I would like to create a signed dSPM map visualizing the difference
between two trial types (within subjects).
The steps seem pretty straightforward:
1) create difference waves for each subject using xplotter.
2) compute the forward solution, inverse operator, and signed dSPM stc
file for each subject based on these difference waves.
3) average the resulting stc files across subjects and load in mne_analyze
However, I'm not sure about the proper computation to perform to create
the difference waves in xplotter. I believe the proper calculation is either
Condition1 - Condition2
or
abs(Condition1) - abs(Condition2)
I'm leaning towards the latter. I reasoned that if I just took the
simple difference, the sign of the difference wave wouldn't consistently
correspond to a greater Condition1 or Condition2 deflection. A large,
negative Condition1 deflection and a smaller negative Condition2
deflection at the same time points would produce a positive difference
wave, for example. By taking absolute value, positive values should
consistently reflect greater Cond1 deflections and negative values
should reflect greater Cond2 deflection.
This method rests on the assumption, though, that greater deflection
from baseline always equals greater underlying neural response. I'm not
sure if this is the case (for example, what to make of a
positive-directed deflection that occurs completely below baseline?).
Could anyone please clarify if creating difference waves and then
projecting these into source space is indeed a reasonable method for
comparing conditions, and, if so, which calculation is more appropriate?
Hi Matt,
If you just want to visualize the difference between 2 conditions in
source space I think the mne_average_estimates (using weights of +1 and
-1 for the stc files of the 2 conditions of interest) method would be
very straightforward...
Based on your questions I would like to mention a couple of issues to keep
in mind:
(1) Subtraction in sensor space and then inversion is equivalent to
inversion of each condition and then subtraction is source space as long
as you use a common inverse operator and you don't do any non-linear
transformations - So taking abs() in sensor space is not appropriate
(2) If you choose to do a simple subtraction in sensor space i.e WITHOUT
abs(), then it is important to make sure that the number of trials going
into the 2 conditions are similar and to adjust the nave for scaling the
noise covariance.
Manipulating the weights fed into mne_average_estimates sounds like a
simple solution.
Sheraz, can you explain more what you mean by subtracting in source
space? Simply subtracting the dSPM matrix for one condition from the
other? Does manipulating the weights in average_estimates essentially
accomplish the same thing?
As Hari mentioned they are exactly same, also beside subtraction you can
always do statistical contrast between the conditions using epoch as
observation, which might shows better difference between conditions as
mean can be dominated by few noisy trials.