Robust Propensity Score Matching
Robust Propensity Score Matching Estimator · Also known as: robust PSM, PSM with robust variance, bias-corrected PSM, matching with robust inference
Robust Propensity Score Matching (robust PSM) is a quasi-experimental causal inference method that pairs treated and control units on their estimated probability of receiving treatment (the propensity score), then estimates the average treatment effect using variance estimators that account for the uncertainty introduced by estimating the propensity score itself. The correction, developed by Abadie and Imbens (2016), prevents misleading inference that standard bootstrap or analytic formulas produce when applied naively after matching.
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 robust PSM when you want a causal treatment-effect estimate from observational data with a binary treatment and you have a rich set of pre-treatment covariates that plausibly satisfy unconfoundedness. It is particularly appropriate when the matched-sample size is modest (< 500 matched pairs) and when you need reliable standard errors for inference rather than just a point estimate. Prefer robust PSM over naive PSM whenever you plan to report confidence intervals or p-values. Do not use it when key confounders are unobserved — matching only removes observed selection bias. Avoid it when the treatment and control covariate distributions barely overlap (poor common support), or when the outcome or treatment is measured with substantial error.
Strengths & limitations
- Corrects the known downward bias in standard PSM variance estimators, producing confidence intervals with correct coverage.
- Balances observed confounders without requiring a parametric outcome model, making the estimate more robust to outcome-model misspecification.
- Transparent: balance diagnostics (SMD plots, love plots) allow direct inspection of how well confounding is removed.
- Applicable in many disciplines without large-sample requirements compared to doubly robust or semiparametric alternatives.
- Flexible caliper and matching-ratio choices let practitioners trade bias against variance according to the data structure.
- Unconfoundedness (no unmeasured confounders) is untestable from data alone; the method cannot remove hidden bias.
- Matching discards unmatched control units, potentially wasting information and reducing effective sample size.
- The Abadie-Imbens robust variance applies to nearest-neighbour matching; it does not extend automatically to kernel or radius matching without additional adaptation.
- Performance degrades when the propensity model is heavily misspecified, even though the outcome model is not assumed.
Frequently asked
How does robust PSM differ from standard PSM?
The matching procedure and the ATT point estimate are identical. The difference is in the standard errors: standard PSM typically uses naive analytic formulas or bootstrapping that treat propensity scores as known, understating uncertainty. Robust PSM adds a correction term for the variance introduced by estimating propensity scores, yielding valid confidence intervals and p-values.
Can I use bootstrapping instead of the Abadie-Imbens correction?
Bootstrapping the final matched dataset is not valid because each bootstrap resample re-estimates the propensity scores differently, yet the matches from the original sample are fixed. Abadie and Imbens showed that even wild bootstrap methods are asymptotically invalid for nearest-neighbour matching. Use the analytical variance correction instead.
What if my covariates do not balance well after matching?
Tighten the caliper (e.g., to 0.1 or 0.2 standard deviations of the log-odds score), increase the number of matches, or re-specify the propensity model by adding interactions or non-linear terms. If balance remains poor, consider entropy balancing or doubly robust estimation as alternatives.
How many matches per treated unit should I use?
One-to-one matching minimises bias but maximises variance. Increasing to 2-to-1 or 3-to-1 reduces variance at the cost of potentially worse balance because more distant controls are used. In practice, start with one-to-one and assess balance; add more matches only if variance is problematic and balance diagnostics remain acceptable.
Does robust PSM handle unobserved confounders?
No. Like all matching methods, robust PSM relies on the unconfoundedness assumption — that all confounders are measured and included in the propensity model. Unobserved confounders remain a threat. Sensitivity analysis (e.g., Rosenbaum bounds) can quantify how strong a hidden confounder would need to be to overturn the conclusion.
Sources
- Abadie, A., & Imbens, G. W. (2016). Matching on the Estimated Propensity Score. Econometrica, 84(2), 781-807. DOI: 10.3982/ECTA11293 ↗
- 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). Robust Propensity Score Matching Estimator. ScholarGate. https://scholargate.app/en/causal-inference/robust-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.
- Coarsened Exact MatchingCausal inference↔ compare
- Doubly Robust EstimationCausal inference↔ compare
- Inverse Probability WeightingCausal inference↔ compare
- Matching EstimatorCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare
- Propensity Score WeightingCausal inference↔ compare