那个比值的误差根本没被放大:误差翻倍的是你以为自己在比较的东西,不是比值本身

Written by

in

Language

Evidence level: C

我们把怕的东西搞错了整整一轮。比值 Δ/D_min 的抽样误差不但没被放大,反而被系统性压小;真正会翻转判决的,是分母里那个你从来没报过的固定项 B。

正文

一、先说结论:怕反了

上一轮我给出的操作规则是「先算比值 R = Δ 除以 D_min,落在 0.7 到 2.5 之间才谈打折」。我还留了一条待验证:D_min 里的 δ 和 ρ 都是估的,两个估计量一除,误差会怎么走。

现在算完了,结论跟预想的相反。比值不但没有把误差翻倍,它的相对误差永远比分子自己更小,而且有一个干净的上界。

设 R = Δ /(B + 1.96·u·SE),其中 u 是根号下 2 减 2ρ,B 是边际带宽这个固定项。做一次 delta 方法展开,R 的相对标准误恰好是

相对 SE = 1 / (1.96 + B/(u·SE))

这是本文最要紧的一条。分母里那个 1.96 是拍下来的 95% 系数,B/(u·SE) 恒为正。所以相对 SE 永远小于 1/1.96 = 0.510。比值做不出一条 50% 以上的相对误差带。所谓「一除就把噪声翻倍」在代数上不成立。

n 等于 60、m 等于 0.85 时实测 0.482,n 等于 600 时降到 0.431,n 等于 2600 时 0.369。从 60 例加到 2600 例,四十三倍的样本,相对精度只从 48% 改善到 37%。

二、为什么压不下去:结构项在当家

看那个式子的分母。B/(u·SE) 里,SE 随根号 n 缩小,B 是常数。样本量一大,B 项就接管了分母。

m 等于 0.85 时 B 是 0.0166,n 等于 60 时 B/(u·SE) 只有 0.114,B 在 D_min 里占 5.5%;n 等于 2600 时占 27.7%;n 等于 50000 时占 62.7%。B 从配角变成主角。

这解释了一件反直觉的事:比值的精度不是「样本够就准」,而是被一个跟样本量无关的结构常数锁住。要让半宽降到 0.45,ρ 等于 0.5 时需要约两万六千例;降到 0.30 要九万六千例。而整个判读文献的样本量中位数只有 71 —— 那是从 2022 年 82 项一致性研究的综述里读出来的数。

三、判决其实只有一个旋钮,它叫 λ

把 R 的式子重排一下,会浮出一个恒等式:

Δ / D_min = z / λ,其中 z = Δ/(u·SE),λ = 1.96 + B/(u·SE)

我用四组 n 乘三组 ρ 共十二个点验算过,R 与 z 除以 λ 的差最大 8.9×10⁻¹⁶,纯浮点。

这个恒等式的后果比它的形式重要得多。它说明「R 在 0.7 到 2.5 之间」这件事完全等价于「z 在 0.7λ 到 2.5λ 之间」。 而 λ 只由 (B, SE, ρ) 决定,与你要判的那个差 Δ 一点关系都没有。

于是那条判决带,在显著性尺度上是一条会滑动的窗口。n 等于 60 时窗口是 z 落在 1.45 到 5.19;n 等于 600 时 1.62 到 5.80;n 等于 10000 时 2.40 到 8.58。同一个真差,样本量一大,就不再落进「可打折」这一档。

具体点:m 等于 0.85、ρ 等于 0.5、真差 0.15。n 等于 60 判「太细、别打」,n 等于 125 翻成「可打折」,一路到 n 等于 2589 还在这一档,n 等于 5000 再翻成「不必打折」。三段判决、同一个真差、只换样本量。

两个研究各做各的、真差一模一样,判决不一致的概率:n 取 60 对 600 是 64.9%,60 对 2000 是 75.7%,200 对 10000 是 100%。

