Power Tradeoffs Under Asymmetric Boundaries

Why Boundary Asymmetry Changes Power

In a two-sided group sequential design, the upper boundary can represent evidence of benefit and the lower boundary evidence of harm. If the limits differ in stringency, the design need not have equal chances of crossing in the two directions. That difference matters when the scientific consequences of a missed benefit and an undetected harm are not the same. The key is to distinguish how the rule allocates rejection probability from whether it controls the overall null rejection probability.

Throughout, take a positive standardized effect to indicate benefit and a negative effect to indicate harm. At look \(k\), let \(Z_k\) be the standardized test statistic, and let \(l_k\) and \(u_k\) be the lower and upper efficacy limits, with \(l_k<u_k\). The rule rejects at the first look for which \(Z_k\leq l_k\) or \(Z_k\geq u_k\). If neither condition holds, monitoring continues when another look is planned.

For a standard group-sequential normal approximation, suppose information at look \(k\) is \(I_k\), with \(0<I_1<\cdots<I_K\), and let \(t_k=I_k/I_K\). If \(\eta=\delta\sqrt{I_K}\) is the standardized effect at the final information level, then \(Z_k\) has mean \(\eta\sqrt{t_k}\) and variance 1. Under the canonical independent-increments model, the covariance of \(Z_i\) and \(Z_j\) is \(\sqrt{t_i/t_j}\) for \(i<j\). These approximations require an appropriate large-sample normal model, correctly specified information times, and a fixed monitoring rule. The looks are dependent; treating their statistics as independent gives incorrect first-crossing probabilities.

First-Crossing Probabilities and Directional Power

Write \(U_k\) for the event that the first efficacy crossing occurs at look \(k\) and is upper, and \(L_k\) for the corresponding lower event. Then the upper and lower powers at effect \(\eta\) are sums of mutually exclusive first-crossing probabilities:

The subscript \(\eta\) indicates that the probability is evaluated under that alternative, not under \(H_0\). Under \(H_0\), the sum is the type I error probability for the complete monitoring rule. The two directional null error probabilities add to the overall null rejection probability, but neither is generally equal to a nominal per-look tail probability: the looks are dependent and only first crossings count.

At one look, where \(Z\sim N(\eta,1)\), the calculations reduce to normal tail areas. For lower limit \(l\) and upper limit \(u\),

Thus the boundary locations determine the probability mass assigned to each direction. Raising \(u\) reduces the upper crossing probability at a fixed effect; lowering \(l\) reduces the lower crossing probability. At multiple looks, however, changing an early boundary can also alter which paths remain under observation, so the effect on the other direction's eventual probability is not captured by a single tail calculation.

Comparisons Must Hold Null Error Control Fixed

An asymmetric rule can appear to have more power in one direction simply because it has been given more null error probability in that tail. A fair comparison therefore states what is being held fixed. A common comparison matches the overall null rejection probability and asks how the alternative rejection probabilities differ. Another comparison holds directional null error allocations fixed. These are distinct design questions; matching only one tail does not imply that the overall null error probabilities match.

For one look, a symmetric pair of limits \(-c\) and \(c\) has null rejection probability \(2[1-\Phi(c)]\). Matching an overall null rejection probability \(\alpha\) requires \(c=\Phi^{-1}(1-\alpha/2)\). At multiple looks, the corresponding matching must use the joint first-crossing probability of the entire design, not this one-look formula. The correlation across looks and the full sequences of limits both matter.

Conditional Calculation at a Later Look

In a group sequential design, power at a later look must account for both the chance of reaching that look and the conditional chance of crossing once there. Under the canonical model, for \(i<j\),

The conditional mean includes the observed interim statistic and the expected drift from the remaining information. A later-look tail probability conditional on \(Z_i=z\) is not itself a power: to obtain unconditional first-crossing power, it must be averaged over interim values that did not already cross, and the paths must be counted only at their first crossing.

Justification, Diagnostics, and Limitations

The first-crossing decomposition follows directly from the stopping rule. For any realized path, either the first boundary crossing is upper, the first crossing is lower, or no efficacy boundary is crossed. The upper and lower first-crossing events are disjoint, as are the events for different first-crossing looks. Adding their probabilities therefore gives the probability of any efficacy rejection. At a single look, the events are simply \(Z\geq u\) and \(Z\leq l\); standardizing \(Z\sim N(\eta,1)\) yields the tail formulas above. At multiple looks, the same event logic applies, but joint normal dependence is needed to calculate each probability.

Useful diagnostics include checking that the directional probabilities sum to the total, that their sum is no greater than 1, and that the null total agrees with the design's calibrated error probability. Under a coherent normal location model with fixed limits, increasing \(\eta\) shifts the statistic distribution upward, so upper-crossing probability should not decrease and lower-crossing probability should not increase. A result that violates this check may indicate a sign error, a boundary-indexing error, or a calculation that counts paths after an earlier stop.

A symmetric-boundary comparison is useful as a reference, but it does not answer every design question. Unequal tail error allocation and unequal boundary shape can yield different power profiles even when the overall null rejection probability is the same. Further, power depends on the effect size: a rule favoring upper crossings may perform differently for a small beneficial effect than for a large harmful effect. A single alternative value cannot summarize the entire operating characteristic.

The normal approximation may be inadequate with small samples, sparse outcomes, misspecified information, or substantial departures from the assumed independent-increments structure. In those settings, power should be evaluated under a model appropriate to the trial and checked with suitable numerical methods. Finally, an efficacy rule alone does not quantify expected sample size or stopping time, nor does a rejection probability measure clinical value. Those are separate design considerations.

Common Mistake

Wrong reasoning: “The upper limit is more stringent, so the design has less power overall.” This conclusion does not follow from the upper limit alone. The lower limit may be less stringent, and total power combines the first-crossing probabilities in both directions. Changes to early limits can also change which paths remain eligible to cross later.

Correction: Calculate upper and lower first-crossing probabilities under each stated alternative, then add them for total rejection probability. Compare designs only after specifying which null error constraint is held fixed, and report the direction of rejection rather than relying on total power alone.