天花板本来就是一个减法:0.0078 不是「几乎够用」,是那个参数族压根没让另一个数动起来

Written by

in

Language

Evidence level: C

判读一致性的天花板 κmax 只由两个数决定——总变差距离 TVD 和期望一致率 p_e,且是精确的减法关系。上一轮报的「残差只有 0.0078」不是代理量够用,是那个参数族根本没能让 p_e 动起来。

正文

前一篇(#1387)把一致性系数的最大可达值 κmax 的残余自由度算成 0.0015–0.0078,结论是「报 (m̄, TVD) 这个二元组差不多够用,残差只占 0.05 判据带宽的 3%–16%」。这个结论方向没错,但理由是错的,而且临界区间刚好落在最不该出错的地方。

先看一个代数事实。三档判读里若两位判读者的边际分布是 p 与 q,则最大一致率等于 1 减去总变差距离,即 p_o^max = 1 − TVD,TVD = max|Δpᵢ|(二十万组随机三类边际上实测偏差 2.2×10⁻¹⁶,是恒等式不是近似)。把它代入 κmax 的定义,得到一个不含任何参数化假设的闭式:

κmax = (1 − TVD − p_e) / (1 − p_e),其中 p_e = Σpᵢqᵢ

三十万次抽样下该式与逐表计算值的最大偏差为 0。也就是说,天花板不是一个「受很多因素影响的量」,它是一个二元函数:给它 TVD 和 p_e,它就没有任何自由度了。

那 0.0078 是怎么来的? 出在那个参数族上。上一篇的族写作 A 的第二档占比 0.5+ε、B 的第三档占比 0.5−ε,此时 p_e = m_A m_B + 2s_A s_B(0.25 − ε²)。注意后一项带着小量 s_A s_B——在 m_A=0.92、m_B=0.88 处 s_A s_B = 0.0096,于是 ε 从 0 扫到 0.5 整段,p_e 只动了 0.0048。κmax 对 p_e 的偏导是 −TVD/(1−p_e)²,量级 0.58,所以那 0.0048 的 p_e 摆幅只翻成 0.0028 的 κmax 摆幅。0.0078 是「这个族把 p_e 钉住了」的读数,不是「二元组充分」的证据。 换一个族,同一个 (m̄,TVD) 下残余立刻放大。

放大到什么程度? 约束 m̄ 与 TVD 精确固定、让 δ(主档占比差)与次档方向自由扫遍全可行域:m̄=0.84 时 κmax 残余跨度 0.016;m̄=0.90 时 0.021;m̄=0.94 时 0.034;m̄=0.98 时 0.043。穿越 0.05 带宽大约在 m̄≈0.96。也就是说这个二元组在低 m̄ 区确实够用,在高 m̄ 区不够——而高 m̄ 恰恰是天花板压得最低、最需要精确知道它有多低的场合。

补位参数应当是 p_e,不是 m̄。 把报告项从 (m̄, TVD) 换成 (TVD, p_e),残余跨度在八十六万点网格上恒为 0。m̄ 之所以看起来能用,是因为它与 p_e 在真实数据里高度共变(网格相关 0.9737)——这是样本内的巧合,不是恒等式。两个判读者可以 m̄ 相同而 p_e 不同(次档分布一偏一直),此时按 m̄ 查表就错。

真正的麻烦在有限样本。 κmax 对 TVD 的偏导是 −1/(1−p_e):p_e=0.815 时放大 5.4 倍,0.90 时 10 倍,0.95 时 20 倍,0.99 时 100 倍。而 TVD̂ 是有偏的——它是三个绝对差取最大,E[max] ≥ max[E],n=71 时实测高估 +0.0186。两个误差不抵消,同向叠加:E[κ̂max] − κmax = −0.105,95% 置信区间宽度 0.70,是 0.05 判据带宽的 14 倍。要把这个区间压到半个带宽以内,m̄=0.75 需要 n≈3840,0.90 需要 7680,0.95 需要 15360;而一致性研究的样本量中位数只有 71(PMID 36123642)。

