I’m working on a multi-group CFA for a psych tool using Likert scale data. I’ve set it up as ordinal and I’m using WLSMV estimation with theta parameterization. Here’s a simplified version of my model:
model_simplified <- '
factor1 =~ item1 + item2 + item3
factor2 =~ item4 + item5 + item6
factor3 =~ item7 + item8 + item9
factor1 ~~ factor2
factor2 ~~ factor3
factor1 ~~ factor3
'
fit_simplified <- cfa(model_simplified, data=mydata, estimator='WLSMV', std.lv=TRUE, parameterization='theta')
The initial CFA runs, but I get a warning about the variance-covariance matrix. The real problem comes when I try to do group comparisons:
group_cfa <- cfa(model_simplified, data=mydata, estimator='WLSMV', std.lv=TRUE, group='gender', parameterization='theta')
This throws an error about empty categories in one group. Oddly, when I use a different grouping variable, I get the same error but the CFA still runs.
Any ideas on why this happens with one grouping but not another? How can I fix this issue with empty categories? Thanks for any help!
hey there, i’ve run into similar issues before. sounds like u might have some low frequency categories in ur likert data for certain groups. have u tried collapsing categories? like combining ‘strongly disagree’ and ‘disagree’ into one? that could help with the empty category problem. just a thought!
I’ve encountered this problem in my research as well. It’s often due to sparse data in certain response categories for specific groups. One approach that’s worked for me is to use a robust ML estimator instead of WLSMV. Try setting estimator='MLR' in your CFA function. This treats the data as continuous rather than ordinal, which can be more stable with small sample sizes or infrequent categories.
Another option is to explore your data thoroughly before running the CFA. Use frequency tables or visualizations to identify which items and groups are causing issues. You might find patterns that suggest recoding or collapsing categories as a last resort.
Remember, though, that changing your data or estimation method can affect your results, so document any modifications clearly in your methods section.
Hey ExploringForest, that’s a tricky situation you’re in! 
I’m curious, have you taken a close look at the distribution of responses across your Likert scale for each group? Sometimes these issues pop up when you’ve got really skewed data in one group but not the other.
Maybe you could try running some descriptive stats or even plotting the response distributions for each item by group? That might give you a clearer picture of where the empty categories are coming from.
Also, I wonder if the sample sizes for your groups are pretty different? That could definitely contribute to this kind of problem.
Have you considered using a Bayesian approach instead? It might handle the sparse data better, though it’s a bit more complex to set up.
What kind of psych tool are you validating, if you don’t mind me asking? I’m always interested in learning about new measures!