Do t -Statistic Hurdles Need to Be Raised?

This article investigates the empirical justification for raising statistical significance thresholds, or "t-hurdles," to reduce false discoveries in academic research, focusing on the cross-sectional stock return predictability literature. It demonstrates that due to publication bias—where results failing to meet current significance levels are often unreported—the key parameter governing the need to raise t-hurdles (the share of false factors) is weakly identified, leading to substantial uncertainty about whether these thresholds should be increased, maintained, or lowered. In contrast, alternative multiple testing statistics that focus on published findings, such as empirical Bayes shrinkage and the false discovery rate (FDR), are strongly identified and suggest that published results are at least 75% true with high confidence. The study employs a structural model estimated via quasi-maximum likelihood on a comprehensive dataset of published predictors, confirming robustness across various assumptions and highlighting that debates on scientific credibility should prioritize these more strongly identified statistics rather than solely focusing on raising t-hurdles.