|
|
LCA with known classes using trainin... |
|
Message/Author |
|
|
Unusual as it may sound I’m trying to run a LCA knowing exactly both the number of latent classes and the class membership of each individual. The purpose of this already "solved" LCA is to compare the model fit of alternative latent classes restricting them to a set of different observed classes using the same data. Model 1, for example, groups the data into 2 known classes ("high" "low"), whereas Model 2 groups individuals into 4 known classes ("high" "middle" "middle-low" "low"). How to specify such a LCA using Mplus? I’ve already figured out a solution but I’m not sure if it’s right und just wanted to hear your expert opinion about it (I didn’t found any comparable example on the manual). I did the following: using the option "training" with (MEMBERSHIP) I defined, for each model, dummy variables (t1 t2 and t1,t2, t3,t4 respectively) and ran a standard LCA with the same categorical indicators in each case. By doing so, I suppose, is the latent class membership restricted to the known classes captured by the dummy-coded t-variables defined in "training". Is that right? Below is a shortened syntax for Model 1. Thanks for your help! (...) Variable: CATEGORICAL is var1 var2; CLASSES = c (2); TRAINING = t1 t2 (MEMBERSHIP); Analysis: Type=mixture; (...) |
|
|
The training data need to have (with 2 classes as an example) t1=1 t2=0 for a person to be a known member of class 1 and t1=0 t2=1 for a person to be a known member of class 2. The two classes would then have to have distinguishable features in the MODEL command in terms of parameter restrictions so that the classes have different meanings. I am not sure, however, if this is what you are asking. |
|
|
Thanks for your answer. What I'm trying to do is to fit a LCA knowing a priori both the number of classes and the class membership of every individual. The purpose of such an analysis is just to be able to compare the fit of different models with different observed classes. The model I wanted to run is the one explained in Example 7.3, however I'd like to restrict the class membership of every individual to match known classes in the data. These known classes should identify exactly the latent classes in the model. A very simple task indeed. I thought this could be easily solved by introducing, in the case of 2 classes, "TRAINING = t1 t2 (MEMBERSHIP)" in the VARIABLE command of Example 7.3. Like in this example, the MODEL command would be empty. In short: I'm fitting Example 7.3 restrictig latent classes to perfectly identify known classes in the data.....is "Training" the right option to introduce this restriction or is there any other way to fit such a model? Thanks again for your help. |
|
ywang posted on Tuesday, August 18, 2009 - 7:10 am
|
|
|
Dear Drs. Muthen: I have a follow-up question. Can the "multiple-group LCA analysis" be used instead of "LCA with training variables" to fit "the model with known classes"? What is the difference between 'multiple-group LCA" and 'LCA with training variables" in this particular situation as Andres described? Thank you very much! |
|
|
Training data can accomplish the same as Knownclass (multiple-group analysis). Neither is available when you have several latent class variables, but then you can use an observed indicator for known class membership (see second approach to CACE modeling in the UG). |
|
Back to top |
|
|