Dear Ladies and Gentlemen,
I am currently working as a Master's student on several publications that methodically rely on latent profile analyses. In the context of these LPAs, I have repeatedly encountered the problem, that calculating the BLRT test with the help of Mplus (TECH14) is both time-consuming and often unsuccessful. In this specific case, an LPA with 30 Likert-scaled items (1-5) of the "Big Five" (OCEAN model) with a sample size of N= 738 was conducted. I would be interested to know, which approach you prefer in your research work.
Q1: Do you increase the number of bootstraps and LRT starts as the number of profiles increases, or do you exclude the BLRT test when you encounter error messages and instead refer to the Loglik value, LMR test, and the BIC fit indicator?
So far I have tried the following settings for TECH14 according to the recommendations of the forum entries by Muthen & Muthen:
LRTSTARTS 0 0 100 20 / LRTSTARTS 0 0 500 50.
Both of these options will result in unsuccessful bootstrap draws if more than three profiles are calculated for the model.
Q2: Do you treat your Likert scaled items as interval scaled variables and use the MLR Estimator or do you treat your indicator items as ordinal variables and use the WLSMV Estimator?
In my case, attributing the items as categorical with the WLSMV Estimator leads already with two profiles to a "non-positive definite first-order derivative product matrix".
There seem to be conflicting opinions here. Brown (2006) writes "The WLSMV is a robust estimator which does not assume normally distributed variables and provides the best option for modeling categorical or ordered data".
On the other hand, Bengt.O. Muthen (2006) :
The most important aspect is how strong floor and/or ceiling effects you have. If they are strong, you may want to use a categorical approach.
Q3: Would any of you be willing to cross-check my syntax for comparing distal outcomes with the BCH approach? (See appendix)
Thanks in advance for your help.
Philipp Schulz
References:
Brown, T. (2006). Confirmatory factor analysis for applied research. New York: Guildford.
Bengt. O. Muthen (2006): http://www.statmodel.com/discussion/messages/13/1236.html?1145023154
Wow will you get answers. Good luck with your research.for me likert is ordinal and should be treated that way. Best of luck to you, David Booth
Philipp Schulz I agree with David Eugene Booth . Since your items are ordinal (ordered categorical), you would typically want to use classical latent class analysis (LCA), not latent profile analysis (LPA). LPA is designed for continuous (metrical, interval) indicators, whereas LCA is for categorical (binary, ordinal) indicators. The problems that you encountered with the BLRT may be related to the large number of items (30 items is a lot for both LCA and LPA). This does not necessary mean that using fewer items is better (in principle, it is good to have many indicators, as long as they are good class indicators), but it may explain the problems in successfully conducting the bootstrap. The BIC is a simple and good alternative in my experience. And yes, as the number of classes / profiles goes up, you should definitely increase the number of starts for both the target (estimated) class model and the bootstrap. You should also check that the best loglikelihood value for each model can be replicated at least a few times so as to minimize the risk of local likelihood maxima (potentially invalid solutions and/or invalid bootstrap results).
With regard to ML vs. WLSMV, you may be confusing latent class/latent profile analysis with confirmatory factor analysis (CFA) and structural equation modeling (SEM). Both LPA and LCA typically make use of maximum likelihood (ML) estimation, not WLSMV. WLSMV is used for ordinal (ordered categorical) variables used as indicators of continuous latent variables ("factors") in CFA and SEM.
If you have not seen this , you may find it useful - you are clearly not alone in having trouble in getting these complex and demanding models to fit in Mplus:
http://www.statmodel.com/discussion/messages/13/4529.html?1249068377
I think it would be a good idea to contact them directly for their advice https://www.statmodel.com/support/
Their replies are generally helpful and thorough.
Dear Kelvyn Jones , David Eugene Booth and Christian Geiser
Thank you for your assessment and recommendations.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling