Event

Large non-experimental studies are increasingly used to evaluate the benefits and harms of medical interventions. One of the principal challenges in such studies is confounding - systematic differences between patients exposed to an intervention of interest versus the chosen comparator.

Recently, instrumental variable approaches have been proposed that use variation in treatment preference between providers to estimate treatment effects. Under certain assumptions, these methods allow unbiased estimation of treatment effects even if important confounders are unmeasured.

If the healthcare provider fails to satisfy assumptions of an instrumental variable, it should instead be treated as a potential confounding factor. I will illustrate how the results of non-experimental studies can depend strongly on how one handles the healthcare provider in the analysis.

Although the property of being a confounder or an instrumental variable is not statistically testable, I will provide some ideas about how one can determine the appropriate use of the healthcare provider in non-­experimental studies.