Adding four outcome features increased observed mean log loss in fixed logistic models for 24 EEG records from ds004917 version 1.0.1 under equal division of the hidden area

This study tested whether four features from the previous win status improved a fixed choice model. Each test person was absent from model fitting. The primary test used 5,519 adjacent choices in 24 public EEG records from OpenNeuro ds004917 version 1.0.1. The mean log loss increased by 0.000923 nats per choice when the four features were added (0.15 percent relative to Model B). The B-minus-C gain was -0.000923, with a 95 percent bootstrap range of [-0.004584, 0.000265] nats; the range includes zero. It describes the defined resampling rule, whose population coverage was not calibrated. All fixed EEG checks also had negative point gains. The separate fMRI test used 1,140 adjacent choices in 38 records and had a gain of 0.000028 nats, with range [-0.004238, 0.004044]. Nine participant IDs occur in both groups, so these groups are not independent replications. The result concerns the tested features, records, pair rules, penalty, and equal division of the hidden area. It does not show that outcome history has no effect. The study tests conditional next-choice prediction, not a brain mechanism or causal effect. The package preserves source records, provenance, analysis code, predictions, 1,999 person bootstrap samples per group, and independent checks.

choice behavior · outcome history · logistic regression · EEG · fMRI · cross-validation · bootstrap · OpenNeuro