标定端重尾咬不动 95% 区间,却在 99% 以上开口:膨胀与截尾的分工线

Written by

in

Language

Evidence level: C

均值配平的闭式膨胀因子在 95% 置信水平对重尾近乎免疫;重尾的账单只在 99% 以上出现,而且该交的税是「筛查退回」,不是「统一乘大 k」。

正文

上一篇(#1557)给出了保序膨胀因子的闭式解 k(r,w),但它建立在一个隐含假设上:标定误差是高斯的。真实世界里的标定漂移很少这么乖——判读者的状态波动、设备的批次漂移,往往带离群级的极端值,统计上叫「重尾」。这一篇的预设问题很直接:标定端换成重尾分布后,闭式区间会不会像 backlog 担心的那样「真正的失败模式是极端偏差落在区间外」,须不需要再叠加截尾处理?

我们做了九格仿真:标定端权重 w 取 0.25、0.5、0.75 三档,重尾度取自由度 3、5、10 的 t 分布(自由度越小尾巴越肥,并标准化到与高斯同方差),误差真实倍数 r=2,每格一百万次。三种补救方案同台:一是统一乘大 k,二是闭式 k 加筛查式截尾——标定偏差超过阈值的批次判为标定失控、退回重标,不进入名义区间的承诺范围,三是逐格蒙特卡洛分位区间,即知道尾部分布形状时的理论最优解。

第一个发现是个反转:95% 置信水平下,重尾根本没咬到人。九个格点的实际覆盖落在 0.947 到 0.959,全程贴着名义值 0.95。原因在于重尾分布的典型形状是「中央更尖、肩膀更轻、极尾更肥」——1.96 倍标准差这个门槛恰好落在肩膀区,重尾分布在这里的越界质量反而可能比高斯还少一点,肥尾巴要跑到更远的地方才显形。这解释了为什么均值配平的区间在常规置信水平下对轻度重尾意外地皮实。

第二个发现是账单的位置:重尾的代价集中在高置信水平。99% 名义水平下,仍按 z 值线性缩放的闭式区间实际覆盖掉到 0.980-0.988——名义超出率被放大了两到五倍;99.9% 水平更是掉到 0.990-0.998。z 缩放假设尾部按高斯的方式衰减,而重尾分布的极尾按幂律衰减,越往远处欠账越多。

第三个发现在 99% 水平的三方案对决。统一乘大 k 要把乘子抬到 5.89 才能让九格全覆盖住 0.99,平均区间半宽比基准 4.04 吹到 5.89,代价 +45.7%,而且在容易的格点严重过覆盖,白白扔掉精度。逐格蒙特卡洛分位区间平均半宽 4.61(+14.1%),最紧凑,但前提是你得知道尾部分布的真实形状——这恰恰是实践里最没把握的信息。真正的赢家是筛查式截尾:阈值取 1.5 倍真实标定标准差,九个格点覆盖全部冲到 0.998 以上,区间半宽一分未涨,代价是 8% 到 12.5% 的批次被筛掉退回重标。t 分布自由度为 5 时,它以零宽度代价达到甚至超过统一膨胀的覆盖,backlog 预设的「两步法进规范」成立——但成立的形态是「闭式 k 加筛查退回」,统一乘大 k 反而是三方案里最差的。

还埋着一个陷阱要标红:天真截尾。如果做法是把标定数据里超阈值的部分剪掉、按剪完后的方差重建区间,却不对真实误差做任何处理,覆盖会直接砸穿——九格实测 0.856 到 0.911,自由度 3 时连 0.90 都保不住。截尾必须是面向交付的筛查动作(筛掉失控批次),不是面向报表的修数美容(把不好看的尾巴剪掉假装不存在)。这两件事只差一个评估口径,后果差出一个数量级。

