Frank Harrell

Anyone who gets to work with Aki is very lucky …

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

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

Good grief ICE!

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!)

"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.

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

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.

August 2025 Top 40 New CRAN Packages – R Works Here are my picks for the best from the ninety-seven new packages that landed on CRAN in August, organized in eighteen categories: Causal Inference, Data, Differential Privacy, Ecology, Environmental ...

August 2025's Top 40 picks are out, highlighting excellent #RStats packages from Differential Privacy to Toxicology to Causal Inference.

This marks 1 year of Top 40 posts since we started R Works! 🎉

Check out the post here: rworks.dev/posts/august...

After decades as a teacher experiencing such issues, my conclusion is to have a single pre-course personal laptop R setup help session and put an absolute requirement that setup be completed a week before class starts.

Welcome! You are invited to join a webinar: Research Methods Academy: Seminar - The Importance of Sample Size for the Development and Evaluation of Clinical Prediction Models. After registering, you w... Led by Professor Richard Riley Join this Online Seminar on clinical prediction models estimate individual risk for outcomes from disease. Many studies lack rigorous design; this session will focus o...

Giving a talk today which anyone can watch online: 1pm - 2pm
"The Importance of Sample Size for the Development and Evaluation of Clinical Prediction Models"
- intended for a broad audience (anyone doing or using prediction model research)
imperial-ac-uk.zoom.us/webinar/regi...

But I’d rather see such laziness than have researchers say “consuming soy lowers the risk of breast cancer” etc. And I’d like to see the modest “predicts” used more often, supplemented with “further understanding of pathways is required for a causal conclusion about the predictor”.

Thanks. It may be interesting to study the meaning of estimands when not conditioning on a strong factor when the estimator is non-collapsible as with binary Y.

It would be interesting to see an example where that is reasonable. I know this situation exists but can’t believe it would be very common. We also have the age variable even though it raises a privacy issue, hence age is sometimes truncated at 80 or so to prevent re-identification.

If X is predictive of outcome and X was used in probability sampling but X was not stored in the dataset, isn’t that a major design flaw?

Illogic of Weighting While Estimating Model Parameters Consider a situation where some types of subjects are over- or under-sampled from a population where the sampling probabilities are known. For example, to increase efficiency we might sample a greate...

But see discourse.datamethods.org/t/illogic-of...

fharrell.com/post/marg

That’s the main reference, a simple demonstration of how unadjusted estimates are ill-defined and easily manipulated by subject inclusion criteria even with complete homogeneity of effects.

The main reason Revenge Dump's in office again and not in prison where he belongs is cos Mitch McConnell violated his oath of office and didn't lead the Senate to vote for impeachment. Thanks Mitch that decision will haunt you to your grave and as long as there's still recorded history afterwards.

Reviewer notes: In a randomized experiment, the pre-post differences are not effect estimates Reviewer notes are a new short format with brief explanations of basic ideas that might come in handy during (for example) the peer-review process. They are a great way to keep Julia from writing 10,0...

Psychologists are also definitely guilty of confusing change with treatment response, which is why I wrote about it here:

www.the100.ci/2025/01/22/r...

It’s not just confounding; when the outcome is binary or time-to-event, failure to condition on important prognostic factors results in ill-defined effect estimates even with zero confounding.

The fact that most meta-analyses use unadjusted estimates casts doubt on the meta-analysis field.

“Doing research is a choice, and unless you’re involved in some urgent project—curing a disease or winning a war or righting some injustice or raising living standards or whatever—or some interesting project—baseball statistics or the theory of random walks or whatever—you shouldn’t do it.”

Always a good read

I am sickened that journals are still letting authors compute change from baseline in parallel group studies. hbiostat.org/bbr/change

Avoiding One-Number Summaries of Treatment Effects for RCTs with Binary Outcomes – Statistical Thinking This article presents an argument that for RCTs with a binary outcome the primary result should be a distribution and not any single number summary. The GUSTO-I study is used to exemplify risk differe...

Join the club. There is nothing wrong with ORs. Like all other one-number summaries it’s best to accompany them with n-number summaries: www.fharrell.com/post/rdist/i... #EpiSky #StatsSky

What a refreshing change from the Gideon bible I find in so many U.S. hotel rooms.

#Statistics thought of the day: To the entire non-statistical world, 'error' means 'mistake' but not in frequentism. Type I assertion probability α isn't P(mistake in asserting θ>0), i.e., P(θ <= 0 | p < 0.05); the assertion is made without knowing θ. hbiostat.org/bbr/alpha #StatsSky #bayes

E Fieller also had a famous saying called “Fieller’s forecasting lemma”: “Given them a number or give them a date, but never both.”

And I thought it was only 92% :-)