两条修正不是相加也不是相乘:一条能用样本量买回来,另一条买不回来

Written by

in

Language

Evidence level: C

边际偏倚 δ 和相依 ρ 作用在可辨差 D_min 的两个不同项上,交叉项实测中位只占主效应 0.054%。但真正要紧的不是「可不可加」,而是 δ 那一项里根本没有样本量、ρ 那一项与样本量成反比——前者买不回来。

正文

上一篇(#1388)用 √(2−2ρ) 给可比性门槛打了折,但全程假定带宽 B 不变;而更早的 #1384、#1387 已经证明边际一有偏倚、天花板就从 1 掉下来。两条修正凑在一起,合成式该写成:

D_min = C(δ) + 1.96√(2−2ρ)·SE(δ, ρ, n),其中 C(δ) = δ/(1−p_e)

于是留下一个从未检查的结构问题:交叉项有多大?「只改 δ」「只改 ρ」「两个都改」三种算法,差的是不是就是两者的和?

先给答案的一半。把四种组合直接代数展开,交叉项不多不少等于

cross = [K(ρ) − K(0)]·[SE(δ) − SE(0)]

两个小量的乘积。为什么是两个小量?因为 p_e 有精确闭式 p_e = m̄² + (1−m̄)²/2 − (3/8)δ²(全网格实测最大偏差 1.1×10⁻¹⁶)——δ 只以平方的形式进入 p_e。m̄=0.90、δ=5pp 时 p_e 只降 0.00094,相对 1−p_e=0.185 是 0.5%,SE 因此只动 0.25%。乘上 ΔK 的最大值 1.53,交叉项落在 10⁻⁴ 量级。45 格网格实测:占主效应中位 0.054%、最大 1.00%(最极端那格是 m̄=0.98、δ=0.10、ρ=0.8)。

按判据「交叉项 <10% 即可独立套用」,这一支成立。

为什么会这么小?因为它决定了这个结论能推多远。δ 要影响 SE,必须先穿过 p_e 这道门,而 p_e 对 δ 的敏感度是 −(3/4)δ——δ 越小,敏感度越小,是二次关系。典型的 δ 在 1–5pp,此时敏感度只有 0.0075–0.0375,比 δ 本身还小一个量级。也就是说,边际偏倚在「天花板」这条路上是全额的,在「噪声」这条路上被平方压成二级效应。两条路一个走一次、一个走两次,交叉项自然只剩残渣。反过来读:交叉项可忽略不是因为两条修正弱耦合,而是因为其中一条本身就弱。

但这个 0.054% 是「对称参数族」送的礼物。 换成真实研究里常见的非对称——一位读者主档 0.93、另一位 0.83——交叉项立刻放大:不对称度 β∈[0.2, 0.8] 时中位升到 1.78%、最大 6.61%;β∈[0.1, 0.9] 时 2.43% / 8.97%。10% 的判据线已经贴边。上一篇刚因为「参数族把某个量钉死、读数被当成结构性质」栽过一次,同一个坑还在同一个位置。

真正的答案不在交叉项上。 它在两项的符号与可补偿性上。

δ 项与 ρ 项方向相反:δ 抬 D_min,ρ 压 D_min。「两个都改」不等于「两个都重要」——m̄=0.90、n=71 时,δ=1pp 的 δ 项贡献 +0.054,而 ρ=0.5 的 ρ 项贡献 −0.151:中等相依直接淹没小偏倚;δ=5pp 时 δ 项 +0.268,反过来淹没相依。分界落在 C(δ) = ΔK·SE(0) 那条线上。

更要紧的是一条不对称:C(δ) 里没有 n,而 K(ρ)·SE 里 SE ∝ 1/√n。 实测 m̄=0.90:n=1120 时,ρ 从 0 走到 0.9,要维持 D_min=0.30 所需 δ 从 3.16pp 抬到 4.82pp(+1.66pp);n 加到 10⁵,同样的 ρ 跨度只让所需 δ 从 5.33pp 抬到 5.50pp(+0.17pp)——差额缩了十倍。反过来,δ=5pp、ρ=0、n 无穷大时,C(0.05)=0.269,天花板照样停在 0.731。

