Some Practical Guidance for the Implementation of Propensity Score Matching
Causal Question / Estimand
The ATT of a program (typically a labour-market policy) under selection on observables — and a step-by-step protocol for the implementation choices PSM forces on the researcher.
Identification Strategy
Selection on observables: conditional on covariates, treatment is independent of potential outcomes (Ignorability), so matched untreated units supply the treated counterfactual. Because conditioning on many covariates is infeasible, match on the scalar Propensity-Score. The paper structures the full pipeline: (1) estimate the propensity score (model and variable choice); (2) choose a matching algorithm (nearest-neighbour, caliper/radius, kernel, stratification); (3) impose the common support / Overlap region; (4) assess matching quality / covariate balance; (5) estimate effects and standard errors; (6) conduct sensitivity analysis to hidden bias (Rosenbaum bounds).
Key Assumptions
Ignorability (conditional independence / unconfoundedness), Overlap (common support), SUTVA, and a Propensity-Score specification that achieves balance.
Threats to Validity
Unobserved heterogeneity violating conditional independence (probed via Rosenbaum sensitivity bounds); failure of common support; bias–variance trade-offs across matching algorithms; specification of the score. The paper is explicit that “each step involves a lot of decisions” — researcher degrees of freedom.
Setting / Data
n/a — methodological survey/guide; oriented to programme (labour-market policy) evaluation.
Key Claims
- PSM identification is no stronger than unconfoundedness + overlap; the value added is a disciplined sequence of implementation decisions.
- No matching algorithm dominates; choices trade bias against variance and should be reported and checked for robustness.
- Because unconfoundedness is untestable, sensitivity analysis to unobserved confounding is an essential part of credible PSM.
Connections
- Builds on: Ignorability, Overlap, Propensity-Score, Imbens2004-NonparametricATEReview
- Practitioner sibling: Imbens2015-MatchingMethodsInPractice
- Cautionary evidence on robustness: SmithTodd2005-ReconcilingPSMEvidence. See also PSM
Citation
Caliendo, M., & Kopeinig, S. (2008). Some Practical Guidance for the Implementation of Propensity Score Matching. Journal of Economic Surveys, 22(1), 31–72.