Controlling the Danger of False Discoveries in Estimating Multiple Treatment Effects

I expose the risk of false discoveries in the context of multiple treatment effects. A false discovery is a nonexistent effect that is falsely labeled as statistically significant by its individual t-value. Labeling nonexistent effects as statistically significant has wide-ranging academic and policy-related implications, like costly false conclusions from policy evaluations. I reexamine an empirical labor market model by using state-of-the art multiple testing methods and I provide simulation evidence. By merely using individual t-values at conventional significance levels, the risk of labeling probably nonexistent treatment effects as statistically significant is unacceptably high. Individual t-values even label a number of treatment effects as significant, whereas multiple testing indicates false discoveries in these cases. Tests of a joint null hypothesis such as the well-known F-test control the risk of false discoveries only to a limited extent and do not optimally allow for rejecting individual hypotheses. Multiple testing methods control the risk of false discoveries in general while allowing for individual decisions in the sense of rejecting individual hypotheses