故事 · Neyman 的"频率主义"发明 Origin Story · Neyman's Frequentist Invention
1937 年,波兰统计学家 Jerzy Neyman 正式提出置信区间。他特意强调:置信区间的"95%"是关于抽样过程的, 不是关于某一个具体区间的。一旦你算出了 [12.3, 15.7] 这个具体数字,真值要么在里面、要么不在,没有"95% 概率"可言。 95% 描述的是"如果我反复用这个方法造区间,长期看 95% 会命中"。这是频率主义的核心,也是最常被误读的一句话。 In 1937, the Polish statistician Jerzy Neyman formally defined the confidence interval. He took pains to stress that the "95%" describes the sampling procedure, not any single computed interval. Once you crunch the numbers and get [12.3, 15.7], the true value is either inside or it isn't — no "95% probability" applies to it. The 95% says: "if I keep building intervals this way forever, 95% of them will hit the target." That is the heart of frequentist inference — and the sentence most often misquoted.

1 反复抽样:一条条区间,约 5% 会漏掉真值 Repeated Sampling: About 5% of Intervals Miss the Truth

覆盖率 —Coverage —

竖直紫线是固定的真值 μ。每条横线是一次抽样算出的置信区间,中点是样本均值 x̄。 绿色=盖住了真值,红色=漏掉了。点"持续抽样"看覆盖率逼近置信水平。 The vertical purple line is the fixed true mean μ. Each horizontal line is one confidence interval from one sample, centered on the sample mean x̄. Green = the interval covers μ; red = it misses. Hit "Keep sampling" and watch coverage drift toward the confidence level.

2 覆盖的是"过程",不是"某一个区间" The Procedure Has Coverage, Not Any One Interval

横轴是已抽样次数,纵轴是累计覆盖率。蓝线随次数增加,会在目标置信水平附近抖动收敛。这就是"95%"真正的含义 —— 一个长期频率,不是单个区间的概率。 x-axis is the number of samples drawn so far; y-axis is cumulative coverage. The blue line jitters but converges around the target confidence level as draws pile up. That is what "95%" really means — a long-run frequency, not the probability of any single interval.

3 现实里的置信区间 Confidence Intervals in the Real World

民调误差:"支持率 52%,误差 ±3%"就是一个 95% 置信区间,半宽 3% 即 margin of error。 Poll margin of error: "52% support, ±3 points" is a 95% confidence interval. The half-width of 3% is the margin of error.
区间估计:报告均值别只给点估计,附上区间才说清精度有多高、不确定有多大。 Interval estimation: never report a mean as a single point. Pair it with an interval so the precision and uncertainty are visible.
置信水平误解:把"95% 区间"读成"真值有 95% 概率在内"是经典错误,真值并无概率。 Classic misreading: treating "95% interval" as "the true value has a 95% chance of being inside" is the textbook mistake. The truth has no probability — it just is.
样本量与精度:想让区间缩一半,n 要变成 4 倍 —— 精度的代价是平方级的样本。 Sample size vs precision: to halve the interval width, you need 4× the sample size. Precision costs samples quadratically.
一句话In One Line
置信区间公式 x̄ ± z·(σ/√n) 里,x̄ 和区间端点都是随机的,因为它们随样本而变; 真值 μ 却是固定不动的。所以"概率"属于造区间这个过程,不属于造好的某一个区间。 置信水平就是这个随机过程的长期命中率。把它读成"真值落在这个区间的概率",是把随机的对象和固定的对象弄反了 —— 看完上面的动画,这一点就再也忘不了。 In the formula x̄ ± z·(σ/√n), x̄ and the endpoints are random — they shift with every fresh sample. But the true mean μ is fixed. So the "probability" lives in the procedure that builds the interval, never in any one interval after it's built. The confidence level is that procedure's long-run hit rate. Reading it as "the chance the truth lies in this interval" swaps the random object with the fixed one — once you've watched the animation above, you won't make that swap again.
常见误用Common Mistakes
"真值有 95% 概率落在这个区间"。真值固定无概率;95% 是过程的长期覆盖率。 "The true value has a 95% probability of being in this interval". The truth is fixed — no probability attaches to it. The 95% is the procedure's long-run coverage.
区间越窄越好就拼命求窄。窄要靠加大 n,硬降置信水平换来的窄是骗自己。 Chasing narrowness for its own sake. Earn narrower intervals by raising n. Buying narrowness by lowering the confidence level is fooling yourself.
小样本、σ 未知仍用 z。该用 t 分布,t 比 z 略宽,补偿用 s 估 σ 的不确定。 Using z with a small sample and unknown σ. Switch to the t distribution. Critical value t α/2 is slightly larger than z to absorb the extra uncertainty from estimating σ with s.

置信区间