Message/Author 


Will like to know if you have an example in your achieves on how to test the invariance of class proportions, and class probabilities in LCA. Trying to test the gender invariance of class proportions and probabilities but cannot find a clear example to give me. Will appreciate if a clear example in your achieves is provided or any other link to help me. 


When you have c ON cg in the model the class probabilities of c are free. When you do not have c ON cg in the model, the class probabilities of c are constrained to be equal across the cg classes. You can compare these two models.



hi are the following syntax commutative. CLASSES = cg (2) c (2); CLASSES = c (2) cg (2); and which of them apply to your analogy above. 


You need to use CLASSES = cg (2) c (2); 


Dear Drs Muthen, I want to test the invariance across gender of my 3 class LPA model. I'm comparing the unconstrained model to a model having the means of my indicators constrained to equality in men and women using 2*LogLik difference (I also taking into account the scaling correction factors obtained with the MLR estimator, au you descibed in the web site). My questions: 1) This test is significant (suggesting noninvariance), but the BIC decreases (suggesting that the constrained model is better). Which I should trust the most? 2) Is the 2*Loglik difference test affected by sample size? (like DIFFTEST or Chisquare in measurement invariance). In my case I have 500 subjects in a dataset and 830 in a second one (I want to keep the 2 dataset separate). 3) Can I test partial invariance releasing some indicators from the equality constraint? Are there any statistical indices to "guide" this process (such as the modification indices in measurement invariance)? Thanks in advance for your help! 


1) I would go by the testing when that can be done (as here) more than BIC. The only disadvantage with such testing for mixtures is that the class formation may change noticeably. 2) To some extent, but these n's aren't that large. 3) Q1: Yes. Q2: Modindices don't work that well with mixtures. Not sure what is a good approach. 


Thanks for your answer. Do you know any reference I can use to support the choice of using testing instead of BIC? or, in general, any study that discuss this subject... Thank you again. 


No, I can't think of any. Others? 


I only found this paper for the moment (and it is about LCA, not LPA): Finch, H. (2015) A Comparison of Statistics for Assessing Model Invariance in Latent Class Analysis. Open Journal of Statistics, 5, 191210. http://dx.doi.org/10.4236/ojs.2015.53022 


Good. 

Back to top 