Hi Martyn and Guillaume,
Thank you, you two, for your responses!
Martyn, I had indeed already come across your FiN and NeuroImage papers
and a number of your posts on the SPM forum. As I am only interested in
group x condition and not some restricted level of my condition, I
understood that the F-ratio error term only depends on the overall model
error, i.e. is independent of subject effects, as you have very nicely
stepped through in your FiN paper (very much appreciated, by the way!)
-- and therefore I wouldn't risk inflated type I error rates and hence
using SPM's flexible factorial is kosher?
The example I posted is me comparing my two-sample t-test, as you have
pointed out Guillaume as the right approach (yippee; where I took the
difference of my condition at my time points as contrasts in
subject-level models so I only have one con image per subject) with
flexible factorial out of pure curiosity and didactic reasons before
going on and adding my additional 2 other between-subject variables --
and indeed, I wanted to use either Martyn's MRM toolbox or SwE (also for
purely didactic reasons) or GLM Flex, possibly with the lme4-based/mixed
effects model approach in AFNI which I would be familiar with though
balking at defining an appropriate random term (but I see SPM12 also has
a mixed effects analysis module??)....depending on how motivated (how
much time) I have...so actually maybe just one of these after I figure
out which one of these would allow me pose my specific questions in my
longitudinal data most directly (I'm thinking in terms of hierarchical
regression modeling).
The motivation for my post is to understand some of the very basic
technical details/what you can draw from looking at the whitened design
matrix and covariance/orthogonality matrices -- sorry, prob too general
of a question and hence I didn't originally pose it that way but in
context of a toy problem. In a way, this is also a vote for a
"troubleshooting"/"diagnostics"/"data checking" section in the manual.
I'm trying to compile info from various lab worksheets people have made
available (but often they're workbooks and so missing the "teacher's
notes to the students").
Some of the very basic things I'm not sure about reading is, for
example, in the middle model in the figure of my original post, what can
I conclude from the covariance structure that variances/the diagonals
for the left half is that much more prominent than the bottom right
(reasons for concern?)? That the variances are so minimal in the third
model from the left which do not seem to correspond to the high
covariance between some subjects to my very naive eye? I understand the
number of hyperparameters that need to be calculated increases when the
model has to deal with non-sphericity -- but what are what seems like 4
hyperparameters in the 2 leftmost models correspond to? Sorry quite
naive questions.
Thanks again,
Gina
|