The Selection of Control Variables in Capital Structure Research With Machine Learning

Abstract

The previous literature on capital structure has produced plenty of potential determinants of leverage over the last decades. However, their research models usually cover only a restricted number of explanatory variables, and many suffer from omitted variable bias. This study contributes to the literature by advocating a sound approach to selecting the control variables for empirical capital structure studies. We applied two linear LASSO inference approaches and the double machine learning (DML) framework to the LASSO, random forest, decision tree, and gradient boosting learners to evaluate the marginal contributions of three proposed determinants; cash holdings, non-debt tax shield, and current ratio. While some studies did not use these variables in their models, others obtained contradictory results. Our findings have revealed that cash holdings, current ratio, and non-debt tax shield are crucial factors that substantially affect the leverage decisions of firms and should be controlled in empirical capital structure studies.