Robust lasso for model selection and aplications

Estimation of model parameters and selection of variables in the model are important issues in regression analysis. In recent years, shrinkage estimators, which can simultaneously make parameter estimation and selection of variables in the model, come to the fore. These estimators are grouped according to their specific characteristics. In this study, Least Absolute Shrinkage and Selection Operator-LASSO estimator, which is one of the shrinkage estimators, is discussed. Since LASSO does not have oracle feature, the adaptive LASSO (ALASSO) estimator has been proposed in the literature. However, these estimators are sensitive to the outliers. In this context, robust ALASSO estimators are developed in the literature based on M-estimators. In this study, robust ALASSO estimators are considered using the ρ and ψ functions of Cauchy, Andrews, Fair, Welsch, Logistics and Talwar as an alternative to the robust ALASSO estimators based on the ρ and ψ functions of the Huber and Tukey existing in the literature. The performances of these estimators are compared via a simulation study. For comparing the performance of the estimators, Mean Squared Error (MSE) and Mean Squared Prediction Error (MSPE) criteria are used. Furthermore, model selection accuracies of the methods are measured. In addition to the simulation study, two different real data, one of is available in the literature and the other is compiled by the researcher, are applied.