Multilevel Bayesian Model Averaging
Also known as: ML-BMA, hierarchical Bayesian model averaging, multilevel BMA, Bayesian model averaging in multilevel models
Multilevel Bayesian model averaging (ML-BMA) extends classical Bayesian model averaging to grouped or hierarchically structured data. Rather than committing to a single multilevel model specification, it computes a weighted average of predictions and parameter estimates across a set of candidate multilevel models, weighting each model by its posterior probability given the data. The result accounts simultaneously for uncertainty in the grouping structure, fixed effects, random effects, and covariate selection.
Key highlights
- Explicitly propagates model-structure uncertainty into final estimates and predictions, preventing over-confident conclusions.
- Naturally handles variable selection and random-effects structure selection within a principled Bayesian framework.
- Weighted ensemble predictions tend to be better calibrated than any single candidate model, especially out-of-sample.
- Posterior model probabilities provide an interpretable summary of relative model support.
- Compatible with standard multilevel model samplers (Stan, JAGS, brms) via bridge sampling or WAIC-based approximations.
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 ML-BMA when data have a clear hierarchical or grouped structure and the correct multilevel model specification is genuinely uncertain — for example, when it is unclear which predictors belong at which level, or whether random slopes are needed. It is especially valuable in small-to-medium samples where choosing incorrectly among plausible models inflates Type I error or biases effect estimates. Do not use it when the model space is so large that marginal likelihoods cannot be computed reliably, when a single well-justified model is strongly preferred on theoretical grounds, or when computational resources are very limited and a single well-fitted multilevel model suffices.
Strengths & limitations
- Explicitly propagates model-structure uncertainty into final estimates and predictions, preventing over-confident conclusions.
- Naturally handles variable selection and random-effects structure selection within a principled Bayesian framework.
- Weighted ensemble predictions tend to be better calibrated than any single candidate model, especially out-of-sample.
- Posterior model probabilities provide an interpretable summary of relative model support.
- Compatible with standard multilevel model samplers (Stan, JAGS, brms) via bridge sampling or WAIC-based approximations.
- Marginal likelihood computation is expensive and sensitive to the choice of priors on within-model parameters; diffuse priors can inadvertently penalise complex models (Bartlett's paradox).
- Requires specifying a manageable candidate model set; if the space grows exponentially with predictors, exhaustive enumeration becomes infeasible and stochastic search heuristics are needed.
- Inference is only as good as the coverage of the candidate model set — if the true model is absent, BMA averages over the wrong space.
- Communicating BMA results to applied audiences is harder than reporting a single regression table.
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 multilevel BMA different from standard BMA?
Standard BMA averages over fixed-effects regression models. Multilevel BMA applies the same logic to models that also contain random effects and group-level structure. The candidate models differ not only in which covariates are included but also in which random effects are present and how the hierarchy is specified, making marginal likelihood computation more demanding.
How are marginal likelihoods computed in practice?
Exact computation is rarely feasible. Common approximations include the Laplace approximation, the Bayesian information criterion (BIC) as a crude proxy, WAIC (widely applicable information criterion), and bridge sampling via MCMC. Bridge sampling is the most accurate and is available through the bridgesampling R package, which interfaces with Stan and brms.
What if I have too many candidate models to enumerate?
When the candidate space is large, stochastic search methods such as Markov chain Monte Carlo model composition (MC3) or reversible-jump MCMC can explore the model space without evaluating every candidate, focusing computation on high-posterior-probability regions.
Should I report individual model results alongside BMA estimates?
Yes. It is good practice to report both the BMA-averaged estimates and the posterior model probabilities, and to show how much the top-ranked individual models differ. This transparency helps readers understand how much model uncertainty actually affects the conclusions.
Can I use information criteria like AIC instead of marginal likelihoods?
AIC-based pseudo-BMA weights are sometimes used as a frequentist approximation but they do not correspond to true Bayesian posterior model probabilities. BIC-based weights are a closer approximation but still depend on asymptotics. For genuine Bayesian model averaging, marginal likelihoods estimated via bridge sampling or WAIC with Bayesian leave-one-out cross-validation are preferred.
Sources
- 1.Hoeting, J. A., Madigan, D., Raftery, A. E. & Volinsky, C. T. (1999). Bayesian model averaging: A tutorial. Statistical Science, 14(4), 382-401.
- 2.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
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Multilevel Bayesian Model Averaging. ScholarGate. https://scholargate.app/bayesian/multilevel-bayesian-model-averaging