What a Zero-Spending Look Means
A monitoring schedule may include a planned analysis at which the futility-spending function allocates no additional probability to stopping for futility. The analysis can still be operationally useful: the data may be reviewed, recorded, or used for decisions governed by other rules. But if the futility rule is to add no chance of stopping at this look, its futility boundary must not exclude any continuing trial history.
The key distinction is between cumulative and incremental spending. If the cumulative target is unchanged from the preceding look, the increment is zero. That does not mean the boundary should be set to the preceding numerical boundary. At a new look with more information, the test statistic contains new random information, so reusing a finite boundary can create new futility stops.
For a one-sided lower futility rule, the trial stops for futility at look \(i\) if it has continued to that look and \(Z_i\leq l_i\), where \(l_i\) is the lower boundary. To state the event precisely, let \(R_{i-1}\) mean that the trial is eligible to make a futility decision at look \(i\): it has not already stopped, and it has not been removed from consideration by any competing rule as defined in the design. Then the first-crossing event is \(R_{i-1}\cap\{Z_i\leq l_i\}\). Zero incremental spending requires this event to have probability zero under the design alternative.
Why a Distinct Look Has No Finite Boundary
Use the canonical group sequential normal model: under a fixed standardized effect \(\theta\), the successive statistics have means \(E_\theta(Z_i)=\theta\sqrt{t_i}\), variances \(1\), and correlations \(\operatorname{Corr}_\theta(Z_i,Z_j)=\sqrt{t_i/t_j}\) for \(i<j\). Thus, when \(t_{i-1}<t_i\), the conditional distribution of \(Z_i\) given the immediately preceding statistic is normal with positive variance.
In particular, for a distinct-information look, the conditional variance given \(Z_{i-1}\) is \(1-t_{i-1}/t_i>0\). A normal distribution with positive variance assigns positive probability to every interval extending to negative infinity. Therefore, for every finite \(l_i\), each continuing history with finite \(Z_{i-1}\) has a positive conditional probability that \(Z_i\leq l_i\). If continuation has positive probability, averaging over continuing histories gives a positive probability of a new futility crossing. This contradicts a zero incremental target.
The appropriate boundary convention is therefore \(l_i=-\infty\), which makes the futility event empty. This is a statement about the futility decision at that look, not a statement that the data are unavailable or that no analysis can occur. If the design has additional stopping rules, those rules remain in force.
Worked Examples
Worked Example: Reading a Spending Schedule
A hypothetical design has three looks at information fractions \(t_1=0.30\), \(t_2=0.60\), and \(t_3=1.00\). Its cumulative futility-spending targets under the design alternative are \(0.02\), \(0.02\), and \(0.05\). Find the incremental targets and specify the second-look futility boundary.
By definition, subtract each preceding cumulative target:
The second look has more information than the first, since \(0.60>0.30\). Under the canonical model its conditional variance is \(1-0.30/0.60=0.50\), which is positive. Consequently, any finite lower boundary at look 2 would allow some histories that continued past look 1 to stop for futility at look 2. The zero-increment specification is \(l_2=-\infty\).
The first look is calibrated to contribute \(0.02\) of first-crossing probability, and the third to contribute \(0.03\), under the design's stated calibration rule. The second look contributes none. The schedule's final cumulative target is \(0.05\), not \(0.02+0+0.05\): the \(0.05\) is already cumulative.
Worked Example: A Finite Boundary Creates New Stops
For a simple demonstration of the geometry, suppose a hypothetical two-look design has \(t_1=0.50\), \(t_2=1.00\), and zero drift \(\theta=0\). This is a probability illustration, not a recommendation to use zero drift as a design alternative. Suppose the first-look lower boundary is \(l_1=0\). Calculate the probability of a new futility stop at look 2 if a finite boundary \(l_2=0\) is used.
