Papers · William Alexander Aldred

The Estimator Is a Choice: How Many of the Earlier Papers' Verdicts Survive a Change of Convention

Publicly deposited November 1, 2026 · Zenodo · 6 pp · DOI: 10.5281/zenodo.23865043

Abstract

Every number the earlier papers published passed through a statistician's choice: fit the scale in log space or in linear space, quote the scatter as a population deviation or a robust one, regress with an errors-in-variables estimator or with one of the two ordinary least squares. The choices were made for stated reasons and fixed in the written registration before the data were examined. This paper asks what they are worth. It takes the eight forms of the first round and their three tests, holds the data and the method fixed, and changes one convention at a time. The count: the scatter test is robust (no verdict moves under the RMS or sample-deviation changes; exactly one moves under a robust MAD scatter, the marginal form, in the forms' favor). The scale-agreement test is fragile, against the forms: fit the scale in linear space instead of log space and six of eight forms lose a pass they had. The slope test's verdict is robust to ordinary least squares and dead under inverse least squares: regress the wrong way round and seven of eight slope exclusions evaporate, the flagship McGaugh form landing at 3.986 with its interval straddling 4, and its intercept offset collapsing from +0.378 dex to +0.010 dex, inside the band. No single change was hidden and no verdict was averaged away. The program's defense of the conventions it fixed in advance is printed in §5 as what it is: an argument about the data's error structure, which the EIV estimator honors and both least-squares directions violate, one each way. The no-closure result survives this audit. It survives with its convention-dependence priced and named, which is where such a result should live.

The result, in plain language

Every number in papers 82 and 83 passed through a statistician's choice — log space or linear, population scatter or robust, errors-in-variables regression or ordinary least squares. The choices were fixed in the written registration before the data were examined. This paper asks what they are worth, changing one convention at a time on the sealed results, with data and method held fixed.

The scatter test is robust. The scale test is fragile against the forms — a linear-space fit costs six of eight forms a pass. The slope test dies under inverse least squares: regress the wrong way round and the flagship McGaugh form lands at 3.986, its offset collapsing from +0.378 dex to +0.010 dex, inside the band. Of 48 cells, 14 flip — and every flip toward closure comes from an estimator that zeroes an error column the data carry. The no-closure result survives the audit. It survives with its convention-dependence priced and named, which is where such a result should live.

Program note

An audit of the verdicts of papers 82–83 under rival statistical conventions, run on the sealed results, with the computation sealed and reproduced before the audit ran. Verdict, as the paper states it: the negative verdict stands, with its convention-dependence priced — 48 single-swap verdict cells; 14 flip, and every flip toward closure comes from an estimator that zeroes an error column the data carry. The vacuum catastrophe remains unsolved, and this paper claims no result on it.

The program