Thank you again for your prompt response Anderson.
But consider the context of strictly the single sample tbss_sym FA test (n=9) with one covariate age (demeaned):
-the randomize call tested for L>R and included the sign flip option (randomise -i all_FA_skeletonised_left_minus_right.nii.gz -o tbss_LgR –m mean_FA_symmetrised_skeleton_mask.nii.gz -d design.mat -t design.con -1 --T2);
- testing R>L by multiplying the input by -1 or alternatively multiplying the contrasts by -1
- the contrasts were C1: group mean (significance is > 0); C2 pos eft age; C3 neg eff age (EVs and contrast are below).
In this context C1 tests if significance is > 0 adjusting for age; C2 and C3 indicate if the covariate age is positively on negatively related to FA. As such corrp_tstat2 is not a mirror opposite of corrp_tstat1 Correct?
As your previous email made lucid, given a set of significant L> R coordinates, the same set of coordinates could be reported as R < L in a between group test.
However, in the single sample tbss_sym test the cluster reports (method : http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Randomise/UserGuide#Getting_Cluster_and_Peak_Information_from_Randomise_Output) for the L>R and R>L show different sets of significant coordinates. For example, here are the first 2 lines of L>G and R>L cluster reports for the expert sample; I have added the JHU-ICBM info associated with the coordinates in fslview.
L>R
Cluster Index Voxels MAX MAX X (mm) MAX Y (mm) MAX Z (mm) L>R, JHU-ICBM Intensity: 1-p
14 323 11.2 -30 41 -3 5% IFOF; 3% ATR 0.9628
13 40 9.42 -29 30 24 3% UN; 3% IFOF; 3% ATR 0.9687
R> L
14 3529 17.8 -31 -43 28 32% SLF; 11% SLF (temporal part) 0.998
13 434 9.23 -49 -35 -11 21% SLF (temporal part); 21% SLF; 3% ILF; 3% IFOF 0.9707
So while a given set of L> R coordinates could be reported also as R < L, the cluster reports (and the images associated with cluster report significant coordinates corrp_tstat images in fslview) indicate different main tracts, or aspects of main tracts, have significantly different FA in the left and right hemispheres. Is this a reasonable conclusion?
Single group, tbss_sym
Group EV1 EV2
1 1 -1
1 1 27
1 1 -4
1 1 -2
1 1 -4
1 1 -5
1 1 -3
1 1 -5
1 1 -3
Contrasts
group mean 1 0
pos eff age 0 1
neg eff age 0 -1
Hi Charles,Hmm, possibly, although the -1 should produce both contrasts (+1 and -1, as tstat1 and tstat2 respectively), so I think you could safely ignore R>L (or L>R, whichever you prefer).
All the best,
AndersonOn 5 June 2014 14:57, Charles Leger <[log in to unmask]> wrote:
Thank you Anderson. It would seem that the R>L test is really only useful in the context where the -1 option is used, mainly in the single sample tbss_sym analysis.
On Thu, Jun 5, 2014 at 2:38 AM, Anderson M. Winkler <[log in to unmask]> wrote:Hi CharliePlease, see below:On 4 June 2014 20:35, charlie <[log in to unmask]> wrote:
A between group tbss_sym FA analysis revealed just a single cluster (made of 15 significant voxels) for L>R for experts (EXP) vs controls (CON); contrast C2 below for corrp_tstat2. The max x, y and z mm coordinates were -21, -56, 39, C1 was ns for L>R.Ok, so this means that at (-21,-56,39) the left side has higher FA on experts when compared to controls.
By contrast , the R>L test showed C1 CON>EXP, corrp_tstat1, but at the same coordinates as C1 from the L>R test (-21, -56, 39). The C2 R>L (Exp>CON) test was ns. The coordinates -21, 56, 39 identified as the inferior longitudinal fasciculus (posterior portion) and the posterior part of the anterior thalamic radiation by the JHU white matter tractography atlas.
Ok, so this means that at (-21,-56,39), the right side has higher FA on controls when compared to experts. This is the same as saying that the left has lower FA on controls when compared to experts, which is the same as that left has higher FA on experts than controls, which is the same as above...
It is reasonable to expect some FA difference between experts and controls but it seemed strange that the difference would occur at the same coordinates. I thought I could have made an error multiplying the randomize L>G input by -1 to arrive at the R - L input for randomize. But I also tried a separate analysis deriving the R - L by multiplying the contrast by negative 1 with the same results.
Does this seem an appropriate result???Looks correct to me!In fact, no need to test both R>L and L>R. Just one of these is fine, because as you've seen earlier, these two images are identical, except for the opposite signs. You can invert the tests by testing the opposite contrasts, which you are also doing (hence the duplicate results).
All the best,
Anderson
All tests were run at 5000 permutations, cluster reports peak coordinates (following http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Randomise/UserGuide#Getting_Cluster_and_Peak_Information_from_Randomise_Output) agreed with visual inspection in fslview.
Evs and Contrast are below: there were not correlations between significant FA differences and age or gender (age and gender were demeaned, age-mean(age))
Group EV1 EV2 EV3 EV4
CON EXP Age Gender
1 1 0 -1.94 -0.22
1 1 0 26.06 -0.22
1 1 0 -4.94 -0.22
1 1 0 -2.94 -0.22
1 1 0 -4.94 -0.22
1 1 0 -5.94 -0.22
1 1 0 -3.94 -0.22
1 1 0 -5.94 -0.22
1 1 0 -3.94 -0.22
1 0 1 3.06 -0.22
1 0 1 3.06 -0.22
1 0 1 1.06 -0.22
1 0 1 2.06 -0.22
1 0 1 -0.94 -0.22
1 0 1 0.06 0.78
1 0 1 -0.94 0.78
1 0 1 -0.94 0.78
1 0 1 2.06 0.78
Contrasts
EV1 EV2 EV3 EV4
C1 CON > EXP 1 -1 0 0
C2 EXP> CON -1 1 0 0
Pos eff age 0 0 1 0
Neg eff age 0 0 -1 0
Eff of gender 0 0 0 1