Author: joe

  • 保守口径值多少钱:按例数报自由度,置信区间到底比标定折减宽多少

    Evidence level: C

    例数口径的代价不是 9%,而是 2% 到 3%;它不是由相关性强弱决定,而是由例数多少决定——把区间做窄的最快办法是加例数,不是标定折减系数。

    正文

    上一轮 #1559 给出了一个让人安心的结论:气功外气判读的群体均值检验,从「按例数报自由度」这个保守口径起步,置信区间覆盖率在所有会话内相关水平上都稳稳停在 0.95,标定错也不会伤到结论。但保守是有价格的:把每例的重复遍数只当成一个数来用,那些遍数携带的信息是不是被白白扔掉了?backlog 的预设估算是「当会话内相关只有 0.1、折减系数 0.83 时,CI 比标定后折减口径宽约 9%」。本轮就用解析推导加蒙特卡洛仿真,把这个价格精确量出来,并给出「标定实验到底值不值得做」的量化门槛。

    先看价格的结构。两个口径估计的是同一个量——群体均值的标准误,而标准误的方差标度两侧完全相同:例数口径用每例均值,其方差自动携带设计效应 (1+(r−1)ρ)/r;折减口径用全部读数加设计效应校正,方差标度也是同一个式子。也就是说,保守口径的「浪费」不在方差标度上,只在一个地方:自由度。例数口径 df = n−1,折减口径 df = n·r/(1+(r−1)ρ) − 1,两者的 t 分位数不同,宽度比几乎就等于两个 t 分位数的比值。这是一个可以手算的结论,仿真只是确认它。

    仿真网格覆盖 r = 2/3/5/10、ρ = 0.05/0.1/0.2/0.3、n = 5 到 50,每组 20 万次重复,另加一层「覆盖率校平」的公平比较——把折减口径的区间统一缩放到实际覆盖 0.95 后再比宽度,消除任何虚胖精度。结果分三层。第一层,主工作区(n=30、r=3、ρ<0.3)的代价只有 2.0% 到 2.2%,即使把遍数拉到 10、相关压到 0.05,也才 3.0%;n=20 时不超过 5%。backlog 预设的「约 9%」明显高估了——9% 那个数字其实是设计效应相对「把遍数当独立观测」这一错误名义口径的膨胀率,不是保守口径相对标定口径的代价,两者被混成了一条线。

    第二层,代价几乎与 ρ 无关,只与 n 有关。同一个 (n, r) 下,ρ=0.05 和 ρ=0.3 的宽度比相差不足 1 个百分点。原因还是结构性的:ρ 以同样的方式进入两侧口径的方差标度,互相抵消,剩下的只有它对折减自由度的微弱拉动。换句话说,「高遍数低相关场景代价最大」这个直觉只说对了一半——高遍数确实放大代价(r=10 比 r=3 贵约 0.8 个百分点),但低相关本身几乎不抬价。真正让代价起飞的是例数太少:n=10 时代价跳到 8% 到 11%,n=7 时 12% 到 22%,n=5 时高达 21% 到 37%。

    第三层,代价超过 25% 的区域确实存在,但只出现在 n 不超过 5 的极端小样本(r≥5 时 n=5 超线,r=3 时 n=4 超线)。这个区域里最便宜的解决方案不是标定折减系数,而是加例数:n 从 5 加到 7,代价立刻从 25% 以上掉到 12% 到 20%,加两个例子的收益远超任何一次标定实验。而且要注意,上述对比还假设标定后的 ρ 是已知常数——真实标定实验里 ρ 本身带估计误差,折减口径的实际收益还要再打折扣,保守口径的相对优势只会更大。

    落到立项判断上,backlog 的两条判据有明确答案。常用工作区(r=3、ρ<0.3、n≥20)实测代价 2% 到 5%,远低于 10% 门槛——「例数口径起步规则全场景可用」成立,标定实验正式降级为研究性工作,不设立项优先级。代价超 25% 的区域存在但极其狭窄(n≤5),且该区域的正确解是先把例数补到 7 以上而不是去标定。把两条合起来:标定实验从「最好有」降为「研究性」,新研究一律例数口径起步,样本量按 n(例数)而不是 n×r(例数乘遍数)计算。信度样本量规划领域确有成熟精度框架(PMID 24338600,2014),样本量对重测信度估计的实质影响也有 2025 年的直接证据(PMID 40119099),这两个方法恰好可以反过来用于回答「多少例才够」:把仿真里的 n 换成招募成本,就是每个项目自己的标价表。

    这个结论有一层更稳的底。本系列 #1559 已经证明例数口径的覆盖率不依赖 ρ,本轮又证明它的宽度代价也几乎不依赖 ρ——两个性质加在一起,意味着这套规则对「判读者的会话内相关到底是多少」这个未知数完全免疫。在一个连 ρ 取值范围都只能靠猜的研究领域(#1559 引用的压力痛阈文献报过 0.73 的会话内重复相关,PMID 33767329),能有一把对未知参数天然免疫的尺子,本身就是最值钱的资产的来源。剩下的唯一缺口是:例数口径把判读遍数的贡献压缩在每例均值里,代价以 t 分位数的形式存在——这个缺口是数学结构,不是工程缺陷,补它只能靠例数,这也是全系列最终给出的一致处方。

    证据与局限

    已验证: 蒙特卡洛(20 万次/组,覆盖率校平后公平比较)给出全网格代价:n≥20 全部低于 10%,主工作区 2% 到 5%;代价随 n 单调上升、随 ρ 近似不变;超 25% 区仅存于 n≤5。解析层确认宽度比≈两口径 t 分位数之比,方差标度两侧相同。
    部分验证: 折减口径假设 ρ 已标定且为常数;真实标定带估计误差,实际折减收益只会更小,方向性结论模型无关。
    待验证: 气功外气判读场景的会话内相关实测值仍未取得(同 #1559/#1564);若未来实测 ρ 显示判读者间折减差异悬殊,「折减随读者经验漂移」分支仍需单独裁决。

    相关阅读

    L1 网络状态:该主题(例数口径 vs 标定折减口径的精度代价)在 CORRESPONDENCE_NETWORK / EDGE_TABLE 中无对应边;检索「折减/有效自由度/例数/精度代价/标定」零命中。最近邻为 E-060 标签前提边(conflict,讲人际一致性低,PMID 30383398)与 KB-510(深呼吸 HRV 静息 ICC 0.50–0.83,讲读数稳定性),均为「读数信度」邻近议题而非本命题的会话内结构命题。该主题尚未进入 L1 对应网络,本文证据等级封顶 C。

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


    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 the conservative limit cost: reporting degrees of freedom by count, and how much wider is the confidence interval compared to the calibrated reduction

    Evidence level: C

    The cost of the sample-count approach is not 9%, but rather 2% to 3%; it is determined by the number of samples, not by the strength of correlation—the fastest way to narrow the interval is to add samples, not to apply a calibration discount factor.

    Body

    The previous post #1559 provided a reassuring conclusion: the population mean test for Qigong external air judgment, starting from the conservative “sample-count” approach, maintained a confidence interval coverage rate of 0.95 across all levels of within-session correlation, and calibration errors did not harm the conclusion. However, conservatism comes at a price: treating the repetition count per sample merely as a number means that the information carried by those repetitions is seemingly wasted. The backlog’s preset estimate was “when within-session correlation is only 0.1 and the discount factor is 0.83, the CI is about 9% wider than the calibrated-discounted approach.” This round uses analytical derivation plus Monte Carlo simulation to precisely quantify this cost and provide the quantitative threshold for “whether calibration experiments are actually worth doing.”

    First, let’s examine the structure of the cost. Both approaches estimate the same quantity—the standard error of the population mean—and the variance scaling factor is identical on both sides: the sample-count approach uses the mean per sample, whose variance automatically carries the design effect (1+(r−1)ρ)/r; the calibrated-discount approach uses all readings plus design effect correction, and the variance scaling factor is the same formula. Thus, the “waste” of the conservative approach does not lie in the variance scaling factor, but solely in one aspect: degrees of freedom. For the sample-count approach, df = n−1; for the calibrated-discount approach, df = n·r/(1+(r−1)ρ) − 1. Since the t-distribution quantiles differ, the width ratio is almost exactly the ratio of these two t-quantiles. This is a conclusion that can be calculated by hand; the simulation merely confirms it.

    The simulation grid covers r = 2/3/5/10, ρ = 0.05/0.1/0.2/0.3, n = 5 to 50, with 200,000 repetitions per group, plus an additional layer of “coverage leveling” for fair comparison—scaling the intervals of the calibrated-discount approach uniformly to actually cover 0.95 before comparing widths, eliminating any apparent precision inflation. The results fall into three layers.

    Layer 1: In the primary working area (n=30, r=3, ρ<0.3), the cost is only 2.0% to 2.2%; even when pulling repetitions up to 10 and correlation down to 0.05, it is only 3.0%; for n=20, it does not exceed 5%. The backlog’s preset “about 9%” clearly overestimates—the 9% figure actually represents the inflation rate of the design effect relative to the erroneous nominal approach of treating repetitions as independent observations, not the cost of the conservative approach relative to the calibrated-discount approach; these two were conflated into a single line.

    Layer 2: The cost is almost independent of ρ and depends primarily on n. Under the same (n, r), the width ratio difference between ρ=0.05 and ρ=0.3 is less than 1 percentage point. The reason remains structural: ρ enters the variance scaling factor of both approaches in the same way, canceling each other out, leaving only its weak pull on the discounted degrees of freedom. In other words, the intuition that “high repetitions with low correlation incur the highest cost” is only half right—high repetitions do amplify the cost (r=10 costs about 0.8 percentage points more than r=3), but low correlation itself hardly raises the price. What truly drives the cost up is too few samples: at n=10, the cost jumps to 8%–11%; at n=7, it ranges from 12% to 22%; at n=5, it reaches as high as 21%–37%.

    Layer 3: Regions where the cost exceeds 25% do exist, but only in extreme small-sample cases where n ≤ 5 (when r≥5, n=5 crosses the line; when r=3, n=4 crosses the line). In this region, the cheapest solution is not calibrating the discount factor, but adding samples: increasing n from 5 to 7 immediately drops the cost from over 25% down to 12%–20%; the benefit of adding two samples far exceeds any single calibration experiment. Moreover, note that the above comparison assumes ρ is a known constant after calibration—in real calibration experiments, ρ itself carries estimation error, so the actual gain of the calibrated-discount approach is further discounted, making the relative advantage of the conservative approach even greater.

    Turning to project initiation decisions, the backlog’s two criteria have clear answers. In the common working area (r=3, ρ<0.3, n≥20), the measured cost is 2%–5%, far below the 10% threshold—”the sample-count starting rule is applicable across all scenarios” holds true, and calibration experiments formally downgrade to research-level work, not setting project priority. The region where cost exceeds 25% exists but is extremely narrow (n≤5), and the correct solution there is to first supplement samples to more than 7 rather than calibrate. Combining these two points: calibration experiments drop from “best to have” to “research-level,” and all new studies should start with the sample-count approach, calculating sample size based on n (number of samples) rather than n×r (number of samples multiplied by repetitions per sample).

    The reliability sample size planning field indeed has mature precision frameworks (PMID 24338600, 2014), and there is direct evidence from 2025 on the substantive impact of sample size on retest reliability estimation (PMID 40119099); these two methods can be turned around to answer “how many samples are enough”: replacing n in the simulation with recruitment cost yields each project’s own price list.

    This conclusion has an even more solid foundation. Post #1559 in this series already proved that the coverage rate of the sample-count approach does not depend on ρ; this round further proves that its width cost is also almost independent of ρ—these two properties combined mean this set of rules is completely immune to the unknown parameter “what is the within-session correlation for the judge.” In a research field where even the range of ρ can only be guessed (the pressure pain threshold literature cited in #1559 reported within-session repetition correlation of 0.73, PMID 33767329), having a ruler naturally immune to unknown parameters is itself the source of the most valuable asset. The only remaining gap is: the sample-count approach compresses the contribution of judgment repetitions into the mean per sample, with the cost manifesting as t-distribution quantiles—this gap is a mathematical structure, not an engineering defect; filling it can only be done by adding samples, which is also the consistent prescription given throughout this entire series.

