Variable-Ratio Matching and Matched-Set Targets

Why the Number of Controls Can Vary

A matching design need not assign the same number of controls to every treated unit. Some treated units may have many suitable controls nearby in covariate space, while others have only one or a few. A variable-ratio design allows the number of controls to differ across matched sets. The resulting estimate depends not only on which units are matched, but also on how outcomes are averaged within sets and how set contrasts are combined.

The central question is therefore both computational and causal: what population does the matched-set estimate describe? In the common design with one treated unit in each set, averaging the set contrasts equally gives each retained treated unit equal weight. The target is then the average treatment effect for those matched treated units, provided the matched controls validly represent their untreated outcomes.

Definition: A variable-ratio matched set $j$ contains one treated unit, indexed $i_j$, and $m_j$ control units, where $m_j$ can differ across sets. Write $C_j$ for the set of controls. The within-set contrast is the treated unit’s observed outcome minus the mean observed outcome of its matched controls: $$ D_j = Y_{i_j} - \frac{1}{m_j}\sum_{k\in C_j}Y_k. $$ With $J$ matched treated units, the treated-unit-weighted matched-set estimate is $$ \widehat{\tau}_{M,ATT} = \frac{1}{J}\sum_{j=1}^{J}D_j. $$

Here $A_i=1$ denotes treatment and $A_i=0$ denotes control. Each unit has potential outcomes $Y_i(1)$ and $Y_i(0)$, and the observed outcome is $Y_i=A_iY_i(1)+(1-A_i)Y_i(0)$. The finite-sample target for this estimate is the average treatment effect among the matched treated units:

\tau_{M,ATT} = \frac{1}{J}\sum_{j=1}^{J}\left\{Y_{i_j}(1)-Y_{i_j}(0)\right\}.

This target is specific to the treated units retained in the matched sample. It is not automatically the ATT for every treated unit in the original data, nor the ATE for the full study population. If some treated units cannot be matched, removing them changes the population represented by the estimate.

Assumptions Behind the Matched-Set Contrast

Matching is a design step; it does not itself guarantee causal identification. For a causal interpretation, the usual potential-outcome conditions must be plausible in the matched population. Consistency requires that each person’s observed outcome equal the potential outcome under the treatment actually received. No interference requires that one unit’s treatment not change another unit’s outcome. Conditional exchangeability requires that, given the covariates used for matching, treatment assignment be independent of the potential outcomes. For an ATT target, a practical overlap requirement is that each retained treated unit have eligible controls with comparable covariates.

Under exact covariate matching, the identification argument is direct: conditional on the matched covariates, exchangeability implies that the mean outcome among controls identifies the treated unit’s mean untreated potential outcome. In practice, matching is often approximate. Then the comparison is credible only to the extent that the matched controls are sufficiently similar on confounders, and any remaining covariate differences are small enough not to create consequential bias. A caliper or other eligibility rule can improve match quality, but may also exclude treated units and thereby change the target.

How the Estimate Represents the Matched Treated Units

Within each set, the controls are averaged before the set contrasts are combined. If every set contains one treated unit and the contrasts are averaged equally, each treated unit contributes $1/J$ to the final estimate, regardless of whether its set contains one control or several. The controls serve as comparisons for those treated units; they do not become additional treated units in the target population.

This order of operations matters. A control in a set of four receives one-quarter of that set’s control contribution, while the single control in a set receives the full control contribution for its set. If a control is reused across sets, it can contribute to more than one set contrast. Reuse does not change which treated units define the target, although it can induce dependence among the contrasts and should be reflected in uncertainty calculations.

Worked Example: Calculating an Estimate With Different Set Sizes

Suppose a hypothetical study forms three matched sets. Each set has one treated unit, but the numbers of controls are one, two, and three. The observed outcomes are:

Set	Treated outcome	Control outcomes	Control mean
1	7	5	5
2	12	7, 9	8
3	14	3, 4, 5	4

The within-set contrasts are:

D_1=7-5=2,\qquad D_2=12-\frac{7+9}{2}=4,\qquad D_3=14-\frac{3+4+5}{3}=10.

There are $J=3$ matched treated units, so the treated-unit-weighted estimate is

\widehat{\tau}_{M,ATT}=\frac{2+4+10}{3} =\frac{16}{3}=5.333.

In the hypothetical outcome units, the estimate is 5.333 units per treated unit. Its target is the average effect among these three matched treated units, if the controls adequately identify their untreated outcomes. The six controls improve the comparisons within the sets; they do not make the target an effect for six or nine people.

Why Set Construction Can Shift the Target

The number of controls in a set is a design feature, but the rules that determine whether a treated unit receives a set are also consequential. For example, a strict caliper may leave some treated units without an eligible control. The estimate then concerns only the matched treated units. Even if matching retains all treated units, the covariate distribution of those units may differ from that of all treated people in the source population. Under treatment-effect heterogeneity, these changes can alter the average effect being represented.

A useful justification for the treated-unit-weighted formula can be seen by examining one set. If its controls have the same covariates as the treated unit, conditional exchangeability gives

E\left[\frac{1}{m_j}\sum_{k\in C_j}Y_k\;\middle|\;X_{i_j}\right] =E\left[Y_{i_j}(0)\mid X_{i_j}\right].

The treated unit’s observed outcome is $Y_{i_j}(1)$, by consistency. Thus, conditional on its covariates, the expected set contrast is

E[D_j\mid X_{i_j}] =E\left[Y_{i_j}(1)-Y_{i_j}(0)\mid X_{i_j}\right].

