起源故事 · 高尔顿的散点,皮尔逊的尺子
Origin Story · Galton's Scatter, Pearson's Ruler
19 世纪末,Francis Galton 研究父母身高与子女身高,发现把两者画成散点能看出「一起变化」的趋势,
并提出了「回归」的雏形。他的学生Karl Pearson 把这种关系量化成今天的相关系数 —— 关键一步是标准化:
协方差除以两个变量各自的标准差,量纲被约掉,结果永远落在 [−1, 1],于是「身高 vs 体重」和「价格 vs 销量」终于能放在同一把尺子上比较。
但 Pearson 也反复强调:这把尺子量的是线性同步,绝不是谁导致谁。
In the late 19th century, Francis Galton studied how tall children grew compared with their parents.
Plotting the pairs on a scatter chart, he saw a clear "move-together" trend and sketched the first idea of regression.
His student Karl Pearson turned that visual hunch into the correlation coefficient we still use today.
The key move was standardization: divide the covariance by both standard deviations and the units cancel,
so the score is always in [−1, 1]. From that day on, "height vs weight" and "price vs sales" could be measured on the same ruler.
But Pearson kept reminding everyone: the ruler measures linear synchrony, never who causes whom.
1 散点图 · 拖 r 看点云聚散 Scatter plot · drag r to watch the cloud reshape
强正相关strong positive2 现实里的相关分析 Correlation in the Real World
相关 ≠ 因果:冰淇淋销量与溺水数都随夏天升高,r 很高,但谁也不导致谁 —— 背后是「气温」这个潜伏变量。
Correlation ≠ causation: ice-cream sales and drowning deaths both spike in summer. r looks huge, but neither causes the other — temperature is the lurking variable.
散点图先行:r 只测线性关系,完美的抛物线 r 也可能≈0。永远先画散点,别只看一个数。
Plot first, score second: r only captures linear structure. A perfect parabola can score r ≈ 0. Always eyeball the scatter before trusting a single number.
标准化:r 无量纲,所以「价格 vs 销量」和「温度 vs 用电」能横向比强弱。
Scale-free: r is dimensionless, so "price vs sales" and "temperature vs power use" can be compared head-to-head for strength.
离群点:一两个极端点能把 r 从 0 拉到 0.8 或反过来,分析前务必先排查异常值。
Outliers: one or two extreme points can yank r from 0 to 0.8 or vice versa. Screen for outliers before reporting any correlation.
一句话In One Line
协方差是「两变量一起偏离各自均值」的平均,乘积为正说明同向、为负说明反向,
但它的大小被量纲绑架,没法横向比。皮尔逊 r 把它除以两个标准差,量纲约掉、范围锁进 [−1, 1] ——
这才成了一把通用的「关联强度尺」。可这把尺子有两条铁律:它只量线性关系(非线性会被低估),
而且相关永远不等于因果 —— 高 r 可能是巧合、是反向、或藏着第三个潜伏变量。
Covariance is the average of "how far X and Y wander from their own means, multiplied together":
positive products mean they swing the same way, negative ones mean opposite ways,
but the magnitude is hostage to whatever units the data carry. Pearson r divides by both standard deviations,
cancels the units, and locks the score into [−1, 1] — a universal ruler for linear association strength.
Two unbreakable rules: this ruler only sees linear patterns (non-linear ones get under-counted),
and correlation is never causation — a high r can be coincidence, reverse causation, or the shadow of a hidden third variable.
常见误用Common Mistakes
看到高 r 就断定 X 导致 Y。相关 ≠ 因果,要排除潜伏变量、反向因果,必要时做实验。
Reading a high r as "X causes Y". Correlation is not causation — rule out lurking variables and reverse causation, and run a controlled experiment when the stakes are real.
r≈0 就认定两者无关。r 只测线性,非线性(如抛物线)关系 r 也可能为 0,先画散点图。
Treating r ≈ 0 as "no relationship". r only measures linear structure. A clean parabolic link can still score zero — always plot the scatter first.
不查离群点就信 r。少数极端点能彻底扭曲 r,分析前先做异常值排查。
Trusting r without screening outliers. A handful of extreme points can completely distort r — run an outlier check before publishing the number.