The Gold Standard of Experimental Design | Conditional Power, Predictive Power and Futility Boundaries

1. Start with the final success criterion

Back to contents

Define the primary hypothesis, test, analysis population, information target and significance threshold before calculating interim probabilities. Conditional or predictive power is always relative to a final success event. If the endpoint, estimand or multiplicity plan changes, the quantity changes. State whether futility is binding or non-binding and who can override the rule. A percentage without its assumed effect, information fraction and future-data model is not a decision statistic.

2. Define conditional power precisely

Back to contents

Conditional power is the probability of final success given current data under a specified future effect. The assumed effect might be the design alternative, the current interim estimate, a clinically relevant effect or another value. These choices can produce very different answers. Observed conditional power inherits the noise of the interim estimate and can fall sharply after a modest early fluctuation. Report the assumption beside the number and avoid speaking of one unique conditional power.

3. Understand predictive power

Back to contents

Predictive power treats the treatment effect or future data as uncertain and averages conditional success probability over a predictive distribution. In Bayesian form, that distribution combines a prior with interim data; frequentist predictive approaches can use estimates and uncertainty differently. Results depend on prior strength, model and endpoint. Predictive power is not automatically more objective. It can be more stable than plugging in a noisy effect, but its assumptions must be visible and stress-tested.

4. Time the review by information, not convenience

Back to contents

A very early review has little ability to distinguish weak treatment from noise; a late review saves few resources. Use information fraction based on events or variance, not simply enrolment. Delayed outcomes mean many enrolled participants contribute no interim information. Plan data cleaning, adjudication and follow-up cutoffs. For time-varying or delayed effects, an interim model assuming immediate proportional benefit may underestimate future success.

5. Choose futility boundaries through operating characteristics

Back to contents

A threshold such as conditional power below 20 percent is not universal. Simulate the probability of stopping under no effect, target effect, smaller meaningful effects, delayed effects and nuisance misspecification. Examine expected sample size, lost power, false-futility probability and remaining safety information. Binding boundaries affect the formal testing procedure if crossed and ignored; non-binding boundaries permit continuation without inflating type I error under specified designs but may reduce expected efficiency. Governance still matters.

6. Worked example: the same interim data, three powers

Back to contents

A trial reaches 50 percent information with an interim effect half the planned target. Conditional power under the original target is 55 percent because future data are assumed to follow the design effect. Observed conditional power using the interim estimate is 24 percent. Predictive power with a sceptical but weak prior is 31 percent. A prespecified non-binding boundary is 15 percent, so the monitoring committee recommends continuation while reviewing delayed-effect assumptions. The example shows that the measures are not interchangeable forecasts; each encodes a different model of the future.

7. Protect blinding and operational integrity

Back to contents

Comparative interim statistics should normally be restricted to an independent data monitoring committee or statistical group. Investigators receive the recommendation, not the effect estimate or exact power unless the charter allows it. A visible stop-or-continue decision can still reveal information, so communications should be limited. Record data cut, code, assumptions, committee deliberation and deviations. Safety monitoring and efficacy or futility monitoring answer different questions even when reviewed together.

8. Changed-context transfer check: an education intervention

Back to contents

A cluster-randomised study of a school programme reviews futility after half the schools complete one term. Scores may improve only after teacher training matures, while intracluster correlation and school-level degrees of freedom make information lower than the student count suggests. The transferable method defines success, computes actual information, compares conditional and predictive assumptions and simulates false futility. The changed context requires a delayed implementation model and a boundary that respects cluster count, not a clinical-trial threshold copied unchanged.

9. Know the limits and common misconceptions

Back to contents

Low conditional power does not prove no effect. It may reflect an early negative fluctuation, delayed benefit, wrong nuisance assumptions or a misspecified endpoint model. Predictive power is not a posterior probability that the treatment works unless success and model are defined that way. Futility monitoring cannot repair poor adherence, missing data or outcome bias. Continuing despite a boundary may affect interpretation and operations. Stopping early reduces knowledge about safety and subgroups.

10. Audit assumptions, simulations and decisions

Back to contents

Publish the monitoring plan, success event, information calculation, effect assumptions, priors, thresholds, binding status and operating-characteristic simulations before unblinding. Retain interim datasets, cleaning status, code and committee records. Recalculate under design, observed, clinically relevant and delayed-effect scenarios. Report the probability of stopping under each true effect and the consequences of overruling the boundary. The final paper should state that the trial stopped for futility, not that ineffectiveness was proven.

11. Frequently asked questions

Back to contents

Is conditional power a forecast? It is a model-based probability under a specified future effect. Why can values differ? Design-effect, observed-effect and predictive assumptions differ. Does a non-binding boundary mean optional evidence? It permits continuation under the planned testing framework, but governance and interpretation still matter. Can futility be assessed for safety? Safety stopping uses different risks and rules. Is 20 percent the standard boundary? No; thresholds should follow simulations and context.

12. The gold standard makes the future assumption explicit

Back to contents

A credible futility design connects the interim statistic to the final success criterion, uses true information, exposes the assumed future effect and calibrates boundaries through simulation. It protects unblinded information and preserves the distinction between conserving resources and proving absence of benefit. The decision is strongest when readers can see not only the power number, but also the model of the future that produced it.