Estimation algorithm with numerical i...
Message/Author
 carla rampichini posted on Tuesday, August 10, 2004 - 6:56 am
Dear Muthen and Muthen,
I'm trying to fit a two-level factor model with five ordinal variables. With the ALGORITHM=INTEGRATION option, the TECH8 OUTPUT shows in the column labelled ALGORITHM either EM or QN (very rarely) for different iterations. Particularly, I would like to know where I can find a description of the estimation algorithm used with numerical integration.
Moreover, from the description of the ALGORITHM option on page 404, it seems that if you specify ALGORITHM=INTEGRATION you cannot choose the optimization method. Is this true?

Carla
 bmuthen posted on Friday, August 13, 2004 - 5:07 pm
The default algorithm with algorithm=integration is "EMA", that is an EM algorithm with Accelerations using QN (quasi-Newton) or FS (Fisher scoring) steps when EM is slow. You can request algo= EM for a pure EM algorithm, or ODLL for a direct likelihood optimization. A forthcoming paper by Asparouhov-Muthen will detail this.
 davide morselli posted on Monday, April 20, 2009 - 2:22 am
Hallo,
I'm testing a Two-level CFA model with random slopes. I obtain very different results in the variance using ALGORITHM=INTEGARTION or EM. The estimates are congruent, but variances and residual variances are very different. Where EM gives a significant residual variance of the latent variable at within level and a not signifcant at between level, INTEGRETION estimates the exact opposite. I wonder why and which integration I should choose
 Linda K. Muthen posted on Monday, April 20, 2009 - 10:08 am
It is likely the convergence criteria are different for the two tracks of the program that you are using. You probably have not obtained the same loglikelihood for both analyses. You can sharpen MCONVERGENCE.
 davide morselli posted on Tuesday, April 21, 2009 - 2:31 am
Yes the loglikelihoods are different. By sharping Mconvergence you mean decreasing it?
 Linda K. Muthen posted on Tuesday, April 21, 2009 - 7:24 am
Check to see what the MCONVERGENCE value is for the analysis with the best loglikelihood. Use that number for the other analysis.