报出 kappa、一致率和主档占比,还剩多少没说:那条带宽就是辨识的天花板

Written by

in

Language

Evidence level: C

一研究报了三档裁决的 kappa、观察一致率、主档占比三个数,看起来该说的都说了。但给定这三个数,剩下那个自由度能让 kappa 摆动 0.016 到 0.156——摆动幅度随主档占比单调收窄,且完全由算术决定。这条带宽,就是数据能告诉你的上限。

正文

上一篇(kappa-marginal-concentration-p2-dominance)得出「p₀ 固定后 kappa 只有一个自变量 Σp²」。它的待验证清单里留了三条,其中最扎人的一条是:主档占比 m 与主档判对率 q 是否独立。如果这两个量绑在一起(主档占比高只因为「大家都在说没变化」),那么上一篇那张「m 从 0.33 扫到 0.95」的膝部表就只是反事实。

这篇先把这一条做实,再顺手算出一个上一篇没算的量:给定 (p₀, m) 之后,还剩多少自由度。结果比预想的干净——它有一个闭式解,而且这个解给出了三档裁决数据的能力上限。

一、先给 m 与 q 松绑:它们不绑定

上一篇的担心是合理的猜想,但它是错的。我做了一次随机参数扫描:让判读能力(非主档判对率 c、主档误判到相邻档的单侧率 h、跨两档误判率 g)与真实病情构成(τ₁/τ₂/τ₃)四个量自由取值,各抽 40 万组,只看那些落在同一个主档占比附近的样本:

锁定 m ≈ 样本数 q 的实测范围 跨幅
0.30 8422 0.012 – 0.991 0.979
0.50 14020 0.195 – 0.996 0.801
0.70 9256 0.665 – 1.000 0.335

同一个「判读者说主档占三成」,主档判对率可以从 1.2% 一直走到 99.1%。m 与 q 不绑定。 上一篇那条「没人敢判变化 → 主档占比虚高 → 判对率被一起抬高」的自我强化机制,在任何合理的判读模型里都不成立:因为 m 高可以由「样本里真没变化」造出来,也可以由「判读者不敢判」造出来,这两条路径给出的 q 截然相反。

这就救回了上一篇的膝部表:m 确实可以当作一个近似独立的设计变量来扫。上一轮的担心作废。

不过注意跨幅随 m 单调收窄(0.98 → 0.80 → 0.34)。这不是巧合,下面第三节会给出它的解析来源。

二、真正锁住的不是 m 与 q,是 Σp² 的带宽

三档、两位判读者、边际相等。这一点是三档裁决的默认情形(两位判读者对同一批人,通常不会一个人判出三成改善、另一个人判出三成加重)。边际相等时,Σp² 被 m 完全夹住:

Σp² ∈ [ m² + (1−m)²/2 ,  m² + (1−m)² ]

下界对应「非主档的两档完全均分」,上界对应「非主档全部压在同一个档」。这个式子不依赖 p₀,只依赖 m。

再套回 kappa = (p₀ − Σp²)/(1 − Σp²)(单调递减),立刻得到给定 (p₀, m) 时 kappa 的完整区间。脚本实算(p₀ = 0.90):

m Σp² 带宽 kappa 区间 kappa 带宽
0.20 0.320 [0.688, 0.844] 0.156
0.30 0.245 [0.762, 0.850] 0.088
0.40 0.180 [0.792, 0.849] 0.057
0.50 0.125 [0.800, 0.840] 0.040
0.70 0.045 [0.762, 0.785] 0.023
0.85 0.0112 [0.608, 0.624] 0.017
0.90 0.0050 [0.444, 0.460] 0.015
0.95 0.0013 [−0.053, −0.039] 0.014

