Back to PLS Help

Group Behavioral PLS
khiersche
Posted on 03/18/20 16:18:19
Number of posts: 8
khiersche posts:

Hello everyone,

I am interested in running a group behavioral PLS correlation to compare two age groups. Does anyone know of a software package or a matlab script that can accomplish this? 

For context, I am a research assistant at Ohio State University, looking to analyze health and cognitive data we have collected, but we are not using any imaging data. 

Thank you,

Kelly Hiersche 

 

Replies:

Untitled Post
rmcintosh
Posted on 03/18/20 16:21:46
Number of posts: 385
rmcintosh replies:

quote:

Hello everyone,

I am interested in running a group behavioral PLS correlation to compare two age groups. Does anyone know of a software package or a matlab script that can accomplish this? 

For context, I am a research assistant at Ohio State University, looking to analyze health and cognitive data we have collected, but we are not using any imaging data. 

Thank you,

Kelly Hiersche 

 

You can do this with our PLS package.  The command-line version has all the scripts to do this. You just need to get the data into matlab matrices.

 

Let me know if you need more detailed help

 

Randy



Group Behavioral PLS
khiersche
Posted on 03/19/20 14:51:29
Number of posts: 8
khiersche replies:

Hello Dr. McIntosh & others,

Thank you for your response. After reading through the user manual and the comments at the top of the plsanalysis.m script I believe I was able to correctly run a group analysis on my dataset. But I am having trouble understanding the results. 

The command I ran was: 

result=pls_analysis({datamat1_group1,datamat1_group2},[nsubjects1,nsubjects2], 1, opt);

where method = 4 (regular multiblock PLS)

opt.stacked_behavdata = datamat2

Datamat1_group1 is size (33 x 36)

Datamat1_group2 is size (31 x 36)

nsubjects1 = 33

nsubjects2 = 31 

Datamat2 is size (64 x 5), with all participants from group 1 listed first in the same order as they are in datamat1_group1 followed by the participants from group 2 in the same order as they are in Datamat_group2. 

 

I have run non-group, behavioral PLS before, and I have always had the same number of LVs as columns in datamat2. In the results from my group analysis, I have 12 LVs, even though there are only 5 columns in my dataset2. I assumed that it may be 5 LVs for each group, but I am still not sure what the additional two LVs represent. 

Here is some information on the sizes of the result structures produced by this multiblock pls analysis:  

singular values: size (12 x 1)

lvcorrs: size (10 x 12)

perm_result.sprob: size (12 x 1)

After reading through the documentation, I looked at TBv, which is the task and behavior v values separately. In my case, I believe task refers to groups. TBv contains a 2x12 matrix and a 10x12 matrix. Does this mean that the first two latent variables are task (aka group latent variables? maybe one overall for each?) and the other ten are the 5 for group 1 and the 5 for group 2? I am really not sure how to interpret what each of these latent variables represents in this group analysis. 

It is very possible I am misunderstanding this. I am relatively new to PLS. I have read a few papers and spent a lot of time looking at the matlab code, but conceptually, I am definitely still learning.

Any help in interpreting these results would be greatly apprecaited! 

Best,

Kelly Hiersche 



Untitled Post
rmcintosh
Posted on 03/19/20 15:04:26
Number of posts: 385
rmcintosh replies:

quote:

Hello Dr. McIntosh & others,

Thank you for your response. After reading through the user manual and the comments at the top of the plsanalysis.m script I believe I was able to correctly run a group analysis on my dataset. But I am having trouble understanding the results. 

The command I ran was: 

result=pls_analysis({datamat1_group1,datamat1_group2},[nsubjects1,nsubjects2], 1, opt);

where method = 4 (regular multiblock PLS)

opt.stacked_behavdata = datamat2

Datamat1_group1 is size (33 x 36)

Datamat1_group2 is size (31 x 36)

nsubjects1 = 33

nsubjects2 = 31 

Datamat2 is size (64 x 5), with all participants from group 1 listed first in the same order as they are in datamat1_group1 followed by the participants from group 2 in the same order as they are in Datamat_group2. 

 

I have run non-group, behavioral PLS before, and I have always had the same number of LVs as columns in datamat2. In the results from my group analysis, I have 12 LVs, even though there are only 5 columns in my dataset2. I assumed that it may be 5 LVs for each group, but I am still not sure what the additional two LVs represent. 

