Just How Sensitive are Instrumental Variable Estimates

Abstract

Researchers regularly use instrumental variables to resolve concerns about regressor endogeneity. The existing literature has correctly emphasized that the choice of instrumental variables matter for the resulting estimates. This paper shows that researchers should also be concerned that the functional form of the instrument matters as well for the resulting estimates. For example, simply changing an instrumental variable from the level to the logarithm can change estimates directly. This article documents the problem, suggests why the problem occurs and suggests different approaches to the problem.