On approximate robust confidence distributions

TL;DR

This paper explores the development of approximate robust confidence distributions using unbiased M-estimating functions, enhancing frequentist inference robustness when the model is only approximately correct or data contains outliers.

Contribution

It introduces methods to derive robust confidence distributions from asymptotic theory and simulation, extending existing results for robust scoring rules.

Findings

Robust confidence distributions can be derived from asymptotic theory.

Simulation methods effectively produce robust confidence distributions.

Application to non-inferiority testing demonstrates practical utility.

Abstract

A confidence distribution is a complete tool for making frequentist inference for a parameter of interest $\psi$ based on an assumed parametric model. Indeed, it allows to reach point estimates, to assess their precision, to set up tests along with measures of evidence for statements of the type "$\psi > \psi_0$" or "$\psi_1 \leq \psi \leq \psi_2$", to derive confidence intervals, comparing the parameter of interest with other parameters from other studies, etc. The aim of this contribution is to discuss robust confidence distributions derived from unbiased $M-$estimating functions, which provide robust inference for $\psi$ when the assumed distribution is just an approximate parametric model or in the presence of deviant values in the observed data. Paralleling likelihood-based results and extending results available for robust scoring rules, we first illustrate how robust confidence distributions can be derived from the asymptotic theory of robust pivotal quantities. Then, we discuss the derivation of robust confidence distributions via simulation methods. An application and a simulation study are illustrated in the context of non-inferiority testing, in which null hypotheses of the form $H_0: \psi \leq \psi_0$ are of interest.