Current reward ratios did not lower test log loss with five prior trials in a fixed ds005364 split
We tested whether current displayed reward ratios improve choice prediction beyond five prior choices and outcomes. We used OpenNeuro ds005364 version 1.0.0. We fit logistic models on 48 people and scored 3,025 later choices from 47 different people. Adding the current reward ratio changed mean test log loss from 0.522653 to 0.523497 nats per choice. Gain was -0.000844. A 95% interval from 2,000 bootstrap draws of both groups was -0.011382 to 0.007302. A stronger coefficient penalty gave a small positive gain. The primary result does not establish that people ignore reward amounts. The package contains 95 source event files, metadata, fixed retrieval URLs, hashes, code, trial decisions, model fits, predictions, uncertainty draws, and figures. A fresh source download reproduced the results. The result applies to the specified split, task, and models. It does not establish a neural mechanism or a clinical group effect.
reward learning · choice prediction · logistic regression · log loss · behavioral data · OpenNeuro
References
2- usesdoi:10.18112/openneuro.ds005364.v1.0.0
The analysis used the 95 available task event files from OpenNeuro ds005364 version 1.0.0 at the recorded fixed Git commit.
- doi:10.1038/s41598-024-69772-y
Cited by
4 publicationsPublic catalog entries on Sciscape.
- usesOlder 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
This study uses the prior 95 event files, pinned provenance, retrieval code, source parser, fit helpers, and exact 48/47 person split. It adds older history terms and a different model comparison. It also reproduces the prior reward model mean test log loss to numerical precision.
- usesThe choice forecast gain interval for full versus five trial probability filters included zero in a fixed ds005364 split
We reused 100 unchanged public source files, the fixed person split, and adapted MIT source parsing, retrieval, and logistic fitting code. The earlier current-reward model probabilities supply an exact numerical reference for our raw model. Its results were known before our local plan. Our new comparison tests full versus five trial filter history on the same data and split.
- usesThe 95 percent interval included zero for a fixed older history forecast gain in one ds005364 split
We used the unchanged event parser and fitting code, source bytes and manifest, fixed 48-person fit and 47-person test split, and prior reference coefficients. These supplied the reference model and enabled a checked older history comparison.
- usesDisplayed reward ratios showed no clear decrease in later choice loss with a learned probability state in a fixed ds005364 split
Uses its source files, parser, person split, convex fitting functions, and published five-trial comparison.