Here is some information on the sizes of the result structures produced by this multiblock pls analysis:  

singular values: size (12 x 1)

lvcorrs: size (10 x 12)

perm_result.sprob: size (12 x 1)

After reading through the documentation, I looked at TBv, which is the task and behavior v values separately. In my case, I believe task refers to groups. TBv contains a 2x12 matrix and a 10x12 matrix. Does this mean that the first two latent variables are task (aka group latent variables? maybe one overall for each?) and the other ten are the 5 for group 1 and the 5 for group 2? I am really not sure how to interpret what each of these latent variables represents in this group analysis. 

It is very possible I am misunderstanding this. I am relatively new to PLS. I have read a few papers and spent a lot of time looking at the matlab code, but conceptually, I am definitely still learning.

Any help in interpreting these results would be greatly apprecaited! 

Best,

Kelly Hiersche 

Hi Kelly -  I think you want a regular behavior PLS (method = 3).  Multiblock combines task (mean-centerd) and behavior PLS, which is why you get twice as many LVs

 

Randy



Group Behavioral PLS
khiersche
Posted on 03/19/20 15:33:47
Number of posts: 8
khiersche replies:

Hello again,

Thank you for your quick reply. I have used the method=3 on each of my groups, by running separate PLS analysis. I was hoping that by combining my two groups and using method=4 I would be able to test the group differences in my data. Is that correct? 

If so, how do I go about interpretting the results of the 12 LVs?

 

Thank you for all of your help! 

Kelly 



Untitled Post
rmcintosh
Posted on 03/19/20 15:48:59
Number of posts: 385
rmcintosh replies:

quote:

Hello again,

Thank you for your quick reply. I have used the method=3 on each of my groups, by running separate PLS analysis. I was hoping that by combining my two groups and using method=4 I would be able to test the group differences in my data. Is that correct? 

If so, how do I go about interpretting the results of the 12 LVs?

 

Thank you for all of your help! 

Kelly 

you can test for group differences using both method 3 and 4.  Method 3 will identify similarities and differences in a data-driven mode, whereas method 4 you'll need to provide contrasts. 



Group Behavioral PLS
khiersche
Posted on 03/19/20 16:13:27
Number of posts: 8
khiersche replies:

Thank you very much for the clarification! 



Group Behavioral PLS
khiersche
Posted on 03/24/20 14:12:28
Number of posts: 8
khiersche replies:

Hello again,

First, thank you for all of your help up to this point. I have two questions, regarding interpretting my analysis. 

Question 1: I have a bar graph which has ten bars: the height of each bar corresponds to the correlation values in the first column of result.boot_result.orig_corr. Each bar has an error bar, calculated using result.boot_result.orig_corr, result.boot_result.llcorr and result.boot_result.ulcorr. Am I correct in assuming that these are the correlations for the 5 cognitive measures, in the same order as they are given in datamat2, and that the first five bars refer to the cognition scores of older adults and the second five refer to those of young adults? 

Question 2: How do I use this graph (or other information from the results) to interpret group differences? 

Additional information: I ran the analysis with method = 3. I had two datamat1 inputs, one for older adult (size 33 x 35) health data and one for young adult (size 31 x 35) health data. I used one datamat2 (size 64 x 5) which has all of the cognition information for both groups, with the older adult subjects first, followed by the young adult subjects. 

 



Untitled Post
rmcintosh
Posted on 03/25/20 09:38:34
Number of posts: 385
rmcintosh replies:

quote:

Hello again,

First, thank you for all of your help up to this point. I have two questions, regarding interpretting my analysis. 

Question 1: I have a bar graph which has ten bars: the height of each bar corresponds to the correlation values in the first column of result.boot_result.orig_corr. Each bar has an error bar, calculated using result.boot_result.orig_corr, result.boot_result.llcorr and result.boot_result.ulcorr. Am I correct in assuming that these are the correlations for the 5 cognitive measures, in the same order as they are given in datamat2, and that the first five bars refer to the cognition scores of older adults and the second five refer to those of young adults? 

Question 2: How do I use this graph (or other information from the results) to interpret group differences? 

Additional information: I ran the analysis with method = 3. I had two datamat1 inputs, one for older adult (size 33 x 35) health data and one for young adult (size 31 x 35) health data. I used one datamat2 (size 64 x 5) which has all of the cognition information for both groups, with the older adult subjects first, followed by the young adult subjects. 

