What Match Loss Means
A caliper restriction sets a maximum allowed distance between a treated unit and a control unit for them to be eligible to match. Tightening the caliper removes some candidate links. It does not necessarily remove the same number of units from the final matched sample: a unit may have several eligible partners, and an assignment may still match it after other links disappear. Conversely, a unit that has at least one candidate partner can still be left unmatched if that partner is needed elsewhere.
The distinction matters when evaluating a matching design. A smaller caliper may improve closeness on the chosen distance measure, but it can also reduce the number of feasible matches and alter which treated units and controls remain. The resulting matched sample may therefore represent a narrower or different portion of the original study population. Link counts alone do not describe that change.
The distance might be an absolute difference in propensity scores, an absolute difference in logit propensity scores, or another prespecified matching distance. A threshold is meaningful only with its distance definition and scale stated. For example, a caliper of \(0.2\) on one scale is not automatically equivalent to \(0.2\) on another.
This is a design-stage restriction, not by itself a causal identification assumption. A caliper can enforce a chosen degree of closeness, but it does not establish that measured covariates suffice to control confounding or that treatment assignment is unconfounded. Those assumptions must be considered separately when drawing causal conclusions.
Why Link Loss Is Monotone but Match Loss Is Not a Simple Count
If \(c_1 < c_2\), every link eligible under \(c_1\) is also eligible under \(c_2\). Thus the smaller threshold's edge set is a subset of the larger threshold's edge set. The number of eligible links cannot increase when the caliper tightens. The maximum number of one-to-one pairs also cannot increase, because every matching feasible under the smaller edge set remains feasible under the larger one.
This gives a useful short justification. A one-to-one matching is a set of eligible links in which no unit appears in more than one link. If \(M_1\) is feasible at \(c_1\), all its links are also eligible at \(c_2\), so \(M_1\) is feasible at \(c_2\). The set of feasible matchings at \(c_1\) is therefore contained in the set at \(c_2\). Taking the largest feasible matching in each set yields:
where \(E(c)\) is the eligible-link set and \(m_{\max}(c)\) is the maximum number of one-to-one pairs. This result concerns the maximum possible matched-pair count. A particular algorithm's selected pairs can still change as the threshold changes, especially when it optimizes distance or another criterion among feasible solutions. Nor does the result imply that every unit retained at a tighter threshold must appear in a particular optimal solution.
Worked Example: Counting Links at Three Calipers
Consider four hypothetical treated units, \(T1\) through \(T4\), and four hypothetical controls, \(C1\) through \(C4\). The entries below are distances, in the chosen matching-distance units.
| C1 | C2 | C3 | C4 | |
|---|---|---|---|---|
| T1 | 0.05 | 0.18 | 0.31 | 0.48 |
| T2 | 0.14 | 0.09 | 0.22 | 0.39 |
| T3 | 0.27 | 0.19 | 0.08 | 0.16 |
| T4 | 0.42 | 0.34 | 0.21 | 0.11 |
At \(c=0.30\), count entries no greater than \(0.30\). The treated-unit link counts are \(2, 3, 4,\) and \(2\), respectively, for \(2+3+4+2=11\) eligible links. At \(c=0.20\), the counts are \(2, 2, 3,\) and \(1\), giving \(8\) links. At \(c=0.12\), only \(T1\)-\(C1\), \(T2\)-\(C2\), \(T3\)-\(C3\), and \(T4\)-\(C4\) remain, giving \(4\) links.
Thus tightening from \(0.30\) to \(0.20\) removes \(11-8=3\) candidate links, or \(3/11 \approx 27.3\%\) of the original links. Tightening from \(0.20\) to \(0.12\) removes \(8-4=4\), or \(50.0\%\) of the links then available. These are link-loss percentages, not percentages of units lost.
For one-to-one matching, the four diagonal links remain eligible at every listed threshold. They provide four distinct pairs, so the maximum matched-pair count is four at all three thresholds. For example, their total distance is \(0.05+0.09+0.08+0.11=0.33\). Link loss is substantial, but this particular example has no loss of matched units under those thresholds. The candidate partner sets nonetheless contract, particularly for \(T4\), whose only eligible partner at \(c=0.12\) is \(C4\).
When Candidate Links Compete
A treated unit can have an eligible link and still be unmatched. This happens when eligible links compete for the same control, or when the matching objective selects a different set of links. The following example separates the existence of a candidate from the feasibility of retaining every unit.
Worked Example: A Feasible Link Does Not Guarantee a Match
Suppose three hypothetical treated units and three hypothetical controls have these distances:
| C1 | C2 | C3 | |
|---|---|---|---|
| T1 | 0.04 | 0.16 | 0.34 |
| T2 | 0.08 | 0.19 | 0.40 |
| T3 | 0.23 | 0.31 | 0.45 |
At \(c=0.45\), all nine links are eligible, so three one-to-one pairs are feasible. At \(c=0.25\), the eligible links are \(T1\)-\(C1\), \(T1\)-\(C2\), \(T2\)-\(C1\), \(T2\)-\(C2\), and \(T3\)-\(C1\): five links. There are only two controls, \(C1\) and \(C2\), available to match \(T1\) and \(T2\), while \(T3\)'s only eligible partner is \(C1\). A three-pair matching is impossible because \(C3\) has no eligible links. A two-pair matching is feasible.
