[VoxBo] fdr in vbstatmap and f-tests in rfx analysis

[Daniel Y Kimberg]

David January wrote:
> 1) I'm trying to generate a false discovery rate statmap using
> vbstatmap for an rfx analysis.  At the command line, I'm entering the
> following:
> >vbstatmap <GLMdir> -o <outputfile> -c 1 -s tfdr
> This yields the following error:
> [E] vbstatmap: invalid scale tfdr.
> What am I doing wrong?

This is a little bit of a mess right now, the code tries to support a
variety of old and new ways to do these things.  In any case, "-s
tfdr" hasn't been supported for a while.  I think the best way is to
specify the relevant contrasts in your contrasts.txt file, something
like this:

tcontrast t vec 1
tpcontrast tp vec 1

and then do something like this:

vbstatmap glmdir -c tpcontrast -q 0.01
vbstatmap glmdir -c tcontrast -o map.cub

Remember that an FDR map is just a stat map that has been thresholded
to control the false discovery rate.  Typically it's a t map, although
to calculate the FDR threshold, you actually need p values.  I think
the last released version of vbstatmap is clever enough to do the
right thing when you request FDR and a t scale.  But I'm not 100%
sure, so the above method should work regardless.  When you use the -q
flag, vbstatmap knows to calculate a threshold that controls FDR, and
to print out the corresponding t threshold (not the p threshold).  So
if you want an expected false discovery rate of no more than 0.01, you
would use -q 0.01 (as in the example above).  When you run it,
vbstatmap will give you the t threshold that you can use to threshold
your t map.  Sorry this is so confusing.  I'll try to remember to
clean this up at some point soon, and to update the usage information
to be more helpful (and current).

> 2) For a different rfx analysis, I want to do an F test of the
> entire model (of 3 covariates).  However, when I specify scale = f
> in my vbprep script that generates the rfx .tes file, I get a
> uniform red brain.  When I specify t as the scale, I get a standard
> statmap.  I'm not mean norming the data; any ideas on what's going
> wrong in the f-version?

Where are the red values showing up, when you try to run your second
tier analysis?  If you look at the rfx tes file, are you getting bad
stat maps for each subject, or just for some?  Are the red values 0 or
NaN or something else?

dan