Dear Ruth and PPI experts,
I am a PPI user, not an
expert. There are two points I concentrated on. One is that you convolved the
psychophysiological interaction with HRF. I don’t think it’s necessary. Maybe
you can read the SPM manual and there is an example which treat the
psycholphysiological interaction as a regressor instead of a condition.
The other is the method you
used to extract VOIs. I am always puzzled by the most appropriate approach to
define VOIs when I use DCM and PPI. You used an extremely low threshold and an
extremely big sphere. Maybe that won’t miss any information, but redundant
information, even wrong information, will be included into your ROIs. It’s a dilemma.
If I use a low threshold or a long distance between individual peak loci and
group-level peak loci, I won’t be confident in the reliability of information. However,
if I use a high threshold, I won’t find so many subjects with activated VOIs (maybe
20 in 30).
On more question, what is
the relationship between the individual peak voxel and the group-level peak
loci?
Best Wishes!
Lu Feng
From: SPM (Statistical Parametric Mapping)
[mailto:[log in to unmask]] On Behalf Of Ruth van Holst
Sent: 2010年11月22日 20:32
To: [log in to unmask]
Subject: [SPM] PPI question
Dear PPI experts,
I am currently working with
Psychophysiological interaction (PPI) analyses and have some question about my
applied method so far.
I did the following:
To identify the left IFG activation for
each participant we examined the F contrast of all conditions at p<0.999
uncorrected, to obtain similar amount of voxels in or VOI across all
subjects. The deconvolved time series from a 16 mm radius sphere around the
individually defined peak activated voxel within the left IFG was extracted (30
participants). The PPI was calculated as the element by element product of the
left IFG time series (the first eigenvariate from all voxels' time series) and
a vector coding for the effect of task. These products were subsequently
re-convolved with the hemodynamic response function (HRF). The interaction term
was then entered as a regressor in a first level model together with the time
series of the left IFG and the vector coding for the task effect. The models
were estimated and contrasts generated to test the effects of positive PPI.
These contrast images for the PPI effects were then entered in a second level
analysis.
These choices were based on he SPM forum
list, were I understood the following:
“you should think of the contrast you
use to display the activations as simply a window to which voxels will be
extracted. SPM goes to the xyz locations of the active voxels in your VOI, gets
the actual data from your processed images and then adjusts the data for any
effects of interest you specify. The contrast you use to display activations
merely identifies the voxels to extract data from it does not affect the data
that are extracted.” By Darren Gitelman
Therefore I used a absurdly low threshold
of p<0.999 to capture all activity in the VOI and to make sure that the
below mentioned advise could be assured.
“Always use ALL the voxels in the
ROI. You don't want to compare a subset of voxels in one subject to a subset in
other subjects, then you might be getting apples and oranges. This is of course
a bigger issue if you were to use a larger ROI.” By Donald McLaren
I do think that a sphere of 16 radius is
quite large, but when using a smaller radius I could not detect activity in
this region for each participant. This would be a problem because excluding
participant form this analysis would be a waste as well.
The results from this analyses look
sensible, but I am unsure if my method is decent enough to pursue writing it down.
I hope that you are able to give your
opinion about my chosen PPI strategy.
Thanks in advance for your time,
Best wishes
Ruth