multiple y variables & Interaction variable

Posted on 02/25/20 20:11:40

Number of posts: 10

Hello Baycrest PLS community!

I have three questions regarding the variables that can be handled using this package.

1) My understanding is that using the command line I can use the PLS package with the data table, not only with imaging as Y. Is this correct?

2) If so, would it be possible to use this package with multiple y variables in the model?

3) In my research, I observe significant interaction between two variables on my dependent (Y) variable when running a regression model. Is it possible to include an interaction variable in PLS in the package? If so, then how?

Thank you very much for your help and expertise.

Peter

Untitled Post

Posted on 02/26/20 07:35:27

Number of posts: 392

quote:

Hello Baycrest PLS community!

I have three questions regarding the variables that can be handled using this package.

1) My understanding is that using the command line I can use the PLS package with the data table, not only with imaging as Y. Is this correct?

2) If so, would it be possible to use this package with multiple y variables in the model?

3) In my research, I observe significant interaction between two variables on my dependent (Y) variable when running a regression model. Is it possible to include an interaction variable in PLS in the package? If so, then how?

Thank you very much for your help and expertise.

Peter

Hi Peter - the answer to your questions is 'yes' but can you give me a better idea of what you are analyzing so I can give you more details?

Randy

Untitled Post

Posted on 02/29/20 18:12:42

Number of posts: 10

quote:

Hi Peter - the answer to your questions is 'yes' but can you give me a better idea of what you are analyzing so I can give you more details?

Randy

Hello Randy!

Thank you for your prompt reply and sorry for the delay. It's been a busy week of experiments.

Most definitely! I am investigating Alzheimer's disease pathophysiology using multiple PET tracers for amyloid, tau and neuroinflammation. I have done some analyses using regression models and found significant interactions between amyloid and neuroinflammation. Thus, it motivated me to see in multivariate levels how amyloid, neuroinflammation, and amyloidxNeuroinflammation interaction covary together on tau. I have 35 different ROIs currently with all the of the 3 PET tracers. Based on these data, I conducted PLS path analysis using R including the interactions between all the possible 35 amyloid and 35 neuroinflammation ROIs and added 35 ROIs of tau as multiple y variables. This gave me results that supports the regression models as well.

Then, I thought using your package will be complementary and enables me to do PLS using tau images as the y variable since path analysis heavily depends on my assumptions/hypotheses. I would love to include your PLS package to show within significant latent variables (LVs) and which (behaviour) variables contribute significantly and how they correlate with the brain scores. I think using tau as an image PLS will show the covariance as a topography easily.

My goals are #1) voxel-based PLS using tau image and 35 amyloid ROIs and 35 neuroinflammation ROIs as "behaviour" data. #2) replicate #1 but using 35 tau ROIs instead of tau images. #3) repeat #1 & #2 but also including the interaction terms. I have successfully done the #1. But I am stuck with #2 and #3 and my previous questions were asked with these goals in mind.

PS. I want to also ask how I could generate the Singular value plot, Correlation Overview plot of significant LVs, and BSR as overlayed with brain image for #1 (ex. Tau images) purpose, as well as BSR results to represent for #2 (ROIs) purpose using command lines? So far, I am using the GUI and there seems to be a bug with the GUI on our Matlab (R2018 or R2015b) that does not allow selection of other LVs results. Also, command lines would be more flexible to use.

Thank you very much for your help! It is very much appreciated.

Best regards,

Peter

Untitled Post

Posted on 03/01/20 08:56:04

Number of posts: 392

Hi Peter,

The easiest way to run your models is to use the command line version, which is called "pls_analysis". It is called in the GUI version so the results file that is created is identical. The help text for the script indicates what each input and output field is, so if you are facile in Matlab, it should be easy to run models 2 and 3 in your list.

One question: how do you generate the interaction terms for your model? Is it the product of the variables?

Untitled Post

Posted on 03/01/20 13:43:37

Number of posts: 10

quote:

Hello Randy,
Essentially, yes. In the regression models, I use the following syntax to get the interaction variables. Tau ~ amyloid*Neuroinflammation. In the pls path analysis, I generated the interaction variables by multiplying the amyloid ROI values with Neuroinflammation ROI values and input the product as the interaction variables in the models. Do you think it's also applicable with this PLS as well?
Thank you for your help. I will also look into the pls_analysis command line!
Best regards,
Peter
Hi Peter,

The easiest way to run your models is to use the command line version, which is called "pls_analysis". It is called in the GUI version so the results file that is created is identical. The help text for the script indicates what each input and output field is, so if you are facile in Matlab, it should be easy to run models 2 and 3 in your list.

One question: how do you generate the interaction terms for your model? Is it the product of the variables?

