起源故事 · Neyman 与一个被误读了八十年的概念
Origin Story · Neyman and a Concept Misread for Eighty Years
1937 年,波兰裔统计学家 Jerzy Neyman 提出置信区间,并反复强调它的真正含义:
"95% 置信"说的不是某一个具体区间,而是这套抽样方法的长期表现 —— 重复一万次,约 9500 次造出的区间会盖住真值。
真值 μ 是一个固定的、不动的常数,谈不上"它有多大概率在区间里"。
会动的是区间(每次样本不同,区间就左右漂移)。Neyman 把不确定性放在"方法"上,而不是"真值"上 —— 这个微妙的视角转换,是频率派推断的基石。
In 1937 the Polish statistician Jerzy Neyman formally introduced the confidence interval, and spent the rest of his career insisting on what it really meant:
"95% confidence" is not a statement about any single interval, but about the long-run behavior of the procedure — run it ten thousand times and roughly 9,500 of the intervals will cover the true value.
The true mean μ is a fixed, stationary constant; it makes no sense to ask "what is the probability it lies inside the interval?"
What actually moves is the interval itself — each new sample shifts it left or right. Neyman placed the uncertainty on the method, not on the unknown parameter. That subtle reframing is the foundation of frequentist inference.
1 反复抽样:会动的是区间,不是真值 Repeated Sampling: The Intervals Move, the Truth Doesn't
尚未抽样no samples yet2 样本均值从哪来:抽样分布与区间宽度 Where x̄ Comes From: The Sampling Distribution Sets the Width
样本均值 x̄ 围绕真值 μ 波动,服从标准误为 σ/√n 的钟形。区间宽度 = z 值 × 标准误 —— 所以提高置信水平(z 变大)或减小样本量(标准误变大),区间都会变宽。 The sample mean x̄ jitters around the true μ on a bell with standard error σ/√n. The interval half-width — the margin of error — is z × SE, so raising the confidence level (bigger z) or shrinking the sample (bigger SE) both make the interval wider.
3 现实里的置信区间 Confidence Intervals in the Real World
民调误差范围:"支持率 52%,误差 ±3%"就是个 95% 置信区间。±3% 正是半个区间宽度。
Poll margin of error: "52% support, ±3 points" is a 95% confidence interval in disguise. The ±3 is exactly the half-width.
A/B 测试:转化率提升的区间若跨过 0,说明这点提升可能只是抽样噪声,别急着上线。
A/B testing: if the confidence interval for the conversion-rate lift crosses zero, the apparent gain may just be sampling noise — don't ship the change yet.
质量验收:均值的置信区间是否整段落在规格内,决定这批货能不能放行。
Acceptance sampling: whether the confidence interval for the lot mean sits entirely inside spec decides whether the batch ships or gets quarantined.
临床试验:疗效差异的置信区间不含 0,才算统计上证明了药物有效。
Clinical trials: only when the confidence interval for the treatment effect excludes zero do regulators accept it as statistical evidence of efficacy.
一句话In One Line
置信区间把单个点估计("均值是 52%")升级成带不确定性的区间估计("52% ± 3%")。
它的可信度不在某一段区间上,而在方法的长期覆盖率上 —— 这就是为什么不能说"真值有 95% 概率在这段里"。
想要区间更窄(更精确),只有两条路:增大样本量(标准误降)或降低置信水平(z 变小)。
置信区间与下一页的假设检验是同一枚硬币:区间不含某个值,等价于在该值上拒绝原假设。
A confidence interval upgrades a bare point estimate ("the mean is 52%") into an interval estimate that carries its own uncertainty ("52% ± 3%").
Its credibility does not live in any single bar — it lives in the long-run coverage rate of the procedure. That is exactly why "the truth has a 95% probability of being in this interval" is the wrong sentence.
To tighten the interval (more precision) you have only two levers: grow the sample size (shrinks the standard error) or lower the confidence level (smaller z).
Two-sided confidence intervals and the hypothesis test on the next page are two faces of the same coin: if the interval excludes a value, the corresponding null hypothesis is rejected at the matching α level.
常见误用Common Mistakes
说"真值有 95% 概率落在此区间"。真值是常数;95% 描述的是方法的长期覆盖率,不是这一段的概率。
Saying "the true value has a 95% probability of lying in this interval". The truth is a constant. The 95% describes the procedure's long-run coverage rate, not a probability statement about this particular interval.
把更宽的区间当成更好。区间越宽越保险但越没用;要靠加大样本量来收窄,而非抬高置信水平。
Treating a wider interval as a "better" one. Wider is safer but less useful. Tighten the interval by collecting more data, not by cranking the confidence level up to 99%.
区间含 0 仍宣称"有差异"。差值区间跨过 0,等价于不显著,不能下"有效"的结论。
Claiming "there's a difference" while the interval still contains 0. A difference interval that crosses zero is equivalent to a non-significant test — you cannot conclude the effect is real.