Hi Patricia,
it looks ok to me, though I am not sure why you would choose to follow the second model rather than the first one based on the previous answer, especially if you're only looking at the contrasts between the healthy group and every other disease group... You might want to add the -x option in randomise to output the uncorrected p-values.
Cheers,
Gwenaelle
De: Patricia Pires <[log in to unmask]>
Objet: Re: [FSL] Re : [FSL] Glm
Ą: [log in to unmask]
Date: Mardi 23 novembre 2010, 9h55
Hi Gwenaėlle,
thanks for your quick reply. Then I assume second approach is better than the first one for my model.
I would be safe if I have done well, could you please help me with this?
I have 4 groups to contrast FA values:
1.- Healthy (EV3)2.- Disease 1 (EV1)3.- Disease 2 (EV4)4.- Disease 3 (EV2)
If i choose second approach i have to run in Glm an ANOVA 4-group unpaired and not a t-test (as it would be the case of the firts approach), is that right?
My principal question is about the design matrix. I think that I have run well the EV's (4 EV's with their respectives 1 and 0 according to their corresponding group). I didn't add any covariate.
However I am not pretty sure if i did well in the "Contrast and F-test" option:
I did 6 contrast -->
/ContrastName1 Healthy (EV3) > Disease1 (EV1)
/ContrastName2 Healthy (EV3) < Disease1 (EV1)/ContrastName3 Healthy (EV3) >Disease2 (EV4)/ContrastName4 Healthy (EV3) <Disease2 (EV4)/ContrastName5 Healthy (EV3) >Disease3 (EV2)
/ContrastName6 Healthy (EV3) <Disease3 (EV2)
Matrix
EV1 EV2 EV3 EV4
-1.000000e+00 0.000000e+00 1.000000e+00 0.000000e+00 1.000000e+00 0.000000e+00 -1.000000e+00 0.000000e+00 0.000000e+00 0.000000e+00 1.000000e+00 -1.000000e+00 0.000000e+00 0.000000e+00 -1.000000e+00 1.000000e+00
0.000000e+00 -1.000000e+00 1.000000e+00 0.000000e+00 0.000000e+00 1.000000e+00 -1.000000e+00 0.000000e+00
I have not selected anything in the F-test option (i.e. F-test=0). Is that correct?
Then I saved the matrix.
Next, I have run randomise with this format:
randomise
-i all_FA_skeletonised.nii -m mean_FA_skeleton_mask.nii -o
resultados_TBSS -d matriz.mat
-t matriz.con
-n 5000
--T2
-V
Did I follow well all procedures? I appreciate very much your help.
Bests regards,
Patricia.
2010/11/16 Gwenaėlle DOUAUD <[log in to unmask]>
Hi Patricia,
here is what Tom (Nichols) said on this earlier this year:
"The only difference between the two approaches is the assumption of common error variance over the 3rd group if included (possibly bad), and a corresponding increase in DF (always good).
So there's no right answer...
The safe way is to only study the data needed (2nd approach) because if it happens that group C has wildly smaller variance you can get inflated significances (or reduced power if it has wildly larger variance, but still not 'accurate' inferences relative to 'truth').
However, if the 'master' inference is determined by the F-test across all groups, then it's fine to work with the big model - 1st approach - since you're depending on it's validity anyway."
Hope this helps,
Gwenaelle
--------------------------------------------------------------------
Gwenaėlle Douaud, PhD
FMRIB Centre, University of Oxford
John Radcliffe Hospital, Headington OX3 9DU Oxford UK
Tel: +44 (0) 1865 222 523 Fax: +44 (0) 1865 222 717
www.fmrib.ox.ac.uk/~douaud
--------------------------------------------------------------------
