Dear Marcus,
> I have done two different TBSS analysis of one groupe of patients
> using each time one of two subscores of a neuropsychologcal score as
> a regressor. I found sign. difference modulation of FA in some
> regions for subscore A, and some other sign modulation for subscore B.
>
> Now I would like to compare these two p_images, i.e.
> subscoreA_corrp_tstat.nii.gz and subscoreB_corrp_tstat.nii.gz
>
> More general my question would thus be how to do this.
> I would like to know if these two neurospychological subscoresA and
> B really measure different effects in different brain regions or
> not. Of course I could compare the p-images by simply looking at
> them, but I think that there must be a more precise way to do this.
I think that the best way for you to do this is by using a model
containing both scores. By performing an F-test on one of them you
will then see what regions are significantly "correlated" with that
regressor after everything that can be explained by the first
regressor has been removed.
Let me give an example to explain the principle. Let's say you have
some study in some group of subjects with some disease. Let us further
say you want to see what areas of the brain correlated with disease
duration (dd). But you would also like to know what areas correlate
with subject age, so you enter that too as a covariate. Chances are
that age and dd will in turn be highly correlated. That means that
there will be areas of the brain that correlate with dd and at the
same time with age. The question then is "was it dd or age that caused
this?". Our data cannot really help us with that, for all we know it
could have been dd, but it could equally well have been age. GLM will
then play it safe. So if you e.g. perform an F-test for age GLM will
first remove anything that could be explained by dd, and then look for
any remaining correlation with age. And vice versa if we test for dd.
This means that one can have the paradoxical situation that one has a
brain are that is highly correlated with both dd and age, but when we
test for either dd or age we get nothing.
This behaviour will of course mean that we suffer a loss of
sensitivity when we have multiple correlated regressors. But at the
same time it is the desired behaviour since we will not make "too
bold" claims.
I hope all this was clear.
> That was the moment when I was trying the "cluster" command in order
> to have "harder data" than just looking at colours. However I seem
> not to understand the values its output gives and I could not find
> any explanation anywhere.
It is in general not a good idea to compare statistics (a statistic on
a statistic is a "meta-statistic" and might be used e.g. when
comparing different statistical methods). The better way is to do as I
described above. What you suggest below is to base your test on
"difference" between your regressors on some summary statistics of
super-threshold voxels, and that will NOT be kosher. In general you
should always "compare then threshold", never the other way around.
> cluster --in=subscoreA_corrp_tstat.nii.gz --thresh=0.99
> gives me:
>
> Cluster - Index - Voxels - Z-MAX -
> 3 42 0.993 85
> 2 12 0.997 142
> 1 1 0.992 52
>
> which I read as
> cluster number - clustersize - ?? - ??
I'm no expert on the cluster utility, but it looks to me like 3 is
cluster index (arbitrary), 42 is the number of voxels with 1-p > 0.99
and 0.993 is the greatest 1-p value in that cluster.
> and then the
> Z-MAX X - Z-MAX Y - Z-MAX Z- Z-COG X - Z-COG Y- Z-COG Z
>
> which seem to be both coordinates, however I do not know what the
> MAX and the COG mean.
MAX indicates it is the coordinates of the maximum 1-p value. COG is
center of gravity.
Good Luck Jesper
|