Policy Evaluation Matching Estimator
Also known as: matching estimator, program evaluation matching, treatment effect matching, Abadie-Imbens estimator
The policy evaluation matching estimator estimates the causal effect of a program or policy on treated units by pairing each participant with one or more non-participants who share similar pre-treatment characteristics. Developed rigorously by Heckman, Ichimura & Todd (1998) and Abadie & Imbens (2006), it avoids parametric outcome models and is the standard non-parametric tool for program and policy evaluation.
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 the policy evaluation matching estimator when you must estimate a treatment or program effect from observational data and you can assume selection on observables — that is, all variables driving program participation are measured. It is best suited for cross-sectional or short-panel data where treated and untreated units overlap substantially in covariate space. Do not use it when important confounders are unobserved (an instrument or DiD design is then preferable), when overlap between the groups is very thin, or when the sample is smaller than roughly 50 per group, as sparse matches produce large variance.
Strengths & limitations
- Non-parametric: does not impose a functional form on the outcome equation, making it robust to model misspecification.
- Transparent: the matched pairs are observable, so the quality of counterfactual construction can be directly inspected.
- Straightforward ATT interpretation: answers 'what did the program do for those who participated?'
- Can be combined with regression adjustment (bias-corrected matching) to reduce remaining bias from imperfect matches.
- Widely accepted in program evaluation, economics, and policy research with well-established inference theory.
- Requires the conditional independence assumption (selection on observables): if unobserved confounders drive selection, the estimator is biased and no diagnostic can fully reveal this.
- Demands sufficient overlap between treated and control covariate distributions; units in thin regions of the common support must be trimmed, potentially limiting generalizability.
- Matching on many covariates in small samples leads to poor matches, high variance, and unreliable estimates.
- Standard bootstrap inference is invalid for nearest-neighbor matching; analysts must use the Abadie-Imbens analytic variance or bias-corrected variants.
Frequently asked
What is the conditional independence assumption and can I test it?
It states that, conditional on observed covariates, the potential outcomes are independent of treatment assignment. It cannot be tested directly because the counterfactual is unobserved. You can support it by including all plausible confounders, checking balance, and conducting a Rosenbaum bounds sensitivity analysis to gauge how large an unobserved confounder would need to be to overturn your conclusion.
How many matches per treated unit should I use?
Using one match (k=1) minimizes bias but maximizes variance. Using more matches (k=3 or 5) reduces variance at the cost of slightly worse matches and some additional bias. For most policy-evaluation samples, k=1 to 4 with a caliper on the propensity score is a reasonable default, combined with the bias-corrected Abadie-Imbens estimator.
Is propensity-score matching the same as the policy evaluation matching estimator?
Propensity-score matching is one variant: matches are found based on the estimated probability of treatment (propensity score) rather than raw covariate distances. The Abadie-Imbens framework typically uses Mahalanobis distance on covariates directly, which avoids the added estimation uncertainty from the propensity model. Both aim at the same ATT estimand.
What if treated and control groups barely overlap?
Poor overlap (thin common support) forces matches across large covariate distances, producing biased counterfactuals. The standard remedy is to trim units outside a common-support region — those treated units whose covariate values have no comparable controls — and restrict inference to the overlapping subpopulation.
Can I combine matching with regression adjustment?
Yes. The bias-corrected matching estimator (Abadie & Imbens 2011) adds a regression-based correction within matched groups to remove residual bias from imperfect matches. This doubly robust approach is recommended whenever the number of continuous covariates is large relative to sample size.
Sources
- Abadie, A., & Imbens, G. W. (2006). Large sample properties of matching estimators for average treatment effects. Econometrica, 74(1), 235-267. DOI: 10.1111/j.1468-0262.2006.00655.x ↗
- Heckman, J. J., Ichimura, H., & Todd, P. (1998). Matching as an econometric evaluation estimator. Review of Economic Studies, 65(2), 261-294. DOI: 10.1111/1467-937X.00044 ↗
How to cite this page
ScholarGate. (2026, June 3). Policy Evaluation Matching Estimator. ScholarGate. https://scholargate.app/en/causal-inference/policy-evaluation-matching-estimator
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
- Difference-in-DifferencesEconometrics↔ compare
- Instrumental Variables in Health ResearchHealth Economics↔ compare
- Inverse Probability WeightingCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare
- Regression DiscontinuityCausal inference↔ compare