Two-Sided Group Sequential Decision Regions

Why Two-Sided Decisions Have Three Regions

In a group sequential test, data are examined at planned analysis looks rather than only once at the end. For a two-sided test of a point null, evidence in either direction can support rejection: the estimate may be sufficiently above the null value or sufficiently below it. At an interim look, however, neither tail may have been reached. That possibility is the continuation region: the test has not yet met a rejection rule and proceeds to another planned look.

These regions describe the decision made from the statistic at a particular look. They do not, by themselves, specify how likely the test is to reject across all looks. The latter depends on the joint behavior of the statistics over time and on the complete boundary sequence.

The lower and upper limits need not have the same absolute value. The definition only requires that the lower limit be below the upper limit. Which numerical direction of a particular statistic supports which effect depends on how that statistic is defined; here, the direction has been stated explicitly to make the region labels unambiguous.

At an interim look, a statistic in either rejection region calls for stopping and rejecting \(H_0\), in the corresponding direction. A statistic in the continuation region calls for collecting data to the next planned look, provided another look remains. At the final look, there is no next look: a statistic between the limits does not lead to rejection under this rule. It is a final nonrejection, not acceptance or proof of \(H_0\).

From Look-Specific Regions to the Sequential Rule

A look-specific region classifies one value of \(Z_k\). A group sequential decision, in contrast, depends on the order in which the data cross the limits. Define the stopping look as the first look at which either limit is reached. If neither limit is reached before the final look, the test ends at look \(K\).

If this set is empty among the interim looks, set \(\tau=K\). Rejection at the stopping look occurs when \(Z_\tau\le l_\tau\) or \(Z_\tau\ge u_\tau\). If the test reaches look \(K\) without crossing either limit, it does not reject. This description assumes the limits are prespecified and the analysis is conducted at the planned looks; changing the looks or the rule after seeing results changes the procedure that is being applied.

For example, a value above the upper limit at look 2 is an upper-tail crossing at that look. But if the test already crossed a limit at look 1, the look-2 value is not part of the actual stopping decision: under the rule, the test has already stopped. It is therefore useful to distinguish the look-specific events \(Z_k\ge u_k\) and \(Z_k\le l_k\) from first-crossing events, which also require that no earlier boundary was crossed.

Deriving the Partition and the First-Crossing Interpretation

The three regions follow directly from ordering the real line around two limits. Since \(l_k<u_k\), any real value \(z\) satisfies exactly one of the following: \(z\le l_k\), \(l_k<z<u_k\), or \(z\ge u_k\). These cases cannot overlap, and together they cover every possible value. Therefore, at a single look, the lower-tail rejection region, continuation region, and upper-tail rejection region form a disjoint partition of the statistic’s possible values.

The sequential rule adds the requirement that a crossing be the first one. At look \(k\), an upper-tail first crossing requires the statistic to be in the upper region at \(k\) and in the continuation region at every earlier look. The same logic applies to a lower-tail first crossing. In symbols, for \(k>1\), the upper first-crossing event is:

Replacing \(Z_k\ge u_k\) with \(Z_k\le l_k\) gives the lower first-crossing event. At the first look, there are no earlier looks to check. This justification explains why a set of boundaries alone is not a complete description of the observed stopping decision: the path through the earlier looks matters.

Applying the Regions to an Estimate and Standard Error

A common form of signed statistic is an estimate’s distance from the null value, measured in standard-error units. For example, when the normal approximation is appropriate, one may use \(Z_k=(\hat{\theta}_k-\theta_0)/SE_k\), where \(\hat{\theta}_k\) is the estimate available at look \(k\). The units in the numerator and standard error must agree, so the resulting statistic is dimensionless. The boundary comparison is then made on the standardized scale, not directly on the estimate scale.

Diagnostics, Comparisons, and Limitations

Several checks help prevent a correct-looking classification from representing the wrong procedure:

• Check the ordering of limits. The continuation interval exists only when \(l_k<u_k\). Reversed or equal limits indicate a malformed set of three regions.
• Check the statistic’s direction. Confirm that the definition of \(Z_k\) matches the stated meaning of upper and lower. A reversed sign swaps which tail supports a positive effect.
• Check boundary inclusion. This tutorial uses inclusive rejection inequalities, \(Z_k\le l_k\) and \(Z_k\ge u_k\), and a strict continuation interval. For a continuous statistic, exact equality usually has probability zero; for a discrete statistic, equality conventions can change decisions and must be specified.
• Check the whole path. A later crossing is not the stopping decision if an earlier look already crossed a limit. Keep the observed first crossing, not merely the most extreme value, as the decision point.
• Check calibration separately. Correctly classifying a statistic does not show that the boundaries control the desired overall Type I error. Repeated looks produce dependent evidence, so the complete design needs appropriate sequential calibration.

A fixed-sample two-sided test has two rejection tails at one final analysis, but no interim continuation decisions. A one-sided group sequential test uses a rejection tail in only the direction specified by its alternative; the other side is not automatically a second rejection region. In contrast, the two-sided rule defined here has both tails as rejection regions and an interval between them for continuation at interim looks.

A lower-tail rejection region should not be confused with a futility region. Both can be located on the lower side of a statistic’s scale, but they represent different decisions: lower-tail rejection supports the alternative in the lower direction, whereas a futility rule is a reason to stop without rejecting for insufficient evidence of benefit under the design’s specified objective. If a design uses additional futility limits, it must state how those limits combine with the two efficacy rejection regions; the three-region definition above concerns the two-sided rejection rule itself.

Finally, the regions are not enough to calculate power, an overall error rate, or the probability of stopping at a particular look. Those quantities require a model for the joint statistics across looks, the complete boundary sequence, and any additional stopping rules. A useful diagnostic is to distinguish the question “Where does this look’s statistic fall?” from “What is the probability that the full procedure first stops here?” The first is answered by the inequalities; the second is an operating-characteristic calculation.