Multilevel Bootstrap Simulation
Also known as: hierarchical bootstrap, cluster bootstrap, stratified bootstrap for multilevel data, multilevel resampling
Multilevel bootstrap simulation is a resampling technique designed for clustered or hierarchically structured data. It preserves the nested data structure by resampling at each level independently — first drawing clusters (e.g., schools, hospitals), then drawing observations within each sampled cluster — so that bootstrap replicate datasets reflect the same multilevel organisation as the original data.
Key highlights
- Respects the nested data structure by resampling at each level, preserving intraclass correlation patterns.
- Makes no distributional assumptions about the sampling distribution of the estimand.
- Provides valid standard errors and confidence intervals even for complex multilevel statistics with no closed-form expression.
- Straightforward to implement once the level structure is identified.
- Can be combined with any multilevel model or estimator without modifying the model itself.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use multilevel bootstrap simulation when you have clustered or nested data and need standard errors or confidence intervals for an estimand that spans levels — for example, intraclass correlations, random effect variances, or fixed effects from a mixed model — and you do not want to rely on asymptotic normality. It is especially useful with small numbers of clusters, where standard asymptotic approximations can be unreliable. Do not use it as a substitute for a properly specified multilevel model: the bootstrap describes uncertainty around an estimand but does not itself specify the data-generating mechanism. Avoid it when the number of clusters is very small (fewer than 10–15) because the bootstrap distribution over clusters will be too coarse.
Strengths & limitations
- Respects the nested data structure by resampling at each level, preserving intraclass correlation patterns.
- Makes no distributional assumptions about the sampling distribution of the estimand.
- Provides valid standard errors and confidence intervals even for complex multilevel statistics with no closed-form expression.
- Straightforward to implement once the level structure is identified.
- Can be combined with any multilevel model or estimator without modifying the model itself.
- Requires a sufficient number of clusters at the highest level; with very few clusters the bootstrap distribution is discrete and inaccurate.
- Computationally intensive: fitting the model B times can be slow for large datasets or complex models.
- Does not model the data-generating process — it only characterises sampling variability around a point estimate.
- Choice of resampling scheme (which levels to resample, whether to resample residuals vs. cases) can affect results.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How is the multilevel bootstrap different from an ordinary bootstrap?
An ordinary bootstrap resamples rows without regard to group membership, which breaks the cluster structure and misrepresents between-cluster variability. The multilevel bootstrap resamples clusters first and then observations within clusters, so each replicate dataset has the same nested structure as the original.
How many bootstrap replications do I need?
For standard error estimation, B = 1000 is usually adequate. For percentile confidence intervals, B = 2000–5000 is recommended. For bias-corrected and accelerated (BCa) intervals, which require influence values, B = 5000 or more is advisable.
Can I use this method with only a few clusters?
The multilevel bootstrap becomes unreliable when the number of top-level clusters is small (roughly fewer than 15). With only a handful of clusters the bootstrap distribution over clusters is very coarse and intervals will not have the correct coverage. In such cases, consider a parametric bootstrap or Bayesian posterior simulation instead.
Should I resample cases or residuals?
Case resampling (resampling entire observations) is the most common and most robust approach, as it makes no assumptions about the model form. Residual resampling can be more efficient when the model is correctly specified but can fail under misspecification. For multilevel data, case-level resampling at each level is generally preferred.
Is the multilevel bootstrap Bayesian?
No. The multilevel bootstrap is a frequentist resampling method that approximates the sampling distribution of a statistic. It does not require or produce a posterior distribution. However, it can be used alongside Bayesian models as a non-parametric alternative for checking frequentist coverage of Bayesian credible intervals.
Sources
- 1.Efron, B. (1979). Bootstrap methods: Another look at the jackknife. The Annals of Statistics, 7(1), 1–26.
- 2.Davison, A. C. & Hinkley, D. V. (1997). Bootstrap Methods and their Application. Cambridge University Press.ISBN 978-0521574716
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Multilevel Bootstrap Simulation. ScholarGate. https://scholargate.app/bayesian/multilevel-bootstrap-simulation