The multivariate Behrens-Fisher region, recalibrated: honest bootstrap coverage where Hotelling and Welch-James drift (m03cd) -- Reproducibility bundle
Year: 2026
Type: software
Abstract: When two multivariate groups are compared, the fuller summary of how they differ is a confidence region for the mean-difference vector. The textbook region is the Hotelling T-squared ellipsoid built from a pooled covariance, exact under multivariate normality and a common covariance; the standard heteroscedastic repair is a Welch-James ellipsoid that adds the two per-group covariances and refers the quadratic form to a chi-square on p degrees of freedom. This deposit accompanies the paper showing that at the corner applied multivariate work most often produces — groups with unequal covariances, skewed marginals, and small unequal samples — both regions under-cover: at a four-dimensional design where the smaller group carries the larger covariance, with a lognormal parent, the realized coverage of the pooled Hotelling region falls to 0.818 and of the Welch-James chi-square region to 0.764, against a nominal 0.95. The pooled region fails because its equal-covariance premise breaks in the anti-conservative direction of the multivariate Behrens-Fisher problem; the Welch-James region fails further because its chi-square reference is liberal when the covariances are estimated from few observations. The paper gives a region that keeps the Welch-James quadratic form but replaces its critical value with the upper quantile of a studentized bootstrap of that same statistic, each resample recomputing both group means and both group covariances, so the critical value carries the small-sample, skew-driven inflation the chi-square omits. Its worst-case coverage over the same grid is 0.936, and it is near nominal, 0.94 to 0.99, on every cell, where the two classical regions fall to 0.82 and 0.76. The honesty costs region size, largest exactly where the classical regions under-cover, so the recommendation is a routing one: report the Hotelling region as the default and the calibrated region beside it when the groups have visibly unequal covariances and skewed marginals at small unequal n, reading a disagreement as a warning that the reported region is smaller than the data support. The bundle contains the built manuscript (author and anonymized, docx + pdf), the derivations companion (D1-D5, with a resolving verification hook per result), the figure and its generator, and the deterministically-seeded runnerrerun/rc_m03cd_coverage.py (seed 20260826, 500 replications per cell, 399 bootstrap draws, self-contained numpy/scipy) with its locked CSV. Every coverage and size number in the paper regenerates from the deposited code at the recorded seed. New to this record are the diagnosis that both standard mean-difference regions under-cover at the unequal-covariance, skewed, small-unequal-n corner and that the naive Welch-James chi-square region under-covers more than the pooled Hotelling region, together with the studentized-bootstrap-calibrated Welch-James region that restores near-nominal coverage. Code — the .py runner and figure generator — is released under the MIT License (LICENSE-MIT.txt), © 2026 William J. Dwyer. All non-software content — the manuscript, the derivation companion, tables, and rendered figure — is released under the Creative Commons Attribution 4.0 International Public License (CC BY 4.0) (LICENSE-CC-BY-4.0.txt): "You are free to share and adapt the material for any purpose, including commercially, provided appropriate credit is given." Canonical URL: https://creativecommons.org/licenses/by/4.0/legalcode. Attribution: Dwyer, W. J. (2026). The multivariate Behrens-Fisher region, recalibrated: honest bootstrap coverage where Hotelling and Welch-James drift.
Author: Dwyer, William J.
Institution: University of Massachusetts Lowell
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204220069
