Monotone first-visit fits with fair ties had higher expected later accuracy than replay for repeated risky offers in 120 ds002843 v1.0.1 adults
We tested a monotone rule for later risky choices in OpenNeuro ds002843 v1.0.1. We fitted each person with first-visit choices. The rule requires acceptance to stay the same or increase when the amount and win probability increase. An exact minimum cut found the least and greatest optimal acceptance sets. A fixed equal mixture gave fair ties. For 14,250 matched later choices in 120 adults, mean expected accuracy was 80.63% for this rule and 79.49% for earlier-choice replay. The mean gain was 1.14 percentage points. A paired person bootstrap gave a 95% interval from 0.75 to 1.52 points. Both fixed extreme rules also gave positive mean gains. An exact hindsight audit gave mean test accuracy from 79.42% to 81.91% among all training-optimal rules. These bounds use later answers. They show that the rule for ties is part of the result. The claim concerns these people and repeated offers. It does not test new offers or a training effect. This is a new risk endpoint in a cohort used for prior intertemporal-choice work. The package preserves source data, hashes, code, certificates, predictions, bootstrap draws, figures, and repeat instructions. A separate solver and a complete rerun passed. A fresh retrieval matched all 1,175 source files.
monotone prediction · risky choice · repeated offers · isotonic regression · minimum cut · choice consistency · bootstrap
References
4- usesHyperbolic fits did not establish higher accuracy than replay on repeated offers: 118 ds002843 v1.0.1 adults, linear utility, fixed 2% lapse sci:sha256:e8c7ed10d4ce57adc5a521c5d31f21eecff6453d5ef4c665af0f912cbeb423f1
Its earlier-choice replay test motivated our comparison with replay on exact repeated offers. That report used intertemporal choices. We test a monotone rule for risky choices in the same source cohort. We did not reuse its code or derived data.
- usessci:sha256:9fa0ce68b90344ed7c34fd1ac8f08275070168c2f4e7c41247e1e867cd9397e2 not known to Sciscape
The design of its test across visits motivated our use of first-visit fits and later choices in the same dataset. Its repeated-offer audit informed our scope statement. We use the separate risk task and newly retrieved source files. No code or derived choices were reused. This is not an independent cohort replication.
- usesdoi:10.18112/openneuro.ds002843.v1.0.1
This exact dataset version supplied the risk task event files and metadata analyzed here.
- doi:10.1523/JNEUROSCI.2832-16.2017
Background publication describing the source study setting and gamble outcomes.
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesMonotone grid rules with the least early choice error allowed both answers for 12% of later offers in 108 selected NARPS records
Its audit of extreme training optimal rules, equal extreme mixtures, and hindsight later accuracy bounds informed this investigation. We used another task with distinct later offers and added exact counts and uniform fractions for all optimal full grid rules. No source files, fitted results, or code were reused from that publication.
- usesProbability weights and choice offsets reduced later log loss in 165 ds002843 first visits with fixed bounds and 2% lapse
This archive supplied 664 raw first-visit risk event files, six metadata files, and a source manifest. New code made all fits and results. All used source bytes were also retrieved from the fixed public Git commit and verified.