[VoxBo] group F-test failure and FDR

[David January]

Hi Dan,

I'm running VoxBo off the Penn CfN cluster, so I assume the versions
are up to date.  That said, I'm running vbstatmap 1.8.5pre/Aug 30
2007.  To change the scale from t to F, I take the same file and just
change the letter following scale in the .vbp script, so everything
else is the same.

When I use voxel surfer and display the F ratio for same contrast, I
don't have any subjects that turn up as NaN.  In general, when I try
to view a glm in vbview, I get a uniform color brain, whereas when I
view them in voxel surfer I get something intelligible.  What would
that mean?  I will say, however, that I'm getting a uniform red brain
from this group F-test in voxel surfer as well.

-David


On Nov 29, 2007 10:11 PM, Daniel Y Kimberg <kimberg at mail.med.upenn.edu> wrote:
> David January wrote:
> > Does anyone know what would cause a subject's statmap made as part of
> > a group-level/random effects analysis to read NaN when the scale for
> > the statmap is set to F (and not, for what it's worth, to t or
> > beta--which I'm taking to indicate that all of the low-level structure
> > things like have the covariates in the same order across subjects
> > (which I checked independently anyway) are ok)?
> >
> > Also, does anyone know why vbstatmap would give me that same threshold
> > for an fdr of .05 as for .5?
>
> I'm a little stumped.  Could you check a few things?  First, see if
> you get the same behavior or different generating the stat maps in the
> IDL-based voxel surfer.  Second, could you double-check that you're
> running the latest version of vbstatmap (run it with no arguments,
> version is up top).  It's possible the red map problem would clear up
> with a newer version.  Also, just to be sure, you're using the same
> contrast vector for the t and F maps, right, it's just the scale
> that's changing?
>
> Weird things can happen with FDR thresholds sometimes, it depends on
> what your stat maps look like.  But if you're running an old
> vbstatmap, there was a fix at one point that improved the FDR
> calculation quite a bit.
>
> dan
>



-- 
David January
Department of Psychology
University of Pennsylvania