What is the best way to deal with missing data in latent class models?
For example, Latent GOLD can estimate models with missing data (LG technical manual page 51). Simply put, the likelihood contribution is based on the observed indicators only, which means that for an observation with missing data, the likelihood function is maximized with only the remaining data available.
OK-that's fine, but does this mean for the parameters? Am I going to get better classifications because the parameters were estimated with as much information as possible (i.e. if the observation say only has a missing value), or am I going to get worse classifications because of the missing value?