5 个点里到底有几个独立点:φ 轨迹的 lag-1 自相关有多高,以及它把 MDBC 表整体放宽几成

Written by

in

Language

Evidence level: C

一句话核心命题:5 个随访点 ≠ 5 个独立观测——φ 是逐次呼吸算出的统计量,同一场记录内的呼吸本就高度相关,随访间又共享体位、时段、操作员这些慢变量;自相关把有效样本数从 5 打回 2-3,整张 MDBC 表就得按量级放宽,而不是微调。

正文

一、问题:整张表都建在一个没人验过的假设上

丹田呼吸的纵向研究常用 0、1、3、6、12 个月五点随访方案,每次测呼吸相位离散度 φ。上一篇文章(#1184)用这 5 个点算出了「最小可检测拐点位移」(MDBC)表,每一格写着”这个幅度、这个时间点的位移,最少要几次随访才能分辨”。整张表都默认了一件事:5 个随访点 = 5 个独立观测。

这个假设恰恰最可疑。φ 不是”点一下出一读数”的简单测量——它是从一段连续呼吸记录里算出来的统计量,本身已内含几十上百次呼吸的信息。同一次记录内的呼吸之间本就高度相关(这一口和上一口只差几秒,体位、潮气量几乎没变),不同随访的 φ 又共享同一批慢变量:同一把椅子、同一个时段、同一个操作员。相关性一高,”5 个点”的名义信息量就要大打折——不是微调,是量级上的放宽。

二、机制:自相关把独立样本数打成几分之几

“自相关使有效样本数小于名义样本数”有明确的量化框架。一组 n 个观测、lag-1 自相关为 ρ 的序列,有效样本数近似为:

n_eff = n × (1 − ρ) / (1 + ρ)

把 n=5 代进去:ρ=0 时 n_eff=5(假设成立);ρ=0.3 时 n_eff≈2.7;ρ=0.5 时 n_eff≈1.7;ρ=0.7 时 n_eff≈0.9。只要随访点间 lag-1 自相关到 0.3-0.5,5 个点的有效信息就只有 2-3 个独立观测。

对生理信号时序做相关性检验的研究反复确认:不考虑自相关会系统性高估信息量、低估置信区间宽度(PMID 39846001)。纵向试验设计里,自相关结构直接改写样本量公式——Kirby 等用 γ^s·θ 双参数相关模型证明:重复测量的自相关越强,达成同样功效所需的等效独立样本越少,忽略自相关会让功效预算系统性乐观(PMID 7913674)。更晚近的整群随机试验方法学把这条推得更细:个体自相关是功效公式里的显式输入项,必须单独建模(PMID 32133688);估计这些相关参数还有专门教程,且多数已发表试验根本不上报这些参数(PMID 37196320)——”5 个点算 5 个独立观测”的普遍做法,几乎没经过自查。

三、但这不能拿文献系数直接代入

那去查文献找典型 lag-1 自相关代进去算,行吗?不行。呼吸变异的相关结构并非单一常系数。对感觉运动变异做时间结构分析发现:简单短期结构(相邻观测相关)重测性高、接近稳定特质,而复杂长期记忆成分重测性差(PMID 36618326)。”这 5 个点之间的相关”既含稳定的短程相关、又含飘忽的长程成分,对有效样本量的贡献完全不同。用文献里任何”典型值”代入,都是在用一个未必适用本人群、本协议、本指标的常数,去决定一整张表要不要放宽——这正是本主题必须实测的原因。

四、标定设计:一次零额外招募的纯相关结构测定

标定与 #1184 主实验共用受试者,零额外招募:

  1. ≥40 人按 0/1/3/6/12 月五点各测一次 φ(每次 ≥10 分钟);
  2. 逐人拟合 AR(1) 取 lag-1 ρ,算 n_eff = 5 × (1 − ρ) / (1 + ρ);
  3. 把 n_eff 代回 #1184 仿真框架,重算 6-12 月段拐点区间半宽,与按名义 n=5 的结果比对,得放宽倍数。

小样本下区间估计方法本身会偏(PMID 19010644),故 n_eff 扣减后须用自助法重算区间。

五、三档判据

判据一(独立假设崩塌):若 lag-1 中位数 >0.4 且 n_eff 中位数 <3,则”5 点 = 5 独立观测”不成立,整张 MDBC 表须按 n_eff 重算并整体放宽,此前任何按名义点数做的功效预算一律作废。

判据二(表可用但须附注):若 lag-1 落在 0.1-0.4、n_eff 落在 3-4,则表保留,但每格须附 n_eff 与放宽倍数,报结论时只能报”落在哪个可分辨区间”,不能报具体月数。

