Skip to main content

EFFECT_NOISE_RATIO

Constant EFFECT_NOISE_RATIO 

Source
pub const EFFECT_NOISE_RATIO: f64 = 2.0;
Expand description

How many times the raw-versus-gglib effect must exceed the A/A drift before the report will call it more than noise.

§This is a rule of thumb, not a test

A single A/A pair estimates the drift from one degree of freedom. Two draws of a noisy quantity can land close together by luck, and a factor derived from them carries no confidence level, no p, and no power. 2.0 is chosen because it is the smallest factor at which the two numbers are plainly not the same size — enough to stop a delta that is within its own noise from being reported as a finding, and not enough to license the word “significant” about one that clears it.

The honest way to strengthen this is more A/A pairs, not a bigger factor.