Category: 科学易学

  • 同一个 σ 区间,A 档门槛只有一端能过:区间端点方向闸门

    Evidence level: B

    本文性质:方法学 + 库内审计,draft,未进 accepted,未直写 779/782。
    一句话结论:σ ≥ 10⁻² → 1 km 山体 A ≥ 10 dB 这条 A 档门槛,在 σ 区间的最不利端只有 4.83 dB;它只在区间另一端成立。“区间能过”不等于”门槛成立”——这是本库此前 13 轮审计工具都没检出来的一类缺陷。


    一、问题从哪来:一个被反复引用、但没人验过方向的门槛

    池④ 这条线上,mountain-sample-priority-sigma-pre-screening-sop.md §二 的三档表是下地采样决策的总闸门:

    档位 山体类型 σ (S/m) 1 km A(dB) @ 7.83 Hz 建议
    A 档(优先测) 湿玄武岩、含金属硫化矿/磁铁矿山体、断裂带蚀变岩 ≥ 10⁻² ≥ 10 dB 第一批测点
    B 档(条件性测) 含水/含裂隙花岗岩、含黏土断层带、含盐水侵染灰岩 10⁻³–10⁻¹ 1–10 dB A 档测完仍有预算再上
    C 档(不测或仅作对照) 干燥未蚀变花岗岩、火成岩体外缘 ≤ 10⁻⁴ < 1 dB 跳过

    这张表已经被后面几篇文章反复引用(high-sigma-lithology-elf-shielding-path.md 的 K1 门槛、elf-skin-depth-table-numeric-integrity-audit.md §5、多份实验设计)。

    前几轮已经查出过它的一处硬错:A 档「σ ≥ 10⁻² → A ≥ 10 dB」在 σ=10⁻² 处只有 4.83 dB,被记为”门槛高估 ~2×”。

    但那个判定本身有个盲点:它拿 σ 区间的一个端点(10⁻²)去判,然后直接下”门槛错”的结论。σ 在这里是开区间(≥ 10⁻²,上限在表里没写),区间内 σ 越大 A 越大 —— 那么”门槛能否成立”就取决于你取区间哪一端。

    这正是本轮要堵的口子。


    二、物理事实:σ 与 δ、A 的方向是相反的

    三个式子决定了全部推理,不需要任何文献:

    δ = √(2 / (ω μ σ))        →  σ ↑ 则 δ ↓(开方反比)
    A(dB) ≈ 8.7 · d / δ        →  δ ↓ 则 A ↑(反比)
    ∴ A ∝ √σ(d、f、μ 固定时)
    

    所以在同一条 σ 区间里:

    • 取 σ 最小端 → δ 最大 → A 最小(门槛最容易不达标)
    • 取 σ 最大端 → δ 最小 → A 最大(门槛最容易达标)

    对 A ≥ 10 dB 这种下界断言,最不利端永远是 σ 最小端。反之,对 δ ≤ 67 km 这种上界断言,最不利端是 σ 最大端(真值要么本来就小,要么被压得更小)。

    换句话说:判一个带区间的门槛,必须先想清楚”最不利端点在哪一侧”。 这一步此前从来是隐式的,靠人记。


    三、把 A 档门槛两端的数算出来

    固定 f = 7.83 Hz、d = 1 km、μr = 1:

    A(σ=10⁻²) = 4.829 dB      ← 区间最小端(若 σ 真取到 10⁻²)
    A(σ=10⁻¹) = 15.271 dB
    A(σ=1)    = 48.29  dB
    

    反解「1 km 处 A 正好 10 dB」所需的 σ:

    δ_need = 8.7 × 1000 / 10 = 870 m
    σ_need = 2 / (2π × 7.83 × μ0 × 870²) ≈ 4.29 × 10⁻² S/m
    

    于是 A 档那条门槛可以写成一个精确的等价命题:

    σ ≥ 10⁻² → 1 km A ≥ 10 dB 成立,当且仅当 σ 实际落在 ≥ 4.29×10⁻²。
    若 σ 落在 [10⁻², 4.29×10⁻²) 这一段,1 km 处 A 不足 10 dB。

    也就是:这条门槛在它自己声明的区间里,只有约 0.63 个数量级(10⁻² → 4.29×10⁻²)是”不成立”的,其余都成立。

    这不是”门槛错了 2 倍”,而是一个更精确、也更有操作价值的描述:门槛的真实分界是 4.29×10⁻²,不是 10⁻²。差的那 0.63 个数量级,恰好是野外 σ 实测的分辨率边缘 —— 这正是它危险的地方。


    四、这不是孤例:15 条断言卡在同一处

    本轮把 delta-table-scan.py 从 v1.2 升到 v1.3,加了一个此前没有的判定:当 σ 是区间时,把声明值分别在两个端点下过一遍,看结论是否一致。四种结果:

    结果 含义 可下的结论
    both 两端都成立 无条件成立(最硬的一档)
    signed 只有一端成立 不能判对、也不能判错 —— 属”未量化方向”,必须报出来
    none 两端都不成立 无条件不成立
    None σ 是单点 无需端点判读

    全库跑下来(扫描根 .,780 全库):

    断言总数 344|minmatch 200|endpoint: both=9, signed=15, 单点=320
    

    15 条 signed。逐条看,最关键的是这三类:

    位置 声明 最不利端真值 结论
    mountain-sample-priority-sigma-pre-screening-sop.md:55(中英各 1) σ ≥ 10⁻² → 1 km A ≥ 10 dB 4.829 dB 端点不足,需改为 σ ≥ 4.29×10⁻²
    schumann-station-candidate-tier-a-wet-basalt.md:45(中英各 1)+ 3 处引用 σ ≤ 10⁻⁴ → δ ≥ 670 km 17.99 km 端点不成立(第 9–13 轮已判,本次由工具自动复现)
    elf-skin-depth-table-numeric-integrity-audit.md 等审计表 引用上两行的判读 — 属”引用了错值”,随源一起改

    注意第二条的意义:δ ≥ 670 km 这条,前 4 轮是靠人工逐格复算才判死的(且成因追了 4 轮仍未定)。v1.3 的端点判读是零人工、自动复现——这是工具第一次独立命中它。


    五、一个反直觉的副作用:门槛”太松”比”太严”更危险

    前面 elf-skin-depth-table-numeric-integrity-audit.md 把 A 档门槛记为「高估 ~2×」,方向读起来像是”门槛太严、把好东西挡在外面”。

    但按端点算清楚之后,实际效果是双向的:

    • 若 σ = 10⁻²(区间最不利端):A = 4.83 dB,达不到 10 dB → 这条”值得优先测”的标签给早了;
    • 若 σ = 10⁻¹(区间中间):A = 15.27 dB,远超 10 dB → 门槛又显得过严。

    所以这张表的真问题不是”偏 2 倍”,而是:它在用一个区间定义一个应该由单点决定的分档。σ 是采样后才知道的量(该 SOP 自己写”按 10 g 岩石样品实测 σ 决定分档”),却用采样前的文献区间去写判据 —— 逻辑上是”用粗粒度输入下细粒度结论”。

    这与本库此前记下的两类缺陷同族:

    • 池④ #51「区间跨档」(判据分辨率不够却给了单一结论)
    • 池③ #192「一条伪造引文改写核心数值结论」(结论精度超过输入精度)

    本轮的贡献是给出了第三个具体形态:区间端点的方向性被忽略——不是区间太宽,也不是数值算错,而是没说明用的是哪一端。


    六、把它变成可执行的三条规矩(R7–R9)

    写入 08-ops/experiments/pool4-sigma-endpoint-direction-governance.md(P4-AUDIT-009),摘要如下:

    R7(端点判读):凡 σ(或任何进公式的参数)以区间/开区间给出,而声明值是不等式断言时,必须在两端各判一次,并报告 both / signed / none。只判一端即视为未完成判定。已固化进 delta-table-scan.py 的 endpoint_check()。

    R8(signed 不得升级为”已修正”或”已证伪”):signed 只说明”方向未量化”,处置动作是补写端点(把 ≥ 10 dB 改成 σ ≥ 4.29×10⁻² 时 ≥ 10 dB),不是判错、也不是改数值。此条防止把”没写清”当成”算错了”去改正文。

    R9(判据参数与采样顺序必须一致):若某分档的判据参数(σ)在流程上是采样后才测得,则文献区间只能用于预筛层级,不得直接写成”测得值 Δ 门槛”。本条专门针对 SOP 类文件。


    七、证据与局限

    证据等级 E3(结构类比 + 解析复算),非 E1。

    • 全部为公式复算(δ=√(2/(ωμσ))、A≈8.7d/δ)与库内文本解析,无任何实测;
    • 4.29×10⁻² 是在 μr=1、均匀半空间、无位移电流假设下的自洽阈值;真实山体 σ 随频率弥散(Revil 2013)、μr 可 >1(Kariya & Shankland 1983 对下地壳干岩石的测量即显示强频散),故野外阈值会偏离这个数,方向不保证;
    • 本文判”门槛分界应为 4.29×10⁻²”仅指与表格自述 σ 自洽,不指野外真实分界;
    • 15 条 signed 中只精算了 3 类(A 档门槛、670 km、审计表引用),其余未逐条复算;
    • both 只有 9 条、signed 15 条,说明本库带区间的断言本来就少(320 条是单点 σ),因此”15 条”这个数不能外推为”库内缺陷率”;
    • 端点判读依赖 σ 区间的正确解析。开区间(≥/≤)的上/下界是工具人为给 3 个数量级余量的约定,不是文件写明的,这是本文最脆的一环;
    • 工具仍只认 markdown 表格 + 文本正则,正文散句中的区间断言覆盖不全(覆盖率 38/58 表 = 65.5%,含 F0 剔除 4 张后的分母口径)。

    未做:未改任何已发布正文(A 档门槛的文、670 km 的源文及其 5 处引用全部未动,等 JOE 批);未做 σ 实测;未新增「靠山→健康」通路(净增通路 0);未标 accepted;未直写 779/782。


    参考文献

    1. Revil A. Effective conductivity and permittivity of unsaturated porous materials in the frequency range 1 mHz–1 GHz. Water Resources Research, 2013. doi:10.1029/2012WR012700
      —— 用于说明 σ 的频率弥散:真实岩石电导率不是常数,本文的单一 σ 取值是简化。
    2. Laštovičková M, Bochníček J. A contribution to the measurement of frequency dependence of electrical conductivity of rocks. Studia Geophysica et Geodaetica, 1992. doi:10.1007/BF01613900
      —— 同上,岩石电导率频散的直接测量证据。
    3. Kariya KA, Shankland TJ. Electrical conductivity of dry lower crustal rocks. Geophysics, 1983;48:52–61. doi:10.1190/1.1441407
      —— 干燥下地壳岩石的电导率量级,用于界定 10⁻⁴–10⁻² 这档 σ 的合理性边界。
    4. Mehran M, Arulanandan K. Low frequency conductivity dispersion in clay-water-electrolyte systems. Clays and Clay Minerals, 1977;25:39–48. doi:10.1346/CCMN.1977.0250107
      —— 黏土/页岩(本表 B 档主要岩性之一)的低频电导率弥散,说明 B 档 “1–10 dB” 的区间宽度有物理来源。

    以上 4 个 DOI 均经 Crossref 逐条回查(标题/期刊/年份/作者一致)。本文无 PMID:4 篇均属地球物理与材料学期刊,不在 PubMed 索引范围,故按宪章 §2 只能记 E3,不构成 E1。


    附:复算脚本(可复现)

    MU0 = 4π×10⁻⁷ ; DBN = 8.685889638
    δ(f,σ,μr) = √( 2 / (2π f MU0 μr σ) )
    A(d,σ)    = DBN · d / δ
    
    A(1 km, 10⁻²) = 4.829 dB
    A(1 km, 10⁻¹) = 15.271 dB
    δ_need(A=10dB, d=1km) = 8.7×1000/10 = 870 m
    σ_need = 2/(2π·7.83·MU0·870²) = 4.29×10⁻² S/m
    

    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。

    Same σ interval, A-gear threshold passable only at one end: interval endpoint directional gate

    Evidence level: B

    Nature of this paper: Methodology + internal audit, draft, not yet accepted, not directly written for 779/782.
    One-sentence conclusion: For the A-threshold condition σ ≥ 10⁻² → 1 km mountain A ≥ 10 dB, the least favorable end of the σ interval yields only 4.83 dB; it holds true only at the opposite end of the interval. “Interval passability” does not equal “threshold validity”—this is a class of defect missed by the previous 13 rounds of audit tools in this repository.


    I. Where the Problem Originates: A Repeatedly Cited, Yet Unverified Directional Threshold

    On Pool ④, the three-tier table in §II of mountain-sample-priority-sigma-pre-screening-sop.md serves as the master gate for on-site sampling decisions:

    Tier Lithology Type σ (S/m) 1 km A(dB) @ 7.83 Hz Recommendation
    Tier A (Priority Measurement) Wet basalt, metal sulfide-bearing ore bodies, magnetite mine bodies, alteration rocks within fault zones ≥ 10⁻² ≥ 10 dB First batch of measurement points
    Tier B (Conditional Measurement) Water-bearing or fractured granite, clay-bearing fault zones, limestone infiltrated by saline water 10⁻³–10⁻¹ 1–10 dB Proceed only if budget remains after Tier A measurements
    Tier C (Skip or Reference Only) Dry unaltered granite, outer margins of igneous rock bodies ≤ 10⁻⁴ < 1 dB Skip

    This table has been repeatedly cited in subsequent articles (the K1 threshold in high-sigma-lithology-elf-shielding-path.md, §5 of elf-skin-depth-table-numeric-integrity-audit.md, and multiple experimental designs).

    Previous rounds identified a critical error in this table: the Tier A claim “σ ≥ 10⁻² → A ≥ 10 dB” yields only 4.83 dB at σ = 10⁻², leading to the conclusion of “threshold overestimation by ~2×”.

    However, that judgment itself contains a blind spot: it uses one endpoint of the σ interval (10⁻²) to evaluate the threshold, then directly concludes “the threshold is wrong.” Here, σ represents an open interval (≥ 10⁻², with the upper limit not specified in the table); within this interval, larger σ values correspond to larger A values. Therefore, whether “the threshold holds” depends entirely on which end of the interval is selected.

    This is precisely the gap we must address in this round.

    II. Physical Facts: σ and δ, A Direction Opposite

    Three equations determine the entire inference; no literature is required:

    δ = √(2 / (ω μ σ))        →  σ ↑ 则 δ ↓(开方反比)
    A(dB) ≈ 8.7 · d / δ        →  δ ↓ 则 A ↑(反比)
    ∴ A ∝ √σ(d、f、μ 固定时)
    

    Therefore, within the same σ interval:

    • Selecting the minimum σ end → δ is maximum → A is minimum (the threshold is most likely to fail)
    • Selecting the maximum σ end → δ is minimum → A is maximum (the threshold is most likely to be met)

    For assertions of a lower bound such as A ≥ 10 dB, the most unfavorable end is always the minimum σ end. Conversely, for assertions of an upper bound such as δ ≤ 67 km, the most unfavorable end is the maximum σ end (the true value is either already small or suppressed further).

    In other words: To judge a threshold with an interval, one must first clarify “which side is the most unfavorable endpoint.” This step was previously implicit, relying on human memory.

    III. Calculating the Threshold Values at Both Ends of Level A

    With fixed parameters f = 7.83 Hz, d = 1 km, and μr = 1:

    A(σ=10⁻²) = 4.829 dB      ← minimum end of the interval (if σ truly reaches 10⁻²)
    A(σ=10⁻¹) = 15.271 dB
    A(σ=1)    = 48.29  dB
    

    Solving for the σ required when A exactly equals 10 dB at 1 km:

    δ_need = 8.7 × 1000 / 10 = 870 m
    σ_need = 2 / (2π × 7.83 × μ0 × 870²) ≈ 4.29 × 10⁻² S/m
    

    Thus, the Level A threshold can be stated as a precise equivalent proposition:

    σ ≥ 10⁻² → 1 km A ≥ 10 dB holds if and only if σ actually falls in ≥ 4.29×10⁻².
    If σ lies in the interval [10⁻², 4.29×10⁻²), then A at 1 km is less than 10 dB.

    This means: within the interval it declares, only approximately 0.63 orders of magnitude (10⁻² → 4.29×10⁻²) render the threshold “not holding”; the rest do hold.

    This is not “the threshold being off by a factor of 2,” but rather a more precise and operationally valuable description: the true dividing line is 4.29×10⁻², not 10⁻². That missing 0.63 orders of magnitude corresponds precisely to the resolution edge of field measurements of σ — and this is exactly where the danger lies.

    IV. This is not an isolated case: 15 assertions converge at the same point

    This round upgrades delta-table-scan.py from v1.2 to v1.3, adding a previously absent validation: when σ is an interval, the declared value is tested at both endpoints to check for consistency. Four outcomes emerge:

    Outcome Meaning Conclusion
    both Both endpoints hold Unconditionally true (the strongest tier)
    signed Only one endpoint holds Cannot be confirmed nor rejected — belongs to “unquantified direction” and must be reported
    none Neither endpoint holds Unconditionally false
    None σ is a single point No endpoint evaluation needed

    Running the full repository (scanning root . across 780 files):

    Total assertions: 344 | minmatch: 200 | endpoint: both=9, signed=15, single-point=320
    

    15 signed assertions. Examining each individually, the most critical ones fall into these three categories:

    Location Declaration Worst-case endpoint truth value Conclusion
    mountain-sample-priority-sigma-pre-screening-sop.md:55 (1 Chinese, 1 English) σ ≥ 10⁻² → 1 km A ≥ 10 dB 4.829 dB Endpoint insufficient; must revise to σ ≥ 4.29×10⁻²
    schumann-station-candidate-tier-a-wet-basalt.md:45 (1 Chinese, 1 English) + 3 citations σ ≤ 10⁻⁴ → δ ≥ 670 km 17.99 km Endpoint fails (already determined in rounds 9–13; this time automatically reproduced by the tool)
    Audit tables such as elf-skin-depth-table-numeric-integrity-audit.md Referencing the two rows above — Belong to “citing incorrect values”; updated along with the source

    Note on the significance of the second item: The claim δ ≥ 670 km was previously invalidated only after manual recalculation of each cell over the first four rounds (and the underlying cause remained unresolved after four iterations). The v1.3 endpoint evaluation is zero manual effort, automatically reproduced — marking the first time the tool independently hit this result.

    V. A Counterintuitive Side Effect: “Too Loose” Thresholds Are More Dangerous Than “Too Strict”

    The preceding elf-skin-depth-table-numeric-integrity-audit.md records the A-grade threshold as “overestimated by ~2ד, which initially suggests the threshold is “too strict, blocking good things outside.”

    However, calculating precisely at the endpoints reveals the actual effect is bidirectional:

    • If σ = 10⁻² (the least favorable end of the interval): A = 4.83 dB, falling short of 10 dB → the label “worth prioritizing for measurement” was applied prematurely;
    • If σ = 10⁻¹ (the middle of the interval): A = 15.27 dB, far exceeding 10 dB → the threshold now appears overly strict.

    Thus, the true issue with this table is not “off by 2×”, but rather: it uses an interval to define a grade that should be determined by a single point. σ is a quantity known only after sampling (the SOP itself states “classify based on measured σ from 10 g rock samples”), yet the criterion is written using pre-sampling literature intervals — logically, this constitutes “deriving fine-grained conclusions from coarse-grained inputs.”

    This aligns with two previously recorded defects in this repository:

    • Pool ④ #51 “Interval spanning grades” (insufficient criterion resolution yet yielding a single conclusion);
    • Pool ③ #192 “A fabricated citation rewriting the core numerical conclusion” (conclusion precision exceeds input precision).

    This round contributes a third specific form: the directional nature of interval endpoints is ignored — not because the interval is too wide, nor because the calculation is wrong, but because it fails to specify which endpoint is used.

    VI. Executable Rules (R7–R9)

    Write to 08-ops/experiments/pool4-sigma-endpoint-direction-governance.md (P4-AUDIT-009), summary as follows:

    R7 (Endpoint Assessment): Whenever σ (or any parameter entering the formula) is given as an interval or open interval, and the declared value is an inequality assertion, both endpoints must be evaluated, reporting both / signed / none. Evaluating only one endpoint is considered incomplete. This rule is now fixed into delta-table-scan.py via endpoint_check().

    R8 (signed may not be upgraded to “corrected” or “falsified”): signed merely indicates “direction not quantified”; the remedial action is supplementing endpoints (changing ≥ 10 dB to σ ≥ 4.29×10⁻² when ≥ 10 dB), not correcting errors or altering numerical values. This rule prevents treating “unclear specification” as “calculation error” and thereby modifying the main text.

