Evidence level: C
相关矩阵本身是估出来的,估不准的时候,「独立 vs 相关」两个设定的份额差里混进来的,可能不是共线强度,而是估计噪声。
正文
上一篇文章(#1178)把「谐波方差份额」的整条判据,压在一个前提上:要么按独立输入算份额,要么按「实测相关结构」构造相关输入再算。两套数字一对比,差值落在哪个区间,就决定旧结论是保留、双报还是作废。这条判据看起来无懈可击,但它默认了一件事:「相关结构」是一个已知的、固定的设定。可惜它不是——相关矩阵是从样本里估出来的,而样本只有 n=60。这一篇要问的问题很朴素:同一批人重抽一次,相关矩阵会变多少?变出来的这部分,会不会正好把判据的 5 个百分点阈值吃掉?
先看相关估计在 n=60 时到底有多不稳。统计学里这是有硬数字的老问题:在小样本下,样本相关系数是总体相关的有偏估计,偏到什么程度?贝叶斯框架下的模拟显示,用均匀先验与 arc-sine 先验估计小样本相关,明显优于经验估计量与极大似然估计量(PMID 23378667)——也就是说,连「怎么估」这个环节都还没定论,默认的 Pearson 估计量在小样本下本身就带系统性偏差。更直观的量级证据来自一篇 2024 年的脑影像可重复性研究:真实相关系数 r=0.63 的两个变量,在 N=25 的小样本里,观测到的相关系数可以从 0.25 一路抖到 0.95(PMID 39215207)。n=60 虽然比 25 好,但绝没有好到「估计稳」的程度——这篇文献用 1000 次 bootstrap 画出「抽样变异性 vs 样本量」曲线,曲线在几百人的量级才明显收窄。把「抖 0.3~0.7」这个幅度换算到本文的场景:三个变量两两相关只有 3 个系数,任何一个系数在 0.3 与 0.7 之间跳动,就足以让「相关输入」这一设定在两次抽样里给出完全不同的联合分布。
那「相关结构变了」会怎样传导到份额差上?GSA 文献对这一点同样有明确结论:GSA 结果强受相关输入结构影响,而现有考虑相关性的方法大多计算昂贵、难以实施(PMID 35095342)。更细的机制来自两篇药代动力学方法学:潜变量方法把两个相关输入拆成一个公共潜变量加两个独立残差,发现「是否考虑相关」直接改变每个参数的敏感性指标解读(PMID 34032996);两阶段 δ 指标法则更进一步,先在无相关假设下识别因果效应参数,再引入真实相关结构考察「间接效应」,结论是同一批参数在两步里的排名可以不同(PMID 37422844)。翻译成脉象的话:「几何占几成、弹性占几成」这个数字,在相关矩阵抖一下之后,可能就把两个器官的排名弄反。而 #1178 的判据恰恰用「份额差」来判定旧结论命运——若差值本身含有一个与共线强度同量级的噪声项,那么「差值 <5pp 判共线弱」这条规则在 n=60 下,很可能只是在读噪声。
这就是本文的核心主张,也是与母文 #1178 的差别所在:#1178 讨论的是「相关结构该不该被考虑」,本文讨论的是「相关结构本身能不能被估准」。前者是建模选择问题,后者是估计问题——后者解决不了,前者就是空中楼阁。系统生物学里早已示范过完整的 GSA 方法论(PMID 18572196),但它的标准流程里,相关矩阵是作为「已知输入」给定的;而当我们没有 150 人、只有 60 人时,这个「已知」就不成立了。好在补救路径是现成的:bootstrap 重抽样可以直接给每个相关系数一个抽样分布与置信区间(PMID 21973092),代价只是计算量;而把相关性当作后验分布去学、而非当作预设点值的贝叶斯校准路线(PMID 41838759),更是把「相关结构不确定」这件事本身纳入模型——这两条路都是可执行的。
证据与局限
已验证: 小样本下样本相关是有偏估计、贝叶斯先验可改进(PMID 23378667);GSA 结果强受相关输入影响、相关方法大多昂贵难实施(PMID 35095342);「是否考虑相关」直接改变敏感性指标的解读与参数排名(PMID 34032996、PMID 37422844);N=25 时真实 r=0.63 可观测为 0.25~0.95(PMID 39215207);bootstrap 可给出相关估计的抽样分布(PMID 21973092)。
部分验证: 「n=60 时相关估计噪声接近 5pp 判据阈值」——量级推断成立(由 PMID 39215207 的变异性曲线外推),但尚无人在脉象的几何-弹性-边界三参数上直接测过这个噪声到底多大。
待验证: ① n=60 与 n=150 两样本量下,三参数相关矩阵的 bootstrap 95% 区间宽度是多少;② 两设定份额差的抽样分布,其半宽是否超过 5pp;③ 相关强度人为设为 0.2/0.4/0.6/0.8 四档时,区间宽度如何随真实相关变化;④ 让 5pp 阈值可靠区分共线强弱所需的最小样本量。
相关阅读
证据等级: C
参考文献: PMID 39215207, PMID 23378667, PMID 34032996, PMID 37422844, PMID 18572196, PMID 35095342, PMID 41838759L1 网络状态:该主题对应边「谐波方差份额 / 输入相关性 / 共线强度 ↔ 测量学与外效度维度(相关结构估计)」在 CORRESPONDENCE_NETWORK 中整簇缺失——节点表无「方差份额 / 参数相关性 / 归因权重」被测量本身,物理侧仅「相位 / 节律」一行,边表止于 E-050 且其现代证据栏原文「尚无任何可测来源」;本文为「测量学与外效度」簇的深化层(从「相关结构是假设产物」深化到「相关结构本身估计不稳」),属方法学推进,非已验证机制。
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。
Re-estimation of the same sample yields different results: Instability of the geometric-elastic-boundary correlation matrix estimation at n=60
