Category: 科学易学

  • 「以资深者为准」会不会错,问的不是频率:校正后的残差恒等于资深者自己那口偏差,翻倍只在偏倚比到 2 时出现

    Evidence level: C

    一个恒等式就把这个设问作废了:把新手挪到资深者身上之后,剩下的偏差就是资深者自己那口偏差——它跟「这次方向判断对没对」无关,也跟「隔多久错一次」无关。

    一、先看这条约定到底做了什么

    两位判读者看同一批病例,各带一口自己的零点偏移:资深者 A 读出「真值 + a」,新手 B 读出「真值 + b」。两人对同一病例的差就是 b − a。

    「以资深者为准」的做法是:把 B 整体减去这口差。代进去一步就出来:

    B − (b − a) = 真值 + a

    也就是说,校正之后剩下的偏差,精确等于资深者自己的偏倚 a。一分不多,一分不少。这与本次方向判断对错毫无关系——方向判断对,残差是 a;判断反了,残差变成 2b − a。两者都不是「错了一次」这么轻的说法,前者是「白校正」,后者是「把偏倚从 b 放大到 2b − a」。

    二、「方向错误率」这个量在数学上没有位置

    设 b = 5 pp。a 从 0 到 4 pp 逐个代进去,衡量「校正后误差平方」与「不校正误差平方」的比值:0.000、0.040、0.250、0.640。

    到 a = 5 pp(也就是 a 与 b 一样大)时,这个比值恰为 1.000。这意味着校正跑了整整一套流程,结果和不校正一模一样,连百分之一的改善都没有。再往上,a = 7.5 pp 时比值 2.25,a = 10 pp 时 4.00。

    这条曲线的形状说明:真正的开关不是一个概率,而是 a 与 b 的比值。它小于 1 就赚,等于 1 就白干,大于 1 就赔。而「错一次翻倍」这件事,只在 a = 2b 这个特殊点上才成立——那时比值恰好是 4。

    三、抽样噪声造成的方向错误率,小到可以不谈

    有人会说:上面假定 b − a 估得准,可它是估出来的,估歪了方向不就反了?这个担心在数值上站不住。

    两人的差值均值,其标准误是根号二乘单次噪声除根号 n。单次噪声取 0.5 pp、n = 120 时,标准误只有 0.0645 pp。真差 5 pp 对应的 z 值是 77 个标准差——方向判错的概率小于十的负十三次方。这不是「很少错」,是「错不了」。

    要让它开始错,真差得小到什么程度?反解出来:错误率降到 15% 需要真差小于 0.067 pp;降到 30% 需要小于 0.034 pp。而这个量级的差,本来就落在「校不校正都无所谓」的区间里。实测蒙特卡洛:真差 0.1 pp 时错误率 6.12%(解析值 6.07%),真差 0.05 pp 时 21.9%。

    换句话说,backlog 里设想的那三档——5%、20% 至 30%、接近 50%——没有一档能由抽样噪声产生。

    四、真正会出错的是换场景,错误率等于 2p(1−p)

    第二个来源才有意义:在 A 场景标定,拿到 B 场景去用。各场景的真实偏倚差 δ 服从均值 μ、标准差 τ 的分布。两个场景方向相反的概率是

    2Φ(μ/τ) × [1 − Φ(μ/τ)]

    代入:μ = 5 pp、τ = 5 pp,错误率 26.7%;τ = 3.6 pp,15.1%;τ = 3.1 pp,10.1%;τ = 2 pp,1.2%。要压到 15% 以下,τ 必须不超过 3.59 pp;压到 10%,不超过 3.09 pp。

    τ 本身估得出来吗?可以。15 份研究的条件下,真 τ = 4 pp 时它的 95% 置信区间半宽约 1.48 pp。所以「错误率 15%」这条线是可检验的,但它检验的对象是跨场景异质性,不是「这条约定本身可不可信」。

    五、两种失败必须分开,混在一起判据就乱了

    第一种是系统性错:资深者自己就比新手偏得更多(a 大于 b)。这时每一次的残差都大于 b,误差平方比是比值的平方。比值 1.5 得 2.25,比值 2 得 4.00。这不是「多少比例会错」,是百分之百都错,而且平均不掉——固定同一位资深者,算 1 份、5 份、15 份、50 份研究,残差恒为 1.000、1.000、1.000、1.000 pp。

    第二种是随机错:每份研究换一对读者,a 围绕 2.5 pp 波动、标准差 2.5 pp。这时平均才有效:1 份研究残差 3.54 pp,5 份 2.74 pp,15 份 2.58 pp,50 份 2.53 pp。压掉的只是随机那部分,那 2.5 pp 的系统项一直留在那里。

    把这两种当成同一个「随机错误率」,就会得出「错 20% 还能接受」的结论——而那正是系统性错时最危险的说法。

    六、外部锚的自指困境

    要判比值 a/b,就得知道 a。而知道 a 需要一个比两人都准的外部尺——那恰恰是校正想省掉的东西。

    量化一下:外部量的偏倚取 1 pp(b = 5 pp),比值估计就偏 0.20;取 3 pp,偏 0.60;取 5 pp,偏 1.00。也就是说,只要外部量的偏倚达到新手的一半以上,「比值小于 0.9 可以校正」这个判断本身就判不出来了。

    这不是悲观,是边界。这条约定不能被验证,只能被限定。

    七、可以执行的报法

    三项必须同时写出来:① b 的点估计与置信区间;② 比值的上界,由外部量精度反推;③ 是固定一位资深者,还是每份研究换人。

    判据:上界小于 0.9,可用;跨在 0.9 至 1.1 之间,判定为打平,直接报未校正值,别把「校正后」当成果;大于 1.1,禁用,因为它不是偶尔错,是每次都错。

    文献里正反两面都有支撑。正面:用多面 Rasch 模型估出并移除评分者效应后,分数与外部变量的相关性确实提高(PMID 36591072);同一量表跨自评与医生评的可比性检验也走这条路,且发现自评系统性高于医生、但幅度小(PMID 25546697)。反面:简单比值型校正会给终身关联高估 29%,原因是当前水平并非唯一驱动(PMID 16051613)——这正是比值大于 1 的真实例子。而方法比较本身,原作者早就说过应当报差值而不是报相关(PMID 22129533)。

    证据与局限

    已验证:校正后残差恒等于资深者自身偏倚,是代数恒等式,不需要数据支撑;误差平方比等于两人偏倚比的平方(多组参数核到小数点后四位)。

    部分验证:抽样噪声导致的方向错误率极小(n ≥ 60 时低于十的负十三次方),已用解析式与四十万次蒙特卡洛互校一致。

    待验证:跨场景异质性 τ 在真实判读领域的取值完全没测过,本文所有 26.7%、15.1%、10.1% 都建在假设的 μ 与 τ 上;比值的上界如何由外部量精度反推,也还没做过实测标定。

    相关阅读

    证据等级: C
    参考文献: PMID 22129533, PMID 25546697, PMID 42306452, PMID 36591072, PMID 16051613, PMID 37336529, PMID 33853549, PMID 30383398

    L1 网络状态:EDGE_TABLE 第 202 行「E-060 标签前提」——「KB-292(conflict):中医医师间诊断一致性在开放人群中『低(low)』,作者称为『令人担忧』(PMID 30383398);同系列第二篇指出偶然校正一致性统计常被误用、建议 Gwet AC2;但训练可改善(PMID 19538100)」——状态标为「冲突(未裁决)」。本文按冲突双方并列处理:一致性低是实测(30383398),训练可改善也是实测(19538100),两者并不互斥,因为前者测的是开放人群的原始一致性、后者测的是训练后的变化量。本文的恒等式只在「两人各自带一口恒定偏移」这个模型下成立,若偏移本身随病例亚型变化(PMID 42306452 的差异性评分者功能正是此情形),残差就不再是一个常数。

    证据等级说明:A 表示有 PubMed 直接支撑该结论,B 表示科学类比或间接证据,C 表示传统文化共识或初步探索。本文标 C,是因为该边在 L1 内核中处于冲突未裁决状态,且本文核心结论是代数推导而非实测。


    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。

    Based on the Chinese text provided, here is the translation of the title into English:

    Evidence level: C

    A single identity equation invalidates this question: shifting the novice onto the senior expert leaves behind only the senior expert’s own bias—it has nothing to do with “whether this directional judgment was right or wrong,” nor with “how often they err.”

    I. What This Convention Actually Does

    Two raters assess the same set of cases, each carrying their own zero-point offset: Senior Expert A reads “True Value + a”, Novice B reads “True Value + b”. The difference between them on the same case is b − a.

    The “rely on the senior expert” approach: subtract this entire difference from B. Substituting once yields:

    B − (b − a) = True Value + a

    Thus, after correction, the remaining bias equals exactly the senior expert’s own bias a. Not one bit more, not one bit less. This is entirely unrelated to whether the directional judgment was correct—if the direction is right, the residual is a; if the direction is wrong, the residual becomes 2b − a. Neither represents “an error once” so lightly; the former is “white correction,” the latter “amplifying the bias from b to 2b − a.”

    II. “Direction Error Rate” Has No Mathematical Standing

    Set b = 5 pp. Substituting a from 0 to 4 pp stepwise, measuring the ratio of “corrected error squared” to “uncorrected error squared”: 0.000, 0.040, 0.250, 0.640.

    At a = 5 pp (i.e., a equals b), this ratio exactly equals 1.000. This means the correction runs through the full procedure yet yields results identical to no correction, with not even a one-hundredth improvement. Beyond that, at a = 7.5 pp the ratio is 2.25, and at a = 10 pp it is 4.00.

    The shape of this curve indicates: the true switch is not a probability, but the ratio of a to b. Below 1 it gains, equal to 1 it is wasted effort, above 1 it loses. And “doubling the error once” holds only at the special point a = 2b—where the ratio is exactly 4.

    III. Sampling Noise Makes Direction Error Rate Negligible

    One might argue: above we assumed b − a is estimated accurately, but since it’s estimated, if the estimate is off, the direction flips? This concern fails numerically.

    The standard error of the mean difference between two raters is √2 times single-rater noise divided by √n. With single-rater noise at 0.5 pp and n = 120, the standard error is only 0.0645 pp. A true difference of 5 pp corresponds to a z-score of 77 standard deviations—probability of directional error less than 10⁻¹³. This is not “rarely wrong,” it is “cannot be wrong.”

    To make error begin, how small must the true difference be? Solving backwards: error rate drops to 15% only if true difference < 0.067 pp; to 30% only if < 0.034 pp. Such magnitudes fall squarely into the “correct or not, doesn’t matter” zone. Monte Carlo simulation confirms: true difference 0.1 pp yields error rate 6.12% (analytical value 6.07%); true difference 0.05 pp yields 21.9%.

    In short, none of the three backlog tiers—5%, 20–30%, near 50%—can arise from sampling noise.

    IV. Real Errors Come from Cross-Scenario Shifts; Error Rate Equals 2p(1−p)

    The second source matters: calibrate in Scenario A, apply to Scenario B. True bias differences δ across scenarios follow a distribution with mean μ and standard deviation τ. Probability the two scenarios have opposite direction is

    2Φ(μ/τ) × [1 − Φ(μ/τ)]

    Substituting: μ = 5 pp, τ = 5 pp → error rate 26.7%; τ = 3.6 pp → 15.1%; τ = 3.1 pp → 10.1%; τ = 2 pp → 1.2%. To keep error below 15%, τ must not exceed 3.59 pp; to keep below 10%, τ ≤ 3.09 pp.

    Can τ be estimated? Yes. Under 15 studies, with true τ = 4 pp, its 95% confidence interval half-width is about 1.48 pp. Thus “error rate 15%” is testable, but it tests cross-scenario heterogeneity, not “whether this convention itself is trustworthy.”

    V. Two Failures Must Be Separated; Mixing Them Confuses Criteria

    First, systematic error: the senior expert is inherently more biased than the novice (a > b). Then every residual exceeds b, and the error-squared ratio is the square of the ratio. Ratio 1.5 gives 2.25, ratio 2 gives 4.00. This is not “what proportion will err,” it is 100% always wrong, and never averages down—fixing the same senior expert, computing 1, 5, 15, or 50 studies, the residual stays constant at 1.000, 1.000, 1.000, 1.000 pp.

    Second, random error: each study pairs different raters, with a fluctuating around 2.5 pp and standard deviation 2.5 pp. Then averaging matters: 1 study residual 3.54 pp, 5 studies 2.74 pp, 15 studies 2.58 pp, 50 studies 2.53 pp. Only the random part gets suppressed; the 2.5 pp systematic term remains.

    Treating both as one “random error rate” leads to the conclusion “20% error is acceptable”—which is precisely the most dangerous claim under systematic error.

    VI. The Self-Referential Dilemma of External Anchors

    To judge the ratio a/b, we must know a. Knowing a requires an external scale more accurate than either—exactly what correction aims to avoid.

    Quantify: external bias taken as 1 pp (b = 5 pp), ratio estimate biased by 0.20; taken as 3 pp, biased by 0.60; taken as 5 pp, biased by 1.00. Thus, once external bias exceeds half the novice’s, “ratio < 0.9 allows correction” cannot even be judged.

    This is not pessimism, it’s a boundary. This convention cannot be validated, only bounded.

    VII. Executable Reporting Format

    Three items must be reported simultaneously: ① point estimate and confidence interval of b; ② upper bound of the ratio, inferred from external scale precision; ③ whether one fixed senior expert or a new person per study.

    Criteria: upper bound < 0.9 → usable; spanning 0.9 to 1.1 → call it a tie, report the uncorrected value directly, do not treat “after correction” as achievement; > 1.1 → forbidden, because it never errs randomly—it always errs.

    Literature supports both sides. Positive: after estimating and removing rater effects using multi-faceted Rasch models, correlation with external variables indeed improves (PMID 36591072); cross-comparability of self-ratings vs physician ratings also follows this path, finding self-rating systematically higher but smaller magnitude (PMID 25546697). Negative: simple ratio-based correction overestimates lifetime association by 29%, because current level is not the sole driver (PMID 16051613)—this is exactly the ratio > 1 case. Moreover, method comparison itself, authors long ago stated should report difference rather than correlation (PMID 22129533).

    Evidence and Limitations

    Verified: After correction, residual equals the senior expert’s own bias—a purely algebraic identity, requiring no data; error-squared ratio equals the square of the bias ratio (multi-parameter kernels to four decimal places).

    Partially Verified: Sampling noise-induced direction error rate is negligible (n ≥ 60, below 10⁻¹³), confirmed by analytical formula and 400,000 Monte Carlo runs.

    To Be Verified: Cross-scenario heterogeneity τ has never been measured in real diagnostic domains; all our 26.7%, 15.1%, 10.1% rest on assumed μ and τ; how the ratio’s upper bound is inferred from external scale precision also lacks empirical calibration.

    Related Reading

    Evidence Level: C
    References: PMID 22129533, PMID 25546697, PMID 42306452, PMID 36591072, PMID 16051613, PMID 37336529, PMID 33853549, PMID 30383398

    L1 Network Status: EDGE_TABLE row 202 “E-060 label premise”—”KB-292 (conflict): diagnostic consistency among TCM physicians in open populations is ‘low (low)’, authors call it ‘concerning’ (PMID 30383398); the second paper in the series notes that chance-corrected agreement statistics are often misused, recommends Gwet AC2; yet training can improve (PMID 19538100)”—status marked “Conflict (undecided)”. This paper treats both sides of the conflict equally: low consistency is empirically measured (30383398), training can improve it is also empirically measured (19538100); they do not contradict because the former measures raw consistency in open populations, the latter measures change after training. Our identity equation holds only under the model “two raters each carry a constant offset”; if the offset itself varies by case subtype (PMID 42306452’s differential rater function is exactly this scenario), the residual is no longer constant.

    Evidence Level Explanation: A indicates direct PubMed support for this conclusion; B indicates scientific analogy or indirect evidence; C indicates traditional consensus or preliminary exploration. This paper is marked C because this edge lies in a conflicted, undecided state within the L1 core, and our core conclusion is algebraic derivation rather than empirical measurement.


    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.

  • 同一批数据、五个起点、五个答案、似然一模一样:位置漂移的可估性,比它的精度先垮

    Evidence level: C

    上一篇算出了「位置漂移到多少会撞线」,这一篇要问:那个漂移量本身能不能从数据里解出来。答案是——放开一个从没人明说过的假设,它根本没有唯一解。

    正文

    一、这条论证卡在一个从没被估过的量上

    前几篇的链条是:判读者的尺子会在两个方向歪——长度(尺度 a)和零点(位置 b)。尺度错有天花板,位置错没有;十二分之一标准差的零点偏移,就足以吃掉 5 个百分点那条判据线。所以「训练到底改没改到该改的地方」全押在 b 上。

    问题是 b 从来没被估过。不是没人愿意,是这条论证停在一个假设上——「只要同一批受试者被 ≥2 位读者打分,b 就能估出来」。这话听着像 many-facet Rasch 里 severity 参数的现成搬用(Myford & Wolfe 那两篇经典综述确实把 severity 立为标准 facet,PMID 14523257 / 15064538)。但「有参数」和「参数可估」是两件事。本文用最朴素的潜变量阈值模型把它算一遍:真值 x~N(0,1),读者 j 把 a·x+b+ε 按自己的一组切点切成 K 级。

    二、硬约束:切点一放开,b 没有唯一解

    同一份合成数据(n=480,参照读者 b=0,第二位读者真 b=0.30,5 级),只改优化起点,别的都不动。

    起点 b=0.00 收敛到 0.000,0.20 到 0.170,0.40 到 0.358,0.70 到 0.728,1.00 到 1.007。五个答案,负对数似然全部是 1460.4372,一位小数都不差。

    不是优化器不收敛,是目标函数在这个方向上本来就平。把 b 固定在 0.00 至 0.90 六个数上、每次让切点组重估到最优,profile 似然一字不差。原因很直白:b 整体加 Δ,四个切点也整体加 Δ,每个受试者落进哪一级完全不变。b 与切点组是一条共线的平移方向,这条方向上没有信息。

    这不是精度问题,是「有没有答案」的问题。它直接作废一种很自然的实验设计:训练前后各判一轮、每轮切点各自自由,然后比两轮 b。这么跑出来的数,取决于优化器从哪出发。

    三、要救,得加一个与目的相冲突的假设

    唯一救法是假定两轮共享同一组切点——「训练没改分档边界,只改了整体松紧」。加上这条,profile 立刻有唯一极小:真值 0.30 处最低,0.15 处差 2.86,0.45 处差 1.65,两头都掉出 95% 区间。

    但请看清它在说什么:它假定训练没动切点。而「训练把判读标准教对了」恰恰是判读训练最主要的机制之一,也是本系列要检测的对象。用一个「训练没改边界」的假设,去检验「训练改了什么」,逻辑上是循环的。做了之后结论只能写成条件句,而这个前提得先独立验证。

    另一条路本文没走但值得记:把时间维显式建模。Uto 的贝叶斯 MFRM 用马尔可夫先验绑住相邻时点的 severity,实测比各时点独立估计更准(PMID 36284065)。它不消除共线,但用「相邻时点不该跳变」的先验把方差压住。前提是得有时序,不是两轮。

    四、就算可识别,精度也差一倍

    共享切点这个可识别版本上,把样本量拉一遍(每档 120~150 次重复):n=120 时 b̂ 标准差 0.140、95% 置信区间半宽 0.275;n=240 是 0.096 / 0.188;n=480 是 0.070 / 0.145;n=960 是 0.051 / 0.091。收敛速度约 1.55/√n。

    对照上一轮立的判据:「n=120 时 b̂ 的 95% 半宽超过 0.13 就说明不可测」。实测 0.275,是 0.13 的 2.1 倍——判据触发,而且不是刚超线。换句话说,那条判据自己把 n=120 判了死刑,它同时又把这档写成实验起点,两句话不能同时成立。

    要把半宽压到 0.126,需 n≈580;只要求标准差达标,需 n≈151。「120 例起步」在任何一档上都不够。

    更麻烦的一层:这些数还是把切点当已知算的。真实分析里切点也得估,b 的方差再涨一截——n=480 时从 0.060 到 0.070(1.17 倍),n=2000 时从 0.030 到 0.039(1.27 倍)。折成等效样本量,n=480 只顶 655 例,掉近两成。

