FSL Listserve:

I have an event related design in which I am modeling effects corresponding
to 3 separate events/phases of a task.  I have 2 groups (patients vs.
controls) with different Ns and for some subjects in each group, two
consecutive task runs (unfortunately, not all subjects received 2 runs, but
there are an equal number of total runs between groups). I would like to
include both runs for those subject that have them (create a mean for that
subject) and combine these with subjects who received only one run, and
then conduct a group comparison.  If not for different number of runs between
subjects, I think that this would be a straightforward example of a multi-
session, multi-subject repeated measure analysis. Since it is not, below is the
analysis strategy that I am planning.

1.) Run a 1st level analysis for each run for all subjects (some will have a
single run, some will have two), with MCFLIRT on to conduct motion correction
specific to each run.

2.) For subjects with 2 runs, conduct a 2nd level analysis in which the lower
level feat directories from run1 and run2 are the specified inputs.  A fixed
effect model is specified and a single EV is coded as follows:

Group EV1
Input 1 (run1) 1 1
Input 2 (run2) 1 1

This is I believe will provide a subject’s average activation across the 2 runs,
generating 3 cope.feat directories corresponding to the 3 event/phases I am
interested in modeling.  

3.) Conduct a 3rd level analysis using a mixed effect model with inputs from
the 1st (for single run subjects) and 2nd (for two run subjects) level
analyses.  Group membership is now included as an EV and relevant contrasts
are coded to examine Control > Patient and Patient>Control differences.

What I am not clear on is how I can combine the *.feat directories from the
1st level analyses for subjects who have a single run with the cope*.feat
directories from the 2nd level analyses for subjects with 2 runs as the inputs
to this 3rd level analysis.

To generate cope*.feat directories for subject with only one run, I had
considered entering run1 twice and then zeroing this out as below, but I am
not sure what consequences that will have on variance estimates etc and it
just doesn’t seem correct.
Group EV1
Input 1 (run1) 1 1
Input 2 (run1) 1 0

I’d appreciate any guidance on this and/ or any suggestions or comments if
this analytic approach seems off base.

Thanks,
Jim