Very small error variances with gsem

[Ross Macmillan]

I am using gsem to estimate a LC model with 5 quasi continuous covariates [family(gaussian) used]. All covariates have the same basic distribution but when I run the model I get a really, really small error variance for one of the covariates. Never seen this in 20 odd years. Anybody have any thoughts on what is going on?

[Weiwen Ng]

If you can show full output, it could help. Otherwise, my comments are based only on my limited prior knowledge (I don't run LCA with gaussian indicators as much - btw I assume that by covariates, you mean indicators; I'd normally understand covariates as things that aren't indicators but are related to latent class membership).

How small is "really, really small"? Also, is there a standard error of the variance, or is the SE missing? When you put it that way, I wonder if it's trending towards zero, i.e. it's near the boundary of the parameter space. I think I'd view that solution as questionable (and if the SE is missing, it's likely invalid). If you used the nonrtolerance option, I'd try again without it enabled - that disables a secondary convergence criterion, and it's used more with binary indicators, but I wonder what happens here. If you didn't use multiple start values, I would go ahead and enable that - people tend to use at least 100; I don't think they're super helpful with 2-3 latent classes but you might as well do it.

Actually, it's better if you can save the results of the random starts and examine them in Excel. Base Stata doesn't have a prepackaged option for this. I demonstrate one here. It requires saving things to Excel.