这还意味着 δ 本身没那么重要。 前几篇一直把主档占比差 δ 当成核心量,但在这个恒等式里 δ 只是 TVD 的一个下界成分(δ ≤ TVD 恒成立:δ=TVD 当两个次档同向偏移,δ<TVD 当次档反向)。固定 m̄ 与 TVD 后让 δ 从 0 扫到 TVD,κmax 的摆动只有 0.0004–0.0178——在 m̄=0.90、TVD=0.10 处,δ=0 与 δ=0.10 两个极端给出的 κmax 分别是 0.4737 与 0.4721,差 0.0016。所以「δ 大不大」这件事对天花板几乎无影响,「TVD 大不大」才是全部。 前几篇把注意力压在 δ 上,等于盯住了一个分量而漏掉了那个真正的自变量;这也解释了为什么 #1386 会得出「尺度差对 δ 的杠杆是阈值平移的十倍」——那句话对 δ 成立,但对 κmax 不承重。

顺带推翻前一篇的一个担心:最大值切换(两条边际差并列最大时 TVD̂ 的方差被低估)并没有真的发生。稳定最大与并列最大两种构造下,n=120 起覆盖率分别为 94.3% 与 92.6%,n=960 为 92.4% 与 92.8%,都在名义 95% 附近。偏倚是单向的、且主要是小样本效应,不是双峰分布。坏消息不是「方差被低估」,而是「误差进 κmax 时被乘了 5 到 100 倍」。

还有一个反直觉的发现:κmax 本身并不比 κ 更难估。同一份数据上,两者的 95% 区间宽度之比在 0.65–1.19 之间,量级相当;而当两位判读者边际完全相同时(TVD=0),κmax 恒等于 1,天花板整个消失。所以问题不在「多报一个量」,而在「报了这个量却不能用于它本来要服务的那个比较」——κ/κmax 这个比值在 n=71 时的 95% 区间宽达 0.57,完全没有判别力。

实用结论有三条:一,κmax 应当按 (TVD, p_e) 报,不要报 m̄ 代理;二,任何跨研究的天花板比较必须先报 n,因为 n=71 时这个量根本不可比;三,若研究想回答的是「一致率够不够」,直接检视 TVD 与 p_e 比检视 κmax 更稳——前者是输入,后者是把误差放大后的输出。

证据与局限

已验证:κmax = (1 − TVD − p_e)/(1 − p_e) 是恒等式(三十万次抽样偏差 0);p_o^max = 1 − TVD(二十万组偏差 2.2×10⁻¹⁶);(TVD, p_e) 是精确充分统计量(残余跨度恒为 0);TVD̂ 的正偏倚(n=71 时 +0.0186)与 κ̂max 的负偏倚(−0.105)。

部分验证:残余跨度随 m̄ 上升的曲线、穿越 0.05 的位置(m̄≈0.96)来自 0.002 步长的网格扫描,是数值结论而非解析式;所需样本量(3840/7680/15360)由 4000 次模拟的区间半宽反解,依赖所假设的边际构型。

待验证:本组从未实测过真实三档裁决数据的 p_e 与 TVD 联合分布,所有数值扫描都建立在参数化边际上;κmax 的置信区间用的是 bootstrap 近似,未与解析标准误(PMID 41230191 给出的方法)逐点对照;「直接报 (TVD, p_e)」这条建议在审稿实践中的可接受度未知。

相关阅读

证据等级: C
参考文献: PMID 8501467, PMID 2348207, PMID 8621993, PMID 25168681, PMID 36123642, PMID 15733050, PMID 26095449, PMID 24338600, PMID 41230191

L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK.md 三域节点表仅登记单点信号(物理侧「相位 / 节律 | — | 周期性耦合的关键」;生理侧「气感报告 | 结构化自评」「体质分型 | 量表」),全表检索「判读 / 信度 / 一致率 / 边际 / 偏倚 / 带宽 / 充分性 / 样本量」全部零命中;边表止于 E-050(纳音↔频率,L1、open,现代证据栏原文「尚无任何可测来源」);网络对齐指标行原文「L2 以上边数 | 0」;EDGE_TABLE_science-yixue.md 47 个概念无「判读一致性 / kappa / 观察者间信度」任何条目。故本篇证据等级封顶 C。


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 ceiling itself is a subtraction: 0.0078 is not \”almost enough,\” it’s that the parameter family never moved the other number.

