Bayesian Coarsened Exact Matching
Bayesian Coarsened Exact Matching Estimator · Also known as: Bayesian CEM, BCEM, Bayesian monotonic imbalance bounding matching
Bayesian Coarsened Exact Matching (Bayesian CEM) combines the coarsening-and-exact-matching framework of Iacus, King, and Porro with Bayesian posterior inference. Covariates are discretised into coarser bins so that treated and control units can be matched exactly within those bins, and Bayesian priors are then placed on the treatment-effect parameters to produce full posterior distributions over the causal estimand rather than a single point estimate.
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 CEM when you have observational data with a binary treatment, several continuous or categorical pre-treatment covariates, and a continuous or binary outcome, and you want both strong confounder control (from the CEM step) and an explicit uncertainty quantification (from the Bayesian step). It is particularly appropriate when sample sizes are modest and frequentist confidence intervals would be unreliable, or when you want to incorporate prior information. It is not appropriate when treatment is not clearly binary, when covariates are high-dimensional and many strata would be empty, or when a large share of units would be discarded by matching, leaving a sample too small to update the posterior meaningfully.
Strengths & limitations
- Bounds imbalance on observed covariates by construction, unlike propensity-score methods that merely balance on average.
- Produces a full posterior distribution over the treatment effect, quantifying uncertainty more completely than a point estimate and confidence interval.
- Allows incorporation of prior substantive knowledge about the effect size or outcome model.
- Robust to model misspecification in the propensity model because matching is covariate-based rather than propensity-based.
- The matching step is transparent and easy to report: matched strata are explicitly defined bins.
- Discards unmatched units, potentially reducing sample size substantially and narrowing external validity.
- Coarsening choices (bin boundaries) can materially affect which units are matched and therefore the posterior; sensitivity analyses are required.
- Bayesian computation (MCMC or variational inference) adds software complexity relative to standard CEM.
- As with all matching methods, it cannot control for unobserved confounders.
- In high-dimensional covariate settings, many strata may be empty and matching rates can be very low.
Frequently asked
How does Bayesian CEM differ from standard CEM?
Standard CEM reports a weighted difference-in-means or a regression estimate on the matched sample using frequentist inference. Bayesian CEM replaces that last step with a Bayesian outcome model, yielding a posterior distribution over the ATT rather than a point estimate with a confidence interval. The matching step is identical in both.
How do I choose the coarsening bins?
Bins should reflect substantive knowledge about where covariate values meaningfully differ (e.g., age decades, income quintiles). Automatic cutpoints based on quantiles are a common default, but sensitivity of the matched sample and the posterior to bin choices should always be reported.
What prior should I use for the treatment effect?
A weakly informative prior such as a normal distribution centred at zero with a scale that reflects plausible effect magnitudes is a reasonable default. If domain knowledge suggests a direction or magnitude, a more informative prior is appropriate. Always run a sensitivity analysis varying the prior to check how much it influences the posterior.
What if many units are discarded by the matching step?
A high discard rate signals poor covariate overlap between treated and control groups. Options include relaxing the coarsening (wider bins), restricting the analysis to the region of common support, or switching to a method that does not require exact stratum membership, such as propensity-score weighting with Bayesian inference.
Can Bayesian CEM handle continuous treatments?
The standard CEM framework is designed for binary treatments. Extending it to continuous treatments requires discretising the treatment variable, which changes the causal estimand and interpretation. For continuous treatments, Bayesian dose-response or Bayesian marginal structural models are more natural choices.
Sources
- Iacus, S. M., King, G., & Porro, G. (2012). Causal Inference without Balance Checking: Coarsened Exact Matching. Political Analysis, 20(1), 1-24. DOI: 10.1093/pan/mpr013 ↗
- Hill, J. L. (2011). Bayesian Nonparametric Modeling for Causal Inference. Journal of Computational and Graphical Statistics, 20(1), 217-240. DOI: 10.1198/jcgs.2010.08162 ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Coarsened Exact Matching Estimator. ScholarGate. https://scholargate.app/en/causal-inference/bayesian-coarsened-exact-matching
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 Matching EstimatorCausal inference↔ compare
- Bayesian Propensity Score MatchingCausal inference↔ compare
- Coarsened Exact MatchingCausal inference↔ compare
- Entropy BalancingCausal inference↔ compare
- Matching EstimatorCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare