Convergence not achieved - SEM with second-order latent variable

[Ted Nelson]

Hey!

I'm currently working with the SEM Builder in Stata 16.1 trying to estimate a model including a second-order latent variable (see picture what i'm trying to estimate). The construction of the whole model is based on theoretical arguments and was already tested in slightly different contexts. All the variables (Q3-Q22) are numeric on an ordinal scale (1-5) and there are almost 2000 observations. The majority of the data is either value 3, 4 or 5. From the 9th iteration, the log likelihood doesn't change anymore and all the values are 'not concave'. In addition, many standard errors are missing and i don't know how to fix this (i don't know how to make the parameters identified) and couldn't figure it out through Google/YouTube/reddit/Stata manuals. Does anybody have an idea on how to continue with this?

Any help is appreciated, thanks!

[Weiwen Ng]

I'm not an expert on model identification, but it appears that you have two latent variables in your model (loyal and satisf) that you are attempting to estimate with only two items each. I don't think those two variables are identified. The only way to solve that problem would be to add more items to each latent variable. That may mean that the survey originator should have added the items from the start, but that they did not do so. If that's the case, then there's no way to proceed without a time machine.

[Ted Nelson]

Weiwen Ng Thanks a lot! And i was thinking there's a problem with the second-order variable exp... Unfortunately, i do not own a time machine, but i can blame it on me, as i originated the survey. Your comment and hint solved my problem. Will work around this problem by somehow merging the two items or reducing it to only one, which should be fine with the questions asked. Cheers!