Post

💬 Viewpoint: In #Bayesian trials, the choice of prior information and careful interpretation of treatment benefit probabilities are central to valid conclusions, especially in adaptive and pediatric studies. ja.ma/45tlFYL

17/8/26, 4:00 pm0 reposts3 likes

David Epstein

@davidnotdave.bsky.social

31 Aug

Hmm: "...a Bayesian posterior probability of treatment benefit of 0.92 in a large trial with a neutral prior corresponds approximately to a 1-sided frequentist P value of .08 or a 2-sided frequentist P value of .16, which would not meet conventional thresholds for statistical significance." (1/5)

David Epstein

@davidnotdave.bsky.social

31 Aug

That comparison doesn't hold up. The Bayes .92 means that, if the assumptions of the model (including the neutral prior) are correct, then, given the trial results, there’s a 92% probability that the treatment really had benefit. That's what clinicians/patients want to know. (2/5)

David Epstein

@davidnotdave.bsky.social

31 Aug

The p values say nothing about the probability of benefit. They're the long-run frequency with which the observed results would occur over many trials in the absence of a real benefit. That's not a question anyone asked, and it's not very helpful for decisionmaking by clinicians or patients. (3/5)

David Epstein

@davidnotdave.bsky.social

31 Aug

Also, the 1-sided frequentist test is problematic because, when you do one, you're not permitted to draw conclusions in the opposite of your hypothesized direction. If the treatment seems to reduce benefit (i.e., to do harm), your 1-sided test requires you to conclude only that it didn't help. (4/5)