 |
 |
I thought it would be a good idea to take the within-subject nature of the FIR basis EVs into account. In other words, if one subject had an elevated baseline response to a whole condition, we expect the 5 FIR PEs that are passed up to be inflated for this subject compared to other subjects. But we don't want to introduce that variance into our f-test, which is just asking if the 5 FIR bases are different from each other, regardless of the individual subject baselines. Is that not how FIR analyses are usually done? I thought they were essentially a repeated-measures ANOVA, where you want to remove subject means from your input data so that your error term will be smaller and your stats more powerful? Thanks, Todd On Fri, Aug 7, 2009 at 6:50 AM, Eugene Duff<[log in to unmask]> wrote: > Hi Todd - > I'm not sure why you need the subject evs at all? If you are trying to > test whether there is any consistent (across subjects) non-zero signal > across any the 5 time points, then you just need the 5 evs for each of the > timepoints. In your current model, the subject ev tries to model a mean > level of signal across the 5 timepoints, so your f-test is testing for > deviation from a square response, which I dont think is interesting for you. > Eugene > 2009/8/7 Todd Thompson <[log in to unmask]> >> >> Hi, all. I can't quite figure out how to do this, but it seems like it >> should be something FSL can do. Perhaps someone can give me a pointer? >> >> I have 18 subjects, each of whom has 3 runs. I used an FIR model at >> the first levels (5 basis functions per condition), then collapsed >> across the three runs at the second level (fixed effects), so now I >> have one gfeat per subject, each of which has a COPE for each basis >> function in each condition. >> >> I'd like to analyze the condition f-tests at the group level, now, but >> I don't know how. My naive assumption was that I'd set up the >> contrasts like this example with 2 subjects, and 5 basis functions: >> 1 0 1 0 0 0 0 >> 1 0 0 1 0 0 0 >> 1 0 0 0 1 0 0 >> 1 0 0 0 0 1 0 >> 1 0 0 0 0 0 1 >> 0 1 1 0 0 0 0 >> 0 1 0 1 0 0 0 >> 0 1 0 0 1 0 0 >> 0 1 0 0 0 1 0 >> 0 1 0 0 0 0 1 >> >> >> The first two columns are the subject EVs, and the last five columns >> show the basis functions. >> >> I'd hoped to set up the contrasts like this: >> 0 0 1 0 0 0 0 >> 0 0 0 1 0 0 0 >> 0 0 0 0 1 0 0 >> 0 0 0 0 0 1 0 >> 0 0 0 0 0 0 1 >> >> So that I'd get one COPE for each basis function, and then I could >> just run an f-test across them all to see where my group activations >> for that condition were. From this, I could get estimated HRFs by >> merging the first 5 COPEs, and so on. >> >> Of course, the problem here is that in my EV matrix, the sum of the >> first 2 columns is the sum of the last 5, which makes the stats >> grouchy. >> >> How do I do this? >> >> Thanks! >> Todd >> >> >> p.s. Oddly enough, SPM doesn't seem to mind this setup at the higher >> levels, which I thought was a basic problem with linear algebra. Any >> idea how it avoids this problem? > > > > -- > Eugene Duff > > Centre for Functional MRI of the Brain (FMRIB) > University of Oxford > John Radcliffe Hospital, Headington OX3 9DU Oxford UK > > Ph: +44 (0) 1865 222 545 Fax: +44 (0) 1865 222 717 > > -- >