5 个百分点和 15 个百分点是从哪儿来的:份额差判据的标定曲线

Written by

in

Language

Evidence level: C

“差值小于 5 个百分点结论保留,5 到 15 个百分点双报,超过 15 个百分点作废”——这三条线本身有没有出处?如果没有,同一批数据落在哪一格就取决于阈值而不是共线强度。

正文

一、先问一个不讨喜的问题:这三条线是谁定的

上一篇文章给”独立 vs 相关”两种设定下的谐波方差份额立了三格判据:差值 <5 个百分点算共线弱、结论保留;5–15 个百分点必须双设定并列报告;>15 个百分点一律作废重算。数字看起来很精确,但追下去会发现一个尴尬的事实:5、15 这两个数没有出处——它们是从方法学写作习惯里取的整十数(5 的倍数),而不是从任何实测或仿真标定出来的。

这在方法学上是硬伤。阈值不标定,判据的格子边界就是任意的:同一批数据落在哪一格,取决于阈值定在哪,而不取决于共线强度到底多强。两个研究各自取了 4/12 和 6/14 两组线,同一份结果会被一份判成”保留”、另一份判成”作废”。数字越精确,越不该是拍出来的。

二、阈值该对准什么:结论翻转率,而不是一个好看的百分点

真正该问的问题是:份额差达到多大,才足以让一个结论大概率被翻过来?阈值应当定在”结论翻转概率可接受”的那个点上,而不是定在一个好看的整数上。

这个思路在生物力学领域有成熟示范。全局敏感性分析(GSA)基于高斯过程元建模的那条线,明确要求”透明声明估计过程中引入的各类不确定项”(PMID 36546844)——因为元建模本身会引入额外不确定性,不声明,敏感度指数的置信就无从谈起。这恰好就是”份额数字必须带不确定区间”的行业先例。

更关键的一条来自动脉壁力学模型的不确定性量化研究:模型不确定性足以掩盖真实的组间效应,而增加测量重复次数可以降低它(PMID 28755237)。这给了阈值一个物理含义——阈值不是统计惯例,而是”把结论翻转率压到可接受水平所需的份额差”,这个份额差反过来可以由测量重复次数来兑换。

三、为什么这个问题长期没人处理:主流做法根本不检验

阈值标定迟迟没人做的原因,藏在敏感性分析的实践调查里:过去十年绝大多数已发表的敏感性分析仍是单因子轮换(OAT),并建立在未经检验的线性与可加性假设上,全局方法只有少数人在用(PMID 26934843)。

也就是说,”份额数字”在领域里长期被当成一个可以直接比较的数——大家默认输入独立、默认线性可加,于是从未被要求回答”这个数字在多宽的噪声带上晃动”。没有噪声带的数字,自然也就没有”多大差值才算出圈”的标定需求。问题不是技术做不到,是惯例里根本不存在这个追问。

四、怎么标定:纯仿真,不必等真人数据

好消息是,这条标定曲线可以纯仿真拉出来,不需要新增受试者:

  1. 构造已知真值的份额:用一维血流模型合成谱数据,人为设定真实的”几何 / 弹性 / 边界”份额比,如 50/30/20、40/20/40、25/25/50 三档;
  2. 铺开共线强度与样本量:按四档相关强度(0.2/0.4/0.6/0.8)与三档样本量(30/60/150),每格生成 ≥500 次重复;
  3. 记录两个”翻车”事件:每次按独立设定算份额,与真值比,记录”份额排序被弄反”与”最大项被认错”两个事件的发生率;
  4. 反解阈值:以发生率为纵轴、份额差为横轴画结论翻转概率曲线,反解翻转率 10% 与 25% 对应的份额差。

五、三种可能的结局,三种不同的处方

若 10% 翻转率对应的份额差落在 4–7 个百分点(即 5 这条线大体靠谱),只需按实证值微调数字、判据结构不动;

若该值 <3 个百分点(即原 5 这条线过于宽松、共线稍强就会翻转结论),则”<5pp 保留”这一格必须收紧,且既往所有单设定份额结论的可靠性须整体下调——过去按”差值小”留下的结论里,有一批其实早该作废;

若该值 >10 个百分点(即原线过严、把可用的结论误判为不可用),则须放宽,并把”作废”改为”双报”——代价是把更多含混结论放进”需并列报告”的池子。

六、与母文和上一篇的关系

母文立了”在什么相关性结构下占几成”这个问法,并给出 5/15 两线的暂定值;上一篇(n=60 相关矩阵估计不稳性)指出这条判据在现样本量下可能只是在读噪声。本文补上最后一环:即便样本量够、相关矩阵稳,5/15 这两条线本身的刻度也没标定过。三篇合起来是一个完整的”判据审查”序列——问法、噪声、刻度。

注意证据薄弱:本文是纯方法学推演,没有真实数据支撑 5/15 的具体刻度,标定曲线本身也尚未跑出结果;以下结论只能当”阈值必须有出处”这一操作规范的论证,不能当”5 和 15 是错的对”的答案。

L1 网络状态:该主题(谐波方差份额的输入相关性 / 共线强度 / 阈值标定)尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK 边表止于 E-050「六十甲子纳音 | 对应 | 60 个频率」(open,现代证据栏原文「尚无任何可测来源」),节点表无「方差份额 / 参数相关性 / 归因权重」被测量本身,全文件 grep「谐波 / 方差 / 份额 / 共线 / 标定」零命中;EDGE_TABLE_science-yixue 仅列概念计数(脉象 73 篇 / A级5),无「阈值 / 标定 / 翻转率」条目。已登记入 L1 gaps 表(「谐波方差份额 / 输入相关性 / 共线强度」簇第 3 条)。

证据与局限

已验证:生物力学 GSA 要求透明声明估计中的不确定项(PMID 36546844);模型不确定性可掩盖真实效应、增加测量重复可降低之(PMID 28755237);主流敏感性分析仍以 OAT 与线性可加假设为主(PMID 26934843);软组织生物力学研究存在明确的样本量考量框架(PMID 37517620)。

部分验证:阈值应定在”结论翻转率可接受点”这一原则有方法学依据,但尚无研究把该原则落地到”份额差 → 翻转率”的具体曲线上。

待验证:10% 与 25% 翻转率各自对应的份额差到底是 3、5、7 还是 10 个百分点;翻转率曲线在低相关(0.2–0.4)与高相关(0.6–0.8)区间的形状是否一致;”排序弄反”与”最大项认错”两个事件是否随阶次(C1–C10)单调变化。

相关阅读

证据等级: C
参考文献: PMID 36546844, PMID 28755237, PMID 26934843, PMID 37517620


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。

5 percentage points and 15 percentage points: calibration curve of the share difference criterion

Evidence level: C

“If the difference is less than 5 percent points, retain the conclusion; if between 5 and 15 percent points, report both settings; if exceeding 15 percent points, discard and recalculate.” — Do these three thresholds have an origin? If not, which bin a given dataset falls into depends on the threshold rather than the collinearity strength.

Main Text

I. First, Ask an Unpleasant Question: Who Defined These Three Lines?

The previous article established three criteria for the harmonic variance share under two settings: “independent vs. correlated.” Specifically, a difference <5 percent points indicates weak collinearity and the conclusion is retained; 5–15 percent points requires reporting both settings simultaneously; >15 percent points means the result is discarded and must be recalculated. These numbers appear precise, yet tracing them reveals an awkward fact: the numbers 5 and 15 have no origin. They are simply round numbers (multiples of 5) drawn from methodological writing conventions, not values derived from any empirical measurement or simulation calibration.

This is a hard flaw in the methodology. Without calibrated thresholds, the boundaries of the decision bins are arbitrary: which bin a given dataset falls into depends on where the threshold is set, not on how strong the collinearity actually is. Two studies might adopt different lines (e.g., 4/12 and 6/14); the same result could be judged “retain” in one and “discard” in the other. The more precise the numbers, the less they should be guessed.

II. What Should the Threshold Target: Conclusion Flip Rate, Not a Pretty Percentage Point?

The real question to ask is: at what share difference does it become probable enough for a conclusion to flip? The threshold should be set at the point where “the probability of conclusion flipping is acceptable,” not at a visually appealing integer.

This approach has established precedents in the biomechanics field. Global Sensitivity Analysis (GSA) based on Gaussian process metamodeling explicitly requires “transparent declaration of all uncertainty terms introduced during the estimation process” (PMID 36546844)—because metamodeling itself introduces additional uncertainty; without declaration, the confidence of sensitivity indices cannot be established. This is exactly the industry precedent for “share numbers must carry uncertainty intervals.”

Even more critical is the uncertainty quantification study of carotid wall mechanics models: model uncertainty can mask true inter-group effects, and increasing the number of measurement repetitions reduces it (PMID 28755237). This gives the threshold a physical meaning—the threshold is not merely a statistical convention, but “the share difference required to suppress the conclusion flip rate to an acceptable level,” which can in turn be traded for measurement repetition counts.

III. Why This Question Has Long Been Ignored: Mainstream Practices Do Not Test

The reason threshold calibration has long gone unaddressed lies in the practice of sensitivity analysis: over the past decade, the vast majority of published sensitivity analyses have relied on One-At-A-Time (OAT) single-factor rotation, built upon untested assumptions of linearity and additivity, with only a few researchers using global methods (PMID 26934843).

