|
Message/Author |
|
Anonymous posted on Monday, June 06, 2005 - 8:54 am
|
|
|
Two years ago I ran the following code with Mplus 2: MODEL: %OVERALL% c#1 ON educ; f1 by madher1-madher6@1; f2 by madher1@-2.5 madher2@-1.5 madher3@-.5 madher4@.5 madher5@1.5 madher6@2.5; f3 by madher1@6.25 madher2@2.25 madher3-madher4@.25 madher5@2.25 madher6@6.25; [madher1-madher6@0]; %c#1% [f1*-0.03409 f2*0.00019 f3*0.000403]; %c#2% [f1*-0.3461 f2*0.03901 f3*0.05065]; OUTPUT: tech8; The sample size is 69. I obtained a reasonable grouping of the two groups (29/40) with very nice interpretation of the mean values of the two groups. The loglikelihood is about -151.383. Now I re-ran the exactly same code with Mplus 3, because the code automatically searches the starting values of the largest loglikelihood, it chose the starting value with the loglikelihood at about -127.675 and continued. But this time the algorithm does not have a valid Fisher information matrix. I also tried a simplified Mplus 3 code as following (no class specific starting values): MODEL: %OVERALL% f1 f2 f3 | madher1@0 madher2@1 madher3@2 madher4@3 madher5@4 madher6@5; c#1 ON educ; It seems that Mplus cannot choose from the ten random trials for a good Fisher Information matrix and ask for user-supplied starting values. I would like to know if my old results with smaller loglikelihood are qualified as MLE through Mplus 2. For the improvement of the loglikelihood, should I choose the larger loglikelihood with bad Fisher information or stay with the old results? Thanks. |
|
|
This question should addressed to support@statmodel.com. Please send both outputs, data, and your license number. Some defaults have changed between Version 2 and Version 3. This is likely the cause but we can determine exactly what is happening if we can have the above information. |
|
Back to top |
|
|