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