In other words, “share numbers” have long been treated in the field as quantities directly comparable—everyone assumes independent inputs and linear additivity, so there has never been a requirement to answer “how much does this number fluctuate within a noise band?” Without numbers in a noise band, there is naturally no need for calibration of “what difference counts as significant.” The problem is not technical impossibility, but the convention itself does not include this inquiry.

IV. How to Calibrate: Pure Simulation, No Need for New Human Data

The good news is that this calibration curve can be pulled out via pure simulation, without needing to recruit new subjects:

  1. Construct shares with known true values: Synthesize spectral data using a one-dimensional blood flow model, artificially setting true “geometry/elasticity/boundary” share ratios such as 50/30/20, 40/20/40, and 25/25/50;
  2. Spread collinearity strength and sample size: According to four levels of correlation strength (0.2/0.4/0.6/0.8) and three levels of sample size (30/60/150), generate ≥500 repetitions for each cell;
  3. Record two “failure” events: For each run, calculate shares under the independent setting and compare with true values, recording the incidence rates of “share ranking reversed” and “maximum term misidentified”;
  4. Invert the thresholds: Plot the conclusion flip probability curve with incidence rate on the vertical axis and share difference on the horizontal axis, then invert to find the share differences corresponding to flip rates of 10% and 25%.

V. Three Possible Outcomes, Three Different Prescriptions

If the share difference corresponding to a 10% flip rate falls between 4–7 percent points (meaning the line at 5 is roughly reliable), simply fine-tune the numbers based on empirical values while keeping the decision structure unchanged;

If this value is <3 percent points (meaning the original 5-line is too loose and even slight collinearity flips the conclusion), then the “<5pp retain” bin must be tightened, and the reliability of all past single-setting share conclusions must be globally downgraded—among the conclusions previously retained due to “small difference,” a batch should actually have been discarded;

If this value is >10 percent points (meaning the original line is too strict and usable conclusions are misjudged as unusable), then it must be relaxed, and “discard” should be changed to “report both”—the cost is putting more ambiguous conclusions into the pool requiring simultaneous reporting.

VI. Relationship to the Parent Paper and Previous Article

The parent paper established the question “what proportion under what correlation structure” and provided tentative values of 5/15; the previous article (n=60 correlation matrix estimation instability) pointed out that this criterion might merely be reading noise under current sample sizes. This article completes the final link: even if sample size is sufficient and the correlation matrix is stable, the scales of the 5/15 lines themselves have never been calibrated. Together, these three papers form a complete “criterion review” sequence—questioning the formulation, analyzing noise, and calibrating scales.

Note on weak evidence: This article is pure methodological deduction; it does not provide real data supporting the specific scales of 5/15, nor has the calibration curve itself been run yet; the following conclusions can only serve as argumentation for the operational norm that “thresholds must have an origin,” not as answers to whether “5 and 15 are wrong or right.”

L1 Network Status: This topic (harmonic variance share input correlation / collinearity strength / threshold calibration) has not yet entered the L1 corresponding network—the CORRESPONDENCE_NETWORK edge table stops at E-050 “Sixty Jiazi Na Yin | Correspondence | 60 frequencies” (open, original text in modern evidence column states “no measurable source yet”), and the node table lacks “variance share / parameter correlation / attribution weight” being measured itself; full file grep for “harmonic / variance / share / collinearity / calibration” yields zero hits; EDGE_TABLE_science-yixue only lists concept counts (73 pulse diagnosis articles / A-level 5), with no entries for “threshold / calibration / flip rate.” Already registered in the L1 gaps table (cluster #3: harmonic variance share / input correlation / collinearity strength).

Evidence and Limitations

Verified: Biomechanics GSA requires transparent declaration of uncertainty terms in estimation (PMID 36546844); model uncertainty can mask true effects, and increasing measurement repetitions reduces it (PMID 28755237); mainstream sensitivity analysis still relies primarily on OAT and linear additivity assumptions (PMID 26934843); soft tissue biomechanics research has clear sample size consideration frameworks (PMID 37517620).

Partially Verified: The principle that thresholds should be set at “points where conclusion flip rates are acceptable” has methodological basis, but no study has yet grounded this principle onto the specific curve of “share difference → flip rate.”

To Be Verified: Whether the share differences corresponding to 10% and 25% flip rates are 3, 5, 7, or 10 percent points; whether the shape of the flip rate curve is consistent across low correlation (0.2–0.4) and high correlation (0.6–0.8) intervals; whether the two events—”ranking reversed” and “maximum term misidentified”—change monotonically with order (C1–C10).

Further Reading

Evidence Level: C
References: PMID 36546844, PMID 28755237, PMID 26934843, PMID 37517620


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 *