Full delays removed oracle bias but gave opposite mean absolute error changes in 347 stationary lognormal go models with fixed delays and 250 ms stopping
We tested delay averaging in 347 model profiles from two public stop datasets. Each model used stationary lognormal go times, fixed recorded delays, and a stop time of 250 ms. The full delay method used the empirical quantile of all go time minus delay values. It had zero oracle bias. Oracle bias is error when the true model probabilities are known. The usual method had mean absolute oracle bias of 50.40 ms in 226 AOMIC profiles and 10.66 ms in 121 ds008037 profiles. We simulated 2,000 records per profile. Using full delays increased mean absolute error from 82.01 to 83.31 ms in AOMIC. It decreased this error from 22.14 to 21.07 ms in ds008037. Profile bootstrap gain intervals were -5.13 to 2.75 ms and 0.58 to 1.62 ms. The full delay method reduced squared bias and increased variance in both cohorts. These results do not estimate human stop times. We preserve all source files, simulation outputs, code, proofs, checks, and reproduction instructions. All 482 source files matched fresh downloads. A complete second run reproduced all numerical outputs.
stop-signal task · response inhibition · integration method · delay averaging · race model · oracle bias · sample error · simulation benchmark
References
11- usesCount delay control gave lower mean absolute error in 100 trial AOMIC PIOP2 simulations with 200 ms start delays and fixed 250 ms stop times sci:sha256:28ce2dc3ead545704669cc5ec76850693decbe63a39350436b82b5bcfd3ad509
Its fixed 250 ms stop simulations informed our benchmark stop time. We held recorded delays fixed to test delay averaging instead of controller choice.
- usesTime changes gave negative pooled screens in 29 of 223 constructed AOMIC PIOP2 models with five blocks and conditionally independent 200 ms stopping sci:sha256:5d153cb72eee1959f5c88631fa6c87f4af2ae5c7d0a1a4dda653263706418fe0
Its fixed-stop race constructions showed effects of time changes on a pooled screen. They informed our choice to hold go speed stationary and test delay averaging.
- usesIntegration estimates of stopping time correlate less across halves than go response times in 132 selected AOMIC PIOP2 runs sci:sha256:f8c60202fcda29057738eb637e45397b0734a14048aaf77d7483b78722c0a330
We used its 233 preserved AOMIC source files and fixed retrieval manifest. We checked all bytes against fresh downloads. We wrote new analysis code.
- usesIntegration estimates of stopping time correlate less than go response times across task halves of 96 trials in 121 ds008037 runs sci:sha256:ae7b8bd6b118ba7c0f207f111d9897f9bad6653aa577bce663c70ac202116287
We used its 249 preserved ds008037 source files and fixed retrieval manifest. Its source audit informed removal of rows with missing block fields. We checked all bytes against fresh downloads.
- doi:10.7554/eLife.46323
- doi:10.1016/j.neubiorev.2008.08.014
- usesdoi:10.18112/openneuro.ds002790.v2.0.0
The benchmark uses the recorded delays and go-response profiles from this exact AOMIC PIOP2 release.
- usesdoi:10.18112/openneuro.ds008037.v1.0.1
The benchmark uses the recorded delays and go-response profiles from this exact dataset release.
- doi:10.1038/s41597-021-00870-6
- doi:10.64898/2026.07.12.738085
- doi:10.1162/imag_a_00349
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesFour delay groups narrowed capped median stop time bounds in 347 fixed independent race models with known lognormal go laws
This publication supplied unchanged source files, fixed manifests, and the 347-design scope. We rebuilt and checked its go-profile table before we calculated the new median bounds.
- usesFixed mean 250 ms and SD 100 ms allowed full delay integration widths of 0.12 to 96.10 ms in 347 stationary lognormal go models with fixed delays and independent integer stop times from 0 to 500 ms
We used its 482 source files, pinned retrieval manifests, MIT retrieval code, and 347-profile table. Its full delay identity for variable stop times defined the audited value. We extend its constant 250 ms benchmark with fixed mean and variance bounds, endpoint laws, and dual checks.