Evidence level: C
排序不是测出来的,是比较出来的;而比较用的那把尺子,本身也有刻度。
一、信噪比是比值,分母小的时候它不老实
命题要求把五个参数按「设备偏倚 ÷ 效应量」排个高低。这个量是比值,分子分母都带误差,比值估计量的方差不是两者方差之和——两个估计相关时还要减去协方差项,而分母接近零时它会爆炸。X5 与 AX 恰恰是分母最不稳定的一类:强迫振荡的电抗类参数重复性明显弱于电阻类,学龄前儿童实测 X5 的组内相关仅 0.55、95% 置信区间下限掉到 −0.17,相对重复系数 91.8%,而同批人的 R5 是 0.82 与 35.5%(PMID 35927844)。分母自己就抖成这样,比出来的比值自然也抖。
二、排序不稳定有实测:单队列的第十名,重抽能进前三
英国 Trent 癌症登记处做过一次直接的检验:把地区生存率排名反复随机重抽,一个在 1 年乳腺癌生存率上排第十的地区,6% 的重抽里能进前三;同一地区在三个相邻队列里分别排第十、第五、第十(PMID 11429710)。作者的原话值得记住——排名差异很大,可能只反映实际结局上微不足道的差别。五个参数排五个名次,比几十个地区排名更容易翻转。
三、「谁最大」是选出来的,选出来的极值天生偏大
「X5/AX 偏倚最大」这句话,是从五个数里挑出最大的那个得来的。凡是先选后报的极值,都带着赢家诅咒:被选中的估计系统性偏高,重测会向均值回落(PMID 41869282 讨论的正是这个机制)。所以即使第一轮真测出 X5 在顶端,第二轮它掉到第二、第三也是预期之内,不构成对命题的反驳,也不构成对命题的证实——单轮排序根本没有检验力。
四、排序还取决于用哪个指标当裁判
人工晶体度数公式的横向比较给了一个干净的示范:八个公式在 330 只超长眼轴上比精度,按均方根绝对误差排,Zhu-Lu 最低(0.562 D),Cooke K6 紧随(0.563 D);可换成中位绝对误差,Cooke K6 反而最低(0.271 D),前者的优势消失(PMID 42177933)。同一批数据、同一批公式,换个评价指标冠军就换人。丹田五参数若分别用 ICC、LoA 宽度、相对重复系数当尺子,排出来的序很可能不是同一个。
五、三个模态共用同一批人,一致可能是假的
让 RIP、FOT、EIT 各自排一次序,再算排序一致性,看起来很稳。麻烦在于三个模态测的是同一批受试者:体型、年龄、呼吸模式同时影响三处的效应量,会制造出共同的相关结构,让三个排序「自发地」相像。这类假设违背会把一致性系数推高——有研究实测,忽略方差异质性时组内相关被高估到 0.640,忽略研究间变异再抬 0.02–0.07(PMID 30208858)。一致本身不能证明命题,只能证明三个模态被同一批人牵着走。
六、设备偏倚属于设备对,不属于参数
这是最硬的一处。两台商用振荡仪对 49 例患者加 27 名健康人的配对测量显示:全部阻抗参数都显著不同,AX 的系统偏倚 −4.10 cmH₂O/L,一致性界限 −14.99 到 6.80,ICC 却还在 0.59–0.85 之间;结论是这两台机器不能互换使用(PMID 42871008)。请注意「这两台」。换一台参考机、换一条呼吸带、换一个电极环,偏倚就换了。所以「X5/AX 偏倚在所有参数中最大」这句话缺一个主语——哪一对设备。
反转:该报的是双栏表加置信集,不是排序
把命题改写成可回答的形式:对指定的设备对,逐参数同时报偏倚及其置信区间、效应量及其置信区间,再报排序的 bootstrap 置信集。排序只在设备对层面成立,不能上升为参数属性。
四格判据:三个模态都排出 X5/AX 在顶端、且相邻名次差的区间不跨零——排序可报,但限定于本设备对;三模态一致而前两名差跨零——只能报「X5/AX 属偏高组」,不许报「最大」;一致性系数显著但换设备对后翻转——排序是设备对属性,禁止上升;一致性系数不显著——退回逐参数原文偏倚与一致性界限,排序这条线直接关闭。
证据与局限
已验证:各参数设备偏倚与一致性界限确实差异很大,且电抗类参数的重复性弱于电阻类(PMID 42871008、35927844、22960661);单队列排名不稳定、极值估计偏高、指标替换会改变名次,这三条在别的领域都有实测(PMID 11429710、41869282、42177933)。
部分验证:排序确定性本身已有量化工具,网络 meta 分析里提出过 POTH 指标与秩聚类方法(PMID 40542622、41635948),小评判者数下的一致性系数区间估计也有专门比较(PMID 24605969);但这些工具都还没用在丹田参数上。
待验证:X5/AX 是否真的在五参数中偏倚最大——现有文献只能证明某两台机器之间 AX 偏倚大,无法证明它在所有设备对、所有人群里都最大。跨模态排序是否复现,从未被检验过。
L1 网络状态:该主题尚未进入 L1 对应网络。网络节点表只覆盖易学/中医侧、物理侧、生理/体验侧三类节点,无「设备偏倚 / 信噪比 / 排序一致性」等测量学维度节点;边表当前只到 3.4 天干地支↔五行;「丹田」仅在概念计数中出现(51 篇,最高 A 级 11 篇),未形成对应边。
相关阅读
证据等级: C
参考文献: PMID 42871008, PMID 35927844, PMID 22960661, PMID 30208858, PMID 41869282, PMID 42177933, PMID 11429710, PMID 40542622, PMID 41635948, PMID 24605969
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。
Can the \”Signal-to-Noise Ratio Ranking\” of Dan Tian Parameters Be Reproduced Across Modalities: Is the Device Bias of X5/AX Truly the Largest?
Evidence level: C
Ranking is not measured; it is comparative. And the ruler used for comparison also has its own scale.
I. Signal-to-Noise Ratio is a Ratio; When the Denominator is Small, It Becomes Unstable
