2026-10-11T04:31:55Z https://api.figshare.com/v2/oai

oai:figshare.com:article/34036522 2026-09-30T17:52:19Z category_15 category_24 category_45 category_734 category_64 portal_5 item_type_6 month_year_09_2026

<p>Correlation structure of model features across all trials and trade-off trials.</p> Sergej A. E. Golowin (25145509) Niall W. Duncan (9177024) Faizan Shaikh (25145512) Christoph W. Korn (11657787) Neuroscience Evolutionary Biology Sociology Biological Sciences not elsewhere classified Cancer without requiring assumptions enabled formal comparison choices showed evidence sequential foraging task integrated value signals approach &# 8211 g ., reward humans adjust decision avoidance &# 8211 threat probabilities ), higher reward probabilities optimal action values individual decision features higher predation risk predation risk action values value structure sequential decision optimal policy relevant features decision features reward probability probabilistic reward xlink "> strongly reflected shared model remains unclear reduced reliance policies based parsimonious account optimization principles increased alignment immediate rewards future risks findings suggest feature policy enhanced integration creating contexts commonly framed avoidance contexts <p>Pairwise Pearson correlations between optimal policy values () and candidate decision variables. Color intensity indicates correlation strength ( to 1). <b>(A)</b> Correlations across all trials. -values show relatively weak associations with individual environmental features (e.g., : <i>r</i> = 0.05; : ), but stronger correlations with state-dependent variables, particularly wait when safe (<i>r</i> = 0.70) and binary energy state (<i>r</i> = 0.37). Several heuristic and value-based predictors are more strongly intercorrelated (e.g., marginal value and expected gain: <i>r</i> = 0.68), reflecting shared structure among simpler decision variables. <b>(B)</b> Correlations restricted to trade-off trials ( near zero). In these difficult decision states, associations between and value-based predictors increase (e.g., expected gain: <i>r</i> = 0.53; marginal value: <i>r</i> = 0.50), while correlations with state-dependent features are substantially reduced or absent. The multi-feature policy was not included, as it is a composite function of the predictors and therefore shares variance with them by construction. Correlations between such composite outputs and their constituent features are not informative for assessing multicollinearity among predictors. Accordingly, this analysis is intended as a descriptive check for dependencies among candidate features and does not replace formal model comparison or model recovery procedures. Overall, the pattern indicates that correlations observed across all trials are partly driven by extreme (easy) states, whereas in trade-off trials, predictors align more strongly with value-based representations, consistent with increased alignment with integrated value signals in difficult decisions.</p> <p>(PDF)</p> 2026-09-30T18:00:31Z Text Journal contribution 10.1371/journal.pcbi.1014793.s003 https://figshare.com/articles/journal_contribution/_p_Correlation_structure_of_model_features_across_all_trials_and_trade-off_trials_p_/34036522 CC BY 4.0