Thanks for the Mplus V7.1 which has a lot of new and useful features. I have a question related to covariate testing in mixture model (factor mixture model).
Originally I used DU3STEP/DE3STEP to analyze the covariates. However, the new V7.1 gave warning messages for a lot of my auxiliary variables, namely, that the classification error between Step 1 and Step 3 exceeded 20%. Apparently there were no results for those variables.
I understand that this check is a new feature for V7.1. In this case, should I use the new DCON/DCAT for distal outcomes? This new approach did not give this warning message (except for listwise deletion).
It is relieving that I can go for the DCON/DCAT auxiliary approach in analyzing the distal outcomes, because I do not know how to deal with the many error warnings associated with the DU3STEP/DE3STEP approach.
Thanks for your help. I have a follow up question regarding the dcon/dcat. I am interested in using this analysis, but have Mplus7, not Mplus7.1. Is there a manual code that I can use to run the same analysis?
Laina Isler posted on Wednesday, June 12, 2013 - 2:38 pm
Thanks, I will look into updating it.
I am getting an entropy of 0.611, which I understand is not strong at all, but my average latent class probabilities for most likely class membership range from 0.815 to 0.74. And the classes obtain good chi-square values in the 3 step and 1 step equality tests.
I am not sure whether I should interpret the data as demonstrating distinctive classes, or whether the entropy is too low. I have tried changing number of classes, and this does not improve entropy. Is there another way to improve entropy?
Laina Isler posted on Wednesday, June 12, 2013 - 3:12 pm
Also - I am dealing with data sets of 4500 to 6900. I am not certain whether entropy may be negatively effected by class size.
An entropy of 0.6 is ok. You should not decide on the number of classes based primarily on entropy - use BIC instead. Sample size does not have to do with entropy.
Laina Isler posted on Wednesday, June 12, 2013 - 4:15 pm
Thanks for the quick reply. The BIC are very large, so I am not certain how much of a difference is necessary to support one number of classes over another. Furthermore, the BIC keeps decreasing when I add classes long after further tests (Vuong-lo-mendell-runin likelihood ratio test) indicate that fewer classes would be sufficiant.
For instance, a BIC of 322937.566 vs. 322838.965 vs. 322769.701 were obtained for 4, 5, 6 classes, respectively. While the tech11 output (lo-menell-ruban adjusted LRT) is 0.008, 0.0102, and 0.0922, respectively.
Does this indicate the a 5 class solution is attained, despite BIC values continuing to decrease, (up to 9 classes continue to show a decrease in BIC)
Yes, Tech11 gives support to 5 classes, but given that BIC continues to decline, the conclusion isn't clear. I would recommend what we teach in Topic 5, namely to look at how different the solutions are for 5 and higher number classes. Look at the profile plots - perhaps 6 and more classes are merely uninteresting variations on 5-class themes.
Thanks for your help. The 5 class seems the most reasonable. However, I am getting altered models (class size, average latent class probabilities and plot) depending on whether I run the data with syntax to test covariates (du3step), or likelihood ratio tests (tech11 and tech14).
I thought these tests were not supposed to effect the classes themselves, and was wondering which output would be more accurate, or whether there is a way to run the tests wihtout impactin class results?
Ali posted on Tuesday, February 23, 2016 - 8:03 am
Hello, I fit covaraites(country,and gender) into LCA separately, and each model selects 3 classes as the same as without covariate. I check if the log odds is significant on covaraites.As for the country, it has 6 categories, and I coded them as five dummy variables.From the results, in each class, four out of five are significant. As for the gender covariate, there is only one dummy variable and it is significant. Could I choose the model with country covariate as the final model ? Because it has lowest AIC,BIC and A_BIC compared with LCA without covariate and LCA with gender covariate