rgrove posted on Thursday, December 13, 2012 - 3:54 am
I'm running a mixture model with 2 factors and varying classes. I noticed that the default settings for this are to restrict the factor structure to be equal across classes. Is this how it is meant to be done or is it meaningful to determine whether the structure differs across classes?
The other issue I have is that when this restriction is relaxed there are huge discrepancies between the numbers in each class compared with those when the structure is restricted. Is this due to the number of parameters?
The default is meant as a baseline model that is parsimonious and converges easily. The need for relaxing the model should be explored as you mention. It is to be expected that a more relaxed model can change the class percentages. You would go by BIC to determine which model is best.