Multilevel CFA w/Ordered Categorical Indicators + WLSMV Estimation

96 views
Skip to first unread message

Kara Styck

unread,
Mar 23, 2018, 3:23:31 PM3/23/18
to lavaan
Is there a way to conduct multilevel CFA with ordered categorical indicators and WLSMV estimation in lavaan? If so, I'd greatly appreciate example code (ideally with a sample dataset, so I can reproduce the output). Thank-you in advance for your help! 

Mark Seeto

unread,
Mar 23, 2018, 4:41:04 PM3/23/18
to lavaan
According to http://lavaan.ugent.be/tutorial/before.html lavaan currently does not do multilevel CFA.

Terrence Jorgensen

unread,
Mar 27, 2018, 7:20:12 AM3/27/18
to lavaan
Is there a way to conduct multilevel CFA with ordered categorical indicators and WLSMV estimation in lavaan?

lavaan does handle some basic multilevel situations (only 2-level models with complete data, no random slopes)


Only in the development version.  See how to install:


But no categorical variables at this point.

Terrence D. Jorgensen
Postdoctoral Researcher, Methods and Statistics
Research Institute for Child Development and Education, the University of Amsterdam

Kara Styck

unread,
Mar 27, 2018, 12:42:24 PM3/27/18
to lavaan
Ok, thank-you! I have a follow-up question. In Mplus, you can use sampling weights to compute standard errors and the chi-square test statistic to account for clustered data. Is there a similar function in lavaan for ordered categorical data with WLSMV estimation?

Thank-you in advance for your help! 

Terrence Jorgensen

unread,
Mar 28, 2018, 7:23:04 AM3/28/18
to lavaan
In Mplus, you can use sampling weights to compute standard errors and the chi-square test statistic to account for clustered data. Is there a similar function in lavaan for ordered categorical data with WLSMV estimation?

The survey package can adjust for clustering, stratification, and sampling weights, and the lavaan.survey package provides an easy way to pass information to/from the lavaan and survey packages to get robust SEs for your SEMs.  Unfortunately, the survey package does not provide the tools to deal with categorical outcomes.

How many categories are there?  Some research indicates robust MLE  (e.g., estimator = "MLM") should yield sufficiently unbiased results when there are at least 5-7 categories, depending on how asymmetric the thresholds are.  But they only investigated regular CFA, not multilevel CFA.


If you have several items per factor, parceling would also be provide a way to items with more categories and therefore more likely for MLE to work.  But I do not recommend parceling if your research questions are about the measurement model itself.  Even if your research questions are about the structural level, you should take great care in how you allocate items to parcels in a theory-driven way:


or to take uncertainty into account if you allocate parcels randomly.  See the references on the ?parcelAllocation help page in the semTools package for more information.
Reply all
Reply to author
Forward
0 new messages