Fig. 4: Modelling Confidence.
From: Prior information differentially affects discrimination decisions and subjective confidence reports
a Model simulations of confidence. Confidence ratings following correct and incorrect trials were simulated from the model across different posterior information levels, between the two conditions. Simulations used the wchoice value from the model fit, 2.17, and four different values of wconf: 1 (optimal weighting of prior), 0.33 (3-fold overweighting of the prior), 3 (3-fold underweighting of the prior), and 1.27 (the value obtained from the model fit, underweighting the prior in confidence less than in the decision), shown from left to right respectively. The resulting mean confidence values are shown here. The model predicts mean confidence to increase with increased available posterior information following correct decisions and decrease with increased available posterior information following incorrect trials. As the prior is increasingly overweighted, the model predicts higher mean confidence in the Stronger-Lead condition following correct decisions, and lower in the Stronger-Lead condition following incorrect decisions (middle-left panel). As the prior is increasingly underweighted, the opposite pattern is predicted (middle-right panel). Due to the suboptimal weighting at the decision level, the model predicts differences in mean confidence when wconf is optimal (far-left panel). When the prior is underweighted in confidence, but less so than in the decision, the model can produce the pattern seen behaviourally (far-right panel). b Data and predictions of the individually fit model. On the left are the data, showing confidence following correct and incorrect decisions per posterior level and condition. On the right is mean confidence generated from sampling each individual participant’s fit posterior parameter distributions 10 times and simulating 720 complete trials for each of those sampled parameters, also using that participant’s staircased coherences, internal noise and decision bias. Error bars capture standard deviation (SD) of mean confidence across participants (N = 20). c Hierarchical model posterior distribution for wconf and b. The top posterior distribution is for the group mean parameter of wconf. The lower posterior distribution is for the group mean parameter of b. The blue shaded regions show the 89% credible intervals and the vertical black dashed line corresponds to the parameter values of an optimal observer. Source data are provided as a Source Data file.
