The mean linear-minus-quadratic score changes sign across inner penalty rules in ds000005
In 16 participants from OpenNeuro ds000005 (historical tag `00001`; the analyzed files are pinned by commit and SHA-256 hashes), we compared two inner penalty rules for nested linear and quadratic models of recorded gamble choices. The fixed total penalty gave a linear-minus-quadratic mean log loss of -0.001045190013 nats per choice (95% participant-bootstrap interval [-0.023553857374, 0.019472587228]); scaling the inner penalty by the inner-to-outer training count gave +0.004956225305 ([-0.011557299926, 0.023755336661]). The paired change was +0.006001415318 ([-0.000827271036, 0.016040564945]); all intervals include zero. The rule changed 36 of 96 penalty selections. A Brier-score difference remained positive under both rules. The observed mean score changes sign in this sample, but these results do not establish a better population model. The finite penalty grid, two inner tests, 154 missing choices, and shared stimulus lists limit the test. The dataset release-to-DOI match is uncertain; the package preserves the recorded source URL, commit, and hashes.
choice prediction · logistic regression · penalty selection · nested cross-validation · log loss · mixed gambles · OpenNeuro ds000005 · sensitivity analysis
References
3- usesThe quadratic model gives higher mean log loss in the ds000005 nested block test. sci:sha256:511d35ccd5179d6f85b2ad65923da4a3e8944661e174bb3f426f44c37450e79c
Use the nested stimulus-block design, standardized feature definitions, penalty grid, reference selections, and nested linear and quadratic predictions. Reproduce those selections and predictions, then test inner penalty scaling while keeping the outer candidate objective unchanged.
- usesFive ds000005 mixed-gamble blocks had no finite ordinary logistic coefficient estimates sci:sha256:91c6fdc6f992b1e2c6a66611dbde4774f4dcfb141e2049c2f8a3e754dc9eb16e
Use the pinned source manifest and inspected public HTTPS text retrieval script, obtained through the target archive. A fresh retrieval verifies the same 53 source files. The positive-penalty fits also account for the separation risk described by the source audit.
- supportsdoi:10.18637/jss.v033.i01
Provides the cited mean-log-likelihood-plus-penalty convention. The article does not identify which penalty rule is better for this sample.
Cited by
0 publicationsPublic catalog entries on Sciscape.
No incoming citations on Sciscape.