Comparison of Latent Class Structures PreviousNext
Mplus Discussion > Categorical Data Modeling >
 Anonymous posted on Tuesday, May 11, 2004 - 10:42 am
I have run my LCA analyses separately by gender and have found some differences in the conditional probabilities and class sizes. Is there a way I can test to see if the differences in the measurement models are statistically significant? They are all in the same ballpark in terms of probability size, so I wanted to see if there's a way to test this.

Thank you.
 Linda K. Muthen posted on Tuesday, May 11, 2004 - 10:46 am
You could do a multiple group LCA using the KNOWNCLASS option of the VARIABLE command. Then you could do difference testing as in regular multiple group analysis to test your hypotheses of invariance.
 S. Oesterle posted on Wednesday, July 21, 2004 - 9:45 am
I estimated a multiple group LCA with the KNOWNCLASS option. Since Mplus gives me a Pearson and a Likelihood Ratio Chi-Square, I am wondering which one to use to do the difference testing.
 Linda K. Muthen posted on Wednesday, July 21, 2004 - 12:04 pm
I would use the H0 loglikelihood for difference testing. -2 times the loglikelihood difference is distributed as chi-square.
 Heather Orton posted on Thursday, February 23, 2006 - 7:56 am
I am working on an LCA and need to determine if I need to run the LCA by gender, or if I can combine the genders. When I run the LCA for each gender I found 3 classes for each model, and the classes look fairly comparable. To "officially" test that I can combine the genders and run one LCA, I ran the LCA with 3 classes and included the gender variable in the KNOWNCLASS statement (H0 loglikelihood=-13622.577). I then ran the LCA with 3 classes without the KNOWNLCLASS statement (H0 loglikelihood=-11724.179). I want to be sure I am calculating the likelihood ratio test correctly. I took -2*(the difference between the loglikehood H0 for each model) and get 3796.8 for my chi-square value. My confusion is with the degrees of freedom.

1. Do I subtract the DFs reported with the Likelihood ratio chi-square for each model? If so, I get DF=2010, and the p-value for the likelihood ratio test is 0.0.

2. If I've done this correctly, does the test indicate that I should run the LCA separately for each gender becasue the LCA structure is significantly different for each gender?

Thanks for you help!
 Linda K. Muthen posted on Thursday, February 23, 2006 - 10:44 am
You would use KNOWNCLASS in both analyses. In one analysis the parameters would not be held equal across classes and in the other anslysis the parameters would be held equal across classes.
 Heather Orton posted on Monday, February 27, 2006 - 1:12 pm
Linda, thank you for your reply. I'm very new to Mplus and am struggling with the code that would restrict the parameters across classes. Would it be possible for you to guide me through the MODEL statements that would do this? Thanks.

 Linda K. Muthen posted on Monday, February 27, 2006 - 1:38 pm
I would suggest that you read the section in Chapter 13 on multiple group analysis and the section on equalities in Chapter 16. The same rules apply to classes. Basically, if you specify:

y ON x (1);

this will hold the regression coefficient equal across classes. The same applies to all parameters.
 Heather Orton posted on Monday, February 27, 2006 - 3:22 pm
Thank you again for your quick response. I have read both chapters and consulted with a colleague who attended the Mplus training in Alexandria. She helped me write me some code that she thinks is assigning the appropriate constraints. However, it does not seem to me to be doing what you described. I would greatly appreciate it if you could review this code and let me know if it is correct. I will be sending my program and data set to the support e-mail address. Thank you in advance for your help!

 Sebastian Weirich posted on Wednesday, May 28, 2008 - 5:32 am
Hello. I try to specify a model with several classes by the KNOWNCLASS option. If there are only two classes (e.g. "gender"), I can estimate the probability of one class by estimating the mean of this class in the whole group (%OVERALL%) and inserting it into (1/(1+exp(-mean))). If there are more than two classes, this method fails. Is there a possibility to get the probabilities of the (more than two) classes via their latent means?

