Response-Adaptive Randomization as a Sequential Rule

Why Allocation Is a Sequential Rule

In a fixed-ratio randomized trial, the planned probability of assignment to each arm stays the same throughout randomization. In a response-adaptive trial, the probability for the next participant can change as outcomes from earlier participants accumulate. This makes the allocation mechanism a rule applied repeatedly over time, rather than a single ratio chosen once at the start.

The distinction matters for both implementation and analysis. The trial team must know which information is available at each assignment and how that information determines the next probabilities. A probability calculated from earlier outcomes is not the same thing as a guarantee about how many participants will ultimately be assigned to each arm. Even with fixed probabilities, random variation means realized arm counts need not match their planned ratio exactly.

Definition: Let $A_i$ denote the treatment assigned to participant $i$, and let $H_{i-1}$ be the information available immediately before that participant is assigned. A sequential randomization rule specifies, for each arm $k$, an assignment probability $$ q_{ik}(H_{i-1})=P(A_i=k\mid H_{i-1}), $$ where $q_{ik}(H_{i-1})\geq 0$ and $\sum_k q_{ik}(H_{i-1})=1$. The rule is response-adaptive when these probabilities can depend on outcomes observed from participants enrolled earlier in the trial.

The history $H_{i-1}$ may include earlier assignments, outcomes that have become available, and other information specified by the protocol. It must not include a future outcome that has not yet been observed at the time of assignment. An outcome-based update also needs a defined response, such as response at a specified assessment time, and a rule for deciding whether that outcome is sufficiently complete to use.

Response-adaptive randomization does not mean assigning the next participant deterministically to the arm with the best current result. The probabilities can change with the history while a random draw still determines each assignment. The particular function that converts accumulated outcomes into probabilities is a design choice; the sequential definition alone does not prescribe a preferred function.

Fixed-Ratio and Response-Adaptive Allocation

With fixed-ratio randomization, the allocation probabilities are constant: for every assignment $i$, $q_{ik}(H_{i-1})=q_k$, where the planned probabilities $q_k$ sum to one. For example, a 1:1 allocation to two arms has $q_A=q_B=0.5$ at every assignment. A response-adaptive rule instead allows $q_{ik}$ to vary with the outcomes available before assignment $i$.

A fixed ratio describes probabilities, not a promise that the final numbers in the arms will be exactly in that ratio. Likewise, a response-adaptive rule describes the probabilities used at each step; it does not specify the final allocation counts in advance. Those counts depend on the sequence of random assignments and, for an adaptive rule, on the outcomes that influence later probabilities.

The essential operational assumptions are that the rule is specified clearly enough to calculate the probabilities at each assignment, the relevant history is recorded in the order it becomes available, and the randomization mechanism actually uses the stated probabilities. To interpret an update as response-adaptive, outcomes must be measured consistently across arms and participants. If an outcome is used before comparable follow-up is complete, the update can reflect differences in observation time or missingness rather than differences in response.

Worked Example: Updating Probabilities From Two Response Rates

Consider a hypothetical two-arm trial with a fully observed binary response. For illustration, its protocol uses the rule $q_A=0.5+0.2(\hat p_A-\hat p_B)$ and $q_B=1-q_A$, where $\hat p_A$ and $\hat p_B$ are the observed response proportions among participants assigned to the respective arms. Suppose the outcomes available before the next assignment are 3 responses among 4 participants in A and 1 response among 4 in B.

The observed proportions are $\hat p_A=3/4=0.75$ and $\hat p_B=1/4=0.25$. Substituting them into the specified rule gives

q_A=0.5+0.2(0.75-0.25)=0.5+0.10=0.60,\qquad q_B=1-0.60=0.40.

Thus the next participant has a 0.60 probability of assignment to A and a 0.40 probability of assignment to B. These are probabilities for the next randomization, not a claim that 60% of all participants will be assigned to A. This arithmetic illustrates a rule; it does not establish that this particular rule is suitable for a clinical trial.

How the Sequential Probabilities Form a Trial Path

For a realized sequence of assignments, the randomization mechanism contributes one assignment probability at each step: the probability specified by the rule using the history available immediately before that assignment. Multiplying those probabilities gives the product of the sequential assignment kernels evaluated along that realized history. It is important to describe this product accurately: it is not generally the probability of the assignment sequence conditional on the entire history of assignments and outcomes, including outcomes observed later.

A joint path includes outcomes as well as assignments. To calculate its probability under a model, both the assignment probabilities and the outcome probabilities are needed. If $Y_i$ denotes participant $i$'s outcome, the sequential factorization is

P(A_1=a_1,Y_1=y_1,\ldots,A_m=a_m,Y_m=y_m) = \prod_{i=1}^{m} P(A_i=a_i\mid G_{i-1})\, P(Y_i=y_i\mid G_{i-1},A_i=a_i).

Here, $G_{i-1}=(A_1,Y_1,\ldots,A_{i-1},Y_{i-1})$ is the full preceding path, with $G_0$ empty, as required by the chain rule. If the randomization mechanism depends on the preceding path only through $H_{i-1}$, the assignment factor can be written $q_i(H_{i-1})$; the outcome factor can be conditioned only on $H_{i-1}$ if the outcome model justifies that reduction. The factorization does not say that the assignment-path probability conditional on a complete outcome path is the product of the assignment probabilities.

Worked Example: Assignment-Path and Joint-Path Probabilities

Suppose a hypothetical two-participant trial uses $P(A_1=A)=0.5$. After participant 1 responds, the rule assigns participant 2 to A with probability 0.6; after participant 1 does not respond, that probability is 0.4. Assume participant 1 responds with probability 0.6 if assigned A and 0.3 if assigned B. Participant 2, if assigned A, responds with probability 0.6 regardless of the first outcome. Calculate the probability of the path $A_1=A,\;Y_1=1,\;A_2=A,\;Y_2=1$, where $Y_i=1$ means response.

The first assignment has probability 0.5. Given assignment to A, the first response has probability 0.6. Because that response is observed before participant 2 is assigned, the next assignment probability is 0.6. Given assignment to A, participant 2's response probability is 0.6. Multiplying these sequential factors gives

0.5\times 0.6\times 0.6\times 0.6=0.108.

So the specified joint path has probability 0.108 under the stated randomization rule and outcome model. The assignment factors on that path multiply to $0.5\times0.6=0.30$, but 0.30 is the product of the assignment kernels evaluated at the path's pre-assignment histories; it is not, in general, the conditional probability of the assignment sequence given the complete outcome sequence.

For instance, if instead the first outcome were defined to be a response if and only if A was assigned, then $P(A_1=A)=0.5$, but $P(A_1=A\mid Y_1=1)=1$. Conditioning on the outcome changes the assignment probability. This is why assignment kernels must not be confused with assignment probabilities conditional on future outcomes.

Derivation: Factorizing a Joint Path

The path factorization follows by applying the chain rule in temporal order. First, the probability of a joint path is the product of the conditional probability of each successive event given the events already specified. At step $i$, assignment $A_i$ occurs before outcome $Y_i$. Thus the two factors for that step condition on the full preceding path $G_{i-1}$, and then on that path and the assignment:

P(A_i=a_i,Y_i=y_i\mid G_{i-1}) = P(A_i=a_i\mid G_{i-1})\, P(Y_i=y_i\mid G_{i-1},A_i=a_i).

Applying this decomposition successively for $i=1,\ldots,m$ yields the displayed joint-path product, where $G_{i-1}$ contains all preceding assignments and outcomes. The assignment factor reduces to $q_i(H_{i-1})$ when the randomization mechanism depends only on $H_{i-1}$; the outcome factor can be conditioned only on $H_{i-1}$ if the outcome model justifies that reduction. It does not require the outcomes to have identical probabilities across arms or across histories.