四、还有一条更硬的天花板

R 的上界是 Δ 除以 B。因为 1.96·u·SE 恒为正,R 永远小于 Δ/B。

m 等于 0.85 时 B 是 0.0166,所以真差 0.02 的 R 上限只有 1.205、真差 0.03 上限 1.807、真差 0.04 上限 2.410 —— 这三种情况无论做多少人、ρ 多高,永远进不了「不必打折」那一档。规则给了一档,数据里根本没有那一档。

反过来说,真差 0.05 就能到 2.5,0.15 能到 9.0。2.5 这条上界对小于 0.0415 的真差是一句空话。

这里要补一句方法学上的旧账:比值型统计量的区间从来是个坑,不是本篇才发现的。Fieller 定理的名字就来自这件事——它的区间有三种形态,闭区间、闭区间的补集、以及整条实轴,具体是哪一种取决于分母能不能显著拒绝零。1999 年一项一万次模拟的对照研究直接说,当效果差接近零时,Fieller 与三种 bootstrap 全都给不出可用结果。工具变量文献里的说法更直白:比值估计量的均值不存在有限值,因为分母有非零概率贴着零,样本一多均值就发散。

本篇的分母好在不接近零——B 与 1.96·u·SE 都是正数之和,D_min 恒有下界 B。这正是它比典型比值估计安全的地方。 反过来说,一旦有人把分母换成不带宽带底的项目(比如只用 SE),那套无界区间的麻烦立刻回来。

五、ρ 的误差:比想象的小得多

原假设担心 δ 与 ρ 正相关会「系统性互相放大」。实测:n 等于 60 时相关系数从 负 0.5 走到 正 0.5,半宽从 0.930 变成 1.029,只差 10.6%;n 等于 600 时差 8.9%。

原因在方差分解。R 等于 1 处,ρ 项对总方差的贡献 n 等于 60 时是 3.6%、n 等于 600 时 0.36%、n 等于 2600 时 0.08%。ρ 估不准这件事,在比值这个用途上几乎可以忽略。 跟上一轮 #1410 的结论对上了:ρ 的抽样误差在那条判据里也从来不是瓶颈。

六、真正该改的三处

第一,判决必须带 λ,不能只报 R。λ 一个数就把 (B, SE, ρ) 三样东西压进去了,而它是唯一决定判决带位置的东西。报告里不写 λ,等于不写这条规则的适用条件。

第二,0.7 与 2.5 这两条线是按「R 的尺度」画的,而 R 的尺度本身随 n 滑动。正确写法是把它翻译成 z 尺度:z 要落在 0.7λ 到 2.5λ 之间。原规则在 n 从 60 变到 10000 的过程里,实际要求的 z 下界从 1.45 涨到 2.40 —— 严了六成半,而没人知道这件事。

第三,n 小的时候不是「区间宽一点」的问题,是「判决本身不存在」的问题。n 等于 60 时半宽 0.945、区间总宽 1.89,比整条判决带(宽 1.8)还宽 —— 区间把三档全盖住了。这种情况下给一个「可打折」或「太细」的结论,是在报噪声。

顺带说一句 n 该怎么规划。既然相对精度等于 1.96 除以 z,那「区间总宽不超过判决带宽」等价于半宽不超过 0.9,m 等于 0.85 时需约 219 例。但那只保证区间窄于判决带,不保证能分清三档。 要半宽压到 0.45,ρ 等于 0.5 时需两万六千例。这两个数差两个量级,文献里的实际做法是把它当第一个数在用。

证据与局限

已验证:① 相对 SE 等于 1 除以(1.96 加 B/(u·SE))这条式子在四组 n 乘三组 ρ 上逐点吻合,上界 0.510 是代数结论;② R 恒等于 z 除以 λ 十二个格点最大偏差 8.9×10⁻¹⁶;③ R 的 95% 区间上限是 Δ 除以 B,与 n、ρ 无关;④ ρ 项对总方差的贡献在各档均低于 6%;⑤ Δ 与 ρ 的相关性在正负 0.5 范围内对半宽的影响不超过 11%(二十万次蒙特卡洛)。

部分验证:① SE 的闭式沿用前作 Delta 方法口径,边际相等假设未在本篇重检;② B 取自 m 等于 0.60、0.85、0.90、0.95 四点的插值,中间值有插值误差;③ 判决带 0.7 与 2.5 本身是上一轮拍的,本篇只算它的误差、不辩护它的取值。

待验证:① ρ̂ 的抽样标准差用了 (1 减 ρ 平方) 除以根号 n 的近似,小样本下未必准;② 本篇假定 Δ 与 ρ 独立可分解,若两者真同源(即 #1436 那个问题),方差结构要重算;③ 比值形式在分母可能接近零时 Fieller 型区间会变成无界集,本文的线性展开在 B/(u·SE) 很小时可能低估尾部;④ 三档之外的有序数据未涉及。

L1 网络状态:L1 对应网络中未找到与本主题(比值估计的误差传播与判决带稳定性)对应的边。两个源文件内检索「比值 / 置信区间 / 误差 / 阈值概率 / 决策曲线 / 净收益 / 判读 / 一致性」均无命中。唯一结构相关的是 §3.8 至 §3.11 反复出现的「E-060 标签前提」,状态为 conflict(未裁决):「中医医师间诊断一致性在开放人群中低(low),作者称为令人担忧(PMID 30383398);同系列第二篇指出偶然校正一致性统计常被误用、建议 Gwet AC2(PMID 31730403、31634000、31730404);但训练可改善(PMID 19538100)、量表化可行」。冲突双方在本篇的意义是:这条测量链的输入端本身未被裁决,本篇讨论的比值再准,也只是在一个尚未定论的量表上做精度分析。该冲突不构成对本篇数学结论的否证,但限定了它的适用范围。

相关阅读

证据等级: C
参考文献: PMID 32865227, PMID 10874387, PMID 26282889, PMID 11512137, PMID 22560852, PMID 11890316, PMID 36123642


免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

The error of the ratio itself was never amplified: what doubled in error was what you thought you were comparing, not the ratio itself.

Evidence level: C

We got the thing we feared wrong by exactly one full cycle. The sampling error of the ratio Δ/D_min was not only not amplified; it was systematically suppressed. What actually flips the verdict is the fixed term B in the denominator, which you never reported.

Body

I. Conclusion First: The Fear Was Reversed

In the previous round, my operational rule was “first calculate the ratio R = Δ divided by D_min, and only discuss discounting if it falls between 0.7 and 2.5.” I also left one item to verify: since δ and ρ inside D_min are both estimates, dividing two estimators will affect the error in what way.

Now that the calculation is done, the conclusion is opposite to expectation. The ratio did not double the error; its relative error is always smaller than that of the numerator itself, and it has a clean upper bound.

Let R = Δ /(B + 1.96·u·SE), where u is the square root of 2 minus 2ρ, B is the fixed term for marginal bandwidth. Expanding via the delta method, R’s relative standard error is exactly

Relative SE = 1 / (1.96 + B/(u·SE))

This is the most critical equation in this paper. The 1.96 in the denominator is the captured 95% coefficient; B/(u·SE) is always positive. Thus, the relative SE is always less than 1/1.96 = 0.510. The ratio cannot produce a relative error band exceeding 50%. The claim that “dividing doubles the noise” fails algebraically.

When n equals 60 and m equals 0.85, the measured value is 0.482; when n equals 600 it drops to 0.431; when n equals 2600 it is 0.369. Increasing the sample from 60 cases to 2,600 cases—a forty-three-fold increase—improves relative precision only from 48% to 37%.

II. Why It Cannot Be Suppressed: Structural Terms Dominate

Look at the denominator of that equation. In B/(u·SE), SE shrinks with the square root of n, while B is constant. As sample size grows, the B term takes over the denominator.

When m equals 0.85, B is 0.0166; when n equals 60, B/(u·SE) is only 0.114, meaning B accounts for 5.5% of D_min; when n equals 2600 it accounts for 27.7%; when n equals 50,000 it accounts for 62.7%. B shifts from supporting role to lead.

This explains a counterintuitive fact: the precision of the ratio is not “accurate when sample size is sufficient,” but locked by a structural constant independent of sample size. To bring the half-width down to 0.45, ρ must equal 0.5, requiring approximately 26,000 cases; to reach 0.30 it requires 96,000 cases. Yet the median sample size across all judgment literature is only 71 — a number read from a 2022 review of 82 consistent studies.

III. The Verdict Actually Has Only One Knob, Called λ

Rearranging the equation for R reveals an identity:

Δ / D_min = z / λ, where z = Δ/(u·SE), λ = 1.96 + B/(u·SE)

I verified this across four groups of n times three groups of ρ, totaling twelve points; the maximum difference between R and z divided by λ is 8.9×10⁻¹⁶, pure floating-point noise.

The consequence of this identity is far more important than its form. It shows that “R falls between 0.7 and 2.5” is completely equivalent to “z falls between 0.7λ and 2.5λ.” Since λ depends only on (B, SE, ρ), it has nothing to do with the difference Δ you are judging.

Thus, that judgment band is a sliding window on the significance scale. When n equals 60, the window is z falling between 1.45 and 5.19; when n equals 600, between 1.62 and 5.80; when n equals 10,000, between 2.40 and 8.58. The same true difference, with larger sample size, no longer falls into the “discountable” category.

Specifically: m equals 0.85, ρ equals 0.5, true difference 0.15. When n equals 60 it is judged “too fine, do not discount”; when n equals 125 it flips to “discountable”; continuing up to n equals 2,589 it remains in this category; when n equals 5,000 it flips again to “no need to discount.” Three verdicts, the same true difference, changing only sample size.

If two studies proceed independently with identical true differences, the probability of inconsistent verdicts: n taking 60 versus 600 is 64.9%, 60 versus 2,000 is 75.7%, 200 versus 10,000 is 100%.

IV. Another Harder Ceiling

R’s upper bound is Δ divided by B. Since 1.96·u·SE is always positive, R is always less than Δ/B.

When m equals 0.85, B is 0.0166, so for true differences of 0.02, 0.03, and 0.04, the R upper limits are only 1.205, 1.807, and 2.410 respectively — in no case do these enter the “no need to discount” category, regardless of how many people are tested or how high ρ is. The rule provides one tier, but the data lacks that tier entirely.

Conversely, a true difference of 0.05 can reach 2.5, and 0.15 can reach 9.0. The upper bound of 2.5 is an empty statement for true differences less than 0.0415.

Here, let me add a methodological debt: ratio-type statistics intervals have always been a pitfall, not something discovered in this paper. The name Fieller’s theorem comes precisely from this issue — its interval has three forms: closed interval, complement of a closed interval, and the entire real axis, depending on whether the denominator can significantly reject zero. A 1999 study with 10,000 simulations directly stated that when the effect difference approaches zero, neither Fieller nor three bootstrap methods yield usable results. The statement in instrument variable literature is even clearer: the mean of the ratio estimator does not have a finite value, because the denominator has non-zero probability of touching zero, causing the mean to diverge as sample size increases.

The denominator in this paper is fortunate in not approaching zero — B and 1.96·u·SE are sums of positive numbers, so D_min always has a lower bound B. This is exactly where it is safer than typical ratio estimators. Conversely, once someone replaces the denominator with items without bandwidth bottoms (e.g., using only SE), the trouble of unbounded intervals immediately returns.

V. ρ’s Error: Much Smaller Than Imagined

The original hypothesis worried that positive correlation between δ and ρ would “systematically amplify each other.” Measured results: when n equals 60, the correlation coefficient moves from negative 0.5 to positive 0.5, half-width shifts from 0.930 to 1.029, a difference of only 10.6%; when n equals 600 the difference is 8.9%.

The reason lies in variance decomposition. At R equals 1, the ρ term’s contribution to total variance is 3.6% when n equals 60, 0.36% when n equals 600, and 0.08% when n equals 2,600. The uncertainty in estimating ρ is almost negligible for this ratio application. This matches the conclusion from previous #1410: ρ’s sampling error was never the bottleneck in that judgment criterion.

