A0788
Title: Bayesian variable selection for skew normal models
Authors: Andriette Bekker - University of Pretoria (South Africa) [presenting]
Mohammad Arashi - Ferdowsi University of Mashhad (Iran)
Janet Van Niekerk - King Abdullah University of Science and Technology (Saudi Arabia)
Arnold van Wyk - University of Pretoria (South Africa)
Abstract: Variable selection is one of the most commonly faced problems in statistical analysis. In the frequentist paradigm, penalized regression methods such as L1 regularization and LASSO are used to induce sparsity in high-dimensional settings. In the Bayesian setting, sparsity can be induced by means of a two-component mixture prior with sufficient probability mass at zero. There has also been a recent development that uses global-local shrinkage priors for high-dimensional Bayesian variable selection. The Dirichlet-Laplace (DL) prior is a popular example of this and has shown promising results compared to existing feature selection methods in the Bayesian framework. Incorporating an asymmetrical component into the variable selection framework is proposed. This is showcased by incorporating a skew-normal random error component into the Dirichlet-Laplace prior to linear regression. A framework for prior selection and hyperparameter tuning of the proposed model is also proposed. The performance of the proposed model is assessed and compared with its symmetrical counterpart in both simulated and real-data examples. It is found to not only perform well but also identify certain non-zero signals due to the inclusion of skewness in the proposed model.
