Hi Shengwei,

I don't see anything incorrect in your design or in the randomise options. It looks like there is no significant effect with this design.

However, I also note that there is a considerable correlation between the EV of interest and various others, so that any true effect, if existing, may not be disambiguated. For instance, the diagnosis in your sample has a considerable correlation with age (-0.62) and with sex (0.35). It has also some correlation with race (-0.26) and WMH (-0.17).

The issue with age, sex and race can be solved by recruiting more balanced groups. The WMH cannot be checked when recruiting, but maybe it doesn't need to be included in the design, as FA, WMH and, perhaps, the diagnosis being investigated may all share a common pathophysiological process in the WM. Or maybe the diagnosis simply doesn't explain FA reduction any more than what is already explained by the WMH.

All the best,

Anderson



On 12 May 2014 15:28, Shengwei Zhang <[log in to unmask]> wrote:
Hi Anderson,

Here's the link with all the information:
https://drive.google.com/file/d/0B1-hPZ22Vq1VOHRtaGZydDVsLUk/edit?usp=sharing

The randomise command was:
randomise -i all_FA_skeletonised -o tbss -m mean_FA_skeleton_mask -d design.mat -t design.con --T2

Thank you!
Shengwei


On Sat, May 10, 2014 at 6:25 AM, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Shengwei,
Could you please attach the design files (.mat, .con, etc) and paste here the exact randomise call you're using?
Thanks
All the best,
Anderson


On Friday, May 9, 2014, Shengwei Zhang <[log in to unmask]> wrote:
Hello,

I'm running randomise in the last step of TBSS but not sure how to configure the design matrix using GLM. The _corrp image showed nothing significant, which is unexpected.

The tbss_1 to 4 steps were run successfully. The model used is that FA is linearly related to the a) presence of disease (0 or 1), b) age, c) sex, d) education, e) race (1 or 2), and f) specific brain region volume normalised.

Here's the input to Glm using the latest version of FSL V5.0.6:
1. GLM Setup: "higher level/ non-timeseries design" was chosen, and # inputs equals the number of participants (about 100);
2. "EVs" tag in GLM: # EVs = 7 (6 covariates in previous paragraph and 1 error term), no voxel-dependent EVs, corresponding values were pasted in the order that the participants appeared in the all_FA_skeletonise (for the error term it's all ones);
3. "Contrast & F-tests" tag in GLM: 1 contrast, 0 F-tests, only the covariate of interest (e.g. a or e) set to 1 (assume positive correlation) or -1 (negative correlation) and other EVs set to 0.
4. repeat 1-3 for different covariates of interest and different type of correlation.

Then the design matrix was saved and randomise run.

Is this correct? I read the Fslwiki page about GLM but didn't find the answer. Any help is appreciated.

Shengwei