Blog

  • 跨文件自相矛盾:两张 δ 表互相打架时,谁对谁错不用查参数就能判

    Evidence level: C

    一句话结论:同一个材料、同一个频点、同一张”厚度—衰减”表,本库里有两份正本给出了互不相容的数字。判它们打架不需要知道地下真实电导率是多少——只需要一条恒等式:A ∝ d。这是第 15 轮那条”同文件内互检”的跨文件版,也是它第一次真正抓到本库的主体缺陷形态。

    一、为什么”跨文件”比”同文件”重要

    第 15 轮做了一件很值钱的事:把 A(dB) ≈ 8.7·d/δ 里的 A ∝ d 抽出来,做成一条零参数检查——同一个文件里,同一材料同一频点下若出现两组”厚度—衰减”,两者之比必须相等。

    它判死了本库最硬的一处错(skin-depth-basement 里散文写「1 km 花岗岩 A ≈ 0.27 dB」、表格写「100 m 花岗岩 27 dB」,厚度差 10 倍而衰减差 100 倍)。

    但它报出的违反只有 1 处。原因不是本库干净,而是它的判据里有一条硬门槛:

    厚度至少差 2 倍,才值得比(更小的差不构成可判信号)
    

    这条门槛在数学上是对的(同厚度时比例检验退化),在工程上却把最硬的一类矛盾挡在门外:

    同厚度、不同 A —— 这根本不需要比例,只需要”相等”。而恰恰是这一类,在双语镜像 + 审计文互引的库里才是主体。

    二、这一轮把范围推到跨文件,先撞了两次墙

    把”同材料同频点”的配对范围从单文件扩到全库,第一次跑出来的违者是 32 对。这个数字不可信——按第 15 轮的纪律(每条新出现的违反都必须逐条归因才可采信),逐条看下来,32 对里绝大多数是假阳性,来源只有两个:

    墙一:审计文与台账被当成”主张方”。
    pool4-unparsed-table-classification-governance.md 里有一段并排写的是:

    散文:  1 km 花岗岩 → A ≈ 0.27 dB
    §3 表: 100 m 花岗岩 → A = 27 dB
    

    这两行本来就该不同——它是在引用一处矛盾,不是在主张一个数值。同理,backlog 里也抄了同一段。把它们与正本互比,等于把”有人在引用错值”当成”这个文件也错了”。

    墙二:散文里的材料名跨行串味。
    一份 SOP 里有两行相邻的列表:

    - 30 cm 混凝土墙 @ 50 Hz:δ ≈ 80 cm → A ≈ 3.3 dB
    - 2 mm μ-metal @ 50 Hz:δ ≈ 0.5 mm → A ≈ 35 dB
    

    第一行的 A ≈ 3.3 dB 被工具认成了 μ-metal——因为”最近材料词”是按字符距离算的,而换行不是边界,下一行的 μ-metal 字符反而更近。于是在”同厚度 2 mm”下,凭空多出一条「μ-metal 3.3 vs 35 dB,差 10.6 倍」的假矛盾。中英双语镜像各中一次,共 4 条假阳性。

    两道墙都堵上之后,32 对收敛到 6 对。 这是本轮第二次验证”放宽匹配必须同步收紧排除”——第 14 轮踩了三次、第 15 轮踩了三次,本轮再踩两次。

    三、留下的 6 对,全部指向同两处伤口

    # 材料 @ 频点 厚度比 A 比 偏差 A 侧 B 侧
    1 花岗岩 @ 7.83 Hz 10× 100× 10× mountain-elf-typical-granite-transparent 1000 m / 0.27 dB skin-depth-basement L76 100 m / 27 dB
    2 μ-metal @ 7.83 Hz 20× 8.65× 0.43× skin-depth-basement L72 2 mm / 30 dB skin-depth-reference-table L131 0.1 mm / 3.47 dB
    3 钢板 @ 7.83 Hz 2× 5.46× 2.73× skin-depth-basement L73 5 mm / 83 dB skin-depth-reference-table L129 10 mm / 15.2 dB
    4 钢板 @ 7.83 Hz 5× 54.6× 10.9× skin-depth-basement L73 5 mm / 83 dB skin-depth-reference-table L130 1 mm / 1.52 dB
    5 花岗岩 @ 7.83 Hz 10× 180× 18× skin-depth-basement L76 100 m / 27 dB skin-depth-reference-table L125 1000 m / 0.15 dB
    6 花岗岩 @ 7.83 Hz 10× 100× 10× skin-depth-basement L76 100 m / 27 dB pool4-mountain-elf-typical-granite-transparent L19 1000 m / 0.27 dB

    看这张表有一个很清晰的信号:6 对里有 4 对的一侧是同一个文件、同一行——skin-depth-basement 的 L76「100 m 花岗岩 27 dB」。

    第 13–15 轮已经在同文件内判过这一行(它和本文自己的散文 0.27 dB 差 1000 倍)。本轮的意义是:换一个方向再判,还是它。它不只与自己的散文打架,还与另外两份正本(skin-depth-reference-table 0.15 dB、mountain-elf-typical-granite-transparent 0.27 dB)打架。

    而这两份正本彼此是相容的:0.15 dB vs 0.27 dB,差 1.8 倍,在 2 倍容差内。两条独立来源说”1 km 花岗岩衰减不到 1 dB”,一条说”100 m 花岗岩衰减 27 dB”——按 A ∝ d,100 m 应当只有 1 km 的十分之一,即 0.015–0.027 dB。

    结论不依赖任何 σ 取值:只要承认 A ∝ d,L76 就错了。而 A ∝ d 在 d ≪ δ 的近似下是恒等式,不依赖 σ。

    四、必须说清的三条边界

    第一,这不证明”另外两份是对的”。 A ∝ d 只约束同一材料内部的相对关系,它没有说 0.15 dB 或 0.27 dB 哪个才是真值。两处都错的”错错相消”同样能通过比例检验——第 15 轮那张 μ-metal 行就是这样:它在逐格复算里判 OK,恰恰因为表的上下两节用了同一个错的 δ。

    第二,本检验只可证伪,不可验证。 比例对得上 ≠ 数值对。所以本轮的 6 条违反可以当作”至少有 6 处不相容”的证据;但“只有 6 处”不能当作质量指标——它只反映本次识别规则能看见多少(见下条)。

    第三,覆盖率必须报三项(R12 规矩):

    项 值
    ① 扫描根 780-网站与内容中枢(全库 2,403 个 md)
    ② 识别规则 等式断言的”厚度–A(dB)”点;同材料 + 同频点 + 同 μr;符号距离 ≤ 2×;转述/审计文不作主张方
    ③ 静默跳过 含可判点的 md 共 11 个,其中 5 个是转述/审计文被排除;正本仅 6 个文件 / 28 个点

    第三项是本轮最该被记住的数字:整个库能参与跨文件互比的,只有 28 个点、14 组配对。 这个规模小到什么程度?—— 6 条违反是从 14 组配对里出来的,违反率听起来很高,但样本量根本不足以推出任何”全库矛盾率”。这一点必须在写进任何结论前先说。

    五、顺带修掉的第 8 类失败模式:区间被压成左端点

    做跨文件配对时撞到一件更基础的事:工具把区间格压成了等式。

    parse_num_ex("1–10 dB")  →  (1.0, "1", "eq")
    

    也就是说,表格里写 A 1–10 dB(B 档范围)的格子,被当成 A = 1 dB 的等式断言送去比对。全库命中 δ/A 列的区间格共 42 处,其中 20 处明确属于”整格就是一个区间”。

    这类格不是”值算错了”,而是判读型归属错了——它表达的是档位范围,不是一个测量值。把它当等式判,等于凭空造出一条错误断言。

    修复方式刻意保守:不做区间端点判读,只做”不冒充等式”。识别出区间后,比较符记为 range,判读结果记为”不可判”,单独列一档 RANGE-unjudged,不进 MISMATCH。

    为什么不做双端点判读?因为 A 与 δ 的”好/坏方向”相反,且这里的区间表达的是档位而非测量不确定度。强行判端点会重演第 12 轮那件事:判了 A 档、下一轮又自我证伪改成 C 档。区间只可报”不可单点判”,不可报”通过”。

    修完后 MISMATCH 总数一条未变(24 + 34 + 196),OK 从 146 降到 138,差额 8 条正是从”被当成等式且恰好落在容差内”翻成”不可判”的那批。

    六、这一轮真正学到的

    教训一:门槛本身就是盲区。
    d_ratio ≥ 2 是上一轮为了让比例检验有信号而设的,它没错;但它把”同厚度不同 A”整类挡在门外。任何为了降噪而设的门槛,都会同时定义一类永远看不见的缺陷。 跨文件这一轮的第一件事,就是把门槛拆成两条:同厚度 → 判”相等”;不同厚度 → 判”成比例”。

    教训二:引用不是主张。
    在一个”审计文互相引用”的库里,把审计文当主张方,会把”有人在引用错值”读成”这个文件也错了”。32 → 6 的收敛,一半来自这一条。

    教训三:换行是语义边界。
    nearest_material 用字符距离是合理的,但它默认”文本是连续的”。列表里每行一个材料时,这个假设就破了。工具里所有”就近取”的逻辑,都要问一句:跨行还算不算同一个主语?

    七、证据与局限

    外部文献支撑的是公式与参数前提,不是本库的数字。

    • A(dB) ≈ 8.686·d/δ 与 δ = √(2/(ωμσ)) 是均匀导电半空间下的标准结果;其在岩石上的适用性受电导率频散限制:Kariya & Shankland (1983, Geophysics, doi:10.1190/1.1441407) 报告干燥下地壳岩石的电导率随频率变化;Laštovičková & Bochníček (1992, Stud. Geophys. Geod., doi:10.1007/BF01613900) 专门测量岩石电导率的频率依赖性;Revil (2013, Water Resour. Res., doi:10.1029/2012WR012700) 给出非饱和多孔介质有效电导率与介电常数在宽频段的模型;Mehran & Arulanandan (1977, Clays Clay Miner., doi:10.1346/CCMN.1977.0250107) 报告黏土—水—电解质体系的低频电导率频散。
    • 四篇均在 Crossref 回查通过。均为地物/材料期刊,不在 PubMed 收录范围 → 无 PMID → 如实标 E3,不构成 E1。

    局限(必须与结论同读):

    1. 本检验判”不相容”仅指文件内两处数字不满足 A ∝ d,不指真实地下 A 为另一值,也不指谁的 σ 引用错了。
    2. A ≈ 8.7d/δ 假设均匀半空间、无位移电流、μr/σ 为常数。真实山体分层、各向异性、σ 随频率弥散——上列四篇均支持频散存在,故公式本身在宽频段只是量级估算。
    3. 有效配对仅 14 组、正本仅 6 文件 28 点,不足以外推任何”全库矛盾率”。
    4. 只覆盖等式断言;不等式断言(A ≥ 10 dB)与文字区间(”数百米–数千米”)不参与。
    5. 转述/审计文被整体排除,意味着若某处矛盾只存在于审计文与正本之间、且正本本身不自相矛盾,本检验看不见。
    6. ASCII 图内的标注仍在盲区;本轮 skin-depth-basement L80 那条散文点被”同行材料”规则保留(材料与 A 值确在同一行),但同文件的 ASCII 图段仍未纳入。

    免责声明:本文是库内文本一致性的方法学审计,不是地球物理测量结论,不构成任何”靠山能否屏蔽 ELF”的实证判断,更不构成健康或医疗建议。任何据此做的采样决策,须先经 #14 阶段3 实测与 779 终版审核。

    八、下一步

    1. #64:把 R5 / R5X / R7 三条零参数检查合并成一张”断言充分性矩阵”,统一回答”这批断言里有几成是可判的”。
    2. D1′/D1/D2 单独提请 JOE:mountain-sop A 档门槛 σ ≥ 10⁻² → 1 km A ≥ 10 dB 的最不利端只有 4.83 dB,真实分界是 σ ≥ 4.29×10⁻²。这是唯一影响采样决策的一项,已连续 3 轮挂起。
    3. #59:F2 余 4 + F3 余 6 逐格复算(仍是 P2)。
    4. 不要做:把 R5X 扩成跨材料或跨 μr(物理上不成立,第 15 轮已立铁律);用”通过数”当质量指标(R11)。

    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。

    Cross-file Contradiction: When Two Delta Tables Clash, You Can Determine Who Is Right Without Checking Parameters

    Evidence level: C

    One-sentence conclusion: Two official tables within this repository for the same material, frequency, and “thickness–attenuation” relationship yield mutually incompatible numbers. Resolving this conflict does not require knowing the true subsurface conductivity—it relies solely on the identity A ∝ d. This is the cross-file iteration of the “intra-file cross-check” introduced in Round 15, and it is the first time this repository’s core structural defect has been definitively identified.

    I. Why “Cross-file” is More Critical Than “Intra-file”

    Round 15 performed a valuable operation: it extracted the A ∝ d proportionality from the formula A(dB) ≈ 8.7·d/δ to create a zero-parameter check. Within any single file, if two entries exist for the same material and frequency, the ratio of their thicknesses must equal the ratio of their attenuation values.

    This check definitively identified the repository’s most glaring error: in skin-depth-basement, prose states “1 km granite A ≈ 0.27 dB” while the table lists “100 m granite 27 dB”—a thickness ratio of 10 yielding an attenuation ratio of 100.

    However, this check reported only one violation. This is not because the repository is clean; rather, its criterion includes a hard threshold:

    Thickness must differ by at least 2× to be comparable (smaller differences do not constitute a detectable signal).
    

    Mathematically, this threshold is correct (the proportionality test degenerates when thicknesses are equal). However, engineering-wise, it excludes the most severe class of contradictions:

    Same thickness, different A — This fundamentally does not require proportionality; it requires “equality.” And precisely this class constitutes the majority of conflicts in a bilingual mirror + audit-text cross-reference repository.

    II. Extending the Scope to Cross-file: First Two Walls Collided

    Expanding the pairing scope from “same file” to “entire repository” yielded 32 violations. This number is unreliable—per Round 15 discipline (each newly detected violation must be individually attributed to be credible), most of the 32 pairs are false positives, originating from two sources:

    Wall One: Audit texts and ledgers mistaken for “claimants.”
    In pool4-unparsed-table-classification-governance.md, two lines appear side-by-side:

    Prose:  1 km granite → A ≈ 0.27 dB
    §3 Table: 100 m granite → A = 27 dB
    

    These two lines were meant to be different—they are citing a contradiction, not asserting a single value. Similarly, the backlog copied the same segment. Comparing them against official tables treats “someone citing a wrong value” as “this file is also wrong.”

    Wall Two: Material names crossing row boundaries in prose.
    One SOP contains two adjacent list items:

    - 30 cm concrete wall @ 50 Hz: δ ≈ 80 cm → A ≈ 3.3 dB
    - 2 mm μ-metal @ 50 Hz: δ ≈ 0.5 mm → A ≈ 35 dB
    

    The tool misidentified A ≈ 3.3 dB as belonging to μ-metal—because “nearest material word” is calculated by character distance, and newlines are not boundaries; the μ-metal text in the next row is closer. Under “same thickness 2 mm,” this created a phantom contradiction: “μ-metal 3.3 vs 35 dB, differing by 10.6×.” The bilingual mirror caught this once in each language, totaling 4 false positives.

    After blocking both walls, the 32 pairs converged to 6 pairs. This marks the second verification of “relaxing matching must simultaneously tighten exclusion”—Round 14 tripped three times, Round 15 tripped three times, and this round trips twice.

    III. The Remaining 6 Pairs All Point to Two Wounds

    # Material @ Frequency Thickness Ratio A Ratio Deviation Side A Side B
    1 Granite @ 7.83 Hz 10× 100× 10× mountain-elf-typical-granite-transparent 1000 m / 0.27 dB skin-depth-basement L76 100 m / 27 dB
    2 μ-metal @ 7.83 Hz 20× 8.65× 0.43× skin-depth-basement L72 2 mm / 30 dB skin-depth-reference-table L131 0.1 mm / 3.47 dB
    3 Steel plate @ 7.83 Hz 2× 5.46× 2.73× skin-depth-basement L73 5 mm / 83 dB skin-depth-reference-table L129 10 mm / 15.2 dB
    4 Steel plate @ 7.83 Hz 5× 54.6× 10.9× skin-depth-basement L73 5 mm / 83 dB skin-depth-reference-table L130 1 mm / 1.52 dB
    5 Granite @ 7.83 Hz 10× 180× 18× skin-depth-basement L76 100 m / 27 dB skin-depth-reference-table L125 1000 m / 0.15 dB
    6 Granite @ 7.83 Hz 10× 100× 10× skin-depth-basement L76 100 m / 27 dB pool4-mountain-elf-typical-granite-transparent L19 1000 m / 0.27 dB

    A clear signal emerges from this table: 4 of the 6 pairs involve the same file and line—skin-depth-basement L76 “100 m granite 27 dB.”

    Rounds 13–15 already judged this line within the same file (it differs by 1000× from its own prose value of 0.27 dB). The significance of this round is: judging from the opposite direction still yields the same result. It not only conflicts with its own prose but also with two other official tables (skin-depth-reference-table 0.15 dB and mountain-elf-typical-granite-transparent 0.27 dB).

    These two official tables are consistent with each other: 0.15 dB vs 0.27 dB differs by 1.8×, within the 2× tolerance. Two independent sources state “1 km granite attenuation is less than 1 dB,” while one states “100 m granite attenuation is 27 dB”—per A ∝ d, 100 m should be only one-tenth of 1 km, i.e., 0.015–0.027 dB.

    Conclusion does not depend on any σ value: as long as A ∝ d is accepted, L76 is wrong. And A ∝ d holds as an identity under the approximation d ≪ δ, independent of σ.

    IV. Three Boundaries That Must Be Clarified

    First, this does not prove “the other two are correct.” A ∝ d only constrains relative relationships within the same material; it does not specify which of 0.15 dB or 0.27 dB is the true value. Even “double errors canceling each other” can pass proportionality tests—this is exactly what happened with the μ-metal row in Round 15: it passed incremental recalculation precisely because both upper and lower sections of the table used the same erroneous δ.

    Second, this test can only falsify, not verify. Matching proportions ≠ matching values. Therefore, the 6 violations in this round serve as evidence of “at least 6 incompatibilities”; but “only 6” cannot serve as a quality metric—it merely reflects how many the current identification rules can detect (see next section).

    Third, coverage must report three items (per R12 rules):

    Item Value
    ① Scan root 780-Website and Content Hub (2,403 md files in full repository)
    ② Identification rule “Thickness–A(dB)” points of the equality assertion; same material + same frequency + same μr; symbol distance ≤ 2×; transcribed/audit texts excluded as claimants
    ③ Silent skip 11 md files containing judgable points, of which 5 were excluded as transcribed/audit texts; official tables cover only 6 files / 28 points

    The third item is the most critical number of this round: only 28 points across 14 pairs in the entire repository can participate in cross-file comparison. How small is this scale?—6 violations emerged from 14 pairs, making the violation rate sound high, but the sample size is insufficient to infer any “full-repository contradiction rate.” This must be stated before writing any conclusion.

    V. Incidentally Fixed Failure Mode #8: Intervals Compressed to Left Endpoints

    During cross-file pairing, we encountered a more fundamental issue: the tool compressed interval cells into equations.

    parse_num_ex("1–10 dB")  →  (1.0, "1", "eq")
    

    That is, table cells stating A 1–10 dB (B-grade range) were treated as the equation assertion A = 1 dB. Across the entire repository, 42 interval cells hit the δ/A column, of which 20 are explicitly “entire cell is an interval.”

    These cells are not “value calculation errors”; rather, judgmental attribution is wrong—they express grade ranges, not measurement values. Treating them as equations creates phantom assertions.

    The fix is deliberately conservative: do not interpret interval endpoints; only avoid pretending to be equations. After identifying intervals, the comparison symbol is marked range, the judgment result is “unjudgable,” and they are listed separately under RANGE-unjudged, not entering MISMATCH.

    Why not interpret both endpoints? Because A and δ have opposite “good/bad directions,” and here intervals express grades rather than measurement uncertainty. Forcing endpoint interpretation would repeat Round 12’s error: judging A-grade, then self-falsifying to C-grade in the next round. Intervals can only report “unjudgable at single points,” not “passed.”

    After the fix, MISMATCH total remained unchanged (24 + 34 + 196), OK dropped from 146 to 138, with the difference of 8 coming from those “mistakenly treated as equations but falling within tolerance.”

    VI. What This Round Truly Taught Us

    Lesson One: Thresholds are blind spots.
    The d_ratio ≥ 2 threshold was set in the previous round to give proportionality tests a signal; it is not wrong, but it excludes “same thickness, different A” entirely. Any threshold set to reduce noise simultaneously defines a class of defects that can never be seen. The first task of this cross-file round was to split the threshold into two: same thickness → judge “equality”; different thickness → judge “proportionality.”

    Lesson Two: Citations are not claims.
    In a repository where “audit texts cross-cite each other,” treating audit texts as claimants causes “someone citing a wrong value” to be read as “this file is also wrong.” The convergence from 32 to 6 is half due to this rule.

    Lesson Three: Newlines are semantic boundaries.
    nearest_material using character distance is reasonable, but it assumes “text is continuous.” When each list item represents a separate material, this assumption breaks. All “nearest-taking” logic in the tool must ask: does crossing lines still count as the same subject?

    VII. Evidence and Limitations

    External literature supports formulas and parameter premises, not the repository’s numbers.

    • A(dB) ≈ 8.686·d/δ and δ = √(2/(ωμσ)) are standard results for uniform half-space; their applicability to rock is limited by conductivity dispersion: Kariya & Shankland (1983, Geophysics, doi:10.1190/1.1441407) reports frequency-dependent conductivity of dry lower crust rocks; Laštovičková & Bochníček (1992, Stud. Geophys. Geod., doi:10.1007/BF01613900) specifically measured frequency dependence of rock conductivity; Revil (2013, Water Resour. Res., doi:10.1029/2012WR012700) provides a model for effective conductivity and dielectric constant of unsaturated porous media across wide bandwidths; Mehran & Arulanandan (1977, Clays Clay Miner., doi:10.1346/CCMN.1977.0250107) reports low-frequency conductivity dispersion of clay–water–electrolyte systems.
    • All four papers passed Crossref verification. All are earth/materials journals, not within PubMed scope → no PMID → correctly marked E3, not E1.

    Limitations (must be read alongside conclusions):

    1. This test judges “incompatibility” only when two numbers within the same file do not satisfy A ∝ d; it does not mean the true subsurface A is another value, nor that whose σ citation is wrong.
    2. A ≈ 8.7d/δ assumes uniform half-space, no displacement current, constant μr/σ. Real layered mountains, anisotropy, and σ dispersion with frequency—these four papers all support dispersion existence, so the formula itself is only an order-of-magnitude estimate across wide bandwidths.
    3. Only 14 valid pairs, 6 official files with 28 points, insufficient to extrapolate any “full-repository contradiction rate.”
    4. Only covers equality assertions; inequality assertions (A ≥ 10 dB) and textual intervals (“hundreds of meters to thousands of meters”) are excluded.
    5. Transcribed/audit texts are entirely excluded, meaning if a contradiction exists only between an audit text and an official table, and the official table itself is not internally contradictory, this test cannot see it.
    6. Annotations within ASCII diagrams remain in blind spots; this round’s skin-depth-basement L80 prose point was retained by the “same-row material” rule (material and A value indeed appear on the same line), but the ASCII diagram segment of the same file was not included.

    Disclaimer: This document is a methodological audit of textual consistency within the repository, not a geophysical measurement conclusion. It does not constitute any empirical judgment on “whether ELF can be shielded by mountains,” nor does it provide health or medical advice. Any sampling decisions based on this must first undergo #14 Stage 3 measurement and the final 779 review.

    VIII. Next Steps

    1. #64: Merge R5 / R5X / R7 three zero-parameter checks into a single “Assertion Sufficiency Matrix,” uniformly answering “what fraction of these assertions are judgable.”
    2. D1′/D1/D2 separately request JOE: mountain-sop A-grade threshold σ ≥ 10⁻² → 1 km A ≥ 10 dB has its worst endpoint at only 4.83 dB, while the true boundary is σ ≥ 4.29×10⁻². This is the only item affecting sampling decisions, and it has been pending for three consecutive rounds.
    3. #59: Incremental recalculation of remaining F2 (4) and F3 (6) cells (still P2).
    4. Do not do: Expand R5X across materials or μr (physically invalid, Round 15 established ironclad rules); use “passed count” as quality metric (R11).
  • 肯德尔 W 的「刻度」其实不存在——排序统计量的分辨率,不由它的可达值个数决定

    Evidence level: C

    W 只有 44 个可达值,听起来像一把只有 44 个刻度的尺子。但真正的麻烦不是刻度太粗,是这把尺子上刻的压根不是「份额差」这个量。

    正文

    一、先问一个更前面的问题

    前一篇已经说清:W 不显著不等于三个模态不一致,因为它只有 44 个可达值。顺理成章的下一步是问——那这 44 个格点,能分辨多小的份额差?如果连「体位 28% 对带位 27%」这一个百分点的差都分不出,那这套设计就不该用。

    但这个问题问法本身有陷阱。W 是秩排列的函数,不是份额差的函数。要判断它能不能分辨 δ,得先确认 δ 有没有进入它的计算。答案是:没有,一个数都没进。所以「W 的分辨率是多少」这句话,严谨地说没有答案——不是分辨率不够,是它不测这个量。

    下面把这两件事分开算清楚:格点有多粗(能算),δ 要多大多才能被检出(也能算,但要靠仿真),以及两者的关系(不是同一个数)。

    二、格点算术:能算的那一半

    W = 12S / [k²(n³−n)],S 是秩和离均差平方和。S 只能取整数,且可证为偶数,所以最小非零步长恒为

    ΔW_min = 24 / [k²(n³−n)]

    这个式子只含 n 和 k,与份额无关。n=5、k=3 时 ΔW_min = 0.0222,可达值 44 个;n=3、k=3 时 ΔW_min = 0.111,可达值只剩 6 个(0、0.111、0.333、0.444、0.778、1.0)。

    穷举 k=3、n=5 的全部 172.8 万种排序组合后,α=0.05 的精确临界落在 W=0.7111,实际 α=0.0455,只有 12 个格点能判显著。n=3 更极端:唯一显著值是 W=1.0,要求三个模态的排序完全一致,α 仅 0.0278——这是一个「全票或不算」的设计。

    三、δ 根本不在这条格点系上

    把两个极端数据摆一起:份额 (99%, 0.25%, 0.25%, 0.25%, 0.25%) 与 (21%, 20%, 20%, 20%, 19%),只要三个模态给出同一个排序,W 都等于 1.0。

    前者 δ 是 98.75 个百分点,后者 δ 是 1 个百分点——相差近百倍,W 一模一样。反过来,同一个小到可忽略的 δ 也可能让 W 从 0.6889 掉到 0.6667,仅因为某个模态把一个相邻名次换了。

    所以「44 个可达值」和「44 档份额差」不是一回事。前者是排列空间的一个商集,后者是连续单纯形上的距离。两者之间没有单调映射,更没有等距映射。W 的格点密度完全不携带关于 δ 大小的信息。

    四、瓶颈在秩变换这一步

    n 个对象的份额向量有 n−1 个自由度,是连续无穷多;排序只有 n! 种。n=5 时 120 种,k=3 个模态组合出 172.8 万种,W 再把它们压成 44 个值。

    信息不是被 W 压掉的,是被排序这一步压掉的。172.8 万 → 44 这个 3.9 万倍的压缩比看着惊人,但真正的坍缩发生在更前面:无穷 → 120。W 只是给这 120 种排列贴了个编号。

    这不是 W 的缺陷,是任何秩统计量的固有性质。代价有文献佐证:有序响应模型的渐近效率与检验力随类别数单调上升,合并类别(尤其二分成两类)损失信息最重(PMID 34547198);连续结局二分类的检验力损失从可忽略到相当大,取决于切点与均值差的关系(PMID 9598427);连续预测变量二分被认为是「一个坏主意」,还引入残余混杂(PMID 16217841);整群试验里 ICC 衰减带来的增益补偿不了信息损失(PMID 22733454)。

    五、δ_min:这一半只能靠仿真

    绕过解析、直接模拟:设真实份额 top1 领先 top2 一个 δ,评判者观测时叠加对数噪声 σ,n 个源、k 个模态,取 80% 检验力为门槛,结果是一张表。

    n=5、k=3、σ=0.05(评判者几乎完美)时 δ_min=0.05;σ=0.30(现实)时 δ_min=0.16。也就是要 16 个百分点才够。而源设计里体位与带位的真实差被设成 28% 对 27%,即 0.01——差 16 倍。

    n=3、k=3 更绝:σ=0.10 时 δ_min 就顶到 0.30,σ 再大则 80% 检验力在任何 δ 下都达不到。三个源三个模态,这套设计在物理上无法回答「谁最大」。

    六、加 n 比加 k 有用

    表的规律很干净。n 决定格点粗细(ΔW_min ∝ n⁻³),k 只负责压噪声(W 的期望是 1/k,k 增大零分布中心下移,临界值随之从 0.711 降到 0.550)。

    所以 n=5 时把模态从 3 个加到 4 个,σ=0.30 下 δ_min 从 0.16 降到 0.12,只改善 25%;而 n 从 5 加到 6(k 仍为 3),同样从 0.16 降到 0.12。两者打平——但加第 4 个模态要新仪器、新协议、新成本,多找一个变异源往往只是多设一个条件。

    这与少评判者场景的一般结论一致:方差分量分析的估计精度同时受评判者数与刺激数约束,仿真给出的下限是 60 个评判者、30 个刺激(PMID 31898288);少评判者时 ICC 与 CCC 的区间宽度问题尤其突出(PMID 24605969、PMID 25417040);多评判者一致性研究应按「预设区间精度」反解样本量,而不是按「够不够做检验」(PMID 22560852)。

    七、判据与做法

    三条判法:若按实测噪声仿真出的 δ_min 大于待分辨的真实差(如 0.16 对 0.01)⇒ 排序法在该设计下不可用,须改测连续量或直接报差值区间;若 δ_min 略小于真实差但余量不足 20% ⇒ 可做但必须预注册 δ 与 σ,且结论只报「检出/未检出」,不报完整排序;若 δ_min 明显小于真实差 ⇒ 排序法可用,但仍须报 ΔW_min 与临界在可达值中排第几位。

    配套两条:报告 W 时同时报 k²(n³−n)/24(即可达值个数),让读者知道尺子有多粗;把「最小可分辨份额差」写进方法学预注册,而不是事后算。

    证据与局限

    已验证: W 的可达值个数与最小步长由 ΔW_min = 24/[k²(n³−n)] 决定(本文穷举验证:n=3/k=3 为 6 个、n=4/k=3 为 17 个、n=5/k=3 为 44 个、n=5/k=4 为 78 个、n=6/k=3 为 77 个,与公式逐一符合);W 的零分布期望为 1/k,临界值随 k 下降(n=5 时 k=3 为 0.7111、k=4 为 0.5500);有序响应类别数越多、估计效率与检验力越高,二分类损失最重(PMID 34547198);连续变量二分的检验力损失随切点与均值差变化(PMID 9598427);二分连续预测变量引入残余混杂与偏倚(PMID 16217841);整群试验二分的检验力净损失(PMID 22733454);纵向数据二分的实质检验力损失(PMID 19904810);方差分量分析精度同时受评判者数与刺激数约束,仿真下限 60 评判者 / 30 刺激(PMID 31898288);少评判者时 ICC/CCC 区间方法差异显著(PMID 24605969、PMID 25417040);多评判者一致性研究应按区间精度反解样本量(PMID 22560852)。

    部分验证: δ_min 表为本文蒙特卡洛结果(每格 900 次),依赖设定的对数噪声结构与「其余 n−2 个源等距分布」这一份额构型,换构型数值会动;σ=0.05 至 0.50 这五档是作者设定,未对应任何实测的评判者误差分布;「加 n 与加 k 打平」只是 n=5,k=3 → n=6,k=3 与 n=5,k=4 两个点的比较,不是一般性结论。

    待验证: 丹田呼吸场景下三个模态(RIP/FOT/EIT)各自的份额观测噪声 σ 真实取值(无任何实测);体位、带位、操作员、个体、残差五个源的真实份额范围(无直接研究);评判者误差是对数正态还是别的形态;把 W 换成连续量的直接方差分量检验后,δ_min 能改善多少倍。

    L1 网络状态:「该主题尚未进入 L1 对应网络」——CORRESPONDENCE_NETWORK.md 逐词检索「排序 / 名次 / 份额 / 变异 / 刻度 / 分辨率 / 一致性」仅命中表 4.2 的指标行「生克一致性 | 相生/相克方向能否复现 | 未测」与节点表「心率变异性 HRV」中的「变异」二字,均非边;边表 21 条(E-001~E-005、E-010~E-014、E-030~E-034、E-040~E-044、E-050)全部为五行/五脏/五色/五音/五方/五季/五志/八卦/干支/纳音之间的对应关系,无任何以排序统计量、可达值、分辨率或份额差为源或目标的边。EDGE_TABLE_science-yixue.md 仅有概念文章计数(丹田 51 篇 / 气功 48 篇 / 呼吸 189 篇),无层级与状态判定。故本文封顶 C。

    相关阅读

    证据等级: C
    参考文献: PMID 34547198, PMID 9598427, PMID 16217841, PMID 22733454, PMID 19904810, PMID 31898288, PMID 24605969, PMID 25417040, PMID 22560852


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Kendall’s W has no \”scale\”—the resolution of rank statistics is not determined by the number of attainable values

    Evidence level: C

    W has only 44 reachable values, sounding like a ruler with only 44 marks. But the real trouble is not that the scale is too coarse; it is that this ruler does not measure “share difference” at all.

    Body

    I. First, Ask a More Fundamental Question

    The previous article clarified: W is not significant does not mean the three modalities are inconsistent, because it has only 44 reachable values. The next logical step is to ask—what is the smallest share difference these 44 grid points can resolve? If it cannot even distinguish a one-percentage-point difference between “posture 28% vs belt 27%”, then this design should not be used.

    But the question itself contains a trap. W is a function of rank order, not share difference. To determine whether it can resolve δ, one must first confirm whether δ enters its calculation. The answer is: no, not a single number does. Therefore, “What is the resolution of W?” strictly speaking has no answer—not because the resolution is insufficient, but because it does not measure this quantity.

    Below we separate these two issues clearly: how coarse the grid is (calculable), how large δ must be to be detected (calculable, but via simulation), and the relationship between them (not the same number).

    II. Grid Arithmetic: The Calculable Half

    $$W = \frac{12S}{k^2(n^3-n)}$$

    where $S$ is the sum of squared deviations of ranks from their means. Since $S$ can only take integer values and is provably even, the minimum non-zero step size is always

    $$\Delta W_{min} = \frac{24}{k^2(n^3-n)}$$

    This formula depends only on $n$ and $k$, not on share. When $n=5, k=3$, $\Delta W_{min} = 0.0222$ with 44 reachable values; when $n=3, k=3$, $\Delta W_{min} = 0.111$ with only 6 reachable values (0, 0.111, 0.333, 0.444, 0.778, 1.0).

    Exhausting all 1.728 million ranking combinations for $k=3, n=5$, the exact critical value at $\alpha=0.05$ falls at $W=0.7111$, yielding an actual $\alpha=0.0455$, with only 12 grid points able to judge significance. For $n=3$ it is even more extreme: the only significant value is $W=1.0$, requiring the rank order of all three modalities to be completely identical, with $\alpha$ only 0.0278—a “all-or-nothing” design.

    III. δ Is Not on This Grid at All

    Placing two extreme datasets together: shares (99%, 0.25%, 0.25%, 0.25%, 0.25%) versus (21%, 20%, 20%, 20%, 19%), as long as the three modalities yield the same ranking, W equals 1.0.

    In the former case δ is 98.75 percentage points; in the latter δ is 1 percentage point—a difference nearly 100-fold, yet W is identical. Conversely, an extremely small δ can cause W to drop from 0.6889 to 0.6667, simply because one modality swaps two adjacent ranks.

    Thus “44 reachable values” and “44 levels of share difference” are not the same. The former is a quotient set in ranking space; the latter is distance on the continuous simplex. There is no monotonic mapping between them, nor an isometric one. The grid density of W carries absolutely no information about the magnitude of δ.

    IV. The Bottleneck Lies in the Rank Transformation

    The share vector of $n$ objects has $n-1$ degrees of freedom, forming a continuous infinity; ranking has only $n!$ possibilities. For $n=5$ there are 120 such possibilities; combining $k=3$ modalities yields 1.728 million combinations, which W compresses into just 44 values.

    Information is not lost by W; it is lost by ranking. Compressing 1.728 million into 44—a ratio of 39,000—seems astonishing, but the true collapse happens earlier: infinity into 120. W merely assigns numbers to these 120 arrangements.

    This is not a flaw of W, but an inherent property of any rank statistic. Literature supports the cost: asymptotic efficiency and power of ordered response models increase monotonically with the number of categories, while merging categories (especially into two) loses information most heavily (PMID 34547198); power loss from dichotomizing continuous outcomes ranges from negligible to substantial, depending on the relationship between cutpoint and mean difference (PMID 9598427); dichotomizing continuous predictors is considered “a bad idea” and introduces residual confounding (PMID 16217841); in cluster trials, ICC decay cannot compensate for information loss (PMID 22733454).

    V. δ_min: This Half Can Only Be Simulated

    Bypassing analytic derivation, direct simulation: let the true share of top1 lead top2 by δ; evaluators observe with log-normal noise σ, $n$ sources and $k$ modalities, taking 80% power as the threshold. The result is a table.

    When $n=5, k=3, \sigma=0.05$ (evaluators nearly perfect), $\delta_{min}=0.05$; when $\sigma=0.30$ (realistic), $\delta_{min}=0.16$. Thus, 16 percentage points are required. Yet in the source design, the true difference between posture and belt is set at 28% vs 27%, i.e., 0.01—a factor of 16 smaller.

    For $n=3, k=3$ it is even more extreme: when $\sigma=0.10$, $\delta_{min}$ reaches 0.30; increasing σ further makes 80% power unattainable for any δ. With three sources and three modalities, this design physically cannot answer “who is largest.”

    VI. Adding n Is More Useful Than Adding k

    The pattern in the table is clean. $n$ determines grid coarseness ($\Delta W_{min} \propto n^{-3}$); $k$ only suppresses noise (the expected value of W is $1/k$, so as $k$ increases the zero-distribution center shifts down, and critical values drop from 0.711 to 0.550).

    Thus, increasing modalities from 3 to 4 when $n=5$ lowers $\delta_{min}$ from 0.16 to 0.12 under $\sigma=0.30$, improving only 25%; whereas increasing $n$ from 5 to 6 (keeping $k=3$) also lowers $\delta_{min}$ from 0.16 to 0.12. They tie—but adding a fourth modality requires new instruments, protocols, and costs, while finding an additional variance source often merely adds another condition.

    This aligns with general conclusions for fewer evaluators: estimation precision in variance component analysis is constrained by both evaluator and stimulus counts, with simulation yielding a lower bound of 60 evaluators and 30 stimuli (PMID 31898288); with fewer evaluators, interval width issues for ICC and CCC are especially prominent (PMID 24605969, PMID 25417040); multi-evaluator consistency studies should solve sample size based on “preset interval precision” rather than “whether enough for testing” (PMID 22560852).

    VII. Criteria and Practices

    Three judgment rules: if the simulated $\delta_{min}$ exceeds the true difference to be resolved (e.g., 0.16 vs 0.01) ⇒ ranking methods are unusable under this design; must measure continuous quantities or directly report difference intervals; if $\delta_{min}$ is slightly below the true difference but margin is less than 20% ⇒ it can be done but must preregister δ and σ, and conclusions should report only “detected/not detected,” not full ranking; if $\delta_{min}$ is clearly smaller than the true difference ⇒ ranking methods are usable, but still must report $\Delta W_{min}$ and the rank of the critical value among reachable values.

    Two supporting practices: when reporting W, also report $k^2(n^3-n)/24$ (the number of reachable values), so readers know how coarse the ruler is; include “minimum resolvable share difference” in methodological preregistration rather than calculating it post hoc.

    Evidence and Limitations

    Verified: The number of reachable values and minimum step size of W are determined by $\Delta W_{min} = 24/[k^2(n^3-n)]$ (exhaustive verification in this article: $n=3/k=3$ yields 6, $n=4/k=3$ yields 17, $n=5/k=3$ yields 44, $n=5/k=4$ yields 78, $n=6/k=3$ yields 77, all matching the formula); the expected value of W’s zero distribution is $1/k$, with critical values decreasing as $k$ increases (for $n=5$, $k=3$ gives 0.7111, $k=4$ gives 0.5500); more categories in ordered responses increase estimation efficiency and power, with dichotomization causing the greatest loss (PMID 34547198); power loss from dichotomizing continuous variables depends on the cutpoint and mean difference (PMID 9598427); dichotomizing continuous predictors introduces residual confounding and bias (PMID 16217841); dichotomization in cluster trials results in net power loss (PMID 22733454); longitudinal data dichotomization incurs substantial power loss (PMID 19904810); variance component analysis precision is constrained by both evaluator and stimulus counts, with simulation lower bounds of 60 evaluators / 30 stimuli (PMID 31898288); with fewer evaluators, ICC/CCC interval methods show significant differences (PMID 24605969, PMID 25417040); multi-evaluator consistency studies should solve sample size based on interval precision (PMID 22560852).

    Partially Verified: The $\delta_{min}$ table represents Monte Carlo results from this article (900 runs per cell), depending on the assumed log-normal noise structure and the “remaining $n-2$ sources equally spaced” share configuration; changing configurations would alter values; the five levels of $\sigma$ from 0.05 to 0.50 are author-set and do not correspond to any measured evaluator error distributions; “adding n ties with adding k” is merely a comparison of two points ($n=5,k=3 \to n=6,k=3$ and $n=5,k=4$), not a general conclusion.

    Pending Verification: the true values of share observation noise σ for each of the three modalities (RIP/FOT/EIT) in the Dantian breathing scenario (no measurement yet); the true share ranges for posture, belt, operator, individual, and residual sources (no direct study); whether evaluator error follows log-normal or another distribution; how much $\delta_{min}$ improves when replacing W with direct variance component testing on continuous quantities.

    L1 Network Status: “This topic has not yet entered the L1 corresponding network”—CORRESPONDENCE_NETWORK.md word-by-word search for “ranking / rank / share / variance / scale / resolution / consistency” only hits the indicator row of Table 4.2 “Sheng-Ke Consistency | Can Sheng-Xiang/Sheng-Chi directions be reproduced | Not measured” and the node table “Heart Rate Variability HRV” containing the word “variance,” neither of which are edges; Edge Table 21 (E-001~E-005, E-010~E-014, E-030~E-034, E-040~E-044, E-050) all consist of correspondences among Five Phases / Five Organs / Five Colors / Five Tones / Five Directions / Five Seasons / Five Emotions / Eight Trigrams / Heavenly Stems and Earthly Branches / Na Yin, with no edges originating from or targeting rank statistics, reachable values, resolution, or share difference. EDGE_TABLE_science-yixue.md contains only concept article counts (Dantian 51 / Qigong 48 / Breathing 189), with no level or status determination. Hence this article caps at C.

    Related Reading

    Evidence Level: C
    References: PMID 34547198, PMID 9598427, PMID 16217841, PMID 22733454, PMID 19904810, PMID 31898288, PMID 24605969, PMID 25417040, PMID 22560852


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 「差值区间跨了零」不是方法分歧,是检验力换了件衣服穿

    Evidence level: C

    用「份额差的置信区间是否跨零」替代 Kendall W,看起来是换了一把更严的尺子;实际只换了读数方式,它量的仍是同一份信息,能不能分辨出来由一个与「W 对不对」无关的数决定。

    一、这条替代路线在验什么

    Kendall W 的毛病是:显著只说明排序不像随机,说不明名次可信。于是有了更聪明的替代——不排秩,直接算份额差。体位 28%、带位 27%,差 1 个百分点,给差值做 bootstrap 区间,看它跨不跨零。

    逻辑看着无懈可击。但有个东西被塞进了判据却没写出来:跨不跨零说的不是两法对不对,是这个设计的分辨力够不够。这两件事在数学上是同一件事。

    二、20% 的跨零率,就是 80% 的检验力

    设差值真值 δ、标准误 SE,d/SE 是标准化效应量。区间跨零当且仅当 |d̂|<1.96SE,故

    P(跨零) = Φ(1.96−d/SE) − Φ(−1.96−d/SE)

    检验力为 Φ(d/SE−1.96)。两式对解:跨零率恰为 20% 时 d/SE = 1.96+0.842 = 2.802,检验力恰 80.00%。这是恒等式而非近似——跨零率 = 1 − 检验力。所以把判据写成「跨零率 >20% 即两法不一致」,等价于说「设计检验力不到 80%」。判据以为自己在讲分歧,实际在讲功效。

    三、源设计的刻度:差 1 个百分点 vs 要 14 个百分点

    体位 28% 对带位 27%,δ=0.01。三模态 n=5 的方差分量设计,单份额 bootstrap SE 保守取 0.05,则 d/SE=0.20,跨零率 94.5%,检验力 3.9%。

    反解:让跨零率降到 20%,同 SE 下 δ 需 0.140,是实际的十四倍;把 SE 压到 0.03 也还要 δ=0.084。这一对 28/27 离「跨得出零」差一个数量级。

    四、必然触发的判据,不构成证据

    于是「W 显著而差值 CI 跨零的案例占比 >20%」这条判据,在源设计上会以接近 100% 触发而非 20%。原因不是 W 错了,是 d/SE 小于 1 时跨零率必然在 93% 以上——换 bootstrap、换解析式、换 2000 次重抽样,结论都一样。

    它把「什么都没测出来」包装成「方法有矛盾」。拆包装只要一步:报跨零率的同时报出 d/SE。没有这个分母,跨零率是个无含义的百分比。

    五、换的不是严度,是表达力

    那这条路白走了?不。两法同数据、同方差分量、同重抽样;W 是秩统计量、CI 是连续量线性组合,但在「δ=0 与否」上它们是同一检验的两种表述。故 CI 不可能更严——检验力几乎相同,因为有效信息是同一份。

    真正变的是信息类型:CI 给符号与量级,可预注册、可跨研究累加;W 只给一个换不回份额的 0–1 系数。这是表达力升级,值得做,但别当成分辨力升级。

    六、整条排序可信还要再抬门槛

    k 个模态要比 k−1 次。要求整条排序同时成立,两次区间须同时不跨零:联合覆盖率 k=3 时 0.95²=0.9025,k=5 掉到 0.8145。Bonferroni 把 z 抬到 k=3 的 2.241、k=5 的 2.498;同时满足「不假阳性」与「80% 检验力」所需 d/SE 升到 k=3 的 3.083、k=5 的 3.339。

    越是把「替代 W」做认真,对数据要求越高。源设计连 2.802 都差十四倍——现实结论只有一个:先扩样本。

    七、边界:差值正则,绝对份额不正则

    差值 δ 在 ilr 坐标下是线性组合,正则、可用 bootstrap,这点优于 W。但一旦问「体位份额本身是否大于零」,就踏进参数空间边界:真值为 0 时估计量分布退化,百分位 bootstrap 会塌向观测值(重抽样生成不出比样本更小的方差)。文献记录明确:零边界处的方差估计会让其余参数区间一并失真(PMID 23670939);按广义异质性统计量反解的区间覆盖率对,但宽到没有信息(PMID 27604952)。同为「区间」,一个能报一个不能报,区别全在被测的量是否正则在空间内部。

    证据与局限

    已验证:「跨零率 = 1 − 检验力」是恒等式;跨零率 20% 精确对应 d/SE=2.802、检验力 80.00%,已用正态分位数反解核验(误差 <1e-6)。联合覆盖率 0.95^(k−1) 与 Bonferroni 门槛均为解析计算。组成数据走 ilr 有明确文献(PMID 27450923、25411321),固定/可变总和路径差异有仿真比较(PMID 40247167)。

    部分验证:「跨零率不构成分歧证据」依赖 d/SE 可估,而 SE 由 bootstrap 给出,小样本与边界附近偏窄,d/SE 会偏大、跨零率被低估。此层无法用解析消掉,只能并列报告解析式、jackknife、参数 bootstrap 三种 SE(PMID 24605969 记录小评判者数下各法覆盖率差异明显)。多重比较下区间覆盖不足有直接记录(PMID 2049512)。「重叠区间不等于不显著」在独立两组已有量化规则(PMID 18991332、11800251、30119088),但配对/相关结构下的对应规则未建立——本篇用近似,不是定理。

    待验证:① 源设计真实 SE 没估过;② 「W 显著而 CI 跨零」在真实数据中占比多少,没跑过;③ 若把 d/SE≥2.8 设为前置闸门,会切掉多少本可做的分析,成本没算过。

    L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK.md 三域节点表(易学/中医侧、物理侧、生理/体验侧)登记的全是信号名——物理侧仅有「相位 / 节律 —— 周期性耦合的关键」,生理侧仅有「呼吸率 / 呼吸模式」「气感报告」等观测项,没有任何一栏登记「统计判定框架」这一层(区间覆盖率、检验力、联合覆盖、标准化效应量 d/SE 均无处挂载);边表 3.1–3.5 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」),无以「判定框架 / 检验力 / 覆盖率」为源或目标的边。

    相关阅读

    证据等级: C
    参考文献: PMID 27450923, PMID 25411321, PMID 40247167, PMID 23670939, PMID 27604952, PMID 2049512, PMID 18991332, PMID 11800251, PMID 30119088, PMID 24605969

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


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    The gap interval crossing zero is not methodological divergence but the validation force changing clothes.

    Evidence level: C

    Using “whether the confidence interval of the difference crosses zero” instead of Kendall’s W appears to be swapping for a stricter ruler; in reality, it only changes the reading method. It still measures the same information, and whether one can distinguish it depends on a number unrelated to “whether W is correct.”

    I. What This Alternative Route Validates

    The flaw in Kendall’s W is that significance only indicates the ranking is not random; it does not clarify how credible the ranks are. Thus, a smarter alternative emerged: do not rank, but directly calculate the difference in shares. With body posture at 28% and arm position at 27%, the difference is 1 percentage point. Perform a bootstrap interval on this difference and check if it crosses zero.

    The logic appears unassailable. But something was embedded in the criterion without being written out: crossing zero does not indicate whether the two methods are right or wrong; it indicates whether the design’s sensitivity is sufficient. These two things are mathematically the same.

    II. A 20% Crossing Rate Equals 80% Power

    Let the true difference be $\delta$, the standard error $SE$, and $d/SE$ the standardized effect size. The interval crosses zero if and only if $|\hat{d}| < 1.96 SE$, hence

    $$
    P(\text{crossing}) = \Phi(1.96 – d/SE) – \Phi(-1.96 – d/SE)
    $$

    Power is $\Phi(d/SE – 1.96)$. Solving both equations: the crossing rate is exactly 20% when $d/SE = 1.96 + 0.842 = \mathbf{2.802}$, and power is exactly 80.00%. This is an identity, not an approximation—crossing rate = 1 − power. Thus, writing the criterion as “crossing rate >20% implies methods disagree” is equivalent to saying “design power is less than 80%.” The criterion thinks it is talking about discrepancy, but actually talks about power.

    III. Scale of the Source Design: Difference of 1 vs. Required 14 Percentage Points

    Body posture at 28% versus arm position at 27%, $\delta=0.01$. In a three-modal design with $n=5$, the variance component for a single share bootstrap SE is conservatively taken as 0.05, so $d/SE = \mathbf{0.20}$, crossing rate 94.5%, and power 3.9%.

    Working backwards: to bring the crossing rate down to 20%, under the same SE, $\delta$ must be 0.140, fourteen times the actual value; even pressing SE to 0.03 still requires $\delta=0.084$. This pair of 28/27 is off by an order of magnitude from “crossing zero.”

    IV. A Criterion That Always Triggers Does Not Constitute Evidence

    Thus, the criterion “proportion of cases where W is significant but difference CI crosses zero >20%” will trigger at nearly 100% rather than 20% under the source design. The reason is not that W is wrong, but that when $d/SE < 1$, the crossing rate is inevitably above 93%—regardless of swapping bootstrap, analytical formulas, or resampling 2000 times.

    It packages “nothing was measured” as “methods contradict each other.” Unpacking it takes one step: report the crossing rate together with $d/SE$. Without this denominator, the crossing rate is a meaningless percentage.

    V. What Changed Is Not Stringency, But Expressiveness

    So this route was in vain? No. Both methods use the same data, same variance components, same resampling; W is a rank statistic, CI is a linear combination of continuous quantities, but regarding “whether $\delta=0$,” they are two expressions of the same test. Hence CI cannot be stricter—power is almost identical because effective information is the same.

    What truly changes is the type of information: CI provides sign and magnitude, allowing pre-registration and cross-study accumulation; W only provides a 0–1 coefficient that cannot be converted back to shares. This is an upgrade in expressiveness, worth doing, but do not mistake it for an upgrade in sensitivity.

    VI. The Entire Ranking Credibility Must Raise the Threshold

    For $k$ modalities, one must compare against $k-1$. Requiring the entire ranking to hold simultaneously means both intervals must not cross zero: joint coverage for $k=3$ is $0.95^2 = \mathbf{0.9025}$, dropping to 0.8145 for $k=5$. Bonferroni raises the $z$ threshold to 2.241 for $k=3$ and 2.498 for $k=5; simultaneously satisfying “no false positives” and “80% power” requires $d/SE$ to rise to 3.083 for $k=3$ and 3.339 for $k=5$.

    The more seriously one treats “replacing W,” the higher the data requirements. The source design is off by fourteen times from 2.802—the real conclusion is only one: expand the sample size first.

    VII. Boundary: Difference Is Regular, Absolute Share Is Not

    The difference $\delta$ is a linear combination in ilr coordinates, regular and usable with bootstrap, an advantage over W. But once asking “whether the body posture share itself is greater than zero,” one steps into the parameter space boundary: when the true value is 0, the estimator distribution degenerates, and percentile bootstrap collapses toward observed values (resampling cannot generate variance smaller than the sample). Literature records clearly: variance estimation at the zero boundary causes distortion of all other parameter intervals (PMID 23670939); interval coverage solved via generalized heterogeneity statistics is wide but uninformative (PMID 27604952). Both are “intervals,” one can report, the other cannot; the difference lies entirely in whether the measured quantity is regular within the space.

    Evidence and Limitations

    Verified: “Crossing rate = 1 − power” is an identity; crossing rate 20% precisely corresponds to $d/SE=2.802$ and power 80.00%, verified via normal quantile inversion (error <1e-6). Joint coverage $0.95^{k-1}$ and Bonferroni thresholds are analytical calculations. Compositional data via ilr has clear literature (PMID 27450923, PMID 25411321); fixed/variable sum path differences have simulation comparison (PMID 40247167).

    Partially Verified: “Crossing rate does not constitute discrepancy evidence” depends on $d/SE$ being estimable, while SE is given by bootstrap; small samples and boundaries yield narrow SE, causing $d/SE$ to be overestimated and crossing rate underestimated. This layer cannot be eliminated analytically; only reporting analytical, jackknife, and parametric bootstrap SEs together (PMID 24605969 records clear coverage differences under small judge counts). Multiple comparison leads to insufficient interval coverage (PMID 2049512). “Overlapping intervals do not mean non-significance” has quantified rules for independent groups (PMID 18991332, PMID 11800251, PMID 30119088), but no corresponding rules have been established for paired/correlated structures—this paper uses approximation, not theorem.

    To Be Verified: ① the true SE of the source design has never been estimated; ② what proportion of real data shows “W significant but CI crosses zero” has not been run; ③ if setting $d/SE \geq 2.8$ as a prerequisite gate, how much analysis would be cut off and what cost has not been calculated.

    L1 Network Status: This topic has not yet entered the L1 corresponding network. The three-domain node table in CORRESPONDENCE_NETWORK.md (Yijing Studies/TCM side, physics side, physiology/experience side) registers only signal names—physics side has only “phase / rhythm — key of periodic coupling,” physiology side has only “respiration rate / respiration pattern,” “qi sensation report” and other observational items, no column registers the layer of “statistical decision framework” (interval coverage, power, joint coverage, standardized effect size $d/SE$ have nowhere to attach); edge table 3.1–3.5 stops at E-050 (“60 Heavenly Stems and Earthly Branches Na Yin | Correspondence | 60 Frequencies”), level L1, status open, modern evidence column original text “no measurable source yet,” with no edges sourced or targeted by “decision framework / power / coverage.”

    Related Reading

    Evidence Level: C
    References: PMID 27450923, PMID 25411321, PMID 40247167, PMID 23670939, PMID 27604952, PMID 2049512, PMID 18991332, PMID 11800251, PMID 30119088, PMID 24605969

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


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 「腹式化」和「协调化」这两个词,是解释还是命名:4 个指标撑不起 2 个构造的算术真相

    Evidence level: C

    一个只有 1 个自由度的检验,装不下「要不要拆成两个构造」这个问题。

    正文

    上一轮问的是「12 格够不够判方向」,这一轮问得更靠前一层:%AB、CoV、φ、X5 这四个参数,到底该算 1 个构造,还是 2 个构造。 设想的研究很干净——大样本横断面同步采四个参数,做二因子模型,看它是否显著优于单因子。但这套设计拆成算术就露馅:四指标、每构造两个,自由度只剩 1 个。一个自由度,要去回答一个关于「世界上有几个东西」的问题。 这不是样本量不够,是模型装不下问题。

    先把账算清。四个指标提供 4×5/2=10 个已知量。二因子相关模型需要估:2 个载荷(另 2 个固定为 1 以定标)、4 个误差方差、2 个因子方差、1 个因子间相关,共 9 个。10−9=1。单因子版本则是 3 个载荷加 4 个误差加 1 个因子方差,共 8 个,df=2。所以「二因子 vs 单因子」的 χ² 差检验只有 1 个自由度:它只能答一道是非题,而且答错的方式不对称。

    不对称在哪里?二指标因子的识别规则有个附加条件:该因子必须与模型中另一个因子有非零相关。翻成大白话——只有当两个构造确实相关时,这个模型才估得出来。若真答案是「两者无关」,每个因子只剩两个指标、相关为零,模型当场经验不可识别。而纵向 CFA 的通用要求本来就写明:每个潜变量至少三个指标才可识别,两个指标的模型「不可估」(PMID 41511670)。于是这个检验只在中度相关那一段能跑通:真值靠近 0 时估不出来,靠近 1 时又出不当解。它恰好只在你不必问的区间里工作。

    真答案是「一个构造」时,二因子模型也不会礼貌地失败。它给出的是不当解:负误差方差、载荷大于 1、或者干脆不收敛——而软件报的是「估计问题」,不是「你的二因子假设被推翻」。后果是研究者去修它:删掉一个指标、让两个误差项相关、或者换个参数化。代谢综合征那组八指标三模型就是活教材:青春期模型直接不收敛,其余阶段反复出 Heywood case,删了离群值也不解决(PMID 19732562)。更要紧的一条是:若真相关矩阵本身就蕴含不当解,加大样本只会让它更常出现(PMID 36875295)。所以「再招点人」这条路,方向是反的。

    还有一层方法混淆。%AB 与 φ 都出自同一台 RIP、同一次佩戴、同一次校准;CoV 来自 EIT,X5 来自 FOT。也就是说,那个唯一自由度所检验的跨构造相关(%AB×φ),恰恰是同一方法内部的相关。按多质多法模型的结构,同方法相关只能往上加,不会往下减,它把「一个构造」这个方向推得更像。但反方也有硬证据:直接测过的研究没能支持「相同作答格式会显著虚高收敛效度」(PMID 30284857);而共同方法方差与模型设定错误在实践中常常分不开(PMID 10908073)。两边一起摆:这一格的相关不能当「构造重叠」的干净证据,可也不能一口咬定它是假的。

    检验力这一关同样过不去。按 n=120、1 个自由度算,若真模型与单因子的差距是 ΔRMSEA=0.10,检验力只有 19.5%;要到 80%,样本量得约 785。所以「χ² 不显著 → 判一个因子」在这个规模下等于没测。相关侧更直白:n=120 时 Fisher z 区间半宽 0.181,真值 0.86 的 95% 区间上界正好压在 0.900;要宣称「相关低于 0.90」,真值至多 0.859。也就是说,n=120 只在共享方差已达 74% 时,才勉强说得上「没到 0.9」。

    出路不是加人,是加指标和加方法。把指标扩到 6 个(3+3,df=8),或者让每个构造各由两种方法测;并且预注册一条退化处理规则:出现不当解时不许删指标、不许放误差相关,必须原样记为「二因子不可估」。三格判据:① 六指标 3+3、Δχ² 显著且两因子相关的 95% 区间上界低于 0.85 → 拆成两构造成立,「三模态同向」命题废止,改为分别验证;② 区间含 1.0 或出现不当解 → 不许宣布「一个构造」,只能写「在当前指标数下不可分辨」,转做单构造内部的多方法验证;③ Δχ² 不显著但检验力低于 0.5 → 记为「未检出」,不得记为「单因子被支持」。须自认:0.85 与 0.5 两个门槛无出处,须由本队列的等效界反解替换。另外,即便建成二因子,能不能比均值还另说——标量不变性在实践中只是可逼近的理想(PMID 28080078),而违反它会让潜在的组间差异虚高到 63% 的假阳性率(PMID 34997192);模型错配时常用拟合指标往往也检不出来(PMID 26283977),双因子与二阶因子在杂质多时同样难分(PMID 29230085、PMID 35677113)。

    证据与局限

    已验证:每个潜变量至少三个指标才可识别,两指标模型不可估(PMID 41511670);四指标二因子相关模型的自由度为 1 而单因子为 2,Δχ² 只有 1 个自由度,是我们自算的代数结论;大型因子模型反复出现不收敛与 Heywood case,且删离群值无效(PMID 19732562);样本量增大不能消除不当解、反而增加其出现概率(PMID 36875295);模型错配时指标常检不出(PMID 26283977);标量不变性是理想而非现状(PMID 28080078),违反它可使潜在组间差异的假阳性率达 63%(PMID 34997192)。

    部分验证:共同方法方差会虚高相关的结构性推论成立,但一项直接检验未发现有意义增幅(PMID 30284857),与共同方法方差和设定错误难分(PMID 10908073)一起看,只能判「方向可能被推、幅度未定」。双因子与二阶因子的可分辨性随未建模复杂度上升而下降(PMID 29230085、PMID 35677113),提示任何「拆成两个」的胜利都可能只是模型更松。

    待验证:① 丹田四参数的跨构造相关里有多大来自 RIP 这一共同方法,从未拆过;② 六指标 3+3 方案下四个参数的因子归属是否稳定;③ 「腹式化」「协调化」两个词是否只是给两组参数起了名字,而非发现了两个机制。

    L1 网络状态:边表止于 E-050(「六十甲子纳音 ↔ 60 个频率」,L1、open,现代证据栏原文「尚无任何可测来源」);节点表生理侧仅登记「呼吸率 / 呼吸模式|可穿戴」,无「腹式化 / 协调化」等构造字段,全表逐词检索「构造 / 因子 / 效度 / 不变 / 多质」均 0 命中——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 41511670, PMID 19732562, PMID 36875295, PMID 26283977, PMID 30284857, PMID 10908073, PMID 28080078, PMID 34997192, PMID 29230085, PMID 35677113


    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。

    \”Abdominalization\” and \”Coordination\”: The Arithmetic Truth of Two Constructs Supported by Only Four Indicators

    Evidence level: C

    A test with only 1 degree of freedom cannot accommodate the question of “whether to split into two constructs.”

