GMM with skew t distribution
Message/Author
 Siny Tsang posted on Monday, April 03, 2017 - 9:01 am
Hello,

I am trying to modify the Mplus code in Muthen & Asparouhov (2015) to fit some longitudinal BMI data, but I'm a bit confused with the Mplus codes provided in the Appendix.

The data is positively skewed, so we're looking to use DISTRIBUTION = SKEWT. With time varying covariates (i.e., age), I am assuming that we should use the codes in Tables B1 and B2 as an example. Instead of using the AT command to specify time-varying covariates, I see that the ON command is used. However, I don't quite understand why the regression paths are constrained to be equal for the different time points? Is there a way to get the equivalent of the slope estimates as in GMM with normal distribution?

Thanks.
 Siny Tsang posted on Monday, April 03, 2017 - 10:03 am
Follow up to my previous message,

I am guessing that the estimated coefficient for the ON path is the same as the estimated means for the slope if we use the usually GMM script:

i s | w1 w2 w3 w4 AT age1 age2 age3 age4;

So what happened to the variances of S if we use the ON method like this?

i BY w1-w4@1;
w1 ON age1;
w2 ON age2;
w3 ON age3;
w4 ON age4;

Is this essentially a fixed slope within class (but vary between class)? If so, can we model a class-varying slope effect as well?

Thanks.
 Bengt O. Muthen posted on Thursday, April 06, 2017 - 5:39 pm
Which Tables B1 and B2 are you referring to - in which document?

For the second question, say

s | w1 ON age1;
s | w2 ON age2;
s | w3 ON age3;
s | w4 ON age4;
 Siny Tsang posted on Monday, April 10, 2017 - 6:54 am
The Tables B1 and B2 were in the "Growth mixture modeling with non-normal distributions" paper (Muthen & Asparuhov, 2014).

Is there a way to get the class-varying slope effect with DISTRIBUTION = SKEWT?

Thanks!
 Bengt O. Muthen posted on Tuesday, April 11, 2017 - 6:06 pm
I'm sorry but I don't see such tables in that paper - can you email me what you are looking at?

Class-varying slopes can be handled. If not directly, then indirectly using a factor with class-varying loadings.