闭式与 9 格穷举(60×60×60 网格)的最大偏差 0.0013,出现在 m = 0.20 处;m ≥ 0.5 后偏差 ≤ 0.0007。式子成立。

三、这条带宽的含义,比它看起来重要

三个推论,一个比一个不客气。

第一,报 m 比报 Σp² 更划算,但两者不可互换。 同样的 m 精度,Σp² 的区间宽度在 m 小时可达 0.32(m = 0.20),在 m 大时压到 0.0013(m = 0.95)。带宽随 m 单调收窄,且收缩极快——从 m = 0.20 到 m = 0.90,带宽掉了 64 倍。这意味着:m 越大,kappa 越是被算术钉死,「判得更好」越插不上手。这正是上一篇那条解析零点(κ<0 ⇔ Σp² > p₀)的定量补充:p₀ = 0.90 时,m 过了 0.90 附近,kappa 必为负,且带宽只剩 0.015——你连「靠改进判读把 kappa 拉回正数」的理论空间都没有。

第二,那种「p₀ 几乎不动而 kappa 走完 0.43 全程」的现象,是参数对称性造出来的,不是机制。 我第一版算的时候,三档判对率全部设成 0.90(相邻错位 0.05、跨档误判 0.01),于是三行的平方和都等于 0.8182,p₀ 在整条扫描上纹丝不动(极差 0.0029),kappa 却从 0.30 走到 0.73。那个「放大 144 倍」看起来很惊人,但它只是「三个对角元相等」这一个设定的产物。换成更真实的一组(主档好判 0.94、非主档 0.82),p₀ 极差立刻变成 0.176,放大倍数掉到 2.1 倍,kappa 的峰也从 m ≈ 0.33 移到 m ≈ 0.55。

这条是我自己踩的坑,得写下来:仿真里任何「p₀ 几乎不变而 kappa 大变」的结论,第一件事是查 p₀ 为什么不变——它多半是参数对称性造出来的,不是机制。

第三,最硬的落点:三个数报完,8% 到 15% 的 kappa 区间还是空的。 m = 0.20 时带宽 0.156,等于说 kappa 报 0.77 的研究,真实值可能在 0.69 到 0.84 之间。而 kappa 的常规解读档(Landis-Koch 那套 0.41/0.61/0.81)宽度只有 0.20——一条带宽就能吃掉大半档位。所以三档裁决的 kappa,在 m 小的时候连跨档判断都不该做。

四、跟文献对得上

这条闭式的实质,Lantz 与 Nebenzahl 早在 1996 年就说清楚了:给定观察一致率 p₀,kappa 有三个特征值 kappa_max / kappa_min / kappa_nor,全部只是 p₀ 的函数(PMID 8621993)。他们据此建议:报 kappa 必须同时报 p₀、不一致的对称性 S_D、以及在「有该特质」上的一致率 P₊₊。

本文做的是把这条建议在三档、边际相等这个具体情形下算成闭式,并指出它等价于「Σp² 被 m 夹在一条带宽里」。Byrt 等人的三指数分解(偏倚 / 流行度 / 观察一致率,PMID 8501467)是同一件事的另一条路径——他们从「kappa 该拆成三个独立指数」出发,本文从「Σp² 的可行域是个区间」出发,落点相同。Feinstein 与 Cicchetti 1990 年那两个悖论(PMID 2348207)在这里也没被绕过:本文第一、二节的表里,kappa 从 0.69 走到 0.84 再掉到负值,全程判读能力不变,只是 m 在动。

五、给三档裁决报告规范的具体改动

上一篇提的三联报(p₀, Σp², kappa)继续有效,但要加一条:必须报 Σp² 的带宽,或者等价地报 m。理由是:

  1. 只报 Σp² 的点值,读者无法知道这个点值是被钉死的还是浮动的。m = 0.90 时点值精度 ±0.0025 就够,m = 0.20 时同样精度毫无意义。
  2. 带宽本身是个诊断量。带宽 > 0.05(约 m < 0.45)时,kappa 的任何跨档比较都不成立,只能同 m 区间内比。
  3. 带宽随 m 单调,所以「m 大」这件事既是坏消息(kappa 趋负)也是好消息(读数确定)——这两件事必须一起说,不能只报其中一个。