    Main Text

    The previous round asked whether “12 cells are enough to judge direction”; this round asks a layer deeper: whether %AB, CoV, φ, and X5 should be calculated as 1 construct or 2 constructs. The envisioned study is clean—large-sample cross-sectional simultaneous measurement of four parameters, two-factor model, testing whether it significantly outperforms the single-factor. But this design, when broken down arithmetically, reveals the flaw: four indicators, two per construct, leaves only 1 degree of freedom. One degree of freedom is insufficient to answer a question about “how many things exist in the world.” This is not a matter of insufficient sample size; it is a matter of the model being unable to accommodate the question.

    Let us balance the accounts. Four indicators provide 4×5/2 = 10 known quantities. A two-factor correlation model requires estimation of: 2 loadings (the other 2 fixed at 1 for scaling), 4 error variances, 2 factor variances, and 1 inter-factor correlation, totaling 9. 10 − 9 = 1. The single-factor version has 3 loadings plus 4 errors plus 1 factor variance, totaling 8, df = 2. Thus, the χ² difference test for “two-factor vs single-factor” has only 1 degree of freedom: it can answer only one yes/no question, and the way it fails is asymmetric.

    Where does the asymmetry lie? The identification rule for two-indicator factors has an additional condition: the factor must have non-zero correlation with another factor in the model. In plain terms—only when two constructs are indeed correlated can this model be estimated. If the true answer is “they are unrelated,” each factor has only two indicators, correlation zero, and the model becomes empirically unidentified. The general requirement for longitudinal CFA explicitly states: each latent variable must have at least three indicators to be identified; a two-indicator model is “unestimable” (PMID 41511670). Thus, this test works only in the moderate correlation range: it cannot estimate when the true value is near 0, and produces improper solutions when near 1. It operates exactly in the interval you do not need to ask about.

    When the true answer is “one construct,” the two-factor model does not politely fail. It yields improper solutions: negative error variances, loadings greater than 1, or outright non-convergence—while the software reports “estimation problems,” not “your two-factor hypothesis was refuted.” Consequently, researchers attempt to fix it: remove an indicator, correlate two error terms, or change parameterization. The metabolic syndrome set of eight indicators and three models serves as a textbook example: the puberty model failed to converge entirely, while other stages repeatedly produced Heywood cases; removing outliers did not resolve it (PMID 19732562). More critically: if the true correlation matrix itself implies improper solutions, increasing the sample size only makes them more frequent (PMID 36875295). Thus, “recruiting more people” is the wrong direction.

    There is also methodological confusion. %AB and φ both originate from the same RIP, same wearing session, same calibration; CoV comes from EIT, X5 from FOT. That is, the only degree of freedom testing cross-construct correlation (%AB×φ) is actually correlation within the same method. According to the structure of multitrait-multimethod models, same-method correlations can only increase, never decrease; it pushes the “one construct” direction even more strongly. Yet counter-evidence exists: direct measurement studies failed to support “same response format significantly inflates convergent validity” (PMID 30284857); and common method variance and model misspecification are often inseparable in practice (PMID 10908073). Putting both sides together: this cell’s correlation cannot be taken as clean evidence of “construct overlap,” nor can it be definitively declared false.

    Power is also insufficient. With n = 120 and 1 degree of freedom, if the true difference between models is ΔRMSEA = 0.10, power is only 19.5%; to reach 80%, sample size must be approximately 785. Thus, “χ² not significant → conclude one factor” in this scale is equivalent to no measurement. The correlation side is more direct: at n = 120, Fisher z interval half-width is 0.181; the upper bound of the 95% interval for true value 0.86 is exactly 0.900; to claim “correlation below 0.90,” true value must be at most 0.859. That is, n = 120 can only weakly claim “not reaching 0.9” when shared variance is already 74%.

    The solution is not adding people, but adding indicators and methods. Expand indicators to 6 (3+3, df = 8), or have each construct measured by two methods; and preregister a degradation rule: when improper solutions occur, do not remove indicators, do not correlate error terms; record them as “two-factor unestimable.” Three-cell criteria: ① Six indicators 3+3, Δχ² significant and upper bound of 95% CI for two-factor correlation < 0.85 → splitting into two constructs holds, “trimodal directionality” hypothesis abolished, replaced by separate validation; ② Interval contains 1.0 or improper solutions appear → cannot declare “one construct,” must write “unresolvable under current indicator count,” proceed with multimethod validation within single construct; ③ Δχ² not significant but power < 0.5 → record as “not detected,” do not record as “single-factor supported.” Must acknowledge: thresholds 0.85 and 0.5 have no source; must be replaced by equivalent bounds derived from this cohort. Furthermore, even if two-factor is established, whether it exceeds mean invariance remains uncertain—scalar invariance in practice is only an ideal approximation (PMID 28080078); violating it can inflate false positive rate of latent group differences to 63% (PMID 34997192); model misspecification often goes undetected by fit indices (PMID 26283977); two-factor and second-order factors are similarly hard to distinguish when complexity is high (PMID 29230085, PMID 35677113).

    Evidence and Limitations

    Verified: Each latent variable requires at least three indicators to be identified; two-indicator models are unestimable (PMID 41511670); four-indicator two-factor correlation model has df = 1 while single-factor has df = 2, Δχ² only 1 degree of freedom, which is our own algebraic conclusion; large factor models repeatedly show non-convergence and Heywood cases, and removing outliers is ineffective (PMID 19732562); increasing sample size does not eliminate improper solutions but increases their probability (PMID 36875295); model misspecification often goes undetected by indicators (PMID 26283977); scalar invariance is ideal rather than reality (PMID 28080078); violating it can inflate false positive rate of latent group differences to 63% (PMID 34997192).

    Partially Verified: Common method variance does inflate structural correlation, but one direct test found no meaningful increase (PMID 30284857); combined with common method variance and model misspecification being inseparable (PMID 10908073), we can only conclude “direction may be pushed, magnitude undetermined.” Distinguishability of two-factor and second-order factors decreases with unmodeled complexity (PMID 29230085, PMID 35677113), suggesting any “split into two” victory may merely reflect looser modeling.

    To Be Verified: ① How much of the cross-construct correlation among the four Dantian parameters originates from RIP, this common method, has never been decomposed; ② Under the six-indicator 3+3 scheme, whether factor assignment of four parameters remains stable; ③ Whether “abdominalization” and “coordination” are merely names given to two sets of parameters, rather than discoveries of two mechanisms.

    L1 Network Status: Edge table ends at E-050 (“Sixty Jiazi Yin-Ying ↔ 60 frequencies”, L1, open, original text of modern evidence column states “no measurable source yet“); node table physiological side only registers “respiration rate / respiration pattern | wearable”, no construct fields such as “abdominalization / coordination”; full table word search for “construct / factor / validity / invariant / multitrait” yields 0 hits—this topic has not yet entered the L1 corresponding network.

    Further Reading

    Evidence Level: C
    References: PMID 41511670, PMID 19732562, PMID 36875295, PMID 26283977, PMID 30284857, PMID 10908073, PMID 28080078, PMID 34997192, PMID 29230085, PMID 35677113


    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.

  • 「7/12」这句话什么都没说:方向投票的检验力下限,与 12 格设计在数学上够不够格

    Evidence level: C

    一个连「随机」和「高度一致」都分不开的区间,被拿来当三角验证的判据,等于掷骰子。

    一、本篇不查假设,只查尺子

    三模态同向这个假设本身有没有依据,上一篇查过——同一批早产儿数据里,相位角 60.6° 对 12.5° 分得极开,胸廓贡献比却 25.5% 对 36.3%、p=0.11 毫无差异(PMID 9387956)。那是问”三个模态会不会真同向”。

    本篇问下一层:假定它们真同向,你手上这套 3 中心 × 4 模态 = 12 格的投票法,有没有本事把它认出来。 两件事必须分开。假设可以千真万确,而检验它的设计烂到永远说不出话。方法学界把这点叫检验力——它不改变真值,只改变你能不能看见真值。而方向投票最容易漏算它,因为它看起来只是在数数。

    二、12 格不是 12 票,门槛在 10/12

    先算最基础的一笔。若真值是同向率 50%,12 格出 7 个同向纯属噪声。双侧精确二项检验下:7/12 的 p=0.77,8/12 的 p=0.39,9/12 的 p=0.15,要到 10/12 才第一次跌破 0.05(p=0.039)。

    故 12 格的最低门槛是 83%,不是直觉里的”过半就行”。这条线随格数变且不单调:10 格要 9/10,20 格要 15/20(75%),30 格要 21/30(70%)。格数越少,门槛比例越离谱——6 格和 8 格都要全票。因为二项分布在小 n 下方差极大,12 次抛硬币出 8 次正面毫不稀奇;可一旦把”8 次正面”翻译成”三分之二的器官同向变”,话就变重了,分布并没变。

    三、7/12 的真实含义:一个宽到无用的区间

    比 p 值更该看区间。7/12 点估计 58.3%,Clopper-Pearson 95% 区间是 [0.277, 0.848],宽 0.572。含义要说得直白:真值可能是 28%(基本随机),也可能是 85%(高度一致),数据都兼容。

    一个连”随机”和”高度一致”都分不开的区间,拿来当三角验证的判据,等于掷骰子。四种常见读法全不成立:读”失败”——排除不了 85%;读”部分成功”——兼容 28%;读”趋势”——趋势需方向性先验,投票法没有;读”补几格就好”——需要的是四倍以上格数,见第五节。

    四、三重共线:12 个格子里有几个是独立的

    12 格本身不独立,至少三重打折。

    模态之间:%AB、φ、X5、CoV 测的是同一批人同一套呼吸,共享受试者、共享时段、共享”他今天状态如何”。三项水平模型的理由正是这类依赖效应量(PMID 23055166);当四个独立证据用,等于一份数据抄四遍。

    中心之内:每个中心只有一批人。设计效应 1+(m−1)×ICC,m=4 时 ICC 取 0.3 得 1.9、取 0.5 得 2.5——12 格折成 6.3 格、4.8 格。折到 6 格后门槛不是 10/12,是 6/6 全票。

    方向之内:若真实同向率只有 0.7(已算不错),12 格达 10/12 的概率 25.3%;有效格折到 6.3 掉到 11.8%。八成情况下,真存在的同向信号被判”失败”。

    五、反过来一样糟:12/12 也只支持 73.5%

    假阳性那侧同样不能放心。12 格全同向看似无可挑剔,但全同向的 95% 区间下界只有 0.735(=0.025^(1/12))。你想说”高度一致”,数据只支持”至少 73.5% 同向”,且不含方向外的任何效应量信息。

    压到 0.90 下界需 65 格,0.927 需 49 格。3×4=12 与 49 差四倍——不是”多做一个中心”能填的坑,是设计量级不对。同样的账别人算过:Many Labs 2 常被引作”地点不影响可复现性”,事后功效模拟却发现它查文化异质性的检验力本就不够,所以”没查到”不等于”不存在”(PMID 39379734)。

    六、这个坑不是我们独有的

    100 篇公共卫生系统综述里,60 篇用了非 meta 合成,其中 54 篇用投票计数——只有 3 篇按方向投,36 篇按统计显著性投,15 篇没说清(PMID 36758885)。

    按显著性投更糟:它把”样本够不够”混进”效应在不在”。按方向投也不安全——平票概率不确定时,符号检验在 64 种设计条件里 46 种(71.9%)功效低于 80%,10 种低于 70%(PMID 20974291)。平票就是投不出方向的格子,多到能掏空设计。再加一层:小样本效应量系统性偏大(PMID 23302257),而方向投票给每项研究同等权重,等于把最不可靠的抬到和可靠的一样高。

    七、判据三格与做法

    第一格——设计不够格,结论作废。 若实测 ICC ≥ 0.3 或模态间相关 ≥ 0.5,有效格折到 6-7 格,12 格设计在任何真实同向率下功效均低于 30%——”7/12 判失败””12/12 判通过”都不得写入结论,只能报”未检出”并附有效格数与功效。

    第二格——够格但改口径。 若 ICC < 0.2 且相关 < 0.3、有效格 ≥ 9,可保留 12 格,但报告须三件齐备:门槛写清(10/12)、区间报出(含宽度)、平票单列(不并入任一侧)。缺一项即视为不完整。

    第三格——扩到够为止。 要坚持”三角验证通过”,按 p=0.7 达 80% 功效需 49 格、90% 需 65 格。做不到就明写”本研究不具备判定能力”,改报逐模态方向加各自区间,不做跨模态合并。

    做法:先用 ≥40 人同批数据估模态间相关矩阵与中心内 ICC,折出有效格,再按折后格数重算门槛与功效,最后落判。明确自认:”ICC 0.3″”相关 0.5″”功效 30%”三个数取自通用惯例、无出处,须由本队列实测反解替换。

    证据与局限

    已验证:12 格双侧精确二项检验的门槛是 10/12(p=0.039),7/12 的 p=0.77、区间 [0.277, 0.848] 宽 0.572,全同向下界 0.735——均为本次自算,公式与 Clopper-Pearson 法可复算;早产儿与足月儿对照中相位角 60.6° 对 12.5°(p<0.0001)、胸廓贡献比 25.5% 对 36.3%(p=0.11)确为同一批数据(PMID 9387956);依赖效应量须用三项水平模型处理(PMID 23055166);100 篇系统综述中 54 篇用投票计数、仅 3 篇按方向投(PMID 36758885);平票概率不确定时符号检验 71.9% 的设计条件功效低于 80%(PMID 20974291);小样本研究效应量系统性偏大(PMID 23302257);Many Labs 2 查文化异质性的检验力本就不足(PMID 39379734)。

    部分验证:上述方法学证据全部来自其他学科(心理学复现、重症 meta 分析、公共卫生综述),无一在气功三模态场景落地;设计效应公式本身是标准式,但 m=4、ICC=0.3-0.5 这两个取值是代入假设,不是实测。

    待验证:本方案三模态(%AB/φ/X5/CoV)的模态间相关矩阵与中心内 ICC 实测值——它直接决定有效格是 12 还是 5;平票(方向判不出)的真实发生率,以及该不该并入任一侧;把格数从 12 扩到 49 或 65 在经费与中心数上是否可行;门槛 83% 与”有效格 6-7″之间的判据线在实际数据上会不会出现中间态。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK 节点表三域登记的全是「信号名」(相位 / 节律、呼吸率 / 呼吸模式、气感报告),无「研究设计 / 检验力 / 样本单位 / 多中心」任何字段;逐词 grep「方向 / 一致性 / 检验力 / 功效 / 样本 / 多中心 / 中心 / 模态 / 投票 / 假阳性」除结构说明里的零星文字外全部 0 命中;边表止于 E-050,状态 open,现代证据栏为「尚无任何可测来源」。故本文证据等级封顶 C。
    证据等级: C
    参考文献: PMID 9387956, PMID 23055166, PMID 36758885, PMID 20974291, PMID 23302257, PMID 39379734


    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。

    7/12 Says Nothing: The Lower Bound of Power for Directional Voting and Whether the 12-Grid Design is Mathematically Qualified

    Evidence level: C

    An interval that cannot distinguish between “random” and “highly consistent” is used as the criterion for triangulation; this is equivalent to rolling dice.

    I. This Paper Does Not Test Hypotheses, It Tests the Ruler

    Whether the hypothesis of three-modal concordance has any basis was examined in the previous paper—the phase angle 60.6° versus 12.5° is extremely divergent, while the thoracic contribution ratio 25.5% versus 36.3% shows no difference (p=0.11) in the same cohort of preterm infants (PMID 9387956). That addressed whether “the three modalities truly move in the same direction.”

    This paper asks the next level: Assuming they do move in the same direction, does your current 3-center × 4-modalities = 12-grid voting method have the power to detect it. Two things must be separated. The hypothesis can be true, while the design to test it is so poor it cannot speak. Methodologists call this power—it does not change the truth, only whether you can see the truth. Directional voting most easily overlooks this because it appears to be merely counting.

    II. 12 Grids Are Not 12 Votes; the Threshold Is 10/12

    First, the most basic calculation. If the true concordance rate is 50%, getting 7 out of 12 concordant is pure noise. Under two-sided exact binomial testing: 7/12 yields p=0.77, 8/12 yields p=0.39, 9/12 yields p=0.15, reaching 10/12 is the first time it drops below 0.05 (p=0.039).

    Thus the minimum threshold for 12 grids is 83%, not the intuitive “more than half.” This line varies with grid count and is non-monotonic: 10 grids require 9/10, 20 grids require 15/20 (75%), 30 grids require 21/30 (70%). Fewer grids make the threshold proportion absurd—6 grids and 8 grids both require unanimous votes. Because binomial distribution variance is huge for small n, getting 8 heads in 12 coin tosses is not remarkable; but translating “8 heads” into “two-thirds of organs moving concordantly” changes the weight, while the distribution itself remains unchanged.

    III. The True Meaning of 7/12: An Interval Too Wide to Be Useful

    More than the p-value, one should look at the interval. The point estimate for 7/12 is 58.3%, but the Clopper-Pearson 95% interval is [0.277, 0.848], width 0.572. The meaning must be stated plainly: the true value could be 28% (basically random) or 85% (highly consistent); the data are compatible with both.

    An interval that cannot separate “random” from “highly consistent” used as the criterion for triangulation is equivalent to rolling dice. All four common readings fail: reading “failure”—cannot exclude 85%; reading “partial success”—compatible with 28%; reading “trend”—trends require directional priors, which voting lacks; reading “add a few grids”—requires four times more grids, see Section V.

    IV. Triple Collinearity: How Many of the 12 Grids Are Independent

    The 12 grids themselves are not independent; at least three layers of discount apply.

    Between modalities: %AB, φ, X5, CoV measure the same people breathing the same way, sharing subjects, sharing time periods, sharing “how he felt today.” The rationale for three-level models is precisely such dependency effect sizes (PMID 23055166); when four independent evidences are used, it equals copying one dataset four times.

    Within centers: Each center has only one batch of people. Design effect is 1+(m−1)×ICC; with m=4, ICC=0.3 gives 1.9, ICC=0.5 gives 2.5—12 grids fold to 6.3 grids, 4.8 grids. Folding to 6 grids makes the threshold not 10/12, but 6/6 unanimous.

    Within directions: If the true concordance rate is only 0.7 (already quite good), the probability of reaching 10/12 in 12 grids is 25.3%; effective grids fold to 6.3 drops this to 11.8%. In eight cases, the true concordant signal is judged “failed.”

