Bayesian Hierarchical Model with Missing Data
Also known as: BHM missing data, multilevel Bayesian missing data model, hierarchical Bayesian imputation, Bayesian multilevel model with incomplete data
A Bayesian hierarchical model with missing data treats unobserved values as additional unknowns and samples them jointly with all model parameters from the posterior. The nested structure of the hierarchy borrows strength across groups, while the Bayesian framework naturally propagates uncertainty from missingness through every estimate and prediction.
Key highlights
- Integrates imputation and estimation in one coherent model, avoiding the ad hoc combining rules of multiple imputation.
- Hierarchical partial pooling borrows strength across groups, improving estimates for groups with sparse or highly missing data.
- Full posterior uncertainty naturally propagates through every estimate and prediction, including uncertainty from missingness.
- Handles complex missingness patterns that depend on observed covariates (MAR) without requiring a separate missingness model.
- Flexible: extends straightforwardly to non-normal outcomes, time series, and MNAR sensitivity analyses.
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 this approach when data are nested within groups and missingness is non-trivial, especially when the missing-data fraction varies across groups and complete-case analysis would lose too many observations. It is particularly valuable when missingness is at random (MAR) and the covariates that predict missingness are already in the model. Avoid it when missingness is likely MNAR and you have no credible model for the missingness mechanism, since the posterior will be sensitive to unverifiable assumptions. For very large datasets with minimal missingness, simpler imputation followed by standard multilevel models may be adequate.
Strengths & limitations
- Integrates imputation and estimation in one coherent model, avoiding the ad hoc combining rules of multiple imputation.
- Hierarchical partial pooling borrows strength across groups, improving estimates for groups with sparse or highly missing data.
- Full posterior uncertainty naturally propagates through every estimate and prediction, including uncertainty from missingness.
- Handles complex missingness patterns that depend on observed covariates (MAR) without requiring a separate missingness model.
- Flexible: extends straightforwardly to non-normal outcomes, time series, and MNAR sensitivity analyses.
- Computationally demanding — the joint posterior over parameters and all missing values may require long MCMC chains, especially in large datasets.
- Results under MNAR are sensitive to untestable assumptions about the missingness mechanism.
- Model specification becomes complex when missingness affects multiple variables with different types (binary, continuous, count).
- Requires expertise in prior selection and convergence diagnostics; poor choices lead to unreliable posteriors.
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 does this differ from standard multiple imputation followed by multilevel modeling?
Multiple imputation (MI) separates imputation from analysis and uses Rubin's combining rules to pool estimates across imputed datasets. The Bayesian approach performs both simultaneously in one posterior, avoiding approximate combining rules and naturally propagating all uncertainty. With large samples and simple models the two methods give similar answers; the Bayesian approach is more coherent and more flexible for complex hierarchical structures.
Can I use this method when data are missing not at random (MNAR)?
Yes, but you must specify an explicit model for the missingness mechanism — for example a selection model or pattern-mixture model. MNAR parameters are not identified by the observed data alone, so the posterior is sensitive to prior assumptions. Sensitivity analyses over plausible MNAR models are essential before drawing conclusions.
What MCMC sampler should I use?
Hamiltonian Monte Carlo (HMC/NUTS, implemented in Stan) is generally preferred for continuous parameters because it explores the posterior efficiently and provides reliable convergence diagnostics. Gibbs sampling (JAGS, WinBUGS) is an alternative for conditionally conjugate models. Always run multiple chains and check R-hat values below 1.01 and adequate effective sample size.
How do I check whether my imputation is sensible?
Use posterior predictive checks: draw replicated datasets from the posterior (including imputed values) and compare their distribution to the observed data. Check that imputed values fall in plausible ranges and that patterns across groups are consistent with domain knowledge. Also examine whether results change substantially if complete cases only are used, as a sensitivity check.
Does the model require the missing data fraction to be small?
No. The Bayesian framework handles high missing-data fractions, but with more missing data the posterior will be wider and more prior-dependent. When missingness is very high in some groups, posterior estimates for those groups will rely heavily on the hierarchical pooling from other groups and on the prior, so transparent reporting of this dependence is important.
Sources
- 1.Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A. & Rubin, D. B. (2013). Bayesian Data Analysis (3rd ed.). CRC Press.ISBN 978-1439840955
- 2.Little, R. J. A. & Rubin, D. B. (2002). Statistical Analysis with Missing Data (2nd ed.). Wiley.ISBN 978-0471183860
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Bayesian Hierarchical Model with Missing Data. ScholarGate. https://scholargate.app/bayesian/bayesian-hierarchical-model-with-missing-data