CFE 2019: Start Registration
View Submission - CMStatistics
B1323
Title: Tuning and smoothing parameter selection for robust estimation of non-parametric effects Authors:  William Aeberhard - Stevens Institute of Technology (United States) [presenting]
Eva Cantoni - University of Geneva (Switzerland)
Rosalba Radice - Cass Business School (United Kingdom)
Giampiero Marra - University College London (United Kingdom)
Abstract: Deviations from model assumptions are known to throw off any likelihood-based estimation and inference, and hinder penalization schemes meant to ensure some degree of smoothness for non-parametric (additive, non-linear) effects approximated by linear combinations of basis functions. Robust estimation methods are a reliable alternative, but the usual tuning of an efficiency-robustness trade off based on asymptotic covariances is not meaningful any more since the achieved smoothness is generally not comparable between estimation methods. To address this, we propose a median downweighting proportion criterion which is simple, general, and fast to compute. Furthermore, we extend the Fellner-Schall smoothing parameter selection method to robust estimation and compare it to robust versions of the Akaike Information Criterion and Schwarz's Bayesian Information Criterion. Various generalized additive models are used for illustration.