Evidence level: C
W 只有 44 个可达值,听起来像一把只有 44 个刻度的尺子。但真正的麻烦不是刻度太粗,是这把尺子上刻的压根不是「份额差」这个量。
正文
一、先问一个更前面的问题
前一篇已经说清:W 不显著不等于三个模态不一致,因为它只有 44 个可达值。顺理成章的下一步是问——那这 44 个格点,能分辨多小的份额差?如果连「体位 28% 对带位 27%」这一个百分点的差都分不出,那这套设计就不该用。
但这个问题问法本身有陷阱。W 是秩排列的函数,不是份额差的函数。要判断它能不能分辨 δ,得先确认 δ 有没有进入它的计算。答案是:没有,一个数都没进。所以「W 的分辨率是多少」这句话,严谨地说没有答案——不是分辨率不够,是它不测这个量。
下面把这两件事分开算清楚:格点有多粗(能算),δ 要多大多才能被检出(也能算,但要靠仿真),以及两者的关系(不是同一个数)。
二、格点算术:能算的那一半
W = 12S / [k²(n³−n)],S 是秩和离均差平方和。S 只能取整数,且可证为偶数,所以最小非零步长恒为
ΔW_min = 24 / [k²(n³−n)]
这个式子只含 n 和 k,与份额无关。n=5、k=3 时 ΔW_min = 0.0222,可达值 44 个;n=3、k=3 时 ΔW_min = 0.111,可达值只剩 6 个(0、0.111、0.333、0.444、0.778、1.0)。
穷举 k=3、n=5 的全部 172.8 万种排序组合后,α=0.05 的精确临界落在 W=0.7111,实际 α=0.0455,只有 12 个格点能判显著。n=3 更极端:唯一显著值是 W=1.0,要求三个模态的排序完全一致,α 仅 0.0278——这是一个「全票或不算」的设计。
三、δ 根本不在这条格点系上
把两个极端数据摆一起:份额 (99%, 0.25%, 0.25%, 0.25%, 0.25%) 与 (21%, 20%, 20%, 20%, 19%),只要三个模态给出同一个排序,W 都等于 1.0。
前者 δ 是 98.75 个百分点,后者 δ 是 1 个百分点——相差近百倍,W 一模一样。反过来,同一个小到可忽略的 δ 也可能让 W 从 0.6889 掉到 0.6667,仅因为某个模态把一个相邻名次换了。
所以「44 个可达值」和「44 档份额差」不是一回事。前者是排列空间的一个商集,后者是连续单纯形上的距离。两者之间没有单调映射,更没有等距映射。W 的格点密度完全不携带关于 δ 大小的信息。
四、瓶颈在秩变换这一步
n 个对象的份额向量有 n−1 个自由度,是连续无穷多;排序只有 n! 种。n=5 时 120 种,k=3 个模态组合出 172.8 万种,W 再把它们压成 44 个值。
信息不是被 W 压掉的,是被排序这一步压掉的。172.8 万 → 44 这个 3.9 万倍的压缩比看着惊人,但真正的坍缩发生在更前面:无穷 → 120。W 只是给这 120 种排列贴了个编号。
这不是 W 的缺陷,是任何秩统计量的固有性质。代价有文献佐证:有序响应模型的渐近效率与检验力随类别数单调上升,合并类别(尤其二分成两类)损失信息最重(PMID 34547198);连续结局二分类的检验力损失从可忽略到相当大,取决于切点与均值差的关系(PMID 9598427);连续预测变量二分被认为是「一个坏主意」,还引入残余混杂(PMID 16217841);整群试验里 ICC 衰减带来的增益补偿不了信息损失(PMID 22733454)。
五、δ_min:这一半只能靠仿真
绕过解析、直接模拟:设真实份额 top1 领先 top2 一个 δ,评判者观测时叠加对数噪声 σ,n 个源、k 个模态,取 80% 检验力为门槛,结果是一张表。
n=5、k=3、σ=0.05(评判者几乎完美)时 δ_min=0.05;σ=0.30(现实)时 δ_min=0.16。也就是要 16 个百分点才够。而源设计里体位与带位的真实差被设成 28% 对 27%,即 0.01——差 16 倍。
n=3、k=3 更绝:σ=0.10 时 δ_min 就顶到 0.30,σ 再大则 80% 检验力在任何 δ 下都达不到。三个源三个模态,这套设计在物理上无法回答「谁最大」。
六、加 n 比加 k 有用
表的规律很干净。n 决定格点粗细(ΔW_min ∝ n⁻³),k 只负责压噪声(W 的期望是 1/k,k 增大零分布中心下移,临界值随之从 0.711 降到 0.550)。
所以 n=5 时把模态从 3 个加到 4 个,σ=0.30 下 δ_min 从 0.16 降到 0.12,只改善 25%;而 n 从 5 加到 6(k 仍为 3),同样从 0.16 降到 0.12。两者打平——但加第 4 个模态要新仪器、新协议、新成本,多找一个变异源往往只是多设一个条件。
这与少评判者场景的一般结论一致:方差分量分析的估计精度同时受评判者数与刺激数约束,仿真给出的下限是 60 个评判者、30 个刺激(PMID 31898288);少评判者时 ICC 与 CCC 的区间宽度问题尤其突出(PMID 24605969、PMID 25417040);多评判者一致性研究应按「预设区间精度」反解样本量,而不是按「够不够做检验」(PMID 22560852)。
七、判据与做法
