Why the Timing of Dose Decisions Matters
In dose finding, an allocation rule can use both efficacy and toxicity outcomes to guide the next assignment. But the rule cannot use information that has not yet arrived. If several patients are assigned together, their doses are fixed before any of their outcomes can influence the next decision. If allocation is updated after each patient, new information can affect the next assignment sooner—provided the outcomes are observed in time.
The distinction is about the timing of allocation decisions, not about changing the outcome model. Both strategies can use the same joint dose model and the same benefit-risk rule. Cohort allocation may be operationally convenient, while individual updating can respond more quickly to concerning evidence. Neither is inherently safer or more informative in every setting: the result depends on cohort size, outcome delay, model assumptions, and the decision rule.
A joint model represents the four category probabilities at dose \(d\), for example \(p_{10}(d)=P(E=1,T=0\mid d)\). Their sum is one. The marginal efficacy probability is \(p_{10}(d)+p_{11}(d)\), and the marginal toxicity probability is \(p_{11}(d)+p_{01}(d)\). Keeping the joint categories matters: efficacy and toxicity may occur together, and treating their marginal probabilities as if they described separate patients can misrepresent benefit-risk.
For a transparent comparison, suppose the allocation policy selects the eligible dose with the greatest posterior expected utility. Use utility \(u(E,T)=E-0.8T\), so efficacy without toxicity has utility 1, no efficacy and no toxicity has utility 0, efficacy with toxicity has utility 0.2, and toxicity without efficacy has utility \(-0.8\). A dose is eligible only if its posterior predictive toxicity probability is at most 0.25. These are illustrative policy choices, not universal clinical thresholds.
What Changes When Allocation Is Batched?
At a cohort boundary, the design uses all outcomes available at that time to update the joint model and apply the allocation rule. Within the next cohort, assignments are held fixed. With individual updating, the same model and rule are applied after each evaluable outcome. Thus, if the cohort size is \(m\), as many as \(m-1\) assignments may occur before the first cohort outcome can affect allocation. That delay can matter when evidence changes the dose’s eligibility or ranking.
This comparison assumes that the same outcome definition, model, prior information, utility, and toxicity constraint are used in both strategies. It also assumes that dose assignments follow the stated rule, patients are comparable for the model’s purposes, and joint outcomes are observed reliably. If outcomes are delayed, “patient-by-patient” updating is possible only after the relevant outcome becomes available; one patient enrolling per decision does not mean one patient’s outcome is immediately known.
The allocation schedule alone does not create extra outcome information. It changes which dose later patients receive and when the model can respond. Under conditional independence given dose and model parameters, the outcome likelihood for assigned doses \(d_1,\ldots,d_n\) is
Here \(\theta\) denotes the joint-model parameters. If dose assignment is a deterministic function of previously observed data and the rule does not depend on unknown \(\theta\), the assignment rule adds no parameter-dependent likelihood factor. The outcomes still update the model through the probabilities at the doses actually assigned. This does not remove the need to account for delayed, missing, or selectively observed outcomes in a real analysis.
Worked Example: A Cohort Update Versus an Individual Update
Consider two doses and the policy defined above. Suppose the current posterior category counts are \((1,7,0,2)\) at dose 1 and \((3,5,1,1)\) at dose 2, in the order \((1,0),(0,0),(1,1),(0,1)\). For this example, assume independent Dirichlet priors by dose and use the posterior category means as the predictive probabilities. Dose 1 has total count 10: its toxicity probability is \((0+2)/10=0.20\), efficacy probability is \((1+0)/10=0.10\), and expected utility is \(0.10-0.8(0.20)=-0.0600\). It is eligible.
At dose 2, the total count is also 10. Its toxicity probability is \((1+1)/10=0.20\), efficacy probability is \((3+1)/10=0.40\), and expected utility is \(0.40-0.8(0.20)=0.240\). It is eligible and has the higher utility, so the policy assigns the next patient or cohort to dose 2.
Suppose the first dose-2 patient has outcome \((0,1)\). With individual updating, the dose-2 counts become \((3,5,1,2)\), total 11. Its predictive toxicity probability is \((1+2)/11=0.273\), which exceeds the 0.25 limit. Dose 2 is no longer eligible, so dose 1 is selected from the two available doses.
With a cohort of three, all three assignments to dose 2 are made before any cohort outcome informs the next decision. If all three outcomes are \((0,1)\), the updated counts are \((3,5,1,4)\), total 13, and the predictive toxicity probability is \((1+4)/13=0.385\). Dose 2 is then ineligible, and the next allocation is dose 1. The two approaches can therefore reach the same subsequent decision, but the cohort policy exposed two additional patients to dose 2 before the first observed toxicity could affect allocation. This example assumes outcomes are available immediately; delays would postpone either update.
Safety Exposure and a Short Calculation
An allocation policy can limit exposure only through what it changes. If every patient in a cohort receives the same dose regardless of interim outcomes, a cohort of size \(m\) cannot react to those outcomes until the cohort has been assigned. Conversely, individual updating can change later assignments after early evidence—but only if the outcome is observed, the model responds to it, and the decision rule permits a change.
A simple calculation illustrates the possible scale of the difference. Assume independent patient outcomes at a high dose with toxicity probability 0.40. A cohort of three assigned there has expected toxicity count \(3(0.40)=1.20\), and probability \(1-(1-0.40)^3=0.784\) of at least one toxicity. Now consider individual allocation: assign the first patient high; if that patient has toxicity, assign the next two to a lower dose with toxicity probability 0.10; otherwise keep them at the high dose. Assume outcomes are immediately known and independent conditional on assigned dose.
Under these assumptions, individual updating reduces the expected number of toxicities from 1.20 to 0.960 in three assignments. It does not change the probability of at least one toxicity: the first patient is at the high dose in both strategies, and if that patient has no toxicity, the policy keeps the remaining patients at the high dose. The probability remains \(0.40+0.60\{1-(0.60)^2\}=0.784\). The calculation demonstrates a possible reduction in subsequent exposure, not a guaranteed reduction in risk or a general safety result.
Diagnostics, Alternatives, and Limitations
Before interpreting a cohort-versus-individual comparison, check the assumptions that make its numerical conclusions meaningful. Confirm that the joint categories are mutually exclusive and exhaustive, that the efficacy and toxicity assessment windows are defined, and that the model’s probabilities are coherent across doses. Check that every allocation decision uses only information available at that time. For delayed outcomes, record which patients’ outcomes are pending; treating pending outcomes as if they were complete can make an update appear faster or more reassuring than it is.
Useful design diagnostics include the number of assignments made before each outcome-based update, the number of patients assigned to each dose, and the frequency with which the policy changes dose after an individual outcome or cohort. Simulations under plausible joint dose-response patterns can reveal whether a policy spends many assignments at a dose that later becomes ineligible, or changes dose repeatedly in response to small amounts of information. A mismatch between observed patterns and model assumptions—such as an unexpected efficacy-toxicity relationship—should prompt caution about the resulting allocations, not a claim that a smaller cohort has fixed the model.
| Allocation approach | Potential advantage | Important limitation |
|---|---|---|
| Patient-by-patient updating | Can respond to each newly observed joint outcome | Requires timely outcomes and can make decisions sensitive to sparse data |
| Cohort-based updating | Provides a scheduled point to review accumulated outcomes and can reduce dose changes between reviews | Assignments within a cohort cannot respond to its outcomes |
| Fixed allocation throughout | Simple assignment schedule | Does not adapt assignments to accumulating joint evidence |
The principal pitfall is to assume that a smaller cohort is automatically safer. That reasoning overlooks whether outcomes are available before the next assignment and whether the rule changes dose when they are. If outcomes arrive only after a full cohort has enrolled, patient-by-patient allocation cannot use them in real time. A second mistake is to infer that a cohort policy learns less in every sense: it may eventually observe the same number of outcomes, but it cannot use a cohort’s outcomes to alter assignments within that cohort. These are different claims about timing and exposure.
Individual updating also has limitations. With few observations, one unusual joint outcome can produce a large posterior change, particularly if prior information is weak or the model does not borrow information across doses. Cohort boundaries can make decisions less reactive, but they do not guarantee stable estimates. Neither approach substitutes for checking the joint model, defining acceptable toxicity, or specifying what happens when evidence is uncertain. The comparison here concerns allocation timing; separate safety stopping policies address whether enrollment should cease under specified evidence.
Worked Example: Choosing a Cohort Size Under Delayed Outcomes
Suppose a center can enroll three patients quickly, but each patient’s joint efficacy-toxicity outcome is assessed only after 21 days. A nominally patient-by-patient rule would assign patient 2 before patient 1’s outcome is known, and patient 3 before patient 2’s outcome is known. In this operational setting, all three assignments are effectively made without feedback. Calling the schedule “patient-by-patient” does not make it adaptive during those 21 days.
If enrollment can instead be staggered so the next assignment waits for the prior outcome, the rule can update between patients. That changes calendar time and may lengthen recruitment. The relevant design choice is therefore not just cohort size: it is the number of assignments made while outcomes remain unavailable. The design team should compare those schedules under the same joint model and policy, reporting both expected assignments by dose and the delays that make feedback possible.
Define efficacy, toxicity, their assessment windows, and when each outcome is considered available for an allocation decision.
State the model, utility or ranking criterion, eligibility constraints, and how ties or unavailable outcomes are handled.
Apply the same policy to cohort and individual schedules; count assignments made before feedback and summarize dose exposure and outcomes.
Check Your Understanding
Answer each question, then compare with the solutions below.
- In the joint-outcome notation used here, express the marginal toxicity probability using category probabilities.
- A cohort has four patients, all assigned before any outcomes are available. How many assignments can be made before the first outcome could change a dose decision?
- At a dose, the four posterior category means are \(0.30, 0.40, 0.10, 0.20\), in the stated category order. Calculate the predictive toxicity probability and determine eligibility under the 0.25 limit.
- Explain why patient-by-patient allocation may fail to provide faster feedback when outcomes are delayed.
Solutions. (1) Toxicity is \(p_{11}+p_{01}\), because it occurs in categories \((1,1)\) and \((0,1)\). (2) Three assignments can follow the first assignment before any outcome is available to alter the remaining doses; in this stated schedule, all four are already assigned, so no assignment in that cohort can change. (3) Toxicity probability is \(0.10+0.20=0.30\), which exceeds 0.25, so the dose is ineligible. (4) If the next patient is assigned before the previous patient’s relevant outcome is observed, the allocation rule has no new outcome information to use; the nominally individual schedule then has no faster outcome-based feedback.