Choose the Design Before the Algorithm
Choosing a matching method is not simply choosing the algorithm that produces the smallest distances. A method must fit the study’s design: whether units may be reused, how many controls should be assigned to each treated unit, which distances are acceptable, and whether the available units can form the required matched sets. These choices also affect which units contribute to the analysis and therefore the population to which the result applies.
Matching is used to construct comparable groups from treated and control units using observed covariates. Its causal interpretation still depends on the relevant identification assumptions, including conditional exchangeability and adequate overlap; matching does not remove bias from unmeasured confounding or create comparable units where none exist. The method should be chosen using design information and covariates, without selecting it to obtain a preferred outcome estimate.
Suppose \(d_{ij}\) is the distance between treated unit \(i\) and control unit \(j\), where smaller values indicate greater similarity under the chosen distance measure. The distance definition must be appropriate for the study, but the choice of distance for mixed-scale covariates is a separate issue. Here, assume a distance has already been specified.
Translate Study Requirements into Constraints
Begin by asking what the design must accomplish, rather than by naming an algorithm. A requirement that each treated unit receive exactly one distinct control suggests one-to-one matching without replacement, provided there are enough feasible controls. A requirement to allow the same control to serve several treated units calls for matching with replacement. A requirement that every matched set respect a maximum distance calls for a caliper or another explicit eligibility restriction.
Ratio requirements need the same precision. “Use two controls per treated unit” usually means a fixed 1:2 treated-to-control ratio for each matched treated unit, not merely two controls somewhere in the overall sample. If that exact structure cannot be formed, the design must state whether to leave units unmatched, relax the ratio, permit reuse, or choose a different structure. Full matching is useful when variable ratios are acceptable: each matched set contains at least one treated unit and at least one control, but the ratio can vary across sets.
A distance objective also needs to be explicit. Greedy nearest-neighbor matching chooses the closest currently eligible control for a treated unit, so the result can depend on the order of treated units. Optimal pair matching instead minimizes the total distance across a feasible assignment, subject to its constraints. Neither objective automatically enforces good covariate balance in every relevant sense, and neither overrides design restrictions such as a caliper.
Worked Example: Why Greedy Nearest Neighbor Can Miss the Best Pairing
Suppose two treated units, \(T_1\) and \(T_2\), must each be paired with a different control, without replacement. The hypothetical distances are:
| \(C_1\) | \(C_2\) | |
|---|---|---|
| \(T_1\) | 1 | 4 |
| \(T_2\) | 2 | 100 |
If greedy nearest-neighbor matching processes \(T_1\) first, it assigns \(T_1\) to its nearest control, \(C_1\), at distance 1. The only remaining control for \(T_2\) is \(C_2\), at distance 100. The total distance is
The alternative assignment pairs \(T_1\) with \(C_2\) and \(T_2\) with \(C_1\). Its total distance is
The second assignment is globally better for the stated objective because \(6<101\). Optimal pair matching selects it. This does not mean greedy matching is always poor: it means that a locally closest choice can use a control that is much more valuable to another treated unit. The example assumes both assignments are allowed and that total distance is the objective; a caliper or another constraint could change the feasible assignments.
When Distance Restrictions Matter
A caliper sets a maximum allowed distance for a match. A candidate pair with \(d_{ij}\) greater than the threshold is ineligible, even if it is the closest available pair. This guards against accepting very dissimilar matches, but it can reduce the number of matched units. The design decision is therefore a trade-off: stricter restrictions can improve the similarity of retained pairs while changing the composition and size of the matched sample.
Worked Example: A Caliper and Control Reuse Change Feasibility
Consider two treated units and two controls, with hypothetical distances \(d(T_1,C_1)=0.08\), \(d(T_1,C_2)=0.16\), \(d(T_2,C_1)=0.11\), and \(d(T_2,C_2)=0.30\). Set a caliper of 0.12. A pair is eligible only if its distance is at most 0.12.
For \(T_1\), only \(C_1\) is eligible because \(0.08\leq0.12\), whereas \(0.16>0.12\). For \(T_2\), only \(C_1\) is eligible because \(0.11\leq0.12\), whereas \(0.30>0.12\). Thus, the eligible-pair set is \(\{(T_1,C_1),(T_2,C_1)\}\).
If controls cannot be reused, both treated units cannot be matched: there is only one eligible control. At most one treated unit can be paired, so the design must accept match loss or revise a prespecified constraint. If reuse is allowed, both treated units can be paired to \(C_1\), but the same control contributes to two matches. Reuse satisfies the caliper and retains both treated units, but it does not create two distinct controls or two independent copies of the control information.
This example shows why feasibility must be checked against the complete set of restrictions, not just by counting the total numbers of treated and control units. There are two controls in the sample, but \(C_2\) cannot meet the distance restriction for either treated unit.
When Ratios Vary or Counts Are Unequal
Fixed-ratio matching is attractive when a study needs a uniform structure, such as one treated unit matched to one control or one treated unit matched to two controls. Without replacement, a fixed ratio can be infeasible because the sample has too few eligible controls or because the eligible-pair pattern does not permit the required assignments. Allowing replacement may make the requested ratio attainable, but it changes how often controls are used and should be an explicit design decision.
Full matching offers a different structure. It groups units into matched sets, each containing treated and control units, while allowing the ratio to vary between sets. This can use units efficiently when the treated and control sample sizes differ, but it does not guarantee that every unit can be included under additional restrictions such as calipers or ratio limits. Nor should a variable-ratio set be mistaken for a fixed-ratio design: the number of controls per treated unit may differ from one subclass to another.
Worked Example: Choosing Full Matching When Ratios May Vary
Suppose a hypothetical study has three treated units, \(T_1,T_2,T_3\), and two controls, \(C_1,C_2\). Distances are \(d(T_1,C_1)=0.10\), \(d(T_2,C_2)=0.20\), and \(d(T_3,C_2)=0.30\); assume these are the relevant within-set distances and there are no additional restrictions.
A one-to-one design without replacement cannot match all three treated units: two distinct controls can support at most two treated-control pairs. A full-matching allocation can instead form one set containing \(T_1,C_1\) and another containing \(T_2,T_3,C_2\). The first set has a 1:1 treated-to-control ratio, and the second has a 2:1 ratio. Every unit is included, and each set contains both treatment groups.
If the within-set objective is the sum of treated-control distances, the stated allocation has total distance
This sum uses the three stated treated-control distances: \(T_1\) to \(C_1\), and \(T_2\) and \(T_3\) to \(C_2\). The example establishes that a variable-ratio structure is feasible and uses all units; it does not establish that this allocation is optimal among all possible full matchings. If a maximum set ratio or caliper were imposed, feasibility and the selected allocation would need to be checked again.
A Practical Selection Rule
The following sequence makes the choice transparent. It does not replace checking the output; it makes the design decisions explicit before matching.
Specify which treated or control units may be excluded. A method that retains fewer units may produce a result for a narrower matched population.
State whether controls may be reused and whether each treated unit needs a fixed number of controls. If variable ratios are acceptable, consider a full-matching structure.
If so, define a caliper and treat pairs beyond it as ineligible. Anticipate that this may leave units unmatched.
Use nearest-neighbor matching when sequential closest eligible choices fit the design. Prefer optimal matching when the design calls for minimizing total distance across a global assignment.
Inspect unmatched units, reuse counts, achieved ratios, distances, and covariate balance. Revise a design constraint only with a defensible, documented rationale.
The reason this sequence matters can be stated directly: an algorithm can optimize only over assignments it is allowed to make. If \(\mathcal{A}\) is the set of assignments satisfying the design restrictions, an optimal assignment minimizes total distance over \(\mathcal{A}\). Tightening the restrictions can only leave \(\mathcal{A}\) unchanged or make it smaller. If no assignment satisfies them, the problem is infeasible; no algorithm can produce a valid solution without relaxing a restriction or changing the design.
This expression also clarifies why “optimal” does not mean best in every respect. It means optimal for the specified objective over the specified feasible assignments. A method can have a smaller total distance but still have undesirable balance on a covariate, discard a substantively important group of units, or fail a practical ratio requirement. Those features must be assessed separately.
Diagnostics, Comparisons, and Limitations
| Design need | Method to consider | Important check |
|---|---|---|
| Sequential closest eligible pairs; reuse rule specified | Nearest-neighbor matching | Order dependence and whether later units are left without eligible controls |
| No pair may exceed a chosen distance | Caliper matching, combined with a selection rule | Number and characteristics of unmatched units |
| Minimize total pair distance under one-to-one or other stated constraints | Optimal matching | Feasible assignment, total objective, and individual distances |
| Variable treated-control ratios across sets are acceptable | Full matching | Set composition, ratio restrictions, and whether required units can be included |
A matching result should be checked against the design specification. Confirm that no forbidden reuse occurred, the requested ratios were achieved, and no pair violated a caliper. Then examine covariate balance and the distribution of match distances, not only their average or total. Also report how many units were excluded and whether exclusion was concentrated in particular covariate regions. A design that meets its formal constraints can still yield poor balance or a highly selected matched sample.
These methods are alternatives for constructing assignments, not substitutes for one another in every setting. A caliper is a restriction that can accompany nearest-neighbor or optimal assignment. Full matching changes the matched-set structure and permits variable ratios. Greedy nearest neighbor is often straightforward, but its order can matter; optimal matching addresses the global objective but cannot compensate for a poor distance definition or inappropriate constraints. All approaches are limited by observed covariate information and available overlap.
- Pitfall: “Optimal” means the best causal comparison. This is wrong because optimal refers to a stated mathematical objective, such as total distance, not to causal validity or balance on every feature. Check the design, balance, and target population separately.
- Pitfall: A caliper guarantees that every treated unit is matched well. It guarantees only that accepted pairs meet the chosen distance limit. Some units may have no eligible partner and be excluded.
- Pitfall: Two controls in the data are enough for two treated units under no replacement. Not necessarily. The eligible-pair pattern may leave both treated units competing for the same control, as in the caliper example.
- Pitfall: Reuse creates extra distinct controls. Reuse assigns the same control to multiple matches; it does not increase the number of unique control units.
Check Your Understanding
Answer each question using the stated design constraints. Brief solutions follow the questions.
- A study requires one distinct control for each treated unit and seeks to minimize the sum of pair distances. Which assignment approach directly matches that global objective?
- With a caliper, two treated units each have only the same single control as an eligible partner. Can both be matched without replacement? What design change could permit both matches?
- A design has more treated than control units, and it permits different treated-control ratios across matched sets. Which structure should be considered, and what must still be checked?
- Why does minimizing total distance not establish that the resulting matched sample has adequate covariate balance?
Solutions. (1) Optimal matching directly minimizes the total distance over feasible assignments; the one-to-one, no-reuse rule defines the feasible set. (2) No. The single eligible control cannot be assigned to both treated units without replacement. Permitting control reuse could allow both matches, if reuse is substantively acceptable and the caliper is met. (3) Consider full matching, which permits variable ratios across sets. Check set composition, ratio restrictions, caliper eligibility if imposed, and whether all required units can be included. (4) The objective is only the specified sum of distances. It does not guarantee small distances on every covariate or adequate balance; those require separate diagnostics.