We have run a latent growth curve model with covariates using 5 binary observed variables. The model is fairly straight-forward in that we are specifying an intercept factor and a linear growth slope factor. After the unconditional model was fit, we added gender as a covariate to predict both the intercept and slope factors. This model converged and makes sense. Now we removed gender as a covariate and add a binary predictor representing ethnicity (0,1). This model does not converge in the 1000 iterations that MPlus used. This ethnicity variable has roughly an equal split between Caucasian and minority levels of this binary covariate. I was surprised that the model did not converge in light of how the previous analysis with gender converged nicely. Could this just be a case of having to submit better starting values for the relation between the ethnicity covariate and the intercept and slope factors, respectively.
Starting values are usually not needed so I doubt that this is the case here. Is the ethnicity variable also binary? It sounds like it may not be. If there are more than two categories of ethnicity, you need to create a set of dummy variables to represent the categories. You need k-1 dummy variables to represent k categories. If you can't solve the problem, you should send your input/output, data, and license number to email@example.com.