Fig. 5: Multivariate data analysis.

Classifiers were trained on the MEG field patterns acquired while participants were experiencing the outcomes (grey segment). For each participant, this was a data structure containing the magnetic field at each of the 275 sensors acquired in each of the 540 trials. (i) Channel selection. The number of channels was reduced to 135 by selecting the participant-specific subset containing the least eyeblink artefacts⁴¹. (ii) We then computed the cross-validated accuracy of the classification at each time bin. As training set, we retained approach trials in which either the positive (P, green) or the negative (N, red) outcome was presented, and discarded neutral and avoidance trials (hyphen, white). At this stage the regularisation λ coefficient was set to 0.025. (iii) The time bin of peak accuracy was then selected to build the training set of the classifiers, which (iv) were defined as the 135 weights associated with each channel resulting from a lasso-regularised logistic regression. The λ coefficient used at this stage was optimised with a second cross-validation procedure. (v) Analysis of the deliberation phase: the classifiers were then used to estimate the relative probability that either outcome was being represented during deliberation aligned either to trial start (cyan segment) or token appearance (yellow segment). (vi) The classification resulted in outcome representation probability (p(o)) time series of which we considered one time bin at the time to (vii) compute a LME and extract fixed-effects statistics (loss probability: low (L), medium (M), high (H); loss magnitude: 0–5; choice: approach, avoidance). (viii) For statistical inference, we applied a non-parametric cluster-level correction over the F-values of the main effects resulting from the LME.