

Nonlinear constraints and complex sa... 

Message/Author 

Gabriel Nagy posted on Wednesday, September 01, 2004  10:29 am



One of the great features of Mplus 3 is the possibility to specify nonlinear constraints. This approach works well in the traditional SEM framework where no dataclustering is assumed. However, I have made the experience that the inclusion of nonlinear constraints in conjunction with the ‘complex data’ option reveals the wrong / uncorrected test statistics. The fit indices are equivalent to the ML results when no clustering is assumed. This problem seems not to be specific to the data I use. Equivalent models (with and without nonlinear constraints) revealed different fit indices when using the ‘complex’ option. When using the nonlinear constraints the modelfit was exactly the same for ‘type = general’ and ‘type = complex’,respectively. This was not true for the unconstrained model. What can I do to get the correct fit statistics for models that include nonlinear constraints in a situation where the data is clustered? 

bmuthen posted on Wednesday, September 01, 2004  11:11 am



Please send your inputs, outputs, and data to support@statmodel.com. 

Gabriel posted on Thursday, September 02, 2004  11:26 am



Thanks for the offer, but I first have to ask my boss... However, I think that the problem involves mainly situations where latent variances are constrained. The ‘complex’ procedure does not work in situations involving direct estimates of standardized paths. An example is a regression between two standardized latent variables with both total variances fixed to 1. This procedure works well when the variance of the predictor variable is fixed to 1 and the residual of the dependent variable is constrained to be 1 – b**2. This kind of model works in the general modeling framework. The fit is equivalent to a model with reference loadings fixed to a value of 1 and freely estimated variances. However, when I fit the constrained model using the complex option, I always get a scaling correction factor of 1. This is not the case when I use the ‘traditional’ model specification. Maybe I’m wrong, but my understanding is that both models are (or should be) equivalent and therefore the fit statistics / correction factors should be equivalent as well. Thanks Gabriel 

bmuthen posted on Friday, September 03, 2004  8:48 am



Perhaps you can send at least the output so we can help you. 

Back to top 

