Explaining impact of predictors in rankings: an illustrative case of states rankings

ABSTRACT This study presents an approach that can be used to identify important predictors used incalculating performance rankings and gauge their sensitivities. Random Forests is a powerful machine learning tool well known for their predictive powers. It is especially suited to broach the small-n, large-p problem usually found in rankings procedures. However, random forests are unable to shed any insight intohow the examined predictors affect individual entries in the ranked set. A procedure calledLocal Interpretable Model-Agnostic Explanations (LIME) enables decision-makers to discernthe most important individual variables and their relative contributions to the outcome ofeach element in the ranked set. To explain this procedure, we use the 2016 edition of theALEC-Laffer State Rankings data. With the method proposed in this study, a state’s policymakerswould have specific knowledge on how to improve their state’s ranking. This method is ofgeneral applicability to any policy domain.