Bayesian Model Averaging with Missing Data
Also known as: BMA with missing data, Bayesian model averaging under missingness, BMA-MI, model-averaged imputation
Bayesian Model Averaging with missing data (BMA-MD) simultaneously addresses two sources of uncertainty: which model best describes the data, and what the unobserved values are. Rather than selecting a single imputed dataset and a single model, the approach averages predictions across the full space of candidate models and plausible completions of the missing values, propagating both sources of uncertainty into every estimate and prediction.
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 BMA with missing data when you have a non-trivial number of candidate models or variable subsets and your dataset has partially observed variables, and you want estimates that honestly propagate both model uncertainty and missingness uncertainty into inference. It is especially valuable in exploratory or variable-selection settings where no single model is clearly correct. Do not use it when the missingness mechanism is MNAR and you cannot model that mechanism explicitly; when the model space is enormous without natural enumeration; or when computational budget prohibits fitting and imputing across many models. For purely confirmatory analysis with a pre-specified model, standard multiple imputation alone may suffice.
Strengths & limitations
- Propagates both model-selection uncertainty and missing-data uncertainty into every posterior summary, avoiding double-confident false precision.
- Produces better-calibrated predictions than single-model analysis, particularly when no single model dominates.
- Naturally incorporates prior information on both model plausibility and the imputation model.
- Avoids ad-hoc variable selection decisions that can inflate false-positive rates.
- Posterior model probabilities provide an interpretable summary of evidence for different model structures.
- Computationally expensive: requires fitting and imputing across potentially many candidate models.
- Posterior model probabilities can be sensitive to prior model probabilities and the prior on parameters within each model.
- When the model space is not enumerable (e.g., very high-dimensional predictors), only approximations such as Occam's Window or MC3 can be used.
- Assumes MAR unless the missingness mechanism is explicitly modeled, which is unverifiable from data alone.
- Interpreting averaged coefficients across models with different variable sets requires care, as a coefficient can be zero in models that exclude the variable.
Frequently asked
How does BMA with missing data differ from running BMA on a singly-imputed dataset?
Single imputation treats the imputed values as if they were true, underestimating uncertainty and biasing model weights. BMA with missing data draws multiple imputations within each model, so both the imputation uncertainty and the model uncertainty are propagated into the final averaged estimate.
What if the number of candidate models is too large to enumerate?
Use stochastic search approaches such as Markov Chain Monte Carlo Model Composition (MC3) or Occam's Window pruning. These explore the high-posterior-probability region of model space without exhaustive enumeration, giving approximate BMA weights that are adequate for most practical purposes.
Can I use BMA with missing data when data are Missing Not At Random?
Standard BMA-MD assumes MAR. Under MNAR you must specify an explicit model for the missingness mechanism (a selection model or a pattern-mixture model) and include it as part of the model space. This is substantially more complex and requires sensitivity analyses over plausible MNAR assumptions.
How are posterior model probabilities affected by the prior on models?
With a uniform prior over models, posterior probabilities are driven entirely by marginal likelihoods. However, the marginal likelihood for each model depends on the prior placed on the model's parameters, so flat or diffuse priors can inadvertently penalise complex models (Bartlett's paradox). Weakly informative parameter priors and a prior sensitivity check are recommended.
What software implements BMA with missing data?
The R package BMA implements core BMA for regression. For integration with missing data, the mice or Amelia packages can be combined with BMA to produce imputation-averaged model weights. Stan and JAGS allow fully Bayesian joint models for parameters, missing values, and model indicators simultaneously.
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 ↗
- Rubin, D. B. (1987). Multiple Imputation for Nonresponse in Surveys. John Wiley & Sons, New York. ISBN: 978-0471655749
How to cite this page
ScholarGate. (2026, June 3). Bayesian Model Averaging with Missing Data. ScholarGate. https://scholargate.app/en/bayesian/bayesian-model-averaging-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.
- Approximate Bayesian Computation with Missing DataBayesian↔ compare
- Bayesian Hierarchical Model with Missing DataBayesian↔ compare
- Bayesian Inference with Missing DataBayesian↔ compare
- Bayesian Model AveragingBayesian↔ compare
- Multiple ImputationStatistics↔ compare
- Sequential Monte Carlo with Missing DataBayesian↔ compare