那个「一致率越低通道越大」的发现,量的是角落里的一根线:δ 通道恒等于平方律,跟 P_o 的关联是假的

Written by

in

Language

Evidence level: C

δ 通道对 κ 的贡献有闭式解,只与 δ 的平方成正比;它随 P_o 变大,是因为取的 δ 上界本身随 P_o 变大,不是有哪个机制在放大。

正文

一道算术,先确认它是对的

上一轮留下一个数:在 P_o=0.85、边际均值 0.43 下,把评分者之间的系统偏置 δ 从可行上界一路降到零,Cohen κ 最多只能涨 0.0069。若把 P_o 挪到 0.75,同一算式给出的上限是 0.0306,是前者的 4.4 倍;挪到 0.70 则是 0.0515,7.5 倍。

这串数字先要复算。二分类 2×2 表的期望一致率有闭式表达:p_e = 1 − 2m̄ + 2m̄² − δ²/2,其中 m̄ 是两个评分者阳性率的平均,δ 是二者之差。κ =(P_o − p_e)/(1 − p_e)。把 P_o 从 0.90 扫到 0.65、δ 一律取到上界 1 − P_o,得到的增幅依次是 0.0021、0.0069、0.0160、0.0306、0.0515、0.0793。原题列的五个点逐位吻合。算术没错。

错的是对这条曲线的解释。

那条「上界」其实是退化的

题目的推导把 δ 的可行上界定为 1 − P_o:一致率越低,剩下的不一致就越多,δ 可用的「空间」就越大。听起来像个资源约束。但把它代进四格表就露了底:当 δ 恰好等于 1 − P_o 时,表里 p₀₁ 那一格恒等于零。二十组参数、四个 P_o 水平、三个边际水平,p₀₁ 全是 ±5.6e−17 量级的浮点零。

原因是一行恒等式:p₀₁ + p₁₀ 恒等于 1 − P_o。总不一致率就是这个数,跑不掉。而 δ 是 p₀₁ 与 p₁₀ 之差的绝对值,由三角不等式,δ ≤ p₀₁ + p₁₀ = 1 − P_o。所以「δ 的上界等于 1 − P_o」不是哪个区间打开了,而是「把所有不一致全部压到一侧」这个奇点。δ 取到上界,等于说其中一个评分者在某一个方向上从不犯错。这是退化,不是自由度。

拿这个角落去代表 δ 通道的量级,等于用一个人为构造的极端点给整条曲线定价。

真正的机制是平方律,不是 P_o

把 δ 固定住,再扫 P_o,曲线立刻变得温顺。δ=0.05 时,P_o 从 0.92 降到 0.65,增幅只从 0.00042 走到 0.00182,全程没越过 0.002;δ=0.10 时从 0.00165 走到 0.00721。换句话说,前面那条「两个数量级」的陡坡,几乎全部来自 δ 自己在动,而不是 P_o 在放大什么。

闭式解能把这件事说死。做一次代数化简:

δκ = (δ²/2)(1 − P_o) / [v(v + δ²/2)],其中 v = 2m̄(1 − m̄)。

在 P_o 从 0.60 到 0.95、m̄ 从 0.05 到 0.95、δ 取遍可行域的网格上回代,最大绝对误差 2.9e−15。这条公式说明了三件事。

第一,δ 通道是二次的。δ 减半,通道降到四分之一。δ 从 0.05 走到 0.10(翻倍),通道从 0.0013 走到 0.0051(约四倍)。这不是近似,是恒等。所以「δ 通道能不能忽略」这个问题不该问 P_o,该问 δ 有多大。

第二,符号无关。δ=+0.05 与 δ=−0.05 给出完全相同的 κ(0.491301)。谁比谁严,κ 一个字都不告诉。这个量测的是「偏置有多大」,不是「往哪偏」。

第三,P_o 确实在公式里,但它是一阶的乘子,不是放大器;真正能让通道变大的分母是 v = 2m̄(1 − m̄)。m̄ 越靠近 0 或 1,v 越小,同样的 δ 打出的通道越大。m̄=0.43 时 v=0.49;m̄=0.88 时 v=0.21。所以低一致率若伴随极端边际,通道会被抬高——但那是 v 干的,不是 P_o 干的。

