Two distinct kinds of adaptation
A doubly adaptive biased coin design (DBCD) adapts in two connected ways. First, observed patient responses inform an updated target allocation: the allocation proportions the design currently aims to achieve. Second, the current allocation counts inform the randomization for the next patient, so that the design can respond to whether an arm is under- or overrepresented relative to that target. The first loop changes the destination; the second guides movement toward it.
These loops answer different questions. The response loop asks, “Given the response information available now, what allocation target does the prespecified target rule recommend?” The allocation loop asks, “Given that target and the patients already assigned, how should the next assignment be randomized?” Keeping the questions separate makes a DBCD easier to describe, audit, and diagnose.
For \(K\) treatment arms, write \(N_k(n)\) for the number assigned to arm \(k\) among the first \(n\) patients, and \(x_k(n)=N_k(n)/n\) for its current allocation proportion when \(n>0\). Let \(q_k(n)\) be the current target proportion for arm \(k\), and let \(\pi_k(n+1)\) be the conditional probability that patient \(n+1\) is assigned to arm \(k\). Each of the vectors \(x(n)\), \(q(n)\), and \(\pi(n+1)\) has nonnegative entries summing to one. The response information used to calculate \(q(n)\) is the information available under the design’s prespecified update schedule.
The sequence of the two feedback loops
A useful way to understand the design is to follow its timeline. After an outcome is available, the response-based procedure updates its estimates or other response summaries. A prespecified target rule uses those summaries to produce the current target allocation. The allocation-count component then combines that target with the assignments already made to determine the randomization for the next patient. The assignment is drawn using that randomization, and the sequence continues as further outcomes become available.
Use only response information that has been observed by the update time. The design must specify how delayed or missing outcomes are handled.
Apply the prespecified target rule to the response information to obtain \(q(n)\), a vector of desired proportions.
Combine the current target with the counts \(N_1(n),\ldots,N_K(n)\) to obtain the next-patient randomization probabilities.
Assign the next patient according to those probabilities, record the assignment, and update again when new responses are available.
The response-driven target rule embodies the allocation objective, such as a prespecified ethical target. The allocation-count mechanism does not itself decide which treatment is better; it uses the current target and assignment history to guide target tracking. Conversely, the response rule does not erase the fact that earlier random assignments have already produced a particular allocation history.
Worked Example: Responses Revise the Target
Consider a hypothetical two-arm trial with binary responses, where a response is coded as 1 and no response as 0. After eight patients, five have been assigned to A and three to B. Suppose the observed outcomes are four responses among the five patients on A and one response among the three on B. Thus the sample response proportions are \(4/5=0.800\) for A and \(1/3=0.333\) for B.
Assume a prespecified target rule, applied to these response summaries, returns \(q_A(8)=0.65\) and \(q_B(8)=0.35\). These target values are stipulated here to illustrate the feedback structure; they are not a claim that this particular response summary must produce those values under every target rule. The current allocation proportions are
The response loop has therefore set a target of 65% for A and 35% for B, while the allocation history currently stands at 62.5% for A and 37.5% for B. The allocation-count loop uses both pieces of information when generating the next randomization. It does not simply replace the current counts with the target, nor does it assign the next patient deterministically to A.
Why the allocation history matters
A target is a proportion to be tracked over a sequence, not a record of the assignments already made. If an arm has received fewer assignments than its current target would suggest, the design’s allocation-count feedback can account for that shortfall when forming the next randomization. If the target changes after new responses arrive, the comparison also changes. Thus assignment history and response information have separate roles, even though both can affect the next assignment.
A short calculation illustrates why a randomization can provide restoring movement toward a target. Temporarily hold the two-arm target \(q_A\) fixed, let \(x_A(n)=N_A(n)/n\), and let \(\pi_A(n+1)\) be the conditional probability of assigning the next patient to A. If \(I_{n+1}\) is 1 when the next patient receives A and 0 otherwise, then \(N_A(n+1)=N_A(n)+I_{n+1}\), with conditional expectation \(E[I_{n+1}\mid\text{history}]=\pi_A(n+1)\). It follows that
Subtracting the fixed target gives
For example, if \(x_A(n)<q_A\) and the design chooses \(\pi_A(n+1)\) between \(x_A(n)\) and \(q_A\), the conditional expected next allocation proportion lies between the current proportion and the target. This calculation explains the purpose of count feedback without specifying a particular DBCD formula. In an adaptive design, the target may itself change after further responses, so this fixed-target calculation describes one step rather than guaranteeing a path to an unchanging destination.
Worked Example: Counts Are Below a Fixed Target
Suppose a hypothetical trial currently has \(n=10\) patients, with \(N_A(10)=3\). For this calculation, assume the current target for A is held fixed at \(q_A=0.60\), and the design’s count-feedback component gives the next patient a 0.70 probability of assignment to A. This probability is supplied as an example of an allocation-stage output, not derived from a particular rule.
