LAUSR

Health economists are often interested in the effects of provider-level attributes (e.g., nonprofit status or quality rating) on patient outcomes, but estimation is subject to selection bias due to correlation with other omitted provider-level attributes that also affect patient outcomes. Recently, researchers have attempted to use patient-level instrumental variables, such as differential distance, to solve this problem of a provider-level endogenous treatment variable in settings where patients are nested within providers. However, to satisfy validity assumptions, an instrumental variable for a provider attribute must be at the provider level or a larger unit of aggregation, not at the patient level. A patient-level instrument cannot predict variation in a provider attribute separately from other, potentially unmeasured, provider attributes. In this paper, we explain this misapplication, review the extent of this problem in recent literature, and offer alternative approaches to avoid this misapplication of patient-level instrumental variables.