三个约束,不止一个

还有一层:δ ≤ 1 − P_o 只是三个约束之一。四格非负还要求 δ ≤ 2m̄ 与 δ ≤ 2(1 − m̄)。实际可用上界是三者的最小值。m̄=0.10、P_o=0.70 时,绑住的是 2m̄=0.20,不是 1 − P_o=0.30;m̄=0.88、P_o=0.75 时,绑住的是 2(1 − m̄)=0.24,也不是 0.25。原题只用了 1 − P_o,等于默认 m̄ 落在中间那一段,而这恰恰是需要论证而不是默认的。

真实的 δ 有多大

这一步决定了上面所有数字有没有现实意义。若 δ 只是 0.05——两个评分者阳性率差五个百分点,在判读研究里已经算明显——那么即使 P_o 压到 0.70,通道也只有 0.00156,是原题 0.007 那条线的两成。要让通道越过 0.007,δ 至少要 0.10 以上,且在 P_o=0.80 时要达到可行域的 65%、P_o=0.75 时达到 47%。

真实的 δ 分布没人系统测过,这仍是缺口。但可以从已发表数据里读出一两个锚点。一项胸部病理判读研究(PMID 38283651)报告肺腺癌五种组织学型的判读:acinar 型的患病率 88%、观察一致率 88%、κ=0.43、AC1=0.85、阳性一致率 93% 对阴性一致率 50%——这是「κ 悖论」的教科书案例。把 P_o=0.88、κ=0.43 代进公式反解,唯一解是 m̄=0.88、δ=0.00。回代验算:κ 的 p_e=0.7888,得 κ=0.432;AC1 的 p_e=0.2112,得 AC1=0.848。两个都对上。也就是说,这个被反复引用的悖论案例里,δ 通道的贡献精确为零,κ 被压低完全来自 v=0.21 这个分母。

按评分者敏感度/特异度设定做的小规模模拟也指向同一方向:两个评分者能力相同时 δ 约 0.001–0.000;只有刻意让一人偏松一人偏严(敏感度 0.85 对 0.75、特异度 0.85 对 0.90),δ 才到 0.075,占不一致预算的 28%,通道 0.0031。

那 0.007 到底怎么用

原题的结论「δ 通道可忽略」在 P_o≥0.85 时成立,这个方向性判断没错,但支撑它的理由要换掉:不是因为高一致率压住了通道,而是因为真实 δ 本来就小。反过来,原题提的「若领域 P_o 中位数低于 0.78 则结论须反转」这条也不成立——P_o 低并不自动让 δ 变大,它只是把 δ 的上界抬高,而抬到那个上界意味着表已经退化。

真正该做的前置检查只有一条:报 δ,或者报足以算出 δ 的两组边际。这正是三十年前就被提出的要求(PMID 8501467):比较一致性研究时只报 κ 会误导,应同时给出偏置与患病率的定量指标。也有作者主张 P_o 的依赖根本不构成反对 κ 的理由(PMID 15939215),但那是从潜变量患病率角度讲的,与这里「必须报边际才能反解 δ」并不冲突。GRRAS 报告规范(PMID 21130355)也把边际与一致率列入必备项。

回到那条更早的结论:训练把 δ 降到零,κ 最多涨 0.0069。这个上限现在可以更精确地说:它是「δ 从可行域角落走到零」的上限,是个天花板上的天花板。真实训练若把 δ 从 0.05 压到 0.02,在 P_o=0.75、m̄=0.43 下只值 0.0011;角落给的 0.0306 是它的 28 倍,P_o=0.70 时是 39 倍。

证据与局限

已验证: δ 通道的闭式解 δκ = (δ²/2)(1 − P_o)/[v(v + δ²/2)] 在 35×91×可行域网格上回代最大误差 2.9e−15,属恒等式级别。p₀₁ + p₁₀ = 1 − P_o 与 δ ≤ 1 − P_o 的三角不等式来源亦为精确推导。PMID 38283651 的 acinar 案例反解出 δ=0.000,κ 与 AC1 两个值同时复现。

部分验证: 真实判读研究中 δ 的分布没有系统测量,本文的 0.02–0.08 区间来自评分者敏感度/特异度模拟与一项研究的反解,属估计而非普查。κ 在极小样本下的抽样波动参考 PMID 15733050 与 PMID 21969984 的区间估计框架,本文的 SE 数值来自 6000 次自助模拟,未与解析公式逐点对表。

待验证: ① 对 ≥30 项判读训练研究做边际回溯,算出 δ 的实测中位数与四分位,才能判定 0.007 这个量级是普遍成立还是只在特定子域成立;② 三评分者及多档情形下,δ 通道是否仍守平方律、v 是否换成广义形式;③ 当 P_o 与 m̄ 同时被报告时,(P_o, κ, m̄) 三元组能否唯一定出 δ——本文只验证了二分类两点情形,一般情形需要可辨识性证明。

相关阅读

证据等级: C
参考文献: PMID 8501467, PMID 15939215, PMID 38283651, PMID 19159474, PMID 21130355, PMID 25090041

L1 网络状态:该主题尚未进入 L1 对应网络 —— 边表 63 行无「一致性系数 / 评分者间信度 / 判读训练」相关边;网络 1881 行中「一致率」仅 1 处、「判读」仅 1 处、「κ」3 处,全部落在 E-060 标签前提(conflict 未裁决,涉及 Gwet AC1/AC2 与 Brennan-Prediger 的精度区间错配),与本条讨论的 δ 通道量级无对应关系。

证据等级说明:A=有 PubMed 直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索。本文属方法学推演,核心结论为本文自行推导的闭式解与模拟,PubMed 文献仅提供报告规范与实例锚点,故按 C 处理。


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 finding that \”the lower the consistency rate, the larger the channel\” measures only a single line in the corner: the δ channel is constantly equal to the square law, and its…

Evidence level: C

The closed-form solution for the δ-channel’s contribution to κ scales strictly with δ²; it increases with P_o solely because the chosen upper bound for δ itself grows with P_o, not because any mechanism amplifies it.

Body

One Arithmetic Step to Confirm It Is Correct

A number left over from the previous round: with P_o = 0.85 and marginal means of 0.43, driving the systematic bias δ between raters down from its feasible upper bound to zero can increase Cohen’s κ by at most 0.0069. If P_o is shifted to 0.75, the same formula yields an upper limit of 0.0306, which is 4.4 times the former; at 0.70 it becomes 0.0515, or 7.5 times.

This sequence of numbers must first be re-calculated. The expected agreement for a 2×2 contingency table has a closed-form expression: p_e = 1 − 2m̄ + 2m̄² − δ²/2, where m̄ is the average of the two raters’ positive rates and δ is their difference. κ = (P_o − p_e)/(1 − p_e). Sweeping P_o from 0.90 down to 0.65 while setting δ always to its upper bound of 1 − P_o yields increments of 0.0021, 0.0069, 0.0160, 0.0306, 0.0515, and 0.0793 respectively. The five points listed in the original problem match digit-for-digit. The arithmetic is correct.

The error lies in the interpretation of this curve.

That “Upper Bound” Is Actually Degenerate

The original derivation sets the feasible upper bound for δ at 1 − P_o: lower agreement means more remaining disagreement, so the available “space” for δ is larger. This sounds like a resource constraint. But substituting it into the four-cell table reveals the bottom: when δ exactly equals 1 − P_o, the cell p₀₁ in the table is constantly zero. Across twenty parameter sets and four P_o levels, p₀₁ is a floating-point zero of ±5.6e−17 magnitude.

The reason is an identity: p₀₁ + p₁₀ always equals 1 − P_o. Total disagreement is this number, which cannot escape. Since δ is the absolute difference between p₀₁ and p₁₀, the triangle inequality gives δ ≤ p₀₁ + p₁₀ = 1 − P_o. Thus “δ’s upper bound equals 1 − P_o” is not an interval opening up, but the singularity where “all disagreement is pushed to one side.” When δ reaches its upper bound, it means one rater never makes a mistake in one direction. This is degeneracy, not degrees of freedom.

