Thanks, Christian! Great explanation.
Emily
----- Original Message -----
From: "Christian F. Beckmann" <[log in to unmask]>
Date: Tuesday, June 24, 2008 8:44 pm
Subject: Re: [FSL] Melodic Components
> Hi Emily
>
> >
> > 1) I have tried running Melodic on one session, with the
> appropriate
> > FEAT design (3 column format
> > for custom timing), but since the full model fit covers ALL the
> > copes (is that right),
>
> No, the total model fit combines all EVs, it does not consider
> copes
> at all.
>
> > I have no idea
> > how to pick the relevant component out of the 183 it gives me.
> >
>
> If you specify the design.mat as the model _and_ the design.con as
>
> the set of contrasts in the melodic GUI then the GLM table
> underneath
> the plot should contain significance for each one of the individual
>
> contrasts (rightmost column). Given that ICA time courses might be
> inverted it always makes sense to include each contrast in it's
> positive and negative form (or look for p<0.01 _and_ p>0.99 to see
> if
> a particular contrast is significantly related to the IC time course)
>
>
> > 2) I then ran ONLY one contrast (2 EV), with same problem
>
> > 3) I have no run only one Contrast with one EV, to see if I can
> > find the timecourses in Melodic that
> > look like the timecourse from FEAT for that EV. I am assuming
> that
> > is how I identify what
> > components are related to that contrast in ICA?
> >
>
> They might not, ICA will attempt to show you what's in the data.
> Imagine a finger tapping experiment where you ask subjects to use
> their index finger in condition A and perform random finger tapping
> in
> condition B. In the GLM the canonical way of modelling this is to
> have
> 2 EVs, one for conditionA and one for condition B (assuming that
> you
> also have rest in between). In ICA, however, the associated time
> course should be one which shows both active conditions in the same
>
> time course (at different levels, hopefully). Therefore, the ICA
> time
> course will not look like any one single EV but instead be a linear
>
> combination of the two. The full model fit does show you if it is
> task
> related.
>
> > 4) If the full model fit is all the inputs, why does it change
> > between components?
>
> Because the data changes (i.e. the tested time courses change).
>
> >
> >
> > 5) Is the IC ## time course supposed to fit that full model line
> > really closely? NONE of ours do, but
> > our data is from a self paced task, and the timecourse is not
> > blocked predictably-
>
> The data is never wrong, the ICA time courses might be poorly
> estimated and/or your design might be badly wrong
>
> > is that why it
> > doesn't?
> >
>
> maybe
>
> > 6) Is it possible to run that many COPES in ICA and have any
> > meaningful data come out, or should
> > I be running 1 EV at a time, or just one contrast at a time?
>
> The number of contrasts does not matter. In fact, compared to a GLM
>
> you might actually want to add more contrasts to e.g. ook at de-
> activations or at signifcace of nuisance regressors that you'd
> typically not include in your COPEs for a standard GLM
>
> hth
> Christian
>
|