The BerHu penalty and the grouped effect

The Huber's criterion is a useful method for robust regression. The adaptive least absolute shrinkage and selection operator (lasso) is a popular technique for simultaneous estimation and variable selection. In the case of small sample size and large covariables numbers, this penalty is not very satisfactory variable selection method. In this paper, we introduce an adaptive reversed version of Huber's criterion as a penalty function. We call this penalty adaptive Berhu penalty. As for elastic net penalty, small coefficients contribute their $\ell_1$ norm to this penalty while larger coefficients cause it to grow quadratically (as ridge regression). We show that the estimator associated with criterion such that ordinary least square or Huber's one combining with adaptive Berhu penalty enjoys the oracle properties. In addition, this procedure encourages a grouping effect. This approach is compared with adaptive elastic net regularization. Extensive simulation studies demonstrate satisfactory finite-sample performance of such procedure. A real example is analyzed for illustration purposes. Keywords : Adaptive Berhu penalty; concomitant scale; elastic net penalty; Huber's criterion; oracle property; robust estimation.