筛查式截尾的阈值-成本曲线:安全带 [2.0, 3.0],下界白花钱、上界翻车

Written by

in

Language

Evidence level: C

筛查阈值 c 从 1.2 扫到 4.0:覆盖-筛掉率之间隔着一条平坦的安全带,c 小于 2 是拿真金白银买多余的安全,c 大于 3 是让筛查条款形同虚设。

正文

上一篇(#1561)裁决了 99% 置信水平下的三方案对决:筛查式截尾以零宽度代价拿下全场——阈值取 1.5 倍真实标定标准差,覆盖冲到 0.998 以上,代价是 8% 到 12.5% 的批次退回重标。但这个赢家身上有个悬案:阈值 1.5 是硬编的。阈值调松一点(比如 2.0),筛掉率降下来、成本省下来,覆盖还保不保?调得更松(3.0、4.0),筛查是不是就筛不到东西了?这条阈值-成本曲线的形状,之前从没人扫出来过。这一篇把它扫出来。

我们做了五十四格仿真:阈值 c 取 1.2 到 4.0 六档,重尾度取自由度 3、5、10 的 t 分布(标准化到与高斯同方差),标定端方差占比 w 取 0.25、0.5、0.75 三档,真实标定精度按宣称的两倍生成,每格三十万次。每格记录两件事:筛掉率(多少批次因标定偏差越界被退回)和条件覆盖(通过筛查的批次里,真实误差落在闭式区间内的比例)。在此基础上叠加成本系数——每筛掉一个批次就要付一次重标工时,按「单位交付成本 = 1 + λ×筛掉率/(1−筛掉率)」折算,重标成本系数 λ 取 0.5。

第一条发现是安全带的存在:c 在 2.0 到 3.0 之间,九个重尾格点的条件覆盖全部 ≥0.99,筛掉率压在 1.4% 到 5.0% 之间。c=2.0 的边界格刚好压线(筛掉率最高 5.0%、覆盖最低 0.9957),c=2.5 最从容(筛掉率 1.9%-2.3%,覆盖 ≥0.9933),c=3.0 覆盖开始贴近 0.99 的地板(0.9910 起)。也就是说,backlog 预设的判据一成立:申报一条「c 不小于 2」就能全格达标——但更稳的写法是「c 取 2.5」,给自己留出覆盖与筛掉率两个方向的余量。

第二条发现是判据二不成立:最优 c 不随重尾度显著漂移。自由度 3(最肥的尾)与自由度 10(接近高斯)的可用阈值窗口漂移不足 0.5——重尾度改变的是曲线两端的失效方式,不改安全带的位置。申报规则不需要按尾部分型分档,一条 c 值走遍 ν。这省掉了报告规范里最麻烦的一类条款:先估计尾部形状、再查分档表——尾部形状恰恰是实践里最没把握的量。

第三条发现是判据三要改判:帕累托前沿在筛掉率 10% 附近没有陡坎——c 从 1.5 降到 1.2,筛掉率从 12.5% 爬到 21%,覆盖几乎不动(0.998 上下)。真正的陡坎在另一头:c 越过 3.0 之后覆盖跳水,3.5 时跌到 0.986-0.992、4.0 时 0.984-0.990,比不筛查还糟。原因是筛查形同虚设:阈值松到重尾批次的漏网概率升高,失控批次混进承诺范围,而闭式区间对它们的极尾误差无能为力。所以批次预算上限条款(判据三的预设)不必设——预算上限防的是筛太多,实测筛太多只发生在 c<1.5 的自费区,不设上限也没人会停在那里;真正要设的是阈值上界:c 不得超过 3.0,否则筛查条款反而制造虚假承诺。

成本侧的图景随之清晰:重标成本系数 0.5 时,单位交付成本随 c 单调下降(c=1.2 时 1.13,c=3.0 时 1.007),约束端只剩覆盖——成本最优点就是覆盖约束的边界 c=3.0。但这个「最优点」是刀尖上的最优点:压着覆盖地板没有余量,重尾度估计稍有偏差就翻车。实践里正确的选点是安全带中段:c=2.5 时多付 1.2% 的重标成本,换覆盖余量 0.3 个百分点以上。文献侧与此呼应:证据合成里的离群检测阈值设定本就是灵敏度分析的标准对象(PMID 32921002),含离群数据时朴素均值估计的误导幅度可以大到让结论反转(PMID 32267856),少研究场合的稳健区间修正须谨慎标定否则覆盖失准(PMID 26573817)。