证据与局限

已验证:
– 闭式 Σp² ∈ [m² + (1−m)²/2, m² + (1−m)²] 在边际相等条件下由「非主档两档如何分配」直接推出;与 9 格穷举的最大偏差 0.0013(m = 0.20),m ≥ 0.5 后偏差 ≤ 0.0007。
– 带宽随 m 单调收窄:m 从 0.20 到 0.95,Σp² 带宽从 0.320 降到 0.0013,kappa 带宽从 0.156 降到 0.014。
– m 与 q 不绑定:随机参数扫描 40 万组,锁定 m ≈ 0.30 时 q 的跨幅 0.979(0.012–0.991),锁定 m ≈ 0.70 时跨幅 0.335。上一篇担心的「m 与 q 耦合」不成立。
– 「p₀ 恒定而 kappa 大变」可以由参数对称性伪造:两套参数(三档判对率全 0.90,三行平方和同为 0.8182 vs 主档 0.94/非主档 0.82)分别给出 p₀ 极差 0.0029 与 0.176,是同一现象的两个来源。

部分验证:
– 闭式只在边际相等时成立。边际不等时 Σp² 的可行域会变宽(多一个偏倚维度),带宽会更大,闭式失效。下一篇要处理这个(#1383 加权 kappa 的等值线)。
– 「kappa 的常规解读档宽 0.20」这句话依赖 Landis-Koch 那套分级,而该分级本身在文献里长期被批评任意。本文只把它当量级参照,不当阈值。
– 随机扫描用的是单一有序三档模型(相邻错位率 h、跨档错率 g、非主档判对率 c)。更复杂的错误结构(比如两位判读者能力不等)没有覆盖。

待验证:
– 该主题尚未进入 L1 对应网络:CORRESPONDENCE_NETWORK 的三域节点表只有单点信号节点(物理侧「相位/节律」、生理侧「呼吸率/呼吸模式」「气感报告」),全表检索「判据 / 裁决 / 一致率 / 判读 / 信度 / 边际 / 观察者间」全部零命中;边表止于 E-050(纳音↔频率,open,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」。故本文封顶 C,全部结论只作方法学讨论,不得读作已确立机制。
– 需要一次实测:同一批 ≥60 名受试者、3 位判读者独立三档判读,量出真实的 (p₀, m, Σp², kappa) 四点,才能检验本文带宽是否覆盖实测点。若实测 Σp² 落在带宽外,说明两位判读者的边际不等,闭式须加偏倚项。
– 边际不等情形下的 Σp² 可行域尚未推出闭式,只知道它比相等情形宽,宽多少取决于偏倚量级。

相关阅读

证据等级: C
参考文献: PMID 2348207, PMID 8501467, PMID 8621993
L1 网络状态: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」,语义为两振荡锁相、不带一致性/判读/边际语义;生理侧「呼吸率 / 呼吸模式 | 可穿戴」「气感报告 | 结构化自评」),全表检索「判据 / 裁决 / 一致率 / 判读 / 信度 / 边际 / 观察者间」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」。→ 该边在 L1 无对应承重边,本文封顶 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。

Report kappa, agreement, and proportion of the main file; what remains unsaid is that bandwidth defines the ceiling of identification.

Evidence level: C

A study reporting three metrics—kappa, observed agreement ($p_0$), and main-tier proportion ($m$)—appears comprehensive. However, given these three values, the remaining degree of freedom allows kappa to swing between 0.016 and 0.156. This range narrows monotonically as $m$ increases, determined entirely by arithmetic. This bandwidth represents the upper limit of what the data can tell you.

Body

The previous article (kappa-marginal-concentration-p2-dominance) established that “once $p_0$ is fixed, kappa has only one independent variable, $\Sigma p^2$.” Its checklist of items to verify included three entries, with the most pressing being: whether the main-tier proportion $m$ and the main-tier correct rate $q$ are independent. If these two quantities are bound together (i.e., a high main-tier proportion exists merely because “everyone said there was no change”), then the previous article’s knee-table (sweeping $m$ from 0.33 to 0.95) would represent only counterfactuals.

This article first solidifies that independence between $m$ and $q$, then calculates an additional quantity the previous article did not address: the remaining degree of freedom given $(p_0, m)$. The result is cleaner than expected—it admits a closed-form solution, which also defines the upper limit of capability for three-tier verdict data.

