Visit 1 response time did not establish lower loss for later choice changes in 118 ds002843 v1.0.1 adults under fixed logistic models

We tested whether visit 1 response time improves predictions of changed choices at a later visit after recorded choice and fitted difficulty are included. OpenNeuro ds002843 version 1.0.1 supplied 14,044 matched offers from 118 adults, with 2,316 changed choices. All rows from a test person stayed out of logistic model training. Difficulty used a visit 1 hyperbolic discount fit with linear money utility and a fixed 2% lapse term. The base model loss was 0.380830 nats per pair. Adding response time gave loss 0.379641, a gain of 0.001189 (0.31%). The 95% interval from 2,000 person resamples that refit both models was [-0.001205, 0.003087]. The primary interval included zero. This test did not establish a gain from the response time term under the fixed models. It did not establish equal performance. The interval is conditional on five fixed folds and the recorded visits. This is a test of prediction of repeated choices in a selected cohort with training between visits. It does not test a neural cause or a treatment effect. The package contains pinned source files, code, all predictions, all resamples, independent checks, and a complete second run with identical numerical files.

response time · choice stability · delay discounting · predictive modeling · repeated choices · person-level cross-validation · behavioral neuroscience