MerQuR+Statistics Glossary

Statistics Glossary → Propensity score matching (nearest neighbour, caliper)

Propensity score matching (nearest neighbour, caliper)concept

A method for observational data that builds comparable groups by matching treated and control units whose probability of receiving treatment (the propensity score) is similar.

The propensity score is the probability of being in the treatment group given a unit's observed characteristics: e(x) = P(T = 1 | X), usually estimated by logistic regression. Units with the same score are, on average, balanced on the observed characteristics. In nearest-neighbour matching each treated unit is paired with the control unit with the closest score (or with k units in 1:k matching); a caliper excludes pairs whose score difference exceeds a set threshold.

In matching with replacement a control can be used more than once, which reduces bias but increases variance. After matching, balance should be checked with standardized mean differences (usually aiming for below 0.1). The method generally estimates the average effect in the treated (ATT) and can only balance measured confounders; unmeasured confounders still cause bias. Dropping unmatched units narrows how far the result can be generalized.

An everyday example

A hospital studies whether people who join a smoking-cessation programme gain weight a year later. Participants are older and better educated, so a direct comparison would mislead. A propensity score is estimated from age, education and cigarettes per day, and each participant is matched with a non-participant whose score differs by less than 0.05; in the matched groups age and education are now balanced.

In MerQur

The 'Method' (nearest-neighbour matching, IPW, subclassification), 'Matching ratio (1:k)', 'Matching with replacement' and 'Caliper (SD units)' options in Propensity Score Analysis; the caliper excludes pairs whose score difference exceeds c × SD(PS).

Analyses where this concept appears

Propensity Score Analysis

Related concepts

Inverse probability weighting (IPW)Standardized mean difference (SMD) and covariate balanceAverage treatment effect (ATE, ATT, ATC)