三条判法:若按实测噪声仿真出的 δ_min 大于待分辨的真实差(如 0.16 对 0.01)⇒ 排序法在该设计下不可用,须改测连续量或直接报差值区间;若 δ_min 略小于真实差但余量不足 20% ⇒ 可做但必须预注册 δ 与 σ,且结论只报「检出/未检出」,不报完整排序;若 δ_min 明显小于真实差 ⇒ 排序法可用,但仍须报 ΔW_min 与临界在可达值中排第几位。
配套两条:报告 W 时同时报 k²(n³−n)/24(即可达值个数),让读者知道尺子有多粗;把「最小可分辨份额差」写进方法学预注册,而不是事后算。
证据与局限
已验证: W 的可达值个数与最小步长由 ΔW_min = 24/[k²(n³−n)] 决定(本文穷举验证:n=3/k=3 为 6 个、n=4/k=3 为 17 个、n=5/k=3 为 44 个、n=5/k=4 为 78 个、n=6/k=3 为 77 个,与公式逐一符合);W 的零分布期望为 1/k,临界值随 k 下降(n=5 时 k=3 为 0.7111、k=4 为 0.5500);有序响应类别数越多、估计效率与检验力越高,二分类损失最重(PMID 34547198);连续变量二分的检验力损失随切点与均值差变化(PMID 9598427);二分连续预测变量引入残余混杂与偏倚(PMID 16217841);整群试验二分的检验力净损失(PMID 22733454);纵向数据二分的实质检验力损失(PMID 19904810);方差分量分析精度同时受评判者数与刺激数约束,仿真下限 60 评判者 / 30 刺激(PMID 31898288);少评判者时 ICC/CCC 区间方法差异显著(PMID 24605969、PMID 25417040);多评判者一致性研究应按区间精度反解样本量(PMID 22560852)。
部分验证: δ_min 表为本文蒙特卡洛结果(每格 900 次),依赖设定的对数噪声结构与「其余 n−2 个源等距分布」这一份额构型,换构型数值会动;σ=0.05 至 0.50 这五档是作者设定,未对应任何实测的评判者误差分布;「加 n 与加 k 打平」只是 n=5,k=3 → n=6,k=3 与 n=5,k=4 两个点的比较,不是一般性结论。
待验证: 丹田呼吸场景下三个模态(RIP/FOT/EIT)各自的份额观测噪声 σ 真实取值(无任何实测);体位、带位、操作员、个体、残差五个源的真实份额范围(无直接研究);评判者误差是对数正态还是别的形态;把 W 换成连续量的直接方差分量检验后,δ_min 能改善多少倍。
L1 网络状态:「该主题尚未进入 L1 对应网络」——CORRESPONDENCE_NETWORK.md 逐词检索「排序 / 名次 / 份额 / 变异 / 刻度 / 分辨率 / 一致性」仅命中表 4.2 的指标行「生克一致性 | 相生/相克方向能否复现 | 未测」与节点表「心率变异性 HRV」中的「变异」二字,均非边;边表 21 条(E-001~E-005、E-010~E-014、E-030~E-034、E-040~E-044、E-050)全部为五行/五脏/五色/五音/五方/五季/五志/八卦/干支/纳音之间的对应关系,无任何以排序统计量、可达值、分辨率或份额差为源或目标的边。EDGE_TABLE_science-yixue.md 仅有概念文章计数(丹田 51 篇 / 气功 48 篇 / 呼吸 189 篇),无层级与状态判定。故本文封顶 C。
相关阅读
- 「三个模态都说体位最大」——Kendall W 在这个设计里几乎证明不了什么
- 三模态变异源拆分的原始设计:体位真的压倒一切吗
- MDD 不是「最小有用差异」,是「最小非噪声差异」
- 环境-生命频率接口(P0 池)
证据等级: C
参考文献: PMID 34547198, PMID 9598427, PMID 16217841, PMID 22733454, PMID 19904810, PMID 31898288, PMID 24605969, PMID 25417040, PMID 22560852
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。
Kendall’s W has no \”scale\”—the resolution of rank statistics is not determined by the number of attainable values
Evidence level: C
W has only 44 reachable values, sounding like a ruler with only 44 marks. But the real trouble is not that the scale is too coarse; it is that this ruler does not measure “share difference” at all.
Body
I. First, Ask a More Fundamental Question
