Information matrix: expected vs. obse... PreviousNext
Mplus Discussion > Multilevel Data/Complex Sample >
 Anonymous posted on Wednesday, August 24, 2005 - 7:35 am
hi Prof. Muthen

I am running a two-level model using mplus.
Mplus allow users to use OBSERVED vs. EXPECTED to calculate information matrix in two-level modeling.

So what will you suggest when to use which?
which one is more robust?

any ref. about the calculation/formulae of these information matrices you can recommend? I couldn't find it from the tech. appendix.

Thanks a lot!!
 Linda K. Muthen posted on Wednesday, August 24, 2005 - 8:37 pm
We recommend observed with missing data and otherwise expected. These are the Mplus defaults. The other options are avaiable for studying the effects of using different information matrices.

See the Kenward and Molenberghs reference in the back of the Mplus User's Gudie.
 Janine Neuhaus posted on Friday, January 09, 2009 - 1:08 pm
Dear Linda,

I ran a multilevel CFA twice, once a half year ago with version 3.1 and right now, with version 5.1.
I got different results, although the data were absolutely the same. I noticed that in the first case "expected" and in the second "observed" were denoted for calculation of the information matrix. Has the default setting changed from version 3 to 5?
As I do not have any missings (working with MI) I guess "expected" is in my case the correct setting, right?
Thank you very much for your help!
 Linda K. Muthen posted on Friday, January 09, 2009 - 1:15 pm
The default in Version 3 was listwise deletion of the data with missing values. In Version 5, the default to to use all information. This is why the information matrices changed. If you want listwise deletion, add LISTWISE=ON; to the DATA command.
 jml posted on Saturday, September 26, 2015 - 2:26 pm

I am fitting a single-level SEM model without any missing data and thinking of how to explain the difference between the ML and MLR estimators for my write-up. Is it correct to say that Mplus uses the expected information matrix to compute SEs and test statistic for ESTIMATOR = ML, and uses the observed information matrix to compute SEs and test statistic for ESTIMATOR = MLR?

 Bengt O. Muthen posted on Sunday, September 27, 2015 - 12:16 am
Not quite. See our tech appendix 8 from Version 2 on our website.
 jml posted on Sunday, September 27, 2015 - 2:36 pm
Thanks for the response Dr. Muthen. One more question: Asparouhov (2005) states that the asymptotic covariance matrix used by the MLM and MLMV estimators is explained in Muthen & Satorra (1995), the "technical aspects of Muthen's LISCOMP approach" paper. Is there another paper that discusses this acov matrix that is a little easier to understand? The 1984 paper is too technical for me to be able to follow. I really just want to write out the equation for this acov matrix, I don't need to explain how it's derived.
 Bengt O. Muthen posted on Monday, September 28, 2015 - 12:22 am
Muthen-Satorra 1995 concern WLSMV, not MLM/MLMV.
 jml posted on Monday, September 28, 2015 - 12:17 pm
Hi Dr. Muthen,

Thanks again for the response. Is there something that you could point me to that shows the acov matrix for MLM/MLMV, such as tech appendix 8 does for MLR?
 Tihomir Asparouhov posted on Tuesday, September 29, 2015 - 3:50 am
See (NTML estimator) in Section 3.2 in

MLM and MLMV have the same acov matrix.
 jml posted on Tuesday, September 29, 2015 - 9:54 am
Thank you!
Back to top
Add Your Message Here
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Options: Enable HTML code in message
Automatically activate URLs in message