F. Greselin, L. Á. García-Escudero, A. Mayo Íscar, G. McLachlan
Finite mixtures of skew distributions have emerged as an effective tool to model heterogeneous data with asymmetric features. Many proposals appeared in the literature, with different tail behaviors. Aiming at a robust methodology, we consider adding trimming and constraints to the estimation of mixtures of skew normal.
Trimming is a very flexible tool to protect against any form of contamination that could occur during data collection and that may spoil the inferential results. The constrained ML estimation of the covariance matrices reduces spurious solutions and avoids singularities. In this way the applicability of the most basic skew model, the skew normal, has been remarkably widened. We present an illustration of such non-elliptically contoured clustering method and associated algorithm for its implementation. Finally, some issues related to truncated moment estimation along the EM algorithm are presented, and alternative methods to overcome their awkward estimation are discussed.
Palabras clave / Keywords: skewness, clustering, robust estimation, finite mixture
Programado
Sesión invitada SI08 Métodos Estadísticos Robustos y Aplicaciones II (Organizadores: Luis Ángel García Escudero y Agustín Mayo Íscar)
1 de junio de 2018 16:00
Sala Cristal
