Question: ML CFA with ordered categorical data

84 views
Skip to first unread message

Christi Phillips

unread,
Apr 15, 2020, 3:48:57 PM4/15/20
to lav...@googlegroups.com
Does anyone know if lavaan can run maximum likelihood estimates with ordered-categorical data under a CFA framework? 

If so how? I am under the impression that I can only do this with OpenMx, but I hope that is not the case.


Thanks!
Christi Phillips

Edward Rigdon

unread,
Apr 15, 2020, 4:38:40 PM4/15/20
to lav...@googlegroups.com
     Maximum likelihood estimation assumes (conditional) multivariate normality, but ordered categorical data are not normally distributed. So using ML estimation would mean that you were assuming the data were not ordered categorical.
--Ed Rigdon

--
You received this message because you are subscribed to the Google Groups "lavaan" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lavaan+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lavaan/CAM9wzCJGQ2HoL4KU1eFc-Sg61qb0tmS_1VtaA4w-0jXzB6tTXw%40mail.gmail.com.

Patrycja Klimek

unread,
Apr 15, 2020, 4:49:35 PM4/15/20
to lavaan
I've had the same problem. In Mplus you can run the robust maximum likelihood estimator with ordered data...but not yet in R.
Also in response to the last person's comment, here is a helpful article justifying and supporting the use of MLR estimator with ordinal data:https://link.springer.com/article/10.3758/s13428-015-0619-7

Gavin T. L. Brown

unread,
Apr 15, 2020, 5:22:43 PM4/15/20
to lavaan
and to further muddy the waters about ordinal variables.
this chapter demonstrated that if the ordinal response scale had 5 or 6 ordered options, the data behave as if they are continuous. This can make life a lot simpler if you have a long enough set of options that people actually use.

Finney, S. J., & DiStefano, C. (2006). Non-normal and categorical data in structural equation modeling. In G. R. Hancock & R. D. Mueller (Eds.), Structural equation modeling: A second course (pp. 269-314). Greenwich, CT: Information Age Publishing.

Edward Rigdon

unread,
Apr 15, 2020, 5:49:36 PM4/15/20
to lav...@googlegroups.com
     If you choose MLR estimation, then you are treating the observed values as if they are actual numbers, not just ordered labels ("Never" "Sometimes" "Always"), as they are if the data are only ordinal. You can't use MLR with ordinal data. You either treat the data as ordinal and don't use MLR, or you treat the data as continuous but nonnormal and use MLR or some other continuous data estimator.
--Ed Rigdon

--
You received this message because you are subscribed to the Google Groups "lavaan" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lavaan+un...@googlegroups.com.

car...@web.de

unread,
Apr 15, 2020, 5:59:06 PM4/15/20
to lav...@googlegroups.com
Ed is of course right and on closer inspection this is also mentioned in the quoted article. As soon as you use ML(R), you assume that the data is continuous, even if it is not.
Am 15.04.20, 23:49 schrieb Edward Rigdon <edward...@gmail.com>:

Terrence Jorgensen

unread,
Apr 16, 2020, 9:24:18 AM4/16/20
to lavaan
I've had the same problem. In Mplus you can run the robust maximum likelihood estimator with ordered data...but not yet in R.

I addressed your misunderstanding in your other thread:  https://groups.google.com/d/msg/lavaan/LunOL4ZctV4/Mb7z0JqgAAAJ

Short answer: MLR is not available in Mplus for ordinal data, that is just a misleading keyword they use to trigger MML estimation, which IS available in R (lavaan, mirt, other IRT packages).  When using the probit link, they all assume a normally distributed latent item response underlying ordinal observations, and there is no robust option to free us from that assumption with moving to Bayesian MCMC estimation.

Also in response to the last person's comment, here is a helpful article justifying and supporting the use of MLR estimator with ordinal data:https://link.springer.com/article/10.3758/s13428-015-0619-7

They were less supportive than you imply, but Rhemtulla et al. (2012) were more supportive and informative when they published almost the same study a few years prior.   The ordinal data should have at least 5 categories (when approximately normal) or at least 7 categories (if thresholds are asymmetric) to treat as continuous using MLR.
  

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

Reply all
Reply to author
Forward
0 new messages