Chris Tosh | The piranha problem in statistics
The piranha problem (too many large, independent effect sizes influence the same outcome) has received some attention on Andrew Gelman’s blog. But now it’s a paper! Chris Tosh (Memorial Sloan Kettering) talks about multiple views of the piranha problem and detecting the implausible scientific claims that are published. The butterfly effect makes an appearance.
If you enjoyed the science-vs-pseudoscience topics, you’ll enjoy this one.
#statistics #datascience #science
0:00 - Coming up in the episode
2:35 - What is the Piranha Problem?
19:54 - Confusing effect sizes
23:11 - The "words & walking speed" study
26:22 - Declaration of independent variables
30:58 - Piranha theorems for correlations
37:07 - Piranha theorems for linear regression
40:37 - Piranha Theorems for mutual information
44:13 - Bounds on the independence of the covariates
46:12 - Applying the piranha theorem to real data
50:12 - Applying the piranha theorem across studies
54:05 - A Bayesian detour
1:00:12 - The butterfly effect & chaos
1:04:26 - Applying the piranha theorem to cancer research