Choosing a Boundary Family Is a Design Decision
A classical group sequential design can stop a trial early for efficacy, but the boundary family determines how demanding that early evidence must be. With a fixed one-sided significance level, maximum information, and planned looks, a design team still has to decide how to distribute its willingness to stop across those looks. A design that makes early stopping relatively accessible may be attractive when early evidence could change practice quickly; one that makes early stopping difficult may better match a priority for strong early evidence and a large final analysis.
The choice must be made before examining interim results. It is not a way to select whichever boundary gives the more favorable decision after data have accumulated. The goal here is to compare candidate classical designs against the same constraints and explain which best fits stated priorities.
Make the Comparison on Common Ground
Candidate boundaries are comparable only if they use the same testing direction, one-sided \(\alpha\), maximum information, and planned look schedule. For the standard canonical model under \(H_0\), the look statistics have a joint multivariate normal distribution with mean zero, unit variances, and correlations
This model is appropriate when the standardized statistics have independent increments on the information scale and the planned information fractions correctly describe the looks. A fixed boundary vector can then be calibrated under the joint null distribution so the probability of crossing at least one efficacy boundary is the specified \(\alpha\). The marginal probability \(P_0(Z_i\ge c_i)\) at a particular look is not, by itself, the design’s overall type I error.
For a concrete comparison, suppose a hypothetical trial has one-sided \(\alpha=0.025\), a maximum information corresponding to 600 participants, and planned looks at 300, 450, and 600 participants. Assume information is proportional to sample size, so the information fractions are \(0.50\), \(0.75\), and \(1.00\). A numerical calibration for this schedule might report the following rounded candidate boundaries. The entries are illustrative calibrated cutoffs, not universal constants; actual cutoffs depend on the specified model and schedule.
| Candidate family | Look 1, \(t_1=0.50\) | Look 2, \(t_2=0.75\) | Final look, \(t_3=1.00\) | Typical pattern |
|---|---|---|---|---|
| Pocock-like | 2.18 | 2.18 | 2.18 | Similar stringency at each look |
| O’Brien–Fleming-like | 2.80 | 2.29 | 1.98 | Very stringent early, less stringent near the maximum information |
Both rows are understood to have been calibrated to the same overall one-sided \(\alpha\) under the same null model. Their numerical patterns reveal their different priorities: the Pocock-like design has a lower early cutoff, whereas the O’Brien–Fleming-like design requires more extreme early evidence and has a lower final cutoff.
Why Overall Alpha Depends on the Whole Path
The looks are correlated: a high statistic early in the trial tends to be associated with a high statistic later. Thus, adding the marginal tail probabilities \(P_0(Z_i\ge c_i)\) would count some possible paths more than once. Instead, partition rejection by the first look at which a boundary is crossed. These first-crossing events are disjoint, giving the exact identity
For example, at the second look the contribution to overall alpha is not simply \(P_0(Z_2\ge c_2)\). It is the probability that \(Z_2\) crosses its boundary and no earlier boundary was crossed. This is why the joint null model, rather than a list of isolated nominal cutoffs, is required for calibration.
The design choice also depends on the alternative. A boundary family that facilitates earlier stopping may reduce expected information when a beneficial effect produces early crossings. But power and expected information are operating characteristics of the whole design, not properties read off one cutoff alone. They should be evaluated under clinically relevant alternatives using a prespecified analysis method.
Worked Example: Prioritizing the Chance of an Early Stop
Suppose the trial described above has reached its first look, at 300 participants, and the hypothetical observed statistic is \(Z_1=2.40\). Determine what each candidate design would do.
For the Pocock-like design, the first-look cutoff is \(c_1=2.18\). Since \(2.40\ge2.18\), the efficacy boundary is crossed and the design stops for efficacy.
For the O’Brien–Fleming-like design, \(c_1=2.80\). Since \(2.40<2.80\), there is no efficacy crossing, so the trial continues, subject to any separately prespecified stopping rules.
This comparison makes the early-stopping tradeoff concrete. If the design priority is to make a favorable early decision possible at less extreme evidence, the Pocock-like candidate better matches that preference. This single hypothetical statistic does not justify changing a design after the trial begins; the choice should have been made before interim data were examined.
Use Stated Priorities, Not a Label
The family name alone does not determine the appropriate design. A defensible choice starts with the constraints and priorities: the maximum information cannot be exceeded, the planned schedule must be feasible, and the overall type I error must be controlled. Within those constraints, the team can compare the early-stopping behavior, power, expected information, and clinical consequences of stopping at different looks.
For instance, a sponsor might set a minimum acceptable power under a clinically relevant effect and then prefer the candidate with the lower expected sample size among those meeting that requirement. Another team might give greater weight to requiring compelling early evidence, even if that means fewer early stops. These are distinct design objectives; neither is answered by observing which cutoff is numerically smaller at the final look.