The two statistics are standard normal with correlation \(\rho=\sqrt{0.50/1.00}=\sqrt{0.5}\). A new stop at look 2 occurs when \(Z_1>0\) and \(Z_2\leq0\). For two standard normal variables with correlation \(\rho\), the probability of this opposite-sign quadrant is \(1/4-\arcsin(\rho)/(2\pi)\). Substitution gives:
Thus this finite boundary creates a 12.5% chance of a first futility stop at look 2 in the illustrative model. It cannot represent zero incremental spending. The positive probability is not an artefact of the zero-drift choice: with a nondegenerate conditional normal distribution and a finite lower boundary, the conditional crossing probability remains positive under any finite design drift.
Worked Example: A Repeated Look With No New Information
A hypothetical monitoring system calculates \(Z_1\) at one information time and, at a later scheduled review, repeats exactly the same calculation without collecting additional information. Thus \(Z_2=Z_1\), not a new noisy statistic. Suppose the first-look futility boundary is \(l_1=-1\), and the repeated look also uses \(l_2=-1\). Under a standard normal reference distribution, the probability of stopping at the first look is:
A trial reaching the repeated look must have \(Z_1>-1\). The probability of a new stop at look 2 is therefore:
Here a finite repeated boundary adds no chance of futility stopping because the statistic is identical and all values at or below \(-1\) were already stopped at look 1. This degenerate case differs from an ordinary new-information look, where the conditional variance is positive. In practice, a repeated review can be documented as having no new futility decision; it should not be treated as evidence that a finite boundary is safe at a distinct-information look.
Implementation, Diagnostics, and Alternatives
A zero-spending look should be represented in the monitoring plan in a way that prevents a finite lower limit from being mistaken for an active futility rule. For a distinct-information look, record that futility stopping is disabled there, equivalently \(l_i=-\infty\). If a software table cannot display negative infinity, use an explicit status such as “no futility boundary” and ensure the implementation treats it as no futility stopping region.
Useful checks include:
- Compute each increment as the difference between consecutive cumulative targets. Check that a zero increment is not mistakenly replaced by the cumulative target itself.
- Confirm that information fractions increase strictly for looks treated as distinct. If they do, verify that the conditional variance is positive.
- Check that the event used to calibrate spending is the same event the design intends to control, including the treatment of earlier futility and any competing stopping rules.
- Inspect software output for finite lower boundaries at zero-increment looks. A finite value may indicate a small positive target, numerical truncation, or an implementation error; investigate rather than interpreting it automatically as zero spending.
Two common alternatives must be distinguished. Omitting the futility decision at a look leaves the futility stopping set unchanged and is the usual choice for a zero-increment look. Assigning a small positive increment permits a finite boundary, but it is a different design: some additional histories can stop for futility. A third possibility, reusing the preceding finite boundary, is not generally equivalent to omitting the decision. With new information, the statistic can move below that boundary among trials that previously continued.
A zero-spending futility rule also does not eliminate other reasons to review or stop a trial. Efficacy, safety, data-quality, or operational decisions are separate and must follow their own prespecified rules. The futility-spending statement alone says nothing about those decisions or about the overall type I error rate.
Check Your Understanding
Answer each question using the distinction between cumulative targets, incremental targets, and first-crossing events.
- Cumulative targets at three looks are \(0.01\), \(0.04\), and \(0.04\). What is the incremental target at the third look, and what futility boundary convention is appropriate if it has distinct new information?
- At a distinct-information look, why does a finite lower boundary generally imply positive incremental futility-crossing probability?
- A team uses a finite boundary at a zero-increment look, arguing that the cumulative target remains unchanged. What is the error in this reasoning?
- Can a finite boundary at a repeated, identical-statistic look add no new futility stops? State the condition that makes this possible.
Solutions. (1) The increment is \(0.04-0.04=0\), so use \(l_3=-\infty\), or explicitly record that no futility boundary applies. (2) The conditional statistic has positive variance, so for every finite lower boundary its conditional normal distribution assigns positive probability to values below that boundary; averaging over continuing histories gives positive crossing probability. (3) An unchanged cumulative target means the increment is zero, not that a previous finite boundary can be carried forward: new information can move a continuing trial below that boundary. (4) Yes. If the statistic is exactly repeated and the same boundary was already applied, every value that meets the repeated boundary was already stopped, so no continuing value can newly cross it.