起源故事 · 回归平庸,与 "regression" 一词的由来
Origin Story · Regression to Mediocrity, and Where "Regression" Comes From
19 世纪末,Francis Galton 先研究豌豆种子大小的亲子遗传,又测量了上千对父子的身高,发现一个反直觉的规律:
高个父亲的儿子,倾向于"没那么高";矮个父亲的儿子,又往往比父亲高一点 —— 后代总在向平均值靠拢。
他称之为"回归平庸"(regression to mediocrity),这就是统计学里 "regression" 一词的来历。
他的学生 Karl Pearson 1896 年把这种关系形式化,写下了至今通用的相关系数 r。
In the late 19th century, Francis Galton first studied how pea-seed size passed from parent to offspring, then measured thousands of father–son height pairs — and found a counterintuitive pattern:
tall fathers tended to have sons who were "not quite as tall", while short fathers often had sons a touch taller than themselves. Offspring kept drifting back toward the population average.
Galton called this "regression to mediocrity" — the origin of the word "regression" in statistics.
His student Karl Pearson formalized the relationship in 1896, writing down the correlation coefficient r we still use today.
1 散点与回归线:拖一个点,r 跟着动 Scatterplot & Regression Line: Drag a Dot, Watch r React
强正相关strong positive2 协方差的几何:四个象限的"正负投票" The Geometry of Covariance: Four Quadrants Voting + or −
以两条均值线把图切成四块。落在右上 / 左下(绿)的点贡献正协方差,落在左上 / 右下(红)的点贡献负协方差。哪种颜色压倒性更多,符号就归谁 —— 这就是协方差的本质:一场关于"同向还是反向"的投票。 The two mean lines slice the plot into four quadrants. Points in the top-right / bottom-left (green) cast positive votes for covariance; points in the top-left / bottom-right (red) cast negative ones. Whichever color wins by majority sets the sign — that's all covariance is: a vote on whether the two variables move the same way or opposite ways.
3 现实里的相关性 Correlation in the Real World
相关 ≠ 因果:冰淇淋销量与溺水人数高度相关,背后真凶是"夏天"。r 大只说明同进退,不证明谁导致谁。
Correlation ≠ causation: ice-cream sales correlate strongly with drowning deaths — the lurking variable is summer. A high r only says they move together; it never identifies which one drives which (a textbook spurious correlation).
质量因子筛选:把各工艺参数与缺陷率算相关,r 大的优先排进 DOE,砍掉不相干的噪声因子。
Screening quality factors: correlate each process parameter with defect rate. Promote the high-|r| factors into a follow-up DOE and drop the noise factors that barely move the needle.
回归分析前置:建模前先看相关矩阵。两个自变量彼此 r 太高 = 多重共线性,模型系数会乱。
Pre-flight for regression: inspect the correlation matrix before fitting. Two predictors with a very high pairwise r signal multicollinearity — the regression coefficients will become unstable and hard to interpret.
投资组合:资产间相关性越低,分散风险效果越好。低 r 的资产,才是真正的"鸡蛋放不同篮子"。
Portfolio construction: the lower the pairwise correlation between assets, the better diversification works. Low-r holdings are the only ones that truly put your eggs in different baskets.
一句话In One Line
协方差先衡量两个变量是否"同向波动",但它带量纲:用厘米还是米算,数值天差地别,没法横向比较。
于是 Pearson 把它除以两者的标准差,压进 [−1, +1] 这把无量纲的统一尺子里 —— 这就是相关系数 r。
它只捕捉线性关系:一条完美的抛物线 r 也可能接近 0。看懂 r,是进入回归分析、DOE、SPC 多变量分析的第一道门槛。
Covariance first measures whether two variables drift the same direction, but it carries units — switch from centimeters to meters and the number changes wildly, killing any cross-comparison.
Pearson's fix was to divide by the product of the two standard deviations, squeezing everything onto the unitless [−1, +1] ruler we call the Pearson correlation coefficient r.
It only captures the linear piece of the relationship: a perfect parabola can still give r ≈ 0. For ranked or monotonic-but-nonlinear data, reach for Spearman's ρ instead. Reading r fluently is the first gate into regression, DOE, and multivariate SPC.
常见误用Common Mistakes
把高相关当成因果。相关只说同进退;因果要靠实验(DOE)或控制混杂变量。
Treating a high r as proof of causation. Correlation only says the two variables co-move. To claim causation you need a designed experiment (DOE) or proper control of confounders.
r≈0 就断言"没关系"。r 只测线性;曲线关系下 r 可能为 0,务必先画散点图看一眼。
Reading r ≈ 0 as "no relationship". Pearson's r only measures linearity. A clean quadratic curve can give r near zero — always eyeball the scatterplot before concluding independence.
被一个离群点带偏。单个极端点能把 r 拉高或拉反,先检查异常值再下结论。
Letting a single outlier hijack r. One extreme point can inflate r, flip its sign, or kill it entirely. Vet outliers before drawing any conclusion.