A0621
Title: Evaluating instrumental variable assumptions using randomization tests
Authors: Zach Branson - Carnegie Mellon University (United States) [presenting]
Abstract: Instrumental variable (IV) analyses are a common approach for estimating causal effects in observational studies, where units are non-randomized to treatment. Even though treatment is non-randomized, estimating causal effects is tenable if there is an instrument that affects outcomes only through its correlation with covariates. Thus, a fundamental assumption of IV approaches is that the instrument is as-if randomly assigned to units. There are several falsification tests in the literature for this assumption, which compare balance on observed covariates by IV status to balance on observed covariates by treatment status, with the hope that the former is better than the latter. Adding to this literature, we propose a randomization test for this assumption. We use the balance that would have been produced under randomization as a standard by which to compare IV balance and treatment balance. A benefit of our test over other tests is that it can be used to perform a global balance assessment across covariates as well as assessments on individual covariates. Furthermore, these tests can incorporate blocking information if IV assignment depends on blocks. We demonstrate this approach using a recent application using bed availability in the ICU as an instrument for admission to the ICU.
