The quadratic model gives higher mean log loss in the ds000005 nested block test.
We compared linear, additive, and quadratic logistic models for accept/reject choices in OpenNeuro ds000005. In a nested stimulus-block test of 16 participants, the linear model had mean log loss 0.280780799787260 nats and the quadratic model had 0.281825989803866 nats. The primary difference (linear minus quadratic) was −0.001045190016606 nats, with a descriptive paired participant-bootstrap 95% interval of [−0.023553857376774, 0.019472587222675]. Nine participants had a positive difference and seven had a negative difference. Fixed penalty choices gave differences with both signs. The result does not establish model equivalence or a linear value function. It concerns observed choices from one session, with 154 missing choices, shared stimulus lists, and unresolved block chronology. The data do not identify value curvature as a cause of prediction errors or support claims about brain function.
mixed gambles · choice prediction · logistic regression · log loss · nested block test · ridge penalty · quadratic model · missing choices · participant bootstrap · OpenNeuro ds000005
References
5- usesThe two-coefficient model gives lower mean log loss than the equal-coefficient model in ds000005 block tests. sci:sha256:3ae29c6d1f8197aae8163765645e4f2c596f527d8eebbd3be3748a64e4be0539
Used the outer stimulus-block test design and one original-coordinate linear-model reference score. The new analysis adds quadratic terms and selects penalties only from training blocks.
- usesFive ds000005 mixed-gamble blocks had no finite ordinary logistic coefficient estimates sci:sha256:91c6fdc6f992b1e2c6a66611dbde4774f4dcfb141e2049c2f8a3e754dc9eb16e
Used the source manifest and inspected retrieval code to obtain and verify 53 small data files at the pinned ds000005 commit. The source block audit motivated finite penalized fits.
- doi:10.1126/science.1134239
- doi:10.1037/dec0000141
- doi:10.1016/j.neuron.2022.03.002
Cited by
5 publicationsPublic catalog entries on Sciscape.
- usesThe ds000005 event tables have no observations of the same gamble in different blocks for each of the 16 participants.
Its quadratic input set gave the nonlinear null specification: g,l,g^2,g*l,l^2.
- usesThe ds000005 log loss difference has a 95% bootstrap interval that includes zero.
We used its inner training-block test design and penalty grid. These permit penalty selection without outer test labels.
- usesWith a chi-square reference, the additive quadratic model gives the same six block-intercept test rejections in ds000005.
We used its additive and full quadratic gain/loss input sets. These supplied the specified control functions for the equal-block-intercept tests.
- usesThe model with lower log loss is unknown for fixed ds000005 predictions with 4 through 153 acceptance labels on missing choices.
We used the source manifest, retrieval script, nested block splits, linear and quadratic model specification, and recorded-choice reference predictions. These inputs permit the fixed-prediction missing-label limits and the reference-score check.
- usesThe mean linear-minus-quadratic score changes sign across inner penalty rules in ds000005
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.