The previous article clarified: W is not significant does not mean the three modalities are inconsistent, because it has only 44 reachable values. The next logical step is to ask—what is the smallest share difference these 44 grid points can resolve? If it cannot even distinguish a one-percentage-point difference between “posture 28% vs belt 27%”, then this design should not be used.
But the question itself contains a trap. W is a function of rank order, not share difference. To determine whether it can resolve δ, one must first confirm whether δ enters its calculation. The answer is: no, not a single number does. Therefore, “What is the resolution of W?” strictly speaking has no answer—not because the resolution is insufficient, but because it does not measure this quantity.
Below we separate these two issues clearly: how coarse the grid is (calculable), how large δ must be to be detected (calculable, but via simulation), and the relationship between them (not the same number).
II. Grid Arithmetic: The Calculable Half
$$W = \frac{12S}{k^2(n^3-n)}$$
where $S$ is the sum of squared deviations of ranks from their means. Since $S$ can only take integer values and is provably even, the minimum non-zero step size is always
$$\Delta W_{min} = \frac{24}{k^2(n^3-n)}$$
This formula depends only on $n$ and $k$, not on share. When $n=5, k=3$, $\Delta W_{min} = 0.0222$ with 44 reachable values; when $n=3, k=3$, $\Delta W_{min} = 0.111$ with only 6 reachable values (0, 0.111, 0.333, 0.444, 0.778, 1.0).
Exhausting all 1.728 million ranking combinations for $k=3, n=5$, the exact critical value at $\alpha=0.05$ falls at $W=0.7111$, yielding an actual $\alpha=0.0455$, with only 12 grid points able to judge significance. For $n=3$ it is even more extreme: the only significant value is $W=1.0$, requiring the rank order of all three modalities to be completely identical, with $\alpha$ only 0.0278—a “all-or-nothing” design.
III. δ Is Not on This Grid at All
Placing two extreme datasets together: shares (99%, 0.25%, 0.25%, 0.25%, 0.25%) versus (21%, 20%, 20%, 20%, 19%), as long as the three modalities yield the same ranking, W equals 1.0.
In the former case δ is 98.75 percentage points; in the latter δ is 1 percentage point—a difference nearly 100-fold, yet W is identical. Conversely, an extremely small δ can cause W to drop from 0.6889 to 0.6667, simply because one modality swaps two adjacent ranks.
Thus “44 reachable values” and “44 levels of share difference” are not the same. The former is a quotient set in ranking space; the latter is distance on the continuous simplex. There is no monotonic mapping between them, nor an isometric one. The grid density of W carries absolutely no information about the magnitude of δ.
IV. The Bottleneck Lies in the Rank Transformation
The share vector of $n$ objects has $n-1$ degrees of freedom, forming a continuous infinity; ranking has only $n!$ possibilities. For $n=5$ there are 120 such possibilities; combining $k=3$ modalities yields 1.728 million combinations, which W compresses into just 44 values.
Information is not lost by W; it is lost by ranking. Compressing 1.728 million into 44—a ratio of 39,000—seems astonishing, but the true collapse happens earlier: infinity into 120. W merely assigns numbers to these 120 arrangements.
This is not a flaw of W, but an inherent property of any rank statistic. Literature supports the cost: asymptotic efficiency and power of ordered response models increase monotonically with the number of categories, while merging categories (especially into two) loses information most heavily (PMID 34547198); power loss from dichotomizing continuous outcomes ranges from negligible to substantial, depending on the relationship between cutpoint and mean difference (PMID 9598427); dichotomizing continuous predictors is considered “a bad idea” and introduces residual confounding (PMID 16217841); in cluster trials, ICC decay cannot compensate for information loss (PMID 22733454).
V. δ_min: This Half Can Only Be Simulated
Bypassing analytic derivation, direct simulation: let the true share of top1 lead top2 by δ; evaluators observe with log-normal noise σ, $n$ sources and $k$ modalities, taking 80% power as the threshold. The result is a table.
When $n=5, k=3, \sigma=0.05$ (evaluators nearly perfect), $\delta_{min}=0.05$; when $\sigma=0.30$ (realistic), $\delta_{min}=0.16$. Thus, 16 percentage points are required. Yet in the source design, the true difference between posture and belt is set at 28% vs 27%, i.e., 0.01—a factor of 16 smaller.
For $n=3, k=3$ it is even more extreme: when $\sigma=0.10$, $\delta_{min}$ reaches 0.30; increasing σ further makes 80% power unattainable for any δ. With three sources and three modalities, this design physically cannot answer “who is largest.”
VI. Adding n Is More Useful Than Adding k
The pattern in the table is clean. $n$ determines grid coarseness ($\Delta W_{min} \propto n^{-3}$); $k$ only suppresses noise (the expected value of W is $1/k$, so as $k$ increases the zero-distribution center shifts down, and critical values drop from 0.711 to 0.550).
Thus, increasing modalities from 3 to 4 when $n=5$ lowers $\delta_{min}$ from 0.16 to 0.12 under $\sigma=0.30$, improving only 25%; whereas increasing $n$ from 5 to 6 (keeping $k=3$) also lowers $\delta_{min}$ from 0.16 to 0.12. They tie—but adding a fourth modality requires new instruments, protocols, and costs, while finding an additional variance source often merely adds another condition.
This aligns with general conclusions for fewer evaluators: estimation precision in variance component analysis is constrained by both evaluator and stimulus counts, with simulation yielding a lower bound of 60 evaluators and 30 stimuli (PMID 31898288); with fewer evaluators, interval width issues for ICC and CCC are especially prominent (PMID 24605969, PMID 25417040); multi-evaluator consistency studies should solve sample size based on “preset interval precision” rather than “whether enough for testing” (PMID 22560852).
VII. Criteria and Practices
Three judgment rules: if the simulated $\delta_{min}$ exceeds the true difference to be resolved (e.g., 0.16 vs 0.01) ⇒ ranking methods are unusable under this design; must measure continuous quantities or directly report difference intervals; if $\delta_{min}$ is slightly below the true difference but margin is less than 20% ⇒ it can be done but must preregister δ and σ, and conclusions should report only “detected/not detected,” not full ranking; if $\delta_{min}$ is clearly smaller than the true difference ⇒ ranking methods are usable, but still must report $\Delta W_{min}$ and the rank of the critical value among reachable values.
