Why the Control Must Be Addressed Explicitly
When an interim decision removes an experimental arm, the decision changes more than which treatments are offered to new participants. It also affects the comparison group: whether the control continues enrolling, which later participants contribute to each comparison, and how results are summarized across stages. These choices should be specified in advance. Otherwise, a trial can end up with an ambiguous control group or comparisons that mix different participant groups without a clear interpretation.
In a multi-arm multi-stage (MAMS) trial, several experimental arms may share one control arm. If at least one experimental arm continues, the control will generally need to continue too: participants assigned to it provide a concurrent comparison for the retained arm or arms. Dropping an experimental arm does not, by itself, justify dropping the control. Conversely, if no experimental arm remains, continuing to recruit control participants alone ordinarily creates no randomized treatment comparison for the trial’s primary objective.
This policy is about trial conduct, not a new arm-selection rule. The earlier material on selection and stopping establishes how a set of experimental arms may be retained or eliminated. Here the focus is what happens to control enrollment, participant follow-up, and subsequent comparisons once that set is known.
Specify What “Retain the Control” Means
A protocol should distinguish three decisions. First, does the control remain open to new participants while any experimental arm continues? Second, how are new participants allocated among the control and the retained arms? Third, what follow-up and outcome collection continue for participants assigned before an arm was dropped? These decisions are related, but none determines the others automatically.
A simple policy is to keep the control open for every stage in which at least one experimental arm is retained, and to continue outcome collection for all randomized participants according to the planned follow-up schedule. At each stage, the allocation ratio can be defined for the active arms only. For example, after one experimental arm is dropped, a protocol may allocate new participants equally between the remaining experimental arm and control. If two experimental arms remain, it might instead use equal allocation across those two arms and control, or a different prespecified ratio.
There is no universal requirement that the control-to-experimental allocation ratio stay constant across stages. A change may be operationally reasonable when the number of active arms changes. It should, however, be documented, implemented through the randomization system, and accounted for in the analysis. The control should be enrolled concurrently with the retained arms, using the same eligibility criteria and outcome-assessment schedule, so that each comparison is anchored to participants treated during a comparable period.
Control participants randomized before an arm is dropped remain in the trial and should ordinarily be followed as planned. Their data do not become invalid because a treatment arm closes. Similarly, participants already assigned to a dropped arm should not be silently removed from data collection: their outcomes may be needed for safety, reporting, or other prespecified analyses, even though that arm no longer receives new participants.
Worked Example: Keeping a Shared Control After One Arm Is Dropped
Suppose a hypothetical trial begins with control \(C\) and experimental arms \(A\) and \(B\). In the first stage, 40 participants are randomized to each arm. At the interim decision, \(A\) is retained and \(B\) is dropped. The protocol calls for 60 additional participants to be randomized equally between the active arms, now \(C\) and \(A\).
The second-stage allocation is 30 to \(C\) and 30 to \(A\). Across the trial, the planned randomized totals are therefore \(70\) in \(C\), \(70\) in \(A\), and \(40\) in \(B\). The control is not closed when \(B\) is dropped: it continues concurrently with \(A\), and the new \(C\) participants can contribute to the later comparison with \(A\).
The change from three active arms in stage 1 to two in stage 2 is explicit in the allocation plan. All 40 participants already assigned to \(B\) remain subject to the protocol’s follow-up schedule. No new participant is assigned to \(B\). This separates the operational decision to close \(B\) from the obligation to complete follow-up for its existing participants.
If the outcome is a continuous measure for which higher values are better, suppose the stage-2 means are \(12.4\) units for \(A\) and \(10.8\) units for \(C\). The stage-2 raw mean difference is \(\bar{x}_{A,2}-\bar{x}_{C,2}=12.4-10.8=1.6\) units in favor of \(A\). This is a comparison among stage-2 participants; by itself it is not a summary of the entire trial or an adjusted estimate across stages.
Comparisons Across Stages Need Care
A shared control can be used in comparisons with multiple retained experimental arms, but each comparison should be defined clearly. At a given stage, the raw difference between an experimental-arm mean and the control mean is a concurrent stage-specific contrast. To summarize across stages, an analysis may combine stage information or adjust for stage. The choice and its assumptions must be stated; merely pooling all observations in each arm does not guarantee a stage-adjusted comparison.
To see why, let \(n_{E,k}\) and \(n_{C,k}\) be the sample sizes in experimental arm \(E\) and control at stage \(k\), and let \(\bar{x}_{E,k}\) and \(\bar{x}_{C,k}\) be their stage-specific sample means. Define each stage-specific difference as \(d_k=\bar{x}_{E,k}-\bar{x}_{C,k}\). The pooled difference between the two arm means can be written as
The first term is a weighted summary of the stage-specific differences, using the experimental arm’s stage weights. The second term is a composition term. It vanishes if the experimental and control arms have the same proportions of participants at every stage; it also vanishes if the control mean is constant across stages. Otherwise, the pooled contrast can reflect both treatment differences and unequal stage composition. In particular, even constant treatment effects at all stages do not by themselves make the raw pooled difference a weighted average of those effects.
This decomposition is an algebraic identity: write each experimental mean as \(\bar{x}_{E,k}=\bar{x}_{C,k}+d_k\), weight those means by \(n_{E,k}/n_E\), and subtract the control mean \(\sum_k(n_{C,k}/n_C)\bar{x}_{C,k}\). The terms involving \(\bar{x}_{C,k}\) separate into the composition term shown above. Thus, the problem is not that pooling is always wrong; it is that a pooled marginal contrast has a particular interpretation that depends on allocation across stages and stage-specific control outcomes.
Worked Example: Different Stage Weights Can Reverse a Pooled Contrast
Consider hypothetical data from two stages. The outcome is measured in points, with higher values better. At both stages, the experimental arm’s mean is 2 points higher than the control mean. However, stage-specific means and sample sizes differ as follows.
| Stage | Control mean | Experimental mean | Control sample size | Experimental sample size |
|---|---|---|---|---|
| 1 | 10 | 12 | 20 | 80 |
| 2 | 20 | 22 | 80 | 20 |
The pooled experimental mean is \(\bar{x}_{E}=(80\times12+20\times22)/100=(960+440)/100=14\) points. The pooled control mean is \(\bar{x}_{C}=(20\times10+80\times20)/100=(200+1600)/100=18\) points. Therefore, the raw pooled difference is \(14-18=-4\) points, despite the experimental mean exceeding the control mean by 2 points in each stage.
The composition term explains the result. Experimental-arm stage weights are \(0.8\) and \(0.2\), while control-arm stage weights are \(0.2\) and \(0.8\). The weighted stage-specific treatment contrast is \(0.8(2)+0.2(2)=2\) points. The composition term is \((0.8-0.2)(10)+(0.2-0.8)(20)=6-12=-6\) points. Adding the terms gives \(2+(-6)=-4\), exactly the pooled difference.
A stage-adjusted summary using equal stage weights is \(0.5(12-10)+0.5(22-20)=2\) points. That is a different estimand from the raw marginal contrast: it averages the two stage-specific contrasts equally. Other stage weights could be chosen if justified by the target population or design. The important point is to define the intended summary and not describe the unadjusted pooled difference as a treatment-effect average when the stage weights differ.
When No Experimental Arm Remains
If all experimental arms are dropped, there is ordinarily no ongoing randomized comparison that requires new control participants. A common prespecified policy is therefore to stop new enrollment to all arms and complete follow-up for participants already randomized. This is distinct from stopping all trial activity immediately: safety monitoring, outcome collection, data cleaning, and reporting may continue as specified.
A protocol might instead continue some data collection for a separate, explicitly defined purpose. For example, it might collect long-term outcomes from already-enrolled participants or maintain a registry. That does not make control-only recruitment a continuation of the original randomized comparison. Any such activity needs its own rationale, consent and operational plan, and should not be represented as evidence from a concurrent comparison when none exists.
Worked Example: Closing Enrollment After All Experimental Arms Are Dropped
Suppose a hypothetical trial has control \(C\) and experimental arms \(A\) and \(B\). At an interim decision, neither experimental arm meets the prespecified continuation criterion. The protocol says that when no experimental arm remains, randomization and enrollment stop.
