I’m stuck on a multi-group CFA for a psych tool using Likert data. I’ve made the data ordinal and am using WLSMV with theta parameterization. Here’s a bit of my code:
model_psych <- '
mood_low =~ q9_ord + q16_ord + q17_ord + q18_ord + q35_ord + q50_ord
worry =~ q1_ord + q12_ord + q19_ord + q38_ord + q45_ord + q49_ord
physical_symptoms =~ q2_ord + q7_ord + q23_ord + q29_ord + q33_ord + q37_ord
mood_low ~~ worry
worry ~~ physical_symptoms
mood_low ~~ physical_symptoms
'
fit_psych <- cfa(model_psych, data=d_ordinal, estimator='WLSMV', std.lv=TRUE, parameterization="theta")
The initial CFA works, but I get a warning about the vcov matrix. The real problem comes when I try to do it by groups:
config_psych <- cfa(model_psych, data=d_ordinal, estimator='WLSMV', std.lv=TRUE, group="gender", parameterization="theta")
I get an error about empty categories in one group. Oddly, with a different grouping, I get the same error but the CFA still runs. Any ideas on why this happens or how to fix it? Is it because of the low frequencies in some response categories?
The issue you’re encountering is likely due to sparse data in some of your response categories, particularly when splitting by groups. This is a common problem with Likert scale data, especially when sample sizes are smaller or response distributions are skewed.
One approach to mitigate this is to collapse your Likert scale categories. For instance, if you’re using a 5-point scale, you could consider reducing it to a 3-point scale by combining the extreme categories with their adjacent ones. This can help address the empty cell problem without completely sacrificing the ordinal nature of your data.
Another strategy is to use bootstrapping methods, which can be more robust to sparse data. You might want to explore the lavaan.survey package, which allows for bootstrapping in CFA models.
If these approaches don’t resolve the issue, you may need to reconsider your model specification or data collection strategy. Sometimes, problematic items might need to be dropped, or additional data may need to be collected to ensure sufficient responses across all categories and groups.
Hey there Owen_Galaxy! 
I’ve run into similar issues with Likert data before, and it can be pretty frustrating. Have you considered collapsing some of your response categories? Sometimes when you have low frequencies in certain categories, it can cause these kinds of errors.
What’s the distribution of your responses looking like? It might be worth checking if you have any categories with really low counts. If you do, maybe try combining them with adjacent categories and see if that helps?
Also, I’m curious about your sample size. How many participants do you have in each group? Sometimes these issues pop up when you have a small sample size in one of your groups.
Have you tried running a descriptive analysis on your data, broken down by group? That might give you some insights into where the empty categories are coming from.
Oh, and one more thing - have you considered using a different estimator? WLSMV is great for ordinal data, but sometimes ULSMV can be more robust with small sample sizes or when you have issues with empty cells.
Let me know if any of this helps or if you want to bounce around some more ideas!
yo Owen_Galaxy, that sucks man. i’ve had similar probs w/ likert data. have u tried collapsing some categories? like combine the ones w/ low counts. might help w/ those empty category errors.
also, whats ur sample size like? if its small in one group, that could be why. maybe check ur data distribution by group?
good luck dude, hope u figure it out!