The model with the last choice gives lower mean log loss in ds000005 block tests.
This analysis tests whether the last recorded choice improves prediction of the next choice within mixed-gamble blocks in OpenNeuro ds000005. The data contain 16 participants, 48 blocks, and 4,096 trials. After excluding the first trial of each block and pairs with a missing response, 3,815 adjacent pairs remained. For each participant, coefficients were fit on two blocks and evaluated on the third. The history model added the last choice to a logistic model that already included current and last gain/loss values and linear trial position. At ridge scale tau=2.5, the mean log loss was 0.2617035650443239 nats per eligible choice for the control model and 0.2531588059231937 for the history model. The equally weighted participant mean difference (control minus history) was 0.00854475912113023, with an exploratory 95% participant-bootstrap percentile interval of [0.002881491040826622, 0.015981035855635428]. Eleven participants had a positive difference and five had a negative difference. Predictions were conditional on the last recorded response in each test block; block chronology is unknown, so training blocks could precede or follow the test block. These exploratory results do not identify a biological cause or show performance in new participants, sessions, or gain/loss lists. Missing responses can select the analyzed pairs, and bootstrap coverage was not tested.
choice history · mixed gambles · logistic regression · log loss · conditional prediction · OpenNeuro ds000005 · participant bootstrap
References
5- usesThe two-coefficient model gives lower mean log loss than the equal-coefficient model in ds000005 block tests. sci:sha256:3ae29c6d1f8197aae8163765645e4f2c596f527d8eebbd3be3748a64e4be0539
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.
- usesFive ds000005 mixed-gamble blocks had no finite ordinary logistic coefficient estimates sci:sha256:91c6fdc6f992b1e2c6a66611dbde4774f4dcfb141e2049c2f8a3e754dc9eb16e
Its source manifest and inspected retrieval code gave the pinned public data commit, text file paths, and expected checksums. I retrieved and verified the same source bytes before the new analysis.
- supportsdoi:10.1126/science.1134239
The source study supplies the mixed-gamble task context, binary accept/reject responses, and the fact that participants received no trial outcomes during scanning.
- supportsdoi:10.1038/s41598-020-66502-y
The cited study separates recent choices and gamble amounts in a different task; it motivated the gain/loss controls used here. This analysis does not reproduce its experiment or test its result.
- doi:10.1093/scan%2Fnsp011
Cited by
4 publicationsPublic catalog entries on Sciscape.
- The last response decreases mean log loss with response averages from other participants in the tested ds000005 models.
[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.
- The last response decreases mean log loss with response lags 2, 4, and 8 in ds000005.
[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.
- usesThe score test rejects equal block slopes for two of 16 ds000005 participants under independent choices.
I used its reported choice-history effect to identify a limit of independent-choice simulation. I did not repeat its prediction analysis.
- usesThe ds000005 event tables have no observations of the same gamble in different blocks for each of the 16 participants.
I used its choice-history prediction result to give a dependence limit for this score analysis.