Hello,
I am performing Structural mean.centering task PLS analyses using DTI. I have 5 conditions (FA,MD,MO,L1,L23) and 2 groups(patients,controls). I am trying to detect group differences. I am wondering about the usage of different mean-centering type options.
1) The user manual says that, setting mean centering type to 1 boosts group differences. Can anyone explain what is happening behind this? is the grand mean of all 5 conditions (FA,MD,MO,L1,L23) for both the groups calculated and subtracted from each condition?
2) If the above is true, then how are the ranges of different conditions handled? Because FA ranges from 0 to 1; MO ranges from -1 to 1; L1 has a very low value (in the order of 0.004). How is the grand mean calculated and subtracted from each individual condition?
Thank you for your time
Best
Jay
Hi Jay,
the mean-centering approach is not as mysterious as it sounds. Its basically just removing the grand mean from the data set to focus on differences between groups.
consider the following scenario
mean group1 = 10
mean group2 = 20
If you run this through SVD you will essentially get one LV that will reflect the mean of both groups (15) and one the reflects the difference from the grand mean (5). Mean centreing focuses the solution on the differences. It also helps with the permutation test to reduce the resampling space to differences.
As for the different image types, the centering is done for each image separately in the structural MRI module so scaling differences between images should not matter.
Hi Randy,
Thank you for the reply. Now, it is clear how mean-centering focuses on the group differences.
After i perform structural PLS analyses using 5 conditions and 2 groups,, 5000 permutations, 500 bootstrappings, i see that LV1 is significant. But when i look at 'task PLS brainscore with CI' bar plot, i see that only two conditions show the group difference (example the bars corresponding to FA and MO are inverted in both groups) and the rest of the conditions are not visible at all (the height of the bar is close to 0).
1) The heights of the bars in the two conditions which show a group difference are different (FA has a bar with high value e.g 15 and MO has low value e.g 4, CI of both do not cross 0). The PLS tutorial says that the height of the bar is arbitrary and i should interpret only the direction. So, do i give equal weightage to FA and MO while interpreting the brain voxel pattern corresponding to this LV?
2)Why are the bars corresponding to the rest of the three condition close to 0 (almost 0)? I would expect that if FA goes high, MD and L2L3 would go low.
I am having some difficulty in attaching the image of the 'task PLS brainscore with CI' bar plot (which i am trying to describe) but i will try to attach it below.
BW
Jay
Hi Randy,
Thank you for the reply. Now, it is clear how mean-centering focuses on the group differences.
After i perform structural PLS analyses using 5 conditions and 2 groups,, 5000 permutations, 500 bootstrappings, i see that LV1 is significant. But when i look at 'task PLS brainscore with CI' bar plot, i see that only two conditions show the group difference (example the bars corresponding to FA and MO are inverted in both groups) and the rest of the conditions are not visible at all (the height of the bar is close to 0).
1) The heights of the bars in the two conditions which show a group difference are different (FA has a bar with high value e.g 15 and MO has low value e.g 4, CI of both do not cross 0). The PLS tutorial says that the height of the bar is arbitrary and i should interpret only the direction. So, do i give equal weightage to FA and MO while interpreting the brain voxel pattern corresponding to this LV?
2)Why are the bars corresponding to the rest of the three condition close to 0 (almost 0)? I would expect that if FA goes high, MD and L2L3 would go low.
I am having some difficulty in attaching the image of the 'task PLS brainscore with CI' bar plot (which i am trying to describe) but i will try to attach it below.
BW
Jay
Jay - can you send the image of the CI bar plot to my email account?
rmcintosh@research.baycrest.org
Baycrest is an academic health sciences centre fully affiliated with the University of Toronto
Privacy Statement - Disclaimer - © 1989-2024 BAYCREST HEALTH SCIENCE. ALL RIGHTS RESERVED