Laplace, it turns out, was the real Bayesian villain, according to Fisher. From this fun paper projecteuclid.org/journals/bay...
24.02.2026 19:17
👍 0
🔁 0
💬 0
📌 0
Laplace, it turns out, was the real Bayesian villain, according to Fisher. From this fun paper projecteuclid.org/journals/bay...
Somewhat related to this, it turns out R.A. Fisher considered Bayes a frequentist? 🧐
Nice perspectives, thanks! Re. risk aversion, I'd add that drugs have been approved on the basis of Bayesian methods for decades. So some companies are happily using Bayes, and others now have to play catch up. Hopefully the guidance can help normalize Bayes more widely.
Yep, that's one example. Useful e.g. for precision if you have one or more prognostic baseline variables (common in pharma RCTs and elsewhere)
G-comp can be necessary even in RCTs if the estimand is a marginal effect and the model has nonlinearities
Silly me! Thanks so much! 🙌
Hi Richard, could you share an example of frequentists proving good properties of Bayesian estimators?
For a step by step intro to poststratification, we cover it here (again, code-heavy; conceptual explanation in the book): theissbendixen.com/dag-book/cha...
We have an intro book on causal inference as part of Sage's QASS series coming out next year, so won't be of use for you right now, but the companion website (code-heavy) is live and might provide some inspiration
theissbendixen.com/dag-book/
Seems analogous to how, in the missing data literature, reasonable folks recommend a complete case analysis alongside any fancy imputation? Can be useful to see what the adjustment is doing descriptively, or even in terms of broader sensitivity/robustness checks ("totality of evidence" and all that)
Other applications of hierarchical models here for borrowing across different regions of the world.
cdn.who.int/media/docs/d...
Sweet, thanks!
Thanks!
Don't remember having seen Dirichlet processes used this way, any chance you could point to a rationale and/or implementation? Thanks!
Added a small section on tau
Depends on the model, of course, but with a simple post ~ pre + treatment, I don't immediately see why you couldn't put an informative prior on pre (and/or treatment, given relevant data)?
Sure, I think you could just borrow information for the controls on the raw scale?
... @fusaroli.bsky.social @nathanielforde.bsky.social @natehaines.bsky.social @tcarpenter.bsky.social @mzloteanu.bsky.social @jordannafa.bsky.social @tjmahr.com @mjskay.com @cameronpat.bsky.social @marvin-schmitt.com @vincentab.bsky.social @rmkubinec.bsky.social @paulbuerkner.com, etc.
Possibly of interest to @rmcelreath.bsky.social @solomonkurz.bsky.social @dingdingpeng.the100.ci @ruben.the100.ci @chelseaparlett.bsky.social @phdemetri.bsky.social @stephenjwild.bsky.social @matti.vuorre.com @healthstatsdude.bsky.social @jonaslindeloev.bsky.social @andrew.heiss.phd
"Being Bayesian in a Frequentist World"
New post on "Bayesian dynamic borrowing" in R 📚
Link 👇
Final manuscript submitted to the publisher! 📚
