Dear Maria,
hope that someone else could comment on my suggestion....
to put both groups in one simple regression model is definitely NOT what you
want, because you would not be able to differentiate between the two groups.
You can fit such a model for each of the groups separately, then with the
drawback that you cannot make inferences about differences between the
groups.
My guess is that the multiple regression option is the best one, where the
spm-image data should be ONE of the first level contrasts of interest (per
subject): either A-B or B-A (it was not quite clear to me, if you meant
groups A and B or two within-subject conditions A and B, where I am talking
about the latter).
I am not sure if you need 3 or 4 covariates: At least you have to model the
psychological score separately for each group. For modeling this you would
have to enter zeros as "psych-score value" for each of the subjects of the
respective other group. This implies that the value "zero" must not be a
valid psych-score, so that eventually a constant has to be added to the
psych-scores, leaving only positive numbers for this score (excluding zero).
Then you need at least a third column modeling the overall mean, which
consists of ones only.
My guess is that this last suggestion is "right", but maybe it is better to
model the means separately for each group (two columns consisting of ones and
zeros only). (???) This would be your suggestion with 4 covariates.
After that the contrast [1 -1 0] would tell you where in the brain the
covariance between psych-score and con-image was bigger in the first group as
compared to the second group. Conversely the contrast [-1 1 0] would show you
"the same" for the comparison group 2 > group 1. Depending on your choice at
the first level (contrast A-B or B-A), the results are exactly the same,
differing in their sign only. This means that the interpretation is not that
easy, but it is no worse than in a two-sample t-test, where you do not
(necessarily) know if A-B was bigger in group 1 because A was so large in
group 1 and so low in group 2, or was it a low response during B (compared to
A) in group 1 and a high response in A (compared to B) in group 2 (with
several other at least theoretical possibilities.....).
By the way, the contrast would almost not differ for the 4 covariates model,
including only one extra-zero: [1 -1 0 0] and [-1 1 0 0], respectively.
Best wishes,
Thilo
On Tuesday 28 February 2006 17:19, Maria Densmore wrote:
> Dear SPMers,
>
> I am running a correlation analysis, I have 2 groups and I would like to
> compare the 2 groups, A>B and B>A and correlate this comparison with a
> psych score.
>
> Would I in a fixed effects model: enter all the subjects, then create a
> con.img for each subject 1-26 for A-B and B-A then enter these into a
> SIMPLE REGRESSION model to correlate with the psych score [1].
>
>
> I'm not sure what contrast I would use to show A>B correlated with a pych
> score if I used a MULTIPLE REGRESSION analysis with 4 covariates. Could it
> be [1 -1 1 -1](A>B) and [-1 1 -1 1](B>A) for:
> 1) 15X1 15x0
> 2) 15x0 15X1
> 3) 15 psych scores 15X0
> 4) 15X0 followed by 15 psych scores
>
>
> Comments on the above and the best approach would be greatly appreciated.
>
> Thank you for your time,
> Maria
--
Thilo Kellermann
RWTH Aachen University
Pauwelstr. 30
52074 Aachen
Tel.: +49 (0)241 / 8089977
Fax.: +49 (0)241 / 8082401
E-Mail: [log in to unmask]
|