Unequal Upper And Lower Efficacy Limits

Why Unequal Limits Need a Signed Scale

A two-sided group sequential design can stop early when evidence favors either direction of a treatment contrast. If the treatment effect has a meaningful sign convention, however, the two directions need not use limits of equal magnitude. A large positive statistic may indicate benefit, while a sufficiently negative statistic may indicate harm or a meaningful effect in the opposite direction. Unequal limits make these stopping regions different in the two directions.

The interpretation depends on fixing the contrast orientation before looking at the data. In this tutorial, a positive contrast means an effect in the direction called favorable for treatment, and a negative contrast means an effect in the opposite direction. Reversing that convention reverses the interpretation of every sign.

These are look-specific limits: each may vary with \(k\). A statistic between the limits at a look does not trigger an efficacy stop there. If the trial has another planned look, the design continues, subject to any separately specified rules such as safety stopping. The lower efficacy limit is not automatically a futility limit: crossing it is evidence for the opposite signed direction, whereas a futility rule concerns whether continuing is worthwhile under a power or benefit criterion.

Model and Decision Regions

The standard working model for this discussion is a prespecified sequence of planned analyses with increasing information. Under the global null \(H_0\), the standardized statistics are approximately jointly normal with mean zero, unit variance, and canonical correlations. In particular, for looks \(i<j\),

where \(I_k\) is the information available at look \(k\). This structure arises, for example, when the underlying score has independent increments and each statistic is the accumulated score divided by the square root of its information. Under a fixed signed effect \(\theta\), the canonical approximation has \(E(Z_k)=\theta\sqrt{I_k}\); positive \(\theta\) favors the designated treatment direction. These approximations and the planned analysis schedule must be suitable for the particular test and data.

The stopping rule is sequential: it uses the first look at which either limit is crossed. Define \(C_k\) as the event that all earlier looks continued, meaning \(L_j<Z_j<U_j\) for every \(j<k\). Then the two first-crossing events at look \(k\) are

These events are mutually exclusive: at a given look, a statistic cannot be at least \(U_k\) and at most \(L_k\) when \(L_k<U_k\). First-crossing events at different looks are also disjoint, because a path that stops at one look does not proceed to a later efficacy decision.

Why the Regions Are Asymmetric

At one look, the rejection region is the union \((-\infty,L_k]\cup[U_k,\infty)\). With more than one look, the overall rejection region consists of all paths that first enter one of these regions at any planned look. The lower and upper parts are defined separately; they need not be mirror images.

A simple sign-change check shows the distinction. If every statistic on a path is replaced by its negative, then the original upper crossing \(Z_k\geq U_k\) becomes \(-Z_k\leq -U_k\). For this to be exactly the lower crossing under the same limits, it must be that \(L_k=-U_k\). Thus, when the magnitudes differ, reversing the sign of the data does not generally preserve the stopping decision. This is a statement about the geometry of the rule, not a statement that one direction has a particular probability of rejection.

The limits must still be assessed as a sequential design. Under \(H_0\), the overall type I error is the probability of crossing either boundary at any look. The marginal tail probability at one look is not, by itself, that overall error probability. The boundaries therefore need to be calibrated for the intended error control using the joint null distribution and the planned looks. Unequal limits specify the shape of the decision regions; they do not automatically guarantee a chosen error rate.

Worked Decisions in Both Directions

Assumptions, Diagnostics, and Alternatives

The rule is meaningful only if its ingredients are coherent. The signed contrast and favorable direction must be defined before interpreting results; the limits must be specified for every planned look; and the statistic must use the same orientation and standardization throughout. The approximate joint distribution used to calibrate the limits must also be appropriate. In practice, this requires attention to the analysis model, information timing, endpoint, and any design features that affect the null distribution.

Several checks help detect implementation errors:

• Check ordering. At each look, verify \(L_k<U_k\). If the limits overlap or are reversed, the continuation interval is not defined as intended.
• Check signs and labels. Confirm that positive Z values still mean the designated favorable direction and that lower-limit crossings have the intended interpretation.
• Check the first crossing. Once a limit is crossed, later hypothetical statistics cannot change which look first triggered stopping.
• Check calibration separately. A plausible-looking pair of limits is not proof of overall type I error control. The complete sequential rule, including all looks, must be evaluated under the null model.
• Check what the rule does not say. Crossing a Z limit is a decision under the design, not a direct statement about clinical importance or a guarantee of replication.

A symmetric two-sided rule uses limits with equal magnitude and opposite signs at each look. It has sign-mirrored decision regions, unlike an asymmetric rule. A separate one-sided efficacy rule instead targets one direction and ordinarily has no opposite-direction efficacy boundary. Unequal two-sided limits retain stopping regions in both directions but allow different thresholds for them. These are different specifications of the decision rule; none can be substituted for another without considering the intended hypotheses and error control.

Unequal limits have limitations. Their use alone does not explain why one direction has a less stringent threshold, nor does it quantify clinical consequences. Nor does a more extreme numerical limit necessarily translate into a particular directional error probability when there are multiple correlated looks. The interpretation should therefore report the signed contrast convention, the limit crossed, and the look at which the first crossing occurred.

Common Mistake