    V. Conversely, 12/12 Also Only Supports 73.5%

    The false-positive side is equally unworthy of trust. All 12 grids concordant appears flawless, but the lower bound of the 95% interval for all-concordant is only 0.735 (=0.025^(1/12)). To claim “highly consistent,” the data only support “at least 73.5% concordant,” and contain no information about effects outside the direction.

    Reaching a lower bound of 0.90 requires 65 grids, 0.927 requires 49 grids.3×4=12 differs from 49 by four times—not something filled by “adding one more center,” but a mismatch in design scale. Others have calculated this: Many Labs 2 is often cited as “location does not affect replicability,” but post-hoc power simulation found its ability to detect cultural heterogeneity has insufficient power, so “not detected” does not mean “does not exist” (PMID 39379734).

    VI. This Pit Is Not Unique to Us

    Among 100 public health systematic reviews, 60 used non-meta synthesis, of which 54 used vote counting—only 3 voted by direction, 36 by statistical significance, 15 did not clarify (PMID 36758885).

    Voting by significance is worse: it mixes “sample size sufficient” into “effect present.” Voting by direction is also unsafe—when tie probability is uncertain, the sign test has power below 80% in 46 of 64 design conditions (71.9%), and below 70% in 10 conditions (PMID 20974291). A tie means no grid can determine direction, enough to empty the design. Add another layer: small-sample effect sizes are systematically inflated (PMID 23302257), while directional voting gives equal weight to each study, elevating the least reliable to the same level as the most reliable.

    VII. Criteria for Three Grids and Procedures

    First grid—design insufficient, conclusion void. If measured ICC ≥ 0.3 or inter-modal correlation ≥ 0.5, effective grids fold to 6–7; the 12-grid design has power below 30% under any true concordance rate—”7/12 judged failed,” “12/12 judged passed” cannot be written in conclusions, only report “not detected” with effective grid count and power.

    Second grid—sufficient but adjust criteria. If ICC < 0.2 and correlation < 0.3, effective grids ≥ 9, keep 12 grids but reporting must include three items: threshold clearly stated (10/12), interval reported (with width), ties listed separately (not merged to either side). Missing any item renders it incomplete.

    Third grid—expand until sufficient. To maintain “triangulation passed,” achieving 80% power at p=0.7 requires 49 grids, 90% requires 65 grids. If not possible, clearly state “this study lacks adjudication capacity,” report directional addition per modality with respective intervals, without cross-modal merging.

    Procedure: First estimate inter-modal correlation matrix and within-center ICC using ≥40 subjects from the same batch, fold to effective grids, then recalculate threshold and power based on folded grids, finally render judgment. Explicitly acknowledge: “ICC 0.3,” “correlation 0.5,” “power 30%” are three numbers taken from general convention without source, must be replaced by empirical measurement from this cohort.

    Evidence and Limitations

    Verified: The threshold for 12-grid two-sided exact binomial test is 10/12 (p=0.039), 7/12 yields p=0.77, interval [0.277, 0.848] width 0.572, all-concordant lower bound 0.735—all self-calculated here, formulas and Clopper-Pearson method can be re-verified; phase angle 60.6° versus 12.5° (p<0.0001), thoracic contribution ratio 25.5% versus 36.3% (p=0.11) are indeed from the same dataset of preterm versus term infants (PMID 9387956); dependency effect size must be handled via three-level model (PMID 23055166); among 100 systematic reviews, 54 used vote counting, only 3 by direction (PMID 36758885); when tie probability is uncertain, sign test power falls below 80% in 71.9% of design conditions (PMID 20974291); small-sample studies show systematic inflation of effect sizes (PMID 23302257); Many Labs 2 has insufficient power to detect cultural heterogeneity (PMID 39379734).

    Partially verified: All methodological evidence comes from other disciplines (psychology replication, intensive care meta-analysis, public health reviews); none has been implemented in the Qigong three-modal scenario; the design effect formula itself is standard, but m=4, ICC=0.3–0.5 are assumed values, not measured.

    To be verified: The inter-modal correlation matrix and within-center ICC for this three-modal scheme (%AB/φ/X5/Cov)—this directly determines whether effective grids are 12 or 5; the true incidence of ties (direction cannot be judged), and whether they should merge to either side; feasibility of expanding grids from 12 to 49 or 65 given budget and center count; whether the threshold line between 83% and “effective grids 6–7” appears as an intermediate state in actual data.

    Related Reading

    L1 Network Status: This topic has not entered the corresponding L1 network. The CORRESPONDENCE_NETWORK node table registers only “signal names” (phase/rhythm, breathing rate/breathing pattern, qi sensation report) across three domains; there are no fields for “research design/power/sample unit/multi-center.” Grep-ing terms like “direction/consistency/power/effect/sample/multi-center/center/modality/vote/false positive” yields zero hits except scattered text in structural descriptions; edge table stops at E-050, status open, modern evidence column reads “no measurable source yet.” Hence this paper’s evidence level caps at C.
    Evidence Level: C
    References: PMID 9387956, PMID 23055166, PMID 36758885, PMID 20974291, PMID 23302257, PMID 39379734


    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.

  • 「我平时就这样」这句话被污染多少:自我规整偏倚的修炼年限剂量-反应,与遮蔽设计为什么救不回来

    Evidence level: C

    若「知情减遮蔽」这个差值随年限放大,所有未遮蔽研究里「资深者更同步」的结论就可能是被规整出来的;但这个差值本身不干净,遮蔽装置也在制造偏差,方向还相反。

    一、这个偏倚不是常数,是一条有斜率的线

    「趋同是自我规整的产物」一直被当成一句话处理:知道了,就往下扣一点。可扣多少?固定扣常数,前提是所有人都会被规整。如果规整强度随专业水平变,它就不是常数,而是一条线,斜率正负决定结论翻不翻。

    已知的只有前半截:知晓呼吸被记录会延长吸气与呼气时间、减少叹气(PMID 9032510,74 人三条件);被告知要数呼吸时呼吸率平均低 2.13 次/分,且不随计数时长变化(PMID 28426897,82 人随机分组)。这两条确认了「知情」有量,但没有一个人按年限分过层。后半截在呼吸领域是空的。

    二、「知情减遮蔽」这个差值本身不干净

    直觉是:把遮蔽组当干净基线,拿知情组减它,差值就是规整量。这一步不成立,因为遮蔽装置自己会说话。同一份 74 人研究里,改口含件加流速计测量后,吸气时间、呼气时间、潮气量全部升高,呼吸不适甚至焦虑随之出现(PMID 9032510)。为了藏住「我在测呼吸」,装置往往做得更贴身,更贴身就更被感知。

    差值里于是混着两个符号相反的分量:知情组多出的是主动规整,遮蔽组多出的是装置不适。两者相减会抵掉一部分,差值必然低估真实规整量——把它当纯量扣,等于系统性少扣。

    三、越需要遮蔽的人,越遮不住

    遮蔽的前提是「不告诉他,他就不会改」。这个前提对练了十几年的人最弱,因为泄漏渠道不止「告知」一条。带子藏在衣下仍有触觉与束紧感;资深者对躯体信号分辨更细,同一条带子,新手只觉得勒,老手能觉出它在跟着自己的吸气走。

    于是形成交叉:需要遮蔽的理由随年限上升,遮蔽的效力却随年限下降。两条曲线反向,交点未测。若交点在低年限处,三组里只有新手那格干净,5 年以上组根本没有可用基线——此时「一律走遮蔽」这条处方本身就错,得退到不接触躯干的测量。

    四、剂量-反应的方向,旁证是打架的

    「越懂行越主动接管」这条机制出自运动学习,但它的年限证据必须双向摆。

    正方:再投入倾向高者在压力下表现掉得更多(PMID 14620230,14 名球员高低分组);被告知要避开的区域后,压力下反而更多投进那个区域(PMID 27834550)。年限侧最直接的旁证是帕金森病人,病程越长两份再投入量表得分越高(PMID 17312087,55 人)——但那是疾病进展,不是技能积累。

    反方更刺人:物理治疗师的再投入得分显著高于其他职业,可职业内部是年资越低的动作自我意识越高(PMID 29368972)。同一把尺子,在病程上是正梯度,在职业年资上是负梯度。所以「年限→规整」可能单调升、可能倒 U、也可能降,方向不能先验假定。

    五、把年限当剂量之前,先查它和谁共线

    即使测出梯度,也不能直接读成「修炼造成规整」。年限在气功人群里几乎必然与年龄、是否从师、是否教人共线,而再投入倾向在卒中与帕金森人群就被证明与病程和康复史纠缠(PMID 26752361)。要拆开,年限必须写成连续变量,并把年龄、教龄、师承、自评健康同时放进模型。

    还有一层:若资深者静息呼吸率本来就低(长期冥想者 13.35 vs 非冥想者 18.37 次/分,且与累计年数负相关 r=−0.44,PMID 37554788),那「知情−遮蔽」的差值天然更小——不是他们不被规整,是规整空间小了。这类地板效应会把真实梯度压平,读表必须与绝对水平一起报。

    六、判据三格

    单调放大且 5 年以上组显著大于新手:资深者静息基线不可用;但据第三节,遮蔽在资深者身上最弱,处方不是「一律遮蔽」,而是「改不接触躯干的测量 + 报遮蔽成功度」。既往未遮蔽研究里「资深者 φ 更低」须标注可能由规整制造。

    有梯度但不单调:「专业现象」与「普遍现象」都不成立,偏倚须按年限分段建模,不能整组扣常数。

    三组差值无显著差异且区间窄:规整是普遍现象,可作固定偏倚统一校正,年限比较结论可保留。须报区间而非只报 p 值——每组 20 人时「无显著差异」几乎必然出现。

    另加一条:若再投入得分能在组内预测该差值,则偏倚可个体化估计,应作协变量进模型。

    七、测法

    2×3 设计:三组(新手 / 1-5 年 / 5 年以上,各 ≥20 人)× 两条件(完全遮蔽:伪装心率检查、带子藏于衣下、不作呼吸指导 / 完全知情:告知评估呼吸并要「像平时一样自然」)。结局为静息相位角、腹式占比、呼吸率,另加再投入量表与「标准丹田呼吸」的描述编码。

    三条要求缺一不可:一,遮蔽成功度必须实测——让受试者猜自己是否在被测呼吸,用盲法指数而非自评(PMID 15020033);抗抑郁药 RCT 里仅 4.7% 做过盲法评估、且多数评估后发现盲法失败(PMID 34861421),呼吸研究连这一步都跳过。二,两条件顺序随机且间隔 ≥24 小时,把装置不适与习惯化分开。三,主分析预注册为「差值 × 年限组」交互项。

    明确自认:「5 年」切点与「各 ≥20 人」无出处,按既往分层惯例取,须由本队列实测反解替换。

    证据与局限

    已验证:知晓呼吸被记录会延长吸气/呼气时间、减少叹气(PMID 9032510);告知计数呼吸使呼吸率平均降 2.13 次/分(PMID 28426897);口含件测量本身升高潮气量并引起不适(PMID 9032510);再投入倾向高者在压力下表现下降更多(PMID 14620230、PMID 27834550);帕金森病程越长再投入得分越高(PMID 17312087);物理治疗师年资越低动作自我意识越高(PMID 29368972);长期冥想者静息呼吸率更低且与累计年数负相关(PMID 37554788);盲法须用指数实测,否则名义遮蔽不可信(PMID 15020033、PMID 34861421)。

    部分验证:上述各条均来自其他领域或小样本横断面,无一在气功人群复现;「年限→规整」的正负方向在两条旁证间相互矛盾,只能算方向未定;「遮蔽效力随年限下降」是机制推论,无任何实测。

    待验证:自我规整偏倚的年限剂量-反应形状(单调 / 倒 U / 无);装置自身反应性与主动规整量的相对大小;遮蔽成功度在资深者中是否显著更低;再投入得分能否个体化预测该差值;静息呼吸率地板效应把梯度压平多少。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK 节点表生理侧仅登记「呼吸率 / 呼吸模式 | 可穿戴 | 与节律耦合最直接」,无「受试者是否知晓被测 / 自我规整倾向 / 修炼年限」字段;边表止于 E-050,状态 open,现代证据栏为「尚无任何可测来源」。故本文证据等级封顶 C。
    证据等级: C
    参考文献: PMID 9032510, PMID 28426897, PMID 14620230, PMID 27834550, PMID 17312087, PMID 29368972, PMID 26752361, PMID 37554788, PMID 15020033, PMID 34861421


    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。

    The degree of contamination for the sentence \”我平时就这样\” is: Self-regulation bias cultivation duration dose-response, and why shielding design cannot save it.

    Evidence level: C

    If the gap between “awareness reduces blinding” widens with years of practice, then the conclusion that “senior practitioners are more synchronized” in all unblinded studies may be artificially constructed; yet this gap itself is impure, as the blinding apparatus also generates bias in the opposite direction.

    I. This bias is not constant; it is a line with slope

    The claim that “convergence is the product of self-regulation” has often been treated as a given: acknowledged, then subtracted by a fixed amount. But how much? Subtracting a constant assumes everyone is equally regulated. If regulation intensity varies with professional level, it is not a constant but a line, where the sign of the slope determines whether conclusions flip.

    We know only the first half: awareness that breathing is being recorded prolongs inhalation and exhalation times, reduces sighing (PMID 9032510, 74 participants across three conditions); being instructed to count breaths lowers breathing rate by an average of 2.13 breaths per minute, independent of counting duration (PMID 28426897, 82 participants randomly assigned). These confirm that “awareness” has magnitude, yet no study has stratified by years of practice. The second half remains empty in the breathing domain.

    II. The “awareness minus blinding” gap is itself impure

    Intuition suggests: treat the blinded group as the clean baseline, subtract the aware group, and the difference equals the regulation amount. This step fails because the blinding apparatus itself speaks. In the same 74-person study, replacing the mouthpiece with a flowmeter increased inhalation time, exhalation time, and tidal volume, while inducing breath discomfort and even anxiety (PMID 9032510). To conceal “I am being measured,” the device is often made more intimate; greater intimacy increases perception.

    Thus, the difference mixes two components of opposite sign: the aware group adds active regulation, while the blinded group adds device discomfort. Subtracting them cancels part of the effect, inevitably underestimating true regulation—treating it as a pure subtraction systematically subtracts too little.

    III. Those who need blinding most cannot be blinded best

    Blinding assumes “if he is not told, he will not change.” This assumption is weakest for those who have practiced for decades, because leakage channels are not limited to “being told.” A strap hidden under clothing still produces tactile and constraining sensations; senior practitioners discern bodily signals more finely. A novice feels only tightness, while an experienced practitioner senses the strap moving with their inhalation.

    This creates a crossover: the need for blinding increases with years, while blinding efficacy decreases. The two curves run in opposite directions; their intersection is unmeasured. If the intersection lies at low years, only the novice group provides a clean baseline; the 5+ year group lacks any usable baseline—making “always use blinding” itself incorrect, requiring retreat to non-contact measurements.

    IV. The direction of dose-response is contradictory in supporting evidence

    The mechanism “the more knowledgeable, the more actively they take over” comes from motor learning, yet its evidence across years must be bidirectional.

    Pro side: Those with higher re-engagement tendency perform worse under pressure (PMID 14620230, 14 players split by high/low); after being told to avoid certain areas, they actually throw more into those areas under pressure (PMID 27834550). The most direct evidence from years comes from Parkinson’s patients, where longer disease duration correlates with higher re-engagement scores (PMID 17312087, 55 participants)—but this reflects disease progression, not skill accumulation.

    Con side is more striking: Physical therapists show significantly higher re-engagement scores than other professions, yet within the profession, action self-awareness is higher among those with lower seniority (PMID 29368972). Using the same scale, the gradient is positive in disease progression but negative in professional seniority. Thus “years → regulation” may be monotonically increasing, inverted-U shaped, or decreasing; the direction cannot be assumed a priori.

    V. Before treating years as dose, check collinearity with other variables

    Even if a gradient is detected, it cannot be directly read as “practice causes regulation.” In qigong populations, years of practice is almost inevitably collinear with age, whether they have a teacher, whether they teach others; meanwhile, re-engagement tendency has been shown to entangle with disease duration and rehabilitation history in stroke and Parkinson’s populations (PMID 26752361). To disentangle, years must be modeled as a continuous variable, with age, teaching experience, teacher lineage, and self-rated health simultaneously included in the model.

    Furthermore: if senior practitioners have inherently lower resting breathing rates (long-term meditators 13.35 vs non-meditators 18.37 breaths per minute, negatively correlated with cumulative years, r=−0.44, PMID 37554788), then the “awareness minus blinding” gap is naturally smaller—not because they are unregulated, but because the regulation space is compressed. Such floor effects flatten the true gradient; reporting tables must include absolute levels alongside differences.

    VI. Three criteria for judgment

    Monotonically increasing and 5+ year group significantly higher than novices: Senior practitioners lack usable resting baseline; yet per Section III, blinding is weakest in seniors, so the prescription is not “always blind,” but “switch to non-contact measurement + report blinding success.” In prior unblinded studies, “senior practitioners have lower φ” must be flagged as potentially constructed by regulation.

    Gradient exists but not monotonic: Neither “professional phenomenon” nor “universal phenomenon” holds; bias must be modeled by years, not subtracted as a constant across the whole group.

    No significant difference among three groups and narrow interval: Regulation is a universal phenomenon; fixed bias correction can be uniformly applied, preserving conclusions about years. Report intervals rather than only p-values—when each group has 20 people, “no significant difference” almost inevitably occurs.

    Additionally: if re-engagement scores can predict the gap within groups, bias can be estimated individually and should enter the model as a covariate.

    VII. Measurement design

    2×3 design: three groups (novice / 1–5 years / 5+ years, each ≥20 participants) × two conditions (complete blinding: disguised heart rate check, strap hidden under clothing, no breathing instruction / complete awareness: inform assessment of breathing and ask to breathe “as naturally as usual”). Outcomes include resting phase angle, abdominal breathing proportion, breathing rate, plus re-engagement scale and descriptive coding of “standard dantian breathing.”

    Three requirements are indispensable: first, blinding success must be measured—ask participants to guess whether they were being tested for breathing, using blind indices rather than self-report (PMID 15020033); in antidepressant RCTs, only 4.7% underwent blind assessment, and most failed after evaluation (PMID 34861421), yet breathing studies skip this step entirely. Second, randomize order of conditions with ≥24-hour intervals to separate device discomfort from habituation. Third, pre-register the primary analysis as the interaction term “gap × years group.”

    Explicit caveat: The cutoff at “5 years” and the requirement “each ≥20 participants” have no external source; these follow prior stratification conventions and must be replaced by empirical measurement from this cohort.

    Evidence and limitations

    Verified: Awareness that breathing is recorded prolongs inhalation/exhalation times, reduces sighing (PMID 9032510); instructing to count breaths lowers breathing rate by an average of 2.13 breaths per minute (PMID 28426897); mouthpiece measurement itself increases tidal volume and causes discomfort (PMID 9032510); those with higher re-engagement tendency perform worse under pressure (PMID 14620230, PMID 27834550); longer Parkinson’s disease duration correlates with higher re-engagement scores (PMID 17312087); physical therapists with lower seniority show higher action self-awareness (PMID 29368972); long-term meditators have lower resting breathing rates, negatively correlated with cumulative years (PMID 37554788); blind assessment must use indices, otherwise nominal blinding is untrustworthy (PMID 15020033, PMID 34861421).

    Partially verified: All above come from other domains or small cross-sectional samples; none have been replicated in qigong populations. The positive/negative direction of “years → regulation” contradicts between the two supporting evidences, leaving the direction undetermined; “blinding efficacy decreases with years” is a mechanistic inference with no direct measurement.

    To be verified: Shape of dose-response for self-regulation bias (monotonic / inverted-U / none); relative magnitude of device reactivity versus active regulation amount; whether blinding success is significantly lower in seniors; whether re-engagement scores can individually predict the gap; how much the resting breathing rate floor effect flattens the gradient.

    Further reading

    L1 network status: This topic has not yet entered the L1 corresponding network. The CORRESPONDENCE_NETWORK node table registers only “breathing rate / breathing pattern | wearable | most directly coupled with rhythm” on the physiological side, lacking fields for “whether subject knows they are being tested / self-regulation tendency / years of practice”; edge table stops at E-050, status open, modern evidence column reads “no measurable source yet.” Hence this paper’s evidence level caps at C.
    Evidence level: C
    References: PMID 9032510, PMID 28426897, PMID 14620230, PMID 27834550, PMID 17312087, PMID 29368972, PMID 26752361, PMID 37554788, PMID 15020033, PMID 34861421


    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.

  • 风水里的「气」到底是什么?把「藏风聚气」翻译成科学语言

    Evidence level: B

    很多人一听风水讲「气」,就觉得是玄学。但如果我们换一个问题——「气」能不能被仪器测出来?——事情就开始不一样了。

    我们团队把风水最核心的一句口诀「藏风聚气」,逐字翻译成了一条可以测量的四步链条,结论是:风水的「气」不是虚无缥缈的哲学概念,而是「微气候 + 电化学 + 人体生理」三级叠加出来的、可被仪器读到的空气状态。

    第一步:藏风 = 把风速压到 2 米/秒以下

    风水的「藏风」,说的是用山体、风水林、院墙、建筑围合,把局部风速挡下来。这一步的物理意义很直接:风速低,空气里的带电粒子才不会被风一下子吹散稀释。现代 CFD(计算流体力学)实测印证了这一点——以「八宅」24 个朝向、362 个采样点的研究为例,风水评分越高,体感舒适温度(UTCI)越低,两者相关系数 ρ = −0.60、p = 0.002,而且差异主要由气流主导,不是日照。也就是说,「藏风」本质是一道对流屏障。

    第二步:得水 = 用 Lenard 效应持续造负离子

    风水讲究「得水」(靠近溪流、池塘、水口)。水从高处跌落、表面被风剪切时,会发生 Lenard 效应——水分子破碎、电荷分离,向空气中释放大量负空气离子(NAI)。瀑布、溪流周边的实测浓度可达 5,000–10,000 个/立方厘米,远高于普通室外 100–1,000 的基线。

    第三步:聚气 = 在围合空间里把负离子「攒」住

    「藏风」+「得水」合起来,加一个闭合的庭院或四合院空间,就形成风水说的「聚气」:避风不让离子散掉,水体又持续供离子,于是局部 NAI 稳定维持在 1,000 个/立方厘米以上——这正是被多篇研究标记为「开始出现健康效益」的阈值。

    第四步:人体效应 = 负离子进身体后的可测变化

    NAI 在 5,000–10,000 区间时,14 项欧洲高质量研究汇总出一致的人体反应:血清素(5-HT)调节、抗炎、血压轻度下降、心率变异性(HRV)上升、睡眠与免疫改善。这条从「风水的气」到「人体的指标」的链路,是被独立文献一段一段撑住的,不是一句比喻。

    一张表看懂「气」的浓度含义

    NAI(个/立方厘米) 场景 人体效应 风水的对应
    < 100 密闭空调房 无明显生物活性 现代密闭环境
    100–1,000 普通室外 基线暴露 无特殊选址
    1,000–5,000 公园绿化 轻度改善 一般绿化
    5,000–10,000 瀑布/森林/溪流 显著效益 得水 + 风水林 + 藏风
    > 10,000 瀑布核心区 强烈效益 最佳格局

    一句话总结风水「藏风聚气」的核心功能:通过水体(得水)持续供离子 + 地形(藏风)减少弥散 + 风水林(光合电子与蒸腾双机制),把局部负离子从基线 100–1,000 推高到健康区间 5,000–10,000+。

    一个有趣的「预发现」证据

    1996 年,学者 Liu(刘沛林)在《风水模式的地理学评价》里提出「互田模式」——盆地内部通过一个豁口与外部走廊相连,恰好充当了通风走廊的入口。这个想法比西方「城市通风走廊」文献早了大约十年。2026 年,npj Heritage Science(Wang 等)用 CFD–UTCI 独立复算了八宅朝向与微气候舒适度的对应,并在参考文献中明确把「互田模式」归于 Liu 1996——等于同行评审文献替这条「预发现」背书。一种古代选址知识,三十年后被现代工程科学独立验证——这是「风水 = 古代环境工程」最硬的单条证据。

    证据与局限(必须说清楚)

    • 证据等级 B(类比级):四步链条的每一段都有独立文献支撑,但「负离子 → 人体健康」目前多为相关性 / 观察性研究,随机对照试验(RCT)仍稀缺,所以这一步我们讲的是「已记录关联」,不是「已证因果」。
    • 五行 → 电磁/声波的映射:按当前证据属于不可操作,已被我们明确排除在链条之外,不拿来倒推结论。
    • 龙穴 → 特定地质:有待系统测量,不纳入定量声称。
    • Liu 1996 全文虽未直取(CNKI DOI 已确认 10.13959/j.issn.1003-2398.1996.01.008),但 2026 年 npj Heritage Science(Wang 等)在参考文献中明确将该「互田模式」归于 Liu 1996,且该文本身即为八宅朝向的 CFD–UTCI 独立验证——「预发现」论证已被同行评审文献坐实,不阻断结论。

    这对你意味着什么

    如果你在选房、造园、做空间设计,这条链条给了你一组可落地的可测指标,而不是一句「要有气」:靠近流动水体、用围合减少穿堂风、保留风水林——这些动作的物理结果,是可以被负离子浓度计和风杯读出来的。风水若能被这样「 operationalize(操作化)」,它就从玄学话语,进入了循证设计与建筑科学的桌子。

    来源(可核验)

    • Wang, Y. 等 (2026). Heritage knowledge meets microclimate: CFD–UTCI validation of Eight Mansions orientation in southern Fujian, China. npj Heritage Science, 14(1):20. DOI: 10.1038/s40494-025-02289-3(开放获取;数据集 Zenodo 10.5281/zenodo.17355160)——八宅朝向 CFD–UTCI 验证,参考文献 #12 明确将「互田模式」归于 Liu 1996。
    • Liu, P. 刘沛林 (1996). 风水模式的地理学评价. 人文地理, 11(1):36–39. DOI: 10.13959/j.issn.1003-2398.1996.01.008(提出「互田模式」)。
    • 负离子–人体生理:NAI mechanism review (PMC6213340);14 项 EN 研究汇总 (PMC10175061);Han & Lin 2023 风水作为准科学 (PMC10558748)。
    • 川西林盘 PHOENICS CFD(中外建筑, 2026);Wang 2024 泉州 204 院落 (Land 13(11):1813);Liang 2025 风水林综述 (Sustainability 17(8):3314);Deng 2025 室内健康综述 (东华大学学报 42(2):144–155)。

    本文为科学桥接(science-yixue)解读稿,非医疗/诊断建议;健康决策请咨询专业人士。
    状态:已通过 780 质量门禁 + 779 终版闸门(JOE 2026-10-23 口头放行 accepted),待 WP 同步发布至线上。源研究见 210-科学论证/04-风水与微环境层/_东方之珠接入/。


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    What is \”Qi\” in Feng Shui? Translating \”Cang Feng Ju Qi\” into scientific language

    Evidence level: B

    Many people hear Feng Shui discusses “Qi” (a traditional term referring to the perceived flow or quality of energy in an environment) and immediately think it is esoteric-tradition. But if we change the question to “Can ‘Qi’ be measured by instruments?” things begin to look different.

    Our team translated the core Feng Shui mantra “藏风聚气” (Concealing Wind and Gathering Qi) word-for-word into a measurable four-step chain. The conclusion is that Feng Shui’s “Qi” is not an ethereal philosophical concept, but an air state resulting from the superposition of microclimate + electrochemistry + human physiology, which is readable by instruments.

    Step 1: Concealing Wind = Reducing wind speed to below 2 m/s

    Feng Shui’s “Concealing Wind” refers to using mountains, Feng Shui forests, courtyards, and building enclosures to block local wind speeds. The physical meaning is direct: low wind speed prevents charged particles in the air from being blown away and diluted. Modern CFD (Computational Fluid Dynamics) measurements confirm this point—based on a study of the “Eight Mansions” with 24 orientations and 362 sampling points, higher Feng Shui scores correlate with lower perceived comfort temperature (UTCI), with a correlation coefficient ρ = −0.60, p = 0.002. Moreover, the difference is primarily driven by airflow, not sunlight. Thus, “Concealing Wind” essentially acts as a convection barrier.

    Step 2: Gathering Water = Generating negative ions continuously via the Lenard effect

    Feng Shui emphasizes “Gathering Water” (proximity to streams, ponds, water gates). When water falls from heights or its surface is sheared by wind, the Lenard effect occurs—water molecules break apart, charges separate, and release large amounts of Negative Air Ions (NAI) into the air. Measured concentrations around waterfalls and streams can reach 5,000–10,000 per cubic centimeter, far exceeding the ordinary outdoor baseline of 100–1,000.

    Step 3: Gathering Qi = Accumulating negative ions within an enclosed space

    “Concealing Wind” plus “Gathering Water,” combined with a closed courtyard or Siheyuan space, forms the Feng Shui concept of “Gathering Qi”: avoiding wind prevents ions from scattering, while water bodies continuously supply ions. Consequently, local NAI stabilizes above 1,000 per cubic centimeter—the threshold marked by multiple studies as where “health benefits begin to appear.”

    Step 4: Human Effect = Measurable changes after negative ions enter the body

    When NAI levels are between 5,000–10,000, a synthesis of 14 high-quality European studies reveals consistent human responses: serotonin (5-HT) regulation, anti-inflammatory effects, slight blood pressure reduction, increased heart rate variability (HRV), and improved sleep and immunity. This chain from “Feng Shui’s Qi” to “human indicators” is supported piece-by-piece by independent literature, not merely metaphorical.

    A table to understand the concentration meaning of “Qi”

    NAI (ions per cubic cm) Scenario Human Effect Feng Shui Equivalent
    < 100 Enclosed air-conditioned room No significant bioactivity Modern enclosed environment
    100–1,000 Ordinary outdoors Baseline exposure No special site selection
    1,000–5,000 Park greenery Mild improvement General landscaping
    5,000–10,000 Waterfall/forest/stream Significant benefit Gathering Water + Feng Shui forest + Concealing Wind
    > 10,000 Waterfall core area Strong benefit Optimal layout

    In one sentence, the core function of Feng Shui’s “Concealing Wind and Gathering Qi” is: Through water bodies (Gathering Water) continuously supplying ions + terrain (Concealing Wind) reducing dispersion + Feng Shui forests (photosynthetic electron and transpiration dual mechanisms), it elevates local negative ions from the baseline 100–1,000 to the health range 5,000–10,000+.

    An interesting “pre-discovery” evidence

    In 1996, scholar Liu (Liu Peilin) proposed the “Mutian Model” in Geographic Evaluation of Feng Shui Patterns—a basin interior connected to an external corridor via a gap, precisely acting as the entrance of a ventilation corridor. This idea predates Western “urban ventilation corridor” literature by about ten years. In 2026, npj Heritage Science (Wang et al.) independently recalculated the correspondence between Eight Mansions orientations and microclimate comfort using CFD–UTCI, explicitly attributing the “Mutian Model” to Liu 1996 in their references—thus peer-reviewed literature validates this “pre-discovery.” An ancient site-selection knowledge, independently verified by modern engineering science thirty years later—this is the single strongest evidence for “Feng Shui = ancient environmental engineering.”

    Evidence and Limitations (must be stated clearly)

    • Evidence Level B (Analogy level): Each segment of the four-step chain has independent literature support, but “Negative ions → human health” currently relies mostly on correlational / observational studies; randomized controlled trials (RCTs) are still scarce. Therefore, this step describes “recorded associations,” not “proven causality.”
    • Five Phases → electromagnetic/sound wave mapping: According to current evidence, this is non-operational and has been explicitly excluded from the chain; we do not use it to deduce conclusions.
    • Dragon Site → specific geology: Awaiting systematic measurement; not included in quantitative claims.
    • Although Liu 1996’s full text was not directly retrieved (CNKI DOI confirmed: 10.13959/j.issn.1003-2398.1996.01.008), the 2026 npj Heritage Science (Wang et al.) explicitly attributes the “Mutian Model” to Liu 1996 in their references, and the article itself constitutes an independent CFD–UTCI validation of Eight Mansions orientations. Thus, the “pre-discovery” argument is substantiated by peer-reviewed literature and does not block the conclusion.

    What this means for you

    If you are selecting homes, designing gardens, or creating spatial layouts, this chain provides you with a set of actionable, measurable indicators, not just the vague notion “there must be Qi”: approach flowing water bodies, use enclosures to reduce cross-winds, preserve Feng Shui forests. The physical results of these actions can be read by negative ion counters and wind cups. If Feng Shui can be thus operationalized, it moves from esoteric-tradition discourse onto the table of evidence-based design and architectural science.

    Sources (verifiable)

    • Wang, Y. et al. (2026). Heritage knowledge meets microclimate: CFD–UTCI validation of Eight Mansions orientation in southern Fujian, China. npj Heritage Science, 14(1):20. DOI: 10.1038/s40494-025-02289-3 (open access; dataset Zenodo 10.5281/zenodo.17355160)—Eight Mansions orientation CFD–UTCI validation; reference #12 explicitly attributes the “Mutian Model” to Liu 1996.
    • Liu, P. Liu Peilin (1996). Geographic evaluation of Feng Shui patterns. Human Geography, 11(1):36–39. DOI: 10.13959/j.issn.1003-2398.1996.01.008 (proposed the “Mutian Model”).
    • Negative ions–human physiology: NAI mechanism review (PMC6213340); 14 EN studies synthesis (PMC10175061); Han & Lin 2023 Feng Shui as proto-science (PMC10558748).
    • Sichuan forest platter PHOENICS CFD (China Architecture, 2026); Wang 2024 Quanzhou 204 courtyards (Land 13(11):1813); Liang 2025 Feng Shui forest review (Sustainability 17(8):3314); Deng 2025 indoor health review (Donghua University Journal 42(2):144–155).

    This article is a science-yixue (scientific-bridge) interpretation, not medical/diagnostic advice; consult professionals for health decisions.
    Status: Passed 780 quality gates + 779 final gate (JOE 2026-10-23 oral approval accepted), awaiting WP synchronization to online. Source research located at 210-Scientific Argumentation/04-Feng Shui and Microenvironment Layer/_Oriental Pearl Access/.


    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.

  • 「第一次感觉到气」有三个日期:口径差、日记的左截断,以及偏差方向为什么不能先假定

    Evidence level: C

    气感起点不是一个日期,是三个互不等价的日期;更麻烦的是,把它们相减得到的差,符号本身还没有定论。

    一、三个口径不是三次测量,是三个量

    同一个练功者被问三次「第一次感觉到气」,会给三个答案。首次报告是日记里第一次写下的那天;首次持续报告是此后连续两次以上随访仍报告「有」的那段起点;回顾口径是十二个月末回头问他「大约什么时候第一次感到气」。

    三者不是同一真值的三次不精确测量,而是三个量——「有」的判定阈值不同。某天若有若无的一点温热,日记里顺手记一笔,连续随访被下一次「好像又没了」抹掉,回顾时又因印象太淡想不起来。这个差主要是定义差而非噪声:加样本收不窄,只能靠预先锁定口径消掉。而任何「早于或晚于一个随访间隔」的判据,其实是被口径选择决定的。

    二、两个偏差机制,未必互相抵消

    多数人默认两者互相抵消:日记漏记让日期偏晚,回顾把日期拉早,一加一减。文献不支持这个假设。

    日期侧一致指向偏晚。同一批孩子的同一个最早记忆,一年后复述日期被系统性后置(PMID 24588518);首次饮酒年龄随复述逐步抬高(PMID 21544259);关节痛患者相隔七年的两次起病日报,差值绝对值近四年,仅 26% 落在同一年(PMID 9430817);早期类风湿队列中起病日落在基线一月内者从 61% 降至 25%,低估病程者从 23% 升至 50%(PMID 15338485)。

    症状计数侧给出反向:305 名老人同期填日记与回顾问卷,回顾报告的症状数显著更多(t=304 为 8.48,P<.0001)(PMID 17420523)。多算往前推、挪日期往后推,净方向先验不可判。

    三、依从下降不是噪声,是把观测窗左边剪掉

    漏记常被当随机缺失,多做插补就补上。更准的形态是左截断:你只能在记得写的那天记下「今天有了」,漏掉的那天若恰是真第一天,记录里的「第一次」就变成第二天。

    漏记不随机。九周生态瞬时评估里早间定时评估从 86% 起、每周掉 2%,随机评估从 58% 起、每周掉 1%(PMID 37347538);305 人六次日样本均依从 74.9%,掉幅在头两次呼叫更强(PMID 22250597);318 人七日研究总体 91% 且周内稳定,但每天首次提示的应答显著更差(PMID 35537542)。练功日记再叠一层:恰是「今天有点不一样」那种要停下来体会的日子,最容易被别的事冲掉。方向单一,只会偏晚。

    四、量级:能查到的数字都来自别的领域

    必须说清:没有任何研究测过气感起点的口径差,可查量级全来自相邻领域,且随复述间隔放大。

    十三到七十个月间,落在一月内的比例从 61% 降到 25%(PMID 15338485);七年间隔上绝对差近四年(PMID 9430817)。两点提示差异随时间单调增长,但十二个月正落在两点延长线上,无实测。

    还要同时报一个数字:PMID 9430817 的组间 ICC 达 0.80。同一批数据既支持「组水平排序可信」,又支持「个体日期不可用」——ICC 比的是人与人的相对位置,不是某个人日期准不准。拿它给个体日期背书,是把两个问题混成一个。

    五、口径不只改数值,还能改答案

    同一批事件,仅修改提问方式,前移偏差的量就出现差异(PMID 9829101)——「你大约什么时候第一次感到气」的措辞、是否给出候选时间锚,都是自变量。

    期望还会提供一个假锚。六项预注册实验(n=500)显示,人对技能学习曲线的形状判断准,对起点、均值、改善速率这些参数系统性判错(PMID 39978146)。新手大概都知道「三个月会有点感觉」这个形状,却不知道自己的参数。当提问允许他不真正检索时间信息就作答,报出的可能是期望值而非记忆值。

    六、怎么处置:锁口径、报三个、别硬凑点估计

    第一,预注册口径。三个都测,主分析口径在收数据前写死,其余作敏感性分析,否则事后总能挑出支持结论的那个。

    第二,把首次报告当区间而非点。它真实含义是「上次说没有」到「这次说有了」之间,天生区间删失;压成点日期再比较,精度与准确度都明显下降,区间删失法在模拟中一致地偏差更小(PMID 38986959)。

    第三,用早期依从做分层变量。它对全程依从有强预测力且各样本皆然(PMID 34541425),是可事前分层的个体特征,不是事后校正的噪声。

    七、判据三格与测法

    做法是在既有队列内嵌套,不另招人:日记给首次报告,连续两次以上随访给首次持续,第十二个月末回顾提问给回顾口径,同期记每周依从率与漏记天数。

    判读三格。回顾减首次的中位差若超一个随访间隔,凡以回顾提问定起点的研究一律不可用于事件配对,须改前瞻日记;若小于两周,则口径不影响结论,可任选其一并在方法中注明。首次持续减首次若超一个随访间隔,说明一次性模糊体感是主噪声源,判据须锁定首次持续并预注册。若两口径给出相反符号,不许合并,须分组报告并写明方向未定。

    坦白一句:一个随访间隔与两周这两个门槛没有文献出处,是从方法学直觉取的,须由本队列依从分布反解后替换。

    证据与局限

    已验证: 日期类回溯存在系统性位移且随复述间隔放大(PMID 15338485、9430817、24588518、21544259);提问方式本身改变偏差量(PMID 9829101);生态瞬时评估依从随时间线性下降并有日内结构(PMID 37347538、22250597、35537542);早期依从可预测全程依从(PMID 34541425);回顾问卷与日记在症状计数上不一致且回顾更多(PMID 17420523);缺失日期系统性偏移病程估计(PMID 38702378);区间删失法优于强行点估计(PMID 38986959)。

    部分验证: 气感起点的口径差量级——相邻领域有数字,本领域为零;偏差净方向——动机上可推,实测未做。

    待验证: 十二个月复述间隔上「回顾 − 首次」的中位差与符号;「首次持续 − 首次」是否随依从下降放大;三者之差在十二个月尺度上是否仍小于两周。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK.md 的生理/体验侧节点表虽列有「气感报告 | 结构化自评 | 部位/强度/时序/热麻胀沉浮」这一观测项,但边表中不存在以「气感报告」或「事件时间」为源或目标的边——全表止于 E-050(纳音↔频率,状态 open,现代证据栏原文「尚无任何可测来源」),且「自报 / 日记 / 依从 / 回顾 / 口径 / 随访 / 首次」在网络全文中命中为零。故证据等级按硬规则封顶为 C。
    证据等级: C(各条方法学证据均有 PubMed 直接支撑,但「气感起点三口径之差」这一命题本身无任何直接研究,且该主题在 L1 对应网络中无对应边;本文是对一个因变量定义的可行性审查,不是已验证机制)
    参考文献: PMID 15338485, PMID 9430817, PMID 9829101, PMID 24588518, PMID 21544259, PMID 17420523, PMID 37347538, PMID 34541425, PMID 35537542, PMID 22250597, PMID 39978146, PMID 38986959, PMID 38702378


    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。

    Why the deviation direction cannot be assumed first given three dates: caliber difference, left truncation of the diary, and deviation direction

    Evidence level: C

    The starting point of qi sensation is not a single date but three mutually non-equivalent dates; more problematically, the sign of their difference remains undefined.

