Table 2.

Methods of Customizing a Pair-Matching Specification and Their Implications

Option	Benefits	Cautions
Matching on the covariates directly (e.g., Mahalanobis distance matching)	Can better balance the joint distribution of covariates; does not require an exposure model	May not perform well with many covariates, due to curse of dimensionality
Matching on the propensity score	Requires matching only on a single dimension; has theoretical balancing properties; tends to perform well empirically	Relies on specification of exposure model, pairs may not be close on covariates
Restrictions on closeness of matches	Can improve balance; yields close pairs; improves robustness to assumptions about outcome model	Dropping units decreases precision and can change the target population/estimand
Matching with replacement	Better balance than without replacement; good with small unexposed samples or when ratio of exposed to unexposed is high	Reusing units decreases precision; increases reliance on a few units
k:1 matching	Retains more units, thereby increasing precision	Balance can be worse