Evidence level: C
三个模态排序一致,只说明它们像;不说明它们对。而在这套设计里,连「像不像」都量不出来。
正文
一、先看这把尺子有几个刻度
检验「体位是否主导变异」的常见做法,是让 RIP、FOT、EIT 三个模态各自给五个变异源(体位、带位、操作员、个体、残差)排序,再算 Kendall 一致性系数 W。k=3 个评判者,n=5 个对象。
把这套设计的全部可能性穷举一遍——每个模态 5! = 120 种排序,三个模态合计 172.8 万种组合——能得到的 W 只有 44 个不同取值。α=0.05 的精确临界是 W ≥ 0.6889,44 个可达值里只有 13 个能判显著。
n 更小时更糟:若只比三个源(体位、操作员、残差),可达 W 只剩 6 个值——0、0.111、0.333、0.444、0.778、1.0,其中仅 2 个能显著。
含义很直接:W 不显著,不等于三个模态不一致。它可能只是这把尺子没有中间刻度。
二、W 显著,也不等于排序站得住
反过来的坑更深。W 只问「三个模态的排序像不像」,从不问「这个排序本身稳不稳」。
做个模拟:设真实份额为体位 28%、带位 27%(几乎并列),即使有效重复次数涨到 500,体位被误判成非第一的概率仍有 36%。若真实差拉大到 40% 对 20%,重复 30 次就能把误判率压到 7.4%。
W = 1.0 完全可以是三个模态一起在噪声里抖。 一致性(agreement)和正确性(accuracy)是两件事,前者可以完美,后者可以全错。
三、份额是成分数据,五个秩不独立
变异源占比必须加起来等于 100%,这叫闭合数据(closed data)。闭合约束会让各份额天然负相关:一个涨,别的必须跌。
有研究用垃圾组分数据演示过后果:同一批数据按绝对质量算,可避免蔬菜垃圾与塑料包装正相关;改用质量分数算,相关变成负的(PMID 28882426);算术均值甚至能给出负的比例,置信区间 (-4.03, 8.45)。
这直接打掉 W 的前提——W 的零分布假设 k 个排序彼此独立,而成分数据的五个秩被「总和恒定」绑在一起。正确做法是先做对数比变换(clr/ilr)再谈方差(PMID 38806627)。
四、水平数太少,方差分量自己在飘
体位只有 2 个水平(坐/卧),带位 2-3 个,操作员 2-3 个。水平少是所有麻烦的源头。
负的方差估计和边界估计在这种设计里很常见。有方法学工作专门讨论「伪方差」:某些数据的边际分布与「负方差随机截距斜率模型」在数学上完全等同,而测量次数少、随访短会扩大这片区域(PMID 38695394)。另有模拟比较各类估计量「出现负估计的概率」(PMID 26843971)。
检验「某方差分量为零」更麻烦:零落在参数空间边界上,似然比与得分统计量的常规卡方分布是错的,得靠置换法近似有限样本分布(PMID 22930674)。
而 REML 的抽样方差走大样本渐近——有研究要样本量到 1413 才认为渐近近似够用(PMID 35197554)。二十几个人的样本里,份额的抽样误差很可能大于要比较的那点差距。
五、「有体位效应」≠「体位是最大源」
体位影响呼吸的证据其实很扎实:系统综述 43 项研究,立位到卧位功能残气量显著下降(PMID 30305051);62 人 IOS 测试,卧位 R5、R5-R20 全升、X5 更负,肥胖组更明显(PMID 36787314);17 例 ARDS 患者坐起 30 分钟后,EIT 的 ROI 通气分布整体重排(PMID 39116511)。
但这些都只说明体位有效应,不说明体位是最大的源。中间隔着一个跨源头比大小,而这恰恰是最不稳的一步。
顺带一提:EIT 区域通气分布在固定通气设置下重复性极好(10 头猪,通气重心随 PEEP 从 45±3% 移到 49±3%,区间极窄,PMID 17664640)。重复性好只代表同一体位内稳,与跨源头比大小无关。
六、还有一个隐藏前提:异方差
G 理论经典假设各条件下残差方差齐性。生理数据里这个假设常不成立。有工作用位置-尺度混合模型扩展 G 理论,发现不建模异方差会掩盖真实的测量质量差异,须按条件分别计算 G 与 D 系数才看得出来(PMID 41549064)。多层级评分者信度同理——评判者数量少是 ICC 偏倚与低效率的主因(PMID 33818118)。
三个模态就是三个评判者,本来就少;再叠上异方差,「平均份额」这个数代表不了任何单一条件。
反转:该报的不是 W,是相邻差值的区间
整个设计的毛病是问错了问题。真正要答的是「体位这份额比第二名大多少,这个差有没有超出不确定度」,而不是「三个模态的排序像不像」。做法是把排序换成两两差值加 bootstrap 置信区间:只有 CI 不跨零的相邻对,才算分得出高低。W 一个这样的问题都答不了。
四条判法:三模态 W 均显著且相邻差 CI 全不跨零 ⇒ 排序可信,可谈体位主导;W 显著但相邻差 CI 跨零 ⇒ 只能报「体位高于其余」,不能报完整排序;W 不显著但某两源差 CI 稳定不跨零 ⇒ 是粒度问题不是不一致,改报差值;W 不显著且所有差 CI 都跨零 ⇒ 该数据支撑不了任何排序结论,「体位主导」退回假设。
配套三条:报 W 必须同时报可达值个数与临界值在其中的排位;成分数据先 ilr/clr 变换;样本量按「要分辨的最小份额差」反解,别按「够不够做 W」。
证据与局限
已验证: Kendall W 在 k=3、n=5 时的可达取值离散性与精确排列临界值(本文穷举计算);闭合数据(成分数据)会引入伪相关、算术均值可给出负比例(PMID 28882426);对数比变换处理成分数据的必要性(PMID 38806627);负方差与伪方差模型的成因,小测量次数与短随访扩大该区域(PMID 38695394);零方差检验在参数边界上常规卡方分布失效、需置换法(PMID 22930674);REML 方差分量抽样方差的渐近性对小样本可能有偏(PMID 35197554);体位对肺功能的影响(系统综述 43 项,PMID 30305051);IOS 卧位上阻抗升高、X5 更负(PMID 36787314);EIT 坐位改变通气分布(PMID 39116511);EIT 区域通气分布重复性(PMID 17664640);异方差下 G 理论需扩展(PMID 41549064);评判者数量少导致 ICC 偏倚(PMID 33818118)。
部分验证: 上述体位证据多来自机械通气患者、肥胖者或 ARDS 人群,健康修炼者在半卧/端坐下的体位效应幅度是否同量级未知;份额排序的翻转概率为本文蒙特卡洛模拟结果,依赖设定的真实份额与误差结构,非实测;成分数据在变异源拆分场景下的伪相关幅度未直接测量。
待验证: 丹田呼吸场景下五个变异源的真实份额范围(无任何直接研究);体位与带位是否存在交互(卧位腹带受压后张力变化是否简单可加);三模态各自的残差方差是否满足齐性;把 W 换成差值区间后,原「体位主导」结论是否仍成立。
L1 网络状态:「该主题尚未进入 L1 对应网络」——CORRESPONDENCE_NETWORK.md 节点表把「相位 / 节律」列为物理侧节点、「呼吸率 / 呼吸模式」列为生理侧观测项,均未与易学侧任何节点连成边;节点表中完全没有「变异源 / 方差分量 / 一致性系数 / 排序」这类测量学维度节点,边表 3.1–3.4 只登记五行↔五脏/五色/五音/五方/五季/五志、八卦↔五行、天干地支↔五行,无任何以变异源拆分、体位主导或 Kendall W 为源或目标的边。EDGE_TABLE_science-yixue.md 仅有概念计数(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
证据等级: C
参考文献: PMID 28882426, PMID 38806627, PMID 38695394, PMID 26843971, PMID 22930674, PMID 35197554, PMID 30305051, PMID 36787314, PMID 39116511, PMID 17664640, PMID 41549064, PMID 33818118
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。
\”Three Modalities Say Posture is Biggest\” — Kendall W Can’t Prove Much in This Design
Evidence level: C
All three modalities rank consistently; this only indicates similarity, not correctness. In this design, even the degree of “similarity” is unmeasurable.
Main Text
I. First, How Many Tick Marks Does This Ruler Have?
A common approach to testing whether “posture dominates variation” is to have RIP, FOT, and EIT each rank five sources of variation (posture, belt position, operator, individual, residual), then calculate Kendall’s coefficient of concordance W. With $k=3$ raters and $n=5$ objects.
Exhaustively enumerate all possibilities for this design—each modal has $5! = 120$ rankings, totaling over 1.7 million combinations for three modalities—only 44 distinct values of W can be obtained. The exact critical value at $\alpha=0.05$ is $W \ge 0.6889$. Of the 44 attainable values, only 13 are significant.
It is worse when $n$ is smaller: if comparing only three sources (posture, operator, residual), the attainable W values drop to just 6: 0, 0.111, 0.333, 0.444, 0.778, and 1.0. Only 2 of these are significant.
The implication is direct: Non-significant W does not mean the three modalities are inconsistent. It may simply mean this ruler lacks intermediate tick marks.
II. Significant W Does Not Mean the Ranking Is Stable
The reverse pitfall is deeper. W only asks “how similar are the three rankings,” never “is this ranking itself stable.”
Run a simulation: Assume true proportions are 28% for posture and 27% for belt position (almost tied). Even with effective replication increasing to 500, the probability of misjudging posture as non-first remains 36%. If the true difference widens to 40% vs. 20%, only 30 replications are needed to reduce the misjudgment rate to 7.4%.
W = 1.0 can simply mean all three modalities are shaking in the noise. Agreement and accuracy are two separate matters; the former can be perfect while the latter is entirely wrong.
III. Proportions Are Compositional Data; Five Ranks Are Not Independent
The proportions of variation sources must sum to 100%; this is closed data. The closure constraint naturally induces negative correlation among shares: as one increases, others must decrease.
Studies have demonstrated consequences using compositional data on garbage components: calculating by absolute mass avoids a positive correlation between vegetable waste and plastic packaging; switching to mass fractions turns the correlation negative (PMID 28882426); even the arithmetic mean can yield negative proportions with a confidence interval of (-4.03, 8.45).
This directly invalidates W’s premise—the assumption that $k$ rankings are independent. The five ranks of compositional data are bound together by the “constant sum” constraint. The correct approach is to perform log-ratio transformations (clr/ilr) before discussing variance (PMID 38806627).
IV. Too Few Levels Cause Variance Components to Float
Posture has only 2 levels (sitting/lying), belt position 2-3, and operator 2-3. Few levels are the source of all troubles.
Negative variance estimates and boundary estimates are common in such designs. Methodological work specifically discusses “pseudo-variance”: certain data marginal distributions are mathematically equivalent to a “negative variance random intercept slope model,” and few measurements plus short follow-ups expand this region (PMID 38695394). Other simulations compare estimators regarding “the probability of yielding negative estimates” (PMID 26843971).
Testing “whether a variance component is zero” is more complicated: zero lies on the boundary of the parameter space, so the standard chi-square distribution for likelihood ratio and score statistics is incorrect; permutation methods are needed to approximate the finite sample distribution (PMID 22930674).
REML sampling variance relies on large-sample asymptotics—studies suggest sample sizes up to 1413 are needed for the asymptotic approximation to be adequate (PMID 35197554). In a sample of only dozens, the sampling error for proportions may exceed the small differences being compared.
V. “Posture Effect Exists” $\neq$ “Posture Is the Largest Source”
Evidence that posture affects breathing is actually quite solid: a systematic review of 43 studies found significant decreases in functional residual capacity from sitting to lying (PMID 30305051); in a 62-person IOS test, lying showed increased R5 and R5-R20 with more negative X5, especially in the obese group (PMID 36787314); in 17 ARDS patients, sitting for 30 minutes caused a global rearrangement of EIT ROI ventilation distribution (PMID 39116511).
However, these only show posture has an effect, not that it is the largest source. Separating them requires cross-source comparison, which is precisely the most unstable step.
Incidentally: EIT regional ventilation distribution shows excellent repeatability under fixed ventilation settings (10 pigs; ventilation centroid shifted from $45 \pm 3\%$ to $49 \pm 3\%$ with a very narrow interval, PMID 17664640). Good repeatability only indicates stability within the same posture, unrelated to cross-source magnitude comparison.
VI. Another Hidden Premise: Heterogeneity of Variance
The G-theory classic assumption is homoscedasticity of residuals across conditions. This assumption often fails in physiological data. Work using location-scale mixed models extended G-theory, finding that failing to model heterogeneity masks true differences in measurement quality; G and D coefficients must be calculated separately by condition to reveal them (PMID 41549064). Multi-level rater reliability is similar—few raters are the main cause of ICC bias and low efficiency (PMID 33818118).
Three modalities are three raters, already few; stacking heterogeneity means the “average proportion” represents nothing for any single condition.
