Sign Conventions For Upper And Lower Boundaries

Why a Sign Convention Matters

A group-sequential decision is often described as crossing an upper or lower boundary. Those words identify positions on the signed Z scale; they do not, by themselves, say which treatment performed better. The scientific interpretation depends on how the treatment contrast was defined. If positive values mean treatment benefit, an upper crossing points in that direction. If the contrast is reversed, the same evidence appears as a lower crossing instead.

This distinction matters whenever results are calculated, reported, or displayed using different contrast orders. A boundary-crossing decision can be translated correctly only when the statistic, contrast definition, and boundary signs all refer to the same orientation. The decision regions themselves are familiar from two-sided group sequential decision regions: at each planned look, the statistic can cross the upper boundary, cross the lower boundary, or remain between them. Here the focus is the sign convention that gives those outcomes their direction.

Definition: At look $k$, let $\hat{\theta}_k$ estimate a treatment contrast $\theta$, and let $SE_k>0$ be its standard error. For testing the null value $\theta_0$, define the signed statistic $$ Z_k=\frac{\hat{\theta}_k-\theta_0}{SE_k}. $$ The sign is determined by the specified order and coding of the contrast. Given an upper boundary $u_k$ and a lower boundary $l_k$, with $l_k<u_k$, an upper crossing occurs if $Z_k\geq u_k$, a lower crossing if $Z_k\leq l_k$, and continuation at that look if $l_k<Z_k<u_k$. The boundary labels refer to the numerical scale of $Z_k$.

This definition assumes that the analysis uses a signed, approximately standard-normal statistic on the scale for which the boundaries were specified. In a typical large-sample analysis, the estimate and its standard error are calculated under the prespecified model, and the statistic is standardized against the null value. The estimate must represent the same contrast at every look, and the standard error must be positive. A group-sequential design may use correlated statistics across looks; the sign translation below does not require the looks to be independent. The boundaries must nonetheless be those calibrated for the planned design and statistic.

For interpretation, suppose $\theta=\mu_T-\mu_C$, where $T$ denotes treatment and $C$ control, and suppose larger outcome values are defined as more favorable. Then a positive $\theta$ means a larger mean outcome in the treatment group. If instead the endpoint is one for which smaller values are favorable, that contrast definition does not make positive values beneficial automatically: the endpoint’s meaning must also be stated. Neither the word “upper” nor the sign of $Z_k$ substitutes for that context.

Reading the Decision at a Look

With a fixed contrast orientation, the rule is direct: compare the observed signed statistic with the two boundary values. Crossing the upper boundary means the statistic is sufficiently high on that signed scale to meet the upper stopping criterion. Crossing the lower boundary means it is sufficiently low to meet the lower stopping criterion. If neither inequality holds, the statistic is in the continuation region at that look. The interpretation of an upper or lower crossing as evidence in a particular scientific direction comes from the contrast definition, not from the stopping rule alone.

For example, if treatment minus control is the defined contrast and larger outcomes favor treatment, then a positive Z statistic corresponds to an estimate above the null value in the treatment-favorable direction. A negative Z statistic corresponds to an estimate below the null value. If the contrast is defined as control minus treatment instead, those directional interpretations reverse. The numerical boundary comparison remains valid only if the boundary signs are expressed in the same orientation as the statistic.

Worked Example: Classifying a Statistic at One Look

In a hypothetical randomized trial, the endpoint is a score for which larger values are defined as more favorable. The analysis uses treatment minus control, with null contrast $\theta_0=0$. At look 2, the estimated contrast is $4.8$ score points and its standard error is $2.0$ points. The supplied boundaries are $u_2=2.30$ and $l_2=-2.10$. Assume the estimate and standard error come from the prespecified analysis, are appropriate for this look, and the boundaries were calibrated for this signed statistic.

First calculate the statistic: $$ Z_2=\frac{4.8-0}{2.0}=2.40. $$

Since $2.40\geq 2.30$, the statistic crosses the upper boundary. It does not cross the lower boundary, because $2.40$ is not less than or equal to $-2.10$. Under the stated coding, the estimate is positive in the treatment-minus-control direction, and larger scores are favorable. Thus the upper crossing is in the treatment-favorable direction. This is a boundary-based stopping decision; the sign convention alone does not establish the size or practical importance of the effect.

What Reversing the Contrast Does

Reversing the order of the contrast negates the estimate relative to the null and leaves its standard error unchanged. For example, changing from treatment minus control to control minus treatment changes $\hat{\theta}_k-\theta_0$ to its negative, provided the null value is reversed along with the contrast. Therefore the signed statistic changes sign. To describe the same decision using the reversed statistic, the upper and lower boundary roles must also be exchanged and negated.

Result: If the reversed contrast and its null are both the negatives of the original contrast and null, then $Z_k^*=-Z_k$. A decision rule expressed in the reversed orientation has upper boundary $u_k^*=-l_k$ and lower boundary $l_k^*=-u_k$. Consequently, an original upper crossing corresponds to a reversed lower crossing, and an original lower crossing corresponds to a reversed upper crossing.

The transformed boundaries follow from translating the inequalities, not from a change in the underlying evidence. An original upper crossing satisfies $Z_k\geq u_k$. Since $Z_k^*=-Z_k$, this is equivalent to $Z_k^*\leq -u_k$, which is a lower crossing at $l_k^*=-u_k$. Likewise, an original lower crossing satisfies $Z_k\leq l_k$, equivalent to $Z_k^*\geq-l_k$, which is an upper crossing at $u_k^*=-l_k$. The continuation condition transforms as well: $$ l_k<Z_k<u_k \quad\Longleftrightarrow\quad -u_k<Z_k^*<-l_k. $$

This algebra also covers boundaries that are not equal in absolute value. Reversing a contrast does not justify simply swapping the words “upper” and “lower” while retaining the old numerical values. The signs and values must be transformed as shown. If a design uses symmetric boundaries, the transformed values may look especially simple, but symmetry is not required for the sign translation.

Worked Example: Reversing the Contrast and the Boundaries

Consider a hypothetical analysis at look 1 with contrast treatment minus control, null value zero, estimate $0.42$ units, and standard error $0.18$ units. Suppose the boundaries are $u_1=2.25$ and $l_1=-2.05$. Assume the prespecified model supports the standard-error calculation, and that the boundaries apply to this contrast and look.

In the original orientation, $$ Z_1=\frac{0.42-0}{0.18}=2.333\ldots\approx 2.33. $$ Because $2.33\geq2.25$, this is an upper crossing. Now reverse the contrast to control minus treatment. The estimate is $-0.42$ units, the null remains zero, and the standard error remains $0.18$ units: $$ Z_1^*=\frac{-0.42-0}{0.18}=-2.333\ldots\approx -2.33. $$

The correctly transformed boundaries are $u_1^*=-l_1=2.05$ and $l_1^*=-u_1=-2.25$. The reversed statistic satisfies $-2.33\leq-2.25$, so it crosses the lower boundary. The same data therefore produce an upper crossing under treatment-minus-control coding and a lower crossing under control-minus-treatment coding. In both cases the direction of the estimated contrast is unchanged scientifically: treatment has the higher estimated outcome by $0.42$ units.

A Short Derivation and a Direction Check

Let the original contrast be $\theta=\mu_T-\mu_C$ with null value $\theta_0$. Define the reversed contrast by $\theta^*=-\theta$, so its null value is $\theta_0^*=-\theta_0$ and its estimate is $\hat{\theta}_k^*=-\hat{\theta}_k$. Negating an estimate does not change its variance, because $\operatorname{Var}(-X)=\operatorname{Var}(X)$; hence $SE_k^*=SE_k$. Substitution into the statistic gives $$ Z_k^*= \frac{\hat{\theta}_k^*-\theta_0^*}{SE_k^*} = \frac{-\hat{\theta}_k-(-\theta_0)}{SE_k} = -\frac{\hat{\theta}_k-\theta_0}{SE_k} = -Z_k. $$

The derivation requires the entire contrast to be reversed consistently. Reversing only the label in a table while leaving the calculated estimate, null value, or boundaries untouched is not a valid transformation. Nor does the result say that every observed result must stop the trial: it only establishes how a specified boundary decision maps between two orientations.

Worked Example: Interpreting a Lower Crossing

In a hypothetical study, the endpoint is a measure for which larger values are favorable. The contrast is treatment minus control, and at look 3 the estimate is $-3.6$ points with standard error $1.5$ points. The supplied boundaries are $u_3=2.40$ and $l_3=-2.00$. Assume the estimator and standard error follow the planned model and that the boundaries are stated for treatment minus control.

The statistic is $$ Z_3=\frac{-3.6-0}{1.5}=-2.40. $$ Since $-2.40\leq-2.00$, this is a lower crossing. In this orientation, the estimate is below the null in the treatment-minus-control direction, so the observed estimate is unfavorable to treatment relative to control for this endpoint. “Lower crossing” describes the numerical decision; the endpoint definition and contrast order supply its scientific interpretation.

If the analysis were instead reported as control minus treatment, the statistic would be $Z_3^*=2.40$. The transformed upper boundary would be $u_3^*=-l_3=2.00$, so $2.40\geq2.00$ is the corresponding upper crossing. Keeping the original upper and lower boundaries after reversing the statistic would mislabel the decision.

Diagnostics, Alternatives, and Limitations

A practical check is to trace the sign from the raw contrast to the decision. Verify which group is subtracted from which, how larger and smaller endpoint values are interpreted, what null value is used, and whether the statistic and boundaries share the same orientation. Then recompute the statistic from the estimate and standard error. If the estimate is on one scale but the boundary is for a differently coded statistic, the comparison is invalid even if the arithmetic is performed correctly.

Common warning signs include a reported positive estimate paired with a negative Z statistic when the null is zero and the standard error is positive; a change in treatment/control ordering between looks; or a monitoring table whose upper and lower labels disagree with the analysis code. Such discrepancies should be resolved before interpreting a stop decision. A useful audit is to negate the estimate and verify that the statistic also negates, while the standard error remains unchanged.

There are two common reporting conventions. One is to keep a single prespecified contrast orientation throughout and describe the upper and lower crossings relative to it. Another is to reverse the contrast for a different display or report. The latter can be valid, but the statistic and boundaries must be transformed together. Reporting the signed statistic, the contrast order, and the endpoint direction is generally more informative than reporting only $|Z_k|$. An absolute statistic discards which side of the null was observed and cannot, by itself, distinguish the two directional stopping outcomes.

A sign convention does not determine whether the boundaries are statistically calibrated, whether the model’s standard error is suitable, or what scientific conclusion is justified. Those questions depend on design and analysis choices beyond sign translation. In particular, the algebraic equivalence under contrast reversal does not remove the need to use the intended boundaries at each planned look, nor does a crossing alone establish a causal effect or clinical importance. It gives a consistent way to express the same boundary decision under different contrast orientations.

Common mistake: “The statistic changed from positive to negative, so the evidence changed from an upper crossing to a lower crossing and the stopping conclusion is different.” The correction is that reversing the contrast changes the statistic’s sign and requires the boundary transformation $u_k^*=-l_k,\; l_k^*=-u_k$. When both are transformed, the corresponding decision is the same; only its upper/lower label and directional wording change.

Check Your Understanding

For each item, use the stated contrast orientation and compare the signed statistic with the supplied boundaries. Solutions follow the questions.

At a look, $Z_k=2.05$, $u_k=2.10$, and $l_k=-2.20$. Is there an upper crossing, a lower crossing, or continuation?
A contrast is treatment minus control, and larger outcomes are favorable. The estimate is $-1.2$ units with $SE=0.6$ units. What is the signed Z statistic, and which direction does its sign indicate relative to the null value zero?
If the original boundaries are $u_k=2.30$ and $l_k=-1.90$, what are the boundaries after reversing the contrast?
An original analysis has $Z_k\leq l_k$. Is that same decision an upper or lower crossing after reversing the contrast? State the transformed boundary it uses.

Solutions. (1) Since $-2.20<2.05<2.10$, the result is in the continuation region. (2) $Z_k=(-1.2-0)/0.6=-2.00$; the negative sign means the estimate is below the null in the treatment-minus-control orientation, which is the unfavorable direction under the stated endpoint definition. (3) $u_k^*=-l_k=1.90$ and $l_k^*=-u_k=-2.30$. (4) It becomes an upper crossing, because $Z_k^*=-Z_k\geq-l_k$; the transformed upper boundary is $u_k^*=-l_k$.