See responses below. It depends on what you want to see...

On Tue, Sep 6, 2011 at 12:33 AM, Bob Spunt <[log in to unmask]> wrote:
> Assume I have a significant PPI effect of a seed region X on target region Y
> in the task contrast A > B. What is a valid procedure for plotting this
> effect? I know of multiple procedures:
> 1. If you've done gPPI, you could plot the betas from A and B.

This will work. This would be a bar graph for A and B. This is done at
the group level with the beta's for A and B from the first level.


> 2. If you did traditional PPI, you could run two additional models, each
> modeling the interaction among the seed and one of the conditions. Then,
> extract and plot the betas.

This is described in the manual. You regress/ plot A*seed against
A*target. It is done for each subject individually. Finally, this
doesn't represent what you've actually modelled and analyzed from the
first-level to the second-level. With gPPI, this would be easy to do
since the conditions are all ready seperated for the seed region,
you'd just need to repeat for the target region to get the second set
of values to do the regression or plotting.

> 3. Regress data from the target region onto the seed region, separated by
> condition; then compare the slopes. I think this is the procedure used in
> the original Friston et al. (1997) paper. With this, it's unclear to me how
> best to separate the conditions. For instance, if one has 5 second stimuli
> in an event-related design with a TR of 2 seconds, would one extract the
> first 4 or so observations (corresponding to the 8 second period following
> stimulus onset)?

The original data was from a block design. For an event-related design
it is unclear how you would label the points, or which point to
choose. Remember that the interaction is at the neural level and in a
rapid-event related design is very quick. You'd have 1 point per
trial. The problem becomes how to pick the one TR. No easy solution,
which I think is why people opt for #1 or #2.


> 4. A similar procedure to 3 which I saw used by Lombardo and colleagues
> (2010, J Cog Neuro). They extracted the timecourses of the seed and target
> regions as in 3, but instead of separating the timecourses by condition,
> they multiply them by an HRF-convolved task vector for each condition. Then,
> they compute correlations among the seed and target ROIs, then convert the
> correlations into Fisher's z scores to permit visual comparison.

I have absolutely no idea what this would even represent in terms of
physiology or with respect to PPI. I'll let someone explain the
physiological relevance. With regards to PPI, it is very different.
(1) the timecourses used to form the interaction are the estimated
neural activity, not the extracted BOLD signal. (2) The interaction is
not of a convolved HRF with anything, but of the event times with the
neural activity. This is then convolved with the HRF. (3) PPI isn't a
correlation, it is difference in slopes between two event types.

While this is an interesting metric, it doesn't correspond well to the
PPI method that you'd implement in SPM nor gPPI.


> If any one has any thoughts on these procedures or recommendations for
> others, I would really appreciate it.

I'd use #1. If you need to illustrate at the single subject level (for
illustrative purposes only) #2 would work.

#3 would work if you knew the TRs that corresponded to the
interations, but since you don't know which points to choose, you
should avoid this method.

#4 is an interesting metric, but might not correlate with PPI. The
direction should be the same though.


> Sincerely,
> Bob Spunt
> Postdoctoral Fellow
> Social Cognitive Neuroscience Lab - www.scn.ucla.edu
> Department of Psychology
> University of California, Los Angeles
>