Background/Aims: Bayesian adaptive designs quantify belief in treatment effect based on the observed trial data and empirically unverifiable prior beliefs. Such belief statements are intuitive and appealing to clinicians and patients alike but not typically consistent with traditional frequentist designs, which prioritize control of the overall false positive rate (Type I error) and quantify uncertainty using confidence intervals. We propose instead the novel Frequentist Confidence-Adaptive design illustrated by a two-arm five-stage trial to evaluate pre-hospital blood pressure management for haemorrhagic stroke patients.
Methods: We use confidence distributions constructed from confidence intervals at every confidence level (0-100) to formulate a design with options to stop early for high confidence of efficacy, harm, or lack of meaningful benefit (LMB, defined as insufficient improvement over control). Confidence decision thresholds are chosen to control Type I error, derived from O’Brien-Fleming-type alpha-spending. Final analysis expresses confidence in treatment benefit, harm, and LMB. We conducted 10,000 simulations each under positive, negative and neutral treatment effects to investigate operational characteristics.
Results: Using formulated confidence-based decision thresholds, there was a 2.47% and 84.1% chance of finding treatment efficacy when there a neutral (one-sided Type I Error) and positive effect (Power) respectively. In presence of a negative effect, there was a 99.5% chance of stopping early whether for LMB or inferiority.
Conclusion: Frequentist Confidence-Adaptive design enables frequentist quantification of belief in any treatment effect of interest that: a) does not require specification of empirically unverifiable prior beliefs and b) provides appropriate Type I and II error control.