Lns posted on Tuesday, September 08, 2015 - 1:31 pm
I have three-level data for which I initially ran a single-level LCA, which gave me a Vuong's p = 0.0013. However, when I add clustering and run as complex, Vuong's test changes to p = 0.5910. How should I interpret this discrepancy?
I wouldn't call this a discrepancy. The data are not the same when you take the clustering into account versus when you do not. You should expect the results to differ.
Lns posted on Tuesday, September 08, 2015 - 7:52 pm
Although the Vuong's changes when the data is clustered, the BIC does not...thus, when clustered, Vuong's indicates that I should accept a 3 class model, but BIC indicates 4 class model. Which should I take into account?
Just to add to Linda's answer that when you use a more elaborate model the statistical power to detect an additional class decreases because the model is less parsimonious. So it is not unexpected that Vuong's test changes. Also you are adding additional covariates essentially (the random effects) - that can easily change the number of classes. Finally - if adding multilevel does not improve BIC - I would worry about the need for multilevel modeling more rather than the number of classes.
Lns posted on Wednesday, September 09, 2015 - 11:15 am
Thank you for your response! Can you explain a little more what you mean by "would worry about the need for multilevel modeling more rather than the number of classes." Why would I expect the BIC to change when clustering and what does it not changing suggest?
If adding multilevel effects does not improve BIC then you don't have cluster effects and you should not include cluster effects in the model.
Lns posted on Wednesday, September 09, 2015 - 12:54 pm
Thank you. One more question. I want to use my latent categories to predict y and also moderation. I am very confused on the literature I've found (auxiliary vs distal variables? I do not understand these terms) on how this works in the syntax. My entropy is low (.67) so I do not want to use hard class codes as predictors. What would the most basic syntax look like to a) directly predict LCA class -> y and b) moderate LCA class -> M -> y, and for these, how do I add control variables?