See inline responses below.
On Fri, Oct 5, 2012 at 11:30 AM, Andre Knops <[log in to unmask]> wrote:
> Dear SPMers,
>
> I have a few questions about how to run a within-subject – GLM (aka repeated measures ANOVA in experimental psychology) in SPM8.
> I have a 3 factorial design (2 x 2 x 2) and all factors are within-subject, i.e. each participant takes part in all 8 conditions.
> First question: Is it correct that for the moment being there is no way in SPM to include all three factors plus a SUBJECT factor in a single GLM?
This is correct. However, GLM Flex by Aaron Schultz can handle your
more complex model.
>
> When restricting myself to a 2 x 2 design (collapsing over one factor) I did the following to set up the within-subject GLM (flexible factorial):
> 1st factor:
> Name: Subject
> Independence: Yes
> Variance: Unequal
> Grand Mean Scaling: No
> ANCOVA: No
>
> 2nd factor:
> Name: SESSION (sess)
> Independence: No
> Variance: Unequal
> Grand Mean Scaling: No
> ANCOVA: No
>
> 3rd factor:
> Name: OPERATION (oper)
> Independence: No
> Variance: Unequal
> Grand Mean Scaling: No
> ANCOVA: No
I would set all your variances to 'equal' as has been suggested in the
past. While it might be tempting to set the variances to unequal,
these setting are used to estimate the covariance patterns and correct
for non-sphericity. As they are estimated parameters, one has to be
careful that noise in the data isn't causing a poor estimate. I
suspect that more work on the variance correction approaches will be
published soon with the field moving toward more complicated
within-subject designs.
>
> Then, under ‘specify Subjects or all Scans & Factors’ I define for each subject the four contrast images corresponding to the factorial levels of the above factors, i.e.
> con_sess1oper1.nii
> con_sess1oper2.nii,
> con_sess2oper1.nii
> con_sess2oper2.nii (in this order).
>
> Under ‘Conditions’ I enter a 4x2 double
> 1 1
> 1 2
> 2 1
> 2 2
>
> I tell SPM to compute the main effects #1, #2, and #3, as well as the interaction term between #2 and #3 (my within-subject factors).
>
> As a result I get a design matrix with 8 conditions
> Sess1, Sess2, Oper1, Oper2, sess1oper1, sess1oper2, sess2oper1, sess2oper2
> and as many 'block regressors' as I have participants. No nuisance, no covariates defined.
>
> Question 2: Is this the correct way of setting up the GLM?
Yes.
>
> Then I compute the following t-contrasts to test the differential effects of SESS1/SESS2:
> 1 -1 0 0 .5 .5 -.5 -.5 (what is more active in sess1 than in sess2)
> -1 1 0 0 -.5 -.5 .5 .5 (what is more active in sess2 than in sess1)
> Or to see the impact of OPER1/OPER2:
> 0 0 1 -1 .5 -.5 .5 -.5 (what is more active in oper1 than in oper2)
> 0 0 -1 1 -.5 .5 -.5 .5 (what is more active in oper2 than in oper1)
>
> Question 2: Is this the correct way of computing the mentioned contrasts?
Yes. If you want to see how to form all of these contrasts, then look
at some of previous posts on generating contrasts as I go through them
step by step.
>
> Question3: What would I have to enter for the respective F-contrast to test, for example, what regions are affected by OPERATION?
Use the above T-contrast, but tell SPM they are an F-contrast. Because
you have a 2x2 design, the size of the F-contrast for each factor will
be 1 (n-1).
>
> Best,
> André
|