Hi Mark,

tbss_tfce_p_tstat.nii is indeed the map of uncorrected p-values, but
those p-values are not specific to a "cluster" in the traditional sense.
Rather, they are the p-values for a VOXEL.  This is why the p-value
differs from voxel to voxel.  In contrast, if you run a cluster extent
(-c option) or cluster mass (-C option) permutation analysis, you'll
find that the p-values are identical in adjacent voxels for which the
test statistic (t or f) was greater than the specified threshold (a
"cluster" in the traditional sense).

I wonder if part of the confusion is related to the following: 
While the TFCE test statistic (contained in your case in
tbss_tfce_tstat.nii) at each voxel is computed with "support" from the
surrounding voxels (with a tradeoff between height and extent determined
by the H and E values), the statistics are computed on voxel-values.
Thus, it is possible to have a p-value (corrected or not) that exceeds a
given threshold within a single isolated voxel, even though the
computation of the underlying test statistic was computed with support
from surrounding voxels.

Note also that the t-value represented by your design (in "tbss_tstat1")
is totally different from the TFCE test statistic itself.  This latter
test statistic can be obtained if you include the -R option in the
command line.  The tfce p-values are computed from the TFCE test
statistic, NOT from the t-value.  So yes, it is possible to have a t-
value at vox1 that is higher than vox2, yet the tfce p-value at vox1 is
less significant than at vox2.  It all depends on the integral of the
height times extent function that is computed to derive the TFCE test
statistic (see Smith and Nichols 2009).  The TFCE test statistic is not
a t-value, and does not have an associated degree of freedom.

Power is indeed probably your problem in losing "significance" going
from the uncorrected to corrected p-values (assuming an effect is indeed
present if you had a sufficient number of subjects).

As an aside, note that all those p-values are one-sided tests.  If you
did not hypothesize a direction a priori for the effect, then you would
need to account for the two-sided nature of your original hypotheses.

Hope that helps,
-Mike H.

On Thu, 2009-11-19 at 22:49 -0800, Mark Shen wrote: 
> Dear FSL experts,
> 
> I am still trying to understand whether the tbss_tfce_p_tstat output  
> from TFCE is returning uncorrected p-values that are specific to each  
> cluster that is found in a tract?  I am assuming this because some  
> voxels have a higher uncorrected t-value but yet a higher uncorrected  
> p-value than other voxels located in other tracts.  So each found  
> cluster follows its own t-distribution with its own degrees of freedom?
> 
> Also, does thresholding the tbss_tcfe_corrp_tstat to 1-P=.95 give the  
> clusters in tracts that survive a "p<.05 threshold, corrected for  
> multiple-comparisons?"
> 
> Finally, with a sample size of N=22 (11 in clinical group, 11  
> controls), I found several clusters in tracts that survive at the  
> uncorrected 1-P=.95 level for tcfe_p_tstat, but no clusters in tracts  
> that survive at the corrected 1-P=.95 level for tcfe_corrp_tstat.  Is  
> this simply a power issue?  I added more subjects to obtain N=34 (21  
> clinical, 13 controls), but I still only find tracts that survive at  
> corrected 1-P=0.8 level (but none at .95) for the tcfe_corrp_tstat.
> 
> Thank you for your time,
> Mark Shen
> M.I.N.D. Institute
> University of California, Davis
> 
> 
> On Nov 19, 2009, at 9:52 AM, Michael Harms wrote:
> 
> Re (1): Yes, it is the raw t-value, and will be the same as if you used
> if the -x (voxelwise) and -c (cluster) options.  In fact, if you use all
> options simultaneously, you get just a single <root>_tstat1.nii.gz file.
> 
> Re (2): Yes, <root>_tfce_p_tstat is the map of uncorrected p-values for
> the TFCE statistic (although I'm not sure how that is calculated given
> that the TFCE doesn't have a parametric distribution).
> 
> Re (3): Yes, <root>_tfce_corrp_tstat is the FWE corrected p-value.
> According to the Smith and Nichols (2009) paper on TFCE, it is based on
> the maximum (voxel-wise) TFCE value (across permutations).  Thresholding
> at 0.95 to 1 will give you VOXELS (which may or may not be contiguous to
> other voxels within that range) that survive the voxel-wise FWE test of
> the TFCE value.
> 
> cheers,
> Mike H.
> 
> On Thu, 2009-11-19 at 17:23 +0000, Mark Shen wrote:
> > I followed the TBSS v1.2 manual, and after creating .mat & .con  
> > files for a
> > two-sample unpaired t test, I ran the following randomise script  
> > for TFCE:
> >
> > randomise -i all_FA_skeletonised -o tbss -m mean_FA_skeleton_mask -d
> > design.mat -t design.con -n 500 --T2 -V
> >
> > Outputs generated:
> > tbss_tstat1
> > tbss_tstat2
> > tbss_tfce_p_tstat1
> > tbss_tfce_p_tstat2
> > tbss_tfce_corrp_tstat1
> > tbss_tfce_corrp_tstat2
> >
> > Questions:
> >
> > (1) Is tbss_tstat the raw, uncorrected t values?  If it is raw and
> > uncorrected, should this t statistical map be identical to the  
> > tstat outputs
> > from using voxel-wise or cluster-based thresholding options?
> >
> > (2) Is tbss_tfce_p_tstat the uncorrected p values?  Am I right to  
> > say that
> > this is not corrected for FWE multiple comparisons, but TFCE is  
> > returning
> > uncorrected p-values that are specific to each cluster that is  
> > found?  (I am
> > assuming this because some voxels have a higher uncorrected t-value  
> > but also
> > higher uncorrected p-value than other voxels located in other  
> > tracts.  So
> > each found cluster follows its own t-distribution?)
> >
> > (3) Is tbss_tcfe_corrp_tstat the FWE corrected p values?  Thus,  
> > thesholding
> > this to .95-1 gives the clusters that survive a "p<.05 threshold,  
> > corrected
> > for multiple-comparisons?
> >
> > I found several clusters that at 1-P=.95 level for tcfe_p_tstat,  
> > but no
> > clusters that survive at the 1-P=.95 level for tcfe_corrp_tstat.   
> > Is this
> > simply a power issue?  This analysis has N=22 (11 in clinical  
> > group, 11
> > controls), but I am adding subjects to obtain N=34 (21 clinical, 13
> > controls) and will ultimately have N=100 (70, 30).
> >
> > Thank you!
> >
> > Mark Shen
> > UC Davis M.I.N.D. Institute