Second Order Approximations for Slightly Trimmed Sums

Abstract

We investigate the second order asymptotic behavior of trimmed sums $T_n=\frac 1n \sum_{i=\kn+1}^{n-\mn}\xin$, where $\kn$, $\mn$ are sequences of integers, $0\le \kn < n-\mn \le n$, such that $\min(\kn, \mn) \to \infty$, as $\nty$, the $\xin$'s denote the order statistics corresponding to a sample $X_1,...,X_n$ of $n$ i.i.d. random variables. In particular, we focus on the case of slightly trimmed sums with vanishing trimming percentages, i.e. we assume that $\max(\kn,\mn)/n\to 0$, as $\nty$, and heavy tailed distribution $F$, i.e. the common distribution of the observations $F$ is supposed to have an infinite variance. We derive optimal bounds of Berry -- Esseen type of the order $O\bigl(r_n^{-1/2}\bigr)$, $r_n=\min(\kn,\mn)$, for the normal approximation to $T_n$ and, in addition, establish one-term expansions of the Edgeworth type for slightly trimmed sums and their studentized versions. Our results supplement previous work on first order approximations for slightly trimmed sums by Csorgo, Haeusler and Mason (1988) and on second order approximations for (Studentized) trimmed sums with fixed trimming percentages by Gribkova and Helmers (2006, 2007).