判据三(担忧撤销):若 lag-1 中位数 <0.1 且 n_eff >4,则独立假设近似成立,原表可原样使用,本文担忧全部撤销。

六、对易学—科学桥接的意义

丹田呼吸研究要回答”长期修炼到底在生理上改变了什么”,这依赖纵向随访的可靠性,而可靠性又依赖”每个随访点贡献多少独立信息”。若 5 个点实际只有 2-3 个独立观测,此前按 5 点算出的”拐点出现时间””最小可检测位移”都要重新标定——这不是否定丹田呼吸研究,而是把它从”做了 5 次测量”的乐观拉回”只有 2-3 个独立证据”的现实。测量学这一步不补,后面所有”几年练出了什么”的结论都站不稳。

证据与局限

已验证: 自相关使时序数据的有效样本数小于名义样本数,且有成熟的 n_eff 近似公式(PMID 39846001);重复测量自相关强度直接改写纵向试验的样本量公式(PMID 7913674);个体自相关是纵向整群试验功效计算的显式输入项(PMID 32133688);估计这些相关参数已有标准教程(PMID 37196320)。

部分验证: 呼吸/感觉运动变异的时间结构重测性高、但复杂长记忆成分重测性差(PMID 36618326),提示相关结构非单一常系数——这是”必须实测、不能借文献”的依据,但该研究并非丹田呼吸人群的 φ 轨迹数据。

待验证: 丹田呼吸 φ 轨迹在 0/1/3/6/12 月五点方案下的 lag-1 自相关真实值(本文全部判据都建立在这个尚未测过的数字上);n_eff 扣减后 6-12 月段区间半宽的放宽倍数;小样本下自助法重算的区间是否稳定。

相关阅读

证据等级: C(方法学设计裁决:机制有文献支撑,但判据所依赖的 φ 轨迹自相关实测值尚未测出,属测量学标定类探索)
参考文献: PMID 39846001, PMID 7913674, PMID 32133688, PMID 37196320, PMID 36618326, PMID 19010644, PMID 40119099

L1 网络状态: 该主题尚未进入 L1 对应网络——「丹田参数 ↔ 测量学维度」簇目前无任何以”随访点 lag-1 自相关 / 有效样本数 n_eff / 独立观测假设”为源或目标的边,已登记入 science-yixue-l1-gaps.md(簇第 19 条)。


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 many independent points are there among the five: what is the lag-1 autocorrelation of the φ trajectory, and by what factor does it relax the entire MDBC table?

Evidence level: C

Core proposition in one sentence: 5 follow-up points ≠ 5 independent observations—φ is a statistic computed from successive breaths; breaths within the same recording are inherently highly correlated, and follow-ups share slow variables such as posture, time of day, and operator. Autocorrelation reduces the effective sample size from 5 back to 2–3, so the entire MDBC table must be scaled up by order of magnitude rather than fine-tuned.

Main Text

I. Problem: The Entire Table Rests on an Unverified Assumption

Longitudinal studies of dantian breathing commonly use a five-point follow-up schedule at 0, 1, 3, 6, and 12 months, measuring the breathing phase dispersion φ each time. The previous article (#1184) used these 5 points to compute the “Minimum Detectable Change in Breathing Phase” (MDBC) table, where each cell states “this magnitude, at this time point, requires at least this many follow-ups to distinguish.” The entire table assumes: 5 follow-up points = 5 independent observations.

This assumption is precisely the most suspicious part. φ is not a simple measurement that yields one reading per tap—it is a statistic derived from a continuous breathing record, inherently incorporating information from dozens or hundreds of breaths. Breaths within the same recording are already highly correlated (each breath differs from the previous by only a few seconds, with posture and tidal volume nearly unchanged). Moreover, φ values across different follow-ups share the same batch of slow variables: the same chair, the same time slot, the same operator. High correlation means the nominal information content of “5 points” is heavily discounted—not fine-tuned, but scaled down by orders of magnitude.

II. Mechanism: Autocorrelation Reduces the Effective Sample Size to a Fraction

“The effective sample size is less than the nominal sample size due to autocorrelation” has a clear quantitative framework. For a sequence of n observations with lag-1 autocorrelation ρ, the effective sample size is approximately:

n_eff = n × (1 − ρ) / (1 + ρ)

Substituting n = 5: if ρ = 0, n_eff = 5 (the assumption holds); if ρ = 0.3, n_eff ≈ 2.7; if ρ = 0.5, n_eff ≈ 1.7; if ρ = 0.7, n_eff ≈ 0.9. Once the lag-1 autocorrelation between follow-up points reaches 0.3–0.5, the effective information from 5 points amounts to only 2–3 independent observations.

Studies on temporal correlation in physiological signals repeatedly confirm: ignoring autocorrelation systematically overestimates information content and underestimates confidence interval width (PMID 39846001). In longitudinal trial design, the autocorrelation structure directly rewrites the sample size formula—Kirby et al. used a two-parameter correlation model γ^s·θ and proved: the stronger the autocorrelation of repeated measures, the fewer equivalent independent samples are needed to achieve the same power; ignoring autocorrelation makes the power budget systematically optimistic (PMID 7913674). More recent cluster randomized trial methodology refines this further: individual autocorrelation is an explicit input in the power formula and must be modeled separately (PMID 32133688); there are dedicated tutorials for estimating these correlation parameters, yet most published trials do not report these parameters (PMID 37196320)—the common practice of “calculating 5 independent observations from 5 points” has almost never been self-checked.

