A1150
Title: What is in the model? A comparison of variable selection criteria and model search approaches
Authors: Marco Ferreira - Virginia Tech (United States)
Allison Tegge - Virginia Tech (United States)
Shuangshuang Xu - Virginia Tech (United States) [presenting]
Abstract: Variable selection methods are crucial for screening and identifying the most associated regression variables to a dependent variable of interest. In particular, a parsimonious model helps interpretation. Among the plethora of variable selection methods, BIC, AIC, and LASSO are three of the most widely used. The aim is to provide a comprehensive comparison among these three methods through a simulation study. For small numbers of regressors, the AIC and BIC are implemented with an exhaustive search of the model space. For a large number of regressors, the model space is explored using a genetic algorithm. The simulation study considers variable selection in linear models and generalized linear models. The results show that when compared with AIC and LASSO, for both small and large number of regressors, the BIC provides better performance in terms of correct identification rate and false discovery rate, while still having high recall.
