Back to PLS HelpPosted on 04/25/14 11:19:12
Number of posts: 36
Greetings PLS experts,
I recently ran a seed PLS analysis focused on resting state and the DMN, with 7 subjects, just as an exercise for myself. I followed the Grigg and Grady (2010) paper, in the sense that I used the PCC coordinates as well as split by data in blocks of 5 TR (10 seconds), which provided me 35 blocks in total.
I thus ran 1000 permutations and 1000 bootstrap, but when viewing my results, my first 6 latent variables were shown to be significant.
My first LV very much looked like the standard DMN reported as determined by visual comparison as well as going through the cluster report.
However, the next 5 LV's, although showing grey matter clusters (I'm assuming not related to noise or other artefacts), were not expected. From the papers I've gone through, I can't recall seeing something to the extent of 6 or 7 significant LV's (Grigg and Grady I believe reported their first 2 LV's).
I should note that the cross block covariance value for the first LV was 29%, and then a drop where the 2nd LV is around 8% and from here a very gradual decline.
I just wanted to ask if seeing so many LV's as significant should raise any alarms (potentially poor preprocessing?). The sample size is low (Grigga and Grady had 18 I belive compared to my 7). Also, there's the issue of the other significant LV's explaining a relatively low amount of cross block covariance. I was just hoping for some insight into this issue of having something like 6 LV's come out as significant.
Any help would be great. Thank you.
Paul
Replies:
Posted on 04/26/14 02:18:29
Number of posts: 394
Hi Paul,
Can I get a bit more information? What is the total number of LVs you get from your analysis? 35? If so, it may be expected that you would get many significant LVs since you will likely be capturing changes in the functional connections across blocks. While these may be small, relative to the first LV which would represent the common functional connections across all blocks, these other LVs are no less important. As to whether they might reflect artifact, that will depend on the spatial pattern you see on the bootstrap image. For example, if you get a lot of edge effects, it may represent residual motion between runs. Its hard to say for sure without seeing the results. If you can post a few snapshots of these results I would be happy to take a look.
Randy
Posted on 04/26/14 07:51:56
Number of posts: 36
Hi Paul,
Can I get a bit more information? What is the total number of LVs you get from your analysis? 35? If so, it may be expected that you would get many significant LVs since you will likely be capturing changes in the functional connections across blocks. While these may be small, relative to the first LV which would represent the common functional connections across all blocks, these other LVs are no less important. As to whether they might reflect artifact, that will depend on the spatial pattern you see on the bootstrap image. For example, if you get a lot of edge effects, it may represent residual motion between runs. Its hard to say for sure without seeing the results. If you can post a few snapshots of these results I would be happy to take a look.
Randy
Hi Randy,
thank you very much for the reply. Yes, the total number of LVs is 35. It was my initial thought that since this analysis was being split into 35 blocks (and thus LVs) with a seed analsyis, that it was OK to have the significant LVs reported be beyond 1 or 2, but the Grigg and Grady (2010) paper only focused on the first two LVs, as well as most papers only report the 1st or 2nd LV as significant (although those are usually task related PLS, so nothing like 35 LV's in total is ever involved).
In regards to artifacts, I don't see the typical edge/rim effects related to motion artifacts, and it seems as though majority of the voxels in the BSR images can be attribtued to grey matter clusters (although some hints of voxels in ventricle related areas, I don't think it would make the data completely nonsense).
I've posted below LVs 3-5 with their crossblock covariance all set to a BSR of 3.3
Thank you very much for the help.
-Paul
LV3 (4.85 %)
LV4 (4.21 %)
LV5 (3.72%)
I'm Online
Posted on 04/26/14 07:56:59
Number of posts: 229
Hi Paul..
Unfortunately, the links to the images are broken, so we can't view them...
It may well be that there were additional significant LVs in the Grigg/Grady paper. They appear to have chosen to focus on those that were substantial contributors:
We report here the temporal and spatial patterns of the two
primary LVs, accounting for 42% of the covariance in the data
(each of the remaining LVs accounted for <5% of the covariance.
You may want to contact Dr. Grady to find out if my supposition is true..
cheers,
Nancy
Posted on 04/26/14 08:24:32
Number of posts: 36
Hi Paul..
Unfortunately, the links to the images are broken, so we can't view them...
It may well be that there were additional significant LVs in the Grigg/Grady paper. They appear to have chosen to focus on those that were substantial contributors:
We report here the temporal and spatial patterns of the two
primary LVs, accounting for 42% of the covariance in the data
(each of the remaining LVs accounted for <5% of the covariance.
You may want to contact Dr. Grady to find out if my supposition is true..
cheers,
Nancy
Hi Nancy,
great, thanks for the help. I think the links should be working now (I hope...). Going back over the paper, that very well could be it, since it looks as though it was never explicitly said that LVs 1 and 2 were the only LVs to come out as significant, just that LV1 and LV2 were looked at. Thank you!
Paul
Posted on 04/26/14 08:28:56
Number of posts: 394
Hi Nancy,
great, thanks for the help. I think the links should be working now (I hope...). Going back over the paper, that very well could be it, since it looks as though it was never explicitly said that LVs 1 and 2 were the only LVs to come out as significant, just that LV1 and LV2 were looked at. Thank you!
Paul
Paul - can you post the images elsewhere? Our forum program does not allow embedded graphics. Also, it might be nice to see the correlation bar graphs for some of these LVs to help interpret the effect.
Posted on 04/26/14 09:35:08
Number of posts: 36
Paul - can you post the images elsewhere? Our forum program does not allow embedded graphics. Also, it might be nice to see the correlation bar graphs for some of these LVs to help interpret the effect.
Hi Randy,
would sharing a link to an image sharing host be okay, such as a photobucket account?
Paul
Posted on 04/26/14 09:42:39
Number of posts: 394
Hi Randy,
would sharing a link to an image sharing host be okay, such as a photobucket account?
Paul
sure
Posted on 04/26/14 09:46:53
Number of posts: 36
sure
Great, hopefully this works...(sorry for this hassle with sharing the images).
I've included the r values in the title of the correlation bar plots shown from the scatter plots of the PCC seed vs brain scores.
They should be in order from LV1 to LV6 with their corresponding correlation bar plots.
http://s63.photobucket.com/user/pdhami06/library/PLS_results?sort=2&page=1
Thanks again for help.
-Paul
I'm Online
Posted on 04/30/14 17:06:16
Number of posts: 229
Hi Paul..
those look fine... Note however, that you would only be able to consider the first two LVs. If you look at the correlations with each block, many of those are not stable (CI includes zero) on LVs >3. (LV2 is a bit unstable, but there are some sequential blocks with stable correlations with the brain pattern)
The LV1/LV2 patterns do not match those of Grigg (e.g., very few positive saliences in LV1, and those are mainly in CSF & WM), but you have fewer subjects, and a different experimental design.
nancy
Posted on 05/04/14 09:07:37
Number of posts: 36
Hi Paul..
those look fine... Note however, that you would only be able to consider the first two LVs. If you look at the correlations with each block, many of those are not stable (CI includes zero) on LVs >3. (LV2 is a bit unstable, but there are some sequential blocks with stable correlations with the brain pattern)
The LV1/LV2 patterns do not match those of Grigg (e.g., very few positive saliences in LV1, and those are mainly in CSF & WM), but you have fewer subjects, and a different experimental design.
nancy
Hi Nancy,
great, thank you for the response.
Just had a question in regards to your previous post. In regards to the salience, is there an issue whether it is positive or negative? Or is it a warning or flag in a sense that the noise (white matter/ventricles) appeared with positive brain scores and the actual grey matter regions of interest as negative brain scores. From my understanding, the brain score is just a way to contrast between different regions as they relate to the conditions.
From my interpretation of the correlation plot for each negative block (PCC activity vs brain score), the greater the PCC activity is, the greater activity there is in regards to the negative/blue brain score regions, all of which represented (or looked similar in this case) to the DMN, so I assumed that having the negative/blue brain scores was OK, as long as it fit the context.
Thanks again.