Assume 75 participants have already been randomized: 25 to \(C\), 26 to \(A\), and 24 to \(B\). The trial stops recruiting new participants to all three arms. The 75 existing participants continue the planned outcome follow-up, and the investigators complete the specified safety and data-reporting procedures. No additional control participants are recruited, because they would not be randomized against a continuing experimental arm under this design.
The operational conclusion is not that the treatment effects are exactly zero, nor that the control participants should be discarded. It is that the prespecified selection rule leaves no experimental comparison to continue. The interpretation of the observed interim results and any final analyses remains governed by the trial’s analysis plan.
Checks, Alternatives, and Limitations
Before implementation, check that the control-retention rule can be followed in the randomization system and that the analysis plan recognizes stage and allocation changes. Useful checks include comparing planned and realized allocation by arm and stage, verifying that control enrollment overlaps in calendar time with each retained arm, and checking that follow-up completion is not inadvertently curtailed for participants in dropped arms. An unexpected gap in control recruitment or a change in eligibility criteria may undermine the intended concurrent comparison and should be documented.
The main alternative to a continuing shared control is a separate control group for each experimental arm or stage. Separate controls can make comparisons operationally distinct, but may require more participants and do not automatically solve changes in patient mix over time. A shared control can be efficient because the same control participants inform more than one comparison, but those comparisons are statistically dependent through their shared control data. The analysis must account for that dependence when joint conclusions are made.
A second alternative is to stop recruiting to the control at the same time as a particular experimental arm. This can be coherent if no other experimental arm remains, or if the protocol explicitly uses another valid comparison source. It is generally not coherent to close a shared control while continuing an experimental arm that relies on that control, unless the design has prespecified a different comparison strategy. Historical or nonconcurrent control data may differ in calendar time, eligibility, care, or outcome measurement; they should not be treated as interchangeable with concurrent randomized controls without a justified design and analysis.
These operational rules do not themselves establish that a chosen selection rule has desirable error rates, adequate power, or unbiased final inference. Those properties depend on the full design and analysis. Nor does a stage-adjusted analysis automatically answer every scientific question: the target weighting across stages should match the intended interpretation. If treatment effects or participant populations differ by stage, a single summary may conceal that variation. Report stage-specific information where relevant and avoid claiming that a single pooled contrast applies uniformly.
Common Pitfalls
- Closing the control whenever one experimental arm is dropped. If another experimental arm continues, this removes its planned concurrent comparison. Keep the shared control open for retained arms unless a different comparison strategy was specified.
- Stopping outcome collection in a dropped arm. Dropping stops new assignments; it does not automatically cancel follow-up for participants already randomized.
- Assuming pooled means are stage-adjusted. Unequal stage proportions can make the raw pooled contrast depend on stage-specific control means. Use a clearly defined stage-adjusted summary or describe the pooled contrast as marginal and composition-dependent.
- Continuing control-only recruitment without a comparison objective. If no experimental arm remains, specify why further control enrollment is needed; otherwise, stop enrollment and complete planned follow-up.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- After arm \(B\) is dropped, arm \(A\) continues. What should happen to a shared control under a concurrent-control policy?
- Why should participants assigned to a dropped arm usually remain in planned follow-up?
- In one stage, the control mean is 5 and the experimental mean is 7; in another, they are 9 and 11. What is the equally weighted stage-adjusted difference?
- Can the raw pooled difference be interpreted as a weighted average of stage-specific treatment differences when the arms have different stage proportions? State a condition under which the composition term vanishes.
- When all experimental arms are dropped and the protocol stops enrollment, what happens to participants already randomized?
Solutions. (1) Keep the control open and enroll it concurrently with \(A\), using the allocation plan specified for the remaining arms. (2) Stopping new assignment does not erase the participants’ randomized status or the need for planned outcomes and safety information. (3) The stage-specific differences are \(7-5=2\) and \(11-9=2\), so the equally weighted summary is \(0.5(2)+0.5(2)=2\) outcome units. (4) Not in general: unequal stage weights can add a composition term involving stage-specific control means. That term vanishes when both arms have the same stage proportions, or when the control mean is constant across stages. (5) No new participants are enrolled under that rule, but existing participants continue the protocol-specified follow-up and reporting procedures.