Coarsened Exact Matching (CEM)

• Coarsened Exact Matching is a stratification method that discretizes covariate space into bins to create balanced strata for comparing treated and control units.
• It employs a closed-form weighting scheme derived from dual-norm operator theory to achieve explicit error control, though it faces challenges such as sample attrition and residual confounding.
• Practical applications and extensions, including clustering-based coarsening and survival analysis adaptations, expand CEM’s utility despite limitations in high-dimensional contexts.

Coarsened Exact Matching (CEM) is a nonparametric, monotonic imbalance bounding (MIB) approach to covariate adjustment in causal inference and observational studies. CEM assigns units to strata based on discretized covariate bins and restricts matched samples to those strata containing both treated and control units. Subsequent estimation leverages this stratification to compare outcomes across balanced, coarsened populations. The method is notable for its conceptual simplicity, closed-form implementation, and direct connection to empirical process and operator-norm frameworks. However, CEM is fundamentally limited by the curse of dimensionality and residual confounding from inexact matching on coarse bins as opposed to original variables.

1. Formal Definition and Algorithmic Foundations

Coarsened Exact Matching operates by mapping each unit’s covariate vector $X \in \mathbb{R}^d$ to a discrete stratum via coordinate-wise partitioning. Each covariate $X_j$ is binned into $m_j$ intervals $\mathcal{C}_j = \{C_{j,1}, \ldots, C_{j,m_j}\}$, yielding the coarsening map $c : \mathbb{R}^d \rightarrow S$ where

$$c(X) = (c_1(X_1), \ldots, c_d(X_d)), \quad c_j(X_j) = \ell \ \text{if} \ X_j \in C_{j,\ell}.$$

The full set of strata is $$S = \{1, \ldots, m_1\} \times \cdots \times \{1, \ldots, m_d\}$$.

Units with treatment indicator $T_i \in \{0,1\}$ are retained only if their stratum $s = c(X_i)$ contains at least one unit from both arms. Each treated unit is matched to all control units sharing the same stratum; unmatched units are discarded. The analyst specifies bin cutpoints (subject-matter or quantiles) prior to observing outcomes. CEM can be parameterized either via custom binning or automated rules (e.g., Sturges’ rule $m_j \approx \lceil \log_2 n + 1 \rceil$).

A closed-form weighting scheme is derived from dual-norm operator theory (Kallus, 2016): for each control unit $i$ in stratum $k$,

$w_i^* = \frac{n_{1k} / n_1}{n_{0k}},$

where $n_{1k}, n_{0k}$ are the treated and control counts in stratum $k$, respectively. The Average Treatment effect on the Treated (ATT) estimator is then

$$\hat\tau_{\text{CEM}} = \frac{1}{n_1}\sum_{i\in T} Y_i - \sum_{i\in C} w_i^* Y_i = \sum_{k=1}^K \frac{n_{1k}}{n_1}(\bar{Y}_{T,k} - \bar{Y}_{C,k}).$$

These weights enforce exact balance in stratum frequencies.

2. Theoretical Properties and Empirical Process Characterization

CEM can be formally embedded in an empirical process framework (Cortés et al., 2023). The balance function class $\mathcal{H}_{\text{CEM}} = \{h_s : s\in S\}$ consists of indicator functions for each stratum, $h_s(x) = 1\{c(x) = s\}$. Covariate balance after matching is measured as

$$\sup_{h \in \mathcal{H}_{\text{CEM}}} |\hat{G}_1(h) - \hat{G}_0(h)|,$$

with $\hat{G}_t(h) = (1/n_t)\sum_{i:T_i = t} h(X_i)$.

The complexity of $\mathcal{H}_{\text{CEM}}$ is $$\left|\mathcal{H}_{\text{CEM}}\right| = |S| = \prod_{j=1}^d m_j$$, and the VC-dimension is bounded by

$$V(\mathcal{H}_{\text{CEM}}) \leq \sum_{j=1}^d \log_2 m_j.$$

Uniform imbalance is controlled via a finite-sample concentration inequality (Thm 3.1), showing with high probability

$$\sup_{s \in S} |\hat{G}_1(h_s) - \hat{G}_0(h_s)| \le 2 \mathcal{R}_n(\mathcal{H}_{\text{CEM}}) + \sqrt{\frac{\log(2/\delta)}{2n}}$$

where $\mathcal{R}_n(\mathcal{H}_{\text{CEM}})$ is the Rademacher complexity. A further upper bound is

$$\mathcal{R}_n(\mathcal{H}_{\text{CEM}}) \le \sqrt{\frac{2V(\mathcal{H}_{\text{CEM}})\log(en/V(\mathcal{H}_{\text{CEM}}))}{n}},$$

implying uniform post-matching imbalance scales as $O\left(\sqrt{V/n}\right)$. These results establish CEM as an MIB method with explicit error control.

3. Bias-Variance Trade-Off, Bin Selection, and Dimensionality Constraints

The granularity of coarsening (number of bins per covariate) fundamentally controls CEM’s bias-variance tradeoff. Finer coarsening ($m_j \uparrow$) yields smaller bias but more empty strata, reducing effective sample size and inflating variance. A practical rule is to set $m_j \asymp n^{1/(2+d)}$, which balances $\sqrt{V/n}$, with $V = \sum_{j=1}^d \log_2 m_j$. Minimum stratum cell-size rules (≥5 units/stratum) are strongly recommended.

