I am trying to use TBSS with "randomise" and have some problems understanding my results.
Essentially I get significant differences over my entire FA-skeleton with a two sample t-test
between my patient and control group (I don’t believe on such strong significant difference –
the differences, if at all, should be very subtle.)
To better get a grip on what is going on, I decided to create some simulated FA-data
(not on skeletonised but the “raw” FA-data! )
1. I chose one 3D-FA data set with good SNR and filtered it further to minimize variance
across all voxels (filename: FA_001)
2. generated 10 “subjects” with the same FA-data from 1.:
fslmaths FA_001 –add 0 FA_002 (and so on to FA_010)
3. merged all individual 3D-FA data into one 4D data set:
fslmerge –t all_FA `$FSLDIR/bin/imglob FA_*`
4. used the “randn” in fslmaths to add Gaussian noise to “all_FA” with SNR=100
fslmaths FA_001 –mul 100 –randn –div 100 all_FA_noise
5. checked all_FA in fslview -> looks ok
6. created mean and mask from all_FA_noise:
mean_FA; mean_FA_mask
7. did a one sample t-test (mean)
randomise –i all_FA_noise –m mean_FA_mask –o OneSampT_FA -1 –T (5040 realizations)
I actually used randomize_parallel on a 4 core machine:
randomize_parallel –i all_FA_noise –m mean_FA_mask –o OneSampT_FA -1 –T
(21 fragments of 240 permutations; 5040)
8. results from 7. :
OneSampT_FA_tfce_corrp_tstat1
OneSampT_FA_tfce_p_tstat1
OneSampT_FA_tfce_tstat1
And here is where my confusion starts:
According to the FSL-manuals this should be a “test for activation” – so in my interpretation a test to check if
there are no significant differences of the mean across all “timepoints” (or “subjects” in case of TBSS).
Does the OnesampleT-test also produces (1-p) or p-values?
OneSampT_FA_tstat1 is supposed to contain the probability distribution for the one sample test, is this
correct? The histogram looks ok and stretches in a asymmetric Rayleigh-like distribution from 0 to
ca. 200. High values are in WM, low values are in GM and CSF. Essentially it looks like a noisy FA-map.
So I assume this is not "demeaned" is it?
There is almost no virtual difference between “*_tfce_p_tstat1” and “*_tfce_corrp_tstat1”
except that the first (are those the uncorrected p-values???) has a smooth histogram and
the latter a “discrete” binarised histogram between 0…1 but almost all values are at 1.0 or very closed to it.
They both look almost like the mean_FA_mask (all pixels are closed to 1 and there is no contrast)
Now if this is supposed to be (1-p) I am confused here. My data are virtually identical except for noise
(but with a very high SNR=100).
That means I should find a large p for testing the difference of the mean between all data (timepoints), am I right?
If “*_tfce_corrp_tstat1” is really (1-p)as the manual states, the opposite would be true for the simulated data – a strong differences between the mean of my data - which is impossible since they are generated from the same source
with very high SNR.
Any help highly appreciated,
Burkhard.
|