Hierarchical Bayesian Model Averaging
Also known as: HBMA, hierarchical BMA, multilevel Bayesian model averaging, Bayesian model averaging in hierarchical models
Hierarchical Bayesian model averaging (HBMA) combines Bayesian model averaging with hierarchical model structure, averaging posterior quantities over a set of candidate models weighted by each model's posterior probability. Rather than selecting a single best model, HBMA propagates model uncertainty through a hierarchical framework, producing predictions and parameter estimates that honestly reflect uncertainty about which model is correct.
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 HBMA when the data have a clustered or multilevel structure and several plausible hierarchical model specifications exist — for example, different covariate sets or random-effects formulations. It is especially valuable when the sample size within groups is small, because both hierarchical pooling and model averaging help stabilise estimates. Also use it when downstream decisions depend on predictions and the cost of acting on the wrong model is high. Do not use HBMA when the model space is extremely large (the number of candidate models grows exponentially with covariates), when marginal likelihoods are computationally intractable and no reliable approximation is available, or when the data are simple and flat, making hierarchical structure unnecessary.
Strengths & limitations
- Propagates model-selection uncertainty into every prediction and inference, avoiding overconfident conclusions from a single chosen model.
- Combines the bias-reduction benefits of hierarchical pooling with the robustness of model averaging.
- Posterior model probabilities provide an interpretable ranking of competing model structures.
- Naturally handles clustered, grouped, or multilevel data through the hierarchical component.
- Reduces sensitivity to arbitrary modelling choices such as covariate selection or random-effects specification.
- Marginal likelihoods are often analytically intractable and must be approximated (e.g., via BIC, Laplace approximation, or bridge sampling), introducing approximation error.
- Computational cost scales with the number of candidate models; exhaustive averaging over large model spaces requires specialised strategies such as MC3 or stochastic search.
- Results depend on the prior over models; vague or poorly justified model priors can inflate weight on overly complex models.
- Interpretation of averaged coefficients can be difficult when the same variable appears in some models but not others (inclusion probability must be reported).
Frequently asked
How is HBMA different from just picking the best hierarchical model by BIC?
BIC-based selection picks one model and discards all others, ignoring the uncertainty about whether that model is correct. HBMA retains all candidate models weighted by their posterior probability, so predictions reflect both parameter uncertainty within each model and uncertainty about which model is right. When models are nearly equally supported, averaging substantially widens credible intervals compared to selection.
What is a variable's inclusion probability in HBMA?
The inclusion probability of a variable is the sum of posterior model probabilities over all models that include that variable. A high inclusion probability (e.g., above 0.5 or 0.9) indicates the data strongly support the variable's presence regardless of the other modelling choices, while a low inclusion probability suggests the variable's contribution is weak or redundant.
Can I use HBMA with MCMC if marginal likelihoods are intractable?
Yes. Markov chain Monte Carlo model composition (MC3) or reversible-jump MCMC allows the sampler to move between models as well as between parameter values. The proportion of MCMC iterations spent in each model estimates its posterior probability. This avoids explicit marginal likelihood computation but requires care in designing the between-model moves.
How large can the candidate model set be?
With p predictors the full model space has 2^p candidates, which becomes intractable beyond roughly p = 25 for exhaustive enumeration. Stochastic search, MC3, or focused candidate sets (e.g., models defined by domain theory) are the standard remedies. The hierarchical component compounds complexity if different random-effects structures are also being averaged.
When do all model weights concentrate on one model?
With a very large sample the marginal likelihood becomes highly discriminating and the posterior probability of the best model approaches 1, reducing HBMA to ordinary model selection. In this regime the gain from averaging is small, but the procedure remains valid and conservative.
Sources
- Hoeting, J. A., Madigan, D., Raftery, A. E., & Volinsky, C. T. (1999). Bayesian model averaging: A tutorial. Statistical Science, 14(4), 382–417. link ↗
- Fragoso, T. M., Bertoli, W., & Louzada, F. (2018). Bayesian model averaging: A systematic review and conceptual classification. International Statistical Review, 86(1), 1–28. DOI: 10.1111/insr.12243 ↗
How to cite this page
ScholarGate. (2026, June 3). Hierarchical Bayesian Model Averaging. ScholarGate. https://scholargate.app/en/bayesian/hierarchical-bayesian-model-averaging
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 Information CriterionModel Evaluation↔ compare
- Bayesian Model AveragingBayesian↔ compare
- Bayesian RegressionBayesian↔ compare
- Hierarchical Bayesian InferenceBayesian↔ compare
- Hierarchical Markov Chain Monte CarloBayesian↔ compare