Hello,

We are running an ROI-ROI, gPPI analysis and have noticed that each edge/connection is being counted twice, and we hoped to get some advice on this. When doing the gPPI analysis, both terms' edges (i.e. both node A time course * task predicts node B time course and B * task predicts A time course) go into the group-level analysis; so, for example, when looking at 27 ROIs, the result should be 351 connectors, but instead we get double that (so 702). This really limits the ability to run this kind of analysis at a network level.

Is there any way to resolve this issue, or perhaps integrate a number of potential solutions (eg, add options to include the mean of the edges as one edge, or the max)? Given that integrating task-based analyses and network-level analyses could really be an important way forward in fMRI, we think that offering ways to address this limitation could be really useful for many groups.

Thank you for your help!
Laura

Any chance this could be acknowledged?...
We have a couple of studies that depend on this kind of analysis.

Thank you for your help,
Rany

Originally posted by Laura Jett:

Hi Laura,

I am not sure if I am understanding correctly your question, but if you are referring to the 'alternative settings for network-based inferences: Network Based Statistics' option in the ROI-to-ROI results explorer window, you are right that the current implementation of NBS uses an undirected graph when grouping connections into networks (and if, as in your case, the original adjacency matrix A is asymmetric, CONN will consider the matrix A|A' as the new symmetric adjacency matrix when determining groups of connections conforming a 'network') but after that the computation of both network-size and network-mass use the original (possibly asymmetric) graph and connectivity values (so that each "direction" of an edge is counted once if appropriate). I am not sure I understand why you find this limits the ability to run analyses at the network level, if you could please clarify.

Best
Alfonso
 
Originally posted by Laura Jett:

Thank you for your reply! We understand why we end up with an asymmetric matrix; however, we are still unsure on a few points:

• Does the matrix need to be symmetric for network-based statistics to be valid? If network-based statistics are valid with an asymmetric matrix, then how do the statistics work if both "directed edges" are counted?
• What is the interpretation? For example, if A → B is part of a significant cluster, can we infer the directionality of this edge?

Do you know of any implementations of network-based analysis that use gPPI as inputs so that we may see how others have handled these kinds of questions? We really appreciate all of your assistance with this matter!

Kind regards,
Laura

Hi,

I randomly came across a paper in which the upper and lower triangles were averaged in a PPI ROIxROI matrix (https://doi.org/10.1016/j.jad.2021.05.08...).
It might be that this is the most practical solution if one wants to run network analyses for a PPI design (with the other option of entering both sets of directed edges seeming potentially invalid).

Alfonso: what needs to be done in order to end up with a matrix of averaged triangles (resulting in one triangle and NAs?) that would then allow for network analyses?

Thanks,
Rany

Just bringing this up again, in case anyone has any experience with running network analyses with a gPPI approach.

Thanks,
Rany

Originally posted by Rany Abend:

Hi Laura,

Regarding (1), yes, NBS statistics are derived from the results of randomization/permutation simulations so they remain valid irrespective of how the networks are defined (e.g. as sets of connected nodes from directed or indirected graphs). I am not sure what you mean by "how do the statistics work if both directed edges are counted", the network-intensity measure from a network is computed by summing the number of suprathreshold edges within the network. For example, if a network of two areas A&B is identified and A->B is significant and B->A is not then the network intensity is 1, if both are significant then the network intensity is 2. The same logic applies to any other network computed during the permutation/randomization runs, so the final network-intensity statistics are derived by comparing the intensity of this network to the distribution of network intensities expected if the null hypothesis was true. 

And regarding (2), the interpretation of directionality differences in gPPI analyses is debated, unless you want to make a specific case about the directionality of specific connections within any of your significant networks I would generally recommend simply describing each network in terms of the nodes/ROIs within each network and the connections between them, rather than focusing on the specific directionality of these edges/connections (but it is perfectly fine either way)

Hope this helps
Alfonso
 
Originally posted by Laura Jett:

Thanks, Alfonso. This is helpful.

In case we wanted to average the top and bottom triangles, rather than enter both triangles into network analyses, could you help us figure out which data elements to average?

Thanks again,
Rany

Originally posted by Alfonso Nieto-Castanon:

Hi Rany,

Yes, one relatively simple way would be the following:

1) use the 'export data' button on the 2nd-level ROI-to-ROI results window to export the original (asymmetric) ROI-to-ROI matrices for all subjects into a single datafile

2) then take the data in those matrices and average it with its transpose, e.g. using something like:

  [data,names,coords,samples] = conn_mtx_read('/myfolder/data.nii');     % read from file
  data = (data + permute(data,[2,1,3])) / 2;                                            % average upper and lower triangular parts
  conn_mtx_write('/myfolder/symdata.nii', data, names, coords, samples); % write to file

and 3) then enter your new (symmetric) ROI-to-ROI matrices into a new 2nd-level analysis, e.g. using something like:

conn_module( 'glm' , ...
   'data', '/myfolder/symdata.nii', ...
   'design_matrix', X, ...
   'contrast_between', C, ...
   'contrast_within', M, ...
   'folder', '/myfolder/symanalyses/' )

entering in the X/C/M matrices the 2nd-level analysis associated design matrix / between-subjects contrast / within-subjects contrast, respectively (see https://web.conn-toolbox.org/resources/g... for details about the GLM CONN module)

Hope this helps
Alfonso
Originally posted by Rany Abend:

Thanks so much, Alfonso.
we'll give it a try.
We really appreciate it.

Rany
Originally posted by Alfonso Nieto-Castanon: