Characterize when fridge outperforms cross-validation

Determine for which covariate values the focused ridge (fridge) approach will outperform ridge regression tuned by cross-validation, thereby establishing whether the performance advantage of individualized tuning can be predicted in advance.

Background

The paper introduces focused ridge regression (fridge), which selects a tuning parameter separately for each target covariate vector by minimizing an estimated mean squared prediction error. This individualized choice is intended to improve predictions for particular patients compared with the single tuning parameter selected by cross-validation. The simulations show that the advantage of focusing is not universal and can disappear when the plug-in estimates used to construct the focused tuning parameter are insufficiently precise. The unresolved issue is therefore to characterize, before applying the method, the covariate values for which fridge is expected to outperform conventional cross-validated ridge regression.