From a Clinical Objective to a Statistical Design
A trial objective such as “determine whether the new treatment improves outcomes” is not yet a design specification. It does not identify which outcome defines improvement, how outcomes will be compared, how large an improvement matters, or how much evidence the design should require. Those choices must be explicit before a group sequential design can be planned.
A design specification is a compact, reviewable statement of the statistical and operational targets that a planned design must meet. It connects the protocol’s clinical objective to quantities such as maximum information, planned analysis times, type I error, and power. It is not itself a set of stopping boundaries or a software configuration: those are developed from the specification.
The specification should be traceable to the protocol. In particular, it should make clear who is being compared, what outcome is measured and when, what summary or contrast is primary, and what magnitude of effect the design is intended to detect. If any of these elements is unresolved, a numerical sample-size target can create a false impression of precision.
State the Endpoint, Contrast, and Target Effect
Begin with the endpoint as it will be analyzed, not merely a broad clinical concept. For a continuous endpoint, specify the measurement, time point, and treatment contrast, such as the difference in mean change from baseline at week 12. For a binary endpoint, specify the event definition and assessment window, and say whether the principal contrast is a risk difference, risk ratio, or another measure. For a time-to-event endpoint, specify the event, time origin, censoring rules, and the contrast, often a hazard ratio under a proportional-hazards model.
The statistical hypothesis should match that endpoint and contrast. For example, let \(\theta\) denote a treatment-minus-control mean difference. A two-sided superiority objective might test \(H_0:\theta=0\) against \(H_a:\theta\ne 0\). A one-sided objective is appropriate only when the protocol’s scientific claim and testing plan genuinely exclude effects in the other direction. A smaller p-value threshold is not a substitute for a clearly stated directional claim.
The target effect is the effect at which the design’s power goal is evaluated. It should be clinically interpretable and expressed on the endpoint’s scale. It is not automatically the effect most likely to occur, nor is it necessarily the smallest effect that would be clinically worthwhile. The specification should distinguish the null value, the target effect, and any clinical threshold used to justify treatment.
Maximum Information and the Analysis Schedule
Information describes the precision available for estimating the primary treatment contrast. Its units depend on the endpoint and statistical model. For some outcomes, a planned sample size is a useful proxy for information; for time-to-event outcomes, the number of observed events is often more directly relevant. The specification should say which quantity defines the maximum information and what assumptions connect it to recruitment and follow-up.
Let \(I_{\max}\) be the planned maximum information. If there are \(K\) analyses, let \(I_k\) be the information available at analysis \(k\), and define its information fraction as \(t_k=I_k/I_{\max}\). The final analysis has \(t_K=1\). An analysis schedule may therefore be stated using information fractions, or using endpoint-relevant quantities such as event counts, provided the mapping between them is clear.
The schedule also needs an operational interpretation. A target of \(t_k=0.50\) does not, by itself, tell a trial team when to prepare an interim report. For an event-driven trial it may correspond to a target number of observed events; for a fixed-follow-up endpoint it may be tied to the number of participants with completed assessments. Accrual, follow-up, and data-cleaning delays affect calendar timing, but they should not be confused with the information target.
A group sequential design controls the planned overall type I error across its analyses, rather than treating each look as a fresh test at level \(\alpha\). Power is the chance of rejecting \(H_0\) at one or more planned analyses under the target effect and the specified design assumptions. Consequently, an error-rate and power statement must refer to the full planned sequence of analyses, including the final analysis.
Worked Examples
Worked Example: A Continuous Primary Endpoint
Suppose a hypothetical two-arm trial compares mean symptom-score improvement at week 12. Lower scores are better, and the primary contrast is treatment minus control. The protocol team regards a 5-point greater mean improvement as the target effect, and plans a two-sided test with overall \(\alpha=0.05\) and 90% power. Assume independent observations, equal allocation, a common standard deviation of 10 points, and a normal approximation.
