Hello,
I would like to run a correlation between the FBA metrics and some cognitive score. I have already run this for standard DTI maps (e.g. FA) by Randomise, however, I do not know if possible to run it by fixelcfestats.
I thought a method that can help this kind of analysis. I would like to describe here the pipeline and understand what you think.
I transformed the FBA metrics in voxel-based metrics in this way:

fixel2voxel [input.mif] none [output.nii.gz]

In this case, for each subject in the study I have images with different volumes (e.g. 12).
For each volume I created a 4D images and I ran the correlation with my score for each volume (12 times) by using different software, for instance Matlab or R by a linear model. As results, I had 12 correlations maps in nifti format (in one 4D file). After I converted this file to mif:
2) voxel2fixel [correlation.nii.gz] dir dir [correlation.mif]
For the correlation I need only the t-value and not the p-value with its corrections.

Because I am not able to run correlations by fixelcfestats I tried this method by using a simple t-test between two groups. Before I ran the t-test by fixelcfestats, after by Matlab by using the nifti file. The results were the same.

Why do you think you cannot run correlations with fixelfcestats? The command can handle any GLM that can be defined by a design matrix. You can run a regression between a fixel metric and your cognitive score, Using the traditional FBA approach (e.g. here), you can preprocess your data in preparation for regressions.

Hi Steven,
thanks a lot for your reply.
Ok. I can run directly the correlation by using fixelfcestats, however I am not sure about the design matrix.
E.g. I have 10 subjects and I want to correlate a FBA metric with a cognitive score by using age and gender as covariate.
Is it correct this matrix:

The matrix looks mostly fine, but you may want to consider adding a global intercept column (all 1s), and scaling your score and age columns. That is, center them to mean 0 and rescale to a variance of 1. It helps the matrix conditioning for modeling to run better, but for interpreting the beta coefficients, you will need to keep in mind that you appleid this scaling.

For the contrasts, I think you only need the first line.

randomize will be performing some form of cluster enhancement (presumably TFCE by default), and it will presumably be doing so on each volume individually. But for two fixels in adjacent voxels that belong to the same bundle, they could be assigned to different volumes, depending on the relative sizes of other fixels in each of those respective voxels. So the statistical enhancement would not behave as one would want.

For the contrasts, I think you only need the first line.

Only if you are exclusively interested in pathways in which there is a positive correlation between FBA metric and cognitive score. If you wish to be sensitive to the prospect of a negative correlation, you need both lines, to perform two independent t-tests.

randomize will be performing some form of cluster enhancement (presumably TFCE by default)

Yes, it is true, but if I convert the fixels to voxels in different volumes and I run, by Matlab for example, the correlations (or linear model) in each voxel and in each volume, and then I re-convert my outputs to fixel?
I am not interested to p values corrections, but only to the t values in each voxels.

Again, I tried this by using a t-test between two groups. Before I run it by fixelcfestats and then by converting all volumes to voxel and by run the t-test by Matlab. I re-converted my output (from Matlab) to fixel and I had the same results for t values (I used mrview). Logically, not for the p-values, because the corrections is different between the two methods.

For the contrasts, I think you only need the first line

Yes, I know that. I will try to run both contrasts, therefore, 1 and -1

If I convert the fixels to voxels in different volumes and I run, by Matlab for example, the correlations (or linear model) in each voxel and in each volume, and then I re-convert my outputs to fixel?

You would need custom code to re-project the voxel-wise values back into the fixel directory format, preserving the order of fixels within each voxel and skipping voxel values that are null-filled in the process of producing the 4D image; but yes, it could be done (and it sounds from your later text that you already have?).

If you are doing a strictly fixel-wise operation, with zero knowledge of spatial location or connectivity, then you could simply feed the 1D fixel data files into Matlab and operate on those directly; there’s actually no theoretical need to project those data back to “a spatial parametric map”.

Yes, the key motivation for using the GLM in fixelcfestats rather than an alternative is the fact that the resulting test statistics go straight into CFE, which is not implemented elsewhere to my knowledge.

and it sounds from your later text that you already have?
yes, I have already tried it by with a t-test, not with correlations, because I did not know if was possible t use fixelcfestats for correlations

If you are doing a strictly fixel-wise operation, with zero knowledge of spatial location or connectivity, then you could simply feed the 1D fixel data files into Matlab and operate on those directly; there’s actually no theoretical need to project those data back to “a spatial parametric map”.
Yes, I could also try it…

Ans, also, yes, I know about the CFE. In Matlab is quite impossible to use it.