Why Directional Power Needs Its Own Check
A two-sided group-sequential design may use mirrored efficacy limits: at look \(k\), it stops for an upper-direction result if \(Z_k \geq c_k\), or for a lower-direction result if \(Z_k \leq -c_k\). This symmetry is attractive because the rule treats positive and negative test statistics alike. But a symmetric rule does not guarantee that the design has adequate power for every effect that matters. Power depends on the size of the effect being targeted as well as on the monitoring rule.
The key question is not simply whether the limits are mirrored. It is whether the design has enough probability of crossing the relevant limit under each clinically important alternative. Unequal plausible effect magnitudes can create unequal directional power even with mirrored limits. Unequal clinical consequences can also make equal power an inappropriate design objective: a direction associated with serious harm may deserve a different operating target from a direction associated with benefit.
The comparisons below use the standard canonical Gaussian model. At look \(k\), \(Z_k\) is the standardized statistic, \(I_k\) is the information, and \(\theta\) is the standardized effect per square root of information. The model assumes that information increases across looks and that the increments contain independent information. Under this model, the joint distribution has means and covariances
In particular, under \(H_0:\theta=0\), each \(Z_k\) has a standard normal marginal distribution. The canonical model is a working approximation when the test statistic is approximately normal and its information scale is appropriate; conclusions based on it should be checked if those conditions are doubtful.
When Mirrored Limits Do Give Equal Directional Power
Mirrored limits are not inherently biased toward one sign. Under the canonical Gaussian model, if the two alternatives have the same magnitude and opposite signs, and the limits are mirrored at every look, the directional operating characteristics match. This gives a useful benchmark: if reported powers differ in that setting, check the implementation, the model, or whether some other part of the design is asymmetric.
The reason is sign reflection. If the process under effect \(+\theta\) is \(Z_1,\ldots,Z_K\), then the process \(-Z_1,\ldots,-Z_K\) has the same distribution as the process under effect \(-\theta\). A mirrored rule maps every upper crossing under \(+\theta\) to a lower crossing under \(-\theta\), at the same look. Thus the probabilities of those paired events are equal.
Worked Example: Unequal Alternatives Under the Same Limits
Worked Example: Unequal Alternatives Under the Same Limits
Suppose a hypothetical trial has one final analysis and uses mirrored two-sided limits of \(c=1.96\). The standard error is \(0.10\) outcome units. Consider a clinically plausible positive effect of \(0.30\) units and a negative effect of \(-0.15\) units. Under the normal model, the standardized means are the effects divided by the standard error:
For a positive effect, upper directional power is \(P(Z\geq 1.96)\) when \(Z\sim N(3.00,1)\):
For the negative effect, lower directional power is
The rule therefore has about 85.1% power in the specified positive direction and 32.3% in the specified negative direction. The difference does not show that mirrored limits intrinsically favor positive effects: the alternatives are not equal in magnitude. It shows that this design is much less likely to detect the smaller negative effect. If that negative effect is important, its power needs to be addressed explicitly.
A Sign-Reflection Justification
The symmetry result can be stated for a full sequence of looks, not only a single final analysis. Let the upper and lower boundaries at look \(k\) be \(c_k\) and \(-c_k\). Under the canonical model, the mean vector under \(+\theta\) is \((\theta\sqrt{I_1},\ldots,\theta\sqrt{I_K})\), while under \(-\theta\) it is its negative. The covariance matrix is unchanged by changing the sign of every component. Consequently, negating the entire statistic vector under \(+\theta\) gives the distribution of the vector under \(-\theta\).
For any sample path, negation exchanges \(Z_k\geq c_k\) with \(-Z_k\leq -c_k\) at every look. It also preserves the first look at which either boundary is crossed. Therefore the path-level events of upper crossing under \(+\theta\) and lower crossing under \(-\theta\) have equal probabilities. This argument depends on all the stated symmetries: equal-magnitude alternatives, a sign-symmetric joint model, and mirrored limits. It does not imply equal power for alternatives of different magnitudes, nor does it establish equal clinical value for the two outcomes.
Worked Example: Checking the Symmetry Benchmark
Worked Example: Checking the Symmetry Benchmark
At a hypothetical single final analysis, let the mirrored limits be \(1.96\) and \(-1.96\). Suppose the two alternatives are standardized effects \(+2.40\) and \(-2.40\), with unit variance for \(Z\). Under the positive alternative, upper directional power is
Under the negative alternative, lower directional power is
The opposite-tail probability under either alternative is negligible here: for example, under \(\delta=2.40\), it is \(\Phi(-1.96-2.40)=\Phi(-4.36)\), approximately \(0.0000065\). The example illustrates the symmetry benchmark, not a claim that the chosen power is adequate. Both directions have about 67.0% directional power, so both may be underpowered if the design target is higher.
Power Targets and Clinical Priorities
When effect magnitudes differ, symmetric limits can leave one important direction with inadequate power. One possible response is to consider different upper and lower limits. Another is to revisit the effect sizes used to plan the study or the overall design. The choice depends on the scientific question: a design should not be made asymmetric merely because its initial power calculations are inconvenient.
Clinical priorities are a separate issue from the probability calculations. If missing a harmful effect has greater consequences than missing a beneficial effect, equal directional power need not be the appropriate objective. That judgment may motivate different power targets, different stopping limits, or different benefit-harm thresholds. Whichever choice is made, the resulting rule must still be calibrated so that its overall type I error rate under \(H_0\) meets the design requirement. Moving one limit without recalculating error probabilities can change the chance of a false rejection, including through repeated looks.
Worked Example: Limits Implied by Unequal Power Targets
For a transparent one-analysis illustration, suppose the positive alternative is \(\delta_+=3.00\), the negative alternative is \(\delta_-=-2.30\), and the desired directional power is 90% in both directions. For a normal statistic with unit variance, an upper cutoff \(u\) gives power \(P(Z\geq u)=\Phi(\delta_+-u)\). Since the 90th percentile of the standard normal distribution is approximately \(1.2816\), the upper cutoff needed for 90% power is
For the lower direction, power is \(P(Z\leq \ell)=\Phi(\ell-\delta_-)\). Solving for 90% power gives
These cutoffs are not mirrored. They are the cutoffs that separately attain the stated directional power targets under the two assumed alternatives. They do not, by themselves, establish a valid two-sided testing procedure: the combined type I error under \(H_0\) must also be checked, and in a group-sequential design the error calculation must account for all looks. The calculation identifies a power conflict that a complete design must resolve.
Diagnostics, Alternatives, and Limitations
A useful diagnostic is a directional power table. For each clinically important effect, record its sign and magnitude, the probability of crossing the corresponding efficacy limit, and the probability of crossing the opposite limit. For a group-sequential design, calculate crossing probabilities using the joint distribution across looks; marginal power at the final look alone does not include early crossings. Compare the results with the prespecified power targets and clinical priorities, not just with one another.
If the computed powers differ where equal-and-opposite effects should match, investigate the assumptions and calculations. Check that the boundaries really are mirrored at every look, that the alternatives have equal standardized magnitudes, and that the covariance and information inputs are consistent. Also consider whether the statistic is adequately approximated by the canonical Gaussian model. A skewed sampling distribution, an incorrectly specified information scale, or differing nuisance-parameter behavior across directions can undermine the reflection argument.
The main alternatives answer different questions. Mirrored limits with equal-magnitude alternatives provide a symmetric reference design. Unequal limits may target unequal directional power or reflect different priorities, but their error rates must be recalibrated. Alternatively, a design can retain mirrored limits while choosing a common target based on the direction that is harder to detect; this can require more information than planning around the easier direction. Neither approach removes the need to specify which effects are clinically important.
A common mistake is to conclude, “The limits are symmetric, so power must be the same in both directions.” That is correct only for paired, equal-magnitude alternatives under the relevant sign-symmetric model and mirrored rule. If the positive alternative is \(+3\) standard-error units and the negative alternative is only \(-1.5\), those are not a sign-reversed pair. Another mistake is to lower a boundary to improve power in one direction and assume the error rate is unchanged. The correction is to evaluate directional power for the stated alternatives and recalibrate the entire group-sequential rule under the null.
Check Your Understanding
Answer each question using the definitions and assumptions in this tutorial. Short solutions follow the questions.
- With mirrored limits and a sign-symmetric canonical Gaussian model, why do alternatives \(+\theta\) and \(-\theta\) have equal directional power when their magnitudes match?
- At one analysis with limits \(\pm 1.96\), what is the upper directional power under a standardized effect of \(+2.00\)?
- Does equal directional power imply that both directions have adequate power? Explain.
- Why is it not enough to lower one group-sequential boundary and leave the error-rate calculation unchanged?
Solutions. (1) Negating every statistic at every look maps the process under \(+\theta\) to the process under \(-\theta\), and maps every upper crossing to a lower crossing at the same look. (2) The power is \(\Phi(2.00-1.96)=\Phi(0.04)\approx 0.5160\), or 51.6%. (3) No. Equal power can still be below the prespecified target, as in the example with about 67.0% power in each direction. (4) Changing a boundary changes the probability of crossing under \(H_0\); with repeated looks, the joint crossing probability across all looks must be recalculated to verify the overall type I error rate.