「这个方差是不是零」问错了:体位因子只有两个水平,它永远测不出来

Written by

in

Language

Evidence level: C

方差分量的检验不是普通检验:真值卡在零这条墙上,p 值要按混合卡方算,临界值 2.706 而不是 3.841。但更要命的是——当体位只有坐/卧两个水平时,这个分量结构上就不可检出,加大样本量一分钱效用都没有。

正文

上一轮(#1194)算清了「占比 vs 绝对量」会互相翻多少。这一轮要往前退一步:在翻之前,那个被拿来比较的体位方差,到底有没有被证明过大于零?如果它本身就在零附近,后面所有组间比较都是在比较两个噪声。

先说一件被普遍忽略的事:方差分量的零假设落在参数空间的边界上。普通检验假设真值在内部,零假设下统计量服从中心卡方;但方差不可能为负,真值等于零时它就贴在墙上,似然比统计量的零分布是两个卡方的等权混合——一半概率取 0,一半概率取 $\chi^2_1$(PMID 16474914 专门讨论了这个混合分布及行为遗传学里由此产生的 p 值偏倚;PMID 25106394 在双变量 ACE 模型里给出同一结论;PMID 22822249 从渐近理论处理了边界下的伪似然比;PMID 22930674 与 PMID 41891701 则给出多分量和贝叶斯替代方案)。

这不是学究气。若你按 3.841 取临界值,实际犯一类错的概率只有 2.5%,不是你想要的 5%——等于白送一半严度。要 5%,临界值是 2.706(0.5×χ²₁ 的 95 分位);要 10%,是 1.642。两边的功效代价也不对称:80% 功效下,正确临界值只需非中心参数 λ=6.18,用错 3.841 却要 λ=7.85,多花 26.9% 的信息量。

但真正的麻烦在第二层。体位因子在设计里只有两个水平(坐位、仰卧)。方差分量的估计精度由水平数 k 决定,不是受试者数:平衡单向设计下估计量的相对标准误约 $\sqrt{2/(k-1)}$。k=2 时它是 1.414,也就是说标准误和它要估的那个量一样大。换算成信号噪声比:$\tau/SE = \sqrt{(k-1)/2} = 0.707$。

0.707 意味着什么?要让 95% 置信区间下界排除零,需要 z>1.96;0.707 永远到不了。这不是样本量问题,是把受试者从 60 人加到 6000 人也不会变——加人只缩小受试者层面的噪声,缩不了「只有两种体位」这个自由度上限。要 z>1.96,得 (k−1)/2 > 3.8416,即 k ≥ 9:至少九种体位/角度档位。

于是 #1195 原设计的判据「体位分量 95%CI 下界含零者占比 >40%」在数学上有个尴尬结局:如果体位因子真的只有两档,这个占比必然接近 100%,不是 40%。而 100% 会被读成「天花板效应得到证实、组间设计不可测」,实际上它只证明了一件事——设计里体位那根轴太短。这是把一个设计缺陷误诊成人群属性。

再往下还有第三层:判据本身的抽样噪声。n=60 时,若真比例是 40%,其 95%CI 是 ±12.4 个百分点;换成一批人,同一人群可能读出 28% 也可能读出 52%——正好横跨 40% 这条判定线。而若真比例是 10%,噪声只有 ±7.6 个百分点。把判定线放在噪声最大处(40%),等于让结论由抽样运气决定。缺口表的四十八条里反复出现同一件事:判据写得像结论,其实是把一个统计量的一次抽样当成了属性。

顺带一提,混合卡方只适用于单一边界分量的检验;若同时检验多个分量(比如体位+操作员都可能是零),零分布变成卡方变量的混合,权重是按各分量是否为零的组合概率算的,简单等权 0.5/0.5 就失效了(PMID 22930674)。而自助法给出的 CI 在方差分量上也有自己的病:估计量本身偏态、且被零截断,百分位法在真值接近零时覆盖率严重不足(PMID 14673940 专门比较了几种自助 CI 在 ICC 上的表现);样本量对信度估计的影响比通常以为的大得多(PMID 40119099、PMID 9463853、PMID 24338600 从设计角度给了所需精度与样本量的对应)。天花板/地板效应的报告也有成熟口径:应报落在最高(最低)分的比例,并说明该比例是否足以限制对变化的检出(PMID 26519187 在牛津膝评分上做了示范;PMID 32043052 在 PROMIS 各域上比较了不同关节的分布)。

一句话收束:「体位方差是否显著大于零」这个问题,先要问设计给了它几个自由度。两档体位配六十个人,得到的不是「没有效应」,而是「这个问题在本设计里没被问过」。

L1 网络状态:CORRESPONDENCE_NETWORK.md 三域节点表(易学/中医侧、物理侧、生理/体验侧)无任何以「方差分量 / 信度 / 边界检验 / 自由度 / 设计分辨率」为源或目标的边;EDGE_TABLE_science-yixue.md 只统计概念被写过几篇(频率 407 / 呼吸 189 / 丹田 51)与最高证据级。该主题尚未进入 L1 对应网络。

证据与局限

已验证: 方差分量零假设落在参数空间边界时,似然比统计量的零分布为 0.5χ²(0)+0.5χ²(1),α=0.05 对应临界值 2.7055(非 3.841)——此为经典结果,多篇方法学文献一致(PMID 16474914、25106394、22822249)。本例数值为解析自算:80% 功效 c=2.706 需 λ=6.183,c=3.841 需 λ=7.848,比值 1.269。方差分量估计量精度由水平数 k 决定,k=2 时相对标准误 √2≈1.414,τ/SE=0.707<1.96——纯代数结论,与样本量无关。

部分验证: k≥9 才可使单一分量的 CI 下界稳定排除零(由 (k−1)/2>3.8416 反解),此推导假设平衡单向设计、正态近似、单一边界分量;非平衡设计、多分量同时检验、或含协变量的模型下阈值会变。多分量边界检验的实际零分布需按分量组合概率加权(PMID 22930674),本文未展开计算。

待验证: ① 真实 SVC-FOT 数据里,体位分量的点估计在一批健康人中究竟落在什么位置——若点估计普遍远小于其标准误,则无论门槛怎么设都测不出,「天花板效应」这个词用在你身上是错的,用错了对象;② 把体位档位从 2 扩到 9 以上(坐 90°/75°/60°/45°/30°、仰卧 0°/15°/30°/45°)后,体位分量是否真的浮现——若浮现,则此前”体位无可测变异”的结论只是设计太短;③ 40% 与 10% 这两条判据线的选择依据从未被论证,应改为基于 MDC95 与预期组间差的比值,而非固定百分比。

相关阅读

证据等级: C
参考文献: PMID 16474914, PMID 25106394, PMID 22822249, PMID 22930674, PMID 41891701, PMID 14673940, PMID 40119099, PMID 9463853, PMID 24338600, PMID 26519187, PMID 32043052

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。

This variance is not zero: the posture factor has only two levels, so it can never be estimated

Evidence level: C

Testing variance components is not an ordinary test: the true value is pinned against the zero wall, so the p-value must be computed using a mixed chi-square distribution, with a critical value of 2.706 rather than 3.841. But more critically, when the posture factor has only two levels (sitting/lying), this component is structurally undetectable; increasing the sample size yields no utility.

Body

In the previous round (#1194), we calculated how much “proportion vs. absolute amount” flips against each other. This round steps back: before flipping, was the posture variance component ever proven to be greater than zero? If it itself lies near zero, all subsequent between-group comparisons are merely comparing two noises.

First, a widely overlooked fact: the null hypothesis for the variance component lies on the boundary of the parameter space. Ordinary tests assume the true value is internal; under the null, the statistic follows a central chi-square. But variance cannot be negative; when the true value equals zero, it sticks to the wall. The null distribution of the likelihood ratio statistic is an equal-weight mixture of two chi-squares: half probability at 0, half at $\chi^2_1$ (PMID 16474914 specifically discusses this mixed distribution and the resulting p-value bias in behavior genetics; PMID 25106394 reaches the same conclusion in the bivariate ACE model; PMID 22822249 handles boundary pseudo-likelihood ratios via asymptotic theory; PMID 22930674 and PMID 41891701 offer multi-component and Bayesian alternatives).

This is not pedantry. If you use 3.841 as the critical value, the actual Type I error rate is only 2.5%, not your desired 5%—you effectively lose half the stringency for free. To achieve 5%, the critical value is 2.706 (the 95th percentile of $0.5\chi^2_1$); for 10%, it is 1.642. The power costs are also asymmetric: at 80% power, the correct critical value requires a non-centrality parameter $\lambda=6.18$; using the wrong 3.841 requires $\lambda=7.85$, costing 26.9% more information.

But the real trouble lies in the second layer. The posture factor has only two levels (sitting, supine) by design. The precision of the variance component estimate depends on the number of levels $k$, not the number of subjects: under a balanced one-way design, the relative standard error of the estimator is approximately $\sqrt{2/(k-1)}$. When $k=2$, this is 1.414, meaning the standard error is as large as the quantity being estimated. Converted to signal-to-noise ratio: $\tau/SE = \sqrt{(k-1)/2} = 0.707$.

What does 0.707 mean? To exclude zero in the lower bound of a 95% confidence interval, we need $z > 1.96$; 0.707 can never reach this. This is not a sample size issue; adding subjects from 60 to 6,000 will not change it—adding subjects only reduces subject-level noise, but cannot expand the degree-of-freedom ceiling imposed by “only two posture levels.” To achieve $z > 1.96$, we need $(k-1)/2 > 3.8416$, i.e., $k \geq 9$: at least nine posture/angle levels.

Thus, the original criterion in #1195—”proportion of cases where the posture component’s 95% CI lower bound includes zero exceeds 40%”—has a mathematically awkward outcome: if the posture factor truly has only two levels, this proportion must approach 100%, not 40%. A 100% rate is read as “ceiling effect confirmed, between-group design unmeasurable,” but actually it proves only one thing—the posture axis in the design is too short. This misdiagnoses a design flaw as a population property.

Next comes the third layer: sampling noise of the criterion itself. With $n=60$, if the true proportion is 40%, its 95% CI spans $\pm 12.4$ percentage points; with another batch, the same population might yield 28% or 52%—exactly crossing the 40% decision line. If the true proportion is 10%, the noise shrinks to $\pm 7.6$ percentage points. Placing the decision line at the point of maximum noise (40%) means conclusions are decided by sampling luck. The forty-eight entries in the gap table repeatedly highlight this: the criterion is written like a conclusion, yet it treats a single sampling of a statistic as an attribute.

Incidentally, the mixed chi-square applies only to a single boundary component; if testing multiple components simultaneously (e.g., posture + operator both possibly zero), the null distribution becomes a mixture of chi-square variables weighted by the combination probabilities of which components are zero, so simple equal weighting 0.5/0.5 fails (PMID 22930674). Bootstrap CIs for variance components also have their own issues: the estimator is biased and truncated at zero, so percentile methods severely under-cover when the true value is near zero (PMID 14673940 compares several bootstrap CI methods for ICC); the impact of sample size on reliability estimation is far greater than usually assumed (PMID 40119099, PMID 9463853, PMID 24338600 provide design-based correspondences between required precision and sample size). Reporting of ceiling/floor effects also has established conventions: report the proportion falling at the highest (lowest) score and clarify whether this proportion is sufficient to limit detection of change (PMID 26519187 demonstrates this on the Oxford Knee Score; PMID 32043052 compares joint distributions across PROMIS domains).

In summary: “Is the posture variance significantly greater than zero?” must first ask how many degrees of freedom the design provides. Two posture levels with sixty subjects does not yield “no effect”; it yields “this question was never asked in this design.”

L1 Network Status: The three-domain node table in CORRESPONDENCE_NETWORK.md (Yijing Studies/TCM side, Physics side, Physiology/Experience side) contains no edges sourced or targeted by “variance component / reliability / boundary test / degrees of freedom / design resolution”; EDGE_TABLE_science-yixue.md only counts how many times concepts were written (frequency: 407 for variance, 189 for breathing, 51 for dantian) and the highest evidence level. This topic has not yet entered the L1 corresponding network.

Evidence and Limitations

Verified: When the null hypothesis for a variance component lies on the parameter space boundary, the null distribution of the likelihood ratio statistic is $0.5\chi^2(0)+0.5\chi^2(1)$, with $\alpha=0.05$ corresponding to a critical value of 2.7055 (not 3.841)—this is a classic result consistently supported by multiple methodological literature (PMID 16474914, 25106394, 22822249). The numerical values here are analytically computed: for 80% power, $c=2.706$ requires $\lambda=6.183$, while $c=3.841$ requires $\lambda=7.848$, yielding a ratio of 1.269. The precision of the variance component estimator is determined by the number of levels $k$; when $k=2$, the relative standard error is $\sqrt{2}\approx1.414$, so $\tau/SE=0.707<1.96$—a purely algebraic conclusion independent of sample size.

Partially Verified: $k \geq 9$ is required for the lower bound of a single-component CI to stably exclude zero (solved from $(k-1)/2 > 3.8416$), assuming a balanced one-way design, normal approximation, and a single boundary component; thresholds would shift under unbalanced designs, multi-component simultaneous testing, or models with covariates. The actual null distribution for multi-component boundary tests must be weighted by component combination probabilities (PMID 22930674), which is not expanded here.

To Be Verified: ① In real SVC-FOT data, where exactly does the point estimate of the posture component fall across batches of healthy individuals—if point estimates are generally far smaller than their standard errors, no threshold can detect it, and the term “ceiling effect” is misapplied to you; ② Expanding posture levels from 2 to over 9 (sitting at 90°/75°/60°/45°/30°, supine at 0°/15°/30°/45°)—does the posture component truly emerge? If it does, the prior conclusion that “posture has no measurable variation” was merely due to insufficient design length; ③ The basis for choosing the 40% and 10% decision lines was never argued and should instead be based on the ratio of MDC95 to expected between-group differences, not fixed percentages.

Further Reading

Evidence Level: C
References: PMID 16474914, PMID 25106394, PMID 22822249, PMID 22930674, PMID 41891701, PMID 14673940, PMID 40119099, PMID 9463853, PMID 24338600, PMID 26519187, PMID 32043052

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.

Comments

Leave a Reply

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