Diagnostics, Alternatives, and Limitations

A practical check is to reconstruct each assignment probability from the recorded history and compare it with the probability used by the randomization system. The probabilities for each assignment must be nonnegative and sum to one. The trial record should also make it possible to determine which outcomes were available at each update. Discrepancies can reveal implementation errors, data recorded out of sequence, or an update rule that was not sufficiently specified.

Other warning signs include updates based on different assessment schedules across arms, substantial missing or delayed responses, or abrupt probability changes driven by very few observed outcomes. These issues can make an outcome-based allocation rule behave differently from its intended specification. In particular, if response ascertainment is incomplete when an update occurs, the observed response proportions may not be comparable. The protocol should state whether an update waits for a defined follow-up window and how unavailable outcomes are handled.

Fixed-ratio randomization is the main contrast: its probabilities do not depend on accumulating outcomes, so its assignment mechanism is simpler to describe and audit. Response-adaptive randomization permits probabilities to change as outcomes arrive, but it makes allocation depend on the history. That dependence must be accounted for when describing the design and analyzing the trial. Neither design guarantees equal numbers across arms, and neither allocation probabilities nor observed response differences alone establish a treatment effect.

Response-adaptive allocation also has limits. Early outcomes can be noisy, and an update based on a small amount of information can make later allocation sensitive to chance variation. Unequal allocation may reduce the number of participants contributing to comparisons in some arms. Delayed outcomes complicate the definition of the available history. The sequential rule therefore needs to be judged as part of the full trial design, with attention to its operating characteristics and the scientific objective. The definition of response adaptation does not itself solve these design questions.

Common pitfall: “The assignment-path probability is the product of the allocation probabilities, even after conditioning on all the outcomes.” This is generally false: later outcomes can be related to earlier assignments, so conditioning on them can change the conditional probabilities of those assignments. The product describes the sequential assignment kernels evaluated at the histories available immediately before each assignment. To obtain a joint assignment-and-outcome path probability, include the corresponding outcome probabilities as well.

Worked Example: A Three-Arm Probability Update

Suppose a hypothetical protocol defines three allocation scores from observed response proportions using $s_k=1+\hat p_k$, then sets $q_k=s_k/\sum_j s_j$. Before an assignment, the observed response proportions are 0.50 in arm A, 0.25 in arm B, and 0.00 in arm C. Apply this specified rule.

The scores are $s_A=1+0.50=1.50$, $s_B=1+0.25=1.25$, and $s_C=1+0=1.00$. Their sum is $1.50+1.25+1.00=3.75$. Therefore,

q_A=\frac{1.50}{3.75}=0.400,\qquad q_B=\frac{1.25}{3.75}=0.3333,\qquad q_C=\frac{1.00}{3.75}=0.2667.

The probabilities sum to $0.400+0.3333+0.2667=1.000$, up to rounding. They are the probabilities for the next assignment under this illustrative rule. The score formula is not implied by response-adaptive randomization as a general definition, and its use would require separate design assessment.

Check Your Understanding

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

A two-arm trial assigns every participant with probability 0.5 to each arm. Is it response-adaptive if the observed response rates differ? Explain.
Under the rule $q_A=0.5+0.2(\hat p_A-\hat p_B)$, let $\hat p_A=2/5$ and $\hat p_B=1/5$. Calculate both next-assignment probabilities.
Why is the product of sequential assignment probabilities not generally the assignment-path probability conditional on all outcomes?
Name two features of a trial history that should be checked when auditing an outcome-based update.

Solutions. 1. No. The probabilities remain fixed at 0.5 regardless of the observed response rates, so this is fixed-ratio randomization. 2. The difference is $2/5-1/5=0.2$, so $q_A=0.5+0.2(0.2)=0.54$ and $q_B=1-0.54=0.46$. 3. Conditioning on outcomes observed after assignments can change the conditional probabilities of earlier assignments; the product instead evaluates each assignment kernel using the history available immediately before that assignment. A joint-path probability also includes outcome probabilities. 4. Check that the outcomes used in each update were actually available before the assignment, and that the recorded history reproduces the probabilities implemented by the randomization system. Also verify that probabilities are nonnegative and sum to one.