Mingnan Liu posted on Thursday, November 14, 2013 - 8:39 am
Hi, I am new to Mplus and there is a not so fundamental model that I don't know how to set up. Here is the situation: I have 12 Likert scales (y1-y12), each with 5 categories (strongly agree--strongly disagree). I want to run a LCA and I believe the y1-y4 belong to latent class variable f1, y5-y8 belong to f2, and y9-y12 belong to f3. The theory also says that y1-y12 also belong to another latent variable f4. At the same time, y1-y12 belong to f5 as well. The differences are for f1-f4, the y1-y12 are treated as ordinal scale while for f5 y1-y12 are treated as nominal scale. f1-f3 are correlated with each other whereas f4 and f5 are not correlated with each other, nor with f1-f3. How should I set up the Analysis and Model? Any suggestions are appreciated!
Do you want to use factors as latent class indicators for the LCA?
Mingnan Liu posted on Thursday, November 14, 2013 - 9:07 am
Do you mean I should do CFA first before LCA? I do want to get the conditional probability of each response category category, particularly for f4 and f5. My understanding is that I will one for f4 since I treat the indicator as ordinal and five for f5 since they are nominal. One thing I forgot to mention is that I need to add a constrain so that the beta weights (or conditional probability) for f4 are the same across all 12 questions. Similarly, the constrain is needed for f5 so that conditional probabilities on the five categories are the same across 12 questions. Thank you.
It is not clear to me what you want to do. Are f1-f5 continuous latent variables, that is, factors or categorical latent variables, that is, latent class variables. Please note that the same variables cannot be treated as ordinal and nominal.
Mingnan Liu posted on Thursday, November 14, 2013 - 10:59 am
My apologizes for the unclear questions. To put it simpler, I am trying to replicate the findings from
Kieruj and Moors, 2013 "Response style behavior: question format dependent or personal style?" Quality and Quantity 47: 193-211
particularly the model they specified on p.200. As you can see, F_1i to F_3i are content relevant whereas A_i and E_i are content irrelevant. I am interested in A and E latent variables. I want to treat the indicator variables as nominal when estimating the latent variable E because, on p.201, the the beta weights are different for each response category. This is why there is a subscript c for the coefficients of E. The indicator variables are treated as nominal when estimating latent variables A and F1-F3. Therefore, only one beta weight for each question. Also, the beta weights for E and A are the same for all questions, as you can see from p.201.
I think this can be done in Mplus throught Model Constraint and using nominal link, but don't have the time to explore how right now.
lisa Car posted on Tuesday, June 27, 2017 - 8:24 am
In reading elsewhere in this discussion (not this particular thread) it was suggested to do a confirmatory LCA in the same dataset prior to doing so in a different dataset. I have read the article by Finch and Bronk as well as the UG and was able to run a confirmatory LCA using this guidance on the same dataset in which I ran an exploratory LCA. It has confirmed my findings. My question is whether there is a reference that says to use confirmatory LCA in this way for the same dataset?
I would not confirm on the same data set that I explored. Although I could use a confirmatory model on the same data set, I would not interpret that as a confirmation - a new data set is needed for that.
lisa Car posted on Wednesday, June 28, 2017 - 6:00 am
Thank you for the clarification. Could a confirmatory model on the same dataset be considered as internal validation of the model?