Fig. 3.

Reconstruction of a five-dipole network: Comparison of ability to infer a simulated network structure under different corrections for source leakage. Five connected dipoles were simulated in the left and right occipital lobes, the right and left frontal lobes, and the right somatosensory cortex (see Fig. 2), for 15 experiments of duration 10 min, at an SNR of 1.0. Group-level z-statistics for the full and partial correlations between the power time-courses of 38 cortical ROIs (Fig. 2B) are shown, after the application of corrections for source leakage. We combine two metrics into one heat map with full correlations below the diagonal and (unregularised) partial correlations above. We compare uncorrected data (B), data adjusted using the symmetric multivariate correction presented in this paper (C), and data corrected for zero-lag correlations pairwise between ROIs (D). C is the only set of matrices which can be thresholded to achieve perfect reconstruction. The higher arrow highlights an example false positive, and the lower a weak true positive with a neighbouring strong false positive (see text). A symmetrical version of the simulated network matrix is shown (A), indicating the relationships between ROIs which would constitute a perfect network discovery for these (undirected) network analysis methods.