Abstract:
|
Empirical analyses in economics focusing on estimating policy effects typically rely on regression approaches that include a variety of control variables. However, applied researchers rarely know the exact identities of variables that should be controlled for in their empirical models before seeing the data and likely engage in some model search in presenting their results. We illustrate how some intuitive model selection procedures may result in undesirable properties in estimation and inference about treatment parameters of interest even in settings with randomly assigned treatments. We then adapt approaches developed in econometrics and statistics for doing inference in high-dimensional models and show how these approaches may be employed for inference in the canonical economic problems of estimating treatment effects in randomized control trials, in group randomized trials, and in observational studies. We illustrate the use of the methods through simulation and several empirical examples.
|