Why a Shared Control Needs an Explicit Allocation Rule
In a multiarm trial, each experimental treatment may be compared with the same control group. The control is “shared” because participants assigned to control contribute information to more than one treatment-control comparison. The randomization rule must therefore specify the probability of assigning the next eligible participant to control and how the remaining probability is divided among the active experimental arms.
There is no single allocation rule implied by the phrase “shared control.” A design can keep the control probability fixed as the set of active treatments changes, or it can keep a specified allocation ratio between each experimental arm and control. Those choices generally produce different probabilities. The rule should be stated explicitly rather than inferred from a desired control-to-treatment ratio.
The probabilities may depend on the randomization history, interim information, or a prespecified stratum. For a fixed-probability rule, the central feature is that the control probability is specified and held at its target value while the active experimental probabilities are recalculated. The probabilities describe the chance of the next assignment, not guaranteed enrollment counts.
Fixed Control Probability With Weighted Experimental Arms
Suppose the target probability of assignment to control is \(c\), where \(0<c<1\). Give each active experimental arm \(j\) a positive allocation weight \(w_j\). These weights can represent equal allocation or a prespecified relative preference among active arms. To preserve the control probability, distribute the remaining probability \(1-c\) among the active arms in proportion to their weights.
The weights need not sum to one: only their relative sizes matter. Multiplying every active weight by the same positive constant leaves each assignment probability unchanged. The required conditions are that the control target is between zero and one, that the active set is nonempty, and that the active weights are positive with a finite, positive sum. These conditions ensure positive assignment probability for every arm and make the formula well-defined.
This construction is conditional on the active set at the time of assignment. If \(c=0.25\), for example, control remains a 25% chance for the next eligible participant both before and after an experimental arm is removed. The total experimental probability remains 75%; the active experimental arms divide that 75% according to their current weights.
Worked Example: Three Equally Weighted Experimental Arms
A hypothetical platform has active experimental arms A, B, and D, with a shared control C. The design specifies \(p_C=0.25\), and all three experimental arms have equal weights \(w_A=w_B=w_D=1\). Calculate the assignment probabilities.
The experimental probability available for allocation is \(1-0.25=0.75\). The sum of the active experimental weights is \(1+1+1=3\). Thus, for each experimental arm,
The four assignment probabilities are therefore \(p_C=0.25\), \(p_A=0.25\), \(p_B=0.25\), and \(p_D=0.25\). They sum to \(0.25+0.25+0.25+0.25=1.00\), as required.
Now suppose B is removed, so the active set is \(\{A,D\}\). The control target remains 0.25. The remaining weights sum to 2, so
After removal, the probabilities are \(p_C=0.25\), \(p_A=0.375\), and \(p_D=0.375\), which again sum to 1. The removal increases the probability for each remaining experimental arm, not the control probability.
Derivation and a Weighted Example
The fixed-control formula follows from two requirements. First, control receives probability \(c\). Second, each active experimental arm receives a share of the total experimental probability \(1-c\) proportional to its weight. If arm \(j\)'s share is \(q_j\), proportional allocation requires \(q_j/q_k=w_j/w_k\) for any two active arms. Therefore \(q_j=\lambda w_j\) for a common constant \(\lambda\). Since experimental probabilities must sum to \(1-c\),
Substituting this value of \(\lambda\) gives \(q_j=(1-c)w_j/\sum_k w_k\), the stated rule. This derivation also shows why the active weights must be renormalized after an arm leaves: the denominator is the sum over the arms that can still receive assignments.
Worked Example: Unequal Weights After an Arm Is Removed
In a hypothetical trial, the control probability is fixed at 0.30. Three experimental arms have weights \(w_A=2\), \(w_B=1\), and \(w_D=3\). Calculate all probabilities, then recalculate after B is removed.
Initially, the experimental probability is \(1-0.30=0.70\), and the total experimental weight is \(2+1+3=6\). The probabilities are
Together with \(p_C=0.3000\), these sum to \(0.2333+0.1167+0.3500+0.3000=1.0000\), up to rounding. If B is removed, the remaining weights sum to \(2+3=5\), while the control probability remains 0.30. Hence
The new probabilities are \(p_C=0.30\), \(p_A=0.28\), and \(p_D=0.42\), which sum to 1. The relative allocation between A and D is preserved: \(0.28/0.42=2/3=w_A/w_D\). Their probabilities rise because they now divide the same 0.70 experimental probability using the remaining weights only.
Fixed Control Probability Is Not a Fixed Per-Arm Ratio
A common alternative is to specify an experimental-to-control allocation ratio \(r\) for each active experimental arm. Under equal per-arm ratios, each experimental arm has probability \(r p_C\). With \(m\) active experimental arms, the probabilities must satisfy \(p_C+m r p_C=1\), giving
Here, removing an experimental arm changes the control probability because the number \(m\) changes. In contrast, with a fixed-control rule, the control probability stays at its target and the per-arm ratios change as the active set changes.
Worked Example: Comparing Two Allocation Policies
A hypothetical trial has three equally weighted experimental arms and one shared control. Compare a fixed control probability of 0.25 with a fixed 1:1 allocation ratio between each active experimental arm and control. Then compare the probabilities after one experimental arm is removed.
Under the fixed-control policy, \(p_C=0.25\), and the remaining 0.75 is divided equally among three arms. Each experimental arm has probability \(0.75/3=0.25\). After one arm is removed, control remains at 0.25 and each of the two active arms receives \(0.75/2=0.375\).
