Word terms with penalty 10 reduced recall log loss in one fixed ds004789 split of people. The refit interval included zero.
We tested recall probabilities in a fixed split of OpenNeuro ds004789 version 3.1.0. There were 87 training people and 86 test people. Test features included recall from five early lists. With a word penalty of 10, word identity terms reduced mean test log loss from 0.536341 to 0.532043 nats per word. The gain was 0.004299 nats. A fixed-model interval excluded zero. The primary interval refitted both models in 1,999 person draws. It was [-0.000798, 0.004569] and included zero. The refit distribution also shifted toward smaller gains. A word penalty of 1 gave no observed gain. The result does not establish a positive gain across new training samples. It applies to separate people on the same 300-word English pool after five early lists. The package preserves 551 source files, all list decisions, labels, model fits, scores, resampling counts, code, figures, and reproduction commands. It provides a checked behavioral reference for later memory prediction studies.
free recall · word identity · memory prediction · logistic regression · person bootstrap · OpenNeuro ds004789
References
3- usesdoi:10.18112/openneuro.ds004789.v3.1.0
Behavioral event tables and metadata from this exact dataset release supplied the recall labels and prediction features.
- doi:10.3758/s13423-020-01820-w
- doi:10.1038/s41562-020-0901-2
Cited by
5 publicationsPublic catalog entries on Sciscape.
- usesForward recall exceeded a uniform order reference with the recalled set, first recall, and total absolute distance fixed in 135 people from ds004789 v3.1.0
This publication supplied the pinned ds004789 source subset and full Git tree to the publication that we retrieved. We used those source resources to preserve and check the session selection.
- usesRecall amount had a higher correlation between odd and even lists than the order score in 135 selected ds004789 people
This publication supplied the archived ds004789 event files, dataset metadata, full Git tree, source manifest, and lowest numbered session rule. We wrote new parsing and analysis code. We did not use its fitted models or prediction results.
- usesAdjacent forward recall exceeded uniform orders with fixed recalled words and endpoints in 89 people from ds004789 v3.1.0
The full pinned source tree and retrieval helper in the two-session archive came from this upstream work. We preserved them and their source credit. They enabled path selection and fresh downloads with Git blob and SHA-256 checks.
- usesRecall fraction had a higher repeat correlation than a conditional order score in 89 people from ds004789 v3.1.0; stricter record rules gave difference intervals that included zero
We used its full source tree, version record, retrieval helper, and valid-list parser rules. These materials enabled source selection, retrieval, and list checks. The conditional order score and two-session comparison are new analyses.
- usesAdjacent forward recall exceeded uniform orders with fixed recalled words and endpoints in 89 people from ds004789 v3.1.0
The full pinned source tree and retrieval helper in the two-session archive came from this upstream work. We preserved them and their source credit. They enabled path selection and fresh downloads with Git blob and SHA-256 checks.