Recursive Calculation Of Binding Futility Boundaries

Why Binding Futility Requires Recursive Integration

A lower futility boundary is not calibrated from the marginal distribution at a look alone when the rule is binding. A participant whose test statistic crossed an earlier efficacy boundary has already stopped for success; one who crossed an earlier futility boundary has already stopped for futility. Neither path can contribute to the probability of reaching a later look and crossing its futility boundary. The calculation must therefore carry forward only paths that remained active.

The cumulative and incremental beta-spending ideas provide the target probabilities. Here the central task is to calculate those probabilities when successive test statistics are correlated and earlier stopping rules truncate the paths. The resulting boundary equations are recursive: each new equation uses the boundaries already determined at earlier looks.

Definition: Consider analyses at information fractions $0<t_1<\cdots<t_K=1$. Under a specified design alternative, let $Z_i$ be the standardized test statistic at look $i$, and let $\delta$ be the standardized effect at the final analysis. In the canonical model, $$ E(Z_i)=\delta\sqrt{t_i},\qquad \operatorname{Var}(Z_i)=1,\qquad \operatorname{Corr}(Z_i,Z_j)=\sqrt{\frac{t_i}{t_j}}\quad (i<j). $$ At look $i$, let $u_i$ be the upper efficacy boundary and, for a nonfinal look, let $l_i$ be the lower futility boundary. A binding rule stops for futility when $Z_i\le l_i$, and stops for efficacy when $Z_i\ge u_i$. A path continues only when $l_i<Z_i<u_i$.

This canonical model assumes the information fractions are correctly specified, the standardized statistics have the stated joint normal distribution, and the design alternative fixes the drift $\delta$ used for beta spending. The stopping rule also requires nonoverlapping boundaries, $l_i<u_i$. These assumptions make the conditional distribution at a later look available directly from the joint normal model. If the statistic or information process does not support this model, the resulting boundary calculation may not describe the operating characteristics of the actual design.

Write $b_i$ for the planned cumulative futility-spending probability by look $i$, evaluated under the design alternative. The incremental target is $q_i=b_i-b_{i-1}$, with $b_0=0$. This is the probability assigned to a futility stop at look $i$, among paths not stopped earlier. It is not, by itself, the overall type II error probability: paths that continue to the final analysis and do not cross an efficacy boundary also contribute to overall failure to reject.

The Recursive Probability Equation

At the first look, the probability of a futility stop is a marginal normal probability. At later looks, the event must include survival through every preceding look. Define the region of prior continuation as $$ \mathcal{C}_{i-1} = \{(z_1,\ldots,z_{i-1}): l_j<z_j<u_j\text{ for }j=1,\ldots,i-1\}. $$ The incremental spending equation for look $i$ is

q_i = P_{\delta}\!\left( (Z_1,\ldots,Z_{i-1})\in\mathcal{C}_{i-1},\ Z_i\le l_i \right) = \int_{\mathcal{C}_{i-1}}\int_{-\infty}^{l_i} f_i(z_1,\ldots,z_i)\,dz_i\,dz_1\cdots dz_{i-1},

where $f_i$ is the joint density of $(Z_1,\ldots,Z_i)$ under the design alternative. For the first look, the prior-continuation region is empty and the equation reduces to

q_1=P_{\delta}(Z_1\le l_1) =\Phi\!\left(l_1-\delta\sqrt{t_1}\right).

Once $l_1,\ldots,l_{i-1}$ have been found, the right-hand side at look $i$ is a nondecreasing function of $l_i$. Increasing $l_i$ includes additional values of $Z_i$ in the futility region, without changing which paths reached the look. Thus a one-dimensional root-finding method can solve for $l_i$, while the probability at each trial value of $l_i$ is evaluated by integrating over prior continuation.

Justification: The event of a futility stop at look $i$ is the intersection of two events: the trial continued at every earlier look, and $Z_i\le l_i$. The joint density integral above is the probability of exactly that intersection. Adding probability from paths outside $\mathcal{C}_{i-1}$ would count paths that the binding stopping rule has already removed.

A useful computational form conditions on the statistics at earlier looks. Given a surviving history $(Z_1,\ldots,Z_{i-1})=(z_1,\ldots,z_{i-1})$, the canonical joint normal model gives a normal conditional distribution for $Z_i$. The inner integral can therefore be evaluated as a conditional normal cumulative probability, and the outer integral averages that probability over surviving histories. For two looks, for example,

Z_2\mid Z_1=z \sim N\!\left( \delta\sqrt{t_2} +\sqrt{\frac{t_1}{t_2}}\left(z-\delta\sqrt{t_1}\right), \ 1-\frac{t_1}{t_2} \right).

The second argument of $N$ here is the variance. This conditional expression is often more convenient for numerical quadrature than integrating the full bivariate density directly.

Worked Example: Calibrating a First-Look Boundary

Suppose a hypothetical design has $t_1=0.50$, design alternative $\delta=2.00$, and cumulative futility spending $b_1=0.0278$. Find the first lower boundary. Assume the proposed efficacy boundary is $u_1=2.50$.

At the first look, $Z_1\sim N(2.00\sqrt{0.50},1)=N(1.414,1)$. Since there is no earlier stopping, the entire incremental target $q_1=b_1=0.0278$ is a marginal lower-tail probability: $$ 0.0278=\Phi(l_1-1.414). $$ The standard normal quantile for 0.0278 is approximately $-1.914$, so $$ l_1=1.414-1.914=-0.500. $$ Thus the futility rule is $Z_1\le-0.500$, which has probability approximately 0.0278 under this alternative. The boundaries are ordered correctly because $-0.500<2.50$. For context, the upper-boundary crossing probability is $P(Z_1\ge2.50)=1-\Phi(1.086)\approx0.139$; this efficacy event is distinct from the lower-tail spending event.

A Two-Look Recursive Calculation

At a second look, the lower boundary is not obtained by applying a normal quantile to the incremental target alone. The earlier continuation interval determines which first-look values may be integrated. The next example shows how a candidate second boundary implies an incremental probability, which can then be used as the target in the recursive equation.

Worked Example: Integrating Over Surviving Paths

Continue the hypothetical design above with $t_2=1$, $l_1=-0.500$, and $u_1=2.50$. Suppose the proposed second-look futility boundary is $l_2=0.200$. Calculate the incremental futility probability at look 2 under $\delta=2.00$, and hence the cumulative futility spending through look 2.

The first-look statistic has mean $1.414$, and the correlation between $Z_1$ and $Z_2$ is $\sqrt{0.50}=0.7071$. The conditional distribution is $$ Z_2\mid Z_1=z\sim N(1.000+0.7071z,\ 0.500). $$ Therefore, the conditional probability of crossing the second lower boundary is $$ P(Z_2\le0.200\mid Z_1=z) = \Phi\!\left(\frac{0.200-(1.000+0.7071z)}{0.7071}\right) = \Phi(-1.131-z). $$ Only paths with $-0.500<Z_1<2.50$ reach look 2. The incremental probability is consequently $$ q_2=\int_{-0.500}^{2.50} \phi(z-1.414)\Phi(-1.131-z)\,dz. $$ Numerical quadrature gives $q_2\approx0.0247$. The cumulative futility-crossing probability through look 2 is therefore $b_2=b_1+q_2\approx0.0278+0.0247=0.0525$. In a calibration problem, if the planned cumulative target is 0.0525, the recursive equation is solved for $l_2$; this calculation shows that $l_2\approx0.200$ meets that target. The final analysis has no futility boundary in this example.

The integration limits matter. Integrating from $-\infty$ instead of $l_1$ would include first-look paths already stopped for futility. Integrating beyond $u_1$ would include paths already stopped for efficacy. Neither set of paths reaches the second analysis under the binding rule.

Worked Example: Handling a Zero-Spending Look

Consider a hypothetical first look with $t_1=0.50$, $\delta=2.00$, and efficacy boundary $u_1=2.50$, but planned futility spending $b_1=0$. What is the lower-boundary convention, and what probability of paths remains eligible to continue?

A zero lower-tail target requires $P(Z_1\le l_1)=0$, so the boundary is $l_1=-\infty$, interpreted as no futility stop at that look. The efficacy crossing probability is still $$ P(Z_1\ge2.50)=1-\Phi(2.50-1.414) =1-\Phi(1.086)\approx0.139. $$ Thus the probability reaching the next look is $1-0.139=0.861$. The later integration must include all first-look values below 2.50, not all real values: the upper efficacy crossing still removes paths, even though the futility-spending target is zero. This distinction is essential when a spending schedule includes a look with no planned futility crossing.

Computational Workflow and Diagnostics

In practice, the recursion separates the calculation into a boundary-solving step and a probability-evaluation step. The probability evaluation may use multivariate normal integration, conditional normal quadrature, or a validated numerical routine. Whatever method is used, it must apply the same continuation region at every previously completed look.

Specify the design model.
Record the information fractions, design alternative, efficacy boundaries, and cumulative futility-spending targets. Check that the information fractions increase and end at 1.

Convert cumulative targets to increments.
Compute $q_i=b_i-b_{i-1}$. A negative increment signals that the proposed cumulative schedule decreases and cannot represent nonnegative incremental crossing probabilities.

Start at the first look.
Use the marginal normal lower-tail equation to find $l_1$, or set $l_1=-\infty$ for zero spending. Check that the lower and upper boundaries do not overlap.

Integrate over prior continuation.
At look $i$, evaluate the joint probability of prior continuation and $Z_i\le l_i$, using already calibrated boundaries $l_j$ and specified efficacy boundaries $u_j$ for all $j<i$.

Solve and verify numerically.
Find $l_i$ that makes the probability equal to $q_i$. Re-evaluate the probability at the reported boundary and check the residual, monotonicity, and the ordering $l_i<u_i$.

A useful feasibility check is that the requested increment cannot exceed the probability mass available among paths that reach the look and remain below its efficacy boundary. If the target is too large, no lower boundary satisfying $l_i<u_i$ can produce it. A numerical solver that returns a boundary at or above $u_i$, or fails to bracket a root within the allowed range, is a warning to inspect the spending schedule and prior continuation probabilities rather than to accept the returned value.

Other warning signs include a computed incremental probability that decreases when the trial lower boundary is raised, an accumulated probability that does not equal the sum of increments, and substantial disagreement between independent integration methods. These can indicate reversed inequality limits, an incorrect conditional variance, or accidental inclusion of paths stopped at earlier looks. Numerical quadrature should also be sufficiently accurate near boundaries where the integrand changes rapidly.

Interpretation, Alternatives, and Limitations

The recursive method is the direct calculation for a binding rule: earlier decisions remove paths from later analyses. A marginal calculation instead uses $P_{\delta}(Z_i\le l_i)$ without conditioning on prior continuation. That can be useful for describing a marginal statistic, but it does not calibrate the incremental probability of a binding futility stop at look $i$. A nonbinding rule is another distinct design: futility crossings may be recorded but do not remove paths from subsequent efficacy analyses. Its path probabilities therefore use a different continuation rule.

The recursion also relies on a specified spending target and design alternative. Changing $\delta$, information fractions, or earlier boundaries changes the probability of reaching later looks and can change the calibrated lower boundaries. The calculation does not establish, by itself, the overall type II error or the complete operating characteristics of the design; those require accounting for all paths that fail to reject, including paths that continue to the final analysis. The interpretation of beta spending as futility-crossing probability should therefore be kept separate from overall beta.

Key takeaway: For a binding futility design, calibrate each lower boundary using the probability of crossing it among paths that survived every earlier efficacy and futility decision. The first boundary is marginal; each later boundary is recursive.

Check Your Understanding

Answer each question using the canonical model and the binding stopping rule.

Why does the second-look futility probability integrate only over values of $Z_1$ between $l_1$ and $u_1$?
If $b_1=0.03$ and $b_2=0.05$, what is the incremental target $q_2$?
At a first look, $t_1=0.25$, $\delta=2$, and the futility target is 0.05. Find $l_1$ approximately.
What does a negative incremental target indicate?
Why is cumulative futility spending not generally equal to overall type II error?

Solutions. (1) Values at or below $l_1$ have already stopped for futility, and values at or above $u_1$ have already stopped for efficacy; neither reaches look 2. (2) $q_2=b_2-b_1=0.05-0.03=0.02$. (3) The first-look mean is $2\sqrt{0.25}=1$. Thus $0.05=\Phi(l_1-1)$, so $l_1=1-1.6449\approx-0.645$. (4) The cumulative target has decreased, which is incompatible with nonnegative incremental crossing probability. (5) Overall type II error also includes paths that continue through the futility looks but do not cross an efficacy boundary at the final analysis.