Tabu Search for Variable Selection in Multiple Regression With Higher-Order Terms

• In Event: AERA Roundtable Session 15
  In Roundtable Session: Applications, Demonstrations, and Evaluations of Statistical Models and Methods

Abstract

Variable selection procedures for multiple regression are not typically designed with higher order regression terms in mind. Higher order terms require special attention because hierarchical validity requires that higher order terms be accompanied by all associated lower order terms. To address this, we developed a new exploratory variable selection procedure based on the Tabu search algorithm, called higher order Tabu search (HOTS). In this study, we propose and evaluate the performance of HOTS using simulations and compare its performance to an existing algorithm designed for the same purpose. The results show that HOTS generally performs well under the studied conditions but that there is also room for improvement.

Authors

• Eric Wright, University of Florida

• Walter L. Leite, University of Florida

• Katerina M. Marcoulides, University of Minnesota - Twin Cities

• Corinne Huggins-Manley, University of Florida