Three-Arm Dose-Selection Design Example

Designing the Selection and Confirmation Stages

A three-arm dose-selection design begins with a phase 2 stage that randomizes participants to control, a lower dose, or a higher dose. The phase 2 data guide which dose proceeds. The trial then conducts a confirmatory comparison of that selected dose with control. The central design challenge is to distinguish the data used to choose a dose from the data used to test its confirmatory claim.

One transparent design uses phase 2 outcomes only for selection and reserves independent stage 2 participants for confirmation. This approach does not extract confirmatory information from the phase 2 observations, but it makes the effect of selection on the confirmatory test explicit. It also gives a clear example of accounting for the shared phase 2 control: the two dose-versus-control estimates are correlated, so they should not be treated as independent evidence when selecting a dose or describing phase 2 results.

The protocol must specify the endpoint, analysis population, selection rule, transition timing, stage 2 allocation, and confirmatory test before examining the phase 2 results. For example, the rule might first exclude any dose failing a safety criterion, then select the remaining dose with the larger phase 2 mean difference from control; ties could go to the lower dose. A rule that leaves no dose eligible should stop without a confirmatory claim. The particular rule is a design choice, not a conclusion to improvise after seeing outcomes.

This example assumes independent participants across stages, randomization within each stage, a continuous endpoint measured on the same scale in both stages, and normally distributed outcomes with a common variance across the two stage 2 groups for the pooled t test. The primary calculation also assumes complete endpoint data; missingness that changes the analysis population or invalidates the planned model requires a prespecified method. These assumptions should be assessed using design knowledge and diagnostics, not asserted merely because a test can be calculated.

What the Shared Control Changes

In phase 2, both dose contrasts use the same control observations. If the phase 2 group means are \(\bar{X}_L,\bar{X}_H,\bar{X}_C\), the estimated contrasts are \(\hat{\Delta}_L=\bar{X}_L-\bar{X}_C\) and \(\hat{\Delta}_H=\bar{X}_H-\bar{X}_C\). A favorable or unfavorable fluctuation in the control mean moves both contrasts in the same direction. Thus, even with independent randomized treatment groups, the two contrast estimates are dependent.

Under independent observations with common variance \(\sigma^2\) and equal phase 2 group sizes \(n\), each contrast has variance \(2\sigma^2/n\). Their covariance is \(\sigma^2/n\), because the shared control mean appears with a minus sign in both. Their correlation is therefore \(1/2\). With unequal group sizes, the covariance remains \(\sigma^2/n_C\), while the contrast variances become \(\sigma^2(1/n_L+1/n_C)\) and \(\sigma^2(1/n_H+1/n_C)\).

The correlation calculation explains the shared-control issue; it does not require a complicated selection statistic. A simple, fully specified rule can select the dose with the larger observed mean difference, subject to safety eligibility. Inference about the selected dose then relies on the independent stage 2 data in the design described here, not on pretending that the phase 2 contrasts were independent or that the winner's phase 2 estimate is unbiased.

Confirmatory Test and Its Type I Error

After selection, stage 2 randomizes new participants between the selected dose \(J\) and control. For a continuous endpoint, the pooled two-sample t statistic is appropriate under independent normal observations with a common within-group variance. With \(n_J\) and \(n_C\) stage 2 participants, respective means \(\bar{x}_J,\bar{x}_C\), and pooled standard deviation \(s_p\), the statistic and degrees of freedom are:

The confirmatory test rejects \(H_0\) when \(T>t_{1-\alpha,df}\), where \(t_{1-\alpha,df}\) is the \(1-\alpha\) quantile of the central t distribution with \(df\) degrees of freedom. Because the stage 2 participants are independent of phase 2, conditioning on the phase 2 data fixes the selected dose without selecting stage 2 outcomes. Under the null boundary \(\mu_J-\mu_C=0\), the conditional statistic has a central t distribution. For a strictly negative difference, its distribution is noncentral t, not central t; its right-tail rejection probability is no greater than at the boundary.

This argument also gives a strong type I error guarantee for claims about the selected dose. For any dose \(j\) whose null is true, the probability of selecting \(j\) and then falsely rejecting its null is at most \(\alpha P(J=j)\), because the stage 2 test has conditional size at most \(\alpha\). Summing over doses whose nulls are true gives a total false-claim probability no greater than \(\alpha\sum_{j:\,H_{0j}\text{ true}}P(J=j)\leq\alpha\). This guarantee depends on the independence and prespecification conditions; it does not justify using the selected phase 2 estimate as if selection had not occurred.

Diagnostics, Alternatives, and Limitations

Before making a confirmatory claim, check that the actual conduct matches the design: the endpoint and analysis population were not changed in response to results; the chosen dose followed the written rule; and the stage 2 comparison uses only its assigned participants as planned. Inspect group-level distributions and residuals for severe departures from normality or influential observations, and compare variability across groups. Marked heteroscedasticity undermines the exact pooled t procedure; a suitable alternative, such as a prespecified Welch test, changes the variance calculation and degrees of freedom and should be specified in advance.

The independent stage 2 test is a simple alternative to approaches that combine phase 2 and stage 2 evidence using a prespecified combination test or conditional-error method. Those approaches can use more information, but require a carefully specified procedure that preserves type I error despite selection and dependence. An ordinary pooled analysis that treats selected phase 2 data as independent confirmatory evidence is not a substitute. The design here trades some efficiency for a direct conditional validity argument.

Important limitations remain. The selected dose may not be the best dose in truth, and the winner's phase 2 effect estimate is typically optimistic because selection favors high observed values. The confirmatory result applies to the selected dose and the tested endpoint and population; it does not establish that the selected dose is superior to the unselected dose. A shared-control selection rule can also be sensitive to an unusually favorable or unfavorable control mean. If the stage 2 population, endpoint measurement, or treatment implementation differs materially from phase 2, the intended interpretation may no longer hold.

A common mistake is to reason, “Only one dose is tested, so selection cannot affect the false-positive rate,” and then include the selected dose's phase 2 observations in an ordinary pooled test. The correction is to ask whether the confirmatory data are independent of the selection. In this design they are, so the conditional stage 2 test controls its size; if phase 2 outcomes are reused, a method that accounts for selection and dependence is needed.