Jeudi 10 mars 2022 | Tim Hesterberg (Google) : Bootstrap Surprises

Heure : 15:30 ; Par zoom : https://uqam.zoom.us/j/83797904757

Résumé : Resampling methods are easier to use and more accurate than classical formula-based statistical methods, but computationally expensive. Whoa – wait a minute. You can go wrong if you don’t understand the idea behind the bootstrap. You might think of the bootstrap for small samples where you don’t trust the central limit theorem, but the most common bootstrap methods are less accurate in small samples than classical methods. There are simple variations that are dramatically more accurate, and these show that the old n > 30 rule is just wrong; try n > 5000 instead (theory confirms this). Finally, we bootstrap for big data at Google because it is faster than using formulas. I hope to change not only the way you think about the bootstrap, but about statistical practice.