Blog

  • 零差异地板能不能在收材料之前算出来:判读门禁对边际先验精度的依赖裁决

    Evidence level: C

    门禁要拦人,就得在人进门之前算得出地板——而地板公式吃的是各档真实比例,这个比例恰恰是研究还没开始测的东西。

    正文

    前一篇(#1432)立了一条门禁:任何用「伤害近似 TVD 除以 H」做五百分点级裁决的判读设计,必须先算零差异地板——两份完全相同材料的预期伤害值。地板超过判据一半,这个设计就不能判,应该拦在招募之前,而不是事后找借口。这条门禁的逻辑本身没毛病,但它藏着一个没验过的前提:地板是事前可算的吗?

    问题出在地板公式的输入上。地板要吃各档的真实比例 p——哪个难度档占多少。而 p 恰恰是这个研究要测的东西。你还没收材料,手里根本没有 p。实践中只有三种替代:一、先做预试验,用小样本估 p;二、翻同一批材料的历史判读记录,用旧边际当新边际;三、干脆用经验构型假设,比如「主档占九成」。三种替代都有误差,误差顺着公式传到地板,地板不准,门禁就可能是拦错了人或放错了人。

    误差传导的数学骨架是现成的:当分子和分母都是随机量时,比值的方差不能把分母当常数算,必须用 delta 方法做一阶近似——这在致癌性检验的比值估计文献里有标准处理(PMID 8241374)。而「小样本估 p」这条路有一个更隐蔽的坑:当类别数超过样本量时,多样性的朴素估计会严重偏离真值,Simpson 指数的抽样方差需要闭式无偏修正才可信(PMID 39020976)。这直接打脸替代方案一:预试验 30 例去估一个五档分布,低频档的估计噪声大到可以吞掉地板对判据的优势。替代方案三的病根在另一处:kappa 一致性研究早就证明,流行度(边际分布的偏斜)必须和偏差、真实一致拆开单独报告,不能混进单一系数(PMID 8501467)——「主档占九成」这种单一假设,把边际形状压扁成了一个数,地板预测自然跟着压扁的形状一起错。

    更麻烦的是地板对 p 的依赖不是单调的。主档比例越接近 1,H(归一化熵)越小,分母折出来的伤害越大——门禁显得越严;但与此同时,小档比例越小,其根号项贡献越低,两者可能部分抵消。也就是说,在最需要门禁的高一致区,地板对 p 误差的敏感度可能反而下降。这既是好消息(先验粗一点也够用)也是坏消息(曲面上存在拐点,必须标出来)。

    于是本篇的设计是纯解析加仿真:在「材料难度梯度(连续阈值切档)× 主档比例 × 档数」三维网格上,把地板折伤害画成曲面;扫替代 p 的误差从 5% 到 50%,标出「地板预测误差不超过 20%」所需的 p 精度门槛;最后用 3 至 5 批真实材料的判读边际做外推检验,看曲面预测能否落到实测点上。

    裁决标准三分支:若主档比例误差在 10% 以内就能把地板预测误差压到 20% 以内,则门禁可以事前执行,只须强制报告边际先验来源(预试验 / 历史 / 经验假设,三者标记不同置信级);若需要 p 的误差在 3% 以内才够,则门禁事实上不可事前执行,必须改成「预试验强制 30 例算地板」,承认这条门禁是半事后的;若折伤害对 p 的响应在近饱和区(主档超 0.95)反而变平,则高一致区的地板可当作常数,门禁简化为按档数查表——查表门禁执行成本最低,代价是只对高一致区有效。

    这条裁决决定的是 #1432 门禁的执行时机:能事前算,门禁就是设计审查工具;不能事前算,它就只是事后归因工具。两者的科研管理学地位完全不同,不能用同一个名字混着叫。

    证据与局限

    已验证: 比值随机分母的 delta 方法方差近似(PMID 8241374);kappa 流行度效应须单独报告(PMID 8501467)。
    部分验证: 类别数超样本量时朴素多样性估计严重偏离、需闭式修正(PMID 39020976,物理文献场景,迁移到判读档位分布是类比)。
    待验证: 三维网格地板曲面的形状本身;10% / 3% 两条 p 精度门槛的实测数值;主档>0.95 时响应是否真的变平;3-5 批真实材料边际的外推检验。均未做,本篇全部结论均为设计裁决待考。

    相关阅读

    证据等级: C
    参考文献: PMID 39020976, PMID 8241374, PMID 8501467

    L1 网络状态:该主题尚未进入 L1 对应网络。EDGE_TABLE 中最近邻为「气功」边(48 条 A 级证据,PMID 18565332,slug ayurveda-three-doshas-frequency),但该边对应气功外气文献支撑,与本篇纯统计方法学主题(TVD 地板边际先验精度)不构成同一条边;CORRESPONDENCE_NETWORK 无 TVD/地板/边际先验对应条目。故证据等级封顶 C。


    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。

    Can Zero-Difference Flooring Be Calculated Before Material Receipt: A Decision on the Dependence of Access Control on Marginal Prior Precision

    Evidence level: C

    To gatekeep, one must calculate the floor before participants enter—yet the floor formula consumes true proportions across difficulty tiers, which are precisely what the study has not yet measured.

    Body

    The previous entry (#1432) established a gatekeeping rule: any adjudication design using “harm approximation TVD divided by H” for five-percentage-point decisions must first calculate the zero-difference floor—the expected harm values for two identical materials. If the floor exceeds half the criterion, this design cannot adjudicate and should be blocked before recruitment rather than post-hoc rationalized. The logic of this gatekeeping rule itself is sound, but it conceals an unverified premise: Is the floor calculable in advance?

    The problem lies in the inputs of the floor formula. The floor consumes true proportions $p$ across tiers—which difficulty tier occupies what share. Yet $p$ is precisely what this study must measure. Before collecting data, one possesses no $p$. Practically, only three alternatives exist: (1) conduct a pilot study to estimate $p$ with a small sample; (2) flip historical adjudication records of the same materials, using old margins as new margins; (3) simply adopt empirical configuration assumptions, such as “the primary tier occupies ninety percent.” All three alternatives introduce error, which propagates through the formula to distort the floor. If the floor is inaccurate, gatekeeping may wrongly exclude or include participants.

    The mathematical skeleton for error propagation is ready: when both numerator and denominator are random quantities, the variance of their ratio cannot treat the denominator as constant; one must apply the delta method for first-order approximation—a standard treatment in ratio estimation literature for carcinogenicity tests (PMID 8241374). The “small-sample estimation of $p$” route has a more insidious pitfall: when the number of categories exceeds the sample size, naive diversity estimation severely deviates from truth; Simpson’s index sampling variance requires closed-form unbiased correction to be credible (PMID 39020976). This directly refutes alternative (1): a pilot of 30 cases estimating a five-tier distribution, the noise in low-frequency tiers is so large it swallows the floor’s advantage over the criterion. The root cause of alternative (3) lies elsewhere: kappa consistency studies have long proven that prevalence (skewness of the marginal distribution) must be reported separately from bias and true agreement, not mixed into a single coefficient (PMID 8501467)—”the primary tier occupies ninety percent” flattens the marginal shape into a single number, so the floor prediction naturally follows this flattened shape and errs.

    More troubling is that the floor’s dependence on $p$ is not monotonic. As the primary tier proportion approaches 1, $H$ (normalized entropy) shrinks, inflating harm via the denominator—gatekeeping appears stricter; yet simultaneously, as the small-tier proportion shrinks, its square-root contribution diminishes, potentially partially offsetting the effect. Thus, in the high-consistency region where gatekeeping is most needed, the floor’s sensitivity to $p$ error may actually decrease. This is both good news (coarse priors suffice) and bad news (the surface contains inflection points that must be marked).

    Therefore, this paper’s design is purely analytical plus simulation: on a three-dimensional grid of “material difficulty gradient (continuous threshold tiering) × primary tier proportion × number of tiers,” we plot the floor-derived harm as a surface; we sweep $p$ estimation error from 5% to 50%, marking the $p$ precision threshold required for “floor prediction error within 20%”; finally, we use adjudication margins from 3 to 5 batches of real materials for extrapolation testing, checking whether surface predictions fall on measured points.

    The verdict criteria branch into three: If primary tier proportion error within 10% suffices to keep floor prediction error under 20%, then gatekeeping can execute pre-data collection, requiring mandatory reporting of margin prior sources (pilot / historical / empirical assumption, marked with different confidence levels); If $p$ error must be under 3% to suffice, then gatekeeping cannot practically execute pre-data collection, and must shift to “mandatory 30-case pilot to compute the floor,” acknowledging this gatekeeping is semi-post-hoc; If harm response flattens in near-saturated regions (primary tier >0.95), then the floor in high-consistency zones can be treated as constant, simplifying gatekeeping to table lookup by tier count—table lookup incurs the lowest execution cost but applies only to high-consistency zones.

    This verdict determines the timing of #1432 gatekeeping execution: if calculable pre-data, gatekeeping serves as design review tool; otherwise it remains post-hoc attribution tool. Their scientific management status differs fundamentally and cannot share the same name.

    Evidence and Limitations

    Verified: Delta-method variance approximation for ratio with random denominator (PMID 8241374); kappa prevalence effect must be reported separately (PMID 8501467).
    Partially Verified: Naive diversity estimation severely deviates when categories exceed sample size, requiring closed-form correction (PMID 39020976, physics literature context; transferred to adjudication tier distribution via structural analogy).
    Pending Verification: The shape of the three-dimensional floor surface itself; empirical values for the 10%/3% $p$ precision thresholds; whether response truly flattens when primary tier >0.95; extrapolation testing using margins from 3–5 batches of real materials. None have been performed; all conclusions in this paper remain design verdicts pending verification.

    Related Reading

    Evidence Level: C
    References: PMID 39020976, PMID 8241374, PMID 8501467

    L1 Network Status: This topic has not yet entered the L1 corresponding network. The nearest neighbor in EDGE_TABLE is the “Qigong external qi” edge (48 A-level evidence items, PMID 18565332, slug ayurveda-three-doshas-frequency), but this edge corresponds to Qigong external qi literature support, not the pure statistical methodology theme of this paper (TVD floor margin prior precision); CORRESPONDENCE_NETWORK lacks TVD/floor/margin prior entries. Hence evidence level capped at C.


    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.

  • 校正不是开关是旋钮:把扣掉多少节律当可调参数,看识别力会不会先升后降

    Evidence level: B

    一致性指标永远欢迎校正,识别力指标未必——两个指标对”扣多狠”的响应曲线形状不同,这条曲线的峰值位置才是”该扣多少”的定量答案。

    正文

    前面几篇把节律校正的账算了两轮:扣掉昼夜与季节节律之后,脉搏谐波谱的重测一致性会涨,且涨在哪几阶有分阶裁决。但所有这些设计都把校正当成了开关——要么扣,要么不扣。真实世界里校正是一个连续的旋钮:记 λ 为扣掉的比例,λ=0 是原始谱,λ=1 是全扣,中间可以只扣一部分。缩尾估计、经验贝叶斯收缩这类现成方法都能实现部分扣除。

    这个旋钮之所以关键,是因为一致性指标和识别力指标对 λ 的响应曲线形状必然不同。ICC 是个体间方差除以总方差:扣节律削的是个体内方差,只要个体间方差不同时塌掉,ICC 就随 λ 单调上升——扣得越狠涨得越多。但识别力(AUC)问的是另一件事:同一个人的两次测量,相似度超过不同人那对的概率。它只看”个体间隔相对个体抖动的比值”。若节律里混着个体签名的一部分,λ 增大就会先降噪(AUC 升)后削签名(AUC 降)——曲线上必然出现一个内部峰值。这个峰值位置 λ* 就是”该扣多狠”的定量答案,而现有文献要么全扣要么不扣,从没人画过这条曲线。

    ICC 对 λ 单调上升这件事还有个更深的坑:ICC 本身高度依赖取值分布。仿真研究显示,同样质量的数据,受试者分布取凸形、均匀或凹形,算出的 ICC 大小完全不同(PMID 29707825);对同一指标的实测也发现,同一变量在不同数据集里的 ICC 可以差几个数量级(PMID 15485730)。这意味着 λ 变化时 ICC 与 AUC 完全可能脱钩——一条单调升、一条先升后降——此前所有”ICC 涨了所以校正合理”的论证都欠一条检验。

    脉搏谐波这边,节律成分真实存在是有系统综述背书的:低频谐波随体力活动与交感兴奋升高、随静息与副交感兴奋降低,疾病与生理状态会改写第一到第十阶的强度分布(PMID 38481553)。而”个体签名”这个识别目标本身就要求跨周稳定——颈动脉血流音作为生物特征的研究里,同一人间隔数周的多条录音保持了可区分的谱特征,分类准确率超过 95%(PMID 34640975)。如果脉搏谐波要走到同样的应用,就必须回答:哪些成分是签名、哪些是节律噪声、以及校正时该把它们各自保留多少。

    可做的检验很直接:在既有三臂设计的数据上令 λ 从 0 到 1 步长 0.1,每个点上同时算 ICC 与开集识别 AUC(训练季与测试季严格分开,冒充者按年龄性别血压匹配),画出两条曲线。判据分四格:若 AUC 存在内部峰值且 λ<1,则全扣是过度校正,最优策略是部分扣除,λ 本身可作群体/个体节律成分比的估计量;若 AUC 单调不降,则节律纯为噪声、全扣最优;若 AUC 从 λ=0 起单调下降,则节律主要是个体签名、任何校正都有害;若两条曲线峰值位置相差超过 0.3,则 ICC 与 AUC 不可互换,以 ICC 论证校正合理性的既往研究需要重做。

    证据与局限

    已验证: 脉搏谐波低阶成分随生理状态波动有系统综述支撑(PMID 38481553);ICC 依赖取值分布、跨研究不可直接比较有仿真与实证支撑(PMID 29707825、15485730);脉搏类信号可作为跨周稳定的生物特征有直接先例(PMID 34640975)。

    部分验证: “ICC 与 AUC 对校正强度的响应曲线形状不同”是统计结构推断,尚未在脉搏谐波数据上实测。

    待验证: AUC(λ) 是否存在内部峰值;λ* 是否可作为节律成分比的估计量;两指标峰值位置差是否超过 0.3。

    相关阅读

    证据等级: B
    参考文献: PMID 38481553, PMID 29707825, PMID 15485730, PMID 34640975

    L1 网络状态:该主题尚未进入 L1 对应网络。L1(CORRESPONDENCE_NETWORK / EDGE_TABLE_science-yixue)中最近邻为 E-061「呼吸节律(0.1 Hz 谱峰)↔ Mayer 波/自主神经谱」,状态 open、层级 L2+,但该边针对呼吸节律与自主神经谱峰,非”节律校正强度 × 个体识别 AUC”同一条边。EDGE_TABLE 中「脉象」概念文章 73 篇、A级 5,「时辰」A级 9,均未覆盖校正强度扫描设计。故本文按方法学推断处理,证据等级封顶 B,不得写作已验证机制。

    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 is not a switch but a knob: removing rhythm as a fixed parameter to see if recognition power first rises then falls

    Evidence level: B

    Consistency metrics always welcome calibration, but recognition metrics do not necessarily—because the two metrics respond differently to “how much rhythm is removed.” The peak location of this response curve provides the quantitative answer for “how much should be removed.”

