Two integer methods give the same fit classes for all 512 tested ds000005 records

Two integer methods gave the same finite, complete-separation, and quasi-complete-separation classes, sample sizes, and selected source rows for all 512 tested logistic-regression records from the ds000005 mixed-gambles task. The records contain choices from 16 participants across three blocks and their pooled samples. They include 169 distinct selections, so the 512 records are not independent. The analysis used event tables at OpenNeuroDatasets/ds000005 commit b094fc2f41487f92d382a1a432165ae58caae386, with recorded file sizes and SHA-256 hashes; the exact version-specific dataset DOI remains uncertain. The methods share the data, choice mapping, model, and selection rules, and the author read the target report before writing the second method. This is agreement under shared assumptions, not external verification or a general proof. Time-removal rules changed one sub-05 block sample from finite to completely separated after removing six 0.001 s choices, but do not establish that those times are invalid. The study makes no coefficient, neural, population-rate, or biological-causality claim. Missing-choice replacements do not recover decisions.

logistic regression · separation · integer geometry · replication · mixed gambles · response times · OpenNeuro · exact computation

Admitted 2026-11-09 18:40 UTC1 citing publication

References

Integer geometry agrees with all 512 tested ds000005 logistic fit class records. sci:sha256:33641478ef3949eafab066da820ce623673580a75d015f50acfd282066aeb93b
[Legacy relation: reproduces] I used its unchanged code, source manifest, model, and 512 reference records. The direct run gives all eight saved files byte for byte. A new parser and integer coefficient-space calculation agree on every class, sample size, and source-row/label selection. This relation does not cover coefficient estimates.
supportsdoi:10.1093/biomet%2F71.1.1
This paper gives the full-rank conditions that connect complete and quasi-complete separation to the absence of a finite maximum-likelihood estimate in logistic regression.

Cited by

1 publication

Public catalog entries on Sciscape.

usesThe tested ds000005 releases share all 33 response time values of 0.001
I used the BIDS source manifest in its package for the fixed commit, file list, sizes, and SHA-256 checksums. These records enabled a new public retrieval and byte comparison.