Reversal: What Should Be Reported Is Not W, But Confidence Intervals for Adjacent Differences
The flaw in the entire design is asking the wrong question. The real question should be “by how much is the posture proportion larger than the second place, and does this difference exceed uncertainty,” not “how similar are the three rankings.” The solution is to replace rankings with pairwise differences plus bootstrap confidence intervals: only adjacent pairs with CIs not crossing zero count as distinguishable. W cannot answer such a question.
Four judgment criteria:
1. All three-modal W significant AND all adjacent difference CIs non-crossing zero $\Rightarrow$ ranking is credible; posture dominance can be discussed.
2. W significant BUT adjacent difference CIs cross zero $\Rightarrow$ only report “posture higher than others,” not a full ranking.
3. W non-significant BUT some source pair difference CI stable non-crossing zero $\Rightarrow$ it is a granularity issue, not inconsistency; report differences.
4. W non-significant AND all difference CIs cross zero $\Rightarrow$ data does not support any ranking conclusion; “posture dominance” reverts to hypothesis.
Three supporting rules: Always report the number of attainable values and their rank relative to the critical value when reporting W; perform ilr/clr transformation for compositional data first; calculate sample size based on the “minimum proportion difference to resolve,” not based on “enough for W.”
Evidence and Limitations
Verified: Kendall W attainable value discreteness and exact permutation critical values for $k=3, n=5$ (exhaustive calculation in this paper); compositional data introduces pseudo-correlation and arithmetic means can yield negative proportions (PMID 28882426); necessity of log-ratio transformations for compositional data (PMID 38806627); causes of negative and pseudo-variance models, where few measurements and short follow-ups expand this region (PMID 38695394); standard chi-square distribution failure for zero-variance tests at parameter boundaries requiring permutation (PMID 22930674); REML sampling variance asymptotics may be biased for small samples (PMID 35197554); posture effects on lung function (systematic review of 43 studies, PMID 30305051); IOS impedance increases and X5 becomes more negative in lying position (PMID 36787314); EIT alters ventilation distribution in sitting position (PMID 39116511); EIT regional ventilation distribution repeatability (PMID 17664640); G-theory requires extension under heterogeneity (PMID 41549064); few raters cause ICC bias (PMID 33818118).
Partially Verified: Most posture evidence comes from mechanically ventilated patients, obese individuals, or ARDS populations; the magnitude of posture effects in semi-recumbent/sitting healthy practitioners remains unknown; flipping probabilities for proportion rankings are Monte Carlo simulation results from this paper, dependent on assumed true proportions and error structures, not measured; magnitude of pseudo-correlation in compositional data for variation source decomposition was not directly measured.
To Be Verified: True proportion ranges for the five variation sources in the Dantian breathing scenario (no direct research exists); whether posture and belt position interact (whether tension changes after lying with abdominal belts are simply additive); whether residual variances for the three modalities satisfy homogeneity; whether replacing W with difference intervals preserves the original “posture dominance” conclusion.
L1 Network Status: “This topic has not yet entered the L1 corresponding network”—The CORRESPONDENCE_NETWORK.md node table lists “Phase/Rhythm” as physical-side nodes and “Respiration Rate/Pattern” as physiological-side observations, but neither connects to any Yijing nodes; the node table contains no measurement-dimension nodes such as “Variation Source/Variance Component/Concordance Coefficient/Ranking,” and Edge Tables 3.1–3.4 only register Five Phases $\leftrightarrow$ Five Organs/Five Colors/Five Tones/Five Directions/Five Seasons/Five Emotions, Eight Trigrams $\leftrightarrow$ Five Phases, Heavenly Stems/Branches $\leftrightarrow$ Five Phases. No edges exist with variation source decomposition, posture dominance, or Kendall W as source or target. EDGE_TABLE_science-yixue.md only contains concept counts (Dantian 51 / Qigong 48 / Breathing 189), with no hierarchy or status determination. Thus, this paper is capped at Level C.
Related Reading
- Original Design for Tri-Modal Variation Source Decomposition: Does Posture Truly Overrule Everything
- “All Three Modalities Move in the Same Direction” — Actually Never Verified
- MDD Is Not “Smallest Useful Difference,” But “Smallest Non-Noise Difference”
- Environment-Life Frequency Interface (P0 Pool)
Evidence Level: C
References: PMID 28882426, PMID 38806627, PMID 38695394, PMID 26843971, PMID 22930674, PMID 35197554, PMID 30305051, PMID 36787314, PMID 39116511, PMID 17664640, PMID 41549064, PMID 33818118
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