Active-Control Trials and Reference-Arm Definition

Why Reference-Arm Definition Matters

In an active-control trial, participants may be randomized to an experimental treatment or to one of several established therapies. If an experimental-arm participant drops out before the outcome is fully observed, a reference-based analysis needs a distribution for the unobserved outcomes. Choosing a comparator is therefore part of specifying the missing-data assumption: it determines which observed response pattern supplies the predictive model after dropout.

An active control is not automatically an appropriate reference merely because it is an active treatment. Comparators can differ in dose, delivery, population, follow-up schedule, or expected response. A statistically convenient arm may not represent the clinical scenario the analysis intends to assume. The reference should be selected before examining how alternative choices affect the estimated treatment effect, and the rationale should state which participants, outcomes, and follow-up period the arm is meant to represent.

Definition: A reference arm is the trial arm whose specified outcome distribution is used to predict missing outcomes under a reference-based assumption. In an active-control trial, an eligible reference arm is an active comparator judged relevant to the target clinical scenario and sufficiently compatible with the analysis in population, treatment regimen, outcome definition, and follow-up.

The choice is not a claim that a participant who discontinued experimental treatment actually received the comparator. It is a modeling assumption about the distribution of outcomes that are not observed. The earlier material on selecting a reference arm in multiarm trials motivates choosing among trial arms; here the emphasis is on what that choice means when the candidate arms are active therapies.

Specify the Predictive Distribution

Let $H$ denote the participant’s observed outcome history up to dropout, $F$ the vector of later outcomes that are missing, $X=x$ the participant’s baseline covariates, and $r$ the selected reference arm. A common specification uses the reference arm’s joint outcome model, conditional on the same baseline covariates, to obtain a predictive distribution for $F$ given the observed history $H=h$.

For a multivariate normal model within reference arm $r$, write the conditional mean vector and covariance matrix as follows. The covariance blocks refer to the joint distribution of history and future outcomes under that arm.

\begin{pmatrix}H\\F\end{pmatrix}\,\bigg|\,X=x,\;r \sim N\left[ \begin{pmatrix}\mu_{H,r}(x)\\\mu_{F,r}(x)\end{pmatrix}, \begin{pmatrix} \Sigma_{HH,r} & \Sigma_{HF,r}\\ \Sigma_{FH,r} & \Sigma_{FF,r} \end{pmatrix} \right]

Conditional normal result: If $\Sigma_{HH,r}$ is nonsingular, the reference-arm distribution for the future outcomes given observed history $h$ is $$ F\mid H=h,\;X=x,\;r \sim N\left( \mu_{F,r}(x)+\Sigma_{FH,r}\Sigma_{HH,r}^{-1}\{h-\mu_{H,r}(x)\}, \;\Sigma_{FF,r}-\Sigma_{FH,r}\Sigma_{HH,r}^{-1}\Sigma_{HF,r} \right). $$

The first term in the mean is the reference arm’s expected future outcome at covariate value $x$. The adjustment uses how the observed history differs from that arm’s expected history, together with the reference arm’s association between history and future outcomes. The conditional covariance also comes from the reference model; it is not obtained by changing only the predicted mean.

This formula assumes the stated joint normal model is a reasonable description on the outcome scale being analyzed, the covariance blocks are estimated for the relevant history and visits, and the participant’s covariates and observed history can be evaluated under the reference model. Baseline covariates should be handled consistently across arms. In particular, the reference-arm prediction should be evaluated at the participant’s $x$, rather than silently substituting an average covariate profile. If the model is not normal, the same conceptual task remains—specify and use the reference arm’s conditional predictive distribution—but the normal formula does not apply automatically.

Worked Arm-Specific Predictions

Worked Example: Choosing Between Two Active Comparators

Hypothetical data. A trial compares an experimental therapy with two eligible active comparators, C1 and C2. The outcome is change from baseline in a symptom score, in points; higher values indicate greater improvement. An experimental-arm participant drops out after the week-4 assessment, which is $h=9$ points. At this participant’s baseline covariates, the fitted reference-arm models give the following summaries for week 4 ($H$) and week 8 ($F$):

For C1, $\mu_H=5$, $\mu_F=7$, $\operatorname{Var}(H)=16$, $\operatorname{Cov}(F,H)=4$, and $\operatorname{Var}(F)=9$. For C2, $\mu_H=3$, $\mu_F=5$, $\operatorname{Var}(H)=9$, $\operatorname{Cov}(F,H)=3$, and $\operatorname{Var}(F)=9$. Assume the normal models are adequate and these covariance estimates are positive and internally consistent.

With one observed history value, the conditional mean is $\mu_F+\operatorname{Cov}(F,H)\operatorname{Var}(H)^{-1}(h-\mu_H)$, and the conditional variance is $\operatorname{Var}(F)-\operatorname{Cov}(F,H)^2/\operatorname{Var}(H)$.

Using C1:

E(F\mid H=9,\mathrm{C1})=7+\frac{4}{16}(9-5)=7+1=8.00\text{ points}

The predictive variance is $9-4^2/16=9-1=8$ points squared, so the predictive standard deviation is $\sqrt{8}=2.828$ points.

Using C2:

E(F\mid H=9,\mathrm{C2})=5+\frac{3}{9}(9-3)=5+2=7.00\text{ points}

The predictive variance is $9-3^2/9=8$ points squared, again giving a predictive standard deviation of $2.828$ points. The two reference choices thus predict week-8 improvement of 8.00 points under C1 and 7.00 points under C2 for the same observed history and covariates. This difference arises from the reference arms’ means and history-to-future associations; it is not evidence that either prediction is the participant’s actual unobserved outcome.

Pooling Is a Separate Modeling Choice

A trial may have more than one clinically eligible active comparator. Selecting one arm is not the same operation as pooling arms. A pooled common model assumes that the data from the selected controls can support a shared reference distribution under the model’s specified covariate and treatment structure. If the arms differ meaningfully, an unadjusted pooled model can conceal those differences.

Pooling also needs a precise definition. A common multivariate normal model fitted to combined control data is not generally the same distribution as a mixture of arm-specific distributions. In a mixture, the predictive distribution can remain a mixture after conditioning on history; the relative contributions of the arms may depend on that history. The next example shows the result for one explicitly specified pooled model.

Worked Example: A Pooled Common-Normal Reference

Hypothetical calculation. Use the C1 and C2 summaries in the previous example, and suppose the analysis prespecifies a common bivariate normal reference model formed by combining the two arms with equal weights. For this calculation, the between-arm differences contribute to the pooled covariance. The pooled means are $\mu_H=(5+3)/2=4$ points and $\mu_F=(7+5)/2=6$ points.

The pooled variance of $H$ is the average within-arm variance plus the variance of the arm-specific means:

\operatorname{Var}_{\mathrm{pool}}(H) =\frac{16+9}{2}+\frac{(5-4)^2+(3-4)^2}{2} =12.5+1=13.5

The pooled covariance is the average within-arm covariance plus the covariance of the arm-specific means:

\operatorname{Cov}_{\mathrm{pool}}(F,H) =\frac{4+3}{2} +\frac{(7-6)(5-4)+(5-6)(3-4)}{2} =3.5+1=4.5

Thus the conditional pooled prediction at $h=9$ is

E_{\mathrm{pool}}(F\mid H=9) =6+\frac{4.5}{13.5}(9-4) =6+1.6667 =7.667\text{ points}

For completeness, the pooled future-outcome variance is $9+1=10$ points squared: the within-arm average is 9, and the arm means differ by one point on either side of 6. The pooled conditional variance is therefore $10-4.5^2/13.5=8.50$ points squared. This common-normal pooled model predicts 7.667 points, between the arm-specific predictions of 8.00 and 7.00 points, but it is not automatically the correct compromise. It is justified only if a common pooled reference distribution is a defensible assumption for the analysis.

Multiple Future Visits and Reference-Arm Choice

When several post-dropout visits are missing, the reference choice determines a vector of predicted means and a joint covariance matrix. Preserving that covariance matters because repeated outcomes from the same participant are related. The following calculation uses one observed history value and two future visits.

Worked Example: Predicting a Two-Visit Trajectory

Hypothetical model summaries. At a fixed baseline profile, let $H$ be the week-4 change from baseline, and let $F=(F_8,F_{12})^\mathsf{T}$ contain week-8 and week-12 changes, all in symptom-score points. The participant has $H=8$. For C1, $\mu_H=6$, $\mu_F=(7,8)^\mathsf{T}$, $\operatorname{Var}(H)=16$, $\operatorname{Cov}(F,H)=(4,6)^\mathsf{T}$, and $\Sigma_{FF}=\begin{pmatrix}9&6\\6&16\end{pmatrix}$. For C2, $\mu_H=4$, $\mu_F=(5,5.5)^\mathsf{T}$, $\operatorname{Var}(H)=9$, $\operatorname{Cov}(F,H)=(3,4)^\mathsf{T}$, and $\Sigma_{FF}=\begin{pmatrix}9&5\\5&12\end{pmatrix}$. Assume each joint normal covariance matrix is valid and the measurement schedule and outcome definition match across arms.

