I'm trying to understand some puzzling results from my FSL analyses and,
specifically, I want to know more about how FSL handles empty EV files
(where there are no events of specific condition during that run). In my
analysis, I am sorting trials into different conditions based upon the
trial-to-trial subjective report of the subject. So a subject performs a
task and then depending on their rating, that trial gets sorted into
Condition 1, 2, or 3. We were originally interested in regions that were
parametrically active across the the 3 conditions, but many subjects (more
than half) never used the 3 response. So we have decided just to look for
regions that are more active in Condition 2 and 3 than in Condition 1. My
first approach at doing this was to code each condition in a separate EV
file and then combine across them in a contrast later (so the contrast
would be -2 1 1 for conditions 1<2&3). Since then I have changed to an
approach where I just include the 2s and 3s together in the same EV file
with no differentiation between them. My problem is that these two
approaches yield very different patterns of activity. To be specific, the
first approach (modelling 2s and 3s separately and then combining together
later) yields much more activity then when I model 2s and 3s together.
Looking at the difference between these activation maps it appears that
the extra regions correspond to voxels that are less active on 1s than
baseline. Is it possible that all those empty EVs for condition 3 are
being treated as baseline (i.e., a bunch of consistent 0 values are being
added in)? So my original attempt at a 2&3>1 map was actually regions that
were more active during a mixture of baseline and activity during
conditions 2 & 3. Does this make sense? If this isn't what is going on,
can anyone explain to me what is? I think the key here is what FSL does
when it has an empty EV.
Any advice would be greatly appreciated.
Ben
|