Yikes Instagram. An Advertisement for a company that promises to learn how to do meta-analysis: “In just a matter of weeks you can get a great study [systemic review/ meta analysis] done published in top journals…without a statistician or [slow collaboration process]” #academicsky #metascience

Look at the data! It's all over the map. If you drink half a cup per day, your risk is a little higher than none at all. If it's 2.5 cups, risk is 20% lower. But if it's 4.5 cups, risk is only 13% lower. There's no universe where this makes causal sense. jamanetwork.com/journals/jam...

You know that coffee & dementia study that's going around? It's hogwash. For starters, it's not pre-registered (that I could find), which means it may have been the product of a fishing expedition looking for dementia associations. But even if it's not ... www.nytimes.com/2026/02/09/h...

@nature.org again functioning as a marketing platform rather than upholding scientific standards. It is misleading at best to describe this as a "randomized controlled trial" AND there is no disclosure that these are Google researchers "studying" a Google product. www.nature.com/articles/s41...

the rating scale was a slider with tick marks on the 10s

Don't you f**king dare.

Quite interesting read. I wish I got to hear a little bit more about FDA decisions like this in more detail as well. evidenceinthewild.com/about/ #statssky #skystats #academicsky

It still shocks me how prevalent the Hosmer–Lemeshow gof test is, especially considering Hosmer, Lemeshow and Le Cessie published a comparison in 1997 (17 years after the HL test came out) stating its poor performance! and it has been almost 30 years since! #academicsky #skystats #statssky

"in particular mice methods tend to have the best performance. Despite their growing popularity, deep learning-based joint imputation methods, such as miwae, gain, miracle, and remasker consistently underperformed in our analysis." #statistics #statsky #academicsky #skystats

Trying to work out why all peer reviewers of systematic reviews now suggest including every sensitivity analysis ever devised (do a leave-one-out, do a sequential analysis, do p-curve, how about some outlier detection?) Is the answer ‘they are all now written by LLMs?’

This is why we need forums like bluesky and discourse.datamethods.org #academicsky #statssky #statistics #skystats

2/2 Contrast that with Bayesian designs where magic happens: If the analysis at the planned study end is valid, all interim Bayesian analyses must be equally valid regardless of the stopping rule, and they require no modification, no consideration of sampling distributions.

The estimator is fine; the assumed sampling distribution is not. If continuation depends on interim results, this isn’t a fixed-sample experiment. Reporting a conventional 95% CI implies uncond. coverage, which is false unless the futility rule was irrelevant.Would love @f2harrell.bsky.social ideas

Trial statisticians! Should RCT pubs report conventional 95%-confidence intervals if they did a futility interim and continued the trial? No sampling distribution can possibly be Gaussian if randomly better than average data reaches study completion while randomly worse than average doesn't.

If we understood statistics, we would use a many many more one-tailed statistical tests.

Great opportunity to join us at JHU!

Why not use partial R^2 and partial chi-square statistics from likelihood-based tests, which directly quantify a predictor’s marginal contribution to model fit? #rstats #statssky #academicsky #skystats #statistics

As a non-academic, these cultures fascinate me: Academic 1: Omg someone asked for my slides & I'd rather die than share them Academic 2: just posted all my slides & notes. Lecture videos, too, but I had bronchitis all year so they're not great. Email me and I'll clarify what I said in btw coughs!

Yo what the fuck short form videos are bad Short-Form Videos Degrade Our Capacity to Retain Intentions arxiv.org/pdf/2302.03714 Ht: HowTown

Phenomenal 3-part lecture series by Dr. Hong Wang hosted by Institut des Hautes Etudes Scientifiques (IHES) going through the proof for 2D, sticky, and general Kakeya sets. Probably the biggest math breakthrough in 2025! #math #academicsky #mathsky www.youtube.com/watch?v=B9xK...

The web version of this text is incredibly high yield. Block quotes from other statisticians. Audio-inputs. Also an excellent table at the end (which I can only partly screen cap on my phone). #academicsky #statistics #statssky

#Statistics thought of the day: For longitudinal studies, don't treat the baseline measurement as the first follow-up measurement. Baseline cannot be considered a response to treatment. You'd be surprised how often this in done in #pharma. hbiostat.org/rmsc/long#mo... #StatsSky

Anyway, I really feel like the whole environment, including the state of the published literature, is really colluding against students learning how to do proper research. That’s a shame; I always have the impression our students really do want to learn how to do things the right way.

End of year reminder: the TRIPOD+AI reporting guideline is reporting standard for all clinical prediction model studies, including those using machine learning and AI. --> www.bmj.com/content/385/...

A first for both of us, @alicemeadows.bsky.social! 🎉 In it, I talk about how ORCID is making it possible to share other types of research outputs relevant to the arts and humanities scholars (like musical compositions, imagery and film, and teaching materials.) info.orcid.org/new-work-typ...

Our guidance regarding performance measures for medical AI models is finally out! - Stop bashing AUROC, although it does not settle things - Calibration and clinical utility are key - Show risk distributions - Classification statistics (e.g. F1) are improper www.thelancet.com/journals/lan...

A good title for a chapter on estimation: "What to expect when you're expectation value"

Translating written / whiteboard statistics to Latex is still an incredibly painful experience. #academicsky #mathsky #statssky

One thing @statberry.bsky.social does a lot that we should all be doing more is showing investigators the results of simulating their trial several times. When you know the true treatment effect but you get wild estimates with non-large N it usually gets their attention.

I didn't realize this until the other day but I usually keep matches and a candle on my coffee table (an old wooden chest) and it is essentially all the ingredients you need to estimate pi and solve Buffons Needle problem. I can't unsee it now. My poor future houseguests. #statssky #statistics #math

#rstats It is with profound sadness I heard that my long-time friend and colleague, John Fox passed away this week. He was the author of {car}, {effects}, {Rcmdr}, ... and numerous influential books. I will miss him greatly. www.john-fox.ca

Interesting move of a preprint server to ban papers on AI, apparently due to record numbers of submissions of which low quality content was generated using AI #academicsky #AI #statssky

This whole interaction. Is exactly why I am on this platform. Experts, discussing and debating. Sharing resources. 👏 👏 More of this please

Because @laurewynants.bsky.social and @benvancalster.bsky.social do no get on here much #DataScience #MachineLearning #MissingData #Imputation #Rstats #Python #AI journals.plos.org/plosone/arti...

One of the most upsetting articles I've read in a long time www.theargumentmag.com/p/when-grade... UCSD report senate.ucsd.edu/media/740347... We are failing a generation of kids.

@f2harrell.bsky.social you were right. Coverage for TMLE xgboost grid search size 5 appeared better than 20 but still the coverage is disappointing. I’ve updated this examination for my learning. Thanks again for the guidance. www.kenkoonwong.com/blog/tmle/

I was thinking about the efficiency of nonparametric tests and remembered this SO post about the Wilcox being 96% as efficient as T test, even in small samples. For those who have a copy, Lehman & Romano did an thorough job in <5 pages detailing situation. #statssky #statistics #rstats #AcademicSky

We wrote the Strain on scientific publishing to highlight the problems of time & trust. With a fantastic group of co-authors, we present The Drain of Scientific Publishing: a 🧵 1/n Drain: arxiv.org/abs/2511.04820 Strain: direct.mit.edu/qss/article/... Oligopoly: direct.mit.edu/qss/article/...

I have been a reviewer on 5 papers this year for AI/ML validation and each time they used R or R Squared. Had to request a Bier Score, Emax, even slope and intercept would be more ideal. Would be nice if ICI was more popular too

Survey Statistics: weights and MRP for voters statmodeling.stat.columbia.edu/2025/11/11/s...

It’s fascinating how cohorts of statisticians have parallel pockets of hyper-specific knowledge due to textbooks, like the collective familiarity with blue fiddler crab mating patterns, courtesy of Agresti. Palmer penguins, iris data, ect, any others I'm missing? #statssky #statistics #rstats

Well, it's official. #UNL Chancellor Bennett submitted his final proposal and it includes eliminating the #statistics department. budgetprocess.unl.edu/final-budget... Guess it's time to go look for jobs. Anyone looking for a couple of very talented #datavis researchers?

Quick thread on the BBC and the political and societal significance of recent developments: One of the main reasons the UK has historically been so much less polarised than the US, is that Britain has a shared source of information, consumed and trusted by most people regardless of their politics.

One thing I have noticed a lot in meta-analysis (thinking of those involving comparison of prognostic models) is that most focus on Embase and Medline, but no mention of Compendium (think IEEE) WoS or Scopus. Granted the sheer volume for data extraction would be astronomical.

Does anyone else automictically reach for the PDF symbol when opening a new article on a journal site? It just seems like I can read the them faster. #AcademicSky

It's a shame I didn't find this before Oct. 31st. And also a shame the term "Ghost-time bias" never caught on.

NEW PAPER: "Reporting guidelines for studies involving generative artificial intelligence applications: what do I use, and when?" --> www.nature.com/articles/s41... #HealthTech #ClinicalAI #MachineLearning #MedAI #AIinMedicine #TransparentAI #HealthInnovation #GenAI

Probably the papers I share the most often are @f2harrell.bsky.social 1996 on how to effectively develop multivariate models and @benvancalster.bsky.social 2019 paper on Calibration. Both extremely well written and informative. What are you most shared stats paper? #statistics #statssky #academicsky

I recently rederived the poisson pdf distribution from the binomial. Reminds me of @kareemcarr.bsky.social recent blog about memorization vs proof. No computer around and needed it for something. Certainly there are times where knowing something will be time saving then re-deriving it from scratch.