Covariates & distals in LPA mixture C... PreviousNext
Mplus Discussion > Latent Variable Mixture Modeling >
 Andrew Young Choi posted on Wednesday, November 23, 2016 - 4:12 pm
Hi Bengt & Linda, and happy Thanksgiving week.

I am in the process of modeling covariates and distal outcomes with a latent profile model (3-classes). As in p.182 of the Mplus User's Guide (example 7.17), my distal outcomes are latent factors.

I used the BCH method to simultaneously estimate covariates and distal outcomes in the latent profile model (Asparouhov & Muthen, October 7, 2014 - auxiliary variables in mixture modeling: Using the BCH method in Mplus to estimate a distal outcome model and an arbitrary secondary model).

I fixed the mean of the distal outcomes to 0 for one of the classes to examine differences in mean across classes. Unit variance identification was used to set the scale for the distal outcome factors. The class-specific means, SEs, and significance tests appear and the needed information is provided.

The problem begins when I attempt to include a command to regress distal outcomes on the covariates in the %OVERALL% section (e.g., distal ON covariate). I want to control for the relations among covariates and distals before class-specific means are estimated.

Although the model runs, it no longer provides class-specific estimates of distal outcome means. It only provides the mean and relevant estimates for the observed indicators for each class.

I was wondering if I could get your thoughts on how to address this issue? Thanks in advance!
 Bengt O. Muthen posted on Wednesday, November 23, 2016 - 4:32 pm
Have you tried TECH4?
 Andrew Y Choi posted on Wednesday, November 30, 2016 - 4:27 pm
Thanks-- yes, Tech4 provides the distal means for each profile. I'm curious why the mean estimates from tech4 and the class-specific intercept values (that represent distal outcome means now that they are endogenous variables) are different. This seems to be due to the distal outcome means (intercepts) now being conditioned on class AND the three binary covariates that I have in the model. Is that correct?

Also, in this model (similar to figure 2 in webnote 15, plus a path from c to x), do you have any suggestions in how to summarize distal mean differences conditioned on class and covariates? I am most interested in mean differences across classes, but can't ignore the covariates (and centering doesn't seem appropriate since they are binary).
 Bengt O. Muthen posted on Thursday, December 01, 2016 - 10:49 am
Q1. Yes, it is the usual difference between means and intercepts. You might center the covariates.

Q2. For ideas, see e.g. the paper on our website

Muthén, B., Brown, C.H., Masyn, K., Jo, B., Khoo, S.T., Yang, C.C., Wang, C.P., Kellam, S., Carlin, J., & Liao, J. (2002). General growth mixture modeling for randomized preventive interventions. Biostatistics, 3, 459-475. Mplus inputs and outputs used in this paper can be viewed and/or downloaded from the Examples page.
 Andrew Y Choi posted on Wednesday, March 01, 2017 - 3:12 pm
Hi Bengt & Linda,

Thanks for these prior responses. I have another question.

When simultaneously modeling covariates and distal outcomes in a latent class model using the BCH method, is there a way to quantify the additional variance in distal outcomes explained by C, beyond what is explained by D on X?

Thanks in advance.
 Bengt O. Muthen posted on Wednesday, March 01, 2017 - 6:05 pm
C plays the same role as a dummy variable (or dummy variables for more than 2 classes) in regular linear regression. The regression slopes for the dummy variables are obtained from the difference in the distal means for the different classes. Thereby it should be possible to compute an R-square. But that seems like a lot of work when you already see that C makes a significant difference.
 Tim Manfred Götz posted on Tuesday, July 16, 2019 - 12:31 am
Dear Bengt

I use a LCA(3-Step-method) w. an arbitrary secondary model - I have 5 classes and want to estimate a regression model. I want the effect of a manifest exog. predictor on a latent var. I assume a quadr/cubic effect. Up to now I did not specify centering of my predictor - I assume that I need to group mean center my predictor within classes prior to defining quadratic/ cubic transforms. Is there a way to group mean center with mixture? Or is grand mean ok? Default in mixture?

X = a10b;
X1 = X/1000;
X2 = (X*X)/1000;
X3 = (X*X*X)/1000;

Type = mixture random;
Starts = 0 ;

INTd | P WITH Xdum; !Dummy interaction
INT1 | P WITH X1; !linear interaction
INT2 | P WITH X2; !quadratic interaction
INT3 | P WITH X3; !cubic interaction

P BY P1@0 P2 P3; [INTER@0];
L BY L1@0 L2 L3; [L@0];
L ON X1 X2 X3 P INTd INT1 INT2 INT3 Covariates;


Thanks for any advice :-) TG
 Tim Manfred Götz posted on Tuesday, July 16, 2019 - 12:54 am
@Correction above: [INTER@0] -> must be [P@0]

One additional question: If group mean centering is the way to go for biasing the estimations the least, would it be an appropriate way to do the group mean centering manually within classes by using the Clustering option, save the data and then generate the quadr/cubic terms?

I just believe that this is not an adequate way to do, as this does not incorporate the "class-missspecification" parameters.

Thanks again!
 Bengt O. Muthen posted on Tuesday, July 16, 2019 - 6:29 pm
I would subtract the mean of X before you create X1, X2, X3.
Back to top
Add Your Message Here
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Options: Enable HTML code in message
Automatically activate URLs in message