 

Howdy

 

For Q1: the order is same as in datamat2.  The errorbars correspond to the "location" of the 95% confidence interval bounds

Q2: this will depend on whether the give LV shows similarities or differences - this you can infer from orig_corr and whether the confidence limits overlap. A lack of overlap suggests a difference, which you can confirm.  if you want to send me a screenshot of some of these resutls I can provide a bit more guidance: rmcintosh@research.baycrest.org



Group Behavioral PLS
khiersche
Posted on 04/24/20 10:31:15
Number of posts: 8
khiersche replies:

Hello again,

The lab I work in has continued to use this PLS software for group analyses. We ran a behavoral PLS comparing males and females on a variety of health and cognition measures and the results showed five significant LVs. We determined that there were not group differences between males and females, because the 95% confidence intervals overlapped for all of the cognition measures. However, there were some cognition measures in which both (or just one of) the male and female confidence intervals did not cross zero. Would we infer from that that we have a significant latent variable comprised of cognition measures (where the confidence interval does not cross zero) and health measures (with bootstrap ratios > 1.96), but that latent variable does not have any group differences? Basically, we are wondering if a latent variable shows no group differences, are the LVs still interpretable? 

Any insight would be greatly appreciated!

Thank you,

Kelly 



Untitled Post
rmcintosh
Posted on 04/24/20 10:37:10
Number of posts: 385
rmcintosh replies:

quote:

Hello everyone,

I am interested in running a group behavioral PLS correlation to compare two age groups. Does anyone know of a software package or a matlab script that can accomplish this? 

For context, I am a research assistant at Ohio State University, looking to analyze health and cognitive data we have collected, but we are not using any imaging data. 

Thank you,

Kelly Hiersche 

 

Hi Kelly - thanks for the question.  Behavior PLS when you have more than one group will identify latent variables that captures similarities and differences between groups. Thus, the LV you describe is basically telling you that female and males show the same brain-behavior mapping for the spatial pattern in the and for the behaviors that are reliable by confidence intervals. So, yes very interpretable

An illustration of this is in the appendix of this paper: https://www.dropbox.com/s/7m3mkpdkn54qydm/cipaper.pdf?dl=0



Group Behavioral PLS
khiersche
Posted on 04/24/20 11:31:12
Number of posts: 8
khiersche replies:

Thank you for your quick response and for the reference article - it is a helpful example!

I have a question about intepretting whether or not cognitive variables are significant, within a LV, under a few different conditions. Are my assumptions correct / How are these scenerios interpretted? 

1. Male and Female Error bars overlap eachother, but do not overlap with zero --> This is cognition variable is significant, but does not show sex differences.  

2. Male and Female Error bars overlap each other and also both overlap with zero --> This cognition variable is not significant and there are no sex differences. 

3. Male and Female Error bars do not overlap each other and both do not overlap with zero --> This cognition variable is significant, and shows sex differences. 

4. Male and Female Error bars do not overlap each other and one of the genders does not overlap with zero --> How is this interpretted? 

5. Male and Female Error bars overlap each other and one of the genders does not overlap with zero --> How is this interpretted?  

 



Untitled Post
rmcintosh
Posted on 04/24/20 12:19:19
Number of posts: 385
rmcintosh replies:

quote:

Thank you for your quick response and for the reference article - it is a helpful example!

I have a question about intepretting whether or not cognitive variables are significant, within a LV, under a few different conditions. Are my assumptions correct / How are these scenerios interpretted? 

1. Male and Female Error bars overlap eachother, but do not overlap with zero --> This is cognition variable is significant, but does not show sex differences.  

2. Male and Female Error bars overlap each other and also both overlap with zero --> This cognition variable is not significant and there are no sex differences. 

3. Male and Female Error bars do not overlap each other and both do not overlap with zero --> This cognition variable is significant, and shows sex differences. 

4. Male and Female Error bars do not overlap each other and one of the genders does not overlap with zero --> How is this interpretted? 

5. Male and Female Error bars overlap each other and one of the genders does not overlap with zero --> How is this interpretted?  

 

You should think of the first PLS you ran as a sort of omnibus test, which you woud follow up with specific comparisons. What I might suggest next is:

 

1) run the analysis within group

2) run a few analyses focusing on one or two behaviors to compare the groups directly.

 

The sensitivity of these will vary a bit, so we may need to do some comparisons back to the omnibus test, which I can tell you how to do if need be (basically you correlate the U matrices between analyses).




Login to reply to this topic.

  • Keep in touch

Enter your email above to receive electronic messages from Baycrest, including invitations to programs and events, newsletters, updates and other communications.
You can unsubscribe at any time.
Please refer to our Privacy Policy or contact us for more details.

  • Follow us on social
  • Facebook
  • Instagram
  • Linkedin
  • Pinterest
  • Twitter
  • YouTube

Contact Us:

3560 Bathurst Street
Toronto, Ontario
Canada M6A 2E1
Phone: (416) 785-2500

Baycrest is an academic health sciences centre fully affiliated with the University of Toronto