Multivariate statistical analysis of magnetoencephalography data

UnrestrictedI describe methods for the detection of brain activation and functional connectivity in cortically constrained maps of current density computed from magnetoencephalography (MEG) data using multivariate statistical analysis. I apply time-frequency (wavelet) analysis to individual epochs to produce dynamic images of brain signal power on the cerebral cortex in multiple time-frequency bands, and I form observation matrices by putting together the power from all frequency bands and all trials. To detect changes in brain activity, I fit these observations into separate multivariate linear models for each time band and cortical location with experimental conditions as predictor variables; the resulting Roy’s maximum statistic maps are thresholded for significance using permutation tests and the maximum statistic approach. A source is considered significant if it exceeds a statistical threshold, which is chosen to control the familywise error rate, or the probability of at least one false positive, across the cortical surface. As follow-up techniques to identify individual frequencies that contribute significantly to experimental effects, I further describe protected F-tests and linear discriminant analysis. To detect functional interactions in the brain, I take these observations and compute the canonical correlation between a chosen reference voxel and every other voxel in the brain. The canonical correlation maps are also thresholded for significance, but here I use parametric asymptotic approximations. Based on collinearity properties of the vectors associated to the canonical correlations, I implement procedures to discard voxels whose interaction with the reference is due to linear mixing, and describe approximate F-tests to identify individual frequencies that contribute significantly to detected interactions. I evaluate these multivariate approaches both on simulated data and experimental data from a MEG visuomotor task study.; My results indicate that Roy’s maximum root is more powerful than univariate approaches in detecting experimental effects when correlations exist between power across frequency bands; as for canonical correlation analysis, I find that it detects experimental effects, allowing the simultaneous evaluation of several possible combinations of cross-frequency interactions in a single test