    Evidence and Limitations

    Verified: Monte Carlo (200,000 per group, with coverage leveling for fair comparison) yields full-grid costs: n≥20 all below 10%, primary working area 2%–5%; cost increases monotonically with n and remains approximately constant with ρ; regions exceeding 25% exist only for n≤5. Analytical layer confirms width ratio ≈ ratio of t-quantiles, variance scaling factors identical on both sides.
    Partially Verified: The calibrated-discount approach assumes ρ is already calibrated and constant; real calibration carries estimation error, so actual discounted gain will be even smaller, making the directional conclusion model-independent.
    Pending Verification: Measured within-session correlation values for Qigong external air judgment scenarios have not yet been obtained (same as #1559/#1564); if future measured ρ shows large differences in discounting between judges, the “discount varies with reader experience” branch still requires separate adjudication.

    Related Reading

    L1 Network Status: This topic (cost of sample-count vs. calibrated-discount approaches) has no corresponding edge in CORRESPONDENCE_NETWORK / EDGE_TABLE; search for “discount/effective degrees of freedom/sample-count/precision cost/calibration” yields zero hits. Nearest neighbors are E-060 label premise edge (conflict, discussing low interpersonal consistency, PMID 30383398) and KB-510 (deep breathing HRV resting ICC 0.50–0.83, discussing reading stability), both concerning “reading reliability” rather than this topic’s within-session structural proposition. This topic has not yet entered the L1 corresponding network; evidence level capped at C.

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


    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.

  • 标定出来的信度本身也有误差:折减口径的收益会被侵蚀掉几成

    Evidence level: B

    折减口径把”标定出的相关系数”当成已知数代入公式省钱,但标定实验自己也是个样本——这个系数本身就带估计误差,误差一旦代入,折减省下来的样本量会被吃掉几成,这条线还能不能走要靠仿真标价。

    正文

    上一篇文章说过,我们的”有效自由度折减口径”有一张全网格标价表:只要知道判读者之间(或会话之间)的相关系数 ρ,就能按设计效应把例数口径的样本量打一个折,省下的钱就是折减口径的全部卖点。但那张表有一个隐藏假设——ρ 是已知常数。

    真实世界里 ρ 从天上掉不下来,它必须由一次”标定实验”估出来:招一批受试者、安排若干个会话、重复读数,算出一个 ρ̂。而 ρ̂ 本身是个统计量,会话数少的时候误差不小。信度方法学里这恰恰是老问题:Walter 与 Bonett 的精度框架给出了”想要把 ρ 的置信区间压到指定宽度需要多少受试者与多少次重复”的闭式算法(PMID 24338600),2025 年的行为科学综述进一步提醒:样本量不足时重测信度的估计会系统性失真(PMID 40119099)。换句话说,标定实验自己也要算样本量,省钱的折减口径在入口处就先花了一笔钱。

    更麻烦的是误差的传导路径。群随机试验文献早就踩过这个坑:样本量公式里必须代入一个 ICC,而 ICC 往往只来自小规模预试验,估计的不确定性对最终样本量和功效影响很大——Lewis 与 Julious 的数值方法显示,若不管不确定性,主试验可能严重欠功效;而常用的”取置信上限”做法虽然安全,却经常把试验设计得过度膨胀、低效(PMID 34477455)。2025 年提出的混合保证(assurance)方法干脆把 ICC 的不确定性用先验分布喂进设计,用”平均成功概率”替代单点功效(PMID 39935065);更早的贝叶斯分层模型则把多个历史试验的 ICC 信息合并出后验分布来定样本量(PMID 37036110)。Walter 团队 2001 年的经典结论同样指向这一点:信度研究应该按”ρ 置信区间的预期宽度”来规划,而不是按点估计(PMID 11746313)。2024 年的一篇系统综述把这些方法在单因素方差分析模型下排了个队,指出最常用的 Wald 大样本区间恰恰是统计性质最差的那种(PMID 38320802)。

    把这些经验折回我们的问题:折减口径把 ρ̂ 代入设计效应后,SE 的标度不再是对真 ρ 的精确校正——ρ̂ 偏高时折减过猛、区间偏窄、覆盖率掉;ρ̂ 偏低时折减过弱、白白多花钱。折减口径对例数口径的宽度优势,会随标定误差增大而单调被侵蚀。所以核心判据必须量化:在原有仿真网格上让 ρ̂ 从”以 n_s 个标定会话估出的近似正态分布”里抽取(se_ρ 按精度框架闭式给出),对比”ρ 已知 / ρ̂ 带误差”两种折减口径相对例数口径的宽度比。若 n_s≥20 时带误差的折减口径仍保留已知 ρ 情形收益的七成以上,标定实验维持研究性定位、并给出最小标定样本量 n_s≥20;若收益保留不足一半,标定路线整体出局,例数口径成为唯一推荐口径。

    侵蚀的形态还可以预判三档,免得仿真跑完才发现问题问错了方向。第一档是不对称侵蚀:ρ̂ 的抽样分布在低 ρ 区是右偏的,标定会话越少,ρ̂ 偶然偏高的机会越大,而设计效应对 ρ 的放大恰好是非线性的——偏高带来的”过度折减”对区间的伤害,比偏低带来的”少省几个样本”更重,所以侵蚀大概率不是对称的均值问题而是尾部问题。第二档是覆盖与宽度的解耦:折减省下的宽度若以牺牲覆盖率为代价,名义上更窄的置信区间实际上不再有 95% 的信誉,这时”宽度优势”本身就不再是收益,判据里必须同时盯宽度和覆盖两条线,只看宽度会奖励偷工减料。第三档是最小标定样本量的台阶效应:精度框架(PMID 24338600)给出的 se_ρ 随会话数衰减是先陡后缓的,n_s 从 5 加到 20 时误差砍掉大半,从 20 加到 40 却所省无几——若仿真实测出这条台阶,最小标定样本量就不该拍脑袋定 20,而应落在台阶拐点附近。这三档形态哪一档主导,直接决定标定路线是”加几场会话就能救”还是”结构性没戏”。

    这与易学测量线(气功外气判读、丹田三模态、得气量表)的关系是同一句话:用第一人称与生理读数去标定构型相关系数这条路,成本不只是标定实验本身,还包括误差传导进下游每一个样本量决策。东方修炼传统里”功夫深浅”历来难以量化,现代测量学给出的答案是:量化可以,但每个量化环节的误差都要进账本——省样本量的捷径,要先付标定的学费。

    证据与局限

    已验证: ICC 估计的抽样误差对样本量/功效决策有实质影响(PMID 34477455、39935065、37036110);信度研究的样本量可按精度闭式规划(PMID 24338600、11746313);小样本重测信度估计会失真(PMID 40119099)。

    部分验证: “取置信上限”保守但低效的折中策略在群随机试验中已量化(PMID 34477455);折减口径的侵蚀幅度需要本项目仿真标价,方向可由文献支撑、数值结果待产出。

    待验证: n_s≥20 是否是标定实验的最小样本量;侵蚀后收益保留率的具体曲线形态;折减口径与混合保证方法(PMID 39935065)在本项目语境下的优劣对比。

    相关阅读

    证据等级: B
    参考文献: PMID 34477455, PMID 39935065, PMID 38320802, PMID 24338600, PMID 11746313, PMID 40119099
    L1 网络状态:该主题尚未进入 L1 对应网络(EDGE_TABLE 与 CORRESPONDENCE_NETWORK 中无「自由度折减/标定误差」对应边;最近邻为 E-060 标签信度,状态 conflict——中医医师间诊断一致性低且一致性统计常被误用,PMID 30383398/31730403)。


    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 reliability itself has error: how much of the reduced caliber’s gain will be eroded?

    Evidence level: B

    The reduction scale substitutes the “calibrated correlation coefficient” as a known constant into the formula to save costs. However, the calibration experiment itself constitutes a sample, and this coefficient inherently carries estimation error. Once this error is introduced, the sample size savings achieved by the reduction scale are eroded by a certain proportion. Whether this path remains viable depends on simulation-based pricing.

    Main Text

    As mentioned in the previous article, our “effective degrees of freedom reduction scale” includes a full-grid pricing table: knowing the correlation coefficient ρ between raters (or sessions) allows one to discount the sample size based on the design effect, saving money which represents the sole selling point of the reduction scale. However, that table relies on a hidden assumption—ρ is a known constant.

    In reality, ρ does not fall from the sky; it must be estimated via a “calibration experiment”: recruiting a batch of subjects, arranging several sessions, repeating readings, and calculating an estimated ρ̂. Since ρ̂ itself is a statistic, the error can be substantial when the number of sessions is small. In reliability methodology, this is precisely a longstanding issue: Walter and Bonett’s precision framework provides a closed-form algorithm for “how many subjects and repetitions are needed to shrink the confidence interval of ρ to a specified width” (PMID 24338600). A 2025 behavioral science review further warns that insufficient sample sizes lead to systematic distortion in the estimation of test-retest reliability (PMID 40119099). In other words, the calibration experiment itself requires sample size calculation, meaning the cost-saving reduction scale incurs an upfront expense at the entrance.

    Complicating matters further is the propagation path of error. The literature on cluster randomized trials has long encountered this pitfall: sample size formulas must substitute an ICC, yet the ICC often originates from small-scale pilot studies, and estimation uncertainty significantly impacts the final sample size and power. Lewis and Julious’s numerical methods show that ignoring uncertainty may result in the main trial being severely underpowered; while the common practice of “taking the confidence upper limit” is safe, it frequently leads to an overly inflated and inefficient trial design (PMID 34477455). A hybrid assurance method proposed in 2025 feeds the uncertainty of the ICC into the design via prior distributions, replacing single-point power with “average success probability” (PMID 3993605); earlier Bayesian hierarchical models combine ICC information from multiple historical trials to derive posterior distributions for sample size determination (PMID 37036110). Walter’s team also reached a classic conclusion in 2001: reliability studies should be planned based on the “expected width of the ρ confidence interval” rather than point estimates (PMID 11746313). A systematic review from 2024 ranks these methods under a single-factor ANOVA model, noting that the commonly used Wald large-sample interval is actually the one with the worst statistical properties (PMID 38320802).

    Translating these experiences back to our problem: after the reduction scale substitutes ρ̂ into the design effect, the scaling of the standard error (SE) is no longer an exact correction for the true ρ. When ρ̂ is high, the reduction is excessive, leading to narrower intervals and reduced coverage; when ρ̂ is low, the reduction is insufficient, wasting money. The width advantage of the reduction scale over the example scale is monotonically eroded as calibration error increases. Therefore, the core criterion must be quantified: on the original simulation grid, draw ρ̂ from an approximate normal distribution estimated by n_s calibration sessions (where se_ρ is given by the closed-form precision framework), and compare the width ratio of the reduction scale relative to the example scale under “known ρ” versus “ρ̂ with error.” If n_s ≥ 20, the reduction scale with error retains over 70% of the benefit of the known ρ scenario, supporting the research-oriented positioning of the calibration experiment and establishing a minimum calibration sample size of n_s ≥ 20; if the retained benefit falls below half, the calibration route is entirely ruled out, and the example scale becomes the sole recommended option.

    Three eroded forms can be anticipated to avoid discovering the wrong direction only after running simulations. The first is asymmetric erosion: the sampling distribution of ρ̂ is right-skewed in low ρ regions; with fewer calibration sessions, the chance of ρ̂ being spuriously high increases, while the design effect’s amplification of ρ is nonlinear—so the “over-reduction” caused by high estimates inflicts greater harm on the interval than the “under-saving” caused by low estimates. Thus, erosion is likely not a symmetric mean issue but a tail problem. The second is decoupling of coverage and width: if the width saved by reduction comes at the cost of reduced coverage, then although the nominal confidence interval appears narrower, it no longer carries 95% credibility; here, “width advantage” itself ceases to be a benefit, and the criterion must monitor both width and coverage simultaneously—focusing only on width rewards cutting corners. The third is the step effect of minimum calibration sample size: the precision framework (PMID 24338600) shows that se_ρ decays with the number of sessions in a pattern that is steep initially then flattens; increasing n_s from 5 to 20 cuts error by half, but going from 20 to 40 saves almost nothing. If simulation reveals this step, the minimum calibration sample size should not be arbitrarily set at 20 but placed near the inflection point of the step. Which of these three forms dominates directly determines whether the calibration route is “saved by adding a few sessions” or “structurally hopeless.”

    This connects to the Yijing measurement line (qigong external qi reading, dantian tri-modal, qi-acquisition scale) in one sentence: using first-person physiological readings to calibrate the configuration correlation coefficient incurs costs not only from the calibration experiment itself but also from error propagation into every downstream sample size decision. In the Eastern cultivation tradition, “depth of practice” has long been difficult to quantify; modern metrology offers the answer: quantification is possible, but every quantification step’s error must be recorded in the ledger—the shortcut to saving sample size requires first paying the tuition for calibration.

    Evidence and Limitations

    Verified: Sampling error in ICC estimation substantially affects sample size/power decisions (PMID 34477455, 3993605, 37036110); reliability study sample sizes can be planned using precision closed-form equations (PMID 24338600, 11746313); small-sample test-retest reliability estimates are prone to distortion (PMID 40119099).

    Partially Verified: The conservative yet inefficient compromise of “taking the confidence upper limit” has been quantified in cluster randomized trials (PMID 34477455); the degree of erosion for the reduction scale requires simulation-based pricing in this project, though direction is supported by literature and numerical results pending output.

    To Be Verified: Whether n_s ≥ 20 is indeed the minimum sample size for the calibration experiment; the specific curve shape of retained benefit after erosion; and the comparative advantages/disadvantages of the reduction scale versus hybrid assurance methods (PMID 3993605) within this project’s context.

    Further Reading

    Evidence Level: B
    References: PMID 34477455, PMID 3993605, PMID 38320802, PMID 24338600, PMID 11746313, PMID 40119099
    L1 Network Status: This topic has not yet entered the L1 corresponding network (no edges for “degrees of freedom reduction/calibration error” in EDGE_TABLE and CORRESPONDENCE_NETWORK; nearest neighbor is E-060 label reliability, status conflict—diagnostic consistency among TCM physicians is low and consistency statistics are often misused, PMID 30383398/31730403).


    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.

  • 维持组「没掉」里有多少是肌肉保住了、多少是第二次测得更准

    Evidence level: B

    减量训练研究里”每周练一次就够保住”的结论,可能有一部分是测量学习效应伪装的——因为所有”保住”的判断都建立在第二次测量上,而第二次测量天然比第一次测得准。

    练功的人最怕一个字:退。气功讲”功不练则散”,健身讲”停练就掉肌肉”。于是现代研究给了一个让人安心的答案:减量训练(tapering/detraining)研究发现,从每周多次降到每周一次,力量和耐力大部分能”保住”。这听起来像”留一手”的科学背书——但其实这个结论有一个很少被戳破的软肋:它全部建立在”第二次测量”上。

    问题出在哪:第二次测量天然更准

    呼吸肌训练研究里最常用的指标是最大吸气压(PImax)——使劲吸一口气,看能产生多大负压。这个动作高度依赖技巧:嘴唇怎么封住接口、多深预备、多快发力,都是练出来的。1998 年的经典实验(PMID 9769281)让 18 名健康人隔天做三次呼吸肌耐力测试,结果耐力成绩从第一次到第三次系统性上升 69%→77%→86%(占 PImax 的比例),而呼吸肌力量本身(PImax)没变——上升的部分不是肌肉变强,是受试者学会了怎么测,主观吃力度(Borg 评分)在相同负荷下随次数下降,说明早期成绩是被”不舒服”压住的,不是被肌肉能力压住的。

    在慢阻肺患者身上同样如此(PMID 14959807):十名 COPD 患者做四次测试,PImax 四次里涨了 21%,耐力指标涨了 32%,到第三次才平台。也就是说,一个人的”进步”里混着两种东西:训练带来的真进步,和学会测量带来的假进步。

    现在把这个放进减量训练研究的框架里看:典型的设计是”训练 12 周 → 测一次(基线)→ 减量或停练 → 再测一次(随访)”,比较两次的差。训练组随访时已经测过很多次,动作熟练;但如果把整个”测量学习”都发生在训练期里,那基线那次(第一次或前几次)恰恰是最不准的——基线被低估,随访显得”保住了”,其中有多少是真的肌肉保持、多少只是第二次测得更准?已发表的减量训练研究,没有一个系统扣过这一项。热身实验给了佐证:先做亚极量呼吸肌活动,能把重复测量的学习效应从 11.4% 压到 3.2%(PMID 11223726)——说明学习效应是个可以被操纵的实在变量,不是统计噪声。

    用易学的话说:考功和养功不是一回事

    传统武术和气功里其实早有这个区分的直觉:“师父考功”和”日常养功”是两件事。老师傅考核徒弟时不看一次发挥,要看”稳定盘”——因为一次测试里有临场发挥、紧张、生疏的成分。用现代话讲:传统体系默认”测量结果里有测量者与被测者的共同学习痕迹”,所以用重复考核、多场景考核来把技巧性成分摊平。减量训练研究恰恰在这里偷了懒:它把”一次会话内的最好成绩”当成肌肉能力的无偏读数,而呼吸肌测试文献自己早就证明了这不是无偏的。

    对气功研究还有一个更直接的推论:凡是用”修炼年限”做横断面分层的研究(练 10 年的人比新手呼吸肌强多少),都同时混着”练出来的真差异”和”测得多的熟练差异”——修炼者一年测几十次,新手一辈子第一次进实验室。这不只是减量训练的问题,是所有气功生理测量的通病。

    怎么把两种”保住”拆开

    设计上并不难:取足够人数(如 ≥90 人)完成 12 周训练后,随机分维持组与完全停练组,两组在停练前、停练 6/12/24 周各测三次,而且每次会话内连续三次尝试单独记录成序列。然后用线性混合模型把”会话序号”(跨天的学习)和”尝试序号”(会话内的学习)两个学习项先估计出来,再从组间差里扣。核心判据:若扣掉学习项后维持组的优势从显著变成不显著,”每周一次就够”就是测量伪迹;若扣完优势仍在且保留七成以上,减量训练才算真正成立;若维持组的学习项反而比停练组大(他们一直在练、更熟练),说明学习效应被误当成了组间差异,所有观察性减量研究都得重估。

    这个设计的价值不止于呼吸肌:它是给”传统体系里’功夫会不会退’这个老问题”配的一个测量学防呆装置——先承认测量本身会被学习污染,再谈保住了多少。

    证据与局限

    已验证: 呼吸肌耐力与力量测量存在系统性学习效应,且来源是呼吸模式改变与主观吃力下降而非肌肉条件化(PMID 9769281);COPD 人群同样存在,PImax 四次测试升 21%、第三次才平台(PMID 14959807);亚极量热身可把重复测量学习效应从 11.4% 压到 3.2%,证明学习效应可操纵(PMID 11223726)。

    部分验证: 学习效应”伪装成维持”的具体幅度在减量训练文献中从未被量化——上述判据(扣除后优势是否消失)目前只是设计提案,没有已发表的验证结果。

    待验证: ① 已发表减量训练研究在扣除会话序号与尝试序号学习项后,”每周一次维持”结论的存活率;② 气功修炼年限分层研究中,熟练测试者偏差对”功龄效应量”的污染比例;③ 停练组与维持组学习曲线斜率是否系统不同。

    相关阅读

    证据等级: B
    参考文献: PMID 9769281, PMID 14959807, PMID 11223726
    L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 与 EDGE_TABLE_science-yixue.md 检索「呼吸肌/减量训练/detraining/最大吸气压/MIP/学习效应」零命中;最近邻边为「丹田↔呼吸共振」(EDGE_TABLE 收录 51 条、A 级证据),仅覆盖丹田呼吸的共振测量线,不覆盖”减量训练维持率的学习效应混杂”命题。本篇按规则以方法学批判立场写作,未将任何桥接写成已验证机制。


    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 many of the maintained group’s results are due to preserved muscle mass versus improved accuracy on the second measurement

    Evidence level: B

    In reduction training research, the conclusion that “training once per week is sufficient to maintain” may be partially masked by measurement learning effects—because all judgments of “maintenance” rely on the second measurement, which is inherently more accurate than the first.

    Practitioners fear one word above all: decline. Qigong states “功不练则散” (skill dissipates without practice), while fitness says “停练就掉肌肉” (stopping training leads to muscle loss). Modern research thus offers a reassuring answer: reduction training (tapering/detraining) studies find that most strength and endurance can “be maintained” when shifting from multiple weekly sessions to just one. This sounds like scientific backing for “keeping something in reserve”—but this conclusion has a soft spot rarely exposed: it relies entirely on the “second measurement”.

    Where the Problem Lies: The Second Measurement Is Naturally More Accurate

    The most common metric in respiratory muscle training studies is maximal inspiratory pressure (PImax)—the negative pressure generated by taking a deep breath as hard as possible. This action is highly skill-dependent: how the lips seal the interface, how deep the preparation, and how fast the effort—all require practice. A classic 1998 experiment (PMID 9769281) had 18 healthy participants perform respiratory muscle endurance tests three times every other day. Endurance scores systematically rose from 69% to 77% to 86% of PImax across the three trials, while PImax itself remained unchanged—the increase was not due to stronger muscles but rather participants learned how to measure, with subjective effort (Borg rating) decreasing under the same load, indicating early scores were suppressed by discomfort, not muscle capacity.

    The same holds true in COPD patients (PMID 14959807): ten COPD patients underwent four tests, with PImax rising 21% over the four trials, endurance metrics up 32%, plateauing only by the third test. Thus, an individual’s “progress” mixes two components: real improvement from training and false improvement from learning to measure.

    Placing this into the tapering framework: typical design is “train 12 weeks → measure once (baseline) → reduce or stop training → measure again (follow-up)”, comparing the difference. The training group has already measured repeatedly, so their technique is refined; but if all “measurement learning” occurs during the training period, then the baseline (first or early tests) is actually the least accurate—baseline is underestimated, making follow-up appear “maintained.” How much of this is true muscle preservation and how much is just better measurement? No published tapering study has systematically accounted for this. A warm-up experiment provides support: pre-testing with submaximal respiratory activity reduces the repeated-measurement learning effect from 11.4% to 3.2% (PMID 11223726)—showing learning effects are manipulable real variables, not statistical noise.

    In Yijing Studies Terms: Examining Skill and Cultivating Skill Are Not the Same

    Traditional martial arts and qigong already have an intuitive distinction: “Master examines disciple” vs. “daily cultivation” are two different things. When a master evaluates a disciple, they do not judge based on a single performance but look at the “stable plate”—because one test includes momentary exertion, nervousness, and unfamiliarity. In modern terms: traditional systems assume “measurement results contain shared learning traces between measurer and measured,” so they use repeated assessments across multiple scenarios to average out technical components. Tapering studies cut corners here: they treat the “best score within one session” as an unbiased reading of muscle capacity, yet respiratory muscle literature itself proves this is not unbiased.

    For qigong research, there is an even more direct implication: any cross-sectional study stratified by “years of practice” (e.g., comparing those who trained for 10 years to novices) simultaneously mixes “real differences from training” with “familiarity differences from repeated testing”—practitioners test dozens of times per year, while novices enter the lab only once in their lifetime. This is not just a tapering issue but a universal flaw in qigong physiological measurement.

    How to Separate the Two Types of “Maintenance”

    The design is straightforward: recruit sufficient participants (e.g., ≥90) to complete 12 weeks of training, then randomly assign them to a maintenance group or a complete stop group. Both groups undergo three measurements before stopping, and again at 6/12/24 weeks post-stop. Crucially, within each session, the three consecutive attempts are recorded as a sequence. Linear mixed models then estimate two learning terms—”session number” (cross-day learning) and “attempt number” (within-session learning)—before subtracting them from the group difference. The core criterion: if the maintenance group’s advantage becomes non-significant after removing learning terms, then “once per week is enough” is a measurement artifact; if the advantage remains and accounts for over 70%, tapering truly holds; if the maintenance group’s learning term is larger than the stop group (they kept training and became more skilled), then learning effects were mistaken for group differences, requiring re-evaluation of all observational tapering studies.

    This design’s value extends beyond respiratory muscles: it provides a metrological fail-safe for the age-old question “Will traditional skill regress?”—first acknowledge that measurement itself is contaminated by learning, then discuss how much was truly preserved.

