Hi Behrang,
> In other words, if I do one-sample t-test on every voxel (assuming
> different subjects are different observations) in Matlab, should it be the
> same image as SPM_XT.img generated by SPM5?
Yes, I think so, to within some reasonable numerical error. SPM uses a
more general statistical model than MATLAB's ttest, but it's
mathematically equivalent. There are lots of nice sets of notes on the
GLM, the ones I glanced at most recently are:
http://www2.imm.dtu.dk/~fn/ps/Nielsen2004General_slide.pdf
And SPM does evaluate its General Linear Model at each voxel
independently, computing beta and residual mean square images, which
are later used to produce the contrast and statistic images (I think
spm_contrasts contains the actual code that derives the t-value from
the contrast and its standard error).
I think the only places where voxels can't be considered independently
are things for multiple comparison correction, e.g.
- smoothness estimation includes spatial derivatives (see Kiebel paper)
- FDR uses sorted p-values over all voxels in the search region
- cluster-based correction looks for spatially connected components
Hope that helps,
Ged
|