Binding Futility In Joint Crossing Probabilities

Why Binding Futility Changes a Crossing Probability

A group-sequential design can stop early for either efficacy or futility. With a binding futility rule, crossing a lower boundary ends the trial, so that path cannot later cross an efficacy boundary. Consequently, the probability of efficacy at a later look is not merely the probability that the later test statistic exceeds its upper boundary. It is the probability of exceeding that boundary and having avoided both earlier efficacy stopping and earlier futility stopping.

This tutorial focuses on that joint event. It assumes a prespecified sequence of analyses and boundaries, with the canonical joint-normal model for the sequential test statistics. The information times are fixed for the probability calculation; changing the timing or adapting the boundaries would require recalculating the joint distribution and events.

Definition: At look \(i\), let \(Z_i\) be the standardized test statistic, \(c_i\) the upper efficacy boundary, and \(l_i\) the lower futility boundary. A binding rule stops for efficacy if \(Z_i\ge c_i\), stops for futility if \(Z_i\le l_i\), and continues only if \(l_i<Z_i<c_i\). A later efficacy-crossing event includes the requirement that the statistic fell in the continuation interval at every earlier look.

For the calculations below, write \(t_i=I_i/I_K\) for the information fraction at look \(i\), where \(I_i\) is the information at that look and \(I_K\) is the maximum information. Under \(H_0\), the canonical model is \(Z_i\sim N(0,1)\), with \(\operatorname{Corr}(Z_i,Z_j)=\sqrt{t_i/t_j}\) when \(i<j\). In particular, for two looks, \(Z_2\mid Z_1=z\sim N(\sqrt{t_1/t_2}\,z,\ 1-t_1/t_2)\). This model captures the dependence caused by using accumulating data; the two statistics are not independent.

Deriving the Two-Look Event

Suppose look 1 has boundaries \(l_1\) and \(c_1\), and look 2 has efficacy boundary \(c_2\). The binding-rule event of efficacy at look 2 is \(l_1<Z_1<c_1\) together with \(Z_2\ge c_2\). Conditioning on \(Z_1=z\) gives the probability as an integral of a conditional tail probability over the permitted continuation region.

P_0(E_2) = P_0(l_1<Z_1<c_1,\ Z_2\ge c_2) = \int_{l_1}^{c_1}\phi(z) \left[ 1-\Phi\left( \frac{c_2-\sqrt{t_1/t_2}\,z}{\sqrt{1-t_1/t_2}} \right) \right]\,dz .

Here \(\phi\) and \(\Phi\) are the standard normal density and cumulative distribution function. The lower integration limit is essential: values \(Z_1\le l_1\) are stopped paths under a binding rule, not eligible paths that might later reach efficacy. The upper limit \(c_1\) excludes paths that already stopped for efficacy at look 1. The conditional tail inside the integral accounts for dependence between looks.

Derivation: The joint density factors as \(f_{Z_1,Z_2}(z_1,z_2)=f_{Z_1}(z_1)f_{Z_2\mid Z_1}(z_2\mid z_1)\). Integrating first over \(z_2\ge c_2\) yields \(P(Z_2\ge c_2\mid Z_1=z_1)\). Integrating that conditional probability against \(\phi(z_1)\) over \(l_1<z_1<c_1\) gives the displayed expression. Thus the lower boundary changes the joint event by removing the earlier-futility region from the integral.

Worked Example: Two-Look Crossing Under Binding Futility

Consider a hypothetical two-look design with \(t_1=0.5\), \(t_2=1\), \(l_1=0.5\), \(c_1=2.4\), and \(c_2=2.0\). Calculate the probability under \(H_0\) of efficacy at look 2 under the binding rule.

The correlation is \(\sqrt{0.5}=0.7071\), and the conditional standard deviation of \(Z_2\) given \(Z_1\) is \(\sqrt{0.5}=0.7071\). Substitution into the joint-event formula gives

P_0(E_2) = \int_{0.5}^{2.4}\phi(z) \left[ 1-\Phi\left(\frac{2.0-0.7071z}{0.7071}\right) \right]\,dz \approx 0.0177 .