The proposition requires ranking five parameters by “Device Bias ÷ Effect Size.” As a ratio, both numerator and denominator carry error. The variance of the ratio estimator is not merely the sum of the two variances; when estimates are correlated, a covariance term must be subtracted. Furthermore, as the denominator approaches zero, the variance explodes. X5 and AX belong precisely to the category with the most unstable denominators: reactance-type parameters forced into oscillation exhibit significantly weaker repeatability than resistance-type parameters. In pre-school children, the intraclass correlation (ICC) for X5 was only 0.55 with a 95% confidence interval lower bound dropping to −0.17, and the relative repeatability coefficient reached 91.8%. In contrast, R5 from the same cohort showed an ICC of 0.82 and a repeatability coefficient of 35.5% (PMID 35927844). If the denominator itself is so unstable, the resulting ratio will naturally fluctuate.
II. Ranking Instability Has Empirical Evidence: A Single Queue’s Tenth Place Can Enter the Top Three Upon Resampling
The UK Trent Cancer Registry conducted a direct test: by repeatedly randomly resampling regional survival rates, a region ranked tenth in one-year breast cancer survival appeared in the top three in 6% of resamples. The same region ranked tenth, fifth, and tenth across three adjacent cohorts (PMID 11429710). The author’s own words are worth remembering: ranking differences can be large, potentially reflecting only trivial differences in actual outcomes. Ranking five parameters against five ranks is far more prone to flipping than ranking dozens of regions.
III. “Who is Largest” Is Selected; Selected Extremes Are Naturally Biased High
The statement “X5/AX bias is the largest” is derived by selecting the maximum value from five numbers. Any extreme selected after the fact carries the winner’s curse: the selected estimate is systematically biased upward, and re-measurement will regress toward the mean (PMID 41869282 discusses this mechanism). Therefore, even if X5 truly appears at the top in the first round, dropping to second or third in a second round is expected and does not refute the proposition nor confirm it. Single-round ranking lacks statistical power for validation.
IV. Ranking Depends on Which Metric Serves as the Judge
A clean demonstration comes from comparing artificial lens power formulas across eight formulas in 330 ultra-long axial lengths. Ranked by root mean square absolute error, Zhu-Lu was lowest (0.562 D), followed closely by Cooke K6 (0.563 D). However, using median absolute error, Cooke K6 became the lowest (0.271 D), causing Zhu-Lu’s advantage to vanish (PMID 42177933). With the same data and formulas, changing the evaluation metric changes the champion. If the five Dantian parameters are measured using ICC, LoA width, or relative repeatability coefficient as the ruler, the resulting ranking will likely differ.
V. Three Modalities Share the Same Subjects; Consistency May Be False
Allowing RIP, FOT, and EIT to each generate a ranking independently, then calculating ranking consistency, appears stable. The complication lies in the fact that three modalities measure the same cohort of subjects: body habitus, age, and breathing patterns simultaneously affect effect sizes across all three modalities, creating a shared correlation structure that causes the three rankings to resemble each other “spontaneously.” Such assumption violations inflate consistency coefficients—studies have measured intraclass correlation overestimated to 0.640 when variance heterogeneity is ignored, and further inflated by 0.02–0.07 when between-study variation is ignored (PMID 30208858). Consistency alone cannot prove the proposition; it only proves that the three modalities are being led by the same cohort.
VI. Device Bias Belongs to Device Pairs, Not Parameters
This is the most rigorous point. Paired measurements of 49 patients and 27 healthy subjects using two commercial oscillators showed that all impedance parameters were significantly different. AX exhibited a systematic bias of −4.10 cmH₂O/L with limits of agreement ranging from −14.99 to 6.80, yet the ICC remained between 0.59 and 0.85; the conclusion was that these two machines cannot be used interchangeably (PMID 42871008). Note the phrase “these two machines.” Changing a reference machine, a breathing belt, or an electrode ring changes the bias. Therefore, the statement “X5/AX bias is the largest among all parameters” lacks a subject—specifically, which device pair.
Reversal: Report Dual-Column Tables with Confidence Sets, Not Rankings
Reformulate the proposition into a testable form: For a specified device pair, simultaneously report bias and its confidence interval for each parameter, effect size and its confidence interval, and then report the bootstrap confidence set for the ranking. Ranking holds only at the device-pair level and cannot be elevated to a parameter attribute.
Four-Grid Criterion: If all three modalities rank X5/AX at the top AND the interval for adjacent rank differences does not cross zero—ranking can be reported but must be limited to this specific device pair. If three modalities agree but the interval for the top two ranks crosses zero—only report “X5/AX belongs to the high-bias group”; do not report “largest.” If consistency coefficients are significant but flip upon changing the device pair—ranking is a device-pair attribute; prohibit elevation. If consistency coefficients are not significant—revert to raw parameter bias and limits of agreement; close the ranking line entirely.
Evidence and Limitations
Verified: Significant differences exist in device bias and limits of agreement across parameters, with reactance-type parameters showing weaker repeatability than resistance-type (PMID 42871008, 35927844, 22960661); single-cohort ranking instability, winner’s curse in extreme estimates, and metric substitution altering ranks are empirically verified in other fields (PMID 11429710, 41869282, 42177933).
Partially Verified: Tools for quantifying ranking determinism exist, including the POTH index and rank clustering methods proposed in network meta-analysis (PMID 40542622, 41635948), and specialized comparisons for consistency coefficient interval estimation with small numbers of judges (PMID 24605969); however, none of these tools have yet been applied to Dantian parameters.
To Be Verified: Whether X5/AX truly exhibits the largest bias among the five parameters—existing literature can only prove AX bias is large between two specific machines, unable to prove it is maximal across all device pairs and populations. Whether cross-modal ranking replicates has never been tested.
L1 Network Status: This topic has not yet entered the L1 corresponding network. The node table covers only Yijing/TCM, Physics, and Physiological/Experiential nodes; there are no measurement dimensions nodes for “Device Bias / Signal-to-Noise Ratio / Ranking Consistency.” The edge table currently only covers the 3.4 Heavenly Stems and Earthly Branches ↔ Five Phases mapping; “Dantian” appears only in concept counting (51 papers, highest A-level 11), without forming a corresponding edge.
Related Reading
- Cross-Modal Consistency Test for Ranking Variance Source Proportions
- Quantification of Interchangeability Rate Across Devices for Dantian Parameters
- Minimum Detectable Difference and Signal-to-Noise Threshold for CoV Displacement
- Environment-Life Frequency Interface (Pool 4)
Evidence Level: C
References: PMID 42871008, PMID 35927844, PMID 22960661, PMID 30208858, PMID 41869282, PMID 42177933, PMID 11429710, PMID 40542622, PMID 41635948, PMID 24605969
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.