Avaliação de critérios para a seleção do número de componentes em misturas finitas de normais assimétricas

Abstract

The present work aims to evaluate some information criteria for the selection of models in the context of finite mixtures of skew-normal distributions. The analyzed criteria are the Akaike s Information Criterion - AIC, the Bayesian Information Criterion - BIC and the Efficient Detection Criterion - EDC. The evaluation concerning the performance presented by these criteria was obtained through a simulation study, on which the EM algorithm is required to find the maximum likelihood estimates of for the parameters of the model where the criteria are applied. It was also performed an experiment for the application of the theory developed, modeling a real data set previously analyzed in the specific literature. The results obtained point that, in an asymptotic sense, the three criteria tend to correctly evaluate the number of necessary components, but for small samples the AIC presents inferior performance than BIC or EDC.