Convergence problem

[Chao Xu]

Hi all,

I ran a fairly large bifactor model (40 items loading on 7 factors) on ordinal scales (7-point scale). Here is the problem.  When I specified residual terms to be uncorrelated (i.e., the residual covariance matrix is diagonal), estimation of this model didn't converge. However, if I specified one (and only one) pair of items to be correlated, the model converges rapidly. 

What could be the possible causes that are associated with this issue? Any thoughts will be enormously appreciated. Thank you.

Best,

Chao

[Terrence Jorgensen]

What could be the possible causes that are associated with this issue? Any thoughts will be enormously appreciated.

I would recommend posting on SEMNET, where there are several users familiar with bifactor models that might have thoughts (and because this is not a lavaan issue).

http://www2.gsu.edu/~mkteer/semnet.html

Terrence D. Jorgensen

Postdoctoral Researcher, Methods and Statistics

Research Institute for Child Development and Education, the University of Amsterdam

UvA web page: http://www.uva.nl/profile/t.d.jorgensen

[Chao Xu]

Thanks, Terrence.

My data were collected in a pre-post manner. What is wired is that, when I fit bifactor model to the post data, it converges rapidly in both cases regardless of specification of correlated residuals. But when it comes to the pre data, it converges only when a pair of residuals are specified to be correlated. That's why I am getting curious as to what might have caused this divergence. Do you think it could be a numerical issue relevant to lavaan due to the way it optimizes?

Chao

[Terrence Jorgensen]

Do you think it could be a numerical issue relevant to lavaan due to the way it optimizes?

Possibly, but seeing both your model and your data would be necessary to track down the problem.