Machine Learning-Augmented Coarsened Exact Matching (ML-CEM)
Machine Learning-Augmented Coarsened Exact Matching Estimator · Also known as: ML-augmented CEM, ML-CEM, automated coarsened exact matching, ML-assisted CEM
Machine Learning-Augmented Coarsened Exact Matching extends Coarsened Exact Matching (Iacus, King & Porro, 2012) by using supervised machine learning to automate and optimise the coarsening step — the discretisation of continuous covariates into bins — rather than relying on researcher-specified cutpoints. This reduces both ad hoc subjectivity in coarsening decisions and residual imbalance, while preserving CEM's core logic of exact matching within coarsened strata.
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 ML-augmented CEM when you have observational data with several continuous covariates, lack strong domain knowledge for choosing coarsening cutpoints, and want a more data-adaptive, replicable matching design. It is particularly valuable when covariate distributions are skewed, multimodal, or have complex non-linear relationships with the outcome. It is not appropriate when the sample is very small (fewer than ~80 units), when interpretability of the matching rule must be fully transparent to non-technical audiences, or when exact treatment-covariate overlap is already clear from domain knowledge. Do not use it as a substitute for addressing fundamental unconfoundedness violations — no matching method eliminates unmeasured confounding.
Strengths & limitations
- Automates coarsening cutpoint selection, removing a key source of researcher-degree-of-freedom bias in standard CEM.
- Can improve covariate balance relative to manually specified CEM when covariate-outcome relationships are non-linear or high-dimensional.
- Inherits CEM's transparency: matched strata are discrete and auditable, unlike propensity-score matching on a continuous score.
- Flexible choice of ML learner allows adaptation to the data structure (trees for interactions, LASSO for sparsity, etc.).
- Reduces matched-sample imbalance while retaining the simplicity of exact-within-stratum matching.
- ML-selected coarsening may sacrifice interpretability — the reasons for bin boundaries may not be easily communicated to practitioners.
- When many covariates are used, the curse of dimensionality can still leave large fractions of units unmatched after pruning.
- The choice of ML algorithm introduces a new tuning-parameter decision (tree depth, regularisation strength) that requires justification.
- Like all matching methods, ML-CEM cannot adjust for unmeasured confounders; unconfoundedness must be plausibly satisfied.
- Bootstrapping or other resampling for standard errors can be computationally demanding with large datasets.
Frequently asked
How does ML-augmented CEM differ from standard propensity score matching?
Standard propensity score matching collapses all covariates into a single estimated probability score and matches on that scalar. CEM (and its ML variant) instead coarsens each covariate separately and requires exact agreement across all coarsened values within a stratum. This makes covariate balance more transparent and avoids the risk of exact balance on the score masking imbalance on individual covariates.
Which ML algorithm should I use for the coarsening step?
Regression trees or random forests are natural choices because their split points directly translate into coarsening cutpoints. LASSO or ridge regression can be used to build a prognostic score whose quantiles define bins. The choice should be guided by your data size and the interpretability required; cross-validate to avoid overfitting.
What happens to units that cannot be matched?
They are pruned. ML-CEM, like standard CEM, discards treated units with no control counterpart in their stratum and vice versa. Always report the fraction pruned and assess whether the matched sample remains representative of your target population.
Can I use ML-CEM with panel data?
Yes. Covariates measured at baseline can be coarsened using ML, and the matched strata can then be used in a difference-in-differences or fixed-effects estimator applied to the matched panel, combining the benefits of matching-based balance and panel-based trend control.
Does ML-CEM eliminate unmeasured confounding?
No. Like all matching estimators, ML-CEM requires the unconfoundedness (ignorability) assumption: conditional on observed covariates, treatment assignment is as good as random. If important confounders are unobserved, no matching method — including ML-augmented variants — can remove the resulting bias.
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 ↗
- Imai, K., & Ratkovic, M. (2014). Covariate balancing propensity score. Journal of the Royal Statistical Society: Series B, 76(1), 243-263. DOI: 10.1111/rssb.12027 ↗
How to cite this page
ScholarGate. (2026, June 3). Machine Learning-Augmented Coarsened Exact Matching Estimator. ScholarGate. https://scholargate.app/en/causal-inference/machine-learning-augmented-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.
- Coarsened Exact MatchingCausal inference↔ compare
- Doubly Robust EstimationCausal inference↔ compare
- Entropy BalancingCausal inference↔ compare
- Machine Learning-Augmented Propensity Score MatchingCausal inference↔ compare
- Matching EstimatorCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare