标准化之后还剩什么:三质判定的族裔残余差异是位置、离散还是形状

Written by

in

Language

Evidence level: B

z 分数只干了两件事——把均值挪到零、把方差压到一。如果残余的族裔差异还剩下,那它只可能藏在第三样东西里:分布的形状。

正文

上一篇我们把”相对模式匹配”当成一把自带的尺子,问它能不能消除三质判定的跨族裔漂移。这一篇接着往下问一层:消除不掉的那部分,到底是什么成分?

先看 z 分数实际做了什么。对每个族裔,它减去该组的均值、再除以该组的标准差。于是位置(均值)被强行归零,离散(方差)被强行归一。做完这两步,两个族裔的分布看起来”同一把尺子”了——但只对前两阶矩成立。分布的形状:偏度、峰度、模态数、分位数之间的非线性关系,全都原封不动留着。z 分数是平移加缩放,而平移和缩放改变不了形状。

那么”残余 = 形状”这个推论成立吗?不一定,这里有个容易被忽略的坑:z 分数只在你能准确估计每个族裔的均值和方差时才真正抹平位置和离散。小样本、强非正态、异方差的情形下,估计误差会以”伪形状”的形式混进残余里。也就是说,残余空间里同时躺着两种东西——真的形状差异,和标准化的估计噪声。这两者必须分开,否则会把方法误差误读成族裔生理差异。

要分开它们,需要一把能”拆解差异”的尺子。统计上有现成工具:2-Wasserstein 距离的三项分解。Schefzik 等 2021 年在单细胞分析里给出精确分解(PMID 33792651),把两个分布的平方距离拆成三项——位置项 (μA−μB)²、离散项 (σA−σB)²、形状项 2σAσB(1−ρ),其中 ρ 是分位数-分位数图的相关系数,专门度量形状。这个方法的价值在于:它不满足于说”两个分布不同”,而是回答”不同在哪个成分、各占多少”。同一个框架稍加推广就能用在 HRV 上。

更直接的建模工具是 GAMLSS(广义可加模型的位置-尺度-形状框架)。2023 年的一篇综述系统介绍了它(PMID 37502671):传统回归只建模均值,GAMLSS 可以把分布的均值、方差、偏度、峰度分别建模成协变量的函数。换到我们的问题上,就是直接问一句:”族裔这个变量,对偏度和峰度有没有独立效应?”——这是把”残余是位置、离散还是形状”从描述性问题变成可检验的统计假设。

为什么形状成分值得认真对待?因为对 HRV 来说,形状从来不是小事。HRV 指标(SDNN、RMSSD、频域功率)的分布是强烈右偏的,这是教科书级的事实,也是为什么几十年来都要先做对数变换再统计(PMID 29034226、PMID 26849653)。形状成分大,意味着”只对齐均值方差”的标准化天然会漏掉一大块信息。而且形状本身携带生理信号:Hayano 等 2011 年发现,非高斯性指数 λ(本质上是分布尾部和峰部形状的度量)能独立于所有传统 HRV 指标预测心梗后心源性死亡(PMID 21994495)。形状不是噪声,它是信号。

族裔差异在形状层面也确实有痕迹。SWAN 睡眠研究用归一化高频功率(HF nu,本身已经是相对量、天然扣掉了绝对功率尺度)仍然测到显著族裔差异(PMID 24077772)——既然相对量还有差异,那差异就不全在绝对尺子上。另一个更精巧的线索来自亚裔与白人儿童的对照:时域指标 SDNN、RMSSD 没有族裔差异,但频域 LF:HF 比值有差异(PMID 16770356)。同一个体、同一段记录,换个指标族裔差异就从无变有——说明”哪个成分携带族裔信息”是高度指标特异的。南亚儿童的长期 HRV(SDANN)显著低于白人儿童,且在统计调整了体力活动、体脂、最大摄氧量之后依然存在(PMID 23444234),进一步排除了”残余只是生活方式混杂”的解释。工程实践其实早就默认了这一点:MESA 直接给出四个族裔、分性别的 HRV 参考范围(PMID 27396499),等于是承认”一把绝对尺子量不了所有人”。

把这些拼起来,可检验的预测就清楚了:z 标准化之后,残余族裔差异中形状成分的占比显著大于零;而且随标准化层次加深(原始 → z 分数 → 分位数/秩归一化),位置项和离散项会趋近于零,形状项占比则相对上升,成为唯一剩下的成分。分位数归一化是个关键对照——它按定义把位置和离散完全抹掉,所以它之后的残余就是”形状成分”的纯测量。

裁决实验可以这样设计:取多族裔、同三质分型队列,逐 HRV 指标计算 Wasserstein 三项分解,记录位置/离散/形状占比;在三种标准化层次上各做一遍,画出”形状占比 vs 标准化层次”的曲线;再用 GAMLSS 检验族裔对偏度和峰度的独立效应;最后用置换检验(打乱族裔标签)建立零分布,确认形状成分不是估计噪声。若形状占比确实随层次单调上升并稳定非零,那就说明——族裔感知校准如果只重标定均值和方差,做的恰好是 z 分数已经做过的事,残余永远清不干净;真正需要的可能是分布层面的映射(分位数映射、形状校正),而不是简单的线性重标定。

对三质判定而言,这一步的意义在于划清责任边界:相对模式匹配负责干掉位置和离散,而形状这一块要么承认它必须靠族裔分层兜底,要么就得用更重的分布映射去碰——但碰之前得先确认,形状里装的到底是族裔信息,还是体质信息,抑或两者纠缠在一起。

证据与局限

已验证: HRV 指标分布强烈右偏、需对数变换(PMID 29034226、26849653);2-Wasserstein 距离可精确分解为位置/离散/形状三项(PMID 33792651);GAMLSS 可分别建模均值、方差、偏度、峰度(PMID 37502671);HRV 分布形状(非高斯性)独立预测心源性死亡(PMID 21994495);族裔差异在归一化相对量上仍存在(PMID 24077772),且呈指标特异性(PMID 16770356),部分独立于体力活动与体脂(PMID 23444234);四族裔分性别参考范围已建立(PMID 27396499)。

部分验证: z 标准化能抹平位置与离散成分(由定义与 PMID 33792651 的分解框架支持),但”残余是否以形状成分为主”尚无直接针对族裔 HRV 的研究,属方法学外推。

待验证: 多族裔 HRV 数据经 z 标准化后,残余差异中位置/离散/形状三项的占比各是多少;该占比随标准化层次(原始→z→分位数归一化)如何变化;小样本估计误差贡献的”伪形状”占比随样本量的衰减曲线。

相关阅读

证据等级: B
参考文献: PMID 33792651, PMID 37502671, PMID 21994495, PMID 24077772, PMID 16770356, PMID 23444234, PMID 27396499, PMID 29034226, PMID 26849653


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。

What Remains After Standardization: Are Ethnic Residual Differences in Three Quality Judgments Due to Location, Dispersion, or Shape

Evidence level: B

z-scores accomplish only two things: they shift the mean to zero and compress the variance to one. If ethnic differences persist in the residuals, they must be hidden in the third element: shape.

Main Text

In the previous article, we treated “relative pattern matching” as a built-in ruler and asked whether it could eliminate cross-ethnic drift in tridosha classification. This article asks one layer deeper: What constitutes the portion that cannot be eliminated?

First, let us examine exactly what z-scores do. For each ethnic group, they subtract the group mean and divide by the group standard deviation. Consequently, location (mean) is forced to zero, and dispersion (variance) is forced to unity. After these two steps, the distributions of two ethnic groups appear “measurable by the same ruler”—but this holds only for the first two moments. The shape of the distribution—skewness, kurtosis, modality counts, and non-linear relationships between quantiles—remains completely unchanged. Z-scores involve translation and scaling; neither translation nor scaling alters shape.

So, does the inference “Residual = Shape” hold? Not necessarily. There is a pitfall often overlooked: z-scores only truly flatten location and dispersion when the mean and variance for each ethnic group can be accurately estimated. Under conditions of small samples, strong non-normality, or heteroscedasticity, estimation errors can infiltrate the residuals as “pseudo-shape.” Thus, the residual space contains two distinct entities simultaneously: true shape differences and standardized estimation noise. These must be separated; otherwise, methodological errors may be misread as ethnic physiological differences.

To separate them, we need a ruler capable of “decomposing differences.” Statistics offers a ready-made tool: the three-term decomposition of the 2-Wasserstein distance. Schefzik et al. (2021) provided an exact decomposition in single-cell analysis (PMID 33792651), splitting the squared distance between two distributions into three terms: a location term $(\mu_A-\mu_B)^2$, a dispersion term $(\sigma_A-\sigma_B)^2$, and a shape term $2\sigma_A\sigma_B(1-\rho)$, where $\rho$ is the correlation coefficient of quantile-quantile plots, specifically measuring shape. The value of this method lies in its refusal to merely state “two distributions are different”; instead, it answers “in which component do they differ and what proportion does each occupy.” With slight extension, this same framework can be applied to HRV.

A more direct modeling tool is GAMLSS (Generalized Additive Models for Location, Scale, and Shape). A 2023 review systematically introduces it (PMID 37502671): while traditional regression models only the mean, GAMLSS can model the distribution’s mean, variance, skewness, and kurtosis separately as functions of covariates. Applied to our question, this allows us to directly ask: “Does the ethnic variable have an independent effect on skewness and kurtosis?”—transforming the descriptive question “Is the residual due to location, dispersion, or shape?” into a testable statistical hypothesis.

