A Variable Selection in Multiple Linear Regression Models Based on Tabu Search

Abstract

This research has proposed a variable selection method based on the Tabu Search for multiple linear regression models. In this study two objective functions used in the Tabu Search are mean square error (MSE) and the mean absolute error (MAE). The results of Tabu Search are compared with the results obtained by the stepwise regression method based on the hit percentage criterion. The simulations cover both cases, without and with multicollinearity problems. There are six independent variables in full models with four independent variables that influence the dependent variable. The sample sizes are 25 and 100, repeated 500 times for
each situation. Without the multicollinearity problem, the hit percentages of the stepwise regression method and Tabu Search using both objective functions are almost the same. With the multicollinearity problem, the hit percentages of Tabu Search using both objective functions are more than the hit percentage of the stepwise regression method. In particular, the regression coefficients have the wrong sign from simulation by the stepwise regression method.