
Message/Author 

Marianne SB posted on Tuesday, January 12, 2010  3:36 am



Hi Linda and Bengt I am doing a quadratic GMM with longitudinal data from five time points (no outcomes are included so far). The indicators seem to favour a threeclass solution. I read that to do the LMRLRT test, the biggest group should be the last class, and I have problems with setting starting values that will get the classes in the right order. I have tried to specify seed and set starting values in all groups, but the biggest group still won’t go to class 3. Thus, this: Analysis: type is mixture; estimator is mlr; STARTS = 200 50; STITERATIONS = 10; optseed = 948615; lrtstarts = 0 0 100 50; Model: %OVERALL% i s q  gse13@0 gse15@.2 gse18@.5 gse21@.8 gse23@1.0; q@0; %c#1% [i*4.075]; [s*6.634]; [q*7.317]; %c#2% [i*3.135]; [s*0.848]; [q*1.525]; %c#3% [i*4.366]; [s*1.079]; [q*0.187]; still gives me the right solution, but the wrong order. So I read something about giving the starting values for the class probabilities in logit scale if the groups are of very different sizes. Is this the right thing to do here, and how is this done? 

Gareth posted on Tuesday, January 12, 2010  6:09 am



I am having a similar problem, and wondered if it would be possible in a future version of Mplus, to request the reference class in the input file (e.g. based on factor means, or class size)? 


You should not use OPTSEED. You should use STARTS=0; and get rid of STITERATIONS. 

Marianne SB posted on Wednesday, January 13, 2010  2:05 am



Now I got it right! Thank you very much! 

Back to top 

