Robustification of estimators by Winsorizing on ellipsoids

For many of the more complex models in statistical genetics and measurement error analysis (such as, e.g., variance component models) robust methods are obviously needed, but as yet hardly used. This is partly due to the fact that the parameters of such models often have an indispensable subject-matter interpretation which potential robust substitutes might lack. In this paper some tools are presented to robustify given procedures rather than to construct new ones. The main technical devices are generalized moments and a homotopy argument; they appear to be of interest for any approach to robustness.