What a Matched Set Represents
Matching organizes treated and control units into groups intended to make comparisons more relevant. The structure of those groups matters: a group containing one treated unit and one control unit is not the same design as a group containing one treated unit and several controls, or a group whose treated-to-control ratio varies from set to set. These structures determine which units are grouped together and how many units of each treatment status are retained. They do not, by themselves, establish that the groups are comparable or specify a unique treatment-effect estimator.
Let \(A_i=1\) indicate that unit \(i\) received the treatment and \(A_i=0\) indicate that it did not. A matched set \(S_k\) is a group containing both treated and control units. Write \(n_{1k}\) for the number of treated units and \(n_{0k}\) for the number of control units in set \(k\). A structurally valid set has \(n_{1k}\geq 1\) and \(n_{0k}\geq 1\). Its size is \(n_{1k}+n_{0k}\).
In the usual partition representation used here, every retained unit belongs to exactly one set. Units not retained in any set are unmatched. This convention makes it straightforward to count the sample and describe each set. If a design reuses a unit in multiple comparisons, simple counts of distinct units and counts of appearances in sets are no longer the same; the distinction must be made explicit.
Pair, Fixed-Ratio, and Full Matching
The names of common matching structures describe how treated and control counts vary across sets. “Ratio” refers to the relative counts within a set, not to a claim that the units have identical characteristics.
| Structure | Composition of each set | What varies across sets? |
|---|---|---|
| Pair matching | One treated and one control: \((1,1)\) | Neither count; each set has two units |
| Fixed-ratio matching | A prespecified count pattern, commonly one treated and \(m\) controls: \((1,m)\) | Neither count when the same ratio is used throughout |
| Full matching | At least one treated and at least one control in every set | Both counts, and therefore the ratio and set size, may vary |
Pair matching is the special case in which each set has exactly one unit from each group. In fixed-ratio matching, a common structure is \(1:m\): each set has one treated unit and the same prespecified number \(m\) of controls. The reverse orientation, \(m:1\), is also possible, depending on which group is assigned multiple units. If the available units do not permit every set to have the intended composition, the proposed fixed-ratio structure cannot be completed as stated.
Full matching permits different compositions across sets, such as one treated with three controls in one set and two treated with one control in another. Its defining structural requirement is that each set contains both treatment groups; there is no single treated-to-control ratio that every set must share. “Full” describes this flexible set structure, not a guarantee that every unit in the original data is retained.
Worked Example: Counting Pair-Matched Sets
Suppose a hypothetical analysis retains six units and arranges them as follows. The labels T and C indicate treated and control units.
| Set | Treated unit | Control unit |
|---|---|---|
| 1 | T1 | C1 |
| 2 | T2 | C2 |
| 3 | T3 | C3 |
Each row is a set with composition \((n_{1k},n_{0k})=(1,1)\). Thus there are three sets, three retained treated units, and three retained controls. The total number of retained units is \(3(1+1)=6\). Every set contains both treatment groups, and every set has the same two-unit composition, so this is pair matching. The table describes group membership; it does not show whether the treated and control units are similar on relevant pretreatment covariates.
Worked Example: Checking a Fixed One-to-Two Ratio
In a hypothetical design, each set is intended to contain one treated unit and two controls. The proposed groups are \(\{T1,C1,C2\}\), \(\{T2,C3,C4\}\), and \(\{T3,C5,C6\}\).
For each set, \(n_{1k}=1\) and \(n_{0k}=2\), so its composition is \((1,2)\) and its size is \(1+2=3\). Across the three sets, the retained treated count is \(3\times1=3\), and the retained control count is \(3\times2=6\). The total is \(3+6=9\), also equal to \(3\times3=9\) units across three sets. The proposed structure therefore satisfies the one-to-two fixed-ratio requirement.
If the second set instead contained only one control, it would have composition \((1,1)\), not \((1,2)\). The sample could still be described as a collection of matched sets if that set contained both groups, but it would not satisfy the stated fixed one-to-two ratio throughout.
Worked Example: Reading Variable Compositions in Full Matching
Consider these hypothetical full-matching sets: \(S_1=\{T1,C1,C2,C3\}\), \(S_2=\{T2,T3,C4\}\), and \(S_3=\{T4,C5,C6\}\).
The set compositions are \((1,3)\), \((2,1)\), and \((1,2)\). Their sizes are \(1+3=4\), \(2+1=3\), and \(1+2=3\), respectively. Summing the treated counts gives \(1+2+1=4\) treated units; summing the control counts gives \(3+1+2=6\) controls. The total retained count is \(4+6=10\), which agrees with the set-size check \(4+3+3=10\). Every set contains both groups, and the compositions vary, so the structure is full matching rather than pair matching or a single fixed ratio.
The fact that one set has two treated units does not invalidate the structure. Full matching requires both groups in each set; it does not require exactly one treated unit per set.
Why Set Structure Matters for Comparisons
For an outcome \(Y_i\), a set containing both groups allows a within-set descriptive contrast. For example, the treated and control outcome means in set \(k\) can be written as
The denominators are positive precisely because a valid matched set contains at least one unit from each group. In a pair, each mean is the outcome of a single unit. In a larger set, each mean averages multiple outcomes from its respective group. This establishes how a within-set descriptive contrast can be formed; it does not establish that \(\Delta_k\) is a causal effect.
To see why structure does not determine a single overall summary, suppose a design has several sets with different sizes. One could describe an overall result by giving each set equal influence or by allowing sets with more units to contribute more. These choices need not produce the same number. The scientific target, the design, and the chosen estimator must therefore be made explicit rather than inferred from the label “full matching.” Set structure is information about grouping, not a complete analysis plan.
A causal interpretation requires more than valid set counts. Matching is commonly based on pretreatment covariates, and causal conclusions still depend on assumptions such as consistency, no interference between units, conditional exchangeability given the covariates used for adjustment, and adequate treatment overlap in the population targeted by the analysis. A particular set structure does not prove those assumptions. Nor does merely placing units in the same set guarantee exact covariate balance.
Diagnostics, Choices, and Limitations
Begin by checking the structure directly. For each set, list \(n_{1k}\), \(n_{0k}\), and the set size. Confirm that both counts are positive, that the observed composition matches the design claimed, and that the sum of set sizes agrees with the number of retained units when units appear only once. Also report how many treated and control units were not matched. A label such as “fixed ratio” is not an adequate substitute for these counts.
Next, examine whether the grouping is useful for the intended comparison. Assess covariate balance in the matched sample, including within-set differences where relevant, rather than assuming that a set’s membership guarantees similarity. Look for unusually large sets or markedly uneven treated-control counts, and consider whether those features leave some comparisons dependent on a small number of units. A structurally valid set may still have poor covariate balance or weak support for the comparison.
The alternatives represent different design choices. Pair matching yields simple two-unit groups but can require discarding available units that do not fit pairs. A fixed ratio uses a constant composition and can include multiple controls per treated unit when that is feasible. Full matching accommodates varying compositions and can retain heterogeneous set sizes, but the resulting structure may be less uniform and requires care when defining summaries and interpreting the target population. No structure is automatically best: the appropriate choice depends on overlap, available units, the scientific target, and the analysis plan.
Matching also has limitations common to all three structures. It cannot balance unmeasured covariates by construction, and post-treatment variables should not be treated as baseline matching variables for estimating a treatment effect. Restricting the analysis to matched units can change the population represented by the analysis. The resulting estimate should not be generalized to all treated units without considering who was retained and the assumptions needed for that generalization.
Common Mistake
Wrong reasoning: “Every group has a treated unit and a control, so the groups must be comparable and the estimate must be causal.” The first part checks only the minimum structure of a matched set. It says nothing about how similar the units are on measured covariates, whether important confounders remain, or whether the causal assumptions hold.
Correction: Treat set membership and causal validity as separate questions. First verify the counts and matching structure. Then examine covariate balance and overlap, state the target population, and assess the assumptions required for a causal interpretation. A structurally valid set is necessary for a within-set treated-control comparison, but it is not sufficient for causal identification.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A design has four sets, each containing one treated unit and one control. How many treated units, controls, and total units are retained? What structure is this?
- Three sets each contain one treated unit and three controls. State the composition and size of each set, and calculate the total retained counts.
- A proposed set has two treated units and no controls. Is it a valid matched set under the definition here? Why or why not?
- Three sets have compositions \((1,2)\), \((2,1)\), and \((1,1)\). Is this pair matching, fixed one-to-two matching, or full matching?
- Why does a valid matched-set structure not, by itself, establish a causal effect?
Solutions. (1) There are \(4\times1=4\) treated units and \(4\times1=4\) controls, for \(4+4=8\) total units; this is pair matching. (2) Each set has composition \((1,3)\) and size \(1+3=4\). Across three sets there are \(3\) treated units, \(9\) controls, and \(12\) total units. (3) No. Its control count is zero, so it does not contain both treatment groups and cannot support a within-set treated-control comparison. (4) It is full matching: every set contains both groups, but the compositions vary, so neither pair matching nor a fixed one-to-two ratio describes all sets. (5) Set structure records which units are grouped and in what counts. It does not ensure covariate balance or establish causal assumptions such as conditional exchangeability, consistency, and no interference.