Propensity Score Matching

Propensity Score Matching: Concept Architecture

Orientation and learning roadmap

This page moves from the study question and matching design to the conditions required for a causal interpretation. It then shows how to assess balance and overlap, interpret an estimate from matched patients, and use the result in an economic evaluation.

Study question and design

Start by specifying the treatment contrast, eligibility criteria, treatment assignment time, outcome window, and population whose effect matters. These choices determine which variables precede treatment and whether the available comparator patients can support the intended comparison.

The propensity score is the probability of receiving treatment conditional on measured baseline covariates:

$$e(X_i)=P(T_i=1\mid X_i),$$

where $T_i$ indicates treatment and $X_i$ contains pretreatment characteristics of patient $i$. Matching pairs or groups treated and comparator patients with similar estimated scores, sometimes also requiring exact matches on selected factors or a maximum score distance (a caliper). The original propensity-score result explains why conditioning on a suitable balancing score can balance the measured covariates when its assumptions hold.

Causal conditions and covariate choice

A matched comparison can estimate a causal effect only when treatment and comparator potential outcomes are exchangeable conditional on the measured baseline covariates, eligible patients have a positive probability of either treatment, and treatment versions and outcomes are defined consistently. These conditions cannot be established by a good-looking balance table alone, because unmeasured confounding and violations of consistency remain possible.

Covariates should be selected using the clinical and causal structure of treatment assignment and outcomes, with particular attention to baseline prognostic variables and confounders. Do not adjust for variables caused by treatment as though they were baseline confounders, since this can introduce bias or remove part of the effect of interest. The propensity model's ability to predict treatment is less important for this purpose than covariate balance after matching.

Matching choices and diagnostics

Specify the score-estimation method, matching ratio, distance scale, caliper, replacement rule, and handling of patients without matches before inspecting outcomes. These choices determine which patients contribute to the estimate and whether the comparison remains credible.

Assess the distributions of scores in both groups, the number of unmatched patients, and balance on each important baseline covariate after matching. Standardised mean differences, variance ratios, and distribution plots are useful diagnostics; a small propensity-score difference by itself does not prove that individual covariates are balanced. A common descriptive target is an absolute standardised mean difference below 0.1, but this heuristic is not a universal validity threshold and should be considered alongside subject-matter importance and fuller distribution checks.

Estimand and worked interpretation

Matching should be interpreted against the population actually retained by the design. One-to-one matching of treated patients to controls commonly targets the average treatment effect among the matched treated patients, which need not equal the effect among all treated or all eligible patients.

Suppose 100 treated and 140 comparator patients are eligible, but only 80 treated patients can be matched to 80 comparators with adequate overlap. If 16 of the 80 matched treated patients and 24 of their 80 matched comparators have an adverse event, the matched risks are $16/80=0.20$ and $24/80=0.30$, giving a descriptive risk difference of $0.20-0.30=-0.10$. Under the stated causal assumptions and adequate balance, this suggests ten fewer events per 100 patients in the matched treated population; the 20 excluded treated patients are outside that estimate's direct target. Inference should account for matched sets and any reuse or weighting rather than treating observations as an ordinary independent random sample.

Role in health economic evaluation

Matched observational cohorts can supply comparative evidence on clinical outcomes, resource use, costs, and quality-adjusted life-years when randomised evidence is unavailable or less representative of the decision setting. The analysis should carry forward the matched population, follow-up, outcome definitions, and residual uncertainty into the economic model rather than silently generalising to patients with no comparator support.

Propensity score matching addresses observed baseline differences in treatment selection; it does not itself solve missing outcome data, differential follow-up, time-related biases, or unmeasured confounding. Analysts should explain whether the treatment-effect estimate, the cost and utility estimates, and the decision population refer to the same people, and assess sensitivity to plausible remaining bias.

Distinctions and evidence base

The propensity score is a patient-level assignment probability, whereas propensity score matching is one design strategy that uses that probability to construct a comparison group. Weighting, stratification, and regression adjustment use different mechanisms and may target different populations even when based on the same baseline data.

The foundational balancing-score argument is developed by Rosenbaum and Rubin (1983). Practical matching design and balance assessment are discussed in Austin (2011) and Austin's balance-diagnostics study; the NICE Decision Support Unit observational-data guidance places comparative patient-level methods in the context of technology appraisal.