Frank Harrell

Researcher ‘honestly shocked’ to discover name on paper, editor claims misunderstanding While reviewing her Google Scholar profile to prepare a list of her publications, psychologist Maryam Farhang came across a paper she didn’t recognize.  The article, in the Journal of Research…

While reviewing her Google Scholar profile to prepare a list of her publications, psychologist Maryam Farhang came across a paper she didn’t recognize.

The article included her name and affiliation, but shevhadn’t written or contributed to the paper in any way.

The next time they tell you there’s no money for healthcare, remember there was money to start a war with Iran.

The next time they tell you there’s no money for housing or social supports, remember there was money to bomb a girls elementary school.

There’s always money for war.

In many observational studies it is a push to even call the sample a 'cohort'. For example in electronic health record-based studies we seldom know what makes a patient enter our health system, and anything about the patient that occurred while in a previous system is unknown.

7  Nonparametric Statistical Tests – Biostatistics for Biomedical Research

There might be something gained by framing this as a rank difference test, e.g., stacking the data and using an ordinal model with random intercepts for persons as exemplified in hbiostat.org/bbr/nonpar#s...

If by name you are referring to TTE I strongly disagree. Emulation means to copy, and blinding, randomization, simple time zero, and prospective follow-up are not being copied.

How does that help control for unmeasured confounders? How do most observational data collections even think about confounders?

What's new is valuable but has nothing to do with accounting for confounding.

How is it that a paper that claims to show how to do target trial emulation does not address confounding by indication and its ramifications for data collection? www.jclinepi.com/article/S089... #EpiSky #StatsSky

Much better, since ultra-fast data.table has a concise algebra for almost everything you need, with no package dependencies: Use LLM as a helper in constructing data.table commands as in this advanced example: hbiostat.org/rflow/long#s... #RStats

Let me just ask for an example like the one I posed, where the new method is used to draw inference about the unknown treatment effect.

Just took a quick look. Hard to see what the actual inference is, and how it would apply to longitudinal ANCOVA with a treatment x time interaction in the model + baseline covariates.

This sounds too good. Wouldn't us Americans miss our traffic jams, rampant consumerism and resulting trash and recycling challenges, power grids being hurt by AI data farms, and all the interesting compounds we have in our food, water, and air?

Isn't that about interpreting the process that generates confidence intervals, but doesn't help with interpreting the single computed CI in front of you? #StatsSky #Statistics

They cannot reliably estimate sample avg Tx effects (single differences) since they don't address confounding by indication, so why would they allow estimation of differential treatment effects (double diffs)? Contrast with Bayesian models admitting uncertainty: fharrell.com/post/hxcontrol

I am completely convinced that it is detrimental to teach tidyverse to students before they learn base R and perhaps data.table. I have looked at students' homeworks when they haven't followed that advice. Phew! #RStats

US men's olympic hockey team: cowards when it counts

My effort to reproduce this paper began as part of the Institute for Replication’s ongoing project to systematically examine the reproducibility and robustness of papers in Nature Human Behaviour3; my participation in this endeavour was approved by the Ethical Review Board of Vrije Universiteit Amsterdam’s School of Business and Economics. Inspecting the paper’s first two figures revealed a mathematical impossibility. There are nine EU countries that experienced zero terror attacks during the study’s time frame. However, the paper reports that the inverse hyperbolic sine of these countries’ per capita attack rates are positive, and increase or decrease over time. This is impossible; the inverse hyperbolic sine of zero is zero4. The main outcome variable displayed in the paper’s second figure is hard-coded in the replication data as ‘DVSin’. Figure 1’s top row of plots shows that DVSin is negatively correlated with both terrorist attack rates (r = −0.107, two-sided P = 0.024) and their inverse hyperbolic sine (r = −0.108, two-sided P = 0.022). These plots also show that in the 305/420 country-year observations after 2006 experiencing zero terror attacks (72.6%), DVSin takes on 292 different positive values. This implies that the paper’s main outcome variable cannot possibly be constructed as described in the paper.

"the paper’s main outcome variable cannot possibly be constructed as described in the paper."

Retraction of 2023 paper that did not use the reported variables. The replication report is astonishing www.nature.com/articles/s41...

Pre-publication peer review remains undefeated in laundering bullshit

@thelancet.com continues to fail in its role as a gatekeeper of good science, not only by a 2 week rule but by limiting letters to 250 words. How do you expose a severe but very technical flaw in a study's design or analysis in 250 words? #StatsSky

Brilliantly written with one small caveat: Towards the end the article implies that observational data may be useful for studying heterogeneity in treatment effects. Don't know why that would be the case. But great work Darren! #StatsSky #EpiSky

@thelancet.com continues to make a mockery of peer review at times.

I tried to tell y'all.

To clarify, you have to show excellent Bayesian or frequentist power at the max affordable N. But that N in no way should be considered "the" sample size.

hbiostat.org/bayes/bet/de...

Sorry - should be hbiostat.org/bayes/bet/de...

An alternate way to read this is to say that sample size calculations should be abandoned in favor of sequential analysis once the maximum affordable N is determined. hbiostat.org/bayes/beta/d...

The Republican Party Has a Nazi Problem How did the GOP become a haven for slogans and ideas straight out of the Third Reich?

Terrific piece by former Republican, Tom Nichols
"The Republicans have a Nazi problem, yes. But this means that the United States also has a Nazi problem.. . ."
www.theatlantic.com/magazine/202...

Justice Department withheld and removed some Epstein files related to Trump An NPR investigation finds the public database of Epstein files is missing dozens of pages related to sexual abuse accusations against President Trump.

NPR - DoJ withheld some Epstein files related to allegations that President Trump sexually abused a minor. www.npr.org/2026/02/24/nx-s1-...

This reminds me of an article I recently read about how liberal arts colleges in the US that do not have graduate students are giving their students better education by relying on nothing but real professors to teach. Not being "research universities" is a plus. Contrast with rampant use of TAs.

Goal-Driven Flexible Bayesian Design presentation updated w/comparison of performance of frequentist group sequential designs: hbiostat.org/bayes/design . Frequentist approach takes far too long to make a decision by controlling something that is NOT an error prob. #Statistics #StatsSky

It is vitally important that human statisticians stand up to collaborators who don't understand statistical principles or who do bad science (e.g., hope to learn a lot when p >> N). But those are precisely the skills NOT being taught in #Statistics programs. www.fharrell.com/talk/collab #StatsSky