故事 · 一个被偷偷假设的前提 Story · A Quietly Assumed Premise
一位工程师拿着新工序的 30 个尺寸,直接套了过程能力公式,算出 Cpk = 1.5,开心地报告"能力达标"。 可客户抽检却频频超差。问题出在哪?这批数据其实右偏得厉害 —— 长尾全压在规格那一侧, 而 Cpk 的公式默认数据是对称正态的,算出的"能力"是一个海市蜃楼。 这就是为什么统计学家把正态性检验放在所有参数方法的最前面:自 1900 年代 Karl Pearson 起, 人们就用正态概率图让肉眼判断;后来 Anderson-DarlingShapiro-Wilk 把"像不像直线" 量化成一个 p 值。用任何假设正态的工具之前,先问这一句,能省下后面无数返工。 An engineer grabbed 30 measurements from a new process, plugged them straight into the capability formula, got Cpk = 1.5, and cheerfully reported "capable." Then customer inspections kept flagging out-of-spec parts. What went wrong? The data was severely right-skewed — the long tail sat squarely on the spec side, and Cpk's formula presumes symmetric normal data. The "capability" it returned was a mirage. This is exactly why statisticians put the normality test at the front of every parametric method: since Karl Pearson in the early 1900s, people have used the probability plot for an eyeball check; later Anderson-Darling and Shapiro-Wilk turned "how close to a straight line" into a single p-value. Asking this one question before any normality-dependent tool saves a mountain of rework downstream.

1 正态概率图:点贴直线就正态 Probability Plot: Points on the Line Mean Normal

点贴线on the line

2 直方图叠正态曲线:形状对得上吗 Histogram + Theoretical Bell: Do the Shapes Line Up?

蓝柱是真实数据的分布,红线是用同一均值与标准差画出的理论正态曲线。正态时两者贴合;右偏时柱子会向右拖出长尾,双峰时会冒出红线根本盖不住的第二座山。 Blue bars are the actual data; the red line is the theoretical normal curve drawn from the same mean and standard deviation. When data is normal the two hug each other; right-skewed data drags a long tail to the right; bimodal data sprouts a second hill the red line can't begin to cover.

3 现实里的正态性检验 Normality Tests in the Real World

t 检验前的体检:双样本 t、配对 t 都假设正态,n 小又偏态时先验一把,否则结论可能反向。 The pre-flight check before a t-test: two-sample and paired t-tests both assume normality. With small, skewed samples, check first — otherwise your conclusion can flip the wrong way.
过程能力前提:Cp / Cpk 的良率换算建立在正态上,数据偏态会高估或低估能力,必须先验。 The capability prerequisite: Cp / Cpk translate spread into yield via the normal curve. Skewed data over- or under-states real capability — check before you trust the number.
Q-Q 图先于 p 值:检验只告诉你"像不像",概率图告诉你"哪里不像" —— 尾部翘起还是中间鼓起。 Plot before p-value: the test only tells you whether it looks normal; the probability plot tells you where it fails — tails curling up, or the middle bulging.
不正态怎么办:换非参数方法、做 Box-Cox 变换,或拆分混入的两批数据,而不是硬套正态公式。 When it isn't normal: switch to a non-parametric test, apply a Box-Cox transform, or split out the two mixed batches — don't force a normal formula on non-normal data.
一句话In One Line
正态性检验不是一道选做题,而是参数统计的门票。t 检验、方差分析、Cpk、控制限, 它们的公式里都焊死了一个"数据正态"的假设;假设错了,公式照样吐出一个漂亮数字,只是这个数字不对。 记住两件事:第一,先看图再看 p —— 概率图告诉你偏在哪里,p 值只给一个是非判断; 第二,p > 0.05 不是"证明正态",只是"没有证据说它不正态"。 样本越大,越小的偏离也会被检验抓出来,所以图与数要一起读。 Normality testing isn't optional — it's the entry ticket for parametric statistics. t-tests, ANOVA, Cpk, control limits — their formulas all have "data is normal" welded in. Get the assumption wrong and the formula still spits out a pretty number, but that number is wrong. Remember two things. First, look at the plot, then read the p-value — the plot tells you where the deviation lives; the p-value only gives a yes/no. Second, p > 0.05 doesn't "prove normal" — it just says "no evidence to reject normal." The larger your sample, the smaller the deviation the test will catch — so always read the plot and the number together.
常见误用Common Mistakes
跳过检验直接套正态公式用 Cpk / t 检验前先做正态性检验,省下后面返工。 Skipping the check and plugging into normal formulas anyway. Run a normality test before Cpk or any t-test — it saves rework later.
把 p > 0.05 当成"数据是正态的"它只是"没有足够证据拒绝正态",不等于证明。 Reading p > 0.05 as "the data is normal". It only means "not enough evidence to reject normality" — that is not the same as proof.
只看 p 值不看概率图大样本下微小偏离也会让 p 变小,看图才知道偏离是否要紧。 Reading the p-value alone, ignoring the probability plot. With large n even tiny deviations push p down. The plot tells you whether the deviation actually matters.

正态性检验