Resolving Ties in Nearest-Neighbor Matching

Why Tie-Breaking Matters

Nearest-neighbor matching selects controls that are closest to treated units under a chosen distance. Sometimes two or more eligible controls have exactly the same smallest distance to a treated unit. The matching rule identifies all of them as nearest; it does not, by itself, say which one to select when the required number of matches is smaller than the tied group.

A tie may seem like a minor computing detail, but resolving it determines which people enter the matched sets. Under matching without replacement, choosing one control also removes it from the pool for later treated units. A tie rule can therefore affect not only the current match but also subsequent matches. It should be stated explicitly and applied consistently, rather than left to software defaults or the order in which rows happen to appear.

Definition: Let $e_i$ denote the propensity score for treated unit $i$, and $e_j$ the score for control $j$. For a specified distance $d_{ij}$, the tied nearest controls for $i$ are the eligible controls whose distances equal the smallest eligible distance: $$ \mathcal{N}_i=\left\{j:d_{ij}=\min_{k\in\mathcal{C}_i}d_{ik}\right\}, $$ where $\mathcal{C}_i$ is the set of controls eligible for this match. For example, one possible distance is $d_{ij}=|e_i-e_j|$. The set $\mathcal{N}_i$ contains more than one control when the minimum distance is attained by more than one.

The distance must be specified before ties can be identified. It could be an absolute difference in propensity scores or another declared distance used by the matching procedure. The examples below use absolute score differences. They also hold fixed the treated-unit order, matching ratio, and replacement rule so that the effects shown are attributable to tie resolution.

Rules for Resolving a Tie

A deterministic rule always selects the same control when given the same data and matching setup. A common simple choice is to select the tied control with the smallest immutable identifier. Another is to compare a pre-specified secondary criterion, such as the absolute difference on a particular baseline covariate, and then use an identifier if that criterion is also tied. These rules can make the procedure reproducible, but they do not make the selected control uniquely correct: all tied controls are equally close under the primary distance.

A row-order rule, such as selecting whichever tied control appears first in the input file, is deterministic only if the data order is kept fixed. Reordering the file can change the match. A randomized choice among tied controls is an alternative, but it is not deterministic; documenting the random seed makes the selection reproducible for a given implementation, not identical across all possible seeds. A tie rule should not use observed outcomes, since choosing matches based on outcomes can bias the subsequent comparison.

For a fixed-ratio procedure, a tie can occur at the selection boundary. If a treated unit needs two controls and the closest control is unique but two other controls tie for the second position, selecting both tied controls would produce three matches rather than the required two. A tie-breaking rule must select the required number, unless the procedure explicitly changes its ratio or weighting rule.

Worked Example: A Tie in One-to-One Matching

A treated unit has estimated propensity score $0.50$. Three eligible controls have scores $0.47$, $0.53$, and $0.56$, with identifiers A, B, and C. Using absolute score difference, identify the tied nearest controls and compare two deterministic rules.

The distances are: $$ d_{iA}=|0.50-0.47|=0.03,\quad d_{iB}=|0.50-0.53|=0.03,\quad d_{iC}=|0.50-0.56|=0.06. $$

The minimum distance is $0.03$, so $\mathcal{N}_i=\{\mathrm{A},\mathrm{B}\}$. With one-to-one matching, only one control can be selected. If the rule is “smallest identifier,” A is matched. If a pre-specified secondary rule selects the control whose age is closest to the treated unit’s age, suppose the treated unit is 60 years old, A is 42, and B is 59. The age differences are $|60-42|=18$ years and $|60-59|=1$ year, so this rule selects B.

Both choices have the same primary propensity-score distance, $0.03$. The secondary rule changes the matched set, not the score distance of the selected match. It also does not prove that B is preferable overall: that conclusion would require justification of the secondary criterion and consideration of other measured covariates.

Fixed-Ratio Matching at a Tie Boundary

When the matching ratio requires several controls per treated unit, ties can arise at the last available position. A deterministic rule then chooses a subset of the tied controls. The same principle applies as in one-to-one matching: first identify the minimum distances and the tie group, then apply the declared rule only to the tied candidates.

Worked Example: Choosing the Second Control in a Two-Control Match

A treated unit with score $0.40$ is to be matched to two controls. The eligible control scores are: A, $0.39$; B, $0.37$; C, $0.43$; and D, $0.35$. Use absolute score difference and, among tied candidates at the boundary, choose the control with the smallest identifier.

Calculate every distance: $$ d_{iA}=|0.40-0.39|=0.01,\quad d_{iB}=|0.40-0.37|=0.03, $$ $$ d_{iC}=|0.40-0.43|=0.03,\quad d_{iD}=|0.40-0.35|=0.05. $$

A is the unique closest control, so it takes the first position. B and C tie at distance $0.03$ for the second position. The stated identifier rule selects B, giving the matched set $\{\mathrm{A},\mathrm{B}\}$. Its two distances are $0.01$ and $0.03$, with total distance $0.01+0.03=0.04$. Selecting C instead would give the same total distance, $0.01+0.03=0.04$, but a different matched set. Selecting both B and C would yield three controls and violate the specified two-control ratio.

The calculation shows what is and is not determined by the primary distance: the total distance for this set is fixed, but the identity of the second control is not. If B and C differ on baseline covariates, their selection can still affect the covariate composition of the matched sample.

