Joint Futility and Continuation Screening

Why Use Two Cutoffs?

At an interim analysis, a treatment arm can look unpromising without being conclusively futile, or promising without yet meeting a threshold for strong evidence. A single cutoff forces these different situations into the same category. A joint futility and continuation rule makes the distinction explicit: it identifies arms to drop, arms with sufficiently strong interim evidence to continue automatically, and arms whose results lie between those boundaries.

The intermediate region is not a statistical result that can be left undefined. The design must state what happens there. In the rule developed here, an arm in that region is retained provisionally for the next planned stage, but is not labelled evidence-qualified. This default avoids dropping an arm solely for failing to reach the stronger evidence cutoff. A protocol may attach a prespecified review to provisional retention, but it must state the review and its consequence in advance.

The score and cutoffs must use the same scale. For example, \(S_j\) might be an estimated treatment-versus-control contrast in a specified unit, or a standardized interim statistic. The numerical cutoffs are not interchangeable across these scales. The score must also have a consistent direction: if positive values favor treatment, both cutoffs must be interpreted on that orientation. The earlier material on continuation thresholds for strong interim evidence motivates the upper cutoff; the lower cutoff adds a distinct futility decision.

Operating Probabilities Under a Normal Score Model

For a simple calculation, suppose that under a specified treatment effect the interim score for one arm satisfies \(S_j\sim N(\delta_j,1)\), where \(\delta_j\) is its mean on the standardized score scale. This model assumes that the score is approximately normal, that its standard deviation is one on this scale, and that the mean and variance used for the calculation are appropriate at the planned interim look. It is an operating-characteristic model, not a claim that every realized score must be normal.

Let \(\Phi\) denote the standard normal cumulative distribution function. The three outcomes follow directly from the rule:

These probabilities sum to one because the three intervals partition the score scale. Under the stated default, the probability of any continuation is \(1-\Phi(f-\delta_j)\): the upper cutoff separates the two kinds of continuation, while the lower cutoff determines whether the arm is dropped. That distinction matters when reporting design performance. The probability of evidence-qualified continuation is not the same as the probability of retaining an arm.

Choosing and Checking the Intermediate-Zone Policy

The cutoffs and the consequence for the middle region should be specified before examining interim results. A useful design statement names the score, its direction, both cutoffs, the equality conventions, and the action for each region. If provisional continuation requires a review or another prespecified condition, that condition must be stated as well. Otherwise, two analysts could classify the same score identically but implement different trial decisions.

The ordering \(f<c\) is essential. If \(f=c\), there is no intermediate region; if \(f>c\), some scores would satisfy both the drop and continuation descriptions unless a separate precedence rule were added. Neither situation implements the three-region rule defined above. The cutoffs also need to be on a scale where a score comparison has a clear interpretation. Applying a standardized-score cutoff directly to an unstandardized effect estimate, for example, can produce meaningless classifications.

The following derivation gives a simple way to calculate the expected numbers of arms in each category. Let \(I_{Dj}\) equal 1 when arm \(j\) is dropped and 0 otherwise, and define \(I_{Pj}\) and \(I_{Ej}\) analogously for provisional and evidence-qualified continuation. The total number dropped is \(D=\sum_j I_{Dj}\). By linearity of expectation,

This identity does not require the arm scores to be independent. The same argument gives \(E(P)=\sum_jP(f<S_j<c)\) and \(E(E)=\sum_jP(S_j\ge c)\), where \(P\) is the number provisionally continued and \(E\) the number evidence-qualified. However, dependence does matter for probabilities involving several arms at once, such as the probability that all arms are dropped or that at least two are evidence-qualified. Shared-control contrasts can be correlated, so multiplying marginal probabilities as if arms were independent is generally not justified.

Diagnostics, Alternatives, and Limitations

Before using the rule, check the design and analysis for the following failure modes:

• Reversed score direction. If larger values actually indicate less benefit, applying the rule as written reverses the intended decisions. Confirm the contrast direction using a concrete example.
• Cutoffs on inconsistent scales. A cutoff for standardized scores cannot be applied to effects in clinical units without a valid conversion. Record the score definition, units, and standardization in the protocol.
• An undefined gray-zone action. “Between the thresholds” is not an implementable decision. State whether such arms are provisionally retained and any conditions attached to that retention.
• Uncalibrated operating characteristics. The normal-score formulas depend on the assumed interim distributions. Check those assumptions and evaluate the rule under plausible treatment effects and interim sample sizes.
• Ignoring shared-control dependence. Marginal probabilities and expected counts can be calculated arm by arm, but joint probabilities cannot generally be obtained by treating shared-control scores as independent.

A single-threshold futility rule answers whether an arm is dropped below one boundary, but does not distinguish strong evidence for continuation from the intermediate range. A single continuation threshold identifies evidence-qualified scores but, by itself, does not say what happens to scores below it. The joint rule supplies both distinctions and makes the intermediate action explicit. A policy that simply continues every arm avoids interim dropping but does not use the interim score to screen for futility.

The two cutoffs do not, by themselves, establish efficacy, guarantee a clinically important effect, or control the trial’s overall error rates. They are selection and continuation rules based on interim information. Their operating characteristics should be assessed for the intended design, including the dependence induced by shared controls. Also, provisional continuation can consume additional participants or resources; its practical cost should be considered when setting the cutoffs and defining the review process.