Pairwise Matching Design in Experiments

• Pairwise matching design is an experimental strategy that partitions units into disjoint pairs based on pre-treatment characteristics to reduce confounding and variance.
• In one-dimensional settings, adjacent pairing of sorted units minimizes within-pair differences exactly for both absolute and squared loss objectives with O(N log N) complexity.
• Extensions of pairwise matching integrate randomized block design and hybrid imbalance optimization to balance observed and unobserved heterogeneity, achieving robust asymptotic optimality.

Searching arXiv for the core matched-pair experimental-design papers and the one-dimensional optimal matching result. Pairwise matching design is an experimental-design strategy that partitions units into disjoint pairs so that members of each pair are as similar as possible on pre-treatment information, and then uses those pairs as the basis for treatment assignment, comparison, or inference. In its most classical form, the design is motivated by the desire to reduce confounding and variance by comparing like with like, whether the relevant pre-treatment information is a scalar such as age or prognostic score, a multivariate covariate vector, or an unknown prognostic quantity approximated from observed covariates. Across the literature, pairwise matching appears both as a combinatorial optimization problem on observed units and as a blocking design for randomized experiments. In one dimension, the optimization problem collapses to sorting and pairing adjacent values, giving exact global optimality under absolute- or squared-difference objectives (Bukac, 2018). In randomized two-arm experiments, prior pair matching emerges as the finest balanced block design and, under a Bernoulli incidence model with risk-difference estimation, is optimal among block designs and minimax over all balanced designs (Kapelner et al., 2022). Under a broader simultaneous-tail criterion that accounts jointly for assignment randomness and unobserved subject-level noise, pairwise matching is asymptotically optimal among block designs for continuous, incidence, proportion, count, and uncensored survival outcomes (Azriel et al., 2024). A further development hybridizes optimal nonbipartite matching with imbalance optimization, preserving the robustness advantages of matched pairs while improving global covariate balance (Krieger et al., 2020).

1. One-dimensional pair formation and exact optimality

Bukac’s “Matching on a line” formalizes the simplest pairwise matching design problem: given an even number of real-valued observations,

$x_1, x_2, \dots, x_{2n},$

partition them into $n$ disjoint pairs so that within-pair dissimilarity is minimized (Bukac, 2018). The design interpretation is direct. If subjects can be represented by a single scalar covariate or score—such as age, a prognostic score, or a propensity-score-like scalar—then matched-pair design becomes a one-dimensional partitioning problem.

For pairwise matching, the paper studies two within-pair criteria. Under absolute differences,

$A(x_i,x_j)=|x_j-x_i|,$

the objective is

$\min \sum_{r=1}^n |x_{i_r}-x_{j_r}|.$

Under squared differences,

$S(x_i,x_j)=(x_j-x_i)^2,$

the objective is

$\min \sum_{r=1}^n (x_{i_r}-x_{j_r})^2.$

The core structural result is that on the real line, crossing pairs are never better than adjacent pairs. For a sorted 4-tuple $x_1\le x_2\le x_3\le x_4$, the absolute-difference uncrossing inequalities are

$$A(x_1,x_3)+A(x_2,x_4)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0,$$

$$A(x_1,x_4)+A(x_2,x_3)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0.$$

The squared-difference analogue is

$$S(x_1,x_3)+S(x_2,x_4)-\bigl(S(x_1,x_2)+S(x_3,x_4)\bigr) =2(x_4-x_1)(x_3-x_2)\ge 0,$$

$n$0

These inequalities imply that any deviation from adjacent pairing can be locally improved by uncrossing (Bukac, 2018).

The global theorem then follows by induction: for sorted points

$n$1

an optimal pair matching is

$n$2

and this holds for both

$n$3

(Bukac, 2018). The resulting algorithm is immediate: sort the $n$4 scalars and pair adjacent values. Since sorting dominates, the complexity is

$n$5

This exact result is specific to one dimension and to the stated loss functions. The paper explicitly contrasts it with general nonbipartite matching, for which weighted matching methods such as Edmonds’ algorithm are needed and are cited with complexity $n$6 (Bukac, 2018). This suggests a central organizing principle for pairwise matching design: when all matching information can be reduced to a single ordered scalar and the objective is based on absolute or squared within-pair differences, the combinatorial search over pairings becomes unnecessary.

2. Pairwise matching as a randomized block design

In randomized experiments, pairwise matching is not only a partitioning rule but also a randomization design. The paper “The Role of Pairwise Matching in Experimental Design for an Incidence Outcome” studies a two-arm randomized experiment with $n$7 fixed subjects, balanced assignment, and binary outcomes under a nonparametric general response model (Kapelner et al., 2022). Each subject has covariates $n$8, assignment $n$9, and Bernoulli outcome with arm-specific probability

$A(x_i,x_j)=|x_j-x_i|,$0

The estimand is the sample average treatment effect,

$A(x_i,x_j)=|x_j-x_i|,$1

where

$A(x_i,x_j)=|x_j-x_i|,$2

The estimator is the difference in means,

$A(x_i,x_j)=|x_j-x_i|,$3

Under balanced assignment and equal treatment probability, the estimator is unbiased, and the mean squared error satisfies

$A(x_i,x_j)=|x_j-x_i|,$4

where

$A(x_i,x_j)=|x_j-x_i|,$5

Because the second term does not depend on the design, comparing designs reduces exactly to comparing

$A(x_i,x_j)=|x_j-x_i|,$6

(Kapelner et al., 2022).

Within the class of balanced block designs, pair matching is the finest possible blocking scheme. If there are $A(x_i,x_j)=|x_j-x_i|,$7 blocks of equal size $A(x_i,x_j)=|x_j-x_i|,$8, then within each block exactly half the subjects are assigned treatment and half control. Two cases are central:

Design	Structure	Within-block covariance
Balanced complete randomization design (BCRD)	One block of size $A(x_i,x_j)=|x_j-x_i|,$9	Off-diagonals $\min \sum_{r=1}^n |x_{i_r}-x_{j_r}|.$0
Pair matching (PM)	$\min \sum_{r=1}^n |x_{i_r}-x_{j_r}|.$1 blocks of size $\min \sum_{r=1}^n |x_{i_r}-x_{j_r}|.$2	$\min \sum_{r=1}^n |x_{i_r}-x_{j_r}|.$3

The theoretical optimum is defined by sorting subjects by

$\min \sum_{r=1}^n |x_{i_r}-x_{j_r}|.$4

and pairing adjacent units. Each pair is then randomized uniformly between $\min \sum_{r=1}^n |x_{i_r}-x_{j_r}|.$5 and $\min \sum_{r=1}^n |x_{i_r}-x_{j_r}|.$6 (Kapelner et al., 2022). In the idealized one-covariate monotone case, sorting by the covariate suffices; in higher dimensions, the paper recommends approximate matching using covariate similarity and uses optimal nonbipartite matching with Mahalanobis distance in simulations (Kapelner et al., 2022).

The first main theorem states that pair matching is optimal among block designs with even block sizes: $\min \sum_{r=1}^n |x_{i_r}-x_{j_r}|.$7 for any $\min \sum_{r=1}^n |x_{i_r}-x_{j_r}|.$8 and any $\min \sum_{r=1}^n |x_{i_r}-x_{j_r}|.$9 (Kapelner et al., 2022). For pair matching,

$S(x_i,x_j)=(x_j-x_i)^2,$0

whereas for larger blocks the corresponding quantity averages over all within-block pairwise squared differences. Since sorted adjacent differences are the smallest available, finer blocking monotonically improves the design criterion, and pair matching is best.

The second main result is minimax optimality over all balanced designs: $S(x_i,x_j)=(x_j-x_i)^2,$1 where

$S(x_i,x_j)=(x_j-x_i)^2,$2

and $S(x_i,x_j)=(x_j-x_i)^2,$3 is the class of designs satisfying balanced support and marginal treatment probability $S(x_i,x_j)=(x_j-x_i)^2,$4 (Kapelner et al., 2022). The proof reduces the worst-case problem to extreme-point step vectors and shows that pair matching attains the universal lower bound of $S(x_i,x_j)=(x_j-x_i)^2,$5 on the design-dependent quadratic form.

These results place pairwise matching at the center of randomized design for incidence outcomes: it is both the best balanced block design under the stated criterion and the safest one in a worst-case sense when the prognostic quantity $S(x_i,x_j)=(x_j-x_i)^2,$6 is unknown but ordered.

3. Robustness, randomness, and tail-optimality

A later paper extends the case for pairwise matching beyond mean MSE under a binary outcome model. “The Pairwise Matching Design is Optimal under Extreme Noise and Assignments” studies equally allocated two-arm experiments with continuous, incidence, proportion, count, and uncensored survival outcomes, modeling both assignment randomness and unobserved subject-specific noise (Azriel et al., 2024). Potential outcomes are decomposed as

$S(x_i,x_j)=(x_j-x_i)^2,$7

with zero-mean independent noise components and bounded fourth moments.

The estimator remains

$S(x_i,x_j)=(x_j-x_i)^2,$8

and the design-dependent mean MSE after averaging over noise is

$S(x_i,x_j)=(x_j-x_i)^2,$9

where

$\min \sum_{r=1}^n (x_{i_r}-x_{j_r})^2.$0

and $\min \sum_{r=1}^n (x_{i_r}-x_{j_r})^2.$1 does not depend on the design (Azriel et al., 2024). However, the paper argues that averaging over assignment first and then over noise is unrealistic, because a real experiment realizes only one assignment. It therefore proposes a simultaneous tail criterion based on the upper quantile of the squared estimation error: $\min \sum_{r=1}^n (x_{i_r}-x_{j_r})^2.$2 approximated by

$\min \sum_{r=1}^n (x_{i_r}-x_{j_r})^2.$3

Within homogeneous block designs, the paper proves a universal asymptotic lower bound: $\min \sum_{r=1}^n (x_{i_r}-x_{j_r})^2.$4 Pairwise matching attains this lower bound under the additional matching condition

$\min \sum_{r=1}^n (x_{i_r}-x_{j_r})^2.$5

Under this assumption,

$\min \sum_{r=1}^n (x_{i_r}-x_{j_r})^2.$6

and

$\min \sum_{r=1}^n (x_{i_r}-x_{j_r})^2.$7

so pairwise matching asymptotically minimizes both the mean term and the variance term of the simultaneous-tail criterion among block designs (Azriel et al., 2024).

A second conceptual result compares pairwise matching to deterministic perfect-balance allocation. Perfect balance minimizes observed imbalance almost completely, but the paper shows that its asymptotic variance of squared error is

$\min \sum_{r=1}^n (x_{i_r}-x_{j_r})^2.$8

which is four times larger than pair matching’s

$\min \sum_{r=1}^n (x_{i_r}-x_{j_r})^2.$9

(Azriel et al., 2024). This yields a clear randomness–imbalance tradeoff: complete randomization is too random, deterministic optimized allocation is too nonrandom, and pairwise matching occupies the intermediate regime that best manages both observed and unobserved heterogeneity under the joint-tail criterion.

This suggests that pairwise matching design should not be understood merely as variance reduction through blocking. In this asymptotic framework it is also a controlled-randomization device that preserves enough randomization to protect against unlucky realizations of unobserved factors while still exploiting observed covariate structure.

4. Hybridization with imbalance optimization

The paper “Better Experimental Design by Hybridizing Binary Matching with Imbalance Optimization” proposes a different extension: preserve matched pairs, but optimize treatment assignment across pairs using within-pair covariate differences (Krieger et al., 2020). The setting is a balanced two-arm experiment with $x_1\le x_2\le x_3\le x_4$0 units, covariates $x_1\le x_2\le x_3\le x_4$1, and assignment

$x_1\le x_2\le x_3\le x_4$2

The estimator is the difference in means,

$x_1\le x_2\le x_3\le x_4$3

The design begins with optimal nonbipartite matching. For a distance

$x_1\le x_2\le x_3\le x_4$4

the matched-pair structure is

$x_1\le x_2\le x_3\le x_4$5

The main implementation uses Mahalanobis distance, and the ordinary matched-pair assignment space is

$x_1\le x_2\le x_3\le x_4$6

The innovation is to retain the matched pairs and define within-pair difference vectors

$x_1\le x_2\le x_3\le x_4$7

The orientation of each pair then determines whether $x_1\le x_2\le x_3\le x_4$8 or $x_1\le x_2\le x_3\le x_4$9 contributes to treatment-minus-control imbalance. Since the global Mahalanobis imbalance is

$$A(x_1,x_3)+A(x_2,x_4)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0,$$0

one can improve overall balance by choosing pair orientations so that the signed sum of the $$A(x_1,x_3)+A(x_2,x_4)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0,$$1 vectors is small (Krieger et al., 2020).

The paper studies two such second-stage procedures:

Hybrid design	Second stage	Constraint
MR	Rerandomization within $$A(x_1,x_3)+A(x_2,x_4)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0,$$2	Keep assignments with $$A(x_1,x_3)+A(x_2,x_4)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0,$$3
MG	Greedy pair-of-pairs switching on $$A(x_1,x_3)+A(x_2,x_4)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0,$$4	Preserve matched-pair structure

For MR,

$$A(x_1,x_3)+A(x_2,x_4)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0,$$5

For MG, the greedy switching heuristic is applied not to raw units but to the pair-difference vectors $$A(x_1,x_3)+A(x_2,x_4)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0,$$6, using pair-of-pairs switches that preserve the within-pair treatment–control constraint (Krieger et al., 2020).

The asymptotic result emphasized in the paper is for one uniformly distributed covariate. Matching alone shrinks within-pair gaps at rate

$$A(x_1,x_3)+A(x_2,x_4)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0,$$7

while greedy switching alone achieves

$$A(x_1,x_3)+A(x_2,x_4)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0,$$8

imbalance reduction. Combined, MG attains

$$A(x_1,x_3)+A(x_2,x_4)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0,$$9

imbalance (Krieger et al., 2020). For MR, if $$A(x_1,x_4)+A(x_2,x_3)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0.$$0 candidate assignments are searched and the best retained, the rate becomes

$$A(x_1,x_4)+A(x_2,x_3)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0.$$1

Simulation results support the conceptual tradeoff. Under pure linear outcome models, pure imbalance-optimizing designs are strongest; under pure nonlinear models, matching-based designs are strongest; under mixed linear/nonlinear models, the hybrids MR and MG are best (Krieger et al., 2020). This suggests that matched-pair design and covariate-imbalance optimization address distinct aspects of the response surface. Matching protects against nonlinear structure by aligning nearby units, while orientation optimization reduces residual global imbalance relevant under linear components.

A common misconception is that matching and imbalance optimization are substitutes. The hybrid construction indicates the opposite: they can be compounded, with matching shrinking the problem and optimization solving the shrunk problem better.

5. Algorithmic and inferential variants

Pairwise matching design appears in several additional forms that broaden its methodological scope. One line concerns active enrollment in matched-pair experiments. “Matched-Pair Experimental Design with Active Learning” treats treatment-effect heterogeneity as a classification problem and uses thresholded within-pair treatment–control differences to label covariate points as belonging to a high-effect region (Li et al., 12 Sep 2025). The target region is

$$A(x_1,x_4)+A(x_2,x_3)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0.$$2

and under balanced matching the Bayes classifier

$$A(x_1,x_4)+A(x_2,x_3)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0.$$3

satisfies

$$A(x_1,x_4)+A(x_2,x_3)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0.$$4

The active enrollment region is

$$A(x_1,x_4)+A(x_2,x_3)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0.$$5

and the paper proves, under its assumptions, that

$$A(x_1,x_4)+A(x_2,x_3)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0.$$6

with high probability (Li et al., 12 Sep 2025). This transforms matched-pair design from a passive blocking device into an adaptive enrichment mechanism.

Another line concerns pairwise comparison data rather than paired assignment. “Accelerated Experimental Design for Pairwise Comparisons” studies which $$A(x_1,x_4)+A(x_2,x_3)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0.$$7 pairwise comparisons should be elicited from $$A(x_1,x_4)+A(x_2,x_3)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0.$$8 possibilities, using a D-optimal criterion

$$A(x_1,x_4)+A(x_2,x_3)-\bigl(A(x_1,x_2)+A(x_3,x_4)\bigr)=2(x_3-x_2)\ge 0.$$9

with

$$S(x_1,x_3)+S(x_2,x_4)-\bigl(S(x_1,x_2)+S(x_3,x_4)\bigr) =2(x_4-x_1)(x_3-x_2)\ge 0,$$0

Although the objective concerns comparison elicitation rather than experimental blocking, the paper exploits the pairwise-difference structure to accelerate greedy design from

$$S(x_1,x_3)+S(x_2,x_4)-\bigl(S(x_1,x_2)+S(x_3,x_4)\bigr) =2(x_4-x_1)(x_3-x_2)\ge 0,$$1

to

$$S(x_1,x_3)+S(x_2,x_4)-\bigl(S(x_1,x_2)+S(x_3,x_4)\bigr) =2(x_4-x_1)(x_3-x_2)\ge 0,$$2

(Guo et al., 2019). This suggests a broader algorithmic theme: pairwise design problems often admit substantial simplification when the paired object can be expressed as a structured difference.

The pairwise-constraint literature offers a combinatorial analogue. In 3D registration, tentative correspondences are vertices and pairwise geometric consistency defines edges: $$S(x_1,x_3)+S(x_2,x_4)-\bigl(S(x_1,x_2)+S(x_3,x_4)\bigr) =2(x_4-x_1)(x_3-x_2)\ge 0,$$3 Seeking the largest pairwise-consistent subset is equivalent to maximum clique in the consistency graph (Parra et al., 2019). This is not a matched-pair randomized design, but it reflects the same structural principle: pairwise feasibility can replace explicit optimization over a latent transformation.