Using this corner to represent the scale of the δ-channel is like pricing the entire curve using an artificially constructed extreme point.

The True Mechanism Is a Square Law, Not P_o

Fixing δ and sweeping P_o immediately makes the curve gentle. At δ = 0.05, lowering P_o from 0.92 to 0.65 increases the increment only from 0.00042 to 0.00182, never exceeding 0.002; at δ = 0.10 it goes from 0.00165 to 0.00721. In other words, the steep slope of “two orders of magnitude” previously attributed mostly comes from δ moving itself, not P_o amplifying anything.

The closed-form solution makes this definitive. After algebraic simplification:

δκ = (δ²/2)(1 − P_o) / [v(v + δ²/2)], where v = 2m̄(1 − m̄).

Back-substituting across the grid of P_o from 0.60 to 0.95, m̄ from 0.05 to 0.95, and δ over its feasible domain yields a maximum absolute error of 2.9e−15. This formula establishes three facts.

First, the δ-channel is quadratic. Halving δ reduces the channel to one-quarter; doubling δ from 0.05 to 0.10 increases the channel from 0.0013 to 0.0051 (roughly four times). This is not an approximation, but an identity. Therefore the question “Can the δ-channel be ignored?” should not ask about P_o, but about how large δ is.

Second, the sign is irrelevant. δ = +0.05 and δ = −0.05 yield identical κ values (0.491301). Who is stricter tells κ nothing. This measurement captures “how much bias exists,” not “which direction it biases toward.”

Third, while P_o does appear in the formula, it is a first-order multiplier, not an amplifier; the denominator v = 2m̄(1 − m̄) is what truly allows the channel to grow. As m̄ approaches 0 or 1, v shrinks, so the same δ produces a larger channel. When m̄ = 0.43, v = 0.49; when m̄ = 0.88, v = 0.21. Thus low agreement accompanied by extreme margins elevates the channel—but that is due to v, not P_o.

Three Constraints, Not Just One

There is another layer: δ ≤ 1 − P_o is only one of three constraints. Non-negativity in the four-cell table also requires δ ≤ 2m̄ and δ ≤ 2(1 − m̄). The actual usable upper bound is the minimum of these three. When m̄ = 0.10 and P_o = 0.70, the binding constraint is 2m̄ = 0.20, not 1 − P_o = 0.30; when m̄ = 0.88 and P_o = 0.75, it is 2(1 − m̄) = 0.24, not 0.25. The original problem used only 1 − P_o, effectively assuming m̄ falls in the middle segment, which is precisely what needs argumentation rather than default assumption.

How Large Is the Real δ

This step determines whether all the above numbers have practical meaning. If δ is merely 0.05—meaning the difference between two raters’ positive rates is five percentage points, which is already noticeable in diagnostic studies—then even with P_o pressed down to 0.70, the channel is only 0.00156, just twenty percent of the original 0.007 line. To push the channel beyond 0.007, δ must be at least 0.10, reaching 65% of the feasible domain when P_o = 0.80 and 47% when P_o = 0.75.

No systematic measurement of the true δ distribution exists; this remains a gap. However, one or two anchors can be read from published data. A chest pathology diagnosis study (PMID 38283651) reports the diagnosis of five histological types of lung adenocarcinoma: acinar type has a prevalence of 88%, observed agreement of 88%, κ = 0.43, AC1 = 0.85, positive agreement of 93% versus negative agreement of 50%—this is the textbook case of the “κ paradox.” Substituting P_o = 0.88 and κ = 0.43 into the formula yields a unique solution: m̄ = 0.88 and δ = 0.00. Back-substitution verifies: κ’s p_e = 0.7888, giving κ = 0.432; AC1’s p_e = 0.2112, giving AC1 = 0.848. Both match perfectly. Thus, in this repeatedly cited paradox case, the δ-channel’s contribution is precisely zero; κ is suppressed entirely due to the denominator v = 0.21.

Small-scale simulations based on rater sensitivity/specificity settings also point in the same direction: when both raters have equal ability, δ is approximately 0.001–0.000; only when one is deliberately set loose and the other tight (sensitivity 0.85 vs 0.75, specificity 0.85 vs 0.90) does δ reach 0.075, accounting for 28% of the disagreement budget, yielding a channel of 0.0031.

So How Is That 0.007 Actually Used

The original conclusion that “the δ-channel is negligible” holds when P_o ≥ 0.85, and this directional judgment is correct, but the supporting reason must be replaced: it is not because high agreement suppresses the channel, but because the true δ is inherently small. Conversely, the claim in the original that “if the domain’s P_o median falls below 0.78, the conclusion must reverse” is also invalid—low P_o does not automatically make δ larger; it merely raises δ’s upper bound, and reaching that upper bound implies the table has degenerated.

The only prerequisite check that truly matters is this: report δ, or report the two marginal values sufficient to calculate δ. This requirement was proposed thirty years ago (PMID 8501467): comparing consistency studies reporting only κ is misleading; quantitative measures of bias and prevalence must also be provided. Some authors argue that dependence on P_o fundamentally does not oppose κ (PMID 15939215), but this is stated from the perspective of latent prevalence and does not conflict with “marginals must be reported to solve for δ.” The GRRAS reporting guidelines (PMID 21130355) also list marginals and agreement as mandatory items.

Returning to the earlier conclusion: training δ down to zero can increase κ by at most 0.0069. This upper limit can now be stated more precisely: it is the ceiling for “δ moving from the feasible corner to zero,” a ceiling-on-a-ceiling. Real-world training reducing δ from 0.05 to 0.02, with P_o = 0.75 and m̄ = 0.43, yields only 0.0011; the corner-derived 0.0306 is 28 times larger, and at P_o = 0.70 it is 39 times larger.

Evidence and Limitations

Verified: The closed-form solution for the δ-channel, δκ = (δ²/2)(1 − P_o)/[v(v + δ²/2)], was back-substituted across a 35×91×feasible-domain grid with a maximum absolute error of 2.9e−15, establishing it as an identity-level result. The derivations of p₀₁ + p₁₀ = 1 − P_o and δ ≤ 1 − P_o via the triangle inequality are also exact. The acinar case from PMID 38283651 was solved backwards to yield δ = 0.000, successfully reproducing both κ and AC1 values.

Partially Verified: The distribution of δ in real diagnostic studies has not been systematically measured; the 0.02–0.08 interval in this paper stems from rater sensitivity/specificity simulations and the back-solution of one study, representing estimation rather than census. Sampling variability of κ under extremely small samples references the interval estimation frameworks of PMID 15733050 and PMID 21969984; the SE values in this paper derive from 6,000 bootstrap simulations and were not matched point-by-point against analytical formulas.

Pending Verification: ① Conduct marginal back-calculation on ≥30 diagnostic training studies to compute the empirical median and quartiles of δ, thereby determining whether the 0.007 magnitude is universally valid or restricted to specific subdomains; ② In multi-rater and multi-level scenarios, does the δ-channel still obey the square law, and does v adopt a generalized form; ③ When both P_o and m̄ are reported, can the triplet (P_o, κ, m̄) uniquely determine δ—this paper verified only the binary two-point case, while general cases require identifiability proofs.

Related Reading

Evidence Level: C
References: PMID 8501467, PMID 15939215, PMID 38283651, PMID 19159474, PMID 21130355, PMID 25090041

L1 Network Status: This topic has not yet entered the corresponding L1 network—the edge table row 63 lacks edges related to “consistency coefficient / inter-rater reliability / diagnostic training”; among 1,881 rows in the network, “agreement rate” appears only once, “diagnosis” once, and “κ” three times, all falling under the E-060 label premise (conflict unresolved, involving mismatched precision intervals between Gwet’s AC1/AC2 and Brennan-Prediger), which bears no correspondence to the δ-channel magnitude discussed here.

Evidence Level Explanation: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration. This paper belongs to methodological derivation; core conclusions are the closed-form solution and simulations derived herein, while PubMed literature provides only reporting guidelines and example anchors, hence classified as C.


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 *