What Does Treatment-Effect Surrogacy Across Arms Mean?
An intermediate outcome can help select treatments before the final endpoint is mature. But a useful selection outcome and a treatment-effect surrogate are not the same claim. Treatment-effect surrogacy asks whether differences between treatments in their effects on the intermediate outcome track differences between treatments in their effects on the final endpoint. It is a statement about treatment effects across arms, not simply about whether people with favorable intermediate outcomes tend to have favorable final outcomes.
For example, a marker may be strongly associated with survival among patients, yet a treatment that changes the marker may have little or no effect on survival. Conversely, an intermediate outcome need not explain every pathway to the final endpoint to be useful for prioritizing arms. The strength and purpose of the claim must therefore be stated: evidence that arm effects move together is not proof that changing the intermediate outcome causes the final benefit.
The scales and signs matter. Suppose larger values of an intermediate score are better. One possible effect is \(\theta_{jS}=E(S\mid j)-E(S\mid 0)\), in score points. For a time-to-event endpoint analyzed with a hazard ratio under a proportional-hazards model, define \(\theta_{jF}=-\log(HR_j)\). Then positive values of both effects indicate benefit. The final effect is dimensionless, whereas the intermediate effect has score units; the slope \(b\) consequently has final-effect units per score point. If the intermediate is a time, binary outcome, or other measure, its treatment effect and scale must be defined accordingly rather than assumed to be a mean difference.
What Can Be Learned From Several Arms?
In a multi-arm trial, the experimental arms provide several treatment-effect contrasts against the same control. Plotting estimated intermediate effects against estimated final effects can reveal whether the arms rank similarly and whether larger intermediate effects tend to accompany larger final effects. This is a useful description of the trial’s treatment pattern. It is not, by itself, a reliable estimate of a general relationship across treatments.
The arm-level effects are usually estimates, not known values. Their precision can differ, and their errors are dependent when they use a shared control. A regression of estimated final effects on estimated intermediate effects that ignores these features may misrepresent the relationship. Moreover, with only a few experimental arms, one arm can determine the apparent slope or correlation. A visually clear trend can coexist with substantial uncertainty.
Worked Example: Describing a Trend Across Three Arms
Suppose a hypothetical randomized trial has three experimental arms and a control. The intermediate measure is a symptom score for which higher is better. For the final time-to-event endpoint, assume proportional hazards and use \(-\log(HR)\), so positive values represent benefit. The observed arm-level estimates are:
| Arm | Intermediate effect (score points) | Final HR | Final effect, \(-\log(HR)\) |
|---|---|---|---|
| A | 2 | 0.9512 | 0.0500 |
| B | 4 | 0.8607 | 0.1500 |
| C | 6 | 0.7788 | 0.2500 |
For example, arm B has \(-\log(0.8607)\approx0.1500\). The three plotted points are \((2,0.05)\), \((4,0.15)\), and \((6,0.25)\). Their means are \(\bar{x}=4\) score points and \(\bar{y}=0.15\). The descriptive least-squares slope is
The intercept is \(\hat a=\bar{y}-\hat b\bar{x}=0.15-(0.05)(4)=-0.0500\). In this hypothetical dataset, each additional score point in the intermediate treatment effect accompanies \(0.05\) more units of estimated final effect. The fitted line describes these three estimates only. It does not establish that the relationship holds for other treatments, and the small number of arms does not support a confident general surrogacy claim.
Why a Shared Control Matters
When every experimental effect is a treatment-minus-control contrast, estimated effects share the control estimate. This creates dependence even if the experimental-arm observations are independent. Treating the contrasts as independent can therefore distort uncertainty calculations for their across-arm relationship.
Worked Example: Covariance From a Shared Control
Suppose two experimental arms and a control each have \(n=50\) independent participants, and the outcome has variance \(\sigma^2=100\) in each arm. For an outcome mean, write the estimated contrasts as \(\hat\theta_1=\bar X_1-\bar X_0\) and \(\hat\theta_2=\bar X_2-\bar X_0\). Assume equal variances and independent observations between arms.
The variance of each estimated contrast is \[ \operatorname{Var}(\hat\theta_j)=\frac{100}{50}+\frac{100}{50}=4. \] The covariance is \[ \operatorname{Cov}(\hat\theta_1,\hat\theta_2) =\operatorname{Cov}(\bar X_1-\bar X_0,\bar X_2-\bar X_0) =\operatorname{Var}(\bar X_0) =\frac{100}{50}=2. \] Thus their correlation is \(2/\sqrt{4\cdot4}=0.500\). The positive covariance arises because a control mean that happens to be low makes both treatment-minus-control estimates larger. These calculations illustrate the shared-control mechanism; in a real analysis, the covariance must match the actual design, endpoint, and estimator.
The same issue can arise for contrasts on both endpoints. Their joint sampling uncertainty, including dependence across arms and possibly across endpoints, should be considered when making formal statements about a relationship. A simple plot remains useful for exploration, but it does not account for this uncertainty. In particular, ordinary regression on noisy estimated effects is not automatically a valid inferential analysis.
A Short Justification for the Across-Arm Relationship
The reason to compare treatment effects, rather than patient outcomes alone, follows from the scientific question. Let \(S\) denote the intermediate outcome and \(F\) the final endpoint. An individual-level association concerns how \(S\) and \(F\) vary among individuals. Treatment-effect surrogacy instead compares the causal contrasts induced by assigning each treatment arm versus control. Those are different quantities: the first can arise from patient characteristics or disease severity, while the second concerns treatment-induced differences.