VI. Three Places That Truly Need Revision

First, the verdict must include λ, not just report R. A single number λ compresses (B, SE, ρ), and it is the only factor determining the position of the judgment band. Omitting λ from the report equals omitting the applicability conditions of this rule.

Second, the lines at 0.7 and 2.5 are drawn according to “R’s scale,” but R’s scale itself slides with n. The correct way is to translate it into z scale: z must fall between 0.7λ and 2.5λ. The original rule, as n increases from 60 to 10,000, actually requires the z lower bound to rise from 1.45 to 2.40 — a tightening of sixty-five percent, yet no one knew this.

Third, when n is small, the problem is not “the interval is wider,” but “the verdict itself does not exist.” When n equals 60, the half-width is 0.945 and the total interval width is 1.89, which is wider than the entire judgment band (width 1.8) — the interval covers all three tiers. Under such circumstances, giving a verdict of “discountable” or “too fine” is merely reporting noise.

Incidentally, let me say how n should be planned. Since relative precision equals 1.96 divided by z, “total interval width not exceeding judgment band width” is equivalent to “half-width not exceeding 0.9,” requiring approximately 219 cases when m equals 0.85. But this only may support the interval is narrower than the judgment band, not that it can distinguish the three tiers. To press the half-width down to 0.45, ρ must equal 0.5, requiring 26,000 cases. These two numbers differ by two orders of magnitude; actual practice in the literature treats it as the first number.

Evidence and Limitations

Verified: ① The equation “Relative SE equals 1 divided by (1.96 plus B/(u·SE))” matches point-by-point across four groups of n times three groups of ρ; the upper bound 0.510 is an algebraic conclusion; ② R always equals z divided by λ, with maximum deviation of 8.9×10⁻¹⁶ across twelve grid points; ③ R’s 95% interval upper bound is Δ divided by B, independent of n and ρ; ④ the ρ term’s contribution to total variance is below 6% in all tiers; ⑤ the influence of correlation between Δ and ρ on half-width, within ±0.5 range, does not exceed 11% (200,000 Monte Carlo simulations).

Partially Verified: ① The closed-form for SE follows the Delta method convention of prior work; the assumption of equal margins was not rechecked in this paper; ② B is interpolated from four points with m equal to 0.60, 0.85, 0.90, and 0.95, introducing interpolation error in intermediate values; ③ the judgment bands of 0.7 and 2.5 themselves were set in the previous round; this paper only calculates their error, not defends their values.

To Be Verified: ① the sampling standard deviation of ρ̂ uses the approximation (1 minus ρ squared) divided by square root of n, which may not be accurate for small samples; ② this paper assumes Δ and ρ are independent and decomposable; if they truly share the same source (i.e., the issue in #1436), the variance structure must be recalculated; ③ when the denominator in ratio form approaches zero, Fieller-type intervals become unbounded sets; this paper’s linear expansion may underestimate tails when B/(u·SE) is very small; ④ ordered data outside the three tiers were not involved.

L1 Network Status: L1 corresponds to edges not found in the network matching this topic (error propagation of ratio estimation and stability of judgment bands). Searching “ratio / confidence interval / error / threshold probability / decision curve / net gain / judgment / consistency” in both source files yielded no hits. The only structurally related element is the recurring “E-060 label premise” in §3.8 to §3.11, status conflict (unresolved): “Diagnostic consistency among TCM physicians is low in open populations (low), authors call this worrying (PMID 30383398); the second paper in the same series notes that chance-corrected agreement statistics are often misused, recommending Gwet AC2 (PMID 31730403, 31634000, 31730404); but training can improve (PMID 19538100), and scaling is feasible.” The significance of this conflict for this paper is: the input end of this measurement chain itself remains undecided; the ratio precision discussed here is merely an accuracy analysis on a scale not yet finalized. This conflict does not refute the mathematical conclusions of this paper, but limits its scope of application.

Related Reading

Evidence Level: C
References: PMID 32865227, PMID 10874387, PMID 26282889, PMID 11512137, PMID 22560852, PMID 11890316, PMID 36123642


Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *