Message/Author |
|
Anonymous posted on Tuesday, May 10, 2005 - 5:07 am
|
|
|
I work on a sem mixture model. Model with four classes is better than the three classes model. AIC and BIC are lower and entropy greater, but in this four classes model the size of the third classe is only 13 cases, [4146,223,13,1832]. Is it correct to choose such a model ? Is it preferable to choose the three classes model ? [4209, 1853, 152]. Thank you for your help. |
|
bmuthen posted on Tuesday, May 10, 2005 - 5:55 am
|
|
|
If the small class in the 4-class solution is clearly interpretable, it seems you can go with 4 classes. In this context, I would also bring in covariates before I make the decision. |
|
Anonymous posted on Tuesday, May 10, 2005 - 6:53 am
|
|
|
Thank you very much Dr Muthén |
|
|
Dear mr Muthen, I have a similar question about sample sizes. I don't have data yet but I'm thinking about how many respondents I need in my project. I will use general growth mixture modeling to search for subgroups in the sample, based on intercept and slope. It's more or less an explorative research. My question is how many respondents there have to be minimal in a subgroup? When having a subgroup with, for example, 10 subjects isn't that resulting in a very low power? Thanks for helping |
|
bmuthen posted on Tuesday, July 26, 2005 - 7:57 am
|
|
|
10 subjects in a class can be sufficient or inadequate - it depends very much on the model and its parameter values. For example, you may be in the advantageous situation that your model has only 2 parameters specific to the small class with 10 subjects - e.g. the means of the intercept and slope - while other parameters of that class are the same as those of other classes. In this case you may for example have enough power to reject that the slope mean is zero if the slope mean estimate has a low enough SE - which is a function of how clearly the data determines this slope. You can do a Monte Carlo simulation study to shed light on this, but if you have no pilot data nor strong theory, it is hard to choose parameter values for such a study. |
|
|
Hi, I have a question about assignation to classes. I use the mixture model and I want to know if it's possible to know for everybody in witch classes they are assigned. Thank you. analysis: type = mixture missing; starts = 20 2; model: %overall% i s q | hyb1@0 hyb2@1 hyb3@2 hyb4@3; plot: type = plot3; series = hyb1-hyb4(*); |
|
|
Use the CPROBABILITIES option of the SAVEDATA command. |
|
Back to top |