Error due to sem option from()?

[Kai Leitemo]

I have the BE version of Stata 18 and am experiencing an error code "Matrix ___000003" after using the from option attempting to estimate at 3rd Order CFA using SEM.

Since my 3rd Order CFA does not converge using the default initial values, I am therefore attempting to first to estimate a 1st order CFA version of the model that does converge, holding on to the estimation results using "matrix beta=e(b)" and then estimate the 3rd factor CFA using the from option in SEM.

...from(beta,skip).

After starting to initializing the 3rd factor model estimation, the program stops returning the above error code. What can be the problem?

The CFA model is used to model a 108-item personality questionnaire, so by many standards, it is a large model. There are 108 indicators in the model, 9 1st order factors, 3 2nd factors and 1 3rd order factor. I have 269 observations.

I would be interested in hearing your ideas about what might be wrong.

Thanks.

Kai

[Kai Leitemo]

The correct error term is "matrix __000003 not found"

[Joseph Coveney]

What can be the problem?

I don't know about that problem, because you don't really show enough to say, and because what you do show—matrix beta=e(b) and ...from(beta,skip)—is not consistent with the error message that you claim to get, which is for a temporary matrix name.

But you might have another problem should you ever get that one sorted out, namely, that when you take the nine exogenous latent factors from the first fitted model and use their variance coefficients as starting values in the second, you need to rename them in the matrix stripe in order to account for the fact that they are endogenous latent factors in the latter. See the code following the comment "// Exogenous latent factors will now be endogenous' in this post for guidance.

More advice that you didn't ask for: Stata sometimes has difficulty in fitting a higher-order CFA when the equivalent lower-order CFA with all pairwise covariance terms in lieu of the corresponding latent factors converges nicely. You might want to consider that approach instead.