 |
 |
See inline responses below. On Wed, Jan 25, 2012 at 5:01 PM, Matt Weber <[log in to unmask]> wrote: > Hi all. I'm new to FSL, so I've got a rookie question that I haven't quite > seen through based on the answers posted here. (Quick pre-digression: I've > noticed that Dan Keeser has posted to this list a couple of times -- Dan, I > think I'm doing something very similar to what you did in your 2011 J. > Neurosci paper.) > > I need to set up a design matrix for dual-regression on concatenated ICA > of resting-state data. The experimental design incorporates a > between-subjects factor (Group: true vs. sham brain stimulation) and a > within-subjects factor (Time: pre- and post-treatment). My effect of > interest is a Group x Time interaction. > > My initial impulse is that the design matrix should look like the > following: > > Grp TRPR TRPO SHPR SHPO > 1 1 0 0 0 > 1 0 1 0 0 > 2 0 0 1 0 > 2 0 0 0 1 > > ... with the first row being a truly stimulated subject pre-stimulation, > the second a truly stimulated subject post-stimulation, and the next two > rows pre- and post-stimulation for a sham subject. Then the contrast I want > looks like this: > > Title TRPR TRPO SHPR SHPO > C1 IXN -1 1 1 -1 > > ... or, phrased conceptually, (TRPO-TRPR)-(SHPO-SHPR). > > Two problems. First, this matrix doesn't incorporate the repeated measures > nature of the design. It seems like the usual solution is to add one EV per > subject. That can be done, but does it then compromise the between-groups > part of the interaction? You need to include one EV per subject. It will not compromise your interaction effects, it will actually make them better. When you add the subject EV, you are then comparing the group differences between the change in conditions. An alternative, equally valid and identical, is to do the pre-post subtract prior to the group model. The new model would be a two-sample t-test of the difference images. > Do I need to add an EV for each group mean as well? (That seems redundant > with the "Group" designation, but I don't really understand what groups do > in this context.) I like to include it as it helps me build the contrasts, but it isn't necessary in FSL. In SPM, when you are estimating the covariance structure, then it can have subtle effects. > And second... it seems like no one else's solutions look like this. I > can't tell if this is because interactions like these with within- and > between-subjects components are actually not really handled in FEAT yet > (some emails on the list indicate they aren't), or if it's just because I > don't really understand how FEAT works yet. > time*group is a within-subject effect time is a within-subject effect group is a between-subject effect (this model will not produce the correct statistics for this effect) > > It seems like a common approach people around here use for problems like > these is to compute pre-post differences offline and then do basically an > unpaired t-test on the differences, putting it on a totally between-subject > basis. But I think the fact that I'm using dual_regression takes that off > the table, because dual_regression needs time course inputs. I would > really, really rather not try to modify dual_regression, though I'll try if > that's what's required. > The input into this model, which you say is through FEAT, should be a single image per condition, not time courses (unless I am missing something). As such, you can compute the differences at any point. If you do it outside of a model, then you won't get the variances that could be used for FLAME, but if you are using OLS, then a simple difference image (A-B) for each subject will suffice. You are already entering A and B into the model for each subject. > > I hope this all makes a bit of sense. Happy to expand on any of it. Thanks > in advance! > > > Matt >