Maximum Likelihood Estimation for Mixtures of Distributions
Ludolf Erwin Meester · 1984
This paper describes a simulation study of (small sample) maximum likelihood estimation for mixtures of distributions. Maximum likelihood (ML) leads to a multidimensional maximization problem for which a modified Newton method and the EM algorithm are used. The moment and Kabir method are applied to provide initial points. For twenty-seven different mixture distributions a number of samples is generated by means of a pseudo-random number generator. Frequently the supplied initial points are not admissible. Although an ad hoc method for initial points behaves well, still no estimates are obtained for a number of samples.