起源故事 · 从"遗传回归"到万能模型
Origin Story · From "Regression to the Mean" to a Universal Model
"回归"这个词来自 19 世纪的 Francis Galton:他发现高个父母的孩子身高会"回归"到平均值附近。
但真实世界里,结果往往受多个因素同时影响,简单回归不够用。于是统计学家把它推广成多元回归 ——
一次性放进面积、位置、楼层等所有变量,用最小二乘法找到那张最贴合数据云的平面。
从房价评估到产品良率分析,它都是回答"到底哪个因素最关键"的主力工具。
The word "regression" comes from 19th-century scientist Francis Galton, who noticed that children of unusually tall parents tended to regress toward the mean in height.
But real-world outcomes are usually driven by several variables at once, and simple regression isn't enough. Statisticians extended it into multiple regression —
throw every predictor (size, location, floor, age) into one model and let ordinary least squares find the plane that hugs the data cloud most tightly.
From housing valuations to manufacturing yield analysis, it remains the workhorse for answering "which driver actually matters most?"
1 拟合平面:穿过数据云的那块"斜板" The Fitted Plane: A Tilted Slab Through the Data Cloud
拟合中fitting…2 系数与显著性:谁的影响大、谁可信 Coefficients & Significance: Who Matters, Who You Can Trust
条越长=净效应越大。实心=系数显著(可信),半透明=不显著。 共线性升高时,系数估计会变得不稳定,显著性掉得最快。 Longer bar = bigger net effect. Solid = coefficient is significant (trustworthy), translucent = not significant. As collinearity rises, coefficient estimates get unstable and significance is the first thing to collapse.
3 现实里的多元回归 Multiple Regression in the Real World
房价模型:面积、地段、楼层、房龄一起进模型,得出每个因素的边际价格。
Housing model: feed size, neighborhood, floor, and age into one model to read off the marginal price of each driver.
良率影响因素:温度、压力、转速对良率的净效应,定位最关键的工艺参数。
Yield drivers: quantify the net effect of temperature, pressure, and spindle speed on yield to pinpoint the critical process parameter.
销量驱动因子:价格、广告、季节、渠道同时建模,量化各自的贡献度。
Sales drivers: model price, ad spend, seasonality, and channel together to quantify how much each one really contributes.
关键 X 筛选:六西格玛分析阶段用回归从一堆候选 X 里筛出真正影响 Y 的少数几个。
Vital-few X screening: in the Six Sigma Analyze phase, regression sifts a long list of candidate X's down to the handful that actually move Y.
一句话In One Line
多元回归最值钱的概念是"净效应"—— 控制住其他变量后,单看一个因素的纯影响。
这正是它比一次只看一个变量强的地方:能拆开"面积"和"楼层"各自的功劳。
但它有个软肋:当自变量本身高度相关(共线性),模型分不清功劳该记给谁,系数会乱跳、符号甚至反转。
所以建模铁律是 ——先看 R² 整体好不好,再看 VIF 有没有共线性,最后才信单个系数。
The crown jewel of multiple regression is the net effect — the pure contribution of one driver once everything else is held constant.
That's exactly why it beats looking at one variable at a time: it can separate the credit owed to "size" from the credit owed to "floor".
Its soft spot: when predictors are highly correlated with each other (collinearity), the model can't decide who deserves the credit — coefficients swing wildly and signs can even flip.
So the modeling discipline is firm — first check overall R², then check VIF for collinearity, and only then trust an individual coefficient.
常见误用Common Mistakes
R² 高就当模型可信。还要看残差是否随机、有无共线性(下一节诊断)。
Trusting a model just because R² is high. Also check that residuals look random and that there's no hidden collinearity (covered in the diagnostics page next).
无视共线性硬解读系数。VIF>10 时系数不可信,应剔除或合并冗余变量。
Reading coefficients straight off while ignoring collinearity. Once VIF crosses 10 the coefficients are no longer trustworthy — drop or merge the redundant predictor.
把相关当成因果。回归只说关联,因果需实验设计(DOE)佐证。
Treating association as causation. Regression only shows association — you need a designed experiment (DOE) to back a causal claim.