Why Match Allocation to Posterior Probabilities?
In a response-adaptive trial, allocation can depend on the outcomes observed so far. For a binary outcome such as response by a specified time, one possible rule uses the current posterior probability that each treatment arm has the highest response probability. It gives the next patient a higher chance of receiving an arm that is more likely to be best, while retaining randomization when the evidence is uncertain.
This rule is called posterior probability matching. It does not assign every patient to the arm with the largest posterior mean response probability. Instead, it makes the arm’s chance of receiving the next patient equal to its posterior chance of being best, subject to a precise definition of “best” and appropriate handling of ties.
Model and Definition
Suppose there are \(K\) treatment arms. Let \(\theta_j\) denote the probability of response on arm \(j\), for \(j=1,\ldots,K\). For the basic model, assume each eligible patient’s response on arm \(j\), if assigned that arm, is a Bernoulli outcome with probability \(\theta_j\). Conditional on the arm-specific response probabilities, outcomes are independent across patients. The response definition and assessment time are the same across arms, and the response probability is stable over the patients represented by the model.
Let \(D\) denote the observed trial data available at an allocation update, including the arm assignments and observed binary outcomes. A Bayesian model and prior distribution produce a posterior distribution for \((\theta_1,\ldots,\theta_K)\) given \(D\). Here, “best” means having the largest response probability: arm \(j\) is best when \(\theta_j\) exceeds the response probability of every other arm.
The tie condition matters: without it, the probabilities \(P(\theta_j=\max_k\theta_k\mid D)\) need not sum to one. With a tie-breaking rule, the events “arm \(j\) is selected as best” are mutually exclusive and exhaustive. Their posterior probabilities therefore sum to one, so they define valid assignment probabilities. For example, a rule might choose uniformly among the tied best arms. A deterministic rule that favors the lowest-numbered arm also produces valid probabilities, but can favor that arm when ties occur.
A common working model for binary outcomes assigns independent Beta priors to the arm-specific response probabilities. If arm \(j\) has prior \(\theta_j\sim\operatorname{Beta}(a_j,b_j)\), and the data contain \(s_j\) responses among \(n_j\) patients on that arm, then \(\theta_j\mid D\sim\operatorname{Beta}(a_j+s_j,b_j+n_j-s_j)\). With positive Beta parameters, the joint posterior is continuous, so posterior ties have probability zero. This convenient tie property does not remove the need to state a tie rule when using a model that can have point masses or other discrete components.
Worked Example: Two-Arm Probability Matching
Suppose two hypothetical arms, A and B, each have a \(\operatorname{Beta}(1,1)\) prior for their response probability. After the first observations, A has one response among one patient and B has zero responses among one patient. Find the matching allocation probabilities for the next patient.
The Beta–binomial update gives \(\theta_A\mid D\sim\operatorname{Beta}(2,1)\) and \(\theta_B\mid D\sim\operatorname{Beta}(1,2)\). Their posterior densities on \(0<x<1\) are \(f_A(x)=2x\) and \(f_B(x)=2(1-x)\). The posterior cumulative distribution function for \(\theta_B\) is \(F_B(x)=2x-x^2\). Because both distributions are continuous, the probability of a tie is zero.
Condition on \(\theta_A=x\). The probability that B’s response probability is below \(x\) is \(F_B(x)\). Integrating over A’s posterior gives
Thus \(q_B(D)=1-q_A(D)=1/6\approx0.1667\). The next patient is assigned to A with probability about 83.33% and to B with probability about 16.67%. These are randomization probabilities conditional on the observed data, not claims that A has an 83.33% response rate or that this probability of allocation guarantees a better outcome for the patient.
Why the Probabilities Form an Allocation Rule
The justification is a probability partition. Under a continuous posterior with zero probability of ties, exactly one arm has the largest \(\theta_j\) in each posterior draw. Therefore the events that the individual arms are best partition the posterior sample space. In symbols,
If ties are possible, use instead the events that each arm is selected as best by the stated tie-breaking rule. Those events still partition the possibilities, including tied outcomes, and their probabilities sum to one. A practical implementation can draw one arm from the categorical distribution with probabilities \(q_1(D),\ldots,q_K(D)\), then record the actual assignment.
This argument establishes that the rule is a valid randomization distribution, not that it is optimal for every clinical or scientific objective. Matching posterior probabilities gives an arm a share equal to its chance of being best. The rule does not directly account for how much better one arm might be than another, the clinical importance of a response, or a separate objective such as learning about treatment effects.
Worked Example: An Explicit Tie-Breaking Rule
Consider a hypothetical model in which the posterior assigns probability one to \(\theta_A=\theta_B=0.50\). Both arms are tied for best with certainty. If one incorrectly defines each allocation probability as \(P(\theta_j=\max_k\theta_k\mid D)\), then \(q_A(D)=1\) and \(q_B(D)=1\), which sum to 2 and cannot be assignment probabilities.
