External and Real-World Controls · Tutorial 395 of 696

Integrating External Controls With a Concurrent Randomized Arm

You will learn to define the target comparison, adjust external controls, and assess how assumptions about source differences affect the combined posterior evidence.

Advanced 12 min read

What You'll Learn

Define the treatment effect that remains anchored by randomization.
Distinguish adjusted external-control evidence from a concurrent randomized control.
Specify a Bayesian model for borrowing with an explicit source-difference parameter.
Calculate a posterior treatment estimate and interval from concurrent and external controls.
Assess how covariate adjustment, overlap, and borrowing assumptions affect the result.
Compare borrowing with concurrent-only analysis and naïve pooling.

Why Combine External and Concurrent Controls?

A randomized trial’s concurrent control group supports a direct comparison with the treated group because treatment assignment is randomized within the trial. External controls may add information about the control outcome, but they come from a different data source and were not randomized alongside the treated patients. Combining the groups can improve precision only if the analysis addresses differences in patient mix, outcome measurement, and care over time. The external data do not acquire the protection of randomization merely by being included in the same model.

The central question is therefore not simply how many control observations are available. It is how much the external controls should inform the control outcome for the trial population, given the evidence that the populations and data-generating processes are comparable. The target population should be specified first. In this tutorial, the target is the eligible population represented by the randomized trial, and the treatment effect is the average difference between assignment to the trial treatment and assignment to concurrent control in that population.

Definition: Integration of external controls means using eligible nontrial control observations, after appropriate adjustment and under explicit assumptions, to inform the control outcome in a randomized trial. The primary treatment comparison remains the effect of randomized treatment assignment versus concurrent control in the specified trial population. External controls supplement the estimate; they do not create a randomized comparison with the treated group.

Define the Comparison Before Borrowing

Let \(A_i=1\) indicate assignment to the trial treatment and \(A_i=0\) indicate assignment to concurrent control. Let \(S_i=0\) identify trial participants and \(S_i=1\) identify external controls. Let \(X_i\) denote measured baseline covariates, and let \(Y_i\) be the outcome at the prespecified assessment time. The treatment effect of interest may be an average mean difference, a risk difference, or another clinically defined contrast. The estimand must specify the outcome, time point, treatment strategy, and target population.

A useful outcome model for a continuous endpoint is

Y_i = \alpha + \beta A_i + \boldsymbol{\gamma}^{\mathsf T}X_i + \delta S_i + \epsilon_i,\qquad \epsilon_i \sim N(0,\sigma^2).

Here \(\beta\) is the treatment contrast conditional on the included covariates under the model, while \(\delta\) represents a source difference between external and trial controls after covariate adjustment. The model assumes a common treatment effect across covariate values unless treatment-by-covariate interactions are added. It also assumes that the source difference can be represented adequately by the stated source term. A more elaborate model may be needed if outcomes are clustered, binary, or time-to-event.

For a population-average effect, the analysis should align comparisons to the trial target population. One approach is to model outcomes conditional on \(X_i\) and standardize predicted outcomes to the trial covariate distribution. Another is to construct transport weights for eligible external controls so their measured covariate distribution resembles that of the target trial population. Either approach requires adequate overlap: covariate patterns common in the trial must be represented in the external data. A conditional regression coefficient is not automatically the same as a marginal population-average effect, particularly for nonlinear outcome models.

Assumptions to assess: The treatment effect is identified by randomization within the trial, with appropriate handling of missing outcomes and protocol deviations for the chosen estimand. External controls are eligible and have harmonized outcomes and covariates. After adjustment, remaining outcome differences between sources are sufficiently small for the planned degree of borrowing. The external data have support in the trial target population, and relevant time trends or care differences are addressed. These assumptions are substantive; a prior distribution cannot make them true.

Adjustment and the Role of the Source Difference

Covariate adjustment is intended to make the external controls informative about the control outcome in the trial population, not to erase every distinction between data sources. Baseline severity, age, prior treatment, calendar time, and eligibility criteria may differ. Adjusting for measured covariates can address observed differences, but residual confounding and source-specific measurement or care remain possible. The earlier material on transport weights and evaluating external-control selection bias provides relevant context for assessing these issues.

A Bayesian analysis can encode uncertainty about the residual source difference. For example, assign \(\delta \sim N(0,\tau^2)\). A small \(\tau\) says that large source differences are considered unlikely and allows stronger borrowing; a large \(\tau\) allows substantial source differences and reduces borrowing. This is a modeling choice that should be clinically justified and examined in sensitivity analyses. It is not evidence that the sources are exchangeable.

The following simplified normal model illustrates the calculation after external-control outcomes have been adjusted to the trial covariate distribution. Let \(T\) be the estimated treatment-group mean, \(C\) the concurrent-control mean, and \(E\) the adjusted external-control mean. Suppose their sampling variances are \(v_T\), \(v_C\), and \(v_E\). Write

T \sim N(\mu+\beta,v_T),\qquad C \sim N(\mu,v_C),\qquad E \sim N(\mu+\delta,v_E),\qquad \delta \sim N(0,\tau^2).

Here \(\mu\) is the target-population control mean, and \(\beta\) is the treatment-minus-control mean difference. This model assumes independent sampling errors across the three disjoint groups, known or suitably estimated variances, and a common target-population mean after adjustment. With a flat prior for \(\mu\), integrating over \(\delta\) gives \(E\mid\mu \sim N(\mu,v_E+\tau^2)\). The external estimate thus contributes less precision when the assumed source difference is more uncertain.

Deriving the Combined Posterior

The concurrent and adjusted external control estimates are conditionally independent normal observations of \(\mu\), with variances \(v_C\) and \(v_E+\tau^2\). Multiplying their normal likelihoods and completing the square yields the posterior for \(\mu\). With a flat prior for \(\beta\), the treated-group mean contributes information about \(\beta\), not about \(\mu\): for any \(\mu\), the unrestricted \(\beta\) can account for the observed treated mean.

\mu\mid C,E \sim N(m_\mu,V_\mu),\qquad V_\mu=\left(\frac{1}{v_C}+\frac{1}{v_E+\tau^2}\right)^{-1},

m_\mu=V_\mu\left(\frac{C}{v_C}+\frac{E}{v_E+\tau^2}\right).

Conditional on \(\mu\), the treatment contrast is \(T-\mu\) with sampling variance \(v_T\). Averaging over uncertainty in \(\mu\) gives

\beta\mid T,C,E \sim N(T-m_\mu,\;v_T+V_\mu).

The posterior variance contains both uncertainty in the treatment mean and uncertainty in the target control mean. The formulas show precisely how borrowing operates: the adjusted external estimate receives precision \(1/(v_E+\tau^2)\), rather than its sampling precision alone. When \(\tau\) is large, its contribution approaches zero.

Worked Example: Posterior Estimate With Partial Borrowing

Suppose a hypothetical randomized trial measures a symptom score at six months, where lower scores are better. The treatment group has \(n=100\), mean \(T=10.2\), and standard deviation \(4\). The concurrent-control group has \(n=50\), mean \(C=12.0\), and standard deviation \(4\). After covariate adjustment and standardization to the trial population, hypothetical external controls have \(E=12.5\), \(n=100\), and standard deviation \(4\). For illustration, treat these standard deviations as known and set \(\tau=1.5\) score points. Assume the adjustment is adequate, the outcome measurement is harmonized, and the independent normal sampling model is reasonable.

The sampling variances are \(v_T=4^2/100=0.16\), \(v_C=4^2/50=0.32\), and \(v_E=4^2/100=0.16\). The external-control variance after allowing for source difference is \(v_E+\tau^2=0.16+2.25=2.41\). Therefore,

V_\mu=\left(\frac{1}{0.32}+\frac{1}{2.41}\right)^{-1} =(3.125+0.4149)^{-1}=0.2825.

The combined target-control mean is

m_\mu=0.2825\left(\frac{12.0}{0.32}+\frac{12.5}{2.41}\right) =0.2825(37.50+5.187)=12.06.

Thus the posterior mean treatment contrast is \(10.2-12.06=-1.859\) score points, and its posterior standard deviation is \(\sqrt{0.16+0.2825}=0.6652\). A normal 95% credible interval is \(-1.859\pm1.96(0.6652)\), or approximately \((-3.163,-0.555)\) points. Under this model and prior, the posterior probability that treatment lowers the mean score is \(P(\beta<0\mid\text{data})\), approximately 0.997. This is a probability statement about the parameter under the specified Bayesian model, not a frequentist coverage claim. The result concerns a mean difference in symptom-score points; its clinical importance requires a clinically meaningful threshold.

Worked Example: Why Naïve Pooling Can Mislead

Consider a different hypothetical setting with a continuous outcome for which lower is better. The randomized trial has 40 concurrent controls with mean 12.0 and 50 treated participants with mean 10.5. Sixty external controls have mean 15.0 because, before adjustment, they are older and more severely ill. Simply pooling all 100 controls gives

\bar{Y}_{\mathrm{pooled\ control}} =\frac{40(12.0)+60(15.0)}{100} =\frac{1380}{100}=13.8.

