 |
 |
Hi Kavous, Please see below: On 2 June 2017 at 02:12, Kavous Salehzadeh <[log in to unmask]> wrote: > Hi Anderson, > > I am so grateful for your prompt response. I am new to FSL and in > particular to PALM. So, sorry for any stupid questions. > > Let me explain more about the experiment and the hypothesis. My time > points are: > > t1: pre-intervention > t2: mid-intervention > t3: post-intervention > t4: follow-up (8 weeks after stopping intervention to see whether effect > is persisting or not) > > And my three groups are: > > G1: control (no effect hypothesized within time points) > G2: intervention1 (small effect hypothesized within t1-t2-t3 + *NO* persisting > effect on t4) > G3: intervention2 (greater effect hypothesized within t1-t2-t3 + > persisting effect on t4) > > So according to my hypothesis, it seems I should follow your advice on > using PALM and -ISE (without the assumption of compound symmetry). > > So now a couple of questions arose in my mind: > > 1) Can I still use semi-autonomous FSLVBM process (i.e., fslvbm_1_bet > - fslvbm_2_template - fslvbm_3_proc) and later use PALM? > Yes and no. PALM won't know where the data is coming from. Now I notice you are using VBM, which has skewed data, so the ISE assumption won't hold. That leaves you with a choice between the ugly and the bad: compound symmetry (unlikely to be valid) versus symmetric errors (certainly invalid). Violating ISE is on average slightly conservative (we have seen with quite extensive simulations), whereas violating compound symmetry leads to excess false positives. So, if I were to choose, I'd go with ISE. > > 2) I know that in the longitudinal VBM analysis with two time-points, I > should treat each time point as an individual subject, and later before > randomize subtract GM_mod_merg_s* files between two time-points. I wonder > in my case should I compute all pairwise differences and do PALM for each? > (i.e., t1-t2, t1-t3, t1-t4, t2-t3, t2-t4, t3-t4). Does the following > command seem right? > > *palm -i diff_t1_t2_GM_mod_merg_s3.nii -d design.mat -t design.con -m > GM_mask.nii -f design.fts -eb eb_file.csv -vg vg_file.csv -T -C 3.1 -n 5000 > -save1-p -corrcon -o myresults -ise* > > It wouldn't be crazy to test the various pairwise differences and do a joint test with NPC, but that wouldn't be specific for trends evolving over timepoints, and for your particular hypothesis it would amount to overfitting. > 3) Should I use [Pset, VG] = palm_quickperms(M, EB, P, ISE) or can I make > my own vg.csv file (likewise attachment). If this command is necessary can > I use design.mat for M? > You don't need palm_quickperms at all. This command is only if you want to generate the set of permutations and VG to use in some different software (e.g., to use with your own code for some unrelated task). The VGs can be constructed manually or, in most cases, created automatically with "-vg auto". > > 4) Can I still use FSL GLM command for making the design matrix? > Yes. Design matrix, contrasts, and EB (file *.grp) can be created with the Glm command. > Please see attached design matrix. Is it right matrix in my case? > This design isn't correct I'm afraid. Please see a different one here: https://dl.dropboxusercontent.com/u/2785709/outbox/mailinglist/ design_kavous.ods > > 5) Should I run PALM separately for permuting within-block (using > -eb -vg and -within options) and permuting whole-block (using -eb -vg and > -whole options)? > If you take my suggestion of not violating compound symmetry, then there will be no -within, only -whole. That said, in general, it's possible to do both things simultaneously (see the Multi-level block permutation paper <http://www.sciencedirect.com/science/article/pii/S105381191500508X>). Hope this helps! All the best, Anderson > > Best regards, > Kavous > > > On Thu, Jun 1, 2017 at 11:07 AM, Anderson M. Winkler < > [log in to unmask]> wrote: > >> PS: If you choose to use PALM, with the 4 timepoints, consider using the >> "-ise" option. It doesn't require compound symmetry, but requires that the >> errors themselves are symmetrical (i.e., have a symmetrical distribution >> around zero). >> >> On 30 May 2017 at 23:23, Anderson M. Winkler <[log in to unmask]> >> wrote: >> >>> Hi Kavous, >>> >>> What are the research hypotheses? If it's about changes over timepoints, >>> and interaction group by timepoint, then this needs the assumption of >>> compound symmetry, which is fine for 2 timepoints, but becomes harder with >>> 4. If you want to make that assumption, then you can use PALM, defining one >>> exchangeablity block per subject, and permuting within-block and also >>> whole-block (for the interactions. >>> >>> If you can't make the compound symmetry assumption, consider Bryan >>> Guillaume's toolbox called SwE. It seems the most recent version is on >>> GitHub: https://github.com/BryanGuillaume/SwE-toolbox >>> >>> If, however, the hypothesis is about group differences regardless of >>> changes in time, then randomise can be used directly, with the option >>> --permuteBlocks. This is the Example 6 of the randomise paper: >>> http://www.sciencedirect.com/science/article/pii/S1053811914000913 >>> >>> Hope this helps! >>> >>> All the best, >>> >>> Anderson >>> >>> >>> On 30 May 2017 at 05:05, Kavous <[log in to unmask]> wrote: >>> >>>> Hi FSLers, >>>> >>>> I'm going to analyze a longitudinal VBM between 3 groups and within 4 >>>> time-points. >>>> >>>> By reading following discussions, I know that in 2 time-points analysis >>>> I have to use all my subjects' data to make a study-specific template and >>>> subtract pre-post data after smoothing for randomise analysis. >>>> >>>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=FSL;d6651f48.1008 >>>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=FSL;3dc8868b.1408 >>>> >>>> However, my question is that is there any way to consider all 4 >>>> time-points in a single analysis? >>>> >>>> Thanks, >>>> Kavous >>>> >>> >>> >> >