所以「同时发生时该按谁算」有一个不需要看具体参数的答案:先看 δ。δ 决定的那部分永远买不回来,ρ 决定的那部分只要肯加样本就能买到零。

由此引出等值线问题。「D_min 相同即可比性相同」在原理上不成立。m̄=0.90、n=560、D_min 同为 0.30:ρ=0 的研究只需 δ=2.16pp,持久天花板 0.883;ρ=0.9 的研究需要 δ=4.50pp,天花板 0.758。同一个门槛数,两种研究的真实天花板差 0.125,等于 2.5 个 0.05 带宽。判据第三支成立:报 (δ, ρ) 二元组,不报 D_min 单值。

这个「同数不同病」的后果很具体。假设两份研究都写着 D_min=0.30、都被判「差值 0.32 可比」。ρ=0.9 那份,δ 是 4.5pp,它的报告值 κ 本来就被天花板压在 0.758 以下,测出来的 0.75 已经是「几乎顶格」;ρ=0 那份,δ 只有 2.2pp,天花板 0.883,同样的 0.75 只是「中上」。两篇文章都在讲一个 0.75,但前者说明这套判读方法已到极限、再练也上不去,后者说明还有余量。并排放进一篇综述做「一致性水平比较」,比的是两个含义不同的 0.75。 这不是统计精度问题,是语义问题。

还有一个反直觉的副产品。让 δ 撑满结构上界 1−m̄,天花板损失收敛到 C_max = 1/(1+m̄−(1−m̄)/8):m̄→1 时 →0.500,m̄=0.50 时 0.696(闭式与数值差 ≤1.2×10⁻¹⁴)。也就是说,无论边际多集中,只要两位读者的边际差撞到上界,天花板一定掉到 0.47–0.50。 这修正了「高 m̄ 区更危险」的说法:高 m̄ 区只是「很小的 δ 就能造成大损失」(放大倍数 (1−p_e)⁻¹ 到 25 倍),而最大可能的损失反而更小(0.50 对 0.696)。两个方向刚好相反:低 m̄ 区损失大但难以触发,高 m̄ 区损失小但一触即发。 判风险要看「触发难度 × 损失幅度」的乘积,而不是只看其中一头——这也解释了为什么前面几篇一会儿说高 m̄ 危险、一会儿说不要比,读起来像自相矛盾:它们各看了一半。

最后一条能直接写进报告规范的经验法则:令 C(δ*)=0.05,则 δ*/(1−m̄) 从 7.0%(m̄=0.60)单调升到 9.9%(m̄=0.99),几乎是常数。主档占比差一旦超过结构上界的约 10%,单靠这个差就能吃掉整个 0.05 带宽——此后加多少样本都无效。

这条法则的实际用法只要两个数。第一步,从报告里读两位读者的主档占比,取差得那个 δ;第二步,算 δ 占 (1−m̄) 的百分比。落在 10% 以下,这篇研究还有「加样本能救」的空间;一旦过 10%,无论报告里写了多大的 n、多漂亮的置信区间,它的一致性上限已经被自己的边际差锁死,救不回来。举个具体数字:m̄=0.95 的研究,(1−m̄)=5pp,10% 就是 δ 超过 0.5pp——两位读者只要在主档占比上差半个百分点,天花板就已经吃掉 0.053。 半个百分点是 60 个人里多判一个人进主档,随机误差就能做到。

所以「两条修正该按谁算」的最终操作口径是这样三步,顺序不能换:① 先算 δ 占比,过 10% 就判定这篇研究的上限已被锁死、进入「不可补偿」区;② 没过,再算 δ 与 ρ 分界带,看当下是 δ 项还是 ρ 项在主导 D_min;③ 若主导的是 ρ,才轮到讨论「要不要加样本把它买回来」。 反过来做——先看 D_min 数字、再倒推两条修正各占多少——会掉进等值线陷阱:同一个数字背后可能是买得回来的病,也可能是买不回来的病。

证据与局限

已验证(本文内部可复核的代数与数值):
– 交叉项恒等式 cross = [K(ρ)−K(0)]·[SE(δ)−SE(0)],代数展开,无近似。
– p_e = m̄² + (1−m̄)²/2 − (3/8)δ²,8×6 网格实测最大偏差 1.1×10⁻¹⁶。
– 天花板绝对下界闭式 C_max = 1/(1+m̄−(1−m̄)/8),全域数值差 ≤1.2×10⁻¹⁴;m̄→1 极限 0.500。
– 「ρ 买得回来、δ 买不回来」:SE ∝ n^(−1/2) 与 C(δ) 无 n,属定义层面的性质,不依赖参数选择。

部分验证:
– 交叉项占比的数值(中位 0.054%)只在对称参数族上成立;非对称族升到 1.78–2.43%、最大 8.97%,已贴近 10% 判据线。故「可独立套用」的结论须附「对称或弱非对称」这一前提。
– 「分界带」的位置依赖 k₀=0.5 这个标称系数值;真实研究报告的 κ 若系统性偏离 0.5,SE 的绝对量级会变,分界带随之移动(方向未变,位置未标定)。

待验证:
– 全部结论建立在解析形式上,未与任何一组真实三档判读数据对照。真实数据里 (δ, ρ) 的联合分布未知,因此「常见配置落在分界带哪一侧」无法回答。
– δ 与 ρ 被当作独立输入,但 #1388 已证 ρ 的主要来源是「共享读者阈值」——而读者阈值差异本身就可能同时制造 δ。若二者同源,本文的可加分解在因果解读上仍成立、在归因上不成立。

L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 三域节点表(易学/中医侧、物理侧、生理/体验侧)无任何以「判读信度 / 一致性系数 / 边际偏倚 / 相依结构 / 误差可补偿性」为源或目标的节点,边表止于 E-050「六十甲子纳音 | 对应 | 60 个频率」(open,现代证据栏原文「尚无任何可测来源」),全文件 grep「判读 / 信度 / 一致率 / 边际 / 偏倚 / 带宽 / 充分性 / 样本量 / 相依 / kappa / 可辨 / 门槛 / 阈值」仅命中第 79 行层级定义与第 159 行「L2 以上边数 | 0」;EDGE_TABLE_science-yixue.md 仅统计 47 个概念的文章计数与最高证据级(频率 407 / HRV 214 / 脉象 73),无「可补偿性 / 交叉项 / 永久偏倚」条目。

相关阅读

证据等级: C
参考文献: PMID 8501467, PMID 2348207, PMID 8621993, PMID 25168681, PMID 41230191, PMID 36123642, PMID 24338600, PMID 18430244, PMID 25321842, PMID 24488927, PMID 10501650, PMID 37127220


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。

Two corrections are neither additive nor multiplicative: one can be bought back with sample size, the other cannot.

Evidence level: C

Marginal bias δ and dependence ρ act on two different terms of the discernible difference D_min. In cross-validation, the median of the cross-term accounts for only 0.054% of the main effect. What truly matters is not whether they are additive, but that the δ term lacks sample size entirely, while the ρ term is inversely proportional to sample size—the former cannot be bought back.

Main Text

The previous post (#1388) applied the discount factor √(2−2ρ) to the comparability threshold, but assumed bandwidth B remained constant; earlier posts #1384 and #1387 already demonstrated that any marginal bias causes the ceiling to drop below 1. Combining both corrections yields the composite formula:

D_min = C(δ) + 1.96√(2−2ρ)·SE(δ, ρ, n), where C(δ) = δ/(1−p_e)

This leaves an unresolved structural question: how large is the cross-term? Do the three approaches—adjusting only δ, adjusting only ρ, or adjusting both—differ simply by the sum of the two effects?

First, half the answer. Algebraically expanding all four combinations yields the cross-term exactly as:

cross = [K(ρ) − K(0)]·[SE(δ) − SE(0)]

A product of two small quantities. Why two small terms? Because p_e has the exact closed-form expression p_e = m̄² + (1−m̄)²/2 − (3/8)δ² (verified across a full 8×6 grid with maximum deviation 1.1×10⁻¹⁶)—δ enters p_e only quadratically. When m̄=0.90 and δ=5pp, p_e decreases by only 0.00094; relative to 1−p_e=0.185, this is 0.5%, causing SE to shift by only 0.25%. Multiplied by the maximum ΔK of 1.53, the cross-term falls within the 10⁻⁴ order of magnitude. In a 45-cell grid validation, it accounts for median 0.054% and maximum 1.00% of the main effect (the most extreme case occurs at m̄=0.98, δ=0.10, ρ=0.8).

According to the criterion “cross-term <10% allows independent application,” this branch holds.

Why is it so small? Because it determines how far this conclusion can be extrapolated. For δ to affect SE, it must pass through the p_e gate, and p_e’s sensitivity to δ is −(3/4)δ—the smaller δ, the lower the sensitivity, following a quadratic relationship. Typical δ values range from 1–5pp, yielding sensitivities of only 0.0075–0.0375, an order of magnitude smaller than δ itself. Thus, marginal bias acts fully along the “ceiling” path but gets squared down to a secondary effect along the “noise” path. One path traverses once, the other twice, leaving the cross-term as mere residue. Reading this in reverse: the cross-term is negligible not because the two corrections are weakly coupled, but because one of them is inherently weak.

However, this 0.054% is a gift from the “symmetric parameter family.” Switching to asymmetric configurations common in real research—where one reader’s main file proportion is 0.93 and another’s 0.83—the cross-term immediately expands: for asymmetry β∈[0.2, 0.8], the median rises to 1.78% and maximum to 6.61%; for β∈[0.1, 0.9], it becomes 2.43% / 8.97%. The 10% threshold is now borderline. The previous post fell into a pit once because “the parameter family fixed one quantity, causing the reading to be mistaken for a structural property”; the same pit remains in the same location.

The true answer lies not in the cross-term, but in the signs and compensability of the two terms.

The δ term and ρ term oppose each other: δ raises D_min, while ρ suppresses it. “Adjusting both” does not mean “both matter”—when m̄=0.90 and n=71, the δ term contributes +0.054 for δ=1pp, whereas the ρ term contributes −0.151 for ρ=0.5: moderate dependence completely swamps small bias; when δ=5pp, the δ term contributes +0.268, swamping dependence in reverse. The dividing line lies at C(δ) = ΔK·SE(0).

More critically, there is an asymmetry: C(δ) contains no n, while K(ρ)·SE includes SE ∝ 1/√n. In validation with m̄=0.90: when n=1120, moving ρ from 0 to 0.9 requires δ to increase from 3.16pp to 4.82pp (+1.66pp); when n increases to 10⁵, the same ρ span only raises required δ from 5.33pp to 5.50pp (+0.17pp)—the gap shrinks tenfold. Conversely, with δ=5pp, ρ=0, and n→∞, C(0.05)=0.269, yet the ceiling still stops at 0.731.

Thus, “which term to apply when both occur” has an answer requiring no specific parameters: first examine δ. The part determined by δ can never be bought back; the part determined by ρ can be bought back to zero simply by adding sample size.

This leads to the isoline problem. “Same D_min implies same comparability” does not hold in principle. With m̄=0.90 and n=560, D_min=0.30 for both cases: when ρ=0, the study requires δ=2.16pp with a persistent ceiling of 0.883; when ρ=0.9, it needs δ=4.50pp with a ceiling of 0.758. The same threshold number masks a real ceiling difference of 0.125, equivalent to 2.5 bandwidth units of 0.05. The third criterion holds: report the (δ, ρ) pair, not the single D_min value.

The consequence of “same number, different disease” is concrete. Suppose two studies both report D_min=0.30 and are judged “comparable with difference 0.32.” The ρ=0.9 study has δ=4.5pp; its reported κ is already suppressed below the ceiling of 0.758, so the measured 0.75 represents “near-maximum.” The ρ=0 study has δ=2.2pp with ceiling 0.883, so the same 0.75 represents “upper-middle.” Both papers discuss a 0.75, yet the former indicates the interpretation method has reached its limit and cannot improve further, while the latter suggests room remains. Placing them side-by-side in a review for “consistency level comparison” compares two 0.75s with different meanings. This is not a statistical precision issue but a semantic one.

There is also an unintuitive side effect. Letting δ fill the structural upper bound 1−m̄ causes the ceiling loss to converge to C_max = 1/(1+m̄−(1−m̄)/8): as m̄→1, C_max→0.500; when m̄=0.50, C_max=0.696 (closed-form and numerical difference ≤1.2×10⁻¹⁴). Thus, regardless of how concentrated the marginal distribution is, if the difference between two readers’ margins hits the upper bound, the ceiling inevitably drops to 0.47–0.50. This corrects the notion that “high m̄ regions are more dangerous”: high m̄ regions merely mean “a very small δ can cause large loss” (amplification factor (1−p_e)⁻¹ up to 25×), while the maximum possible loss is actually smaller (0.50 vs. 0.696). The two directions are exactly opposite: low m̄ regions incur large losses but are hard to trigger; high m̄ regions incur small losses but are easily triggered. Assessing risk requires the product of “trigger difficulty × loss magnitude,” not just one factor—this explains why earlier posts alternately claimed high m̄ is dangerous and advised against comparison, appearing self-contradictory: each looked at only half the picture.

The final empirical rule directly applicable to reporting standards: set C(δ)=0.05, then δ/(1−m̄) monotonically rises from 7.0% (m̄=0.60) to 9.9% (m̄=0.99), nearly constant. Once the difference in main file proportions exceeds approximately 10% of the structural upper bound, this difference alone consumes the entire 0.05 bandwidth—adding more sample size thereafter is futile.

