A0455
Title: Resistant inference for complex and large models
Authors: Maria-Pia Victoria-Feser - University of Geneva (Switzerland) [presenting]
Yuming Zhang - University of Geneva (Switzerland)
Abstract: At this moment in time, data not only come in huge quantities, but they also come with features that are not desirable. These include outliers (that are typically hard to detect), missing data, selection bias, measurement errors, and so on. Although there exists an abundance of (easily accessible) statistical and computational methods, these methods tend to address these features separately. Among others, one potential reason for this situation is that accounting for these data features simultaneously can encounter enormous hindrances in the computational aspects. We propose an alternative robustness framework that allows the construction of resistant estimators and associated inferential procedures that have desirable finite sample properties and are computationally tractable with complex and large models. This framework furthermore allows including additional data features such as informative missingness, selection bias and/or measurement errors, with a small additional price in terms of computational costs. More specifically, we employ a two-step approach consisting of considering first a naive estimator that is easy to compute, and derive from it, using a simulation-based approach, a final estimator with desirable asymptotic and finite sample properties. We apply the proposed methodology to GLM and MLM with censored or misclassified data and with outliers.
