I have been running a set of grwoth curve analyses including interactions between latent variables and observed variables, using the Algorithm=Integration command. In one particular set of analyses, the models frequently will not converge. However, when I specify a much greater number of integration points than the default 15, at some point the models will converge.
My question is, what exactly am I doing when I specify more integration points? Is this generally considered a valid method of obtaining convergence? Do more points cause the output to be more suspect in any way? I noticed that some of the output may change slightly with a different # of intergration points.
You are increasing precision when you add more integration points so this is not a bad thing. Your model may be difficult to estimate and therefore need more integration points. There is a brief description of numerical integration in the user's guide.