Robust Matching Estimator (Bias-Corrected Matching)
Bias-Corrected Robust Matching Estimator · Also known as: bias-corrected matching, Abadie-Imbens matching, AI matching estimator, robust nearest-neighbor matching
The robust matching estimator, developed by Abadie and Imbens (2006, 2011), extends nearest-neighbor matching by adding a regression-based bias correction that removes the finite-sample bias arising when matched units are not perfectly alike. It yields consistent, asymptotically normal estimates of average treatment effects with a heteroskedasticity-robust variance formula that is valid regardless of the number of continuous covariates.
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 robust matching estimator when you need to estimate a causal treatment effect from observational data and the strong ignorability assumption (treatment assignment independent of potential outcomes conditional on covariates) is plausible. It is especially valuable when you have a moderate number of continuous covariates that make exact or coarsened exact matching difficult. Prefer it over standard matching when you want formal bias corrections and robust inference. It is not appropriate when the overlap assumption fails badly (treatment and control covariate distributions do not overlap), when the number of covariates is very high (the curse of dimensionality degrades match quality and bias correction), or when treatment assignment depends on variables not observed in your data.
Strengths & limitations
- Provides a formal bias correction that removes finite-sample matching bias caused by imperfect matches, a deficiency absent in standard matching.
- Robust variance estimator is heteroskedasticity-consistent and accounts for observations being used as matches multiple times.
- Does not require a correctly specified propensity score model — matching is performed directly on covariates.
- Supports estimation of both ATE and ATT under a unified framework.
- Asymptotically normal, enabling standard confidence intervals and hypothesis tests.
- Curse of dimensionality: as the number of continuous covariates grows, match quality degrades and the bias correction may be insufficient.
- Requires the strong ignorability assumption — unobserved confounders that affect both treatment and outcome invalidate causal interpretation.
- Choice of M (number of matches) and distance metric introduces researcher discretion that can affect results.
- Computationally more demanding than simple nearest-neighbor matching due to the local regression step.
Frequently asked
How is the robust matching estimator different from standard nearest-neighbor matching?
Standard nearest-neighbor matching takes the raw outcome difference between matched pairs as the treatment effect. The robust version adds a regression-based bias correction that accounts for the residual covariate difference within each matched pair, and uses a variance formula that correctly handles multiplicity (the same control being matched to multiple treated units). This makes it valid for inference in the presence of continuous covariates.
How many matches M should I use?
Using M = 1 gives minimum bias but maximum variance; larger M reduces variance but can increase bias if distant matches are forced. Common practice is M in {1, 2, 4}. Abadie and Imbens suggest checking sensitivity across values of M, since the bias correction partially offsets the bias introduced by less-than-perfect matches.
Do I need a propensity score for this method?
No. The robust matching estimator matches directly on covariates using a distance metric (typically Mahalanobis), bypassing the propensity score. This can be an advantage when the propensity score model is uncertain, but it is also why curse-of-dimensionality concerns are relevant with many covariates.
When should I prefer doubly robust estimation instead?
Doubly robust methods are preferable when you have many covariates or when you want protection against misspecification of either the outcome or propensity model. The robust matching estimator is competitive when covariate dimension is low to moderate and you want a nonparametric, model-light approach to causal inference.
Is the bias correction always sufficient?
Not when overlap is poor or the number of continuous covariates is large. The bias correction uses a local regression that itself can be imprecise if matched units are far apart. In such cases, trimming on common support or switching to propensity-score weighting methods is advisable.
Sources
- Abadie, A., & Imbens, G. W. (2011). Bias-Corrected Matching Estimators for Average Treatment Effects. Journal of Business & Economic Statistics, 29(1), 1-11. DOI: 10.1198/jbes.2009.07333 ↗
- 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 ↗
How to cite this page
ScholarGate. (2026, June 3). Bias-Corrected Robust Matching Estimator. ScholarGate. https://scholargate.app/en/causal-inference/robust-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
- Doubly Robust EstimationCausal inference↔ compare
- Inverse Probability WeightingCausal inference↔ compare
- Matching EstimatorCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare