I s there any way that I can get histograms for the overall latent class probabilities in M-plus? I seem to be getting them only for individual items in my LCA graph output.
Thuy Nguyen posted on Wednesday, August 08, 2007 - 12:09 pm
There is no way to get this histogram from your LCA model. But as a workaround, you can use SAVEDATA to save the data and class probabilities from the LCA model. Then run a type=basic model on the saved data and include the class probabilities as regular variables. Request TYPE=PLOT1 in the type=basic run. This will allow you to view histograms on the class probabilities.
So you mean i would run a typoe=basic model with the original data as well as the cprobs as regular variables?
Thuy Nguyen posted on Wednesday, August 08, 2007 - 6:54 pm
The saved data file will include the original data for all the analysis variables in the LCA model along with the class probabilities (make sure to request SAVE=CPROB in the SAVEDATA command). Then run type=basic on this data. You can either use all variables or just the cprobs.
I ran a R3STEP analysis and I am looking for descriptive stats of the auxiliary variables for the sample and for each class. Is the procedure described above a way to achieve that? If so, where can I find more information to do it. I'm not sure how to include the class probabilities in the type=basic syntax.
Thank you Tihomir for your quick and clear answer! I finally ran the auxiliary variables as a DCAT or BCH because there were message errors when using DU3step (classification error in step 1, variance zero).
A lot of the auxiliary variables are sociodemographic characteristics. Are you saying that you would not recommend to present a table of sociodemographic characteristics that would include %/mean for each class because the descriptive stats are not observed? Would you suggest to only include the descriptive stats for the overall sample?
Also, does it make sense to compare the latent classes on the variables used to build LCA model? Can I use the "Latent class odds ratio results" to do so? And what can I use for the continuous variables?
Finally, is there a paper in which you discussed the use of p values correction for the R3step/DCAT/BCH analysis?
You can present mean for each class for the sociodemographic characteristics, however, keep in mind that the main model is a statement in the opposite direction, i.e., [C|X] rather than [X|C]. You can present both if that is useful. The model that R3step estimates is model that computes P(C=1|X).
If you want to compare the latent class indicators I would recommend looking at the entropy for each variable https://www.statmodel.com/download/UnivariateEntropy.pdf You can also look at the means for the continuous variables or odds ratios or results in probability scale for categorical variables. The p-values are based on the sandwich SE estimator.
I have another question. I would like to test the local independence assumption. Our LCA model has 4 categorical and 2 continuous indicators of latent class. For the categorical variables, I analysed the standardized bivariate residuals. For the continuous variables, I am not sure how to test the assumption.
What do you think of this procedure found in an article (Dembo et al., 2012)? "For the continuous variables, the local independence assumption was tested by introducing the observed variables as a latent factor in the LCA analysis and comparing the obtained BIC from this model with the BIC from the selected 2-class model. A smaller BIC for the selected 2-class model was obtained supporting the local independence of the indicators."
I tried the RESIDUAL command and TECH12, as well as UG ex. 7.16 and 7.22 but I am not sure how to interpret the output.
**Dembo R, Briones-Robinson R, Ungaro R, et al. Emotional/psychological and related problems among truant youths: An exploratory latent class analysis. Journal of emotional and behavioral disorders. 2012;20(3):157-168. doi:10.1177/1063426610396221.
Putting a factor behind the continuous outcomes and checking BIC is a reasonable approach. That's a kind of factor mixture model for which you see several papers posted as well as included in our short courses.