故事 · 被忽视的「前提检查」
Origin Story · The Forgotten Precondition Check
统计史上,人们花了大量精力研究怎么比较均值,却常常忘了这些方法都站在一个假设上:各组波动一样大。
Bartlett(1937)提出了一个检验方差齐性的方法,可它对数据偏离正态极其敏感,一旦数据不正态就容易误判。
后来 Levene(1960)换了个思路:不直接比方差,而是比「每个点到组中心的绝对偏差」的均值 ——
这样一来即使数据不那么正态也稳健得多,成了今天工程师最常用的齐性检验。它们看似不起眼,
却是 t 检验、ANOVA 能否站得住脚的地基检查:地基歪了,上面的均值结论就靠不住。
Statisticians poured decades into perfecting tests for means, but those tests all rest on a hidden plank: every group spreads by the same amount.
Bartlett (1937) offered a way to test that plank, yet his statistic is brutally sensitive to non-normality — any skew and it cries wolf.
Levene (1960) reframed the question: don't compare variances directly, compare the mean absolute deviation of each point from its group center.
That tweak makes the test far more robust when the data isn't textbook-normal, and it's now the engineer's default homogeneity check.
Modest as they look, these tests are the foundation check for t-tests and ANOVA: tilt the foundation, and every mean-based conclusion above it leans with it.
1 两组分布:一样胖,还是一胖一瘦? Two distributions: equally fat, or one fat one lean?
σ₁=8 σ₂=82 F = s₁²/s₂² 落在 F 分布的哪一段 Where F = s₁²/s₂² lands on the F-distribution
若两组方差真相等,F = s₁²/s₂² 服从 F 分布、围着 1 波动。 两条红虚线是双侧临界值;黄线(当前 F)落在两线之间判「齐」,落到外侧判「不齐」。 If the variances are truly equal, F = s₁²/s₂² follows an F-distribution centered on 1. The two red dashed lines are the two-sided critical values; the yellow line (current F) inside the band reads "equal", outside reads "unequal".
3 现实里的方差检验 Variance Tests in the Real World
方差齐性:两条产线均值差不多,但一条波动大一条波动小,方差检验直接量出谁更稳。
Homogeneity: two production lines have similar means but one jitters and one stays steady — the variance test calls out which line is the stable one.
t 检验前提:跑双样本 t 前先验齐性;不齐就改用不假设等方差的 Welch t,避免 p 值失真。
Precondition for t-tests: check homogeneity before a two-sample t-test; if it fails, switch to Welch's t-test (no equal-variance assumption) to keep the p-value honest.
Levene 检验:多组、数据非正态时的首选齐性检验,比 Bartlett 稳健,DOE 残差诊断常用。
Levene's test: the go-to homogeneity check for many groups or non-normal data — more robust than Bartlett and a staple of DOE residual diagnostics.
过程波动对比:改善前后的方差是否真的降了?方差检验是验证「变稳了」的硬证据。
Before/after process spread: did variance actually drop after the improvement? The variance test is the hard evidence that the process truly became steadier.
一句话In One Line
均值告诉你「中心在哪」,方差告诉你「散得多开」。比均值的检验之所以要求方差齐性,
是因为它们把各组的波动合并成一个共同的「噪声尺度」来算标准误 —— 一旦各组胖瘦悬殊,这个合并就名不副实,
算出的 t 或 F 不再服从假定的分布,p 值随之失真。方差检验干的就是「先验地基」:
两组用 F = s₁²/s₂² 量比值偏离 1 多远;多组用 Levene 把每个点到中心的绝对距离再跑一遍 ANOVA。
结论很实用 —— 齐,就放心用经典 t / ANOVA;不齐,就换 Welch 或非参数方法。
在六西格玛里,方差本身往往就是改善目标:把波动压下来,比把均值挪一点更有价值。
The mean tells you where the center sits; the variance tells you how far it spreads. Mean-based tests demand homogeneity of variance
because they pool every group's spread into one shared "noise scale" to build the standard error. The moment groups disagree on spread, that pool is
misnamed: the resulting t or F no longer follows its assumed distribution, and the p-value drifts off course.
Variance tests do the foundation check: F = s₁²/s₂² measures how far the ratio strays from 1 for two groups; Levene runs ANOVA on the absolute deviations from each center for many.
The takeaway is practical — equal spread → use classic t / ANOVA with confidence; unequal → swap in Welch or a non-parametric test.
And in Six Sigma, variance is often the improvement target itself: knocking the spread down beats nudging the mean.
常见误用Common Mistakes
不验方差齐性就直接比均值。方差悬殊时经典 t / ANOVA 的 p 值会失真,先做齐性检验再决定方法。
Comparing means without testing homogeneity first. When variances disagree, classic t / ANOVA p-values are distorted — run the homogeneity test first, then pick your method.
对明显非正态的数据用 Bartlett / F。它们对正态假设很敏感,偏态数据改用对非正态稳健的 Levene。
Using Bartlett or F on clearly non-normal data. Both are touchy about normality — for skewed data, reach for Levene or Brown-Forsythe, which hold up under non-normality.
方差不齐就硬套等方差公式。改用 Welch t / Welch ANOVA(不合并方差),结论才可信。
Forcing the equal-variance formula when variances clearly differ. Switch to Welch's t-test / Welch ANOVA, which never pools variances — only then is the conclusion trustworthy.