Worked Example: Applying a Prespecified Planning Criterion
Consider a second hypothetical design review with the same maximum sample size of 600. A planning simulation under a prespecified beneficial effect reports power of \(0.90\) and expected sample size of 468 for the Pocock-like candidate, and power of \(0.89\) and expected sample size of 516 for the O’Brien–Fleming-like candidate. Assume the review criterion, set in advance, is power of at least \(0.88\), followed by choosing the candidate with the smaller expected sample size.
Both candidates meet the power requirement: \(0.90\ge0.88\) and \(0.89\ge0.88\). Their expected savings relative to the maximum sample size are \(600-468=132\) participants and \(600-516=84\) participants, respectively. The Pocock-like candidate has the smaller expected sample size, 468 rather than 516, so it is preferred under this criterion.
This conclusion is conditional on the stated effect, simulation model, and criterion. It does not establish that the Pocock-like design is universally better, nor does the expected sample size replace the maximum planned sample size. The trial must still be designed to accommodate the maximum of 600 participants if it does not stop early.
Worked Example: Interpreting a Look Without Confusing Nominal and Overall Evidence
Suppose the trial reaches its second look with \(Z_2=2.24\), and no boundary was crossed at the first look. Compare the decision under the two candidate families.
The Pocock-like second-look cutoff is 2.18, and \(2.24\ge2.18\). The trial crosses that efficacy boundary and stops for efficacy. The O’Brien–Fleming-like second-look cutoff is 2.29, and \(2.24<2.29\). That boundary is not crossed, so the trial continues.
The one-sided nominal p-value associated with \(Z_2=2.24\) is approximately \(1-\Phi(2.24)=0.0125\), where \(\Phi\) is the standard normal cumulative distribution function. This describes the tail probability for a single standard normal statistic under the null. It is not the overall p-value or type I error calculation for the sequential design. The design decision above follows the calibrated boundary for the assigned family and look.
Diagnostics, Alternatives, and Limitations
Before selecting a candidate, verify that each design uses the same alpha, maximum information, look fractions, and one-sided efficacy direction. Check that the boundary vector was calibrated for the actual planned schedule, rather than borrowed from a different number or timing of looks. A discrepancy between the intended and implemented look schedule, information measure, or statistic can make a printed boundary table inappropriate.
A useful comparison should include more than cutoff values. Examine power under relevant alternatives, expected sample size or information, and the probability of crossing at each look. If a candidate’s early-stopping behavior conflicts with the trial’s practical or clinical priorities, that is a reason to reconsider it before enrollment, not a reason to reinterpret an observed interim result.
| Design feature | Pocock-like pattern | O’Brien–Fleming-like pattern |
|---|---|---|
| Early efficacy evidence | Less stringent than the O’Brien–Fleming-like candidate in this comparison | Very stringent |
| Later efficacy evidence | Cutoff remains similar across looks | Cutoff decreases as information accumulates |
| Potential design appeal | Earlier stopping may be more accessible when early evidence is favorable | Early stopping is reserved for more extreme evidence |
| What still needs evaluation | Power and expected information under relevant alternatives | Power and expected information under relevant alternatives |
These are patterns, not guarantees about a particular trial’s outcomes. Both candidates have limitations: a classical fixed-boundary design relies on its specified model and look schedule, and its cutoffs do not automatically handle unplanned looks or materially changed information timing. The comparison here concerns efficacy boundaries; any futility rules or other stopping criteria need separate specification and evaluation.
A common mistake is to compare only the final cutoffs and select the candidate with the smaller one, claiming it must be more powerful or more liberal overall. That reasoning ignores the earlier boundaries and the correlation among looks. The correction is to compare complete, calibrated boundary vectors and evaluate whole-design operating characteristics under the same model, schedule, and alternatives.
Check Your Understanding
Answer each question using the stated design constraints. Short solutions follow the questions.
- At the first look, \(Z_1=2.60\). Using the illustrative table, which candidate crosses its efficacy boundary, and why?
- Why is the overall type I error not the sum of the marginal probabilities \(P_0(Z_i\ge c_i)\)?
- A team wants very stringent evidence before allowing an early efficacy stop. Which of the two candidate patterns better reflects that priority?
- Why can a boundary table for information fractions \(0.50, 0.75, 1.00\) not automatically be used for a different look schedule?
- What is wrong with choosing between boundary families after seeing which one declares efficacy at an interim look?
Solutions. (1) The Pocock-like candidate crosses because \(2.60\ge2.18\); the O’Brien–Fleming-like candidate does not because \(2.60<2.80\). (2) The look statistics are correlated, so marginal crossing events overlap. Overall alpha is computed from disjoint first-crossing events under the joint null distribution. (3) The O’Brien–Fleming-like pattern, with its more stringent early cutoff, better reflects that preference. (4) Calibration depends on the joint distribution at the planned information fractions, so a changed schedule requires an appropriate recalibration. (5) The selection would depend on interim data and could invalidate the prespecified error control; the design choice must be made before examining those results.