First calculate a fixed-sample reference size. With \(m\) participants per arm, the variance of the difference in sample means is \(2\sigma^2/m\). The usual normal-approximation calculation gives
where \(\delta=5\) points, \(\sigma=10\) points, \(\beta=1-0.90=0.10\), \(z_{0.975}=1.960\), and \(z_{0.90}=1.282\). Thus
Rounding up gives 85 participants per arm, or 170 total, as a fixed-sample reference. Now suppose a separately calibrated group sequential design requires a hypothetical 2% increase in maximum sample size relative to this rounded reference. Then \(170(1.02)=173.4\), so a balanced operational target is 174 participants, 87 per arm. Under the equal-allocation normal model, information is \(I=N/(4\sigma^2)\), so this maximum corresponds to \(I_{\max}=174/(4(10^2))=0.435\) per squared score point.
A specification could schedule analyses at \(t=0.50, 0.75,\) and \(1.00\), corresponding approximately to 87, 131, and 174 participants with completed primary assessments. These are information targets, not a claim that precisely those numbers will be available on particular calendar dates. The 2% increase is an explicit hypothetical planning input; a real design must obtain its maximum information from calibration of the chosen sequential design.
Worked Example: A Time-to-Event Endpoint
Consider a hypothetical trial of time to first hospitalization, comparing treatment with control under equal randomization. Suppose the target hazard ratio is 0.70, the proportional-hazards model is considered appropriate for planning, censoring is noninformative, and the log-rank comparison is the primary analysis. With two-sided \(\alpha=0.05\) and 80% power, the usual event-count approximation is
where \(D\) is the total number of events and \(q=0.50\) is the treatment allocation fraction. Here \(\log(0.70)\approx -0.3567\), \(z_{0.975}=1.960\), and \(z_{0.80}=0.8416\). Substitution gives
The fixed-sample reference is therefore about 247 events. Suppose, for illustration, that a calibrated sequential design calls for a 3% increase in maximum events. Then \(247(1.03)=254.41\), which rounds up to a maximum target of 255 events. Analyses at information fractions \(0.50, 0.75,\) and \(1.00\) can be specified approximately at 128, 191, and 255 observed events.
The event target does not determine the number of participants or the calendar duration by itself. Those also depend on the accrual pattern, follow-up, and event and censoring rates. Thus the specification should name 255 events as the maximum information target and document the assumptions used to estimate the recruitment and follow-up needed to reach it.
Worked Example: A Binary Endpoint and an Ambiguous Objective
A hypothetical protocol says, “The treatment should reduce complications.” To make this designable, the team defines the primary endpoint as at least one complication by day 30 and specifies the treatment-minus-control risk difference. Assume independent binary outcomes, equal allocation, complete day-30 follow-up, control risk \(p_C=0.30\), target treatment risk \(p_T=0.20\), two-sided \(\alpha=0.05\), and 80% power.
For a normal-approximation reference calculation, let \(\bar p=(p_C+p_T)/2=0.25\). The approximate required number per arm is
Using \(z_{0.975}=1.960\) and \(z_{0.80}=0.8416\), the two terms in brackets are \(1.960\sqrt{0.375}\approx1.200\) and \(0.8416\sqrt{0.37}\approx0.512\). Therefore
Rounding up gives 294 participants per arm, or 588 total, as a fixed-sample reference. If the hypothetical sequential design requires a 4% maximum-sample increase, \(588(1.04)=611.52\), so the balanced maximum is 612 participants, 306 per arm. A schedule at information fractions \(0.50, 0.75,\) and \(1.00\) would correspond approximately to 153, 230, and 306 participants per arm with completed day-30 outcome data.
The resulting specification is more precise than “reduce complications”: it names the event definition, time window, contrast, target risks, maximum evaluable sample, analysis schedule, and error-rate and power goals. If follow-up is incomplete or the target risks change, the sample-size calculation and operational plan must be revisited.
Justifying the Information Schedule
Information fractions are useful because they describe the relative precision available at a look, rather than only the fraction of calendar time or participants enrolled. For a simple normal contrast with known variance, information is the reciprocal of the variance of its estimator. In the continuous-outcome example, equal group sizes give
This explains why, under those assumptions, doubling the total sample size doubles information. The same simple proportionality need not hold for every endpoint or analysis. For a time-to-event endpoint, event count is often a practical information measure; for binary outcomes, the information contributed depends in part on event rates. A schedule stated only as “at 50% of enrollment” may therefore fail to represent 50% of the planned statistical information.