Averaging these conditional contrasts equally across matched treated units corresponds to averaging their conditional treatment effects equally. Approximate rather than exact matching adds a potential discrepancy: the control mean may reflect covariates that differ from those of the treated unit. The derivation therefore explains the target under valid comparisons; it does not make covariate imbalance harmless.

Worked Example: Excluding Treated Units Changes the Target

Consider a hypothetical population of four treated units: two in a mild-disease subgroup and two in a severe-disease subgroup. Suppose their treatment effects are 2 outcome units for each mild-disease unit and 8 outcome units for each severe-disease unit. If all four treated units are represented, their average effect is

\frac{2+2+8+8}{4}=\frac{20}{4}=5 \text{ outcome units}.

Now suppose an eligibility rule leaves one severe-disease treated unit without a suitable control. The three retained treated units have effects 2, 2, and 8, so their target average is

\frac{2+2+8}{3}=\frac{12}{3}=4 \text{ outcome units}.

This hypothetical calculation assumes the stated subgroup effects are correctly identified by the matches. It shows that the change from 5 to 4 outcome units need not indicate a computational error: the matched estimate targets a different treated population. The exclusion has changed the composition of the target, and the subgroups have different effects.

Aggregation Choices and Alternatives

The default estimate above gives each matched treated unit equal weight. A different calculation can change the represented target. For instance, weighting each set contrast by its number of controls gives more influence to treated units with larger sets. Unless treatment effects are constant across treated units, that calculation does not generally estimate the same matched-treated ATT. A raw difference between the pooled treated and pooled control means is also not the same operation as averaging within-set contrasts: it ignores which controls were assigned to which treated units and can give a different result.

Fixed-ratio matching is a simpler alternative when each treated unit can be assigned the same number of suitable controls. One-to-one matching may be easier to describe, but it can leave useful controls unused. Variable-ratio matching can use more eligible controls for some treated units while keeping the treated-unit target explicit through equal weighting of set contrasts. Its advantage is not automatic: a larger set is helpful only if its added controls are credible comparisons. Regression adjustment or weighting are other possible strategies, but they use different modeling or weighting choices and do not remove the need to specify the target population.

Worked Example: A Different Aggregation Gives a Different Summary

Suppose three hypothetical matched sets have control counts $m_1=1$, $m_2=3$, and $m_3=2$, and set contrasts $D_1=1$, $D_2=5$, and $D_3=9$ outcome units. Equal averaging across the three treated units gives

\frac{1+5+9}{3}=5 \text{ outcome units}.

If instead the contrasts are weighted by their numbers of controls, the result is

\frac{1(1)+3(5)+2(9)}{1+3+2} =\frac{34}{6}=5.667 \text{ outcome units}.

The two values differ because the second calculation assigns more weight to the treated units in larger sets. The control-count-weighted result is not the equal-treated-unit average, and it would require a separately justified target interpretation. A larger $m_j$ alone is not a reason to give the corresponding treated unit more weight.

Diagnostics and Common Pitfalls

Before interpreting a variable-ratio estimate, inspect both the matching structure and the covariate comparisons it creates. Check how many treated units were retained, the distribution of $m_j$, whether any controls were reused, and whether sets contain controls that are close to their treated units on important confounders. Summarize covariate balance for the matched treated and control units, and examine whether poor matches cluster in particular covariate regions or subgroups. A large number of controls cannot compensate for a lack of overlap or systematically poor comparability.

Mistake: “There are more controls, so the target includes more people.” The number of controls affects the comparison for each treated unit; with equal averaging of one-treated-unit sets, the target remains the matched treated units.
Mistake: “Every control outcome can be pooled because all controls were matched.” Pooling discards set membership and may alter the contrast and its weighting. Calculate each within-set control mean first.
Mistake: “A tight match guarantees no confounding.” Covariate similarity on observed variables does not rule out unmeasured confounding. The causal interpretation still depends on exchangeability given the measured covariates.
Mistake: “Variable set sizes are only a precision choice.” The number of controls affects each set’s comparison, while match eligibility can determine which treated units remain. Both the comparison quality and the target require attention.

Variable-ratio matching is most useful when the supply of credible controls differs across treated units. Report the matching eligibility rules, the number of treated units retained, the set-size pattern, control reuse, the within-set contrast, and the aggregation rule. These details make the represented target visible and help readers distinguish a matched-sample effect from a broader population effect.

Check Your Understanding

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

A set contains one treated unit with outcome 11 and two controls with outcomes 6 and 8. What is the set contrast?
Two matched treated units have set contrasts 3 and 7. What is their equal-treated-unit estimate, and what population does it target?
Why does a set with four controls not automatically receive four times the weight of a set with one control in the treated-unit-weighted estimate?
A caliper excludes treated units with covariate values outside the region where eligible controls are available. What target change may result?
State one diagnostic that can reveal a problem with variable-ratio matching and explain what it indicates.

Solutions. (1) The control mean is $(6+8)/2=7$, so the contrast is $11-7=4$ outcome units. (2) The estimate is $(3+7)/2=5$ outcome units; it targets the average treatment effect among those two matched treated units, if their controls identify their untreated outcomes. (3) The estimate averages one contrast per treated unit, so each treated unit receives equal weight; the four controls are averaged within their set. (4) The retained treated units may have a different covariate and treatment-effect distribution from all treated units, so the estimate may represent a restricted matched-treated ATT. (5) For example, poor covariate balance after matching signals that treated units and their controls remain different on measured characteristics; this weakens the comparison and may leave confounding bias.