Nicole,
You are correct in thinking that the parameter estimate for the session with
many fewer instances of the event will tend to be have a larger variance
than that of other sessions with more instances. One solution more easily
implementable within the SPM framework would be to use contrasts that weight
the parameter estimates from each session optimally (i.e., inversely by
their variances). These variances (not the estimated varainces). to within a
scaling factor are saved somewhere in the SPM output (I am still in the
bronze age with SPM99, so I wouldn't want to point you in the wrong
direction for SPM5, but no doubt someone here will be able to direct you).
Best,
Eric
----- Original Message -----
From: "Nicole David" <[log in to unmask]>
To: <[log in to unmask]>
Sent: Wednesday, July 27, 2005 11:18 AM
Subject: [SPM] treating multiple sessions as one
> Thanks, Anja!
>
> Concerning the second issue, what if a condition/type of event only
> appears once in one scanning run and then 10 times in another scanning
> run? doesn't that mean that the parameter estimates will be highly
> unstable and very influenced by noise. Would it be possible to analyze
> everything as one run and include extra covariates for each scanning
> session? And, if so, how would I do this or is there another solution
> (e.g. including global signal scaling)?
>
> Nicole
>
>
>
> <[log in to unmask]>:
>
> > Hi Nicole,
> > ad 1) in my experience, it is better, if you model errors as an
> > additional error condition (but do not differentiate it further
> > into different errors or you lose too many degrees of freedom)
> > you coudl also choose to ignore errors and only model correct trials.
> > I usually try both and choose the better versions (i.e. the one with
> > higher T-values). Leaving the errors in is usually the worst option.
> >
> > ad 2) it might be difficult to get good results, because when
> > the scanner starts anew it will have a different setoff/mean value.
> > So the analysis over all runs will not yield much. So you should do it
> > as a FFX Model with 4 different sessions. Then SPM will estimate all
> > predictors for each session separately and insert a session
> > mean as a covariate of no interest.
> >
> > Hope that helps
> >
> > Anja
>
> >
> > >>> Nicole David <[log in to unmask]> 27.07.2005 13:00 >>>
> > hi,
> > i have two questions:
> >
> > 1) i have an event-related design. each event (a stimulus) is coupled
> with
> > a rating and categorized according to it (into two conditions).
> sometimes,
> > subjects missed a rating, i.e. yielded no score. How do i deal with
> these
> > events that weren't rated? do i have to model them out, i.e. as a
> > third "error" condition, so that they don't fall into the iti (and thus
> > the baseline)? fortunately, they didn't occur often... or would it be
> > legitimate to ignore those only specifying a model with the events that
> > were rated?
> >
> > 2) i had 4 scanning sessions. would it be o.k. to treat everything as
> > one long scan?
> >
> > any ideas?
> > thanks,
> > nicole
|