Dear Daniel,
Thanks for your email.
> Thanks for your advice below. I actually tried modelling each of the 8
>trial types by specifying 2 onsets - one for the cue and another for the
>target - each time any trial occurred. That is, I modelled each of the 8
>trial types as a separate regressor, but specified two onsets for each event
>to try to best capture the event-related activity produced by the cue and
>target stimuli within each trial. Unfortunately, this didn't work too well.
I am not entirely clear on how you specified 2 onsets. Do you mean that you
had 8 regressors for your cue and 8 more for the target? If so, this would
result in all of the cue regressors being 100% correlated with your target
regressors (since each target follows every cue by a (very small) fixed
interval). In your previous email to the SPM99 help list, you rightly
expressed the concern that the cues might interact with the targets to
produce complex haemodynamic responses. I would suggest that you model
activity that is time-locked to the cues, and that you implement a method
that models flexibly and makes few assumptions about the shape of the
haemodynamic response. The best way of doing this in SPM99 is to use basis
functions. Instead of convolving your onset times with the hrf, you might
want to convolve with gamma or fourier functions. This strategy works best
if you have lots of data and can afford to spend degrees of freedom on more
columns in the design matrix. F-contrasts may then be used to test for
differential effects between conditions.
>Even testing for a main effect of visual field of target gives very
>statistically weak results, and I would think that would be a very powerful
>effect in early visual areas. So, it seems like I'm probably not capturing
>all of the variance here. Can you recommend a different way to model the
>data? Should I have just specified a single onset time at the time of
>target presentation? Should I model each trial as a small block? Or, would
>you recommend a different approach?
Hope this helps.
Very best wishes,
Narender
>Thanks as usual,
>:> Daniel
>
> > Dear Daniel,
> >
> > I thought I might have a go at answering this one...
> >
> > At 13:10 23/10/2001 -0400, you wrote:
> > >Dear SPMers
> > >
> > >I have a question about how to best model a particular design. In this
> > >design,
> > > there are 8 trial types. Each trial type has 2 stimuli -- one cue
> > >and one target, which are separated by 1 TR (1.5 seconds). This is a
>cued
> > >attention task, in which there are two types of cues (task 1 cue and task
> > >2 cue)
> > >and eight types of targets, which are defined by crossing Task (task 1,
> > >task 2), Visual
> > >Field (LVF, RVF) and Congruency (congruent, incongruent). I've
> > >listed the eight types of randomized events below.
> > >
> > >1. Task1 Cue + LVF congruent Task 1 target
> > >2. Task1 Cue + LVF incongruent Task1 target
> > >3. Task1 Cue + RVF congruent Task1 target
> > >4. Task1 Cue + RVF incongruent Task1 target
> > >5. Task2 Cue + LVF congruent Task2 target
> > >6. Task2 Cue + LVF incongruent Task2 target
> > >7. Task2 Cue + RVF congruent Task2 target
> > >8. Task2 Cue + RVF incongruent Task2 target
> > >
> > >
> > > My question concerns the best way to model these events. One way is
>to
> > >assign one regressor to all Task1 cues, a second regressor to all Task2
> > >cues, and then eight more regressors, one for each type of target. I'm
> > >concerned, though, that since there are no cue-only trials (i.e., trials
>in
> > >which only a cue appears) it might be difficult for the multiple
>regression
> > >technique to distinguish between responses to cue and target stimuli.
> > >Moreover, the SOA between cue and target stimuli is fixed (i.e., not
> > >jitterred).
> > >Would the fact that a given cue can be followed by 4 different types of
> > >targets
> > >help with distinguishing cue and target responses if the data were
>modeled
> > >in this way?
> >
> > I agree with the point that you will not be able to reliably disambiguate
> > the cue and target, and would go further to suggest that this is the case
> > whatever modeling method you choose. However, you have a balanced 2 x 2 x
>2
> > factorial design. Given that you have a sufficient number of trials and a
> > rapid enough TR, you will be able to test for the main effects (and the
> > interaction if this is relevant to your study). If you were to model the 8
> > event types individually, then you will be able to test for the main
> > effects of visual field laterality using the t-contrast [-1 -1 1 1 -1 -1 1
> > 1]. This will localise voxels in which the mean activity for
>presentations
> > in the left visual field is greater than the mean activity for
> > presentations in the right, over and above all of the other factors in
>your
> > design. Obviously, you could also test for the main effect of congruence
> > over and above all other components using [1 -1 1 -1 1 -1 1 -1]. In answer
> > to your question, this is possible because the contrast is balanced across
> > both levels of cue, and thus the activity from each level of cue
> > contributes equally to each type of target.
> >
> > > A second modeling option is to model each of the 8 event-related
> > > trial types listed above
> > >separately. This could be done by specifying an onset at the time of cue
> > >presentation and then using temporal and dispersion derivatives to catch
>the
> > >target response. However, I'm concerned that this might not capture
>enough
> > >of the target response. It could also be done by specifying an onset at
> > >the time of target presentation.
> > >In this case, however, it seems like each target regressor would probably
> > >model variance that was due to the target
> > >as well as the immediately preceding cue. Is that correct?
> >
> > That is absolutely correct. There would be little point in using a the
> > temporal derivative to capture target-related activity. You might, for
> > example, end up modeling cue-related activity that occurred late (remember
> > that the forms of haemodynamic responses are heterogeneous), and this
>would
> > be mis-attributed to the target. There is also no point in modeling the
> > targets separately for the reasons you mention.
> >
> > With best wishes,
> >
> > Narender
> >
> >
> > ********************************************************************
> > Dr Narender Ramnani
> >
> > Sensorimotor Control Group
> > Department of Physiology
> > University of Oxford
> > Parks Road
> > Oxford OX1 3TP
> > UK
> >
> > Oxford University Centre for
> > Functional Magnetic Resonance Imaging of the Brain,
> > John Radcliffe Hospital,
> > Headington,
> > Oxford OX3 9DU
> > UK
> >
> > Tel. 01865 222704 (Direct)
> > 01865 222729 (Admin)
> > mob. 0771 2632785
> > Fax. 01865 222717
> > email [log in to unmask]
> >
> > *******************************************************************
> >
> >
> >
********************************************************************
Dr Narender Ramnani
Sensorimotor Control Group
Department of Physiology
University of Oxford
Parks Road
Oxford OX1 3TP
UK
Oxford University Centre for
Functional Magnetic Resonance Imaging of the Brain,
John Radcliffe Hospital,
Headington,
Oxford OX3 9DU
UK
Tel. 01865 222704 (Direct)
01865 222729 (Admin)
mob. 0771 2632785
Fax. 01865 222717
email [log in to unmask]
*******************************************************************
|