Hello, I am running an LCGA using a count outcome (number of days worked in the past 3 months) measured at five timepoints. The data is heavily skewed with a large number of zeros at each timepoint, and many people (approx 40%) who have zero outcomes at all timepoints. I have been running an LCGA model with a fixed zero class, but am wondering if I should be using a ZIP model? Should I be using: 1. a zip model without specifying a zero class 2. a zip model with specifying a zero class 3. a regular lcga (not zip) with a zero class
Also, I would like to explore the possibility of using a gmm instead of lcga to allow within-class variation. Can I compare the fit of the gmm model to the lcga using the bic and loglikelihood? If so, do I have to compare the models with the same number of classes (in other words compare the 3-class lcga to the 3-class gmm)?
I would explore both 1. and 2. and compare via LL and BIC. In this regard, you may want to take a look at
Kreuter, F. & Muthen, B. (2007). Longitudinal modeling of population heterogeneity: Methodological challenges to the analysis of empirically derived criminal trajectory profiles. Forthcoming in Hancock, G. R., & Samuelsen, K. M. (Eds.). (2007). Advances in latent variable mixture models. Charlotte, NC: Information Age Publishing, Inc.
on our web site.
Yes, you can compare these models via LL and BIC, also under different number of classes. And, using Tech10 as you mention in your other message.
Thank you for your response. I compared the zip model with and without the zero class, and based on the LL and BIC, the zip model without the zero class is a better fit. However, the classification quality is much better with the zero class. Entropy is higher (.84 vs. .57) and when I look at each individual's probability of being in each class, they are much more clearly classified when there is a zero class. In the model without the zero class, many people have .33 probability in being in each of the three classes. Do you know why this would be the case? Does this indicate that, despite what the fit statistics suggest, the model with the zero class is actually a better model?
The added zero class often does not improve BIC. A low class is sufficient.
Although tempting, I would not choose between models based on entropy. Entropy is like r-square in SEM - it can be good for poorly fitting models and poor for well-fitting models. I don't know why you have poor classification in your 3-class model. Have you established that you need 3 classes rather than 2, and that you need 2 classes rather than 1?
Also, when you get Version 5, looking at the fit statistics in Tech10 for say the 10 most frequent response patterns can make choices between models easier.