The last response decreases mean log loss with response averages from other participants in the tested ds000005 models.
In 16 participants completing a fixed set of mixed gambles, adding the last response to held-block ridge logistic models with current and previous response averages from other participants decreased mean test log loss by 0.008933143963118709 nat per choice on 3,815 eligible rows. The difference was positive for 11 of 16 participants; all 16 participant-removal estimates were positive. The approximate 95% jackknife interval was [0.0012655042388341918, 0.016600783687403226] nat per choice under an independent-participant sampling assumption. The averages use other participants’ responses from the test blocks, so the result depends on those records and fixed stimulus lists. This single-sample score comparison does not identify a biological cause or remove criterion changes, button effects, missing-response effects, or model error; performance on a new sample and different gamble orders remains untested.
choice behavior · logistic regression · log loss · response history · mixed gambles · OpenNeuro ds000005 · jackknife
References
3- usesThe last response decreases mean log loss with response lags 2, 4, and 8 in ds000005. sci:sha256:431956868d514aa58dd34786660ece4a7f1394b14d20bdc907588b88d19fc9dd
I adapted the inspected source file list, retrieval code, data loader, and damped Newton solver. They enabled the source hash checks and the new response-average comparison.
- The model with the last choice gives lower mean log loss in ds000005 block tests. sci:sha256:ecdbfc894a670fa01456dd0b6f6446526ed0a4e2edcc12914f8334cedbc25981
[Legacy relation: reproduces] The base and last-response models use the same 3,815 trial rows. All 7,630 reference predictions agree within 1e-7. The mean log loss decrease is 0.00854475912113023 nat per choice. This is a calculation with the same data, not another sample.
- doi:10.1126/science.1134239
Cited by
1 publicationPublic catalog entries on Sciscape.
- usesThe previous four-level response code increases mean log loss in 12 ds000005 model comparisons.
The inspected model specification, CC0 loader and numerical code, source manifest, and 3,815 reference peer_own predictions enabled the new four-code comparison and its calculation checks.