带宽不到 0.05 也不代表能比:可比性取决于「带宽 + 2.77 倍标准误」,不是带宽本身

Written by

in

Language

Evidence level: C

「带宽超过 0.05 就不许跨档比较」这条线,是在边际相等的假设下拍的。把它和样本量放在一起算,会发现它的真身是「带宽 + 2.77 倍标准误」——而在真实的三档裁决里,抽样误差比带宽大十倍,带宽这一项根本没资格当家。

正文

一、这条线连自己指哪个量都没说清

先说一个不该出现的硬伤。0.05 这条线在两篇前作里被反复引用,但「带宽」指的是哪个量,前后并不一致。

1382 说的是 kappa 的带宽——固定 (p₀, m) 之后 kappa 的可行区间宽度。复算下来,它等于 0.05 的位置在 m ≈ 0.434(p₀ = 0.90 时)。

1384 说的是 Σp² 的带宽——边际集中度可行域的宽度 (1−m)²/2,它等于 0.05 的位置在 m < 1−√0.1 ≈ 0.684。

两个口径差 0.25 个主档占比。也就是说,同一份数据,按一个口径判「可以比」,按另一个口径判「不许比」。一条判据在生效前先要决定自己量的是什么,这是它最该先解决的问题。 本文统一采用 kappa 口径——因为「两个 kappa 能不能比」问的是 kappa 的差,不是 Σp² 的差。

二、0.05 想防的到底是什么

把它翻译成人话:如果一份读数本身允许在 0.05 范围内浮动,那么两份读数差 0.03 就不是真差。这个直觉是对的。但它漏了第二件事:读数还带抽样误差。

固定边际、只算抽样误差(Delta 法,SE = √[p₀(1−p₀)/(n(1−Σp²)²)]),我用蒙特卡洛四千次核对过,两者比值 0.97–1.10,公式靠得住。于是两条 kappa 之差要显著,需要

D ≥ B + 2.772 · SE

其中 B 是带宽、2.772 = 1.96×√2 是两条独立区间之差的标准误倍数。右边那一整块,才是「可辨差下限」D_min。 0.05 线只盯住了 B,把 SE 当不存在。

三、真实区间里,SE 比 B 大十倍

取真实的三档裁决场景:p₀ = 0.90,主档占比 m = 0.85–0.90。此时 B 只有 0.0166–0.0150,占一个 Landis-Koch 档位(0.20)的 8.3%–7.5%——带宽本身其实很乖。

但 n = 60 时 SE = 0.15–0.21,是 B 的 9–14 倍。合成之后:

m B D_min (n=60) 折合档位 D_min (n=200) 折合档位
0.60 0.0298 0.236 1.18 0.143 0.71
0.85 0.0166 0.428 2.14 0.242 1.21
0.90 0.0150 0.603 3.02 0.337 1.69
0.95 0.0068 0.822 4.11 0.454 2.27

m 越大,B 越小,可辨差反而越差。 这不是笔误:m 大时 Σp² 接近 1,分母 (1−Σp²)² 把 SE 炸开。于是那个「m 越大读数越确定」的好消息(#1382 的核心发现),在跨样本比较这个用途上完全反转——m 越大,你越没有资格拿两个 kappa 去比。

四、临界比不是 1.5,是 1.4 到 58

按原设想,「带宽须小于差值的 2/3」(临界比 R = D_crit/B ≈ 1.5)该是个稳定常数,能替代 0.05。实测在 (m, p₀, n) 网格上,R 从 1.4 走到 57.9,跨两个数量级。

R = 1.5 这条线要达成,需要 B 远大于 SE。反解:

m B 达成 R=1.5 所需 n
0.20 0.1563 491
0.40 0.0568 2,630
0.60 0.0298 11,517
0.90 0.0150 约 36.7 万

在最需要它的区间,R=1.5 是个永远达不到的条件。 所以「带宽 < 差值×0.67」这个方案不但不能替代 0.05,它连自己都执行不了。

五、但 B 不会被样本量抹掉

如果只是「样本量不够」,那加人就完了。这里有个不该忽略的结构:

