Heterogeneous Treatment Effect Coarsened Exact Matching
Also known as: HTE-CEM, CEM with CATE estimation, subgroup CEM, coarsened exact matching with effect heterogeneity
Heterogeneous treatment effect coarsened exact matching (HTE-CEM) extends the coarsened exact matching framework to estimate how treatment effects vary across subgroups or individual characteristics. After CEM creates balanced strata by coarsening continuous covariates into bins and exactly matching units within each bin, conditional average treatment effects (CATEs) are computed within or across these strata, revealing where treatment works, for whom, and by how much.
Key highlights
- Provides transparent, stratum-level balancing that is easy to audit and communicate to non-technical stakeholders.
- Avoids model dependence in the balancing stage — no propensity model needs to be correctly specified.
- Naturally partitions the sample into strata that serve as intuitive subgroups for heterogeneity analysis.
- Restricts inference to the region of common support, reducing bias from extrapolation.
- CEM weights integrate seamlessly into standard weighted regression frameworks for CATE estimation.
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 HTE-CEM when you suspect treatment effects differ across measurable subgroups and you have an observational dataset where randomisation was not feasible. It is most appropriate when covariates include a mix of continuous and categorical variables, when the common-support region is of primary interest, and when transparency of the matching algorithm is required for applied or policy audiences. Prefer it over propensity-score-based HTE methods when you want to avoid model dependence in the first-stage balancing step. Do not use it when the covariate space is so high-dimensional that most strata become empty after matching, when the treatment is continuous rather than binary, or when the primary goal is an overall ATE rather than subgroup-specific estimates.
Strengths & limitations
- Provides transparent, stratum-level balancing that is easy to audit and communicate to non-technical stakeholders.
- Avoids model dependence in the balancing stage — no propensity model needs to be correctly specified.
- Naturally partitions the sample into strata that serve as intuitive subgroups for heterogeneity analysis.
- Restricts inference to the region of common support, reducing bias from extrapolation.
- CEM weights integrate seamlessly into standard weighted regression frameworks for CATE estimation.
- Coarsening choices (bin width, cutpoints) affect which units match and which strata are retained, introducing researcher degrees of freedom.
- In high-dimensional covariate spaces, most strata become empty, drastically reducing the matched sample and statistical power.
- Subgroup CATE estimates within narrow strata can be imprecise when stratum sample sizes are small.
- The method does not directly handle continuous treatments; it requires a binary or discretised treatment variable.
- Pruning unmatched units changes the estimand — results apply to the matched subsample, not necessarily the full target population.
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 do I choose the coarsening cutpoints for each covariate?
CEM defaults to automatic cutpoints (e.g., Sturges rule for histogram bins), but user-defined cutpoints based on theory or policy-relevant thresholds are preferable. The key trade-off is bias versus sample size: finer bins improve within-stratum balance but increase the fraction of unmatched units.
How is HTE-CEM different from running a subgroup analysis after ordinary CEM?
Standard post-CEM subgroup analysis splits the matched sample by a single variable and computes within-split ATTs. HTE-CEM is more systematic: it estimates CATEs as a function of multiple covariates simultaneously, either by leveraging the natural stratum structure or by fitting a weighted interaction model, and it explicitly reports and tests effect heterogeneity rather than treating subgroups as an afterthought.
What if many strata are empty after matching?
This is a sign of poor common support or overly fine coarsening. Widen the bin boundaries, reduce the number of matching covariates to those that are most strongly confounding, or switch to a method better suited to high-dimensional settings such as a causal forest.
Can I use HTE-CEM with panel data?
Yes, with modification. Match units at baseline on pre-treatment covariates using CEM, then model the outcome in a panel difference-in-differences framework on the matched-and-weighted sample. This gives stratum-specific DiD estimates that combine the balancing strength of CEM with the temporal controls of panel methods.
How many observations do I need for reliable CATE estimates?
As a rough guide, each subgroup stratum should contain at least 20-30 matched pairs to produce stable stratum-level estimates. Overall, CATE analyses typically require larger samples than simple ATT estimation because the effective sample is partitioned across subgroups.
Sources
- 1.Iacus, S. M., King, G., & Porro, G. (2012). Causal Inference without Balance Checking: Coarsened Exact Matching. Political Analysis, 20(1), 1-24.
- 2.Imai, K., & Ratkovic, M. (2013). Estimating treatment effect heterogeneity in randomized program evaluation. Annals of Applied Statistics, 7(1), 443-470.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Heterogeneous Treatment Effect Coarsened Exact Matching. ScholarGate. https://scholargate.app/causal-inference/heterogeneous-treatment-effect-coarsened-exact-matching