Skip to content
Back to search results

Propensity Score Matching

Introduced by
Paul Rosenbaum and Donald Rubin
Year
1983
Type
Method
Family
causal-inference
Font size

What does it do?

Propensity score matching (PSM) is a method for reducing confounding bias in observational studies by balancing baseline characteristics between treatment groups, simulating randomization.

Developed by Rosenbaum and Rubin (1983), it estimates the probability of receiving treatment given observed covariates, then matches or weights treated and control individuals with similar treatment probabilities. Widely used in medicine, epidemiology, and policy evaluation when randomized trials are infeasible or unethical, enabling estimation of treatment effects while controlling for selection bias.

The intuition behind the method

Randomized trials balance baseline characteristics automatically, enabling causal inference. Observational studies lack randomization; treated individuals often differ systematically from untreated (selection bias). Comparing outcomes between treatment groups confounds treatment effects with preexisting differences. Propensity score matching (PSM) estimates the probability that each person receives treatment given their baseline characteristics. By matching treated and untreated individuals with similar propensity scores (treatment probabilities), PSM creates two groups comparable at baseline, mimicking randomization. This removes confounding bias for measured covariates, allowing valid causal inference. The key assumption: conditional independence—given observed covariates, treatment assignment is random (no unmeasured confounding).

History

Paul Rosenbaum and Donald Rubin's 1983 foundational paper unified matching and causal inference via potential outcomes framework (Rubin, 1974). They proved propensity score matching achieves covariate balance, reducing bias. Peter Austin's work (2010s) advanced balance diagnostics and weighting methods. Modern extensions include: generalized propensity score for continuous treatments (Hirano & Imbens, 2004), doubly robust methods combining propensity and outcome models (van der Laan & Robins, 2003), targeted maximum likelihood estimation (TMLE; van der Laan & Rose, 2011).

When to use it

Use propensity score matching when conducting observational studies and unable to randomize, but all confounders are measured. Medical research: evaluating treatment efficacy (surgery vs conservative care) comparing outcomes for matched treated/untreated patients with similar baseline risk. Policy evaluation: assessing program impact (e.g., vocational training on employment) comparing participants and matched nonparticipants. Epidemiology: estimating disease risk associated with exposure while controlling baseline characteristics. Prerequisite: must have collected sufficient baseline data on confounders before treatment assignment; unmeasured confounding (e.g., unmeasured disease severity) remains a threat.

Limitations and considerations

  • Conditional independence assumption: valid only if all relevant confounders are measured. Unmeasured confounding (e.g., unmeasured disease severity, patient preference) remains a threat; PSM cannot detect or correct it.
  • Inefficiency: matching reduces sample size (unmatched individuals discarded); matched analyses are less statistically powerful than unmatched regression.
  • Limited overlap: if treated and untreated populations have very different covariate distributions, overlap region is small; estimates restricted to overlap region only.
  • Model dependence: propensity score estimates depend on correct specification of logistic regression; misspecification (omitted interactions, nonlinear relationships) yields poor balance.
  • Loss of information: one-to-one matching discards potentially informative unmatched individuals; many-to-one matching or weighting retain all data.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Reduces confounding bias: by creating groups balanced on measured confounders, PSM mimics randomization for those variables, enabling causal inference.
  • Interpretable: propensity score is a single scalar collapsing many covariates; transparent how covariate balance achieved.
  • Flexible: works with binary or multiple treatments; extends to continuous treatments (generalized propensity score).
  • Diagnostic: covariate balance assessment (SMD, love plots) provides clear evidence of bias reduction; imbalance visible before outcome analysis.
  • Doubly robust: when combined with outcome regression, the estimator is consistent if either the propensity model or outcome model is correct (resilience to misspecification).

Applications

  • Comparative effectiveness research: evaluating treatment outcomes (medication A vs B, surgical vs medical management) in electronic health records without experimental randomization.
  • Policy evaluation: assessing program impact (job training, school interventions) on outcomes (employment, earnings) comparing program participants with matched nonparticipants.
  • Epidemiological cohort studies: estimating disease risk from exposures (smoking, diet) in observational data while adjusting for demographic and lifestyle confounders.
  • Economic impact studies: evaluating effects of regulatory changes, tax policies, or healthcare reforms on individual outcomes; balancing treated and control states.
  • Vaccine effectiveness: comparing outcomes (infection rates, disease severity) in vaccinated vs unvaccinated matched on baseline risk factors when RCTs infeasible.

How it works

How it works

1

Administer propensity score matching by: (1) identifying the outcome of interest, treatment variable (binary or multiple), and all relevant confounders (baseline characteristics that affect both treatment and outcome), (2) collecting covariate data on all participants, (3) checking the overlap assumption: treatment and control groups must have overlapping ranges of propensity scores; lack of overlap indicates fundamental incomparability, (4) defining the sample: intention-to-treat (analyze in original groups) or per-protocol (analyze matched sample; causality claims require caution). Ensure adequate sample size: PSM reduces effective sample size via matching, so begin with larger N than equivalent RCT.

