Yuchun Peng posted on Monday, September 28, 2009 - 7:49 am
I have tried several mixture models. However, it is almost impossible to have a global solution. I have tried to increase the starts sets up to 1000 100 with 20 stiterations. However, it does not solve my problem!!! the final loglikelihood was not still a global solution. I have followed the user book and checked whether the parameter estimates are similar except two parameters (the mean of SI and QI). Can I use the solution with the highest loglikelihood??
Amber Fahey posted on Wednesday, June 06, 2018 - 7:30 pm
I am using two-level growth modeling to model the trajectory of cognitive recovery during inpatient rehabilitation. I would like to compare a linear to quadratic model. I initially received the following warning when added the quadratic function: THE MODEL ESTIMATION DID NOT TERMINATE NORMALLY DUE TO AN ILL-CONDITIONED FISHER INFORMATION MATRIX. CHANGE YOUR MODEL AND/OR STARTING VALUES.
I followed suggestions in other posts and increased the starting values to 50 5, then received this warning:
THE BEST LOGLIKELIHOOD VALUE HAS BEEN REPLICATED. RERUN WITH AT LEAST TWICE THE RANDOM STARTS TO CHECK THAT THE BEST LOGLIKELIHOOD IS STILL OBTAINED AND REPLICATED.
I ran it again with starting values 100 10 and received the exact same Loglikelihood as the model with 50 5 starting values. What I can't figure out is if this means I'm good, or if I have received a local solution only? If so, what else can I do to make this work and verify that it is not a local solution only? I read in the aforementioned post to check TECH8 but I am wondering is there a tutorial or reference you can recommend to learn how to compare solutions?
RECEIVED: Loglikelihood H0 Value -2136.678
Amber Fahey posted on Thursday, June 07, 2018 - 12:57 pm
The best logliklihood and corresponding seed values, and all parameter estimates are are exactly the same for the models with 50 5 and 100 10 starting values. Can I assume then I did not receive a local solution or is it still necessary to run OPTSEED with the seed value obtained on the best LL? And if so, am I understanding the article right that I check that the parameter estimates are the same as what I received when not specifying OPTSEED?
I assume you are doing growth mixture modeling, not just growth modeling. The aim of Starts is to get several best logL values (several almost identical logL values that are listed at the top as the best ones) - the more you get, the more confident you can feel that you have hit a global, not a local optimum. Starts = 100 10 isn't a very high number - you might want to use Starts = 400 100 for example.
Amber Fahey posted on Thursday, June 07, 2018 - 7:06 pm
I am using Type = two-level random.
I used starts 400 100 as you suggested but I am not sure what I am supposed to do after that. The parameter estimates for the best LL using 400 100 is almost identical to those obtained with a slightly different LL (once I round to two decimal places) using starts 100 10, except the variance of my intercept is different by 0.072. In your estimation, does this mean that I have hit a global optimum?
The primary concerns is if the best LL is replicated. You don't say how many times the best LL is replicated among the 100 analyses you asked for. With Type=Random, it is sometimes hard to get exact replication with respect to all the decimals. A secondary concern is if parameter estimates are different.