I am using the newly released beta version (2.0) to run a multivariate twin analysis and have some questions. Due to rather low prevalence (but high number of observations, so have good power anyway) I am having some numerical issues but theses seems to be handled by using the options mvnMaxPointsA/B/C and mvnAbsEps in mxOption. However, I am curious of what these actually do, and cannot find it in the documentation. I am assuming the mvnMaxPointsA/B/C adjust the number of integration points in the numerical integration of the multivariate normal density. But what about mvnAbsEps (or mvnRelEps for that matter), I do not understand how these tolerance conditions works in the numerical integration, for example could I expect any problems if setting them too low, except time that is?
Further, what does the “Function precision” option actually do? All I can find as an explanation is: “a measure of accuracy with which f and c can be computed” but I do not understand the implications of this and cannot find out what f and c is. Is f the logL?
And finally, the obtained SEs are rather small (10-4), but the CIs are broad. I am aware that the CIs are likelihood based and cannot be directly compared with the SE, but they are both measure of precision in the estimates and when the SEs indicates the estimates are rather accurate but the CI gives you another picture it makes you start to wonder. I am thinking we can trust the CI more if the SEs are derived from calculated Hessian, is that correct? If I have a reason to believe Hessian is not valid is there any option that could be used to provide a more valid SEs?