Lightnews — Scholar-powered news

Frank Harrell @f2harrell.bsky.social · 13h

Yes the main idea is to write about things others are interested in or will benefit from, which motivates them to take part after the first draft is posted online.

1

Frank Harrell @f2harrell.bsky.social · 13h

If you have non-stratified-on covariates the 0.5 threshold is for the covariate setting you request to be predicted.

2

Frank Harrell @f2harrell.bsky.social · 16h

Getting rid of academic journals in favor of capturing ongoing comments and suggestions from experts would make the world a better place.

Matti Vuorre @matti.vuorre.com · 18h

Against Publishing: universonline.nl/nieuws/2025/...

Preprints are read, shared, and cited, yet still dismissed as incomplete until blessed by a publisher. I argue that the true measure of scholarship lies in open exchange, not in the industry’s gatekeeping of what counts as published.

2 4 35

Frank Harrell @f2harrell.bsky.social · 16h

To estimate the mean the highest time value must be uncensored. To estimate the median for a covariate value X=x the survival curve for X=x must drop below 0.5. This is for semiparametric/nonparametric estimators like Cox/K-M.

1 4

Frank Harrell @f2harrell.bsky.social · 1d

Yes for (only) the very special case of a linear model, if pre is linearly related to post and pre is put in both the left and right side of the model, this will rescue the error made in analyzing change. Why rely on a special case? Use a principled method that works for all models.

3

Frank Harrell @f2harrell.bsky.social · 1d

You're welcome. A major manifestation of the problem is the inability to interpret results when the rigid assumptions are not met, and having the same amount of change mean completely different things to different persons.

1 1

Frank Harrell @f2harrell.bsky.social · 1d

Briefly the two measures must already be perfectly transformed, there be no floor or ceiling effects, and difference must be unrelated to baseline. Details at hbiostat.org/bbr/change #Statistics #StatsSky

14 Transformations, Measuring Change, and Regression to the Mean – Biostatistics for Biomedical Research

hbiostat.org

1 1 9

Reposted by Frank Harrell

Darren Dahly @statsepi.bsky.social · 2d

I read and write, I explore and I question, I design and script and analyse, I interpret and communicate. I do this to train my mind in the hopes of one day generating new knowledge. New knowledge that might even be useful, and that no algorithm can yet be trained on.

Terry McGlynn @hormiga.bsky.social · 6d

Y'all. I just got ChatGPT to do everything in R for this manuscript. I mean EVERYTHING. And it's all legit and reproducible. I'm shook.

How are we mentoring our trainees in statistics now? Who needs to learn coding in R line by line, and who doesn't?

scienceforeveryone.science/statistics-i...

Statistics in the era of AI

How do we mentor, teach, and do stats when AI can do so much of the work?

scienceforeveryone.science

4 17 81

Reposted by Frank Harrell

Raider @iwillnotbesilenced.bsky.social · 3d

This Is Fascism

690 12K 30K

Frank Harrell @f2harrell.bsky.social · 2d

At @vanderbilt.edu I used to teach a biostat course for interdisciplinary biological sciences PhD students but then (1) a senior leader told all the students to drop the course and take an online one and (2) no one took the online course anyway.

1 4

Frank Harrell @f2harrell.bsky.social · 2d

Glad there are exceptions. Want to describe what you program teaches for confounder specification?

Frank Harrell @f2harrell.bsky.social · 3d

Agreed but also see major deficits in statistical training of epidemiologists. Just one example: their courses say that stepwise regression is OK.

3 1

Frank Harrell @f2harrell.bsky.social · 3d

And one of the biggest problems in epi is the large number of researchers who won't use causal methods (and won't even do sensitivity analysis to unmeasured confounders!) but who use causal conclusions in their papers.

1 1 5

Frank Harrell @f2harrell.bsky.social · 3d

Yes! Coding is thinking. My new rule of thumb: Use a LLM only to write code to do things that either are low priority or that I don’t have time to do and wouldn’t have done without LLM (while assuming the generated code is only 80% right).

3 9

Frank Harrell @f2harrell.bsky.social · 3d

Unfortunately true. Many researchers see statisticians as speedbumps. To me, worst offenders have been animal researchers and epidemiologists. An animal researcher once told me “There is a reason I don’t come to the free daily biostat clinics: I know you will tell me that 6 dogs isn’t enough.”

2 1 16

Frank Harrell @f2harrell.bsky.social · 7d

Yongxi Long and colleagues have written an excellent post about when and when not to worry about the proportional odds assumption: discourse.datamethods.org/t/when-and-w... #StatsSky #EpiSky #Statistics #RStats

When and why (not) to worry about the PO assumption

Aim We wrote an article (Long, Wiegers, Jacobs, Steyerberg, & Van Zwet, 2025) about the proportional odds (PO) assumption in the analysis of ordinal outcomes. we use various examples from neurological...

discourse.datamethods.org

9 35

Frank Harrell @f2harrell.bsky.social · 7d

That's largely true. The biggest problem is the amount of medical practice that has not been researched.

1 1

Frank Harrell @f2harrell.bsky.social · 8d

Anyone who gets to work with Aki is very lucky …

Aki Vehtari @avehtari.bsky.social · 8d

I'm looking for a doctoral student with Bayesian background to work on Bayesian workflow and cross-validation (see my publication list users.aalto.fi/~ave/publica... for my recent work) at Aalto University.

Apply through the ELLIS PhD program (dl October 31) ellis.eu/news/ellis-p...

ELLIS PhD Program: Call for Applications 2025

The ELLIS mission is to create a diverse European network that promotes research excellence and advances breakthroughs in AI, as well as a pan-European PhD program to educate the next generation of AI...

ellis.eu

1 1 3

Frank Harrell @f2harrell.bsky.social · 8d

The extremely problematic use of change scores is so poorly understood by researchers that it’s almost sickening. Most don’t even understand what is needed for the subtraction operator to work. hbiostat.org/bbr/change

2 3 18

Frank Harrell @f2harrell.bsky.social · 10d

An excellent argument. One related way I try to get course managers’ attention is to quote the cost of SPSS licenses to students once they leave the comfort of the university site license. The cost is obsene, for a product that does far less than R, & LLM can even help students learn to program R

1 3 14

Reposted by Frank Harrell

David Colquhoun @davidcolquhoun.bsky.social · 10d

Good grief ICE!

Mike Galsworthy @mikegalsworthy.bsky.social · 10d

History repeats.

2 3 11

Reposted by Frank Harrell

Andrew Heiss @andrew.heiss.phd · 11d

If you've ever wanted to learn how to make beautiful websites with #QuartoPub and #rstats , check out this workshop I'm giving in a couple weeks! It'll be a blast (and we're covering Quarto's brand new _brand dot yaml system!)

Statistical Horizons @stathorizons.bsky.social · 26d

Learn to create and publish a professional, data-focused website in “Create an Online Presence with Quarto Websites” on October 16-17, with @andrew.heiss.phd‬! Discover how to use #Quarto to build a variety of websites like personal portfolios, research compendiums, and interactive dashboards.

Quarto Websites | Online Seminar | Code Horizons

This online course taught by Andrew Heiss, Ph.D., teaches you how to use Quarto to build a variety of data-focused websites.

codehorizons.com

3 29 86

Reposted by Frank Harrell

Peter Tennant @pwgtennant.bsky.social · 11d

"Uncooperative statistician": the term used (typically by a senior clinician) to describe a well-trained and knowledgeable statistician who refuses to conduct flawed or fraudulent research.

3 14 53

Frank Harrell @f2harrell.bsky.social · 14d

Flexible Goal-Drive Bayesian Design - video of presentation is now available at www.fharrell.com/talk/gdesign/ #Statistics #StatsSky #bayes #rct

Goal-Driven Flexible Bayesian Design – Statistical Thinking

The majority of clinicals trials that are successfully launched end with equivocal results, with confidence intervals that are too wide to allow drawing a conclusion other than “the money was spent”. ...

www.fharrell.com

3 6

Frank Harrell @f2harrell.bsky.social · 14d

In addition an equally important requirement is that the study is actually designed, i.e., that the investigators cared enough about the question to get funding for prospective data collection, QC, and unbiased outcome evaluation.