Evidence level: C
一个比值有没有用,不看分子准不准,看分母有多大——R₆ 的分母只有 8%,所以它在大样本里是个好指标,在真实样本量里是一台造「形状」的机器。
正文
上一篇留下一把尺子:先从一阶合向量 S₁ 反解离散度 σ̂,再用该 σ̂ 下的正态分布预测各阶应为多少,把实测值除以预测值,得到无量纲残差谱 R_k。正态族恒等于 1,偏离都归给形状。在那批数里这把尺子很锋利:σ=1.43 小时、n=200 时正态族 R₆ 是 1.15±0.52,等权双峰族 8.26±0.11,几乎不重叠。
问题出在那个「n=200」。真实研究不做 200 人。
一、先看分母有多小
σ=1.43 小时时,正态分布第六阶预测值是 8.02%。R₆ 就是「拿百分之几去除以百分之八」——分母本身是个位数量级,任何估计误差在除法里都被放大一个数量级。
放大多少可以算死。第六阶预测值对 σ 的对数敏感度是 k²ω²σ̂²,k=6 时等于 5.05:σ̂ 偏 10%,分母就偏 50%。而 σ̂ 从 S₁ 反解,S₁ 在 n=60 时的标准差约 9.3%——光这一条通路就给分母带来近 47% 的抖动。第八阶敏感度 8.97,分母只剩 1.13%。
二、我把它算了一遍:n=60 的误判率不是两成,是四成七
用参数 bootstrap 在正态族上生成 R₆ 零分布(每档 40000 次重复,σ=1.43,周期 24 小时):n=60 中位数 1.437、四分位距 0.951–2.138、P(R₆>1.5)=47.1%;n=200 中位 1.148、P(>1.5)=25.3%;n=480 中位 1.061、P(>1.5)=11.2%。
母文说 n=60 时「约两成正态人群会被 1.5 误判」。实际 47.1%,差 2.4 倍。 中位数也对不上:母文报 1.49,我复现 1.437;四分位距母文 0.97–2.20,我 0.95–2.14。方向量级都对,但「两成」会被当成安全边际,实际接近抛硬币。
为什么错这么多?1.5 这条线根本不是分位数,是凭感觉定的整数。正态族 n=60 时 R₆ 的 95 分位是 3.82,99 分位 5.86。1.5 落在中位数和 25 分位之间,连「高于一半人」都算不上。
三、偏倚的方向不是「抬高」,是「往 1 挤」
直觉说「小样本噪声大,R 被抬高」,实测不是。把等权双峰族(真 R₆=7.79)放进同一口径:n=60 时正态 1.437 比双峰 8.145,比值 5.67;n=200 比值 7.10;n=480 比值 7.78。
真比值 7.79。n=60 时两族差距被压掉 27%,而正态族自己从 1 抬到 1.44。小样本同时干两件事:把无形状的抬上去、把有形状的压下来——它在往 1 挤,不是单向抬高。 这对「这人有形状」的结论危害最大:它会把大量无形状的人塞过线;而「这人是双峰」的结论只会被低估。
四、文档里那条「安全阈值改 4」只在极小样本才安全
母文补救是「阈值设 R₆>4 并同时报告 n」。每档假阳性:n=22 时 19.3%,n=30 是 14.4%,n=50 是 6.4%,n=71 是 2.9%,n=100 是 1.0%,n≥200 归零。
所以 4 这条线在 n≥100 时确实能用,但 n=22(正是那篇脉搏谐波信度研究的规模,PMID 27170904 只有 22 人)时五次里有一次把纯正态判成有形状。代价是检出力:真形状族 R₆=2.26(均匀族,六族里最靠近 1 的真形状)在 n=200 时 P(R₆>4)=0%。
固定阈值不可能同时管住假阳性和检出力,因为零分布宽度本身随 n 变。 唯一正确做法是按 n 分档报分位数。
五、能不能用 bootstrap 把这个偏倚修掉
这是最容易想到的补救,也是最容易骗人的一步。
取 bootstrap 的 95 分位当阈值。 n=60 上做 150 次重抽样,95 分位中位数是 4.20。看着合理,但它只是把同一条偏倚线用重抽样复现一遍——它不回到 1.0,它回到 4.2。分位数只说得清「这个 n 下多大算异常」,讲不出「真值应该是 1」。
基本 bootstrap 偏差校正(2R̂−重抽样均值)。三组独立种子、n=60/120/200 各 400 组,校正后中位数落在 0.95–1.05(真值 1.0)——真修好了。代价是方差:n=60 校正后四分位距 0.29–1.65,比未校正更宽。
BCa 这条路已有定论(PMID 26074184 / 35712166):bootstrap 区间覆盖率系统性低于名义水平,小样本下所有方法都不可靠;BCa 中第一类错误率偏高、百分位法偏低,中间是一整个连续谱,没有哪个是「对的」。
六、最要紧的一层:这个比值根本不是「比值」
前五节都按「R = 分子/分母」处理,好像两者各自有噪声。但真实构造里分母不是独立估出来的——它由 S₁ 反解的 σ̂ 算出,σ̂ 只由 S₁ 决定。所以分子 S₆ 与分母 pred₆ 共享同一个随机源。
我量了这个相关:n=60 时 +0.333,n=200 时 +0.377,n=480 时 +0.402。正相关且随 n 增强——n 大时 S₁ 估得越准,pred₆ 越贴着 S₆ 背后的真实 σ 走。
这意味着:把 R₆ 当「两个独立估计量的比值」算标准误一定算错。 独立假设预测的变异系数在 n=200 时是 0.513、实测 0.449;n=480 时预测 0.383、实测 0.332——高估 14%–15%。谁按独立假设做功效计算会以为需要更多样本、实际不用那么多,但那份「不用」建立在他同时接受 47% 假阳性率的基础上。
更麻烦的是,这层相关让 bootstrap 校正变得可疑。基本 bootstrap 假设「偏倚光滑、重抽样能复现它」。当分子分母共享同一个 S₁ 时,重抽样同时扰动两者、偏倚结构被一起复制——校正之所以数值上有效,可能只是碰巧在这个构造下成立。
七、落到可执行的层面
禁报单点阈值。 任何报 R_k 的地方必须同时报 n,并按 n 查分档阈值。我把 n=22 到 1000 的 95 分位算出来了,可以直接抄:22→9.0,30→6.8,40→5.3,50→4.5,60→3.9,80→3.2,100→2.9,120→2.7,150→2.4,200→2.2,300→1.9,480→1.7,700→1.6,1000→1.5。注意表末端:n=1000 时阈值才降到 1.5。1.5 这条线只在千人样本上才有 5% 假阳性,在小样本上的真实含义只是「高于一半人」。
弱形状要单独定 n。 真形状越弱需要的样本越大。均匀族真 R₆=2.26(六族里最靠近 1 的真形状),用各 n 下 95 分位作阈,检出力:n=150→47%,200→59%,300→71%,400→81%,500→86%,800→96%。要做到 80% 检出力得 400 人——而这正撞上现实:一致性研究连续终点样本量中位数只有 50、κ 型终点 71(PMID 36123642),那篇脉搏谐波信度研究只有 22 人(PMID 27170904)。
报 R 必须报「不可判定区间」。 n=200 时 R₆ 的 95 分位在 2.1 以上,那么 1.0 到 2.1 这一段根本无法区分「没形状」和「有一点形状」。任何「R₆ 轻度升高说明有轻微结构」的结论,在这个样本量下都是把噪声当信号。
证据与局限
已验证(本文自算,可复现):
– 正态族 R₆ 零分布:n=60 中位 1.437、IQR 0.951–2.138、P(>1.5)=47.1%;n=200 中位 1.148、P(>1.5)=25.3%;n=480 中位 1.061、P(>1.5)=11.2%。母文「n=60 约两成」低估 2.4 倍。
– 分母敏感度解析式 dln(pred_k)/dln(σ)=k²ω²σ²:k=6 时 5.05,k=8 时 8.97。
– 小样本偏倚是「向 1 收缩」而非单向抬高:等权双峰/正态两族比值从 n=480 的 7.78 压到 n=60 的 5.67(真值 7.79)。
– 母文推荐的阈值 4:n=22 假阳性 19.3%、n=30 14.4%、n=100 1.0%、n≥200 归零;但对真形状 R₆=2.26 的族在 n=200 时检出力为 0。
– 分子分母相关 corr(S₆,pred₆):n=60 为 +0.333、n=480 为 +0.402;独立假设高估变异系数 14%–15%。
– 按 n 的 95 分位阈值表(n=22→9.0 至 n=1000→1.5),可直接照用。
– 基本 bootstrap 偏差校正把 n=60 中位数从 1.44 拉回 0.95–1.05(三组种子),代价是四分位距变宽。
部分验证:
– 「真实研究样本量偏小」有实测支撑:一致性研究连续终点中位 50(IQR 25–100)、κ 终点 71(PMID 36123642,82 篇研究、仅 33% 给了样本量依据);脉搏谐波信度研究 22 人(PMID 27170904,前 11 阶 ICC 0.83–0.96)。但这两批研究不是同一领域,不能直接说「脉搏谐波研究的中位样本量是 50」。
– 「bootstrap 置信区间覆盖率系统性不足」「BCa 第一类错误偏高」有方法学文献(PMID 26074184 / 35712166),但都是在其他领域的模拟,未在相位谐波数据上验过。
– 「比值型估计量在小样本下有偏」这一普遍现象有独立证据:log 响应比的蒙特卡洛显示小样本有偏、需专门校正估计量(PMID 26405731);HRV 谱指标因有限采样率产生偏倚、HF 被高估而 LF/HF 被低估(PMID 15132314)。方向与本文一致(都是分母/低频端被高估),但那是采样率导致、不是样本量导致,机制不同。
待验证:
– 全部结论建立在 σ=1.43 小时、δ 峰(无核宽)的正态族上;换成有核宽的真实分布时,S_k 的真值会变,零分布的绝对位置会平移。核宽取多少,没有实测。
– 我只测了正态族作零分布。真实人群是「主群正态 + 约 15% 非单峰亚群」(上一轮的混合场景),这个混合情形的零分布没算,它会比纯正态更宽,也就是说真实的假阳性率可能比我报的 47% 更高。
– 三种偏差校正我只测了基本 bootstrap 一种,没测 BCa 与 Stine 型;它们在这个「共享 S₁」构造下的表现是未知的。
– 均匀族真 R₆=2.26 这个标定依赖我按母文口径复现的族定义;母文表里均匀族 R₆ 是 2.26,我复现 2.26,两者对上了,但峰宽、峰数等其他族定义的可比性未逐族核。
相关阅读
L1 网络状态:该主题尚未进入 L1 对应网络 —— EDGE_TABLE 中「谐波」概念登记 147 篇、最高证据级 A(代表 PMID 21622820),但边表止于 E-061(呼吸节律 0.1Hz — Mayer 波,open),全表无任何与「比值型估计量 / 零分布 / 阈值分档 / 小样本偏倚」相关的边;CORRESPONDENCE_NETWORK 节点表仅登记「相位 / 节律 — 周期性耦合的关键」一行(有对象、无层级、无状态)。按硬规则证据等级封顶 C。
证据等级: C
参考文献: PMID 36123642, PMID 27170904, PMID 26405731, PMID 26074184, PMID 35712166, PMID 15132314
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。
The criterion that \”R₆ exceeding 1.5 implies shape\” misclassifies nearly half of real samples: the denominator is only 8%, and no one has calculated the bias of division.
Evidence level: C
Whether a ratio is useful depends not on the numerator’s accuracy, but on the denominator’s magnitude. Since R₆’s denominator is only 8%, it serves as a good indicator in large samples but acts as a machine generating “shape” in real-world sample sizes.
Main Text
The previous section left a ruler: first, invert the discrete variance σ̂ from the first-order coherence vector S₁, then predict the expected values for each order under that σ̂ using the normal distribution. Divide the measured values by these predicted values to obtain the dimensionless residual spectrum R_k. Since the normal family is always equal to 1, any deviation is attributed to shape. Among those numbers, this ruler is sharp: with σ=1.43 hours and n=200, the normal family R₆ is 1.15±0.52, while the equal-weighted bimodal family is 8.26±0.11, with almost no overlap.
The problem lies in “n=200”. Real studies do not use 200 people.
I. First, Examine How Small the Denominator Is
When σ=1.43 hours, the predicted sixth-order value of the normal distribution is 8.02%. R₆ is essentially “dividing a percentage by eight percent”—the denominator itself is of unit order, so any estimation error in the division is amplified by an order of magnitude.
How much amplification can be calculated exactly. The logarithmic sensitivity of the sixth-order prediction to σ is k²ω²σ̂², which equals 5.05 when k=6: if σ̂ is off by 10%, the denominator is off by 50%. Since σ̂ is inverted from S₁, and S₁ has a standard deviation of about 9.3% when n=60—this single pathway alone introduces nearly 47% jitter in the denominator. The eighth-order sensitivity is 8.97, leaving the denominator at only 1.13%.
II. I Calculated It Again: The Misclassification Rate for n=60 Is Not 20%, It Is 47%
Using parameter bootstrap on the normal family to generate the zero distribution of R₆ (40,000 repetitions per tier, σ=1.43, 24-hour cycle): the median for n=60 is 1.437, interquartile range 0.951–2.138, P(R₆>1.5)=47.1%; for n=200 the median is 1.148, P(>1.5)=25.3%; for n=480 the median is 1.061, P(>1.5)=11.2%.
The original paper states that for n=60, “about 20% of the normal population is misclassified by 1.5.” The actual rate is 47.1%, nearly 2.4 times higher. The medians also do not match: the paper reports 1.49, while I reproduced 1.437; the interquartile ranges are 0.97–2.20 in the paper versus 0.95–2.14 here. The direction and magnitude are correct, but “20%” is treated as a safety margin when in reality it is close to flipping a coin.
Why is there so much error? The line at 1.5 is not a quantile; it is an integer set by intuition. For the normal family with n=60, the 95th percentile of R₆ is 3.82, and the 99th percentile is 5.86. 1.5 falls between the median and the 25th percentile, not even reaching “higher than half the people.”
III. The Direction of Bias Is Not “Elevating,” It Is “Squeezing Toward 1”
Intuition suggests “small samples have more noise, so R is elevated,” but the measurement shows otherwise. Placing the equal-weighted bimodal family (true R₆=7.79) into the same framework: for n=60, the normal 1.437 versus bimodal 8.145 gives a ratio of 5.67; for n=200 the ratio is 7.10; for n=480 the ratio is 7.78.
The true ratio is 7.79. For n=60, the gap between the two families is compressed by 27%, while the normal family itself rises from 1 to 1.44. Small samples do two things simultaneously: they lift the shapeless up and press the shaped down—it squeezes toward 1, not just elevates unidirectionally. This is most harmful to the conclusion “this person has shape”: it allows many shapeless individuals to cross the line; meanwhile, the conclusion “this person is bimodal” will only be underestimated.
IV. The Document’s “Safe Threshold Changed to 4” Is Only Safe for Extremely Small Samples
The paper’s remedy is “set the threshold at R₆>4 and report n.” False positive rates: n=22 is 19.3%, n=30 is 14.4%, n=50 is 6.4%, n=71 is 2.9%, n=100 is 1.0%, and n≥200 drops to zero.
Thus, the line at 4 is indeed usable for n≥100, but for n=22 (the exact scale of the pulse harmonic reliability study, PMID 27170904 with only 22 people) one out of five times a pure normal is judged to have shape. The cost is sensitivity: the true shape family with R₆=2.26 (uniform family, the closest true shape to 1 among six families) has P(R₆>4)=0% at n=200.
Fixed thresholds cannot simultaneously control false positives and sensitivity, because the width of the zero distribution itself varies with n. The only correct approach is to report quantiles stratified by n.
V. Can Bootstrap Correct This Bias?
This is the easiest fix to think of, and also the most deceptive step.
Take the bootstrap 95th percentile as the threshold. For n=60, performing 150 resamples, the median of the 95th percentile is 4.20. It looks reasonable, but it merely reproduces the same bias line via resampling—it does not return to 1.0, it returns to 4.2. Quantiles only clarify “what counts as abnormal for this n,” not “the true value should be 1.”
Basic bootstrap bias correction (2R̂−resample mean). Three independent seeds, n=60/120/200 each with 400 groups; after correction, the median falls in 0.95–1.05 (true value 1.0)—truly fixed. The cost is variance: for n=60, the corrected interquartile range widens to 0.29–1.65, broader than uncorrected.
BCa has already been settled (PMID 26074184 / 35712166): bootstrap interval coverage is systematically below the nominal level, making all methods unreliable for small samples; BCa has an elevated Type I error rate, percentile methods are too low, and there is a whole continuum in between—none of them is “correct.”
VI. The Most Critical Layer: This Ratio Is Not Even a “Ratio”
The first five sections treat “R = numerator/denominator,” as if both have independent noise. But in the true construction, the denominator is not estimated independently—it is calculated from σ̂ inverted from S₁, and σ̂ depends solely on S₁. Thus, the numerator S₆ and the denominator pred₆ share the same random source.
I measured this correlation: +0.333 for n=60, +0.377 for n=200, and +0.402 for n=480. It is positively correlated and strengthens with n—as n grows, S₁ estimates more accurately, so pred₆ adheres closer to the true σ behind S₆.
This means: calculating the standard error of R₆ as “the ratio of two independent estimates” is necessarily wrong. The coefficient of variation predicted under the independence assumption is 0.513 for n=200 (actual 0.449) and 0.383 for n=480 (actual 0.332)—overestimated by 14%–15%. Anyone performing power calculations under the independence assumption will think more samples are needed, when in fact fewer suffice, but that “fewer” is built on accepting a 47% false positive rate.
More problematically, this correlation makes bootstrap correction suspicious. Basic bootstrap assumes “bias is smooth and resampling can reproduce it.” When numerator and denominator share the same S₁, resampling disturbs both simultaneously, copying the bias structure together—the numerical effectiveness of correction may merely be coincidental under this construction.
VII. Down to Executable Levels
Prohibit single-point thresholds. Any report of R_k must simultaneously report n and consult stratified thresholds by n. I have computed the 95th percentile for n=22 to 1000, ready for direct use: 22→9.0, 30→6.8, 40→5.3, 50→4.5, 60→3.9, 80→3.2, 100→2.9, 120→2.7, 150→2.4, 200→2.2, 300→1.9, 480→1.7, 700→1.6, 1000→1.5. Note the end of the table: the threshold only drops to 1.5 at n=1000. The line at 1.5 yields only 5% false positives on a thousand-sample, while its real meaning for small samples is merely “higher than half the people.”
Weak shape requires separate n. The weaker the true shape, the larger the required sample. For the uniform family with true R₆=2.26 (the closest true shape to 1 among six families), using the 95th percentile as threshold across n: sensitivity is n=150→47%, 200→59%, 300→71%, 400→81%, 500→86%, 800→96%. Achieving 80% sensitivity requires 400 people—which coincides with reality: the median sample size for continuous endpoints in consistency studies is only 50 (IQR 25–100), and κ-type endpoints is 71 (PMID 36123642, 82 studies, only 33% provided sample size basis); the pulse harmonic reliability study had only 22 people (PMID 27170904).
Report R with “Indecisive Interval.” For n=200, the 95th percentile of R₆ is above 2.1; thus, the range from 1.0 to 2.1 cannot distinguish “no shape” from “some shape.” Any conclusion that “slight elevation of R₆ indicates slight structure” is treating noise as signal at this sample size.
Evidence and Limitations
Verified (self-calculated in this paper, reproducible):
– Normal family R₆ zero distribution: median 1.437 for n=60, IQR 0.951–2.138, P(>1.5)=47.1%; median 1.148 for n=200, P(>1.5)=25.3%; median 1.061 for n=480, P(>1.5)=11.2%. The paper’s “about 20%” for n=60 underestimates by 2.4 times.
