Hi Antonia,
I think one clarifying idea is that the largest statistical effect sizes
might not necessarily inhere in the tests that are relevant for one's
hypotheses. So even though one can conceptually treat the coefficients from
the derivatives as repeated measures, it might not be rational to use them
as such unless there is a scientific hypothesis concerning them.
That being said, it is not reasonable to assume that these repeated measures
are i.i.d., and so their second-level covariance structure should be
estimated if one is going to use them as second-level data. Also, there is
no point to include them as second-level data unless one is going to
explicitly test hypotheses with them. That is, they can not act as
"regressors of non-interest" as they are not regressors at the second level,
but data.
Eric
----- Original Message -----
From: "antonia hamilton" <[log in to unmask]>
To: <[log in to unmask]>
Sent: Wednesday, June 22, 2005 5:24 PM
Subject: [SPM] 2nd level analysis with deriv and disp
> Hello,
>
> I'm sure this is a foolish question, but I've been getting myself
> confused. I've got data from an fMRI experiment with a standard 2x2
> factorial event related design and I've modeled each subject's data
> with an HRF and its derivative and dispersion in SPM2. So I get 12
> beta images (2x2 factorial x 3) for each of my 20 subjects. Now I
> want to go to the second level and find the main effects of my
> factors. I don't have any specific hypotheses about temporal effects,
> but I'm expecting to get a basic HRF-shaped response in all the
> regions I'm interested in, and I need to know if this activation is
> bigger in some regions than in others, using the most powerful
> analysis I possible.
>
> So should I ...
>
> a) Calculate contrasts for each subject for each main effect with
> each basis function and then do something with them ... ??
>
> b) Ignore my deriv and dispersion columns and just do the main
> effects using t-tests on the HRFs?
>
> c) Take all 12 betas for each subject into a big ANOVA and do some
> kind of F test? I tried getting a main effect of A with:
> hAB hAb haB hab dAB dAb daB dab sAB sAb saB sab
> 1 1 -1 -1 0 0 0 0 0 0 0 0
> 0 0 0 0 1 1 -1 -1 0 0 0 0
> 0 0 0 0 0 0 0 0 1 1 -1 -1
> where h = hrf, d = deriv, s = dispersion, and A and B are my factors
> But here I'm not sure if I need to this as a repeated measures ANOVA
> or just a straight forward one.
>
> Methods b and c give fairly similar results, but I don't know which is
> right or why. I guess the real question is are my Deriv and Disp
> columns a useful signal in themselves, or should I consider them as
> regressors of non-interest and focus just on my HRF column.
>
> Can anyone comment?
>
> Antonia.
|