    Evidence and Limitations

    Verified: Respiratory muscle endurance and strength measurements exhibit systematic learning effects, stemming from altered breathing patterns and reduced subjective effort rather than muscle conditioning (PMID 9769281); COPD populations similarly show this, with PImax rising 21% over four tests and plateauing only by the third (PMID 14959807); submaximal warm-up reduces repeated-measurement learning effects from 11.4% to 3.2%, proving learning effects are manipulable (PMID 11223726).

    Partially Verified: The specific magnitude by which learning effects “masquerade as maintenance” has never been quantified in tapering literature—the above criteria (whether advantage disappears after subtraction) remain design proposals without published verification.

    To Be Verified: ① Survival rate of the “once per week maintenance” conclusion after subtracting session and attempt learning terms in published tapering studies; ② Proportion of “practice-year effect size” contamination by test-familiarity bias in qigong cross-sectional studies; ③ Whether learning curve slopes differ systematically between stop and maintenance groups.

    Further Reading

    Evidence Level: B
    References: PMID 9769281, PMID 14959807, PMID 11223726
    L1 Network Status: This topic has not yet entered the L1 correspondence network—zero hits in CORRESPONDENCE_NETWORK.md and EDGE_TABLE_science-yixue.md for “respiratory muscle/tapering/detraining/MIP/learning effect”; nearest neighbors are “Dantian ↔ respiratory resonance” (51 entries in EDGE_TABLE, Level A evidence), covering only the resonance measurement line of Dantian breathing, not the proposition “learning effect contamination in tapering maintenance rate.” This piece is written from a methodological critique stance, without writing any bridge as a verified mechanism.


    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.

  • 得气量表的会话内相关会不会被气感强度分层搅动

    Evidence level: C

    强气感与弱气感子群的读数稳定性可能根本不是同一量级——不分层就把得气量表当一条构型处理,算出的会话内相关可能是「平均数骗人」。

    正文

    跨构型标定体系里,得气量表作为一条「构型」参与比较,其折减系数由会话内相关推出。这里有一个被默认掉的前提:整条构型的会话内相关是一个单一量,不随人群分层而变。但得气体验的强度在个体之间横跨极大。一项 527 名健康志愿者的研究显示,针刺足三里时,平和体质者的酸痛强度显著高于气虚、阳虚、痰湿、气郁等体质人群(PMID 29071897);一篇得气临床研究综述也明确指出,同样的针刺操作下,不同个体感到的得气程度差异很大,评估工具理应把个体特征纳入(PMID 23486017)。

    强度差异为什么会威胁到信度估计?机制上的担忧是:强气感读者的「信号」强,读数天然更稳定,会话内一致性好;弱气感读者接近在猜,读数接近随机波动。把两层混在一起估计,得到的会话内相关是两层的加权平均——它会被弱组拉低。此时报告出来的「得气构型的会话内相关」不再是构型本身的性质,而是人群构成比的函数:换一批人,同一个量表的同一条信度数字就变了。跨构型比较若一端是得气量表、另一端是生理信号构型,这种人群依赖性会直接污染折减系数的公平性。

    L1 对应网络上,这条边正好骑在一个未裁决的冲突上。工具现成一侧:中文版 MASS 量表重测信度达 ICC 0.92(PMID 22617434),SNSQ 量表在 227 人样本上完成验证(PMID 18576920)——量表「能用」有据可查。跨情境失效一侧:SNSQ 德语版在实验条件下因子结构无法复现,且无法把得气与疼痛区分开,重测相关降至 0.54 到 0.60(PMID 22189363);还有研究提醒得气反应不能当盲法成功与否的检验指标(PMID 40619110)。关键在于:两侧都没有报过分层分析。ICC 0.92 是全样本数字,它既不反驳也不支持「强气感子群更稳、弱气感子群接近猜」这个假设——这条具体的研究在文献里至今是空白。

    可做的设计因此很明确:在跨构型标定的得气量表数据上,按基线气感强度把读者分成强弱两组,用分层混合模型各自估计会话内分量。方法学工具也已备好:贝叶斯层次框架里检验个体内方差是否同质已有成熟范式(PMID 34816384),它允许每个人的方差不同,而不是硬套同方差假设——这正好对应「强弱两层方差结构可能不同」的问题形态。

    核心判据事先写死:若强弱两组会话内相关的置信区间重叠且中位数差小于 0.1,则单档折减系数可用,得气构型继续按整体申报;若强组比弱组高 0.2 以上,则得气构型的折减系数必须按体验强度分表申报,样本量表也要加一层体验档位——否则跨构型比较的口径就不成立。中间地带(差 0.1 到 0.2)则需要扩大标定样本再裁决,而不是悄悄选边。

    证据与局限

    已验证:得气体验强度存在显著的个体与体质差异(PMID 29071897,n=527);得气量表工具层面有验证先例,中文 MASS 重测 ICC 0.92(PMID 22617434)。
    部分验证:量表信度跨情境不稳定(德语版 SNSQ 因子结构不可复现、重测相关下降,PMID 22189363);心理因素可影响得气报告。
    待验证:按气感强度分层后,强弱子群的会话内相关是否同一量级——目前无任何直接研究,本文判据为预注册式设计,等待数据裁决。

    相关阅读

    证据等级: C
    参考文献: PMID 29071897, PMID 23486017, PMID 22617434, PMID 22189363, PMID 34816384

    L1 网络状态:E-060 体验通道(得气第一人称报告→量表信度)状态 open,内嵌冲突未裁决——KB-318「工具箱现成,不必自造量表」(C-MMASS 重测 ICC=0.92,PMID 22617434)对撞 KB-323「跨情境不可靠」(SNSQ 德语版因子结构无法复现且无法区分痛,重测降至 r=0.538/0.603,PMID 22189363;得气不能当盲法成功检验指标,PMID 40619110)。本文主题「按气感强度分层估计会话内相关」尚未进入 L1 对应网络,已登记 science-yixue-l1-gaps.md。


    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。

    Will session-relatedness of the Qi-amount scale be disrupted by stratification of Qi-sensation intensity

    Evidence level: C

    The stability of readings within strong-sensation and weak-sensation subgroups may not be on the same magnitude. Treating the Qi-Sensation Scale as a single configuration without stratification could result in an intra-session correlation that is “deceived by the average.”

    Main Text

    Within the cross-configuration calibration framework, the Qi-Sensation Scale participates as a “configuration,” and its reduction coefficient is derived from intra-session reliability. A default assumption underlies this approach: the intra-session correlation of the entire configuration is treated as a single quantity, invariant across population stratification. However, the intensity of Qi sensation varies enormously among individuals. A study involving 527 healthy volunteers found that the soreness intensity during acupuncture at Zusanli was significantly higher in individuals with a balanced constitution compared to those with Qi deficiency, Yang deficiency, Phlegm-Dampness, or Qi stagnation (PMID 29071897). A review of clinical studies on Qi sensation also explicitly states that under identical acupuncture procedures, the degree of Qi sensation experienced differs greatly among individuals, and assessment tools should incorporate individual characteristics (PMID 23486017).

    Why does this intensity variation threaten reliability estimation? The mechanistic concern is that readers with strong Qi sensation produce a stronger “signal,” leading to inherently more stable readings and better intra-session consistency, whereas readers with weak Qi sensation are essentially guessing, resulting in readings close to random fluctuation. Combining these two layers into a single estimate yields an intra-session correlation that is a weighted average of both groups, which will be pulled down by the weak group. Consequently, the reported “intra-session correlation of the Qi configuration” is no longer a property of the configuration itself but rather a function of population composition: changing the sample would alter the same reliability number for the same scale. If cross-configuration comparison involves one end as the Qi-Sensation Scale and the other as a physiological signal configuration, this population dependency directly contaminates the fairness of the reduction coefficient.

    L1 corresponds to the network; this edge straddles an unresolved conflict. On the tool side, the Chinese version of the MASS scale demonstrates a retest reliability ICC of 0.92 (PMID 22617434), and the SNSQ scale has been validated on a sample of 227 individuals (PMID 18576920)—evidence that the scales “work” is documented. On the cross-context failure side, the German version of SNSQ failed to replicate its factor structure under experimental conditions and could not distinguish Qi sensation from pain, causing retest correlations to drop to 0.54–0.60 (PMID 22189383); furthermore, research warns that Qi sensation cannot serve as a criterion for blind success (PMID 40619110). Crucially, neither side reported stratified analysis. The ICC of 0.92 is a whole-sample figure; it neither refutes nor supports the hypothesis that “the strong-sensation subgroup is more stable while the weak-sensation subgroup is near guessing”—this specific gap remains blank in the literature.

    Therefore, the design is clear: for Qi-Sensation Scale data within cross-configuration calibration, readers should be divided into strong and weak groups based on baseline sensation intensity, and intra-session components should be estimated separately using stratified mixed models. Methodological tools are already available: Bayesian hierarchical frameworks have established paradigms for testing whether within-individual variances are homogeneous (PMID 34816384), allowing each person to have different variance rather than rigidly assuming homogeneity—this precisely matches the problem structure of “variance structures may differ between strong and weak layers.”

    The core criterion is fixed in advance: if the confidence intervals of intra-session correlations for the strong and weak groups overlap and the median difference is less than 0.1, a single reduction coefficient applies, and the Qi configuration continues to be declared as a whole; if the strong group exceeds the weak group by more than 0.2, the reduction coefficient for the Qi configuration must be declared separately by sensation intensity, and the sample scale must add an extra sensation tier—otherwise, the cross-configuration comparison framework fails. The intermediate zone (difference between 0.1 and 0.2) requires expanding the calibration sample before decision, rather than silently choosing a side.

    Evidence and Limitations

    Verified: Significant individual and constitutional differences exist in Qi sensation intensity (PMID 29071897, n=527); validated tools exist at the scale level, with the Chinese MASS retest ICC reaching 0.92 (PMID 22617434).
    Partially Verified: Scale reliability is unstable across contexts (German SNSQ factor structure could not be replicated and failed to distinguish pain, causing retest correlation to drop, PMID 22189383); psychological factors can influence Qi sensation reporting.
    To Be Verified: After stratification by sensation intensity, whether the intra-session correlations of strong and weak subgroups are on the same magnitude—currently no direct study exists; this criterion is a pre-registered design awaiting data resolution.

    Further Reading

    Evidence Level: C
    References: PMID 29071897, PMID 23486017, PMID 22617434, PMID 22189383, PMID 34816384

    L1 Network Status: E-060 Experience Channel (Qi sensation first-person report → scale reliability) is open, with embedded conflict unresolved—KB-318 “Tools are ready, no need to create scales” (C-MMASS retest ICC=0.92, PMID 22617434) collides with KB-323 “Unreliable across contexts” (German SNSQ factor structure could not be replicated and failed to distinguish pain, retest dropped to r=0.538/0.603, PMID 22189383; Qi sensation cannot serve as blind success criterion, PMID 40619110). This topic “Estimating intra-session correlation by stratifying sensation intensity” has not yet entered the L1 corresponding network and is registered in science-yixue-l1-gaps.md.


    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.

  • 修正版地板闭式标定:ρ 轴解析级归零,α 的残余漂移全数转移到 m 轴

    Evidence level: C

    把地板公式换成根号(1 加 m 减 1 乘 ρ)形态之后,α 跨 ρ 的漂移被代数精确对消——但「一根线走天下」仍然不成立:残余漂移完整转移到构型数 m 上,m 从 2 到 16 标尺掉了三倍多。

    正文

    先把欠账摆清楚。上一轮重建给出两个硬事实:合并地板判据不跨 ρ 保形(保形偏差 D 从 0.40 到 2.25),而且 D 只依赖(ρ,m)两个量、与自由度 df 无关。当时留了两条出路:保守的一条是按 ρ 分档配倍率表;进取的一条是把地板公式里的相关结构换掉——上轮明确指出,正确的 m 与 ρ 依赖藏在(1 加 m 减 1 乘 ρ)的平方根里,而不是原公式的(根号 c 减 ρ 乘根号 m)里。本文把进取那条路走完:从单位均值方差出发解析重推修正地板,代回原参数网格,逐格检验修正版 α 能不能拉回常数。

    推导从最朴素的统计事实开始。m 个等相关读数——两两相关为 ρ、各自方差归一——取均值,其方差是(1 加 m 减 1 乘 ρ)再除以 m。等相关结构下这是教科书结果,本文用 200 万次蒙特卡洛逐格核验,十二个格子的偏差全部小于千分之二。修正地板 F’ 就是把这一项开方、乘上自由度因子:根号(2 除以 df)乘根号(1 加 m 减 1 乘 ρ)再除以根号 m。它跟原公式 F 的差别不在自由度,全在相关结构的写法上:原式用 ρ 的平方(c 等于 1 加 ρ 平方乘 m 减 1),正确结构用的是 ρ 本身——ρ 小于 1 时后者永远更大,这正是原式在高 ρ 端把地板报低、判据给出虚假承诺的代数根源。

    代回原网格需要 ε 定义层级的完整继承,本文不重造 ε,而是用族内两条已实测的关系把它拼出来:α 在 ρ 等于 0.4 基准格的值(m 不大于 8 时约 2.0,m=16 异常档 2.72 到 2.90),乘以上轮 D 表反推出的 ε 跨 ρ 比值——根号(1 加 m 减 1 乘 ρ)除以根号(1 加 0.4 乘 m 减 1),该形态与上轮 D 表八格全部吻合。代完之后代数上发生了一件干净的事:修正版 α 表达式里所有带 ρ 的项精确对消,剩下 α’ 等于基准格 α 乘一个只含 m 的因子 g——g 等于(根号 c 减 0.4 乘根号 m)除以根号(1 加 0.4 乘 m 减 1)。ρ 轴被修正地板完整吸收,这是解析级归零,不是数值近似。

    数出来的标尺:m=2 时 α’ 约 0.89,m=4 约 0.58,m=8 约 0.34,m=16 约 0.26(若 m=16 取保守档则 0.18)。预先设好的三档判据,裁决是「一败两不触发」。判据一(全格 α’ 落基准加减 0.15)失败:跨 m 全距 0.63,超线一倍,闭式不能一根线走天下。判据二(ρ=0.6 低地板格漂移超 0.5 则闭式退役)不触发:ρ 轴根本没有残余漂移,闭式不用退役。判据三(m 轴与 ρ 轴交叉敏感、等值线扭曲须二维查表)也不成立:两根轴干脆地解耦了,ρ 被 F’ 吃干净,m 轴剩一根独立的一维残余。

    这对报告规范是好消息里带一个尾巴。怕的是最坏情形——ρ 乘 m 的二维分档表,十二格起步还要防交叉项;实际拿到的是一维修正:一根线乘一张只有四个格子的 m 查表(0.89、0.58、0.34、0.26),比二维表便宜一个量级。但两条诚实的边界必须写明。其一,ε* 跨 ρ 比值的形态只在 ρ 等于 0.2 和 0.6 两个端点被 D 表验证过,中间取值若偏离根号形态,α’ 会带小的残余 ρ 漂移,本文无法排除。其二,m 轴残余是从基准格 α 的实测值继承来的,而那份实测自己带着 m=16 网格量化疑点——残余是真实结构还是部分伪影,要把 ε 步长加密一个量级重扫才能定死。测量学文献对这两条边界不陌生:方差比区间的精确性对相关结构敏感是四十多年的老问题,非平衡混合模型下的专门处理 1985 年就有;闭式区间要靠正确的方差结构才能保住精度,近年方法学也有背书;聚类设计里忽略组内相关会把样本量算错的教训同源。脉诊三部九候、会诊多师合议,骨子里都是 m 个相关读数取均值——修正后的地板公式告诉你的第一件事是:相关性每高一分,均值方差就多抬(m 减 1)乘 ρ 那么多,这笔账躲不开。

    证据与局限

    已验证(本文重建模型内部):单位均值方差 b 公式十二格蒙特卡洛核验偏差小于千分之二;修正版 α’ 的 ρ 依赖项精确对消(代数恒等式,三档 ρ 逐格断言通过);上轮 D 表八格用 ε* 比值形态重建全部吻合。

    部分验证:修正地板 F’ 的方差核心(b 形态)是教科书结果加本文 MC 核验;但 ε* 机器本身继承自原网格实测,本文未重造其原始仿真。

    待验证:g(m) 残余能否用单参数闭式拟合(1 除以 m 型还是别的形态,决定 m 查表能不能再压成公式);中间 ρ 取值(0.3、0.5)下 ε* 比值的根号形态保真度;m=16 残余是真实结构还是网格量化伪影(步长加密一个量级重扫)。

    相关阅读

    证据等级: C
    参考文献: PMID 4005371, PMID 36725775, PMID 32823372

    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

    L1 网络状态:2026-11-04 复检 EDGE_TABLE_science-yixue.md 与 CORRESPONDENCE_NETWORK.md——「闭式/标定/地板/保形/方差比」零命中(仅命中 EEG alpha 网络重组等无关条目);最近邻 E-060 标签信度边处于「冲突(未裁决)」状态,但本题「地板公式形态修正 × α 常数性」与该边无对应关系——该主题尚未进入 L1 对应网络,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。

    Revised Floor Closed-Loop Calibration: Zeroing the ρ-axis via analytical series and transferring all residual α drift to the m-axis

    Evidence level: C

    After reformulating the floor equation into a square-root form—specifically √(1 + m − 1 × ρ)—the algebraic drift of α across ρ is precisely canceled. However, the principle that “one line fits all” remains invalid: residual drift is fully transferred to the configuration count m, which scales by more than three times across its range from 2 to 16.

    Body

    Let us first clarify the outstanding accounts. The previous reconstruction yielded two hard facts: merging the floor criterion fails to preserve conformality across ρ (conformal deviation D ranges from 0.40 to 2.25), and D depends solely on the two quantities (ρ, m), independent of degrees of freedom df. At that time, two paths remained: the conservative route involves stratifying by ρ and assigning a multiplier table; the progressive route replaces the relevant structure within the floor formula itself. As previously indicated, the correct dependence of m on ρ is hidden within the square root of (1 + m − 1 × ρ), rather than in the original formula’s √(c − ρ × √m). This paper completes the progressive path: starting from the unit mean variance, we analytically re-derive the corrected floor and substitute it back into the original parameter grid to verify, cell by cell, whether the revised α can be pulled back to a constant.

    The derivation begins with the most elementary statistical fact. For m correlated readings—pairwise correlation ρ, each variance normalized—the mean has variance (1 + m − 1 × ρ)/m. Under this correlated structure, this is a textbook result; we verified it via 2 million Monte Carlo runs across all twelve cells, with deviations in all cases less than two thousandths. The corrected floor F’ is obtained by taking the square root of this term and multiplying by the degrees-of-freedom factor: √(2/df) × √(1 + m − 1 × ρ) / √m. The difference between F’ and the original formula F lies not in degrees of freedom but entirely in the expression of the correlated structure: the original uses ρ² (where c = 1 + ρ² × m − 1), while the correct form employs ρ itself. Since ρ < 1, the latter is always larger, precisely explaining why the original formula underestimates the floor at high ρ, leading to false promises from the criterion.

    Substituting back into the original grid requires full inheritance of the ε definition hierarchy; we do not reconstruct ε anew but instead assemble it from two already-measured relationships within the family: the value of α at the ρ = 0.4 baseline cell (approximately 2.0 when m ≤ 8, and the anomalous range 2.72–2.90 when m = 16), multiplied by the ε ratio across ρ inferred from the previous D table: √(1 + m − 1 × ρ) / √(1 + 0.4 × m − 1). This form matches all eight cells of the previous D table. Algebraically, a clean transformation occurs: in the revised α expression, all terms involving ρ precisely cancel, leaving α’ equal to the baseline α multiplied by a factor g depending only on m: g = (√c − 0.4 × √m) / √(1 + 0.4 × m − 1). The ρ axis is completely absorbed by the revised floor—this is analytical zeroing, not numerical approximation.

    The resulting scale: α’ ≈ 0.89 for m = 2, 0.58 for m = 4, 0.34 for m = 8, and 0.26 for m = 16 (or 0.18 if the conservative tier is taken). Against the three pre-set criteria, the verdict is “one failure, two non-triggers.” Criterion 1 (all cells α’ falling within baseline ±0.15) fails: the full range across m is 0.63, exceeding the line by a factor of two; thus, the closed form cannot be “one line fits all.” Criterion 2 (drift in low-floor cells with ρ = 0.6 exceeding 0.5 triggers retirement of the closed form) does not trigger: there is no residual drift along the ρ axis, so retirement is unnecessary. Criterion 3 (cross-sensitivity between m and ρ axes, requiring two-dimensional lookup due to distorted contour lines) also does not hold: the two axes cleanly decouple, with ρ fully consumed by F’ and m retaining only a single independent one-dimensional residual.

    This is good news for the report, though with a caveat. The concern lies in the worst-case scenario—a two-dimensional stratification table of ρ × m, starting with twelve cells and requiring cross-term prevention. In reality, we obtain a one-dimensional correction: simply multiply the line by a four-cell m lookup table (0.89, 0.58, 0.34, 0.26), which is an order of magnitude cheaper than the two-dimensional table. Nevertheless, two honest boundaries must be stated. First, the form of the ε* ratio across ρ was validated only at the endpoints ρ = 0.2 and ρ = 0.6 via the D table; if intermediate values deviate from the square-root form, α’ may carry small residual ρ drift, which we cannot exclude. Second, the residual along m is inherited from the measured α values at the baseline cell, which themselves carry quantization uncertainty in the m = 16 grid—uncertainty regarding whether the residual reflects true structure or partial artifact, resolvable only by re-running with ε step size refined by one order of magnitude. Measurement literature is familiar with these boundaries: the sensitivity of variance ratio precision to correlated structure has been a longstanding issue for over forty years, with specialized treatment under unbalanced mixed models dating back to 1985; closed-form intervals require correct variance structure to preserve precision, a point supported by recent methodological advances; the lesson from cluster designs—that ignoring within-group correlation leads to incorrect sample size calculation—is of the same origin. Pulse diagnosis’s three sections and nine pulses, or multi-expert consultation, are fundamentally about taking the mean of m correlated readings—the revised floor formula tells us immediately: for every increase in correlation, the mean variance rises by (m − 1) × ρ; this account cannot be avoided.

    Evidence and Limitations

    Verified (within this paper’s reconstructed model): the unit mean variance b formula was validated via Monte Carlo across twelve cells, with deviations less than two thousandths; the ρ-dependence in revised α’ precisely cancels (algebraic identity, confirmed across three tiers of ρ cell-by-cell); the ε* ratio form reconstructed from the previous D table’s eight cells fully matches.

    Partially Verified: the variance core (b form) of the revised floor F’ is a textbook result plus Monte Carlo verification in this paper; however, the machine ε* itself inherits from original grid measurements, and we did not re-simulate its original structure.

    To Be Verified: whether the residual g(m) can be fitted by a single-parameter closed form (1/m type or another form, determining whether the m lookup table can be further compressed into a formula); fidelity of the ε* ratio’s square-root form at intermediate ρ values (0.3, 0.5); whether the m = 16 residual reflects true structure or grid quantization artifact (resolvable only by re-running with ε step size refined by one order of magnitude).

    Further Reading

    Evidence Level: C
    References: PMID 4005371, PMID 36725775, PMID 32823372

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

    L1 Network Status: Rechecked on 2026-11-04 against EDGE_TABLE_science-yixue.md and CORRESPONDENCE_NETWORK.md—zero hits for “closed form / calibration / floor / conformal / variance ratio” (only unrelated entries such as EEG alpha network reorganization matched); the nearest neighbor E-060 label credibility edge is in “conflict (unresolved)” state, but this topic’s “floor formula form correction × α constancy” has no correspondence with that edge—thus this topic has not yet entered the L1 corresponding network, and evidence level is 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.

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

    Evidence level: C

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

    正文

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

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

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

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

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

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

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

    证据与局限

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

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

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

    相关阅读

    证据等级: C
    L1 网络状态:该主题尚未进入 L1 对应网络——EDGE_TABLE_science-yixue.md 检索「截尾、winsorize、重尾、heavy tail、筛查、筛掉、帕累托、pareto、离群、outlier」零命中,CORRESPONDENCE_NETWORK.md 亦无对应边(最近邻 E-061 属 HRV 静息 ICC 命题,仅类比旁证,不覆盖筛查阈值-成本命题)
    参考文献: PMID 32921002, PMID 32267856, PMID 26573817

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


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

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

    Evidence level: C

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

    Main Text

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

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

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

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

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

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

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

    Evidence and Limitations

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

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

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

    Related Reading

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

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


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 折减系数跨判读构型可不可迁移:脉象、得气、HRV 三种读数是不是同一个折减

    Evidence level: C

    折减系数如果随判读构型改变,它就不是统计常数而是「构型函数」——想把一个构型上标定出来的折减数当通用常数用,必须先做跨构型平行标定。

    正文

    上一篇(会话内相关性把有效自由度吃掉多少)解决了一个具体问题:同一批判读读两遍,两遍读数因为共享同一个会话而不独立,会话内相关越高,有效自由度越接近「例数」而不是「例数乘遍数」。解法是实测标定一个折减系数,把名义自由度除以它,样本量表才算数。但那次标定有个隐藏前提:只针对单一判读构型——读者用手指读脉象,同一批例、同一个会话内反复读。

    问题在于,气功外气判读线不止脉象一种读数。至少还有两条常用构型:得气自评量表(读者把主观体验按部位和强度打分)和 HRV 读数(从心电或脉搏波里提取心率变异性的谱分量)。三种构型的物理通道完全不同:脉象是手指触觉的模式识别,得气量表是主观体验的语言化打分,HRV 是仪器信号的算法提取。会话内相关由「同一个人在同一个会话里读数的自我一致性」决定,而这个一致性没有理由跨通道天然一致——触觉系统、体验报告系统、自主神经信号提取系统的稳定性是三套机制。

    旁证确实指向「横跨很宽」。压力痛阈这种仪器辅助读数的会话内重复信度约 ICC 0.73(PMID 33767329);HRV 静息重测信度在不同综述里报告的范围大体落在 0.50 到 0.83 之间。更麻烦的是,个体内方差的异质性本身已被实证:贝叶斯层级模型显示,同一个人在不同场合的读数波动并不是人人相同,存在系统性的个体差异(PMID 34816384)。这意味着折减系数既可能是构型函数,还可能叠加个体函数——如果标定时只测脉象构型,就把这两个自由度全部忽略了。

    信度研究的方法学文献也早就提醒过:信度估计的精度强烈依赖样本量和设计(PMID 24338600),重测信度估计的可靠性随样本量下降明显恶化(PMID 40119099)。把这三件事拼起来,结论是:在一个构型上花大力气标定出的折减数,搬到另一个构型上用,精度和可迁移性都没有担保。

    判据可以这样设计。跨构型平行标定:同一批判读者(不少于六人)对同一批例,三种构型(脉象判读、得气量表、HRV 提取)各做会话内三遍重复;用混合模型分别分解每种构型的会话内方差分量,得到三个会话内相关;再用跨构型一致性系数检验这三个数是不是同一个量级。若三个会话内相关的九成五置信区间彼此重叠、且中位数两两之差小于 0.1,则折减系数可以作为单一常数进入样本量表,各构型共用;只要有一对不重叠或差值不小于 0.1,折减系数就必须按构型分表申报——脉象一张表、得气一张表、HRV 一张表,跨研究比较时先对齐构型再谈自由度。

    研究内核(L1 对应网络)里,这道题还没有自己的边。最接近的一条是「五脏辨证标签的信度前提」边,而且状态标的是冲突,不是确认:一边是开放人群里中医医师间诊断一致性被评为「低」、作者称之为令人担忧(PMID 30383398);另一边是训练可以改善一致性(PMID 19538100)、量表化方案也可行(PMID 15165421、23762155)。这条冲突边讲的是「不同人之间」的一致性,本题问的是「同一个人在会话内」的读数稳定性,属于同一个信度家族的两端,谁也不能替谁作证。这就是本题必须实测的原因。

    对普通读者说清楚利害:折减系数听起来是纯统计的东西,但它标定的对象是「人的读数」。读数通道一换,标定就作废——这和换一把尺子要重新校准是一个道理。没有人会拿体重秤的校准证书去担保血压计,判读构型之间的差别也一样实在。

    证据与局限

    已验证: 会话内重复读数不独立、信度估计精度依赖样本量与设计(PMID 33767329、24338600、40119099)。
    部分验证: 个体内方差异质性存在于其他读数任务(PMID 34816384);中医判读一致性低但可训练改善(PMID 30383398、19538100)——均非气功外气判读本身的直接测量。
    待验证: 三种判读构型的会话内相关是否同量级(跨构型平行标定待做);折减系数是常数还是构型函数;构型间是否会话内相关一致但个体间方差异质,出现「平均数骗人」的分层格局。

    相关阅读

    证据等级: C
    参考文献: PMID 33767329, PMID 34816384, PMID 24338600, PMID 40119099, PMID 30383398, PMID 19538100

    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

    L1 网络状态:该主题尚未进入 L1 对应网络(EDGE_TABLE 与 CORRESPONDENCE_NETWORK 检索「折减 / 有效自由度 / 判读构型 / 会话内相关 / ICC」零命中)。最近邻为「E-060 标签前提」边,原文摘录:「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)……训练可改善(PMID 19538100)、量表化可行」——状态:conflict(未裁决)。该边讲人际一致性,本题讲会话内稳定性,非同一条边。


    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。

    Is the reduction coefficient transferable across diagnostic configurations: Do pulse, deqi, and HRV readings represent the same reduction factor?

    Evidence level: C

    Note: If the reduction coefficient varies with the interpretation configuration, it is not a statistical constant but a “configuration function.” Applying a coefficient calibrated on one configuration as a universal constant requires parallel cross-configuration calibration.

    Body

    The previous discussion (on how within-session correlation consumes effective degrees of freedom) addressed a specific problem: reading the same critical set twice renders the two readings non-independent because they share the same session. Higher within-session correlation means the effective degrees of freedom approach the “sample count” rather than “sample count multiplied by the number of readings.” The solution was to empirically calibrate a reduction coefficient, divide the nominal degrees of freedom by it, and only then does the sample scale count. However, that calibration had a hidden assumption: it applied only to a single interpretation configuration—where the reader uses finger touch to read the pulse, repeatedly on the same batch of cases within one session.

    The problem is that Qigong external Qi interpretation lines include more than just pulse readings. At least two other common configurations exist: the Qi-Self-Assessment Scale (where the reader scores subjective experience by body part and intensity) and HRV readings (extracting heart rate variability spectral components from ECG or pulse wave signals). The physical channels for these three configurations are entirely different: pulse reading relies on tactile pattern recognition via fingers, the Qi scale involves linguistic scoring of subjective experience, and HRV uses algorithmic extraction from instrument signals. Within-session correlation is determined by “the self-consistency of the same person reading within the same session,” and there is no reason for this consistency to be naturally consistent across channels—the stability of the tactile system, the experience reporting system, and the autonomic signal extraction system are three distinct mechanisms.

    Collateral evidence indeed points to “spanning a wide range.” Instrument-assisted readings like pressure pain threshold show within-session repeatability reliability of approximately ICC 0.73 (PMID 33767329); HRV resting retest reliability ranges broadly between 0.50 and 0.83 across different reviews. More problematically, heterogeneity of within-person variance has been empirically demonstrated: Bayesian hierarchical models show that fluctuation in readings across different occasions for the same individual is not uniform, exhibiting systematic individual differences (PMID 34816384). This means the reduction coefficient could be either a configuration function or potentially compounded with an individual function—if calibration measures only the pulse configuration, both degrees of freedom are ignored.

    Methodological literature on reliability studies also long warned: precision of reliability estimation strongly depends on sample size and design (PMID 24338600), and reliability of retest estimation deteriorates significantly as sample size decreases (PMID 40119099). Combining these three points, the conclusion is: a reduction coefficient carefully calibrated on one configuration cannot be transferred to another with may support precision or transferability.

    The criterion can be designed as follows. Cross-configuration parallel calibration: the same critical readers (at least six) perform three within-session repetitions on the same batch of cases using each of the three configurations (pulse interpretation, Qi scale, HRV extraction); use mixed models to decompose the within-session variance component for each configuration, yielding three within-session correlations; then test whether these three numbers are of the same magnitude using cross-configuration consistency coefficients. If the 95% confidence intervals of the three within-session correlations overlap in 95% of cases and pairwise differences of medians are less than 0.1, the reduction coefficient can enter the sample scale as a single constant, shared across configurations; if any pair does not overlap or differs by at least 0.1, the reduction coefficient must be declared separately by configuration—pulse one table, Qi scale another, HRV another—and when comparing across studies, align configurations before discussing degrees of freedom.

    Within the research core (the L1 corresponding network), this question currently has no direct edge. The closest is the “reliability prerequisite for Five-Organ differentiation labels” edge, and the status marked is conflict, not confirmation: on one side, diagnostic consistency among TCM practitioners in open populations is rated as “low,” termed “concerning” by authors (PMID 30383398); on the other, training can improve consistency (PMID 19538100) and scale-based schemes are feasible (PMID 15165421, 23762155). This conflict edge concerns consistency “between different people,” whereas this topic asks about reading stability “of the same person within a session”—belonging to opposite ends of the same reliability family, neither can testify for the other. This is precisely why empirical measurement is required.

    To clarify the stakes for ordinary readers: the reduction coefficient sounds like pure statistics, but what it calibrates is “human readings.” When the reading channel changes, the calibration becomes void—just like changing a ruler requires recalibration. No one would use a weight scale calibration certificate to may support a blood pressure monitor; the difference between interpretation configurations is equally real.

    Evidence and Limitations

    Verified: Within-session repeated readings are non-independent; reliability estimation precision depends on sample size and design (PMID 33767329, 24338600, 40119099).
    Partially Verified: Heterogeneity of within-person variance exists in other reading tasks (PMID 34816384); TCM diagnostic consistency is low but trainable (PMID 30383398, 19538100)—neither directly measures Qigong external Qi interpretation itself.
    To Be Verified: Whether within-session correlations of the three interpretation configurations are of the same magnitude (cross-configuration parallel calibration pending); whether the reduction coefficient is a constant or a configuration function; whether configurations have consistent within-session correlation but heterogeneous between-person variance, creating an “average deceives” stratified pattern.

    Further Reading

    Evidence Level: C
    References: PMID 33767329, PMID 34816384, PMID 24338600, PMID 40119099, PMID 30383398, PMID 19538100

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

    L1 Network Status: This topic has not yet entered the L1 corresponding network (zero hits in EDGE_TABLE and CORRESPONDENCE_NETWORK searches for “reduction / effective degrees of freedom / interpretation configuration / within-session correlation / ICC”). The nearest neighbor is the “E-060 label prerequisite” edge, with original text excerpt: “Diagnostic consistency among TCM practitioners in open populations is ‘low,’ termed ‘concerning’ by authors (PMID 30383398) … training can improve (PMID 19538100), scale-based schemes feasible” — status: conflict (undecided). This edge addresses interpersonal consistency, whereas this topic addresses within-session stability; they are not the same 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.

  • 标定端重尾咬不动 95% 区间,却在 99% 以上开口:膨胀与截尾的分工线

    Evidence level: C

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

    正文

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

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

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

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

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

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

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

    证据与局限

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

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

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

    相关阅读

    证据等级: C
    L1 网络状态:该主题尚未进入 L1 对应网络——EDGE_TABLE_science-yixue.md 检索「重尾、截尾、膨胀、覆盖、稳健、方差比」零命中,CORRESPONDENCE_NETWORK.md 亦无对应边(最近邻 E-061 属 HRV 静息 ICC 命题,仅类比旁证)
    参考文献: PMID 32267856, PMID 26573817, PMID 32921002, PMID 27114218

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


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Calibration of Heavy Tails: Inflation and Truncation Division Lines

    Evidence level: C

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

    Body

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

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

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

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

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

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

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

    Evidence and Limitations

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

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

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

    Related Reading

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

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


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 合并地板判据跨 ρ 保不保形:F≤0.047 一根线要不要按 ρ 分档

    Evidence level: C

    一根在 ρ=0.4 标定出来的地板线,放到 ρ=0.2 会过度保守、放到 ρ=0.6 会给出”比实际更可用”的虚假承诺——F 判据不跨 ρ 保形,报告规范必须分档。

    正文

    先讲背景。平衡注入实验里有一个绕不开的问题:多大的注入量才”检得出来”。上一轮(三轴塌缩那篇)发现了一个漂亮的规律——把判读构型数 m、自由度 df、注入量 ε 三个轴一起扫,”最小可检出注入 ε* 除以闭式地板 F”这个比值 α 稳稳落在 2.0 附近,于是有人想把这根线一根走天下:只要算出 F,乘 2 就是安全线。但那次的塌缩律是在判读相关 ρ=0.4 这一个点上标定出来的。ρ 变了,这根线还站得住吗?

