Min-Max Bias Robust Regression

Abstract

This paper considers the problem of minimizing the maximum asymptotic bias of regression estimates over $\\varepsilon$-contamination neighborhoods for the joint distribution of the response and carriers. Two classes of estimates are treated: (i) $M$-estimates with bounded function $\\rho$ applied to the scaled residuals, using a very general class of scale estimates, and (ii) bounded influence function type generalized $M$-estimates. Estimates in the first class are obtained as the solution of a minimization problem, while estimates in the second class are specified by an estimating equation. The first class of $M$-estimates is sufficiently general to include both Huber Proposal 2 simultaneous estimates of regression coefficients and residuals scale, and Rousseeuw-Yohai $S$-estimates of regression. It is shown than an $S$-estimate based on a jump-function type $\\rho$ solves the min-max bias problem for the class of $M$-estimates with very general scale. This estimate is obtained by the minimization of the $\\alpha$-quantile of the squared residuals, where $\\alpha = \\alpha(\\varepsilon)$ depends on the fraction of contamination $\\varepsilon$. When $\\varepsilon \\rightarrow 0.5, \\alpha(\\varepsilon) \\rightarrow 0.5$ and the min-max estimator approaches the least median of squared residuals estimator introduced by Rousseeuw. For the bounded influence class of $GM$-estimates, it is shown the "sign" type nonlinearity yields the min-max estimate. This estimate coincides with the minimum gross-error sensitivity $GM$-estimate. For $p = 1$, the optimal $GM$-estimate is optimal among the class of all equivariant regression estimates. The min-max $S$-estimator has a breakdown point which is independent of the number of carriers $p$ and tends to 0.5 as $\\varepsilon$ increases to 0.5, but has a slow rate of convergence. The min-max $GM$-estimate has the usual rate of convergence, but a breakdown point which decreases to zero with increasing $p$. Finally, we compare the min-max biases for both types of estimates, for the case where the nominal model is multivariate normal.