Asymptotic normality of Winsorized means (Q1105936): Difference between revisions

Given independent, identically distributed random variables \(X_ 1,...,X_ n\), let \(r_ n\) and \(s_ n\) be specified integers satisfying \(1\leq r_ n, s_ n\leq n/2\). A Winsorized mean is constructed by replacing the \(s_ n\) smallest observations and the \(r_ n\) largest observations by the \(s_ n th\) smallest observation and the \(r_ n th\) largest observation, respectively, before calculating an arithmetic mean. Suppose that \(r_ n,s_ n\to \infty\), \(r_ n/n,s_ n/n\to 0\), and that the underlying distribution function of the \(X_ i\) is ``convex at infinity''. This paper gives a necessary and sufficient condition for the Winsorized mean to be asymptotically normal. Further, it is shown that asymptotic normality can fail when the convexity assumption is dropped, even if the \(X_ i\) are bounded.