For a finite set of treatments, the working line \(\theta_{jF}=a+b\theta_{jS}+\varepsilon_j\) summarizes how their effects align. If \(b>0\), larger favorable intermediate effects tend to accompany larger favorable final effects, given the chosen orientations. The residual \(\varepsilon_j\) represents departures from that trend, including treatment-specific effects on the final endpoint not captured by the intermediate effect. A small residual pattern in one trial is not proof that such departures will remain small for new treatments.
This is an across-treatment argument, not a mediation proof. Even if the estimated arm effects align, the intermediate outcome may not be the mechanism through which treatment affects the final endpoint. To make a causal mediation claim requires additional assumptions and evidence beyond an across-arm relationship.
Single-Trial Evidence: Useful but Limited
A single multi-arm trial can provide a preliminary check of whether its treatments show compatible effects on the intermediate and final endpoints. It can expose obvious contradictions—for example, an arm with a substantial intermediate benefit but no corresponding final benefit. But it usually contains few distinct treatment effects, all tested in one population and under one trial’s implementation. It cannot, on its own, establish that the relationship will generalize to new treatments, doses, populations, or settings.
Precision is a further limitation. The final endpoint may have few events at the time of analysis, and the intermediate effect estimates can also be noisy. Inferences about the across-arm slope must reflect uncertainty in both axes rather than treating intermediate effects as error-free predictors. A shared control adds covariance, and selecting or emphasizing arms because their observed estimates look favorable can make the displayed relationship more optimistic than the underlying pattern. Therefore, distinguish a descriptive trend in observed estimates from evidence for a reproducible treatment-level relationship.
Worked Example: A Contradictory Arm Limits the Claim
Consider a second hypothetical trial with four experimental arms. The intermediate effect is again measured in score points with higher values favorable, and final benefit is \(-\log(HR)\) under a proportional-hazards model.
| Arm | Intermediate effect (score points) | Final HR | Final effect, \(-\log(HR)\) |
|---|---|---|---|
| A | 1 | 0.95 | 0.0513 |
| B | 2 | 0.90 | 0.1054 |
| C | 3 | 0.85 | 0.1625 |
| D | 4 | 1.05 | -0.0488 |
For arm D, \(-\log(1.05)\approx-0.0488\), so its estimated final effect points toward harm even though its intermediate effect is the largest. Arms A through C show a favorable ordering on both endpoints, but arm D contradicts a simple monotone pattern. The right conclusion is not that the intermediate outcome is a validated surrogate. Rather, this trial contains mixed arm-level evidence that warrants investigation of uncertainty, endpoint timing, and treatment-specific effects. The estimates alone do not identify which explanation is correct.
Diagnostics, Alternatives, and Common Mistakes
Before interpreting an across-arm pattern, check whether the treatment effects are defined on comparable scales and in consistent directions. Confirm that the same control and endpoint estimands are used across arms, that the intermediate and final outcomes refer to appropriate populations and times, and that follow-up and outcome ascertainment do not differ in ways that create artificial contrasts. For a hazard-ratio effect, the proportional-hazards interpretation should be defensible; otherwise, a different final-effect measure may be more suitable.
Inspect the arm-level estimates together with their uncertainty, not just a correlation or fitted line. Look for influential arms, a restricted range of intermediate effects, opposing directions, and unequal precision. If the intermediate effects are nearly identical across arms, the trial offers little information about whether variation in those effects tracks variation in final effects. If a conclusion changes when one arm is removed, the relationship is fragile. These checks diagnose limitations; they do not turn a small set of arms into a validation study.
The main alternatives answer different questions. Individual-level prognostic association asks whether people with better intermediate outcomes tend to have better final outcomes; it does not establish treatment-effect surrogacy. A treatment-by-intermediate comparison may describe differences within a trial but is not the same as a relationship across treatment effects. External evidence across multiple trials can provide more distinct treatment contrasts and settings for evaluating generalizability, although comparability and uncertainty still matter. A joint model can represent linked outcomes and their dependencies, but modeling dependence alone does not establish surrogate validity. In this tutorial, the central comparison is the treatment effects themselves and the limits of interpreting their pattern.
- Common mistake: “The intermediate outcome predicts survival for individual patients, so it is a treatment-effect surrogate.”
- Correction: Individual-level prediction is not evidence that treatments producing larger intermediate effects will produce larger final effects. Evaluate the arm-level treatment contrasts directly.
- Common mistake: “The three treatment points lie on a line, so the surrogate is validated.”
- Correction: A small number of arms in one trial gives limited information about uncertainty and generalizability. Describe the observed pattern as preliminary evidence, not validation.
- Common mistake: “Each treatment-versus-control estimate is independent because the experimental arms are separate.”
- Correction: Contrasts sharing a control are generally correlated. Account for that covariance in formal inference.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- A biomarker is strongly associated with event-free survival among patients. Does that establish treatment-effect surrogacy across arms?
- For a final endpoint with \(HR=0.80\), calculate \(-\log(HR)\) and interpret its direction.
- Two mean treatment contrasts use the same control. What feature of their estimates must be considered when assessing an across-arm relationship?
- Three arms have almost identical intermediate effects but different estimated final effects. What can the trial establish about the slope relating the effects?
Solutions. (1) No. That is an individual-level association; treatment-effect surrogacy requires comparing treatment effects across arms. (2) \(-\log(0.80)\approx0.2231\). Under the stated proportional-hazards scale, this is a positive estimated final benefit relative to control. (3) Their sampling errors are generally correlated because both contrasts include the same control estimate; the covariance should be reflected in formal inference. (4) Very little about the slope: with almost no spread in intermediate effects, the trial provides little information about how variation in those effects relates to variation in final effects.