Defining the External-Control Estimand

Why the Estimand Comes First

An external-control analysis compares trial participants receiving an investigational treatment with patients from outside the trial who received a comparator treatment. Before selecting records or fitting a model, specify what contrast is of interest, for whom, on which outcome, and over what period. Together, these choices define the estimand: the target quantity the analysis is intended to learn about.

This distinction matters because two analyses can use the same treatment and the same external database but answer different questions. For example, a 6-month risk difference among all eligible patients is not the same quantity as a 12-month risk difference among patients with a particular disease stage. Statistical adjustment or Bayesian borrowing does not make these estimands interchangeable.

Writing the Estimand Formally

Let \(A=1\) denote assignment to the treatment strategy of interest and \(A=0\) assignment to the comparator strategy. For a person in the target population, let \(Y^a(\tau)\) denote the outcome that would be observed under strategy \(a\) by horizon \(\tau\). This notation describes potential outcomes; for each person, only the outcome under the strategy actually received is observed.

For a binary endpoint, such as death by 12 months, a common estimand is the population risk difference. With \(S^*=1\) indicating membership in the target population, it is

If the endpoint is an adverse event, a negative value means lower risk under treatment. If the endpoint is a favorable event, the same sign convention may have the opposite clinical interpretation, so the outcome and contrast direction must be stated explicitly. Other contrasts, such as a risk ratio, mean difference, or difference in restricted mean survival time, are possible; they target different summaries even when based on the same outcomes.

The notation also makes clear that the estimand is not an observed-data statistic. A sample proportion, adjusted estimate, or Bayesian posterior summarizes information about the target quantity under a model and its assumptions. It does not redefine the target quantity.

Four Choices That Determine Who Can Be a Control

Each part of the estimand places substantive constraints on external patients. These are not merely data-management choices: they determine whether a patient’s observed experience is relevant to the counterfactual comparator outcome in the estimand.

• Treatment contrast. Specify the treatment strategy and the comparator strategy, including clinically important differences in regimen or usual care. Patients who received a materially different comparator do not directly represent the comparator strategy as defined.
• Outcome. Specify the event or measurement, its definition, and how it is summarized. A database endpoint that is not comparable in meaning or ascertainment may not measure the outcome in the estimand.
• Time horizon. Specify when the outcome is assessed or the period over which it is summarized. A patient without observations covering the relevant period may not provide the required outcome information; shorter follow-up cannot simply be treated as a negative event.
• Target population. State the population to which the contrast applies, using clinically meaningful characteristics. External patients outside that population are not automatically suitable controls merely because their records are available.

Eligibility and the definition of the treatment strategies also have to be clinically coherent. If the target is patients meeting a particular disease-stage and biomarker profile, comparator records from patients with a different profile do not become relevant simply by increasing the external sample size. Detailed operational alignment of eligibility, time zero, treatment, and outcome definitions requires careful design; the estimand states what those design choices must serve.

Why the Target Population Changes the Contrast

A treatment contrast is generally an average over a population. If treatment effects or baseline outcome risks differ across patient groups, changing the population’s group proportions can change the average contrast. The standardization identity makes this explicit. For discrete covariate strata \(x\), write \(w_x=P(X=x\mid S^*=1)\). Then the target risk under strategy \(a\) is

The identity follows by partitioning the target population into its mutually exclusive strata and summing each stratum’s outcome risk multiplied by its share of the population. It shows why an external cohort’s unweighted outcome rate need not equal the target comparator risk: its mix of patients may differ from the target mix.

Assumptions Behind an External Comparator

A clearly stated estimand is necessary but not sufficient for identifying it from trial and nontrial data. The following assumptions describe why external outcomes might inform the comparator quantity; they are not guaranteed by having a large database.

• Consistency. A patient’s observed outcome under the strategy received corresponds to the potential outcome for that same strategy. This requires the strategy and outcome to be specified sufficiently well; materially different versions of “standard care” can undermine the interpretation.
• Conditional comparability. After accounting for relevant measured characteristics, comparator outcomes in the external source are informative about comparator outcomes in the target population. If important factors related to both source membership and outcome are unmeasured, the external comparison may remain biased.
• Positivity and overlap. Relevant patient profiles in the target population must be represented in the external data, and the comparator strategy must be observed for those profiles. If some target subgroup has no comparable external patients, its comparator outcome cannot be learned from that source without additional assumptions.
• Comparable measurement and follow-up. The external data must support a sufficiently similar definition and assessment of the outcome over the specified horizon. Differences in recording or incomplete follow-up can make an apparent contrast reflect measurement rather than treatment.

In a randomized trial, randomization supports comparison of trial treatment groups within the trial population, subject to the trial’s design and conduct. It does not automatically make external controls exchangeable with trial participants. The external comparison additionally relies on assumptions about measured and unmeasured differences between sources and about how outcomes transport to the target population.

Deriving the Transported Comparator Risk

A short derivation shows how target-population composition enters an external-control analysis. Suppose \(X\) is a set of measured baseline characteristics, and let \(S=0\) identify the external source. If the conditional comparator risk is transportable across source and target after conditioning on \(X\), then

Substitute this equality into the standardization identity for the target population:

Thus the external source supplies conditional comparator risks, while the target population supplies the weights. This result requires the stated conditional transportability assumption and adequate overlap; it is not a guarantee that adjustment removes bias. If key prognostic characteristics are unavailable or target strata are absent from the source, the equality or the calculation may not be defensible.

Diagnostics, Alternatives, and Common Pitfalls

Before interpreting an analysis, check whether the data can answer the stated question. Confirm that the comparator treatment, outcome definition, horizon, and target population are explicit. Compare important baseline characteristics and their ranges across sources, inspect whether target subgroups have external representation, and check whether outcome capture extends to the required horizon. A lack of overlap or incompatible outcome definitions is a substantive warning, not a problem solved merely by adding a covariate to a model.

The main alternatives differ in what they target or what data they use. A randomized concurrent control can provide a direct comparison within the trial population, while an external control may broaden information but requires additional comparability and transport assumptions. A crude external rate targets the source cohort’s observed mix, not necessarily the trial’s target population. Standardized or model-based comparisons can target a specified population, but depend on measured covariates, model adequacy, and overlap. Bayesian borrowing changes how information contributes to uncertainty and estimation; it does not change the treatment contrast or make an incompatible source relevant.

A common mistake is to define the estimand as “the difference between the trial arm and everyone available in the database.” This reverses the design logic: data availability does not define the scientific target. The correction is to specify the target population and comparator strategy first, then determine whether particular nontrial patients can supply relevant comparator information. Patients who fail the clinical target definition, received another strategy, or lack outcome information at the chosen horizon cannot be made suitable by increasing their number.

A further limitation is that even a well-specified and carefully aligned estimand may not be identifiable from available data. Unmeasured confounding, poor overlap, outcome misclassification, or unsupported extrapolation can prevent a credible causal interpretation. In that case, the analysis should report the limitation and narrow its claims rather than imply that the posterior distribution captures all sources of uncertainty.