A1051
Title: Inference after multiple hypothesis testing
Authors: Andreas Dzemski - University of Gothenburg (Sweden) [presenting]
Wenjie Wang - Hiroshima University (Japan)
Ryo Okui - University of Tokyo (Japan)
Abstract: Some empirical studies estimate multiple treatment effects corresponding to different sub-populations, different treatments or different outcome variables and focus on interpreting the results of the significant specifications. We develop new estimators that are unbiased under this kind of data-driven model selection along with corresponding valid confidence intervals. Our framework admits a large class of rules for determining significance, including many step-down and step-up methods. In an empirical application, we compare our estimator to a naive estimator that does not account for the selection step. For effects that are significant but close to the threshold of insignificance, our approach detects large selection bias and produces estimates that are very different from the estimates obtained by a naive approach.
