Dear David,
> I have a question, particularly for Dr. Friston, on AnCOVA/correlation
> (could the responder please respond directly in addition to the mail
> group, as I'm having trouble getting mail from the list!).
>
> I have the following PET design: 2 groups, 4 conditions (in 4 scans),
> and a subject characteristic measured outside the scanner, on which the
> groups happen to differ. I noted a discrepancy in some results that I
> don't understand.
>
> First, in order to make a condition, I used the means utility to create
> proportionally scaled images for the subjects. Then I used the image
> calculator to subract condition 1 from condition 2 (I don't know if
> there is an easier way to do this, but since I didn't have replication
> scans, I could not get *con.imgs for subjects from a sub x condition
> interaction).
>
> Then, I used a basic correlation model to correlate these subtraction
> (condition) images with my external characteristic (co-variate; design
> matrix= covariate mu). Finally, I used AnCOVA to compare the two
> groups, adjusting for differences in the co-variate (design matrix=
> group1 group2 mu covariate).
>
> At this point I assume that in the correlation analysis, a contrast of
> [1 0] tells me the areas where my co-variate correlates with the
> condition effect represented by the image subtraction (Condition1 -
> Condition2).
>
> I also assume that in the AnCOVA, the contrast [0 0 0 1] should tell me
> the identical information as the above contrast for the correlation.
> However, it doesn't. The results are similar in some areas, and
> dramatically different in others.
The difference between the two results is simply due to the fact that
the first analysis includes within- and between-group effects of the
covariate, whereas the second discounts between-group contributions
because you have modeled the group effect.
> Do I misunderstand AnCOVA?
>
> I also assume that the contrast [1 -1 0 0] tells me were group 1 is
> greater than group 2, with this difference having variance due to the
> covariate already subtracted out. Is this correct?
Absolutely.
> A quick incidental question: Assuming that I have made the condition
> images correctly (mean utility, image calculator), how can I apply the
> regular mask to exclude areas outside the brain?
I think you use the 'explicit masking' option and a mask from an
appropriate first-level analysis.
I hope this helps - Karl
|