Not Frequentist Enough | Hacker News Reader

Not Frequentist Enough

statmodeling.stat.columbia.edu

23 pointsTomte8 comments

The root challenge of all modern frequentist statistics is "does the selection criteria for the sample actually allow me to generalize the results". Experiments and post-hoc analysis are really sensitive to selection biases and researchers (and much worse the media) tend to overgeneralize the results of a study. Its particularly bad in medicine. My personal bugbear is that you can only participate in a study for an autoimmune disorder if you only have the 1 disorder of focus, but real people with autoimmune disorders tend to have more than one (diagnosed or not). So studies on treatment efficacy basically can't actually generalize to the populations they intend to treat because they select against them.

> The root challenge of all modern frequentist statistics is "does the selection criteria for the sample actually allow me to generalize the results".

Why did you single out frequentists here? This is a challenge for all inference, prediction, classification, and any other "forecasting".

All those things are either frequentist or baysian statistical models. Baysian models are much harder (often impossible) to use but at least pay attention to intersectionality.

Frequentists define this as their challenge (and in very specific ways).

Some other subfields doing similar things don't.

Wow, I could have sworn I read this a while back, but the post date is today! What?

For anyone else that this occurs to, here's the explanation: a similar post with only a few wording differences was posted back in October: https://statmodeling.stat.columbia.edu/2022/10/05/not-freque...

Yes: https://news.ycombinator.com/item?id=33114403

I had exactly the same reaction. I wonder what happened. Thanks for sleuthing.

I came to the comment section to look for this comment! Thanks.

Maybe the reposting is part of a frequentist study. I'll call it article laundering.

Reply on news.ycombinator.com