Matched-Pair Effects Versus Population Effects

Why a Pair Contrast Is Not Automatically a Population Effect

Matching creates close comparisons, but it does not by itself determine which population an effect describes. A difference between the treated and control outcomes in one pair is an observed pair contrast. A population-average treatment effect, by contrast, averages each target unit’s causal effect over a specified population. These quantities can be related, but they are not automatically the same.

The distinction matters even when every pair contains exactly one treated and one control unit. A pair contrast compares outcomes from two different people. A unit’s individual causal effect compares that same person’s two potential outcomes, only one of which is observed. The aggregation rule then determines how pair-level comparisons contribute to an overall estimate and which units the resulting summary represents.

A population-average effect requires specifying the target population and its averaging weights. For a finite target population \(Q\) of \(N_Q\) units, the equally weighted average treatment effect is

More generally, if target weights \(q_i\) are nonnegative and sum to one, the weighted target effect is \(\tau_Q=\sum_{i\in Q}q_i\tau_i\). The matched-pair average is another aggregation: with \(J\) pairs and equal pair weights, it is \(\bar D_{\text{pair}}=J^{-1}\sum_{j=1}^J D_j\). A causal interpretation of that average requires assumptions connecting the observed contrasts to the effects in the target being averaged.

When Pair Contrasts Have a Causal Interpretation

A useful way to see the distinction is to imagine a randomized design in which one member of each pair is assigned treatment and the other control. Suppose assignment is equally likely to select either member, and consider the average contrast over these two possible assignments. The expected contrast is the pair’s average of its members’ individual treatment effects.

To verify this, let the two members be \(a\) and \(b\). If \(a\) is treated, the contrast is \(Y_a(1)-Y_b(0)\). If \(b\) is treated, it is \(Y_b(1)-Y_a(0)\). Averaging these possibilities gives:

Thus, under this particular within-pair randomization, the expected pair contrast equals the average causal effect for the two members. This is an expectation over assignment, not a claim that the realized contrast equals either member’s effect. Observational matching does not randomly assign treatment within a pair, so this argument alone does not justify a causal interpretation there.

For observational matching, the required justification instead comes from assumptions about treatment assignment and the matched comparison. In a common treated-unit matching design, the intended target is the average treatment effect among the matched treated units. A control matched to treated unit \(t_j\) is being used to represent that treated unit’s unobserved untreated outcome. With exact covariate matching and conditional exchangeability, their untreated outcomes have the same conditional mean. The expected pair contrast can then represent the treated unit’s treatment effect, on average over the matched treated units.

If matches are approximate rather than exact, the untreated outcome means for treated and control members may differ even after conditioning on the matching variables. The pair contrast can then combine a treatment effect with residual baseline differences. In addition, a causal interpretation applies to the units represented by the matched design; it does not automatically extend to units that were excluded or to a broader population.

Aggregation Determines Which Effect Is Summarized

After calculating one contrast per pair, the analysis still needs an aggregation rule. With equally weighted pairs, the summary is the arithmetic mean of the \(D_j\). That summary gives each pair the same contribution. It describes an average over the pairs included in the analysis, provided the assumptions needed for a causal interpretation hold. It need not describe the full eligible population.

More generally, write an aggregated contrast as \(\hat\tau_w=\sum_{j=1}^J w_jD_j\), where \(w_j\geq0\) and \(\sum_j w_j=1\). The weights say how much each pair contributes. If the \(D_j\) estimate causal effects for the units represented by their pairs, the weights determine how those effects are averaged. A change in weights can therefore change the estimand, even if the underlying pair contrasts do not change.

In the special case of one treated and one control per pair, equal weights across \(J\) pairs give each pair a contribution of \(1/J\). If the target is defined as the equal-weight average effect among the \(J\) matched treated units, and each pair contrast validly represents its treated member’s effect, the equal-pair average is aligned with that target. This is not automatically the ATE for all eligible units: unmatched units are not represented by the same average. Nor does equal weighting make each pair’s observed contrast a literal individual effect.

Diagnostics, Failure Modes, and Alternatives

Several checks help determine whether an aggregated pair contrast is credible for its stated target. First, inspect covariate balance within pairs and across the matched sample. Large within-pair differences in important pretreatment covariates suggest that controls may not represent treated units’ untreated outcomes well. Second, report which units were retained and discarded; this reveals whether the matched target is narrower than the population named in the research question. Third, examine whether any pairs dominate the aggregate because of the chosen weights. Each check addresses a different issue: balance concerns comparability, retention concerns the target population, and weighting concerns how represented effects are combined.

A common mistake is to say, “Each pair has one treated and one control unit, so the average pair difference is the population ATE.” The reasoning confuses the structure of the data with the target of the average. Having one unit per treatment group within each pair does not ensure that all target-population units are represented, that treatment is exchangeable within pairs, or that the chosen pair weights match the desired population weights. The correction is to name the units represented by the pairs, state the assumptions that justify the contrasts, and specify how the contrasts are aggregated.

Matched-pair aggregation is not the only way to summarize effects. A weighted estimator or a standardized outcome model can target a different population or use a different weighting rule; a regression coefficient may also summarize effects differently when effects vary across units. These approaches are comparable only after their target populations and effect scales are made explicit. Matching can make comparisons more transparent, but it cannot remove unmeasured confounding or supply information about units with no suitable matches. A precise interpretation therefore states both the causal assumptions and the population represented.