– Analytical expression for denominator sensitivity dln(pred_k)/dln(σ)=k²ω²σ²: k=6 gives 5.05, k=8 gives 8.97.
– Small-sample bias is “contraction toward 1,” not unidirectional elevation: the ratio of equal-weighted bimodal/normal families drops from 7.78 at n=480 to 5.67 at n=60 (true value 7.79).
– Paper-recommended threshold 4: false positives are 19.3% at n=22, 14.4% at n=30, 1.0% at n=100, and zero for n≥200; but sensitivity for the true shape family with R₆=2.26 is 0% at n=200.
– Correlation between numerator and denominator corr(S₆,pred₆): +0.333 for n=60, +0.402 for n=480; independence assumption overestimates coefficient of variation by 14%–15%.
– Table of 95th percentile thresholds stratified by n (n=22→9.0 to n=1000→1.5) can be directly applied.
– Basic bootstrap bias correction pulls the median for n=60 from 1.44 back to 0.95–1.05 (three seed groups), at the cost of widened interquartile range.
Partially Verified:
– “Real studies have smaller sample sizes” has empirical support: median 50 for continuous endpoints (IQR 25–100), κ endpoints 71 (PMID 36123642, 82 studies, only 33% provided sample size basis); pulse harmonic reliability study had 22 people (PMID 27170904, ICC for first 11 orders 0.83–0.96). However, these two batches are not from the same domain, so one cannot directly state “the median sample size of pulse harmonic studies is 50.”
– “Bootstrap confidence interval coverage is systematically insufficient” and “BCa has elevated Type I error” have methodological literature (PMID 26074184 / 35712166), but these are simulations in other domains, not verified on phase harmonic data.
– “Ratio-type estimators are biased in small samples” is a general phenomenon supported by independent evidence: Monte Carlo of log response ratio shows bias in small samples, requiring specific correction (PMID 26405731); HRV spectral indices are biased due to finite sampling rate, with HF overestimated and LF/HF underestimated (PMID 15132314). The direction matches this paper (denominator/low-frequency end overestimated), but that is due to sampling rate, not sample size, so the mechanism differs.
Pending Verification:
– All conclusions are based on the normal family with σ=1.43 hours and δ peak (no kernel width); switching to a true distribution with kernel width will shift the true value of S_k and the absolute position of the zero distribution. How much kernel width to choose has no empirical basis yet.
– I only measured the normal family for the zero distribution. Real populations consist of “main group normal + about 15% non-unimodal subgroup” (the previous mixed scenario); the zero distribution for this mixed case was not calculated, and it will be broader than pure normal, meaning the real false positive rate may be higher than the reported 47%.
– Among three bias corrections, I only tested basic bootstrap; BCa and Stine-type methods were not evaluated. Their performance under this “shared S₁” construction is unknown.
– The calibration of uniform family true R₆=2.26 depends on reproducing the family definition per the paper’s scale; the paper’s table lists uniform family R₆ as 2.26, which I reproduced, so they match, but comparability of other family definitions (peak width, peak count, etc.) was not checked per family.
Related Reading
- Dispersion is Myopic, Shape Is Audible
- The List of “Which Order Is Decidable,” Half Is Set by Cleaning Thresholds
- Biological Frequency Code (Research Pool 3)
L1 Network Status: This topic has not yet entered the corresponding L1 network. The EDGE_TABLE registers 147 articles on “harmonic” concepts, with maximum evidence level A (representing PMID 21622820), but the edge table stops at E-061 (respiratory rhythm 0.1Hz — Mayer wave, open); the entire table contains no edges related to “ratio-type estimators / zero distribution / threshold stratification / small-sample bias”; the CORRESPONDENCE_NETWORK node table only registers “phase/rhythm — periodic coupling key” (with objects, no hierarchy, no state). Hard rules cap evidence level at C.
Evidence Level: C
References: PMID 36123642, PMID 27170904, PMID 26405731, PMID 26074184, PMID 35712166, PMID 15132314
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.
Leave a Reply