On the bootstrap and the trimmed mean (Q1186780)

The authors show that the bootstrap approximation to the distribution of the Studentized trimmed mean performs better than the normal approximation, in that it is in error by terms of smaller order than \(n^{-1/2}\), where \(n\) denotes sample size. No assumptions are made about symmetry and only mild smoothness conditions are imposed. The method of proof has two parts. First they develop a representation for the Studentized trimmed mean, which enables Edgeworth expansions to be derived under minimal conditions. Then they smooth the bootstrap distribution of the trimmed mean in a differential way, using substantial smoothing near the trimmed quantiles and very little smoothing elsewhere. This enables them to supply enough smoothing to overcome the problem of discreteness of the bootstrap distribution, but not so much that they influence the term of order \(n^{-1/2}\) in the expansion. The same technique may be used to solve other problems, for example, to study bootstrap approximations to the distribution of trimmed variance estimates.