Philipp,
> using VBM we are looking genotype dependent effects in structural
> scans (spm5).
>
> The genotype can be encoded as a factor with three levels (different
> group sizes in our case) - Variances were assumed to be unequal, a
> set of mean centred covariates (no interactions) is also entered.
>
> Assuming that the group levels are in the first 3 regressors, how is
> a T-contrast correctly defined that compares p o o l e d groups 1
> and 2 against 3?
>
> (+0.5 +0.5 -1?)
This is correct.
> In what way is such a test (if allowed at all) different from
> pooling the groups 1 and 2 from the start (giving one df more)? I
> expected subtle results between the two models, but the differences
> were larger than just some voxels.
As you've observed, the cost of modeling groups 1 & 2 separately is a
reduction of error DF by one. The benefit is that any systematic
difference between groups 1 & 2 is modeled; if a 1 vs 2 effect is
present and not modeled, it just inflates the residual variance and
can decrease sensitivity.
So if the 1/2/3 model is better than the 12/3 model, it suggests the
presence of an appreciable 1 vs 2 effect. I honestly won't expect
12/3 to be better unless you had a tiny amount of data where that 1 DF
would really made a difference.
-Tom
-- Thomas Nichols -------------------- Department of Biostatistics
http://www.sph.umich.edu/~nichols University of Michigan
[log in to unmask] 1420 Washington Heights
-------------------------------------- Ann Arbor, MI 48109-2029
|