Jon Heron posted on Friday, September 29, 2006 - 10:52 am
I think I may have found a problem and although I need to run a few more checks on this, I thought I'd post my findings to see if anyone else has found the same.
I've been running a 5-class LCA on 5 time points (binary vars) and a sample of 10,000 using the 'noserr' and 'nochi' options to improve running time.
In the most recent run I used starts = 20000 10000; stiterations = 10; stscale = 20; (just to be doubly sure).
I've been finding that the optimal model is repeated with a good number of starting values but that using the optseed option will not recreate the model which appears as optimal.
After much headscratching I think this is due to the 'noserr' option. With this turned off I am presented with a warning about positive definite matrices and hence it seems quite clear to me why the optseeds are not doing their job.
Like I say, this is still work in progress. It may turn out to be nothing to do with 'noserr', but either way I have something strange going on here.
I have perhaps experienced this same issue. I wrote to Mplus support and was informed that I need to specify the same STSCALE that was used to obtain the optimal solution. What I had done was when I specified OPTSEED, I deleted the STSCALE line - and thus it defaulted to a value of 5.
So if you add STSCALE=20 next to your OPTSEED command, you might fix your problem.