Methods

This page explains when to use each method.

IID methods

BCa (method="bca")

Best default when you want strong coverage properties for a scalar statistic.

Percentile (method="percentile")

Good when you want a simple interval and your bootstrap distribution is already informative.

Basic (method="basic")

Useful as a classic alternative to percentile intervals.

Studentized (method="studentized")

Useful when you need bootstrap-t intervals and can afford higher compute cost.

Bayesian (method="bayesian")

Useful for Bayesian bootstrap style uncertainty without assuming a parametric model.

Poisson / Bernoulli / Subsampling

Useful for weighted or subsampling-based workflows, especially under heavy tails or sampling constraints.

Time-series methods

Moving Block Bootstrap (mbb)

Preserves local autocorrelation by resampling contiguous blocks.

Circular Block Bootstrap (cbb)

Like MBB, but wraps around the series to reduce edge effects.

Stationary Bootstrap (stationary)

Uses random block lengths and is often a strong default for stationary dependent data.

Tapered Block Bootstrap (tapered)

Reduces block-boundary artifacts with tapering windows.

Sieve Bootstrap (sieve)

Fits an AR approximation and resamples residual-driven trajectories. Use it when an AR representation is reasonable.

Wild Bootstrap (wild)

Useful for heteroscedastic residual structures.

Hierarchical methods

Cluster Bootstrap (cluster)

Resample entire clusters, not rows, when observations within a cluster are dependent.

Stratified Bootstrap (strata)

Use when the sampling design or inference target is stratified.

Practical defaults

• General-purpose scalar statistic: bca
• Model metric on iid holdout set: bca or percentile
• Financial / dependent series: stationary or mbb
• Panel / user-session data: cluster
• Survey-like grouped data: strata