The naïve treatment-minus-control contrast is \(10.5-13.8=-3.3\) outcome units. The concurrent randomized comparison is \(10.5-12.0=-1.5\) units. The difference between these estimates is not evidence that external controls reveal a larger treatment effect; it can arise because the unadjusted external group has a different prognosis. Suppose, for illustration, covariate adjustment standardizes the external mean to 12.4 in the trial population. That adjusted value is closer to the concurrent mean, but it does not alone establish exchangeability or justify giving the external group full weight. A borrowing model should still account for residual uncertainty about source differences.

Worked Example: Sensitivity to the Borrowing Assumption

Return to the six-month symptom-score example. Changing \(\tau\) changes the influence of the adjusted external estimate while leaving the observed trial data fixed. If \(\tau=0\), the model assumes exact equality of the adjusted external and trial control means. Then \(v_E+\tau^2=0.16\), so \(m_\mu=(12.0/0.32+12.5/0.16)/(3.125+6.25)=12.33\), and \(V_\mu=1/9.375=0.1067\). The treatment posterior mean becomes \(10.2-12.33=-2.133\) points, with posterior standard deviation \(\sqrt{0.16+0.1067}=0.5164\).

With \(\tau=1.5\), the earlier calculation gives a posterior mean of \(-1.859\) and standard deviation \(0.6652\). If \(\tau=5\), the external variance is \(0.16+25=25.16\), so its precision is only \(1/25.16=0.03975\). The resulting \(m_\mu\) is approximately 12.01 and \(V_\mu\) approximately 0.3160. The treatment posterior mean is about \(-1.806\), with standard deviation \(\sqrt{0.16+0.3160}=0.6899\). As the allowed source difference grows, the result approaches the concurrent-only analysis: \(T-C=-1.8\), with standard deviation \(\sqrt{0.16+0.32}=0.6928\). This sensitivity is important because the data alone do not determine the appropriate value of \(\tau\).

Diagnostics, Alternatives, and Limitations

Before interpreting a combined posterior, check whether the planned adjustment and model are credible. Compare trial and external covariate distributions, verify overlap, and investigate calendar-time and endpoint-ascertainment differences. Examine whether the adjusted external outcome is compatible with the concurrent control outcome, while recognizing that agreement on measured summaries cannot rule out unmeasured differences. Posterior predictive checks can reveal model misfit, but they cannot prove the absence of selection bias or residual confounding.

Concurrent-only analysis relies on the randomized comparison and avoids assumptions about external controls, but may be less precise when the concurrent control group is small.
Naïve pooling treats external controls as if they were trial controls. It can give excessive weight to systematic source differences and is generally not defensible without strong justification.
Adjusted borrowing uses external outcomes after measured-covariate adjustment and represents residual source uncertainty explicitly. Its validity still depends on model form, overlap, and assumptions about unmeasured differences.

Warning signs include poor covariate overlap, unstable or extreme transport weights, substantial disagreement between adjusted external and concurrent outcomes, sensitivity to reasonable modeling choices, and endpoint definitions that differ across sources. If overlap is weak, the analysis may rely on extrapolation rather than observed comparisons. If care or outcome ascertainment changed over time, a single source offset may not be adequate. If treatment effects vary across patient groups, a single common \(\beta\) may not describe the target population effect. In such situations, report the limitation and consider reducing or omitting external borrowing rather than treating a narrow interval as reassurance.

A common mistake is to say, “The external controls were adjusted, so pooling them with trial controls is equivalent to having randomized more patients.” Adjustment addresses measured differences under a specified model; it does not randomize the external patients or eliminate unmeasured differences. The correction is to retain the concurrent randomized comparison as the anchor, specify how external outcomes are made relevant to the trial population, and show how conclusions change when residual source differences are allowed to be larger.

Key takeaway: External controls can sharpen a randomized trial’s estimate when their outcomes are aligned to the trial population and their residual source difference is represented transparently. The amount of borrowing is an assumption-sensitive modeling decision, not a consequence of sample size alone.

Check Your Understanding

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

In the normal model, what does \(\tau=0\) assert about the adjusted external and trial control means?
Why can naïve pooling create an exaggerated treatment contrast when external controls have worse baseline prognosis?
If \(\tau\) increases substantially, what happens to the external controls’ contribution to the posterior precision for \(\mu\)?
State one reason covariate adjustment may fail to make an external control outcome comparable to the trial control outcome.

Solutions. (1) It asserts exact equality of the adjusted external and trial control means, \(\delta=0\); it is a strong assumption. (2) Worse-prognosis external controls may have a higher outcome mean, shifting the pooled control mean away from the trial’s target population and making a treatment-minus-control contrast appear more favorable when lower outcomes are better. (3) The external variance \(v_E+\tau^2\) increases, so its precision \(1/(v_E+\tau^2)\) decreases. (4) Adjustment may not account for unmeasured prognostic factors, inadequate covariate overlap, changing care, or differences in outcome ascertainment.

Back to Tutorials