pibglynnrueda.pdf | 253 KB |
Abstract:
Post-instrument covariates are often included in IV analyses to address a violation of the exclusion restriction. We demonstrate that even in linear constant-effects models with large samples: 1) invariance between IV estimates (with and without post-instrument covariates) does not imply that the exclusion restriction holds with respect to the post-instrument covariate, 2) OLS with an omitted variable will often have less bias than IV with the post-instrument covariate, 3) measurement error in the post- instrument covariate does not necessarily lead to attenuation, and 4) the bias of OLS and IV are related. Therefore, if used, IV with a post-instrument covariate should always be paired with OLS, and results should be discussed in concert. We illustrate these points with a re-analysis of Acemoglu, Johnson, and Robinson (2001), showing that for the paper’s claims to be valid, at least 35% of the variance in the causal variable must be due to measurement error.