The test did not show lower log-value loss: 164 ds002843 first visits, 90-day split, hyperbolic discounting, linear utility, 2% lapse

We tested two choice rules in 164 first visits from OpenNeuro ds002843 version 1.0.1. The main models used hyperbolic discounting, linear utility, and a fixed 2% uniform lapse. They fit 9,074 choices at delays up to 90 days. They tested 10,572 choices at delays from 91 to 180 days. Mean test loss was 0.411401 for the value-difference rule and 0.410293 for the log-value rule. The mean gain was 0.001108 nats per choice. Its 95% paired person bootstrap interval was [-0.004849, 0.007670]. The test did not show lower log-value loss. It did not show equal loss. Five fixed secondary checks show sensitivity to the noise assumption. In the zero-lapse check, an analysis after scoring found that three people account for 98.52% of the net gain. A proof shows that an exponential log-value rule depends on only two products of three free positive parameters. Thus choices alone cannot identify the utility exponent, discount rate, and choice slope in that model. The package preserves raw event tables, source hashes, retrieval code, predictions, resamples, and reproduction commands. All independent checks passed. A full repeat produced identical analysis files.

intertemporal choice · discounting · choice rules · log-value model · value-difference model · hyperbolic discounting · exponential discounting · parameter identifiability · bootstrap · OpenNeuro ds002843