Common average subtraction changes posterior theta-to-alpha ranks in the first four minutes of 88 cleaned ds004504 v1.0.9 EEG files

We tested EEG reference choice in the first four minutes of all 88 cleaned ds004504 v1.0.9 files. The fixed score was posterior log10(theta power / alpha power). Common average subtraction reversed 565 of 3,828 participant pair orders, or 14.8%. Rank correlation was 0.884. The median absolute rank change was eight places. Alzheimer-control score AUC rose from 0.811 to 0.869. The paired change was 0.057, with a 95% participant bootstrap interval of -0.013 to 0.131. The interval includes zero, so a population AUC gain remains uncertain. Both half-file checks and the full-file check gave similar rank sensitivity. Saved real cross spectra support further fixed linear reference checks. Channel powers alone lack the required cross-channel terms. Independent numerical checks passed for all 88 source files and all 34 AUC rows. A repeat offline analysis matched all ten result files exactly. A fresh source replay matched one spectral archive exactly. The package preserves source metadata, hashes, versioned retrieval records, spectral data, bootstrap samples, code, results, and instructions. The result applies to these cleaned signals and source groups from one hospital. No classifier was trained.

EEG · common average reference · spectral power · theta-to-alpha ratio · cross spectra · OpenNeuro · Alzheimer disease · rank sensitivity