Breakdown properties of location estimates based on halfspace depth and projected outlyingness (Q1208644)

The authors investigate multivariate generalizations of the median, trimmed mean and so-called \(W\)-estimates introduced by \textit{C. F. Mosteller} and \textit{J. W. Tukey} [Data analysis and regression. (1977)]. The estimates are based on a geometric construction related to ``projection pursuit'' by applying the concept of the depth of a point \(x\) in a \(d\)-dimensional data set. All the estimates are affine equivariant and have high breakdown point. This combination of equivariance and robustness is interesting because many of the classical location estimators lack one or both of these properties. The generalization of the median has a breakdown point of at least \(1/(1+d)\) and it can be as high as 1/3 assuming a centrosymmetric distribution. Under very mild conditions the \(W\)-estimates have a breakdown point close to 1/2, even in high dimensions. In contrast, various estimators based on rejecting outliers and taking the mean of the remaining observations (such as iterative ellipsoidal trimming, convex hull peeling) never have a breakdown point exceeding \(1/(1+d)\).