ROI connectivity threshold

[amnon yacoby]

Dear Nick,

 I would appreciate some advice on a topic that could be beyond those that are discussed in the  forum.

I am currently doing a task fMRI analysis, based on pre-defined anatomic ROIs. The main analysis is RSA within each ROI (I am using matlab code). In addition, I planned to test the functional connectivity between the ROIs. This analysis is based on 2^nd-order RSA between the ROIs (Kriegescorte et al, 2008, Pillet et al, 2020). My question refers to testing the significance of the resulting functional connectivity:

What would be the recommended test/simulation to test the significance of a pair-ROI connectivity: the problem I have is what threshold could be set for testing the significance? (I intend of course to apply multiple comparisons correction once such threshold is set). I thought of setting the mean/median connectivity between all ROIs as this threshold, but I could not find any reference for this or for other threshold.    

 Thanks a lot for your assistance,

Best regards,

Amnon

[Nick Oosterhof]

Greetings,

On 20 Apr 2024, at 10:55, amnon yacoby <amnon....@gmail.com> wrote:

I am currently doing a task fMRI analysis, based on pre-defined anatomic ROIs. The main analysis is RSA within each ROI (I am using matlab code). In addition, I planned to test the functional connectivity between the ROIs. This analysis is based on 2^nd-order RSA between the ROIs (Kriegescorte et al, 2008, Pillet et al, 2020). My question refers to testing the significance of the resulting functional connectivity:

What would be the recommended test/simulation to test the significance of a pair-ROI connectivity: the problem I have is what threshold could be set for testing the significance? (I intend of course to apply multiple comparisons correction once such threshold is set). I thought of setting the mean/median connectivity between all ROIs as this threshold, but I could not find any reference for this or for other threshold.    

I don’t have much experience with functional connectivity, and less when it comes to testing for significance. 

That said, have you thought about a Monte Carlo simulation with randomization? In general the approach is:

1) define the null and alternative hypothesis.

2) compute some summary value on the original data, so that the alternative hypothesis results in a larger summary value than the null hypothesis. 

3) shuffle (randomize) the data in a way that is consistent with the null hypothesis to get null data, and compute the summary value each time. Repeat N_ITER times.

4) see if the original summary value is within the top 5% (or whatever you choose for the alpha level) in the distribution of the null data.

Could this work for your analyses? I can imagine that finding a good way to shuffle is not trivial. Due to spatial and temporal correlation (dependence) in fRMI data, a random shuffle of voxel locations or volumes over time does not seem valid to me.

best,

Nick

[amnon yacoby]

Dear Nick,

Thanks a lot for your response.

Yes, I think I can use the monte-carlo simulation method. I thought of using the whole 2nd-order RSA values among all ROI-pairs data-set, from which I exclude the values that refer to the tested ROI-pairs. Thus, in this reference data-set the similarities that pertain to the voxels of the test-set are not present. I will, then, sample in each permutation from this reference data-set.

Many thanks!

Amnon    

[amnon yacoby]

Hi Nick,

I implemented and ran the simulations and obtained the functional network P-values results. I, then, corrected these results for the multiple comparisons of all netwoking pairs combinations.

Thanks a lot to your advise!

Amnon