In the ideal world of clinical research, every question would be answered by a Randomized Controlled Trial (RCT). By randomly assigning patients to a treatment or a control group, researchers ensure that the only systematic difference between the two groups is the treatment itself. Any difference in outcome can then be confidently attributed to that treatment.
But the real world is rarely ideal. RCTs are expensive, time-consuming, and often ethically impossible. Instead, researchers are frequently left with "observational data"—records of what happened to patients in the normal course of clinical practice. The problem? Observational data is riddled with selection bias.
The Problem of Selection Bias
Imagine comparing two surgical techniques for colon cancer: a less aggressive partial colectomy (PC) and a more extensive extended hemicolectomy (HC). If you simply look at the data, you might find that patients who received the less aggressive surgery lived longer. But does the surgery cause the survival, or did surgeons simply choose the less aggressive option for patients who were already healthier?
As Xu Guan, Zhixun Zhao, and colleagues explored in their study of transverse colon cancer, these baseline differences (covariates) like age, tumor size, and comorbidities act as "confounders." They muddy the waters, making it impossible to tell if the treatment or the patient's initial health drove the outcome.
The Solution: Propensity Score Matching
Propensity Score Matching (PSM) is a statistical "time machine" that attempts to reconstruct the conditions of an RCT from messy, non-randomized data.
The core intuition is simple: Find statistical twins.
Instead of comparing all treated patients to all untreated patients, PSM looks for pairs of individuals—one from each group—who were equally likely to have received the treatment based on their baseline characteristics.
1. Calculating the Propensity Score
First, researchers use a statistical model (usually logistic regression) to calculate a single number for every patient: the Propensity Score. This score represents the probability (from 0 to 1) that a patient would receive the treatment, given everything we know about them before the treatment happened.
For example, in a study by Aakriti Gupta and Björn Redfors on statin use in COVID-19 patients, the propensity score would account for age, sex, and pre-existing conditions like diabetes. A 70-year-old male with hypertension might have a high propensity score (a high likelihood of being on a statin), while a healthy 20-year-old would have a low one.
2. Matching the Twins
Once every patient has a score, the researcher matches a "treated" patient with a "control" patient who has an nearly identical score.
By doing this, you create two groups that are balanced. In the matched "statin" and "non-statin" groups, you will have the same proportion of 70-year-olds, the same proportion of diabetics, and so on. As Fan Yang and Nan Luo demonstrated when comparing dialysis outcomes, this process effectively "cancels out" the observed variables that would otherwise bias the results.
Why It Works: Balancing the Scales
The power of PSM lies in its ability to reduce a high-dimensional problem (matching on dozens of different traits) into a one-dimensional problem (matching on a single score).
When the groups are balanced, the "selection" part of the bias is removed. If a patient with a 0.75 propensity score received the treatment and their "twin" with a 0.75 score did not, the assignment between those two individuals is essentially random. When you aggregate thousands of such pairs, you have simulated a randomized trial.
This technique has allowed researchers to find clarity in complex clinical scenarios:
- Surgical Safety: Rui Wang and Xin Chen used PSM to show that on-pump vs. off-pump heart surgery resulted in similar risks of kidney injury once patient age and health were balanced.
- Drug Repurposing: Liam Rose and Allison Koenecke used it to identify that alpha-1 blockers were associated with lower COVID-19 mortality.
- Rare Diseases: In fields where patients are scarce, Mercedeh Ghadessi and Rui Tang have highlighted PSM as a vital tool for using historical controls to supplement small trial groups.
The "Observed" Limitation
It is important to remember that PSM is not magic. It can only balance observed covariates—the data points actually recorded in the database. If there is a "hidden" factor that influenced the doctor's decision but wasn't recorded (like a patient's "frailty" or "will to live"), PSM cannot account for it.
Despite this, PSM remains the gold standard for extracting causal truth from the vast wealth of observational data, providing a rigorous framework to ensure that when we see a difference in patient outcomes, we are looking at the effect of the medicine, not just the shadow of the patient's history.