Choosing Allocation Update Timing

Why Update Timing Is a Design Choice

An adaptive randomization rule determines how trial information translates into allocation probabilities. A separate design decision determines when those probabilities are recalculated and put into use. The same probability-matching or expected-utility rule can therefore produce different assignment sequences if one trial recalculates after every patient and another recalculates only at scheduled reviews.

Timing is not merely administrative. It determines how rapidly new information can influence enrollment, how often staff must implement a change, and which participants are randomized under a given probability vector. It can also affect how transparent and reproducible trial conduct is. The design should state the allocation rule and the update schedule as separate, prespecified components.

For a participant randomized while update \(k\) is in force, a valid randomization can be implemented by drawing a uniform random variable \(U\) on \([0,1)\) and assigning the participant according to the cumulative probabilities. For two arms, for example, assign to arm 1 if \(U<\pi_{1,k}\), and to arm 2 otherwise. This procedure gives each participant the stated allocation probability; it does not guarantee that observed arm counts will equal their expected values.

This tutorial compares three common schedules. The update mechanism is distinct from the method used to calculate the probabilities, such as posterior probability matching or allocation by posterior expected utility. The design may also impose allocation floors, caps, or a minimum control allocation; the chosen schedule must implement those constraints as well.

Three Ways to Schedule Updates

With patient-by-patient updating, each new allocation probability vector is calculated before the next participant is assigned, using the information permitted by the protocol at that time. This can make the design responsive, but requires a reliable data pipeline and a clear process for authorizing and activating each change. A protocol must specify what happens if a calculation or approval is not ready when the next participant is eligible.

With a fixed-cohort schedule, a vector is calculated at the start of a cohort and remains in force for a specified number of assignments. The trial then reviews the prespecified information and calculates the next vector. Cohort size is fixed by design, not by how quickly participants enroll. This makes the number of assignments under each vector easier to state, while allowing probabilities to remain unchanged as information accumulates within the cohort.

With a calendar-time schedule, recalculation takes place at fixed dates or intervals—for example, on the first business day of each month—rather than after a fixed number of assignments. The number of participants assigned under each vector therefore depends on the pace and pattern of enrollment. If no eligible participant is randomized during an interval, the scheduled review can still occur if the protocol requires it, but no assignments will have used the new vector yet.

Worked Examples

What the Schedules Imply

A short justification helps clarify why expected allocation counts are useful but are not assignment guarantees. If a fixed cohort contains \(m\) assignments and arm \(a\) has constant probability \(\pi_a\) throughout, define \(I_j=1\) when participant \(j\) is assigned to arm \(a\), and \(I_j=0\) otherwise. The total count is \(N_a=\sum_{j=1}^{m}I_j\). Since \(E(I_j)=\pi_a\), linearity of expectation gives

If the randomization draws are independent and the probability remains fixed, then \(N_a\) is binomial and \(\operatorname{Var}(N_a)=m\pi_a(1-\pi_a)\). The variance describes ordinary random variation in counts; it is not a feature that can be eliminated by choosing a convenient update date.

Under patient-by-patient updating, let \(\pi_{a,j}\) be the probability for assignment \(j\), determined from information available before that assignment. Conditional on that information, \(E(I_j\mid\text{information before assignment }j)=\pi_{a,j}\). Taking expectations gives

The probabilities can change as information changes, so a single binomial calculation with one constant probability generally does not describe the total count over an adaptively updated sequence. Nor should a design report the expected count as if it were a promised allocation ratio.

For calendar-time updates, the same fixed-probability calculation can be applied within each interval when the number of participants and the vector in force are specified. But the interval sizes may vary, and in planning they may themselves be uncertain. This makes the expected overall allocation a function of both the schedule and accrual. A schedule comparison should therefore consider plausible enrollment patterns, not just the intended probability vectors.

Operational Consequences and Diagnostics

Patient-by-patient updates can align allocation closely with the most recently processed information, but frequent changes increase operational demands. The allocation system must apply a single authorized vector to each participant, record the version used, and prevent a change from taking effect inconsistently across sites. If decision review takes time, the protocol must define a practicable cut-off and contingency rather than leave staff to decide informally.

Fixed cohorts are easier to communicate and audit by assignment count. Their main trade-off is that all participants in a cohort use the same probabilities, even if the information supporting a future update accumulates before the cohort is complete. Cohorts should be chosen with trial workflow and the intended update frequency in mind; “cohort” should mean a defined number of randomizations, not an informal period of enrollment.

Calendar-time updates can fit recurring data-review meetings and standard operational cycles. However, a date-based schedule can create uneven numbers of assignments under each vector when accrual is uneven. It can also produce a review with little new information or a busy interval with many participants assigned under a vector calculated earlier. These are consequences of the schedule, not evidence that the randomization has failed.

Useful monitoring checks include confirming that every assignment has a timestamp, the probability-vector version in force, and a reproducible randomization record; checking that scheduled updates occurred only under the protocol’s stated conditions; and comparing actual counts with the variation expected under the probabilities used. A discrepancy from expected counts alone is not proof of a malfunction. Investigate whether the implemented vector, random-number procedure, eligibility sequence, or recordkeeping departed from the design.

The design also needs to distinguish a scheduled update from an unscheduled operational correction. For example, a system failure should not silently change allocation probabilities or cause staff to choose an arm. The protocol should define who can authorize a correction, how the event is documented, and whether randomization pauses until the intended procedure is restored.

Choosing and Specifying a Schedule

There is no universally best update frequency. Patient-by-patient updating emphasizes responsiveness; fixed cohorts emphasize count-based review points; calendar-time updating emphasizes a predictable timetable. The appropriate choice depends on the expected accrual rate, the feasibility of timely data processing and review, and how frequently the design is intended to react to new information. Compare schedules using the same allocation rule and plausible enrollment scenarios so the comparison isolates timing rather than changing several design features at once.

A protocol should state the update trigger, the information cut-off, the method for calculating and approving a new vector, the time at which it becomes active, and the process for recording assignments made under each version. For a cohort or calendar schedule, it should also state what happens if the planned review cannot be completed on time. These details make the randomization rule auditable and prevent ambiguous transitions between probability vectors.

In this comparison, assume the trial has a prespecified allocation rule and valid randomization procedure, the review schedule is followed, and every active probability vector is applied consistently. The examples assume probabilities are fixed within a cohort or calendar interval where stated. They do not address how pending or delayed outcomes enter the information used at an update; that requires separate design provisions.

Common Pitfalls

• Treating a schedule as an allocation rule. Saying “we update monthly” does not specify how the monthly probabilities are calculated. State both the schedule and the rule.
• Expecting exact balance. A 0.60 probability for A does not guarantee that 60% of a small cohort will receive A. It specifies the randomization probability for each assignment while that vector is active.
• Confusing fixed cohorts with calendar periods. A cohort of four always contains four assignments; a month can contain any number, including zero.
• Changing probabilities without a defined activation point. Specify which vector applies to the next participant and how the change is recorded.
• Comparing schedules under different rules. To assess timing alone, keep the allocation calculation and constraints consistent across the schedules being compared.