Finite pool correction reduced mean absolute 8–13 Hz contrast bias and raised mean squared error for uniform 20-window samples without replacement from fixed ds004148 pools

We tested uniform window selection without replacement from fixed posterior EEG power pools. The source was OpenNeuro ds004148 version 1.0.1, with 60 adults, three sessions, and both eye states. The main score was the closed-eye minus open-eye 8–13 Hz log arithmetic mean power contrast. Each state supplied 20 selected windows. The mean expected shift from the full-pool contrast was 0.01975 dB. Its descriptive 95% person bootstrap interval was 0.00748 to 0.03419 dB. A finite pool delete-one correction reduced the shift to 0.00245 dB. Mean absolute pair bias fell from 0.03717 to 0.00334 dB. Mean squared contrast error rose from 0.65563 to 0.68395 dB squared in simulation. This was a 4.32% increase. We give an exact expectation identity, a numerical tool, and a second-order correction proof. The proof does not guarantee zero bias or lower error for general pools. The package preserves all 360 spectra, source hashes, code, results, bootstrap draws, and reproduction instructions. Two fresh source records rebuilt exactly. A full repeat and separate numerical checks passed. The result concerns selection from recorded pools. It does not establish a neural cause, clinical use, or performance on new recordings.

EEG · alpha power · finite population · selection bias · delete-one correction · logarithmic mean · mean squared error · uniform sampling without replacement