Under the fixed per-arm 1:1 ratio, \(r=1\). With three experimental arms, \(m=3\), so
At this initial stage the two policies coincide. After one arm is removed, the fixed-ratio policy has \(m=2\), and therefore
Thus, fixed control allocation yields probabilities \(0.25, 0.375, 0.375\) for control and the two active treatments, whereas the fixed-ratio policy yields \(0.3333\) for each. The distinction is a design choice, not a calculation discrepancy.
Operational Checks and Interpretation
Allocation probabilities are probabilities for individual assignments. If a probability is constant for \(N\) assignments, the expected number assigned to an arm is \(N p_j\); the realized count is random and need not equal that expectation. If probabilities change during enrollment, expected counts must be calculated across the periods or assignments with their corresponding probabilities, rather than by multiplying the final probability by the full sample size.
For example, under the fixed-control policy in the first example, if 120 hypothetical participants are randomized while three equal-weight experimental arms are active, the expected control count is \(120(0.25)=30\). If another 120 participants are randomized after B is removed, the expected control count in that second period is also \(120(0.25)=30\). In the second period, the expected counts for A and D are each \(120(0.375)=45\). These are expectations, not promised exact counts.
A shared control may contribute to several comparisons, but it remains one group of participants. Consequently, treatment-control estimates that use the same control observations are statistically related; they should not be treated as independent merely because they are separate comparisons. Allocation probabilities alone do not resolve how comparisons should be analyzed.
The rule also relies on operational assumptions. The listed arms must be the arms that are eligible to receive the next assignment; the randomization system must use the same active set and weights as the design specifies; and each probability must be positive for an arm intended to enroll. If probabilities are stratified, the target and weights must be interpreted within the relevant stratum, with probabilities summing to one there. A fixed control probability in one stratum does not by itself imply the same realized overall proportion if stratum sizes or enrollment rates differ.
Diagnostics, Alternatives, and Limitations
A simple audit can catch many implementation errors. At every update, verify that the control probability equals its specified target, that the active experimental probabilities sum to \(1-c\), and that the full set sums to one. Also verify that the active-arm denominator excludes removed arms and includes every active arm exactly once. Check the actual randomization specification against the intended target: a system that holds each arm-to-control ratio fixed will generally change the control probability as arms are removed.
Common warning signs include probabilities that do not sum to one, a control probability that changes under a supposedly fixed-control rule, a removed arm retaining positive probability, or a remaining arm's relative share no longer matching its stated weight. Rounding should be applied carefully: calculate using full precision, then round for display. If probabilities are rounded before assignment, the implemented probabilities may no longer sum exactly to one.
| Allocation approach | What is held fixed? | Effect when an arm is removed |
|---|---|---|
| Fixed control probability with normalized weights | Control probability \(c\) | Remaining experimental probabilities are recalculated to sum to \(1-c\) |
| Equal allocation across all active arms | Equal probability for each active arm | Control probability rises as the number of active arms falls |
| Fixed per-arm experimental-to-control ratio | Each active arm's ratio to control | Control probability changes with the number of active experimental arms |
The fixed-control approach is useful when the design requires a stable chance of receiving control throughout changes in the experimental arm set. Its limitation is that the ratio of each remaining experimental arm's probability to control grows as arms are removed under equal weighting. Conversely, a fixed per-arm ratio preserves that comparison-level allocation ratio but does not preserve the control probability. Neither policy is automatically best: the choice depends on the design objective and should be stated before implementation.
A frequent mistake is to divide the control probability among the treatment comparisons, as if a participant assigned to control belonged to a separate control arm for each comparison. That is incorrect: control is assigned once, with probability \(c\). Under a fixed-control rule, it is the experimental remainder \(1-c\) that is divided among active arms. Another mistake is to conclude that removing an arm must increase the control probability. That conclusion holds for some policies, such as equal allocation among all active arms or a fixed per-arm ratio, but not for a rule that explicitly holds \(p_C=c\) fixed.
Check Your Understanding
Use the stated allocation rule in each question. The solutions are included so that each probability calculation can be checked.
- A trial has \(p_C=0.20\) and four equally weighted active experimental arms. What is each experimental assignment probability?
- Under a fixed-control rule with \(p_C=0.30\), two active arms have weights 1 and 3. Find their probabilities. What are the probabilities if the weight-1 arm is removed?
- With three active experimental arms and a fixed per-arm ratio \(r=2\), calculate the control probability and each experimental probability.
- Why does removal of an arm not change the control probability under the fixed-control weighted rule?
Solutions. (1) The experimental remainder is \(1-0.20=0.80\), divided equally across four arms: \(0.80/4=0.20\) per arm. (2) The experimental remainder is 0.70 and total weight is \(1+3=4\), so the probabilities are \(0.70(1/4)=0.175\) and \(0.70(3/4)=0.525\). After the weight-1 arm is removed, the weight-3 arm receives the full experimental remainder, 0.70, while control remains 0.30. (3) Here \(m=3\) and \(r=2\), so \(p_C=1/(1+3(2))=1/7=0.1429\); each experimental probability is \(2/7=0.2857\). (4) The rule sets \(p_C=c\) separately from the experimental weights. Removing an arm changes the set of weights used to divide \(1-c\), but does not change \(c\).