Re-Estimation With Group Sequential Boundaries

Why Sample-Size Changes Affect Sequential Decisions

In a group sequential trial, efficacy and sometimes futility are assessed before the planned final analysis. The trial’s stopping boundaries are set in advance to control the chance of a false efficacy conclusion while allowing earlier decisions when the evidence is strong. A sample-size re-estimation can change how much information will be available at a later analysis. That can change the joint distribution of the test statistics on which the boundary calibration depended.

The key distinction is between a decision already made at an unchanged interim analysis and a decision planned for a later analysis whose timing or information has shifted. If an efficacy boundary has been crossed at the prespecified look, a later sample-size increase does not undo that crossing. But keeping the same numerical threshold at a later look does not, by itself, establish that the original overall error control still holds.

Definition: A group sequential boundary is a prespecified threshold for a test statistic at one or more interim analyses. In a one-sided efficacy test where positive values favor treatment, an upper efficacy boundary \(u_j\) at look \(j\) calls for stopping and declaring efficacy if \(Z_j \geq u_j\). A lower futility boundary \(l_j\) may call for stopping for lack of sufficient promise if \(Z_j \leq l_j\). The protocol must specify which boundaries are binding and what action follows crossing each one.

Here \(Z_j\) denotes the standardized treatment comparison at look \(j\). To discuss how boundaries are calibrated, use the usual canonical approximation: under the null hypothesis \(H_0\), each \(Z_j\) has an approximately standard normal distribution, and the statistics across looks have the dependence implied by accumulating information. This is an approximation for many large-sample designs, not a universal model. The actual test, outcome distribution, analysis method, and design assumptions must support its use.

In a simple setting with equal allocation, a normally distributed outcome, and known variance \(\sigma^2\), information is proportional to the number of participants with observed outcomes. If \(I_j\) denotes information at look \(j\), then the canonical joint distribution has

\operatorname{Corr}(Z_i,Z_j)=\sqrt{\frac{I_i}{I_j}}, \qquad i<j.

This dependence matters: a boundary sequence is calibrated for the joint chance of crossing an efficacy boundary at any look, not just for the probability of crossing a threshold at one isolated analysis. Re-estimating the sample size can change the planned information at the final look and therefore the joint distribution used for calibration.

What a Prespecified Boundary Does—and Does Not—Guarantee

Suppose a trial has two planned looks, at interim information \(I_1\) and final information \(I_2\), with upper boundaries \(u_1\) and \(u_2\). The overall one-sided type I error is the probability, assuming \(H_0\) is true, of crossing an efficacy boundary at either look:

P_{H_0}(Z_1 \geq u_1 \text{ or } Z_2 \geq u_2).

A design chooses the boundaries so that this crossing probability meets the prespecified significance level \(\alpha\), subject to its stated assumptions. It is not generally enough to check each look separately. Nor is it safe to assume that a boundary sequence calibrated for one information schedule remains calibrated after that schedule changes.

A sample-size change made after an interim analysis does not alter the data or test statistic already observed at that analysis. If the interim analysis occurs at its planned time and uses the prespecified statistic and boundary, its decision rule remains the rule specified for that look. The issue is what happens next: the final look may occur at a different information level, an analysis may be added or removed, or the future statistic may be interpreted using a boundary calibrated for a different plan.

One useful justification follows directly from the canonical model. Write the score-like process at information \(I_j\) as \(S(I_j)\), with independent increments and variance \(I_j\) under \(H_0\). Then \(Z_j=S(I_j)/\sqrt{I_j}\). For \(I_i<I_j\), the covariance of the two standardized statistics is

\operatorname{Cov}(Z_i,Z_j) = \frac{\operatorname{Cov}(S(I_i),S(I_j))}{\sqrt{I_i I_j}} = \frac{I_i}{\sqrt{I_i I_j}} = \sqrt{\frac{I_i}{I_j}}.

The middle equality holds because the later score includes the earlier score plus an independent increment, so their covariance is the variance \(I_i\) of the earlier score. If the planned final information changes, this correlation changes. Since the probability of crossing any boundary depends on the joint distribution, an unchanged list of numerical cutoffs does not automatically imply unchanged type I error. A design-specific justification or recalibration may be needed.

Worked Examples

Worked Example: An Interim Efficacy Boundary Is Crossed

Hypothetical design. A one-sided trial has a planned interim analysis after 120 participants have complete outcomes and a final analysis after 240. The prespecified interim efficacy boundary is \(u_1=2.90\). At the interim analysis, the estimated treatment effect is 4.50 points, with standard error 1.45 points. Assume the normal test statistic is appropriate under the design.

The interim statistic is

Z_1=\frac{4.50\text{ points}}{1.45\text{ points}}=3.103.

Since \(3.103 \geq 2.90\), the prespecified efficacy boundary is crossed. The design’s rule is to stop for efficacy. A proposed increase in the final sample size is not applied: the trial has already met its stated stopping criterion. This example assumes that the look, test statistic, and boundary were implemented as prespecified; it does not establish that an arbitrary post hoc boundary would control error.

Worked Example: A Binding Futility Boundary Takes Priority

Hypothetical design. At the same planned interim look, suppose the protocol specifies a binding futility boundary \(l_1=0.30\), meaning that the trial stops for futility when \(Z_1 \leq 0.30\). The interim estimated treatment effect is 0.20 points and its standard error is 0.80 points. A sample-size rule would otherwise recommend increasing the target after a low conditional-power result.

The observed statistic is

Z_1=\frac{0.20\text{ points}}{0.80\text{ points}}=0.250.

Because \(0.250 \leq 0.30\), the binding futility condition is met. Under the stated protocol, the trial stops; it does not proceed to increase the sample size. The sample-size recommendation cannot override a binding boundary unless the design explicitly permits that action. If the futility boundary had instead been designated nonbinding, the protocol would need to state whether stopping is optional and how a decision to continue affects the planned analyses.