If the procedure first maximizes the number of pairs and then minimizes their total distance, it selects \(T1\)-\(C1\) and \(T2\)-\(C2\). Their total distance is \(0.04+0.19=0.23\). The alternative \(T3\)-\(C1\) and \(T1\)-\(C2\) also has two pairs, but its total is \(0.23+0.16=0.39\). The distance-minimizing choice therefore leaves \(T3\) unmatched even though \(T3\) has an eligible link.
At \(c=0.10\), only \(T1\)-\(C1\) and \(T2\)-\(C1\) remain eligible. They compete for the same control, so at most one pair can be formed. If the objective is to minimize distance among one-pair solutions, \(T1\)-\(C1\), at distance \(0.04\), is selected; \(T2\) is then unmatched. The maximum pair count has fallen from three to two to one as the caliper tightens. The identity of retained units depends on both the eligible-link graph and the matching objective.
Match Loss in Full Matching
In full matching, a matched set contains at least one treated unit and at least one control unit. A caliper removes links that could support such sets. Consequently, tightening the caliper can make a previously feasible set impossible, even if the total number of remaining units appears adequate. If ratio limits are imposed, a set must also satisfy those limits; an eligible edge alone does not guarantee that all units can be placed in feasible sets.
Worked Example: A Tight Caliper Makes a Treated Unit Infeasible
Consider two hypothetical treated units, \(T1\) and \(T2\), and two controls, \(C1\) and \(C2\). Their distances are \(d_{T1,C1}=0.04\), \(d_{T1,C2}=0.18\), \(d_{T2,C1}=0.09\), and \(d_{T2,C2}=0.16\). Suppose the design requires one treated unit per control in each matched set.
At \(c=0.20\), all four links are eligible. One feasible full-matching set contains both treated units and both controls, with a treated-to-control ratio of \(2:2\), which is \(1:1\). At \(c=0.08\), only \(T1\)-\(C1\) remains eligible: \(0.04 \leq 0.08\), while \(0.09\), \(0.16\), and \(0.18\) all exceed \(0.08\). Neither \(T2\) nor \(C2\) has an eligible partner. The set containing all four units is no longer feasible, and the only possible matched set is \(T1\) with \(C1\). Thus one treated unit and one control are lost from the matched sample, even though the remaining set still satisfies the specified \(1:1\) ratio.
This example assumes that units without eligible partners may be left out of the matched sample. If a design instead requires every unit to be matched, the tightened-caliper problem has no feasible solution. Reporting which convention is used is part of interpreting match loss.
Diagnostics, Alternatives, and Limitations
A useful caliper assessment reports more than the threshold and the final sample size. For each candidate threshold, examine the following:
- Eligible-link count: how many candidate treated-control links remain?
- Unit-level degree: how many eligible partners does each treated unit and each control have? A degree of zero identifies a unit with no possible partner.
- Feasible matched count: how many treated units, controls, or pairs can the chosen matching structure retain, given any no-reuse, set-size, or ratio restrictions?
- Composition: which units or covariate profiles disappear as the threshold tightens? Summarize this separately for treated and control units.
- Balance and distance: among the selected matches, inspect relevant covariate balance and the distances actually used. A small caliper does not guarantee good balance on covariates not represented adequately by the matching distance.
These checks help distinguish a sensible restriction from a failing design. A rapid rise in zero-degree units, a sharp drop in the maximum feasible matched count, or selective loss of an important treated subgroup signals that the chosen threshold has a substantial sample-composition cost. There is no universal acceptable percentage of match loss: the consequences depend on the study question, the target population, the matching structure, and the units excluded.
The main alternatives answer different design questions. A looser caliper retains more candidate links but permits larger distances. Nearest-neighbor matching without a caliper can assign a partner even when that partner is relatively distant; that avoids a hard cutoff but does not ensure closeness. Constrained optimal matching can choose a globally coordinated assignment among eligible links, rather than making each choice in isolation, but it cannot restore links excluded by the caliper. Changing the distance measure can alter which links are removed, so comparisons should state both the measure and threshold.
Finally, a tighter threshold can change the population to which a matched-sample estimate pertains. If units with certain covariate profiles are disproportionately removed, the retained sample is selected by feasibility under the design. A result for that sample should not automatically be described as an effect for all treated units or the original study population. Match loss is therefore not merely an algorithmic inconvenience; it informs what population the design can compare.
Common Mistake
A related error is to infer that every unit with at least one eligible link will be matched. As the competing-link example shows, a unit can lose access to a control because the assignment uses that control in another pair. The correction is to assess the feasible matching problem under the actual restrictions and objective, not just to count each unit's candidate partners.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- If a caliper is tightened from \(c=0.25\) to \(c=0.15\), can the number of eligible links increase? Explain.
- In the first worked example, how many links are eligible at \(c=0.20\), and what proportion of the \(c=0.30\) links have been lost?
- At \(c=0.25\) in the competition example, why can three treated units not all be matched one-to-one?
- In the full-matching example, what happens at \(c=0.08\) if every unit is required to be matched?
Solutions. (1) No. Every link with distance at most \(0.15\) also has distance at most \(0.25\), so tightening can only preserve or remove links. (2) There are \(8\) links at \(c=0.20\), compared with \(11\) at \(c=0.30\). The proportion lost is \((11-8)/11=3/11 \approx 27.3\%\). (3) At that threshold, \(C3\) has no eligible links, so only two controls can be used; no one-to-one assignment can form three pairs. (4) The problem is infeasible: \(T2\) and \(C2\) have no eligible partners, so not every unit can be placed in a matched set.