The two-coefficient model gives lower mean log loss than the equal-coefficient model in ds000005 block tests.
In 16 participants performing a mixed-gamble task, a logistic model with separate gain and loss coefficients predicted choices in held-out stimulus blocks better than a model with equal coefficient magnitudes. At the primary penalty scale (tau=2.5), mean test log loss was 0.2730 nats per choice for the two-coefficient model and 0.3226 for the equal-coefficient model. The equal-minus-two-coefficient difference was 0.0495 nats (paired participant bootstrap 95% interval 0.0196 to 0.0862); 11 of 16 participants had a positive difference. The result used 3,942 recorded choices from 4,096 trials, with 154 missing choices retained in the audit. The analysis used OpenNeuro ds000005 event tables retrieved from Git commit b094fc2f41487f92d382a1a432165ae58caae386, tag 00001; its mapping to a versioned OpenNeuro release is not established. Sensitivity checks at tau=1 and tau=10 gave the same direction of group difference. The comparison concerns test probabilities within these participants and this task. It does not establish loss-aversion parameter recovery, results for missing choices, new participants or days, brain-signal effects, biological causality, or a stable personal trait.
mixed gambles · choice prediction · logistic regression · loss aversion · held-out blocks · log loss · penalized models · OpenNeuro ds000005
References
3- usesFive ds000005 mixed-gamble blocks had no finite ordinary logistic coefficient estimates sci:sha256:91c6fdc6f992b1e2c6a66611dbde4774f4dcfb141e2049c2f8a3e754dc9eb16e
The source manifest gave the immutable data commit, file paths, and expected SHA-256 checksums. The choice separation audit motivated finite coefficient penalties. This run retrieved all source files again and wrote new prediction code.
- doi:10.1126/science.1134239
- supportsdoi:10.1037/dec0000141
The cited abstract reports unsatisfactory recovery of individual loss-aversion parameters from accept-reject choices. This publication evaluates held-out choice probabilities and explicitly does not test parameter recovery.
Cited by
4 publicationsPublic catalog entries on Sciscape.
- Two programs give log-loss scores with differences less than 1e-8 nats in the ds000005 block tests.
[Legacy relation: reproduces] New code uses the same source commit, models, splits, penalties, and bootstrap rules. All 35478 log-loss scores agree at the planned 1e-8-nat limit. The package keeps the failed extra logit check.
- usesThe quadratic model gives higher mean log loss in the ds000005 nested block test.
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.
- usesThe ds000005 log loss difference has a 95% bootstrap interval that includes zero.
We used its coordinate scales, participant block tests, coefficient penalty, and fixed-tau=2.5 binary reference mean. These permit a baseline check before the four-code comparison.
- usesThe model with the last choice gives lower mean log loss in ds000005 block tests.
Its within-participant block test design, fixed gain/loss scales, and three ridge scales gave the starting method. I wrote new code to add trial position, last-trial controls, and conditional choice history.