Non-Parametric Uncertainty Bias and Variance Estimation via Nested Bootstrapping and Influence Functions

In using limited datasets, modeling the uncertainty via non-parametric methods arguably provides more robust estimators of the unknown value of interest. We propose a novel nested bootstrap method that accounts for the uncertainty from various sources (input data, model, and estimation) more robustly. The nested bootstrap is particularly apt to the more nuanced conditional settings in constructing prediction rules but is easily generalizable. We utilize influence functions to estimate the bias due to input uncertainty and devise a procedure to correct the estimators' bias in a simulation optimization routine. Implementations in the context of feature selection via simulation optimization on two simulated datasets prove a significant improvement in robustness and accuracy.