III. But We Cannot Directly Plug Literature Coefficients

So, why not look up typical lag-1 autocorrelation values in the literature and plug them in? No. The correlation structure of breathing variability is not a single constant coefficient. Time-structure analysis of somatosensory-motor variability reveals: simple short-term structure (adjacent observations) has high test-retest reliability and approximates stable traits, whereas complex long-memory components have poor test-retest reliability (PMID 36618326). “The correlation among these 5 points” includes both stable short-range correlations and fluctuating long-range components, contributing differently to the effective sample size. Using any “typical value” from the literature to decide whether the entire table must be scaled is like applying a constant that may not apply to this specific population, protocol, or metric—this is precisely why this topic requires empirical calibration.

IV. Calibration Design: Pure Autocorrelation Structure Measurement with Zero Additional Recruitment

The calibration shares subjects with the main experiment in #1184, requiring zero additional recruitment:

  1. ≥40 participants each measure φ at the five points (0/1/3/6/12 months), with each measurement lasting ≥10 minutes;
  2. Fit an AR(1) model for each individual to obtain lag-1 ρ, then compute n_eff = 5 × (1 − ρ) / (1 + ρ);
  3. Plug n_eff back into the #1184 simulation framework to recompute the half-width of the inflection interval for the 6–12 month segment, comparing it with the nominal n = 5 result to obtain the broadening factor.

Under small sample sizes, interval estimation methods themselves are biased (PMID 19010644), so after subtracting n_eff, bootstrap resampling must be used to recompute intervals.

V. Three-tier Criteria

Criterion 1 (Collapse of the independence assumption): If the median lag-1 autocorrelation > 0.4 and the median n_eff < 3, then “5 points = 5 independent observations” fails; the entire MDBC table must be recalculated using n_eff and scaled up accordingly, invalidating any power budgets previously made based on nominal point counts.

Criterion 2 (Table usable but with annotations): If lag-1 falls between 0.1–0.4 and n_eff lies between 3–4, the table remains valid but each cell must include n_eff and the broadening factor; conclusions reported can only state “falls within which detectable interval,” not specific months.

Criterion 3 (Concerns withdrawn): If the median lag-1 autocorrelation < 0.1 and n_eff > 4, then the independence assumption approximately holds; the original table can be used as-is, and all concerns are withdrawn.

VI. Significance for Bridging Yijing Studies and Science

Dantian breathing research aims to answer “what physiological changes result from long-term cultivation,” which depends on the reliability of longitudinal follow-ups, in turn depending on “how much independent information each follow-up point contributes.” If the 5 points actually yield only 2–3 independent observations, previously computed “inflection time” and “minimum detectable displacement” must be recalibrated—this does not negate dantian breathing research but pulls it back from the optimism of “5 measurements” to the reality of “only 2–3 independent pieces of evidence.” Without this metrological step, all subsequent conclusions about “what was achieved after several years” will be unstable.

Evidence and Limitations

Verified: Autocorrelation reduces the effective sample size of temporal data below the nominal sample size, with established n_eff approximation formulas (PMID 39846001); the strength of autocorrelation in repeated measures directly rewrites the sample size formula for longitudinal trials (PMID 7913674); individual autocorrelation is an explicit input in power calculations for longitudinal cluster trials (PMID 32133688); standard tutorials exist for estimating these correlation parameters (PMID 37196320).

Partially verified: The test-retest reliability of breathing/somatosensory-motor variability time structures is high, but complex long-memory components have low reliability (PMID 36618326), indicating the correlation structure is not a single constant coefficient—this is the basis for “must be measured empirically, cannot borrow from literature,” though this study does not provide φ trajectory data for the dantian breathing population.

To be verified: The true lag-1 autocorrelation values of the dantian breathing φ trajectory under the 0/1/3/6/12 month five-point schedule (all criteria in this paper depend on this yet-to-be-measured number); the broadening factor of the interval half-width for the 6–12 month segment after n_eff subtraction; whether bootstrap-resampled intervals remain stable under small samples.

Related Reading

Evidence level: C (methodological design decision: the mechanism is supported by literature, but the φ trajectory autocorrelation values required for the criteria have not yet been measured; this belongs to metrological calibration exploration).
References: PMID 39846001, PMID 7913674, PMID 32133688, PMID 37196320, PMID 36618326, PMID 19010644, PMID 40119099

L1 network status: This topic has not yet entered the L1 corresponding network—the “dantian parameters ↔ metrological dimensions” cluster currently has no edges sourced from or targeting “follow-up point lag-1 autocorrelation / effective sample size n_eff / independent observation assumption,” and has been registered in science-yixue-l1-gaps.md (cluster item 19).


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 *