Hi Todd,
Hope you don't mind me weighing in here, your questions caught my eye because I've been working with PPI in both cross-sectional and longitudinal designs quite a bit lately.
As for the section in Jill's PPI page about "PHYS- only" models, my take is that she's criticizing the approach that treats an fMRI timecourse during which participants were performing a task as if it was a resting state scan (essentially a seed-based analysis during a task scan). Basically, constructing a GLM that includes only the PHYS timecourse of interest and ignoring that there were X blocks of task performance (let's just say 6 30s blocks of finger-tapping) leaves open the issue that regions that 'come online' to perform the task will have correlated timecourses simply because they are involved in fingertapping (i.e. the main-effect of task). This would likely influence what is being construed as 'connectivity' because the PHYS regressor not only has the subtle volume by volume information that you're interested in extracting connectivity from, but also contains task-activation information, which in spirit a PPI is fundamentally interested in ignoring.
As you point out, that type of model may be impoverished because it does not account for the upshoot and downshoot in signal leading in and out of blocks (which my guess, would contribute quite a bit of 'connectivity' to your PHYS regressor).
I'd suggest having a look at Donald McLaren's generalized PPI paper (2012 I believe) -- the implementation he describes is lent more toward SPM than FSL, but to my understanding a gPPI can be constructed in FSL in the same way as the PPI model, just with extra interaction terms.
To get at your specific questions about your implementation. I'd consider creating a model that is identical to the standard block analysis of your data (you say you have 4 conditions, so 4 task regressors), then add the PHYS timecourse you're interested in (just as prescribed in Jill's tutorial), and then create a PPI (Feat interaction term) for each of the 4 conditions X PHYS regressor. You'll have 9 regressors in the model (4 conditions, 1 PHYS, and 4 PPIs: 1 per condition). As per Donald's paper, including all the data in the model like this should actually increase your sensitivity, despite the colinearity inherent to PPI designs. Have a look at how the PPI regresors turn out in FEAT's model PNG file and I think it will clear up your concern about the lead-in 4-6s of each block. You'll see that the block regressors absorb all the up-down variance due to task-related activation. After modeling, you'll see that the PHYS regressor will actually account for nearly all the variance in the model (unfortunate reality of the design). But your 4 PPI regressors will tell you what was connected, discretely during the condition it was associated with. Each of those maps should be capable of being entered into any cross-sectional/longitudinal group comparisons you care to look into. With a PPI per condition you can also look at condition x timepoint interactions as well, which is where I think the real value comes in with the gPPI.
Best,
Jonathan Hakun
________________________________________
From: FSL - FMRIB's Software Library [[log in to unmask]] on behalf of Todd Thompson [[log in to unmask]]
Sent: Tuesday, January 28, 2014 5:29 PM
To: [log in to unmask]
Subject: [FSL] Functional connectivity changes/longitudinal PPI
Hi, all. I'm currently analyzing data from a training paradigm and
could use some advice, please.
In the "Between subjects designs: an alternative approach" section of
Jill's generally excellent PPI page, here:
http://www2.fmrib.ox.ac.uk/Members/joreilly/what-is-ppi
, she talks about a style of "PPI" where the "Psychological" regressor
is a constant "group A" or "group B" indicator, which makes the
physiological regressor and the interaction regressor entirely
redundant. The upshot is that you end up with simply a physiological
regressor in your model, and you can directly compare statistical maps
from that regressor to the other groups'. She then suggests that you
can also use this approach to compare connectivity before and after
training/TMS/whatever.
To do that analysis correctly, Jill suggests a regressor that includes
the timecourse only during the blocks of interest, and to de-mean that
timecourse in order to partial out the main effect of task (which we
assume is similar from pre- to post-training). I'd like to do that,
but have some clarification questions to make sure I'm not doing
something absurd.
So, here are my two questions:
1) What does "during the blocks of interest" actually mean? Presumably
I don't want to include the portion of the time course that is rising
from baseline activity, since this dramatic shift in activation will
dwarf the variability observed during the block and create
artificially high "connectivities" with all of the other
task-responding regions? Is simply ignoring the first 6s of the block
activity safe, or is there a better approach?
2) What else goes into the model? I have 4 conditions, and am
interested in the connectivity changes from pre- to post- on only one
of them. Presumably I should clean up the unexplained variance in the
model with a normal convolved regressor for the other three conditions
and nuisance regressors? Does that sound right? Just a model with one
seed timecourse that is 0's everywhere except for the blocks of
interest and has the demeaned signal in the blocks of interest, then 3
convolved condition regressors, then nuisance regressors for motion
and WM/CSF signal?
Thanks so much, in advance, and if there's a paper that goes into more
detail on this topic I'd love to read it!
Todd
----
Todd Thompson
Doctoral Candidate, Dept. of Brain and Cognitive Sciences
Massachusetts Institute of Technology, 46-4037C
|