Extended Data Fig. 9. Model comparison (Exp. 2-4) analogous to Fig. 3f, but additionally allowing for differential learning from confirmatory vs. disconfirmatory choice feedback.

Here, all models included an additional parameter ω ∈ (0;10), by which belief-confirming learning rates were modelled as: α_conf = α * ω. Same conventions as in Fig. 3f. Markers show model fits using a pseudo-R-squared (Rsq, left y-axis; diamonds and error bars show mean ± s.e.m., dots show individual participants). Overlaid red bar graphs indicate each model’s probability of describing the majority of subjects best (right y-axis, pxp: protected exceedance probability). While the extra parameter ω led to general improvements in fit (note overall higher Rsq compared to Fig. 3f), the model comparison result with respect to winner/loser asymmetries was identical, both in terms of Rsq/BIC and pxp. The estimates of parameter ω in the winning models were larger than 1 (mean=4.72, SE=0.27, p < 0.001, r=0.83, Wilcoxon signed-rank test against 1, collapsed across Exp. 2-4), indicating an overall bias towards confirmatory feedback.