Dear Colm,
> 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
