Comparing four bootstrap methods for stratified three-stage sampling

For a stratified three-stage sampling design with simple random sampling without replacement at each stage, only the Bernoulli bootstrap is currently available as a bootstrap for design-based inference under arbitrary sampling fractions. This article extends three other methods (the mirror-match bootstrap, the rescaling bootstrap, and the without-replacement bootstrap) to the design and conducts simulation study that estimates variances and constructs coverage intervals for a population total and selected quantiles. The without-replacement bootstrap proves the least biased of the four methods when estimating the variances of quantiles. Otherwise, the methods are comparable.