I. Decoupling $m$ and $q$: they are not bound

The concern in the previous article was a reasonable conjecture, but it is incorrect. I performed a random parameter sweep: allowing the diagnostic capability (non-main-tier correct rate $c$, single-side error rate of misjudging the main tier as adjacent $h$, cross-tier error rate $g$) and the true condition composition ($\tau_1/\tau_2/\tau_3$) to take free values, sampling 400,000 groups each, and examining only those samples falling near the same main-tier proportion:

Locked $m \approx$ Sample count Observed range of $q$ Span
0.30 8422 0.012 – 0.991 0.979
0.50 14020 0.195 – 0.996 0.801
0.70 9256 0.665 – 1.000 0.335

For the same statement “the adjudicators say the main tier accounts for 30%,” the main-tier correct rate can range from 1.2% all the way to 99.1%. $m$ and $q$ are not bound. The self-reinforcing mechanism described in the previous article—that “no one dares to judge change $\rightarrow$ main-tier proportion is artificially inflated $\rightarrow$ correct rate is jointly elevated”—does not hold in any reasonable adjudication model: because a high $m$ can arise either from “true absence of change in the sample” or from “adjudicators’ reluctance to judge,” these two pathways yield diametrically opposite $q$ values.

This rescues the previous article’s knee-table: $m$ can indeed be treated as an approximately independent design variable to sweep. The previous concern is invalidated.

Note, however, that the span narrows monotonically with $m$ (0.98 $\rightarrow$ 0.80 $\rightarrow$ 0.34). This is not a coincidence; Section III below provides its analytical source.

II. What is truly locked is not $m$ and $q$, but the bandwidth of $\Sigma p^2$

Three tiers, two adjudicators, equal margins. This is the default scenario for three-tier verdicts (two adjudicators rarely assign 30% improvement to one person and 30% worsening to another for the same batch). When margins are equal, $\Sigma p^2$ is completely constrained by $m$:

Σp² ∈ [ m² + (1−m)²/2 ,  m² + (1−m)² ]

The lower bound corresponds to “the two non-main tiers are perfectly split,” while the upper bound corresponds to “all non-main tiers pile into a single tier.” This formula does not depend on $p_0$; it depends only on $m$.

Substituting back into kappa = $(p_0 − \Sigma p^2)/(1 − \Sigma p^2)$ (monotonically decreasing), we immediately obtain the full interval for kappa given $(p_0, m)$. Script calculation (with $p_0 = 0.90$):

$m$ $\Sigma p^2$ bandwidth kappa interval kappa bandwidth
0.20 0.320 [0.688, 0.844] 0.156
0.30 0.245 [0.762, 0.850] 0.088
0.40 0.180 [0.792, 0.849] 0.057
0.50 0.125 [0.800, 0.840] 0.040
0.70 0.045 [0.762, 0.785] 0.023
0.85 0.0112 [0.608, 0.624] 0.017
0.90 0.0050 [0.444, 0.460] 0.015
0.95 0.0013 [−0.053, −0.039] 0.014

The maximum deviation between the closed form and the exhaustive 9-grid search (60×60×60 grid) is 0.0013, occurring at $m = 0.20$; for $m \geq 0.5$, the deviation is $\leq 0.0007$. The formula holds.

III. The meaning of this bandwidth is more significant than it appears

Three implications follow, each more uncompromising than the last.

First, reporting $m$ is more cost-effective than reporting $\Sigma p^2$, but the two are not interchangeable. For the same precision of $m$, the interval width of $\Sigma p^2$ can reach 0.32 when $m$ is small ($m = 0.20$), but shrinks to 0.0013 when $m$ is large ($m = 0.95$). The bandwidth narrows monotonically with $m$, and contracts extremely rapidly—from $m = 0.20$ to $m = 0.90$, the bandwidth drops by a factor of 64. This means: the larger $m$ is, the more kappa is pinned down by arithmetic; “better adjudication” has less room to maneuver. This precisely quantifies the analytical zero point mentioned in the previous article ($\kappa < 0 \iff \Sigma p^2 > p_0$): when $p_0 = 0.90$, once $m$ passes near 0.90, kappa must be negative, and the bandwidth shrinks to only 0.015—you lack even the theoretical space to “pull kappa back to positive numbers by improving adjudication.”

