Fig. 8. Modeling the reduction in context-dependent bias.

Shown is bias (Eq. (1)) as a function of RT, averaged across observers, in models and in the behavioral context experiments (gray lines, reproduced from Fig. 4 cd). a–c Bounded model (DDM, red lines), measured in four RT bins, where the influence of a context change (−20° vs. +20°) is modeled as a (a) drift rate change, b starting point change (same model as in Fig. 7), c drift rate and starting point change (see the Methods). d Unbounded model (purple lines), where the influence of a context change is modeled as a change in starting point in an unbounded diffusion process. This model was analyzed by fitting bias to a $\sqrt{\mathrm{RT}-t_0}$ decay rate function (see the Methods). The t₀ measures the non-decision time (10–550 ms; using a fixed value of 350 ms obtained from the RT distributions gives similar results). Note that the x-axis in panels a–c is for relative RTs (i.e., cumulative RT probability), while panel d is for physical RT (time difference from t₀), which is reasonable considering the different models. Fitting was done on the group level. Overall, the different modeling approaches (panels c, d) can qualitatively account for a slow rate of reduction in bias.