I am trying to estimate a quadratic growth model using gsem, and am encountering convergence difficulties with what I think should be a relatively simple model. The data are repeated observations (N = 27,439, range 1/7 obs per subject, mean 5.5 obs per subject) nested within subjects (N = 4,979). I can fit the same model in 'mixed' without any problem, so to be clear the ultimate motivation for using gsem is to extend the model in some other ways that aren't possible in mixed (adding a measurement model/joint modelling of dropout).

Code:
gsem outcome <- timevar timevar_squared c.timevar#RS[idvar] timevar_squared#RS2[idvar]  RI[idvar]
What I think is the equivalent model using 'mixed' converges very quickly:

Code:
mixed outcome timevar timevar_squared || idvar: timevar timevar_squared , cov(uns)
I've tried a number of the estimation problem approaches suggested in the gsem documentation, including:

1) estimating a simpler model first (which converges no problem), and using the parameter estimates as starting values for the full model:

Code:
gsem outcome <- timevar timevar_squared c.timevar#RS[idvar] RI[idvar]
matrix b = e(b)
gsem outcome <- timevar timevar_squared c.timevar#RS[idvar] timevar_squared#RS2[idvar]  RI[idvar], from(b)
2) Reducing the number of integration points to attempt to find better starting values option 'intpoints(3)'
3) Using alternative numerical integration methods, 'intmethod(ghermite)' and 'intmethod(laplace)'
4) adding the 'difficult' option

In all of these methods there is an initial improvement in the likelihood, followed by many iterations of 'not concave' with no improvement in the likelihood.

The only way I've been able to get gsem to converge is to:

1) run gsem with the 'noestimate' option to get a matrix with the appropriate structure for the parameters
2) run the same model in mixed
3) modify the matrix from 'gsem... , noestimate' with the final values from mixed
4) re-run gsem with the final values from mixed as starting values

This works, and gives me back the same parameters from mixed, but obviously isn't a general solution, and won't be very useful for cases where I'm extending the model.

Is there something in this sort of model likely to create problems for gsem consistently? And are there other general solutions I should look to in these situations?