Worked Selection-Rule Design for a Three-Arm Trial

From Interim Estimates to a Complete Decision

At an interim analysis, a three-arm trial with two experimental treatments and one shared control must translate observed data into a decision: which experimental treatment, if any, should continue? A useful design does not leave the answer to an informal comparison made after seeing the results. It defines the effect measure, threshold, continuation constraint, fallback, and tie-break in advance.

This tutorial constructs one such policy. It uses the difference in interim mean outcomes as the effect estimate and a fixed threshold for eligibility. A resource constraint permits exactly one experimental arm to continue, while requiring that at least one continue. The rule therefore has to specify what to do both when an arm meets the threshold and when neither does.

The example policy below uses a clinically meaningful interim threshold of 3 outcome points and a one-arm continuation constraint: exactly one experimental arm must be retained. An arm is eligible if its estimated effect is at least 3 points. If one or more arms are eligible, retain the eligible arm with the larger estimated effect. If neither is eligible, retain the arm with the larger estimated effect anyway, to satisfy the minimum-continuation constraint. Whenever the two estimates are tied in either part of the rule, retain \(A\).

The rule is deliberately complete even for an unpromising interim result. If both estimates are below zero, it still retains the less unfavorable arm, because this particular design requires one experimental arm to continue. That is a design constraint, not evidence that the selected arm is beneficial. A trial that permits stopping all experimental treatment would need a different, explicitly stated rule.

Assumptions and What the Estimates Mean

Suppose the outcome is measured on the same scale in all three arms, with larger values representing better outcomes. At the interim analysis, each participant contributes one outcome, treatment assignment is randomized, and the observed participants are analyzed according to their assigned arms. For the calculations below, assume independent participants and use each arm's sample mean and sample standard deviation. These conditions make the sample mean differences interpretable as estimates of mean treatment differences in the interim data.

The estimated differences are descriptive summaries for the selection policy. The threshold of 3 points is a cutoff on an estimate; it is not a test of a null hypothesis and does not by itself control a type I error rate. In particular, a selected estimate above 3 does not establish that the true effect exceeds 3. Selection can favor arms whose interim estimates happen to be high, even when their underlying effects are similar.

For a simple independent-groups calculation, the estimated standard error of an arm-versus-control difference is \(\sqrt{s_j^2/n_j+s_C^2/n_C}\), where \(s_j,s_C\) are the interim sample standard deviations and \(n_j,n_C\) the corresponding sample sizes. This provides a measure of sampling variability; it is not part of the stated selection rule. Since both treatment estimates use the same control group, they are not independent of one another. The operational comparison remains well-defined, but analyses of selection probabilities or error rates must account for the trial's sampling structure.

The rule also assumes that the measurement scale and direction are fixed before the interim data are reviewed. If lower values are better, the effect must be defined with the direction reversed (for example, control mean minus treatment mean), so that positive values still represent benefit. Thresholds cannot be compared meaningfully across differently scaled outcomes without a prespecified conversion.

Worked Applications

Why the Rule Is Complete

A selection policy should give exactly one answer for every possible pair of interim estimates that meets the stated data requirements. The definition above can be checked by splitting all cases into two mutually exclusive possibilities: at least one arm meets the threshold, or neither does.

To verify completeness, first suppose \(E\neq\varnothing\). With two arms, \(E\) contains either one arm or both. If it contains one, that arm is the unique choice; if it contains both, their estimates are either unequal, giving a unique larger estimate, or equal, resolved in favor of \(A\). Now suppose \(E=\varnothing\). Both arms are in the fallback comparison, and again either one estimate is larger or the tie-break selects \(A\). Thus the policy always selects exactly one arm. The construction also shows why a tie-break only for eligible arms is insufficient: equal estimates below threshold require a decision too.

Diagnostics, Alternatives, and Limitations

Before applying the rule, check that the inputs support the intended comparison. Confirm that the outcome definition, direction, units, interim time point, and analysis population are consistent across arms. Check the arm-specific sample sizes and summaries for data errors, implausible values, or missingness that could make the observed means misleading. The earlier material on protecting selection decisions from data-quality problems addresses such checks in greater depth; a numerical rule cannot repair corrupted or incomparable inputs.

Also inspect how close the estimates are to the threshold and to each other. A choice between 3.01 and 2.99 points is determined by the rule but may be sensitive to small data changes. Likewise, if two estimates are nearly equal, the selected arm can switch with minor sampling variation. Such sensitivity is a reason to evaluate the design across plausible effect patterns and interim sample sizes, not to change the rule after seeing the observed estimates.

A fixed estimate threshold is straightforward to explain and implement, but it does not directly incorporate estimate precision. A conditional-power criterion instead uses a model for future data and a target for the probability of eventual success; it requires assumptions about future effect and information. A rule based on standardized statistics weights estimated effects by their uncertainty, but it answers a different question and still needs a complete continuation and tie policy. Neither alternative should be substituted after interim results are known.

The forced fallback is an important limitation. If both estimates are poor, the design continues the less poor arm because it requires one experimental arm to proceed. If continuing no experimental arm is acceptable, the protocol can instead allow an empty continuation set under a separately specified stopping condition. A further limitation is that the 3-point threshold is a design choice requiring clinical justification; it is not universally appropriate.

Finally, this tutorial's selection policy does not establish confirmatory evidence or provide familywise error control for later testing. Selecting an arm and testing a treatment effect are distinct operations. The selected estimate is affected by the selection process, and a later inferential procedure must be designed for that sequential, multi-arm setting rather than treating the chosen arm as if it had been selected independently of the data.