From Cumulative Spending to an Increment at One Look
In a group sequential design, the overall type I error budget is controlled across a sequence of interim looks. A cumulative spending function describes the amount of that budget intended to have been used by a given information fraction. The increment at look \(k\) is the increase in this cumulative target since look \(k-1\).
The distinction is important: an increment is not automatically the probability that the test statistic crosses its boundary at look \(k\). A boundary crossing at that look may include trials that would already have crossed an earlier boundary. The relevant event for allocating error to look \(k\) is a first crossing there, after no earlier efficacy crossing.
This definition describes the spending schedule; by itself, it does not specify a test statistic boundary. To interpret the increment as an actual probability under \(H_0\), the boundaries must be calibrated so that the probability of at least one efficacy crossing by look \(k\) is \(A(t_k)\), using the joint null distribution of the statistics and the stated stopping rules. That distinction separates a planned budget from a probability achieved by a particular set of boundaries.
Why the Difference Is the Natural Allocation
Let \(E_k\) denote the event that the efficacy boundary is crossed at look \(k\), and define \(C_k\) as the event of a first efficacy crossing at look \(k\). In symbols, \(C_k\) requires a crossing at \(k\) and no crossing at any earlier look. The first-crossing events are disjoint: a trial can stop for efficacy at no more than one first-crossing look.
Write \(R_k\) for the event of at least one efficacy crossing by look \(k\). Then \(R_k=C_1\cup\cdots\cup C_k\). Under \(H_0\), if the boundaries achieve the cumulative targets at the planned looks, \(P(R_k\mid H_0)=A(t_k)\). Subtracting the corresponding result through look \(k-1\) gives the key justification.
Thus, when the cumulative targets are achieved, the increment equals the unconditional null probability of a first crossing at look \(k\). “Unconditional” here means that the probability is calculated over all trials under \(H_0\), including those that crossed earlier; trials that crossed earlier simply are not in \(C_k\).
This is not usually the same as \(P(E_k\mid H_0)\), the marginal probability of crossing the look-\(k\) boundary considered on its own. That marginal event includes trials that crossed earlier as well as trials still at risk of a first crossing. Nor is \(\Delta\alpha_k\) automatically the conditional probability of a crossing at look \(k\) among trials that reached it without an earlier crossing. The conditional probability divides the first-crossing probability by the probability of surviving earlier looks.
Worked Example: Calculating Increments From a Spending Schedule
Suppose a hypothetical one-sided design has \(\alpha=0.025\). Its cumulative spending targets at four planned looks are \(A(t_1)=0.001\), \(A(t_2)=0.006\), \(A(t_3)=0.015\), and \(A(t_4)=0.025\). Calculate the incremental allocation at each look and check the total.
At the first look, \(A(t_0)=A(0)=0\), so $$ \Delta\alpha_1=0.001-0=0.001. $$ At the remaining looks, subtract the preceding cumulative target: $$ \begin{aligned} \Delta\alpha_2&=0.006-0.001=0.005,\\ \Delta\alpha_3&=0.015-0.006=0.009,\\ \Delta\alpha_4&=0.025-0.015=0.010. \end{aligned} $$
The increments sum to \(0.001+0.005+0.009+0.010=0.025\), the planned overall one-sided level. Under boundaries calibrated to achieve these cumulative targets, the increments are the respective null probabilities of a first crossing at each look. They are not four separate tests that may each use its increment as an ordinary standalone significance level without accounting for the sequential design.
Marginal, First-Crossing, and Conditional Probabilities
The probability distinction can be made explicit with a simple null model. Suppose two hypothetical test statistics have a bivariate standard normal distribution under \(H_0\), each with variance 1 and correlation \(0.5\). For illustration, set both efficacy cutoffs to zero. This deliberately simple setup is useful for separating the events; it is not a proposed clinical trial design.
Worked Example: A Look-Specific Tail Probability Is Not a First-Crossing Probability
Let \(Z_1\) and \(Z_2\) have the joint distribution just described. A crossing occurs when the statistic is at least zero. At look 2, the marginal crossing probability is $$ P(Z_2\geq 0\mid H_0)=0.5, $$ because \(Z_2\) has a standard normal distribution.
A first crossing at look 2 requires \(Z_1<0\) and \(Z_2\geq0\). For a bivariate standard normal pair with correlation \(\rho\), the probability of this particular opposite-sign quadrant is $$ P(Z_1<0,Z_2\geq0)=\frac{\arccos(\rho)}{2\pi}. $$ Here \(\rho=0.5\), so $$ P(Z_1<0,Z_2\geq0)=\frac{\arccos(0.5)}{2\pi} =\frac{\pi/3}{2\pi} =\frac{1}{6}\approx0.1667. $$
The marginal look-2 crossing probability is \(0.5\), but only \(0.1667\) is a first crossing at look 2. The difference consists of paths with \(Z_1\geq0\) that already crossed at look 1 and also have \(Z_2\geq0\). In this example, the cumulative crossing probability by look 1 is \(0.5\), and by look 2 it is \(2/3\). Their difference, \(2/3-0.5=1/6\), is exactly the look-2 first-crossing probability. The numerical values illustrate the event distinction, not a clinically useful alpha level.
There is a second distinction: a first-crossing probability is unconditional, whereas a conditional crossing probability is calculated only among paths that have not crossed earlier. If cumulative targets are attained, then \(P(\text{no earlier crossing}\mid H_0)=1-A(t_{k-1})\). Provided this probability is positive, the conditional crossing probability at look \(k\), given survival through look \(k-1\), is $$ P(C_k\mid \text{no earlier crossing},H_0) =\frac{\Delta\alpha_k}{1-A(t_{k-1})}. $$ This is generally different from both \(\Delta\alpha_k\) and the marginal probability of crossing the look-\(k\) boundary.
