The last choice and its win interaction gave no clear gain in fixed choice models of 38 ds004917 fMRI records

We tested whether the last choice and its win interaction improve prediction for people outside the fitting data. We used repaired choices from OpenNeuro ds004917 version 1.0.1. The fixed primary model used 1,140 adjacent choices from 38 fMRI records. The gain in log loss was -0.00004 nats per choice. The 95% person bootstrap interval was [-0.00730, 0.00400]. The secondary EEG test used 5,094 choices from 24 records. Its gain was 0.00042, with interval [-0.00529, 0.00385]. Nine public IDs occur in both groups. All five fMRI checks had intervals that included zero. Each bootstrap sample refit both models. The full second run gave the same numeric bytes. The package preserves source records, code, predictions, and checks. The result applies to the fixed models and fold rules. It does not prove that choice history has no effect.

choice history · decision making · logistic regression · cross-validation · log loss · fMRI · EEG · OpenNeuro ds004917