    Body

    Previous articles have calculated the rhythm-calibration account twice: after removing circadian and seasonal rhythms, the retest consistency of the pulse harmonic spectrum increases, and the increase varies by harmonic order. Yet all these designs treat calibration as a switch—either remove or do not remove. In reality, calibration is a continuous dial: let λ denote the proportion removed, where λ=0 represents the original spectrum and λ=1 represents full removal; intermediate values allow partial removal. Established methods such as shrinkage estimation and empirical Bayesian shrinkage can achieve partial removal.

    This dial is critical because the response curves of consistency metrics and recognition metrics to λ must differ. ICC (Intraclass Correlation Coefficient) is defined as between-subject variance divided by total variance: removing rhythm reduces within-subject variance, so as long as between-subject variance does not collapse simultaneously, ICC rises monotonically with λ—the more aggressively removed, the greater the increase. However, recognition (AUC) addresses a different question: whether the similarity between two measurements of the same individual exceeds the similarity between pairs of different individuals. It examines only “the ratio of between-subject separation to within-subject jitter.” If rhythm contains part of the individual signature, increasing λ will first reduce noise (AUC rises) then erode the signature (AUC falls)—necessarily producing an internal peak on the curve. The location of this peak, λ*, is the quantitative answer for “how much should be removed,” yet existing literature either removes everything or nothing; no one has plotted this curve.

    The fact that ICC rises monotonically with λ presents a deeper pitfall: ICC itself heavily depends on the distribution of values. Simulation studies show that for data of equal quality, choosing convex, uniform, or concave distributions of subjects yields completely different ICC values (PMID 29707825); empirical measurements of the same metric also reveal that ICC for the same variable can differ by several orders of magnitude across datasets (PMID 15485730). This means that as λ varies, ICC and AUC may completely decouple—one rising monotonically, the other rising then falling—rendering prior arguments that “ICC increased, therefore calibration is justified” incomplete without this test.

    Regarding pulse harmonics, the existence of rhythmic components is supported by systematic reviews: low-frequency harmonics increase with physical activity and sympathetic excitation, decrease with rest and parasympathetic excitation, and disease or physiological states reshape the intensity distribution across the first to tenth orders (PMID 38481553). Meanwhile, the recognition target “individual signature” itself requires cross-week stability: in studies of carotid blood flow as a biometric, multiple recordings spanning weeks from the same individual maintained distinguishable spectral features, achieving classification accuracy exceeding 95% (PMID 34640975). If pulse harmonics are to achieve similar applications, they must answer: which components constitute the signature, which are rhythmic noise, and how much of each should be retained during calibration.

    The test to perform is direct: on existing three-arm design data, let λ step from 0 to 1 in increments of 0.1, computing both ICC and open-set recognition AUC at each point (with training and testing seasons strictly separated, impostors matched by age, sex, and blood pressure), then plot both curves. The criteria divide into four cases: if AUC exhibits an internal peak with λ<1, then full removal is over-calibration; the optimal strategy is partial removal, and λ itself serves as an estimator of the ratio of group/individual rhythmic components; if AUC monotonically does not decrease, then rhythm is purely noise and full removal is optimal; if AUC monotonically decreases from λ=0, then rhythm primarily constitutes the individual signature and any calibration is harmful; if the peak locations of the two curves differ by more than 0.3, then ICC and AUC are not interchangeable, and prior studies relying on ICC to justify calibration must be redone.

    Evidence and Limitations

    Verified: Low-order pulse harmonic components fluctuate with physiological states, supported by systematic reviews (PMID 38481553); ICC dependence on value distribution and lack of direct comparability across studies are supported by both simulation and empirical evidence (PMID 29707825, 15485730); pulse-like signals can serve as cross-week stable biometrics, with direct precedent (PMID 34640975).

    Partially Verified: “ICC and AUC have different response curve shapes to calibration intensity” is a statistical structural inference, not yet empirically tested on pulse harmonic data.

    To Be Verified: Whether AUC(λ) exhibits an internal peak; whether λ* can serve as an estimator of the rhythmic component ratio; whether the difference between peak locations exceeds 0.3.

    Further Reading

    Evidence Level: B
    References: PMID 38481553, PMID 29707825, PMID 15485730, PMID 34640975

    L1 Network Status: This topic has not yet entered the corresponding L1 network. In L1 (CORRESPONDENCE_NETWORK / EDGE_TABLE_science-yixue), the nearest neighbor is E-061 “Respiratory rhythm (0.1 Hz spectral peak) ↔ Mayer wave / autonomic spectrum,” with status open and level L2+, but this edge pertains to respiratory rhythm and autonomic spectral peaks, not “calibration intensity × individual recognition AUC” on the same edge. The EDGE_TABLE contains 73 articles on “pulse diagnosis,” 5 at Level A, and 9 on “time of day,” all of which do not cover calibration intensity scanning designs. Therefore, this article is treated via methodological inference, with evidence level capped at B and not written as a verified mechanism.

    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.

  • 三轴塌缩成一根轴:m 从 2 扫到 16,可用域下界始终贴着两倍统计地板走,df≥50 且 ε≥0.10 的统一判据退役

    Evidence level: C

    上一篇文章把 df 当第二根轴、把判读构型数 m 钉死在 4;本文把 m 解冻从 2 扫到 16,发现 ε×df×m 三轴可用域整体塌缩到一根合并轴——统计地板系数 F 上:窗口下界 ε* ≈ 2.0×F,发散线 ≈ 1.0–1.4×F,「df≥50 且 ε≥0.10」的统一判据退役,换成「F≤0.047」一根线。

    正文(1000-1500字)

    df×ε 窗口那篇文章留下明确欠账:全部窗口边界是在 m=4(每例四个判读构型)下扫出来的,而 m 本身是设计变量——两点注入的省预算方案 m=2,和加构型换精度的 m=16,窗口下界不可能在同一位置。本文沿用父文全部模型(总误差 T(τ₂)=ετ₂/√m+√(2/df)·(τ+c/τ−2ρ√m)/(2√m),c=1+ρ²(m−1),ρ=0.4;闭式为定点迭代 τ=√(v(τ)·√(2/df)/(2ε)),可用标准为闭式点位与自洽最优点的偏差 ≤15%),做一轮 df∈{20,30,50,100,200}×m∈{2,4,8,16}×ε 步长 0.0025 的三轴网格。

    第一层读数是 m 轴的移动量。同一 df 档内,窗口下界随 m 单调下压:df=20 时从 m=2 的 0.225 压到 m=4 的 0.135、m=8 的 0.075、m=16 的 0.055;df=50 时从 0.140 压到 0.035。父文预设的裁决线是「m 轴上窗口下界变化 <0.02 个 ε 则统一判据成立」——实测最大移动 0.090,超线四倍半,统一判据不成立。m=2 对 m=4 的抬高在 df=20/30/50 分别为 +0.090/+0.075/+0.055,全部越过 0.05 的「两点注入单独定线」触发线;df=100/200 时降到 +0.040/+0.025,触发线内侧。所以分档答案很干脆:df≤50 的紧预算区,两点注入必须单独定线;df≥100 的宽松区,m=2 与 m=4 可共用一张表。

    第二层读数才是本文真正的发现:三轴不用分档,因为它们塌缩了。把每个格子的下界除以统计地板系数 F=√(2/df)·(√c−ρ√m)/√m,得到 α=ε*/F:m∈{2,4,8} 的全部 15 个格子里 α 落在 1.94–2.18,几乎是一根水平线。也就是说,df 与 m 这两个看似独立的旋钮,只通过乘积 F 起作用——df 砍半与 m 从 16 换成 4 在地板上等价,窗口下界、发散边界全被 F 一根轴钉死。这不是巧合而是代数必然:地板公式里 df 与 m 本来就以乘积形式出现(√(2/df) 乘 (√c−ρ√m)/√m),闭式迭代和最优点比较都建立在这条地板之上,下界自然继承同一结构。发散线同样塌缩:各格闭式发散的 ε 上限乘 √df 后是常数(m=2 约 0.50、m=4 约 0.39),换算成 F 的倍数即 1.0–1.4 倍——父文那条「ε·√df≤0.71」的对角线,在 m 轴上是一族平行对角线,族参数还是 F。

    塌缩律有一条诚实的边界:m=16 档的 α 系统性抬到 2.72–2.90,不再贴着 2.0 走。这一档的地板 F 只有 0.006–0.019,与 ε 扫描步长 0.0025 同量级,抬升既可能只是网格量化伪影(步长占了下界的一到四成),也可能是一阶地板公式在高斯近似之外被高阶项接管的真实边界。好消息是抬升方向偏保守——α 偏大意味着实际可用域比判据承诺的更宽,照 F 判据执行不会把可用的设计误判出局;坏消息是 F<0.01 区间内判据给出的「富余」不可当真。裁决留给后续:把 ε 步长加密一个量级重扫 m=16 档,α 回落则为伪影,不回落则一阶公式须加修正项。在这之前,操作上把塌缩律的适用区间标为 F≥0.01 即可,m=16 的高构型方案本来就不是省预算路线的主力。

    这组塌缩直接改写三条操作守则。其一,样本量表从 df 一根轴换成 F 一根轴:要保住 ε≥0.10 的可用窗口,判据统一为 F≤0.047(=0.10/2.1),查表只需算一个乘积。其二,例数预算的换算有了定量版:m=4、df=50 时 F=0.042 达标,而 m=2 要压到同样的 F 需 df≈148——两点注入想要与四构型同等的统计地板,自由度要翻三倍,这就是「最小构型单独定线」的代数内容。其三,敏感性排序的裁决:紧 df 端 m 轴比 df 轴更敏感(m 从 2 到 16 下界移动 0.170,超过 m=4 时 df 从 20 到 200 的移动 0.090),「例数预算优先于标定精度投入」方向成立,但要改一个字——优先补的不是例数本身,是合并地板 F,加构型、加遍数、加重复哪个便宜补哪个,ε 端砸钱换不来窗口。

    测量学文献对「构型数与自由度共同决定精度」有独立支撑:方差比置信区间的精确性同时依赖自由度结构与平衡设计(PMID 4005371);方差被估计时朴素区间覆盖率系统性失准、须按自由度修正(PMID 36725775);信度估计的精度随样本量剧烈变化,样本量规划是设计的一部分(PMID 40119099、24338600)。

    证据与局限

    已验证: ε(m,df) 全网格数值为直接实算(df∈{20,30,50,100,200}×m∈{2,4,8,16}×ε 步长 0.0025,ρ=0.4,可用=闭式与自洽最优偏差≤15%);α=ε/F 在 m∈{2,4,8} 全部 15 格落在 1.94–2.18 为同一扫描输出;发散线 ε_div·√df 为常数(m=2 约 0.50、m=4 约 0.39)与 k>k_crit=c/(c−ρ²m) 的代数条件一致。
    部分验证: 塌缩律只在 ρ=0.4 单一取值下成立,ρ∈{0.2,0.6} 的 F 公式里 (√c−ρ√m) 随 ρ 变化、m 敏感性可能反转,未扫;m=16 档 α 系统性抬到 2.72–2.90,地板已小到与网格步长同量级,成因(量化伪影 vs 高阶项接管)未裁决。
    待验证: 合并判据 F≤0.047 的跨 ρ 保形性;m=16 档 α 抬升的成因;会话内相关对有效自由度的折减(承接父文,另行立项)。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络(CORRESPONDENCE_NETWORK.md / EDGE_TABLE_science-yixue.md 检索「方差/区间/覆盖/膨胀/标定/自由度/df/样本量」EDGE_TABLE 零命中;CORRESPONDENCE_NETWORK 最近邻 E-061 生理注入与脾容积标定等生理学条目,均不覆盖注入法可用域×构型数命题)。

    证据等级: C
    参考文献: PMID 4005371, PMID 36725775, PMID 40119099, PMID 24338600

    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。

    Three-axis collapse into a single axis: the usable domain lower bound consistently tracks twice the statistical floor as m sweeps from 2 to 16, retiring the unified criterion of…

    Evidence level: C

    In the previous article, degrees of freedom (df) served as the second axis, and the interpretation configuration count m was fixed at 4. This article unfreezes m, sweeping it from 2 to 16, and finds that the three-axis usable domain collapses entirely into a single merged axis—the statistical floor coefficient F. The unified criterion “df≥50 and ε≥0.10” is retired, replaced by the single line “F≤0.047”.

    Body (1000-1500 words)

    The df×ε window article left a clear accounting gap: all window boundaries were swept under m=4 (four interpretation configurations per case), while m itself is a design variable. A two-injection budget-saving scheme uses m=2, whereas adding configurations for precision uses m=16; the window lower bound cannot remain at the same position. This article retains all parent models (total error T(τ₂)=ετ₂/√m+√(2/df)·(τ+c/τ−2ρ√m)/(2√m), c=1+ρ²(m−1), ρ=0.4; closed-form fixed-point iteration τ=√(v(τ)·√(2/df)/(2ε)), with the usable standard being that the deviation between closed-form point and self-consistent optimum ≤15%), and performs a three-axis grid sweep: df∈{20,30,50,100,200}×m∈{2,4,8,16}×ε step 0.0025.

    The first reading concerns the movement along the m axis. Within each df tier, the window lower bound monotonically decreases as m increases: for df=20, it drops from 0.225 at m=2 to 0.135 at m=4, 0.075 at m=8, and 0.055 at m=16; for df=50, it drops from 0.140 to 0.035. The parent article’s preset decision line states “the unified criterion holds if the window lower bound change along the m axis is <0.02 ε”—yet the actual maximum shift is 0.090, more than four times the threshold, so the unified criterion fails. The elevation of m=2 over m=4 for df=20/30/50 is +0.090/+0.075/+0.055, all exceeding the 0.05 “two-injection alone sets the line” trigger; for df=100/200 it drops to +0.040/+0.025, inside the trigger zone. Thus the tiered answers are clear: for df≤50 (tight budget), two-injection must set its own line; for df≥100 (loose zone), m=2 and m=4 can share one table.

    The second reading is this article’s true discovery: the three axes do not need tiering because they collapse. Dividing each cell’s lower bound by the statistical floor coefficient F=√(2/df)·(√c−ρ√m)/√m yields α=ε*/F: across all 15 cells with m∈{2,4,8}, α falls between 1.94 and 2.18, nearly a horizontal line. In other words, df and m—two seemingly independent knobs—act only through the product F: halving df is equivalent to changing m from 16 to 4 on the floor; the window lower bound and divergence boundary are both pinned by F alone. This is not coincidence but algebraic necessity: the floor formula inherently contains df and m as a product (√(2/df) multiplied by (√c−ρ√m)/√m), and both the closed-form iteration and optimum comparison build upon this floor, so the lower bound naturally inherits the same structure. The divergence line likewise collapses: multiplying each cell’s closed-form divergence ε upper bound by √df yields a constant (m=2 ≈0.50, m=4 ≈0.39); converting to F multiples gives 1.0–1.4 times—parent article’s “ε·√df≤0.71” diagonal becomes, along the m axis, a family of parallel diagonals whose parameter is still F.

    The collapse law has one honest boundary: the m=16 tier’s α systematically rises to 2.72–2.90, no longer tracking near 2.0. This tier’s floor F is only 0.006–0.019, comparable to the ε scan step 0.0025; the rise may be merely grid quantization artifact (step occupies one-fourth of the lower bound) or the true boundary where higher-order terms take over beyond Gaussian approximation in the first-order floor formula. Good news: the rise direction is conservative—α larger means actual usable domain is wider than the criterion promises, so executing by F will not erroneously exclude viable designs. Bad news: within F<0.01, the criterion’s “margin” cannot be trusted. Decision deferred: re-sweep m=16 with ε step refined by one order; if α falls back, it’s artifact; otherwise the first-order formula needs correction terms. Until then, operationally, the collapse law applies for F≥0.01; m=16 high-configuration schemes were never the main budget-saving route.

    This collapse directly rewrites three operational rules. First, the sample scale shifts from df axis to F axis: to preserve ε≥0.10 usable window, the unified criterion is F≤0.047 (=0.10/2.1); table lookup requires computing only one product. Second, case budget conversion gains a quantitative version: with m=4, df=50, F=0.042 meets the standard, whereas m=2 must drop to the same F requiring df≈148—two-injection needs three times the degrees of freedom to match four-configuration statistical floor, which is the algebraic content of “minimum configuration alone sets the line”. Third, sensitivity ranking verdict: the tight df end’s m axis is more sensitive than the df axis (m from 2 to 16 shifts lower bound by 0.170, exceeding df from 20 to 200 shift of 0.090 when m=4), so “case budget prioritized over calibration precision investment” holds, but with one word change—what fills first is not cases themselves but the merged floor F; adding configurations, iterations, or repeats—whichever is cheaper fills that gap; spending on ε does not expand the window.

    Measurement literature independently supports “configuration count and degrees of freedom jointly determine precision”: variance ratio confidence interval exactness depends on both df structure and balanced design (PMID 4005371); when variance is estimated, naive interval coverage systematically miscovers and must be corrected by df (PMID 36725775); reliability estimation precision varies sharply with sample size, making sample planning part of design (PMID 40119099, 24338600).

    Evidence and Limitations

    Verified: ε(m,df) full grid values computed directly (df∈{20,30,50,100,200}×m∈{2,4,8,16}×ε step 0.0025, ρ=0.4, usable=closed form vs self-consistent optimum deviation ≤15%); α=ε/F across all 15 cells with m∈{2,4,8} falls 1.94–2.18 as same scan output; divergence line ε_div·√df constant (m=2 ≈0.50, m=4 ≈0.39) matches algebraic condition k>k_crit=c/(c−ρ²m).
    Partially Verified: Collapse law holds only for ρ=0.4 single value; for ρ∈{0.2,0.6}, F formula’s (√c−ρ√m) varies with ρ and m sensitivity may reverse, unswept; m=16 tier α rises systematically to 2.72–2.90, floor shrinks to grid step scale, cause (quantization artifact vs higher-order takeover) undecided.
    Pending: Cross-ρ conformality of merged criterion F≤0.047; cause of m=16 tier α rise; session correlation reduction of effective df (carrying forward parent article, separate project).

    Related Reading

    L1 Network Status: This topic has not yet entered the corresponding L1 network (searching CORRESPONDENCE_NETWORK.md / EDGE_TABLE_science-yixue.md for “variance/interval/coverage/inflation/calibration/df/sample size” yields zero hits in EDGE_TABLE; nearest neighbors in CORRESPONDENCE_NETWORK are E-061 physiological injection and spleen volume calibration, neither covering the injection usable domain×configuration count proposition).

    Evidence Level: C
    References: PMID 4005371, PMID 36725775, PMID 40119099, PMID 24338600

    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.

  • √(σ_v/2ε) 这个最优点自己站不站得住:σ_v 随水位自洽后,闭式权衡只在中等 ε 区间可用

    Evidence level: C

    上一篇文章的闭式最优注入水位 τ₂* ≈ √(σ_v/2ε) 把 σ_v 当常数、把 ε 当宣称值;让 σ_v 按方差估计的卡方尺度族随水位自洽后,闭式只在 ε≈0.10–0.30 的中段可用(偏差 ≤11%),ε 缩到 0.05 以下直接失灵——因为统计误差有一道地板,闭式看不见。

    正文(1000-1500字)

    上一篇文章给出过一条闭式权衡:两点注入反解误差相关系数 ρ 时,第二级注入水位每升高一档,标定偏置项 ετ₂/√m 线性上涨、统计噪声项 σ_v/(2√m τ₂) 双曲下落,令两者相等就得到最优水位 τ₂* ≈ √(σ_v/(2ε))。这个式子很好用,但它脚下垫着两个没交代的假设:σ_v 被当成与水位无关的常数;ε 被当成研究者自己说了算的「宣称标定精度」。本文把这两个假设逐一拆开做自洽裁决。

    先看 σ_v。它不是常数——v 是实测方差比,而方差估计服从卡方尺度族:df 个自由度下方差的相对标准差恒为 √(2/df),绝对标准差随水位等比放大。水位 τ₂ 处 v 的期望约为 1+ρ²(m−1)−2ρτ₂√m+τ₂²,于是 σ_v(τ₂) ≈ v(τ₂)·√(2/df)。代回总误差 T(τ₂) = ετ₂/√m + σ_v(τ₂)/(2√m τ₂),统计项不再单纯随 τ₂ 下落:v/τ₂ = τ₂ + c/τ₂ − 2ρ√m(其中 c = 1+ρ²(m−1))在 τ₂ = √c 处触底后回升——「锚加得越大统计越准」的单调直觉在这里死掉了,统计误差有一道地板。

    数值扫描把后果算清楚了。沿用父文参数(m=4、ρ=0.4、df=100、c=1.48),20 万点网格扫 ε∈[0.01, 0.30],比较「σ_v 当常数的闭式 τ₂」与「自洽 σ_v 的数值最优 τ₂」。结果分三段:ε=0.10–0.30 的中段,两者偏差仅 −1.5% 到 −11%,闭式可当设计初值;ε=0.05 时偏差拉大到 27%;ε=0.01–0.03 时闭式直接失灵——定点迭代下 τ₂* 飙到 6.6–11.4,而自洽最优稳稳停在 1.02–1.14。原因正是统计地板:闭式以为 ε 越小越该加大锚去换统计精度,但自洽统计项的地板 √(2/df)·(√c−ρ√m)/√m ≈ 0.03 压在那里,锚再大也穿不透,继续加锚只会让标定项接管误差曲线——「最优点」恰好挪到总误差爬坡的位置。

    再看 ε 宣称×2 的情形。按宣称 ε 设计、真实 ε 翻倍时,总误差只放大 1.02–1.15 倍,远低于预设判据的 1.5 倍熔断线——这是「平底」救的场:T(τ₂) 在最优点附近是平缓的 U 形,点位放偏 20–30% 误差只涨几个百分点,标定精度的宣称误差对「锚放哪里」相当宽容。但点位稳健不等于推断有效:报给读者的可识别区间宽度直接正比于真实 ε,ε 被低估一倍、区间就窄一倍,覆盖率损失一分不少。测量学文献里同样的结构有现成教训:方差比的置信区间要用专用方法构造才有精确覆盖(PMID 4005371);把方差不确度当已知量的朴素区间,在方差本身被估计时覆盖率就不达标(PMID 36725775);GUM 不确定度框架与蒙特卡洛方法在非线性、输出分布不对称的场合会给出不同答案、后者更可信(PMID 21238753、24659835),被测量越接近零非对称越严重、甚至冒出负下界这种非物理结果(PMID 27114218);而不确定度分量合成时每一层——包括标定修正本身的不确定度——都必须计入(PMID 22896744)。

    敏感性抽查把这幅图的边界又推了一把。ρ 降到 0.2 时,ε=0.10 的闭式偏差从 −1.5% 恶化到 +59%,闭式点处的总误差已是最优点的 1.13 倍;df 从 100 压到 30 时,ε=0.10 偏差恶化到 +93%、总误差放大 1.36 倍;两者叠加到 ε=0.05 时,闭式点直接落在标定爬坡段上,总误差是最优点的 3.1–4.1 倍——「平底」的保护不是无条件的,ρ 越小、df 越小,统计地板越高、U 形越窄,闭式的失效带就越往中段侵吞。反过来 ρ=0.6 时中段闭式偏差收窄到 ±11% 以内,点位几乎无损。换言之,「中段可用」不是 ε 一根轴上的事,而是 ε×ρ×df 三根轴围出的一块窗口:ρ 越小、自由度越紧,窗口越窄,闭式能撑住的 ε 下限越高。

    落到裁决:两条预设判据各砍一半。判据一「ε≤0.10 偏差<15% ⇒ 闭式可当初值」方向对、边界反——可用区在 ε≥0.10 的中段,ε 越小越不可用;判据二「ε×2 放大>1.5× ⇒ 须第三方复核」在本参数化下不触发,点位设计免强制第三方复核,但区间报告端必须用经核验的 ε。操作守则三条:① ε≤0.05 的「高精度宣称」反而最该警惕——那正是闭式最不可信的区域,此时应对自洽总误差曲线直接做数值求最小,别信闭式;② ε 落在 0.10–0.20 时闭式当初值、数值收尾;③ 无论哪档,ε 本身的核验等级要写进报告,因为区间宽窄取决于它而不是点位。

    证据与局限

    已验证: 自洽总误差曲线与「中段可用/小 ε 失灵」三段分档为 20 万点网格数值扫描结果(m=4、ρ=0.4、df=100);ε 宣称×2 的总误差放大 1.02–1.15 倍为同一扫描输出;统计项地板 v/τ₂ 在 τ₂=√c 处取极小为代数推导。
    部分验证: 统计地板闭式 √(2/df)·(√c−ρ√m)/√m 依赖一阶近似与高斯假设;ρ∈{0.2, 0.4, 0.6} × df∈{30, 50, 100} 的敏感性抽查已完成、全网格未扫。
    待验证: df 更小时地板抬高、中段可用窗口是否收窄;标定误差取加性口径时自洽解须重推;第三方核验 ε 的成本收益与区间膨胀因子的校准曲线。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络(CORRESPONDENCE_NETWORK.md / EDGE_TABLE_science-yixue.md 检索「方差比/自洽/卡方/σ_v/最优水位/权衡/注入/标定」,仅命中 E-061 生理注入与脾容积标定等生理学条目,无「注入法最优水位自洽性」对应边)。

    证据等级: C
    参考文献: PMID 4005371, PMID 36725775, PMID 21238753, PMID 24659835, PMID 27114218, PMID 22896744

    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。

    Whether the optimal point √(σ_v/2ε) stands on its own: closed-form trade-offs are only viable for moderate ε after σ_v becomes self-consistent with water level

    Evidence level: C

    上一篇文章的闭式最优注入水位 τ₂* ≈ √(σ_v/2ε) 把 σ_v 当常数、把 ε 当宣称值;让 σ_v 按方差估计的卡方尺度族随水位自洽后,闭式只在 ε≈0.10–0.30 的中段可用(偏差 ≤11%),ε 缩到 0.05 以下直接失灵——因为统计误差有一道地板,闭式看不见。

    正文(1000-1500字)

    上一篇文章给出过一条闭式权衡:两点注入反解误差相关系数 ρ 时,第二级注入水位每升高一档,标定偏置项 ετ₂/√m 线性上涨、统计噪声项 σ_v/(2√m τ₂) 双曲下落,令两者相等就得到最优水位 τ₂* ≈ √(σ_v/(2ε))。这个式子很好用,但它脚下垫着两个没交代的假设:σ_v 被当成与水位无关的常数;ε 被当成研究者自己说了算的「宣称标定精度」。本文把这两个假设逐一拆开做自洽裁决。

    先看 σ_v。它不是常数——v 是实测方差比,而方差估计服从卡方尺度族:df 个自由度下方差的相对标准差恒为 √(2/df),绝对标准差随水位等比放大。水位 τ₂ 处 v 的期望约为 1+ρ²(m−1)−2ρτ₂√m+τ₂²,于是 σ_v(τ₂) ≈ v(τ₂)·√(2/df)。代回总误差 T(τ₂) = ετ₂/√m + σ_v(τ₂)/(2√m τ₂),统计项不再单纯随 τ₂ 下落:v/τ₂ = τ₂ + c/τ₂ − 2ρ√m(其中 c = 1+ρ²(m−1))在 τ₂ = √c 处触底后回升——「锚加得越大统计越准」的单调直觉在这里死掉了,统计误差有一道地板。

    数值扫描把后果算清楚了。沿用父文参数(m=4、ρ=0.4、df=100、c=1.48),20 万点网格扫 ε∈[0.01, 0.30],比较「σ_v 当常数的闭式 τ₂」与「自洽 σ_v 的数值最优 τ₂」。结果分三段:ε=0.10–0.30 的中段,两者偏差仅 −1.5% 到 −11%,闭式可当设计初值;ε=0.05 时偏差拉大到 27%;ε=0.01–0.03 时闭式直接失灵——定点迭代下 τ₂* 飙到 6.6–11.4,而自洽最优稳稳停在 1.02–1.14。原因正是统计地板:闭式以为 ε 越小越该加大锚去换统计精度,但自洽统计项的地板 √(2/df)·(√c−ρ√m)/√m ≈ 0.03 压在那里,锚再大也穿不透,继续加锚只会让标定项接管误差曲线——「最优点」恰好挪到总误差爬坡的位置。

    再看 ε 宣称×2 的情形。按宣称 ε 设计、真实 ε 翻倍时,总误差只放大 1.02–1.15 倍,远低于预设判据的 1.5 倍熔断线——这是「平底」救的场:T(τ₂) 在最优点附近是平缓的 U 形,点位放偏 20–30% 误差只涨几个百分点,标定精度的宣称误差对「锚放哪里」相当宽容。但点位稳健不等于推断有效:报给读者的可识别区间宽度直接正比于真实 ε,ε 被低估一倍、区间就窄一倍,覆盖率损失一分不少。测量学文献里同样的结构有现成教训:方差比的置信区间要用专用方法构造才有精确覆盖(PMID 4005371);把方差不确度当已知量的朴素区间,在方差本身被估计时覆盖率就不达标(PMID 36725775);GUM 不确定度框架与蒙特卡洛方法在非线性、输出分布不对称的场合会给出不同答案、后者更可信(PMID 21238753、24659835),被测量越接近零非对称越严重、甚至冒出负下界这种非物理结果(PMID 27114218);而不确定度分量合成时每一层——包括标定修正本身的不确定度——都必须计入(PMID 22896744)。

    敏感性抽查把这幅图的边界又推了一把。ρ 降到 0.2 时,ε=0.10 的闭式偏差从 −1.5% 恶化到 +59%,闭式点处的总误差已是最优点的 1.13 倍;df 从 100 压到 30 时,ε=0.10 偏差恶化到 +93%、总误差放大 1.36 倍;两者叠加到 ε=0.05 时,闭式点直接落在标定爬坡段上,总误差是最优点的 3.1–4.1 倍——「平底」的保护不是无条件的,ρ 越小、df 越小,统计地板越高、U 形越窄,闭式的失效带就越往中段侵吞。反过来 ρ=0.6 时中段闭式偏差收窄到 ±11% 以内,点位几乎无损。换言之,「中段可用」不是 ε 一根轴上的事,而是 ε×ρ×df 三根轴围出的一块窗口:ρ 越小、自由度越紧,窗口越窄,闭式能撑住的 ε 下限越高。

    落到裁决:两条预设判据各砍一半。判据一「ε≤0.10 偏差<15% ⇒ 闭式可当初值」方向对、边界反——可用区在 ε≥0.10 的中段,ε 越小越不可用;判据二「ε×2 放大>1.5× ⇒ 须第三方复核」在本参数化下不触发,点位设计免强制第三方复核,但区间报告端必须用经核验的 ε。操作守则三条:① ε≤0.05 的「高精度宣称」反而最该警惕——那正是闭式最不可信的区域,此时应对自洽总误差曲线直接做数值求最小,别信闭式;② ε 落在 0.10–0.20 时闭式当初值、数值收尾;③ 无论哪档,ε 本身的核验等级要写进报告,因为区间宽窄取决于它而不是点位。

    证据与局限

    已验证: 自洽总误差曲线与「中段可用/小 ε 失灵」三段分档为 20 万点网格数值扫描结果(m=4、ρ=0.4、df=100);ε 宣称×2 的总误差放大 1.02–1.15 倍为同一扫描输出;统计项地板 v/τ₂ 在 τ₂=√c 处取极小为代数推导。
    部分验证: 统计地板闭式 √(2/df)·(√c−ρ√m)/√m 依赖一阶近似与高斯假设;ρ∈{0.2, 0.4, 0.6} × df∈{30, 50, 100} 的敏感性抽查已完成、全网格未扫。
    待验证: df 更小时地板抬高、中段可用窗口是否收窄;标定误差取加性口径时自洽解须重推;第三方核验 ε 的成本收益与区间膨胀因子的校准曲线。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络(CORRESPONDENCE_NETWORK.md / EDGE_TABLE_science-yixue.md 检索「方差比/自洽/卡方/σ_v/最优水位/权衡/注入/标定」,仅命中 E-061 生理注入与脾容积标定等生理学条目,无「注入法最优水位自洽性」对应边)。

    证据等级: C
    参考文献: PMID 4005371, PMID 36725775, PMID 21238753, PMID 24659835, PMID 27114218, PMID 22896744

    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


    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.

