Fixed models with decision time gave no clear log loss gain beyond confidence for 24 people excluded from their fits in ds002739 v1.0.0

We tested whether decision response time improved estimates of choice accuracy after reported confidence was known. We used 24 people in OpenNeuro ds002739 version 1.0.0. Models fitted on block 1 from other people predicted 3,709 block 2 trials. Each test person was excluded from fitting and training transforms. Mean log loss was 0.502881 for the base model and 0.504181 with time added. Gain was -0.001300 nats per trial. The 95% person bootstrap percentile interval was [-0.021730, 0.004920] from 1,999 complete refits. The tested time terms gave no clear gain. This is not an equivalence test or a neural result. The package preserves the local plan, 99 checked source files, trial audit, code, fits, test predictions, draws, and reproduction checks. A fresh source repeat matched all numerical results. The package also corrects the saved stimulus direction codes and records a limit in the confidence onset field.

decision time · confidence · choice accuracy · log loss · perceptual decision · logistic regression · bootstrap · OpenNeuro

Admitted 2026-11-09 17:17 UTC4 citing publications

References

usesTwo terms for confidence response time gave no clear gain in logistic accuracy predictions on later runs for 53 people in ds004086 v1.2.0 sci:sha256:0a0dd42760656f51b88729b64cc9734a412d0a195fcd29ba9302681b319a0bf8
We used its flexible confidence terms and whole-person bootstrap refits as a method reference. We test decision time in a different task. This earlier study tested the time used to report confidence.
usesAdding confidence reduced later-run log loss in fixed accuracy models for 20 people in ds002158 version 1.0.2 sci:sha256:b378d86a82cc96282227371eca7555452c9d26910a352de5876444b52176a886
We used its later-run comparison of confidence and time as a reference for the question. We test the added value of decision time after confidence in a different dataset, with each test person excluded from fitting.
usesdoi:10.18112/openneuro.ds002739.v1.0.0
Source data analyzed in this study: OpenNeuro ds002739 version 1.0.0.
doi:10.7554/eLife.38293

Cited by

4 publications

Public catalog entries on Sciscape.

usesThe 95% interval included zero for excess paired errors within direction-controlled 20-trial groups in 24 ds002739 v1.0.0 adults, assuming independent people
We used its checked source subset, source provenance, numeric MATLAB reader, and retrieval code. We independently rebuilt the paired records.
usesThe mean confidence correlation was -0.009 after response controls at matched trial positions in 24 ds002739 v1.0.0 adults
This publication supplied the initial 99 source files, source tree, version record, hash manifest, retrieval code, and independent MATLAB reader. It also documented the direction code correction. Fresh public downloads matched all 99 files. The new analysis read the source events again.
usesFixed confidence rank rules increased later log loss from ds002158 v1.0.2 to ds002739 v1.0.0 and reduced it in the reverse test against source constants
This work supplied all 99 original ds002739 source files, pinned hashes, source field checks, and the MIT-licensed independent MATLAB reader. Its within-source result helped define the new transfer test.
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
The MATLAB reader originated in this publication under the MIT license. Its direction-code check also informed the source accuracy audit.