These variants indicate that pairwise matching design spans at least three methodological regimes: blocking and randomization in experiments, adaptive enrichment based on matched-pair outcomes, and combinatorial or information-theoretic selection over pairwise objects.

6. Scope, assumptions, and limitations

The strongest results in the literature depend on restrictive structural assumptions. In one dimension, exact optimality by adjacent pairing requires that all matching information be represented by a single scalar and that the objective be

$$S(x_1,x_3)+S(x_2,x_4)-\bigl(S(x_1,x_2)+S(x_3,x_4)\bigr) =2(x_4-x_1)(x_3-x_2)\ge 0,$$4

(Bukac, 2018). The argument does not extend automatically to multivariate geometry, where nearest-neighbor structure and global optimum can diverge.

In randomized-experiment theory for incidence outcomes, the optimality criterion depends on the unknown prognostic quantity

$$S(x_1,x_3)+S(x_2,x_4)-\bigl(S(x_1,x_2)+S(x_3,x_4)\bigr) =2(x_4-x_1)(x_3-x_2)\ge 0,$$5

Theorems for pair matching assume exact ordering by this $$S(x_1,x_3)+S(x_2,x_4)-\bigl(S(x_1,x_2)+S(x_3,x_4)\bigr) =2(x_4-x_1)(x_3-x_2)\ge 0,$$6, whereas practice typically matches on covariates or proxies such as Mahalanobis distance (Kapelner et al., 2022). The same practical gap appears in the simultaneous-tail analysis, where asymptotic optimality requires

$$S(x_1,x_3)+S(x_2,x_4)-\bigl(S(x_1,x_2)+S(x_3,x_4)\bigr) =2(x_4-x_1)(x_3-x_2)\ge 0,$$7

under the chosen pairing (Azriel et al., 2024).

Hybrid designs inherit the quality of the initial match. If the match is poor, then balancing pair differences may optimize the wrong representation of heterogeneity. Likewise, active matched-pair enrichment presumes that matched-pair threshold labels meaningfully track the high-effect region and that matching is sufficiently balanced to justify the classification reduction (Li et al., 12 Sep 2025).

A further practical limitation is dimensionality. Several papers suggest that approximate matching in high-dimensional covariate spaces can degrade when many included covariates are weakly prognostic (Kapelner et al., 2022). This suggests that pairwise matching is not automatically improved by increasing the covariate set. A plausible implication is that matching quality may depend as much on prognostic variable selection as on the downstream matching algorithm.

Finally, pairwise matching design does not eliminate the need for randomness. The simultaneous-tail analysis shows that over-optimized deterministic allocation can worsen tail risk from unobserved heterogeneity, even when observed covariate balance is nearly perfect (Azriel et al., 2024). The literature therefore does not support the simple view that “more balance is always better.” Rather, the balance–randomness tradeoff is central.

7. Conceptual position within experimental design

Across these papers, pairwise matching design occupies a distinctive position between unrestricted randomization and highly deterministic optimization. Its unifying logic is local comparability: rather than equalizing only arm-level means, it creates or exploits local neighborhoods of similar units and then bases treatment comparison, assignment, or learning on those local structures.

In its simplest exact form, pairwise matching on a line is a deterministic partitioning rule with globally optimal within-pair loss (Bukac, 2018). In randomized experiments, pair matching is the finest balanced block design and can be characterized as both optimal among block designs and minimax over balanced designs for incidence outcomes (Kapelner et al., 2022). Under a joint tail-risk criterion, it also emerges as the asymptotically optimal compromise between full randomization and deterministic perfect balance (Azriel et al., 2024). In hybrid designs, matched pairs become a scaffold on which further imbalance optimization is performed (Krieger et al., 2020). In adaptive designs, matched-pair outcomes generate labels for sequential region discovery (Li et al., 12 Sep 2025).

Taken together, these results suggest a general interpretation. Pairwise matching design is not merely a preprocessing step or a variance-reduction heuristic. It is a structural design paradigm in which pair formation, pairwise assignment, and pairwise comparison encode assumptions about smoothness, prognostic similarity, and the desirable scale at which treatment contrasts should be controlled. When those assumptions are appropriate, pairwise matching can transform difficult global design problems into tractable local ones, sometimes exactly and sometimes asymptotically. When they are not, its apparent simplicity can conceal strong model dependence.