2

Estimate propensity scores: use logistic regression to predict probability of treatment assignment given all measured covariates. p_i = P(Treatment=1 | Covariates). Propensity score is a scalar (0–1) collapsing multiple covariates into one dimension. Then, match treated (score p) with untreated (score q) if |p − q| < caliper (tolerance, typically 0.1 × SD of logit propensity score); common approaches: 1:1 nearest neighbor matching, 1:m nearest neighbor, or caliper matching. Alternative to matching: propensity score weighting (inverse probability of treatment weighting, IPTW). Weight treated individuals by 1/p_i (up-weighting underrepresented treated) and controls by 1/(1−p_i); produces pseudo-population where treatment is independent of covariates. Check covariate balance: standardized mean difference (SMD) in covariates between groups should be < 0.1 (ideally < 0.05) post-matching; if not, increase caliper or include more covariates in propensity model. Report before/after balance via SMD plots.

3

After matching or weighting, compare outcomes between treatment groups using standard methods (t-test, regression, survival analysis) on the matched/weighted sample. The average treatment effect on the treated (ATT) is outcome difference in matched sample; if matching is successful (baseline balance achieved), this difference is unconfounded by measured covariates. Report: (1) covariate balance before and after matching (SMD tables or plots), (2) number of matched/unmatched pairs, (3) treatment effect (difference in outcomes), (4) confidence intervals and p-values, (5) sensitivity analysis to unmeasured confounding (e.g., Rosenbaum bounds). Examine common support: proportion of treated with propensity scores above minimum untreated score and below maximum untreated score; poor overlap invalidates estimates. Use regression adjustment on the matched sample as additional confounder control (doubly robust estimation).

Common pitfalls

  • Matching on the outcome: biases estimates toward zero (reduces power). Propensity score modeling should include confounders (variables predicting both treatment and outcome), not mediators or consequences.
  • Failing to check overlap: proceeding with matching when treated/control propensity distributions don't overlap. Always check and report common support; estimates outside overlap are unreliable extrapolations.
  • Ignoring covariate balance: assuming matching worked without examining standardized mean differences. Always report balance diagnostics (SMD plots, love plots); imbalance indicates need for alternative approach.
  • Analyzing full sample instead of matched sample: if you match then analyze all original data, you ignore the matching (defeats purpose). Analyze matched sample only, or use weighting on full sample.
  • P-hacking with multiple propensity models: fitting different propensity models and selecting the one yielding best balance or strongest treatment effect. Specify propensity model a priori based on theory; use cross-validation.
  • Causal claims without temporal precedence: propensity score reduces confounding by measured covariates, but causality requires treatment before outcome. Reverse causality still threatens validity.

Frequently asked questions

What is the propensity score, and why does matching on it work?

The propensity score is the probability of receiving treatment given observed covariates, estimated via logistic regression: P(T=1|Covariates). Two individuals with the same propensity score have similar baseline characteristics (in expectation) even if treated differently. Matching on propensity scores creates comparable groups: in the matched sample, treatment status is independent of covariates (conditional independence achieved). This is equivalent to having made randomization at baseline, enabling causal inference. The propensity score is a scalar (one number per person), reducing matching from many dimensions to one (curse of dimensionality resolved).

What is the difference between propensity score matching and inverse probability of treatment weighting (IPTW)?

Propensity score matching: pairs treated individuals with untreated individuals having similar propensity scores; analyzes only matched pairs, discarding unmatched individuals. IPTW (inverse probability weighting): weights all individuals by inverse propensity scores (weight treated by 1/p, controls by 1/(1−p)), creating a pseudo-population where treatment is independent of covariates. IPTW retains all data, improving efficiency and power. Both achieve covariate balance; IPTW is statistically more efficient. Trade-off: IPTW can produce extreme weights if some individuals have very high or very low propensity scores, increasing variance; matching avoids this but loses data.

How do I assess covariate balance, and what is acceptable?

Calculate standardized mean difference (SMD) for each covariate: SMD = (mean_treated − mean_control) / pooled SD. SMD < 0.10 is generally accepted as 'balanced'; < 0.05 is excellent. Plot SMDs pre- and post-matching (love plot): before/after balance visible at a glance. Calculate SMD for all confounders; if any remain imbalanced (SMD > 0.10) after matching, consider: widening caliper (match more loosely), including interactions in propensity model, or using alternative methods (stratification, regression adjustment). Never judge balance by p-values (p-values are sample size-dependent); always examine SMD numerically and graphically.

Can propensity score matching address unmeasured confounding?

No. PSM reduces bias from measured confounders (those you've collected data on). Unmeasured confounders (e.g., patient motivation, unmeasured disease severity, hidden selection mechanisms) remain uncontrolled. Conduct sensitivity analysis to assess how sensitive conclusions are to unmeasured confounding (Rosenbaum bounds: how strong an unmeasured confounder must be to change conclusions). Always discuss unmeasured confounding limitations. Consider instrumental variables or natural experiments if unmeasured confounding is a major concern.

Related methods

Sources

  1. 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.

    Open DOI reference
    Copied.
  2. Austin, P. C. (2011). An introduction to propensity score methods for reducing the effects of confounding. Multivariate Behavioral Research, 46(3), 399–424.

    Open DOI reference
    Copied.
  3. Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology, 66(5), 688–701.

    Open DOI reference
    Copied.