体位占比降了,可能只是别人的方差涨了:一个不随样本量消失的偏倚

Written by

in

Language

Evidence level: C

「训练组体位占比更低」这句话,可能一个字都不涉及体位——它可能只是「训练组的个体差异更大」。这两件事在占比这一个数上长得一模一样,而且样本量越大,越分不开。

正文

一、先把「占比」写成公式,问题就露出来了

丹田三模态测量的方差分析把总方差拆成五个源:体位、带位、操作员、个体、残差。设它们的方差分量是 v₁ 到 v₅,那么「体位占 28%」这句话的准确写法是 p₁ = v₁ /(v₁+v₂+v₃+v₄+v₅)。分子是我们要研究的体位,分母是另外四个源加上它自己。

组间比较时,我们算的是训练组与新手组的 p₁ 之差,但我们真正想问的,是两组 v₁ 之差。这两件事之间隔着一个分母 S = Σvᵢ。只要 S 在两组之间不一样,p₁ 的差就不再是 v₁ 的差的影子。这不是统计技巧问题,是「你报的数和你想说的数不是同一个数」。

二、分母不是背景,它是第五个自变量

对 p₁ 取对数再差分,得到一条恒等式:Δln p₁ = Δln v₁ − Δln S,而 Δln S 近似等于 Σpᵢ·Δln vᵢ。也就是说,占比的变化里,只有一部分来自体位本身,另一部分是另外四个源变化的加权和,被减号挂在了前面。

按源设计的份额(0.28 / 0.27 / 0.15 / 0.20 / 0.10)算,Δln v₁ 与 Δln S 的相关系数是 0.59。这个正相关会在某些情形下部分抵消,但它同时也说明:分母不是静止的背景,它跟着所有源一起动,且动得不小。任何只看 p₁ 的结论,都在默认「S 没动」。

三、等价折算:任何源的变化都能伪装成体位的假变化

把上式整理,得到一条可以直接用的通式:当体位方差真值不变时,源 j 变化 δ_j 会在 p₁ 上制造出等价于体位变化 p_j·δ_j/(1−p₁) 的假象。个体方差翻倍(δ=0.693),等价于体位假降 0.139 个对数单位,即 13%。

这条通式有一个反直觉的推论:若其余四个源按原比例同步缩放,那么折算权重 p_j/(1−p₁) 恒等于它的条件份额,与 p₁ 大小完全无关。实测把主源份额从 0.15 一路调到 0.70,第二源翻倍所对应的等价体位变化始终是 0.2599。所以「体位是最大的一头,它应该最不容易被带偏」这个直觉是错的——稳不稳不看主源占多少,看其余源相对它怎么变。

四、抵消点:体位要真涨多少,才能让占比「看不见」

反解通式可得抵消点。个体方差翻 1.5 倍时,体位方差需要真涨 11.9%,占比才会保持不动;翻 2 倍时需涨 21.2%;翻 2.5 倍需涨 29.0%;翻 3 倍需涨 35.7%。这些数字的含义是:一个真实存在的体位改善,完全可以被另一个源的膨胀抹平到「不显著」。

反过来更糟。在体位真值毫无变化的前提下,按名义双侧 5% 水平做检验,占比法的假阳性率随个体方差膨胀迅速失控:涨 10% → 8.2%,涨 20% → 17.0%,涨 50% → 60.5%,翻倍 → 96.8%;而绝对量法始终稳在 5.0%。反解这条曲线:假阳性要摸到 20%,个体方差只需涨 22.5%;要摸到 80%,只需涨 66.6%。门槛低到任何一个真实人群的异质性波动都能跨过去。

五、最要命的一条:加大样本量会让它更糟

通常遇到不确定,第一反应是加人。这个坑不吃这一套。固定个体方差涨 20%,把每组人数从 20 一路加到 320,占比法假阳性率依次是 10.9%、17.0%、29.3%、51.7%、80.8%——单调上升。名义 95% 置信区间的实际覆盖率同步崩塌:涨 20% 时是 83.0%,涨 50% 时只剩 39.5%,翻倍时是 3.3%。

原因很干脆:这是偏倚,不是噪声。噪声随样本量缩小,偏倚不动,于是信噪比里偏倚那一项越来越占主导。换句话说,这项研究做得越认真、样本越大、区间越窄,那个错误的结论就越「显著」。同时,占比法的标准误比绝对量法还小约 19%(0.0364 对 0.0447),它会用更窄的区间把偏倚报出来,看上去更精确。

六、什么时候真的会翻转,而不是仅仅夸大

两法结论真正相反,条件很干净:分母漂移的幅度超过分子本身的漂移,即 |δS| > |δ₁| 且方向一致时,占比法与绝对量法给出相反的符号。实测:体位涨 1.3 倍叠加个体涨 2.5 倍,比值 0.978,两法仍同向;再把带位也设成涨 2 倍,比值升到 1.69,翻转成立——绝对量报「体位升」,占比报「体位降」。

把判据反过来用更有价值:若体位真涨 10%,个体方差只需涨 61% 就能让结论翻转;真涨 5%,个体只需涨 28%;真涨 15%,需涨 101%。而人群中个体差异的组间差,本来就常在 1.5 到 3 倍这个量级。

七、出路:不是换统计量,是换报告对象

三条可执行的规则。第一,组间推断只对绝对方差分量做,报 v₁ 的差值与区间,占比降级为描述性指标,不进假设检验。第二,凡报占比结论,必须同时报分母漂移量 δS;δS 与 δ₁ 同量级时该结论作废。第三,若确实要在组成框架里做,改用等距对数比(ilr)坐标,避免闭合约束带来的伪相关;但要注意这是近似而非恒等,2026 年的方法学综述指出相对值只是「几乎」满足子组成一致性。

最后一点分寸:数学耦合并非总是致命。有实测研究把急性/慢性负荷比拆成耦合与非耦合两版,发现效应量几乎全是微不足道。这恰恰说明判据的价值——耦合致不致命,取决于 δS 相对 δ₁ 有多大,而这个比值是可以事前算出来的。

证据与局限

已验证: 比值型指标与其分母存在数学耦合,会产生伪相关,这是被反复确认的现象(PMID 29101104、PMID 25245392;后者在 315 人样本中显示,调整动脉管径后高/低风险组血流介导扩张的差异被减半);组成数据(各分量非负且和为定值)不能直接套用常规多变量方法,需用对数比变换(PMID 28555522、PMID 34737726);对闭合数据直接取算术均值、标准差与相关系数会产生系统偏误(PMID 28882426);相对丰度会掩盖绝对量的真实变化,需另做绝对定量(PMID 40689579、PMID 34146605);同类「比率随一个与目标无关的量变动」的偏倚在一致性统计量中亦有经典记录(kappa 受患病率影响,PMID 8501467、PMID 34520848);数学耦合的实际危害大小取决于具体场景,已有实测给出效应量微不足道的反例(PMID 31672929)。

部分验证: 本文全部数字来自本人蒙特卡洛与解析计算(每格 40 万次模拟,n=40/组、σ=0.20、五源份额 0.28/0.27/0.15/0.20/0.10)。份额构型与 σ 均为设定值,非实测;「等价折算权重与 p₁ 无关」只在其余源按原比例缩放的假设下成立;δS 的一阶近似忽略了高阶项;假阳性与覆盖率的表格基于正态近似。

待验证: 丹田场景下五个方差分量的真实组间差异量级(无任何实测,个体方差组间差 1.5–3 倍仅为类比猜测);三模态各自的份额观测噪声 σ 真实取值;Δln vᵢ 之间是否独立(若相关,S 的方差需重算);在本领域换用 ilr 后原有结论有多少条会翻转。

L1 网络状态:「该主题尚未进入 L1 对应网络」——CORRESPONDENCE_NETWORK.md 逐词检索「占比 0 / 份额 0 / 比率 0 / 方差 0 / 分量 0 / 组成 0 / 闭合 0 / 分母 0 / 辛普森 0 / 判定 0 / 功效 0」,全部零命中;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,止于 E-050(「六十甲子纳音|对应|60 个频率」,层级 L1、状态 open,现代证据栏原文「尚无任何可测来源」),无任何以比率、分母、方差分量或组间比较为源或目标的边。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 29101104, PMID 25245392, PMID 28555522, PMID 34737726, PMID 28882426, PMID 40689579, PMID 34146605, PMID 8501467, PMID 34520848, PMID 31672929


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。

Positional bias may not vanish with sample size: it might just be others’ variance increasing

Evidence level: C

“The proportion of the training group is lower” might not involve the proportion at all—it may simply mean “the individuals in the training group have greater variability.” These two things look identical when expressed as a single proportion, and larger sample sizes make them even harder to distinguish.

Body

I. Writing the “Proportion” as a Formula Reveals the Problem

The variance analysis of the Dantian three-modal measurement decomposes total variance into five sources: posture, belt, operator, individual, and residual. Let their variance components be v₁ to v₅, then the accurate statement of “posture accounts for 28%” is p₁ = v₁ / (v₁+v₂+v₃+v₄+v₅). The numerator is the posture we study, while the denominator includes the other four sources plus itself.

When comparing groups, we calculate the difference in p₁ between the training and novice groups, but what we truly want to know is the difference between their v₁ values. These two quantities are separated by the denominator S = Σvᵢ. As long as S differs between groups, the difference in p₁ no longer reflects the shadow of the difference in v₁. This is not a statistical technique issue; it is “the number you report and the number you intend to convey are not the same.”

II. The Denominator Is Not Background; It Is the Fifth Independent Variable

Taking the logarithm of p₁ and differencing yields an identity: Δln p₁ = Δln v₁ − Δln S, where Δln S approximates Σpᵢ·Δln vᵢ. Thus, changes in proportion arise partly from posture itself and partly from the weighted sum of changes in the other four sources, subtracted at the front.

Based on source design shares (0.28 / 0.27 / 0.15 / 0.20 / 0.10), the correlation between Δln v₁ and Δln S is 0.59. This positive correlation can partially offset under certain conditions, but it also demonstrates: the denominator is not static background; it moves with all sources and moves significantly. Any conclusion based solely on p₁ assumes “S did not move.”

III. Equivalent Conversion: Any Source Change Can Masquerade as a False Posture Change

Rearranging the equation yields a directly usable general formula: when the true posture variance remains constant, a change δⱼ in source j creates an illusion of posture change equivalent to pⱼ·δⱼ/(1−p₁). Doubling individual variance (δ=0.693) is equivalent to a false posture drop of 0.139 log units, i.e., 13%.

This general formula has a counterintuitive implication: if the other four sources scale synchronously by their original proportions, then the conversion weight pⱼ/(1−p₁) equals its conditional share and is completely independent of p₁ magnitude. Testing adjusted the primary source share from 0.15 up to 0.70; doubling the second source always corresponds to an equivalent posture change of 0.2599. Thus, the intuition “posture is the largest component, so it should be least easily misled” is wrong—stability does not depend on the primary share size, but on how the other sources change relative to it.

IV. The Cancellation Point: How Much Must Posture Truly Rise for the Proportion to Remain Invisible?

Solving the general formula inversely yields the cancellation point. When individual variance increases 1.5-fold, posture variance must truly rise 11.9% for the proportion to remain unchanged; doubling requires 21.2%; 2.5-fold requires 29.0%; tripling requires 35.7%. These numbers mean: a real posture improvement can be completely erased by expansion of another source, rendering it “insignificant.”

Conversely, the situation is worse. Under the assumption that posture remains unchanged, applying nominal two-sided 5% significance testing to the proportion method causes false positive rates to rapidly lose control as individual variance expands: rising 10% → 8.2%, 20% → 17.0%, 50% → 60.5%, doubling → 96.8%; while the absolute quantity method remains steady at 5.0%. Solving this curve inversely: to reach a false positive of 20%, individual variance needs only rise 22.5%; to reach 80%, just 66.6%. The threshold is so low that any real population heterogeneity fluctuation can cross it.

V. The Most Dangerous Point: Increasing Sample Size Makes It Worse

Usually encountering uncertainty, the first reaction is to add people. This pitfall does not follow that rule. Fixing individual variance increase at 20%, increasing group size from 20 up to 320, the proportion method false positive rates are sequentially 10.9%, 17.0%, 29.3%, 51.7%, 80.8%—monotonically rising. The actual coverage of nominal 95% confidence intervals simultaneously collapses: at 20% rise it is 83.0%, at 50% rise only 39.5%, doubling it is 3.3%.

The reason is clear: this is bias, not noise. Noise shrinks with sample size, bias remains constant, so in the signal-to-noise ratio the bias term increasingly dominates. In other words, the more carefully this study is conducted, the larger the sample, the narrower the interval, the more “significant” the wrong conclusion becomes. Meanwhile, the proportion method’s standard error is about 19% smaller than the absolute quantity method (0.0364 vs 0.0447); it reports the bias with a narrower interval, appearing more precise.

VI. When Will the Conclusion Truly Flip, Rather Than Just Be Exaggerated?

The two methods yield truly opposite conclusions under clean conditions: the denominator drift magnitude exceeds the numerator’s own drift, i.e., |δS| > |δ₁| and they move in the same direction, then the proportion method and absolute quantity method give opposite signs. Testing: posture rises 1.3-fold combined with individual rising 2.5-fold, ratio 0.978, both methods still same direction; setting belt also rising 2-fold, ratio climbs to 1.69, flip occurs—absolute quantity reports “posture up,” proportion reports “posture down.”

Using the criterion inversely is more valuable: if posture truly rises 10%, individual variance needs only rise 61% to flip the conclusion; true rise 5%, individuals need just 28%; true rise 15%, requires 101%. Meanwhile, inter-group differences in individual variance among real populations are often at the level of 1.5 to 3 times.

VII. The Way Out: Not Changing the Statistic, But Changing the Reporting Target

Three actionable rules. First, inter-group inference applies only to absolute variance components; report the difference and interval of v₁, downgrading proportion to a descriptive indicator, excluding hypothesis testing. Second, any proportion conclusion must simultaneously report the denominator drift δS; when δS and δ₁ are of similar magnitude, that conclusion is void. Third, if indeed working within the compositional framework, adopt isometric log ratio (ilr) coordinates to avoid spurious correlation from closure constraints; note this is approximate rather than exact, as the 2026 methodological review indicates relative values are only “almost” satisfying sub-compositional consistency.

One final nuance: mathematical coupling is not always fatal. Measured studies have decomposed acute/chronic load ratios into coupled and uncoupled versions, finding effect sizes almost entirely negligible. This precisely illustrates the criterion’s value—coupling causes non-fatal outcomes depending on how large δS is relative to δ₁, and this ratio can be calculated in advance.

Evidence and Limitations

Verified: Ratio-type indicators are mathematically coupled with their denominators, producing spurious correlation, a repeatedly confirmed phenomenon (PMID 29101104, PMID 25245392; the latter showed in a 315-person sample that after adjusting arterial diameter, the difference in flow-mediated dilation between high/low risk groups was halved); compositional data (non-negative components summing to a constant) cannot directly apply conventional multivariate methods, requiring log-ratio transformation (PMID 28555522, PMID 34737726); directly taking arithmetic mean, standard deviation, and correlation coefficients from closed data produces systematic bias (PMID 28882426); relative abundance masks true absolute quantity changes, requiring separate absolute quantification (PMID 40689579, PMID 34146605); similar “ratio bias with variation in an irrelevant quantity” is also classically recorded in consistency statistics (kappa affected by prevalence, PMID 8501467, PMID 34520848); the actual severity of mathematical coupling depends on specific contexts, with measured studies providing counterexamples where effect sizes are negligible (PMID 31672929).

Partially Verified: All numbers in this paper derive from my Monte Carlo and analytical calculations (400,000 simulations per cell, n=40/group, σ=0.20, five-source shares 0.28/0.27/0.15/0.20/0.10). Share configuration and σ are set values, not measured; “equivalent conversion weight independent of p₁” holds only under the assumption that other sources scale proportionally; δS’s first-order approximation ignores higher-order terms; false positive and coverage tables rely on normal approximation.

To Be Verified: True inter-group magnitude of the five variance components in the Dantian scenario (no measurement, individual variance inter-group difference 1.5–3 times is purely analogical speculation); true values of observation noise σ for each of the three modalities; whether Δln vᵢ are independent (if correlated, S’s variance must be recalculated); how many original conclusions would flip after adopting ilr in this field.

L1 Network Status: “This topic has not yet entered the L1 corresponding network”—CORRESPONDENCE_NETWORK.md word-by-word search for “proportion 0 / share 0 / ratio 0 / variance 0 / component 0 / composition 0 / closed 0 / denominator 0 / Simpson 0 / judgment 0 / power 0”, all zero hits; edge table 21 entries (E-001~E-005, E-010~E-014, E-030~E-034, E-040~E-044, E-050) all represent correspondence relationships among Five Phases/Five Organs/Five Colors/Five Tones/Five Directions/Five Seasons/Five Emotions/Eight Trigrams/Stems and Branches/Nayin, stopping at E-050 (“60 Jiazi Nayin | Corresponds | 60 Frequencies”, level L1, status open, original modern evidence column states “No measurable source yet“), no edges originating from or targeting ratios, denominators, variance components, or inter-group comparisons. EDGE_TABLE_science-yixue.md contains only concept article counts (Dantian 51 / Qigong 48 / Breathing 189), no level and status judgment. Hence this paper caps at C.

Related Reading

Evidence Level: C
References: PMID 29101104, PMID 25245392, PMID 28555522, PMID 34737726, PMID 28882426, PMID 40689579, PMID 34146605, PMID 8501467, PMID 34520848, PMID 31672929


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 *