After having run a CFA model using two different versions of Mplus (4.2/6) I wanted to ask if it is possible that in the earlier version the model won't converge but in the later it does just fine. Nothing about the models changes between runs, but I am wondering what may be the reason for the non-convergence/convergence (ML estimation).
There are so many minor changes that have happened between 4.2 and 6, it is difficult to say. It could have to do with better starting values or small algorithmic changes. There have also been compiler changes. Any one of these could be the reason.
I am performing a multilevel CFA with the MLR estimator in Mplus 7. I have a model with a lot of parameters (+- 460 free parameters, +- 323 degrees of freedom) and a sample size of 1600. I had already had problems with the number of H1iterations, which I increased to 5000.
Now I get a saddle point warning. I've tried the recommendations made here: https://www.statmodel.com/download/Saddle%20point.pdf (so I set the Miterations to 50000, the Mconvergence to .000001, and the logcriterion to .000001). I still get the warning message, but below the warning it says: Model estimation terminated normally. So I'm guessing I can ignore this message. Is that right?
Second question: I've tried the model with and without the higher/lower convergence and iterations criterions, and I get somewhat different results. In this case, would it be better to use the default options or the increased/decreased options?