Dear Anderson,
I got it now. This is really helpful!
Thank you so much for your examples for my questions.
Cheers,
Yuwen
Date: Thu, 10 Aug 2017 09:59:07 -0300
From: "Anderson M. Winkler" <[log in to unmask]>
Subject: Re: Help with GLM for testing Correlation Difference in a single group
Hi Yuwen,
This design & contrasts wouldn't be right even if the symptoms were more
similar to each other (e.g., abdominal pain vs. chest pain, or pain in the
left hand vs. pain in the right hand), the reason being that even these
don't represent the same thing. The case in which you could subtract as you
indicate would be if you were comparing abdominal pain in group 1 vs.
abdominal pain in group 2, or symptom A count in group 1 vs. symptom A
count in group 2, but unfortunately not the two things in the same subjects.
From your messages I believe you'd like to test whether FA correlates more
strongly with symptom A count than with symptom B count, and vice-versa.
This is what is provided by the interaction contrast, which I strongly
recommend you run.
The only case in which your design would be right would be if you wanted to
see if the difference in symptom counts (i.e., #symA - #symB) is associated
with FA, after the overall number of symptoms (i.e., #symA + #symB) is
corrected for. This is what effectively what your design is testing. If
it's something you are interested in, then that design is ok. Otherwise,
it's the interaction you'd want.
About correlations between symptom counts A and B (anxiety co-occurring
with depression symptoms), yes, it can be a problem for investigation, in
that power will be reduced to detect the correlations of either count with
FA. There is nothing you can do, though, unless redefine the problem (e.g.,
use symptoms A and B to derive some latent measure that could be
investigated, etc).
Regarding the interaction contrast, the directions and slopes can be known
from the signs and magnitudes of the regression coefficients of the main
effects (C1 and C2 in the design I proposed). It is also possible to
extract the values from the significant voxels and make a scatter plot in
some other software (e.g., Matlab, R, or even some spreadsheet software).
Hope this helps!
All the best,
Anderson
|