(A) Subjects chose between two pictures of differing value (reward and cost (physical effort, or delay)) by saccade. Neural activity was examined during the 1s choice epoch, whilst monkeys held fixation. (B) Correlates of decision variables in DLPFC single units (n=303). Coefficient of partial determination for multiple regression of left minus right action value (blue), chosen value (green) and left minus right choice (red) onto single neuron activity. Lines show mean /- s.e. across all recorded neurons. See also Figure 1—figure supplement 1. (C) Early action value difference coding predicts late chosen action coding. Z-scored regression coefficients for left minus right action value at 300 ms (ordinate) are plotted against Z-scored regression coefficients for left minus right choice at 700 ms (abscissa), for each DLPFC neuron (R=0.42; p=3.6*10-14). (D) Baseline-corrected ERP during choice epoch, in example subject, split by region. Lines denote mean /- s.e. across n = 125 (DLPFC)/n = 43 (OFC)/n = 85 (ACC) electrodes. Other subjects plotted in Figure 1—figure supplement 2. Vertical dashed lines are added to allow latency comparison with (E) and (F). (E) Z-statistic of regression of LFP data from example subject onto chosen and unchosen value. Note that at timepoint (i), overall value =chosen+unchosen; at timepoint (ii), value difference = chosen-unchosen. Lines denote mean /- s.e. across electrodes. In Figure 1—figure supplement 3, figure is split into different brain regions and cost/benefit. (F) Temporal derivative of evoked potential in part (D) (averaged across regions). Comparing parts (D), (E) and (F) shows that value correlates at timepoints (i) and (ii) occur when the LFP is ramping (derivative is non-zero), rather than peaking (derivative is near or at zero).