  • 点位稳健救不了区间:用宣称 ε 报的可识别区间,在真实 ε 翻倍时覆盖率掉到多少、要配多大的膨胀因子才保住名义覆盖

    Evidence level: C

    按宣称标定精度报的区间,真实标定精度差一倍时覆盖率只剩六到八成;想用一个只看倍数的膨胀因子把覆盖拉回来,仿真判死——因子必须携带误差预算构成信息,而这恰好是设计者手里本来就有的量。

    正文(1000-1500字)

    上一篇文章裁决过一件事:标定精度 ε 宣称值被低估一倍时,最优注入水位的点位设计几乎不受伤——总误差只放大 1.02 到 1.15 倍,靠的是总误差曲线在最优附近一道平缓的「平底」。但点位稳健不等于推断有效:报给读者的可识别区间宽度直接正比于宣称 ε,区间端不设防。本文把「不设防」量化成两个数:真实覆盖率到底掉到多少?补一个膨胀因子 k 能不能救?

    仿真沿用父文参数化:判读构型数 m=4,误差相关系数 ρ 取 0.2、0.4、0.6,方差比的自由度 df 取 30、100,宣称精度 ε_claim 取 0.10、0.20,注入水位取自洽最优。每格十万次模拟:总偏差由两部分相加,标定部分按真实精度生成(真实与宣称之比 r 从 1 扫到 3),统计部分按方差比的卡方尺度族逐次抽取——也就是说统计端的不确度是估出来的、区间里已经算了它,裸奔的只有被低估的标定端。报告区间一律按宣称 ε 构造,名义覆盖 95%。共一百零八个格子。

    先看不补因子的下场。r=2 时覆盖率只剩 0.68 到 0.85,r=3 时掉到 0.49 到 0.72——比「区间窄一倍」的直觉更狠。掉多少由误差预算构成决定:统计项占得越多的格子掉得越少,因为统计端受保护、只有标定端裸奔。ρ=0.2、df=30、ε_claim=0.10 的格子里统计项占七成,r=2 时覆盖还剩 0.845;ρ=0.6、df=100、ε_claim=0.20 的格子里统计项只占百分之一点四,r=2 时覆盖 0.676——几乎分毫不差地等于纯标定被低估一倍时的理论值。区间失真的账本和点位稳健的账本是两本,构成不同,罚也不同。

    再看补救。每格二分求「把覆盖拉回 95% 所需的膨胀因子」,结果显示 k 除了随倍数比 r 增长之外,还强烈依赖误差构成:r=2 时 k 跨格从 1.38 到 2.00,离散度随 r 爬升,r=1.25 时约 15%,r=3 时达 60%。取跨格中位数拟一条「单一 k 曲线」套回全部格子:中位覆盖确实回到 0.950,但一百零八格里五十五个越出验收带,而且错的方向是过覆盖——统计项占大头的格子被吹到 0.99,区间白白宽了三成,用精度损失换名义合规。单一 k 倍数曲线,判死。

    但死亡证书上写了活路。把每格的误差预算构成记作 w:标定项方差占总方差的比例,它是设计期就能算出来的量——标定项是宣称 ε 与水位的函数,统计项是自由度的函数,两者都不用等数据。k 的全部跨格变异被一个闭式吃掉:k 等于「r 方乘 w 加一减 w」再开方。代回仿真,全格点覆盖率 0.944 到 0.952,中位 0.950,偏离名义不超过零点八个百分点。换句话说,保序膨胀可行,但因子必须带构成信息——而这不需要迭代求解 ρ,w 是报告者手里已有的账。

    落到裁决,两条预设判据各改一半。判据一「单一 k 倍数曲线全格点拉回」不成立,维持死刑;判据二「k 必须带 ρ 估计迭代」过重,改判为「k 必须附带误差预算构成 w,一次换算即可」。守则三条:第一,报告可识别区间必须附 w;统计项占半壁以上的格子禁用任何只看倍数的统一膨胀——那种格子里统一乘 r 会把区间吹宽三成以上。第二,倍数比 r 的核验等级与 w 一起写进报告,覆盖率的承诺只在两者都成立时兑现。第三,w 算不出来的场合(标定误差与统计误差相关、或标定端重尾),闭式失效,退回第三方复核,别硬套。测量学文献对此早有平行教训:被测量接近零时区间会冒出非物理的负下界、GUM 框架必须让位给蒙特卡洛(PMID 27114218);展开不确定度的覆盖验收用蒙特卡洛比解析传播更可信(PMID 21238753);方差比这类量的区间要用专用构造才有精确覆盖(PMID 4005371);方差本身被估计时,朴素区间失准、精确方法才达标(PMID 36725775)。区间这件事上,「点位稳」从来不是免死金牌,账本透明才是。

    证据与局限

    已验证: 未校正覆盖率坍塌(r=2 时 0.68-0.85、r=3 时 0.49-0.72)、k 跨格离散度随 r 从 15% 爬到 60%、单一中位数曲线套用 55/108 格越界(且以过覆盖为主)、闭式 k(r,w) 全格点覆盖 0.944-0.952——均为一百零八格×每格十万次、固定种子的仿真结果。
    部分验证: 模型假定标定误差与统计误差独立且高斯;统计项逐次抽取已计入,但换算式中的构成 w 用统计项中位数代入,带来不超过 0.8 个百分点的覆盖偏差。
    待验证: 标定误差与统计误差相关时闭式须重推;标定端重尾分布下 k(r,w) 的稳健性;w 本身的测量误差对两遍法承诺的侵蚀幅度。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络(EDGE_TABLE_science-yixue.md 检索「方差/区间/覆盖/膨胀/标定」零命中;CORRESPONDENCE_NETWORK.md 最近邻为 E-060 标签信度边【conflict 状态,但属「五脏标签分组」命题,不覆盖本篇「区间覆盖率×膨胀因子」】与 E-061 0.1Hz 生理注入边,均无对应关系)。

    证据等级: C
    参考文献: PMID 27114218, PMID 21238753, PMID 4005371, 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。

    Pointwise Robustness Cannot Save Interval Coverage: When the True ε Doubles, How Much Does Coverage Drop and What Expansion Factor Is Needed to Preserve Nominal Coverage?

    Evidence level: C

    When an identifiable interval is declared using a claimed calibration precision ε, the true coverage drops to between 68% and 85% if the actual calibration precision is twice the claimed value. Attempting to restore coverage using an inflation factor k that depends solely on this ratio fails; simulations show such a single-factor approach is untenable. The inflation factor must instead incorporate the error budget composition, which is precisely the quantity available to the designer.

    Body (1000–1500 words)

    A previous article ruled on a specific matter: when the claimed calibration precision ε is underestimated by a factor of two, the optimal injection level design for points remains remarkably robust—the total error increases only by a factor of 1.02 to 1.15, relying on the flat “bottom” of the total error curve near the optimum. However, point robustness does not equate to inference validity: the width of the identifiable interval reported to readers is directly proportional to the claimed ε, with no guardrails at the interval endpoints. This article quantifies that “no guardrails” into two numbers: how far does the true coverage drop? Can an inflation factor k restore it?

    The simulation follows the parameterization of the parent text: judgment configuration count m=4, error correlation coefficient ρ takes values 0.2, 0.4, 0.6, degrees of freedom df for the variance ratio takes 30, 100, claimed precision ε_claim takes 0.10, 0.20, and injection levels are drawn from the self-consistent optimum. Each grid involves 100,000 simulations: total deviation consists of two parts added together—the calibration part is generated according to true precision (ratio r from 1 sweeping to 3), and the statistical part is sequentially drawn via the chi-square scale family of variance ratios. In other words, uncertainty on the statistical side is estimated and already accounted for within the interval; only the underestimated calibration side runs naked. All reported intervals are constructed using the claimed ε, yielding nominal coverage of 95%. There are 108 grids in total.

    First, consider the outcome without adding an inflation factor. When r=2, coverage drops to between 0.68 and 0.85; when r=3, it falls to between 0.49 and 0.72—far more severe than the intuition that “the interval is half as wide.” How much it drops depends on the error budget composition: grids where the statistical component occupies a larger share drop less, because the statistical side is protected while only the calibration side runs naked. In the grid with ρ=0.2, df=30, ε_claim=0.10, the statistical component accounts for 70% of the error; when r=2, coverage remains at 0.845. In contrast, in the grid with ρ=0.6, df=100, ε_claim=0.20, the statistical component accounts for only 1.4%; when r=2, coverage drops to 0.676—almost exactly matching the theoretical value for pure calibration underestimated by a factor of two. The ledger of interval distortion and the ledger of point robustness are two separate accounts with different structures, leading to different penalties.

    Next, consider the remedy. For each grid, we compute in binary form the inflation factor k required to restore coverage to 95%. The results show that besides increasing with the ratio r, k strongly depends on error composition: when r=2, k varies across grids from 1.38 to 2.00, with dispersion rising as r increases—about 15% when r=1.25 and reaching 60% when r=3. Fitting a “single k curve” using the median across grids and applying it to all grids: indeed, median coverage returns to 0.950, but 55 out of 108 grids exceed the acceptance band, and they overshoot—grids where the statistical component dominates are inflated to 0.99, unnecessarily widening the interval by 30%, trading precision loss for nominal compliance. A single k ratio curve is ruled out.

    Yet the death certificate includes a path to life. Let w denote the error budget composition for each grid: the proportion of total variance occupied by the calibration term, which can be computed during design—the calibration term is a function of claimed ε and injection level, while the statistical term depends on degrees of freedom; neither requires waiting for data. All variation in k across grids is consumed by a closed-form expression: k equals the square root of (r² × w + 1 − w). Substituting back into simulations, full-grid coverage ranges from 0.944 to 0.952, median at 0.950, deviating from nominal by no more than 0.8 percentage points. In other words, ordered inflation is feasible, but the factor must carry composition information—this does not require iterative estimation of ρ, since w is already available in the report.

    To summarize the ruling, both preset criteria are revised by half. Criterion one—”a single k ratio curve restores all grids”—is ruled out, maintaining the death sentence. Criterion two—”k must include iterative estimation of ρ”—is deemed too heavy; it is revised to “k must carry the error budget composition w, allowing a single conversion.” Three rules follow: First, reported identifiable intervals must include w; grids where the statistical component exceeds half must not use any uniform inflation based solely on ratio—multiplying uniformly by r would inflate such intervals by more than 30%. Second, the verification level of ratio r must be written into the report alongside w; coverage commitment is fulfilled only when both hold. Third, where w cannot be computed (e.g., calibration and statistical errors are correlated, or the calibration end has heavy tails), the closed form fails; revert to third-party review rather than forcing application. Measurement literature has long provided parallel lessons: when the measured quantity approaches zero, intervals may produce non-physical negative lower bounds, so the GUM framework must yield to Monte Carlo (PMID 27114218); for coverage acceptance of expanded uncertainty, Monte Carlo is more credible than analytical propagation (PMID 21238753); intervals for quantities like variance ratios require specialized construction for accurate coverage (PMID 4005371); when variance itself is estimated, naive intervals are inaccurate and exact methods are required (PMID 36725775). Regarding intervals, “point stability” is never a get-out-of-jail card; transparent accounting is essential.

    Evidence and Limitations

    Verified: Uncorrected coverage collapse (0.68–0.85 at r=2, 0.49–0.72 at r=3), k dispersion across grids rising from 15% to 60% with r, median-only curve causing 55/108 grids to overshoot (primarily over-coverage), and closed-form k(r,w) achieving full-grid coverage 0.944–0.952—all based on 108 grids × 100,000 simulations per grid with fixed seeds.
    Partially Verified: Model assumes calibration error and statistical error are independent and Gaussian; sequential drawing of the statistical component is accounted for, but substituting the median of the statistical component into w in the conversion formula introduces up to 0.8 percentage point coverage deviation.
    To Be Verified: When calibration and statistical errors are correlated, the closed form must be re-derived; robustness of k(r,w) under heavy-tailed calibration distributions; and the erosion of the two-pass commitment due to measurement error in w itself.

    Related Reading

    L1 Network Status: This topic has not yet entered the corresponding L1 network (EDGE_TABLE_science-yixue.md yields zero hits for “variance/interval/coverage/inflation/calibration”; CORRESPONDENCE_NETWORK.md has nearest neighbors E-060 label credibility edge [conflict status, but belongs to the “five-organs label group” proposition, not covering this article’s “interval coverage × inflation factor”] and E-061 0.1Hz physiological injection edge, neither matching).

    Evidence Level: C
    References: PMID 27114218, PMID 21238753, PMID 4005371, 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.

  • 自由度这张底牌:df 从 200 压到 20,中段可用窗口从 0.24 收窄到 0.14,df 不够时闭式直接出局

    Evidence level: C

    上一篇文章的「中段可用」是在 df=100 下扫出来的;把方差比自由度从 200 一路压到 20,统计地板按 √(2/df) 抬升 2.2 倍,可用窗口下界从 ε=0.05 挪到 0.16——df≤50 且 ε=0.10 时闭式解干脆不存在,df≥50 与 ε≥0.10 必须同格申报。

    正文(1000-1500字)

    自洽 σ_v 那篇文章留下一个悬案:三段分档是 df=100 下扫出来的,而真实注入实验的方差比自由度常常吃紧——m 例判读、每例两三遍,df 可能只有 20 到 50。自由度一紧,√(2/df) 这道统计地板随之上抬,中段可用窗口会不会整个关门?本文把 df 当成第二根轴,做一轮 df×ε 二维扫描。

    先把地板的尺度钉死。方差估计服从卡方尺度族,相对标准差恒为 √(2/df)。沿用父文参数(m=4、ρ=0.4、c=1.48),统计项地板 √(2/df)·(√c−ρ√m)/√m 随 df 下降按反比爬升:df=200 时 0.021,df=100 时 0.029,df=50 时 0.042,df=30 时 0.054,df=20 时 0.066——df 从 100 砍到 20,地板翻 2.2 倍。地板抬高不改变总误差曲线的 U 形形状,但它把曲线最小值整体顶起来,同时让闭式 τ₂* ≈ √(σ_v/(2ε)) 的前提更早崩塌:这个式子要有解,需要 2ε 大过地板系数 √(2/df),两者在 df 轴上直接对撞。

    网格扫描把「窗口」画了出来。以「闭式点位与自洽最优点的偏差不超过 15%」为可用标准,各 df 档的可用 ε 区间是:df=200 为 [0.05, 0.30],df=100 为 [0.07, 0.30],df=50 为 [0.10, 0.30],df=30 收窄到 [0.13, 0.30],df=20 只剩 [0.16, 0.30]。三条读数值得记住。其一,窗口下界随 df 单调上抬,df 每收紧一档,下限抬高约 0.03 个 ε,从 df=200 的 0.05 一路挪到 df=20 的 0.16。其二,窗口宽度从 0.24 缩到 0.14,收紧近一半。其三,也是最难看的一条:在 ε=0.10 这个父文判据的中线点上,df=50 及以下闭式解根本不存在——定点迭代在 2ε ≤ √(2/df) 时发散,这正是父文「df=100、ε≤0.03 时 τ₂* 飙到 6.6 到 11.4」在 df 轴上的镜像:不降 ε、只压 df,发散照样发生。所以预设判据里「df=30、ε=0.10 偏差超 25%」这一句在实算里还要加重——不是偏差超线,是闭式直接出局。

    这条发散条件值得单独说一句,因为它是一条对角线而不是一条竖线。2ε ≤ √(2/df) 整理后等价于 ε·√df ≤ 0.71:ε 与 df 在这条线上互换角色,降 ε 砍窗口、压 df 同样砍窗口,两条路通向同一个死点。而真实实验的 df 恰恰最容易在这条线附近吃紧:判读例数受招募与预算限制,每例遍数受判读疲劳限制,预实验阶段 df 掉到 20 到 30 是常态而非例外;更隐蔽的是脱落——设计的 30 例走掉 5 例,df 就从 60 掉到 50,窗口下界立刻从 0.10 抬到 0.13,原定的 ε=0.10 设计瞬间出窗。也就是说,df 不该被当成设计书里随手填的一个背景数,它和 ε 一样是要申报、要复核、要留余量的设计变量;预算紧的时候,宁可砍 ε 的宣称精度,也不要把 df 压到窗口外,因为前者伤的是区间宽度,后者伤的是整个点位设计的可用性。

    统计项在总误差里的占比解释了地板为什么不能只看绝对值。最优点处统计项占总误差 43% 到 72%:df=20、ε=0.05 时占 72%,df=200、ε=0.20 时占 43%。df 越紧、ε 越小,误差预算里越大的一块被统计地板锁死,标定端宣称得再好也穿不透。这直接给例数预算定了线:判据「df≥50 窗口保持开放」方向成立,但要加边界——df=50 的窗口从 ε=0.10 起步,低于该值无窗。按 df 约等于例数乘构型的口径反推,两点注入、每例两遍的构型下,df≥50 需要约 25 例判读;每例三遍则约 17 例。测量学文献对同一结构有独立支撑:方差比的精确置信区间依赖自由度结构,不能用朴素口径(PMID 4005371);方差本身被估计时,把不确度当已知量的朴素区间覆盖率系统性失准(PMID 36725775);信度估计的精度随样本量剧烈变化,样本量规划是设计的一部分而非事后事项(PMID 40119099、24338600)。

    落到裁决与守则。父文判据一须改写为:中段可用窗口是 ε×df 二维区域,下界随 df 上抬,df≥50 与 ε≥0.10 须同时声明、缺一不可;判据二(ε 宣称翻倍点位放大不超 1.5 倍)在全部 df 档下不触发,区间报告端口径不变。操作守则四条:设计书里 df 与 ε 同格申报,凡报 ε≤0.05 且 df<100 的宣称一律数值求最小、闭式禁用;df<50 而 ε<0.13 的方案视为窗口未开放,判读结果只能作探索性表述;例数预算按「df≥50 或 ε≥0.16」反推最低判读例数;报告模板新增 df 字段,跨研究比较只在同 df 档内进行。这与父文「ε 越小越该警惕」是同一条教训在 df 轴上的复现:资源紧张时先砍例数、先紧自由度,而闭式最不可信的恰恰是这一档。

    证据与局限

    已验证: 统计地板 √(2/df)·(√c−ρ√m)/√m 在五个 df 档的数值为直接代入计算;可用窗口边界(闭式偏差 ≤15% 的 ε 区间)为 df×ε 网格实算结果(m=4、ρ=0.4、c=1.48,df∈{20,30,50,100,200}、ε 步长 0.005);df≤50 且 ε=0.10 时闭式无解为 2ε≤√(2/df) 的代数条件。
    部分验证: 窗口边界只在 ρ=0.4 单一取值下扫描,ρ∈{0.2, 0.6} 的 df×ε 等高线未扫;df 与判读例数的换算按「例数×构型」粗口径,未含会话内相关对有效自由度的折减。
    待验证: m(判读构型数)轴上窗口边界如何移动,三轴联合的完整可用域;给定目标总误差反解最小 df 的查用表;会话内相关性对有效自由度的折减系数标定。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络(CORRESPONDENCE_NETWORK.md / EDGE_TABLE_science-yixue.md 检索「方差比/自由度/自洽/卡方/σ_v/最优水位/权衡/地板」,仅命中 E-061 生理注入与脾容积标定等生理学条目,无「注入法窗口×自由度」对应边)。

    证据等级: C
    参考文献: PMID 4005371, PMID 36725775, PMID 40119099, PMID 24338600

    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。

    The ace of degrees of freedom: compressing df from 200 to 20 shrinks the mid-range usable window from 0.24 to 0.14, forcing immediate exit when df is insufficient

    Evidence level: C

    The “mid-range usable” window in the previous article was scanned under df=100; compressing the variance ratio degrees of freedom from 200 down to 20 raises the statistical floor by √(2/df) by a factor of 2.2, shifting the lower bound of the usable window from ε=0.05 to 0.16—when df≤50 and ε=0.10, closed-form solutions simply do not exist, and df≥50 with ε≥0.10 must be declared jointly.

    Main Body (1000-1500 words)

    The self-consistent σ_v article left an outstanding question: the three-tier classification was scanned under df=100, but real injection experiments often face tight variance ratio degrees of freedom—m cases judged, each case two to three passes, df possibly only 20 to 50. When degrees of freedom tighten, √(2/df) this statistical floor rises accordingly; will the mid-range usable window close entirely? This paper treats df as the second axis, performing a df×ε two-dimensional scan.

    First, fix the scale of the floor. Variance estimation follows the chi-square scale family, with constant relative standard deviation √(2/df). Using parent article parameters (m=4, ρ=0.4, c=1.48), the statistical term floor √(2/df)·(√c−ρ√m)/√m rises inversely as df decreases: at df=200 it is 0.021, at df=100 it is 0.029, at df=50 it is 0.042, at df=30 it is 0.054, at df=20 it is 0.066—reducing df from 100 to 20 raises the floor by 2.2 times. Raising the floor does not change the U-shape of the total error curve, but it lifts the entire minimum curve upward while causing the closed-form τ₂* ≈ √(σ_v/(2ε)) premise to collapse earlier: this formula requires a solution only if 2ε exceeds the floor coefficient √(2/df), and these two quantities collide directly on the df axis.

    The grid scan draws out the “window.” Using “closed-form point deviation from self-consistent optimum not exceeding 15%” as the usable criterion, the usable ε intervals for each df tier are: df=200 [0.05, 0.30], df=100 [0.07, 0.30], df=50 [0.10, 0.30], df=30 narrowed to [0.13, 0.30], df=20 only [0.16, 0.30]. Three readings are worth remembering. First, the window lower bound rises monotonically with df tightening; each df tier contraction lifts the lower limit by approximately 0.03 ε, moving from df=200’s 0.05 all the way to df=20’s 0.16. Second, window width shrinks from 0.24 to 0.14, nearly halved. Third, and most troubling: at ε=0.10, the mid-point of the parent article criterion, closed-form solutions for df≤50 simply do not exist—fixed-point iteration diverges when 2ε ≤ √(2/df), precisely mirroring the parent article’s “df=100, ε≤0.03 causes τ₂* to spike from 6.6 to 11.4” on the df axis: without reducing ε, merely compressing df still causes divergence. Therefore, in real calculation, the preset criterion “df=30, ε=0.10 deviation exceeds 25%” must be strengthened—not exceeding the line, but closed-form directly exiting.

    This divergence condition deserves special mention because it is a diagonal rather than a vertical line. Rearranging 2ε ≤ √(2/df) yields ε·√df ≤ 0.71: ε and df swap roles along this line, reducing ε shrinks the window, compressing df also shrinks the window, both paths leading to the same dead end. Real experimental df is precisely most likely to be tight near this line: case numbers constrained by recruitment and budget, passes per case limited by judgment fatigue, pre-experimental phase df dropping to 20–30 is normal rather than exceptional; more subtly, attrition—designed 30 cases losing 5, df drops from 60 to 50, window lower bound immediately rises from 0.10 to 0.13, original ε=0.10 design instantly falls out of window. Thus df should not be treated as a background number casually filled in the design document; like ε, it must be declared, reviewed, and reserved with margin. When budget is tight, cut ε’s claimed precision rather than compress df out of the window, because the former harms interval width while the latter destroys overall point design usability.

    The proportion of statistical terms in total error explains why the floor cannot be judged by absolute value alone. At optimum, statistical terms account for 43% to 72% of total error: at df=20, ε=0.05 it accounts for 72%, at df=200, ε=0.20 it accounts for 43%. The tighter df and smaller ε, the larger portion of the error budget locked by the statistical floor, making calibration-end claims no matter how well-stated fail to penetrate. This directly sets the line for case budget: the criterion “df≥50 keeps window open” holds directionally but requires boundary—df=50’s window starts at ε=0.10, below which no window exists. Inverting via df≈cases×configuration, under two-point injection with two passes per case, df≥50 requires approximately 25 cases; three passes per case then approximately 17. Measurement literature provides independent support for the same structure: exact confidence intervals of variance ratios depend on degrees of freedom structure, cannot use naive bandwidth (PMID 4005371); when variance itself is estimated, treating uncertainty as known causes systematic undercoverage of naive interval coverage (PMID 36725775); reliability estimation precision varies dramatically with sample size, so sample size planning is part of design rather than post-hoc matter (PMID 40119099, 24338600).

    Down to verdict and rules. Parent article criterion one must be rewritten as: mid-range usable window is the ε×df two-dimensional region, lower bound rises with df, df≥50 and ε≥0.10 must be declared jointly, neither missing; criterion two (ε claimed doubling point amplification not exceeding 1.5×) does not trigger across all df tiers, interval report bandwidth unchanged. Operational rules four: declare df and ε jointly in design document; any claim with ε≤0.05 and df<100 must numerically minimize, closed-form disabled; schemes with df<50 and ε<0.13 deemed window not open, judgment results only exploratory expression; case budget inverted from “df≥50 or ε≥0.16” to minimum judgment cases; report template adds df field, cross-study comparison only within same df tier. This mirrors the parent article’s “smaller ε warrants more caution” lesson on the df axis: when resources are tight, cut cases first, tighten degrees of freedom first, while closed-form is precisely the least trustworthy tier.

    Evidence and Limitations

    Verified: Statistical floor √(2/df)·(√c−ρ√m)/√m values across five df tiers computed by direct substitution; usable window boundaries (closed-form deviation ≤15% ε intervals) derived from df×ε grid real calculation (m=4, ρ=0.4, c=1.48, df∈{20,30,50,100,200}, ε step 0.005); closed-form unsolvable when df≤50 and ε=0.10 confirmed by algebraic condition 2ε≤√(2/df).
    Partially Verified: Window boundaries scanned only at single ρ=0.4 value; df×ε contour lines for ρ∈{0.2, 0.6} not scanned; df-to-case conversion uses coarse “cases×configuration” bandwidth, omitting intra-session correlation reduction of effective degrees of freedom.
    To Be Verified: How window boundaries shift along m (judgment configuration count) axis, complete usable domain under three-axis joint; lookup table for minimum df inversion given target total error; intra-session correlation coefficient calibration for effective degrees of freedom reduction.

    Related Reading

    L1 Network Status: This topic has not yet entered the corresponding L1 network (search CORRESPONDENCE_NETWORK.md / EDGE_TABLE_science-yixue.md for “variance ratio/degrees of freedom/self-consistency/chi-square/σ_v/optimal water level/trade-off/floor”, only hit E-061 physiological injection and spleen volume calibration entries, no “injection method window×degrees of freedom” corresponding edge).

    Evidence Level: C
    References: PMID 4005371, PMID 36725775, PMID 40119099, PMID 24338600

    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.

  • 会话内相关性把有效自由度吃掉多少:同一批判读读两遍,df 是翻倍还是打折

    Evidence level: C

    同一读者在同一个会话内把同一例判读两遍,这两遍并不是两次独立观察;会话内相关性越高,有效自由度越接近「例数」而不是「例数×遍数」,样本量表必须按折减系数修订才算数。

    正文

    气功外气研究中,脉象/得气类判读常采用「同一位读者、同一会话内、对同一批例子重复判读多遍」的设计,然后按「例数 × 遍数」申报自由度,希望用翻倍的 df 换取更窄的置信区间和更小的样本量下限。前一轮 #1556 就按「df≈例数×构型」的粗口径反推过所需例数。这个口径有一个隐藏前提:同一读者在同一会话内读同一例的两遍,是两次互相独立的测量。但真实判读流程里,两遍之间隔着的时间只有几秒到几分钟,读者的注意状态、判读习惯、对前一判读的记忆几乎完全共享——这两遍高度相关。

    统计学对此有一个现成的量化工具,即设计效应(design effect):当组内观测相关为 ρ、每组重复 r 遍时,有效样本量等于名义样本量除以 (1 + (r – 1)ρ)。把这个公式套到自由度上,折减系数就是它的倒数。按三遍重复(r=3)算:会话内相关 ρ=0.1 时折减到 0.83,ρ=0.3 时折减到 0.63,ρ=0.5 时只剩 0.50,ρ=0.7 时只剩 0.42。backlog 里预估「实际 df 可能只有名义的 60%-80%」,对应 ρ 落在 0.1 到 0.3 之间,属于完全合理的区间——问题不是会不会折减,而是折减后样本量表还作不作数。

    我们用蒙特卡洛仿真把两套口径的置信区间覆盖率比了一遍:30 例、每例 3 遍、检验群体均值,重复 4 万次。结果非常清楚。按名义口径(df = 例数 × 遍数)构造的 95% 置信区间,在 ρ=0 时覆盖率是 0.951,一切正常;一旦会话内相关升到 0.3,覆盖率掉到 0.876;0.5 时掉到 0.832;0.7 时只剩 0.791。翻译成人话:如果你把三遍当三次独立观察,声称 95% 的把握,实际只有 80% 左右——每一次「达标」的结论都更可能是运气。而换用保守口径,只按每例的均值、以 df=例数-1 构造区间,覆盖率在所有 ρ 水平都稳定在 0.948 到 0.951 之间,一分不多一分不少。这个结果的实践含义很硬:与其去精确标定折减系数,不如直接从「例数」口径起步,覆盖率天然有保证,ρ 拍不准也伤不到结论。

    那折减系数到底要不要标定?要,但目的变了。仿真给出的裁决是:ρ≤0.1(折减≥0.83)时名义口径勉强可用但已偏薄,不建议依赖;ρ 落在 0.1 到 0.5 之间时,直接采用例数口径最稳,样本量表按「有效 df≈例数」重估;真正危险的场景是 ρ 在个体之间不稳定——文献已经有证据表明这一点。压力痛阈测量研究用线性混合模型分解方差分量后发现,同一测量者在会话内重复的读数之间存在不可忽略的相关(inter-session-repetition ICC 0.73,明显低于人们习惯报告的 inter-session ICC 0.88)(PMID 33767329)——重复次数越多,相关越藏不住。贝叶斯框架下的研究更进一步指出,个体内的方差并非同质,个体层面的 ICC 可能相差六倍,传统单一 ICC 掩盖了这种个体差异(PMID 34816384)。这意味着折减系数不是一个大锅饭常数:新手读者、熟手读者、同一读者的不同状态,折减可能不同。backlog 预设的第三种分支「折减随读者经验漂移,则 df 申报必须附带经验档位」因此不是过度设计,而是被现有信度文献支持的必要动作。

    落到可执行层面,标定协议与 #1556 的框架一致:≥20 例、同一读者同会话内每例读 3 遍、隔周再读 3 遍,用混合模型把会话内、会话间、读者间三个方差分量拆开,算出有效 df 与名义 df 之比。判据分三档:比值≥0.85,粗口径可用,样本量表不动;比值落在 0.6 到 0.85 且跨批稳定,样本量表按折减系数统一上调、例数下限按 1/折减放大;比值随读者经验显著漂移,df 申报必须带经验档位,并写进方法学一节供跨研究比较。需要提醒的是「≥0.85 就不修表」这一档别急着当绿灯:仿真的 0.83 折减档(ρ=0.1)覆盖已经到 0.927,离名义 0.95 有差距。更稳妥的做法是无论标定结果如何,新研究的申报 df 一律从例数口径起步,把「翻倍」当作红利而不是底数。

    这里还给样本量规划领域一个直接的验证支点:信度研究样本量规划确有成熟精度框架(PMID 24338600,2014),且样本量对重测信度估计有实质影响(PMID 40119099,2025)。把这两个方法与上述方差分量分解接上,就可以回答「到底要多少人、多少遍,才能把折减系数标定到能用的精度」——这正是本主题的下一个可测问题。

    证据与局限

    已验证: 设计效应公式给出折减系数的闭式解;蒙特卡洛仿真(4 万次重复)显示名义口径覆盖率在 ρ≥0.3 时跌破 0.90,例数口径全程稳定在 0.95 附近。
    部分验证: 重复读数相关性与个体方差异质性有信度研究文献支撑(PMID 33767329 / 34816384),但对象是压力痛阈与认知任务,不是脉象/得气判读本身。
    待验证: 气功外气判读场景下的会话内相关实测标定尚未进行;「折减系数跨人群可迁移」与「随读者经验漂移」两个竞争假设还没有本领域数据裁决。

    相关阅读

    L1 网络状态:该主题(会话内相关对有效自由度的折减)在 CORRESPONDENCE_NETWORK / EDGE_TABLE 中无对应边;最近邻条目 E-061 仅涉深呼吸 HRV 稳定性(静息 ICC 0.50-0.83、频域差于时间域),可作「读数可靠性随构型漂移」的类比旁证,但非本命题的直接证据。该边尚未进入 L1 对应网络,本文证据等级封顶 C。

    证据等级: C
    参考文献: PMID 33767329, PMID 34816384, PMID 24338600, PMID 40119099


    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 does within-session correlation consume effective degrees of freedom: doubling or halving df when reading the same critique twice?

    Evidence level: C

    When the same reader judges the same set of cases twice within a single session, these two readings are not independent observations. The higher the intra-session correlation, the closer the effective degrees of freedom (df) approach the raw number of cases rather than “cases × repetitions.” Consequently, sample size scales must be revised using a reduction factor to remain valid.

    Body

    In Qi功 external Qi research, pulse diagnosis / “getting Qi” type judgments often employ a design where “the same reader, within the same session, repeatedly judges the same batch of cases multiple times,” then declares degrees of freedom as “cases × repetitions.” This approach hopes to double the df to achieve narrower confidence intervals and lower sample size thresholds. Previous entry #1556 already back-calculated required case numbers using this coarse definition of “df ≈ cases × configuration.” This definition carries a hidden assumption: that two readings of the same case by the same reader within the same session represent two mutually independent measurements. However, in real judgment workflows, the time gap between readings is only a few seconds to minutes; the reader’s attentional state, judgment habits, and memory of the previous judgment are almost entirely shared—making these two readings highly correlated.

    Statistics offers a ready quantification tool for this: the design effect. When intra-group observations correlate at ρ and each case is repeated r times, the effective sample size equals the nominal sample size divided by (1 + (r – 1)ρ). Applying this formula to degrees of freedom, the reduction factor is its reciprocal. Calculating for three repetitions (r=3): with intra-session correlation ρ=0.1, the reduction drops to 0.83; at ρ=0.3, it becomes 0.63; at ρ=0.5, only 0.50 remains; and at ρ=0.7, just 0.42. The backlog estimated that “actual df may be only 60%-80% of the nominal value,” corresponding to ρ falling between 0.1 and 0.3—a fully reasonable range. The issue is not whether reduction occurs, but whether the revised sample size scale still holds water.

    Monte Carlo simulations compared confidence interval coverage under both scales: 30 cases, each read 3 times, testing population means, repeated 40,000 times. The results are clear. Constructing 95% confidence intervals using the nominal scale (df = cases × repetitions) yields coverage of 0.951 when ρ=0—everything works as expected. Once intra-session correlation rises to 0.3, coverage drops to 0.876; at 0.5 it falls to 0.832; and at 0.7 it plummets to only 0.791. In plain terms: if you treat three readings as three independent observations and claim 95% confidence, the actual coverage is only around 80%—every “successful” conclusion is more likely due to chance. Switching to the conservative scale, which uses only the mean per case and constructs intervals with df = cases – 1, stabilizes coverage between 0.948 and 0.951 across all ρ levels—no more, no less. The practical implication is hard: Rather than precisely calibrating the reduction factor, start directly from the “cases” scale; coverage is naturally may support, and inaccurate ρ estimates won’t harm conclusions.

    So does the reduction factor need calibration? Yes, but its purpose shifts. The simulation verdict states: when ρ ≤ 0.1 (reduction ≥ 0.83), the nominal scale is barely usable but already too thin; do not rely on it. When ρ falls between 0.1 and 0.5, directly adopting the case scale is most stable; re-estimate sample size scales using “effective df ≈ cases.” The truly dangerous scenario is when ρ varies unstably across individuals—literature already provides evidence for this. Pressure pain threshold studies decomposed variance components using linear mixed models and found non-negligible correlation between repeated readings by the same measurer within a session (inter-session-repetition ICC 0.73, significantly lower than the commonly reported inter-session ICC of 0.88) (PMID 33767329)—the more repetitions, the harder to hide correlation. Bayesian framework research further indicates that within-individual variance is not homogeneous; individual-level ICCs may differ by up to sixfold, with traditional single ICCs masking such individual differences (PMID 34816384). This means the reduction factor is not a one-size-fits-all constant: novice readers, experienced readers, and different states of the same reader may experience different reductions. The backlog’s third branch—”reduction shifts with reader experience, so df declaration must include experience tiers”—is therefore not over-designed but supported by existing reliability literature.

    At the actionable level, the calibration protocol aligns with #1556’s framework: ≥20 cases, each read 3 times by the same reader within the same session, then re-read 3 times weekly apart. Use mixed models to decompose variance components into intra-session, inter-session, and inter-reader parts, calculating the ratio of effective df to nominal df. Criteria fall into three tiers:
    – Ratio ≥ 0.85: coarse scale usable; sample size scales remain unchanged.
    – Ratio between 0.6 and 0.85 with stability across batches: sample size scales uniformly scaled up by the reduction factor; case number thresholds multiplied by 1/reduction.
    – Ratio significantly shifts with reader experience: df declaration must include experience tiers, written into the methodology section for cross-study comparison.

    Caution: do not rush to treat “≥0.85 = no adjustment” as a green light. The simulated 0.83 reduction tier (ρ=0.1) already yields coverage of 0.927, falling short of the nominal 0.95. A safer approach is to always start new study df declarations from the case scale, treating “doubling” as a bonus rather than the baseline.

    This also provides direct validation for sample size planning in reliability research: established precision frameworks exist (PMID 24338600, 2014), and sample size substantially affects reliability estimation via retest (PMID 40119099, 2025). Connecting these methods with the above variance decomposition answers “how many people and how many repetitions are needed to calibrate the reduction factor to usable precision”—which is the next measurable question of this topic.

    Evidence and Limitations

    Verified: The design effect formula yields a closed-form solution for the reduction factor; Monte Carlo simulations (40,000 repetitions) show nominal-scale coverage drops below 0.90 when ρ ≥ 0.3, while the case-scale remains stable near 0.95 throughout.
    Partially Verified: Reliability literature supports correlations between repeated readings and heterogeneity in individual variance (PMID 33767329 / PMID 34816384), but these studies focused on pressure pain thresholds and cognitive tasks, not pulse diagnosis / “getting Qi” judgments themselves.
    Pending Verification: Actual calibration of intra-session correlation in Qi功 external Qi judgment scenarios has not yet been conducted; competing hypotheses—”reduction factor is transferable across populations” vs. “reduction shifts with reader experience”—lack domain-specific data to adjudicate them.

    Related Reading

    L1 Network Status: This topic (intra-session correlation’s reduction of effective degrees of freedom) has no corresponding edge in the CORRESPONDENCE_NETWORK / EDGE_TABLE; nearest neighbor entry E-061 only involves breathing HRV stability (resting ICC 0.50–0.83, frequency domain worse than time domain), serving as an analog side evidence for “reading reliability drifting with configuration,” but not direct evidence for this proposition. This edge has not yet entered the L1 correspondence network; thus, this topic’s evidence level caps at C.

    Evidence Level: C
    References: PMID 33767329, PMID 34816384, PMID 24338600, PMID 40119099


    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.

  • 标定误差与统计误差相关时,闭式膨胀因子还保不保序:相关系数只要到 0.4,全部十二格一起漏水

    Evidence level: C

    上一篇文章的闭式保序膨胀因子建在一个没说出口的前提上——标定误差与统计误差各管各的;本文把这个前提拆了:相关一上 0.4,十二个格子全线跌破验收下限,而且正负相关的后果根本不对称。

    正文(1000-1500字)

    父文给可识别区间配了一枚闭式膨胀因子:k 等于「r 方乘 w 加一减 w」再开方,r 是真实与宣称标定精度之比,w 是标定项方差占误差预算的比例。代回仿真,全格点覆盖 0.944 到 0.952——但这枚闭式从推导的第一行起就假设标定误差与统计误差独立。真实注入实验里两者未必各管各的:同一批判读若系统性宽松,既抬高标定偏差又拉大方差比偏离,两条误差通过同一个源头挂钩。相关一旦非零,闭式的覆盖承诺还剩几成?

    仿真沿用父文全部口径:df 取 30、100,宣称精度 ε 取 0.10、0.20,倍数比 r 取 1、2、3,每格二十万次。新增一个旋钮——cal-stat 相关系数 ρ_cs,从负 0.8 扫到正 0.8,用高斯 copula 把统计端的卡方尺度族抽样和标定端的正态抽样搅在一起。先验基线:ρ_cs=0 时闭式覆盖 0.950 到 0.952,与父文严丝合缝,模型对齐无误。

    然后漏水,而且是整齐地漏。ρ_cs=+0.4 时,十二个格子的覆盖全部掉进 0.904 到 0.929——没有一格守住 93% 的验收下限;ρ_cs=+0.8 时最低 0.859,最好的一格也只有 0.907。backlog 里预设的两条判据,第一条(中等相关 ±0.4 以内闭式稳健)直接判死:0.4 不是安全区,是全线漏水的起点。第二条(+0.8 跌破 90% 才升级)倒是触发,但比预期提前——升级不该等到 0.8,0.4 就该动手。

    漏水的机制一句话能说清:相关的代价走的是交叉项 2 乘 ρ_cs 乘真实标定误差乘统计误差的积,这一项直接加进总方差,而闭式只按独立情形的平方和配宽度。正相关把总方差抬起来、区间却没跟着宽,于是欠覆盖;负相关把总方差压下去、区间显得过宽,于是过覆盖——ρ_cs=−0.4 时覆盖虚胖到 0.969 到 0.988,−0.8 时接近全包。正负两个方向不对称:正相关是有效性风险(区间说谎),负相关只是精度浪费(区间啰嗦)。而真实判读场景里,宽松/严苛同源恰恰产生正相关——最危险的方向正是最常见的那一个。还有一层没料到的:交叉项的相对大小取决于两个误差分量的均衡度,哪头独大都伤不到区间,两头差不多大时才最疼——本文格子里两头的比例本来就在同一量级,所以漏得整齐;预算畸形(一头占九成以上)的实验反而对相关免疫。

    补救比想象的便宜。既然损伤走的是交叉项,升级也是一行代数:k 改成携带 ρ_cs 的总方差形式,把交叉项显式加回平方根里。代回仿真,整张 ρ_cs 网格(含正负 0.8)覆盖全部回到 0.949 到 0.953——无需迭代、无需蒙特卡洛,闭式升级为 k(r,w,ρ_cs) 一次换算到位。测量学文献早有平行教训:输入量相关时 GUM 律的传播公式会失准,蒙特卡洛才是裁判(PMID 27114218);展开不确定度的覆盖验收上,蒙特卡洛与 ISO-GUM 的系统差本身就值得专文对照(PMID 21238753);衍生气相色谱里共用的衍生化步骤让多个不确定度分量天然相关、须单独建账(PMID 31845399)——「同一来源养活两条误差」在分析化学里是常态,不是例外。

    落到守则,三条。第一,父文的闭式 k(r,w) 只在「相关可忽略」时兑现覆盖承诺,报告里 w 与 r 之外必须附第三项:声明 cal-stat 相关的实测值或豁免理由。第二,相关为正是主要威胁,覆盖沿 ρ_cs 线性下漏,0.4 即越线;相关为负只浪费宽度,最多白付近半区间,验收时可放行。第三,升级式 k(r,w,ρ_cs) 的覆盖承诺以「相关系数实测准确」为前提——本文 copula 模型里两者天然相等是模型的内禀性质,真实实验里必须直接量测,这是本文结论最大的软肋,也是下一篇要拆的口子。

    证据与局限

    已验证: 独立基线闭式覆盖 0.950-0.952(十二格,与父文对齐);ρ_cs=+0.4 全部格子跌破 93%(0.904-0.929);+0.8 最低 0.859;负相关过覆盖 0.969-1.000;k(r,w,ρ_cs) 升级式全网格拉回 0.949-0.953——均为十二格×每格二十万次、固定种子的仿真结果。
    部分验证: 相关经高斯 copula 注入,模型内 γ(标定端与统计端的实测相关)恒等于 ρ_cs 是构造性质;交叉项解析式 2ρ_cs·ε_true·s 在该模型内精确成立,真实实验的相关结构未必如此干净。
    待验证: 真实判读数据里 γ 对 ρ_cs 的衰减与映射关系;ρ_cs 本身用样本相关估计时的估计噪声对升级式覆盖的侵蚀;预算畸形(单分量占比九成以上)格子对相关的免疫边界。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络(EDGE_TABLE_science-yixue.md 检索「覆盖/膨胀/标定/相关/方差」零命中;CORRESPONDENCE_NETWORK.md 最近邻为 E-060 标签信度边【conflict,涉及蒙特卡洛与加权系数精度区间,但属「五脏标签分组」命题,不覆盖本篇「cal-stat 相关×闭式保序」】与 E-061 0.1Hz 生理注入边,均无对应关系)。

    证据等级: C
    参考文献: PMID 21238753, PMID 27114218, PMID 31845399

    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。

    When calibration error correlates with statistical error, does the closed-form inflation factor preserve ordering: once the correlation coefficient reaches 0.4, all twelve cells…

    Evidence level: C

    The closed-form inflation factor presented in the preceding article rests on an unstated premise: calibration error and statistical error are treated as independent. This article dismantles that assumption. Once the correlation coefficient between them reaches 0.4, all twelve grid points collectively fail to meet the acceptance threshold, and the consequences of positive versus negative correlation prove fundamentally asymmetric.