对判读实践的三条申报规则:一,99% 水平筛查条款报「c=2.5」(或至少「c 不小于 2 且不大于 3」);二,无需按尾部分型分档,尾部形状未知的场合安全带照用;三,条款必须双边界——只写「阈值不小于 2」不写上界,等于给后人留了一个把筛查调到名存实亡的口子。筛查这条命是阈值给的,也能被阈值收回去。

证据与局限

已验证:五十四格(c×ν×w)、每格三十万次的仿真下,安全带 c∈[2.0, 3.0] 内全格条件覆盖 ≥0.99 且筛掉率 ≤5.0%;c=2.5 时筛掉率 1.9%-2.3%、覆盖 ≥0.9933;c=3.5 覆盖跌破 0.99、c=4.0 跌到 0.984-0.990;最优窗跨 ν 漂移 <0.5;筛掉率 10% 附近前沿平缓无陡坎——均与 #1561 九格结果量级吻合(c=1.5 时筛掉率 8-12.5%、覆盖 0.998+)。

部分验证:t 分布是重尾的理想化模型,真实标定漂移的尾形未实测;重标成本系数 λ=0.5 是外生假设,未按实际工时标定;统计端按 df=30 卡方尺度族逐次抽取,df 更小时曲线可能整体下移。

待验证:重标成本系数的真实标定与成本最优点的敏感性(λ 从 0.2 到 2 扫描,看最优点是否仍在 [2, 3] 内);标定端与统计端相关(#1560 的 ρ_cs 大于零)时安全带是否保形——相关会让失控批次的统计端噪声同向放大,覆盖端失效可能提前到 3.0 之前;90% 与 95% 水平下安全带是否同样存在还是整体左移。

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「截尾、winsorize、重尾、heavy tail、筛查、筛掉、帕累托、pareto、离群、outlier」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,不覆盖筛查阈值-成本命题)
参考文献: PMID 32921002, PMID 32267856, PMID 26573817

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。

Threshold-Cost Curves for Screening Truncation: Seatbelt [2.0, 3.0], Lower Bound White Spending, Upper Bound Crashing

Evidence level: C

Screening threshold $c$ sweeps from 1.2 to 4.0: coverage vs. rejection rate is separated by a flat safety belt; $c < 2$ buys excess safety at real cost, $c > 3$ renders the screening clause nominal.

Main Text

The previous post (#1561) adjudicated a three-way contest at 99% confidence: screening with truncation wins outright with zero width penalty—threshold set at 1.5 times the true calibration standard deviation yields coverage above 0.998, at the cost of 8% to 12.5% batch re-calibration. But this winner carries an open question: threshold 1.5 is hard-coded. Loosening the threshold (e.g., to 2.0) lowers rejection rate and saves cost—does coverage hold? Loosening further (3.0, 4.0)—does screening stop working entirely? The shape of this threshold-cost curve has never been mapped before. This post maps it.

We ran fifty-four simulations: threshold $c$ at six levels (1.2 to 4.0), tail heaviness via t-distributions with df = 3, 5, 10 (standardized to match Gaussian variance), calibration-end variance weight $w$ at three levels (0.25, 0.5, 0.75), true calibration precision generated at twice the claimed level, thirty thousand iterations per cell. Each cell records two metrics: rejection rate (fraction of batches rejected due to calibration drift exceeding bounds) and conditional coverage (proportion of passed batches whose true error falls within the closed interval). Superimposed is the cost coefficient—each rejected batch incurs re-calibration labor, scaled by unit delivery cost = 1 + $\lambda \times$ rejection rate / (1 − rejection rate), with re-calibration cost coefficient $\lambda = 0.5$.

First finding: the safety belt exists. For $c \in [2.0, 3.0]$, all nine heavy-tail cells achieve conditional coverage ≥ 0.99, with rejection rates bounded between 1.4% and 5.0%. At $c = 2.0$, the boundary cell just meets the line (max rejection 5.0%, min coverage 0.9957); $c = 2.5$ is most comfortable (rejection 1.9%–2.3%, coverage ≥ 0.9933); $c = 3.0$ brings coverage down to the 0.99 floor (starting at 0.9910). Thus, the backlog-prescribed criterion holds: declaring “c ≥ 2” ensures all cells meet standards—but a more robust formulation is “c = 2.5”, reserving margin in both coverage and rejection directions.

Second finding: criterion two fails. Optimal $c$ does not shift significantly with tail heaviness. The usable threshold window between df = 3 (heaviest tails) and df = 10 (near-Gaussian) drifts less than 0.5—tail heaviness alters failure modes at the curve ends, not the safety belt position. Reporting rules need not stratify by tail type; a single $c$ value suffices across $\nu$. This eliminates the most cumbersome class of reporting clauses: first estimate tail shape, then consult stratification tables—tail shape is precisely the least certain quantity in practice.

