Why adapt dose allocation?
A dose-ranging trial must do two things at once: estimate how responses differ across doses and gather enough information at promising doses to support a decision. A fixed allocation preserves a planned comparison, but it does not respond to accumulating evidence. A purely greedy rule, in contrast, can send nearly everyone to the dose that looks best early on, before the less-favored doses have been adequately assessed. Bayesian adaptive dose-ranging uses updated probability distributions to guide allocation while explicitly protecting continued learning.
The allocation rule and the final selection rule are related but distinct. The allocation rule determines who is assigned to each dose as data accumulate. The final rule applies to the complete trial evidence and should reflect the clinical decision, such as requiring a sufficiently high probability of a clinically meaningful response before choosing the lowest qualifying dose. Neither rule is determined by Bayesian analysis alone: both depend on prespecified clinical objectives and design choices.
Model, updating, and prediction
For a simple illustration, let \(Y_{ij}\) be the efficacy outcome for participant \(i\) assigned to active dose \(d_j\), expressed on a scale where larger values indicate better response. Suppose the placebo-referenced outcomes follow the model \(Y_{ij}\mid\theta_j\sim N(\theta_j,\sigma^2)\), independently across participants and doses, with known residual standard deviation \(\sigma\). Here \(\theta_j\) is the mean response at dose \(d_j\). For each dose, assign an independent normal prior \(\theta_j\sim N(\mu_0,\tau_0^2)\). These assumptions are deliberately simple; an applied analysis must use a model appropriate to its endpoint, trial design, and available information.
After observing \(n_j\) participants at dose \(d_j\), with sample mean \(\bar{x}_j\), the posterior distribution is normal. Its variance and mean are:
The posterior describes uncertainty about the dose’s mean response. A future participant’s response is more variable: integrating over uncertainty in \(\theta_j\) gives \(Y_{\mathrm{new},j}\mid\text{data}\sim N(m_j,\sigma^2+v_j)\). The term \(\sigma^2\) is individual outcome variability; \(v_j\) is uncertainty about the dose mean. These distributions answer different questions. For example, the posterior probability that a dose’s mean exceeds a meaningful threshold is not the same as the predictive probability that one new participant exceeds that threshold.
A brief derivation of the posterior update
For one dose, the likelihood of the observed sample mean is proportional to \(\exp[-n_j(\bar{x}_j-\theta_j)^2/(2\sigma^2)]\), while the prior density is proportional to \(\exp[-(\theta_j-\mu_0)^2/(2\tau_0^2)]\). Multiplying them adds the two quadratic terms in \(\theta_j\). The coefficient of \(\theta_j^2\) is \(1/v_j=1/\tau_0^2+n_j/\sigma^2\); completing the square gives the posterior mean \(m_j=v_j(\mu_0/\tau_0^2+n_j\bar{x}_j/\sigma^2)\). Thus the posterior precision is prior precision plus data precision. As \(n_j\) increases, the data contribute more precision and the posterior variance decreases.
Worked Example: Updating a Dose-Specific Response
Suppose the hypothetical outcome is measured in response units, \(\mu_0=0\), \(\tau_0=8\) response units, and \(\sigma=10\) response units. At one dose, \(n_j=4\) participants have a sample mean \(\bar{x}_j=6\) response units. The normal model assumes independent outcomes and a known residual standard deviation, as specified above.
The posterior variance is \(v_j=(1/64+4/100)^{-1}=17.98\) response units squared, so the posterior standard deviation is \(\sqrt{17.98}=4.240\) response units. The posterior mean is \(m_j=17.98(4\times6/100)=4.315\) response units. Therefore, \(\theta_j\mid\text{data}\sim N(4.315,17.98)\), where the second normal parameter is a variance.
The predictive distribution for one future participant is \(N(4.315,100+17.98)=N(4.315,117.98)\). Its standard deviation is \(10.86\) response units. If a response above 5 units is of interest, the predictive probability for one future participant is \(P(Y_{\mathrm{new},j}>5\mid\text{data})=1-\Phi((5-4.315)/10.86)\approx0.475\). By contrast, the posterior probability that the dose mean exceeds 5 is \(P(\theta_j>5\mid\text{data})=1-\Phi((5-4.315)/4.240)\approx0.436\). The two probabilities differ because individual outcomes vary around the dose mean.