Worked Example: Converting an Increment to a Conditional Probability
Suppose a hypothetical spending schedule has cumulative targets \(A(t_1)=0.004\) and \(A(t_2)=0.012\). The second-look increment is $$ \Delta\alpha_2=0.012-0.004=0.008. $$ If boundaries achieve those targets, the probability under \(H_0\) of no first-look crossing is \(1-0.004=0.996\). Therefore, among paths that have not crossed at look 1, the conditional probability of a first crossing at look 2 is $$ \frac{0.008}{0.996}=0.008032\text{ (approximately)}. $$
The unconditional first-crossing probability is \(0.008\); the conditional probability among paths reaching look 2 without a crossing is about \(0.008032\). The difference is small here because earlier spending was small, but it need not be small when earlier cumulative spending is larger.
Assumptions, Checks, and Common Misreadings
The probability interpretation relies on a specified null hypothesis, the joint null distribution of the sequential test statistics, the planned look schedule, and the efficacy stopping rule. It also relies on comparing like with like: the cumulative probability \(A(t_k)\) and the difference \(A(t_k)-A(t_{k-1})\) must refer to the same one-sided error criterion and the same design. If the boundaries have not been calibrated to the targets, the differences remain increments of the planned spending function, but need not equal the actual probabilities of first crossing under those boundaries.
A useful diagnostic is to inspect the schedule before interpreting any increments. For a valid nondecreasing cumulative spending schedule, \(A(0)=0\), each increment is nonnegative, and the increments telescope: $$ \sum_{j=1}^{k}\Delta\alpha_j =\sum_{j=1}^{k}\{A(t_j)-A(t_{j-1})\} =A(t_k)-A(0) =A(t_k). $$ At the final planned look, this sum should equal the planned total \(\alpha\) when \(A(t_K)=\alpha\). A negative increment signals a decreasing cumulative target or a calculation/indexing error. A sum that fails to match the cumulative target signals a subtraction or look-labeling error.
A further check is to distinguish the schedule from the achieved operating characteristics. If boundaries are available, the relevant null probability through look \(k\) is the probability of crossing at least one boundary up to that look, accounting for their joint distribution and earlier stopping. A marginal tail probability at look \(k\) alone does not make this check, because it ignores earlier crossing paths. Likewise, using \(\Delta\alpha_k\) as a conditional level among surviving trials is incorrect unless it is converted to the conditional probability using the survival probability.
| Quantity | Meaning under the null | Relationship to the increment |
|---|---|---|
| \(A(t_k)\) | Target probability of at least one crossing by look \(k\) | Cumulative target through look \(k\) |
| \(\Delta\alpha_k\) | Increase in cumulative target since look \(k-1\) | Equals first-crossing probability at \(k\) when targets are achieved |
| \(P(E_k\mid H_0)\) | Marginal probability of crossing the look-\(k\) boundary | Usually not equal; includes paths that crossed earlier |
| \(P(C_k\mid\text{no earlier crossing},H_0)\) | Conditional first-crossing probability among paths still at risk | Equals \(\Delta\alpha_k/[1-A(t_{k-1})]\) when targets are achieved |
A common mistake is to say, “The increment at look \(k\) is the probability of crossing the boundary at look \(k\), so it equals the null tail area beyond that boundary.” The correction is that the increment is defined from cumulative spending. If the cumulative targets are attained, it equals the probability of a first crossing at look \(k\), which excludes prior-crossing paths; it is not generally the marginal tail area or the conditional probability among surviving paths.
The main alternative is to think in terms of nominal per-look significance levels or fixed boundary cutoffs. Those quantities can describe a boundary at an individual look, but do not by themselves express how much cumulative type I error is targeted by that look. Cumulative spending and its increments make that bookkeeping explicit. They do not, alone, determine the boundary values: the joint null distribution and boundary calibration are still needed. The detailed construction of boundaries from a spending schedule is a separate step.
Finally, the simple increment formula presumes that the cumulative targets are evaluated at the stated planned looks and that the error criterion is specified consistently. Additional stopping rules, such as futility stopping, can change which paths remain eligible to cross later. Two-sided designs also require a clear convention for how the total error is represented across the tails. In each case, the event definition and the probability being controlled must be stated before interpreting an increment.
Check Your Understanding
Answer each question, then check the solutions below.
- A schedule has \(A(t_1)=0.002\), \(A(t_2)=0.009\), and \(A(t_3)=0.020\). Find the three increments and verify their sum.
- In one sentence, distinguish a first crossing at look \(k\) from a marginal crossing at look \(k\).
- If \(A(t_{k-1})=0.04\) and \(\Delta\alpha_k=0.006\), what is the conditional crossing probability at look \(k\) among paths that have not crossed earlier, assuming the cumulative targets are achieved?
- A proposed cumulative schedule decreases from \(0.010\) at one look to \(0.008\) at the next. What diagnostic fails, and what should be checked?
Solutions. (1) The increments are \(0.002\), \(0.009-0.002=0.007\), and \(0.020-0.009=0.011\). Their sum is \(0.002+0.007+0.011=0.020\), the final cumulative target. (2) A first crossing at \(k\) requires no earlier crossing, while a marginal crossing at \(k\) includes paths that may already have crossed earlier. (3) The probability is \(0.006/(1-0.04)=0.006/0.96=0.00625\). (4) The schedule is not nondecreasing, so it yields a negative increment; check the target values, look ordering, and subtraction before interpreting the schedule.