Under C1, the history adjustment is $(4,6)^\mathsf{T}(8-6)/16=(0.5,0.75)^\mathsf{T}$, so the conditional mean trajectory is $(7.5,8.75)^\mathsf{T}$ points. Its conditional covariance is

\begin{pmatrix}9&6\\6&16\end{pmatrix} -\frac{1}{16} \begin{pmatrix}4\\6\end{pmatrix} \begin{pmatrix}4&6\end{pmatrix} = \begin{pmatrix}8&4.5\\4.5&13.75\end{pmatrix} \text{ points}^2

Under C2, the adjustment is $(3,4)^\mathsf{T}(8-4)/9=(1.333,1.778)^\mathsf{T}$, giving a conditional mean trajectory of $(6.333,7.278)^\mathsf{T}$ points. Its conditional covariance is

\begin{pmatrix}9&5\\5&12\end{pmatrix} -\frac{1}{9} \begin{pmatrix}3\\4\end{pmatrix} \begin{pmatrix}3&4\end{pmatrix} = \begin{pmatrix}8&3.667\\3.667&10.222\end{pmatrix} \text{ points}^2

Both predictions use the same observed week-4 value, but their later trajectories and their week-8/week-12 dependence differ. A reference-arm definition therefore affects more than the predicted mean at one visit: it can alter the joint distribution used to generate complete missing-outcome trajectories.

Eligibility, Diagnostics, and Alternatives

Before using an active comparator as a reference, check whether the arm provides a coherent basis for the intended prediction. Relevant questions include whether its regimen and treatment setting represent the clinical scenario of interest, whether outcome definitions and visit times align, and whether the model can evaluate the participant’s covariates and observed history. These are substantive and statistical checks; a good numerical fit alone does not establish clinical relevance.

Useful diagnostics include comparing fitted arm-specific means and covariance patterns over the observed visits, examining predicted conditional means across the range of observed histories, and checking whether predictions are implausible on the outcome’s scale. Large discrepancies between eligible arms are a warning that a single pooled reference may hide important heterogeneity. If histories or visit schedules differ, ensure the conditional prediction uses the appropriate observed-history components rather than treating non-comparable records as identical.

Reference specification	What it assumes	Key implication
One active comparator	The selected arm is the relevant reference distribution for the target prediction.	Predictions retain that arm’s conditional means and covariance.
Pooled common model	The selected arms can be represented by the specified shared model.	Between-arm differences may affect pooled estimates; pooling is not neutral.
Mixture of arm-specific models	The reference distribution is a defined combination of arm-specific distributions.	The result can remain a mixture, and conditioning on history can change arm contributions.

These approaches answer different scientific questions. A single comparator is often easiest to interpret when one active therapy is the intended reference. Pooling can be useful when a common distribution is clinically and statistically defensible. A mixture may be appropriate when the target scenario explicitly comprises several alternatives, but its mixing weights and conditioning rules must be specified. None is universally preferred; the analysis should make the choice and its rationale transparent.

A frequent mistake is to say, “Both controls are active, so combine them and use the average response.” That reasoning ignores differences in response, covariance, and the definition of the target reference. The correction is to define clinical eligibility first, state whether the model selects one arm or combines arms, and then calculate predictions from that exact specification. A reference-based prediction is assumption-dependent, not identified solely by the observed data among participants who dropped out.

Key takeaway: Specify the active-control reference distribution before calculating post-dropout predictions. State the eligible comparator or the exact pooling or mixture rule, evaluate the model at the participant’s covariates and observed history, and use the chosen reference distribution’s conditional covariance as well as its mean.

Check Your Understanding

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

A reference arm has $\mu_H=4$, $\mu_F=6$, $\operatorname{Var}(H)=9$, $\operatorname{Cov}(F,H)=3$, and $\operatorname{Var}(F)=10$. For a participant with $H=7$, calculate the conditional mean and variance of $F$.
Why is a common pooled reference model not generally equivalent to choosing either of two active comparator arms?
Give two reasons that an active-control arm might be clinically ineligible as a reference even if its outcomes are measured in the trial.
In the conditional normal formula, what role does $\Sigma_{FH}\Sigma_{HH}^{-1}(h-\mu_H)$ play?

Solutions. (1) The conditional mean is $6+(3/9)(7-4)=6+1=7.00$ outcome units. The conditional variance is $10-3^2/9=10-1=9$ outcome units squared. (2) Pooling imposes a specified shared model and can incorporate between-arm differences into its means and covariances; an arm-specific model retains that arm’s own conditional distribution. (3) For example, the regimen may not represent the clinical scenario of interest, or the arm may have incompatible outcome definitions or follow-up times. Differences in population or dose can also make it unsuitable. (4) It adjusts the reference arm’s expected future outcome for how the participant’s observed history differs from that arm’s expected history, using the reference arm’s history-to-future association.