When evaluating the impact of a policy, a medical treatment, or a corporate strategy, the gold standard is the Randomized Controlled Trial (RCT). In an RCT, participants are randomly assigned to either a treatment group or a control group. Because assignment is random, both groups share nearly identical distributions of baseline characteristics. Any subsequent difference in outcomes can be confidently attributed to the treatment itself.
In the real world, however, running an RCT is often impossible, unethical, or prohibitively expensive. Instead, researchers must rely on observational data. The fundamental challenge of observational data is selection bias: the reasons why an entity receives a treatment are often systematically correlated with the outcome.
For example, in clinical settings, healthier patients might be selected for aggressive surgeries, while frailer patients receive conservative care. If we simply compare the survival rates of these two groups, the surgical treatment will appear highly effective, but this difference may just reflect the baseline health of the patients. Propensity Score Matching (PSM) is a statistical framework designed to resolve this dilemma by reconstructing a randomized experiment from observational data.
The Intuition of Propensity Scores
To eliminate selection bias, we want to compare treated and untreated units that are otherwise identical. If we have only one or two baseline characteristics (covariates)—such as age and biological sex—we can easily perform exact matching: find a treated patient and an untreated patient who are both 45-year-old males.
However, as the number of covariates grows to include income, location, comorbidities, and historical trends, we encounter the curse of dimensionality. Exact matches become impossible to find because the multi-dimensional space is too sparse.
Propensity score matching solves this by compressing high-dimensional baseline characteristics into a single scalar value: the propensity score. The propensity score is the predicted probability that a unit receives the treatment, given its observed covariates:
$$e(X) = P(D = 1 \mid X)$$
Where $D$ is the treatment indicator ($1$ for treated, $0$ for control) and $X$ is the vector of baseline characteristics. Typically, this probability is estimated using a logistic regression or machine learning model.
The core insight of this approach is that we do not need to match units on every individual covariate. Instead, if a treated unit and a control unit share the exact same propensity score, they had the same ex-ante probability of receiving the treatment. If one was treated and the other was not, that assignment becomes effectively random conditional on the observed characteristics.
Creating a Pseudo-Population
Once propensity scores are calculated, researchers use matching algorithms (such as nearest-neighbor matching or kernel matching) to pair treated units with control units that have similar scores. This process discards unmatched units and balances the covariates across the remaining groups, effectively creating a "pseudo-population" that mimics a controlled trial.
This methodology has found widespread application across diverse domains:
- Clinical Interventions: In a study on prostate cancer treatments, Tewari et al. utilized propensity scoring to balance baseline patient characteristics, allowing for a reliable comparison of long-term survival between conservative management, radiation therapy, and radical prostatectomy. Similarly, Zhang et al. applied PSM to evaluate surgical approaches for esophageal carcinoma, ensuring that patient selection factors did not skew recovery and survival outcomes.
- Real-World Evidence and Drug Repurposing: Rather than waiting years for new clinical trials, Yan et al. leveraged generative AI to identify drug candidates for Alzheimer's disease and then validated those candidates using propensity score matching on retrospective electronic health records of over 200,000 patients.
- Auditing Algorithmic Systems: Beyond medicine and economics, the framework can isolate bias in complex software ecosystems. For instance, Jiang et al. used PSM to audit YouTube's comment moderation system. By matching comments with similar predicted probabilities of being flagged based on content features, the authors could distinguish between direct political bias and indirect effects driven by the baseline distribution of policy-violating content.
By forcing the treatment and control groups to have the same distribution of propensity scores, the distribution of the underlying covariates $X$ becomes balanced between the two groups.
The Hybrid Powerhouse: PSM-DID
While standard PSM is highly effective, it suffers from a major limitation: it can only balance observed covariates. If there are unobserved variables that influence both treatment selection and the outcome (such as a firm's internal culture or a patient's unrecorded genetic factors), PSM cannot eliminate the resulting bias.
To overcome this, researchers frequently combine PSM with the Difference-in-Differences (DID) framework, creating a hybrid approach known as PSM-DID.
Standard DID Assumption:
Treated: o-------------------o (Parallel to control trend)
Control: o-------------------o
With Selection Bias (Violated Parallel Trends):
Treated: o-------------------/ (Different baseline characteristics
Control: o-------------------o cause divergent trajectories)
PSM-DID Solution:
1. Use PSM to select a control group with identical baseline characteristics.
2. Apply DID to compare their parallel trajectories over time.
The DID framework tracks units over time and measures the change in the treatment group relative to the change in the control group. DID relies on the parallel trends assumption—the premise that, in the absence of treatment, the trajectories of both groups would have moved in parallel. If the treatment and control groups have vastly different baseline characteristics, this assumption is likely to fail.
PSM-DID solves this by using PSM as a preprocessing step to select a control group whose baseline characteristics match the treatment group perfectly. Once the groups are aligned, the DID model is applied to calculate the treatment effect over time. This dual approach ensures that the estimation is robust against both selection bias on observables (via PSM) and time-invariant unobserved confounding (via DID).
This hybrid framework is the standard tool for evaluating regional policies and macro-economic interventions:
- Policy Evaluation: Sun et al. applied PSM-DID to evaluate whether the "Belt and Road Initiative" promoted economic growth among participating nations, ensuring that the participating and non-participating countries were structurally comparable prior to the initiative.
- Corporate and Market Dynamics: Cui deployed a multi-period PSM-DID framework to demonstrate how mergers and acquisitions boost enterprise-level productivity, ensuring that the observed improvements were not confounded by pre-existing traits that made certain firms more likely to engage in M&A.
- Environmental and Sustainability Initiatives: Li integrated PSM with difference-in-differences models to evaluate how inclusive green financial policies support regional green technology innovation, aligning treatment and control regions based on observable economic indicators before tracking their innovation trajectories.
Methodological Caveats
While propensity score matching is a powerful tool for causal inference, it is not a magical fix for flawed data.
First, the quality of the match depends entirely on the quality of the propensity score model. If critical confounding variables are omitted from the model, the scores will not reflect the true assignment probability, and selection bias will remain.
Second, as noted by Onishi et al. in their study of real estate markets, matching methodologies can sometimes introduce a subtle bias toward prime market segments where data is abundant, potentially reducing the generalizability of the findings across more diverse or niche populations.
When implemented carefully—ensuring common support between groups and verifying covariate balance after matching—propensity score methods allow researchers to extract clear, actionable causal insights from complex, non-randomized data.
Key Foundational Work
Short- and Long-Term Outcomes of Laparoscopic Versus Open Resection for Gastric Gastrointestinal Stromal Tumors (2016) — contributed 1 concept elements to this direction.
Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach (2007) — contributed 1 concept elements to this direction.