A concentration approach to sensitivity studies in statistical estimation problems

International audienceIt is shown that the concept of concentration is of potential interest in the sensitivity study of some parameters and related estimators. Basic ideas are introduced for a real parameter >0 together with graphical representations using Lorenz curves of concentration. Examples based on the mean, standard deviation and variance are provided for some classical distributions. This concentration approach is also discussed in relation with influence functions. Special emphasis is given to the average concentration of an estimator which provides a sensitivity measure allowing one to compare several estimators of the same parameter. Properties of this measure are investigated through simulation studies and its practical interest is illustrated by examples based on the trimmed mean and the Winsorized variance