Bayesian Propensity Score Weighting
Also known as: Bayesian PSW, Bayesian IPW, Bayesian inverse probability weighting, Bayesian propensity weighting
Bayesian Propensity Score Weighting estimates causal treatment effects in observational data by combining a Bayesian model for the propensity score with inverse probability weighting. By placing a prior over propensity-score parameters and propagating posterior uncertainty through the weighting step, this approach yields fully probabilistic uncertainty intervals for the average treatment effect, accounting for the uncertainty in both the score model and the outcome.
Key highlights
- Propagates propensity-score estimation uncertainty into the treatment-effect posterior, producing credible intervals that reflect all sources of uncertainty.
- Allows incorporation of substantive prior knowledge about which covariates influence treatment assignment.
- Naturally handles small samples where maximum-likelihood propensity models are unstable or near-separation occurs.
- Yields a full posterior distribution over the ATE, enabling direct probability statements such as P(ATE > 0).
- Compatible with marginal structural models and doubly robust estimators within a unified Bayesian framework.
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 propensity score weighting when you have observational data with a binary (or categorical) treatment, a set of pre-treatment confounders, and a continuous or binary outcome, and you want fully probabilistic uncertainty quantification rather than confidence intervals derived from asymptotic theory. It is especially valuable in small-to-moderate samples where propensity-score estimation uncertainty is non-trivial, when you have substantive prior knowledge about treatment selection, or when you need to propagate uncertainty through a downstream analysis such as a marginal structural model. Do not use it as a substitute for strong covariate balance — Bayesian weighting still requires that all important confounders are measured (no unmeasured confounding). It is computationally demanding and may be excessive when samples are large and the propensity model is well-specified.
Strengths & limitations
- Propagates propensity-score estimation uncertainty into the treatment-effect posterior, producing credible intervals that reflect all sources of uncertainty.
- Allows incorporation of substantive prior knowledge about which covariates influence treatment assignment.
- Naturally handles small samples where maximum-likelihood propensity models are unstable or near-separation occurs.
- Yields a full posterior distribution over the ATE, enabling direct probability statements such as P(ATE > 0).
- Compatible with marginal structural models and doubly robust estimators within a unified Bayesian framework.
- Computationally intensive: MCMC sampling over propensity-score parameters can be slow for large datasets or complex models.
- Prior sensitivity: results may depend on the chosen priors, especially with small samples or weak identification.
- Requires specifying the treatment model explicitly, and misspecification biases the ATE estimate despite Bayesian framing.
- Extreme weights from near-zero or near-one propensity scores remain a practical problem even in Bayesian implementations.
- Interpreting and communicating posterior distributions requires Bayesian literacy that may be absent in applied policy or clinical audiences.
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 PSW differ from standard (frequentist) propensity score weighting?
Standard PSW fits a single logistic model by maximum likelihood, plugs in the predicted probabilities as weights, and constructs bootstrap or sandwich standard errors for the outcome model separately. Bayesian PSW treats the logistic model parameters as random, samples from their posterior, and propagates that uncertainty through the weights into a posterior distribution over the treatment effect — capturing propensity-score estimation uncertainty that frequentist two-stage approaches usually ignore.
Do I still need to check covariate balance after Bayesian weighting?
Yes. Bayesian estimation of the propensity score does not guarantee that the resulting weights produce balance on observed covariates. You should still compute standardised mean differences between weighted treated and control groups and verify that they are close to zero.
What prior should I use for the propensity-score model?
Weakly informative priors — for example, normal(0, 2.5) on log-odds coefficients — are generally recommended. They provide enough regularisation to stabilise estimates in moderate samples without imposing strong beliefs about covariate effects. Completely flat priors can lead to extreme weights when the propensity model is near separation.
Is Bayesian PSW valid when there is unmeasured confounding?
No. Like all propensity-score methods, Bayesian PSW relies on the assumption that treatment assignment is ignorable given the observed covariates. If important confounders are unmeasured, the estimated ATE is biased regardless of the estimation framework. Sensitivity analysis methods (e.g., Bayesian sensitivity analysis for unmeasured confounding) are needed to assess this threat.
When should I prefer Bayesian PSW over doubly robust estimation?
Bayesian PSW is preferable when you want a fully probabilistic treatment-effect posterior and have substantive priors about treatment selection. Doubly robust estimation offers protection against misspecification of either the treatment or outcome model but yields point estimates with asymptotic standard errors. If sample size is large and both models are well-specified, the two approaches will give similar results; Bayesian PSW adds value mainly through its uncertainty quantification and prior incorporation.
Sources
- 1.McCandless, L. C., Gustafson, P., & Austin, P. C. (2009). Bayesian propensity score analysis for observational data. Statistics in Medicine, 28(1), 94–112.
- 2.Saarela, O., Stephens, D. A., Moodie, E. E. M., & Klein, M. B. (2015). On Bayesian estimation of marginal structural models. Biometrics, 71(2), 279–288.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Bayesian Propensity Score Weighting. ScholarGate. https://scholargate.app/causal-inference/bayesian-propensity-score-weighting