Why Match Trial Participants to External Controls?
An external-control cohort can contribute information about outcomes under a comparator treatment when a concurrent trial control group is small or unavailable. But trial participants and patients observed in routine care may differ at baseline. Matching selects external patients who are similar to trial participants on measured baseline characteristics, creating a more comparable sample for an outcome comparison.
Matching is a design choice, not a guarantee of causal validity. It can improve comparability on recorded covariates, but it does not make unmatched covariates disappear, correct errors in the recorded data, or establish that patients from different sources would have had the same outcomes absent treatment. The earlier material on estimating propensity scores with external controls provides the background for the score-based method below.
Before matching, define the treatment strategies, time zero, follow-up, outcomes, and baseline covariates consistently across sources. The covariates used for matching should be measured before treatment and chosen based on their relevance to confounding and outcome differences. Matching on post-treatment variables can introduce bias rather than remove it.
A causal interpretation also needs assumptions beyond the matching algorithm: consistency of the treatment strategies and outcomes; no important unmeasured confounding of the trial-treatment versus external-control comparison, conditional on \(X\); and adequate overlap, so relevant trial participants have comparable external controls. These assumptions can be especially demanding when data sources differ in care setting, measurement, or selection into observation.
Propensity-Score Matching and Calipers
A propensity score compresses the measured covariates into one number, but similar scores do not ensure that every covariate is well balanced in a finite matched sample. A common distance is the absolute difference in logit scores:
A caliper is a maximum permitted distance. A pair is eligible only if its distance is no larger than the chosen caliper. Calipers are often specified on the logit-score scale, sometimes relative to the standard deviation of the logit scores in the comparison sample. The caliper should be declared before inspecting outcomes. A narrow caliper can improve closeness but leave more trial participants unmatched; a wide caliper retains more participants but permits poorer matches. There is no universally optimal width.
Worked Example: Applying a Logit-Score Caliper
Suppose a hypothetical comparison has four trial participants, \(T1\) through \(T4\), and five external controls, \(C1\) through \(C5\). Their estimated propensity scores are shown below. Use a caliper of 0.30 on the absolute logit-score difference and match one control per trial participant, without replacement.
| Participant | \(\hat e(X)\) | Participant | \(\hat e(X)\) |
|---|---|---|---|
| T1 | 0.70 | C1 | 0.68 |
| T2 | 0.62 | C2 | 0.60 |
| T3 | 0.45 | C3 | 0.48 |
| T4 | 0.35 | C4 | 0.10 |
| C5 | 0.20 |
For T1 and C1, the logits are \(\log(0.70/0.30)=0.84730\) and \(\log(0.68/0.32)=0.75377\). Using the unrounded values, the absolute difference is approximately \(0.09353\), which is within the caliper. For T2 and C2, the logits are \(0.48955\) and \(0.40547\), giving a difference of approximately \(0.08408\). For T3 and C3, the logits are \(-0.20067\) and \(-0.08004\), giving a difference of approximately \(0.12063\). Each of these three pairs meets the caliper.
For T4, the logit is \(\log(0.35/0.65)=-0.61904\). C4 has logit \(\log(0.10/0.90)=-2.19722\), a difference of \(1.57819\); C5 has logit \(\log(0.20/0.80)=-1.38629\), a difference of \(0.76726\). Neither control is within 0.30, so T4 is unmatched. Under this rule, the matched sample contains three of the four trial participants and their three controls. This result illustrates a design trade-off: excluding T4 enforces the caliper but changes which trial participants the matched comparison represents.
The caliper does not determine every detail of the matching procedure. A nearest-neighbor algorithm selects the closest eligible control for each trial participant; the order in which participants are processed can matter. An optimal algorithm instead seeks the set of matches minimizing a total distance, subject to the specified constraints. With matching without replacement, a control can be used only once. With replacement, a particularly close control may be used more than once, which can improve closeness but requires the analysis to account for reuse.
Matching on Covariate Distances
Matching can also use distances calculated directly from baseline covariates. A simple standardized Euclidean distance divides each covariate difference by a scale chosen before matching, then combines the scaled differences. Scaling is important when covariates have different units: a raw difference of five years cannot be meaningfully combined with a raw difference of two severity points without a scale choice. Mahalanobis distance is another option; it accounts for the covariance among covariates. Whatever distance is chosen, the covariates and scaling rule should be clinically and statistically justified.
Worked Example: Matching on Standardized Covariate Distance
In a hypothetical example, trial participants and external controls have baseline age and disease-severity scores. For illustration, standardize age differences by 5 years and severity differences by 2 points, and calculate Euclidean distance. The trial participants are T1 \((60\text{ years}, 8\text{ points})\) and T2 \((55\text{ years}, 6\text{ points})\). The controls are C1 \((62,7)\), C2 \((56,5)\), and C3 \((58,9)\). Use a distance caliper of 1.0 and match without replacement.
The distance formula is \(d=\sqrt{(\Delta\text{age}/5)^2+(\Delta\text{severity}/2)^2}\). For T1 and C1, this gives \(\sqrt{(2/5)^2+(-1/2)^2}=\sqrt{0.41}=0.6403\). For T1 and C3, it gives \(\sqrt{(-2/5)^2+(1/2)^2}=0.6403\) as well. These are tied nearest eligible controls. For T2 and C2, the distance is \(\sqrt{(1/5)^2+(-1/2)^2}=\sqrt{0.29}=0.5385\).
One valid assignment is T1 to C1 and T2 to C2; both distances are below 1.0 and no control is reused. The same total distance results from assigning T1 to C3 and T2 to C2. A prespecified tie-breaking rule is needed to make the choice reproducible. This example shows why the distance definition and tie handling belong in the design specification, rather than being improvised after outcomes are examined.
A covariate-distance match can target closeness on specific variables, while propensity-score matching uses a summary of the measured covariates. Neither approach guarantees balance in all covariates or unmeasured factors. Direct distance matching can be preferable when a small set of covariates must be closely aligned; score matching can be convenient with many covariates, but can obscure which individual characteristics remain imbalanced. Researchers may also impose exact matching on a particularly important categorical covariate before applying another distance rule.
Matching Ratios and the Target of the Estimate
A matching ratio states how many controls are selected per trial participant. A 1:1 match uses one control per participant; a 1:2 match uses two. More controls can increase precision when suitable controls are available, but additional controls are typically less similar on average. The ratio is not a substitute for checking balance and overlap. The analysis must also reflect the matched-set structure and any repeated use of controls.
When the scientific target is the average treatment effect among matched trial participants, an intuitive matched estimate gives each retained trial participant equal influence and compares that participant’s outcome with the mean outcome of their matched controls. This defines an estimate for the retained participants, not automatically for every participant in the original trial.
Worked Example: A 1:2 Match and Its Target
Suppose a hypothetical study matches each of three trial participants to two distinct external controls. The outcome is a change in a clinical score, measured in points; higher values indicate greater improvement. The matched outcomes are: T1 has 10 points and controls have 7 and 8; T2 has 8 points and controls have 6 and 7; T3 has 9 points and controls have 8 and 7.
For each set, subtract the average control outcome from the trial participant’s outcome. For T1 the contrast is \(10-(7+8)/2=2.5\) points. For T2 it is \(8-(6+7)/2=1.5\) points. For T3 it is \(9-(8+7)/2=1.5\) points. Averaging the three set contrasts gives \((2.5+1.5+1.5)/3=1.833\) points.
This is a matched-sample outcome contrast. Its causal interpretation as an effect for these retained trial participants requires the causal assumptions, adequate covariate balance (or appropriate adjustment for remaining imbalance), and comparable outcome measurement. The arithmetic alone does not establish those conditions. If some trial participants were excluded because no suitable controls existed, the result does not directly estimate an effect for the full trial population.
Why Approximate Matching Does Not Prove Comparability
A short derivation makes the limitation explicit. Let \(Y(0)\) denote the outcome that would be observed under the control strategy, and suppose for simplicity that the outcome regression under control is \(m_0(X)=E\{Y(0)\mid X\}\). Under conditional exchangeability and consistency, an external control’s observed outcome can inform the control outcome at its own covariate values. But if matched trial and control covariate distributions differ, their expected control outcomes can still differ:
The right-hand side is zero if the matched groups have the same covariate distribution, or if the control outcome is unaffected by the remaining covariate differences. Approximate score closeness alone does not ensure either condition. Consequently, matching must be followed by assessment of covariate balance, and residual imbalance may require an appropriate adjustment strategy. Matching is also not a remedy for unmeasured confounding: if an important cause of both cohort membership and outcome is absent from \(X\), the displayed conditional comparison may remain biased.
Diagnostics, Alternatives, and Limitations
After matching, examine balance in each important baseline covariate, not only the propensity-score distribution. Compare covariate summaries and distributions between matched groups, inspect score or distance overlap, and report the number of trial participants retained and controls used. A small average difference can conceal substantial imbalance in a tail or subgroup. If important variables remain imbalanced, consider whether the matching specification is inadequate and whether an analysis that adjusts for the residual differences is justified.
Matching differs from inverse probability weighting and outcome regression. Matching forms a restricted sample of pairs or sets; weighting retains observations but changes their contribution to the analysis; outcome regression models the outcome conditional on covariates. These are distinct approaches with different assumptions, targets, and sensitivity to limited overlap. Matching is attractive for transparent design and direct inspection of comparability, but it can discard data and cannot guarantee that a treatment-effect estimate is unbiased.
The target population should be named explicitly. If matching discards trial participants, a simple matched contrast pertains to the retained trial participants, subject to the required causal and balance conditions. It cannot be generalized to excluded participants solely because they were originally randomized in the trial. Generalizing beyond the matched sample requires additional assumptions and an analysis that addresses the change in target population.
Check Your Understanding
Answer each question, then compare with the solutions below.
- Why does a caliper reduce the number of possible matches, and what is one consequence of choosing a tighter caliper?
- In a 1:2 match, what does the “2” represent? Does it mean the matched estimate automatically targets all trial participants?
- A trial participant has no external control within the prespecified caliper. What can be said about the matched comparison’s target population?
- Why is balance assessment needed even when propensity-score matching was used?
Solutions. (1) The caliper rules out pairs whose specified distance exceeds its maximum. A tighter caliper may improve closeness among retained pairs but can leave more trial participants unmatched. (2) It means two controls are selected per trial participant. It does not establish that the full trial population is represented; exclusions can change the target. (3) That participant is excluded under the stated matching rule, so the simple matched estimate describes retained participants rather than automatically describing the full trial. (4) Similar scores do not guarantee equality of every measured covariate, and they say nothing by themselves about unmeasured confounding. Covariate balance and overlap must be checked directly.