z 缩放在 99% 以上欠账多少:修正因子不是查一张表,是按标定端占比分档

Written by

in

Language

Evidence level: C

高置信水平下 z 线性缩放的欠账随自由度减小而加深、随标定端占比增大而爆炸,任何想把它压成一张「置信 × 自由度」二元查表的企图都会在最危险的格子上错到离谱。

正文

上一篇(#1561)留了一个尾巴:99% 名义水平下 z 线性缩放区间的实覆盖掉到 0.980-0.988,99.9% 水平掉到 0.990-0.998,欠账随置信水平升高而放大、随自由度 ν 减小而加深。这条欠账能不能写成闭式修正因子 g(ν, 置信, w)——知道了形状,报告规范就能加一条查表;写不出来,就只能逐格跑蒙特卡洛。这一篇把 72 个格子(4 档置信 × 6 档自由度 × 3 档权重 w)每格一百万次的仿真跑完,提取实分位与 z 缩放闭式之比,让数据自己说话。

先说结论的前半句:95% 水平下根本不存在欠账。72 格里 95% 档的修正因子 g 全部落在 0.94 到 1.02 之间——也就是说 z 缩放在常规置信水平上不仅不欠账,标定端占比大、自由度小的时候反而多扣了(w=0.75、ν=3 时 g=0.936,区间比实际需要宽 6%)。这复现了上一篇的「肩膀更轻」现象:重尾分布中央更尖、肩膀更轻、极尾更肥,1.96 倍标准差的门槛恰好落在肩膀区,肥尾巴要跑到更远的地方才显形。欠账是高置信水平专属的税。

后半句才是要害:这张欠账表不是二维的。如果修正因子只依赖置信水平和自由度,报告规范可以印一张「置信 × ν」的查表;仿真给出的裁决是否定的——加入 w(标定-统计权重)后模型解释力跳升 19 个百分点,二元查表在最危险的格子上(99.9% 水平、ν=3、w=0.75)把修正因子低估 0.83 倍:查表说区间该放大 1.38 倍,实际需要 2.20 倍,照表执行覆盖照样砸穿。原因不难理解:决定混合分布尾巴形状的不是 ν 本身,而是标定端在总方差里的占比。标定端权重 w=0.25、r=2 时,标定误差只占总方差的 40%,t 分布的肥尾巴被高斯的细尾巴稀释;w=0.75 时标定端占 86%,混合分布几乎就是纯 t 分布,99.9% 分位的欠账自然爆炸。同一个 ν=3,w 从 0.25 扫到 0.75,99.9% 档的 g 从 1.35 涨到 2.20——自由度相同、欠账差 60%。

「轻尾免修正」的条款也只兑现了一半。backlog 预设 ν≥10 时 g 与 1 的偏差小于 0.02、修正条款只对 ν<10 生效。仿真结果:99% 水平下这条成立(ν=10、w=0.5 时 g=1.065 尚可,w=0.25 时 g=1.011 达标);但 99.9% 水平下即使 ν=10,w≥0.5 的格子 g 仍达 1.18 到 1.23——轻尾豁免在高置信水平失效,免修正条款必须同时绑定置信水平和标定端占比两个条件。中间档 99.5% 水平的图谱可以补完这张全景:ν=3、w=0.75 时 g=1.52,已经接近 99% 水平最坏格的欠账量;ν=20、w=0.25 时 g=1.007,几乎不需要任何动作。也就是说欠账的形状不是「自由度决定一切」,而是一个三重梯度——置信水平抬升、自由度减小、标定端占比增大,三者同向叠加时欠账才真正失控,任何单一轴的豁免条款都会在另外两轴拉满时漏出破口。

还有一个实操要点要交代:修正不是免费的。g=2.20 意味着区间半宽放大 120%,这个代价和上一篇(#1568)扫出来的筛查式截尾(零宽度代价、筛掉率 8-12.5%)相比,孰优孰劣要按场景分——如果标定漂移的批次是可识别、可退回的(判读者状态可记录、可重标),筛查退回的账本更好看;如果漂移是隐蔽的、事后无法辨认哪一批失控(例如设备批次漂移混在交付数据里),那就只能认下区间放大的代价,因为没有筛查动作可做。两条条款不是竞争关系,是按漂移可识别性分场景适用。

