Alex,
On Wed, 29 Mar 2006 10:50:29 +0100 "Dresner, Alex"
<[log in to unmask]> wrote:
> Dear Tom,
> I've been looking at the ResMS images for a couple of models, and was
trying to also do a manual calculation of T-scores from the beta values.
What specifically are the 'error degrees of freedom'? My first guess would
have been the df of the model, (# scans) - (# regressors).
>
> Also, I noticed a completely different formula from the one Christian
seems to be using, in an older e-mail on the list:
> > Only when dividing the ResMS map by the effective df's (SPM.xX.erdf)
did I
> > get a correct estimation of the individual SD,
> > (for a particular contrast) using:
The short answer is that xX.erdf is the right thing to divide by. Your
"n-p" calulation can be off for two reasons. First, if you specified a
high-pass filter, there are actually extra columns in your design matrix
that you never see (see xX.K.X0) and hence your "p" must account for these.
The second reason, which only applies if you using SPM99, comes from an
effective degrees of freedom calculation; if a low-pass filter is used or
if intrinsic autocorrelations are modeled, then error DF are calculated as
per tr(RV)/tr(RVRV) (see Worsley & Friston (1995) for details), which may
well be fractional.
Hope this helps!
-Tom
|