Robust estimation in the linear model with asymmetric error distributions (Q1082742)

This article considers robust estimation of a location parameter and robust estimation of a linear regression model when the observations and error terms, respectively, are independent and identically distributed random variables, each with distribution function F. As in Huber's theory of robust estimation F is an unknown member of a specified class of distribution functions \({\mathcal F}.\) Two particular classes of distribution functions are considered, namely \({\mathcal F}_{a_ 0,\epsilon}\) and \({\mathcal F}_{[-a_ 0,b_ 0]}\). F is in \({\mathcal F}_{a_ 0,\epsilon}\) if it is governed on the set \([-a_ 0,a_ 0]\) by the standard normal distribution function contaminated, with probability \(\epsilon\), by an unknown distribution which is symmetric about 0; outside the interval \([-a_ 0,a_ 0]\), F is completely unknown. The parameters \(a_ 0\) \((a_ 0>0)\) and \(\epsilon\) \((0<\epsilon <1)\) are assumed to be known. The class \({\mathcal F}_{[-a_ 0,b_ 0]}\) consists of all distributions which on a fixed set \([-a_ 0,b_ 0]\) are governed by a known distribution, outside the interval \([- a_ 0,b_ 0]\) the distribution is completely arbitrary. In this article a sequence of M-estimators for the location problem is found which is optimal (in the sense of minimizing the asymptotic mean squared error as F ranges over \({\mathcal F}_{a_ 0,\epsilon})\) within the class of all M-estimators. This result is then generalized to the problem of estimating a linear regression model. Finally, the optimal M- estimators among the class of consistent and asymptotically normal M- estimators of the linear model parameters are found under the error distribution models \({\mathcal F}_{a_ 0,\epsilon}\) and \({\mathcal F}_{[- a_ 0,b_ 0]}.\) For the model \({\mathcal F}_{a_ 0,\epsilon}\) the strong assumption that the design matrix is constructed by repeating p fixed linearly independent rows as \(n\to \infty\) is made. The authors suggest that the same results can be achieved under weaker conditions on the design matrix, but with some additional complications in the derivations. However, the authors argue that this approach would result in a decrease in the small-sample efficiency.