Bolstered error estimator with feature selection

Classification and error estimation are fundamental problems in genomic applications which are typically characterized by large numbers of variables and small numbers of samples. A previously proposed bolstered error estimator was found to work well in the small-sample settings with modest numbers of features not requiring feature selection. In this simulation study, we have improved the method for estimation of the bolstering kernels, which leads to an improved bolstered error estimator that has significantly reduced root mean square error compared to widely-accepted cross-validation error estimator, and performed well over a range of models and model complexities.