I run a multi-group CFA with lavaan in R. Several categorical variables with 11 levels yield an incorrect final threshold ordering. Revised sample code follows:
Hey Luke87, I’m really intrigued by this unexpected final threshold ordering issue. I’ve played around with lavaan a bit and wonder if the problem could be related to sparse category frequencies or non-linear threshold spacing in your data. It’s kind of interesting because sometimes these ordering issues pop up if the underlying distribution in one group is not quite aligning with the default assumptions of the model. Have you tried looking at the empirical threshold estimates before any constraints are applied? I remember that in one case, manually inspecting those estimates shed some light on what might be forcing a reordering upon the final output. Also, any insights on whether this behaviour persists even when you adjust sample sizes or drop some levels could add another layer of clarification. Would love to hear your thoughts or any experiments you’ve conducted to work around this!
I encountered a similar issue while running multi-group CFAs with categorical thresholds. In my experience, slight differences in data distribution among groups can lead to unexpected threshold anomalies. One approach that helped was to experiment with the optimization settings, as convergence issues sometimes resulted in reorderings. It was also useful to manually inspect the threshold estimates and consider finer adjustments in constraint specifications. Ensuring that the empirical distributions were adequately represented in the model was key, and experimenting with different estimators occasionally provided more stable solutions.
hey luke87, i had a simlar prob. i sorted my factor levels explicitly and that solved the reordering issue. check your level counts too - low freqs can mess things up sometimes. hope this help!