Evidence level: C
一个判据如果对任何真值都返回「否」,它就没在测东西——而逐阶样本量不等,恰好把高阶的「不可判定」做成了这种状态。
正文
一、门槛公式里其实有两个自变量,只写了一个
上一轮给的三档门槛写成 LoB = 2.448/√n,LoD 与 LoQ 同构,自变量只有 n。这个写法隐含一个前提:零假设下相位在圆上完全均匀。但真实的脉搏相位记录从来不是纯均匀——每一阶都有自己的分散度 σ_k,k=1 时相位集中、k=6 时散得很开,这是同一批数据的常态。一旦 σ_k 不同,「无共享」这个零假设就不再是均匀分布,而是某个宽度的对称分布。Rayleigh 统计量对「有宽度但无方向」的分布并不免疫:宽度越大,样本估计的合向量长度越容易在纯噪声里偶然长出一截。也就是说,门槛里其实有两个自变量——样本量与该阶的固有分散度——而公式只承认了前一个。把所有阶套在同一个数值门槛上,这本身就是一处未经检验的借用。
二、2.448 这个常数已经吃掉了多重性,吃的是没人写明的 20 次
把常数反推一下:2.448² = 5.993,而 Rayleigh 统计量 Z = n·R̄² 在零假设下服从指数分布,P(Z > 5.993) = 0.00249。也就是说,2.448 等价于把单次检验的 α 压到 0.0025,恰好是 0.05 除以 20。这个数字很可疑:它不像从 6 个阶次推出来的(6 次比较应给 2.19),也不像从 18 次(三档门槛 × 六阶)推出来的(应给 2.41)。它对应的正是 20 次这个从没出现在任何一张表里的口径。后果分两头:若实际只报 6 阶,门槛偏严,白丢功效;若实际比较次数超过 20(加上跨组、跨时点很容易超),门槛偏松,假阳性失控。多重性校正必须与比较次数绑定,现在这个绑定被藏进了一个四位小数里。
三、150 人是算术副产品,不是设计:判定线就是 50% 功效线
把 2.448/√n 反过来解,就得到「刚好能判定」的样本量 n = (2.448/R̄)²。若某阶真值 R̄=0.20,n = 149.8——正好 150 人。这个数字看起来像个漂亮的设计目标,其实只是除法结果。更关键的是,n 不是「判定正确」的线,而是「判定等于抛硬币」的线:在 Z 的均值约为 nR̄²+1、方差约为 4nR̄²+2 的近似下,n=150、R̄=0.20 时功效恰为 50%。要 80% 功效,需要的样本量是 (3.27/R̄)²,即 268 人。判定线与决策线差 1.8 倍样本量。* 这意味着现有逐阶表里,凡真值贴近门槛的阶,其状态几乎完全由那一次抽样的运气决定,而报出来的却是一个确定的状态标签。
四、伪迹剔除与真值同向:越弱的阶越容易被剔掉
逐阶剔除之所以危险,不在于它让样本量变小,而在于这个变小与我们要测的东西同向。剔除规则本质上是信噪比规则:某阶幅度小、谱峰不突出 → 信噪比低 → 被判伪迹剔除。而「幅度小」正是「该阶没有共享」的表现。于是真值越小的阶被剔得越狠,剩下样本越少,门槛越高。这不是随机缺失,是典型的信息性缺失,后果是系统性偏倚而非仅仅损失精度。这条链在实测层面有据可查:腕部光电容积脉搏波在站立时信噪比只有 9.0 dB,仰卧时 18.6 dB,同一批 1142 人、只换体位就差了近 10 dB(PMID 40577339);心电设备之间的拒收率差异同样显著(PMID 40872047)。剔除率是体位与设备的函数,不是生理的函数——那么「哪一阶不可判定」里就必然混着体位与设备的贡献。
五、逐阶样本量各不同,跨阶比较双向吃亏
逐阶剔除还制造了一个更隐蔽的后果:各阶样本量互不相等,而「低阶共享高于高阶」这类跨阶比较,用的是几个不同的 n。R̄ 本身对 n 无偏,但它的估计精度随 n 变化:n=100 时 R̄ 的标准误约为 n=200 时的 1.4 倍。小样本阶因此双向吃亏——既更难被判为「有共享」,也更容易被判为「与别的阶不同」。若改用共同剔除(各阶只保留全部有效的同一批人)换取可比性,代价立刻出现:200 人里若高阶只剩 100,共同剔除后低阶也从 200 掉到 100,门槛从 0.173 升到 0.245,升幅 42%。保可比性与保样本量不可兼得,必须二选一,而现在的表两头都不占。
六、「下降 30%」这条警报线没有依据,该报的是安全边际
上一轮建议 n_k 降幅超过 30% 就须逐阶单列。但门槛按平方根走:样本量掉三成,门槛只升 19.6%。这个升幅在真值远离门槛时完全无害,在真值贴着门槛时却足以翻面。所以降幅本身不是判据,真正的判据是安全边际——该阶的合向量长度估计值与 2.448/√n_k 的比值。比值 1.5 以上,剔掉一半样本也不翻;比值 1.05,剔掉 11% 就翻。同一套清洗阈值,对不同阶的杀伤力差出一个量级,而降幅这个数字对此一无所知。
七、测法
在 ≥200 人同日重复脉搏数据上,固定一套剔除规则跑一遍,逐阶报有效样本量 n_k、分散度 σ̂_k、合向量长度 R̄_k 与安全边际;把清洗阈值从 3 dB 改到 6 dB 再到 10 dB,看状态翻转落在哪几阶;另设一条共同剔除对照臂,量化保可比性所付的样本量代价。判据三条:某阶在宽严两档阈值下状态相反 ⇒ 该阶判「数据不足」,不得写「无共享」;安全边际低于 1.1 的阶,其状态一律标注为不确定;各阶 n_k 极差超过 2 倍 ⇒ 逐阶表禁用统一 n,门槛必须逐阶单列并写明多重性口径。
证据与局限
已验证: Rayleigh 统计量的零分布与临界值换算是标准结果,2.448 对应 α=0.0025 可精确反推;样本量不足与数据缺失会系统性扭曲生理信号的衍生特征(PMID 34450799 报告高频心率变异性成分的中位误差达 64%),且缺失机制与观测条件强相关(PMID 40577339、PMID 40872047 均属此类实测)。圆形数据检验的选型与功效差异已被系统比较(PMID 30100666、PMID 31391040)。
部分验证: 「伪迹剔除与真值同向」这条推断建立在「剔除规则=信噪比规则」这一前提下,方向合理但没有针对相位谐波逐阶剔除的直接实验;功效近似式是渐近结果,小样本与小 n_k 下偏差未标定。
待验证: 真实脉搏相位数据里各阶有效样本量的实际落差、三档清洗阈值导致的判定翻转率、以及共同剔除所付代价的具体量级,都需要在同一批 ≥200 人数据上实测。
相关阅读
证据等级: C
参考文献: PMID 40577339, PMID 40872047, PMID 34450799, PMID 30100666, PMID 31391040, PMID 27170904, PMID 29207961, PMID 20029935, PMID 32029234
L1 网络状态:该主题尚未进入 L1 对应网络(边表止于 E-061 呼吸节律 0.1 Hz 谱峰—Mayer 波,状态 open;节点表物理侧仅登记「相位 / 节律 — — — 周期性耦合的关键」一行,无层级无状态)
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 list of \”decidable at which order\” is half determined by the cleaning threshold: varying sample sizes per order quietly shift the threshold.
Evidence level: C
A criterion that returns “no” for any true value is not measuring anything. The unequal sample sizes across orders precisely convert the higher-order “indecidability” into this state.
Body
I. The Threshold Formula Actually Has Two Independent Variables, Yet Only One Is Written
The three-tier thresholds provided previously are written as LoB = 2.448/√n, with LoD and LoQ being isomorphic, implying the independent variable is only n. This notation implicitly assumes that, under the null hypothesis, phases are perfectly uniformly distributed on a circle. However, real pulse phase records are never purely uniform—each order has its own dispersion σ_k (k=1 when phases are concentrated, k=6 when they scatter widely), which is the norm for the same dataset. Once σ_k differs, the “no shared” null hypothesis is no longer a uniform distribution but rather a symmetric distribution of some width. The Rayleigh statistic is not immune to distributions that “have width but no direction”: the wider the width, the more easily the resultant vector length estimated from samples grows spuriously within pure noise. Thus, the threshold actually has two independent variables—sample size and the inherent dispersion of that order—yet the formula acknowledges only the former. Applying all orders to the same numerical threshold is itself an unverified borrowing.
II. The Constant 2.448 Already Swallows Multiplicity, Specifically for the Unstated 20 Times
Reversing the constant: 2.448² = 5.993, and under the null hypothesis, the Rayleigh statistic Z = n·R̄² follows an exponential distribution, with P(Z > 5.993) = 0.00249. This means 2.448 is equivalent to compressing the single-test α to 0.0025, exactly 0.05 divided by 20. This number is suspicious: it does not derive from the 6 orders (which should yield 2.19 for six comparisons) nor from the 18 times (three tiers × six orders, which should yield 2.41). It corresponds precisely to the 20 times—a scope never appearing in any table. The consequences split into two directions: if only 6 orders are reported, the threshold is overly strict, losing power unnecessarily; if the actual number of comparisons exceeds 20 (easily surpassed with cross-group or cross-time-point analyses), the threshold is too loose, leading to runaway false positives. Multiplicity correction must be bound to the number of comparisons; now this binding is hidden within a four-decimal constant.
III. 150 People Is an Arithmetic Byproduct, Not a Design: The Decision Line Is the 50% Power Line
Solving 2.448/√n backwards yields the sample size n = (2.448/R̄)² required to “just decide.” If the true value for an order is R̄=0.20, then n = 149.8—exactly 150 people. This number looks like a beautiful design target, but it is merely the result of division. More critically, n is not the line for “correct decision” but rather the line where “decision equals a coin flip”: under the approximation that Z’s mean is roughly nR̄²+1 and variance is roughly 4nR̄²+2, when n=150 and R̄=0.20, the power is exactly 50%. To achieve 80% power, the required sample size is (3.27/R̄)², i.e., 268 people. The decision line differs from the power line by a factor of 1.8 in sample size.* This means that for any order where the true value is close to the threshold, its state is almost entirely determined by the luck of that single sampling, yet what is reported is a definite state label.
IV. Artifact Removal and True Values Move in the Same Direction: Weaker Orders Are More Likely to Be Removed
The danger of order-by-order removal lies not in reducing sample size, but in that reduction moving in the same direction as what we aim to measure. The removal rule is essentially a signal-to-noise ratio rule: if an order’s amplitude is small and its spectral peak not prominent → low signal-to-noise ratio → judged as artifact and removed. “Small amplitude” is precisely the manifestation of “no shared” for that order. Thus, orders with smaller true values are removed more aggressively, leaving fewer samples and raising the threshold. This is not random missingness but typical informative missingness, leading to systematic bias rather than merely lost precision. This chain is verifiable at the measurement level: photoplethysmography of the wrist has a signal-to-noise ratio of only 9.0 dB when standing and 18.6 dB when supine; the same batch of 1,142 people changing posture alone results in nearly a 10 dB difference (PMID 40577339); rejection rate differences between ECG devices are similarly significant (PMID 40872047). The rejection rate is a function of posture and device, not physiology—thus “which order is indecidable” inevitably mixes contributions from posture and device.
V. Unequal Sample Sizes Across Orders Cause Double Loss in Cross-Order Comparison
Order-by-order removal also creates a more subtle consequence: sample sizes differ across orders, while cross-order comparisons such as “low-order sharing exceeds high-order” rely on several different n values. While R̄ itself is unbiased with respect to n, its estimation precision varies with n: the standard error of R̄ when n=100 is about 1.4 times that when n=200. Small-sample orders thus suffer double loss—they are harder to be judged as “having shared” and more likely to be judged as “different from other orders.” If common removal is adopted (retaining only the same batch of valid people across all orders) to ensure comparability, the cost immediately emerges: among 200 people, if the high order drops to 100, after common removal the low order also falls from 200 to 100, raising the threshold from 0.173 to 0.245, an increase of 42%. Preserving comparability and preserving sample size cannot be achieved simultaneously; one must choose, and the current tables fail at both ends.
VI. The “Drop 30%” Alert Line Has No Basis; What Should Be Reported Is the Safety Margin
The previous recommendation states that if n_k drops by more than 30%, order-by-order listing is required. However, thresholds follow the square root: a 30% drop in sample size raises the threshold by only 19.6%. This increase is harmless when the true value is far from the threshold but sufficient to flip the result when the true value is close. Thus, the drop itself is not the criterion; the real criterion is the safety margin—the ratio of the estimated resultant vector length for that order to 2.448/√n_k. A ratio above 1.5 means halving the sample won’t flip the result; a ratio of 1.05 means dropping just 11% will flip it. Using the same cleaning threshold, the lethality across different orders differs by an order of magnitude, yet the drop percentage tells us nothing about this.
VII. Measurement Protocol
On ≥200 people with repeated pulse data on the same day, fix a set of removal rules and run once, reporting valid sample sizes n_k, dispersion σ̂_k, resultant vector length R̄_k, and safety margin for each order; adjust the cleaning threshold from 3 dB to 6 dB then to 10 dB, observing which orders the state flips; additionally set up a common removal control arm to quantify the sample size cost for preserving comparability. Three criteria: if an order’s state is opposite under wide and narrow thresholds ⇒ mark that order as “insufficient data,” do not write “no shared”; for orders with safety margin below 1.1, mark their state as uncertain; if the range of n_k across orders exceeds twice ⇒ forbid unified n in order-by-order tables; thresholds must be listed individually per order with the multiplicity scope stated.
Evidence and Limitations
Verified: The null distribution and critical value conversion of the Rayleigh statistic are standard results; 2.448 corresponds to α=0.0025 and can be precisely reversed; insufficient sample size and data missingness systematically distort derived features of physiological signals (PMID 34450799 reports median error up to 64% for high-frequency heart rate variability components), and missingness mechanisms strongly correlate with observation conditions (PMID 40577339, PMID 40872047 both belong to such measured cases). Selection of circular data tests and their power differences have been systematically compared (PMID 30100666, PMID 31391040).
Partially Verified: The inference that “artifact removal and true values move in the same direction” rests on the premise that “removal rule = signal-to-noise ratio rule,” which is directionally reasonable but lacks direct experiments on order-by-order phase harmonic removal; the power approximation formula is an asymptotic result, with deviations under small samples and small n_k not calibrated.
To Be Verified: The actual disparity in valid sample sizes across orders in real pulse phase data, the flip rate caused by three-tier cleaning thresholds, and the specific magnitude of cost for common removal all require measurement on the same batch of ≥200 people data.
Further Reading
- Zero-point reading determined by n, but n differs across orders
- Phase dispersion is separate for each order: spectral nonlinearity and per-order σ
- Biological Frequency Code Research Pool
Evidence Level: C
References: PMID 40577339, PMID 40872047, PMID 34450799, PMID 30100666, PMID 31391040, PMID 27170904, PMID 29207961, PMID 20029935, PMID 32029234
L1 Network Status: This topic has not yet entered the corresponding L1 network (edge table stops at E-061 respiratory rhythm 0.1 Hz spectral peak—Mayer wave, state open; node table physical side only registers “phase/rhythm — — — key periodic coupling” row, no hierarchy, no state)
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