sample.weights with categorical data

[Christopher Bratt]

I'm sorry for asking an elementary question, simply trying to understand current limitations. 

I want to use sampling weights (and incorporate clustering in other analyses). It seems we cannot do that with lavaan or lavaan.survey if the model includes categorical indicators of latent variables?

This restriction is hard to spot in the lavaan.survey pdf (here), where the following is written under a heading saying "Note", rather than "Limitations": "3) Currently only structural equation models for continuous variables are implemented." Notably, I received no warning when running the lavaan model and then using the lavaan.survey() command to apply sampling weights to a model with variables declared as ordered. 

Back to lavaan... I discovered that lavaan recently implemented sampling weights, but apparently not when the model includes categorical data? A model without sampling weights runs fine, adding the sampling.weights argument gives the following error:

 Error in lav_options_set(opt) : 

  lavaan ERROR: estimator ML for ordered data is not supported yet. Use WLSMV instead. 

WLSMV seems not to be supported for sampling.weights. The description of this argument says: 'Only available if estimator is ML in combination with robust standard errors and a robust test statistic. By default, the estimator will be "MLR".'

Am I correct in concluding that using sampling weights (or incorporating information on clustering) is not yet possible with lavaan as long as the model includes ordinal variables?

(This post might be taken as being critical towards lavaan. Quite the contrary, I am very impressed by lavaan and think it is unfortunate that one man has to do nearly all of the work with that package. I am just trying to figure out if/when I can move analyses from Mplus to lavaan.)

Christopher Bratt

[Terrence Jorgensen]

Am I correct in concluding that using sampling weights (or incorporating information on clustering) is not yet possible with lavaan as long as the model includes ordinal variables?

Correct.  Only DWLS estimation is available for categorical outcomes, thus none of the features are available that require MLE (e.g., FIML for missing data, sampling weights). 

Terrence D. Jorgensen

Assistant Professor, Methods and Statistics

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

http://www.uva.nl/profile/t.d.jorgensen