Why a Tie Can Change Later Matches

With replacement, a selected control remains available for later treated units. A tie resolved for one treated unit therefore need not remove a tied candidate from later consideration. Without replacement, a selected control is no longer eligible for later matches. Consequently, two tie rules can produce different later candidate sets, even when they make equally close choices for the first treated unit.

The following calculation isolates that consequence. It holds the treated-unit order fixed and uses one-to-one greedy matching without replacement. The point is not to compare different treated-unit orders, but to show how resolving one tie can alter the remaining pool.

Worked Example: A Tie Changes the Next Match Without Replacement

There are two treated units, T1 and T2, processed in that order, and three controls, C1, C2, and C3. The distances are:

Distance	C1	C2	C3
T1	0.10	0.10	0.30
T2	0.11	0.25	0.30

For T1, C1 and C2 tie at the minimum distance $0.10$. If the identifier rule selects C1, C1 is removed from the pool. T2 then chooses the closer of the remaining controls: C2 at $0.25$ rather than C3 at $0.30$. The total distance is $0.10+0.25=0.35$.

If the tie rule instead selects C2 for T1, C2 is removed. T2 then chooses C1 at $0.11$ rather than C3 at $0.30$. The total distance is $0.10+0.11=0.21$. Thus, each rule makes an equally close match for T1, but the resulting two matched sets have different total distances and different control assignments. The tie decision is consequential because of the no-replacement restriction.

Justification, Diagnostics, and Limitations

The direct mathematical reason for documenting ties is that minimizing distance does not always identify a unique control. If two controls $j_1$ and $j_2$ both attain the minimum $m_i$, then $d_{ij_1}=m_i=d_{ij_2}$. Choosing either has the same distance for that treated unit. Under matching without replacement, however, the choice can change which controls remain eligible for another treated unit. The next match is computed from that changed eligible set, so equality at the first step does not imply equality of the complete matched sets.

This result concerns the matching algorithm, not causal identification. Matching uses estimated propensity scores, and a tie-breaking rule cannot establish that the score model is adequate, that treated and control units are comparable on relevant covariates, or that the assumptions required for a causal interpretation hold. The causal meaning of any estimate must be assessed separately.

Useful checks include recording the number of ties, the tied distances, the selected controls, the rule applied, and whether matching used replacement. After matching, inspect covariate balance and the number of times controls were selected. If many ties appear, check whether scores or distances were rounded before matching: distinct unrounded distances can appear equal after rounding. Ideally, identify ties using the precision of the values actually used by the algorithm and report any numerical tolerance used by the software. A tolerance can treat nearly equal values as tied, but it changes the operational definition of a tie and should be specified rather than chosen after seeing the results.

Deterministic rules also have limitations. An identifier-based choice is reproducible but has no substantive reason to improve covariate balance. A covariate-based secondary rule can target a stated feature, but it may worsen balance on other features and can introduce discretion if selected after examining the results. A randomized rule avoids privileging a particular identifier but can yield different matched sets across random draws. In all cases, assess the matched data rather than assuming that the rule created good balance.

The alternatives should not be conflated. Keeping all tied controls changes the number of controls or the weighting scheme, so it is not a tie-break under a fixed ratio. Choosing a secondary criterion still selects a nearest control only with respect to the primary distance; it does not optimize all covariates. A global optimization method is a different matching approach, not merely a local tie-breaking convention. For a transparent nearest-neighbor analysis, report the primary distance, matching ratio, replacement rule, treated-unit order when relevant, and tie-breaking rule.

Common Mistakes

Treating the first row as inherently closest. Row order is not evidence of smaller distance. Calculate the distances and identify the complete minimum set before applying a rule.
Breaking ties after looking at outcomes. Outcome-based selection can make the comparison outcome-dependent. Choose a rule using identifiers or pre-specified baseline information instead.
Assuming equal distance means equal matched sets. With no replacement, selecting one tied control changes the controls available to later treated units.
Keeping every tied control while claiming a fixed ratio. Including extra controls changes the matching design; select the required number or explicitly define a different procedure.
Ignoring rounding. A displayed tie may not be an exact tie in the values used for matching. State the precision or tolerance used to decide equality.

Key takeaway: A tie is a set of controls sharing the minimum distance, not an instruction about which one to select. Specify and report a reproducible, outcome-independent rule; then evaluate the resulting matched sets, especially when matching without replacement.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A treated score is $0.60$. Three control scores are $0.58$, $0.62$, and $0.64$. Using absolute score distance, identify the tied nearest controls and give their common distance.
A treated unit requires two controls. The sorted distances are $0.02$, $0.05$, $0.05$, and $0.08$. What is tied, and what must a fixed-ratio rule do?
Why can two tie rules that give the same distance for the first treated unit produce different later matches without replacement?
Why is selecting the tied control with the more favorable observed outcome an inappropriate tie-breaking rule?

Solutions. (1) The distances are $|0.60-0.58|=0.02$, $|0.60-0.62|=0.02$, and $|0.60-0.64|=0.04$. The first two controls tie as nearest, each at distance $0.02$. (2) The two controls at distance $0.05$ tie for the second position, after the unique control at $0.02$. A fixed-ratio rule must select one of those tied controls using its declared tie-breaker; selecting both would exceed the two-control ratio. (3) Without replacement, the selected control is removed from the eligible pool. Different tied choices leave different controls available, so later nearest distances and selected controls can differ. (4) Choosing by observed outcome makes match selection depend on the outcome being compared, potentially biasing the analysis. A tie rule should be specified independently of outcomes.