[FieldTrip] Motor beta activity - DICS solution more noisy than sensor data?

[Eelke Spaak]

Fellow FieldTrippers,

Currently I am looking at a contrast for left- versus right-hand index
finger button presses. As expected, on sensor level (combined planar
gradient, grand average) I see a clear lateralisation in beta band
power starting at least 0.5s before the button press (see
https://db.tt/Rtch3Qjy). Both 'blobs' are significant; there is
clearly more beta power ipsilateral to the response hand. I would
prefer to do further analyses on source level, so I attempt to
reconstruct the sources for this effect using DICS beamformer (common
filter, applied to both conditions separately; fixedori and realfilter
= 'yes'). The grand average results for this (again contrast left vs
right response hand) are shown at https://db.tt/IBQZG0d8 . (Ignore the
R/L-flip, this is radiological convention.)

As you can see, the source level solution is much more blurry than on
sensor level. This picture is without using any regularisation (lambda
parameter), the results are even worse when I use lambda = '5%'. The
negative blob (right hand higher power than left) becomes 'marginally
significant' on source level (p ~ 0.06) where it was p < 0.001 on
sensor level. The positive blob is nowhere near significant. Also, the
individual results are much less topographically consistent on source
than on sensor level (explaining the worse statistics).

I have checked the segmentation of my MRIs, the 'gray' seems to be
nicely within the head all the time. Also, I have manually verified
the alignment of headmodel, sourcemodel, and gradiometer information
for all subjects.

As a final note, the above sensor-level plot was taken from a 'slice'
out of a planar-gradient time-frequency analysis (mtmconvol). The
ingredient for the beamformer was an mtmfft fourier spectrum on the
axial gradiometer data, obtained for just the time-frequency range of
interest (subselect toilim [-0.5 0], mtmfft foi = 23, tapsmofrq = 7).
When I compute condition-averaged power based on these fourier spectra
and look at the contrast, the results are again as expected:
https://db.tt/n2P3UKcQ (of course less localised because of axial
gradient vs planar). The freq structures underlying this contrast are
exactly the same as those going into ft_sourceanalysis, so the problem
must be in the source analysis step (and/or in the preparation of the
geometric information, although these seem fine by visual inspection).

Does anyone have any idea that might explain these seemingly
contradictory results? I would have expected demixing to improve
signal-to-noise ratio, rather than worsen it.

Thanks!
Best,
Eelke