Person slope directions reduced test log loss after size adjustment in 107 selected NARPS adults under fixed ridge logistic models

We tested person gain and loss slope directions for later choice prediction. Each model permits person bias and slope size to change. We used 107 selected adults from OpenNeuro NARPS version 1.0.5. Fixed ridge logistic models fit runs 1 and 2 and predict runs 3 and 4. The direction model reduced mean log loss by 0.008365 nats per recorded choice against the scale model. The 95% percentile interval was 0.002645 to 0.014471. It used 5,000 samples of people with complete model refits. The relative reduction was 3.27%. All five planned checks had positive mean gains. The check trained on run 1 had an interval that included zero. The model proof shows an exact scale model restriction with the same penalty. The source cohort was selected using behavior from the full task. This result concerns later binary choices in one visit under the tested models. It does not establish a stable trait or a neural mechanism. The package preserves source files, retrieval hashes, code, fits, predictions, bootstrap samples, results, and figures. It includes instructions to inspect and reproduce the work.

mixed gambles · choice prediction · slope direction · ridge logistic regression · person-level variation · NARPS