The integral can be evaluated by numerical quadrature. It is the probability of a final efficacy crossing among paths that continued after look 1, not the marginal probability of \(Z_2\ge2.0\). The calculation includes both boundaries at look 1: it excludes \(Z_1\le0.5\) for futility and \(Z_1\ge2.4\) for early efficacy.

What Changes When the Lower Boundary Is Omitted?

A useful comparison is to calculate the later crossing probability while excluding early efficacy but not excluding early futility. That event is \(Z_1<c_1,\ Z_2\ge c_2\). It is not the binding-rule event, because it counts paths with \(Z_1\le l_1\) as though they could continue. Its integral has the same conditional-normal form but a different lower limit.

Worked Example: Quantifying the Paths Removed by Futility

Use the hypothetical design from the preceding example. If the lower boundary is omitted, the probability of a look-2 crossing without a look-1 efficacy crossing is

P_0(Z_1<2.4,\ Z_2\ge2.0) = \int_{-\infty}^{2.4}\phi(z) \left[ 1-\Phi\left(\frac{2.0-0.7071z}{0.7071}\right) \right]\,dz \approx 0.0190 .

The binding rule instead gives approximately \(0.0177\). Therefore, the probability removed from this later-crossing event by enforcing \(Z_1>0.5\) is approximately \(0.0190-0.0177=0.0013\), or 0.13 percentage points. The direction is as expected: adding the restriction \(Z_1>0.5\) cannot increase the probability of the event.

For context, the probability of crossing the first-look efficacy boundary under \(H_0\) is \(P_0(Z_1\ge2.4)=1-\Phi(2.4)\approx0.00820\). The overall probability of efficacy at either look under this binding rule is therefore approximately \(0.00820+0.0177=0.0259\); the two events are disjoint because a path that crosses at look 1 does not proceed to look 2. If futility were ignored in this path calculation, the corresponding sum would be approximately \(0.00820+0.0190=0.0272\). These are calculations under the specified boundaries and model, not a general claim about error control for other ways of handling futility.

As a quick check on the numerical results, \(P_0(Z_1<2.4,\ Z_2\ge2.0)\) cannot exceed the marginal probability \(P_0(Z_2\ge2.0)=1-\Phi(2.0)\approx0.02275\). The value \(0.0190\) satisfies this bound. The binding probability must also be no larger than \(0.0190\), since it imposes the additional restriction \(Z_1>0.5\). Such subset checks are valuable safeguards when using numerical integration.

Extending the Calculation to More Looks

With \(K\) looks, a path reaching look \(i\) without stopping must have continued at every earlier look. Define \(E_i\) to be efficacy stopping at look \(i\). For \(i>1\), the event is that \(Z_j\) lies between its boundaries for all \(j<i\), followed by \(Z_i\ge c_i\). The overall probability of efficacy is the sum of the look-specific probabilities because the stopping events are mutually exclusive.

P_0(\text{efficacy}) = P_0(Z_1\ge c_1) + \sum_{i=2}^{K} P_0\left( l_j<Z_j<c_j\ \text{for every }j<i,\ Z_i\ge c_i \right).

Each term is a joint probability under the multivariate canonical normal distribution. It can be evaluated by multivariate-normal integration or recursively by conditioning on the statistics at the preceding look. A lower boundary is applied only to paths that reach its look; it does not retroactively change probabilities of earlier stopping events. If a design has no futility boundary at a particular look, the corresponding continuation restriction must reflect the actual rule rather than introducing a boundary that is not there.

Worked Example: A Three-Look Path and Its Conditional Final Crossing

Consider hypothetical information fractions \(t_1=0.25\), \(t_2=0.50\), and \(t_3=1\). Suppose the first two looks use boundaries \(l_1=l_2=0.5\) and \(c_1=c_2=2.0\), and the final efficacy boundary is \(c_3=2.0\). For a path with \(Z_1=1.0\) and \(Z_2=1.4\), both interim statistics are in their continuation intervals: \(0.5<1.0<2.0\) and \(0.5<1.4<2.0\). Thus this path is eligible for a final crossing under the binding rule.

Under the canonical model, the next increment is independent of the accumulated statistic at look 2, so \(Z_3\mid Z_2=1.4\sim N(\sqrt{0.5}\times1.4,\ 0.5)\). Its conditional mean is \(0.9899\), and its conditional standard deviation is \(0.7071\). Therefore,

P_0(Z_3\ge2.0\mid Z_2=1.4) = 1-\Phi\left(\frac{2.0-0.9899}{0.7071}\right) = 1-\Phi(1.428) \approx0.0767 .

