Cumulative And Incremental Beta Spending

Why Cumulative and Incremental Spending Are Different

A futility spending schedule describes how a planned amount of probability is allocated over the interim looks of a group sequential design. The schedule is usually written cumulatively: it states the amount spent by each look, rather than the amount assigned only to that look. To see what is allocated between successive looks, take differences between cumulative values.

The distinction matters because the interim looks are sequential. A participant or trial that has already stopped for futility cannot stop for futility again at a later look. Thus, the probability newly spent at look $k$ concerns trials that reach that look without an earlier futility stop. It is a first-crossing probability, not simply the probability of being below a lower boundary at look $k$ considered on its own.

Definition: Let $K$ be the number of planned looks, and let $\beta_k$ be the cumulative beta-spending target through look $k$, evaluated under a prespecified design alternative. Set $\beta_0=0$. The incremental beta allocation at look $k$ is $$ \Delta\beta_k=\beta_k-\beta_{k-1},\qquad k=1,\ldots,K. $$ In a design that matches its spending targets, $\beta_k$ is the probability of having stopped for futility by look $k$ under that alternative, and $\Delta\beta_k$ is the probability of first stopping for futility at look $k$.

The design alternative is part of the definition: a probability of futility crossing depends on the distribution of the test statistics, and that distribution depends on the assumed treatment effect and other model features. The schedule is therefore not an alternative-free description of beta.

Conditions for a Coherent Spending Schedule

Before interpreting or using a schedule, check its basic structure. The planned cumulative targets should begin at zero, should not decrease, and should not exceed the total beta budget assigned to the futility-spending plan. If the design assigns a total budget $\beta_{\mathrm{budget}}$ through the final look, then a schedule that uses the entire budget has $\beta_K=\beta_{\mathrm{budget}}$. A plan that intentionally leaves some budget unspent instead has $\beta_K<\beta_{\mathrm{budget}}$.

A useful set of assumptions for interpreting the increments is:

The number and ordering of looks are specified, with $\beta_k$ referring to the $k$th look.
The cumulative targets are chosen before examining the interim results and are nondecreasing.
The probabilities are defined under one specified design alternative and a specified joint model for the sequential test statistics.
Stopping for futility is defined as stopping at the first look where the relevant futility rule is met.
The boundary-calculation procedure is designed to match the cumulative targets under the chosen model.

These conditions explain why simple subtraction is always the arithmetic for an increment, but matching the increment to an actual probability requires more than subtraction. The boundaries and the joint distribution of the statistics must make the first-crossing probabilities equal the planned increments. Determining those boundaries is a separate calculation.

Worked Example: Differencing a Three-Look Schedule

Suppose a hypothetical design uses three looks and specifies cumulative futility-spending targets of $0.04$, $0.12$, and $0.20$ under its design alternative. Find the incremental allocation at each look.

Set $\beta_0=0$. At look 1, $$ \Delta\beta_1=\beta_1-\beta_0=0.04-0=0.04. $$ At look 2, $$ \Delta\beta_2=\beta_2-\beta_1=0.12-0.04=0.08. $$ At look 3, $$ \Delta\beta_3=\beta_3-\beta_2=0.20-0.12=0.08. $$

The increments are $0.04$, $0.08$, and $0.08$, or 4, 8, and 8 percentage points. They sum to the final cumulative target: $$ 0.04+0.08+0.08=0.20=\beta_3. $$ Under a design calibrated to match these targets, the corresponding interpretation is a 4% probability of first futility stopping at look 1, an 8% probability of first stopping at look 2, and an 8% probability of first stopping at look 3, each under the specified design alternative.

Why the Increments Are First-Crossing Probabilities

Let $F_k$ denote the event that the trial has stopped for futility by look $k$, and let $F_0$ be the empty event. Because a trial that has stopped does not continue to later looks, these events are nested: $F_1\subseteq F_2\subseteq\cdots\subseteq F_K$. Let $A_k$ denote the event that the first futility stop occurs at look $k$. Then $A_k=F_k\setminus F_{k-1}$. Since $F_{k-1}$ is contained in $F_k$, the probability of this set difference is the difference in the event probabilities:

P(A_k\mid \text{design alternative}) = P(F_k\mid \text{design alternative}) - P(F_{k-1}\mid \text{design alternative}).

If the cumulative targets match the cumulative crossing probabilities, this becomes $\Delta\beta_k=\beta_k-\beta_{k-1}$. The first-crossing events $A_1,\ldots,A_K$ are mutually exclusive: a trial can first stop at only one look. Their probabilities therefore add to the probability of having stopped for futility by the final look:

P(F_K\mid \text{design alternative}) = \sum_{k=1}^{K}P(A_k\mid \text{design alternative}) = \sum_{k=1}^{K}\Delta\beta_k = \beta_K.

This identity also gives a direct check on arithmetic: the sum of all increments through look $k$ must equal the cumulative target $\beta_k$. It does not say that the event of being below the look-$k$ boundary, without regard to earlier stopping, has probability $\Delta\beta_k$. Earlier crossing and the dependence among sequential statistics matter.

Worked Example: Interpreting Probability Allocations Along the Path

Consider a hypothetical two-look design under its design alternative. Suppose its calibrated cumulative futility-crossing probabilities are $0.03$ by look 1 and $0.11$ by look 2. The probability of being below the look-2 futility boundary, evaluated among all hypothetical trials whether or not they crossed earlier, is $0.14$. Find the incremental allocations and explain the difference.

At the first look, the incremental allocation is $$ \Delta\beta_1=0.03-0=0.03. $$ At the second look, it is $$ \Delta\beta_2=0.11-0.03=0.08. $$ Thus, the probability of first stopping for futility at look 2 is $0.08$, not $0.14$. The $0.14$ is a marginal probability of being below the second-look boundary, counting trials without restricting to those that reached look 2 without an earlier futility stop. It is not the second-look first-crossing probability.

The cumulative total is consistent: $$ \Delta\beta_1+\Delta\beta_2=0.03+0.08=0.11=\beta_2. $$ The figures are hypothetical and illustrate the distinction; the values of the boundaries and probabilities would have to be jointly calibrated for a real design.

Checking a Schedule and Its Total Budget

Cumulative values and increments provide two views of the same allocation. Cumulative values answer, “How much has been allocated by this look?” Increments answer, “How much additional probability is allocated at this look?” Both views are useful for checking the plan, but they should not be confused.

Worked Example: Finding and Correcting an Invalid Schedule

A hypothetical four-look plan reports cumulative targets $0.02$, $0.07$, $0.06$, and $0.12$, with a total futility budget of $0.12$. Check the schedule and compute the apparent increments.

Taking differences gives $$ \Delta\beta_1=0.02-0=0.02,\quad \Delta\beta_2=0.07-0.02=0.05, $$ $$ \Delta\beta_3=0.06-0.07=-0.01,\quad \Delta\beta_4=0.12-0.06=0.06. $$ The negative third increment is not a valid probability allocation. The problem is that the cumulative target decreases from $0.07$ to $0.06$; probabilities of having stopped by successive looks cannot decrease because $F_2\subseteq F_3$.

One possible correction, if consistent with the design’s intended allocation, is to change the third cumulative target to $0.09$. The increments then become $$ 0.02,\quad 0.07-0.02=0.05,\quad 0.09-0.07=0.02,\quad 0.12-0.09=0.03. $$ They are all nonnegative and sum to $0.12$. This repair is an arithmetic illustration, not a recommendation to alter a schedule after seeing interim data. A valid design schedule should be specified in advance.

What Beta Spending Does—and Does Not—Mean

A futility-spending schedule allocates the probability of futility stopping over looks under a specified alternative. It is not automatically the overall type II error probability of the entire trial. Overall type II error concerns failing to reject $H_0$ under the alternative; that outcome can include paths that never stop for futility and do not meet the efficacy criterion by the end. The relation between the futility-crossing probability and overall power depends on the full stopping design.

This distinction is important when reporting a schedule. A statement such as “the final cumulative futility spending is $0.20$” means that the planned probability of a futility stop by the final look is $0.20$ under the chosen design alternative, if the procedure matches its targets. It does not, by itself, mean that the trial’s overall type II error is $0.20$, or that its power is $0.80$.

The increments also are not independent probabilities. They are probabilities of mutually exclusive first-stopping events under one joint sequential model. Changing the design alternative, the joint model, the timing of looks, or the stopping rule can change the probabilities associated with the boundaries. The numerical differences in a fixed spending schedule remain arithmetic differences, but whether actual crossing probabilities match them depends on the design calculation.

Common Mistakes and Diagnostics

Using the cumulative amount as the current-look allocation. If $\beta_2=0.12$ and $\beta_1=0.04$, the second-look increment is $0.12-0.04=0.08$, not $0.12$.
Using a marginal boundary probability as the increment. The increment refers to a first crossing at that look, among paths not already stopped for futility. A marginal probability at one look does not make that adjustment.
Allowing cumulative spending to decrease. A negative increment indicates an invalid schedule or a data-entry/indexing error; it is not a negative probability.
Calling futility spending the overall beta error without qualification. The schedule concerns futility stops under a specified alternative, while overall type II error depends on all ways the design can fail to reject.
Ignoring the design alternative. If the alternative used to evaluate crossing probabilities is not the one for which the plan was calibrated, the target interpretation need not hold.
Misaligning look indices. A value assigned to look $k$ must be differenced against the cumulative value at look $k-1$, not against a value associated with a different information time.

A practical diagnostic is to compute every increment, check that none is negative, and verify that the running sums reproduce the original cumulative targets. Then confirm that the final cumulative target agrees with the stated futility budget or with the explicitly stated amount left unspent. These checks catch schedule arithmetic errors; they do not substitute for checking that the sequential boundary calculation matches the intended probabilities.

Key takeaway: Cumulative beta spending through look $k$ is the total planned futility-crossing probability by that look under a specified design alternative. Subtracting consecutive cumulative values gives the probability allocated to a first futility crossing at the current look. The increments sum to the final cumulative target, but they are not generally the marginal probabilities of falling below each look’s boundary.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A three-look schedule has cumulative targets $0.025$, $0.090$, and $0.160$. What is the incremental allocation at each look?
Why is the probability of first futility stopping at look 3 not generally the same as the marginal probability of being below the look-3 futility boundary?
A proposed cumulative schedule is $0.03$, $0.08$, $0.08$, and $0.15$. Find its increments and state whether it is structurally valid.
Does a final cumulative futility-spending target of $0.18$ by itself establish that overall power is $0.82$? Explain.

Solutions. 1. With $\beta_0=0$, the increments are $0.025$, $0.090-0.025=0.065$, and $0.160-0.090=0.070$. They sum to $0.160$. 2. First crossing at look 3 excludes paths that already stopped for futility at an earlier look; a marginal boundary probability does not make that exclusion and does not represent the same event. 3. The increments are $0.03$, $0.05$, $0$, and $0.07$. They are nonnegative and sum to $0.15$, so the schedule is structurally nondecreasing and valid, though matching targets still requires appropriate calibration. 4. No. The $0.18$ is the planned probability of futility stopping by the final look under the specified design alternative, if targets are matched. Overall power depends on all possible stopping paths and cannot be determined from that value alone.