Hi Tom,
Thanks a lot for making my (scientific) life that simple. Indeed, if
I can carry out the conjunction on the corrected p-maps, everything
becomes quiet easy.
Regarding my bug/feature point, I noticed, that without a brain
mask, the tstats output from a randomise call does not provide a
value for each voxel within the brain, but apparently only for
those, that were above some threshold. I did not look more into it
to be able to tell, whether this threshold is applied on the input
data or the output data.
However, when providing a mask with the -m option, I got a t value
for each voxel, including negative t-values that were missing
before. I created the mask by taking the absolute values of the
input data and binarising it, hence this selects exactly all voxel,
that are not constantly 0 for all time points. Since this simple
change lead to the behavior I wanted, I was happy with it, just
wondering if this might reflect a bug or was by purpose.
best,
wolf
On 05/13/2012 08:57 PM, Thomas Nichols wrote:
[log in to unmask]" type="cite">Dear Wolf,
For reference (& others on the list), here are the basics
of conjunctions: Inference on the 'conjunction null' requires
that all the (say, K) conjoined tests are individually
significant. Equivalently, the minimum statistic over all the K
tests must be significant when judged as a single test (judging
the minimum against a special 'minimum null distribution' gives
you inference on the 'global null', something different; see
Nichols et al, NI 25:653-660, 2005).
This is easy enough to implement for voxel-wise inference, as
you just take the intersection of all K thresholded maps; e.g.
taking the intersection of K maps thresholded at voxel-wise FWE
0.05 produces a FWE 0.05 voxel-wise conjunction inference. If
you are working with P-values instead of statistics, the
combining operation is the maximum over the K tests; with
randomise's convention of writing out 1-minus-P images, you flip
this and again need to take minimums.
Now, TFCE is cluster-informed voxel-wise inference, and so it
is as simple as combining the tfce_corrp images, taking the
maximum over the K 1-minus-P images. So, no need to make Z
images! (If you really needed them, you could convert the
uncorrected tfce_p images into Z's with fslmaths's -ptoz.)
Finally, on your comment
Btw,
is it a bug or an intended feature, that randomise produces an
thresholded output for the t statistics if no mask is
specified, but when for example a brain map is provided, I
would get the complete t-statisticcs, including negative
values?
I'm afraid I can't parse this. Randomise doesn't produce
thresholded output... that's for you to do with visualization or
with fslmaths. Do you mean the analysis mask? It's crucial
that you supply an analysis mask, otherwise you'll be analyzing
everything including non-brain voxels.
-Tom
On Thu, May 10, 2012 at 10:37 AM, wolf
zinke
<[log in to unmask]>
wrote:
Hi,
I want to carry out a conjunction analysis on the results of
randomise, and apply tfce to the results.
To do so, I ran randomise, converted the uncorrected p maps
to z maps, and took the minimum from both compared maps. I
fed this minimum statistics map into fslmaths with the -tfce
option.
So, now my maybe a bit naive questions: How can I convert
the resulting tfce-map into a map of p- or z-values? In the
paper from 2009 it was stated: "For inference, the TFCE
image can easily be turned into voxel-wise p-values (either
uncorrected, or corrected for multiple comparisons across
space) via permutation testing.", but somehow I am missing
the clue how to implement this.
I also want to obtain the corrdected p stats, that I would
get when using the tfce option within randomise, is there an
easy way to process the output from fslmaths in order to get
such maps? I would be happy about any pointer.
Btw, is it a bug or an intended feature, that randomise
produces an thresholded output for the t statistics if no
mask is specified, but when for example a brain map is
provided, I would get the complete t-statisticcs, including
negative values?
Thanks for any suggestions,
wolf
--
__________________________________________________________
Thomas Nichols, PhD
Principal Research Fellow, Head of Neuroimaging Statistics
Department of Statistics & Warwick Manufacturing Group
University of Warwick, Coventry CV4 7AL, United Kingdom