Print

Print
Hi Stephen,

Thanks very much for the reply. To go into a bit more detail: we simulated a "from-voxel" timecourse by making a vector of impulses at random intervals of 8 - 18 s. These impulses were assigned to one of two conditions. The magnitude of each impulse was varied randomly. A "to-voxel" was created as a condition-dependent scaling-factor times the from-voxel. We used scaling factors of 1 and -1 for conditions A and B, respectively, for instance. For other to-voxels we used 1 versus 0 and 10 versus 1. What we wanted was to reconstruct these scaling factors per condition. So for instance, if we know by construction that in condition B more positive responses at voxel 1 are associated with less positive responses at voxel 2, we would like to get a negative regression coefficient when regressing the to-voxel 2 on the from-voxel 1 for condition B.

Using simple home-brewed software (which did things a little differently than SPM) to analyze the simulated data we did get the coefficients we expected, e.g. if the scaling factors per condition were 10 and 1 then these values were indeed found as regression coefficients. Initially, we'd tried to get the same results by entering only a single condition in the standard SPM PPI interface, but this could even change the expected sign of the coefficient. So we weren't sure whether we might be making a mistake somewhere in the analysis, or whether we just shouldn't be using the standard PPI function for this purpose, possibly because of the implicit baseline problem you mentioned, which we'll definitely have to think about.

Best Regards,

Thomas

On Fri, Jul 17, 2009 at 4:57 PM, Stephen J. Fromm <[log in to unmask]> wrote:
On Thu, 16 Jul 2009 21:16:12 +0100, Thomas Gladwin
<[log in to unmask]> wrote:

>Hi,
>
>We're having some trouble with the interpretation of a PPI result. We found
>a positive significant interaction (this was in an event-related design),
>but wanted to know what kind of relationships this resulted from. For
>instance, was there a positive correlation in one condition and a negative
>correlation in the other; or a positive and a zero correlation; or a more
>versus less positive correlation, etc.

First you should ask how meaningful these other PPIs are.  That is, if the
original PPI used "A - B" as the input contrast (the "psycho-" in "PPI"), then
taking it apart would mean doing PPIs for A and B separately.  But everything
in fMRI and (some of) PET is based on differences, meaning you can't interpret
absolute levels.  So a PPI for "A" would really mean a PPI for "A - (some
baseline)".

[Maybe this doesn't apply to your case; maybe your original PPI was for
something like "(A1 - A2) - (B1 - B2)", in which case the two component
contrasts are themselves differences.]

That can be problematic, in that if I recall correctly SPM insists on a
difference contrast (enforced by the PPI user interface), meaning that the
baseline would have to be explicitly modeled.  And then there's the issue of
conceptual interpretations of these "sub-PPIs," which could possibly be
problematic.

>What one of our colleagues tried was to run the PPI with each condition
>separately, and then use the T- and contrast values to determine whether
the
>relationship in that condition was positive, negative or around zero. A
>negative regression weight was found, which we at first thought might imply
>some sort of negative interaction.
>
>However, we're not quite sure whether this is a valid approach. We also used
>simulated data to try to understand how the analysis works, and this
>resulted in the expected PPIs, but the per-condition didn't seem to bear any
>absolute relationship to the simulated regression coefficients. For
>instance, in one condition we modelled impulses in a certain "dependent"
>voxel as 10 times an impulse in the independent voxel (which we used for the
>VOI). In the other condition, the scaling factor was 1. But this resulted in
>a positive and negative "partial PPI" rather than more and less positive values.

I didn't quite follow the above paragraph.

>So our primary question is, should we even be expecting these per-condition
>PPIs to provide a meaningful value, in the sense that negative values for
>one condition imply a true negative physiological interaction in that condition?
>
>Thanks very much in advance.