Model predictions and model lesion analysis for Experiment 1.

(A) Subjects’ errors in Mode and Mean estimates (left panels, the same as Fig 3A & 3C) versus the predictions of the CoS models (right panels). The CoS models could well predict subjects’ errors. Insets refer to the mean across the three local skewness conditions. (B) Summed ΔAICc of the CoS model for Mean estimates and its lesioned models. The lesioned models included models without Lateral Inhibition (“CoS w/o LI”), without Power Transformation (“CoS w/o PT”), without both the components (“CoS w/o LI&PT”), and without clustering (“SW”, which is the subjective weighting model in Fig 5B). The unlesioned CoS model outperformed the lesioned models, implying that power transformation, lateral inhibition and the clustering process were all necessary for the CoS model to explain subjects’ Mean estimates. (C) Predictions of the four lesioned CoS models, which deviated from the data quantitatively or qualitatively. Given the models’ considerable differences in ΔAICc, their differences in the prediction plots might look counterintuitively small, which is partly due to the fact that these predictions are only about the mean trend of data whereas different models can also differ in their noise distributions. See text for the intuitions for these results.