Bayesian Doubly Robust Estimation
Bayesian Doubly Robust Estimation of Average Treatment Effects · Also known as: Bayesian DR, Bayesian AIPW, Bayesian augmented inverse probability weighting, Bayesian semiparametric causal estimation
Bayesian Doubly Robust Estimation combines the classical doubly robust (DR) augmented inverse probability weighting framework with Bayesian inference. It simultaneously models the propensity score and the outcome regression, placing prior distributions over both, and derives a posterior distribution over the average treatment effect that remains consistent even if one of the two component models is misspecified.
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 Bayesian Doubly Robust Estimation when you have observational data and want causal inference that is both robust to model misspecification and fully Bayesian in its uncertainty quantification. It is well-suited when covariates are numerous and informative priors or regularisation are desirable, or when transparent propagation of modelling uncertainty is required (e.g., health-technology assessment, clinical epidemiology). It is not the right choice when the sample is very small (fewer than ~50 observations), since the two-model approach demands enough data to estimate both sub-models; nor is it needed when a fully randomised experiment already balances covariates, making propensity-score adjustment unnecessary.
Strengths & limitations
- Doubly robust: consistent for the average treatment effect even when one of the two component models is misspecified, providing a safety net absent in single-model approaches.
- Full posterior distribution over the treatment effect propagates uncertainty from both sub-models, yielding credible intervals that are more honest than frequentist sandwich standard errors in small or moderate samples.
- Accommodates informative prior knowledge about confounders or effect modifiers, which can substantially sharpen inference when such knowledge is available.
- Naturally handles high-dimensional covariate spaces through shrinkage priors (Horseshoe, LASSO-type), without requiring a separate variable-selection step.
- The posterior predictive framework simplifies sensitivity analyses: changing priors or component models produces an updated posterior, making uncertainty exploration coherent and systematic.
- Computationally demanding: fitting two Bayesian models and combining them via MCMC is substantially slower than frequentist doubly robust estimation.
- Choice of priors for both the propensity model and the outcome regression introduces additional researcher degrees of freedom; poorly chosen priors can distort inference, especially in small samples.
- The double-robustness guarantee is asymptotic: in finite samples, severe misspecification of one model still degrades performance, so it is not a license to ignore model diagnostics.
- Software support is less mature than for frequentist DR estimation; implementation typically requires Stan, JAGS, or custom code rather than off-the-shelf packages.
- Interpreting a posterior distribution over the causal estimand requires familiarity with Bayesian reporting conventions, which can hinder communication with audiences trained only in frequentist frameworks.
Frequently asked
What makes this estimator 'doubly robust'?
The DR estimator contains two working models: one for the treatment propensity and one for the outcome. If either model is correctly specified, the estimator of the average treatment effect is consistent. This means a researcher has two chances to get a model right instead of one, providing protection that single-model methods like propensity-score weighting alone or outcome regression alone cannot offer.
How does the Bayesian version differ from the frequentist doubly robust estimator?
In the frequentist version, both sub-models are fitted by maximum likelihood and the DR estimate is a single number with a sandwich standard error. In the Bayesian version, both sub-models have prior distributions over their parameters, posterior draws propagate through the DR formula, and the final output is a full posterior distribution of the treatment effect — capturing uncertainty from both models jointly.
Do I need both models to be correctly specified?
No — that is the defining advantage. Only one needs to be correct for consistent estimation. However, finite-sample performance degrades when one model is severely wrong, so you should still invest in careful specification and model-checking for both components.
Which software can implement Bayesian doubly robust estimation?
Stan (via RStan or PyStan) and JAGS are the primary platforms, as they support custom Bayesian model code for both sub-models. The R packages brms or rstanarm can fit each component model separately; combining them into a DR estimator requires custom post-processing of posterior draws.
When should I prefer this over standard doubly robust estimation?
Prefer the Bayesian version when you have genuine prior knowledge to incorporate, when sample sizes are moderate and honest uncertainty quantification is critical, or when you intend to perform systematic sensitivity analyses by varying priors. For large samples where asymptotic frequentist inference is reliable and computation is a concern, the standard frequentist DR estimator may be sufficient.
Sources
- Bang, H., & Robins, J. M. (2005). Doubly robust estimation in missing data and causal inference models. Biometrics, 61(4), 962-973. DOI: 10.1111/j.1541-0420.2005.00377.x ↗
- Scharfstein, D., Nabi, R., Kennedy, E. H., Huang, M.-Y., Bonvini, M., & Smid, M. (2021). Semiparametric sensitivity analysis: Unmeasured confounding in observational studies. arXiv:1910.14694. link ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Doubly Robust Estimation of Average Treatment Effects. ScholarGate. https://scholargate.app/en/causal-inference/bayesian-doubly-robust-estimation
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 Causal Impact AnalysisCausal inference↔ compare
- Bayesian Propensity Score MatchingCausal inference↔ compare
- Doubly Robust EstimationCausal inference↔ compare
- Inverse Probability WeightingCausal inference↔ compare
- Marginal Structural ModelCausal inference↔ compare