Penalty choice changed later log loss differences between affine and quadratic models in 108 selected NARPS people

We tested later binary choice prediction in 108 selected NARPS people. Models were fit separately for each person. Training used runs 1 and 2. Testing used runs 3 and 4. With a fixed unit coefficient penalty, full quadratic terms reduced mean person log loss by 0.00638 nats per recorded choice. The 95% person bootstrap interval was 0.00276 to 0.00987. A weaker penalty gave a gain near zero. An added test selected each model penalty with training choices. Its quadratic gain was -0.01135, with interval -0.01446 to 0.01000. Each bootstrap sample selected the penalties again. The measured benefit changed with the penalty. Source inclusion used choices from the whole visit. These results concern the shared cohort and stated models. They do not identify a neural mechanism. The package contains source files, hashes, plans, code, fits, predictions, bootstrap draws, figures, and reproduction steps. Fresh source retrieval and exact repeat runs passed. An independent parser and optimizer checked the results.

NARPS · choice prediction · logistic regression · quadratic model · regularization · log loss · bootstrap