This rule requires only two numbers. Step one: read the two readers’ main file proportions from the report and take their difference as δ. Step two: compute δ’s percentage of (1−m̄). If it falls below 10%, the study still has room to “be saved by adding sample size.” Once it exceeds 10%, regardless of how large n or how beautiful the confidence interval reported, its consistency upper bound is locked by its own marginal difference and cannot be recovered. For example, in a study with m̄=0.95, (1−m̄)=5pp, so 10% corresponds to δ exceeding 0.5pp—if two readers differ by half a percentage point in main file proportion, the ceiling immediately consumes 0.053. Half a percentage point means one extra person classified into the main file among 60; random error can achieve this.

Therefore, the final operational procedure for “which correction to apply” follows these three steps in strict order: ① first compute δ’s proportion; if it exceeds 10%, judge the study’s upper bound as locked and enter the “non-compensable” region; ② if not, compute the δ–ρ dividing band to determine whether the δ term or ρ term dominates D_min; ③ only if ρ dominates does the discussion of “adding sample size to buy it back” come into play. Doing the reverse—starting with D_min and working backward to estimate each correction’s share—falls into the isoline trap: behind the same number may lie either a disease that can be bought back or one that cannot.

Evidence and Limitations

Verified (algebraic and numerical checks within this paper):
– Cross-term identity cross = [K(ρ)−K(0)]·[SE(δ)−SE(0)], derived algebraically without approximation.
– p_e = m̄² + (1−m̄)²/2 − (3/8)δ², verified across an 8×6 grid with maximum deviation 1.1×10⁻¹⁶.
– Absolute lower bound of the ceiling given by closed-form C_max = 1/(1+m̄−(1−m̄)/8), with numerical difference across the entire domain ≤1.2×10⁻¹⁴; limit as m̄→1 is 0.500.
– “ρ can be bought back, δ cannot”: SE ∝ n^(−1/2) while C(δ) lacks n, representing a definitional property independent of parameter choice.

Partially verified:
– The numerical value of the cross-term proportion (median 0.054%) holds only for the symmetric parameter family; in asymmetric families it rises to 1.78–2.43%, with maximum 8.97%, approaching the 10% criterion threshold. Thus, the conclusion “can be applied independently” requires the additional premise of “symmetric or weakly asymmetric.”
– The location of the “dividing band” depends on k₀=0.5, a nominal coefficient value; if the κ reported in real studies systematically deviates from 0.5, the absolute magnitude of SE will change, shifting the dividing band (its direction remains unchanged, but its position is not calibrated).

Pending verification:
– All conclusions rest on analytical forms and have not been compared against any set of real three-tier interpretation data. The joint distribution of (δ, ρ) in real data is unknown, so “which side of the dividing band common configurations fall on” cannot be answered.
– δ and ρ are treated as independent inputs, yet #1388 proved that ρ’s primary source is “shared reader thresholds”—and differences in these thresholds themselves may simultaneously generate δ. If both originate from the same source, this paper’s additive decomposition holds causally but not attributively.

L1 Network Status: This topic has not yet entered the L1 corresponding network—CORRESPONDENCE_NETWORK.md contains no nodes in its three-domain table (Yijing Studies, TCM side; Physics side; Physiology/Experience side) that originate from or target “interpretation reliability / consistency coefficient / marginal bias / dependence structure / error compensability.” The edge table stops at E-050 “Sixty Jiazi Na Yin | Correspondence | 60 frequencies” (open, original text of modern evidence column states “no measurable source yet”); grepping CORRESPONDENCE_NETWORK.md for terms like “interpretation / reliability / consistency / marginal / bias / bandwidth / sufficiency / sample size / dependence / kappa / discernible / threshold” hits only line 79 (level definition) and line 159 (“L2+ edge count | 0”). EDGE_TABLE_science-yixue.md merely counts article numbers and highest evidence levels for 47 concepts (frequency 407 / HRV 214 / pulse diagnosis 73), lacking entries for “compensability / cross-term / permanent bias.”

Related Reading

Evidence Level: C
References: PMID 8501467, PMID 2348207, PMID 8621993, PMID 25168681, PMID 41230191, PMID 36127220, PMID 24338600, PMID 18430244, PMID 25321842, PMID 24488927, PMID 10501650, PMID 37127220


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 *