Bootstrap Simulation with Missing Data
Bootstrap Simulation with Missing Data Handling · Also known as: bootstrap with missing data, bootstrap imputation simulation, resampling under missingness, bootstrap MI
Bootstrap simulation with missing data combines resampling-based variance estimation with principled handling of incomplete observations. Rather than deleting cases or assuming complete data, the method integrates imputation or weighting directly into the bootstrap loop, propagating the additional uncertainty due to missingness into the final standard errors and confidence intervals.
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 this method when you need valid standard errors and confidence intervals in the presence of missing data and either (i) the analytic formula for standard errors is unavailable or unreliable, or (ii) you want a single coherent resampling framework that handles both sampling and imputation uncertainty. It is especially suitable for complex estimators (regression coefficients, mediation effects, quantiles) with MAR data. Avoid it when data are MNAR and no sensitivity model is specified — no resampling method can rescue inference from non-ignorable missingness without additional assumptions. Also avoid when the fraction of missing data is very high (>50%) and the imputation model is poorly specified, as bootstrap estimates can be highly variable.
Strengths & limitations
- Propagates imputation uncertainty into confidence intervals without requiring Rubin's pooling rules as a separate step.
- Works with any estimator, including non-smooth or non-parametric statistics for which analytic standard errors do not exist.
- Flexible choice of imputation engine within the loop: MICE, random forest imputation, or predictive mean matching.
- BCa intervals correct for bias and skewness in the bootstrap distribution, improving coverage.
- Computationally straightforward to implement in standard software (R mice + boot, Python scikit-learn).
- Computationally intensive: B replicates each require a full imputation run, multiplying cost by B.
- Bootstrap variance can underestimate true variance when the missing fraction is high and the imputation model is misspecified.
- Does not rescue inference under MNAR without a correctly specified sensitivity model.
- Requires enough complete-data observations to fit a credible imputation model within each bootstrap sample.
Frequently asked
Should I impute before or inside the bootstrap loop?
Always inside the loop. Imputing once on the full dataset and then bootstrapping treats imputed values as known, artificially reducing variance. Imputing anew within each replicate correctly reflects the additional uncertainty due to missing data.
How many bootstrap replicates do I need?
At least 1000 replicates for standard error estimates and at least 2000 for tail-sensitive intervals (95% percentile or BCa). With expensive imputation models, 500 replicates may be a pragmatic lower bound, but results should be checked for stability.
Is this the same as multiple imputation with Rubin's rules?
They address the same problem but differently. Rubin's rules combine m imputed datasets analytically and require the imputation model to be compatible (proper). The bootstrap-inside-imputation approach estimates total variance non-parametrically and is less sensitive to propriety conditions, making it more robust for complex estimators.
Can I use this when data are MNAR?
Not without an explicit sensitivity model for the missingness mechanism. MNAR means the probability of being missing depends on the unobserved value itself, so no resampling scheme can recover unbiased inference without additional assumptions. Pattern-mixture or selection models should be used alongside bootstrap variance estimation.
What software implements this?
In R, the mice package combined with boot provides a clean workflow; the bootImpute package implements the Bartlett-Hughes bootstrap-MI integrator directly. In Python, sklearn.impute together with a custom bootstrap loop achieves the same result.
Sources
- Efron, B. & Tibshirani, R. J. (1993). An Introduction to the Bootstrap. Chapman and Hall/CRC. ISBN: 978-0412042317
- Little, R. J. A. & Rubin, D. B. (2019). Statistical Analysis with Missing Data (3rd ed.). Wiley. ISBN: 978-0470526798
How to cite this page
ScholarGate. (2026, June 3). Bootstrap Simulation with Missing Data Handling. ScholarGate. https://scholargate.app/en/bayesian/bootstrap-simulation-with-missing-data
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 Inference with Missing DataBayesian↔ compare
- Gibbs Sampling with Missing DataBayesian↔ compare
- Monte Carlo Simulation with Missing DataBayesian↔ compare
- Multiple ImputationStatistics↔ compare
- Sequential Monte Carlo with Missing DataBayesian↔ compare