D_min(n) = B + 2.772·SE(n),n → ∞ 时 收敛到 B,不是 0。

B 是不可约下限,不是精度损失。 它来自「同一份 kappa 读数在边际允许范围内本就可能是另一个值」这件事,跟你有多少人无关。想靠扩样本把可辨差压到 B 以下,方向就错了。这一条比「带宽多宽」重要得多:它说明 0.05 这条线抓住的东西是对的(结构性不可辨),只是量错了位置(该量 D_min 而非 B)。

六、反过来的灾难:0.05 线在真实区间是「太松」

最后一击,方向也反。要让 D_min ≤ 0.05(即 0.05 线真能保证可比),p₀ = 0.90 时需要:

m B 所需 n
0.70 0.0230 4,860
0.85 0.0166 9,109
0.90 0.0150 16,963

真实研究 n = 60–200。此时「B < 0.05 → 允许比」放行的是 D_min = 0.34–0.60 的组合——差三个档位都辨不出来,却被判为合规。 所以 0.05 这条线不是太严,是在最常见的那一档上太松,松到失去意义。

七、该报什么

单一阈值作废,改成合成量。凡跨样本、跨档位比较两个 kappa,报三个数:

B(该 (p₀, m) 下的 kappa 带宽)
SE(该 n 下的 Delta 法标准误)
D_min = B + 2.772·SE

判据只有一条:实测差 ≥ D_min 才谈可辨,否则一律写「不可辨」,不论 B 有多小。 若研究要事先规划,反过来用第二节那条不等式解 n——但记住目标不是「B < 0.05」,是「D_min < 你关心的最小效应」。

顺带一个校准参考:2022 年 82 项一致性研究的系统综述里,kappa 类研究样本量中位数只有 71(PMID 36123642),而 Bujang 等人 2024 年给出的 kappa 可靠性最小样本量是 15(PMID 38449496)。这两个数都远低于本表任何一档——按本文口径,n = 15 时 D_min 会到 0.84–1.19,跨整条 kappa 量程。样本量规划的口径若不与可辨差挂钩,”最小样本 15″ 这类建议在比较用途上等于零。

证据与局限

已验证:B 的闭式与 #1382 原表逐格复算一致(m = 0.20→0.95 八格,偏差 <0.0007);Delta 法 SE 经四千次蒙特卡洛核对,实测/理论比 0.97–1.10;D_min = B + 2.772·SE 的不等式由两条独立 95% 区间不重叠直接推出。

部分验证:0.05 线两个口径(m ≈ 0.434 与 m ≈ 0.684)确实在 #1382 与 #1384 各自成立,但「哪个才是本意」无法从文本判定,本文按 kappa 口径统一处理。

待验证:①「档位宽度」这个约束本文只取了 Landis-Koch 的 0.20,若改用更细档位则 D_min 的档位折算全部改变,而 0.20 本身被长期批评为任意(PMID 25168681 显示不同系数对同一差值分辨力不同);② 本文假设两条 kappa 独立,同一批受试者上的两次判读实为相依,2.772 会偏保守,真实门槛应更低——这个修正量从未被算过;③ 全部推导压在「边际相等」上,δ 一进来(#1384)天花板先塌,D_min 本身也要重写。

相关阅读

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/可比性/抽样误差语义;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 / 偏倚 / 带宽 / 可比 / 样本量 / 标准误」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」。→ 该主题尚未进入 L1 对应网络,本文封顶 C。

证据等级: C
参考文献: PMID 36123642, PMID 38449496, PMID 25168681, PMID 8621993, PMID 25088752


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。

Bandwidth under 0.05 does not imply comparability: comparability depends on ‘bandwidth plus 2.77 times the standard error’, not bandwidth itself

Evidence level: C

The threshold “bandwidth exceeding 0.05 forbids cross-tier comparison” was derived under the assumption of equal marginals. When calculated alongside sample size, its true form is “bandwidth + 2.77 times the standard error.” In real three-tier arbitration scenarios, sampling error is ten times larger than bandwidth; the bandwidth term alone lacks authority to decide.

Body

I. This Line Does Not Even Define Its Own Quantity

First, a glaring error that should not exist. The 0.05 line is repeatedly cited in two prior works, yet the quantity referred to as “bandwidth” is inconsistent between them.

1382 discusses the kappa bandwidth—the width of the feasible interval for kappa after fixing (p₀, m). Recalculated, the position where it equals 0.05 occurs at m ≈ 0.434 (when p₀ = 0.90).

1384 discusses the Σp² bandwidth—the width of the feasible domain for marginal concentration, given by (1−m)²/2. The position where this equals 0.05 occurs at m < 1−√0.1 ≈ 0.684.

These two definitions differ by 0.25 of the main tier proportion. That is, the same dataset may be judged “comparable” under one definition and “incomparable” under the other. A criterion must first decide what quantity it measures before it can take effect; this is the most fundamental issue it must resolve. This paper uniformly adopts the kappa definition—because “whether two kappas can be compared” asks about the difference in kappa, not the difference in Σp².

II. What Exactly Is 0.05 Trying to Prevent?

Translating this into plain terms: if a single reading itself allows fluctuation within 0.05, then a difference of 0.03 between two readings is not a real difference. This intuition is correct. However, it overlooks a second factor: readings also carry sampling error.

Fixing the marginal and calculating only the sampling error (Delta method, SE = √[p₀(1−p₀)/(n(1−Σp²)²)]), I verified this via Monte Carlo simulation with 4,000 runs. The ratio of the two terms ranges from 0.97 to 1.10, confirming the formula’s reliability. Thus, for two kappa values to show a significant difference, we require

D ≥ B + 2.772 · SE

where B is the bandwidth, and 2.772 = 1.96×√2 is the multiplier for the standard error of the difference between two independent intervals. This entire right-hand side constitutes the “minimum detectable difference” D_min. The 0.05 line only focuses on B, treating SE as nonexistent.

III. In Real Intervals, SE Is Ten Times Larger Than B

Consider the realistic three-tier arbitration scenario: p₀ = 0.90, main tier proportion m = 0.85–0.90. Here B ranges only from 0.0166 to 0.0150, representing 8.3%–7.5% of a Landis-Koch tier width (0.20)—the bandwidth itself is actually quite well-behaved.

However, when n = 60, SE ranges from 0.15 to 0.21, which is 9–14 times larger than B. Combining these:

m B D_min (n=60) Equivalent Tier D_min (n=200) Equivalent Tier
0.60 0.0298 0.236 1.18 0.143 0.71
0.85 0.0166 0.428 2.14 0.242 1.21
0.90 0.0150 0.603 3.02 0.337 1.69
0.95 0.0068 0.822 4.11 0.454 2.27

As m increases, B decreases, yet the minimum detectable difference becomes worse. This is not a typo: when m is large, Σp² approaches 1, causing the denominator (1−Σp²)² in the SE formula to explode. Thus, the good news from #1382—that “larger m means more certain readings”—completely reverses when applied to cross-sample comparison: the larger m is, the less qualified you are to compare two kappas.

IV. The Critical Ratio Is Not 1.5, but Ranges from 1.4 to 57.9

According to the original design, “bandwidth must be less than two-thirds of the difference” (critical ratio R = D_crit/B ≈ 1.5) was supposed to be a stable constant, replacing 0.05. However, testing across the (m, p₀, n) grid reveals that R varies from 1.4 to 57.9—spanning two orders of magnitude.

For R = 1.5 to hold, B must be much larger than SE. Solving backwards:

m B Sample size n required for R=1.5
0.20 0.1563 491
0.40 0.0568 2,630
0.60 0.0298 11,517
0.90 0.0150 approximately 367,000

In the region where it is most needed, R=1.5 is an impossible condition. Therefore, the proposal “bandwidth < 0.67 × difference” not only fails to replace 0.05; it cannot even be implemented consistently.

V. Yet B Cannot Be Erased by Sample Size

If the issue were merely “insufficient sample size,” adding more subjects would solve it. However, there is a structural element that cannot be ignored:

D_min(n) = B + 2.772·SE(n), and as n → ∞, this converges to B, not zero.

