Am 14.01.2021 um 02:45 schrieb Reynaldo Senra: The MG estimator as such is just a simple average of OLS regressions, so there cannot be any numerical problems, nothing needs to be maximized iteratively. However, we were talking only of using the MG values to provide different initial parameter guesses for the PMG routine. The local maximum problem can happen for any way of choosing initial values, in principle. That's why we are thinking about providing an option to choose between different ways of initialization. First, as explained above, the MG itself can never be stuck at a local max. So that's the good news. But ML (= max-likelihood!) presupposes that you are able to find the global max, otherwise it's not a well-defined construct. So the short answer is no. In practice, however, whether or not the PMG routine did find the global max is something we can never know for sure. This is a general problem of numerical maximization procedures in all software, but of course some algorithms are better than others. In practice, we just take the converged values and hope they are really the best. cheers sven