Thanks paul, and just to be sure: the mat file with the nuisance variables input into fsl_regfilt, is it demeaned or not?

jeff

From: paul mccarthy
Reply-To: FSL - FMRIB's Software Library
Date: Tuesday, July 28, 2015 at 2:36 AM

To: FSL - FMRIB's Software Library
Subject: Re: [FSL] using fsl_regfilt to regress out confounds

Hi Jeff,

Aah, because you're interested in reporting the absolute FA/F2 values after stats, my previous advice of globally shifting the data doesn't really make much sense.  In this case, a better approach may be  not to use fsl_regfilt, and to just include your confounds in the design matrix that you pass to randomise.

Cheers,

Paul

On 28 July 2015 at 01:33, Jeff Stevenson <[log in to unmask]> wrote:

Hi paul, thanks for getting back to me. Seems like we are using the –f flag for cols correctly then. F2 is the crossing fiber component of FA from bedpost. So it can be small and noisy. Default thresh is 0.05. In our case values just above the default thresh go negative when filtered, Which is a nonsense value. Does randomise demean the in file? I thought not. We do demean the design matrix file with the behavior and covariates.

Some of those results, when covaried the traditional way, are significant with positive cluster means. When filtered and then tested the regions with low values are still significant but some of the mean values for the clusters are negative. Which is a problem to report. Are you suggesting that I add back in the lowest negative from fslstats -R to give a range >0? For all measures? GM WM F2 etc…

Cheers jeff

From: paul mccarthy
Reply-To: FSL - FMRIB's Software Library
Date: Friday, July 24, 2015 at 4:22 AM
To: FSL - FMRIB's Software Library
Subject: Re: [FSL] using fsl_regfilt to regress out confounds

Hi Jeff,

On 13 July 2015 at 17:02, Jeff Stevenson <[log in to unmask]> wrote:

Hi folks, I am trying to use fsl_regfilt on some tbss and vbm data and have 2 questons:

When removing 2 (or more) nuisance variables (gender etc) I demean both into a 2 col design.mat format, then use the flag –f “1,2” to apply both cols ( or –f “1,2,3” for 3 cols etc)?

In other words are components “1,2,3” from ICA the same as columns “1,2,3” in a GLM design.mat?

Yes. If you are regressing using a melodic mixing matrix, '-f 1,2,3' refers to components 1,2,3. If you are regressing using a GLM design matrix, then '-f 1,2,3' refers to columns 1,2,3 of the matrix.

Secondly, even with only 1 variable when applied to F2 some of the voxels go negative. More so with 2. Should those simply be thresholded to 0 after or is there a constraint flag? How does fthresh and fthresh2 fit into this issue if at all?

I'm not sure what 'F2' is referring to, but the regression may result in voxel values dipping into the negative - it may or may not happen, depending on the regressor values, and on the raw data values.  The fthresh and fthresh2 options are irrelevant here - they are only applicable when regressing against MELODIC components.

You shouldn't threshold the regressed data, but if you want your data to be > 0, it should be safe to apply a global shift to move all of the voxel values above zero (assuming that the data will be de-meaned during stats anyway).

Cheers,

Paul

 

Jeff