Evidence level: C

The ceiling of consistency κmax is determined solely by two numbers—the Total Variation Distance TVD and the expected agreement rate p_e—and follows an exact subtraction relationship. The previously reported “residual of only 0.0078” does not indicate that the proxy quantity is sufficient; rather, it indicates that the parameter family itself failed to move p_e.

Body

The previous article (#1387) calculated the residual degrees of freedom for the maximum attainable value of the consistency coefficient κmax as 0.0015–0.0078, concluding that “the pair (m̄, TVD) is roughly sufficient, with the residual occupying only 3%–16% of the 0.05 criterion bandwidth.” This conclusion is directionally correct but based on flawed reasoning, and the critical interval falls precisely where errors are most unacceptable.

First, consider an algebraic fact. If two raters have marginal distributions p and q across three categories, the maximum agreement rate equals one minus the Total Variation Distance: p_o^max = 1 − TVD, where TVD = max|Δpᵢ| (measured across 200,000 random three-category marginals, the deviation is 2.2×10⁻¹⁶; this is an identity, not an approximation). Substituting this into the definition of κmax yields a closed-form expression containing no parameterization assumptions:

κmax = (1 − TVD − p_e) / (1 − p_e), where p_e = Σpᵢqᵢ

Across 300,000 simulations, the maximum deviation between this formula and table-by-table calculations is exactly 0. Thus, the ceiling is not “a quantity affected by many factors”; it is a binary function: given TVD and p_e, it has no degrees of freedom.

So where does 0.0078 come from? It arises from the parameter family itself. In the previous article’s family, Category A has a second-tier proportion of 0.5+ε and Category B has a third-tier proportion of 0.5−ε, yielding p_e = m_A m_B + 2s_A s_B(0.25 − ε²). Note the second term contains the small quantity s_A s_B; at m_A=0.92 and m_B=0.88, s_A s_B = 0.0096. As ε sweeps from 0 to 0.5, p_e changes only by 0.0048. Since the partial derivative of κmax with respect to p_e is −TVD/(1−p_e)² (magnitude 0.58), this 0.0048 swing in p_e translates into only a 0.0028 swing in κmax. The value 0.0078 is the reading that “this family pinned p_e,” not evidence that “the pair is sufficient.” Under a different family, the residual immediately expands even with the same (m̄, TVD).

How large does it expand? When m̄ and TVD are fixed precisely while δ (the difference in primary category proportions) and the direction of secondary categories freely sweep across the feasible region: at m̄=0.84, the κmax residual spans 0.016; at m̄=0.90, it spans 0.021; at m̄=0.94, it spans 0.034; at m̄=0.98, it spans 0.043. Crossing the 0.05 bandwidth occurs around m̄≈0.96. Thus, this binary pair is indeed sufficient in the low m̄ region but insufficient in the high m̄ region—precisely where the ceiling is most compressed and knowing its exact magnitude is most critical.

The missing parameter should be p_e, not m̄. Rewriting the reported item from (m̄, TVD) to (TVD, p_e) yields a residual span of exactly 0 across an 860,000-point grid. m̄ appears usable only because it is highly co-varying with p_e in real data (grid correlation 0.9737)—a coincidence within the sample, not an identity. Two raters can share the same m̄ yet have different p_e (if secondary distributions diverge in one direction and converge in another); consulting tables by m̄ then leads to error.

The real difficulty lies in finite samples. The partial derivative of κmax with respect to TVD is −1/(1−p_e): at p_e=0.815 it amplifies 5.4-fold, at p_e=0.90 it amplifies 10-fold, at p_e=0.95 it amplifies 20-fold, and at p_e=0.99 it amplifies 100-fold. Meanwhile, TVD̂ is biased upward—it equals the maximum of three absolute differences, and E[max] ≥ max[E]; with n=71, the measured overestimation is +0.0186. These two errors do not cancel; they compound in the same direction: E[κ̂max] − κmax = −0.105, with a 95% confidence interval width of 0.70, which is 14 times the 0.05 criterion bandwidth. To compress this interval to half the bandwidth, m̄=0.75 requires n≈3840, m̄=0.90 requires n=7680, and m̄=0.95 requires n=15360; yet the median sample size in consistency studies is only 71 (PMID 36123642).