    拆开看机制就知道有麻烦。闭式地板 F 里有一项(根号 c 减 ρ 乘根号 m),这一项随 ρ 增大而缩小:ρ 越高,公式宣称的”地板”越低。可是真实世界里,判读构型之间越相关,读数冗余越多、共模噪声越重,每个样本的单位均值方差反而被(1 加 m 减 1 乘 ρ)这一项越推越高——注入信号被稀释,检出越来越难,真实的 ε 只会往上走。一个往下沉,一个往上升,α 等于两者之比,不存在自动抵消的美事:α 必然随 ρ 单调抬升*。这个方向性判断不依赖任何具体模型细节,是结构性的。

    口说无凭,跑网格。沿用上轮全部设置,把 ρ 从 0.4 扩到 0.2 和 0.6,m 取 2、4、8、16,df 取 20 到 200,共 60 格。每格用解析式算最小可检出注入,再抽三格做 20 万次蒙特卡洛核验,偏差不超过千分之一。结果用”保形偏差 D”(该格 α 除以 ρ=0.4 基准格的 α)来读:ρ=0.2 一侧 D 落在 0.40 到 0.64——α 比基准低三成六到六成;ρ=0.6 一侧 D 落在 1.72 到 2.25——α 比基准高七成二到一倍二五。更扎眼的是低地板端:ρ=0.6 且 m 不小于 8 时,地板公式里那一项缩到 0.13 到 0.18,地板趋零的预言兑现,α 被抬到基准的 2.1 到 2.2 倍。还有一条干净的副产品:D 只跟 ρ 和 m 有关,跟 df 无关——df 那根轴在比值里约掉了,ρ 漂移是(ρ,m)的固有属性,加样本量救不了。

    裁决落在预先设好的三档判据里的第二、三档:α 不落 2.0 加减 0.3,F 判据不跨 ρ 保形;低地板格确实发散。方向上还不对称:把 ρ=0.4 标出的线用到 ρ=0.2,你会要求比实际多 1.6 到 2.5 倍的注入量——过度保守,浪费但安全;用到 ρ=0.6,你会以为很小的注入就够——判据给出”比实际更可用”的危险承诺,平衡验收会放水。危险方向恰好落在高相关、多构型的场景,也就是”多个懂行人看同一件事”的现实配置。

    修正有两条路。保守一条:报告规范按 ρ 分三档(0.2 以下、0.2 到 0.4、0.6 及以上),每档各配一张倍率表。进取一条:把地板公式里那项换成含真实相关结构的形态——本文重建显示,正确的 m 与 ρ 依赖藏在(1 加 m 减 1 乘 ρ)的平方根里,而不是原公式的(根号 c 减 ρ 乘根号 m)里;若修正版闭式能在原网格上把 α 拉回常数,一根线就还能要回来,只是线要重画。两条路都要先过原口径复核这一关。

    落到操作上,分档表可以读成三档倍率:判读相关在 0.2 以下的场合,把按 0.4 标定出来的安全线放宽——少要一点注入,验收不至于无谓卡人;0.2 到 0.4 之间维持原线不动;0.6 及以上的场合,把线收紧到 1.7 至 2.25 倍——宁可多要注入,也不给”地板很低、随便注注就检得出”的错觉。分档表每格只查 ρ 和 m 两个量,不用动样本量,这正是”D 与 df 无关”这条副产品的实用价值:表一旦定版,不随例数变,跨研究也好对表。

    这事跟易学有什么关系?脉诊讲三部九候、会诊讲多师合议,骨子里都是”多构型判读”——但多个视角若高度相关,样本量不能按人头乘。五个意见一致的先生看同一处宅基,信息量不等于五份独立鉴定。方差比区间对相关结构的敏感性是统计学里四十多年的老问题(1985 年就有非平衡混合模型下方差比区间的专门处理),相关输入下解析传播失准、须退回蒙特卡洛的边界也在计量学里写明了。易学多法门互参要”参”得真,先得把”参”字后面的相关性算清楚。

    证据与局限

    已验证(本文重建模型内部):α 随 ρ 单调抬升;60 格网格 D 值 0.40 至 2.25;低地板端(ρ=0.6、m≥8)α 抬 2.1 倍以上;解析与 20 万次蒙特卡洛偏差小于千分之一;D 与 df 无关。

    部分验证:方向性结论(一降一升、必然漂移)不依赖重建细节,属结构性质;但 α 的绝对标尺依赖上轮原口径的 ε 定义层级,本文重建未完全对齐,绝对倍率须以原代码复核为准。

    待验证:修正版地板闭式(根号(1 加 m 减 1 乘 ρ)形态)能否在原网格上把 α 拉回常数;判读者间相关 ρ 的实测经验分布(现实中到底落在 0.2 还是 0.4 还是 0.6,决定分档表哪一档最常用)。

    相关阅读

    证据等级: C
    参考文献: PMID 4005371, PMID 36725775, PMID 21238753, PMID 40119099

    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

    L1 网络状态:EDGE_TABLE_science-yixue.md 与 CORRESPONDENCE_NETWORK.md 检索「塌缩/地板/方差比/判读构型/平衡注入」零命中;最近邻 E-060(内感受量表单因子塌缩结构)与本题”判读构型相关 × 方差比检出地板”命题无关——该主题尚未进入 L1 对应网络,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。

    Merging Floor Criterion Across ρ: Should a Single Line Be Categorized by ρ When F ≤ 0.047?

    Evidence level: C

    A floor line calibrated at ρ=0.4 becomes excessively conservative when applied to ρ=0.2 and offers a false promise of “more usable than reality” when applied to ρ=0.6—the F criterion does not preserve shape across ρ; reporting standards must be stratified.

    Body

    First, the background. A central issue in balance injection experiments is determining the minimum injection amount that can be “detected.” The previous study (on triaxial collapse) revealed a striking pattern: sweeping the axes of judgment configuration count $m$, degrees of freedom $df$, and injection amount $\varepsilon$ together, the ratio $\alpha$—defined as the minimum detectable injection $\varepsilon^*$ divided by the closed-form floor $F$—stably centers around 2.0. This led to the idea of a universal single line: simply calculate $F$ and multiply by 2 to obtain the safety threshold. However, that collapse law was calibrated at a single point of judgment correlation $\rho=0.4$. If $\rho$ changes, does this line still hold?

    Examining the mechanism reveals complications. The closed-form floor $F$ includes a term ($\sqrt{c} – \rho\sqrt{m}$) that shrinks as $\rho$ increases: higher $\rho$ lowers the formula’s declared “floor.” Yet in reality, as judgment configurations become more correlated, reading redundancy increases and common-mode noise intensifies, pushing the unit mean variance of each sample upward by the factor $(1 + m – 1 \times \rho)$—diluting the injected signal and making detection harder, so the true $\varepsilon^$ rises. One term sinks, the other rises; $\alpha$, their ratio, cannot automatically cancel out: $\alpha$ necessarily increases monotonically with $\rho$*. This directional judgment relies on no specific model details; it is structural.

    Words alone are insufficient; we run the grid. Using all settings from the previous round, we expand $\rho$ from 0.4 to 0.2 and 0.6, set $m$ at 2, 4, 8, 16, and $df$ from 20 to 200, creating 60 cells. For each cell, we compute the minimum detectable injection via the analytical formula and verify it with 200,000 Monte Carlo simulations across three cells, achieving deviations under one thousandth. Results are read via “shape-preserving deviation $D$” (the ratio of $\alpha$ in each cell to the baseline $\alpha$ at $\rho=0.4$): on the $\rho=0.2$ side, $D$ ranges from 0.40 to 0.64—$\alpha$ is 36% to 64% lower than baseline; on the $\rho=0.6$ side, $D$ ranges from 1.72 to 2.25—$\alpha$ is 72% higher to double the baseline. More strikingly, at the low-floor end: when $\rho=0.6$ and $m \geq 8$, the term in the floor formula shrinks to 0.13–0.18, fulfilling the prediction of the floor approaching zero, and $\alpha$ is lifted to 2.1–2.2 times the baseline. A clean side product emerges: $D$ depends only on $\rho$ and $m$, not on $df$—the $df$ axis cancels out in the ratio, and $\rho$ drift is an intrinsic property of $(\rho, m)$; adding sample size cannot save it.

    The verdict falls into the second and third tiers of the pre-set three-tier criterion: $\alpha$ does not stay within $2.0 \pm 0.3$, so the $F$ criterion fails to preserve shape across $\rho$; the low-floor cells indeed diverge. The direction is asymmetric: applying the line calibrated at $\rho=0.4$ to $\rho=0.2$ requires 1.6 to 2.5 times more injection than actually needed—overly conservative, wasteful but safe; applying it to $\rho=0.6$ assumes very small injections suffice—the criterion gives the dangerous promise of “more usable than reality,” and balance acceptance will let water flow. The dangerous direction coincides with high-correlation, multi-configuration scenarios—i.e., the real-world setup of “multiple experts observing the same matter.”

    Correction has two paths. Conservative: reporting standards stratify by $\rho$ into three tiers (below 0.2, 0.2–0.4, and 0.6 and above), each with its own multiplier table. Progressive: replace the term in the floor formula with one reflecting true correlation structure—this paper’s reconstruction shows the correct dependence of $m$ and $\rho$ lies in the square root of $(1 + m – 1 \times \rho)$, not in the original formula’s $(\sqrt{c} – \rho\sqrt{m})$; if the corrected closed-form can pull $\alpha$ back to a constant on the original grid, the single line can be recovered, though it must be redrawn. Both paths must first pass the original aperture review.

    Operationally, the stratification table reads as three multiplier tiers: for cases where judgment correlation is below 0.2, relax the safety line calibrated at 0.4—require less injection, avoiding unnecessary blocking during acceptance; keep the original line unchanged for 0.2–0.4; tighten the line to 1.7–2.25 times for $\rho \geq 0.6$—better to require more injection than to create the illusion of “a very low floor where any injection works.” Each cell in the stratification table checks only $\rho$ and $m$, not sample size—this is the practical value of the side product “$D$ independent of $df$”: once the table is finalized, it does not vary with case count, enabling cross-study alignment.

    How does this relate to Yijing Studies? Traditional pulse diagnosis speaks of “three sections, nine pulses” and consultation involves “multi-expert consensus”—essentially both are forms of “multi-configuration judgment.” However, if multiple perspectives are highly correlated, sample size cannot simply multiply by headcount. Five experts agreeing on the same foundation site do not yield five times the information. The sensitivity of variance ratio intervals to correlation structure is a 40-year-old problem in statistics (with dedicated handling of variance ratio intervals under unbalanced mixed models as early as 1985); under correlated inputs, analytical propagation fails and must revert to Monte Carlo boundaries, as explicitly stated in metrology. For Yijing’s multi-method cross-reference to be truly “cross-referenced,” the correlation behind the word “cross” must first be calculated precisely.

    Evidence and Limitations

    Verified (within this paper’s reconstructed model): $\alpha$ increases monotonically with $\rho$; the 60-cell grid yields $D$ values from 0.40 to 2.25; at the low-floor end ($\rho=0.6, m \geq 8$), $\alpha$ rises more than 2.1 times; analytical and 200,000 Monte Carlo results differ by less than one thousandth; $D$ is independent of $df$.

    Partially verified: The directional conclusion (one down, one up; inevitable drift) does not depend on reconstruction details and is structural in nature; however, the absolute scale of $\alpha$ relies on the original definition tier of $\varepsilon$ from the previous round, which this paper’s reconstruction does not fully align with—absolute multipliers must be verified against the original code.

    To be verified: Whether the corrected floor closed-form (in the shape of $\sqrt{1 + m – 1 \times \rho}$) can pull $\alpha$ back to a constant on the original grid; and the empirical distribution of actual judgment correlation $\rho$ in practice (whether it falls at 0.2, 0.4, or 0.6, which determines which tier of the stratification table is most commonly used).

    Further Reading

    Evidence Level: C
    References: PMID 4005371, PMID 36725775, PMID 21238753, PMID 40119099

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

    L1 Network Status: EDGE_TABLE_science-yixue.md and CORRESPONDENCE_NETWORK.md yield zero hits for “collapse/floor/variance ratio/judgment configuration/balance injection”; the nearest neighbor E-060 (interoceptive quantity form factor collapse structure) is unrelated to this paper’s proposition of “judgment configuration correlation × variance ratio detection floor”—this topic has not yet entered the L1 corresponding network, so evidence level is 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.

  • 同一个师傅教出来的两个判读者,那道「一模一样也判出0.29伤害」的地板还作数吗

    Evidence level: C

    地板公式假设两位判读者彼此独立抽样;共享培训和共享参考答案会制造正相关,把地板往下压——压多少没人量过,而它决定小样本判读研究是「全灭」还是「同训可救」。

    正文

    先回顾上一篇文章的结论:在零真实差异的极限情形下,两份判读分布的总变差(TVD)期望不为零,其「地板」等于 0.5 乘以各档二项波动之和,三档 71 例的门槛构型给出 0.0442——换算成伤害指标约 0.29。据此我们判定:几十到几百例的判读设计「不能判、只能描述」。

    但这套推导有一块地基没人敲过:它假设两位判读者对同一批材料的误差是独立的。真实的判读场景恰恰相反——两位判读者上过同一套培训课,对同一类难点学会做同样的系统性取舍;他们看过同一份参考答案,对同一批争议案会做出同样方向的偏移。这些共享成分让两人的误差正相关。

    相关性如何改写地板?两位判读者各自估计同一组档比例 p,两人估计之差的方差是 2(1−ρ)·Var,其中 ρ 是两人估计的组内相关。独立时 ρ=0,方差全额入账;共享训练越强,ρ 越高,方差按 2(1−ρ) 缩水。地板正比于方差的平方根:ρ=0.5 时地板降到约七成,ρ=0.8 时只剩约四成五。也就是说,那道 0.29 的伤害线,在高度同训的团队里可能被系统性高估,把本可成立的内部判读研究误判为无效。

    这不是纯理论担忧。方法学文献早有信号:成对一致性度量的重抽样区间性质强烈依赖判读者之间的依赖结构——在低一致区间,jackknife 与 bootstrap 的表现分歧本身就在提示「分母里藏着相关性」(PMID 21611165);层级二值结局的组内相关必须用非参数 bootstrap 才能稳健估计,正是反解 ρ 的现成工具(PMID 16416402);而 κ 对患病率的依赖只有在敏感度与特异度不变时才成立、真实研究中改变患病率必然连带你者特性(PMID 15939215),这提示「依赖结构」不能当常数写进公式。

    怎么裁决?设计一个三组对照的实测:招募至少 6 组已成对判读者,覆盖三类情形——同中心同培训、同中心不同培训、跨中心。同一批不少于 100 例材料,每人独立重复判读 3 遍。用重复判读的篇内闭式解出每对判读者估计之间的 ρ,再实测每对的零差异地板,看「地板实测值」与「ρ 折算后的解析预测」对不对得上。

    判据三分支:第一,若同中心同训组的 ρ 中位数超过 0.5、且实测地板比独立假设解析值低三成以上,则「地板不可救」必须限定为跨中心或独立判读场景,同训团队允许用更小样本。第二,若三组的 ρ 都低于 0.3、实测地板与解析值差在 15% 以内,则独立性假设站得住,原判据原样保留——共享培训的作用没有想象中大。第三,也是最有趣的可能:若 ρ 与地板的关系在高 ρ 区出现反弹(共享的不只是「正确取舍」,还有「共享错误」被同步放大),则地板根本不能写成单一 ρ 的函数,必须改报「共享误差结构」,并且禁止用 ρ 做样本量折扣——因为你折扣掉的可能不是噪声,而是两个判读者共同的系统性盲区。

    还要说清这道题和前一篇的分工。之前那一问查的是「共同偏移」:同一份参考答案会不会在两位判读者身上制造同方向的偏差,从而生成相关性——那是均值层面的问题。本篇问的是「共同训练压缩分子噪声」:同培训让两人对同一难度的案子做同样的系统性取舍,这是方差层面的问题。一个查偏差、一个查方差,两条线合起来才是「共享成分对地板」的完整账本,任何一条单独都不够。

    第三种情形对这套方法论才是真威胁:正相关缩掉的是「随机项」,但共享错误里如果混着系统性偏置,缩掉方差的同时把偏差留下了——地板降了,偏倚没降,「判得比地板准」可能是幻觉。这正是必须用三组对照而不是单一团队自测的原因。对这套判读方法论的下游用户,操作含义也直接:样本量核算表必须区分「跨中心判读」与「同训团队判读」两栏;在拿到三组实测之前,任何想用「我们俩培训一样」来折半样本量的说法,都先按独立假设保守处理。

    证据与局限

    已验证: 两比例之差的方差在正相关下按 2(1−ρ) 缩(标准统计结果);一致性度量的重抽样区间依赖判读者依赖结构(PMID 21611165)。
    部分验证: 层级二值结局的非参数 bootstrap 反解 ρ 工具链成熟(PMID 16416402);共享训练可改善一致性(E-060 相邻边有 PMID 19538100 支撑)。
    待验证: ρ 与地板压缩的定量对应、高 ρ 区是否反弹、三组构型的 ρ 实际分布——全部为本篇提出、未经实测的仿真/实验设计。

    相关阅读

    证据等级: C
    参考文献: PMID 21611165, PMID 16416402, PMID 15939215

    L1 网络状态:该主题尚未进入 L1 对应网络。最近邻边为 E-060 标签信度(KB-292 conflict:中医医师间诊断一致性在开放人群中「低」、偶然校正一致性统计常被误用建议 Gwet AC2,PMID 30383398/31730403;KB-317 conflict:样本量单位错配 n≥50 指评定者 vs 120 指病例 + 精度区间错配,PMID 40298274/33853549)——均为相邻边,本边(TVD 地板 × 判读者共享训练/共享参考答案的 ρ 方差压缩)无对应条目。故本文证据等级封顶 C,地板压缩机制为推演,非已验证结论。


    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。

    Two graders trained by the same master, does the floor that \”looks identical but yields 0.29 damage\” still count?

    Evidence level: C

    The floor formula assumes the two interpreters draw samples independently. Shared training and shared reference answers create positive correlation, which depresses the floor—by how much has not been measured, yet this determines whether small-sample interpretation studies are “all-or-nothing” or “salvageable through shared training.”

    Body

    First, recall the conclusion of the previous article: in the limit of zero true difference, the expected total variation distance (TVD) between two interpretation distributions is not zero. Its “floor” equals 0.5 times the sum of binomial fluctuations across categories; with a three-tier configuration of 71 cases, this yields 0.0442—translating to a damage metric of approximately 0.29. Thus, we concluded that interpretation designs involving dozens to hundreds of cases “cannot interpret; they can only describe.”

    However, this derivation rests on an unstated foundation: it assumes the errors of two interpreters on the same materials are independent. Real-world interpretation scenarios are precisely the opposite—two interpreters attend the same training course, learn to make identical systematic trade-offs on the same types of difficult cases; they consult the same reference answers, shifting their judgments in the same direction for the same disputed cases. These shared components make their errors positively correlated.

    How does correlation reshape the floor? Each interpreter estimates the proportion of cases falling into each category, $p$. The variance of the difference between their two estimates is $2(1-\rho)\cdot\text{Var}$, where $\rho$ denotes the intraclass correlation of their estimates. When independent, $\rho=0$, so the full variance counts; stronger shared training raises $\rho$, shrinking the variance by the factor $2(1-\rho)$. Since the floor scales with the square root of variance, $\rho=0.5$ reduces the floor to about 70% of its independent value, while $\rho=0.8$ leaves it at roughly 45%. Consequently, the 0.29 damage threshold may be systematically overestimated in highly synchronized teams, causing valid internal interpretation studies to be wrongly deemed ineffective.

    This is not merely theoretical concern. Methodological literature already signals this: resampling intervals for paired-agreement measures strongly depend on the dependency structure between interpreters—discrepancies between jackknife and bootstrap performance in low-agreement regions themselves hint that “correlation lurks in the denominator” (PMID 21611165); nonparametric bootstrap is required to robustly estimate intraclass correlation for hierarchical binary outcomes, serving as a ready tool for inverting $\rho$ (PMID 16416402); and the dependence of Cohen’s $\kappa$ on prevalence holds only when sensitivity and specificity remain constant, whereas real studies altering prevalence inevitably shift these characteristics (PMID 15939215), indicating that “dependency structure” cannot be treated as a constant in formulas.

    How to adjudicate? Design a three-arm empirical trial: recruit at least six pairs of paired interpreters covering three scenarios—same center same training, same center different training, cross-center. Use a batch of no fewer than 100 cases; each interpreter independently reinterprets the same set three times. Solve for $\rho$ between each pair’s estimates using the within-repetition closed-form solution, then measure the zero-difference floor empirically to test whether “empirical floor” matches “theoretical prediction after $\rho$ adjustment.”

    The criterion splits into three branches: First, if the median $\rho$ for the same-center-same-training group exceeds 0.5 and the empirical floor falls more than 30% below the independent-interpretation theoretical value, then “floor not salvageable” must be restricted to cross-center or independent-interpretation scenarios; same-training teams may use smaller sample sizes. Second, if $\rho$ for all three groups stays below 0.3 and the empirical floor differs from the theoretical value by less than 15%, then the independence assumption holds, and the original evidence base remains valid—shared training’s effect is not as large as imagined. Third, and most interestingly, if the relationship between $\rho$ and floor compression rebounds in high-$\rho$ regions (because shared components include not just “correct trade-offs” but also “shared errors” amplified synchronously), then the floor cannot be expressed as a single function of $\rho$; we must report “shared error structure” instead and prohibit using $\rho$ to discount sample size—because what we discount may not be noise, but the systematic blind spots shared by both interpreters.

    We must also clarify how this question differs from the previous one. The earlier inquiry examined “shared bias”: whether a single reference answer creates same-direction deviations in both interpreters, generating correlation—that is a mean-level issue. This current question addresses “shared training compressing numerator noise”: same training leads both interpreters to make identical systematic trade-offs on cases of the same difficulty, which is a variance-level issue. One probes bias, the other variance; together they constitute the complete ledger of “shared components’ impact on the floor.” Any single line alone is insufficient.

    The third scenario poses the true threat to this methodology: positive correlation removes “random terms,” but if shared errors contain systematic bias, reducing variance while leaving bias intact means the floor drops yet bias does not—”being more accurate than the floor” may be an illusion. This is precisely why we need three arms of comparison rather than a single self-test by the same team. For downstream users of this interpretation methodology, operational implications are direct: sample-size calculation tables must distinguish “cross-center interpretation” from “same-training-team interpretation” columns; until the three empirical results are obtained, any claim to halve sample size based on “we share the same training” must be conservatively treated under the independence assumption.

    Evidence and Limitations

    Verified: The variance of the difference between two proportions shrinks by $2(1-\rho)$ under positive correlation (standard statistical result); resampling intervals for agreement measures depend on interpreter dependency structure (PMID 21611165).
    Partially verified: Nonparametric bootstrap inversion of $\rho$ for hierarchical binary outcomes is a mature toolchain (PMID 16416402); shared training can improve agreement (E-060 adjacent edge supported by PMID 19538100).
    To be verified: quantitative correspondence between $\rho$ and floor compression, whether high-$\rho$ regions rebound, and the actual distribution of $\rho$ across the three configurations—all proposed here but not yet empirically tested.

    Further Reading

    Evidence Level: C
    References: PMID 21611165, PMID 16416402, PMID 15939215

    L1 Network Status: This topic has not yet entered the L1 corresponding network. The nearest adjacent edge is the E-060 label credibility (KB-292 conflict: diagnostic agreement among TCM physicians in open populations is “low,” while chance-corrected agreement statistics are often misapplied, per Gwet AC2, PMID 30383398/31730403; KB-317 conflict: sample-size unit mismatch $n\geq50$ refers to raters vs. 120 cases, with precision interval mismatch, PMID 40298274/33853549)—both are adjacent edges. This edge (TVD floor $\times$ $\rho$ variance compression from shared training/shared reference answers) lacks a corresponding entry. Hence, this article’s evidence level caps at C; the floor-compression mechanism is derived, not empirically verified.


    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.