I'm working with a growth mixture model and I'm comparing a 2 and 3 class model. Each class represents a different curve shape (linear, log, sinusoidal).
The 3 class model includes both curves from the 2-class (lin and log) plus the sinusoidal curve as the third class.
There are LRT results available in the tech11 output. If classes were created without a predefined curve, these results would be readily interpretable for model selection, but since I have predefined curves for each class, what exactly is used as the k-1 mdoel in the LRTs? Is it actually the same model with the last class removed (so just lin and log), or is it something else?
It is actually with the first class removed and as you noted indeed this is important in this situation and you may need to reorder the classes to obtained the test you want. BIC comparison could be used as well.