Empirical findings confirm that CEM achieves lowest $L_1$ imbalance across sample sizes $n \in \{200, 500, 1000\}$ and dimensions $d \in \{5, 10, 20\}$, but discards 5–15% of units in moderate-to-high dimension scenarios. Marginal improvements in balance from additional coarsening come at the cost of higher data loss. The curse of dimensionality is substantial: for $p=5,7$ and modest $m_j$, simulation studies report retention below 20%, whereas PSM retains 70–90% (Wan, 25 Jan 2026). In high dimensions, residual confounding and unstable estimates are prominent.

4. Comparison with Related Matching and Balancing Methods

4.1 CEM versus Propensity Score Matching (PSM)

While CEM is conceptually simple and nonparametric, it is inexact: matching is performed only on discretized bins, not the continuous covariates themselves. This yields persistent residual confounding unless bins are sufficiently fine—a condition rarely tenable in high dimensions due to data attrition (Wan, 25 Jan 2026). PSM, in contrast, leverages the balancing-score property $\bm{X} \perp W \mid e(\bm{X})$, achieving unbiased ATT estimates (via mean difference) and robustness to post-matching model misspecification. Wan’s simulations (2026) demonstrate:

• Unadjusted bias: CEM(auto) ≈ 0.10; PSM ≈ 0.03
• Retention: CEM(auto) ≈ 40%; PSM ≈ 80%
• Multivariate SMD: PSM yields lower imbalance

Post-matching outcome model misspecification affects CEM but not PSM, unless CEM bins have been specifically tailored to sufficient statistics for outcome.

4.2 Operator-Norm and Kernel Matching

In operator-norm terms, CEM minimizes the dual norm over piecewise-constant function classes, yielding closed-form exact stratum balancing (Kallus, 2016). More general kernel-based matching estimators can be constructed, inheriting universal consistency properties.

4.3 Generalized Coarsened Confounding

Recent developments generalize CEM to allow clustering-based coarsening (e.g., k-means, random forest proximity) rather than fixed, Cartesian product binning (Ghosh et al., 6 Jan 2025). The estimator and variance formulas remain unchanged, extending large-sample consistency and asymptotic normality to data-driven partitions.

5. Extensions, Survival Analysis, and Monotonic Imbalance Bounding

CEM forms the basis for monotonic imbalance bounding (MIB) methods: as bin widths $\Delta_j \downarrow 0$, marginal imbalance in each covariate vanishes, but practical sample sizes become limiting (Che et al., 2024). Extensions such as Caliper Synthetic Matching (CSM) replace bins with radius-based calipers in metric space, improving joint covariate balance and allowing local synthetic controls. CSM recovers CEM in the limit of uniform calipers and fixed bin widths, and sharpens bias bounds from $O(c)$ to $o(c)$ under Lipschitz outcome surfaces.

In survival analysis, CEM serves as the basis for weighted log-rank statistics. Control weights are time-varying and ensure exactly matched risk counts within strata; the resulting log-rank process admits asymptotic normality under the null and consistency under alternatives (Baba et al., 2024). Simulation studies show superior robustness of CEM to misspecified propensity models.

6. Implementation Guidelines and Limitations

Best practice recommendations include:

• Begin with coarse binning (2–3 per covariate) to control VC dimension and avoid empty strata.
• Use quantile-based cutpoints for equitable stratum sizes.
• Check marginal SMDs post-matching; refine binning or switch methods if $\max_j|\text{SMD}_j| > 0.10$.
• Employ minimum cell-size rules (≥5 units/stratum).
• In moderate to high dimensions ($p \ge 5$) or when covariate-outcome relationships are complex/nonlinear, prefer methods such as PSM or kernel balancing.

Limitations of CEM include:

• Exponential growth of strata with increasing $d$ and $m_j$, resulting in sample attrition.
• Residual confounding and model dependence if outcome models are not correctly specified for the original covariates.
• Data loss and unstable ATT estimates in high dimensionality unless coarse bins are used, which increases bias.

7. Applications and Empirical Findings

CEM has been widely adopted in policy evaluation, clinical study analysis, and computational benchmarking. Simulation studies confirm its superiority versus nearest-neighbor and optimal matching in low dimensions, and its competitive balance versus PSM when bin counts and sample sizes are moderate (Cortés et al., 2023). Generalized coarsened confounding approaches (e.g., k-means or RF proximity coarsening (Ghosh et al., 6 Jan 2025)) yield improved precision and bias control in canonical datasets, including the LaLonde job training and birth weight studies.

In survival analysis (Baba et al., 2024), CEM-weighted log-rank tests outperform IPTW-based approaches, particularly under covariate-treatment assignment misspecification. Empirical applications (e.g., Brazilian educational interventions (Che et al., 2024)) demonstrate that local synthetic controls can further reduce bias while retaining effective sample sizes.

---

In summary, Coarsened Exact Matching provides a rigorous, transparent, and computationally tractable method for causal effect estimation via stratified covariate balance. Its empirical process and dual-norm operator foundations yield explicit bounds and closed-form estimators, yet the method is fundamentally limited by the curse of dimensionality and inexactness of coarsened matching. Extensions via clustering, metric-space calipers, and synthetic controls expand its applicability and precision, but tradeoffs between granularity, bias, and retained sample size remain central in practice.