Frank Harrell

Big Book of R looks like an amazing resource in general too

This should be a bright, blinking light catching everyone’s attention.

YouTube video by Forbrukerrådet - Norwegian Consumer Council A Day in the Life of an Ensh*ttificator

The Norwegian Consumer Council with an amazing video on enshittification and how to resist it.

www.youtube.com/watch?v=T4Up...

Suggested criterion for federal funding cuts: Any university that puts "The" in front of its name (you don't see Oxford or Cambridge doing that).

Kristi Noem has done a stunning amount of damage and it’s good she’s gone.

But this doesn’t change the fact that we need a complete overhaul of DHS, impartial investigations into the killings of two American citizens, and information on children that were taken from Minnesota.

If you think that's a major problem, trace where the money goes. The fraction supporting full-time faculty teachers and their direct administrative support is probably embarrassing (at almost all universities including mine).

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

The results depend on linearity&additivity of effects on the original scale, &absence of floor and ceiling effects. These are unlikely to be satisfied. Change analyses can only be equated to ANCOVA in linear models. I wish the linear model had not been the starting point. hbiostat.org/bbr/change

This is one reason I wish that we had more globally accepted patient functional measures that were routinely collected. These can be overridden by clinical events (like ED visits) in an ordinal scale. hbiostat.org/endpoint

Beautiful web site by @andrew.heiss.phd

Quarto has been an incredible platform for my teaching material and I highly recommend mirroring to the web using Netlify (static html hosting). I have all public material in a local ~/web directory and deploy that with Netlify command line tool. Many there people find Github pages satisfactory.

Don't we know this already? Cluster analysis answers the question of how I can add to my CV. :-)

Well put. One of the worst things I've seen in multiverse analyses is authors giving equal weight to unprincipled analyses as to principled ones, e.g., including least squares analysis for a binary outcome.

You can use multiple imputation with Bayes, possibly making use of posterior stacking as you do during model fitting. I don't know the workflow for the better full Bayesian modeling approach when doing out-of-sample prediction but I'll bet it's possible. @avehtari.bsky.social any pointers?

Anthropic are heroes. Claude is the best.

After nearly 400 episodes, "Stats + Stories" has become a space in which statisticians, journalists, and the curious come together to make sense of the world around them. Read this behind-the-scenes look at what it takes to produce a podcast. magazine.amstat.org/blog/2026/03... #statssky

Very nice blog post. Doesn't dissuade me from my belief that using the most flexible model that the information content in the data will support is a better approach. And that approach gives you accurate uncertainty intervals right out of the box.

Beautiful!! Related:locking up of clinical trial data by NIH&industry is tragic. Individual pt data from 1 large well done randomized trial can be worth all the meta-analyses you want & can be many times better than the largest observational study. Data completeness and quality are tops for RCTs.

Quietly, calmly and forensically, BBC just dismantled the Trump communications shitshow on Iran.

No hyperbole, just laying out an unprecedented military, diplomatic and reputational shambles.

Worth a watch.

(🎥 BBC News/BBC Verify)

Jesus in a MAGA hat “One day I will return but only once you bomb a bunch of kids because a pedophile told you to”

Cartoon by Jesse Duquette

That approach has much to offer; I'm just a bit concerned with the treatment of the data generating process, which in Bayes can involves a single dataset in isolation (one-time events). And for Bayes is's preferable to think of more general/flexible models instead of always simulating things.

Bayesian thinking: specify a model that has extra parameters for known unknowns, e.g., ratio of variances of two groups, degree of non-normality, nonlinearities, interactions. Put skeptical priors on these "hope you don't need" parameters. Avoids discontinuous thinking from jumping between models.

2/2 When you are first introduced to #Statistics through Bayes, impatience with indirect and usually ill-defined concepts such as confidence intervals, NHST, and p-values grows exponentially. #StatsSky

1/2 One of the many beauties of @rmcelreath.bsky.social 's Statistical Rethinking is that it is 100% Bayesian, i.e., no frequentist ideas. If only that was the first #Statistics book anyone sees - the analytical world would be so much farther along. #StatsSky

A great question. To me teaching the CLT makes things worse, not better. One thing it does is to give people the idea that uncertainty intervals should be symmetric. The biggest harm is assuming that the CLT leads to accurate intervals.

By not catching the defects in change from baseline, and not caring about the assumed correlation pattern in longitudinal measurements, this SAP is a joke.

The first thing I evaluation when reading a statistical analysis plan is whether the outcome variable is proper. I'll bet that AI didn't even consider that.

This has most of the errors I'm always harping about.

On the last part I hesitate to recommend a suboptimal imputation method just because it's aligned with future use. Our goal should be to have an excellent imputation (or non-imputation Bayesian) method then find out how to use it in the field, if possible.

Right; you create a circular multiple imputation, nothing wrong with that.

That's my first choice, because it fully informs you of the importance of the value that is missing for the person being predicted.