Full Matching and Its Implicit Target

Why Full Matching Has an Implicit Target

Full matching groups treated and control units into matched sets, with at least one unit from each treatment group in every set. A set may contain one treated unit and several controls, several treated units and one control, or larger numbers of both. This flexibility can retain units that a fixed-ratio design might leave unmatched, but it also raises an important question: how much does each set contribute to the estimate?

The answer depends on how the within-set contrasts are combined. If sets with many treated units receive more weight, the estimate emphasizes the retained treated population. If each set receives the same weight, a small set can matter as much as a large one. The matched sets alone do not select one universal target or one universal estimator.

Definition: Let \(A_i=1\) indicate that unit \(i\) is treated and \(A_i=0\) that it is a control. Let \(Y_i(1)\) and \(Y_i(0)\) be the unit’s potential outcomes under treatment and control, and let \(S_i=s\) identify its matched set. In set \(s\), write \(n_{1s}\) for the number of treated units and \(n_{0s}\) for the number of controls. A full-matched set has \(n_{1s}\geq 1\) and \(n_{0s}\geq 1\).

A common starting point is the observed outcome contrast in each set:

\hat{\tau}_s=\bar{Y}_{1s}-\bar{Y}_{0s}, \qquad \bar{Y}_{1s}=\frac{1}{n_{1s}}\sum_{i:A_i=1,S_i=s}Y_i, \qquad \bar{Y}_{0s}=\frac{1}{n_{0s}}\sum_{i:A_i=0,S_i=s}Y_i.

The contrast compares the mean observed outcome among treated units with the mean among controls in the same set. It is an observed-data contrast; it becomes evidence about a causal effect only under suitable assumptions. In particular, having both treatment groups in a set is not itself proof that their outcomes are exchangeable.

Assumptions for a Causal Interpretation

The causal interpretation uses the usual potential-outcomes conditions. Consistency means that a unit’s observed outcome equals its potential outcome under the treatment it actually received. No interference means one unit’s treatment does not change another unit’s potential outcome. In addition, treated and control units must be comparable with respect to their potential outcomes within the matched sets, after accounting for the measured covariates used in matching and any appropriate adjustment.

For a simple set-specific contrast to identify the treated units’ average effect in set \(s\), the relevant mean-exchangeability condition is that the controls provide the treated units’ mean untreated outcome: \(E[Y(0)\mid A=1,S=s]=E[Y(0)\mid A=0,S=s]\). This condition may be plausible with exact matching on sufficient covariates, or after adequate within-set adjustment. In approximate matching, merely sharing a set does not guarantee it. There must also be treated and control units in the relevant sets; full matching ensures this at the set level, but it cannot create comparable observations where the data lack overlap.

Under these conditions, the expected within-set contrast corresponds to the average treatment effect among treated units in that set:

E[\hat{\tau}_s] = E[Y(1)\mid A=1,S=s]-E[Y(0)\mid A=1,S=s].

The first mean is observed among treated units. Mean exchangeability allows the controls’ observed mean to represent the treated units’ otherwise-unobserved mean outcome under control. If that representation fails, the contrast can reflect baseline differences as well as treatment effects.

How Set Contributions Determine the Target

One common summary gives each treated unit equal influence at the set level: weight each set contrast by its treated count. With \(K\) matched sets and \(N_1=\sum_{s=1}^K n_{1s}\) retained treated units, this estimate is

\hat{\tau}_{T} = \sum_{s=1}^{K}\frac{n_{1s}}{N_1}\hat{\tau}_s.

The weights sum to one. If the set contrasts identify the effects among their treated units, the corresponding target is the average effect for the retained treated population, not automatically for every treated unit in the original population or for the full study population.

Justification: Substituting the set contrast into the treated-count-weighted summary gives

\hat{\tau}_{T} = \sum_{s=1}^{K}\frac{n_{1s}}{N_1} \left(\bar{Y}_{1s}-\bar{Y}_{0s}\right) = \frac{1}{N_1}\sum_{s=1}^{K}n_{1s}\bar{Y}_{1s} - \sum_{s=1}^{K}\frac{n_{1s}}{N_1}\bar{Y}_{0s}.

Since \(n_{1s}\bar{Y}_{1s}\) is the sum of treated outcomes in set \(s\), the first term is the mean observed outcome among all retained treated units. The second term combines each set’s control mean using that set’s share of the retained treated units. Under the stated comparability assumptions, that second term represents the mean untreated outcome for the retained treated target. This explains the target without implying that each control has equal influence in the estimate: a control mean in a set with many treated units supports more of the treated target than a control mean in a set with few treated units.

Other summaries answer different questions. An equally weighted average of the set contrasts gives every set the same influence, regardless of its treated count. A control-count-weighted average instead emphasizes sets according to their numbers of controls and, under analogous assumptions, may align with an effect for the retained controls. These are different summaries when treatment effects vary across sets. Thus, stating only that “full matching was used” is not enough to describe the estimand; the aggregation rule must also be stated.

Worked Example: A Treated-Weighted Full-Matching Estimate

Suppose a hypothetical study creates three full-matched sets. The outcome is a score, measured in points. Set 1 has two treated units with outcomes 10 and 12, and one control with outcome 8. Set 2 has one treated unit with outcome 15, and three controls with outcomes 13, 14, and 16. Set 3 has two treated units with outcomes 7 and 9, and two controls with outcomes 6 and 7. Assume, for this illustration, that outcomes are comparable within sets under the causal assumptions above.

The set contrasts are:

\hat{\tau}_1=\frac{10+12}{2}-8=11-8=3,

\hat{\tau}_2=15-\frac{13+14+16}{3}=15-\frac{43}{3}\approx 0.667,

\hat{\tau}_3=\frac{7+9}{2}-\frac{6+7}{2}=8-6.5=1.5.

There are \(N_1=2+1+2=5\) retained treated units. The treated-count-weighted estimate is

\hat{\tau}_{T} = \frac{2}{5}(3)+\frac{1}{5}(0.667)+\frac{2}{5}(1.5) \approx 1.933\text{ points}.

The interpretation, conditional on the assumptions, is an estimated average treatment effect of about 1.93 points for the five retained treated units represented by these sets. It is not automatically an estimate for all treated units who were eligible before matching. An equally weighted set average would be \((3+0.667+1.5)/3\approx1.722\) points, a different summary because it gives set 2 the same set-level influence as sets 1 and 3 despite its smaller treated count.

Worked Example: Changing the Aggregation Changes the Emphasis

Consider two hypothetical full-matched sets. In set 1, three treated units have outcomes 8, 10, and 12, while one control has outcome 6. In set 2, one treated unit has outcome 7, while three controls have outcomes 6, 7, and 8. Outcomes are in points, and assume within-set comparability for purposes of interpreting these contrasts.

The first set’s contrast is \((8+10+12)/3-6=10-6=4\) points. The second set’s contrast is \(7-(6+7+8)/3=7-7=0\) points. The treated-count-weighted average is

\frac{3}{4}(4)+\frac{1}{4}(0)=3\text{ points}.

An equal-set average is \((4+0)/2=2\) points. A control-count-weighted average is \((1/4)(4)+(3/4)(0)=1\) point. The three numbers differ because the set effects differ and the rules assign different influence to each set. The treated-weighted summary emphasizes the set containing three of the four retained treated units; the control-weighted summary emphasizes the set containing three of the four controls. The matched design has not changed, but the implicit target of the summary has.

Diagnostics and Failure Modes

The target interpretation is only as credible as the comparisons that support the set contrasts. Before presenting a causal estimate, inspect the matched data at the set level and across the matched sample. Useful checks include:

Covariate comparability within sets: examine whether treated and control units are close on important prognostic covariates, rather than relying only on the fact that they share a set.
Set composition: report the treated and control counts across sets. A few large sets may contribute much of the estimate under count-weighted aggregation.
Overlap and discarded units: describe which units did not enter the matched sample. The target of a matched estimate generally pertains to retained units, so exclusion can limit its relevance to the original population.
Dependence on the aggregation rule: compare reasonable set summaries when effects may vary. Large differences indicate that the result is sensitive to which sets are emphasized.

A diagnostic can reveal a problem without resolving it. For example, good average balance across all matched units can conceal poor comparisons in a particular influential set. Conversely, a set with many controls is not necessarily more informative about the treated target if it contains few treated units. Inspect both the set-specific contrasts and the covariate distributions that support them.

Worked Example: A Matched Set That Does Not Support a Causal Claim

Suppose one hypothetical full-matched set contains one treated unit with a baseline risk score of 9 and an observed outcome of 20 points, plus two controls with risk scores of 2 and 3 and outcomes of 12 and 13 points. The set contrast is

20-\frac{12+13}{2}=20-12.5=7.5\text{ points}.

The arithmetic is correct, but the causal interpretation is not justified by the set membership alone. The treated unit’s baseline risk score is far above either control’s score. If risk score predicts the outcome, the control mean may be a poor estimate of that treated unit’s untreated outcome. The observed 7.5-point contrast may therefore combine treatment effect and baseline risk differences.

The appropriate conclusion is not that the treatment increased the outcome by 7.5 points. First investigate the covariate mismatch and whether adjustment or a different design can produce credible comparisons. If no suitable controls exist for this kind of treated unit, the available data may not identify its effect without additional assumptions. Full matching can organize observed comparisons; it cannot supply missing overlap or remove unmeasured confounding by itself.

How Full Matching Relates to Alternatives

Full matching is one way to create comparison sets, not a guarantee of a particular causal estimand. Fixed-ratio matching imposes a chosen number of controls per treated unit and may leave some units unmatched; variable-ratio matching allows that number to vary. Full matching is more flexible about set composition, but its estimate still depends on which units are retained, how well they compare, and how set contrasts are aggregated.

A regression-adjusted analysis can account for residual measured differences, but its target depends on the population over which effects are summarized and on the model’s adequacy. Weighting approaches also define contributions through their weights. These approaches should be compared by their assumptions, diagnostics, and target—not simply by which estimate is numerically closest to another.

The central limitation is that a matched estimate is not automatically transportable to units excluded from the matched sample. Nor does a treated-weighted set summary become the overall average treatment effect merely because both treatment groups appear in every set. State the retained population and aggregation rule, and qualify the causal interpretation by the comparability assumptions.

Common pitfall: “Every full-matched set contains treated and control units, so the estimate is the ATE for the whole sample.” This is incorrect. Set membership does not ensure exchangeability, and the estimate’s target depends on which units are retained and how set contrasts are combined. A treated-count-weighted summary under appropriate assumptions targets the retained treated population; another aggregation can emphasize a different population.

Check Your Understanding

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

A full-matched set contains four treated units and one control. Under a treated-count-weighted summary, how does this set’s weight compare with a set containing one treated unit? What does that weight mean?
Two sets have estimated contrasts of 5 and 1 points. The first contains three treated units and the second contains one. Calculate the treated-count-weighted estimate.
Why does having at least one treated and one control in every set not, by itself, establish a causal interpretation?
When might an equally weighted set average differ from a treated-count-weighted estimate, and what does the difference indicate?

Solutions. (1) The first set receives four times the weight: its treated-count share is four units, compared with one in the other set. This reflects its larger share of the retained treated population, not proof that the set is better matched. (2) The estimate is \((3/4)(5)+(1/4)(1)=4\) points. (3) Treated and control units may still differ in outcome-related covariates; exchangeability and adequate overlap are needed for the controls to represent the treated units’ counterfactual outcomes. (4) The summaries differ when set contrasts differ and set sizes are unequal. The difference shows that the aggregation rule changes which sets—and thus which part of the matched population—has more influence.