Second-order properties of an extrapolated bootstrap without replacement under weak assumptions (Q1363400)

The paper investigates a variant of the bootstrap, namely one which samples less than the actual sample size (undersampling) and without replacement (WoRB). It shows that the WoRB admits an Edgeworth expansion of the standardized statistic when some assumptions hold. The order of the error is \(O(b_n^{-1/2})\) when one samples \(b_n\) out of \(n\) observations. This is worse than \(O(n^{-1/2})\). But a convex combination of the undersampled WoRB and the usual Gaussian approximation leads to an improvement. This gives second order correct confidence intervals when the sample size is chosen adequately. The results are extended to strong mixing random fields using the idea of moving block bootstrap. The last section gives some simulation results.