PAC, surrogate data, and group level inference

Dear BST community,

I have an EEG dataset of 100 participants. I am interesting in calculating the PAC at each of the 68 regions as defined by the DK atlas. I would like to see which regions exhibit significant PAC values (relative to surrogate data).

I am trying to follow the paper below, but am unclear on how to make a group level inference as to whether a certain element of the PAC matrix (of a region) is significant?

https://www.sciencedirect.com/science/article/pii/S1053811915000804

From my understanding, for example for one region, I create 500 pinknoise signals, run PAC on each of the signals, then for each element, compare an individual participant's element PAC value to the distribution of PAC values created by the 500 pinknoise signals, and see if it significant as a chosen p-value.

However, this procedure gives significance at an individual/participant level if I am understanding correctly. How can I do this but on a group level inference?

My initial thought is to pretty much create a mirror dataset (i.e., 100 fake subjects, each with 68 regions, with each region having a pinknoise signal, and calculating PAC on this signal), and then compare (for each region, for each frequency element of the PAC matrix), the PAC values from the simulated data versus the real subject data using an independent t-test. Then, correct for multiple comparisons.

I honestly have no idea if this is close to being correct, so any feedback/suggestions would be greatly appreciated.

Best,
Paul

@Sylvain @pantazis?

Hi Paul:

I think what you suggest with the surrogate group of participants is valid. I would however suggest that the pink signals be those of EEG sensors and that you apply the same imaging kernel as for the test data to estimate source times series. Doing so, they will be affected by the same amount of crosstalk imposed by the source modeling step as your empirical data.

Thank you Sylvain!

Hi Sylvain,

I just had a quick follow-up question I would really appreciate your thoughts on.

Let's say I have two groups, and I would like to compare PAC values between these two groups.

Would it be okay to just compare the 2 groups directly?

An alternative I was thinking about is following the 2-step GLM analysis typically done in fMRI, with a first-level and a higher-level approach (e.g., https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FEAT/UserGuide). I was thinking the first level could be comparing each individual, within each group, to a surrogate PAC dataset. Then, the resulting statistics would be fed into higher-level (group-level) for group inferences. Although typically the first-level design contrast involves a contrast between tasks, I was thinking in this case, it could be done between the real data of a participant and surrogate data.

I could be totally wrong here, so apologizes in advance!

Best,
Paul