The last response decreases mean log loss with response lags 2, 4, and 8 in ds000005.
In 16 participants performing the ds000005 mixed gambles task, adding the immediately preceding accept or reject response to a ridge logistic model with response lags 2, 4, and 8 reduced the participant weighted mean test log loss by 0.006246811919520186 nat per choice (95% paired participant bootstrap interval 0.003090701241585878 to 0.009849542245150791; 14 of 16 participant differences were positive). The comparison used 3,319 eligible rows and tau=2.5. The full response history model did not outperform the stimulus only model, and the lag 1 versus lag 8 comparison had an interval that included zero. Tests were conditional on recorded response history, used one participant sample and fixed stimulus lists, and did not establish biological cause or memory duration. See the full methods, additional comparisons, reproduction checks, source hashes, and limitations in README.md.
choice behavior · response history · log loss · ridge logistic regression · OpenNeuro ds000005 · gamble task
References
3- The model with the last choice gives lower mean log loss in ds000005 block tests. sci:sha256:ecdbfc894a670fa01456dd0b6f6446526ed0a4e2edcc12914f8334cedbc25981
[Legacy relation: reproduces] The code calculates the tau=2.5 primary comparison again on the same 3,815 eligible response pairs. All 7,630 reference predictions agree within 3.6e-15 in logit. The source list, retrieval code, and model design also support the added lag comparisons.
- doi:10.1126/science.1134239
- doi:10.1038/s41598-020-66502-y
Cited by
3 publicationsPublic catalog entries on Sciscape.
- usesThe last response decreases mean log loss with response averages from other participants in the tested ds000005 models.
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 ds000005 models for 3,700 trials include no latent criterion state.
[Legacy relation: reproduces] The new code calculates the primary older-versus-all comparison on the same 3,319 eligible trials. All 32 reference participant scores agree within 4.661826480401032e-12 nats. The reference manifest and model specification enabled this check.
- The log loss difference changes sign with the penalty scale in the ds000005 two-state model tests.
[Legacy relation: reproduces] The code calculates all 32 primary participant scores again on the same 3,319 trial rows. The largest difference is 8.326672684688674e-17 nat per choice.