Paper 4, Study 2 — predeclared plan v1.0: a model-diagnostic guard calibrated at the actual sample size (Behrens–Fisher programme)
Prospective registration of the predeclared plan (v1.0, frozen 16 Nov 2026) for Study 2 of Paper 4 in a programme on the direction of small-sample interval failure for a difference in arithmetic means. Study 1 of Paper 4 registered a guard that switches from a lognormal or skew-normal model to the gamma model when the model's posterior mean departs from the sample mean by more than two standard errors. Study 2 tests a guard whose reference scale is estimated at the actual sample size: the gap between the model's posterior median of the mean and the sample mean is studentised by its jackknife standard error, computed from exact leave-one-out posteriors on the grid, and compared with the t quantile. The study uses the first 4,000 datasets of each of Study 1's 156 settings, regenerated exactly from the registered seeds and checked against the archived Welch indicators. Hypotheses G1–G3 concern the skew-normal guard's false-alarm rate where the model's first-order bias is negligible, its discrimination where the bias is substantial, and the safety of the guarded interval relative to the registered guard; the lognormal jackknife guard and a parametric-bootstrap guard are reported as exploratory. The package contains the frozen plan and its SHA-256 fingerprint, the runner, the frozen confirmatory analysis, the frozen inputs, and the pre-freeze validation on non-study seeds that informed the design (disclosed in the plan, §9a). No Study 2 computation on study data, and no Study 1 analysis, had been run at the time of deposit. AI-use note (paste, or use the programme's final wording): The plan, code and checks were developed with Claude (Anthropic) as a computational assistant under the author's direction; the author approved the design decisions and hypotheses and is responsible for the registration.
Authors
- Johan G.L. Verheyden (ORCID: https://orcid.org/0000-0002-6129-2794)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24292182
- Primary Topic
- Statistical Methods and Bayesian Inference
- Type
- article
- Field-Weighted Citation Impact
- 0.00