    五、那个「自由度来自读者数」的推论,实测是反的

    上一轮留了一句处方:「要精度就加读者,不加病例,因为 b 的自由度来自读者数。」这句话在两个意义上都错。

    第一层,读者确实提供配对自由度——没有第二位读者,b 与切点连共线都谈不上。所以「必须 ≥2 位读者」是对的。但配对只解决有无解,不解决解得准不准。

    第二层才是要害。固定总判读次数 480,让 R 位读者各评同一批病例:R=2、每读者 240 例 → 0.055;R=4、120 例 → 0.085;R=8、60 例 → 0.134;R=16、30 例 → 0.210。读者从 2 位加到 16 位,精度掉了 3.8 倍。每增一位读者都在摊薄每人的病例数,而精度只看「这一位读者评了多少例」。

    「每人重复看 2 次」也救不了:总判读 480 次下,240 例×2 遍得 0.086,120 例×4 遍得 0.120,60 例×8 遍得 0.143。优先级是病例数 > 重复次数 > 读者数,与上一轮正好倒过来。

    六、b 到 δ 的斜率不是常数,是切点处的密度

    上一轮换算用中位切点那档,得「b=0.3 → 10.6pp、0.6 → 20.4pp、1.5 → 41.0pp」,据此说「十二分之一标准差就吃掉整条线」。这个系数不是常数,它等于切点处的标准正态密度 φ(c),而 c 由阳性率决定。

    按「δ=5pp 需要多大 b」反解:阳性率 50% 要 0.126,40% 要 0.128,30% 要 0.139,20% 要 0.167,10% 要 0.245,5% 要 0.363,2% 要 0.578。低阳性率任务里,「撞线」需要的位置偏移是中位场景的 2.9 到 4.6 倍。

    这对「气功外气变化」这类判读特别要紧:它多半不是对半分的人工二分类,阳性事件本来就稀少。在高门槛场景,那句「小到没人当回事」的 0.126σ 反而真不算大事——要 0.36 到 0.58 个标准差才够。同一个数,中位切点下是警报,2% 阳性率下是常态噪声。

    顺带一个好消息:上一轮担心「a 与 b 高度相关(abs(r)>0.7)就不可分」,实测远没到。n=120 时 r=0.035,n=480 时 0.202,n=1920 时 0.161。两个自由度数值上是分开的,那条担心不成立。

    证据与局限

    已验证:b 与切点组的平移共线是模型的结构性质,profile 似然在 b 方向严格平坦(六点扫描 ΔNLL 全为 0.00000),多起点收敛到 0.000/0.170/0.358/0.728/1.007 而似然同为 1460.4372;共享切点后唯一可识别(0.15 与 0.45 处 ΔNLL 各 2.86 / 1.65,均超 95% 界);b̂ 精度按约 1.55/√n 收缩,n=120 半宽 0.275;固定总判读次数下读者数从 2 增到 16,b̂ 散布从 0.055 涨到 0.210;切点自由估计使 b 方差膨胀 1.17~1.27 倍;b→δ 换算斜率等于 φ(c),低阳性率下所需 b 放大 2.9~4.6 倍。

    部分验证:全部数字来自潜变量阈值模型的合成数据(5 级、等距真值、正态噪声),未在真实判读数据上复现。模型假设「读者只经 a、b 两个自由度作用于真值」,真实判读者还有中心化倾向、晕轮效应等额外 facet(PMID 14523257 列了五类),会额外增加方差,故本文精度数字应视为上限(乐观侧)。MFRM 的 severity 与本文的 b 概念对应,但前者用 logit 尺度、后者用真值标准差尺度,数值不可直接互换。

    待验证:「切点未被训练改变」这个可识别性前提在真实判读数据上是否成立,无任何实测;若训练确实动了切点,本文全部 b̂ 都是条件估计。真实判读任务的阳性率分布(决定 φ(c)、进而决定「5pp 需多大 b」)在本领域从未报告,本文只能给跨基率换算表。气功外气判读的等级数 K 也未定,K 由 3 增到 6 时 b̂ 散布从 0.108 降到 0.092(约 15% 改善),但 K 越多各等级越薄,可行性未知。

    L1 网络状态:L1 中对应本主题的是边表 E-060「标签前提」,状态为「冲突(未裁决),且新增两处单位错配/精度区间错配」——原文记「中医医师间诊断一致性在开放人群中『低』(PMID 30383398);同系列指出偶然校正一致性统计常被误用、建议 Gwet AC2(PMID 31730403 / 31634000 / 31730404);但训练可改善(PMID 19538100)」。冲突双方即:一致性低 vs 训练可改善。本文结论要求两边都成立才讲得通——「一致性低」对应 b 与切点共线导致分不开,而「训练可改善」若成立则恰好破坏本文的可识别性前提(训练动了切点)。故本文不判定该冲突,只指出冲突里有一层从没被分开:两个观察都可以由「一致率 P_o 变化」单独造出,而不动切点分毫。(就「位置参数的可估性、共线方向、精度归属」这一层而言,该主题尚未进入 L1 对应网络,已在 science-yixue-l1-gaps.md 登记为「测量一致性」簇第 34 条。)

    相关阅读

    证据等级: C
    参考文献: PMID 14523257, PMID 15064538, PMID 36284065, PMID 33169286, PMID 20473138, PMID 8501467, 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


    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。

    Estimability of Position Drift Collapses Before Its Accuracy

    Evidence level: C

    The previous article calculated “how much position drift causes crossing the decision line.” This article asks: can that drift magnitude itself be solved from the data? The answer is—by relaxing an assumption never explicitly stated before, there is no unique solution.

    Body

    I. This Argument Stalls on a Quantity Never Previously Estimated

    The chain in previous articles is: the rater’s scale tilts in two directions—length (scale parameter $a$) and zero-point (position parameter $b$). Scale error has a ceiling; position error does not. A zero-point shift of just one-twelfth of a standard difference consumes 5 percent of that decision line. Thus, whether “training actually changed what needed changing” hinges entirely on $b$.

    The problem is that $b$ has never been estimated. Not because no one was willing, but because this argument stops at an assumption—that “as long as the same subjects are rated by ≥2 raters, $b$ can be estimated.” This sounds like a ready application of the severity parameter in many-facet Rasch models (Myford & Wolfe’s classic reviews indeed establish severity as the standard facet, PMID 14523257 / 15064538). But “having a parameter” and “being estimable” are two different matters. This paper computes it using the simplest latent variable threshold model: true values $x \sim N(0,1)$, and rater $j$ cuts $a \cdot x + b + \varepsilon$ into $K$ levels according to their own set of cutpoints.

    II. Hard Constraint: Once Cutpoints Are Freed, $b$ Has No Unique Solution

    Using the same synthetic data ($n=480$, reference rater $b=0$, second rater true $b=0.30$, 5 levels), we only change the optimization starting point, leaving everything else fixed.

    Starting at $b=0.00$ converges to 0.000, 0.20 to 0.170, 0.40 to 0.358, 0.70 to 0.728, 1.00 to 1.007. Five answers, yet all negative log-likelihoods are exactly 1460.4372, differing not even in the first decimal place.

    It is not that the optimizer fails to converge; the objective function is inherently flat in this direction. Fixing $b$ at six values from 0.00 to 0.90 and re-optimizing the cutpoints each time yields profile likelihoods identical to a single decimal. The reason is straightforward: adding $\Delta$ to $b$ overall also adds $\Delta$ to all four cutpoints, so each subject falls into exactly the same level. $b$ and the cutpoint group lie on the same collinear translation direction; there is no information along this direction.

    This is not a precision issue, but a “whether an answer exists” problem. It directly invalidates a very natural experimental design: run judgments before and after training, with each round’s cutpoints freely optimized, then compare the two $b$ values. The numbers produced depend entirely on where the optimizer starts.

    III. To Save It, We Must Add an Assumption That Conflicts with the Goal

    The only rescue is to assume both rounds share the same cutpoints—”training did not change the grading boundaries, only the overall tightness.” Adding this constraint immediately yields a unique minimum: lowest at true value 0.30, off by 2.86 at 0.15, off by 1.65 at 0.45, with both tails falling outside the 95% interval.

    But please note what this implies: it assumes training did not move the cutpoints. Yet “training teaching the judgment standard correctly” is precisely one of the main mechanisms of judgment training, and also the object this series aims to detect. Using an assumption that “training did not change boundaries” to test “what training changed” is logically circular. Afterward, conclusions can only be written as conditional statements, and this premise must first be independently verified.

    Another path not taken here but worth noting: explicitly modeling the time dimension. Uto’s Bayesian MFRM uses Markov priors to bind severity across adjacent time points, empirically outperforming independent estimates at each point (PMID 36284065). It does not eliminate collinearity, but uses the prior that “adjacent time points should not jump” to suppress variance. The prerequisite is having a time series, not just two rounds.

    IV. Even If Identifiable, Precision Is Twice as Poor

    On this identifiable version with shared cutpoints, we pull the sample size through (120~150 repetitions per level): $n=120$ gives $\hat{b}$ standard deviation 0.140, 95% CI half-width 0.275; $n=240$ is 0.096 / 0.188; $n=480$ is 0.070 / 0.145; $n=960$ is 0.051 / 0.091. Convergence rate is approximately $1.55/\sqrt{n}$.

    Comparing to the decision criterion set earlier: “when $n=120$, if $\hat{b}$’s 95% half-width exceeds 0.13, it is unmeasurable.” The actual 0.275 is 2.1 times 0.13—the criterion triggers, and not barely. In other words, that criterion itself sentenced $n=120$ to death, while simultaneously using it as the experimental starting point; these two statements cannot both hold.

    To press the half-width down to 0.126, we need $n \approx 580$; requiring only standard deviation to meet the mark, we need $n \approx 151$. “Starting with 120 cases is insufficient for any level.”

    A more troubling layer: these numbers treat cutpoints as known. In real analysis, cutpoints must also be estimated, inflating $b$’s variance further—by 1.17 times when $n=480$ (from 0.060 to 0.070), and by 1.27 times when $n=2000$ (from 0.030 to 0.039). Converted to equivalent sample size, $n=480$ only equals 655 cases, dropping nearly 20%.

    V. The Inference That “Degrees of Freedom Come from Number of Raters” Is Actually Reversed

    The previous article left a prescription: “for precision, add more raters, not more cases, because $b$’s degrees of freedom come from the number of raters.” This is wrong in two senses.

    First, raters do provide paired degrees of freedom—without a second rater, $b$ and cutpoints cannot even be collinear. So “must have ≥2 raters” is correct. But pairing only solves whether a solution exists, not how accurate it is.

    The second point is the crux. Fixing total judgment count at 480, let $R$ raters each rate the same batch of cases: $R=2$, 240 cases each → 0.055; $R=4$, 120 cases each → 0.085; $R=8$, 60 cases each → 0.134; $R=16$, 30 cases each → 0.210. Increasing raters from 2 to 16 reduces precision by 3.8 times. Each additional rater dilutes the number of cases per person, while precision depends only on “how many cases this rater judged.”

    “Each person repeats twice” also fails: under total 480 judgments, 240 cases × 2 rounds gives 0.086, 120 cases × 4 rounds gives 0.120, 60 cases × 8 rounds gives 0.143. Priority is cases > repetitions > raters, exactly the reverse of the previous claim.

    VI. The Slope from $b$ to $\delta$ Is Not Constant; It Equals Density at the Cutpoint

    The previous conversion used the median cutpoint level, yielding ” $b=0.3 \to 10.6\text{pp}$, $0.6 \to 20.4\text{pp}$, $1.5 \to 41.0\text{pp}$,” hence claiming “one-twelfth of a standard difference eats the whole line.” This coefficient is not constant; it equals the standard normal density $\phi(c)$ at the cutpoint, where $c$ is determined by the positive rate.

    Solving backwards for “what $b$ is needed for $\delta=5\text{pp}$”: positive rate 50% requires 0.126, 40% requires 0.128, 30% requires 0.139, 20% requires 0.167, 10% requires 0.245, 5% requires 0.363, 2% requires 0.578. In low-positive-rate tasks, the position shift needed to “hit the line” is 2.9 to 4.6 times the median scenario.

    This is especially critical for judgments like “qigong external qi changes”: they are rarely artificial binary splits with 50% positivity, since positive events are inherently scarce. In high-threshold scenarios, that “tiny 0.126σ nobody cares about” is actually not a big deal—it requires 0.36 to 0.58 standard differences to matter. The same number is an alarm under median cutpoints but normal noise under 2% positive rate.

    A side note of good news: the previous worry that “if $|r(a,b)| > 0.7$, they are inseparable” never materialized in practice. At $n=120$, $r=0.035$; at $n=480$, $r=0.202$; at $n=1920$, $r=0.161$. The two degrees of freedom are numerically separable, so that worry does not hold.

    Evidence and Limitations

    Verified: $b$’s collinearity with the cutpoint group is a structural property of the model; profile likelihood is strictly flat in the $b$ direction (six-point scan $\Delta\text{NLL}$ all 0.00000); multiple starting points converge to 0.000/0.170/0.358/0.728/1.007 with identical likelihood 1460.4372; after sharing cutpoints, only identifiable (ΔNLL at 0.15 and 0.45 are 2.86 / 1.65, both exceeding 95% bounds); $\hat{b}$ precision contracts at rate ~1.55/√n, half-width 0.275 at $n=120$; under fixed total judgments, increasing raters from 2 to 16 inflates $\hat{b}$ spread from 0.055 to 0.210; free estimation of cutpoints inflates $b$ variance by 1.17~1.27 times; the $b \to \delta$ conversion slope equals $\phi(c)$, requiring $b$ to be 2.9~4.6 times larger under low positive rates.

    Partially Verified: All numbers come from synthetic data of the latent variable threshold model (5 levels, equally spaced true values, normal noise), not replicated on real judgment data. The model assumes “raters affect true values only via $a$ and $b$ degrees of freedom,” whereas real raters have additional facets like central tendency bias and halo effect (PMID 14523257 lists five types), which would further increase variance; thus these precision figures should be viewed as upper bounds (optimistic side). MFRM’s severity corresponds to this paper’s $b$, but the former uses logit scale while the latter uses true-value standard deviation scale, so values cannot be directly swapped.

    To Be Verified: Whether the identifiability premise “cutpoints remain unchanged by training” holds for real judgment data has no empirical verification; if training indeed moved cutpoints, all $\hat{b}$ here are conditional estimates. The positive rate distribution of real judgment tasks (which determines $\phi(c)$, hence “what $b$ is needed for 5pp”) has never been reported in this field; this paper can only provide cross-rate conversion tables. The number of levels $K$ for qigong external qi judgment is also undetermined; when $K$ increases from 3 to 6, $\hat{b}$ spread drops from 0.108 to 0.092 (~15% improvement), but with more levels each grade becomes thinner, feasibility unknown.

    L1 Network Status: In L1, the entry corresponding to this topic is edge table E-060 “Label Premise”, status “Conflict (Unresolved), with two new unit mismatches/precision interval mismatches—original record states “diagnostic consistency among TCM physicians is ‘low’ in open populations (PMID 30383398); the same series notes that chance-corrected consistency statistics are often misused, recommending Gwet AC2 (PMID 31730403 / 31634000 / 31730404); yet training can improve (PMID 19538100).” The conflicting parties are: low consistency vs. trainable improvement. This paper’s conclusion requires both to hold for coherence—”low consistency” corresponds to $b$ and cutpoints being collinear and inseparable, while “trainable improvement” if true would exactly destroy this paper’s identifiability premise (training moved cutpoints). Hence this paper does not resolve the conflict, only points out one layer within it never separated: both observations can be generated solely by “consistency rate $P_o$ change” without moving cutpoints at all.** (Regarding this layer of “position parameter estimability, collinear direction, precision attribution,” this topic has not yet entered the corresponding L1 network; it is registered in science-yixue-l1-gaps.md as cluster #34 of “measurement consistency”.)

    Related Reading

    Evidence Level: C
    References: PMID 14523257, PMID 15064538, PMID 36284065, PMID 33169286, PMID 20473138, PMID 8501467, PMID 30383398, PMID 19538100

    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.

  • 两位判读者是一起漂还是各漂各的:这个相关系数不是「没人量过」,而是「两个人的数据里量不出来」

    Evidence level: C

    想量「两人漂移的相关系数」,用两个人的分数是量不出来的——三个量里只有两个独立信息。要量它,必须往设计里加一个不是人的东西。

    一、先把这个设问的算式摆正

    上一批文章里有一条结论:实测漂移是临界限的 1.5 到 3.8 倍,撞线概率约七成。这条结论站在一个式子上——两人分差的标准差等于单个漂移标准差乘根号二。根号二从哪来?从「两人各漂各的、互不相干」这个前提来。

    这个前提从没被检验过,一松就全翻。若两人一起往同方向松(同一套培训、同一时段、同一份标准一起走样),分差被削掉大半,撞线概率掉一个量级。

    但有一件事必须先说清:「量这个相关系数」这个动作,用两位判读者自己的数据做不到。

    二、三个量、两个方程:为什么不可辨识

    把每次评分的严厉度拆成三块:读者自己的常数偏倚、随时间一起走的共同成分、读者与时间的交互(各漂各的)。可观测的只有两个统计量——两人分差的方差只由交互那一块决定,共同成分在里面被差掉了;两人均值的方差则由共同成分、交互成分的一半、以及案例难度漂移三者相加。

    两个方程、三个未知数。解不唯一。数值核算:固定交互项 0.30,让共同成分与案例漂移从零往上走,估出的相关系数会从 0.000 涨到 0.500、再到 0.667,而真值一直是 0。只要案例难度也在漂,估出来的「共同模」有一半是案例漂移冒充的。

    这正是多面 Rasch 模型里反复记录的一件事:严厉度与题目难度的漂移在观测上混在一起,看分数分不开(PMID 37283589 指出忽略评分顺序会产生有偏的模型结构与能力估计;PMID 36284065 给严厉度加马尔可夫先验,正因时点独立假设不够用)。

    三、原文那两个数字对不上

    顺手核了一下上一批给出的两个数:分差标准差 0.44、共同模下 0.098。按根号二关系反解,0.098 对应相关系数 0.9504,不是文中暗示的 0.9。而按七成撞线率反解临界限(约 0.170),相关系数 0.9 对应的分差标准差是 0.139,撞线率落在 22.3%,也不是 8.8%。

    方向没错——共同模确实把撞线率大幅压低——但幅度被夸大了。这类「方向对、数字没核」的算式,正是这条支线反复出问题的地方。

    四、外部证据怎么说

    一项持续近两年、458 名教师的观察评分研究,直接测了「评分者会不会收敛到同一水平」:结论是没有收敛,评分者之间的差异在整个研究期内反而在扩大(PMID 29795823)。这支持「各漂各的」这一侧。

    另一项 16 名考官的前后对照研究,把严厉度反馈给考官本人,期望他们向共同标准靠:干预没有显著效应,且严厉度在干预前就已不稳定;作者写「校准对考官而言似乎很难」(PMID 28587597)。这削弱了「共享干预产生共享漂移」这条机制。

    还有一项指出:漂移检测依赖「重评旧材料」的设计,而重评样本的抽样方式会实质影响结论(PMID 29881127)。连带「漂移有多大」这个上游量都还没被稳稳测住。

    五、那要怎么量

    加人。 三位以上判读者、构成连通网络,共同成分才可能从两两分差的方程组里分离出来。

    加锚。 同一批材料在不同时点由固定一组校准读者重评,用他们的漂移当共同成分读数。但锚自己也在漂,这一步只是把不可辨识变成条件可辨识。

    加非人通道。 案例难度漂移若能由材料本身的客观指标(时长、复杂度、客观检查条目数)预估,就能从均值方差里扣掉,剩下的才归给共同模。

    判据建议:若三条路都做不到,「共同模有多大」在任何单一研究中都不该给点估计,只能报区间,且区间须宽到同时包含「完全独立」与「完全共模」两个端点。 这不是保守,是数据里真的没那么多信息。

    证据与局限

    已验证:分差方差只反映交互成分、不反映共同成分,这是代数事实(本文数值核算,固定交互项 0.30 时真值 0 可估出 0.500);持续两年的观察评分研究实测到差异扩大而非收敛(PMID 29795823);严厉度反馈干预无显著效应(PMID 28587597)。

    部分验证:漂移确实存在、量级不小,这在多面 Rasch 与概化理论的多次应用中反复出现(PMID 36284065、PMID 37283589);但这些研究都没把共同成分单独分离出来报过。

    待验证:共同模的真实大小;案例难度漂移与评分者漂移的相对量级;加入第三位读者或客观锚之后区间能收窄多少。三项全部未被实测,本文所有数字均为模型推演,不可当实测值引用。

    相关阅读

    证据等级: C
    参考文献: PMID 29795823, PMID 28587597, PMID 37283589, PMID 36284065, PMID 29881127

    L1 网络状态(原文摘录,E-060 标签前提,CORRESPONDENCE_NETWORK.md 第 202 行):「KB-292(conflict):中医医师间诊断一致性在开放人群中『低(low)』,作者称为『令人担忧』(PMID 30383398)……但训练可改善(PMID 19538100)、量表化可行……→ 『用五脏标签做分组』这一步本身带噪声」,层级「冲突(未裁决)」,状态 conflict。该边只裁决「标签整体信度低不低、训练能不能改善」,从未涉及「漂移是共同还是独立」这一层,故本文证据等级封顶 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。

    Are the two raters drifting together or drifting apart: this correlation coefficient is not \”unmeasured,\” but \”impossible to measure from the two people’s data.\

    Evidence level: C

    To quantify the “correlation coefficient of drift between two individuals,” using the scores of the two people alone is insufficient—only two independent pieces of information exist among three quantities. To measure it, one must introduce a non-human element into the design.