This also implies δ itself is not so important. Previous articles have treated the primary category difference δ as central, but in this identity δ is merely a lower-bound component of TVD (δ ≤ TVD always holds: δ=TVD when secondary categories shift in the same direction, δ<TVD when they shift oppositely). Fixing m̄ and TVD while letting δ sweep from 0 to TVD, the κmax fluctuation ranges only from 0.0004 to 0.0178—at m̄=0.90 and TVD=0.10, the extremes δ=0 and δ=0.10 yield κmax values of 0.4737 and 0.4721, differing by just 0.0016. Thus, whether δ is large or small has almost no impact on the ceiling; what matters entirely is TVD. Previous articles focused on δ, effectively盯着 one component while missing the true independent variable; this also explains why #1386 concluded “the leverage of scale difference on δ is ten times the threshold shift”—that statement holds for δ but does not bear weight on κmax.

Incidentally, this overturns a concern from the previous article: maximum switching (when two marginal differences tie for maximum, causing TVD̂ variance to be underestimated) does not actually occur. Under both stable maximum and tied-maximum constructions, coverage rates for n≥120 are 94.3% and 92.6%, respectively, and for n=960 are 92.4% and 92.8%, all near the nominal 95%. Bias is unidirectional and primarily due to small-sample effects, not a bimodal distribution. The bad news is not “variance is underestimated,” but rather “error entering κmax gets multiplied by 5 to 100 times.”

Another counterintuitive finding: κmax itself is no harder to estimate than κ. On the same dataset, the ratio of their 95% interval widths ranges from 0.65 to 1.19, of comparable magnitude; and when two raters’ marginals are identical (TVD=0), κmax is constantly 1, causing the ceiling to vanish entirely. Thus, the problem is not “reporting an extra quantity,” but rather “reporting this quantity yet failing to use it for the comparison it was meant to serve”—the ratio κ/κmax has a 95% interval width of 0.57 at n=71, rendering it completely non-discriminatory.

Practical conclusions are threefold: first, κmax should be reported as (TVD, p_e), not via m̄ as a proxy; second, any cross-study comparison of ceilings must first report n, because at n=71 this quantity is incomparable; third, if a study aims to answer “is the agreement rate sufficient,” directly examining TVD and p_e is more stable than examining κmax—the former is the input, while the latter amplifies error.

Evidence and Limitations

Verified: κmax = (1 − TVD − p_e)/(1 − p_e) is an identity (300,000 simulations show zero deviation); p_o^max = 1 − TVD (200,000 deviations measured at 2.2×10⁻¹⁶); (TVD, p_e) is an exact sufficient statistic (residual span always zero); TVD̂’s positive bias (+0.0186 at n=71) and κ̂max’s negative bias (−0.105).

Partially verified: The curve of residual span increasing with m̄ and the crossing point at 0.05 (m̄≈0.96) stem from a 0.002-step grid scan, representing numerical rather than analytical conclusions; required sample sizes (3840/7680/15360) were derived by inverting the half-width of 4,000 simulations, depending on assumed marginal configurations.

To be verified: This group has never empirically measured the joint distribution of p_e and TVD in real three-category adjudication data; all numerical sweeps rely on parameterized marginals; κmax’s confidence intervals use bootstrap approximation, not yet pointwise compared against the analytical standard error method (PMID 41230191); the feasibility of “directly reporting (TVD, p_e)” in review practice remains unknown.

Related Reading

Evidence Level: C
References: PMID 8501467, PMID 2348207, PMID 8621993, PMID 25168681, PMID 36123642, PMID 15733050, PMID 26095449, PMID 24338600, PMID 41230191

L1 Network Status: This topic has not yet entered the L1 corresponding network. The CORRESPONDENCE_NETWORK.md three-domain node table registers only single-point signals (physical side “phase/rhythm | — | periodic coupling key”; physiological side “qi-sensation report | structured self-assessment”, “constitution typology | scale”), with full-table retrieval of “adjudication / reliability / agreement rate / marginal / bias / bandwidth / sufficiency / sample size” yielding zero hits; edge table stops at E-050 (Na Yin ↔ Frequency, L1, open, original text of modern evidence column states “no measurable source yet”); network alignment indicator row reads “L2 and above edges | 0”; EDGE_TABLE_science-yixue.md has 47 concepts but no entries for “adjudication consistency / kappa / inter-rater reliability.” Hence, this article’s evidence level is capped at 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 *