Optimal Matching With Variable Treatment Ratios

Why Use Variable-Ratio Matching?

Matching pairs treated units with similar untreated units so that observed covariates are more comparable between groups. A fixed 1:1 match uses one control for each treated unit. But the number and quality of available controls can vary: one treated unit may have several close matches, while another has only one. Variable-ratio matching allows the number of controls to differ across treated units, subject to prespecified limits.

Using more than one control can make better use of available comparison units and may improve precision, especially when an additional control is close to the treated unit. It does not guarantee better balance or a more precise estimate in every data set. The distance criterion, ratio limits, allowable matches, retained population, and outcome analysis must all be specified and checked.

For a causal interpretation, the usual assumptions include consistency, conditional exchangeability, and positivity. Consistency means the observed outcome equals the potential outcome under the treatment actually received. Conditional exchangeability means \((Y(1),Y(0))\) is independent of \(A\) given measured \(X\); positivity requires \(0<e(X)<1\) for covariate profiles in the target population. These assumptions are not established by matching or by good balance diagnostics. Matching can improve comparability on measured covariates, but it cannot remove bias from unmeasured confounding.

Choosing and Minimizing a Distance Criterion

A matching algorithm needs a numerical rule for how close a treated-control pair is. One common choice is the absolute difference in estimated propensity scores. Another is the absolute difference in estimated logit propensity scores, \(\operatorname{logit}\{\hat e(X)\}=\log\{\hat e(X)/(1-\hat e(X))\}\), often divided by a scale such as the standard deviation of the logits. The scale and distance must be stated because the magnitude of a distance is meaningful only relative to its definition.

With variable ratios, the algorithm typically chooses both which controls go to each treated unit and how many controls each treated unit receives within allowed bounds. A common objective is to minimize the sum of treated-control distances across all assigned pairs, subject to constraints such as a minimum and maximum number of controls per treated unit, unique use of controls, and any caliper restriction. This is a global assignment problem: the best overall configuration need not result from giving each treated unit its individually closest control first.

Check Balance and the Matched Sets

After matching, assess balance on the covariates used to estimate the propensity score and on other important pretreatment covariates. A common summary is the standardized mean difference (SMD), the difference in group means divided by a standard deviation used as the scale. For a continuous covariate, a common denominator is the square root of the average of the two group variances. For a binary covariate, the difference in proportions can be standardized on a corresponding proportion-based scale. Report the definition used, and examine distributions as well as means when nonlinearities or extreme values matter.

For variable ratios, control summaries should respect the matched-set design. If each treated unit is to count equally for an ATT, each control in set \(i\) receives within-set weight \(1/k_i\), so that the controls in every set together have weight one. Assess balance using these weights or an equivalent matched-set summary, not by treating all matched controls as equally influential regardless of set size. Also inspect the distribution of pair distances, any caliper violations, the numbers of treated units retained and excluded, and whether some sets have unusually large ratios.

Estimating the Matched-Sample ATT

For a matched set \(i\) containing one treated unit and \(k_i\) controls, calculate the treated outcome minus the average control outcome. The matched-sample ATT estimate is the average of these set-level contrasts across the \(M\) retained treated units. This gives each treated unit equal weight, whether its set has one control or several. It targets the average treatment effect among the treated units retained by the matching design; if treated units were excluded, it need not target the ATT among all originally treated units.

The equal-set weighting matters. Pooling all treated-control pairs and averaging their outcome differences can give more influence to treated units with larger \(k_i\). That is a different weighting scheme and generally does not equal the ATT estimator defined above.

Uncertainty, Diagnostics, and Alternatives

The set-level standard-error calculation is most defensible when sets are independent, there are enough sets for an approximation to be useful, and the distribution of set-level contrasts is not severely irregular. If controls are reused, sets share observations, or matching creates dependence, the independence calculation does not apply as written. Even without replacement, a small number of retained sets can make a \(t\) approximation unstable. The propensity score is estimated, and the matching operation itself is data-dependent; an interval that ignores these features may not fully represent sampling uncertainty. The uncertainty method should match the design and estimand, and its assumptions should be reported.

Useful failure signals include poor weighted balance on important covariates, a concentration of large pair distances, many treated units left unmatched, highly variable set sizes, or sensitivity of the match to reasonable changes in the distance or ratio limits. A caliper can prevent matches that are too distant, but may also exclude treated units. A low total distance does not certify covariate balance: the optimization targets its chosen distance, not every measured covariate or the outcome.

Fixed-ratio matching uses the same number of controls for every treated unit, making its weighting and set structure simpler but potentially discarding useful controls or forcing weak matches. Variable-ratio matching relaxes that constraint while retaining an ATT-oriented equal-treated-unit weighting rule. Full matching, treated in a separate context, permits sets with different numbers of treated and control units and can use different weights for comparative estimands; it is not interchangeable with the one-treated-per-set ATT estimator described here. Weighting methods avoid explicit pair assignment but rely on their own overlap and weight-stability diagnostics.

Common Pitfall

A frequent error is to conclude that an optimal match must be balanced because it minimizes total propensity-score distance. The conclusion does not follow: the objective may be small while a covariate remains imbalanced, especially if that covariate contributes little to the chosen distance or if a few poor matches are offset by many very close ones. The correction is to treat optimization and diagnostics as distinct tasks: minimize the prespecified criterion, then report balance across relevant covariates, distance distributions, retention, and sensitivity to design choices.

A second error is to average all pair differences equally when some treated units have more controls. For an ATT that gives each retained treated unit equal influence, first average controls within each matched set, then average the set contrasts. Finally, describe the estimate as applying to the retained treated population unless the design supports a broader target.