伪形状成分识别与分离点定位:残余形状占比随样本量的衰减曲线,多少样本量后形状占比才可信

Written by

in

Language

Evidence level: C

跨族裔比较三质分布时,即使形状项真的非零,也要先问一句:这里面有多少是族裔间的真形状差异,有多少只是小样本经验分布自己的随机起伏。

正文

上一轮已经裁决:Wasserstein 形状项的取舍必须放在同批配对数据里检验。但在这之前还横着一个更根本的问题——形状项本身可不可信。经验分布只是总体的抽样快照:样本小时,它的偏度、峰度、多峰结构本身就带着随机起伏。这部分起伏会原样混进形状项,形成”伪形状”——看起来像族裔间分布形状不同,实际只是每边样本量不够、经验分布没站稳。位置项和离散项各有大样本理论兜底,形状结构对样本量最敏感,所以这个问题在形状项上最尖锐。

好消息是,伪形状和真形状有一个可分开的数学性质:估计噪声随样本量单调衰减,真形状差异不衰减。样本量往上加,经验分布逐步贴近总体分布,伪形状占比一路下降;而族裔间真实的形状差异不随样本量变化。于是可以画一条”形状项占比对样本量”的衰减曲线:曲线的下降段是估计噪声,曲线最终走平的位置就是真形状,曲线由降到平的转折点就是分离点——分离点之后的样本量,形状占比读数才开始可信。

这个思路在生理测量里已有成熟的先例链:估计量的稳定性依赖数据量,是可测、且被反复测过的量。心梗患者心率与 QT 变异性的短时重现性随记录时长系统变化(PMID 27681306);HRV 各指标的重复性受测量窗长度制约(PMID 15286929);血压与心率短时变异性的重现性实测(PMID 7981010);超短程与标准程 HRV 分析的一致性边界(PMID 39063591);心血管变异性分析中传递函数估计本身的可靠性也随数据量变化(PMID 11465889)。归一化一侧同样如此:经典分位归一化以”各分布相同”为前提(PMID 12925520),组间分布确实不同时需要平滑处理(PMID 29036413),非平衡转录组的归一化选择本身就是一个综述级别的难题(PMID 32039167),而分位回归建立连续常模时估计偏差与精度会直接传导进下游读数(PMID 32483976),分位映射偏置校正的比较研究也确认了这一点(PMID 33670819)。最优传输方法在单细胞组学上做经验测度比较时(PMID 35157031),以及用它学习单细胞扰动响应时(PMID 37770709),估计噪声对比较结果的污染都是已知要面对的问题。本主题缺的不是”现象存在”的证据,而是把衰减曲线在多族裔 HRV 场景下定量画出来的那一次实测。

设计分两步。第一步合成数据校准:构造已知形状差异的总体,在 n=20/50/100/200/500 梯度上反复抽样,拿到纯估计噪声下的占比衰减曲线模板;第二步实测多族裔 HRV 子采样,同样五个梯度逐档重算形状项占比,与模板对照——实测曲线在哪个样本量档离开纯噪声模板、开始走平,那里就是分离点。三情形裁决预注册:若到 n=500 仍在降,说明该量级下形状占比整体不可信,族裔形状结论冻结;若 n≤100 就进入平台,形状项可放心使用;若出现先降后平再降的双平台结构,须按指标分层分别报告。

三条实施细节先行锁定,防裁决被污染。第一,形状占比里的估计噪声有两个来源:族裔内抽样的采样噪声,以及子采样与标准化管线自身引入的处理噪声——合成模板必须把两条噪声源都覆盖,否则实测曲线的下降段会被误读成纯采样效应。第二,每个样本量档的子采样要重复足够次数(预注册 200 次)并报置信带,用置信带宽度收敛与否做辅助判据:带宽也走平的档位才算真分离点,只看占比点估计会把抖动误判成平台。第三,子采样时保留配对结构——同批设备、同记录窗口、同协变量调整——沿用上一轮”同批配对”的裁决口径,避免把管线差异误算进族裔形状差异。

这一步做完,前几轮的形状项裁决才真正站得住:分离点之内的所有族裔形状差异,都应按伪形状处理,不进结论。

证据与局限

已验证: 估计量稳定性依赖数据量/测量窗长度,在 HRV、血压变异性、传递函数估计上有直接实测文献(PMID 27681306、15286929、7981010、11465889);归一化方法的分布假设与估计误差传导有系统研究(PMID 12925520、29036413、32039167、32483976)。

部分验证: 最优传输方法在经验测度比较中受估计噪声影响,已有方法学文献确认(PMID 35157031、37770709),但针对一维 Wasserstein 形状项的”占比随样本量衰减曲线”未见现成实测。

待验证: 多族裔 HRV 子采样下伪形状占比的衰减曲线与分离点位置;三情形判决式对应的族裔形状结论冻结/解冻规则,均为本研究设计提案,尚无直接实测。

相关阅读

证据等级: C
参考文献: PMID 27681306, PMID 15286929, PMID 7981010, PMID 39063591, PMID 11465889, PMID 12925520, PMID 29036413, PMID 32039167, PMID 32483976, PMID 33670819, PMID 35157031, PMID 37770709

L1 网络状态:本主题尚未进入 L1 对应网络(EDGE_TABLE 对形状/Wasserstein/样本量/子采样零命中;CORRESPONDENCE_NETWORK 仅有 MSE 曲线伪影、PRQ 同步区形状等相邻方法学背景边,无本主题裁决边),证据等级按规则封顶 C。


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。

Pseudo-shape component identification and separation point localization: residual shape proportion decay curve with sample size, beyond what sample size does the shape proportio…

Evidence level: C

When conducting cross-ethnic comparisons of triad distributions, even if the shape component is genuinely non-zero, one must first ask: how much of this reflects true shape differences between ethnic groups, and how much is merely random fluctuation inherent in small-sample empirical distributions.

Body

The previous round adjudicated that the inclusion of Wasserstein shape terms must be tested within batches of paired data. However, before that, a more fundamental issue remains: whether the shape term itself is trustworthy. Empirical distributions are merely sampling snapshots of the population; with small samples, their skewness, kurtosis, and multimodal structures inherently carry random fluctuations. This noise blends directly into the shape term, creating “pseudo-shapes”—differences that appear to be variations in distribution shape between ethnic groups but are actually just insufficient sample sizes on each side, leaving the empirical distributions unstable. While location terms and discrete terms have large-sample theories to support them, shape structure is most sensitive to sample size; thus, this problem is sharpest for the shape term.

The good news is that pseudo-shapes and true shapes possess a separable mathematical property: estimation noise monotonically decreases with sample size, whereas true shape differences do not. As sample size increases, the empirical distribution gradually approaches the population distribution, causing the proportion of pseudo-shape to decline continuously; conversely, the true shape differences between ethnic groups remain constant regardless of sample size. Consequently, one can plot a “proportion of shape term versus sample size” decay curve: the descending segment represents estimation noise, and the point where the curve flattens out corresponds to the true shape. The transition point from decline to plateau is the separation point—sample sizes beyond this point yield shape term readings that are trustworthy.

This approach has established precedents in physiological measurement: the stability of estimators depends on data volume, a measurable and repeatedly tested quantity. For instance, short-term reproducibility of heart rate and QT variability in myocardial infarction patients systematically changes with recording duration (PMID 27681306); the repeatability of HRV indices is constrained by measurement window length (PMID 15286929); empirical measurements of short-term reproducibility for blood pressure and heart rate variability (PMID 7981010); consistency boundaries for ultra-short-term versus standard HRV analysis (PMID 39063591); and reliability of transfer function estimation in cardiovascular variability analysis also varies with data volume (PMID 11465889). Normalization follows similar principles: classical quantile normalization assumes “identical distributions across groups” (PMID 12925520), requiring smoothing when group distributions genuinely differ (PMID 29036413); normalization choices for unbalanced transcriptomes constitute a review-level challenge (PMID 32039167); and quantile regression, when establishing continuous norms, transmits estimation bias and precision directly into downstream readings (PMID 32483976), a fact confirmed by comparative studies on quantile mapping bias correction (PMID 33670819). Optimal transport methods face similar issues when performing empirical measure comparisons in single-cell genomics (PMID 35157031) or learning single-cell perturbation responses (PMID 37770709), where estimation noise contaminating comparison results is a known problem. This topic lacks not evidence of “phenomenon existence,” but rather the single empirical measurement that quantitatively draws this decay curve in a multi-ethnic HRV context.

The design proceeds in two steps. First, synthetic data calibration: construct populations with known shape differences and repeatedly sample at gradients of n=20/50/100/200/500 to obtain template curves of proportion decay under pure estimation noise. Second, real-world multi-ethnic HRV subsampling: recalculate the shape term proportion at each of the five gradients and compare against the template—the sample size level where the empirical curve diverges from the pure noise template and begins to plateau marks the separation point. Three scenarios for adjudication are pre-registered: if proportion continues declining at n=500, the shape term is overall untrustworthy at this scale, freezing ethnic shape conclusions; if the plateau occurs at n≤100, the shape term is reliable; if a dual-plateau structure emerges (decline followed by plateau then decline), reporting must be stratified by metric.

Three implementation details are locked in advance to prevent adjudication contamination. First, estimation noise within the shape proportion stems from two sources: sampling noise from intra-ethnic subsampling and processing noise introduced by the subsampling and normalization pipeline. The synthetic template must cover both noise sources; otherwise, the descending segment of the empirical curve may be misinterpreted as pure sampling effects. Second, subsampling at each sample size level must repeat sufficiently (pre-registered 200 times) with confidence bands reported; the width of these bands serves as an auxiliary criterion: only levels where bandwidth convergence occurs represent true separation points, whereas relying solely on point estimates risks mistaking jitter for plateaus. Third, subsampling must preserve paired structure—same equipment batch, same recording window, same covariate adjustment—adhering to the “same-batch pairing” adjudication criteria from the previous round, thereby avoiding misattributing pipeline differences to ethnic shape variations.

With this step completed, the shape term adjudications from earlier rounds truly stand: all ethnic shape differences within the separation point must be treated as pseudo-shapes and excluded from conclusions.

Evidence and Limitations

Verified: Estimator stability relies on sample size and measurement window length, with direct empirical literature in HRV, blood pressure variability, and transfer function estimation (PMID 27681306, 15286929, 7981010, 11465889); distributional assumptions of normalization methods and error propagation have systematic study (PMID 12925520, 29036413, 32039167, 32483976).

Partially Verified: Optimal transport methods are affected by estimation noise in empirical measure comparisons, confirmed by methodological literature (PMID 35157031, 37770709), but no direct empirical data exists for the “proportion decay curve with respect to sample size” of the one-dimensional Wasserstein shape component.

Pending Verification: The decay curve of pseudo-shape proportion and the separation point under multi-ethnic HRV subsampling; the freeze/unfreeze rules for ethnic shape conclusions corresponding to the three decision cases—all are design proposals of this study and lack direct empirical measurement.

Related Reading

Evidence Level: C
References: PMID 27681306, PMID 15286929, PMID 7981010, PMID 39063591, PMID 11465889, PMID 12925520, PMID 29036413, PMID 32039167, PMID 32483976, PMID 33670819, PMID 35157031, PMID 37770709

L1 Network Status: This topic has not yet entered the L1 corresponding network (EDGE_TABLE shows zero hits for shape/Wasserstein/sample size/subsampling; CORRESPONDENCE_NETWORK contains only adjacent methodological background edges such as MSE curve artifacts, PRQ synchronous region shape, etc., but lacks the verdict edge for this topic). Consequently, the evidence level is capped at C per the rules.


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 *