Jon Heron posted on Monday, December 08, 2014 - 6:25 am
I'm attempting to derive the appropriate H0 model prior to fitting a 2-class GMM with two covariates.
Since the Bayes estimator does not work with latent class regression I have been forced to remove the line "c on x1 x2;" from my imputation. I'm left wondering how to preserve the covariate effect during the imputation stage, and have opted for
%c#1% [x1 x2];
%c#2% [x1 x2];
although this feels less than ideal. Do you perhaps have any other suggestions?
I would also include class specific variance covariance for x1 and x2 and also use x1 and x2 as predictors of the intercept and slope of the growth model.
I would first try a simulation with no covariates and make sure you can see benefits of the H0 imputation over FIML in these simpler settings.
Jon Heron posted on Wednesday, December 10, 2014 - 12:57 am
Thanks Tihomir, I tried to include co/var terms but was informed that this would require Metropolis Hastings. Does this indicate I may have done something wrong, or would you be happy to embrace MH?
Jon Heron posted on Wednesday, December 10, 2014 - 1:15 am
Ignore me, I had been mentioning variances but not covariances. I can now run this without MH.
I guess I am now making distributional assumptions regarding X that would not be a requirement for my subsequent latent class regression. But those same assumptions will need to be made anyway once I've induced missingness in X.