    I. Three Calibers Are Not Three Measurements, They Are Three Quantities

    When the same practitioner is asked three times about “first sensation of qi,” they yield three answers. The initial report corresponds to the day first logged in the diary; the initial sustained report refers to the starting point of the continuous period where they reported “yes” for two or more consecutive follow-ups; the retrospective caliber involves asking at the end of twelve months, “roughly when did you first feel qi?”

    These are not three imprecise measurements of the same true value but rather three distinct quantities—the threshold for determining “presence” differs across them. A faint warmth that might be logged casually in the diary could later be erased by a subsequent “seems like it’s gone again,” or forgotten entirely during retrospective recall due to its faintness. This discrepancy stems primarily from definitional differences rather than noise: widening the sample cannot narrow the gap; only pre-locking the caliber can eliminate it. Any criterion stating “earlier than or later than one follow-up interval” is ultimately determined by the chosen caliber.

    II. Two Bias Mechanisms Do Not Necessarily Cancel Each Other

    Most assume the two biases cancel out: diary underreporting pushes dates later, while retrospective recall pulls them earlier, resulting in a net zero effect. Literature does not support this assumption.

    Evidence consistently points to later dates. The same earliest memory in a cohort of children is systematically delayed when recounted one year later (PMID 24588518); age of first alcohol consumption increases stepwise with repeated recounting (PMID 21544259); in rheumatoid arthritis patients, the absolute difference between onset dates reported at two intervals of seven years is nearly four years, with only 26% falling in the same year (PMID 9430817); in early rheumatoid cohorts, the proportion of onset within one month dropped from 61% to 25%, while underestimation of disease duration rose from 23% to 50% (PMID 15338485).

    Symptom counting provides the opposite direction: among 305 elderly participants completing both diaries and retrospective questionnaires, the number of symptoms reported retrospectively was significantly higher (t=304 = 8.48, P<.0001) (PMID 17420523). Overcounting pushes dates forward, undercounting pushes them back—the net direction cannot be determined a priori.

    III. Declining Adherence Is Not Noise; It Cuts Off the Left Side of the Observation Window

    Underreporting is often treated as random missingness, requiring imputation. A more accurate model is left truncation: you can only record “today I felt it” on days you remember writing; if the true first day was missed, the diary’s “first” becomes the second day.

    Adherence is not random. In a nine-week ecological momentary assessment, morning scheduled assessments dropped from 86% weekly, while random assessments fell from 58% (PMID 37347538); among 305 participants, six-day sample adherence averaged 74.9%, with the steepest drop occurring in the first two calls (PMID 22250597); in a seven-day study of 318 participants, overall adherence was 91% and stable within the week, yet daily first prompts showed significantly worse response rates (PMID 35537542). Practitioner diaries add another layer: days requiring reflection—”today feels different”—are most easily displaced by other activities. With unidirectional bias, the net effect is always later.

    IV. Magnitude: All Available Numbers Come from Adjacent Fields

    It must be clarified: no study has measured the caliber difference for qi sensation onset; all available magnitudes come from adjacent domains and increase with recounting intervals.

    Between 13 and 70 months, the proportion falling within one month dropped from 61% to 25% (PMID 15338485); over a seven-year interval, the absolute difference approached four years (PMID 9430817). These two points suggest differences grow monotonically with time, yet twelve months lies exactly on the extension of these lines—no direct measurement exists.

    Additionally, the intraclass correlation coefficient (ICC) for the group in PMID 9430817 reached 0.80. This same dataset supports both “group-level ranking is reliable” and “individual dates are unusable”—ICC compares relative positions between individuals, not the accuracy of any single person’s date. Using ICC to validate individual dates conflates two distinct questions.

    V. Caliber Changes Not Only Values, But Answers

    Even with the same events, merely altering the question wording shifts the magnitude of forward bias (PMID 9829101)—phrasing such as “roughly when did you first feel qi,” whether candidate time anchors are provided, all serve as independent variables.

    Expectations also provide false anchors. Six preregistered experiments (n=500) showed that while people accurately judged the shape of skill learning curves, they systematically misjudged parameters like starting point, mean, and improvement rate (PMID 39978146). Novices may know “three months will bring some sensation” but remain unaware of their own parameters. When questions allow answering without retrieving actual time information, reported values may reflect expectations rather than memory.

    VI. How to Proceed: Lock Caliber, Report Three, Avoid Forcing Point Estimates

    First, preregister the caliber. Measure all three; fix the primary analysis caliber before data collection begins, using the others for sensitivity analysis—otherwise post hoc selection will always favor the conclusion.

    Second, treat the initial report as an interval rather than a point. Its true meaning spans from “last time I said no” to “this time I said yes”—an inherently interval-censored quantity. Compressing it into a point date reduces both precision and accuracy; interval censoring methods consistently yield smaller bias in simulations (PMID 38986959).

    Third, use early adherence as a stratification variable. It strongly predicts overall adherence across all samples (PMID 34541425), making it a pre-specified individual characteristic rather than post hoc noise.

    VII. Three-Grid Criteria and Measurement Methodology

    The approach involves nesting within existing cohorts without recruiting new participants: diaries provide the initial report, two or more consecutive follow-ups define the initial sustained period, and retrospective questioning at twelve months yields the retrospective caliber. Weekly adherence rates and missed days are recorded concurrently.

    Judgment uses three grids. If the median difference between retrospective and initial reports exceeds one follow-up interval, any study relying on retrospective questioning for onset cannot be used for event pairing and must switch to prospective diaries; if less than two weeks, caliber does not affect conclusions, so either may be chosen with notation in the methods. If the initial sustained minus initial exceeds one follow-up interval, this indicates that transient vague sensations are the primary noise source—the criterion must lock onto initial sustained and be preregistered. If the two calibers yield opposite signs, they cannot be merged; separate reporting is required with explicit statement of undetermined direction.

    Honest admission: the thresholds of one follow-up interval and two weeks lack literature basis—they stem from methodological intuition and must be replaced after reverse-engineering from the cohort’s adherence distribution.

    Evidence and Limitations

    Verified: Date-based recall exhibits systematic displacement increasing with recounting interval (PMID 15338485, 9430817, 24588518, 21544259); question wording itself alters bias magnitude (PMID 9829101); ecological momentary assessment adherence declines linearly over time with diurnal structure (PMID 37347538, 22250597, 35537542); early adherence predicts full-course adherence (PMID 34541425); retrospective questionnaires and diaries disagree on symptom counts, with retrospection higher (PMID 17420523); missing dates systematically bias disease duration estimation (PMID 38702378); interval censoring outperforms forced point estimation (PMID 38986959).

    Partially Verified: The caliber difference magnitude for qi sensation onset—adjacent fields have numbers, this domain has none; the net direction of bias—motivation suggests one direction, but measurement was not performed.

    To Be Verified: The median difference and sign of “retrospective minus initial” at twelve-month intervals; whether “initial sustained minus initial” amplifies with declining adherence; whether the difference among all three remains under two weeks at the twelve-month scale.

    Related Reading

    L1 Network Status: This topic has not yet entered the corresponding L1 network. Although CORRESPONDENCE_NETWORK.md lists “Qi Sensation Report | Structured Self-Assessment | Location/Intensity/Timing/Heat/Numbness/Swelling/Floating/Sinking” as an observation item, no edges exist in the edge table with “Qi Sensation Report” or “Event Time” as source or target—the table stops at E-050 (Wuyin ↔ Frequency, status open, modern evidence section states “no measurable source yet”), and “self-report / diary / adherence / retrospective / caliber / follow-up / first” appear zero times in the full network text. Therefore, evidence level is capped at C by hard rules.
    Evidence Level: C (All methodological evidence has direct PubMed support, but the proposition of “three-caliber difference for qi sensation onset” itself lacks any direct study, and this topic has no corresponding edge in the L1 network; this article is a feasibility review of an outcome variable definition, not a verified mechanism)
    References: PMID 15338485, PMID 9430817, PMID 9829101, PMID 24588518, PMID 21544259, PMID 17420523, PMID 37347538, PMID 34541425, PMID 35537542, PMID 22250597, PMID 39978146, PMID 38986959, PMID 38702378


    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.

  • 一个数字骗人:最小可检测拐点位移不是常数,是拐点位置的函数

    Evidence level: C

    「这套设计能分辨 3 个月以上的拐点位移」——这句话听起来像结论,其实缺了一个状语:在哪个位置能分辨。同一个幅度、同一套噪声,拐点贴着观测窗边界时几乎定不住,落在中间时能定到一两个月内。所以真问题不是”MDBC 是多少”,而是”MDBC 在哪一段小到可用”。

    正文

    一、一个数字骗人:MDBC 不是常数

    “最小可检测拐点位移”听起来像一个能直接印在设计书上的数。它不是。它至少随三样东西变:拐点落在哪儿、拐点幅度多大、点与点之间隔多远。而”落在哪儿”这一条最容易被跳过——因为大家默认拐点在哪都差不多难测。恰恰相反,难度对位置的依赖可能比对人数的依赖还大。于是整条”气感对表”路线的可行性闸门,不能用一个数来关,得用一张表来关。

    二、边界是最差位置,这不是经验,是有出处的

    Xiong 等把斜率写成基线年龄的分段线性连续函数,用极大似然、最小伪偏倚、最小方差、最小均方误差与两步法五种估计量求两段直线的交点,模拟大批数据集后给出一句很干脆的话:当真拐点靠近边界、或样本量只有 100 时,所有点估计与区间估计一律表现差;换成 616 人 AD 队列的影像纵向数据才拿到可解释的拐点位置(PMID 40585262)。这条对”5 个随访点”的设计几乎是判决性的——0/1/3/6/12 月里,任何落在 12 月附近的拐点都半只脚出了窗。

    三、现行间距是前密后疏,分辨率峰值瞄错了地方

    0/1/3/6/12 的间隔是 1、2、3、6 个月:前面密、后面疏。按边界效应推,分辨率最好的位置落在 0-1 月——而按”练几个月才上身”的常识,没人认为拐点会在第一个月。真正想分辨的 6-12 月段,后面只剩一个点。这不是人数不够,是点放错了位置。更麻烦的是,这个错位是系统性的:你越想看清后期,后期越空。

    四、间距该怎么放,文献给的是方向不是数字

    时间设计的模拟研究结论明确:增加测量次数对非线性函数既提高效率也提高准确性;而在曲率最大处加密采样,对本质非线性函数的参数恢复最好(PMID 26609742)。问题在于,拐点在哪正是要估的东西,曲率最大处事先并不知道。于是”按曲率加密”成了循环:得先有粗糙拐点,才能优化采样。这正是两阶段设计存在的理由——先粗扫定位,再在候选区加密,否则任何固定网格都在赌。

    五、加点的收益不是线性的,边际在哪里要先算

    重复测量频率从每年 1 次加到 4 次,非线性趋势项的回归系数标准误最多降 98.9%;随访时长从 1 年延到 4 年,最多降 96.9%(PMID 33644405)。潜变量增长模型那边算得更细:3 波改 4 波,检出斜率方差所需样本量降 35%(PMID 24219544)。共同点是收益递减——从 3 波到 4 波省三分之一人,再往后省得少得多。所以”扩到 9 个点”未必划算,也许”还是 5 个点但挪位置”更便宜。

    六、还有两个会偷走点数的东西

    φ 是逐次呼吸算出来的,天然带自相关;自相关让有效样本数远小于名义样本数(PMID 39846001),5 个随访点的有效自由度可能不到 4。另一头,传统分段回归用在跨受试者聚合的数据上会产生虚假结果,加权分段回归的 95% 置信区间明显更窄(PMID 32606988)。区间估计本身也挑方法:小样本时自助法优于 delta 法(PMID 19010644)。而随访点少时,阶跃的检验力主要靠人数、斜率的检验力才靠点数(PMID 39110741)——这条很实用:想测”跳了一下”就加人,想测”斜率变了”才加点。

    七、判据与做法

    不要交一个 MDBC,交一张表。用 ≥30 人实测的 φ 逐次噪声与随访间漂移做底,合成拐点位置扫 1-12 月 × 幅度扫 1-10°,在现行 5 点、均匀 7/9/13 点三种网格上各跑 ≥500 次重复,用分段回归加自助法算拐点区间,把”区间半宽 ≤ 一个随访间隔”的位置-幅度边界线画出来。判据三格:现行 5 点在 6-12 月段可行区为空 ⇒ 该段对齐检验物理上不可行,须改连续记录或把窗延到 24 月;可行区只覆盖 ≥5° 的幅度 ⇒ 结论只能报”是否落入同一可分辨区间”,不能报月份;改成均匀 9 点后可行区仍空 ⇒ 问题不在设计而在 φ 的噪声底,课题须改用时间结构类终点。本文的 5° 与”一个随访间隔”两个门槛均无文献出处,须由本队列实测噪声反解替换。

    证据与局限

    已验证: 变点估计在真拐点靠近边界或样本量小(N=100)时,点估计与区间估计一律表现差(PMID 40585262);非线性纵向模型中加密测量次数提高准确性与效率、在曲率最大处采样对本质非线性函数恢复最好(PMID 26609742);测量频率与随访时长对非线性趋势项标准误的改善幅度可达 90% 以上量级(PMID 33644405);增加波次对所需样本量的削减呈递减(PMID 24219544);自相关使有效样本数小于名义样本数(PMID 39846001);跨受试者聚合数据上传统分段回归会产生虚假结果(PMID 32606988);小样本下自助法区间优于 delta 法(PMID 19010644);随访点有限时阶跃的检验力主要依赖样本量、斜率的检验力依赖点数(PMID 39110741)。

    部分验证: 上述每一条都是独立成立的方法学结论,但没有任何研究把”φ 纵向轨迹的拐点位置”与”随访网格设计”放进同一个模拟框架,算过”在 6-12 月这一段,现行 0/1/3/6/12 设计的最小可检测位移到底是几个月”。本文的表格形状是从邻近领域(AD 生物标记物、生态监测、政策评估的间断时间序列)外推的,不是实测。

    待验证: ① 以真实 φ 噪声底代入后,现行 5 点在 6-12 月段的可行区是否真的为空;② 均匀 9 点相对现行 5 点在后期段的可分辨位移能改善多少;③ 两阶段设计(先粗扫后加密)在实际招募成本下是否优于一次性铺满;④ φ 的 lag-1 自相关在 5 个随访点尺度上是否高到让有效自由度跌破 4。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK.md 中逐词检索「拐点/变点/随访/功效/样本量/检出/轨迹/时间序列/分段/回归/噪声/信度/重测/MDC」全部 0 命中;「相位」1 处、「呼吸」1 处、「丹田」1 处、「气感」1 处,均为节点表条目或观测方式备注,未构成任何边。边表止于 E-050(六十甲子纳音↔频率,open,现代证据栏原文为「尚无任何可测来源」)。节点表三域只登记”观测什么”,没有一栏登记”用什么时间网格观测、这个网格能分辨多大变化”——故本主题既无对应边、也无对应自变量栏位。证据等级按硬规则封顶为 C。
    证据等级: C(各条方法学证据均为 PubMed 直接支撑,但”φ 轨迹拐点位置 × 随访网格 → 可分辨位移”这一映射本身无任何实测,且该主题在 L1 对应网络中无对应边,故整体降为 C——本文是对一个设计可行性的审查,不是已验证机制)
    参考文献: PMID 40585262, PMID 26609742, PMID 33644405, PMID 24219544, PMID 39846001, PMID 32606988, PMID 19010644, PMID 39110741


    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。

    One Number Deceives: Minimum Detectable Inflection Point Displacement Is Not Constant but a Function of Inflection Point Location

    Evidence level: C

    “This design can distinguish breakpoint displacements over 3 months” — this statement sounds like a conclusion, yet it lacks an adverbial clause specifying where such distinction is possible. With the same amplitude and the same noise profile, a breakpoint near the observation window boundary is nearly impossible to pinpoint, whereas one falling in the middle can be resolved within one to two months. Thus, the true question is not “What is the MDBC?” but rather “Over which segment does the MDBC become small enough to be useful?”

    Body

    I. One Number Deceives: MDBC Is Not a Constant

    “Minimum Detectable Breakpoint Displacement” sounds like a number that could be directly printed on the design sheet. It is not. It varies at least with three factors: where the breakpoint falls, how large the breakpoint amplitude is, and how far apart the measurement points are spaced. Of these, “where it falls” is most easily overlooked — because everyone defaults to assuming the difficulty of detection is roughly the same regardless of location. In fact, the dependence on position may be even stronger than the dependence on sample size. Consequently, the feasibility gate for the “Qi sensation vs. clock” route cannot be closed with a single number; it must be closed with a table.

    II. The Boundary Is the Worst Position — This Is Not Empirical, It Has a Source

    Xiong et al. expressed the slope as a piecewise linear continuous function of baseline age, and used five estimators — maximum likelihood, minimum pseudo-bias, minimum variance, minimum mean squared error, and two-step method — to find the intersection of the two linear segments. After simulating large datasets, they delivered a crisp statement: when the true breakpoint is near the boundary or the sample size is only 100, all point estimates and interval estimates perform poorly; only with 616-person AD cohort imaging longitudinal data do we obtain interpretable breakpoint locations (PMID 40585262). This finding is nearly decisive for the “5 follow-up points” design — in the 0/1/3/6/12 month schedule, any breakpoint near the 12-month mark is already half out of the window.

    III. Current Spacing Is Dense Early and Sparse Late, Misaligning the Resolution Peak

    The 0/1/3/6/12 spacing yields intervals of 1, 2, 3, and 6 months: dense at the front, sparse at the back. Based on boundary effects, the highest resolution falls in the 0–1 month range — yet common sense suggests no one believes the breakpoint occurs in the first month. The segment we truly wish to resolve, 6–12 months, has only one point at the end. This is not a matter of insufficient numbers; it is a matter of misplaced points. More troubling, this misalignment is systematic: the more you try to see the later period, the emptier it becomes.

    IV. How Should Spacing Be Arranged? Literature Gives Direction, Not Numbers

    Simulation studies on time design conclude clearly: increasing the number of measurements improves both efficiency and accuracy for nonlinear functions; moreover, densifying sampling at the point of maximum curvature yields the best parameter recovery for inherently nonlinear functions (PMID 26609742). The problem is that the location of the breakpoint — precisely what we aim to estimate — is unknown beforehand, so the point of maximum curvature cannot be known in advance. Thus, “densify at maximum curvature” becomes circular: we must first have a rough breakpoint estimate to optimize sampling. This is exactly why two-stage designs exist — first a coarse sweep to locate the region, then densification within the candidate zone; otherwise any fixed grid is merely gambling.

    V. The Benefit of Adding Points Is Not Linear; Marginal Gains Must Be Calculated First

    Increasing repeated measurement frequency from once per year to four times per year reduces the standard error of the nonlinear trend regression coefficient by at most 98.9%; extending follow-up duration from 1 to 4 years reduces it by at most 96.9% (PMID 33644405). Latent growth models calculate even more finely: changing from 3 waves to 4 waves reduces the sample size required to detect slope variance by 35% (PMID 24219544). The common thread is diminishing returns — going from 3 to 4 waves saves one-third of the people, but subsequent additions save far less. Therefore, “expanding to 9 points” may not be cost-effective; perhaps “staying at 5 points but shifting their positions” is cheaper.

    VI. Two More Things That Will Steal Degrees of Freedom

    φ (phi) is computed from successive breath measurements, naturally carrying autocorrelation; this autocorrelation makes the effective sample size far smaller than the nominal sample size (PMID 39846001), so with five follow-up points the effective degrees of freedom may be less than four. On the other end, traditional segmented regression applied to cross-subject aggregated data produces spurious results; weighted segmented regression yields significantly narrower 95% confidence intervals (PMID 32606988). Interval estimation itself is method-sensitive: bootstrap methods outperform delta methods in small samples (PMID 19010644). And when follow-up points are few, the power of detecting jumps relies primarily on sample size, while the power of detecting slope changes relies primarily on the number of points (PMID 39110741) — this is practical: to test “a jump,” add people; to test “a slope change,” add points.

