On Thu, Oct 20, 2011 at 6:06 AM, Robert Schulz <[log in to unmask]> wrote:
> Hello,
>
> I try to perform a voxelwise simple regression analysis on normalized (MNI) FA maps (derived by FSL, non TBSS) of a group of chronic stroke patients. The aim is to identify (cluster of) voxels that are significantly correlated with a certain behavioural measure:
> One number for each subjects.
>
> What I did so far:
>
> Specify 2nd level analysis: One-sample t-test
> Scans: individual FA maps (smoothed 3x3x3)
> Covariates: Vector nx1 with relevant behav measure, no interactions, centering overall mean, no masking, no norm
>
> So GLM: Individual scans - Beh. Covariate
>
> No I am not sure about the contrast to apply:
>
> 1) t [0 1] to address correlation there are clusters outside the brain p uncorr 0.001; FWE 0.05 not a single cluster. Shouldn't this contrast particularly ask for sign. voxels over the group whose FA values are positively correlated with the covariate?
Yes. That is what the contrast is testing. The reason you get stuff
outside the brain is because the values outside the brain are not 0.
If they were all 0, then you would not get anything. I'd just use a
brain mask and look at the voxels inside the brain mask. Since your
looking at FA, you might want to use a mask where the probability of
WM is higher than GM or CSF. Alternatively, you can create a mask
based on the average FA image across your subjects.
>
> 2) t [1 1]: what does this mean, because here I get quite plausible clusters that could be correlated with motor outcome.
It means the average of the slope of the covariate and the intercept
of the group, which is meaningless. It can not be interpreted.
>
> If I missed the correct contrast so far, what would it be?
>
> I am looking forward to your answers and comments.
>
> Kind regards
>
> Robert
>
|