The previous four-level response code increases mean log loss in 12 ds000005 model comparisons.

In 16 participants performing 48 blocks of a mixed-gambles task, a logistic model that used the previous binary response had lower held-block mean log loss than a nested model that added two contrasts of the previous four-level response code. The primary difference (binary minus code loss) was -0.013006255777611707 nat per response (approximate 95% participant-deletion interval [-0.01947364533753523, -0.006538866217688183]); the code model had higher loss for 14 of 16 participants. The difference was negative in all 12 planned model and penalty settings. The comparison used 3,815 rows from the same 16 participants, and peer predictors included other participants’ responses in the held block. It is therefore a conditional, same-sample comparison, not a forecast without peer data. The interval assumes independent participant selection at fixed task sequences; its small-sample coverage is unknown. The result applies to these models and this sample. It does not show that response codes contain no information or identify a psychological or biological cause.

response history · logistic regression · log loss · mixed gambles · model comparison · OpenNeuro ds000005