Dear Sharpley,
> I am using FSLVBM to correlate a variable of interest with areas of
> grey matter volume loss. I have 36 scans, 1 variable of interest and
> 1 demeaned variable of non-interest in the design. The correlation
> is being tested using a [1] t-contrast where decreasing scores on
> the behavioural measure is associated with decreased grey matter
> volumes.
>
> At the moment, I do not get areas of significance when viewing the
> images using the following command:
> fslview $FSLDIR/data/standard/MNI152_T1_2mm fslvbm_tfce_corrp_tstat1
> -l Red-Yellow -b 0.949,1
>
> However, I DO get areas of significance when viewing the
> tfce_p_tstat images even with a p value of .005, i.e.:
> fslview $FSLDIR/data/standard/MNI152_T1_2mm fslvbm_tfce_p_tstat1 -l
> Red-Yellow -b 0.9949,1
>
> Is it possible to report these results? If not, what can I do?
>
> I have tried voxel-based thresholding, and get similar findings
> where I get significant findings using _vox_p_tstat with a p value
> of .005 but not _vox_corrp_tstat.
it is of course "possible" to report, the question is if you want to
and if the reviewers let you get away with it.
I would personally be very reluctant to report uncorrected 0.005
results. If you lower the threshold on your fslvbm_tfce_corrp_tstat1,
at what level does this are show up? My guess would be that it
corresponds to a corrected level of ~0.5, or even lower. That means
that there is a rather great risk that this is a false positive.
My suggestion to you, and pretty much everyone, is to use directed
hypotheses. I.e. sit down and think hard about where you expect to
find your changes/activations before you do the analysis. You can then
use this to limit the part of the brain you are looking in through
masking in randomise. My naive guesstimate is that 90% of everyone
doing an imaging study has a quite good idea where they expect to find
something, but at the same time 90% perform their analysis as if they
didn't (i.e. they search everywhere in the brain).
My personal choice would be to perform 2 analyses, one with a directed
hypothesis/hypotheses through a quite restrictive mask covering only
the area/areas I have an a priori hypothesis about. And then to also
perform an analysis without a directed hypothesis. From both of these
analyses I would only report corrected results. Strictly speaking one
should perform yet another correction because one is reporting from
two analyses, but I think one can be a little permitting there.
> Are there papers which have published results using FSLVBM with a
> covariate of interest? I am having difficulty finding these. Perhaps
> these papers would be helpful in giving me some advice too...?
I am sure there are, though none comes to mind right now. However,
there is no real difference between a "covariate of interest" and a
variable coding for group.
> At the moment, I have found something very very similar to what I am
> doing but the authors used SPM2. They published their findings using
> "a threshold of significance of p < .001 uncorrected for multiple
> comparisons".
As I said above, there is a chance you would get away with it.
Especially if you could argue that this area is what one would expect
to find. But then we are again back at the directed hypothesis which
is better tested as suggested above.
Good luck Jesper
|