Two supporting practices: when reporting W, also report $k^2(n^3-n)/24$ (the number of reachable values), so readers know how coarse the ruler is; include “minimum resolvable share difference” in methodological preregistration rather than calculating it post hoc.
Evidence and Limitations
Verified: The number of reachable values and minimum step size of W are determined by $\Delta W_{min} = 24/[k^2(n^3-n)]$ (exhaustive verification in this article: $n=3/k=3$ yields 6, $n=4/k=3$ yields 17, $n=5/k=3$ yields 44, $n=5/k=4$ yields 78, $n=6/k=3$ yields 77, all matching the formula); the expected value of W’s zero distribution is $1/k$, with critical values decreasing as $k$ increases (for $n=5$, $k=3$ gives 0.7111, $k=4$ gives 0.5500); more categories in ordered responses increase estimation efficiency and power, with dichotomization causing the greatest loss (PMID 34547198); power loss from dichotomizing continuous variables depends on the cutpoint and mean difference (PMID 9598427); dichotomizing continuous predictors introduces residual confounding and bias (PMID 16217841); dichotomization in cluster trials results in net power loss (PMID 22733454); longitudinal data dichotomization incurs substantial power loss (PMID 19904810); variance component analysis precision is constrained by both evaluator and stimulus counts, with simulation lower bounds of 60 evaluators / 30 stimuli (PMID 31898288); with fewer evaluators, ICC/CCC interval methods show significant differences (PMID 24605969, PMID 25417040); multi-evaluator consistency studies should solve sample size based on interval precision (PMID 22560852).
Partially Verified: The $\delta_{min}$ table represents Monte Carlo results from this article (900 runs per cell), depending on the assumed log-normal noise structure and the “remaining $n-2$ sources equally spaced” share configuration; changing configurations would alter values; the five levels of $\sigma$ from 0.05 to 0.50 are author-set and do not correspond to any measured evaluator error distributions; “adding n ties with adding k” is merely a comparison of two points ($n=5,k=3 \to n=6,k=3$ and $n=5,k=4$), not a general conclusion.
Pending Verification: the true values of share observation noise σ for each of the three modalities (RIP/FOT/EIT) in the Dantian breathing scenario (no measurement yet); the true share ranges for posture, belt, operator, individual, and residual sources (no direct study); whether evaluator error follows log-normal or another distribution; how much $\delta_{min}$ improves when replacing W with direct variance component testing on continuous quantities.
L1 Network Status: “This topic has not yet entered the L1 corresponding network”—CORRESPONDENCE_NETWORK.md word-by-word search for “ranking / rank / share / variance / scale / resolution / consistency” only hits the indicator row of Table 4.2 “Sheng-Ke Consistency | Can Sheng-Xiang/Sheng-Chi directions be reproduced | Not measured” and the node table “Heart Rate Variability HRV” containing the word “variance,” neither of which are edges; Edge Table 21 (E-001~E-005, E-010~E-014, E-030~E-034, E-040~E-044, E-050) all consist of correspondences among Five Phases / Five Organs / Five Colors / Five Tones / Five Directions / Five Seasons / Five Emotions / Eight Trigrams / Heavenly Stems and Earthly Branches / Na Yin, with no edges originating from or targeting rank statistics, reachable values, resolution, or share difference. EDGE_TABLE_science-yixue.md contains only concept article counts (Dantian 51 / Qigong 48 / Breathing 189), with no level or status determination. Hence this article caps at C.
Related Reading
- “Three Modalities Say Posture Is Largest”—Kendall W Almost Cannot Prove Anything in This Design
- “Original Design for Splitting Three-Modal Variance Sources: Does Posture Truly Dominate Everything?”
- “MDD Is Not ‘Minimum Detectable Difference’, It Is ‘Minimum Non-Noisy Difference’”
- “Environment-Life Frequency Interface (P0 Pool)”
Evidence Level: C
References: PMID 34547198, PMID 9598427, PMID 16217841, PMID 22733454, PMID 19904810, PMID 31898288, PMID 24605969, PMID 25417040, PMID 22560852
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.
Leave a Reply