故事 · 啤酒厂里诞生的 Student
Origin Story · "Student" Was Born in a Brewery
1908 年,吉尼斯酒厂的化学家 William Gosset 为小样本啤酒质检发明了 t 分布:t = (x̄ − μ) / (s/√n),
用样本标准差 s 替代未知的总体 σ,代价是尾部更厚。酒厂禁止员工发表,他只好化名"Student",t 检验从此叫学生 t 检验。
另两根支柱:卡方 χ² = Σ(Oᵢ − Eᵢ)²/Eᵢ 衡量观测与期望的偏离(拟合优度、独立性);Fisher 的 F = s₁²/s₂² 比较两个方差,是方差分析(ANOVA)的核心。
三者都随自由度 ν 变形,ν 足够大时统统趋近正态。
In 1908, Guinness brewery chemist William Gosset invented the t distribution to inspect small batches of beer:
t = (x̄ − μ) / (s/√n), substituting the sample standard deviation s for the unknown population σ — at the cost of heavier tails.
Guinness forbade employees from publishing, so he signed his paper "Student", and ever since we call it Student's t-test.
The other two pillars: χ² = Σ(Oᵢ − Eᵢ)² / Eᵢ measures how far observed counts wander from expected (goodness-of-fit, independence); and Fisher's F = s₁² / s₂² compares two variances — the core test statistic in analysis of variance (ANOVA).
All three deform with the degrees of freedom ν, and as ν grows large each one converges back to the normal.
1 t 分布:厚尾 → 正态(小样本均值检验) t Distribution: Heavy Tails → Normal (Small-Sample Mean Tests)
厚尾修正Heavy-tail correction2 卡方 χ² 分布:右偏 → 对称(方差 / 拟合优度) Chi-Square χ² Distribution: Right-Skewed → Symmetric (Variance / Goodness-of-Fit)
3 F 分布:两方差之比(ANOVA 引擎) F Distribution: Ratio of Two Variances (The ANOVA Engine)
4 现实里的三大抽样分布 The Three Sampling Distributions in Practice
t 检验:小样本下比较均值(新工艺 vs 旧工艺)、给均值算置信区间,用 t 替正态修正厚尾。
t-test: compare means on a small sample (new process vs. old) and build confidence intervals for the mean — using t in place of the normal accounts for the heavy tails.
卡方检验:拟合优度(实际分布像不像理论)、列联表独立性(两分类变量是否相关)。
Chi-square test: goodness-of-fit (does the observed distribution match theory?) and contingency-table independence (are two categorical variables related?).
ANOVA:F 比较"组间方差/组内方差",一次检验多组均值是否有差异,是 DOE 的基础。
ANOVA: F compares "between-group variance / within-group variance" to test several group means at once — the statistical backbone of designed experiments (DOE).
自由度:三者的形状都由 ν 控制,本质是"被数据约束后还剩多少独立信息"。
Degrees of freedom: ν governs the shape of all three, and intuitively it counts how much independent information remains after the data have already imposed their constraints.
一句话In One Line
正态分布是"已知 σ"的理想世界,但现实里 σ 通常未知,只能用样本估计 —— 这一估计就引入了额外的不确定性。
t、卡方、F 就是为"用样本估总体"量身定做的修正分布,由自由度 ν 控制偏差大小。
记住一句话:样本越大,ν 越大,三者越像正态 —— 小样本时千万别拿正态临界值硬套,那会让你高估自己的把握。
The normal lives in an idealised world where σ is already known. Real life never is — you estimate σ from the sample, and that single act injects extra uncertainty.
t, χ² and F are the corrections custom-built for that "estimate the population from a sample" reality, with the degrees of freedom ν dialling how big the correction needs to be.
One thing to remember: larger sample → larger ν → all three look more like the normal. With small samples, never reach for the normal critical value — you will badly overstate your confidence.
常见误用Common Mistakes
小样本用正态临界值 1.96。σ 未知、n 小要用 t 分布,临界值更大。
Using the normal critical value 1.96 on a small sample. With σ unknown and n small, use the t distribution — the critical value is larger.
卡方期望频数太小(<5)还硬算。期望频数过小要合并类别或换精确检验。
Running chi-square with expected frequencies below 5. Collapse sparse categories or switch to Fisher's exact test when expected counts get too small.
F 检验对正态偏离很敏感却不查前提。先验证正态性,必要时改用稳健 / 非参方法。
Trusting an F-test without checking its assumptions. The F-test is sensitive to departures from normality — verify normality first, and reach for a robust or non-parametric alternative when needed.