Now specify that ties are broken uniformly at random. Conditional on this posterior state, the tie-breaking step selects A with probability \(1/2\) and B with probability \(1/2\). Hence the matching probabilities, defined as the probabilities of being selected as best, are \(q_A(D)=1/2\) and \(q_B(D)=1/2\). They sum to 1 and give a valid allocation rule.
This example is deliberately discrete. Under the independent Beta posteriors in the preceding example, an exact tie has probability zero, so this particular correction is not needed for the numerical calculation there. It remains essential to define the rule for models or posterior distributions where ties can occur.
Updating the Allocation After an Outcome
The assignment probabilities are recalculated when new outcome data become available. The model must specify which outcomes are available at each update; if a patient’s binary outcome has not yet been assessed, it should not be counted as an observed response or nonresponse. This basic illustration assumes outcomes are available before the next allocation decision.
Worked Example: A Sequential Update Changes the Match
Continue the hypothetical two-arm trial in the first example. Given its initial data, the next patient has probability \(1/6\) of being assigned to B. Suppose that assignment occurs and the patient responds. With the same \(\operatorname{Beta}(1,1)\) priors, A still has posterior \(\operatorname{Beta}(2,1)\), while B now has one response and one nonresponse, giving posterior \(\operatorname{Beta}(2,2)\).
For \(\theta_B\sim\operatorname{Beta}(2,2)\), its density is \(6y(1-y)\), and its cumulative distribution function is \(F_B(x)=3x^2-2x^3\). A’s density remains \(2x\). The probability that A is best is therefore
With no posterior tie, \(q_B(D_{\mathrm{new}})=1-0.7000=0.3000\). Conditional on this particular observed response and the resulting data, the next patient would be assigned to A with probability 70.00% and to B with probability 30.00%. One outcome changes the evidence and thus the allocation probabilities; it does not establish that B is superior or inferior.
Assumptions, Diagnostics, and Alternatives
The calculation is only as useful as the model and the outcome information supplied to it. Before using the rule, check that the binary response has a clinically meaningful definition and common assessment window, that the recorded outcomes match those definitions, and that the likelihood is appropriate for the trial. The simple Beta–binomial model treats outcomes as conditionally independent with a constant arm-specific response probability. Strong patient heterogeneity, time trends, differences in follow-up, or treatment effects that vary by patient may make that model inadequate. A model that ignores such structure can yield allocation probabilities that are numerically well-defined but poorly calibrated to the trial.
Diagnostics should include checking arm-specific response and nonresponse counts, verifying the posterior update from those counts, and confirming that the calculated \(q_j(D)\) values are between zero and one and sum to one within numerical tolerance. If probabilities are estimated by posterior simulation, check that enough draws were used for stable estimates, particularly when one arm’s probability of being best is small. In each draw, count exactly one selected best arm; apply the stated tie rule if necessary. Also check that only outcomes available under the protocol have entered the update. Data errors, inconsistent endpoint coding, or silently treating pending outcomes as failures can distort the allocation.
A fixed randomization scheme uses prespecified assignment probabilities that do not change in response to accumulating outcomes. Posterior probability matching is adaptive: its probabilities change with the posterior evidence. Both are randomization procedures, but the adaptive rule can produce greater allocation imbalance and can make the assignment mechanism depend on earlier outcomes. Response-adaptive randomization is a broader class of rules; posterior probability matching is one particular way to set its probabilities. Another possible design objective is to allocate according to posterior expected utility, which is a different criterion and should not be treated as equivalent to probability of being best.
Posterior probability matching also has limitations as a clinical-trial design choice. The arm with the greatest chance of being best need not have a clinically meaningful advantage, and the rule does not by itself protect a control-arm sample size, guarantee adequate information for comparisons, or ensure acceptable patient-level risk. Those concerns require explicit design decisions and monitoring, rather than an assumption that matching probabilities automatically address them. The rule is best understood as a transparent link between posterior ranking uncertainty and randomization, whose suitability depends on the trial’s objectives and safeguards.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- Three arms have posterior probabilities of being best \(0.20\), \(0.50\), and \(0.30\), with no ties. What are the next-patient assignment probabilities?
- Why can the probabilities \(P(\theta_j=\max_k\theta_k\mid D)\) fail to define an allocation rule when ties are possible?
- In a two-arm trial, the posterior probability that A is best is \(0.72\), with zero posterior probability of a tie. What is the matching probability for B?
- Name one model assumption in the simple Beta–binomial setup that could be questionable in a clinical trial, and explain why checking it matters.
Solutions. (1) Assign with probabilities \(0.20\), \(0.50\), and \(0.30\); they sum to one. (2) The best-arm events can overlap when two or more arms tie, so their probabilities may sum to more than one. A tie-breaking rule makes the events of being selected as best mutually exclusive. (3) \(q_B(D)=1-0.72=0.28\), so B receives the next patient with probability 0.28. (4) For example, the model assumes a constant response probability within each arm. If response chances change over calendar time or vary substantially with patient characteristics, a constant-probability model may misrepresent the evidence and produce misleading allocation probabilities.