 |
 |
Hi Gwenaëlle, Thank you so much for your help. The design you suggested is attached, hopefully I understood that correctly. So would the difference between the two timepoints be the new DV, i.e. would I have a compute, for each subject, a new structural image defined as timepoint1 - timepoint2? (if so, I assume that's done with fslmaths) Is the fact that this design doesn't have each subject as its own EV (as suggested in the examples on the website) not a problem? I thought that this is done to model inter-subject variability as RFX. Also, I was wondering whether you (or anyone else..) has any thoughts about any of the other questions in my previous email.. Many thanks! Tudor On 10 June 2013 11:40, Gwenaëlle DOUAUD <[log in to unmask]> wrote: > Hi Tudor, > > here are some links that might help you further... > > For the 2 groups 2 timepoints in randomise, it is better to take the > difference between the two timepoints: > > https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1008&L=FSL&P=R8137&I=-3&d=No+Match%3BMatch%3BMatches > > For the two "nuisance" covariates, if you don't expect an interaction with > your two group labels, then add the demeaned age and demeaned gender each > in one EV and the corresponding two columns in your contrast file should > then be filled with 0. A useful link on when it is necessary to demean your > covariates: http://mumford.fmripower.org/mean_centering/ > > Cheers, > Gwenaëlle > > > ------------------------------------------------------------------------------------ > Gwenaëlle Douaud, PhD > MRC Career Development Fellow > > FMRIB Centre, University of Oxford > John Radcliffe Hospital, Headington > OX3 9DU Oxford UK > Tel: +44 (0) 1865 222 523 > Fax: +44 (0) 1865 222 717 > > www.ndcn.ox.ac.uk/team/researchers/gwenaelle-douaud > > ----------------------------------------------------------------------------------- > > ------------------------------ > *De :* Tudor Popescu <[log in to unmask]> > *À :* [log in to unmask] > *Envoyé le :* Dimanche 9 juin 2013 16h04 > *Objet :* Re: [FSL] VBM: Help with design and contrasts > > Thanks Mark. In that example, the interaction factor (EV2) seems to change > alternations between -1 1 to 1 -1, to enable the subsequent comparison > between the 2 groups. When having 3 groups though, how should this factor > be built, in order to have all possible comparisons between the groups, and > maintaining the sum of all its elements zero? > > And what is the difference between doing the inference via FSL/GLM versus > extracting grey matter volume values in fslstats and just doing a > mixed-model ANOVA on them in SPSS? The latter method seems easier to do, > but is it the case that it is not correct because it is skipping the > permutations done by randomise? > > - If YES: what exactly is being "permuted" by randomise but is not > permuted in SPSS? > - If NO: should fslstats be applied to structurals as last processed > by fslvbm_2_template, or do they also need to go through fslvbm_3_proc? > > Also, few more questions: > > 1) If it the case that *all *EVs should be demeaned, then how come > EV3..EV6, in the design<https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM?action=AttachFile&do=get&target=two_grp_two_lev.png>you linked to, are not demeaned, i.e. sum of all elements across the entire > EV (column) is not zero? > > 2) Where are the corresponding error/log files created after each FSLVBM > command is run? I couldn't find any fslvbm*.e*or fslvbm*.o* files in my VBM > directory – or is it the case that these files are only created in case the > command exits with an error code? > > 3) (from the fslvbm page) "To be able to compare all the GM images on a > voxelwise basis, they need to be transformed into a standard space, which > involves the use of non-linear registration (..) You want all the > structures across your subjects to match (that's why you use a non-linear > registration), but not "too much" or you would not be able to see any > difference, if all these structures were perfectly aligned across the > subjects." – what does "not too much" mean, and how exactly is it avoided > in fslvbm? > > Thanks!! > > On 9 June 2013 00:40, Mark Jenkinson <[log in to unmask]> wrote: > > Hi, > > Look at: > > https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#ANOVA:_2-groups.2C_2-levels_per_subject_.282-way_Mixed_Effect_ANOVA.29 > and follow the advice for a randomise analysis. > > Covariates can be added to the design easily, and they should be > demeaned (including gender). See > https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM/Faq?highlight=%28demeaning%29 > > All the best, > Mark > > > > > On 7 Jun 2013, at 13:39, Tudor Popescu <[log in to unmask]> wrote: > > Hi everyone, > > I have pre and post scans of 72 subjects divided into three groups, so 24 > in each group and 144 scans altogether. I would like to do VBM to look for > an effect of the "between" factor (group), as well as of the "within" > factor (pre/post), on grey matter volume, while co-varying out gender and > age. I am at the stage of the VBM pipeline where I need to create my > design.mat and design.con, and I'm a bit confused as to how these should be > created. > > From a general GLM perspective, I guess both questions, i.e. both the > effect of group and the effect of scan time (pre/post), *could* be > answered by a single design; however I know that the order in which > subjects are entered into design.mat has to be the same as the order in > template_list, and I've named my structurals according to the template > groupName_subjNr_preOrPost.nii.gz, which makes it difficult to see how the > two factors should be coded into the same design matrix. > > Also, I know factor demeaning is sometimes a contentious issue - which > factors (if any) should be demeaned, exactly? And should categorical > factors such as gender be demeaned as well? > > Regarding contrasts: assuming that only one design.mat is enough to test > both factors, is it correct that the design.con for the within factor > should have -1 for pre and 1 for post (or vice-versa), with all other > columns in the matrix receiving a 0? And that the contrast for the between > factor should have a 2 for one group and -1 for each of the other two > groups (with all three possible permutations of where the "2" goes)? > Also, what value should the covariate columns (age, gender) receive in the > contrast? > > Thanks very much in advance for any help, I really appreciate it! > Tudor > > > > > >