Second, the phenomenon where “$p_0$ remains almost static while kappa traverses the full 0.43 range” is created by parameter symmetry, not mechanism. In my first calculation, I set all three-tier correct rates to 0.90 (adjacent misalignment 0.05, cross-tier error 0.01), so the sums of squares for all three rows equal 0.8182, $p_0$ remains unchanged across the entire sweep (range 0.0029), yet kappa moves from 0.30 to 0.73. That “144-fold amplification” appears startling, but it is merely the product of “all three diagonal elements being equal.” Switching to a more realistic set (main tier easy to judge at 0.94, non-main tiers at 0.82), the range of $p_0$ immediately becomes 0.176, the amplification factor drops to 2.1, and the peak of kappa shifts from $m \approx 0.33$ to $m \approx 0.55$.

This is a pitfall I stepped into; I must record it: in any simulation, the conclusion that “$p_0$ remains nearly constant while kappa changes dramatically” must first check why $p_0$ remained constant—it is mostly fabricated by parameter symmetry, not mechanism.

Third, the hardest point: after reporting the three numbers, an 8% to 15% kappa interval remains empty. When $m = 0.20$, the bandwidth is 0.156, meaning a study reporting kappa = 0.77 could have a true value anywhere between 0.69 and 0.84. The conventional interpretation bands for kappa (Landis-Koch’s 0.41/0.61/0.81) have widths of only 0.20—one bandwidth can swallow most of these bands. Therefore, for three-tier verdicts, kappa should not even attempt cross-tier comparisons when $m$ is small.

IV. Alignment with literature

The essence of this closed form was clarified by Lantz and Nebenzahl back in 1996: given the observed agreement $p_0$, kappa has three eigenvalues $\kappa_{\text{max}} / \kappa_{\text{min}} / \kappa_{\text{nor}}$, all of which are functions solely of $p_0$ (PMID 8621993). They recommended reporting kappa must simultaneously include $p_0$, the symmetry of disagreement $S_D$, and the agreement on “having that trait” $P_{++}$.

This article takes that recommendation and computes it into a closed form under the specific scenario of three tiers, equal margins, noting its equivalence to “$\Sigma p^2$ is squeezed into a bandwidth by $m$.” Byrt et al.’s three-index decomposition (bias / prevalence / observed agreement, PMID 8501467) is another path to the same thing—they start from “kappa should be decomposed into three independent indices,” while this article starts from “the feasible region of $\Sigma p^2$ is an interval,” yet they converge on the same point. Feinstein and Cicchetti’s 1990 two paradoxes (PMID 2348207) are also not bypassed here: in the tables of Sections I and II, kappa moves from 0.69 to 0.84 then drops to negative values, while diagnostic capability remains unchanged—only $m$ is shifting.

V. Specific revisions to the three-tier verdict reporting standard

The triple-reporting scheme ($p_0$, $\Sigma p^2$, kappa) proposed in the previous article remains valid, but one additional requirement must be added: the bandwidth of $\Sigma p^2$ must be reported, or equivalently, $m$ must be reported. The reasons are:

  1. Reporting only the point value of $\Sigma p^2$ leaves readers unable to know whether this point is fixed or floating. When $m = 0.90$, a precision of $\pm 0.0025$ is sufficient; when $m = 0.20$, the same precision is meaningless.
  2. The bandwidth itself serves as a diagnostic measure. When bandwidth > 0.05 (roughly $m < 0.45$), any cross-tier comparison of kappa is invalid; comparisons can only be made within the same $m$ interval.
  3. Since bandwidth narrows monotonically with $m$, the fact that “$m$ is large” is both bad news (kappa tends toward negative) and good news (readings are certain)—these two aspects must be stated together; reporting only one is insufficient.