Untitled Post

Posted on 03/01/20 14:27:36

Number of posts: 392

quote:

Hello Randy, Essentially, yes. In the regression models, I use the following syntax to get the interaction variables. Tau ~ amyloid*Neuroinflammation. In the pls path analysis, I generated the interaction variables by multiplying the amyloid ROI values with Neuroinflammation ROI values and input the product as the interaction variables in the models. Do you think it's also applicable with this PLS as well? Thank you for your help. I will also look into the pls_analysis command line! Best regards, Peter

Hello Randy, Essentially, yes. In the regression models, I use the following syntax to get the interaction variables. Tau ~ amyloid*Neuroinflammation. In the pls path analysis, I generated the interaction variables by multiplying the amyloid ROI values with Neuroinflammation ROI values and input the product as the interaction variables in the models. Do you think it's also applicable with this PLS as well? Thank you for your help. I will also look into the pls_analysis command line! Best regards, Peter

I don't think including an interaction term as you do in a regression model will do much b/c PLS does not adjust for the correlations amongst predictor variables as is done in regression. The interaciton may come out on its own with just the regular variables where you could have latent variables that are primarily driven by one set of measures or the other, and another latent variable that reflects their joint prediction.

Untitled Post

Posted on 03/03/20 15:28:05

Number of posts: 10

quote:

I don't think including an interaction term as you do in a regression model will do much b/c PLS does not adjust for the correlations amongst predictor variables as is done in regression. The interaciton may come out on its own with just the regular variables where you could have latent variables that are primarily driven by one set of measures or the other, and another latent variable that reflects their joint prediction.

Hi Randy,

I see. Could you elaborate more on "The interaciton may come out on its own with just the regular variables where you could have latent variables that are primarily driven by one set of measures or the other, and another latent variable that reflects their joint prediction"? I think I could see but below is my train of thoughts that is holding a break.

My understanding of interaction effect is the effect of variable A in the presence of variable B is different when variable A is alone. How will this be described by having two latent variables that show different patterns of covariance from the behavior variables? In PLS, the latent variables are orthogonal to each other?

Thank you for your clarification!

Best regards,

Peter

Untitled Post

Posted on 03/03/20 17:57:16

Number of posts: 392

quote:

Hi Randy,

I see. Could you elaborate more on "The interaciton may come out on its own with just the regular variables where you could have latent variables that are primarily driven by one set of measures or the other, and another latent variable that reflects their joint prediction"? I think I could see but below is my train of thoughts that is holding a break.

My understanding of interaction effect is the effect of variable A in the presence of variable B is different when variable A is alone. How will this be described by having two latent variables that show different patterns of covariance from the behavior variables? In PLS, the latent variables are orthogonal to each other?

Thank you for your clarification!

Best regards,

Peter

You won't get an interaction term in the classic sense, but rather latent variables (LV) that may reflect unique and shared prediction of your outcome measure. For example, if we have your model 1 with tau images as outcome and the 35 amyloid ROIs and 35 neuroinflammation ROIs as predictors you could get something like:

one LV that reflect tau uptake that is uniquely predicted by amyloid (only the amyloiod ROIs have non-zero weights)

one LV that reflect tau uptake that is uniquely predicted by neuroinflammation ROIS (only they have non-zero weights)

one LV that reflect tau uptake that is a combinations of both ROI (where some subset of each have non-zero weights)

Given the dimensionality of your data you're like to have more than just three and probably not as clean as my scenario, but hopefully you get the idea.

Untitled Post

Posted on 07/02/20 18:38:41

Number of posts: 10

quote:

You won't get an interaction term in the classic sense, but rather latent variables (LV) that may reflect unique and shared prediction of your outcome measure. For example, if we have your model 1 with tau images as outcome and the 35 amyloid ROIs and 35 neuroinflammation ROIs as predictors you could get something like:

one LV that reflect tau uptake that is uniquely predicted by amyloid (only the amyloiod ROIs have non-zero weights)

one LV that reflect tau uptake that is uniquely predicted by neuroinflammation ROIS (only they have non-zero weights)

one LV that reflect tau uptake that is a combinations of both ROI (where some subset of each have non-zero weights)

Given the dimensionality of your data you're like to have more than just three and probably not as clean as my scenario, but hopefully you get the idea.

Dear Randy,

Sorry for the delay in response as my attempts to reply kept failed due to some technical difficulties.

Yes, it makes complete sense!

For the sake of completion, then, it is still possible to create an interaction term by multiplying two variables of interests and use this interaction variables (35 x 35 = 1225) ?

Thank you for your incredible support!

Best regards,

Peter

Untitled Post

Posted on 07/02/20 18:48:50

Number of posts: 392

quote:

Dear Randy,

Sorry for the delay in response as my attempts to reply kept failed due to some technical difficulties.

Yes, it makes complete sense!

For the sake of completion, then, it is still possible to create an interaction term by multiplying two variables of interests and use this interaction variables (35 x 35 = 1225) ?

Thank you for your incredible support!

Best regards,

Peter

Hi Peter - you can certainly create such a set of interactions - I would be concerned about excess collinearity and interpretabilty, but otherwise, go for it :)

Untitled Post

Posted on 07/08/20 19:41:25

Number of posts: 10

quote:

Hi Peter - you can certainly create such a set of interactions - I would be concerned about excess collinearity and interpretabilty, but otherwise, go for it :)

Haha, yes! I will keep that in mind.

Thank you for your amazing support! It's much appreciated!

Best regards,

Peter

Login to reply to this topic.

Toronto, Ontario

Canada M6A 2E1

Phone: (416) 785-2500

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

Privacy Statement - Disclaimer - © 2020 BAYCREST HEALTH SCIENCE. ALL RIGHTS RESERVED