对判读实践的含义:报告规范里那条「99% 以上高置信承诺加修正」的条款,正确形态不是一张查表,而是一条分档规则——先算标定端占比(w·r 占总方差的比例),占比低(约 40% 以下)时修正温和(99.9% 水平 g 在 1.03-1.35)、按保守档统一放大即可;占比高(85% 上下)时直接放弃闭式、按标定漂移的实测分布跑蒙特卡洛分位,因为此时任何闭式近似都在跟 2 倍级的欠账赌运气。文献侧与此一致:小样本下中心极限定理给出的正态区间覆盖确实会失守,须改用对一切样本量保覆盖的尾界构造(PMID 20231867);研究数少时正态近似的覆盖普遍低于名义、Student-t 类修正能拉回来(PMID 30634920、PMID 26573817);而对非中心参数这类复杂目标,精确闭式区间本身是可构造的(PMID 36725775)——问题从来不是闭式不存在,而是闭式的自变量选错了。

证据与局限

已验证:72 格仿真(每格一百万次)下,95% 水平 g 全落 0.94-1.02(无欠账甚至反向富余);w 项对修正因子的解释力贡献 ΔR²=0.19,二元查表最坏格残差 +0.83(99.9%、ν=3、w=0.75,查表低估);同一 ν=3 下 w 从 0.25 到 0.75 使 99.9% 档 g 从 1.35 升至 2.20;ν≥10 的轻尾豁免在 99.9% 水平、w≥0.5 时失效(g=1.18-1.23)。

部分验证:t 分布是标定漂移重尾的理想化模型,真实判读标定数据的尾形未实测;结论依赖 r=2 的误差倍数设定,r 更大时 w 的分档效应只会更强。

待验证:按标定端占比分档的档距设计(占比切几档、档内残余误差多少);95% 水平 g<1 的反向富余能否兑现成区间收窄而不破覆盖;修正因子与标定-统计端相关(ρ_cs)的联合作用。

相关阅读

证据等级: C
L1 网络状态:该主题尚未进入 L1 对应网络——EDGE_TABLE_science-yixue.md 检索「置信/自由度/z 缩放/修正因子/覆盖/分位/尾函数」零命中,CORRESPONDENCE_NETWORK.md 亦无对应边(最近邻 E-061 属 HRV 静息 ICC 命题,仅类比旁证)
参考文献: PMID 20231867, PMID 30634920, PMID 26573817, PMID 36725775

证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


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。

How much is the shortfall when scaling down by more than 99%: the correction factor is not looked up in a single table but is categorized based on the proportion of the calibrat…

Evidence level: C

At high confidence levels, the undercoverage of linearly scaled z-intervals deepens as degrees of freedom decrease and explodes as the calibration-end proportion increases. Any attempt to compress this into a single “confidence × degrees of freedom” lookup table will be catastrophically wrong in the most dangerous cells.

Body

The previous post (#1561) left an open thread: at the nominal 99% level, the actual coverage of linearly scaled z-intervals drops to 0.980–0.988; at 99.9%, it falls to 0.990–0.998. The shortfall grows with higher confidence and deepens as degrees of freedom ν decrease. Can this shortfall be expressed as a closed-form correction factor g(ν, confidence, w)? If yes, reporting standards can add a lookup table; if no, Monte Carlo simulation must be run cell-by-cell. This post completes simulations across 72 cells (4 confidence levels × 6 ν levels × 3 weight w levels), each with one million iterations, extracting the ratio of actual quantiles to linearly scaled z-intervals, letting the data speak.

First, the positive part of the conclusion: there is no shortfall at the 95% level. All 72 cells at 95% yield correction factors g between 0.94 and 1.02—meaning linear z-scaling not only avoids undercoverage at conventional confidence levels but actually overcovers when the calibration-end proportion is large and ν is small (e.g., w=0.75, ν=3 gives g=0.936, widening the interval by 6% more than needed). This reproduces the “lighter shoulders” phenomenon from the previous post: heavy-tailed distributions have sharper centers, lighter shoulders, and fatter extremes; the 1.96σ threshold falls squarely in the shoulder region, so the fat tails only emerge at much higher quantiles. Shortfall is a tax exclusive to high confidence levels.

The critical part comes next: this shortfall table is not two-dimensional. If the correction factor depended only on confidence and ν, reporting standards could print a “confidence × ν” lookup table. The simulation verdict is negative: adding w (the calibration-to-statistical weight) boosts model explanatory power by 19 percentage points. A two-variable table underestimates the correction factor by 0.83× in the most dangerous cell (99.9% confidence, ν=3, w=0.75): the table suggests widening intervals by 1.38×, but the actual requirement is 2.20×—following the table would still result in catastrophic undercoverage. The reason is straightforward: what determines the tail shape of the mixed distribution is not ν itself, but the proportion of total variance contributed by the calibration end. When w=0.25 and r=2, calibration error accounts for only 40% of total variance; the fat tails of the t-distribution are diluted by the thin Gaussian tails. When w=0.75, the calibration end comprises 86% of variance, making the mixed distribution nearly pure t; hence, the shortfall at the 99.9% quantile explodes. For the same ν=3, increasing w from 0.25 to 0.75 raises g at the 99.9% level from 1.35 to 2.20—a 60% difference despite identical degrees of freedom.

