cluster, robust and fitmeasures(), for ordered/ordinal vars and its estimator

[joh4nd]

I am doing something like this, thanks for this software(!):

fit <- cfa(efa.model, data = data,
           sampling.weights = "pw", cluster = "clusters", ordered = TRUE,
           missing = "pairwise",
)

The cluster imposes robust standard errors as usual. In another (non-SEM) context, robust/sandwich st. err. can help indicate modelling problems, the more the significance of the results depend on them. SEM is different with all its GoF criteria. And lavaan is able to return the up to 3 versions--plain, robust, and scaled--that fitmeasures() of some criteria, whose values vary slightly but perhaps substantially, depending on the eye of the beholder. I have read in this lavaan google group that criteria labelled 'robust', not the rest, should be used for interpretation when returned (and the scaled version is old whereas the robust are in line with the literature). I don't understand why but noticed that references have been posted; OK. But in the context of 'cluster' I thought: Is this approach different? So should I discard the plain and scaled versions (when there are robust) simply because "cluster" is specified? Note, the resulting fits of the given type all seem fine, but there are differences.

Furthermore, since the model specifies ordered = TRUE , this implies estimator = DWLS--and as far as I have understood--se = "robust.sem", test = "scaled.shifted" (alternatively, test = "mean.var.adjusted"), because this is what WLSMV is. But that does not not result in robust fit criteria by fitmeasures(), which I find somewhat confusing. So why does lavaan not do that? If I refrain from specifying the cluster, there are no robust fit results and the plain even gets slightly better (toward outstanding/excellent). Of course, what matters is getting it right, not desirable/comfortable, which is also why I ask these questions.

One last thing, must parameterization="theta" be specified with ordinal data? I read somewhere here that the option was made available for ordinal/ordered data, but it seems to be the case that it is not forced/imposed automatically. So I guess it primarily contributes to rescaling. 

Johan

[Terrence Jorgensen]

           sampling.weights = "pw", cluster = "clusters", ordered = TRUE,

Cluster-robust results are not available for models with ordinal outcomes.  There is an open feature request:

https://github.com/yrosseel/lavaan/issues/254

If/when the feature is ever added, you will see that issue updated as "resolved".

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

[joh4nd]

Why do I get results then? And thanks for the link - I noticed it in another thread.