Worked Example: A Larger Final Sample Changes the Calibration Problem

Hypothetical design. A two-look trial plans an interim analysis after 120 participants and a final analysis after 240 participants, all with complete outcomes. The efficacy boundaries \(u_1=2.90\) and \(u_2=2.05\) were prespecified and calibrated for that plan. At the interim look, \(Z_1=1.10\), so neither the efficacy boundary nor a hypothetical binding futility boundary of \(0.30\) is crossed. A permitted re-estimation increases the final target to 300 participants.

Under the simple equal-information-per-participant assumption, the planned interim-to-final information ratio was \(120/240=0.500\). After the increase, it is \(120/300=0.400\). The canonical correlation between the interim and final statistics would therefore be approximately

\sqrt{0.500}=0.7071 \qquad\text{under the original plan, compared with}\qquad \sqrt{0.400}=0.6325 \text{ after the increase.}

Suppose the observed final estimate is 0.104 points with standard error 0.0500 points. Then \(Z_2=0.104/0.0500=2.080\), which exceeds the old numerical final boundary \(2.05\). That arithmetic shows a boundary crossing under the old cutoff; it does not prove that the original overall type I error is preserved under the changed design. The final boundary was calibrated for a different information schedule. The re-estimated design needs a valid, prespecified justification for its final testing rule before that crossing can support a confirmatory conclusion.

Diagnostics and Common Failure Modes

Before interpreting a post-re-estimation boundary crossing, check the actual analysis path against the design assumptions. A useful review asks whether the same interim look occurred at the planned information, whether the same statistic and efficacy direction were used, whether the final analysis is at its planned information, and whether any additional analyses were introduced. It also checks whether futility boundaries are binding or nonbinding and whether the implemented stopping action matches the protocol.

Moved or added looks. An extra efficacy analysis creates another opportunity to cross a boundary. It cannot simply be treated as if it were one of the original looks unless its role is covered by the design’s error-control plan.
A different final information level. A larger or smaller sample can change the joint distribution of interim and final statistics. Reusing the original final cutoff without a design-specific argument is not sufficient evidence of preserved error control.
Changing a threshold after seeing interim results. Choosing a more favorable cutoff because the observed result is close to the old boundary makes the decision rule data-dependent. That choice must be accounted for in the design’s error control; it is not a neutral operational adjustment.
Ignoring the type of futility rule. A binding rule is part of the required stopping procedure. A nonbinding rule may be ignored under its stated design, but crossing it does not automatically make continuing correct or incorrect; the prespecified procedure governs.
Confusing sample size with information. Participant count is a useful proxy only under suitable assumptions. Incomplete outcomes, changing variance, or unequal information per participant can mean that the realized information does not match a simple count-based calculation.

A frequent mistake is: “The final \(Z\)-statistic exceeded the same numerical cutoff as before, so the trial still has the original significance level.” The correction is that the original significance level applies to a calibrated sequence of analyses under stated assumptions. When the sequence or information changes, check whether that calibration still applies; do not infer it from a single final statistic.

How the Main Approaches Differ

Approach	What it does	Main caution
Keep the planned sample size and look schedule	Uses the original boundary sequence under its planned analysis schedule.	It may not meet the revised precision or power goal that prompted re-estimation.
Change the sample size and use the original cutoffs	Retains numerical boundaries while changing the amount of final information.	Validity is not automatic; the revised joint testing procedure needs justification.
Recalibrate or otherwise justify the revised procedure in advance	Specifies boundaries and stopping rules for the permitted adaptation.	The justification must match the adaptation rule, analysis timing, and assumptions actually used.
Add an unadjusted efficacy look	Tests at another time without incorporating it into the planned sequence.	It adds an opportunity to reject \(H_0\) and can invalidate the intended type I error control.

These approaches are not interchangeable. A group sequential design is more than a set of appealing cutoffs: it is a rule linking the analyses, stopping decisions, and error rate. A prespecified re-estimation procedure should state what changes are allowed, which boundary decisions take priority, and how the future efficacy test remains interpretable. The precise method for maintaining a final test after adaptation is a design-specific matter; the central point here is that the original boundary calibration must not be presumed to survive a change.

Key takeaway: A sample-size update does not erase an interim decision already made at a prespecified look. But a changed future information schedule can change the joint distribution underlying group sequential boundaries. Treat later cutoffs as valid only when the revised analysis path is covered by a prespecified, justified error-control procedure.

Check Your Understanding

Answer each question, then compare your reasoning with the short solutions below.

A trial’s binding interim futility rule is \(Z_1 \leq 0.40\). The interim statistic is \(0.35\), and a separate rule recommends doubling the sample size. What should happen under the stated protocol?
Why can changing the planned final sample size affect the calibration of an unchanged final boundary?
In the simple canonical model, interim information is 100 units and final information is 400 units. What is the correlation between the two standardized statistics?
A team adds an unplanned efficacy analysis but uses the original final cutoff at both analyses. Why is that not automatically valid?
At an unchanged interim look, \(Z_1\) exceeds the prespecified efficacy boundary. Does a later sample-size update reverse that boundary crossing?

Solutions. (1) Stop for futility: \(0.35 \leq 0.40\), and the rule is binding, so the sample-size recommendation does not override it. (2) The final information changes the dependence between interim and final statistics; the joint probability of crossing any efficacy boundary can therefore change. (3) The correlation is \(\sqrt{100/400}=\sqrt{0.25}=0.500\). (4) The additional analysis creates another opportunity to cross an efficacy boundary and was not necessarily included in the original calibration. (5) No. The observed interim crossing remains; a later update cannot undo the decision at that look.