Abstract:
This article derives an equation characterizing the difference between OLS and IV coefficients under potentially heterogenous treatment effects. This leads to what I call the Coefficient Difference Check, which consists of checking that the difference between the estimated OLS and IV coefficients has the same sign as the expected selection effect. I show failures of this check can arise because: IV is invalid, the expected selection story is incorrect, or there are particular heterogenous treatment effects that imply the IV estimate is both "fragile" and that it provides a more biased estimate of the ATT than OLS. Failures of this check are relatively common in the literature. I describe best practices given such failures.