Fig. 2.

Drift-diffusion model fits and parameters. The drift-diffusion model (DDM) included eight free parameters (a). Apart from the standard-parameters reflecting drift rate (v), boundary (a), and non-decision time (Ter), we included variance parameters that determined how much standard-parameters varied from trial-to-trial (sv, sz, st, cf. Supplementary Fig. 1). Model comparisons (b), revealed that two additional free parameters for distractor weighting (f, red) and it's trialwise variance (sf, yellow) increased model fit measured as approximate BIC (see Supplementary Figure 2 for results of worse fitting models). f varies with variance sf over trials and modulates flanker processing by scaling the drift rate during distractor presentation. (c–e) shows quantile fits of the model (dark blue) against human RT data and (f) shows model and human accuracy. In all conditions (congruent & incongruent correct as well as incongruent error), the model captures the RT data in each quantile, suggesting a good fit to the data. Plot conventions as in Fig. 1b. Note that we removed congruent errors from the analysis as these were very rare (<2.5% of trials). g–m Displays the relationship between model parameters and behaviour across subjects. Displayed are regression coefficients and 99.9% confidence intervals. g Faster participants were fit by higher drift rates, lower decision boundaries, and lower non-decision times. h–j RT on error trials h was most strongly dependent on the non-decision time because errors are usually very fast, whereas drift rates are more closely associated with RT on correct congruent (i) and incongruent (j) trials. k Accuracy was most strongly reflected in the height of the boundary (a) but additionally dependent on how much distractors were processed (f) and how variable their suppression from trial-to-trial was (sf). Higher variance reduces accuracy, because in more trials distractors are likely to not be suppressed. l Interference and the ratio of errors between incongruent and congruent trials (m) additionally covaried with distractor weighting and its variance (f, sf): the less flankers were suppressed in the model, the higher was the interference and the more incongruent compared to congruent errors a participant made