GMM with distal outcome
Message/Author
 Koko Adal posted on Friday, August 06, 2010 - 9:18 am
We are working on Growth Mixture Model. We have continuous indicators (repeated measurement), covariates and a categorical distal outcome variable (with three sub-groups).

Hypothetically, we have two types of covariates: those affect the continuous indicators through the latent classes and our distal outcome (group-1); and those have an effect only on our distal outcome (group-2).

Here is what we did:

1. i s| wt1@0 wt2@1 wt3@2 wt4@3 wt5@4 wt6@5
2. C#1-C#2 ON group-1
3. Distal outcome ON group-1 group-2

The model is identified. And we have odds ratio of our distal outcome on the covariates. But we wanted to know the individual odds ratio for the sub-groups of our distal outcome.

1. Is it possible to disentangle the crude odds ratio into sub-groups?

2. Can we do multinomial logistic regression of our distal outcome on the covariates?

Thank you.
 Koko Adal posted on Friday, August 06, 2010 - 9:47 am

We are working on Growth Mixture Model. We have continuous indicators (repeated measurement), covariates and a categorical distal outcome variable (with three categories).

Hypothetically, we have two types of covariates: those affect the continuous indicators ( i.e. through the latent classes) and our distal outcome--group-1; and those have an effect only on our distal outcome--group-2.

Here is what we did:

1. i s| wt1@0 wt2@1 wt3@2 wt4@3 wt5@4 wt6@5
2. C#1-C#2 ON group-1
3. Distal outcome ON group-1 group-2

The model is identified. And we have odds ratio of our distal outcome on the covariates. But we wanted to know the individual odds ratio for the sub-groups of our distal outcome.

1. Is it possible to disentangle the crude odds ratio into sub-groups?

2. Can we do multinomial logistic regression of our distal outcome on the covariates?

Thank you.
 Linda K. Muthen posted on Friday, August 06, 2010 - 9:54 am
1. If by sub-groups you mean classes, then you can mention the ON statements in the class-specific parts of the MODEL command to free these parameters. They are held equal across classes as the default.

2. If the distal outcome is nominal and is place on the NOMINAL list, the regression is a multinomial logistic regression.
 Koko Adal posted on Friday, August 06, 2010 - 11:02 am
Thank you.

One more question about the distal outcome:

Our distal outcome has three values( 0,1 and 2). We tried to change the 'CATEGORICAL' into 'NOMINAL'. But, every thing changed. The model became unidentifieable and we have got stars(******) for some of our coefficients. We no more have also Thresholds in the outcome.

Are we on the wrong direction?
 Koko Adal posted on Friday, August 06, 2010 - 1:40 pm
Sorry, we have figured out our problem. We were supposed to regress the latent classes on the covariates as indicated in Topic-6 of the handouts and Video.

We have a question about fixing of the factor loadings. On Slide no. 152 of Topic-6, "Input Excerpts NLSY Growth Mixture Model With Covariates And A Distal Outcome", we can see that you have fixed the factor loadings.

2. Why is it needed and how do we do that in our model? we have simply fixed our factor loadings (0-5)

Thank you.
 Linda K. Muthen posted on Sunday, August 08, 2010 - 11:16 am
1. These are explained on Slide 153.

2. Use time scores of 0-5.
 Michelle Maier posted on Monday, September 13, 2010 - 7:41 am
I am working on a growth mixture model where I have three continuous distal outcomes (two observed and one latent). I have been able to test for mean differences across the classes in the observed variables by using the auxiliary e option but not in the latent variable (error says unknown variable in a BY statement). Is it possible to use this option to test for mean differences in the latent variable, as well?
 Linda K. Muthen posted on Monday, September 13, 2010 - 8:52 am
The AUXILIARY (e) option is an approximate screening tool. To test the equality of means of your distal outcomes across classes, you should use either MODEL TEST or loglikelihood difference testing. Conditional independence is assumed when you have more than one distal outcome. You should use WITH statements to covary the distal outcomes if needed.
 Michelle Maier posted on Thursday, September 16, 2010 - 1:03 pm
Thank you for your quick response. I actually did try using MODEL TEST first. However, I did not get a wald statistic in my output. I am using the syntax below to test for one outcome at a time (all three outcomes are in my use variables list). Is there another way to define m1 and m2 when you have a two class model?

%C#1%
[i s q];
[Sci] (m1);

%C#2%
[i s q];
[Sci] (m2);

MODEL TEST:
0=m1-m2;

Thank you!
 Linda K. Muthen posted on Thursday, September 16, 2010 - 1:42 pm
 Rebecca Shearer posted on Friday, October 22, 2010 - 10:54 am
Hello, I have been reading this very helpful post. However, I am trying to use the Model Test command to examine whether there are mean differences among a distal outcome for a 4-class latent profile analysis. Is there a way to use the model test for the distal outcome without including the distal outcome in the estimation of the classes? When I include the outcome in the usevariable statement then it appears to be included in the estimation of the classes (along with the other variables I am interested in). I tried to the use the auxiliary command, and then I am able to get mean differences across the classes cleanly. However, is this acceptable or do I also need to follow up with the Wald test? Thank you.
 Linda K. Muthen posted on Sunday, October 24, 2010 - 9:11 am
As is shown in the paper by Clark using AUXILIARY (e) performs well only if entropy is at least .8. See the following paper on the website for a fuller discussion:

Clark, S. & Muthén, B. (2009). Relating latent class analysis results to variables not included in the analysis.
 Rebecca Shearer posted on Tuesday, October 26, 2010 - 7:42 am
thank you!
 Rebecca Shearer posted on Sunday, November 28, 2010 - 8:46 pm
Hello again, to follow up with my previous post re: 4 class latent profile analysis. We have some missing data on our distal outcomes and are using the auxiliary command to examine mean differences among some of these outcomes. Is there a way to determine from the output the sample size that Mplus uses to estimate these mean differences? Thank you, Rebecca
 Linda K. Muthen posted on Monday, November 29, 2010 - 10:43 am
All available data for each variable is used. We don't give the sample size for each variable as part of the output.
 Rebecca Shearer posted on Friday, December 03, 2010 - 5:52 pm
thank you very much for your help- to clarify, does this mean that FIML is used for the distal outcomes if the auxiliary command is used? I am still a little confused
 Linda K. Muthen posted on Friday, December 03, 2010 - 6:13 pm
No, it means that for the AUXILIARY command, if y1 has 100 people without missing, 100 are used. If y2 has 110 without missing, 110 are used.