Dear gPPI users,

Recently I have run a gPPI analysis (for the first time) on our data, using an anatomical mask from the AAL atlas as seed region. Currently, I am looking at the second level analysis and notice some autocorrelation in the results (E.g. if I would choose an amygdala AAL atlas mask as input seed region for the gPPI, in my second level I'd find some amygdala activity -> functional connectivity between the amygdala and the amygdala). My question is; does the gPPI analysis usually exclude the input/seed region in its output results/should there be no auto-correlation/connectivity? 

If it normally does, then I'd be interested in knowing how come I do see (limited) autocorrelation in our data. One thing I could imagine is that our data has a voxel size of 3x3x3 and the AAL mask is 2x2x2 (i.e. distinct voxel sizes). Would you expect this to be related to the autocorrelation? Besides, if you'd think it's not feasible to run the analysis this way with the distinct voxel sizes I'd also be happy to hear your thoughts on this.

Thanks in advance for your input!
ACVDH

Originally posted by acvdh: I would not expect autocorrelation in the data. PPI is about the difference in connectivity with the seed between tasks. I suppose it's theoretically possible if the ROI is big enough and the connectivity between the mean time course and the seed changes when completing the task. 

Could you provide more details about the first-level models and contrast that you used from the 1st level. How is the seed ROI defined? Did you create a separate image of it? What is the duration of your events? I'm wondering if there is a possibility that events with duration = 0 are causing this effect. Another option may be that there are not enough events.

The mask voxel size should not make a difference as it's resized on the fly to generate the data. 

Dear Donald,

Thanks a lot for your reply and thinking along with us!

Here are some answers to your questions:
First level contrast: "Audio (30 seconds) > Baseline (30 seconds)" 30 seconds of listening to an audioscript versus 30 seconds of rest. This is repeated 3 times (so there are 3 audio blocks and 3 baseline blocks).
ROI definition: .nii image of the region in the AAL atlas, in this case it was the amygdala. It indeed is a separate image (same image for each subject)

A small amount of events causing this effect, which you mentioned, triggered me, as I imagine that we may have little events as compared to e.g. a stop signal task. How many events would you recommend as a minimum for gPPI? Something we discussed in our team is that the event we have might be too long to probe e.g. the amygdala effect, as it may elicit only a brief response. Would it be worth considering cutting the data into smaller fragments/shorter events - or would your expectation be that this would only mess up the data/analysis?

In case you have any additional thoughts on the occurrence of the autocorrelation based on the information I provided, I would of course be glad to hearing them!

Thanks again!
ACVDH
 
Originally posted by Donald McLaren:

Hi,

Two general comments:
(1) The number of events is not an issue with block designs as you have long enough events that there shouldn’t be correlations between the columns of the design matrix. Can you check the correlations of the design matrix?
(2) Baseline should not be modeled as a condition. If it is modeled, I think this could be the source.

-Donald



Originally posted by acvdh: