Δ 的方差里,有多少是「隔天再测」抖出来的:个体内与个体间方差成分拆分,与年限斜率还剩多少可测空间

Written by

in

Language

Evidence level: C

「练得越久 Δ 越窄」这句话要成立,先得回答一个问题:同一个人隔天再测一次,Δ 能抖出多大动静?这个数字一天没测出来,年限斜率就一天说不清「可测空间」有多大。

一、问题的形状:两个「已知数字」,没有一个来自 φ 或 %AB

上一轮(#1181)论证年限斜率的功效陷阱时,引了两个关键数字:场次之间的变异平均相当于个体间变异的 25%–50%,与横断面比较中跨越 12–27 岁年龄范围的差异相当(PMID 15014087);以及个体斜率估计的方差由两部分构成——个体内变异(测量误差、随访长度与频率决定)与真实的个体间斜率差异(PMID 2362980)。

这两个数字支撑了「场次噪声能伪造十几年功力」的核心担忧。但这里有一个必须戳破的缺口:这两个数字,没有一个是在胸腹相位角 φ 或腹式占比 %AB 上测出来的。PMID 15014087 测的是知觉—运动任务,PMID 2362980 是纵向研究功效的通用统计框架——它们给出的是「认知任务世界」的方差结构,不是「呼吸力学参数世界」的。

而呼吸力学参数自己的方差结构,可能完全不同。同一批早产儿在静息睡眠里,φ 的个体内变异系数(intrasubject CV)可以横跨 11%–123%(PMID 24716708),而且作者明确归因于胸腹相位关系本身的不稳定,不是记录或分析方法造成的。如果 φ/%AB 的个体内成分占总量 60% 以上,那么年限斜率的可测空间几乎为零——这不是”样本量不够”能救的,是分子结构的问题。

二、为什么必须做方差成分拆分,而不是直接算功效

把总方差直接当功效公式的分母,等于假设「总方差 = 个体间方差」。但母文已经点破:横断面四层设计把个体内成分推到了最大——每人只测一次,个体内抖动和个体间差异在同一个数里无法分离,真实斜率差异被埋在下面。

方差成分拆分(variance component decomposition)做的正是这件事:用线性混合模型把总方差拆成 个体间(between-person)、个体内/场次(within-person/session)、残余(residual) 三层。拆完之后的关键输出是两个数:

  1. 个体内成分占总方差的比例——它直接决定 ICC 的高低。个体内占比 >60%(ICC <0.4)时,按年限分层的任何斜率结论都落在噪声里。
  2. 设计效应(design effect = 1 + (m−1)×ICC)——它决定「重复测 m 次取均值」能把有效样本量放大多少。设计效应越大,同样的 n=120 能检出的斜率越小。

这两个数互为表里:ICC 是方差的比值,design effect 是这个比值在设计里的放大倍数。不拆方差,两个都算不出来;而这两个数在 φ/%AB 上至今没有任何实测——这就是「先算功效」还是「先拆方差」的分水岭。

三、φ 的个体内 CV 11%–123%:方差结构可能完全不同于认知任务

PMID 24716708 的量级提示值得单独说。同一批受试者、同一场记录、同一个算法口径,个体内 CV 可以从 11% 到 123%——这不是”测量不准”,而是胸腹相位关系本身就是个不稳定的生理量(作者结论原话:variability arises from the instability of the rib cage and abdominal phase relationship, not from the recording and analytical methods)。

把这个数字放进 #1181 的框架里,后果是具体的:如果 φ 的个体内成分在成人丹田练习者身上也占到 60% 以上,那么「场次噪声等量于十几年」在 φ 上字面成立——而 #1181 引用的 25%–50% 只是认知任务的数字,在 φ 上可能更高。反过来,若实测个体内占比 <30%(ICC >0.7),”噪声等量于十几年”在 φ 上不成立,#1181 的核心担忧可撤回。这正是方差成分拆分这个纯测量学实验的价值:它能把「担忧」从修辞变成可判定的数。

四、设计:纯方差成分研究,零干预、零分组

这个研究不需要干预、不需要分组、不涉及气功练习本身,只需要一份干净的重复测量数据:

  • ≥40 名健康成人,覆盖新手到 10 年以上(不按年限层等分,连续覆盖即可)
  • 同体位、同倾角(体位和倾角会改变 φ,必须锁死)
  • 隔日 ×3 + 隔周 ×2,共 5 次,每次 ≥10 分钟
  • 每次取稳定段算 %AB 与 φ,互相关法、固定单一算法口径(口径不同会让 φ 差出几十度)
  • 用线性混合模型拆三层方差,算个体内占比、ICC、design effect
  • 最后据拆分结果反推「四层年限设计 n=120 时的最小可检出斜率」

五、核心判据:三档裁决

  1. 个体内成分 >60%(ICC <0.4):按年限分层的任何斜率结论都落在噪声内——现有横断面证据整体降级为「未可解释」,功效计算必须改用 ICC 校正后的有效样本量(n_eff = n / design effect)。
  2. 个体内占 30%–60%(ICC 0.4–0.7):斜率可测,但必须报 design effect 与有效 n,且重复测量次数须 ≥3 次取均值才允许进回归——单次测量不能进。
  3. 个体内 <30%(ICC >0.7):§二 所引的「噪声等量于十几年」在 φ 上不成立,#1181 的核心担忧可撤回,年限斜率按常规功效计算即可。

六、与 #1255 的关系:一个数据的两次使用

1254(方差成分拆分)与 #1255(retest gain 饱和曲线)共用同一批 40 人 × 5 次数据,零额外采集成本。方差成分拆分回答「Δ 里多少是隔天抖的」,retest gain 回答「抖的方向是不是单向抬升、几次饱和」。两者合起来才是完整的测量学体检:一个管方差结构(ICC 决定可测空间),一个管系统性偏移(retest 决定基线定义)。只拆方差不测 retest,会把”练习效应”当成”个体差异”;只测 retest 不拆方差,会把”噪声底”当成”可测空间”。

证据与局限

已验证: 场次间变异在知觉—运动任务上相当于个体间的 25%–50%、与跨 12–27 岁横断面差相当(PMID 15014087);斜率方差由个体内与个体间两成分构成(PMID 2362980);φ 在静息睡眠中个体内 CV 横跨 11%–123% 且源于相位关系不稳而非测量方法(PMID 24716708);纵向设计中 retest gain 与年龄-队列差异的混淆已被仿真证实(PMID 21639642)。

部分验证: 「场次噪声等量于十几年」在认知任务域成立,但在 φ/%AB 上尚无任何直接测量——25%–50% 与 11%–123% 这两个数字来自不同人群与不同任务,无法直接外推到成人丹田练习者。

待验证: φ/%AB 的个体内/个体间方差成分拆分(≥40 人 × 5 次、混合模型三层拆分);个体内占比三档判据(>60% / 30–60% / <30%)的落地数值;ICC 校正后四层年限设计的实际最小可检出斜率;design effect 在 φ 上的实测值。

L1 网络状态: 「丹田」概念在 L1 对应网络中有 51 篇文章、最高证据级 A(代表 PMID 10593633,腹部呼吸体腔共振);但 「方差成分 / 个体内-个体间拆分 / ICC / design effect」这一测量学维度整条边缺失——CORRESPONDENCE_NETWORK.md 节点表三域只登记「观测什么量」,无任何位置承载「这个量能被估到多准、方差里多少是隔天再测抖出来的」;边表止于 E-050(六十甲子纳音↔频率,L1、open,现代证据栏「尚无任何可测来源」)。本条已登记 L1 gap 表(第 70 条,「丹田参数 ↔ 测量学维度」簇第 13 条)。

相关阅读

证据等级: C
参考文献: PMID 15014087, PMID 2362980, PMID 24716708, PMID 21639642


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

How much of the variance in Δ is due to ‘retest after one day’: decomposition of within-subject and between-subject variance components, and how much measurable space remains fo…

Evidence level: C

For the statement “the longer one practices, the narrower Δ becomes” to hold, we must first answer: if the same person is tested again the next day, how much movement can Δ produce? Until this number is measured, we cannot clarify how large the “measurable space” is for the tenure slope.

I. The Shape of the Problem: Two “Known Numbers,” Neither from φ or %AB

In the previous round (#1181), when arguing about the efficacy trap of the tenure slope, two key numbers were cited: between-session variability averages 25%–50% of between-person variability, and is comparable to differences across the 12–27 year age range in cross-sectional comparisons (PMID 15014087); and the variance of individual slope estimates consists of two components—within-person variability (determined by measurement error, follow-up length, and frequency) and true between-person slope differences (PMID 2362980).

These two numbers support the core concern that “session noise can fake decades of practice.” However, there is a gap that must be broken here: neither of these numbers was measured on the thoraco-abdominal phase angle φ or the abdominal ratio %AB. PMID 15014087 measures perceptual-motor tasks, and PMID 2362980 is a general statistical framework for longitudinal study efficacy—they provide the variance structure of the “cognitive task world,” not the “respiratory mechanics parameter world.”

The variance structure of respiratory mechanics parameters themselves may be entirely different. In the same cohort of preterm infants at rest during sleep, the within-subject coefficient of variation (CV) for φ can span 11%–123% (PMID 24716708), and the authors explicitly attribute this to instability in the thoraco-abdominal phase relationship itself, not to recording or analytical methods. If the within-person component of φ/%AB accounts for more than 60% of the total, then the measurable space for the tenure slope is nearly zero—this cannot be saved by “insufficient sample size”; it is a problem of molecular structure.

II. Why Variance Component Decomposition Is Necessary, Rather Than Direct Power Calculation

Treating total variance directly as the denominator of the power formula assumes “total variance = between-person variance.” But the original text already reveals: the cross-sectional four-level design pushes the within-person component to its maximum—each person is measured only once, so within-person jitter and between-person differences cannot be separated in the same number, burying true slope differences underneath.

Variance component decomposition does exactly this: using linear mixed models, it splits total variance into between-person, within-person/session, and residual layers. The key outputs after decomposition are two numbers:

  1. The proportion of within-person component to total variance—this directly determines the ICC. When the within-person proportion exceeds 60% (ICC <0.4), any slope conclusion stratified by tenure falls into noise.
  2. The design effect (design effect = 1 + (m−1)×ICC)—this determines how much the effective sample size is amplified by “taking the mean of m repeated measurements.” The larger the design effect, the smaller the slope detectable with the same n=120.

These two numbers are interdependent: ICC is the ratio of variances, and the design effect is the amplification factor of this ratio in the design. Without decomposing variance, neither can be calculated; and these two numbers have never been directly measured on φ/%AB—this is the watershed between “calculating power first” and “decomposing variance first.”

III. φ’s Within-Subject CV of 11%–123%: Variance Structure May Differ Completely from Cognitive Tasks

The magnitude reported in PMID 24716708 warrants separate discussion. For the same subjects, the same recording session, and the same algorithmic scope, the within-subject CV ranges from 11% to 123%—this is not “imprecise measurement,” but rather the thoraco-abdominal phase relationship itself is an unstable physiological quantity (the authors’ conclusion: “variability arises from the instability of the rib cage and abdominal phase relationship, not from the recording and analytical methods”).

Placing this number into the framework of #1181 yields concrete consequences: if the within-person component of φ accounts for more than 60% in adult dantian practitioners, then “session noise equals decades” holds literally for φ—and the 25%–50% cited in #1181, being from cognitive tasks, may be even higher for φ. Conversely, if direct measurement shows the within-person proportion is less than 30% (ICC >0.7), then “noise equals decades” does not hold for φ, and the core concern of #1181 can be withdrawn. This is precisely the value of variance component decomposition as a pure measurement experiment: it transforms “concern” from rhetoric into determinable numbers.

