Propensity matching or weighting?
The decision
A propensity score describes treatment assignment given measured baseline covariates. Matching constructs comparable treated and untreated groups; weighting changes each observation’s contribution. First specify the intervention contrast, eligibility, time origin and population whose average effect matters.
What overlap looks like
When matching is useful
Matching can make the compared groups concrete and reveal treated participants without suitable controls. Nearest-neighbor matching to treated participants often targets an effect among the treated who remain represented. Specify the matching ratio, caliper, replacement, exact-matching variables and how unmatched participants alter the target. Other matching designs can target other populations.
When weighting is useful
Weights can target an average treatment effect, an effect among the treated, or an overlap population, depending on their construction. Ordinary ATE inverse-probability weights can become extreme when treatment is nearly deterministic. Stabilization, truncation and overlap weighting are different choices, not interchangeable fixes; explain their effect on the estimand.
What both require
- Choose baseline confounders using substantive knowledge and a causal argument. Avoid automatically adjusting for post-treatment variables, colliders or every available predictor.
- Inspect overlap and covariate balance after adjustment, including important nonlinear terms and interactions. A propensity-model AUC is not a balance diagnostic.
- Use uncertainty methods appropriate to the estimator, matching dependence and estimated weights. A naive unweighted regression standard error is not automatically valid.
- Prespecify outcome analysis and sensitivity to residual confounding. Good measured balance cannot test the absence of unmeasured confounding.
Example
A treatment is rarely offered to frail patients. Large ATE weights and many unmatched frail participants both reveal limited information for an all-patient effect. Restricting to a clinically justified overlap population may be more credible, but the investigator must accept the changed population.
Record in the workspace
Record ATE, ATT or another target, covariate timing, overlap and balance criteria, matching or weight rules, uncertainty method and sensitivity analyses. Compare these approaches where more than one can answer the same defensible question.
What to watch out for
Matching and weighting address measured confounding under assumptions; neither removes unmeasured confounding or makes a lack of overlap disappear. Trimming and discarded matches can change the target population.
- An Introduction to Propensity Score Methods for Reducing the Effects of Confounding in Observational Studies (opens in a new tab) · Multivariate Behav Res · doi 10.1080/00273171.2011.568786
- Principles of confounder selection (opens in a new tab) · Eur J Epidemiol · doi 10.1007/s10654-019-00494-6