Spending Shape and Power Tradeoffs

Why Spending Shape Matters

Two group sequential designs can use the same significance level, maximum sample size, and planned look times yet have different power. The reason is that they may distribute their type I error differently across the interim analyses. A design that spends more alpha early typically uses a less stringent early efficacy boundary, creating more opportunity to stop early for benefit. To keep the total type I error fixed, however, its later boundaries must also be calibrated to the alpha remaining.

The tradeoff is about when evidence can trigger rejection, not about changing the overall type I error target. An early opportunity can help when an effect is large enough to produce strong interim evidence. A design that reserves more alpha for later looks can instead place relatively more emphasis on evidence accumulated by the end. Which design has greater power depends on the effect size and the full sequence of boundaries.

Definition: An alpha-spending shape specifies the cumulative amount of type I error allocated by each planned information time. For looks $k=1,\ldots,K$, let $t_k=I_k/I_{\max}$, where $I_k$ is the information at look $k$ and $I_{\max}$ is the maximum information. A spending schedule $A(t_k)$ is nondecreasing, with $A(0)=0$ and $A(1)=\alpha$. The incremental spend at look $k$ is $A(t_k)-A(t_{k-1})$, taking $A(t_0)=0$.

Cumulative spending is a target for the probability of having crossed an efficacy boundary by that look under $H_0$. It is not generally the probability of crossing at that look alone. Because the test statistics across looks are correlated, boundary calibration accounts for the probability of having crossed earlier. The earlier material on boundary calibration and crossing probabilities provides the numerical machinery; the question here is how different valid spending targets compare after calibration.

Comparison Conditions and Power

To attribute power differences to spending shape, hold the rest of the design fixed. In particular, use the same one-sided $H_0$ and $H_a$, significance level $\alpha$, information times, maximum information, test statistic, and analysis rules. Compare designs that are each calibrated to the same cumulative type I error $\alpha$. Assume a regular large-sample setting in which the canonical group sequential model is adequate, information times are correctly specified, and there are no futility stops or other changes to the analysis plan.

Model: Under a standardized effect $\theta$, the test statistics at the planned looks follow the canonical model $$ (Z_1,\ldots,Z_K)\sim N_K\left((\theta\sqrt{t_1},\ldots,\theta\sqrt{t_K}),\,\Sigma\right), \qquad \Sigma_{ij}=\sqrt{\frac{t_i}{t_j}}\quad (i\le j), $$ with $\Sigma_{ji}=\Sigma_{ij}$. Here $\theta$ is the effect measured in final-information standard-error units; under $H_0$, $\theta=0$. A design with upper efficacy boundaries $b_1,\ldots,b_K$ rejects $H_0$ if $Z_k\ge b_k$ at any look $k$.

For a specified alternative $\theta>0$, power is the probability of crossing at least one boundary:

\operatorname{Power}(\theta) =P_{\theta}\left(\bigcup_{k=1}^{K}\{Z_k\ge b_k\}\right) =1-P_{\theta}(Z_1<b_1,\ldots,Z_K<b_K).

The spending schedule determines the calibrated boundaries; the joint distribution and those boundaries determine power. Thus, comparing only the first-look boundaries, or only the final boundaries, does not establish which design has greater overall power. The probability of crossing at any look is the relevant comparison.

From Spending Targets to a Power Tradeoff

A larger cumulative spend at an early look usually requires a less stringent boundary there, since the design must allow a larger null crossing probability by that time. But the boundary sequence is calibrated jointly: the probability of crossing earlier affects how much additional type I error can be spent later. A spending schedule alone is therefore not a list of per-look critical values. The boundaries must be calibrated under the joint null distribution.

Worked Example: Comparing Cumulative and Incremental Spend

Suppose three candidate designs use four planned looks at information fractions $0.25, 0.50, 0.75,$ and $1.00$, with one-sided $\alpha=0.025$. Their hypothetical cumulative spending targets are listed below. Each target is nondecreasing and ends at the same total alpha.

Shape	$A(0.25)$	$A(0.50)$	$A(0.75)$	$A(1.00)$
Front-loaded	0.008	0.014	0.020	0.025
Gradual	0.004	0.010	0.018	0.025
Back-loaded	0.001	0.004	0.012	0.025

For the front-loaded design, the incremental spends are $0.008$, $0.014-0.008=0.006$, $0.020-0.014=0.006$, and $0.025-0.020=0.005$. They sum to $0.025$. For the back-loaded design, the increments are $0.001$, $0.004-0.001=0.003$, $0.012-0.004=0.008$, and $0.025-0.012=0.013$, also summing to $0.025$.

The front-loaded target allows a greater cumulative null crossing probability by the first look than the back-loaded target: $0.008$ versus $0.001$. This indicates the intended difference in early emphasis. It does not mean that $0.008$ and $0.001$ are the probabilities of crossing at the first look conditional on not having crossed before; nor do these values themselves give the boundaries. Each schedule still needs boundary calibration.

A useful justification for examining power over several effect sizes is that, with fixed upper boundaries, power cannot decrease as the positive effect increases under this model. Couple the statistics at effect $\theta$ to those at effect zero by writing $Z_k(\theta)=Z_k(0)+\theta\sqrt{t_k}$. For $\theta_2>\theta_1$, each look satisfies $Z_k(\theta_2)\ge Z_k(\theta_1)$. Therefore any path that crosses a boundary under $\theta_1$ also crosses under $\theta_2$. This establishes monotonicity for each design, but it does not establish that one spending shape has greater power than another at every effect size.