This is the conditional chance of a final crossing for the specified interim values; it is not the unconditional three-look crossing probability. To obtain that probability, integrate over all pairs \((Z_1,Z_2)\) satisfying both continuation restrictions, weighting the conditional final tail by their joint density. Pairs outside either continuation interval contribute zero to final efficacy under the binding rule.

Assumptions, Diagnostics, and Limitations

The calculations are valid when the sequential statistics follow the stated canonical joint-normal model, the information fractions and boundaries are correctly specified, and the binding stopping rule is implemented as defined. In applications, the model is an approximation whose adequacy depends on the test statistic, data structure, and information calculation. Boundary values also need to be expressed on the same statistic scale used in the probability calculation.

Several checks help diagnose mistakes:

Check event limits. For a binding later crossing, earlier continuation requires the lower as well as the upper restriction. An integral beginning at \(-\infty\) has omitted futility stopping.
Check set inclusion. Adding an earlier stopping restriction cannot increase a later-crossing probability. The binding event is a subset of the event that omits futility.
Check marginal bounds. A joint crossing probability cannot exceed the marginal probability of crossing the relevant upper boundary at that look.
Check disjointness. Overall efficacy probability is the sum of look-specific stopping probabilities, not the sum of overlapping events that count a path more than once.
Check the dependence structure. Treating looks as independent ignores the canonical correlation and generally gives the wrong joint probability.

The main alternative calculation is to report a marginal crossing probability at a later look, or a joint probability that omits one or more earlier stopping restrictions. Such quantities can be useful for answering different questions, but they are not the probability of later efficacy under a binding rule. The binding calculation is directly tied to the specified stopping process. Its limitations are equally direct: it depends on the assumed joint distribution, correct information times, and correct boundary implementation. When these inputs are uncertain or the joint model is not well represented by the canonical approximation, simulation under a carefully specified data-generating model can provide a complementary check.

Common Mistake: Counting Stopped Paths as Continuing

A common error is to calculate \(P_0(Z_1<c_1,\ Z_2\ge c_2)\) and call it the binding-rule probability of final efficacy. That calculation removes paths that crossed the first efficacy boundary, but it still includes paths with \(Z_1\le l_1\), which would already have stopped for futility. The correction is to integrate only over \(l_1<Z_1<c_1\), or, with more looks, to require continuation at every preceding look. A second error is to assume the later statistic is independent of the interim statistic; the conditional mean and variance in the canonical model show why that shortcut is not justified.

Key takeaway: Under binding futility, a later efficacy crossing is a joint event restricted to paths that survived every earlier look. The lower boundary changes the event itself, not merely the interpretation of a marginal crossing probability.

Check Your Understanding

Answer each question using the binding-rule event definitions above. Brief solutions follow the questions.

At a two-look design, what range of \(Z_1\) contributes to look-2 efficacy when both an earlier futility and efficacy boundary apply?
For \(t_1=0.4\) and \(t_2=1\), give the conditional mean and variance of \(Z_2\) given \(Z_1=z\) under \(H_0\).
Why must the binding-rule look-2 crossing probability be no greater than the corresponding probability that omits futility stopping?
If the look-2 crossing probability under a binding rule is \(0.012\) and the first-look efficacy probability is \(0.006\), what is the overall probability of efficacy at either look, assuming those are the only efficacy looks?
Explain why \(P(Z_2\ge c_2)\) alone is not the probability of look-2 efficacy under a binding design.

Solutions. (1) Integrate over \(l_1<Z_1<c_1\), while requiring \(Z_2\ge c_2\). (2) The conditional mean is \(\sqrt{0.4}\,z\approx0.6325z\), and the variance is \(1-0.4=0.6\). (3) The binding event adds the restriction \(Z_1>l_1\), so it is a subset of the event that omits futility. (4) The stopping events are disjoint, so the total is \(0.012+0.006=0.018\), or 1.8%. (5) The marginal probability includes paths that crossed efficacy earlier and paths stopped for futility earlier; neither belongs to the event of reaching look 2 and crossing efficacy there.