Primary Submission Category: Heterogeneous Treatment Effects
Conditionally Normalized Two-Way Fixed Effects Estimation with Sub-Gaussian Rates
Authors: Adam Bouyamourn,
Presenting Author: Adam Bouyamourn*
I describe TWFE estimation as a nuisance parameter estimation problem, and describe heterogeneity in terms of a location-scale approximation that shows simply why TWFE estimation breaks down in the context of heterogeneity or imbalanced panel designs. This suggests a general approach to resolving the problem — conditional Studentization in the directions of the nuisance dimensions, which yields estimators that are ancillary for nuisance parameters, and hence consistent for the Average Treatment Effect. I show that this approach is applicable to a wide variety of designs. I explain why these estimators resolve the so-called `forbidden comparisons’ problem, and why the problem of `negative weights’ is not, in fact, a problem. I then propose and evaluate the performance of three conditionally Studentized estimators consistent for the ATE with sub-Gaussian rates: trimmed means, median of means, and an Edgeworth expansion of the sample mean. I assess their empirical performance through simulation and the reanalysis of two papers.