Missing means in grouped configural i...
Message/Author
 Jonathan Boudreau posted on Wednesday, March 25, 2015 - 1:39 pm
Hello Mplus community,

I am running a test of configural invariance on a series of continuous variables on multiple factors, grouped by occasion of measurement, and have encountered missing means/intercepts. I have compared two methods, the first of which is outlined in the Mplus 7 User's guide on p.485, and the second is a group based version of the script available at:

http://www.lesahoffman.com/948/948_Example9b_CFA_Longitudinal_Invariance.pdf

The main difference between the two models is that the first sets the factor intercepts to be zero, whereas the second estimates the factor intercepts by restricting an equivalent number of item intercepts to be zero. Both models are properly identified and should be equivalent.

The problem is that the second model requires global factor intercepts (or relaxing the restiction on the group factor intercepts) which are seemingly absent from the model (the first model has p more degrees of freedom than the second, where p is the number of factors).

Is there a way to relax the restriction on the group factor intercepts / add the global factor intercepts for the second model?

Regards,
J
 Bengt O. Muthen posted on Wednesday, March 25, 2015 - 5:04 pm
The 2 approaches should have the same number of parameters and be equivalent.

I don't know what you mean by "global factor intercepts".
 Jonathan Boudreau posted on Thursday, March 26, 2015 - 5:56 am
Hello Dr. Muthén,

Thank for taking the time to reply to my post. Perhaps "global factor intercepts" was a poor choice of words. Each of the, say k, groups is supposed to have a set of p factor intercepts that are "free", and a set of p item intercepts that are constrained to be zero. When I run the second model, k-1 groups are properly estimated. However, the remaining group has factor intercepts, as well as item intercepts that are mysteriously constrained to be zero. If we consider this last group as a reference group, then p baseline intercepts are missing from the model. This is what I meant by missing "global factor intercepts".

This is very puzzling because the general and k-1 group specific parts of the MODEL command have identical terms in regards to factor structures and intercepts, notably: (continued)
 Jonathan Boudreau posted on Thursday, March 26, 2015 - 5:56 am
(continued)
MODEL: !General model
f1 BY V1@1 V4* V7*;
f2 BY V2@1 V5* V8*;
f3 BY V3@1 V6* V9*;
[V1-V2@0 V4-V9*];
[f1-f3*];
f1-F3*;
f1 WITH f2* f3*;
f2 WITH f3*;

MODEL GX: !Group model
f1 BY V1@1 V4* V7*;
f2 BY V2@1 V5* V8*;
f3 BY V3@1 V6* V9*;
[V1-V2@0 V4-V9*];
[f1-f3*];

where X takes on values from 2 to k, and groups are appropriately defined as G1-Gk. Any explanation as to why the alternative specifications yield different models is very appreciated.

Best regards,
J
 Bengt O. Muthen posted on Thursday, March 26, 2015 - 10:24 am
The link you sent refers to longitudinal analysis, that is, a single group of individuals measured several times. But you talk about k groups and your input also looks like a multiple-group setup. Longitudinal analysis is not done as multiple-group analysis. Perhaps you have different groups of people measured at different time points. To make the situation clear, I think you need to send the 2 outputs to support@statmodel.com and point to the parameters that puzzle you.

Also, we ask that posts be limited to one window.