The Effective Statistician - in association with PSI How to communicate results from adaptive studies simple, but still correct
Oct 27, 2025
In this engaging discussion, Kaspar Rufibach, an experienced biostatistician and authority on adaptive clinical trials, shares his insights. He explains the challenges of communicating results from adaptive studies and emphasizes the importance of using clear, defensible language. Listeners learn about the difference between conditional and unconditional bias, the implications for point estimates, and the significance of median-unbiased estimation. Kaspar also addresses the complexities surrounding secondary endpoints and the necessity of pre-specifying adjustments to ensure trust and reproducibility.
AI Snips
Chapters
Transcript
Episode notes
Estimation vs Hypothesis Testing
- Adaptive trials protect type I error for hypothesis tests but complicate estimation of treatment effect size for patients and clinicians.
- Unbiased estimation after interims is nontrivial because sampling distributions become mixtures of truncated normals.
Decide Which Bias To Target
- Ask which bias matters: conditional (given stopping stage) or unconditional (repeated-sampling property).
- Prefer targeting unconditional bias for frequentist inference to retain desired repeated-sampling properties.
Stage-Wise Ordering Simplifies Inference
- Multi-stage sampling requires an ordering of the sample space to define p-values and confidence intervals.
- Stage-wise ordering treats earlier stopping as more extreme and enables median-unbiased estimators and p-values that don't depend on future looks.
