Extended Data Fig. 1: Timeseries demonstrating value-related variance.

The mean across neurons of the coefficient of partial determination (CPD) for value (cued probability) over time, following cue onset. The CPD measures how much variance in each neuron’s firing is explained by a given regressor (see below). This timeseries validates the 200–600 ms post-onset window that we used in order to match that used in Dabney, Kurth-Nelson et al.¹⁵, because that window very closely matches the peak value-related coding in the timeseries (as in Kennerley et al.⁸). We therefore used 200–600 ms and did not try any other time windows in order to avoid any possible p-hacking. Nonetheless we note that in a window that captures this peak, defined as when the CPD is higher than two thirds of the maximum CPD (270–620 ms), the core correlation between reversal point and asymmetric scaling was significant, R = 0.38, P = 0.019. Shaded region is the SEM across neurons. Note, as in Kennerley et al.⁸, the CPD for regressor X_i is defined by \({{CPD}(X}_{i})=\left[{SSE}\left({X}_{-i}\right)-{SSE}\left({X}_{-i},{X}_{i}\right)\right]/{SSE}\left({X}_{-i}\right)\), where SSE(X) is the sum of squared errors in a regression model that includes a set of regressors X, and \({X}_{-i}\) is a set of all the regressors included in the model except X_i. The CPD for X_i is more positive if X_i explains more variance in neuronal firing.