C. Sullivan posted on Tuesday, September 04, 2007 - 12:33 pm
I'm working with a data set that has several measures representing different types of criminal offenses.
I'm trying to get a sense of underlying offense patterns using LCA, but have a lot of cases with no offenses at all. Is it possible to fix values in the MPlus input so that those cases are captured in a single class? This strikes me as similar to the setting of the "zero class" in the Kreuter and Muthen paper, but I'm having trouble operationalizing it in my LCA models.
Yes, you take the zero class approach. If you have categorical outcomes, you fix the thresholds to +15 for all items in the zero class.
C. Sullivan posted on Wednesday, September 05, 2007 - 1:52 pm
I used that approach and it appears that no cases are being placed in that class (based on the most likely class membership assignment). That seems counterintuitive since the 0 0 0 0 0 0 0 response pattern is so prevalent in the data.
Lucy Barnard posted on Wednesday, September 05, 2007 - 10:11 pm
Is there any recommendation when to use "zero classes"? I'm analyzing longitudinal cigarette smoking data of preadolescents and I'm already using two-part (mixture) modeling. 6% of my sample has structural zeros over time, but probably there are more "structurals" since some have many zeros and one or two missings over time. Is a "low" or "normative" class sufficient in this case? Can one compare the BIC between solutions with zero classes and without, just in order to decide on that issue, or is there an alternative? thank's in advance!
It's a good question. We debated this issue in our two Kreuter-Muthen papers on crime curves (see Mplus web site under Growth Mixture Modeling). One could argue that a zero class is substantively warranted and may have its own relationship to covariates. One the other hand, in practice it seems that the zero class is often indistinguishable from a very low class. BIC seems like a reasonable gauge here.
Rosie Green posted on Tuesday, February 15, 2011 - 9:34 am
I am having a similar problem to that posted by C Sullivan above. When I try to create a 0 class by fixing all thresholds to +15 no cases are placed in that class based on most likely class membership assignment.
I have found in some applications a fairly prevalent class with very few items endorsed. The items endorsed are commonly endorsed, representing for instance mild forms of misbehavior. So in your case, this "almost zero" class may absorbe the subjects who have all zeros. This then indicates that the all zero subjects are no different from the almost all zero subjects.
Rosie Green posted on Wednesday, February 16, 2011 - 1:27 am
Thanks, this does indeed seem to be the problem in my data.
Rosie Green posted on Wednesday, February 16, 2011 - 9:58 am
Having looked into this slightly further, it appears that ideally we do want to separate our all zero subjects from our almost-all zero subjects if possible.
This is because all the forms of disadvantage identified in our variables are relatively severe and none are particularly commonly endorsed. However, the solution with one class constrained to zero is still creating one (very small) zero class and another (much larger) almost-zero class, which is making interpretation of our models difficult.
Could you advise on any possible solutions to this problem?
Perhaps your model does not fit the data well. If your variables are categorical you can look at the bivariate tables of Tech10. You can also try a slightly different model, namely a factor mixture model. There are several papers on our web site using this type of model. See for instance,
Muthén, B. (2008). Latent variable hybrids: Overview of old and new models. In Hancock, G. R., & Samuelsen, K. M. (Eds.), Advances in latent variable mixture models, pp. 1-24. Charlotte, NC: Information Age Publishing, Inc.