An application of shrinkage to dealing with extreme values
N. T. Longford

Abstract

Trimming and winsorisation are commonly used devices for reducing the influence of exceptionally large observations in positively skewed data. We explore a flexible version winsorisation based on the general idea of shrinkage. It can be interpreted as a compromise between retaining and truncating an outlying observation. We show that the details of winsorisation should be informed not only by the distribution of the data but also by the target of estimation.
 
  To appear in Statistics in Transition 12


June 2008