Design Around the Decision
A Phase II dose-ranging study is useful when its results can separate plausible dose-response patterns that would lead to different development choices. A large sample is not automatically informative: if all active doses lie in a narrow range, the study may estimate responses precisely there while failing to distinguish a low-dose threshold from a plateau or continued improvement at higher doses. Design therefore starts with the decision the study must support, not with a convenient number of arms.
The earlier material on dose-selection estimands and decision criteria provides the framework for specifying the population, endpoint, treatment contrast, and rule for selecting a dose. Here the emphasis is on translating that decision into dose arms, allocation, sample size, and duration. The design should be assessed under several plausible response patterns, including patterns near the boundary where the preferred dose might change.
Choosing Dose Arms to Separate Response Patterns
Dose selection for the study should use the available pharmacology, predicted dose-response information, clinically meaningful response thresholds, and safety constraints. The dose range should include placebo or another appropriate reference, doses expected to be below and above a clinically meaningful response threshold when feasible, and doses that can reveal whether response is approaching a plateau. An arm above the likely plateau may be valuable if it clarifies the shape or safety trade-off, but an implausibly high or unsafe dose is not informative enough to justify its inclusion.
The spacing of doses matters as much as the number of arms. Closely spaced doses can characterize a narrow part of the curve, but may be poor at separating a threshold pattern from a gradual increase. Wider spacing can distinguish broad patterns but may leave uncertainty about where a clinically useful dose lies. Logarithmic spacing is often a sensible starting point when changes in dose are more naturally interpreted as proportional than additive; it is not a universal rule. Candidate schedules should be checked against predicted responses and the actual selection criterion.
Worked Example: Placing Arms to Distinguish a Threshold From a Plateau
Suppose a hypothetical drug can be studied at 0, 10, 30, 60, or 120 mg once daily. Prior evidence suggests that a clinically meaningful response might begin between 10 and 30 mg, while the response could plateau between 60 and 120 mg. The prespecified rule is to select the lowest dose expected to achieve the meaningful response, provided its safety profile is acceptable.
A design with placebo and 10, 30, and 60 mg samples the region below the anticipated threshold, near the threshold, and near the expected plateau. It can potentially distinguish a pattern in which 10 mg is already useful from one in which response emerges only around 30 mg. The 60 mg arm provides information about the higher end of the likely useful range.
If the safety profile at 120 mg is acceptable and uncertainty about the plateau could change the decision, adding 120 mg may be justified. It would help distinguish continued improvement beyond 60 mg from a plateau, and could expose a dose-related tolerability cost. If 120 mg is not a plausible treatment option or is not ethically acceptable, including it solely to make the dose range look broad would not improve the decision.
The study should not assume in advance that the anticipated pattern is correct. Its design scenarios should also include a flat response, a response that continues to increase across the range, and responses close to the meaningful-response boundary. These scenarios reveal whether the selected arms can distinguish decisions that matter.
Allocation and Precision for Dose Comparisons
For a continuous endpoint, a simple planning approximation can compare arm means. Suppose each arm has a common standard deviation \(\sigma\), observations are independent between participants, and each participant contributes one endpoint measurement. For an active-dose mean minus the placebo mean, the variance is approximately \(\sigma^2(1/n_0+1/n_A)\), where \(n_0\) and \(n_A\) are the placebo and active-dose sample sizes. This approximation assumes no material difference in endpoint variance across arms and no adjustment for other sources of uncertainty.
Equal allocation is straightforward and provides similar precision for comparisons among equally sized arms. If several active doses will each be compared with a shared placebo, assigning somewhat more participants to placebo can reduce the combined variance of those contrasts. That trade-off is relevant only if the placebo comparisons are important enough to justify fewer participants per active dose.
Worked Example: Comparing Equal and Placebo-Weighted Allocation
A hypothetical study has 300 evaluable participants, one placebo arm, and three active-dose arms. Assume a common endpoint standard deviation of 15 points. First allocate 75 participants to each of the four arms. The standard error for any active-dose versus placebo difference is
For the stated objective of minimizing the sum of the three contrast variances, \(k=3\), so the optimal continuous ratio is \(n_0/n_A=\sqrt{3}\approx1.732\). With \(n_0+3n_A=300\), this gives \(n_A=300/(3+\sqrt{3})\approx63.40\) and \(n_0\approx109.81\). A practical integer allocation totaling 300 is 111 on placebo and 63 on each active dose. Its contrast standard error is
The placebo-weighted allocation gives a smaller standard error for each active-versus-placebo contrast than equal allocation under these assumptions. The difference is modest, and the weighted allocation has fewer participants in each active arm. If the decision depends mainly on comparing active doses with one another, or if dose-specific safety estimates require larger active groups, equal allocation or another allocation may be preferable.
Sample Size Must Match the Decision
A sample-size calculation should use a clinically relevant difference, a defensible variability assumption, a planned error rate, and the comparisons or selection rule that will actually be used. For an individual two-sided comparison under a normal approximation, a test statistic has approximate mean \(\delta/SE\) under a true difference \(\delta\). Power is the probability of rejecting the null hypothesis under that specified alternative. For several dose comparisons, the design must also address multiplicity if the intended conclusions rely on a family of tests.
For example, consider the weighted allocation above and a clinically relevant placebo-adjusted difference of 8 points. The noncentrality value is \(8/2.366\approx3.381\). If three two-sided comparisons are protected by a Bonferroni familywise significance level of 0.05, each comparison uses \(\alpha=0.05/3\), with an approximate critical value \(z_{1-\alpha/2}=2.394\). The approximate power for each comparison is
This is power for detecting that specified difference in one comparison under the stated approximation. It is not the probability of selecting the correct dose. Correct selection depends on the responses at all doses, their uncertainty, the selection rule, and safety information. If the goal is to choose among doses rather than simply to reject a set of null hypotheses, assess operating characteristics under complete dose-response scenarios—for example, the probability of selecting a clinically useful dose, selecting a dose that is too high, or making no selection.
Normal approximations are useful for early planning when sample sizes are moderate or large and the assumed endpoint distribution is reasonable. Final planning may require a \(t\)-based calculation, simulation, or a model that reflects the planned analysis, baseline adjustment, missing data, and multiplicity procedure. A calculation based on a single assumed standard deviation should be checked across plausible values rather than treated as known.
Duration, Dropout, and Operational Feasibility
Study duration determines whether the chosen endpoint can represent the treatment response of interest. A short treatment period may be sufficient for a rapidly equilibrating pharmacologic effect, but not for delayed response, accumulation, titration, or an endpoint that changes slowly. The planned measurement time should allow adequate exposure and response development while avoiding an unnecessarily long study. Safety follow-up should also be long enough to capture risks that could alter the dose decision.
Recruitment time and participant follow-up jointly determine calendar duration. Dropout should be accounted for when converting an evaluable sample-size target into a randomized target. The simple adjustment below assumes the same dropout proportion in every arm; if dropout is expected to differ by dose, arm-specific inflation or simulation is more appropriate.
Worked Example: Inflating for Dropout and Estimating Calendar Time
Suppose the design needs 300 evaluable participants with the 111:63:63:63 allocation above, and 10% of randomized participants are expected not to provide the primary endpoint. The approximate total enrollment required is \(300/(1-0.10)=333.3\), so at least 334 participants should be randomized. Scaling the allocation gives approximately 123.3 placebo participants and 70 in each active arm. One integer allocation totaling 334 is 124 on placebo and 70 per active dose.
Assume, hypothetically, that recruitment proceeds at 20 participants per month, treatment lasts 12 weeks, and safety follow-up lasts a further 4 weeks. Recruitment then takes about \(334/20=16.7\) months. The last participant completes 16 weeks after entering, so a rough minimum from first enrollment to last follow-up is \(16.7+16/4.35\approx20.4\) months, before allowing for start-up, recruitment variability, data cleaning, or delays.
This calendar estimate is conditional on the stated recruitment rate and follow-up schedule. Before finalizing duration, confirm that 12 weeks is appropriate for the expected onset and stabilization of response, and that 4 weeks of follow-up is sufficient for relevant safety evaluation. If either assumption fails, the nominal sample size can be met while the study remains uninformative for selection.
Diagnostics, Alternatives, and Limitations
Before launch, check whether the design can answer its intended question under realistic scenarios. Vary the assumed response shape, endpoint variability, dropout, and recruitment rate. Look for doses whose expected responses are nearly indistinguishable, a range that omits plausible minimum effective or plateau doses, and a sample size that gives adequate pairwise power but poor probability of making the intended selection. For duration, verify that the primary endpoint is measured after sufficient treatment and that follow-up captures decision-relevant safety.
Equal allocation is simple and protects precision across arms; placebo-weighted allocation can improve shared-placebo comparisons when those comparisons dominate the objective. A placebo-free active-control design may be appropriate when placebo is not acceptable or the decision is explicitly comparative, but it does not estimate response relative to placebo in the same way. A crossover design can reduce between-participant variability in some settings, yet is unsuitable when carryover, long-lasting effects, disease progression, or a long washout makes period comparisons unreliable. These choices should follow the estimand and practical constraints rather than a general preference for one design.
A fixed parallel-group design also has limitations: it cannot guarantee that the chosen dose will be optimal, and uncertainty may remain even with adequate planning. The design depends on assumptions about endpoint variability, adherence, dose-response behavior, and treatment duration. If results will be interpreted only at tested doses, ensure that the arms cover plausible choices; if interpolation is central to the decision, the prespecified analysis and its assumptions must support it. The methods for formal multiple-comparison dose-response modeling and adaptive dose-ranging address distinct design and analysis questions and should not be substituted for clear specification of the decision itself.
Common mistake: “The study has high power to show that the highest dose differs from placebo, so it will identify the best dose.” This does not follow. A significant high-dose contrast may coexist with similar responses at lower doses, unacceptable safety, or inadequate precision to distinguish the preferred doses. The correction is to evaluate the full prespecified selection rule across plausible efficacy and safety scenarios, not to equate one significant contrast with a successful dose-selection study.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Why might a design with several closely spaced active doses fail to distinguish a threshold response from a plateau?
- With one placebo and three equally sized active arms, what is the continuous allocation ratio \(n_0/n_A\) that minimizes the sum of the active-versus-placebo contrast variances under equal variances and equal costs?
- Why is power for a single active-versus-placebo contrast not the same as the probability of selecting the correct dose?
- A study requires 240 evaluable participants and anticipates 20% endpoint loss. Approximately how many participants should be randomized?
- Name two reasons a 12-week treatment period might be inadequate for dose selection.
Solutions. (1) The arms may cover only a narrow region, leaving little information below a possible threshold or above a possible plateau; include doses that distinguish those patterns. (2) The ratio is \(n_0/n_A=\sqrt{3}\approx1.732\), under the assumptions and objective stated. (3) A contrast tests one specified difference, whereas correct selection depends on responses and uncertainty across all doses, the selection rule, and safety. (4) \(240/(1-0.20)=300\), so approximately 300 participants should be randomized if the loss proportion applies uniformly. (5) Response may emerge slowly, exposure may not have stabilized, the endpoint may change slowly, or safety risks relevant to the decision may require longer observation.