Economics Seminars: Andreas Dzemski (Gothenburg)
Andreas Dzemski (Gothenburg)
1B6, Priory Road Complex
Title: Inference After Multiple Hypothesis Testing
Abstract: In empirical research, it is often desirable to estimate an array of treatment effects and then discuss the estimated effect sizes of those effects that are found to be significant. The raw estimates of effect size do not take into account the correlation between effect size and finding significance. Effects that are overestimated are more likely to be significant ("winner's curse" of empirical research). We consider selective inference on the effect size of the significant treatment effects when joint significance of the effects is determined by a step-down rule. Our approach conditions on finding significance. We provide conditionally median-unbiased estimators of effect size and construct conditional confidence intervals.