CFA with ordinal (Likert) items
Hi JASP team!
Second pasdag in the Netherlands, I hope you will not answer to my question today :-)
I recently found out that CFA with ordinal items is different than the one with continuous items. Revelation! I went back to JASP to test my model based on this new knowledge and these are the two problems I faced:
- One error message reads "estimator ML for ordered data is not supported yet. Use WLSMV instead". However, the estimator options that I have do not include WLSMV, but Auto, NL, GLS, WLS, ULS, and DWLS
- To solve (1), I relied on WLS (starts the same, should be similar...), and the following error message that I get reads "The model could not be estimated. Error message: the leading minor of order 217 is not positive definite". Any clue what this may mean?
I also have a statistics question. Reading through literature, I read in DiStefano (2005) that "Ordered categorical (i.e., Likert) item-level data are not continuous, but may be treated as such under certain distributional conditions (i.e., skewness and kurtosis |2.0| and at least five ordered scale points (DiStefano, 2002; Muthén & Kaplan, 1985; West, Finch, Curran, 1995)". All my items meet these recommendations, apart from one, that has Kurtosis>2. Is this one item a reason to not treat my data as continuous? Would it make sense, e.g. to exclude it from further analysis?
Many thanks!
Georgios.
Comments
Hi Georgios,
Issue 1 looks more like a bug (unless it is explained in the help file). If it is not explained in the help file, please post it on our GitHub page (for details see https://jasp-stats.org/2018/03/29/request-feature-report-bug-jasp/).
Issue 2 generally means the model is overparameterized with respect to the data, that is, there exist multiple maxima in the likelihood landscape.
As far as the statistics question goes, the royal road is to model the ordinal scale as ordinal (possibly involve a statistician; I believe the brms package can handle these structures well). Or you could just set these items to continuous in JASP -- honestly, I would be surprised if it would matter much.
E.J.