Hi - I am attempting to use the BCH approach to carry out an arbitrary secondary model within classes as described in Webnote 21 part 3. The model seems to be running, but I notice that the class proportions are fairly different from the first step model that generated the BCH weights. Is this cause for concern?
Iím sorry I also forgot to add: In the secondary model, Iím wanting to specify a random effect on the overall model. I used type = RANDOM MIXTURE, and it appeared to work with the BCH weights specified in the TRAINING command - but I want to make sure it was ok to use them with the random effect.
1) I don't think it is of concern. What you see in the second stage is simply the sum of the BCH weights. These are determined in the first step, i.e., it is not as if somehow the class allocation changed in the second step estimation. These are entirely determined in the first step and shouldn't really be interpreted as class proportions. Only the first step class proportions should be interpreted as such.