 |
 |
Hi, If you are doing with outside of FEAT with randomise (as you should for VBM - I didn't realise this was what you were doing) then you just need to calculate the individual first-level results with fslmaths before running the randomise analysis (which will be doing the group average). You only have one contrast to feed up (the negative version can be taken care of in the randomise analysis by using a -1 contrast on the group mean), and the weightings are: 0.75 0.75 -1 -1 -1 0.75 0.75 which you can calculate with: fslmaths image1 -add image2 -add image6 -add image7 -mul 0.75 -sub image3 -sub image4 -sub image5 contrastimage Do this for each subject, combine the contrastimage results into a 4D image, and then run randomise with just a group mean design (like the one you had for step 2) but include both a +1 and a -1 contrast. This will then allow you to do your statistical test validly. All the best, Mark On 21 Nov 2013, at 02:20, Vincent Koppelmans <[log in to unmask]<mailto:[log in to unmask]>> wrote: Dear Mark, Thank you very much for your reply. The 7 conditions do differ, and I would not want to group them. The first two (1,2) measurements are pre treatment measurements, the following three (3-5) are 'during treatment' measurements, and the last two (6,7) measurements are post treatment measurements. One of many questions that I would like to answer is: is there a difference between the treatment measurements and the no-treatment (pre + post) measurements in GM density (this is a VBM analysis). I am not sure if I follow your reply, so please bear with me. What I understand is that I need to do the following: 1) look at the effects of time within each subject separately 2) pool the results 1) if this is correct, then - I have to create a 4D file for each subject with all of his/her scans/time points - the following model would be appropriate for this job: <01_subject.png> 2) pool the results by creating a mean contrast using the stats images from step 1) as input: <02_group.png> I tried the first step on each subject, but I get no results: randomise -i GM_sub_1.nii.gz -m mask.img -o fslvbm -d design.mat -t design.con -T -n 500 randomise options: -i GM_sub_1.nii.gz -m mask.img -o fslvbm -d design.mat -t design.con -T -n 500 Loading Data: Data loaded 140 permutations required for exhaustive test of t-test 1 Doing all 140 unique permutations Starting permutation 1 (Unpermuted data) Warning: The unpermuted statistic image for the current image contains no positive values, and cannot be processed with TFCE. A blank output image will be created. Starting permutation 2 Starting permutation 3 (etc) Could you please point me in the right direction? Thanks! - Vincent Op 11 nov. 2013, om 05:59 heeft Mark Jenkinson <[log in to unmask]<mailto:[log in to unmask]>> het volgende geschreven: Hi, If you are just repeating the same condition 7 times in each subject then the Three Level Analysis is fine, but I'm guessing from the contrasts in your design that this is not the case (although you haven't really explained this). The design you attach is problematic because it is rank deficient (it models the mean value in two different ways) and also cannot model the between-subject and within-subject variances correctly (as the residuals will be a mixture of these variances and the GLM does not know this). Instead I think you would be better off doing a two level analysis where you calculate a single measure of interest from each subject and then feed this up to the higher level analysis. Given your contrasts, it seems that you have a fairly simple question that could be set up as a contrast for each subject, and then you can easily feed this up to a very straightforward higher-level analysis that just averages across the group. All the best, Mark On 10 Nov 2013, at 02:04, Vincent <[log in to unmask]<mailto:[log in to unmask]>> wrote: Dear FSL experts, I am trying to figure out the best way to create a model for a one-group, 7 time points repeated measures ANOVA. The FSL GLM wiki states that: "FEAT has the ability to accommodate independent data with heterogeneous variance (the mixture of common between subject variance and subject-varying measurement error noise), but cannot account for arbitrary repeated measures correlation. Likewise, randomise cannot accommodate general repeated measures designs. That said, there are several very special cases where the GLM (and Feat, and randomise) can model repeated measures, though usually with assumptions and caveats." The "Multi-Session & Multi-Subject (Repeated Measures - Three Level Analysis)" (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Multi-Session_.26_Multi-Subject_.28Repeated_Measures_-_Three_Level_Analysis.29) however, seems eligible as a model. Is this the only right model for a one-group, 7 time points, within subjects repeated measures ANOVA, if I want to use FSL randomise? Why can I not just use 'this' (see attachment) design? Thanks, - Vincent <Model.png>