    Body (1000–1500 words)

    The parent text assigns a closed-form inflation factor to identifiable intervals: $k = \sqrt{r^2 w + 1 – w}$, where $r$ denotes the ratio of true to claimed calibration precision, and $w$ represents the proportion of the error budget allocated to the calibration term. Substituting these parameters into simulations yields full-grid coverage between 0.944 and 0.952. However, this closed-form derivation assumes independence from the first line of deduction. In real-world injection experiments, errors do not necessarily operate in isolation; a single critical reading that appears systematically lenient simultaneously elevates calibration bias and inflates the variance ratio, linking both errors to a common source. If correlation is non-zero, how much of the coverage promise remains?

    Simulations adhere strictly to the parent text’s specifications: degrees of freedom ($df$) set at 30 and 100; claimed precision $\varepsilon$ at 0.10 and 0.20; ratio $r$ at 1, 2, and 3; each grid point running twenty million trials. A new control knob is introduced: the cal-stat correlation coefficient $\rho_{cs}$, sweeping from $-0.8$ to $+0.8$. A Gaussian copula blends the chi-square scale family from the statistical end with normal sampling from the calibration end. The prior baseline—$\rho_{cs}=0$—yields closed-form coverage of 0.950 to 0.952, aligning perfectly with the parent text; model alignment is exact.

    Then the leak begins, and it leaks uniformly. At $\rho_{cs}=+0.4$, coverage for all twelve grids drops entirely into the 0.904–0.929 range—no single grid maintains the 93% acceptance floor. At $\rho_{cs}=+0.8$, the lowest coverage reaches 0.859, and even the best-performing grid falls to 0.907. Two pre-set criteria in the backlog are immediately invalidated: the first (closed-form robustness within $\pm0.4$ moderate correlation) fails outright—0.4 is not a safe zone but the starting point of total breakdown. The second (upgrade triggered only when $+0.8$ drops below 90%) does trigger, but earlier than expected—the upgrade should not wait until 0.8; action is required at 0.4.

    The mechanism of leakage reduces to one clear principle: the cost of correlation flows through the cross-term $2 \rho_{cs} \varepsilon_{\text{true}} s$, where $\varepsilon_{\text{true}}$ is the true calibration error and $s$ is the statistical error. This term directly adds to total variance, while the closed-form width allocates only the sum of squares under independence. Positive correlation inflates total variance without widening the interval, resulting in under-coverage; negative correlation suppresses total variance while the interval appears overly wide, yielding over-coverage. At $\rho_{cs}=-0.4$, coverage becomes inflated to 0.969–0.988; at $-0.8$, it approaches full coverage. Positive and negative directions are asymmetric: positive correlation poses validity risk (the interval lies), whereas negative correlation merely wastes precision (the interval is verbose). In real diagnostic scenarios, lenient/strict conditions stemming from the same source generate positive correlation—the most dangerous direction coincides with the most common one.

    An unexpected layer emerges: the relative magnitude of the cross-term depends on the balance between the two error components. If one dominates, the interval suffers little; maximum damage occurs when both are comparable in scale. In this study’s grids, the two ratios lie within the same order of magnitude, causing uniform leakage; experiments with skewed budgets (one component exceeding 90%) prove immune to correlation.

    Remediation is cheaper than anticipated. Since damage arises via the cross-term, upgrading requires a single algebraic step: redefine $k$ to carry $\rho_{cs}$ in the total variance form, explicitly adding the cross-term back into the square root. Re-running simulations across the full $\rho_{cs}$ grid (including $\pm0.8$) restores coverage to 0.949–0.953—no iteration, no Monte Carlo needed; the closed-form upgrade $k(r,w,\rho_{cs})$ computes directly. Measurement literature offers parallel lessons: when input quantities correlate, the GUM propagation formula becomes inaccurate, and Monte Carlo serves as the arbiter (PMID 27114218); in expanding uncertainty coverage, systematic differences between Monte Carlo and ISO-GUM warrant dedicated comparison (PMID 21238753); in derivative gas chromatography, shared derivatization steps make multiple uncertainty components naturally correlated, requiring separate accounting (PMID 31845399)—“one source feeding two errors” is the norm, not the exception, in analytical chemistry.

    Three rules follow:
    1. The parent’s closed-form $k(r,w)$ delivers coverage only when correlation is negligible; reports must include a third item alongside $w$ and $r$: either the measured cal-stat correlation or justification for exemption.
    2. Positive correlation is the primary threat; coverage leaks linearly along $\rho_{cs}$, crossing the 0.4 threshold; negative correlation merely wastes width, at most inflating nearly half the interval, acceptable under review.
    3. The upgraded closed-form $k(r,w,\rho_{cs})$ assumes accurate measurement of the correlation coefficient. In this article’s copula model, both are inherently equal by construction; in real experiments, they must be directly measured—a critical soft spot and the next frontier to explore.

    Evidence and Limitations

    Verified: Independent baseline closed-form coverage 0.950–0.952 (twelve grids, aligned with parent text); $\rho_{cs}=+0.4$ causes all grids to drop below 93% (0.904–0.929); $+0.8$ yields minimum 0.859; negative correlation produces over-coverage 0.969–1.000; upgraded $k(r,w,\rho_{cs})$ restores full grid coverage to 0.949–0.953—all based on twelve grids × twenty million trials per grid with fixed seeds.
    Partially Verified: Correlation injected via Gaussian copula; within the model, $\gamma$ (measured correlation between calibration and statistical ends) equals $\rho_{cs}$ by construction; the cross-term analytical expression $2\rho_{cs}\cdot\varepsilon_{\text{true}}\cdot s$ holds exactly in this model, though real experimental correlation structures may not be so clean.
    Pending Verification: The attenuation and mapping relationship of $\gamma$ to $\rho_{cs}$ in real diagnostic data; erosion of upgraded-form coverage due to estimation noise when $\rho_{cs}$ itself is estimated from sample correlation; immunity boundaries of grids with skewed budgets (single component exceeding 90%).

    Related Reading

    L1 Network Status: This topic has not yet entered the corresponding L1 network. Search in EDGE_TABLE_science-yixue.md for “coverage/inflation/calibration/correlation/variance” yields zero hits; the nearest neighbor in CORRESPONDENCE_NETWORK.md is E-060 label confidence edge【conflict, involving Monte Carlo and weighted coefficient precision intervals, but belonging to the “five-organ label grouping” proposition, not covering this article’s “cal-stat correlation × closed-form order preservation”】and E-061 0.1Hz physiological injection edge; neither corresponds to this topic.

    Evidence Level: C
    References: PMID 21238753, PMID 27114218, PMID 31845399

    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.

  • 那些算出来的门槛为什么普遍差两三成,答案可能藏在没人报的东西里:已发表判读研究几乎只给κ与区间,不给两个读者各自的完整边际,于是「分歧摊在几条线上」这个量从未被观测过

    Evidence level: C

    同一个偶然一致率、同一个真实分歧总量,仅因分歧落在哪几档不同,门槛就差两成上下——而决定这件事的关键量,恰恰是判读研究不报告的那一个。

    一个被卡住的链条

    判读一致性研究里有一条环环相扣的链:想定样本量门槛,先要算方差系数;想算方差系数,先要知道两位读者的分歧按什么形状摊在各个类别上。上一篇文章已经证明,这条链的最后一环不是可有可无的细节——同样的偶然一致率、同样的真实分歧总量,仅改变差落在哪几档,方差系数就能跨一点一四到一点三二倍,折算成样本量门槛是两成上下的差别;而实践里当唯一旋钮用的「偶然一致率单变量简化式」,在三百六十个格点的检验里只有不到两成能落到一成误差以内,最坏偏四点五倍。

    问题在于:这个决定性的量——分歧的档位落点——在已发表文献里几乎找不到。主流报告格式是「kappa 等于某值、九成五区间多少」,再讲究一点补一个别的系数作对照;两位读者各自的分类别边际属于少数派做法,连观察一致率本身都常常缺席。也就是说,「文献里的门槛偏乐观还是偏保守」这个问题,今天在数据层面就无法回答——不是没人算过,是算它需要的原料从来没被系统地报出来。

    为什么边际落点不是技术细节

    有人会问:门槛差两成,多招两成人不就行了?这个直觉漏了两件事。第一,偏差没有固定方向。边际集中(分歧全压在一两档)与边际摊开(分歧均匀铺满各档)会把方差系数往相反方向拉,同样一份「看起来合理」的门槛,对不同研究可能一边过剩、一边严重不足。第二,单参数简化式的误差可以大到四点五倍,意味着照它算出来的门槛可能与真实需要差出一个数量级——这种误差不会靠「多留点余量」自然吸收。

    方法学文献早有伏笔。kappa 系列系数对边际形态的敏感性自上世纪九十年代起就被反复指出:同样的边际结构能让系数在「看似很强」与「令人担忧」之间跳档,样本量公式的输入正是完整边际而非某个汇总值;二〇二二年一项蒙特卡洛研究进一步量化了边际不对称分布下各 kappa 型系数行为的漂移。链条早已写明,缺的只是文献层面那一手实测。

    把「没被观测过的量」测出来

    可行的做法是一次文献计量审计。筛选至少四十项报告了完整分类别边际(四格表或边际比例)的判读研究,逐项重建两位读者的边际与差落点,算出该研究的真实方差系数,再与它若按单参数简化式估出的系数相比,得到「文献门槛偏差」的比值分布。乳腺密度四档判读这类报告格式完整的研究是现成模板。

    关键不在算,而在三条判据怎么切。若比值中位落在零点九到一点一,说明单参数简化在实践中大体够用,只是在极端构型上出错,处理办法是要求并报构型敏感性;若中位系统性偏离一超过两成,说明偏差有固定方向,那么「判读研究必须报告完整边际」应升级为强制规范,否则其门槛不可复核;若偏差在两个方向上各占一半,说明偏差不可先验修正,只能逐研究重算——门槛的「事前公示」目标就须降级为「事前给出区间」。

    这一步为什么必须走在设计前面

    我们自己的多池设计马上要用这些门槛。如果文献门槛的偏差结构摸不清,等于把设计手册架在一个没校准过的刻度尺上:门槛偏松,欠功效研究照常通过;门槛偏紧,样本量翻倍却没人知道为什么。先花一轮文献计量的成本,把「没人报的量」的偏差分布测出来,后面所有判读设计的样本量核算才有可追溯的依据——这正是这一系列文章反复强调的次序:判据的来历必须比判据本身更早被审。

    证据与局限

    已验证:kappa 类系数受边际分布支配(PMID 8501467 拆解偏倚、流行度与 kappa 三成分);边际不对称下各 kappa 型系数行为漂移已由蒙特卡洛量化(PMID 35124830);一致性研究的报告规范本身薄弱,样本量依据常缺(PMID 36123642)。

    部分验证:完整边际是样本量公式的必要输入(PMID 22560852、11512137 的公式结构均以此为前提),但「文献门槛偏差」的实测分布尚无人做过,本文设计属待执行方案。

    待验证:四十项以上研究的比值分布、三条判据的实际落点、以及「强制报边际」规范的可执行成本。注意:本文属统计方法学推演,与易学概念无直接对应,其价值在于为后续辨证判读类设计提供门槛校准的前置工具。

    相关阅读

    L1 网络状态:本主题(文献门槛偏差 × 方差系数边际形态依赖)在 CORRESPONDENCE_NETWORK 与 EDGE_TABLE_science-yixue 中均无对应边;最近邻为 E-060 标签前提边,其原文摘录:「KB-292(conflict):中医医师间诊断一致性在开放人群中『低(low)』,作者称为『令人担忧』(PMID 30383398)……标签信度,冲突(未裁决)」——该边处理标签信度本身,与本篇的门槛偏差审计不构成同一条边,故本文证据等级封顶 C。

    证据等级: C
    参考文献: PMID 8501467, PMID 35124830, PMID 36123642, PMID 22560852, PMID 11512137, PMID 37127220

    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。

    Why calculated thresholds universally differ by 20–30%: the answer likely lies in what is never reported; published diagnostic studies almost exclusively provide κ and intervals…

    Evidence level: C

    The reason calculated thresholds are universally off by roughly 20–30% likely lies in the unreported data: published interpretation studies almost exclusively report κ (kappa) and confidence intervals, but never the complete marginal distributions for both readers. Consequently, the quantity “divergence spread across specific categories” has never been observed.

    With the same chance agreement rate and the same total amount of real divergence, the threshold differs by about 20% simply because the divergence falls into different categories. The key quantity determining this difference is precisely the one not reported in interpretation studies.

    A Stalled Chain

    In consistency studies, there is an interlinked chain: to set the sample size threshold, one must first calculate the variance coefficient; to calculate the variance coefficient, one must know how the divergence between two readers is distributed across categories. As proven in the previous article, this final link is not a trivial detail—the same chance agreement rate and the same total amount of real divergence can result in variance coefficients ranging from 1.14 to 1.32 times, translating to a sample size threshold difference of about 20%. In practice, the “single-variable simplified formula for chance agreement,” used as the sole control knob, fails in nearly 80% of cases across 360 grid points, with the worst error reaching 4.5 times.

    The problem is that this decisive quantity—the specific category distribution of divergence—is virtually absent in published literature. The standard reporting format is “kappa equals X, with a 95% confidence interval of Y,” sometimes supplemented by another coefficient for comparison; the individual marginal distributions for both readers belong to a minority approach, and even the observed agreement rate itself is often missing. Therefore, the question of whether “thresholds in the literature are overly optimistic or overly conservative” cannot be answered at the data level today—not because no one has calculated it, but because the raw materials required for such calculation have never been systematically reported.

    Why Marginal Distribution Is Not a Technical Detail

    One might ask: if the threshold differs by 20%, simply recruiting 20% more participants would suffice. This intuition overlooks two critical points. First, the bias does not have a fixed direction. When margins are concentrated (divergence compressed into one or two categories) versus when they are spread out (divergence evenly distributed across all categories), the variance coefficient is pulled in opposite directions. Thus, the same “seemingly reasonable” threshold may be excessively high for one study and severely insufficient for another. Second, the error of the single-parameter simplified formula can be as large as 4.5 times, meaning the calculated threshold could differ from the true requirement by an entire order of magnitude—an error that cannot be naturally absorbed by “adding extra margin.”

    Methodological literature has long hinted at this. Since the 1990s, it has been repeatedly noted that kappa-series coefficients are highly sensitive to marginal distribution: the same marginal structure can cause a coefficient to jump between “appearing strong” and “concerning,” while the input for sample size formulas must be the complete marginal distribution rather than some aggregated value. A 2022 Monte Carlo study further quantified the drift in behavior of various kappa-type coefficients under asymmetric marginal distributions. The chain was already laid out; what was missing was the empirical measurement at the literature level.

