Hello Chris,
I don't think this approach is going to work - your single regressor is essentially just fitting the difference between the first few timepoints ( depending on the convolution used ) and the remaining timepoints, which is unlikely to generate meaningful results. Linking this to your  question below - it is fine to have just one EV for a session, but the condition modelled must have variation ( like a standard block-design ). If your ultimate goal is to find group differences between multiple before and after sessions, then I think it might be worth looking at Melodic/dual_regression.

Hope this helps,
Kind Regards
Matthew
Anderson and FSLers-

I've run into a technical, and philosophical problem analyzing the data this way.

The Experiment:
Session A - resting state
leave scanner; change in mental state; return;
Session B - resting state


desired: what brain regions are more/less active between mental state A & mental state B?

1 - technical) Each of my sessions is entirely one and only one EV.
a) Entering a square wave with no off period causes FSL's model setup to add some junk at the beginning of the scan so I'm using an 'off' period of .1s at the beginning.  I get the same result with just a single entry in the tradition 3 column format. Is this correct?
b) the session is 350s; does a HPF of 800s sound reasonable, so remove drift but keep any baseline offset? 
c) is the logic of the second-level feat different?  we're not doing Fixed Effects, since there are no repeated measures, which makes me think we should just be forming copes at this point.  However, adding 2 input directories, 1 for each EV, and specifying a cope doesn't lead to any copes being created.

2) Philosophically, I'm unsure if it's even possible to detect what we're after.  In each block, given our model, we're essentially looking for voxels with constant activity over the entire block.  Without any variation in any EV within a block, however, is it even possible to compare across sessions?  Any constant difference between sessions might be do to random scanner drift as much as an effect.  Essentially my concern is a lack of a baseline, since we're detrending.  Is the only change in our t-statistic due to differences in variance?


Perhaps functional connectivity analysis is the only way forward?

sincerely,

Chris Hardy
Catalyst Agency LLC

On Thu, Feb 11, 2016 at 1:48 AM, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Chris,

Please see below:


On 11 February 2016 at 00:20, Chris Hardy <[log in to unmask]> wrote:
Hey FSLers,

I have a multi-session question, for which I could not find an exact answer in the archives nor the multi-session / repeated measures example given in the documentation: http://fsl.fmrib.ox.ac.uk/fslcourse/lectures/practicals/feat2/index.html#multisession , although it is closely related to the example.

Specifically, for each subject, I have three different conditions (A, B, & C), and each is it's own entire functional session (one run for A w/ no other EVs, one run for B and no EVs, etc).  I simply want to end up with the group level copes A-B, B-C, & A-C.  Judging from the example, I should run 3 levels of FEATs (since there is no multi-session functionality built in):  
1) 3 separate FEATs for each subject - 1 for each session - and make a 'contrast' that is just 1 for the one regression since there's nothing to subtract against.  This essentially is asking if the parameters measured are greater than 0.
2) a FEAT within each subject using these 3 1st level FEATs to create the contrasts.
3) a group-level feat

Essentially just a paired t-test where the variables are different session blocks.

Yes, everything fine until here. Note that the multi-level in FEAT is precisely for multi-session, so there is a such built-in functionality. At any rate, please see below:
 

My concern is that upon following this logic and running the level 1 for each subject is that my A-nothing contrast showed no significant voxels, but it seems like there absolutely should be, given the person was alive.

Subject living or not, there may be no significant activity, even with a simple contrast as this. This isn't really a problem.

A suggestion is to continue with the analysis. It may well be that none or very few of the subjects actually display activity, but at the group level, some results may be seen (or still not).
 

Previously, I had tried concatenating the runs into one functional, and then having A, B, and C in each subject's initial model, but from what I can tell about FSL (I come from SPM), that creates issues with movement b/n sessions and temporal filtering problems.

Exactly.

All the best,

Anderson

 

Thank you for any input!


sincerely,

Chris