    R9 (Criterion parameters and sampling order must align): If the criterion parameter (σ) for a given tier is measured after sampling in the workflow, then literature intervals can only be used for pre-screening levels, and cannot be directly written as “measured value Δ threshold.” This rule specifically targets SOP-style documents.

    VII. Evidence and Limitations

    Evidence Level E3 (Structural Analogy + Analytical Recalculation), not E1.

    • All values are derived from formulaic recalculation (δ=√(2/(ωμσ)), A≈8.7d/δ) and in-text parsing, with no field measurements;
    • The value 4.29×10⁻² represents the self-consistent threshold under the assumptions of μr=1, a uniform half-space, and zero displacement current; true mountain conductivity σ exhibits frequency dispersion (Revil 2013), and μr can exceed 1 (Kariya & Shankland 1983 measured strong dispersion in dry lower-crust rocks), so the field threshold will deviate from this number and the direction is not may support;
    • The statement that “the threshold boundary should be 4.29×10⁻²” refers only to self-consistency with the σ values stated in the table, not to the actual field boundary;
    • Among the 15 signed entries, only three categories were precisely recalculated (E3 threshold, 670 km, and audit table citations); the remainder were not individually verified;
    • The count of both is only 9 entries, while signed totals 15, indicating that this repository contains few interval-based assertions (320 entries are single-point σ values); therefore, the number “15” cannot be extrapolated to represent the “in-house defect rate”;
    • Endpoint interpretation depends on correct parsing of σ intervals. The upper/lower bounds for open intervals (≥/≤) are conventions set by the tool with a human-imposed margin of three orders of magnitude, not explicitly stated in the files; this is the most fragile link in this paper;
    • The tool still recognizes only markdown tables and text regex, leaving interval assertions scattered in the prose undercovered (coverage 38/58 tables = 65.5%, with the denominator excluding the 4 F0 tables).

    Not done: no changes to any published text (the E3 threshold text, the source for 670 km and its five citations remain untouched pending JOE approval); no σ field measurements; no new “backrest→health” pathway added (net increase in pathways = 0); not marked accepted; did not directly write 779/782.

    References

    1. Revil A. Effective conductivity and permittivity of unsaturated porous materials in the frequency range 1 mHz–1 GHz. Water Resources Research, 2013. doi:10.1029/2012WR012700
      —— Used to explain the frequency dispersion of σ: the true electrical conductivity of real rock is not constant; the single σ value used here is a simplification.
    2. Laštovičková M, Bochníček J. A contribution to the measurement of frequency dependence of electrical conductivity of rocks. Studia Geophysica et Geodaetica, 1992. doi:10.1007/BF01613900
      —— Same as above, direct measurement evidence of frequency dispersion in rock conductivity.
    3. Kariya KA, Shankland TJ. Electrical conductivity of dry lower crustal rocks. Geophysics, 1983;48:52–61. doi:10.1190/1.1441407
      —— Magnitude of electrical conductivity for dry lower crustal rocks, used to define the reasonable boundary of the 10⁻⁴–10⁻² range for σ.
    4. Mehran M, Arulanandan K. Low frequency conductivity dispersion in clay-water-electrolyte systems. Clays and Clay Minerals, 1977;25:39–48. doi:10.1346/CCMN.1977.0250107
      —— Low-frequency conductivity dispersion in clay/shale (one of the main lithologies in this table’s Category B), indicating that the “1–10 dB” interval width for Category B has a physical basis.

    All four DOIs above have been individually verified via Crossref (title/journal/year/authors consistent). This paper has no PMID: all four articles belong to geophysical and materials science journals, outside the PubMed indexing scope; therefore, per Charter §2 they can only be recorded as E3 and do not constitute E1.

    Appendix: Recalculation Script (Reproducible)

    MU0 = 4π×10⁻⁷ ; DBN = 8.685889638
    δ(f,σ,μr) = √( 2 / (2π f MU0 μr σ) )
    A(d,σ)    = DBN · d / δ
    
    A(1 km, 10⁻²) = 4.829 dB
    A(1 km, 10⁻¹) = 15.271 dB
    δ_need(A=10dB, d=1km) = 8.7×1000/10 = 870 m
    σ_need = 2/(2π·7.83·MU0·870²) = 4.29×10⁻² S/m
    

    Disclaimer: This document is a compilation, comparison, and research design discussion of traditional knowledge and scientific literature, intended solely for understanding and reference. It does not constitute medical advice and cannot replace professional diagnosis or treatment. The effect sizes, frequency ranges, and exposure conclusions mentioned herein are largely inferences at the research report level; individual variations can be significant. If you experience physical discomfort or health concerns, please seek medical attention and have them evaluated by a qualified physician.

  • 「掉了多少」和「掉了百分之几」会不会给出相反结论:同一批停练数据在三种口径下的翻转率

    Evidence level: C

    同一批停练数据,换个算法就把结论翻过来——这件事到底有多容易发生,从来没人量过。

    正文

    上一轮我们已经指出:保留率这个比值,近似于在检验基线差。同等绝对降幅下,起点高的一组必然保留率更高。但那只说明「可能会偏」,没说明偏多少。差 0.03 是记法问题,差到让一半的组间比较换向,就是既往文献全部要重算的问题。这一轮要把它量出来。

    第一层是数学耦合。 保留率 = 末值 ÷ 峰值。峰值既是分子的一部分,又是分母。这种「比率变量进回归」的问题,统计学界讨论了一个多世纪,医学界却几乎没理会。Tu 等人重分析了两篇牙周病学论文,发现在比率变量上建相关与回归严重违反方法假设,两篇论文的结论实质上误导,其推荐意见不能成立(PMID 14749086)。停练研究里保留率常常直接被当作因变量放进组间比较,这与被重分析的那两篇是同一个结构。

    第二层是天花板压缩,它和耦合叠加。 中风后「比例恢复规则」说多数人恢复固定比例(约 70%)的功能损失,其经验支持就是初始分与后续变化量之间的强负相关。Bowman 等人专门论证:这个相关被数学耦合与向天花板压缩共同混淆,且这种混淆在个体层推断和组层推断里都存在,不因为把样本聚合就消失(PMID 33827246)。停练数据里「峰值」正是天花板本身——练到顶了,再怎么练也上不去,一停练的下降幅度天然被峰值封住。天花板的连带后果还有:最小临床重要改善会随基线升高而变大,原因不是患者期望不同,而是地板与天花板(PMID 24556220);功能独立性量表的原始增益被天花板支配,改成归一到 0–1 的可获得空间占比才把这块偏差消掉(PMID 25573600)。

    第三层:真实数据上的翻转量级,其实已经有实测锚点。 骨密度监测是最干净的案例。指南建议用绝对值(g/cm²)判断随访变化,临床却普遍用百分比。Leslie 用同一个数据库 1317 对脊柱、1259 对髋部扫描比较两种口径:用合并的相对最小显著变化判「变化超限」,低骨密度组 33.5%、高组 20.0%,差异极显著(P<0.0001);换成绝对口径,27.8% 对 23.4%,差异消失(P=0.13)。更关键的是分类不一致率——相对口径下 13.6% 的人被分进另一类,绝对口径只有 1.7%(PMID 16537116)。也就是说,口径一换,显著性翻转,约七分之一的人换了归类。这就是我们要在停练数据上测的那个「翻转率」的现成量级。

    第四层:变化分数与协方差分析本身就会打架。 Lin 与 Larzelere 在 4753 名美国青少年的真实数据上复现了 Lord 悖论:差分分数与协方差分析给相反结果。他们还做了一件更重要的事——检验两种「一致化」方法(双中心化协方差分析与前测匹配),结果两者各自复现一套结论,彼此之间的差距和原始矛盾一样大,并没有谁更接近真相(PMID 33242670)。Cappelleri 与 Cislo 把机制说透了:一旦对基线做调整,得到的临床意义阈值反映的是一个「人为让各锚定转换层基线相同」的场景,未调整版才保住了外部效度(PMID 37526447)。Sjölander 等进一步指出,当分数带测量误差时,调整与不调整各自对准不同的目标参数,不存在哪一个是天然正确的(PMID 37043152)。三篇合起来的意思是:「三口径一致」不是默认状态,是需要被证明的结论。

    但也有反方,必须写出来。 耦合存在不等于结论必翻。Coyne 等在精英运动员的急慢性负荷比上做了耦合/非耦合对照,实测效应量几乎全是微不足道(g=0.04–0.21),耦合版与非耦合版相关高达 0.88–0.99,结论是耦合对负荷监控的实际影响可能很小(PMID 31672929)。所以不能从「这个指标有耦合」直接推到「这个指标不可用」。翻转率只能实测,不能演绎。

    怎么测。 取 ≥90 人的停练队列(年限分三层各 ≥30),停练前 ≥3 次重复基线取均值以压低测量误差,停练 6/12/24 周各测一次。同一批数据跑三遍完全相同的年限×时点混合模型,只换因变量:① 绝对变化量;② 保留率(末值/峰值);③ 可获得空间占比(变化量 ÷(理论上限 − 基线))。再用协方差分析替代变化分数重跑一遍,作为第四路对照。报告时必须给全三件事:每个口径的效应量与置信区间、翻转的条目数与翻转方向、以及三口径对年限的排序是否一致。

    四格判据。 至少两口径的交互项方向或显著性相反 ⇒ 老本效应在现有文献里不可判定,既往按百分比得出的结论须全部重算;只有保留率口径显著、另两口径不显著 ⇒ 判为算术伪迹,该口径不得单独作终点;三口径一致显著且绝对量口径的效应量不低于保留率口径的一半 ⇒ 老本成立,可进入机制追问;三口径排序完全一致但绝对量效应量不足保留率口径的三分之一 ⇒ 结论成立而效应被系统性夸大,所有已发表的效应量须打折引用。

    证据与局限

    已验证: 比率变量进回归会因数学耦合产生伪结果,重分析可推翻原结论(PMID 14749086,实测重分析);相对口径与绝对口径在同一批骨密度数据上给出显著与不显著两种结论,分类不一致率 13.6% vs 1.7%(PMID 16537116,1317+1259 对实测);比例恢复规则的经验支持被数学耦合与天花板压缩混淆(PMID 33827246,方法学论证);差分分数与协方差分析在同一数据上结论相反且两种修正法不能收敛(PMID 33242670,4753 人实测)。

    部分验证: 最小临床重要改善的基线依赖源自地板天花板而非期望差异(PMID 24556220,单一量表系列);归一到可获得空间占比可消除天花板效应(PMID 25573600,单中心样本);测量误差下调整与不调整对准不同目标参数(PMID 37043152,解析推导为主);耦合的实际影响可能很小(PMID 31672929,两个运动项目案例,与主流批评相反)。

    待验证: 停练场景下三口径的翻转率本身,无任何已发表实测;呼吸侧参数(%AB、φ、X5、CoV)是否存在与骨密度同量级的口径敏感性,未知;可获得空间占比在呼吸参数上的「理论上限」如何界定,目前纯属约定;四格判据的阈值(一半、三分之一)为本文设定,未经数据校准。本文证据等级定为 C,非自愿下调:该主题未进入 L1 对应网络,且核心的「翻转率」数值在文献中完全空缺,全部论证靠邻近领域外推。

    本主题尚未进入 L1 对应网络,缺口已登记至 08-ops/science-yixue-l1-gaps.md。

    相关阅读

    证据等级: C
    参考文献: PMID 14749086, PMID 16537116, PMID 33827246, PMID 33242670, PMID 37526447, PMID 37043152, PMID 24556220, PMID 25573600, PMID 31672929

    L1 网络状态:「该主题尚未进入 L1 对应网络」——CORRESPONDENCE_NETWORK.md 节点表三个域(易学/中医侧、物理侧、生理/体验侧)全部只登记「观测什么」,没有任何节点对应「指标口径/变化量/保留率/可获得空间占比」这类测量学维度;生理侧最接近的是「呼吸率 / 呼吸模式」(备注「与节律耦合最直接」)与「气感报告」(结构化自评),均只到观测层。边表 3.1–3.5 止于 E-050(六十甲子纳音↔频率,现代证据栏原文「尚无任何可测来源」),无任何以停练、保留或口径为源或目标的边。EDGE_TABLE_science-yixue.md 仅有概念计数(丹田 51 篇 / 气功 48 篇 / 呼吸 189 篇),无层级与状态判定。故本文封顶 C。


    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。

    Will \”How Many\” and \”What Percentage\” Yield Opposite Conclusions: Flip Rates of the Same Batch of Training Data Under Three Metrics

    Evidence level: C

    The same set of detraining data, swapped for a different algorithm, flips the conclusion—how often does this actually happen? No one has ever measured it.

    Main Text

    In the previous round, we pointed out that the retention rate ratio approximates a test of baseline differences. Under equal absolute declines, the group with the higher starting point will naturally have a higher retention rate. However, this only indicates that bias might exist; it does not quantify the magnitude of the bias. A difference of 0.03 is a notational issue, but a difference large enough to reverse the direction of half of the between-group comparisons requires recalculating all prior literature. This round aims to quantify that gap.

    First Layer: Mathematical Coupling. Retention rate = Final Value ÷ Peak Value. The peak is both part of the numerator and the denominator. This issue of “ratio variables entering regression” has been discussed in the statistical community for over a century, yet the medical community has largely ignored it. Tu et al. re-analyzed two periodontology papers and found that building correlations and regressions on ratio variables severely violates methodological assumptions; the conclusions of both papers were substantively misleading, and their recommendations cannot stand (PMID 14749086). In detraining studies, the retention rate is frequently directly treated as a dependent variable in between-group comparisons, mirroring the same structure identified in the re-analyzed papers.

    Second Layer: Ceiling Compression, which stacks with coupling. The post-stroke “proportion recovery rule” posits that most individuals recover a fixed proportion (approx. 70%) of functional loss, supported empirically by the strong negative correlation between initial scores and subsequent change amounts. Bowman et al. specifically argued that this correlation is confounded by both mathematical coupling and ceiling compression; such confounding exists in both individual-level inference and group-level inference and does not disappear merely by aggregating samples (PMID 33827246). In detraining data, the “peak” is indeed the ceiling itself—once trained to the limit, further training yields no increase, and the natural decline upon detraining is inherently capped by this peak. Additional consequences of the ceiling include: minimal clinically important improvement scales with baseline elevation not due to differing patient expectations, but due to the floor and ceiling (PMID 24556220); raw gains on the Functional Independence Measure are dominated by the ceiling, and shifting to a normalization into 0–1 space resolves this bias only when expressed as available spatial proportion (PMID 25573600).

    Third Layer: The magnitude of reversal on real data already has empirical anchors. Bone density monitoring offers the cleanest case. Guidelines recommend using absolute values (g/cm²) to judge follow-up changes, yet clinical practice predominantly uses percentages. Leslie compared two metrics using the same database: 1317 spine scans and 1259 hip scans. Using a combined relative minimum significant change to judge “excessive change,” 33.5% of the low bone density group and 20.0% of the high group exceeded thresholds—a highly significant difference (P<0.0001). Switching to an absolute metric yielded 27.8% vs. 23.4%, with the difference vanishing (P=0.13). More critically, the inconsistency rate in classification: under the relative metric, 13.6% of individuals were reclassified into another category, whereas under the absolute metric, only 1.7% (PMID 16537116). In other words, changing the metric flips significance, reclassifying approximately one-seventh of the population. This provides the ready magnitude for the “reversal rate” we aim to measure in detraining data.

    Fourth Layer: Change scores and ANOVA inherently clash. Lin and Larzelere replicated Lord’s Paradox in real data from 4,753 US adolescents: differential scores and ANOVA yielded opposite results. They also performed a more significant task—testing two “harmonization” methods (double-centered covariance analysis and pre-test matching). Both methods reproduced distinct conclusions, with the gap between them as large as the original contradiction, with neither approaching the truth (PMID 33242670). Cappelleri and Cislo clarified the mechanism: once baseline adjustment is applied, the resulting clinical significance threshold reflects a scenario where “all anchor conversion layers are forced to have identical baselines”; the unadjusted version preserves external validity (PMID 37526447). Sjölander et al. further noted that when scores carry measurement error, adjustment and non-adjustment target different parameters, with no one being naturally correct (PMID 37043152). Taken together, these three papers mean: “Three-port alignment” is not the default state but a conclusion requiring proof.

    However, counterarguments must be presented. The existence of coupling does not imply conclusions must flip. Coyne et al. conducted coupled vs. non-coupled controls on acute and chronic load ratios in elite athletes; observed effect sizes were almost negligible (g=0.04–0.21), with correlations between coupled and non-coupled versions reaching 0.88–0.99. The conclusion is that the practical impact of coupling on load monitoring may be minimal (PMID 31672929). Therefore, one cannot deduce “this metric is unusable” solely from “this metric exhibits coupling.” Reversal rates can only be measured empirically, not deduced.

    How to measure. Select a detraining cohort of ≥90 individuals (split into three tenure layers, each ≥30). Take the mean of ≥3 repeated baselines prior to detraining to suppress measurement error. Measure once at 6/12/24 weeks post-detraining. Run three identical tenure × timepoint mixed models on the same dataset, swapping only the dependent variable: ① Absolute change amount; ② Retention rate (Final/Peak); ③ Available spatial proportion (Change ÷ (Theoretical Max − Baseline)). Then re-run using ANOVA instead of change scores as a fourth control. Reporting must include all three items: effect sizes and confidence intervals for each metric, the count and direction of reversals, and whether the three metrics align in their ranking of tenure.

    Four-Grid Criterion. If interaction terms show opposite directions or significance in at least two metrics ⇒ the “old stock” effect remains indeterminate in current literature; conclusions derived previously via percentages must be recalculated entirely. If only the retention rate metric is significant while the other two are not ⇒ classify as an arithmetic artifact; this metric should not be used alone as an endpoint. If all three metrics show consistent significance and the absolute amount effect size is at least half of the retention rate effect size ⇒ old stock holds; proceed to mechanism inquiry. If all three metrics rank consistently but the absolute amount effect size is less than one-third of the retention rate effect size ⇒ the conclusion holds but the effect is systematically exaggerated; all published effect sizes must be cited with discounting.

    Evidence and Limitations

    Verified: Ratio variables entering regression produce pseudo-results due to mathematical coupling; re-analysis can overturn original conclusions (PMID 14749086, empirical re-analysis); relative and absolute metrics yield significant vs. non-significant conclusions on the same bone density dataset with a classification inconsistency rate of 13.6% vs. 1.7% (PMID 16537116, empirical data from 1317+1259 pairs); empirical support for the proportion recovery rule is confounded by mathematical coupling and ceiling compression (PMID 33827246, methodological argument); differential scores and ANOVA yield opposite conclusions on the same data, and two correction methods fail to converge (PMID 33242670, empirical data from 4753 individuals).

    Partially Verified: Baseline dependency of minimal clinically important improvement stems from floor/ceiling effects rather than expectation differences (PMID 24556220, single scale series); normalization to available spatial proportion eliminates ceiling effects (PMID 25573600, single-center sample); under measurement error, adjustment and non-adjustment target different parameters (PMID 37043152, primarily analytical derivation); the actual impact of coupling may be small (PMID 31672929, two sports cases, contrary to mainstream criticism).

    To Be Verified: The reversal rate across three metrics in the detraining scenario itself has no published empirical measurement; it remains unknown whether respiratory parameters (%AB, φ, X5, CoV) exhibit caliber sensitivity comparable to bone density; the definition of “theoretical maximum” for available spatial proportion in respiratory parameters is currently purely conventional; the thresholds in the four-grid criterion (half, one-third) are set by this paper and lack data calibration. This paper assigns an evidence level of C, not voluntarily downgraded: this topic has not entered the L1 corresponding network, and the core “reversal rate” value is completely absent from literature; all arguments rely on extrapolation from adjacent fields.

    This topic has not yet entered the L1 corresponding network; the gap is registered in 08-ops/science-yixue-l1-gaps.md.

    Related Reading

    Evidence Level: C
    References: PMID 14749086, PMID 16537116, PMID 33827246, PMID 33242670, PMID 37526447, PMID 37043152, PMID 24556220, PMID 25573600, PMID 31672929

    L1 Network Status: “This topic has not yet entered the L1 corresponding network”—the CORRESPONDENCE_NETWORK.md node table registers only “what is observed” across three domains (Yijing/TCM side, Physics side, Physiological/Experiential side), with no nodes corresponding to measurement dimensions such as “metric caliber,” “change amount,” “retention rate,” or “available spatial proportion.” The physiological side is closest with “respiratory rate / respiratory pattern” (noted as “most directly coupled to rhythm”) and “qi sensation reports” (structured self-assessment), both reaching only the observation layer. Edge tables 3.1–3.5 end at E-050 (Six Heavenly Stems and Earthly Branches Na Yin ↔ Frequency; modern evidence column states “no measurable sources yet”), with no edges sourced or targeted by detraining, retention, or caliber. EDGE_TABLE_science-yixue.md contains only concept counts (Dantian 51 papers / Qigong 48 papers / Breathing 189 papers), lacking hierarchy and status determination. Thus, this paper caps at Level C.


    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.

  • 按人切还是按次切:同一个「分半系数」,可能根本在量两样东西

    Evidence level: C

    分半稳不稳定,先要问清「一半」是什么的一半;单位一换,系数就不在同一个坐标系里了。

    一、问题是怎么欠下来的

    上一轮只确认了分半口径会决定 Jaccard 数值,没说变多少。而「差 0.03」与「差 0.20」后果完全不同:前者只是记法不统一,后者意味着过去按不同口径报的稳定性数字放在一起比就是错的。丹田数据每人多次重复测量,两种切法都合法、都不缺数据——所以这个差必须量出来。

    二、分半系数从来不是「同一个量的两种算法」

    Green 在 2003 年拆掉了那个默认:时间间隔会引入瞬态误差(心情、状态、昼夜漂移),一旦存在,alpha 就是信度的高估;他为此推出 test-retest alpha,并指出「test-retest 分半系数」是所有可能的跨时间分半系数的均值(PMID 12741675)。跨时间分半得到的是被重新定义的量,把瞬态误差算进了分母——不是原系数的另一种算法。按次切与按人切,从一开始就不在估同一个东西。

    三、按次切会把「同一个人」拆到两边

    同一个人的第一次进左半、第二次进右半,这两次本就高度相似,而这个相似来自他这个人、不来自筛选规则。这是伪重复:同一个体被当成两个独立观测单位。Becker 指出,测量误差一旦相关,信度估计被系统性抬高;他给的对策正是把等价半构造在两批不同测量场合上,但警告构造不当时失真方向可以反转、幅度在单维性不足时相当大(PMID 11783569)。所以按次切的偏高不是「略微」,偏差方向可以两头跑。

    四、概化理论把这件事讲清楚了

    概化理论把总方差拆成人、场合、条目及其交互,拆开之后信度就不是一个数。2026 年一项用该框架评估贝克抑郁与绝望量表的研究给出量化:个体差异变化的信度(Rc)与人内变异的信度(Rcn)是两个独立系数,方差分解显示前者占 15%、后者占 17%–19%(PMID 41801754)。这提供了一条读法——按人切大致问「换一批人,规则还成立吗」,按次切大致问「换一次测量,还复现吗」。两个都是真问题,只是不同的问题。

    五、差值会跟着人内相关放大,还有第三层

    若虚高机制确是人内相关被当成规则证据,则两套切法的差值应随 ICC_person 单调放大,差值本身就成了诊断量。还有一层容易漏的:密集纵向数据是「每人每天多次」的三层结构(次嵌套于天、天嵌套于人),常规做法只当两层。模拟显示层次数设错会造出虚假结论——方差被膨胀、自回归估计被扭曲,且存在自相关时「天层方差是否为零」的检验本身不可靠(PMID 27378986)。丹田数据若一天测多次,按次切恰好就是把这个三层压成两层。

    六、但按次切不能一概判错

    若研究问题本就是「单次测量之间能否复现」,按次切是对的,删掉它反而删掉要测的东西。真正的问题不是哪种切法正确,而是报数字时不写清单位。可操作顺序:先写清要回答「换人」还是「换次」,再选单位;报数时单位与切刀数一起写;跨研究比较只在同单位同刀数之间进行。另有一个不可比来源:即便同为按人切,ICC 也有十种形式,必须报型号(PMID 27330520)。

    七、换问法:四个动作,四条判据

    动作:同一批数据并行跑两套切法,各报 Jaccard、超额稳定性与入选频率;同时估 ICC_person;做差值对 ICC_person 的回归;用分位数下界而非最优切点报下界(最优分半在小样本下显著高估信度,PMID 24306557)。另有脑电研究直接示范了「换个单位当 item 会低估信度」(PMID 19055503)。

    判据:按次切高出按人切 >0.15 ⇒ 判为虚高,既往按次切结论一律不可采信;差 <0.05 ⇒ 单位不影响结论,统一按人切以简化;差值随 ICC_person 单调放大 ⇒ 该差可作人内相关代理读法,须逐数据集校正后再比;按次切稳定性更高而两者选对率相同 ⇒ 指标被口径污染,只能同口径内比。

    证据与局限

    已验证: 跨时间分半引入瞬态误差、得到的是被重新定义的量(PMID 12741675);测量误差相关抬高信度、等价半构造不当可使失真反向(PMID 11783569);概化理论下「个体差异变化」与「人内变异」是两个独立系数(PMID 41801754);密集纵向数据层次数设错造成方差膨胀与估计误导(PMID 27378986);最优分半小样本下显著高估信度(PMID 24306557);ICC 十种形式不报型号则数值不可比(PMID 27330520);把分段当 item 会低估信度(PMID 19055503)。

    部分验证: 上述结论均出自心理测量、脑电与临床量表场景。把单位从条目换成受试者或测量次数,机制上同构,但没有文献直接在「参数入选集合的 Jaccard」上验证过差量级——属类比迁移。

    待验证: ① 丹田参数上两套切法的实际差值从未测过,0.05 与 0.15 两条线是从信度文献平移来的、未对真值标定;② 差值随 ICC_person 单调放大只是推演,无实测曲线;③ 切四份、切八份时差值如何变化未知;④ 按次切虚高部分能否用 ICC_person 精确扣回,取决于人内相关是否稳定;⑤ 本篇证据薄弱,主论点属可测假设,不宜当已确认机制引用。

    相关阅读

    证据等级: C
    参考文献: PMID 12741675, PMID 11783569, PMID 41801754, PMID 27378986, PMID 24306557, PMID 27330520, PMID 19055503
    L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 节点表按易学/中医侧、物理侧、生理/体验侧三类域分派(原文「| 节点 | 各域的概念/量(五行、五脏、频率、颜色…) | 按域分派(见下表) |」),三类域均无「分半单位 / 人内相关 / 方差来源 / 信度系数」这类测量学节点;边表止于 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」),无任何以测量单位为源或目标的边。EDGE_TABLE_science-yixue.md 为概念反抽统计,仅列概念计数与最高证据级,无 L1/L2/L3 层级与 active/open/conflict 状态判定,故本条边状态为缺。


    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。

    Cut by person or cut by frequency: the same \”split coefficient\” might actually be measuring two different things

    Evidence level: C

    Whether the split-half reliability is stable depends first on defining what constitutes “half.” Changing the unit shifts the coefficient out of the same coordinate system.

    I. How the Problem Accumulates

    The previous round confirmed that split-half methodology determines the Jaccard value but did not specify the magnitude of change. A difference of “0.03” versus “0.20” yields entirely different consequences: the former suggests inconsistent notation, while the latter implies that comparing stability figures reported under different methodologies is fundamentally erroneous. Dantian data allow for repeated measurements per individual; both splitting methods are valid and do not compromise data availability. Therefore, this discrepancy must be quantified.

    II. The Split-Half Coefficient Is Never “Two Algorithms for the Same Quantity”

    Green dismantled this default in 2003 by noting that time intervals introduce transient errors (mood, state fluctuation, circadian drift). If such errors exist, alpha constitutes an overestimation of reliability. He proposed the test-retest alpha and indicated that the “split-half coefficient across time” represents the mean of all possible cross-temporal split-half coefficients (PMID 12741675). The cross-temporal split yields a redefined quantity where transient errors are incorporated into the denominator—it is not merely an alternative algorithm for the original coefficient. Split-by-item and split-by-individual have never estimated the same underlying construct from the outset.

    III. Split-by-Item Fragments “The Same Individual” Across Splits

    In split-by-item, an individual’s first measurement enters the left half and the second enters the right half. These two measurements are inherently highly similar due to the individual themselves, not the selection rules. This constitutes pseudo-replication: treating the same subject as two independent observational units. Becker noted that when measurement errors are correlated, reliability estimates are systematically inflated; his remedy involves constructing equivalent halves across two distinct measurement occasions, yet he warned that improper construction can reverse the direction of distortion and amplify magnitude when unidimensionality is insufficient (PMID 11783569). Thus, the inflation in split-by-item is not merely “slight”; the bias direction can swing either way.

    IV. Generalizability Theory Clarifies the Matter

    Generalizability theory decomposes total variance into persons, occasions, items, and their interactions. Once decomposed, reliability is not a single number. A 2026 study applying this framework to the Beck Depression and Hopelessness Scales provided quantification: reliability for individual-difference change ($R_c$) and reliability for within-person variance ($R_{cn}$) are two independent coefficients. Variance decomposition showed the former accounts for 15% and the latter for 17%–19% (PMID 41801754). This offers a reading perspective: split-by-item roughly asks, “Does the rule hold with a new batch of people?” while split-by-item roughly asks, “Is it reproducible with a new measurement occasion?” Both are valid questions, but they address different constructs.

    V. The Difference Amplifies with Within-Subject Correlation and Adds a Third Layer

    If the inflation mechanism relies on within-subject correlation being treated as rule-based evidence, the difference between the two splitting methods should monotonically increase with $ICC_{person}$, making the difference itself a diagnostic metric. There is another layer often overlooked: dense longitudinal data possess a three-level structure (items nested within days, days nested within persons), yet standard practice often treats it as two levels. Simulation shows that mis-specifying the number of levels creates false conclusions—variance is inflated, autoregressive estimates are distorted, and when autocorrelation exists, the test for whether “day-level variance is zero” itself becomes unreliable (PMID 27378986). If Dantian data are collected multiple times daily, split-by-item effectively compresses this three-level structure into two.

    VI. But Split-by-Item Is Not Universally Wrong

    If the research question is inherently “can replication occur between single measurements,” split-by-item is correct; removing it deletes the very thing to be measured. The core issue is not which splitting method is correct, but whether units are explicitly stated when reporting numbers. A practical sequence: first clarify whether one is answering “changing people” or “changing occasions,” then select the unit; report numbers alongside the unit and the number of cuts; cross-study comparisons should only occur within identical units and cut counts. Another source of non-comparability exists: even with split-by-individual, there are ten forms of ICC, and the model type must be reported (PMID 27330520).

    VII. Rephrasing the Question: Four Actions, Four Criteria

    Actions: Run both splitting methods on the same dataset in parallel, reporting Jaccard values, excess stability, and inclusion frequencies; simultaneously estimate $ICC_{person}$; perform regression of the difference against $ICC_{person}$; report lower bounds using percentile cutoffs rather than optimal split points (optimal splitting significantly overestimates reliability in small samples, PMID 24306557). Additionally, EEG studies directly demonstrate that “changing the unit to an item can underestimate reliability” (PMID 19055503).

    Criteria: If split-by-item exceeds split-by-individual by >0.15 ⇒ classify as inflated; all prior conclusions based on split-by-item are unacceptably invalid. If difference <0.05 ⇒ units do not affect conclusions; standardize to split-by-individual for simplicity. If the difference monotonically amplifies with $ICC_{person}$ ⇒ treat this difference as a proxy read for within-subject correlation, requiring calibration per dataset before comparison. If split-by-item shows higher stability while selection accuracy remains identical ⇒ the metric is contaminated by definition; comparison should only occur within the same definition.

    Evidence and Limitations

    Verified: Cross-temporal splitting introduces transient errors yielding a redefined quantity (PMID 12741675); correlated measurement errors inflate reliability, and improper equivalent half construction can reverse distortion direction (PMID 11783569); under Generalizability Theory, “individual-difference change” and “within-person variance” are independent coefficients (PMID 41801754); mis-specifying levels in dense longitudinal data causes variance inflation and estimation bias (PMID 27378986); optimal splitting significantly overestimates reliability in small samples (PMID 24306557); failure to report ICC model type renders values incomparable (PMID 27330520); treating segments as items underestimates reliability (PMID 19055503).

    Partially Verified: The above conclusions derive from psychometrics, EEG, and clinical scale contexts. Mechanistically isomorphic when switching units from items to subjects or measurement counts, but no literature directly validates the magnitude of differences on “Jaccard values within parameter selection sets”—this relies on analogical transfer.

    To Be Verified: ① The actual difference between the two splitting methods on Dantian parameters has never been measured; the 0.05 and 0.15 thresholds are translated from reliability literature without true value calibration; ② The monotonic amplification of difference with $ICC_{person}$ is a deduction without empirical curves; ③ How differences change when splitting into four or eight parts remains unknown; ④ Whether the inflated portion in split-by-item can be precisely recovered via $ICC_{person}$ depends on the stability of within-subject correlation; ⑤ Evidence in this paper is weak; main arguments constitute testable hypotheses and should not be cited as confirmed mechanisms.

    Related Reading

    Evidence Level: C
    References: PMID 12741675, PMID 11783569, PMID 41801754, PMID 27378986, PMID 24306557, PMID 27330520, PMID 19055503
    L1 Network Status: This topic has not yet entered the L1 corresponding network—the CORRESPONDENCE_NETWORK.md node table assigns nodes to Yijing/TCM, Physics, and Physiological/Experiential domains (original text: “| Node | Concepts/Quantities per Domain (Five Phases, Five Organs, Frequencies, Colors…) | Domain Assignment (see table below) |”), but none of the three domains contain measurement nodes such as “split-half units,” “within-subject correlation,” “variance sources,” or “reliability coefficients.” The edge table terminates at E-050 (“Six Heavenly Stems and Earthly Branchs Na-Yin | Corresponds | 60 Frequencies,” Level L1, Status open; modern evidence column originally states “No measurable sources at all“). There are no edges with measurement units as source or target. EDGE_TABLE_science-yixue.md is a conceptual back-extraction statistic listing only concept counts and highest evidence levels, lacking L1/L2/L3 hierarchies and active/open/conflict status determinations; therefore, the status of this edge is missing.


    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.

  • 三条判据线到底会不会打架:别用 kappa 去判,它自己就带着同一个病

    Evidence level: C

    要判断三条判据是否一致,最直觉的做法是算一个 Cohen’s kappa;但 kappa 恰恰在分布最不均衡的时候最不可信——而这正是本项目的常态。

    一、这个问题是怎么来的

    上一轮迭代里,范围损失的判据从「合成加权指数」改成了「逐变量诊断」:标准化均值差小于 0.1 则保留合并;标准化均值差超过 0.25 且交互检验功效高于 80% 则分层报告;交互检验功效低于 30% 则报未知。新判据干净,但从未在同一份数据上与旧的合成指数做过交叉验证。

    于是问题变成:同一批参数放到三条判据线下,会不会给出相同的裁定?直觉上的答案是算一个 Cohen’s kappa。但这条路有个坑,而且在本项目的语境下坑特别深。

    二、kappa 会把你想要的答案读反

    kappa 的定义是:(观察一致率 − 期望一致率)除以(1 − 期望一致率)。分母表示「还剩多少分歧可以被解释」,分子表示「实际一致比瞎猜好多少」。关键因子是「期望一致率」。

    当边际分布高度不均衡——绝大多数参数被划进同一档——期望一致率趋近 1,分母趋近 0,kappa 塌向 0,哪怕观察一致率高达 0.9。这就是「高一致、低 kappa」的悖论(PMID 2348207)。同一篇论文还指出第二个悖论:在边际不对称的情形下,kappa 可能比在对称失衡的情形下人为更高,看上去「更好」,而失衡其实更严重。

    放到本项目上:三条判据的期望裁定分布恰恰是高度不均衡的——绝大多数参数应当落入「保留合并」,而这正是悖论最严重的区域。直接算 kappa 只会得出「三条判据严重不一致」,而这可能只是「裁定分布太偏」的另一种说法。

    三、三分类又多加了一层复杂性

    裁定不是二值而是三值:保留、分层、未知。在三个类别下,期望一致率由三个边际比例共同决定,比二值情形更敏感;此外,算加权 kappa 还要选权重(线性或二次),多出一个会改变数值的自由度。方法学比较给出的结论很明确:在 2×2 表的多种边际情形下,kappa 与 alpha 受边际分布影响最大,而 AC1 与 B 统计量更稳健,其中 B 统计量在两种悖论下表现最好(PMID 25168681);偏倚校正与流行率校正的 kappa(PABAK)直接校正了这两重偏倚(PMID 8501467)。这不只是理论:在高特征流行率的病理诊断中,kappa 会显著偏离观察一致率给人的印象(PMID 38283651)。

    四、一致不等于正确

    三条判据共享一个未经检验的前提:真值空间是被构造出来的。合成指数里的权重、逐变量诊断里的 0.1 与 0.25 阈值、比例度量背后的真值分布,全都是建模者填进去的。它们彼此一致,只说明它们在同一套假设下对齐,并不能证明其中任何一条正确。

    所以 kappa 高不是好信号:它意味着三条线没有提供独立信息。真正该担心的是三条一起同意却一起错——这种偏差,一致率永远测不出来。

    五、第三条判据根本不是同一个量

    前两条判据给出的是逐参数的类别裁定,第三条「结论成立的比例」却是连续量:真值空间中结论仍然成立的比例。要把这个连续量塞进三档裁定,就必须选一个阈值。阈值是自由参数:取 20% 还是 30%,一致率就变了。

    因此「三幅画像的 kappa」其实是两个类别量与一个被任意切了阈的连续量之间的一致性。挪一下阈值,结论就变。这与上一轮批评「把判别力改名叫可推广性」是同一个病,只是入口换了一个。

    六、一致性是对称的,后果不是

    判据 A 判「降级」而 B 判「保留」,与 A 判「保留」而 B 判「降级」,在一致率上完全相同。但后果差别极大:前者多花样本成本,后者导致结论过度外推。应该报告的是不一致的条数与方向,而不是幅度。

    kappa 本身也有相当可观的抽样波动。一个旁证:连「算 kappa 需要多少样本」这个问题,文献里也是用观察一致率的比例而非 kappa 值本身来处理的(PMID 25088752)。拿一把自己会晃的尺子去判断另外两把尺子是否一致,得到的精度不会超过那把尺子自己的不稳定性。

    七、把问题换掉

    四个动作可以替代「算一个 kappa」:

    1. 逐参数分别报告三条原始裁定,不做合并、不做分数求和——这与平衡诊断的提议一致:逐变量报告标准化均值差与方差比,而不是压成单一指标(PMID 30788363)。
    2. 同时报告观察一致率、PABAK 与 B 统计量;任何只依赖 kappa 得出的结果一律作废。
    3. 报告不一致的条数与方向,并计算「会改变设计决策的参数线条数」。这才是与决策相关的量:交互检验功效不足会使调节幅度不可估(PMID 15066682),边际估计与条件估计甚至可能给出相反的决策(PMID 41626969)。
    4. 用自助法给一致率生成置信区间。若区间宽度超过 0.3,每组 60 个样本不足以支撑一致性结论,必须扩样。

    四条判据对应四种结论:若观察一致率高于 85% 而 kappa 低于 0.4,则悖论得到确认,kappa 是误报,必须撤回「三幅画像不可互换」这一结论;若校正后 PABAK 仍低于 70%,则确实存在不一致,三幅画像必须分开报告,且禁止用范围损失数值做跨研究比较;若第三条判据的阈值在 20% 到 30% 之间移动时,导致超过 20% 的参数跨档,则它不具备判据资格,只能降级为敏感性分析;若自助法区间宽度超过 0.3,则样本量问题先于一致性问题。

    一句话:判据之间是否一致,必须用一把不受分布形状影响的尺子去量;用一把有偏的尺子量,量出来的永远是尺子自己的偏。

    证据与局限

    已验证:Cohen’s kappa 存在两种结构性悖论——在边际不均衡下,观察一致率高也会被大幅压低;在边际不对称下,kappa 可能人为更高,且这两种悖论都不能靠把 kappa 换成 kappa max 来修掉(PMID 2348207);对 2×2 表中各种一致性系数的系统比较显示,kappa 与 alpha 受边际分布影响最大,AC1 与 B 统计量更稳健,其中 B 统计量在两种悖论下表现最好(PMID 25168681);PABAK 直接校正了这两重偏倚(PMID 8501467);kappa 的样本量规划实践已转向使用观察一致率的比例而非 kappa 值本身(PMID 25088752)。

    部分验证:高流行率特征下 kappa 显著偏离观察一致率的现象,已在病理诊断实证数据中被量化(PMID 38283651),但其对象是肺组织学特征、读者是病理医师,与丹田参数的裁定场景不同;平衡诊断应逐变量报告标准化均值差与方差比而非压成单一指标(PMID 30788363),这一提议源自倾向得分匹配领域,移植到本项目构成类比;交互检验功效不足会使调节幅度不可估(PMID 15066682),效应修饰与不可压缩性并存可导致边际与条件估计给出相反决策(PMID 41626969),两者都支持「方向比幅度更重要」,但均基于临床试验与综述语境。

    待验证:三条判据在同一份丹田数据上的实际裁定分布从未被测量过——「多数参数落入保留合并」只是预期;若实测分布其实是均衡的,悖论风险显著下降,本文的主论证随之减弱;三档裁定下 kappa 的具体偏离幅度没有模拟数据;PABAK 与 B 统计量在本项目数据上的数值差异没有实测;第三条判据的阈值敏感性曲线(在 20%、25%、30% 三档上跨档的参数比例)从未算过;「一致不等于正确」这一层完全依赖论证,任何一致率指标都测不了它。

    相关阅读

    证据等级: C
    参考文献: PMID 2348207, PMID 25168681, PMID 8501467, PMID 25088752, PMID 38283651, PMID 30788363, PMID 15066682, PMID 41626969
    L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 节点表按易学/中医侧、物理侧、生理/体验侧三类域分派节点(依概念与量,如五行、五脏、频率、颜色等),三类域中均无「判据一致性 / 裁定类别 / 一致性系数 / 边际分布偏倚」这类测量学与判读层的节点;边表止于 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」),不存在以判据、裁定或一致率为源或目标的边。EDGE_TABLE_science-yixue.md 为反抽统计,仅列概念计数(丹田 51 篇 / 气功 48 篇 / 呼吸 189 篇 / 频率 407 篇)与最高证据级,无 L1/L2/L3 层级与 active/open/conflict 状态判定,故本条边状态为缺。


    免责说明:本文属方法学与统计口径的讨论,仅供理解与参考,不构成医疗建议,不能替代执业医师的判断与诊疗。文中出现的阈值、百分比与样本量建议均为判读一致性这一方法学层面的推演,不适用于任何个体情形;如有健康疑虑,请及时就医并由专业医师判断。

    Will the Three Criteria Lines Clash? Don’t Use Kappa to Judge It, as It Carries the Same Flaw

    Evidence level: C

    To determine if three criteria are consistent, the most intuitive approach is to calculate Cohen’s kappa; however, kappa is least trustworthy precisely when distribution imbalances are highest—a condition central to this project.

