Exploring comparative effect heterogeneity with instrumental variables: prehospital intubation and mortality

We highlight the role of local instrumental variable (LIV) methods in exploring treatment effect heterogeneity using an empirical example of evaluating the use versus non-use of prehospital intubation (PHI) in patients with traumatic injury on inpatient mortality. We find evidence that the effect of PHI on inpatient mortality varies over levels of unobserved confounders giving rise to a phenomenon known as essential heterogeneity. Under essential heterogeneity, the traditional instrumental variable (IV) method, when using a continuous IV, estimates an effect that is an arbitrary weighted average of the casual effects for marginal groups of patients whose PHI receipt are directly influenced by the IV levels. Instead, the LIV methods estimate the distribution of treatment effects for every margin that is identified by data and allow for predictable aggregation to recover estimates for meaningful treatment effect parameters such as the Average Treatment Effect (ATE) and the Effect on the Treated (TT). LIV methods also allow exploring heterogeneity in treatment effects over levels of observed confounders. In the PHI analysis, we estimate an ATE of 0.074. We find strong evidence of positive self-selection in practice based on observed and unobserved characteristics, whereby patients who were most likely to be harmed by PHI were also less likely to receive PHI. However, the degree of positive self-selection mitigates in regions with higher rates of PHI use. We also explore factors associated with the prediction of significant harm by PHI. We provide clinical interpretation of results and discuss the importance of these methods in the context of comparative effectiveness research.