B is an irreducible lower bound, not a loss of precision. It arises from the fact that “the same kappa reading, within allowable marginal variation, could inherently correspond to another value,” regardless of how many subjects you have. Attempting to suppress the detectable difference below B by expanding the sample size is going in the wrong direction. This insight is far more important than “how wide the bandwidth is”: it shows that the 0.05 line captures the right structural concept (structural undetectability) but measures it at the wrong location (should measure D_min, not B).

VI. The Reverse Disaster: 0.05 Is “Too Loose” in Real Intervals

The final blow reverses the direction entirely. To achieve D_min ≤ 0.05 (so that the 0.05 line truly may support comparability), with p₀ = 0.90, we require:

m B Required n
0.70 0.0230 4,860
0.85 0.0166 9,109
0.90 0.0150 16,963

Real studies typically have n = 60–200. Under these conditions, allowing “B < 0.05 → comparable” actually permits D_min = 0.34–0.60 combinations—differences spanning three tiers remain undetectable yet are judged compliant. Thus, the 0.05 line is not too strict; it is too loose on the most common tier, becoming meaningless.

VII. What Should Be Reported

The single threshold is abolished; replace it with composite metrics. For any cross-sample, cross-tier comparison of two kappas, report three numbers:

B (kappa bandwidth under this (p₀, m))
SE (Delta-method standard error under this n)
D_min = B + 2.772·SE

The criterion is singular: only discuss detectability if the observed difference ≥ D_min; otherwise, write “undetectable,” regardless of how small B is. If planning studies in advance, use the inequality from Section II to solve for n—but remember the goal is not “B < 0.05,” but “D_min < your minimum effect of interest.”

As a calibration reference: in the 2022 systematic review of 82 consistency studies, the median sample size for kappa-type studies is only 71 (PMID 36123642), while Bujang et al. (2024) report a minimum kappa reliability sample size of 15 (PMID 38449496). Both numbers are far below any tier in this table. Under this paper’s definition, with n = 15, D_min reaches 0.84–1.19, spanning the entire kappa range. If sample size planning is not linked to detectable difference, recommendations like “minimum sample 15” are effectively zero for comparative purposes.

Evidence and Limitations

Verified: B’s closed-form matches the full-grid recalculation from #1382 (eight tiers from m = 0.20 to 0.95, deviation <0.0007); Delta-method SE confirmed via 4,000 Monte Carlo runs, observed/theoretical ratio 0.97–1.10; the inequality D_min = B + 2.772·SE directly follows from non-overlapping 95% confidence intervals of two independent estimates.

Partially verified: The two definitions of the 0.05 line (m ≈ 0.434 and m ≈ 0.684) indeed hold in #1382 and #1384 respectively, but “which was the intended meaning” cannot be determined from the text; this paper uniformly adopts the kappa definition.

Pending verification: ① “Tier width” constraint: this paper only uses Landis-Koch’s 0.20; adopting finer tiers would alter all tier-equivalent D_min values, while 0.20 itself has long been criticized as arbitrary (PMID 25168681 shows different coefficients yield varying resolution for the same difference); ② this paper assumes two kappas are independent, but repeated readings on the same subjects are inherently dependent, making 2.772 conservative; the true threshold should be lower—this correction has never been calculated; ③ all derivations rest on “equal marginals”; introducing δ (#1384) collapses the ceiling first, requiring D_min itself to be rewritten.

Related Reading

L1 Network Status: The three-domain node table in CORRESPONDENCE_NETWORK.md registers only single-point signals (physical side “phase / rhythm | — | key for periodic coupling”, semantics of two oscillators phase-locked, without consistency/interpretation/comparability/sampling error semantics; physiological side “breathing rate / breathing pattern | wearable”, “qi sensation report | structured self-assessment”). A full search for “criterion / verdict / consistency rate / interpretation / reliability / marginal / inter-observer / bias / bandwidth / comparable / sample size / standard error” yields zero hits; the edge table stops at E-050 (Na Yin ↔ frequency, L1, open, original evidence column states “no measurable source yet”); network alignment indicator row reads “L2+ edges | 0”. → This topic has not yet entered the corresponding L1 network; this paper caps at evidence level C.

Evidence Level: C
References: PMID 36123642, PMID 38449496, PMID 25168681, PMID 8621993, PMID 25088752


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 *