Group Sequential Design

Group Sequential Design: Concept Architecture

Orientation and learning roadmap

This page follows a trial through planned interim looks at accumulating outcome data. It explains the stopping boundaries, demonstrates why repeated testing needs error control, and connects early stopping to the evidence available for clinical and economic decisions.

Plan the looks before the trial starts

Specify the primary outcome, hypotheses, maximum sample size or information target, and the timing of interim and final analyses in the protocol. The plan should also define who may see comparative results and which prespecified rules can lead to stopping for strong benefit, harm, or lack of benefit.

Information is often measured by the precision of the primary treatment-effect estimate rather than by the raw number enrolled. A data monitoring committee may review unblinded evidence while preserving the trial team's blinding and conduct. The FDA guidance on adaptive clinical trial designs describes group sequential trials as having prospectively planned interim analyses and stopping criteria.

Why repeated significance tests need a boundary

Testing the same null hypothesis at several opportunities and declaring success whenever a conventional threshold is crossed can inflate the chance of a false-positive conclusion. A group sequential design instead sets boundaries for the planned looks so that the overall, or familywise, type I error probability meets the chosen level under the relevant null hypothesis.

For a simple conservative illustration with two looks, one can require a two-sided $p$ value no greater than $0.025$ at either look. The union bound guarantees:

$$P_0(\text{reject at look 1 or look 2}) \leq P_0(p_1\leq0.025)+P_0(p_2\leq0.025) \leq0.025+0.025=0.05.$$

This bound does not assume independence between looks, but it can be more conservative than boundaries designed for their actual dependence. O'Brien–Fleming-type boundaries typically demand stronger evidence early and less extreme evidence later, while other designs distribute error differently. State whether boundaries are one-sided or two-sided and whether a futility rule is binding before interpreting the trial's error properties.

Worked two-look decision

Imagine a trial with a planned interim analysis at 100 completed primary outcomes and a final analysis at no more than 200. Under the illustrative $0.025$ threshold at each look, an interim two-sided $p$ value of $0.012$ crosses the efficacy boundary, so the prespecified efficacy rule allows the trial to stop.

An interim $p$ value of $0.030$ does not cross that boundary, so enrolment or follow-up continues unless a different prespecified safety or futility rule applies. If the eventual final $p$ value is $0.020$, it crosses the final illustrative boundary; this example explains decisions at the two looks, not a complete protocol or an adjusted effect estimate. The trial still needs valid interval estimation and reporting methods that account for its sequential design.

Choose boundaries and stopping purposes

Early stopping for efficacy can reduce exposure to an inferior strategy, while stopping for futility can avoid further recruitment when the planned evidence is unlikely to support the target conclusion. Safety monitoring serves a different clinical purpose and needs an appropriate governance plan rather than being treated as a routine efficacy test.

The Adaptive Designs CONSORT Extension asks trials to report prespecified interim rules, adaptations, and actual stopping decisions. A schedule for looks without defined decision rules is insufficient to establish the intended operating characteristics.

What early stopping means for economic evidence

The number and timing of observed events determine how precisely treatment effects can be estimated for a cost-effectiveness model. A trial stopped early for apparent benefit may offer less information about long-term outcomes, rare harms, resource use, and health-related quality of life than a trial completing follow-up.

Use design-aware effect estimates and uncertainty intervals, and separate a trial's efficacy stopping rule from the economic model's later extrapolation assumptions. Earlier availability of results can have value, but shorter observation and selected early stopping can also increase uncertainty about lifetime costs and outcomes. These consequences should be considered when comparing trial designs or interpreting the resulting economic evaluation.

Validate and report the sequential analysis

Check the prespecified number and timing of looks, information fractions, spending or critical values, multiplicity across hypotheses, and simulated or analytical type I error and power. Record any changes made after interim comparative data became available, and distinguish a planned stop from an unplanned decision.

Report the look at which the trial stopped, the reason, actual sample size, treatment effect, and interval estimate using methods appropriate to the stopping rule. A conventional unadjusted final-analysis confidence interval may not retain its advertised coverage after selection by an early boundary. Protocol transparency and independent monitoring are essential when interim knowledge could influence recruitment, measurement, or analysis.