I was wondering if anyone has done some experimentation with TFCE
parameters for MEG data?
I'm familiar with TFCE in FSL and they have their suggested H, E, sigma,
and connectivity clustering values for skeleton and other analyses. Looking
at the TFCE paper, the results seem quite sensitive to the chosen
parameters.
Eric's tutorial on the website is quite nice and uses pre-defined
parameters, so I was wondering if anyone has done any in-depth testing to
suggest an acceptable range for MEG data. And, not to make it more
difficult, any clues on how they would vary based on the analysis used
(e.g. power analysis, temporal clustering, different source localization
methods, etc)?
I'd recommend trying different ranges of parameters on the example file
first. This will give you some sense of how the minimum value and step
parameter affect the outcome for that dataset, where the correct/ideal
outcome is already known.
When it comes to extending these ideas to real datasets, the ideal
parameters could vary a bit as you suggest. The lower you set the initial
threshold, the longer the clustering will take (a practical consideration)
since you're including more points in the clustering step. Similarly, as
step size is decreased, the resolution of your significance values is
increased, but this comes at the cost of doing more clustering iterations
(computational time).
If FSL provides suggestions for analyses parameters in particular
situations, they should translate over to mne-python nicely since the
underlying algorithms are designed to be the same. Beyond that, I
personally can't make a strong recommendation at this point, since my
experience with TFCE is still somewhat limited (although I hope to work
with it in-depth soon). Perhaps someone else has some ideas, though.