故事 · Stevens 的四把尺子
Origin Story · Stevens' Four Rulers
1946 年,哈佛心理学家 S.S. Stevens 在《Science》上发表一篇短文,把所有测量数据分成四级:
名义、序数、区间、比率.他的核心洞见是:数字写出来都长一样,但背后允许的运算天差地别.
球衣上印着「10 号」和「7 号」,你不会说「10 号比 7 号多 3」;可体重 10 公斤确实比 7 公斤多 3 公斤.
同样三个阿拉伯数字,一个是名义标签、一个是比率测量 —— 用错尺度,再漂亮的统计也是空中楼阁.
这就是六西格玛 Measure 阶段第一课:先认清你手里的数据,是哪一种尺.
In 1946 Harvard psychologist S.S. Stevens published a short paper in Science sorting every kind of measurement into four levels:
Nominal, Ordinal, Interval, Ratio. His core insight: numbers may look identical on paper, but the operations they license differ enormously.
Jersey numbers 10 and 7 don't make player #10 "three more" than #7; yet 10 kg really is 3 kg heavier than 7 kg.
Same three digits — one is a label, the other a ratio measurement. Pick the wrong scale and even the slickest statistic collapses.
This is the very first lesson of the Six Sigma Measure phase: before anything else, ask which ruler you're holding.
1 四把尺子:能做什么 · 不能做什么 Four Rulers: What You Can Do · What You Can't
名义 · 信息量最低名义尺度只是「贴标签」:能数频次、能算众数、能画饼图,但不能比大小、不能算平均. 给颜色或邮编算平均数,纯属胡来.
2 信息量阶梯:越往上能做的越多 The Information Ladder: Higher = More Allowed
四种尺度是嵌套递进的:比率包含区间的全部能力,区间包含序数,序数包含名义. 越往右,允许的运算越多、信息越丰富.能用比率尽量用比率. The four scales are strictly nested: Ratio inherits everything Interval can do, Interval inherits Ordinal, Ordinal inherits Nominal. The further right you go, the richer the information and the more legal operations you gain. Whenever you can collect Ratio data, collect Ratio data.
3 现实里的测量尺度 Scales in the Real World
连续 vs 离散:比率/区间常是连续型(尺寸、温度),名义/序数多为离散型(等级、类别).连续数据信息量大,制程偏移更早被发现.
Continuous vs Discrete: Ratio and Interval data are typically continuous (dimensions, temperature); Nominal and Ordinal are discrete (grades, categories). Continuous data carries more information, so process drift shows up earlier.
选对图表:名义→饼图/条形图;序数→堆叠条形;区间/比率→直方图/箱线图/散点图.尺度错,图就误导.
Pick the right chart: Nominal → pie / bar; Ordinal → stacked bar; Interval/Ratio → histogram / boxplot / scatter. Wrong scale, misleading chart.
选对检验:名义→卡方;序数→秩和(Mann-Whitney);区间/比率→t 检验、ANOVA、回归.尺度是选检验的第一道分叉.
Pick the right test: Nominal → chi-square; Ordinal → rank-sum (Mann-Whitney); Interval/Ratio → t-test, ANOVA, regression. Scale is the very first fork in the test-selection tree.
量表陷阱:满意度 1~5 是序数,严格说不能直接算平均;但工程实务常近似当区间用,前提是间距大致等距,需谨慎说明.
The Likert trap: A 1–5 satisfaction scale is strictly Ordinal, so averaging it is technically illegal. Engineering practice often treats it as Interval anyway — but only if the spacing is roughly equal, and you flag the assumption clearly.
一句话In One Line
测量尺度不是学术分类,是分析方法的入场券.它一锤定音地决定:
能算哪个集中趋势(众数/中位数/均值)、画哪种图、跑哪类检验.
区别区间与比率的关键是有没有「绝对零」 —— 摄氏 0 度不是「没有温度」,所以「20 度是 10 度的两倍热」是错的;
而重量 0 公斤是真没有,「20 公斤是 10 公斤的两倍」才成立.
六西格玛 Measure 阶段拿到任何数据,第一件事就是问:这是哪把尺?问清楚了,后面才不会全盘皆错.
Measurement scales aren't academic trivia — they are the entry ticket to every analysis.
They decide, once and for all, which measure of central tendency (mode / median / mean) is allowed, which chart fits, and which test family applies.
The line between Interval and Ratio is one question: is there a true zero? 0 °C is not "no temperature" — so "30 °C is twice as hot as 15 °C" is just wrong;
0 kg really is no mass, so "20 kg is twice 10 kg" holds up.
In the Measure phase, the very first question to ask of any dataset is "which ruler is this?" — get that wrong and everything downstream goes wrong with it.
常见误用Common Mistakes
给名义数据算平均(邮编、工号、颜色编码).名义只能数频次、取众数,平均数没有意义.
Averaging Nominal data (zip codes, employee IDs, color codes). Nominal data only supports frequency counts and the mode — averages are meaningless.
把序数当区间硬算倍数(说「满意度 4 是 2 的两倍」).序数只保证顺序,间距未必相等,慎用均值与比值.
Treating Ordinal as Interval and claiming ratios ("a 4 is twice as satisfied as a 2"). Ordinal preserves rank only — gaps aren't guaranteed equal, so means and ratios deserve skepticism.
说「30°C 是 15°C 的两倍热」.摄氏是区间尺度无绝对零,倍数无意义;要谈倍数得用开尔文.
"30 °C is twice as hot as 15 °C". Celsius is Interval with no true zero, so ratios don't apply — if you really need a "twice as hot" claim, switch to Kelvin.