Propensity Score Matching

Propensity score matching (PSM) is a statistical technique that pairs treated and untreated individuals who have similar estimated propensity scores, creating a matched sample in which the distribution of observed baseline covariates is balanced between groups. By constructing comparable groups from observational data, PSM reduces confounding and enables more credible estimates of treatment effects in the absence of randomization.

When randomized controlled trials are impractical, unethical, or insufficient to address a research question, observational studies provide the primary evidence base. However, treatment groups in observational data typically differ in ways that confound the treatment effect estimate. PSM addresses this by creating a matched comparison that mimics the balance achieved by randomization, at least for measured covariates. It has become one of the most commonly applied methods in clinical epidemiology, health services research, and social science for estimating causal effects from nonexperimental data.

The PSM process proceeds in several steps. First, the propensity score is estimated for each individual, typically using a logistic regression model in which treatment assignment is regressed on observed baseline covariates. The resulting predicted probability represents each individual's propensity for receiving the treatment.

Second, treated and untreated individuals are matched based on the proximity of their propensity scores. The most common approaches are nearest neighbor matching (pairing each treated individual with the untreated individual whose propensity score is closest), caliper matching (restricting matches to pairs whose scores differ by less than a specified caliper width, typically 0.2 standard deviations of the logit propensity score), and optimal matching (minimizing the total distance across all matched pairs). Matching can be done without replacement or with replacement (which allows an untreated individual to serve as a match for more than one treated individual).

Third, covariate balance must be assessed in the matched sample. Standardized mean differences are calculated for each covariate, with values below 0.10 generally indicating adequate balance. If balance is poor, the propensity score model should be respecified, for example by adding interaction terms or higher order terms.

Fourth, the treatment effect is estimated in the matched sample. For a binary outcome, this is often the difference in proportions or a conditional logistic regression. For a time to event outcome, a Cox model that accounts for the matched structure (for example, stratified on matched pair or with a robust variance) is often used. Standard errors must account for the matched structure of the data, either through paired analyses or cluster robust variance estimators.

PSM has important limitations. It can only balance measured covariates; unmeasured confounders remain a source of bias. Matching may discard a large proportion of the sample if treated and untreated groups have poor overlap (limited common support region), reducing generalizability and statistical power. Sensitivity analyses, such as Rosenbaum bounds, quantify how strong an unmeasured confounder would need to be to overturn the findings. King and Nielsen (2019) also argue that matching on the propensity score itself, rather than on the covariates, can increase imbalance, model dependence and bias, particularly when units are pruned after good balance has been reached; this is one reason some researchers prefer covariate based matching methods such as Mahalanobis distance or coarsened exact matching.

Use PSM when the goal is to estimate a causal treatment effect from observational data and when a large set of measured covariates is available. It is most effective when the treated and untreated groups overlap substantially in their covariate distributions. PSM is not appropriate when overlap is minimal, as many treated individuals will be unmatchable. Poor overlap also produces extreme weights in inverse probability weighting, so no method can supply comparisons the data do not contain; options include restricting the analysis to the region of common support and stating the narrower population, or weighting approaches such as overlap weights. PSM is unnecessary in randomized trials where balance is achieved by design.

A common error is matching on propensity scores without checking covariate balance afterward, assuming that matching automatically produces balanced groups. Another mistake is including variables in the propensity score model that occur after treatment assignment, which can introduce bias. Researchers sometimes match with too wide a caliper, accepting poor matches that leave substantial covariate imbalance, or too narrow a caliper, discarding large portions of the sample and reducing power. Finally, failing to conduct a sensitivity analysis for unmeasured confounding leads to overconfident causal claims.

A study uses electronic health records to compare one year readmission rates between 3,200 heart failure patients who received a clinical pharmacist consultation at discharge and 9,800 who did not. Before matching, consulted patients are older (mean 72 versus 68 years), have more comorbidities, and use more medications. A propensity score model with 18 baseline covariates is estimated, and 1:1 nearest neighbor matching with a caliper of 0.2 standard deviations of the logit of the propensity score yields 2,850 matched pairs. After matching, all standardized mean differences are below 0.08. The one year readmission rate is 28.4% in the consultation group versus 33.1% in the matched control group (risk difference = 4.7 percentage points, 95% CI: 2.3 to 7.1, p < 0.001).