Fig. 5: Model comparison.
From: Neuro-computational mechanisms and individual biases in action-outcome learning under moral conflict
a Model formalization for the chosen (red contour) option. In all models, decisions are based on a categorical logit function (aka softmax), contrasting the expected value across the alternatives with inverse temperature τ∈[0,5] and learning rate LR∈[0,1]. EV = Expected Value, PE = Prediction Error, Out = Outcome. Subscript M = money and S = shocks. Outcomes are coded by value: high-shock OutS = −1, low-shock OutS = +1, high-Money OutM = +1, low-money OutM = −1. If outcome is withheld during Dropout, OutS or OutM = 0. Outcomes are set to the same values (+1 and −1) for money and shock because of the optimization task (Supplementary Note §4). b Decision Model at the 11th trial for the Dropout condition. c Choices in the first 10 trials of the ConflictDropout blocks (Online experiment) as a function of preference together with the predictions by M1, M2Dec, and M2Out. M0 always predicts 0.5 and is not shown. Choices and predictions averaged over all blocks, error bars: s.e.m. across participants (N = 29 Considerate, N = 24 Lucrative, N = 26 Ambiguous). d Leave-one-out information criterion (LOOIC34) of the models over the first 10 trials. Error bars: standard error of the estimation LOOICerror. M0 performs poorly, the three learning models perform similarly. The information criterion (IC) scale captures how much information is lost when comparing model predictions with actual choices, and smaller IC thus characterize models that better describe behavior. N = 79. e Change of participant choices (N = 29 Considerate, N = 24 Lucrative, N = 26 Ambiguous) and model predictions for the 10th to the 11th trial. Dashed lines connect 10th to 11th trials when money is removed (gray background), dotted lines, when shocks are removed (yellow background). 11th trial not included in model fitting. M2Out (black arrowheads) makes the best predictions for the 11th trial. Error bars: standard error of the mean across participants. f Distribution over 4000 posterior draws of the summed log-likelihood of the 11th trial over all participants multiplied by −2 to place values on the information criterion scale as for LOOIC. M2Out outperformed all other models. We use LOOIC in d, because these first 10 trials were included in the fitting of the model, but log-likelihood in f because it is not included in the model fit. Source data are provided as a Source Data file.
