Older feedback lowered mean log loss by 0.000375 nats for later choices in 47 ds005364 test people under fixed models; the 95% interval included zero

We tested whether feedback from more than five trials ago adds choice prediction value after recent history and older choices are included. The data were OpenNeuro ds005364 version 1.0.0. Fixed logistic models used 48 people for fitting and 47 people for testing. The test used 3,025 choices after trial 60. Both models used five recent trials, the current reward ratio, and older choice and missing-feedback terms. One model also used older signed feedback with decay 0.9. Each person had equal score weight. The added term lowered observed mean log loss by 0.000375 nats per choice. A 95% interval from 2,000 draws that sampled people and fitted both models again was -0.000311 to +0.001472. All eight planned setting checks also had intervals that included zero. The test people had been used in earlier work. This test did not give clear evidence of a prediction gain under the fixed models. It cannot identify a brain mechanism or prove that older feedback has no value. The package preserves 100 source files, pinned URLs and hashes, code, features, predictions, bootstrap draws, and repeat instructions. A new download matched all source hashes. A full repeat matched all 18 scientific output files.

reinforcement learning · choice prediction · feedback history · log loss · logistic regression · bootstrap · OpenNeuro ds005364