Worked Power Comparison

The following comparison shows why the alternative effect must be specified. Its figures are hypothetical rounded outputs from calibrating each design to the same null crossing targets and evaluating the joint crossing probability. They are illustrative numerical results, not universal properties of these three labels.

Worked Example: Power Across Three Effect Sizes

Assume the four information fractions and the three spending schedules in the preceding example. Suppose calibration under the canonical model and numerical evaluation of the joint crossing probability give the following powers:

Standardized effect $\theta$	Front-loaded	Gradual	Back-loaded
1	0.220	0.230	0.240
2	0.640	0.630	0.610
3	0.920	0.900	0.870

For $\theta=1$, the back-loaded design has the largest displayed power, $0.240$. Its advantage over the front-loaded design is $0.240-0.220=0.020$, or 2 percentage points. For $\theta=3$, the front-loaded design has the largest displayed power, $0.920$, exceeding the back-loaded design by $0.920-0.870=0.050$, or 5 percentage points. The ranking changes with the effect size.

In this example, $\theta=2$ means that the alternative effect is two final-information standard errors from zero. A power of $0.640$ means that, if this alternative and the model hold, the front-loaded design rejects $H_0$ in about 64% of repeated trials. It is not the probability that $H_a$ is true. The comparison supports a specific statement about these calibrated designs at these effect sizes; it does not show that front-loading is always preferable.

The calculation being evaluated is $$ \operatorname{Power}(\theta) =1-P_{\theta}(Z_1<b_1,\ldots,Z_4<b_4). $$ For example, at $\theta=2$, the stated numerical output for the front-loaded design is $1-0.360=0.640$. The no-crossing probability $0.360$ is the joint probability over all four correlated looks, not a product of four separate marginal probabilities.

Making a Design Choice

A power comparison should begin with the alternatives that matter scientifically. If early evidence and early stopping for efficacy are priorities, a front-loaded design may be attractive, provided its power at the design alternative is adequate. If the main objective is power for effects that need more information to become apparent, reserving more spending for later looks may be preferable. These are design considerations, not guarantees: the calibrated power curve is the evidence for the particular schedule and model.

A fair comparison should report, at minimum, the total $\alpha$, information schedule, maximum information, spending targets, calibration method, and power at clinically relevant alternatives. A table or curve over a range of $\theta$ is generally more informative than a single power value. The effect scale should also be stated clearly; standardized effect sizes are not directly interpretable in clinical units unless translated using the final standard error or the model’s effect parameter.

Power is not the only design criterion. Spending shape can also affect how often a trial stops early and the expected amount of information used, even when maximum information is fixed. A power comparison alone does not answer those questions. Nor does a numerical power calculation protect against a misspecified information schedule or an unsuitable canonical approximation; sensitivity checks may be needed when those assumptions are doubtful.

Diagnostics and Common Pitfalls

Check null calibration. Each design should attain the same cumulative type I error at the final look, and its interim cumulative crossing probabilities should agree with its stated spending targets to numerical tolerance.
Check the inputs. Verify that the compared designs use the same look times, maximum information, test direction, and analysis rules. Otherwise, power differences cannot be attributed to spending shape alone.
Check the full curve. A ranking at one effect size can reverse at another. Report powers at a meaningful range of alternatives rather than declaring one shape universally more powerful.
Check numerical stability. Joint crossing probabilities can be sensitive to integration accuracy and boundary-calibration tolerances. Repeat calculations at tighter tolerances if results are close or change unexpectedly.

Common mistake: “The front-loaded design spends more alpha early, so it must have more power.” This confuses an early crossing opportunity with total power under a particular alternative. Earlier boundaries and later boundaries are calibrated together; the correct comparison is the probability of crossing any boundary under that alternative.

Key takeaway: Spending shape reallocates the timing of type I error, not its total amount. Under fixed maximum information and look times, compare fully calibrated designs using their joint crossing probabilities across the effect sizes that matter; no spending shape is automatically best for every alternative.

Check Your Understanding

Use the definitions and hypothetical power table above. The questions emphasize how to interpret a comparison rather than how to calibrate boundaries.

For the gradual schedule, calculate the incremental spend at each of the four looks.
At $\theta=2$, which design has the largest displayed power, and what is its difference from the back-loaded design?
Why is the cumulative spend $A(0.50)$ not generally the probability of crossing exactly at the second look?
What does power $0.240$ at $\theta=1$ mean in the back-loaded design?
Give two conditions that should be held fixed to attribute power differences to spending shape.

Solutions. (1) The increments are $0.004$, $0.010-0.004=0.006$, $0.018-0.010=0.008$, and $0.025-0.018=0.007$; they sum to $0.025$. (2) The front-loaded design is largest at $0.640$, a difference of $0.640-0.610=0.030$, or 3 percentage points. (3) $A(0.50)$ is the cumulative probability of having crossed by the second look under $H_0$; it includes paths that crossed earlier. (4) If the model and $\theta=1$ alternative hold, about 24% of repeated trials using that design reject $H_0$. (5) Examples include holding total $\alpha$, information times, maximum information, test direction, and analysis rules fixed.