Back to PLS HelpPosted on 05/22/14 02:49:14
Number of posts: 16
Hi,
I have run a task-PLS analysis contrasting an experimental memory condition and a control condition that produced a set of regions more active during the experimental condition. Two of these regions (left and right hippocampus) were theoretically interesting, so I decided to do a seed-PLS analysis to see if these two regions showed differential functional connectivity during the memory condition.
A significant LV was produced showing one network of regions with positive saliences (corresponding to regions "functionally connected" to the right HC) and another with negative saliences (corresponding to regions "functionally connected" to the left HC).
My question is how to interpret these positive and negative saliences. Intuitively, it seems to suggest that the regions with positive saliences positively correlate with the RHC (and not the LHC) and vice versa for the negative saliences.
But when you look at the r-values for regions, it seems that what it's really saying is that for regions with positive saliences, the r-value for a particular voxel is greater for the RHC seed relative to the LHC seed. This allows for the possibility that a voxel showing a reliable positive salience is *not* functionally connected to the RHC (e.g. r = .04), but is anti-correlated with the LHC (e.g. r = -.81)
If this is the case, it seems that one should avoid making a generalised statement like "Regions shown in warm colours are functionally connected to the RHC". Instead, one should really check the actual r-values for all the regions showing reliable effects to determine how each region is contributing to the overall effect.
Does this seem like a reasonable approach?
Apologies for the long-winded post.
Many thanks.
Reece Roberts
Research Fellow
School of Psychology, The University of Auckland
Centre for Brain Research
Email: r.roberts@auckland.ac.nz
Tel: +649 3737599 x86793
Replies:
Posted on 05/23/14 01:21:59
Number of posts: 394
Hi Reece,
your suggestion to look at the correlations is exactly right. Remember that LVs identify the differences but don't tell you the nature of the difference. You need to look at the raw data, in this case the correlations. Note the same ambiguity exists in analysis of mean changes, where you won't know if a difference is "more activation" or "less deactivation" unless you look at the actual mean data.
Posted on 06/12/14 19:47:47
Number of posts: 16
Thanks for your reply Randy!
I have a further question regarding seed-PLS that perhaps you guys could shed some light on: When running a seed-PLS analysis, is it ideal to remove the seed-voxel from the analysis?
According to Krishnan, Williams, McIntosh and Abdi (2011, p.464), we should be removing the voxel -- "Because Yseed already accounts for seed activity, the seed voxels are removed from X."
There is however, no option to do this in the PLS software.
I have played around with this and removed the voxel manually (from the st_coords and st_datamat variables for each subject). The results are quite interesting, and I'm not sure what to make of them.
When I run a regular behav-PLS analysis with a single seed, a single group and a single condition, removing the seed-voxel makes no difference at all to the p-value of the LV. Presumably this has something to do with the fact that the BSR of the voxel in question is 0.00 when it is left in the analysis (the salience for this voxel, as observed when clicking "View --> Brain LV", is the highest in the brain, however)
When I add another condition to the same analysis, there is now a difference in the p-value for LV1 (removing the voxel increases the p-value from .014 to .024). In addition, for the analysis that includes the seed-voxel, the BSR of the seed-voxel is very high (625.07). This voxel also has the highest salience in the brain.
What is best practice for seed-PLS analyses in terms of whether the seed-voxel should be included in the analysis?
Many thanks,
Reece
Posted on 06/15/14 16:19:08
Number of posts: 394
Thanks for your reply Randy!
I have a further question regarding seed-PLS that perhaps you guys could shed some light on: When running a seed-PLS analysis, is it ideal to remove the seed-voxel from the analysis?
According to Krishnan, Williams, McIntosh and Abdi (2011, p.464), we should be removing the voxel -- "Because Yseed already accounts for seed activity, the seed voxels are removed from X."
There is however, no option to do this in the PLS software.
I have played around with this and removed the voxel manually (from the st_coords and st_datamat variables for each subject). The results are quite interesting, and I'm not sure what to make of them.
When I run a regular behav-PLS analysis with a single seed, a single group and a single condition, removing the seed-voxel makes no difference at all to the p-value of the LV. Presumably this has something to do with the fact that the BSR of the voxel in question is 0.00 when it is left in the analysis (the salience for this voxel, as observed when clicking "View --> Brain LV", is the highest in the brain, however)
When I add another condition to the same analysis, there is now a difference in the p-value for LV1 (removing the voxel increases the p-value from .014 to .024). In addition, for the analysis that includes the seed-voxel, the BSR of the seed-voxel is very high (625.07). This voxel also has the highest salience in the brain.
What is best practice for seed-PLS analyses in terms of whether the seed-voxel should be included in the analysis?
Many thanks,
Reece
Hi Reece,
First, when you do your seed analysis, are you selecting just a single voxel or the area around a coordinate? You may want to do the latter as that will stabilize the signal across subjects.
There is a mask that gets applied in seed PLS for where the correlation between variiables in and X and Y are perfect. If you select a region rather than a voxel, this is less likely to happen. In practice, with >20K voxels, having the seed in X and Y doesn't affect things too much. If you are dealing with ROI data, or a subset of voxels, then you definitely want to remove the seed from one side.
Login to reply to this topic.
