Smoothing Allocation Updates to Limit Abrupt Time Shifts

Why Smooth an Allocation Update?

In response-adaptive randomization, incoming outcome information can change the allocation probability used for later participants. A direct update may move that probability sharply when an interim estimate changes, especially when relatively little response information is available. If enrollment conditions also change over calendar time, a sharp allocation shift can leave the treatment arms with quite different mixes of early and late participants. Smoothing offers a way to moderate the size of each update.

Smoothing is an update rule for the allocation probability; it is not the same as smoothing an outcome estimate. It does not make the arms contemporaneous, remove calendar-time drift, or guarantee balance in any realized trial. Its purpose is narrower: when the response-based target changes, it can make the probability used for randomization move more gradually.

Definition: At enrollment step $t$, let $q_t$ be the raw response-based target probability of assigning the next participant to arm A, calculated using information available before that assignment. Let $\pi_t$ be the probability actually used to assign that participant to A. Exponential smoothing with parameter $\lambda$, where $0<\lambda\leq 1$, updates the probability by $$ \pi_t=(1-\lambda)\pi_{t-1}+\lambda q_t. $$ The initial probability $\pi_1$ must be specified by the design. When $\lambda=1$, the smoothed rule uses the raw target immediately; smaller $\lambda$ gives the previous allocation probability more influence.

For this rule to describe a valid sequential randomization procedure, each $q_t$ must be calculated from information available before participant $t$ is randomized, and the actual assignment must be randomized with probability $\pi_t$ given the information then available. The update must produce a probability in $[0,1]$. If the design also imposes allocation floors or ceilings, the probability actually used must respect them; a constrained or clipped update may no longer follow the unconstrained recursion exactly. Smoothing does not replace those safeguards.

What the Smoothing Rule Does

The update is a convex combination: for $0<\lambda\leq1$, $\pi_t$ lies between the previous probability $\pi_{t-1}$ and the current raw target $q_t$. The amount of movement is exactly

\pi_t-\pi_{t-1}=\lambda(q_t-\pi_{t-1}).

Thus, if the target is far from the current probability, a smaller $\lambda$ limits the step taken toward it. This is a statement about the planned probability sequence, not a guarantee about the number of participants actually assigned to either arm.

A Short Derivation: Earlier Targets Still Matter

For a fixed $\lambda$ and no clipping, repeated substitution shows that the current probability incorporates earlier raw targets with geometrically decreasing weights. Starting with $\pi_1$ and applying the rule for $t\geq2$,

\begin{aligned} \pi_t &=(1-\lambda)\pi_{t-1}+\lambda q_t\\ &=(1-\lambda)^{t-1}\pi_1+ \lambda\sum_{k=2}^{t}(1-\lambda)^{t-k}q_k. \end{aligned}

The expression follows by induction. It holds at $t=2$ by the update rule. If it holds at $t-1$, multiplying that expression by $1-\lambda$ and adding $\lambda q_t$ gives the displayed expression at $t$. The coefficients are nonnegative and sum to one, so the current probability is a weighted average of the initial probability and earlier targets. When a target remains constant at $q$, the difference from it contracts:

\pi_t-q=(1-\lambda)^{t-1}(\pi_1-q).

So a persistent target change is not ignored: without constraints, the probability approaches the new target. But with smaller $\lambda$, it approaches more slowly. If raw targets fluctuate in response to noisy interim data, smoothing can also damp the size of the resulting probability fluctuations.

Worked Example: A Single Large Target Change

Consider a hypothetical trial in which the initial probability of assigning to A is $\pi_1=0.50$. After the first period, the response-based procedure supplies a raw target of $q_t=0.80$ for each of the next three periods. Use $\lambda=0.40$ and calculate the probabilities used in those periods.

For period 2, $$ \pi_2=(1-0.40)(0.50)+0.40(0.80)=0.30+0.32=0.62. $$

For period 3, $$ \pi_3=0.60(0.62)+0.40(0.80)=0.372+0.32=0.692. $$

For period 4, $$ \pi_4=0.60(0.692)+0.40(0.80)=0.4152+0.32=0.7352. $$

The direct, unsmoothed rule would use 0.80 in each of periods 2–4. The smoothed rule instead moves from 0.50 toward 0.80 in stages. It does not stay at 0.50: as long as the raw target remains 0.80, the used probability continues to rise toward it.

Calendar-Time Composition: Expected Allocation Masses

A probability sequence can be compared across enrollment periods to see how much calendar-time separation it induces in expectation. The calculation must be described carefully. Dividing an arm’s expected time-weighted allocation total by its expected allocation count gives the mean enrollment period of that arm’s expected allocation mass. It is generally not the expectation of the realized arm-specific sample mean, because the realized count is random and is the denominator of that sample mean.

The next hypothetical calculation assumes 20 participants enroll in each of four periods. For illustration, treat the probabilities in each period as fixed at the values shown. In an actual adaptive trial, probabilities can depend on the observed response history, so their future values may themselves be random. The expected counts below are allocation masses under this specified sequence, not guaranteed realized counts.

Worked Example: Comparing Smoothed and Direct Updates Across Periods

Use the probabilities from the previous example: for A, the smoothed sequence is $0.50, 0.62, 0.692, 0.7352$, while the direct sequence is $0.50, 0.80, 0.80, 0.80$. The probability of assigning to B is one minus the probability of assigning to A. Each of the four periods contains 20 hypothetical participants.

For any arm, the expected allocation mass in a period is the number enrolling in that period multiplied by that arm’s assignment probability. Index the periods as 1 through 4. Under the smoothed sequence, A’s total expected allocation mass is $$ 20(0.50+0.62+0.692+0.7352)=20(2.5472)=50.944. $$

Its time-weighted expected allocation total is $$ 20\{1(0.50)+2(0.62)+3(0.692)+4(0.7352)\} =20(6.7568)=135.136. $$

Therefore, the mean enrollment period of A’s expected allocation mass is $135.136/50.944=2.6526$. Across all arms, expected allocation mass is $4(20)=80$, and the time-weighted total is $20(1+2+3+4)=200$. B’s expected count is $80-50.944=29.056$, and its time-weighted total is $200-135.136=64.864$. Its mean enrollment period of expected allocation mass is $64.864/29.056=2.2324$. The difference between these two allocation-mass means is about $2.6526-2.2324=0.4203$ periods.

Under the direct sequence, A’s expected count is $20(0.50+0.80+0.80+0.80)=58$, and its time-weighted total is $20\{1(0.50)+2(0.80)+3(0.80)+4(0.80)\}=154$. The corresponding mean is $154/58=2.6552$. B’s expected count is $80-58=22$, and its time-weighted total is $200-154=46$, giving $46/22=2.0909$. The difference is about $2.6552-2.0909=0.5643$ periods.

In this example, smoothing reduces the separation between the two allocation-mass means, from about 0.5643 to 0.4203 periods. It does not eliminate the separation: the probability of assignment to A still increases over time, so the arms’ expected enrollment distributions still differ. These summaries describe expected allocation masses; they should not be reported as expectations of realized arm-specific mean enrollment periods.

Fluctuating Targets and Update Size

Smoothing can also moderate short-term oscillations. The next example isolates the arithmetic of the update; it does not claim that a particular sequence of raw targets would arise in any real trial. In practice, those targets would be generated from response information available at each update.

Worked Example: Damping Alternating Targets

Suppose the initial probability is $\pi_1=0.50$, the successive raw targets are $0.70, 0.30, 0.70$, and $\lambda=0.25$. The first updated probability is $$ \pi_2=0.75(0.50)+0.25(0.70)=0.55. $$

The second update gives $$ \pi_3=0.75(0.55)+0.25(0.30)=0.4125+0.075=0.4875, $$ and the third gives $$ \pi_4=0.75(0.4875)+0.25(0.70)=0.365625+0.175=0.540625. $$

The raw targets range from 0.30 to 0.70, a range of 0.40. The updated probabilities here range from 0.4875 to 0.55, a range of 0.0625. This illustrates how a small update parameter can damp rapid changes. It does not show that smoothing always reduces every measure of imbalance, or that it will be preferable when a genuine, sustained target change should be followed promptly.

Diagnostics, Alternatives, and Limitations

The allocation-probability timeline is a direct diagnostic: compare the raw target $q_t$ with the probability actually used $\pi_t$. Large, sudden changes in $q_t$ paired with smaller changes in $\pi_t$ are consistent with smoothing. Also examine arm-by-time enrollment tables, as established in the earlier material on allocation probability timelines and arm-by-time enrollment. These show whether the intended gradual updates have translated into less abrupt temporal separation in the observed trial. Random allocation means the realized counts can differ from their expected allocation masses.

Several failure modes deserve attention. If the response target is noisy, smoothing may damp erratic movement, but a very small $\lambda$ can make the allocation probability slow to respond even when evidence for a changed target persists. If prognosis changes rapidly over calendar time, a gradual allocation update may still produce substantial time imbalance. If a target jumps because a meaningful design condition has changed, automatically smoothing through the jump may be inappropriate; the reason for the target change must be considered, not just its size. Finally, smoothing can make probability trajectories look less abrupt without ensuring sufficient treatment overlap in every period.

The main alternatives answer different design needs. A direct update uses the latest raw target immediately, which responds quickly but can create abrupt probability changes. A fixed allocation probability does not react to response data and can avoid response-driven shifts, but it forgoes response adaptation. Less frequent, prespecified update times can limit how often probabilities change, though each update can still be large. These approaches are not interchangeable: the choice depends on how quickly the design should react to response information and how important gradual changes are for the intended enrollment pattern.

Common pitfall: “The probabilities change gradually, so calendar-time imbalance has been removed.” This is incorrect. Smoothing limits the step between probabilities; it does not force the arms to have the same enrollment-period distribution. A probability that gradually moves upward can still allocate more participants to A later in enrollment. Check the actual probability timeline and period-specific enrollment counts, and treat smoothing as one design feature rather than a guarantee of temporal balance.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

If $\pi_{t-1}=0.60$, $q_t=0.40$, and $\lambda=0.25$, what probability is used at step $t$?
For fixed $\lambda=0.30$, how much can one update move from $\pi_{t-1}$ if $q_t$ and $\pi_{t-1}$ are both probabilities?
Why is an expected time-weighted allocation total divided by an expected count not generally the expectation of a realized arm-specific mean enrollment period?
Does smoothing guarantee that both arms have participants in every enrollment period? Explain.

Solutions. (1) $\pi_t=0.75(0.60)+0.25(0.40)=0.45+0.10=0.55$. (2) Since $|q_t-\pi_{t-1}|\leq1$, the update equation gives $|\pi_t-\pi_{t-1}|=0.30|q_t-\pi_{t-1}|\leq0.30$. (3) The realized mean has a random denominator, the realized arm count; in general, a ratio of expected totals is not the expectation of a ratio. The expected allocation-mass ratio should be named and interpreted as such. (4) No. Smoothing changes assignment probabilities, not realized assignments, and it does not guarantee any specific arm-by-period count. Random variation, as well as the probability sequence itself, can leave a period with few or no participants in one arm.