Thomas and Russ,
Global normalization is an interesting topic. For single subject analyses, it is
possible to eliminate global normalization. However, if you are analyzing a group,
some type of normalization must be used. For fixed effects analyses, you have to
use global scaling. For random effects analyses you have other options such as
grand mean scaling.
However, there are cases when global normalization is required. One example would
be when the global signal changes with the task. With our global normalization,
such a data set would show whole brain activation. Unfortunately, region
activations can influence the global calculation and global normalization can alter
the significance of "real" activations. There is currently no perfect solution, but
Jesper Andersson has been working on this issue. He recently published a paper
addressing global changes in PET (Andersson et. al, Neuroimage 13, 1193-1206, 2001)
and previously suggested a method (Andersson et. al Neuroimage 6, 237-244, 1997)
using a masking technique. A comparison of techniques was presented at OHBM by
Maria Gavrilescu et. al (NeuroImage 13(6) S122, 2001).
So, in short, there are conditions where global normalization is very important but
it remains debatable as to which method should be used.
Paul Laurienti
Russ Poldrack wrote:
> Dear Thomas,
> it's not clear to me why you would evern want to do global signal normalization
> for fMRI data. If you are including a high-pass filter in your design then
> things like low-frequency drifts will be taken care of on a voxel-by-voxel
> basis, and if you include motion parameters as regressors then you will also
> get rid of any intensity artifacts related to motion. Perhaps someone else
> could comment on why one would want to do global normalization for fMRI - we
> never do it these days and we get perfectly reasonable results.
>
> cheers,
> russ
>
> Thomas Stephan wrote:
>
> > Dear List,
> >
> > to deal with the problems regarding correlations between global brain signal
> > and the task in
> > fMRI experiments we use the tool rd_taskcorr provided by Kalina Christoff to
> > compute
> > the significance of those correlations.
> > If we have very simple experiments with an ABABAB design, we can just skip
> > global scaling
> > if we detect a significant correlation.
> > Now we analyze data of an experiment with 3 different stimulation conditions
> > and corresponding
> > three different rest conditions. What should be done if there is a
> > correlation between global signal
> > and e.g. Rest3, but no significant correlation for Stim1-3 and Rest1-2 ?
> > Should we compute some contrasts with global scaling and other contrasts
> > without ?
> >
> > Thanks for your opinions
> > Thomas Stephan
> > ======================================================================
> > Thomas Stephan Email: [log in to unmask]
> > Klinikum Grosshadern
> > Neurologisches Forschungshaus
> > Marchioninistr. 23 Fon: +49 089 / 7095-4819
> > 81377 Muenchen Fax: +49 089 / 7095-4801
> > ======================================================================
>
> --
> Russell A. Poldrack, Ph. D.
> MGH-NMR Center
> Building 149, 13th St.
> Charlestown, MA 02129
>
> Phone: 617-726-4060
> FAX: 617-726-7422
> Email: [log in to unmask]
> Web Page: http://www.nmr.mgh.harvard.edu/~poldrack
|