    I. How This Question Arises

    In the previous iteration, the criterion for range loss shifted from a “synthesized weighted index” to a “per-variable diagnosis”: retain merging if the standardized mean difference is less than 0.1; report stratification if the standardized mean difference exceeds 0.25 and interaction power surpasses 80%; report unknown if interaction power is below 30%. The new criterion is clean, yet it has never been cross-validated against the old synthesized index on the same dataset.

    Thus, the question becomes: If the same parameter is subjected to three criteria lines, will they yield identical verdicts? The intuitive answer is to calculate Cohen’s kappa. However, this approach contains a pitfall, and it is particularly deep in the context of this project.

    II. Kappa Reads Your Desired Answer Backwards

    Kappa is defined as: (Observed Agreement – Expected Agreement) divided by (1 – Expected Agreement). The denominator represents “how much disagreement remains explainable,” while the numerator represents “how much better actual agreement is than guessing.” The critical factor is “Expected Agreement.”

    When marginal distributions are highly imbalanced—where the vast majority of parameters are categorized into a single tier—the expected agreement approaches 1, the denominator approaches zero, and kappa collapses toward 0, even if observed agreement reaches 0.9. This is the paradox of high agreement but low kappa (PMID 2348207). The same paper identifies a second paradox: kappa can be artificially higher under marginal asymmetry compared to symmetric imbalance, appearing “better” despite the imbalance.

    Applied to this project: The expected verdict distribution across the three criteria is precisely highly imbalanced—most parameters should fall into “retain merging,” which is exactly the region where the paradox is most severe. Calculating kappa directly will yield “severe inconsistency among three criteria,” which may merely be another phrasing for “verdict distribution is too skewed.”

    III. Three Categories Add Another Layer of Complexity

    Verdicts are not binary but ternary: retain, stratify, or unknown. Under three categories, expected agreement is jointly determined by three marginal proportions, making it more sensitive than in binary cases; furthermore, calculating weighted kappa requires selecting weights (linear or quadratic), introducing another degree of freedom that alters the numerical value. Methodological comparisons conclude clearly: across various marginal scenarios in 2×2 tables, kappa and alpha are most affected by marginal distributions, while AC1 and B statistics are more robust, with the B statistic performing best under both paradoxes (PMID 25168681); the bias-adjusted and prevalence-adjusted kappa (PABAK) directly corrects for these two biases (PMID 8501467). This is not merely theoretical: in pathological diagnoses with high feature prevalence, kappa deviates significantly from the impression given by observed agreement (PMID 38283651).

    IV. Agreement Does Not Equal Correctness

    The three criteria share an unverified premise: the true value space is constructed. The weights in the synthesized index, the 0.1 and 0.25 thresholds in the per-variable diagnosis, and the true value distribution underlying the proportion metrics are all filled by modelers. Agreement among them only indicates they align under the same set of assumptions; it does not validate any one of them as correct.

    Therefore, high kappa is not a positive sign: it implies the three lines provide no independent information. The real concern is when all three agree but are all wrong—a discrepancy that agreement rates can never detect.

    V. The Third Criterion Is Not the Same Quantity

    The first two criteria yield categorical verdicts per parameter, while the third—”proportion of conclusions holding”—is a continuous quantity: the proportion within the true value space where the conclusion remains valid. To force this continuous quantity into a three-tier verdict, a threshold must be selected. The threshold is a free parameter; setting it at 20% versus 30% alters the agreement rate.

    Thus, “three-portrait kappa” is actually consistency between two categorical quantities and a continuous quantity thresholded arbitrarily. Shifting the threshold shifts the conclusion. This is the same disease as the previous criticism of renaming “discriminative ability” to “generalizability,” merely with a different entry point.

    VI. Agreement Is Symmetric, But Consequences Are Not

    Criterion A judges “downgrade” while B judges “retain,” versus A judging “retain” while B judges “downgrade”—these are identical in terms of agreement rates. Yet the consequences differ vastly: the former incurs extra sample costs, while the latter leads to over-extrapolation of conclusions. What should be reported is the count and direction of inconsistencies, not their magnitude.

    Kappa itself suffers from substantial sampling variability. A corroborating example: even the question of “how many samples are needed to calculate kappa” has been addressed in literature using proportions of observed agreement rather than the kappa value itself (PMID 25088752). Using a ruler that shakes itself to judge whether two other rulers are consistent will not yield precision exceeding the ruler’s own instability.

    VII. Change the Question

    Four actions replace calculating a single kappa:

    1. Report all three original verdicts per parameter separately, without merging or summing scores—this aligns with the balanced diagnosis proposal: report standardized mean differences and variance ratios per variable without compressing them into a single metric (PMID 30788363).
    2. Report observed agreement, PABAK, and the B statistic simultaneously; discard any results based solely on kappa.
    3. Report the count and direction of inconsistencies, and calculate “the number of parameter lines that would alter design decisions.” This is the decision-relevant quantity, whereas insufficient interaction power renders modulation magnitude unestimable (PMID 15066682), and marginal/conditional estimates may even yield opposite decisions (PMID 41626969).
    4. Use bootstrapping to generate confidence intervals for agreement rates. If the interval width exceeds 0.3, a sample size of 60 per group is insufficient to support consistency conclusions; samples must be expanded.

    Four criteria correspond to four outcomes: if observed agreement exceeds 85% while kappa is below 0.4, the paradox is confirmed, kappa is a false alarm, and the conclusion that “three portraits are not interchangeable” must be retracted; if PABAK remains below 70% even after correction, there is indeed inconsistency, and the three portraits must be reported separately with range loss values prohibited for cross-study comparison; if shifting the third criterion’s threshold between 20% and 30% causes over 20% of parameters to cross tiers, it lacks criterion qualification and should only serve as a sensitivity analysis; if the bootstrap interval width exceeds 0.3, sample size issues precede consistency issues.

    In one sentence: Whether criteria agree must be measured with a ruler unaffected by distribution shape; measuring with a biased ruler always yields the ruler’s own bias.

    Evidence and Limitations

    Verified: Cohen’s kappa exhibits two structural paradoxes—high observed agreement is drastically suppressed under marginal imbalance, and kappa can be artificially higher under marginal asymmetry, neither of which is fixed by replacing kappa with kappa max (PMID 2348207); systematic comparisons of various agreement coefficients in 2×2 tables show kappa and alpha are most affected by marginal distributions, with AC1 and B statistics being more robust and the B statistic performing best under both paradoxes (PMID 25168681); PABAK provides direct correction for these two biases (PMID 8501467); sample size planning practices for kappa have already shifted to using proportions of observed agreement rather than the kappa value itself (PMID 25088752).

    Partially Verified: The phenomenon of kappa deviating significantly from observed agreement under high-prevalence features has been quantified in empirical pathological diagnosis data (PMID 38283651), but the subjects were lung tissue histological features and readers were pathologists, differing from the verdict scenario of Dantian parameters; balanced diagnosis should report per-variable standardized mean differences and variance ratios rather than compressing them into a single metric (PMID 30788363), a proposal originating from propensity score matching fields, which constitutes an analogy when transferred to this project; insufficient interaction test power renders modulation magnitude unestimable (PMID 15066682), and the coexistence of effect modification and incompressibility can cause marginal and conditional estimates to yield opposite decisions (PMID 41626969), both supporting “direction matters more than magnitude,” yet based on clinical trial and review contexts.

    To Be Verified: The actual verdict distribution of the three criteria on the same Dantian dataset has never been measured—”most parameters fall into retain merging” is merely an expectation; if the empirical distribution proves balanced, paradox risk drops significantly, weakening this paper’s main argument; no simulation data exists for the specific deviation magnitude of kappa under three-tier verdicts; no empirical data exists for the numerical difference between PABAK and B statistics in this project’s data; the sensitivity curve for the third criterion’s threshold (proportion of parameters crossing tiers at 20%, 25%, and 30%) has never been calculated; the layer “agreement does not equal correctness” relies solely on argumentation and cannot be tested by any agreement rate metric.

    Related Reading

    Evidence Level: C
    References: PMID 2348207, PMID 25168681, PMID 8501467, PMID 25088752, PMID 38283651, PMID 30788363, PMID 15066682, PMID 41626969
    L1 Network Status: This topic has not yet entered the L1 corresponding network—the CORRESPONDENCE_NETWORK.md node table assigns nodes by domain (Yijing/TCM side, Physics side, Physiological/Experiential side) based on concepts/quantities (Five Phases, Five Organs, Frequencies, Colors, etc.). Currently, none of the three domains contain measurement or judgment layer nodes related to “Criterion Consistency / Verdict Categories / Agreement Coefficients / Marginal Distribution Bias.” The edge table ends at E-050 (“Six Heavenly Stems and Earthly Branchs Na-Yin | Corresponds | 60 Frequencies,” Level L1, Status open; modern evidence column originally states “No measurable sources at all“). There are no edges originating from or targeting criteria, verdicts, or agreement rates. EDGE_TABLE_science-yixue.md is a reverse statistical extraction listing only concept counts (51 Dantian articles / 48 Qigong articles / 189 Breathing articles / 407 Frequency articles) and highest evidence levels, lacking L1/L2/L3 hierarchies or active/open/conflict status determinations; therefore, the status of this edge is marked as missing.


    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.

  • 重叠系数在偏态数据上还认不认得 AUC:那条「单调对应」只在正态下成立

    Evidence level: C

    重叠系数与 AUC 能互相换算,靠的是正态假设给的一个巧合;假设一撤,两者连方向都不一定同。

    一、这个「等价」是怎么来的

    上一轮把重叠系数(OVL)判成「AUC 的非线性重标」,听着像省了件事:算出一个,另一个就知道。但这换算有前提。OVL 是两条密度曲线逐点取小后的面积,AUC 是一组分数高于另一组的概率。两组都服从正态、只差均值与方差时,两者都是那两个参数的函数,构成一一对应。

    问题在「只差均值与方差」这句。正态分布由均值和方差完全决定,没有第三个自由度;偏态分布有。多一个自由度,一个 OVL 值就能对应一串 AUC 值——映射不再是单射,换算表从一条线变成一片区域。这不是精度问题,是有没有资格换算。

    二、丹田参数几乎不可能落在正态里

    丹田五参数的实际分布:呼吸比、相位同步率是有界比例量,天然挤在 0 和 1 之间;X5 电抗可以跨零,一跨零分布就分叉;CoV 天然右偏。四类量没有一类是「均值加减方差能描述完」的。

    文献侧态度很清楚:二正态参数法在宽范围情形下表现最好,但这恰恰说明非正态时需另作处理,而非照用(PMID 34693817);另一项比较五种 OVL 估计的研究发现,固定带宽核密度的非参数估计在几乎所有情形下均方误差更小、覆盖率更接近名义水平(PMID 16279139)。参数法能赢的场合,正是分布离正态不太远的场合。

    三、AUC 与 OVL 在变换下命运不同

    这一点不需要模拟就能确定,也是最硬的一刀。AUC 是秩统计量:把所有人的测量值做任何单调变换(开方、取对数、Box-Cox),排序不变,AUC 一丝不动。有研究直接借这个不变性做回归建模,处理非正态、有序、删失数据,估计无偏且覆盖率落在名义水平(PMID 37278185)。

    OVL 不是。它是密度量,逐点取小再积分,依赖数值的具体位置与曲线形状。同一批数据做一次 Box-Cox,AUC 不变,OVL 会变。

    这条直接判了「重标」死刑:若 OVL 真是 AUC 的单调重标,AUC 不变时 OVL 也不该变。它变了,说明它带着 AUC 里没有的信息——那就不是重标,是另一个量。

    四、相关近乎完美,精度可以天差地别

    有人可能退一步:那至少两者高度相关,能互相代表吧。这项系统模拟正好回答了这个退让。研究用偏态正态生成可控的偏度、方差比、均值差与样本量,比较 Cohen’s d、共用语效应量(CLES)、由 d 推的参数重叠系数、非参数重叠指数。结论:前三个描述性相关近乎完美,但表现差得很远——CLES 与参数重叠系数在偏度与异方差下偏差大、覆盖率低;非参数指数在形状差异下无偏,可两组真重叠时反而不可靠(PMID 42094317)。

    作者自己的落点句最值得抄:效应量指标不可互换,高经验相关不保证同等精度。套到本文:OVL 与 AUC 的相关性再高,也不能当等价的证明。相关衡量「一起动」,精度衡量「动得准」。

    五、AUC 本身在偏态下不偏,偏的是阈值

    还有一个容易搞混的分工。偏态下双正态 ROC 的 AUC 倾向无偏,仍可作全局性能的合理度量;偏的是判定阈值。有研究因此改用能吸收偏度与厚尾的偏指数幂分布建 ROC 模型,既改善 AUC 估计效率,也把切点分类率拉回来(PMID 37260468)。

    分界由此清楚:全局量在偏态下能撑住,局部量撑不住。OVL 属全局形状量,它不能替代阈值信息,也不能借 AUC 的稳健性给自己背书——两个全局量一个稳一个不稳,本身就说明它们不是同一个东西。

    六、误差棒的口径也各走各的

    即使退到「只比数值、不换算」,区间构造也完全不同源。OVL 要专门方法:参数法、非参数法、自助法各有适用面(PMID 34693817、PMID 16279139),处理协变量还得先 Box-Cox 再自助(PMID 40852736)。AUC 那侧是另一套:同一批数据上九种解析法与三种自助法给出的区间可以差很多(PMID 26734848)。

    两套区间的宽度、覆盖率、对样本量的敏感度都不一样。「用 OVL 的区间推 AUC 的区间」没有可依据的推导链。

    七、把问题换掉

    真正该做的不是验证「OVL 能不能替代 AUC」,而是三个前置动作:

    一是先报每个参数的偏度与方差比,再做 Box-Cox,记录 OVL 变了多少、AUC 变了多少。AUC 的变化量应当为零,这是仪器校准;不为零说明算法写错了。
    二是同批数据上算 OVL 与 AUC 的秩相关,变换后再算一遍。两条相关并排看,差多少就是偏态的代价。
    三是定用途:要全局判别力就报 AUC;要形状描述就报 OVL,但须声明未做变换、只在同一口径内比较。

    四条判据对应四种结局:秩相关低于 0.8,则「重叠度可替代判别力」不成立、OVL 只能作形状描述;偏离方向随方差比翻转,则关系非单调、任何反推作废;Box-Cox 后秩相关回升到 0.9 以上,则 OVL 可用但须声明已变换;实测偏离小于 5% 而模拟偏离大于 30%,则现有丹田参数分布意外接近正态、偏态档不适用——那反倒是好消息,但得先测出来才算。

    落点一句话:靠假设撑起来的换算表,先测假设,再谈换算。

    证据与局限

    已验证: AUC 作为秩统计量对单调变换不变,变换模型可借此处理非正态、有序与删失数据且估计无偏、覆盖率落在名义水平(PMID 37278185);OVL 的推断方法学已系统建立,参数法(二正态)在宽范围分布情形下优于非参数法并提供 R 实现(PMID 34693817);五种 OVL 估计的比较显示固定带宽核密度非参数法在多数情形均方误差更小、覆盖率更接近名义(PMID 16279139);协变量特异 OVL 需回归加 Box-Cox 并配自助法区间(PMID 40852736);同一批数据上不同 AUC 置信区间方法结果差异显著,偏差校正加速自助法被推荐(PMID 26734848)。

    部分验证: 偏态正态系统模拟显示 Cohen’s d、CLES、参数重叠系数三者描述性相关近乎完美但偏差与覆盖率表现差异很大,参数重叠系数在偏度与异方差下偏差大、覆盖率低(PMID 42094317)——这直接支撑「高相关不等于同等精度」,但其结局变量是心理学量表分,非丹田生理参数;非参数重叠指数在形状差异下无偏、在两组真重叠时不可靠,这一条提示非参数法也有自己的失效区;偏态下双正态 ROC 的 AUC 倾向无偏而阈值偏,偏指数幂分布模型可同时改善(PMID 37260468),说明全局量与局部量在偏态下命运不同,但该结论建立在单一临床数据集(肺炎克雷伯菌感染)上。

    待验证: 丹田五参数上 OVL–AUC 秩相关的实测值从未算过;Box-Cox 前后 OVL 的变动幅度无数据;四种偏态(对数正态、Beta 双峰、截断正态)配三种方差比的偏离方向是否随方差比翻转,属模拟预期而非观测;「实测偏离小于 5%」这一分支纯属假设——若实测偏离大于 30%,则前述所有换算类做法在本项目内一律作废,这条正是本篇最主要的可证伪点。

    相关阅读

    证据等级: C
    参考文献: PMID 37278185, PMID 34693817, PMID 16279139, PMID 40852736, PMID 26734848, PMID 42094317, PMID 37260468
    L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 节点表按易学/中医侧、物理侧、生理/体验侧三类域分派(原文「| 节点 | 各域的概念/量(五行、五脏、频率、颜色…) | 按域分派(见下表) |」),三类域均无「重叠系数 / 判别力 / 分布偏度 / 单调变换不变性 / 误差棒口径」这类测量学与外效度维度节点;边表止于 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」),无任何以分布形状或判别力为源或目标的边。EDGE_TABLE_science-yixue.md 为反抽统计,仅列概念计数(丹田 51 篇 / 气功 48 篇 / 呼吸 189 篇 / 频率 407 篇)与最高证据级,无 L1/L2/L3 层级与 active/open/conflict 状态判定,故本条边状态为缺。


    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 Overlap Coefficient Still Recognize AUC in Skewed Data: The \”Monotonic Correspondence\” Only Holds Under Normality

    Evidence level: C

    The overlap coefficient (OVL) and AUC are convertible only due to a coincidence granted by the normality assumption; without it, they may not even share the same direction.

    I. How This “Equivalence” Emerges

    In the previous round, classifying OVL as a “nonlinear re-labeling of AUC” sounded like saving effort: calculate one, and the other is known. However, this conversion has prerequisites. OVL is the area under the pointwise minimum of two density curves, whereas AUC is the probability that a score from one group exceeds that of another. When both groups follow a normal distribution differing only in mean and variance, both metrics are functions of those two parameters, forming a one-to-one correspondence.

    The issue lies in the phrase “differing only in mean and variance.” The normal distribution is fully determined by these two parameters with no third degree of freedom; skewed distributions have one. Adding a degree of freedom means a single OVL value can correspond to a series of AUC values—the mapping is no longer injective, and the conversion table transforms from a line into an area. This is not a precision issue but a question of eligibility for conversion.

    II. Dantian Parameters Almost Never Fall Within Normality

    The actual distribution of the five Dantian parameters: the breathing ratio and phase synchronization rate are bounded proportions naturally squeezed between 0 and 1; X5 reactance can cross zero, causing a bifurcation in distribution; CoV is naturally right-skewed. None of the four categories can be fully described by “mean plus or minus variance.”

    The literature stance is clear: the two-normal-parameter method performs best over a wide range, but this precisely indicates that alternative handling is required under non-normal conditions rather than blind application (PMID 34693817). Another study comparing five OVL estimation methods found that fixed-bandwidth kernel density nonparametric estimation yielded lower mean squared error and coverage closer to nominal levels in almost all scenarios (PMID 16279139). Parameter methods excel only when distributions are not far from normal.

    III. Different Fates Under Transformation for AUC and OVL

    This point requires no simulation to determine and represents the hardest cut. AUC is a rank statistic: applying any monotonic transformation (square root, logarithm, Box-Cox) to all measurement values preserves the ranking, leaving AUC unchanged. Research directly leverages this invariance for regression modeling to handle non-normal, ordinal, and censored data, yielding unbiased estimates with coverage falling within nominal levels (PMID 37278185).

    OVL is not. It is a density quantity involving pointwise minimization followed by integration, dependent on specific numerical positions and curve shapes. Applying a Box-Cox transformation to the same dataset leaves AUC unchanged but alters OVL.

    This directly sentences “re-labeling” to death: if OVL were truly a monotonic re-labeling of AUC, OVL should not change when AUC remains constant. Since it does change, it carries information absent in AUC—meaning it is not a re-labeling but a distinct quantity.

    IV. Near-Perfect Correlation Yet Potentially Divergent Precision

    One might retreat: at least they are highly correlated and represent each other. This systematic simulation answers that concession directly. Researchers used skewed-normal generation to control skewness, variance ratios, mean differences, and sample sizes, comparing Cohen’s d, Common Language Effect Size (CLES), parameter-based OVL derived from d, and nonparametric overlap indices. Conclusion: the first three show near-perfect descriptive correlation, yet performance varies drastically—CLES and parameter-based OVL exhibit large biases and low coverage under skewness and heteroscedasticity; nonparametric indices are unbiased under shape differences but unreliable when true overlap exists (PMID 42094317).

    The author’s concluding remark is worth quoting: effect size metrics are not interchangeable; high empirical correlation does not may support equivalent precision. Applied here: even if the correlation between OVL and AUC is high, it cannot serve as proof of equivalence. Correlation measures “moving together,” while precision measures “moving accurately.”

    V. AUC Itself Is Unbiased Under Skewness; The Threshold Is Biased

    Another point prone to confusion regarding division of labor. Under skewness, the dual-normal ROC AUC tends toward unbiasedness and can still serve as a reasonable measure of global performance; what is biased is the decision threshold. Research therefore adopts skewed power distribution models capable of absorbing skewness and heavy tails to build ROC models, improving AUC estimation efficiency while restoring cut-point classification rates (PMID 37260468).

    The boundary is thus clear: global metrics can sustain under skewness, while local metrics cannot. OVL belongs to global shape metrics; it cannot substitute threshold information nor borrow AUC’s robustness for itself. Two global metrics—one stable, one unstable—already indicate they are not the same thing.

    VI. Error Bars Follow Separate Calibers

    Even retreating to “comparing values without conversion,” interval construction differs fundamentally in origin. OVL requires specialized methods: parametric, nonparametric, and bootstrap approaches each have specific applicability (PMID 34693817, PMID 16279139); handling covariates requires Box-Cox transformation followed by bootstrapping (PMID 40852736). The AUC side follows another set: nine analytical methods and three bootstrap approaches on the same dataset can yield intervals differing significantly (PMID 26734848).

    The widths, coverage rates, and sensitivity to sample size of the two intervals differ. “Inferring AUC intervals from OVL intervals” lacks a derivable logical chain.

    VII. Changing the Problem

    The true task is not verifying “Can OVL replace AUC” but performing three preparatory actions:

    First, report the skewness and variance ratios for each parameter, apply Box-Cox, and record how much OVL changes versus how much AUC changes. The change in AUC should be zero; this is instrument calibration. Non-zero change indicates an algorithmic error.
    Second, calculate the rank correlation between OVL and AUC on the same dataset, then recalculate after transformation. Compare the two correlations side-by-side; the difference represents the cost of skewness.
    Third, define the purpose: report AUC for global discrimination; report OVL for shape description but must declare no transformation was applied and comparisons remain within the same caliber.

    Four criteria correspond to four outcomes: if rank correlation is below 0.8, “overlap degree can replace discriminative power” does not hold and OVL serves only shape description; if deviation direction flips with variance ratio, the relationship is non-monotonic and any reverse inference is void; if rank correlation recovers above 0.9 after Box-Cox, OVL is usable but must declare transformation; if observed deviation is less than 5% while simulated deviation exceeds 30%, the current Dantian parameter distribution unexpectedly approaches normality, making the skewness category inapplicable—that would be good news but must first be measured to count.

    Final conclusion: Conversion tables sustained by assumptions must first test the assumptions before discussing conversion.

    Evidence and Limitations

    Verified: AUC as a rank statistic is invariant to monotonic transformations; transformation models leverage this to handle non-normal, ordinal, and censored data with unbiased estimates and nominal-level coverage (PMID 37278185); OVL inference methodology is systematically established, with the two-normal-parameter method outperforming nonparametric methods across wide distribution ranges and providing R implementation (PMID 34693817); comparisons of five OVL estimations show fixed-bandwidth kernel density nonparametric methods yield lower mean squared error and closer coverage to nominal levels in most scenarios (PMID 16279139); covariate-specific OVL requires regression plus Box-Cox with bootstrap intervals (PMID 40852736); different AUC confidence interval methods on the same dataset show significant differences, with bias-corrected accelerated bootstrap recommended (PMID 26734848).

    Partially Verified: Skewed-normal system simulations show Cohen’s d, CLES, and parameter-based overlap coefficients exhibit near-perfect descriptive correlation but vary greatly in bias and coverage performance; parameter-based OVL shows large bias and low coverage under skewness and heteroscedasticity (PMID 42094317)—directly supporting “high correlation does not equal equivalent precision,” though its outcome variable is psychological scale scores, not Dantian physiological parameters; nonparametric overlap indices are unbiased under shape differences but unreliable when true overlap exists, indicating nonparametric methods also have failure zones; under skewness, dual-normal ROC AUC tends toward unbiasedness while thresholds are biased, and skewed power distribution models can improve both simultaneously (PMID 37260468), indicating different fates for global versus local metrics under skewness, though this conclusion rests on a single clinical dataset (Klebsiella pneumoniae infection).

    To Be Verified: The observed rank correlation between OVL and AUC on the five Dantian parameters has never been calculated; no data exists on the magnitude of OVL variation before and after Box-Cox; whether deviation directions flip with variance ratios across four skewness types (log-normal, Beta bimodal, truncated normal) paired with three variance ratios is a simulation expectation rather than observation; the branch “observed deviation less than 5%” is purely hypothetical—if observed deviation exceeds 30%, all conversion-based practices within this project are invalidated, which is the primary falsifiable point of this paper.

    Related Reading

    Evidence Level: C
    References: PMID 37278185, PMID 34693817, PMID 16279139, PMID 40852736, PMID 26734848, PMID 42094317, PMID 37260468
    L1 Network Status: This topic has not yet entered the L1 correspondence network—the CORRESPONDENCE_NETWORK.md node table assigns nodes by domain (Yijing/TCM side, Physics side, Physiological/Experiential side) based on concepts/quantities (Five Phases, Five Organs, Frequencies, Colors…). None of the three domains contain measurement validity or external validity dimension nodes such as “Overlap Coefficient / Discriminative Power / Distribution Skewness / Monotonic Transformation Invariance / Error Bar Caliber”; the edge table ends at E-050 (“Six Heavenly Stems and Earthly Branchs Na Yin | Corresponds | 60 Frequencies”, Level L1, Status open, Modern Evidence column originally “No measurable sources at all“), with no edges sourced or targeted by distribution shape or discriminative power. EDGE_TABLE_science-yixue.md is a reverse statistical count, listing only concept counts (Dantian 51 / Qigong 48 / Breathing 189 / Frequency 407) and highest evidence level, lacking L1/L2/L3 hierarchy and active/open/conflict status judgments; thus, the state of this edge is missing.


    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.

  • 超额稳定性的 0.6 和 0.3 是借来的尺子:该标定的不是分数,是「分数到选对率」的映射曲线

    Evidence level: C

    0.6 站得住、0.3 真脆——这两条线从来没对着任何已知真值量过,它们是别处一把尺子上刻的记号。

    一、两个数的来历

    说法是「超额稳定性 ≥0.6 算稳,<0.3 算脆」。这两个数的出处,是评分者一致性里 kappa 的经验区间:0.6 上下大致对应「尚可」,0.3 以下对应「差」。而 kappa 那套区间本身是口头传统,且在低 base rate 下会过度惩罚一致(PMID 25090041)。把一条从一致性研究借来的尺子插到参数筛选上,中间没有任何一步对着已知真值标过。

    二、稳定性与准确性本来就是两个维度

    这一点在特征选择文献里是被明确分开写的。FREL 那篇开篇就说,好的特征选择算法由 accuracy 与 stability 两个因素刻画;它的集成版本把稳定性提上去之后,准确率只是「更好或相当」(PMID 25134091)。另一篇做集成选择的,稳定性分数明显改善,分类性能同样只是相当或略好(PMID 35016102)。更直接的一句在第三篇:小样本下不同方法的分类准确率可以彼此相当,但它们选出的特征在精度与稳定性上差得很远(PMID 22403633)。稳定,不等于对。

    三、这个领域其实标定过门槛,标的是错误率

    stability selection 的正统做法,是把选择概率的阈值与「期望假选个数」挂钩,用一个理论上界来控制。后来有人改用置换来估 FDR 定阈值,结果比 FWER 上界宽松、能捞回更多真信号(PMID 23089816)。再往后有人把它做成自动校准,把期望假阳性压到 20 个以下(PMID 38143734)。三条路指向同一件事:阈值该锚在「会选错几个」上,而不是锚在一个 0 到 1 之间、没有单位的分数上。

    四、同一个分数,换个选择器,选对率差几十倍

    而「会选错几个」这个锚,随选择器剧烈变动。一项 3300 次模拟的研究里,混合效应模型的 stability selection 把假发现率稳定压在 0.02 以下、真阳性率 0.73 到 0.97;同一个模拟里换成 Lasso,假发现率 0.59 到 0.72,真阳性率却 ≥0.93(PMID 35843027)。两者都在「选出参数」,假阳性比例差三十倍。所以「稳定性 0.6 对应多少选对率」这句话缺个主语:得先说清是哪个选择器。映射曲线不是一条,是一族。

    五、门槛里还有「先选后报」的水分

    你算的稳定性,是在「已经选出这 5 个」之后算的,而被选中的那几个稳定性天然偏高。Heinze 等的综述明确写着:变量选择会损害最终模型的稳定性、系数无偏性与 p 值有效性,因此建议把整个选择过程重抽样后的量作为常规报告项(PMID 29292533)。先选后报,门槛会显得更容易过。

    六、这里没有可算的零分布

    还有个硬约束。knockoff 那类方法之所以能给有限样本下的精确 FDR 保证,是因为它构造了一组可交换的对照变量,让真信号与假信号在交换下不可分辨(PMID 36854821)。分半 Jaccard 没有这种构造,它没有原假设下的精确分布,门槛只能靠经验标定或模拟——而模拟里填什么真值,它就答什么。

    七、反转:不标定分数,改报选对率

    反转不在「门槛调高还是调低」,在「该不该拿一个分数当主判据」。

    第一步,把二值化往后挪。empirical Bayes 那条线给的方案是不做 0-1 判决,直接对每个参数报 local fdr 与 local false sign rate,且输入用效应量与标准误两个数,而非一个 p 值或 z 值(PMID 27756721)。这样「选对率」这个概念本身就被拆细了。

    第二步,若一定要保留二值判决,就报两条直接估计:选对率(入选集合里真信号所占比例)与漏选率,而不是报一个间接的稳定性分数。配套做法是模拟端设 30 个候选、真信号 5 个,效应量、偏置、信度各三档,跑 1000 次分半与独立抽,拟合「超额稳定性 → 选对率」曲线并反解两个门槛,再用既有多设备丹田数据按文献共识参数做外部核对。

    四格判据:超额稳定性 ≥0.6 且对应选对率 ≥70%,则 0.6 可留用;0.6 对应选对率 <70%,则线太松、须上移;0.3 对应选对率已 >50%,则 0.3 太严、应改判「降级为探索性」;若 0.4 到 0.7 区间曲线近乎平坦,则稳定性与原命题脱钩、该指标不能作主判据。

    证据与局限

    已验证: 稳定性与准确性是可分离的两个维度,集成方法提稳而性能仅相当或略好(PMID 25134091、PMID 35016102);小样本下方法间准确率相当而选出特征的精度与稳定性差异很大(PMID 22403633);stability selection 的阈值可用置换估 FDR 或自动校准到给定假阳性数(PMID 23089816、PMID 38143734);同一数据上不同选择器的假发现率可相差三十倍(PMID 35843027)。

    部分验证: 上述均为方法学文献,其「选对率」是在模拟真值已知条件下算的;把同一套映射搬到丹田五参数上,须假定真实信号集已知,而这一点在现有多设备数据里并不成立。

    待验证: 0.6 与 0.3 对应的选对率究竟是多少,从未量过;映射曲线是否在 0.4 到 0.7 之间平坦,未知;模拟曲线与实测核对之间的偏差量级未知。

    相关阅读

    证据等级: C
    参考文献: PMID 25134091, PMID 35016102, PMID 22403633, PMID 23089816, PMID 38143734, PMID 35843027, PMID 29292533, PMID 36854821, PMID 27756721, PMID 25090041
    L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 节点表仅按「易学/中医侧」「物理侧」「生理/体验侧」三类域分派(原文「| 节点 | 各域的概念/量(五行、五脏、频率、颜色…) | 按域分派(见下表) |」),三类域均无「筛选阈值 / 稳定性分数 / 选对率 / 假发现率 / 局部 fdr / 阈值标定」这类测量学维度节点;边表止于 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」),无任何以筛选规则、稳定性分数或错误率门槛为源或目标的边。EDGE_TABLE_science-yixue.md 为反抽统计,仅有概念计数(丹田 51 篇 / 气功 48 篇 / 呼吸 189 篇 / 频率 407 篇)与最高证据级,无 L1/L2/L3 层级与 active/open/conflict 状态判定。


    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。

    Overstability’s 0.6 and 0.3 are borrowed rulers: what needs calibration is not the score, but the mapping curve from \”score\” to \”accuracy\

    Evidence level: C

    0.6 holds up; 0.3 is brittle—these two lines have never been calibrated against any known ground truth value. They are merely markings carved onto a ruler from elsewhere.

    I. The Origin of the Two Numbers

    The claim is that “excess stability ≥0.6 indicates stability” and “<0.3 indicates brittleness.” The origin of these numbers lies in the empirical ranges for kappa within inter-rater reliability: values around 0.6 roughly correspond to “fair,” while values below 0.3 correspond to “poor.” However, the kappa framework itself is rooted in oral tradition and tends to over-penalize agreement under low base rates (PMID 25090041). Inserting a ruler borrowed from consistency research into parameter screening involves no step that has ever been calibrated against known ground truth.

    II. Stability and Accuracy Are Inherently Two Distinct Dimensions

    This distinction is explicitly separated in the feature selection literature. The FREL paper opens by stating that good feature selection algorithms are characterized by two factors: accuracy and stability; its ensemble version increases stability while accuracy merely becomes “better or equivalent” (PMID 25134091). Another study on ensemble selection shows a marked improvement in stability scores, yet classification performance remains only equivalent or slightly better (PMID 35016102). A more direct statement from a third study notes that under small sample sizes, classification accuracies across different methods can be equivalent, yet the selected features differ significantly in precision and stability (PMID 22403633). Stability does not equal correctness.

    III. This Field Has Actually Calibrated Thresholds, But Targets Error Rates

    The orthodox approach to stability selection links the selection probability threshold to the “expected number of false selections,” using a theoretical upper bound for control. Later, some researchers adopted permutation methods to estimate FDR and set thresholds, resulting in bounds looser than the FWER upper bound and allowing recovery of more true signals (PMID 23089816). Subsequently, others automated the calibration to cap expected false positives below 20 (PMID 38143734). All three paths point to the same conclusion: thresholds should be anchored on “how many errors will occur,” not anchored on a unitless score between 0 and 1.

    IV. The Same Score Yields Selection Accuracy Differing by Orders of Magnitude Across Different Selectors

    The anchor of “how many errors will occur” varies drastically depending on the selector. In a simulation study involving 3,300 iterations, stability selection using mixed-effects models kept the false discovery rate (FDR) stable below 0.02 with true positive rates ranging from 0.73 to 0.97; switching to Lasso within the same simulation yielded an FDR of 0.59 to 0.72 but a true positive rate ≥0.93 (PMID 35843027). Both methods “select parameters,” yet the false positive proportions differ by a factor of thirty. Therefore, the statement “stability 0.6 corresponds to what selection accuracy” lacks a subject: one must first specify which selector is being used. The mapping curve is not a single line, but a family of curves.

    V. There Is “Post-Selection Reporting” Bias Within the Thresholds

    The stability you calculate is performed after “these 5 have already been selected,” meaning the selected variables naturally exhibit higher stability. Heinze et al.’s review explicitly states that variable selection compromises the stability of the final model, the unbiasedness of coefficients, and the validity of p-values; therefore, it is recommended to report quantities derived from resampling the entire selection process as a standard metric (PMID 29292533). Post-selection reporting makes thresholds appear easier to pass.

    VI. There Is No Calculable Null Distribution Here

    There is also a hard constraint. Methods like knockoff can provide exact FDR may support in finite samples because they construct a set of exchangeable control variables, rendering true signals indistinguishable from false ones under exchangeability (PMID 36854821). Half-split Jaccard lacks this construction; it possesses no exact distribution under the null hypothesis, so thresholds must rely on empirical calibration or simulation—and in simulation, whatever truth values are filled in will dictate the answer.

    VII. Reversal: Do Not Calibrate the Score; Report Selection Accuracy Instead

    The reversal is not about “raising or lowering the threshold,” but whether “a single score should serve as the primary criterion.”

    Step 1: Shift the Binary Decision. The empirical Bayes approach proposes not performing a 0-1 decision, but directly reporting the local false discovery rate (local fdr) and local false sign rate for each parameter, using effect size and standard error as inputs rather than a single p-value or z-score (PMID 27756721). This inherently fragments the concept of “selection accuracy.”

    Step 2: If Binary Decisions Must Be Retained, Report Direct Estimates. Instead of reporting an indirect stability score, report selection accuracy (the proportion of true signals within the selected set) and omission rates. Accompanying procedures involve setting up a simulation with 30 candidates, 5 true signals, and three levels each for effect size, bias, and reliability; run 1,000 half-split and independent draws to fit the “excess stability → selection accuracy” curve, solve for two thresholds inversely, and then perform external validation using multi-device Dantian data according to literature-consensus parameters.

    Four-Grid Criterion:
    * If excess stability ≥0.6 AND corresponding selection accuracy ≥70%, then 0.6 may be retained.
    * If excess stability 0.6 corresponds to selection accuracy <70%, then the line is too loose and must be raised.
    * If excess stability 0.3 corresponds to selection accuracy already >50%, then 0.3 is too strict and should be reclassified as “exploratory.”
    * If the curve between 0.4 and 0.7 is nearly flat, stability is decoupled from the proposition, and this metric cannot serve as a primary criterion.

    Evidence and Limitations

    Verified: Stability and accuracy are separable dimensions; ensemble methods increase stability while performance remains only equivalent or slightly better (PMID 25134091, PMID 35016102); under small samples, accuracies across methods are equivalent while precision and stability of selected features differ significantly (PMID 22403633); stability selection thresholds can be controlled via permutation estimation of FDR or automated calibration to a specific false positive count (PMID 23089816, PMID 38143734); false discovery rates across different selectors on the same data can differ by a factor of thirty (PMID 35843027).

    Partially Verified: The above are methodological literature where “selection accuracy” is calculated under conditions of known simulated truth; applying the same mapping to multi-device Dantian parameters requires assuming the true signal set is known, which does not hold for current multi-device data.

    To Be Verified: The exact selection accuracy corresponding to 0.6 and 0.3 has never been measured; whether the mapping curve is flat between 0.4 and 0.7 is unknown; the magnitude of deviation between simulated curves and empirical validation remains unknown.