    VII. Criteria and Procedures

    Do not deliver a single MDBC; deliver a table. Using φ noise measured from ≥30 real subjects and drift between follow-ups as the base, synthesize breakpoint location sweeps across 1–12 months and amplitude sweeps across 1–10°. Run ≥500 repetitions on each of the current 5-point grid, the uniform 7/9/13-point grids, and compute breakpoint intervals via segmented regression plus bootstrap. Draw the position-amplitude boundary line where “interval half-width ≤ one follow-up interval.” The criteria have three tiers: (1) the current 5-point design leaves the feasible region empty for the 6–12 month segment ⇒ this segment is physically infeasible for alignment testing; must switch to continuous recording or extend the window to 24 months; (2) the feasible region covers only amplitudes ≥5° ⇒ conclusions can report only “whether it falls into the same resolvable interval,” not specific months; (3) even with uniform 9-point spacing, the feasible region remains empty ⇒ the problem lies not in the design but in φ’s noise floor; the study must switch to a time-structural endpoint. The 5° threshold and “one follow-up interval” in this text have no literature source; they must be replaced by values reverse-solved from实测 noise of this cohort.

    Evidence and Limitations

    Verified: Breakpoint estimation performs poorly when the true breakpoint is near boundaries or sample size is small (N=100) for both point and interval estimates (PMID 40585262); in nonlinear longitudinal models, increasing measurement frequency improves accuracy and efficiency, with sampling at maximum curvature yielding best recovery of inherently nonlinear functions (PMID 26609742); measurement frequency and follow-up duration can reduce standard errors of nonlinear trend coefficients by over 90% (PMID 33644405); increasing waves shows diminishing returns in reducing required sample size (PMID 24219544); autocorrelation reduces effective sample size below nominal (PMID 39846001); traditional segmented regression on cross-subject aggregated data produces spurious results (PMID 32606988); bootstrap intervals outperform delta methods in small samples (PMID 19010644); with limited follow-up points, jump detection power depends mainly on sample size while slope change power depends mainly on number of points (PMID 39110741).

    Partially Verified: Each of the above is an independent methodological conclusion, but no study has placed “the breakpoint location of the φ trajectory” and “the follow-up grid design” into the same simulation framework to calculate “what is the minimum detectable displacement in months for the current 0/1/3/6/12 design over the 6–12 month segment.” The table shape in this text is extrapolated from adjacent fields (AD biomarkers, ecological monitoring, policy evaluation discontinuous time series), not measured directly.

    Pending Verification: (1) After substituting real φ noise floor, is the feasible region for the current 5-point design truly empty over the 6–12 month segment? (2) How much improvement does uniform 9-point spacing offer over current 5-point in resolvable displacement during the later segment? (3) Under actual recruitment cost, does the two-stage design (coarse sweep then densification) outperform one-time full coverage? (4) Is the lag-1 autocorrelation of φ so high at the 5-point scale that effective degrees of freedom drop below four?

    Related Reading

    L1 Network Status: This topic has not yet entered the L1 corresponding network. A word-by-word search in CORRESPONDENCE_NETWORK.md for “breakpoint / change point / follow-up / power / sample size / detection / trajectory / time series / segmented / regression / noise / reliability / retest / MDC” yields zero hits; “phase” appears once, “breath” once, “dantian” once, “qi sensation” once — all as node table entries or observation method notes, not forming any edges. The edge table stops at E-050 (Sixty Jiazi Yin-Yang ↔ Frequency, open, original text of modern evidence column states “no measurable source yet”). The node table’s three domains only register “what to observe,” with no column registering “what time grid to use for observation, and what magnitude of change this grid can resolve” — hence this topic has neither corresponding edges nor corresponding independent variable columns. Evidence level capped at C by hard rules.
    Evidence Level: C (all methodological evidence is directly supported by PubMed, but the mapping itself — “φ trajectory breakpoint location × follow-up grid → resolvable displacement” — has no direct measurement, and this topic lacks corresponding edges in the L1 network, so overall it is downgraded to C — this text is a review of design feasibility, not verification of mechanism)
    References: PMID 40585262, PMID 26609742, PMID 33644405, PMID 24219544, PMID 39846001, PMID 32606988, PMID 19010644, PMID 39110741


    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.

  • 「够稳」和「够灵」可以是反向的:φ-SD 够不够格当协调稳定性终点,取决于它的 ICC 高在哪一半

    Evidence level: C

    判断一个指标能不能当终点,不该只问它「稳不稳」,而要问它稳在哪一半——是测量精度撑起来的稳,还是样本里人跟人差得远撑起来的稳。前者是资格,后者是幻觉。而这两件事,靠同一个 ICC 数字根本分不开。

    一、先把一个判决拆成两个

    「φ-SD 够不够格当协调稳定性终点」这句话里,其实压着两条互不等价的判决:一是它测得准不准,二是它动不动。临床测量学把这两件事分得很清楚——一个量表可以稳定得无可挑剔,却对任何干预都不动分毫。把「稳」当作唯一资格线,选出来的可能是一个稳得像块石头、因此什么都测不到的指标。所以这一节先把判决拆成两条:信度(重复测量是否一致)与反应性(真实变化能否被照出来)。后面五节分别看这两条线各自会怎么坑人,以及它们为什么会一起坑人。

    二、ICC 是比率,分子是「人和人差多少」

    ICC 的算法是「个体间方差 ÷(个体间方差 + 误差方差)」。分子的名字叫个体间方差,实际含义是采样人群有多杂,不是测量有多准。481 项影像组学研究的系统综述把这条说透了:高 ICC 可能主要反映样本里个体差异大(比如恶性肿瘤),并不保证量化精准;反过来,在同样精准的测量下,同质样本(如正常组织)会给出低 ICC(PMID 34603997)。这对 φ-SD 是直接的:只招新手,个体差异本来就窄,ICC 被压到不合格;混招新手与老手,ICC 被抬到合格。同一台带子、同一套算法,资格结论由抽样决定。ICC 估错的代价也有硬账——某研究按文献 ICC=0.02 规划样本,实测 0.12–0.17,有效样本缩水 17 倍(PMID 20167897)。

    三、同一件事,两篇论文给出相反结论

    信度能不能替代反应性,文献本身是打起来的。一派把量表逐项拆解,发现内部一致性系数与标准化反应均值(SRM)的相关高达 0.90–1.00,结论是两者可互换(PMID 15691385)。另一派在 516 名慢阻肺与膝损伤患者、79 个量表域上重做,得到的相关只有 0.10(95%CI −0.12–0.32),并指出前者的做法只在一个量表内部拆项,天然会看到同步下降;他们举出反例:SF-36 躯体疼痛 α=0.93,SRM 只有 0.29(PMID 15877824)。这场争论对本文的意义不是谁对,而是:任何人都不能仅凭信度高就宣布某指标够格当终点,尤其在跨组、跨条件比较时。φ-SD 恰好要跨的正是「修炼者 vs 新手」这种组。

    四、MDC 与 MID 是两条线,交叉才叫能测

    最小可检测变化(MDC)回答「多大变化超出噪声」,最小重要差异(MID)回答「多大变化才有意义」。前者由测量误差算出:MDC=1.96×√2×SEM;后者要靠外部锚定(PMID 16925807)。两者不能互相顶替——有研究直接测过:MDC 无法可靠替代 MID(PMID 19800198)。而现实里最常见的失败形态是 MDC 比 MID 还大:某工作功能量表 SDC 为 7.2–17 分,14 个 MIC 值里只有 1 个超过 SDC,换成高基线亚组则有 10 个超过(PMID 22974498)。翻成直白的话:这个量表在大多数人身上根本测不出「有意义的变化」。φ-SD 若走上同一条路,结局就是课题做完拿不到结论,而不是拿到否定结论。

    五、φ-SD 自己的麻烦:离散度的噪声底本来就厚

    φ-SD 还有个别人没有的额外负担:它本身是个离散度统计量,其误差同时被相位均值与呼吸次数推着走(跨零与样本量对圆形统计的偏差,已在别篇拆过)。承重的事实是:呼吸模式测量的信度报告极不统一,绝对容积随校准与体位漂移(PMID 42862511);分区贡献的重测信度虽高,却只在运动态验证过(PMID 35226168);而呼吸变异的时间结构(lag-1 自相关、谱斜率)重测性高,复杂的 ARFIMA 长记忆成分却很差(PMID 36618326)。这三条摆在一起指向一个具体处方:如果 φ-SD 不合格,别急着重招人,先试「换形态」——用变异的时间结构代替变异的大小。

    六、补救路径:平均 k 次抬得动信度,抬不动反应性

    信度是可以买来的。Spearman-Brown 预言公式给出:把 k 次测量取平均,ICC 从 ρ 抬到 kρ/[1+(k−1)ρ](PMID 28342902)。ρ=0.5 时 k=3 抬到 0.75,k=5 抬到 0.83。这条公式好用,但只作用于误差方差那一半,对「分子不够大」毫无办法——同质样本里平均一百次,ICC 照样低。更要紧的是它完全不碰反应性:反应性取决于干预打没打在指标敏感的那一段上,跟测量重复几次无关(PMID 15877824)。所以补救只能救「测不准」,救不了「测不到」。这一条要写进设计里,否则会变成用增加测量次数掩盖指标选错。

    七、判据与测法

    三格判据,前两格看信度、第三格看反应性:(一)ICC 高但 MDC95 大于可得组间差——不是好指标,是「稳而无用」,终点须换;(二)ICC 高且 MDC95 小于组间差的一半——够格,但必须同时报出抽样异质性(受试者年限跨度、φ-SD 的个体间 SD),否则 ICC 无法解释;(三)ICC 中等(0.5–0.75)——只能做组均值比较,四格判据里的「个体归类」全部降级。测法:≥40 人(刻意含新手与资深两档并分层报 ICC)、同体位、隔日×3 与隔周×2 共 5 次、每次≥10 分钟;同时算 φ-SD 与时间结构指标(lag-1 自相关、谱斜率)两套;另设一次已知能改变胸腹协调的呼吸训练作为反应性探针(效应量参照 PMID 25590447),看 φ-SD 是否真的动。判据里 0.75 与「一半」这两个数没有出处,须由本队列自己的 MID 反解替换。

    证据与局限

    已验证:ICC 会被样本异质性抬高、且高 ICC 不等于高精度(PMID 34603997,481 项研究系统综述);ICC 估错可使有效样本缩水一个数量级(PMID 20167897,实测);MDC 无法可靠替代 MID(PMID 19800198);MDC 大于 MID 导致「测不出有意义变化」的实证存在(PMID 22974498);平均 k 次测量提升信度的公式成立(PMID 28342902)。

    部分验证:信度与反应性的关系在文献中相互矛盾——一派报相关 0.90–1.00(PMID 15691385),一派报 0.10(PMID 15877824)。此处按冲突双方并列,不取单边结论。呼吸模式测量的信度报告不统一、φ 类参数日间稳定性差,属邻近证据外推(PMID 42862511、35226168、36618326),不是 φ-SD 自身的实测值。

    待验证:φ-SD 的 ICC、SEM、MDC95 至今无人测过;「φ-SD 对呼吸训练有反应」只是设计假设,未验证;0.75 门槛与「MDC95 < 组间差一半」两条线是借用通例,本领域无出处;若 φ-SD 的个体间方差主要来自年龄与体型而非「协调能力」,则第二格判据失效。

    相关阅读

    证据等级: C
    参考文献: PMID 34603997, PMID 20167897, PMID 15691385, PMID 15877824, PMID 16925807, PMID 19800198, PMID 22974498, PMID 42862511, PMID 35226168, PMID 36618326, PMID 28342902, PMID 25590447

    L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK.md 中「信度 / 重测 / 最小可检测 / MDC / ICC / 终点 / 离散度 / 标准差 / SD / 胸腹 / 稳定性 / 组内 / 气功 / 最小可检测」14 词全份检索均为 0 命中,仅「相位 / 节律」作为物理侧节点名、「呼吸率 / 呼吸模式」作为生理侧节点名各出现 1 次(备注仅「与节律耦合最直接」);边表止于 E-050(六十甲子纳音↔频率,状态 open,证据栏原文「尚无任何可测来源」)。故本文证据等级封顶 C,不作机制性断言。

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


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    \”Stable enough\” and \”responsive enough\” can be inversely related: whether φ-SD qualifies as the endpoint for coordination stability depends on which half of its ICC is high.

    Evidence level: C

    Determining whether an indicator qualifies as an endpoint should not ask only “Is it stable?” but rather “Stable in which half?”—is the stability supported by measurement precision or by large differences between individuals in the sample? The former constitutes qualification; the latter constitutes illusion. And these two matters cannot be distinguished by the same ICC number alone.

    I. Splitting One Judgment into Two

    The statement “Is φ-SD qualified to serve as a coordination stability endpoint” actually compresses two non-equivalent judgments: first, whether it measures accurately, and second, whether it responds. Clinical measurement science distinguishes these two matters clearly—a scale can be stable beyond reproach yet remain unresponsive to any intervention whatsoever. Treating “stability” as the sole qualification criterion may select an indicator that is stable like a stone and therefore measures nothing. Thus, this section splits the judgment into two: reliability (consistency across repeated measurements) and responsiveness (ability to reflect true change). The following five sections examine how each of these lines can lead us astray individually, and why they jointly lead us astray.

    II. ICC Is a Ratio; the Numerator Reflects “How Much People Differ”

    The ICC algorithm is “between-subject variance ÷ (between-subject variance + error variance).” The name of the numerator, between-subject variance, actually signifies how heterogeneous the sampled population is, not how precise the measurement is. A systematic review of 481 imaging radiomics studies clarifies this point: a high ICC may primarily reflect large individual differences within the sample (e.g., malignant tumors), without guaranteeing quantitative precision; conversely, under equally precise measurement conditions, homogeneous samples (e.g., normal tissue) yield a low ICC (PMID 34603997). For φ-SD, this is direct: recruiting only novices, whose individual differences are inherently narrow, depresses the ICC to an unqualified level; mixing novices and experts lifts the ICC to a qualified level. With the same tape and the same algorithm, the qualification conclusion is determined by sampling. The cost of misestimating ICC also has hard accounting—some studies planned samples based on literature ICC=0.02 but observed 0.12–0.17, resulting in an effective sample shrinkage of 17-fold (PMID 20167897).

    III. On the Same Matter, Two Papers Give Opposite Conclusions

    Whether reliability can substitute for responsiveness is itself contested in the literature. One camp decomposes scales item-by-item and finds that the correlation between internal consistency coefficient and standardized response mean (SRM) reaches 0.90–1.00, concluding that the two are interchangeable (PMID 15691385). Another camp retests across 516 patients with COPD and knee injury on 79 scale domains, obtaining a correlation of only 0.10 (95%CI −0.12–0.32), and points out that the former approach decomposes items only within a single scale, naturally yielding synchronized declines; they offer counterexamples: SF-36 bodily pain α=0.93 but SRM only 0.29 (PMID 15877824). The significance of this debate for our work is not which side is right, but that no one can declare an indicator qualified as an endpoint solely on the basis of high reliability, especially when comparing across groups and conditions. φ-SD precisely must cross the “practitioners vs. novices” group boundary.

    IV. MDC and MID Are Two Lines; Crossing Them Means It Can Be Measured

    The minimum detectable change (MDC) answers “How large a change exceeds noise,” while the minimum important difference (MID) answers “How large a change is meaningful.” The former is calculated from measurement error: MDC=1.96×√2×SEM; the latter relies on external anchoring (PMID 16925807). The two cannot substitute for each other—studies have directly tested this: MDC cannot reliably replace MID (PMID 19800198). The most common failure mode in reality is when MDC exceeds MID: for one work function scale, SDC ranges 7.2–17 points, and among 14 MIC values, only 1 exceeds SDC, whereas in a high-baseline subgroup, 10 exceed it (PMID 22974498). In plain terms: this scale fundamentally cannot detect “meaningful change” in most people. If φ-SD follows the same path, the outcome is that the study ends without a conclusion rather than with a negative conclusion.

    V. φ-SD’s Own Trouble: Its Noise Floor Is Already Thick

    φ-SD carries an additional burden not shared by other indicators: it is itself a dispersion statistic, whose error is simultaneously pushed by phase mean and breathing count (cross-zero bias and sample size bias in circular statistics, already dissected elsewhere). The bearing facts are: breathing pattern measurement reliability reports are highly inconsistent; absolute volume drifts with calibration and posture (PMID 42862511); retest reliability of partition contributions is high but verified only in motion (PMID 35226168); while the temporal structure of breathing variation (lag-1 autocorrelation, spectral slope) shows high retestability, complex ARFIMA long-memory components perform poorly (PMID 36618326). These three facts together point to a specific prescription: if φ-SD fails qualification, do not rush to recruit more people; instead try “changing form”—using variation in temporal structure rather than variation in magnitude.

    VI. Remediation Path: Averaging k Measurements Can Lift Reliability, Not Responsiveness

    Reliability can be bought. The Spearman-Brown prophecy formula gives: averaging k measurements lifts ICC from ρ to kρ/[1+(k−1)ρ] (PMID 28342902). When ρ=0.5, k=3 lifts it to 0.75, k=5 lifts it to 0.83. This formula is useful but acts only on the error-variance half, leaving “numerator not large enough” completely unsolved—even averaging 100 times in homogeneous samples still yields low ICC. More critically, it does not touch responsiveness at all: responsiveness depends on whether the intervention hits the sensitive segment of the indicator, regardless of how many times it is measured (PMID 15877824). Thus, remediation can only save “cannot measure accurately,” not “cannot measure at all.” This must be written into the design, otherwise it becomes masking indicator selection errors by increasing measurement repetitions.

    VII. Criteria and Measurement Methods

    Three-grid criteria, with the first two grids examining reliability and the third grid examining responsiveness: (I) ICC is high but MDC95 exceeds available between-group differences—not a good indicator, it is “stable yet useless”; the endpoint must be changed; (II) ICC is high and MDC95 is less than half the between-group difference—qualified, but must simultaneously report sampling heterogeneity (subject age span, φ-SD between-subject SD), otherwise ICC cannot be explained; (III) ICC is moderate (0.5–0.75)—only suitable for group mean comparison; all “individual classification” items in the four-grid criterion must be downgraded. Measurement method: ≥40 participants (deliberately including both novice and expert tiers, reporting ICC stratified), same posture, 3 days apart ×3 and 1 week apart ×2 totaling 5 times, each ≥10 minutes; simultaneously calculate φ-SD and temporal structure indicators (lag-1 autocorrelation, spectral slope) in two sets; additionally set up one breathing training known to alter thoraco-abdominal coordination as a responsiveness probe (effect size referenced PMID 25590447), to see whether φ-SD truly moves. The numbers 0.75 and “half” in the criteria have no source; they must be replaced by back-solving from this cohort’s own MID.

    Evidence and Limitations

    Verified: ICC is elevated by sample heterogeneity, and high ICC does not equal high precision (PMID 34603997, systematic review of 481 studies); misestimating ICC can shrink effective sample by an order of magnitude (PMID 20167897, observed); MDC cannot reliably replace MID (PMID 19800198); empirical evidence exists that MDC > MID leads to “cannot detect meaningful change” (PMID 22974498); the formula for averaging k measurements to improve reliability holds (PMID 28342902).

    Partially Verified: The relationship between reliability and responsiveness is contradictory in the literature—one camp reports correlation 0.90–1.00 (PMID 15691385), another reports 0.10 (PMID 15877824). Here both conflicting sides are listed side-by-side, without taking a unilateral conclusion. Breathing pattern measurement reliability reports are inconsistent; φ-type parameters show poor day-to-day stability, constituting extrapolation beyond proximal evidence (PMID 42862511, 35226168, 36618326), not φ-SD’s own measured values.

    To Be Verified: φ-SD’s ICC, SEM, MDC95 have never been measured; “φ-SD responds to breathing training” is only a design hypothesis, unverified; the 0.75 threshold and “MDC95 < half of between-group difference” are borrowed from convention, with no source in this field; if φ-SD’s between-subject variance stems primarily from age and body size rather than “coordination ability,” then the second-grid criterion fails.

    Related Reading

    Evidence Level: C
    References: PMID 34603997, PMID 20167897, PMID 15691385, PMID 15877824, PMID 16925807, PMID 19800198, PMID 22974498, PMID 42862511, PMID 35226168, PMID 36618326, PMID 28342902, PMID 25590447

    L1 Network Status: This topic has not yet entered the L1 corresponding network. In CORRESPONDENCE_NETWORK.md, all 14 terms—”reliability / retest / minimum detectable / MDC / ICC / endpoint / dispersion / standard deviation / SD / thoraco-abdominal / stability / within-group / qigong / minimum detectable”—have zero hits; only “phase / rhythm” appears once as a physical-side node name, and “breathing rate / breathing pattern” appears once as a physiological-side node name (note: only “most directly coupled with rhythm”); edge table stops at E-050 (Sixty Jiazi Yin-Yang ↔ Frequency, status open, evidence column original text “no measurable source yet“). Therefore, this article’s evidence level caps at C, making no mechanistic assertions.


    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.