Turning posterior predictions into allocation
One transparent allocation strategy is to calculate the posterior probability that each candidate dose has the highest prespecified utility, then combine that probability with a uniform exploration component. For an efficacy-only illustration, define the utility at each active dose as its mean response \(\theta_j\). For each posterior draw, identify the dose with the largest utility; the fraction of draws in which dose \(d_j\) is largest estimates its posterior probability of being best, denoted \(p_j^{\mathrm{best}}\). This calculation uses the joint posterior distribution, not just the posterior means.
For \(K\) active doses, an exploration mixture is \(q_j=(1-\rho)p_j^{\mathrm{best}}+\rho/K\), where \(0<\rho\leq1\) is prespecified. Since the probabilities of being best sum to one, the \(q_j\) also sum to one. Moreover, each active dose has \(q_j\geq\rho/K\). The exploration component therefore prevents the allocation probability for any candidate dose from falling below that floor. A protocol can add minimum accrued sample sizes, maximum allocation caps, or fixed placebo allocation as further operational safeguards.
This rule is an example, not a universal optimum. The utility can instead reflect a prespecified clinical criterion, and an allocation rule can use posterior predictive probabilities of clinically relevant outcomes. What matters is that the rule is defined before outcomes are observed, its clinical meaning is clear, and it continues to collect information at doses needed for the final decision.
Worked Example: A Posterior-Guided Cohort Allocation
Consider three active doses and a new hypothetical cohort of 30 participants. At an interim analysis, posterior sampling gives estimated probabilities of being the best dose of 0.10, 0.35, and 0.55 for the low, middle, and high doses, respectively. These estimates are assumed to come from sufficiently many draws of the prespecified joint posterior. Set \(\rho=0.30\), so each dose receives an exploration contribution of \(0.30/3=0.10\).
The allocation probabilities are \(q_{\mathrm{low}}=0.70(0.10)+0.10=0.170\), \(q_{\mathrm{middle}}=0.70(0.35)+0.10=0.345\), and \(q_{\mathrm{high}}=0.70(0.55)+0.10=0.485\). They sum to 1.000, and each is at least 0.100.
The expected cohort allocations are \(30(0.170)=5.10\), \(30(0.345)=10.35\), and \(30(0.485)=14.55\) participants. An implementable allocation might target 5, 10, and 15 participants, respectively, with the randomization procedure handling integer assignments. These are expected or target allocations, not a guarantee of those exact counts. In particular, the lowest dose still receives participants even though its current probability of being best is only 0.10.
Final selection is a separate decision
The final rule should express what counts as adequate evidence for a clinical choice. One possible rule is to choose the lowest tested dose for which the posterior probability of exceeding a clinically meaningful response threshold \(\delta\) is at least a prespecified value \(\gamma\). Formally, select the lowest \(d_j\) satisfying \(P(\theta_j\geq\delta\mid\text{all trial data})\geq\gamma\). If no dose meets the criterion, the rule returns no qualifying dose. This rule is appropriate only when lower dose is preferred among doses that meet the efficacy requirement and other relevant considerations do not change that ordering.
Worked Example: Applying a Posterior Selection Criterion
Suppose a hypothetical trial has three ordered doses with final posterior distributions for mean response \(N(4.8,1.2^2)\), \(N(6.4,1.5^2)\), and \(N(7.0,2.0^2)\), respectively, in response units. The prespecified meaningful-response threshold is \(\delta=5\) response units, and the required posterior probability is \(\gamma=0.80\). Assume the posterior distributions come from the prespecified analysis and that choosing the lowest qualifying dose is clinically appropriate.
For the lowest dose, \(P(\theta_1\geq5\mid\text{data})=1-\Phi((5-4.8)/1.2)=1-\Phi(0.167)\approx0.434\), so it does not qualify. For the middle dose, the probability is \(1-\Phi((5-6.4)/1.5)=\Phi(0.933)\approx0.825\). For the highest dose, it is \(1-\Phi((5-7.0)/2.0)=\Phi(1.000)\approx0.841\).
Both the middle and highest doses meet the 0.80 criterion, so this rule selects the middle dose as the lowest qualifying dose. The result is conditional on the model, prior, data, threshold, and decision rule. It is not a statement that there is an 82.5% probability that the fixed middle dose is truly effective in some model-free sense.
