 found that Black patients had higher rates of complications and health care utilization followin...