起源故事 · 化名 "Student" 的酿酒师
Origin Story · The Brewer Who Published as "Student"
1908 年，都柏林健力士(Guinness)啤酒厂的酿酒师 William Sealy Gosset 遇到一个难题：判断一批大麦好不好，
只能拿到几个样本，正态分布在小样本下严重低估误差。他硬是推导出一条更厚尾的新曲线来补偿。
可健力士禁止员工发论文（怕泄露商业机密），他只好化名 "Student" 偷偷发表 —— 这就是 Student's t 的由来。
后来 Fisher 把它的两个亲戚补全：χ² 管方差与拟合，F（纪念 Fisher）管两组方差之比，也就是方差分析的引擎。
In 1908, William Sealy Gosset, a brewer at Dublin's Guinness brewery, had a problem: he had to judge a barley batch from just a handful of samples, and the normal distribution severely understated the error at such small sample sizes.
He worked out a new, heavier-tailed curve to make up for it. Guinness forbade staff from publishing — they were terrified of leaking trade secrets — so he submitted the paper under the pseudonym "Student". That is where the name Student's t comes from.
Ronald Fisher later rounded out the family: χ² handles variance and goodness-of-fit, and F — named after Fisher himself — handles the ratio of two variances, which is the engine inside analysis of variance.
1t 分布：样本越少，曲线越矮胖、尾巴越厚Student's t: fewer samples → shorter, wider curve, heavier tails
t 分布t-distribution2 厚尾的代价：要达到 95% 置信，门槛被推得更远 The cost of fat tails: 95% confidence pushes the threshold farther out
3 三兄弟各管一摊：现实里的 t / χ² / F Three siblings, three jobs: t / χ² / F in practice
t 检验：小样本比较均值。两条产线的平均强度有没有差？改善前后均值动了没？样本少时的首选。
t-test: comparing means on small samples. Do the two production lines really have different mean strength? Did the average move between before-and-after the improvement? The default choice when n is small.
χ² 检验：拟合优度与独立性。实际缺陷分布是否符合预期？两个分类变量是否相关（列联表）？
χ² test: goodness-of-fit and independence. Does the observed defect mix match what was expected? Are two categorical variables related in a contingency table?
F 检验：方差之比。两台设备波动谁更大？方差分析(ANOVA)判断多组均值是否齐，引擎就是 F。
F-test: ratio of variances. Which of two machines jitters more? Analysis of variance (ANOVA) — which decides whether several group means are equal — is powered by F.
自由度 df：算出某个统计量后还剩多少"自由变动的信息"。它是这三条曲线唯一的形状旋钮。
Degrees of freedom (df): how many pieces of information are still free to vary after you compute a statistic. It is the single shape knob that controls all three curves.
一句话
In One Line
正态分布假设你早就知道总体标准差；现实里我们只有样本、连标准差都得估，这份额外的不确定就化成了 t 分布那条更厚的尾巴。
df 越大，等于样本越多、估计越靠谱，曲线就一点点收回正态 —— t 分布是正态在"信息不足"时的诚实版本。
χ² 把方差信息攒起来，F 把两份 χ² 相除比方差 —— 三者同源，都是抽样分布，是下游所有假设检验（第 33 页）的弹药库。
The normal distribution assumes you already know the population standard deviation. In real life we only have a sample — even the standard deviation has to be estimated — and that extra uncertainty is what becomes those heavier tails on the t curve.
The larger df is, the more samples you have and the more reliable your estimate, so the curve slowly contracts back into the normal — Student's t is the honest version of the normal when information is scarce.
χ² aggregates variance information; F divides one χ² by another to compare variances. All three are sampling distributions cut from the same cloth, and together they form the ammunition depot behind every hypothesis test downstream (see page 33).
常见误用
Common Mistakes
小样本硬用 z（正态）检验。总体标准差未知、样本小，就该用 t 检验，否则低估误差、假阳性飙升。
Forcing a z-test on a small sample. If the population standard deviation is unknown and n is small, you need a t-test — otherwise you underestimate the error and your false-positive rate explodes.
把自由度算错（如忘了减 1）。单样本 df = n−1；查表与软件输出的 df 必须对上。
Getting df wrong — typically forgetting to subtract one. A one-sample test uses df = n−1. Make sure the df on your lookup table matches the df your software is reporting.
χ² 检验期望频数太小还硬算。每格期望频数一般需 ≥5，否则合并类别或换精确检验。
Running χ² when expected cell counts are tiny. Each expected count should generally be ≥ 5 — otherwise merge categories or switch to an exact test such as Fisher's exact.