multigroup SEM with missing data

[Ivan Jacob Pesigan]

Has anyone tried to perform the multigroup SEM procedure described by Graham (2012) in lavaan? If yes, how did you prepare your input data? How did you specify the constraints?

Thanks very much.

Graham, J. W. and D. L. Coffman (2012). Structural Equation Modeling with Missing Data. Handbook of structural equation modeling. R. H. Hoyle. New York, NY, Guilford Press.

-Jek

[Terrence Jorgensen]

Has anyone tried to perform the multigroup SEM procedure described by Graham (2012) in lavaan? If yes, how did you prepare your input data? How did you specify the constraints?

That method is tedious, and it is much easier to use FIML, which provides the same results.  The only reason to use the MGSEM approach (one group per missing-data pattern) is if you do not have access to the raw data, but only the summary statistics in each group/pattern.  If you want to implement it, you can just specify the same label across groups, for each parameter in the model:

http://lavaan.ugent.be/tutorial/groups.html

Graham's chapter assumed software needed to fit the same variables in each group's model, but lavaan is more flexible.  You can specify different group's models in different blocks, the same way that different levels (in a multilevel SEM) can different models:

http://lavaan.ugent.be/tutorial/multilevel.html

Instead of "level: 1" and "level: 2" in the syntax, you would specify blocks for groups defined by missing-data patterns ("group: 1", "group: 2", "group: 3", etc.).  

Terrence D. Jorgensen

Postdoctoral Researcher, Methods and Statistics

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

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