| Message/Author |
|
|
|
|
| I apologize in advance for perhaps a stupid question. After LCA, how do I save the class membership (i.e., which participant belongs to which class)? What would be the syntax and where should it appear in the syntax file? |
|
|
|
|
Savedata:
save = cprob;
file = ... |
|
|
|
|
Dear Dr. Muthen,
Many thanks. May I ask you a related question about no. of class decision?
In my analysis, BIC suggests 3 classes whereas aBIC suggests 4 classes. At the same time, BLRT keeps suggesting further differentiating the sample to 5 classes or more.
Under the circumstances, what should I do to determine the optimal no. of classes?
Tak |
|
|
|
|
Sorry, I am new to LCA. I have now also referred to the Lo-Mendell-Rubin test which surprisingly was not significant in the comparison between 4 and 3 classes. But BLRT has a p < .001. Why is there such a big discrepancy between the two tests? What should I do given the discrepancy?
So my questions are twofold: 1) How do I reconcile the differences between BIC and aBIC (or more commonly labeled ssaBIC); 2) How do I reconcile the differences between BLRT and LMR?
Apology for so many questions. |
|
|
|
|
| These fit measures often give different results. I would simply use BIC. |
|
|
|
|
Many thanks for your reply. The problem is that journal reviewers often want a range of statistics... Anyway, I have further questions and hope you don't mind:
1) The indicators have 3 levels (scored 0, 1 and 2) that reflect an ordinal scale. Would it be inappropriate to use LPA instead of LCA? I found results stemming from 3 categories difficult to interpret in terms of describing the features of each class.
2) If I understand correctly, LPA does not require multivariate normality but BLRT and LMR do. Is there a way to test for multivariate normality in MPLus? I have searched the User Manual as well as on web but cannot find a clue.
3) If BLRT and LMR assume multivariate normality, does it mean they are not suitable for LCA? It's curious that I have seen people reporting these statistics in LCA papers.
Tak |
|
|
|
|
1) This depends on how large a percentage is at the end points. If greater than say 25%, I think you should treat them as categorical (ordinal). The solution can be interpreted via indicator probabilities.
2) - 3) Mixture modeling needs non-normally distributed variables - otherwise, a mixture cannot be found. You are probably thinking of within-class normality which is a different matter. There is no need to test for normality. |
|
|
|
|
Many thanks again for your reply. I have read many discussion threads in this forum and you keep suggesting to use BIC. But in the Nylund et al. simulation study, it says that aBIC performs better than BIC when the indicators are categorical and class membership is unequal in size. Can you comment further on the relative merits of BIC and aBIC?
Tak |
|
|
|
|
| Maybe you refer to the 2007 Nylund et al article and its Table 7 where aBIC does better than BIC in a couple of cases with categorical outcomes and low sample sizes. In Table 7 page 557 (categorical outcomes), BIC beats aBIC in 2 of the 4 panels and in Table 7 page 558 (continuous outcomes) in almost all cases. My feeling is that it is difficult for simulation studies to cover realistic real-data settings in a comprehensive and good way and therefore difficult to set rules for what should be reported when. Perhaps aBIC is better in a few cases but I prefer to go with a single measure all the time so I never use anything but BIC these days (I report the loglikelihood as well to see which model is closest to the data - and of course the # parameters, but nothing else). |
|
|
|
|
| By loglikelihood, do you mean the likelihood ratio chi square? How do you interpret the value as to which one is closest to the data? Just smaller is better? But this value keeps decreasing with more and more classes being specified. |
|
|
|
|
| No, I mean the log likelihood value that is printed - the higher the better with maximum likelihood. It does get trivially better when increasing the number of classes but for two models with the same number of classes, their LL difference is of interest. |
|
|
|
|
| Really thankful for answering my question. I have now proceeded to calling TECH10 with the Auxiliary command (about 10 variables specified on this command). However I find it difficult to read the results and understand what analysis was being done. Is there any reference for me so I can read and interpret the analysis properly? |
|
|
|
|
| TECH10 is discussed in Topic 2 of our Short Course videos and handouts on our web site. It is hard to learn about these topics on Mplus Discussion by asking one question at a time. If you run into issues, send your output to Support along with your license number. |
|
|
| Back to top |