    I. Correcting the Equation for This Question

    A conclusion from a previous article states: measured drift ranges from 1.5 to 3.8 times the threshold, with a collision probability of approximately seventy percent. This conclusion rests on an equation—the standard deviation of the difference between two raters equals the single drift standard deviation multiplied by the square root of two. Where does the square root of two come from? It stems from the premise that “the two raters drift independently, without influencing each other.”

    This premise has never been tested; once relaxed, everything collapses. If the two drift together in the same direction (same training regimen, same time period, same standard deforming simultaneously), the difference is reduced by half, and the collision probability drops by an order of magnitude.

    However, one thing must be clarified: measuring this correlation coefficient cannot be done using data from the two raters themselves.

    II. Three Quantities, Two Equations: Why They Are Unidentifiable

    Let us decompose the severity of each rating into three parts: the rater’s own constant bias, the common component drifting over time, and the interaction between rater and time (drifting independently). Only two statistical quantities are observable—the variance of the difference between raters is determined solely by the interaction component, as the common component cancels out in the difference; the variance of the mean is composed of the common component, half the interaction component, and the case difficulty drift.

    Two equations, three unknowns. The solution is not unique. Numerical calculation: fixing the interaction term at 0.30 and letting the common component and case drift increase from zero, the estimated correlation coefficient rises from 0.000 to 0.500 and then to 0.667, while the true value remains 0. As long as case difficulty is also drifting, half of the estimated “common mode” is actually masquerading as case drift.

    This is precisely what has been repeatedly recorded in the Multi-Facet Rasch model: severity and item difficulty drift are confounded in observation, making scores indistinguishable (PMID 37283589 points out that ignoring rating order produces biased model structure and ability estimates; PMID 36284065 adds a Markov prior to severity precisely because the independence-of-time assumption is insufficient).

    III. The Original Two Numbers Do Not Match

    Checking the two numbers given previously: difference standard deviation 0.44, common mode under 0.098. Solving backwards via the square root of two relationship, 0.098 corresponds to a correlation coefficient of 0.9504, not the 0.9 implied in the text. Solving backwards via the seventy percent collision rate (threshold approximately 0.170), a correlation of 0.9 yields a difference standard deviation of 0.139, placing the collision rate at 22.3%, not 8.8%.

    The direction is correct—the common mode does indeed significantly suppress the collision rate—but the magnitude is exaggerated. Such “correct direction, unverified numbers” equations are precisely where this branch repeatedly fails.

    IV. What Do External Evidence Say?

    A two-year observational study of 458 teachers directly measured “whether raters converge to the same level”: the conclusion is no convergence; differences between raters actually widened over the study period (PMID 29795823). This supports the “drifting independently” side.

    Another before-and-after study of 16 examiners, where severity feedback was given to the examiners themselves, expecting them to move toward the common standard: the intervention had no significant effect, and severity was already unstable before the intervention; authors wrote “calibration seems difficult for examiners” (PMID 28587597). This weakens the mechanism of “shared intervention producing shared drift.”

    There is also a study indicating: drift detection depends on the design of “re-rating old materials,” and the sampling method of the re-rated sample substantially affects conclusions (PMID 29881127). Even “how much drift there is,” this upstream quantity, has not been firmly measured.

    V. How Then to Measure It?

    Add people. With three or more raters forming a connected network, the common component can only be separated from pairwise difference equations.

    Add anchors. The same batch of materials is re-rated at different times by a fixed set of calibrated raters, using their drift as the common component reading. But anchors themselves also drift; this step merely transforms the unidentifiable into conditionally identifiable.

    Add non-human channels. If case difficulty drift can be estimated by objective indicators of the material itself (duration, complexity, number of objective check items), it can be subtracted from the mean variance, leaving only what belongs to the common mode.

    Judgment recommendation: If none of these three paths is feasible, “how large the common mode is” should not be given a point estimate in any single study; only intervals should be reported, and these intervals must be wide enough to encompass both “completely independent” and “completely common mode” endpoints. This is not conservatism; it is simply that the data contains far less information than assumed.

    Evidence and Limitations

    Verified: Variance of differences reflects only the interaction component, not the common component, which is algebraic fact (this article’s numerical calculation shows that with fixed interaction 0.30, true value 0 yields estimated 0.500); the two-year observational study measured widening rather than convergence of differences (PMID 29795823); severity feedback intervention had no significant effect (PMID 28587597).

    Partially verified: Drift does exist and is substantial, repeatedly appearing in applications of Multi-Facet Rasch and Generalizability Theory (PMID 36284065, PMID 37283589); however, none of these studies separately reported the common component.

    Pending verification: True magnitude of the common mode; relative magnitude between case difficulty drift and rater drift; how much the interval narrows after adding a third rater or objective anchors. All three remain unmeasured; every number in this article is model-derived and cannot be cited as observed values.

    Related Reading

    Evidence Level: C
    References: PMID 29795823, PMID 28587597, PMID 37283589, PMID 36284065, PMID 29881127

    L1 Network Status (excerpt from original text, E-060 tag premise, CORRESPONDENCE_NETWORK.md line 202): “KB-292 (conflict): Diagnostic consistency among TCM physicians in open populations is ‘low’ (PMID 30383398), termed ‘concerning’ by authors… but training can improve (PMID 19538100), scaling is feasible… → ‘grouping by five-organ labels’ itself introduces noise’, level ‘conflict (unresolved)’, status conflict. This edge only adjudicates ‘whether label reliability is low and whether training can improve’, never addressing ‘whether drift is common or independent’, hence this article’s evidence level caps 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.

  • 定期重校准到底管不管用:两时点分不出「越判越松」和「每次乱抖」,而两者的下限差一个量级

    Evidence level: C

    重校准能压住多少漂移,不取决于校准多勤,取决于漂移按什么规律长出来 —— 而这件事,看两次数据永远看不出来。

    一、这个处方是怎么来的

    真人判读的位置漂移是五个百分点报警线的一点五到三点八倍。判前校准压不住判读过程中的漂移,那就退一步,隔一段时间重校准一次。这问题被问过但没答清:原因不在校准动作够不够标准,而在它落在什么形态上——同一条措施遇到不同形态,效果不是差一点,是差一个层级。

    二、两时点分不开:不是测不到,是解不唯一

    设严厉度随时间走成三段之和:每轮固定偏移 c、每轮随机抖动 s、单次判读的瞬时噪声 τ。两次判读只给一个增量,方差要同时承担三件事。数矩条件,两时点只有 1 个增量、1 个可算方差,未知数 3 个,冗余负 2。落到数字:十轮后观测到 0.167 个标准差的位移,可行解是一整条曲线 —— 纯游走步长 0.0528 且零趋势,或零游走且每轮 0.0167 的趋势,中间每档都成立。常引用的「步长 0.05 时约十一轮漂回报警线」,十一是纯游走假设下的解,换成纯趋势是每轮 0.015。

    三、三时点就够,但靠的不是看曲线

    加一个时点,增量变 2 个,可算的协方差元素变 3 个,冗余刚好归零。关键是分辨形态靠的不是「走势图像直线还是像毛刺」,而是相邻两段增量的协方差符号。趋势加瞬时噪声时,相邻增量共用同一个噪声项,协方差等于 τ 的平方取负;纯游走时各走各的,协方差为零。模拟每格四千次:15 人时 87.6% 与 88.3%,30 人时 95.3% 与 96.0%,60 人时双双超过 99%。代价是约一半抽样会给出负的方差分量,必须用约束估计,不能把点估计当结论。

    四、实测把几种形态都碰上了,而且互相打架

    几条实证摆一起,谁也没压下谁。五次同版本十二站 OSCE、228 个评分机会里只有 1 个显示显著漂移,且先被评的考生反而占便宜,与「越判越松」正相反(PMID 30955444)。另一次单场 OSCE 里时段与评分显著正相关,系数 0.88(95% 置信区间 0.38 到 1.38,P=0.001),与站点难度有交互,指向疲劳(PMID 19769648)。第三份是 8 名评分者对同一批视频判四轮、间隔至少一周,总分在 3.0 到 35.7 之间摆,第四轮显著抬高(PMID 26837086)。样本最大那份 458 名教师两年,漂移在最初几天极大并持续近两年,评分者没收敛到共同严厉度、彼此变异反而增大(PMID 29795823)—— 这条跳出趋势与游走的老框子:发散型既拦不住也不会自己稳定。

    五、量级差在地板,不在降幅

    把「间隔减半」当可做的实验,两种形态的降幅是 50.0% 与 29.3%,只差 1.7 倍,说差一个数量级是夸大。差一个量级的是可达下限:趋势型随校准加密趋近于零,每例都校准后残留 0.0152,是报警线的 9%;游走型无论校准多密都卡在步长本身,步长 0.05、0.10、0.15 时残留分别是报警线的 30%、60%、90%。趋势型的问题是校准够不够勤,游走型的问题是校准根本改不了它。两者并存时降幅落在 30% 到 32%,靠降幅分不出来,所以协方差符号那条检验不能省。

    六、怎么测、怎么写进规范

    最低设计:固定锚定材料、至少六名评分者、至少六个均匀间隔时点、顺序随机。先算三个量:相邻增量的协方差、间隔减半前后的偏移比、拟合优度(六时点冗余十二)。判据:协方差显著为负且降幅近 50%,写「间隔不超过 T 则偏移不超过 cT」;协方差为零且降幅落在 29% 附近,写「校准只重置零点、漂移按根号 T 重累积」并明写残留地板等于步长;两者并存则两参数并报。

    证据与局限

    已验证:三时点起趋势型与游走型在矩条件上可分,分离靠相邻增量的协方差符号而非曲线形状;模拟在 30 人以上给出九成以上判别率。校准频率对两类形态的可达下限不同,是代数结论,不依赖数据。

    部分验证:实测证据互相冲突。「不到百分之一显著漂移」与「时段系数 0.88 显著」并存,说明形态可能因人因场景而异,不是领域共性;

    待验证:几种形态在真实判读数据里各占多少,没有任何研究报过。六时点以上的纵向序列在中文医学判读研究里几乎不存在,本判据只能作设计规范。

    L1 网络状态

    L1 网络状态:E-060 标签前提 原文记为「五脏辨证(作为分组标签)— 前提 — 标签信度 …… 冲突(未裁决)」,该边只裁决「标签整体信度低不低、训练能不能改善」。对「漂移按趋势累积还是按游走扩散」,CORRESPONDENCE_NETWORK.md 与 EDGE_TABLE_science-yixue.md 全表检索「游走」「趋势」「累积」「疲劳」「时点」「马尔可夫」「状态空间」「纵向」均 0 命中。该主题尚未进入 L1 对应网络,故证据等级封顶 C,缺口已登记至 science-yixue-l1-gaps.md。

    相关阅读

    证据等级: C
    参考文献: PMID 30955444, PMID 19769648, PMID 26837086, PMID 29795823, PMID 36284065, PMID 37283589, PMID 29881127, PMID 28587597


    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。

    Does Regular Recalibration Work: Two Timepoints Cannot Distinguish \”Gradually Looser\” from \”Chaotic Fluctuations,\” Yet Their Lower Bounds Differ by an Order of Magnitude

    Evidence level: C

    Recalibration suppresses drift only to the extent dictated by the underlying growth pattern of the drift, not merely by how frequently it is performed. This distinction cannot be discerned from two data points alone.

    I. Origin of This Prescription

    The observed positional drift among real raters is 1.5 to 3.8 times the five-percent alarm threshold. Since pre-assessment calibration fails to suppress drift occurring during the assessment process, we must step back and consider periodic recalibration. This issue has been asked but not fully answered: the cause lies not in whether the calibration action is standard enough, but in what morphological pattern it falls upon—the same measure encounters different forms, and the effect differs not slightly, but by an entire order of magnitude.

    II. Two Timepoints Are Insufficient: Not Unmeasurable, But Non-Unique Solutions

    Let severity evolve over time as the sum of three components: fixed offset $c$ per round, random jitter $s$ per round, and instantaneous noise $\tau$ per single judgment. Two assessments provide only one increment; the variance must simultaneously account for three unknowns. Under moment conditions, two timepoints yield 1 increment and 1 calculable variance, yet 3 unknowns, resulting in a redundancy of -2. Numerically: after ten rounds, an observed displacement of 0.167 standard deviations yields a feasible solution as an entire curve—pure random walk step size 0.0528 with zero trend, or zero random walk with a trend of 0.0167 per round; both hold true for intermediate intervals. The commonly cited “step size 0.05 returns to the alarm line after about eleven rounds” is solved under the pure random walk assumption; switching to pure trend yields 0.015 per round.

    III. Three Timepoints Suffice, But Not by Observing the Curve

    Adding a third timepoint increases increments to 2 and calculable covariance elements to 3, making redundancy exactly zero. Crucially, distinguishing the pattern does not rely on whether the trend image looks straight or spiky, but on the sign of the covariance between adjacent increments. When trend plus instantaneous noise is present, adjacent increments share the same noise term, so covariance equals $-\tau^2$; under pure random walk, each moves independently, so covariance is zero. Simulations with 4,000 trials per cell: at 15 raters, discrimination rates are 87.6% vs. 88.3%; at 30 raters, 95.3% vs. 96.0%; at 60 raters, both exceed 99%. The cost is that about half of sampled cases yield negative variance components, requiring constrained estimation rather than treating point estimates as conclusions.

    IV. Real-World Evidence Encounters Multiple Patterns That Conflict

    Several empirical studies coexist, none suppressing the others. In five OSCE sessions with twelve stations and 228 scoring opportunities, only 1 showed significant drift—and the first-assessed candidates actually gained advantage, contradicting “loosening over time” (PMID 30955444). In another single OSCE session, time interval and score were significantly positively correlated (coefficient 0.88; 95% CI 0.38 to 1.38, $P=0.001$), interacting with station difficulty, pointing to fatigue (PMID 19769648). A third study had 8 raters assess the same video set over four rounds with at least one week between, total scores ranging from 3.0 to 35.7, with the fourth round significantly elevated (PMID 26837086). The largest sample, 458 teachers over two years, showed drift extremely high initially and persisting for nearly two years; raters failed to converge on common severity, and mutual variation increased instead (PMID 29795823)—this breaks the old framework of trend vs. random walk: divergent patterns neither stabilize nor self-correct.

    V. The Difference Lies in the Floor, Not the Reduction Rate

    Treating “halving the interval” as an experiment: the reduction rates for the two patterns are 50.0% and 29.3%, differing only by 1.7 times—claiming an order-of-magnitude difference is exaggerated. The true order-of-magnitude gap lies in the achievable lower bound: trend-type drift approaches zero with denser calibration, leaving 0.0152 after full recalibration—9% of the alarm line; random-walk-type drift remains stuck at its step size regardless of calibration density—residuals are 30%, 60%, and 90% of the alarm line at step sizes 0.05, 0.10, and 0.15 respectively. For trend-type, the issue is calibration frequency; for random-walk-type, the issue is that calibration cannot alter it at all. When both coexist, reduction falls between 30% and 32%, making it impossible to distinguish by reduction rate alone, so the covariance-sign test cannot be omitted.

    VI. How to Measure and Write into Standards

    Minimum design: fixed anchor material, at least six raters, at least six evenly spaced timepoints, random order. First compute three quantities: covariance of adjacent increments, offset ratio before and after halving the interval, goodness-of-fit (six timepoints yield twelve redundancies). Criteria: if covariance is significantly negative and reduction nears 50%, write “offset ≤ cT when interval ≤ T”; if covariance is zero and reduction nears 29%, write “calibration only resets the zero point, drift re-accumulates as $\sqrt{T}$” and explicitly state residual floor equals step size; if both coexist, report both parameters.

    Evidence and Limitations

    Verified: Starting from three timepoints, trend-type and random-walk-type can be distinguished under moment conditions, relying on the sign of covariance between adjacent increments rather than curve shape; simulations with 30+ raters yield over 90% discrimination rate. Calibration frequency affects the achievable lower bound differently for each pattern—an algebraic conclusion independent of data.

    Partially Verified: Empirical evidence conflicts. “Less than 1% significant drift” coexists with “time-interval coefficient 0.88 significantly positive,” suggesting patterns may vary by person and context, not as domain-wide commonalities.

    Pending Verification: No study has reported the proportion of each pattern among real assessment data. Longitudinal sequences beyond six timepoints are virtually absent in Chinese medical assessment research; thus this criterion serves only as a design specification.

    L1 Network Status

    L1 Network Status: E-060 Label Premise original text recorded as “Five-Zang differentiation (as grouping label) — premise — label reliability … conflict (unresolved).” This edge only resolves “whether overall label reliability is low or not, and whether training can improve it.” Regarding “whether drift accumulates via trend or spreads via random walk,” CORRESPONDENCE_NETWORK.md and EDGE_TABLE_science-yixue.md searched all tables for “random walk,” “trend,” “accumulate,” “fatigue,” “timepoint,” “Markov,” “state space,” “longitudinal”—zero hits. This topic has not yet entered the L1 corresponding network, so evidence level capped at C; gap registered to science-yixue-l1-gaps.md.

