On Thursday, October 3, 2013, Caspar M. Schwiedrzik wrote:

Hi Doug,

On Thursday, October 3, 2013, Douglas N Greve wrote:

It sounds like two issues:
1. p-values not consistent with your program. What did you use to compute? Did you do a two-sided (which is what fsfast uses)?

I used ttest in Matlab, two sided.

2. Using pcc maps. Why not use -m pcc?

Isn't that giving me a map per subject? How do I get the group map that is consistent with the results of mri_glmfit run on ces.nii?

Thanks, Caspar

doug

On 10/03/2013 01:10 PM, Caspar M. Schwiedrzik wrote:

Hi Doug,
I loaded the pcc.nii file that I got from isxconcat-sess into Matlab and then ran a t-test against 0 over the 4th dimension. I converted the resulting p-values to -log10 and then compared them to the output of mri_glmfit, namely sig.vol.
This was the mri_glmfit command:
mri_glmfit \
--surf averagesubject hemisphere \
--y pcc.nii \
--no-cortex \
--osgm \
--glmdir analysisname
I was expecting the p-values to be the same, which apparently is not the case, unless I am doing/understanding something wrong.

By now, I am actually more inclined to use the regression coefficients instead. However, I'd still like to get pcc maps from them, if there is a way to do so in FSFAST.
Thanks, Caspar

2013/10/3 Douglas N Greve <greve@nmr.mgh.harvard.edu <mailto:greve@nmr.mgh.harvard.edu>>

On 10/03/2013 10:39 AM, Caspar M. Schwiedrzik wrote:

Hi Doug,

when I run a two-tailed t-test against 0 in Matlab on the Rs
in pcc.nii that I get out of isxconcat-sess with -m pcc, and
DOF from ffxdof.dat, I get different -log10(p) values than the
ones that come out of mri_glmfit.

I don't understand what you mean. Can you elaborate?

I am not sure why this is happening.
In principle, I just want pcc maps as final output to show
them on the surface (instead of p-values). So I'd be happy to
follow your advice regarding the biasing effects of noise and
autocorrelation and use the regression coefficients. However,
mri_glmfit (v5.1) does not seem to output pcc maps of the
contrasts (contrary to selxavg3-sess on the single subject
level). How would I get those?

Thanks, Caspar

2013/10/1 Douglas N Greve <greve@nmr.mgh.harvard.edu
<mailto:greve@nmr.mgh.harvard.edu>
<mailto:greve@nmr.mgh.harvard.edu
<mailto:greve@nmr.mgh.harvard.edu>>>

On 10/01/2013 01:13 PM, Caspar M. Schwiedrzik wrote:
> Hi Doug,
> it would be great if you could give me some further
advise on the
> group analysis of functional connectivity maps.
> Specifically, I am trying to get PCC maps for certain
seeds, and am
> not planning any comparison between groups.
> Following your previous advise, I am running
isxconcat-sess with -m
> pcc to get the PCC maps.
> I would then run
>
> mri_glmfit \
> --surf averagesubject hemisphere \
> --y pcc.nii \
> --no-cortex \
> --osgm \
> --glmdir analysisname
>
> *Could you please provide some more detail on what kind of
analysis is
> performed when I provide pcc.nii as an input for
mri_glmfit? Is it a
> t-test of the Fisher-transformed r-values against 0?
I just run a t-test of the r-values. I don't have a
program to convert
them to z-values, however, there are z-values that are
created in the
first level analysis. These are generated from the
p-values but I
bet it
would give you the same thing. Use -m z with
isxconcat-sess if you
want
to use the z.
> *Is the average r-value or z-value saved somewhere?
Which level? For mri_glmfit, they are not, but it is not
hard to get
them with matlab.
> *Do you take the autocorrelation into account (as in
Vincent JL et
> al., 2007. Intrinsic functional architecture in the
anaesthetized
> monkey brain. Nature. 447:83-86)?
Not usually, but it could be done by not including
-no-whiten when you
run mkanalysis-sess. I usually use the regression
coefficients instead
of correlation coefficients because that they are at least
unbiased with
respect to noise level and autocorrelation.
doug

> I'd also be happy to look this up but I'd need to know
where I can