起源故事 · Fisher 与 Rothamsted 的麦田
Origin Story · Fisher and the Rothamsted Wheat Plots
1920 年代,统计学家 R.A. Fisher 在英国 Rothamsted 农业试验站研究不同肥料对小麦产量的影响。
他要同时比较好几种处理 —— 如果两两做 t 检验,比 4 组就要做 6 次,每次都有 5% 出错的机会,整体早就错得离谱。
Fisher 灵光一闪:与其逐对比较,不如把方差拆开 —— 看"组与组之间的差距"相对于"组内本身的波动"是否足够大。
这就是 方差分析(ANOVA),今天每一次 DOE、每一份田间试验报告里都有它。
In the 1920s, the statistician R.A. Fisher was studying how different fertilizers affect wheat yield at the Rothamsted Experimental Station in England.
He needed to compare several treatments at once — and pairwise t-tests would mean six tests for four groups, each with a 5% chance of a false positive. The overall error rate would be a disaster.
Fisher's insight was to decompose variance instead of comparing pairs: ask whether "the spread between group means" is big relative to "the variation within each group."
That's Analysis of Variance (ANOVA), and it sits at the heart of every DOE and field-trial report written today.
1 三组数据:差距大,还是只是巧合? Three Groups: A Real Gap, or Just Coincidence?
观察散点scatter view2 方差分解:F 就是信号比噪声 Variance Decomposition: F Is Signal Over Noise
总变异被劈成两块:组间(处理造成的,分子)和 组内(随机噪声,分母)。 F = MSB / MSW。F 远大于 1,说明信号压过噪声,差异是真的。 Total variation splits in two: between-group (driven by the treatments — the numerator) and within-group (pure random noise — the denominator). F = MSB / MSW. An F well above 1 says the signal beats the noise and the difference is real.
3 现实里的单因素 ANOVA One-Way ANOVA in the Wild
多机台对比:三台注塑机的尺寸有没有系统差异?一次 ANOVA 替代多次两两 t。
Machine-to-machine comparison: do three injection-molding machines drift on dimensions? One ANOVA replaces a pile of pairwise t-tests.
多配方筛选:四种催化剂的得率谁更高?先 ANOVA 判"有没有差",再做事后比较找"谁最好"。
Screening formulations: four catalysts, which one delivers the highest yield? ANOVA first answers "is there any difference," then post-hoc tests find "which one wins."
多供应商质量:来料强度按供应商分组比较,识别哪家波动大或均值偏。
Multi-supplier quality: group incoming strength by vendor to spot which one runs wider, or sits off-target.
多组改善效果:三个班组的改善后产出是否真有差距,而非仅靠运气。
Cross-team improvement: post-Improve output across three crews — is the gap real performance, or just luck?
一句话In One Line
ANOVA 的名字叫"方差分析",听起来像在比方差,其实是在用方差去比均值。
它把"组间差距"和"组内噪声"放进同一把尺子(F 比),一次回答"这些组里至少有一组不一样吗"。
但它只告诉你有没有差异,不告诉你谁和谁不同 —— 那要靠事后的多重比较(Tukey 等)。F 显著只是故事的开头。
ANOVA literally means "analysis of variance," but it isn't really comparing variances — it's using variance to compare means.
It puts between-group gaps and within-group noise on the same ruler (the F-ratio) and answers, in one shot, "is at least one group different from the others?"
What it doesn't tell you is which groups differ from which — that's the job of post-hoc multiple-comparison procedures (Tukey HSD and friends). A significant F is only the opening line of the story.
常见误用Common Mistakes
用多次两两 t 检验代替 ANOVA。多组一次性用 ANOVA,避免错误率累积膨胀。
Replacing ANOVA with a stack of pairwise t-tests. For three or more groups, run one ANOVA — the family-wise error rate doesn't inflate.
F 显著就宣布"某两组不同"。F 只说"至少一组不同",谁不同要做事后多重比较。
Reading a significant F as "groups X and Y differ". F only says "at least one group is different." Use a post-hoc test (Tukey HSD) to find which pairs actually differ.
无视前提硬套(组内非正态、方差极不齐)。先查正态与方差齐性,不满足改用 Welch 或非参数。
Forcing ANOVA without checking the assumptions (non-normal residuals, badly unequal variances). Check normality and homogeneity of variance (Levene's test) first; if they fail, switch to Welch's ANOVA or a non-parametric option like Kruskal–Wallis.