    Related Reading

    Evidence Level: C
    References: PMID 25134091, PMID 35016102, PMID 22403633, PMID 23089816, PMID 38143734, PMID 35843027, PMID 29292533, PMID 36854821, PMID 27756721, PMID 25090041
    L1 Network Status: This topic has not yet entered the L1 Correspondence Network—the CORRESPONDENCE_NETWORK.md node table assigns nodes only by domain categories (“Yijing/TCM side,” “Physics side,” “Physiological/Experiential side”) based on concepts/quantities (Five Phases, Five Organs, Frequencies, Colors, etc.). None of the three domains contain measurement-level nodes related to “screening thresholds,” “stability scores,” “selection accuracy,” “false discovery rate,” “local fdr,” or “threshold calibration.” The edge table ends at E-050 (“Six Heavenly Stems and Earthly Branchs Na Yin | Corresponds | 60 Frequencies,” Level L1, status open; modern evidence column states “No measurable sources yet”). There are no edges originating from or terminating at screening rules, stability scores, or error rate thresholds. EDGE_TABLE_science-yixue.md is a reverse statistical count containing only concept counts (Dantian 51 articles / Qigong 48 articles / Breathing 189 articles / Frequency 407 articles) and highest evidence level, lacking L1/L2/L3 hierarchy and active/open/conflict status determination.


    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.

  • 样本量不是一个数:受试者、重复次数、设备数三条支路各有各的膝部,而且互不独立

    Evidence level: C

    把”要多少人”当成一个数字,等于把三种不同货币兑成一堆再数——加人数、加次数、加设备,摊的根本不是同一份方差。

    一、那个被漏掉的两个乘数

    设计讨论里的总样本 n,其实是三样东西的乘积:受试者数 k、每人重复次数 m、设备数 d。三者的功效机制完全不同。加受试者,增加的是独立信息单位;加重复次数,摊的是随机误差;加设备数,是把系统偏倚请进模型。把这三件事压成一个 n,后面的”曲线平台期”就必然说不清——因为你不知道是哪个乘数到顶了。

    二、加受试者:唯一没有内在上限的一条

    功效大致随受试者数的平方根增长,加人永远有效。这一条支路的「膝部」不是生理的,是预算的、伦理的。所以它不该被当成一个”天然上限”来找。真正会自己到顶的是另外两条——重复次数和设备数。把它们混在一起找拐点,最后找到的往往只是成本曲线。

    三、加重复:膝部位置由 ICC 决定,且可以算

    平均 m 次后的标准误按 σ/√m 收缩,前提是这 m 次之间的误差互相独立。但真实测量里的”日内波动”并非全是随机误差,含不可平摊的状态漂移,于是出现平台。Spearman-Brown 给出饱和速率:信度 = mρ/(1+(m−1)ρ)。PMID 38472632 证明了收敛速率正是这个公式,同时提醒短多维测验会正向偏倚、照此外推会过度预测。PMID 18470757 说得最直白:重复之间相关接近 0 时加次很划算,接近 1 时几乎不赚,并给出固定预算下人与次的最优配比。

    四、加设备:不是摊薄,是暴露

    一台设备的偏倚是固定偏移,重复测同一台摊不掉。多台设备在模型里相当于一个”设备”随机因子,会新增一个方差分量、占掉自由度。PMID 41380987 显示多扫描仪的异质性可以直接改变关联分析的结论,ComBat 之类的校正也只是部分补偿。PMID 32970522 算得更具体:簇大小不等时,要把簇数增加四成以上才补得回效率损失。设备数不等于更多样本。

    五、三条曲线不独立,这才是最大的坑

    加设备会移动重复支路的膝部:设备偏倚与重复误差在方差分解里会合成一项,设备数一变,组内方差的估计就变,膝部随之搬家。所以”三条支路各自找拐点再相加”这个做法本身是错的。PMID 41520835 把正确路径写成两阶段:G 研究先估出各来源的方差分量,D 研究再在指定设计下优化配比。PMID 22049666 给了一个可比的量级——受试者数对概化系数的作用明显大于题目数。

    六、设计阶段分量未知,所以别拿点估计去优化

    D 研究要求方差分量已知,可设计阶段恰恰不知道。正解是 maximin:在合理参数区间内保证最差情形下的功效。PMID 42479089 与 PMID 37270287 都给了这套做法与现成工具。更早的 PMID 3629046 画出信度研究的等值线:受试者数与重复次数是两条轴,任意一对组合都能达标——这本身就说明”最少要多少人”没有唯一答案。

    七、一个能直接执行的定位算法

    与其拟合三条边际曲线找膝部,不如先做一次小规模 G 研究:抽少量受试者,每人重复若干次,跨若干台设备,估出个体间、个体内、设备三个方差分量。然后代入 D 研究公式,算三条支路各自的边际功效增益。判据很简单——哪条支路每增加一单位资源的功效增益最大,就先加那条;增益最小的那条,说明瓶颈在别处分量上,加它只是烧钱。PMID 29718726 还提醒,忽略某些误差来源会系统性地扭曲信度与相关估计,等于在错误的方差结构上做优化。

    八、把问法换掉:先定位瓶颈方差分量

    三问:组间差主要被个体间真差异淹,还是被日内漂移淹,还是被设备偏倚淹?组内方差里有多少是可平摊的随机误差、多少是不可平摊的状态项?各台设备的偏倚是同向(可校正)还是异向(须分层)?答完这三问,”要多少人”才有确定含义。

    证据与局限

    已验证:概化理论把总变异拆成受试者、场合、设备等多个可分别估计的分量,并据此做设计优化,这条方法论路线在心理生理测量上已有系统综述(PMID 41520835);重复测量对功效的增益强烈依赖重复间相关,相关高时几乎无收益(PMID 18470757);多中心、多设备设计的最佳中心数与每中心人数需联合优化(PMID 42479089);簇数不足或簇大小不等会实质损失效率(PMID 32970522、PMID 21163852)。

    部分验证:Spearman-Brown 的收敛速率在随机抽样题目下成立(PMID 38472632),但短而多维的测验会高估,外推需谨慎;设备异质性可被统计校正部分抵消,但校正后关联分析的结论是否稳定,仍取决于异质性的结构(PMID 41380987)。

    待验证:本项目的丹田类参数(呼吸比、相位同步率、电抗)各自的三条支路拐点位置完全未知;组内方差里”可平摊随机误差”与”不可平摊状态项”的比例从未分离过;设备偏倚是同向还是异向也未测过。本节全部结论目前只能作为设计方法学的类比,不能当作已成立的实测规律。

    相关阅读

    证据等级: C
    参考文献: PMID 41520835, PMID 18470757, PMID 38472632, PMID 22049666, PMID 42479089, PMID 37270287, PMID 3629046, PMID 41380987, PMID 32970522, PMID 21163852, PMID 29718726, PMID 38878117
    L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 节点表只设易学/中医侧、物理侧、生理体验侧三域,无「样本量 / 重复次数 / 设备 / 方差分量 / 概化系数」等测量学与设计维度节点;EDGE_TABLE_science-yixue.md 边表止于 E-050(六十甲子纳音→60 频率,状态 open,现代证据栏原文”尚无任何可测来源”),无任何以测量误差或设计参数为端点的边。故本文证据等级封顶 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。

    Sample Size Is Not a Single Number: Three Distinct and Non-Independent Paths Including Subjects, Replication Counts, and Equipment

    Evidence level: C

    Treating “how many people” as a single number is like converting three different currencies into a pile and counting them—the sum of subjects, repetitions, and devices does not share the same variance.

    I. The Two Multipliers That Were Missed

    The total sample size $n$ discussed in design debates is actually the product of three distinct factors: the number of subjects ($k$), the number of repetitions per subject ($m$), and the number of devices ($d$). Their mechanisms for generating power are fundamentally different. Increasing subjects adds independent information units; increasing repetitions spreads out random error; increasing devices brings systematic bias into the model. Collapsing these three distinct actions into a single $n$ inevitably obscures the “plateau phase” of the curve—because you cannot determine which multiplier has reached its limit.

    II. Adding Subjects: The Only Path Without an Intrinsic Ceiling

    Power generally scales with the square root of the number of subjects; adding people is always effective. The “knee” of this path is not physiological but budgetary and ethical. Therefore, it should not be treated as a “natural ceiling” to search for. The other two paths—repetitions and devices—are the ones that naturally plateau. Mixing them together to find an inflection point often yields only a cost curve.

    III. Adding Repetitions: Knee Position Determined by ICC and Calculable

    The standard error after an average of $m$ repetitions contracts according to $\sigma/\sqrt{m}$, provided the errors across these $m$ instances are mutually independent. However, “intra-day fluctuation” in real-world measurements is not purely random error; it contains unspreadable state drifts, leading to a plateau. Spearman-Brown provides the saturation rate: $\text{Reliability} = \frac{m\rho}{1+(m-1)\rho}$. PMID 38472632 proves that the convergence rate follows exactly this formula, while simultaneously warning that short, multi-dimensional tests exhibit positive bias and that extrapolation leads to overprediction. PMID 18470757 states the most plainly: adding repetitions is highly cost-effective when inter-repetition correlation approaches 0, but yields almost no gain when it approaches 1, and provides the optimal ratio of people to repetitions under a fixed budget.

    IV. Adding Devices: Not Dilution, but Exposure

    The bias of a single device is a fixed offset; repeating measurements on the same device does not dilute it. Multiple devices in a model function like a “device” random factor, adding a variance component and consuming degrees of freedom. PMID 41380987 shows that heterogeneity across multiple scanners can directly alter conclusions in association analyses, and corrections like ComBat only provide partial compensation. PMID 32970522 calculates more specifically: when cluster sizes are unequal, cluster numbers must be increased by over 40% to recover efficiency losses. The number of devices is not equivalent to a larger sample size.

    V. The Three Curves Are Not Independent; This Is the Biggest Pitfall

    Adding devices shifts the knee of the repetition path: device bias combines with repetition error in variance decomposition; as the number of devices changes, the estimate of within-subject variance changes, and the knee moves accordingly. Therefore, the practice of “finding an inflection point for each path separately and then summing them” is fundamentally flawed. PMID 41520835 writes the correct path in two stages: G studies first estimate variance components from each source, and D studies then optimize allocation ratios under a specified design. PMID 22049666 provides a comparable magnitude—the effect of subject numbers on the generalizability coefficient is clearly greater than that of item numbers.

    VI. Design-Stage Variance Components Are Unknown, So Do Not Optimize Using Point Estimates

    D studies require known variance components, yet these are precisely unknown at the design stage. The correct approach is maximin: may support power under the worst-case scenario within a reasonable parameter range. PMID 42479089 and PMID 37270287 both provide this methodology along with ready-made tools. Earlier, PMID 3629046 plotted the iso-reliability contours for reliability studies: subject numbers and repetitions form two axes, and any pair of combinations can meet the standard—this itself demonstrates that “how many people are needed” has no unique answer.

    VII. An Executable Positioning Algorithm

    Rather than fitting three marginal curves to find knees, conduct a small-scale G study first: recruit a few subjects, have each repeat several times, and use multiple devices. Estimate the three variance components: between-individual, within-individual, and device-specific. Then substitute these into the D study formula to calculate the marginal power gain for each path separately. The criterion is simple: whichever path offers the greatest power gain per additional unit of resource should be increased first; if the gain is minimal on a path, the bottleneck lies in another variance component, and adding to it merely burns money. PMID 29718726 also warns that ignoring certain error sources systematically distorts reliability and correlation estimates, equivalent to optimizing under an incorrect variance structure.

    VIII. Change the Question: First Locate the Bottleneck Variance Component

    Three questions: Is between-subject variance primarily drowned out by true between-individual differences, intra-day drift, or device bias? Within-subject variance: what proportion is spreadable random error versus unspreadable state terms? Are device biases unidirectional (correctable) or multidirectional (requiring stratification)? Once these three questions are answered, “how many people” acquires a definite meaning.

    Evidence and Limitations

    Verified: Generalizability theory decomposes total variance into estimable components such as subjects, occasions, and devices, using these for design optimization; this methodological route has a systematic review in psychophysiological measurement (PMID 41520835); the gain from repeated measurements for power strongly depends on inter-repetition correlation, yielding almost no benefit when correlation is high (PMID 18470757); optimal number of centers and subjects per center in multi-center, multi-device designs must be jointly optimized (PMID 42479089); insufficient cluster numbers or unequal cluster sizes result in substantial efficiency losses (PMID 32970522, PMID 21163852).

    Partially Verified: Spearman-Brown’s convergence rate holds under random sampling of items (PMID 38472632), but short, multi-dimensional tests may overestimate it, requiring caution in extrapolation; device heterogeneity can be partially offset by statistical correction, but the stability of association analysis conclusions after correction still depends on the structure of heterogeneity (PMID 41380987).

    To Be Verified: The knee positions for the three paths of our project’s Dantian parameters (breathing ratio, phase synchronization rate, impedance) are entirely unknown; the proportion of “spreadable random error” versus “unspreadable state terms” within intra-subject variance has never been separated; whether device biases are unidirectional or multidirectional has not been measured. All conclusions in this section currently serve only as analogies for design methodology and cannot be treated as established empirical laws.

    Related Reading

    Evidence Level: C
    References: PMID 41520835, PMID 18470757, PMID 38472632, PMID 22049666, PMID 42479089, PMID 37270287, PMID 3629046, PMID 41380987, PMID 32970522, PMID 21163852, PMID 29718726, PMID 38878117
    L1 Network Status: This topic has not yet entered the L1 corresponding network—the CORRESPONDENCE_NETWORK.md node table only sets domains for Yijing/TCM, Physics, and Physiological Experience, lacking nodes for “Sample Size / Repetitions / Devices / Variance Components / Generalizability Coefficient” measurement and design dimensions; the EDGE_TABLE_science-yixue.md edge table ends at E-050 (Six Heavenly Stems and Earthly Branches Na Yin → 60 Frequencies, status open, modern evidence column text states “No measurable sources yet”), with no edges terminating at measurement error or design parameters. Thus, the evidence level for this article is capped at C.


    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.

  • 范围损失不该用「结论还成立的比例」来数:换成分布重叠度,反而更靠不住

    Evidence level: C

    分布重叠度量的是一群人长得像不像,范围损失问的是结论搬过去还灵不灵——这两件事之间缺一整座桥。

    一、这个替换想解决什么,又解决了什么

    上一篇把降级损失拆成两笔账:功效损失能靠加样本补,范围损失补不回。但「补不回多少」是用「参数真值空间中结论仍成立的比例」算的。这个定义有毛病:真值空间是自己设的,模拟端填什么它就答什么,等于自出自判。

    于是有了这个想法:改用修炼者与新手两群的协变量分布重叠度来度量——重叠越低,人群差越远,结论越不可靠。听上去很顺,但顺的地方往往就是坑。

    二、先分清两个词:重叠是形状,损失是可用性

    分布重叠系数(overlap coefficient, OVL)是两个分布密度函数逐点取小后的积分,取值 0 到 1。它是纯几何量:只描述两条曲线的交叠面积,不含任何关于「结论」的信息。

    范围损失问的是另一件事:把在 A 群成立的结论拿到 B 群,方向和数值还对不对。一个变量可以两群分布天差地别(OVL 0.4),但它根本不是效应修饰变量,搬过去毫无损失;也可以两群分布几乎一样(OVL 0.95),而它恰好是强效应修饰变量,一搬就翻车。OVL 只看得见前者,看不见后者。

    三、重叠系数本身并不好算

    这个量在文献里长期被冷落,主因是缺软件、缺覆盖全分布情形的推断方法。近年才有系统处理:参数法(二正态模型)在大范围分布情形下优于非参数法,并提供 R 代码(PMID 34693817);另有工作把它扩展为「协变量特异」版本,用线性回归加可能的 Box-Cox 变换估计,自助法构造置信区间(PMID 40852736)。

    后一篇的关键含义:OVL 本身可以被协变量改写。年龄一变,重叠系数就变。那么把「某变量上的 OVL」当作固定的人群差指标,前提就已经不成立。

    四、更要紧的是:它不是新信息

    在二正态假设下,OVL 与曲线下面积(AUC)之间存在确定的单调对应——知道一个就能算出另一个。这意味着 OVL 并不携带判别力之外的独立信息,只是 AUC 的一个非线性重标。

    把一个本质上是判别力的量改名叫「可推广性」,正是这类口径替换最典型的自欺:数字变了,含义没变。判别力回答「能不能分开两类人」,可推广性回答「换个群体结论还对不对」——后者还取决于效应修饰结构,那部分信息 OVL 里一个比特都没有。

    五、加权合成那一步最脆

    第二步是「重叠系数 × 该变量对效应量的调节幅度」再加权合成。这里有两个独立的塌方点。

    其一,调节幅度估不准。在按总体效应设计的试验里,交互检验的功效远低于总体检验:总体效应有 80% 功效的试验,检出同量级交互只有约 29% 功效;要同等功效,样本量须扩到四倍(PMID 15066682)。亚组分析假阳性与假阴性的量级也早有量化(PMID 11701102)。乘子里最不确定的那一项,恰好是权重最大的一项。

    其二,量纲不可通约。重叠系数是无量纲比例,调节幅度是效应量单位,两者相乘得到的指数没有物理含义,只能内部比较——它答不了「损失百分之几」,只能答「比另一组大还是小」。

    六、单变量重叠不等于联合重叠

    多个协变量各自重叠 0.9,联合起来可能高度不重叠,这是高维几何的常识。而多变量的「重叠系数」没有唯一定义:逐点取小在高维仍可写,但估计方差爆炸;改用能量距离或核两样本检验,得到的是检验统计量而非 0-1 比例,不能乘调节幅度。

    更要紧的是方向性。真正的可推广性识别条件要求「参与×处理」在绝对或相对尺度上无交互(PMID 40266689),这是关于结构的主张,不是关于分布的描述;后分层与 g 方法能借目标人群协变量加权(PMID 41992116),前提仍是效应修饰变量集合正确。分布差异只是必要条件的一个代理,从来不是充分条件。

    七、反转:不问损失多大,问损失落在谁身上

    与其造一个合成指数,不如把问题改成可执行的三问:

    其一,先做平衡诊断。标准化均值差、方差比逐变量报(PMID 30788363),别急着压成一个数。其二,把交互检验的功效算出来写进方案——若某变量的交互在现有样本下检出功效不足 30%(PMID 15066682),该变量的「调节幅度」应标为不可估,而不是填个点估计去乘。其三,报告适用范围时只报「本研究人群的协变量区间」,不报「可外推全体」——社区抽样与试验入选标准的落差可以大到只有中位 4% 的患者符合条件(PMID 17105779)。

    判据也据此改写:若逐变量诊断显示关键效应修饰变量标准化均值差 <0.1,则降级过度保守,可保留合并;若 >0.25 且该变量交互功效 >80%,则降级必要,应按该变量分层报告;若交互功效 <30%,则无论重叠多高,都只能报「该变量调节幅度未知、结论适用范围未经检验」——这时任何合成指数都是把未知数乘上一个系数,得到的仍是未知数。

    证据与局限

    已验证: OVL 的推断方法学已系统建立,参数法(二正态)在大范围分布情形下优于非参数法,已有 R 实现(PMID 34693817);协变量特异 OVL 可用回归加 Box-Cox 估计、自助法给置信区间(PMID 40852736);按总体效应设计的试验对交互检验功效严重不足,同量级交互仅约 29% 功效、需四倍样本(PMID 15066682);亚组分析的假阳性与假阴性风险已量化(PMID 11701102);倾向得分匹配后的平衡诊断应以标准化均值差与方差比逐变量报告(PMID 30788363);试验入选标准与社区人群的落差可量化,哮喘领域仅中位 4% 符合(PMID 17105779);从试验外推需「参与×处理」无交互的结构性假设(PMID 40266689),后分层可借目标人群协变量加权(PMID 41992116)。

    部分验证: 效应修饰与不可压缩性并存时,边际估计与条件估计可给出方向相反的治疗决策(PMID 41626969)——这支持「口径选择本身会改变结论」,但其结论建立在二值结局与比数比尺度上,连续型丹田参数是否同样敏感尚未检验;无正性假设时重叠权重法把目标人群限定为「有足够重叠者」(PMID 38847059),思路与本文一致,但那是估计处理效应的方法,不是度量范围损失的方法;加法与乘法尺度下交互的度量不可互换(PMID 16620385),提示「调节幅度」必须先声明尺度。

    待验证: 「加权范围损失指数」在丹田五参数上从未算过,其与「结论成立比例」口径的差值方向与量级均无数据;OVL 与 AUC 在二正态下的对应关系是否可外推到实际偏态的丹田参数分布,尚未核;修炼者与新手两群在候选效应修饰变量上的重叠系数实测值(提案预期普遍 >0.8)纯属假设;若实测确认普遍 >0.8,则两笔账的二分在本项目内失效、须整条撤回——这一条本身就是本篇最重要的可证伪点。

    相关阅读

    证据等级: C
    参考文献: PMID 40852736, PMID 34693817, PMID 15066682, PMID 11701102, PMID 30788363, PMID 17105779, PMID 40266689, PMID 41992116, PMID 41626969, PMID 38847059, PMID 16620385
    L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 节点表按易学/中医侧、物理侧、生理/体验侧三类域分派(原文「| 节点 | 各域的概念/量(五行、五脏、频率、颜色…) | 按域分派(见下表) |」),三类域均无「范围损失 / 效应修饰变量 / 协变量分布重叠 / 目标人群」这类测量学与外效度维度节点;边表止于 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」),无任何以分布重叠、效应修饰或可推广性为源或目标的边。EDGE_TABLE_science-yixue.md 为反抽统计,仅列概念计数(丹田 51 篇 / 气功 48 篇 / 呼吸 189 篇 / 频率 407 篇)与最高证据级,无 L1/L2/L3 层级与 active/open/conflict 状态判定,故本条边状态为缺。


    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。

    Range loss should not be counted by \”proportion of valid conclusions\”: using distribution overlap is even less reliable

    Evidence level: C

    Distribution overlap measures how similar a group of people looks; range loss asks whether a conclusion remains valid when transferred to another group—there is a missing bridge between these two concepts.

    I. What This Replacement Aims to Solve, and What It Actually Solves

    The previous post decomposed degradation loss into two accounts: efficacy loss can be compensated by increasing sample size, but range loss cannot. However, “how much cannot be compensated” was calculated using the “proportion of conclusions that still hold within the true value space.” This definition has flaws: the true value space is self-defined, and the simulation endpoint answers whatever it is fed, amounting to self-validation.

    Thus arose this idea: use the overlap of covariate distributions between practitioners and novices to measure it—lower overlap implies greater population difference and less reliable conclusions. It sounds smooth, but smoothness often hides pitfalls.

    II. First Distinguish Two Terms: Overlap Is Shape, Loss Is Usability

    The overlap coefficient (OVL) is the integral of the pointwise minimum of two distribution density functions, ranging from 0 to 1. It is a purely geometric quantity: it describes only the intersection area of two curves and contains no information regarding “conclusions.”

    Range loss asks about a different matter: whether the direction and magnitude of a conclusion valid in Group A remain correct when applied to Group B. One variable may have vastly different distributions between two groups (OVL 0.4), yet not be an effect modifier at all, resulting in no loss upon transfer; conversely, two groups may have nearly identical distributions (OVL 0.95), yet the variable happens to be a strong effect modifier, causing failure upon transfer. OVL sees only the former, not the latter.

    III. The Overlap Coefficient Itself Is Difficult to Calculate

    This quantity has long been neglected in the literature primarily due to a lack of software and inferential methods covering all distribution scenarios. Systematic handling has only recently emerged: parametric methods (two-normal models) outperform non-parametric methods in broad distribution scenarios and provide R code (PMID 34693817); other work extends it to a “covariate-specific” version using linear regression with potential Box-Cox transformations for estimation and bootstrap methods to construct confidence intervals (PMID 40852736).

    The key implication of the latter post is: OVL itself can be rewritten by covariates. If age changes, the overlap coefficient changes. Therefore, treating “OVL on a specific variable” as a fixed metric of population difference presupposes a condition that is already invalid.

    IV. More Importantly: It Is Not New Information

    Under the assumption of bivariate normality, OVL has a deterministic monotonic correspondence with the area under the curve (AUC)—knowing one allows calculation of the other. This means OVL does not carry independent information beyond discriminative power; it is merely a non-linear rescaling of AUC.

    Renaming a quantity that is essentially about discriminative power as “generalizability” is a classic form of self-deception in such metric substitutions: the number changes, but the meaning does not. Discriminative power answers “Can we distinguish two types of people?” while generalizability asks “Is the conclusion still correct in a different population?”—the latter depends on effect modification structures, information of which OVL contains zero bits.

    V. The Weighted Synthesis Step Is Most Fragile

    The second step is “overlap coefficient × the magnitude of modulation by that variable on the effect size,” then weighted synthesis. Here lie two independent collapse points.

    First, modulation magnitude is estimated inaccurately. In trials designed for overall effects, the power for interaction tests is far lower than for overall tests: a trial with 80% power for overall effects detects interactions of similar magnitude with only about 29% power; to achieve equivalent power, sample size must be quadrupled (PMID 15066682). The magnitudes of false positives and false negatives in subgroup analysis have also been quantified for some time (PMID 11701102). The least certain term in the multiplier is precisely the one with the largest weight.

    Second, units are incommensurate. The overlap coefficient is a dimensionless ratio, while modulation magnitude is in effect size units; their product yields an index with no physical meaning, usable only for internal comparison—it cannot answer “what percentage of loss,” only “larger or smaller than another group.”

    VI. Univariate Overlap Does Not Equal Joint Overlap

    With multiple covariates, each having an overlap of 0.9, they may jointly be highly non-overlapping—a common knowledge of high-dimensional geometry. Furthermore, the “overlap coefficient” for multivariate data lacks a unique definition: taking the pointwise minimum in high dimensions can still be written, but estimation variance explodes; using energy distance or kernel two-sample tests yields a test statistic rather than a 0-1 proportion, which cannot be multiplied by modulation magnitude.

    More importantly is directionality. True conditions for generalizability identification require no interaction between “participation × treatment” on absolute or relative scales (PMID 40266689)—a claim about structure, not description of distributions; post-stratification and the g-method can weight by target population covariates (PMID 41992116), but only if the set of effect modifiers is correct. Distribution differences are merely a proxy for a necessary condition, never a sufficient one.

    VII. Inversion: Do Not Ask How Large the Loss Is, But Where It Falls

    Rather than constructing a synthetic index, reframe the problem into three executable questions:

    First, conduct balance diagnostics. Report standardized mean differences and variance ratios variable by variable (PMID 30788363); do not rush to compress them into a single number. Second, calculate the power for interaction tests and incorporate it into the protocol—if the interaction for a specific variable has insufficient detection power (<30%) under current sample size (PMID 15066682), mark that variable’s “modulation magnitude” as unestimable rather than filling in a point estimate to multiply. Third, when reporting applicability, report only the “covariate range of this study population,” not “extrapolable to all”—the gap between community sampling and trial inclusion criteria can be so large that only the median 4% of patients meet eligibility (PMID 17105779).

    Criteria are also rewritten accordingly: if variable-by-variable diagnostics show standardized mean differences for key effect modifiers <0.1, degradation is overly conservative, and pooling may be retained; if >0.25 and interaction power >80%, degradation is necessary, and stratification by that variable should be reported; if interaction power <30%, regardless of how high the overlap is, only report “modulation magnitude for this variable unknown, applicability of conclusions not verified”—at this point, any synthetic index multiplies an unknown by a coefficient, yielding still an unknown.

    Evidence and Limitations

    Verified: Inferential methodologies for OVL are systematically established; parametric methods (two-normal) outperform non-parametric methods in broad distribution scenarios with R implementation available (PMID 34693817); covariate-specific OVL can use regression with Box-Cox estimation and bootstrap for confidence intervals (PMID 40852736); trials designed for overall effects suffer severely from insufficient power for interaction tests, with similar interactions having only ~29% power and requiring quadrupled sample sizes (PMID 15066682); false positive and false negative risks in subgroup analysis have been quantified (PMID 11701102); balance diagnostics after propensity score matching should report standardized mean differences and variance ratios variable by variable (PMID 30788363); the gap between trial inclusion criteria and community populations is quantifiable, with only the median 4% eligible in asthma domains (PMID 17105779); extrapolation from trials requires structural assumptions of no interaction between “participation × treatment” (PMID 40266689), and post-stratification can weight by target population covariates (PMID 41992116).

    Partially Verified: When effect modification and incompressibility coexist, marginal and conditional estimates can yield opposite treatment decisions (PMID 41626969)—supporting the notion that “metric choice itself alters conclusions,” but its conclusions are built on binary outcomes and odds ratio scales; whether continuous-type丹田 parameters are similarly sensitive remains untested; without positivity assumptions, overlap weighting methods restrict the target population to “those with sufficient overlap” (PMID 38847059), aligning with this paper’s思路 but serving as an estimation method for treatment effects rather than a metric for range loss; interaction metrics are not interchangeable under additive and multiplicative scales (PMID 16620385), suggesting “modulation magnitude” must first declare its scale.

    To Be Verified: The “weighted range loss index” has never been calculated on丹田 five parameters; the direction and magnitude of difference between this metric and the “proportion of conclusions holding” lack data; whether the correspondence between OVL and AUC under bivariate normality can be extrapolated to actual skewed丹田 parameter distributions remains unverified; measured OVL values between practitioners and novices on candidate effect modifiers are purely hypothetical (proposal expects generally >0.8); if measured confirmation shows generally >0.8, the binary accounting within this project fails and must be withdrawn entirely—this point itself is the most falsifiable aspect of this paper.

    Related Reading

    Evidence Level: C
    References: PMID 40852736, PMID 34693817, PMID 15066682, PMID 11701102, PMID 30788363, PMID 17105779, PMID 40266689, PMID 41992116, PMID 41626969, PMID 38847059, PMID 16620385
    L1 Network Status: This topic has not yet entered the L1 corresponding network—the CORRESPONDENCE_NETWORK.md node table assigns nodes by Yijing/TCM side, physics side, and physiological/experiential side (original text “| Node | Concept/Quantity in each domain (Five Phases, Five Organs, Frequency, Color…)| Assigned by domain (see table below)|”), with no nodes in any of the three domains measuring “range loss / effect modifier / covariate distribution overlap / target population”; edge tables stop at E-050 (“60 Heavenly Stems and Earthly Branchs Na Yin | Corresponds | 60 Frequencies”, Level L1, status open, modern evidence column originally “No measurable sources at all“), with no edges sourced or targeted by distribution overlap, effect modification, or generalizability. EDGE_TABLE_science-yixue.md is a reverse statistical count, listing only concept counts (丹田 51 papers / Qigong 48 papers / Breathing 189 papers / Frequency 407 papers) and highest evidence level, lacking L1/L2/L3 hierarchy and active/open/conflict status determination; thus, the status of this edge is missing.


    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.

  • 阈值该离众数多远才配二分:真正决定 Jaccard 的不是百分位,是稀有侧还剩几个样本

    Evidence level: C

    阈值定在分布第几分位,不是一个能自由拧的旋钮;它是被样本量锁死的秩序统计量。

    一、这个问法预设了一个不存在的旋钮

    上一轮把验证顺序倒过来:先看 ≥2 落在第几分位,再决定二分值不值得留。听起来合理,但它默认「距众数多远」可以自由选——想落 85 分位就落 85 分位。

    分位位置不是自由参数。95 分位在 n=20 里就是第 19 个秩序统计量,其值由排序后倒数第二个观测单独决定;换个人、换个带位,这个观测换成另一个数,切点就整个搬家。横轴看似连续,实际只有几个格点,每格自带一份由 n 决定的抖动。

    二、二分的账先记清

    连续变量切成两类会丢信息、掉功效,这是有关系统研究的老结论:多变量回归中把连续预测因子分组被判为「坏主意」(PMID 16217841);整群随机试验里二分连续结局的功效损失可算得出(PMID 22733454);二分的统计性质与连续不可互换(PMID 18389492);连年龄二分后得到的比数比都有偏(PMID 17066378)。

    替代路线也已存在:同一批原始研究用分布法可同时给出连续与二分两种估计,且推断一致(PMID 24920271)。方向提示很清楚——不是「二分怎么切更稳」,而是能不能不切。

    三、切点偏倚的实测账:主因是样本量

    真正量过这件事的一项模拟研究设敏感性、特异性真值都是 84%,用 Youden 指数数据驱动选切点(PMID 18258670)。n=40 时估成约 90%,n=200 时约 86%,偏倚与样本量成反比。

    同一段结果里还有一句更该记住:样本量固定时,结果分布形态对该偏倚几乎没有影响。这句直接顶掉「按分布形态分别定线」这个前置任务——至少在这类偏倚上,形状不是自变量。研究还试了三种更稳的做法(假定分布、留一法、平滑 ROC),它们更不易偏,但假设不满足时性能会崩。对应的修补工具是:选完切点后的效应须用收缩加自助法才能恢复置信区间覆盖率(PMID 15160403)。

    四、小样本里这两个百分位分辨不出来

    还有更基础的问题:这个 n 下,70 与 85 分位能不能分开。

    分位数与百分位的置信区间有专门算法(PMID 30591808);参考限本质是对极端百分位的估计,精度由样本量决定,且有现成的最少受试者数算法(PMID 25029084)。落到格点:n=20 时 85 分位只有 3 个观测在其上,95 分位只剩 1 个。两个切点值的区间几乎必然重叠——「放 85 还是放 95」在这个样本量下不是可执行的操作,是同一个模糊地带的两种叫法。

    小样本参考区间的实践建议同向:把数值画成点图,用多种方法各算参考限,看是否互相靠得住(PMID 19473330);个体内变异大、强偏态、样本又小时,分布估计尤须谨慎(PMID 21731012)。

    五、「双峰」这个前提自己先要过关

    「距众数多远」隐含一个形态判断。双峰意味存在天然分界,二分才有道理,但形态判断很脆。

    三种常用度量(双峰系数 BC、Hartigan dip、单成分对双成分的 AIC 差)被系统比对过(PMID 22806703):距离影响总强于比例,比例的影响在三种度量间差异极大;偏度会虚高双峰性检测,某些情形引发反常交互;AIC 差极度敏感,几乎把一切分布判成双峰。BC 与 dip 大体一致,但存在若干重要分歧。

    意思是:一个偏态单峰分布很可能被读成双峰,于是你以为该用双峰线,其实数据从头到尾只有一个峰。把形态放在推理链最上游,等于用一个判别力有限的判断去支撑全部结论。

    六、把横轴换成稀有侧计数 k

    改造的入手点:百分位难用,因为它把距离和样本量搅在一起;把横轴换成阈值以上的样本个数 k,两者就分开了。

    同一条 15% 经验线,n=100 是 15 人,n=20 只剩 3 人。前者切点稳、分半后每半仍有七八人;后者每半只剩一两人,Jaccard 的分子分母都由这几个观测决定。这解释了为什么同一条经验线在不同研究里表现迥异——它根本不是同一个门槛。

    换成 k 后得到一个可检验推论:若稳定性主要由此决定,则四种分布形态的「k → Jaccard」曲线应当收敛。

    七、反转与四条判据

    反转是:不问「阈值该离众数多远」,改问「稀有侧至少留几个样本才够」。

    四条判据:横轴换 k 后四形态最大差 <0.05,则分位与形态皆降为伪变量,规则改写成「稀有侧 k≥X」;换 k 后仍明显分形态(>0.05),则形状真起作用、须按形态分线,但形态判定须先做偏度校正、三种度量并列报;k 固定而 Jaccard 仍随 n 下降,则问题在个体内相关与分半结构,回到分半单位那条线;算出 70 与 85 分位切点的置信区间发现重叠,则该样本量下任何「前 15-20%」经验线都不可执行,只能报 k 与它的实际人数。

    证据与局限

    已验证: 数据驱动选最优切点致敏感性特异性高估,偏倚随样本量减小放大,n=40 时真值 84% 估成约 90%(PMID 18258670);样本量固定时形态对该偏倚影响很小(同前);选切点后置信区间覆盖不足,须收缩加自助法(PMID 15160403);二分损信息、掉功效、产生有偏比数比(PMID 16217841、PMID 22733454、PMID 18389492、PMID 17066378);偏度虚高双峰检测,三种度量结论不一致(PMID 22806703);极端百分位估计精度由样本量决定,有最少受试者数算法(PMID 25029084、PMID 30591808)。

    部分验证: 分布法可在同一批原始研究中同时给出连续与二分估计且一致(PMID 24920271),但属对已发表研究的二次分析,非前瞻验证;小样本参考区间须多方法并列(PMID 19473330)、个体内变异大与强偏态须谨慎(PMID 21731012),均为方法学建议而非本参数的实测。

    待验证: 「阈值分位 → Jaccard 中位数」曲线从未在丹田五参数上画过,拐点在 80-85 分位只是待检验假设;换成计数 k 后四形态曲线是否收敛,是本篇核心可检验命题,尚无数据;70 与 85 分位切点置信区间是否实际重叠,也尚未算。

    相关阅读

    证据等级: C
    参考文献: PMID 18258670, PMID 15160403, PMID 16217841, PMID 22733454, PMID 18389492, PMID 17066378, PMID 24920271, PMID 22806703, PMID 25029084, PMID 30591808, PMID 19473330, PMID 21731012
    L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 节点表仅按易学/中医侧、物理侧、生理/体验侧三类域分派(原文「| 节点 | 各域的概念/量(五行、五脏、频率、颜色…) | 按域分派(见下表) |」),三类域均无「筛选阈值 / 切点 / 分位位置 / 分布形态判定 / 入选集合」这类测量学维度节点;边表止于 E-050(「六十甲子纳音 | 对应 | 60 个频率」,状态 open),无任何以阈值、切点或分布形态为源或目标的边。EDGE_TABLE_science-yixue.md 亦仅有概念计数(丹田 51 篇 / 气功 48 篇 / 呼吸 189 篇 / 频率 407 篇),无 L1/L2/L3 层级与 active/open/conflict 状态判定。


    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。

    How Far Should the Threshold Be from the Median for Binary Classification: What Really Determines Jaccard Similarity Is Not the Percentile but How Many Samples Remain on the Rar…

    Evidence level: C

    Setting a threshold at a specific percentile is not a freely adjustable knob; it is an order statistic locked by sample size.

    I. This Question Presupposes a Nonexistent Knob

    In the previous round, we reversed the verification sequence: first check where $\ge 2$ falls in the percentile distribution, then decide that a binary split is not worth retaining. This sounds reasonable, but it assumes that “how far from the mode” can be freely selected—choosing the 85th percentile if desired.

    Percentile positions are not free parameters. The 95th percentile with $n=20$ is the 19th order statistic, a value determined solely by the second-to-last observation in the sorted sequence; changing an individual or the grouping shifts this observation, causing the cutoff point to relocate entirely. The horizontal axis appears continuous, but it actually consists of only a few grid points, each carrying its own jitter determined by $n$.

    II. Keep the Binary Ledger Clear First

    Splitting a continuous variable into two categories loses information and reduces statistical power; this is an established conclusion in systems research: grouping continuous predictors in multivariate regression has been deemed a “bad idea” (PMID 16217841); power loss in binary continuous outcomes in cluster randomized trials can be calculated (PMID 22733454); the statistical properties of binary variables are not interchangeable with continuous ones (PMID 18389492); even odds ratios derived from binary age groups are biased (PMID 17066378).

    Alternative routes already exist: the same original studies can use distribution methods to simultaneously provide both continuous and binary estimates with consistent inference (PMID 24920271). The directional hint is clear—it is not “how to cut more stably,” but rather whether one can avoid cutting altogether.

    III. Empirical Accounting for Cutoff Bias: Main Driver is Sample Size

    One simulation study that truly quantified this set true sensitivity and specificity at 84%, using data-driven selection via the Youden index (PMID 18258670). With $n=40$, the estimate became approximately 90%; with $n=200$, it was about 86%, indicating bias inversely proportional to sample size.

    Also worth remembering from the same results: when sample size is fixed, the distribution shape has little effect on this bias. This directly negates the prerequisite task of “drawing lines based on distribution shape”—at least regarding this type of bias, shape is not the independent variable. Studies also tested three more robust approaches (assuming distribution, leave-one-out, and smoothed ROC); these are less prone to bias but performance collapses when assumptions are violated. The corresponding correction tool is that effect estimates after selecting a cutoff must use shrinkage with bootstrapping to restore confidence interval coverage (PMID 15160403).

    IV. These Percentiles Are Indistinguishable in Small Samples

    There is a more fundamental issue: can the 70th and 85th percentiles be distinguished at this $n$.

    Confidence intervals for quantiles and percentiles have specific algorithms (PMID 30591808); reference limits are essentially estimates of extreme percentiles, their precision determined by sample size, and there are existing algorithms for minimum subject numbers (PMID 25029084). On the grid: with $n=20$, the 85th percentile rests on only 3 observations, while the 95th percentile rests on just 1. The intervals for these two cutoff values almost certainly overlap—”placing at 85 or 95″ under this sample size is not an executable operation but two names for the same fuzzy zone.

    Practical suggestions for small-sample reference intervals align: plot values as dot plots, calculate reference limits using multiple methods to see if they hold up (PMID 19473330); when within-individual variability is large, distributions are strongly skewed, and sample sizes are small, distribution estimation requires extreme caution (PMID 21731012).

    V. The “Bimodal” Premise Must First Pass Its Own Test

    “How far from the mode” implies a morphological judgment. Bimodality suggests a natural boundary exists, making binary splitting logical, but morphological judgments are fragile.

    Three common metrics (bimodal coefficient BC, Hartigan dip, and AIC difference for single vs. two components) have been systematically compared (PMID 22806703): distance affects overall strength more than proportion, while the effect of proportion varies drastically across metrics; skewness can artificially inflate bimodality detection, causing anomalous interactions in certain cases; AIC difference is extremely sensitive, almost classifying all distributions as bimodal. BC and dip are generally consistent but have several important divergences.

    The implication is: a skewed unimodal distribution is likely to be read as bimodal, leading one to think a bimodal line should be used, when in fact the data has only one peak from start to finish. Placing morphology at the very top of the reasoning chain equals using a judgment with limited discriminative power to support all conclusions.

    VI. Replace the Horizontal Axis with Rare-Side Count $k$

    The point of intervention: percentiles are difficult to use because they conflate distance and sample size; replacing the horizontal axis with the count $k$ of samples above the threshold separates the two.

    The same 15% empirical line: with $n=100$, it is 15 people; with $n=20$, only 3 remain. The former yields a stable cutoff with roughly seven or eight people in each half; the latter leaves only one or two per half, meaning Jaccard’s numerator and denominator are determined by these few observations. This explains why the same empirical line performs so differently across studies—it is not the same threshold.

    Replacing with $k$ yields a testable inference: if stability is primarily determined by this, then the “$k \to$ Jaccard” curves for the four distribution shapes should converge.

    VII. Inversion and Four Criteria

    The inversion is: do not ask “how far from the mode should the threshold be,” but rather “how many samples must remain on the rare side to suffice.”

    Four criteria: if the maximum difference among the four shapes after switching to $k$ is $<0.05$, then percentiles and morphology are reduced to pseudo-variables, and the rule is rewritten as “rare-side $k \ge X$”; if they still clearly differ by shape ($>0.05$) after switching to $k$, then shape truly plays a role and lines must be drawn by shape, but morphological judgment must first undergo skewness correction and report all three metrics concurrently; if $k$ is fixed but Jaccard still declines with $n$, then the issue lies in within-individual correlation and splitting structure, returning to the split-unit line; if calculating confidence intervals for 70th and 85th percentile cutoffs reveals overlap, then any “top 15-20%” empirical line is unexecutable at this sample size, and only report $k$ and its actual count.

    Evidence and Limitations

    Verified: Data-driven selection of optimal cutoffs leads to overestimation of sensitivity and specificity, with bias amplified as sample size decreases ($n=40$ estimates true value of 84% as ~90%) (PMID 18258670); shape has little effect on this bias when sample size is fixed (same source); confidence interval coverage is insufficient after selecting a cutoff, requiring shrinkage with bootstrapping (PMID 15160403); binary splitting loses information, reduces power, and produces biased odds ratios (PMID 16217841, PMID 22733454, PMID 18389492, PMID 17066378); skewness artificially inflates bimodality detection with inconsistent conclusions across three metrics (PMID 22806703); precision of extreme percentile estimates is determined by sample size with existing algorithms for minimum subject numbers (PMID 25029084, PMID 30591808).

    Partially Verified: Distribution methods can provide both continuous and binary estimates consistently from the same original studies (PMID 24920271), but this is secondary analysis of published studies, not prospective validation; small-sample reference intervals require multiple concurrent methods (PMID 19473330), and large within-individual variability with strong skewness require caution (PMID 21731012), both being methodological suggestions rather than empirical tests of this parameter.

    To Be Verified: The “threshold percentile $\to$ Jaccard median” curve has never been plotted on the Dan Tian five parameters; the inflection point at the 80-85th percentile is merely a hypothesis awaiting testing; whether the four shape curves converge after switching to count $k$ is the core testable proposition of this paper, for which no data exists yet; whether confidence intervals for 70th and 85th percentile cutoffs actually overlap has also not been calculated.

    Related Reading

    Evidence Level: C
    References: PMID 18258670, PMID 15160403, PMID 16217841, PMID 22733454, PMID 18389492, PMID 17066378, PMID 24920271, PMID 22806703, PMID 25029084, PMID 30591808, PMID 19473330, PMID 21731012
    L1 Network Status: This topic has not yet entered the L1 corresponding network—the CORRESPONDENCE_NETWORK.md node table assigns nodes only by Yijing/TCM side, Physics side, and Physiological/Experiential side categories (original text “| Node | Concepts/Quantities in each domain (Five Phases, Five Zangs, Frequency, Color…) | Assigned by domain (see table below) |”), with no measurement dimensions nodes such as “selection threshold / cutoff / percentile position / distribution morphology judgment / selected set” in any of the three domains; the edge table stops at E-050 (“Six Heavenly Stems and Earthly Branchs Na Yin | Corresponds | 60 frequencies”, status open), with no edges originating or terminating at thresholds, cutoffs, or distribution shapes. EDGE_TABLE_science-yixue.md only contains concept counts (Dan Tian 51 articles / Qigong 48 articles / Breathing 189 articles / Frequency 407 articles), lacking L1/L2/L3 hierarchy and active/open/conflict status determination.


    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.

  • 「两态差值」自己测得准不准,没人验过:Δφ 的 MDC95 与等效界的可行性闸门

    Evidence level: C

    在两态趋同这件事上,先要问的不是「差值有多小」,而是「差值小到多少还能被分辨」——如果测量误差比等效界还宽,整条路线在数学上就不成立。

    一、整条路线压在一个没测过的量上

    前一篇把判据挂在「两态 φ 差值(Δφ)是否落在等效界内」上,但这个差值自身的测量误差,此前没人标定过。而差值分数有一个反直觉的性质:两个成分各自越一致,差值越不可靠。

    n=139 的 fMRI 实测里,两个条件各自的组内一致性都超过 0.94,两者之差却只有 −0.06,因为两条件相关高达 0.97(PMID 29458188)。同一现象在 ERP 上复现:增益和损失各自的内部一致性都在 0.70 以上,相减后掉到 0.36 以下(PMID 29057369);反应时任务的一周重测也显示,单条件均值信度可接受,而对比分数只有中等偏下(PMID 27720867)。

    换句话说,两把尺子各自准,不代表「两把尺子量出来的差」准。

    二、MDC 管的是「不是噪声」,不是「有用」

    MDC95 的算法是固定的:1.96 × √2 × SEM,那个 √2 就是因为一个差值含两次测量。它只回答「多大的变化超出测量误差」,不回答「多大的变化有生理意义」——后者必须由专业判断事先给定,且等效性检验的方法学要求先设界、再看数据(PMID 29135817)。

    两者差一个量级也不奇怪:某功能测试 ICC 高到 0.96、SEM 仅 0.59 秒,MDC95 仍有 1.62 秒(PMID 31928090)。更细的一层是 MDC 随分数水平变化:中段最小、两端最大,量表中段的临界值 3.6 分、两端可达 21.6 分(PMID 26740062)。如果经验者的 φ 恰落在分布低值端,用全样本平均的 MDC 就会低估误差。

    三、ICC 会替你把话说满

    ICC 是「个体间方差 ÷ 总方差」的比。测量误差完全不变时,样本越均质它越低,样本越悬殊它越高。481 项影像组学研究的系统综述把这层直接写进结论:高 ICC 可能只反映样本异质,并不保证测量精确(PMID 34603997)。方差齐性假设被违反时 ICC 还会被抬高,同一批数据换个模型从 0.821 掉到 0.295(PMID 30208858)。

    所以「在这批经验者里 ICC=0.8」不能外推给新手。更要命的是 ICC 与 Bland-Altman 界限会互相矛盾:同一批一致性数据,ICC 判「差到中等」、LoA 判「良到好」(PMID 20189765;另见 PMID 35184729,高一致时 LoA 仍可能很宽)。而小样本 ICC 尤其不牢:10 名受试者的研究里 ICC(3,1)=0.92,95%CI 却是 0.72–0.98(PMID 9688034)——区间宽到跨掉了整个判读档位。两个指标都报,缺一不可。

    四、φ 这条线上已知的三条坏消息

    其一,相位角的日间信度本来就弱。 OEP 标记法对照研究里,时间比与相位角的日间 ICC 只有 0.4–0.7,tI/tE、tI/tTot 等参数日间不可靠,唯独胸腹容积贡献比能到 0.9 以上(PMID 40732591)。

    其二,RIP 测出的 φ 个体内变异系数跨 11%–123%。 作者用机械模型把记录与分析软件的贡献压到 2.3% 以下、呼吸选取的贡献压到 8.1%,判定变异来自胸腹相位关系本身的不稳(PMID 24716708)。而 RIP 与 OEP 的 φ 一致性尚可(PMID 28767680),但那是同一天、同一条件下比出来的。

    其三,算法选型能一口吃掉整个等效界。 同一组 φ 数据,环形图法在部分情形误差超过 30°,互相关与最大线性相关法能压到 5° 以内(PMID 12422344)。

    五、反推样本量的公式要换一条

    若 MDC95 能测出来,样本量的反解方向就变了:它不该按组间效应量反解,而该按 MDC 反解。

    原因在差值方差的结构。两成分独立时,差值方差是成分方差的两倍;相关为 r 时收缩成 2σ²(1−r)。r=0.97 的情形下,差值 SD 只有成分 SD 的 0.24 倍——看着精度很高,其实信号同步缩小,信噪比未必改善。真正能自主调节的只剩两件事:重复测量次数(SEM 近似按 1/√k 下降),和遮蔽设计(把「知情」这一项系统偏倚从基线里拿掉)。

    六、这道闸门的判法

    把它落成一个可行性不等式:等效界必须同时大于 MDC95(否则差值永远判不进等效区),又小于预期的真实效应(否则任何差值都算「等效」,判据失去分辨力)。这个夹逼区间不空,路线才成立;为空,就说明当前测量手段在物理上撑不起这个命题。

    四格落点:MDC95 大于等效界 ⇒ 不可检验,降级为方法学报告并换终点(容积贡献比);MDC95 在等效界一半以内 ⇒ 可行,样本量按 MDC 反解;「知情−遮蔽」的静息 φ 差大于等效界三分之一 ⇒ 未遮蔽的静息基线一律不可用;差值 ICC 低于 0.5 而两个成分 ICC 高于 0.8 ⇒ 减法本身是错的,应改成分层混合效应模型,不做逐人相减。

    证据与局限

    已验证: 差值分数的信度低于其成分、且成分越相关塌得越狠,这是跨 fMRI、ERP、反应时三类任务的重复确认(PMID 29458188 / 29057369 / 27720867);MDC95 与 SEM 的换算关系、以及 MDC 不等于最小重要差异,属方法学共识(PMID 29135817 / 31928090);ICC 受样本异质性与方差齐性影响(PMID 34603997 / 30208858 / 9688034)。

    部分验证: OEP 相位角与时间比的日间信度弱、仅容积贡献比过 0.9,来自单一对照研究(PMID 40732591);φ 个体内 CV 达 11%–123% 来自早产儿样本(PMID 24716708),成人经验者的量级尚未实测;算法选型可致 30° 误差来自模拟加动物实验(PMID 12422344)。

    待验证: Δφ 的 MDC95 在成人丹田呼吸经验者身上到底是多少,目前没有任何实测值;圆形统计量(1−R、von Mises κ)与算术 MDC 在跨零 φ 上的换算关系未建立;「遮蔽−知情」差值本身是否大于等效界三分之一,属纯推测。

    L1 网络状态:CORRESPONDENCE_NETWORK.md 节点表(2.2 物理侧)仅列「相位 / 节律 —— 周期性耦合的关键」,2.3 生理侧仅列「呼吸率 / 呼吸模式」与「气感报告(部位/强度/时序/热麻胀沉浮)」,三者均未与任何易学/中医节点成边;边表 3.1–3.4 只登记五行↔五脏 / 五色 / 五音 / 五方 / 五季 / 五志、八卦↔五行、天干地支↔五行,无任何以 Δφ、测量误差、MDC / ICC 为源或目标的边。该主题尚未进入 L1 对应网络。

    相关阅读

    证据等级: C
    参考文献: PMID 29458188, PMID 29057369, PMID 27720867, PMID 29135817, PMID 31928090, PMID 26740062, PMID 34603997, PMID 30208858, PMID 20189765, PMID 35184729, PMID 9688034, PMID 40732591, PMID 24716708, PMID 28767680, PMID 12422344


    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。

    How Accurate Is the \”Two-State Difference\” Self-Measurement? Feasibility Gates for MDC95 and Equivalent Boundaries of Δφ

    Evidence level: C

    In the convergence of two states, the question is not “how small is the difference,” but “how small can the difference be while still being distinguishable”—if the measurement error exceeds the equivalence margin, the entire methodology fails mathematically.

    I. The Entire Route Rests on an Unmeasured Quantity

    The previous article anchored the criterion on “whether the two-state φ difference (Δφ) falls within the equivalence margin,” yet the measurement error of this difference itself has never been calibrated. Notably, difference scores possess a counterintuitive property: the more consistent each component is individually, the less reliable the difference becomes.

    In an n=139 fMRI study, within-subject consistency for both conditions exceeded 0.94, yet their difference was only −0.06 due to a correlation of 0.97 (PMID 29458188). This phenomenon replicates in ERP data: internal consistency for gain and loss individually exceeds 0.70, but drops below 0.36 upon subtraction (PMID 29057369); weekly retest reliability in reaction time tasks shows acceptable mean reliability for single conditions, while contrast scores remain only moderately low (PMID 27720867).

    In short, two accurate rulers do not may support that the “difference measured by two rulers” is accurate.

    II. MDC95 Governs “Not Noise,” Not “Usefulness”

    The MDC95 algorithm is fixed: $1.96 \times \sqrt{2} \times \text{SEM}$. The $\sqrt{2}$ accounts for the two measurements contained in a difference. It answers only “how large a change exceeds measurement error,” not “how large a change has physiological significance”—the latter must be predefined by expert judgment, and equivalence testing methodological requirements mandate setting boundaries before examining data (PMID 29135817).

    It is unsurprising that the two differ by an order of magnitude: a functional test with an ICC as high as 0.96 and an SEM of only 0.59 seconds still yields an MDC95 of 1.62 seconds (PMID 31928090). A finer layer involves MDC varying with score levels: minimum in the middle, maximum at the extremes; critical values of 3.6 points in the middle range can reach 21.6 points at the extremes (PMID 26740062). If an experiential φ falls precisely at the low end of the distribution, using the average MDC from the full sample will underestimate error.

    III. ICC Will Overstate Your Case

    ICC is the ratio of “between-individual variance ÷ total variance.” When measurement error remains constant, ICC decreases as samples become more homogeneous and increases as they become more heterogeneous. A systematic review of 481 imaging studies writes this directly into its conclusion: high ICC may only reflect sample heterogeneity and does not may support measurement precision (PMID 34603997). When the assumption of homoscedasticity is violated, ICC can be inflated; switching models on the same dataset can drop ICC from 0.821 to 0.295 (PMID 30208858).

    Therefore, an ICC of 0.8 in this specific group cannot be extrapolated to novices. More critically, ICC and Bland-Altman limits can contradict each other: the same consistency data may have ICC judging “moderate difference” while LoA judges “good to excellent” (PMID 20189765; see also PMID 35184729, where LoA can still be wide even with high consistency). Small-sample ICC is particularly unstable: in a study with 10 subjects, ICC(3,1)=0.92 yet the 95%CI is 0.72–0.98 (PMID 9688034)—the interval is wide enough to span an entire interpretation tier. Both metrics must be reported; neither alone suffices.

    IV. Three Known Bad News Items on the φ Line

    First, daily reliability of phase angles is inherently weak. In OEP marker comparison studies, the daily ICC for time ratios and phase angles ranges only from 0.4 to 0.7; parameters such as tI/tE and tI/tTot are unreliable on a daily basis, whereas thoraco-abdominal volume contribution ratios can reach above 0.9 (PMID 40732591).

    Second, RIP-measured φ shows an intraclass coefficient of variation spanning 11%–123%. Authors used mechanical models to suppress the contribution of recording and analysis software below 2.3% and respiratory selection below 8.1%, concluding that variability stems from instability in the thoraco-abdominal phase relationship itself (PMID 24716708). While consistency between RIP and OEP φ is acceptable (PMID 28767680), this comparison was conducted on the same day under identical conditions.

    Third, algorithm selection can swallow the entire equivalence margin. With the same set of φ data, the circular plot method yields errors exceeding 30° in some scenarios, whereas cross-correlation and maximum linear correlation methods can compress errors within 5° (PMID 12422344).

    V. The Sample Size Formula Must Change Direction

    If MDC95 can be measured, the inverse direction for sample size calculation changes: it should not be derived from between-group effect sizes but from MDC.

    The reason lies in the structure of difference variance. When two components are independent, difference variance is double the component variance; with correlation $r$, it contracts to $2\sigma^2(1-r)$. In the case of $r=0.97$, difference SD is only 0.24 times component SD—appearing high precision, yet the signal synchronously shrinks, and the signal-to-noise ratio may not improve. The only variables truly adjustable are repeated measurement counts (SEM approximates a $1/\sqrt{k}$ decline) and masking designs (systematically removing the “informed” bias from the baseline).

    VI. Judgment of This Gate

    Cast this into a feasibility inequality: the equivalence margin must simultaneously be greater than MDC95 (otherwise the difference can never fall into the equivalence zone) and less than the expected true effect (otherwise any difference counts as “equivalent,” rendering the criterion non-discriminatory). This squeezing interval must not be empty; only then does the route hold. If it is empty, it indicates that current measurement手段 cannot physically support this proposition.

    Four-point decision logic:
    – MDC95 > Equivalence Margin ⇒ Not testable; downgrade to methodological reporting and switch endpoints (volume contribution ratio).
    – MDC95 < Half of Equivalence Margin ⇒ Feasible; calculate sample size based on MDC.
    – Unmasked resting φ difference > 1/3 of Equivalence Margin ⇒ Unmasked resting baselines are universally unusable.
    – Difference ICC < 0.5 while component ICCs > 0.8 ⇒ Subtraction itself is incorrect; should adopt stratified mixed-effects models instead of individual subtraction.

    Evidence and Limitations

    Verified: The lower reliability of difference scores compared to their components, with sharper collapse as components correlate higher, is cross-confirmed across fMRI, ERP, and reaction time tasks (PMID 29458188 / 29057369 / 27720867); the conversion relationship between MDC95 and SEM, and that MDC ≠ Minimum Clinically Important Difference, are methodological consensus (PMID 29135817 / 31928090); ICC is affected by sample heterogeneity and variance homogeneity (PMID 34603997 / 30208858 / 9688034).

    Partially Verified: Weak daily reliability of OEP phase angles and time ratios, with only volume contribution ratios exceeding 0.9, comes from a single comparison study (PMID 40732591); φ intraclass CV of 11%–123% comes from preterm infant samples (PMID 24716708), and the magnitude for adult experienced practitioners remains unmeasured; algorithm selection causing 30° errors comes from simulation and animal experiments (PMID 12422344).

    To Be Verified: The actual value of Δφ MDC95 in adult dantian breathing experienced practitioners remains unmeasured; conversion relationships between circular statistics ($1-R$, von Mises $\kappa$) and arithmetic MDC across zero-crossing φ have not been established; whether the “masked-informed” difference itself exceeds one-third of the equivalence margin is pure speculation.

    L1 Network Status: The CORRESPONDENCE_NETWORK.md node table (Section 2.2 Physical Side) lists only “Phase/Rhythm — Key Periodic Coupling,” while Section 2.3 Physiological Side lists only “Respiration Rate/Pattern” and “Qi Sensation Reports (Location/Intensity/Timing/Thermal/Numbness/Swelling/Sinking/Rising).” None of these three connect to any Yijing/TCM nodes. Tables 3.1–3.4 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. There are no edges sourced or targeted by Δφ, measurement error, MDC, or ICC. This topic has not yet entered the L1 Correspondence Network.

    Related Reading

    Evidence Level: C
    References: PMID 29458188, PMID 29057369, PMID 27720867, PMID 29135817, PMID 31928090, PMID 26740062, PMID 34603997, PMID 30208858, PMID 20189765, PMID 35184729, PMID 9688034, PMID 40732591, PMID 24716708, PMID 28767680, PMID 12422344


    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.