    Related Reading

    Evidence Level: C
    References: PMID 30955444, PMID 19769648, PMID 26837086, PMID 29795823, PMID 36284065, PMID 37283589, PMID 29881127, PMID 28587597


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

  • 那个决定折不打折的比例,其实推不动判决:真正推它的是那个从来没人问过的代价比

    Evidence level: C

    上一篇把「打折划不划算」的判决全压在一个从未被估过的比例 π 上。把这张曲面翻开来算,会发现:π 在这里的分量小到几乎可以忽略,真正在推判决的是另一个量。

    正文

    一、把公式拆开,看 π 站在哪儿

    「打折划算」的条件是 π 超过临界值 π*。这个 π* 没有闭式,但拆开来源后形式很简单:打折的功效增益记 ΔP,多付的假阳性记 Δα,则

    π* = k·Δα / (k·Δα + ΔP)

    k 是「一次假阳性」相对于「一次真检出」的代价比。用上一篇表内数值反解:ρ=0.5 时 ΔP=20.57 个百分点、Δα=4.44 个百分点,取 k=1 得 0.1775,文中报 0.178;ρ 从 0.2 到 0.9 五点逐一核对,最大偏差 0.0005。公式和表都没错。

    问题在于这三个量里只有一个被认真对待过:ΔP 和 Δα 是定积分、算得准;而 k 和 π 都是拍出来的,k 甚至从未被提起。

    二、π 不是分布参数,是尾概率

    π 定义为「被比较的两份判读里真有差的比例」,听起来像总体参数,其实取决于你把多小的差算作「有差」。

    设真实差值服从尺度为 s 的半正态(以独立标准误为单位)。取 s=0.5——典型差异只有半个标准误,判读研究里很现实——把「算有差」的门槛从 0.01 挪到 0.30 个标准误,π 就从 0.984 掉到 0.549。同一份数据、同一个分布,只因刻度线右移,π 掉了一半。 它不是能被估计的参数,而是必须连分辨率一起声明的导出量——而文献里两个都没报过。

    三、就算去估,它和「差异多大」也分不开

    用 spike-and-slab 混合模型直接数 π:模拟 40 个比较,真值 π=0.25、差异尺度 1.5 个标准误。极大似然给出 π=0.60、尺度 1.35,两个数都离真值很远。剖面似然更说明问题——尺度在 0.2 到 4.0 上扫、每个 π 取最优,对数似然下降不到 2 的区间覆盖 π 从 0.10 到 0.75。这一整条脊上,数据分不出哪个更合理。 重复 200 次,π 的估计均值 0.42、标准差 0.28,三成次数直接顶到搜索上界。

    原因是「少量真差、每个都很大」与「很多真差、每个都很小」产生几乎相同的观测分布。要拆开必须先知道其中一个——而它恰是要估的那个。

    四、最反直觉的一步:π 根本推不动判决

    到这里结论似乎是「π 估不出,规则不能用」。但把整张曲面翻开来算,方向整个反了。

    π* 的跨度确实吓人,从 0.051 到 0.990 近二十倍,可绝大部分落在没人会走的地方。取硬下界:只要分辨率不超过 0.30 个标准误,任何差异分布形状下 π 都不低于 0.549。在这个现实带内,ρ 从 0.20 到 0.95、z 从 0.5 到 3.0 的 832 个格点上,π 能推翻「打折」判决的只有 51 格,占 6%,且百分之百落在 z≥2.5 的反弹区;扣掉这块,624 个格点里一格都没有。π 取 0.55 还是 0.99,判决不同的格点只有 5.9%。对照:同样格点上把 ρ 从 0.3 变到 0.9,判决翻转 21%,是 π 的将近四倍。

    因为 π* 在会做判决的整片区域里都低于 0.2,而现实的 π 高于 0.5。门槛在地板上,判决在房顶上,中间隔着一整个楼层。 真要翻案得让 π 掉到 0.05,那意味着二十个比较里只有一个真有差——这种研究本就不该谈打折。

    五、真正在推判决的是那个没人问过的代价比

    回到第一节的式子。取 ρ=0.5,把 k 从 0.25 挪到 10(前者「假阳性很轻」、后者「极重」),π* 从 0.051 变到 0.679,跨度 13 倍。同一行数据上,π 只挪动 5.9% 的格点,k 能把整片翻过来。

    k 是什么?「一次误判相对于一次真检出值多少」,纯临床判断,数据里没有它的位置。净收益框架的原始文献说得很清楚:这个交换率必须由人指定,可以靠一个简单问题问出来——「为了找到一个真病人,最多愿让几个健康人做不必要的检查」——但答案主观,所以要把一整段合理取值都画成曲线看,不能由数据自证。

    这条规则的真正支点,不在被反复担心的 π 上,而在从未被写出来的 k 上。 上一篇把 π 当软肋,是看错了位置。

    π 仍有用途,但用途不同:它决定「赚多少」,不决定「赚不赚」。ρ=0.5、z=1.0 时每 100 次比较的净增正确判定:π=0.10 时 −1.94 次(净亏),0.25 时 +1.81 次,0.50 时 +8.06 次,0.80 时 +15.57 次。同一个「该打折」判决,收益能从亏两次到赚十五次。

    六、这条线上现有研究的状态

    必须说清:上面全是方法学推演,不是实证结论,而支撑它的经验基础薄得需要直说。

    中医五脏辨证作为分组标签,信度本身处于未裁决的冲突。一项研究直接报告:中医师之间诊断一致性在开放人群中「低」,作者用「令人担忧」形容。同系列另一篇则指出问题可能出在统计工具——偶然校正类系数在低基率、多类别场合会被严重惩罚,建议改用对基率不敏感者;且有研究显示经问卷化诊断与培训可明显改善。两边都有理,也都没能推翻对方。放射影像领域的系统评价给了另一侧面:观察者间变异研究样本量普遍偏小且无依据,多数报了一致性系数,而这些系数常常与研究结论对不上。

    至于「被判可辨的比较里真有差的比例」,方法学与临床文献都没有任何一篇报过。不是检索不全,是这个量从未被当成需要报告的东西。

    L1 网络状态:E-060 标签前提 — conflict(未裁决)。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);同系列第二篇指出偶然校正一致性统计常被误用、建议 Gwet AC2(PMID 31730403 / 31634000 / 31730404);但训练可改善(PMID 19538100)」→ 冲突双方:低一致性实测量 — 对 — 统计工具误用致低估、且训练可改善。该边指向「标签信度」,非「π 的可估性」,本文不据此边任一方作结论。

    证据与局限

    已验证:① π* 的闭式结构为 k·Δα/(k·Δα+ΔP),用表内五行数值反解最大偏差 0.0005;② π 是尾概率而非分布参数,门槛从 0.01 挪到 0.30 个标准误(s=0.5)可使 π 从 0.984 降至 0.549,属恒等式性质;③ π 与差异尺度在 spike-and-slab 下联合不可识别,剖面似然脊上 ΔlogL<2 覆盖 π∈[0.10,0.75];④ 在 π≥0.549 现实带内,832 个格点中 π 能推翻判决的仅 51 格且全在 z≥2.5,扣除后为 0/624;⑤ k 从 0.25 到 10 可使 π* 跨度 13 倍,判决翻转能力远超 π。

    部分验证:① 格点统计依赖双线性插值,原表仅 6×4 节点,插值误差未量化;② 「π≥0.549」的推导假定差值服从半正态,若重尾(少数极大差异 + 多数微小差异)下界会下移;③ 定量换算用 k=1,k≠1 时数值需缩放,但定性结论不变;④ 不可识别结论在正态 slab 下成立,换重尾 slab 后脊形变、宽度未必收窄。

    待验证:① 真实研究里 π 落在哪个量级全无数据,本文论断皆为理论下界非实测;② 分辨率该由谁定、定在哪,没有任何规范,而它直接决定 π;③ k 在判读研究语境下该取多少从未被讨论,而本文证明它比 π 重要得多;④ 「方法学与临床验证两类研究的 π 结构不同」无法用现有文献检验,因为两类都不报 π。

    相关阅读

    证据等级: C
    参考文献: PMID 36123642, PMID 26810254, PMID 17099194, PMID 18177782, PMID 37399082, PMID 34062626, PMID 28626349, PMID 30383398, PMID 31730403, PMID 19538100, PMID 16483120


    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 decision on whether to discount or not actually fails to drive the judgment; what truly drives it is the cost ratio that has never been asked about.

    Evidence level: C

    In the previous article, the decision of whether discounting is cost-effective was entirely pressed onto a ratio π that had never been estimated. Unfolding this surface for calculation reveals: the weight of π here is so small it is almost negligible; what truly drives the decision is another quantity.

    Main Text

    I. Unpacking the Formula to See Where π Stands

    The condition for “discounting is cost-effective” is that π exceeds the critical threshold π*. This π* has no closed form, but after unpacking its source, the structure is simple: let ΔP denote the efficacy gain from discounting and Δα denote the false positives incurred by paying more; then

    π* = k·Δα / (k·Δα + ΔP)

    where k is the cost ratio of “one false positive” relative to “one true detection.” Working backwards using the numerical values in the previous table: when ρ=0.5, ΔP=20.57 percentage points and Δα=4.44 percentage points; taking k=1 yields 0.1775, while the article reports 0.178; checking ρ from 0.2 to 0.9 at five points shows a maximum deviation of 0.0005. The formula and the table are both correct.

    The problem lies in these three quantities: only one is taken seriously. ΔP and Δα are definite integrals, calculated precisely; whereas k and π are both ad hoc, with k never even mentioned.

    II. π Is Not a Distribution Parameter, It Is a Tail Probability

    π is defined as “the proportion of true differences among the two sets of readings being compared,” which sounds like a population parameter, but in fact it depends on how small a difference you count as “a difference.”

    Suppose the true difference follows a half-normal distribution with scale s (in units of independent standard error). Taking s=0.5—typical differences amount to half a standard error, which is very realistic in reading studies—moving the threshold for “counting as different” from 0.01 to 0.30 standard errors causes π to drop from 0.984 to 0.549. The same data, the same distribution, yet because the scale line shifts right, π drops by half. It is not an estimable parameter; it is a derived quantity that must be declared together with the resolution—and neither has ever been reported in the literature.

    III. Even If We Estimate It, It Is Inseparable from “How Large the Difference Is”

    Using a spike-and-slab mixture model to directly count π: simulate 40 comparisons, true π=0.25, difference scale 1.5 standard errors. Maximum likelihood gives π=0.60 and scale 1.35, both far from the true values. Profile likelihood makes this even clearer—sweeping the scale from 0.2 to 4.0 and taking optimal π at each point, the log-likelihood drops by less than 2 over the interval covering π from 0.10 to 0.75. Along this entire ridge, the data cannot distinguish which is more reasonable. Repeating 200 times, the estimated mean of π is 0.42 with standard deviation 0.28; in one-third of cases it hits the upper bound of the search space.

    The reason is that “few true differences, each large” and “many true differences, each small” produce nearly identical observed distributions. To disentangle them, one must know one of them beforehand—and that is precisely the quantity we wish to estimate.

    IV. The Most Counterintuitive Step: π Cannot Drive the Decision at All

    Up to here, the conclusion seems to be “π cannot be estimated, so the rule cannot be used.” But unfolding the entire surface for calculation reverses the direction entirely.

    The span of π* is indeed startling, from 0.051 to 0.990—nearly twentyfold—but the vast majority falls in regions nobody would ever go to. Taking the hard lower bound: as long as resolution does not exceed 0.30 standard errors, under any difference distribution shape π never drops below 0.549. Within this realistic band, across the 832 grid points where ρ ranges from 0.20 to 0.95 and z from 0.5 to 3.0, π can overturn the “discounting” decision at only 51 grid points, accounting for 6%, and all of them fall entirely within the rebound zone where z≥2.5; after removing this region, none of the 624 remaining grid points can overturn it. Whether π is taken as 0.55 or 0.99, only 5.9% of the grid points yield different decisions. In contrast: on the same grid points, changing ρ from 0.3 to 0.9 flips the decision 21%, which is nearly four times what π does.

    Because π* stays below 0.2 across the entire region where decisions are actually made, while the realistic π exceeds 0.5, the threshold lies on the floor, the decision lies on the ceiling, with an entire floor in between. To overturn the verdict would require π to drop to 0.05, meaning only one out of twenty comparisons truly differs—a study that should not even discuss discounting.

    V. What Truly Drives the Decision Is That Cost Ratio Nobody Has Asked About

    Returning to the formula in Section I. Taking ρ=0.5, moving k from 0.25 to 10 (the former implies “false positives are light,” the latter “extremely heavy”), π* shifts from 0.051 to 0.679—a span of 13 times. On the same row of data, π only shifts 5.9% of the grid points, whereas k can flip the entire landscape.

    What is k? “How many times worse is one misjudgment compared to one true detection?” This is pure clinical judgment; there is no place for it in the data. The original literature on the net-gain framework states clearly: this exchange rate must be specified by humans, and can be elicited via a simple question—”To find one true patient, how many healthy people are you willing to subject to unnecessary testing?”—but the answer is subjective, so one must draw a curve over the entire range of reasonable values and cannot let the data speak for itself.

    The true pivot of this rule lies not on π, which is repeatedly worried about, but on k, which has never been written down. In the previous article, treating π as the weak link was looking at the wrong place.

    π still has use, but its role is different: it determines “how much to gain,” not “whether to gain.” For ρ=0.5 and z=1.0, the net increase in correct judgments per 100 comparisons is: −1.94 when π=0.10 (net loss), +1.81 when π=0.25, +8.06 when π=0.50, and +15.57 when π=0.80. The same “should discount” decision can yield gains ranging from losing twice to gaining fifteen times.

    VI. The Current State of Research on This Line

    It must be made clear: all the above are methodological derivations, not empirical conclusions, and the empirical foundation supporting them is so thin it needs to be stated outright.

    The traditional consensus on TCM five-organ differentiation as grouping labels remains unresolved; reliability itself is in an undecided conflict. One study directly reports: among practitioners, diagnostic consistency in the open population is “low,” with authors describing it as “concerning.” Another paper in the same series points out that the problem may lie in statistical tools—coefficient of agreement corrected for chance is severely penalized under low base rates and multi-class settings, suggesting switching to measures insensitive to base rate; furthermore, studies show that questionnaire-based diagnosis and training can significantly improve it. Both sides have merit, yet neither has overturned the other. A systematic review in radiology provides another perspective: observer variation studies generally have undersized samples without justification, most report consistency coefficients, yet these coefficients often do not match the study conclusions.

    As for “the proportion of true differences among comparisons deemed distinguishable,” neither methodology nor clinical literature has ever reported it. This is not due to incomplete searching; this quantity has never been regarded as something needing reporting.

    L1 Network Status: E-060 label premise — conflict (undecided). Original excerpt: “Consistency among TCM practitioners in the open population is ‘low,’ described by authors as ‘concerning’ (PMID 30383398); the second paper in the same series indicates that chance-corrected agreement statistics are often misused and recommends Gwet AC2 (PMID 31730403 / 31634000 / 31730404); however, training can improve it (PMID 19538100).” → The two conflicting sides: low consistency measured directly — yes; statistical tools misused leading to underestimation, and training can improve. This side points to “label reliability,” not “estimability of π”; this article does not draw conclusions based on either side.

    Evidence and Limitations

    Verified: ① The closed-form structure of π* is k·Δα/(k·Δα+ΔP); back-solving with five numerical values from the table yields a maximum deviation of 0.0005; ② π is a tail probability, not a distribution parameter; shifting the threshold from 0.01 to 0.30 standard errors (s=0.5) causes π to drop from 0.984 to 0.549, which is an identity property; ③ Under spike-and-slab, π and difference scale are jointly non-identifiable; along the profile likelihood ridge, ΔlogL<2 covers π∈[0.10,0.75]; ④ Within the realistic band where π≥0.549, among 832 grid points only 51 can overturn the decision, all lying in z≥2.5; after removing this region, none of the 624 remaining points can overturn it; ⑤ Moving k from 0.25 to 10 expands π* by 13 times, giving decision-flipping power far exceeding that of π.

    Partially Verified: ① Grid-point statistics rely on bilinear interpolation; the original table has only 6×4 nodes, so interpolation error is not quantified; ② The derivation of “π≥0.549” assumes differences follow a half-normal; under heavy tails (few extreme differences plus many tiny ones), the lower bound would shift downward; ③ Quantitative conversion uses k=1; when k≠1, numerical values need scaling, but qualitative conclusions remain unchanged; ④ The non-identifiability conclusion holds under normal slab; switching to heavy-tail slab distorts the ridge shape, and width may not narrow.

    To Be Verified: ① No data exists on what magnitude π actually falls in real studies; all assertions here are theoretical lower bounds, not empirical findings; ② There is no standard specifying who should set the resolution and where, yet it directly determines π; ③ The appropriate value of k in the context of reading studies has never been discussed, yet this article proves it is far more important than π; ④ The claim that “π structure differs between methodology and clinical validation studies” cannot be tested with existing literature, because neither type reports π.

    Further Reading

    Evidence Level: C
    References: PMID 36123642, PMID 26810254, PMID 17099194, PMID 18177782, PMID 37399082, PMID 34062626, PMID 28626349, PMID 30383398, PMID 31730403, PMID 19538100, PMID 16483120


    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.

  • 「两个参数要并排测」是个假问题:p_e 根本不是第二个参数,它是异质性的影子——真正该测的是「随机抽两个案子不一样」的那个概率

    Evidence level: C

    上一篇把报告规范压在 κmax = (1−TVD−p_e)/(1−p_e) 上,说 TVD 与 p_e 是「两个精确充分统计量」,于是担心实测里两者相关多少。这一轮把代数学完:p_e 被 TVD 和异质性 H 完全锁死,只剩 TVD²/8 量级的自由度;真正的第二个自变量从来不是 p_e,而是 H。

    正文

    上一篇(#1395)留了个待验证:真实三档裁决研究里 (TVD, p_e) 落在哪个区域、两者相关多少,一片空白;若强负相关,高一致区的警告可以降级;若近乎独立,警告要整体加重。这个设计有个前提没被检查——p_e 到底能不能算作独立于 TVD 的第二个参数。

    先把三位判读者的边际分布记成 p 与 q,写 p_e = Σpᵢqᵢ。用恒等式 (p−q)² = p²+q²−2pq,把 pq 解出来再求和,得到一个纯代数关系:p_e = S − ‖p−q‖²/4,其中 S = Σmᵢ²,m = (p+q)/2 是合并边际。这一步没有任何假设,是把 p_e 换了个写法。

    关键在第二项。给定 TVD,差向量 v = p−q 满足「正部之和 = 负部之和 = TVD」且和为 0,于是 ‖v‖² 只能在一个确定的区间里跑:最集中时 ‖v‖² = 4TVD²(全部差额压在一格),最分散时 ‖v‖² = 4TVD²·c_min,c_min 由档数决定(三档是 3/8,四档 1/4,五档 5/24)。代入上式,p_e 就被夹在 [S − TVD², S − TVD²/3] 之间,宽度是 TVD² 量级。

    所以 p_e 不是第二个独立参数,它等于「异质性」减去「分歧」。 令 H = 1 − S,H 就是「随机抽两个案子,它们的真实判读档不同的概率」——判读问题里最常见的异质性指标(simpson 意义上的 diversity)。实测最能说明问题:同一组 H = 0.5592、TVD = 0.05,分别把差额压在三个不同的格上,算出的 κmax 全是 0.9108,一位小数都不差。集中在哪一格、哪个档位是多数——全都不影响。κmax 只是 (H, TVD) 的二元函数,没有第三个自变量在起作用。

    这就把上一篇的担心反转了。它怕的是「两参数相关会不会让高一致区的警告失真」。实际上两者不是「相关」,是代数的同一件事:在 H 固定时,TVD 一升 p_e 必降(因为 ‖v‖² 必须跟着涨),斜率是 −TVD/2 量级,这不是经验相关而是恒等式的推论。而我按「研究实际分布」做的 2 万组仿真给出 r(TVD, p_e) = −0.074——几乎无相关。但那个数没有信息含量:它完全由我假设的抽样范围决定,换个范围就换个数。当一个相关系数可以被构造分布任意摆布时,把判据挂在它身上是错的。

    反过来说,二分类情形最能露出底牌。两个档时 c 恒等于 1/2(五万组随机边际实测 c 的方差为 0),于是 p_e = S − TVD²/2 是恒等式,p_e 在 (m, TVD) 之外零自由度。也就是说,越简单的判读设计,p_e 越冗余;它看起来像个独立参数,只是档数多了以后代数松了一点。

    那「松了一点」有多大? 三档、H = 0.45、TVD = 0.10 时,形状带来的 κmax 摆动是 0.0006;TVD = 0.20 时也才 0.0046。作为对照,同样是 TVD = 0.10,只把 H 从 0.80 降到 0.15,κmax 从 0.876 掉到 0.352。换形状动 0.0006,换异质性动 0.52——差三个数量级。 上一篇花在「残余跨度随 m̄ 上升」上的那一整节,量级上抓错了对象:真正决定浪费多少一致性的不是判读档位集中在哪,而是这批案子本身有多难分。

    放大关系还有个干净的读法:同样的 TVD,H 从 0.80 压到 0.15,伤害放大 5.1–5.3 倍,几乎恒定。原因是伤害约等于 TVD/H——把 TVD 除以「案子之间的差异」,得到的是一个无量纲的分歧率。H = 0.45 时,TVD/H = 0.03 / 0.06 / 0.12 / 0.25 / 0.50 分别对应伤害 0.030 / 0.060 / 0.120 / 0.247 / 0.478。这是一条近似直线,斜率约 1。 判读研究里 5 个百分点的容差,到 H = 0.15 的样本上只等于 0.75 个百分点的 TVD 容差——这就是「高一致区警告」的真实来源,跟边际形状无关。

    上一轮的另一半担心也顺手解决了。它怕 TVD̂ 正偏与 p̂_e 的偏倚在 κmax 里部分对冲,使「n = 71 时偏倚 −0.105」这个数要整体修正。把偏倚按两路拆开:TVD 路(固定 p_e,只换成 TVD̂)与 p_e 路(固定 TVD,只换成 p̂_e)。真表 TVD = 0.03、n = 71 时,总偏倚 −0.1076,其中 TVD 路 −0.1066、p_e 路 −0.0004——TVD 路占 99.6%。TVD = 0.08 时占 98.2%。n = 400 时仍是 98.8%。所谓对冲比不是 50%,是 1.3%:p_e 这一路小到不构成一个可以对冲的量。 原因是 pe 路对 κmax 的偏导被 (1−p_e)² 压住,而 p_e 的量级本身又只有 0.3–0.5。

    顺带发现 TVD̂ 的偏倚是相当大的:真 TVD = 0 时,n = 71 下平均估出 0.0884,n = 3000 才降到 0.0135。这意味着一个真正完全对齐的判读系统,在 71 例的样本里会被报告成「边际有 8.8 个百分点的分歧」。这一项直接进 κmax,按 −TVD/H 折算,H = 0.45 时就是 −0.196 的天花板——比「−0.105」还重。上一篇报的 −0.105 对应的是 TVD = 0.03 这个特定真值,不是普遍情形。

    最后是上一篇那条最硬的判据本身没法执行。「取 ≥8 份三档判读原始表逐份算 (TVD, p_e) 打点,判相关系数的绝对值大于 0.6 则重算曲线、小于 0.3 则独立报」。八份表下 r 的 95% 区间是 [−0.703, +0.705];真 r = 0.6 时判对概率 55.2%,真 r = 0.3 时 40.8%——判对的概率比抛硬币好一点点。 扩到 120 份,真 r = 0.6 判对 51.1%、真 r = 0.3 判对 49.6%,因为判据线正好压在零假设分布的两侧尾上,加样本不解决问题。而按本轮的代数,这个相关本来就不该测——它测的是一个被构造分布支配的量。

    证据与局限

    已验证(纯代数,可复算):
    – 恒等式 p_e = S − ‖p−q‖²/4,三十万组随机三档边际零偏差。
    – p_e 的可行区间 [S − TVD², S − 3TVD²/8](三档),四十万组网格零越界。档数 K 的一般式为 c_min = (1/4)(1/⌈K/2⌉ + 1/⌊K/2⌋),K = 2…8 全部与实测吻合(K = 2 实测 c 方差为 0)。
    – κmax 只依赖 (H, TVD):同 (H, TVD) 下不同集中格给出完全相同的 κmax;形状带宽在三档 TVD = 0.20 时仅 0.0046。
    – 伤害 ≈ TVD/H,斜率约 1;H 从 0.80 到 0.15 放大 5.1–5.3 倍。
    – 偏倚分解:TVD 路占 98–99.7%,p_e 路 < 2%,对冲比接近 1(即无对冲)。

    部分验证:
    – TVD̂ 正偏的量级(n = 71 时真值 0 估成 0.088)来自独立二项抽样模拟,与 #1404 的漂移地板是同源推理,但未经真实三档表的联合 bootstrap 核验。
    – H = 1 − S 作为「案子难度」的代理,与经典 simpson diversity 同构,但判读问题里「真实档」不可观测,H 只能从合并边际估,这一步的估计误差本轮未算。

    待验证:
    – H 的抽样误差有多大、它占 κmax 误差预算多少——若 H 本身估得不准,「伤害 = TVD/H」这个比值会把两路误差放大成一个更糟的量。
    – 判读档位是有序的(轻/中/重),而 TVD 与 H 都是无序指标,丢失了「错一格 vs 错两格」的区分;有序情形下 H 是否还够用、还是必须换成累积分布的差异,未验。
    – 8 份表的可分辨性结果是按 Fisher z 近似算的,真实三档表的 (TVD, p_e) 联合分布不是二元正态,实际判决力可能更差。

    L1 网络状态:CORRESPONDENCE_NETWORK.md 与 EDGE_TABLE_science-yixue.md 全表检索「TVD / 边际 / 天花板 / p_e / 一致性 / 信度 / 判读」无对应边:唯一相关命中为 §3.8「E-060 标签前提」行(状态 conflict,讨论「中医医师间诊断一致性低 vs 统计误用致低估」这一标签信度前提),语义是「分组标签可不可靠」,不是「天花板由哪几个量决定」;其余命中均无关(第 192 行「生克一致性」指五行生克方向、物理侧第 57 行「相位 / 节律」指两振荡锁相)。边表止于 E-050,对齐指标行原文「L2 以上边数:0」。该主题尚未进入 L1 对应网络。

    相关阅读

    证据等级: C
    参考文献: PMID 8501467, PMID 8621993, PMID 23776034, PMID 25168681, PMID 37127220, PMID 38283651

    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。

    Two parameters side-by-side is a false problem: p_e isn’t the second parameter; it’s the shadow of heterogeneity—the true measure is the probability that two randomly drawn case…

    Evidence level: C

    In the previous post, the report specification was pressed against $\kappa_{\text{max}} = (1-\text{TVD}-p_e)/(1-p_e)$, stating that TVD and $p_e$ are “two exact sufficient statistics,” thus raising concern about how much they correlate in actual measurement. This round completes the algebra: $p_e$ is completely locked by TVD and heterogeneity $H$, leaving only degrees of freedom on the order of $\text{TVD}^2/8$; the true second independent variable is never $p_e$, but rather $H$.

    Body

    The previous post (#1395) left an unverified point: in real three-tier adjudication studies, which region $(\text{TVD}, p_e)$ falls into and how much they correlate remains blank; if the correlation is strongly negative, warnings for high-consistency regions can be downgraded; if nearly independent, warnings must be overall intensified. This design has a premise that was not checked—whether $p_e$ can truly be counted as the second parameter independent of TVD.

    Let us denote the marginal distributions of the three adjudicators as $p$ and $q$, writing $p_e = \sum p_i q_i$. Using the identity $(p-q)^2 = p^2 + q^2 – 2pq$, we solve for $pq$ and sum, yielding a purely algebraic relation: $p_e = S – |p-q|^2/4$, where $S = \sum m_i^2$ and $m = (p+q)/2$ is the merged marginal. This step involves no assumption; it merely rewrites $p_e$.

    The key lies in the second term. Given TVD, the difference vector $v = p-q$ satisfies “sum of positive parts = sum of negative parts = TVD” and sums to 0, so $|v|^2$ can only run within a fixed interval: most concentrated, $|v|^2 = 4\text{TVD}^2$ (all differences compressed into one cell); most dispersed, $|v|^2 = 4\text{TVD}^2 \cdot c_{\min}$, where $c_{\min}$ is determined by the number of tiers (for three tiers $c_{\min} = 3/8$, for four tiers $1/4$, for five tiers $5/24$). Substituting into the equation, $p_e$ is squeezed between $[S – \text{TVD}^2, S – \text{TVD}^2/3]$, with width on the order of $\text{TVD}^2$.

    Thus $p_e$ is not the second independent parameter; it equals “heterogeneity minus divergence.” Let $H = 1 – S$; $H$ is “the probability that two randomly drawn cases have different true adjudication tiers”—the most common heterogeneity indicator in adjudication problems (diversity in the Simpson sense). Actual measurement best illustrates the point: with the same $H = 0.5592$ and $\text{TVD} = 0.05$, compressing the difference into three different cells respectively, the calculated $\kappa_{\text{max}}$ is all $0.9108$, not differing by even one decimal place. Where it concentrates, which tier is the majority—all do not affect anything. $\kappa_{\text{max}}$ is simply a binary function of $(H, \text{TVD})$; no third independent variable plays a role.

    This reverses the previous concern. It feared “whether correlation between two parameters would distort warnings for high-consistency regions.” In fact, they are not “correlated”; they are algebraically the same thing: when $H$ is fixed, as TVD rises, $p_e$ must fall (because $|v|^2$ must increase accordingly), with slope on the order of $-\text{TVD}/2$. This is not empirical correlation but a consequence of the identity. Our 20,000 simulations based on “actual study distribution” yield $r(\text{TVD}, p_e) = -0.074$—almost no correlation. But that number carries no information content: it is entirely determined by the assumed sampling range; change the range, and the number changes. When a correlation coefficient can be arbitrarily manipulated by constructed distribution, hanging the criterion on it is wrong.

    Conversely, the two-tier case most clearly reveals the bottom line. With two tiers, $c$ is constantly $1/2$ (in 50,000 random marginal tests, the variance of $c$ is 0), so $p_e = S – \text{TVD}^2/2$ is an identity, and $p_e$ has zero degrees of freedom outside $(m, \text{TVD})$. That is, the simpler the adjudication design, the more redundant $p_e$ becomes; it looks like an independent parameter only because algebra loosens slightly as the number of tiers increases.

    How much does it loosen? With three tiers, $H = 0.45$, $\text{TVD} = 0.10$, the shape-induced swing in $\kappa_{\text{max}}$ is $0.0006$; even with $\text{TVD} = 0.20$, it is only $0.0046$. As contrast, with the same $\text{TVD} = 0.10$, reducing $H$ from $0.80$ to $0.15$ causes $\kappa_{\text{max}}$ to drop from $0.876$ to $0.352$. Changing shape moves $0.0006$, changing heterogeneity moves $0.52$—a difference of three orders of magnitude. The entire section in the previous post spent on “residual span increasing with $\bar{m}$” caught the wrong object in terms of magnitude: what truly determines how much consistency is wasted is not where the adjudication tiers concentrate, but how hard these cases themselves are to distinguish.

    The amplification relationship also has a clean reading: with the same TVD, compressing $H$ from $0.80$ to $0.15$ amplifies harm by $5.1–5.3$ times, almost constant. This is because harm is approximately $\text{TVD}/H$—dividing TVD by “difference between cases” yields a dimensionless divergence rate. When $H = 0.45$, $\text{TVD}/H = 0.03/0.06/0.12/0.25/0.50$ correspond to harms $0.030/0.060/0.120/0.247/0.478$. This is nearly a straight line with slope about 1. In adjudication studies, a 5-percentage-point tolerance corresponds to only $0.75$-percentage-point TVD tolerance when $H = 0.15$—this is the true source of “warnings for high-consistency regions,” independent of marginal shape.

    The other half of the previous concern is also resolved. It feared that positive bias in $\widehat{\text{TVD}}$ and bias in $\widehat{p_e}$ would partially offset each other in $\kappa_{\text{max}}$, requiring overall correction of the number “bias $-0.105$ when $n = 71$. Splitting the bias into two paths: TVD path (fix $p_e$, replace with $\widehat{\text{TVD}}$) and $p_e$ path (fix TVD, replace with $\widehat{p_e}$). With true $\text{TVD} = 0.03$ and $n = 71$, total bias is $-0.1076$, where TVD path contributes $-0.1066$ and $p_e$ path $-0.0004$—the TVD path accounts for 99.6%. When $\text{TVD} = 0.08$, it accounts for 98.2%. Even at $n = 400$, it remains 98.8%. The so-called offset ratio is not 50%, but 1.3%: the $p_e$ path is too small to constitute an offset quantity. This is because the partial derivative of $\kappa_{\text{max}}$ with respect to $p_e$ is suppressed by $(1-p_e)^2$, while the magnitude of $p_e$ itself is only $0.3–0.5$.

    Incidentally, we discover that the bias in $\widehat{\text{TVD}}$ is quite large: when true $\text{TVD} = 0$, with $n = 71$ the average estimate is $0.0884$, dropping to $0.0135$ only at $n = 3,000$. This means a truly perfectly aligned adjudication system would be reported as “marginal divergence of 8.8 percentage points” in a sample of 71 cases. This term directly enters $\kappa_{\text{max}}$; converting via $-\text{TVD}/H$, when $H = 0.45$ it reaches a ceiling of $-0.196$—heavier than the reported “$-0.105$.” The $-0.105$ reported previously corresponds to the specific true value $\text{TVD} = 0.03$, not the general case.

    Finally, the hardest criterion itself from the previous post cannot be executed: “Take $\ge 8$ three-tier adjudication raw tables, calculate $(\text{TVD}, p_e)$ point-by-point, if absolute value of correlation coefficient $> 0.6$ recalculate curve, $< 0.3$ report independently.” With eight tables, the 95% interval of $r$ is $[-0.703, +0.705]$; when true $r = 0.6$, probability of correct judgment is $55.2\%$, when true $r = 0.3$ it is $40.8\%$—probability of correct judgment is slightly better than flipping a coin. Expanding to 120 tables, true $r = 0.6$ yields $51.1\%$ correct judgment, true $r = 0.3$ yields $49.6\%$, because the criterion line sits exactly on both tails of the null hypothesis distribution; adding samples does not solve the problem. And according to this round’s algebra, this correlation should not be measured at all—it measures a quantity dominated by constructed distribution.

    Evidence and Limitations

    Verified (pure algebra, reproducible):
    – Identity $p_e = S – |p-q|^2/4$, zero bias across 300,000 random three-tier marginals.
    – Feasible interval for $p_e$: $[S – \text{TVD}^2, S – 3\text{TVD}^2/8]$ (three tiers), zero out-of-bound across 400,000 grid points. General formula for $K$ tiers: $c_{\min} = (1/4)(1/\lceil K/2 \rceil + 1/\lfloor K/2 \rfloor)$, all $K = 2\ldots8$ match actual measurement (for $K=2$, variance of $c$ is 0).
    – $\kappa_{\text{max}}$ depends only on $(H, \text{TVD})$: different concentrated cells under same $(H, \text{TVD})$ yield identical $\kappa_{\text{max}}$; shape bandwidth is only $0.0046$ at three-tier $\text{TVD} = 0.20$.
    – Harm $\approx \text{TVD}/H$, slope about 1; $H$ from $0.80$ to $0.15$ amplifies harm by $5.1–5.3$ times.
    – Bias decomposition: TVD path accounts for $98–99.7\%$, $p_e$ path $< 2\%$, offset ratio nearly 1 (i.e., no offset).

    Partially verified:
    – Magnitude of positive bias in $\widehat{\text{TVD}}$ (true value $0$ estimated as $0.088$ when $n=71$) comes from independent binomial sampling simulation, sharing the same drift-floor reasoning as #1404, but not yet cross-validated by joint bootstrap on real three-tier tables.
    – $H = 1 – S$ as proxy for “case difficulty” is isomorphic to classic Simpson diversity, but in adjudication problems “true tiers” are unobservable; $H$ can only be estimated from merged marginals, and this estimation error is not calculated in this round.

    To be verified:
    – How large is the sampling error of $H$, and what proportion of $\kappa_{\text{max}}$ error budget does it consume—if $H$ itself is poorly estimated, the ratio “harm = TVD/H” will amplify errors from both paths into an even worse quantity.
    – Whether adjudication tiers are ordered (light/medium/heavy), while TVD and $H$ are both unordered indicators, losing the distinction between “off by one tier vs off by two tiers”; in ordered cases, is $H$ still sufficient or must it be replaced by cumulative distribution differences, unverified.
    – The resolvability result for 8 tables was calculated using Fisher $z$ approximation; the joint distribution of $(\text{TVD}, p_e)$ on real three-tier tables is not bivariate normal, so actual adjudicative power may be even worse.

    L1 Network Status: No corresponding edges found in full retrieval of CORRESPONDENCE_NETWORK.md and EDGE_TABLE_science-yixue.md for terms “TVD / marginal / ceiling / $p_e$ / consistency / reliability / adjudication”. The only relevant hit is §3.8 “E-060 label premise” row (status conflict, discussing “low diagnostic consistency among TCM physicians vs statistical misuse leading to underestimation”—this label reliability premise), whose semantics is “whether grouping labels are reliable,” not “which quantities determine the ceiling”; all other hits are irrelevant (row 192 “sheng-ke consistency” refers to Five Phases sheng-ke direction, physical side row 57 “phase/rhythm” refers to two oscillators phase-locked). Edge table stops at E-050; alignment indicator row states “L2 and above edge count: 0.” This topic has not yet entered the L1 correspondence network.

    Related Reading

    Evidence Level: C
    References: PMID 8501467, PMID 8621993, PMID 23776034, PMID 25168681, PMID 37127220, PMID 38283651

    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.

  • 「m≥0.95 判安全」这条线是用估出来的数画的:真到门槛的人一半能过,加多少样本都压不到零

    Evidence level: C

    一条判据线如果画在它自己那个「被估出来的门槛」上,那么无论加多少样本,它的通过率都只能趋近 50%,永远回不到 0。

    一、门槛是从一个恒等式里来的,而恒等式偷偷把 m 当成了已知数

    1393 的核心门槛是「m≥0.95 ⇒ δ>5pp 数学上不可达」,依据是一条恒等式:(1−m)×100 就是 δ 的上界,m=0.95 时恰好等于 5。这条式子本身没错。但它有一个没说出口的前提——m 是已知的。

    实际的 m 是数出来的:n 个人判读,m̂ = 合格数 / n。只要 n 有限,m̂ 就带抽样误差,SE = √(m(1−m)/n);m=0.95 时 SE = √(0.0475/n)。代进常见样本量:

    n m̂ 的 SE 95% 区间半宽 半宽 ÷ 5pp 预算
    30 3.98pp 7.80pp 1.56
    60 2.81pp 5.51pp 1.10
    73 2.55pp 5.00pp 1.00
    120 1.99pp 3.90pp 0.78
    240 1.41pp 2.76pp 0.55
    480 0.99pp 1.95pp 0.39
    960 0.70pp 1.35pp 0.27

    要让 m̂ 的区间半宽正好等于 5pp(也就是把整个 δ 预算吃光),需要 n=73。低于 73 例的研究,它的 m̂ 误差比它能担保的余量还大。 这个比值有闭式:噪声 ÷ 余量 = √(19/n),n=19 时恰好等于 1——样本量低于 19,后面关于「安全」的话都不成立。

    二、判据线压在门槛上,就永远不比抛硬币更准

    现行规则是「m̂ ≥ 0.95 判安全」。算真 m 正好等于 0.95(就在门槛上)时的通过率:n=60 是 64.7%,n=240 是 57.6%,n=480 是 55.4%,n=960 是 53.8%,n=5000 是 51.7%。

    单调趋向 50%,永远回不到 0。原因是结构性的:判据线与真值门槛重合时,误差对称地分在两边,一半在上、一半在下。 被它判为「安全」的那一半,恰好是运气好的那一半。这说明 #1404 发现的「判据线正好压在噪声底上」不是巧合,而是通病:拿一个带误差的估计,去跟它自己的定义线比。

    三、误差不是对称的:既放行两倍超标者,又拒掉合格者

    n=60 时的完整通过率(真 m → 通过率):0.90→13.7%,0.93→38.7%,0.96→78.1%,0.97→89.4%,0.99→99.7%。两端都有问题。

    真 m=0.90 意味着 δ 上界 10pp,是预算的两倍,仍有 13.7% 被放行;真 m=0.93(7pp,超标 40%)有 38.7% 被放行。反过来,真 m=0.97(3pp,余量充足)只有 89.4% 通过——合格者被拒掉十分之一。 同时具备「放行两倍超标者」和「驳回合规者」两个缺陷,说明它不是保守或激进的偏斜,而是根本没有分辨力。

    抽样模拟(真 m 在 0.90–0.99 间均匀取)更直接:n=60 时被放行的研究里有 33.8% 实际 δ>5pp;n=240 时仍有 17.8%。被放行者的真实上界 P90,n=60 是 7.62pp,n=240 是 5.71pp。这批研究报告上写着「≤5pp」,但其中三成到一成七,真实值已经越线。

    四、诚实的数:报告「5pp」的时候,真值可能是 12.4pp

    m̂ 恰好等于 0.95(刚好卡线过关)时,m 的 95% 单侧下界(精确二项法)给出的诚实 δ 上界是:

    n m 的单侧下界 诚实 δ 上界 是名义 5pp 的几倍
    60 0.8758 12.4pp 2.5
    120 0.9037 9.6pp 1.9
    240 0.9202 8.0pp 1.6
    480 0.9304 7.0pp 1.4
    960 0.9368 6.3pp 1.3

    即便做到 n=960,卡线过关也只能担保 6.3pp。想担保到 5pp 理论上要 n 趋于无穷——5pp 是这条线的渐近值,不是任何有限样本能报出的数。 凡报告写「δ 上界 ≤5pp」,实际含义是「m̂ 刚好过了 0.95」,而不是「δ 被证明 ≤5pp」,两者差 1.3 到 2.5 倍。

    五、改法:用精确单侧上界,不用点估计过线

    正确判据不是「m̂ 过线」,而是「合格率的下界过线」,等价于「失败率的精确单侧上界 ≤5%」。这个量可精确算(Clopper-Pearson)。

    (1)零失败时最简单。 全合格时失败率上界 = 1−0.05^(1/n)。要它 ≤5%,需 n ≥ ln0.05/ln0.95 = 58.4,即 59 例。n=30 时零失败上界 9.50pp,永远过不了关——30 例这个规模从一开始就没资格谈 5pp。n=60 给 4.87pp,n=120 给 2.47pp,n=240 给 1.24pp。

    (2)允许失败时按 n 定 c_max。 n=120 允许 1 例,n=240 允许 6 例,n=480 允许 15 例,n=960 允许 36 例,n=3000 允许 130 例。换算成等价名义门槛 t = 1 − c_max/n:0.9917 / 0.9750 / 0.9688 / 0.9625 / 0.9567。n=6000 才降到 0.9548——名义门槛应抬到 0.96–0.99,不是 0.95。*

    代价是真实的。精确判据很保守:真 m=0.97 时,n=240 通过率仅 41.8%,n=480 是 63.0%,n=960 才到 92.3%。要「守住 5pp、又有八成把握放行合规研究」,n 得在 1000 上下。

    六、真正的瓶颈:这笔样本量买在了错的那一项上

    1404 已算出:漂移导致的 δ̂ 标准误,n 无穷时停在 5.31pp,不随样本量下降。把 m̂ 的误差按方差与之相加:

    n 漂移项 m̂ 项 合计
    60 5.31pp 2.81pp 6.01pp
    240 5.31pp 1.41pp 5.49pp
    960 5.31pp 0.70pp 5.36pp
    2000 5.31pp 0.49pp 5.33pp

    从 60 例加到 2000 例,总误差只从 6.01pp 降到 5.33pp,降了 11%:n 只能治 m̂ 那一项(门槛侧),治不了漂移那一项(输入侧),而漂移本来就是大项。

    所以「为满足精确判据把 n 抬到 1000」这条建议,买的是较小那一项。抬到 1000 后误差结构变成 5.31 + 0.70,瓶颈完全落在漂移上,再抬 n 一分钱效果都没有。样本量能买到的,只是把门槛侧误差抹平;它抹不平尺子本身在动。

    证据与局限

    已验证:m̂ 抽样误差公式、各样本量下的通过率、Clopper-Pearson 精确上界、n≥59 的零失败下限、t* 序列——全部是纯数学关系,可直接复算,不依赖任何实测数据。

    部分验证:#1404 的漂移地板 5.31pp 是换算链产物,非实测值;两处误差按方差相加的前提是二者独立,与 #1424 要量的相关系数直接相关,目前未测。

    待验证:① 漂移与 m̂ 误差是否真独立;② 贝叶斯收缩估计能否在 n 较小时把「放行超标者」比例压下来,若不能说明是信息不足只能加 n;③ 若研究者看到 m̂ 后追加样本(可选停止),上述精确上界失效,须改用序贯设计上界。

    该主题尚未进入 L1 对应网络:边表止于 E-050,检索「门槛 / 阈值 / 样本量 / 置信 / 抽样 / 比例 / 基准 / kappa / 一致性 / 判读 / 信度」均无对应边,仅第 192 行「生克一致性」字面相近但所指不同。按硬约束,本篇证据等级封顶为 C。

    L1 网络状态:无对应边(边表末项 E-050,概念表内无「一致性/门槛/样本量」类节点;第 192 行仅有「生克一致性|相生/相克方向能否复现|未测」,与判读一致性无关)

    相关阅读

    证据等级: C
    参考文献: PMID 24338600, PMID 15733050, PMID 8501467, PMID 36123642, PMID 29707825, PMID 26519187


    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 line \”m≥0.95 is safe\” was drawn using an estimated number: half of those who truly reach the threshold pass, and adding more samples never drives it to zero.

