A1600
Title: Comparisons of variable selection methods in mixtures of linear regression models
Authors: Susana Faria - Universidade do Minho (Portugal) [presenting]
Ana Moreira - Universidade do Minho (Portugal)
Abstract: Finite mixture regression models provide a flexible tool for modelling data that arise from a heterogeneous population, where the relationship between the dependent variable and the explanatory variables varies across different subpopulations. In the applications of these models, a large number of explanatory variables is often considered; for this reason, variable selection assumes great relevance for mixture models. However, since all subset selection methods are computationally intensive, more efficient methodologies were developed to overcome this problem, such as methods based on penalty functions. The least absolute shrinkage and selection operator (LASSO) method, the Adaptive Least Absolute Shrinkage and Selection Operator (ALASSO) method and the relaxed least absolute shrinkage and selection operator (RLASSO) method are some examples of these methods. The problem of variable selection in mixtures of linear regression models in the presence of a large number of explanatory variables is analyzed by comparing the performance of LASSO, ALASSO and RLASSO in the selection of explanatory variables, employing the expectation-maximization and classification expectation-maximization algorithms for parameter estimation. An extensive simulation study reveals how different scenarios affect the performance of these algorithms and penalization functions in selecting explanatory variables.
