I am curious of the difference between the Vuong-Lo-Mendell-Rubin LRT test versus the Lo-Mendell-Rubin adjusted LRT test. What does the adjusted test adjust for, and is one preferred when making class enumeration decisions (obviously in conjunction with other criteria such as ABIC, BIC, and BLRT).
X is not related – we remove the first class where all the beta coefficients are (not the last where all the coefficients are fixed to 0).
To understand how tech 11 works you must carefully consider the reported log-likelihood value for the k-1 class (that will be in this case a single class with no effect of X). That k-1 class run should be run separately to ensure and verify that the test is for the correct k v.s. k-1 classes.
Thank you. So I simply leave out the x variable in the separate k-1 run (this is a simple growth curve model). However, the LL from the separate run and the LL reported in TECH11 output for the 2-class model do not match. Does x have to appear somewhere in the input syntax? This is what I did:
2 class model: Classes=c(2); Usevariables are y1-y4 x; %Overall% i s q | y1@0y2@1y3@2y4@3; s-q@0; c#1 on x; (TECH k-1 LL=-26518.010)
1 class model: Classes=c(1); Usevariables are y1-y4; %Overall% i s q | y1@0y2@1y3@2y4@3; s-q@0; (LL=-26552.888)
The code is correct and you should have obtained the same LL. You can use the data from User's guide example 8.1 and see that it gives the same LL.
You can try using starts=100 with the 1 class model, or use some stricter convergence criterion such as mconv=0.0000001. If you still don't get the same LL send your runs and data to firstname.lastname@example.org