Zad Rafi (Sir Panda)

When planning for drop outs: If you need n patients and expect some proportion x to drop out, you don't inflate n by 1.x but rather divide by 1-x.

Example: n = 200 and expected 20% drop out

200 * 1.20 = 240 (incorrect, as 80% of 240 = 192)

200 / 0.8 = 250 (correct, as 80% of 250 = 200)

People can’t hand out randomized envelopes properly and accidentally randomize entire villages instead of people and were supposed to believe in a large randomized study with no other issues

He was right, and with large language AI models you don’t even need to conduct studies

In that case, with a large sample size and other biases, you have a narrow interval around the wrong value, making you overconfident in range that could be off to make a practical difference

Plus, the idea of a very large trial with a precise estimate and no other biases is fairy tale. If people accidentally randomize entire villages or can’t hand out envelopes right, imagine how perfectly they’re conducting all the other protocols of the study

It’s not a bad article but there are some misunderstanding of how certain procedures work, which makes it frustrating when seeing an authority of EBM creating occasional blanket rules of thumb for what’s desirable

If you wanna balance prognostic factors so badly, why not measure them beforehand and divide the subjects into two groups to achieve perfect balance and the sought after “no differences between groups”?

Naturally the only thing to do after seeing this is to suggest that you actually test whether randomization has succeeded via a statistical test for balance, and if it’s not, just rerandomize forever until you achieve perfect balance and the holy p =1

The Use of Historical Controls in Clinical Trials, With Dr Althouse JAMAevidence JAMA Guide to Statistics and Methods · Episode

"Trials" with historical controls - the worst of all worlds.

open.spotify.com/episode/08PC...

RIP Edward Leamer. Specification Searches was ahead of its time

Nice summary of our recent paper on SMOTE w/ @alcarriero.bsky.social @benvancalster.bsky.social

RIP Edward Leamer. Specification Searches was ahead of its time

Screenshot of http://jmlr.org/papers/v26/23-1317.html

My work on network regression and mediation in latent space models is now published at JMLR!

jmlr.org/papers/v26/2...