故事 · Student 的另一半智慧
Origin Story · Student's Other Stroke of Genius
当年酿酒师 William Gosset（笔名 Student）发明 t 检验，是为了在小样本下比较两批大麦、两种工艺。
可工程师很快发现一个更巧的招数：与其找两群互不相干的对象，不如让同一批对象各自当自己的对照 ——
减肥药试验中量同一个人的前后体重，机加工里测同一台机床改造前后的尺寸。每个对象自带「基线」，
做差之后，人与人之间的天然差异（噪声）被彻底抵消，只剩处理真正带来的变化。这就是配对设计：
它不增加成本，却能让微弱的真效应从噪声里浮出水面 —— 是实验设计里性价比最高的一招。
The brewer William Gosset (pen name "Student") invented the t-test to compare two barley batches or two brewing processes with small samples.
Engineers soon spotted a cleverer move: instead of finding two unrelated groups, let each subject serve as its own control —
weigh the same person before and after a diet drug, measure the same machine before and after a retrofit. Each subject brings its own "baseline",
and once you subtract, all the natural variation between subjects (the noise) cancels out, leaving only the change the treatment really caused. That's paired design:
no extra cost, yet a weak true effect can rise out of the noise. It's the highest-leverage trick in all of experimental design.
1 前测 → 后测：每条连线是同一个人 Pre → Post: Each Line Is One Subject
n=122 同一份数据，两种算法，两条命运 Same Data, Two Methods, Two Fates
两条 t 分布判定带：配对 t把个体差异从分母里踢掉，t 值冲过临界线、显著；独立 t却把个体差异当成误差，分母被撑大、t 值缩水，常常落在临界线内、判「不显著」。 Two verdict bars: the paired t kicks between-subject spread out of the denominator, so its t shoots past the critical line — significant. The independent t rolls that spread into the error term; the denominator balloons, the t-value shrinks, and it usually lands inside the critical line — "not significant".
3 现实里的双样本与配对 Two-Sample vs Paired in the Real World
前后对比：减肥前后体重、培训前后成绩、改善前后周期时间，同一批对象两次测，配对 t 是默认选择。
Before vs after: weight before/after a diet, scores before/after training, cycle time before/after a kaizen — same subjects measured twice, paired t is the default.
AB 同源测试：同一组用户先看 A 再看 B，同人对比剔除用户间偏好差异，比两群人 AB 更灵敏。
Within-subject A/B: have the same users see A then B; comparing within a person strips out taste differences between users and is far more sensitive than comparing two separate cohorts.
配对降噪：左右手握力、同板两点温度、同批料两种处理，配对把「来源差异」清零，小样本也能出结论。
Pairing kills noise: left vs right grip strength, two points on the same PCB, two treatments on the same lot of material — pairing zeroes out the source-level differences and lets a small n reach a conclusion.
独立两组：A 班 vs B 班、新供应商 vs 老供应商，对象本就是两群不相干的人，才用双样本 t。
Truly independent groups: Class A vs Class B, new supplier vs old supplier — when the subjects really are two unrelated groups, the two-sample t is what you want.
一句话In One Line
t 检验的本质是「信号 ÷ 噪声」：分子是均值差（信号），分母是它的标准误（噪声）。
双样本 t 的分母里塞进了两组各自的全部波动 —— 包括人和人之间天生的差异，这部分噪声你并不关心，却被迫一起承担。
配对 t 的高明之处在于：先对每个对象做差，个体差异在做差那一刻就被消掉了，
分母里只剩「差值的波动」这个干净得多的噪声。同样的信号，更小的噪声，t 值自然更大、功效更高。
所以判断用哪个的唯一标准是：两组数据之间是否存在天然的配对关系。有，就别浪费它。
Every t-test is "signal ÷ noise": the numerator is a mean difference (signal), the denominator is its standard error (noise).
The two-sample t crams all the spread in both groups into that denominator — including the innate variation between people, which you don't care about but get charged for anyway.
The paired t's elegance: subtract within each subject first, and between-subject variation evaporates in that single step;
the denominator only carries the "spread of the differences" — a far cleaner noise term. Same signal, smaller noise, naturally larger t, naturally higher power.
So the only question that decides which test to use is: is there a natural pairing between the two columns of data? If yes, don't waste it.
常见误用Common Mistakes
明明是配对数据，却用双样本 t。同一对象的前后测、同源样本必须用配对 t，否则个体差异淹没真效应、检不出。
Running a two-sample t on paired data. Before/after on the same subject or within-subject samples must use paired t — otherwise between-subject variation buries the real effect and the test misses it.
把两群不相干的人硬凑成「对」。配对要求逐一对应（同人、同批、匹配特征），随意配对会引入假关联。
Forcing two unrelated groups into artificial "pairs". Pairing must have a one-to-one link (same person, same lot, matched on key traits). Pairing at random injects spurious correlation.
双样本 t 直接套等方差公式。两组方差差很多时改用 Welch t（不假设等方差），先做方差齐性检验。
Defaulting to the pooled-variance two-sample formula. When the two group variances are clearly unequal, switch to Welch's t (no equal-variance assumption) — check homogeneity of variance first.