Merging latent classes PreviousNext
Mplus Discussion > Latent Variable Mixture Modeling >
Message/Author
 Anonymous posted on Thursday, July 12, 2012 - 3:56 am
I am hoping someone can advise me how to merge latent classes in Mplus. For example, I have a situation with 4 classes, with probabilities of class membership as 40%, 30%, 25%, and 5%. For the purposes of estimating predictors of class membership, I wish to amalgamate the third and fourth classes so that the first two classes are unaffected. What syntax would I use in Mplus? Would I use the training option and if so would the membership or probabilities sub-option be most appropriate?
 Linda K. Muthen posted on Thursday, July 12, 2012 - 9:55 am
I would not recommend merging classes.
 Anonymous posted on Thursday, July 12, 2012 - 1:36 pm
Thanks for that response. My reason for asking was I have seen some multiple group latent class analyses in the literature where, when a model was fitted and some classes in certain groups had 0 (or very close to 0) probability of class membership, the model was refitted forcing membership in these classes to be 0, on the argument this improves model fit. I was interested in learning how to do this, or your opinion on this approach. Thank you.
 Linda K. Muthen posted on Friday, July 13, 2012 - 10:36 am
If you want one less class, run the model with one less class. If the classes don't combine the way you would have forced, it may not be wise to force the two classes together. That does not reflect what the data says.
 Jody Mdoda posted on Monday, August 24, 2015 - 5:21 am
My situation:

1. I ran a LCA to model consumption patterns of foodstuff from a large number of indicators. Fit statistics point towards the 5-k-solution which is substantively meaningful.

2. Interpreting the 5 classes, two classes can be clearly defined as healthy and three as unhealthy.

3. I now would like to add covariates (e.g., SES) in order to predict having a 'healthy consumption pattern'.

My aim is to qualitatively evaluate the overall pattern. I do not have information to quantify the healthiness of individual products. Moreover, theoretically many goods can only be evaluated as healthy in consideration to the remaining consumption. Also, in my case, it would not make sense to use the 2-, 3- or 4-k-solution. Only, the 5-k-solution differentiates the consumption clearly in terms of healthiness. And fit criteria identifiy this solution too.

Being only interested in predicting 'healty patterns' vs. 'unhealthy patterns', a merge would make the analyses much more straightforward and readable.

Is there a solution to do it in MPLUS (6.1) that I do not see? I refrain from exporting the class probabilies to Stata for the analysis (- class differentiation is only okay with entropy =0.84 and average posterior probs > 0.88).

Many thanks!
 Jody Mdoda posted on Tuesday, August 25, 2015 - 6:24 am
As additional information to my question above: I have just found that Collins and Lanza 2010, p. 172-176 describe something akin to what I want to do: Combining two latent classes into one single reference class, so that a binominal logistic regression can be run to highlight a particular research question. The only difference, I would like to combine classes on both ends (three classes into one reference class and two classes into one target class).

(They say it would be possible in Proc LCA. However, I cannot see any explications whether it boils down to the categorize-analyze-approach that I would like to avoid if somehow possible.)
 Bengt O. Muthen posted on Tuesday, August 25, 2015 - 8:24 am
You can do second-order LCA in Mplus, where you add a latent class variable (say c2) that influences the original (1st-order) latent class variable (say c1). You do "c1 on c2" specifying that c2 class 1 has probability 1 of resulting in the 2 healthy c1 classes and zero probability of resulting in the 3 unhealthy classes. And you specify the reverse for c2 class 2. then you regress c2 on SES.
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: