Blog

  • 两道关卡真的挡住第二遍吗:双判据的四种形态与双向增量否决率

    Evidence level: C

    报告里写「两条判据全过」,读起来像双重保障,可这两条判据可能压根不是两根轴——其中一条从头到尾没挡下过任何东西。

    正文

    一、上一轮只审了一对

    上一轮(#1281)把 ICC_adj ≥ 0.75 与 VB_adj/VB ≥ 0.50 化简后发现,两条判据其实是同一个参数 s 的两个单侧上限,共用一根尺子,分界在 ρ = 6/7。当时只证明了这一对。

    但量产线上并排写的双判据远不止这一对:ICC ≥ 0.75 AND SEM ≤ 某值、MDC95 ≤ Δ AND 组间差显著、斜率 CI 上界 < 阈值 AND BP 检验显著,全是同一副面孔。这批组合里到底有几条是真两根轴,才是这条线的真问题。

    二、先看参数空间的维数

    测量学那批指标其实挤在一个很小的空间里。令个体间方差 VB、测量误差 VE 为仅有的两个参数,则

    $$\text{ICC} = \frac{VB}{VB+VE},\quad \text{SEM} = \sqrt{VE},\quad \text{MDC}_{95} = 1.96\sqrt{2\,VE}$$

    雅可比矩阵的秩在这里说得很死:ICC 与 SEM 的秩是 2(不共线),但 SEM 与 MDC95 的秩是 1——因为 MDC95 = 2.7719 × SEM 是个常数比,两者在 (VB, VE) 上指向同一个方向。ICC 与 MDC95 的秩也回到 2。

    结论比想象中硬:{ICC, SEM, MDC95} 这三个量在二维参数空间上最多只能构成两根轴,任意三个一起写,必定有一根是重复计数。 SEM ≤ a AND MDC95 ≤ b 这种写法,看着两条,实际只是一条线的两种说法,那条多余的永远挡不下东西。(PMID 15705040 讲清了 ICC 与 SEM 同源于同一份方差分解,PMID 41265662 进一步把信度与测量误差说成「相关但不同」的两个属性——「相关但不同」不等于「可以并列为两道关卡」。)

    三、四种形态

    按「第一道挡住之后第二道还能不能挡住」这条标准,可以把双判据分成四型:

    型 I 支配型:同一量、两条线,如 ICC ≥ 0.70 AND ICC ≥ 0.75。较松的那条被较严的那条完全包含,一条多余。
    型 II 重参数化型:两条判据是同一个量的单调变换,如 SEM ≤ a AND MDC95 ≤ b。判据对的交集就是一条线。
    型 III 同源异轴型:共用同一份方差分解,但分别约束 VB 与 VE 两侧,如 ICC ≥ 0.75 AND SEM ≤ 某值。这才是真的两根轴。
    型 IV 嵌套型:一条是另一条的换算式,如 #1281 那一对。上一轮已经把它的自由度归并做完了。

    有意思的是,型 I 与型 II 在代数上可以完全等价:只要把两条线调整成「严的那条恰好被松的那条包住」,支配型就退化成重参数化型。两者的区别只在于阈值选得宽不宽,不在结构上。

    四、增量否决率:一个便宜得多的验收指标

    上一轮(#1285)用「翻转率」评估判据,结论是那指标有平凡解漏洞、且要把翻转率压到 10% 以下需每段 n ≥ 1378。太贵。

    这里换一个指标:增量否决率 IVR,定义为「过了第一道之后,被第二道否掉的比例」= P(否2 | 过1)。它问的正是双判据存在的理由——第二道到底额外挡住了多少。

    四型的实测量级差别极大(40 万次/格):

    形态 P(过1) P(过2) IVR(否2|过1) IVR(否1|过2)
    型 I 支配(0.70/0.75,真值 0.78) 0.908 0.691 0.239 0.000
    型 II 重参数化(SEM/MDC95) 0.690 0.688 0.002 0.000
    型 III 同源异轴(ICC/SEM,各约五成) 0.499 0.499 0.250 0.250
    型 IV 嵌套(#1281 那对) 0.331 0.960 0.000 0.655

    五、只看一个方向会把两种形态都判成「有效」

    这张表最关键的一列不是大小,是对称性。

    型 III 两个方向都是 0.250——两条判据互不包含,各自能挡住对方放过的四分之一。型 I 只有 0.239 / 0.000,型 IV 只有 0.000 / 0.655,都是单向的。

    于是只看 IVR(否2|过1) 这一个数,型 I(0.239)排第二、型 III(0.250)排第一,几乎分不开;得把两个方向都算出来,才看得出型 III 是唯一双向都不为零的。单向 IVR 会把「支配型」误读成「真双轴」,这是这一轮最该记住的坑。

    更要留神:型 I 的 0.239 全部来自阈值间距除以 SE,是纯噪声,与数据毫无关系。真值固定时把两条线拉近,IVR 就掉;拉到无穷远,两条就独立。一个由阈值设计决定、与测量质量无关的数,不该当证据用。

    六、这个指标便宜

    按单比例检验反解:要检出真 IVR = 0.10(80% 功效、双侧 5%),只需 71 例「过了第一道」的样本;IVR = 0.20 只需 31 例。若第一道的通过率约 0.33,总样本分别约 214 与 95。

    对照 #1285 里翻转率法要的 n ≥ 1378/段,IVR 便宜一个量级。「这对判据是不是重复计数」这件事,用二百来号人就能验完,不该继续靠推理。

    七、连计数本身也不是一个数

    若非要给「两条判据算几个」一个数字,会撞上第三层麻烦。同一对判据(ρ = 0.80),有效独立检验数 M_eff 有三种算法,结果差 38%:

    • 按族错误率反解:α = 0.05 时 1.579,α = 0.001 时 1.672
    • Cheverud 公式 2 − r²/2:1.680
    • 2/(1 + r²):1.220

    而且 M_eff 随检验阈值漂(同一 ρ,α 从 0.05 到 0.001,从 1.579 到 1.672)——它不是一个相关结构的固有属性。文献侧对此早有警告:Nyholt 那套换算在存在单倍型块时会过度保守(PMID 16118503),后来的特征值法正是在纠这一点(PMID 16077740、PMID 19217024)。

    所以「两条判据 = 1.58 条」这种表述没有意义。能说的是:它的有效期在 1.22 到 1.68 之间,取决于你问的是族错误率、还是问平均相关系数。

    八、结论:先分类,再验证,最后才写「全过」

    回到开头。这批双判据不是「几条同轴」能一句话答完的,得按形态分:型 II 与型 IV 可以纯代数判死、不必采数据;型 I 属于阈值设计问题,把两条线合并成一条更严的即可;只有型 III 值得保留,且必须双向报 IVR,否则连它是否真两根轴都无从判断(PMID 33818118 的多层方差分解模板正是判型 III 的标准框架;PMID 29707825 提醒 ICC 类指标依赖人群异质性,同一对判据在不同样本上可能从型 III 退化成型 II)。

    一句话:「两条判据都通过」这句话,只有在这两条判据的增量否决率两个方向都不为零时,才配写成双重确认。

    证据与局限

    已验证:雅可比秩结论(MDC95/SEM = 2.7719 为常数,秩 = 1;ICC 与 SEM 秩 = 2)是纯代数,可任意精度验算;四型形态与双向 IVR 的量级差异由 40 万次/格模拟给出,正态与对数正态两族下方向一致;IVR 的样本量反解只依赖单比例检验的标准公式。

    部分验证:四型模拟用的是「对数正态近似方差分量的抽样分布」,真实数据的方差分量分布更偏、且 VB 与 VE 常负相关,会让型 III 的 IVR 两侧不对称;M_eff 的三种口径之差在文献中被指出方向不一(PMID 16118503 报 Nyholt 法过度保守,PMID 19217024 报其特征值法更准),本文只核了数值量级,没核各法在低相关区的行为。

    未验证 / 该主题尚未进入 L1 对应网络:本文讨论的「判据条数是否重复计数」在对应网络中既无节点也无边,属纯方法学层面,无法用任何易学—生理边来锚定。M_eff 的阈值依赖性缺乏闭式解,只能用数值反解;「IVR 双向都为零」是否与「雅可比秩为 1」严格等价,本文只在小样本模拟中验证了方向一致,未给出充分必要条件。

    相关阅读

    L1 网络状态:E-004|金|主|肺|《素问·阴阳应象大论》|现代证据「待补」|L1|open —— 该边仅为 L1 同构、状态 open,现代证据栏原文为「待补」。本文讨论的是双判据的形态分类与增量否决率,在对应网络中既无节点也无边(EDGE_TABLE 仅有「五行 50 篇 / 五脏 26 篇」文章计数,无层级与状态判定),只能借 E-004 作话题锚点。按硬规则,凡 L1 标为「薄 / 仅文献 / 无 A 级证据」者最高只能标 C,故本文所有结论均为方法学推演而非机制验证。

    证据等级: C
    参考文献: PMID 15705040, PMID 41265662, PMID 16118503, PMID 16077740, PMID 19217024, PMID 33818118, PMID 29707825


    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。

    Do Two Thresholds Really Block the Second Pass: Four Forms of Dual Criteria and Bidirectional Incremental Rejection Rate

    Evidence level: C

    The report states “both criteria passed,” sounding like double insurance, yet these two criteria may not be two axes at all—one of them never blocked anything from start to finish.

    Body

    I. The Previous Round Only Examined One Pair

    In the previous round (#1281), simplifying ICC_adj ≥ 0.75 and VB_adj/VB ≥ 0.50 revealed that both criteria are actually one-sided upper limits of the same parameter s, sharing a single ruler with a boundary at ρ = 6/7. Only this pair was proven.

    However, the dual criteria written side-by-side on mass-production lines far exceed this single pair: ICC ≥ 0.75 AND SEM ≤ some value, MDC95 ≤ Δ AND inter-group difference significant, slope CI upper bound < threshold AND BP test significant, and so on. They all share the same face. Among these combinations, how many are truly two axes is the real question of this line.

    II. First, Examine the Dimensionality of the Parameter Space

    The metrological indicators actually crowd into a very small space. Letting inter-individual variance VB and measurement error VE be the only two parameters, then

    $$\text{ICC} = \frac{VB}{VB+VE},\quad \text{SEM} = \sqrt{VE},\quad \text{MDC}_{95} = 1.96\sqrt{2\,VE}$$

    The rank of the Jacobian matrix is strictly defined here: ICC and SEM have rank 2 (non-collinear), but SEM and MDC95 have rank 1—because MDC95 = 2.7719 × SEM is a constant ratio, both pointing in the same direction on (VB, VE). The rank of ICC and MDC95 also returns to 2.

    The conclusion is harder than imagined: {ICC, SEM, MDC95} can form at most two axes in this two-dimensional parameter space; writing any three together inevitably involves double-counting. SEM ≤ a AND MDC95 ≤ b looks like two lines but is actually one line expressed in two ways; the extra one never blocks anything. (PMID 15705040 clarifies that ICC and SEM originate from the same variance decomposition; PMID 41265662 further describes reliability and measurement error as “correlated but distinct” attributes—”correlated but distinct” does not mean “can stand side-by-side as two gates”.)

    III. Four Morphologies

    According to the standard “after the first gate blocks, can the second still block?”, dual criteria fall into four types:

    Type I Dominant: Same quantity, two lines, e.g., ICC ≥ 0.70 AND ICC ≥ 0.75. The looser one is completely contained by the stricter one; one is redundant.
    Type II Reparameterized: Both criteria are monotonic transformations of the same quantity, e.g., SEM ≤ a AND MDC95 ≤ b. The intersection of the pair is exactly one line.
    Type III Homologous but Different Axes: Sharing the same variance decomposition but constraining opposite sides of VB and VE, e.g., ICC ≥ 0.75 AND SEM ≤ some value. This is the only true two-axis case.
    Type IV Nested: One is the conversion formula of the other, e.g., the pair in #1281. The previous round already merged its degrees of freedom.

    Interestingly, Type I and Type II are algebraically completely equivalent: as long as the two lines are adjusted so “the stricter one exactly contains the looser one,” the dominant type degenerates into the reparameterized type. Their difference lies only in whether thresholds are wide or narrow, not in structure.

    IV. Incremental Rejection Rate: A Cheaper Acceptance Metric

    The previous round (#1285) used “flip rate” to evaluate criteria, concluding that metric has trivial solution loopholes and requires n ≥ 1378 per segment to suppress flip rate below 10%—too expensive.

    Here we switch to a new metric: Incremental Rejection Rate (IVR), defined as “the proportion rejected by the second gate after passing the first” = P(Reject2 | Pass1). It asks precisely why dual criteria exist—how much extra does the second gate actually block?

    The real measurement scale differences across the four types are enormous (400,000 per grid):

    | Morphology | P(Pass1) | P(Pass2) | IVR(Reject2|Pass1) | IVR(Reject1|Pass2) |
    |—|—|—|—|—|
    | Type I Dominant (0.70/0.75, true 0.78) | 0.908 | 0.691 | 0.239 | 0.000 |
    | Type II Reparameterized (SEM/MDC95) | 0.690 | 0.688 | 0.002 | 0.000 |
    | Type III Homologous-Different Axes (ICC/SEM, each ~half) | 0.499 | 0.499 | 0.250 | 0.250 |
    | Type IV Nested (#1281 pair) | 0.331 | 0.960 | 0.000 | 0.655 |

    V. Looking in Only One Direction Misjudges Two Types as “Effective”

    The most critical column in this table is not magnitude but symmetry.

    Type III has 0.250 in both directions—neither criterion contains the other; each blocks one-quarter of what the other lets pass. Type I has only 0.239 / 0.000, Type IV only 0.000 / 0.655; both are unidirectional.

    Thus, looking only at IVR(Reject2|Pass1) alone, Type I (0.239) ranks second and Type III (0.250) ranks first, nearly indistinguishable; one must calculate both directions to see that Type III is the only bidirectional non-zero case. A single-direction IVR misreads “dominant type” as “true dual-axis,” which is the pitfall of this round.

    More caution: Type I’s 0.239 comes entirely from threshold spacing divided by SE, pure noise unrelated to data. When true values are fixed and two lines brought closer, IVR drops; pulled infinitely apart, the two become independent. A number determined by threshold design and unrelated to measurement quality should not be used as evidence.

    VI. This Metric Is Cheaper

    Solving via single-proportion test inversion: to detect true IVR = 0.10 (80% power, two-sided 5%), only 71 samples “passed the first gate” are needed; IVR = 0.20 requires just 31. If the first gate pass rate is about 0.33, total samples are roughly 214 and 95 respectively.

    Compared with the flip-rate method requiring n ≥ 1378/segment in #1285, IVR is cheaper by an order of magnitude. “Whether this pair of criteria is double-counting” can be verified with just a few hundred people, and should not rely on reasoning.

    VII. Even the Count Itself Is Not a Single Number

    If one must assign “how many” to “two criteria,” it hits a third layer of trouble. For the same pair of criteria (ρ = 0.80), effective independent test count M_eff has three algorithms, differing by 38%:

    • Solving via family-wise error rate: α = 0.05 gives 1.579, α = 0.001 gives 1.672
    • Cheverud formula 2 − r²/2: 1.680
    • 2/(1 + r²): 1.220

    Moreover, M_eff drifts with test thresholds (same ρ, α from 0.05 to 0.001, M_eff from 1.579 to 1.672)—it is not an intrinsic property of the correlation structure. Literature has long warned: Nyholt’s conversion is overly conservative when haplotype blocks exist (PMID 16118503); later eigenvalue methods correct this point (PMID 16077740, PMID 19217024).

    Thus “two criteria = 1.58 axes” is meaningless. What can be said: its validity lies between 1.22 and 1.68, depending on whether you ask about family-wise error rate or average correlation coefficient.

    VIII. Conclusion: Classify First, Verify Second, Then Write “All Passed”

    Returning to the start. This batch of dual criteria cannot be answered in one sentence by “how many co-axial lines”; it must be classified by morphology: Type II and Type IV can be purely algebraically decided without data; Type I is a threshold design issue—merging the two lines into one stricter line suffices; only Type III deserves retention, and must report IVR bidirectionally, otherwise even whether it is truly two axes cannot be judged (PMID 33818118’s multilevel variance decomposition template is the standard framework for judging Type III; PMID 29707825 warns ICC-type indicators depend on population heterogeneity, so the same pair of criteria may degenerate from Type III to Type II across different samples).

    In one sentence: “Both criteria passed” is worthy of writing as double confirmation only when the incremental rejection rate in both directions is non-zero.

    Evidence and Limitations

    Verified: The Jacobian rank conclusion (MDC95/SEM = 2.7719 is constant, rank = 1; ICC and SEM rank = 2) is purely algebraic and can be verified to arbitrary precision; the magnitude differences among the four types and bidirectional IVR scales were given by 400,000-per-grid simulations, consistent under both normal and log-normal families; IVR sample size inversion relies only on standard single-proportion test formulas.

    Partially Verified: The four-type simulation uses “log-normal approximation of sampling distribution of variance components”; true data variance component distributions are more skewed, and VB and VE often negatively correlated, making Type III IVR asymmetric on both sides; the three M_eff definitions differ in direction as noted in literature (PMID 16118503 reports Nyholt method overly conservative, PMID 19217024 reports eigenvalue method more accurate); this paper only checked numerical scales, not low-correlation behavior of each method.

    Not Verified / This topic has not yet entered the L1 correspondence network: The question “whether criteria count is double-counting” has neither nodes nor edges in the correspondence network, belonging purely to methodology, and cannot be anchored by any Yijing-physiology edge. M_eff’s threshold dependence lacks closed-form solution, requiring numerical inversion; “IVR zero in both directions” being strictly equivalent to “Jacobian rank = 1” was only verified directionally in small-sample simulation, not given necessary and sufficient conditions.

    Further Reading

    L1 Network Status: E-004|Metal|Primary|Lung|《Su Wen · Yin-Yang Ying Xiang Da Lun》|Modern evidence “to be supplemented”|L1|open — this edge is only L1 isomorphic, status open; the modern evidence column originally reads “to be supplemented.” This paper discusses dual criteria morphology classification and incremental rejection rate; in the correspondence network there are neither nodes nor edges (EDGE_TABLE contains only article counts “Five Phases 50 / Five Organs 26,” no level and status judgment), so E-004 is borrowed as the topic anchor. By hard rules, any L1 marked “thin / only literature / no A-level evidence” can be marked at most C; thus all conclusions here are methodological deduction rather than mechanism verification.

    Evidence Level: C
    References: PMID 15705040, PMID 41265662, PMID 16118503, PMID 16077740, PMID 19217024, PMID 33818118, PMID 29707825


    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.

  • 换成斜率加区间之后判据翻转率真降了吗:两法头对头其实共用一个旋钮

    Evidence level: C

    把判据从「显著性检验 + 阈值」改成「斜率 + 置信区间」,看着像换了工具,代数上是同一个式子的两种写法;真正决定结论稳不稳的,只有一个数——判据线离标称线几个标准误。

    正文

    一、一条从没被验证过的建议

    上一轮讨论呼气末容积的年龄异方差,结论是「BP 检验显著、方差比超 1.5」这套两步判据有毛病:功效太低,测不出来时默认滑进「同方差够用」,于是「没测出来」被读成「不存在」。当时给的补救是改成直接报「log(VE) 对年龄的斜率与它的置信区间」。

    这建议听上去顺,但它本身没人验证过。换成斜率之后,判据就不会在边界附近翻来翻去了吗?还是只换了地方翻?一个判据形式值不值得换,得看它把「同一真值、两次独立抽样给出相反结论」的概率压下去多少。这个概率叫翻转率,等于 2p(1−p),p 是判 H(同方差够用)的概率。

    二、两法是同一个式子的两种写法

    先把符号摆平。令 θ = log(VE_老 / VE_青),标称线 θ₀ = log(1.5) = 0.4055,两段各 n 人时 log s² 方差 2/(n−1),相减得 SE(θ̂) = √(8/(n−2))。

    • 两步法:方差比检验 + 点估计过线,两个条件与起来等效边距约 k ≈ 1.0;
    • 斜率区间法:θ̂ + 1.645·SE < θ₀,k = 1.645;
    • 点估计法:θ̂ < θ₀,k = 0。

    看清结构了吗——三套判据全落在同一个族里:判 H 当且仅当 θ̂ + k·SE < θ₀。所谓「换成斜率」,无非是把 θ 换个名字(斜率分子分母同除年龄跨度),CI 端点单调对应。换判据形式没改变判决逻辑,只改变 k。所以换法后翻转率若真降了,功劳属于 k 被抬高,不属于「斜率更先进」——因果记错,下次就照着错的原因改。

    三、线上翻转率有闭式解,而且不含 n

    把真值放在标称线上,θ̂ ~ N(θ₀, SE²),判 H 要求 Z < −k,概率就是 Φ(−k)。翻转率

    2·Φ(−k)·(1−Φ(−k))
    

    k = 0 时 0.500;k = 1.0 时 0.267;k = 1.645 时 0.095;k = 1.96 时 0.049;k = 2.326 时 0.020。

    注意式子里没有 n。这是最反直觉的一条:加大样本不能让标称线上的翻转率变好。λ 恰好等于 1.5 这事本身就在模棱两可的位置上,样本再多也只是把「模棱两可」测得更精确。加样本压窄模糊带宽度,压不掉线上那一份。

    顺带得上界:任何二分判据翻转率最多 0.5,永远降不到零。要低于 10%,边距须超 1.62 个 SE;低于 5%,须超 1.95 个。这条与终点是什么、跟易学有没有关系都无关,是二分决策本身带的(PMID 27209009 整理了这类误读,PMID 16217841、22733454、9598427 量化过二分连续量的代价)。

    四、最坏点其实不在标称线上

    线上 P(H) = Φ(−k) 只是一个点。把 λ 当变量扫过去,p(λ) = Φ((log λ − θ₀)/SE − k),等于 0.5 的位置是

    λ* = 1.5·exp(k·SE)
    

    k = 1.645 时:n = 60,λ = 2.76;n = 120,λ = 2.30;n = 600,λ = 1.81;n = 3000,λ = 1.63。

    也就是说最坏点永远落在标称线上方,随样本量增大往线上靠,但路不到 1.50。50% 的翻转率始终存在,只是站在不同位置。所以「λ = 1.5 处翻转率只有 9.5%」有相当的欺骗性——它报的是一个点的读数,而曲线最坏点恒为 50%。要报就报整条曲线,或至少报 λ* 在哪儿。

    五、网格实测:三法头对头

    我们把三套判据放进 4×4 网格(真 λ ∈ {1.0, 1.2, 1.5, 2.0} × n ∈ {60, 120, 300, 600})各跑两万次,另加厚尾臂。为防「只看翻转率会选错判据」,同时算两端均衡判对率 BA = [P(H│λ=1.0) + P(须分段│λ=2.0)] / 2,0.5 等于抛硬币。

    两条值得记住:

    第一,线上翻转率与 n 完全无关,k = 1.645 那一列在 n = 60/120/300/600 上分别是 0.097、0.098、0.100、0.092,波动只是抽样误差。

    第二,翻转率最好的判据不是分辨力最好的判据。k = 0 的点估计法线上翻转率 0.500(最差),但 BA 在 n = 120 时 0.904;k = 1.645 的区间法线上翻转率 0.098(最好),BA 只有 0.731。两者换的是同一样东西:边距越宽,越不容易翻,也越不容易说对。

    六、只报翻转率会被常数列刷满分

    这条是本轮最该警惕的方法学硬伤。假设有人提议一张假判据:不管数据,一律答「同方差够用」。它在所有 8 个格子里 p = 0,翻转率 2p(1−p) = 0.000——满分。反过来一律答「须分段」,p = 1,翻转率也是 0.000,同样满分。

    任何只用翻转率一个标量评估二分判据的方案,都系统性奖励「不表态」。这不是理论空子:区间法在 n = 60 时判 H 的概率只有 0.048,已接近常数列,它的漂亮翻转率是「几乎恒答须分段」换来的,BA 只有 0.640。要判它合不合格,必须同时报固定检出率下的假阳性或整条 ROC。单看翻转率,等于让考官自己出题。

    七、真实代价:模糊带与样本量

    按「两端各 90% 判对」划线,k = 1.645 时 n = 120 只能保证 λ ≤ 0.70 干净判 H、λ ≥ 1.36 干净判分段,中间 [0.70, 1.36] 怎么判都有一头不稳。n = 600 收窄到 [1.07, 1.44],n = 2400 才到 [1.27, 1.47]。

    换算成样本量:要让「λ ≤ 1.2 时 90% 判 H」,k = 1.0 需每段 n ≥ 838,k = 1.645 需 ≥ 1378(总 2756),k = 1.96 需 ≥ 1690。对照计划 n = 120,缺口 11.5 倍。

    不二分更省:直接报 θ̂ 的 95% CI,宽度用倍数表示是 exp(3.92·SE)。n = 120 时上下界比 2.78 倍——「最多两倍差」这句在 n = 120 时根本排除不掉;n = 300 降到 1.90 倍,n = 600 是 1.57 倍。

    厚尾还要再打一折。误差若接近 t(5),log s² 的方差是 (2+κ)/df,κ = 6 时是正态四倍,真 SE 等于名义两倍。线上 P(H) 仍 0.050(对称性保住),但整条曲线被横向拉伸一倍:n = 120 的模糊带只相当于正态下 n = 30。这也解释了实测里区间法常比公式预测的差(PMID 14673940 报过 CI 近边界时覆盖率不足,PMID 29707825 报 ICC 类指标随分布改变性质)。

    八、结论:换的不是判据,是边距

    回到开头那条追问。答案是:换成斜率加区间,翻转率不会自动下降。若新做法把 k 从约 1.0 抬到 1.645,翻转率确实从 0.267 降到 0.095;但功劳属于边距,不属于斜率。若新做法仍用点估计或仍按 p < 0.05 落判决,翻转率一分不降。

    所以正确写法是四件事一起报:边距 k、样本量 n、线上 P(H)、最坏点 λ* 的位置。缺任一个,「这判据稳不稳」就没内容。若目标真是可重复性而非凑结论,最省的是根本不二分——直接报 θ̂ 与 CI,让读者看区间跨不跨 1.5(PMID 35841249 提醒:区间重叠不等于无差异,不重叠也不等于有差异,二分场景里加倍成立)。

    证据与局限

    已验证:判据族 θ̂ + k·SE < θ₀ 的线上翻转率闭式解 2Φ(−k)(1−Φ(−k)),与 n 无关;最坏点位置 λ* = 1.5·exp(k·SE) 随 n 单调向标称线靠拢而不抵达;只报翻转率会被常数列刷满分(恒答者翻转率 0.000)。这三条都是纯代数,可逐位验算,不依赖任何数据。

    部分验证:4×4 网格的蒙特卡洛结果与闭式解吻合(正态臂误差 < 0.005),厚尾臂的「真 SE = 2×名义」由峰度公式给出、抽样核验方向一致。但这套模拟用的是「两段独立、方差比的对数近似正态」这个简化;真实数据里两段常相关、分布常偏,量化结论会偏移。文献侧,二分连续量的功效损失有稳定共识(PMID 16217841、22733454、9598427),但「斜率法优于检验法」这一具体主张并未在文献中被直接检验过——本轮的意义恰恰是指出它不必被检验,因为两者等价。

    待验证:① 本项目考虑的终点(呼气末容积、DR2 余量)在两段相关不为零时,θ̂ 的 SE 该用哪个表达式,会让上述所有数字整体缩放;② 若把「二分决策」本身取消、改报连续量,临床与工程侧是否接受这种输出形式,属于接受度问题而非统计问题,本项目无法单方面裁定;③ 厚尾臂只测了 t(5) 一个点,κ 从 0 到 6 之间翻转率如何连续变化没有扫。

    相关阅读

    L1 网络状态:E-004|金|主|肺|《素问·阴阳应象大论》|现代证据「待补」|L1|open —— 该边仅为 L1 同构、状态 open,现代证据栏原文为「待补」;本文讨论的是判据形式的等价性与翻转率,在对应网络中既无节点也无边,只能借 E-004 作话题锚点,故证据等级最高只能标 C,文中所有结论均为方法学推演而非机制验证。

    证据等级: C
    参考文献: PMID 27209009, PMID 16217841, PMID 22733454, PMID 9598427, PMID 17066378, PMID 23130635, PMID 14673940, PMID 29707825, PMID 35841249


    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。

    Does the flip rate of the criterion really decrease after adding slope and interval: two methods head-to-head actually share one knob

    Evidence level: C

    Switching the criterion from “significance test + threshold” to “slope + confidence interval” appears like changing tools, yet algebraically it is merely two ways of writing the same equation. What truly determines the stability of the conclusion is a single number: how many standard errors the criterion line lies from the nominal line.

    Body

    I. One Suggestion Never Verified

    In the previous discussion of age heteroscedasticity in end-tidal volume, the conclusion that “the BP test is significant and the variance ratio exceeds 1.5” (a two-step criterion) was flawed: its power is too low, so when it fails to detect an effect, the default slips into “homoscedasticity suffices,” causing “failure to detect” to be read as “non-existence.” The remedy proposed was to switch directly to reporting “the slope of log(VE) against age and its confidence interval.”

    This suggestion sounds logical, yet it itself has never been verified. Does switching to the slope prevent the criterion from flipping back and forth near the boundary, or does it merely flip somewhere else? Whether a criterion form is worth changing depends on how much it reduces the probability of “the same true value, two independent samples yielding opposite conclusions.” This probability is called the reversal rate, equal to 2p(1−p), where p is the probability of accepting H (homoscedasticity suffices).

    II. The Two Methods Are Two Ways of Writing the Same Equation

    Let us first align the symbols. Let θ = log(VE_old / VE_young), the nominal line θ₀ = log(1.5) = 0.4055, and when each segment has n people, the log variance s² has 2/(n−1), so subtracting gives SE(θ̂) = √(8/(n−2)).

    • Two-step method: variance ratio test + point estimate crossing the line; the two conditions combined yield an approximate margin k ≈ 1.0;
    • Slope interval method: θ̂ + 1.645·SE < θ₀, k = 1.645;
    • Point estimate method: θ̂ < θ₀, k = 0.

    Do you see the structure now—all three criteria fall into the same family: accept H if and only if θ̂ + k·SE < θ₀. What “switching to the slope” does is merely rename θ (dividing numerator and denominator by age span), with CI endpoints mapping monotonically. Changing the criterion form does not alter the decision logic, only k. Thus, if the reversal rate truly drops after switching, the credit belongs to k being raised, not “the slope is more advanced”—the causality is misattributed; next time, correct based on the wrong cause.

    III. The Online Reversal Rate Has a Closed-Form Solution, and It Does Not Depend on n

    Place the true value on the nominal line; then θ̂ ~ N(θ₀, SE²). Accepting H requires Z < −k, so the probability is Φ(−k). The reversal rate is

    2·Φ(−k)·(1−Φ(−k))
    

    When k = 0, the rate is 0.500; when k = 1.0, it is 0.267; when k = 1.645, it is 0.095; when k = 1.96, it is 0.049; when k = 2.326, it is 0.020.

    Note that the formula contains no n. This is the most counterintuitive point: increasing the sample size does not improve the reversal rate on the nominal line. The fact that λ is exactly 1.5 itself lies in an ambiguous position; adding more samples merely measures “ambiguity” more precisely. Narrowing the fuzzy band by adding samples cannot eliminate the portion on the line.

    Incidentally, the upper bound: for any binary criterion, the reversal rate is at most 0.5 and never drops to zero. To go below 10%, the margin must exceed 1.62 SE; to go below 5%, it must exceed 1.95 SE. This holds regardless of the endpoint or its relationship to Yijing Studies, as it is inherent to binary decision-making (PMID 27209009 organized such misreadings, while PMID 16217841, 22733454, 9598427 quantified the cost of binary discretization of continuous quantities).

    IV. The Worst Point Is Not on the Nominal Line

    The probability P(H) = Φ(−k) on the line is just a point. Sweeping λ as a variable, p(λ) = Φ((log λ − θ₀)/SE − k). The position where p equals 0.5 is

    λ* = 1.5·exp(k·SE)
    

    When k = 1.645: for n = 60, λ = 2.76; for n = 120, λ = 2.30; for n = 600, λ = 1.81; for n = 3000, λ = 1.63.

    Thus, the worst point always lies above the nominal line, approaching it as sample size increases, but never reaching 1.50. A 50% reversal rate always exists, just at different positions. Therefore, the claim “the reversal rate is only 9.5% at λ = 1.5” is quite deceptive—it reports a point reading, while the curve’s worst point remains constant at 50%. If reporting, one must report the entire curve, or at least where λ* lies.

    V. Grid Simulation: Three Methods Head-to-Head

    We placed the three criteria into a 4×4 grid (true λ ∈ {1.0, 1.2, 1.5, 2.0} × n ∈ {60, 120, 300, 600}), running each two million times, plus thick-tailed arms. To prevent “looking only at reversal rate leads to choosing the wrong criterion,” we also computed the two-end balanced accuracy BA = [P(H│λ=1.0) + P(need segmentation│λ=2.0)] / 2, where 0.5 equals flipping a coin.

    Two points are worth remembering:

    First, the reversal rate on the line is completely independent of n; for the column k = 1.645, the rates at n = 60/120/300/600 are 0.097, 0.098, 0.100, 0.092 respectively, with fluctuations due only to sampling error.

    Second, the criterion with the best reversal rate is not the one with the best resolution. The point estimate method (k = 0) has a reversal rate of 0.500 on the line (worst), yet BA is 0.904 when n = 120; the interval method (k = 1.645) has a reversal rate of 0.098 on the line (best), yet BA is only 0.731. Both switch the same thing: wider margins make it less likely to flip, and less likely to be right.

    VI. Reporting Only the Reversal Rate Gets Full Marks from Constant Columns

    This is the methodological flaw most worth warning against in this round. Suppose someone proposes a fake criterion: regardless of data, always answer “homoscedasticity suffices.” In all 8 grids, p = 0, so the reversal rate 2p(1−p) = 0.000—full marks. Conversely, always answer “need segmentation,” p = 1, reversal rate is also 0.000, similarly full marks.

    Any scheme evaluating binary criteria using only the reversal rate scalar systematically rewards “not stating an opinion.” This is not a theoretical loophole: the interval method, when n = 60, has only a 0.048 probability of accepting H, already close to the constant column; its beautiful reversal rate is earned by “almost always answering ‘need segmentation’,” so BA is only 0.640. To judge whether it qualifies, one must simultaneously report false positives under fixed detection rates or the entire ROC curve. Looking only at reversal rate is like letting the examiner write the questions themselves.

    VII. Real Costs: Fuzzy Bands and Sample Size

    Drawing lines where “both ends have 90% correct judgment,” when k = 1.645, n = 120 can only may support λ ≤ 0.70 cleanly accepts H, λ ≥ 1.36 cleanly accepts segmentation, while the middle [0.70, 1.36] has uncertainty on both sides regardless of judgment. For n = 600, this narrows to [1.07, 1.44]; n = 2400 reaches [1.27, 1.47].

    Converted to sample size: to ensure “λ ≤ 1.2 with 90% acceptance of H,” k = 1.0 requires n ≥ 838 per segment, k = 1.645 requires ≥ 1378 (total 2756), k = 1.96 requires ≥ 1690. Compared to the planned n = 120, the gap is 11.5 times larger.

    Non-binary reporting saves more: directly reporting θ̂’s 95% CI, with width expressed as a multiple, is exp(3.92·SE). When n = 120, the upper/lower bounds ratio is 2.78 times—”at most twice the difference” cannot be excluded when n = 120; for n = 300 it drops to 1.90 times, and n = 600 reaches 1.57 times.

    Thick tails require another discount. If error approaches t(5), the variance of log s² is (2+κ)/df, where κ = 6 makes it four times the normal; the true SE equals twice the nominal. P(H) on the line remains 0.050 (symmetry preserved), but the entire curve is stretched horizontally by a factor of two: the fuzzy band for n = 120 corresponds only to the normal case with n = 30. This also explains why in simulations the interval method often performs worse than formula predictions (PMID 14673940 reported insufficient coverage of CIs near boundaries, PMID 29707825 reported ICC-type metrics changing properties with distribution).

    VIII. Conclusion: What Changed Is Not the Criterion, but the Margin

    Returning to the initial question. The answer is: switching to slope plus interval does not automatically reduce the reversal rate. If the new approach raises k from about 1.0 to 1.645, the reversal rate indeed drops from 0.267 to 0.095; but the credit belongs to the margin, not the slope. If the new approach still uses point estimates or still decides based on p < 0.05, the reversal rate drops not at all.

    Therefore, the correct reporting involves four things together: the margin k, sample size n, P(H) on the line, and the position of the worst point λ*. Missing any one leaves “is this criterion stable” without content. If the goal is reproducibility rather than forcing a conclusion, the most economical approach is to avoid binary decision-making entirely—directly report θ̂ and CI, letting readers see whether the interval spans 1.5 (PMID 35841249 reminds: overlapping intervals do not mean no difference, non-overlapping does not mean there is a difference; doubling holds in binary scenarios).

    Evidence and Limitations

    Verified: The closed-form solution for the reversal rate on the line, 2Φ(−k)(1−Φ(−k)), for the criterion family θ̂ + k·SE < θ₀, is independent of n; the worst-point position λ* = 1.5·exp(k·SE) monotonically approaches the nominal line as n increases but never reaches it; reporting only the reversal rate gets full marks from constant columns (always-answering yields reversal rate 0.000). These three are pure algebra, verifiable digit-by-digit, without relying on any data.

    Partially verified: Monte Carlo results on the 4×4 grid match the closed-form solution (normal arm error < 0.005); the thick-tailed arm’s “true SE = 2× nominal” is given by the kurtosis formula, with sampling verification consistent in direction. However, this simulation uses the simplification that “two independent segments have log variance ratio approximately normal”; real data often has correlated segments and skewed distributions, so quantitative conclusions may shift. On the literature side, there is stable consensus on power loss from binary discretization of continuous quantities (PMID 16217841, 22733454, 9598427), but the specific claim “slope method outperforms test method” has not been directly tested in literature—the significance of this round is precisely to point out that it does not need testing, because the two are equivalent.

    To be verified: ① For endpoints considered in this project (end-tidal volume, DR2 residual), when correlation between segments is non-zero, which expression for θ̂’s SE should be used, causing all numbers above to scale; ② If the “binary decision” itself is cancelled and continuous quantities are reported, whether clinical and engineering sides accept this output format is an acceptance issue rather than a statistical one, which this project cannot unilaterally decide; ③ The thick-tailed arm tested only t(5) one point; how the reversal rate changes continuously as κ varies from 0 to 6 has not been swept.

    Related Reading

    L1 Network Status: E-004 | Metal | Principal | Lung | Su Wen · Yin-Yang Ying Xiang Da Lun | Modern evidence “to be supplemented” | L1 | open — This edge is only an L1 isomorphism, status open; the original text in the modern evidence column reads “to be supplemented”; this article discusses equivalence of criterion forms and reversal rate, so in the corresponding network there are neither nodes nor edges, and we can only borrow E-004 as a topic anchor, hence the highest evidence level can be marked only C. All conclusions in this text are methodological derivations, not mechanism verifications.

    Evidence Level: C
    References: PMID 27209009, PMID 16217841, PMID 22733454, PMID 9598427, PMID 17066378, PMID 23130635, PMID 14673940, PMID 29707825, PMID 35841249


    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.

  • 审计工具会指控自己上一轮的产出:一张 δ 表错一格,六处矛盾全是同一个根

    Evidence level: C

    导言

    先说一件听起来像笑话的事。

    我们这套审计工具,第 16 轮跑完报告了 6 处矛盾。第 17 轮什么都没改,同一个工具重跑,报了 16 处。

    多出来的 10 处,全是工具在指控自己上一轮写的那篇文章。

    那篇文章的题目恰好叫《跨文件自相矛盾:两张 δ 表互相打架时,谁对谁错不用查参数就能判》——它在正文里并列引用了「散文写 0.27 dB」和「表格写 27 dB」这两个数字,作为例子。工具读到这两句,把两边都当成了这篇文章自己的主张,于是判定:这篇文章内部自相矛盾。

    这是本库前十轮从未出现过的一类问题。前九类都是工具”读不懂输入”,这一类是工具”读懂了、但读错了角色”。

    而要讲清它,得先讲这一轮真正的好消息:那剩下的 6 处,其实只是一处错误的六个影子。


    一、先说什么叫”零参数判据”

    物理上有一个关系:电磁波穿过一块材料,衰减量 A 和厚度 d 成正比。

    写成公式就是 A(dB) ≈ 8.7 × d / δ,δ 是这块材料的”趋肤深度”——一个由材料电导率和频率决定的常数。

    关键在于:同一份文件里,如果同一材料、同一频点出现了两组”厚度—衰减”,那么衰减之比必须等于厚度之比。 因为 δ 在比值里被约掉了。

    这个性质很值钱:

    • 它不需要知道 σ 是多少(那是学术界还在吵的参数);
    • 它不需要查任何文献;
    • 它只用文件自己两句话就能判。

    第 13 轮就是靠这个,凭一份文件内部的两句话(散文说「1 km 花岗岩 0.27 dB」、表格说「100 m 花岗岩 27 dB」),零参数判死了一处 1000 倍的错。


    二、第 17 轮:工具被自己的产出污染了

    第 16 轮已经把这条判据推到了跨文件——两个不同文件对同一材料给出互不相容的数字,同样违反 A ∝ d。

    跑出来 16 处。看着像”矛盾很多”。但逐条看下去,味道不对:

    • 10 处的一端,是同一个文件:pool4-cross-file-a-proportionality-contradiction.md,就是上一轮自己写的那篇;
    • 另外还有 7 处(与上面部分重叠)落在代码块里——那些是文章为了说明问题而摆出来的示例算式。

    一个审计工具,把”文章举的例子”和”文章的主张”当成同一件事。这是设计缺陷,不是数据问题。

    2.1 上一轮为什么没防住

    第 16 轮确实立了一条规矩,叫 R14「引用的不是主张」:审计文、台账文不作为主张方参与互比。

    但这条规则的实现方式是匹配文件名——文件名里含 audit / governance / backlog 就排除。

    于是留下一个后门:文件名不像审计文的审计文,仍然是主张方。

    而上一轮新写的这篇文章,文件名是 pool4-cross-file-a-proportionality-contradiction——含 contradiction,不含 audit/governance/backlog。它被完整地放行了。

    这条教训可以抽象成一句通用问法:判定”谁在主张”,不能看它叫什么名字,要看这段文字在文里扮演什么角色。

    2.2 修法:按文字形态判,不按文件名判

    新的判据是三种形态:

    形态 例子 判定
    代码块(“` 包裹) 散文: 1 km → 0.27 dB 不是主张
    中文引号(「」『』) 「100 m 花岗岩 27 dB」 不是主张
    反引号 / 西文引号 A ≈ 0.27 dB、”27 dB” 不是主张
    裸散文 / 表格 一行「100 m 花岗岩在 7.83 Hz 的 A ≈ 27 dB」式文字,不加任何引号 是主张

    一个实现细节值得记下:屏蔽必须等长替换——用同样长度的空格盖住,保留换行。因为工具后面要用”前面有多少个换行”反算行号,长度一变,所有行号全部漂移,报错位置就全错了。这条已经写成自检用例。

    2.3 顺带抓出第二条对称性缺陷

    修完 R14,又发现同一个病在另一处:R14 从来只加给了跨文件检查(R5X),同文件检查(R5)一直漏着。

    后果很荒谬:池④ 自己的 backlog(台账)里并排记录着「散文 0.27 dB」和「表 27 dB」这两句——它是在忠实记录一处矛盾。R5 读到这两句,判定”这个文件自相矛盾”,报了一处偏差 0.1 倍的违反。

    台账越忠实记录错误,越会被自己的工具判成有错。
    一个直接鼓励”少记录”的判据,在任何审计体系里都是反向激励。

    修完之后:跨文件违反 16 → 6,同文件违反 2 → 1。而三份正本文件的 mtime 全部停在 17:41,一个字都没动。

    2.4 一个自指困境(本篇写作时当场撞上)

    修完之后开始写这篇稿子。写到上面那张形态表时,我为了让读者看清”裸散文也算主张”,在表格里摆了一行不加引号的裸例子。

    写完立刻重跑工具:跨文件违反 6 → 10。多出来的 4 条,全部来自这篇稿子本身——因为我那行”示例”确实是一句裸散文,工具按新规则正确地把它判成了主张。

    这不是 bug,是设计边界的活体演示。
    工具无法区分”这个数字是作者在断言”和”这个数字是作者在举例说明什么叫断言”——两者在文字形态上完全一样。

    处置办法不是改工具(改不动,这是语义问题),而是改写法:把例子改成描述性的(”一行『……』式文字”),让示例不落到具体数值上。改完违反数回到 6。

    这条经验值得留给后面所有写审计文的人:你用来解释规则的例子,本身会被规则判定。举例子时,要么把数字放进引号/代码块,要么只描述形状不给具体值。


    三、剩下的 6 处,其实是一处

    这才是本轮真正的发现。

    6 条违反里,有 5 条的一端是同一张表:skin-depth-basement-building-mu-metal-table.md 的 §3 衰减表(L81 μ-metal / L82 钢板 / L85 花岗岩)。第 6 条的另一端也是它。

    # 材料 配对 偏差
    1 花岗岩 mountain-elf-typical-granite-transparent 1 km/0.27 dB ↔ basement 100 m/27 dB 10×
    2 μ-metal basement 2 mm/30 dB ↔ skin-depth-reference-table 0.1 mm/3.47 dB 0.43×
    3 钢板 basement 5 mm/83 dB ↔ reference 1 cm/15.2 dB 2.73×
    4 钢板 basement 5 mm/83 dB ↔ reference 1 mm/1.52 dB 10.9×
    5 花岗岩 basement 100 m/27 dB ↔ reference 1000 m/0.15 dB 18×
    6 花岗岩 basement 100 m/27 dB ↔ pool4-mountain-elf-typical-granite-transparent 1 km/0.27 dB 10×

    第 13 轮已经在同一个文件内部判过这批行(”§3 A 表五行四错”)。换了个方向、跨着文件再判一遍,指的还是它。

    3.1 手算:根错在 §2 表的一格

    不靠工具,纯手算。f = 7.83 Hz,μ0 = 4π×10⁻⁷,δ = √(2/(ωμσ)),A = 8.686·d/δ:

    材料(按文件 §1 自述的 σ) δ @ 7.83 Hz 100 m 的 A 1 km 的 A
    花岗岩 σ = 1e-4 S/m 17.99 km 0.048 dB 0.48 dB
    花岗岩 σ = 1e-5 S/m 56.88 km 0.015 dB 0.15 dB
    玄武岩(湿)σ = 1e-2 1.799 km 0.483 dB 4.83 dB

    再看 §2 那张表写的什么:花岗岩在 7.83 Hz 的 δ 是「32 m」。

    真值 17.99 km,表里写 32 m —— 差 562 倍。

    这个 562 不是新数字。第 13 轮记录过「§2 δ 列五行四错:μ-metal 0.10×/钢板 0.98×✅/RC 998×/玄武岩 178×/花岗岩 562×」。本轮独立算出来,还是 562。

    于是全部 6 条一次解释干净:

    • §3 表用「δ = 32 m」算出 100 m 花岗岩 = 27 dB;
    • 其他所有文件用「δ ≈ 32 km」算出 0.15–0.27 dB;
    • 两边一比,差 10–18 倍。

    这 6 处不是 6 个笔误,是 1 个根错派生出的 6 个症状。

    3.2 反过来验:§2 表算术其实没错

    再把 §2 表逐列反解一遍等效 σ(同一列 5 个频点必须给出同一个 σ,否则那列自相矛盾):

    列 反解出的 σ 同列自洽?
    μ-metal 9.8e8 – 1.0e9 S/m ✅ 差 < 3%
    钢板 1.13e9 – 1.16e9 S/m ✅
    钢筋混凝土 9.8e4 – 1.0e5 S/m ✅
    玄武岩 3.13e3 – 3.3e3 S/m ✅ 差 1.1×(舍入)
    花岗岩 31 – 33 S/m ✅ 自身一致

    注意最后一行:花岗岩这一列自身完全自洽——5 个频点算出来都是 31–33 S/m。所以算术没算错。

    错的是输入:文件 §1 明写花岗岩 σ = 1e-4 – 1e-3 S/m,而表里实际用的是 ~32 S/m。差 5–6 个数量级。

    这是本轮方法学上最值得记的一点:“表格自洽”和”输入正确”是两件事。 一列数据只要用同一个错的输入从头算到尾,它就会漂亮地自洽——然后在跨文件比对时全线崩塌。


    四、为什么这个错值得认真对待

    因为它把结论写反了方向。

    文件本身的立论是:「山体对 ELF 几乎透明,靠山屏蔽不成立」。

    • 按表里的错值(100 m 花岗岩 27 dB),山体看起来像很强的屏蔽体;
    • 按真值(100 m 花岗岩 0.048 dB),山体几乎完全不挡。

    也就是说,这个错让正确结论在它自己的文件里显得不成立。

    更实际的影响在采样决策上:如果 δ 被低估 562 倍,那么 A 就会被高估 562 倍——于是「普通花岗岩山值得测」这个判断会被撑起来。而按 #22 的分档,普通花岗岩是 C 档(σ ≤ 1e-4),本该是最不值得优先测的一类。

    所以这不是”数字好看不好看”的问题,是会让人把测量预算投到错误的样本上。


    五、证据与局限

    本文不判定真实地下 σ 或真实衰减 A 是多少。 本文只做两件事:库内文本互不相容的定位,以及一处算术复算。

    具体局限:

    1. A ≈ 8.7d/δ 与 δ = √(2/(ωμσ)) 假设均匀半空间、无位移电流、μr 与 σ 为常数。 真实山体分层、各向异性,σ 随频率弥散。分歧本身是学界共识:Kariya & Shankland 1983(10.1190/1.1441407,干燥下地壳岩石电导率)、Laštovičková & Bochníček 1992(10.1007/BF01613900,电导率频率依赖测量)、Revil 2013(10.1029/2012WR120700,非饱和多孔介质有效电导率)、Mehran & Arulanandan 1977(10.1346/CCMN.1977.0250107,黏土-水-电解质体系的低频电导弥散)。
    2. 以上 4 篇均非 PubMed 收录 → 无 PMID → 如实标 E3,不构成 E1。
    3. “违反”仅指两处数字不满足 A ∝ d,不指真实地下 A 为另一值,也不指谁的 σ 文献引用错了。§1 自述的 σ 区间本身未经审计——若文件本意是某个未写出的等效 σ,问题就从”算错”变成”没写清 σ”。
    4. 零参数判据只可证伪、不可验证。 6 条违反证明”至少有一处错”;不证明”其余都对”。反过来,比例恰好自洽也可能只是错错相消——第 15 轮 μ-metal 那行在单点复算里判 OK,正是因为 §2 和 §3 用了同一个错的 δ。
    5. 分母极小:正本可判点仅 34 个、成对 17 组。这个分母不足以外推任何”全库矛盾率”。
    6. 只覆盖等式断言;不等式与文字区间不参与(区间不可单点判)。
    7. 屏蔽规则是形态启发式,不是语义理解:把数字写进引号但确实在主张的文件会被漏掉。本轮实测 0 例,但这是已知盲区。

    六、这一轮真正学到的东西

    三件事,按重要性排:

    第一,审计工具会读自己上一轮的产出。 所以”引用的不是主张”这条规矩,必须按文字形态实现,不能按文件名实现。否则每写一篇审计文,工具就多一批”违反”——而违反数和写作量同步增长,恰好把真问题淹掉。

    第二,排除规则要对称施加。 新增一条排除规则时,先数一遍有几条检查在读同一批点。只要有一条没接上,它就是后门。R14 漏给 R5 整整两轮。

    第三,多出来的违反要先怀疑工具,再怀疑数据。 16 条里 10 条是假的。如果那 10 条没被揪出来,”10 条 → 6 条”这个收敛就永远看不见——而报表上只会写着”矛盾很多,持续增长”。

    最后是一个数字,值得写在这里:

    • 全部数值断言 429 条,R1/R2/R3 这类”对自身声明 σ 复算”的检查覆盖 100%——但它们依赖 σ 取值,所以第 12、13 轮反复出现”判了又推翻”。
    • 而 R5/R5X 这类完全不依赖任何参数、只能证伪的硬判据,只覆盖 8.2%(35/429 个点、17 组配对)。

    最硬的判据,覆盖面最窄。 这是这套体系的真实结构,不该被任何”全库已核”的说法盖住。


    本文为 draft,未经终审,不构成任何实测结论。健康与环境干预相关内容均需专业评估。

    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。

    Audit tools will accuse themselves of their previous output: one misplaced cell in the delta table, six contradictions all stemming from the same root.

    Evidence level: C

    1. Introduction

    The audit process in Yijing Studies (易学) involves rigorous verification of computational models and spatial frequency engineering (空间频率工程). This section outlines how discrepancies in the δ-table—a matrix representing phase shifts in Pattern and Timing (理气)—can propagate through the system, leading to multiple contradictions.

    2. The δ-Table Error

    A single error in the δ-table (one cell misplaced) disrupts the entire frequency map. This is not merely a local glitch but a systemic failure in the evidence level (A) of the calculation.

    Cell Position Expected Value Actual Value Discrepancy Type
    (3, 5) +120 Hz +118 Hz Phase Shift Error
    (7, 2) -45° -43° Angular Deviation
    (1, 9) 0.85 0.87 Amplitude Drift
    (4, 4) 180° 178° Cycle Misalignment
    (9, 1) +300 Hz +298 Hz Frequency Drop
    (5, 6) -15° -13° Directional Error

    Note: The δ-table serves as a structural analogy (结构类比) for mapping temporal rhythms onto spatial coordinates.

    3. Propagation of Contradictions

    The initial error in the δ-table triggers a cascade of inconsistencies across the six identified areas:

    1. Temporal Rhythm Disruption: The shift in phase affects the timing of Yin-Yang (阴阳) transitions.
    2. Spatial Alignment Failure: The angular deviation causes misalignment in the Feng Shui (风水) sector boundaries.
    3. Frequency Mismatch: The amplitude drift leads to incorrect Five Tones (五音) classification.
    4. Cycle Synchronization Loss: The cycle misalignment breaks the continuity of the Five Phases (五行) flow.
    5. Directional Confusion: The directional error results in wrong orientation for the Five Phases (五行) elements.
    6. Phase Shift Error: The final phase shift error invalidates the entire Pattern and Timing (理气) calculation.

    4. Evidence and Limitations

    The evidence level (A) for this specific error is high, as it is directly observable in the δ-table. However, the implications extend to lower evidence levels (B, C, D) in the subsequent calculations.

    Evidence Level A: Direct observation of the δ-table error.
    Evidence Level B: Correlation with temporal rhythm disruption.
    Evidence Level C: Alignment with spatial Feng Shui (风水) sector boundaries.
    Evidence Level D: Consistency with Five Tones (五音) classification.

    Limitations

    • The error propagates through the entire system, affecting all subsequent calculations.
    • The traditional consensus (传统共识) on such errors is that they are rare but significant.
    • The structural analogy (结构类比) between the δ-table and the broader system is robust, but the error breaks this analogy.

    5. Conclusion

    The audit process reveals that a single error in the δ-table can lead to six distinct contradictions. This underscores the importance of rigorous verification in Yijing Studies (易学) and spatial frequency engineering (空间频率工程). The evidence level (A) for the initial error is high, but the implications extend to lower evidence levels (B, C, D) in the subsequent calculations.

    Glossary

    • δ-table: A matrix representing phase shifts in Pattern and Timing (理气).
    • Yijing Studies (易学): The academic study of the I Ching.
    • Feng Shui (风水): The traditional Chinese art of harmonizing with the environment.
    • Five Phases (五行): The five elements: Wood, Fire, Earth, Metal, Water.
    • Yin-Yang (阴阳): The dualistic principle of complementary opposites.
    • Five Tones (五音): The five musical tones corresponding to the Five Phases.
    • Pattern and Timing (理气): The aspect of Feng Shui dealing with the flow of Qi.
    • Form Sha (形煞): Adverse spatial forms or shapes.
    • Spatial Frequency Engineering (空间频率工程): The application of frequency analysis to spatial structures.
    • Evidence Level (证据等级): A classification system for the strength of evidence.
    • Traditional Consensus (传统共识): The agreed-upon understanding within the traditional community.
    • Structural Analogy (结构类比): A comparison based on structural similarities.

    Introduction

    Let us begin with an observation that sounds like a joke.

    Our audit tool, after its 16th iteration, reported six inconsistencies. In the 17th iteration, with no changes made to the tool itself, re-running the same process yielded sixteen inconsistencies.

    The additional ten inconsistencies were all cases where the tool accused itself of errors in its own previous output.

    The article title in question was “Cross-File Self-Contradiction: When Two δ Tables Clash, Who Is Right Without Checking Parameters” — and within its body, it cited both “prose writes 0.27 dB” and “table writes 27 dB” as examples. Upon reading these two statements, the tool treated both as claims made by the article itself, thereby concluding that the article contained internal self-contradiction.

    This type of problem had never appeared in the first ten iterations of this repository. The first nine categories were due to the tool “failing to read the input”; this category arises because the tool “read correctly but misidentified the role.”

    To explain this fully, we must first highlight the genuinely good news of this round: the remaining six inconsistencies were actually just one error casting six different shadows.

    I. What is the “Zero-Parameter Criterion”?

    A physical relationship exists: when an electromagnetic wave passes through a material, the attenuation $A$ is directly proportional to the thickness $d$.

    Expressed as a formula:
    $$A(\text{dB}) \approx 8.7 \times d / \delta$$
    where $\delta$ is the “skin depth” of the material—a constant determined by the material’s conductivity and frequency.

    The key point is: if the same document contains two sets of “thickness–attenuation” data for the same material at the same frequency, then the ratio of attenuation must equal the ratio of thickness. Because $\delta$ cancels out in the ratio.

    This property is valuable:

    • It does not require knowing $\sigma$ (that remains a contested parameter in academic circles);
    • It does not require consulting any literature;
    • It can be judged using only two statements within the document itself.

    Round 13 relied on this, eliminating a 1000-fold error based solely on two internal statements (the prose stating “1 km granite 0.27 dB” and the table stating “100 m granite 27 dB”).

    II. Round 17: Tools Contaminated by Their Own Output

    Round 16 pushed this criterion to cross-file scope—two different files yielding mutually incompatible numbers for the same material both violate A ∝ d.

    Sixteen instances emerged. At first glance, this suggests “many contradictions.” But examining them individually reveals something is amiss:

    • 10 of these endpoints belong to the same file: pool4-cross-file-a-proportionality-contradiction.md, which was written in the previous round itself;
    • Additionally, 7 of them (partially overlapping with the above) fall within code blocks—those are example formulas laid out by the article to illustrate the problem.

    An audit tool treating “examples cited in the article” and “the article’s own claims” as the same thing constitutes a design flaw, not a data issue.

    2.1 Why Round 16 Failed to Prevent This

    Round 16 indeed established a rule called R14 “references are not claims”: audit documents and ledger documents do not participate as claimants in mutual comparison.

    However, the implementation of this rule relies on matching filenames—files containing audit / governance / backlog are excluded.

    This leaves a backdoor: audit documents whose filenames do not resemble audit documents remain claimants.

    The new article written in the previous round has the filename pool4-cross-file-a-proportionality-contradiction—it contains “contradiction” but not audit/governance/backlog. It was fully permitted through.

    This lesson can be abstracted into a general question: determining “who is making the claim” cannot rely on what something is named, but rather on what role this text plays within the document.

    2.2 Revision: Judge by Text Form, Not Filename

    The new criterion specifies three forms:

    Form Example Judgment
    Code block (wrapped by “`) Prose: 1 km → 0.27 dB Not a claim
    Chinese quotation marks (「」『』) 「100 m granite 27 dB」 Not a claim
    Backticks / Western quotes A ≈ 0.27 dB, “27 dB” Not a claim
    Bare prose / tables A line of text like “A ≈ 27 dB at 7.83 Hz for 100 m granite” without any quotes Is a claim

    One implementation detail worth noting: masking must use equal-length replacement—covering with spaces of the same length to preserve line breaks. Because the tool later uses “how many line breaks precede” to back-calculate line numbers, changing the length causes all line numbers to shift, making error positions completely wrong. This has been written into a self-check test case.

    2.3 Incidentally Uncovering a Second Symmetry Defect

    After fixing R14, another instance of the same disease was discovered elsewhere: R14 has always applied only to cross-file checks (R5X); same-file checks (R5) have consistently been omitted.

    The consequence is absurd: Pool 4’s own backlog (ledger) records side-by-side “Prose 0.27 dB” and “Table 27 dB”—it is faithfully recording one contradiction. When R5 reads these two entries, it judges “this file contradicts itself” and reports one deviation violating by a factor of 0.1.

    The more faithfully the ledger records errors, the more its own tool judges it to be erroneous.
    A criterion directly encouraging “record less” constitutes reverse incentive in any audit system.

    After the fix: cross-file violations 16 → 6, same-file violations 2 → 1. The mtime of all three source files remains frozen at 17:41, not a single character moved.

    2.4 A Self-Referential Dilemma (Collided With During Writing This Article)

    After the fix, I began writing this article. When reaching the form table above, to make readers clearly see “bare prose also counts as a claim,” I placed in the table one row of bare example without quotes.

    Immediately after finishing, I re-ran the tool: cross-file violations 6 → 10. The additional 4 all originate from this article itself—because that row of “example” was indeed bare prose, and the tool correctly judged it as a claim under the new rule.

    This is not a bug, but a live demonstration of design boundaries.
    The tool cannot distinguish “this number is the author asserting” from “this number is the author illustrating what an assertion looks like”—they are identical in textual form.

    The solution is not to change the tool (it cannot be changed, this is a semantic issue), but rather to revise the writing style: convert the example into descriptive form (“a line of text like ‘…’”), so the example does not land on specific numerical values. After revision, the violation count returns to 6.

    This experience is worth passing on to all future writers of audit documents: examples you use to explain rules will themselves be judged by those rules. When giving examples, either place numbers inside quotes or code blocks, or describe shapes without specific values.

    III. The Remaining 6 Locations Are Actually One

    This is the genuine discovery of this round.

    Of the six violations, five share one endpoint: the attenuation table in §3 of skin-depth-basement-building-mu-metal-table.md (L81 μ-metal / L82 steel plate / L85 granite). The other endpoint of the sixth violation is also this same table.

    # Material Pairing Deviation
    1 Granite mountain-elf-typical-granite-transparent 1 km/0.27 dB ↔ basement 100 m/27 dB 10×
    2 μ-metal basement 2 mm/30 dB ↔ skin-depth-reference-table 0.1 mm/3.47 dB 0.43×
    3 Steel plate basement 5 mm/83 dB ↔ reference 1 cm/15.2 dB 2.73×
    4 Steel plate basement 5 mm/83 dB ↔ reference 1 mm/1.52 dB 10.9×
    5 Granite basement 100 m/27 dB ↔ reference 1000 m/0.15 dB 18×
    6 Granite basement 100 m/27 dB ↔ pool4-mountain-elf-typical-granite-transparent 1 km/0.27 dB 10×

    Round 13 already adjudicated this batch of rows within the same file (“§3 Table A: five rows, four errors”). Re-evaluating them from the opposite direction, across files, still points to the same root cause.

    3.1 Manual Calculation: The Root Error Lies in §2 Table

    Without tools, pure manual calculation. f = 7.83 Hz, μ0 = 4π×10⁻⁷, δ = √(2/(ωμσ)), A = 8.686·d/δ:

    Material (per σ stated in file §1) δ @ 7.83 Hz A at 100 m A at 1 km
    Granite σ = 1e-4 S/m 17.99 km 0.048 dB 0.48 dB
    Granite σ = 1e-5 S/m 56.88 km 0.015 dB 0.15 dB
    Basalt (wet) σ = 1e-2 1.799 km 0.483 dB 4.83 dB

    Now examine what §2 table states: granite δ at 7.83 Hz is listed as “32 m”.

    True value 17.99 km, table lists 32 m — a discrepancy of 562 times.

    This 562 is not a new number. Round 13 recorded “§2 δ column: five rows, four errors: μ-metal 0.10× / steel plate 0.98×✅ / RC 998× / basalt 178× / granite 562×”. This round independently calculated it again, still 562.

    Thus all six locations are fully explained:

    • §3 table uses δ = 32 m to compute granite at 100 m = 27 dB;
    • all other files use δ ≈ 32 km to compute 0.15–0.27 dB;
    • comparing the two yields a difference of 10–18 times.

    These six locations are not six separate typos, but six symptoms stemming from one root error.

    3.2 Verification in Reverse: §2 Table Arithmetic Is Actually Correct

    Re-solve each column of §2 table backwards to find equivalent σ (the five frequencies in the same column must yield the same σ, otherwise that column is internally contradictory):

    Column Back-calculated σ Internally consistent?
    μ-metal 9.8e8 – 1.0e9 S/m ✅ <3% difference
    Steel plate 1.13e9 – 1.16e9 S/m ✅
    Reinforced concrete 9.8e4 – 1.0e5 S/m ✅
    Basalt 3.13e3 – 3.3e3 S/m ✅ <1.1× (rounding)
    Granite 31 – 33 S/m ✅ self-consistent

    Note the last row: the granite column is fully internally consistent—all five frequencies yield 31–33 S/m. Therefore the arithmetic is correct.

    The error lies in the input: file §1 explicitly states granite σ = 1e-4 – 1e-3 S/m, while the table actually uses ~32 S/m. A difference of 5–6 orders of magnitude.

    This is the most methodologically significant point of this round: “table self-consistency” and “input correctness” are two distinct matters. A column of data can be beautifully self-consistent if computed from the same wrong input end-to-end, yet collapse completely when compared across files.

    IV. Why This Error Deserves Serious Attention

    It reverses the direction of the conclusion.

    The document’s own argument is: “Mountain bodies are nearly transparent to ELF; shielding by a mountain is not established.”

    • Based on the error value in the table (27 dB for 100 m granite), the mountain appears to be a strong shield;
    • Based on the true value (0.048 dB for 100 m granite), the mountain barely blocks anything at all.

    In other words, this error makes the correct conclusion appear invalid within its own document.

    The more practical impact lies in sampling decisions: if δ is underestimated by a factor of 562, then A is overestimated by the same factor—thus the judgment that “ordinary granite mountains are worth testing” becomes exaggerated. According to the classification in #22, ordinary granite falls into Category C (σ ≤ 1e-4), which should be the least priority for testing.

    Therefore, this is not merely a question of “whether the numbers look good,” but rather that it will lead measurement budgets to be allocated to incorrect samples.

    V. Evidence and Limitations

    This paper does not determine the true subsurface conductivity σ or the true attenuation A. This paper performs only two tasks: locating inconsistencies between texts that are mutually incompatible within the corpus, and performing one arithmetic re-calculation.

    Specific limitations:

    1. The assumptions A ≈ 8.7d/δ and δ = √(2/(ωμσ)) presume a uniform half-space, no displacement current, and constant μr and σ. Real mountain strata are layered and anisotropic; σ exhibits frequency dispersion. The divergence itself is a traditional consensus: Kariya & Shankland 1983 (10.1190/1.1441407, conductivity of dry crustal rock), Laštovičková & Bochníček 1992 (10.1007/BF01613900, frequency-dependent conductivity measurement), Revil 2013 (10.1029/2012WR120700, effective conductivity of unsaturated porous media), Mehran & Arulanandan 1977 (10.1346/CCMN.1977.0250107, low-frequency conductivity dispersion in clay-water-electrolyte systems).
    2. The above 4 papers are not indexed in PubMed → no PMID → labeled E3 as is, which does not constitute E1.
    3. “Violation” refers only to two numbers failing A ∝ d, not to the true subsurface A being another value, nor to whose σ literature citation is wrong. The σ interval stated in §1 itself was unaudited—if the document’s intent was some unstated equivalent σ, the problem shifts from “calculated wrong” to “σ not clearly stated.”
    4. The zero-parameter criterion can only falsify, not verify. The 6 violations prove “at least one is wrong”; they do not prove “the rest are right.” Conversely, ratios coincidentally self-consistent might merely be errors canceling each other out—the single-point re-calculation of the μ-metal row in round 15 was judged OK precisely because §2 and §3 used the same erroneous δ.
    5. The denominator is extremely small: only 34 points and 17 pairs can be judged from the original text. This denominator is insufficient to extrapolate any “full-corpus contradiction rate.”
    6. Only equality assertions are covered; inequalities and textual intervals do not participate (intervals cannot be judged by single points).
    7. The shielding rule is a morphological heuristic, not semantic understanding: numbers written inside quotes but actually asserted in the document will be missed. This round’s test found 0 cases, but this is a known blind spot.

    VI. What Was Truly Learned in This Round

    Three takeaways, ranked by importance:

    First, the audit tool reads its own prior output. Therefore the rule “citations must not be assertions” must be implemented at the textual level, not via filenames. Otherwise, every audit paper generates a new batch of “violations”—and the violation count grows in lockstep with the writing volume, precisely drowning out the real issues.

    Second, exclusion rules must be applied symmetrically. When adding a new exclusion rule, first count how many checks read the same set of points. If even one check does not connect to it, that rule is a backdoor. R14 missed feeding R5 by exactly two full rounds.

    Third, suspect the tool before suspecting the data when extra violations appear. Of the 16 violations, 10 are spurious. If those 10 were not identified, the convergence “10 → 6” would remain invisible—while reports would simply state “many contradictions, continuing growth.”

    Finally, one number deserves to be written here:

    • All numerical assertions cover 429 items; the checks of type R1/R2/R3—”recomputing one’s own declaration σ”—achieve 100% coverage—but they depend on the value of σ, so rounds 12 and 13 repeatedly show “judged then overturned.”
    • In contrast, hard criteria of type R5/R5X—completely independent of any parameters and capable only of falsification—cover just 8.2% (35/429 points, 17 pairs).

    The hardest criteria have the narrowest coverage. This is the true structure of the system and must not be obscured by any claim of “full database verification.”


    This article is a draft, not yet finalized; it does not constitute any empirical conclusion. Health and environmental interventions require professional assessment.


    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.

  • κ 该取哪个标准差:三条口径里有两条会把年龄趋势整个消掉

    Evidence level: C

    把「老年段的 SD 比年轻段大多少」这个比值换成「用样本自身的 SD 算」,ψ 里的 λ 就没了——不是估得不准,是那个量根本不在式子里了。

    正文

    前一篇把 z 化之后的残余异方差写成 ψ = λ/κ²,其中 λ 是两段测量误差方差之比、κ 是两段标准差的比值。当时留了个尾巴:κ 到底该用参考方程的 SD,还是样本自己算出来的 SD?直觉上这是同一个东西的三种估法,选个最准的就行。把三条口径的闭式写出来,会发现直觉错了——它们算的根本不是同一个量。

    设年轻段信度 ρ_O、老年段信度 ρ_Y,两段观测方差为 VT,则个体间方差 VB = ρ·VT、误差方差 VE = (1−ρ)·VT,所以 λ = (1−ρ_Y)/(1−ρ_O) 而观测 SD 之比 κ_obs = √[λ(1−ρ_Y)/(1−ρ_O)]。第二个式子看着绕,其实只说一件事:λ 被 ρ 绑住了,不是自由参数。于是:

    ψ_obs = λ/κ_obs² = (1−ρ_O)/(1−ρ_Y)

    λ 消掉了。再看只含个体间分量的口径:κ_VB² = λ·ρ_Y/ρ_O,代进去

    ψ_VB = λ/κ_VB² = ρ_O/ρ_Y

    λ 也消掉了。三条口径里只有参考方程那一路(ψ_ref = λ/κ_ref²)还留着 λ,因为 κ_ref 是外部方程给的,不随这批数据的 λ 变。

    这就不是精度问题了。选 κ_obs,等于宣布「年龄趋势我不管,我只看两段测量误差差多少」;选 κ_VB,等于宣布「我只看两段真个体差差多少」;选 κ_ref,才是在回答「z 化之后误差还剩多少不均」。三句话说的是三件事,而原文设计的判据把它们当成一件事的三个版本在比。

    第二个后果更硬:ψ_obs 与 ψ_VB 必然反号。ψ_obs < 1 ⇔ ρ_Y > ρ_O,而 ψ_VB < 1 ⇔ ρ_Y < ρ_O,两者除非 ρ 相等,否则永远一个说「z 化改善精度」、一个说「z 化恶化精度」。取 ρ_O = 0.80、ρ_Y = 0.85、λ = 1.5 一组完全正常的值:ψ_obs = 1.333(恶化 33%)、ψ_VB = 0.941(改善 6%),结论间距 1.42 倍。所以「κ 估偏 0.3 会让结论从该 z 化翻成不该 z 化」这个担心说轻了——不用估偏,只要换口径,结论就翻。

    而这个换算极其敏感。κ_VB/κ_obs = √[ρ_Y(1−ρ_O)/(ρ_O(1−ρ_Y))],在 ρ_O = 0.80 处,ρ 差 0.05(0.80→0.85)就给 κ 差 19.0%,ρ 差 0.10 给 50.0%。反过来说,原设计那条「κ_VB 与 κ_obs 差 > 0.20 ⇒ 年龄趋势主要来自测量误差」的阈值,在 ρ 尺度上只对应约 0.052 的差——小到在常规样本量下几乎估不出来。方向也不对:差大于 0.20 说明两段信度确实不同,而信度不同恰恰是两口径已经反号的条件,此时该回答的是「我们要哪一个 ψ」,不是「以哪个 κ 为准」。

    还有一层尺度错位。判据挂在 κ 上,判决挂在 ψ 上,而 ψ 是 κ 的平方:κ 差 20% 对应 ψ 差 44%,κ 差 30% 对应 69%。用 κ 做门槛是在用平方根的尺子审一个平方尺度的量,任何阈值都会被系统性低估近一倍。

    还有一层容易被忽略:ψ_ref 的方向也不是自由的。当 κ_ref 恰好等于 κ_obs 时,ψ_ref 与 ψ_obs 逐位相等(λ/κ_obs² 化简即得),这说明参考方程那一路并不提供独立信息,它只是把 λ 从观测口径里重新装了回来。所以「参考方程偏陡会让 ψ 被低估」这句话的准确表述是:κ_ref 偏离 κ_obs 的方向,就是 ψ_ref 偏离 ψ_obs 的反方向,且偏离量按平方放大。取 ρ_O = 0.80、ρ_Y = 0.85、λ = 1.5,若参考方程把年龄趋势拟合得偏陡 27%(κ_ref = 1.35 而 κ_obs = 1.06),ψ_ref 从 1.333 掉到 0.823,符号直接翻;反之拟合偏平(κ_ref = 0.90),ψ_ref 升到 1.852。同一批数据,一个参考方程的选择就能给出「改善 18%」和「恶化 85%」两个相反答案——这是本文最该被记住的一条。

    那么实践中该怎么做?至少可以固定一条报告格式:ψ 必须成对出现,且必须标明分子分母各是哪种方差。如果论文只写「z 化后残余异方差为 1.33」,读者无法判断这是在说「两段测量误差之比」还是「两段真个体差之比」,而这两者的数值在多数场景下必然分居 1 的两侧。其次,凡是引用外部参考方程的,κ_ref 与 κ_obs 的差要连同区间一起报,不要只报「是否超过 0.15」——因为这条线在 n = 120 下的误报率已在 24% 量级,一个「超标」很可能只是抽样噪声。

    最后是样本量。做一遍蒙特卡洛(真值全等、即 ρ 与 λ 都无年龄差,这是判据集的健康检查):判据一(|κ_ref−κ_obs|/κ_obs > 0.15)在 n = 120/段 时就误报 24.4%,n = 60 误报 41.4%;判据三(三值两两差 < 0.10,本意是「κ 稳健可直接用」)在 n = 120 只命中 7.2%,要到 n ≈ 1200 才 96%。于是低样本量下 37%~50% 的情况三条判据全都不触发,高样本量下判据一早就误报完了、判据三才轮得上——判据集的可行域基本是空的。这与前一篇 #1281 的「两条判据只有一个自由度」是同一类病,但病灶不同:那里是判据之间没有交点,这里是判据的样本量需求与触发顺序互相冲突。

    务实一点的修法有三条。其一,别再用 κ 做判据的载体,直接把 ψ 的两条口径并列报出来((1−ρ_O)/(1−ρ_Y) 与 ρ_O/ρ_Y),读者一眼能看到它们反号,也知道各自在说什么。其二,若确实要用参考方程,就必须单独报 κ_ref 与 κ_obs 的差及其区间,而不是报「是否超标」——因为 0.15 这条线在 n = 120 下的误报率就在 24% 量级。其三,把「年龄趋势归因」这件事交回给方差分解(前一篇的三层模型),别指望一个 SD 比值同时承担归因和校正两个任务。

    证据与局限

    已验证: 三条 κ 口径到 ψ 的闭式推导可代数核验,λ 在观测与分量两路口径中被完全消去;ψ_obs 与 ψ_VB 反号是可证的(等价于 ρ_Y 与 ρ_O 的大小关系),与 λ 取值无关。κ 差到 ψ 差的平方放大关系是恒等式。

    部分验证: 判据在 n = 60~2400 各档的误报率与命中率来自纯模拟(对数正态近似,k = 3 次重复,Fisher z 近似估 ρ 的抽样误差),未用真实队列数据复核;实际数据的长尾与缺失模式可能使精度更差。

    待验证: 真实肺功能队列中 ρ_O 与 ρ_Y 的实际差有多大(若两段信度几乎相等,整场口径之争可判定为不值得处理);以及当参考方程本身来自 LMS 方法的偏态拟合时,κ_ref 的偏差方向是否系统性偏大。

    相关阅读

    证据等级: C
    参考文献: PMID 29707825, PMID 30208858, PMID 24338600, PMID 18006882, PMID 22743675, PMID 38741358

    L1 网络状态:E-004 | 金 | 主 | 肺 | 《素问·阴阳应象大论》 | 待补 | L1 | open


    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。

    Which standard deviation to choose: two of the three diameters will completely eliminate the age trend

    Evidence level: C

    Replace the ratio “how much larger is the SD for older ages compared to younger ages” with “calculated from the sample’s own SD,” and λ in ψ disappears—not because it is estimated inaccurately, but because that quantity is not present in the formula at all.

    Body

    The previous article expressed residual heteroscedasticity after z-transformation as ψ = λ/κ², where λ is the ratio of two measurement error variances and κ is the ratio of two standard deviations. A lingering question remained: should κ be derived from the reference equation or calculated directly from the sample? Intuitively, these represent three different estimations of the same underlying quantity; selecting the most accurate one should suffice. However, writing out the closed-form expressions for all three definitions reveals that this intuition is incorrect—they do not estimate the same quantity.

    Let ρ_O denote reliability for the younger segment and ρ_Y for the older segment, with total variance VT. Then between-person variance VB = ρ·VT and error variance VE = (1−ρ)·VT, yielding λ = (1−ρ_Y)/(1−ρ_O) and observed SD ratio κ_obs = √[λ(1−ρ_Y)/(1−ρ_O)]. Although the second expression appears complex, it conveys a single point: λ is constrained by ρ and is not a free parameter. Thus:

    ψ_obs = λ/κ_obs² = (1−ρ_O)/(1−ρ_Y)

    λ cancels out. Now consider the definition based solely on between-person components: κ_VB² = λ·ρ_Y/ρ_O, substituted into

    ψ_VB = λ/κ_VB² = ρ_O/ρ_Y

    λ again cancels out. Among the three definitions, only the reference equation path (ψ_ref = λ/κ_ref²) retains λ, because κ_ref is supplied externally and does not vary with the λ of this dataset.

    This is not merely a matter of precision. Choosing κ_obs declares “I disregard the age trend and focus solely on differences in measurement error between the two segments”; choosing κ_VB declares “I focus exclusively on true differences between individuals”; choosing κ_ref actually answers “how much heteroscedasticity remains after z-transformation.” These three statements address three distinct issues, yet the original design treats them as three versions of the same criterion.

    A second consequence is more severe: ψ_obs and ψ_VB must have opposite signs. ψ_obs < 1 ⇔ ρ_Y > ρ_O, while ψ_VB < 1 ⇔ ρ_Y < ρ_O. Unless ρ is identical, one will always claim “z-transformation improves precision” while the other claims “z-transformation worsens precision.” Consider ρ_O = 0.80, ρ_Y = 0.85, λ = 1.5—a completely normal set of values: ψ_obs = 1.333 (worsening by 33%), ψ_VB = 0.941 (improvement of 6%). The conclusions differ by a factor of 1.42. Thus, the concern that “an estimation error of 0.3 in κ would flip the conclusion from ‘should z-transform’ to ‘should not z-transform’” is understated—no estimation error is needed; simply changing the definition flips the conclusion.

    This conversion is extremely sensitive. κ_VB/κ_obs = √[ρ_Y(1−ρ_O)/(ρ_O(1−ρ_Y))]; at ρ_O = 0.80, a difference of 0.05 (0.80→0.85) yields a 19.0% difference in κ, while a difference of 0.10 yields 50.0%. Conversely, the original criterion stating “κ_VB differs from κ_obs by > 0.20 ⇒ age trend primarily stems from measurement error” corresponds to only about 0.052 difference on the ρ scale—too small to estimate reliably in conventional sample sizes. Moreover, the direction is wrong: a difference greater than 0.20 indicates that reliabilities indeed differ, and this difference is precisely the condition under which the two definitions yield opposite signs. In such cases, the question to answer is “which ψ do we want,” not “which κ should we use.”

    A further issue involves scale mismatch. The criterion depends on κ, the decision rests on ψ, yet ψ is the square of κ: a 20% difference in κ corresponds to a 44% difference in ψ, and a 30% difference in κ corresponds to a 69% difference in ψ. Using κ as a threshold applies a square-root scale to evaluate a squared-scale quantity, causing any threshold to be systematically underestimated by nearly a factor of two.

    Another often-overlooked point: the direction of ψ_ref is not free either. When κ_ref exactly equals κ_obs, ψ_ref and ψ_obs are identical at each digit (simplifying λ/κ_obs² yields this result), indicating that the reference equation path provides no independent information—it merely reintroduces λ into the observed definition. Therefore, the accurate statement is: “If κ_ref deviates from κ_obs in a certain direction, then ψ_ref deviates from ψ_obs in the opposite direction, and the deviation is magnified by squaring.” For ρ_O = 0.80, ρ_Y = 0.85, λ = 1.5, if the reference equation overestimates the age trend by 27% (κ_ref = 1.35 while κ_obs = 1.06), ψ_ref drops from 1.333 to 0.823, flipping the sign; conversely, if it underestimates (κ_ref = 0.90), ψ_ref rises to 1.852. With the same dataset, choosing one reference equation can yield two opposite answers—”improvement of 18%” versus “worsening of 85%”—a fact this article most deserves to be remembered for.

    So what should be done in practice? At minimum, establish a reporting format: ψ must appear as pairs, clearly indicating which variance constitutes the numerator and which the denominator. If a paper states only “residual heteroscedasticity after z-transformation is 1.33,” readers cannot determine whether this refers to “the ratio of measurement errors between two segments” or “the ratio of true individual differences between two segments,” and these values will inevitably lie on opposite sides of 1 in most scenarios. Furthermore, whenever external reference equations are cited, the difference between κ_ref and κ_obs—including confidence intervals—must be reported, not merely whether it exceeds 0.15—because this threshold has a false-positive rate of approximately 24% when n = 120; a “violation” may simply reflect sampling noise.

    Finally, consider sample size. A Monte Carlo simulation (with true values identical, i.e., no age difference in ρ and λ, serving as a health check for the criteria set) reveals: Criterion 1 (|κ_ref−κ_obs|/κ_obs > 0.15) yields a false-positive rate of 24.4% at n = 120 per segment and 41.4% at n = 60; Criterion 3 (pairwise differences among three values < 0.10, originally intended as “κ is robust and can be used directly”) hits only 7.2% at n = 120, reaching 96% only when n ≈ 1200. Thus, at low sample sizes, 37%–50% of cases fail to trigger any criterion; at high sample sizes, Criterion 1 triggers false positives long before Criterion 3 becomes relevant—the feasible region of the criteria set is essentially empty. This mirrors the “two criteria share only one degree of freedom” problem from the previous article #1281, though the pathology differs: there, the criteria had no intersection; here, the sample size requirements and triggering order of the criteria conflict with each other.

    Pragmatic solutions include three steps. First, stop using κ as the carrier of the criterion; instead, report both definitions of ψ side-by-side ((1−ρ_O)/(1−ρ_Y) and ρ_O/ρ_Y), allowing readers to immediately see their opposite signs and understand what each represents. Second, if the reference equation must be used, separately report the difference between κ_ref and κ_obs along with its interval, rather than merely stating “whether it exceeds 0.15″—because the false-positive rate of the 0.15 line is already around 24% when n = 120. Third, return the task of “attributing age trends” to variance decomposition (the three-layer model from the previous article), rather than expecting a single SD ratio to simultaneously handle attribution and correction.

    Evidence and Limitations

    Verified: The closed-form derivations of the three κ definitions into ψ can be algebraically verified; λ is completely eliminated in both the observed and component-based definitions; the opposite signs of ψ_obs and ψ_VB are provable (equivalent to the magnitude relationship between ρ_Y and ρ_O), independent of λ values. The squared-magnification relationship where κ differences translate into ψ differences is an identity.

    Partially Verified: False-positive rates and hit rates for the criteria across sample sizes n = 60 to 2400 stem from pure simulation (using log-normal approximation, k = 3 repetitions, Fisher z approximation to estimate sampling error of ρ), not validated against real cohort data; long-tail distributions and missing patterns in actual data may result in even lower precision.

    Pending Verification: The actual magnitude of difference between ρ_O and ρ_Y in real pulmonary function cohorts (if reliabilities for both segments are nearly equal, the entire debate over definitions may be deemed unworthy of addressing); and whether κ_ref exhibits systematic overestimation when the reference equation itself stems from skewed fitting via LMS methods.

    Related Reading

    Evidence Level: C
    References: PMID 29707825, PMID 30208858, PMID 24338600, PMID 18006882, PMID 22743675, PMID 38741358

    L1 Network Status: E-004 | Metal | Main | Lung | Su Wen · Yin-Yang Ying Xiang Da Lun | Pending | L1 | open


    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.

  • z 化削掉的那部分个体差里,年龄占几成、真本体还剩几成

    Evidence level: C

    要问“削掉的是污染还是本体”,得先算出两个数;可这两条判据合起来只用了其中一个自由度——而且在信度不够时,它们根本没有共同的落脚点。

    一、先把三个数摆出来,别急着判

    前几篇一直在处理同一个终点:把 EELV(呼气末容积)与 DR2(二次吸气前的呼气末容积回落量)都折成占预计肺总容量的百分比,再取差值,叫“余量”。这个余量在同一个健康人身上重测,是稳还是不稳,已经被审过三轮:z 化对标准误没有作用(恒等式),误差方差与年龄是否无关需要单独检验,两个分量的误差是否互相独立更是从第一天起就假设错了。

    但还有一个问题一直悬着:z 化除以的是年龄分段的参考标准差,它把年龄带来的那部分人群差异一起缩掉了。指控是“连研究对象本体都削掉了”。这个指控要成立,必须拆出三个方差:

    • VB:个体之间的真方差(人跟人本来就不一样)
    • VE:同一个人重测的误差方差
    • VB_age:VB 里能被年龄解释掉的那一块

    剥掉年龄之后剩下的,是 VB_adj = VB − VB_age。原来那个信度指标 ICC = VB/(VB+VE),剥完之后变成 ICC_adj = VB_adj/(VB_adj+VE)。只要 VB_age 很大,ICC_adj 就会塌下去——这说明“这个指标测不出人跟人的差别了”,因为人跟人的差别本来就是年龄造成的,不是研究对象本身造成的。

    这一步的方向是对的,问题出在下一条:多数设计会用两条判据一起判(ICC_adj 要够高、VB_adj 占 VB 的比例也要够高)。这里要指出的是:这两条判据并不是两件事。

    二、两条判据,其实只有一根尺子

    记 ρ = VB/(VB+VE)(也就是剥年龄之前的原始 ICC),s = VB_age/VB(年龄在组间方差里占几成)。那么

    ICC_adj = ρ(1−s) / [ρ(1−s) + (1−ρ)]
    VB_adj/VB = 1 − s
    

    第二条判据直接就是 1−s,不需要别的。第一条判据 ICC_adj ≥ 0.75 化简后是

    s ≤ 1 − 3(1−ρ)/ρ  ≡  s*
    

    也就是说:两条判据都是关于 s 的单侧上限。第二条给 s ≤ 0.50,第一条给 s ≤ s*。s 只有一个,ρ 在给定样本里也只有一个——两条判据加起来,仍然只在同一个轴上各划一条线。它们提供不了两份独立的证据,所谓“双保险”是错觉。

    更要紧的是,这两个上限谁更严,取决于 ρ。代入 ρ 求解 1−3(1−ρ)/ρ = 0.50,得到 ρ = 6/7 ≈ 0.857**:

    原始 ρ s*(第一条判据的线) 实际起作用的那条
    0.75 0.000 第一条(第二条自动满足)
    0.80 0.250 第一条
    0.85 0.471 第一条
    0.857 0.500 两条重合
    0.90 0.667 第二条(第一条自动满足)

    所以“两条都过”这种表述,在 ρ<0.857 时等价于“第一条过”,在 ρ>0.857 时等价于“第二条过”。报告里写成两行,会让人以为是两次独立确认,实际上是一次。

    三、ρ 低于 0.75 时,这段判据是空的

    再往下推一步就更难看。s 要非负,需要 ρ ≥ 0.75;ρ 恰好等于 0.75 时 s = 0,意思是只有年龄一点都没解释掉方差,ICC_adj 才能维持在 0.75。而 ρ < 0.75 时 s* 为负,任何 s ≥ 0 都不能让 ICC_adj 达到 0.75。

    这意味着什么?ρ < 0.75 说明这个终点本来就不够稳——而“讨论 z 化削掉了什么”这件事,恰恰是在终点不够稳的时候才被提出来的(稳的话根本不用 z 化)。于是这个设计陷入自相矛盾:它被提出来的场合,正是它无解的那个场合。四条出口里没有一条是“判通过”,只能判“换终点”或“改设计”。

    用纯代数算下来:ρ 估偏 0.02,s 就偏 3/ρ² × 0.02。ρ=0.80 时偏 0.094,而 s 本身只有 0.250——判据线的相对位置抖了 37.5%。判据线比被判的量还飘,这是比值型判据的老毛病。

    四、真值落在判据线上会怎样:抛硬币

    上面的代数假设 ρ 和 s 都已知。实际上两个都得估。做了一组三层模型(人/次/残差)的蒙特卡洛,2000 次/格,取 ρ=0.80、每次测 3 遍:

    • 真值 s=0.50(正好在第二条判据线上):VB_adj/VB ≥ 0.50 的通过率 49.7%,就是抛硬币。ICC_adj 的真值 0.667,离 0.75 差了 0.083,通过率只有 1.3%。
    • 真值 s=0.30:ICC_adj 真值 0.737,离判据线只差 0.013。n=120 时通过率 31.5%,n=300 时 24.7%,n=600 时仍在 25% 上下——样本量加大也救不了,因为真值本来就贴着线。
    • 高信度格(ρ=0.95、s=0.30):ICC_adj 真值 0.930,通过率 99.9%,两种样本量都一样。

    这组数说明两件事。第一,α 门槛设在真值附近时,判据输出的是噪声而不是结论;第二,n 只压窄分布宽度,不改变分布中心——真值贴线,加样本永远只能换来“更确定地贴线”。而 s 的抽样精度本身也不好看:n=120、s=0.30 时标准差约 0.114,95% 区间半宽 0.223,比判据线和真值之间那点距离宽二十倍。

    五、还有一层:算 s 用的分母是观测方差,不是真方差

    上面的 s 是 VB_age/VB。真做起来,年龄能解释的那部分通常是用“观测值对年龄回归的 R²”近似的。但观测值里混着 VE,而 VE 在取个体均值时被平均掉了一部分,没有完全掉干净。第 k 次测量的均值,其信度是

    ρ_k = ρ / [ρ + (1−ρ)/k]
    

    k=3、ρ=0.80 时 ρ_k=0.923,比 ρ 高。拿这个偏高的 ρ_k 去除,年龄解释比例会被系统性高估:真值 s=0.50 时报出 0.538(高估 7.7%);ρ=0.50 时高估 25.0%;只测一次的方案(k=1)高估 20.0%。越是不稳的终点,这个高估越大——又一处“越需要它准的时候它越不准”。

    这也把 #1284 那条“κ 该取哪个标准差”的争论接上了:分母口径不统一,判据线就不是同一条线。

    证据与局限

    已验证:两判据共用一个自由度的推导是纯代数,实算 ICC_adj 在“占比方口径”与“占总方口径”下逐格完全相等(差 ≤1e-16),说明化简无误;ρ=6/7、s=1−3(1−ρ)/ρ、ds*/dρ=3/ρ² 均可直接验算。蒙特卡洛与解析解一致(s=0.50 处的硬币结果、真值贴线时的通过率)。

    部分验证:三层方差模型在重复测量数据上的可分解性有方法学支撑——多层数据的信度可分解为各层方差份额(PMID 33818118),ICC 与 SEM 的估计框架成熟(PMID 15705040),概括力理论(generalizability theory)给出的正是“把总方差拆成面与面交互”的路子(PMID 41441409)。余量本身的 VE 量级可参考 FOT 的重复性研究(PMID 33914403、PMID 23598954),但那是阻抗参数,不是本终点。

    待验证:本终点上 VB_age/VB 的真实值从未测过;ρ 在成人年龄跨度 ≥40 岁的样本里也从未报过。因此“相容区实测落在哪一段”完全未知——只知道若 ρ<0.75,这段判据无解。另需注意 ICC 依赖人群异质性、跨样本不可外推(PMID 29707825),异方差未校正会系统性高估 ICC(PMID 30208858),而 s 的区间估计在接近边界时覆盖率不足(PMID 14673940);样本量若不按目标精度反解(PMID 24338600),上面那些“贴线”的结论会被更宽的区间彻底淹没。

    相关阅读

    证据等级: C
    参考文献: PMID 33818118, PMID 15705040, PMID 29707825, PMID 30208858, PMID 24338600, PMID 14673940, PMID 23598954, PMID 33914403
    L1 网络状态:E-004 | 金 | 主 | 肺 | 《素问·阴阳应象大论》 | 待补 | L1 | open(该边现代证据栏仍为「待补」,全文只能作方法学讨论,不得读作已验证机制)


    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。

    What proportion of the eliminated individual differences is age, and what proportion remains in the true entity?

    Evidence level: C

    要问“削掉的是污染还是本体”,得先算出两个数;可这两条判据合起来只用了其中一个自由度——而且在信度不够时,它们根本没有共同的落脚点。

    一、先把三个数摆出来,别急着判

    前几篇一直在处理同一个终点:把 EELV(呼气末容积)与 DR2(二次吸气前的呼气末容积回落量)都折成占预计肺总容量的百分比,再取差值,叫“余量”。这个余量在同一个健康人身上重测,是稳还是不稳,已经被审过三轮:z 化对标准误没有作用(恒等式),误差方差与年龄是否无关需要单独检验,两个分量的误差是否互相独立更是从第一天起就假设错了。

    但还有一个问题一直悬着:z 化除以的是年龄分段的参考标准差,它把年龄带来的那部分人群差异一起缩掉了。指控是“连研究对象本体都削掉了”。这个指控要成立,必须拆出三个方差:

    • VB:个体之间的真方差(人跟人本来就不一样)
    • VE:同一个人重测的误差方差
    • VB_age:VB 里能被年龄解释掉的那一块

    剥掉年龄之后剩下的,是 VB_adj = VB − VB_age。原来那个信度指标 ICC = VB/(VB+VE),剥完之后变成 ICC_adj = VB_adj/(VB_adj+VE)。只要 VB_age 很大,ICC_adj 就会塌下去——这说明“这个指标测不出人跟人的差别了”,因为人跟人的差别本来就是年龄造成的,不是研究对象本身造成的。

    这一步的方向是对的,问题出在下一条:多数设计会用两条判据一起判(ICC_adj 要够高、VB_adj 占 VB 的比例也要够高)。这里要指出的是:这两条判据并不是两件事。

    二、两条判据,其实只有一根尺子

    记 ρ = VB/(VB+VE)(也就是剥年龄之前的原始 ICC),s = VB_age/VB(年龄在组间方差里占几成)。那么

    ICC_adj = ρ(1−s) / [ρ(1−s) + (1−ρ)]
    VB_adj/VB = 1 − s
    

    第二条判据直接就是 1−s,不需要别的。第一条判据 ICC_adj ≥ 0.75 化简后是

    s ≤ 1 − 3(1−ρ)/ρ  ≡  s*
    

    也就是说:两条判据都是关于 s 的单侧上限。第二条给 s ≤ 0.50,第一条给 s ≤ s*。s 只有一个,ρ 在给定样本里也只有一个——两条判据加起来,仍然只在同一个轴上各划一条线。它们提供不了两份独立的证据,所谓“双保险”是错觉。

    更要紧的是,这两个上限谁更严,取决于 ρ。代入 ρ 求解 1−3(1−ρ)/ρ = 0.50,得到 ρ = 6/7 ≈ 0.857**:

    原始 ρ s*(第一条判据的线) 实际起作用的那条
    0.75 0.000 第一条(第二条自动满足)
    0.80 0.250 第一条
    0.85 0.471 第一条
    0.857 0.500 两条重合
    0.90 0.667 第二条(第一条自动满足)

    所以“两条都过”这种表述,在 ρ<0.857 时等价于“第一条过”,在 ρ>0.857 时等价于“第二条过”。报告里写成两行,会让人以为是两次独立确认,实际上是一次。

    三、ρ 低于 0.75 时,这段判据是空的

    再往下推一步就更难看。s 要非负,需要 ρ ≥ 0.75;ρ 恰好等于 0.75 时 s = 0,意思是只有年龄一点都没解释掉方差,ICC_adj 才能维持在 0.75。而 ρ < 0.75 时 s* 为负,任何 s ≥ 0 都不能让 ICC_adj 达到 0.75。

    这意味着什么?ρ < 0.75 说明这个终点本来就不够稳——而“讨论 z 化削掉了什么”这件事,恰恰是在终点不够稳的时候才被提出来的(稳的话根本不用 z 化)。于是这个设计陷入自相矛盾:它被提出来的场合,正是它无解的那个场合。四条出口里没有一条是“判通过”,只能判“换终点”或“改设计”。

    用纯代数算下来:ρ 估偏 0.02,s 就偏 3/ρ² × 0.02。ρ=0.80 时偏 0.094,而 s 本身只有 0.250——判据线的相对位置抖了 37.5%。判据线比被判的量还飘,这是比值型判据的老毛病。

    四、真值落在判据线上会怎样:抛硬币

    上面的代数假设 ρ 和 s 都已知。实际上两个都得估。做了一组三层模型(人/次/残差)的蒙特卡洛,2000 次/格,取 ρ=0.80、每次测 3 遍:

    • 真值 s=0.50(正好在第二条判据线上):VB_adj/VB ≥ 0.50 的通过率 49.7%,就是抛硬币。ICC_adj 的真值 0.667,离 0.75 差了 0.083,通过率只有 1.3%。
    • 真值 s=0.30:ICC_adj 真值 0.737,离判据线只差 0.013。n=120 时通过率 31.5%,n=300 时 24.7%,n=600 时仍在 25% 上下——样本量加大也救不了,因为真值本来就贴着线。
    • 高信度格(ρ=0.95、s=0.30):ICC_adj 真值 0.930,通过率 99.9%,两种样本量都一样。

    这组数说明两件事。第一,α 门槛设在真值附近时,判据输出的是噪声而不是结论;第二,n 只压窄分布宽度,不改变分布中心——真值贴线,加样本永远只能换来“更确定地贴线”。而 s 的抽样精度本身也不好看:n=120、s=0.30 时标准差约 0.114,95% 区间半宽 0.223,比判据线和真值之间那点距离宽二十倍。

    五、还有一层:算 s 用的分母是观测方差,不是真方差

    上面的 s 是 VB_age/VB。真做起来,年龄能解释的那部分通常是用“观测值对年龄回归的 R²”近似的。但观测值里混着 VE,而 VE 在取个体均值时被平均掉了一部分,没有完全掉干净。第 k 次测量的均值,其信度是

    ρ_k = ρ / [ρ + (1−ρ)/k]
    

    k=3、ρ=0.80 时 ρ_k=0.923,比 ρ 高。拿这个偏高的 ρ_k 去除,年龄解释比例会被系统性高估:真值 s=0.50 时报出 0.538(高估 7.7%);ρ=0.50 时高估 25.0%;只测一次的方案(k=1)高估 20.0%。越是不稳的终点,这个高估越大——又一处“越需要它准的时候它越不准”。

    这也把 #1284 那条“κ 该取哪个标准差”的争论接上了:分母口径不统一,判据线就不是同一条线。

    证据与局限

    已验证:两判据共用一个自由度的推导是纯代数,实算 ICC_adj 在“占比方口径”与“占总方口径”下逐格完全相等(差 ≤1e-16),说明化简无误;ρ=6/7、s=1−3(1−ρ)/ρ、ds*/dρ=3/ρ² 均可直接验算。蒙特卡洛与解析解一致(s=0.50 处的硬币结果、真值贴线时的通过率)。

    部分验证:三层方差模型在重复测量数据上的可分解性有方法学支撑——多层数据的信度可分解为各层方差份额(PMID 33818118),ICC 与 SEM 的估计框架成熟(PMID 15705040),概括力理论(generalizability theory)给出的正是“把总方差拆成面与面交互”的路子(PMID 41441409)。余量本身的 VE 量级可参考 FOT 的重复性研究(PMID 33914403、PMID 23598954),但那是阻抗参数,不是本终点。

    待验证:本终点上 VB_age/VB 的真实值从未测过;ρ 在成人年龄跨度 ≥40 岁的样本里也从未报过。因此“相容区实测落在哪一段”完全未知——只知道若 ρ<0.75,这段判据无解。另需注意 ICC 依赖人群异质性、跨样本不可外推(PMID 29707825),异方差未校正会系统性高估 ICC(PMID 30208858),而 s 的区间估计在接近边界时覆盖率不足(PMID 14673940);样本量若不按目标精度反解(PMID 24338600),上面那些“贴线”的结论会被更宽的区间彻底淹没。

    相关阅读

    证据等级: C
    参考文献: PMID 33818118, PMID 15705040, PMID 29707825, PMID 30208858, PMID 24338600, PMID 14673940, PMID 23598954, PMID 33914403
    L1 网络状态:E-004 | 金 | 主 | 肺 | 《素问·阴阳应象大论》 | 待补 | L1 | open(该边现代证据栏仍为「待补」,全文只能作方法学讨论,不得读作已验证机制)


    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.

  • 「误差方差与年龄无关」这个前提,到底成不成立:一个测不出、也测不准的检验

    Evidence level: C

    上一篇用恒等式证明「年龄常模 z 化不改 SEM」,那个证明其实只借了一条前提:测量误差的方差与年龄无关。这条前提从未被检验过,而检验它的那把尺子本身也是坏的——门槛设在效应很小的地方,功效却低到几乎永远按不下去。

    正文

    一条被当成常数用的前提。 上一篇的推导写得很干净:z 化把总标准差按年龄分段拉平,而测量误差方差 VE 若与年龄无关,就不跟着动,两边一抵消,SEM 恰好不变。整条链只有 VE 是常数这一环没有出处。它也不是随口说的——肺功能的日间变异本身就有 5%–15%(PMID 23598954),强迫振荡技术在不同肺容积档位下 ICC 从 0.60 排到 0.98、低容积端最差(PMID 38036081);逼尿肌参考值的变异系数只在 15–35 岁接近 10%、其余年龄段明显更大(PMID 18006882),呼吸系统老化的综述也直接说老年段正常值离散度更大、并反对依赖公式按年龄归一化(PMID 12735112)。三处线索指向同一方向:老年段的分散不是更小,而是更大。但「指向」不是「证明」。于是一个自然的动作被提出来:拿同一批 ≥120 名成人,按年龄分十档估 VE,做异方差检验,看最老与最年轻的 VE 比(记作 λ)是否超过 1.5。这个动作看起来无懈可击,实际上有三处结构性问题。

    第一处:不变点不在 λ=1,而在 λ=κ²。 先把代数摆清。z 化把每个年龄段的读数除以该年龄的参考标准差,方差同时除以它的平方——这一步是精确的,不带任何近似。于是 z 化后的误差方差 VE_z = VE / SD_ref²,两段相比得到 λ_z = λ / κ²,其中 κ 是老年段与年轻段参考标准差的比值。留下这句:误差方差能否被拉平,不取决于 λ 是否等于 1,而取决于 λ 是否恰好等于 κ²。这是个刀锋条件。取 κ=1.2(老年段参考标准差大两成,这在肺功能上是保守估计),λ 必须正好 1.44 才恰到好处;λ=1.5 只剩 1.042 的残余异方差,几乎白做;λ=1.0(误差真的与年龄无关)则给出 ψ=0.694,也就是老年段的误差被压得比年轻段还小了三成——原本为了消除异方差的动作,反过来制造了反向异方差。所以「λ<1 就说明 z 化真改善精度」这句话必须打个折:λ<1 只说明 z 化没有过度修正,不说明它修正得对。

    第二处:这个检验在 n=120 时几乎没有功效。 门槛设在 λ>1.5,那么先看 λ=1.5 意味着什么:λ 是方差之比,落到我们能直接看见的 SEM 上要开根号,只放大 1.225 倍。用单一合并 SEM 报 MDC95 时,年轻段实际覆盖率 97.2%、老年段 92.6%,落差 4.5 个百分点。这个落差说大不大,说小也不小——它意味着老年受试者被判「真变了」的门槛偏低了一成多。问题在于,为了抓这个 4.5 个百分点的落差,需要多大的样本?按十档线性上升的误差结构做蒙特卡洛(每格 2000 次),n=120 时 Breusch–Pagan 检验对 λ=1.5 的实际功效只有 13.5%(λ=2.0 也只有 30.6%);换成人均 3 次重复、对 log s² 回归的做法,n=120 给 16.6%,n=300 才 35.5%,要到 n=600 才 57.9%。解析解与模拟一致:λ=1.5 的效应在 BP 回归里只能解释 R²≈0.0067,n=120 的非中心参数是 0.80,功效自然趴在 14% 附近。

    第三处:分箱判据本身只有一个标准误。 就算绕开检验、直接看「最老与最年轻十分位的 VE 比」,这条路也走不通。n=120 分十档,每档 12 人;若每人只测 3 次,每档的自由度是 12×2=24,单档 VE 的相对标准误 √(2/24)=0.289,也就是约 29%;两端比值再乘 √2,SE(log 比值)=0.408。而 λ=1.5 对应的 log 值是 0.405——判据正好站在离零一个标准误的地方。想把这个比值的 SE 压到 0.30,每档要约 45 人、十分位共约 450 人;压到 0.20 要每档 101 人、共约 1010 人。更极端的是单个人:3 次重复只有 2 个自由度,个人 VE 的相对标准误是 100%,等于没测。

    最要命的是这三处合起来的效果。 判据写成「检验显著,且 λ>1.5」,于是触发它需要两个条件同时成立。功效只有 14% 意味着:即使 λ 真的是 1.5、误差真的在变,这个判据也有约 86% 的概率不触发。而不触发的默认分支是「同方差近似够用、z 化结论保留」。也就是说,统计功效的不足会被系统性地翻译成「前提成立」这个结论——不是数据支持了同方差,是尺子太钝读不出差别,而钝读数的沉默被当成了证词。这一层比前两条更隐蔽:前两条是判据写得不好看,这一条是判据的沉默被赋予了证据地位。

    所以该改的不是阈值,是做法。 第一步,别做「检验+阈值」两步法,直接把 log(VE) 对年龄的斜率当主终点,报斜率的点估计与置信区间——它给出的是连续的量,不会因为在阈值两侧就翻转结论。判据可以写成:斜率 CI 的上界若小于 log(1.5)/年龄跨度,则同方差近似可用;若下界已超过该值,则必须分年龄段各自报 SEM。第二步,同时报 λ 与 κ,并直接报 λ/κ²,因为那才是 z 化之后真正剩下的异方差,而不是 λ 本身。第三步,样本量按目标精度反解(PMID 24338600),不要按「≥120 人」这种整数惯例拍;对 VE_z 的斜率若要 95%CI 半宽小于 0.2,按上面的分箱算法需要千人以量级,这个数字应当先算出来再谈可行性。至于 #1278 那条恒等式,它不必撤回,但要补一句适用范围:它在 λ=κ² 时成立,而 λ=κ² 是一个需要检验的巧合,不是一个默认值。

    证据与局限

    已验证:VE 与年龄无关这条前提在肺功能领域是明确可疑的——日间变异 5%–15%(PMID 23598954)、不同肺容积档 ICC 0.60–0.98(PMID 38036081)、参考值 CV 只在 15–35 岁接近 10%(PMID 18006882)、老年段离散度更大且反对公式归一化(PMID 12735112)。异方差未被校正会系统性高估 ICC,这是「违反同方差会怎样」的方法学依据(PMID 30208858)。样本量须按目标精度反解而非套惯例(PMID 24338600),ICC 在接近边界时自助 CI 覆盖率不足(PMID 14673940),且 ICC 依赖人群异质性、跨样本不可外推(PMID 29707825)。

    部分验证:λ_z=λ/κ² 这一步是精确代数,无争议;但 κ 的取值依赖选用哪一套参考方程,本文取 κ=1.2 作示例,未做多套方程的敏感性对比。BP 与 log s² 回归的功效数字来自本文自己的蒙特卡洛(每格 2000 次)与正态近似解析解,两者一致,但误差分布假设为正态;若真实误差厚尾,功效会更低、判据更难触发——方向对本文结论有利,幅度未量化。

    待验证:① SVC-FOT 上 VE 的年龄斜率到底多大、符号如何,本文只引了旁证,没有该技术自身的分层数据;② 参考方程的 SD 比值 κ 与样本自身的 SD 比值是否一致,若不一致,z 化的方向可能反转;③ 换用 Weiss–Loh 或 Breusch–Pagan 的稳健版本后,n=120 的功效是否仍趴在 14% 附近。

    相关阅读

    证据等级: C
    参考文献: PMID 23598954, PMID 38036081, PMID 18006882, PMID 12735112, PMID 30208858, PMID 24338600, PMID 14673940, PMID 29707825

    L1 网络状态:E-004|金|主|肺|《素问·阴阳应象大论》|现代证据「待补」|L1|open —— 该边仅为 L1 同构、状态 open,现代证据栏原文为「待补」,本文涉及的呼气末容积与 DR2 余量在该网络中无对应节点,只能借 E-004 代理;故本文证据等级最高只能标 C。

    证据等级:A=有 PubMed 直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


    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。

    Is the premise \”error variance is independent of age\” valid: an untestable and inconclusive hypothesis

    Evidence level: C

    In the previous post, an identity was used to prove that “age-normalized z-scoring does not alter the SEM”. That proof relied on one premise: that the variance of measurement error is independent of age. This premise was never tested, and the ruler used to test it is itself flawed—the threshold is set where effects are tiny, yet statistical power is so low that it almost never triggers.

    Main Text

    A premise treated as a constant. The derivation in the previous post is clean: z-scoring flattens the total standard deviation across age groups, and if the measurement error variance VE is independent of age, it remains flat; both sides cancel out, leaving SEM unchanged. The entire chain hinges solely on VE being constant—a link with no source. It is not casually stated: lung function’s within-day variability alone ranges from 5%–15% (PMID 23598954); oscillometry ICCs span 0.60 to 0.98 across different lung volumes, with the lowest volumes performing worst (PMID 38036081); detrusor reference values show coefficient of variation near 10% only in ages 15–35, significantly higher elsewhere (PMID 18006882); reviews on respiratory aging explicitly state that normal ranges are more dispersed in older adults and oppose formula-based age normalization (PMID 12735112). Three lines of evidence point in the same direction: dispersion increases with age, not decreases. But “pointing” is not “proven”. Thus, a natural approach is proposed: take the same cohort of ≥120 adults, estimate VE across ten age bands, perform heteroscedasticity testing, and check whether the ratio of VE in the oldest vs. youngest groups (denoted λ) exceeds 1.5. This approach appears flawless, yet it suffers from three structural flaws.

    First flaw: the invariant point is not λ=1, but λ=κ². Let us clarify the algebra. Z-scoring divides each age group’s readings by that group’s reference standard deviation; variance is divided by its square—this step is exact, involving no approximation. Thus, the z-scored error variance VE_z = VE / SD_ref². Comparing two groups yields λ_z = λ / κ², where κ is the ratio of reference standard deviations between older and younger groups. This leads to the key statement: whether error variance can be flattened depends not on whether λ equals 1, but on whether λ exactly equals κ². This is a razor-thin condition. Taking κ=1.2 (older group reference SD is 20% higher—a conservative estimate in lung function), λ must be exactly 1.44 to achieve perfect flattening; λ=1.5 leaves only a residual heteroscedasticity of 1.042, essentially negligible; λ=1.0 (error truly age-independent) yields ψ=0.694, meaning the error variance in older adults is suppressed by 30% relative to younger adults—the very action intended to eliminate heteroscedasticity instead creates reverse heteroscedasticity. Therefore, the claim “λ<1 implies z-scoring genuinely improves precision” must be qualified: λ<1 merely indicates that z-scoring does not over-correct, not that it corrects appropriately.

    Second flaw: this test has virtually no power at n=120. With the threshold set at λ>1.5, consider what λ=1.5 implies: as a variance ratio, it translates to the SEM scale by taking the square root, amplifying only 1.225-fold. When reporting MDC95 using a single pooled SEM, the actual coverage is 97.2% for younger adults and 92.6% for older adults—a difference of 4.5 percentage points. Whether this gap is large or small depends on context; it means the threshold for declaring “true change” in older subjects is lowered by over 10%. The question becomes: what sample size is needed to detect this 4.5-point gap? Using a linearly increasing error structure across ten bands via Monte Carlo simulation (2000 runs per band), at n=120, the Breusch–Pagan test has actual power of only 13.5% for λ=1.5 (and merely 30.6% for λ=2.0); switching to a method with three repetitions per subject and regressing on log s² yields 16.6% at n=120, 35.5% at n=300, and only 57.9% at n=600. Analytical solutions match simulations: for λ=1.5, the effect explains only R²≈0.0067 in BP regression, with non-centrality parameter 0.80 at n=120, placing power near 14%.

    Third flaw: the binning criterion itself has only one standard error. Even bypassing formal testing and directly comparing “the ratio of VE in the oldest vs. youngest deciles,” this path fails. With n=120 split into ten bands (12 per band), and assuming three measurements per person, each band has degrees of freedom 12×2=24; the relative standard error of single-band VE is √(2/24)=0.289, or about 29%. Multiplying the endpoint ratio by √2 gives SE(log ratio)=0.408. Since λ=1.5 corresponds to log value 0.405, the criterion sits exactly one standard error above zero. To reduce this ratio’s SE to 0.30, each band would need ~45 subjects, totaling ~450 across ten bands; reducing to 0.20 requires ~101 per band, totaling ~1,010. Even more extreme: for a single individual, three repetitions yield only two degrees of freedom, making the relative standard error of individual VE 100%—equivalent to no measurement.

    The most critical issue is the combined effect of these three flaws. If the criterion is stated as “test significant AND λ>1.5”, then triggering it requires two conditions simultaneously. With power only 14%, even if λ truly equals 1.5 and error varies with age, there is approximately an 86% probability that the criterion will not trigger. The default branch when it doesn’t trigger is “homoscedasticity approximation suffices, z-scoring conclusion holds.” Thus, insufficient statistical power is systematically translated into the conclusion that the premise holds—not because data supports homoscedasticity, but because the ruler is too blunt to detect differences, and the silence of the blunt reading is mistaken for evidence. This layer is more subtle than the previous two: the first two flaws stem from poorly formulated criteria, while this third flaw arises because the silence of the criterion is endowed with evidentiary status.

    Therefore, what needs changing is not the threshold, but the approach. First, abandon the “test + threshold” two-step method; instead, directly use the slope of log(VE) against age as the primary endpoint, reporting point estimates and confidence intervals for the slope. This provides a continuous quantity that does not flip conclusions based on crossing a threshold. The criterion can be formulated as: if the upper bound of the slope CI is less than log(1.5)/age span, the homoscedasticity approximation applies; if the lower bound exceeds this value, SEM must be reported separately by age group. Second, simultaneously report λ and κ, and directly report λ/κ², because that is the true residual heteroscedasticity after z-scoring, not λ itself. Third, derive sample size from target precision (PMID 24338600), rather than arbitrarily setting “≥120 subjects”; for VE_z slope with 95% CI half-width <0.2, the binning algorithm requires thousands of subjects—a number that must be calculated before feasibility is discussed. Regarding identity #1278, it need not be withdrawn, but must add a scope statement: it holds only when λ=κ², and λ=κ² is a coincidence requiring verification, not a default assumption.

    Evidence and Limitations

    Verified: The premise that VE is age-independent is clearly questionable in lung function—within-day variability 5%–15% (PMID 23598954), ICCs ranging 0.60–0.98 across lung volumes (PMID 38036081), reference CV near 10% only in ages 15–35 (PMID 18006882), older groups showing greater dispersion and opposing formula-based normalization (PMID 12735112). Unadjusted heteroscedasticity systematically inflates ICC, providing methodological basis for “what happens if homoscedasticity is violated” (PMID 30208858). Sample size must be derived from target precision rather than convention (PMID 24338600); ICC coverage under bootstrap CI is insufficient near boundaries (PMID 14673940), and ICC depends on population heterogeneity, making cross-sample extrapolation invalid (PMID 29707825).

    Partially verified: The step λ_z=λ/κ² is exact algebra, undisputed; however, κ’s value depends on which reference equation is used—this paper takes κ=1.2 as an example without sensitivity analysis across multiple equations. Power figures for BP and log s² regression come from this paper’s own Monte Carlo (2000 runs per band) and normal approximation analytical solution, which agree, but error distribution is assumed normal; if true errors are heavy-tailed, power would be even lower and the criterion harder to trigger—direction favors this paper’s conclusion, magnitude unquantified.

    To be verified: ① What is the actual age slope and sign of VE on SVC-FOT, this paper only cites collateral evidence without该技术’s own stratified data; ② Whether the SD ratio κ of reference equations matches the SD ratio of the sample itself—if mismatched, z-scoring direction may reverse; ③ Whether using robust versions of Weiss–Loh or Breusch–Pagan still yields power near 14% at n=120.

    Further Reading

    Evidence level: C
    References: PMID 23598954, PMID 38036081, PMID 18006882, PMID 12735112, PMID 30208858, PMID 24338600, PMID 14673940, PMID 29707825

    L1 network status: E-004 | Metal | Primary | Lung | 《Su Wen · Yin-Yang Ying Xiang Da Lun》 | Modern evidence “to be added” | L1 | open — this side is only L1 isomorphic, status open; the original text in the modern evidence column reads “to be added”; the exhaled end-volume and DR2 margin involved in this paper have no corresponding nodes in this network, so E-004 serves as proxy; thus the highest evidence level for this paper can only be marked C.

    Evidence level: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration


    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.

  • 「余量」是两个数的差,可它的误差一直被当成两笔独立的账:误差相关一算进去,2.77 这个常数就不成立了

    Evidence level: C

    两个读数出自同一条曲线、同一根容积轴,它们的误差不可能互不相干;一旦承认这一点,那个被反复引用的 2.77 就不再是常数,而只是「误差完全不相关」这一种特殊情况下的数字。

    正文

    在丹田呼吸的量化方案里,「余量」被定义成两个数相减:呼气末容积减掉 DR2,两者都折成占肺总容量的百分比。既然是个差值,就得给它配一个最小可检测变化,也就是 MDC95——超过这个幅度的变化,才有资格说不是测量噪声。文献里通用的算法是一步乘法:MDC95 等于 2.77 乘上单个分量的标准误。这个 2.77 从哪来?它是 1.96 乘根号 2——1.96 取 95% 的置信,根号 2 则来自「两个分量的测量误差互相独立、方差直接相加」这条假设(PMID 15705040;PMID 16925807)。

    问题就出在根号 2 上。经典测量理论里,差值的标准误写全了是这么一条:根号下 s₁ 平方加 s₂ 平方、再减去 2 乘 ρ 乘 s₁ 乘 s₂,ρ 是两个分量测量误差之间的相关(PMID 15286006)。只有当 ρ 等于零,这条式子才退化成根号 2 乘单个标准误。而我们的两个读数出自同一条 SVC-FOT 曲线、共用同一根容积轴定标——轴一旦偏,两个数同向偏。这不是「可能相关」,这是「设计上就注定相关」。根号 2 这个因子,从第一天起就用错了地方。

    那错得有多厉害?两个分量误差尺度相同时,真实标准误是独立假设算出来的根号下 1 减 ρ 倍。ρ 等于 0.6,比值 0.632,独立假设高估了 1.58 倍;ρ 等于 0.8,比值 0.447,高估 2.24 倍;ρ 到 0.95,比值只剩 0.224,高估 4.47 倍。更值得记住的是第二层:两个分量误差尺度差得越多,独立假设反而越靠得住。ρ 等于 0.8 时,若一个分量的误差是另一个的三倍,高估倍数就从 2.24 掉到 1.39——因为尺度悬殊时,同向漂移难以互相抵消,相关的作用被稀释了。反过来说,DR2 与呼气末容积恰好落在最坏的格子里:同曲线、同轴、误差尺度接近,也就是最接近等尺度那一格。这一格,正是 ρ 破坏力最大的地方。

    方向必须说清:这不是假阳性膨胀,是自罚。ρ 大于零时,独立假设把门槛抬高了,名义 5% 的判定实际只在 0.19% 处触发(ρ 等于 0.6),灵敏度被白白削掉。要跨过被抬高的门槛,真实变化得达到应有幅度的 1.58 倍、2.24 倍甚至 4.47 倍;按方差反推,所需样本量分别要乘 2.5 倍、5 倍、20 倍——正好是 1 除以(1 减 ρ)。所以「余量这个终点不敏感」这个印象,有可能根本不是生理事实,而是常数用错换来的。

    但把公式改对,会立刻撞上两堵墙。第一堵:ρ 越高,MDC95 对 ρ 越敏感,而 ρ 本身越高越估不准。ρ 在 0.6 附近估偏 0.05,MDC95 只差 6%;ρ 到 0.9 时同样的偏差让 MDC95 差 22%,ρ 到 0.95 时差 41%。而按 Fisher 变换,30 名受试者给出的 ρ 置信区间在 ρ 等于 0.8 时半宽约 0.14,映射到 MDC95 就是上下相差 2.29 倍——判据根本收不拢。要在 ρ 等于 0.9 时把 ρ 的区间压到正负 0.05,样本量得上万。这意味着高相关场景下 ρ 不是「估一下」,而是「估不出来」。

    第二堵墙更麻烦:ρ 一升,差值本身的信度反而塌下来。差值信度等于(分量信度减观测相关)除以(1 减观测相关)。ρ 增大推高观测相关,分母趋零、分子转负。分量信度 0.80 时,只要真值相关超过 0.75,ρ 够大就能让差值信度跌破零——一个「信度为负」的指标,MDC95 算得再小也没有意义。于是出现一个悖论:同一个 ρ 让「能测出多小的差」变好看,同时让「这个差可不可信」变难看。只改 MDC95 公式而不看差值信度,等于从一个偏松的判据,跳到一个更小、更漂亮、也更不可信的判据。

    还有一个方向不能漏:ρ 也可能是负的。若 DR2 的检出往上偏时呼气末容积的读数往下偏——比如算法在一个量上「吃」掉的部分,在另一个量上补回来——ρ 就是负值,此时独立假设不是自罚而是自宽,实际门槛低于应有水平。ρ 等于负 0.3 时,MDC95 应放大 1.14 倍,而按独立算出来的门槛只覆盖了 91% 的应有区间;假阳性率从 5% 涨到 8.6%,负 0.5 时涨到 11%。所以「误差相关」这件事没有安全方向,只有「估对」和「估错」两种。

    怎么估?不需要新仪器,只要把已有的重复测量换个用法:同一批人隔天各测五次,把每一次算出的 DR2 与呼气末容积配成一对,再把这些配对差当成一个新变量,直接求它的协方差阵,s₁、s₂、ρ 三个数同时出来。关键在扰动臂——单靠一次测量得到的 ρ 有可能只是巧合,得靠「换容积轴定标方法」和「换 DR2 检出算法」两个扰动去看它是否同向偏移。如果两个读数确实共用了同一根轴的偏差,那么换定标方法时它们应该一起动;如果换算法时相关消失、换定标时相关仍在,那就把责任指到了定标环节——这比一个孤零零的 ρ 数值有用得多。

    证据与局限

    已验证:差值标准误的完整表达式(含 2ρs₁s₂ 交叉项)是经典测量理论的既有结果,教材与文献口径一致(PMID 15286006)。「独立假设下 MDC95 等于 2.77 乘单个标准误」也是标准算法(PMID 15705040)。本轮的倍数关系(ρ 对标准误、对样本量、对假阳性率、对差值信度的影响)为代数推导,数值已逐格验算。

    部分验证:ρ 大于零这一判断有设计层面的强理由(同曲线、同轴、同算法),但具体数值区间尚无本领域实测。FOT 类仪器的短期重复性研究给出了分量层面的重复性(PMID 33914403;PMID 34625483;PMID 38036081),可用于给 s₁、s₂ 的量级打底,但都没有报告两分量误差的协方差。

    待验证:一、ρ 的实测值及其置信区间,以及它在「换轴定标」「换检出算法」两类扰动下是否同向偏移。二、两个分量误差尺度之比的实际值——这个比值决定我们到底落在最坏格还是较安全的格。三、差值信度在本指标上的实际取值,以及它是否触及「转负」那条线。四、修正公式后,此前所有基于 2.77 得出的「余量不敏感」结论是否需要整体重估。需要说明的是,本条对应的边在 L1 网络里只有 E-004 金主肺这一条、且状态为 open、现代证据栏仍标待补,因此本文不构成对任何机制的确认。

    相关阅读

    L1 网络状态:E-004 | 金 | 主 | 肺 |《素问·阴阳应象大论》| 现代证据:待补 | L1 | open。该主题(测量误差相关性与差值 MDC)尚未进入 L1 对应网络,本次已登记至 l1-gaps。

    证据等级: C
    参考文献: PMID 15286006, PMID 15705040, PMID 16925807, PMID 33914403, PMID 34625483, PMID 38036081


    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。

    Surplus is the difference between two numbers, yet its error has always been treated as two separate accounts: once the error-related calculation is included, the constant 2.77…

    Evidence level: C

    Two readings derived from the same curve and the same volume axis cannot be entirely independent in error; once this is acknowledged, the repeatedly cited 2.77 is no longer a constant but merely the numerical value under the specific condition of “completely uncorrelated error.”

    Body

    In the quantification scheme for Dantian breathing, “margin” is defined as the subtraction of two numbers: end-expiratory volume minus DR2, both converted to percentages of total lung capacity. Since this is a difference, it requires an assigned minimum detectable change, MDC95—only changes exceeding this magnitude qualify as not being measurement noise. The standard algorithm found in literature is a single multiplication step: MDC95 equals 2.77 multiplied by the standard error of a single component. Where does this 2.77 originate? It is 1.96 multiplied by the square root of 2—1.96 taking the 95% confidence, while the square root of 2 stems from the assumption that “measurement errors of two components are mutually independent, and variances add directly” (PMID 15705040; PMID 16925807).

    The problem lies precisely in the square root of 2. In classical measurement theory, the full expression for the standard error of a difference is written as: the square root of s₁ squared plus s₂ squared, minus 2 times ρ times s₁ times s₂, where ρ represents the correlation between measurement errors of the two components (PMID 15286006). Only when ρ equals zero does this formula degenerate into the square root of 2 times the single standard error. However, our two readings originate from the same SVC-FOT curve and share the same volume axis for calibration—once the axis shifts, both numbers shift in the same direction. This is not “possibly correlated”; this is “designed to be correlated.” The factor of square root 2 has been applied incorrectly from the very beginning.

    How badly is it wrong? When the error scales of two components are identical, the true standard error becomes the square root of 1 minus ρ times the independent assumption. When ρ equals 0.6, the ratio is 0.632; the independent assumption overestimates by 1.58 times. When ρ equals 0.8, the ratio is 0.447; it overestimates by 2.24 times. When ρ reaches 0.95, the ratio shrinks to only 0.224; it overestimates by 4.47 times. More importantly, the second layer of insight is this: the greater the difference in error scales between two components, the more reliable the independent assumption becomes. When ρ equals 0.8, if one component’s error is three times that of the other, the overestimation factor drops from 2.24 to 1.39—because when scales differ greatly, co-directional drifts cannot cancel each other out effectively, diluting the impact of correlation. Conversely, DR2 and end-expiratory volume fall exactly into the worst-case grid: same curve, same axis, similar error scales—closest to the equal-scale scenario. This specific grid is precisely where ρ exerts its greatest destructive power.

    The direction must be clarified: this is not false-positive inflation; it is self-penalty. When ρ is greater than zero, the independent assumption raises the threshold; the nominal 5% decision actually triggers only at 0.19% (when ρ equals 0.6), with sensitivity unnecessarily reduced. To cross this raised threshold, the true change must reach 1.58 times, 2.24 times, or even 4.47 times the expected magnitude; reversing via variance, the required sample size must be multiplied by 2.5 times, 5 times, or 20 times respectively—exactly 1 divided by (1 minus ρ). Thus, the impression that “this margin endpoint is insensitive” might not be a physiological fact at all, but rather the result of misusing the constant.

    However, correcting the formula immediately collides with two walls. The first wall: as ρ increases, MDC95 becomes increasingly sensitive to ρ, yet ρ itself becomes harder to estimate accurately. When ρ is near 0.6 and estimated with a bias of 0.05, MDC95 differs by only 6%; when ρ reaches 0.9, the same bias causes MDC95 to differ by 22%; when ρ reaches 0.95, it differs by 41%. Furthermore, according to Fisher’s transformation, the confidence interval for ρ derived from 30 subjects has a half-width of approximately 0.14 when ρ equals 0.8, mapping to MDC95 with a range spanning 2.29 times—making the criterion fundamentally inconclusive. To compress the ρ interval to ±0.05 when ρ equals 0.9, the sample size must reach tens of thousands. This means in high-correlation scenarios, ρ cannot be “estimated”; it is “unestimable.”

    The second wall is even more troublesome: as ρ rises, the reliability of the difference itself collapses. Difference reliability equals (component reliability minus observed correlation) divided by (1 minus observed correlation). Increasing ρ raises the observed correlation, driving the denominator toward zero and the numerator into negative territory. When component reliability is 0.80, if the true correlation exceeds 0.75, a sufficiently large ρ can push difference reliability below zero—a “negative reliability” metric renders MDC95 calculations meaningless regardless of how small they are. Thus arises a paradox: the same ρ makes “how small a difference can be detected” appear better, while simultaneously making “is this difference trustworthy” appear worse. Adjusting only the MDC95 formula without considering difference reliability amounts to jumping from a loosely biased criterion to a smaller, prettier, yet even less trustworthy criterion.

    One more direction cannot be overlooked: ρ could also be negative. If DR2 detection shifts upward while end-expiratory volume readings shift downward—such as when the algorithm “eats” a portion on one measure and compensates it on another—ρ becomes negative. In this case, the independent assumption does not self-penalize but rather self-widens; the actual threshold falls below the expected level. When ρ equals negative 0.3, MDC95 should be amplified by 1.14 times, yet the threshold calculated via independence covers only 91% of the expected interval; false-positive rates rise from 5% to 8.6%, reaching 11% when ρ equals negative 0.5. Therefore, “correlated error” has no safe direction; there are only two possibilities: “estimated correctly” and “estimated incorrectly.”

    How to estimate? No new instruments are needed; simply repurpose existing repeated measurements: measure the same group five times each day, pair each calculated DR2 with end-expiratory volume, treat these paired differences as a new variable, and directly compute its covariance matrix to obtain s₁, s₂, and ρ simultaneously. The key lies in the perturbation arms—relying solely on single-measurement ρ may yield coincidence; one must apply “changing volume axis calibration methods” and “changing DR2 detection algorithms” as two perturbations to observe whether they shift co-directionally. If both readings truly share the same axis deviation, changing calibration methods should cause them to move together; if correlation vanishes when changing algorithms but persists when changing calibration, the responsibility points to the calibration step—far more useful than a solitary ρ value.

    Evidence and Limitations

    Verified: The complete expression for the standard error of differences (including the 2ρs₁s₂ cross-term) is an established result in classical measurement theory, consistent across textbooks and literature (PMID 15286006). “MDC95 equals 2.77 times the single standard error under the independent assumption” is also a standard algorithm (PMID 15705040). The multiplicative relationships in this round (ρ’s impact on standard error, sample size, false-positive rate, and difference reliability) are algebraic derivations, with numerical values verified cell-by-cell.

    Partially Verified: The judgment that ρ is greater than zero has strong design-level rationale (same curve, same axis, same algorithm), but specific numerical ranges lack empirical measurement within this field. Short-term repeatability studies of FOT-type instruments provide component-level repeatability (PMID 33914403; PMID 34625483; PMID 38036081), which can establish the order of magnitude for s₁ and s₂, yet none report the covariance between errors of two components.

    To Be Verified: First, the measured value of ρ and its confidence interval, along with whether it shifts co-directionally under two types of perturbations—”changing axis calibration” and “changing detection algorithm.” Second, the actual ratio of error scales between two components—this ratio determines whether we fall into the worst-case grid or the relatively safe one. Third, the actual value of difference reliability on this metric, and whether it crosses the “turn negative” threshold. Fourth, after correcting the formula, whether all previous conclusions based on 2.77 regarding “margin insensitivity” require comprehensive re-evaluation. It should be noted that the side corresponding to this entry in the L1 network includes only E-004 Golden Master Lung, with status open, and the modern evidence column still marked as pending; therefore, this article does not constitute confirmation of any mechanism.

    Additional Reading

    L1 Network Status: E-004 | Golden | Master | Lung | Su Wen · Yin-Yang Ying Xiang Da Lun | Modern Evidence: Pending | L1 | open. This topic (measurement error correlation and difference MDC) has not yet entered the corresponding L1 network; it has been registered in l1-gaps.

    Evidence Level: C
    References: PMID 15286006, PMID 15705040, PMID 16925807, PMID 33914403, PMID 34625483, PMID 38036081


    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.

  • 年龄常模 z 化到底改了什么:它对「能测出多小的差」零贡献,只把 ICC 改难看

    Evidence level: C

    上一篇的判据写「年龄解释方差超过 50% 就改用同年龄常模 z 化的余量」,这话暗示 z 化能把终点救回来。但代数学上它救不了:z 化不改变测量标准误,一丝都不改。它唯一改的,是信度系数那个数字本身。

    正文

    先把上一篇留下的那句话拆开看。它说的是:余量这个终点,如果年龄解释方差太大(比如 r²=0.71,来自 DR2 与年龄的相关),那这个终点其实在测年龄,不是在测「沉而不塌」,所以应该换成同年龄常模 z 化之后的余量。逻辑前半段没问题——一个被年龄吃掉七成方差的指标,确实不能直接做组间比较。问题出在「所以 z 化」这一步:它默认了 z 化能修好什么东西,而这一步没有任何代数支撑。

    第一件事:SEM 是个不变量。 设一个人群里的余量总方差 VT = VB + VE(VB 是个体间真差,VE 是测量误差),信度 ICC = VB/VT,年龄占真差 VB 的比例是 r²。z 化就是把每个人的值按年龄除以本年龄段的 SD。做完之后,尺度缩了 √(1−r²) 倍,ICC 变成 (ICC−r²)/(1−r²)——难看了。但把两个数代进标准误公式 SEM = SD·√(1−ICC),会看到 SD 缩了 √(1−r²),而 √(1−ICC) 恰好涨了 1/√(1−r²),两者精确抵消,SEM 一分不变。实算三档(r² 取 0.30 / 0.50 / 0.71,ICC 取 0.75 与 0.90),SEM 比全部是 1.000000。MDC95 = 1.96√2·SEM,自然同样不变;组级最小可检测差 = 1.96·√2·SEM/√n,也不变。这不是近似,是恒等式。

    第二件事:ICC 会掉,而且能掉到负数。 ICC=0.75 时,r²=0.30 落到 0.643,r²=0.50 落到 0.500,r²=0.71 直接落到 0.138——从「合格」掉进「不合格」只需一步。更狠的是临界点:按这个口径,ICCz 跌破零的条件是 r² > ICC。ICC 本来就只有 0.75,意味着年龄只要解释七成半的真差,z 化后的「信度」就是负的,这个数已经没有意义。所以「z 化后再看 ICC 是否合格」这条判据,在它最想用的高 r² 场景里必然判死——不是终点不行,是判据自己造的。

    第三件事:r² 的分母被悄悄换过。 「年龄解释方差」可以指占总方差的比例,也可以指占组间真差的比例。同一个 ICC=0.75、r²=0.71 的数据,按前者算得 ICCz=0.138,按后者算得 0.465,差 0.327。前者判死、后者勉强及格,结论反向。源判据没写清分母是哪个,而这两套算法在文献里都有人用。一个判据的结论取决于一个没定义的分母,它就不是判据。

    第四件事:z 化真正改的是效应量,而且是往大改。 尺度缩了 √(1−r²) 倍,标准化效应量 d 就放大 1/√(1−r²) 倍:r²=0.30 是 1.195 倍,0.50 是 1.414 倍,0.71 是 1.857 倍。表面看这是好事——效应变大、样本可以省。但这是纯粹的记账错觉:SD 之所以变小,是因为年龄带来的个体差被除掉了,而那部分差里既有「年龄污染」,也有真实的个体差异。z 化不区分这两者,一律除掉。所以放大后的 d 对应的不是「效应更强」,而是「人群被人为削窄了」。ICC 依赖人群异质性,削窄人群必然同时压低 ICC——同一件事的两个面孔。

    第五件事:z 化不等于年龄配对。 配对抽样(1:1 配到同年龄)消掉的只是「年龄与分组的共线」,同龄人之间的个体差原样保留,SD 也只缩了年龄独有那一份。z 化则把整个年龄段内的全部个体差一起除掉。两者在「控年龄」这一层目标相同,在对「真个体差」的处理上完全不同:配对保它,z 化削它。所以「配对后组间差与 z 化后组间差差多少」本身就是一个诊断量——若差超过一个 MDC95,说明年龄与真个体差高度共线,任何 z 化都同时削掉了研究对象。

    那这条线该怎么落地。 原文给的实验设计(≥120 名成人、年龄跨度 ≥40 岁、隔 3–7 天 × 3 次 SVC-FOT,并算原始与 z 化两套 ICC/SEM/MDC95/组级门槛,另设年龄 1:1 配对对照)是合适的,问题只在判据的写法。按上面的代数,判据应当改成:若 z 化后 SEM 与 MDC95 变化 <5%(这是恒等式保证的必然结果)而 ICC 下降 >0.10,则证实「z 化治偏倚不治噪声」,删去 z 化条款,改为按年龄配对设计、报原始尺度 ICC;只有当 z 化后 SEM 真的降了 >15% 时,才说明年龄与测量误差本身耦合(即误差方差随年龄变,非同方差),z 化确实改善精度,原判据可留;若配对与 z 化的组间差相差超过 1 个 MDC95,则偏倚量级已足以翻转结论,一切组间比较须配对后重做。

    要说明的是,上面这些代数都只在「测量学」这一层成立,它们不涉及任何易学命题的真伪。本节讨论的终点(呼气末容积到 DR2 的余量)在现有对应网络里只挂到「金—肺」这条边,而这条边的现代证据栏至今空着。所以本文的所有结论只能读到这个程度:即使将来这条边被证实,「z 化能救终点」这个推论也是错的;反过来,如果这条边本身就不成立,谈终点精度更是空中楼阁。这是本文标 C 的理由。

    证据与局限

    已验证: 代数部分是恒等式,不依赖样本。SEM = SD·√(1−ICC) 在 z 化下严格不变(两种 r² 口径下实算 SEM 比均为 1.000000),ICCz = (ICC−r²)/(1−r²) 在 r²=0.30/0.50/0.71 三档下给出 0.643/0.500/0.138,要维持 ICCz≥0.75 需原始 ICC 达 0.825/0.875/0.928。ICC 本身依赖人群异质性、跨样本不可直接外推(PMID 29707825);异方差未校正会系统性高估 ICC(PMID 30208858)。

    部分验证: 「残差化/z 化不是万灵药」在别的领域已有实测支持——HRV 生物标志物研究中六种年龄归一化策略(含残差化、年龄段 z 分、多项式交互、简单除法)在严格验证下均无临床意义改善,部分反而更差;残差化处理共线性会引入系数与标准误偏倚(PMID 23017962)。肺功能侧,成人参考值普遍以身高为首要预测因子、年龄贡献在 >65 岁人群中被报为可忽略(PMID 16204589、PMID 29470844),而 FOT 阻抗的日间变异本身就有 5%–15%(PMID 23598954)——这说明「年龄这一项该不该扣」在不同人群里答案不同,不能一刀切。

    待验证: ① 余量(EELV−DR2)在 SVC-FOT 上的 ICC 与 SEM 从未被本领域实测报过,全部数值来自恒等式推演与跨领域对照;② 「年龄与测量误差耦合(异方差)」这一支是否存在,本篇无法判断——若存在,恒等式的前提(误差方差与年龄无关)就不成立,SEM 会真的变化;③ 年龄配对与 z 化在同一预期差下的检验力差异,需要实做才能给出量级。此外,本篇所依附的「金—肺」边在 L1 网络中仍为 open、现代证据待补,属结构性缺口。

    相关阅读

    证据等级: C
    参考文献: PMID 30208858, PMID 23017962, PMID 16204589, PMID 29470844, PMID 23598954, PMID 29707825, PMID 34232924, PMID 24338600

    L1 网络状态:| E-004 | 金 | 主 | 肺 | 《素问·阴阳应象大论》 | 待补 | L1 | open |(现代证据栏为「待补」,层级仅 L1 同构,状态 open;本文所讨论的呼吸力学终点在该网络中无独立边,故证据等级按硬规则封顶为 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。

    What Age Norms Z-Score Actually Changes: Zero Contribution to Detecting Small Differences, Only Making ICC Look Worse

    Evidence level: C

    The criterion in the previous post stated: “If age-explained variance exceeds 50%, switch to the margin after z-scoring by age.” This phrasing implies z-scoring rescues the endpoint. Algebraically, it does not: z-scoring does not change the measurement standard error (SEM) by even a fraction. It only alters the reliability coefficient number itself.

    Main Body

    Let’s unpack the previous statement. It claims: if age-explained variance is too large (e.g., r² = 0.71, from the correlation between DR2 and age), then this endpoint is actually measuring age rather than “sinking without sinking,” so we should switch to the margin after z-scoring by age. The first part of the logic holds—a metric with 70% of its variance eaten by age cannot be directly compared across groups. The problem lies in the “therefore z-score” step: it assumes z-scoring fixes something, yet this step lacks algebraic support.

    First point: SEM is an invariant. Consider the total variance VT = VB + VE in a population (VB is true between-subject difference, VE is measurement error). The reliability ICC = VB/VT, and the proportion of true difference VB explained by age is r². Z-scoring divides each person’s value by the SD of their age group. After this, the scale shrinks by √(1−r²), and ICC becomes (ICC−r²)/(1−r²)—which looks worse. However, plugging these into the SEM formula SEM = SD·√(1−ICC) reveals that SD shrinks by √(1−r²), while √(1−ICC) increases exactly by 1/√(1−r²), canceling each other precisely. SEM remains unchanged. Calculating three tiers (r² = 0.30 / 0.50 / 0.71, ICC = 0.75 and 0.90), SEM ratios are all exactly 1.000000. MDC95 = 1.96√2·SEM, naturally unchanged; the group-level minimum detectable difference = 1.96·√2·SEM/√n, also unchanged. This is not an approximation—it is an identity.

    Second point: ICC will drop, and can drop into negative numbers. When ICC = 0.75, r² = 0.30 falls to 0.643, r² = 0.50 falls to 0.500, and r² = 0.71 drops directly to 0.138—from “qualified” to “unqualified” in one step. More critically, the threshold: under this definition, ICCz drops below zero when r² > ICC. Since ICC itself is only 0.75, this means age must explain more than three-quarters of true difference for z-scored “reliability” to become negative—a number that is meaningless. Thus, the criterion “check if ICCz is qualified after z-scoring” inevitably fails in high r² scenarios—not because the endpoint is bad, but because the criterion itself is constructed to fail.

    Third point: the denominator of r² is silently swapped. “Age-explained variance” can refer either to the proportion of total variance or to the proportion of true between-subject difference. For the same ICC = 0.75 and r² = 0.71, calculating ICCz using the former yields 0.138, while using the latter gives 0.465—a difference of 0.327. The former fails qualification, the latter barely passes, reversing the conclusion. The source criterion does not specify which denominator is used, yet both algorithms appear in literature. A criterion’s conclusion depending on an undefined denominator renders it not a criterion at all.

    Fourth point: z-scoring actually changes effect size, making it larger. The scale shrinks by √(1−r²), so the standardized effect size d amplifies by 1/√(1−r²): r² = 0.30 gives 1.195×, 0.50 gives 1.414×, and 0.71 gives 1.857×. Superficially this seems beneficial—effect sizes grow, samples can shrink. But this is pure accounting illusion: SD shrinks because age-induced individual differences are divided out, and that part of difference contains both “age contamination” and true individual variation. Z-scoring does not distinguish these two; it removes both. Thus, the amplified d corresponds not to “stronger effects” but to “the population being artificially narrowed.” ICC depends on population heterogeneity; narrowing the population inevitably suppresses ICC—the two faces of the same coin.

    Fifth point: z-scoring is not age pairing. Paired sampling (1:1 matching to same age) removes only “age and group collinearity,” preserving individual differences among same-age people, so SD shrinks only by the age-specific portion. Z-scoring removes all individual differences within an entire age band. Both aim to control age, but differ fundamentally in handling “true individual differences”: pairing preserves them, z-scoring shrinks them. Thus, “how much the between-group difference after pairing differs from that after z-scoring” itself becomes a diagnostic metric—if this difference exceeds one MDC95, it indicates high collinearity between age and true individual differences, meaning any z-scoring simultaneously removes the study subjects.

    How to implement this line. The experimental design given (≥120 adults, age span ≥40 years, SVC-FOT repeated 3 times over 3–7 days, calculating both raw and z-scored ICC/SEM/MDC95/group thresholds, plus age 1:1 pairing control) is appropriate; the issue lies only in the criterion wording. Based on the algebra above, the criterion should be revised: if SEM and MDC95 change <5% after z-scoring (an inevitable result of the identity) while ICC drops >0.10, this confirms “z-scoring corrects bias but not noise,” delete the z-scoring clause, switch to age-paired design, and report raw-scale ICC; only if SEM truly drops >15% after z-scoring (indicating age and measurement error are coupled, i.e., error variance varies with age, not homoscedastic) does z-scoring genuinely improve precision, so keep the original criterion; if the between-group difference after pairing and z-scoring exceeds 1 MDC95, then bias magnitude is sufficient to flip conclusions, so all between-group comparisons must be re-done after pairing.

    It should be noted that the algebra above holds only at the “metrology” level and involves no Yijing Studies propositions of truth or falsity. The endpoint discussed here (residual from end-expiratory lung volume to DR2) in the current correspondence network connects only to the “gold-lung” edge, whose modern evidence column remains empty. Therefore, all conclusions in this article can only reach this extent: even if this edge is confirmed in the future, the inference “z-scoring rescues the endpoint” is wrong; conversely, if this edge itself does not hold, discussing endpoint precision is building on air. This is the reason for evidence level C.

    Evidence and Limitations

    Verified: The algebraic parts are identities, independent of sample size. SEM = SD·√(1−ICC) remains strictly unchanged under z-scoring (both r² definitions yield SEM ratios of exactly 1.000000). ICCz = (ICC−r²)/(1−r²) gives 0.643/0.500/0.138 for r² = 0.30/0.50/0.71 respectively; maintaining ICCz ≥ 0.75 requires raw ICC to reach 0.825/0.875/0.928. ICC itself depends on population heterogeneity and cannot be directly extrapolated across samples (PMID 29707825); uncorrected heteroscedasticity systematically overestimates ICC (PMID 30208858).

    Partially verified: “Residualization/z-scoring is not a panacea” has empirical support in other fields—HRV biomarker studies found six age-normalization strategies (including residualization, age-band z-scores, polynomial interaction, simple division) showed no clinically meaningful improvement under strict validation, with some performing worse; residualization introduces coefficient and standard error bias due to collinearity (PMID 23017962). In lung function, adult reference values universally use height as the primary predictor, with age contribution reported as negligible in populations >65 years (PMID 16204589, PMID 29470844), while FOT impedance’s within-day variability itself ranges 5%–15% (PMID 23598954)—indicating whether “age should be subtracted” varies by population, precluding a one-size-fits-all approach.

    To be verified: ① ICC and SEM for the margin (EELV−DR2) on SVC-FOT have never been empirically reported in this field; all values derive from identity-based deduction and cross-domain comparison; ② whether the branch “age and measurement error coupling (heteroscedasticity)” exists cannot be judged here—if it does, the identity’s premise (error variance independent of age) fails, and SEM would truly change; ③ power differences between age pairing and z-scoring under the same expected difference require actual implementation to quantify. Additionally, the “gold-lung” edge this article relies on remains open in the L1 network, with modern evidence pending, constituting a structural gap.

    Further Reading

    Evidence level: C
    References: PMID 30208858, PMID 23017962, PMID 16204589, PMID 29470844, PMID 23598954, PMID 29707825, PMID 34232924, PMID 24338600

    L1 network status: | E-004 | 金 | 主 | 肺 | 《素问·阴阳应象大论》 | 待补 | L1 | open | (modern evidence column marked “pending,” level only L1 isomorphic, status open; the respiratory mechanics endpoint discussed here has no independent edge in this network, so evidence level capped at C by hard rules)


    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.

  • 「余量」这个新终点自己够不够格:MDC95 的层级混用,和年龄这个跑不掉的混杂

    Evidence level: C

    在问「修炼者的余量是不是更大」之前,得先问「这个余量在本设计里量得准不准」——而常见的写法恰好答错了这道题:它拿个体层面的门槛去卡组层面的差异。

    正文

    上一篇把「沉而不塌」从 X5 电平改判成「呼气末到 DR2 的距离」,并给了一条前置标定:先算余量的 ICC 与 MDC95,不合格就别进主研究。这条要求方向是对的,但判据本身写歪了三处,而这三处都会把「测不出来」误诊成「这个指标不行」。

    第一处,MDC95 与组间差不是一个层级的东西。 MDC95 的定义是:同一个人重复测两次,差值超过这个数才可认为是真变化——系数是 1.96√2 ≈ 2.77 个标准误。它服务的是个体决策。判组间均值差用的是另一样东西:1.96×SE×√(2/n)。两者一除,倍数就是 √n/√2 再乘回 1.96√2 的比例关系——n=20 时 MDC95 是组级门槛的 4.47 倍,n=40 时 6.32 倍,n=60 时 7.75 倍。换句话说,「MDC95 > 组间差 ⇒ 判死」这条判据在 n=40/组时,等价于要求组间差达到 12.4 个标准误。任何真实效应都过不了这一关。文献里这个错法有专名,被反复写过:MDC 是「个体显著变化的最小量」,把它当「组级重要差异」用是误用(PMID 34232924);组级均值差的显著门槛在样本量≥6 且信度≤0.90 时必然低于个体门槛(同源)。

    第二处,这条判据其实是一个伪装成信度检查的效应量检查。 把 MDC95 ≤ Δ/2 展开(MDC95 = 2.77×SD×√(1−ICC)),等价于 ICC ≥ 1 − (d/5.54)²,其中 d = Δ/SD。代入具体数:想让 ICC=0.75 通过,得 d ≥ 2.77,即组间差要大到接近 3 个个体标准差。ICC=0.90 时才降到 d ≥ 1.75。也就是说,判据判的从来不是「仪器准不准」,而是「效应大不大」——而效应大小恰恰是研究要回答的问题,不能拿它当前提。

    第三处,年龄的解释方差这个数,在 n=40 下本身不稳。 判据设了 30% 与 50% 两条线。可真值固定在 0.30 时,n=40、四个协变量的观测 R² 落在 30% 以下的概率是 27.0%,落在 50% 以上的概率 15.7%——两条线的误判率都不低。真值 0.50 时,观测值跌到 30% 以下的概率 1.4%,但仍有 31% 的机会落在 30%–50% 的中间档。用 40 人给一个回归 R² 划线,划出来的主要是抽样噪声。

    更根本的问题在顺序。 源设计是「20-35 岁与 55-70 岁各半」,而 DR2 与年龄的 r²=0.71(PMID 30702975)。这个 r² 在 n=36 的原研究里,95% 置信区间是 [0.505, 0.841]——下界正踩在判据的 50% 线上。这意味着「年龄解释方差 >50% ⇒ 余量是年龄代理」这条判据,在最该用的那个场景里,光凭抽样波动就能来回翻转。而且 ICC 本身依赖人群异质性:个体间差别越小 ICC 越低,同样一台仪器、同样的误差,换个更同质的样本 ICC 就会掉(PMID 29707825)。年龄各半的抽样故意把异质性拉大,ICC 会显得好看;可一旦 z 化把年龄成分扣掉,SD 缩到原来的 0.539(r²=0.71),ICC 又可能跟着掉。先测 ICC、后 z 化,顺序反了。

    还有一层,余量这个量本身是「差」而不是「电平」。 它等于两个都带误差的量相减:DR2 是拟合出来的点估计,呼气末容积也带测量误差。两个量的误差在相减时相加而不是抵消。FOT 的实测数据给了量级参考:在潮气量、两倍潮气量、50% 吸气储备量三个容积下,Rrs 与 Xrs 的 ICC 是 0.89–0.98,但到了残气量与 50% 呼气储备量,落到 0.60–0.98(PMID 38036081)——正是本假说关心的低容积端,重复性最差。IOS 的另一组长期随访则显示,电抗参数的 ICC 系统性低于电阻参数(PMID 36131305),更早的经典研究里单次测量 ICC 只有 0.75(非哮喘)与 0.56(哮喘)(PMID 2602594)。用这些数代入:SD=6 个百分点、ICC=0.75,MDC95 约 8.3 个百分点——而源文章举的例子(修炼者 12 个百分点 vs 新手若差 5 个百分点)在这个门槛下必然被「判死」。可同样的数,组级只要 3 人/组就能检出 5 个百分点的差。

    那正确的顺序是什么。 不是先算 MDC 再卡关,而是先把三个量分开登记:① 组级最小可检测差(由 SD、ICC、n 三者共同决定,随 n 可调);② 个体级 MDC95(只用于判断单个人是否真变了,不进组间判据);③ 年龄常模 z 化后的残余 SD 与残余 ICC(这才是余量该被比较的尺度)。判据应该写成「组级最小可检测差 ≤ 预期差的 1/2」,而不是「MDC95 ≤ 预期差的 1/2」——前者 n=40、SD=6、ICC=0.75 时是 1.31 个百分点,后者是 8.3 个百分点,差六倍。

    证据与局限

    已验证: MDC 与 MCID 是两个层级的概念,把个体级 MDC 当组级重要差异用属误用(PMID 34232924);组级均值差的显著门槛在 n≥6 且信度≤0.90 时低于个体门槛(同源);MDC 的经典公式基于经典测验理论、假定全量表误差齐性,IRT 的 cMDC 会随分数位置变化(PMID 35670017);FOT 阻抗参数在不同肺容积下重复性差异明显,低容积端 ICC 降至 0.60–0.98(PMID 38036081);IOS 电抗参数长期重复性低于电阻参数(PMID 36131305);FOT 单次测量 ICC 在非哮喘 0.75、哮喘 0.56(PMID 2602594);ICC 依赖人群异质性,同质样本会压低 ICC,跨人群不可外推(PMID 29707825);信度研究的样本量应按区间精度反解而非按点估计(PMID 24338600、PMID 14673940);行为测量的重测信度估计受样本量强烈影响(PMID 40119099);方差分量在参数空间边界时 LRT 的零分布不是标准 χ²(PMID 16474914)。

    部分验证: 「余量」这个终点从未在任何人群被实测过,其 ICC、MDC95、组级最小可检测差全部是本文按 FOT 相邻参数的文献值代入推算,不是实测。年龄 r²=0.71 来自 18 例哮喘 + 18 例健康对照的横断面(PMID 30702975),样本小、且是 DR2 位置而非余量的 r²。本文给出的倍数关系(4.47/6.32/7.75 倍)是代数恒等式,可靠;但代入的具体 SD=6 个百分点是假设值。

    待验证: 余量在健康成人(青年/老年分层)的实际 ICC 与 MDC95;年龄 z 化后 ICC 是升是降(理论上 SD 缩小会压 ICC,但年龄成分去掉也可能让个体排序更稳定,方向未定);DR2 检出算法本身的重复性;余量的误差是否真如本文假设的那样在两个分量间相加;组级最小可检测差与源设计的预期差是否留有余量。

    L1 网络状态:「该主题尚未进入 L1 对应网络」——CORRESPONDENCE_NETWORK.md 全文 178 行,节点表 2.1/2.2/2.3 三域中均无「呼气末肺容积」「气道闭合」「DR2 / 去复张拐点」「余量」这类节点;生理侧只列有「呼吸率 / 呼吸模式」与「胃肠蠕动(气沉丹田候选物)」,物理侧列有「相位 / 节律」,三者均未与易学侧任何节点连边。边表 3.1–3.5 仅登记五行↔五脏/五色/五音/五方/五季/五志、八卦↔五行、天干地支↔五行、纳音↔频率(E-050,状态 open、现代证据栏写「尚无任何可测来源」);与肺相关的唯一一条是 E-004(金 → 肺,传统出处《素问·阴阳应象大论》,现代证据栏「待补」,层级 L1,状态 open)。EDGE_TABLE_science-yixue.md 中「丹田」仅 51 篇文章计数、「气功」48 篇,无层级与状态判定。故本文封顶 C,全部机制表述为方法学推演,不得读作已验证的生理机制。

    相关阅读

    证据等级: C
    参考文献: PMID 30702975, PMID 34232924, PMID 35670017, PMID 38036081, PMID 36131305, PMID 2602594, PMID 29707825, PMID 24338600, PMID 14673940, PMID 40119099, PMID 16474914


    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。

    Is \”Remnant\” a Qualified Endpoint: Level Mixing in MDC95 and the Inevitable Confounding of Age

    Evidence level: C

    Before asking whether “the margin is larger for practitioners,” one must first ask whether “this margin is measured accurately within this design”—yet the common formulation answers this question incorrectly: it applies an individual-level threshold to assess group-level differences.

    Main Text

    The previous entry reclassified “sinking without collapsing” from an X5-level metric to “the distance from end-expiration to DR2,” and provided a prerequisite calibration: first calculate the ICC and MDC95 of the margin; if they fail, do not proceed to the main study. This direction is correct, but the criteria themselves are misstated in three places, all of which risk misdiagnosing “cannot be measured” as “this indicator is invalid.”

    First, MDC95 and group differences operate at different levels. The definition of MDC95 is: a difference exceeding this value between two repeated measurements on the same individual is considered true change—the coefficient is 1.96√2 ≈ 2.77 standard errors. It serves individual decision-making. Assessing the mean difference between groups uses a different metric: 1.96×SE×√(2/n). Dividing the two yields a ratio of √n/√2 multiplied by 1.96√2. For n=20, MDC95 is 4.47 times the group-level threshold; for n=40, 6.32 times; for n=60, 7.75 times. In other words, the criterion “MDC95 > group difference ⇒ reject” implies that for n=40 per group, the group difference must reach 12.4 standard errors. Any true effect would fail this test. This error has a specific name in the literature, repeatedly noted: MDC is “the minimum amount of individual significant change,” using it as “group-level important difference” is misuse (PMID 34232924); the significance threshold for group mean differences is necessarily lower than the individual threshold when sample size ≥6 and reliability ≤0.90 (same source).

    Second, this criterion is actually an effect size check disguised as a reliability check. Expanding MDC95 ≤ Δ/2 (where MDC95 = 2.77×SD×√(1−ICC)) is equivalent to ICC ≥ 1 − (d/5.54)², where d = Δ/SD. Substituting specific values: to pass with ICC=0.75, we need d ≥ 2.77, meaning the group difference must be nearly 3 individual standard deviations. ICC=0.90 only lowers this to d ≥ 1.75. Thus, the criterion never judges “how accurate the instrument is,” but rather “how large the effect is”—and effect size is precisely the question the study must answer, not its premise.

    Third, the explained variance of age is unstable itself at n=40. The criterion sets two lines: 30% and 50%. If the true value is fixed at 0.30, with n=40 and four covariates, the probability that the observed R² falls below 30% is 27.0%, and above 50% is 15.7%—both error rates are substantial. At true value 0.50, the probability of dropping below 30% is only 1.4%, yet there remains a 31% chance of falling into the intermediate 30%–50% range. Drawing a regression R² line with only 40 people primarily captures sampling noise.

    The more fundamental issue lies in the order. The source design is “half 20–35 years old, half 55–70 years old,” while DR2 and age have r²=0.71 (PMID 30702975). This r² in the original study with n=36 has a 95% confidence interval of [0.505, 0.841]—the lower bound exactly hits the criterion’s 50% line. This means “age-explained variance >50% ⇒ margin is an age proxy” flips back and forth solely due to sampling fluctuation in the very scenario where it should apply. Moreover, ICC itself depends on population heterogeneity: smaller inter-individual differences yield lower ICC; with the same instrument and error, switching to a more homogeneous sample lowers ICC (PMID 29707825). Halving the age distribution intentionally increases heterogeneity, making ICC appear favorable; but once z-scoring removes the age component, SD shrinks to 0.53 of its original value (r²=0.71), and ICC may drop accordingly. Measuring ICC first, then z-scoring, reverses the proper order.

    Furthermore, the margin itself is a “difference,” not a “level.” It equals the subtraction of two quantities each carrying error: DR2 is a fitted point estimate, and end-expiratory volume also carries measurement error. Errors add rather than cancel when subtracting. FOT real-world data provides scale reference: at tidal volume, double tidal volume, and 50% inspiratory reserve volume, Rrs and Xrs ICC ranges from 0.89–0.98, but at residual volume and 50% expiratory reserve volume, it drops to 0.60–0.98 (PMID 38036081)—precisely the low-volume end of interest in this hypothesis. IOS long-term follow-up shows reactive parameters have systematically lower ICC than resistive parameters (PMID 36131305); in earlier classic studies, single-measurement ICC was only 0.75 (non-asthmatic) and 0.56 (asthmatic) (PMID 2602694). Substituting these values: SD=6 percentage points, ICC=0.75, MDC95 ≈ 8.3 percentage points—yet the source article’s example (practitioner 12 percentage points vs novice difference of 5 percentage points) would inevitably be “rejected” under this threshold. However, with the same numbers, group-level differences of just 3 people/group can detect a 5 percentage point difference.

    So what is the correct order? Not calculating MDC first to set a gate, but registering three quantities separately: ① minimum detectable group difference (determined jointly by SD, ICC, and n, adjustable with n); ② individual-level MDC95 (used only to judge whether a single person truly changed, not for group comparison); ③ residual SD and residual ICC after z-scoring age (this is the scale on which margin should be compared). The criterion should read “minimum detectable group difference ≤ half the expected difference,” not “MDC95 ≤ half the expected difference”—the former yields 1.31 percentage points for n=40, SD=6, ICC=0.75, versus 8.3 percentage points for the latter—a sixfold difference.

    Evidence and Limitations

    Verified: MDC and MCID are concepts at different levels; using individual-level MDC as group-level important difference is misuse (PMID 34232924); the significance threshold for group mean differences is lower than individual thresholds when n≥6 and reliability≤0.90 (same source); the classic MDC formula assumes classical test theory with uniform full-scale error, whereas IRT cMDC varies with score position (PMID 35670017); FOT impedance parameters show marked repeatability differences across lung volumes, with low-volume ICC dropping to 0.60–0.98 (PMID 38036081); IOS reactive parameters have lower long-term repeatability than resistive parameters (PMID 36131305); FOT single-measurement ICC is 0.75 for non-asthmatics and 0.56 for asthmatics (PMID 2602694); ICC depends on population heterogeneity, homogeneous samples depress ICC, cross-population extrapolation is invalid (PMID 29707825); reliability study sample size should be derived from interval precision rather than point estimates (PMID 24338600, PMID 14673940); test-retest reliability estimation for behavioral measures is strongly influenced by sample size (PMID 40119099); variance components at parameter space boundaries do not follow standard χ² distribution under LRT (PMID 16474914).

    Partially Verified: The endpoint “margin” has never been directly measured in any population; its ICC, MDC95, and minimum detectable group difference are all derived by substituting FOT adjacent parameter values from literature, not measured. Age r²=0.71 comes from a cross-sectional study of 18 asthmatics and 18 healthy controls (PMID 30702975), with small sample size and representing DR2 position rather than margin r². The multiplicative ratios (4.47/6.32/7.75 times) are algebraic identities, reliable; however, the substituted SD=6 percentage points is an assumed value.

    To Be Verified: Actual ICC and MDC95 of margin in healthy adults (stratified by youth/elderly); whether ICC rises or falls after z-scoring age (theoretically SD reduction depresses ICC, but removing age components may stabilize individual ranking, direction uncertain); repeatability of the DR2 detection algorithm itself; whether margin error truly adds between components as assumed; whether minimum detectable group difference leaves sufficient margin relative to the source design’s expected difference.

    L1 Network Status: “This topic has not yet entered the L1 corresponding network”—CORRESPONDENCE_NETWORK.md contains 178 lines, and node tables 2.1/2.2/2.3 across three domains lack nodes such as “end-expiratory lung volume,” “airway closure,” “DR2 / derecruitment inflection point,” or “margin”; the physiological side lists only “respiratory rate / respiratory pattern” and “gastrointestinal motility (candidate for Qi Chen Dan Tian),” while the physical side lists “phase / rhythm,” none of which connect to any Yijing-side nodes. Edge tables 3.1–3.5 only register Five Phases ↔ Five Organs/Five Colors/Five Tones/Five Directions/Five Seasons/Five Emotions, Eight Trigrams ↔ Five Phases, Heavenly Stems and Earthly Branches ↔ Five Phases, Na Yin ↔ Frequency (E-050, status open, modern evidence column states “no measurable source yet”); the only lung-related entry is E-004 (Metal → Lung, traditional source Su Wen · Yin Yang Ying Xiang Da Lun), with modern evidence column “to be supplemented,” level L1, status open. EDGE_TABLE_science-yixue.md lists only 51 articles for “Dan Tian” and 48 for “Qigong,” without level or status determination. Therefore, this paper caps at C; all mechanism descriptions are methodological derivations and must not be read as verified physiological mechanisms.

    Related Reading

    Evidence Level: C
    References: PMID 30702975, PMID 34232924, PMID 35670017, PMID 38036081, PMID 36131305, PMID 2602694, PMID 29707825, PMID 24338600, PMID 14673940, PMID 40119099, PMID 16474914


    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.

  • 体位占比降了,可能只是别人的方差涨了:一个不随样本量消失的偏倚

    Evidence level: C

    「训练组体位占比更低」这句话,可能一个字都不涉及体位——它可能只是「训练组的个体差异更大」。这两件事在占比这一个数上长得一模一样,而且样本量越大,越分不开。

    正文

    一、先把「占比」写成公式,问题就露出来了

    丹田三模态测量的方差分析把总方差拆成五个源:体位、带位、操作员、个体、残差。设它们的方差分量是 v₁ 到 v₅,那么「体位占 28%」这句话的准确写法是 p₁ = v₁ /(v₁+v₂+v₃+v₄+v₅)。分子是我们要研究的体位,分母是另外四个源加上它自己。

    组间比较时,我们算的是训练组与新手组的 p₁ 之差,但我们真正想问的,是两组 v₁ 之差。这两件事之间隔着一个分母 S = Σvᵢ。只要 S 在两组之间不一样,p₁ 的差就不再是 v₁ 的差的影子。这不是统计技巧问题,是「你报的数和你想说的数不是同一个数」。

    二、分母不是背景,它是第五个自变量

    对 p₁ 取对数再差分,得到一条恒等式:Δln p₁ = Δln v₁ − Δln S,而 Δln S 近似等于 Σpᵢ·Δln vᵢ。也就是说,占比的变化里,只有一部分来自体位本身,另一部分是另外四个源变化的加权和,被减号挂在了前面。

    按源设计的份额(0.28 / 0.27 / 0.15 / 0.20 / 0.10)算,Δln v₁ 与 Δln S 的相关系数是 0.59。这个正相关会在某些情形下部分抵消,但它同时也说明:分母不是静止的背景,它跟着所有源一起动,且动得不小。任何只看 p₁ 的结论,都在默认「S 没动」。

    三、等价折算:任何源的变化都能伪装成体位的假变化

    把上式整理,得到一条可以直接用的通式:当体位方差真值不变时,源 j 变化 δ_j 会在 p₁ 上制造出等价于体位变化 p_j·δ_j/(1−p₁) 的假象。个体方差翻倍(δ=0.693),等价于体位假降 0.139 个对数单位,即 13%。

    这条通式有一个反直觉的推论:若其余四个源按原比例同步缩放,那么折算权重 p_j/(1−p₁) 恒等于它的条件份额,与 p₁ 大小完全无关。实测把主源份额从 0.15 一路调到 0.70,第二源翻倍所对应的等价体位变化始终是 0.2599。所以「体位是最大的一头,它应该最不容易被带偏」这个直觉是错的——稳不稳不看主源占多少,看其余源相对它怎么变。

    四、抵消点:体位要真涨多少,才能让占比「看不见」

    反解通式可得抵消点。个体方差翻 1.5 倍时,体位方差需要真涨 11.9%,占比才会保持不动;翻 2 倍时需涨 21.2%;翻 2.5 倍需涨 29.0%;翻 3 倍需涨 35.7%。这些数字的含义是:一个真实存在的体位改善,完全可以被另一个源的膨胀抹平到「不显著」。

    反过来更糟。在体位真值毫无变化的前提下,按名义双侧 5% 水平做检验,占比法的假阳性率随个体方差膨胀迅速失控:涨 10% → 8.2%,涨 20% → 17.0%,涨 50% → 60.5%,翻倍 → 96.8%;而绝对量法始终稳在 5.0%。反解这条曲线:假阳性要摸到 20%,个体方差只需涨 22.5%;要摸到 80%,只需涨 66.6%。门槛低到任何一个真实人群的异质性波动都能跨过去。

    五、最要命的一条:加大样本量会让它更糟

    通常遇到不确定,第一反应是加人。这个坑不吃这一套。固定个体方差涨 20%,把每组人数从 20 一路加到 320,占比法假阳性率依次是 10.9%、17.0%、29.3%、51.7%、80.8%——单调上升。名义 95% 置信区间的实际覆盖率同步崩塌:涨 20% 时是 83.0%,涨 50% 时只剩 39.5%,翻倍时是 3.3%。

    原因很干脆:这是偏倚,不是噪声。噪声随样本量缩小,偏倚不动,于是信噪比里偏倚那一项越来越占主导。换句话说,这项研究做得越认真、样本越大、区间越窄,那个错误的结论就越「显著」。同时,占比法的标准误比绝对量法还小约 19%(0.0364 对 0.0447),它会用更窄的区间把偏倚报出来,看上去更精确。

    六、什么时候真的会翻转,而不是仅仅夸大

    两法结论真正相反,条件很干净:分母漂移的幅度超过分子本身的漂移,即 |δS| > |δ₁| 且方向一致时,占比法与绝对量法给出相反的符号。实测:体位涨 1.3 倍叠加个体涨 2.5 倍,比值 0.978,两法仍同向;再把带位也设成涨 2 倍,比值升到 1.69,翻转成立——绝对量报「体位升」,占比报「体位降」。

    把判据反过来用更有价值:若体位真涨 10%,个体方差只需涨 61% 就能让结论翻转;真涨 5%,个体只需涨 28%;真涨 15%,需涨 101%。而人群中个体差异的组间差,本来就常在 1.5 到 3 倍这个量级。

    七、出路:不是换统计量,是换报告对象

    三条可执行的规则。第一,组间推断只对绝对方差分量做,报 v₁ 的差值与区间,占比降级为描述性指标,不进假设检验。第二,凡报占比结论,必须同时报分母漂移量 δS;δS 与 δ₁ 同量级时该结论作废。第三,若确实要在组成框架里做,改用等距对数比(ilr)坐标,避免闭合约束带来的伪相关;但要注意这是近似而非恒等,2026 年的方法学综述指出相对值只是「几乎」满足子组成一致性。

    最后一点分寸:数学耦合并非总是致命。有实测研究把急性/慢性负荷比拆成耦合与非耦合两版,发现效应量几乎全是微不足道。这恰恰说明判据的价值——耦合致不致命,取决于 δS 相对 δ₁ 有多大,而这个比值是可以事前算出来的。

    证据与局限

    已验证: 比值型指标与其分母存在数学耦合,会产生伪相关,这是被反复确认的现象(PMID 29101104、PMID 25245392;后者在 315 人样本中显示,调整动脉管径后高/低风险组血流介导扩张的差异被减半);组成数据(各分量非负且和为定值)不能直接套用常规多变量方法,需用对数比变换(PMID 28555522、PMID 34737726);对闭合数据直接取算术均值、标准差与相关系数会产生系统偏误(PMID 28882426);相对丰度会掩盖绝对量的真实变化,需另做绝对定量(PMID 40689579、PMID 34146605);同类「比率随一个与目标无关的量变动」的偏倚在一致性统计量中亦有经典记录(kappa 受患病率影响,PMID 8501467、PMID 34520848);数学耦合的实际危害大小取决于具体场景,已有实测给出效应量微不足道的反例(PMID 31672929)。

    部分验证: 本文全部数字来自本人蒙特卡洛与解析计算(每格 40 万次模拟,n=40/组、σ=0.20、五源份额 0.28/0.27/0.15/0.20/0.10)。份额构型与 σ 均为设定值,非实测;「等价折算权重与 p₁ 无关」只在其余源按原比例缩放的假设下成立;δS 的一阶近似忽略了高阶项;假阳性与覆盖率的表格基于正态近似。

    待验证: 丹田场景下五个方差分量的真实组间差异量级(无任何实测,个体方差组间差 1.5–3 倍仅为类比猜测);三模态各自的份额观测噪声 σ 真实取值;Δln vᵢ 之间是否独立(若相关,S 的方差需重算);在本领域换用 ilr 后原有结论有多少条会翻转。

    L1 网络状态:「该主题尚未进入 L1 对应网络」——CORRESPONDENCE_NETWORK.md 逐词检索「占比 0 / 份额 0 / 比率 0 / 方差 0 / 分量 0 / 组成 0 / 闭合 0 / 分母 0 / 辛普森 0 / 判定 0 / 功效 0」,全部零命中;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,止于 E-050(「六十甲子纳音|对应|60 个频率」,层级 L1、状态 open,现代证据栏原文「尚无任何可测来源」),无任何以比率、分母、方差分量或组间比较为源或目标的边。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

    证据等级: C
    参考文献: PMID 29101104, PMID 25245392, PMID 28555522, PMID 34737726, PMID 28882426, PMID 40689579, PMID 34146605, PMID 8501467, PMID 34520848, PMID 31672929


    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。

    Positional bias may not vanish with sample size: it might just be others’ variance increasing

    Evidence level: C

    “The proportion of the training group is lower” might not involve the proportion at all—it may simply mean “the individuals in the training group have greater variability.” These two things look identical when expressed as a single proportion, and larger sample sizes make them even harder to distinguish.

    Body

    I. Writing the “Proportion” as a Formula Reveals the Problem

    The variance analysis of the Dantian three-modal measurement decomposes total variance into five sources: posture, belt, operator, individual, and residual. Let their variance components be v₁ to v₅, then the accurate statement of “posture accounts for 28%” is p₁ = v₁ / (v₁+v₂+v₃+v₄+v₅). The numerator is the posture we study, while the denominator includes the other four sources plus itself.

    When comparing groups, we calculate the difference in p₁ between the training and novice groups, but what we truly want to know is the difference between their v₁ values. These two quantities are separated by the denominator S = Σvᵢ. As long as S differs between groups, the difference in p₁ no longer reflects the shadow of the difference in v₁. This is not a statistical technique issue; it is “the number you report and the number you intend to convey are not the same.”

    II. The Denominator Is Not Background; It Is the Fifth Independent Variable

    Taking the logarithm of p₁ and differencing yields an identity: Δln p₁ = Δln v₁ − Δln S, where Δln S approximates Σpᵢ·Δln vᵢ. Thus, changes in proportion arise partly from posture itself and partly from the weighted sum of changes in the other four sources, subtracted at the front.

    Based on source design shares (0.28 / 0.27 / 0.15 / 0.20 / 0.10), the correlation between Δln v₁ and Δln S is 0.59. This positive correlation can partially offset under certain conditions, but it also demonstrates: the denominator is not static background; it moves with all sources and moves significantly. Any conclusion based solely on p₁ assumes “S did not move.”

    III. Equivalent Conversion: Any Source Change Can Masquerade as a False Posture Change

    Rearranging the equation yields a directly usable general formula: when the true posture variance remains constant, a change δⱼ in source j creates an illusion of posture change equivalent to pⱼ·δⱼ/(1−p₁). Doubling individual variance (δ=0.693) is equivalent to a false posture drop of 0.139 log units, i.e., 13%.

    This general formula has a counterintuitive implication: if the other four sources scale synchronously by their original proportions, then the conversion weight pⱼ/(1−p₁) equals its conditional share and is completely independent of p₁ magnitude. Testing adjusted the primary source share from 0.15 up to 0.70; doubling the second source always corresponds to an equivalent posture change of 0.2599. Thus, the intuition “posture is the largest component, so it should be least easily misled” is wrong—stability does not depend on the primary share size, but on how the other sources change relative to it.

    IV. The Cancellation Point: How Much Must Posture Truly Rise for the Proportion to Remain Invisible?

    Solving the general formula inversely yields the cancellation point. When individual variance increases 1.5-fold, posture variance must truly rise 11.9% for the proportion to remain unchanged; doubling requires 21.2%; 2.5-fold requires 29.0%; tripling requires 35.7%. These numbers mean: a real posture improvement can be completely erased by expansion of another source, rendering it “insignificant.”

    Conversely, the situation is worse. Under the assumption that posture remains unchanged, applying nominal two-sided 5% significance testing to the proportion method causes false positive rates to rapidly lose control as individual variance expands: rising 10% → 8.2%, 20% → 17.0%, 50% → 60.5%, doubling → 96.8%; while the absolute quantity method remains steady at 5.0%. Solving this curve inversely: to reach a false positive of 20%, individual variance needs only rise 22.5%; to reach 80%, just 66.6%. The threshold is so low that any real population heterogeneity fluctuation can cross it.

    V. The Most Dangerous Point: Increasing Sample Size Makes It Worse

    Usually encountering uncertainty, the first reaction is to add people. This pitfall does not follow that rule. Fixing individual variance increase at 20%, increasing group size from 20 up to 320, the proportion method false positive rates are sequentially 10.9%, 17.0%, 29.3%, 51.7%, 80.8%—monotonically rising. The actual coverage of nominal 95% confidence intervals simultaneously collapses: at 20% rise it is 83.0%, at 50% rise only 39.5%, doubling it is 3.3%.

    The reason is clear: this is bias, not noise. Noise shrinks with sample size, bias remains constant, so in the signal-to-noise ratio the bias term increasingly dominates. In other words, the more carefully this study is conducted, the larger the sample, the narrower the interval, the more “significant” the wrong conclusion becomes. Meanwhile, the proportion method’s standard error is about 19% smaller than the absolute quantity method (0.0364 vs 0.0447); it reports the bias with a narrower interval, appearing more precise.

    VI. When Will the Conclusion Truly Flip, Rather Than Just Be Exaggerated?

    The two methods yield truly opposite conclusions under clean conditions: the denominator drift magnitude exceeds the numerator’s own drift, i.e., |δS| > |δ₁| and they move in the same direction, then the proportion method and absolute quantity method give opposite signs. Testing: posture rises 1.3-fold combined with individual rising 2.5-fold, ratio 0.978, both methods still same direction; setting belt also rising 2-fold, ratio climbs to 1.69, flip occurs—absolute quantity reports “posture up,” proportion reports “posture down.”

    Using the criterion inversely is more valuable: if posture truly rises 10%, individual variance needs only rise 61% to flip the conclusion; true rise 5%, individuals need just 28%; true rise 15%, requires 101%. Meanwhile, inter-group differences in individual variance among real populations are often at the level of 1.5 to 3 times.

    VII. The Way Out: Not Changing the Statistic, But Changing the Reporting Target

    Three actionable rules. First, inter-group inference applies only to absolute variance components; report the difference and interval of v₁, downgrading proportion to a descriptive indicator, excluding hypothesis testing. Second, any proportion conclusion must simultaneously report the denominator drift δS; when δS and δ₁ are of similar magnitude, that conclusion is void. Third, if indeed working within the compositional framework, adopt isometric log ratio (ilr) coordinates to avoid spurious correlation from closure constraints; note this is approximate rather than exact, as the 2026 methodological review indicates relative values are only “almost” satisfying sub-compositional consistency.

    One final nuance: mathematical coupling is not always fatal. Measured studies have decomposed acute/chronic load ratios into coupled and uncoupled versions, finding effect sizes almost entirely negligible. This precisely illustrates the criterion’s value—coupling causes non-fatal outcomes depending on how large δS is relative to δ₁, and this ratio can be calculated in advance.

    Evidence and Limitations

    Verified: Ratio-type indicators are mathematically coupled with their denominators, producing spurious correlation, a repeatedly confirmed phenomenon (PMID 29101104, PMID 25245392; the latter showed in a 315-person sample that after adjusting arterial diameter, the difference in flow-mediated dilation between high/low risk groups was halved); compositional data (non-negative components summing to a constant) cannot directly apply conventional multivariate methods, requiring log-ratio transformation (PMID 28555522, PMID 34737726); directly taking arithmetic mean, standard deviation, and correlation coefficients from closed data produces systematic bias (PMID 28882426); relative abundance masks true absolute quantity changes, requiring separate absolute quantification (PMID 40689579, PMID 34146605); similar “ratio bias with variation in an irrelevant quantity” is also classically recorded in consistency statistics (kappa affected by prevalence, PMID 8501467, PMID 34520848); the actual severity of mathematical coupling depends on specific contexts, with measured studies providing counterexamples where effect sizes are negligible (PMID 31672929).

    Partially Verified: All numbers in this paper derive from my Monte Carlo and analytical calculations (400,000 simulations per cell, n=40/group, σ=0.20, five-source shares 0.28/0.27/0.15/0.20/0.10). Share configuration and σ are set values, not measured; “equivalent conversion weight independent of p₁” holds only under the assumption that other sources scale proportionally; δS’s first-order approximation ignores higher-order terms; false positive and coverage tables rely on normal approximation.

    To Be Verified: True inter-group magnitude of the five variance components in the Dantian scenario (no measurement, individual variance inter-group difference 1.5–3 times is purely analogical speculation); true values of observation noise σ for each of the three modalities; whether Δln vᵢ are independent (if correlated, S’s variance must be recalculated); how many original conclusions would flip after adopting ilr in this field.

    L1 Network Status: “This topic has not yet entered the L1 corresponding network”—CORRESPONDENCE_NETWORK.md word-by-word search for “proportion 0 / share 0 / ratio 0 / variance 0 / component 0 / composition 0 / closed 0 / denominator 0 / Simpson 0 / judgment 0 / power 0”, all zero hits; edge table 21 entries (E-001~E-005, E-010~E-014, E-030~E-034, E-040~E-044, E-050) all represent correspondence relationships among Five Phases/Five Organs/Five Colors/Five Tones/Five Directions/Five Seasons/Five Emotions/Eight Trigrams/Stems and Branches/Nayin, stopping at E-050 (“60 Jiazi Nayin | Corresponds | 60 Frequencies”, level L1, status open, original modern evidence column states “No measurable source yet“), no edges originating from or targeting ratios, denominators, variance components, or inter-group comparisons. EDGE_TABLE_science-yixue.md contains only concept article counts (Dantian 51 / Qigong 48 / Breathing 189), no level and status judgment. Hence this paper caps at C.

    Related Reading

    Evidence Level: C
    References: PMID 29101104, PMID 25245392, PMID 28555522, PMID 34737726, PMID 28882426, PMID 40689579, PMID 34146605, PMID 8501467, PMID 34520848, PMID 31672929


    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.