Assumptions, diagnostics, and alternatives
The calculations above rely on a specified probability model, a prior distribution, and a rule that can be implemented as written. In the normal example, the outcome scale is treated as continuous, residual variation is normal with known variance, and observations are independent conditional on dose means. The independent dose priors also imply no borrowing of information across doses. Those assumptions may be unsuitable when outcomes are binary, participants have repeated measurements, exposure differs from assigned dose, or dose response is expected to change smoothly across the range. The model must reflect the actual design and estimand; adaptive allocation does not repair a misspecified model.
Useful operational checks include reviewing whether observed outcomes are compatible with the model’s predictive distributions, confirming that randomization probabilities and accrued arm counts match the protocol, and tracking how posterior uncertainty changes at each dose. Investigate unexpectedly large residuals, poor predictive fit, or instability in allocation probabilities. Check that minimum sample-size requirements are being met and that the adaptive process is not leaving a dose with too little information for the final rule. If outcomes depend on calendar time or delayed follow-up, an allocation update based on incomplete or systematically different information can be misleading.
| Approach | Strength | Main trade-off |
|---|---|---|
| Fixed allocation | Predictable arm sizes and straightforward comparisons | Does not redirect allocation as evidence accumulates |
| Greedy allocation to the current best estimate | Concentrates participants on the dose that currently appears most promising | Can suppress learning about alternatives and amplify early random variation |
| Bayesian allocation with an exploration component | Uses updated uncertainty while retaining a prespecified share for learning | Depends on the model, prior, utility, and allocation choices |
The posterior probability of a dose being best is not itself a clinical selection criterion. A dose could have the largest expected response while failing a meaningful-response threshold, or a higher dose could offer little added benefit. Similarly, allocating more participants to a dose does not establish that it is the best dose. Allocation serves both learning and decision support; the final choice must follow the prespecified clinical rule.
Common mistake and limitations
Common mistake: “The high dose currently has the largest posterior mean, so all remaining participants should receive it.” This reasoning treats an uncertain early estimate as if it were known, abandons comparison with other doses, and may make the final selection depend heavily on chance imbalances early in the trial. A correction is to use the prespecified posterior-based allocation rule, retain its exploration floor and sample-size safeguards, and apply the separate final decision criterion to all trial data.
Bayesian adaptation does not remove uncertainty or guarantee an efficient decision. Results can be sensitive to the prior when data are sparse, and the choice of utility, exploration fraction, cohort size, and stopping rules affects both allocation and information. An allocation that performs well for one assumed response pattern may perform poorly for another. These designs also require reliable interim data handling, simulation under plausible scenarios, and clear communication of how randomization changes. The simple calculations here assume a fixed dose-response model; comparing alternative dose-response models is a separate issue.
Check Your Understanding
Answer each question using the stated assumptions. Solutions follow each item.
- If posterior uncertainty about a dose mean decreases, does the predictive variance for a future individual necessarily become zero? Solution: No. In the normal model it is \(\sigma^2+v_j\). Even if \(v_j\) approaches zero, individual variability \(\sigma^2\) remains.
- Three active doses have posterior probabilities of being best equal to 0.20, 0.30, and 0.50. With \(\rho=0.30\), what are the allocation probabilities under the exploration mixture? Solution: The uniform contribution is \(0.30/3=0.10\). Thus the probabilities are \(0.70(0.20)+0.10=0.240\), \(0.70(0.30)+0.10=0.310\), and \(0.70(0.50)+0.10=0.450\), which sum to 1.
- Why can the posterior probability that a dose mean exceeds a threshold differ from the probability that one future participant exceeds it? Solution: The first concerns uncertainty in \(\theta_j\); the second includes both uncertainty in \(\theta_j\) and individual outcome variability. In the normal model, their variances are \(v_j\) and \(v_j+\sigma^2\), respectively.
- Two doses meet a prespecified posterior threshold for meaningful response, and the rule is to choose the lowest qualifying dose. What additional condition makes selecting the lower dose consistent with the rule? Solution: The doses must be ordered so that the lower dose is preferred among qualifying options, and no other decision consideration in the prespecified rule reverses that preference.
- What is one risk of allocating all new participants to the dose with the largest current posterior mean? Solution: Early random variation can make a dose appear best; concentrating allocation there can leave alternatives insufficiently studied and weaken the final comparison.