A recipe for robust estimation using pseudo data

In this paper we study the utility of the pseudo data for robust smoothing and its applications. The idea of the pseudo data is to transform the original data to those with bounded errors, to which one can apply the usual fitting methods to get the robust estimator. This motivates a fast computational algorithm for the robust estimator. The main contribution of the paper is to extend the application scope of the pseudo data idea to various least-squares estimators such as the linear regression, kernel smoothing and Kriging, while the previous studies are limited to the smoothing spline estimator. We demonstrate the promising empirical evidences of the pseudo data approach through various experiments under different situations.