IV. Design: Pure Variance Component Study, Zero Intervention, Zero Grouping

This study requires no intervention, no grouping, and does not involve qigong practice itself; it only needs clean repeated measurement data:

  • ≥40 healthy adults, covering novices to over 10 years (not evenly stratified by tenure, but continuously covered)
  • Same posture, same tilt angle (posture and tilt angle affect φ, so these must be locked)
  • Every other day ×3 + every week ×2, totaling 5 sessions, each ≥10 minutes
  • Calculate %AB and φ from stable segments each time, using mutual correlation method with a fixed single algorithmic scope (different scopes would cause φ to differ by tens of degrees)
  • Use linear mixed models to decompose the three layers of variance, calculating the within-person proportion, ICC, and design effect
  • Finally, reverse-calculate from the decomposition results to determine “the minimum detectable slope for a four-level tenure design with n=120”

V. Core Criteria: Three-Tier Verdicts

  1. Within-person component >60% (ICC <0.4): Any slope conclusion stratified by tenure falls into noise—existing cross-sectional evidence overall downgrades to “uninterpretable,” and power calculation must use ICC-corrected effective sample size (n_eff = n / design effect).
  2. Within-person accounts for 30%–60% (ICC 0.4–0.7): Slope is measurable, but the design effect and effective n must be reported, and repeated measurements must average ≥3 times before entering regression—single measurements are not allowed.
  3. Within-person <30% (ICC >0.7): The “noise equals decades” claim cited in §II does not hold for φ, so the core concern of #1181 can be withdrawn, and the tenure slope can be calculated using conventional power methods.

VI. Relationship with #1255: Two Uses of One Dataset

1254 (variance component decomposition) and #1255 (retest gain saturation curve) share the same dataset of 40 people × 5 sessions, with zero additional acquisition cost. Variance component decomposition answers “how much of Δ is due to testing again the next day,” while retest gain answers “does the direction of ‘jitter’ unidirectionally rise, and does it saturate after several times.” Together they constitute a complete measurement examination: one manages variance structure (ICC determines measurable space), and the other manages systematic shift (retest determines baseline definition). Decomposing variance without measuring retest would mistake “practice effect” for “individual difference”; measuring retest without decomposing variance would mistake “noise floor” for “measurable space.”

Evidence and Limitations

Verified: Between-session variability in perceptual-motor tasks equals 25%–50% of between-person variability and matches cross-sectional differences across 12–27 years (PMID 15014087); slope variance consists of within-person and between-person components (PMID 2362980); φ’s within-subject CV spans 11%–123% at rest during sleep, attributed to phase relationship instability rather than measurement methods (PMID 24716708); retest gain confounding with age-cohort differences in longitudinal designs has been simulated and confirmed (PMID 21639642).

Partially Verified: “Session noise equals decades” holds in the cognitive task domain, but no direct measurement exists for φ/%AB—the figures 25%–50% and 11%–123% come from different populations and tasks, so they cannot be directly extrapolated to adult dantian practitioners.

Pending Verification: Decomposition of within-person/between-person variance components for φ/%AB (≥40 people × 5 sessions, three-layer mixed model); the three-tier thresholds for within-person proportion (>60% / 30–60% / <30%); the actual minimum detectable slope for the four-level tenure design after ICC correction; and the measured design effect for φ.

L1 Network Status: The concept of “Dantian” corresponds to 51 articles in the L1 network, with the highest evidence level A (representing PMID 10593633, abdominal breathing cavity resonance); however, the entire measurement dimension of “variance component / within-person-between-person decomposition / ICC / design effect” is missing—the CORRESPONDENCE_NETWORK.md node table’s three domains only record “what quantity is observed,” with no position carrying “how accurately this quantity can be estimated, and how much of the variance is due to testing again the next day”; the edge table stops at E-050 (Sixty Jiazi Na Yin ↔ Frequency, L1, open, modern evidence column states “no measurable source yet”). This gap has been registered in the L1 gap table (item 70, “Dantian parameters ↔ measurement dimension” cluster item 13).

Further Reading

Evidence Level: C
References: PMID 15014087, PMID 2362980, PMID 24716708, PMID 21639642


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.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *