Choosing The Alternative For Beta Spending

Why the Design Alternative Matters

A futility boundary is calibrated under a specified alternative, not under every possible treatment effect at once. The alternative represents an effect for which the design aims to control the probability of stopping for futility according to its beta-spending targets. If that effect is changed, the same lower boundary generally has a different crossing probability; recalibrating it to the same target generally produces a different boundary.

The choice is therefore a design decision with clinical consequences. An alternative that is too large may make a design tolerant of futility stopping when the true effect is smaller but still worthwhile. An alternative that is too small may make the design reluctant to stop even when the effect is negligible. The alternative should be chosen from the clinical objective and scientific context before outcomes are observed.

Model and Calibration Target

Use the canonical normal group-sequential model. At look \(i\), let \(Z_i\) be the standardized treatment-effect statistic, let \(I_i\) be the cumulative information, and let \(\theta\) be the treatment effect per unit of the parameter scale. Assume \(0<I_1<\cdots<I_K\), a known or suitably estimated variance, and the usual independent-increments normal approximation. Under a fixed effect \(\theta\), the joint model is

Thus, at a single look, \(Z_i\sim N(\theta\sqrt{I_i},1)\). These assumptions are the basis for the calculations below; if the statistic or its sampling distribution does not follow this model adequately, the resulting boundaries and stopping probabilities may not be reliable.

Let \(l_i\) denote the lower futility boundary at look \(i\): a sufficiently low \(Z_i\) triggers the planned futility action. A beta-spending design specifies cumulative futility targets \(\beta(t_i)\) at the planned information fractions \(t_i=I_i/I_K\). The calibration uses \(\theta^*\) to choose boundaries that meet those targets under the design’s specified stopping rule. In a multi-look design, the relevant probability is generally the probability of a first futility stop along paths that have not already stopped; the treatment of efficacy crossings and whether the futility rule is binding must match the design being implemented.

This distinction matters: beta spending describes a calibration target under \(\theta^*\), not a universal probability of futility stopping for every effect. Nor is a futility-crossing probability automatically the overall type II error probability. Overall failure to reject also depends on efficacy boundaries, the full sequence of looks, and the stopping rule.

One-Look Derivation

The one-look case shows the role of the alternative directly. Suppose the futility rule is to stop when \(Z\leq l\), and the desired probability of futility stopping under \(\theta^*\) is \(\beta\). Because \(Z\sim N(\theta^*\sqrt I,1)\),

Applying the inverse standard normal distribution function gives the calibrated boundary:

This is a complete one-look calibration under the stated normal model. Holding \(I\) and \(\beta\) fixed, a larger \(\theta^*\) gives a higher numerical boundary: the entire statistic distribution has shifted upward, so the cutoff must also move upward to leave the same probability \(\beta\) below it. Conversely, at a fixed boundary, a larger true effect reduces the futility-crossing probability.

What Happens Under Other Effects?

A calibrated boundary controls its target probability at the selected alternative, subject to the stated model and rule. To understand the consequences of choosing that alternative, also calculate the stopping probability at other effects of clinical interest. For a one-look boundary \(l\), this probability is

This calculation is an operating characteristic at \(\theta\); it does not change the original calibration target. In a group-sequential design, the corresponding calculation must use the joint distribution across looks and account for prior stopping. The same principle applies: evaluate the planned rule at each relevant effect, rather than assuming the beta target holds everywhere.

Choosing an Alternative That Matches the Design Goal

The alternative should reflect the effect for which the design’s futility behavior is intended to be acceptable. In practice, this usually requires clinical and scientific judgment rather than a purely mathematical choice. A useful rationale identifies the smallest effect worth pursuing, an effect expected from prior evidence, or another prospectively justified target. The rationale should make clear which question the design is answering.

For example, suppose a hypothetical trial regards \(\theta=0.12\) as a modest but still worthwhile effect and \(\theta=0.20\) as a larger anticipated effect. With \(I=100\) and \(\beta=0.10\), calibrating at 0.20 gives

At the smaller effect, however, the probability of crossing that boundary is

Thus, the 10% target at \(\theta^*=0.20\) does not mean that the chance of futility stopping is 10% at \(\theta=0.12\); in this example it is about 31.51%. If the design instead calibrates at 0.12, its boundary is

which gives a 10% futility probability at 0.12. These calculations help expose the design trade-off: choosing the larger effect as \(\theta^*\) makes the rule more willing to stop when the effect is smaller. Whether that is acceptable depends on the clinical importance of those smaller effects.

When more than one effect is important, a single alternative cannot make one beta target hold at all of them. Instead, prespecify a primary design alternative for calibration, then assess stopping probabilities and overall operating characteristics at additional effects that matter clinically. These extra evaluations inform the choice; they do not silently redefine the calibration alternative.

Multi-Look Designs and Practical Checks

With multiple looks, each \(Z_i\) is correlated with earlier statistics. Consequently, a marginal calculation such as \(P_{\theta^*}(Z_i\leq l_i)\) is not generally the probability of a first futility stop by look \(i\). A path may already have stopped for efficacy or futility, and its later statistics do not lead to further decisions. Boundary calibration must therefore use the joint normal model and the exact binding or nonbinding rule specified for the design.

The alternative affects this joint calculation through the mean vector \((\theta^*\sqrt{I_1},\ldots,\theta^*\sqrt{I_K})\). For each candidate \(\theta^*\), the design calculation must use that mean vector while preserving the planned information times, spending targets, efficacy rules, and futility interpretation. A valid comparison changes the alternative deliberately and keeps those other design choices fixed.

• Check the effect scale. Confirm that \(\theta^*\sqrt{I_i}\) is dimensionless and uses the same treatment-effect definition and direction as \(Z_i\).
• Check the clinical rationale. State why the selected effect is relevant to a decision to continue or stop.
• Check the stopping rule. Confirm whether efficacy crossings are included in the calibration probability and whether the futility boundary is binding.
• Check other effects. Calculate operating characteristics at clinically meaningful effects besides \(\theta^*\), especially effects smaller than the primary design alternative.

A warning sign is reporting a beta-spending target as though it were the probability of stopping under all plausible treatment effects. Another is recalibrating from an interim estimate without treating that change as a change to the design. The design alternative is prespecified; using unplanned outcome information to choose it can alter the operating characteristics and invalidate the original interpretation.

Comparison and Limitations

The normal-model calculations also have limitations. Approximation error, misspecified information, noncanonical statistics, or a mismatch between the assumed and implemented stopping rule can make calculated stopping probabilities inaccurate. A clinically appropriate alternative cannot compensate for a model or implementation that does not represent the trial. For multi-look designs, verify the computed operating characteristics under the actual planned rules.

Common Mistake

Wrong reasoning: “The design targets beta equal to 0.10, so the futility-stop probability must be 10% whether the true effect is 0.12, 0.15, or 0.20.” This confuses a calibration target at one specified alternative with a probability that is constant over the parameter space.

Correction: State \(\theta^*\), the spending target, and the stopping-rule convention together. Then evaluate the probability under each other effect of interest. In the one-look example, a boundary calibrated to 0.10 at \(\theta^*=0.20\) produced a 31.51% futility probability at \(\theta=0.12\), not 10%.