Choosing a Boundary Is a Design Decision
A classical group sequential boundary determines how strong the evidence must be at each planned analysis before a trial can stop for efficacy. The choice is not simply between a “strict” and a “lenient” test. It is a choice about where across the analysis schedule to place the evidence requirement, subject to controlling the trial’s overall Type I error.
Pocock boundaries use similar critical values at each look. O’Brien-Fleming boundaries require very strong evidence early and become less stringent near the final analysis. The Wang-Tsiatis family offers shapes between these patterns. The appropriate choice depends on what the trial prioritizes: the opportunity to act on early evidence, the evidence required at the end, or the practical consequences of repeated interim review.
For the comparisons below, suppose the trial has \(K\) planned looks at information fractions \(0<t_1<\cdots<t_K=1\), uses a one-sided efficacy test, and analyzes a canonical group-sequential statistic \(Z_k\). Under \(H_0\), the statistics have standard normal marginal distributions and the usual canonical joint dependence arising from accumulating information. The treatment effect and variance assumptions used to construct that model must be appropriate, and the look schedule must be specified. A design’s critical values \(b_k\) are calibrated together so that the probability of crossing at least one efficacy boundary under \(H_0\) is the target \(\alpha\).
This is an overall, not a per-look, error requirement. A critical value cannot be transferred uncritically to a different number or timing of looks, a different test direction, or a different target \(\alpha\). The discussion here concerns efficacy-boundary selection, not the design of a futility rule.
What the Three Families Prioritize
The design choice is clearest when expressed as a trial priority rather than as a vague preference for “early stopping.” Under Pocock, the critical values are approximately constant across the planned looks. This makes the early efficacy threshold less extreme than under an O’Brien-Fleming design, but the final threshold is generally more stringent than the corresponding O’Brien-Fleming final threshold. Pocock can therefore be attractive when the trial values a reasonably similar evidence bar at each scheduled review.
O’Brien-Fleming boundaries place a very high bar on early efficacy stopping and relax that bar as information accumulates. This can be appealing when an early stop should require exceptionally compelling evidence, while avoiding as much penalty at the final analysis. It does not mean that early stopping is impossible; it means that an early crossing must be unusually strong.
The Wang-Tsiatis family provides a continuum of shapes. In the parameterization used here, the boundary has the form \(b_k=c\,t_k^{\delta-1/2}\), with \(c\) calibrated to the planned design and \(\alpha\). Values of \(\delta\) nearer zero give an O’Brien-Fleming-like shape; values nearer one-half give a flatter, Pocock-like shape. The parameter controls relative shape, while calibration determines the scale \(c\).
These are tendencies in boundary geometry, not guarantees about when a trial will stop. Actual stopping depends on the observed sequence of statistics, which are correlated across looks, as well as on the chosen boundary. A flatter threshold can make crossing at an early look less demanding, but it cannot ensure that early evidence will arise.
A Short Justification for Comparing Shapes
For a Wang-Tsiatis design, compare the boundary at an interim information fraction \(t\) with its value at the final analysis. Since \(t_K=1\), the final boundary is \(b(1)=c\). Dividing the interim boundary by this final value removes the calibration constant:
For example, at \(t=0.25\), \(\delta=0\) gives a relative boundary of \(0.25^{-1/2}=2\); \(\delta=0.25\) gives \(0.25^{-1/4}=\sqrt{2}\approx1.414\); and \(\delta=0.5\) gives \(0.25^0=1\). Thus the interim-to-final contrast is largest for the O’Brien-Fleming-like shape and disappears for the flat Pocock-like shape. This calculation justifies comparing the shape through \(\delta\), but it does not supply calibrated critical values: \(c\) still has to be chosen for the overall error requirement and the actual look schedule.
Worked Examples for Stated Priorities
Worked Example: Prioritizing a Similar Evidence Bar at Each Look
A hypothetical trial has four equally spaced information looks and a one-sided overall \(\alpha=0.025\). Its team wants the evidence requirement to be similar at interim and final analyses, rather than reserving most of the leniency for the final look. Which family best matches this stated priority?
A Pocock design is the direct match because its critical values are approximately constant across looks. For this four-look schedule and target \(\alpha\), the calibrated one-sided Pocock critical value is approximately \(2.36\) at each look. The overall Type I error is controlled by calibrating this common value under the joint distribution of the four statistics; it is not obtained by assigning a separate \(0.025\) error rate to each look.
For illustration, suppose the observed statistic at the second look is \(Z_2=2.40\). The comparison is \(2.40>2.36\), so the Pocock efficacy boundary is crossed at that look. This conclusion concerns the specified hypothetical design and its stated calibration. It does not establish that the trial should stop regardless of clinical context or other prespecified procedures.
An O’Brien-Fleming shape would not match the stated preference as closely: its early threshold is deliberately more demanding than its final threshold. A Wang-Tsiatis shape could be selected if the team wanted an intermediate pattern, but “similar evidence bar at each look” points most directly to Pocock.
Worked Example: Requiring Especially Strong Early Evidence
A hypothetical confirmatory trial has four planned looks. Its committee states that it is willing to continue to the final analysis unless interim evidence is exceptionally strong; it also wants the final critical value to be relatively less stringent than the early ones. The trial plans a one-sided test at overall \(\alpha=0.025\). Which boundary shape is the more natural starting point?
O’Brien-Fleming is the natural starting point because its early critical values are much higher than its final critical value. This directly encodes the committee’s stated preference: an early efficacy conclusion requires a large statistic, while evidence accumulated by the final analysis is assessed against a less extreme boundary.
For a concrete decision calculation, suppose the calibrated first-look critical value for the actual design is \(b_1=4.00\), and the observed first-look statistic is \(Z_1=3.20\). Then \(3.20<4.00\), so the first-look efficacy boundary is not crossed. The comparison alone does not say whether the result is statistically unpromising or whether the trial should continue under other monitoring rules; it only answers the efficacy-boundary question.
The value \(4.00\) here is an explicitly hypothetical first-look cutoff for illustrating the decision, not a universal O’Brien-Fleming value. In an actual design, the calibrated cutoff depends on the information schedule, number of looks, and target error level. The team should not adopt a remembered early cutoff without verifying that those design conditions match.
Worked Example: Choosing an Intermediate Wang-Tsiatis Shape
A hypothetical trial team wants early evidence to face a higher bar than final evidence, but regards the extreme early-to-final contrast of an O’Brien-Fleming-like shape as too strong. It considers a Wang-Tsiatis shape with \(\delta=0.25\) and an interim analysis at \(t=0.25\). How does this shape compare with the two endpoints?
The relative interim boundary is \(t^{\delta-1/2}\). Substituting \(t=0.25\) and \(\delta=0.25\) gives
Thus, before calibration of the scale, the interim boundary is about \(1.414\) times the final boundary. For comparison, the corresponding ratios are \(2\) when \(\delta=0\) and \(1\) when \(\delta=0.5\). The candidate \(\delta=0.25\) therefore has an intermediate shape: it demands more relative evidence early than at the end, but is less steep than the \(\delta=0\) endpoint.
This calculation supports choosing an intermediate shape to investigate, not declaring a final design. The constant \(c\) must still be calibrated so the complete design controls overall \(\alpha\) for the specified looks. The team should then check whether the resulting actual critical values meet its operating and clinical priorities.
A Practical Selection Process
A disciplined choice begins by translating broad preferences into comparisons across the analysis schedule. The following sequence helps prevent a boundary family from being chosen solely by name or by a single remembered cutoff.
State the test direction, target overall \(\alpha\), planned information fractions, and number of looks. These conditions determine the calibration problem.
Decide whether early efficacy stopping should require exceptional evidence, whether similar thresholds across looks are desirable, or whether an intermediate profile is preferred.
O’Brien-Fleming fits a strong early-evidence requirement; Pocock fits a relatively flat threshold profile; Wang-Tsiatis can represent intermediate shapes.
Calculate the critical values for the proposed schedule and verify overall Type I error. Compare the resulting thresholds at every look, not only the first and last.
Confirm that review timing, data availability, decision governance, and the consequences of stopping align with the design. Boundary shape alone does not settle those issues.
Comparisons, Diagnostics, and Limitations
| Design feature | Pocock | O’Brien-Fleming | Wang-Tsiatis |
|---|---|---|---|
| Typical boundary profile | Approximately flat across looks | Very high early; relaxes toward the final analysis | Flexible profile between familiar shapes |
| Priority it naturally matches | Similar evidence requirement at each review | Exceptional evidence required for early stopping | A chosen intermediate early-to-final contrast |
| Important caution | Final threshold can be more stringent than with O’Brien-Fleming | Early crossing is possible but requires very strong evidence | The shape parameter does not itself calibrate overall \(\alpha\) |
A useful diagnostic is to inspect the calibrated \(b_k\) values against the intended priority. If a design intended to demand exceptional early evidence has a profile that does not show a substantially higher early threshold, recheck the parameterization, look schedule, and calibration. If a supposedly flat design has noticeably varying cutoffs, first determine whether the schedule or the chosen family actually implies that profile rather than assuming a computational error.
Other warning signs are using a cutoff calibrated for a different number of looks, treating a one-sided cutoff as if it applied to a two-sided design, or reading nominal p-values at separate looks as though they were the overall Type I error. Calibration should be checked using the joint behavior of the statistics across all planned analyses. If the model assumptions or timing plan change, the design should be reassessed rather than relying on its former cutoffs.
Boundary selection also has limits. It does not determine the probability of stopping early under a particular treatment effect; that requires considering the assumed effect and the joint distribution of the statistics. Nor does a boundary specify whether stopping early is clinically appropriate, whether the available data are mature enough, or how non-efficacy monitoring decisions are made. These matters should inform the design, but they are not encoded by the efficacy boundary alone.
Common Mistake
Check Your Understanding
Answer each question, then compare with the brief solutions.
- A trial team wants early efficacy stopping to require unusually strong evidence and prefers a less stringent final boundary. Which classical shape is the most natural candidate?
- For a Wang-Tsiatis design with \(\delta=0.25\), calculate the interim-to-final boundary ratio at \(t=0.25\).
- Why is a Pocock critical value not automatically valid when the planned number of looks changes?
- A hypothetical calibrated Pocock boundary is \(2.36\) at a look, and the observed statistic is \(2.31\). Has the efficacy boundary been crossed?
- Does choosing a flat boundary guarantee that a trial will stop early? Explain.
Solutions. (1) O’Brien-Fleming, because it requires stronger evidence early and relaxes the requirement toward the final analysis. (2) \(0.25^{0.25-0.5}=0.25^{-0.25}=\sqrt{2}\approx1.414\). (3) The overall Type I error depends on the joint crossing probability across all looks, which changes with the schedule; the critical value must be recalibrated. (4) No: \(2.31<2.36\), so the boundary is not crossed. (5) No. A flat boundary describes the critical values, not the observed statistics; early stopping occurs only if an observed statistic crosses the boundary.