Bayesian Inference with Missing Data
Also known as: Bayesian missing data analysis, Bayesian data augmentation, Bayesian imputation, missing data Bayesian model
Bayesian inference with missing data treats unobserved values as unknown parameters and integrates them out of the posterior distribution. Rather than deleting or ad hoc imputing incomplete records, the method jointly models observed and missing data under an explicit missing-data mechanism, producing fully calibrated posterior uncertainty that honestly reflects what the data cannot tell us.
Key highlights
- Propagates missingness uncertainty into all parameter estimates and predictions through principled marginalisation.
- Handles MAR and MNAR mechanisms explicitly within a single coherent model.
- Avoids bias introduced by listwise deletion or single deterministic imputation.
- Posterior distributions are fully calibrated and provide honest credible intervals even under partial observation.
- Can incorporate informative priors to stabilise estimates when observed data are sparse after missingness.
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 Bayesian inference with missing data when incomplete records cannot be discarded without inducing substantial bias, when a MAR or MNAR mechanism must be modelled and justified, or when you need uncertainty from missingness to propagate fully into downstream inferences rather than being suppressed by a single imputed dataset. It is particularly valuable in longitudinal studies, clinical trials, and surveys with systematic dropout. Avoid it when missingness is genuinely MCAR and the fraction is negligible (complete-case analysis is then unbiased and simpler), when the missing-data model cannot be specified with any credibility, or when computational resources do not permit full MCMC runs on large datasets.
Strengths & limitations
- Propagates missingness uncertainty into all parameter estimates and predictions through principled marginalisation.
- Handles MAR and MNAR mechanisms explicitly within a single coherent model.
- Avoids bias introduced by listwise deletion or single deterministic imputation.
- Posterior distributions are fully calibrated and provide honest credible intervals even under partial observation.
- Can incorporate informative priors to stabilise estimates when observed data are sparse after missingness.
- Requires a credible model for the joint distribution of observed and missing data, which may be hard to specify correctly.
- Under MNAR the missingness mechanism must itself be modelled, and results can be sensitive to unverifiable assumptions.
- Computationally demanding: sampling the joint posterior over parameters and missing values scales poorly with the volume of missing data.
- Convergence diagnostics are more complex because the sampler explores a higher-dimensional space including the latent missing values.
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 Bayesian imputation differ from multiple imputation (MI)?
Multiple imputation draws several completed datasets separately and analyses each, combining results with Rubin's rules — a two-stage approximation. Bayesian inference with missing data integrates imputation and estimation in a single MCMC algorithm, so uncertainty propagation is exact within the model and no combining rules are needed. The Bayesian approach is more coherent but requires fitting one joint model rather than an imputation model and an analysis model separately.
Can I use this approach when the missing-data mechanism is MNAR?
Yes, but you must specify and fit a model for the missingness mechanism itself, linking the probability of being missing to the unobserved values. Because MNAR assumptions are untestable from the observed data, sensitivity analysis over plausible MNAR models is essential. Simply ignoring the mechanism and hoping results are robust is not acceptable under MNAR.
What algorithm is typically used to sample the joint posterior?
Data augmentation (Tanner and Wong 1987) is the classical approach, alternating between imputing missing values and sampling parameters. In modern probabilistic programming systems (Stan, PyMC, BUGS) Hamiltonian Monte Carlo or Gibbs sampling handle the augmented parameter space automatically when the missing data model is declared as part of the model.
How do I assess whether my MCMC has converged when missing values are part of the sampler?
Monitor R-hat and effective sample size for the model parameters of primary interest as well as for a representative sample of the imputed values. Trace plots for the imputed variables should show good mixing. If the fraction of missing data is high the effective sample size can drop sharply, requiring longer chains.
Is Bayesian inference with missing data appropriate for large datasets?
Full MCMC over a joint posterior that includes all missing values can be prohibitively slow for large datasets with high missingness rates. Variational inference or amortised approaches can reduce cost, but may underestimate posterior variance. For very large problems, multiple imputation followed by analysis is often a pragmatic compromise.
Sources
- 1.Little, R. J. A. & Rubin, D. B. (2002). Statistical Analysis with Missing Data (2nd ed.). Wiley-Interscience.ISBN 978-0471183860
- 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). Bayesian Inference with Missing Data. ScholarGate. https://scholargate.app/bayesian/bayesian-inference-with-missing-data