The current proportion on A is \(x_A(10)=3/10=0.30\), below the target. The conditional expected number assigned to A after one more patient is
Consequently, the conditional expected allocation proportion after that assignment is
This expected proportion has moved from 0.30 toward 0.60, but it remains below the target. The realized count must be an integer: with probability 0.70 the count becomes four, and with probability 0.30 it remains three. The expectation describes repeated randomizations under the same stated history; it is not a fractional assignment in the actual trial.
Assumptions, diagnostics, and failure modes
The two-loop description is meaningful only when the design specifies how responses enter the target and how allocation history enters randomization. Important conditions include a well-defined response measure, a prespecified target rule, a clear update schedule, and a valid randomization mechanism. If response estimates are based on a statistical model, the model and its assumptions also matter; poor estimation can make the target unstable or systematically misleading. For causal interpretation, assignment should follow the stated randomization procedure, and departures from that procedure need to be recorded and addressed.
Practical diagnostics should examine the two loops separately. For the response loop, check whether response data are available as assumed, whether estimates are based on enough observations to be informative, and whether the target changes sharply after small amounts of new data. For the allocation loop, compare observed allocation proportions with the contemporaneous targets over time, and verify that the recorded assignments match the randomization probabilities the design specified. Review very small arm counts and any probabilities near zero, since they can leave little opportunity to learn about an arm or to correct an allocation imbalance.
- Target instability: Large target changes after a few outcomes can reflect noisy response estimates rather than strong evidence of a real difference.
- Delayed outcomes: Updating as if every patient’s response were observed can be wrong when some outcomes have not yet matured. The update schedule must reflect what is actually available.
- Confusing target and assignment history: A target of 60% for A does not mean that 60% of patients already received A. Report both quantities when describing the design.
- Near-deterministic assignment: If randomization probabilities become too close to zero or one, chance imbalances and limited future information can become difficult to correct.
- Unrecorded rule changes: Changing the target rule or the update schedule during a trial can make the implemented design differ from the prespecified one.
Worked Example: The Target Changes After New Responses
Suppose a hypothetical two-arm trial has assigned six patients to each arm, so \(x_A(12)=x_B(12)=0.50\). At one update, the response procedure produces a target of \(q_A(12)=0.55\) and \(q_B(12)=0.45\). After two additional responses become available, the prespecified target rule produces \(q_A(14)=0.45\) and \(q_B(14)=0.55\). The values are illustrative outputs of a target rule, not a universal mapping from response counts to targets.
At the first update, A is below its 0.55 target, while the current allocation is evenly split. At the second update, the target has changed direction: A’s target is now 0.45. This does not mean that the earlier assignment record was wrong. It means the response loop revised the destination using newly available information, while the allocation loop must work from the assignments actually made. A diagnostic review should ask whether the responses were correctly recorded and whether the target change follows the prespecified rule, rather than treating a target change alone as proof of an error.
How DBCD differs from simpler approaches
With fixed-target randomization, the desired allocation proportions are set in advance and do not change in response to outcomes. Such a design can still use allocation counts to track its fixed target, but it lacks the response-driven target loop. With response adaptation without count feedback, outcomes may change the target or directly influence assignment, but the design does not separately correct for the realized allocation history in the same way. A DBCD combines response-based target updating with count-based guidance toward the current target.
These approaches answer different design needs. A fixed target avoids response-driven changes in the allocation objective, while a DBCD can respond to observed outcomes. But response adaptation can make allocation more dependent on interim data, and count feedback cannot repair a poor target rule or biased response estimates. A DBCD therefore does not automatically guarantee ethical superiority, balanced information, or improved inference. Those properties depend on the target, the randomization rule, the response process, and the analysis plan.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In a DBCD, what information drives the target update, and what information drives allocation-count feedback?
- A two-arm trial has assigned 8 of 15 patients to A, while its current target for A is 0.45. Calculate the current allocation proportion for A and say whether it is above or below the target.
- If the response-based target changes after newly observed outcomes, does that change the assignments already made? Explain.
- Name one diagnostic for each feedback loop and explain what it can reveal.
Solutions. (1) The target update uses observed response information through a prespecified target rule; allocation-count feedback uses the current arm counts together with the current target to determine the next randomization. (2) \(x_A=8/15=0.5333\), or about 53.3%; this is above the target of 45%. (3) No. The target is revised for subsequent decisions, but the assignments already made remain part of the observed allocation history. (4) For the response loop, check whether target values change sharply after small amounts of response data; this can reveal instability. For the allocation loop, compare recorded assignments and allocation proportions with the design’s specified randomization behavior; discrepancies can reveal implementation or recording problems.