Hi Alain,

Thanks for sending.

What happens is that even in the "best" voxel, F1 is too far from the tail (i.e., the significant part of the distribution). The peak voxel has an F-statistic of just 0.388. With df1=2 and df2=475, this means that when this statistic is converted to a z-score before the computation of TFCE, the max has z-stat of -0.441, which is negative, and so TFCE can't even be computed.

This suggests lack of effects for this contrast, which can probably be removed from the analysis.

All the best,

Anderson

On 23 October 2015 at 08:44, Alain Imaging <[log in to unmask]> wrote:

Hi Anderson, 

I'be just uploaded a rar file with the images and the matrix I used.

Please note that due to the inclusion of covariable of no interest in the matrix, the voxel-dependent EV are now the 10th,11th and 12th.

Alain

---

De : FSL - FMRIB's Software Library <[log in to unmask]> de la part de Anderson M. Winkler <[log in to unmask]>
Envoyé : vendredi 23 octobre 2015 09:13

À : [log in to unmask]
Objet : Re: [FSL] voxelwise EV in randomise

 

Hi Alain,

Could you please (quite exceptionally) upload your input files? You should receive in a few minutes an email with instructions on how to do it.

Thanks.

All the best,

Anderson

On 22 October 2015 at 13:30, Alain Imaging <[log in to unmask]> wrote:

It seems that I can not stop seeking help.

I launched randomise with the following command

-i Cognition.nii.gz -m L_Puta_mask.nii.gz -o ./L_Puta_randomise/L_Puta -d Group_model.mat -t Group_model.con -f Group_model.fts --T2 --vxl=4,5,6 --vxf=L_Puta_Int_alelle0Shape.nii.gz,L_Puta_Int_alelle1Shape.nii.gz,L_Puta_Int_alelle2Shape.nii.gz

When the program start the permutation, after having calculated the statistic for the unpermuted data it gives me a warning saying:

"Warning: The unpermuted statistic image for the current image contains no positive values, and cannot be processed with TFCE. A blank output image will be created."

This happen for the first of the contrast that you suggested, that is the f contrast based on the following t contrast

1 -1 0 0 0 0

1 0 -1 0 0 0

But not if I skip the first contrast and I run directly the second f contrast

0 0 0 1 -1 0

0 0 0 1 0 -1

If I understood correctly, this kind of warning is related either to bad contrast or to something wrong in the images. In this case I don't think the contrast is wrong, and the fact that the second f contrast works make me think that the images are fine too.

Just for clarification, the input image is a 4d volume containing the mask of the putamen multiplied by the (demeaned) cognitive score of each subject.

Have you got any idea about what could have gone wrong ?

Thanks

A

---

De : FSL - FMRIB's Software Library <[log in to unmask]> de la part de Alain Imaging <[log in to unmask]>
Envoyé : jeudi 22 octobre 2015 14:07

À : [log in to unmask]
Objet : Re: [FSL] voxelwise EV in randomise

 

Hi Anderson, 

just a quick followup question.

I decided to use the design you suggested (6 columns, 1 for each of the 3 group and 3 for the interactions group x shape). To do this, I used the fsl_glm gui, entering 3 for the opetion "additional voxel-dependent EVs" and choosing the correct images. I then entered the proper EVs, the contrasts and saved the design.

If I open the .mat file, I see that there are 6 columns: the first 3 are the columns that I entered in the gui, but I wonder from where the value in the last 3 columns came, and if they have any meaning or importance.

Do they interact with the 4d images that I entered in the randomise command line  (--vxl=4,5,6 --vxf=myfile1,myfile2,myfile3) ?

Thank in advance

A

---

De : FSL - FMRIB's Software Library <[log in to unmask]> de la part de Anderson M. Winkler <[log in to unmask]>
Envoyé : jeudi 22 octobre 2015 09:01
À : [log in to unmask]
Objet : Re: [FSL] voxelwise EV in randomise

 

Hi Alain,

Please, see below:

On 21 October 2015 at 16:05, Alain Imaging <[log in to unmask]> wrote:

Hi everybody,

in the context of a shape analysis using FIRST, I want to observe the effect of shape, SNPs and their interaction on a cognitive variable.

