Binding Futility Rules

Why a Binding Futility Decision Changes the Budget

A futility rule can determine whether a trial is allowed to continue, not merely advise the team that continuing is unlikely to be useful. When the rule is binding, a history that triggers stopping cannot lead to rejection of \(H_0\). Its conditional rejection probability is therefore zero. On histories that do continue, the remaining rejection budget is the probability that the original procedure would reject, conditional on the interim data observed.

This budget is history-specific. It is generally not equal to the nominal significance level \(\alpha\), nor is it necessarily the same for every trial that continues. Conditional-error preservation compares a proposed continuation rule with the original rule at the same observed history. The relevant comparison is not just whether the new design has a small average rejection probability.

The subscript \(0\) indicates that the probability is evaluated under the null model. In a composite null, a procedure must control the error for every null parameter value, or use a justified least-favourable value; the calculations below are exact for the specified Gaussian null model. A binding rule can make the original design's actual type I error smaller than its nominal \(\alpha\), because some possible null histories are stopped before rejection can occur.

Deriving the Conditional Error in a Two-Stage Test

Consider a one-sided, two-stage Gaussian test under \(H_0\). Let \(Z_1\) be the standardized interim statistic and \(Z_2\) the standardized contribution from the future stage. Assume that under the null, \(Z_1\) and \(Z_2\) are independent standard normal random variables. Let \(t\), with \(0<t<1\), be the fraction of total information available at interim. The final standardized statistic is

This model makes \(Z\) standard normal under \(H_0\). Suppose the original rejection rule is \(Z>c\), where \(c\) is the one-sided critical value. The binding futility rule stops when \(Z_1\le f\), for a prespecified cutoff \(f\). Thus, rejection requires both continuation and crossing the final critical value.

For an observed interim value \(Z_1=x\), the future increment \(Z_2\) remains standard normal by the independent-increments assumption. On a continuation history \(x>f\), rejection occurs precisely when \(Z_2>(c-\sqrt{t}\,x)/\sqrt{1-t}\). Writing \(\Phi\) for the standard normal cumulative distribution function, the conditional error is therefore

The first line is a consequence of the binding decision: once stopped, the trial does not reject, regardless of what future data might hypothetically have been observed. The second line is the null probability that the future data cross the original final boundary, conditional on the observed interim value. In particular, the continuation budget is determined by \(x\), not just by the fact that \(x\) passed the futility cutoff.

To see the averaging argument, let \(R_{\mathrm{new}}\) be rejection under the modified rule. Conditional on the interim data, its rejection probability is \(P_0(R_{\mathrm{new}}\mid Z_1)\). If this is at most \(C(Z_1)\) for all histories, then taking expectations under the null gives

The final equality is the law of total probability. Because \(C(x)=0\) for \(x\le f\), those histories contribute nothing to the original design's rejection probability. If the original design was calibrated to have size at most \(\alpha\), then a rule bounded by its conditional error also has size at most \(\alpha\); it need not use all of that nominal level.

Worked Calculations

How the Budget Varies and How to Check It

For a continuation history, increasing \(x\) lowers the future-stage threshold \((c-\sqrt{t}\,x)/\sqrt{1-t}\). The original conditional error therefore increases with stronger interim evidence. In the Gaussian model this can be seen by differentiating the continuation expression:

where \(\phi\) is the standard normal density. This does not mean that the trial has a higher nominal \(\alpha\) at a high interim value. It means only that, conditional on that particular history, the original final test would have a higher probability of rejection.

A practical check of a proposed continuation procedure is:

Diagnostics, Pitfalls, and Alternatives

The Gaussian calculation depends on the stated model: the null distribution is correctly specified, stage increments are independent, the information fraction \(t\) is known, and the original final statistic and boundary are the ones used to define \(C(x)\). If stage increments are dependent, the conditional distribution of future data must reflect that dependence; inserting the independent-increments formula would give the wrong budget. Likewise, a changed interim statistic or a change in the model requires recalculating the conditional error from the appropriate conditional distribution.

In a numerical implementation, check that conditional errors are between zero and one, equal zero on binding-stop histories, and generally increase with \(x\) in this one-sided Gaussian setup. Recalculate near the futility cutoff and near histories where the proposed rule's conditional probability approaches its budget. Rounding should not determine whether a rule passes: compare unrounded probabilities or use a suitably conservative margin. For a composite null, also verify control over the null parameter values rather than assuming the displayed simple-null calculation is automatically sufficient.

The main contrast is with a nonbinding futility rule. Under a nonbinding rule, a recommendation to stop may be ignored for the purpose of type I error control; conditional error is based on the original rejection rule even on histories that would have prompted futility. Under a binding rule, the stop is enforced, so rejection on those histories is impossible and their conditional error is zero. The introductory material on nonbinding futility distinguishes these operational choices; for conditional-error calculations, the essential point is to use the rule that actually governs rejection.

• Pitfall: using \(\alpha\) as every history's budget. Nominal \(\alpha\) is an unconditional error bound, whereas \(C(x)\) is conditional on the observed interim history. Use the derived history-specific value.
• Pitfall: ignoring binding stops when averaging. A stopped trial cannot reject. Assigning it the continuation formula overstates the original design's error and describes a different rule.
• Pitfall: conditioning only on the word “continue.” The average error among all continued trials is not the conditional error at a particular \(x\). A proposed rule should be compared with the budget given the conditioning information that defines the adaptation.
• Pitfall: checking one history and declaring the whole rule valid. A proposal can fit the budget at one value and exceed it elsewhere. Verify the inequality throughout the relevant continuation histories.

Binding futility can improve operational clarity by making stopping decisions enforceable, but it may reduce power and make the original design's actual type I error conservative relative to nominal \(\alpha\). Conditional-error preservation does not itself guarantee adequate power, clinical usefulness, or robustness to model misspecification; it addresses the null rejection probability under the assumptions used to calculate the function.