In praise of partially interpretable predictors

Authors

Tri Le, Mercer UniversityFollow
Bertrand S. Clarke, University of Nebraska-LincolnFollow

ORCID IDs

Tri Le https://orcid.org/0000-0001-8454-0220

Date of this Version

2020

Citation

Statistical Analysis and Data Mining: The ASA Data Science Journal. 2020;13:113–133.

Comments

© 2020 The Authors. This is an open access article under the terms of the Creative Commons Attribution License

Abstract

Often there is an uninterpretable model that is statistically as good as, if not better than, a successful interpretable model. Accordingly, if one restricts attention to interpretable models, then one may sacrifice predictive power or other desirable properties. A minimal condition for an interpretable, usually parametric, model to be better than another model is that the first should have smallermean-squared error or integratedmean-squared error.We show through a series of examples that this is often not the case and give the asymptotic forms of a variety of interpretable, partially interpretable, and noninterpretable methods. We find techniques that combine aspects of both interpretability and noninterpretability in models seem to give the best results.