    Measuring the “Unobserved Quantity”

    A feasible approach is a bibliometric audit. Select at least 40 studies reporting complete marginal distributions (four-fold tables or marginal proportions) and reconstruct the margins and divergence points for both readers, calculate the true variance coefficient for each study, then compare it with the coefficient estimated using the single-parameter simplified formula to derive the distribution of “literature threshold bias.” Studies on breast density with four categories provide ready-made templates.

    The key is not the calculation itself, but how to apply three criteria:
    – If the median ratio falls between 0.9 and 1.1, the single-parameter simplification is largely adequate in practice, though it fails under extreme configurations; the solution is to require and report configuration sensitivity.
    – If the median systematically deviates by more than 20%, the bias has a fixed direction; thus, “interpretation studies must report complete margins” should become a mandatory standard, otherwise their thresholds cannot be verified.
    – If the bias is split equally in both directions, it cannot be corrected a priori; each study must be recalculated individually—the goal of “pre-publication disclosure” must therefore downgrade to “pre-publication interval reporting.”

    Why This Step Must Precede Design

    Our own multi-pool design immediately requires these thresholds. If the bias structure of literature thresholds remains unclear, we are essentially building our design manual on an uncalibrated ruler: overly loose thresholds allow underpowered studies to pass; overly tight thresholds double the sample size without anyone knowing why. Spending one round of bibliometric cost to measure the bias distribution of “unreported quantities” ensures that all subsequent sample size calculations have traceable foundations—this is precisely the sequence repeatedly emphasized in this series of articles: the origin of the criteria must be audited before the criteria themselves.

    Evidence and Limitations

    Verified: Kappa-type coefficients are governed by marginal distribution (PMID 8501467 decomposes bias, prevalence, and kappa into three components); under asymmetric margins, drift in behavior of various kappa-type coefficients has been quantified via Monte Carlo (PMID 350124830); reporting standards for consistency studies are inherently weak, often lacking sample size basis (PMID 36123642).

    Partially Verified: Complete margins are necessary inputs for sample size formulas (PMID 22560852, 11512137 both use this as a structural prerequisite), but the empirical distribution of “literature threshold bias” has never been measured; this article proposes a pending execution plan.

    Pending Verification: The ratio distribution of over 40 studies, the actual landing points of the three criteria, and the implementation cost of mandating margin reporting. Note: This article belongs to statistical methodology derivation and does not directly correspond to Yijing Studies concepts; its value lies in providing a prerequisite tool for threshold calibration of future diagnostic interpretation designs.

    Related Reading

    L1 Network Status: This topic (literature threshold bias × variance coefficient marginal shape dependence) has no corresponding edge in CORRESPONDENCE_NETWORK or EDGE_TABLE_science-yixue; the nearest neighbor is the E-060 label prerequisite edge, whose original excerpt states: “KB-292 (conflict): diagnostic consistency among TCM practitioners in open populations is ‘low (low)’, authors call it ‘concerning’ (PMID 30383398) … label credibility, conflict (unresolved)”—this edge handles label credibility itself and does not constitute the same edge as this article’s threshold bias audit, hence this article’s evidence level caps at C.

    Evidence Level: C
    References: PMID 8501467, PMID 350124830, PMID 36123642, PMID 22560852, PMID 11512137, PMID 37127220

    Evidence Level: A = directly supported by PubMed, 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.

  • 档数越多样本量越爆炸,但发表的论文照样几十例:是真能扛,还是换了系数把欠功效盖住了

    Evidence level: C

    前一篇算出的样本量表显示四档要 1850 例、五档要 34225 例,而真实发表的等级一致性研究大多只有几十到几百例——要么这些研究集体欠功效,要么它们靠换系数把爆炸部分绕过去了。这一篇要审的就是这条逃生通道。

    正文

    先把上一条的账摆出来。在”两次评级次档质量均匀摊开”的假设下,测一个 0.90 的中高一致率,若要求估计精度 δ=0.05,四档量表需要约 1850 例,五档需要约 34225 例。这个增长不是线性的,是档数每加一档就翻一个量级的爆炸式增长。可去翻任何一本发表过的等级一致性研究——脉诊辨证、影像分级、症状量表——样本量几乎清一色是几十到几百例。

    这产生了一个尖锐的问题:如果样本量表是对的,那现存所有四档以上、报告”一致性中等、可用于临床”的研究,全都在严重欠功效的状态下发表了,其系数区间宽到撑不起任何裁决。但这些论文照样过审、照样发表、照样被引用。两种解释必须选一个,或者都要:第一,它们确实欠功效,结论不可靠,只是审稿人没算这笔账;第二,它们换了某个对档数不敏感的系数,把爆炸部分绕过去了。

    第二条通道是真实存在的,文献里甚至写得明明白白。方法学研究的经典结论是:二次加权 κ 随档数增加系统性升高,而线性加权 κ 对档数基本不敏感(PMID 8834562)。换句话说,同一批数据,从二次加权换成线性加权,档数惩罚就大半消失了。蒙特卡洛研究还补了一刀:档数与样本量同时增大时,加权系数的估计精度反而下降——”多招点受试者”并不自动换来精度(PMID 33853549)。更扎眼的是一个中医场景的实例:五位中医师对 42 例做自由边际评级,Fleiss κ 与 Gwet AC2 算出的结论差了整整一个等级(PMID 29851495)。系数选型不是技术细节,它直接决定你论文最后一行写”一致性良好”还是”一致性堪忧”。

    有趣的是,这条逃生通道本身在研究内核里就是一处公开冲突。一边的观点是:抗灰区(中低一致率)最优的恰恰是 Gwet AC2 和 Brennan-Prediger κ,但所有加权系数在”中低一致性 + 大样本 + 类别数多”的组合下精度显著下降——也就是档数惩罚躲不掉,AC2 只是相对最好的那个。另一边的模拟结论方向相反:Fleiss κ 和 Conger κ 在序数分类上反而更高,而尺度维数增加只对名义分类有助、对序数分类无助,所以像五脏辨证这种多等级序数场景,Gwet AC2 恰是比 κ 更稳健的选择。两派共同承认的是”换系数改变结论”,分歧在于哪个方向换才是对的——这正是欠功效审计最该吃进去的背景:如果连选哪个系数都未裁决,那”低 n 报出的系数”的参考价值要再打一层折扣。

    还有一个常被漏掉的层次:分层设计下 AC1 的样本量单位与齐性检验根本对不上——有的研究算的是评定者数,有的算的是病例数,两者不可直接互校(PMID 40298274)。所以做审计时,”该有多少样本”这个问题本身要先统一单位,否则实际 n 与应有 n 的比值会算出假象。

    怎么裁决?做法是文献计量:检索不少于 30 篇报告四档及以上有序一致性系数的临床研究,逐篇提取样本量、档数、系数类型、报告区间宽度,然后按样本量表反推”这个设计应有的最小 n”,计算实际 n 与应有 n 的比值分布。判据分三路:若七成以上研究的实际 n 低于应有值的一半,欠功效立得住,此后任何多档判读设计必须先过功效核算;若比例低于三成,说明存在模型没捕捉到的降维机制(比如名义五档实际只用了三档),样本量因子要补一个”有效档数”修正项;若分布呈双峰——一部分严重不足、一部分充裕——那差异来源就要归因到系数类型或材料难度梯度,顺手产出一份”什么条件下可以省样本”的清单。

    这条审计对本项目的意义很直接:脉诊、辨证、穴位判定这些核心研究线,全部依赖”几位评定者对同一批病例分等级”的一致性数据。如果那条逃生通道被证实,那么引用链条里所有”一致性中等”的文献,都要先问一句:这几十例,是真能扛,还是系数换出来的体面。

    证据与局限

    已验证: 二次加权 κ 随档数升高而线性加权 κ 不敏感(PMID 8834562);灰区场景下系数精度随档数与样本量交互变化(PMID 33853549);中医师评级中 Fleiss κ 与 AC2 结论差一个等级(PMID 29851495)——”换系数改变结论”这条通道文献证据充分。

    部分验证: 分层设计中样本量单位错配问题有专门研究但覆盖面窄(PMID 40298274);”换系数绕开档数惩罚”在已发表文献中实际发生的频率,目前没有任何直接统计。

    待验证: ≥30 篇多档一致性研究的欠功效比例;实际 n / 应有 n 比值分布的形状;系数类型能否解释双峰。以上全部为提案,未执行。

    相关阅读

    证据等级: C
    参考文献: PMID 8834562, PMID 33853549, PMID 29851495, PMID 40298274

    L1 网络状态:E-060 标签信度(KB-317,conflict,未裁决)——「闸门规格撞上方法学反例」:AC1 分层设计样本量 n≥50 与 KB-301 的「120 例」单位错配不可互校;「所有加权系数在中低一致性 + 大样本 + 类别数多时精度显著下降」(PMID 33853549);KB-333(conflict,中)另立第三处错配——「Fleiss κ 与 Conger κ 在序数分类上悖论性更高,尺度维数增加只对名义分类有助、对序数分类无助,Gwet AC2 恰是序数场景更稳健的选择」。本主题(多档一致性研究文献层面的欠功效审计)尚未进入 L1 对应网络,E-060 为最近相邻边。


    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。

    The more variables, the more samples explode, yet published papers still have dozens of cases: is it truly robust or did they swap coefficients to cover up low power?

    Evidence level: C

    The previous calculation showed that a four-tier scale requires 1,850 cases and a five-tier scale requires 34,225 cases to achieve a moderate-to-high consistency rate of 0.90 with an estimation precision of δ=0.05, assuming “uniform allocation across secondary tiers.” However, most published studies on grade consistency—such as pulse diagnosis, imaging grading, or symptom scales—typically involve only dozens to hundreds of cases. Either these studies collectively suffer from underpowering, or they circumvented the exponential growth by swapping coefficients. This paper audits that escape route.

    Main Text

    Let us first lay out the arithmetic from the previous section. Under the assumption of “uniform allocation across secondary tiers,” achieving a moderate-to-high consistency rate of 0.90 with an estimation precision of δ=0.05 requires approximately 1,850 cases for a four-tier scale and about 34,225 cases for a five-tier scale. This growth is not linear; it is an explosive increase, doubling in magnitude with each additional tier. Yet, any review of published grade consistency studies—pulse diagnosis, imaging grading, symptom scales—reveals that sample sizes are almost exclusively in the range of dozens to hundreds.

    This raises a sharp question: if the sample scale is correct, then all existing studies reporting “moderate consistency, clinically useful” with four or more tiers have been published under severe underpowering conditions, with coefficient intervals too wide to support any decision. Nevertheless, these papers were accepted, published, and cited. We must choose between two explanations, or accept both: First, they indeed suffer from underpowering; their conclusions are unreliable, but reviewers failed to account for this cost. Second, they swapped to a coefficient that is insensitive to the number of tiers, thereby circumventing the explosion.

    The second route is real; the literature explicitly documents it. A classic methodological conclusion states that quadratic-weighted κ systematically increases with the number of tiers, whereas linear-weighted κ is largely insensitive to tier count (PMID 8834562). In other words, switching from quadratic-weighting to linear-weighting for the same dataset eliminates most of the tier penalty. Monte Carlo simulations further sharpen this point: when both the number of tiers and sample size increase, the estimation precision of weighted coefficients actually declines—”enrolling more subjects” does not automatically improve precision (PMID 33853549). Even more striking is a clinical example from Traditional Chinese Medicine: five practitioners independently graded 42 patients using free margins; Fleiss κ and Gwet AC2 yielded conclusions differing by exactly one grade (PMID 29851495). Coefficient selection is not merely a technical detail; it directly determines whether the last sentence of your paper reads “good consistency” or “poor consistency.”

    Interestingly, this escape route itself represents an open conflict within the research core. On one side, the optimal metrics for gray-zone (low-to-moderate consistency) scenarios are Gwet AC2 and Brennan-Prediger κ, yet all weighted coefficients suffer significant precision loss under the combination of “low-to-moderate consistency + large sample size + many categories”—meaning the tier penalty cannot be avoided, and AC2 is only relatively best. On the other side, simulation results point in the opposite direction: Fleiss κ and Conger κ perform better for ordinal classification, whereas increasing scale dimensionality helps only nominal classification and harms ordinal classification; thus, for multi-grade ordinal scenarios like five-phase differentiation, Gwet AC2 is more robust than κ. Both camps agree that “swapping coefficients changes the conclusion,” but they disagree on which direction is correct—precisely the background that underpowering audits must consume: if even the choice of coefficient remains undecided, the value of coefficients reported in low-n studies must be discounted further.

    There is also a frequently overlooked layer: under stratified designs, the sample size units for AC1 do not align with homogeneity tests—some studies calculate based on the number of raters, others based on the number of cases, making direct comparison impossible (PMID 40298274). Therefore, during auditing, the question “how many samples are needed” must first unify its units; otherwise, the ratio of actual n to required n will yield spurious results.

    How do we resolve this? The approach is bibliometric: retrieve at least 30 clinical studies reporting ordered consistency coefficients for four or more tiers, extract sample size, number of tiers, coefficient type, and reported interval width for each, then reverse-calculate the “minimum required n” based on the sample scale and compute the distribution of actual n / required n ratios. The criteria fall into three categories: if more than 70% of studies have actual n less than half the required value, underpowering is confirmed; any future multi-tier design must undergo power calculation first. If the proportion is less than 30%, it indicates a dimensionality-reduction mechanism not captured by current models (e.g., nominal five tiers effectively using only three), requiring a “effective tier count” correction term in the sample factor. If the distribution is bimodal—some severely underpowered, others adequately powered—the source of difference must be attributed to coefficient type or material difficulty gradient, yielding a checklist of “under what conditions can sample size be reduced.”

    The significance of this audit for our project is direct: core research lines such as pulse diagnosis, syndrome differentiation, and acupoint determination all rely on consistency data where “several raters grade the same set of cases.” If this escape route is confirmed, then every “moderate consistency” paper in the citation chain must first be asked: do these dozens of cases truly hold up, or is this respectability manufactured by swapping coefficients?

    Evidence and Limitations

    Verified: Quadratic-weighted κ increases linearly with tier count while linear-weighted κ remains insensitive (PMID 8834562); coefficient precision in gray-zone scenarios varies with the interaction of tier count and sample size (PMID 33853549); in Traditional Chinese Medicine practitioner grading, Fleiss κ and AC2 differ by one grade (PMID 29851495)—the “swapping coefficients changes conclusions” route is well-supported in the literature.

    Partially Verified: The mismatch of sample size units in stratified designs has been studied but covers limited scope (PMID 40298274); the actual frequency of “escaping tier penalties via coefficient swapping” in published literature lacks direct statistical evidence.

    Pending Verification: The proportion of underpowered studies among ≥30 multi-tier consistency studies; the shape of the actual n / required n ratio distribution; whether coefficient type explains bimodality. All are proposals, not yet executed.

    Further Reading

    Evidence Level: C
    References: PMID 8834562, PMID 33853549, PMID 29851495, PMID 40298274

    L1 Network Status: E-060 label credibility (KB-317, conflict, undecided)—”tier specification collides with methodological counterexample”: AC1 stratified design sample size n≥50 mismatches KB-301’s “120 cases” unit; “all weighted coefficients lose precision significantly under low-to-moderate consistency + large sample size + many categories” (PMID 33853549); KB-333 (conflict, medium) introduces a third mismatch—”Fleiss κ and Conger κ paradoxically perform better for ordinal classification, while increasing scale dimensionality aids only nominal classification and harms ordinal classification, making Gwet AC2 the more robust choice for ordinal scenarios.” This topic (underpowering audit of multi-tier consistency studies at the literature level) has not yet entered the corresponding L1 network; E-060 is the nearest adjacent edge.


    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.