Mapping Z Limits To Effect-Scale Thresholds

Why Translate a Z Boundary?

A group sequential rule is often stated in standardized units: at an interim look, stop for efficacy if the test statistic crosses an upper or lower boundary. That rule is convenient for design and error-rate calculations, but it can be hard to interpret in the units of the outcome. Translating each boundary into an estimated treatment-effect threshold answers a concrete question: how large, in the outcome’s units, would the observed effect estimate have to be at this look to cross the boundary?

The translation depends on the standard error at that look. As information accumulates, standard errors commonly decrease, so the effect-scale threshold can change even when the Z-scale boundary stays the same. This is a conversion of the existing stopping rule, not a new rule and not a statement about the treatment effect that is clinically important.

Definition and Assumptions

Here \(\theta\) denotes the treatment-effect parameter, with \(H_0:\theta=0\). The estimate and standard error must refer to the same treatment contrast, analysis population, outcome scale, and look. The sign convention must also be consistent: positive estimates are in the direction represented by positive Z values. The upper boundary is crossed when \(Z_k\ge U_k\), and the lower boundary when \(Z_k\le L_k\).

The conversion itself is algebraic: it requires a positive standard error and the stated definition of \(Z_k\). It does not by itself establish that the boundaries control the intended type I error or power. Those properties depend on the design’s assumptions and calibration, including the joint behavior of the statistics across looks. The mapping also assumes the boundary and statistic are expressed in the same Z coordinate system. The earlier material on sign conventions for upper and lower boundaries establishes how the directions are labelled.

Deriving the Effect-Scale Thresholds

At a given look, substitute the definition \(Z_k=\hat{\theta}_k/SE_k\) into the upper-boundary crossing condition. Since \(SE_k>0\), multiplying an inequality by it preserves the inequality direction:

The lower-boundary condition works the same way. Positivity of \(SE_k\) again means that multiplying does not reverse the inequality:

Thus, if \(L_k< U_k\), the continuation condition at look \(k\) is \(L_k SE_k<\hat{\theta}_k<U_k SE_k\), using strict inequalities when boundary equality counts as crossing. The two effect-scale thresholds need not be equal in magnitude. They inherit any asymmetry in the Z boundaries, and both are scaled by the standard error at that particular look.

Using the Mapping at Each Look

A reliable calculation keeps the look index attached to every quantity. Pair \(U_k\) and \(L_k\) with \(SE_k\), not with a standard error from another look, and multiply before comparing the observed estimate with the thresholds. The following steps make the conversion explicit.

Effect-Scale Thresholds Are Look-Specific

A Z boundary is dimensionless, while an effect-scale threshold has the units of the estimated effect. Even if the same upper Z limit is used at two looks, its effect-scale value generally differs when the standard errors differ. Conversely, a boundary that becomes less stringent in Z units might still correspond to a smaller effect-scale threshold if the standard error has fallen sufficiently. For each look, the product \(U_k SE_k\) or \(L_k SE_k\), rather than either component alone, determines the corresponding threshold.

This distinction is useful when communicating what an interim rule means in outcome units. It does not make the threshold a fixed clinical target. A clinical target is specified on the effect scale for its substantive meaning; a sequential crossing threshold is obtained from the design boundary and current standard error. They can differ, and one should not be substituted for the other.

Diagnostics, Alternatives, and Limitations

Several quick checks can reveal a conversion error. The product of a Z boundary and standard error must have the effect’s units. If a positive standard error is used, a negative lower boundary must remain negative; if the lower threshold becomes positive without a change in the boundary’s sign convention, check the arithmetic. Finally, recomputing \(Z_k=\hat{\theta}_k/SE_k\) from the estimate used for the decision should reproduce the same crossing conclusion as the effect-scale comparison.

The direct alternative is to retain the Z-scale rule and compare \(Z_k\) with \(U_k\) and \(L_k\). This is often the clearest method for implementing the formal stopping rule. The effect-scale version is a useful translation for interpretation and reporting, but it is algebraically equivalent only when the stated Z statistic is exactly the estimate divided by the stated positive standard error. A confidence interval or a fixed effect target is not a replacement for applying the sequential boundaries.

The mapping alone does not address uncertainty in an estimated standard error, changes in the analysis estimator, missing data, or whether the assumed sequential model is adequate. If an analysis uses a different statistic, a different effect measure, or a transformed coordinate, the formula must match that definition rather than be applied mechanically. The subsequent treatment of nonzero null margins uses a shifted coordinate; here the stated mapping is specifically for the zero-effect null and \(Z_k=\hat{\theta}_k/SE_k\).

Common Mistake

Wrong reasoning: “The upper limit is \(2.2\), so the estimate must exceed \(2.2\) outcome units to stop.” This confuses a dimensionless Z boundary with an effect measured in outcome units. Correction: multiply the Z boundary by the standard error at that look. If \(SE_k=0.70\) outcome units, an upper limit of \(2.2\) corresponds to \(2.2(0.70)=1.54\) outcome units, not \(2.2\) outcome units. Also avoid describing this crossing threshold as the treatment effect under the null or as a clinical importance threshold; it is the observed-estimate value equivalent to the specified standardized boundary at that look.