Double (Iterated) Bootstrap
Also known as: iterated bootstrap, nested bootstrap, calibrated bootstrap, Çift Bootstrap (Double / Iterated Bootstrap)
The double bootstrap is a resampling method that calibrates a bootstrap confidence interval with a second, nested layer of bootstrap to bring its actual coverage closer to the nominal level. Introduced by Hall (1986) and Beran (1987), it is especially valuable for small samples and skewed distributions where a single-layer bootstrap under-covers.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use the double bootstrap when you need an accurate confidence interval on a continuous statistic and an ordinary bootstrap is likely to under-cover — typically with small samples (about 20 to a few hundred observations) or asymmetric, skewed distributions. It assumes the observations are independent and that you can afford the O(B²) computation. With at least 20 observations the calibration is dependable; below 10 the second-level calibration becomes unreliable.
Strengths & limitations
- Improves the actual coverage of bootstrap confidence intervals toward the nominal level, especially in small samples and skewed distributions.
- Reduces the level error of the underlying single bootstrap without requiring distributional assumptions such as normality.
- Applies to a wide range of statistics and data structures, since it only needs independent observations.
- Computationally expensive: the nested resampling scales as O(B²), so B is practically capped around 500-1000.
- Below about 20 observations the extra calibration layer is not worth its cost, and a single-layer bootstrap is preferable.
- Requires independent observations; dependent data (e.g. time series) need a block-based variant instead.
Frequently asked
How does the double bootstrap differ from an ordinary bootstrap?
An ordinary bootstrap resamples once to build an interval. The double bootstrap adds a second, nested resampling layer that measures how well those intervals actually cover, then recalibrates the nominal level so the final coverage matches what is claimed.
Why is it so much slower?
Each of the B₁ outer resamples triggers its own batch of inner resamples, so the work scales as O(B²). That is why B is usually kept around 500-1000 rather than the much larger counts a single bootstrap can afford.
When is the extra layer not worth it?
For samples below about 20 observations the calibration gain rarely justifies the O(B²) cost — a single-layer bootstrap is enough. Below 10 observations the calibration itself becomes unreliable and a permutation test is preferable.
Can I use it on time series data?
Not directly. The double bootstrap assumes independent observations. For autocorrelated series you need a block bootstrap that preserves the serial dependence structure.
Sources
- Hall, P. (1986). On the Bootstrap and Confidence Intervals. Annals of Statistics, 14(4), 1431-1452. DOI: 10.1214/aos/1176350168 ↗
- Beran, R. (1987). Prepivoting to Reduce Level Error of Confidence Sets. Biometrika, 74(3), 457-468. DOI: 10.1093/biomet/74.3.457 ↗
How to cite this page
ScholarGate. (2026, June 1). Double (Iterated) Bootstrap. ScholarGate. https://scholargate.app/en/statistics/double-bootstrap
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Bayesian BootstrapStatistics↔ compare
- Block BootstrapStatistics↔ compare
- Bootstrap InferenceStatistics↔ compare
- Permutation TestStatistics↔ compare
- Wild BootstrapStatistics↔ compare