Loglikelihood PreviousNext
Mplus Discussion > Latent Variable Mixture Modeling >
Message/Author
 Vlad posted on Wednesday, August 04, 2010 - 10:25 am
Hello,
I have a message from Mplus "WARNING: THE BEST LOGLIKELIHOOD VALUE WAS NOT REPLICATED. THE SOLUTION MAY NOT BE TRUSTWORTHY DUE TO LOCAL MAXIMA. INCREASE THE NUMBER OF RANDOM STARTS."
I tried to increase the number of classes it didnt help. I would like to know what the potential sources of this message (besides local max.) and how to solve it.

regards,
V
 Jon Heron posted on Wednesday, August 04, 2010 - 10:34 am
Hi Vlad,

do you really mean that you tried to increase the number of classes??
 Vlad posted on Wednesday, August 04, 2010 - 11:06 am
hi Jon,
yes.
 Jon Heron posted on Wednesday, August 04, 2010 - 11:32 am
OK, thanks for the clarification

The more complex your model, the higher the chance you will get your error about the best log-likelihood not being replicated.

By more complex I mean
- more measures (e.g. more time points)
- more growth factors e.g. quadratic instead of linear
- more classes
- missing data (FIML) instead of complete-case
- probably depends on sample size too

As you pile on the complexity, you need to do more runs with different starting values to avoid the warning you mentioned.

For instance, for a 2 class model I might use

ANALYSIS:
proc = 2(starts);
type=mixture;
starts=200 20;

but I would incrementally increase the numbers on the starts row such that a 5-class model I might have

ANALYSIS:
proc = 2(starts);
type=mixture;
starts=2000 200;


Hope that helps. By increasing the number of classes, you were actually increasing the chance of your error, rather than reducing it.


best wishes, Jon
 Vlad posted on Wednesday, August 04, 2010 - 11:42 am
Thank you,
I will try.
regards,
V
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: