A reward and transition term improved later choice prediction for 164 people under both ds007474 reward rules with fixed ridge models

We tested whether the previous reward and transition term improves later choice prediction after five trials of history. We used public ds007474 version 1.0.5 behavior data for 164 people. Models used fixed ridge penalties and earlier trials for fitting. The added term reduced mean test log loss by 0.0773 natural log units per choice under the state reward rule. The 95 percent person bootstrap interval was 0.0523 to 0.1031. The gain under the stimulus reward rule was 0.0686, with an interval of 0.0446 to 0.0938. The paired gain difference interval included zero. The gain alone does not identify planning. A separate audit found that all 1,160 reward probability resets occurred on rare transitions. The smaller model had lower absolute test error. The 189 OSF behavior files are omitted because their recorded CC BY-NC-SA 4.0 terms conflict with SCI distribution terms. The scientific text and result figure are retained, but the analysis cannot be rerun from the SCI package alone.

choice prediction · reward · transition · ridge regression · behavioral control · two-step task · log loss