Dear Suzanne
> Hi Dr. Friston, I am currently doing a ppi analysis for
>submission of an abstract to society for neuroscience and am hoping to
>discover the relationship between activation in the amygdala during a
>certain task (or contrast as you embed it into the analysis) and activation
>in an orbital frontal region.
> You have in your introduction (the 3rd
>paragraph) that this analysis is regressing activity of one region on
>activity in the second region and looking at the slope of that line which
>would differ under different conditions.
This is certainly a useful way to understand PPIs when the psychological
factor has two levels (i.e. affords two regressions). It gets more complicated
when the factor has many levels or is parametric.
>My question comes from
>interpretation of our results. I have followed the on-line instructions by
>Will Penny. At the end, I am still left with the question of what our
>results actually mean. After I estimate all the individual models and do a
>group level contrast as these directions indicate, the activated voxels I
>assume correlate with the brain region we embedded into the design matrix.
I am not absolutely sure what you have done, so forgive me if I
have misunderstood you. First, the significant voxels testing for
a PPI at the second (between-subject) level are not voxels showing
an activation but voxels showing a interaction (this is the PPI).
This effect is orthogonal to the main effects of the psychological
and physiological variables. Usually, PPIs are used to explain
interactions in terms of changes in coupling.
This calls for a factorial design. In other words you have to have
one task to elicit co-activation so that the regression slope
can be measured and another task to change the coupling (i.e.
regression slope). It is not clear from your description whether
you have a factorial design. If you did, one usually chooses the
physiological variable from a region showing a significant main effect
of one factor and use the other factor as the psychological variable.
This will produce SPMs of regionally-specific interactions that are
explained by a modulation of coupling with the index or seed region by
the second factor. In short, the voxels showing a PPI are not necessarily
those expressing main effects (although they can be).
>Is the p value an indication of how much these areas are correlated? How
>do I know if they are positively or negatively correlated? Lastly, how do
>I know if the regions are activated or deactivated? It seems as if your
>graphs on page 222 would be the answers to these questions, but I don't
>know how you got there. Any help for a confused researcher would be
>greatly appreciated.
Following on from the above points; The interpretation of a PPI is only
clear in relation to some a prior hypothesis. In other words, you state
your hypothesis in terms of an increase, or decrease, in coupling to, or from,
the index area and look for the hypothesized PPI. You do not try to interpret
the direction and sign of the PPI post hoc (because it is not uniquely
determined). To see the slope of the regression coefficients and how they
differ you would simply plot the response of an area showing a PPI against
the response of the index region for both task conditions that constitute
the two level of the psychological factor. This illustrates quantitatively
the regression analysis above.
I hope this help - Karl
|