Estimation of Continuous Regression w... PreviousNext
Mplus Discussion > Latent Variable Mixture Modeling >
 David Rein posted on Friday, April 26, 2002 - 9:09 am
What method does Mplus use to estimate continuous regression models within classes?

If it is maximum likelihood, is the distributional assumption one of normality? Gamma, whatever?

If it is quasi-liklihood, how is the link and the variance parameters defined?

Getting more to the point, for the model I am examining, the best estimates for a non-class model are derived from a quasi-likelihood approach in which the distribution is not specificed, the and the link and variance functions are specified based on a prior exploaration of the data.

Given that I find a link function = mean**.03 and variance function equal to mean**2.4, is there any way to specify this in Mplus to get model estimates?

One alternative I see is to use the FMM to find classes. Then use Mplus, or another program to estimate each regression seperately by class using GLM and the link and variance function.

Any ideas?
 bmuthen posted on Friday, April 26, 2002 - 4:33 pm
Mplus uses ML, assuming within each class conditional normality given covariates. This implies that the overall (mixture, not class-specific) distribution is very non-normal.

I don't know if the link fcn mean and the variance fcn values can be used somewhow.
 David Rein posted on Monday, April 29, 2002 - 8:54 am
Thanks Bengt, this is what I needed to know.
 David Rein posted on Friday, July 11, 2003 - 1:10 pm
Hi Bengt,

Is there a standard way to test if the coefficients on the same exogneous variable are statistically different between classes? For example to test the null that the difference between age on wealth is different between class one (the middle class) and class two (the wealthy) is equal to zero?

The quick and dirty way seems to be to calculate the confidence intervals for each variable and just see ifthey overlap, but I'm thinking there must be another way to do this.
 bmuthen posted on Friday, July 11, 2003 - 4:21 pm
Just do two runs, one where they are equal and one where they are not. Take 2 times the loglikelihood difference and test by chi-square using as df the difference in number of parameters.
 David Rein posted on Monday, July 14, 2003 - 8:55 am
Sounds like a good test, except that I have around a 40 variables, and 4 classes - If I understand you correctly, I'll need to run 40 sets of comparison models to identify statistical differences - and thats just between no class differences and any class differences, not mentioning differences between class 1 and 2, and class 1 and 3, 1 and 4, 2 and 3, 2 and 4, etc.

I'm following your logic, just wish there was an easier or automated way.
 bmuthen posted on Monday, July 14, 2003 - 3:36 pm
Can't think of an easy way.
 Anonymous posted on Thursday, July 17, 2003 - 1:13 pm
I'm attempting to estimate a SEM of the below form in Mplus:

X ----> L ----> Y
X ------------> Y.

Is it possible to specify that one of the X's is correlated with Y, but not the intervening latent class measurement model L ?

Whenever I attempt to exclude one of the X's from the first set of equations (i.e., where L is in intervening variable), Mplus returns the warning that the "excluded" X will be treated as a Y variable, and then produces error msgs for the regressions of the "excluded" X on Y.
 bmuthen posted on Thursday, July 17, 2003 - 1:30 pm
Yes, you can handle the model you show in the figure. The second line is given as "y on x" and this x need not be part of "l on x". If this doesn't clear it up, please send your input, output, and data to
 David Rein posted on Wednesday, August 06, 2003 - 10:03 am
Is there a way to specify a Poisson or NB distribution for a regression of count data within a mixture model?
 Linda K. Muthen posted on Wednesday, August 06, 2003 - 10:52 am
Not in the current version.
 Alexander Kapeller posted on Friday, December 12, 2008 - 12:23 am

I am configuring a latent regression model like below.

NAMES ARE ret105 ret205 ret106 ret206;
usevariable ret105 ret205 ret106 ret206;
missing = all (999);
CLASSES = c (2);


ret05 by ret105;
ret05 by ret205;
ret06 by ret106;
ret06 by ret206;

ret06 on ret05 ;

What is the basis on that the classes are derived:
a) the indicators ret105 ret205 ret106 ret206
b) the regression ret06 on ret05
c) both

I want to have a separation in classes derived from the regression coeffficent as a first separater and then describing different classes with help of the indicators. How can I do this?

 Bengt O. Muthen posted on Friday, December 12, 2008 - 6:38 am
You can look at Tech1 and see which parameters vary across classes - those are the model features that describe the class differences. The Mplus default is that means vary across classes, except for factor indicators in BY statements. So in your case the factor means vary across classes by default. In addition, your regression slope varies because you have specified that. If you don't want the factor means to vary across classes you say


in the overall part of the model.
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