Hi Mike,
Thanks for emailing the figures. The difference is just that SPM2
models a mean/constant column in addition to two columns for groups,
whereas SPM5 just models the group columns. As Volkmar said, the DF
are the same as the constant ones column is just the sum of the two
individual group columns (the rank of both design matrices is 3).
The extra constant column in SPM2 means that the individual group
columns are no longer estimable on their own. A contrast of [1 0 0]'
over the spm5 design is equivalent to a contrast of [1 0 1 0]' over
the spm2 one, while [1 0 0 0] for spm2 is invalid.
For the difference of the two groups (which I'm guessing is what you
are interested in), the spm5 contrast would be e.g. for A>B [1 0 0]-[0
1 0] = [1 -1 0]. While for SPM2: [1 0 1 0] - [0 1 1 0] = [1 -1 0 0].
So in both versions of SPM, a zero-padded contrast [1 -1] will give
the same correct answer for A>B.
I worked through a related example, but for the case where someone was
testing the covariate, which might either be helpful or confusing:
http://www.cs.ucl.ac.uk/staff/gridgway/ancova/
but I hope this email is helpful anyway.
Ged.
Mike Glabus wrote:
> Firstly, I confess to a little ignorance for the rationale for modeling
> the block effect but assume this is equivalent to the "DC" or offset in
> the GLM, i.e. the y intercept (?).
>
> With that in mind, I have been attempting to replicate an SPM2 design in
> SPM5 for a two-group VBM analysis.
>
> In SPM2 I used the "compare populations 1-scan per subject" with ANCOVA.
> In SPM5 I have tried using both "independent t-test" and "full factorial"
> models, but in both cases, there is no modeled block effect. However, the
> df in both SPM2 and SPM5 designs is the same! (see attached)
>
> Is there an explantation for i) the absence of a modeleed block effect in
> SPM5; ii) the df being the same in SPM2 and SPM5 designs, where one
> (putatively) should have one df less (SPM2).
>
> Regards - MFG
>
|