Modelo de mistura de regressão: uma abordagem bayesiana

Abstract

In the current dissertation, we study the mixture regression models and present two Bayesian methodologies for their estimation. The first one considers the number of components is known and we propose the use of two Bayesian model selection criteria, DIC and EBIC, to identify the number of components. In the other one, we propose a reversible jump algorithm with splitmerge steps that estimates parameters and the number of components. We apply the proposed methodologies and also the EM algorithm, already available in R package, for simulated dataset and for Brazilian educational data, studying the relationship among the Basic Education Development Index and some socioeconomic and demographic data.