Anderson Winkler and Mark Jenkinson suggested that the correct way of doing this would be to test a model in randomise using cognition as a dependent variable.

In particular, the model will look like this

Cognition = SNP + Shape + SNPxShape + a bunch of covariables like sex etc

To do so, i need to use some voxelwise EV in my models, and being the first time, I am getting confused.

The first thing that I did is multiplying the structure of interest mask for the cognitive measure of intereset for each subject, and then create a 4D volume. This volume will be the dependent variable in the randomise model (i.e. the input volume).

Then, for the interaction between shape and SNP, I would need to multiply the shape volume for the SNP, again for each subject.

Everything up to here is fine. And it's also possible to test if there is a linear increase/decrease based on the number of copies of one of the alleles, i.e., coding SNP as a single EV with 0, 1 and 2.

 

Now, in this analysis SNP is coded as group (3 levels), rather than as a continuous variable.

This is fine, and then the test will be of all possible pairwise combinations of the three levels, which can be done with 6 t-tests, or with an overall F-test.

 

This make unclear to me what I should do. I am thinking of coding the three groups with a dummy coding (i.e. two columns, 0 0 for group 1, 1 0 for group 2, 0 1 for group 3) and then multiplying the shape volume for these two columns.

This works, but I think a simpler solution would be:

EV1: 1 for group 1, for no copies of the minor allele, 0 otherwise.

EV2: 1 for group 2, for one copy of the minor allele, 0 othewise.

EV3: 1 for group 3, for two copies of the minor allele, 0 otherwise.

EV4: Shape * EV1

EV5: Shape * EV2

EV6: Shape * EV3.

For the contrasts:

C1: 1 -1 0 0 0 0 (for group 1 > group 2) [F1]

C2: 1 0 -1 0 0 0 (for group 1 > group 3) [F1]

C3: 0 1 -1 0 0 0 (for group 2 > group 3)

C4: 0 0 0 1 -1 0 (for Shape in group 1 > Shape in group 2) [F2]

C5: 0 0 0 1 0 -1 (for Shape in group 1 > Shape in group 3) [F2]

C6: 0 0 0 0 1 -1 (for Shape in group 2 > Shape in group 3)

C7: 0 0 0 1 1 1 (for positive correlation of shape and cognition)

C8: 0 0 0 -1 -1 -1 (for positive correlation of shape and cognition)

All the t contrasts above can be repeated in the opposite direction. The two F-tests are marked (F1 and F2), and these will test respectively any difference in cognition due to the SNPs in any direction (F1) or interaction SNPs and Shape (F2).

This model is similar to this one from the manual: http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Two_Groups_with_continuous_covariate_interaction

The difference is that there are 3 groups instead of 2.

 

In this way, the model in randomise would look like this:

Cognition(4D volume) = SNPDummy1 + SNPDummy2 + shapeXSNPDummy1 (4D volume) + shapeXSNPDummy2 (4D volume) + covariates

This model needs an extra EV with just the shape (without splitting). It also needs an intercept, or that the EVs are all mean-centered, as well as the data (you'd generally mean-center before making the products). But use the above example as it's probably simpler.

 

Does this make sense ?

I have more doubts:

1) Should I demean the two dummy columns that code for the group ?

As above.

 

2) Is correct to leave SNPDummy1, SNPDummy2 and all the other covariables as usual EV in the .mat file ?

Yes, just the Shape and the Shape*SNP would need to be voxelwise.

 

3) Since I am mixing EV in the .mat file and voxelwise EV, how can I construct the contrasts of interest ? In which order the voxelwise EV are entered in the model ? 

As above.

Cheers,

Anderson

 

Thank for your time

Alain