John Woo posted on Sunday, September 26, 2010 - 9:47 am
I have two categorical latent class variables c1 and c2. c1 is indicated by binary repeated measures y1-y4. c2 is indicated by binary repeated measures z1-z4. c1 has two categories. c2 has two categories. (I also have covariates predicting various elements of the model.)
When I run the model with "c1 ON c2", using Start=500 20, I get no replication of best loglikelihood.
When I run the model with "c1 WITH c2", I get asked to use "loglinear parametrization". When I do so, I get a result for the best loglikelihood from the "unperturbed" starting value, but this best LL value is so much better than the next best loglikelihood. And, obviously, no replication.
My questions are, (1) can I trust that this model has converged to the global maximum, even though the (best) LL from the unperturbed starting value is not replicated--but so much better than the next best ones (I included Start=500 20)?
(2) in terms of plotting c1, i want to plot c1 as TWO classes. But, I am given information for FOUR classes because of the joint information with c2 (i.e., 11, 12, 21, 22). For c1#1, can I weight average the 11 and 12 according to the sample weight given for the four classes?
1. It sounds like you allow too strong perturbation of starting values, which can happen with several latent class variables. Choose STSCALE=1 instead of the default 5.
2. I don't know what you nean by plotting c1. Perhaps you mean item probability profiles for c1 classes and you don't like that the profiles also change as a function of c2? If that's so, you can change the model using Model c1, Model c2 (see UG for examples).
John Woo posted on Monday, September 27, 2010 - 7:31 pm
You said above, "I don't know what you nean by plotting c1. Perhaps you mean item probability profiles for c1 classes and you don't like that the profiles also change as a function of c2? If that's so, you can change the model using Model c1, Model c2 (see UG for examples)."
Thank you. That was exactly what I meant.
I have one follow-up question. I am planning to use LCGA to identify the trajectories instead of my original plan for using repeated LCA.
Yet, I still want to have two trajectories for c1 and two trajectories for c2. But because I am doing c1 WITH c2, I will get four trajectories (i.e., four sets of growth factors).
Is there a way to hold constant the growth factors for c1#1 across the categories of c2, and hold constant the growth factors for c1#2 across the categories of c2, etc?
If so, could you tell me the commands or direct me to User Guide examples?