I am so skeptical about Bayesian analysis because the priors will be manipulated.
There is a live example: the Tigris trial of polymyxin hemoperfusion for sepsis. They took one of the many negative RCTs and post-hoc cherry-picked a “positive” subgroup and declared it as their prior. The Tigris study itself had just enough patients to not turn the tide.
Sensitivity analyses are common in current clinical trials, and I have to assume they will be utilized in Bayesian trials as well. If there is anything unique about Bayesian methods, it's how easy the impact of their assumptions are to scrutinize!
It is also worth noting that similar tricks can be performed without using Bayesian methods at all by e.g. choosing an inappropriate external/historical comparison arm, inclusion/exclusion criteria, or even the control treatment itself.
Great piece. I agree Bayesian CTs will become a much bigger part of drug approvals. The advantage of Bayesian methods when using adaptive designs will also be a factor.
However, I disagree that the Bayesian prior threatens the idea of reviewer clinical judgement. The subjectivity in Bayesian priors allows for more clinical judgement not less. At the design stage the reviewers can have their clinical judgements incorporated (and quantified) into the supplementary and sensitivity analysis sections of the statistical analysis plan in the form of different priors and prior weights etc. At authorisation application stage the reviewers still have clinical judgement, e.g., where Bayesian dynamic borrowing is used, whether 65% borrowing or 70% borrowing is acceptable relies on clinical input more than whether .048 or .055 (or .027) is acceptable imo.
I am so skeptical about Bayesian analysis because the priors will be manipulated.
There is a live example: the Tigris trial of polymyxin hemoperfusion for sepsis. They took one of the many negative RCTs and post-hoc cherry-picked a “positive” subgroup and declared it as their prior. The Tigris study itself had just enough patients to not turn the tide.
Sensitivity analyses are common in current clinical trials, and I have to assume they will be utilized in Bayesian trials as well. If there is anything unique about Bayesian methods, it's how easy the impact of their assumptions are to scrutinize!
It is also worth noting that similar tricks can be performed without using Bayesian methods at all by e.g. choosing an inappropriate external/historical comparison arm, inclusion/exclusion criteria, or even the control treatment itself.
Great piece. I agree Bayesian CTs will become a much bigger part of drug approvals. The advantage of Bayesian methods when using adaptive designs will also be a factor.
However, I disagree that the Bayesian prior threatens the idea of reviewer clinical judgement. The subjectivity in Bayesian priors allows for more clinical judgement not less. At the design stage the reviewers can have their clinical judgements incorporated (and quantified) into the supplementary and sensitivity analysis sections of the statistical analysis plan in the form of different priors and prior weights etc. At authorisation application stage the reviewers still have clinical judgement, e.g., where Bayesian dynamic borrowing is used, whether 65% borrowing or 70% borrowing is acceptable relies on clinical input more than whether .048 or .055 (or .027) is acceptable imo.
Why don't we just pick non informative priors with null effect? That's the same assumption that p values are based on.