Thank you.
 Linda K. Muthen posted on Wednesday, May 28, 2008 - 1:37 pm
There is an example in Chapter 13 -- Calculating Probabilities From Logistic Regression Coefficients. See the example where all covariates are at the value of zero.
 Miguel Villodas posted on Thursday, December 17, 2009 - 7:08 am
Hello Linda,

I am building an LTA model with three latent classes at the first two time points and four classes at the third time point. I would like to establish partial measurement invariance (i.e., full invariance between the first two time points as well as two of the classes from the third time point). The way that it appears from the baseline models is that the third class splits into two classes at the third time point.

Is it possible to test this simultaneously or does it require sequential testing of the full invariance across the two time points followed by partial invariance of only two of the classes across all three time points?

Would the KNOWNCLASS statement be used for both of these analyses?

Also, the code for such analyses is becoming increasingly complicated. For the three time points, is it necessary to specify separate models for each time point?

Thank you very much in advance for you help with these issues.
 Linda K. Muthen posted on Thursday, December 17, 2009 - 12:48 pm
The KNOWNCLASS option is not used to compare across time. For the two classes that do not change, you can compare across time using equality constraints. Examples of these equalities can be see in Examples 8.13 and 8.14. This requires a model without equalities being compared to a model with equalities.
 db40 posted on Thursday, August 06, 2015 - 5:56 am

I have ran a multi-group analysis using the KNOWNCLASS command and have settled upon a 4 class solution, grouping the model by gender.

The Classes upon inspection appear to be qualitatively and quantitatively different. Are there any examples or codes in the manual which can help me learn how to test if the models are statistically significant from each other.

I might add the gender split is M=43% /F=56% if that makes any difference.
 Bengt O. Muthen posted on Thursday, August 06, 2015 - 2:55 pm
You can label the outcome means/thresholds in each class and then express differences across classes in Model Constraint to get tests.
 db40 posted on Friday, August 07, 2015 - 5:04 am
Sorry Bengt I'm not familiar with this. Is this in the manual anywhere?
 Linda K. Muthen posted on Friday, August 07, 2015 - 6:34 am
See MODEL CONSTRAINT in the user's guide.
 Sunil Sahadev posted on Monday, October 29, 2018 - 2:28 pm
Hi I am considering a mixture model with binary dependent variable and six binary independent variable. Is it possible to do this in mplus. Also is there any special example in the mplus guide that you could point me to so that I can write the code. I am new to MPlus
 Bengt O. Muthen posted on Monday, October 29, 2018 - 4:22 pm
See the User's Guide on our website, ex 7.12.
 Sunil Sahadev posted on Tuesday, October 30, 2018 - 6:04 am
Thanks a lot for the prompt response. My only doubt with example 7.12 is that the latent variable 'c' is the dependent variable in this example. While the independent variable is 'x'. The equation is c on x . can we also have a x on c equation. That is the latent variable is the independent variable?
 Bengt O. Muthen posted on Tuesday, October 30, 2018 - 5:12 pm
You don't say y on c. Instead, the relationship c--> y is captured by the y means/intercepts varying across the c classes just as if c was a set of dummy variable predictors.
 Sunil Sahadev posted on Wednesday, October 31, 2018 - 1:27 am
ok thanks a lot. I will try this out
 Betty Lin posted on Tuesday, February 05, 2019 - 9:47 am
I've run an LPA using standardized raw and change scores and would like to test mean differences in unstandardized raw scores across profiles. Can I use the BCH method to do this? Thank you!
 Bengt O. Muthen posted on Tuesday, February 05, 2019 - 5:11 pm
BCH can be used together with LPA if the distal outcomes are continuous.
 BL posted on Wednesday, February 06, 2019 - 10:22 am
Thanks so much, Dr. Muthen. Just to be clear, can BCH can also be used to probe for mean differences in continuous concurrent variables? The indicators used in the LPA are derived from the unstandardized raw scores of interest.
 Bengt O. Muthen posted on Wednesday, February 06, 2019 - 5:17 pm
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