Why Specify Contrasts Before an Arm Is Dropped?
In a multi-arm trial, an interim decision may discontinue one treatment condition while other arms continue. The decision changes which treatments are assigned to future participants, but it does not automatically change the scientific question the trial set out to answer. A contrast specified in advance can remain exactly the same target even when one of the trial’s arms stops enrolling.
The key distinction is between a contrast’s definition and the data available to estimate it. If a discontinued arm is not part of a contrast, dropping it need not alter that contrast. If the contrast includes the discontinued arm, the target still has its original meaning, but ongoing randomization no longer supplies new observations under that treatment condition. Replacing the discontinued arm with a different comparator or deleting its contribution changes the contrast rather than preserving it.
For example, if \(C\) is control and \(A\) and \(B\) are experimental treatments, the contrast \(\psi_A=\mu_A-\mu_C\) asks about treatment \(A\) versus control. Its interpretation depends on those named conditions and the fixed population and outcome summary. It does not become a contrast of \(A\) versus whichever arm remains available.
What Arm Dropping Changes—and What It Does Not
An arm can be removed from future assignment without changing the treatment condition as originally defined. The planned contrast coefficients likewise do not change merely because the allocation scheme changes. Thus, if a discontinued arm has coefficient zero in a contrast, it contributes nothing to that contrast both before and after discontinuation.
If the discontinued arm has a nonzero coefficient, the contrast remains a well-defined target under the original treatment conditions. However, once no participants are assigned to that treatment condition, the trial does not collect new randomized outcomes under it. This distinction matters: “the estimand still has a meaning” does not imply “the continuing trial can estimate it from new participants without additional assumptions or data.”
Worked Examples
Worked Example: Dropping an Arm Outside the Contrast
Suppose a hypothetical three-arm trial prespecifies a symptom score at 12 weeks, with lower scores indicating better outcomes. The treatment conditions are control \(C\), treatment \(A\), and treatment \(B\), and the target population and intercurrent-event strategy are fixed. The two planned contrasts are \(\psi_A=\mu_A-\mu_C\) and \(\psi_B=\mu_B-\mu_C\). At an illustrative stage, the corresponding target-population means are 10.0, 7.5, and 8.2 points.
For treatment \(A\) versus control, the contrast is:
Suppose \(B\) is discontinued. The coefficient of \(\mu_B\) in \(\psi_A\) is zero, so the calculation remains \(7.5-10.0=-2.5\) points. The contrast still refers to \(A\) versus \(C\) on the same 12-week score in the same target population; discontinuing \(B\) does not turn it into another comparison.
The number is illustrative, not evidence from a real study. Whether the continuing trial can estimate \(\psi_A\) depends on its design and data; the point here is that the contrast’s definition does not depend on the presence of \(B\).
Worked Example: A Discontinued Comparator Remains Part of the Target
Consider a separate hypothetical trial with control \(C\), treatment \(X\), and treatment \(Y\), all assessed on the same 12-week symptom score in the same target population. Before enrollment, the primary contrast is fixed as \(\psi_{XC}=\mu_X-\mu_C\). Suppose illustrative target-population means are \(\mu_C=12.0\), \(\mu_X=9.0\), and \(\mu_Y=7.5\) points.
The prespecified contrast is:
Now suppose the trial discontinues control \(C\), while \(X\) and \(Y\) remain available. The original contrast still means the mean outcome under \(X\) minus the mean outcome under \(C\). Its illustrative value remains \(-3.0\) points; discontinuing \(C\) does not erase the \(C\) term from the definition. But future randomized participants can no longer provide new outcomes under \(C\), so future data alone do not directly supply a concurrent randomized comparison for this contrast.
Replacing control with \(Y\) would answer a different question:
Because higher scores are worse, this illustrative contrast favors \(Y\) over \(X\) by 1.5 points. It is not a revised calculation of \(\psi_{XC}\); it is a different treatment comparison. The illustrative means demonstrate the arithmetic and interpretation, not what an actual trial would observe or establish.
Worked Example: A Fixed Weighted Contrast After Dropping One Arm
Suppose a hypothetical trial prespecifies a contrast comparing a new treatment \(N\) with a fixed weighted average of two reference strategies, \(R_1\) and \(R_2\). The weights are 0.6 and 0.4, respectively. On a scale where lower is better, illustrative means are \(\mu_N=6.0\), \(\mu_{R_1}=8.0\), and \(\mu_{R_2}=10.0\) points.
The coefficients sum to zero: \(1-0.6-0.4=0\). The contrast and its illustrative value are:
If \(R_2\) is discontinued, the fixed contrast remains the same target, including the 0.4 coefficient on \(R_2\). Removing that term and using \(N-0.6\mu_{R_1}\) would not be a valid preserved contrast: the coefficients would no longer sum to zero, and the expression would not represent the prespecified comparison. Renormalizing the surviving reference weight to 1 would instead define \(N-R_1\), a different contrast. In either case, the original target’s meaning is retained, but ongoing randomization no longer adds new observations under \(R_2\).
Why the Invariance Holds
The algebra makes the preservation rule precise. Let \(D\) be the set of treatment conditions discontinued for future assignment, and let \(S\) be the remaining set. The original contrast is \(\psi=\sum_j c_j\mu_j\). If every discontinued condition has coefficient zero, then:
Thus, removing those zero-coefficient terms leaves both the value and meaning of the contrast unchanged. If some discontinued condition \(k\) has \(c_k\ne0\), its term \(c_k\mu_k\) remains part of the specified contrast. Deleting it changes the expression by \(-c_k\mu_k\); changing other coefficients to restore a desired scale changes the contrast as well. The algebra does not guarantee that the contrast can be estimated from available data. It only shows whether the target itself has changed.
Conditions for a Stable Interpretation
A contrast can be preserved only if its ingredients retain their intended meanings. The assumptions and protocol details should be stated before arm discontinuation, not reconstructed afterward. In particular:
- Treatment conditions are well defined. Each arm denotes a specified intervention strategy. If the version of a treatment changes across stages, the treatment condition may no longer be the same one in the contrast.
- The target population and outcome summary are fixed. The population, outcome variable, summary measure, and assessment time used to define each \(\mu_j\) must retain their stated meanings.
- The intercurrent-event strategy is unchanged. The estimand’s strategy for events such as treatment discontinuation or rescue therapy must remain as specified. Changing that strategy changes what the treatment means in the contrast.
- Data support is distinguished from definition. For causal interpretation from randomized data, assignment and measurement must support the treatment comparisons being made. After an arm is discontinued, there is no ongoing randomized assignment to it; preservation of its estimand does not restore that support.
These conditions align with the earlier material on treatment conditions, outcome summaries, intercurrent events, and fixed-population estimands. They also make clear why a contrast can keep its definition while the trial’s ability to estimate it changes.
Auditing Changes and Diagnosing Drift
A practical audit begins with the original contrast written in terms of named treatment conditions and fixed coefficients. After an arm is discontinued, compare the revised protocol with that definition. Ask whether the same treatment conditions, population, outcome summary, event strategy, and coefficients still appear. If a component has been removed or reweighted, describe the resulting target as a new contrast rather than presenting it as unchanged.
Several warning signs indicate that the interpretation may be drifting:
- The control group is replaced by a surviving active arm without explicitly naming a new contrast.
- A dropped arm’s coefficient is silently set to zero, or surviving weights are renormalized.
- A treatment’s formulation or delivery changes, so the arm label no longer denotes the original treatment condition.
- Eligibility criteria or the outcome assessment schedule change, but results are still described as if the original population or summary were used.
- A contrast involving a discontinued arm is described as newly estimated from continuing participants, despite no assignment to that arm.
The main alternative is to define a new, stage-specific question using only the arms that remain. That can be scientifically useful, but it must be distinguished from the original contrast. A protocol may report both targets if each is clearly specified. The limitation is that the new target does not answer the original question, while the original target may have less information available after discontinuation.
Common Mistake
Wrong reasoning: “The control arm was dropped, so the treatment effect is now the difference between the surviving treatment arms.” This silently changes the comparator and therefore changes the contrast. Correction: Keep the original treatment-versus-control contrast named and defined as specified. If the trial also wants to compare the surviving treatments, define and label that as a separate contrast, and be explicit about which data support each target.
Check Your Understanding
Answer each question using the distinction between a contrast’s definition and the data available to estimate it.
- A trial prespecifies \(\mu_A-\mu_C\) and \(\mu_B-\mu_C\). Arm \(B\) is discontinued. Which contrast is unchanged by removing \(B\) from future assignment?
- A prespecified contrast is \(\mu_X-\mu_C\). Arm \(C\) is discontinued. Does the contrast become \(\mu_X-\mu_Y\) if \(Y\) continues?
- For \(\psi=\mu_N-0.7\mu_{R_1}-0.3\mu_{R_2}\), what happens to the target if \(R_2\) is dropped and its coefficient is simply deleted?
- Name two protocol features, other than arm availability, that must remain fixed to preserve a contrast’s interpretation.
Solutions. (1) \(\mu_A-\mu_C\) is unchanged because \(B\) has coefficient zero in that contrast; \(\mu_B-\mu_C\) remains defined but includes the discontinued condition. (2) No. \(\mu_X-\mu_C\) remains the original target, although future randomization no longer supplies outcomes under \(C\); \(\mu_X-\mu_Y\) is a different contrast. (3) Deleting the term changes the target from \(N\) versus the fixed weighted reference to an expression involving only \(R_1\); renormalizing the remaining weight would define \(N-R_1\), also a different contrast. (4) Examples include the target population, treatment condition versions, outcome and assessment time, and intercurrent-event strategy.