Why is the shape component worthy of serious attention? Because for HRV, shape is never trivial. The distributions of HRV metrics (SDNN, RMSSD, frequency-domain power) are strongly right-skewed; this is textbook fact and explains why logarithmic transformation has been standard practice for decades (PMID 29034226, PMID 26849653). A large shape component implies that standardization aligning only means and variances naturally misses a substantial block of information. Furthermore, shape carries physiological signals: Hayano et al. (2011) found that the non-Gaussianity index $\lambda$ (essentially a measure of distribution tail and peak shape) can predict post-myocardial infarction cardiac death independently of all traditional HRV metrics (PMID 21994495). Shape is not noise; it is signal.

Ethnic differences indeed leave traces at the shape level. The SWAN sleep study measured significant ethnic differences using normalized high-frequency power (HF nu, which is itself a relative quantity and naturally removes absolute power scale) (PMID 24077772)—since relative quantities still show differences, the differences cannot be entirely on absolute scales. Another more refined clue comes from comparisons between Asian and White children: time-domain metrics SDNN and RMSSD showed no ethnic differences, but the frequency-domain LF:HF ratio did show differences (PMID 16770356). In the same individual with the same recording, changing the metric turns an absence of ethnic difference into a presence—indicating that “which component carries ethnic information” is highly metric-specific. Long-term HRV (SDANN) in South Asian children was significantly lower than in White children and persisted even after statistical adjustment for physical activity, body fat, and VO2 max (PMID 23444234), further excluding the explanation that “residuals are merely lifestyle confounders.” Engineering practice has long defaulted to this point: MESA directly provides HRV reference ranges for four ethnicities, stratified by sex (PMID 27396499), effectively acknowledging that “a single absolute ruler cannot measure everyone.”

Putting these pieces together yields a clear testable prediction: The proportion of the shape component within residual ethnic differences after z-standardization is significantly greater than zero; moreover, as standardization depth increases (raw → z-score → quantile/rank normalization), the location and dispersion terms approach zero, while the shape component proportion rises relatively to become the sole remaining element. Quantile normalization serves as a critical control—it, by definition, eliminates both location and dispersion entirely, so its residuals constitute a pure measurement of the “shape component.”

The adjudication experiment can be designed as follows: Take a multi-ethnic cohort with the same three-tridosha phenotypes, calculate the three-term Wasserstein decomposition for each HRV metric individually, and record the proportions of location/dispersion/shape. Repeat this across three standardization levels and plot a “Shape Proportion vs. Standardization Level” curve. Then, use GAMLSS to test the independent effects of ethnicity on skewness and kurtosis. Finally, use permutation tests (shuffling ethnic labels) to establish a null distribution and confirm that the shape component is not estimation noise. If the shape proportion indeed rises monotonically with increasing standardization levels and stabilizes at a non-zero value, then it indicates that if ethnic perception calibration focuses only on re-calibrating means and variances, it is doing exactly what z-scores have already done; the residual will never be cleaned up. What is truly needed may be distribution-level mapping (quantile mapping, shape correction) rather than simple linear re-calibration.

For tridosha classification, the significance of this step lies in delineating responsibility boundaries: relative pattern matching is responsible for eliminating location and dispersion, while the shape component must either be acknowledged as requiring ethnic stratification as a safety net or addressed through heavier distribution mapping—but only after confirming whether the shape contains ethnic information, physiological information, or a complex entanglement of both.

Evidence and Limitations

Verified: HRV metric distributions are strongly right-skewed and require logarithmic transformation (PMID 29034226, PMID 26849653); the 2-Wasserstein distance can be precisely decomposed into location/dispersion/shape terms (PMID 33792651); GAMLSS can model mean, variance, skewness, and kurtosis separately (PMID 37502671); HRV distribution shape (non-Gaussianity) independently predicts cardiac death (PMID 21994495); ethnic differences persist in normalized relative quantities (PMID 24077772) and exhibit metric specificity (PMID 16770356), partially independent of physical activity and body fat (PMID 23444234); four-ethnicity sex-stratified reference ranges have been established (PMID 27396499).

Partially Verified: Z-standardization can flatten location and dispersion components (supported by definition and the decomposition framework in PMID 33792651), but whether “residuals are primarily composed of shape” lacks direct research specifically targeting ethnic HRV, representing a methodological extrapolation.

To Be Verified: The specific proportions of location/dispersion/shape within residual differences after z-standardization in multi-ethnic HRV data; how these proportions change across standardization levels (raw → z → quantile normalization); the decay curve of “pseudo-shape” contribution from small-sample estimation error as sample size decreases.

Related Reading

Evidence Level: B
References: PMID 33792651, PMID 37502671, PMID 21994495, PMID 24077772, PMID 16770356, PMID 23444234, PMID 27396499, PMID 29034226, PMID 26849653


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 *