Fig. 6: Model-based FMRI results for prediction error coding (n = 31).
From: Dopamine regulates decision thresholds in human reinforcement learning in males
Striatal regions coding for model-based prediction errors were identified by computing a main effect of prediction error across all drug conditions using GLM1 (a, flexible factorial model in SPM12 with within-subjects factor of drug condition). Correction for multiple comparisons was performed using a meta-analysis-based region-of-interest mask (see Table 4 and methods section). To reproduce the analysis of Pessiglione et al.17, we then extracted parameter estimates at left and right striatal peak voxels (see Table 5) from GLM2 (flexible factorial model in SPM12 with within-subjects factors of drug condition and prediction error sign) to obtain parameter estimates for positive (+) and negative (−) prediction errors, separately for each drug condition (b). The map in (a) is thresholded at p < .001 uncorrected for display purposes, and projected onto the group mean T1 scan. Pl – Placebo, L – Levodopa, H – Haloperidol; +: positive prediction error, −: negative prediction error. For boxplots, lines represent the median, the box covers the upper and lower quartiles, and the whiskers denote the range of datapoints falling within 1.5 times the interquartile range.
