Thanks Anderson! On Tue, Jun 14, 2016 at 7:19 PM, Anderson M. Winkler <[log in to unmask] > wrote: > Hi Sheena, > > Please, see below: > > > > On 14 June 2016 at 02:13, Sheena Sharma <[log in to unmask]> wrote: > >> Hi FSL experts, >> >> I have just carried out TBSS analysis of 20 subjects that were analysed >> pre and post intervention. I set up a GLM for a paired ttest as described >> in the GLM User Guide. There were 2 contrasts (Pre-Post , and Post -Pre) >> >> Next I carried out randomise: >> >> randomise -i all_FA_skeletonised.nii.gz -o tbss -m mean_FA_skeleton_mask >> -d design_pairedt.mat -t design_pairedt.con --T2 >> >> Finally, I viewed the resulting tfce files (tbss_tfce_corrp_tstat1 and >> tbss_tfce_corrp_tstat2). Nothing survived the significance testing (fslview >> $FSLDIR/data/standard/MNI152_T1_1mm mean_FA_skeleton -l Green -b .3,.7 >> tbss_tfce_corrp_tstat1.nii.gz -l Blue-Lightblue -b 0.949,1) >> >> However, at this point I began experimenting with the cluster command and >> varying thresholds to get a feel for the level at which the two groups are >> significantly different. Just to play around I entered 50% (cluster -i >> tbss_tfce_corrp_tstat1.nii.gz -t 0.50 --mm > cluster_t1_50.txt). >> >> At 50%, the two groups of course have some differences. However, the >> significant clusters are not the same for both contrasts, T1 and T2 (Just >> to check: Does T1 indeed correspond with contrast1 and T2 with contrast2?). >> > > Yes. > > >> >> Why is this? I expected cluster_t2_50 to be the same as cluster_t1_50 >> but with the opposite sign. Perhaps I have made an error in calculations or >> is this what you would expect? >> > > The test statistic (t) for a contrast such as [1 -1] has the opposite sign > when compared to a contrast such as [-1 1]. The p-values not. If the > distribution of the test statistic happens to be perfectly symmetric around > zero, then the p-value for [1 -1] is the same as 1-p+1/nPerm for [-1 1], > although there might be slight variation given that we rarely do all > permutations exhaustively. > > Now, these relationships don't hold anymore for the corrected p-values, as > the distribution of the maximum is skewed, and don't hold for TFCE, which > is a non-linear function of the original test statistic. > > All the best, > > Anderson > > > >> >> Thanks very much, >> Sheena >> > >