故事 · Fisher 在农场里发明的方法 Origin Story · The Method Fisher Invented on a Farm
1920 年代,统计学家 Ronald A. Fisher 在英国 Rothamsted 农业试验站工作,要回答一个朴素的问题: 不同肥料、不同品种的小麦,产量到底有没有真差别?田里同样的地块产量也会忽高忽低, 他需要一把尺子把「品种带来的差异」和「地块本身的随机波动」分开。 于是他发明了方差分析:把总变异拆成组间与组内,用二者之比(后人为纪念他命名为 F 值)来判断。 这套思想后来成了实验设计(DOE)的基石 —— 六西格玛改善里判断「哪个因子真正影响输出」, 靠的正是 Fisher 当年在麦田里磨出来的这把刀。 In the 1920s, statistician Ronald A. Fisher was working at Britain's Rothamsted agricultural research station, facing a deceptively simple question: do different fertilizers and wheat varieties really differ in yield? Even identical plots in the same field swing high and low, so he needed a ruler that could separate "the effect of the variety" from "the random wobble of the soil itself." His invention was Analysis of Variance: split the total variation into between-group and within-group, and judge by their ratio (later named the F-statistic in his honor). That idea became the foundation of Design of Experiments (DOE) — and when a Six Sigma project asks "which factor truly drives the output?", the answer still comes from the blade Fisher first sharpened in those wheat fields.

1 组间拉得开吗?组内挤得紧吗? Are Groups Pulled Apart? Are Points Within a Group Tight?

k=4

每根实心横线是一组的均值,灰色虚线是总均值。组均值离总均值越远 → 组间变异(信号)越大; 同组内部点散得越开 → 组内变异(噪声)越大。F 值就在量这一拉一散的对比。 Each solid horizontal line is a group mean; the gray dashed line is the grand mean. The farther a group mean drifts from the grand mean → larger between-group variation (signal); the wider points scatter inside one group → larger within-group variation (noise). F is exactly this pull-versus-spread contest.

2 总变异如何劈成两半,F 落在哪 How Total Variation Splits and Where F Lands

上方堆叠条把 SS总 切成 SS组间 与 SS组内;下方是 F 分布, 黄线是当前 F、红虚线是临界值 —— F 越过红线,落进右尾,才判「组别均值存在显著差异」。 The stacked bar above slices SS_total into SS_between and SS_within; below is the F distribution. The amber line is the current F, the red dashed line is the critical value — only when F crosses red into the right tail do we conclude "the group means differ significantly."

3 现实里的 ANOVA ANOVA in the Real World

多组比较:三种配方、四台机床、五个班次的良率谁更高,一次 ANOVA 全比完,不必两两 t。 Many-group comparison: three formulas, four machines, five shifts — one ANOVA settles whose yield wins, no pairwise t-tests required.
方差分解:把输出波动归因到「因子带来的」和「随机噪声」,是测量系统分析与稳健设计的语言。 Variance decomposition: attributing output swings to "factor-driven" vs "random noise" — the native language of measurement-system analysis and robust design.
Fisher 与 DOE:田间试验、全因子/部分因子实验,ANOVA 是判断主效应与交互效应的核心工具。 Fisher & DOE: field trials, full and fractional factorial designs — ANOVA is the core engine for testing main effects and interactions.
因子显著性:DOE 输出表里每个因子的 p 值,背后都是一次组间÷组内的 F 检验在打分。 Factor significance: every p-value in a DOE results table is just one more between-over-within F-test, scoring whether a factor truly matters.

4 动手算一遍 · 3 组数据 → SST/SSB/SSW → F Work It Out · 3 Groups → SST/SSB/SSW → F

数据(3 组,每组 n=4,N=12):Data (3 groups, n=4 each, N=12):

组1: 5,6,6,7 (x̄₁=6)  ·  组2: 8,9,9,10 (x̄₂=9)  ·  组3: 11,12,12,13 (x̄₃=12)
总均值grand mean x̿ = 9
SSB = Σn(x̄ⱼ−x̿)² = 4[(6−9)²+(9−9)²+(12−9)²] = 4·18 = 72  (df=k−1=2)
SSW = ΣΣ(x−x̄ⱼ)² = 2+2+2 = 6  (df=N−k=9)
SST = 72+6 = 78 ✓
MSB = 72/2 = 36  ·  MSW = 6/9 = 0.667
F = MSB/MSW = 36/0.667 = 54.0   vs   F0.05, 2, 9 = 4.26 → 54 ≫ 4.26,组间差异高度显著

5 效应量 · 显著之外,组别解释了多少变异 Effect Size · Beyond Significance, How Much Variation Groups Explain

F 显著只说「差异存在」,不说「差异有多大」。效应量回答后者 —— 组间差异占了总变异的多大比例。 A significant F says "a difference exists," not "how large." Effect size answers the latter — what fraction of total variation the grouping accounts for.

η² (eta²) = SSB/SST = 72/78 = 0.92  (样本内解释比例,偏乐观)(sample share explained; optimistic)
ω² (omega²) = [SSB−(k−1)MSW]/(SST+MSW) = [72−2·0.667]/(78+0.667) = 0.90  (去偏后)(bias-corrected)
η² 会高估:它把样本里的随机波动也算进「被解释」,组数或噪声一大就虚高;小样本尤甚。 η² overstates: it counts sample noise as "explained," inflating with more groups or noise — worst in small samples.
ω² 更诚实:扣掉一份噪声估计后再算比例,恒 ≤ η²,是向他人汇报效应大小的首选。 ω² is more honest: it subtracts a noise estimate before taking the ratio, is always ≤ η², and is preferred when reporting effect size to others.

经验档:η²/ω² ≈ 0.01 小、0.06 中、0.14 大。本例 0.90 属于罕见的极强效应(教学数据刻意分得很开)。务必和 p 值一起报 —— 大样本能让微不足道的差异也「显著」,唯有效应量能拦住它。 Rules of thumb: η²/ω² ≈ 0.01 small, 0.06 medium, 0.14 large. The 0.90 here is a rare, very strong effect (teaching data deliberately well-separated). Always report it alongside p — a large sample can make a trivial difference "significant," and only effect size catches that.

一句话In One Line
ANOVA 的名字叫「方差分析」,但它真正比的是均值 —— 靠的却是把方差(变异)拆开这个巧思。 总变异永远等于组间加组内(SS总 = SS组间 + SS组内),这是一道恒等式。问题只剩:组间这块(信号)是不是显著压过 组内那块(噪声)?把两块各自除以自由度变成「均方」,相除就是 F。 如果各组其实没差别,组间均方和组内均方都只在估计同一个噪声,F 应该在 1 附近徘徊; 一旦组间远大于 1、冲进 F 分布右尾,就说明差异大到不像随机能解释。 这就是为什么一个比均值的问题,最后用方差来回答。 ANOVA is called "analysis of variance," but what it really compares are means — using the clever trick of decomposing the variance. Total variation always equals between plus within (SS_total = SS_between + SS_within); that's an identity, not an assumption. The only open question is whether the between piece (signal) significantly overwhelms the within piece (noise). Divide each by its degrees of freedom to get "mean squares," divide one by the other, and you have F. If the groups are truly identical, both mean squares are estimating the same noise floor, and F should hover near 1. Once F shoots well above 1 into the right tail of its distribution, the difference is too large to be blamed on chance. That is why a question about means is, in the end, answered through variance.
常见误用Common Mistakes
对多组反复做两两 t 检验。这会让总体假阳性暴增,先用一次 ANOVA 总判,再做多重比较(Tukey)控制误差。 Running pairwise t-tests across many groups. That inflates the family-wise false-positive rate. Run one ANOVA first; then use a post-hoc procedure (Tukey HSD) to control the error.
F 显著就以为「每组都不同」。ANOVA 只说「至少有一组不同」,具体哪几组要靠事后两两比较定位。 Reading a significant F as "every group is different". ANOVA only says "at least one group differs." Identifying which ones takes a post-hoc pairwise comparison.
不检查前提就上 ANOVA。它假设各组正态、方差齐性、独立;方差不齐时用 Welch ANOVA 或非参数 Kruskal-Wallis。 Running ANOVA without checking its assumptions. It assumes normality, equal variances, and independence. When variances are unequal, switch to Welch's ANOVA or the non-parametric Kruskal-Wallis test.

ANOVA