.jpg){kind=link}
- Published
- Updated
Good performance in one favourable scenario does not establish that an adaptive design is reliable across relevant conditions. A simulation plan should reveal situations in which the design could fail, change decisions or deliver less efficiency than expected.
Define the question and assumptions
Define plausible ranges for effects, event frequency, recruitment, data delays and missingness. Evaluate error rates, power, expected sample size and decision probabilities under null and alternative effect patterns. Derive scenarios from scientific uncertainty rather than restricting the assessment to combinations that make the proposed design appear especially attractive.
Hypothetical study example
In a hypothetical study, simulations assume rapid information flow while site endpoints arrive slowly. Adding delay scenarios shows that many participants would already be enrolled before an adaptation could be implemented. The team reassesses expected efficiency using these more realistic operational conditions instead of retaining the original favourable summary.
Design review checks
- Justify scenario ranges using clinical evidence and operational information sources.
- Display adverse and boundary cases alongside the central favourable planning scenario.
- Explain which decision each reported performance measure is intended to inform.
Record the decision and its limits
Accepted simulation reporting makes methods, assumptions and uncertainty reproducible. Estimated performance proportions themselves have numerical uncertainty depending on simulation repetitions. Results are not guarantees for the actual trial: discuss unmodelled behaviour changes, measurement problems and clinical uncertainty. Guidance-based planning does not create a promise of regulatory acceptance for a particular design.
Frequently asked questions
- Is one central scenario enough?
- Alternative conditions are generally needed to reveal design fragility.
- Do simulations provide a certain probability of success?
- They evaluate performance only under the specified model and assumptions.