Brendan Alexander

Computing tool of the day: I've needed to have an elegant way to stitch together pieces of Quarto books to make a unified html file for teaching. In collaboration with Claude I now have a great yaml-driven tool for this. #rstats

got a framed copy to hang by the ai team

@breannetidemann Seems more reasonable on the log scale.

@_bcalexan It's amazing how years add up, isn't it? lol!!

Life sure leads you down unexpected roads. I've had an opportunity come up recently that was too good to walk away from. And so this summer I will be leaving my current position and my family and I will be moving to Norway. It's bittersweet leaving a career that I've built for 1/

@breannetidemann It hasn’t been 15 years… oh wait…

15 years, but right now it's the right opportunity for us and I'm excited for the adventures ahead. I'm so grateful for all the learning from fellow scientists, techs, farmers, agronomists, commission staff and everyone else in this crazy ag industry. But right now? 2/

@MacrinePhD @matloff @MIT @skdh How does it rank on the ol’ meter?

Physics anyone? The gap between the "weird" quantum world and our everyday reality just got a lot smaller. Researchers at @MIT have built a mathematical bridge showing that quantum behavior—like particles being in two places at once—can be calculated using the classical "principle of least action." By applying the same math we use to predict a falling ball to the double-slit experiment and quantum tunneling, they’ve found a common language for physics at all scales. news.mit.edu/2026/new-study… #Physics #QuantumMechanics #ScienceNews #MIT #Feynman #QuantumPhysics #Innovation #STEM #ScienceTwitter

@learnfromerror @f2harrell @AprileBernardi @PavlosMsaouel @elmir1omerovic @StatModeling Aren’t they both error probs in a way? Bayesian provides errors probs for the data at hand and freq provides it over ‘all such decisions.’ I have read and listened to your webpage on alpha, @f2harrell. I have questions about it, maybe for another time if you’re willing.

@f2harrell @AprileBernardi @PavlosMsaouel @elmir1omerovic @StatModeling Posterior probs aren't error probs, and what's more, they're not given error probabilities. As David Cox said: "they are what they are".

@f2harrell How would you respond to the criticism by Evans et al. in 10.1001/jama.2026.4175? Are the arguments made there against Bayesian methods in phase 3 RCTs valid in any meaningful methodological sense? In particular, what do you make of this claim: “Use of Bayesian methods in late-phase or confirmatory clinical trials has generally been limited to supplementary analyses, given recognition that their implementation can compromise evidentiary and integrity standards and the reliability of results through (1) concession of the benefits of randomization through their inclusion of external (prior) information; (2) loss of objectivity by incorporating sponsor- or investigator-specific priors; and (3) reduced robustness via reliance on strong and sometimes unverifiable assumptions.”

2/2 in the same way, betting on good poker hands isn’t an error even if you lose the hand.

@f2harrell I’ve been trying to understand your view on type 1 errors. Is your view that acting as though a treatment has an effect is not an error if the data suggests it because it’s simply the most logical move? 1/2

@RobCalver5 @rlmcelreath @StatModeling @avehtari @mcmc_stan Does it lean more towards worked practical examples like in statistical rethinking? Or is it more like bda3 by gelman?

Very excited to announce that the #BayesianWorkflow book by @StatModeling, @avehtari, @rlmcelreath et al publishes in June! routledge.com/9780367490140 #RStats #DataScience #Bayesian @mcmc_stan

@JohnCleese No Edmonton date!

Join me this fall for a screening of “Monty Python and the Holy Grail” followed by a live Q&A session. Prepare your silliest questions for me and get your tickets now on johncleeselive.com OR I SHALL TAUNT YOU FOR A SECOND TIME! ☠️🏰🩸🐇🪾 #montypython #holygrail

@f2harrell @learnfromerror @PavlosMsaouel @elmir1omerovic @StatModeling Priors seem similarly unknowable in that there are infinite options for any situation. I agree that the Bayesian posterior would be fantastic to have.

@_bcalexan @learnfromerror @PavlosMsaouel @elmir1omerovic @StatModeling No, that would be like studying oracle properties in predictive models, which are not relevant to me. Conditioning on unknowable quantities is a recipe for cloudy decision making, and it disrespects the workflow. We want to know P(data-based conclusion is correct).

@f2harrell @learnfromerror @PavlosMsaouel @elmir1omerovic @StatModeling So no errors, just bayesian updating all the way down?

@f2harrell @learnfromerror @PavlosMsaouel @elmir1omerovic @StatModeling And that it is sensible to talk about the probability of making errors in such a situation? 2/2

@f2harrell @learnfromerror @PavlosMsaouel @elmir1omerovic @StatModeling I admit, I don’t completely understand your point. I’ve read this page a few times before. Surely you would agree that under the null hypothesis that (either point or range) that it would be an error to conclude that there was a treatment effect? Works in simulations…1/2

@AdamFrank4 @CColose @AstroKatie @WKCosmo The scientific method does not depend on consensus, nor does it necessarily produce it.

1) @CColose @AstroKatie @WKCosmo Yes. This idea that consensus isn't science is so profoundly stupid. It is for scientists! it's how we, as a community, decide what's understood and what is not. And science is always about communities of scientists.

@AdamFrank4 @CColose @AstroKatie @WKCosmo no it isn't, and it shouldn't be either

@5_utr Would you elaborate on this? I don’t think I’ve seen this topic explained in such a way before.

4/ Bayesians have a finite "belief budget": ∫ p(θ)dθ = 1. This is critical. This forces trade-offs—confidence in one idea costs confidence elsewhere. It's basically a built-in guard against overconfidence

After recent discussion around RTOG 1112, where the authors misinterpreted their own main results and conclusion due to multiplicity concerns, and called a “positive” trial “negative”, I think it’s worthwhile for a thread on multiplicity and controlling α… 🧵

@matloff @f2harrell @US_FDA I’m very interested in this answer as well. I still don’t see how multiple comparisons are adjusted for in cases where the treatments are assumed to be non-exchangeable. Sceptical priors only?

@f2harrell @_bcalexan @US_FDA This was a serious question. No negative consequences at all?

To see support and advocacy for Bayesian design and analysis at the highest level of @US_FDA is a thrill.

@f2harrell @matloff @US_FDA I’m all for full pre-specification, that sounds great. I’ve yet to see it happen in real life…

@_bcalexan @matloff @US_FDA This is a key unanswered question. I can see allowing a decision maker to use her own prior if she was forced to set it in writing before seeing the results. But there is also an argument for full pre-specification before the study.

@matloff @f2harrell @US_FDA That answers who has to make the decision but not how they will do it. If you are comfortable with a final arbiter to make the decision then would it also be allowable for that arbiter to set their own prior for the analysis? Or they could ask for p-values of course.

That's a standard response I've received over the years I've been criticizing p-values, "But you have to make a decision!" As I've said more than once in this thread, I certainly am pro-decision making. But I want the decision to be made by the relevant people, NOT to be usurped by p-values or priors. The FDA must ultimately must approve a drug or not, and they should do so with full information. If a statistician uses priors, they are impinging on the FDA's right/responsibility to make a fully informed decision. As to lack of objectivity, we can never have perfection in that regard, but we certainly should not make it worse, which the Bayesian approach does.