「体位占比降了」不等于「练得更稳」:方差分量结构的组间比较有一个比率的坑

Written by

in

Language

Evidence level: C

把总方差拆成「体位、带位、操作员、个体、残差」之后,最诱人的一步是拿五个占比去比修炼者与新手。这一步几乎肯定会出错,因为占比是比率:分子降、分母升,都能让它降。

正文

一、占比是比率,不是量

五个变异源占比加起来必须等于 1。这个约束带来两个后果:任何两个占比天然负相关;「体位占比从 40% 降到 25%」本身没有信息,因为体位方差可能降了 37%,也可能一分没动、只是个体间方差涨了 60%。前者是”人对姿势不敏感了”,后者是”人跟人差得更远了”,方向完全相反,却给出同一个数字。

成分数据在统计上早有定论:算术均值给不出单纯形内的结果,须先做对数比变换(PMID 38806627);份额数据里绝对量上升也可能表现为份额下降(PMID 28882426)。所以要报的是绝对方差分量,不是占比。

二、占比不能直接做组间推断

占比的抽样分布依赖五个分量共享的分母,五个都不独立。对占比直接跑 t 检验,等于假装分母是常数。临床试验里方差齐性检验本身有专门方法——Jackknife、Cochran 一类在重尾与偏态数据下更稳(PMID 37143826);混合模型里若方差随组别变化,正确做法是直接给方差建模(location-scale 型扩展),而非只检验固定效应(PMID 32666394)。

「两组的方差结构不同」是关于方差函数的假设,得在方差那一层测,不能靠占比描述性对比。

三、体位效应本身就不是人人都有

健康人可能根本没有体位效应可降。23 例重度 COPD 患者坐位出现胸廓异步、仰卧消失,上胸廓与腹部坐位同步而仰卧强异步;12 例年龄匹配健康对照,两个体位都没有任何区间异步(PMID 23412901)。

如果健康修炼者与新手的体位方差都接近零,”修炼者体位占比更低”就是分母造出来的假象。这是典型天花板效应:指标在目标人群没有变异空间,组间比较毫无意义。设计前必须先确认该人群的体位方差显著大于零。

四、这么小的设计,方差分量本身估不准

体位只有 2 个水平,操作员往往 1 到 2 个。测量次数少、随访短、残差大方差,都会扩大”非正则”参数区域,使负方差与边界估计变成常态而非异常(PMID 38695394);不同估计量的负估计概率差别很大(PMID 26843971)。

推断更麻烦:参数边界上似然比、Wald、Score 的渐近分布失效,须改置换法或贝叶斯法(PMID 31460683、PMID 16474914)。加上评判者数量少本身就是 ICC 偏倚的主因(PMID 33818118),三个模态的置信区间很可能宽到跨零——”有差异”和”无差异”都测不出来。

五、训练降的是哪种「稳」

技能研究确实支持”专长的人更稳”,但要看清是哪种稳。9 名外科医生各测 4 次:初级住院医分数标准差 5.8(满分 30),高级住院医与主治都是 1.8(PMID 31517685);模拟器训练中练习变异每降 1%,后测升 3.8 分、活体迁移升 3.0 分(PMID 30898374)。

但这些都是同一个人的重复测量,属交叉设计里的”残差 + 交互”,不是”体位”这个设计因子。手枪射击的对照更说明问题:专家 COP 位移更小、多尺度熵更低,可两组上身-枪相对相位都接近同相,真正差别在结构——专家是枪滞后于身,新手是枪领先于身(PMID 28628398)。稳不等于每个维度都变小,而是结构换了位置。

六、组间比较的根本麻烦:不是随机分组

修炼者与新手不是随机分配的,是自选择。能练十年的人,本来就更坐得住、呼吸控制更好——横断面设计分不清”练出来的”和”能练下去的”。志愿者偏差是系统性的:UK Biobank 参与者比一般人群更健康、社会经济状况更好(PMID 28641372)。

另有一个读法错误须提防。SD-OCT 测儿童与成人:两组视网膜厚度均值没有差异,但儿童 COV 是成人的 3 到 4 倍、ICC 显著更低(0.73–0.99 vs 0.986–0.993,PMID 26808961)。同样的均值、差很多的精度——这种”组间精度差异”真实存在,却最容易被误写成”这组人能力更强”。

反转:该报的不是占比,是两个绝对量

整套设计的错在于拿一个比率的分子占比当质量指标。真正该报的是:① 组内信度——用 G 理论的 D 研究算各自的条件信度,看修炼组的 D 系数是否真的更高;② 残差方差的绝对值——绝对噪声到底降了多少。

这里还有个已知陷阱:G 理论经典版假设残差方差齐性,实际数据常常违背,不建模异方差会把测量质量的差异整个掩盖掉,须按条件算 G 与 D(PMID 41549064)。

所以,若修炼组的 D 系数没升、只是体位占比降了,那说明变异只是从”体位”搬到了”人身上”——练十年的人彼此差得更远,这完全可预期,与”测量更稳”毫无关系。

四条判法:一组体位绝对方差显著更低、D 系数更高、残差绝对方差也更低,三处同向 ⇒ 真稳健性提升;体位绝对方差没降但占比降了 ⇒ 比率假象(分母变大),不可报”训练降低体位敏感”;残差方差反而更大而 D 系数不变 ⇒ 变异只是搬了家,是”换了一种不稳”;两组体位绝对方差都接近零 ⇒ 天花板效应,该指标在此人群不可用,须换带位或操作员。

配套三条:体位×组、带位×组、操作员×组三个交互项都要报,因为组间方差差异可能只是某因子对某组根本不起作用;报方差分量必须同时报置信区间与估计方法,落在边界上的分量不许报 p<0.05;样本量按”要分辨的方差比”反解,不按均值差反解。

证据与局限

已验证: 成分数据的闭合约束导致伪相关、份额算术均值可越界(PMID 28882426);对数比变换处理份额数据的必要性(PMID 38806627);方差齐性检验的多种方法及其在重尾/偏态数据下的表现差异(PMID 37143826);混合模型异方差下须对异方差本身建模而非仅检验固定效应(PMID 32666394);小测量次数与短随访扩大负方差/边界区域(PMID 38695394);不同方差分量估计量的负估计概率差异(PMID 26843971);参数边界上经典检验失效、须置换或贝叶斯法(PMID 31460683、PMID 16474914);评判者少导致 ICC 偏倚(PMID 33818118);G 理论需扩展以处理异方差(PMID 41549064);COPD 与健康对照在体位上的胸廓异步差异(PMID 23412901);外科住院医个体内变异随年资下降(PMID 31517685);模拟器练习变异与技能迁移的关联(PMID 30898374);手枪射击专家与新手的协调结构差异(PMID 28628398);UK Biobank 参与者的健康志愿者偏差(PMID 28641372);儿童与成人 OCT 均值相同而精度不同(PMID 26808961)。

部分验证: 上述技能变异证据来自外科模拟训练与射击任务,与丹田呼吸的”训练”在任务结构、反馈方式、评估对象上均不同构,只能作方向性参照;OCT 的组间精度差异是仪器测量场景,与”人主动控制呼吸”的机制不同;体位效应的组间异质性证据来自 COPD 与健康对照,未涵盖长期呼吸训练人群。

待验证: 丹田呼吸场景下体位绝对方差在修炼者与新手间是否真有差异(无任何直接研究);修炼者的 G/D 系数是否真的更高;”修炼者主动固定姿势”造成的方差压缩量级;体位×组的交互是否存在;五个方差分量在 2 水平设计下的置信区间宽度是否已宽到无法支撑组间推断。

L1 网络状态:「该主题尚未进入 L1 对应网络」——CORRESPONDENCE_NETWORK.md 节点表把「相位 / 节律」列为物理侧节点(备注「周期性耦合的关键」)、「呼吸率 / 呼吸模式」列为生理侧观测项(备注「与节律耦合最直接」)、「气感报告」列为体验侧观测项,三者均未与易学侧任何节点连成边;节点表中完全没有「方差分量 / 变异源占比 / 信度系数 / 测量设计」这类测量学维度节点。边表 3.1–3.4 只登记五行↔五脏/五色/五音/五方/五季/五志、八卦↔五行、天干地支↔五行,无任何以变异源占比、组间方差结构或 G 理论为源或目标的边。EDGE_TABLE_science-yixue.md 仅有概念计数(丹田 51 篇 / 气功 48 篇 / 呼吸 189 篇),无层级(L1/L2/L3)与状态(active/open/conflict)判定。故本文封顶 C,所述机制均属方法学推演,不得读作已验证的生理机制。

相关阅读

证据等级: C
参考文献: PMID 38806627, PMID 28882426, PMID 37143826, PMID 32666394, PMID 23412901, PMID 38695394, PMID 26843971, PMID 31460683, PMID 16474914, PMID 33818118, PMID 41549064, PMID 31517685, PMID 30898374, PMID 28628398, PMID 28641372, PMID 26808961


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。

\”Decreased Proportion of Body Position\” Does Not Equal \”More Stable Training\”: A Pitfall in Between-Group Comparisons Using Variance Component Structure

Evidence level: C

Decomposing total variance into “posture, belt, operator, individual, and residual” components, the most tempting step is comparing the five proportions between practitioners and novices. This step almost certainly leads to error because proportions are ratios: a decrease in the numerator or an increase in the denominator can both cause a drop.

Main Text

I. Proportions Are Ratios, Not Quantities

The five variance sources must sum to 1. This constraint brings two consequences: any two proportions are naturally negatively correlated; a “posture proportion dropping from 40% to 25%” conveys no information on its own, because the posture variance could have dropped by 37%, or it could have remained unchanged while individual variance increased by 60%. The former indicates “reduced sensitivity to posture,” while the latter indicates “increased divergence between individuals.” These are opposite directions yet yield the same number.

There is long-standing statistical consensus on component data: arithmetic means do not yield valid results within the simplex and require log-ratio transformation first (PMID 38806627); in share data, an absolute increase may manifest as a decrease in the share (PMID 28882426). Therefore, one should report absolute variance components, not proportions.

II. Proportions Cannot Be Directly Used for Between-Group Inference

The sampling distribution of proportions depends on the shared denominator across five components; none are independent. Running a t-test directly on proportions assumes the denominator is constant. In clinical trials, variance homogeneity tests have specialized methods—Jackknife, Cochran’s Q—which are more robust for heavy-tailed and skewed data (PMID 37143826); in mixed models, if variance varies by group, the correct approach is to model the variance directly (location-scale extensions), rather than only testing fixed effects (PMID 32666394).

“Different variance structures between two groups” is a hypothesis about the variance function and must be tested at the variance level, not via descriptive comparisons of proportions.

III. Posture Effects Do Not Exist in Everyone

Healthy individuals may not have a posture effect to reduce at all. In 23 severe COPD patients, sitting showed thoracic asynchrony while supine eliminated it; upper thorax and abdomen were synchronous in sitting but strongly asynchronous in supine. In 12 age-matched healthy controls, neither posture showed any interval asynchrony (PMID 23412901).

If both healthy practitioners and novices have posture variance near zero, a “lower posture proportion for practitioners” is an artifact of the denominator. This is a classic ceiling effect: the indicator has no variation space in the target population, making between-group comparisons meaningless. Before design, one must confirm that posture variance is significantly greater than zero in the population.

IV. Small Designs Yield Unreliable Variance Component Estimates

Posture has only 2 levels, and operators often have 1 to 2. Few measurements, short follow-ups, and large residual variances all expand the “non-regular” parameter region, making negative variances and boundary estimates the norm rather than anomalies (PMID 38695394); different estimators vary greatly in the probability of negative estimates (PMID 26843971).

Inference is more complicated: asymptotic distributions for likelihood ratio, Wald, and Score tests fail at parameter boundaries, requiring permutation or Bayesian methods (PMID 31460683, PMID 16474914). Few raters themselves are a primary cause of ICC bias (PMID 33818118); confidence intervals for three modes may be so wide they span zero—”significant difference” and “no difference” become indistinguishable.

V. What Kind of “Stability” Does Training Reduce?

Skill research does support “greater stability for experts,” but one must clarify which kind. In 9 surgeons measuring 4 times each: junior residents had a score standard deviation of 5.8 (out of 30), while senior residents and attending physicians were both 1.8 (PMID 31517685); in simulator training, every 1% decrease in practice variance led to a 3.8-point increase on post-test and a 3.0-point increase in live transfer (PMID 30898374).

However, these are repeated measures on the same individual, belonging to “residual + interaction” in crossed designs, not the “posture” design factor. A pistol shooting control group illustrates this better: expert COP displacement was smaller and multiscale entropy lower, yet both groups showed near-in-phase upper body-to-gun relative phase. The true difference lay in structure—experts had the gun lagging behind the body, novices had it leading (PMID 28628398). Stability does not mean every dimension shrinks; it means the structure shifts position.

VI. Fundamental Trouble in Between-Group Comparison: No Random Assignment

Practitioners and novices are not randomly assigned; they are self-selected. Those capable of training for ten years were likely already more able to sit still and control breathing—cross-sectional designs cannot distinguish “what is learned” from “who can persist.” Volunteer bias is systematic: UK Biobank participants are healthier and have better socioeconomic status than the general population (PMID 28641372).

Another reading error must be guarded against. In SD-OCT measurements of children vs. adults: the two groups showed no difference in mean retinal thickness, but children’s COV was 3 to 4 times that of adults, and ICC was significantly lower (0.73–0.99 vs 0.986–0.993, PMID 26808961). Same mean, vastly different precision—this “between-group precision difference” is real yet most easily miswritten as “one group has greater ability.”

Reversal: What Should Be Reported Are Not Proportions, But Two Absolute Quantities

The error in the entire design lies in using a numerator proportion as a quality metric. What should truly be reported is: ① Intra-group reliability—calculate conditional reliability using G-theory’s D study for each group to see if the practitioner group truly has a higher D coefficient; ② Absolute residual variance—how much absolute noise actually decreased.

There is also a known trap here: the classic G-theory assumes homoscedastic residuals, but actual data often violate this. Failing to model heterogeneity masks differences in measurement quality; one must calculate G and D conditionally (PMID 41549064).

Therefore, if the practitioner group’s D coefficient did not rise but only the posture proportion dropped, it means variance merely shifted from “posture” to “person”—people trained for ten years differ more from each other. This is fully expected and unrelated to “more stable measurement.”

Four Criteria: If posture absolute variance is significantly lower, D coefficient is higher, and absolute residual variance is also lower (all three aligned) ⇒ true robustness improvement; if posture absolute variance did not drop but proportion dropped ⇒ ratio artifact (denominator inflated), do not report “training reduces posture sensitivity”; if residual variance increased while D coefficient remained unchanged ⇒ variance merely moved locations, it is “a different kind of instability”; if both groups have posture absolute variance near zero ⇒ ceiling effect, this indicator is unusable for this population; switch to belt or operator factors.

Three accompanying reports: report interactions Posture×Group, Belt×Group, Operator×Group, because between-group variance differences may simply mean a factor is ineffective for a specific group; when reporting variance components, always report confidence intervals and estimation methods; do not report p<0.05 for components falling on boundaries; calculate sample size based on “resolvable variance ratios,” not mean differences.

Evidence and Limitations

Verified: Closure constraints in component data lead to spurious correlations; arithmetic means of shares can exceed bounds (PMID 28882426); necessity of log-ratio transformation for share data (PMID 38806627); multiple methods for variance homogeneity tests and their performance differences in heavy-tailed/skewed data (PMID 37143826); need to model heterogeneity itself rather than only fixed effects in mixed models under heteroscedasticity (PMID 32666394); small measurement counts and short follow-ups expand negative variance/boundary regions (PMID 38695394); differences in negative estimation probabilities across estimators (PMID 26843971); classical tests fail at parameter boundaries requiring permutation or Bayesian methods (PMID 31460683, PMID 16474914); few raters cause ICC bias (PMID 33818118); G-theory requires extension to handle heteroscedasticity (PMID 41549064); differences in thoracic asynchrony between COPD and healthy controls regarding posture (PMID 23412901); decrease in individual variability with seniority among surgical residents (PMID 31517685); association between simulator practice variance reduction and skill transfer (PMID 30898374); coordination structure differences between expert and novice pistol shooters (PMID 28628398); volunteer bias in UK Biobank participants (PMID 28641372); same mean but different precision in OCT between children and adults (PMID 26808961).

Partially Verified: The above skill variability evidence comes from surgical simulation training and shooting tasks; the “training” in Dantian breathing differs structurally in task design, feedback mechanisms, and assessment targets, serving only as directional reference; the between-group precision difference in OCT is an instrument measurement scenario, differing from “active human control of breathing”; evidence for between-group heterogeneity in posture effects comes from COPD and healthy controls, not covering long-term breathing training populations.

To Be Verified: Whether there is a true difference in posture absolute variance between practitioners and novices in the Dantian breathing scenario (no direct research exists); whether practitioners truly have higher G/D coefficients; the magnitude of variance compression caused by “practitioners actively fixing posture”; whether Posture×Group interaction exists; whether confidence interval widths for five variance components under a 2-level design are already so wide as to preclude between-group inference.

L1 Network Status: “This topic has not yet entered the L1 corresponding network”—the CORRESPONDENCE_NETWORK.md node table lists “Phase/Rhythm” as a physical-side node (note: “key for periodic coupling”), “Breathing Rate/Breathing Pattern” as a physiological-side observation (note: “most directly coupled to rhythm”), and “Qi Sensation Reports” as an experiential-side observation; none connect to any Yijing-side nodes. The node table contains no measurement dimension nodes such as “Variance Components/Variance Source Proportions/Reliability Coefficients/Measurement Design.” Edge tables 3.1–3.4 only register Five Phases↔Five Organs/Five Colors/Five Tones/Five Directions/Five Seasons/Five Emotions, Eight Trigrams↔Five Phases, Heavenly Stems/Earthly Branches↔Five Phases; no edges exist with variance source proportions, between-group variance structures, or G-theory as sources or targets. EDGE_TABLE_science-yixue.md only contains concept counts (Dantian 51 / Qigong 48 / Breathing 189), with no hierarchy (L1/L2/L3) or status (active/open/conflict) determination. Thus, this paper caps at Level C; described mechanisms belong to methodological deduction and should not be read as verified physiological mechanisms.

Related Reading

Evidence Level: C
References: PMID 38806627, PMID 28882426, PMID 37143826, PMID 32666394, PMID 23412901, PMID 38695394, PMID 26843971, PMID 31460683, PMID 16474914, PMID 33818118, PMID 41549064, PMID 31517685, PMID 30898374, PMID 28628398, PMID 28641372, PMID 26808961


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 *