Researchers Propose Two-Team Cross-Screening To Assess Replicability

Roy et al. (2025) presented at the Harvard Data Science Initiative propose a two-team cross-screening design that nonrandomly splits observational data and assigns separate discovery and confirmation teams to assess replicability across distinct subgroups. They apply the approach to the Wisconsin Longitudinal Study examining unwanted pregnancy effects and argue nonrandom splitting (e.g., Catholic versus non-Catholic women) strengthens robustness checks against unmeasured confounding and multiplicity.