The information calculation is a planning model, not a guarantee that each interim data set will have exactly the assumed precision. The specification should say how actual analysis timing will be determined if information accrues differently than expected, and which endpoint or event-count rule governs the look.
Assumptions, Diagnostics, and Alternatives
A usable specification makes its assumptions visible. Check that the endpoint definition and analysis population are operationally clear; the target effect is on the same scale as the stated contrast; and the assumed variability, event rate, or allocation ratio is plausible for planning. For repeated assessments, time-to-event outcomes, or incomplete follow-up, state the model assumptions that support the chosen information measure. When several primary endpoints, treatment comparisons, or claims are present, identify how the overall error rate is controlled across those tests as well as across repeated looks. A group sequential plan for one hypothesis does not automatically control multiplicity across several hypotheses.
Warning signs include a target effect that has no clinical justification, a maximum sample size with no stated endpoint or information units, interim looks described only by calendar dates, and power reported without naming the effect at which it is calculated. Another warning sign is treating \(\alpha\) as a per-look threshold: using that same nominal level repeatedly does not, in general, preserve the intended overall type I error. The error-rate goal must apply to the complete planned analysis sequence.
The main alternative is a fixed-sample design with one planned primary analysis. It may be simpler to specify and operate, but it does not provide the same planned opportunities for interim decisions. A group sequential design adds interim analyses while controlling the overall error rate according to its design. The maximum information needed can depend on the chosen sequential plan, so the fixed-sample reference calculations in the examples are starting points, not substitutes for sequential design calibration.
A specification also has limits. It cannot make a poorly chosen endpoint clinically relevant, remove uncertainty in nuisance parameters, or guarantee that recruitment and follow-up will match projections. Nor does it by itself establish the stopping boundaries, monitoring procedures, or data-handling rules needed for implementation. Its purpose is to make the design objective explicit and checkable before those details are finalized.
Common Mistake
Wrong reasoning: “We want 90% power, so we have specified the design.” Power is incomplete without the target effect, endpoint, analysis contrast, type I error, maximum information, and analysis schedule. Two designs can both claim 90% power while targeting different effects or using different error-rate and interim-analysis plans.
Correction: State power as a probability under a particular alternative and the full planned analysis process. For example: “The design targets 90% power to detect a 5-point treatment-minus-control difference in mean week-12 score, with two-sided overall \(\alpha=0.05\), at three analyses scheduled at information fractions 0.50, 0.75, and 1.00.” Then state the endpoint model and maximum information assumptions that support the calculation.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A protocol states “test whether treatment improves survival.” Name two details that must be specified before this is a complete endpoint and contrast.
- For a design with analyses at information fractions 0.40, 0.70, and 1.00, what is the final analysis fraction, and why are these fractions not necessarily fractions of calendar time?
- A binary-outcome design has a fixed-sample reference of 400 participants. A hypothetical calibrated sequential design uses 5% more maximum sample size. What maximum sample size results after rounding up?
- Why is it incorrect to interpret 80% power as an 80% probability that the treatment is effective?
- A trial has two primary hypotheses and several interim looks. Does specifying the overall error rate across looks for one hypothesis automatically control the error rate across both hypotheses? Explain.
Solutions. (1) Specify the survival endpoint, such as time to death or time to a defined event, and the time origin and analysis contrast, such as a hazard ratio or difference in survival at a stated time. (2) The final fraction is 1.00 because it represents maximum planned information. Calendar time can advance at a rate different from information, which depends on outcome and event accumulation. (3) \(400(1.05)=420\), so the maximum is 420 participants. (4) Power is the probability of rejecting \(H_0\) under a specified alternative and design; it is not a probability assigned to the hypothesis after seeing data. (5) No. Multiplicity across hypotheses requires its own error-control plan; repeated-look control for one hypothesis does not automatically control the combined error rate across multiple hypotheses.