Though common in the applied literature, it is known that fixed effects regressions with a constant treatment effect generally do not consistently estimate the sample-weighted treatment effect. This paper demonstrates the extent of the difference between the fixed effect estimate and the sample-weighted effect by replicating nine influential papers from the American Economic Review. We propose a model with fixed effects interactions to identify the sample-weighted treatment effect and derive a test that discriminates between this estimate and the standard fixed effects estimate. For all 9 papers in our replication, at least one set of fixed effects interactions is jointly significant; in 6 of 9 papers, there is a sample-weighted estimate that is statistically different from the standard fixed effects estimate. In 7 of 9 papers, the differences are economically significant (larger than 10%); the average of the largest difference between the estimators from each paper is over 50% and the median is 19.5%. Our procedure does not markedly increase the variance of the estimators in 7 of 9 papers.
- Fixed effects,
- average treatment effect,
- treatment effect heterogeneity
Available at: http://works.bepress.com/cgibbons/1/