Hi All,
I have run a default model and it can be calculated and has good fit both in two data sets (different topics) separately. Now I want to check for the moderating effect using multiple group analysis with AMOS. However, the unconstrained model, measurement weights model, structural weights model and structural covariances model are not identified.
My question is:
1. What is the reason for the unidentifiability of the unconstrained model? I am a bit confused since the default model without multiple group analysis definitely works.
2. What should I do to make unconstrained model, measurement weights model and structural weights model identified?
3. If unconstrained model, measurement weights model, structural weights model cannot be identified, is there some other way to test the moderating effect?
Hope someone can enlighten me, thanks so much for the help!
In general, if a model runs fine in both groups separately, then it should also run as a multigroup model. So there must be something about your current multigroup specification that makes the model underidentified.
Here is my guess from the info that you provided: It looks like the path W1 from the structural (latent variable) residual e13 is a free parameter in your path diagram. This path should typically be fixed to 1 in both groups (in the same way as for the observed variable residuals) rather than freely estimated. This could explain the message about "2 additional constraints needed"--the model would not be identified with this path freely estimated in either group. Could you try fixing the path W1 to 1 and both groups and re-running the model?
Christian Geiser
Hi Christian,
Thank you very much for your quick answer and sorry I'm late.
I have fixed the path W1 to 1 and re-run the model. I am happy to find that measurement weights model, structural weights model and structural covariances model can be identified. Your advice is very useful! Only the unconstrained model is still unidentified. Its result says that iteration limit reached. I have tried to specify a larger iteration limit (Menu: View→Analysis Properties) but it didn’t work.
Best,
Bo Bo
With "unconstrained", do you mean the configural measurement invariance model? I suspect that there is still some kind of error in your model specification. It's impossible to know without knowing more details about the model specification and/or output.
Christian Geiser
Hi Christian,
Thank you very much for your quick answer.
Yes, unconstrained model is baseline model without any parameter set to be invariant across groups. I have checked my model specification many times but as a complete beginner to SEM, I failed to find out what’s wrong. I attached some pictures about model specification and output. And if there is anything more you need, please let me know. I am extremely grateful for your kind help (^o^)／ .
Best,
Bo Bo
It looks like you are getting a negative variance estimate for the exogeneous factor LS (-0.013). This can have multiple causes. One common cause are non-substantial correlations of the indicators of the factor in question. I would check the correlations between the 3 indicators of the LS factor. Are they substantial? If not, this could explain that the factor is weak. The factor variance essentially reflects the covariances among its indicators. If the indicators aren't substantially correlated (meaning their covariances/correlations are close to zero), the factor variance would be near zero also, which can lead to an improper solution (negative factor variance estimate).
Also, what do you get when you simply estimate the same model as a single-group analysis (in each group separately)? When you estimate the model freely in each group, this is equivalent to the configural model in terms of the parameter estimates. Do you get the same problems/error messages/negative factor variance? If not, then there must be a problem in your multigroup specification of the configural model.
- Badges
- Science topic
- Similar topics
- Philosophy
- Esthetics
- Interpretation