New paper by Blandhol, Bonney, Mogstad, and Torgovitsky:

www.restud.com/when-is-tsls...

#REStud
#EconSky

Unfalsifiable by Design: A Year of Trying and Failing to Reproduce a Human Microbiome and Autism Study The myth of open data, reproducibility, responsibility, and accountability in science, and your role in it

How every layer of science's "self-correcting machinery" failed when Iva Veseli and I simply wanted to reproduce the findings of a high-profile study on gut microbiome and autism:

merenlab.org/2026/04/15/u...

RIP Archippus, you would have loved Haskell.

Does anyone in #StatsSky or #EconSky or whatever know of anyone has published about an analogue of the competing risks identification problem in space rather than time?

Brief fun survey from Jessica, Andrew & myself:

If you are a faculty member, research scientist, postdoc, or senior Ph.D. student in any area of science, please take five minutes and fill it out. We’ll share the results widely along with some reflections.

docs.google.com/forms/d/e/1F...

For reasons that are certainly unrelated to grant applications, does anyone know of (ideally with some evidence) uses of tools I have contributed to (scoringutils, epinow2, epinowcast, epidist, primarycensored, episoon, epinow, censoreddistributions ...) from different parts of the world?

Survey research is often interpreted as showing that belief in conspiracy theories can be surprisingly widespread, including belief in conspiracy theories that would be astonishing if true. For example, in The Atlantic we learn that “12 million Americans believe lizard people run our country”

The fun problem is that usually, only specimen that are in bloom would be collected and enter the dataset, so you need to handle this selection effect in your likelihood. You also need to be careful to not mistake changes in flowering _duration_ with changes in _onset_. 2/2

Unified theory resolves phenological paradoxes in biocollection data by modeling phenophase duration during Bayesian inference Phenology is increasingly studied using digitized biocollection data, yet no general theory links observed specimen collection dates to the unobserved processes governing phenological events. We der.....

New paper: Some things to consider when you want to use old samples of flowers from herbaria to make inferences about changes in flowering onset over time (or similar). Had a blast building some Bayesian models for this. Thx David and Daniel for letting me help doi.org/10.1111/nph.... 1/2

Has anyone done the opposite rating? I.e. which popular YT history channels get things mostly right?

Makes sense - the ECDF diff plot for sigma[2] tends towards overestimation, its just not enough to prove a problem. SBC requires quite a lot of simulations to be sensitive to subtle problems and it appears the overestimation affects +/- a dozen of what appears a 100-200 simulations.

I'm a data scientist @ourworldindata.org and I need help from a botanist or someone local to Kyoto, Japan! 🌸

We present one of the world’s longest climate records: 1,200 years of peak cherry blossom dates in Kyoto.

The researcher who maintained it, Prof. Yasuyuki Aono, sadly passed away last year.

If you believe that omega[1,2] and sigma[2] _should_ be informed by data, then you may try to design a derived quant sensitive to that. If sigma[2] should correlate with the total variability of y, then e.g. SBC for sd(y) * sigma[2] could be useful. Or use full model log_lik (if feasible) 3/3

E.g. in the extreme, if my model is just mu ~ N(0,1) and I don't see any data, then a correct posterior for mu is N(0,1), my posteriors will not be correlated with the simulated value and that is the only correct answer. 2/3

SBC will check if your X% intervals contain the true value X% of the time. This appears to be the case (at least for the X = 90 case shown). The fact that the estimates are not informed by the data does not mean there is a problem with the model 1/3

That's the one thing I've ever actually tried in this area...

My _impression_ is that since you (AFAIK) need a ton of simulations anyway, ABC often looses out to amortized inference and the latter works better and is where the cool stuff is. Though, I just have a very shallow view of the literature.

trying to learn more about Approximate Bayesian Computation, to fit less tractable models (say ABMs, or rule-based categorization systems) to data; or fit models to aggregate data reported in other papers. but so many approaches out there! what are your favorite tutorials/papers?

Several groups submitted plots that didn't match their data table. All LLM generated (they quickly admitted it). the experiments appeared to have been actually done, but obviously the work was not checked and had important errors -> low grade (and a bunch of students learning they can't trust LLMs)

Its rare but it happens. We've had students submit PDF that retained an initial part of the prompt in the title metadata. We didn't limit LLM usage, but they used LLM to fake data for an experiment they claimed to have done. There was other stuff weird with the data. 1/2

More details about the Bayesian Workflow book and case studies now available on the book web site avehtari.github.io/Bayesian-Wor... (but you still need to wait a bit for the book)

The New York Times Got Played By A Telehealth Scam And Called It The Future Of AI

Since the New York Times published its semi-viral big profile of Medvi last week — the "AI-powered" telehealth startup that it breathlessly described as a "$1.8 billion company" supposedly run by just two brothers —…

cover of the book "Bayesian Workflow" by Gelman, Vehtari, et al. Coming out later this year, in the summer probably.

I would have preferred to have the "draw the rest of the owl" meme on the cover, but this will do. Seems like it is on schedule, and we'll leave some typos so you know we didn't write it with AI.

The world looks on and shrugs as over half a million people are expelled from their homes - again - in clear disregard and violation of every relevant norm. Larger in scale than what Russia did in Mariupol (which the whole Western world condemned).

Imagine absolutely trashing a female lead author, who leads up a women's research institute, who is undoubtedly going to face unreal threats for doing this work, for correctly describing the fact that previous theories predicted a different predictor here

One of the patterns I have realized in reactions to the last Nature metascience papers is that many seem to observe an increase in reproducibility/replicability estimates and perceive it as a sign of progress. I regret to inform you that the variability is too high to support that conclusion.

Russian embassy in Prague is at Ukranian hereos street (Ukrajinských hrdinů) and just behind the Boris Nemtsov square.

Part 2 of my shrinkage estimator series is out! Part 1 covered the univariate case, but now we dive into multivariate shrinkage 🤓

We cover Spearman's classic correlation disattenuation formula, multivariate James-Stein estimators, and hierarchical methods too

haines-lab.com/post/how-to-...

Why are we still repeating the same rhetoric ("disturbing") after over a decade as if we haven't learned anything in the meantime? Metascience seems so enamored with its original anchors that it can't have enough of repeating them.

what a charming tradition, nailing a completed PhD thesis up!