Back to PLS HelpPosted on 11/07/21 18:03:45
Number of posts: 9
I want to confirm that my comprehension of clusters is correct:
Based on other posts on this message board, I understand that PLS gives us the possibility to output clusters to help us with summarizing data. Clustering is not done as a test of significance. This is done through permutation/bootstrapping.
Thus, if I have a voxel that survives BSR thresholding (e.g., at plus or minus 3 BSR), then it can be interpreted as contributing to the LV even it is not the peak of the cluster. Is the logic correct?
We've been concerned about this because clustering needs to be interpreted with caution in fMRI research in general: "If the initial height threshold results in very large clusters that encompass several brain regions then anatomical specificity is compromised, as inferences cannot be made about individual areas within clusters." (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5153601/) . Even though it doesn't appear to apply to PLS, I want to be extra sure.
And finally, just to give a bit of background into how I got here: I did a ROI analysis with including only a key region of interest (e.g., the hippocampus) and it gave pretty much the same result than the main PLS. Based on the general purpose of PLS and the above, I wonder if ROI analyses (like I did) is beneficial at all.
Your thoughts about this and your help are very much appreciated, as always.
Replies:
Posted on 11/08/21 14:14:20
Number of posts: 394
I want to confirm that my comprehension of clusters is correct:
Based on other posts on this message board, I understand that PLS gives us the possibility to output clusters to help us with summarizing data. Clustering is not done as a test of significance. This is done through permutation/bootstrapping.
Thus, if I have a voxel that survives BSR thresholding (e.g., at plus or minus 3 BSR), then it can be interpreted as contributing to the LV even it is not the peak of the cluster. Is the logic correct?
We've been concerned about this because clustering needs to be interpreted with caution in fMRI research in general: "If the initial height threshold results in very large clusters that encompass several brain regions then anatomical specificity is compromised, as inferences cannot be made about individual areas within clusters." (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5153601/) . Even though it doesn't appear to apply to PLS, I want to be extra sure.
And finally, just to give a bit of background into how I got here: I did a ROI analysis with including only a key region of interest (e.g., the hippocampus) and it gave pretty much the same result than the main PLS. Based on the general purpose of PLS and the above, I wonder if ROI analyses (like I did) is beneficial at all.
Your thoughts about this and your help are very much appreciated, as always.
Hi Annick,
Cluster inference is very tricky at the best of times. Its part of the PLS report mainly for convenience, but you are correect that any voxel that exceeds your BSR threshold is a reliable contributor to the LV. The cluster issue has a bit to do with the spatial autocorrelation - you can play with you peak threshold and get either one big cluster or many little ones. This is partly why we have the abiity to view the voxel weights (saliences) to give you a better idea of the spatial distribution of the effects without obscuring things with the statistical threshold. This helps me - at least - interpret the spatial distribution.
Login to reply to this topic.
