Previous choice terms increased mean prediction loss in later runs for 53 people in ds004086 v1.2.0 under fixed logistic models

We tested whether previous word choices improve prediction of later recognition choices. Both models used true word status, separate person terms, trial position, and previous confidence. We fitted runs 1 and 2 and tested runs 3 and 4 of the same 53 people in OpenNeuro ds004086 version 1.2.0. The test contained 5,857 valid adjacent word pairs. Mean log loss increased from 0.646216 to 0.649071 nats per pair when previous choice and its product with current true status were added. The increase was 0.002855 nats, or 0.442%. A 1,999-sample person bootstrap refitted both models. Its 95% interval for the increase was 0.000737 to 0.006207. All six forward prediction checks also showed higher loss. The reverse phase check was uncertain. The added terms improved the early training fit, but this gain did not carry to the later runs. A separate reader and a dense equation solver checked the results. A complete repeat from fresh source downloads matched all numerical results. The package preserves source bytes, hashes, the local plan, code, predictions, fitted coefficients, all resampling draws, figures, and reproduction instructions. It also proves that counts give the same person bootstrap objective as separate sampled copies. The result concerns these fixed models and this task sample. It does not establish a neural cause or exclude other history effects.

recognition memory · choice history · prediction · logistic regression · person bootstrap · OpenNeuro · ds004086