I have a growth curve with time-invariant and time-varying covariates. My model worked well until I tried to analyse gender subgroups separately, but now 'the model may not be identified'. I've noticed that the observed indicators of the growth curve are highly correlated among females, while the correlations were lower among the whole population. Is this likely to prevent model identification?
Size of correlations is unlikely to affect the identification status. It is probably something different in the setup. Check that the number of parameters is the same when you analyze gender groups separately as when you analyzed them as one group - or are you comparing a single-group and a 2-group run?
Yes, you're right, the model seems to be estimating alphas for some latent constructs where alpha was not estimated before.
In the simple growth model with no covariates, the variance on linear and quadratic slope for females was coming out negative, so I constrained these to 0. Which allowed the model to run. Could it be that these constraints are changing other parts of the model, when I make it more complex?
How do I specify that alpha of latent variables (which represent an additional intercept) should not be estimated?