gsem fail to converge with cov option

[Lucie Piaser]

Hello everyone,

I try to estimate a generalized structural equation model. I have multilevel data (individuals nested within municipality). My dependent variable is the individual feeling of fear that I estimate by individual (sex, age, education, work) and municipality level factors (homicide rate, security and justice spendings, gini index and its square form (possible non-linear effect)). I'm particularly interested in the impact of the gini index, that I suspect to be endogenous. gini_90 is my instrument variable. My code is as following:

gsem (fear <- sex age educ work homicide_rate security justice gini gini_sqr M1[municipality]) (gini <- gini_90 homicide_rate security justice)(gini_sqr <- gini_90 homicide_rate security justice), cov(e.fear*e.gini) cov(e.fear*e.gini_sqr) latent(M1) vce(cluster municipality)

However, this model fails to converge. I noticed that it gave some results when I remove one of the cov options, leaving only

cov(e.fear*e.gini)

or

cov(e.fear*e.gini_sqr)

.
Could someone explain me why this problem emerge and how I could solve it ?

Thank you so much,
Lucie

[Phil Bromiley]

You didn't get a quick answer. You'll increase your chances of a helpful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output, and sample data using dataex. A failure to converge could reflect an interaction of your model with your actual data - the same model could converge for some data sets and not others.

However, it is also possible that the model is simply not identified with the cov option. Identification in this kind of model is tricky but I note you have municipality dummies for one equation but not the others, and the iv's in two equations are identical. This may not apply here, but in many simultaneous equation situations, you need to have an exogenous variable in the system that does not appear in the equation for identification.

Since everything is observed in your model, you could look at 2SLS or 3SLS as alternative estimators.

[Lucie Piaser]

Thanks for your useful comment. After some research, it seems that indeed 2SLS is better suited for my goal and easier to implement. I now have to specify it correctly !