    Evidence level: C

    A decision line drawn at its own “estimated threshold” will have a pass rate approaching 50% regardless of sample size, never returning to zero.

    I. The Threshold Derives from an Identity That Secretly Treats m as Known

    The core threshold of #1393, “m≥0.95 ⇒ δ>5pp mathematically unreachable,” relies on the identity: (1−m)×100 is the upper bound of δ; when m=0.95, this equals exactly 5. This equation itself is correct. However, it has an unstated premise—m is known.

    The actual m is estimated: n people judge, m̂ = qualified count / n. As long as n is finite, m̂ carries sampling error, SE = √(m(1−m)/n); at m=0.95, SE = √(0.0475/n). Substituting common sample sizes:

    n m̂ SE 95% CI half-width half-width ÷ 5pp budget
    30 3.98pp 7.80pp 1.56
    60 2.81pp 5.51pp 1.10
    73 2.55pp 5.00pp 1.00
    120 1.99pp 3.90pp 0.78
    240 1.41pp 2.76pp 0.55
    480 0.99pp 1.95pp 0.39
    960 0.70pp 1.35pp 0.27

    To make the CI half-width exactly equal 5pp (consuming the entire δ budget), n must be 73. Studies with fewer than 73 cases have m̂ error larger than the margin they can may support. This ratio has a closed form: noise ÷ margin = √(19/n); at n=19, it equals exactly 1—below this sample size, all subsequent claims about “safety” do not hold.

    II. When the Decision Line Presses Against the Threshold, Accuracy Never Exceeds Coin Flipping

    Current rules state “m̂ ≥ 0.95 implies safety.” Calculating the pass rate when true m equals exactly 0.95 (right at the threshold): n=60 yields 64.7%, n=240 yields 57.6%, n=480 yields 55.4%, n=960 yields 53.8%, and n=5000 yields 51.7%.

    Monotonically approaching 50%, never returning to zero. The reason is structural: when the decision line coincides with the true threshold, error splits symmetrically—half above, half below. The half judged “safe” is precisely the lucky half. This indicates that #1404’s finding of “the decision line presses exactly on the noise floor” is not coincidental but a systemic flaw: comparing an estimated quantity with its own definition line.

    III. Error is Asymmetric: It Releases Twice the Exceeders and Rejects Qualified Cases

    At n=60, the complete pass rate (true m → pass rate): 0.90→13.7%, 0.93→38.7%, 0.96→78.1%, 0.97→89.4%, 0.99→99.7%. Both ends exhibit problems.

    True m=0.90 implies δ upper bound 10pp, twice the budget; yet 13.7% are released. True m=0.93 (7pp, exceeding by 40%) has 38.7% released. Conversely, true m=0.97 (3pp, ample margin) only passes 89.4%—qualified cases rejected at one-tenth. Simultaneously exhibiting “releasing twice the exceeders” and “rejecting compliant cases” reveals it is neither conservative nor radical bias, but lacks discriminative power entirely.

    Simulation with uniform true m between 0.90–0.99: at n=60, 33.8% of released studies actually have δ>5pp; at n=240, still 17.8%. The true P90 upper bound for released studies: n=60 is 7.62pp, n=240 is 5.71pp. These reports claim “≤5pp,” yet one-third to seventeen percent actually exceed the line.

    IV. Honest Numbers: When Reporting “5pp,” True Value May Be 12.4pp

    When m̂ exactly equals 0.95 (just crossing the line), the honest δ upper bound given by m’s 95% one-sided lower bound (exact binomial method):

    n m one-sided lower bound honest δ upper bound multiple of nominal 5pp
    60 0.8758 12.4pp 2.5
    120 0.9037 9.6pp 1.9
    240 0.9202 8.0pp 1.6
    480 0.9304 7.0pp 1.4
    960 0.9368 6.3pp 1.3

    Even at n=960, crossing the line may support only 6.3pp. To may support 5pp theoretically requires n approaching infinity—5pp is the asymptotic value of this line, not a number any finite sample can report. Any report stating “δ upper bound ≤5pp” actually means “m̂ just passed 0.95,” not “δ proven ≤5pp”—a difference of 1.3 to 2.5 times.

    V. Correction: Use Exact One-Sided Upper Bound, Not Point Estimate Crossing Line

    The correct criterion is not “m̂ crosses the line,” but “lower bound of pass rate crosses the line,” equivalent to “exact one-sided upper bound of failure rate ≤5%.” This quantity can be precisely calculated (Clopper-Pearson).

    (1) Zero failures is simplest. When all pass, failure rate upper bound = 1−0.05^(1/n). To keep it ≤5%, need n ≥ ln0.05/ln0.95 = 58.4, i.e., 59 cases. At n=30, zero-failure upper bound is 9.50pp—never passes the test—30 cases lack qualification for 5pp from the start. At n=60, gives 4.87pp; n=120 gives 2.47pp; n=240 gives 1.24pp.

    (2) When allowing failures, set c_max by n. n=120 allows 1 failure, n=240 allows 6, n=480 allows 15, n=960 allows 36, n=3000 allows 130. Converted to equivalent nominal threshold t = 1 − c_max/n: 0.9917 / 0.9750 / 0.9688 / 0.9625 / 0.9567. Only at n=6000 does it drop to 0.9548—nominal threshold should be raised to 0.96–0.99, not 0.95.*

    The cost is real. Exact criteria are conservative: at true m=0.97, n=240 pass rate is only 41.8%, n=480 is 63.0%, and n=960 reaches 92.3%. To “maintain 5pp with 80% confidence to release compliant studies,” n must be around 1000.

    VI. True Bottleneck: Sample Size Bought the Wrong Term

    1404 calculated: drift-induced δ̂ standard error, as n→∞, stabilizes at 5.31pp—does not decrease with sample size. Adding m̂ error via variance:

    n drift term m̂ term total
    60 5.31pp 2.81pp 6.01pp
    240 5.31pp 1.41pp 5.49pp
    960 5.31pp 0.70pp 5.36pp
    2000 5.31pp 0.49pp 5.33pp

    Increasing from 60 to 2000 cases, total error drops only from 6.01pp to 5.33pp—a 11% reduction: n can only fix the m̂ term (threshold side), not the drift term (input side), which is already the dominant component.

    Thus, “raising n to 1000 for exact criteria” buys the smaller term. After raising to 1000, error structure becomes 5.31 + 0.70—the bottleneck lies entirely in drift; further increasing n yields no effect. Sample size can only flatten the threshold-side error; it cannot correct the ruler itself moving.

    Evidence and Limitations

    Verified: m̂ sampling error formula, pass rates at various sample sizes, Clopper-Pearson exact upper bounds, n≥59 zero-failure lower bound, t* sequence—all pure mathematical relations, directly reproducible without empirical data.

    Partially verified: #1404’s drift floor of 5.31pp is a conversion chain product, not measured; adding two errors via variance assumes independence, directly related to #1424’s correlation coefficient, currently unmeasured.

    Pending verification: (1) whether drift and m̂ error are truly independent; (2) whether Bayesian shrinkage estimation can reduce “releasing exceeders” proportion at small n—if not, information insufficiency requires increasing n; (3) if researchers add samples after observing m̂ (optional stopping), exact upper bounds fail, requiring sequential design upper bounds.

    This topic has not entered the L1 corresponding network: edge table stops at E-050; searching “threshold / threshold / sample size / confidence / sampling / proportion / baseline / kappa / consistency / interpretation / reliability” yields no corresponding edges; only row 192 “生克一致性” (generative-restraining consistency) is superficially similar but refers to something different. Under hard constraints, this topic’s evidence level caps at C.

    L1 Network Status: No corresponding edges (edge table final item E-050, concept table lacks “consistency/threshold/sample size” nodes; row 192 contains only “generative-restraining consistency | whether generative/restraining directions can replicate | unmeasured,” unrelated to interpretation consistency)

    Further Reading

    Evidence Level: C
    References: PMID 24338600, PMID 15733050, PMID 8501467, PMID 36123642, PMID 29707825, PMID 26519187


    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.

  • 两位读者占比不对称,判据线其实一动不动:真正推它的是「两人一起往哪边偏」,倍数正好是 1+2E

    Evidence level: C

    上一轮担心非对称会把判据线挪走。算完发现挪不动——一毫米都不动。但同一次计算翻出了另一处漏项,而它的倍数是干净的 1+2E,跟主档占比和主档差都无关。

    正文

    上一篇留了个悬案。判据线 2ε* = δ/(4(1−m̄)) 是从「两位读者主档占比对称地夹在均值两侧」这个参数族化出来的。真实研究里一位 0.95、另一位 0.85 这种事太常见,于是有个合理的担心:既然不对称,两个剩余空间 (1−m_A) 与 (1−m_B) 就不再相等,判据线会不会被挪?若被压低超过 10%,按对称表设计的那些研究(m̄=0.70 时 n=49 就够)实际上样本量不足,而这批研究已经照着表在做了。

    先把这条线钉死。

    原始闭式是 ε* = δ/(2(s_A+s_B)),分母是两者之和。代入 s_A = 1−m̄−δ/2、s_B = 1−m̄+δ/2,和恒等于 2(1−m̄)——那个 −δ/2 与 +δ/2 正好抵掉。所以 δ/(2(s_A+s_B)) ≡ δ/(4(1−m̄)),这是恒等式,不是近似。全域扫 m̄∈[0.55,0.99]、δ∈[0.02,0.15],偏差为 0;再随机抽 20 万组互不相关的主档占比,最大偏差 1.4×10⁻¹⁵——纯浮点误差。

    结论有点扫兴:非对称根本不挪判据线,一分都不挪。上一篇担心的「朝哪个方向偏」,答案是「没有方向」。

    顺着这个恒等式往下看,会撞见一件更别扭的事。

    参数族 m_A = m̄+δ/2、m_B = m̄−δ/2 被叫作「对称」,可它算出来的 s_A − s_B 恒等于 −δ。δ=0 时它才真的对称;只要两位读者的主档占比不同,两人的剩余空间就必然差 δ,没有例外。也就是说,从来不存在「对称情形」这个特例——那个式子一直是通式,只是被贴错了标签。这恰好解释了它为什么挪不动:它不是从对称假设推出来的,抵掉的那两项是代数结构,不是巧合。

    真正的漏项在另一处。

    把两条次档边际差写开,会得到一个干净的分解:w = s̄·Δ − δ·E。这里 s̄ 是剩余空间均值,Δ=e_A−e_B 是两位读者偏离方向之差,E=(e_A+e_B)/2 是共同偏移——两人一起往第二档偏、还是一起往第三档偏。而 TVD = max(δ, |w|+δ/2),所以「不越线」的条件是 |w| ≤ δ/2。

    现行判据 |Δ| ≤ 2ε* 只写了 w 的第一项,把 −δ·E 整项当零扔掉了。它成立的唯一条件是 E=0——两位读者的偏离正好一正一负、大小相等。这条判据不是错,是只在一种测度为零的巧合下精确。

    把 δE 补回去,解的形貌漂亮得不像话:允许的 Δ 上界从 r/2 变成 r(E+1/2),两者相比正好是 1+2E,r=δ/s̄ 整个消掉。判据该放宽还是该收紧,跟主档占比、跟主档差一点关系都没有,只跟「两个人一起往哪边偏」有关,一倍就是两倍 E。六组不同 (m̄,δ,E) 组合核验,实测比与 1+2E 逐位吻合。

    方向比幅度重要得多。 E>0(两人都往第二档偏)时判据偏保守、多留余量,E=0.10 只放宽 20%,几乎无害。E<0(两人都往第三档偏)时判据偏宽松,按它放行的研究实际不达标:E=−0.05 收 10%,−0.20 收 40%,−0.50 收到 0——判据完全失效,一个安全区都不剩。

    换算到 κ 天花板上更刺眼。m̄=0.90、δ=0.05、Δ 卡在判据线上:E=+0.10 时 κ_max 从 0.7315 掉到 0.7312,跌 0.0003,可忽略;E=−0.10 时掉到 0.7044,跌 0.0271,2.7 个 κ 点。同一个「偏离量级」,换一个方向,代价差九十倍。

    这里必须接上研究内核那条尚未裁决的冲突边:我们整个网络有一个分组前提是「用五脏辨证给受试者贴标签」,而这条前提本身正处在冲突状态。悲观一侧是中医医师间诊断一致性在开放人群中「低」,原作者称之为「令人担忧」(PMID 30383398);乐观一侧来自同一系列——该「低」结论有相当部分来自统计误用,改用 Gwet AC2 重算后线性加权一致率 0.80、偶然校正后的 AC2 落在 0.67–0.73,且训练可改善(PMID 19538100)、量表化可行。双方说的是同一批数据,差别在于用了哪个统计量。 这与本篇的形状完全同构:判据漏掉一个量,结论就从「安全」翻成「不达标」,而数据一动没动。

    所以规矩要改。上一轮设想「改二元表」或「非对称加 n×1.1」——现在看不必要,判据线不动,表不用改。要补的只有一件事:报 E。现行报告只给 m_A、m_B(或 m̄ 与 δ),这两个数加起来等价于 s_A+s_B,但 E 是另一根轴,现有报告项里完全没有它。补法很便宜:两位读者把剩余质量分给第二档的比例,各报一个数,平均值就是 E。

    还有一条更彻底的路,连 ε 和对称假设都不必碰:TVD = δ ⟺ max(|p₂−q₂|, |p₃−q₃|) ≤ δ。三条边际差里主档差 δ 已知,剩下两条只要都不超过它,TVD 就被 δ 焊住。这条判据全程只用观测到的边际分布,不引入未实测参数,也不区分谁是 A 谁是 B,天然对称。40 万组随机边际核验,与原判据零不一致。要多精确有多精确,且不需要谁来先估 ε。

    一句话收口:判据线不是被「两人占比不对等」推走的,是被「两人一起往同一边偏」推走的;前者恒等不变,后者是 1+2E。担心的那个方向不存在,漏掉的那个量一直没人报。

    证据与局限

    已验证:ε* = δ/(2(s_A+s_B)) ≡ δ/(4(1−m̄)) 为恒等式(全域网格偏差 0,20 万组随机主档最大偏差 1.4×10⁻¹⁵);s_A − s_B ≡ −δ,故「对称参数族」只在 δ=0 时成立;分解式 w = s̄Δ − δE 配合 TVD = max(δ, |w|+δ/2) 在 20 万组随机边际上误差 <1.2×10⁻¹⁶;放宽倍数 1+2E 与 r 无关(六组 (m̄,δ,E) 逐位吻合);纯观测判据 max(|p₂−q₂|,|p₃−q₃|) ≤ δ 与 TVD=δ 等价(40 万组零不一致,另 30 万组独立核验亦零);κ_max 跌落 E=+0.10 为 0.0003、E=−0.10 为 0.0271(m̄=0.90, δ=0.05)。

    部分验证:全部推导限于三档、两位读者。四档及以上 w 的分解形式会多一个自由度,1+2E 是否仍成立未推导。E 与 Δ 在真实数据里是否独立未知——若两位读者的偏移由同一批材料难度驱动,E 会与 δ 相关,此时 r 便不能从上界里消掉,1+2E 需重算。

    待验证:真实三档裁决研究中 E 的实测分布,本文只能给「若 |E| ≤ 0.10 则影响在 0.03 个 κ 点内」的判别;E 是否与读者资历、判读顺序、有无回看机制相关,完全没查;填充材料与间隔设计会不会系统性制造 E(与本线上一篇同源,尚未分离)。

    相关阅读

    L1 网络状态:CORRESPONDENCE_NETWORK.md §3.8「E-060 标签前提」原文——「KB-292(conflict):中医医师间诊断一致性在开放人群中「低(low)」,作者称为「令人担忧」(PMID 30383398);同系列第二篇指出偶然校正一致性统计常被误用、建议 Gwet AC2(PMID 31730403 / 31634000 / 31730404);但训练可改善(PMID 19538100)、量表化可行(15165421 阴阳评分法 / 23762155 TEAMSI-TCM) → 「用五脏标签做分组」这一步本身带噪声:观察到的「五脏×HRV」差异可能只是标签噪声」,层级栏「冲突(未裁决)」,状态「conflict」。该边为冲突态,本文已按规则并列写出冲突双方(低一致性 vs 统计误用导致低估),不作单方结论。

    证据等级: C
    参考文献: PMID 30383398, PMID 31730403, PMID 19538100, PMID 26095449, PMID 8501467, PMID 8621993, PMID 2348207


    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。

    Asymmetric reader proportion, static criterion line: the true driver is \”which way both lean,\” with a multiplier of exactly 1+2E

    Evidence level: C

    Previous concerns worried that asymmetry might displace the criterion line. Calculation shows it does not move—not even one millimeter. However, the same calculation reveals another omitted term, whose multiplier is cleanly 1+2E, independent of the primary proportion and primary difference.

    Main Text

    The previous post left an open case. The criterion line $2\epsilon^* = \delta/(4(1-\bar{m}))$ was derived from the parameter family where “two readers’ primary proportions symmetrically bracket the mean.” In real research, scenarios like one reader at 0.95 and another at 0.85 are common, raising a reasonable concern: since they are asymmetric, the residual spaces $(1-m_A)$ and $(1-m_B)$ are no longer equal; could the criterion line shift? If depressed by more than 10%, studies designed according to the symmetric table (where $n=49$ suffices when $\bar{m}=0.70$) would actually have insufficient sample size, yet those studies proceed as if the table applies.

    First, fix this line.

    The original closed form is $\epsilon^ = \delta/(2(s_A+s_B))$, with the denominator being their sum. Substituting $s_A = 1-\bar{m}-\delta/2$ and $s_B = 1-\bar{m}+\delta/2$, the sum always equals $2(1-\bar{m})$—the $-\delta/2$ and $+\delta/2$ exactly cancel. Thus $\delta/(2(s_A+s_B)) \equiv \delta/(4(1-\bar{m}))$, this is an identity, not an approximation*. Scanning the full grid $\bar{m}\in[0.55,0.99]$, $\delta\in[0.02,0.15]$ yields zero deviation; randomly sampling 200,000 independent primary proportion pairs gives a maximum deviation of $1.4\times10^{-15}$—pure floating-point error.

