Dear Steven,
It's indeed not necessary (and SPM does not do it) to estimate a full
and a reduced model in order to compute the extra sum of square of your
F-test - you just need to estimate the full model and use the 'trick'
described in chapter 8 of the "Human Brain Function" book, page 13:
http://www.fil.ion.ucl.ac.uk/spm/doc/books/hbf2/pdfs/Ch8.pdf
The relevant code in SPM is buried in functions spm_contrasts.m and
spm_FcUtil.m if you want to see what happens in practice.
Best regards,
Guillaume.
Steven Liu wrote:
> Dear SPM expert:
>
> As the attached SPM course slides (only F-contrasts in the file for
> simplicity)
> about the contrasts made by Jean-Baptiste Poline, an F contrast is
> computed
> by the following steps:
>
> 1. Esimate the full model
>
> Y = Xb + e,
> based on this estimation, calculate S^2
>
> 2. Estimate the reduced model
> Y = X0b0 + e0
> based on this estimation, calculate S0^2
>
> 3. Then, Calculate F:
> F = (S0^2-S^2)/S^2
>
> But when I run SPM F contrast, I did not see that SPM does another
> estimation
> again using the reduced design matrix X0 as it did for X. And when
> I looked into SPM codes, I did not see that SPM calls spm_spm to
> do the estimation for X0.
> So, why? How SPM gets b0, e0 and then S0^2?
>
> Thank you very much in advance!
>
> Steven
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|