TABLE OF CONTENTS The concept of bootstrap resampling has been well established over the last two decades. As indicated in the introduction, the bootstrap has found a wide range of applications supporting and in some cases replacing traditional techniques which are usually based on asymptotic approximations. This does not mean that the development of the bootstrap theory is complete. There are several interesting derivatives of the classical non-parametric bootstrap.
The balanced bootstrap has been suggested by Davison et al. (1986). In a balanced bootstrap, also known as Latin hypercube resampling (Hall, 2001), the data is resampled in such a way that each of the data points appears the same number of times in the collection of resamples. This reduces the variability of bootstrap resampling, ensuring that the total mean of bootstrap resamples equals the original sample mean.
An alternative to the balanced bootstrap has been proposed by Rao et al. (1997). They note that a bootstrap resample is not equally informative as the original sample, and they propose to resample sequentially with replacement until all data in the original sample is used. This leads to resamples of random lengths.
Several methods have been proposed to minimise the computational cost of bootstrap resampling (Efron, 1990; Karlsson and L thgren, 2000) and efficient parallel implementation of the bootstrap on a hypercube computer was proposed by Xu and Shiue (1991).
