I've been running ANOVA on some VBM in randomise. I'm almost certain I've got the design and contrasts right based on Eugene's email of 27 May last (Re: Correct design setup for ANCOVA with nuisance variable).

So I've got 3 groups L, M, S and they are in the order of the ls command in the GM_mod_merg file.

The (abbreviated to 2 subjects per group) design and contrasts looks like this (I've been following the ANOVA: 1-factor 4-levels example on the webpage):

EV1 EV2 EV3
1 1 0
1 1 0
1 0 1
1 0 1
1 0 0
1 0 0

So assuming EV1 models the last group (S) the EVs are as follows
EV1: mean
EV2: L-S
EV3: M-S

Actually, the interpretation of EV1 is exactly 'S', but as you'll never test it (or the mean) it doesn't matter. Otherwise this is correct.

Contrasts:
EV1 EV2 EV3 F-test
C1 L-M 0 1 0 *
C2 M-S 0 0 1 *

So assuming I've got this right (i'd be grateful if some one could confirm this for me), I then proceed to run randomise using TFCE, threshold (and binarise) the tfce_corrp_fstat1 and tfce_corrp_tstat files at 0.95 and use the resultant masks on to the corresponding fstat and tstat files. Is this correct should I be thresholding the tfce files like this and masking in this fashion? (The exact commands are below).

I'm not sure why you're working so hard. Isn't it sufficient to view the regions of significance in the corrp TFCE image? Or is that you want to view the f-values within the TFCE significant regions? If so, such a masking approach is reasonable.

The problem I then have is that while I get regions showing a group effect from the fstat and some of these are shown in tstat1, the regions identified showing differences from tstat2 are not showing up in the thresholded fstat image.

The is not unusual at all. A F-test is testing more 'degrees of freedom' and will not always detect something that was significant in a single t image alone.

The fact that there regions showing in tstat2 that are not in fstat have me confused. Have I got something wrong with the method below or should non-overlap be expected?

Nope. Just basic stats. Consider even just a single degree of freedom F (i.e. on one t contrast). A one-sided T with significant P-value 0.04 will not be signifiant in the F, as the F is equivalent to the two-sided T, and will have P-value 0.08.

Hope this helps!

-Tom

Any insight on this would be gratefully appreciated.

randomise -i usersOnly_GM_mod_merg_s2 -o longVsMidVsShort -m GM_mask -d longVsMidVsShort.mat -t longVsMidVsShort.con -f longVsMidVsShort.fts -n 5000 -T

fslmaths longVsMidVsShort_tfce_corrp_fstat1 -thr 0.95 -bin mask/longVsMidVsShort_mask_pcorrected_fstat1
fslmaths longVsMidVsShort_fstat1 -mas mask/longVsMidVsShort_mask_pcorrected_fstat1 corrected/longVsMidVsShort_fstat1_corrected

for t in $( seq 1 1 2 ) ; do
if [ -f longVsMidVsShort_tfce_corrp_tstat${t}.nii.gz ] ; then
fslmaths longVsMidVsShort_tfce_corrp_tstat${t} -thr 0.95 -bin mask/longVsMidVsShort_mask_pcorrected_tstat$t
fslmaths longVsMidVsShort_tstat${t} -mas mask/longVsMidVsShort_mask_pcorrected_tstat$t corrected/longVsMidVsShort_tstat${t}_corrected
fi
done

Regards,
--
Colm G. Connolly, Ph. D.
Dept of Psychiatry
University of California, San Diego
9500 Gilman Dr. #0855
La Jolla, CA 92093-0855
Tel: +1-858-246-0608
Fax: +1-858-534-9450

--
____________________________________________
Thomas Nichols, PhD
Principal Research Fellow, Head of Neuroimaging Statistics
Department of Statistics & Warwick Manufacturing Group
University of Warwick
Coventry CV4 7AL
United Kingdom

Email: [log in to unmask]
Phone, Stats: +44 24761 51086, WMG: +44 24761 50752
Fax: +44 24 7652 4532