Paul
I'm Online
Posted on 05/04/14 19:31:48
Number of posts: 229
think of it as the correlation between the design and brain sides of the equation -
for LV1- your "design" indicates negative seed-brain correlations combined with mostly negative brain saliences - those regions are thus positively correlated with your seed - as seed activity increases, so does activity in the other regions.
for LV2- you have both positive and negative brain saliences, and most of your stable correlations are negative - so I'll focus on those. Regions in yellow have a positive relation with the negative correlations - so you are looking at strong negative correlations with the seed across most of the resting state blocks. Regions in blue have a negative relation with the negative correlations, so have positive correlation with the seed at the same timepoints
thus, a combination of strong positive and strong negative correlations defines the relation with your seed for LV2 -
hope this helps
cheers
nancy
Posted on 05/05/14 02:10:38
Number of posts: 36
think of it as the correlation between the design and brain sides of the equation -
for LV1- your "design" indicates negative seed-brain correlations combined with mostly negative brain saliences - those regions are thus positively correlated with your seed - as seed activity increases, so does activity in the other regions.
for LV2- you have both positive and negative brain saliences, and most of your stable correlations are negative - so I'll focus on those. Regions in yellow have a positive relation with the negative correlations - so you are looking at strong negative correlations with the seed across most of the resting state blocks. Regions in blue have a negative relation with the negative correlations, so have positive correlation with the seed at the same timepoints
thus, a combination of strong positive and strong negative correlations defines the relation with your seed for LV2 -
hope this helps
cheers
nancy
Hi Nancy,
yes it does! Thank you very much.
I'm just curious as to whether the direction of brain score (positive or negative) is of importance in this context of the LV1 of the DMN. For example, the Grigg and Grady (2010) paper has positive correlations of the PCC seed with positive BSR values (red regions), where as in mine, there are negative correlations of the PCC seed with the negative BSR values (blue regions).
I know the regions aren't similar, but for example, if my data showed the exact same regions as in the Grigg paper, but with negative correlations and blue/negative BSR values/brain regions, I was under the assumption that both would still represent the regions highlighted as having a positive correlation to the PCC seed. Would that be correct?
I'm curious beacuse I preprocessed my data again, but with focus on removal of ventricle noise since it seemed prominent, and my new results show a "cleaner" looking DMN, but this time, the brain scores are positive, and the remaing sources of noise (white matter) appear as negative brain scores.
Thanks again for all the help.
Paul
I'm Online
Posted on 05/05/14 02:57:36
Number of posts: 229
I'm just curious as to whether the direction of brain score (positive or negative) is of importance in this context of the LV1 of the DMN.
only in the context of the brain saliences - a set of mostly negative patterns (seed/brain) like you find, is the same as a set of mostly positive patterns (seed/brain), like Grigg/Grady found. The interpretation does not change. Any LV can be "flipped" as long as you multiply both sides by -1.
I know the regions aren't similar, but for example, if my data showed the exact same regions as in the Grigg paper, but with negative correlations and blue/negative BSR values/brain regions, I was under the assumption that both would still represent the regions highlighted as having a positive correlation to the PCC seed. Would that be correct?
Correct -
I'm curious beacuse I preprocessed my data again, but with focus on removal of ventricle noise since it seemed prominent, and my new results show a "cleaner" looking DMN, but this time, the brain scores are positive, and the remaing sources of noise (white matter) appear as negative brain scores.
that is quite possible - as long as the brainscore correlation plots are also now positive, then removing the noisy signal from CSF did not change the results.
n
Posted on 05/08/14 01:37:26
Number of posts: 36
I'm just curious as to whether the direction of brain score (positive or negative) is of importance in this context of the LV1 of the DMN.
only in the context of the brain saliences - a set of mostly negative patterns (seed/brain) like you find, is the same as a set of mostly positive patterns (seed/brain), like Grigg/Grady found. The interpretation does not change. Any LV can be "flipped" as long as you multiply both sides by -1.
I know the regions aren't similar, but for example, if my data showed the exact same regions as in the Grigg paper, but with negative correlations and blue/negative BSR values/brain regions, I was under the assumption that both would still represent the regions highlighted as having a positive correlation to the PCC seed. Would that be correct?
Correct -
I'm curious beacuse I preprocessed my data again, but with focus on removal of ventricle noise since it seemed prominent, and my new results show a "cleaner" looking DMN, but this time, the brain scores are positive, and the remaing sources of noise (white matter) appear as negative brain scores.
that is quite possible - as long as the brainscore correlation plots are also now positive, then removing the noisy signal from CSF did not change the results.
n
Great!
Thank you so much for the help Nancy, really appreciate it.
Paul
Login to reply to this topic.
