Johnny Wu posted on Thursday, April 19, 2007 - 2:19 pm
Hi Dr Muthen,
I am looking at drinking trajectories across 15 months. My observed indicators are for month 3, 6, 9, 12, and 15.
A 3-class GMM was the best solution. As people often find, one of the classes is a "Heavy drinking" class.
However, this class contains significantly more missing data than the other classes, especially at the later time points. Substantively it makes sense (the more you are drinking the less likely you are going to show up for the study).
1. Have you encountered this situation before? Is it a major problem?
2. Could you refer me to some articles that address this issue, that I could cite perhaps?
It sounds like you have nonignorable missingness in your data. Although common, this is difficult to deal with. One possibility is to bring covariates into the model that predict this missingness. You can search the literature for nonignorable missing data to see if there are articles that discuss this. I am not aware of any that are accessible.