Robust High-dimensional Tuning Free Multiple Testing

  • 2022-11-23 18:20:18
  • Jianqing Fan, Zhipeng Lou, Mengxin Yu
  • 0

Abstract

A stylized feature of high-dimensional data is that many variables have heavytails, and robust statistical inference is critical for valid large-scalestatistical inference. Yet, the existing developments such as Winsorization,Huberization and median of means require the bounded second moments and involvevariable-dependent tuning parameters, which hamper their fidelity inapplications to large-scale problems. To liberate these constraints, this paperrevisits the celebrated Hodges-Lehmann (HL) estimator for estimating locationparameters in both the one- and two-sample problems, from a non-asymptoticperspective. Our study develops Berry-Esseen inequality and Cram\'{e}r typemoderate deviation for the HL estimator based on newly developed non-asymptoticBahadur representation, and builds data-driven confidence intervals via aweighted bootstrap approach. These results allow us to extend the HL estimatorto large-scale studies and propose \emph{tuning-free} and \emph{moment-free}high-dimensional inference procedures for testing global null and forlarge-scale multiple testing with false discovery proportion control. It isconvincingly shown that the resulting tuning-free and moment-free methodscontrol false discovery proportion at a prescribed level. The simulationstudies lend further support to our developed theory.

 

Quick Read (beta)

loading the full paper ...