Third finding: criterion three must be revised. The Pareto frontier shows no steep cliff near rejection rate 10%—lowering $c$ from 1.5 to 1.2 raises rejection from 12.5% to 21%, with coverage nearly unchanged (around 0.998). The true cliff lies elsewhere: once $c$ exceeds 3.0, coverage plummets—at $c = 3.5$, coverage drops to 0.986–0.992; at $c = 4.0$, to 0.984–0.990, worse than no screening. This occurs because screening becomes nominal: loosening the threshold increases escape probability for heavy-tail batches, allowing out-of-control batches to mix into the promised range, while the closed interval cannot handle their extreme tail errors. Thus, the batch budget upper bound clause (preset in criterion three) is unnecessary—the budget limit prevents over-screening, but empirical over-screening only occurs in the self-funded zone $c < 1.5$; setting no upper bound ensures no one stops there. What truly needs setting is the threshold upper bound: $c$ must not exceed 3.0, or the screening clause itself creates false promises.

The cost picture becomes clear: with re-calibration cost coefficient 0.5, unit delivery cost monotonically decreases with $c$ ($c = 1.2$: 1.13; $c = 3.0$: 1.007). Constraints reduce to coverage alone—the optimal point is the coverage-bound $c = 3.0$. But this “optimal point” is razor-edge optimal: pressed against the coverage floor with no margin, any slight error in tail heaviness estimation causes failure. Practically, the correct choice is mid-safety belt: $c = 2.5$ incurs an extra 1.2% re-calibration cost, trading for over 0.3 percentage points of coverage margin. Literature echoes this: outlier detection threshold setting in evidence synthesis is itself a standard sensitivity analysis target (PMID 32921002); naive mean estimation with outlier data can mislead conclusions so severely that results reverse (PMID 32267856); robust interval correction in less-studied settings must be carefully calibrated, or coverage becomes inaccurate (PMID 26573817).

Three reporting rules for interpretation practice:
1. At 99% level, declare screening clause as “c = 2.5” (or at least “c between 2 and 3 inclusive”);
2. No need to stratify by tail type—use the safety belt even when tail shape is unknown;
3. The clause must have dual boundaries—stating only “threshold ≥ 2” without an upper bound leaves room for future adjustment to render screening nominal. Screening’s life is given by the threshold, and can be taken back by it.

Evidence and Limitations

Verified: Under fifty-four simulations ($c \times \nu \times w$), each with 30,000 iterations, the safety belt $c \in [2.0, 3.0]$ achieves conditional coverage ≥ 0.99 and rejection rate ≤ 5.0% across all cells; at $c = 2.5$, rejection rate is 1.9%–2.3%, coverage ≥ 0.9933; at $c = 3.5$, coverage falls below 0.99, and at $c = 4.0$ to 0.984–0.990; optimal window drifts < 0.5 across $\nu$; the Pareto frontier near rejection rate 10% is flat with no steep cliff—all consistent with #1561’s nine-cell results ($c = 1.5$: rejection 8–12.5%, coverage ≥ 0.998).

Partially verified: t-distribution is an idealized model for heavy tails; true tail shape of calibration drift was not empirically measured; re-calibration cost coefficient $\lambda = 0.5$ is an exogenous assumption, not calibrated to actual labor hours; statistical end uses df = 30 chi-square scale family for sequential sampling, so curves may shift downward overall if df is smaller.

To be verified: true calibration of re-calibration cost coefficient and sensitivity of the optimal point (sweeping $\lambda$ from 0.2 to 2 to see if optimum remains within [2, 3]); whether the safety belt preserves shape when calibration end and statistical end are correlated (#1560’s $\rho_{cs} > 0$)—correlation amplifies statistical noise in the same direction for out-of-control batches, potentially causing coverage failure before 3.0; whether safety belts exist at 90% and 95% levels or shift leftward as a whole.

Related Reading

Evidence level: C
L1 network status: This topic has not yet entered the L1 corresponding network—EDGE_TABLE_science-yixue.md yields zero hits for “truncation, winsorize, heavy tail, heavy tails, screening, reject, Pareto, pareto, outlier, outlier”, and CORRESPONDENCE_NETWORK.md lacks corresponding edges (nearest neighbor E-061 belongs to HRV resting ICC proposition, serving only as analogical corroboration, not covering the screening threshold-cost proposition).
References: PMID 32921002, PMID 32267856, PMID 26573817

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 *