All reward BART choices lowered run 2 log loss versus cashout balloon choices in 187 AgeRisk people under fixed logistic models

We tested a data selection rule for the Balloon Analogue Risk Task in AgeRisk, OpenNeuro ds004711 version 1.0.0. We fit fixed logistic models to each of 187 people. Models trained on all run 1 reward choices had lower run 2 log loss than models trained on cashout balloon choices. The mean gain was 0.007907 natural-log units per choice. Its person bootstrap 95% interval was 0.005859 to 0.009953. This was a 2.83% loss reduction; 154 people had lower loss. The full-data models used 21,841 training choices and predicted 22,494 test choices. A planned check sampled the same training choice count as the cashout data. These random models had greater loss by 0.004294, with a 95% interval of 0.001976 to 0.006533. The Brier score showed no clear gain. The main benefit therefore does not isolate an effect of removing outcome selection. Other planned log loss checks favored the full data. The scope is observed choices in a second run during the same visit, with this fixed model. The package preserves all 567 source files, hashes, the local plan, source code, derived tables, fitted coefficients, predictions, all checks, and figures. Separate source retrieval and score checks passed. A full repeat gave identical core outputs. Hidden risk traits and brain mechanisms need separate tests.

Balloon Analogue Risk Task · choice prediction · log loss · behavioral modeling · OpenNeuro