SEM : I always receive the message of "the model fails to converge" and I see the message "not concave" beside each iteration

[Mohammed Kasbar]

Dear Statalist respected users,

I am trying to run a basic model of 3 variables using SEM.
2 latent variables and one observed variable.
In order to estimate the first latent variable, I am using 25 variables (including only 1 or 2 dummy variables), while I am using 5 variables to estimate the second latent variable.
I contacted Stata technical support and they recommended me to delete some variables and to reduce the # iterations to 80, but the model still fails to converge.

This is the command I am using:

sem (CG -> BrdIn, ) (CG -> NCIn, ) (CG -> EDComp, ) (CG -> NEDCom, ) (CG -> BS, ) (CG -> Qual, ) (CG -> Busy, ) (CG -> FemaleNED, ) (CG -> Foreign, ) ///
(CG -> Overlap, ) (CG -> ACSize, ) (CG -> NCSize, ) (CG -> ExBrdExp, ) (CG -> ExBrdExpYrs, ) (CG -> ExTenure, ) (CG -> CEOTenure, ) (CG -> NEDBrdExp, ) (CG -> NEDBrdExpYrs, ) ///
(CG -> NEDAge, ) (CG -> NED9Yrs, ) (CG -> FemaleED, ) (CG -> NEDTenure, ) (CG -> DualCEO, ) (CG -> AC, ) (CG -> wroa, ) (AC -> ttlassets3, ) ///
(AC -> srleverage, ) (AC -> srfcf, ) (AC -> logbm, ) (AC -> industries, )(AC -> wroa, ), vce(robust) iterate(1000) latent(CG AC ) nocapslatent

Thanks a lot in advance.

[Afonso Rodrigues]

Dear Mohammed Kasbar

I already have faced a similar problem when I was estimating an ARDL regression.
If you have a high number of variables and, thus, a high number of observations, Stata will probably take a while to show you the outcome. However, if somewhere in your dataset, you have one or more variables which have "gaps", It's possible that Stata will not be able to show you the outcome.
Also, multicollinearity problems may also make it harder for Stata to present the outcome of your regression.
I then suggest you to check for gaps in your dataset ant then test your data for multicollinearity.

best regards

[Mohammed Kasbar]

Afonso Rodrigues

Thanks a lot indeed.
I tried reducing the number of variables by taking away the variables with no pronounced variance and I ended up with a reduced version of my variables set. Stata was able to converge the model but I am not sure if what I did sounds ok with statistics.

[Afonso Rodrigues]

Afonso Rodrigues

Thanks a lot indeed.
I tried reducing the number of variables by taking away the variables with no pronounced variance and I ended up with a reduced version of my variables set. Stata was able to converge the model but I am not sure if what I did sounds ok with statistics.

As I said, there are some tests you have to do in order to make sure your results are adequate and trustworthy. I am not the best person to help you since I am new in Econometrics, and I am not familiar with SEM, but I'm sure that there are a lot of pre and post estimation tests that you have to perform in order to check the "correctness" of your estimation and, thus, your results.

best regards

[Weiwen Ng]

Dear Statalist respected users,

I am trying to run a basic model of 3 variables using SEM.
2 latent variables and one observed variable.
In order to estimate the first latent variable, I am using 25 variables (including only 1 or 2 dummy variables), while I am using 5 variables to estimate the second latent variable.
I contacted Stata technical support and they recommended me to delete some variables and to reduce the # iterations to 80, but the model still fails to converge.

...

Per the problem you described in the title, it sounds like your model is not identified. It could be that there is not enough information to estimate some of the parameters. When you run your model, are there any parameters with missing standard errors? There may be a problem with some of those variables. I am not familiar with the specific type of model you are trying to fit, but you may have to constrain or eliminate some parameters. Also, I am not familiar with identification issues. At first glance, it seems like there are enough independent manifest variables for each latent variable, but it's worth checking with someone knowledgeable to see if there are any clear problems with the model.

Side note: both code and results are most readable if presented in code delimiters like below. My signature describes how to use the code delimiter button in the formatting toolbar.

sem (CG -> BrdIn NCIn EDComp NEDCom BS Qual Busy FemaleNED Foreign Overlap ACSize NCSize ExBrdExp ExBrdExpYrs ExTenure CEOTenure /// NEDBrdExp NEDBrdExpYrs NEDAge NED9Yrs FemaleED NEDTenure DualCEO wroa, ) ///
(AC -> ttlassets3 srleverage srfcf logbm industries wroa, ) ///
(CG -> AC, ), vce(robust) iterate(1000) latent(CG AC ) nocapslatent

[Mohammed Kasbar]

Weiwen Ng Afonso Rodrigues

Thanks a lot indeed. Much appreciated.

One more question please, do I have to account for endogeneity when I am using SEM? or, given the fact that we have simultaneous equations, then endogeneity problem is sorted out?

[Weiwen Ng]

Weiwen Ng Afonso Rodrigues

Thanks a lot indeed. Much appreciated.

One more question please, do I have to account for endogeneity when I am using SEM? or, given the fact that we have simultaneous equations, then endogeneity problem is sorted out?

I use SEM for rather different purposes, and I would recommend you ask an economist.

[Mohammed Kasbar]

Thank you