In praise of partially interpretable predictors

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 smaller mean-squared error or integrated mean-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.