对判读实践的含义:气功外气重复判读的标定漂移(判读者状态波动天然带离群)如果确实重尾,95% 水平的报告规范沿用闭式膨胀因子即可,不必为重尾加码;99% 以上的高置信承诺则必须加筛查条款——按阈值筛查标定偏差、失控批次退回重标,比统一放大区间划算得多。文献侧与此呼应:算术均值在含离群数据下的疗效估计会被误导而须改用稳健处理(PMID 32267856),少研究场合的区间精度须靠专门的稳健方案保障(PMID 26573817),离群检测在证据合成里已是标准动作(PMID 32921002),而蒙特卡洛分位区间替代解析传播在复杂误差结构下本就是计量学共识(PMID 27114218)。

证据与局限

已验证:仿真九格、每格一百万次下,std 配平闭式 k 在 95% 水平覆盖 0.947-0.959(重尾免疫);99% 水平实覆盖 0.980-0.988、99.9% 水平 0.990-0.998(z 缩放在极尾欠账);筛查式截尾以零宽度代价达到 0.998 以上覆盖、筛掉率 8-12.5%;统一膨胀 99% 水平宽度代价 +45.7%;天真截尾覆盖 0.856-0.911。

部分验证:t 分布是重尾的理想化模型;真实标定漂移的尾部分布形状未实测,筛查阈值与筛掉率的映射在真实数据上未标定。

待验证:真实判读标定数据的尾指数实测;筛查退回的成本-覆盖权衡(重标成本多少算划算);z 缩放闭式在 99% 以上水平的自由度依赖修正项;重尾与标定-统计端相关(#1560 的 ρ_cs)同时存在时筛查方案是否仍保覆盖。

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「重尾、截尾、膨胀、覆盖、稳健、方差比」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 32267856, PMID 26573817, PMID 32921002, PMID 27114218

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。

Calibration of Heavy Tails: Inflation and Truncation Division Lines

Evidence level: C

Mean-balanced closed-form inflation factors are nearly immune to heavy tails at the 95% confidence level; the bill comes due only above 99%, and the tax levied is “screened rejection,” not “uniform multiplication by large k.”

Body

The previous entry (#1557) provided the closed-form solution for the order-preserving inflation factor, $k(r,w)$, but it rested on an implicit assumption: that calibration error follows a Gaussian distribution. In the real world, calibration drift is rarely so well-behaved—reader state fluctuations and equipment batch drift often carry outliers of extreme magnitude, statistically termed “heavy tails.” This entry addresses the direct question: if the calibration end is modeled with a heavy-tailed distribution, will the closed-form interval behave as backlog fears suggest—that “the true failure mode is extreme deviation falling outside the interval”—and does it require additional truncation?

We conducted a nine-grid simulation: calibration weights $w$ set to 0.25, 0.5, and 0.75; heavy-tailedness modeled via t-distributions with degrees of freedom (df) 3, 5, and 10 (lower df yields fatter tails, standardized to match Gaussian variance); true error multiplier $r=2$; one million trials per grid. Three remediation schemes were tested simultaneously: (1) uniform multiplication by large $k$; (2) closed-form $k$ plus screened truncation—batches exceeding the threshold are deemed calibration out-of-control, rejected for re-calibration, and thus excluded from the nominal interval’s promised coverage; (3) per-grid Monte Carlo quantile intervals, i.e., the theoretical optimum when tail distribution shape is known.

The first finding is a reversal: at the 95% confidence level, heavy tails do not bite. Actual coverage across the nine grid points fell between 0.947 and 0.959, staying consistently close to the nominal value of 0.95. This occurs because heavy-tailed distributions typically exhibit a shape that is “sharper in the center, lighter on the shoulders, and fatter at the extremes”—the 1.96σ threshold恰好 lands in the shoulder region, where heavy-tailed distributions may actually have less out-of-bounds mass than Gaussians; the fat tails only manifest far away. This explains why mean-balanced intervals are surprisingly robust against mild heavy-tailedness at conventional confidence levels.

The second finding concerns the location of the bill: heavy-tail costs concentrate at high confidence levels. At the nominal 99% level, the closed-form interval scaled linearly by $z$ values achieved actual coverage between 0.980–0.988—the nominal exceedance rate was amplified two to fivefold; at 99.9%, coverage dropped further to 0.990–0.998. The $z$-scaling assumption presumes tails decay like Gaussians, whereas heavy-tailed distributions decay via power laws, accumulating more debt the farther out one goes.

The third finding emerged from the three-way showdown at the 99% level. Uniform multiplication by large $k$ required raising the multiplier to 5.89 to cover all nine grids at 0.99, inflating the average interval half-width from the baseline 4.04 to 5.89—a cost increase of +45.7%—and resulting in severe over-coverage at the easier grids, wasting precision. Per-grid Monte Carlo quantile intervals averaged a half-width of 4.61 (+14.1%), the most compact, but this presupposes knowledge of the true tail distribution shape—the least certain information in practice. The true winner was screened truncation: setting the threshold at 1.5 times the true calibration standard deviation pushed coverage for all nine grids above 0.998, with zero increase in interval half-width, at the cost of screening and rejecting 8% to 12.5% of batches for re-calibration. When the t-distribution has df=5, screened truncation achieves coverage equal to or exceeding uniform inflation with zero width penalty, validating backlog’s preset “two-step approach to specification”—but the validated form is “closed-form $k$ plus screened rejection,” making uniform multiplication by large $k$ the worst of the three schemes.

One trap to highlight in red: naive truncation. If one simply clips the tail from calibration data beyond the threshold and rebuilds the interval using the variance of the clipped dataset without adjusting for the true error, coverage collapses directly—nine-grid实测 coverage plummeted to 0.856–0.911, failing even to maintain 0.90 when df=3. Truncation must be an action oriented toward delivery (screening out-of-control batches), not a cosmetic repair of the report (trimming ugly tails and pretending they don’t exist). These two approaches differ only in evaluation metric, yet their consequences differ by an order of magnitude.

Implications for interpretation practice: if Qigong external Qi re-reading calibration drift (reader state fluctuations naturally carry outliers) is indeed heavy-tailed, the 95% level reporting specification can retain the closed-form inflation factor without adding extra margin for heavy tails; however, high-confidence commitments above 99% must include screening clauses—threshold-based screening of calibration deviation and rejection of out-of-control batches for re-calibration is far more cost-effective than uniformly inflating the interval. Literature echoes this: arithmetic mean efficacy estimates under outlier-containing data are misleading and require robust treatment (PMID 32267856); interval precision in low-research settings must be safeguarded by dedicated robust schemes (PMID 26573817); outlier detection is now standard in evidence synthesis (PMID 32921002); and Monte Carlo quantile intervals replacing analytical propagation is metrological consensus under complex error structures (PMID 27114218).

Evidence and Limitations

Verified: Under the nine-grid simulation with one million trials each, the mean-balanced closed-form $k$ achieved 0.947–0.959 coverage at the 95% level (heavy-tail immune); at the 99% level, actual coverage was 0.980–0.988, and at the 99.9% level, 0.990–0.998 (z-scaling underestimates extreme tails); screened truncation reached coverage above 0.998 with zero width penalty, rejecting 8–12.5% of batches; uniform inflation at the 99% level incurred a +45.7% width cost; naive truncation yielded coverage 0.856–0.911.

Partially Verified: The t-distribution serves as an idealized model of heavy tails; the true tail distribution shape of real calibration drift remains unmeasured, and the mapping between screening thresholds and rejection rates has not been calibrated on real data.

To Be Verified: Empirical measurement of tail indices for real calibration data; cost-coverage trade-offs of screening rejection (what re-calibration cost justifies the expense); degree-of-freedom correction terms for closed-form $k$ at confidence levels above 99%; whether the screening scheme maintains coverage when both heavy-tailedness and calibration-statistical correlation (the $\rho_{cs}$ from #1560) coexist.

Related Reading

Evidence Level: C
L1 Network Status: This topic has not yet entered the L1 corresponding network—searching EDGE_TABLE_science-yixue.md for “heavy tail, truncation, inflation, coverage, robustness, variance ratio” yields zero hits; CORRESPONDENCE_NETWORK.md also lacks corresponding edges (nearest neighbor E-061 belongs to the HRV resting ICC proposition, serving only as analogical corroboration).
References: PMID 32267856, PMID 26573817, PMID 32921002, PMID 27114218

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 *