The “light-tail exemption” clause also holds only half the time. The backlog assumes g deviates from 1 by less than 0.02 when ν≥10, so the exemption applies only to ν<10. Simulation results: at 99% confidence, this holds (g=1.065 for ν=10, w=0.5 is acceptable; g=1.011 for ν=10, w=0.25 meets the threshold). But at 99.9% confidence, even with ν=10, cells with w≥0.5 yield g between 1.18 and 1.23—the light-tail exemption fails at high confidence levels. The exemption must be conditioned on both confidence level and calibration-end proportion. The intermediate 99.5% level completes this panoramic picture: at ν=3, w=0.75, g=1.52 (already approaching the worst-case shortfall at 99%); at ν=20, w=0.25, g=1.007 (requiring almost no adjustment). Thus, the shape of shortfall is not “degrees of freedom determine everything,” but rather a triple gradient—confidence level rising, ν decreasing, and calibration-end proportion increasing. When all three align, shortfall truly spirals out of control; any single-axis exemption leaks when the other two axes are maximized.

One practical point must be emphasized: correction is not free. g=2.20 implies widening interval half-widths by 120%. Compared to the screening-style truncation from the previous post (#1568)—zero width cost, screening rate 8–12.5%—which is better depends on the scenario. If calibration drift batches are identifiable and retractable (reader states can be recorded and re-calibrated), the screening-and-retract ledger looks better. If drift is hidden and impossible to identify post-hoc (e.g., equipment batch drift mixed into delivered data), one must accept the cost of interval widening, since no screening action is possible. These two clauses are not competing; they apply based on whether drift is identifiable.

Implications for practice: the reporting standard clause promising “correction for >99% high confidence” should not be a single lookup table, but a tiered rule. First, compute the calibration-end proportion (w·r / total variance). If this proportion is low (around 40% or less), the correction is mild (g at 99.9% ranges from 1.03 to 1.35); uniformly widening intervals conservatively suffices. If the proportion is high (around 85%), abandon closed-form approximation entirely and run Monte Carlo quantiles based on the empirical distribution of calibration drift, because any closed-form approximation is gambling against a shortfall of 2× magnitude. Literature aligns with this: under small samples, the Central Limit Theorem’s normal interval coverage indeed fails, requiring tail bounds that may support coverage for all sample sizes (PMID 20231867); with few studies, normal approximation coverage is generally below nominal, and Student-t corrections restore it (PMID 30634920, PMID 26573817); for complex targets like noncentrality parameters, exact closed-form intervals are constructible (PMID 36725775)—the problem is never the absence of closed forms, but choosing the wrong variables.

Evidence and Limitations

Verified: Across 72 simulation cells (1 million iterations each), g at 95% confidence falls entirely within 0.94–1.02 (no shortfall, even slight overcoverage); the w term explains ΔR²=0.19 of correction factor variance; the worst residual for a two-variable table is +0.83 (99.9%, ν=3, w=0.75, underestimated); for fixed ν=3, increasing w from 0.25 to 0.75 raises g at 99.9% from 1.35 to 2.20; the light-tail exemption (ν≥10) fails at 99.9% confidence when w≥0.5 (g=1.18–1.23).

Partially Verified: The t-distribution serves as an idealized model for heavy-tailed calibration drift; the true tail shape of calibrated data was not empirically measured. Conclusions depend on setting r=2 as the error multiplier; larger r would strengthen the tiered effect of w.

To Be Verified: Optimal tier design based on calibration-end proportion (how many tiers, residual error within tiers); whether g<1 at 95% confidence can translate into interval narrowing without breaking coverage; the combined effect of correction factor and correlation between calibration and statistical ends (ρ_cs).

Related Reading

Evidence Level: C
L1 Network Status: This topic has not yet entered the L1 correspondence network—zero hits for “confidence/degrees of freedom/z scaling/correction factor/coverage/quantile/tail function” in EDGE_TABLE_science-yixue.md, and no corresponding edges in CORRESPONDENCE_NETWORK.md (nearest neighbor E-061 belongs to HRV resting ICC proposition, serving only as analogical corroboration).
References: PMID 20231867, PMID 30634920, PMID 26573817, PMID 36725775

Evidence Level: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration.


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 *