Dear Keith,
> I would like to be able to select a region of interest (ROI) and measure
> the average response of the voxels in this ROI to a number of different
> types of single-event trials. I don't think this is possible with the
> existing SPM code. Has anyone written any routines to do this?
Not that I know of. The simplest way to do this would be to take the
average parameter estimates over the voxels in question and simply
multiply them by the temporal basis functions modeling the response.
To write your own code I would look at spm_graph to see how parameter
estimates and basis functions are retrieved from an analysis.
> Also, we have single-event trials of the same type scattered among a
> number of different scanning sessions (we are limited in the number of
> images we may acquire per scan). I would like to find the average
> response to each type of trial, averaged over the scanning sessions. That
> is, I would like to fit a response function to the all the trials in all
> sessions rather than fitting a response function to the trials in each
> session and then averaging together each of these fitted functions. I
> know that there is some variance between sessions, but in each scan there
> are only a few trials of each type, and I don't think there are enough
> trials to get a reasonable estimate of the response function for the
> trials only in a single scan.
The best approach to this is to model session-specific responses in the
usual way and take the average over sessions using the appropriate
contrast. This will give you the same result as if you had modeled all
the trials in one super-session. This approach assumes you have the
same [similar] number of trials per session. If some sessions contain
very few trials, and others more, then the situation is more
complicated. By modeling session-specific responses one is implicitly
modeling session x trial interactions. These are generally orthogonal
to the simple main effect of trials. However in an unbalanced design
the interaction and main effects can become confounded leading to
inefficient estimates of both. In this instance simply model one
super-session and enter session-specific effects by hand as covariates
of no interest (i.e. columns of ones, and any drift/low frequency
terms, for each session).
I hope this helps - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|