Heterogeneous Treatment Effect Propensity Score Matching
Heterogeneous Treatment Effect Estimation via Propensity Score Matching · Also known as: HTE-PSM, CATE via PSM, subgroup treatment effect matching, conditional average treatment effect matching
Heterogeneous Treatment Effect Propensity Score Matching extends standard PSM to estimate how treatment effects vary across subgroups or individual characteristics. Rather than reporting a single average treatment effect, it uses the matched sample to estimate conditional average treatment effects (CATE), revealing which types of units benefit most or least from a treatment.
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 this method when you have observational data with a binary treatment, sufficient covariates to plausibly satisfy unconfoundedness, and a research question about who benefits from the treatment — not just on average. It is appropriate when sample sizes within meaningful subgroups are large enough for reliable inference (roughly 50+ matched pairs per subgroup). Avoid it when the propensity model cannot achieve good balance, when the number of subgroups is very large relative to sample size (risking multiple-testing inflation), or when treatment is continuous rather than binary — a dose-response framework is more suitable then.
Strengths & limitations
- Recovers treatment effect heterogeneity that a single ATT estimate conceals, supporting personalised or targeted policy recommendations.
- The PSM step removes overt confounding before subgroup analysis, making within-subgroup comparisons more credible than naive stratification.
- Combines the interpretability of matching with the richness of CATE estimation, making findings accessible to applied audiences.
- Can be implemented with pre-specified subgroups (confirmatory) or exploratory causal-tree methods, offering flexibility in research design.
- Directly answers policy-relevant questions such as which subpopulations should be targeted for treatment.
- Relies on the strong ignorability assumption (unconfoundedness): all confounders must be observed and included in the propensity model. Unmeasured confounding invalidates causal interpretation of both average and heterogeneous effects.
- Power is substantially reduced within subgroups; small matched samples within strata produce wide confidence intervals and low reliability.
- Exploratory subgroup discovery risks false positives; without pre-registration or correction for multiple comparisons, reported heterogeneity may be spurious.
- The propensity score must achieve good balance across all covariate dimensions; poor balance in even one subgroup undermines CATE estimates in that stratum.
- Effect heterogeneity detected in an observational study may still reflect residual confounding rather than genuine treatment-effect variation.
Frequently asked
How is this different from standard propensity score matching?
Standard PSM estimates a single average treatment effect on the treated (ATT). HTE-PSM goes further by examining how that effect varies across subgroups or individual characteristics, estimating conditional average treatment effects (CATE) within the matched sample.
What is the risk of exploring many subgroups?
Each subgroup comparison is a hypothesis test, so examining many subgroups without adjusting the significance threshold inflates the false-positive rate. Pre-register subgroups of interest or apply Bonferroni or Benjamini-Hochberg corrections to control multiplicity.
Can I use machine learning for the propensity model?
Yes. Flexible learners such as gradient boosting or random forests often achieve better covariate balance than logistic regression in high-dimensional settings. However, balance should always be checked empirically — a low prediction error does not guarantee good matching balance.
What sample size is sufficient for subgroup analysis?
A rough guideline is at least 50 matched pairs per subgroup for moderately sized effects. For smaller effects or more subgroups, the required sample is larger. Formal power calculations for interaction tests are advisable before committing to an HTE design.
Is unconfoundedness more demanding for HTE than for average effects?
The same ignorability assumption is required. However, because CATE estimates are more local (within subgroups), any residual confounding that happens to correlate with the moderating variable of interest can produce spurious-looking heterogeneity, making robustness checks especially important.
Sources
- Athey, S., & Imbens, G. W. (2016). Recursive Partitioning for Heterogeneous Causal Effects. Proceedings of the National Academy of Sciences, 113(27), 7353-7360. DOI: 10.1073/pnas.1510489113 ↗
- Rosenbaum, P. R., & Rubin, D. B. (1983). The Central Role of the Propensity Score in Observational Studies for Causal Effects. Biometrika, 70(1), 41-55. DOI: 10.1093/biomet/70.1.41 ↗
How to cite this page
ScholarGate. (2026, June 3). Heterogeneous Treatment Effect Estimation via Propensity Score Matching. ScholarGate. https://scholargate.app/en/causal-inference/heterogeneous-treatment-effect-propensity-score-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.
- Difference-in-DifferencesEconometrics↔ compare
- Doubly Robust EstimationCausal inference↔ compare
- Heterogeneous Treatment Effect Difference-in-DifferencesCausal inference↔ compare
- Matching EstimatorCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare