Hi,
This is actually a relatively subtle issue.
You can set the GLM up to work like this (explicit baseline) and have
it be
consistent, as long as the mean value of the signal contains useful
information.
However, in FMRI this is not the case and that is where the complication
comes from.
Since the mean value of the FMRI signal is not useful, it either
needs to be
removed from the EVs and data by demeaning (which is what FEAT does)
or included as a nuisance regressor in the GLM. Either way, it is
then not
allowable in the GLM to have any combination of the remaining EVs which
add up to a constant signal (which would be equivalent of modeling
the mean
a second time). Doing this would make the design rank deficient.
So consider an experiment with one type of stimulation. If you have
one EV
for activation and one EV for baseline then each timepoint will
either be
part of the activation EV or baseline EV. If there was no HRF then
this would
be totally rank deficient as the addition of these two EVs would be
equal to
a constant for all timepoints. However, the HRF delays and blurs to
some
extent which means that the ends of the timeseries don't quite end up
being
a constant (but are close) and so the model is nearly rank deficient,
although
technically you can results from it - they are just poor results.
The situation is equivalent for more than one stimulation, as the sum
of all
the EVs typically adds up to a constant.
The bottom line is that in practice the baseline needs to be left as
implicit,
and not explicitly modeled, due to these rank deficiency issues,
which are
a consequence of the fact that the mean signal contains no useful
information
and is removed. In FEAT the mean is removed automatically, so you do
not
see this typically, although Featquery does use the mean value when
calculating percent signal change.
I hope that all this helps explain why things are as they are.
All the best,
Mark
On 15 Feb 2008, at 18:42, carlos silva pereira wrote:
> I get it, it's quite logical to ignore the baseline when
> contrasting 2 EVs.
> Still it seems the same as considering the baseline an EV and then
> doing EV1-EV2 (baseline) and not the 1 0 0 contrasts.
> But I may be missing the point...
> Best,
> Carlos
>
>
> On 15/02/2008, Mark Jenkinson <[log in to unmask]> wrote: Hi,
>
> If you do a difference between two EVs in your contrast then it
> doesn't matter.
> It is like doing (effect1 - baseline) - (effect2 - baseline) =
> (effect1 - effect2)
>
> Note that this is true for any contrast that adds up to zero.
>
> So in contrasts like that the baseline is irrelevant.
>
> But in the simple "1 0 0 ..." type of contrasts it obviously is
> important and you are
> looking at effect1 wrt baseline.
>
> All the best,
> Mark
>
>
> On 15 Feb 2008, at 18:23, carlos silva pereira wrote:
>
> > Thanks for the quick reply!
> > One last thing: when I specify a contrast between two EVs, will
> > they be previously contrasted to the baseline to?
> > For example: in contrast1 I define a value of 1 for EV1 and in
> > contrast2 I define 1 for EV1 and -1 for EV2.
> > In contrast2 will EV1 and EV2 be contrasted to baseline also or
> > just to each other?
> > Thanks again, this forum is a major help for starters!
> > Carlos
> >
> > On 15/02/2008, Mark Jenkinson <[log in to unmask]> wrote: Hi,
> >
> > You should not use baseline EVs in Feat.
> > It may work, but it will be close to rank deficient normally.
> > The best way is to omit it and then all EVs are implicitly
> contrasted
> > wrt baseline and the modeling inside Feat works optimally.
> >
> > All the best,
> > Mark
> >
> >
> >
> > On 15 Feb 2008, at 18:10, carlos silva pereira wrote:
> >
> > > Hi to all!
> > > Basic question: in the Feat full model setup is it the same to 1)
> > > consider the baseline condition as en EV and contrast it with the
> > > main condition; 2) just omit the baseline and FSL will contrast it
> > > with all the EV's?
> > > I ask this because I tried both options and results are very
> > > similar, although not absolutely equal.
> > > Thanks in advance.
> > > Carlos
> >
>
|