Evidence and Limitations

Verified:
– The closed form $\Sigma p^2 \in [m^2 + (1−m)^2/2, m^2 + (1−m)^2]$ under equal-margin conditions is directly derived from “how the two non-main tiers are allocated”; maximum deviation from the 9-grid exhaustive search is 0.0013 (at $m = 0.20$), and $\leq 0.0007$ for $m \geq 0.5$.
– Bandwidth narrows monotonically with $m$: as $m$ increases from 0.20 to 0.95, the $\Sigma p^2$ bandwidth drops from 0.320 to 0.0013, and the kappa bandwidth drops from 0.156 to 0.014.
– $m$ and $q$ are not bound: random parameter sweep of 400,000 groups shows that when $m \approx 0.30$ is locked, the span of $q$ is 0.979 (0.012–0.991); when $m \approx 0.70$ is locked, the span is 0.335. The “coupling of $m$ and $q$” feared in the previous article does not hold.
– “$p_0$ constant while kappa changes dramatically” can be fabricated by parameter symmetry: two sets of parameters (three-tier correct rates all 0.90, sum of squares of three rows equal 0.8182 vs. main tier 0.94/non-main tiers 0.82) yield $p_0$ ranges of 0.0029 and 0.176 respectively, representing two sources of the same phenomenon.

Partially verified:
– The closed form holds only under equal margins. When margins are unequal, the feasible region of $\Sigma p^2$ widens (adding a bias dimension), making the bandwidth larger and the closed form invalid. The next article will address this (#1383: contour lines of weighted kappa).
– The statement “conventional interpretation bands for kappa have width 0.20” relies on Landis-Koch’s grading scheme, which itself has long been criticized in literature for being arbitrary. This article uses it only as a scale reference, not as thresholds.
– The random sweep employed a single ordered three-tier model (adjacent misalignment rate $h$, cross-tier error rate $g$, non-main-tier correct rate $c$). More complex error structures (e.g., unequal adjudicator capabilities) were not covered.

Pending verification:
– This topic has not yet entered the L1 correspondence network: the three-domain node table of CORRESPONDENCE_NETWORK contains only single-point signal nodes (physical side “phase/rhythm | — | key of periodic coupling,” semantically meaning two oscillators phase-locked, without consistency/adjudication/margin semantics; physiological side “respiratory rate/respiratory pattern | wearable,” “qi-sensation report | structured self-assessment”), and a full-table search for “criteria / verdicts / consistency rate / adjudication / reliability / margins / inter-observer” yields zero hits; the edge table stops at E-050 (Na Yin ↔ frequency, L1, open, original text states “no measurable source yet”); the network alignment indicator row reads “edges above L2 | 0.” Hence this article caps at C, and all conclusions serve only as methodological discussion, not established mechanisms.
– A real-world test is needed: the same batch of $\geq 60$ subjects, 3 adjudicators independently performing three-tier adjudication, to measure the true four-point $(p_0, m, \Sigma p^2, \kappa)$. Only then can we verify whether this article’s bandwidth covers the measured points. If the measured $\Sigma p^2$ falls outside the bandwidth, it indicates unequal margins between adjudicators, requiring the closed form to add a bias term.
– The feasible region of $\Sigma p^2$ under unequal margins has not yet yielded a closed form; we only know it is wider than the equal-margin case, with the extent of widening depending on the magnitude of bias.

Related Reading

Evidence level: C
References: PMID 2348207, PMID 8501467, PMID 8621993
L1 network status: The three-domain node table of CORRESPONDENCE_NETWORK.md registers only single-point signals (physical side “phase/rhythm | — | key of periodic coupling,” semantically meaning two oscillators phase-locked, without consistency/adjudication/margin semantics; physiological side “respiratory rate/respiratory pattern | wearable,” “qi-sensation report | structured self-assessment”), and a full-table search for “criteria / verdicts / consistency rate / adjudication / reliability / margins / inter-observer” yields zero hits; the edge table stops at E-050 (Na Yin ↔ frequency, L1, open, original text states “no measurable source yet”); the network alignment indicator row reads “edges above L2 | 0.” → This edge lacks corresponding load-bearing edges in L1, so this article caps 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 *