Hi, 

I am interested in learning more about how PPI is calculated in Conn14. The questions I have are below:

1) (a) Since I am interested in examining the changes in functional coupling between region A and region B during a task I ran a PPI analysis. I notice that when I run a test with region A as a seed and B as a target that I get slightly different results in comparison to when I run a test with region B as a seed and region A as a target. I am unsure about the asymmetry. If I understand things correctly, the problem might lie in that the PPI regressor is different when I choose A as a seed instead of B as a seed. 
(b) Since there is an asymmetry between A-seed to B-target VS B-seed to A-target then, I'm assuming that I must have clear support for my choice of one or the other?
2) Is the task regressor included in the model in addition to the PPI regressor?
3) Is there a way to extract the correlation between the PPI regressor and task regressor?

Best, 
Michael

Hi Michael,

Yes, PPI analyses are non-symmetric by nature, as they are typically interpreted in the context of "effective" rather than "functional" connectivity. When using region A as a source and region B as target the PPI equation fitted is:

Yb = Ya*k1 + P*k2 + (Ya*P)*k3

(where Ya and Yb are the BOLD timeseries in regions A and B, respectively, P is the task effect, and k1, k2, k3 are the estimated effects; in particular k3 is the interaction effect of interest), while when using region B as source and region A as target the PPI equation fitted is:

Ya = Yb*k1 + P*k2 + (Yb*P)*k3

so the estimated interaction effects (k3 on each equation), while often similar, are not necessarily or simply related to each other.

Regarding your question (2), yes the task/psychological regressor (P in the equations above) is included in the model in addition to the task x physiological interaction (Ya*P or Yb*P in the equations above) term.  And regarding (3), unfortunately that correlation is not returned as part of the standard CONN estimation procedure. You could potentially compute those manually using the ROI timeseries Ya (e.g. in conn_*/preprocessing/ROI_Subject#_Condition#.mat, variable 'data') and the task effects P (e.g. in the same files, variable 'conditionweights{1}')...

Hope this helps
Alfonso

Originally posted by Michael King:

Hi Alfonso, 

Thank you for the reply. quite helpful.

Best
Michael