Partial Identification and Sensitivity Analysis

Abstract

This chapter is concerned with methods of causal inference in the presence of unobserved confounders. Three classes of estimators are discussed, namely, local identification using instrumental variables, sensitivity analysis, and 6 estimation of nonparametric bounds. In each case, the response to the core identification problem is to retreat from the standard focus on point identification of the average treatment effect, yet the three approaches characteristically differ in terms of alternative quantities of interest that are considered empirically estimable under more restrictive circumstances. The chapter develops the basic principles underlying the three classes of partial identification estimators and illustrates their empirical application with an analysis of earnings returns to education.