Evidence level: C
The correlation matrix itself is estimated; when estimates are imprecise, the discrepancy in the proportion of “independent vs. correlated” settings may not reflect collinearity strength but rather estimation noise.
Body
The previous article (#1178) compressed the entire criterion for “harmonic variance share” onto a single premise: either calculate shares based on independent inputs or construct correlated inputs using the “measured correlated structure” before calculation. Comparing these two sets of numbers, the interval in which the difference falls determines whether the old conclusion is retained, reported as dual, or discarded. This criterion appears airtight, yet it presupposes that the “correlated structure” is a known, fixed setting. Unfortunately, this is not the case—the correlation matrix is estimated from the sample, and the sample size is only n=60.
This article addresses a straightforward question: if the same group of subjects is re-sampled, how much does the correlation matrix change? Will this variation consume the 5-percentage-point threshold of the criterion?
First, consider how unstable correlation estimates are at n=60. In statistics, this is a well-documented issue with hard numbers: under small samples, the sample correlation coefficient is a biased estimator of the population correlation. How biased? Bayesian simulations indicate that using uniform priors and arc-sine priors to estimate small-sample correlations performs significantly better than empirical estimators and maximum likelihood estimators (PMID 23378667)—meaning even the method of estimation remains unresolved, and the default Pearson estimator inherently carries systematic bias in small samples.
More intuitive evidence comes from a 2024 neuroimaging reproducibility study: for two variables with a true correlation coefficient of r=0.63, the observed correlation in a small sample of N=25 could fluctuate anywhere from 0.25 to 0.95 (PMID 39215207). While n=60 is better than 25, it is far from stable enough for reliable estimation. That study used 1,000 bootstrap resamples to plot a “sampling variability vs. sample size” curve, which only noticeably narrows at sample sizes in the hundreds. Translating this fluctuation range of ±0.3 to ±0.7 into our scenario: with three variables, there are only three pairwise correlation coefficients. If any single coefficient fluctuates between 0.3 and 0.7 across two resamples, it is sufficient to produce entirely different joint distributions under the “correlated inputs” assumption.
