Carolin posted on Thursday, January 03, 2013 - 9:25 am
I'm running a GMM wit covariates. I know that in TECH4 I can find the means of the growth factors - but how can I know if they are significant?
A second question: how can I interpret the influence of covariates on s and q? For example, depressive symptoms have a significant negative effect on the (negative) linear slope, but a positive significant effect on the quadratic term in youth with externalizing symptoms?
To see if the growth factor means are significant you can run an unconditional model (no covariates).
Perhaps your linear slope s is negative and the quadratic q positive so that the development first goes down and then up. The negative slope on a covariate for s means that as the covariate value increases s gets larger negative values. The positive slope on a covariate for q says that as the cov value increases q increases. So all in all a deeper dip and a stronger upswing as the covariate value increases.
Carolin posted on Friday, January 04, 2013 - 1:43 pm
Thanks a lot for your quick response and the very helpful interpretation!!!
Two related questions concerning the unconditional model: due to missings in my covariates I have a smaller sample size in the conditional model (in comparison with the unconditional model). Is this a problem when I refer to the significance of means in the unconditional model? Furthermore, the sizes of classes change a little bit when I add the covariates. Is that a problem?
I would not worry about those minor issues. But if this bothers you, you may want to think about a new 3-step mixture approach described in Web Note 15 - see our home page.
Carolin posted on Thursday, January 10, 2013 - 10:40 am
Thank you very much, you are very helpful.
I have another question concerning the log-likelihood difference testing. First, I analyzed different models (GMM) for girls and boys separately. Now I want to compare classes with multiple group analysis (two times log-likelihood difference). For example, I am interested if the intercept of class 1 in girls is equal to that of class 1 in boys. My question: Is the H0 model the more restrictive one (constrain two intercepts to be equal) and the H1 model the less rectrictive (no equalities)? Or differently asked: which value do I have to subtract from which one?
I believe this is the default. You can tell by checking your results or TECH1. You free them by mentioning the in the class-specific part of the MODEL command:
i ON x;
i ON x;
Carolin posted on Thursday, January 24, 2013 - 11:56 am
Thank you very much!!!
Can I also free the multinomial logistic regression of latent class variable (c) on covariate x for group 1 (girls) and group 2 (boys) in multiple group analysis? At the moment, I can find only one OR for each class in comparison to the reference class in the results , but I expected to have different values for girls and boys... How can I handle this (syntax)?
I would like to compare the odds ratios via chi-square difference testing... or isn't that possible (sorry for that silly question)
For b), see the UG chapt3er 14, Section Calculating Probabilities From Logistic...
It shows log odds and to get odds you just exponentiate, odds = exp(logodds).
Carolin posted on Wednesday, April 17, 2013 - 12:43 pm
I had to focus on some other things in the meantime but now I am back on this. Unfortunately, I still do not understand how to test if two odds ratios are equal in two known classes (even after I read the recommended section in chapter 14...). Can you tell me the syntax I need?
I would be very happy about your help... Thanks a lot