SEM with censored indicators

[Yi. F]

Hello,

I am wondering whether the latest version of lavaan could handle censored variables? I have only found documents dated back to 2011 saying this feature was not available yet (at that time). I am trying to fit an SEM with censored indicators using the WLSMV estimator. Is this possible with lavaan?

Thanks a lot!!!

Best,

Yi

[Terrence Jorgensen]

I am trying to fit an SEM with censored indicators using the WLSMV estimator. Is this possible with lavaan?

No, not directly.  Is your data censored (i.e., truly normal but with a floor / ceiling effect) or are you looking for a 2-part model like zero-inflated models for counts (e.g., a person is either a smoker or not (binary part), and among the smokers, the second part of the model explains individual differences in how much they smoke).  These imply very different processes and should be handled differently. 

• censored data:  you can specify your model in blavaan, save the JAGS / Stan syntax, and edit it to account for censoring by treating censored values as missing and drawing plausible values from the posterior.  This tutorial should help give you the idea.
• 2-part model:  Brown et al. (2005) describe a latent-growth model with 2-part models for indicators.  The basic idea applied to each endogenous variable can be extended to CFAs or path models.  They used Mplus (and scripts are on the Mplus site), but this should be doable in lavaan using estimator = "MML" and missing = "FIML", although MML is still experimental in lavaan and runs quite slowly.

Terrence D. Jorgensen

Postdoctoral Researcher, Methods and Statistics

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

UvA web page: http://www.uva.nl/profile/t.d.jorgensen

[Yi. F]

Thanks Dr. Jorgensen for your reply! I was referring to the first situation, where a censored normal distribution is involved and the variable is continuous, not categorical. I was thinking about something more like a Tobit factor analysis described by Muthén (1989). But I will try using the imputation approach if this is the only option available with the current version of lavaan. 

Muthén, B. O. (1989). Tobit factor analysis. British journal of mathematical and statistical psychology, 42(2), 241-250.

Thanks again for your input! I greatly appreciate it.

Best,

Yi