B0787
Title: Parsimonious mixture of multivariate mean-mixture of normal distributions
Authors: Mehrdad Naderi - National Chung Hsing University (Taiwan) [presenting]
Andriette Bekker - University of Pretoria (South Africa)
Abstract: Heterogeneity occurs in various problems of multivariate analysis where the data comes from different latent groups. The finite mixture (FM) model is a model-based statistical tool commonly exploited to identify these unobserved groups. Despite the widespread use and attractive properties of the normal distribution, practical researchers believe that the traditional normally-based FM (FM-N) model might not achieve robust inference when asymmetric features exist in data. A mixture of multivariate mean-mixture of normal (FM-MMN) distributions is postulated to address this potential issue. In addition to the parameters of the FM-N model, the proposed FM model has two vector/scaler parameters, in each component, for controlling skewness and mild heavy tails. Maximum likelihood parameter estimates are carried out by implementing an expectation-maximization (EM) type algorithm. The parsimony version of the FM-MMN distributions is also introduced by employing an eigenvalue decomposition of the component covariance matrices. Finally, the utility of the proposed methodology is illustrated by conducting a simulation study and analyzing real data examples.
