Hi Anderson,

Again. Some questions concerning the results.

The study actually is a longitudinal study. Group A is the baseline FA suntract followup FA of control group, and Group B is subtraction of treatment group.  Based on the design.mat and design.con, is my understanding correct? The tbss_tfce_corrp_tstat1 indicates the FA subtract of ctr group is larger than the one of treatment group,  tbss_tfce_corrp_tstat2 means  FA subtract of ctr group is smaller than the one of treatment group, right?

What does the  tbss_tfce_corrp_tstat3 and tbss_tfce_corrp_tstat4 mean?

Best regards,

Yawu  

On Wed, Jul 8, 2015 at 1:34 PM, Liu Y <[log in to unmask]> wrote:

Hi Anderson,
Many thanks for such an awesome advice!
Best regards,
Yawu

On 08/07/2015 1:00 pm, "Anderson M. Winkler" <[log in to unmask]> wrote:

Hi Yawu,

Yes, in that case, then it's possible to compare. And even better if there is an equal number of subjects of each group in each site. ITo get around the warning message you have two mutually exclusive options:

1) Assume that the variances for both sites are the same. In that case, simply ignore the warning, and run randomise making sure that the option "-e" is not used, thus allowing all subjects to be permuted freely across groups and sites.

2) Not assume that the variances for both sites are the same. In that case, use the "group" column to define one such "group" per site (not per experimental group, but actual scanning site). The warning message will appear again: just ignore it. If the number of subjects in each site is almost the same (you have 85 subjects overall, so if about 40-45 subjects are in each site), then you can use randomise, now including the option "-e design.grp". If, however, the number of subjects in each site is too different (e.g., something like about 30 in one site and 50 in the other, or anything more extreme than that) then randomise cannot be used. Instead, use PALM, with the options "-eb design.grp" and also "-vg auto", so that the different variances for each site are taken into account.

For either of these cases, the design and contrasts as you have are fine.

There could be other complicated cases (variances for each experimental group being different, or both groups and sites with different variances, being four variance groups overall). These can be dealt with in PALM if needed, but other modifications would need to be made then.

All the best,

Anderson

On 8 July 2015 at 10:11, Liu Y <[log in to unmask]> wrote:

Hi Anderson,

Maybe I did not say clearly in y previous emails. Our data are from two sites, the Group 1 contains two sites' data, and Group 2 also contains two sites' data. In this situation, we can compare the two groups, right?

Thanks,

Yawu

On Wed, Jul 8, 2015 at 11:26 AM, Anderson M. Winkler <[log in to unmask]> wrote:

Hi Yawu,

If each experimental group was acquired at a different site, it's not possible to separate site effects from group effects. Any difference found could be ascribed to scanner differences, and nothing can be said about the groups.

So, no analysis can be done comparing groups.

This is unrelated to the warning message, but addressing the message isn't relevant because of the issue above.

All the best,

Anderson

On 8 July 2015 at 08:20, Liu Y <[log in to unmask]> wrote:

Dear Anderson,

Thanks for instant reply. I have read the link and followed the instruction, but when I saved the design in Glm, it warned me 'Problem with processing the model: Warning - design matrix uses different groups (for different variances), but these do not contain "separable" EVs for the different groups (...)'

In our study, each group has different site data. 

Best regards,

Yawu

On Wed, Jul 8, 2015 at 9:59 AM, Anderson M. Winkler <[log in to unmask]> wrote:

Hi Yawu,

Please, see this example in the GLM manual:
http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Two-Group_Difference_Adjusted_for_Covariate

Only note, though, that as shown, all subjects from one site seem to be in one group, whereas all subjects from the other site seem to be in the other group. If the design is entirely like that, it isn't possible to disambiguate the effects of group vs site.

All the best,

Anderson

On 8 July 2015 at 07:54, Liu Y <[log in to unmask]> wrote:

Dear FSL experts,

I would like to compare the FA difference between group 1 and group 2 with adjusting age, gender, study site, and education years (all demeaned).

I wonder if my design.mat and design.con are correct? Thanks in advance.

Mat:

/NumWaves 6

/NumPoints 85

/PPheights 1 1 0 0 0 0

/Matrix

group1 group2 site education sex age

1 0 -0.32  3.92 -0.44 4.66

1 0 -0.32  -1.08  0.56 -3.17

1 0 -0.32  2.92 0.56 1.52

.

.

.

0 1 0.68 1.92 -0.44  -1.82

0 1 0.68 1.92   -0.44     -5.54

con:

/ContrastName1 group A > group B /ContrastName2 group B > group A /ContrastName3 group A mean /ContrastName4 group B mean /NumWaves 6 /NumContrasts 4 /PPheights 1.000000e+00 1.000000e+00 1.000000e+00 1.000000e+00 /RequiredEffect 0.848 0.848 0.650 0.544 /Matrix 1.000000e+00 -1.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 -1.000000e+00 1.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 1.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 1.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00

BR,

Yawu