Supplementary Figure 3: Rating variability predictions.

We compared the quality of the efficient-coding model fits (left panels) with a simple and flexible model assuming constant Gaussian noise over the rating scale without posing any prior distribution constraints on the values \(v_{(1, \cdots ,M)}\) (right panels). We found that the efficient-coding model explains the distribution of rating data considerably better than the alternative model for both Experiments 1 (n = 38, panel a) and 2 (n = 37, panel b). The LOO difference is > 500 units in favor of the efficient encoding framework for both models, which statistically confirms that the efficient-coding model captures the empirical variability more accurately. Black dots correspond to the empirical data and error bars in this panel represent the s.e.m. across participants. Model predictions are based on 500 simulated experiments (semi-transparent red lines) where we draw n = 2 ratings for each good and plot rating variability as a function of the mean rating (exactly as derived for the empirical data). The data of each participant is shown in main text Fig. 2b,c.