起源故事 · 当"平均合格"还不够
Origin Story · When "Good on Average" Isn't Enough
早期质量管理只盯均值:平均尺寸对了就算合格。但工程师很快发现,波动本身就是敌人 —— 两批均值相同的零件,
波动大的那批装配时麻烦不断。Fisher 的 F 分布给出了比较两个方差的工具,后来 Levene、Bartlett 又把它推广到多组。
到了六西格玛时代,"减小方差"更被奉为核心 —— 因为顾客感受到的质量,往往不是平均值,而是那一次糟糕的例外。
Early quality control watched only the mean: if the average size was right, the lot passed. Engineers learned the hard way that variation itself is the enemy — two lots with identical means, one wide and one tight, behave very differently on the assembly line.
Fisher's F-distribution gave us the first tool for comparing two variances; Levene and Bartlett later extended it to many groups.
By the Six Sigma era, "shrink the variance" had become a core mantra — because customers don't feel the average. They feel that one bad part.
1 同一个中心,谁宽谁窄? Same Center — Who's Wide, Who's Narrow?
观察宽窄2 实测点:谁的点子甩得更开 Actual Data: Whose Points Scatter More
两条带子是各组的实测值,灰线是公差区间。机器 1 的点挤在中间,机器 2 的点甩到边缘甚至越界 —— 均值达标也救不了波动大。 Each strip is one group's measured values; the dashed lines are LSL/USL. Machine 1 stays clustered near the center while Machine 2 flings points to the edges and out of spec — a good mean cannot rescue bad spread.
3 现实里的方差检验 Equal-Variance Tests in the Real World
设备稳定性:新旧两台设备均值都达标,比方差决定谁更适合精密工序。
Equipment steadiness: when an old and a new machine both hit the mean target, comparing variances decides which one belongs on the precision job.
工艺一致性:改善前后波动有没有真的变小,是六西格玛改进的硬指标。
Process consistency: did variation truly shrink after the improvement? That's the hard metric every Six Sigma project owes its sponsor.
风险波动比较:两个投资组合收益均值相近时,比方差就是比风险大小。
Risk comparison: when two portfolios show similar mean returns, comparing variances is exactly comparing risk.
供应商质量稳定:均值合格还不够,方差小的供应商才省下游麻烦。
Supplier quality: hitting the mean isn't enough — the supplier with smaller variance is the one that saves your downstream line.
一句话In One Line
t 检验比中心,方差检验比离散 —— 它回答的是"谁更稳"这个被低估的关键问题。
很多 ANOVA、t 检验都默默假设"各组方差相等",所以方差检验也常作为前置体检:方差不齐就得换 Welch 或非参数方法。
但要小心:F 检验对正态非常敏感,数据偏一点就会误判,工程实务里更稳的是 Levene 检验。看懂"窄=稳、宽=飘",你就握住了质量改进最朴素的直觉:把分布收窄。
A t-test compares centers; an equal-variance test compares spreads — it answers the underrated question of "who is steadier?"
ANOVA and the standard t-test quietly assume equal variances, so a variance test often runs first as a pre-flight check: if the assumption fails, switch to Welch or a nonparametric route.
One warning: the F-test is extremely sensitive to non-normality, so on real-world skewed data the Levene test is the safer choice. Once you see "narrow = steady, wide = jittery," you've grasped quality improvement's simplest instinct: shrink the distribution.
常见误用Common Mistakes
数据明显非正态还用 F 检验。F 对正态极敏感,偏态数据改用 Levene 检验更稳健。
Running an F-test on visibly non-normal data. The F-test is fragile under non-normality; use the Levene (or Brown-Forsythe) test for skewed data — it's far more robust.
只比均值忽略方差。均值达标但方差大 = 不稳定,照样是质量问题。
Comparing only means and ignoring variance. A good mean with a fat variance is still a quality defect — the unsteady process bites you on the line, not on the report.
跳过方差齐性直接做 ANOVA/t。先查方差齐性,不齐就用 Welch 或非参数方法。
Jumping straight into ANOVA or a pooled t-test without checking equal variance. Always test the equal-variance assumption first; if it fails, switch to Welch's t-test or a nonparametric alternative.