How does this shifting “correlated structure” propagate to the difference in shares? The GSA literature provides clear conclusions on this point: GSA results are strongly influenced by the structure of correlated inputs, and existing methods accounting for correlations are mostly computationally expensive and difficult to implement (PMID 35095342). A finer mechanism emerges from two pharmacokinetic methodologies: latent variable methods decompose two correlated inputs into one common latent variable plus two independent residuals, finding that “whether correlation is considered” directly alters the interpretation of sensitivity indices for each parameter (PMID 34032996); the two-stage δ-index rule further identifies causal effect parameters under no-correlation assumptions before introducing the true correlated structure to examine “indirect effects,” concluding that the ranking of the same parameters can differ between steps (PMID 37422844).
Translating this into our terminology: “what proportion is geometric and what proportion is elastic” can flip the ranking of two organs once the correlation matrix fluctuates. Since #1178 uses this “share difference” to determine the fate of old conclusions, if the difference itself contains a noise term comparable in magnitude to collinearity strength, then the rule “difference <5pp implies weak collinearity” at n=60 may simply be reading noise.
This is the core thesis of this article and its distinction from the parent text #1178: #1178 discusses whether the correlated structure should be considered; this article asks whether the correlated structure itself can be estimated accurately. The former is a modeling choice issue; the latter is an estimation problem—if the latter cannot be solved, the former is built on air. System biology has already demonstrated a complete GSA methodology (PMID 18572196), but its standard workflow treats the correlation matrix as “known input.” When we lack 150 subjects and have only 60, this “known” assumption fails. Fortunately, remediation paths are available: bootstrap resampling can directly provide sampling distributions and confidence intervals for each correlation coefficient (PMID 21973092), at the cost of computation; alternatively, treating correlation as a posterior distribution rather than preset point values via Bayesian calibration (PMID 41838759) incorporates the uncertainty of the “correlated structure” itself into the model—both approaches are feasible.
Evidence and Limitations
Verified: Small-sample correlations are biased estimates; Bayesian priors improve estimation (PMID 23378667); GSA results are strongly affected by correlated input structure, and most correlation-aware methods are computationally expensive (PMID 35095342); considering correlation directly alters sensitivity index interpretation and parameter ranking (PMID 34032996, PMID 37422844); at N=25, true r=0.63 can be observed as 0.25–0.95 (PMID 39215207); bootstrap provides sampling distributions for correlation estimates (PMID 21973092).
Partially Verified: “At n=60, correlation estimate noise approaches the 5pp criterion threshold”—the magnitude inference holds (extrapolated from the variability curve in PMID 39215207), but no one has directly measured this noise level on the geometric-elasticity-boundary three-parameter structure.
To Be Verified: ① What is the width of the bootstrap 95% interval for the three-parameter correlation matrix at n=60 vs. n=150; ② Whether the half-width of the sampling distribution of the share difference exceeds 5pp; ③ How interval width varies with true correlation when artificial correlation strengths are set at 0.2/0.4/0.6/0.8; ④ The minimum sample size required for the 5pp threshold to reliably distinguish strong from weak collinearity.
Related Reading
- Changing the order of ingredients changes the conclusion: bias in harmonic variance share under independent vs. correlated assumptions
- Are residuals “fixable by tuning” or “due to wrong model form”?
- Research pool: Biological frequency code
Evidence Level: C
References: PMID 39215207, PMID 23378667, PMID 34032996, PMID 37422844, PMID 18572196, PMID 35095342, PMID 41838759L1 Network Status: This topic corresponds to the edge “harmonic variance share / input correlation / collinearity strength ↔ measurement and external validity dimensions (correlation structure estimation)” which is entirely missing in the CORRESPONDENCE_NETWORK—no nodes measure variance share / parameter correlation / attribution weights themselves; the physical side only includes “phase / rhythm,” and the edge table stops at E-050 with its modern evidence column stating “no measurable source yet.” This article deepens the “measurement and external validity” cluster (from “correlation structure is a hypothesis product” to “correlation structure itself is unstable in estimation”), representing methodological advancement rather than a verified mechanism.
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