Arm-Specific Thresholds for Unequal Priorities

Why Use Different Thresholds Across Arms?

At an interim analysis, a trial may have reasons to treat its experimental arms differently. One treatment may address a more serious unmet need; another may be expensive to develop or deliver. If every arm must meet the same continuation cutoff, the rule does not directly encode those differing priorities. An arm-specific rule can do so by assigning each arm its own prespecified retention threshold.

A different threshold is a decision-policy choice, not a conclusion that one arm has stronger evidence simply because its cutoff is lower. The interim score describes the evidence or estimated benefit under the chosen scoring rule; the threshold describes how much of that score is required for the arm to continue. These two ingredients should be set and interpreted separately.

The score and threshold must be on the same scale. For example, if \(S_i\) is a standardized statistic, \(c_i\) is a cutoff on that standardized scale. If the score is an estimated treatment effect, the cutoff must be expressed in the corresponding effect units. The rule also requires a consistent direction: if larger estimated effects are favorable, scores should be defined so that larger values are favorable for every arm.

Applying the Rule

The selection set is calculated arm by arm. The comparison is inclusive: a score exactly equal to its threshold qualifies. No ranking is required, and the rule does not require a fixed number of retained arms. Depending on the scores and thresholds, the resulting set can contain all arms, some arms, or none.

A useful check is to write the comparison explicitly for every arm, rather than applying a single cutoff mentally. This makes it less likely that a threshold is accidentally assigned to the wrong treatment, especially when thresholds are close together or the number of arms is large.

Choosing the Scale Before the Threshold

Arm-specific thresholds only make sense after specifying what each score measures. If the scores are standardized statistics, a threshold such as \(1.5\) refers to standardized evidence, not to a treatment effect of 1.5 units. If scores are estimated effects, the cutoff refers directly to an effect difference and its units. A numerical comparison across arms is meaningful only when their scores have a common interpretation or have been deliberately transformed to a comparable scale.

A clinical threshold and a standardized-score threshold answer different questions. A minimum effect threshold asks whether the estimated benefit is large enough in clinical units. A standardized threshold asks whether a statistic on a standard-error scale is large enough. They can select different arms, especially when interim standard errors differ. The earlier material on standardized versus unstandardized selection scores is relevant when deciding which type of score the design should use; the arm-specific feature here is that the cutoff can vary by treatment once that scale is fixed.

How Priorities and Costs Enter the Rule

Clinical priority and development cost can motivate different thresholds, but neither supplies a universal formula for them. A design team might require less interim evidence for a treatment addressing a severe unmet need, or more evidence before continuing an arm that would require a costly development program. The direction and size of a threshold difference depend on the stated decision objective, the consequences of false continuation and premature discontinuation, and the trial’s operating characteristics.

The thresholds should therefore be justified as part of the decision policy, rather than chosen after seeing interim scores. A transparent design records, for each arm, the score definition, threshold, rationale, and operational consequence of meeting or missing it. This helps distinguish a principled priority rule from a post hoc adjustment that favors an observed result.

Operating Characteristics and a Short Derivation

An arm-specific rule changes the regions of interim-score space that lead to continuation. To see this precisely, suppose the vector of scores \(S=(S_1,\ldots,S_K)\) has a joint probability distribution under a specified trial scenario. For any particular subset \(J\subseteq\mathcal A\), the event that exactly the arms in \(J\) are retained is $$ \{i\in J:S_i\ge c_i\}\ \cap\ \{i\in\mathcal A\setminus J:S_i<c_i\}. $$ The probability of that selection set is the probability of this joint event. This follows directly from the definition: an arm belongs to \(\mathcal R\) precisely when its own score passes its cutoff, and an arm outside \(\mathcal R\) precisely when it falls below its cutoff.

For example, with three arms, the event of retaining A and C but not B is $$ \{S_A\ge c_A,\ S_B<c_B,\ S_C\ge c_C\}. $$ Thus its probability depends not only on the three marginal score distributions but also on their joint dependence. In a shared-control design, interim scores may be correlated; the earlier material on shared-control correlation provides the relevant dependence framework. Different thresholds change the boundary of the selection event, while the score distribution determines how much probability lies on each side of that boundary.

Thresholds can be evaluated by calculating or simulating selection probabilities under scenarios of interest, such as no treatment benefit, clinically important effects, or unequal effects across arms. These checks are important because a lower cutoff generally makes its arm easier to retain, but the probability of retaining a specific set of multiple arms also depends on score dependence. An arm-specific threshold does not automatically control a familywise error rate or guarantee a desired probability of continuing a promising treatment; those are operating-characteristic claims that require explicit evaluation under the design’s assumptions.

Diagnostics, Alternatives, and Limitations

Before implementing the rule, verify that all scores have the intended favorable direction, each threshold is on the matching scale, and every arm has exactly one applicable cutoff. Then test the calculation with boundary cases: a score equal to its cutoff must pass, and a score just below it must fail. Recomputing the selection set independently from a written table is a simple way to detect coding or transcription errors.

A particularly important diagnostic is to inspect the probability of retaining each arm and of retaining clinically relevant combinations of arms under plausible trial scenarios. If a nominally high-priority treatment is almost never retained, its threshold may not reflect the intended policy. If an expensive arm is retained too often under scenarios where its effect is negligible, the rule may not be sufficiently selective. These findings do not mechanically dictate new cutoffs; they provide information for a deliberate redesign before the rule is used.

The alternatives encode different policies. A common threshold does not represent different arm priorities unless those priorities enter elsewhere. A ranking rule focuses on relative performance and can retain an arm even if all scores are weak, depending on its exact specification. An arm-specific threshold allows each arm to be judged against its own requirement, but its flexibility makes careful prespecification and calibration especially important.

The rule is limited to the score, look, and decision it defines. It does not account automatically for later interim looks, delayed decisions, confirmatory evidence, or uncertain outcome variance. Those features require additional design specifications. Also, a threshold motivated by cost or clinical value is not itself a statistical adjustment for multiplicity and should not be described as such.

Common Mistake