open-discussion > threshold for functional connectivity
Showing 1-3 of 3 posts
Mar 28, 2017 07:03 PM | Ben Xu - NIH
threshold for functional connectivity
Hi,
I'm running a set of resting-state fMRI scans with two groups (between subject) and two sessions (within subject). At the second level, I used Seed-to-voxel analysis (GLM weighted) with an interaction contrast: (Group1 - Group2) - (Session1 - Session2) using a voxel-level threshold p < 0.05 (uncorrected) and a luster level threshold FWE < 0.05. The resulting connectivity clusters seem reasonable and corroborate the results of a separate task-based fMRI activation.
My question is: Is the voxel-level threshold too liberal or acceptable? Increasing the threshold to p < 0.001 renders no significant clusters.
I understand that using spm, cluster-level threshold is recommended to be set with voxel-level uncorrected p < 0.001, and the cluster-level corrected p < 0.05. Do these thresholds necessarily apply to functional connectivity clusters, too?
Thank you,
Ben
I'm running a set of resting-state fMRI scans with two groups (between subject) and two sessions (within subject). At the second level, I used Seed-to-voxel analysis (GLM weighted) with an interaction contrast: (Group1 - Group2) - (Session1 - Session2) using a voxel-level threshold p < 0.05 (uncorrected) and a luster level threshold FWE < 0.05. The resulting connectivity clusters seem reasonable and corroborate the results of a separate task-based fMRI activation.
My question is: Is the voxel-level threshold too liberal or acceptable? Increasing the threshold to p < 0.001 renders no significant clusters.
I understand that using spm, cluster-level threshold is recommended to be set with voxel-level uncorrected p < 0.001, and the cluster-level corrected p < 0.05. Do these thresholds necessarily apply to functional connectivity clusters, too?
Thank you,
Ben
Mar 30, 2017 09:03 PM | Zeus Gracia-Tabuenca - University of Zaragoza
RE: threshold for functional connectivity
Hi,
According to the simulations in Eklund et al (2016) a cluster-defining threshold (CDT) higher than p<0.001 is very susceptible to false positives, due to the autocorrelation implemented in the most popular softwares is underestimated and the bias is higher as cluster are bigger, so a CDT<0.05 is going to give you big clusters and probably more false positives. They found this bias in task-related and resting-state data, not with a functional connectivity (FC) approach, but you are going to get statistical maps anyway and the computation of the cluster correction is going to be the same, so you will get this bias as well. Also, a non-parametric permutation test will be a robust approach to correct for multiple comparison, and you can apply it if all your participants have the two sessions.
Cheers,
Zeus.
According to the simulations in Eklund et al (2016) a cluster-defining threshold (CDT) higher than p<0.001 is very susceptible to false positives, due to the autocorrelation implemented in the most popular softwares is underestimated and the bias is higher as cluster are bigger, so a CDT<0.05 is going to give you big clusters and probably more false positives. They found this bias in task-related and resting-state data, not with a functional connectivity (FC) approach, but you are going to get statistical maps anyway and the computation of the cluster correction is going to be the same, so you will get this bias as well. Also, a non-parametric permutation test will be a robust approach to correct for multiple comparison, and you can apply it if all your participants have the two sessions.
Cheers,
Zeus.
Apr 6, 2017 03:04 PM | Ben Xu - NIH
RE: threshold for functional connectivity
Thank you very much, Zeus, for your reply. I was at a conference
and couldn't respond to your message earlier.
Ben
Ben