    The conclusion is somewhat anticlimactic: asymmetry does not shift the criterion line at all, not even a fraction. The previous worry about “which direction it shifts” has the answer “no direction.”

    Following this identity further reveals something more awkward.

    The parameter family $m_A = \bar{m}+\delta/2$, $m_B = \bar{m}-\delta/2$ is called “symmetric,” yet it yields $s_A – s_B \equiv -\delta$. Only when $\delta=0$ is it truly symmetric; as long as the two readers’ primary proportions differ, their residual spaces necessarily differ by $\delta$, without exception. That means there never exists a special “symmetric case”—that formula is always the general form, merely mislabeled. This precisely explains why it does not shift: it is not derived from a symmetry assumption; the cancelling terms are algebraic structure, not coincidence.

    The real omission lies elsewhere.

    Expanding the two marginal differences yields a clean decomposition: $w = \bar{s}\Delta – \delta E$. Here $\bar{s}$ is the mean residual space, $\Delta = e_A-e_B$ is the difference in deviation directions, and $E=(e_A+e_B)/2$ is the joint offset—whether both readers shift toward the second tier or both toward the third tier. Since $\text{TVD} = \max(\delta, |w|+\delta/2)$, the condition “not crossing the line” is $|w| \le \delta/2$.

    The current criterion $|\Delta| \le 2\epsilon^$ writes only the first term of $w$, discarding the entire $-\delta E$ term. It holds only if $E=0$—the two readers’ deviations are exactly opposite in sign and equal in magnitude. This criterion is not wrong; it is precise only under a coincidence of zero measure*.

    Restoring $\delta E$ yields an elegant solution shape: the upper bound for allowed $\Delta$ shifts from $r/2$ to $r(E+1/2)$, where $r=\delta/\bar{s}$ cancels entirely. Whether the criterion should relax or tighten depends solely on “which direction both readers shift together,” with a factor of $1+2E$. Six different $(\bar{m},\delta,E)$ combinations verified numerically, matching $1+2E$ digit-for-digit.

    Direction matters far more than magnitude. When $E>0$ (both shift toward the second tier), the criterion becomes conservative, leaving extra margin; $E=0.10$ relaxes it by only 20%, nearly harmless. When $E<0$ (both shift toward the third tier), the criterion becomes overly permissive; studies actually failing standards are allowed through: $E=-0.05$ cuts by 10%, $E=-0.20$ cuts by 40%, $E=-0.50$ cuts to 0—the criterion completely fails, leaving no safety zone.

    Translating to the $\kappa$ ceiling is even more striking. With $\bar{m}=0.90$, $\delta=0.05$, and $\Delta$ pinned at the criterion line: $E=+0.10$ drops $\kappa_{\max}$ from 0.7315 to 0.7312, a drop of 0.0003, negligible; $E=-0.10$ drops it to 0.7044, a drop of 0.0271, 2.7 $\kappa$ points. The same “deviation magnitude,” switched in direction, incurs ninety times the cost.

    Here we must connect to the unresolved conflict edge of the research core: our entire network rests on a grouping premise—”label subjects using five-organ differentiation”—yet this premise itself is in conflict. The pessimistic side is that diagnostic consistency among TCM physicians is “low” in open populations, termed “concerning” by the original authors (PMID 30383398); the optimistic side comes from the same series—that this “low” conclusion stems largely from statistical misuse, and recalculating with Gwet AC2 yields linearly weighted agreement of 0.80, chance-corrected AC2 falling between 0.67–0.73, with training improving it (PMID 19538100) and scale-based assessment feasible (15165421 Yin-Yang Scoring Method / 23762155 TEAMSI-TCM). Both sides refer to the same data; the difference lies in which statistic is used. This structure is fully isomorphic to our current paper: omitting one quantity flips the conclusion from “safe” to “non-compliant,” while the data itself does not move.

    Therefore, the rules must change. Previous proposals suggested “modifying the binary table” or “asymmetric $n\times1.1$”—now these appear unnecessary, since the criterion line does not shift and the table need not change. The only thing to add is: report $E$. Current reports provide only $m_A$, $m_B$ (or $\bar{m}$ and $\delta$); these two numbers sum to $s_A+s_B$, but $E$ is another axis entirely, absent from existing report items. Adding it is cheap: each reader reports the proportion of residual mass assigned to the second tier; their average is $E$.

    There is an even more thorough path, avoiding $\epsilon$ and the symmetry assumption altogether: $\text{TVD} = \delta \iff \max(|p_2-q_2|, |p_3-q_3|) \le \delta$. Among the three marginal differences, the primary difference $\delta$ is known; the other two need only not exceed it for TVD to be locked at $\delta$. This criterion uses only observed marginal distributions throughout, introduces no unmeasured parameters, and is naturally symmetric (does not distinguish who is A or B). Verification of 400,000 random margins shows zero inconsistency with the original criterion. The more precise it gets, the less need for anyone to estimate $\epsilon$ first.

    In summary: the criterion line is not shifted by “unequal reader proportions,” but by “both readers shifting together in the same direction”; the former is invariant, the latter scales as $1+2E$. The feared direction does not exist; the omitted quantity has never been reported.

    Evidence and Limitations

    Verified: $\epsilon^* = \delta/(2(s_A+s_B)) \equiv \delta/(4(1-\bar{m}))$ is an identity (full grid deviation 0, max deviation among 200,000 random primaries $1.4\times10^{-15}$); $s_A – s_B \equiv -\delta$, so “symmetric parameter family” holds only when $\delta=0$; decomposition $w = \bar{s}\Delta – \delta E$ with $\text{TVD} = \max(\delta, |w|+\delta/2)$ yields error $<1.2\times10^{-16}$ on 200,000 random margins; relaxation factor $1+2E$ is independent of $r$ (six $(\bar{m},\delta,E)$ combinations matched digit-for-digit); purely observational criterion $\max(|p_2-q_2|,|p_3-q_3|) \le \delta$ is equivalent to $\text{TVD}=\delta$ (400,000 random margins zero inconsistency, plus independent verification of 300,000 additional cases also zero); $\kappa_{\max}$ drop for $E=+0.10$ is 0.0003, for $E=-0.10$ is 0.0271 ($\bar{m}=0.90$, $\delta=0.05$).

    Partially Verified: All derivations are limited to three tiers and two readers. For four or more tiers, the decomposition of $w$ gains an extra degree of freedom; whether $1+2E$ still holds has not been derived. Whether $E$ and $\Delta$ are independent in real data is unknown—if both readers’ offsets are driven by the same material difficulty, $E$ would correlate with $\delta$, preventing $r$ from cancelling out of the upper bound, requiring recalculation of $1+2E$.

    To Be Verified: The empirical distribution of $E$ in real three-tier adjudication studies; this paper can only state that “if $|E| \le 0.10$, the impact is within 0.03 $\kappa$ points.” Whether $E$ correlates with reader seniority, adjudication order, or presence of re-reading mechanisms remains unexamined; whether filler materials and interval design systematically generate $E$ (sharing origin with the previous line, not yet separated) is unknown.

    Related Reading

    L1 Network Status: CORRESPONDENCE_NETWORK.md §3.8 “E-060 Label Premise” original text—”KB-292 (conflict): Diagnostic consistency among TCM physicians is ‘low’ in open populations, termed ‘concerning’ by authors (PMID 30383398); the second paper in the same series notes that chance-corrected consistency statistics are often misused, recommending Gwet AC2 (PMID 31730403 / 31634000 / 31730404); yet training can improve it (PMID 19538100) and scale-based assessment is feasible (15165421 Yin-Yang Scoring Method / 23762155 TEAMSI-TCM) → “The step of grouping by five-organ labels inherently carries noise: observed ‘five-organ×HRV’ differences may merely reflect label noise“, level column “Conflict (Unresolved)“, status “conflict“. This edge is in conflict state; this paper lists both conflicting parties (low consistency vs. statistical misuse leading to underestimation) side-by-side per rules, without unilateral conclusion.

    Evidence Level: C
    References: PMID 30383398, PMID 31730403, PMID 19538100, PMID 26095449, PMID 8501467, PMID 8621993, PMID 2348207


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

  • 那个「从来没人估过的先验」,其实就是每篇论文表一里的原始一致率:把内生量当成外生输入,才会觉得它永远缺数据

    Evidence level: C

    上一轮说「这个决定一切的比例 π 从未被估过」。它确实没被当成 π 估过——但它的数值每篇一致性研究的表一里都写着,因为它就是 1 减去原始一致率。

    一、先把那个「从没被估过」的量摊开看

    上一轮(#1397)把「该不该给判读差值打折」压在一个叫 π 的占比上:被判读的两份结果里真有差的比例。它给了张曲面,ρ=0.5 时门槛 0.178、ρ=0.9 时 0.054,并自我批评:这个 π 在任何真实研究里都没被估过,是整条推理里唯一没有数据支撑的量。

    这个自我批评很诚实,但只走了一半——它没问这个量在定义上能不能被估。

    判读差 Δ 是连续量(两个 kappa 之差)。若 π 定义为「Δ 不等于零的比例」,那么 Δ 恰好等于零的概率精确为 0——是测度论的必然。于是 π 恒等于 1,整张曲面退化成左端一个点,所有判决都变成「该打折」。

    那么实际在用的 0.25 和 0.5 是什么?是设定,不是估计。原文写「退一步取 π=0.5」「取 π=0.25(偏保守的合理值)」——语气是估计,实质是设定。

    二、降一个层次,π 立刻变成大家都报过的数

    把层次从「两项研究」降到「一对判读」,π 就有了实体:同一批病例上两位读者给出档位不同的比例。它有更常见的名字——1 减去原始一致率(Po)。

    这是恒等式,不是近似:π = P(档位差 ≠ 0) = 1 − Po。

    于是「π 从未被估过」在事实层面不成立。每篇一致性研究的表一都写着 Po,因为它是 kappa 公式的分子项。它不缺数据,只是没被单独命名。

    这也解释了那个奇怪现象:为什么 π 取 0.25 还是 0.5,结论几乎不变。因为 π 的真实值由 Po 决定,而 Po 在各研究里落在 0.60 到 0.85,即 π 落在 0.15 到 0.40。门槛 0.054 和 0.178 都落在此区间下方——判决对 π 的不精确几乎免疫。

    三、这个免疫不是巧合,是结构性的

    Po 对应 kappa π = 1−Po 门槛 π* π 是否过门槛
    0.90 0.847 0.100 0.178 不过
    0.85 0.771 0.150 0.178 不过
    0.80 0.695 0.200 0.178 过
    0.70 0.542 0.300 0.178 过
    0.60 0.389 0.400 0.178 过
    0.55 0.313 0.450 0.178 过

    看走向:Po 越高,kappa 越高,而 π 越低。π 与判读质量是同一根轴上的两个刻度,方向相反。判读可靠时分歧罕见,本来就没什么可打的;判读不可靠时分歧普遍,打折一定划算。

    所以上一轮的两道前置判据其实是同一道:一道用 kappa 之差算比值 z 须落在 0.7 到 2.5,另一道要求 π 超过门槛。前者测「真差够不够大」,后者测「分歧够不够多」,而这两件事由同一个可靠性参数决定。它们是同一扇门的两把锁,锁的是同一根闩。

    因此真正没被报告的不是 π,而是从一致率走到 kappa 的中间量:边际分布。同样 Po=0.80,边际均衡时 kappa 约 0.70,极端偏斜时可低到 0.33。

    四、证据与局限

    已验证:① 恒等式 π = 1 − Po 成立,「该量从未被估」在定义层面不成立;② 上一轮的 π* 曲面可由收益与假阳性代价复现(ρ=0.5 时 4.44/(4.44+20.57)=0.1775;ρ=0.9 时 5.00/(5.00+88.42)=0.0535,与原文一致);③ 判决翻转发生在 Po 高于约 0.82(kappa 约 0.73)处,而实际研究 kappa 中位远低于此,故常见域内判决对 π 免疫;④ kappa 悖论(高 Po 配低 kappa)根源是边际分布而非抽样。

    部分验证:① 上表用三档均等边际作简化,真实边际偏斜时曲线整体平移但方向不变;② 「实际 kappa 中位」依赖样本量调查的 71 例中位,该调查未统计 kappa 分布,属间接推断;③ 两道判据同源的论证用可靠性参数作代理,delta 与 ρ 的相关结构未被直接测过。

    待验证:① 若把 π 定为「|Δ| 超过某等效界 c 的比例」而非「Δ 不为零」,π 就变成 c 的函数,须先定 c,而 c 取多少无任何规范给出——这是新识别的问题;② 两道判据在临界带(Po 约 0.78 到 0.82)可能给出不一致判决,该段未刻画。

    L1 网络状态:EDGE_TABLE 共 47 个概念(频率、HRV、心率、呼吸、自主神经、共振、谐波、五行、五脏等),无「先验概率 / 真差比例 / 原始一致率 / 边际分布」任何条目;CORRESPONDENCE_NETWORK 两轮检索(判读/一致性/信度/混合/成分/真值/先验/比例/金标准)仅命中 §3.8 与 §3.9「E-060 标签前提」条:「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);同系列第二篇指出偶然校正一致性统计常被误用、建议 Gwet AC2(PMID 31730403 / 31634000 / 31730404);但训练可改善(PMID 19538100)」——该边状态为 conflict(未裁决)。本文讨论对象(先验参数是否内生、Po 是否等于 1−π)与该边争议不重叠:那条争的是「标签准不准」,本文问「参数是否独立」。故本文不构成对 E-060 的新证据,亦不改判其状态。

    五、这条线该怎么收

    一句话:不是补一个 π 的估计,是取消 π 这个参数。

    第一,凡引用「打折门槛 π*」处,一律同时报 Po。 π=1−Po,报 Po 就等于报了 π,且每篇都有。禁止报 kappa 却不提 Po。

    第二,把「π 是先验」这说法从规范里删掉。 决策曲线框架说阈值概率该由情境给出,指的是收益与代价的交换率,不是真差比例。后者是内生的,由同一批数据的可靠性决定。把内生量写进「需外部给值」栏,就永远觉得缺数据。

    第三,真正该补的报告项是边际分布。 Po 到 kappa 的映射需要它,而它正是一致性研究最常省的那一栏。这跟前几轮汇合:κ 的上限只由熵与总变差距离决定,而这两者本就是边际的两种写法。

    相关阅读

    证据等级: C
    参考文献: PMID 8621993, PMID 2348207, PMID 25168681, PMID 8501467, PMID 15733050, PMID 17099194, PMID 26810254, PMID 36123642, PMID 30383398, PMID 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。

    That \”prior never estimated before\” is simply the raw consistency rate in Table 1 of every paper: treating endogenous variables as exogenous inputs makes it seem perpetually dat…

    Evidence level: C

    In the previous iteration, it was stated that “the proportion π determining everything has never been estimated.” Indeed, it was not estimated as π—but its numerical value is explicitly written in Table One of every agreement study, because it equals 1 minus the raw agreement rate.

    I. First, Lay Out That “Never-Estimated” Quantity

    The previous iteration (#1397) placed the question of “whether to discount the judgment difference” on a proportion called π: the fraction of judged pairs that truly differ. It provided a surface where, at ρ=0.5, the threshold is 0.178 and at ρ=0.9 it is 0.054, while self-critiquing: this π has never been estimated in any real study and is the only quantity in the entire reasoning chain lacking data support.

    This self-critique is honest but only half-complete—it did not ask whether this quantity can be estimated by definition.

    The judgment difference Δ is a continuous measure (the difference between two kappa values). If π is defined as “the proportion where Δ ≠ 0”, then the probability that Δ exactly equals zero is precisely 0—this is a necessity of measure theory. Thus π is always equal to 1, and the entire surface degenerates into a single point on the left end; all judgments become “should be discounted.”

    So what are the actual values of 0.25 and 0.5 used in practice? They are assumptions, not estimates. The original text states “taking π=0.5 as a step back” and “taking π=0.25 (a conservative reasonable value)”—the tone suggests estimation, but the substance is specification.

    II. Drop One Level, and π Immediately Becomes a Number Everyone Has Reported

    By reducing the level from “two studies” to “one pair of judgments,” π acquires substance: the proportion where two readers give different grades on the same set of cases. It has a more common name—1 minus the raw agreement rate (Po).

    This is an identity, not an approximation: π = P(grade difference ≠ 0) = 1 − Po.

    Thus “π has never been estimated” is factually incorrect. Every Table One in every agreement study lists Po, because it is the numerator term of the kappa formula. It does not lack data; it was simply not named separately.

    This also explains that strange phenomenon: why π=0.25 or π=0.5 yields almost identical conclusions. Because the true value of π is determined by Po, and Po across studies ranges from 0.60 to 0.85, meaning π ranges from 0.15 to 0.40. The thresholds 0.054 and 0.178 both fall below this interval—judgments are nearly immune to imprecision in π.

    III. This Immunity Is Not a Coincidence; It Is Structural

    Po Corresponding kappa π = 1−Po Threshold π* Does π exceed threshold?
    0.90 0.847 0.100 0.178 No
    0.85 0.771 0.150 0.178 No
    0.80 0.695 0.200 0.178 Yes
    0.70 0.542 0.300 0.178 Yes
    0.60 0.389 0.400 0.178 Yes
    0.55 0.313 0.450 0.178 Yes

    Observe the trend: as Po increases, kappa increases, and π decreases. π and judgment quality are two marks on the same axis, moving in opposite directions. When judgments are reliable, disagreements are rare, so there is little to discount; when judgments are unreliable, disagreements are common, so discounting is certainly worthwhile.

    Thus the two preconditions from the previous iteration were actually one and the same: one uses the kappa difference ratio z to fall between 0.7 and 2.5, while the other requires π to exceed the threshold. The former measures “whether the true difference is large enough,” and the latter measures “whether disagreements are frequent enough,” but both are governed by the same reliability parameter. They are two keys to the same door, locking the same latch.

    Therefore what was never reported is not π, but the intermediate quantity between agreement rate and kappa: the marginal distribution. With Po=0.80, the marginal equilibrium yields kappa ≈ 0.70, but under extreme skew it can drop as low as 0.33.

    IV. Evidence and Limitations

    Verified: ① The identity π = 1 − Po holds, so “this quantity was never estimated” is false by definition; ② The previous π* surface can be reproduced from benefit and false-positive cost (ρ=0.5 gives 4.44/(4.44+20.57)=0.1775; ρ=0.9 gives 5.00/(5.00+88.42)=0.0535, matching the original text); ③ Judgment reversal occurs when Po exceeds approximately 0.82 (kappa ≈ 0.73), while actual study kappa medians are far lower, so common-domain judgments are immune to π; ④ The kappa paradox (high Po paired with low kappa) stems from marginal distribution, not sampling.

    Partially Verified: ① The table above uses three equal marginal bins as simplification; under true marginal skew, the curve shifts overall but maintains direction; ② “Actual kappa median” depends on the 71-case median from sample size survey, which did not estimate kappa distribution, making this an indirect inference; ③ The argument that both criteria share origin uses reliability parameter as proxy, but the correlation structure between delta and ρ was not directly measured.

    To Be Verified: ① If π is defined as “the proportion where |Δ| exceeds some equivalent boundary c” rather than “Δ ≠ 0”, then π becomes a function of c, requiring c to be set first, yet no norm specifies how much c should be—this is a new identification problem; ② In the critical band (Po approximately 0.78 to 0.82), the two criteria may yield inconsistent judgments, a region not characterized.

    L1 Network Status: EDGE_TABLE contains 47 concepts (frequency, HRV, heart rate, respiration, autonomic, resonance, harmonic, Five Phases, five organs, etc.), with no entries for “prior probability / true difference proportion / raw agreement rate / marginal distribution”; CORRESPONDENCE_NETWORK two-round retrieval (judgment/agreement/reliability/mixed/component/true value/prior/proportion/gold standard) only hit §3.8 and §3.9 “E-060 label premise” entries: “Diagnostic agreement among TCM physicians in the open population is ‘low (low)’, termed ‘concerning’ (PMID 30383398); the second paper in the series notes that chance-corrected agreement statistics are often misused, recommending Gwet AC2 (PMID 31730403 / 31634000 / 31730404); but training can improve (PMID 19538100)”—this edge status is conflict (unresolved). The subject of this discussion (whether the prior parameter is endogenous, whether Po equals 1−π) does not overlap with this edge controversy: that dispute concerns “are the labels accurate,” while this paper asks “are the parameters independent.” Hence this paper does not constitute new evidence for E-060, nor does it change its status.

    V. How This Line Should Be Closed

    In one sentence: not to add an estimate of π, but to eliminate the parameter π itself.

    First, wherever “discount threshold π*” is cited, always report Po simultaneously. Since π=1−Po, reporting Po equals reporting π, and every study has it. Do not report kappa without mentioning Po.

    Second, remove the claim “π is a prior” from the framework. The decision curve framework states that threshold probability should be given by context, referring to the exchange rate of benefit and cost, not the true difference proportion. The latter is endogenous, determined by the reliability of the same data. Writing an endogenous quantity into “needs external assignment” makes it appear perpetually data-deficient.

    Third, the truly missing report item is the marginal distribution. The mapping from Po to kappa requires it, and this is precisely the column most often omitted in agreement studies. This converges with earlier iterations: the upper limit of κ is determined solely by entropy and total variation distance, both of which are simply different expressions of the marginal.

    Related Reading

    Evidence Level: C
    References: PMID 8621993, PMID 2348207, PMID 25168681, PMID 8501467, PMID 15733050, PMID 17099194, PMID 26810254, PMID 36123642, PMID 30383398, PMID 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.

  • 那把用来分「买得回 / 买不回」的尺子,零刻度压根不在 1:地板是 √(2−2ρ),不是零

    Evidence level: C

    上一篇说 δ 那一半买不回来、ρ 那一半能买回来。要验这条二分,本要拿「批次间方差 ÷ 二项预期」去称 δ 稳不稳。可这个比值在纯噪声下的期望不是 1,是 √(2−2ρ);ρ=0 时地板 1.414,而判据线画在 1.5——离地板只有 6%。

    正文

    上一篇(#1396)给出了一条很硬的分界:可辨差 D_min 拆成 C(δ)=δ/(1−p_e) 和 1.96√(2−2ρ)·SE 两项,前一项里没有 n、后一项与 √n 成反比,所以 δ 买不回来、ρ 买得回来。这条二分若成立,本该能用实测检验:找几位固定读者、跑若干独立判读批次,逐批算 δ,看批次间那点散布是「只有抽样噪声」还是「藏着稳定的个人风格」。

    检验的算法很自然——批次间标准差 ÷ 二项抽样预期。比值接近 1,说明 δ 全是噪声、可补偿;远大于 1,说明有稳定成分、不可补偿。判据也就顺势定在 1.5 和 3.0 两档。

    这个统计量看着无害,但它有一个从头到尾没人算过的性质:它的零刻度不在 1。

    一、地板是 √(2−2ρ),不是零

    两位读者对同一批 n 个案子判档,各自的主档比例是 p̂_A、p̂_B,δ̂ 就是两者相减。这个差的方差是

    Var(δ̂) = [p_A(1−p_A) + p_B(1−p_B) − 2ρ√(p_A(1−p_A)p_B(1−p_B))] / n

    p_A=p_B=m 时收缩成 2m(1−m)(1−ρ)/n。而分母用的「二项抽样预期」通常按 √(2m(1−m)/n) 算,等于假定两位读者的判读互相独立(ρ=0)。于是两者一比:

    比值 = √(2(1−ρ))

    只要 ρ>0,比值就小于 √2,但不等于 1。 反过来看更刺眼:如果分母老老实实按独立算,那么即使 δ 完全没有任何稳定成分,比值也停在 √(2−2ρ)。ρ=0 时是 1.414,ρ=0.5 时是 1.000,ρ=0.8 时只有 0.632。

    也就是说,这个统计量的零点由相依结构决定,而相依结构是另一个参数。 你要用它判断 δ 稳不稳,前提是先知道 ρ;可 ρ 本身在多数研究里也没测过。

    实测核对:按阈值潜变量模型生成两位读者的三档判读(m=0.90、n=120、三千次重抽),σ 观测值与上式预测的比值在 0.988–1.018 之间,公式本身没错。错的是拿它当「有没有稳定成分」的尺子。

    二、1.5 这条线只在地板上方 6%

    把判据线画在 1.5,等于默认零点在 1、留 50% 的余量。真实零点在 1.414(ρ=0 时),余量只剩 1.5/1.414=1.061,即 6%。

    后果是灾难性的。纯噪声数据(τ=0,完全没有稳定成分)在 ρ=0 时越过 1.5 的概率:3 个批次 32.5%,5 个 34.3%,11 个 33.8%,51 个 25.3%,101 个 18.5%。加批次几乎压不下去——因为地板是系统性偏移,不是抽样抖动。到 500 个批次还有 2.7%。

    对照 ρ=0.5:地板 1.000,越 1.5 的概率 3 批 10.5%、11 批 1.3%、51 批基本为零。同样的数据、同样的判据,只因为读者间更同步,误判率差一个数量级。

    三、同一份 δ 会被判成三种结论

    更麻烦的是这个比值把两个量揉在了一起。分子里既有批次间的稳定散布 τ,又有抽样噪声;分母是 n 的函数。于是:

    τ(批次间真实散布) n=120 n=960 n=7200
    0.5pp 1.017 可补偿 1.125 可补偿 1.732 部分
    1.0pp 1.065 可补偿 1.438 可补偿 3.000 稳定
    2.0pp 1.238 可补偿 2.295 部分 5.745 稳定

    (m=0.90、ρ=0.5)τ=1.0pp 这一份数据,在 120 例时判「可补偿」、在 7200 例时判「稳定」。同一批读者、同一个规程、同一个真实世界,结论只随样本量翻面。这不是精度问题,是判据把「效应大小」和「测量精度」两个维度压成了一个数。

    反过来说:想用这条判据判出「稳定风格」,120 例、ρ=0.5 时 τ 得达到 3.06pp;要判到「稳」那一档,τ 需 7.75pp。而 #1396 划的绝对线是 δ = 0.05×(1−p_e),m=0.90 时只有 0.925pp。判据说「不够稳」的门槛,比它自己说「已经吃掉整个带宽」的门槛高 8 倍。*

    四、随机错位本身就在吃带宽

    这一条最要命,也最容易被漏掉。δ 对天花板的伤害不是靠它的均值,是靠它的绝对值——C(δ)=δ/(1−p_e),取的是正的 δ。若 δ 在批次间以 0 为中心随机摆动,它的期望绝对值严格大于零:

    E|δ| = τ√(2/π) ≈ 0.798τ

    m=0.90、1−p_e=0.185 时,换算成 C:τ=0.5pp 就吃掉 0.0216,占总预算 0.05 的 43%;τ=1.0pp 吃掉 0.0431,占 86%;τ=2.0pp 直接 173%,超出预算一倍多。

    这就是「纯随机错位」的代价:它没有均值偏移,看起来无害,但每批都在往天花板锯一刀。#1396 那条「δ 撞到结构上界 1−m̄ 的 10% 就锁死」的经验法则,只对固定 δ 成立;一旦 δ 随机化,小额散布就足以把预算吃干净,根本不用等到 10%。

    五、正确的做法

    三处都要改:

    ① 分母换成含 ρ 的方差。 比值应写成 S_b / √(2m(1−m)(1−ρ)/n),或在 ρ 未知时改用批次内配对差直接估分母,别用独立二项式。

    ② 换成对 τ² 的直接检验。 设 V₀=2m(1−m)(1−ρ)/n 是已知的纯噪声方差,则 (B−1)S_b²/V₀ 在 τ=0 时服从 χ²(B−1)。判 τ>0 的 95% 临界:3 批 2.996、5 批 2.372、8 批 2.010、12 批 1.789。这才是能用的检验。

    ③ 批次数的硬门槛是 18。 τ̂ 的 95% 上界倍数是 (B−1)/χ²₀.₀₅(B−1):3 批 19.5 倍、5 批 5.63 倍、8 批 3.23 倍、12 批 2.40 倍、18 批才降到 1.96 倍。3–5 个批次给出的 τ 上界可以宽到二十倍,对任何判决都没有约束力。原设想「≥3 组批次」远远不够。

    ④ 顺手补一条绝对上限。 判稳之前先算 C(E|δ|)=0.798τ/(1−p_e),超过 0.05 就别谈了——那时已经不是「稳不稳」的问题,是带宽已经被随机摆动吃掉。

    证据与局限

    已验证: 比值公式 √(2−2ρ) 的代数推导无近似;三档阈值模型蒙特卡洛(m=0.90、n=120、三千次)实测 SD 与预测比值 0.988–1.018。χ² 检验与上界倍数均为标准结论,直接可算。

    部分验证: 「δ 的批次间散布里稳定成分占多少」这个问题本身没有任何一组真实三档判读数据回答过。上文的 τ 门槛(120 例、ρ=0.5 时 3.06pp)是解析值,是否落在真实读者群的合理区间,取决于文献里 rater severity 的实测散布——而那一类研究报告的是 ICC 而不是 pp,换算链本身还有一层不确定。

    待验证:
    1. ρ 与 τ 是否独立。若读者间越不同步、δ 的批次散布也越大(同源),那么分母与分子共享同一份变异,比值会系统性偏向 1,地板与判据线一起失去意义。
    2. 批次之间是否真的独立。若批次共享同一批案子或同一套判读培训,批次间方差会低估,τ 上界算得过于乐观。
    3. 本节全部以「两位固定读者」为前提。换成多位读者轮换时,δ 的定义要先从「两人差」推广到「读者边际散布」,那时地板公式要重推。

    相关阅读

    证据等级: C
    参考文献: PMID 8501467, PMID 26925540, PMID 28587597, PMID 30067315, PMID 16955539, PMID 16807131, PMID 17698940, PMID 22769017, PMID 40665614

    L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 三域节点表(易学/中医侧、物理侧、生理/体验侧)无任何以「判读信度 / 边际偏倚 / 批次间方差 / 误差可补偿性」为源或目标的节点,边表止于 E-061「呼吸节律 0.1 Hz 谱峰 — 对应 — Mayer 波」(L2,open,语义为自主神经谱峰对应,与判读误差无关);全文件 grep「kappa / 一致性 / 信度 / 判读 / 偏倚 / 抽样 / 样本量 / 方差 / 批次 / 重复测 / 测量 / ICC / 随机效应 / 方差分解 / 收缩」两轮仅命中第 79 行层级定义、第 174 行 E-060「标签信度前提」(conflict 态,语义为五脏辨证作为分组标签的医师间一致性低,PMID 30383398 vs 31730403)、第 192 行「生克一致性」(指五行生克方向)、第 57 行「相位 / 节律」;EDGE_TABLE_science-yixue.md 仅 63 行,为 47 个概念的计数统计(气功 48 篇 / 最高级 A / A级 9),无「批次 / 可补偿性 / 方差成分」条目。


    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 ruler used to divide \”buyable / not buyable\” has its zero mark nowhere near 1: the floor is √(2−2ρ), not zero.

    Evidence level: C

    In the previous post, we noted that half of δ is irrecoverable while half of ρ is recoverable. To validate this binary division, one would normally weigh the stability of δ using “between-batch variance ÷ binomial expectation.” However, under pure noise, the expected value of this ratio is not 1; it is √(2−2ρ). When ρ=0, the floor is 1.414, yet the decision threshold is drawn at 1.5—only 6% above the floor.

    Main Text

    The previous post (#1396) established a rigid dividing line: the minimum discernible difference, D_min, decomposes into two terms, C(δ)=δ/(1−p_e) and 1.96√(2−2ρ)·SE. The first term contains no n, while the second is inversely proportional to √n; consequently, δ is irrecoverable but ρ is recoverable. If this binary division holds, it should be testable via empirical data: select a fixed set of readers, run several independent batches of readings, compute δ for each batch sequentially, and examine whether the observed scatter between batches reflects “sampling noise alone” or “a stable individual style hidden within.”

    The testing algorithm is straightforward—between-batch standard deviation ÷ binomial sampling expectation. A ratio near 1 indicates δ is purely noise and compensable; a ratio far exceeding 1 indicates a stable component that cannot be compensated. Thus, the criterion naturally falls into two tiers: 1.5 and 3.0.

    This statistic appears innocuous, yet it possesses a property that has never been calculated: its zero point is not 1.

    I. The Floor Is √(2−2ρ), Not Zero

    Two readers assess the same batch of n cases, yielding proportion estimates p̂_A and p̂_B respectively; δ̂ is their difference. The variance of this difference is

    Var(δ̂) = [p_A(1−p_A) + p_B(1−p_B) − 2ρ√(p_A(1−p_A)p_B(1−p_B))] / n

    When p_A=p_B=m, this contracts to 2m(1−m)(1−ρ)/n. The denominator, the “binomial sampling expectation,” is typically computed as √(2m(1−m)/n), which assumes the two readers’ assessments are mutually independent (ρ=0). Comparing the two:

    Ratio = √(2(1−ρ))

    As long as ρ>0, the ratio is less than √2 but not equal to 1. Conversely, this is even more striking: if the denominator is calculated strictly under independence, then even if δ has absolutely no stable component, the ratio settles at √(2−2ρ). When ρ=0, it is 1.414; when ρ=0.5, it is 1.000; when ρ=0.8, it drops to only 0.632.

    Thus, the zero point of this statistic is determined by the dependency structure, which itself is a separate parameter. To use it to judge whether δ is stable, one must first know ρ; yet ρ itself is rarely measured in most studies.

    Empirical verification: using a threshold latent variable model to generate three-tier assessments for two readers (m=0.90, n=120, 3,000 resamples), the observed σ ratio falls between 0.988 and 1.018, confirming the formula itself is correct. The error lies in using it as a ruler for “whether a stable component exists.”

    II. The Line at 1.5 Lies Only 6% Above the Floor

    Setting the decision threshold at 1.5 implicitly assumes the zero point is 1, leaving a 50% margin of safety. In reality, the zero point is 1.414 (when ρ=0), leaving only 1.5/1.414 = 1.061, or 6%.

    The consequences are catastrophic. Pure noise data (τ=0, with no stable component) exceeds 1.5 with probability: 32.5% for 3 batches, 34.3% for 5, 33.8% for 11, 25.3% for 51, and 18.5% for 101. Adding batches barely suppresses the rate—because the floor is a systematic offset, not sampling jitter. Even at 500 batches, 2.7% still exceed the threshold.

    Compare this to ρ=0.5: the floor is 1.000, so exceeding 1.5 occurs with probability 10.5% for 3 batches, 1.3% for 11, and essentially zero for 51. With the same data and the same criterion, merely because readers are more synchronized, the misjudgment rate differs by an order of magnitude.

    III. The Same δ Can Be Judged in Three Ways

    More problematically, this ratio conflates two quantities. The numerator contains both the stable between-batch scatter τ and sampling noise; the denominator is a function of n. Thus:

    τ (true between-batch scatter) n=120 n=960 n=7200
    0.5pp 1.017 compensable 1.125 compensable 1.732 partial
    1.0pp 1.065 compensable 1.438 compensable 3.000 stable
    2.0pp 1.238 compensable 2.295 partial 5.745 stable

    (m=0.90, ρ=0.5) The single dataset τ=1.0pp is judged “compensable” at n=120 but “stable” at n=7200. The same readers, the same protocol, the same real-world scenario yield conclusions that flip solely based on sample size. This is not a precision issue; it is because the criterion compresses “effect size” and “measurement precision”—two distinct dimensions—into a single number.

    Conversely, to judge “stable style” using this criterion: at n=120 and ρ=0.5, τ must reach 3.06pp; to be judged “stable,” τ requires 7.75pp. Yet #1396 sets the absolute line at δ = 0.05×(1−p_e); when m=0.90, this is only 0.925pp. The criterion’s threshold for “not stable enough” is eight times higher than the threshold it itself claims represents “consuming the entire bandwidth.”*

    IV. Random Misalignment Alone Consumes Bandwidth

    This point is the most critical and most easily overlooked. The damage δ inflicts on the ceiling is not via its mean, but via its absolute value—C(δ)=δ/(1−p_e), taking the positive δ. If δ fluctuates randomly around 0 between batches, its expected absolute value is strictly greater than zero:

    E|δ| = τ√(2/π) ≈ 0.798τ

    When m=0.90 and 1−p_e=0.185, converting to C: τ=0.5pp consumes 0.0216, which is 43% of the total budget 0.05; τ=1.0pp consumes 0.0431, or 86%; τ=2.0pp directly consumes 173%, exceeding the budget by more than double.

    This is the cost of “pure random misalignment”: it has no mean shift, appearing harmless, yet every batch saws into the ceiling. #1396’s empirical rule—”δ locks out when it hits 10% of the structural upper bound 1−m̄”—holds only for fixed δ; once δ is randomized, even small scatter can consume the entire budget without waiting for 10%.

    V. The Correct Approach

    Three adjustments are required:

    ① Replace the denominator with a variance containing ρ. The ratio should be written as S_b / √(2m(1−m)(1−ρ)/n), or, when ρ is unknown, estimate the denominator directly from within-batch paired differences rather than using independent binomials.

    ② Switch to a direct test of τ². Let V₀=2m(1−m)(1−ρ)/n be the known pure-noise variance; then (B−1)S_b²/V₀ follows a χ²(B−1) distribution when τ=0. The 95% critical value for τ>0 is: 2.996 for 3 batches, 2.372 for 5, 2.010 for 8, and 1.789 for 12. This is the usable test.

    ③ The hard threshold for batch count is 18. The 95% upper-bound multiplier for τ̂ is (B−1)/χ²₀.₀₅(B−1): 19.5× for 3 batches, 5.63× for 5, 3.23× for 8, 2.40× for 12, and only at 18 batches does it drop to 1.96×. The upper bound for τ given 3–5 batches can be as wide as twenty times, rendering it uninformative for any decision. The original notion of “≥3 batches” is far insufficient.

    ④ Add an absolute upper limit. Before judging stability, compute C(E|δ|)=0.798τ/(1−p_e); if this exceeds 0.05, stop—the issue is no longer “stable or not,” but that the bandwidth has been consumed by random fluctuation.

    Evidence and Limitations

    Verified: The algebraic derivation of the ratio formula √(2−2ρ) involves no approximation; the Monte Carlo simulation of the three-tier threshold model (m=0.90, n=120, 3,000 resamples) confirms the observed SD matches the predicted ratio between 0.988 and 1.018. The χ² test and upper-bound multipliers are standard conclusions, directly calculable.

    Partially Verified: The question “what proportion of the stable component lies within the between-batch scatter of δ” has never been answered by any real three-tier assessment dataset. The τ thresholds above (3.06pp at n=120, ρ=0.5) are analytical values; whether they fall within the reasonable range of actual readers depends on empirical scatter in literature-reported rater severity—yet such reports give ICC rather than pp, introducing an additional uncertainty in the conversion chain.

    Pending Verification:
    1. Whether ρ and τ are independent. If readers become less synchronized, the between-batch scatter of δ increases (homogeneous), meaning the numerator and denominator share the same variation, causing the ratio to systematically approach 1; both the floor and the criterion line lose meaning.
    2. Whether batches are truly independent. If batches share the same cases or the same training protocol, between-batch variance will be underestimated, making the τ upper bound overly optimistic.
    3. This section assumes “two fixed readers.” When switching to rotating multiple readers, the definition of δ must expand from “difference between two people” to “reader marginal scatter,” requiring a re-derivation of the floor formula.

    Related Reading

    Evidence Level: C
    References: PMID 8501467, PMID 26925540, PMID 28587597, PMID 30067315, PMID 16955539, PMID 16807131, PMID 17698940, PMID 22769017, PMID 40665614

    L1 Network Status: This topic has not yet entered the L1 corresponding network—the CORRESPONDENCE_NETWORK.md three-domain node table (Yijing Studies/TCM side, physics side, physiology/experience side) lacks any nodes sourced or targeted by “assessment reliability / marginal bias / between-batch variance / compensability of error.” The edge table stops at E-061 “respiratory rhythm 0.1 Hz spectral peak — corresponds to — Mayer wave” (L2, open, meaning autonomic nervous system spectral peak correspondence, unrelated to assessment error); grepping the full file for “kappa / consistency / reliability / assessment / bias / sampling / sample size / variance / batch / repeated measurement / measurement / ICC / random effect / variance decomposition / shrinkage” yields only the level definition at line 79, line 174 E-060 “label reliability premise” (conflict state, meaning low inter-rater consistency in organ differentiation as grouping labels, PMID 30383398 vs 31730403), line 192 “generating and controlling consistency” (referring to Five Phases generating and controlling direction), and line 57 “phase / rhythm”; EDGE_TABLE_science-yixue.md contains only 63 lines, counting statistics of 47 concepts (Qigong 48 papers / highest level A / Level A 9), with no entries for “batch / compensability / variance component.”


    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.