SF Wang posted on Tuesday, October 08, 2013 - 12:26 pm
The standard 3-step approach is (1)GMM w/o covariates; (2) get class assignment; (3) further analysis, usually logistic regression. Actually to M+ users, steps (1) and (2) are done in one step - one GMM fit provides the class assignments as an output data.
Now standard 3-step approach is replaced by a new 3-step approach by adding "Auxiliary = (R3STEP) x;" statement.
But after reading the webnotes, I am still a little bit confused how many models in M+ we should fit. Please let me know which of the following two understandings is correct: (a)we need to get latent class membership from a GMM w/o covariates and get the error probabilities too, and then feed these into another GMM (with the auxiliary statement) and get the correct class assignment? (b)we only need to replace step (1) in standard 3-step approach by adding auxiliary statement. i.e. the new three steps are: (1)GMM with auxiliary statement; (2) get class assignment; (3) further analysis.
SF Wang posted on Monday, October 21, 2013 - 7:24 am
Thank you for answering my question!
Could you confirm whether my following understandings are correct?
My understanding #1: in the new 3-step approach, auxiliary statement won't change any model results, including the class assignments. It's just logistic regression for class assignment with corrections. Therefore, step(3)-further analysis(logistic regression) has already been done, and the results are in the M-plus output too. (step(2) is done in the same model). So, to users, it's like a one-step procedure.
My understanding #2: if I don't plan to do logistic regressions in step (3)(I am thinking of doing something else, e.g. CART), I wouldn't need the AUXILIARY statement.
Classes do not change only with R3STEP, DCON, and DCAT. They can change with the other settings. This is explained in Web Note 14.
You would not need the AUXILIARY statement in this case.
Natalie posted on Sunday, April 06, 2014 - 3:18 pm
If we want to predict the intercept and slope factor (in addition to the latent class variable C) in the GMM from a covariate, can we do so using the R3step manual or automatic procedure? Will the covariate influence class formation in this case?
Quick question. I'm using the R3step method to estimate predictors for class membership in my GMM. I am seeing some logits above 1 (e.g., 1.052) that are significant. These logits above 1 are offending estimates and, thus, invalidate the R3step results, correct?