New Strategy for Choosing Initial for the EM Algorithm in Poisson Regression Models

Abstract

The EM algorithm is the standard tool for maximum likelihood estimation in finite mixture models. Its most important drawbacks are the slow convergence, the need for a suitable stopping criterion and the choice of the initial values. In this paper, we focus on the issue of selecting initial values for the EM algorithm in mixture Poisson regression models. A new strategy, aiming at overcoming limitations of other approaches, is proposed and a simulation study comparing its performance with two alternative strategies is carried out. In
models with overlapped components and/or not similar mixing proportions, the new strategy has proven to provide more accurate parameter estimates and to require a fewer number of iterations until the EM algorithm convergence saving computing time.