起源故事 · 一个化名"Student"的酿酒厂质检员 Origin Story · A Brewery Chemist Writing as "Student"
1908 年,都柏林吉尼斯酒厂的化学家 William Gosset 面对一个难题:检验大麦品质、酵母浓度时,他只能取到很小的样本, 经典的正态方法在小样本下不靠谱。他推导出一个能修正小样本偏差的分布。可酒厂禁止员工署真名发论文,他便化名 "Student" 发表 —— 这就是 t 分布 (Student's t)。单样本 t 检验,正是他当年要解决的问题:手头这一小批,到底偏没偏离标准? In 1908 at the Guinness brewery in Dublin, the chemist William Gosset had a problem: testing barley quality and yeast concentration, he could only collect tiny samples, and the classical normal-theory methods broke down at small n. He derived a new distribution that corrected for small-sample bias. Guinness forbade staff from publishing under their real names, so he wrote under the pseudonym "Student" — giving us the Student's t distribution. The one-sample t-test answers exactly the question he was facing: has this small batch in front of me really drifted off the standard?

1 你的样本:拖动每个测量值 Your Sample: Drag Each Measurement

可拖动draggable

蓝点是你抽到的样本,黄虚线是目标 μ₀,紫线是样本均值 x̄。x̄ 离 μ₀ 越远、点越集中(s 越小),t 值越大。 Blue dots are your sampled measurements, the dashed yellow line is the target μ₀, the purple line is the sample mean x̄. The farther x̄ sits from μ₀ and the tighter the cluster (smaller s), the larger the t.

2 t 值落在 t 分布的哪里?拒绝域 vs 安全区 Where Does t Land on the t Distribution? Rejection Tails vs Safe Zone

3 现实里的单样本 t 检验 The One-Sample t-Test in the Real World

净含量合规:抽检产品净含量是否达到标称值,判断是否短斤少两、能否过监管。 Net-content compliance: check whether sampled fill weights hit the label claim — decide if a batch is shorting customers or will pass regulators.
工序均值偏移:当前加工尺寸均值是否偏离设计目标,是 Measure / Analyze 的常用工具。 Process mean shift: is the current machining dimension drifting off the design target? A staple tool in DMAIC's Measure and Analyze phases.
单点对标:实测指标(强度、纯度、响应时间)是否达到合同 / 标准规定的单一目标值。 Benchmark check: does a measured metric (strength, purity, response time) hit a single target value set by a contract or standard?
校准核查:仪器测一组标准样,均值是否显著偏离已知真值,判断是否需要校准。 Calibration check: run a reference standard through the instrument — does the mean drift significantly from the known true value? If so, it needs recalibration.
一句话In One Line
t 值的本质是"偏差有几个标准误那么大":分子是 x̄ 与目标的差,分母是均值的不确定性(标准误 = s/√n)。 小样本时,s 估得不准,所以用比正态更"胖尾"的 t 分布来兜底 —— 这正是 Gosset 的贡献。 同样的偏差,样本越多、越集中,就越可信:因为标准误更小,t 值更大,p 值更小。这就是"一小批能不能说话"的数学。 At heart, t answers "how many standard errors is the gap?" — numerator is x̄ minus the target, denominator is the uncertainty in the mean itself (SE = s/√n). With small samples, s is itself a shaky estimate, so we fall back on the fatter-tailed t distribution — Gosset's gift. The same gap becomes more credible as the sample grows and tightens: SE shrinks, t climbs, p drops. That's the math behind "can this small batch speak for itself?"
常见误用Common Mistakes
n 极小还硬下结论。n 太小标准误大、功效低,结论很脆弱,多取几个点。 Forcing a verdict from a tiny n. A small n means large SE and low power — the conclusion is fragile. Collect more measurements.
数据严重偏态 / 有离群点直接用。先看分布与离群,必要时用非参数检验。 Running it on heavily skewed data or with obvious outliers. Inspect the distribution and outliers first — switch to a nonparametric test (Wilcoxon signed-rank) when assumptions break.
把 p > α 当成"证明相等"。不显著 ≠ 等于目标,只是"没有足够证据说不等"。 Reading p > α as "proof of equality". "Not significant" doesn't mean equal to the target — it only means there isn't enough evidence to say it differs.

单样本 t 检验