GSEM - log-likelihood "not concave"

[Krizia Garcia]

Hello

I used sem builder to estimate a model running maximum likelihood algorithm. However, no results come up and iterations just keep on going endlessly. The syntax is as seen below:

gsem (GovExpUSD -> GDPgrowth, ) (TRevUSD -> GDPgrowth, ) (GovEff -> GovExpUSD, ) (GovEff -> TRevUSD, ) (GovEff -> CorrINDX, ) (TaxStruct -> GDPgrowth, ) (TaxStruct -> GovExpUSD, ) (TaxStruct -> TRevUSD, ) (TaxStruct -> CorrINDX, ) (PolStab -> CorrINDX, ) (IndrTaxGDP -> TaxStruct, ) (DirTaxGDP -> TaxStruct, ) (CorrINDX -> GDPgrowth, ) (RegQual -> TRevUS D, ) (Law -> GovExpUSD, ) (Law -> TRevUSD, ) (Law -> CorrINDX, ) (Voice -> GovExpUSD, ) (Voice -> CorrINDX, ) (LFPR -> GDPgrowth, ) (LFPR -> TRevUSD, ), cov( e.GovExpUSD*e.TRevUSD e.CorrINDX*e.GovExpUSD) nocapslatent

Must there be something wrong with the model? I tried the model previously using different variables (e.g.: instead of Government Expenditures USD, I used Government expenditures as a percentage of GDP) and it went well. How can this be remedied?

Hoping for a reply
Thank you

-Krizia

[Marcos Almeida]

You didn't give much detail about the variables.

That said, I have three notes as a suggestion: make sure you selected at least three factors for each latent variable; start with a less complex command; shall you adopt a recursive model, you may test it out as a non-recursive model.

I said three notes but, on second thoughts, there are more: make sure the sample is big enough for the model; the rationale is sound enough; the measures were good enough.

[Krizia Garcia]

I have no latent variables.. I am attaching a screenshot of the diagram created with SEM builder. I do not understand why it does not work when in fact I have done the same previously, only with different variables.. Yes I believe I have sizeable amount of observations.. it is for all countries from 1967-2015..

Thanks!
-Krizia

[Phil Bromiley]

Four suggestions. First, many of us find it easier to work with do files that have reasonable length lines. You can use continuation /// to continue on a new line. For example, we can't see all of your code easily. This also makes it easier for you (see Long's book WorkFlow of Data Analysis Using Stata).

Second, you can estimate some of these with regression to see that you don't have a mess somewhere else.

Third, start with a simple model and work up. These models are tricky - it is easy to mess them up if you don't get all the error covariances right.

Fourth, if everything is observed, you should be able to estimate a simultaneous model with reg3. While there are reasons to go to SEM for a simultaneous model with all observed variables (e.g., if the dv's have unusual properties), if everything is continuous, reg3 is normally easier. Actually, I don't think you even have simultaneity here - regress should give you consistent estimates and suest efficient ones.