Previous confidence reduced mean prediction error for later confidence in 17 MetaRDK patients under fixed ridge models
We tested whether previous confidence helps predict later confidence and accuracy in OpenNeuro ds006253 version 1.0.3. Fixed ridge models used the first half of each patient record for training. The later half supplied 2,273 test pairs in 17 patients. Adding previous confidence reduced mean squared error from 0.040085 to 0.029252. The paired reduction was 0.010834. Its 95% person bootstrap interval was 0.004418 to 0.018435. Error fell in 14 patients. Both main confidence models had more error than a simple training mean. An extra check removed the fitted time trend. Its error reduction was 17%, with a positive interval. This check was added after the main results were read. Accuracy gains changed with the model rule. The main log loss interval crossed zero. The main Brier interval was positive. The extra checks gave slightly worse accuracy predictions with history. Thus, the evidence supports confidence prediction in the tested scope. It leaves the value for accuracy prediction uncertain. The result does not identify a cause or a brain mechanism. The package preserves source data, a checked adjacent-trial benchmark, code, all model outputs, figures, source hashes, and reproduction instructions. An independent checker tested 522 fits. A fresh offline run produced 36 result files with identical bytes.
confidence prediction · serial dependence · epilepsy · ridge regression · behavioral data
References
5- usesAn offline rule that reversed binary choices below 50% confidence reduced mean accuracy in 22 ds006253 v1.0.3 records sci:sha256:15d919e897f825d0e7524832f92635e2fe19fd7266ec97bdee83337d73b5990c
This package supplied the initial source data, retrieval code, and pinned source-author selection facts. All source files were retrieved again and checked. Its accuracy-reversal result was not used as evidence for the new confidence-history result.
- usesdoi:10.18112/openneuro.ds006253.v1.0.3
The OpenNeuro ds006253 version 1.0.3 behavior tables supply the trial values analyzed in this benchmark.
- doi:10.1038/s41467-025-61744-8
- doi:10.3758/s13423-023-02261-x
- doi:10.3758/s13423-025-02748-9
Cited by
5 publicationsPublic catalog entries on Sciscape.
- usesThe 95% interval for confidence prediction gain from posterior alpha power included zero in 34 ds004368 participants under fixed ridge models for later trials
A prior confidence-history result in a different MetaRDK dataset motivated the secondary observed-history control before model fitting. This study does not reuse its data, estimates, or code.
- usesAdding previous confidence increased observed mean test log loss in 20 ds002158 v1.0.2 and 24 ds002739 v1.0.0 people under fixed models
Its history tests in MetaRDK helped distinguish prediction of current confidence from prediction of accuracy after current confidence is known. The present study tests two other datasets with different fixed models.
- usesThe primary 95% interval for current confidence log loss gain included zero in 17 MetaRDK patients under fixed ridge models with later test trials
This publication supplied context about previous confidence and prediction with a time split. That context informed our previous confidence control. We did not reuse its fitted models or analysis code.
- usesThe 95% patient bootstrap interval for current confidence gain included zero on later trials from 17 MetaRDK patients under fixed ridge models
We reused its 68-file source subset, fixed data manifest, Git tree, and extracted author selection facts. We checked the public bytes again. This study tests current confidence added to a correctness model; the cited work tested confidence history.
- usesCurrent confidence lowered observed loss, but the gain remained uncertain after prior-trial adjustment
The previous confidence comparator motivated our history question. That publication uses a different dataset and an epilepsy sample. We did not use its data or code.